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/fifo.h"
29 #include "libavutil/log.h"
30 #include "libavutil/opt.h"
31 #include "libavutil/pixdesc.h"
38 typedef struct CuvidContext
42 CUvideodecoder cudecoder;
43 CUvideoparser cuparser;
47 int drop_second_field;
63 AVBufferRef *hwdevice;
68 AVFifoBuffer *frame_queue;
71 int deint_mode_current;
77 cudaVideoCodec codec_type;
78 cudaVideoChromaFormat chroma_format;
80 CUVIDDECODECAPS caps8, caps10, caps12;
82 CUVIDPARSERPARAMS cuparseinfo;
83 CUVIDEOFORMATEX cuparse_ext;
89 typedef struct CuvidParsedFrame
91 CUVIDPARSERDISPINFO dispinfo;
96 static int check_cu(AVCodecContext *avctx, CUresult err, const char *func)
98 CuvidContext *ctx = avctx->priv_data;
100 const char *err_string;
102 av_log(avctx, AV_LOG_TRACE, "Calling %s\n", func);
104 if (err == CUDA_SUCCESS)
107 ctx->cudl->cuGetErrorName(err, &err_name);
108 ctx->cudl->cuGetErrorString(err, &err_string);
110 av_log(avctx, AV_LOG_ERROR, "%s failed", func);
111 if (err_name && err_string)
112 av_log(avctx, AV_LOG_ERROR, " -> %s: %s", err_name, err_string);
113 av_log(avctx, AV_LOG_ERROR, "\n");
115 return AVERROR_EXTERNAL;
118 #define CHECK_CU(x) check_cu(avctx, (x), #x)
120 static int CUDAAPI cuvid_handle_video_sequence(void *opaque, CUVIDEOFORMAT* format)
122 AVCodecContext *avctx = opaque;
123 CuvidContext *ctx = avctx->priv_data;
124 AVHWFramesContext *hwframe_ctx = (AVHWFramesContext*)ctx->hwframe->data;
125 CUVIDDECODECAPS *caps = NULL;
126 CUVIDDECODECREATEINFO cuinfo;
129 int old_width = avctx->width;
130 int old_height = avctx->height;
132 enum AVPixelFormat pix_fmts[3] = { AV_PIX_FMT_CUDA,
133 AV_PIX_FMT_NONE, // Will be updated below
136 av_log(avctx, AV_LOG_TRACE, "pfnSequenceCallback, progressive_sequence=%d\n", format->progressive_sequence);
138 memset(&cuinfo, 0, sizeof(cuinfo));
140 ctx->internal_error = 0;
142 avctx->coded_width = cuinfo.ulWidth = format->coded_width;
143 avctx->coded_height = cuinfo.ulHeight = format->coded_height;
146 cuinfo.display_area.left = format->display_area.left + ctx->crop.left;
147 cuinfo.display_area.top = format->display_area.top + ctx->crop.top;
148 cuinfo.display_area.right = format->display_area.right - ctx->crop.right;
149 cuinfo.display_area.bottom = format->display_area.bottom - ctx->crop.bottom;
151 // width and height need to be set before calling ff_get_format
152 if (ctx->resize_expr) {
153 avctx->width = ctx->resize.width;
154 avctx->height = ctx->resize.height;
156 avctx->width = cuinfo.display_area.right - cuinfo.display_area.left;
157 avctx->height = cuinfo.display_area.bottom - cuinfo.display_area.top;
160 // target width/height need to be multiples of two
161 cuinfo.ulTargetWidth = avctx->width = (avctx->width + 1) & ~1;
162 cuinfo.ulTargetHeight = avctx->height = (avctx->height + 1) & ~1;
164 // aspect ratio conversion, 1:1, depends on scaled resolution
165 cuinfo.target_rect.left = 0;
166 cuinfo.target_rect.top = 0;
167 cuinfo.target_rect.right = cuinfo.ulTargetWidth;
168 cuinfo.target_rect.bottom = cuinfo.ulTargetHeight;
170 switch (format->bit_depth_luma_minus8) {
172 pix_fmts[1] = AV_PIX_FMT_NV12;
176 pix_fmts[1] = AV_PIX_FMT_P010;
180 pix_fmts[1] = AV_PIX_FMT_P016;
187 if (!caps || !caps->bIsSupported) {
188 av_log(avctx, AV_LOG_ERROR, "unsupported bit depth: %d\n",
189 format->bit_depth_luma_minus8 + 8);
190 ctx->internal_error = AVERROR(EINVAL);
194 surface_fmt = ff_get_format(avctx, pix_fmts);
195 if (surface_fmt < 0) {
196 av_log(avctx, AV_LOG_ERROR, "ff_get_format failed: %d\n", surface_fmt);
197 ctx->internal_error = AVERROR(EINVAL);
201 av_log(avctx, AV_LOG_VERBOSE, "Formats: Original: %s | HW: %s | SW: %s\n",
202 av_get_pix_fmt_name(avctx->pix_fmt),
203 av_get_pix_fmt_name(surface_fmt),
204 av_get_pix_fmt_name(avctx->sw_pix_fmt));
206 avctx->pix_fmt = surface_fmt;
208 // Update our hwframe ctx, as the get_format callback might have refreshed it!
