1 #include "ffmpeg_capture.h"
13 #include <libavcodec/avcodec.h>
14 #include <libavformat/avformat.h>
15 #include <libavutil/avutil.h>
16 #include <libavutil/error.h>
17 #include <libavutil/frame.h>
18 #include <libavutil/imgutils.h>
19 #include <libavutil/mem.h>
20 #include <libavutil/pixfmt.h>
21 #include <libavutil/opt.h>
22 #include <libswscale/swscale.h>
29 #include <unordered_set>
33 #include <movit/colorspace_conversion_effect.h>
35 #include "bmusb/bmusb.h"
36 #include "shared/ffmpeg_raii.h"
37 #include "ffmpeg_util.h"
39 #include "ref_counted_frame.h"
40 #include "shared/timebase.h"
47 using namespace std::chrono;
48 using namespace bmusb;
49 using namespace movit;
50 using namespace Eigen;
52 // Avoid deprecation warnings, but we don't want to drop FFmpeg 5.1 support just yet.
53 #if LIBAVUTIL_VERSION_INT >= AV_VERSION_INT(57, 30, 100)
54 #define pkt_duration duration
59 steady_clock::time_point compute_frame_start(int64_t frame_pts, int64_t pts_origin, const AVRational &video_timebase, const steady_clock::time_point &origin, double rate)
61 const duration<double> pts((frame_pts - pts_origin) * double(video_timebase.num) / double(video_timebase.den));
62 return origin + duration_cast<steady_clock::duration>(pts / rate);
65 bool changed_since(const std::string &pathname, const timespec &ts)
71 if (stat(pathname.c_str(), &buf) != 0) {
72 fprintf(stderr, "%s: Couldn't check for new version, leaving the old in place.\n", pathname.c_str());
75 return (buf.st_mtim.tv_sec != ts.tv_sec || buf.st_mtim.tv_nsec != ts.tv_nsec);
78 bool is_full_range(const AVPixFmtDescriptor *desc)
80 // This is horrible, but there's no better way that I know of.
81 return (strchr(desc->name, 'j') != nullptr);
84 AVPixelFormat decide_dst_format(AVPixelFormat src_format, bmusb::PixelFormat dst_format_type)
86 if (dst_format_type == bmusb::PixelFormat_8BitBGRA) {
87 return AV_PIX_FMT_BGRA;
89 if (dst_format_type == FFmpegCapture::PixelFormat_NV12) {
90 return AV_PIX_FMT_NV12;
93 assert(dst_format_type == bmusb::PixelFormat_8BitYCbCrPlanar);
95 // If this is a non-Y'CbCr format, just convert to 4:4:4 Y'CbCr
96 // and be done with it. It's too strange to spend a lot of time on.
97 // (Let's hope there's no alpha.)
98 const AVPixFmtDescriptor *src_desc = av_pix_fmt_desc_get(src_format);
99 if (src_desc == nullptr ||
100 src_desc->nb_components != 3 ||
101 (src_desc->flags & AV_PIX_FMT_FLAG_RGB)) {
102 return AV_PIX_FMT_YUV444P;
105 // The best for us would be Cb and Cr together if possible,
106 // but FFmpeg doesn't support that except in the special case of
107 // NV12, so we need to go to planar even for the case of NV12.
108 // Thus, look for the closest (but no worse) 8-bit planar Y'CbCr format
109 // that matches in color range. (This will also include the case of
110 // the source format already being acceptable.)
111 bool src_full_range = is_full_range(src_desc);
112 const char *best_format = "yuv444p";
113 unsigned best_score = numeric_limits<unsigned>::max();
114 for (const AVPixFmtDescriptor *desc = av_pix_fmt_desc_next(nullptr);
116 desc = av_pix_fmt_desc_next(desc)) {
117 // Find planar Y'CbCr formats only.
118 if (desc->nb_components != 3) continue;
119 if (desc->flags & AV_PIX_FMT_FLAG_RGB) continue;
120 if (!(desc->flags & AV_PIX_FMT_FLAG_PLANAR)) continue;
121 if (desc->comp[0].plane != 0 ||
122 desc->comp[1].plane != 1 ||
123 desc->comp[2].plane != 2) continue;
125 // 8-bit formats only.
126 if (desc->flags & AV_PIX_FMT_FLAG_BE) continue;
127 if (desc->comp[0].depth != 8) continue;
129 // Same or better chroma resolution only.
130 int chroma_w_diff = src_desc->log2_chroma_w - desc->log2_chroma_w;
131 int chroma_h_diff = src_desc->log2_chroma_h - desc->log2_chroma_h;
132 if (chroma_w_diff < 0 || chroma_h_diff < 0)
135 // Matching full/limited range only.
136 if (is_full_range(desc) != src_full_range)
139 // Pick something with as little excess chroma resolution as possible.
