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"
39 typedef struct CuvidContext
43 CUvideodecoder cudecoder;
44 CUvideoparser cuparser;
48 int drop_second_field;
64 AVBufferRef *hwdevice;
69 AVFifoBuffer *frame_queue;
72 int deint_mode_current;
80 cudaVideoCodec codec_type;
81 cudaVideoChromaFormat chroma_format;
83 CUVIDDECODECAPS caps8, caps10, caps12;
85 CUVIDPARSERPARAMS cuparseinfo;
86 CUVIDEOFORMATEX cuparse_ext;
92 typedef struct CuvidParsedFrame
94 CUVIDPARSERDISPINFO dispinfo;
99 #define CHECK_CU(x) FF_CUDA_CHECK_DL(avctx, ctx->cudl, x)
101 static int CUDAAPI cuvid_handle_video_sequence(void *opaque, CUVIDEOFORMAT* format)
103 AVCodecContext *avctx = opaque;
104 CuvidContext *ctx = avctx->priv_data;
105 AVHWFramesContext *hwframe_ctx = (AVHWFramesContext*)ctx->hwframe->data;
106 CUVIDDECODECAPS *caps = NULL;
107 CUVIDDECODECREATEINFO cuinfo;
110 int old_width = avctx->width;
111 int old_height = avctx->height;
113 enum AVPixelFormat pix_fmts[3] = { AV_PIX_FMT_CUDA,
114 AV_PIX_FMT_NONE, // Will be updated below
117 av_log(avctx, AV_LOG_TRACE, "pfnSequenceCallback, progressive_sequence=%d\n", format->progressive_sequence);
119 memset(&cuinfo, 0, sizeof(cuinfo));
121 ctx->internal_error = 0;
123 avctx->coded_width = cuinfo.ulWidth = format->coded_width;
124 avctx->coded_height = cuinfo.ulHeight = format->coded_height;
127 cuinfo.display_area.left = format->display_area.left + ctx->crop.left;
128 cuinfo.display_area.top = format->display_area.top + ctx->crop.top;
129 cuinfo.display_area.right = format->display_area.right - ctx->crop.right;
130 cuinfo.display_area.bottom = format->display_area.bottom - ctx->crop.bottom;
132 // width and height need to be set before calling ff_get_format
133 if (ctx->resize_expr) {
134 avctx->width = ctx->resize.width;
135 avctx->height = ctx->resize.height;
137 avctx->width = cuinfo.display_area.right - cuinfo.display_area.left;
138 avctx->height = cuinfo.display_area.bottom - cuinfo.display_area.top;
141 // target width/height need to be multiples of two
142 cuinfo.ulTargetWidth = avctx->width = (avctx->width + 1) & ~1;
143 cuinfo.ulTargetHeight = avctx->height = (avctx->height + 1) & ~1;
145 // aspect ratio conversion, 1:1, depends on scaled resolution
146 cuinfo.target_rect.left = 0;
147 cuinfo.target_rect.top = 0;
148 cuinfo.target_rect.right = cuinfo.ulTargetWidth;
149 cuinfo.target_rect.bottom = cuinfo.ulTargetHeight;
151 switch (format->bit_depth_luma_minus8) {
153 pix_fmts[1] = AV_PIX_FMT_NV12;
157 pix_fmts[1] = AV_PIX_FMT_P010;
161 pix_fmts[1] = AV_PIX_FMT_P016;
168 if (!caps || !caps->bIsSupported) {
169 av_log(avctx, AV_LOG_ERROR, "unsupported bit depth: %d\n",
170 format->bit_depth_luma_minus8 + 8);
171 ctx->internal_error = AVERROR(EINVAL);
175 surface_fmt = ff_get_format(avctx, pix_fmts);
176 if (surface_fmt < 0) {
177 av_log(avctx, AV_LOG_ERROR, "ff_get_format failed: %d\n", surface_fmt);
178 ctx->internal_error = AVERROR(EINVAL);
182 av_log(avctx, AV_LOG_VERBOSE, "Formats: Original: %s | HW: %s | SW: %s\n",
183 av_get_pix_fmt_name(avctx->pix_fmt),
184 av_get_pix_fmt_name(surface_fmt),
185 av_get_pix_fmt_name(avctx->sw_pix_fmt));
187 avctx->pix_fmt = surface_fmt;
189 // Update our hwframe ctx, as the get_format callback might have refreshed it!
