libavfilter/vf_thumbnail_cuda.c - manifest_repos/ffmpeg - Git at Google

 /*
 * Copyright (c) 2017, NVIDIA CORPORATION. All rights reserved.
 *
 * Permission is hereby granted, free of charge, to any person obtaining a
 * copy of this software and associated documentation files (the "Software"),
 * to deal in the Software without restriction, including without limitation
 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
 * and/or sell copies of the Software, and to permit persons to whom the
 * Software is furnished to do so, subject to the following conditions:
 *
 * The above copyright notice and this permission notice shall be included in
 * all copies or substantial portions of the Software.
 *
 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
 * DEALINGS IN THE SOFTWARE.
 */

 #include "libavutil/hwcontext.h"
 #include "libavutil/hwcontext_cuda_internal.h"
 #include "libavutil/cuda_check.h"
 #include "libavutil/opt.h"
 #include "libavutil/pixdesc.h"

 #include "avfilter.h"
 #include "internal.h"

 #define CHECK_CU(x) FF_CUDA_CHECK_DL(ctx, s->hwctx->internal->cuda_dl, x)

 #define HIST_SIZE (3*256)
 #define DIV_UP(a, b) ( ((a) + (b) - 1) / (b) )
 #define BLOCKX 32
 #define BLOCKY 16

 static const enum AVPixelFormat supported_formats[] = {
     AV_PIX_FMT_NV12,
     AV_PIX_FMT_YUV420P,
     AV_PIX_FMT_YUV444P,
     AV_PIX_FMT_P010,
     AV_PIX_FMT_P016,
     AV_PIX_FMT_YUV444P16,
 };

 struct thumb_frame {
     AVFrame *buf;               ///< cached frame
     int histogram[HIST_SIZE];   ///< RGB color distribution histogram of the frame
 };

 typedef struct ThumbnailCudaContext {
     const AVClass *class;
     int n;                      ///< current frame
     int n_frames;               ///< number of frames for analysis
     struct thumb_frame *frames; ///< the n_frames frames
     AVRational tb;              ///< copy of the input timebase to ease access

     AVBufferRef *hw_frames_ctx;
     AVCUDADeviceContext *hwctx;

     CUmodule    cu_module;

     CUfunction  cu_func_uchar;
     CUfunction  cu_func_uchar2;
     CUfunction  cu_func_ushort;
     CUfunction  cu_func_ushort2;
     CUstream    cu_stream;

     CUdeviceptr data;

 } ThumbnailCudaContext;

 #define OFFSET(x) offsetof(ThumbnailCudaContext, x)
 #define FLAGS AV_OPT_FLAG_VIDEO_PARAM|AV_OPT_FLAG_FILTERING_PARAM

 static const AVOption thumbnail_cuda_options[] = {
     { "n", "set the frames batch size", OFFSET(n_frames), AV_OPT_TYPE_INT, {.i64=100}, 2, INT_MAX, FLAGS },
     { NULL }
 };

 AVFILTER_DEFINE_CLASS(thumbnail_cuda);

 static av_cold int init(AVFilterContext *ctx)
 {
     ThumbnailCudaContext *s = ctx->priv;

     s->frames = av_calloc(s->n_frames, sizeof(*s->frames));
     if (!s->frames) {
         av_log(ctx, AV_LOG_ERROR,
                "Allocation failure, try to lower the number of frames\n");
         return AVERROR(ENOMEM);
     }
     av_log(ctx, AV_LOG_VERBOSE, "batch size: %d frames\n", s->n_frames);
     return 0;
 }

 /**
  * @brief        Compute Sum-square deviation to estimate "closeness".
  * @param hist   color distribution histogram
  * @param median average color distribution histogram
  * @return       sum of squared errors
  */
 static double frame_sum_square_err(const int *hist, const double *median)
 {
     int i;
     double err, sum_sq_err = 0;

     for (i = 0; i < HIST_SIZE; i++) {
         err = median[i] - (double)hist[i];
         sum_sq_err += err*err;
     }
     return sum_sq_err;
 }

 static AVFrame *get_best_frame(AVFilterContext *ctx)
 {
     AVFrame *picref;
     ThumbnailCudaContext *s = ctx->priv;
     int i, j, best_frame_idx = 0;
     int nb_frames = s->n;
     double avg_hist[HIST_SIZE] = {0}, sq_err, min_sq_err = -1;

     // average histogram of the N frames
     for (j = 0; j < FF_ARRAY_ELEMS(avg_hist); j++) {
         for (i = 0; i < nb_frames; i++)
             avg_hist[j] += (double)s->frames[i].histogram[j];
         avg_hist[j] /= nb_frames;
     }

