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>
30 #include "bmusb/bmusb.h"
31 #include "shared/ffmpeg_raii.h"
32 #include "ffmpeg_util.h"
34 #include "image_input.h"
35 #include "ref_counted_frame.h"
36 #include "shared/timebase.h"
38 #define FRAME_SIZE (8 << 20) // 8 MB.
41 using namespace std::chrono;
42 using namespace bmusb;
43 using namespace movit;
47 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)
49 const duration<double> pts((frame_pts - pts_origin) * double(video_timebase.num) / double(video_timebase.den));
50 return origin + duration_cast<steady_clock::duration>(pts / rate);
53 bool changed_since(const std::string &pathname, const timespec &ts)
59 if (stat(pathname.c_str(), &buf) != 0) {
60 fprintf(stderr, "%s: Couldn't check for new version, leaving the old in place.\n", pathname.c_str());
63 return (buf.st_mtim.tv_sec != ts.tv_sec || buf.st_mtim.tv_nsec != ts.tv_nsec);
66 bool is_full_range(const AVPixFmtDescriptor *desc)
68 // This is horrible, but there's no better way that I know of.
69 return (strchr(desc->name, 'j') != nullptr);
72 AVPixelFormat decide_dst_format(AVPixelFormat src_format, bmusb::PixelFormat dst_format_type)
74 if (dst_format_type == bmusb::PixelFormat_8BitBGRA) {
75 return AV_PIX_FMT_BGRA;
77 if (dst_format_type == FFmpegCapture::PixelFormat_NV12) {
78 return AV_PIX_FMT_NV12;
81 assert(dst_format_type == bmusb::PixelFormat_8BitYCbCrPlanar);
83 // If this is a non-Y'CbCr format, just convert to 4:4:4 Y'CbCr
84 // and be done with it. It's too strange to spend a lot of time on.
85 // (Let's hope there's no alpha.)
86 const AVPixFmtDescriptor *src_desc = av_pix_fmt_desc_get(src_format);
87 if (src_desc == nullptr ||
88 src_desc->nb_components != 3 ||
89 (src_desc->flags & AV_PIX_FMT_FLAG_RGB)) {
90 return AV_PIX_FMT_YUV444P;
93 // The best for us would be Cb and Cr together if possible,
94 // but FFmpeg doesn't support that except in the special case of
95 // NV12, so we need to go to planar even for the case of NV12.
96 // Thus, look for the closest (but no worse) 8-bit planar Y'CbCr format
97 // that matches in color range. (This will also include the case of
98 // the source format already being acceptable.)
99 bool src_full_range = is_full_range(src_desc);
100 const char *best_format = "yuv444p";
101 unsigned best_score = numeric_limits<unsigned>::max();
102 for (const AVPixFmtDescriptor *desc = av_pix_fmt_desc_next(nullptr);
104 desc = av_pix_fmt_desc_next(desc)) {
105 // Find planar Y'CbCr formats only.
106 if (desc->nb_components != 3) continue;
107 if (desc->flags & AV_PIX_FMT_FLAG_RGB) continue;
108 if (!(desc->flags & AV_PIX_FMT_FLAG_PLANAR)) continue;
109 if (desc->comp[0].plane != 0 ||
110 desc->comp[1].plane != 1 ||
111 desc->comp[2].plane != 2) continue;
113 // 8-bit formats only.
114 if (desc->flags & AV_PIX_FMT_FLAG_BE) continue;
115 if (desc->comp[0].depth != 8) continue;
117 // Same or better chroma resolution only.
118 int chroma_w_diff = desc->log2_chroma_w - src_desc->log2_chroma_w;
119 int chroma_h_diff = desc->log2_chroma_h - src_desc->log2_chroma_h;
120 if (chroma_w_diff < 0 || chroma_h_diff < 0)
123 // Matching full/limited range only.
124 if (is_full_range(desc) != src_full_range)
127 // Pick something with as little excess chroma resolution as possible.
