git.sesse.net Git - nageru/blob - nageru/ffmpeg_capture.cpp

   1 #include "ffmpeg_capture.h"
   2
   3 #include <assert.h>
   4 #include <pthread.h>
   5 #include <stdint.h>
   6 #include <stdio.h>
   7 #include <stdlib.h>
   8 #include <string.h>
   9 #include <sys/stat.h>
  10 #include <unistd.h>
  11
  12 extern "C" {
  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>
  23 }
  24
  25 #include <chrono>
  26 #include <cstdint>
  27 #include <utility>
  28 #include <vector>
  29
  30 #include <Eigen/Core>
  31 #include <Eigen/LU>
  32 #include <movit/colorspace_conversion_effect.h>
  33
  34 #include "bmusb/bmusb.h"
  35 #include "shared/ffmpeg_raii.h"
  36 #include "ffmpeg_util.h"
  37 #include "flags.h"
  38 #include "image_input.h"
  39 #include "ref_counted_frame.h"
  40 #include "shared/timebase.h"
  41
  42 #ifdef HAVE_SRT
  43 #include <srt/srt.h>
  44 #endif
  45
  46 #define FRAME_SIZE (8 << 20)  // 8 MB.
  47
  48 using namespace std;
  49 using namespace std::chrono;
  50 using namespace bmusb;
  51 using namespace movit;
  52 using namespace Eigen;
  53
  54 namespace {
  55
  56 steady_clock::time_point compute_frame_start(int64_t frame_pts, int64_t pts_origin, const AVRational &video_timebase, const steady_clock::time_point &origin, double rate)
  57 {
  58         const duration<double> pts((frame_pts - pts_origin) * double(video_timebase.num) / double(video_timebase.den));
  59         return origin + duration_cast<steady_clock::duration>(pts / rate);
  60 }
  61
  62 bool changed_since(const std::string &pathname, const timespec &ts)
  63 {
  64         if (ts.tv_sec < 0) {
  65                 return false;
  66         }
  67         struct stat buf;
  68         if (stat(pathname.c_str(), &buf) != 0) {
  69                 fprintf(stderr, "%s: Couldn't check for new version, leaving the old in place.\n", pathname.c_str());
  70                 return false;
  71         }
  72         return (buf.st_mtim.tv_sec != ts.tv_sec || buf.st_mtim.tv_nsec != ts.tv_nsec);
  73 }
  74
  75 bool is_full_range(const AVPixFmtDescriptor *desc)
  76 {
  77         // This is horrible, but there's no better way that I know of.
  78         return (strchr(desc->name, 'j') != nullptr);
  79 }
  80
  81 AVPixelFormat decide_dst_format(AVPixelFormat src_format, bmusb::PixelFormat dst_format_type)
  82 {
  83         if (dst_format_type == bmusb::PixelFormat_8BitBGRA) {
  84                 return AV_PIX_FMT_BGRA;
  85         }
  86         if (dst_format_type == FFmpegCapture::PixelFormat_NV12) {
  87                 return AV_PIX_FMT_NV12;
  88         }
  89
  90         assert(dst_format_type == bmusb::PixelFormat_8BitYCbCrPlanar);
  91
  92         // If this is a non-Y'CbCr format, just convert to 4:4:4 Y'CbCr
  93         // and be done with it. It's too strange to spend a lot of time on.
  94         // (Let's hope there's no alpha.)
  95         const AVPixFmtDescriptor *src_desc = av_pix_fmt_desc_get(src_format);
  96         if (src_desc == nullptr ||
  97             src_desc->nb_components != 3 ||
  98             (src_desc->flags & AV_PIX_FMT_FLAG_RGB)) {
  99                 return AV_PIX_FMT_YUV444P;
 100         }
 101
 102         // The best for us would be Cb and Cr together if possible,
 103         // but FFmpeg doesn't support that except in the special case of
 104         // NV12, so we need to go to planar even for the case of NV12.
 105         // Thus, look for the closest (but no worse) 8-bit planar Y'CbCr format
 106         // that matches in color range. (This will also include the case of
 107         // the source format already being acceptable.)
