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 typedef struct CuvidContext
49 CUvideodecoder cudecoder;
50 CUvideoparser cuparser;
54 int drop_second_field;
70 AVBufferRef *hwdevice;
75 AVFifoBuffer *frame_queue;
78 int deint_mode_current;
86 cudaVideoCodec codec_type;
87 cudaVideoChromaFormat chroma_format;
89 CUVIDDECODECAPS caps8, caps10, caps12;
91 CUVIDPARSERPARAMS cuparseinfo;
92 CUVIDEOFORMATEX cuparse_ext;
98 typedef struct CuvidParsedFrame
100 CUVIDPARSERDISPINFO dispinfo;
102 int is_deinterlacing;
105 #define CHECK_CU(x) FF_CUDA_CHECK_DL(avctx, ctx->cudl, x)
107 static int CUDAAPI cuvid_handle_video_sequence(void *opaque, CUVIDEOFORMAT* format)
109 AVCodecContext *avctx = opaque;
110 CuvidContext *ctx = avctx->priv_data;
111 AVHWFramesContext *hwframe_ctx = (AVHWFramesContext*)ctx->hwframe->data;
112 CUVIDDECODECAPS *caps = NULL;
113 CUVIDDECODECREATEINFO cuinfo;
117 int old_width = avctx->width;
118 int old_height = avctx->height;
120 enum AVPixelFormat pix_fmts[3] = { AV_PIX_FMT_CUDA,
121 AV_PIX_FMT_NONE, // Will be updated below
124 av_log(avctx, AV_LOG_TRACE, "pfnSequenceCallback, progressive_sequence=%d\n", format->progressive_sequence);
126 memset(&cuinfo, 0, sizeof(cuinfo));
128 ctx->internal_error = 0;
130 avctx->coded_width = cuinfo.ulWidth = format->coded_width;
131 avctx->coded_height = cuinfo.ulHeight = format->coded_height;
134 cuinfo.display_area.left = format->display_area.left + ctx->crop.left;
135 cuinfo.display_area.top = format->display_area.top + ctx->crop.top;
136 cuinfo.display_area.right = format->display_area.right - ctx->crop.right;
137 cuinfo.display_area.bottom = format->display_area.bottom - ctx->crop.bottom;
139 // width and height need to be set before calling ff_get_format
140 if (ctx->resize_expr) {
141 avctx->width = ctx->resize.width;
142 avctx->height = ctx->resize.height;
144 avctx->width = cuinfo.display_area.right - cuinfo.display_area.left;
145 avctx->height = cuinfo.display_area.bottom - cuinfo.display_area.top;
148 // target width/height need to be multiples of two
149 cuinfo.ulTargetWidth = avctx->width = (avctx->width + 1) & ~1;
150 cuinfo.ulTargetHeight = avctx->height = (avctx->height + 1) & ~1;
152 // aspect ratio conversion, 1:1, depends on scaled resolution
153 cuinfo.target_rect.left = 0;
154 cuinfo.target_rect.top = 0;
155 cuinfo.target_rect.right = cuinfo.ulTargetWidth;
156 cuinfo.target_rect.bottom = cuinfo.ulTargetHeight;
158 chroma_444 = format->chroma_format == cudaVideoChromaFormat_444;
160 switch (format->bit_depth_luma_minus8) {
162 pix_fmts[1] = chroma_444 ? AV_PIX_FMT_YUV444P : AV_PIX_FMT_NV12;
166 pix_fmts[1] = chroma_444 ? AV_PIX_FMT_YUV444P16 : AV_PIX_FMT_P010;
170 pix_fmts[1] = chroma_444 ? AV_PIX_FMT_YUV444P16 : AV_PIX_FMT_P016;
177 if (!caps || !caps->bIsSupported) {
178 av_log(avctx, AV_LOG_ERROR, "unsupported bit depth: %d\n",
179 format->bit_depth_luma_minus8 + 8);
180 ctx->internal_error = AVERROR(EINVAL);
184 surface_fmt = ff_get_format(avctx, pix_fmts);
185 if (surface_fmt < 0) {
186 av_log(avctx, AV_LOG_ERROR, "ff_get_format failed: %d\n", surface_fmt);
187 ctx->internal_error = AVERROR(EINVAL);
191 av_log(avctx, AV_LOG_VERBOSE, "Formats: Original: %s | HW: %s | SW: %s\n",
192 av_get_pix_fmt_name(avctx->pix_fmt),
193 av_get_pix_fmt_name(surface_fmt),
194 av_get_pix_fmt_name(avctx->sw_pix_fmt));
196 avctx->pix_fmt = surface_fmt;
198 // Update our hwframe ctx, as the get_format callback might have refreshed it!
