3 * Copyright (c) 2016 Timo Rothenpieler <timo@rothenpieler.org>
5 * This file is part of FFmpeg.
7 * FFmpeg is free software; you can redistribute it and/or
8 * modify it under the terms of the GNU Lesser General Public
9 * License as published by the Free Software Foundation; either
10 * version 2.1 of the License, or (at your option) any later version.
12 * FFmpeg is distributed in the hope that it will be useful,
13 * but WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * Lesser General Public License for more details.
17 * You should have received a copy of the GNU Lesser General Public
18 * License along with FFmpeg; if not, write to the Free Software
19 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
22 #include "compat/cuda/dynlink_loader.h"
24 #include "libavutil/buffer.h"
25 #include "libavutil/mathematics.h"
26 #include "libavutil/hwcontext.h"
27 #include "libavutil/hwcontext_cuda_internal.h"
28 #include "libavutil/cuda_check.h"
29 #include "libavutil/fifo.h"
30 #include "libavutil/log.h"
31 #include "libavutil/opt.h"
32 #include "libavutil/pixdesc.h"
40 #if !NVDECAPI_CHECK_VERSION(9, 0)
41 #define cudaVideoSurfaceFormat_YUV444 2
42 #define cudaVideoSurfaceFormat_YUV444_16Bit 3
45 #if NVDECAPI_CHECK_VERSION(11, 0)
46 #define CUVID_HAS_AV1_SUPPORT
49 typedef struct CuvidContext
53 CUvideodecoder cudecoder;
54 CUvideoparser cuparser;
58 int drop_second_field;
74 AVBufferRef *hwdevice;
77 AVFifoBuffer *frame_queue;
80 int deint_mode_current;
82 int progressive_sequence;
89 cudaVideoCodec codec_type;
90 cudaVideoChromaFormat chroma_format;
92 CUVIDDECODECAPS caps8, caps10, caps12;
94 CUVIDPARSERPARAMS cuparseinfo;
95 CUVIDEOFORMATEX *cuparse_ext;
101 typedef struct CuvidParsedFrame
103 CUVIDPARSERDISPINFO dispinfo;
105 int is_deinterlacing;
108 #define CHECK_CU(x) FF_CUDA_CHECK_DL(avctx, ctx->cudl, x)
110 static int CUDAAPI cuvid_handle_video_sequence(void *opaque, CUVIDEOFORMAT* format)
112 AVCodecContext *avctx = opaque;
113 CuvidContext *ctx = avctx->priv_data;
114 AVHWFramesContext *hwframe_ctx = (AVHWFramesContext*)ctx->hwframe->data;
115 CUVIDDECODECAPS *caps = NULL;
116 CUVIDDECODECREATEINFO cuinfo;
120 int old_width = avctx->width;
121 int old_height = avctx->height;
123 enum AVPixelFormat pix_fmts[3] = { AV_PIX_FMT_CUDA,
124 AV_PIX_FMT_NONE, // Will be updated below
127 av_log(avctx, AV_LOG_TRACE, "pfnSequenceCallback, progressive_sequence=%d\n", format->progressive_sequence);
129 memset(&cuinfo, 0, sizeof(cuinfo));
131 ctx->internal_error = 0;
133 avctx->coded_width = cuinfo.ulWidth = format->coded_width;
134 avctx->coded_height = cuinfo.ulHeight = format->coded_height;
137 cuinfo.display_area.left = format->display_area.left + ctx->crop.left;
138 cuinfo.display_area.top = format->display_area.top + ctx->crop.top;
139 cuinfo.display_area.right = format->display_area.right - ctx->crop.right;
140 cuinfo.display_area.bottom = format->display_area.bottom - ctx->crop.bottom;
142 // width and height need to be set before calling ff_get_format
143 if (ctx->resize_expr) {
144 avctx->width = ctx->resize.width;
145 avctx->height = ctx->resize.height;
147 avctx->width = cuinfo.display_area.right - cuinfo.display_area.left;
148 avctx->height = cuinfo.display_area.bottom - cuinfo.display_area.top;
151 // target width/height need to be multiples of two
152 cuinfo.ulTargetWidth = avctx->width = (avctx->width + 1) & ~1;
153 cuinfo.ulTargetHeight = avctx->height = (avctx->height + 1) & ~1;
155 // aspect ratio conversion, 1:1, depends on scaled resolution
156 cuinfo.target_rect.left = 0;
157 cuinfo.target_rect.top = 0;
158 cuinfo.target_rect.right = cuinfo.ulTargetWidth;
159 cuinfo.target_rect.bottom = cuinfo.ulTargetHeight;
161 chroma_444 = format->chroma_format == cudaVideoChromaFormat_444;
163 switch (format->bit_depth_luma_minus8) {
165 pix_fmts[1] = chroma_444 ? AV_PIX_FMT_YUV444P : AV_PIX_FMT_NV12;
169 pix_fmts[1] = chroma_444 ? AV_PIX_FMT_YUV444P16 : AV_PIX_FMT_P010;
173 pix_fmts[1] = chroma_444 ? AV_PIX_FMT_YUV444P16 : AV_PIX_FMT_P016;
180 if (!caps || !caps->bIsSupported) {
181 av_log(avctx, AV_LOG_ERROR, "unsupported bit depth: %d\n",
182 format->bit_depth_luma_minus8 + 8);
183 ctx->internal_error = AVERROR(EINVAL);
187 surface_fmt = ff_get_format(avctx, pix_fmts);
188 if (surface_fmt < 0) {
189 av_log(avctx, AV_LOG_ERROR, "ff_get_format failed: %d\n", surface_fmt);
190 ctx->internal_error = AVERROR(EINVAL);
194 av_log(avctx, AV_LOG_VERBOSE, "Formats: Original: %s | HW: %s | SW: %s\n",
195 av_get_pix_fmt_name(avctx->pix_fmt),
196 av_get_pix_fmt_name(surface_fmt),
197 av_get_pix_fmt_name(avctx->sw_pix_fmt));
199 avctx->pix_fmt = surface_fmt;
201 // Update our hwframe ctx, as the get_format callback might have refreshed it!
