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 "image_input.h"
40 #include "ref_counted_frame.h"
41 #include "shared/timebase.h"
48 using namespace std::chrono;
49 using namespace bmusb;
50 using namespace movit;
51 using namespace Eigen;
55 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)
57 const duration<double> pts((frame_pts - pts_origin) * double(video_timebase.num) / double(video_timebase.den));
58 return origin + duration_cast<steady_clock::duration>(pts / rate);
61 bool changed_since(const std::string &pathname, const timespec &ts)
67 if (stat(pathname.c_str(), &buf) != 0) {
68 fprintf(stderr, "%s: Couldn't check for new version, leaving the old in place.\n", pathname.c_str());
71 return (buf.st_mtim.tv_sec != ts.tv_sec || buf.st_mtim.tv_nsec != ts.tv_nsec);
74 bool is_full_range(const AVPixFmtDescriptor *desc)
76 // This is horrible, but there's no better way that I know of.
77 return (strchr(desc->name, 'j') != nullptr);
80 AVPixelFormat decide_dst_format(AVPixelFormat src_format, bmusb::PixelFormat dst_format_type)
82 if (dst_format_type == bmusb::PixelFormat_8BitBGRA) {
83 return AV_PIX_FMT_BGRA;
85 if (dst_format_type == FFmpegCapture::PixelFormat_NV12) {
86 return AV_PIX_FMT_NV12;
89 assert(dst_format_type == bmusb::PixelFormat_8BitYCbCrPlanar);
91 // If this is a non-Y'CbCr format, just convert to 4:4:4 Y'CbCr
92 // and be done with it. It's too strange to spend a lot of time on.
93 // (Let's hope there's no alpha.)
94 const AVPixFmtDescriptor *src_desc = av_pix_fmt_desc_get(src_format);
95 if (src_desc == nullptr ||
96 src_desc->nb_components != 3 ||
97 (src_desc->flags & AV_PIX_FMT_FLAG_RGB)) {
98 return AV_PIX_FMT_YUV444P;
101 // The best for us would be Cb and Cr together if possible,
102 // but FFmpeg doesn't support that except in the special case of
103 // NV12, so we need to go to planar even for the case of NV12.
104 // Thus, look for the closest (but no worse) 8-bit planar Y'CbCr format
105 // that matches in color range. (This will also include the case of
106 // the source format already being acceptable.)
107 bool src_full_range = is_full_range(src_desc);
108 const char *best_format = "yuv444p";
109 unsigned best_score = numeric_limits<unsigned>::max();
110 for (const AVPixFmtDescriptor *desc = av_pix_fmt_desc_next(nullptr);
112 desc = av_pix_fmt_desc_next(desc)) {
113 // Find planar Y'CbCr formats only.
114 if (desc->nb_components != 3) continue;
115 if (desc->flags & AV_PIX_FMT_FLAG_RGB) continue;
116 if (!(desc->flags & AV_PIX_FMT_FLAG_PLANAR)) continue;
117 if (desc->comp[0].plane != 0 ||
118 desc->comp[1].plane != 1 ||
119 desc->comp[2].plane != 2) continue;
121 // 8-bit formats only.
122 if (desc->flags & AV_PIX_FMT_FLAG_BE) continue;
123 if (desc->comp[0].depth != 8) continue;
125 // Same or better chroma resolution only.
126 int chroma_w_diff = desc->log2_chroma_w - src_desc->log2_chroma_w;
127 int chroma_h_diff = desc->log2_chroma_h - src_desc->log2_chroma_h;
128 if (chroma_w_diff < 0 || chroma_h_diff < 0)
131 // Matching full/limited range only.
132 if (is_full_range(desc) != src_full_range)
135 // Pick something with as little excess chroma resolution as possible.
136 unsigned score = (1 << (chroma_w_diff)) << chroma_h_diff;
137 if (score < best_score) {
139 best_format = desc->name;
142 return av_get_pix_fmt(best_format);
145 YCbCrFormat decode_ycbcr_format(const AVPixFmtDescriptor *desc, const AVFrame *frame, bool is_mjpeg, AVColorSpace *last_colorspace, AVChromaLocation *last_chroma_location)
148 AVColorSpace colorspace = frame->colorspace;
149 switch (colorspace) {
150 case AVCOL_SPC_BT709:
151 format.luma_coefficients = YCBCR_REC_709;
153 case AVCOL_SPC_BT470BG:
154 case AVCOL_SPC_SMPTE170M:
155 case AVCOL_SPC_SMPTE240M:
156 format.luma_coefficients = YCBCR_REC_601;
158 case AVCOL_SPC_BT2020_NCL:
159 format.luma_coefficients = YCBCR_REC_2020;
161 case AVCOL_SPC_UNSPECIFIED:
162 format.luma_coefficients = (frame->height >= 720 ? YCBCR_REC_709 : YCBCR_REC_601);
165 if (colorspace != *last_colorspace) {
166 fprintf(stderr, "Unknown Y'CbCr coefficient enum %d from FFmpeg; choosing Rec. 709.\n",
169 format.luma_coefficients = YCBCR_REC_709;
172 *last_colorspace = colorspace;
174 format.full_range = is_full_range(desc);
175 format.num_levels = 1 << desc->comp[0].depth;