140 unsigned score = (1 << (chroma_w_diff)) << chroma_h_diff;
141 if (score < best_score) {
143 best_format = desc->name;
146 return av_get_pix_fmt(best_format);
149 YCbCrFormat decode_ycbcr_format(const AVPixFmtDescriptor *desc, const AVFrame *frame, bool is_mjpeg, AVColorSpace *last_colorspace, AVChromaLocation *last_chroma_location)
152 AVColorSpace colorspace = frame->colorspace;
153 switch (colorspace) {
154 case AVCOL_SPC_BT709:
155 format.luma_coefficients = YCBCR_REC_709;
157 case AVCOL_SPC_BT470BG:
158 case AVCOL_SPC_SMPTE170M:
159 case AVCOL_SPC_SMPTE240M:
160 format.luma_coefficients = YCBCR_REC_601;
162 case AVCOL_SPC_BT2020_NCL:
163 format.luma_coefficients = YCBCR_REC_2020;
165 case AVCOL_SPC_UNSPECIFIED:
166 format.luma_coefficients = (frame->height >= 720 ? YCBCR_REC_709 : YCBCR_REC_601);
169 if (colorspace != *last_colorspace) {
170 fprintf(stderr, "Unknown Y'CbCr coefficient enum %d from FFmpeg; choosing Rec. 709.\n",
173 format.luma_coefficients = YCBCR_REC_709;
176 *last_colorspace = colorspace;
178 format.full_range = is_full_range(desc);
179 format.num_levels = 1 << desc->comp[0].depth;
180 format.chroma_subsampling_x = 1 << desc->log2_chroma_w;
181 format.chroma_subsampling_y = 1 << desc->log2_chroma_h;
183 switch (frame->chroma_location) {
184 case AVCHROMA_LOC_LEFT:
185 format.cb_x_position = 0.0;
186 format.cb_y_position = 0.5;
188 case AVCHROMA_LOC_CENTER:
189 format.cb_x_position = 0.5;
190 format.cb_y_position = 0.5;
192 case AVCHROMA_LOC_TOPLEFT:
193 format.cb_x_position = 0.0;
194 format.cb_y_position = 0.0;
196 case AVCHROMA_LOC_TOP:
197 format.cb_x_position = 0.5;
198 format.cb_y_position = 0.0;
200 case AVCHROMA_LOC_BOTTOMLEFT:
201 format.cb_x_position = 0.0;
202 format.cb_y_position = 1.0;
204 case AVCHROMA_LOC_BOTTOM:
205 format.cb_x_position = 0.5;
206 format.cb_y_position = 1.0;
209 if (frame->chroma_location != *last_chroma_location) {
210 fprintf(stderr, "Unknown chroma location coefficient enum %d from FFmpeg; choosing center.\n",
211 frame->chroma_location);
213 format.cb_x_position = 0.5;
214 format.cb_y_position = 0.5;
217 *last_chroma_location = frame->chroma_location;
219 if (is_mjpeg && !format.full_range) {
220 // Limited-range MJPEG is only detected by FFmpeg whenever a special
221 // JPEG comment is set, which means that in practice, the stream is
222 // almost certainly generated by Futatabi. Override FFmpeg's forced
223 // MJPEG defaults (it disregards the values set in the mux) with what
225 format.luma_coefficients = YCBCR_REC_709;
226 format.cb_x_position = 0.0;
227 format.cb_y_position = 0.5;
230 format.cr_x_position = format.cb_x_position;
231 format.cr_y_position = format.cb_y_position;
235 RGBTriplet get_neutral_color(AVDictionary *metadata)
237 if (metadata == nullptr) {
238 return RGBTriplet(1.0f, 1.0f, 1.0f);
240 AVDictionaryEntry *entry = av_dict_get(metadata, "WhitePoint", nullptr, 0);
241 if (entry == nullptr) {
242 return RGBTriplet(1.0f, 1.0f, 1.0f);
245 unsigned x_nom, x_den, y_nom, y_den;
246 if (sscanf(entry->value, " %u:%u , %u:%u", &x_nom, &x_den, &y_nom, &y_den) != 4) {
247 fprintf(stderr, "WARNING: Unable to parse white point '%s', using default white point\n", entry->value);
248 return RGBTriplet(1.0f, 1.0f, 1.0f);
251 double x = double(x_nom) / x_den;
252 double y = double(y_nom) / y_den;
253 double z = 1.0 - x - y;
255 Matrix3d rgb_to_xyz_matrix = movit::ColorspaceConversionEffect::get_xyz_matrix(COLORSPACE_sRGB);
256 Vector3d rgb = rgb_to_xyz_matrix.inverse() * Vector3d(x, y, z);
258 return RGBTriplet(rgb[0], rgb[1], rgb[2]);
263 FFmpegCapture::FFmpegCapture(const string &filename, unsigned width, unsigned height)
264 : filename(filename), width(width), height(height), video_timebase{1, 1}
266 description = "Video: " + filename;
268 last_frame = steady_clock::now();
270 avformat_network_init(); // In case someone wants this.
274 FFmpegCapture::FFmpegCapture(int srt_sock, const string &stream_id)
275 : srt_sock(srt_sock),
276 width(0), // Don't resize; SRT streams typically have stable resolution, and should behave much like regular cards in general.
278 pixel_format(bmusb::PixelFormat_8BitYCbCrPlanar),
281 if (stream_id.empty()) {
282 description = "SRT stream";
284 description = stream_id;
286 play_as_fast_as_possible = true;
288 last_frame = steady_clock::now();
292 FFmpegCapture::~FFmpegCapture()
294 if (has_dequeue_callbacks) {
295 dequeue_cleanup_callback();
297 swr_free(&resampler);
299 if (srt_sock != -1) {
305 void FFmpegCapture::configure_card()
307 if (video_frame_allocator == nullptr) {
308 owned_video_frame_allocator.reset(new MallocFrameAllocator(FRAME_SIZE, NUM_QUEUED_VIDEO_FRAMES));
309 set_video_frame_allocator(owned_video_frame_allocator.get());
311 if (audio_frame_allocator == nullptr) {
312 // Audio can come out in pretty large chunks, so increase from the default 1 MB.
313 owned_audio_frame_allocator.reset(new MallocFrameAllocator(1 << 20, NUM_QUEUED_AUDIO_FRAMES));
314 set_audio_frame_allocator(owned_audio_frame_allocator.get());
318 void FFmpegCapture::start_bm_capture()
324 producer_thread_should_quit.unquit();
325 producer_thread = thread(&FFmpegCapture::producer_thread_func, this);
328 void FFmpegCapture::stop_dequeue_thread()
334 producer_thread_should_quit.quit();
335 producer_thread.join();
338 std::map<uint32_t, VideoMode> FFmpegCapture::get_available_video_modes() const
340 // Note: This will never really be shown in the UI.
