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>
32 #include <movit/colorspace_conversion_effect.h>
34 #include "bmusb/bmusb.h"
35 #include "shared/ffmpeg_raii.h"
36 #include "ffmpeg_util.h"
38 #include "image_input.h"
39 #include "ref_counted_frame.h"
40 #include "shared/timebase.h"
46 #define FRAME_SIZE (8 << 20) // 8 MB.
49 using namespace std::chrono;
50 using namespace bmusb;
51 using namespace movit;
52 using namespace Eigen;
56 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)
58 const duration<double> pts((frame_pts - pts_origin) * double(video_timebase.num) / double(video_timebase.den));
59 return origin + duration_cast<steady_clock::duration>(pts / rate);
62 bool changed_since(const std::string &pathname, const timespec &ts)
68 if (stat(pathname.c_str(), &buf) != 0) {
69 fprintf(stderr, "%s: Couldn't check for new version, leaving the old in place.\n", pathname.c_str());
72 return (buf.st_mtim.tv_sec != ts.tv_sec || buf.st_mtim.tv_nsec != ts.tv_nsec);
75 bool is_full_range(const AVPixFmtDescriptor *desc)
77 // This is horrible, but there's no better way that I know of.
78 return (strchr(desc->name, 'j') != nullptr);
81 AVPixelFormat decide_dst_format(AVPixelFormat src_format, bmusb::PixelFormat dst_format_type)
83 if (dst_format_type == bmusb::PixelFormat_8BitBGRA) {
84 return AV_PIX_FMT_BGRA;
86 if (dst_format_type == FFmpegCapture::PixelFormat_NV12) {
87 return AV_PIX_FMT_NV12;
90 assert(dst_format_type == bmusb::PixelFormat_8BitYCbCrPlanar);
92 // If this is a non-Y'CbCr format, just convert to 4:4:4 Y'CbCr
93 // and be done with it. It's too strange to spend a lot of time on.
94 // (Let's hope there's no alpha.)
95 const AVPixFmtDescriptor *src_desc = av_pix_fmt_desc_get(src_format);
96 if (src_desc == nullptr ||
97 src_desc->nb_components != 3 ||
98 (src_desc->flags & AV_PIX_FMT_FLAG_RGB)) {
99 return AV_PIX_FMT_YUV444P;
102 // The best for us would be Cb and Cr together if possible,
103 // but FFmpeg doesn't support that except in the special case of
104 // NV12, so we need to go to planar even for the case of NV12.
105 // Thus, look for the closest (but no worse) 8-bit planar Y'CbCr format
106 // that matches in color range. (This will also include the case of
107 // the source format already being acceptable.)
108 bool src_full_range = is_full_range(src_desc);
109 const char *best_format = "yuv444p";
110 unsigned best_score = numeric_limits<unsigned>::max();
111 for (const AVPixFmtDescriptor *desc = av_pix_fmt_desc_next(nullptr);
113 desc = av_pix_fmt_desc_next(desc)) {
114 // Find planar Y'CbCr formats only.
115 if (desc->nb_components != 3) continue;
116 if (desc->flags & AV_PIX_FMT_FLAG_RGB) continue;
117 if (!(desc->flags & AV_PIX_FMT_FLAG_PLANAR)) continue;
118 if (desc->comp[0].plane != 0 ||
119 desc->comp[1].plane != 1 ||
120 desc->comp[2].plane != 2) continue;
122 // 8-bit formats only.
123 if (desc->flags & AV_PIX_FMT_FLAG_BE) continue;
124 if (desc->comp[0].depth != 8) continue;
126 // Same or better chroma resolution only.
127 int chroma_w_diff = desc->log2_chroma_w - src_desc->log2_chroma_w;
128 int chroma_h_diff = desc->log2_chroma_h - src_desc->log2_chroma_h;
129 if (chroma_w_diff < 0 || chroma_h_diff < 0)
132 // Matching full/limited range only.
133 if (is_full_range(desc) != src_full_range)
136 // Pick something with as little excess chroma resolution as possible.
