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;
79 cudaVideoCodec codec_type;
80 cudaVideoChromaFormat chroma_format;
82 CUVIDDECODECAPS caps8, caps10, caps12;
84 CUVIDPARSERPARAMS cuparseinfo;
85 CUVIDEOFORMATEX cuparse_ext;
91 typedef struct CuvidParsedFrame
93 CUVIDPARSERDISPINFO dispinfo;
98 static int check_cu(AVCodecContext *avctx, CUresult err, const char *func)
100 CuvidContext *ctx = avctx->priv_data;
101 const char *err_name;
102 const char *err_string;
104 av_log(avctx, AV_LOG_TRACE, "Calling %s\n", func);
106 if (err == CUDA_SUCCESS)
109 ctx->cudl->cuGetErrorName(err, &err_name);
110 ctx->cudl->cuGetErrorString(err, &err_string);
112 av_log(avctx, AV_LOG_ERROR, "%s failed", func);
113 if (err_name && err_string)
114 av_log(avctx, AV_LOG_ERROR, " -> %s: %s", err_name, err_string);
115 av_log(avctx, AV_LOG_ERROR, "\n");
117 return AVERROR_EXTERNAL;
120 #define CHECK_CU(x) check_cu(avctx, (x), #x)
122 static int CUDAAPI cuvid_handle_video_sequence(void *opaque, CUVIDEOFORMAT* format)
124 AVCodecContext *avctx = opaque;
125 CuvidContext *ctx = avctx->priv_data;
126 AVHWFramesContext *hwframe_ctx = (AVHWFramesContext*)ctx->hwframe->data;
127 CUVIDDECODECAPS *caps = NULL;
128 CUVIDDECODECREATEINFO cuinfo;
131 int old_width = avctx->width;
132 int old_height = avctx->height;
134 enum AVPixelFormat pix_fmts[3] = { AV_PIX_FMT_CUDA,
135 AV_PIX_FMT_NONE, // Will be updated below
138 av_log(avctx, AV_LOG_TRACE, "pfnSequenceCallback, progressive_sequence=%d\n", format->progressive_sequence);
140 memset(&cuinfo, 0, sizeof(cuinfo));
142 ctx->internal_error = 0;
144 avctx->coded_width = cuinfo.ulWidth = format->coded_width;
145 avctx->coded_height = cuinfo.ulHeight = format->coded_height;
148 cuinfo.display_area.left = format->display_area.left + ctx->crop.left;
149 cuinfo.display_area.top = format->display_area.top + ctx->crop.top;
150 cuinfo.display_area.right = format->display_area.right - ctx->crop.right;
151 cuinfo.display_area.bottom = format->display_area.bottom - ctx->crop.bottom;
153 // width and height need to be set before calling ff_get_format
154 if (ctx->resize_expr) {
155 avctx->width = ctx->resize.width;
156 avctx->height = ctx->resize.height;
158 avctx->width = cuinfo.display_area.right - cuinfo.display_area.left;
159 avctx->height = cuinfo.display_area.bottom - cuinfo.display_area.top;
162 // target width/height need to be multiples of two
163 cuinfo.ulTargetWidth = avctx->width = (avctx->width + 1) & ~1;
164 cuinfo.ulTargetHeight = avctx->height = (avctx->height + 1) & ~1;
166 // aspect ratio conversion, 1:1, depends on scaled resolution
167 cuinfo.target_rect.left = 0;
168 cuinfo.target_rect.top = 0;
169 cuinfo.target_rect.right = cuinfo.ulTargetWidth;
170 cuinfo.target_rect.bottom = cuinfo.ulTargetHeight;
172 switch (format->bit_depth_luma_minus8) {
174 pix_fmts[1] = AV_PIX_FMT_NV12;
178 pix_fmts[1] = AV_PIX_FMT_P010;
182 pix_fmts[1] = AV_PIX_FMT_P016;
189 if (!caps || !caps->bIsSupported) {
190 av_log(avctx, AV_LOG_ERROR, "unsupported bit depth: %d\n",
191 format->bit_depth_luma_minus8 + 8);
192 ctx->internal_error = AVERROR(EINVAL);
196 surface_fmt = ff_get_format(avctx, pix_fmts);
197 if (surface_fmt < 0) {
198 av_log(avctx, AV_LOG_ERROR, "ff_get_format failed: %d\n", surface_fmt);
199 ctx->internal_error = AVERROR(EINVAL);
203 av_log(avctx, AV_LOG_VERBOSE, "Formats: Original: %s | HW: %s | SW: %s\n",
204 av_get_pix_fmt_name(avctx->pix_fmt),
205 av_get_pix_fmt_name(surface_fmt),
206 av_get_pix_fmt_name(avctx->sw_pix_fmt));
208 avctx->pix_fmt = surface_fmt;
210 // Update our hwframe ctx, as the get_format callback might have refreshed it!
