2 * HW decode acceleration through NVDEC
4 * Copyright (c) 2016 Anton Khirnov
6 * This file is part of FFmpeg.
8 * FFmpeg is free software; you can redistribute it and/or
9 * modify it under the terms of the GNU Lesser General Public
10 * License as published by the Free Software Foundation; either
11 * version 2.1 of the License, or (at your option) any later version.
13 * FFmpeg is distributed in the hope that it will be useful,
14 * but WITHOUT ANY WARRANTY; without even the implied warranty of
15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
16 * Lesser General Public License for more details.
18 * You should have received a copy of the GNU Lesser General Public
19 * License along with FFmpeg; if not, write to the Free Software
20 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
25 #include "libavutil/common.h"
26 #include "libavutil/error.h"
27 #include "libavutil/hwcontext.h"
28 #include "libavutil/hwcontext_cuda_internal.h"
29 #include "libavutil/pixdesc.h"
30 #include "libavutil/pixfmt.h"
37 typedef struct NVDECDecoder {
38 CUvideodecoder decoder;
40 AVBufferRef *hw_device_ref;
48 typedef struct NVDECFramePool {
49 unsigned int dpb_size;
50 unsigned int nb_allocated;
53 static int map_avcodec_id(enum AVCodecID id)
56 case AV_CODEC_ID_H264: return cudaVideoCodec_H264;
57 case AV_CODEC_ID_HEVC: return cudaVideoCodec_HEVC;
58 case AV_CODEC_ID_MJPEG: return cudaVideoCodec_JPEG;
59 case AV_CODEC_ID_MPEG1VIDEO: return cudaVideoCodec_MPEG1;
60 case AV_CODEC_ID_MPEG2VIDEO: return cudaVideoCodec_MPEG2;
61 case AV_CODEC_ID_MPEG4: return cudaVideoCodec_MPEG4;
62 case AV_CODEC_ID_VC1: return cudaVideoCodec_VC1;
63 case AV_CODEC_ID_VP8: return cudaVideoCodec_VP8;
64 case AV_CODEC_ID_VP9: return cudaVideoCodec_VP9;
65 case AV_CODEC_ID_WMV3: return cudaVideoCodec_VC1;
70 static int map_chroma_format(enum AVPixelFormat pix_fmt)
72 int shift_h = 0, shift_v = 0;
74 av_pix_fmt_get_chroma_sub_sample(pix_fmt, &shift_h, &shift_v);
76 if (shift_h == 1 && shift_v == 1)
77 return cudaVideoChromaFormat_420;
78 else if (shift_h == 1 && shift_v == 0)
79 return cudaVideoChromaFormat_422;
80 else if (shift_h == 0 && shift_v == 0)
81 return cudaVideoChromaFormat_444;
86 static int nvdec_test_capabilities(NVDECDecoder *decoder,
87 CUVIDDECODECREATEINFO *params, void *logctx)
90 CUVIDDECODECAPS caps = { 0 };
92 caps.eCodecType = params->CodecType;
93 caps.eChromaFormat = params->ChromaFormat;
94 caps.nBitDepthMinus8 = params->bitDepthMinus8;
96 if (!decoder->cvdl->cuvidGetDecoderCaps) {
97 av_log(logctx, AV_LOG_WARNING, "Used Nvidia driver is too old to perform a capability check.\n");
98 av_log(logctx, AV_LOG_WARNING, "The minimum required version is "
99 #if defined(_WIN32) || defined(__CYGWIN__)
104 ". Continuing blind.\n");
108 err = decoder->cvdl->cuvidGetDecoderCaps(&caps);
109 if (err != CUDA_SUCCESS) {
110 av_log(logctx, AV_LOG_ERROR, "Failed querying decoder capabilities\n");
111 return AVERROR_UNKNOWN;
114 av_log(logctx, AV_LOG_VERBOSE, "NVDEC capabilities:\n");
