Google Git
Sign in
nest-open-source / nest-android-app / ffmpeg / 8e1cc55b733ae723de9931f20a13731bed8162f5 / . / libavcodec / vp9.c
blob: 5d8ad12b211fe34c420a264fc5bd8a6e0b8bdaea [file] [log] [blame]
/*
* VP9 compatible video decoder
*
* Copyright (C) 2013 Ronald S. Bultje <rsbultje gmail com>
* Copyright (C) 2013 Clément Bœsch <u pkh me>
*
* This file is part of FFmpeg.
*
* FFmpeg is free software; you can redistribute it and/or
* modify it under the terms of the GNU Lesser General Public
* License as published by the Free Software Foundation; either
* version 2.1 of the License, or (at your option) any later version.
*
* FFmpeg is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
* Lesser General Public License for more details.
*
* You should have received a copy of the GNU Lesser General Public
* License along with FFmpeg; if not, write to the Free Software
* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
*/
#include "avcodec.h"
#include "get_bits.h"
#include "internal.h"
#include "profiles.h"
#include "thread.h"
#include "videodsp.h"
#include "vp56.h"
#include "vp9.h"
#include "vp9data.h"
#include "vp9dsp.h"
#include "libavutil/avassert.h"
#include "libavutil/pixdesc.h"
#define VP9_SYNCCODE 0x498342
struct VP9Filter {
uint8_t level[8 * 8];
uint8_t /* bit=col */ mask[2 /* 0=y, 1=uv */][2 /* 0=col, 1=row */]
[8 /* rows */][4 /* 0=16, 1=8, 2=4, 3=inner4 */];
};
typedef struct VP9Block {
uint8_t seg_id, intra, comp, ref[2], mode[4], uvmode, skip;
enum FilterMode filter;
VP56mv mv[4 /* b_idx */][2 /* ref */];
enum BlockSize bs;
enum TxfmMode tx, uvtx;
enum BlockLevel bl;
enum BlockPartition bp;
} VP9Block;
typedef struct VP9Context {
VP9SharedContext s;
VP9DSPContext dsp;
VideoDSPContext vdsp;
GetBitContext gb;
VP56RangeCoder c;
VP56RangeCoder *c_b;
unsigned c_b_size;
VP9Block *b_base, *b;
int pass;
int row, row7, col, col7;
uint8_t *dst[3];
ptrdiff_t y_stride, uv_stride;
uint8_t ss_h, ss_v;
uint8_t last_bpp, bpp, bpp_index, bytesperpixel;
uint8_t last_keyframe;
// sb_cols/rows, rows/cols and last_fmt are used for allocating all internal
// arrays, and are thus per-thread. w/h and gf_fmt are synced between threads
// and are therefore per-stream. pix_fmt represents the value in the header
// of the currently processed frame.
int w, h;
enum AVPixelFormat pix_fmt, last_fmt, gf_fmt;
unsigned sb_cols, sb_rows, rows, cols;
ThreadFrame next_refs[8];
struct {
uint8_t lim_lut[64];
uint8_t mblim_lut[64];
} filter_lut;
unsigned tile_row_start, tile_row_end, tile_col_start, tile_col_end;
struct {
prob_context p;
uint8_t coef[4][2][2][6][6][3];
} prob_ctx[4];
struct {
prob_context p;
uint8_t coef[4][2][2][6][6][11];
} prob;
struct {
unsigned y_mode[4][10];
unsigned uv_mode[10][10];
unsigned filter[4][3];
unsigned mv_mode[7][4];
unsigned intra[4][2];
unsigned comp[5][2];
unsigned single_ref[5][2][2];
unsigned comp_ref[5][2];
unsigned tx32p[2][4];
unsigned tx16p[2][3];
unsigned tx8p[2][2];
unsigned skip[3][2];
unsigned mv_joint[4];
struct {
unsigned sign[2];
unsigned classes[11];
unsigned class0[2];
unsigned bits[10][2];
unsigned class0_fp[2][4];
unsigned fp[4];
unsigned class0_hp[2];
unsigned hp[2];
} mv_comp[2];
unsigned partition[4][4][4];
unsigned coef[4][2][2][6][6][3];
unsigned eob[4][2][2][6][6][2];
} counts;
// contextual (left/above) cache
DECLARE_ALIGNED(16, uint8_t, left_y_nnz_ctx)[16];
DECLARE_ALIGNED(16, uint8_t, left_mode_ctx)[16];
DECLARE_ALIGNED(16, VP56mv, left_mv_ctx)[16][2];
DECLARE_ALIGNED(16, uint8_t, left_uv_nnz_ctx)[2][16];
DECLARE_ALIGNED(8, uint8_t, left_partition_ctx)[8];
DECLARE_ALIGNED(8, uint8_t, left_skip_ctx)[8];
DECLARE_ALIGNED(8, uint8_t, left_txfm_ctx)[8];
DECLARE_ALIGNED(8, uint8_t, left_segpred_ctx)[8];
DECLARE_ALIGNED(8, uint8_t, left_intra_ctx)[8];
DECLARE_ALIGNED(8, uint8_t, left_comp_ctx)[8];
DECLARE_ALIGNED(8, uint8_t, left_ref_ctx)[8];
DECLARE_ALIGNED(8, uint8_t, left_filter_ctx)[8];
uint8_t *above_partition_ctx;
uint8_t *above_mode_ctx;
// FIXME maybe merge some of the below in a flags field?
uint8_t *above_y_nnz_ctx;
uint8_t *above_uv_nnz_ctx[2];
uint8_t *above_skip_ctx; // 1bit
uint8_t *above_txfm_ctx; // 2bit
uint8_t *above_segpred_ctx; // 1bit
uint8_t *above_intra_ctx; // 1bit
uint8_t *above_comp_ctx; // 1bit
uint8_t *above_ref_ctx; // 2bit
uint8_t *above_filter_ctx;
VP56mv (*above_mv_ctx)[2];
// whole-frame cache
uint8_t *intra_pred_data[3];
struct VP9Filter *lflvl;
DECLARE_ALIGNED(32, uint8_t, edge_emu_buffer)[135 * 144 * 2];
// block reconstruction intermediates
int block_alloc_using_2pass;
int16_t *block_base, *block, *uvblock_base[2], *uvblock[2];
uint8_t *eob_base, *uveob_base[2], *eob, *uveob[2];
struct { int x, y; } min_mv, max_mv;
DECLARE_ALIGNED(32, uint8_t, tmp_y)[64 * 64 * 2];
DECLARE_ALIGNED(32, uint8_t, tmp_uv)[2][64 * 64 * 2];
uint16_t mvscale[3][2];
uint8_t mvstep[3][2];
} VP9Context;
static const uint8_t bwh_tab[2][N_BS_SIZES][2] = {
{
{ 16, 16 }, { 16, 8 }, { 8, 16 }, { 8, 8 }, { 8, 4 }, { 4, 8 },
{ 4, 4 }, { 4, 2 }, { 2, 4 }, { 2, 2 }, { 2, 1 }, { 1, 2 }, { 1, 1 },
}, {
{ 8, 8 }, { 8, 4 }, { 4, 8 }, { 4, 4 }, { 4, 2 }, { 2, 4 },
{ 2, 2 }, { 2, 1 }, { 1, 2 }, { 1, 1 }, { 1, 1 }, { 1, 1 }, { 1, 1 },
}
};
static void vp9_unref_frame(AVCodecContext *ctx, VP9Frame *f)
{
ff_thread_release_buffer(ctx, &f->tf);
av_buffer_unref(&f->extradata);
av_buffer_unref(&f->hwaccel_priv_buf);
f->segmentation_map = NULL;
f->hwaccel_picture_private = NULL;
}
static int vp9_alloc_frame(AVCodecContext *ctx, VP9Frame *f)
{
VP9Context *s = ctx->priv_data;
int ret, sz;
if ((ret = ff_thread_get_buffer(ctx, &f->tf, AV_GET_BUFFER_FLAG_REF)) < 0)
return ret;
sz = 64 * s->sb_cols * s->sb_rows;
if (!(f->extradata = av_buffer_allocz(sz * (1 + sizeof(struct VP9mvrefPair))))) {
goto fail;
}
f->segmentation_map = f->extradata->data;
f->mv = (struct VP9mvrefPair *) (f->extradata->data + sz);
if (ctx->hwaccel) {
const AVHWAccel *hwaccel = ctx->hwaccel;
av_assert0(!f->hwaccel_picture_private);
if (hwaccel->frame_priv_data_size) {
f->hwaccel_priv_buf = av_buffer_allocz(hwaccel->frame_priv_data_size);
if (!f->hwaccel_priv_buf)
goto fail;
f->hwaccel_picture_private = f->hwaccel_priv_buf->data;
}
}
return 0;
fail:
vp9_unref_frame(ctx, f);
return AVERROR(ENOMEM);
}
static int vp9_ref_frame(AVCodecContext *ctx, VP9Frame *dst, VP9Frame *src)
{
int res;
if ((res = ff_thread_ref_frame(&dst->tf, &src->tf)) < 0) {
return res;
} else if (!(dst->extradata = av_buffer_ref(src->extradata))) {
goto fail;
}
dst->segmentation_map = src->segmentation_map;
dst->mv = src->mv;
dst->uses_2pass = src->uses_2pass;
if (src->hwaccel_picture_private) {
dst->hwaccel_priv_buf = av_buffer_ref(src->hwaccel_priv_buf);
if (!dst->hwaccel_priv_buf)
goto fail;
dst->hwaccel_picture_private = dst->hwaccel_priv_buf->data;
}
return 0;
fail:
vp9_unref_frame(ctx, dst);
return AVERROR(ENOMEM);
}
static int update_size(AVCodecContext *ctx, int w, int h)
{
#define HWACCEL_MAX (CONFIG_VP9_DXVA2_HWACCEL + CONFIG_VP9_D3D11VA_HWACCEL + CONFIG_VP9_VAAPI_HWACCEL)
enum AVPixelFormat pix_fmts[HWACCEL_MAX + 2], *fmtp = pix_fmts;
VP9Context *s = ctx->priv_data;
uint8_t *p;
int bytesperpixel = s->bytesperpixel, res, cols, rows;
av_assert0(w > 0 && h > 0);
if (!(s->pix_fmt == s->gf_fmt && w == s->w && h == s->h)) {
if ((res = ff_set_dimensions(ctx, w, h)) < 0)
return res;
if (s->pix_fmt == AV_PIX_FMT_YUV420P) {
#if CONFIG_VP9_DXVA2_HWACCEL
*fmtp++ = AV_PIX_FMT_DXVA2_VLD;
#endif
#if CONFIG_VP9_D3D11VA_HWACCEL
*fmtp++ = AV_PIX_FMT_D3D11VA_VLD;
#endif
#if CONFIG_VP9_VAAPI_HWACCEL
*fmtp++ = AV_PIX_FMT_VAAPI;
#endif
}
*fmtp++ = s->pix_fmt;
*fmtp = AV_PIX_FMT_NONE;
res = ff_thread_get_format(ctx, pix_fmts);
if (res < 0)
return res;
ctx->pix_fmt = res;
s->gf_fmt = s->pix_fmt;
s->w = w;
s->h = h;
}
cols = (w + 7) >> 3;
rows = (h + 7) >> 3;
if (s->intra_pred_data[0] && cols == s->cols && rows == s->rows && s->pix_fmt == s->last_fmt)
return 0;
s->last_fmt = s->pix_fmt;
s->sb_cols = (w + 63) >> 6;
s->sb_rows = (h + 63) >> 6;
s->cols = (w + 7) >> 3;
s->rows = (h + 7) >> 3;
#define assign(var, type, n) var = (type) p; p += s->sb_cols * (n) * sizeof(*var)
av_freep(&s->intra_pred_data[0]);
// FIXME we slightly over-allocate here for subsampled chroma, but a little
// bit of padding shouldn't affect performance...