190 if (avctx->hw_frames_ctx) {
191 av_buffer_unref(&ctx->hwframe);
193 ctx->hwframe = av_buffer_ref(avctx->hw_frames_ctx);
195 ctx->internal_error = AVERROR(ENOMEM);
199 hwframe_ctx = (AVHWFramesContext*)ctx->hwframe->data;
202 ff_set_sar(avctx, av_div_q(
203 (AVRational){ format->display_aspect_ratio.x, format->display_aspect_ratio.y },
204 (AVRational){ avctx->width, avctx->height }));
206 ctx->deint_mode_current = format->progressive_sequence
207 ? cudaVideoDeinterlaceMode_Weave
210 if (!format->progressive_sequence && ctx->deint_mode_current == cudaVideoDeinterlaceMode_Weave)
211 avctx->flags |= AV_CODEC_FLAG_INTERLACED_DCT;
213 avctx->flags &= ~AV_CODEC_FLAG_INTERLACED_DCT;
215 if (format->video_signal_description.video_full_range_flag)
216 avctx->color_range = AVCOL_RANGE_JPEG;
218 avctx->color_range = AVCOL_RANGE_MPEG;
220 avctx->color_primaries = format->video_signal_description.color_primaries;
221 avctx->color_trc = format->video_signal_description.transfer_characteristics;
222 avctx->colorspace = format->video_signal_description.matrix_coefficients;
225 avctx->bit_rate = format->bitrate;
227 if (format->frame_rate.numerator && format->frame_rate.denominator) {
228 avctx->framerate.num = format->frame_rate.numerator;
229 avctx->framerate.den = format->frame_rate.denominator;
233 && avctx->coded_width == format->coded_width
234 && avctx->coded_height == format->coded_height
235 && avctx->width == old_width
236 && avctx->height == old_height
237 && ctx->chroma_format == format->chroma_format
238 && ctx->codec_type == format->codec)
241 if (ctx->cudecoder) {
242 av_log(avctx, AV_LOG_TRACE, "Re-initializing decoder\n");
243 ctx->internal_error = CHECK_CU(ctx->cvdl->cuvidDestroyDecoder(ctx->cudecoder));
244 if (ctx->internal_error < 0)
246 ctx->cudecoder = NULL;
249 if (hwframe_ctx->pool && (
250 hwframe_ctx->width < avctx->width ||
251 hwframe_ctx->height < avctx->height ||
252 hwframe_ctx->format != AV_PIX_FMT_CUDA ||
253 hwframe_ctx->sw_format != avctx->sw_pix_fmt)) {
254 av_log(avctx, AV_LOG_ERROR, "AVHWFramesContext is already initialized with incompatible parameters\n");
255 av_log(avctx, AV_LOG_DEBUG, "width: %d <-> %d\n", hwframe_ctx->width, avctx->width);
256 av_log(avctx, AV_LOG_DEBUG, "height: %d <-> %d\n", hwframe_ctx->height, avctx->height);
257 av_log(avctx, AV_LOG_DEBUG, "format: %s <-> cuda\n", av_get_pix_fmt_name(hwframe_ctx->format));
258 av_log(avctx, AV_LOG_DEBUG, "sw_format: %s <-> %s\n",
259 av_get_pix_fmt_name(hwframe_ctx->sw_format), av_get_pix_fmt_name(avctx->sw_pix_fmt));
260 ctx->internal_error = AVERROR(EINVAL);
264 if (format->chroma_format != cudaVideoChromaFormat_420) {
265 av_log(avctx, AV_LOG_ERROR, "Chroma formats other than 420 are not supported\n");
266 ctx->internal_error = AVERROR(EINVAL);
270 ctx->chroma_format = format->chroma_format;
272 cuinfo.CodecType = ctx->codec_type = format->codec;
273 cuinfo.ChromaFormat = format->chroma_format;
275 switch (avctx->sw_pix_fmt) {
276 case AV_PIX_FMT_NV12:
277 cuinfo.OutputFormat = cudaVideoSurfaceFormat_NV12;
279 case AV_PIX_FMT_P010:
280 case AV_PIX_FMT_P016:
281 cuinfo.OutputFormat = cudaVideoSurfaceFormat_P016;
284 av_log(avctx, AV_LOG_ERROR, "Output formats other than NV12, P010 or P016 are not supported\n");
285 ctx->internal_error = AVERROR(EINVAL);
289 cuinfo.ulNumDecodeSurfaces = ctx->nb_surfaces;
290 cuinfo.ulNumOutputSurfaces = 1;
291 cuinfo.ulCreationFlags = cudaVideoCreate_PreferCUVID;
292 cuinfo.bitDepthMinus8 = format->bit_depth_luma_minus8;
293 cuinfo.DeinterlaceMode = ctx->deint_mode_current;
295 if (ctx->deint_mode_current != cudaVideoDeinterlaceMode_Weave && !ctx->drop_second_field)
296 avctx->framerate = av_mul_q(avctx->framerate, (AVRational){2, 1});
298 ctx->internal_error = CHECK_CU(ctx->cvdl->cuvidCreateDecoder(&ctx->cudecoder, &cuinfo));
299 if (ctx->internal_error < 0)
302 if (!hwframe_ctx->pool) {
303 hwframe_ctx->format = AV_PIX_FMT_CUDA;
304 hwframe_ctx->sw_format = avctx->sw_pix_fmt;
305 hwframe_ctx->width = avctx->width;
306 hwframe_ctx->height = avctx->height;
308 if ((ctx->internal_error = av_hwframe_ctx_init(ctx->hwframe)) < 0) {
309 av_log(avctx, AV_LOG_ERROR, "av_hwframe_ctx_init failed\n");
317 static int CUDAAPI cuvid_handle_picture_decode(void *opaque, CUVIDPICPARAMS* picparams)
319 AVCodecContext *avctx = opaque;