     // find the frame closer to the average using the sum of squared errors
     for (i = 0; i < nb_frames; i++) {
         sq_err = frame_sum_square_err(s->frames[i].histogram, avg_hist);
         if (i == 0 || sq_err < min_sq_err)
             best_frame_idx = i, min_sq_err = sq_err;
     }

     // free and reset everything (except the best frame buffer)
     for (i = 0; i < nb_frames; i++) {
         memset(s->frames[i].histogram, 0, sizeof(s->frames[i].histogram));
         if (i != best_frame_idx)
             av_frame_free(&s->frames[i].buf);
     }
     s->n = 0;

     // raise the chosen one
     picref = s->frames[best_frame_idx].buf;
     av_log(ctx, AV_LOG_INFO, "frame id #%d (pts_time=%f) selected "
            "from a set of %d images\n", best_frame_idx,
            picref->pts * av_q2d(s->tb), nb_frames);
     s->frames[best_frame_idx].buf = NULL;

     return picref;
 }

 static int thumbnail_kernel(AVFilterContext *ctx, CUfunction func, int channels,
     int *histogram, uint8_t *src_dptr, int src_width, int src_height, int src_pitch, int pixel_size)
 {
     int ret;
     ThumbnailCudaContext *s = ctx->priv;
     CudaFunctions *cu = s->hwctx->internal->cuda_dl;
     CUtexObject tex = 0;
     void *args[] = { &tex, &histogram, &src_width, &src_height };

     CUDA_TEXTURE_DESC tex_desc = {
         .filterMode = CU_TR_FILTER_MODE_LINEAR,
         .flags = CU_TRSF_READ_AS_INTEGER,
     };

     CUDA_RESOURCE_DESC res_desc = {
         .resType = CU_RESOURCE_TYPE_PITCH2D,
         .res.pitch2D.format = pixel_size == 1 ?
                               CU_AD_FORMAT_UNSIGNED_INT8 :
                               CU_AD_FORMAT_UNSIGNED_INT16,
         .res.pitch2D.numChannels = channels,
         .res.pitch2D.width = src_width,
         .res.pitch2D.height = src_height,
         .res.pitch2D.pitchInBytes = src_pitch,
         .res.pitch2D.devPtr = (CUdeviceptr)src_dptr,
     };

     ret = CHECK_CU(cu->cuTexObjectCreate(&tex, &res_desc, &tex_desc, NULL));
     if (ret < 0)
         goto exit;

     ret = CHECK_CU(cu->cuLaunchKernel(func,
                                       DIV_UP(src_width, BLOCKX), DIV_UP(src_height, BLOCKY), 1,
                                       BLOCKX, BLOCKY, 1, 0, s->cu_stream, args, NULL));
 exit:
     if (tex)
         CHECK_CU(cu->cuTexObjectDestroy(tex));

     return ret;
 }

 static int thumbnail(AVFilterContext *ctx, int *histogram, AVFrame *in)
 {
     AVHWFramesContext *in_frames_ctx = (AVHWFramesContext*)in->hw_frames_ctx->data;
     ThumbnailCudaContext *s = ctx->priv;

     switch (in_frames_ctx->sw_format) {
     case AV_PIX_FMT_NV12:
         thumbnail_kernel(ctx, s->cu_func_uchar, 1,
             histogram, in->data[0], in->width, in->height, in->linesize[0], 1);
         thumbnail_kernel(ctx, s->cu_func_uchar2, 2,
             histogram + 256, in->data[1], in->width / 2, in->height / 2, in->linesize[1], 1);
         break;
     case AV_PIX_FMT_YUV420P:
         thumbnail_kernel(ctx, s->cu_func_uchar, 1,
             histogram, in->data[0], in->width, in->height, in->linesize[0], 1);
         thumbnail_kernel(ctx, s->cu_func_uchar, 1,
             histogram + 256, in->data[1], in->width / 2, in->height / 2, in->linesize[1], 1);
         thumbnail_kernel(ctx, s->cu_func_uchar, 1,
             histogram + 512, in->data[2], in->width / 2, in->height / 2, in->linesize[2], 1);
         break;
     case AV_PIX_FMT_YUV444P:
         thumbnail_kernel(ctx, s->cu_func_uchar, 1,
             histogram, in->data[0], in->width, in->height, in->linesize[0], 1);
         thumbnail_kernel(ctx, s->cu_func_uchar, 1,
             histogram + 256, in->data[1], in->width, in->height, in->linesize[1], 1);
         thumbnail_kernel(ctx, s->cu_func_uchar, 1,
             histogram + 512, in->data[2], in->width, in->height, in->linesize[2], 1);
         break;
     case AV_PIX_FMT_P010LE:
     case AV_PIX_FMT_P016LE:
         thumbnail_kernel(ctx, s->cu_func_ushort, 1,
             histogram, in->data[0], in->width, in->height, in->linesize[0], 2);
         thumbnail_kernel(ctx, s->cu_func_ushort2, 2,
             histogram + 256, in->data[1], in->width / 2, in->height / 2, in->linesize[1], 2);
         break;
     case AV_PIX_FMT_YUV444P16:
         thumbnail_kernel(ctx, s->cu_func_ushort2, 1,
             histogram, in->data[0], in->width, in->height, in->linesize[0], 2);
         thumbnail_kernel(ctx, s->cu_func_ushort2, 1,
             histogram + 256, in->data[1], in->width, in->height, in->linesize[1], 2);
         thumbnail_kernel(ctx, s->cu_func_ushort2, 1,
             histogram + 512, in->data[2], in->width, in->height, in->linesize[2], 2);
         break;
     default:
         return AVERROR_BUG;
     }