199 if (avctx->hw_frames_ctx) {
200 av_buffer_unref(&ctx->hwframe);
202 ctx->hwframe = av_buffer_ref(avctx->hw_frames_ctx);
204 ctx->internal_error = AVERROR(ENOMEM);
208 hwframe_ctx = (AVHWFramesContext*)ctx->hwframe->data;
211 ff_set_sar(avctx, av_div_q(
212 (AVRational){ format->display_aspect_ratio.x, format->display_aspect_ratio.y },
213 (AVRational){ avctx->width, avctx->height }));
215 ctx->deint_mode_current = format->progressive_sequence
216 ? cudaVideoDeinterlaceMode_Weave
219 if (!format->progressive_sequence && ctx->deint_mode_current == cudaVideoDeinterlaceMode_Weave)
220 avctx->flags |= AV_CODEC_FLAG_INTERLACED_DCT;
222 avctx->flags &= ~AV_CODEC_FLAG_INTERLACED_DCT;
224 if (format->video_signal_description.video_full_range_flag)
225 avctx->color_range = AVCOL_RANGE_JPEG;
227 avctx->color_range = AVCOL_RANGE_MPEG;
229 avctx->color_primaries = format->video_signal_description.color_primaries;
230 avctx->color_trc = format->video_signal_description.transfer_characteristics;
231 avctx->colorspace = format->video_signal_description.matrix_coefficients;
234 avctx->bit_rate = format->bitrate;
236 if (format->frame_rate.numerator && format->frame_rate.denominator) {
237 avctx->framerate.num = format->frame_rate.numerator;
238 avctx->framerate.den = format->frame_rate.denominator;
242 && avctx->coded_width == format->coded_width
243 && avctx->coded_height == format->coded_height
244 && avctx->width == old_width
245 && avctx->height == old_height
246 && ctx->chroma_format == format->chroma_format
247 && ctx->codec_type == format->codec)
250 if (ctx->cudecoder) {
251 av_log(avctx, AV_LOG_TRACE, "Re-initializing decoder\n");
252 ctx->internal_error = CHECK_CU(ctx->cvdl->cuvidDestroyDecoder(ctx->cudecoder));
253 if (ctx->internal_error < 0)
255 ctx->cudecoder = NULL;
258 if (hwframe_ctx->pool && (
259 hwframe_ctx->width < avctx->width ||
260 hwframe_ctx->height < avctx->height ||
261 hwframe_ctx->format != AV_PIX_FMT_CUDA ||
262 hwframe_ctx->sw_format != avctx->sw_pix_fmt)) {
263 av_log(avctx, AV_LOG_ERROR, "AVHWFramesContext is already initialized with incompatible parameters\n");
264 av_log(avctx, AV_LOG_DEBUG, "width: %d <-> %d\n", hwframe_ctx->width, avctx->width);
265 av_log(avctx, AV_LOG_DEBUG, "height: %d <-> %d\n", hwframe_ctx->height, avctx->height);
266 av_log(avctx, AV_LOG_DEBUG, "format: %s <-> cuda\n", av_get_pix_fmt_name(hwframe_ctx->format));
267 av_log(avctx, AV_LOG_DEBUG, "sw_format: %s <-> %s\n",
268 av_get_pix_fmt_name(hwframe_ctx->sw_format), av_get_pix_fmt_name(avctx->sw_pix_fmt));
269 ctx->internal_error = AVERROR(EINVAL);
273 ctx->chroma_format = format->chroma_format;
275 cuinfo.CodecType = ctx->codec_type = format->codec;
276 cuinfo.ChromaFormat = format->chroma_format;
278 switch (avctx->sw_pix_fmt) {
279 case AV_PIX_FMT_NV12:
280 cuinfo.OutputFormat = cudaVideoSurfaceFormat_NV12;
282 case AV_PIX_FMT_P010:
283 case AV_PIX_FMT_P016:
284 cuinfo.OutputFormat = cudaVideoSurfaceFormat_P016;
286 case AV_PIX_FMT_YUV444P:
287 cuinfo.OutputFormat = cudaVideoSurfaceFormat_YUV444;
289 case AV_PIX_FMT_YUV444P16:
290 cuinfo.OutputFormat = cudaVideoSurfaceFormat_YUV444_16Bit;
293 av_log(avctx, AV_LOG_ERROR, "Unsupported output format: %s\n",
294 av_get_pix_fmt_name(avctx->sw_pix_fmt));
295 ctx->internal_error = AVERROR(EINVAL);
299 cuinfo.ulNumDecodeSurfaces = ctx->nb_surfaces;
300 cuinfo.ulNumOutputSurfaces = 1;
301 cuinfo.ulCreationFlags = cudaVideoCreate_PreferCUVID;
302 cuinfo.bitDepthMinus8 = format->bit_depth_luma_minus8;
303 cuinfo.DeinterlaceMode = ctx->deint_mode_current;
305 if (ctx->deint_mode_current != cudaVideoDeinterlaceMode_Weave && !ctx->drop_second_field)
306 avctx->framerate = av_mul_q(avctx->framerate, (AVRational){2, 1});
308 ctx->internal_error = CHECK_CU(ctx->cvdl->cuvidCreateDecoder(&ctx->cudecoder, &cuinfo));
309 if (ctx->internal_error < 0)
312 if (!hwframe_ctx->pool) {
313 hwframe_ctx->format = AV_PIX_FMT_CUDA;
314 hwframe_ctx->sw_format = avctx->sw_pix_fmt;
315 hwframe_ctx->width = avctx->width;
316 hwframe_ctx->height = avctx->height;
318 if ((ctx->internal_error = av_hwframe_ctx_init(ctx->hwframe)) < 0) {
319 av_log(avctx, AV_LOG_ERROR, "av_hwframe_ctx_init failed\n");
327 static int CUDAAPI cuvid_handle_picture_decode(void *opaque, CUVIDPICPARAMS* picparams)
329 AVCodecContext *avctx = opaque;
