3 * Copyright (c) 2016 Timo Rothenpieler <timo@rothenpieler.org>
5 * This file is part of FFmpeg.
7 * FFmpeg is free software; you can redistribute it and/or
8 * modify it under the terms of the GNU Lesser General Public
9 * License as published by the Free Software Foundation; either
10 * version 2.1 of the License, or (at your option) any later version.
12 * FFmpeg is distributed in the hope that it will be useful,
13 * but WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * Lesser General Public License for more details.
17 * You should have received a copy of the GNU Lesser General Public
18 * License along with FFmpeg; if not, write to the Free Software
19 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
22 #include "compat/cuda/dynlink_loader.h"
24 #include "libavutil/buffer.h"
25 #include "libavutil/mathematics.h"
26 #include "libavutil/hwcontext.h"
27 #include "libavutil/hwcontext_cuda_internal.h"
28 #include "libavutil/cuda_check.h"
29 #include "libavutil/fifo.h"
30 #include "libavutil/log.h"
31 #include "libavutil/opt.h"
32 #include "libavutil/pixdesc.h"
40 #if !NVDECAPI_CHECK_VERSION(9, 0)
41 #define cudaVideoSurfaceFormat_YUV444 2
42 #define cudaVideoSurfaceFormat_YUV444_16Bit 3
45 #if NVDECAPI_CHECK_VERSION(11, 0)
46 #define CUVID_HAS_AV1_SUPPORT
49 typedef struct CuvidContext
53 CUvideodecoder cudecoder;
54 CUvideoparser cuparser;
58 int drop_second_field;
74 AVBufferRef *hwdevice;
77 AVFifoBuffer *frame_queue;
80 int deint_mode_current;
82 int progressive_sequence;
89 cudaVideoCodec codec_type;
90 cudaVideoChromaFormat chroma_format;
92 CUVIDDECODECAPS caps8, caps10, caps12;
94 CUVIDPARSERPARAMS cuparseinfo;
95 CUVIDEOFORMATEX *cuparse_ext;
101 typedef struct CuvidParsedFrame
103 CUVIDPARSERDISPINFO dispinfo;
105 int is_deinterlacing;
108 #define CHECK_CU(x) FF_CUDA_CHECK_DL(avctx, ctx->cudl, x)
110 static int CUDAAPI cuvid_handle_video_sequence(void *opaque, CUVIDEOFORMAT* format)
112 AVCodecContext *avctx = opaque;
113 CuvidContext *ctx = avctx->priv_data;
114 AVHWFramesContext *hwframe_ctx = (AVHWFramesContext*)ctx->hwframe->data;
115 CUVIDDECODECAPS *caps = NULL;
116 CUVIDDECODECREATEINFO cuinfo;
120 int old_width = avctx->width;
121 int old_height = avctx->height;
123 enum AVPixelFormat pix_fmts[3] = { AV_PIX_FMT_CUDA,
124 AV_PIX_FMT_NONE, // Will be updated below
127 av_log(avctx, AV_LOG_TRACE, "pfnSequenceCallback, progressive_sequence=%d\n", format->progressive_sequence);
129 memset(&cuinfo, 0, sizeof(cuinfo));
131 ctx->internal_error = 0;
133 avctx->coded_width = cuinfo.ulWidth = format->coded_width;
134 avctx->coded_height = cuinfo.ulHeight = format->coded_height;
137 cuinfo.display_area.left = format->display_area.left + ctx->crop.left;
138 cuinfo.display_area.top = format->display_area.top + ctx->crop.top;
139 cuinfo.display_area.right = format->display_area.right - ctx->crop.right;
140 cuinfo.display_area.bottom = format->display_area.bottom - ctx->crop.bottom;
142 // width and height need to be set before calling ff_get_format
143 if (ctx->resize_expr) {
144 avctx->width = ctx->resize.width;
145 avctx->height = ctx->resize.height;
147 avctx->width = cuinfo.display_area.right - cuinfo.display_area.left;
148 avctx->height = cuinfo.display_area.bottom - cuinfo.display_area.top;
151 // target width/height need to be multiples of two
152 cuinfo.ulTargetWidth = avctx->width = (avctx->width + 1) & ~1;
153 cuinfo.ulTargetHeight = avctx->height = (avctx->height + 1) & ~1;
155 // aspect ratio conversion, 1:1, depends on scaled resolution
156 cuinfo.target_rect.left = 0;
157 cuinfo.target_rect.top = 0;
158 cuinfo.target_rect.right = cuinfo.ulTargetWidth;
159 cuinfo.target_rect.bottom = cuinfo.ulTargetHeight;
161 chroma_444 = format->chroma_format == cudaVideoChromaFormat_444;
163 switch (format->bit_depth_luma_minus8) {
165 pix_fmts[1] = chroma_444 ? AV_PIX_FMT_YUV444P : AV_PIX_FMT_NV12;
169 pix_fmts[1] = chroma_444 ? AV_PIX_FMT_YUV444P16 : AV_PIX_FMT_P010;
173 pix_fmts[1] = chroma_444 ? AV_PIX_FMT_YUV444P16 : AV_PIX_FMT_P016;
180 if (!caps || !caps->bIsSupported) {
181 av_log(avctx, AV_LOG_ERROR, "unsupported bit depth: %d\n",
182 format->bit_depth_luma_minus8 + 8);
183 ctx->internal_error = AVERROR(EINVAL);
187 surface_fmt = ff_get_format(avctx, pix_fmts);
188 if (surface_fmt < 0) {
189 av_log(avctx, AV_LOG_ERROR, "ff_get_format failed: %d\n", surface_fmt);
190 ctx->internal_error = AVERROR(EINVAL);
194 av_log(avctx, AV_LOG_VERBOSE, "Formats: Original: %s | HW: %s | SW: %s\n",
195 av_get_pix_fmt_name(avctx->pix_fmt),
196 av_get_pix_fmt_name(surface_fmt),
197 av_get_pix_fmt_name(avctx->sw_pix_fmt));
199 avctx->pix_fmt = surface_fmt;
201 // Update our hwframe ctx, as the get_format callback might have refreshed it!