176 format.chroma_subsampling_x = 1 << desc->log2_chroma_w;
177 format.chroma_subsampling_y = 1 << desc->log2_chroma_h;
179 switch (frame->chroma_location) {
180 case AVCHROMA_LOC_LEFT:
181 format.cb_x_position = 0.0;
182 format.cb_y_position = 0.5;
184 case AVCHROMA_LOC_CENTER:
185 format.cb_x_position = 0.5;
186 format.cb_y_position = 0.5;
188 case AVCHROMA_LOC_TOPLEFT:
189 format.cb_x_position = 0.0;
190 format.cb_y_position = 0.0;
192 case AVCHROMA_LOC_TOP:
193 format.cb_x_position = 0.5;
194 format.cb_y_position = 0.0;
196 case AVCHROMA_LOC_BOTTOMLEFT:
197 format.cb_x_position = 0.0;
198 format.cb_y_position = 1.0;
200 case AVCHROMA_LOC_BOTTOM:
201 format.cb_x_position = 0.5;
202 format.cb_y_position = 1.0;
205 if (frame->chroma_location != *last_chroma_location) {
206 fprintf(stderr, "Unknown chroma location coefficient enum %d from FFmpeg; choosing center.\n",
207 frame->chroma_location);
209 format.cb_x_position = 0.5;
210 format.cb_y_position = 0.5;
213 *last_chroma_location = frame->chroma_location;
215 if (is_mjpeg && !format.full_range) {
216 // Limited-range MJPEG is only detected by FFmpeg whenever a special
217 // JPEG comment is set, which means that in practice, the stream is
218 // almost certainly generated by Futatabi. Override FFmpeg's forced
219 // MJPEG defaults (it disregards the values set in the mux) with what
221 format.luma_coefficients = YCBCR_REC_709;
222 format.cb_x_position = 0.0;
223 format.cb_y_position = 0.5;
226 format.cr_x_position = format.cb_x_position;
227 format.cr_y_position = format.cb_y_position;
231 RGBTriplet get_neutral_color(AVDictionary *metadata)
233 if (metadata == nullptr) {
234 return RGBTriplet(1.0f, 1.0f, 1.0f);
236 AVDictionaryEntry *entry = av_dict_get(metadata, "WhitePoint", nullptr, 0);
237 if (entry == nullptr) {
238 return RGBTriplet(1.0f, 1.0f, 1.0f);
241 unsigned x_nom, x_den, y_nom, y_den;
242 if (sscanf(entry->value, " %u:%u , %u:%u", &x_nom, &x_den, &y_nom, &y_den) != 4) {
243 fprintf(stderr, "WARNING: Unable to parse white point '%s', using default white point\n", entry->value);
244 return RGBTriplet(1.0f, 1.0f, 1.0f);
247 double x = double(x_nom) / x_den;
248 double y = double(y_nom) / y_den;
249 double z = 1.0 - x - y;
251 Matrix3d rgb_to_xyz_matrix = movit::ColorspaceConversionEffect::get_xyz_matrix(COLORSPACE_sRGB);
252 Vector3d rgb = rgb_to_xyz_matrix.inverse() * Vector3d(x, y, z);
254 return RGBTriplet(rgb[0], rgb[1], rgb[2]);
259 FFmpegCapture::FFmpegCapture(const string &filename, unsigned width, unsigned height)
260 : filename(filename), width(width), height(height), video_timebase{1, 1}
262 description = "Video: " + filename;
264 last_frame = steady_clock::now();
266 avformat_network_init(); // In case someone wants this.
270 FFmpegCapture::FFmpegCapture(int srt_sock, const string &stream_id)
271 : srt_sock(srt_sock),
272 width(0), // Don't resize; SRT streams typically have stable resolution, and should behave much like regular cards in general.
274 pixel_format(bmusb::PixelFormat_8BitYCbCrPlanar),
277 if (stream_id.empty()) {
278 description = "SRT stream";
280 description = stream_id;
282 play_as_fast_as_possible = true;
284 last_frame = steady_clock::now();
288 FFmpegCapture::~FFmpegCapture()
290 if (has_dequeue_callbacks) {
291 dequeue_cleanup_callback();
293 swr_free(&resampler);
295 if (srt_sock != -1) {
301 void FFmpegCapture::configure_card()
303 if (video_frame_allocator == nullptr) {
304 owned_video_frame_allocator.reset(new MallocFrameAllocator(FRAME_SIZE, NUM_QUEUED_VIDEO_FRAMES));
305 set_video_frame_allocator(owned_video_frame_allocator.get());
307 if (audio_frame_allocator == nullptr) {
308 // Audio can come out in pretty large chunks, so increase from the default 1 MB.
309 owned_audio_frame_allocator.reset(new MallocFrameAllocator(1 << 20, NUM_QUEUED_AUDIO_FRAMES));
310 set_audio_frame_allocator(owned_audio_frame_allocator.get());
314 void FFmpegCapture::start_bm_capture()
320 producer_thread_should_quit.unquit();
321 producer_thread = thread(&FFmpegCapture::producer_thread_func, this);
324 void FFmpegCapture::stop_dequeue_thread()
330 producer_thread_should_quit.quit();
331 producer_thread.join();
334 std::map<uint32_t, VideoMode> FFmpegCapture::get_available_video_modes() const
336 // Note: This will never really be shown in the UI.