     return 0;
 }

 static int filter_frame(AVFilterLink *inlink, AVFrame *frame)
 {
     AVFilterContext *ctx  = inlink->dst;
     ThumbnailCudaContext *s   = ctx->priv;
     CudaFunctions *cu = s->hwctx->internal->cuda_dl;
     AVFilterLink *outlink = ctx->outputs[0];
     int *hist = s->frames[s->n].histogram;
     AVHWFramesContext *hw_frames_ctx = (AVHWFramesContext*)s->hw_frames_ctx->data;
     CUcontext dummy;
     CUDA_MEMCPY2D cpy = { 0 };
     int ret = 0;

     // keep a reference of each frame
     s->frames[s->n].buf = frame;

     ret = CHECK_CU(cu->cuCtxPushCurrent(s->hwctx->cuda_ctx));
     if (ret < 0)
         return ret;

     CHECK_CU(cu->cuMemsetD8Async(s->data, 0, HIST_SIZE * sizeof(int), s->cu_stream));

     thumbnail(ctx, (int*)s->data, frame);

     cpy.srcMemoryType = CU_MEMORYTYPE_DEVICE;
     cpy.dstMemoryType = CU_MEMORYTYPE_HOST;
     cpy.srcDevice = s->data;
     cpy.dstHost = hist;
     cpy.srcPitch = HIST_SIZE * sizeof(int);
     cpy.dstPitch = HIST_SIZE * sizeof(int);
     cpy.WidthInBytes = HIST_SIZE * sizeof(int);
     cpy.Height = 1;

     ret = CHECK_CU(cu->cuMemcpy2DAsync(&cpy, s->cu_stream));
     if (ret < 0)
         return ret;

     if (hw_frames_ctx->sw_format == AV_PIX_FMT_NV12 || hw_frames_ctx->sw_format == AV_PIX_FMT_YUV420P ||
         hw_frames_ctx->sw_format == AV_PIX_FMT_P010LE || hw_frames_ctx->sw_format == AV_PIX_FMT_P016LE)
     {
         int i;
         for (i = 256; i < HIST_SIZE; i++)
             hist[i] = 4 * hist[i];
     }

     CHECK_CU(cu->cuCtxPopCurrent(&dummy));
     if (ret < 0)
         return ret;

     // no selection until the buffer of N frames is filled up
     s->n++;
     if (s->n < s->n_frames)
         return 0;

     return ff_filter_frame(outlink, get_best_frame(ctx));
 }

 static av_cold void uninit(AVFilterContext *ctx)
 {
     int i;
     ThumbnailCudaContext *s = ctx->priv;
     CudaFunctions *cu = s->hwctx->internal->cuda_dl;

     if (s->data) {
         CHECK_CU(cu->cuMemFree(s->data));
         s->data = 0;
     }

     if (s->cu_module) {
         CHECK_CU(cu->cuModuleUnload(s->cu_module));
         s->cu_module = NULL;
     }

     for (i = 0; i < s->n_frames && s->frames[i].buf; i++)
         av_frame_free(&s->frames[i].buf);
     av_freep(&s->frames);
 }

 static int request_frame(AVFilterLink *link)
 {
     AVFilterContext *ctx = link->src;
     ThumbnailCudaContext *s = ctx->priv;
     int ret = ff_request_frame(ctx->inputs[0]);