330 CuvidContext *ctx = avctx->priv_data;
332 av_log(avctx, AV_LOG_TRACE, "pfnDecodePicture\n");
334 ctx->key_frame[picparams->CurrPicIdx] = picparams->intra_pic_flag;
336 ctx->internal_error = CHECK_CU(ctx->cvdl->cuvidDecodePicture(ctx->cudecoder, picparams));
337 if (ctx->internal_error < 0)
343 static int CUDAAPI cuvid_handle_picture_display(void *opaque, CUVIDPARSERDISPINFO* dispinfo)
345 AVCodecContext *avctx = opaque;
346 CuvidContext *ctx = avctx->priv_data;
347 CuvidParsedFrame parsed_frame = { { 0 } };
349 parsed_frame.dispinfo = *dispinfo;
350 ctx->internal_error = 0;
352 if (ctx->deint_mode_current == cudaVideoDeinterlaceMode_Weave) {
353 av_fifo_generic_write(ctx->frame_queue, &parsed_frame, sizeof(CuvidParsedFrame), NULL);
355 parsed_frame.is_deinterlacing = 1;
356 av_fifo_generic_write(ctx->frame_queue, &parsed_frame, sizeof(CuvidParsedFrame), NULL);
357 if (!ctx->drop_second_field) {
358 parsed_frame.second_field = 1;
359 av_fifo_generic_write(ctx->frame_queue, &parsed_frame, sizeof(CuvidParsedFrame), NULL);
366 static int cuvid_is_buffer_full(AVCodecContext *avctx)
368 CuvidContext *ctx = avctx->priv_data;
370 int delay = ctx->cuparseinfo.ulMaxDisplayDelay;
371 if (ctx->deint_mode != cudaVideoDeinterlaceMode_Weave && !ctx->drop_second_field)
374 return (av_fifo_size(ctx->frame_queue) / sizeof(CuvidParsedFrame)) + delay >= ctx->nb_surfaces;
377 static int cuvid_decode_packet(AVCodecContext *avctx, const AVPacket *avpkt)
379 CuvidContext *ctx = avctx->priv_data;
380 AVHWDeviceContext *device_ctx = (AVHWDeviceContext*)ctx->hwdevice->data;
381 AVCUDADeviceContext *device_hwctx = device_ctx->hwctx;
382 CUcontext dummy, cuda_ctx = device_hwctx->cuda_ctx;
383 CUVIDSOURCEDATAPACKET cupkt;
384 AVPacket filter_packet = { 0 };
385 AVPacket filtered_packet = { 0 };
386 int ret = 0, eret = 0, is_flush = ctx->decoder_flushing;
388 av_log(avctx, AV_LOG_TRACE, "cuvid_decode_packet\n");
390 if (is_flush && avpkt && avpkt->size)
393 if (cuvid_is_buffer_full(avctx) && avpkt && avpkt->size)
394 return AVERROR(EAGAIN);
396 if (ctx->bsf && avpkt && avpkt->size) {
397 if ((ret = av_packet_ref(&filter_packet, avpkt)) < 0) {
398 av_log(avctx, AV_LOG_ERROR, "av_packet_ref failed\n");
402 if ((ret = av_bsf_send_packet(ctx->bsf, &filter_packet)) < 0) {
403 av_log(avctx, AV_LOG_ERROR, "av_bsf_send_packet failed\n");
404 av_packet_unref(&filter_packet);
408 if ((ret = av_bsf_receive_packet(ctx->bsf, &filtered_packet)) < 0) {
409 av_log(avctx, AV_LOG_ERROR, "av_bsf_receive_packet failed\n");
413 avpkt = &filtered_packet;
416 ret = CHECK_CU(ctx->cudl->cuCtxPushCurrent(cuda_ctx));
418 av_packet_unref(&filtered_packet);
422 memset(&cupkt, 0, sizeof(cupkt));
424 if (avpkt && avpkt->size) {
425 cupkt.payload_size = avpkt->size;
426 cupkt.payload = avpkt->data;
428 if (avpkt->pts != AV_NOPTS_VALUE) {
429 cupkt.flags = CUVID_PKT_TIMESTAMP;
430 if (avctx->pkt_timebase.num && avctx->pkt_timebase.den)
431 cupkt.timestamp = av_rescale_q(avpkt->pts, avctx->pkt_timebase, (AVRational){1, 10000000});
433 cupkt.timestamp = avpkt->pts;
436 cupkt.flags = CUVID_PKT_ENDOFSTREAM;
437 ctx->decoder_flushing = 1;
440 ret = CHECK_CU(ctx->cvdl->cuvidParseVideoData(ctx->cuparser, &cupkt));
442 av_packet_unref(&filtered_packet);
447 // cuvidParseVideoData doesn't return an error just because stuff failed...
448 if (ctx->internal_error) {
449 av_log(avctx, AV_LOG_ERROR, "cuvid decode callback error\n");
450 ret = ctx->internal_error;
455 eret = CHECK_CU(ctx->cudl->cuCtxPopCurrent(&dummy));
467 static int cuvid_output_frame(AVCodecContext *avctx, AVFrame *frame)
469 CuvidContext *ctx = avctx->priv_data;
470 AVHWDeviceContext *device_ctx = (AVHWDeviceContext*)ctx->hwdevice->data;
471 AVCUDADeviceContext *device_hwctx = device_ctx->hwctx;
472 CUcontext dummy, cuda_ctx = device_hwctx->cuda_ctx;
473 CUdeviceptr mapped_frame = 0;
474 int ret = 0, eret = 0;
476 av_log(avctx, AV_LOG_TRACE, "cuvid_output_frame\n");
478 if (ctx->decoder_flushing) {
479 ret = cuvid_decode_packet(avctx, NULL);
480 if (ret < 0 && ret != AVERROR_EOF)
484 if (!cuvid_is_buffer_full(avctx)) {
486 ret = ff_decode_get_packet(avctx, &pkt);
487 if (ret < 0 && ret != AVERROR_EOF)
489 ret = cuvid_decode_packet(avctx, &pkt);
490 av_packet_unref(&pkt);
491 // cuvid_is_buffer_full() should avoid this.