p = av_malloc(s->sb_cols * (128 + 192 * bytesperpixel +
sizeof(*s->lflvl) + 16 * sizeof(*s->above_mv_ctx)));
if (!p)
return AVERROR(ENOMEM);
assign(s->intra_pred_data[0], uint8_t *, 64 * bytesperpixel);
assign(s->intra_pred_data[1], uint8_t *, 64 * bytesperpixel);
assign(s->intra_pred_data[2], uint8_t *, 64 * bytesperpixel);
assign(s->above_y_nnz_ctx, uint8_t *, 16);
assign(s->above_mode_ctx, uint8_t *, 16);
assign(s->above_mv_ctx, VP56mv(*)[2], 16);
assign(s->above_uv_nnz_ctx[0], uint8_t *, 16);
assign(s->above_uv_nnz_ctx[1], uint8_t *, 16);
assign(s->above_partition_ctx, uint8_t *, 8);
assign(s->above_skip_ctx, uint8_t *, 8);
assign(s->above_txfm_ctx, uint8_t *, 8);
assign(s->above_segpred_ctx, uint8_t *, 8);
assign(s->above_intra_ctx, uint8_t *, 8);
assign(s->above_comp_ctx, uint8_t *, 8);
assign(s->above_ref_ctx, uint8_t *, 8);
assign(s->above_filter_ctx, uint8_t *, 8);
assign(s->lflvl, struct VP9Filter *, 1);
#undef assign
// these will be re-allocated a little later
av_freep(&s->b_base);
av_freep(&s->block_base);
if (s->bpp != s->last_bpp) {
ff_vp9dsp_init(&s->dsp, s->bpp, ctx->flags & AV_CODEC_FLAG_BITEXACT);
ff_videodsp_init(&s->vdsp, s->bpp);
s->last_bpp = s->bpp;
}
return 0;
}
static int update_block_buffers(AVCodecContext *ctx)
{
VP9Context *s = ctx->priv_data;
int chroma_blocks, chroma_eobs, bytesperpixel = s->bytesperpixel;
if (s->b_base && s->block_base && s->block_alloc_using_2pass == s->s.frames[CUR_FRAME].uses_2pass)
return 0;
av_free(s->b_base);
av_free(s->block_base);
chroma_blocks = 64 * 64 >> (s->ss_h + s->ss_v);
chroma_eobs = 16 * 16 >> (s->ss_h + s->ss_v);
if (s->s.frames[CUR_FRAME].uses_2pass) {
int sbs = s->sb_cols * s->sb_rows;
s->b_base = av_malloc_array(s->cols * s->rows, sizeof(VP9Block));
s->block_base = av_mallocz(((64 * 64 + 2 * chroma_blocks) * bytesperpixel * sizeof(int16_t) +
16 * 16 + 2 * chroma_eobs) * sbs);
if (!s->b_base || !s->block_base)
return AVERROR(ENOMEM);
s->uvblock_base[0] = s->block_base + sbs * 64 * 64 * bytesperpixel;
s->uvblock_base[1] = s->uvblock_base[0] + sbs * chroma_blocks * bytesperpixel;
s->eob_base = (uint8_t *) (s->uvblock_base[1] + sbs * chroma_blocks * bytesperpixel);
s->uveob_base[0] = s->eob_base + 16 * 16 * sbs;
s->uveob_base[1] = s->uveob_base[0] + chroma_eobs * sbs;
} else {
s->b_base = av_malloc(sizeof(VP9Block));
s->block_base = av_mallocz((64 * 64 + 2 * chroma_blocks) * bytesperpixel * sizeof(int16_t) +
16 * 16 + 2 * chroma_eobs);
if (!s->b_base || !s->block_base)
return AVERROR(ENOMEM);
s->uvblock_base[0] = s->block_base + 64 * 64 * bytesperpixel;
s->uvblock_base[1] = s->uvblock_base[0] + chroma_blocks * bytesperpixel;
s->eob_base = (uint8_t *) (s->uvblock_base[1] + chroma_blocks * bytesperpixel);
s->uveob_base[0] = s->eob_base + 16 * 16;
s->uveob_base[1] = s->uveob_base[0] + chroma_eobs;
}
s->block_alloc_using_2pass = s->s.frames[CUR_FRAME].uses_2pass;
return 0;
}
// for some reason the sign bit is at the end, not the start, of a bit sequence
static av_always_inline int get_sbits_inv(GetBitContext *gb, int n)
{
int v = get_bits(gb, n);
return get_bits1(gb) ? -v : v;
}
static av_always_inline int inv_recenter_nonneg(int v, int m)
{
return v > 2 * m ? v : v & 1 ? m - ((v + 1) >> 1) : m + (v >> 1);
}
// differential forward probability updates
static int update_prob(VP56RangeCoder *c, int p)
{
static const int inv_map_table[255] = {
7, 20, 33, 46, 59, 72, 85, 98, 111, 124, 137, 150, 163, 176,
189, 202, 215, 228, 241, 254, 1, 2, 3, 4, 5, 6, 8, 9,
10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 21, 22, 23, 24,
25, 26, 27, 28, 29, 30, 31, 32, 34, 35, 36, 37, 38, 39,
40, 41, 42, 43, 44, 45, 47, 48, 49, 50, 51, 52, 53, 54,
55, 56, 57, 58, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69,
70, 71, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84,
86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 99, 100,
101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 112, 113, 114, 115,
116, 117, 118, 119, 120, 121, 122, 123, 125, 126, 127, 128, 129, 130,
131, 132, 133, 134, 135, 136, 138, 139, 140, 141, 142, 143, 144, 145,
146, 147, 148, 149, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160,
161, 162, 164, 165, 166, 167, 168, 169, 170, 171, 172, 173, 174, 175,
177, 178, 179, 180, 181, 182, 183, 184, 185, 186, 187, 188, 190, 191,
192, 193, 194, 195, 196, 197, 198, 199, 200, 201, 203, 204, 205, 206,
207, 208, 209, 210, 211, 212, 213, 214, 216, 217, 218, 219, 220, 221,
222, 223, 224, 225, 226, 227, 229, 230, 231, 232, 233, 234, 235, 236,
237, 238, 239, 240, 242, 243, 244, 245, 246, 247, 248, 249, 250, 251,
252, 253, 253,
};
int d;
/* This code is trying to do a differential probability update. For a
* current probability A in the range [1, 255], the difference to a new
* probability of any value can be expressed differentially as 1-A,255-A
* where some part of this (absolute range) exists both in positive as
* well as the negative part, whereas another part only exists in one
* half. We're trying to code this shared part differentially, i.e.
* times two where the value of the lowest bit specifies the sign, and
* the single part is then coded on top of this. This absolute difference
* then again has a value of [0,254], but a bigger value in this range
* indicates that we're further away from the original value A, so we
* can code this as a VLC code, since higher values are increasingly
* unlikely. The first 20 values in inv_map_table[] allow 'cheap, rough'
* updates vs. the 'fine, exact' updates further down the range, which
* adds one extra dimension to this differential update model. */
if (!vp8_rac_get(c)) {
d = vp8_rac_get_uint(c, 4) + 0;
} else if (!vp8_rac_get(c)) {
d = vp8_rac_get_uint(c, 4) + 16;
} else if (!vp8_rac_get(c)) {
d = vp8_rac_get_uint(c, 5) + 32;
} else {
d = vp8_rac_get_uint(c, 7);
if (d >= 65)
d = (d << 1) - 65 + vp8_rac_get(c);
d += 64;
av_assert2(d < FF_ARRAY_ELEMS(inv_map_table));
}
return p <= 128 ? 1 + inv_recenter_nonneg(inv_map_table[d], p - 1) :
255 - inv_recenter_nonneg(inv_map_table[d], 255 - p);
}
static int read_colorspace_details(AVCodecContext *ctx)
{
static const enum AVColorSpace colorspaces[8] = {
AVCOL_SPC_UNSPECIFIED, AVCOL_SPC_BT470BG, AVCOL_SPC_BT709, AVCOL_SPC_SMPTE170M,
AVCOL_SPC_SMPTE240M, AVCOL_SPC_BT2020_NCL, AVCOL_SPC_RESERVED, AVCOL_SPC_RGB,
};
VP9Context *s = ctx->priv_data;
int bits = ctx->profile <= 1 ? 0 : 1 + get_bits1(&s->gb); // 0:8, 1:10, 2:12
s->bpp_index = bits;
s->bpp = 8 + bits * 2;
s->bytesperpixel = (7 + s->bpp) >> 3;
ctx->colorspace = colorspaces[get_bits(&s->gb, 3)];
if (ctx->colorspace == AVCOL_SPC_RGB) { // RGB = profile 1
static const enum AVPixelFormat pix_fmt_rgb[3] = {
AV_PIX_FMT_GBRP, AV_PIX_FMT_GBRP10, AV_PIX_FMT_GBRP12
};
s->ss_h = s->ss_v = 0;
ctx->color_range = AVCOL_RANGE_JPEG;
s->pix_fmt = pix_fmt_rgb[bits];
if (ctx->profile & 1) {
if (get_bits1(&s->gb)) {
av_log(ctx, AV_LOG_ERROR, "Reserved bit set in RGB\n");
return AVERROR_INVALIDDATA;
}
} else {
av_log(ctx, AV_LOG_ERROR, "RGB not supported in profile %d\n",
ctx->profile);
return AVERROR_INVALIDDATA;
}
} else {
static const enum AVPixelFormat pix_fmt_for_ss[3][2 /* v */][2 /* h */] = {