344 snprintf(buf, sizeof(buf), "%ux%u", sws_last_width, sws_last_height);
347 mode.autodetect = false;
348 mode.width = sws_last_width;
349 mode.height = sws_last_height;
350 mode.frame_rate_num = 60;
351 mode.frame_rate_den = 1;
352 mode.interlaced = false;
354 return {{ 0, mode }};
357 void FFmpegCapture::producer_thread_func()
359 char thread_name[16];
360 snprintf(thread_name, sizeof(thread_name), "FFmpeg_C_%d", card_index);
361 pthread_setname_np(pthread_self(), thread_name);
363 while (!producer_thread_should_quit.should_quit()) {
364 string filename_copy;
366 lock_guard<mutex> lock(filename_mu);
367 filename_copy = filename;
371 if (srt_sock == -1) {
372 pathname = search_for_file(filename_copy);
374 pathname = description;
376 if (pathname.empty()) {
377 send_disconnected_frame();
381 producer_thread_should_quit.sleep_for(seconds(1));
382 fprintf(stderr, "%s not found, sleeping one second and trying again...\n", filename_copy.c_str());
385 should_interrupt = false;
386 if (!play_video(pathname)) {
388 send_disconnected_frame();
392 fprintf(stderr, "Error when playing %s, sleeping one second and trying again...\n", pathname.c_str());
393 producer_thread_should_quit.sleep_for(seconds(1));
398 send_disconnected_frame();
402 // Probably just EOF, will exit the loop above on next test.
405 if (has_dequeue_callbacks) {
406 dequeue_cleanup_callback();
407 has_dequeue_callbacks = false;
411 void FFmpegCapture::send_disconnected_frame()
413 // Send an empty frame to signal that we have no signal anymore.
414 FrameAllocator::Frame video_frame = video_frame_allocator->alloc_frame();
415 size_t frame_width = width == 0 ? global_flags.width : width;
416 size_t frame_height = height == 0 ? global_flags.height : height;
417 if (video_frame.data) {
418 VideoFormat video_format;
419 video_format.width = frame_width;
420 video_format.height = frame_height;
421 video_format.frame_rate_nom = 60;
422 video_format.frame_rate_den = 1;
423 video_format.is_connected = false;
424 if (pixel_format == bmusb::PixelFormat_8BitBGRA) {
425 video_format.stride = frame_width * 4;
426 video_frame.len = frame_width * frame_height * 4;
427 memset(video_frame.data, 0, video_frame.len);
429 video_format.stride = frame_width;
430 current_frame_ycbcr_format.luma_coefficients = YCBCR_REC_709;
431 current_frame_ycbcr_format.full_range = true;
432 current_frame_ycbcr_format.num_levels = 256;
433 current_frame_ycbcr_format.chroma_subsampling_x = 2;
434 current_frame_ycbcr_format.chroma_subsampling_y = 2;
435 current_frame_ycbcr_format.cb_x_position = 0.0f;
436 current_frame_ycbcr_format.cb_y_position = 0.0f;
437 current_frame_ycbcr_format.cr_x_position = 0.0f;
438 current_frame_ycbcr_format.cr_y_position = 0.0f;
439 video_frame.len = frame_width * frame_height * 2;
440 memset(video_frame.data, 0, frame_width * frame_height);
441 memset(video_frame.data + frame_width * frame_height, 128, frame_width * frame_height); // Valid for both NV12 and planar.
444 if (frame_callback != nullptr) {
445 frame_callback(-1, AVRational{1, TIMEBASE}, -1, AVRational{1, TIMEBASE}, timecode++,
446 video_frame, /*video_offset=*/0, video_format,
447 FrameAllocator::Frame(), /*audio_offset=*/0, AudioFormat());
449 last_frame_was_connected = false;
454 if (card_disconnected_callback != nullptr) {
455 card_disconnected_callback();
460 template<AVHWDeviceType type>
461 AVPixelFormat get_hw_format(AVCodecContext *ctx, const AVPixelFormat *fmt)
463 bool found_config_of_right_type = false;
464 for (int i = 0;; ++i) { // Termination condition inside loop.
465 const AVCodecHWConfig *config = avcodec_get_hw_config(ctx->codec, i);
466 if (config == nullptr) { // End of list.
469 if (!(config->methods & AV_CODEC_HW_CONFIG_METHOD_HW_DEVICE_CTX) ||
470 config->device_type != type) {
471 // Not interesting for us.
475 // We have a config of the right type, but does it actually support
476 // the pixel format we want? (Seemingly, FFmpeg's way of signaling errors
477 // is to just replace the pixel format with a software-decoded one,
479 found_config_of_right_type = true;
480 for (const AVPixelFormat *fmt_ptr = fmt; *fmt_ptr != -1; ++fmt_ptr) {
481 if (config->pix_fmt == *fmt_ptr) {
482 fprintf(stderr, "Initialized '%s' hardware decoding for codec '%s'.\n",
483 av_hwdevice_get_type_name(type), ctx->codec->name);
484 if (ctx->profile == FF_PROFILE_H264_BASELINE) {
485 fprintf(stderr, "WARNING: Stream claims to be H.264 Baseline, which is generally poorly supported in hardware decoders.\n");
486 fprintf(stderr, " Consider encoding it as Constrained Baseline, Main or High instead.\n");
487 fprintf(stderr, " Decoding might fail and fall back to software.\n");
489 return config->pix_fmt;
492 fprintf(stderr, "Decoder '%s' supports only these pixel formats:", ctx->codec->name);
493 unordered_set<AVPixelFormat> seen;
494 for (const AVPixelFormat *fmt_ptr = fmt; *fmt_ptr != -1; ++fmt_ptr) {
495 if (!seen.count(*fmt_ptr)) {
496 fprintf(stderr, " %s", av_get_pix_fmt_name(*fmt_ptr));
497 seen.insert(*fmt_ptr);
500 fprintf(stderr, " (wanted %s for hardware acceleration)\n", av_get_pix_fmt_name(config->pix_fmt));
504 if (!found_config_of_right_type) {
505 fprintf(stderr, "Decoder '%s' does not support device type '%s'.\n", ctx->codec->name, av_hwdevice_get_type_name(type));
508 // We found no VA-API formats, so take the first software format.