137 unsigned score = (1 << (chroma_w_diff)) << chroma_h_diff;
138 if (score < best_score) {
140 best_format = desc->name;
143 return av_get_pix_fmt(best_format);
146 YCbCrFormat decode_ycbcr_format(const AVPixFmtDescriptor *desc, const AVFrame *frame, bool is_mjpeg, AVColorSpace *last_colorspace, AVChromaLocation *last_chroma_location)
149 AVColorSpace colorspace = frame->colorspace;
150 switch (colorspace) {
151 case AVCOL_SPC_BT709:
152 format.luma_coefficients = YCBCR_REC_709;
154 case AVCOL_SPC_BT470BG:
155 case AVCOL_SPC_SMPTE170M:
156 case AVCOL_SPC_SMPTE240M:
157 format.luma_coefficients = YCBCR_REC_601;
159 case AVCOL_SPC_BT2020_NCL:
160 format.luma_coefficients = YCBCR_REC_2020;
162 case AVCOL_SPC_UNSPECIFIED:
163 format.luma_coefficients = (frame->height >= 720 ? YCBCR_REC_709 : YCBCR_REC_601);
166 if (colorspace != *last_colorspace) {
167 fprintf(stderr, "Unknown Y'CbCr coefficient enum %d from FFmpeg; choosing Rec. 709.\n",
170 format.luma_coefficients = YCBCR_REC_709;
173 *last_colorspace = colorspace;
175 format.full_range = is_full_range(desc);
176 format.num_levels = 1 << desc->comp[0].depth;
177 format.chroma_subsampling_x = 1 << desc->log2_chroma_w;
178 format.chroma_subsampling_y = 1 << desc->log2_chroma_h;
180 switch (frame->chroma_location) {
181 case AVCHROMA_LOC_LEFT:
182 format.cb_x_position = 0.0;
183 format.cb_y_position = 0.5;
185 case AVCHROMA_LOC_CENTER:
186 format.cb_x_position = 0.5;
187 format.cb_y_position = 0.5;
189 case AVCHROMA_LOC_TOPLEFT:
190 format.cb_x_position = 0.0;
191 format.cb_y_position = 0.0;
193 case AVCHROMA_LOC_TOP:
194 format.cb_x_position = 0.5;
195 format.cb_y_position = 0.0;
197 case AVCHROMA_LOC_BOTTOMLEFT:
198 format.cb_x_position = 0.0;
199 format.cb_y_position = 1.0;
201 case AVCHROMA_LOC_BOTTOM:
202 format.cb_x_position = 0.5;
203 format.cb_y_position = 1.0;
206 if (frame->chroma_location != *last_chroma_location) {
207 fprintf(stderr, "Unknown chroma location coefficient enum %d from FFmpeg; choosing center.\n",
208 frame->chroma_location);
210 format.cb_x_position = 0.5;
211 format.cb_y_position = 0.5;
214 *last_chroma_location = frame->chroma_location;
216 if (is_mjpeg && !format.full_range) {
217 // Limited-range MJPEG is only detected by FFmpeg whenever a special
218 // JPEG comment is set, which means that in practice, the stream is
219 // almost certainly generated by Futatabi. Override FFmpeg's forced
220 // MJPEG defaults (it disregards the values set in the mux) with what
222 format.luma_coefficients = YCBCR_REC_709;
223 format.cb_x_position = 0.0;
224 format.cb_y_position = 0.5;
227 format.cr_x_position = format.cb_x_position;
228 format.cr_y_position = format.cb_y_position;
232 RGBTriplet get_neutral_color(AVDictionary *metadata)
234 if (metadata == nullptr) {
235 return RGBTriplet(1.0f, 1.0f, 1.0f);
237 AVDictionaryEntry *entry = av_dict_get(metadata, "WhitePoint", nullptr, 0);
238 if (entry == nullptr) {
239 return RGBTriplet(1.0f, 1.0f, 1.0f);
242 unsigned x_nom, x_den, y_nom, y_den;
243 if (sscanf(entry->value, " %u:%u , %u:%u", &x_nom, &x_den, &y_nom, &y_den) != 4) {
244 fprintf(stderr, "WARNING: Unable to parse white point '%s', using default white point\n", entry->value);
245 return RGBTriplet(1.0f, 1.0f, 1.0f);
248 double x = double(x_nom) / x_den;
249 double y = double(y_nom) / y_den;
250 double z = 1.0 - x - y;
252 Matrix3d rgb_to_xyz_matrix = movit::ColorspaceConversionEffect::get_xyz_matrix(COLORSPACE_sRGB);
253 Vector3d rgb = rgb_to_xyz_matrix.inverse() * Vector3d(x, y, z);
255 return RGBTriplet(rgb[0], rgb[1], rgb[2]);
260 FFmpegCapture::FFmpegCapture(const string &filename, unsigned width, unsigned height)
261 : filename(filename), width(width), height(height), video_timebase{1, 1}
263 description = "Video: " + filename;
265 last_frame = steady_clock::now();
267 avformat_network_init(); // In case someone wants this.
271 FFmpegCapture::FFmpegCapture(int srt_sock, const string &stream_id)
272 : srt_sock(srt_sock),
273 width(0), // Don't resize; SRT streams typically have stable resolution, and should behave much like regular cards in general.
275 pixel_format(bmusb::PixelFormat_8BitYCbCrPlanar),
278 if (stream_id.empty()) {
279 description = "SRT stream";
281 description = stream_id;
283 play_as_fast_as_possible = true;
285 last_frame = steady_clock::now();
289 FFmpegCapture::~FFmpegCapture()
291 if (has_dequeue_callbacks) {
292 dequeue_cleanup_callback();
294 swr_free(&resampler);
296 if (srt_sock != -1) {
302 void FFmpegCapture::configure_card()
304 if (video_frame_allocator == nullptr) {
305 owned_video_frame_allocator.reset(new MallocFrameAllocator(FRAME_SIZE, NUM_QUEUED_VIDEO_FRAMES));
306 set_video_frame_allocator(owned_video_frame_allocator.get());
308 if (audio_frame_allocator == nullptr) {
309 // Audio can come out in pretty large chunks, so increase from the default 1 MB.