320 CuvidContext *ctx = avctx->priv_data;
322 av_log(avctx, AV_LOG_TRACE, "pfnDecodePicture\n");
324 ctx->key_frame[picparams->CurrPicIdx] = picparams->intra_pic_flag;
326 ctx->internal_error = CHECK_CU(ctx->cvdl->cuvidDecodePicture(ctx->cudecoder, picparams));
327 if (ctx->internal_error < 0)
333 static int CUDAAPI cuvid_handle_picture_display(void *opaque, CUVIDPARSERDISPINFO* dispinfo)
335 AVCodecContext *avctx = opaque;
336 CuvidContext *ctx = avctx->priv_data;
337 CuvidParsedFrame parsed_frame = { { 0 } };
339 parsed_frame.dispinfo = *dispinfo;
340 ctx->internal_error = 0;
342 if (ctx->deint_mode_current == cudaVideoDeinterlaceMode_Weave) {
343 av_fifo_generic_write(ctx->frame_queue, &parsed_frame, sizeof(CuvidParsedFrame), NULL);
345 parsed_frame.is_deinterlacing = 1;
346 av_fifo_generic_write(ctx->frame_queue, &parsed_frame, sizeof(CuvidParsedFrame), NULL);
347 if (!ctx->drop_second_field) {
348 parsed_frame.second_field = 1;
349 av_fifo_generic_write(ctx->frame_queue, &parsed_frame, sizeof(CuvidParsedFrame), NULL);
356 static int cuvid_is_buffer_full(AVCodecContext *avctx)
358 CuvidContext *ctx = avctx->priv_data;
360 int delay = ctx->cuparseinfo.ulMaxDisplayDelay;
361 if (ctx->deint_mode != cudaVideoDeinterlaceMode_Weave && !ctx->drop_second_field)
364 return (av_fifo_size(ctx->frame_queue) / sizeof(CuvidParsedFrame)) + delay >= ctx->nb_surfaces;
367 static int cuvid_decode_packet(AVCodecContext *avctx, const AVPacket *avpkt)
369 CuvidContext *ctx = avctx->priv_data;
370 AVHWDeviceContext *device_ctx = (AVHWDeviceContext*)ctx->hwdevice->data;
371 AVCUDADeviceContext *device_hwctx = device_ctx->hwctx;
372 CUcontext dummy, cuda_ctx = device_hwctx->cuda_ctx;
373 CUVIDSOURCEDATAPACKET cupkt;
374 AVPacket filter_packet = { 0 };
375 AVPacket filtered_packet = { 0 };
376 int ret = 0, eret = 0, is_flush = ctx->decoder_flushing;
378 av_log(avctx, AV_LOG_TRACE, "cuvid_decode_packet\n");
380 if (is_flush && avpkt && avpkt->size)
383 if (cuvid_is_buffer_full(avctx) && avpkt && avpkt->size)
384 return AVERROR(EAGAIN);
386 if (ctx->bsf && avpkt && avpkt->size) {
387 if ((ret = av_packet_ref(&filter_packet, avpkt)) < 0) {
388 av_log(avctx, AV_LOG_ERROR, "av_packet_ref failed\n");
392 if ((ret = av_bsf_send_packet(ctx->bsf, &filter_packet)) < 0) {
393 av_log(avctx, AV_LOG_ERROR, "av_bsf_send_packet failed\n");
394 av_packet_unref(&filter_packet);
398 if ((ret = av_bsf_receive_packet(ctx->bsf, &filtered_packet)) < 0) {
399 av_log(avctx, AV_LOG_ERROR, "av_bsf_receive_packet failed\n");
403 avpkt = &filtered_packet;
406 ret = CHECK_CU(ctx->cudl->cuCtxPushCurrent(cuda_ctx));
408 av_packet_unref(&filtered_packet);
412 memset(&cupkt, 0, sizeof(cupkt));
414 if (avpkt && avpkt->size) {
415 cupkt.payload_size = avpkt->size;
416 cupkt.payload = avpkt->data;
418 if (avpkt->pts != AV_NOPTS_VALUE) {
419 cupkt.flags = CUVID_PKT_TIMESTAMP;
420 if (avctx->pkt_timebase.num && avctx->pkt_timebase.den)
421 cupkt.timestamp = av_rescale_q(avpkt->pts, avctx->pkt_timebase, (AVRational){1, 10000000});
423 cupkt.timestamp = avpkt->pts;
426 cupkt.flags = CUVID_PKT_ENDOFSTREAM;
427 ctx->decoder_flushing = 1;
430 ret = CHECK_CU(ctx->cvdl->cuvidParseVideoData(ctx->cuparser, &cupkt));
432 av_packet_unref(&filtered_packet);
437 // cuvidParseVideoData doesn't return an error just because stuff failed...
438 if (ctx->internal_error) {
439 av_log(avctx, AV_LOG_ERROR, "cuvid decode callback error\n");
440 ret = ctx->internal_error;
445 eret = CHECK_CU(ctx->cudl->cuCtxPopCurrent(&dummy));
457 static int cuvid_output_frame(AVCodecContext *avctx, AVFrame *frame)
459 CuvidContext *ctx = avctx->priv_data;
460 AVHWDeviceContext *device_ctx = (AVHWDeviceContext*)ctx->hwdevice->data;
461 AVCUDADeviceContext *device_hwctx = device_ctx->hwctx;
462 CUcontext dummy, cuda_ctx = device_hwctx->cuda_ctx;
463 CUdeviceptr mapped_frame = 0;
464 int ret = 0, eret = 0;
466 av_log(avctx, AV_LOG_TRACE, "cuvid_output_frame\n");
468 if (ctx->decoder_flushing) {
469 ret = cuvid_decode_packet(avctx, NULL);
470 if (ret < 0 && ret != AVERROR_EOF)
474 if (!cuvid_is_buffer_full(avctx)) {
476 ret = ff_decode_get_packet(avctx, &pkt);
477 if (ret < 0 && ret != AVERROR_EOF)
479 ret = cuvid_decode_packet(avctx, &pkt);
480 av_packet_unref(&pkt);
481 // cuvid_is_buffer_full() should avoid this.
