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 : width(width), height(height), video_timebase{1, 1}
263 filenames.push_back(filename);
264 description = "Video: " + filename;
266 last_frame = steady_clock::now();
268 avformat_network_init(); // In case someone wants this.
271 FFmpegCapture::FFmpegCapture(const std::vector<std::string> &filenames, unsigned width, unsigned height)
272 : filenames(filenames), width(width), height(height), video_timebase{1, 1}
274 description = "Video: " + filenames[0];
276 last_frame = steady_clock::now();
279 avformat_network_init(); // In case someone wants this.
283 FFmpegCapture::FFmpegCapture(int srt_sock, const string &stream_id)
284 : srt_sock(srt_sock),
285 width(0), // Don't resize; SRT streams typically have stable resolution, and should behave much like regular cards in general.
287 pixel_format(bmusb::PixelFormat_8BitYCbCrPlanar),
290 if (stream_id.empty()) {
291 description = "SRT stream";
293 description = stream_id;
295 play_as_fast_as_possible = true;
297 last_frame = steady_clock::now();
301 FFmpegCapture::~FFmpegCapture()
303 if (has_dequeue_callbacks) {
304 dequeue_cleanup_callback();
306 swr_free(&resampler);
308 if (srt_sock != -1) {
314 void FFmpegCapture::configure_card()
316 if (video_frame_allocator == nullptr) {
317 owned_video_frame_allocator.reset(new MallocFrameAllocator(FRAME_SIZE, NUM_QUEUED_VIDEO_FRAMES));
318 set_video_frame_allocator(owned_video_frame_allocator.get());
320 if (audio_frame_allocator == nullptr) {
321 // Audio can come out in pretty large chunks, so increase from the default 1 MB.
322 owned_audio_frame_allocator.reset(new MallocFrameAllocator(1 << 20, NUM_QUEUED_AUDIO_FRAMES));
323 set_audio_frame_allocator(owned_audio_frame_allocator.get());
327 void FFmpegCapture::start_bm_capture()
333 producer_thread_should_quit.unquit();
334 producer_thread = thread(&FFmpegCapture::producer_thread_func, this);
337 void FFmpegCapture::stop_dequeue_thread()
343 producer_thread_should_quit.quit();
344 producer_thread.join();
347 std::map<uint32_t, VideoMode> FFmpegCapture::get_available_video_modes() const
349 // Note: This will never really be shown in the UI.
353 snprintf(buf, sizeof(buf), "%ux%u", sws_last_width, sws_last_height);
356 mode.autodetect = false;
357 mode.width = sws_last_width;
358 mode.height = sws_last_height;
359 mode.frame_rate_num = 60;
360 mode.frame_rate_den = 1;
361 mode.interlaced = false;
363 return {{ 0, mode }};
366 void FFmpegCapture::producer_thread_func()
368 char thread_name[16];
369 snprintf(thread_name, sizeof(thread_name), "FFmpeg_C_%d", card_index);
370 pthread_setname_np(pthread_self(), thread_name);
373 while (!producer_thread_should_quit.should_quit()) {
374 string filename_copy;
375 printf("CAP %zu LEFT\n", filenames.size());
377 lock_guard<mutex> lock(filename_mu);
378 filename_copy = filenames.front();
379 filenames.erase(filenames.begin());
383 if (srt_sock == -1) {
384 pathname = search_for_file(filename_copy);
386 pathname = description;
388 printf("CAP %s\n", pathname.c_str());
389 if (pathname.empty()) {
390 send_disconnected_frame();
394 producer_thread_should_quit.sleep_for(seconds(1));
395 fprintf(stderr, "%s not found, sleeping one second and trying again...\n", filename_copy.c_str());
398 should_interrupt = false;
399 if (!play_video(pathname)) {
401 send_disconnected_frame();
405 fprintf(stderr, "Error when playing %s, sleeping one second and trying again...\n", pathname.c_str());
406 producer_thread_should_quit.sleep_for(seconds(1));
410 if (play_once && filenames.empty()) {
411 send_disconnected_frame();
415 // Probably just EOF, will exit the loop above on next test.
418 if (has_dequeue_callbacks) {
419 dequeue_cleanup_callback();
420 has_dequeue_callbacks = false;
424 void FFmpegCapture::send_disconnected_frame()
426 // Send an empty frame to signal that we have no signal anymore.
