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 int delay = ctx->cuparseinfo.ulMaxDisplayDelay;
382 if (ctx->deint_mode != cudaVideoDeinterlaceMode_Weave && !ctx->drop_second_field)
385 return (av_fifo_size(ctx->frame_queue) / sizeof(CuvidParsedFrame)) + delay >= ctx->nb_surfaces;
388 static int cuvid_decode_packet(AVCodecContext *avctx, const AVPacket *avpkt)
390 CuvidContext *ctx = avctx->priv_data;
391 AVHWDeviceContext *device_ctx = (AVHWDeviceContext*)ctx->hwdevice->data;
392 AVCUDADeviceContext *device_hwctx = device_ctx->hwctx;
393 CUcontext dummy, cuda_ctx = device_hwctx->cuda_ctx;
394 CUVIDSOURCEDATAPACKET cupkt;
395 AVPacket filter_packet = { 0 };
396 AVPacket filtered_packet = { 0 };
397 int ret = 0, eret = 0, is_flush = ctx->decoder_flushing;
399 av_log(avctx, AV_LOG_TRACE, "cuvid_decode_packet\n");
401 if (is_flush && avpkt && avpkt->size)
404 if (cuvid_is_buffer_full(avctx) && avpkt && avpkt->size)
405 return AVERROR(EAGAIN);
407 if (ctx->bsf && avpkt && avpkt->size) {
408 if ((ret = av_packet_ref(&filter_packet, avpkt)) < 0) {
409 av_log(avctx, AV_LOG_ERROR, "av_packet_ref failed\n");
413 if ((ret = av_bsf_send_packet(ctx->bsf, &filter_packet)) < 0) {
414 av_log(avctx, AV_LOG_ERROR, "av_bsf_send_packet failed\n");
415 av_packet_unref(&filter_packet);
419 if ((ret = av_bsf_receive_packet(ctx->bsf, &filtered_packet)) < 0) {
420 av_log(avctx, AV_LOG_ERROR, "av_bsf_receive_packet failed\n");
424 avpkt = &filtered_packet;
427 ret = CHECK_CU(ctx->cudl->cuCtxPushCurrent(cuda_ctx));
429 av_packet_unref(&filtered_packet);
433 memset(&cupkt, 0, sizeof(cupkt));
435 if (avpkt && avpkt->size) {
436 cupkt.payload_size = avpkt->size;
437 cupkt.payload = avpkt->data;
439 if (avpkt->pts != AV_NOPTS_VALUE) {
440 cupkt.flags = CUVID_PKT_TIMESTAMP;
441 if (avctx->pkt_timebase.num && avctx->pkt_timebase.den)
442 cupkt.timestamp = av_rescale_q(avpkt->pts, avctx->pkt_timebase, (AVRational){1, 10000000});
444 cupkt.timestamp = avpkt->pts;
447 cupkt.flags = CUVID_PKT_ENDOFSTREAM;
448 ctx->decoder_flushing = 1;
451 ret = CHECK_CU(ctx->cvdl->cuvidParseVideoData(ctx->cuparser, &cupkt));
453 av_packet_unref(&filtered_packet);
458 // cuvidParseVideoData doesn't return an error just because stuff failed...
459 if (ctx->internal_error) {
460 av_log(avctx, AV_LOG_ERROR, "cuvid decode callback error\n");
461 ret = ctx->internal_error;
466 eret = CHECK_CU(ctx->cudl->cuCtxPopCurrent(&dummy));
478 static int cuvid_output_frame(AVCodecContext *avctx, AVFrame *frame)
480 CuvidContext *ctx = avctx->priv_data;
481 AVHWDeviceContext *device_ctx = (AVHWDeviceContext*)ctx->hwdevice->data;
482 AVCUDADeviceContext *device_hwctx = device_ctx->hwctx;
483 CUcontext dummy, cuda_ctx = device_hwctx->cuda_ctx;
484 CUdeviceptr mapped_frame = 0;
485 int ret = 0, eret = 0;
487 av_log(avctx, AV_LOG_TRACE, "cuvid_output_frame\n");
489 if (ctx->decoder_flushing) {
490 ret = cuvid_decode_packet(avctx, NULL);
491 if (ret < 0 && ret != AVERROR_EOF)
495 if (!cuvid_is_buffer_full(avctx)) {
497 ret = ff_decode_get_packet(avctx, &pkt);
498 if (ret < 0 && ret != AVERROR_EOF)
500 ret = cuvid_decode_packet(avctx, &pkt);
501 av_packet_unref(&pkt);
502 // cuvid_is_buffer_full() should avoid this.