482 if (ret == AVERROR(EAGAIN))
483 ret = AVERROR_EXTERNAL;
484 if (ret < 0 && ret != AVERROR_EOF)
488 ret = CHECK_CU(ctx->cudl->cuCtxPushCurrent(cuda_ctx));
492 if (av_fifo_size(ctx->frame_queue)) {
493 CuvidParsedFrame parsed_frame;
494 CUVIDPROCPARAMS params;
495 unsigned int pitch = 0;
499 av_fifo_generic_read(ctx->frame_queue, &parsed_frame, sizeof(CuvidParsedFrame), NULL);
501 memset(¶ms, 0, sizeof(params));
502 params.progressive_frame = parsed_frame.dispinfo.progressive_frame;
503 params.second_field = parsed_frame.second_field;
504 params.top_field_first = parsed_frame.dispinfo.top_field_first;
506 ret = CHECK_CU(ctx->cvdl->cuvidMapVideoFrame(ctx->cudecoder, parsed_frame.dispinfo.picture_index, &mapped_frame, &pitch, ¶ms));
510 if (avctx->pix_fmt == AV_PIX_FMT_CUDA) {
511 ret = av_hwframe_get_buffer(ctx->hwframe, frame, 0);
513 av_log(avctx, AV_LOG_ERROR, "av_hwframe_get_buffer failed\n");
517 ret = ff_decode_frame_props(avctx, frame);
519 av_log(avctx, AV_LOG_ERROR, "ff_decode_frame_props failed\n");
523 for (i = 0; i < 2; i++) {
524 CUDA_MEMCPY2D cpy = {
525 .srcMemoryType = CU_MEMORYTYPE_DEVICE,
526 .dstMemoryType = CU_MEMORYTYPE_DEVICE,
527 .srcDevice = mapped_frame,
528 .dstDevice = (CUdeviceptr)frame->data[i],
530 .dstPitch = frame->linesize[i],
532 .WidthInBytes = FFMIN(pitch, frame->linesize[i]),
533 .Height = avctx->height >> (i ? 1 : 0),
536 ret = CHECK_CU(ctx->cudl->cuMemcpy2DAsync(&cpy, device_hwctx->stream));
540 offset += avctx->height;
543 ret = CHECK_CU(ctx->cudl->cuStreamSynchronize(device_hwctx->stream));
546 } else if (avctx->pix_fmt == AV_PIX_FMT_NV12 ||
547 avctx->pix_fmt == AV_PIX_FMT_P010 ||
548 avctx->pix_fmt == AV_PIX_FMT_P016) {
549 AVFrame *tmp_frame = av_frame_alloc();
551 av_log(avctx, AV_LOG_ERROR, "av_frame_alloc failed\n");
552 ret = AVERROR(ENOMEM);
556 tmp_frame->format = AV_PIX_FMT_CUDA;
557 tmp_frame->hw_frames_ctx = av_buffer_ref(ctx->hwframe);
558 tmp_frame->data[0] = (uint8_t*)mapped_frame;
559 tmp_frame->linesize[0] = pitch;
560 tmp_frame->data[1] = (uint8_t*)(mapped_frame + avctx->height * pitch);
561 tmp_frame->linesize[1] = pitch;
562 tmp_frame->width = avctx->width;
563 tmp_frame->height = avctx->height;
565 ret = ff_get_buffer(avctx, frame, 0);
567 av_log(avctx, AV_LOG_ERROR, "ff_get_buffer failed\n");
568 av_frame_free(&tmp_frame);
572 ret = av_hwframe_transfer_data(frame, tmp_frame, 0);
574 av_log(avctx, AV_LOG_ERROR, "av_hwframe_transfer_data failed\n");
575 av_frame_free(&tmp_frame);
578 av_frame_free(&tmp_frame);
584 frame->key_frame = ctx->key_frame[parsed_frame.dispinfo.picture_index];
585 frame->width = avctx->width;
586 frame->height = avctx->height;
587 if (avctx->pkt_timebase.num && avctx->pkt_timebase.den)
588 frame->pts = av_rescale_q(parsed_frame.dispinfo.timestamp, (AVRational){1, 10000000}, avctx->pkt_timebase);
590 frame->pts = parsed_frame.dispinfo.timestamp;
592 if (parsed_frame.second_field) {
593 if (ctx->prev_pts == INT64_MIN) {
594 ctx->prev_pts = frame->pts;
595 frame->pts += (avctx->pkt_timebase.den * avctx->framerate.den) / (avctx->pkt_timebase.num * avctx->framerate.num);
597 int pts_diff = (frame->pts - ctx->prev_pts) / 2;
598 ctx->prev_pts = frame->pts;
599 frame->pts += pts_diff;
603 /* CUVIDs opaque reordering breaks the internal pkt logic.
604 * So set pkt_pts and clear all the other pkt_ fields.