211 if (avctx->hw_frames_ctx) {
212 av_buffer_unref(&ctx->hwframe);
214 ctx->hwframe = av_buffer_ref(avctx->hw_frames_ctx);
216 ctx->internal_error = AVERROR(ENOMEM);
220 hwframe_ctx = (AVHWFramesContext*)ctx->hwframe->data;
223 ff_set_sar(avctx, av_div_q(
224 (AVRational){ format->display_aspect_ratio.x, format->display_aspect_ratio.y },
225 (AVRational){ avctx->width, avctx->height }));
227 ctx->deint_mode_current = format->progressive_sequence
228 ? cudaVideoDeinterlaceMode_Weave
231 if (!format->progressive_sequence && ctx->deint_mode_current == cudaVideoDeinterlaceMode_Weave)
232 avctx->flags |= AV_CODEC_FLAG_INTERLACED_DCT;
234 avctx->flags &= ~AV_CODEC_FLAG_INTERLACED_DCT;
236 if (format->video_signal_description.video_full_range_flag)
237 avctx->color_range = AVCOL_RANGE_JPEG;
239 avctx->color_range = AVCOL_RANGE_MPEG;
241 avctx->color_primaries = format->video_signal_description.color_primaries;
242 avctx->color_trc = format->video_signal_description.transfer_characteristics;
243 avctx->colorspace = format->video_signal_description.matrix_coefficients;
246 avctx->bit_rate = format->bitrate;
248 if (format->frame_rate.numerator && format->frame_rate.denominator) {
249 avctx->framerate.num = format->frame_rate.numerator;
250 avctx->framerate.den = format->frame_rate.denominator;
254 && avctx->coded_width == format->coded_width
255 && avctx->coded_height == format->coded_height
256 && avctx->width == old_width
257 && avctx->height == old_height
258 && ctx->chroma_format == format->chroma_format
259 && ctx->codec_type == format->codec)
262 if (ctx->cudecoder) {
263 av_log(avctx, AV_LOG_TRACE, "Re-initializing decoder\n");
264 ctx->internal_error = CHECK_CU(ctx->cvdl->cuvidDestroyDecoder(ctx->cudecoder));
265 if (ctx->internal_error < 0)
267 ctx->cudecoder = NULL;
270 if (hwframe_ctx->pool && (
271 hwframe_ctx->width < avctx->width ||
272 hwframe_ctx->height < avctx->height ||
273 hwframe_ctx->format != AV_PIX_FMT_CUDA ||
274 hwframe_ctx->sw_format != avctx->sw_pix_fmt)) {
275 av_log(avctx, AV_LOG_ERROR, "AVHWFramesContext is already initialized with incompatible parameters\n");
276 av_log(avctx, AV_LOG_DEBUG, "width: %d <-> %d\n", hwframe_ctx->width, avctx->width);
277 av_log(avctx, AV_LOG_DEBUG, "height: %d <-> %d\n", hwframe_ctx->height, avctx->height);
278 av_log(avctx, AV_LOG_DEBUG, "format: %s <-> cuda\n", av_get_pix_fmt_name(hwframe_ctx->format));
279 av_log(avctx, AV_LOG_DEBUG, "sw_format: %s <-> %s\n",
280 av_get_pix_fmt_name(hwframe_ctx->sw_format), av_get_pix_fmt_name(avctx->sw_pix_fmt));
281 ctx->internal_error = AVERROR(EINVAL);
285 if (format->chroma_format != cudaVideoChromaFormat_420) {
286 av_log(avctx, AV_LOG_ERROR, "Chroma formats other than 420 are not supported\n");
287 ctx->internal_error = AVERROR(EINVAL);
291 ctx->chroma_format = format->chroma_format;
293 cuinfo.CodecType = ctx->codec_type = format->codec;
294 cuinfo.ChromaFormat = format->chroma_format;
296 switch (avctx->sw_pix_fmt) {
297 case AV_PIX_FMT_NV12:
298 cuinfo.OutputFormat = cudaVideoSurfaceFormat_NV12;
300 case AV_PIX_FMT_P010:
301 case AV_PIX_FMT_P016:
302 cuinfo.OutputFormat = cudaVideoSurfaceFormat_P016;
305 av_log(avctx, AV_LOG_ERROR, "Output formats other than NV12, P010 or P016 are not supported\n");
306 ctx->internal_error = AVERROR(EINVAL);
310 cuinfo.ulNumDecodeSurfaces = ctx->nb_surfaces;
311 cuinfo.ulNumOutputSurfaces = 1;
312 cuinfo.ulCreationFlags = cudaVideoCreate_PreferCUVID;
313 cuinfo.bitDepthMinus8 = format->bit_depth_luma_minus8;
314 cuinfo.DeinterlaceMode = ctx->deint_mode_current;
316 if (ctx->deint_mode_current != cudaVideoDeinterlaceMode_Weave && !ctx->drop_second_field)
317 avctx->framerate = av_mul_q(avctx->framerate, (AVRational){2, 1});
319 ctx->internal_error = CHECK_CU(ctx->cvdl->cuvidCreateDecoder(&ctx->cudecoder, &cuinfo));
320 if (ctx->internal_error < 0)
323 if (!hwframe_ctx->pool) {
324 hwframe_ctx->format = AV_PIX_FMT_CUDA;
325 hwframe_ctx->sw_format = avctx->sw_pix_fmt;
326 hwframe_ctx->width = avctx->width;
327 hwframe_ctx->height = avctx->height;
329 if ((ctx->internal_error = av_hwframe_ctx_init(ctx->hwframe)) < 0) {
330 av_log(avctx, AV_LOG_ERROR, "av_hwframe_ctx_init failed\n");