128 unsigned score = (1 << (chroma_w_diff)) << chroma_h_diff;
129 if (score < best_score) {
131 best_format = desc->name;
134 return av_get_pix_fmt(best_format);
137 YCbCrFormat decode_ycbcr_format(const AVPixFmtDescriptor *desc, const AVFrame *frame, bool is_mjpeg)
140 AVColorSpace colorspace = frame->colorspace;
141 switch (colorspace) {
142 case AVCOL_SPC_BT709:
143 format.luma_coefficients = YCBCR_REC_709;
145 case AVCOL_SPC_BT470BG:
146 case AVCOL_SPC_SMPTE170M:
147 case AVCOL_SPC_SMPTE240M:
148 format.luma_coefficients = YCBCR_REC_601;
150 case AVCOL_SPC_BT2020_NCL:
151 format.luma_coefficients = YCBCR_REC_2020;
153 case AVCOL_SPC_UNSPECIFIED:
154 format.luma_coefficients = (frame->height >= 720 ? YCBCR_REC_709 : YCBCR_REC_601);
157 fprintf(stderr, "Unknown Y'CbCr coefficient enum %d from FFmpeg; choosing Rec. 709.\n",
159 format.luma_coefficients = YCBCR_REC_709;
163 format.full_range = is_full_range(desc);
164 format.num_levels = 1 << desc->comp[0].depth;
165 format.chroma_subsampling_x = 1 << desc->log2_chroma_w;
166 format.chroma_subsampling_y = 1 << desc->log2_chroma_h;
168 switch (frame->chroma_location) {
169 case AVCHROMA_LOC_LEFT:
170 format.cb_x_position = 0.0;
171 format.cb_y_position = 0.5;
173 case AVCHROMA_LOC_CENTER:
174 format.cb_x_position = 0.5;
175 format.cb_y_position = 0.5;
177 case AVCHROMA_LOC_TOPLEFT:
178 format.cb_x_position = 0.0;
179 format.cb_y_position = 0.0;
181 case AVCHROMA_LOC_TOP:
182 format.cb_x_position = 0.5;
183 format.cb_y_position = 0.0;
185 case AVCHROMA_LOC_BOTTOMLEFT:
186 format.cb_x_position = 0.0;
187 format.cb_y_position = 1.0;
189 case AVCHROMA_LOC_BOTTOM:
190 format.cb_x_position = 0.5;
191 format.cb_y_position = 1.0;
194 fprintf(stderr, "Unknown chroma location coefficient enum %d from FFmpeg; choosing Rec. 709.\n",
195 frame->chroma_location);
196 format.cb_x_position = 0.5;
197 format.cb_y_position = 0.5;
201 if (is_mjpeg && !format.full_range) {
202 // Limited-range MJPEG is only detected by FFmpeg whenever a special
203 // JPEG comment is set, which means that in practice, the stream is
204 // almost certainly generated by Futatabi. Override FFmpeg's forced
205 // MJPEG defaults (it disregards the values set in the mux) with what
207 format.luma_coefficients = YCBCR_REC_709;
208 format.cb_x_position = 0.0;
209 format.cb_y_position = 0.5;
212 format.cr_x_position = format.cb_x_position;
213 format.cr_y_position = format.cb_y_position;
219 FFmpegCapture::FFmpegCapture(const string &filename, unsigned width, unsigned height)
220 : filename(filename), width(width), height(height), video_timebase{1, 1}
222 description = "Video: " + filename;
224 last_frame = steady_clock::now();
226 avformat_network_init(); // In case someone wants this.
229 FFmpegCapture::~FFmpegCapture()
231 if (has_dequeue_callbacks) {
232 dequeue_cleanup_callback();
234 swr_free(&resampler);
237 void FFmpegCapture::configure_card()
239 if (video_frame_allocator == nullptr) {
240 owned_video_frame_allocator.reset(new MallocFrameAllocator(FRAME_SIZE, NUM_QUEUED_VIDEO_FRAMES));
241 set_video_frame_allocator(owned_video_frame_allocator.get());
243 if (audio_frame_allocator == nullptr) {
244 // Audio can come out in pretty large chunks, so increase from the default 1 MB.
245 owned_audio_frame_allocator.reset(new MallocFrameAllocator(1 << 20, NUM_QUEUED_AUDIO_FRAMES));
246 set_audio_frame_allocator(owned_audio_frame_allocator.get());
250 void FFmpegCapture::start_bm_capture()
256 producer_thread_should_quit.unquit();
257 producer_thread = thread(&FFmpegCapture::producer_thread_func, this);
260 void FFmpegCapture::stop_dequeue_thread()
266 producer_thread_should_quit.quit();
267 producer_thread.join();
270 std::map<uint32_t, VideoMode> FFmpegCapture::get_available_video_modes() const
272 // Note: This will never really be shown in the UI.