310 owned_audio_frame_allocator.reset(new MallocFrameAllocator(1 << 20, NUM_QUEUED_AUDIO_FRAMES));
311 set_audio_frame_allocator(owned_audio_frame_allocator.get());
315 void FFmpegCapture::start_bm_capture()
321 producer_thread_should_quit.unquit();
322 producer_thread = thread(&FFmpegCapture::producer_thread_func, this);
325 void FFmpegCapture::stop_dequeue_thread()
331 producer_thread_should_quit.quit();
332 producer_thread.join();
335 std::map<uint32_t, VideoMode> FFmpegCapture::get_available_video_modes() const
337 // Note: This will never really be shown in the UI.
341 snprintf(buf, sizeof(buf), "%ux%u", sws_last_width, sws_last_height);
344 mode.autodetect = false;
345 mode.width = sws_last_width;
346 mode.height = sws_last_height;
347 mode.frame_rate_num = 60;
348 mode.frame_rate_den = 1;
349 mode.interlaced = false;
351 return {{ 0, mode }};
354 void FFmpegCapture::producer_thread_func()
356 char thread_name[16];
357 snprintf(thread_name, sizeof(thread_name), "FFmpeg_C_%d", card_index);
358 pthread_setname_np(pthread_self(), thread_name);
360 while (!producer_thread_should_quit.should_quit()) {
361 string filename_copy;
363 lock_guard<mutex> lock(filename_mu);
364 filename_copy = filename;
368 if (srt_sock == -1) {
369 pathname = search_for_file(filename_copy);
371 pathname = description;
373 if (pathname.empty()) {
374 send_disconnected_frame();
378 producer_thread_should_quit.sleep_for(seconds(1));
379 fprintf(stderr, "%s not found, sleeping one second and trying again...\n", filename_copy.c_str());
382 should_interrupt = false;
383 if (!play_video(pathname)) {
385 send_disconnected_frame();
389 fprintf(stderr, "Error when playing %s, sleeping one second and trying again...\n", pathname.c_str());
390 producer_thread_should_quit.sleep_for(seconds(1));
395 send_disconnected_frame();
399 // Probably just EOF, will exit the loop above on next test.
402 if (has_dequeue_callbacks) {
403 dequeue_cleanup_callback();
404 has_dequeue_callbacks = false;
408 void FFmpegCapture::send_disconnected_frame()
410 // Send an empty frame to signal that we have no signal anymore.
411 FrameAllocator::Frame video_frame = video_frame_allocator->alloc_frame();
412 size_t frame_width = width == 0 ? global_flags.width : width;
413 size_t frame_height = height == 0 ? global_flags.height : height;
414 if (video_frame.data) {
415 VideoFormat video_format;
416 video_format.width = frame_width;
417 video_format.height = frame_height;
418 video_format.frame_rate_nom = 60;
419 video_format.frame_rate_den = 1;
420 video_format.is_connected = false;
421 if (pixel_format == bmusb::PixelFormat_8BitBGRA) {
422 video_format.stride = frame_width * 4;
423 video_frame.len = frame_width * frame_height * 4;
424 memset(video_frame.data, 0, video_frame.len);
426 video_format.stride = frame_width;
427 current_frame_ycbcr_format.luma_coefficients = YCBCR_REC_709;
428 current_frame_ycbcr_format.full_range = true;
429 current_frame_ycbcr_format.num_levels = 256;
430 current_frame_ycbcr_format.chroma_subsampling_x = 2;
431 current_frame_ycbcr_format.chroma_subsampling_y = 2;
432 current_frame_ycbcr_format.cb_x_position = 0.0f;
433 current_frame_ycbcr_format.cb_y_position = 0.0f;
434 current_frame_ycbcr_format.cr_x_position = 0.0f;
435 current_frame_ycbcr_format.cr_y_position = 0.0f;
436 video_frame.len = frame_width * frame_height * 2;
437 memset(video_frame.data, 0, frame_width * frame_height);
438 memset(video_frame.data + frame_width * frame_height, 128, frame_width * frame_height); // Valid for both NV12 and planar.
441 if (frame_callback != nullptr) {
442 frame_callback(-1, AVRational{1, TIMEBASE}, -1, AVRational{1, TIMEBASE}, timecode++,
443 video_frame, /*video_offset=*/0, video_format,
444 FrameAllocator::Frame(), /*audio_offset=*/0, AudioFormat());
446 last_frame_was_connected = false;
451 if (card_disconnected_callback != nullptr) {
452 card_disconnected_callback();
457 AVPixelFormat get_vaapi_hw_format(AVCodecContext *ctx, const AVPixelFormat *fmt)
459 for (const AVPixelFormat *fmt_ptr = fmt; *fmt_ptr != -1; ++fmt_ptr) {
460 for (int i = 0;; ++i) { // Termination condition inside loop.
461 const AVCodecHWConfig *config = avcodec_get_hw_config(ctx->codec, i);
462 if (config == nullptr) { // End of list.
463 fprintf(stderr, "Decoder %s does not support device.\n", ctx->codec->name);
466 if (config->methods & AV_CODEC_HW_CONFIG_METHOD_HW_DEVICE_CTX &&
467 config->device_type == AV_HWDEVICE_TYPE_VAAPI &&
468 config->pix_fmt == *fmt_ptr) {
469 return config->pix_fmt;
474 // We found no VA-API formats, so take the best software format.
478 bool FFmpegCapture::play_video(const string &pathname)
480 // Note: Call before open, not after; otherwise, there's a race.