503 if (ret == AVERROR(EAGAIN))
504 ret = AVERROR_EXTERNAL;
505 if (ret < 0 && ret != AVERROR_EOF)
509 ret = CHECK_CU(ctx->cudl->cuCtxPushCurrent(cuda_ctx));
513 if (av_fifo_size(ctx->frame_queue)) {
514 CuvidParsedFrame parsed_frame;
515 CUVIDPROCPARAMS params;
516 unsigned int pitch = 0;
520 av_fifo_generic_read(ctx->frame_queue, &parsed_frame, sizeof(CuvidParsedFrame), NULL);
522 memset(¶ms, 0, sizeof(params));
523 params.progressive_frame = parsed_frame.dispinfo.progressive_frame;
524 params.second_field = parsed_frame.second_field;
525 params.top_field_first = parsed_frame.dispinfo.top_field_first;
527 ret = CHECK_CU(ctx->cvdl->cuvidMapVideoFrame(ctx->cudecoder, parsed_frame.dispinfo.picture_index, &mapped_frame, &pitch, ¶ms));
531 if (avctx->pix_fmt == AV_PIX_FMT_CUDA) {
532 ret = av_hwframe_get_buffer(ctx->hwframe, frame, 0);
534 av_log(avctx, AV_LOG_ERROR, "av_hwframe_get_buffer failed\n");
538 ret = ff_decode_frame_props(avctx, frame);
540 av_log(avctx, AV_LOG_ERROR, "ff_decode_frame_props failed\n");
544 for (i = 0; i < 2; i++) {
545 CUDA_MEMCPY2D cpy = {
546 .srcMemoryType = CU_MEMORYTYPE_DEVICE,
547 .dstMemoryType = CU_MEMORYTYPE_DEVICE,
548 .srcDevice = mapped_frame,
549 .dstDevice = (CUdeviceptr)frame->data[i],
551 .dstPitch = frame->linesize[i],
553 .WidthInBytes = FFMIN(pitch, frame->linesize[i]),
554 .Height = avctx->height >> (i ? 1 : 0),
557 ret = CHECK_CU(ctx->cudl->cuMemcpy2DAsync(&cpy, device_hwctx->stream));
561 offset += avctx->height;
564 ret = CHECK_CU(ctx->cudl->cuStreamSynchronize(device_hwctx->stream));
567 } else if (avctx->pix_fmt == AV_PIX_FMT_NV12 ||
568 avctx->pix_fmt == AV_PIX_FMT_P010 ||
569 avctx->pix_fmt == AV_PIX_FMT_P016) {
570 AVFrame *tmp_frame = av_frame_alloc();
572 av_log(avctx, AV_LOG_ERROR, "av_frame_alloc failed\n");
573 ret = AVERROR(ENOMEM);
577 tmp_frame->format = AV_PIX_FMT_CUDA;
578 tmp_frame->hw_frames_ctx = av_buffer_ref(ctx->hwframe);
579 tmp_frame->data[0] = (uint8_t*)mapped_frame;
580 tmp_frame->linesize[0] = pitch;
581 tmp_frame->data[1] = (uint8_t*)(mapped_frame + avctx->height * pitch);
582 tmp_frame->linesize[1] = pitch;
583 tmp_frame->width = avctx->width;
584 tmp_frame->height = avctx->height;
586 ret = ff_get_buffer(avctx, frame, 0);
588 av_log(avctx, AV_LOG_ERROR, "ff_get_buffer failed\n");
589 av_frame_free(&tmp_frame);
593 ret = av_hwframe_transfer_data(frame, tmp_frame, 0);
595 av_log(avctx, AV_LOG_ERROR, "av_hwframe_transfer_data failed\n");
596 av_frame_free(&tmp_frame);
599 av_frame_free(&tmp_frame);
605 frame->key_frame = ctx->key_frame[parsed_frame.dispinfo.picture_index];
606 frame->width = avctx->width;
607 frame->height = avctx->height;
608 if (avctx->pkt_timebase.num && avctx->pkt_timebase.den)
609 frame->pts = av_rescale_q(parsed_frame.dispinfo.timestamp, (AVRational){1, 10000000}, avctx->pkt_timebase);
611 frame->pts = parsed_frame.dispinfo.timestamp;
613 if (parsed_frame.second_field) {
614 if (ctx->prev_pts == INT64_MIN) {
615 ctx->prev_pts = frame->pts;
616 frame->pts += (avctx->pkt_timebase.den * avctx->framerate.den) / (avctx->pkt_timebase.num * avctx->framerate.num);
618 int pts_diff = (frame->pts - ctx->prev_pts) / 2;
619 ctx->prev_pts = frame->pts;
620 frame->pts += pts_diff;
624 /* CUVIDs opaque reordering breaks the internal pkt logic.