276 snprintf(buf, sizeof(buf), "%ux%u", width, height);
279 mode.autodetect = false;
281 mode.height = height;
282 mode.frame_rate_num = 60;
283 mode.frame_rate_den = 1;
284 mode.interlaced = false;
286 return {{ 0, mode }};
289 void FFmpegCapture::producer_thread_func()
291 char thread_name[16];
292 snprintf(thread_name, sizeof(thread_name), "FFmpeg_C_%d", card_index);
293 pthread_setname_np(pthread_self(), thread_name);
295 while (!producer_thread_should_quit.should_quit()) {
296 string filename_copy;
298 lock_guard<mutex> lock(filename_mu);
299 filename_copy = filename;
302 string pathname = search_for_file(filename_copy);
303 if (pathname.empty()) {
304 fprintf(stderr, "%s not found, sleeping one second and trying again...\n", filename_copy.c_str());
305 send_disconnected_frame();
306 producer_thread_should_quit.sleep_for(seconds(1));
309 should_interrupt = false;
310 if (!play_video(pathname)) {
312 fprintf(stderr, "Error when playing %s, sleeping one second and trying again...\n", pathname.c_str());
313 send_disconnected_frame();
314 producer_thread_should_quit.sleep_for(seconds(1));
318 // Probably just EOF, will exit the loop above on next test.
321 if (has_dequeue_callbacks) {
322 dequeue_cleanup_callback();
323 has_dequeue_callbacks = false;
327 void FFmpegCapture::send_disconnected_frame()
329 // Send an empty frame to signal that we have no signal anymore.
330 FrameAllocator::Frame video_frame = video_frame_allocator->alloc_frame();
331 if (video_frame.data) {
332 VideoFormat video_format;
333 video_format.width = width;
334 video_format.height = height;
335 video_format.frame_rate_nom = 60;
336 video_format.frame_rate_den = 1;
337 video_format.is_connected = false;
338 if (pixel_format == bmusb::PixelFormat_8BitBGRA) {
339 video_format.stride = width * 4;
340 video_frame.len = width * height * 4;
341 memset(video_frame.data, 0, video_frame.len);
343 video_format.stride = width;
344 current_frame_ycbcr_format.luma_coefficients = YCBCR_REC_709;
345 current_frame_ycbcr_format.full_range = true;
346 current_frame_ycbcr_format.num_levels = 256;
347 current_frame_ycbcr_format.chroma_subsampling_x = 2;
348 current_frame_ycbcr_format.chroma_subsampling_y = 2;
349 current_frame_ycbcr_format.cb_x_position = 0.0f;
350 current_frame_ycbcr_format.cb_y_position = 0.0f;
351 current_frame_ycbcr_format.cr_x_position = 0.0f;
352 current_frame_ycbcr_format.cr_y_position = 0.0f;
353 video_frame.len = width * height * 2;
354 memset(video_frame.data, 0, width * height);
355 memset(video_frame.data + width * height, 128, width * height); // Valid for both NV12 and planar.
358 frame_callback(-1, AVRational{1, TIMEBASE}, -1, AVRational{1, TIMEBASE}, timecode++,
359 video_frame, /*video_offset=*/0, video_format,
360 FrameAllocator::Frame(), /*audio_offset=*/0, AudioFormat());
361 last_frame_was_connected = false;
365 bool FFmpegCapture::play_video(const string &pathname)
367 // Note: Call before open, not after; otherwise, there's a race.
368 // (There is now, too, but it tips the correct way. We could use fstat()
369 // if we had the file descriptor.)
370 timespec last_modified;
372 if (stat(pathname.c_str(), &buf) != 0) {
373 // Probably some sort of protocol, so can't stat.