427 FrameAllocator::Frame video_frame = video_frame_allocator->alloc_frame();
428 size_t frame_width = width == 0 ? global_flags.width : width;
429 size_t frame_height = height == 0 ? global_flags.height : height;
430 if (video_frame.data) {
431 VideoFormat video_format;
432 video_format.width = frame_width;
433 video_format.height = frame_height;
434 video_format.frame_rate_nom = 60;
435 video_format.frame_rate_den = 1;
436 video_format.is_connected = false;
437 if (pixel_format == bmusb::PixelFormat_8BitBGRA) {
438 video_format.stride = frame_width * 4;
439 video_frame.len = frame_width * frame_height * 4;
440 memset(video_frame.data, 0, video_frame.len);
442 video_format.stride = frame_width;
443 current_frame_ycbcr_format.luma_coefficients = YCBCR_REC_709;
444 current_frame_ycbcr_format.full_range = true;
445 current_frame_ycbcr_format.num_levels = 256;
446 current_frame_ycbcr_format.chroma_subsampling_x = 2;
447 current_frame_ycbcr_format.chroma_subsampling_y = 2;
448 current_frame_ycbcr_format.cb_x_position = 0.0f;
449 current_frame_ycbcr_format.cb_y_position = 0.0f;
450 current_frame_ycbcr_format.cr_x_position = 0.0f;
451 current_frame_ycbcr_format.cr_y_position = 0.0f;
452 video_frame.len = frame_width * frame_height * 2;
453 memset(video_frame.data, 0, frame_width * frame_height);
454 memset(video_frame.data + frame_width * frame_height, 128, frame_width * frame_height); // Valid for both NV12 and planar.
457 if (frame_callback != nullptr) {
458 frame_callback(-1, AVRational{1, TIMEBASE}, -1, AVRational{1, TIMEBASE}, timecode++,
459 video_frame, /*video_offset=*/0, video_format,
460 FrameAllocator::Frame(), /*audio_offset=*/0, AudioFormat());
462 last_frame_was_connected = false;
467 if (card_disconnected_callback != nullptr) {
468 card_disconnected_callback();
473 AVPixelFormat get_vaapi_hw_format(AVCodecContext *ctx, const AVPixelFormat *fmt)
475 for (const AVPixelFormat *fmt_ptr = fmt; *fmt_ptr != -1; ++fmt_ptr) {
476 for (int i = 0;; ++i) { // Termination condition inside loop.
477 const AVCodecHWConfig *config = avcodec_get_hw_config(ctx->codec, i);
478 if (config == nullptr) { // End of list.
479 fprintf(stderr, "Decoder %s does not support device.\n", ctx->codec->name);
482 if (config->methods & AV_CODEC_HW_CONFIG_METHOD_HW_DEVICE_CTX &&
483 config->device_type == AV_HWDEVICE_TYPE_VAAPI &&
484 config->pix_fmt == *fmt_ptr) {
485 return config->pix_fmt;
490 // We found no VA-API formats, so take the best software format.
494 bool FFmpegCapture::play_video(const string &pathname)
496 // Note: Call before open, not after; otherwise, there's a race.
497 // (There is now, too, but it tips the correct way. We could use fstat()
498 // if we had the file descriptor.)
499 timespec last_modified;
501 if (stat(pathname.c_str(), &buf) != 0) {
502 // Probably some sort of protocol, so can't stat.
503 last_modified.tv_sec = -1;
505 last_modified = buf.st_mtim;
507 last_colorspace = static_cast<AVColorSpace>(-1);
508 last_chroma_location = static_cast<AVChromaLocation>(-1);
510 AVFormatContextWithCloser format_ctx;
511 if (srt_sock == -1) {
513 format_ctx = avformat_open_input_unique(pathname.c_str(), /*fmt=*/nullptr,
515 AVIOInterruptCB{ &FFmpegCapture::interrupt_cb_thunk, this });
518 // SRT socket, already opened.