625 * So set pkt_pts and clear all the other pkt_ fields.
628 FF_DISABLE_DEPRECATION_WARNINGS
629 frame->pkt_pts = frame->pts;
630 FF_ENABLE_DEPRECATION_WARNINGS
633 frame->pkt_duration = 0;
634 frame->pkt_size = -1;
636 frame->interlaced_frame = !parsed_frame.is_deinterlacing && !parsed_frame.dispinfo.progressive_frame;
638 if (frame->interlaced_frame)
639 frame->top_field_first = parsed_frame.dispinfo.top_field_first;
640 } else if (ctx->decoder_flushing) {
643 ret = AVERROR(EAGAIN);
648 eret = CHECK_CU(ctx->cvdl->cuvidUnmapVideoFrame(ctx->cudecoder, mapped_frame));
650 eret = CHECK_CU(ctx->cudl->cuCtxPopCurrent(&dummy));
658 static int cuvid_decode_frame(AVCodecContext *avctx, void *data, int *got_frame, AVPacket *avpkt)
660 CuvidContext *ctx = avctx->priv_data;
661 AVFrame *frame = data;
664 av_log(avctx, AV_LOG_TRACE, "cuvid_decode_frame\n");
666 if (ctx->deint_mode_current != cudaVideoDeinterlaceMode_Weave) {
667 av_log(avctx, AV_LOG_ERROR, "Deinterlacing is not supported via the old API\n");
668 return AVERROR(EINVAL);
671 if (!ctx->decoder_flushing) {
672 ret = cuvid_decode_packet(avctx, avpkt);
677 ret = cuvid_output_frame(avctx, frame);
678 if (ret == AVERROR(EAGAIN) || ret == AVERROR_EOF) {
680 } else if (ret < 0) {
689 static av_cold int cuvid_decode_end(AVCodecContext *avctx)
691 CuvidContext *ctx = avctx->priv_data;
693 av_fifo_freep(&ctx->frame_queue);
696 av_bsf_free(&ctx->bsf);
699 ctx->cvdl->cuvidDestroyVideoParser(ctx->cuparser);
702 ctx->cvdl->cuvidDestroyDecoder(ctx->cudecoder);
706 av_buffer_unref(&ctx->hwframe);
707 av_buffer_unref(&ctx->hwdevice);
709 av_freep(&ctx->key_frame);
711 cuvid_free_functions(&ctx->cvdl);
716 static int cuvid_test_capabilities(AVCodecContext *avctx,
717 const CUVIDPARSERPARAMS *cuparseinfo,
722 CuvidContext *ctx = avctx->priv_data;
723 CUVIDDECODECAPS *caps;
724 int res8 = 0, res10 = 0, res12 = 0;
726 if (!ctx->cvdl->cuvidGetDecoderCaps) {
727 av_log(avctx, AV_LOG_WARNING, "Used Nvidia driver is too old to perform a capability check.\n");
728 av_log(avctx, AV_LOG_WARNING, "The minimum required version is "
729 #if defined(_WIN32) || defined(__CYGWIN__)
734 ". Continuing blind.\n");
735 ctx->caps8.bIsSupported = ctx->caps10.bIsSupported = 1;
736 // 12 bit was not supported before the capability check was introduced, so disable it.