338 static int CUDAAPI cuvid_handle_picture_decode(void *opaque, CUVIDPICPARAMS* picparams)
340 AVCodecContext *avctx = opaque;
341 CuvidContext *ctx = avctx->priv_data;
343 av_log(avctx, AV_LOG_TRACE, "pfnDecodePicture\n");
345 ctx->key_frame[picparams->CurrPicIdx] = picparams->intra_pic_flag;
347 ctx->internal_error = CHECK_CU(ctx->cvdl->cuvidDecodePicture(ctx->cudecoder, picparams));
348 if (ctx->internal_error < 0)
354 static int CUDAAPI cuvid_handle_picture_display(void *opaque, CUVIDPARSERDISPINFO* dispinfo)
356 AVCodecContext *avctx = opaque;
357 CuvidContext *ctx = avctx->priv_data;
358 CuvidParsedFrame parsed_frame = { { 0 } };
360 parsed_frame.dispinfo = *dispinfo;
361 ctx->internal_error = 0;
363 if (ctx->deint_mode_current == cudaVideoDeinterlaceMode_Weave) {
364 av_fifo_generic_write(ctx->frame_queue, &parsed_frame, sizeof(CuvidParsedFrame), NULL);
366 parsed_frame.is_deinterlacing = 1;
367 av_fifo_generic_write(ctx->frame_queue, &parsed_frame, sizeof(CuvidParsedFrame), NULL);
368 if (!ctx->drop_second_field) {
369 parsed_frame.second_field = 1;
370 av_fifo_generic_write(ctx->frame_queue, &parsed_frame, sizeof(CuvidParsedFrame), NULL);
377 static int cuvid_is_buffer_full(AVCodecContext *avctx)
379 CuvidContext *ctx = avctx->priv_data;
381 return (av_fifo_size(ctx->frame_queue) / sizeof(CuvidParsedFrame)) + 2 > ctx->nb_surfaces;
384 static int cuvid_decode_packet(AVCodecContext *avctx, const AVPacket *avpkt)
386 CuvidContext *ctx = avctx->priv_data;
387 AVHWDeviceContext *device_ctx = (AVHWDeviceContext*)ctx->hwdevice->data;
388 AVCUDADeviceContext *device_hwctx = device_ctx->hwctx;
389 CUcontext dummy, cuda_ctx = device_hwctx->cuda_ctx;
390 CUVIDSOURCEDATAPACKET cupkt;
391 AVPacket filter_packet = { 0 };
392 AVPacket filtered_packet = { 0 };
393 int ret = 0, eret = 0, is_flush = ctx->decoder_flushing;
395 av_log(avctx, AV_LOG_TRACE, "cuvid_decode_packet\n");
397 if (is_flush && avpkt && avpkt->size)
400 if (cuvid_is_buffer_full(avctx) && avpkt && avpkt->size)
401 return AVERROR(EAGAIN);
403 if (ctx->bsf && avpkt && avpkt->size) {
404 if ((ret = av_packet_ref(&filter_packet, avpkt)) < 0) {
405 av_log(avctx, AV_LOG_ERROR, "av_packet_ref failed\n");
409 if ((ret = av_bsf_send_packet(ctx->bsf, &filter_packet)) < 0) {
410 av_log(avctx, AV_LOG_ERROR, "av_bsf_send_packet failed\n");
411 av_packet_unref(&filter_packet);
415 if ((ret = av_bsf_receive_packet(ctx->bsf, &filtered_packet)) < 0) {
416 av_log(avctx, AV_LOG_ERROR, "av_bsf_receive_packet failed\n");
420 avpkt = &filtered_packet;
423 ret = CHECK_CU(ctx->cudl->cuCtxPushCurrent(cuda_ctx));
425 av_packet_unref(&filtered_packet);
429 memset(&cupkt, 0, sizeof(cupkt));
431 if (avpkt && avpkt->size) {
432 cupkt.payload_size = avpkt->size;
433 cupkt.payload = avpkt->data;
435 if (avpkt->pts != AV_NOPTS_VALUE) {
436 cupkt.flags = CUVID_PKT_TIMESTAMP;
437 if (avctx->pkt_timebase.num && avctx->pkt_timebase.den)
438 cupkt.timestamp = av_rescale_q(avpkt->pts, avctx->pkt_timebase, (AVRational){1, 10000000});
440 cupkt.timestamp = avpkt->pts;
443 cupkt.flags = CUVID_PKT_ENDOFSTREAM;
444 ctx->decoder_flushing = 1;
447 ret = CHECK_CU(ctx->cvdl->cuvidParseVideoData(ctx->cuparser, &cupkt));
449 av_packet_unref(&filtered_packet);
454 // cuvidParseVideoData doesn't return an error just because stuff failed...
455 if (ctx->internal_error) {
456 av_log(avctx, AV_LOG_ERROR, "cuvid decode callback error\n");
457 ret = ctx->internal_error;
462 eret = CHECK_CU(ctx->cudl->cuCtxPopCurrent(&dummy));
474 static int cuvid_output_frame(AVCodecContext *avctx, AVFrame *frame)
476 CuvidContext *ctx = avctx->priv_data;
477 AVHWDeviceContext *device_ctx = (AVHWDeviceContext*)ctx->hwdevice->data;
478 AVCUDADeviceContext *device_hwctx = device_ctx->hwctx;
479 CUcontext dummy, cuda_ctx = device_hwctx->cuda_ctx;
480 CUdeviceptr mapped_frame = 0;
481 int ret = 0, eret = 0;
483 av_log(avctx, AV_LOG_TRACE, "cuvid_output_frame\n");
485 if (ctx->decoder_flushing) {
486 ret = cuvid_decode_packet(avctx, NULL);
487 if (ret < 0 && ret != AVERROR_EOF)
491 if (!cuvid_is_buffer_full(avctx)) {
493 ret = ff_decode_get_packet(avctx, &pkt);
494 if (ret < 0 && ret != AVERROR_EOF)
496 ret = cuvid_decode_packet(avctx, &pkt);
497 av_packet_unref(&pkt);
498 // cuvid_is_buffer_full() should avoid this.
