1 #include "ffmpeg_capture.h"
3 #include "shared/shared_defs.h"
11 #include <movit/effect.h>
12 #include <movit/image_format.h>
13 #include <movit/ycbcr.h>
25 #include <libavcodec/avcodec.h>
26 #include <libavcodec/codec.h>
27 #include <libavcodec/codec_id.h>
28 #include <libavcodec/codec_par.h>
29 #include <libavformat/avformat.h>
30 #include <libavutil/avutil.h>
31 #include <libavutil/buffer.h>
32 #include <libavutil/channel_layout.h>
33 #include <libavutil/common.h>
34 #include <libavutil/dict.h>
35 #include <libavutil/error.h>
36 #include <libavutil/frame.h>
37 #include <libavutil/hwcontext.h>
38 #include <libavutil/mathematics.h>
39 #include <libavutil/pixdesc.h>
40 #include <libavutil/pixfmt.h>
41 #include <libavutil/rational.h>
42 #include <libavutil/samplefmt.h>
43 #include <libavutil/version.h>
44 #include <libswresample/swresample.h>
45 #include <libswscale/swscale.h>
52 #include <unordered_set>
56 #include <movit/colorspace_conversion_effect.h>
58 #include "bmusb/bmusb.h"
59 #include "shared/ffmpeg_raii.h"
60 #include "ffmpeg_util.h"
62 #include "ref_counted_frame.h"
63 #include "shared/timebase.h"
70 using namespace std::chrono;
71 using namespace bmusb;
72 using namespace movit;
73 using namespace Eigen;
75 // Avoid deprecation warnings, but we don't want to drop FFmpeg 5.1 support just yet.
76 #if LIBAVUTIL_VERSION_INT >= AV_VERSION_INT(57, 30, 100)
77 #define pkt_duration duration
82 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)
84 const duration<double> pts((frame_pts - pts_origin) * double(video_timebase.num) / double(video_timebase.den));
85 return origin + duration_cast<steady_clock::duration>(pts / rate);
88 bool changed_since(const std::string &pathname, const timespec &ts)
94 if (stat(pathname.c_str(), &buf) != 0) {
95 fprintf(stderr, "%s: Couldn't check for new version, leaving the old in place.\n", pathname.c_str());
98 return (buf.st_mtim.tv_sec != ts.tv_sec || buf.st_mtim.tv_nsec != ts.tv_nsec);
101 bool is_full_range(const AVPixFmtDescriptor *desc)
103 // This is horrible, but there's no better way that I know of.
104 return (strchr(desc->name, 'j') != nullptr);
107 AVPixelFormat decide_dst_format(AVPixelFormat src_format, bmusb::PixelFormat dst_format_type)
109 if (dst_format_type == bmusb::PixelFormat_8BitBGRA) {
110 return AV_PIX_FMT_BGRA;
112 if (dst_format_type == FFmpegCapture::PixelFormat_NV12) {
113 return AV_PIX_FMT_NV12;
116 assert(dst_format_type == bmusb::PixelFormat_8BitYCbCrPlanar);
118 // If this is a non-Y'CbCr format, just convert to 4:4:4 Y'CbCr
119 // and be done with it. It's too strange to spend a lot of time on.
120 // (Let's hope there's no alpha.)
121 const AVPixFmtDescriptor *src_desc = av_pix_fmt_desc_get(src_format);
122 if (src_desc == nullptr ||
123 src_desc->nb_components != 3 ||
124 (src_desc->flags & AV_PIX_FMT_FLAG_RGB)) {
125 return AV_PIX_FMT_YUV444P;
128 // The best for us would be Cb and Cr together if possible,
129 // but FFmpeg doesn't support that except in the special case of
130 // NV12, so we need to go to planar even for the case of NV12.
131 // Thus, look for the closest (but no worse) 8-bit planar Y'CbCr format
132 // that matches in color range. (This will also include the case of
133 // the source format already being acceptable.)
134 bool src_full_range = is_full_range(src_desc);
135 const char *best_format = "yuv444p";
136 unsigned best_score = numeric_limits<unsigned>::max();
137 for (const AVPixFmtDescriptor *desc = av_pix_fmt_desc_next(nullptr);
139 desc = av_pix_fmt_desc_next(desc)) {
140 // Find planar Y'CbCr formats only.
141 if (desc->nb_components != 3) continue;
142 if (desc->flags & AV_PIX_FMT_FLAG_RGB) continue;
143 if (!(desc->flags & AV_PIX_FMT_FLAG_PLANAR)) continue;
144 if (desc->comp[0].plane != 0 ||
145 desc->comp[1].plane != 1 ||
146 desc->comp[2].plane != 2) continue;
148 // 8-bit formats only.
149 if (desc->flags & AV_PIX_FMT_FLAG_BE) continue;
150 if (desc->comp[0].depth != 8) continue;
152 // Same or better chroma resolution only.
153 int chroma_w_diff = src_desc->log2_chroma_w - desc->log2_chroma_w;
154 int chroma_h_diff = src_desc->log2_chroma_h - desc->log2_chroma_h;
155 if (chroma_w_diff < 0 || chroma_h_diff < 0)
158 // Matching full/limited range only.
159 if (is_full_range(desc) != src_full_range)
162 // Pick something with as little excess chroma resolution as possible.