202 if (avctx->hw_frames_ctx) {
203 av_buffer_unref(&ctx->hwframe);
205 ctx->hwframe = av_buffer_ref(avctx->hw_frames_ctx);
207 ctx->internal_error = AVERROR(ENOMEM);
211 hwframe_ctx = (AVHWFramesContext*)ctx->hwframe->data;
214 ff_set_sar(avctx, av_div_q(
215 (AVRational){ format->display_aspect_ratio.x, format->display_aspect_ratio.y },
216 (AVRational){ avctx->width, avctx->height }));
218 ctx->deint_mode_current = format->progressive_sequence
219 ? cudaVideoDeinterlaceMode_Weave
222 ctx->progressive_sequence = format->progressive_sequence;
224 if (!format->progressive_sequence && ctx->deint_mode_current == cudaVideoDeinterlaceMode_Weave)
225 avctx->flags |= AV_CODEC_FLAG_INTERLACED_DCT;
227 avctx->flags &= ~AV_CODEC_FLAG_INTERLACED_DCT;
229 if (format->video_signal_description.video_full_range_flag)
230 avctx->color_range = AVCOL_RANGE_JPEG;
232 avctx->color_range = AVCOL_RANGE_MPEG;
234 avctx->color_primaries = format->video_signal_description.color_primaries;
235 avctx->color_trc = format->video_signal_description.transfer_characteristics;
236 avctx->colorspace = format->video_signal_description.matrix_coefficients;
239 avctx->bit_rate = format->bitrate;
241 if (format->frame_rate.numerator && format->frame_rate.denominator) {
242 avctx->framerate.num = format->frame_rate.numerator;
243 avctx->framerate.den = format->frame_rate.denominator;
247 && avctx->coded_width == format->coded_width
248 && avctx->coded_height == format->coded_height
249 && avctx->width == old_width
250 && avctx->height == old_height
251 && ctx->chroma_format == format->chroma_format
252 && ctx->codec_type == format->codec)
255 if (ctx->cudecoder) {
256 av_log(avctx, AV_LOG_TRACE, "Re-initializing decoder\n");
257 ctx->internal_error = CHECK_CU(ctx->cvdl->cuvidDestroyDecoder(ctx->cudecoder));
258 if (ctx->internal_error < 0)
260 ctx->cudecoder = NULL;
263 if (hwframe_ctx->pool && (
264 hwframe_ctx->width < avctx->width ||
265 hwframe_ctx->height < avctx->height ||
266 hwframe_ctx->format != AV_PIX_FMT_CUDA ||
267 hwframe_ctx->sw_format != avctx->sw_pix_fmt)) {
268 av_log(avctx, AV_LOG_ERROR, "AVHWFramesContext is already initialized with incompatible parameters\n");
269 av_log(avctx, AV_LOG_DEBUG, "width: %d <-> %d\n", hwframe_ctx->width, avctx->width);
270 av_log(avctx, AV_LOG_DEBUG, "height: %d <-> %d\n", hwframe_ctx->height, avctx->height);
271 av_log(avctx, AV_LOG_DEBUG, "format: %s <-> cuda\n", av_get_pix_fmt_name(hwframe_ctx->format));
272 av_log(avctx, AV_LOG_DEBUG, "sw_format: %s <-> %s\n",
273 av_get_pix_fmt_name(hwframe_ctx->sw_format), av_get_pix_fmt_name(avctx->sw_pix_fmt));
274 ctx->internal_error = AVERROR(EINVAL);
278 ctx->chroma_format = format->chroma_format;
280 cuinfo.CodecType = ctx->codec_type = format->codec;
281 cuinfo.ChromaFormat = format->chroma_format;
283 switch (avctx->sw_pix_fmt) {
284 case AV_PIX_FMT_NV12:
285 cuinfo.OutputFormat = cudaVideoSurfaceFormat_NV12;
287 case AV_PIX_FMT_P010:
288 case AV_PIX_FMT_P016:
289 cuinfo.OutputFormat = cudaVideoSurfaceFormat_P016;
291 case AV_PIX_FMT_YUV444P:
292 cuinfo.OutputFormat = cudaVideoSurfaceFormat_YUV444;
294 case AV_PIX_FMT_YUV444P16:
295 cuinfo.OutputFormat = cudaVideoSurfaceFormat_YUV444_16Bit;
298 av_log(avctx, AV_LOG_ERROR, "Unsupported output format: %s\n",
299 av_get_pix_fmt_name(avctx->sw_pix_fmt));
300 ctx->internal_error = AVERROR(EINVAL);
304 cuinfo.ulNumDecodeSurfaces = ctx->nb_surfaces;
305 cuinfo.ulNumOutputSurfaces = 1;
306 cuinfo.ulCreationFlags = cudaVideoCreate_PreferCUVID;
307 cuinfo.bitDepthMinus8 = format->bit_depth_luma_minus8;
308 cuinfo.DeinterlaceMode = ctx->deint_mode_current;
310 if (ctx->deint_mode_current != cudaVideoDeinterlaceMode_Weave && !ctx->drop_second_field)
311 avctx->framerate = av_mul_q(avctx->framerate, (AVRational){2, 1});
313 ctx->internal_error = CHECK_CU(ctx->cvdl->cuvidCreateDecoder(&ctx->cudecoder, &cuinfo));
314 if (ctx->internal_error < 0)
317 if (!hwframe_ctx->pool) {
318 hwframe_ctx->format = AV_PIX_FMT_CUDA;
319 hwframe_ctx->sw_format = avctx->sw_pix_fmt;
320 hwframe_ctx->width = avctx->width;
321 hwframe_ctx->height = avctx->height;
323 if ((ctx->internal_error = av_hwframe_ctx_init(ctx->hwframe)) < 0) {
324 av_log(avctx, AV_LOG_ERROR, "av_hwframe_ctx_init failed\n");
332 static int CUDAAPI cuvid_handle_picture_decode(void *opaque, CUVIDPICPARAMS* picparams)
334 AVCodecContext *avctx = opaque;
335 CuvidContext *ctx = avctx->priv_data;
337 av_log(avctx, AV_LOG_TRACE, "pfnDecodePicture\n");
339 ctx->key_frame[picparams->CurrPicIdx] = picparams->intra_pic_flag;
341 ctx->internal_error = CHECK_CU(ctx->cvdl->cuvidDecodePicture(ctx->cudecoder, picparams));
342 if (ctx->internal_error < 0)
348 static int CUDAAPI cuvid_handle_picture_display(void *opaque, CUVIDPARSERDISPINFO* dispinfo)
350 AVCodecContext *avctx = opaque;
351 CuvidContext *ctx = avctx->priv_data;
352 CuvidParsedFrame parsed_frame = { { 0 } };
354 parsed_frame.dispinfo = *dispinfo;
355 ctx->internal_error = 0;
357 // For some reason, dispinfo->progressive_frame is sometimes wrong.