209 if (avctx->hw_frames_ctx) {
210 av_buffer_unref(&ctx->hwframe);
212 ctx->hwframe = av_buffer_ref(avctx->hw_frames_ctx);
214 ctx->internal_error = AVERROR(ENOMEM);
218 hwframe_ctx = (AVHWFramesContext*)ctx->hwframe->data;
221 ff_set_sar(avctx, av_div_q(
222 (AVRational){ format->display_aspect_ratio.x, format->display_aspect_ratio.y },
223 (AVRational){ avctx->width, avctx->height }));
225 ctx->deint_mode_current = format->progressive_sequence
226 ? cudaVideoDeinterlaceMode_Weave
229 if (!format->progressive_sequence && ctx->deint_mode_current == cudaVideoDeinterlaceMode_Weave)
230 avctx->flags |= AV_CODEC_FLAG_INTERLACED_DCT;
232 avctx->flags &= ~AV_CODEC_FLAG_INTERLACED_DCT;
234 if (format->video_signal_description.video_full_range_flag)
235 avctx->color_range = AVCOL_RANGE_JPEG;
237 avctx->color_range = AVCOL_RANGE_MPEG;
239 avctx->color_primaries = format->video_signal_description.color_primaries;
240 avctx->color_trc = format->video_signal_description.transfer_characteristics;
241 avctx->colorspace = format->video_signal_description.matrix_coefficients;
244 avctx->bit_rate = format->bitrate;
246 if (format->frame_rate.numerator && format->frame_rate.denominator) {
247 avctx->framerate.num = format->frame_rate.numerator;
248 avctx->framerate.den = format->frame_rate.denominator;
252 && avctx->coded_width == format->coded_width
253 && avctx->coded_height == format->coded_height
254 && avctx->width == old_width
255 && avctx->height == old_height
256 && ctx->chroma_format == format->chroma_format
257 && ctx->codec_type == format->codec)
260 if (ctx->cudecoder) {
261 av_log(avctx, AV_LOG_TRACE, "Re-initializing decoder\n");
262 ctx->internal_error = CHECK_CU(ctx->cvdl->cuvidDestroyDecoder(ctx->cudecoder));
263 if (ctx->internal_error < 0)
265 ctx->cudecoder = NULL;
268 if (hwframe_ctx->pool && (
269 hwframe_ctx->width < avctx->width ||
270 hwframe_ctx->height < avctx->height ||
271 hwframe_ctx->format != AV_PIX_FMT_CUDA ||
272 hwframe_ctx->sw_format != avctx->sw_pix_fmt)) {
273 av_log(avctx, AV_LOG_ERROR, "AVHWFramesContext is already initialized with incompatible parameters\n");
274 av_log(avctx, AV_LOG_DEBUG, "width: %d <-> %d\n", hwframe_ctx->width, avctx->width);
275 av_log(avctx, AV_LOG_DEBUG, "height: %d <-> %d\n", hwframe_ctx->height, avctx->height);
276 av_log(avctx, AV_LOG_DEBUG, "format: %s <-> cuda\n", av_get_pix_fmt_name(hwframe_ctx->format));
277 av_log(avctx, AV_LOG_DEBUG, "sw_format: %s <-> %s\n",
278 av_get_pix_fmt_name(hwframe_ctx->sw_format), av_get_pix_fmt_name(avctx->sw_pix_fmt));
279 ctx->internal_error = AVERROR(EINVAL);
283 if (format->chroma_format != cudaVideoChromaFormat_420) {
284 av_log(avctx, AV_LOG_ERROR, "Chroma formats other than 420 are not supported\n");
285 ctx->internal_error = AVERROR(EINVAL);