163 unsigned score = (1 << (chroma_w_diff)) << chroma_h_diff;
164 if (score < best_score) {
166 best_format = desc->name;
169 return av_get_pix_fmt(best_format);
172 YCbCrFormat decode_ycbcr_format(const AVPixFmtDescriptor *desc, const AVFrame *frame, bool is_mjpeg, AVColorSpace *last_colorspace, AVChromaLocation *last_chroma_location)
175 AVColorSpace colorspace = frame->colorspace;
176 switch (colorspace) {
177 case AVCOL_SPC_BT709:
178 format.luma_coefficients = YCBCR_REC_709;
180 case AVCOL_SPC_BT470BG:
181 case AVCOL_SPC_SMPTE170M:
182 case AVCOL_SPC_SMPTE240M:
183 format.luma_coefficients = YCBCR_REC_601;
185 case AVCOL_SPC_BT2020_NCL:
186 format.luma_coefficients = YCBCR_REC_2020;
188 case AVCOL_SPC_UNSPECIFIED:
189 format.luma_coefficients = (frame->height >= 720 ? YCBCR_REC_709 : YCBCR_REC_601);
192 if (colorspace != *last_colorspace) {
193 fprintf(stderr, "Unknown Y'CbCr coefficient enum %d from FFmpeg; choosing Rec. 709.\n",
196 format.luma_coefficients = YCBCR_REC_709;
199 *last_colorspace = colorspace;
201 format.full_range = is_full_range(desc);
202 format.num_levels = 1 << desc->comp[0].depth;
203 format.chroma_subsampling_x = 1 << desc->log2_chroma_w;
204 format.chroma_subsampling_y = 1 << desc->log2_chroma_h;
206 switch (frame->chroma_location) {
207 case AVCHROMA_LOC_LEFT:
208 format.cb_x_position = 0.0;
209 format.cb_y_position = 0.5;
211 case AVCHROMA_LOC_CENTER:
212 format.cb_x_position = 0.5;
213 format.cb_y_position = 0.5;
215 case AVCHROMA_LOC_TOPLEFT:
216 format.cb_x_position = 0.0;
217 format.cb_y_position = 0.0;
219 case AVCHROMA_LOC_TOP:
220 format.cb_x_position = 0.5;
221 format.cb_y_position = 0.0;
223 case AVCHROMA_LOC_BOTTOMLEFT:
224 format.cb_x_position = 0.0;
225 format.cb_y_position = 1.0;
227 case AVCHROMA_LOC_BOTTOM:
228 format.cb_x_position = 0.5;
229 format.cb_y_position = 1.0;
232 if (frame->chroma_location != *last_chroma_location) {
233 fprintf(stderr, "Unknown chroma location coefficient enum %d from FFmpeg; choosing center.\n",
234 frame->chroma_location);
236 format.cb_x_position = 0.5;
237 format.cb_y_position = 0.5;
240 *last_chroma_location = frame->chroma_location;
242 if (is_mjpeg && !format.full_range) {
243 // Limited-range MJPEG is only detected by FFmpeg whenever a special
244 // JPEG comment is set, which means that in practice, the stream is
245 // almost certainly generated by Futatabi. Override FFmpeg's forced
246 // MJPEG defaults (it disregards the values set in the mux) with what
248 format.luma_coefficients = YCBCR_REC_709;
249 format.cb_x_position = 0.0;
250 format.cb_y_position = 0.5;
253 format.cr_x_position = format.cb_x_position;
254 format.cr_y_position = format.cb_y_position;
258 RGBTriplet get_neutral_color(AVDictionary *metadata)
260 if (metadata == nullptr) {
261 return RGBTriplet(1.0f, 1.0f, 1.0f);
263 AVDictionaryEntry *entry = av_dict_get(metadata, "WhitePoint", nullptr, 0);
264 if (entry == nullptr) {
265 return RGBTriplet(1.0f, 1.0f, 1.0f);
268 unsigned x_nom, x_den, y_nom, y_den;
269 if (sscanf(entry->value, " %u:%u , %u:%u", &x_nom, &x_den, &y_nom, &y_den) != 4) {
270 fprintf(stderr, "WARNING: Unable to parse white point '%s', using default white point\n", entry->value);
271 return RGBTriplet(1.0f, 1.0f, 1.0f);
274 double x = double(x_nom) / x_den;
275 double y = double(y_nom) / y_den;
276 double z = 1.0 - x - y;
278 Matrix3d rgb_to_xyz_matrix = movit::ColorspaceConversionEffect::get_xyz_matrix(COLORSPACE_sRGB);
279 Vector3d rgb = rgb_to_xyz_matrix.inverse() * Vector3d(x, y, z);
281 return RGBTriplet(rgb[0], rgb[1], rgb[2]);
286 FFmpegCapture::FFmpegCapture(const string &filename, unsigned width, unsigned height)
287 : filename(filename), width(width), height(height), video_timebase{1, 1}
289 description = "Video: " + filename;
291 last_frame = steady_clock::now();
293 avformat_network_init(); // In case someone wants this.
297 FFmpegCapture::FFmpegCapture(int srt_sock, const string &stream_id)
298 : srt_sock(srt_sock),
299 width(0), // Don't resize; SRT streams typically have stable resolution, and should behave much like regular cards in general.
301 pixel_format(bmusb::PixelFormat_8BitYCbCrPlanar),
304 if (stream_id.empty()) {
305 description = "SRT stream";
307 description = stream_id;
309 play_as_fast_as_possible = true;
311 last_frame = steady_clock::now();
315 FFmpegCapture::~FFmpegCapture()
317 if (has_dequeue_callbacks) {
318 dequeue_cleanup_callback();
320 swr_free(&resampler);
322 if (srt_sock != -1) {
328 void FFmpegCapture::configure_card()
330 if (video_frame_allocator == nullptr) {
331 owned_video_frame_allocator.reset(new MallocFrameAllocator(FRAME_SIZE, NUM_QUEUED_VIDEO_FRAMES));
332 set_video_frame_allocator(owned_video_frame_allocator.get());
334 if (audio_frame_allocator == nullptr) {
335 // Audio can come out in pretty large chunks, so increase from the default 1 MB.
336 owned_audio_frame_allocator.reset(new MallocFrameAllocator(1 << 20, NUM_QUEUED_AUDIO_FRAMES));
337 set_audio_frame_allocator(owned_audio_frame_allocator.get());
341 void FFmpegCapture::start_bm_capture()
347 producer_thread_should_quit.unquit();
348 producer_thread = thread(&FFmpegCapture::producer_thread_func, this);
351 void FFmpegCapture::stop_dequeue_thread()
357 producer_thread_should_quit.quit();
358 producer_thread.join();
361 std::map<uint32_t, VideoMode> FFmpegCapture::get_available_video_modes() const
363 // Note: This will never really be shown in the UI.