202 if (avctx->hw_frames_ctx) {
203 av_buffer_unref(&ctx->hwframe);
205 ctx->hwframe = av_buffer_ref(avctx->hw_frames_ctx);
207 ctx->internal_error = AVERROR(ENOMEM);
211 hwframe_ctx = (AVHWFramesContext*)ctx->hwframe->data;
214 ff_set_sar(avctx, av_div_q(
215 (AVRational){ format->display_aspect_ratio.x, format->display_aspect_ratio.y },
216 (AVRational){ avctx->width, avctx->height }));
218 ctx->deint_mode_current = format->progressive_sequence
219 ? cudaVideoDeinterlaceMode_Weave
222 ctx->progressive_sequence = format->progressive_sequence;
224 if (!format->progressive_sequence && ctx->deint_mode_current == cudaVideoDeinterlaceMode_Weave)
225 avctx->flags |= AV_CODEC_FLAG_INTERLACED_DCT;
227 avctx->flags &= ~AV_CODEC_FLAG_INTERLACED_DCT;
229 if (format->video_signal_description.video_full_range_flag)
230 avctx->color_range = AVCOL_RANGE_JPEG;
232 avctx->color_range = AVCOL_RANGE_MPEG;
234 avctx->color_primaries = format->video_signal_description.color_primaries;
235 avctx->color_trc = format->video_signal_description.transfer_characteristics;
236 avctx->colorspace = format->video_signal_description.matrix_coefficients;
239 avctx->bit_rate = format->bitrate;
241 if (format->frame_rate.numerator && format->frame_rate.denominator) {
242 avctx->framerate.num = format->frame_rate.numerator;
243 avctx->framerate.den = format->frame_rate.denominator;
247 && avctx->coded_width == format->coded_width
248 && avctx->coded_height == format->coded_height
249 && avctx->width == old_width
250 && avctx->height == old_height
251 && ctx->chroma_format == format->chroma_format
252 && ctx->codec_type == format->codec)
255 if (ctx->cudecoder) {
256 av_log(avctx, AV_LOG_TRACE, "Re-initializing decoder\n");
257 ctx->internal_error = CHECK_CU(ctx->cvdl->cuvidDestroyDecoder(ctx->cudecoder));
258 if (ctx->internal_error < 0)
260 ctx->cudecoder = NULL;
263 if (hwframe_ctx->pool && (
264 hwframe_ctx->width < avctx->width ||
265 hwframe_ctx->height < avctx->height ||
266 hwframe_ctx->format != AV_PIX_FMT_CUDA ||
267 hwframe_ctx->sw_format != avctx->sw_pix_fmt)) {
268 av_log(avctx, AV_LOG_ERROR, "AVHWFramesContext is already initialized with incompatible parameters\n");
269 av_log(avctx, AV_LOG_DEBUG, "width: %d <-> %d\n", hwframe_ctx->width, avctx->width);
270 av_log(avctx, AV_LOG_DEBUG, "height: %d <-> %d\n", hwframe_ctx->height, avctx->height);
271 av_log(avctx, AV_LOG_DEBUG, "format: %s <-> cuda\n", av_get_pix_fmt_name(hwframe_ctx->format));
272 av_log(avctx, AV_LOG_DEBUG, "sw_format: %s <-> %s\n",
273 av_get_pix_fmt_name(hwframe_ctx->sw_format), av_get_pix_fmt_name(avctx->sw_pix_fmt));
274 ctx->internal_error = AVERROR(EINVAL);
278 ctx->chroma_format = format->chroma_format;
280 cuinfo.CodecType = ctx->codec_type = format->codec;
281 cuinfo.ChromaFormat = format->chroma_format;
283 switch (avctx->sw_pix_fmt) {
284 case AV_PIX_FMT_NV12:
285 cuinfo.OutputFormat = cudaVideoSurfaceFormat_NV12;
287 case AV_PIX_FMT_P010:
288 case AV_PIX_FMT_P016:
289 cuinfo.OutputFormat = cudaVideoSurfaceFormat_P016;
291 case AV_PIX_FMT_YUV444P:
292 cuinfo.OutputFormat = cudaVideoSurfaceFormat_YUV444;
294 case AV_PIX_FMT_YUV444P16:
295 cuinfo.OutputFormat = cudaVideoSurfaceFormat_YUV444_16Bit;
298 av_log(avctx, AV_LOG_ERROR, "Unsupported output format: %s\n",
299 av_get_pix_fmt_name(avctx->sw_pix_fmt));
300 ctx->internal_error = AVERROR(EINVAL);
304 cuinfo.ulNumDecodeSurfaces = ctx->nb_surfaces;
305 cuinfo.ulNumOutputSurfaces = 1;
306 cuinfo.ulCreationFlags = cudaVideoCreate_PreferCUVID;
307 cuinfo.bitDepthMinus8 = format->bit_depth_luma_minus8;
308 cuinfo.DeinterlaceMode = ctx->deint_mode_current;
310 if (ctx->deint_mode_current != cudaVideoDeinterlaceMode_Weave && !ctx->drop_second_field)
311 avctx->framerate = av_mul_q(avctx->framerate, (AVRational){2, 1});
313 ctx->internal_error = CHECK_CU(ctx->cvdl->cuvidCreateDecoder(&ctx->cudecoder, &cuinfo));
314 if (ctx->internal_error < 0)
317 if (!hwframe_ctx->pool) {
318 hwframe_ctx->format = AV_PIX_FMT_CUDA;
319 hwframe_ctx->sw_format = avctx->sw_pix_fmt;
320 hwframe_ctx->width = avctx->width;
321 hwframe_ctx->height = avctx->height;
323 if ((ctx->internal_error = av_hwframe_ctx_init(ctx->hwframe)) < 0) {
324 av_log(avctx, AV_LOG_ERROR, "av_hwframe_ctx_init failed\n");
332 static int CUDAAPI cuvid_handle_picture_decode(void *opaque, CUVIDPICPARAMS* picparams)
334 AVCodecContext *avctx = opaque;
335 CuvidContext *ctx = avctx->priv_data;
337 av_log(avctx, AV_LOG_TRACE, "pfnDecodePicture\n");
339 ctx->key_frame[picparams->CurrPicIdx] = picparams->intra_pic_flag;
341 ctx->internal_error = CHECK_CU(ctx->cvdl->cuvidDecodePicture(ctx->cudecoder, picparams));
342 if (ctx->internal_error < 0)
348 static int CUDAAPI cuvid_handle_picture_display(void *opaque, CUVIDPARSERDISPINFO* dispinfo)
350 AVCodecContext *avctx = opaque;
351 CuvidContext *ctx = avctx->priv_data;
352 CuvidParsedFrame parsed_frame = { { 0 } };
354 parsed_frame.dispinfo = *dispinfo;
355 ctx->internal_error = 0;
357 // For some reason, dispinfo->progressive_frame is sometimes wrong.