289 ctx->chroma_format = format->chroma_format;
291 cuinfo.CodecType = ctx->codec_type = format->codec;
292 cuinfo.ChromaFormat = format->chroma_format;
294 switch (avctx->sw_pix_fmt) {
295 case AV_PIX_FMT_NV12:
296 cuinfo.OutputFormat = cudaVideoSurfaceFormat_NV12;
298 case AV_PIX_FMT_P010:
299 case AV_PIX_FMT_P016:
300 cuinfo.OutputFormat = cudaVideoSurfaceFormat_P016;
303 av_log(avctx, AV_LOG_ERROR, "Output formats other than NV12, P010 or P016 are not supported\n");
304 ctx->internal_error = AVERROR(EINVAL);
308 cuinfo.ulNumDecodeSurfaces = ctx->nb_surfaces;
309 cuinfo.ulNumOutputSurfaces = 1;
310 cuinfo.ulCreationFlags = cudaVideoCreate_PreferCUVID;
311 cuinfo.bitDepthMinus8 = format->bit_depth_luma_minus8;
312 cuinfo.DeinterlaceMode = ctx->deint_mode_current;
314 if (ctx->deint_mode_current != cudaVideoDeinterlaceMode_Weave && !ctx->drop_second_field)
315 avctx->framerate = av_mul_q(avctx->framerate, (AVRational){2, 1});
317 ctx->internal_error = CHECK_CU(ctx->cvdl->cuvidCreateDecoder(&ctx->cudecoder, &cuinfo));
318 if (ctx->internal_error < 0)
321 if (!hwframe_ctx->pool) {
322 hwframe_ctx->format = AV_PIX_FMT_CUDA;
323 hwframe_ctx->sw_format = avctx->sw_pix_fmt;
324 hwframe_ctx->width = avctx->width;
325 hwframe_ctx->height = avctx->height;
327 if ((ctx->internal_error = av_hwframe_ctx_init(ctx->hwframe)) < 0) {
328 av_log(avctx, AV_LOG_ERROR, "av_hwframe_ctx_init failed\n");
336 static int CUDAAPI cuvid_handle_picture_decode(void *opaque, CUVIDPICPARAMS* picparams)
338 AVCodecContext *avctx = opaque;
339 CuvidContext *ctx = avctx->priv_data;
341 av_log(avctx, AV_LOG_TRACE, "pfnDecodePicture\n");
343 ctx->internal_error = CHECK_CU(ctx->cvdl->cuvidDecodePicture(ctx->cudecoder, picparams));
344 if (ctx->internal_error < 0)
350 static int CUDAAPI cuvid_handle_picture_display(void *opaque, CUVIDPARSERDISPINFO* dispinfo)
352 AVCodecContext *avctx = opaque;
353 CuvidContext *ctx = avctx->priv_data;
354 CuvidParsedFrame parsed_frame = { { 0 } };
356 parsed_frame.dispinfo = *dispinfo;
357 ctx->internal_error = 0;
359 if (ctx->deint_mode_current == cudaVideoDeinterlaceMode_Weave) {
360 av_fifo_generic_write(ctx->frame_queue, &parsed_frame, sizeof(CuvidParsedFrame), NULL);
362 parsed_frame.is_deinterlacing = 1;
363 av_fifo_generic_write(ctx->frame_queue, &parsed_frame, sizeof(CuvidParsedFrame), NULL);
364 if (!ctx->drop_second_field) {
365 parsed_frame.second_field = 1;
366 av_fifo_generic_write(ctx->frame_queue, &parsed_frame, sizeof(CuvidParsedFrame), NULL);
373 static int cuvid_is_buffer_full(AVCodecContext *avctx)
375 CuvidContext *ctx = avctx->priv_data;
377 return (av_fifo_size(ctx->frame_queue) / sizeof(CuvidParsedFrame)) + 2 > ctx->nb_surfaces;
380 static int cuvid_decode_packet(AVCodecContext *avctx, const AVPacket *avpkt)
382 CuvidContext *ctx = avctx->priv_data;