367 snprintf(buf, sizeof(buf), "%ux%u", sws_last_width, sws_last_height);
370 mode.autodetect = false;
371 mode.width = sws_last_width;
372 mode.height = sws_last_height;
373 mode.frame_rate_num = 60;
374 mode.frame_rate_den = 1;
375 mode.interlaced = false;
377 return {{ 0, mode }};
380 void FFmpegCapture::producer_thread_func()
382 char thread_name[16];
383 snprintf(thread_name, sizeof(thread_name), "FFmpeg_C_%d", card_index);
384 pthread_setname_np(pthread_self(), thread_name);
386 while (!producer_thread_should_quit.should_quit()) {
387 string filename_copy;
389 lock_guard<mutex> lock(filename_mu);
390 filename_copy = filename;
394 if (srt_sock == -1) {
395 pathname = search_for_file(filename_copy);
397 pathname = description;
399 if (pathname.empty()) {
400 send_disconnected_frame();
404 producer_thread_should_quit.sleep_for(seconds(1));
405 fprintf(stderr, "%s not found, sleeping one second and trying again...\n", filename_copy.c_str());
408 should_interrupt = false;
409 if (!play_video(pathname)) {
411 send_disconnected_frame();
415 fprintf(stderr, "Error when playing %s, sleeping one second and trying again...\n", pathname.c_str());
416 producer_thread_should_quit.sleep_for(seconds(1));
421 send_disconnected_frame();
425 // Probably just EOF, will exit the loop above on next test.
428 if (has_dequeue_callbacks) {
429 dequeue_cleanup_callback();
430 has_dequeue_callbacks = false;
434 void FFmpegCapture::send_disconnected_frame()
436 // Send an empty frame to signal that we have no signal anymore.
437 FrameAllocator::Frame video_frame = video_frame_allocator->alloc_frame();
438 size_t frame_width = width == 0 ? global_flags.width : width;
439 size_t frame_height = height == 0 ? global_flags.height : height;
440 if (video_frame.data) {
441 VideoFormat video_format;
442 video_format.width = frame_width;
443 video_format.height = frame_height;
444 video_format.frame_rate_nom = 60;
445 video_format.frame_rate_den = 1;
446 video_format.is_connected = false;
447 if (pixel_format == bmusb::PixelFormat_8BitBGRA) {
448 video_format.stride = frame_width * 4;
449 video_frame.len = frame_width * frame_height * 4;
450 memset(video_frame.data, 0, video_frame.len);
452 video_format.stride = frame_width;
453 current_frame_ycbcr_format.luma_coefficients = YCBCR_REC_709;
454 current_frame_ycbcr_format.full_range = true;
455 current_frame_ycbcr_format.num_levels = 256;
456 current_frame_ycbcr_format.chroma_subsampling_x = 2;
457 current_frame_ycbcr_format.chroma_subsampling_y = 2;
458 current_frame_ycbcr_format.cb_x_position = 0.0f;
459 current_frame_ycbcr_format.cb_y_position = 0.0f;
460 current_frame_ycbcr_format.cr_x_position = 0.0f;
461 current_frame_ycbcr_format.cr_y_position = 0.0f;
462 video_frame.len = frame_width * frame_height * 2;
463 memset(video_frame.data, 0, frame_width * frame_height);
464 memset(video_frame.data + frame_width * frame_height, 128, frame_width * frame_height); // Valid for both NV12 and planar.
467 if (frame_callback != nullptr) {
468 frame_callback(-1, AVRational{1, TIMEBASE}, -1, AVRational{1, TIMEBASE}, timecode++,
469 video_frame, /*video_offset=*/0, video_format,
470 FrameAllocator::Frame(), /*audio_offset=*/0, AudioFormat());
472 last_frame_was_connected = false;
477 if (card_disconnected_callback != nullptr) {
478 card_disconnected_callback();
483 template<AVHWDeviceType type>
484 AVPixelFormat get_hw_format(AVCodecContext *ctx, const AVPixelFormat *fmt)
486 bool found_config_of_right_type = false;
487 for (int i = 0;; ++i) { // Termination condition inside loop.
488 const AVCodecHWConfig *config = avcodec_get_hw_config(ctx->codec, i);
489 if (config == nullptr) { // End of list.
492 if (!(config->methods & AV_CODEC_HW_CONFIG_METHOD_HW_DEVICE_CTX) ||
493 config->device_type != type) {
494 // Not interesting for us.
498 // We have a config of the right type, but does it actually support
499 // the pixel format we want? (Seemingly, FFmpeg's way of signaling errors
500 // is to just replace the pixel format with a software-decoded one,
502 found_config_of_right_type = true;
503 for (const AVPixelFormat *fmt_ptr = fmt; *fmt_ptr != -1; ++fmt_ptr) {
504 if (config->pix_fmt == *fmt_ptr) {
505 fprintf(stderr, "Initialized '%s' hardware decoding for codec '%s'.\n",
506 av_hwdevice_get_type_name(type), ctx->codec->name);
507 if (ctx->profile == FF_PROFILE_H264_BASELINE) {
508 fprintf(stderr, "WARNING: Stream claims to be H.264 Baseline, which is generally poorly supported in hardware decoders.\n");
509 fprintf(stderr, " Consider encoding it as Constrained Baseline, Main or High instead.\n");
510 fprintf(stderr, " Decoding might fail and fall back to software.\n");
512 return config->pix_fmt;
515 fprintf(stderr, "Decoder '%s' supports only these pixel formats:", ctx->codec->name);
516 unordered_set<AVPixelFormat> seen;
517 for (const AVPixelFormat *fmt_ptr = fmt; *fmt_ptr != -1; ++fmt_ptr) {
518 if (!seen.count(*fmt_ptr)) {
519 fprintf(stderr, " %s", av_get_pix_fmt_name(*fmt_ptr));
520 seen.insert(*fmt_ptr);
523 fprintf(stderr, " (wanted %s for hardware acceleration)\n", av_get_pix_fmt_name(config->pix_fmt));
527 if (!found_config_of_right_type) {
528 fprintf(stderr, "Decoder '%s' does not support device type '%s'.\n", ctx->codec->name, av_hwdevice_get_type_name(type));
531 // We found no VA-API formats, so take the first software format.