340 snprintf(buf, sizeof(buf), "%ux%u", sws_last_width, sws_last_height);
343 mode.autodetect = false;
344 mode.width = sws_last_width;
345 mode.height = sws_last_height;
346 mode.frame_rate_num = 60;
347 mode.frame_rate_den = 1;
348 mode.interlaced = false;
350 return {{ 0, mode }};
353 void FFmpegCapture::producer_thread_func()
355 char thread_name[16];
356 snprintf(thread_name, sizeof(thread_name), "FFmpeg_C_%d", card_index);
357 pthread_setname_np(pthread_self(), thread_name);
359 while (!producer_thread_should_quit.should_quit()) {
360 string filename_copy;
362 lock_guard<mutex> lock(filename_mu);
363 filename_copy = filename;
367 if (srt_sock == -1) {
368 pathname = search_for_file(filename_copy);
370 pathname = description;
372 if (pathname.empty()) {
373 send_disconnected_frame();
377 producer_thread_should_quit.sleep_for(seconds(1));
378 fprintf(stderr, "%s not found, sleeping one second and trying again...\n", filename_copy.c_str());
381 should_interrupt = false;
382 if (!play_video(pathname)) {
384 send_disconnected_frame();
388 fprintf(stderr, "Error when playing %s, sleeping one second and trying again...\n", pathname.c_str());
389 producer_thread_should_quit.sleep_for(seconds(1));
394 send_disconnected_frame();
398 // Probably just EOF, will exit the loop above on next test.
401 if (has_dequeue_callbacks) {
402 dequeue_cleanup_callback();
403 has_dequeue_callbacks = false;
407 void FFmpegCapture::send_disconnected_frame()
409 // Send an empty frame to signal that we have no signal anymore.
410 FrameAllocator::Frame video_frame = video_frame_allocator->alloc_frame();
411 size_t frame_width = width == 0 ? global_flags.width : width;
412 size_t frame_height = height == 0 ? global_flags.height : height;
413 if (video_frame.data) {
414 VideoFormat video_format;
415 video_format.width = frame_width;
416 video_format.height = frame_height;
417 video_format.frame_rate_nom = 60;
418 video_format.frame_rate_den = 1;
419 video_format.is_connected = false;
420 if (pixel_format == bmusb::PixelFormat_8BitBGRA) {
421 video_format.stride = frame_width * 4;
422 video_frame.len = frame_width * frame_height * 4;
423 memset(video_frame.data, 0, video_frame.len);
425 video_format.stride = frame_width;
426 current_frame_ycbcr_format.luma_coefficients = YCBCR_REC_709;
427 current_frame_ycbcr_format.full_range = true;
428 current_frame_ycbcr_format.num_levels = 256;
429 current_frame_ycbcr_format.chroma_subsampling_x = 2;
430 current_frame_ycbcr_format.chroma_subsampling_y = 2;
431 current_frame_ycbcr_format.cb_x_position = 0.0f;
432 current_frame_ycbcr_format.cb_y_position = 0.0f;
433 current_frame_ycbcr_format.cr_x_position = 0.0f;
434 current_frame_ycbcr_format.cr_y_position = 0.0f;
435 video_frame.len = frame_width * frame_height * 2;
436 memset(video_frame.data, 0, frame_width * frame_height);
437 memset(video_frame.data + frame_width * frame_height, 128, frame_width * frame_height); // Valid for both NV12 and planar.
440 if (frame_callback != nullptr) {
441 frame_callback(-1, AVRational{1, TIMEBASE}, -1, AVRational{1, TIMEBASE}, timecode++,
442 video_frame, /*video_offset=*/0, video_format,
443 FrameAllocator::Frame(), /*audio_offset=*/0, AudioFormat());
445 last_frame_was_connected = false;
450 if (card_disconnected_callback != nullptr) {
451 card_disconnected_callback();
456 template<AVHWDeviceType type>
457 AVPixelFormat get_hw_format(AVCodecContext *ctx, const AVPixelFormat *fmt)
459 bool found_config_of_right_type = false;
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.
465 if (!(config->methods & AV_CODEC_HW_CONFIG_METHOD_HW_DEVICE_CTX) ||
466 config->device_type != type) {
467 // Not interesting for us.
471 // We have a config of the right type, but does it actually support
472 // the pixel format we want? (Seemingly, FFmpeg's way of signaling errors
473 // is to just replace the pixel format with a software-decoded one,
475 found_config_of_right_type = true;
476 for (const AVPixelFormat *fmt_ptr = fmt; *fmt_ptr != -1; ++fmt_ptr) {
477 if (config->pix_fmt == *fmt_ptr) {
478 fprintf(stderr, "Initialized '%s' hardware decoding for codec '%s'.\n",
479 av_hwdevice_get_type_name(type), ctx->codec->name);
480 if (ctx->profile == FF_PROFILE_H264_BASELINE) {
481 fprintf(stderr, "WARNING: Stream claims to be H.264 Baseline, which is generally poorly supported in hardware decoders.\n");
482 fprintf(stderr, " Consider encoding it as Constrained Baseline, Main or High instead.\n");
483 fprintf(stderr, " Decoding might fail and fall back to software.\n");
485 return config->pix_fmt;
488 fprintf(stderr, "Decoder '%s' supports only these pixel formats:", ctx->codec->name);
489 unordered_set<AVPixelFormat> seen;
490 for (const AVPixelFormat *fmt_ptr = fmt; *fmt_ptr != -1; ++fmt_ptr) {
491 if (!seen.count(*fmt_ptr)) {
492 fprintf(stderr, " %s", av_get_pix_fmt_name(*fmt_ptr));
493 seen.insert(*fmt_ptr);
496 fprintf(stderr, " (wanted %s for hardware acceleration)\n", av_get_pix_fmt_name(config->pix_fmt));
500 if (!found_config_of_right_type) {
501 fprintf(stderr, "Decoder '%s' does not support device type '%s'.\n", ctx->codec->name, av_hwdevice_get_type_name(type));
504 // We found no VA-API formats, so take the first software format.
505 for (const AVPixelFormat *fmt_ptr = fmt; *fmt_ptr != -1; ++fmt_ptr) {
506 if ((av_pix_fmt_desc_get(*fmt_ptr)->flags & AV_PIX_FMT_FLAG_HWACCEL) == 0) {
507 fprintf(stderr, "Falling back to software format %s.\n", av_get_pix_fmt_name(*fmt_ptr));
512 // Fallback: Just return anything. (Should never really happen.)