737 ctx->caps12.bIsSupported = 0;
741 ctx->caps8.eCodecType = ctx->caps10.eCodecType = ctx->caps12.eCodecType
742 = cuparseinfo->CodecType;
743 ctx->caps8.eChromaFormat = ctx->caps10.eChromaFormat = ctx->caps12.eChromaFormat
744 = cudaVideoChromaFormat_420;
746 ctx->caps8.nBitDepthMinus8 = 0;
747 ctx->caps10.nBitDepthMinus8 = 2;
748 ctx->caps12.nBitDepthMinus8 = 4;
750 res8 = CHECK_CU(ctx->cvdl->cuvidGetDecoderCaps(&ctx->caps8));
751 res10 = CHECK_CU(ctx->cvdl->cuvidGetDecoderCaps(&ctx->caps10));
752 res12 = CHECK_CU(ctx->cvdl->cuvidGetDecoderCaps(&ctx->caps12));
754 av_log(avctx, AV_LOG_VERBOSE, "CUVID capabilities for %s:\n", avctx->codec->name);
755 av_log(avctx, AV_LOG_VERBOSE, "8 bit: supported: %d, min_width: %d, max_width: %d, min_height: %d, max_height: %d\n",
756 ctx->caps8.bIsSupported, ctx->caps8.nMinWidth, ctx->caps8.nMaxWidth, ctx->caps8.nMinHeight, ctx->caps8.nMaxHeight);
757 av_log(avctx, AV_LOG_VERBOSE, "10 bit: supported: %d, min_width: %d, max_width: %d, min_height: %d, max_height: %d\n",
758 ctx->caps10.bIsSupported, ctx->caps10.nMinWidth, ctx->caps10.nMaxWidth, ctx->caps10.nMinHeight, ctx->caps10.nMaxHeight);
759 av_log(avctx, AV_LOG_VERBOSE, "12 bit: supported: %d, min_width: %d, max_width: %d, min_height: %d, max_height: %d\n",
760 ctx->caps12.bIsSupported, ctx->caps12.nMinWidth, ctx->caps12.nMaxWidth, ctx->caps12.nMinHeight, ctx->caps12.nMaxHeight);
779 if (!ctx->caps8.bIsSupported) {
780 av_log(avctx, AV_LOG_ERROR, "Codec %s is not supported.\n", avctx->codec->name);
781 return AVERROR(EINVAL);
784 if (!caps->bIsSupported) {
785 av_log(avctx, AV_LOG_ERROR, "Bit depth %d is not supported.\n", bit_depth);
786 return AVERROR(EINVAL);
789 if (probed_width > caps->nMaxWidth || probed_width < caps->nMinWidth) {
790 av_log(avctx, AV_LOG_ERROR, "Video width %d not within range from %d to %d\n",
791 probed_width, caps->nMinWidth, caps->nMaxWidth);
792 return AVERROR(EINVAL);
795 if (probed_height > caps->nMaxHeight || probed_height < caps->nMinHeight) {
796 av_log(avctx, AV_LOG_ERROR, "Video height %d not within range from %d to %d\n",
797 probed_height, caps->nMinHeight, caps->nMaxHeight);
798 return AVERROR(EINVAL);
804 static av_cold int cuvid_decode_init(AVCodecContext *avctx)
806 CuvidContext *ctx = avctx->priv_data;
807 AVCUDADeviceContext *device_hwctx;
808 AVHWDeviceContext *device_ctx;
809 AVHWFramesContext *hwframe_ctx;
810 CUVIDSOURCEDATAPACKET seq_pkt;
811 CUcontext cuda_ctx = NULL;
813 const AVBitStreamFilter *bsf;
816 enum AVPixelFormat pix_fmts[3] = { AV_PIX_FMT_CUDA,
820 int probed_width = avctx->coded_width ? avctx->coded_width : 1280;
821 int probed_height = avctx->coded_height ? avctx->coded_height : 720;
822 int probed_bit_depth = 8;
824 const AVPixFmtDescriptor *probe_desc = av_pix_fmt_desc_get(avctx->pix_fmt);
825 if (probe_desc && probe_desc->nb_components)
826 probed_bit_depth = probe_desc->comp[0].depth;
828 // Accelerated transcoding scenarios with 'ffmpeg' require that the
829 // pix_fmt be set to AV_PIX_FMT_CUDA early. The sw_pix_fmt, and the
830 // pix_fmt for non-accelerated transcoding, do not need to be correct
831 // but need to be set to something. We arbitrarily pick NV12.