481 // (There is now, too, but it tips the correct way. We could use fstat()
482 // if we had the file descriptor.)
483 timespec last_modified;
485 if (stat(pathname.c_str(), &buf) != 0) {
486 // Probably some sort of protocol, so can't stat.
487 last_modified.tv_sec = -1;
489 last_modified = buf.st_mtim;
491 last_colorspace = static_cast<AVColorSpace>(-1);
492 last_chroma_location = static_cast<AVChromaLocation>(-1);
494 AVFormatContextWithCloser format_ctx;
495 if (srt_sock == -1) {
497 format_ctx = avformat_open_input_unique(pathname.c_str(), /*fmt=*/nullptr,
499 AVIOInterruptCB{ &FFmpegCapture::interrupt_cb_thunk, this });
502 // SRT socket, already opened.
503 const AVInputFormat *mpegts_fmt = av_find_input_format("mpegts");
504 format_ctx = avformat_open_input_unique(&FFmpegCapture::read_srt_thunk, this,
505 mpegts_fmt, /*options=*/nullptr,
506 AVIOInterruptCB{ &FFmpegCapture::interrupt_cb_thunk, this });
511 if (format_ctx == nullptr) {
512 fprintf(stderr, "%s: Error opening file\n", pathname.c_str());
516 if (avformat_find_stream_info(format_ctx.get(), nullptr) < 0) {
517 fprintf(stderr, "%s: Error finding stream info\n", pathname.c_str());
521 int video_stream_index = find_stream_index(format_ctx.get(), AVMEDIA_TYPE_VIDEO);
522 if (video_stream_index == -1) {
523 fprintf(stderr, "%s: No video stream found\n", pathname.c_str());
527 int audio_stream_index = find_stream_index(format_ctx.get(), AVMEDIA_TYPE_AUDIO);
528 int subtitle_stream_index = find_stream_index(format_ctx.get(), AVMEDIA_TYPE_SUBTITLE);
529 has_last_subtitle = false;
531 // Open video decoder.
532 const AVCodecParameters *video_codecpar = format_ctx->streams[video_stream_index]->codecpar;
533 const AVCodec *video_codec = avcodec_find_decoder(video_codecpar->codec_id);
535 video_timebase = format_ctx->streams[video_stream_index]->time_base;
536 AVCodecContextWithDeleter video_codec_ctx = avcodec_alloc_context3_unique(nullptr);
537 if (avcodec_parameters_to_context(video_codec_ctx.get(), video_codecpar) < 0) {
538 fprintf(stderr, "%s: Cannot fill video codec parameters\n", pathname.c_str());
541 if (video_codec == nullptr) {
542 fprintf(stderr, "%s: Cannot find video decoder\n", pathname.c_str());
546 // Seemingly, it's not too easy to make something that just initializes
547 // “whatever goes”, so we don't get VDPAU or CUDA here without enumerating
548 // through several different types. VA-API will do for now.
549 AVBufferRef *hw_device_ctx = nullptr;
550 if (av_hwdevice_ctx_create(&hw_device_ctx, AV_HWDEVICE_TYPE_VAAPI, nullptr, nullptr, 0) < 0) {
551 fprintf(stderr, "Failed to initialize VA-API for FFmpeg acceleration. Decoding video in software.\n");
553 video_codec_ctx->hw_device_ctx = av_buffer_ref(hw_device_ctx);
554 video_codec_ctx->get_format = get_vaapi_hw_format;
557 if (avcodec_open2(video_codec_ctx.get(), video_codec, nullptr) < 0) {
558 fprintf(stderr, "%s: Cannot open video decoder\n", pathname.c_str());
561 unique_ptr<AVCodecContext, decltype(avcodec_close)*> video_codec_ctx_cleanup(
562 video_codec_ctx.get(), avcodec_close);
564 // Used in decode_ycbcr_format().
565 is_mjpeg = video_codecpar->codec_id == AV_CODEC_ID_MJPEG;
567 // Open audio decoder, if we have audio.
568 AVCodecContextWithDeleter audio_codec_ctx;
569 if (audio_stream_index != -1) {
570 audio_codec_ctx = avcodec_alloc_context3_unique(nullptr);
571 const AVCodecParameters *audio_codecpar = format_ctx->streams[audio_stream_index]->codecpar;
572 audio_timebase = format_ctx->streams[audio_stream_index]->time_base;
573 if (avcodec_parameters_to_context(audio_codec_ctx.get(), audio_codecpar) < 0) {
574 fprintf(stderr, "%s: Cannot fill audio codec parameters\n", pathname.c_str());
577 const AVCodec *audio_codec = avcodec_find_decoder(audio_codecpar->codec_id);
578 if (audio_codec == nullptr) {
579 fprintf(stderr, "%s: Cannot find audio decoder\n", pathname.c_str());
582 if (avcodec_open2(audio_codec_ctx.get(), audio_codec, nullptr) < 0) {
583 fprintf(stderr, "%s: Cannot open audio decoder\n", pathname.c_str());
587 unique_ptr<AVCodecContext, decltype(avcodec_close)*> audio_codec_ctx_cleanup(
588 audio_codec_ctx.get(), avcodec_close);
593 bool first_frame = true;
594 while (!producer_thread_should_quit.should_quit()) {
595 if (process_queued_commands(format_ctx.get(), pathname, last_modified, /*rewound=*/nullptr)) {
598 if (should_interrupt.load()) {
599 // Check as a failsafe, so that we don't need to rely on avio if we don't have to.