499 if (ret == AVERROR(EAGAIN))
500 ret = AVERROR_EXTERNAL;
501 if (ret < 0 && ret != AVERROR_EOF)
505 ret = CHECK_CU(ctx->cudl->cuCtxPushCurrent(cuda_ctx));
509 if (av_fifo_size(ctx->frame_queue)) {
510 CuvidParsedFrame parsed_frame;
511 CUVIDPROCPARAMS params;
512 unsigned int pitch = 0;
516 av_fifo_generic_read(ctx->frame_queue, &parsed_frame, sizeof(CuvidParsedFrame), NULL);
518 memset(¶ms, 0, sizeof(params));
519 params.progressive_frame = parsed_frame.dispinfo.progressive_frame;
520 params.second_field = parsed_frame.second_field;
521 params.top_field_first = parsed_frame.dispinfo.top_field_first;
523 ret = CHECK_CU(ctx->cvdl->cuvidMapVideoFrame(ctx->cudecoder, parsed_frame.dispinfo.picture_index, &mapped_frame, &pitch, ¶ms));
527 if (avctx->pix_fmt == AV_PIX_FMT_CUDA) {
528 ret = av_hwframe_get_buffer(ctx->hwframe, frame, 0);
530 av_log(avctx, AV_LOG_ERROR, "av_hwframe_get_buffer failed\n");
534 ret = ff_decode_frame_props(avctx, frame);
536 av_log(avctx, AV_LOG_ERROR, "ff_decode_frame_props failed\n");
540 for (i = 0; i < 2; i++) {
541 CUDA_MEMCPY2D cpy = {
542 .srcMemoryType = CU_MEMORYTYPE_DEVICE,
543 .dstMemoryType = CU_MEMORYTYPE_DEVICE,
544 .srcDevice = mapped_frame,
545 .dstDevice = (CUdeviceptr)frame->data[i],
547 .dstPitch = frame->linesize[i],
549 .WidthInBytes = FFMIN(pitch, frame->linesize[i]),
550 .Height = avctx->height >> (i ? 1 : 0),
553 ret = CHECK_CU(ctx->cudl->cuMemcpy2D(&cpy));
557 offset += avctx->height;
559 } else if (avctx->pix_fmt == AV_PIX_FMT_NV12 ||
560 avctx->pix_fmt == AV_PIX_FMT_P010 ||
561 avctx->pix_fmt == AV_PIX_FMT_P016) {
562 AVFrame *tmp_frame = av_frame_alloc();
564 av_log(avctx, AV_LOG_ERROR, "av_frame_alloc failed\n");
565 ret = AVERROR(ENOMEM);
569 tmp_frame->format = AV_PIX_FMT_CUDA;
570 tmp_frame->hw_frames_ctx = av_buffer_ref(ctx->hwframe);
571 tmp_frame->data[0] = (uint8_t*)mapped_frame;
572 tmp_frame->linesize[0] = pitch;
573 tmp_frame->data[1] = (uint8_t*)(mapped_frame + avctx->height * pitch);
574 tmp_frame->linesize[1] = pitch;
575 tmp_frame->width = avctx->width;
576 tmp_frame->height = avctx->height;
578 ret = ff_get_buffer(avctx, frame, 0);
580 av_log(avctx, AV_LOG_ERROR, "ff_get_buffer failed\n");
581 av_frame_free(&tmp_frame);
585 ret = av_hwframe_transfer_data(frame, tmp_frame, 0);
587 av_log(avctx, AV_LOG_ERROR, "av_hwframe_transfer_data failed\n");
588 av_frame_free(&tmp_frame);
591 av_frame_free(&tmp_frame);
597 frame->key_frame = ctx->key_frame[parsed_frame.dispinfo.picture_index];
598 frame->width = avctx->width;
599 frame->height = avctx->height;
600 if (avctx->pkt_timebase.num && avctx->pkt_timebase.den)
601 frame->pts = av_rescale_q(parsed_frame.dispinfo.timestamp, (AVRational){1, 10000000}, avctx->pkt_timebase);
603 frame->pts = parsed_frame.dispinfo.timestamp;
605 if (parsed_frame.second_field) {
606 if (ctx->prev_pts == INT64_MIN) {
607 ctx->prev_pts = frame->pts;
608 frame->pts += (avctx->pkt_timebase.den * avctx->framerate.den) / (avctx->pkt_timebase.num * avctx->framerate.num);
610 int pts_diff = (frame->pts - ctx->prev_pts) / 2;
611 ctx->prev_pts = frame->pts;
612 frame->pts += pts_diff;
616 /* CUVIDs opaque reordering breaks the internal pkt logic.
617 * So set pkt_pts and clear all the other pkt_ fields.