516 bool FFmpegCapture::play_video(const string &pathname)
518 // Note: Call before open, not after; otherwise, there's a race.
519 // (There is now, too, but it tips the correct way. We could use fstat()
520 // if we had the file descriptor.)
521 timespec last_modified;
523 if (stat(pathname.c_str(), &buf) != 0) {
524 // Probably some sort of protocol, so can't stat.
525 last_modified.tv_sec = -1;
527 last_modified = buf.st_mtim;
529 last_colorspace = static_cast<AVColorSpace>(-1);
530 last_chroma_location = static_cast<AVChromaLocation>(-1);
532 AVFormatContextWithCloser format_ctx;
533 if (srt_sock == -1) {
535 format_ctx = avformat_open_input_unique(pathname.c_str(), /*fmt=*/nullptr,
537 AVIOInterruptCB{ &FFmpegCapture::interrupt_cb_thunk, this });
540 // SRT socket, already opened.
541 const AVInputFormat *mpegts_fmt = av_find_input_format("mpegts");
542 format_ctx = avformat_open_input_unique(&FFmpegCapture::read_srt_thunk, this,
543 mpegts_fmt, /*options=*/nullptr,
544 AVIOInterruptCB{ &FFmpegCapture::interrupt_cb_thunk, this });
549 if (format_ctx == nullptr) {
550 fprintf(stderr, "%s: Error opening file\n", pathname.c_str());
554 if (avformat_find_stream_info(format_ctx.get(), nullptr) < 0) {
555 fprintf(stderr, "%s: Error finding stream info\n", pathname.c_str());
559 int video_stream_index = find_stream_index(format_ctx.get(), AVMEDIA_TYPE_VIDEO);
560 if (video_stream_index == -1) {
561 fprintf(stderr, "%s: No video stream found\n", pathname.c_str());
565 int audio_stream_index = find_stream_index(format_ctx.get(), AVMEDIA_TYPE_AUDIO);
566 int subtitle_stream_index = find_stream_index(format_ctx.get(), AVMEDIA_TYPE_SUBTITLE);
567 has_last_subtitle = false;
569 // Open video decoder.
570 const AVCodecParameters *video_codecpar = format_ctx->streams[video_stream_index]->codecpar;
571 const AVCodec *video_codec = avcodec_find_decoder(video_codecpar->codec_id);
573 video_timebase = format_ctx->streams[video_stream_index]->time_base;
574 AVCodecContextWithDeleter video_codec_ctx = avcodec_alloc_context3_unique(nullptr);
575 if (avcodec_parameters_to_context(video_codec_ctx.get(), video_codecpar) < 0) {
576 fprintf(stderr, "%s: Cannot fill video codec parameters\n", pathname.c_str());
579 if (video_codec == nullptr) {
580 fprintf(stderr, "%s: Cannot find video decoder\n", pathname.c_str());
584 // Seemingly, it's not too easy to make something that just initializes
585 // “whatever goes”, so we don't get CUDA or VULKAN or whatever here
586 // without enumerating through several different types.
587 // VA-API and VDPAU will do for now. We prioritize VDPAU for the
588 // simple reason that there's a VA-API-via-VDPAU emulation for NVidia
589 // cards that seems to work, but just hangs when trying to transfer the frame.
591 // Note that we don't actually check codec support beforehand,
592 // so if you have a low-end VDPAU device but a high-end VA-API device,
593 // you lose out on the extra codec support from the latter.
594 AVBufferRef *hw_device_ctx = nullptr;
595 if (av_hwdevice_ctx_create(&hw_device_ctx, AV_HWDEVICE_TYPE_VDPAU, nullptr, nullptr, 0) >= 0) {
596 video_codec_ctx->hw_device_ctx = av_buffer_ref(hw_device_ctx);
597 video_codec_ctx->get_format = get_hw_format<AV_HWDEVICE_TYPE_VDPAU>;
598 } else if (av_hwdevice_ctx_create(&hw_device_ctx, AV_HWDEVICE_TYPE_VAAPI, nullptr, nullptr, 0) >= 0) {
599 video_codec_ctx->hw_device_ctx = av_buffer_ref(hw_device_ctx);
600 video_codec_ctx->get_format = get_hw_format<AV_HWDEVICE_TYPE_VAAPI>;
602 fprintf(stderr, "Failed to initialize VA-API or VDPAU for FFmpeg acceleration. Decoding video in software.\n");
605 if (avcodec_open2(video_codec_ctx.get(), video_codec, nullptr) < 0) {
606 fprintf(stderr, "%s: Cannot open video decoder\n", pathname.c_str());
609 unique_ptr<AVCodecContext, decltype(avcodec_close)*> video_codec_ctx_cleanup(
610 video_codec_ctx.get(), avcodec_close);
612 // Used in decode_ycbcr_format().
613 is_mjpeg = video_codecpar->codec_id == AV_CODEC_ID_MJPEG;
615 // Open audio decoder, if we have audio.