115 av_log(logctx, AV_LOG_VERBOSE, "format supported: %s, max_mb_count: %d\n",
116 caps.bIsSupported ? "yes" : "no", caps.nMaxMBCount);
117 av_log(logctx, AV_LOG_VERBOSE, "min_width: %d, max_width: %d\n",
118 caps.nMinWidth, caps.nMaxWidth);
119 av_log(logctx, AV_LOG_VERBOSE, "min_height: %d, max_height: %d\n",
120 caps.nMinHeight, caps.nMaxHeight);
122 if (!caps.bIsSupported) {
123 av_log(logctx, AV_LOG_ERROR, "Hardware is lacking required capabilities\n");
124 return AVERROR(EINVAL);
127 if (params->ulWidth > caps.nMaxWidth || params->ulWidth < caps.nMinWidth) {
128 av_log(logctx, AV_LOG_ERROR, "Video width %d not within range from %d to %d\n",
129 (int)params->ulWidth, caps.nMinWidth, caps.nMaxWidth);
130 return AVERROR(EINVAL);
133 if (params->ulHeight > caps.nMaxHeight || params->ulHeight < caps.nMinHeight) {
134 av_log(logctx, AV_LOG_ERROR, "Video height %d not within range from %d to %d\n",
135 (int)params->ulHeight, caps.nMinHeight, caps.nMaxHeight);
136 return AVERROR(EINVAL);
139 if ((params->ulWidth * params->ulHeight) / 256 > caps.nMaxMBCount) {
140 av_log(logctx, AV_LOG_ERROR, "Video macroblock count %d exceeds maximum of %d\n",
141 (int)(params->ulWidth * params->ulHeight) / 256, caps.nMaxMBCount);
142 return AVERROR(EINVAL);
148 static void nvdec_decoder_free(void *opaque, uint8_t *data)
150 NVDECDecoder *decoder = (NVDECDecoder*)data;
152 if (decoder->decoder)
153 decoder->cvdl->cuvidDestroyDecoder(decoder->decoder);
155 av_buffer_unref(&decoder->hw_device_ref);
157 cuvid_free_functions(&decoder->cvdl);
162 static int nvdec_decoder_create(AVBufferRef **out, AVBufferRef *hw_device_ref,
163 CUVIDDECODECREATEINFO *params, void *logctx)
165 AVHWDeviceContext *hw_device_ctx = (AVHWDeviceContext*)hw_device_ref->data;
166 AVCUDADeviceContext *device_hwctx = hw_device_ctx->hwctx;
168 AVBufferRef *decoder_ref;
169 NVDECDecoder *decoder;
175 decoder = av_mallocz(sizeof(*decoder));
177 return AVERROR(ENOMEM);
179 decoder_ref = av_buffer_create((uint8_t*)decoder, sizeof(*decoder),
180 nvdec_decoder_free, NULL, AV_BUFFER_FLAG_READONLY);
183 return AVERROR(ENOMEM);
186 decoder->hw_device_ref = av_buffer_ref(hw_device_ref);
187 if (!decoder->hw_device_ref) {
188 ret = AVERROR(ENOMEM);
191 decoder->cuda_ctx = device_hwctx->cuda_ctx;
192 decoder->cudl = device_hwctx->internal->cuda_dl;
193 decoder->stream = device_hwctx->stream;
195 ret = cuvid_load_functions(&decoder->cvdl, logctx);
197 av_log(logctx, AV_LOG_ERROR, "Failed loading nvcuvid.\n");
201 err = decoder->cudl->cuCtxPushCurrent(decoder->cuda_ctx);
202 if (err != CUDA_SUCCESS) {
203 ret = AVERROR_UNKNOWN;
207 ret = nvdec_test_capabilities(decoder, params, logctx);
209 decoder->cudl->cuCtxPopCurrent(&dummy);
213 err = decoder->cvdl->cuvidCreateDecoder(&decoder->decoder, params);
215 decoder->cudl->cuCtxPopCurrent(&dummy);
217 if (err != CUDA_SUCCESS) {
218 av_log(logctx, AV_LOG_ERROR, "Error creating a NVDEC decoder: %d\n", err);
219 ret = AVERROR_UNKNOWN;
227 av_buffer_unref(&decoder_ref);