383 AVHWDeviceContext *device_ctx = (AVHWDeviceContext*)ctx->hwdevice->data;
384 AVCUDADeviceContext *device_hwctx = device_ctx->hwctx;
385 CUcontext dummy, cuda_ctx = device_hwctx->cuda_ctx;
386 CUVIDSOURCEDATAPACKET cupkt;
387 AVPacket filter_packet = { 0 };
388 AVPacket filtered_packet = { 0 };
389 int ret = 0, eret = 0, is_flush = ctx->decoder_flushing;
391 av_log(avctx, AV_LOG_TRACE, "cuvid_decode_packet\n");
393 if (is_flush && avpkt && avpkt->size)
396 if (cuvid_is_buffer_full(avctx) && avpkt && avpkt->size)
397 return AVERROR(EAGAIN);
399 if (ctx->bsf && avpkt && avpkt->size) {
400 if ((ret = av_packet_ref(&filter_packet, avpkt)) < 0) {
401 av_log(avctx, AV_LOG_ERROR, "av_packet_ref failed\n");
405 if ((ret = av_bsf_send_packet(ctx->bsf, &filter_packet)) < 0) {
406 av_log(avctx, AV_LOG_ERROR, "av_bsf_send_packet failed\n");
407 av_packet_unref(&filter_packet);
411 if ((ret = av_bsf_receive_packet(ctx->bsf, &filtered_packet)) < 0) {
412 av_log(avctx, AV_LOG_ERROR, "av_bsf_receive_packet failed\n");
416 avpkt = &filtered_packet;
419 ret = CHECK_CU(ctx->cudl->cuCtxPushCurrent(cuda_ctx));
421 av_packet_unref(&filtered_packet);
425 memset(&cupkt, 0, sizeof(cupkt));
427 if (avpkt && avpkt->size) {
428 cupkt.payload_size = avpkt->size;
429 cupkt.payload = avpkt->data;
431 if (avpkt->pts != AV_NOPTS_VALUE) {
432 cupkt.flags = CUVID_PKT_TIMESTAMP;
433 if (avctx->pkt_timebase.num && avctx->pkt_timebase.den)
434 cupkt.timestamp = av_rescale_q(avpkt->pts, avctx->pkt_timebase, (AVRational){1, 10000000});
436 cupkt.timestamp = avpkt->pts;
439 cupkt.flags = CUVID_PKT_ENDOFSTREAM;
440 ctx->decoder_flushing = 1;
443 ret = CHECK_CU(ctx->cvdl->cuvidParseVideoData(ctx->cuparser, &cupkt));
445 av_packet_unref(&filtered_packet);
450 // cuvidParseVideoData doesn't return an error just because stuff failed...
451 if (ctx->internal_error) {
452 av_log(avctx, AV_LOG_ERROR, "cuvid decode callback error\n");
453 ret = ctx->internal_error;
458 eret = CHECK_CU(ctx->cudl->cuCtxPopCurrent(&dummy));
470 static int cuvid_output_frame(AVCodecContext *avctx, AVFrame *frame)
472 CuvidContext *ctx = avctx->priv_data;
473 AVHWDeviceContext *device_ctx = (AVHWDeviceContext*)ctx->hwdevice->data;
474 AVCUDADeviceContext *device_hwctx = device_ctx->hwctx;
475 CUcontext dummy, cuda_ctx = device_hwctx->cuda_ctx;
476 CUdeviceptr mapped_frame = 0;
477 int ret = 0, eret = 0;
479 av_log(avctx, AV_LOG_TRACE, "cuvid_output_frame\n");
481 if (ctx->decoder_flushing) {
482 ret = cuvid_decode_packet(avctx, NULL);
483 if (ret < 0 && ret != AVERROR_EOF)
487 if (!cuvid_is_buffer_full(avctx)) {
489 ret = ff_decode_get_packet(avctx, &pkt);
490 if (ret < 0 && ret != AVERROR_EOF)
492 ret = cuvid_decode_packet(avctx, &pkt);
493 av_packet_unref(&pkt);
494 // cuvid_is_buffer_full() should avoid this.