358 parsed_frame.dispinfo.progressive_frame = ctx->progressive_sequence;
360 if (ctx->deint_mode_current == cudaVideoDeinterlaceMode_Weave) {
361 av_fifo_generic_write(ctx->frame_queue, &parsed_frame, sizeof(CuvidParsedFrame), NULL);
363 parsed_frame.is_deinterlacing = 1;
364 av_fifo_generic_write(ctx->frame_queue, &parsed_frame, sizeof(CuvidParsedFrame), NULL);
365 if (!ctx->drop_second_field) {
366 parsed_frame.second_field = 1;
367 av_fifo_generic_write(ctx->frame_queue, &parsed_frame, sizeof(CuvidParsedFrame), NULL);
374 static int cuvid_is_buffer_full(AVCodecContext *avctx)
376 CuvidContext *ctx = avctx->priv_data;
378 int delay = ctx->cuparseinfo.ulMaxDisplayDelay;
379 if (ctx->deint_mode != cudaVideoDeinterlaceMode_Weave && !ctx->drop_second_field)
382 return (av_fifo_size(ctx->frame_queue) / sizeof(CuvidParsedFrame)) + delay >= ctx->nb_surfaces;
385 static int cuvid_decode_packet(AVCodecContext *avctx, const AVPacket *avpkt)
387 CuvidContext *ctx = avctx->priv_data;
388 AVHWDeviceContext *device_ctx = (AVHWDeviceContext*)ctx->hwdevice->data;
389 AVCUDADeviceContext *device_hwctx = device_ctx->hwctx;
390 CUcontext dummy, cuda_ctx = device_hwctx->cuda_ctx;
391 CUVIDSOURCEDATAPACKET cupkt;
392 int ret = 0, eret = 0, is_flush = ctx->decoder_flushing;
394 av_log(avctx, AV_LOG_TRACE, "cuvid_decode_packet\n");
396 if (is_flush && avpkt && avpkt->size)
399 if (cuvid_is_buffer_full(avctx) && avpkt && avpkt->size)
400 return AVERROR(EAGAIN);
402 ret = CHECK_CU(ctx->cudl->cuCtxPushCurrent(cuda_ctx));
407 memset(&cupkt, 0, sizeof(cupkt));
409 if (avpkt && avpkt->size) {
410 cupkt.payload_size = avpkt->size;
411 cupkt.payload = avpkt->data;
413 if (avpkt->pts != AV_NOPTS_VALUE) {
414 cupkt.flags = CUVID_PKT_TIMESTAMP;
415 if (avctx->pkt_timebase.num && avctx->pkt_timebase.den)
416 cupkt.timestamp = av_rescale_q(avpkt->pts, avctx->pkt_timebase, (AVRational){1, 10000000});
418 cupkt.timestamp = avpkt->pts;
421 cupkt.flags = CUVID_PKT_ENDOFSTREAM;
422 ctx->decoder_flushing = 1;
425 ret = CHECK_CU(ctx->cvdl->cuvidParseVideoData(ctx->cuparser, &cupkt));
430 // cuvidParseVideoData doesn't return an error just because stuff failed...
431 if (ctx->internal_error) {
432 av_log(avctx, AV_LOG_ERROR, "cuvid decode callback error\n");
433 ret = ctx->internal_error;
438 eret = CHECK_CU(ctx->cudl->cuCtxPopCurrent(&dummy));
450 static int cuvid_output_frame(AVCodecContext *avctx, AVFrame *frame)
452 CuvidContext *ctx = avctx->priv_data;
453 AVHWDeviceContext *device_ctx = (AVHWDeviceContext*)ctx->hwdevice->data;
454 AVCUDADeviceContext *device_hwctx = device_ctx->hwctx;
455 CUcontext dummy, cuda_ctx = device_hwctx->cuda_ctx;
456 CUdeviceptr mapped_frame = 0;
457 int ret = 0, eret = 0;
459 av_log(avctx, AV_LOG_TRACE, "cuvid_output_frame\n");
461 if (ctx->decoder_flushing) {
462 ret = cuvid_decode_packet(avctx, NULL);
463 if (ret < 0 && ret != AVERROR_EOF)
467 if (!cuvid_is_buffer_full(avctx)) {
469 ret = ff_decode_get_packet(avctx, &pkt);
470 if (ret < 0 && ret != AVERROR_EOF)
472 ret = cuvid_decode_packet(avctx, &pkt);
473 av_packet_unref(&pkt);
474 // cuvid_is_buffer_full() should avoid this.