     if (ret == AVERROR_EOF && s->n) {
         ret = ff_filter_frame(link, get_best_frame(ctx));
         if (ret < 0)
             return ret;
         ret = AVERROR_EOF;
     }
     if (ret < 0)
         return ret;
     return 0;
 }

 static int format_is_supported(enum AVPixelFormat fmt)
 {
     int i;

     for (i = 0; i < FF_ARRAY_ELEMS(supported_formats); i++)
         if (supported_formats[i] == fmt)
             return 1;
     return 0;
 }

 static int config_props(AVFilterLink *inlink)
 {
     AVFilterContext *ctx = inlink->dst;
     ThumbnailCudaContext *s = ctx->priv;
     AVHWFramesContext     *hw_frames_ctx = (AVHWFramesContext*)inlink->hw_frames_ctx->data;
     AVCUDADeviceContext *device_hwctx = hw_frames_ctx->device_ctx->hwctx;
     CUcontext dummy, cuda_ctx = device_hwctx->cuda_ctx;
     CudaFunctions *cu = device_hwctx->internal->cuda_dl;
     int ret;

     extern char vf_thumbnail_cuda_ptx[];

     s->hwctx = device_hwctx;
     s->cu_stream = s->hwctx->stream;

     ret = CHECK_CU(cu->cuCtxPushCurrent(cuda_ctx));
     if (ret < 0)
         return ret;

     ret = CHECK_CU(cu->cuModuleLoadData(&s->cu_module, vf_thumbnail_cuda_ptx));
     if (ret < 0)
         return ret;

     ret = CHECK_CU(cu->cuModuleGetFunction(&s->cu_func_uchar, s->cu_module, "Thumbnail_uchar"));
     if (ret < 0)
         return ret;

     ret = CHECK_CU(cu->cuModuleGetFunction(&s->cu_func_uchar2, s->cu_module, "Thumbnail_uchar2"));
     if (ret < 0)
         return ret;

     ret = CHECK_CU(cu->cuModuleGetFunction(&s->cu_func_ushort, s->cu_module, "Thumbnail_ushort"));
     if (ret < 0)
         return ret;

     ret = CHECK_CU(cu->cuModuleGetFunction(&s->cu_func_ushort2, s->cu_module, "Thumbnail_ushort2"));
     if (ret < 0)
         return ret;

     ret = CHECK_CU(cu->cuMemAlloc(&s->data, HIST_SIZE * sizeof(int)));
     if (ret < 0)
         return ret;

     CHECK_CU(cu->cuCtxPopCurrent(&dummy));

     s->hw_frames_ctx = ctx->inputs[0]->hw_frames_ctx;

     ctx->outputs[0]->hw_frames_ctx = av_buffer_ref(s->hw_frames_ctx);
     if (!ctx->outputs[0]->hw_frames_ctx)
         return AVERROR(ENOMEM);

     s->tb = inlink->time_base;

     if (!format_is_supported(hw_frames_ctx->sw_format)) {
         av_log(ctx, AV_LOG_ERROR, "Unsupported input format: %s\n", av_get_pix_fmt_name(hw_frames_ctx->sw_format));
         return AVERROR(ENOSYS);
     }

     return 0;
 }

 static int query_formats(AVFilterContext *ctx)
 {
     static const enum AVPixelFormat pix_fmts[] = {
         AV_PIX_FMT_CUDA,
         AV_PIX_FMT_NONE
     };
     AVFilterFormats *fmts_list = ff_make_format_list(pix_fmts);
     if (!fmts_list)
         return AVERROR(ENOMEM);
     return ff_set_common_formats(ctx, fmts_list);
 }

 static const AVFilterPad thumbnail_cuda_inputs[] = {
     {
         .name         = "default",
         .type         = AVMEDIA_TYPE_VIDEO,
         .config_props = config_props,
         .filter_frame = filter_frame,
     },
     { NULL }
 };

 static const AVFilterPad thumbnail_cuda_outputs[] = {
     {
         .name          = "default",
         .type          = AVMEDIA_TYPE_VIDEO,
         .request_frame = request_frame,
     },
     { NULL }
 };