519 const AVInputFormat *mpegts_fmt = av_find_input_format("mpegts");
520 format_ctx = avformat_open_input_unique(&FFmpegCapture::read_srt_thunk, this,
521 mpegts_fmt, /*options=*/nullptr,
522 AVIOInterruptCB{ &FFmpegCapture::interrupt_cb_thunk, this });
527 if (format_ctx == nullptr) {
528 fprintf(stderr, "%s: Error opening file\n", pathname.c_str());
532 if (avformat_find_stream_info(format_ctx.get(), nullptr) < 0) {
533 fprintf(stderr, "%s: Error finding stream info\n", pathname.c_str());
537 int video_stream_index = find_stream_index(format_ctx.get(), AVMEDIA_TYPE_VIDEO);
538 if (video_stream_index == -1) {
539 fprintf(stderr, "%s: No video stream found\n", pathname.c_str());
543 int audio_stream_index = find_stream_index(format_ctx.get(), AVMEDIA_TYPE_AUDIO);
544 int subtitle_stream_index = find_stream_index(format_ctx.get(), AVMEDIA_TYPE_SUBTITLE);
545 has_last_subtitle = false;
547 // Open video decoder.
548 const AVCodecParameters *video_codecpar = format_ctx->streams[video_stream_index]->codecpar;
549 const AVCodec *video_codec = avcodec_find_decoder(video_codecpar->codec_id);
551 video_timebase = format_ctx->streams[video_stream_index]->time_base;
552 AVCodecContextWithDeleter video_codec_ctx = avcodec_alloc_context3_unique(nullptr);
553 if (avcodec_parameters_to_context(video_codec_ctx.get(), video_codecpar) < 0) {
554 fprintf(stderr, "%s: Cannot fill video codec parameters\n", pathname.c_str());
557 if (video_codec == nullptr) {
558 fprintf(stderr, "%s: Cannot find video decoder\n", pathname.c_str());
562 // Seemingly, it's not too easy to make something that just initializes
563 // “whatever goes”, so we don't get VDPAU or CUDA here without enumerating
564 // through several different types. VA-API will do for now.
565 AVBufferRef *hw_device_ctx = nullptr;
566 if (av_hwdevice_ctx_create(&hw_device_ctx, AV_HWDEVICE_TYPE_VAAPI, nullptr, nullptr, 0) < 0 || true) {
567 fprintf(stderr, "Failed to initialize VA-API for FFmpeg acceleration. Decoding video in software.\n");
569 video_codec_ctx->hw_device_ctx = av_buffer_ref(hw_device_ctx);
570 video_codec_ctx->get_format = get_vaapi_hw_format;
573 if (avcodec_open2(video_codec_ctx.get(), video_codec, nullptr) < 0) {
574 fprintf(stderr, "%s: Cannot open video decoder\n", pathname.c_str());
577 unique_ptr<AVCodecContext, decltype(avcodec_close)*> video_codec_ctx_cleanup(
578 video_codec_ctx.get(), avcodec_close);
580 // Used in decode_ycbcr_format().
581 is_mjpeg = video_codecpar->codec_id == AV_CODEC_ID_MJPEG;
583 // Open audio decoder, if we have audio.
584 AVCodecContextWithDeleter audio_codec_ctx;
585 if (audio_stream_index != -1) {
586 audio_codec_ctx = avcodec_alloc_context3_unique(nullptr);
587 const AVCodecParameters *audio_codecpar = format_ctx->streams[audio_stream_index]->codecpar;
588 audio_timebase = format_ctx->streams[audio_stream_index]->time_base;
589 if (avcodec_parameters_to_context(audio_codec_ctx.get(), audio_codecpar) < 0) {
590 fprintf(stderr, "%s: Cannot fill audio codec parameters\n", pathname.c_str());
593 const AVCodec *audio_codec = avcodec_find_decoder(audio_codecpar->codec_id);
594 if (audio_codec == nullptr) {
595 fprintf(stderr, "%s: Cannot find audio decoder\n", pathname.c_str());
598 if (avcodec_open2(audio_codec_ctx.get(), audio_codec, nullptr) < 0) {
599 fprintf(stderr, "%s: Cannot open audio decoder\n", pathname.c_str());
603 unique_ptr<AVCodecContext, decltype(avcodec_close)*> audio_codec_ctx_cleanup(
604 audio_codec_ctx.get(), avcodec_close);
609 bool first_frame = true;
610 while (!producer_thread_should_quit.should_quit()) {
611 if (process_queued_commands(format_ctx.get(), pathname, last_modified, /*rewound=*/nullptr)) {
614 if (should_interrupt.load()) {
615 // Check as a failsafe, so that we don't need to rely on avio if we don't have to.