 108         bool src_full_range = is_full_range(src_desc);
 109         const char *best_format = "yuv444p";
 110         unsigned best_score = numeric_limits<unsigned>::max();
 111         for (const AVPixFmtDescriptor *desc = av_pix_fmt_desc_next(nullptr);
 112              desc;
 113              desc = av_pix_fmt_desc_next(desc)) {
 114                 // Find planar Y'CbCr formats only.
 115                 if (desc->nb_components != 3) continue;
 116                 if (desc->flags & AV_PIX_FMT_FLAG_RGB) continue;
 117                 if (!(desc->flags & AV_PIX_FMT_FLAG_PLANAR)) continue;
 118                 if (desc->comp[0].plane != 0 ||
 119                     desc->comp[1].plane != 1 ||
 120                     desc->comp[2].plane != 2) continue;
 121
 122                 // 8-bit formats only.
 123                 if (desc->flags & AV_PIX_FMT_FLAG_BE) continue;
 124                 if (desc->comp[0].depth != 8) continue;
 125
 126                 // Same or better chroma resolution only.
 127                 int chroma_w_diff = desc->log2_chroma_w - src_desc->log2_chroma_w;
 128                 int chroma_h_diff = desc->log2_chroma_h - src_desc->log2_chroma_h;
 129                 if (chroma_w_diff < 0 || chroma_h_diff < 0)
 130                         continue;
 131
 132                 // Matching full/limited range only.
 133                 if (is_full_range(desc) != src_full_range)
 134                         continue;
 135
 136                 // Pick something with as little excess chroma resolution as possible.
 137                 unsigned score = (1 << (chroma_w_diff)) << chroma_h_diff;
 138                 if (score < best_score) {
 139                         best_score = score;
 140                         best_format = desc->name;
 141                 }
 142         }
 143         return av_get_pix_fmt(best_format);
 144 }
 145
 146 YCbCrFormat decode_ycbcr_format(const AVPixFmtDescriptor *desc, const AVFrame *frame, bool is_mjpeg, AVColorSpace *last_colorspace, AVChromaLocation *last_chroma_location)
 147 {
 148         YCbCrFormat format;
 149         AVColorSpace colorspace = frame->colorspace;
 150         switch (colorspace) {
 151         case AVCOL_SPC_BT709:
 152                 format.luma_coefficients = YCBCR_REC_709;
 153                 break;
 154         case AVCOL_SPC_BT470BG:
 155         case AVCOL_SPC_SMPTE170M:
 156         case AVCOL_SPC_SMPTE240M:
 157                 format.luma_coefficients = YCBCR_REC_601;
 158                 break;
 159         case AVCOL_SPC_BT2020_NCL:
 160                 format.luma_coefficients = YCBCR_REC_2020;
 161                 break;
 162         case AVCOL_SPC_UNSPECIFIED:
 163                 format.luma_coefficients = (frame->height >= 720 ? YCBCR_REC_709 : YCBCR_REC_601);
 164                 break;
 165         default:
 166                 if (colorspace != *last_colorspace) {
 167                         fprintf(stderr, "Unknown Y'CbCr coefficient enum %d from FFmpeg; choosing Rec. 709.\n",
 168                                 colorspace);
 169                 }
 170                 format.luma_coefficients = YCBCR_REC_709;
 171                 break;
 172         }
 173         *last_colorspace = colorspace;
 174
 175         format.full_range = is_full_range(desc);
 176         format.num_levels = 1 << desc->comp[0].depth;
 177         format.chroma_subsampling_x = 1 << desc->log2_chroma_w;
 178         format.chroma_subsampling_y = 1 << desc->log2_chroma_h;
 179
 180         switch (frame->chroma_location) {
 181         case AVCHROMA_LOC_LEFT:
 182                 format.cb_x_position = 0.0;
 183                 format.cb_y_position = 0.5;
 184                 break;
 185         case AVCHROMA_LOC_CENTER:
 186                 format.cb_x_position = 0.5;
 187                 format.cb_y_position = 0.5;
 188                 break;