509 for (const AVPixelFormat *fmt_ptr = fmt; *fmt_ptr != -1; ++fmt_ptr) {
510 if ((av_pix_fmt_desc_get(*fmt_ptr)->flags & AV_PIX_FMT_FLAG_HWACCEL) == 0) {
511 fprintf(stderr, "Falling back to software format %s.\n", av_get_pix_fmt_name(*fmt_ptr));
516 // Fallback: Just return anything. (Should never really happen.)
520 bool FFmpegCapture::play_video(const string &pathname)
522 // Note: Call before open, not after; otherwise, there's a race.
523 // (There is now, too, but it tips the correct way. We could use fstat()
524 // if we had the file descriptor.)
525 timespec last_modified;
527 if (stat(pathname.c_str(), &buf) != 0) {
528 // Probably some sort of protocol, so can't stat.
529 last_modified.tv_sec = -1;
531 last_modified = buf.st_mtim;
533 last_colorspace = static_cast<AVColorSpace>(-1);
534 last_chroma_location = static_cast<AVChromaLocation>(-1);
536 AVFormatContextWithCloser format_ctx;
537 if (srt_sock == -1) {
538 // Regular file (or stream).
539 frame_timeout_started = steady_clock::now();
540 frame_timeout_valid = true;
541 format_ctx = avformat_open_input_unique(pathname.c_str(), /*fmt=*/nullptr,
543 AVIOInterruptCB{ &FFmpegCapture::interrupt_cb_thunk, this });
544 frame_timeout_valid = false;
547 // SRT socket, already opened.
548 const AVInputFormat *mpegts_fmt = av_find_input_format("mpegts");
549 format_ctx = avformat_open_input_unique(&FFmpegCapture::read_srt_thunk, this,
550 mpegts_fmt, /*options=*/nullptr,
551 AVIOInterruptCB{ &FFmpegCapture::interrupt_cb_thunk, this });
556 if (format_ctx == nullptr) {
557 fprintf(stderr, "%s: Error opening file\n", pathname.c_str());
561 if (avformat_find_stream_info(format_ctx.get(), nullptr) < 0) {
562 fprintf(stderr, "%s: Error finding stream info\n", pathname.c_str());
566 int video_stream_index = find_stream_index(format_ctx.get(), AVMEDIA_TYPE_VIDEO);
567 if (video_stream_index == -1) {
568 fprintf(stderr, "%s: No video stream found\n", pathname.c_str());
572 int audio_stream_index = find_stream_index(format_ctx.get(), AVMEDIA_TYPE_AUDIO);
573 int subtitle_stream_index = find_stream_index(format_ctx.get(), AVMEDIA_TYPE_SUBTITLE);
574 has_last_subtitle = false;
576 // Open video decoder.
577 const AVCodecParameters *video_codecpar = format_ctx->streams[video_stream_index]->codecpar;
578 const AVCodec *video_codec = avcodec_find_decoder(video_codecpar->codec_id);
580 video_timebase = format_ctx->streams[video_stream_index]->time_base;
581 AVCodecContextWithDeleter video_codec_ctx = avcodec_alloc_context3_unique(nullptr);
582 if (avcodec_parameters_to_context(video_codec_ctx.get(), video_codecpar) < 0) {
583 fprintf(stderr, "%s: Cannot fill video codec parameters\n", pathname.c_str());
586 if (video_codec == nullptr) {
587 fprintf(stderr, "%s: Cannot find video decoder\n", pathname.c_str());
591 // Seemingly, it's not too easy to make something that just initializes
592 // “whatever goes”, so we don't get CUDA or VULKAN or whatever here
593 // without enumerating through several different types.
594 // VA-API and VDPAU will do for now. We prioritize VDPAU for the
595 // simple reason that there's a VA-API-via-VDPAU emulation for NVidia
596 // cards that seems to work, but just hangs when trying to transfer the frame.
598 // Note that we don't actually check codec support beforehand,
599 // so if you have a low-end VDPAU device but a high-end VA-API device,
600 // you lose out on the extra codec support from the latter.
601 AVBufferRef *hw_device_ctx = nullptr;
602 if (av_hwdevice_ctx_create(&hw_device_ctx, AV_HWDEVICE_TYPE_VDPAU, nullptr, nullptr, 0) >= 0) {
603 video_codec_ctx->hw_device_ctx = av_buffer_ref(hw_device_ctx);
604 video_codec_ctx->get_format = get_hw_format<AV_HWDEVICE_TYPE_VDPAU>;
605 } else if (av_hwdevice_ctx_create(&hw_device_ctx, AV_HWDEVICE_TYPE_VAAPI, nullptr, nullptr, 0) >= 0) {
606 video_codec_ctx->hw_device_ctx = av_buffer_ref(hw_device_ctx);
607 video_codec_ctx->get_format = get_hw_format<AV_HWDEVICE_TYPE_VAAPI>;
609 fprintf(stderr, "Failed to initialize VA-API or VDPAU for FFmpeg acceleration. Decoding video in software.\n");
612 if (avcodec_open2(video_codec_ctx.get(), video_codec, nullptr) < 0) {
613 fprintf(stderr, "%s: Cannot open video decoder\n", pathname.c_str());
616 unique_ptr<AVCodecContext, decltype(avcodec_close)*> video_codec_ctx_cleanup(
617 video_codec_ctx.get(), avcodec_close);
619 // Used in decode_ycbcr_format().
620 is_mjpeg = video_codecpar->codec_id == AV_CODEC_ID_MJPEG;
622 // Open audio decoder, if we have audio.