532 for (const AVPixelFormat *fmt_ptr = fmt; *fmt_ptr != -1; ++fmt_ptr) {
533 if ((av_pix_fmt_desc_get(*fmt_ptr)->flags & AV_PIX_FMT_FLAG_HWACCEL) == 0) {
534 fprintf(stderr, "Falling back to software format %s.\n", av_get_pix_fmt_name(*fmt_ptr));
539 // Fallback: Just return anything. (Should never really happen.)
543 bool FFmpegCapture::play_video(const string &pathname)
545 // Note: Call before open, not after; otherwise, there's a race.
546 // (There is now, too, but it tips the correct way. We could use fstat()
547 // if we had the file descriptor.)
548 timespec last_modified;
550 if (stat(pathname.c_str(), &buf) != 0) {
551 // Probably some sort of protocol, so can't stat.
552 last_modified.tv_sec = -1;
554 last_modified = buf.st_mtim;
556 last_colorspace = static_cast<AVColorSpace>(-1);
557 last_chroma_location = static_cast<AVChromaLocation>(-1);
559 AVFormatContextWithCloser format_ctx;
560 if (srt_sock == -1) {
561 // Regular file (or stream).
562 frame_timeout_started = steady_clock::now();
563 frame_timeout_valid = true;
564 format_ctx = avformat_open_input_unique(pathname.c_str(), /*fmt=*/nullptr,
566 AVIOInterruptCB{ &FFmpegCapture::interrupt_cb_thunk, this });
567 frame_timeout_valid = false;
570 // SRT socket, already opened.
571 const AVInputFormat *mpegts_fmt = av_find_input_format("mpegts");
572 format_ctx = avformat_open_input_unique(&FFmpegCapture::read_srt_thunk, this,
573 mpegts_fmt, /*options=*/nullptr,
574 AVIOInterruptCB{ &FFmpegCapture::interrupt_cb_thunk, this });
579 if (format_ctx == nullptr) {
580 fprintf(stderr, "%s: Error opening file\n", pathname.c_str());
584 if (avformat_find_stream_info(format_ctx.get(), nullptr) < 0) {
585 fprintf(stderr, "%s: Error finding stream info\n", pathname.c_str());
589 int video_stream_index = find_stream_index(format_ctx.get(), AVMEDIA_TYPE_VIDEO);
590 if (video_stream_index == -1) {
591 fprintf(stderr, "%s: No video stream found\n", pathname.c_str());
595 int audio_stream_index = find_stream_index(format_ctx.get(), AVMEDIA_TYPE_AUDIO);
596 int subtitle_stream_index = find_stream_index(format_ctx.get(), AVMEDIA_TYPE_SUBTITLE);
597 has_last_subtitle = false;
599 // Open video decoder.
600 const AVCodecParameters *video_codecpar = format_ctx->streams[video_stream_index]->codecpar;
601 const AVCodec *video_codec = avcodec_find_decoder(video_codecpar->codec_id);
603 video_timebase = format_ctx->streams[video_stream_index]->time_base;
604 AVCodecContextWithDeleter video_codec_ctx = avcodec_alloc_context3_unique(nullptr);
605 if (avcodec_parameters_to_context(video_codec_ctx.get(), video_codecpar) < 0) {
606 fprintf(stderr, "%s: Cannot fill video codec parameters\n", pathname.c_str());
609 if (video_codec == nullptr) {
610 fprintf(stderr, "%s: Cannot find video decoder\n", pathname.c_str());
614 // Seemingly, it's not too easy to make something that just initializes
615 // “whatever goes”, so we don't get CUDA or VULKAN or whatever here
616 // without enumerating through several different types.
617 // VA-API and VDPAU will do for now. We prioritize VDPAU for the
618 // simple reason that there's a VA-API-via-VDPAU emulation for NVidia
619 // cards that seems to work, but just hangs when trying to transfer the frame.
621 // Note that we don't actually check codec support beforehand,
622 // so if you have a low-end VDPAU device but a high-end VA-API device,
623 // you lose out on the extra codec support from the latter.
624 AVBufferRef *hw_device_ctx = nullptr;
625 if (av_hwdevice_ctx_create(&hw_device_ctx, AV_HWDEVICE_TYPE_VDPAU, nullptr, nullptr, 0) >= 0) {
626 video_codec_ctx->hw_device_ctx = av_buffer_ref(hw_device_ctx);
627 video_codec_ctx->get_format = get_hw_format<AV_HWDEVICE_TYPE_VDPAU>;
628 } else if (av_hwdevice_ctx_create(&hw_device_ctx, AV_HWDEVICE_TYPE_VAAPI, nullptr, nullptr, 0) >= 0) {
629 video_codec_ctx->hw_device_ctx = av_buffer_ref(hw_device_ctx);
630 video_codec_ctx->get_format = get_hw_format<AV_HWDEVICE_TYPE_VAAPI>;
632 fprintf(stderr, "Failed to initialize VA-API or VDPAU for FFmpeg acceleration. Decoding video in software.\n");
635 if (avcodec_open2(video_codec_ctx.get(), video_codec, nullptr) < 0) {
636 fprintf(stderr, "%s: Cannot open video decoder\n", pathname.c_str());
639 unique_ptr<AVCodecContext, decltype(avcodec_close)*> video_codec_ctx_cleanup(
640 video_codec_ctx.get(), avcodec_close);
642 // Used in decode_ycbcr_format().
643 is_mjpeg = video_codecpar->codec_id == AV_CODEC_ID_MJPEG;
645 // Open audio decoder, if we have audio.