620 FF_DISABLE_DEPRECATION_WARNINGS
621 frame->pkt_pts = frame->pts;
622 FF_ENABLE_DEPRECATION_WARNINGS
625 frame->pkt_duration = 0;
626 frame->pkt_size = -1;
628 frame->interlaced_frame = !parsed_frame.is_deinterlacing && !parsed_frame.dispinfo.progressive_frame;
630 if (frame->interlaced_frame)
631 frame->top_field_first = parsed_frame.dispinfo.top_field_first;
632 } else if (ctx->decoder_flushing) {
635 ret = AVERROR(EAGAIN);
640 eret = CHECK_CU(ctx->cvdl->cuvidUnmapVideoFrame(ctx->cudecoder, mapped_frame));
642 eret = CHECK_CU(ctx->cudl->cuCtxPopCurrent(&dummy));
650 static int cuvid_decode_frame(AVCodecContext *avctx, void *data, int *got_frame, AVPacket *avpkt)
652 CuvidContext *ctx = avctx->priv_data;
653 AVFrame *frame = data;
656 av_log(avctx, AV_LOG_TRACE, "cuvid_decode_frame\n");
658 if (ctx->deint_mode_current != cudaVideoDeinterlaceMode_Weave) {
659 av_log(avctx, AV_LOG_ERROR, "Deinterlacing is not supported via the old API\n");
660 return AVERROR(EINVAL);
663 if (!ctx->decoder_flushing) {
664 ret = cuvid_decode_packet(avctx, avpkt);
669 ret = cuvid_output_frame(avctx, frame);
670 if (ret == AVERROR(EAGAIN) || ret == AVERROR_EOF) {
672 } else if (ret < 0) {
681 static av_cold int cuvid_decode_end(AVCodecContext *avctx)
683 CuvidContext *ctx = avctx->priv_data;
685 av_fifo_freep(&ctx->frame_queue);
688 av_bsf_free(&ctx->bsf);
691 ctx->cvdl->cuvidDestroyVideoParser(ctx->cuparser);
694 ctx->cvdl->cuvidDestroyDecoder(ctx->cudecoder);
698 av_buffer_unref(&ctx->hwframe);
699 av_buffer_unref(&ctx->hwdevice);
701 av_freep(&ctx->key_frame);
703 cuvid_free_functions(&ctx->cvdl);
708 static int cuvid_test_capabilities(AVCodecContext *avctx,
709 const CUVIDPARSERPARAMS *cuparseinfo,
714 CuvidContext *ctx = avctx->priv_data;
715 CUVIDDECODECAPS *caps;
716 int res8 = 0, res10 = 0, res12 = 0;
718 if (!ctx->cvdl->cuvidGetDecoderCaps) {
719 av_log(avctx, AV_LOG_WARNING, "Used Nvidia driver is too old to perform a capability check.\n");
720 av_log(avctx, AV_LOG_WARNING, "The minimum required version is "
721 #if defined(_WIN32) || defined(__CYGWIN__)
726 ". Continuing blind.\n");
727 ctx->caps8.bIsSupported = ctx->caps10.bIsSupported = 1;
728 // 12 bit was not supported before the capability check was introduced, so disable it.
729 ctx->caps12.bIsSupported = 0;
733 ctx->caps8.eCodecType = ctx->caps10.eCodecType = ctx->caps12.eCodecType
734 = cuparseinfo->CodecType;
735 ctx->caps8.eChromaFormat = ctx->caps10.eChromaFormat = ctx->caps12.eChromaFormat
736 = cudaVideoChromaFormat_420;
738 ctx->caps8.nBitDepthMinus8 = 0;
739 ctx->caps10.nBitDepthMinus8 = 2;
740 ctx->caps12.nBitDepthMinus8 = 4;
742 res8 = CHECK_CU(ctx->cvdl->cuvidGetDecoderCaps(&ctx->caps8));
743 res10 = CHECK_CU(ctx->cvdl->cuvidGetDecoderCaps(&ctx->caps10));
744 res12 = CHECK_CU(ctx->cvdl->cuvidGetDecoderCaps(&ctx->caps12));
746 av_log(avctx, AV_LOG_VERBOSE, "CUVID capabilities for %s:\n", avctx->codec->name);
747 av_log(avctx, AV_LOG_VERBOSE, "8 bit: supported: %d, min_width: %d, max_width: %d, min_height: %d, max_height: %d\n",
748 ctx->caps8.bIsSupported, ctx->caps8.nMinWidth, ctx->caps8.nMaxWidth, ctx->caps8.nMinHeight, ctx->caps8.nMaxHeight);
749 av_log(avctx, AV_LOG_VERBOSE, "10 bit: supported: %d, min_width: %d, max_width: %d, min_height: %d, max_height: %d\n",
750 ctx->caps10.bIsSupported, ctx->caps10.nMinWidth, ctx->caps10.nMaxWidth, ctx->caps10.nMinHeight, ctx->caps10.nMaxHeight);
751 av_log(avctx, AV_LOG_VERBOSE, "12 bit: supported: %d, min_width: %d, max_width: %d, min_height: %d, max_height: %d\n",
752 ctx->caps12.bIsSupported, ctx->caps12.nMinWidth, ctx->caps12.nMaxWidth, ctx->caps12.nMinHeight, ctx->caps12.nMaxHeight);
771 if (!ctx->caps8.bIsSupported) {
772 av_log(avctx, AV_LOG_ERROR, "Codec %s is not supported.\n", avctx->codec->name);
773 return AVERROR(EINVAL);
776 if (!caps->bIsSupported) {
777 av_log(avctx, AV_LOG_ERROR, "Bit depth %d is not supported.\n", bit_depth);
778 return AVERROR(EINVAL);
781 if (probed_width > caps->nMaxWidth || probed_width < caps->nMinWidth) {
782 av_log(avctx, AV_LOG_ERROR, "Video width %d not within range from %d to %d\n",
783 probed_width, caps->nMinWidth, caps->nMaxWidth);
784 return AVERROR(EINVAL);
787 if (probed_height > caps->nMaxHeight || probed_height < caps->nMinHeight) {
788 av_log(avctx, AV_LOG_ERROR, "Video height %d not within range from %d to %d\n",
789 probed_height, caps->nMinHeight, caps->nMaxHeight);
790 return AVERROR(EINVAL);
796 static av_cold int cuvid_decode_init(AVCodecContext *avctx)
798 CuvidContext *ctx = avctx->priv_data;
799 AVCUDADeviceContext *device_hwctx;
800 AVHWDeviceContext *device_ctx;
801 AVHWFramesContext *hwframe_ctx;
802 CUVIDSOURCEDATAPACKET seq_pkt;
803 CUcontext cuda_ctx = NULL;
805 const AVBitStreamFilter *bsf;
808 enum AVPixelFormat pix_fmts[3] = { AV_PIX_FMT_CUDA,
812 int probed_width = avctx->coded_width ? avctx->coded_width : 1280;
813 int probed_height = avctx->coded_height ? avctx->coded_height : 720;
814 int probed_bit_depth = 8;
816 const AVPixFmtDescriptor *probe_desc = av_pix_fmt_desc_get(avctx->pix_fmt);
817 if (probe_desc && probe_desc->nb_components)
818 probed_bit_depth = probe_desc->comp[0].depth;
820 // Accelerated transcoding scenarios with 'ffmpeg' require that the
821 // pix_fmt be set to AV_PIX_FMT_CUDA early. The sw_pix_fmt, and the
822 // pix_fmt for non-accelerated transcoding, do not need to be correct
823 // but need to be set to something. We arbitrarily pick NV12.