623 AVCodecContextWithDeleter audio_codec_ctx;
624 if (audio_stream_index != -1) {
625 audio_codec_ctx = avcodec_alloc_context3_unique(nullptr);
626 const AVCodecParameters *audio_codecpar = format_ctx->streams[audio_stream_index]->codecpar;
627 audio_timebase = format_ctx->streams[audio_stream_index]->time_base;
628 if (avcodec_parameters_to_context(audio_codec_ctx.get(), audio_codecpar) < 0) {
629 fprintf(stderr, "%s: Cannot fill audio codec parameters\n", pathname.c_str());
632 const AVCodec *audio_codec = avcodec_find_decoder(audio_codecpar->codec_id);
633 if (audio_codec == nullptr) {
634 fprintf(stderr, "%s: Cannot find audio decoder\n", pathname.c_str());
637 if (avcodec_open2(audio_codec_ctx.get(), audio_codec, nullptr) < 0) {
638 fprintf(stderr, "%s: Cannot open audio decoder\n", pathname.c_str());
642 unique_ptr<AVCodecContext, decltype(avcodec_close)*> audio_codec_ctx_cleanup(
643 audio_codec_ctx.get(), avcodec_close);
648 bool first_frame = true;
649 int consecutive_errors = 0;
650 while (!producer_thread_should_quit.should_quit()) {
651 if (process_queued_commands(format_ctx.get(), pathname, last_modified, /*rewound=*/nullptr)) {
654 if (should_interrupt.load()) {
655 // Check as a failsafe, so that we don't need to rely on avio if we don't have to.
658 UniqueFrame audio_frame = audio_frame_allocator->alloc_frame();
659 AudioFormat audio_format;
663 frame_timeout_started = steady_clock::now();
664 frame_timeout_valid = true;
665 AVFrameWithDeleter frame = decode_frame(format_ctx.get(), video_codec_ctx.get(), audio_codec_ctx.get(),
666 pathname, video_stream_index, audio_stream_index, subtitle_stream_index, audio_frame.get(), &audio_format, &audio_pts, &error);
667 frame_timeout_valid = false;
668 if (should_interrupt.load()) {
669 // Abort no matter whether we got a frame or not.
673 if (++consecutive_errors >= 100) {
674 fprintf(stderr, "More than 100 consecutive error video frames, aborting playback.\n");
680 consecutive_errors = 0;
682 if (frame == nullptr) {
683 // EOF. Loop back to the start if we can.
684 if (format_ctx->pb != nullptr && format_ctx->pb->seekable == 0) {
685 // Not seekable (but seemingly, sometimes av_seek_frame() would return 0 anyway,
689 if (av_seek_frame(format_ctx.get(), /*stream_index=*/-1, /*timestamp=*/0, /*flags=*/0) < 0) {
690 fprintf(stderr, "%s: Rewind failed, not looping.\n", pathname.c_str());
693 if (video_codec_ctx != nullptr) {
694 avcodec_flush_buffers(video_codec_ctx.get());
696 if (audio_codec_ctx != nullptr) {
697 avcodec_flush_buffers(audio_codec_ctx.get());
699 // If the file has changed since last time, return to get it reloaded.
700 // Note that depending on how you move the file into place, you might
701 // end up corrupting the one you're already playing, so this path
702 // might not trigger.
703 if (changed_since(pathname, last_modified)) {
710 VideoFormat video_format = construct_video_format(frame.get(), video_timebase);
711 if (video_format.frame_rate_nom == 0 || video_format.frame_rate_den == 0) {
712 // Invalid frame rate; try constructing it from the previous frame length.
713 // (This is especially important if we are the master card, for SRT,
714 // since it affects audio. Not all senders have good timebases
715 // (e.g., Larix rounds first to timebase 1000 and then multiplies by
716 // 90 from there, it seems), but it's much better to have an oscillating
717 // value than just locking at 60.
718 if (last_pts != 0 && frame->pts > last_pts) {
719 int64_t pts_diff = frame->pts - last_pts;
720 video_format.frame_rate_nom = video_timebase.den;
721 video_format.frame_rate_den = video_timebase.num * pts_diff;
723 video_format.frame_rate_nom = 60;
724 video_format.frame_rate_den = 1;
727 UniqueFrame video_frame = make_video_frame(frame.get(), pathname, &error);
733 if (last_pts == 0 && pts_origin == 0) {
734 pts_origin = frame->pts;
736 steady_clock::time_point now = steady_clock::now();
737 if (play_as_fast_as_possible) {
738 video_frame->received_timestamp = now;
739 audio_frame->received_timestamp = now;
740 next_frame_start = now;
742 next_frame_start = compute_frame_start(frame->pts, pts_origin, video_timebase, start, rate);
743 if (first_frame && last_frame_was_connected) {
744 // If reconnect took more than one second, this is probably a live feed,
745 // and we should reset the resampler. (Or the rate is really, really low,
746 // in which case a reset on the first frame is fine anyway.)
747 if (duration<double>(next_frame_start - last_frame).count() >= 1.0) {
748 last_frame_was_connected = false;
751 video_frame->received_timestamp = next_frame_start;
753 // The easiest way to get all the rate conversions etc. right is to move the
754 // audio PTS into the video PTS timebase and go from there. (We'll get some
755 // rounding issues, but they should not be a big problem.)
756 int64_t audio_pts_as_video_pts = av_rescale_q(audio_pts, audio_timebase, video_timebase);
757 audio_frame->received_timestamp = compute_frame_start(audio_pts_as_video_pts, pts_origin, video_timebase, start, rate);
759 if (audio_frame->len != 0) {
760 // The received timestamps in Nageru are measured after we've just received the frame.
761 // However, pts (especially audio pts) is at the _beginning_ of the frame.
762 // If we have locked audio, the distinction doesn't really matter, as pts is
763 // on a relative scale and a fixed offset is fine. But if we don't, we will have
764 // a different number of samples each time, which will cause huge audio jitter
765 // and throw off the resampler.
767 // In a sense, we should have compensated by adding the frame and audio lengths
768 // to video_frame->received_timestamp and audio_frame->received_timestamp respectively,
769 // but that would mean extra waiting in sleep_until(). All we need is that they
770 // are correct relative to each other, though (and to the other frames we send),
771 // so just align the end of the audio frame, and we're fine.
772 size_t num_samples = (audio_frame->len * 8) / audio_format.bits_per_sample / audio_format.num_channels;
773 double offset = double(num_samples) / OUTPUT_FREQUENCY -
774 double(video_format.frame_rate_den) / video_format.frame_rate_nom;
775 audio_frame->received_timestamp += duration_cast<steady_clock::duration>(duration<double>(offset));
778 if (duration<double>(now - next_frame_start).count() >= 0.1) {
779 // If we don't have enough CPU to keep up, or if we have a live stream
780 // where the initial origin was somehow wrong, we could be behind indefinitely.