358 parsed_frame.dispinfo.progressive_frame = ctx->progressive_sequence;
360 if (ctx->deint_mode_current == cudaVideoDeinterlaceMode_Weave) {
361 av_fifo_generic_write(ctx->frame_queue, &parsed_frame, sizeof(CuvidParsedFrame), NULL);
363 parsed_frame.is_deinterlacing = 1;
364 av_fifo_generic_write(ctx->frame_queue, &parsed_frame, sizeof(CuvidParsedFrame), NULL);
365 if (!ctx->drop_second_field) {
366 parsed_frame.second_field = 1;
367 av_fifo_generic_write(ctx->frame_queue, &parsed_frame, sizeof(CuvidParsedFrame), NULL);
374 static int cuvid_is_buffer_full(AVCodecContext *avctx)
376 CuvidContext *ctx = avctx->priv_data;
378 int delay = ctx->cuparseinfo.ulMaxDisplayDelay;
379 if (ctx->deint_mode != cudaVideoDeinterlaceMode_Weave && !ctx->drop_second_field)
382 return (av_fifo_size(ctx->frame_queue) / sizeof(CuvidParsedFrame)) + delay >= ctx->nb_surfaces;
385 static int cuvid_decode_packet(AVCodecContext *avctx, const AVPacket *avpkt)
387 CuvidContext *ctx = avctx->priv_data;
388 AVHWDeviceContext *device_ctx = (AVHWDeviceContext*)ctx->hwdevice->data;
389 AVCUDADeviceContext *device_hwctx = device_ctx->hwctx;
390 CUcontext dummy, cuda_ctx = device_hwctx->cuda_ctx;
391 CUVIDSOURCEDATAPACKET cupkt;
392 int ret = 0, eret = 0, is_flush = ctx->decoder_flushing;
394 av_log(avctx, AV_LOG_TRACE, "cuvid_decode_packet\n");
396 if (is_flush && avpkt && avpkt->size)
399 if (cuvid_is_buffer_full(avctx) && avpkt && avpkt->size)
400 return AVERROR(EAGAIN);
402 ret = CHECK_CU(ctx->cudl->cuCtxPushCurrent(cuda_ctx));
407 memset(&cupkt, 0, sizeof(cupkt));
409 if (avpkt && avpkt->size) {
410 cupkt.payload_size = avpkt->size;
411 cupkt.payload = avpkt->data;
413 if (avpkt->pts != AV_NOPTS_VALUE) {
414 cupkt.flags = CUVID_PKT_TIMESTAMP;
415 if (avctx->pkt_timebase.num && avctx->pkt_timebase.den)
416 cupkt.timestamp = av_rescale_q(avpkt->pts, avctx->pkt_timebase, (AVRational){1, 10000000});
418 cupkt.timestamp = avpkt->pts;
421 cupkt.flags = CUVID_PKT_ENDOFSTREAM;
422 ctx->decoder_flushing = 1;
425 ret = CHECK_CU(ctx->cvdl->cuvidParseVideoData(ctx->cuparser, &cupkt));
430 // cuvidParseVideoData doesn't return an error just because stuff failed...
431 if (ctx->internal_error) {
432 av_log(avctx, AV_LOG_ERROR, "cuvid decode callback error\n");
433 ret = ctx->internal_error;
438 eret = CHECK_CU(ctx->cudl->cuCtxPopCurrent(&dummy));
450 static int cuvid_output_frame(AVCodecContext *avctx, AVFrame *frame)
452 CuvidContext *ctx = avctx->priv_data;
453 AVHWDeviceContext *device_ctx = (AVHWDeviceContext*)ctx->hwdevice->data;
454 AVCUDADeviceContext *device_hwctx = device_ctx->hwctx;
455 CUcontext dummy, cuda_ctx = device_hwctx->cuda_ctx;
456 CUdeviceptr mapped_frame = 0;
457 int ret = 0, eret = 0;
459 av_log(avctx, AV_LOG_TRACE, "cuvid_output_frame\n");
461 if (ctx->decoder_flushing) {
462 ret = cuvid_decode_packet(avctx, NULL);
463 if (ret < 0 && ret != AVERROR_EOF)
467 if (!cuvid_is_buffer_full(avctx)) {
469 ret = ff_decode_get_packet(avctx, &pkt);
470 if (ret < 0 && ret != AVERROR_EOF)
472 ret = cuvid_decode_packet(avctx, &pkt);
473 av_packet_unref(&pkt);
474 // cuvid_is_buffer_full() should avoid this.