646 AVCodecContextWithDeleter audio_codec_ctx;
647 if (audio_stream_index != -1) {
648 audio_codec_ctx = avcodec_alloc_context3_unique(nullptr);
649 const AVCodecParameters *audio_codecpar = format_ctx->streams[audio_stream_index]->codecpar;
650 audio_timebase = format_ctx->streams[audio_stream_index]->time_base;
651 if (avcodec_parameters_to_context(audio_codec_ctx.get(), audio_codecpar) < 0) {
652 fprintf(stderr, "%s: Cannot fill audio codec parameters\n", pathname.c_str());
655 const AVCodec *audio_codec = avcodec_find_decoder(audio_codecpar->codec_id);
656 if (audio_codec == nullptr) {
657 fprintf(stderr, "%s: Cannot find audio decoder\n", pathname.c_str());
660 if (avcodec_open2(audio_codec_ctx.get(), audio_codec, nullptr) < 0) {
661 fprintf(stderr, "%s: Cannot open audio decoder\n", pathname.c_str());
665 unique_ptr<AVCodecContext, decltype(avcodec_close)*> audio_codec_ctx_cleanup(
666 audio_codec_ctx.get(), avcodec_close);
671 bool first_frame = true;
672 int consecutive_errors = 0;
673 while (!producer_thread_should_quit.should_quit()) {
674 if (process_queued_commands(format_ctx.get(), pathname, last_modified, /*rewound=*/nullptr)) {
677 if (should_interrupt.load()) {
678 // Check as a failsafe, so that we don't need to rely on avio if we don't have to.
681 UniqueFrame audio_frame = audio_frame_allocator->alloc_frame();
682 AudioFormat audio_format;
686 frame_timeout_started = steady_clock::now();
687 frame_timeout_valid = true;
688 AVFrameWithDeleter frame = decode_frame(format_ctx.get(), video_codec_ctx.get(), audio_codec_ctx.get(),
689 pathname, video_stream_index, audio_stream_index, subtitle_stream_index, audio_frame.get(), &audio_format, &audio_pts, &error);
690 frame_timeout_valid = false;
691 if (should_interrupt.load()) {
692 // Abort no matter whether we got a frame or not.
696 if (++consecutive_errors >= 100) {
697 fprintf(stderr, "More than 100 consecutive error video frames, aborting playback.\n");
703 consecutive_errors = 0;
705 if (frame == nullptr) {
706 // EOF. Loop back to the start if we can.
707 if (format_ctx->pb != nullptr && format_ctx->pb->seekable == 0) {
708 // Not seekable (but seemingly, sometimes av_seek_frame() would return 0 anyway,
712 if (av_seek_frame(format_ctx.get(), /*stream_index=*/-1, /*timestamp=*/0, /*flags=*/0) < 0) {
713 fprintf(stderr, "%s: Rewind failed, not looping.\n", pathname.c_str());
716 if (video_codec_ctx != nullptr) {
717 avcodec_flush_buffers(video_codec_ctx.get());
719 if (audio_codec_ctx != nullptr) {
720 avcodec_flush_buffers(audio_codec_ctx.get());
722 // If the file has changed since last time, return to get it reloaded.
723 // Note that depending on how you move the file into place, you might
724 // end up corrupting the one you're already playing, so this path
725 // might not trigger.
726 if (changed_since(pathname, last_modified)) {
733 VideoFormat video_format = construct_video_format(frame.get(), video_timebase);
734 if (video_format.frame_rate_nom == 0 || video_format.frame_rate_den == 0) {
735 // Invalid frame rate; try constructing it from the previous frame length.
736 // (This is especially important if we are the master card, for SRT,
737 // since it affects audio. Not all senders have good timebases
738 // (e.g., Larix rounds first to timebase 1000 and then multiplies by
739 // 90 from there, it seems), but it's much better to have an oscillating
740 // value than just locking at 60.
741 if (last_pts != 0 && frame->pts > last_pts) {
742 int64_t pts_diff = frame->pts - last_pts;
743 video_format.frame_rate_nom = video_timebase.den;
744 video_format.frame_rate_den = video_timebase.num * pts_diff;
746 video_format.frame_rate_nom = 60;
747 video_format.frame_rate_den = 1;
750 UniqueFrame video_frame = make_video_frame(frame.get(), pathname, &error);
756 if (last_pts == 0 && pts_origin == 0) {
757 pts_origin = frame->pts;
759 steady_clock::time_point now = steady_clock::now();
760 if (play_as_fast_as_possible) {
761 video_frame->received_timestamp = now;
762 audio_frame->received_timestamp = now;
763 next_frame_start = now;
765 next_frame_start = compute_frame_start(frame->pts, pts_origin, video_timebase, start, rate);
766 if (first_frame && last_frame_was_connected) {
767 // If reconnect took more than one second, this is probably a live feed,
768 // and we should reset the resampler. (Or the rate is really, really low,
769 // in which case a reset on the first frame is fine anyway.)
770 if (duration<double>(next_frame_start - last_frame).count() >= 1.0) {
771 last_frame_was_connected = false;
774 video_frame->received_timestamp = next_frame_start;
776 // The easiest way to get all the rate conversions etc. right is to move the
777 // audio PTS into the video PTS timebase and go from there. (We'll get some
778 // rounding issues, but they should not be a big problem.)
779 int64_t audio_pts_as_video_pts = av_rescale_q(audio_pts, audio_timebase, video_timebase);
780 audio_frame->received_timestamp = compute_frame_start(audio_pts_as_video_pts, pts_origin, video_timebase, start, rate);
782 if (audio_frame->len != 0) {
783 // The received timestamps in Nageru are measured after we've just received the frame.
784 // However, pts (especially audio pts) is at the _beginning_ of the frame.
785 // If we have locked audio, the distinction doesn't really matter, as pts is
786 // on a relative scale and a fixed offset is fine. But if we don't, we will have
787 // a different number of samples each time, which will cause huge audio jitter
788 // and throw off the resampler.
790 // In a sense, we should have compensated by adding the frame and audio lengths
791 // to video_frame->received_timestamp and audio_frame->received_timestamp respectively,
792 // but that would mean extra waiting in sleep_until(). All we need is that they
793 // are correct relative to each other, though (and to the other frames we send),
794 // so just align the end of the audio frame, and we're fine.