618 UniqueFrame audio_frame = audio_frame_allocator->alloc_frame();
619 AudioFormat audio_format;
623 AVFrameWithDeleter frame = decode_frame(format_ctx.get(), video_codec_ctx.get(), audio_codec_ctx.get(),
624 pathname, video_stream_index, audio_stream_index, subtitle_stream_index, audio_frame.get(), &audio_format, &audio_pts, &error);
628 if (frame == nullptr) {
629 // EOF. Loop back to the start if we can.
630 if (format_ctx->pb != nullptr && format_ctx->pb->seekable == 0) {
631 // Not seekable (but seemingly, sometimes av_seek_frame() would return 0 anyway,
635 if (true || av_seek_frame(format_ctx.get(), /*stream_index=*/-1, /*timestamp=*/0, /*flags=*/0) < 0) {
636 fprintf(stderr, "%s: Rewind failed, not looping.\n", pathname.c_str());
639 if (video_codec_ctx != nullptr) {
640 avcodec_flush_buffers(video_codec_ctx.get());
642 if (audio_codec_ctx != nullptr) {
643 avcodec_flush_buffers(audio_codec_ctx.get());
645 // If the file has changed since last time, return to get it reloaded.
646 // Note that depending on how you move the file into place, you might
647 // end up corrupting the one you're already playing, so this path
648 // might not trigger.
649 if (changed_since(pathname, last_modified)) {
656 VideoFormat video_format = construct_video_format(frame.get(), video_timebase);
657 if (video_format.frame_rate_nom == 0 || video_format.frame_rate_den == 0) {
658 // Invalid frame rate; try constructing it from the previous frame length.
659 // (This is especially important if we are the master card, for SRT,
660 // since it affects audio. Not all senders have good timebases
661 // (e.g., Larix rounds first to timebase 1000 and then multiplies by
662 // 90 from there, it seems), but it's much better to have an oscillating
663 // value than just locking at 60.
664 if (last_pts != 0 && frame->pts > last_pts) {
665 int64_t pts_diff = frame->pts - last_pts;
666 video_format.frame_rate_nom = video_timebase.den;
667 video_format.frame_rate_den = video_timebase.num * pts_diff;
669 video_format.frame_rate_nom = 60;
670 video_format.frame_rate_den = 1;
673 UniqueFrame video_frame = make_video_frame(frame.get(), pathname, &error);
679 if (last_pts == 0 && pts_origin == 0) {
680 pts_origin = frame->pts;
682 steady_clock::time_point now = steady_clock::now();
683 if (play_as_fast_as_possible) {
684 video_frame->received_timestamp = now;
685 audio_frame->received_timestamp = now;
686 next_frame_start = now;
688 next_frame_start = compute_frame_start(frame->pts, pts_origin, video_timebase, start, rate);
689 if (first_frame && last_frame_was_connected) {
690 // If reconnect took more than one second, this is probably a live feed,
691 // and we should reset the resampler. (Or the rate is really, really low,
692 // in which case a reset on the first frame is fine anyway.)
693 if (duration<double>(next_frame_start - last_frame).count() >= 1.0) {
694 last_frame_was_connected = false;
697 video_frame->received_timestamp = next_frame_start;
699 // The easiest way to get all the rate conversions etc. right is to move the
700 // audio PTS into the video PTS timebase and go from there. (We'll get some
701 // rounding issues, but they should not be a big problem.)
702 int64_t audio_pts_as_video_pts = av_rescale_q(audio_pts, audio_timebase, video_timebase);
703 audio_frame->received_timestamp = compute_frame_start(audio_pts_as_video_pts, pts_origin, video_timebase, start, rate);
705 if (audio_frame->len != 0) {
706 // The received timestamps in Nageru are measured after we've just received the frame.
707 // However, pts (especially audio pts) is at the _beginning_ of the frame.
708 // If we have locked audio, the distinction doesn't really matter, as pts is
709 // on a relative scale and a fixed offset is fine. But if we don't, we will have
710 // a different number of samples each time, which will cause huge audio jitter
711 // and throw off the resampler.