475 if (ret == AVERROR(EAGAIN))
476 ret = AVERROR_EXTERNAL;
477 if (ret < 0 && ret != AVERROR_EOF)
481 ret = CHECK_CU(ctx->cudl->cuCtxPushCurrent(cuda_ctx));
485 if (av_fifo_size(ctx->frame_queue)) {
486 const AVPixFmtDescriptor *pixdesc;
487 CuvidParsedFrame parsed_frame;
488 CUVIDPROCPARAMS params;
489 unsigned int pitch = 0;
493 av_fifo_generic_read(ctx->frame_queue, &parsed_frame, sizeof(CuvidParsedFrame), NULL);
495 memset(¶ms, 0, sizeof(params));
496 params.progressive_frame = parsed_frame.dispinfo.progressive_frame;
497 params.second_field = parsed_frame.second_field;
498 params.top_field_first = parsed_frame.dispinfo.top_field_first;
500 ret = CHECK_CU(ctx->cvdl->cuvidMapVideoFrame(ctx->cudecoder, parsed_frame.dispinfo.picture_index, &mapped_frame, &pitch, ¶ms));
504 if (avctx->pix_fmt == AV_PIX_FMT_CUDA) {
505 ret = av_hwframe_get_buffer(ctx->hwframe, frame, 0);
507 av_log(avctx, AV_LOG_ERROR, "av_hwframe_get_buffer failed\n");
511 ret = ff_decode_frame_props(avctx, frame);
513 av_log(avctx, AV_LOG_ERROR, "ff_decode_frame_props failed\n");
517 pixdesc = av_pix_fmt_desc_get(avctx->sw_pix_fmt);
519 for (i = 0; i < pixdesc->nb_components; i++) {
520 int height = avctx->height >> (i ? pixdesc->log2_chroma_h : 0);
521 CUDA_MEMCPY2D cpy = {
522 .srcMemoryType = CU_MEMORYTYPE_DEVICE,
523 .dstMemoryType = CU_MEMORYTYPE_DEVICE,
524 .srcDevice = mapped_frame,
525 .dstDevice = (CUdeviceptr)frame->data[i],
527 .dstPitch = frame->linesize[i],
529 .WidthInBytes = FFMIN(pitch, frame->linesize[i]),
533 ret = CHECK_CU(ctx->cudl->cuMemcpy2DAsync(&cpy, device_hwctx->stream));
539 } else if (avctx->pix_fmt == AV_PIX_FMT_NV12 ||
540 avctx->pix_fmt == AV_PIX_FMT_P010 ||
541 avctx->pix_fmt == AV_PIX_FMT_P016 ||
542 avctx->pix_fmt == AV_PIX_FMT_YUV444P ||
543 avctx->pix_fmt == AV_PIX_FMT_YUV444P16) {
544 unsigned int offset = 0;
545 AVFrame *tmp_frame = av_frame_alloc();
547 av_log(avctx, AV_LOG_ERROR, "av_frame_alloc failed\n");
548 ret = AVERROR(ENOMEM);
552 pixdesc = av_pix_fmt_desc_get(avctx->sw_pix_fmt);
554 tmp_frame->format = AV_PIX_FMT_CUDA;
555 tmp_frame->hw_frames_ctx = av_buffer_ref(ctx->hwframe);
556 if (!tmp_frame->hw_frames_ctx) {
557 ret = AVERROR(ENOMEM);
558 av_frame_free(&tmp_frame);
562 tmp_frame->width = avctx->width;
563 tmp_frame->height = avctx->height;
566 * Note that the following logic would not work for three plane
567 * YUV420 because the pitch value is different for the chroma
570 for (i = 0; i < pixdesc->nb_components; i++) {
571 tmp_frame->data[i] = (uint8_t*)mapped_frame + offset;
572 tmp_frame->linesize[i] = pitch;
573 offset += pitch * (avctx->height >> (i ? pixdesc->log2_chroma_h : 0));
576 ret = ff_get_buffer(avctx, frame, 0);
578 av_log(avctx, AV_LOG_ERROR, "ff_get_buffer failed\n");
579 av_frame_free(&tmp_frame);
583 ret = av_hwframe_transfer_data(frame, tmp_frame, 0);
585 av_log(avctx, AV_LOG_ERROR, "av_hwframe_transfer_data failed\n");
586 av_frame_free(&tmp_frame);
589 av_frame_free(&tmp_frame);
595 frame->key_frame = ctx->key_frame[parsed_frame.dispinfo.picture_index];
596 frame->width = avctx->width;
597 frame->height = avctx->height;
598 if (avctx->pkt_timebase.num && avctx->pkt_timebase.den)
599 frame->pts = av_rescale_q(parsed_frame.dispinfo.timestamp, (AVRational){1, 10000000}, avctx->pkt_timebase);
601 frame->pts = parsed_frame.dispinfo.timestamp;
603 if (parsed_frame.second_field) {
604 if (ctx->prev_pts == INT64_MIN) {
605 ctx->prev_pts = frame->pts;
606 frame->pts += (avctx->pkt_timebase.den * avctx->framerate.den) / (avctx->pkt_timebase.num * avctx->framerate.num);
608 int pts_diff = (frame->pts - ctx->prev_pts) / 2;
609 ctx->prev_pts = frame->pts;
610 frame->pts += pts_diff;
614 /* CUVIDs opaque reordering breaks the internal pkt logic.
615 * So set pkt_pts and clear all the other pkt_ fields.