602 UniqueFrame audio_frame = audio_frame_allocator->alloc_frame();
603 AudioFormat audio_format;
607 AVFrameWithDeleter frame = decode_frame(format_ctx.get(), video_codec_ctx.get(), audio_codec_ctx.get(),
608 pathname, video_stream_index, audio_stream_index, subtitle_stream_index, audio_frame.get(), &audio_format, &audio_pts, &error);
612 if (frame == nullptr) {
613 // EOF. Loop back to the start if we can.
614 if (format_ctx->pb != nullptr && format_ctx->pb->seekable == 0) {
615 // Not seekable (but seemingly, sometimes av_seek_frame() would return 0 anyway,
619 if (av_seek_frame(format_ctx.get(), /*stream_index=*/-1, /*timestamp=*/0, /*flags=*/0) < 0) {
620 fprintf(stderr, "%s: Rewind failed, not looping.\n", pathname.c_str());
623 if (video_codec_ctx != nullptr) {
624 avcodec_flush_buffers(video_codec_ctx.get());
626 if (audio_codec_ctx != nullptr) {
627 avcodec_flush_buffers(audio_codec_ctx.get());
629 // If the file has changed since last time, return to get it reloaded.
630 // Note that depending on how you move the file into place, you might
631 // end up corrupting the one you're already playing, so this path
632 // might not trigger.
633 if (changed_since(pathname, last_modified)) {
640 VideoFormat video_format = construct_video_format(frame.get(), video_timebase);
641 if (video_format.frame_rate_nom == 0 || video_format.frame_rate_den == 0) {
642 // Invalid frame rate; try constructing it from the previous frame length.
643 // (This is especially important if we are the master card, for SRT,
644 // since it affects audio. Not all senders have good timebases
645 // (e.g., Larix rounds first to timebase 1000 and then multiplies by
646 // 90 from there, it seems), but it's much better to have an oscillating
647 // value than just locking at 60.
648 if (last_pts != 0 && frame->pts > last_pts) {
649 int64_t pts_diff = frame->pts - last_pts;
650 video_format.frame_rate_nom = video_timebase.den;
651 video_format.frame_rate_den = video_timebase.num * pts_diff;
653 video_format.frame_rate_nom = 60;
654 video_format.frame_rate_den = 1;
657 UniqueFrame video_frame = make_video_frame(frame.get(), pathname, &error);
663 if (last_pts == 0 && pts_origin == 0) {
664 pts_origin = frame->pts;
666 steady_clock::time_point now = steady_clock::now();
667 if (play_as_fast_as_possible) {
668 video_frame->received_timestamp = now;
669 audio_frame->received_timestamp = now;
670 next_frame_start = now;
672 next_frame_start = compute_frame_start(frame->pts, pts_origin, video_timebase, start, rate);
673 if (first_frame && last_frame_was_connected) {
674 // If reconnect took more than one second, this is probably a live feed,
675 // and we should reset the resampler. (Or the rate is really, really low,
676 // in which case a reset on the first frame is fine anyway.)
677 if (duration<double>(next_frame_start - last_frame).count() >= 1.0) {
678 last_frame_was_connected = false;
681 video_frame->received_timestamp = next_frame_start;
683 // The easiest way to get all the rate conversions etc. right is to move the
684 // audio PTS into the video PTS timebase and go from there. (We'll get some
685 // rounding issues, but they should not be a big problem.)
686 int64_t audio_pts_as_video_pts = av_rescale_q(audio_pts, audio_timebase, video_timebase);
687 audio_frame->received_timestamp = compute_frame_start(audio_pts_as_video_pts, pts_origin, video_timebase, start, rate);
689 if (audio_frame->len != 0) {
690 // The received timestamps in Nageru are measured after we've just received the frame.
691 // However, pts (especially audio pts) is at the _beginning_ of the frame.
692 // If we have locked audio, the distinction doesn't really matter, as pts is
693 // on a relative scale and a fixed offset is fine. But if we don't, we will have
694 // a different number of samples each time, which will cause huge audio jitter
695 // and throw off the resampler.
697 // In a sense, we should have compensated by adding the frame and audio lengths
698 // to video_frame->received_timestamp and audio_frame->received_timestamp respectively,
699 // but that would mean extra waiting in sleep_until(). All we need is that they
700 // are correct relative to each other, though (and to the other frames we send),
701 // so just align the end of the audio frame, and we're fine.
702 size_t num_samples = (audio_frame->len * 8) / audio_format.bits_per_sample / audio_format.num_channels;
703 double offset = double(num_samples) / OUTPUT_FREQUENCY -
704 double(video_format.frame_rate_den) / video_format.frame_rate_nom;
705 audio_frame->received_timestamp += duration_cast<steady_clock::duration>(duration<double>(offset));
708 if (duration<double>(now - next_frame_start).count() >= 0.1) {
709 // If we don't have enough CPU to keep up, or if we have a live stream
710 // where the initial origin was somehow wrong, we could be behind indefinitely.
711 // In particular, this will give the audio resampler problems as it tries
712 // to speed up to reduce the delay, hitting the low end of the buffer every time.