{ { AV_PIX_FMT_YUV444P, AV_PIX_FMT_YUV422P },
{ AV_PIX_FMT_YUV440P, AV_PIX_FMT_YUV420P } },
{ { AV_PIX_FMT_YUV444P10, AV_PIX_FMT_YUV422P10 },
{ AV_PIX_FMT_YUV440P10, AV_PIX_FMT_YUV420P10 } },
{ { AV_PIX_FMT_YUV444P12, AV_PIX_FMT_YUV422P12 },
{ AV_PIX_FMT_YUV440P12, AV_PIX_FMT_YUV420P12 } }
};
ctx->color_range = get_bits1(&s->gb) ? AVCOL_RANGE_JPEG : AVCOL_RANGE_MPEG;
if (ctx->profile & 1) {
s->ss_h = get_bits1(&s->gb);
s->ss_v = get_bits1(&s->gb);
s->pix_fmt = pix_fmt_for_ss[bits][s->ss_v][s->ss_h];
if (s->pix_fmt == AV_PIX_FMT_YUV420P) {
av_log(ctx, AV_LOG_ERROR, "YUV 4:2:0 not supported in profile %d\n",
ctx->profile);
return AVERROR_INVALIDDATA;
} else if (get_bits1(&s->gb)) {
av_log(ctx, AV_LOG_ERROR, "Profile %d color details reserved bit set\n",
ctx->profile);
return AVERROR_INVALIDDATA;
}
} else {
s->ss_h = s->ss_v = 1;
s->pix_fmt = pix_fmt_for_ss[bits][1][1];
}
}
return 0;
}
static int decode_frame_header(AVCodecContext *ctx,
const uint8_t *data, int size, int *ref)
{
VP9Context *s = ctx->priv_data;
int c, i, j, k, l, m, n, w, h, max, size2, res, sharp;
int last_invisible;
const uint8_t *data2;
/* general header */
if ((res = init_get_bits8(&s->gb, data, size)) < 0) {
av_log(ctx, AV_LOG_ERROR, "Failed to initialize bitstream reader\n");
return res;
}
if (get_bits(&s->gb, 2) != 0x2) { // frame marker
av_log(ctx, AV_LOG_ERROR, "Invalid frame marker\n");
return AVERROR_INVALIDDATA;
}
ctx->profile = get_bits1(&s->gb);
ctx->profile |= get_bits1(&s->gb) << 1;
if (ctx->profile == 3) ctx->profile += get_bits1(&s->gb);
if (ctx->profile > 3) {
av_log(ctx, AV_LOG_ERROR, "Profile %d is not yet supported\n", ctx->profile);
return AVERROR_INVALIDDATA;
}
s->s.h.profile = ctx->profile;
if (get_bits1(&s->gb)) {
*ref = get_bits(&s->gb, 3);
return 0;
}
s->last_keyframe = s->s.h.keyframe;
s->s.h.keyframe = !get_bits1(&s->gb);
last_invisible = s->s.h.invisible;
s->s.h.invisible = !get_bits1(&s->gb);
s->s.h.errorres = get_bits1(&s->gb);
s->s.h.use_last_frame_mvs = !s->s.h.errorres && !last_invisible;
if (s->s.h.keyframe) {
if (get_bits_long(&s->gb, 24) != VP9_SYNCCODE) { // synccode
av_log(ctx, AV_LOG_ERROR, "Invalid sync code\n");
return AVERROR_INVALIDDATA;
}
if ((res = read_colorspace_details(ctx)) < 0)
return res;
// for profile 1, here follows the subsampling bits
s->s.h.refreshrefmask = 0xff;
w = get_bits(&s->gb, 16) + 1;
h = get_bits(&s->gb, 16) + 1;
if (get_bits1(&s->gb)) // display size
skip_bits(&s->gb, 32);
} else {
s->s.h.intraonly = s->s.h.invisible ? get_bits1(&s->gb) : 0;
s->s.h.resetctx = s->s.h.errorres ? 0 : get_bits(&s->gb, 2);
if (s->s.h.intraonly) {
if (get_bits_long(&s->gb, 24) != VP9_SYNCCODE) { // synccode
av_log(ctx, AV_LOG_ERROR, "Invalid sync code\n");
return AVERROR_INVALIDDATA;
}
if (ctx->profile >= 1) {
if ((res = read_colorspace_details(ctx)) < 0)
return res;
} else {
s->ss_h = s->ss_v = 1;
s->bpp = 8;
s->bpp_index = 0;
s->bytesperpixel = 1;
s->pix_fmt = AV_PIX_FMT_YUV420P;
ctx->colorspace = AVCOL_SPC_BT470BG;
ctx->color_range = AVCOL_RANGE_JPEG;
}
s->s.h.refreshrefmask = get_bits(&s->gb, 8);
w = get_bits(&s->gb, 16) + 1;
h = get_bits(&s->gb, 16) + 1;
if (get_bits1(&s->gb)) // display size
skip_bits(&s->gb, 32);
} else {
s->s.h.refreshrefmask = get_bits(&s->gb, 8);
s->s.h.refidx[0] = get_bits(&s->gb, 3);
s->s.h.signbias[0] = get_bits1(&s->gb) && !s->s.h.errorres;
s->s.h.refidx[1] = get_bits(&s->gb, 3);
s->s.h.signbias[1] = get_bits1(&s->gb) && !s->s.h.errorres;
s->s.h.refidx[2] = get_bits(&s->gb, 3);
s->s.h.signbias[2] = get_bits1(&s->gb) && !s->s.h.errorres;
if (!s->s.refs[s->s.h.refidx[0]].f->buf[0] ||
!s->s.refs[s->s.h.refidx[1]].f->buf[0] ||
!s->s.refs[s->s.h.refidx[2]].f->buf[0]) {
av_log(ctx, AV_LOG_ERROR, "Not all references are available\n");
return AVERROR_INVALIDDATA;
}
if (get_bits1(&s->gb)) {
w = s->s.refs[s->s.h.refidx[0]].f->width;
h = s->s.refs[s->s.h.refidx[0]].f->height;
} else if (get_bits1(&s->gb)) {
w = s->s.refs[s->s.h.refidx[1]].f->width;
h = s->s.refs[s->s.h.refidx[1]].f->height;
} else if (get_bits1(&s->gb)) {
w = s->s.refs[s->s.h.refidx[2]].f->width;
h = s->s.refs[s->s.h.refidx[2]].f->height;
} else {
w = get_bits(&s->gb, 16) + 1;
h = get_bits(&s->gb, 16) + 1;
}
// Note that in this code, "CUR_FRAME" is actually before we
// have formally allocated a frame, and thus actually represents
// the _last_ frame
s->s.h.use_last_frame_mvs &= s->s.frames[CUR_FRAME].tf.f->width == w &&
s->s.frames[CUR_FRAME].tf.f->height == h;
if (get_bits1(&s->gb)) // display size
skip_bits(&s->gb, 32);
s->s.h.highprecisionmvs = get_bits1(&s->gb);
s->s.h.filtermode = get_bits1(&s->gb) ? FILTER_SWITCHABLE :
get_bits(&s->gb, 2);
s->s.h.allowcompinter = s->s.h.signbias[0] != s->s.h.signbias[1] ||
s->s.h.signbias[0] != s->s.h.signbias[2];
if (s->s.h.allowcompinter) {
if (s->s.h.signbias[0] == s->s.h.signbias[1]) {
s->s.h.fixcompref = 2;
s->s.h.varcompref[0] = 0;
s->s.h.varcompref[1] = 1;
} else if (s->s.h.signbias[0] == s->s.h.signbias[2]) {
s->s.h.fixcompref = 1;
s->s.h.varcompref[0] = 0;
s->s.h.varcompref[1] = 2;
} else {
s->s.h.fixcompref = 0;
s->s.h.varcompref[0] = 1;
s->s.h.varcompref[1] = 2;
}
}
}
}
s->s.h.refreshctx = s->s.h.errorres ? 0 : get_bits1(&s->gb);
s->s.h.parallelmode = s->s.h.errorres ? 1 : get_bits1(&s->gb);
s->s.h.framectxid = c = get_bits(&s->gb, 2);
/* loopfilter header data */
if (s->s.h.keyframe || s->s.h.errorres || s->s.h.intraonly) {
// reset loopfilter defaults
s->s.h.lf_delta.ref[0] = 1;
s->s.h.lf_delta.ref[1] = 0;
s->s.h.lf_delta.ref[2] = -1;
s->s.h.lf_delta.ref[3] = -1;
s->s.h.lf_delta.mode[0] = 0;
s->s.h.lf_delta.mode[1] = 0;
memset(s->s.h.segmentation.feat, 0, sizeof(s->s.h.segmentation.feat));
}
s->s.h.filter.level = get_bits(&s->gb, 6);
sharp = get_bits(&s->gb, 3);
// if sharpness changed, reinit lim/mblim LUTs. if it didn't change, keep
// the old cache values since they are still valid
if (s->s.h.filter.sharpness != sharp)
memset(s->filter_lut.lim_lut, 0, sizeof(s->filter_lut.lim_lut));
s->s.h.filter.sharpness = sharp;
if ((s->s.h.lf_delta.enabled = get_bits1(&s->gb))) {
if ((s->s.h.lf_delta.updated = get_bits1(&s->gb))) {
for (i = 0; i < 4; i++)
if (get_bits1(&s->gb))
s->s.h.lf_delta.ref[i] = get_sbits_inv(&s->gb, 6);
for (i = 0; i < 2; i++)
if (get_bits1(&s->gb))
s->s.h.lf_delta.mode[i] = get_sbits_inv(&s->gb, 6);
}
}
/* quantization header data */
s->s.h.yac_qi = get_bits(&s->gb, 8);
s->s.h.ydc_qdelta = get_bits1(&s->gb) ? get_sbits_inv(&s->gb, 4) : 0;
s->s.h.uvdc_qdelta = get_bits1(&s->gb) ? get_sbits_inv(&s->gb, 4) : 0;
s->s.h.uvac_qdelta = get_bits1(&s->gb) ? get_sbits_inv(&s->gb, 4) : 0;
s->s.h.lossless = s->s.h.yac_qi == 0 && s->s.h.ydc_qdelta == 0 &&
s->s.h.uvdc_qdelta == 0 && s->s.h.uvac_qdelta == 0;
if (s->s.h.lossless)
ctx->properties |= FF_CODEC_PROPERTY_LOSSLESS;
/* segmentation header info */
if ((s->s.h.segmentation.enabled = get_bits1(&s->gb))) {
if ((s->s.h.segmentation.update_map = get_bits1(&s->gb))) {
for (i = 0; i < 7; i++)
s->s.h.segmentation.prob[i] = get_bits1(&s->gb) ?
get_bits(&s->gb, 8) : 255;
if ((s->s.h.segmentation.temporal = get_bits1(&s->gb))) {
for (i = 0; i < 3; i++)
s->s.h.segmentation.pred_prob[i] = get_bits1(&s->gb) ?