607 FF_DISABLE_DEPRECATION_WARNINGS
608 frame->pkt_pts = frame->pts;
609 FF_ENABLE_DEPRECATION_WARNINGS
612 frame->pkt_duration = 0;
613 frame->pkt_size = -1;
615 frame->interlaced_frame = !parsed_frame.is_deinterlacing && !parsed_frame.dispinfo.progressive_frame;
617 if (frame->interlaced_frame)
618 frame->top_field_first = parsed_frame.dispinfo.top_field_first;
619 } else if (ctx->decoder_flushing) {
622 ret = AVERROR(EAGAIN);
627 eret = CHECK_CU(ctx->cvdl->cuvidUnmapVideoFrame(ctx->cudecoder, mapped_frame));
629 eret = CHECK_CU(ctx->cudl->cuCtxPopCurrent(&dummy));
637 static int cuvid_decode_frame(AVCodecContext *avctx, void *data, int *got_frame, AVPacket *avpkt)
639 CuvidContext *ctx = avctx->priv_data;
640 AVFrame *frame = data;
643 av_log(avctx, AV_LOG_TRACE, "cuvid_decode_frame\n");
645 if (ctx->deint_mode_current != cudaVideoDeinterlaceMode_Weave) {
646 av_log(avctx, AV_LOG_ERROR, "Deinterlacing is not supported via the old API\n");
647 return AVERROR(EINVAL);
650 if (!ctx->decoder_flushing) {
651 ret = cuvid_decode_packet(avctx, avpkt);
656 ret = cuvid_output_frame(avctx, frame);
657 if (ret == AVERROR(EAGAIN) || ret == AVERROR_EOF) {
659 } else if (ret < 0) {
668 static av_cold int cuvid_decode_end(AVCodecContext *avctx)
670 CuvidContext *ctx = avctx->priv_data;
672 av_fifo_freep(&ctx->frame_queue);
675 av_bsf_free(&ctx->bsf);
678 ctx->cvdl->cuvidDestroyVideoParser(ctx->cuparser);
681 ctx->cvdl->cuvidDestroyDecoder(ctx->cudecoder);
685 av_buffer_unref(&ctx->hwframe);
686 av_buffer_unref(&ctx->hwdevice);
688 av_freep(&ctx->key_frame);
690 cuvid_free_functions(&ctx->cvdl);
695 static int cuvid_test_capabilities(AVCodecContext *avctx,
696 const CUVIDPARSERPARAMS *cuparseinfo,
701 CuvidContext *ctx = avctx->priv_data;
702 CUVIDDECODECAPS *caps;
703 int res8 = 0, res10 = 0, res12 = 0;
705 if (!ctx->cvdl->cuvidGetDecoderCaps) {
706 av_log(avctx, AV_LOG_WARNING, "Used Nvidia driver is too old to perform a capability check.\n");
707 av_log(avctx, AV_LOG_WARNING, "The minimum required version is "
708 #if defined(_WIN32) || defined(__CYGWIN__)
713 ". Continuing blind.\n");
714 ctx->caps8.bIsSupported = ctx->caps10.bIsSupported = 1;
715 // 12 bit was not supported before the capability check was introduced, so disable it.
716 ctx->caps12.bIsSupported = 0;
720 ctx->caps8.eCodecType = ctx->caps10.eCodecType = ctx->caps12.eCodecType
721 = cuparseinfo->CodecType;
722 ctx->caps8.eChromaFormat = ctx->caps10.eChromaFormat = ctx->caps12.eChromaFormat
723 = cudaVideoChromaFormat_420;
725 ctx->caps8.nBitDepthMinus8 = 0;
726 ctx->caps10.nBitDepthMinus8 = 2;
727 ctx->caps12.nBitDepthMinus8 = 4;
729 res8 = CHECK_CU(ctx->cvdl->cuvidGetDecoderCaps(&ctx->caps8));
730 res10 = CHECK_CU(ctx->cvdl->cuvidGetDecoderCaps(&ctx->caps10));
731 res12 = CHECK_CU(ctx->cvdl->cuvidGetDecoderCaps(&ctx->caps12));
733 av_log(avctx, AV_LOG_VERBOSE, "CUVID capabilities for %s:\n", avctx->codec->name);
734 av_log(avctx, AV_LOG_VERBOSE, "8 bit: supported: %d, min_width: %d, max_width: %d, min_height: %d, max_height: %d\n",
735 ctx->caps8.bIsSupported, ctx->caps8.nMinWidth, ctx->caps8.nMaxWidth, ctx->caps8.nMinHeight, ctx->caps8.nMaxHeight);
736 av_log(avctx, AV_LOG_VERBOSE, "10 bit: supported: %d, min_width: %d, max_width: %d, min_height: %d, max_height: %d\n",
737 ctx->caps10.bIsSupported, ctx->caps10.nMinWidth, ctx->caps10.nMaxWidth, ctx->caps10.nMinHeight, ctx->caps10.nMaxHeight);
738 av_log(avctx, AV_LOG_VERBOSE, "12 bit: supported: %d, min_width: %d, max_width: %d, min_height: %d, max_height: %d\n",
739 ctx->caps12.bIsSupported, ctx->caps12.nMinWidth, ctx->caps12.nMaxWidth, ctx->caps12.nMinHeight, ctx->caps12.nMaxHeight);
758 if (!ctx->caps8.bIsSupported) {
759 av_log(avctx, AV_LOG_ERROR, "Codec %s is not supported.\n", avctx->codec->name);
760 return AVERROR(EINVAL);
763 if (!caps->bIsSupported) {
764 av_log(avctx, AV_LOG_ERROR, "Bit depth %d is not supported.\n", bit_depth);
765 return AVERROR(EINVAL);
768 if (probed_width > caps->nMaxWidth || probed_width < caps->nMinWidth) {
769 av_log(avctx, AV_LOG_ERROR, "Video width %d not within range from %d to %d\n",
770 probed_width, caps->nMinWidth, caps->nMaxWidth);
771 return AVERROR(EINVAL);
774 if (probed_height > caps->nMaxHeight || probed_height < caps->nMinHeight) {
775 av_log(avctx, AV_LOG_ERROR, "Video height %d not within range from %d to %d\n",
776 probed_height, caps->nMinHeight, caps->nMaxHeight);
777 return AVERROR(EINVAL);
783 static av_cold int cuvid_decode_init(AVCodecContext *avctx)
785 CuvidContext *ctx = avctx->priv_data;
786 AVCUDADeviceContext *device_hwctx;
787 AVHWDeviceContext *device_ctx;
788 AVHWFramesContext *hwframe_ctx;
789 CUVIDSOURCEDATAPACKET seq_pkt;
790 CUcontext cuda_ctx = NULL;
792 const AVBitStreamFilter *bsf;
795 enum AVPixelFormat pix_fmts[3] = { AV_PIX_FMT_CUDA,
799 int probed_width = avctx->coded_width ? avctx->coded_width : 1280;
800 int probed_height = avctx->coded_height ? avctx->coded_height : 720;
801 int probed_bit_depth = 8;
803 const AVPixFmtDescriptor *probe_desc = av_pix_fmt_desc_get(avctx->pix_fmt);
804 if (probe_desc && probe_desc->nb_components)
805 probed_bit_depth = probe_desc->comp[0].depth;
807 // Accelerated transcoding scenarios with 'ffmpeg' require that the
808 // pix_fmt be set to AV_PIX_FMT_CUDA early. The sw_pix_fmt, and the
809 // pix_fmt for non-accelerated transcoding, do not need to be correct
810 // but need to be set to something. We arbitrarily pick NV12.