824 ret = ff_get_format(avctx, pix_fmts);
826 av_log(avctx, AV_LOG_ERROR, "ff_get_format failed: %d\n", ret);
829 avctx->pix_fmt = ret;
831 if (ctx->resize_expr && sscanf(ctx->resize_expr, "%dx%d",
832 &ctx->resize.width, &ctx->resize.height) != 2) {
833 av_log(avctx, AV_LOG_ERROR, "Invalid resize expressions\n");
834 ret = AVERROR(EINVAL);
838 if (ctx->crop_expr && sscanf(ctx->crop_expr, "%dx%dx%dx%d",
839 &ctx->crop.top, &ctx->crop.bottom,
840 &ctx->crop.left, &ctx->crop.right) != 4) {
841 av_log(avctx, AV_LOG_ERROR, "Invalid cropping expressions\n");
842 ret = AVERROR(EINVAL);
846 ret = cuvid_load_functions(&ctx->cvdl, avctx);
848 av_log(avctx, AV_LOG_ERROR, "Failed loading nvcuvid.\n");
852 ctx->frame_queue = av_fifo_alloc(ctx->nb_surfaces * sizeof(CuvidParsedFrame));
853 if (!ctx->frame_queue) {
854 ret = AVERROR(ENOMEM);
858 if (avctx->hw_frames_ctx) {
859 ctx->hwframe = av_buffer_ref(avctx->hw_frames_ctx);
861 ret = AVERROR(ENOMEM);
865 hwframe_ctx = (AVHWFramesContext*)ctx->hwframe->data;
867 ctx->hwdevice = av_buffer_ref(hwframe_ctx->device_ref);
868 if (!ctx->hwdevice) {
869 ret = AVERROR(ENOMEM);
873 if (avctx->hw_device_ctx) {
874 ctx->hwdevice = av_buffer_ref(avctx->hw_device_ctx);
875 if (!ctx->hwdevice) {
876 ret = AVERROR(ENOMEM);
880 ret = av_hwdevice_ctx_create(&ctx->hwdevice, AV_HWDEVICE_TYPE_CUDA, ctx->cu_gpu, NULL, 0);
885 ctx->hwframe = av_hwframe_ctx_alloc(ctx->hwdevice);
887 av_log(avctx, AV_LOG_ERROR, "av_hwframe_ctx_alloc failed\n");
888 ret = AVERROR(ENOMEM);
892 hwframe_ctx = (AVHWFramesContext*)ctx->hwframe->data;
895 device_ctx = hwframe_ctx->device_ctx;
896 device_hwctx = device_ctx->hwctx;
898 cuda_ctx = device_hwctx->cuda_ctx;
899 ctx->cudl = device_hwctx->internal->cuda_dl;
901 memset(&ctx->cuparseinfo, 0, sizeof(ctx->cuparseinfo));
902 memset(&ctx->cuparse_ext, 0, sizeof(ctx->cuparse_ext));
903 memset(&seq_pkt, 0, sizeof(seq_pkt));
905 ctx->cuparseinfo.pExtVideoInfo = &ctx->cuparse_ext;
907 switch (avctx->codec->id) {
908 #if CONFIG_H264_CUVID_DECODER
909 case AV_CODEC_ID_H264:
910 ctx->cuparseinfo.CodecType = cudaVideoCodec_H264;
913 #if CONFIG_HEVC_CUVID_DECODER
914 case AV_CODEC_ID_HEVC:
915 ctx->cuparseinfo.CodecType = cudaVideoCodec_HEVC;
918 #if CONFIG_MJPEG_CUVID_DECODER
919 case AV_CODEC_ID_MJPEG:
920 ctx->cuparseinfo.CodecType = cudaVideoCodec_JPEG;
923 #if CONFIG_MPEG1_CUVID_DECODER
924 case AV_CODEC_ID_MPEG1VIDEO:
925 ctx->cuparseinfo.CodecType = cudaVideoCodec_MPEG1;
928 #if CONFIG_MPEG2_CUVID_DECODER
929 case AV_CODEC_ID_MPEG2VIDEO:
930 ctx->cuparseinfo.CodecType = cudaVideoCodec_MPEG2;
933 #if CONFIG_MPEG4_CUVID_DECODER
934 case AV_CODEC_ID_MPEG4:
935 ctx->cuparseinfo.CodecType = cudaVideoCodec_MPEG4;
938 #if CONFIG_VP8_CUVID_DECODER
939 case AV_CODEC_ID_VP8:
940 ctx->cuparseinfo.CodecType = cudaVideoCodec_VP8;
943 #if CONFIG_VP9_CUVID_DECODER
944 case AV_CODEC_ID_VP9:
945 ctx->cuparseinfo.CodecType = cudaVideoCodec_VP9;
948 #if CONFIG_VC1_CUVID_DECODER
949 case AV_CODEC_ID_VC1:
950 ctx->cuparseinfo.CodecType = cudaVideoCodec_VC1;
954 av_log(avctx, AV_LOG_ERROR, "Invalid CUVID codec!\n");
958 if (avctx->codec->id == AV_CODEC_ID_H264 || avctx->codec->id == AV_CODEC_ID_HEVC) {
959 if (avctx->codec->id == AV_CODEC_ID_H264)
960 bsf = av_bsf_get_by_name("h264_mp4toannexb");
962 bsf = av_bsf_get_by_name("hevc_mp4toannexb");