475 if (ret == AVERROR(EAGAIN))
476 ret = AVERROR_EXTERNAL;
477 if (ret < 0 && ret != AVERROR_EOF)
481 ret = CHECK_CU(ctx->cudl->cuCtxPushCurrent(cuda_ctx));
485 if (av_fifo_size(ctx->frame_queue)) {
486 const AVPixFmtDescriptor *pixdesc;
487 CuvidParsedFrame parsed_frame;
488 CUVIDPROCPARAMS params;
489 unsigned int pitch = 0;
493 av_fifo_generic_read(ctx->frame_queue, &parsed_frame, sizeof(CuvidParsedFrame), NULL);
495 memset(¶ms, 0, sizeof(params));
496 params.progressive_frame = parsed_frame.dispinfo.progressive_frame;
497 params.second_field = parsed_frame.second_field;
498 params.top_field_first = parsed_frame.dispinfo.top_field_first;
500 ret = CHECK_CU(ctx->cvdl->cuvidMapVideoFrame(ctx->cudecoder, parsed_frame.dispinfo.picture_index, &mapped_frame, &pitch, ¶ms));
504 if (avctx->pix_fmt == AV_PIX_FMT_CUDA) {
505 ret = av_hwframe_get_buffer(ctx->hwframe, frame, 0);
507 av_log(avctx, AV_LOG_ERROR, "av_hwframe_get_buffer failed\n");
511 ret = ff_decode_frame_props(avctx, frame);
513 av_log(avctx, AV_LOG_ERROR, "ff_decode_frame_props failed\n");
517 pixdesc = av_pix_fmt_desc_get(avctx->sw_pix_fmt);
519 for (i = 0; i < pixdesc->nb_components; i++) {
520 int height = avctx->height >> (i ? pixdesc->log2_chroma_h : 0);
521 CUDA_MEMCPY2D cpy = {
522 .srcMemoryType = CU_MEMORYTYPE_DEVICE,
523 .dstMemoryType = CU_MEMORYTYPE_DEVICE,
524 .srcDevice = mapped_frame,
525 .dstDevice = (CUdeviceptr)frame->data[i],
527 .dstPitch = frame->linesize[i],
529 .WidthInBytes = FFMIN(pitch, frame->linesize[i]),
533 ret = CHECK_CU(ctx->cudl->cuMemcpy2DAsync(&cpy, device_hwctx->stream));
539 } else if (avctx->pix_fmt == AV_PIX_FMT_NV12 ||
540 avctx->pix_fmt == AV_PIX_FMT_P010 ||
541 avctx->pix_fmt == AV_PIX_FMT_P016 ||
542 avctx->pix_fmt == AV_PIX_FMT_YUV444P ||
543 avctx->pix_fmt == AV_PIX_FMT_YUV444P16) {
544 unsigned int offset = 0;
545 AVFrame *tmp_frame = av_frame_alloc();
547 av_log(avctx, AV_LOG_ERROR, "av_frame_alloc failed\n");
548 ret = AVERROR(ENOMEM);
552 pixdesc = av_pix_fmt_desc_get(avctx->sw_pix_fmt);
554 tmp_frame->format = AV_PIX_FMT_CUDA;
555 tmp_frame->hw_frames_ctx = av_buffer_ref(ctx->hwframe);
556 tmp_frame->width = avctx->width;
557 tmp_frame->height = avctx->height;
560 * Note that the following logic would not work for three plane
561 * YUV420 because the pitch value is different for the chroma
564 for (i = 0; i < pixdesc->nb_components; i++) {
565 tmp_frame->data[i] = (uint8_t*)mapped_frame + offset;
566 tmp_frame->linesize[i] = pitch;
567 offset += pitch * (avctx->height >> (i ? pixdesc->log2_chroma_h : 0));
570 ret = ff_get_buffer(avctx, frame, 0);
572 av_log(avctx, AV_LOG_ERROR, "ff_get_buffer failed\n");
573 av_frame_free(&tmp_frame);
577 ret = av_hwframe_transfer_data(frame, tmp_frame, 0);
579 av_log(avctx, AV_LOG_ERROR, "av_hwframe_transfer_data failed\n");
580 av_frame_free(&tmp_frame);
583 av_frame_free(&tmp_frame);
589 frame->key_frame = ctx->key_frame[parsed_frame.dispinfo.picture_index];
590 frame->width = avctx->width;
591 frame->height = avctx->height;
592 if (avctx->pkt_timebase.num && avctx->pkt_timebase.den)
593 frame->pts = av_rescale_q(parsed_frame.dispinfo.timestamp, (AVRational){1, 10000000}, avctx->pkt_timebase);
595 frame->pts = parsed_frame.dispinfo.timestamp;
597 if (parsed_frame.second_field) {
598 if (ctx->prev_pts == INT64_MIN) {
599 ctx->prev_pts = frame->pts;
600 frame->pts += (avctx->pkt_timebase.den * avctx->framerate.den) / (avctx->pkt_timebase.num * avctx->framerate.num);
602 int pts_diff = (frame->pts - ctx->prev_pts) / 2;
603 ctx->prev_pts = frame->pts;
604 frame->pts += pts_diff;
608 /* CUVIDs opaque reordering breaks the internal pkt logic.
609 * So set pkt_pts and clear all the other pkt_ fields.