713 // In a sense, we should have compensated by adding the frame and audio lengths
714 // to video_frame->received_timestamp and audio_frame->received_timestamp respectively,
715 // but that would mean extra waiting in sleep_until(). All we need is that they
716 // are correct relative to each other, though (and to the other frames we send),
717 // so just align the end of the audio frame, and we're fine.
718 size_t num_samples = (audio_frame->len * 8) / audio_format.bits_per_sample / audio_format.num_channels;
719 double offset = double(num_samples) / OUTPUT_FREQUENCY -
720 double(video_format.frame_rate_den) / video_format.frame_rate_nom;
721 audio_frame->received_timestamp += duration_cast<steady_clock::duration>(duration<double>(offset));
724 if (duration<double>(now - next_frame_start).count() >= 0.1) {
725 // If we don't have enough CPU to keep up, or if we have a live stream
726 // where the initial origin was somehow wrong, we could be behind indefinitely.
727 // In particular, this will give the audio resampler problems as it tries
728 // to speed up to reduce the delay, hitting the low end of the buffer every time.
729 fprintf(stderr, "%s: Playback %.0f ms behind, resetting time scale\n",
731 1e3 * duration<double>(now - next_frame_start).count());
732 pts_origin = frame->pts;
733 start = next_frame_start = now;
734 timecode += MAX_FPS * 2 + 1;
737 bool finished_wakeup;
738 if (play_as_fast_as_possible) {
739 finished_wakeup = !producer_thread_should_quit.should_quit();
741 finished_wakeup = producer_thread_should_quit.sleep_until(next_frame_start);
743 if (finished_wakeup) {
744 if (audio_frame->len > 0) {
745 assert(audio_pts != -1);
747 if (!last_frame_was_connected) {
748 // We're recovering from an error (or really slow load, see above).
749 // Make sure to get the audio resampler reset. (This is a hack;
750 // ideally, the frame callback should just accept a way to signal
751 // audio discontinuity.)
752 timecode += MAX_FPS * 2 + 1;
754 last_neutral_color = get_neutral_color(frame->metadata);
755 if (frame_callback != nullptr) {
756 frame_callback(frame->pts, video_timebase, audio_pts, audio_timebase, timecode++,
757 video_frame.get_and_release(), 0, video_format,
758 audio_frame.get_and_release(), 0, audio_format);
761 last_frame = steady_clock::now();
762 last_frame_was_connected = true;
765 if (producer_thread_should_quit.should_quit()) break;
767 bool rewound = false;
768 if (process_queued_commands(format_ctx.get(), pathname, last_modified, &rewound)) {
771 // If we just rewound, drop this frame on the floor and be done.
775 // OK, we didn't, so probably a rate change. Recalculate next_frame_start,
776 // but if it's now in the past, we'll reset the origin, so that we don't
777 // generate a huge backlog of frames that we need to run through quickly.
778 next_frame_start = compute_frame_start(frame->pts, pts_origin, video_timebase, start, rate);
779 steady_clock::time_point now = steady_clock::now();
780 if (next_frame_start < now) {
781 pts_origin = frame->pts;
782 start = next_frame_start = now;
786 last_pts = frame->pts;
791 void FFmpegCapture::internal_rewind()
793 pts_origin = last_pts = 0;
794 start = next_frame_start = steady_clock::now();
797 bool FFmpegCapture::process_queued_commands(AVFormatContext *format_ctx, const std::string &pathname, timespec last_modified, bool *rewound)
799 // Process any queued commands from other threads.
800 vector<QueuedCommand> commands;
802 lock_guard<mutex> lock(queue_mu);
803 swap(commands, command_queue);
805 for (const QueuedCommand &cmd : commands) {
806 switch (cmd.command) {
807 case QueuedCommand::REWIND:
808 if (av_seek_frame(format_ctx, /*stream_index=*/-1, /*timestamp=*/0, /*flags=*/0) < 0) {
809 fprintf(stderr, "%s: Rewind failed, stopping play.\n", pathname.c_str());
811 // If the file has changed since last time, return to get it reloaded.
812 // Note that depending on how you move the file into place, you might
813 // end up corrupting the one you're already playing, so this path
814 // might not trigger.