374 last_modified.tv_sec = -1;
376 last_modified = buf.st_mtim;
379 auto format_ctx = avformat_open_input_unique(pathname.c_str(), nullptr, nullptr, AVIOInterruptCB{ &FFmpegCapture::interrupt_cb_thunk, this });
380 if (format_ctx == nullptr) {
381 fprintf(stderr, "%s: Error opening file\n", pathname.c_str());
385 if (avformat_find_stream_info(format_ctx.get(), nullptr) < 0) {
386 fprintf(stderr, "%s: Error finding stream info\n", pathname.c_str());
390 int video_stream_index = find_stream_index(format_ctx.get(), AVMEDIA_TYPE_VIDEO);
391 if (video_stream_index == -1) {
392 fprintf(stderr, "%s: No video stream found\n", pathname.c_str());
396 int audio_stream_index = find_stream_index(format_ctx.get(), AVMEDIA_TYPE_AUDIO);
397 int subtitle_stream_index = find_stream_index(format_ctx.get(), AVMEDIA_TYPE_SUBTITLE);
398 has_last_subtitle = false;
400 // Open video decoder.
401 const AVCodecParameters *video_codecpar = format_ctx->streams[video_stream_index]->codecpar;
402 AVCodec *video_codec = avcodec_find_decoder(video_codecpar->codec_id);
403 video_timebase = format_ctx->streams[video_stream_index]->time_base;
404 AVCodecContextWithDeleter video_codec_ctx = avcodec_alloc_context3_unique(nullptr);
405 if (avcodec_parameters_to_context(video_codec_ctx.get(), video_codecpar) < 0) {
406 fprintf(stderr, "%s: Cannot fill video codec parameters\n", pathname.c_str());
409 if (video_codec == nullptr) {
410 fprintf(stderr, "%s: Cannot find video decoder\n", pathname.c_str());
413 if (avcodec_open2(video_codec_ctx.get(), video_codec, nullptr) < 0) {
414 fprintf(stderr, "%s: Cannot open video decoder\n", pathname.c_str());
417 unique_ptr<AVCodecContext, decltype(avcodec_close)*> video_codec_ctx_cleanup(
418 video_codec_ctx.get(), avcodec_close);
420 // Used in decode_ycbcr_format().
421 is_mjpeg = video_codecpar->codec_id == AV_CODEC_ID_MJPEG;
423 // Open audio decoder, if we have audio.
424 AVCodecContextWithDeleter audio_codec_ctx;
425 if (audio_stream_index != -1) {
426 audio_codec_ctx = avcodec_alloc_context3_unique(nullptr);
427 const AVCodecParameters *audio_codecpar = format_ctx->streams[audio_stream_index]->codecpar;
428 audio_timebase = format_ctx->streams[audio_stream_index]->time_base;
429 if (avcodec_parameters_to_context(audio_codec_ctx.get(), audio_codecpar) < 0) {
430 fprintf(stderr, "%s: Cannot fill audio codec parameters\n", pathname.c_str());
433 AVCodec *audio_codec = avcodec_find_decoder(audio_codecpar->codec_id);
434 if (audio_codec == nullptr) {
435 fprintf(stderr, "%s: Cannot find audio decoder\n", pathname.c_str());
438 if (avcodec_open2(audio_codec_ctx.get(), audio_codec, nullptr) < 0) {
439 fprintf(stderr, "%s: Cannot open audio decoder\n", pathname.c_str());
443 unique_ptr<AVCodecContext, decltype(avcodec_close)*> audio_codec_ctx_cleanup(
444 audio_codec_ctx.get(), avcodec_close);
449 bool first_frame = true;
450 while (!producer_thread_should_quit.should_quit()) {
451 if (process_queued_commands(format_ctx.get(), pathname, last_modified, /*rewound=*/nullptr)) {
454 if (should_interrupt.load()) {
455 // Check as a failsafe, so that we don't need to rely on avio if we don't have to.
458 UniqueFrame audio_frame = audio_frame_allocator->alloc_frame();
459 AudioFormat audio_format;
463 AVFrameWithDeleter frame = decode_frame(format_ctx.get(), video_codec_ctx.get(), audio_codec_ctx.get(),
464 pathname, video_stream_index, audio_stream_index, subtitle_stream_index, audio_frame.get(), &audio_format, &audio_pts, &error);
468 if (frame == nullptr) {
469 // EOF. Loop back to the start if we can.