612 FF_DISABLE_DEPRECATION_WARNINGS
613 frame->pkt_pts = frame->pts;
614 FF_ENABLE_DEPRECATION_WARNINGS
617 frame->pkt_duration = 0;
618 frame->pkt_size = -1;
620 frame->interlaced_frame = !parsed_frame.is_deinterlacing && !parsed_frame.dispinfo.progressive_frame;
622 if (frame->interlaced_frame)
623 frame->top_field_first = parsed_frame.dispinfo.top_field_first;
624 } else if (ctx->decoder_flushing) {
627 ret = AVERROR(EAGAIN);
632 eret = CHECK_CU(ctx->cvdl->cuvidUnmapVideoFrame(ctx->cudecoder, mapped_frame));
634 eret = CHECK_CU(ctx->cudl->cuCtxPopCurrent(&dummy));
642 static int cuvid_decode_frame(AVCodecContext *avctx, void *data, int *got_frame, AVPacket *avpkt)
644 CuvidContext *ctx = avctx->priv_data;
645 AVFrame *frame = data;
648 av_log(avctx, AV_LOG_TRACE, "cuvid_decode_frame\n");
650 if (ctx->deint_mode_current != cudaVideoDeinterlaceMode_Weave) {
651 av_log(avctx, AV_LOG_ERROR, "Deinterlacing is not supported via the old API\n");
652 return AVERROR(EINVAL);
655 if (!ctx->decoder_flushing) {
656 ret = cuvid_decode_packet(avctx, avpkt);
661 ret = cuvid_output_frame(avctx, frame);
662 if (ret == AVERROR(EAGAIN) || ret == AVERROR_EOF) {
664 } else if (ret < 0) {
673 static av_cold int cuvid_decode_end(AVCodecContext *avctx)
675 CuvidContext *ctx = avctx->priv_data;
676 AVHWDeviceContext *device_ctx = (AVHWDeviceContext *)ctx->hwdevice->data;
677 AVCUDADeviceContext *device_hwctx = device_ctx->hwctx;
678 CUcontext dummy, cuda_ctx = device_hwctx->cuda_ctx;
680 av_fifo_freep(&ctx->frame_queue);
682 ctx->cudl->cuCtxPushCurrent(cuda_ctx);
685 ctx->cvdl->cuvidDestroyVideoParser(ctx->cuparser);
688 ctx->cvdl->cuvidDestroyDecoder(ctx->cudecoder);
690 ctx->cudl->cuCtxPopCurrent(&dummy);
694 av_buffer_unref(&ctx->hwframe);
695 av_buffer_unref(&ctx->hwdevice);
697 av_freep(&ctx->key_frame);
698 av_freep(&ctx->cuparse_ext);
700 cuvid_free_functions(&ctx->cvdl);
705 static int cuvid_test_capabilities(AVCodecContext *avctx,
706 const CUVIDPARSERPARAMS *cuparseinfo,
711 CuvidContext *ctx = avctx->priv_data;
712 CUVIDDECODECAPS *caps;
713 int res8 = 0, res10 = 0, res12 = 0;
715 if (!ctx->cvdl->cuvidGetDecoderCaps) {
716 av_log(avctx, AV_LOG_WARNING, "Used Nvidia driver is too old to perform a capability check.\n");
717 av_log(avctx, AV_LOG_WARNING, "The minimum required version is "
718 #if defined(_WIN32) || defined(__CYGWIN__)
723 ". Continuing blind.\n");
724 ctx->caps8.bIsSupported = ctx->caps10.bIsSupported = 1;
725 // 12 bit was not supported before the capability check was introduced, so disable it.
726 ctx->caps12.bIsSupported = 0;
730 ctx->caps8.eCodecType = ctx->caps10.eCodecType = ctx->caps12.eCodecType
731 = cuparseinfo->CodecType;
732 ctx->caps8.eChromaFormat = ctx->caps10.eChromaFormat = ctx->caps12.eChromaFormat
733 = cudaVideoChromaFormat_420;
735 ctx->caps8.nBitDepthMinus8 = 0;
736 ctx->caps10.nBitDepthMinus8 = 2;
737 ctx->caps12.nBitDepthMinus8 = 4;
739 res8 = CHECK_CU(ctx->cvdl->cuvidGetDecoderCaps(&ctx->caps8));
740 res10 = CHECK_CU(ctx->cvdl->cuvidGetDecoderCaps(&ctx->caps10));
741 res12 = CHECK_CU(ctx->cvdl->cuvidGetDecoderCaps(&ctx->caps12));
743 av_log(avctx, AV_LOG_VERBOSE, "CUVID capabilities for %s:\n", avctx->codec->name);
744 av_log(avctx, AV_LOG_VERBOSE, "8 bit: supported: %d, min_width: %d, max_width: %d, min_height: %d, max_height: %d\n",
745 ctx->caps8.bIsSupported, ctx->caps8.nMinWidth, ctx->caps8.nMaxWidth, ctx->caps8.nMinHeight, ctx->caps8.nMaxHeight);
746 av_log(avctx, AV_LOG_VERBOSE, "10 bit: supported: %d, min_width: %d, max_width: %d, min_height: %d, max_height: %d\n",
747 ctx->caps10.bIsSupported, ctx->caps10.nMinWidth, ctx->caps10.nMaxWidth, ctx->caps10.nMinHeight, ctx->caps10.nMaxHeight);
748 av_log(avctx, AV_LOG_VERBOSE, "12 bit: supported: %d, min_width: %d, max_width: %d, min_height: %d, max_height: %d\n",
749 ctx->caps12.bIsSupported, ctx->caps12.nMinWidth, ctx->caps12.nMaxWidth, ctx->caps12.nMinHeight, ctx->caps12.nMaxHeight);