 AVFilter ff_vf_thumbnail_cuda = {
     .name          = "thumbnail_cuda",
     .description   = NULL_IF_CONFIG_SMALL("Select the most representative frame in a given sequence of consecutive frames."),
     .priv_size     = sizeof(ThumbnailCudaContext),
     .init          = init,
     .uninit        = uninit,
     .query_formats = query_formats,
     .inputs        = thumbnail_cuda_inputs,
     .outputs       = thumbnail_cuda_outputs,
     .priv_class    = &thumbnail_cuda_class,
     .flags_internal = FF_FILTER_FLAG_HWFRAME_AWARE,
 };
	/*
	* Copyright (c) 2017, NVIDIA CORPORATION. All rights reserved.
	*
	* Permission is hereby granted, free of charge, to any person obtaining a
	* copy of this software and associated documentation files (the "Software"),
	* to deal in the Software without restriction, including without limitation
	* the rights to use, copy, modify, merge, publish, distribute, sublicense,
	* and/or sell copies of the Software, and to permit persons to whom the
	* Software is furnished to do so, subject to the following conditions:
	*
	* The above copyright notice and this permission notice shall be included in
	* all copies or substantial portions of the Software.
	*
	* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
	* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
	* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
	* THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
	* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
	* FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
	* DEALINGS IN THE SOFTWARE.
	*/

	#include "libavutil/hwcontext.h"
	#include "libavutil/hwcontext_cuda_internal.h"
	#include "libavutil/cuda_check.h"
	#include "libavutil/opt.h"
	#include "libavutil/pixdesc.h"

	#include "avfilter.h"
	#include "internal.h"

	#define CHECK_CU(x) FF_CUDA_CHECK_DL(ctx, s->hwctx->internal->cuda_dl, x)

	#define HIST_SIZE (3*256)
	#define DIV_UP(a, b) ( ((a) + (b) - 1) / (b) )
	#define BLOCKX 32
	#define BLOCKY 16

	static const enum AVPixelFormat supported_formats[] = {
	AV_PIX_FMT_NV12,
	AV_PIX_FMT_YUV420P,
	AV_PIX_FMT_YUV444P,
	AV_PIX_FMT_P010,
	AV_PIX_FMT_P016,
	AV_PIX_FMT_YUV444P16,
	};

	struct thumb_frame {
	AVFrame *buf; ///< cached frame
	int histogram[HIST_SIZE]; ///< RGB color distribution histogram of the frame
	};

	typedef struct ThumbnailCudaContext {
	const AVClass *class;
	int n; ///< current frame
	int n_frames; ///< number of frames for analysis
	struct thumb_frame *frames; ///< the n_frames frames
	AVRational tb; ///< copy of the input timebase to ease access

	AVBufferRef *hw_frames_ctx;
	AVCUDADeviceContext *hwctx;

	CUmodule cu_module;

	CUfunction cu_func_uchar;
	CUfunction cu_func_uchar2;
	CUfunction cu_func_ushort;
	CUfunction cu_func_ushort2;
	CUstream cu_stream;

	CUdeviceptr data;

	} ThumbnailCudaContext;

	#define OFFSET(x) offsetof(ThumbnailCudaContext, x)
	#define FLAGS AV_OPT_FLAG_VIDEO_PARAM\|AV_OPT_FLAG_FILTERING_PARAM

	static const AVOption thumbnail_cuda_options[] = {
	{ "n", "set the frames batch size", OFFSET(n_frames), AV_OPT_TYPE_INT, {.i64=100}, 2, INT_MAX, FLAGS },
	{ NULL }
	};

	AVFILTER_DEFINE_CLASS(thumbnail_cuda);

	static av_cold int init(AVFilterContext *ctx)
	{
	ThumbnailCudaContext *s = ctx->priv;

	s->frames = av_calloc(s->n_frames, sizeof(*s->frames));
	if (!s->frames) {
	av_log(ctx, AV_LOG_ERROR,
	"Allocation failure, try to lower the number of frames\n");
	return AVERROR(ENOMEM);
	}
	av_log(ctx, AV_LOG_VERBOSE, "batch size: %d frames\n", s->n_frames);
	return 0;
	}

	/**
	* @brief Compute Sum-square deviation to estimate "closeness".
	* @param hist color distribution histogram
	* @param median average color distribution histogram
	* @return sum of squared errors
	*/
	static double frame_sum_square_err(const int hist, const double median)
	{
	int i;
	double err, sum_sq_err = 0;

	for (i = 0; i < HIST_SIZE; i++) {
	err = median[i] - (double)hist[i];
	sum_sq_err += err*err;
	}
	return sum_sq_err;
	}

	static AVFrame get_best_frame(AVFilterContext ctx)
	{
	AVFrame *picref;
	ThumbnailCudaContext *s = ctx->priv;
	int i, j, best_frame_idx = 0;
	int nb_frames = s->n;
	double avg_hist[HIST_SIZE] = {0}, sq_err, min_sq_err = -1;

	// average histogram of the N frames
	for (j = 0; j < FF_ARRAY_ELEMS(avg_hist); j++) {
	for (i = 0; i < nb_frames; i++)
	avg_hist[j] += (double)s->frames[i].histogram[j];
	avg_hist[j] /= nb_frames;
	}