781 // In particular, this will give the audio resampler problems as it tries
782 // to speed up to reduce the delay, hitting the low end of the buffer every time.
783 fprintf(stderr, "%s: Playback %.0f ms behind, resetting time scale\n",
785 1e3 * duration<double>(now - next_frame_start).count());
786 pts_origin = frame->pts;
787 start = next_frame_start = now;
788 timecode += TYPICAL_FPS * 2 + 1;
791 bool finished_wakeup;
792 if (play_as_fast_as_possible) {
793 finished_wakeup = !producer_thread_should_quit.should_quit();
795 finished_wakeup = producer_thread_should_quit.sleep_until(next_frame_start);
797 if (finished_wakeup) {
798 if (audio_frame->len > 0) {
799 assert(audio_pts != -1);
801 if (!last_frame_was_connected) {
802 // We're recovering from an error (or really slow load, see above).
803 // Make sure to get the audio resampler reset. (This is a hack;
804 // ideally, the frame callback should just accept a way to signal
805 // audio discontinuity.)
806 timecode += TYPICAL_FPS * 2 + 1;
808 last_neutral_color = get_neutral_color(frame->metadata);
809 if (frame_callback != nullptr) {
810 frame_callback(frame->pts, video_timebase, audio_pts, audio_timebase, timecode++,
811 video_frame.get_and_release(), 0, video_format,
812 audio_frame.get_and_release(), 0, audio_format);
815 last_frame = steady_clock::now();
816 last_frame_was_connected = true;
819 if (producer_thread_should_quit.should_quit()) break;
821 bool rewound = false;
822 if (process_queued_commands(format_ctx.get(), pathname, last_modified, &rewound)) {
825 // If we just rewound, drop this frame on the floor and be done.
829 // OK, we didn't, so probably a rate change. Recalculate next_frame_start,
830 // but if it's now in the past, we'll reset the origin, so that we don't
831 // generate a huge backlog of frames that we need to run through quickly.
832 next_frame_start = compute_frame_start(frame->pts, pts_origin, video_timebase, start, rate);
833 steady_clock::time_point now = steady_clock::now();
834 if (next_frame_start < now) {
835 pts_origin = frame->pts;
836 start = next_frame_start = now;
840 last_pts = frame->pts;
845 void FFmpegCapture::internal_rewind()
847 pts_origin = last_pts = 0;
848 start = next_frame_start = steady_clock::now();
851 bool FFmpegCapture::process_queued_commands(AVFormatContext *format_ctx, const std::string &pathname, timespec last_modified, bool *rewound)
853 // Process any queued commands from other threads.
854 vector<QueuedCommand> commands;
856 lock_guard<mutex> lock(queue_mu);
857 swap(commands, command_queue);
859 for (const QueuedCommand &cmd : commands) {
860 switch (cmd.command) {
861 case QueuedCommand::REWIND:
862 if (av_seek_frame(format_ctx, /*stream_index=*/-1, /*timestamp=*/0, /*flags=*/0) < 0) {
863 fprintf(stderr, "%s: Rewind failed, stopping play.\n", pathname.c_str());
865 // If the file has changed since last time, return to get it reloaded.
866 // Note that depending on how you move the file into place, you might
867 // end up corrupting the one you're already playing, so this path
868 // might not trigger.
869 if (changed_since(pathname, last_modified)) {
873 if (rewound != nullptr) {
878 case QueuedCommand::CHANGE_RATE:
879 // Change the origin to the last played frame.
880 start = compute_frame_start(last_pts, pts_origin, video_timebase, start, rate);
881 pts_origin = last_pts;
883 play_as_fast_as_possible = (rate >= 10.0);
894 AVFrameWithDeleter FFmpegCapture::decode_frame(AVFormatContext *format_ctx, AVCodecContext *video_codec_ctx, AVCodecContext *audio_codec_ctx,
895 const std::string &pathname, int video_stream_index, int audio_stream_index, int subtitle_stream_index,
896 FrameAllocator::Frame *audio_frame, AudioFormat *audio_format, int64_t *audio_pts, bool *error)
900 // Read packets until we have a frame or there are none left.
901 bool frame_finished = false;
902 AVFrameWithDeleter audio_avframe = av_frame_alloc_unique();
903 AVFrameWithDeleter video_avframe = av_frame_alloc_unique();
906 bool has_audio = false;
908 AVPacketWithDeleter pkt = av_packet_alloc_unique();
911 if (av_read_frame(format_ctx, pkt.get()) == 0) {
912 if (pkt->stream_index == audio_stream_index && audio_callback != nullptr) {
913 audio_callback(pkt.get(), format_ctx->streams[audio_stream_index]->time_base);
915 if (pkt->stream_index == video_stream_index && video_callback != nullptr) {
916 video_callback(pkt.get(), format_ctx->streams[video_stream_index]->time_base);
918 if (pkt->stream_index == video_stream_index && global_flags.transcode_video) {
919 if (avcodec_send_packet(video_codec_ctx, pkt.get()) < 0) {
920 fprintf(stderr, "%s: Cannot send packet to video codec.\n", pathname.c_str());
922 return AVFrameWithDeleter(nullptr);
924 } else if (pkt->stream_index == audio_stream_index && global_flags.transcode_audio) {
926 if (avcodec_send_packet(audio_codec_ctx, pkt.get()) < 0) {
927 fprintf(stderr, "%s: Cannot send packet to audio codec.\n", pathname.c_str());
929 return AVFrameWithDeleter(nullptr);
931 } else if (pkt->stream_index == subtitle_stream_index) {
932 last_subtitle = string(reinterpret_cast<const char *>(pkt->data), pkt->size);
933 has_last_subtitle = true;
936 eof = true; // Or error, but ignore that for the time being.