768 if (!ctx->caps8.bIsSupported) {
769 av_log(avctx, AV_LOG_ERROR, "Codec %s is not supported.\n", avctx->codec->name);
770 return AVERROR(EINVAL);
773 if (!caps->bIsSupported) {
774 av_log(avctx, AV_LOG_ERROR, "Bit depth %d is not supported.\n", bit_depth);
775 return AVERROR(EINVAL);
778 if (probed_width > caps->nMaxWidth || probed_width < caps->nMinWidth) {
779 av_log(avctx, AV_LOG_ERROR, "Video width %d not within range from %d to %d\n",
780 probed_width, caps->nMinWidth, caps->nMaxWidth);
781 return AVERROR(EINVAL);
784 if (probed_height > caps->nMaxHeight || probed_height < caps->nMinHeight) {
785 av_log(avctx, AV_LOG_ERROR, "Video height %d not within range from %d to %d\n",
786 probed_height, caps->nMinHeight, caps->nMaxHeight);
787 return AVERROR(EINVAL);
790 if ((probed_width * probed_height) / 256 > caps->nMaxMBCount) {
791 av_log(avctx, AV_LOG_ERROR, "Video macroblock count %d exceeds maximum of %d\n",
792 (int)(probed_width * probed_height) / 256, caps->nMaxMBCount);
793 return AVERROR(EINVAL);
799 static av_cold int cuvid_decode_init(AVCodecContext *avctx)
801 CuvidContext *ctx = avctx->priv_data;
802 AVCUDADeviceContext *device_hwctx;
803 AVHWDeviceContext *device_ctx;
804 AVHWFramesContext *hwframe_ctx;
805 CUVIDSOURCEDATAPACKET seq_pkt;
806 CUcontext cuda_ctx = NULL;
812 enum AVPixelFormat pix_fmts[3] = { AV_PIX_FMT_CUDA,
816 int probed_width = avctx->coded_width ? avctx->coded_width : 1280;
817 int probed_height = avctx->coded_height ? avctx->coded_height : 720;
818 int probed_bit_depth = 8;
820 const AVPixFmtDescriptor *probe_desc = av_pix_fmt_desc_get(avctx->pix_fmt);
821 if (probe_desc && probe_desc->nb_components)
822 probed_bit_depth = probe_desc->comp[0].depth;
824 // Accelerated transcoding scenarios with 'ffmpeg' require that the
825 // pix_fmt be set to AV_PIX_FMT_CUDA early. The sw_pix_fmt, and the
826 // pix_fmt for non-accelerated transcoding, do not need to be correct
827 // but need to be set to something. We arbitrarily pick NV12.
828 ret = ff_get_format(avctx, pix_fmts);
830 av_log(avctx, AV_LOG_ERROR, "ff_get_format failed: %d\n", ret);
833 avctx->pix_fmt = ret;
835 if (ctx->resize_expr && sscanf(ctx->resize_expr, "%dx%d",
836 &ctx->resize.width, &ctx->resize.height) != 2) {
837 av_log(avctx, AV_LOG_ERROR, "Invalid resize expressions\n");
838 ret = AVERROR(EINVAL);
842 if (ctx->crop_expr && sscanf(ctx->crop_expr, "%dx%dx%dx%d",
843 &ctx->crop.top, &ctx->crop.bottom,
844 &ctx->crop.left, &ctx->crop.right) != 4) {
845 av_log(avctx, AV_LOG_ERROR, "Invalid cropping expressions\n");
846 ret = AVERROR(EINVAL);
850 ret = cuvid_load_functions(&ctx->cvdl, avctx);
852 av_log(avctx, AV_LOG_ERROR, "Failed loading nvcuvid.\n");
856 ctx->frame_queue = av_fifo_alloc(ctx->nb_surfaces * sizeof(CuvidParsedFrame));
857 if (!ctx->frame_queue) {
858 ret = AVERROR(ENOMEM);
862 if (avctx->hw_frames_ctx) {
863 ctx->hwframe = av_buffer_ref(avctx->hw_frames_ctx);
865 ret = AVERROR(ENOMEM);
869 hwframe_ctx = (AVHWFramesContext*)ctx->hwframe->data;
871 ctx->hwdevice = av_buffer_ref(hwframe_ctx->device_ref);
872 if (!ctx->hwdevice) {
873 ret = AVERROR(ENOMEM);
877 if (avctx->hw_device_ctx) {
878 ctx->hwdevice = av_buffer_ref(avctx->hw_device_ctx);
879 if (!ctx->hwdevice) {
880 ret = AVERROR(ENOMEM);
884 ret = av_hwdevice_ctx_create(&ctx->hwdevice, AV_HWDEVICE_TYPE_CUDA, ctx->cu_gpu, NULL, 0);
889 ctx->hwframe = av_hwframe_ctx_alloc(ctx->hwdevice);
891 av_log(avctx, AV_LOG_ERROR, "av_hwframe_ctx_alloc failed\n");
892 ret = AVERROR(ENOMEM);
896 hwframe_ctx = (AVHWFramesContext*)ctx->hwframe->data;
899 device_ctx = hwframe_ctx->device_ctx;
900 device_hwctx = device_ctx->hwctx;
902 cuda_ctx = device_hwctx->cuda_ctx;
903 ctx->cudl = device_hwctx->internal->cuda_dl;
905 memset(&ctx->cuparseinfo, 0, sizeof(ctx->cuparseinfo));
906 memset(&seq_pkt, 0, sizeof(seq_pkt));