492 if (ret == AVERROR(EAGAIN))
493 ret = AVERROR_EXTERNAL;
494 if (ret < 0 && ret != AVERROR_EOF)
498 ret = CHECK_CU(ctx->cudl->cuCtxPushCurrent(cuda_ctx));
502 if (av_fifo_size(ctx->frame_queue)) {
503 const AVPixFmtDescriptor *pixdesc;
504 CuvidParsedFrame parsed_frame;
505 CUVIDPROCPARAMS params;
506 unsigned int pitch = 0;
510 av_fifo_generic_read(ctx->frame_queue, &parsed_frame, sizeof(CuvidParsedFrame), NULL);
512 memset(¶ms, 0, sizeof(params));
513 params.progressive_frame = parsed_frame.dispinfo.progressive_frame;
514 params.second_field = parsed_frame.second_field;
515 params.top_field_first = parsed_frame.dispinfo.top_field_first;
517 ret = CHECK_CU(ctx->cvdl->cuvidMapVideoFrame(ctx->cudecoder, parsed_frame.dispinfo.picture_index, &mapped_frame, &pitch, ¶ms));
521 if (avctx->pix_fmt == AV_PIX_FMT_CUDA) {
522 ret = av_hwframe_get_buffer(ctx->hwframe, frame, 0);
524 av_log(avctx, AV_LOG_ERROR, "av_hwframe_get_buffer failed\n");
528 ret = ff_decode_frame_props(avctx, frame);
530 av_log(avctx, AV_LOG_ERROR, "ff_decode_frame_props failed\n");
534 pixdesc = av_pix_fmt_desc_get(avctx->sw_pix_fmt);
536 for (i = 0; i < pixdesc->nb_components; i++) {
537 int height = avctx->height >> (i ? pixdesc->log2_chroma_h : 0);
538 CUDA_MEMCPY2D cpy = {
539 .srcMemoryType = CU_MEMORYTYPE_DEVICE,
540 .dstMemoryType = CU_MEMORYTYPE_DEVICE,
541 .srcDevice = mapped_frame,
542 .dstDevice = (CUdeviceptr)frame->data[i],
544 .dstPitch = frame->linesize[i],
546 .WidthInBytes = FFMIN(pitch, frame->linesize[i]),
550 ret = CHECK_CU(ctx->cudl->cuMemcpy2DAsync(&cpy, device_hwctx->stream));
557 ret = CHECK_CU(ctx->cudl->cuStreamSynchronize(device_hwctx->stream));
560 } else if (avctx->pix_fmt == AV_PIX_FMT_NV12 ||
561 avctx->pix_fmt == AV_PIX_FMT_P010 ||
562 avctx->pix_fmt == AV_PIX_FMT_P016 ||
563 avctx->pix_fmt == AV_PIX_FMT_YUV444P ||
564 avctx->pix_fmt == AV_PIX_FMT_YUV444P16) {
565 unsigned int offset = 0;
566 AVFrame *tmp_frame = av_frame_alloc();
568 av_log(avctx, AV_LOG_ERROR, "av_frame_alloc failed\n");
569 ret = AVERROR(ENOMEM);
573 pixdesc = av_pix_fmt_desc_get(avctx->sw_pix_fmt);
575 tmp_frame->format = AV_PIX_FMT_CUDA;
576 tmp_frame->hw_frames_ctx = av_buffer_ref(ctx->hwframe);
577 tmp_frame->width = avctx->width;
578 tmp_frame->height = avctx->height;
581 * Note that the following logic would not work for three plane
582 * YUV420 because the pitch value is different for the chroma
585 for (i = 0; i < pixdesc->nb_components; i++) {
586 tmp_frame->data[i] = (uint8_t*)mapped_frame + offset;
587 tmp_frame->linesize[i] = pitch;
588 offset += pitch * (avctx->height >> (i ? pixdesc->log2_chroma_h : 0));
591 ret = ff_get_buffer(avctx, frame, 0);
593 av_log(avctx, AV_LOG_ERROR, "ff_get_buffer failed\n");
594 av_frame_free(&tmp_frame);
598 ret = av_hwframe_transfer_data(frame, tmp_frame, 0);
600 av_log(avctx, AV_LOG_ERROR, "av_hwframe_transfer_data failed\n");
601 av_frame_free(&tmp_frame);
604 av_frame_free(&tmp_frame);
610 frame->key_frame = ctx->key_frame[parsed_frame.dispinfo.picture_index];
611 frame->width = avctx->width;
612 frame->height = avctx->height;
613 if (avctx->pkt_timebase.num && avctx->pkt_timebase.den)
614 frame->pts = av_rescale_q(parsed_frame.dispinfo.timestamp, (AVRational){1, 10000000}, avctx->pkt_timebase);
616 frame->pts = parsed_frame.dispinfo.timestamp;
618 if (parsed_frame.second_field) {
619 if (ctx->prev_pts == INT64_MIN) {
620 ctx->prev_pts = frame->pts;
621 frame->pts += (avctx->pkt_timebase.den * avctx->framerate.den) / (avctx->pkt_timebase.num * avctx->framerate.num);
623 int pts_diff = (frame->pts - ctx->prev_pts) / 2;
624 ctx->prev_pts = frame->pts;
625 frame->pts += pts_diff;
629 /* CUVIDs opaque reordering breaks the internal pkt logic.
630 * So set pkt_pts and clear all the other pkt_ fields.