495 if (ret == AVERROR(EAGAIN))
496 ret = AVERROR_EXTERNAL;
497 if (ret < 0 && ret != AVERROR_EOF)
501 ret = CHECK_CU(ctx->cudl->cuCtxPushCurrent(cuda_ctx));
505 if (av_fifo_size(ctx->frame_queue)) {
506 CuvidParsedFrame parsed_frame;
507 CUVIDPROCPARAMS params;
508 unsigned int pitch = 0;
512 av_fifo_generic_read(ctx->frame_queue, &parsed_frame, sizeof(CuvidParsedFrame), NULL);
514 memset(¶ms, 0, sizeof(params));
515 params.progressive_frame = parsed_frame.dispinfo.progressive_frame;
516 params.second_field = parsed_frame.second_field;
517 params.top_field_first = parsed_frame.dispinfo.top_field_first;
519 ret = CHECK_CU(ctx->cvdl->cuvidMapVideoFrame(ctx->cudecoder, parsed_frame.dispinfo.picture_index, &mapped_frame, &pitch, ¶ms));
523 if (avctx->pix_fmt == AV_PIX_FMT_CUDA) {
524 ret = av_hwframe_get_buffer(ctx->hwframe, frame, 0);
526 av_log(avctx, AV_LOG_ERROR, "av_hwframe_get_buffer failed\n");
530 ret = ff_decode_frame_props(avctx, frame);
532 av_log(avctx, AV_LOG_ERROR, "ff_decode_frame_props failed\n");
536 for (i = 0; i < 2; i++) {
537 CUDA_MEMCPY2D cpy = {
538 .srcMemoryType = CU_MEMORYTYPE_DEVICE,
539 .dstMemoryType = CU_MEMORYTYPE_DEVICE,
540 .srcDevice = mapped_frame,
541 .dstDevice = (CUdeviceptr)frame->data[i],
543 .dstPitch = frame->linesize[i],
545 .WidthInBytes = FFMIN(pitch, frame->linesize[i]),
546 .Height = avctx->height >> (i ? 1 : 0),
549 ret = CHECK_CU(ctx->cudl->cuMemcpy2D(&cpy));
553 offset += avctx->height;
555 } else if (avctx->pix_fmt == AV_PIX_FMT_NV12 ||
556 avctx->pix_fmt == AV_PIX_FMT_P010 ||
557 avctx->pix_fmt == AV_PIX_FMT_P016) {
558 AVFrame *tmp_frame = av_frame_alloc();
560 av_log(avctx, AV_LOG_ERROR, "av_frame_alloc failed\n");
561 ret = AVERROR(ENOMEM);
565 tmp_frame->format = AV_PIX_FMT_CUDA;
566 tmp_frame->hw_frames_ctx = av_buffer_ref(ctx->hwframe);
567 tmp_frame->data[0] = (uint8_t*)mapped_frame;
568 tmp_frame->linesize[0] = pitch;
569 tmp_frame->data[1] = (uint8_t*)(mapped_frame + avctx->height * pitch);
570 tmp_frame->linesize[1] = pitch;
571 tmp_frame->width = avctx->width;
572 tmp_frame->height = avctx->height;
574 ret = ff_get_buffer(avctx, frame, 0);
576 av_log(avctx, AV_LOG_ERROR, "ff_get_buffer failed\n");
577 av_frame_free(&tmp_frame);
581 ret = av_hwframe_transfer_data(frame, tmp_frame, 0);
583 av_log(avctx, AV_LOG_ERROR, "av_hwframe_transfer_data failed\n");
584 av_frame_free(&tmp_frame);
587 av_frame_free(&tmp_frame);
593 frame->width = avctx->width;
594 frame->height = avctx->height;
595 if (avctx->pkt_timebase.num && avctx->pkt_timebase.den)
596 frame->pts = av_rescale_q(parsed_frame.dispinfo.timestamp, (AVRational){1, 10000000}, avctx->pkt_timebase);
598 frame->pts = parsed_frame.dispinfo.timestamp;
600 if (parsed_frame.second_field) {
601 if (ctx->prev_pts == INT64_MIN) {
602 ctx->prev_pts = frame->pts;
603 frame->pts += (avctx->pkt_timebase.den * avctx->framerate.den) / (avctx->pkt_timebase.num * avctx->framerate.num);
605 int pts_diff = (frame->pts - ctx->prev_pts) / 2;
606 ctx->prev_pts = frame->pts;
607 frame->pts += pts_diff;
611 /* CUVIDs opaque reordering breaks the internal pkt logic.
612 * So set pkt_pts and clear all the other pkt_ fields.