618 FF_DISABLE_DEPRECATION_WARNINGS
619 frame->pkt_pts = frame->pts;
620 FF_ENABLE_DEPRECATION_WARNINGS
623 frame->pkt_duration = 0;
624 frame->pkt_size = -1;
626 frame->interlaced_frame = !parsed_frame.is_deinterlacing && !parsed_frame.dispinfo.progressive_frame;
628 if (frame->interlaced_frame)
629 frame->top_field_first = parsed_frame.dispinfo.top_field_first;
630 } else if (ctx->decoder_flushing) {
633 ret = AVERROR(EAGAIN);
638 av_frame_unref(frame);
641 eret = CHECK_CU(ctx->cvdl->cuvidUnmapVideoFrame(ctx->cudecoder, mapped_frame));
643 eret = CHECK_CU(ctx->cudl->cuCtxPopCurrent(&dummy));
651 static int cuvid_decode_frame(AVCodecContext *avctx, void *data, int *got_frame, AVPacket *avpkt)
653 CuvidContext *ctx = avctx->priv_data;
654 AVFrame *frame = data;
657 av_log(avctx, AV_LOG_TRACE, "cuvid_decode_frame\n");
659 if (ctx->deint_mode_current != cudaVideoDeinterlaceMode_Weave) {
660 av_log(avctx, AV_LOG_ERROR, "Deinterlacing is not supported via the old API\n");
661 return AVERROR(EINVAL);
664 if (!ctx->decoder_flushing) {
665 ret = cuvid_decode_packet(avctx, avpkt);
670 ret = cuvid_output_frame(avctx, frame);
671 if (ret == AVERROR(EAGAIN) || ret == AVERROR_EOF) {
673 } else if (ret < 0) {
682 static av_cold int cuvid_decode_end(AVCodecContext *avctx)
684 CuvidContext *ctx = avctx->priv_data;
685 AVHWDeviceContext *device_ctx = (AVHWDeviceContext *)ctx->hwdevice->data;
686 AVCUDADeviceContext *device_hwctx = device_ctx->hwctx;
687 CUcontext dummy, cuda_ctx = device_hwctx->cuda_ctx;
689 av_fifo_freep(&ctx->frame_queue);
691 ctx->cudl->cuCtxPushCurrent(cuda_ctx);
694 ctx->cvdl->cuvidDestroyVideoParser(ctx->cuparser);
697 ctx->cvdl->cuvidDestroyDecoder(ctx->cudecoder);
699 ctx->cudl->cuCtxPopCurrent(&dummy);
703 av_buffer_unref(&ctx->hwframe);
704 av_buffer_unref(&ctx->hwdevice);
706 av_freep(&ctx->key_frame);
707 av_freep(&ctx->cuparse_ext);
709 cuvid_free_functions(&ctx->cvdl);
714 static int cuvid_test_capabilities(AVCodecContext *avctx,
715 const CUVIDPARSERPARAMS *cuparseinfo,
720 CuvidContext *ctx = avctx->priv_data;
721 CUVIDDECODECAPS *caps;
722 int res8 = 0, res10 = 0, res12 = 0;
724 if (!ctx->cvdl->cuvidGetDecoderCaps) {
725 av_log(avctx, AV_LOG_WARNING, "Used Nvidia driver is too old to perform a capability check.\n");
726 av_log(avctx, AV_LOG_WARNING, "The minimum required version is "
727 #if defined(_WIN32) || defined(__CYGWIN__)
732 ". Continuing blind.\n");
733 ctx->caps8.bIsSupported = ctx->caps10.bIsSupported = 1;
734 // 12 bit was not supported before the capability check was introduced, so disable it.
735 ctx->caps12.bIsSupported = 0;
739 ctx->caps8.eCodecType = ctx->caps10.eCodecType = ctx->caps12.eCodecType
740 = cuparseinfo->CodecType;
741 ctx->caps8.eChromaFormat = ctx->caps10.eChromaFormat = ctx->caps12.eChromaFormat
742 = cudaVideoChromaFormat_420;
744 ctx->caps8.nBitDepthMinus8 = 0;
745 ctx->caps10.nBitDepthMinus8 = 2;
746 ctx->caps12.nBitDepthMinus8 = 4;
748 res8 = CHECK_CU(ctx->cvdl->cuvidGetDecoderCaps(&ctx->caps8));
749 res10 = CHECK_CU(ctx->cvdl->cuvidGetDecoderCaps(&ctx->caps10));
750 res12 = CHECK_CU(ctx->cvdl->cuvidGetDecoderCaps(&ctx->caps12));
752 av_log(avctx, AV_LOG_VERBOSE, "CUVID capabilities for %s:\n", avctx->codec->name);
753 av_log(avctx, AV_LOG_VERBOSE, "8 bit: supported: %d, min_width: %d, max_width: %d, min_height: %d, max_height: %d\n",
754 ctx->caps8.bIsSupported, ctx->caps8.nMinWidth, ctx->caps8.nMaxWidth, ctx->caps8.nMinHeight, ctx->caps8.nMaxHeight);
755 av_log(avctx, AV_LOG_VERBOSE, "10 bit: supported: %d, min_width: %d, max_width: %d, min_height: %d, max_height: %d\n",
756 ctx->caps10.bIsSupported, ctx->caps10.nMinWidth, ctx->caps10.nMaxWidth, ctx->caps10.nMinHeight, ctx->caps10.nMaxHeight);
757 av_log(avctx, AV_LOG_VERBOSE, "12 bit: supported: %d, min_width: %d, max_width: %d, min_height: %d, max_height: %d\n",
758 ctx->caps12.bIsSupported, ctx->caps12.nMinWidth, ctx->caps12.nMaxWidth, ctx->caps12.nMinHeight, ctx->caps12.nMaxHeight);