616 AVCodecContextWithDeleter audio_codec_ctx;
617 if (audio_stream_index != -1) {
618 audio_codec_ctx = avcodec_alloc_context3_unique(nullptr);
619 const AVCodecParameters *audio_codecpar = format_ctx->streams[audio_stream_index]->codecpar;
620 audio_timebase = format_ctx->streams[audio_stream_index]->time_base;
621 if (avcodec_parameters_to_context(audio_codec_ctx.get(), audio_codecpar) < 0) {
622 fprintf(stderr, "%s: Cannot fill audio codec parameters\n", pathname.c_str());
625 const AVCodec *audio_codec = avcodec_find_decoder(audio_codecpar->codec_id);
626 if (audio_codec == nullptr) {
627 fprintf(stderr, "%s: Cannot find audio decoder\n", pathname.c_str());
630 if (avcodec_open2(audio_codec_ctx.get(), audio_codec, nullptr) < 0) {
631 fprintf(stderr, "%s: Cannot open audio decoder\n", pathname.c_str());
635 unique_ptr<AVCodecContext, decltype(avcodec_close)*> audio_codec_ctx_cleanup(
636 audio_codec_ctx.get(), avcodec_close);
641 bool first_frame = true;
642 int consecutive_errors = 0;
643 while (!producer_thread_should_quit.should_quit()) {
644 if (process_queued_commands(format_ctx.get(), pathname, last_modified, /*rewound=*/nullptr)) {
647 if (should_interrupt.load()) {
648 // Check as a failsafe, so that we don't need to rely on avio if we don't have to.
651 UniqueFrame audio_frame = audio_frame_allocator->alloc_frame();
652 AudioFormat audio_format;
656 AVFrameWithDeleter frame = decode_frame(format_ctx.get(), video_codec_ctx.get(), audio_codec_ctx.get(),
657 pathname, video_stream_index, audio_stream_index, subtitle_stream_index, audio_frame.get(), &audio_format, &audio_pts, &error);
659 if (++consecutive_errors >= 100) {
660 fprintf(stderr, "More than 100 consecutive video frames, aborting playback.\n");
666 consecutive_errors = 0;
668 if (frame == nullptr) {
669 // EOF. Loop back to the start if we can.
670 if (format_ctx->pb != nullptr && format_ctx->pb->seekable == 0) {
671 // Not seekable (but seemingly, sometimes av_seek_frame() would return 0 anyway,
675 if (av_seek_frame(format_ctx.get(), /*stream_index=*/-1, /*timestamp=*/0, /*flags=*/0) < 0) {
676 fprintf(stderr, "%s: Rewind failed, not looping.\n", pathname.c_str());
679 if (video_codec_ctx != nullptr) {
680 avcodec_flush_buffers(video_codec_ctx.get());
682 if (audio_codec_ctx != nullptr) {
683 avcodec_flush_buffers(audio_codec_ctx.get());
685 // If the file has changed since last time, return to get it reloaded.
686 // Note that depending on how you move the file into place, you might
687 // end up corrupting the one you're already playing, so this path
688 // might not trigger.
689 if (changed_since(pathname, last_modified)) {
696 VideoFormat video_format = construct_video_format(frame.get(), video_timebase);
697 if (video_format.frame_rate_nom == 0 || video_format.frame_rate_den == 0) {
698 // Invalid frame rate; try constructing it from the previous frame length.
699 // (This is especially important if we are the master card, for SRT,
700 // since it affects audio. Not all senders have good timebases
701 // (e.g., Larix rounds first to timebase 1000 and then multiplies by
702 // 90 from there, it seems), but it's much better to have an oscillating
703 // value than just locking at 60.
704 if (last_pts != 0 && frame->pts > last_pts) {
705 int64_t pts_diff = frame->pts - last_pts;
706 video_format.frame_rate_nom = video_timebase.den;
707 video_format.frame_rate_den = video_timebase.num * pts_diff;
709 video_format.frame_rate_nom = 60;
710 video_format.frame_rate_den = 1;
713 UniqueFrame video_frame = make_video_frame(frame.get(), pathname, &error);
719 if (last_pts == 0 && pts_origin == 0) {
720 pts_origin = frame->pts;
722 steady_clock::time_point now = steady_clock::now();
723 if (play_as_fast_as_possible) {
724 video_frame->received_timestamp = now;
725 audio_frame->received_timestamp = now;
726 next_frame_start = now;
728 next_frame_start = compute_frame_start(frame->pts, pts_origin, video_timebase, start, rate);
729 if (first_frame && last_frame_was_connected) {
730 // If reconnect took more than one second, this is probably a live feed,
731 // and we should reset the resampler. (Or the rate is really, really low,
732 // in which case a reset on the first frame is fine anyway.)
733 if (duration<double>(next_frame_start - last_frame).count() >= 1.0) {
734 last_frame_was_connected = false;
737 video_frame->received_timestamp = next_frame_start;
739 // The easiest way to get all the rate conversions etc. right is to move the
740 // audio PTS into the video PTS timebase and go from there. (We'll get some
741 // rounding issues, but they should not be a big problem.)
742 int64_t audio_pts_as_video_pts = av_rescale_q(audio_pts, audio_timebase, video_timebase);
743 audio_frame->received_timestamp = compute_frame_start(audio_pts_as_video_pts, pts_origin, video_timebase, start, rate);
745 if (audio_frame->len != 0) {
746 // The received timestamps in Nageru are measured after we've just received the frame.
747 // However, pts (especially audio pts) is at the _beginning_ of the frame.
748 // If we have locked audio, the distinction doesn't really matter, as pts is
749 // on a relative scale and a fixed offset is fine. But if we don't, we will have
750 // a different number of samples each time, which will cause huge audio jitter
751 // and throw off the resampler.
753 // In a sense, we should have compensated by adding the frame and audio lengths
754 // to video_frame->received_timestamp and audio_frame->received_timestamp respectively,
755 // but that would mean extra waiting in sleep_until(). All we need is that they
756 // are correct relative to each other, though (and to the other frames we send),
757 // so just align the end of the audio frame, and we're fine.