470 if (format_ctx->pb != nullptr && format_ctx->pb->seekable == 0) {
471 // Not seekable (but seemingly, sometimes av_seek_frame() would return 0 anyway,
475 if (av_seek_frame(format_ctx.get(), /*stream_index=*/-1, /*timestamp=*/0, /*flags=*/0) < 0) {
476 fprintf(stderr, "%s: Rewind failed, not looping.\n", pathname.c_str());
479 if (video_codec_ctx != nullptr) {
480 avcodec_flush_buffers(video_codec_ctx.get());
482 if (audio_codec_ctx != nullptr) {
483 avcodec_flush_buffers(audio_codec_ctx.get());
485 // If the file has changed since last time, return to get it reloaded.
486 // Note that depending on how you move the file into place, you might
487 // end up corrupting the one you're already playing, so this path
488 // might not trigger.
489 if (changed_since(pathname, last_modified)) {
496 VideoFormat video_format = construct_video_format(frame.get(), video_timebase);
497 UniqueFrame video_frame = make_video_frame(frame.get(), pathname, &error);
503 if (last_pts == 0 && pts_origin == 0) {
504 pts_origin = frame->pts;
506 steady_clock::time_point now = steady_clock::now();
507 if (play_as_fast_as_possible) {
508 video_frame->received_timestamp = now;
509 audio_frame->received_timestamp = now;
510 next_frame_start = now;
512 next_frame_start = compute_frame_start(frame->pts, pts_origin, video_timebase, start, rate);
513 if (first_frame && last_frame_was_connected) {
514 // If reconnect took more than one second, this is probably a live feed,
515 // and we should reset the resampler. (Or the rate is really, really low,
516 // in which case a reset on the first frame is fine anyway.)
517 if (duration<double>(next_frame_start - last_frame).count() >= 1.0) {
518 last_frame_was_connected = false;
521 video_frame->received_timestamp = next_frame_start;
523 // The easiest way to get all the rate conversions etc. right is to move the
524 // audio PTS into the video PTS timebase and go from there. (We'll get some
525 // rounding issues, but they should not be a big problem.)
526 int64_t audio_pts_as_video_pts = av_rescale_q(audio_pts, audio_timebase, video_timebase);
527 audio_frame->received_timestamp = compute_frame_start(audio_pts_as_video_pts, pts_origin, video_timebase, start, rate);
529 if (audio_frame->len != 0) {
530 // The received timestamps in Nageru are measured after we've just received the frame.
531 // However, pts (especially audio pts) is at the _beginning_ of the frame.
532 // If we have locked audio, the distinction doesn't really matter, as pts is
533 // on a relative scale and a fixed offset is fine. But if we don't, we will have
534 // a different number of samples each time, which will cause huge audio jitter
535 // and throw off the resampler.
537 // In a sense, we should have compensated by adding the frame and audio lengths
538 // to video_frame->received_timestamp and audio_frame->received_timestamp respectively,
539 // but that would mean extra waiting in sleep_until(). All we need is that they
540 // are correct relative to each other, though (and to the other frames we send),
541 // so just align the end of the audio frame, and we're fine.
542 size_t num_samples = (audio_frame->len * 8) / audio_format.bits_per_sample / audio_format.num_channels;
543 double offset = double(num_samples) / OUTPUT_FREQUENCY -
544 double(video_format.frame_rate_den) / video_format.frame_rate_nom;
545 audio_frame->received_timestamp += duration_cast<steady_clock::duration>(duration<double>(offset));
548 if (duration<double>(now - next_frame_start).count() >= 0.1) {
549 // If we don't have enough CPU to keep up, or if we have a live stream
550 // where the initial origin was somehow wrong, we could be behind indefinitely.
551 // In particular, this will give the audio resampler problems as it tries
552 // to speed up to reduce the delay, hitting the low end of the buffer every time.
553 fprintf(stderr, "%s: Playback %.0f ms behind, resetting time scale\n",
555 1e3 * duration<double>(now - next_frame_start).count());
556 pts_origin = frame->pts;
557 start = next_frame_start = now;
558 timecode += MAX_FPS * 2 + 1;
561 bool finished_wakeup;
562 if (play_as_fast_as_possible) {
563 finished_wakeup = !producer_thread_should_quit.should_quit();
565 finished_wakeup = producer_thread_should_quit.sleep_until(next_frame_start);
567 if (finished_wakeup) {
568 if (audio_frame->len > 0) {
569 assert(audio_pts != -1);
571 if (!last_frame_was_connected) {
572 // We're recovering from an error (or really slow load, see above).