615 FF_DISABLE_DEPRECATION_WARNINGS
616 frame->pkt_pts = frame->pts;
617 FF_ENABLE_DEPRECATION_WARNINGS
620 frame->pkt_duration = 0;
621 frame->pkt_size = -1;
623 frame->interlaced_frame = !parsed_frame.is_deinterlacing && !parsed_frame.dispinfo.progressive_frame;
625 if (frame->interlaced_frame)
626 frame->top_field_first = parsed_frame.dispinfo.top_field_first;
627 } else if (ctx->decoder_flushing) {
630 ret = AVERROR(EAGAIN);
635 eret = CHECK_CU(ctx->cvdl->cuvidUnmapVideoFrame(ctx->cudecoder, mapped_frame));
637 eret = CHECK_CU(ctx->cudl->cuCtxPopCurrent(&dummy));
645 static int cuvid_decode_frame(AVCodecContext *avctx, void *data, int *got_frame, AVPacket *avpkt)
647 CuvidContext *ctx = avctx->priv_data;
648 AVFrame *frame = data;
651 av_log(avctx, AV_LOG_TRACE, "cuvid_decode_frame\n");
653 if (ctx->deint_mode_current != cudaVideoDeinterlaceMode_Weave) {
654 av_log(avctx, AV_LOG_ERROR, "Deinterlacing is not supported via the old API\n");
655 return AVERROR(EINVAL);
658 if (!ctx->decoder_flushing) {
659 ret = cuvid_decode_packet(avctx, avpkt);
664 ret = cuvid_output_frame(avctx, frame);
665 if (ret == AVERROR(EAGAIN) || ret == AVERROR_EOF) {
667 } else if (ret < 0) {
676 static av_cold int cuvid_decode_end(AVCodecContext *avctx)
678 CuvidContext *ctx = avctx->priv_data;
680 av_fifo_freep(&ctx->frame_queue);
683 av_bsf_free(&ctx->bsf);
686 ctx->cvdl->cuvidDestroyVideoParser(ctx->cuparser);
689 ctx->cvdl->cuvidDestroyDecoder(ctx->cudecoder);
693 av_buffer_unref(&ctx->hwframe);
694 av_buffer_unref(&ctx->hwdevice);
696 cuvid_free_functions(&ctx->cvdl);
701 static int cuvid_test_capabilities(AVCodecContext *avctx,
702 const CUVIDPARSERPARAMS *cuparseinfo,
707 CuvidContext *ctx = avctx->priv_data;
708 CUVIDDECODECAPS *caps;
709 int res8 = 0, res10 = 0, res12 = 0;
711 if (!ctx->cvdl->cuvidGetDecoderCaps) {
712 av_log(avctx, AV_LOG_WARNING, "Used Nvidia driver is too old to perform a capability check.\n");
713 av_log(avctx, AV_LOG_WARNING, "The minimum required version is "
714 #if defined(_WIN32) || defined(__CYGWIN__)
719 ". Continuing blind.\n");
720 ctx->caps8.bIsSupported = ctx->caps10.bIsSupported = 1;
721 // 12 bit was not supported before the capability check was introduced, so disable it.
722 ctx->caps12.bIsSupported = 0;
726 ctx->caps8.eCodecType = ctx->caps10.eCodecType = ctx->caps12.eCodecType
727 = cuparseinfo->CodecType;
728 ctx->caps8.eChromaFormat = ctx->caps10.eChromaFormat = ctx->caps12.eChromaFormat
729 = cudaVideoChromaFormat_420;
731 ctx->caps8.nBitDepthMinus8 = 0;
732 ctx->caps10.nBitDepthMinus8 = 2;
733 ctx->caps12.nBitDepthMinus8 = 4;
735 res8 = CHECK_CU(ctx->cvdl->cuvidGetDecoderCaps(&ctx->caps8));
736 res10 = CHECK_CU(ctx->cvdl->cuvidGetDecoderCaps(&ctx->caps10));
737 res12 = CHECK_CU(ctx->cvdl->cuvidGetDecoderCaps(&ctx->caps12));
739 av_log(avctx, AV_LOG_VERBOSE, "CUVID capabilities for %s:\n", avctx->codec->name);
740 av_log(avctx, AV_LOG_VERBOSE, "8 bit: supported: %d, min_width: %d, max_width: %d, min_height: %d, max_height: %d\n",
741 ctx->caps8.bIsSupported, ctx->caps8.nMinWidth, ctx->caps8.nMaxWidth, ctx->caps8.nMinHeight, ctx->caps8.nMaxHeight);