795 size_t num_samples = (audio_frame->len * 8) / audio_format.bits_per_sample / audio_format.num_channels;
796 double offset = double(num_samples) / OUTPUT_FREQUENCY -
797 double(video_format.frame_rate_den) / video_format.frame_rate_nom;
798 audio_frame->received_timestamp += duration_cast<steady_clock::duration>(duration<double>(offset));
801 if (duration<double>(now - next_frame_start).count() >= 0.1) {
802 // If we don't have enough CPU to keep up, or if we have a live stream
803 // where the initial origin was somehow wrong, we could be behind indefinitely.
804 // In particular, this will give the audio resampler problems as it tries
805 // to speed up to reduce the delay, hitting the low end of the buffer every time.
806 fprintf(stderr, "%s: Playback %.0f ms behind, resetting time scale\n",
808 1e3 * duration<double>(now - next_frame_start).count());
809 pts_origin = frame->pts;
810 start = next_frame_start = now;
811 timecode += TYPICAL_FPS * 2 + 1;
814 bool finished_wakeup;
815 if (play_as_fast_as_possible) {
816 finished_wakeup = !producer_thread_should_quit.should_quit();
818 finished_wakeup = producer_thread_should_quit.sleep_until(next_frame_start);
820 if (finished_wakeup) {
821 if (audio_frame->len > 0) {
822 assert(audio_pts != -1);
824 if (!last_frame_was_connected) {
825 // We're recovering from an error (or really slow load, see above).
826 // Make sure to get the audio resampler reset. (This is a hack;
827 // ideally, the frame callback should just accept a way to signal
828 // audio discontinuity.)
829 timecode += TYPICAL_FPS * 2 + 1;
831 last_neutral_color = get_neutral_color(frame->metadata);
832 if (frame_callback != nullptr) {
833 frame_callback(frame->pts, video_timebase, audio_pts, audio_timebase, timecode++,
834 video_frame.get_and_release(), 0, video_format,
835 audio_frame.get_and_release(), 0, audio_format);
838 last_frame = steady_clock::now();
839 last_frame_was_connected = true;
842 if (producer_thread_should_quit.should_quit()) break;
844 bool rewound = false;
845 if (process_queued_commands(format_ctx.get(), pathname, last_modified, &rewound)) {
848 // If we just rewound, drop this frame on the floor and be done.
852 // OK, we didn't, so probably a rate change. Recalculate next_frame_start,
853 // but if it's now in the past, we'll reset the origin, so that we don't
854 // generate a huge backlog of frames that we need to run through quickly.
855 next_frame_start = compute_frame_start(frame->pts, pts_origin, video_timebase, start, rate);
856 steady_clock::time_point now = steady_clock::now();
857 if (next_frame_start < now) {
858 pts_origin = frame->pts;
859 start = next_frame_start = now;
863 last_pts = frame->pts;
868 void FFmpegCapture::internal_rewind()
870 pts_origin = last_pts = 0;
871 start = next_frame_start = steady_clock::now();
874 bool FFmpegCapture::process_queued_commands(AVFormatContext *format_ctx, const std::string &pathname, timespec last_modified, bool *rewound)
876 // Process any queued commands from other threads.
877 vector<QueuedCommand> commands;
879 lock_guard<mutex> lock(queue_mu);
880 swap(commands, command_queue);
882 for (const QueuedCommand &cmd : commands) {
883 switch (cmd.command) {
884 case QueuedCommand::REWIND:
885 if (av_seek_frame(format_ctx, /*stream_index=*/-1, /*timestamp=*/0, /*flags=*/0) < 0) {
886 fprintf(stderr, "%s: Rewind failed, stopping play.\n", pathname.c_str());
888 // If the file has changed since last time, return to get it reloaded.
889 // Note that depending on how you move the file into place, you might
890 // end up corrupting the one you're already playing, so this path
891 // might not trigger.
892 if (changed_since(pathname, last_modified)) {
896 if (rewound != nullptr) {
901 case QueuedCommand::CHANGE_RATE:
902 // Change the origin to the last played frame.
903 start = compute_frame_start(last_pts, pts_origin, video_timebase, start, rate);
904 pts_origin = last_pts;
906 play_as_fast_as_possible = (rate >= 10.0);
917 AVFrameWithDeleter FFmpegCapture::decode_frame(AVFormatContext *format_ctx, AVCodecContext *video_codec_ctx, AVCodecContext *audio_codec_ctx,
918 const std::string &pathname, int video_stream_index, int audio_stream_index, int subtitle_stream_index,
919 FrameAllocator::Frame *audio_frame, AudioFormat *audio_format, int64_t *audio_pts, bool *error)
923 // Read packets until we have a frame or there are none left.
924 bool frame_finished = false;
925 AVFrameWithDeleter audio_avframe = av_frame_alloc_unique();
926 AVFrameWithDeleter video_avframe = av_frame_alloc_unique();
929 bool has_audio = false;
931 AVPacketWithDeleter pkt = av_packet_alloc_unique();
934 if (av_read_frame(format_ctx, pkt.get()) == 0) {
935 if (pkt->stream_index == audio_stream_index && audio_callback != nullptr) {
936 audio_callback(pkt.get(), format_ctx->streams[audio_stream_index]->time_base);
938 if (pkt->stream_index == video_stream_index && video_callback != nullptr) {
939 video_callback(pkt.get(), format_ctx->streams[video_stream_index]->time_base);
941 if (pkt->stream_index == video_stream_index && global_flags.transcode_video) {
942 if (avcodec_send_packet(video_codec_ctx, pkt.get()) < 0) {
943 fprintf(stderr, "%s: Cannot send packet to video codec.\n", pathname.c_str());
945 return AVFrameWithDeleter(nullptr);
947 } else if (pkt->stream_index == audio_stream_index && global_flags.transcode_audio) {
949 if (avcodec_send_packet(audio_codec_ctx, pkt.get()) < 0) {
950 fprintf(stderr, "%s: Cannot send packet to audio codec.\n", pathname.c_str());
952 return AVFrameWithDeleter(nullptr);
954 } else if (pkt->stream_index == subtitle_stream_index) {
955 last_subtitle = string(reinterpret_cast<const char *>(pkt->data), pkt->size);
956 has_last_subtitle = true;
959 eof = true; // Or error, but ignore that for the time being.