965 ret = AVERROR_BSF_NOT_FOUND;
968 if (ret = av_bsf_alloc(bsf, &ctx->bsf)) {
971 if (((ret = avcodec_parameters_from_context(ctx->bsf->par_in, avctx)) < 0) || ((ret = av_bsf_init(ctx->bsf)) < 0)) {
972 av_bsf_free(&ctx->bsf);
976 ctx->cuparse_ext.format.seqhdr_data_length = ctx->bsf->par_out->extradata_size;
977 memcpy(ctx->cuparse_ext.raw_seqhdr_data,
978 ctx->bsf->par_out->extradata,
979 FFMIN(sizeof(ctx->cuparse_ext.raw_seqhdr_data), ctx->bsf->par_out->extradata_size));
980 } else if (avctx->extradata_size > 0) {
981 ctx->cuparse_ext.format.seqhdr_data_length = avctx->extradata_size;
982 memcpy(ctx->cuparse_ext.raw_seqhdr_data,
984 FFMIN(sizeof(ctx->cuparse_ext.raw_seqhdr_data), avctx->extradata_size));
987 ctx->key_frame = av_mallocz(ctx->nb_surfaces * sizeof(int));
988 if (!ctx->key_frame) {
989 ret = AVERROR(ENOMEM);
993 ctx->cuparseinfo.ulMaxNumDecodeSurfaces = ctx->nb_surfaces;
994 ctx->cuparseinfo.ulMaxDisplayDelay = 4;
995 ctx->cuparseinfo.pUserData = avctx;
996 ctx->cuparseinfo.pfnSequenceCallback = cuvid_handle_video_sequence;
997 ctx->cuparseinfo.pfnDecodePicture = cuvid_handle_picture_decode;
998 ctx->cuparseinfo.pfnDisplayPicture = cuvid_handle_picture_display;
1000 ret = CHECK_CU(ctx->cudl->cuCtxPushCurrent(cuda_ctx));
1004 ret = cuvid_test_capabilities(avctx, &ctx->cuparseinfo,
1011 ret = CHECK_CU(ctx->cvdl->cuvidCreateVideoParser(&ctx->cuparser, &ctx->cuparseinfo));
1015 seq_pkt.payload = ctx->cuparse_ext.raw_seqhdr_data;
1016 seq_pkt.payload_size = ctx->cuparse_ext.format.seqhdr_data_length;
1018 if (seq_pkt.payload && seq_pkt.payload_size) {
1019 ret = CHECK_CU(ctx->cvdl->cuvidParseVideoData(ctx->cuparser, &seq_pkt));
1024 ret = CHECK_CU(ctx->cudl->cuCtxPopCurrent(&dummy));
1028 ctx->prev_pts = INT64_MIN;
1030 if (!avctx->pkt_timebase.num || !avctx->pkt_timebase.den)
1031 av_log(avctx, AV_LOG_WARNING, "Invalid pkt_timebase, passing timestamps as-is.\n");
1036 cuvid_decode_end(avctx);
1040 static void cuvid_flush(AVCodecContext *avctx)
1042 CuvidContext *ctx = avctx->priv_data;
1043 AVHWDeviceContext *device_ctx = (AVHWDeviceContext*)ctx->hwdevice->data;
1044 AVCUDADeviceContext *device_hwctx = device_ctx->hwctx;
1045 CUcontext dummy, cuda_ctx = device_hwctx->cuda_ctx;
1046 CUVIDSOURCEDATAPACKET seq_pkt = { 0 };
1049 ret = CHECK_CU(ctx->cudl->cuCtxPushCurrent(cuda_ctx));
1053 av_fifo_freep(&ctx->frame_queue);
1055 ctx->frame_queue = av_fifo_alloc(ctx->nb_surfaces * sizeof(CuvidParsedFrame));
1056 if (!ctx->frame_queue) {
1057 av_log(avctx, AV_LOG_ERROR, "Failed to recreate frame queue on flush\n");
1061 if (ctx->cudecoder) {
1062 ctx->cvdl->cuvidDestroyDecoder(ctx->cudecoder);
1063 ctx->cudecoder = NULL;
1066 if (ctx->cuparser) {
1067 ctx->cvdl->cuvidDestroyVideoParser(ctx->cuparser);
1068 ctx->cuparser = NULL;
1071 ret = CHECK_CU(ctx->cvdl->cuvidCreateVideoParser(&ctx->cuparser, &ctx->cuparseinfo));
1075 seq_pkt.payload = ctx->cuparse_ext.raw_seqhdr_data;
1076 seq_pkt.payload_size = ctx->cuparse_ext.format.seqhdr_data_length;
1078 if (seq_pkt.payload && seq_pkt.payload_size) {
1079 ret = CHECK_CU(ctx->cvdl->cuvidParseVideoData(ctx->cuparser, &seq_pkt));
1084 ret = CHECK_CU(ctx->cudl->cuCtxPopCurrent(&dummy));
1088 ctx->prev_pts = INT64_MIN;
1089 ctx->decoder_flushing = 0;