573 // Make sure to get the audio resampler reset. (This is a hack;
574 // ideally, the frame callback should just accept a way to signal
575 // audio discontinuity.)
576 timecode += MAX_FPS * 2 + 1;
578 frame_callback(frame->pts, video_timebase, audio_pts, audio_timebase, timecode++,
579 video_frame.get_and_release(), 0, video_format,
580 audio_frame.get_and_release(), 0, audio_format);
582 last_frame = steady_clock::now();
583 last_frame_was_connected = true;
586 if (producer_thread_should_quit.should_quit()) break;
588 bool rewound = false;
589 if (process_queued_commands(format_ctx.get(), pathname, last_modified, &rewound)) {
592 // If we just rewound, drop this frame on the floor and be done.
596 // OK, we didn't, so probably a rate change. Recalculate next_frame_start,
597 // but if it's now in the past, we'll reset the origin, so that we don't
598 // generate a huge backlog of frames that we need to run through quickly.
599 next_frame_start = compute_frame_start(frame->pts, pts_origin, video_timebase, start, rate);
600 steady_clock::time_point now = steady_clock::now();
601 if (next_frame_start < now) {
602 pts_origin = frame->pts;
603 start = next_frame_start = now;
607 last_pts = frame->pts;
612 void FFmpegCapture::internal_rewind()
614 pts_origin = last_pts = 0;
615 start = next_frame_start = steady_clock::now();
618 bool FFmpegCapture::process_queued_commands(AVFormatContext *format_ctx, const std::string &pathname, timespec last_modified, bool *rewound)
620 // Process any queued commands from other threads.
621 vector<QueuedCommand> commands;
623 lock_guard<mutex> lock(queue_mu);
624 swap(commands, command_queue);
626 for (const QueuedCommand &cmd : commands) {
627 switch (cmd.command) {
628 case QueuedCommand::REWIND:
629 if (av_seek_frame(format_ctx, /*stream_index=*/-1, /*timestamp=*/0, /*flags=*/0) < 0) {
630 fprintf(stderr, "%s: Rewind failed, stopping play.\n", pathname.c_str());
632 // If the file has changed since last time, return to get it reloaded.
633 // Note that depending on how you move the file into place, you might
634 // end up corrupting the one you're already playing, so this path
635 // might not trigger.
636 if (changed_since(pathname, last_modified)) {
640 if (rewound != nullptr) {
645 case QueuedCommand::CHANGE_RATE:
646 // Change the origin to the last played frame.
647 start = compute_frame_start(last_pts, pts_origin, video_timebase, start, rate);
648 pts_origin = last_pts;
650 play_as_fast_as_possible = (rate >= 10.0);
661 AVFrameWithDeleter FFmpegCapture::decode_frame(AVFormatContext *format_ctx, AVCodecContext *video_codec_ctx, AVCodecContext *audio_codec_ctx,
662 const std::string &pathname, int video_stream_index, int audio_stream_index, int subtitle_stream_index,
663 FrameAllocator::Frame *audio_frame, AudioFormat *audio_format, int64_t *audio_pts, bool *error)
667 // Read packets until we have a frame or there are none left.
668 bool frame_finished = false;
669 AVFrameWithDeleter audio_avframe = av_frame_alloc_unique();
670 AVFrameWithDeleter video_avframe = av_frame_alloc_unique();
673 bool has_audio = false;
676 unique_ptr<AVPacket, decltype(av_packet_unref)*> pkt_cleanup(
677 &pkt, av_packet_unref);
678 av_init_packet(&pkt);
681 if (av_read_frame(format_ctx, &pkt) == 0) {
682 if (pkt.stream_index == audio_stream_index && audio_callback != nullptr) {
683 audio_callback(&pkt, format_ctx->streams[audio_stream_index]->time_base);
685 if (pkt.stream_index == video_stream_index) {
686 if (avcodec_send_packet(video_codec_ctx, &pkt) < 0) {
687 fprintf(stderr, "%s: Cannot send packet to video codec.\n", pathname.c_str());
689 return AVFrameWithDeleter(nullptr);
691 } else if (pkt.stream_index == audio_stream_index) {
693 if (avcodec_send_packet(audio_codec_ctx, &pkt) < 0) {
694 fprintf(stderr, "%s: Cannot send packet to audio codec.\n", pathname.c_str());
696 return AVFrameWithDeleter(nullptr);
698 } else if (pkt.stream_index == subtitle_stream_index) {
699 last_subtitle = string(reinterpret_cast<const char *>(pkt.data), pkt.size);
700 has_last_subtitle = true;
703 eof = true; // Or error, but ignore that for the time being.