811 ret = ff_get_format(avctx, pix_fmts);
813 av_log(avctx, AV_LOG_ERROR, "ff_get_format failed: %d\n", ret);
816 avctx->pix_fmt = ret;
818 if (ctx->resize_expr && sscanf(ctx->resize_expr, "%dx%d",
819 &ctx->resize.width, &ctx->resize.height) != 2) {
820 av_log(avctx, AV_LOG_ERROR, "Invalid resize expressions\n");
821 ret = AVERROR(EINVAL);
825 if (ctx->crop_expr && sscanf(ctx->crop_expr, "%dx%dx%dx%d",
826 &ctx->crop.top, &ctx->crop.bottom,
827 &ctx->crop.left, &ctx->crop.right) != 4) {
828 av_log(avctx, AV_LOG_ERROR, "Invalid cropping expressions\n");
829 ret = AVERROR(EINVAL);
833 ret = cuvid_load_functions(&ctx->cvdl, avctx);
835 av_log(avctx, AV_LOG_ERROR, "Failed loading nvcuvid.\n");
839 ctx->frame_queue = av_fifo_alloc(ctx->nb_surfaces * sizeof(CuvidParsedFrame));
840 if (!ctx->frame_queue) {
841 ret = AVERROR(ENOMEM);
845 if (avctx->hw_frames_ctx) {
846 ctx->hwframe = av_buffer_ref(avctx->hw_frames_ctx);
848 ret = AVERROR(ENOMEM);
852 hwframe_ctx = (AVHWFramesContext*)ctx->hwframe->data;
854 ctx->hwdevice = av_buffer_ref(hwframe_ctx->device_ref);
855 if (!ctx->hwdevice) {
856 ret = AVERROR(ENOMEM);
860 if (avctx->hw_device_ctx) {
861 ctx->hwdevice = av_buffer_ref(avctx->hw_device_ctx);
862 if (!ctx->hwdevice) {
863 ret = AVERROR(ENOMEM);
867 ret = av_hwdevice_ctx_create(&ctx->hwdevice, AV_HWDEVICE_TYPE_CUDA, ctx->cu_gpu, NULL, 0);
872 ctx->hwframe = av_hwframe_ctx_alloc(ctx->hwdevice);
874 av_log(avctx, AV_LOG_ERROR, "av_hwframe_ctx_alloc failed\n");
875 ret = AVERROR(ENOMEM);
879 hwframe_ctx = (AVHWFramesContext*)ctx->hwframe->data;
882 device_ctx = hwframe_ctx->device_ctx;
883 device_hwctx = device_ctx->hwctx;
885 cuda_ctx = device_hwctx->cuda_ctx;
886 ctx->cudl = device_hwctx->internal->cuda_dl;
888 memset(&ctx->cuparseinfo, 0, sizeof(ctx->cuparseinfo));
889 memset(&ctx->cuparse_ext, 0, sizeof(ctx->cuparse_ext));
890 memset(&seq_pkt, 0, sizeof(seq_pkt));
892 ctx->cuparseinfo.pExtVideoInfo = &ctx->cuparse_ext;
894 switch (avctx->codec->id) {
895 #if CONFIG_H264_CUVID_DECODER
896 case AV_CODEC_ID_H264:
897 ctx->cuparseinfo.CodecType = cudaVideoCodec_H264;
900 #if CONFIG_HEVC_CUVID_DECODER
901 case AV_CODEC_ID_HEVC:
902 ctx->cuparseinfo.CodecType = cudaVideoCodec_HEVC;
905 #if CONFIG_MJPEG_CUVID_DECODER
906 case AV_CODEC_ID_MJPEG:
907 ctx->cuparseinfo.CodecType = cudaVideoCodec_JPEG;
910 #if CONFIG_MPEG1_CUVID_DECODER
911 case AV_CODEC_ID_MPEG1VIDEO:
912 ctx->cuparseinfo.CodecType = cudaVideoCodec_MPEG1;
915 #if CONFIG_MPEG2_CUVID_DECODER
916 case AV_CODEC_ID_MPEG2VIDEO:
917 ctx->cuparseinfo.CodecType = cudaVideoCodec_MPEG2;
920 #if CONFIG_MPEG4_CUVID_DECODER
921 case AV_CODEC_ID_MPEG4:
922 ctx->cuparseinfo.CodecType = cudaVideoCodec_MPEG4;
925 #if CONFIG_VP8_CUVID_DECODER
926 case AV_CODEC_ID_VP8:
927 ctx->cuparseinfo.CodecType = cudaVideoCodec_VP8;
930 #if CONFIG_VP9_CUVID_DECODER
931 case AV_CODEC_ID_VP9:
932 ctx->cuparseinfo.CodecType = cudaVideoCodec_VP9;
935 #if CONFIG_VC1_CUVID_DECODER
936 case AV_CODEC_ID_VC1:
937 ctx->cuparseinfo.CodecType = cudaVideoCodec_VC1;
941 av_log(avctx, AV_LOG_ERROR, "Invalid CUVID codec!\n");
945 if (avctx->codec->id == AV_CODEC_ID_H264 || avctx->codec->id == AV_CODEC_ID_HEVC) {
946 if (avctx->codec->id == AV_CODEC_ID_H264)
947 bsf = av_bsf_get_by_name("h264_mp4toannexb");
949 bsf = av_bsf_get_by_name("hevc_mp4toannexb");