758 size_t num_samples = (audio_frame->len * 8) / audio_format.bits_per_sample / audio_format.num_channels;
759 double offset = double(num_samples) / OUTPUT_FREQUENCY -
760 double(video_format.frame_rate_den) / video_format.frame_rate_nom;
761 audio_frame->received_timestamp += duration_cast<steady_clock::duration>(duration<double>(offset));
764 if (duration<double>(now - next_frame_start).count() >= 0.1) {
765 // If we don't have enough CPU to keep up, or if we have a live stream
766 // where the initial origin was somehow wrong, we could be behind indefinitely.
767 // In particular, this will give the audio resampler problems as it tries
768 // to speed up to reduce the delay, hitting the low end of the buffer every time.
769 fprintf(stderr, "%s: Playback %.0f ms behind, resetting time scale\n",
771 1e3 * duration<double>(now - next_frame_start).count());
772 pts_origin = frame->pts;
773 start = next_frame_start = now;
774 timecode += TYPICAL_FPS * 2 + 1;
777 bool finished_wakeup;
778 if (play_as_fast_as_possible) {
779 finished_wakeup = !producer_thread_should_quit.should_quit();
781 finished_wakeup = producer_thread_should_quit.sleep_until(next_frame_start);
783 if (finished_wakeup) {
784 if (audio_frame->len > 0) {
785 assert(audio_pts != -1);
787 if (!last_frame_was_connected) {
788 // We're recovering from an error (or really slow load, see above).
789 // Make sure to get the audio resampler reset. (This is a hack;
790 // ideally, the frame callback should just accept a way to signal
791 // audio discontinuity.)
792 timecode += TYPICAL_FPS * 2 + 1;
794 last_neutral_color = get_neutral_color(frame->metadata);
795 if (frame_callback != nullptr) {
796 frame_callback(frame->pts, video_timebase, audio_pts, audio_timebase, timecode++,
797 video_frame.get_and_release(), 0, video_format,
798 audio_frame.get_and_release(), 0, audio_format);
801 last_frame = steady_clock::now();
802 last_frame_was_connected = true;
805 if (producer_thread_should_quit.should_quit()) break;
807 bool rewound = false;
808 if (process_queued_commands(format_ctx.get(), pathname, last_modified, &rewound)) {
811 // If we just rewound, drop this frame on the floor and be done.
815 // OK, we didn't, so probably a rate change. Recalculate next_frame_start,
816 // but if it's now in the past, we'll reset the origin, so that we don't
817 // generate a huge backlog of frames that we need to run through quickly.
818 next_frame_start = compute_frame_start(frame->pts, pts_origin, video_timebase, start, rate);
819 steady_clock::time_point now = steady_clock::now();
820 if (next_frame_start < now) {
821 pts_origin = frame->pts;
822 start = next_frame_start = now;
826 last_pts = frame->pts;
831 void FFmpegCapture::internal_rewind()
833 pts_origin = last_pts = 0;
834 start = next_frame_start = steady_clock::now();
837 bool FFmpegCapture::process_queued_commands(AVFormatContext *format_ctx, const std::string &pathname, timespec last_modified, bool *rewound)
839 // Process any queued commands from other threads.
840 vector<QueuedCommand> commands;
842 lock_guard<mutex> lock(queue_mu);
843 swap(commands, command_queue);
845 for (const QueuedCommand &cmd : commands) {
846 switch (cmd.command) {
847 case QueuedCommand::REWIND:
848 if (av_seek_frame(format_ctx, /*stream_index=*/-1, /*timestamp=*/0, /*flags=*/0) < 0) {
849 fprintf(stderr, "%s: Rewind failed, stopping play.\n", pathname.c_str());
851 // If the file has changed since last time, return to get it reloaded.
852 // Note that depending on how you move the file into place, you might
853 // end up corrupting the one you're already playing, so this path
854 // might not trigger.
855 if (changed_since(pathname, last_modified)) {
859 if (rewound != nullptr) {
864 case QueuedCommand::CHANGE_RATE:
865 // Change the origin to the last played frame.
866 start = compute_frame_start(last_pts, pts_origin, video_timebase, start, rate);
867 pts_origin = last_pts;
869 play_as_fast_as_possible = (rate >= 10.0);
880 AVFrameWithDeleter FFmpegCapture::decode_frame(AVFormatContext *format_ctx, AVCodecContext *video_codec_ctx, AVCodecContext *audio_codec_ctx,
881 const std::string &pathname, int video_stream_index, int audio_stream_index, int subtitle_stream_index,
882 FrameAllocator::Frame *audio_frame, AudioFormat *audio_format, int64_t *audio_pts, bool *error)
886 // Read packets until we have a frame or there are none left.
887 bool frame_finished = false;
888 AVFrameWithDeleter audio_avframe = av_frame_alloc_unique();
889 AVFrameWithDeleter video_avframe = av_frame_alloc_unique();
892 bool has_audio = false;
895 unique_ptr<AVPacket, decltype(av_packet_unref)*> pkt_cleanup(
896 &pkt, av_packet_unref);
897 av_init_packet(&pkt);
900 if (av_read_frame(format_ctx, &pkt) == 0) {
901 if (pkt.stream_index == audio_stream_index && audio_callback != nullptr) {
902 audio_callback(&pkt, format_ctx->streams[audio_stream_index]->time_base);
904 if (pkt.stream_index == video_stream_index && video_callback != nullptr) {
905 video_callback(&pkt, format_ctx->streams[video_stream_index]->time_base);
907 if (pkt.stream_index == video_stream_index && global_flags.transcode_video) {
908 if (avcodec_send_packet(video_codec_ctx, &pkt) < 0) {
909 fprintf(stderr, "%s: Cannot send packet to video codec.\n", pathname.c_str());
911 return AVFrameWithDeleter(nullptr);
913 } else if (pkt.stream_index == audio_stream_index && global_flags.transcode_audio) {
915 if (avcodec_send_packet(audio_codec_ctx, &pkt) < 0) {
916 fprintf(stderr, "%s: Cannot send packet to audio codec.\n", pathname.c_str());
918 return AVFrameWithDeleter(nullptr);
920 } else if (pkt.stream_index == subtitle_stream_index) {
921 last_subtitle = string(reinterpret_cast<const char *>(pkt.data), pkt.size);
922 has_last_subtitle = true;
925 eof = true; // Or error, but ignore that for the time being.