832 ret = ff_get_format(avctx, pix_fmts);
834 av_log(avctx, AV_LOG_ERROR, "ff_get_format failed: %d\n", ret);
837 avctx->pix_fmt = ret;
839 if (ctx->resize_expr && sscanf(ctx->resize_expr, "%dx%d",
840 &ctx->resize.width, &ctx->resize.height) != 2) {
841 av_log(avctx, AV_LOG_ERROR, "Invalid resize expressions\n");
842 ret = AVERROR(EINVAL);
846 if (ctx->crop_expr && sscanf(ctx->crop_expr, "%dx%dx%dx%d",
847 &ctx->crop.top, &ctx->crop.bottom,
848 &ctx->crop.left, &ctx->crop.right) != 4) {
849 av_log(avctx, AV_LOG_ERROR, "Invalid cropping expressions\n");
850 ret = AVERROR(EINVAL);
854 ret = cuvid_load_functions(&ctx->cvdl, avctx);
856 av_log(avctx, AV_LOG_ERROR, "Failed loading nvcuvid.\n");
860 ctx->frame_queue = av_fifo_alloc(ctx->nb_surfaces * sizeof(CuvidParsedFrame));
861 if (!ctx->frame_queue) {
862 ret = AVERROR(ENOMEM);
866 if (avctx->hw_frames_ctx) {
867 ctx->hwframe = av_buffer_ref(avctx->hw_frames_ctx);
869 ret = AVERROR(ENOMEM);
873 hwframe_ctx = (AVHWFramesContext*)ctx->hwframe->data;
875 ctx->hwdevice = av_buffer_ref(hwframe_ctx->device_ref);
876 if (!ctx->hwdevice) {
877 ret = AVERROR(ENOMEM);
881 if (avctx->hw_device_ctx) {
882 ctx->hwdevice = av_buffer_ref(avctx->hw_device_ctx);
883 if (!ctx->hwdevice) {
884 ret = AVERROR(ENOMEM);
888 ret = av_hwdevice_ctx_create(&ctx->hwdevice, AV_HWDEVICE_TYPE_CUDA, ctx->cu_gpu, NULL, 0);
893 ctx->hwframe = av_hwframe_ctx_alloc(ctx->hwdevice);
895 av_log(avctx, AV_LOG_ERROR, "av_hwframe_ctx_alloc failed\n");
896 ret = AVERROR(ENOMEM);
900 hwframe_ctx = (AVHWFramesContext*)ctx->hwframe->data;
903 device_ctx = hwframe_ctx->device_ctx;
904 device_hwctx = device_ctx->hwctx;
906 cuda_ctx = device_hwctx->cuda_ctx;
907 ctx->cudl = device_hwctx->internal->cuda_dl;
909 memset(&ctx->cuparseinfo, 0, sizeof(ctx->cuparseinfo));
910 memset(&ctx->cuparse_ext, 0, sizeof(ctx->cuparse_ext));
911 memset(&seq_pkt, 0, sizeof(seq_pkt));
913 ctx->cuparseinfo.pExtVideoInfo = &ctx->cuparse_ext;
915 switch (avctx->codec->id) {
916 #if CONFIG_H264_CUVID_DECODER
917 case AV_CODEC_ID_H264:
918 ctx->cuparseinfo.CodecType = cudaVideoCodec_H264;
921 #if CONFIG_HEVC_CUVID_DECODER
922 case AV_CODEC_ID_HEVC:
923 ctx->cuparseinfo.CodecType = cudaVideoCodec_HEVC;
926 #if CONFIG_MJPEG_CUVID_DECODER
927 case AV_CODEC_ID_MJPEG:
928 ctx->cuparseinfo.CodecType = cudaVideoCodec_JPEG;
931 #if CONFIG_MPEG1_CUVID_DECODER
932 case AV_CODEC_ID_MPEG1VIDEO:
933 ctx->cuparseinfo.CodecType = cudaVideoCodec_MPEG1;
936 #if CONFIG_MPEG2_CUVID_DECODER