231 static AVBufferRef *nvdec_decoder_frame_alloc(void *opaque, int size)
233 NVDECFramePool *pool = opaque;
236 if (pool->nb_allocated >= pool->dpb_size)
239 ret = av_buffer_alloc(sizeof(unsigned int));
243 *(unsigned int*)ret->data = pool->nb_allocated++;
248 int ff_nvdec_decode_uninit(AVCodecContext *avctx)
250 NVDECContext *ctx = avctx->internal->hwaccel_priv_data;
252 av_freep(&ctx->bitstream);
253 ctx->bitstream_len = 0;
254 ctx->bitstream_allocated = 0;
256 av_freep(&ctx->slice_offsets);
258 ctx->slice_offsets_allocated = 0;
260 av_buffer_unref(&ctx->decoder_ref);
261 av_buffer_pool_uninit(&ctx->decoder_pool);
266 int ff_nvdec_decode_init(AVCodecContext *avctx)
268 NVDECContext *ctx = avctx->internal->hwaccel_priv_data;
270 NVDECFramePool *pool;
271 AVHWFramesContext *frames_ctx;
272 const AVPixFmtDescriptor *sw_desc;
274 CUVIDDECODECREATEINFO params = { 0 };
276 int cuvid_codec_type, cuvid_chroma_format;
279 sw_desc = av_pix_fmt_desc_get(avctx->sw_pix_fmt);
283 cuvid_codec_type = map_avcodec_id(avctx->codec_id);
284 if (cuvid_codec_type < 0) {
285 av_log(avctx, AV_LOG_ERROR, "Unsupported codec ID\n");
289 cuvid_chroma_format = map_chroma_format(avctx->sw_pix_fmt);
290 if (cuvid_chroma_format < 0) {
291 av_log(avctx, AV_LOG_ERROR, "Unsupported chroma format\n");
292 return AVERROR(ENOSYS);
295 if (!avctx->hw_frames_ctx) {
296 ret = ff_decode_get_hw_frames_ctx(avctx, AV_HWDEVICE_TYPE_CUDA);
301 frames_ctx = (AVHWFramesContext*)avctx->hw_frames_ctx->data;
303 params.ulWidth = avctx->coded_width;
304 params.ulHeight = avctx->coded_height;
305 params.ulTargetWidth = avctx->coded_width;
306 params.ulTargetHeight = avctx->coded_height;
307 params.bitDepthMinus8 = sw_desc->comp[0].depth - 8;
308 params.OutputFormat = params.bitDepthMinus8 ?
309 cudaVideoSurfaceFormat_P016 : cudaVideoSurfaceFormat_NV12;
310 params.CodecType = cuvid_codec_type;
311 params.ChromaFormat = cuvid_chroma_format;
312 params.ulNumDecodeSurfaces = frames_ctx->initial_pool_size;
313 params.ulNumOutputSurfaces = frames_ctx->initial_pool_size;
315 ret = nvdec_decoder_create(&ctx->decoder_ref, frames_ctx->device_ref, ¶ms, avctx);
317 if (params.ulNumDecodeSurfaces > 32) {
318 av_log(avctx, AV_LOG_WARNING, "Using more than 32 (%d) decode surfaces might cause nvdec to fail.\n",
319 (int)params.ulNumDecodeSurfaces);
320 av_log(avctx, AV_LOG_WARNING, "Try lowering the amount of threads. Using %d right now.\n",
321 avctx->thread_count);
326 pool = av_mallocz(sizeof(*pool));
328 ret = AVERROR(ENOMEM);
331 pool->dpb_size = frames_ctx->initial_pool_size;
333 ctx->decoder_pool = av_buffer_pool_init2(sizeof(int), pool,
334 nvdec_decoder_frame_alloc, av_free);
335 if (!ctx->decoder_pool) {
336 ret = AVERROR(ENOMEM);
342 ff_nvdec_decode_uninit(avctx);
346 static void nvdec_fdd_priv_free(void *priv)
348 NVDECFrame *cf = priv;
353 av_buffer_unref(&cf->idx_ref);
354 av_buffer_unref(&cf->decoder_ref);
359 static void nvdec_unmap_mapped_frame(void *opaque, uint8_t *data)
361 NVDECFrame *unmap_data = (NVDECFrame*)data;