633 FF_DISABLE_DEPRECATION_WARNINGS
634 frame->pkt_pts = frame->pts;
635 FF_ENABLE_DEPRECATION_WARNINGS
638 frame->pkt_duration = 0;
639 frame->pkt_size = -1;
641 frame->interlaced_frame = !parsed_frame.is_deinterlacing && !parsed_frame.dispinfo.progressive_frame;
643 if (frame->interlaced_frame)
644 frame->top_field_first = parsed_frame.dispinfo.top_field_first;
645 } else if (ctx->decoder_flushing) {
648 ret = AVERROR(EAGAIN);
653 eret = CHECK_CU(ctx->cvdl->cuvidUnmapVideoFrame(ctx->cudecoder, mapped_frame));
655 eret = CHECK_CU(ctx->cudl->cuCtxPopCurrent(&dummy));
663 static int cuvid_decode_frame(AVCodecContext *avctx, void *data, int *got_frame, AVPacket *avpkt)
665 CuvidContext *ctx = avctx->priv_data;
666 AVFrame *frame = data;
669 av_log(avctx, AV_LOG_TRACE, "cuvid_decode_frame\n");
671 if (ctx->deint_mode_current != cudaVideoDeinterlaceMode_Weave) {
672 av_log(avctx, AV_LOG_ERROR, "Deinterlacing is not supported via the old API\n");
673 return AVERROR(EINVAL);
676 if (!ctx->decoder_flushing) {
677 ret = cuvid_decode_packet(avctx, avpkt);
682 ret = cuvid_output_frame(avctx, frame);
683 if (ret == AVERROR(EAGAIN) || ret == AVERROR_EOF) {
685 } else if (ret < 0) {
694 static av_cold int cuvid_decode_end(AVCodecContext *avctx)
696 CuvidContext *ctx = avctx->priv_data;
698 av_fifo_freep(&ctx->frame_queue);
701 av_bsf_free(&ctx->bsf);
704 ctx->cvdl->cuvidDestroyVideoParser(ctx->cuparser);
707 ctx->cvdl->cuvidDestroyDecoder(ctx->cudecoder);
711 av_buffer_unref(&ctx->hwframe);
712 av_buffer_unref(&ctx->hwdevice);
714 av_freep(&ctx->key_frame);
716 cuvid_free_functions(&ctx->cvdl);
721 static int cuvid_test_capabilities(AVCodecContext *avctx,
722 const CUVIDPARSERPARAMS *cuparseinfo,
727 CuvidContext *ctx = avctx->priv_data;
728 CUVIDDECODECAPS *caps;
729 int res8 = 0, res10 = 0, res12 = 0;
731 if (!ctx->cvdl->cuvidGetDecoderCaps) {
732 av_log(avctx, AV_LOG_WARNING, "Used Nvidia driver is too old to perform a capability check.\n");
733 av_log(avctx, AV_LOG_WARNING, "The minimum required version is "
734 #if defined(_WIN32) || defined(__CYGWIN__)
739 ". Continuing blind.\n");
740 ctx->caps8.bIsSupported = ctx->caps10.bIsSupported = 1;
741 // 12 bit was not supported before the capability check was introduced, so disable it.
742 ctx->caps12.bIsSupported = 0;
746 ctx->caps8.eCodecType = ctx->caps10.eCodecType = ctx->caps12.eCodecType
747 = cuparseinfo->CodecType;
748 ctx->caps8.eChromaFormat = ctx->caps10.eChromaFormat = ctx->caps12.eChromaFormat
749 = cudaVideoChromaFormat_420;
751 ctx->caps8.nBitDepthMinus8 = 0;
752 ctx->caps10.nBitDepthMinus8 = 2;
753 ctx->caps12.nBitDepthMinus8 = 4;
755 res8 = CHECK_CU(ctx->cvdl->cuvidGetDecoderCaps(&ctx->caps8));
756 res10 = CHECK_CU(ctx->cvdl->cuvidGetDecoderCaps(&ctx->caps10));
757 res12 = CHECK_CU(ctx->cvdl->cuvidGetDecoderCaps(&ctx->caps12));
759 av_log(avctx, AV_LOG_VERBOSE, "CUVID capabilities for %s:\n", avctx->codec->name);
760 av_log(avctx, AV_LOG_VERBOSE, "8 bit: supported: %d, min_width: %d, max_width: %d, min_height: %d, max_height: %d\n",
761 ctx->caps8.bIsSupported, ctx->caps8.nMinWidth, ctx->caps8.nMaxWidth, ctx->caps8.nMinHeight, ctx->caps8.nMaxHeight);
762 av_log(avctx, AV_LOG_VERBOSE, "10 bit: supported: %d, min_width: %d, max_width: %d, min_height: %d, max_height: %d\n",
763 ctx->caps10.bIsSupported, ctx->caps10.nMinWidth, ctx->caps10.nMaxWidth, ctx->caps10.nMinHeight, ctx->caps10.nMaxHeight);
764 av_log(avctx, AV_LOG_VERBOSE, "12 bit: supported: %d, min_width: %d, max_width: %d, min_height: %d, max_height: %d\n",
765 ctx->caps12.bIsSupported, ctx->caps12.nMinWidth, ctx->caps12.nMaxWidth, ctx->caps12.nMinHeight, ctx->caps12.nMaxHeight);
784 if (!ctx->caps8.bIsSupported) {
785 av_log(avctx, AV_LOG_ERROR, "Codec %s is not supported.\n", avctx->codec->name);
786 return AVERROR(EINVAL);
789 if (!caps->bIsSupported) {
790 av_log(avctx, AV_LOG_ERROR, "Bit depth %d is not supported.\n", bit_depth);
791 return AVERROR(EINVAL);
794 if (probed_width > caps->nMaxWidth || probed_width < caps->nMinWidth) {
795 av_log(avctx, AV_LOG_ERROR, "Video width %d not within range from %d to %d\n",
796 probed_width, caps->nMinWidth, caps->nMaxWidth);
797 return AVERROR(EINVAL);
800 if (probed_height > caps->nMaxHeight || probed_height < caps->nMinHeight) {
801 av_log(avctx, AV_LOG_ERROR, "Video height %d not within range from %d to %d\n",
802 probed_height, caps->nMinHeight, caps->nMaxHeight);
803 return AVERROR(EINVAL);
809 static av_cold int cuvid_decode_init(AVCodecContext *avctx)
811 CuvidContext *ctx = avctx->priv_data;
812 AVCUDADeviceContext *device_hwctx;
813 AVHWDeviceContext *device_ctx;
814 AVHWFramesContext *hwframe_ctx;
815 CUVIDSOURCEDATAPACKET seq_pkt;
816 CUcontext cuda_ctx = NULL;
818 const AVBitStreamFilter *bsf;
821 enum AVPixelFormat pix_fmts[3] = { AV_PIX_FMT_CUDA,
825 int probed_width = avctx->coded_width ? avctx->coded_width : 1280;
826 int probed_height = avctx->coded_height ? avctx->coded_height : 720;
827 int probed_bit_depth = 8;
829 const AVPixFmtDescriptor *probe_desc = av_pix_fmt_desc_get(avctx->pix_fmt);
830 if (probe_desc && probe_desc->nb_components)
831 probed_bit_depth = probe_desc->comp[0].depth;
833 // Accelerated transcoding scenarios with 'ffmpeg' require that the
834 // pix_fmt be set to AV_PIX_FMT_CUDA early. The sw_pix_fmt, and the
835 // pix_fmt for non-accelerated transcoding, do not need to be correct
836 // but need to be set to something. We arbitrarily pick NV12.