937 case AV_CODEC_ID_MPEG2VIDEO:
938 ctx->cuparseinfo.CodecType = cudaVideoCodec_MPEG2;
941 #if CONFIG_MPEG4_CUVID_DECODER
942 case AV_CODEC_ID_MPEG4:
943 ctx->cuparseinfo.CodecType = cudaVideoCodec_MPEG4;
946 #if CONFIG_VP8_CUVID_DECODER
947 case AV_CODEC_ID_VP8:
948 ctx->cuparseinfo.CodecType = cudaVideoCodec_VP8;
951 #if CONFIG_VP9_CUVID_DECODER
952 case AV_CODEC_ID_VP9:
953 ctx->cuparseinfo.CodecType = cudaVideoCodec_VP9;
956 #if CONFIG_VC1_CUVID_DECODER
957 case AV_CODEC_ID_VC1:
958 ctx->cuparseinfo.CodecType = cudaVideoCodec_VC1;
962 av_log(avctx, AV_LOG_ERROR, "Invalid CUVID codec!\n");
966 if (avctx->codec->id == AV_CODEC_ID_H264 || avctx->codec->id == AV_CODEC_ID_HEVC) {
967 if (avctx->codec->id == AV_CODEC_ID_H264)
968 bsf = av_bsf_get_by_name("h264_mp4toannexb");
970 bsf = av_bsf_get_by_name("hevc_mp4toannexb");
973 ret = AVERROR_BSF_NOT_FOUND;
976 if (ret = av_bsf_alloc(bsf, &ctx->bsf)) {
979 if (((ret = avcodec_parameters_from_context(ctx->bsf->par_in, avctx)) < 0) || ((ret = av_bsf_init(ctx->bsf)) < 0)) {
980 av_bsf_free(&ctx->bsf);
984 ctx->cuparse_ext.format.seqhdr_data_length = ctx->bsf->par_out->extradata_size;
985 memcpy(ctx->cuparse_ext.raw_seqhdr_data,
986 ctx->bsf->par_out->extradata,
987 FFMIN(sizeof(ctx->cuparse_ext.raw_seqhdr_data), ctx->bsf->par_out->extradata_size));
988 } else if (avctx->extradata_size > 0) {
989 ctx->cuparse_ext.format.seqhdr_data_length = avctx->extradata_size;
990 memcpy(ctx->cuparse_ext.raw_seqhdr_data,
992 FFMIN(sizeof(ctx->cuparse_ext.raw_seqhdr_data), avctx->extradata_size));
995 ctx->key_frame = av_mallocz(ctx->nb_surfaces * sizeof(int));
996 if (!ctx->key_frame) {
997 ret = AVERROR(ENOMEM);
1001 ctx->cuparseinfo.ulMaxNumDecodeSurfaces = ctx->nb_surfaces;
1002 ctx->cuparseinfo.ulMaxDisplayDelay = 4;
1003 ctx->cuparseinfo.pUserData = avctx;
1004 ctx->cuparseinfo.pfnSequenceCallback = cuvid_handle_video_sequence;
1005 ctx->cuparseinfo.pfnDecodePicture = cuvid_handle_picture_decode;
1006 ctx->cuparseinfo.pfnDisplayPicture = cuvid_handle_picture_display;
1008 ret = CHECK_CU(ctx->cudl->cuCtxPushCurrent(cuda_ctx));
1012 ret = cuvid_test_capabilities(avctx, &ctx->cuparseinfo,
1019 ret = CHECK_CU(ctx->cvdl->cuvidCreateVideoParser(&ctx->cuparser, &ctx->cuparseinfo));
1023 seq_pkt.payload = ctx->cuparse_ext.raw_seqhdr_data;
1024 seq_pkt.payload_size = ctx->cuparse_ext.format.seqhdr_data_length;
1026 if (seq_pkt.payload && seq_pkt.payload_size) {