362 NVDECDecoder *decoder = (NVDECDecoder*)unmap_data->decoder_ref->data;
363 CUdeviceptr devptr = (CUdeviceptr)opaque;
367 err = decoder->cudl->cuCtxPushCurrent(decoder->cuda_ctx);
368 if (err != CUDA_SUCCESS) {
369 av_log(NULL, AV_LOG_ERROR, "cuCtxPushCurrent failed\n");
373 err = decoder->cvdl->cuvidUnmapVideoFrame(decoder->decoder, devptr);
374 if (err != CUDA_SUCCESS)
375 av_log(NULL, AV_LOG_ERROR, "cuvidUnmapVideoFrame failed\n");
377 decoder->cudl->cuCtxPopCurrent(&dummy);
380 av_buffer_unref(&unmap_data->idx_ref);
381 av_buffer_unref(&unmap_data->decoder_ref);
385 static int nvdec_retrieve_data(void *logctx, AVFrame *frame)
387 FrameDecodeData *fdd = (FrameDecodeData*)frame->private_ref->data;
388 NVDECFrame *cf = (NVDECFrame*)fdd->hwaccel_priv;
389 NVDECDecoder *decoder = (NVDECDecoder*)cf->decoder_ref->data;
391 CUVIDPROCPARAMS vpp = { 0 };
392 NVDECFrame *unmap_data = NULL;
398 unsigned int pitch, i;
399 unsigned int offset = 0;
402 vpp.progressive_frame = 1;
403 vpp.output_stream = decoder->stream;
405 err = decoder->cudl->cuCtxPushCurrent(decoder->cuda_ctx);
406 if (err != CUDA_SUCCESS)
407 return AVERROR_UNKNOWN;
409 err = decoder->cvdl->cuvidMapVideoFrame(decoder->decoder, cf->idx, &devptr,
411 if (err != CUDA_SUCCESS) {
412 av_log(logctx, AV_LOG_ERROR, "Error mapping a picture with CUVID: %d\n",
414 ret = AVERROR_UNKNOWN;
418 unmap_data = av_mallocz(sizeof(*unmap_data));
420 ret = AVERROR(ENOMEM);
424 frame->buf[1] = av_buffer_create((uint8_t *)unmap_data, sizeof(*unmap_data),
425 nvdec_unmap_mapped_frame, (void*)devptr,
426 AV_BUFFER_FLAG_READONLY);
427 if (!frame->buf[1]) {
428 ret = AVERROR(ENOMEM);
432 unmap_data->idx = cf->idx;
433 unmap_data->idx_ref = av_buffer_ref(cf->idx_ref);
434 unmap_data->decoder_ref = av_buffer_ref(cf->decoder_ref);
436 for (i = 0; frame->linesize[i]; i++) {
437 frame->data[i] = (uint8_t*)(devptr + offset);
438 frame->linesize[i] = pitch;
439 offset += pitch * (frame->height >> (i ? 1 : 0));
445 if (!frame->buf[1]) {
446 decoder->cvdl->cuvidUnmapVideoFrame(decoder->decoder, devptr);
447 av_freep(&unmap_data);
449 av_buffer_unref(&frame->buf[1]);
453 decoder->cudl->cuCtxPopCurrent(&dummy);
457 int ff_nvdec_start_frame(AVCodecContext *avctx, AVFrame *frame)
459 NVDECContext *ctx = avctx->internal->hwaccel_priv_data;
460 FrameDecodeData *fdd = (FrameDecodeData*)frame->private_ref->data;
461 NVDECFrame *cf = NULL;
464 ctx->bitstream_len = 0;
467 if (fdd->hwaccel_priv)
470 cf = av_mallocz(sizeof(*cf));
472 return AVERROR(ENOMEM);
474 cf->decoder_ref = av_buffer_ref(ctx->decoder_ref);
475 if (!cf->decoder_ref) {
476 ret = AVERROR(ENOMEM);
480 cf->idx_ref = av_buffer_pool_get(ctx->decoder_pool);
482 av_log(avctx, AV_LOG_ERROR, "No decoder surfaces left\n");
483 ret = AVERROR(ENOMEM);
486 cf->idx = *(unsigned int*)cf->idx_ref->data;
488 fdd->hwaccel_priv = cf;
489 fdd->hwaccel_priv_free = nvdec_fdd_priv_free;
490 fdd->post_process = nvdec_retrieve_data;
494 nvdec_fdd_priv_free(cf);
499 int ff_nvdec_end_frame(AVCodecContext *avctx)