939 // Decode audio, if any.
942 int err = avcodec_receive_frame(audio_codec_ctx, audio_avframe.get());
944 if (*audio_pts == -1) {
945 *audio_pts = audio_avframe->pts;
947 convert_audio(audio_avframe.get(), audio_frame, audio_format);
948 } else if (err == AVERROR(EAGAIN)) {
951 fprintf(stderr, "%s: Cannot receive frame from audio codec.\n", pathname.c_str());
953 return AVFrameWithDeleter(nullptr);
958 // Decode video, if we have a frame.
959 int err = avcodec_receive_frame(video_codec_ctx, video_avframe.get());
961 if (video_avframe->format == AV_PIX_FMT_VAAPI ||
962 video_avframe->format == AV_PIX_FMT_VDPAU) {
963 // Get the frame down to the CPU. (TODO: See if we can keep it
964 // on the GPU all the way, since it will be going up again later.
965 // However, this only works if the OpenGL GPU is the same one.)
966 AVFrameWithDeleter sw_frame = av_frame_alloc_unique();
967 int err = av_hwframe_transfer_data(sw_frame.get(), video_avframe.get(), 0);
969 fprintf(stderr, "%s: Cannot transfer hardware video frame to software.\n", pathname.c_str());
971 return AVFrameWithDeleter(nullptr);
973 sw_frame->pts = video_avframe->pts;
974 sw_frame->pkt_duration = video_avframe->pkt_duration;
975 video_avframe = move(sw_frame);
977 frame_finished = true;
979 } else if (err != AVERROR(EAGAIN)) {
980 fprintf(stderr, "%s: Cannot receive frame from video codec.\n", pathname.c_str());
982 return AVFrameWithDeleter(nullptr);
987 return video_avframe;
989 return AVFrameWithDeleter(nullptr);
992 void FFmpegCapture::convert_audio(const AVFrame *audio_avframe, FrameAllocator::Frame *audio_frame, AudioFormat *audio_format)
994 // Decide on a format. If there already is one in this audio frame,
995 // we're pretty much forced to use it. If not, we try to find an exact match.
996 // If that still doesn't work, we default to 32-bit signed chunked
997 // (float would be nice, but there's really no way to signal that yet).
998 AVSampleFormat dst_format;
999 if (audio_format->bits_per_sample == 0) {
1000 switch (audio_avframe->format) {
1001 case AV_SAMPLE_FMT_S16:
1002 case AV_SAMPLE_FMT_S16P:
1003 audio_format->bits_per_sample = 16;
1004 dst_format = AV_SAMPLE_FMT_S16;
1006 case AV_SAMPLE_FMT_S32:
1007 case AV_SAMPLE_FMT_S32P:
1009 audio_format->bits_per_sample = 32;
1010 dst_format = AV_SAMPLE_FMT_S32;
1013 } else if (audio_format->bits_per_sample == 16) {
1014 dst_format = AV_SAMPLE_FMT_S16;
1015 } else if (audio_format->bits_per_sample == 32) {
1016 dst_format = AV_SAMPLE_FMT_S32;
1020 audio_format->num_channels = 2;
1022 AVChannelLayout channel_layout = audio_avframe->ch_layout;
1023 if (!av_channel_layout_check(&channel_layout) ||
1024 channel_layout.order == AV_CHANNEL_ORDER_UNSPEC) {
1025 av_channel_layout_default(&channel_layout, audio_avframe->ch_layout.nb_channels);
1028 if (resampler == nullptr ||
1029 audio_avframe->format != last_src_format ||
1030 dst_format != last_dst_format ||
1031 av_channel_layout_compare(&channel_layout, &last_channel_layout) != 0||
1032 audio_avframe->sample_rate != last_sample_rate) {
1033 // TODO: When we get C++20, use AV_CHANNEL_LAYOUT_STEREO_DOWNMIX.
1034 AVChannelLayout stereo_downmix;
1035 stereo_downmix.order = AV_CHANNEL_ORDER_NATIVE;
1036 stereo_downmix.nb_channels = 2;
1037 stereo_downmix.u.mask = AV_CH_LAYOUT_STEREO_DOWNMIX;
1039 swr_free(&resampler);
1040 resampler = nullptr;
1041 int err = swr_alloc_set_opts2(&resampler,
1042 /*out_ch_layout=*/&stereo_downmix,
1043 /*out_sample_fmt=*/dst_format,
1044 /*out_sample_rate=*/OUTPUT_FREQUENCY,
1045 /*in_ch_layout=*/&channel_layout,
1046 /*in_sample_fmt=*/AVSampleFormat(audio_avframe->format),
1047 /*in_sample_rate=*/audio_avframe->sample_rate,
1049 /*log_ctx=*/nullptr);
1051 if (err != 0 || resampler == nullptr) {
1052 fprintf(stderr, "Allocating resampler failed.\n");
1056 if (swr_init(resampler) < 0) {
1057 fprintf(stderr, "Could not open resample context.\n");
1061 last_src_format = AVSampleFormat(audio_avframe->format);
1062 last_dst_format = dst_format;
1063 last_channel_layout = channel_layout;
1064 last_sample_rate = audio_avframe->sample_rate;
1067 size_t bytes_per_sample = (audio_format->bits_per_sample / 8) * 2;
1068 size_t num_samples_room = (audio_frame->size - audio_frame->len) / bytes_per_sample;
1070 uint8_t *data = audio_frame->data + audio_frame->len;
1071 int out_samples = swr_convert(resampler, &data, num_samples_room,
1072 const_cast<const uint8_t **>(audio_avframe->data), audio_avframe->nb_samples);
1073 if (out_samples < 0) {
1074 fprintf(stderr, "Audio conversion failed.\n");
1078 audio_frame->len += out_samples * bytes_per_sample;
1081 VideoFormat FFmpegCapture::construct_video_format(const AVFrame *frame, AVRational video_timebase)
1083 VideoFormat video_format;
1084 video_format.width = frame_width(frame);
1085 video_format.height = frame_height(frame);
1086 if (pixel_format == bmusb::PixelFormat_8BitBGRA) {
1087 video_format.stride = frame_width(frame) * 4;
1088 } else if (pixel_format == FFmpegCapture::PixelFormat_NV12) {