777 if (!ctx->caps8.bIsSupported) {
778 av_log(avctx, AV_LOG_ERROR, "Codec %s is not supported.\n", avctx->codec->name);
779 return AVERROR(EINVAL);
782 if (!caps->bIsSupported) {
783 av_log(avctx, AV_LOG_ERROR, "Bit depth %d is not supported.\n", bit_depth);
784 return AVERROR(EINVAL);
787 if (probed_width > caps->nMaxWidth || probed_width < caps->nMinWidth) {
788 av_log(avctx, AV_LOG_ERROR, "Video width %d not within range from %d to %d\n",
789 probed_width, caps->nMinWidth, caps->nMaxWidth);
790 return AVERROR(EINVAL);
793 if (probed_height > caps->nMaxHeight || probed_height < caps->nMinHeight) {
794 av_log(avctx, AV_LOG_ERROR, "Video height %d not within range from %d to %d\n",
795 probed_height, caps->nMinHeight, caps->nMaxHeight);
796 return AVERROR(EINVAL);
799 if ((probed_width * probed_height) / 256 > caps->nMaxMBCount) {
800 av_log(avctx, AV_LOG_ERROR, "Video macroblock count %d exceeds maximum of %d\n",
801 (int)(probed_width * probed_height) / 256, caps->nMaxMBCount);
802 return AVERROR(EINVAL);
808 static av_cold int cuvid_decode_init(AVCodecContext *avctx)
810 CuvidContext *ctx = avctx->priv_data;
811 AVCUDADeviceContext *device_hwctx;
812 AVHWDeviceContext *device_ctx;
813 AVHWFramesContext *hwframe_ctx;
814 CUVIDSOURCEDATAPACKET seq_pkt;
815 CUcontext cuda_ctx = NULL;
821 enum AVPixelFormat pix_fmts[3] = { AV_PIX_FMT_CUDA,
825 int probed_width = avctx->coded_width ? avctx->coded_width : 1280;
826 int probed_height = avctx->coded_height ? avctx->coded_height : 720;
827 int probed_bit_depth = 8;
829 const AVPixFmtDescriptor *probe_desc = av_pix_fmt_desc_get(avctx->pix_fmt);
830 if (probe_desc && probe_desc->nb_components)
831 probed_bit_depth = probe_desc->comp[0].depth;
833 // Accelerated transcoding scenarios with 'ffmpeg' require that the
834 // pix_fmt be set to AV_PIX_FMT_CUDA early. The sw_pix_fmt, and the
835 // pix_fmt for non-accelerated transcoding, do not need to be correct
836 // but need to be set to something. We arbitrarily pick NV12.
837 ret = ff_get_format(avctx, pix_fmts);
839 av_log(avctx, AV_LOG_ERROR, "ff_get_format failed: %d\n", ret);
842 avctx->pix_fmt = ret;
844 if (ctx->resize_expr && sscanf(ctx->resize_expr, "%dx%d",
845 &ctx->resize.width, &ctx->resize.height) != 2) {
846 av_log(avctx, AV_LOG_ERROR, "Invalid resize expressions\n");
847 ret = AVERROR(EINVAL);
851 if (ctx->crop_expr && sscanf(ctx->crop_expr, "%dx%dx%dx%d",
852 &ctx->crop.top, &ctx->crop.bottom,
853 &ctx->crop.left, &ctx->crop.right) != 4) {
854 av_log(avctx, AV_LOG_ERROR, "Invalid cropping expressions\n");
855 ret = AVERROR(EINVAL);
859 ret = cuvid_load_functions(&ctx->cvdl, avctx);
861 av_log(avctx, AV_LOG_ERROR, "Failed loading nvcuvid.\n");
865 ctx->frame_queue = av_fifo_alloc(ctx->nb_surfaces * sizeof(CuvidParsedFrame));
866 if (!ctx->frame_queue) {
867 ret = AVERROR(ENOMEM);
871 if (avctx->hw_frames_ctx) {
872 ctx->hwframe = av_buffer_ref(avctx->hw_frames_ctx);
874 ret = AVERROR(ENOMEM);
878 hwframe_ctx = (AVHWFramesContext*)ctx->hwframe->data;
880 ctx->hwdevice = av_buffer_ref(hwframe_ctx->device_ref);
881 if (!ctx->hwdevice) {
882 ret = AVERROR(ENOMEM);
886 if (avctx->hw_device_ctx) {
887 ctx->hwdevice = av_buffer_ref(avctx->hw_device_ctx);
888 if (!ctx->hwdevice) {
889 ret = AVERROR(ENOMEM);
893 ret = av_hwdevice_ctx_create(&ctx->hwdevice, AV_HWDEVICE_TYPE_CUDA, ctx->cu_gpu, NULL, 0);
898 ctx->hwframe = av_hwframe_ctx_alloc(ctx->hwdevice);
900 av_log(avctx, AV_LOG_ERROR, "av_hwframe_ctx_alloc failed\n");
901 ret = AVERROR(ENOMEM);
905 hwframe_ctx = (AVHWFramesContext*)ctx->hwframe->data;
908 device_ctx = hwframe_ctx->device_ctx;
909 device_hwctx = device_ctx->hwctx;
911 cuda_ctx = device_hwctx->cuda_ctx;
912 ctx->cudl = device_hwctx->internal->cuda_dl;
914 memset(&ctx->cuparseinfo, 0, sizeof(ctx->cuparseinfo));
915 memset(&seq_pkt, 0, sizeof(seq_pkt));
917 switch (avctx->codec->id) {
918 #if CONFIG_H264_CUVID_DECODER
919 case AV_CODEC_ID_H264:
920 ctx->cuparseinfo.CodecType = cudaVideoCodec_H264;