908 switch (avctx->codec->id) {
909 #if CONFIG_H264_CUVID_DECODER
910 case AV_CODEC_ID_H264:
911 ctx->cuparseinfo.CodecType = cudaVideoCodec_H264;
914 #if CONFIG_HEVC_CUVID_DECODER
915 case AV_CODEC_ID_HEVC:
916 ctx->cuparseinfo.CodecType = cudaVideoCodec_HEVC;
919 #if CONFIG_MJPEG_CUVID_DECODER
920 case AV_CODEC_ID_MJPEG:
921 ctx->cuparseinfo.CodecType = cudaVideoCodec_JPEG;
924 #if CONFIG_MPEG1_CUVID_DECODER
925 case AV_CODEC_ID_MPEG1VIDEO:
926 ctx->cuparseinfo.CodecType = cudaVideoCodec_MPEG1;
929 #if CONFIG_MPEG2_CUVID_DECODER
930 case AV_CODEC_ID_MPEG2VIDEO:
931 ctx->cuparseinfo.CodecType = cudaVideoCodec_MPEG2;
934 #if CONFIG_MPEG4_CUVID_DECODER
935 case AV_CODEC_ID_MPEG4:
936 ctx->cuparseinfo.CodecType = cudaVideoCodec_MPEG4;
939 #if CONFIG_VP8_CUVID_DECODER
940 case AV_CODEC_ID_VP8:
941 ctx->cuparseinfo.CodecType = cudaVideoCodec_VP8;
944 #if CONFIG_VP9_CUVID_DECODER
945 case AV_CODEC_ID_VP9:
946 ctx->cuparseinfo.CodecType = cudaVideoCodec_VP9;
949 #if CONFIG_VC1_CUVID_DECODER
950 case AV_CODEC_ID_VC1:
951 ctx->cuparseinfo.CodecType = cudaVideoCodec_VC1;
954 #if CONFIG_AV1_CUVID_DECODER && defined(CUVID_HAS_AV1_SUPPORT)
955 case AV_CODEC_ID_AV1:
956 ctx->cuparseinfo.CodecType = cudaVideoCodec_AV1;
960 av_log(avctx, AV_LOG_ERROR, "Invalid CUVID codec!\n");
964 if (avctx->codec->bsfs) {
965 const AVCodecParameters *par = avctx->internal->bsf->par_out;
966 extradata = par->extradata;
967 extradata_size = par->extradata_size;
969 extradata = avctx->extradata;
970 extradata_size = avctx->extradata_size;
973 ctx->cuparse_ext = av_mallocz(sizeof(*ctx->cuparse_ext)
974 + FFMAX(extradata_size - (int)sizeof(ctx->cuparse_ext->raw_seqhdr_data), 0));
975 if (!ctx->cuparse_ext) {
976 ret = AVERROR(ENOMEM);
980 if (extradata_size > 0)
981 memcpy(ctx->cuparse_ext->raw_seqhdr_data, extradata, extradata_size);
982 ctx->cuparse_ext->format.seqhdr_data_length = extradata_size;
984 ctx->cuparseinfo.pExtVideoInfo = ctx->cuparse_ext;
986 ctx->key_frame = av_mallocz(ctx->nb_surfaces * sizeof(int));
987 if (!ctx->key_frame) {
988 ret = AVERROR(ENOMEM);
992 ctx->cuparseinfo.ulMaxNumDecodeSurfaces = ctx->nb_surfaces;
993 ctx->cuparseinfo.ulMaxDisplayDelay = 4;
994 ctx->cuparseinfo.pUserData = avctx;
995 ctx->cuparseinfo.pfnSequenceCallback = cuvid_handle_video_sequence;
996 ctx->cuparseinfo.pfnDecodePicture = cuvid_handle_picture_decode;
997 ctx->cuparseinfo.pfnDisplayPicture = cuvid_handle_picture_display;
999 ret = CHECK_CU(ctx->cudl->cuCtxPushCurrent(cuda_ctx));
1003 ret = cuvid_test_capabilities(avctx, &ctx->cuparseinfo,
1010 ret = CHECK_CU(ctx->cvdl->cuvidCreateVideoParser(&ctx->cuparser, &ctx->cuparseinfo));
1014 seq_pkt.payload = ctx->cuparse_ext->raw_seqhdr_data;
1015 seq_pkt.payload_size = ctx->cuparse_ext->format.seqhdr_data_length;
1017 if (seq_pkt.payload && seq_pkt.payload_size) {
1018 ret = CHECK_CU(ctx->cvdl->cuvidParseVideoData(ctx->cuparser, &seq_pkt));
1023 ret = CHECK_CU(ctx->cudl->cuCtxPopCurrent(&dummy));
1027 ctx->prev_pts = INT64_MIN;
1029 if (!avctx->pkt_timebase.num || !avctx->pkt_timebase.den)
1030 av_log(avctx, AV_LOG_WARNING, "Invalid pkt_timebase, passing timestamps as-is.\n");
1035 cuvid_decode_end(avctx);
1039 static void cuvid_flush(AVCodecContext *avctx)
1041 CuvidContext *ctx = avctx->priv_data;
1042 AVHWDeviceContext *device_ctx = (AVHWDeviceContext*)ctx->hwdevice->data;
1043 AVCUDADeviceContext *device_hwctx = device_ctx->hwctx;
1044 CUcontext dummy, cuda_ctx = device_hwctx->cuda_ctx;
1045 CUVIDSOURCEDATAPACKET seq_pkt = { 0 };
1048 ret = CHECK_CU(ctx->cudl->cuCtxPushCurrent(cuda_ctx));
1052 av_fifo_freep(&ctx->frame_queue);
1054 ctx->frame_queue = av_fifo_alloc(ctx->nb_surfaces * sizeof(CuvidParsedFrame));
1055 if (!ctx->frame_queue) {
1056 av_log(avctx, AV_LOG_ERROR, "Failed to recreate frame queue on flush\n");
1060 if (ctx->cudecoder) {
1061 ctx->cvdl->cuvidDestroyDecoder(ctx->cudecoder);