1027 ret = CHECK_CU(ctx->cvdl->cuvidParseVideoData(ctx->cuparser, &seq_pkt));
1032 ret = CHECK_CU(ctx->cudl->cuCtxPopCurrent(&dummy));
1036 ctx->prev_pts = INT64_MIN;
1038 if (!avctx->pkt_timebase.num || !avctx->pkt_timebase.den)
1039 av_log(avctx, AV_LOG_WARNING, "Invalid pkt_timebase, passing timestamps as-is.\n");
1044 cuvid_decode_end(avctx);
1048 static void cuvid_flush(AVCodecContext *avctx)
1050 CuvidContext *ctx = avctx->priv_data;
1051 AVHWDeviceContext *device_ctx = (AVHWDeviceContext*)ctx->hwdevice->data;
1052 AVCUDADeviceContext *device_hwctx = device_ctx->hwctx;
1053 CUcontext dummy, cuda_ctx = device_hwctx->cuda_ctx;
1054 CUVIDSOURCEDATAPACKET seq_pkt = { 0 };
1057 ret = CHECK_CU(ctx->cudl->cuCtxPushCurrent(cuda_ctx));
1061 av_fifo_freep(&ctx->frame_queue);
1063 ctx->frame_queue = av_fifo_alloc(ctx->nb_surfaces * sizeof(CuvidParsedFrame));
1064 if (!ctx->frame_queue) {
1065 av_log(avctx, AV_LOG_ERROR, "Failed to recreate frame queue on flush\n");
1069 if (ctx->cudecoder) {
1070 ctx->cvdl->cuvidDestroyDecoder(ctx->cudecoder);
1071 ctx->cudecoder = NULL;
1074 if (ctx->cuparser) {
1075 ctx->cvdl->cuvidDestroyVideoParser(ctx->cuparser);
1076 ctx->cuparser = NULL;
1079 ret = CHECK_CU(ctx->cvdl->cuvidCreateVideoParser(&ctx->cuparser, &ctx->cuparseinfo));
1083 seq_pkt.payload = ctx->cuparse_ext.raw_seqhdr_data;
1084 seq_pkt.payload_size = ctx->cuparse_ext.format.seqhdr_data_length;
1086 if (seq_pkt.payload && seq_pkt.payload_size) {
1087 ret = CHECK_CU(ctx->cvdl->cuvidParseVideoData(ctx->cuparser, &seq_pkt));
1092 ret = CHECK_CU(ctx->cudl->cuCtxPopCurrent(&dummy));
1096 ctx->prev_pts = INT64_MIN;
1097 ctx->decoder_flushing = 0;
1101 av_log(avctx, AV_LOG_ERROR, "CUDA reinit on flush failed\n");
1104 #define OFFSET(x) offsetof(CuvidContext, x)
1105 #define VD AV_OPT_FLAG_VIDEO_PARAM | AV_OPT_FLAG_DECODING_PARAM
1106 static const AVOption options[] = {
1107 { "deint", "Set deinterlacing mode", OFFSET(deint_mode), AV_OPT_TYPE_INT, { .i64 = cudaVideoDeinterlaceMode_Weave }, cudaVideoDeinterlaceMode_Weave, cudaVideoDeinterlaceMode_Adaptive, VD, "deint" },
1108 { "weave", "Weave deinterlacing (do nothing)", 0, AV_OPT_TYPE_CONST, { .i64 = cudaVideoDeinterlaceMode_Weave }, 0, 0, VD, "deint" },
1109 { "bob", "Bob deinterlacing", 0, AV_OPT_TYPE_CONST, { .i64 = cudaVideoDeinterlaceMode_Bob }, 0, 0, VD, "deint" },
1110 { "adaptive", "Adaptive deinterlacing", 0, AV_OPT_TYPE_CONST, { .i64 = cudaVideoDeinterlaceMode_Adaptive }, 0, 0, VD, "deint" },