837 ret = ff_get_format(avctx, pix_fmts);
839 av_log(avctx, AV_LOG_ERROR, "ff_get_format failed: %d\n", ret);
842 avctx->pix_fmt = ret;
844 if (ctx->resize_expr && sscanf(ctx->resize_expr, "%dx%d",
845 &ctx->resize.width, &ctx->resize.height) != 2) {
846 av_log(avctx, AV_LOG_ERROR, "Invalid resize expressions\n");
847 ret = AVERROR(EINVAL);
851 if (ctx->crop_expr && sscanf(ctx->crop_expr, "%dx%dx%dx%d",
852 &ctx->crop.top, &ctx->crop.bottom,
853 &ctx->crop.left, &ctx->crop.right) != 4) {
854 av_log(avctx, AV_LOG_ERROR, "Invalid cropping expressions\n");
855 ret = AVERROR(EINVAL);
859 ret = cuvid_load_functions(&ctx->cvdl, avctx);
861 av_log(avctx, AV_LOG_ERROR, "Failed loading nvcuvid.\n");
865 ctx->frame_queue = av_fifo_alloc(ctx->nb_surfaces * sizeof(CuvidParsedFrame));
866 if (!ctx->frame_queue) {
867 ret = AVERROR(ENOMEM);
871 if (avctx->hw_frames_ctx) {
872 ctx->hwframe = av_buffer_ref(avctx->hw_frames_ctx);
874 ret = AVERROR(ENOMEM);
878 hwframe_ctx = (AVHWFramesContext*)ctx->hwframe->data;
880 ctx->hwdevice = av_buffer_ref(hwframe_ctx->device_ref);
881 if (!ctx->hwdevice) {
882 ret = AVERROR(ENOMEM);
886 if (avctx->hw_device_ctx) {
887 ctx->hwdevice = av_buffer_ref(avctx->hw_device_ctx);
888 if (!ctx->hwdevice) {
889 ret = AVERROR(ENOMEM);
893 ret = av_hwdevice_ctx_create(&ctx->hwdevice, AV_HWDEVICE_TYPE_CUDA, ctx->cu_gpu, NULL, 0);
898 ctx->hwframe = av_hwframe_ctx_alloc(ctx->hwdevice);
900 av_log(avctx, AV_LOG_ERROR, "av_hwframe_ctx_alloc failed\n");
901 ret = AVERROR(ENOMEM);
905 hwframe_ctx = (AVHWFramesContext*)ctx->hwframe->data;
908 device_ctx = hwframe_ctx->device_ctx;
909 device_hwctx = device_ctx->hwctx;
911 cuda_ctx = device_hwctx->cuda_ctx;
912 ctx->cudl = device_hwctx->internal->cuda_dl;
914 memset(&ctx->cuparseinfo, 0, sizeof(ctx->cuparseinfo));
915 memset(&ctx->cuparse_ext, 0, sizeof(ctx->cuparse_ext));
916 memset(&seq_pkt, 0, sizeof(seq_pkt));
918 ctx->cuparseinfo.pExtVideoInfo = &ctx->cuparse_ext;
920 switch (avctx->codec->id) {
921 #if CONFIG_H264_CUVID_DECODER
922 case AV_CODEC_ID_H264:
923 ctx->cuparseinfo.CodecType = cudaVideoCodec_H264;
926 #if CONFIG_HEVC_CUVID_DECODER
927 case AV_CODEC_ID_HEVC:
928 ctx->cuparseinfo.CodecType = cudaVideoCodec_HEVC;
931 #if CONFIG_MJPEG_CUVID_DECODER
932 case AV_CODEC_ID_MJPEG:
933 ctx->cuparseinfo.CodecType = cudaVideoCodec_JPEG;
936 #if CONFIG_MPEG1_CUVID_DECODER
937 case AV_CODEC_ID_MPEG1VIDEO:
938 ctx->cuparseinfo.CodecType = cudaVideoCodec_MPEG1;
941 #if CONFIG_MPEG2_CUVID_DECODER
942 case AV_CODEC_ID_MPEG2VIDEO:
943 ctx->cuparseinfo.CodecType = cudaVideoCodec_MPEG2;
946 #if CONFIG_MPEG4_CUVID_DECODER
947 case AV_CODEC_ID_MPEG4:
948 ctx->cuparseinfo.CodecType = cudaVideoCodec_MPEG4;
951 #if CONFIG_VP8_CUVID_DECODER
952 case AV_CODEC_ID_VP8:
953 ctx->cuparseinfo.CodecType = cudaVideoCodec_VP8;
956 #if CONFIG_VP9_CUVID_DECODER
957 case AV_CODEC_ID_VP9:
958 ctx->cuparseinfo.CodecType = cudaVideoCodec_VP9;
961 #if CONFIG_VC1_CUVID_DECODER
962 case AV_CODEC_ID_VC1:
963 ctx->cuparseinfo.CodecType = cudaVideoCodec_VC1;
967 av_log(avctx, AV_LOG_ERROR, "Invalid CUVID codec!\n");
971 if (avctx->codec->id == AV_CODEC_ID_H264 || avctx->codec->id == AV_CODEC_ID_HEVC) {
972 if (avctx->codec->id == AV_CODEC_ID_H264)
973 bsf = av_bsf_get_by_name("h264_mp4toannexb");
975 bsf = av_bsf_get_by_name("hevc_mp4toannexb");