962 // Decode audio, if any.
965 int err = avcodec_receive_frame(audio_codec_ctx, audio_avframe.get());
967 if (*audio_pts == -1) {
968 *audio_pts = audio_avframe->pts;
970 convert_audio(audio_avframe.get(), audio_frame, audio_format);
971 } else if (err == AVERROR(EAGAIN)) {
974 fprintf(stderr, "%s: Cannot receive frame from audio codec.\n", pathname.c_str());
976 return AVFrameWithDeleter(nullptr);
981 // Decode video, if we have a frame.
982 int err = avcodec_receive_frame(video_codec_ctx, video_avframe.get());
984 if (video_avframe->format == AV_PIX_FMT_VAAPI ||
985 video_avframe->format == AV_PIX_FMT_VDPAU) {
986 // Get the frame down to the CPU. (TODO: See if we can keep it
987 // on the GPU all the way, since it will be going up again later.
988 // However, this only works if the OpenGL GPU is the same one.)
989 AVFrameWithDeleter sw_frame = av_frame_alloc_unique();
990 int err = av_hwframe_transfer_data(sw_frame.get(), video_avframe.get(), 0);
992 fprintf(stderr, "%s: Cannot transfer hardware video frame to software.\n", pathname.c_str());
994 return AVFrameWithDeleter(nullptr);
996 sw_frame->pts = video_avframe->pts;
997 sw_frame->pkt_duration = video_avframe->pkt_duration;
998 video_avframe = move(sw_frame);
1000 frame_finished = true;
1002 } else if (err != AVERROR(EAGAIN)) {
1003 fprintf(stderr, "%s: Cannot receive frame from video codec.\n", pathname.c_str());
1005 return AVFrameWithDeleter(nullptr);
1010 return video_avframe;
1012 return AVFrameWithDeleter(nullptr);
1015 void FFmpegCapture::convert_audio(const AVFrame *audio_avframe, FrameAllocator::Frame *audio_frame, AudioFormat *audio_format)
1017 // Decide on a format. If there already is one in this audio frame,
1018 // we're pretty much forced to use it. If not, we try to find an exact match.
1019 // If that still doesn't work, we default to 32-bit signed chunked
1020 // (float would be nice, but there's really no way to signal that yet).
1021 AVSampleFormat dst_format;
1022 if (audio_format->bits_per_sample == 0) {
1023 switch (audio_avframe->format) {
1024 case AV_SAMPLE_FMT_S16:
1025 case AV_SAMPLE_FMT_S16P:
1026 audio_format->bits_per_sample = 16;
1027 dst_format = AV_SAMPLE_FMT_S16;
1029 case AV_SAMPLE_FMT_S32:
1030 case AV_SAMPLE_FMT_S32P:
1032 audio_format->bits_per_sample = 32;
1033 dst_format = AV_SAMPLE_FMT_S32;
1036 } else if (audio_format->bits_per_sample == 16) {
1037 dst_format = AV_SAMPLE_FMT_S16;
1038 } else if (audio_format->bits_per_sample == 32) {
1039 dst_format = AV_SAMPLE_FMT_S32;
1043 audio_format->num_channels = 2;
1045 AVChannelLayout channel_layout = audio_avframe->ch_layout;
1046 if (!av_channel_layout_check(&channel_layout) ||
1047 channel_layout.order == AV_CHANNEL_ORDER_UNSPEC) {
1048 av_channel_layout_default(&channel_layout, audio_avframe->ch_layout.nb_channels);
1051 if (resampler == nullptr ||
1052 audio_avframe->format != last_src_format ||
1053 dst_format != last_dst_format ||
1054 av_channel_layout_compare(&channel_layout, &last_channel_layout) != 0||
1055 audio_avframe->sample_rate != last_sample_rate) {
1056 // TODO: When we get C++20, use AV_CHANNEL_LAYOUT_STEREO_DOWNMIX.
1057 AVChannelLayout stereo_downmix;
1058 stereo_downmix.order = AV_CHANNEL_ORDER_NATIVE;
1059 stereo_downmix.nb_channels = 2;
1060 stereo_downmix.u.mask = AV_CH_LAYOUT_STEREO_DOWNMIX;
1062 swr_free(&resampler);
1063 resampler = nullptr;
1064 int err = swr_alloc_set_opts2(&resampler,
1065 /*out_ch_layout=*/&stereo_downmix,
1066 /*out_sample_fmt=*/dst_format,
1067 /*out_sample_rate=*/OUTPUT_FREQUENCY,
1068 /*in_ch_layout=*/&channel_layout,
1069 /*in_sample_fmt=*/AVSampleFormat(audio_avframe->format),
1070 /*in_sample_rate=*/audio_avframe->sample_rate,
1072 /*log_ctx=*/nullptr);
1074 if (err != 0 || resampler == nullptr) {
1075 fprintf(stderr, "Allocating resampler failed.\n");
1079 if (swr_init(resampler) < 0) {
1080 fprintf(stderr, "Could not open resample context.\n");
1084 last_src_format = AVSampleFormat(audio_avframe->format);
1085 last_dst_format = dst_format;
1086 last_channel_layout = channel_layout;
1087 last_sample_rate = audio_avframe->sample_rate;
1090 size_t bytes_per_sample = (audio_format->bits_per_sample / 8) * 2;
1091 size_t num_samples_room = (audio_frame->size - audio_frame->len) / bytes_per_sample;
1093 uint8_t *data = audio_frame->data + audio_frame->len;
1094 int out_samples = swr_convert(resampler, &data, num_samples_room,
1095 const_cast<const uint8_t **>(audio_avframe->data), audio_avframe->nb_samples);
1096 if (out_samples < 0) {
1097 fprintf(stderr, "Audio conversion failed.\n");
1101 audio_frame->len += out_samples * bytes_per_sample;
1104 VideoFormat FFmpegCapture::construct_video_format(const AVFrame *frame, AVRational video_timebase)
1106 VideoFormat video_format;
1107 video_format.width = frame_width(frame);
1108 video_format.height = frame_height(frame);
1109 if (pixel_format == bmusb::PixelFormat_8BitBGRA) {
1110 video_format.stride = frame_width(frame) * 4;