952 ret = AVERROR_BSF_NOT_FOUND;
955 if (ret = av_bsf_alloc(bsf, &ctx->bsf)) {
958 if (((ret = avcodec_parameters_from_context(ctx->bsf->par_in, avctx)) < 0) || ((ret = av_bsf_init(ctx->bsf)) < 0)) {
959 av_bsf_free(&ctx->bsf);
963 ctx->cuparse_ext.format.seqhdr_data_length = ctx->bsf->par_out->extradata_size;
964 memcpy(ctx->cuparse_ext.raw_seqhdr_data,
965 ctx->bsf->par_out->extradata,
966 FFMIN(sizeof(ctx->cuparse_ext.raw_seqhdr_data), ctx->bsf->par_out->extradata_size));
967 } else if (avctx->extradata_size > 0) {
968 ctx->cuparse_ext.format.seqhdr_data_length = avctx->extradata_size;
969 memcpy(ctx->cuparse_ext.raw_seqhdr_data,
971 FFMIN(sizeof(ctx->cuparse_ext.raw_seqhdr_data), avctx->extradata_size));
974 ctx->key_frame = av_mallocz(ctx->nb_surfaces * sizeof(int));
975 if (!ctx->key_frame) {
976 ret = AVERROR(ENOMEM);
980 ctx->cuparseinfo.ulMaxNumDecodeSurfaces = ctx->nb_surfaces;
981 ctx->cuparseinfo.ulMaxDisplayDelay = 4;
982 ctx->cuparseinfo.pUserData = avctx;
983 ctx->cuparseinfo.pfnSequenceCallback = cuvid_handle_video_sequence;
984 ctx->cuparseinfo.pfnDecodePicture = cuvid_handle_picture_decode;
985 ctx->cuparseinfo.pfnDisplayPicture = cuvid_handle_picture_display;
987 ret = CHECK_CU(ctx->cudl->cuCtxPushCurrent(cuda_ctx));
991 ret = cuvid_test_capabilities(avctx, &ctx->cuparseinfo,
998 ret = CHECK_CU(ctx->cvdl->cuvidCreateVideoParser(&ctx->cuparser, &ctx->cuparseinfo));
1002 seq_pkt.payload = ctx->cuparse_ext.raw_seqhdr_data;
1003 seq_pkt.payload_size = ctx->cuparse_ext.format.seqhdr_data_length;
1005 if (seq_pkt.payload && seq_pkt.payload_size) {
1006 ret = CHECK_CU(ctx->cvdl->cuvidParseVideoData(ctx->cuparser, &seq_pkt));
1011 ret = CHECK_CU(ctx->cudl->cuCtxPopCurrent(&dummy));
1015 ctx->prev_pts = INT64_MIN;
1017 if (!avctx->pkt_timebase.num || !avctx->pkt_timebase.den)
1018 av_log(avctx, AV_LOG_WARNING, "Invalid pkt_timebase, passing timestamps as-is.\n");
1023 cuvid_decode_end(avctx);
1027 static void cuvid_flush(AVCodecContext *avctx)
1029 CuvidContext *ctx = avctx->priv_data;
1030 AVHWDeviceContext *device_ctx = (AVHWDeviceContext*)ctx->hwdevice->data;
1031 AVCUDADeviceContext *device_hwctx = device_ctx->hwctx;
1032 CUcontext dummy, cuda_ctx = device_hwctx->cuda_ctx;
1033 CUVIDSOURCEDATAPACKET seq_pkt = { 0 };
1036 ret = CHECK_CU(ctx->cudl->cuCtxPushCurrent(cuda_ctx));
1040 av_fifo_freep(&ctx->frame_queue);
1042 ctx->frame_queue = av_fifo_alloc(ctx->nb_surfaces * sizeof(CuvidParsedFrame));
1043 if (!ctx->frame_queue) {
1044 av_log(avctx, AV_LOG_ERROR, "Failed to recreate frame queue on flush\n");
1048 if (ctx->cudecoder) {
1049 ctx->cvdl->cuvidDestroyDecoder(ctx->cudecoder);
1050 ctx->cudecoder = NULL;
1053 if (ctx->cuparser) {
1054 ctx->cvdl->cuvidDestroyVideoParser(ctx->cuparser);
1055 ctx->cuparser = NULL;
1058 ret = CHECK_CU(ctx->cvdl->cuvidCreateVideoParser(&ctx->cuparser, &ctx->cuparseinfo));
1062 seq_pkt.payload = ctx->cuparse_ext.raw_seqhdr_data;
1063 seq_pkt.payload_size = ctx->cuparse_ext.format.seqhdr_data_length;
1065 if (seq_pkt.payload && seq_pkt.payload_size) {
1066 ret = CHECK_CU(ctx->cvdl->cuvidParseVideoData(ctx->cuparser, &seq_pkt));
1071 ret = CHECK_CU(ctx->cudl->cuCtxPopCurrent(&dummy));
1075 ctx->prev_pts = INT64_MIN;
1076 ctx->decoder_flushing = 0;