	// find the frame closer to the average using the sum of squared errors
	for (i = 0; i < nb_frames; i++) {
	sq_err = frame_sum_square_err(s->frames[i].histogram, avg_hist);
	if (i == 0 \|\| sq_err < min_sq_err)
	best_frame_idx = i, min_sq_err = sq_err;
	}

	// free and reset everything (except the best frame buffer)
	for (i = 0; i < nb_frames; i++) {
	memset(s->frames[i].histogram, 0, sizeof(s->frames[i].histogram));
	if (i != best_frame_idx)
	av_frame_free(&s->frames[i].buf);
	}
	s->n = 0;

	// raise the chosen one
	picref = s->frames[best_frame_idx].buf;
	av_log(ctx, AV_LOG_INFO, "frame id #%d (pts_time=%f) selected "
	"from a set of %d images\n", best_frame_idx,
	picref->pts * av_q2d(s->tb), nb_frames);
	s->frames[best_frame_idx].buf = NULL;

	return picref;
	}

	static int thumbnail_kernel(AVFilterContext *ctx, CUfunction func, int channels,
	int histogram, uint8_t src_dptr, int src_width, int src_height, int src_pitch, int pixel_size)
	{
	int ret;
	ThumbnailCudaContext *s = ctx->priv;
	CudaFunctions *cu = s->hwctx->internal->cuda_dl;
	CUtexObject tex = 0;
	void *args[] = { &tex, &histogram, &src_width, &src_height };

	CUDA_TEXTURE_DESC tex_desc = {
	.filterMode = CU_TR_FILTER_MODE_LINEAR,
	.flags = CU_TRSF_READ_AS_INTEGER,
	};

	CUDA_RESOURCE_DESC res_desc = {
	.resType = CU_RESOURCE_TYPE_PITCH2D,
	.res.pitch2D.format = pixel_size == 1 ?
	CU_AD_FORMAT_UNSIGNED_INT8 :
	CU_AD_FORMAT_UNSIGNED_INT16,
	.res.pitch2D.numChannels = channels,
	.res.pitch2D.width = src_width,
	.res.pitch2D.height = src_height,
	.res.pitch2D.pitchInBytes = src_pitch,
	.res.pitch2D.devPtr = (CUdeviceptr)src_dptr,
	};

	ret = CHECK_CU(cu->cuTexObjectCreate(&tex, &res_desc, &tex_desc, NULL));
	if (ret < 0)
	goto exit;

	ret = CHECK_CU(cu->cuLaunchKernel(func,
	DIV_UP(src_width, BLOCKX), DIV_UP(src_height, BLOCKY), 1,
	BLOCKX, BLOCKY, 1, 0, s->cu_stream, args, NULL));
	exit:
	if (tex)
	CHECK_CU(cu->cuTexObjectDestroy(tex));

	return ret;
	}

	static int thumbnail(AVFilterContext ctx, int histogram, AVFrame *in)
	{
	AVHWFramesContext in_frames_ctx = (AVHWFramesContext)in->hw_frames_ctx->data;
	ThumbnailCudaContext *s = ctx->priv;

	switch (in_frames_ctx->sw_format) {
	case AV_PIX_FMT_NV12:
	thumbnail_kernel(ctx, s->cu_func_uchar, 1,
	histogram, in->data[0], in->width, in->height, in->linesize[0], 1);
	thumbnail_kernel(ctx, s->cu_func_uchar2, 2,
	histogram + 256, in->data[1], in->width / 2, in->height / 2, in->linesize[1], 1);
	break;
	case AV_PIX_FMT_YUV420P:
	thumbnail_kernel(ctx, s->cu_func_uchar, 1,
	histogram, in->data[0], in->width, in->height, in->linesize[0], 1);
	thumbnail_kernel(ctx, s->cu_func_uchar, 1,
	histogram + 256, in->data[1], in->width / 2, in->height / 2, in->linesize[1], 1);
	thumbnail_kernel(ctx, s->cu_func_uchar, 1,
	histogram + 512, in->data[2], in->width / 2, in->height / 2, in->linesize[2], 1);
	break;
	case AV_PIX_FMT_YUV444P:
	thumbnail_kernel(ctx, s->cu_func_uchar, 1,
	histogram, in->data[0], in->width, in->height, in->linesize[0], 1);
	thumbnail_kernel(ctx, s->cu_func_uchar, 1,
	histogram + 256, in->data[1], in->width, in->height, in->linesize[1], 1);
	thumbnail_kernel(ctx, s->cu_func_uchar, 1,
	histogram + 512, in->data[2], in->width, in->height, in->linesize[2], 1);
	break;
	case AV_PIX_FMT_P010LE:
	case AV_PIX_FMT_P016LE:
	thumbnail_kernel(ctx, s->cu_func_ushort, 1,
	histogram, in->data[0], in->width, in->height, in->linesize[0], 2);
	thumbnail_kernel(ctx, s->cu_func_ushort2, 2,
	histogram + 256, in->data[1], in->width / 2, in->height / 2, in->linesize[1], 2);
	break;
	case AV_PIX_FMT_YUV444P16:
	thumbnail_kernel(ctx, s->cu_func_ushort2, 1,
	histogram, in->data[0], in->width, in->height, in->linesize[0], 2);
	thumbnail_kernel(ctx, s->cu_func_ushort2, 1,
	histogram + 256, in->data[1], in->width, in->height, in->linesize[1], 2);
	thumbnail_kernel(ctx, s->cu_func_ushort2, 1,
	histogram + 512, in->data[2], in->width, in->height, in->linesize[2], 2);
	break;
	default:
	return AVERROR_BUG;
	}