get_bits(&s->gb, 8) : 255;
}
}
if (get_bits1(&s->gb)) {
s->s.h.segmentation.absolute_vals = get_bits1(&s->gb);
for (i = 0; i < 8; i++) {
if ((s->s.h.segmentation.feat[i].q_enabled = get_bits1(&s->gb)))
s->s.h.segmentation.feat[i].q_val = get_sbits_inv(&s->gb, 8);
if ((s->s.h.segmentation.feat[i].lf_enabled = get_bits1(&s->gb)))
s->s.h.segmentation.feat[i].lf_val = get_sbits_inv(&s->gb, 6);
if ((s->s.h.segmentation.feat[i].ref_enabled = get_bits1(&s->gb)))
s->s.h.segmentation.feat[i].ref_val = get_bits(&s->gb, 2);
s->s.h.segmentation.feat[i].skip_enabled = get_bits1(&s->gb);
}
}
}
// set qmul[] based on Y/UV, AC/DC and segmentation Q idx deltas
for (i = 0; i < (s->s.h.segmentation.enabled ? 8 : 1); i++) {
int qyac, qydc, quvac, quvdc, lflvl, sh;
if (s->s.h.segmentation.enabled && s->s.h.segmentation.feat[i].q_enabled) {
if (s->s.h.segmentation.absolute_vals)
qyac = av_clip_uintp2(s->s.h.segmentation.feat[i].q_val, 8);
else
qyac = av_clip_uintp2(s->s.h.yac_qi + s->s.h.segmentation.feat[i].q_val, 8);
} else {
qyac = s->s.h.yac_qi;
}
qydc = av_clip_uintp2(qyac + s->s.h.ydc_qdelta, 8);
quvdc = av_clip_uintp2(qyac + s->s.h.uvdc_qdelta, 8);
quvac = av_clip_uintp2(qyac + s->s.h.uvac_qdelta, 8);
qyac = av_clip_uintp2(qyac, 8);
s->s.h.segmentation.feat[i].qmul[0][0] = vp9_dc_qlookup[s->bpp_index][qydc];
s->s.h.segmentation.feat[i].qmul[0][1] = vp9_ac_qlookup[s->bpp_index][qyac];
s->s.h.segmentation.feat[i].qmul[1][0] = vp9_dc_qlookup[s->bpp_index][quvdc];
s->s.h.segmentation.feat[i].qmul[1][1] = vp9_ac_qlookup[s->bpp_index][quvac];
sh = s->s.h.filter.level >= 32;
if (s->s.h.segmentation.enabled && s->s.h.segmentation.feat[i].lf_enabled) {
if (s->s.h.segmentation.absolute_vals)
lflvl = av_clip_uintp2(s->s.h.segmentation.feat[i].lf_val, 6);
else
lflvl = av_clip_uintp2(s->s.h.filter.level + s->s.h.segmentation.feat[i].lf_val, 6);
} else {
lflvl = s->s.h.filter.level;
}
if (s->s.h.lf_delta.enabled) {
s->s.h.segmentation.feat[i].lflvl[0][0] =
s->s.h.segmentation.feat[i].lflvl[0][1] =
av_clip_uintp2(lflvl + (s->s.h.lf_delta.ref[0] << sh), 6);
for (j = 1; j < 4; j++) {
s->s.h.segmentation.feat[i].lflvl[j][0] =
av_clip_uintp2(lflvl + ((s->s.h.lf_delta.ref[j] +
s->s.h.lf_delta.mode[0]) * (1 << sh)), 6);
s->s.h.segmentation.feat[i].lflvl[j][1] =
av_clip_uintp2(lflvl + ((s->s.h.lf_delta.ref[j] +
s->s.h.lf_delta.mode[1]) * (1 << sh)), 6);
}
} else {
memset(s->s.h.segmentation.feat[i].lflvl, lflvl,
sizeof(s->s.h.segmentation.feat[i].lflvl));
}
}
/* tiling info */
if ((res = update_size(ctx, w, h)) < 0) {
av_log(ctx, AV_LOG_ERROR, "Failed to initialize decoder for %dx%d @ %d\n",
w, h, s->pix_fmt);
return res;
}
for (s->s.h.tiling.log2_tile_cols = 0;
s->sb_cols > (64 << s->s.h.tiling.log2_tile_cols);
s->s.h.tiling.log2_tile_cols++) ;
for (max = 0; (s->sb_cols >> max) >= 4; max++) ;
max = FFMAX(0, max - 1);
while (max > s->s.h.tiling.log2_tile_cols) {
if (get_bits1(&s->gb))
s->s.h.tiling.log2_tile_cols++;
else
break;
}
s->s.h.tiling.log2_tile_rows = decode012(&s->gb);
s->s.h.tiling.tile_rows = 1 << s->s.h.tiling.log2_tile_rows;
if (s->s.h.tiling.tile_cols != (1 << s->s.h.tiling.log2_tile_cols)) {
s->s.h.tiling.tile_cols = 1 << s->s.h.tiling.log2_tile_cols;
s->c_b = av_fast_realloc(s->c_b, &s->c_b_size,
sizeof(VP56RangeCoder) * s->s.h.tiling.tile_cols);
if (!s->c_b) {
av_log(ctx, AV_LOG_ERROR, "Ran out of memory during range coder init\n");
return AVERROR(ENOMEM);
}
}
/* check reference frames */
if (!s->s.h.keyframe && !s->s.h.intraonly) {
for (i = 0; i < 3; i++) {
AVFrame *ref = s->s.refs[s->s.h.refidx[i]].f;
int refw = ref->width, refh = ref->height;
if (ref->format != ctx->pix_fmt) {
av_log(ctx, AV_LOG_ERROR,
"Ref pixfmt (%s) did not match current frame (%s)",
av_get_pix_fmt_name(ref->format),
av_get_pix_fmt_name(ctx->pix_fmt));
return AVERROR_INVALIDDATA;
} else if (refw == w && refh == h) {
s->mvscale[i][0] = s->mvscale[i][1] = 0;
} else {
if (w * 2 < refw || h * 2 < refh || w > 16 * refw || h > 16 * refh) {
av_log(ctx, AV_LOG_ERROR,
"Invalid ref frame dimensions %dx%d for frame size %dx%d\n",
refw, refh, w, h);
return AVERROR_INVALIDDATA;
}
s->mvscale[i][0] = (refw << 14) / w;
s->mvscale[i][1] = (refh << 14) / h;
s->mvstep[i][0] = 16 * s->mvscale[i][0] >> 14;
s->mvstep[i][1] = 16 * s->mvscale[i][1] >> 14;
}
}
}
if (s->s.h.keyframe || s->s.h.errorres || (s->s.h.intraonly && s->s.h.resetctx == 3)) {
s->prob_ctx[0].p = s->prob_ctx[1].p = s->prob_ctx[2].p =
s->prob_ctx[3].p = vp9_default_probs;
memcpy(s->prob_ctx[0].coef, vp9_default_coef_probs,
sizeof(vp9_default_coef_probs));
memcpy(s->prob_ctx[1].coef, vp9_default_coef_probs,
sizeof(vp9_default_coef_probs));
memcpy(s->prob_ctx[2].coef, vp9_default_coef_probs,
sizeof(vp9_default_coef_probs));
memcpy(s->prob_ctx[3].coef, vp9_default_coef_probs,
sizeof(vp9_default_coef_probs));
} else if (s->s.h.intraonly && s->s.h.resetctx == 2) {
s->prob_ctx[c].p = vp9_default_probs;
memcpy(s->prob_ctx[c].coef, vp9_default_coef_probs,
sizeof(vp9_default_coef_probs));
}
// next 16 bits is size of the rest of the header (arith-coded)
s->s.h.compressed_header_size = size2 = get_bits(&s->gb, 16);
s->s.h.uncompressed_header_size = (get_bits_count(&s->gb) + 7) / 8;
data2 = align_get_bits(&s->gb);
if (size2 > size - (data2 - data)) {
av_log(ctx, AV_LOG_ERROR, "Invalid compressed header size\n");
return AVERROR_INVALIDDATA;
}
ff_vp56_init_range_decoder(&s->c, data2, size2);
if (vp56_rac_get_prob_branchy(&s->c, 128)) { // marker bit
av_log(ctx, AV_LOG_ERROR, "Marker bit was set\n");
return AVERROR_INVALIDDATA;
}
if (s->s.h.keyframe || s->s.h.intraonly) {
memset(s->counts.coef, 0, sizeof(s->counts.coef));
memset(s->counts.eob, 0, sizeof(s->counts.eob));
} else {
memset(&s->counts, 0, sizeof(s->counts));
}
// FIXME is it faster to not copy here, but do it down in the fw updates
// as explicit copies if the fw update is missing (and skip the copy upon
// fw update)?
s->prob.p = s->prob_ctx[c].p;
// txfm updates
if (s->s.h.lossless) {
s->s.h.txfmmode = TX_4X4;
} else {
s->s.h.txfmmode = vp8_rac_get_uint(&s->c, 2);
if (s->s.h.txfmmode == 3)
s->s.h.txfmmode += vp8_rac_get(&s->c);
if (s->s.h.txfmmode == TX_SWITCHABLE) {
for (i = 0; i < 2; i++)
if (vp56_rac_get_prob_branchy(&s->c, 252))
s->prob.p.tx8p[i] = update_prob(&s->c, s->prob.p.tx8p[i]);
for (i = 0; i < 2; i++)
for (j = 0; j < 2; j++)
if (vp56_rac_get_prob_branchy(&s->c, 252))
s->prob.p.tx16p[i][j] =
update_prob(&s->c, s->prob.p.tx16p[i][j]);
for (i = 0; i < 2; i++)
for (j = 0; j < 3; j++)
if (vp56_rac_get_prob_branchy(&s->c, 252))
s->prob.p.tx32p[i][j] =
update_prob(&s->c, s->prob.p.tx32p[i][j]);
}
}
// coef updates
for (i = 0; i < 4; i++) {
uint8_t (*ref)[2][6][6][3] = s->prob_ctx[c].coef[i];
if (vp8_rac_get(&s->c)) {
for (j = 0; j < 2; j++)
for (k = 0; k < 2; k++)
for (l = 0; l < 6; l++)
for (m = 0; m < 6; m++) {
uint8_t *p = s->prob.coef[i][j][k][l][m];
uint8_t *r = ref[j][k][l][m];
if (m >= 3 && l == 0) // dc only has 3 pt
break;
for (n = 0; n < 3; n++) {
if (vp56_rac_get_prob_branchy(&s->c, 252)) {
p[n] = update_prob(&s->c, r[n]);
} else {
p[n] = r[n];
}
}
p[3] = 0;
}
} else {
for (j = 0; j < 2; j++)
for (k = 0; k < 2; k++)
for (l = 0; l < 6; l++)
for (m = 0; m < 6; m++) {
uint8_t *p = s->prob.coef[i][j][k][l][m];
uint8_t *r = ref[j][k][l][m];
if (m > 3 && l == 0) // dc only has 3 pt
break;
memcpy(p, r, 3);
p[3] = 0;
}
}
if (s->s.h.txfmmode == i)
break;
}