742 av_log(avctx, AV_LOG_VERBOSE, "10 bit: supported: %d, min_width: %d, max_width: %d, min_height: %d, max_height: %d\n",
743 ctx->caps10.bIsSupported, ctx->caps10.nMinWidth, ctx->caps10.nMaxWidth, ctx->caps10.nMinHeight, ctx->caps10.nMaxHeight);
744 av_log(avctx, AV_LOG_VERBOSE, "12 bit: supported: %d, min_width: %d, max_width: %d, min_height: %d, max_height: %d\n",
745 ctx->caps12.bIsSupported, ctx->caps12.nMinWidth, ctx->caps12.nMaxWidth, ctx->caps12.nMinHeight, ctx->caps12.nMaxHeight);
764 if (!ctx->caps8.bIsSupported) {
765 av_log(avctx, AV_LOG_ERROR, "Codec %s is not supported.\n", avctx->codec->name);
766 return AVERROR(EINVAL);
769 if (!caps->bIsSupported) {
770 av_log(avctx, AV_LOG_ERROR, "Bit depth %d is not supported.\n", bit_depth);
771 return AVERROR(EINVAL);
774 if (probed_width > caps->nMaxWidth || probed_width < caps->nMinWidth) {
775 av_log(avctx, AV_LOG_ERROR, "Video width %d not within range from %d to %d\n",
776 probed_width, caps->nMinWidth, caps->nMaxWidth);
777 return AVERROR(EINVAL);
780 if (probed_height > caps->nMaxHeight || probed_height < caps->nMinHeight) {
781 av_log(avctx, AV_LOG_ERROR, "Video height %d not within range from %d to %d\n",
782 probed_height, caps->nMinHeight, caps->nMaxHeight);
783 return AVERROR(EINVAL);
789 static av_cold int cuvid_decode_init(AVCodecContext *avctx)
791 CuvidContext *ctx = avctx->priv_data;
792 AVCUDADeviceContext *device_hwctx;
793 AVHWDeviceContext *device_ctx;
794 AVHWFramesContext *hwframe_ctx;
795 CUVIDSOURCEDATAPACKET seq_pkt;
796 CUcontext cuda_ctx = NULL;
798 const AVBitStreamFilter *bsf;
801 enum AVPixelFormat pix_fmts[3] = { AV_PIX_FMT_CUDA,
805 int probed_width = avctx->coded_width ? avctx->coded_width : 1280;
806 int probed_height = avctx->coded_height ? avctx->coded_height : 720;
807 int probed_bit_depth = 8;
809 const AVPixFmtDescriptor *probe_desc = av_pix_fmt_desc_get(avctx->pix_fmt);
810 if (probe_desc && probe_desc->nb_components)
811 probed_bit_depth = probe_desc->comp[0].depth;
813 // Accelerated transcoding scenarios with 'ffmpeg' require that the
814 // pix_fmt be set to AV_PIX_FMT_CUDA early. The sw_pix_fmt, and the
815 // pix_fmt for non-accelerated transcoding, do not need to be correct
816 // but need to be set to something. We arbitrarily pick NV12.
817 ret = ff_get_format(avctx, pix_fmts);
819 av_log(avctx, AV_LOG_ERROR, "ff_get_format failed: %d\n", ret);
822 avctx->pix_fmt = ret;
824 if (ctx->resize_expr && sscanf(ctx->resize_expr, "%dx%d",
825 &ctx->resize.width, &ctx->resize.height) != 2) {
826 av_log(avctx, AV_LOG_ERROR, "Invalid resize expressions\n");
827 ret = AVERROR(EINVAL);
831 if (ctx->crop_expr && sscanf(ctx->crop_expr, "%dx%dx%dx%d",
832 &ctx->crop.top, &ctx->crop.bottom,
833 &ctx->crop.left, &ctx->crop.right) != 4) {
834 av_log(avctx, AV_LOG_ERROR, "Invalid cropping expressions\n");
835 ret = AVERROR(EINVAL);
839 ret = cuvid_load_functions(&ctx->cvdl, avctx);
841 av_log(avctx, AV_LOG_ERROR, "Failed loading nvcuvid.\n");
845 ctx->frame_queue = av_fifo_alloc(ctx->nb_surfaces * sizeof(CuvidParsedFrame));
846 if (!ctx->frame_queue) {
847 ret = AVERROR(ENOMEM);
851 if (avctx->hw_frames_ctx) {
852 ctx->hwframe = av_buffer_ref(avctx->hw_frames_ctx);