815 if (changed_since(pathname, last_modified)) {
819 if (rewound != nullptr) {
824 case QueuedCommand::CHANGE_RATE:
825 // Change the origin to the last played frame.
826 start = compute_frame_start(last_pts, pts_origin, video_timebase, start, rate);
827 pts_origin = last_pts;
829 play_as_fast_as_possible = (rate >= 10.0);
840 AVFrameWithDeleter FFmpegCapture::decode_frame(AVFormatContext *format_ctx, AVCodecContext *video_codec_ctx, AVCodecContext *audio_codec_ctx,
841 const std::string &pathname, int video_stream_index, int audio_stream_index, int subtitle_stream_index,
842 FrameAllocator::Frame *audio_frame, AudioFormat *audio_format, int64_t *audio_pts, bool *error)
846 // Read packets until we have a frame or there are none left.
847 bool frame_finished = false;
848 AVFrameWithDeleter audio_avframe = av_frame_alloc_unique();
849 AVFrameWithDeleter video_avframe = av_frame_alloc_unique();
852 bool has_audio = false;
855 unique_ptr<AVPacket, decltype(av_packet_unref)*> pkt_cleanup(
856 &pkt, av_packet_unref);
857 av_init_packet(&pkt);
860 if (av_read_frame(format_ctx, &pkt) == 0) {
861 if (pkt.stream_index == audio_stream_index && audio_callback != nullptr) {
862 audio_callback(&pkt, format_ctx->streams[audio_stream_index]->time_base);
864 if (pkt.stream_index == video_stream_index && video_callback != nullptr) {
865 video_callback(&pkt, format_ctx->streams[video_stream_index]->time_base);
867 if (pkt.stream_index == video_stream_index && global_flags.transcode_video) {
868 if (avcodec_send_packet(video_codec_ctx, &pkt) < 0) {
869 fprintf(stderr, "%s: Cannot send packet to video codec.\n", pathname.c_str());
871 return AVFrameWithDeleter(nullptr);
873 } else if (pkt.stream_index == audio_stream_index && global_flags.transcode_audio) {
875 if (avcodec_send_packet(audio_codec_ctx, &pkt) < 0) {
876 fprintf(stderr, "%s: Cannot send packet to audio codec.\n", pathname.c_str());
878 return AVFrameWithDeleter(nullptr);
880 } else if (pkt.stream_index == subtitle_stream_index) {
881 last_subtitle = string(reinterpret_cast<const char *>(pkt.data), pkt.size);
882 has_last_subtitle = true;
885 eof = true; // Or error, but ignore that for the time being.
888 // Decode audio, if any.
891 int err = avcodec_receive_frame(audio_codec_ctx, audio_avframe.get());
893 if (*audio_pts == -1) {
894 *audio_pts = audio_avframe->pts;
896 convert_audio(audio_avframe.get(), audio_frame, audio_format);
897 } else if (err == AVERROR(EAGAIN)) {
900 fprintf(stderr, "%s: Cannot receive frame from audio codec.\n", pathname.c_str());
902 return AVFrameWithDeleter(nullptr);
907 // Decode video, if we have a frame.
908 int err = avcodec_receive_frame(video_codec_ctx, video_avframe.get());
910 if (video_avframe->format == AV_PIX_FMT_VAAPI) {
911 // Get the frame down to the CPU. (TODO: See if we can keep it
912 // on the GPU all the way, since it will be going up again later.
913 // However, this only works if the OpenGL GPU is the same one.)
914 AVFrameWithDeleter sw_frame = av_frame_alloc_unique();
915 int err = av_hwframe_transfer_data(sw_frame.get(), video_avframe.get(), 0);
917 fprintf(stderr, "%s: Cannot transfer hardware video frame to software.\n", pathname.c_str());
919 return AVFrameWithDeleter(nullptr);
921 sw_frame->pts = video_avframe->pts;
922 sw_frame->pkt_duration = video_avframe->pkt_duration;
923 video_avframe = move(sw_frame);
925 frame_finished = true;
927 } else if (err != AVERROR(EAGAIN)) {
928 fprintf(stderr, "%s: Cannot receive frame from video codec.\n", pathname.c_str());
930 return AVFrameWithDeleter(nullptr);
935 return video_avframe;
937 return AVFrameWithDeleter(nullptr);
940 void FFmpegCapture::convert_audio(const AVFrame *audio_avframe, FrameAllocator::Frame *audio_frame, AudioFormat *audio_format)
942 // Decide on a format. If there already is one in this audio frame,
943 // we're pretty much forced to use it. If not, we try to find an exact match.