// mode updates
for (i = 0; i < 3; i++)
if (vp56_rac_get_prob_branchy(&s->c, 252))
s->prob.p.skip[i] = update_prob(&s->c, s->prob.p.skip[i]);
if (!s->s.h.keyframe && !s->s.h.intraonly) {
for (i = 0; i < 7; i++)
for (j = 0; j < 3; j++)
if (vp56_rac_get_prob_branchy(&s->c, 252))
s->prob.p.mv_mode[i][j] =
update_prob(&s->c, s->prob.p.mv_mode[i][j]);
if (s->s.h.filtermode == FILTER_SWITCHABLE)
for (i = 0; i < 4; i++)
for (j = 0; j < 2; j++)
if (vp56_rac_get_prob_branchy(&s->c, 252))
s->prob.p.filter[i][j] =
update_prob(&s->c, s->prob.p.filter[i][j]);
for (i = 0; i < 4; i++)
if (vp56_rac_get_prob_branchy(&s->c, 252))
s->prob.p.intra[i] = update_prob(&s->c, s->prob.p.intra[i]);
if (s->s.h.allowcompinter) {
s->s.h.comppredmode = vp8_rac_get(&s->c);
if (s->s.h.comppredmode)
s->s.h.comppredmode += vp8_rac_get(&s->c);
if (s->s.h.comppredmode == PRED_SWITCHABLE)
for (i = 0; i < 5; i++)
if (vp56_rac_get_prob_branchy(&s->c, 252))
s->prob.p.comp[i] =
update_prob(&s->c, s->prob.p.comp[i]);
} else {
s->s.h.comppredmode = PRED_SINGLEREF;
}
if (s->s.h.comppredmode != PRED_COMPREF) {
for (i = 0; i < 5; i++) {
if (vp56_rac_get_prob_branchy(&s->c, 252))
s->prob.p.single_ref[i][0] =
update_prob(&s->c, s->prob.p.single_ref[i][0]);
if (vp56_rac_get_prob_branchy(&s->c, 252))
s->prob.p.single_ref[i][1] =
update_prob(&s->c, s->prob.p.single_ref[i][1]);
}
}
if (s->s.h.comppredmode != PRED_SINGLEREF) {
for (i = 0; i < 5; i++)
if (vp56_rac_get_prob_branchy(&s->c, 252))
s->prob.p.comp_ref[i] =
update_prob(&s->c, s->prob.p.comp_ref[i]);
}
for (i = 0; i < 4; i++)
for (j = 0; j < 9; j++)
if (vp56_rac_get_prob_branchy(&s->c, 252))
s->prob.p.y_mode[i][j] =
update_prob(&s->c, s->prob.p.y_mode[i][j]);
for (i = 0; i < 4; i++)
for (j = 0; j < 4; j++)
for (k = 0; k < 3; k++)
if (vp56_rac_get_prob_branchy(&s->c, 252))
s->prob.p.partition[3 - i][j][k] =
update_prob(&s->c, s->prob.p.partition[3 - i][j][k]);
// mv fields don't use the update_prob subexp model for some reason
for (i = 0; i < 3; i++)
if (vp56_rac_get_prob_branchy(&s->c, 252))
s->prob.p.mv_joint[i] = (vp8_rac_get_uint(&s->c, 7) << 1) | 1;
for (i = 0; i < 2; i++) {
if (vp56_rac_get_prob_branchy(&s->c, 252))
s->prob.p.mv_comp[i].sign = (vp8_rac_get_uint(&s->c, 7) << 1) | 1;
for (j = 0; j < 10; j++)
if (vp56_rac_get_prob_branchy(&s->c, 252))
s->prob.p.mv_comp[i].classes[j] =
(vp8_rac_get_uint(&s->c, 7) << 1) | 1;
if (vp56_rac_get_prob_branchy(&s->c, 252))
s->prob.p.mv_comp[i].class0 = (vp8_rac_get_uint(&s->c, 7) << 1) | 1;
for (j = 0; j < 10; j++)
if (vp56_rac_get_prob_branchy(&s->c, 252))
s->prob.p.mv_comp[i].bits[j] =
(vp8_rac_get_uint(&s->c, 7) << 1) | 1;
}
for (i = 0; i < 2; i++) {
for (j = 0; j < 2; j++)
for (k = 0; k < 3; k++)
if (vp56_rac_get_prob_branchy(&s->c, 252))
s->prob.p.mv_comp[i].class0_fp[j][k] =
(vp8_rac_get_uint(&s->c, 7) << 1) | 1;
for (j = 0; j < 3; j++)
if (vp56_rac_get_prob_branchy(&s->c, 252))
s->prob.p.mv_comp[i].fp[j] =
(vp8_rac_get_uint(&s->c, 7) << 1) | 1;
}
if (s->s.h.highprecisionmvs) {
for (i = 0; i < 2; i++) {
if (vp56_rac_get_prob_branchy(&s->c, 252))
s->prob.p.mv_comp[i].class0_hp =
(vp8_rac_get_uint(&s->c, 7) << 1) | 1;
if (vp56_rac_get_prob_branchy(&s->c, 252))
s->prob.p.mv_comp[i].hp =
(vp8_rac_get_uint(&s->c, 7) << 1) | 1;
}
}
}
return (data2 - data) + size2;
}
static av_always_inline void clamp_mv(VP56mv *dst, const VP56mv *src,
VP9Context *s)
{
dst->x = av_clip(src->x, s->min_mv.x, s->max_mv.x);
dst->y = av_clip(src->y, s->min_mv.y, s->max_mv.y);
}
static void find_ref_mvs(VP9Context *s,
VP56mv *pmv, int ref, int z, int idx, int sb)
{
static const int8_t mv_ref_blk_off[N_BS_SIZES][8][2] = {
[BS_64x64] = {{ 3, -1 }, { -1, 3 }, { 4, -1 }, { -1, 4 },
{ -1, -1 }, { 0, -1 }, { -1, 0 }, { 6, -1 }},
[BS_64x32] = {{ 0, -1 }, { -1, 0 }, { 4, -1 }, { -1, 2 },
{ -1, -1 }, { 0, -3 }, { -3, 0 }, { 2, -1 }},
[BS_32x64] = {{ -1, 0 }, { 0, -1 }, { -1, 4 }, { 2, -1 },
{ -1, -1 }, { -3, 0 }, { 0, -3 }, { -1, 2 }},
[BS_32x32] = {{ 1, -1 }, { -1, 1 }, { 2, -1 }, { -1, 2 },
{ -1, -1 }, { 0, -3 }, { -3, 0 }, { -3, -3 }},
[BS_32x16] = {{ 0, -1 }, { -1, 0 }, { 2, -1 }, { -1, -1 },
{ -1, 1 }, { 0, -3 }, { -3, 0 }, { -3, -3 }},
[BS_16x32] = {{ -1, 0 }, { 0, -1 }, { -1, 2 }, { -1, -1 },
{ 1, -1 }, { -3, 0 }, { 0, -3 }, { -3, -3 }},
[BS_16x16] = {{ 0, -1 }, { -1, 0 }, { 1, -1 }, { -1, 1 },
{ -1, -1 }, { 0, -3 }, { -3, 0 }, { -3, -3 }},
[BS_16x8] = {{ 0, -1 }, { -1, 0 }, { 1, -1 }, { -1, -1 },
{ 0, -2 }, { -2, 0 }, { -2, -1 }, { -1, -2 }},
[BS_8x16] = {{ -1, 0 }, { 0, -1 }, { -1, 1 }, { -1, -1 },
{ -2, 0 }, { 0, -2 }, { -1, -2 }, { -2, -1 }},
[BS_8x8] = {{ 0, -1 }, { -1, 0 }, { -1, -1 }, { 0, -2 },
{ -2, 0 }, { -1, -2 }, { -2, -1 }, { -2, -2 }},
[BS_8x4] = {{ 0, -1 }, { -1, 0 }, { -1, -1 }, { 0, -2 },
{ -2, 0 }, { -1, -2 }, { -2, -1 }, { -2, -2 }},
[BS_4x8] = {{ 0, -1 }, { -1, 0 }, { -1, -1 }, { 0, -2 },
{ -2, 0 }, { -1, -2 }, { -2, -1 }, { -2, -2 }},
[BS_4x4] = {{ 0, -1 }, { -1, 0 }, { -1, -1 }, { 0, -2 },
{ -2, 0 }, { -1, -2 }, { -2, -1 }, { -2, -2 }},
};
VP9Block *b = s->b;
int row = s->row, col = s->col, row7 = s->row7;
const int8_t (*p)[2] = mv_ref_blk_off[b->bs];
#define INVALID_MV 0x80008000U
uint32_t mem = INVALID_MV, mem_sub8x8 = INVALID_MV;
int i;
#define RETURN_DIRECT_MV(mv) \
do { \
uint32_t m = AV_RN32A(&mv); \
if (!idx) { \
AV_WN32A(pmv, m); \
return; \
} else if (mem == INVALID_MV) { \
mem = m; \
} else if (m != mem) { \
AV_WN32A(pmv, m); \
return; \
} \
} while (0)
if (sb >= 0) {
if (sb == 2 || sb == 1) {
RETURN_DIRECT_MV(b->mv[0][z]);
} else if (sb == 3) {
RETURN_DIRECT_MV(b->mv[2][z]);
RETURN_DIRECT_MV(b->mv[1][z]);
RETURN_DIRECT_MV(b->mv[0][z]);
}
#define RETURN_MV(mv) \
do { \
if (sb > 0) { \
VP56mv tmp; \
uint32_t m; \
av_assert2(idx == 1); \
av_assert2(mem != INVALID_MV); \
if (mem_sub8x8 == INVALID_MV) { \
clamp_mv(&tmp, &mv, s); \
m = AV_RN32A(&tmp); \
if (m != mem) { \
AV_WN32A(pmv, m); \
return; \
} \
mem_sub8x8 = AV_RN32A(&mv); \
} else if (mem_sub8x8 != AV_RN32A(&mv)) { \
clamp_mv(&tmp, &mv, s); \
m = AV_RN32A(&tmp); \
if (m != mem) { \
AV_WN32A(pmv, m); \
} else { \
/* BUG I'm pretty sure this isn't the intention */ \
AV_WN32A(pmv, 0); \
} \
return; \
} \
} else { \
uint32_t m = AV_RN32A(&mv); \
if (!idx) { \
clamp_mv(pmv, &mv, s); \
return; \
} else if (mem == INVALID_MV) { \
mem = m; \
} else if (m != mem) { \
clamp_mv(pmv, &mv, s); \
return; \
} \
} \
} while (0)
if (row > 0) {
struct VP9mvrefPair *mv = &s->s.frames[CUR_FRAME].mv[(row - 1) * s->sb_cols * 8 + col];
if (mv->ref[0] == ref) {
RETURN_MV(s->above_mv_ctx[2 * col + (sb & 1)][0]);
} else if (mv->ref[1] == ref) {
RETURN_MV(s->above_mv_ctx[2 * col + (sb & 1)][1]);
}
}
if (col > s->tile_col_start) {
struct VP9mvrefPair *mv = &s->s.frames[CUR_FRAME].mv[row * s->sb_cols * 8 + col - 1];
if (mv->ref[0] == ref) {
RETURN_MV(s->left_mv_ctx[2 * row7 + (sb >> 1)][0]);
} else if (mv->ref[1] == ref) {
RETURN_MV(s->left_mv_ctx[2 * row7 + (sb >> 1)][1]);
}
}
i = 2;
} else {
i = 0;
}
// previously coded MVs in this neighbourhood, using same reference frame
for (; i < 8; i++) {
int c = p[i][0] + col, r = p[i][1] + row;
if (c >= s->tile_col_start && c < s->cols && r >= 0 && r < s->rows) {
struct VP9mvrefPair *mv = &s->s.frames[CUR_FRAME].mv[r * s->sb_cols * 8 + c];
if (mv->ref[0] == ref) {
RETURN_MV(mv->mv[0]);
} else if (mv->ref[1] == ref) {
RETURN_MV(mv->mv[1]);
}
}
}
// MV at this position in previous frame, using same reference frame