501 NVDECContext *ctx = avctx->internal->hwaccel_priv_data;
502 NVDECDecoder *decoder = (NVDECDecoder*)ctx->decoder_ref->data;
503 CUVIDPICPARAMS *pp = &ctx->pic_params;
510 pp->nBitstreamDataLen = ctx->bitstream_len;
511 pp->pBitstreamData = ctx->bitstream;
512 pp->nNumSlices = ctx->nb_slices;
513 pp->pSliceDataOffsets = ctx->slice_offsets;
515 err = decoder->cudl->cuCtxPushCurrent(decoder->cuda_ctx);
516 if (err != CUDA_SUCCESS)
517 return AVERROR_UNKNOWN;
519 err = decoder->cvdl->cuvidDecodePicture(decoder->decoder, &ctx->pic_params);
520 if (err != CUDA_SUCCESS) {
521 av_log(avctx, AV_LOG_ERROR, "Error decoding a picture with NVDEC: %d\n",
523 ret = AVERROR_UNKNOWN;
528 decoder->cudl->cuCtxPopCurrent(&dummy);
533 int ff_nvdec_simple_end_frame(AVCodecContext *avctx)
535 NVDECContext *ctx = avctx->internal->hwaccel_priv_data;
536 int ret = ff_nvdec_end_frame(avctx);
537 ctx->bitstream = NULL;
541 int ff_nvdec_simple_decode_slice(AVCodecContext *avctx, const uint8_t *buffer,
544 NVDECContext *ctx = avctx->internal->hwaccel_priv_data;
547 tmp = av_fast_realloc(ctx->slice_offsets, &ctx->slice_offsets_allocated,
548 (ctx->nb_slices + 1) * sizeof(*ctx->slice_offsets));
550 return AVERROR(ENOMEM);
551 ctx->slice_offsets = tmp;
554 ctx->bitstream = (uint8_t*)buffer;
556 ctx->slice_offsets[ctx->nb_slices] = buffer - ctx->bitstream;
557 ctx->bitstream_len += size;
563 static void nvdec_free_dummy(struct AVHWFramesContext *ctx)
565 av_buffer_pool_uninit(&ctx->pool);
568 static AVBufferRef *nvdec_alloc_dummy(int size)
570 return av_buffer_create(NULL, 0, NULL, NULL, 0);
573 int ff_nvdec_frame_params(AVCodecContext *avctx,
574 AVBufferRef *hw_frames_ctx,
577 AVHWFramesContext *frames_ctx = (AVHWFramesContext*)hw_frames_ctx->data;
578 const AVPixFmtDescriptor *sw_desc;
579 int cuvid_codec_type, cuvid_chroma_format;
581 sw_desc = av_pix_fmt_desc_get(avctx->sw_pix_fmt);
585 cuvid_codec_type = map_avcodec_id(avctx->codec_id);
586 if (cuvid_codec_type < 0) {
587 av_log(avctx, AV_LOG_ERROR, "Unsupported codec ID\n");
591 cuvid_chroma_format = map_chroma_format(avctx->sw_pix_fmt);
592 if (cuvid_chroma_format < 0) {
593 av_log(avctx, AV_LOG_VERBOSE, "Unsupported chroma format\n");
594 return AVERROR(EINVAL);
597 frames_ctx->format = AV_PIX_FMT_CUDA;
598 frames_ctx->width = (avctx->coded_width + 1) & ~1;
599 frames_ctx->height = (avctx->coded_height + 1) & ~1;
600 frames_ctx->initial_pool_size = dpb_size;
602 frames_ctx->free = nvdec_free_dummy;
603 frames_ctx->pool = av_buffer_pool_init(0, nvdec_alloc_dummy);
605 if (!frames_ctx->pool)
606 return AVERROR(ENOMEM);
608 switch (sw_desc->comp[0].depth) {
610 frames_ctx->sw_format = AV_PIX_FMT_NV12;
613 frames_ctx->sw_format = AV_PIX_FMT_P010;
616 frames_ctx->sw_format = AV_PIX_FMT_P016;
619 return AVERROR(EINVAL);
625 int ff_nvdec_get_ref_idx(AVFrame *frame)
627 FrameDecodeData *fdd;
630 if (!frame || !frame->private_ref)
633 fdd = (FrameDecodeData*)frame->private_ref->data;
634 cf = (NVDECFrame*)fdd->hwaccel_priv;