978 ret = AVERROR_BSF_NOT_FOUND;
981 if (ret = av_bsf_alloc(bsf, &ctx->bsf)) {
984 if (((ret = avcodec_parameters_from_context(ctx->bsf->par_in, avctx)) < 0) || ((ret = av_bsf_init(ctx->bsf)) < 0)) {
985 av_bsf_free(&ctx->bsf);
989 ctx->cuparse_ext.format.seqhdr_data_length = ctx->bsf->par_out->extradata_size;
990 memcpy(ctx->cuparse_ext.raw_seqhdr_data,
991 ctx->bsf->par_out->extradata,
992 FFMIN(sizeof(ctx->cuparse_ext.raw_seqhdr_data), ctx->bsf->par_out->extradata_size));
993 } else if (avctx->extradata_size > 0) {
994 ctx->cuparse_ext.format.seqhdr_data_length = avctx->extradata_size;
995 memcpy(ctx->cuparse_ext.raw_seqhdr_data,
997 FFMIN(sizeof(ctx->cuparse_ext.raw_seqhdr_data), avctx->extradata_size));
1000 ctx->key_frame = av_mallocz(ctx->nb_surfaces * sizeof(int));
1001 if (!ctx->key_frame) {
1002 ret = AVERROR(ENOMEM);
1006 ctx->cuparseinfo.ulMaxNumDecodeSurfaces = ctx->nb_surfaces;
1007 ctx->cuparseinfo.ulMaxDisplayDelay = 4;
1008 ctx->cuparseinfo.pUserData = avctx;
1009 ctx->cuparseinfo.pfnSequenceCallback = cuvid_handle_video_sequence;
1010 ctx->cuparseinfo.pfnDecodePicture = cuvid_handle_picture_decode;
1011 ctx->cuparseinfo.pfnDisplayPicture = cuvid_handle_picture_display;
1013 ret = CHECK_CU(ctx->cudl->cuCtxPushCurrent(cuda_ctx));
1017 ret = cuvid_test_capabilities(avctx, &ctx->cuparseinfo,
1024 ret = CHECK_CU(ctx->cvdl->cuvidCreateVideoParser(&ctx->cuparser, &ctx->cuparseinfo));
1028 seq_pkt.payload = ctx->cuparse_ext.raw_seqhdr_data;
1029 seq_pkt.payload_size = ctx->cuparse_ext.format.seqhdr_data_length;
1031 if (seq_pkt.payload && seq_pkt.payload_size) {
1032 ret = CHECK_CU(ctx->cvdl->cuvidParseVideoData(ctx->cuparser, &seq_pkt));
1037 ret = CHECK_CU(ctx->cudl->cuCtxPopCurrent(&dummy));
1041 ctx->prev_pts = INT64_MIN;
1043 if (!avctx->pkt_timebase.num || !avctx->pkt_timebase.den)
1044 av_log(avctx, AV_LOG_WARNING, "Invalid pkt_timebase, passing timestamps as-is.\n");
1049 cuvid_decode_end(avctx);
1053 static void cuvid_flush(AVCodecContext *avctx)
1055 CuvidContext *ctx = avctx->priv_data;
1056 AVHWDeviceContext *device_ctx = (AVHWDeviceContext*)ctx->hwdevice->data;
1057 AVCUDADeviceContext *device_hwctx = device_ctx->hwctx;
1058 CUcontext dummy, cuda_ctx = device_hwctx->cuda_ctx;
1059 CUVIDSOURCEDATAPACKET seq_pkt = { 0 };
1062 ret = CHECK_CU(ctx->cudl->cuCtxPushCurrent(cuda_ctx));
1066 av_fifo_freep(&ctx->frame_queue);
1068 ctx->frame_queue = av_fifo_alloc(ctx->nb_surfaces * sizeof(CuvidParsedFrame));
1069 if (!ctx->frame_queue) {
1070 av_log(avctx, AV_LOG_ERROR, "Failed to recreate frame queue on flush\n");
1074 if (ctx->cudecoder) {
1075 ctx->cvdl->cuvidDestroyDecoder(ctx->cudecoder);
1076 ctx->cudecoder = NULL;
1079 if (ctx->cuparser) {
1080 ctx->cvdl->cuvidDestroyVideoParser(ctx->cuparser);
1081 ctx->cuparser = NULL;
1084 ret = CHECK_CU(ctx->cvdl->cuvidCreateVideoParser(&ctx->cuparser, &ctx->cuparseinfo));
1088 seq_pkt.payload = ctx->cuparse_ext.raw_seqhdr_data;
1089 seq_pkt.payload_size = ctx->cuparse_ext.format.seqhdr_data_length;
1091 if (seq_pkt.payload && seq_pkt.payload_size) {
1092 ret = CHECK_CU(ctx->cvdl->cuvidParseVideoData(ctx->cuparser, &seq_pkt));
1097 ret = CHECK_CU(ctx->cudl->cuCtxPopCurrent(&dummy));
1101 ctx->prev_pts = INT64_MIN;
1102 ctx->decoder_flushing = 0;