706 // Decode audio, if any.
709 int err = avcodec_receive_frame(audio_codec_ctx, audio_avframe.get());
711 if (*audio_pts == -1) {
712 *audio_pts = audio_avframe->pts;
714 convert_audio(audio_avframe.get(), audio_frame, audio_format);
715 } else if (err == AVERROR(EAGAIN)) {
718 fprintf(stderr, "%s: Cannot receive frame from audio codec.\n", pathname.c_str());
720 return AVFrameWithDeleter(nullptr);
725 // Decode video, if we have a frame.
726 int err = avcodec_receive_frame(video_codec_ctx, video_avframe.get());
728 frame_finished = true;
730 } else if (err != AVERROR(EAGAIN)) {
731 fprintf(stderr, "%s: Cannot receive frame from video codec.\n", pathname.c_str());
733 return AVFrameWithDeleter(nullptr);
738 return video_avframe;
740 return AVFrameWithDeleter(nullptr);
743 void FFmpegCapture::convert_audio(const AVFrame *audio_avframe, FrameAllocator::Frame *audio_frame, AudioFormat *audio_format)
745 // Decide on a format. If there already is one in this audio frame,
746 // we're pretty much forced to use it. If not, we try to find an exact match.
747 // If that still doesn't work, we default to 32-bit signed chunked
748 // (float would be nice, but there's really no way to signal that yet).
749 AVSampleFormat dst_format;
750 if (audio_format->bits_per_sample == 0) {
751 switch (audio_avframe->format) {
752 case AV_SAMPLE_FMT_S16:
753 case AV_SAMPLE_FMT_S16P:
754 audio_format->bits_per_sample = 16;
755 dst_format = AV_SAMPLE_FMT_S16;
757 case AV_SAMPLE_FMT_S32:
758 case AV_SAMPLE_FMT_S32P:
760 audio_format->bits_per_sample = 32;
761 dst_format = AV_SAMPLE_FMT_S32;
764 } else if (audio_format->bits_per_sample == 16) {
765 dst_format = AV_SAMPLE_FMT_S16;
766 } else if (audio_format->bits_per_sample == 32) {
767 dst_format = AV_SAMPLE_FMT_S32;
771 audio_format->num_channels = 2;
773 int64_t channel_layout = audio_avframe->channel_layout;
774 if (channel_layout == 0) {
775 channel_layout = av_get_default_channel_layout(audio_avframe->channels);
778 if (resampler == nullptr ||
779 audio_avframe->format != last_src_format ||
780 dst_format != last_dst_format ||
781 channel_layout != last_channel_layout ||
782 audio_avframe->sample_rate != last_sample_rate) {
783 swr_free(&resampler);
784 resampler = swr_alloc_set_opts(nullptr,
785 /*out_ch_layout=*/AV_CH_LAYOUT_STEREO_DOWNMIX,
786 /*out_sample_fmt=*/dst_format,
787 /*out_sample_rate=*/OUTPUT_FREQUENCY,
788 /*in_ch_layout=*/channel_layout,
789 /*in_sample_fmt=*/AVSampleFormat(audio_avframe->format),
790 /*in_sample_rate=*/audio_avframe->sample_rate,
792 /*log_ctx=*/nullptr);
794 if (resampler == nullptr) {
795 fprintf(stderr, "Allocating resampler failed.\n");
799 if (swr_init(resampler) < 0) {
800 fprintf(stderr, "Could not open resample context.\n");
804 last_src_format = AVSampleFormat(audio_avframe->format);
805 last_dst_format = dst_format;
806 last_channel_layout = channel_layout;
807 last_sample_rate = audio_avframe->sample_rate;
810 size_t bytes_per_sample = (audio_format->bits_per_sample / 8) * 2;
811 size_t num_samples_room = (audio_frame->size - audio_frame->len) / bytes_per_sample;
813 uint8_t *data = audio_frame->data + audio_frame->len;
814 int out_samples = swr_convert(resampler, &data, num_samples_room,
815 const_cast<const uint8_t **>(audio_avframe->data), audio_avframe->nb_samples);
816 if (out_samples < 0) {
817 fprintf(stderr, "Audio conversion failed.\n");
821 audio_frame->len += out_samples * bytes_per_sample;
824 VideoFormat FFmpegCapture::construct_video_format(const AVFrame *frame, AVRational video_timebase)
826 VideoFormat video_format;
827 video_format.width = width;
828 video_format.height = height;
829 if (pixel_format == bmusb::PixelFormat_8BitBGRA) {
830 video_format.stride = width * 4;
831 } else if (pixel_format == FFmpegCapture::PixelFormat_NV12) {
832 video_format.stride = width;
834 assert(pixel_format == bmusb::PixelFormat_8BitYCbCrPlanar);
835 video_format.stride = width;
837 video_format.frame_rate_nom = video_timebase.den;
838 video_format.frame_rate_den = frame->pkt_duration * video_timebase.num;
839 if (video_format.frame_rate_nom == 0 || video_format.frame_rate_den == 0) {
840 // Invalid frame rate.