	return 0;
	}

	static int filter_frame(AVFilterLink inlink, AVFrame frame)
	{
	AVFilterContext *ctx = inlink->dst;
	ThumbnailCudaContext *s = ctx->priv;
	CudaFunctions *cu = s->hwctx->internal->cuda_dl;
	AVFilterLink *outlink = ctx->outputs[0];
	int *hist = s->frames[s->n].histogram;
	AVHWFramesContext hw_frames_ctx = (AVHWFramesContext)s->hw_frames_ctx->data;
	CUcontext dummy;
	CUDA_MEMCPY2D cpy = { 0 };
	int ret = 0;

	// keep a reference of each frame
	s->frames[s->n].buf = frame;

	ret = CHECK_CU(cu->cuCtxPushCurrent(s->hwctx->cuda_ctx));
	if (ret < 0)
	return ret;

	CHECK_CU(cu->cuMemsetD8Async(s->data, 0, HIST_SIZE * sizeof(int), s->cu_stream));

	thumbnail(ctx, (int*)s->data, frame);

	cpy.srcMemoryType = CU_MEMORYTYPE_DEVICE;
	cpy.dstMemoryType = CU_MEMORYTYPE_HOST;
	cpy.srcDevice = s->data;
	cpy.dstHost = hist;
	cpy.srcPitch = HIST_SIZE * sizeof(int);
	cpy.dstPitch = HIST_SIZE * sizeof(int);
	cpy.WidthInBytes = HIST_SIZE * sizeof(int);
	cpy.Height = 1;

	ret = CHECK_CU(cu->cuMemcpy2DAsync(&cpy, s->cu_stream));
	if (ret < 0)
	return ret;

	if (hw_frames_ctx->sw_format == AV_PIX_FMT_NV12 \|\| hw_frames_ctx->sw_format == AV_PIX_FMT_YUV420P \|\|
	hw_frames_ctx->sw_format == AV_PIX_FMT_P010LE \|\| hw_frames_ctx->sw_format == AV_PIX_FMT_P016LE)
	{
	int i;
	for (i = 256; i < HIST_SIZE; i++)
	hist[i] = 4 * hist[i];
	}

	CHECK_CU(cu->cuCtxPopCurrent(&dummy));
	if (ret < 0)
	return ret;

	// no selection until the buffer of N frames is filled up
	s->n++;
	if (s->n < s->n_frames)
	return 0;

	return ff_filter_frame(outlink, get_best_frame(ctx));
	}

	static av_cold void uninit(AVFilterContext *ctx)
	{
	int i;
	ThumbnailCudaContext *s = ctx->priv;
	CudaFunctions *cu = s->hwctx->internal->cuda_dl;

	if (s->data) {
	CHECK_CU(cu->cuMemFree(s->data));
	s->data = 0;
	}

	if (s->cu_module) {
	CHECK_CU(cu->cuModuleUnload(s->cu_module));
	s->cu_module = NULL;
	}

	for (i = 0; i < s->n_frames && s->frames[i].buf; i++)
	av_frame_free(&s->frames[i].buf);
	av_freep(&s->frames);
	}

	static int request_frame(AVFilterLink *link)
	{
	AVFilterContext *ctx = link->src;
	ThumbnailCudaContext *s = ctx->priv;
	int ret = ff_request_frame(ctx->inputs[0]);