713 fprintf(stderr, "%s: Playback %.0f ms behind, resetting time scale\n",
715 1e3 * duration<double>(now - next_frame_start).count());
716 pts_origin = frame->pts;
717 start = next_frame_start = now;
718 timecode += MAX_FPS * 2 + 1;
721 bool finished_wakeup;
722 if (play_as_fast_as_possible) {
723 finished_wakeup = !producer_thread_should_quit.should_quit();
725 finished_wakeup = producer_thread_should_quit.sleep_until(next_frame_start);
727 if (finished_wakeup) {
728 if (audio_frame->len > 0) {
729 assert(audio_pts != -1);
731 if (!last_frame_was_connected) {
732 // We're recovering from an error (or really slow load, see above).
733 // Make sure to get the audio resampler reset. (This is a hack;
734 // ideally, the frame callback should just accept a way to signal
735 // audio discontinuity.)
736 timecode += MAX_FPS * 2 + 1;
738 last_neutral_color = get_neutral_color(frame->metadata);
739 if (frame_callback != nullptr) {
740 frame_callback(frame->pts, video_timebase, audio_pts, audio_timebase, timecode++,
741 video_frame.get_and_release(), 0, video_format,
742 audio_frame.get_and_release(), 0, audio_format);
745 last_frame = steady_clock::now();
746 last_frame_was_connected = true;
749 if (producer_thread_should_quit.should_quit()) break;
751 bool rewound = false;
752 if (process_queued_commands(format_ctx.get(), pathname, last_modified, &rewound)) {
755 // If we just rewound, drop this frame on the floor and be done.
759 // OK, we didn't, so probably a rate change. Recalculate next_frame_start,
760 // but if it's now in the past, we'll reset the origin, so that we don't
761 // generate a huge backlog of frames that we need to run through quickly.
762 next_frame_start = compute_frame_start(frame->pts, pts_origin, video_timebase, start, rate);
763 steady_clock::time_point now = steady_clock::now();
764 if (next_frame_start < now) {
765 pts_origin = frame->pts;
766 start = next_frame_start = now;
770 last_pts = frame->pts;
775 void FFmpegCapture::internal_rewind()
777 pts_origin = last_pts = 0;
778 start = next_frame_start = steady_clock::now();
781 bool FFmpegCapture::process_queued_commands(AVFormatContext *format_ctx, const std::string &pathname, timespec last_modified, bool *rewound)
783 // Process any queued commands from other threads.
784 vector<QueuedCommand> commands;
786 lock_guard<mutex> lock(queue_mu);
787 swap(commands, command_queue);
789 for (const QueuedCommand &cmd : commands) {
790 switch (cmd.command) {
791 case QueuedCommand::REWIND:
792 if (av_seek_frame(format_ctx, /*stream_index=*/-1, /*timestamp=*/0, /*flags=*/0) < 0) {
793 fprintf(stderr, "%s: Rewind failed, stopping play.\n", pathname.c_str());
795 // If the file has changed since last time, return to get it reloaded.
796 // Note that depending on how you move the file into place, you might
797 // end up corrupting the one you're already playing, so this path
798 // might not trigger.
799 if (changed_since(pathname, last_modified)) {
803 if (rewound != nullptr) {
808 case QueuedCommand::CHANGE_RATE:
809 // Change the origin to the last played frame.
810 start = compute_frame_start(last_pts, pts_origin, video_timebase, start, rate);
811 pts_origin = last_pts;
813 play_as_fast_as_possible = (rate >= 10.0);
824 AVFrameWithDeleter FFmpegCapture::decode_frame(AVFormatContext *format_ctx, AVCodecContext *video_codec_ctx, AVCodecContext *audio_codec_ctx,
825 const std::string &pathname, int video_stream_index, int audio_stream_index, int subtitle_stream_index,
826 FrameAllocator::Frame *audio_frame, AudioFormat *audio_format, int64_t *audio_pts, bool *error)
830 // Read packets until we have a frame or there are none left.
831 bool frame_finished = false;
832 AVFrameWithDeleter audio_avframe = av_frame_alloc_unique();
833 AVFrameWithDeleter video_avframe = av_frame_alloc_unique();
836 bool has_audio = false;
839 unique_ptr<AVPacket, decltype(av_packet_unref)*> pkt_cleanup(
840 &pkt, av_packet_unref);
841 av_init_packet(&pkt);
844 if (av_read_frame(format_ctx, &pkt) == 0) {
845 if (pkt.stream_index == audio_stream_index && audio_callback != nullptr) {
846 audio_callback(&pkt, format_ctx->streams[audio_stream_index]->time_base);
848 if (pkt.stream_index == video_stream_index && video_callback != nullptr) {
849 video_callback(&pkt, format_ctx->streams[video_stream_index]->time_base);
851 if (pkt.stream_index == video_stream_index && global_flags.transcode_video) {
852 if (avcodec_send_packet(video_codec_ctx, &pkt) < 0) {
853 fprintf(stderr, "%s: Cannot send packet to video codec.\n", pathname.c_str());
855 return AVFrameWithDeleter(nullptr);
857 } else if (pkt.stream_index == audio_stream_index && global_flags.transcode_audio) {
859 if (avcodec_send_packet(audio_codec_ctx, &pkt) < 0) {
860 fprintf(stderr, "%s: Cannot send packet to audio codec.\n", pathname.c_str());
862 return AVFrameWithDeleter(nullptr);
864 } else if (pkt.stream_index == subtitle_stream_index) {
865 last_subtitle = string(reinterpret_cast<const char *>(pkt.data), pkt.size);
866 has_last_subtitle = true;
869 eof = true; // Or error, but ignore that for the time being.