1093 av_log(avctx, AV_LOG_ERROR, "CUDA reinit on flush failed\n");
1096 #define OFFSET(x) offsetof(CuvidContext, x)
1097 #define VD AV_OPT_FLAG_VIDEO_PARAM | AV_OPT_FLAG_DECODING_PARAM
1098 static const AVOption options[] = {
1099 { "deint", "Set deinterlacing mode", OFFSET(deint_mode), AV_OPT_TYPE_INT, { .i64 = cudaVideoDeinterlaceMode_Weave }, cudaVideoDeinterlaceMode_Weave, cudaVideoDeinterlaceMode_Adaptive, VD, "deint" },
1100 { "weave", "Weave deinterlacing (do nothing)", 0, AV_OPT_TYPE_CONST, { .i64 = cudaVideoDeinterlaceMode_Weave }, 0, 0, VD, "deint" },
1101 { "bob", "Bob deinterlacing", 0, AV_OPT_TYPE_CONST, { .i64 = cudaVideoDeinterlaceMode_Bob }, 0, 0, VD, "deint" },
1102 { "adaptive", "Adaptive deinterlacing", 0, AV_OPT_TYPE_CONST, { .i64 = cudaVideoDeinterlaceMode_Adaptive }, 0, 0, VD, "deint" },
1103 { "gpu", "GPU to be used for decoding", OFFSET(cu_gpu), AV_OPT_TYPE_STRING, { .str = NULL }, 0, 0, VD },
1104 { "surfaces", "Maximum surfaces to be used for decoding", OFFSET(nb_surfaces), AV_OPT_TYPE_INT, { .i64 = 25 }, 0, INT_MAX, VD },
1105 { "drop_second_field", "Drop second field when deinterlacing", OFFSET(drop_second_field), AV_OPT_TYPE_BOOL, { .i64 = 0 }, 0, 1, VD },
1106 { "crop", "Crop (top)x(bottom)x(left)x(right)", OFFSET(crop_expr), AV_OPT_TYPE_STRING, { .str = NULL }, 0, 0, VD },
1107 { "resize", "Resize (width)x(height)", OFFSET(resize_expr), AV_OPT_TYPE_STRING, { .str = NULL }, 0, 0, VD },
1111 static const AVCodecHWConfigInternal *cuvid_hw_configs[] = {
1112 &(const AVCodecHWConfigInternal) {
1114 .pix_fmt = AV_PIX_FMT_CUDA,
1115 .methods = AV_CODEC_HW_CONFIG_METHOD_HW_DEVICE_CTX |
1116 AV_CODEC_HW_CONFIG_METHOD_INTERNAL,
1117 .device_type = AV_HWDEVICE_TYPE_CUDA
1124 #define DEFINE_CUVID_CODEC(x, X) \
1125 static const AVClass x##_cuvid_class = { \
1126 .class_name = #x "_cuvid", \
1127 .item_name = av_default_item_name, \
1128 .option = options, \
1129 .version = LIBAVUTIL_VERSION_INT, \
1131 AVCodec ff_##x##_cuvid_decoder = { \
1132 .name = #x "_cuvid", \
1133 .long_name = NULL_IF_CONFIG_SMALL("Nvidia CUVID " #X " decoder"), \
1134 .type = AVMEDIA_TYPE_VIDEO, \
1135 .id = AV_CODEC_ID_##X, \
1136 .priv_data_size = sizeof(CuvidContext), \
1137 .priv_class = &x##_cuvid_class, \
1138 .init = cuvid_decode_init, \
1139 .close = cuvid_decode_end, \
1140 .decode = cuvid_decode_frame, \
1141 .receive_frame = cuvid_output_frame, \
1142 .flush = cuvid_flush, \
1143 .capabilities = AV_CODEC_CAP_DELAY | AV_CODEC_CAP_AVOID_PROBING | AV_CODEC_CAP_HARDWARE, \
1144 .pix_fmts = (const enum AVPixelFormat[]){ AV_PIX_FMT_CUDA, \
1148 AV_PIX_FMT_NONE }, \
1149 .hw_configs = cuvid_hw_configs, \
1150 .wrapper_name = "cuvid", \
1153 #if CONFIG_HEVC_CUVID_DECODER
1154 DEFINE_CUVID_CODEC(hevc, HEVC)
1157 #if CONFIG_H264_CUVID_DECODER
1158 DEFINE_CUVID_CODEC(h264, H264)
1161 #if CONFIG_MJPEG_CUVID_DECODER
1162 DEFINE_CUVID_CODEC(mjpeg, MJPEG)
1165 #if CONFIG_MPEG1_CUVID_DECODER
1166 DEFINE_CUVID_CODEC(mpeg1, MPEG1VIDEO)
1169 #if CONFIG_MPEG2_CUVID_DECODER
1170 DEFINE_CUVID_CODEC(mpeg2, MPEG2VIDEO)
1173 #if CONFIG_MPEG4_CUVID_DECODER
1174 DEFINE_CUVID_CODEC(mpeg4, MPEG4)
1177 #if CONFIG_VP8_CUVID_DECODER
1178 DEFINE_CUVID_CODEC(vp8, VP8)
1181 #if CONFIG_VP9_CUVID_DECODER
1182 DEFINE_CUVID_CODEC(vp9, VP9)
1185 #if CONFIG_VC1_CUVID_DECODER
1186 DEFINE_CUVID_CODEC(vc1, VC1)