if (s->s.h.use_last_frame_mvs) {
struct VP9mvrefPair *mv = &s->s.frames[REF_FRAME_MVPAIR].mv[row * s->sb_cols * 8 + col];
if (!s->s.frames[REF_FRAME_MVPAIR].uses_2pass)
ff_thread_await_progress(&s->s.frames[REF_FRAME_MVPAIR].tf, row >> 3, 0);
if (mv->ref[0] == ref) {
RETURN_MV(mv->mv[0]);
} else if (mv->ref[1] == ref) {
RETURN_MV(mv->mv[1]);
}
}
#define RETURN_SCALE_MV(mv, scale) \
do { \
if (scale) { \
VP56mv mv_temp = { -mv.x, -mv.y }; \
RETURN_MV(mv_temp); \
} else { \
RETURN_MV(mv); \
} \
} while (0)
// previously coded MVs in this neighbourhood, using different reference frame
for (i = 0; i < 8; i++) {
int c = p[i][0] + col, r = p[i][1] + row;
if (c >= s->tile_col_start && c < s->cols && r >= 0 && r < s->rows) {
struct VP9mvrefPair *mv = &s->s.frames[CUR_FRAME].mv[r * s->sb_cols * 8 + c];
if (mv->ref[0] != ref && mv->ref[0] >= 0) {
RETURN_SCALE_MV(mv->mv[0], s->s.h.signbias[mv->ref[0]] != s->s.h.signbias[ref]);
}
if (mv->ref[1] != ref && mv->ref[1] >= 0 &&
// BUG - libvpx has this condition regardless of whether
// we used the first ref MV and pre-scaling
AV_RN32A(&mv->mv[0]) != AV_RN32A(&mv->mv[1])) {
RETURN_SCALE_MV(mv->mv[1], s->s.h.signbias[mv->ref[1]] != s->s.h.signbias[ref]);
}
}
}
// MV at this position in previous frame, using different reference frame
if (s->s.h.use_last_frame_mvs) {
struct VP9mvrefPair *mv = &s->s.frames[REF_FRAME_MVPAIR].mv[row * s->sb_cols * 8 + col];
// no need to await_progress, because we already did that above
if (mv->ref[0] != ref && mv->ref[0] >= 0) {
RETURN_SCALE_MV(mv->mv[0], s->s.h.signbias[mv->ref[0]] != s->s.h.signbias[ref]);
}
if (mv->ref[1] != ref && mv->ref[1] >= 0 &&
// BUG - libvpx has this condition regardless of whether
// we used the first ref MV and pre-scaling
AV_RN32A(&mv->mv[0]) != AV_RN32A(&mv->mv[1])) {
RETURN_SCALE_MV(mv->mv[1], s->s.h.signbias[mv->ref[1]] != s->s.h.signbias[ref]);
}
}
AV_ZERO32(pmv);
clamp_mv(pmv, pmv, s);
#undef INVALID_MV
#undef RETURN_MV
#undef RETURN_SCALE_MV
}
static av_always_inline int read_mv_component(VP9Context *s, int idx, int hp)
{
int bit, sign = vp56_rac_get_prob(&s->c, s->prob.p.mv_comp[idx].sign);
int n, c = vp8_rac_get_tree(&s->c, vp9_mv_class_tree,
s->prob.p.mv_comp[idx].classes);
s->counts.mv_comp[idx].sign[sign]++;
s->counts.mv_comp[idx].classes[c]++;
if (c) {
int m;
for (n = 0, m = 0; m < c; m++) {
bit = vp56_rac_get_prob(&s->c, s->prob.p.mv_comp[idx].bits[m]);
n |= bit << m;
s->counts.mv_comp[idx].bits[m][bit]++;
}
n <<= 3;
bit = vp8_rac_get_tree(&s->c, vp9_mv_fp_tree, s->prob.p.mv_comp[idx].fp);
n |= bit << 1;
s->counts.mv_comp[idx].fp[bit]++;
if (hp) {
bit = vp56_rac_get_prob(&s->c, s->prob.p.mv_comp[idx].hp);
s->counts.mv_comp[idx].hp[bit]++;
n |= bit;
} else {
n |= 1;
// bug in libvpx - we count for bw entropy purposes even if the
// bit wasn't coded
s->counts.mv_comp[idx].hp[1]++;
}
n += 8 << c;
} else {
n = vp56_rac_get_prob(&s->c, s->prob.p.mv_comp[idx].class0);
s->counts.mv_comp[idx].class0[n]++;
bit = vp8_rac_get_tree(&s->c, vp9_mv_fp_tree,
s->prob.p.mv_comp[idx].class0_fp[n]);
s->counts.mv_comp[idx].class0_fp[n][bit]++;
n = (n << 3) | (bit << 1);
if (hp) {
bit = vp56_rac_get_prob(&s->c, s->prob.p.mv_comp[idx].class0_hp);
s->counts.mv_comp[idx].class0_hp[bit]++;
n |= bit;
} else {
n |= 1;
// bug in libvpx - we count for bw entropy purposes even if the
// bit wasn't coded
s->counts.mv_comp[idx].class0_hp[1]++;
}
}
return sign ? -(n + 1) : (n + 1);
}
static void fill_mv(VP9Context *s,
VP56mv *mv, int mode, int sb)
{
VP9Block *b = s->b;
if (mode == ZEROMV) {
AV_ZERO64(mv);
} else {
int hp;
// FIXME cache this value and reuse for other subblocks
find_ref_mvs(s, &mv[0], b->ref[0], 0, mode == NEARMV,
mode == NEWMV ? -1 : sb);
// FIXME maybe move this code into find_ref_mvs()
if ((mode == NEWMV || sb == -1) &&
!(hp = s->s.h.highprecisionmvs && abs(mv[0].x) < 64 && abs(mv[0].y) < 64)) {
if (mv[0].y & 1) {
if (mv[0].y < 0)
mv[0].y++;
else
mv[0].y--;
}
if (mv[0].x & 1) {
if (mv[0].x < 0)
mv[0].x++;
else
mv[0].x--;
}
}
if (mode == NEWMV) {
enum MVJoint j = vp8_rac_get_tree(&s->c, vp9_mv_joint_tree,
s->prob.p.mv_joint);
s->counts.mv_joint[j]++;
if (j >= MV_JOINT_V)
mv[0].y += read_mv_component(s, 0, hp);
if (j & 1)
mv[0].x += read_mv_component(s, 1, hp);
}
if (b->comp) {
// FIXME cache this value and reuse for other subblocks
find_ref_mvs(s, &mv[1], b->ref[1], 1, mode == NEARMV,
mode == NEWMV ? -1 : sb);
if ((mode == NEWMV || sb == -1) &&
!(hp = s->s.h.highprecisionmvs && abs(mv[1].x) < 64 && abs(mv[1].y) < 64)) {
if (mv[1].y & 1) {
if (mv[1].y < 0)
mv[1].y++;
else
mv[1].y--;
}
if (mv[1].x & 1) {
if (mv[1].x < 0)
mv[1].x++;
else
mv[1].x--;
}
}
if (mode == NEWMV) {
enum MVJoint j = vp8_rac_get_tree(&s->c, vp9_mv_joint_tree,
s->prob.p.mv_joint);
s->counts.mv_joint[j]++;
if (j >= MV_JOINT_V)
mv[1].y += read_mv_component(s, 0, hp);
if (j & 1)
mv[1].x += read_mv_component(s, 1, hp);
}
}
}
}
static av_always_inline void setctx_2d(uint8_t *ptr, int w, int h,
ptrdiff_t stride, int v)
{
switch (w) {
case 1:
do {
*ptr = v;
ptr += stride;
} while (--h);
break;
case 2: {
int v16 = v * 0x0101;
do {
AV_WN16A(ptr, v16);
ptr += stride;
} while (--h);
break;
}
case 4: {
uint32_t v32 = v * 0x01010101;
do {
AV_WN32A(ptr, v32);
ptr += stride;
} while (--h);
break;
}
case 8: {
#if HAVE_FAST_64BIT
uint64_t v64 = v * 0x0101010101010101ULL;
do {
AV_WN64A(ptr, v64);
ptr += stride;
} while (--h);
#else
uint32_t v32 = v * 0x01010101;
do {
AV_WN32A(ptr, v32);
AV_WN32A(ptr + 4, v32);
ptr += stride;
} while (--h);
#endif
break;
}
}
}
static void decode_mode(AVCodecContext *ctx)
{
static const uint8_t left_ctx[N_BS_SIZES] = {
0x0, 0x8, 0x0, 0x8, 0xc, 0x8, 0xc, 0xe, 0xc, 0xe, 0xf, 0xe, 0xf
};
static const uint8_t above_ctx[N_BS_SIZES] = {
0x0, 0x0, 0x8, 0x8, 0x8, 0xc, 0xc, 0xc, 0xe, 0xe, 0xe, 0xf, 0xf
};
static const uint8_t max_tx_for_bl_bp[N_BS_SIZES] = {
TX_32X32, TX_32X32, TX_32X32, TX_32X32, TX_16X16, TX_16X16,
TX_16X16, TX_8X8, TX_8X8, TX_8X8, TX_4X4, TX_4X4, TX_4X4
};
VP9Context *s = ctx->priv_data;
VP9Block *b = s->b;
int row = s->row, col = s->col, row7 = s->row7;
enum TxfmMode max_tx = max_tx_for_bl_bp[b->bs];
int bw4 = bwh_tab[1][b->bs][0], w4 = FFMIN(s->cols - col, bw4);
int bh4 = bwh_tab[1][b->bs][1], h4 = FFMIN(s->rows - row, bh4), y;
int have_a = row > 0, have_l = col > s->tile_col_start;
int vref, filter_id;
if (!s->s.h.segmentation.enabled) {
b->seg_id = 0;
} else if (s->s.h.keyframe || s->s.h.intraonly) {
b->seg_id = !s->s.h.segmentation.update_map ? 0 :
vp8_rac_get_tree(&s->c, vp9_segmentation_tree, s->s.h.segmentation.prob);
} else if (!s->s.h.segmentation.update_map ||
(s->s.h.segmentation.temporal &&
vp56_rac_get_prob_branchy(&s->c,
s->s.h.segmentation.pred_prob[s->above_segpred_ctx[col] +
s->left_segpred_ctx[row7]]))) {
if (!s->s.h.errorres && s->s.frames[REF_FRAME_SEGMAP].segmentation_map) {
int pred = 8, x;
uint8_t *refsegmap = s->s.frames[REF_FRAME_SEGMAP].segmentation_map;
if (!s->s.frames[REF_FRAME_SEGMAP].uses_2pass)
ff_thread_await_progress(&s->s.frames[REF_FRAME_SEGMAP].tf, row >> 3, 0);
for (y = 0; y < h4; y++) {
int idx_base = (y + row) * 8 * s->sb_cols + col;
for (x = 0; x < w4; x++)
pred = FFMIN(pred, refsegmap[idx_base + x]);
}
av_assert1(pred < 8);
b->seg_id = pred;
} else {
b->seg_id = 0;
}
memset(&s->above_segpred_ctx[col], 1, w4);
memset(&s->left_segpred_ctx[row7], 1, h4);
} else {