944 // If that still doesn't work, we default to 32-bit signed chunked
945 // (float would be nice, but there's really no way to signal that yet).
946 AVSampleFormat dst_format;
947 if (audio_format->bits_per_sample == 0) {
948 switch (audio_avframe->format) {
949 case AV_SAMPLE_FMT_S16:
950 case AV_SAMPLE_FMT_S16P:
951 audio_format->bits_per_sample = 16;
952 dst_format = AV_SAMPLE_FMT_S16;
954 case AV_SAMPLE_FMT_S32:
955 case AV_SAMPLE_FMT_S32P:
957 audio_format->bits_per_sample = 32;
958 dst_format = AV_SAMPLE_FMT_S32;
961 } else if (audio_format->bits_per_sample == 16) {
962 dst_format = AV_SAMPLE_FMT_S16;
963 } else if (audio_format->bits_per_sample == 32) {
964 dst_format = AV_SAMPLE_FMT_S32;
968 audio_format->num_channels = 2;
970 int64_t channel_layout = audio_avframe->channel_layout;
971 if (channel_layout == 0) {
972 channel_layout = av_get_default_channel_layout(audio_avframe->channels);
975 if (resampler == nullptr ||
976 audio_avframe->format != last_src_format ||
977 dst_format != last_dst_format ||
978 channel_layout != last_channel_layout ||
979 audio_avframe->sample_rate != last_sample_rate) {
980 swr_free(&resampler);
981 resampler = swr_alloc_set_opts(nullptr,
982 /*out_ch_layout=*/AV_CH_LAYOUT_STEREO_DOWNMIX,
983 /*out_sample_fmt=*/dst_format,
984 /*out_sample_rate=*/OUTPUT_FREQUENCY,
985 /*in_ch_layout=*/channel_layout,
986 /*in_sample_fmt=*/AVSampleFormat(audio_avframe->format),
987 /*in_sample_rate=*/audio_avframe->sample_rate,
989 /*log_ctx=*/nullptr);
991 if (resampler == nullptr) {
992 fprintf(stderr, "Allocating resampler failed.\n");
996 if (swr_init(resampler) < 0) {
997 fprintf(stderr, "Could not open resample context.\n");
1001 last_src_format = AVSampleFormat(audio_avframe->format);
1002 last_dst_format = dst_format;
1003 last_channel_layout = channel_layout;
1004 last_sample_rate = audio_avframe->sample_rate;
1007 size_t bytes_per_sample = (audio_format->bits_per_sample / 8) * 2;
1008 size_t num_samples_room = (audio_frame->size - audio_frame->len) / bytes_per_sample;
1010 uint8_t *data = audio_frame->data + audio_frame->len;
1011 int out_samples = swr_convert(resampler, &data, num_samples_room,
1012 const_cast<const uint8_t **>(audio_avframe->data), audio_avframe->nb_samples);
1013 if (out_samples < 0) {
1014 fprintf(stderr, "Audio conversion failed.\n");
1018 audio_frame->len += out_samples * bytes_per_sample;
1021 VideoFormat FFmpegCapture::construct_video_format(const AVFrame *frame, AVRational video_timebase)
1023 VideoFormat video_format;
1024 video_format.width = frame_width(frame);
1025 video_format.height = frame_height(frame);
1026 if (pixel_format == bmusb::PixelFormat_8BitBGRA) {
1027 video_format.stride = frame_width(frame) * 4;
1028 } else if (pixel_format == FFmpegCapture::PixelFormat_NV12) {
1029 video_format.stride = frame_width(frame);
1031 assert(pixel_format == bmusb::PixelFormat_8BitYCbCrPlanar);
1032 video_format.stride = frame_width(frame);
1034 video_format.frame_rate_nom = video_timebase.den;
1035 video_format.frame_rate_den = frame->pkt_duration * video_timebase.num;
1036 video_format.has_signal = true;
1037 video_format.is_connected = true;
1038 return video_format;
1041 UniqueFrame FFmpegCapture::make_video_frame(const AVFrame *frame, const string &pathname, bool *error)