1089 video_format.stride = frame_width(frame);
1091 assert(pixel_format == bmusb::PixelFormat_8BitYCbCrPlanar);
1092 video_format.stride = frame_width(frame);
1094 video_format.frame_rate_nom = video_timebase.den;
1095 video_format.frame_rate_den = frame->pkt_duration * video_timebase.num;
1096 video_format.has_signal = true;
1097 video_format.is_connected = true;
1098 return video_format;
1101 UniqueFrame FFmpegCapture::make_video_frame(const AVFrame *frame, const string &pathname, bool *error)
1105 UniqueFrame video_frame(video_frame_allocator->alloc_frame());
1106 if (video_frame->data == nullptr) {
1110 if (sws_ctx == nullptr ||
1111 sws_last_width != frame->width ||
1112 sws_last_height != frame->height ||
1113 sws_last_src_format != frame->format) {
1114 sws_dst_format = decide_dst_format(AVPixelFormat(frame->format), pixel_format);
1116 sws_getContext(frame->width, frame->height, AVPixelFormat(frame->format),
1117 frame_width(frame), frame_height(frame), sws_dst_format,
1118 SWS_BICUBIC, nullptr, nullptr, nullptr));
1119 sws_last_width = frame->width;
1120 sws_last_height = frame->height;
1121 sws_last_src_format = frame->format;
1123 if (sws_ctx == nullptr) {
1124 fprintf(stderr, "%s: Could not create scaler context\n", pathname.c_str());
1129 uint8_t *pic_data[4] = { nullptr, nullptr, nullptr, nullptr };
1130 int linesizes[4] = { 0, 0, 0, 0 };
1131 if (pixel_format == bmusb::PixelFormat_8BitBGRA) {
1132 pic_data[0] = video_frame->data;
1133 linesizes[0] = frame_width(frame) * 4;
1134 video_frame->len = (frame_width(frame) * 4) * frame_height(frame);
1135 } else if (pixel_format == PixelFormat_NV12) {
1136 pic_data[0] = video_frame->data;
1137 linesizes[0] = frame_width(frame);
1139 pic_data[1] = pic_data[0] + frame_width(frame) * frame_height(frame);
1140 linesizes[1] = frame_width(frame);
1142 video_frame->len = (frame_width(frame) * 2) * frame_height(frame);
1144 const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(sws_dst_format);
1145 current_frame_ycbcr_format = decode_ycbcr_format(desc, frame, is_mjpeg, &last_colorspace, &last_chroma_location);
1147 assert(pixel_format == bmusb::PixelFormat_8BitYCbCrPlanar);
1148 const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(sws_dst_format);
1150 int chroma_width = AV_CEIL_RSHIFT(int(frame_width(frame)), desc->log2_chroma_w);
1151 int chroma_height = AV_CEIL_RSHIFT(int(frame_height(frame)), desc->log2_chroma_h);
1153 pic_data[0] = video_frame->data;
1154 linesizes[0] = frame_width(frame);
1156 pic_data[1] = pic_data[0] + frame_width(frame) * frame_height(frame);
1157 linesizes[1] = chroma_width;
1159 pic_data[2] = pic_data[1] + chroma_width * chroma_height;
1160 linesizes[2] = chroma_width;
1162 video_frame->len = frame_width(frame) * frame_height(frame) + 2 * chroma_width * chroma_height;
1164 current_frame_ycbcr_format = decode_ycbcr_format(desc, frame, is_mjpeg, &last_colorspace, &last_chroma_location);
1167 // FIXME: Currently, if the video is too high-res for one of the allocated
1168 // frames, we simply refuse to scale it here to avoid crashes. It would be better
1169 // if we could somehow signal getting larger frames, especially as 4K is a thing now.
1170 if (video_frame->len > FRAME_SIZE) {
1171 fprintf(stderr, "%s: Decoded frame would be larger than supported FRAME_SIZE (%zu > %u), not decoding.\n", pathname.c_str(), video_frame->len, FRAME_SIZE);
1176 sws_scale(sws_ctx.get(), frame->data, frame->linesize, 0, frame->height, pic_data, linesizes);
1181 int FFmpegCapture::interrupt_cb_thunk(void *opaque)
1183 return reinterpret_cast<FFmpegCapture *>(opaque)->interrupt_cb();
1186 int FFmpegCapture::interrupt_cb()
1188 // If ten seconds is gone without anything happening, we assume that
1189 // we are in a network stream that died and FFmpeg just didn't
1190 // pick it up (or perhaps it just hung, keeping the connection open).
1191 // Called back approximately every 100 ms if something is hanging,
1192 // so we get more than enough accuracy for our purposes.
1193 if (!should_interrupt && frame_timeout_valid &&
1194 duration<double>(steady_clock::now() - frame_timeout_started).count() >= 10.0) {
1195 string filename_copy;
1197 lock_guard<mutex> lock(filename_mu);
1198 filename_copy = filename;
1200 fprintf(stderr, "%s: No frame for more than 10 seconds, restarting stream.\n", filename.c_str());
1201 should_interrupt = true;
1203 return should_interrupt.load();
1206 unsigned FFmpegCapture::frame_width(const AVFrame *frame) const
1209 return frame->width;
1215 unsigned FFmpegCapture::frame_height(const AVFrame *frame) const
1218 return frame->height;
1225 int FFmpegCapture::read_srt_thunk(void *opaque, uint8_t *buf, int buf_size)
1227 return reinterpret_cast<FFmpegCapture *>(opaque)->read_srt(buf, buf_size);
1230 int FFmpegCapture::read_srt(uint8_t *buf, int buf_size)
1232 SRT_MSGCTRL mc = srt_msgctrl_default;
1233 return srt_recvmsg2(srt_sock, reinterpret_cast<char *>(buf), buf_size, &mc);