1062 ctx->cudecoder = NULL;
1065 if (ctx->cuparser) {
1066 ctx->cvdl->cuvidDestroyVideoParser(ctx->cuparser);
1067 ctx->cuparser = NULL;
1070 ret = CHECK_CU(ctx->cvdl->cuvidCreateVideoParser(&ctx->cuparser, &ctx->cuparseinfo));
1074 seq_pkt.payload = ctx->cuparse_ext->raw_seqhdr_data;
1075 seq_pkt.payload_size = ctx->cuparse_ext->format.seqhdr_data_length;
1077 if (seq_pkt.payload && seq_pkt.payload_size) {
1078 ret = CHECK_CU(ctx->cvdl->cuvidParseVideoData(ctx->cuparser, &seq_pkt));
1083 ret = CHECK_CU(ctx->cudl->cuCtxPopCurrent(&dummy));
1087 ctx->prev_pts = INT64_MIN;
1088 ctx->decoder_flushing = 0;
1092 av_log(avctx, AV_LOG_ERROR, "CUDA reinit on flush failed\n");
1095 #define OFFSET(x) offsetof(CuvidContext, x)
1096 #define VD AV_OPT_FLAG_VIDEO_PARAM | AV_OPT_FLAG_DECODING_PARAM
1097 static const AVOption options[] = {
1098 { "deint", "Set deinterlacing mode", OFFSET(deint_mode), AV_OPT_TYPE_INT, { .i64 = cudaVideoDeinterlaceMode_Weave }, cudaVideoDeinterlaceMode_Weave, cudaVideoDeinterlaceMode_Adaptive, VD, "deint" },
1099 { "weave", "Weave deinterlacing (do nothing)", 0, AV_OPT_TYPE_CONST, { .i64 = cudaVideoDeinterlaceMode_Weave }, 0, 0, VD, "deint" },
1100 { "bob", "Bob deinterlacing", 0, AV_OPT_TYPE_CONST, { .i64 = cudaVideoDeinterlaceMode_Bob }, 0, 0, VD, "deint" },
1101 { "adaptive", "Adaptive deinterlacing", 0, AV_OPT_TYPE_CONST, { .i64 = cudaVideoDeinterlaceMode_Adaptive }, 0, 0, VD, "deint" },
1102 { "gpu", "GPU to be used for decoding", OFFSET(cu_gpu), AV_OPT_TYPE_STRING, { .str = NULL }, 0, 0, VD },
1103 { "surfaces", "Maximum surfaces to be used for decoding", OFFSET(nb_surfaces), AV_OPT_TYPE_INT, { .i64 = 25 }, 0, INT_MAX, VD },
1104 { "drop_second_field", "Drop second field when deinterlacing", OFFSET(drop_second_field), AV_OPT_TYPE_BOOL, { .i64 = 0 }, 0, 1, VD },
1105 { "crop", "Crop (top)x(bottom)x(left)x(right)", OFFSET(crop_expr), AV_OPT_TYPE_STRING, { .str = NULL }, 0, 0, VD },
1106 { "resize", "Resize (width)x(height)", OFFSET(resize_expr), AV_OPT_TYPE_STRING, { .str = NULL }, 0, 0, VD },
1110 static const AVCodecHWConfigInternal *const cuvid_hw_configs[] = {
1111 &(const AVCodecHWConfigInternal) {
1113 .pix_fmt = AV_PIX_FMT_CUDA,
1114 .methods = AV_CODEC_HW_CONFIG_METHOD_HW_DEVICE_CTX |
1115 AV_CODEC_HW_CONFIG_METHOD_INTERNAL,
1116 .device_type = AV_HWDEVICE_TYPE_CUDA
1123 #define DEFINE_CUVID_CODEC(x, X, bsf_name) \
1124 static const AVClass x##_cuvid_class = { \
1125 .class_name = #x "_cuvid", \
1126 .item_name = av_default_item_name, \
1127 .option = options, \
1128 .version = LIBAVUTIL_VERSION_INT, \
1130 AVCodec ff_##x##_cuvid_decoder = { \
1131 .name = #x "_cuvid", \
1132 .long_name = NULL_IF_CONFIG_SMALL("Nvidia CUVID " #X " decoder"), \
1133 .type = AVMEDIA_TYPE_VIDEO, \
1134 .id = AV_CODEC_ID_##X, \
1135 .priv_data_size = sizeof(CuvidContext), \
1136 .priv_class = &x##_cuvid_class, \
1137 .init = cuvid_decode_init, \
1138 .close = cuvid_decode_end, \
1139 .decode = cuvid_decode_frame, \
1140 .receive_frame = cuvid_output_frame, \
1141 .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_AV1_CUVID_DECODER && defined(CUVID_HAS_AV1_SUPPORT)
1154 DEFINE_CUVID_CODEC(av1, AV1, NULL)
1157 #if CONFIG_HEVC_CUVID_DECODER
1158 DEFINE_CUVID_CODEC(hevc, HEVC, "hevc_mp4toannexb")
1161 #if CONFIG_H264_CUVID_DECODER
1162 DEFINE_CUVID_CODEC(h264, H264, "h264_mp4toannexb")
1165 #if CONFIG_MJPEG_CUVID_DECODER
1166 DEFINE_CUVID_CODEC(mjpeg, MJPEG, NULL)
1169 #if CONFIG_MPEG1_CUVID_DECODER
1170 DEFINE_CUVID_CODEC(mpeg1, MPEG1VIDEO, NULL)
1173 #if CONFIG_MPEG2_CUVID_DECODER
1174 DEFINE_CUVID_CODEC(mpeg2, MPEG2VIDEO, NULL)
1177 #if CONFIG_MPEG4_CUVID_DECODER
1178 DEFINE_CUVID_CODEC(mpeg4, MPEG4, NULL)
1181 #if CONFIG_VP8_CUVID_DECODER
1182 DEFINE_CUVID_CODEC(vp8, VP8, NULL)
1185 #if CONFIG_VP9_CUVID_DECODER
1186 DEFINE_CUVID_CODEC(vp9, VP9, NULL)
1189 #if CONFIG_VC1_CUVID_DECODER
1190 DEFINE_CUVID_CODEC(vc1, VC1, NULL)