1045 UniqueFrame video_frame(video_frame_allocator->alloc_frame());
1046 if (video_frame->data == nullptr) {
1050 if (sws_ctx == nullptr ||
1051 sws_last_width != frame->width ||
1052 sws_last_height != frame->height ||
1053 sws_last_src_format != frame->format) {
1054 sws_dst_format = decide_dst_format(AVPixelFormat(frame->format), pixel_format);
1056 sws_getContext(frame->width, frame->height, AVPixelFormat(frame->format),
1057 frame_width(frame), frame_height(frame), sws_dst_format,
1058 SWS_BICUBIC, nullptr, nullptr, nullptr));
1059 sws_last_width = frame->width;
1060 sws_last_height = frame->height;
1061 sws_last_src_format = frame->format;
1063 if (sws_ctx == nullptr) {
1064 fprintf(stderr, "%s: Could not create scaler context\n", pathname.c_str());
1069 uint8_t *pic_data[4] = { nullptr, nullptr, nullptr, nullptr };
1070 int linesizes[4] = { 0, 0, 0, 0 };
1071 if (pixel_format == bmusb::PixelFormat_8BitBGRA) {
1072 pic_data[0] = video_frame->data;
1073 linesizes[0] = frame_width(frame) * 4;
1074 video_frame->len = (frame_width(frame) * 4) * frame_height(frame);
1075 } else if (pixel_format == PixelFormat_NV12) {
1076 pic_data[0] = video_frame->data;
1077 linesizes[0] = frame_width(frame);
1079 pic_data[1] = pic_data[0] + frame_width(frame) * frame_height(frame);
1080 linesizes[1] = frame_width(frame);
1082 video_frame->len = (frame_width(frame) * 2) * frame_height(frame);
1084 const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(sws_dst_format);
1085 current_frame_ycbcr_format = decode_ycbcr_format(desc, frame, is_mjpeg, &last_colorspace, &last_chroma_location);
1087 assert(pixel_format == bmusb::PixelFormat_8BitYCbCrPlanar);
1088 const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(sws_dst_format);
1090 int chroma_width = AV_CEIL_RSHIFT(int(frame_width(frame)), desc->log2_chroma_w);
1091 int chroma_height = AV_CEIL_RSHIFT(int(frame_height(frame)), desc->log2_chroma_h);
1093 pic_data[0] = video_frame->data;
1094 linesizes[0] = frame_width(frame);
1096 pic_data[1] = pic_data[0] + frame_width(frame) * frame_height(frame);
1097 linesizes[1] = chroma_width;
1099 pic_data[2] = pic_data[1] + chroma_width * chroma_height;
1100 linesizes[2] = chroma_width;
1102 video_frame->len = frame_width(frame) * frame_height(frame) + 2 * chroma_width * chroma_height;
1104 current_frame_ycbcr_format = decode_ycbcr_format(desc, frame, is_mjpeg, &last_colorspace, &last_chroma_location);
1106 sws_scale(sws_ctx.get(), frame->data, frame->linesize, 0, frame->height, pic_data, linesizes);
1111 int FFmpegCapture::interrupt_cb_thunk(void *opaque)
1113 return reinterpret_cast<FFmpegCapture *>(opaque)->interrupt_cb();
1116 int FFmpegCapture::interrupt_cb()
1118 return should_interrupt.load();
1121 unsigned FFmpegCapture::frame_width(const AVFrame *frame) const
1124 return frame->width;
1130 unsigned FFmpegCapture::frame_height(const AVFrame *frame) const
1133 return frame->height;
1140 int FFmpegCapture::read_srt_thunk(void *opaque, uint8_t *buf, int buf_size)
1142 return reinterpret_cast<FFmpegCapture *>(opaque)->read_srt(buf, buf_size);
1145 int FFmpegCapture::read_srt(uint8_t *buf, int buf_size)
1147 SRT_MSGCTRL mc = srt_msgctrl_default;
1148 return srt_recvmsg2(srt_sock, reinterpret_cast<char *>(buf), buf_size, &mc);