854 ret = AVERROR(ENOMEM);
858 hwframe_ctx = (AVHWFramesContext*)ctx->hwframe->data;
860 ctx->hwdevice = av_buffer_ref(hwframe_ctx->device_ref);
861 if (!ctx->hwdevice) {
862 ret = AVERROR(ENOMEM);
866 if (avctx->hw_device_ctx) {
867 ctx->hwdevice = av_buffer_ref(avctx->hw_device_ctx);
868 if (!ctx->hwdevice) {
869 ret = AVERROR(ENOMEM);
873 ret = av_hwdevice_ctx_create(&ctx->hwdevice, AV_HWDEVICE_TYPE_CUDA, ctx->cu_gpu, NULL, 0);
878 ctx->hwframe = av_hwframe_ctx_alloc(ctx->hwdevice);
880 av_log(avctx, AV_LOG_ERROR, "av_hwframe_ctx_alloc failed\n");
881 ret = AVERROR(ENOMEM);
885 hwframe_ctx = (AVHWFramesContext*)ctx->hwframe->data;
888 device_ctx = hwframe_ctx->device_ctx;
889 device_hwctx = device_ctx->hwctx;
891 cuda_ctx = device_hwctx->cuda_ctx;
892 ctx->cudl = device_hwctx->internal->cuda_dl;
894 memset(&ctx->cuparseinfo, 0, sizeof(ctx->cuparseinfo));
895 memset(&ctx->cuparse_ext, 0, sizeof(ctx->cuparse_ext));
896 memset(&seq_pkt, 0, sizeof(seq_pkt));
898 ctx->cuparseinfo.pExtVideoInfo = &ctx->cuparse_ext;
900 switch (avctx->codec->id) {
901 #if CONFIG_H264_CUVID_DECODER
902 case AV_CODEC_ID_H264:
903 ctx->cuparseinfo.CodecType = cudaVideoCodec_H264;
906 #if CONFIG_HEVC_CUVID_DECODER
907 case AV_CODEC_ID_HEVC:
908 ctx->cuparseinfo.CodecType = cudaVideoCodec_HEVC;
911 #if CONFIG_MJPEG_CUVID_DECODER
912 case AV_CODEC_ID_MJPEG:
913 ctx->cuparseinfo.CodecType = cudaVideoCodec_JPEG;
916 #if CONFIG_MPEG1_CUVID_DECODER
917 case AV_CODEC_ID_MPEG1VIDEO:
918 ctx->cuparseinfo.CodecType = cudaVideoCodec_MPEG1;
921 #if CONFIG_MPEG2_CUVID_DECODER
922 case AV_CODEC_ID_MPEG2VIDEO:
923 ctx->cuparseinfo.CodecType = cudaVideoCodec_MPEG2;
926 #if CONFIG_MPEG4_CUVID_DECODER
927 case AV_CODEC_ID_MPEG4:
928 ctx->cuparseinfo.CodecType = cudaVideoCodec_MPEG4;
931 #if CONFIG_VP8_CUVID_DECODER
932 case AV_CODEC_ID_VP8:
933 ctx->cuparseinfo.CodecType = cudaVideoCodec_VP8;
936 #if CONFIG_VP9_CUVID_DECODER
937 case AV_CODEC_ID_VP9:
938 ctx->cuparseinfo.CodecType = cudaVideoCodec_VP9;
941 #if CONFIG_VC1_CUVID_DECODER
942 case AV_CODEC_ID_VC1:
943 ctx->cuparseinfo.CodecType = cudaVideoCodec_VC1;
947 av_log(avctx, AV_LOG_ERROR, "Invalid CUVID codec!\n");
951 if (avctx->codec->id == AV_CODEC_ID_H264 || avctx->codec->id == AV_CODEC_ID_HEVC) {
952 if (avctx->codec->id == AV_CODEC_ID_H264)
953 bsf = av_bsf_get_by_name("h264_mp4toannexb");
955 bsf = av_bsf_get_by_name("hevc_mp4toannexb");
958 ret = AVERROR_BSF_NOT_FOUND;
961 if (ret = av_bsf_alloc(bsf, &ctx->bsf)) {
964 if (((ret = avcodec_parameters_from_context(ctx->bsf->par_in, avctx)) < 0) || ((ret = av_bsf_init(ctx->bsf)) < 0)) {
965 av_bsf_free(&ctx->bsf);
969 ctx->cuparse_ext.format.seqhdr_data_length = ctx->bsf->par_out->extradata_size;
970 memcpy(ctx->cuparse_ext.raw_seqhdr_data,
971 ctx->bsf->par_out->extradata,
972 FFMIN(sizeof(ctx->cuparse_ext.raw_seqhdr_data), ctx->bsf->par_out->extradata_size));
973 } else if (avctx->extradata_size > 0) {
974 ctx->cuparse_ext.format.seqhdr_data_length = avctx->extradata_size;
975 memcpy(ctx->cuparse_ext.raw_seqhdr_data,
977 FFMIN(sizeof(ctx->cuparse_ext.raw_seqhdr_data), avctx->extradata_size));
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)