b->seg_id = vp8_rac_get_tree(&s->c, vp9_segmentation_tree,
s->s.h.segmentation.prob);
memset(&s->above_segpred_ctx[col], 0, w4);
memset(&s->left_segpred_ctx[row7], 0, h4);
}
if (s->s.h.segmentation.enabled &&
(s->s.h.segmentation.update_map || s->s.h.keyframe || s->s.h.intraonly)) {
setctx_2d(&s->s.frames[CUR_FRAME].segmentation_map[row * 8 * s->sb_cols + col],
bw4, bh4, 8 * s->sb_cols, b->seg_id);
}
b->skip = s->s.h.segmentation.enabled &&
s->s.h.segmentation.feat[b->seg_id].skip_enabled;
if (!b->skip) {
int c = s->left_skip_ctx[row7] + s->above_skip_ctx[col];
b->skip = vp56_rac_get_prob(&s->c, s->prob.p.skip[c]);
s->counts.skip[c][b->skip]++;
}
if (s->s.h.keyframe || s->s.h.intraonly) {
b->intra = 1;
} else if (s->s.h.segmentation.enabled && s->s.h.segmentation.feat[b->seg_id].ref_enabled) {
b->intra = !s->s.h.segmentation.feat[b->seg_id].ref_val;
} else {
int c, bit;
if (have_a && have_l) {
c = s->above_intra_ctx[col] + s->left_intra_ctx[row7];
c += (c == 2);
} else {
c = have_a ? 2 * s->above_intra_ctx[col] :
have_l ? 2 * s->left_intra_ctx[row7] : 0;
}
bit = vp56_rac_get_prob(&s->c, s->prob.p.intra[c]);
s->counts.intra[c][bit]++;
b->intra = !bit;
}
if ((b->intra || !b->skip) && s->s.h.txfmmode == TX_SWITCHABLE) {
int c;
if (have_a) {
if (have_l) {
c = (s->above_skip_ctx[col] ? max_tx :
s->above_txfm_ctx[col]) +
(s->left_skip_ctx[row7] ? max_tx :
s->left_txfm_ctx[row7]) > max_tx;
} else {
c = s->above_skip_ctx[col] ? 1 :
(s->above_txfm_ctx[col] * 2 > max_tx);
}
} else if (have_l) {
c = s->left_skip_ctx[row7] ? 1 :
(s->left_txfm_ctx[row7] * 2 > max_tx);
} else {
c = 1;
}
switch (max_tx) {
case TX_32X32:
b->tx = vp56_rac_get_prob(&s->c, s->prob.p.tx32p[c][0]);
if (b->tx) {
b->tx += vp56_rac_get_prob(&s->c, s->prob.p.tx32p[c][1]);
if (b->tx == 2)
b->tx += vp56_rac_get_prob(&s->c, s->prob.p.tx32p[c][2]);
}
s->counts.tx32p[c][b->tx]++;
break;
case TX_16X16:
b->tx = vp56_rac_get_prob(&s->c, s->prob.p.tx16p[c][0]);
if (b->tx)
b->tx += vp56_rac_get_prob(&s->c, s->prob.p.tx16p[c][1]);
s->counts.tx16p[c][b->tx]++;
break;
case TX_8X8:
b->tx = vp56_rac_get_prob(&s->c, s->prob.p.tx8p[c]);
s->counts.tx8p[c][b->tx]++;
break;
case TX_4X4:
b->tx = TX_4X4;
break;
}
} else {
b->tx = FFMIN(max_tx, s->s.h.txfmmode);
}
if (s->s.h.keyframe || s->s.h.intraonly) {
uint8_t *a = &s->above_mode_ctx[col * 2];
uint8_t *l = &s->left_mode_ctx[(row7) << 1];
b->comp = 0;
if (b->bs > BS_8x8) {
// FIXME the memory storage intermediates here aren't really
// necessary, they're just there to make the code slightly
// simpler for now
b->mode[0] = a[0] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,
vp9_default_kf_ymode_probs[a[0]][l[0]]);
if (b->bs != BS_8x4) {
b->mode[1] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,
vp9_default_kf_ymode_probs[a[1]][b->mode[0]]);
l[0] = a[1] = b->mode[1];
} else {
l[0] = a[1] = b->mode[1] = b->mode[0];
}
if (b->bs != BS_4x8) {
b->mode[2] = a[0] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,
vp9_default_kf_ymode_probs[a[0]][l[1]]);
if (b->bs != BS_8x4) {
b->mode[3] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,
vp9_default_kf_ymode_probs[a[1]][b->mode[2]]);
l[1] = a[1] = b->mode[3];
} else {
l[1] = a[1] = b->mode[3] = b->mode[2];
}
} else {
b->mode[2] = b->mode[0];
l[1] = a[1] = b->mode[3] = b->mode[1];
}
} else {
b->mode[0] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,
vp9_default_kf_ymode_probs[*a][*l]);
b->mode[3] = b->mode[2] = b->mode[1] = b->mode[0];
// FIXME this can probably be optimized
memset(a, b->mode[0], bwh_tab[0][b->bs][0]);
memset(l, b->mode[0], bwh_tab[0][b->bs][1]);
}
b->uvmode = vp8_rac_get_tree(&s->c, vp9_intramode_tree,
vp9_default_kf_uvmode_probs[b->mode[3]]);
} else if (b->intra) {
b->comp = 0;
if (b->bs > BS_8x8) {
b->mode[0] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,
s->prob.p.y_mode[0]);
s->counts.y_mode[0][b->mode[0]]++;
if (b->bs != BS_8x4) {
b->mode[1] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,
s->prob.p.y_mode[0]);
s->counts.y_mode[0][b->mode[1]]++;
} else {
b->mode[1] = b->mode[0];
}
if (b->bs != BS_4x8) {
b->mode[2] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,
s->prob.p.y_mode[0]);
s->counts.y_mode[0][b->mode[2]]++;
if (b->bs != BS_8x4) {
b->mode[3] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,
s->prob.p.y_mode[0]);
s->counts.y_mode[0][b->mode[3]]++;
} else {
b->mode[3] = b->mode[2];
}
} else {
b->mode[2] = b->mode[0];
b->mode[3] = b->mode[1];
}
} else {
static const uint8_t size_group[10] = {
3, 3, 3, 3, 2, 2, 2, 1, 1, 1
};
int sz = size_group[b->bs];
b->mode[0] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,
s->prob.p.y_mode[sz]);
b->mode[1] = b->mode[2] = b->mode[3] = b->mode[0];
s->counts.y_mode[sz][b->mode[3]]++;
}
b->uvmode = vp8_rac_get_tree(&s->c, vp9_intramode_tree,
s->prob.p.uv_mode[b->mode[3]]);
s->counts.uv_mode[b->mode[3]][b->uvmode]++;
} else {
static const uint8_t inter_mode_ctx_lut[14][14] = {
{ 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },
{ 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },
{ 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },
{ 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },
{ 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },
{ 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },
{ 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },
{ 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },
{ 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },
{ 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },
{ 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 2, 2, 1, 3 },
{ 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 2, 2, 1, 3 },
{ 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 1, 1, 0, 3 },
{ 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 3, 3, 3, 4 },
};
if (s->s.h.segmentation.enabled && s->s.h.segmentation.feat[b->seg_id].ref_enabled) {
av_assert2(s->s.h.segmentation.feat[b->seg_id].ref_val != 0);
b->comp = 0;
b->ref[0] = s->s.h.segmentation.feat[b->seg_id].ref_val - 1;
} else {
// read comp_pred flag
if (s->s.h.comppredmode != PRED_SWITCHABLE) {
b->comp = s->s.h.comppredmode == PRED_COMPREF;
} else {
int c;
// FIXME add intra as ref=0xff (or -1) to make these easier?
if (have_a) {
if (have_l) {
if (s->above_comp_ctx[col] && s->left_comp_ctx[row7]) {
c = 4;
} else if (s->above_comp_ctx[col]) {
c = 2 + (s->left_intra_ctx[row7] ||
s->left_ref_ctx[row7] == s->s.h.fixcompref);
} else if (s->left_comp_ctx[row7]) {
c = 2 + (s->above_intra_ctx[col] ||
s->above_ref_ctx[col] == s->s.h.fixcompref);
} else {
c = (!s->above_intra_ctx[col] &&
s->above_ref_ctx[col] == s->s.h.fixcompref) ^
(!s->left_intra_ctx[row7] &&
s->left_ref_ctx[row & 7] == s->s.h.fixcompref);
}
} else {
c = s->above_comp_ctx[col] ? 3 :
(!s->above_intra_ctx[col] && s->above_ref_ctx[col] == s->s.h.fixcompref);
}
} else if (have_l) {
c = s->left_comp_ctx[row7] ? 3 :
(!s->left_intra_ctx[row7] && s->left_ref_ctx[row7] == s->s.h.fixcompref);
} else {
c = 1;
}
b->comp = vp56_rac_get_prob(&s->c, s->prob.p.comp[c]);
s->counts.comp[c][b->comp]++;
}
// read actual references
// FIXME probably cache a few variables here to prevent repetitive
// memory accesses below
if (b->comp) /* two references */ {
int fix_idx = s->s.h.signbias[s->s.h.fixcompref], var_idx = !fix_idx, c, bit;
b->ref[fix_idx] = s->s.h.fixcompref;
// FIXME can this codeblob be replaced by some sort of LUT?