923 #if CONFIG_HEVC_CUVID_DECODER
924 case AV_CODEC_ID_HEVC:
925 ctx->cuparseinfo.CodecType = cudaVideoCodec_HEVC;
928 #if CONFIG_MJPEG_CUVID_DECODER
929 case AV_CODEC_ID_MJPEG:
930 ctx->cuparseinfo.CodecType = cudaVideoCodec_JPEG;
933 #if CONFIG_MPEG1_CUVID_DECODER
934 case AV_CODEC_ID_MPEG1VIDEO:
935 ctx->cuparseinfo.CodecType = cudaVideoCodec_MPEG1;
938 #if CONFIG_MPEG2_CUVID_DECODER
939 case AV_CODEC_ID_MPEG2VIDEO:
940 ctx->cuparseinfo.CodecType = cudaVideoCodec_MPEG2;
943 #if CONFIG_MPEG4_CUVID_DECODER
944 case AV_CODEC_ID_MPEG4:
945 ctx->cuparseinfo.CodecType = cudaVideoCodec_MPEG4;
948 #if CONFIG_VP8_CUVID_DECODER
949 case AV_CODEC_ID_VP8:
950 ctx->cuparseinfo.CodecType = cudaVideoCodec_VP8;
953 #if CONFIG_VP9_CUVID_DECODER
954 case AV_CODEC_ID_VP9:
955 ctx->cuparseinfo.CodecType = cudaVideoCodec_VP9;
958 #if CONFIG_VC1_CUVID_DECODER
959 case AV_CODEC_ID_VC1:
960 ctx->cuparseinfo.CodecType = cudaVideoCodec_VC1;
963 #if CONFIG_AV1_CUVID_DECODER && defined(CUVID_HAS_AV1_SUPPORT)
964 case AV_CODEC_ID_AV1:
965 ctx->cuparseinfo.CodecType = cudaVideoCodec_AV1;
969 av_log(avctx, AV_LOG_ERROR, "Invalid CUVID codec!\n");
973 if (avctx->codec->bsfs) {
974 const AVCodecParameters *par = avctx->internal->bsf->par_out;
975 extradata = par->extradata;
976 extradata_size = par->extradata_size;
978 extradata = avctx->extradata;
979 extradata_size = avctx->extradata_size;
982 ctx->cuparse_ext = av_mallocz(sizeof(*ctx->cuparse_ext)
983 + FFMAX(extradata_size - (int)sizeof(ctx->cuparse_ext->raw_seqhdr_data), 0));
984 if (!ctx->cuparse_ext) {
985 ret = AVERROR(ENOMEM);
989 if (extradata_size > 0)
990 memcpy(ctx->cuparse_ext->raw_seqhdr_data, extradata, extradata_size);
991 ctx->cuparse_ext->format.seqhdr_data_length = extradata_size;
993 ctx->cuparseinfo.pExtVideoInfo = ctx->cuparse_ext;
995 ctx->key_frame = av_mallocz(ctx->nb_surfaces * sizeof(int));
996 if (!ctx->key_frame) {
997 ret = AVERROR(ENOMEM);
1001 ctx->cuparseinfo.ulMaxNumDecodeSurfaces = ctx->nb_surfaces;
1002 ctx->cuparseinfo.ulMaxDisplayDelay = (avctx->flags & AV_CODEC_FLAG_LOW_DELAY) ? 0 : 4;
1003 ctx->cuparseinfo.pUserData = avctx;
1004 ctx->cuparseinfo.pfnSequenceCallback = cuvid_handle_video_sequence;
1005 ctx->cuparseinfo.pfnDecodePicture = cuvid_handle_picture_decode;
1006 ctx->cuparseinfo.pfnDisplayPicture = cuvid_handle_picture_display;
1008 ret = CHECK_CU(ctx->cudl->cuCtxPushCurrent(cuda_ctx));
1012 ret = cuvid_test_capabilities(avctx, &ctx->cuparseinfo,
1019 ret = CHECK_CU(ctx->cvdl->cuvidCreateVideoParser(&ctx->cuparser, &ctx->cuparseinfo));
1023 seq_pkt.payload = ctx->cuparse_ext->raw_seqhdr_data;
1024 seq_pkt.payload_size = ctx->cuparse_ext->format.seqhdr_data_length;
1026 if (seq_pkt.payload && seq_pkt.payload_size) {
1027 ret = CHECK_CU(ctx->cvdl->cuvidParseVideoData(ctx->cuparser, &seq_pkt));
1032 ret = CHECK_CU(ctx->cudl->cuCtxPopCurrent(&dummy));
1036 ctx->prev_pts = INT64_MIN;
1038 if (!avctx->pkt_timebase.num || !avctx->pkt_timebase.den)
1039 av_log(avctx, AV_LOG_WARNING, "Invalid pkt_timebase, passing timestamps as-is.\n");
1044 cuvid_decode_end(avctx);
1048 static void cuvid_flush(AVCodecContext *avctx)
1050 CuvidContext *ctx = avctx->priv_data;
1051 AVHWDeviceContext *device_ctx = (AVHWDeviceContext*)ctx->hwdevice->data;
1052 AVCUDADeviceContext *device_hwctx = device_ctx->hwctx;
1053 CUcontext dummy, cuda_ctx = device_hwctx->cuda_ctx;
1054 CUVIDSOURCEDATAPACKET seq_pkt = { 0 };
1057 ret = CHECK_CU(ctx->cudl->cuCtxPushCurrent(cuda_ctx));
1061 av_fifo_freep(&ctx->frame_queue);
1063 ctx->frame_queue = av_fifo_alloc(ctx->nb_surfaces * sizeof(CuvidParsedFrame));
1064 if (!ctx->frame_queue) {
1065 av_log(avctx, AV_LOG_ERROR, "Failed to recreate frame queue on flush\n");
1069 if (ctx->cudecoder) {
1070 ctx->cvdl->cuvidDestroyDecoder(ctx->cudecoder);
1071 ctx->cudecoder = NULL;
1074 if (ctx->cuparser) {
1075 ctx->cvdl->cuvidDestroyVideoParser(ctx->cuparser);
1076 ctx->cuparser = NULL;