928 // Decode audio, if any.
931 int err = avcodec_receive_frame(audio_codec_ctx, audio_avframe.get());
933 if (*audio_pts == -1) {
934 *audio_pts = audio_avframe->pts;
936 convert_audio(audio_avframe.get(), audio_frame, audio_format);
937 } else if (err == AVERROR(EAGAIN)) {
940 fprintf(stderr, "%s: Cannot receive frame from audio codec.\n", pathname.c_str());
942 return AVFrameWithDeleter(nullptr);
947 // Decode video, if we have a frame.
948 int err = avcodec_receive_frame(video_codec_ctx, video_avframe.get());
950 if (video_avframe->format == AV_PIX_FMT_VAAPI ||
951 video_avframe->format == AV_PIX_FMT_VDPAU) {
952 // Get the frame down to the CPU. (TODO: See if we can keep it
953 // on the GPU all the way, since it will be going up again later.
954 // However, this only works if the OpenGL GPU is the same one.)
955 AVFrameWithDeleter sw_frame = av_frame_alloc_unique();
956 int err = av_hwframe_transfer_data(sw_frame.get(), video_avframe.get(), 0);
958 fprintf(stderr, "%s: Cannot transfer hardware video frame to software.\n", pathname.c_str());
960 return AVFrameWithDeleter(nullptr);
962 sw_frame->pts = video_avframe->pts;
963 sw_frame->pkt_duration = video_avframe->pkt_duration;
964 video_avframe = move(sw_frame);
966 frame_finished = true;
968 } else if (err != AVERROR(EAGAIN)) {
969 fprintf(stderr, "%s: Cannot receive frame from video codec.\n", pathname.c_str());
971 return AVFrameWithDeleter(nullptr);
976 return video_avframe;
978 return AVFrameWithDeleter(nullptr);
981 void FFmpegCapture::convert_audio(const AVFrame *audio_avframe, FrameAllocator::Frame *audio_frame, AudioFormat *audio_format)
983 // Decide on a format. If there already is one in this audio frame,
984 // we're pretty much forced to use it. If not, we try to find an exact match.
985 // If that still doesn't work, we default to 32-bit signed chunked
986 // (float would be nice, but there's really no way to signal that yet).
987 AVSampleFormat dst_format;
988 if (audio_format->bits_per_sample == 0) {
989 switch (audio_avframe->format) {
990 case AV_SAMPLE_FMT_S16:
991 case AV_SAMPLE_FMT_S16P:
992 audio_format->bits_per_sample = 16;
993 dst_format = AV_SAMPLE_FMT_S16;
995 case AV_SAMPLE_FMT_S32:
996 case AV_SAMPLE_FMT_S32P:
998 audio_format->bits_per_sample = 32;
999 dst_format = AV_SAMPLE_FMT_S32;
1002 } else if (audio_format->bits_per_sample == 16) {
1003 dst_format = AV_SAMPLE_FMT_S16;
1004 } else if (audio_format->bits_per_sample == 32) {
1005 dst_format = AV_SAMPLE_FMT_S32;
1009 audio_format->num_channels = 2;
1011 int64_t channel_layout = audio_avframe->channel_layout;
1012 if (channel_layout == 0) {
1013 channel_layout = av_get_default_channel_layout(audio_avframe->channels);
1016 if (resampler == nullptr ||
1017 audio_avframe->format != last_src_format ||
1018 dst_format != last_dst_format ||
1019 channel_layout != last_channel_layout ||
1020 audio_avframe->sample_rate != last_sample_rate) {
1021 swr_free(&resampler);
1022 resampler = swr_alloc_set_opts(nullptr,
1023 /*out_ch_layout=*/AV_CH_LAYOUT_STEREO_DOWNMIX,
1024 /*out_sample_fmt=*/dst_format,
1025 /*out_sample_rate=*/OUTPUT_FREQUENCY,
1026 /*in_ch_layout=*/channel_layout,
1027 /*in_sample_fmt=*/AVSampleFormat(audio_avframe->format),
1028 /*in_sample_rate=*/audio_avframe->sample_rate,
1030 /*log_ctx=*/nullptr);
1032 if (resampler == nullptr) {
1033 fprintf(stderr, "Allocating resampler failed.\n");
1037 if (swr_init(resampler) < 0) {
1038 fprintf(stderr, "Could not open resample context.\n");
1042 last_src_format = AVSampleFormat(audio_avframe->format);
1043 last_dst_format = dst_format;
1044 last_channel_layout = channel_layout;
1045 last_sample_rate = audio_avframe->sample_rate;
1048 size_t bytes_per_sample = (audio_format->bits_per_sample / 8) * 2;
1049 size_t num_samples_room = (audio_frame->size - audio_frame->len) / bytes_per_sample;
1051 uint8_t *data = audio_frame->data + audio_frame->len;
1052 int out_samples = swr_convert(resampler, &data, num_samples_room,
1053 const_cast<const uint8_t **>(audio_avframe->data), audio_avframe->nb_samples);
1054 if (out_samples < 0) {
1055 fprintf(stderr, "Audio conversion failed.\n");
1059 audio_frame->len += out_samples * bytes_per_sample;
1062 VideoFormat FFmpegCapture::construct_video_format(const AVFrame *frame, AVRational video_timebase)
1064 VideoFormat video_format;
1065 video_format.width = frame_width(frame);