1080 av_log(avctx, AV_LOG_ERROR, "CUDA reinit on flush failed\n");
1083 #define OFFSET(x) offsetof(CuvidContext, x)
1084 #define VD AV_OPT_FLAG_VIDEO_PARAM | AV_OPT_FLAG_DECODING_PARAM
1085 static const AVOption options[] = {
1086 { "deint", "Set deinterlacing mode", OFFSET(deint_mode), AV_OPT_TYPE_INT, { .i64 = cudaVideoDeinterlaceMode_Weave }, cudaVideoDeinterlaceMode_Weave, cudaVideoDeinterlaceMode_Adaptive, VD, "deint" },
1087 { "weave", "Weave deinterlacing (do nothing)", 0, AV_OPT_TYPE_CONST, { .i64 = cudaVideoDeinterlaceMode_Weave }, 0, 0, VD, "deint" },
1088 { "bob", "Bob deinterlacing", 0, AV_OPT_TYPE_CONST, { .i64 = cudaVideoDeinterlaceMode_Bob }, 0, 0, VD, "deint" },
1089 { "adaptive", "Adaptive deinterlacing", 0, AV_OPT_TYPE_CONST, { .i64 = cudaVideoDeinterlaceMode_Adaptive }, 0, 0, VD, "deint" },
1090 { "gpu", "GPU to be used for decoding", OFFSET(cu_gpu), AV_OPT_TYPE_STRING, { .str = NULL }, 0, 0, VD },
1091 { "surfaces", "Maximum surfaces to be used for decoding", OFFSET(nb_surfaces), AV_OPT_TYPE_INT, { .i64 = 25 }, 0, INT_MAX, VD },
1092 { "drop_second_field", "Drop second field when deinterlacing", OFFSET(drop_second_field), AV_OPT_TYPE_BOOL, { .i64 = 0 }, 0, 1, VD },
1093 { "crop", "Crop (top)x(bottom)x(left)x(right)", OFFSET(crop_expr), AV_OPT_TYPE_STRING, { .str = NULL }, 0, 0, VD },
1094 { "resize", "Resize (width)x(height)", OFFSET(resize_expr), AV_OPT_TYPE_STRING, { .str = NULL }, 0, 0, VD },
1098 static const AVCodecHWConfigInternal *cuvid_hw_configs[] = {
1099 &(const AVCodecHWConfigInternal) {
1101 .pix_fmt = AV_PIX_FMT_CUDA,
1102 .methods = AV_CODEC_HW_CONFIG_METHOD_HW_DEVICE_CTX |
1103 AV_CODEC_HW_CONFIG_METHOD_INTERNAL,
1104 .device_type = AV_HWDEVICE_TYPE_CUDA
1111 #define DEFINE_CUVID_CODEC(x, X) \
1112 static const AVClass x##_cuvid_class = { \
1113 .class_name = #x "_cuvid", \
1114 .item_name = av_default_item_name, \
1115 .option = options, \
1116 .version = LIBAVUTIL_VERSION_INT, \
1118 AVCodec ff_##x##_cuvid_decoder = { \
1119 .name = #x "_cuvid", \
1120 .long_name = NULL_IF_CONFIG_SMALL("Nvidia CUVID " #X " decoder"), \
1121 .type = AVMEDIA_TYPE_VIDEO, \
1122 .id = AV_CODEC_ID_##X, \
1123 .priv_data_size = sizeof(CuvidContext), \
1124 .priv_class = &x##_cuvid_class, \
1125 .init = cuvid_decode_init, \
1126 .close = cuvid_decode_end, \
1127 .decode = cuvid_decode_frame, \
1128 .receive_frame = cuvid_output_frame, \
1129 .flush = cuvid_flush, \
1130 .capabilities = AV_CODEC_CAP_DELAY | AV_CODEC_CAP_AVOID_PROBING | AV_CODEC_CAP_HARDWARE, \
1131 .pix_fmts = (const enum AVPixelFormat[]){ AV_PIX_FMT_CUDA, \
1135 AV_PIX_FMT_NONE }, \
1136 .hw_configs = cuvid_hw_configs, \
1137 .wrapper_name = "cuvid", \
1140 #if CONFIG_HEVC_CUVID_DECODER
1141 DEFINE_CUVID_CODEC(hevc, HEVC)
1144 #if CONFIG_H264_CUVID_DECODER
1145 DEFINE_CUVID_CODEC(h264, H264)
1148 #if CONFIG_MJPEG_CUVID_DECODER
1149 DEFINE_CUVID_CODEC(mjpeg, MJPEG)
1152 #if CONFIG_MPEG1_CUVID_DECODER
1153 DEFINE_CUVID_CODEC(mpeg1, MPEG1VIDEO)
1156 #if CONFIG_MPEG2_CUVID_DECODER
1157 DEFINE_CUVID_CODEC(mpeg2, MPEG2VIDEO)
1160 #if CONFIG_MPEG4_CUVID_DECODER
1161 DEFINE_CUVID_CODEC(mpeg4, MPEG4)
1164 #if CONFIG_VP8_CUVID_DECODER
1165 DEFINE_CUVID_CODEC(vp8, VP8)
1168 #if CONFIG_VP9_CUVID_DECODER
1169 DEFINE_CUVID_CODEC(vp9, VP9)
1172 #if CONFIG_VC1_CUVID_DECODER
1173 DEFINE_CUVID_CODEC(vc1, VC1)