1106 av_log(avctx, AV_LOG_ERROR, "CUDA reinit on flush failed\n");
1109 #define OFFSET(x) offsetof(CuvidContext, x)
1110 #define VD AV_OPT_FLAG_VIDEO_PARAM | AV_OPT_FLAG_DECODING_PARAM
1111 static const AVOption options[] = {
1112 { "deint", "Set deinterlacing mode", OFFSET(deint_mode), AV_OPT_TYPE_INT, { .i64 = cudaVideoDeinterlaceMode_Weave }, cudaVideoDeinterlaceMode_Weave, cudaVideoDeinterlaceMode_Adaptive, VD, "deint" },
1113 { "weave", "Weave deinterlacing (do nothing)", 0, AV_OPT_TYPE_CONST, { .i64 = cudaVideoDeinterlaceMode_Weave }, 0, 0, VD, "deint" },
1114 { "bob", "Bob deinterlacing", 0, AV_OPT_TYPE_CONST, { .i64 = cudaVideoDeinterlaceMode_Bob }, 0, 0, VD, "deint" },
1115 { "adaptive", "Adaptive deinterlacing", 0, AV_OPT_TYPE_CONST, { .i64 = cudaVideoDeinterlaceMode_Adaptive }, 0, 0, VD, "deint" },
1116 { "gpu", "GPU to be used for decoding", OFFSET(cu_gpu), AV_OPT_TYPE_STRING, { .str = NULL }, 0, 0, VD },
1117 { "surfaces", "Maximum surfaces to be used for decoding", OFFSET(nb_surfaces), AV_OPT_TYPE_INT, { .i64 = 25 }, 0, INT_MAX, VD },
1118 { "drop_second_field", "Drop second field when deinterlacing", OFFSET(drop_second_field), AV_OPT_TYPE_BOOL, { .i64 = 0 }, 0, 1, VD },
1119 { "crop", "Crop (top)x(bottom)x(left)x(right)", OFFSET(crop_expr), AV_OPT_TYPE_STRING, { .str = NULL }, 0, 0, VD },
1120 { "resize", "Resize (width)x(height)", OFFSET(resize_expr), AV_OPT_TYPE_STRING, { .str = NULL }, 0, 0, VD },
1124 static const AVCodecHWConfigInternal *cuvid_hw_configs[] = {
1125 &(const AVCodecHWConfigInternal) {
1127 .pix_fmt = AV_PIX_FMT_CUDA,
1128 .methods = AV_CODEC_HW_CONFIG_METHOD_HW_DEVICE_CTX |
1129 AV_CODEC_HW_CONFIG_METHOD_INTERNAL,
1130 .device_type = AV_HWDEVICE_TYPE_CUDA
1137 #define DEFINE_CUVID_CODEC(x, X) \
1138 static const AVClass x##_cuvid_class = { \
1139 .class_name = #x "_cuvid", \
1140 .item_name = av_default_item_name, \
1141 .option = options, \
1142 .version = LIBAVUTIL_VERSION_INT, \
1144 AVCodec ff_##x##_cuvid_decoder = { \
1145 .name = #x "_cuvid", \
1146 .long_name = NULL_IF_CONFIG_SMALL("Nvidia CUVID " #X " decoder"), \
1147 .type = AVMEDIA_TYPE_VIDEO, \
1148 .id = AV_CODEC_ID_##X, \
1149 .priv_data_size = sizeof(CuvidContext), \
1150 .priv_class = &x##_cuvid_class, \
1151 .init = cuvid_decode_init, \
1152 .close = cuvid_decode_end, \
1153 .decode = cuvid_decode_frame, \
1154 .receive_frame = cuvid_output_frame, \
1155 .flush = cuvid_flush, \
1156 .capabilities = AV_CODEC_CAP_DELAY | AV_CODEC_CAP_AVOID_PROBING | AV_CODEC_CAP_HARDWARE, \
1157 .pix_fmts = (const enum AVPixelFormat[]){ AV_PIX_FMT_CUDA, \
1161 AV_PIX_FMT_NONE }, \
1162 .hw_configs = cuvid_hw_configs, \
1163 .wrapper_name = "cuvid", \
1166 #if CONFIG_HEVC_CUVID_DECODER
1167 DEFINE_CUVID_CODEC(hevc, HEVC)
1170 #if CONFIG_H264_CUVID_DECODER
1171 DEFINE_CUVID_CODEC(h264, H264)
1174 #if CONFIG_MJPEG_CUVID_DECODER
1175 DEFINE_CUVID_CODEC(mjpeg, MJPEG)
1178 #if CONFIG_MPEG1_CUVID_DECODER
1179 DEFINE_CUVID_CODEC(mpeg1, MPEG1VIDEO)
1182 #if CONFIG_MPEG2_CUVID_DECODER
1183 DEFINE_CUVID_CODEC(mpeg2, MPEG2VIDEO)
1186 #if CONFIG_MPEG4_CUVID_DECODER
1187 DEFINE_CUVID_CODEC(mpeg4, MPEG4)
1190 #if CONFIG_VP8_CUVID_DECODER
1191 DEFINE_CUVID_CODEC(vp8, VP8)
1194 #if CONFIG_VP9_CUVID_DECODER
1195 DEFINE_CUVID_CODEC(vp9, VP9)
1198 #if CONFIG_VC1_CUVID_DECODER
1199 DEFINE_CUVID_CODEC(vc1, VC1)