1111 } else if (pixel_format == FFmpegCapture::PixelFormat_NV12) {
1112 video_format.stride = frame_width(frame);
1114 assert(pixel_format == bmusb::PixelFormat_8BitYCbCrPlanar);
1115 video_format.stride = frame_width(frame);
1117 video_format.frame_rate_nom = video_timebase.den;
1118 video_format.frame_rate_den = frame->pkt_duration * video_timebase.num;
1119 video_format.has_signal = true;
1120 video_format.is_connected = true;
1121 return video_format;
1124 UniqueFrame FFmpegCapture::make_video_frame(const AVFrame *frame, const string &pathname, bool *error)
1128 UniqueFrame video_frame(video_frame_allocator->alloc_frame());
1129 if (video_frame->data == nullptr) {
1133 if (sws_ctx == nullptr ||
1134 sws_last_width != frame->width ||
1135 sws_last_height != frame->height ||
1136 sws_last_src_format != frame->format) {
1137 sws_dst_format = decide_dst_format(AVPixelFormat(frame->format), pixel_format);
1139 sws_getContext(frame->width, frame->height, AVPixelFormat(frame->format),
1140 frame_width(frame), frame_height(frame), sws_dst_format,
1141 SWS_BICUBIC, nullptr, nullptr, nullptr));
1142 sws_last_width = frame->width;
1143 sws_last_height = frame->height;
1144 sws_last_src_format = frame->format;
1146 if (sws_ctx == nullptr) {
1147 fprintf(stderr, "%s: Could not create scaler context\n", pathname.c_str());
1152 uint8_t *pic_data[4] = { nullptr, nullptr, nullptr, nullptr };
1153 int linesizes[4] = { 0, 0, 0, 0 };
1154 if (pixel_format == bmusb::PixelFormat_8BitBGRA) {
1155 pic_data[0] = video_frame->data;
1156 linesizes[0] = frame_width(frame) * 4;
1157 video_frame->len = (frame_width(frame) * 4) * frame_height(frame);
1158 } else if (pixel_format == PixelFormat_NV12) {
1159 pic_data[0] = video_frame->data;
1160 linesizes[0] = frame_width(frame);
1162 pic_data[1] = pic_data[0] + frame_width(frame) * frame_height(frame);
1163 linesizes[1] = frame_width(frame);
1165 video_frame->len = (frame_width(frame) * 2) * frame_height(frame);
1167 const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(sws_dst_format);
1168 current_frame_ycbcr_format = decode_ycbcr_format(desc, frame, is_mjpeg, &last_colorspace, &last_chroma_location);
1170 assert(pixel_format == bmusb::PixelFormat_8BitYCbCrPlanar);
1171 const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(sws_dst_format);
1173 int chroma_width = AV_CEIL_RSHIFT(int(frame_width(frame)), desc->log2_chroma_w);
1174 int chroma_height = AV_CEIL_RSHIFT(int(frame_height(frame)), desc->log2_chroma_h);
1176 pic_data[0] = video_frame->data;
1177 linesizes[0] = frame_width(frame);
1179 pic_data[1] = pic_data[0] + frame_width(frame) * frame_height(frame);
1180 linesizes[1] = chroma_width;
1182 pic_data[2] = pic_data[1] + chroma_width * chroma_height;
1183 linesizes[2] = chroma_width;
1185 video_frame->len = frame_width(frame) * frame_height(frame) + 2 * chroma_width * chroma_height;
1187 current_frame_ycbcr_format = decode_ycbcr_format(desc, frame, is_mjpeg, &last_colorspace, &last_chroma_location);
1190 // FIXME: Currently, if the video is too high-res for one of the allocated
1191 // frames, we simply refuse to scale it here to avoid crashes. It would be better
1192 // if we could somehow signal getting larger frames, especially as 4K is a thing now.
1193 if (video_frame->len > FRAME_SIZE) {
1194 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);
1199 sws_scale(sws_ctx.get(), frame->data, frame->linesize, 0, frame->height, pic_data, linesizes);
1204 int FFmpegCapture::interrupt_cb_thunk(void *opaque)
1206 return reinterpret_cast<FFmpegCapture *>(opaque)->interrupt_cb();
1209 int FFmpegCapture::interrupt_cb()
1211 // If ten seconds is gone without anything happening, we assume that
1212 // we are in a network stream that died and FFmpeg just didn't
1213 // pick it up (or perhaps it just hung, keeping the connection open).
1214 // Called back approximately every 100 ms if something is hanging,
1215 // so we get more than enough accuracy for our purposes.
1216 if (!should_interrupt && frame_timeout_valid &&
1217 duration<double>(steady_clock::now() - frame_timeout_started).count() >= 10.0) {
1218 string filename_copy;
1220 lock_guard<mutex> lock(filename_mu);
1221 filename_copy = filename;
1223 fprintf(stderr, "%s: No frame for more than 10 seconds, restarting stream.\n", filename.c_str());
1224 should_interrupt = true;
1226 return should_interrupt.load();
1229 unsigned FFmpegCapture::frame_width(const AVFrame *frame) const
1232 return frame->width;
1238 unsigned FFmpegCapture::frame_height(const AVFrame *frame) const
1241 return frame->height;
1248 int FFmpegCapture::read_srt_thunk(void *opaque, uint8_t *buf, int buf_size)
1250 return reinterpret_cast<FFmpegCapture *>(opaque)->read_srt(buf, buf_size);
1253 int FFmpegCapture::read_srt(uint8_t *buf, int buf_size)
1255 SRT_MSGCTRL mc = srt_msgctrl_default;
1256 return srt_recvmsg2(srt_sock, reinterpret_cast<char *>(buf), buf_size, &mc);