841 video_format.frame_rate_nom = 60;
842 video_format.frame_rate_den = 1;
844 video_format.has_signal = true;
845 video_format.is_connected = true;
849 UniqueFrame FFmpegCapture::make_video_frame(const AVFrame *frame, const string &pathname, bool *error)
853 UniqueFrame video_frame(video_frame_allocator->alloc_frame());
854 if (video_frame->data == nullptr) {
858 if (sws_ctx == nullptr ||
859 sws_last_width != frame->width ||
860 sws_last_height != frame->height ||
861 sws_last_src_format != frame->format) {
862 sws_dst_format = decide_dst_format(AVPixelFormat(frame->format), pixel_format);
864 sws_getContext(frame->width, frame->height, AVPixelFormat(frame->format),
865 width, height, sws_dst_format,
866 SWS_BICUBIC, nullptr, nullptr, nullptr));
867 sws_last_width = frame->width;
868 sws_last_height = frame->height;
869 sws_last_src_format = frame->format;
871 if (sws_ctx == nullptr) {
872 fprintf(stderr, "%s: Could not create scaler context\n", pathname.c_str());
877 uint8_t *pic_data[4] = { nullptr, nullptr, nullptr, nullptr };
878 int linesizes[4] = { 0, 0, 0, 0 };
879 if (pixel_format == bmusb::PixelFormat_8BitBGRA) {
880 pic_data[0] = video_frame->data;
881 linesizes[0] = width * 4;
882 video_frame->len = (width * 4) * height;
883 } else if (pixel_format == PixelFormat_NV12) {
884 pic_data[0] = video_frame->data;
885 linesizes[0] = width;
887 pic_data[1] = pic_data[0] + width * height;
888 linesizes[1] = width;
890 video_frame->len = (width * 2) * height;
892 const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(sws_dst_format);
893 current_frame_ycbcr_format = decode_ycbcr_format(desc, frame, is_mjpeg);
895 assert(pixel_format == bmusb::PixelFormat_8BitYCbCrPlanar);
896 const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(sws_dst_format);
898 int chroma_width = AV_CEIL_RSHIFT(int(width), desc->log2_chroma_w);
899 int chroma_height = AV_CEIL_RSHIFT(int(height), desc->log2_chroma_h);
901 pic_data[0] = video_frame->data;
902 linesizes[0] = width;
904 pic_data[1] = pic_data[0] + width * height;
905 linesizes[1] = chroma_width;
907 pic_data[2] = pic_data[1] + chroma_width * chroma_height;
908 linesizes[2] = chroma_width;
910 video_frame->len = width * height + 2 * chroma_width * chroma_height;
912 current_frame_ycbcr_format = decode_ycbcr_format(desc, frame, is_mjpeg);
914 sws_scale(sws_ctx.get(), frame->data, frame->linesize, 0, frame->height, pic_data, linesizes);
919 int FFmpegCapture::interrupt_cb_thunk(void *unique)
921 return reinterpret_cast<FFmpegCapture *>(unique)->interrupt_cb();
924 int FFmpegCapture::interrupt_cb()
926 return should_interrupt.load();