1079 ret = CHECK_CU(ctx->cvdl->cuvidCreateVideoParser(&ctx->cuparser, &ctx->cuparseinfo));
1083 seq_pkt.payload = ctx->cuparse_ext->raw_seqhdr_data;
1084 seq_pkt.payload_size = ctx->cuparse_ext->format.seqhdr_data_length;
1086 if (seq_pkt.payload && seq_pkt.payload_size) {
1087 ret = CHECK_CU(ctx->cvdl->cuvidParseVideoData(ctx->cuparser, &seq_pkt));
1092 ret = CHECK_CU(ctx->cudl->cuCtxPopCurrent(&dummy));
1096 ctx->prev_pts = INT64_MIN;
1097 ctx->decoder_flushing = 0;
1101 av_log(avctx, AV_LOG_ERROR, "CUDA reinit on flush failed\n");
1104 #define OFFSET(x) offsetof(CuvidContext, x)
1105 #define VD AV_OPT_FLAG_VIDEO_PARAM | AV_OPT_FLAG_DECODING_PARAM
1106 static const AVOption options[] = {
1107 { "deint", "Set deinterlacing mode", OFFSET(deint_mode), AV_OPT_TYPE_INT, { .i64 = cudaVideoDeinterlaceMode_Weave }, cudaVideoDeinterlaceMode_Weave, cudaVideoDeinterlaceMode_Adaptive, VD, "deint" },
1108 { "weave", "Weave deinterlacing (do nothing)", 0, AV_OPT_TYPE_CONST, { .i64 = cudaVideoDeinterlaceMode_Weave }, 0, 0, VD, "deint" },
1109 { "bob", "Bob deinterlacing", 0, AV_OPT_TYPE_CONST, { .i64 = cudaVideoDeinterlaceMode_Bob }, 0, 0, VD, "deint" },
1110 { "adaptive", "Adaptive deinterlacing", 0, AV_OPT_TYPE_CONST, { .i64 = cudaVideoDeinterlaceMode_Adaptive }, 0, 0, VD, "deint" },
1111 { "gpu", "GPU to be used for decoding", OFFSET(cu_gpu), AV_OPT_TYPE_STRING, { .str = NULL }, 0, 0, VD },
1112 { "surfaces", "Maximum surfaces to be used for decoding", OFFSET(nb_surfaces), AV_OPT_TYPE_INT, { .i64 = 25 }, 0, INT_MAX, VD },
1113 { "drop_second_field", "Drop second field when deinterlacing", OFFSET(drop_second_field), AV_OPT_TYPE_BOOL, { .i64 = 0 }, 0, 1, VD },
1114 { "crop", "Crop (top)x(bottom)x(left)x(right)", OFFSET(crop_expr), AV_OPT_TYPE_STRING, { .str = NULL }, 0, 0, VD },
1115 { "resize", "Resize (width)x(height)", OFFSET(resize_expr), AV_OPT_TYPE_STRING, { .str = NULL }, 0, 0, VD },
1119 static const AVCodecHWConfigInternal *const cuvid_hw_configs[] = {
1120 &(const AVCodecHWConfigInternal) {
1122 .pix_fmt = AV_PIX_FMT_CUDA,
1123 .methods = AV_CODEC_HW_CONFIG_METHOD_HW_DEVICE_CTX |
1124 AV_CODEC_HW_CONFIG_METHOD_INTERNAL,
1125 .device_type = AV_HWDEVICE_TYPE_CUDA
1132 #define DEFINE_CUVID_CODEC(x, X, bsf_name) \
1133 static const AVClass x##_cuvid_class = { \
1134 .class_name = #x "_cuvid", \
1135 .item_name = av_default_item_name, \
1136 .option = options, \
1137 .version = LIBAVUTIL_VERSION_INT, \
1139 AVCodec ff_##x##_cuvid_decoder = { \
1140 .name = #x "_cuvid", \
1141 .long_name = NULL_IF_CONFIG_SMALL("Nvidia CUVID " #X " decoder"), \
1142 .type = AVMEDIA_TYPE_VIDEO, \
1143 .id = AV_CODEC_ID_##X, \
1144 .priv_data_size = sizeof(CuvidContext), \
1145 .priv_class = &x##_cuvid_class, \
1146 .init = cuvid_decode_init, \
1147 .close = cuvid_decode_end, \
1148 .decode = cuvid_decode_frame, \
1149 .receive_frame = cuvid_output_frame, \
1150 .flush = cuvid_flush, \
1152 .capabilities = AV_CODEC_CAP_DELAY | AV_CODEC_CAP_AVOID_PROBING | AV_CODEC_CAP_HARDWARE, \
1153 .pix_fmts = (const enum AVPixelFormat[]){ AV_PIX_FMT_CUDA, \
1157 AV_PIX_FMT_NONE }, \
1158 .hw_configs = cuvid_hw_configs, \
1159 .wrapper_name = "cuvid", \
1162 #if CONFIG_AV1_CUVID_DECODER && defined(CUVID_HAS_AV1_SUPPORT)
1163 DEFINE_CUVID_CODEC(av1, AV1, NULL)
1166 #if CONFIG_HEVC_CUVID_DECODER
1167 DEFINE_CUVID_CODEC(hevc, HEVC, "hevc_mp4toannexb")
1170 #if CONFIG_H264_CUVID_DECODER
1171 DEFINE_CUVID_CODEC(h264, H264, "h264_mp4toannexb")
1174 #if CONFIG_MJPEG_CUVID_DECODER
1175 DEFINE_CUVID_CODEC(mjpeg, MJPEG, NULL)
1178 #if CONFIG_MPEG1_CUVID_DECODER
1179 DEFINE_CUVID_CODEC(mpeg1, MPEG1VIDEO, NULL)
1182 #if CONFIG_MPEG2_CUVID_DECODER
1183 DEFINE_CUVID_CODEC(mpeg2, MPEG2VIDEO, NULL)
1186 #if CONFIG_MPEG4_CUVID_DECODER
1187 DEFINE_CUVID_CODEC(mpeg4, MPEG4, NULL)
1190 #if CONFIG_VP8_CUVID_DECODER
1191 DEFINE_CUVID_CODEC(vp8, VP8, NULL)
1194 #if CONFIG_VP9_CUVID_DECODER
1195 DEFINE_CUVID_CODEC(vp9, VP9, NULL)
1198 #if CONFIG_VC1_CUVID_DECODER
1199 DEFINE_CUVID_CODEC(vc1, VC1, NULL)