872 // Decode audio, if any.
875 int err = avcodec_receive_frame(audio_codec_ctx, audio_avframe.get());
877 if (*audio_pts == -1) {
878 *audio_pts = audio_avframe->pts;
880 convert_audio(audio_avframe.get(), audio_frame, audio_format);
881 } else if (err == AVERROR(EAGAIN)) {
884 fprintf(stderr, "%s: Cannot receive frame from audio codec.\n", pathname.c_str());
886 return AVFrameWithDeleter(nullptr);
891 // Decode video, if we have a frame.
892 int err = avcodec_receive_frame(video_codec_ctx, video_avframe.get());
894 if (video_avframe->format == AV_PIX_FMT_VAAPI) {
895 // Get the frame down to the CPU. (TODO: See if we can keep it
896 // on the GPU all the way, since it will be going up again later.
897 // However, this only works if the OpenGL GPU is the same one.)
898 AVFrameWithDeleter sw_frame = av_frame_alloc_unique();
899 int err = av_hwframe_transfer_data(sw_frame.get(), video_avframe.get(), 0);
901 fprintf(stderr, "%s: Cannot transfer hardware video frame to software.\n", pathname.c_str());
903 return AVFrameWithDeleter(nullptr);
905 sw_frame->pts = video_avframe->pts;
906 sw_frame->pkt_duration = video_avframe->pkt_duration;
907 video_avframe = move(sw_frame);
909 frame_finished = true;
911 } else if (err != AVERROR(EAGAIN)) {
912 fprintf(stderr, "%s: Cannot receive frame from video codec.\n", pathname.c_str());
914 return AVFrameWithDeleter(nullptr);
919 return video_avframe;
921 return AVFrameWithDeleter(nullptr);
924 void FFmpegCapture::convert_audio(const AVFrame *audio_avframe, FrameAllocator::Frame *audio_frame, AudioFormat *audio_format)
926 // Decide on a format. If there already is one in this audio frame,
927 // we're pretty much forced to use it. If not, we try to find an exact match.
928 // If that still doesn't work, we default to 32-bit signed chunked
929 // (float would be nice, but there's really no way to signal that yet).
930 AVSampleFormat dst_format;
931 if (audio_format->bits_per_sample == 0) {
932 switch (audio_avframe->format) {
933 case AV_SAMPLE_FMT_S16:
934 case AV_SAMPLE_FMT_S16P:
935 audio_format->bits_per_sample = 16;
936 dst_format = AV_SAMPLE_FMT_S16;
938 case AV_SAMPLE_FMT_S32:
939 case AV_SAMPLE_FMT_S32P:
941 audio_format->bits_per_sample = 32;
942 dst_format = AV_SAMPLE_FMT_S32;
945 } else if (audio_format->bits_per_sample == 16) {
946 dst_format = AV_SAMPLE_FMT_S16;
947 } else if (audio_format->bits_per_sample == 32) {
948 dst_format = AV_SAMPLE_FMT_S32;
952 audio_format->num_channels = 2;
954 int64_t channel_layout = audio_avframe->channel_layout;
955 if (channel_layout == 0) {
956 channel_layout = av_get_default_channel_layout(audio_avframe->channels);
959 if (resampler == nullptr ||
960 audio_avframe->format != last_src_format ||
961 dst_format != last_dst_format ||
962 channel_layout != last_channel_layout ||
963 audio_avframe->sample_rate != last_sample_rate) {
964 swr_free(&resampler);
965 resampler = swr_alloc_set_opts(nullptr,
966 /*out_ch_layout=*/AV_CH_LAYOUT_STEREO_DOWNMIX,
967 /*out_sample_fmt=*/dst_format,
968 /*out_sample_rate=*/OUTPUT_FREQUENCY,
969 /*in_ch_layout=*/channel_layout,
970 /*in_sample_fmt=*/AVSampleFormat(audio_avframe->format),
971 /*in_sample_rate=*/audio_avframe->sample_rate,
973 /*log_ctx=*/nullptr);
975 if (resampler == nullptr) {
976 fprintf(stderr, "Allocating resampler failed.\n");
980 if (swr_init(resampler) < 0) {
981 fprintf(stderr, "Could not open resample context.\n");
985 last_src_format = AVSampleFormat(audio_avframe->format);
986 last_dst_format = dst_format;
987 last_channel_layout = channel_layout;
988 last_sample_rate = audio_avframe->sample_rate;
991 size_t bytes_per_sample = (audio_format->bits_per_sample / 8) * 2;
992 size_t num_samples_room = (audio_frame->size - audio_frame->len) / bytes_per_sample;
994 uint8_t *data = audio_frame->data + audio_frame->len;
995 int out_samples = swr_convert(resampler, &data, num_samples_room,
996 const_cast<const uint8_t **>(audio_avframe->data), audio_avframe->nb_samples);
997 if (out_samples < 0) {
998 fprintf(stderr, "Audio conversion failed.\n");
1002 audio_frame->len += out_samples * bytes_per_sample;