if (have_a) {
if (have_l) {
if (s->above_intra_ctx[col]) {
if (s->left_intra_ctx[row7]) {
c = 2;
} else {
c = 1 + 2 * (s->left_ref_ctx[row7] != s->s.h.varcompref[1]);
}
} else if (s->left_intra_ctx[row7]) {
c = 1 + 2 * (s->above_ref_ctx[col] != s->s.h.varcompref[1]);
} else {
int refl = s->left_ref_ctx[row7], refa = s->above_ref_ctx[col];
if (refl == refa && refa == s->s.h.varcompref[1]) {
c = 0;
} else if (!s->left_comp_ctx[row7] && !s->above_comp_ctx[col]) {
if ((refa == s->s.h.fixcompref && refl == s->s.h.varcompref[0]) ||
(refl == s->s.h.fixcompref && refa == s->s.h.varcompref[0])) {
c = 4;
} else {
c = (refa == refl) ? 3 : 1;
}
} else if (!s->left_comp_ctx[row7]) {
if (refa == s->s.h.varcompref[1] && refl != s->s.h.varcompref[1]) {
c = 1;
} else {
c = (refl == s->s.h.varcompref[1] &&
refa != s->s.h.varcompref[1]) ? 2 : 4;
}
} else if (!s->above_comp_ctx[col]) {
if (refl == s->s.h.varcompref[1] && refa != s->s.h.varcompref[1]) {
c = 1;
} else {
c = (refa == s->s.h.varcompref[1] &&
refl != s->s.h.varcompref[1]) ? 2 : 4;
}
} else {
c = (refl == refa) ? 4 : 2;
}
}
} else {
if (s->above_intra_ctx[col]) {
c = 2;
} else if (s->above_comp_ctx[col]) {
c = 4 * (s->above_ref_ctx[col] != s->s.h.varcompref[1]);
} else {
c = 3 * (s->above_ref_ctx[col] != s->s.h.varcompref[1]);
}
}
} else if (have_l) {
if (s->left_intra_ctx[row7]) {
c = 2;
} else if (s->left_comp_ctx[row7]) {
c = 4 * (s->left_ref_ctx[row7] != s->s.h.varcompref[1]);
} else {
c = 3 * (s->left_ref_ctx[row7] != s->s.h.varcompref[1]);
}
} else {
c = 2;
}
bit = vp56_rac_get_prob(&s->c, s->prob.p.comp_ref[c]);
b->ref[var_idx] = s->s.h.varcompref[bit];
s->counts.comp_ref[c][bit]++;
} else /* single reference */ {
int bit, c;
if (have_a && !s->above_intra_ctx[col]) {
if (have_l && !s->left_intra_ctx[row7]) {
if (s->left_comp_ctx[row7]) {
if (s->above_comp_ctx[col]) {
c = 1 + (!s->s.h.fixcompref || !s->left_ref_ctx[row7] ||
!s->above_ref_ctx[col]);
} else {
c = (3 * !s->above_ref_ctx[col]) +
(!s->s.h.fixcompref || !s->left_ref_ctx[row7]);
}
} else if (s->above_comp_ctx[col]) {
c = (3 * !s->left_ref_ctx[row7]) +
(!s->s.h.fixcompref || !s->above_ref_ctx[col]);
} else {
c = 2 * !s->left_ref_ctx[row7] + 2 * !s->above_ref_ctx[col];
}
} else if (s->above_intra_ctx[col]) {
c = 2;
} else if (s->above_comp_ctx[col]) {
c = 1 + (!s->s.h.fixcompref || !s->above_ref_ctx[col]);
} else {
c = 4 * (!s->above_ref_ctx[col]);
}
} else if (have_l && !s->left_intra_ctx[row7]) {
if (s->left_intra_ctx[row7]) {
c = 2;
} else if (s->left_comp_ctx[row7]) {
c = 1 + (!s->s.h.fixcompref || !s->left_ref_ctx[row7]);
} else {
c = 4 * (!s->left_ref_ctx[row7]);
}
} else {
c = 2;
}
bit = vp56_rac_get_prob(&s->c, s->prob.p.single_ref[c][0]);
s->counts.single_ref[c][0][bit]++;
if (!bit) {
b->ref[0] = 0;
} else {
// FIXME can this codeblob be replaced by some sort of LUT?
if (have_a) {
if (have_l) {
if (s->left_intra_ctx[row7]) {
if (s->above_intra_ctx[col]) {
c = 2;
} else if (s->above_comp_ctx[col]) {
c = 1 + 2 * (s->s.h.fixcompref == 1 ||
s->above_ref_ctx[col] == 1);
} else if (!s->above_ref_ctx[col]) {
c = 3;
} else {
c = 4 * (s->above_ref_ctx[col] == 1);
}
} else if (s->above_intra_ctx[col]) {
if (s->left_intra_ctx[row7]) {
c = 2;
} else if (s->left_comp_ctx[row7]) {
c = 1 + 2 * (s->s.h.fixcompref == 1 ||
s->left_ref_ctx[row7] == 1);
} else if (!s->left_ref_ctx[row7]) {
c = 3;
} else {
c = 4 * (s->left_ref_ctx[row7] == 1);
}
} else if (s->above_comp_ctx[col]) {
if (s->left_comp_ctx[row7]) {
if (s->left_ref_ctx[row7] == s->above_ref_ctx[col]) {
c = 3 * (s->s.h.fixcompref == 1 ||
s->left_ref_ctx[row7] == 1);
} else {
c = 2;
}
} else if (!s->left_ref_ctx[row7]) {
c = 1 + 2 * (s->s.h.fixcompref == 1 ||
s->above_ref_ctx[col] == 1);
} else {
c = 3 * (s->left_ref_ctx[row7] == 1) +
(s->s.h.fixcompref == 1 || s->above_ref_ctx[col] == 1);
}
} else if (s->left_comp_ctx[row7]) {
if (!s->above_ref_ctx[col]) {
c = 1 + 2 * (s->s.h.fixcompref == 1 ||
s->left_ref_ctx[row7] == 1);
} else {
c = 3 * (s->above_ref_ctx[col] == 1) +
(s->s.h.fixcompref == 1 || s->left_ref_ctx[row7] == 1);
}
} else if (!s->above_ref_ctx[col]) {
if (!s->left_ref_ctx[row7]) {
c = 3;
} else {
c = 4 * (s->left_ref_ctx[row7] == 1);
}
} else if (!s->left_ref_ctx[row7]) {
c = 4 * (s->above_ref_ctx[col] == 1);
} else {
c = 2 * (s->left_ref_ctx[row7] == 1) +
2 * (s->above_ref_ctx[col] == 1);
}
} else {
if (s->above_intra_ctx[col] ||
(!s->above_comp_ctx[col] && !s->above_ref_ctx[col])) {
c = 2;
} else if (s->above_comp_ctx[col]) {
c = 3 * (s->s.h.fixcompref == 1 || s->above_ref_ctx[col] == 1);
} else {
c = 4 * (s->above_ref_ctx[col] == 1);
}
}
} else if (have_l) {
if (s->left_intra_ctx[row7] ||
(!s->left_comp_ctx[row7] && !s->left_ref_ctx[row7])) {
c = 2;
} else if (s->left_comp_ctx[row7]) {
c = 3 * (s->s.h.fixcompref == 1 || s->left_ref_ctx[row7] == 1);
} else {
c = 4 * (s->left_ref_ctx[row7] == 1);
}
} else {
c = 2;
}
bit = vp56_rac_get_prob(&s->c, s->prob.p.single_ref[c][1]);
s->counts.single_ref[c][1][bit]++;
b->ref[0] = 1 + bit;
}
}
}
if (b->bs <= BS_8x8) {
if (s->s.h.segmentation.enabled && s->s.h.segmentation.feat[b->seg_id].skip_enabled) {
b->mode[0] = b->mode[1] = b->mode[2] = b->mode[3] = ZEROMV;
} else {
static const uint8_t off[10] = {
3, 0, 0, 1, 0, 0, 0, 0, 0, 0
};
// FIXME this needs to use the LUT tables from find_ref_mvs
// because not all are -1,0/0,-1
int c = inter_mode_ctx_lut[s->above_mode_ctx[col + off[b->bs]]]
[s->left_mode_ctx[row7 + off[b->bs]]];
b->mode[0] = vp8_rac_get_tree(&s->c, vp9_inter_mode_tree,
s->prob.p.mv_mode[c]);
b->mode[1] = b->mode[2] = b->mode[3] = b->mode[0];
s->counts.mv_mode[c][b->mode[0] - 10]++;
}
}
if (s->s.h.filtermode == FILTER_SWITCHABLE) {
int c;
if (have_a && s->above_mode_ctx[col] >= NEARESTMV) {
if (have_l && s->left_mode_ctx[row7] >= NEARESTMV) {
c = s->above_filter_ctx[col] == s->left_filter_ctx[row7] ?
s->left_filter_ctx[row7] : 3;
} else {
c = s->above_filter_ctx[col];
}
} else if (have_l && s->left_mode_ctx[row7] >= NEARESTMV) {
c = s->left_filter_ctx[row7];
} else {
c = 3;
}
filter_id = vp8_rac_get_tree(&s->c, vp9_filter_tree,
s->prob.p.filter[c]);
s->counts.filter[c][filter_id]++;
b->filter = vp9_filter_lut[filter_id];
} else {
b->filter = s->s.h.filtermode;
}
if (b->bs > BS_8x8) {
int c = inter_mode_ctx_lut[s->above_mode_ctx[col]][s->left_mode_ctx[row7]];
b->mode[0] = vp8_rac_get_tree(&s->c, vp9_inter_mode_tree,
s->prob.p.mv_mode[c]);
s->counts.mv_mode[c][b->mode[0] - 10]++;
fill_mv(s, b->mv[0], b->mode[0], 0);
if (b->bs != BS_8x4) {
b->mode[1] = vp8_rac_get_tree(&s->c, vp9_inter_mode_tree,
s->prob.p.mv_mode[c]);
s->counts.mv_mode[c][b->mode[1] - 10]++;
fill_mv(s, b->mv[1], b->mode[1], 1);
} else {
b->mode[1] = b->mode[0];
AV_COPY32(&b->mv[1][0], &b->mv[0][0]);
AV_COPY32(&b->mv[1][1], &b->mv[0][1]);
}
if (b->bs != BS_4x8) {
b->mode[2] = vp8_rac_get_tree(&s->c, vp9_inter_mode_tree,
s->prob.p.mv_mode[c]);
s->counts.mv_mode[c][b->mode[2] - 10]++;
fill_mv(s, b->mv[2], b->mode[2], 2);
if (b->bs != BS_8x4) {
b->mode[3] = vp8_rac_get_tree(&s->c, vp9_inter_mode_tree,
s->prob.p.mv_mode[c]);
s->counts.mv_mode[c][b->mode[3] - 10]++;
fill_mv(s, b->mv[3], b->mode[3], 3);
} else {
b->mode[3] = b->mode[2];
AV_COPY32(&b->mv[3][0], &b->mv[2][0]);
AV_COPY32(&b->mv[3][1], &b->mv[2][1]);
}
} else {
b->mode[2] = b->mode[0];
AV_COPY32(&b->mv[2][0], &b->mv[0][0]);
AV_COPY32(&b->mv[2][1], &b->mv[0][1]);
b->mode[3] = b->mode[1];
AV_COPY32(&b->mv[3][0], &b->mv[1][0]);
AV_COPY32(&b->mv[3][1], &b->mv[1][1]);
}
} else {
fill_mv(s, b->mv[0], b->mode[0], -1);
AV_COPY32(&b->mv[1][0], &b->mv[0][0]);
AV_COPY32(&b->mv[2][0], &b->mv[0][0]);
AV_COPY32(&b->mv[3][0], &b->mv[0][0]);
AV_COPY32(&b->mv[1][1], &b->mv[0][1]);
AV_COPY32(&b->mv[2][1], &b->mv[0][1]);
AV_COPY32(&b->mv[3][1], &b->mv[0][1]);
}
vref = b->ref[b->comp ? s->s.h.signbias[s->s.h.varcompref[0]] : 0];
}
#if HAVE_FAST_64BIT
#define SPLAT_CTX(var, val, n) \
switch (n) { \
case 1: var = val; break; \
case 2: AV_WN16A(&var, val * 0x0101); break; \
case 4: AV_WN32A(&var, val * 0x01010101); break; \
case 8: AV_WN64A(&var, val * 0x0101010101010101ULL); break; \
case 16: { \
uint64_t v64 = val * 0x0101010101010101ULL; \
AV_WN64A( &var, v64); \
AV_WN64A(&((uint8_t *) &var)[8], v64); \
break; \
} \
}
#else
#define SPLAT_CTX(var, val, n) \
switch (n) { \
case 1: var = val; break; \
case 2: AV_WN16A(&var, val * 0x0101); break; \
case 4: AV_WN32A(&var, val * 0x01010101); break; \
case 8: { \
uint32_t v32 = val * 0x01010101; \
AV_WN32A( &var, v32); \
AV_WN32A(&((uint8_t *) &var)[4], v32); \