1066 video_format.height = frame_height(frame);
1067 if (pixel_format == bmusb::PixelFormat_8BitBGRA) {
1068 video_format.stride = frame_width(frame) * 4;
1069 } else if (pixel_format == FFmpegCapture::PixelFormat_NV12) {
1070 video_format.stride = frame_width(frame);
1072 assert(pixel_format == bmusb::PixelFormat_8BitYCbCrPlanar);
1073 video_format.stride = frame_width(frame);
1075 video_format.frame_rate_nom = video_timebase.den;
1076 video_format.frame_rate_den = frame->pkt_duration * video_timebase.num;
1077 video_format.has_signal = true;
1078 video_format.is_connected = true;
1079 return video_format;
1082 UniqueFrame FFmpegCapture::make_video_frame(const AVFrame *frame, const string &pathname, bool *error)
1086 UniqueFrame video_frame(video_frame_allocator->alloc_frame());
1087 if (video_frame->data == nullptr) {
1091 if (sws_ctx == nullptr ||
1092 sws_last_width != frame->width ||
1093 sws_last_height != frame->height ||
1094 sws_last_src_format != frame->format) {
1095 sws_dst_format = decide_dst_format(AVPixelFormat(frame->format), pixel_format);
1097 sws_getContext(frame->width, frame->height, AVPixelFormat(frame->format),
1098 frame_width(frame), frame_height(frame), sws_dst_format,
1099 SWS_BICUBIC, nullptr, nullptr, nullptr));
1100 sws_last_width = frame->width;
1101 sws_last_height = frame->height;
1102 sws_last_src_format = frame->format;
1104 if (sws_ctx == nullptr) {
1105 fprintf(stderr, "%s: Could not create scaler context\n", pathname.c_str());
1110 uint8_t *pic_data[4] = { nullptr, nullptr, nullptr, nullptr };
1111 int linesizes[4] = { 0, 0, 0, 0 };
1112 if (pixel_format == bmusb::PixelFormat_8BitBGRA) {
1113 pic_data[0] = video_frame->data;
1114 linesizes[0] = frame_width(frame) * 4;
1115 video_frame->len = (frame_width(frame) * 4) * frame_height(frame);
1116 } else if (pixel_format == PixelFormat_NV12) {
1117 pic_data[0] = video_frame->data;
1118 linesizes[0] = frame_width(frame);
1120 pic_data[1] = pic_data[0] + frame_width(frame) * frame_height(frame);
1121 linesizes[1] = frame_width(frame);
1123 video_frame->len = (frame_width(frame) * 2) * frame_height(frame);
1125 const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(sws_dst_format);
1126 current_frame_ycbcr_format = decode_ycbcr_format(desc, frame, is_mjpeg, &last_colorspace, &last_chroma_location);
1128 assert(pixel_format == bmusb::PixelFormat_8BitYCbCrPlanar);
1129 const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(sws_dst_format);
1131 int chroma_width = AV_CEIL_RSHIFT(int(frame_width(frame)), desc->log2_chroma_w);
1132 int chroma_height = AV_CEIL_RSHIFT(int(frame_height(frame)), desc->log2_chroma_h);
1134 pic_data[0] = video_frame->data;
1135 linesizes[0] = frame_width(frame);
1137 pic_data[1] = pic_data[0] + frame_width(frame) * frame_height(frame);
1138 linesizes[1] = chroma_width;
1140 pic_data[2] = pic_data[1] + chroma_width * chroma_height;
1141 linesizes[2] = chroma_width;
1143 video_frame->len = frame_width(frame) * frame_height(frame) + 2 * chroma_width * chroma_height;
1145 current_frame_ycbcr_format = decode_ycbcr_format(desc, frame, is_mjpeg, &last_colorspace, &last_chroma_location);
1148 // FIXME: Currently, if the video is too high-res for one of the allocated
1149 // frames, we simply refuse to scale it here to avoid crashes. It would be better
1150 // if we could somehow signal getting larger frames, especially as 4K is a thing now.
1151 if (video_frame->len > FRAME_SIZE) {
1152 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);
1157 sws_scale(sws_ctx.get(), frame->data, frame->linesize, 0, frame->height, pic_data, linesizes);
1162 int FFmpegCapture::interrupt_cb_thunk(void *opaque)
1164 return reinterpret_cast<FFmpegCapture *>(opaque)->interrupt_cb();
1167 int FFmpegCapture::interrupt_cb()
1169 return should_interrupt.load();
1172 unsigned FFmpegCapture::frame_width(const AVFrame *frame) const
1175 return frame->width;
1181 unsigned FFmpegCapture::frame_height(const AVFrame *frame) const
1184 return frame->height;
1191 int FFmpegCapture::read_srt_thunk(void *opaque, uint8_t *buf, int buf_size)
1193 return reinterpret_cast<FFmpegCapture *>(opaque)->read_srt(buf, buf_size);
1196 int FFmpegCapture::read_srt(uint8_t *buf, int buf_size)
1198 SRT_MSGCTRL mc = srt_msgctrl_default;
1199 return srt_recvmsg2(srt_sock, reinterpret_cast<char *>(buf), buf_size, &mc);