1005 VideoFormat FFmpegCapture::construct_video_format(const AVFrame *frame, AVRational video_timebase)
1007 VideoFormat video_format;
1008 video_format.width = frame_width(frame);
1009 video_format.height = frame_height(frame);
1010 if (pixel_format == bmusb::PixelFormat_8BitBGRA) {
1011 video_format.stride = frame_width(frame) * 4;
1012 } else if (pixel_format == FFmpegCapture::PixelFormat_NV12) {
1013 video_format.stride = frame_width(frame);
1015 assert(pixel_format == bmusb::PixelFormat_8BitYCbCrPlanar);
1016 video_format.stride = frame_width(frame);
1018 video_format.frame_rate_nom = video_timebase.den;
1019 video_format.frame_rate_den = frame->pkt_duration * video_timebase.num;
1020 video_format.has_signal = true;
1021 video_format.is_connected = true;
1022 return video_format;
1025 UniqueFrame FFmpegCapture::make_video_frame(const AVFrame *frame, const string &pathname, bool *error)
1029 UniqueFrame video_frame(video_frame_allocator->alloc_frame());
1030 if (video_frame->data == nullptr) {
1034 if (sws_ctx == nullptr ||
1035 sws_last_width != frame->width ||
1036 sws_last_height != frame->height ||
1037 sws_last_src_format != frame->format) {
1038 sws_dst_format = decide_dst_format(AVPixelFormat(frame->format), pixel_format);
1040 sws_getContext(frame->width, frame->height, AVPixelFormat(frame->format),
1041 frame_width(frame), frame_height(frame), sws_dst_format,
1042 SWS_BICUBIC, nullptr, nullptr, nullptr));
1043 sws_last_width = frame->width;
1044 sws_last_height = frame->height;
1045 sws_last_src_format = frame->format;
1047 if (sws_ctx == nullptr) {
1048 fprintf(stderr, "%s: Could not create scaler context\n", pathname.c_str());
1053 uint8_t *pic_data[4] = { nullptr, nullptr, nullptr, nullptr };
1054 int linesizes[4] = { 0, 0, 0, 0 };
1055 if (pixel_format == bmusb::PixelFormat_8BitBGRA) {
1056 pic_data[0] = video_frame->data;
1057 linesizes[0] = frame_width(frame) * 4;
1058 video_frame->len = (frame_width(frame) * 4) * frame_height(frame);
1059 } else if (pixel_format == PixelFormat_NV12) {
1060 pic_data[0] = video_frame->data;
1061 linesizes[0] = frame_width(frame);
1063 pic_data[1] = pic_data[0] + frame_width(frame) * frame_height(frame);
1064 linesizes[1] = frame_width(frame);
1066 video_frame->len = (frame_width(frame) * 2) * frame_height(frame);
1068 const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(sws_dst_format);
1069 current_frame_ycbcr_format = decode_ycbcr_format(desc, frame, is_mjpeg, &last_colorspace, &last_chroma_location);
1071 assert(pixel_format == bmusb::PixelFormat_8BitYCbCrPlanar);
1072 const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(sws_dst_format);
1074 int chroma_width = AV_CEIL_RSHIFT(int(frame_width(frame)), desc->log2_chroma_w);
1075 int chroma_height = AV_CEIL_RSHIFT(int(frame_height(frame)), desc->log2_chroma_h);
1077 pic_data[0] = video_frame->data;
1078 linesizes[0] = frame_width(frame);
1080 pic_data[1] = pic_data[0] + frame_width(frame) * frame_height(frame);
1081 linesizes[1] = chroma_width;
1083 pic_data[2] = pic_data[1] + chroma_width * chroma_height;
1084 linesizes[2] = chroma_width;
1086 video_frame->len = frame_width(frame) * frame_height(frame) + 2 * chroma_width * chroma_height;
1088 current_frame_ycbcr_format = decode_ycbcr_format(desc, frame, is_mjpeg, &last_colorspace, &last_chroma_location);
1090 sws_scale(sws_ctx.get(), frame->data, frame->linesize, 0, frame->height, pic_data, linesizes);
1095 int FFmpegCapture::interrupt_cb_thunk(void *opaque)
1097 return reinterpret_cast<FFmpegCapture *>(opaque)->interrupt_cb();
1100 int FFmpegCapture::interrupt_cb()
1102 return should_interrupt.load();
1105 unsigned FFmpegCapture::frame_width(const AVFrame *frame) const
1108 return frame->width;
1114 unsigned FFmpegCapture::frame_height(const AVFrame *frame) const
1117 return frame->height;
1124 int FFmpegCapture::read_srt_thunk(void *opaque, uint8_t *buf, int buf_size)
1126 return reinterpret_cast<FFmpegCapture *>(opaque)->read_srt(buf, buf_size);
1129 int FFmpegCapture::read_srt(uint8_t *buf, int buf_size)
1131 SRT_MSGCTRL mc = srt_msgctrl_default;
1132 return srt_recvmsg2(srt_sock, reinterpret_cast<char *>(buf), buf_size, &mc);