	if (ret == AVERROR_EOF && s->n) {
	ret = ff_filter_frame(link, get_best_frame(ctx));
	if (ret < 0)
	return ret;
	ret = AVERROR_EOF;
	}
	if (ret < 0)
	return ret;
	return 0;
	}

	static int format_is_supported(enum AVPixelFormat fmt)
	{
	int i;

	for (i = 0; i < FF_ARRAY_ELEMS(supported_formats); i++)
	if (supported_formats[i] == fmt)
	return 1;
	return 0;
	}

	static int config_props(AVFilterLink *inlink)
	{
	AVFilterContext *ctx = inlink->dst;
	ThumbnailCudaContext *s = ctx->priv;
	AVHWFramesContext hw_frames_ctx = (AVHWFramesContext)inlink->hw_frames_ctx->data;
	AVCUDADeviceContext *device_hwctx = hw_frames_ctx->device_ctx->hwctx;
	CUcontext dummy, cuda_ctx = device_hwctx->cuda_ctx;
	CudaFunctions *cu = device_hwctx->internal->cuda_dl;
	int ret;

	extern char vf_thumbnail_cuda_ptx[];

	s->hwctx = device_hwctx;
	s->cu_stream = s->hwctx->stream;

	ret = CHECK_CU(cu->cuCtxPushCurrent(cuda_ctx));
	if (ret < 0)
	return ret;

	ret = CHECK_CU(cu->cuModuleLoadData(&s->cu_module, vf_thumbnail_cuda_ptx));
	if (ret < 0)
	return ret;

	ret = CHECK_CU(cu->cuModuleGetFunction(&s->cu_func_uchar, s->cu_module, "Thumbnail_uchar"));
	if (ret < 0)
	return ret;

	ret = CHECK_CU(cu->cuModuleGetFunction(&s->cu_func_uchar2, s->cu_module, "Thumbnail_uchar2"));
	if (ret < 0)
	return ret;

	ret = CHECK_CU(cu->cuModuleGetFunction(&s->cu_func_ushort, s->cu_module, "Thumbnail_ushort"));
	if (ret < 0)
	return ret;

	ret = CHECK_CU(cu->cuModuleGetFunction(&s->cu_func_ushort2, s->cu_module, "Thumbnail_ushort2"));
	if (ret < 0)
	return ret;

	ret = CHECK_CU(cu->cuMemAlloc(&s->data, HIST_SIZE * sizeof(int)));
	if (ret < 0)
	return ret;

	CHECK_CU(cu->cuCtxPopCurrent(&dummy));

	s->hw_frames_ctx = ctx->inputs[0]->hw_frames_ctx;

	ctx->outputs[0]->hw_frames_ctx = av_buffer_ref(s->hw_frames_ctx);
	if (!ctx->outputs[0]->hw_frames_ctx)
	return AVERROR(ENOMEM);

	s->tb = inlink->time_base;

	if (!format_is_supported(hw_frames_ctx->sw_format)) {
	av_log(ctx, AV_LOG_ERROR, "Unsupported input format: %s\n", av_get_pix_fmt_name(hw_frames_ctx->sw_format));
	return AVERROR(ENOSYS);
	}

	return 0;
	}

	static int query_formats(AVFilterContext *ctx)
	{
	static const enum AVPixelFormat pix_fmts[] = {
	AV_PIX_FMT_CUDA,
	AV_PIX_FMT_NONE
	};
	AVFilterFormats *fmts_list = ff_make_format_list(pix_fmts);
	if (!fmts_list)
	return AVERROR(ENOMEM);
	return ff_set_common_formats(ctx, fmts_list);
	}

	static const AVFilterPad thumbnail_cuda_inputs[] = {
	{
	.name = "default",
	.type = AVMEDIA_TYPE_VIDEO,
	.config_props = config_props,
	.filter_frame = filter_frame,
	},
	{ NULL }
	};

	static const AVFilterPad thumbnail_cuda_outputs[] = {
	{
	.name = "default",
	.type = AVMEDIA_TYPE_VIDEO,
	.request_frame = request_frame,
	},
	{ NULL }
	};

	AVFilter ff_vf_thumbnail_cuda = {
	.name = "thumbnail_cuda",
	.description = NULL_IF_CONFIG_SMALL("Select the most representative frame in a given sequence of consecutive frames."),
	.priv_size = sizeof(ThumbnailCudaContext),
	.init = init,
	.uninit = uninit,
	.query_formats = query_formats,
	.inputs = thumbnail_cuda_inputs,
	.outputs = thumbnail_cuda_outputs,
	.priv_class = &thumbnail_cuda_class,
	.flags_internal = FF_FILTER_FLAG_HWFRAME_AWARE,
	};