1111 { "gpu", "GPU to be used for decoding", OFFSET(cu_gpu), AV_OPT_TYPE_STRING, { .str = NULL }, 0, 0, VD },
1112 { "surfaces", "Maximum surfaces to be used for decoding", OFFSET(nb_surfaces), AV_OPT_TYPE_INT, { .i64 = 25 }, 0, INT_MAX, VD },
1113 { "drop_second_field", "Drop second field when deinterlacing", OFFSET(drop_second_field), AV_OPT_TYPE_BOOL, { .i64 = 0 }, 0, 1, VD },
1114 { "crop", "Crop (top)x(bottom)x(left)x(right)", OFFSET(crop_expr), AV_OPT_TYPE_STRING, { .str = NULL }, 0, 0, VD },
1115 { "resize", "Resize (width)x(height)", OFFSET(resize_expr), AV_OPT_TYPE_STRING, { .str = NULL }, 0, 0, VD },
1119 static const AVCodecHWConfigInternal *cuvid_hw_configs[] = {
1120 &(const AVCodecHWConfigInternal) {
1122 .pix_fmt = AV_PIX_FMT_CUDA,
1123 .methods = AV_CODEC_HW_CONFIG_METHOD_HW_DEVICE_CTX |
1124 AV_CODEC_HW_CONFIG_METHOD_INTERNAL,
1125 .device_type = AV_HWDEVICE_TYPE_CUDA
1132 #define DEFINE_CUVID_CODEC(x, X) \
1133 static const AVClass x##_cuvid_class = { \
1134 .class_name = #x "_cuvid", \
1135 .item_name = av_default_item_name, \
1136 .option = options, \
1137 .version = LIBAVUTIL_VERSION_INT, \
1139 AVCodec ff_##x##_cuvid_decoder = { \
1140 .name = #x "_cuvid", \
1141 .long_name = NULL_IF_CONFIG_SMALL("Nvidia CUVID " #X " decoder"), \
1142 .type = AVMEDIA_TYPE_VIDEO, \
1143 .id = AV_CODEC_ID_##X, \
1144 .priv_data_size = sizeof(CuvidContext), \
1145 .priv_class = &x##_cuvid_class, \
1146 .init = cuvid_decode_init, \
1147 .close = cuvid_decode_end, \
1148 .decode = cuvid_decode_frame, \
1149 .receive_frame = cuvid_output_frame, \
1150 .flush = cuvid_flush, \
1151 .capabilities = AV_CODEC_CAP_DELAY | AV_CODEC_CAP_AVOID_PROBING | AV_CODEC_CAP_HARDWARE, \
1152 .pix_fmts = (const enum AVPixelFormat[]){ AV_PIX_FMT_CUDA, \
1156 AV_PIX_FMT_NONE }, \
1157 .hw_configs = cuvid_hw_configs, \
1158 .wrapper_name = "cuvid", \
1161 #if CONFIG_HEVC_CUVID_DECODER
1162 DEFINE_CUVID_CODEC(hevc, HEVC)
1165 #if CONFIG_H264_CUVID_DECODER
1166 DEFINE_CUVID_CODEC(h264, H264)
1169 #if CONFIG_MJPEG_CUVID_DECODER
1170 DEFINE_CUVID_CODEC(mjpeg, MJPEG)
1173 #if CONFIG_MPEG1_CUVID_DECODER
1174 DEFINE_CUVID_CODEC(mpeg1, MPEG1VIDEO)
1177 #if CONFIG_MPEG2_CUVID_DECODER
1178 DEFINE_CUVID_CODEC(mpeg2, MPEG2VIDEO)
1181 #if CONFIG_MPEG4_CUVID_DECODER
1182 DEFINE_CUVID_CODEC(mpeg4, MPEG4)
1185 #if CONFIG_VP8_CUVID_DECODER
1186 DEFINE_CUVID_CODEC(vp8, VP8)
1189 #if CONFIG_VP9_CUVID_DECODER
1190 DEFINE_CUVID_CODEC(vp9, VP9)
1193 #if CONFIG_VC1_CUVID_DECODER
1194 DEFINE_CUVID_CODEC(vc1, VC1)