 189         case AVCHROMA_LOC_TOPLEFT:
 190                 format.cb_x_position = 0.0;
 191                 format.cb_y_position = 0.0;
 192                 break;
 193         case AVCHROMA_LOC_TOP:
 194                 format.cb_x_position = 0.5;
 195                 format.cb_y_position = 0.0;
 196                 break;
 197         case AVCHROMA_LOC_BOTTOMLEFT:
 198                 format.cb_x_position = 0.0;
 199                 format.cb_y_position = 1.0;
 200                 break;
 201         case AVCHROMA_LOC_BOTTOM:
 202                 format.cb_x_position = 0.5;
 203                 format.cb_y_position = 1.0;
 204                 break;
 205         default:
 206                 if (frame->chroma_location != *last_chroma_location) {
 207                         fprintf(stderr, "Unknown chroma location coefficient enum %d from FFmpeg; choosing center.\n",
 208                                 frame->chroma_location);
 209                 }
 210                 format.cb_x_position = 0.5;
 211                 format.cb_y_position = 0.5;
 212                 break;
 213         }
 214         *last_chroma_location = frame->chroma_location;
 215
 216         if (is_mjpeg && !format.full_range) {
 217                 // Limited-range MJPEG is only detected by FFmpeg whenever a special
 218                 // JPEG comment is set, which means that in practice, the stream is
 219                 // almost certainly generated by Futatabi. Override FFmpeg's forced
 220                 // MJPEG defaults (it disregards the values set in the mux) with what
 221                 // Futatabi sets.
 222                 format.luma_coefficients = YCBCR_REC_709;
 223                 format.cb_x_position = 0.0;
 224                 format.cb_y_position = 0.5;
 225         }
 226
 227         format.cr_x_position = format.cb_x_position;
 228         format.cr_y_position = format.cb_y_position;
 229         return format;
 230 }
 231
 232 RGBTriplet get_neutral_color(AVDictionary *metadata)
 233 {
 234         if (metadata == nullptr) {
 235                 return RGBTriplet(1.0f, 1.0f, 1.0f);
 236         }
 237         AVDictionaryEntry *entry = av_dict_get(metadata, "WhitePoint", nullptr, 0);
 238         if (entry == nullptr) {
 239                 return RGBTriplet(1.0f, 1.0f, 1.0f);
 240         }
 241
 242         unsigned x_nom, x_den, y_nom, y_den;
 243         if (sscanf(entry->value, " %u:%u , %u:%u", &x_nom, &x_den, &y_nom, &y_den) != 4) {
 244                 fprintf(stderr, "WARNING: Unable to parse white point '%s', using default white point\n", entry->value);
 245                 return RGBTriplet(1.0f, 1.0f, 1.0f);
 246         }
 247
 248         double x = double(x_nom) / x_den;
 249         double y = double(y_nom) / y_den;
 250         double z = 1.0 - x - y;
 251
 252         Matrix3d rgb_to_xyz_matrix = movit::ColorspaceConversionEffect::get_xyz_matrix(COLORSPACE_sRGB);
 253         Vector3d rgb = rgb_to_xyz_matrix.inverse() * Vector3d(x, y, z);
 254
 255         return RGBTriplet(rgb[0], rgb[1], rgb[2]);
 256 }
 257
 258 }  // namespace
 259
 260 FFmpegCapture::FFmpegCapture(const string &filename, unsigned width, unsigned height)
 261         : filename(filename), width(width), height(height), video_timebase{1, 1}
 262 {
 263         description = "Video: " + filename;
 264
 265         last_frame = steady_clock::now();
 266
 267         avformat_network_init();  // In case someone wants this.
 268 }
 269
 270 #ifdef HAVE_SRT
 271 FFmpegCapture::FFmpegCapture(int srt_sock, const string &stream_id)
 272         : srt_sock(srt_sock),
 273           width(0),  // Don't resize; SRT streams typically have stable resolution, and should behave much like regular cards in general.
 274           height(0),
 275           pixel_format(bmusb::PixelFormat_8BitYCbCrPlanar),
 276           video_timebase{1, 1}
 277 {
 278         if (stream_id.empty()) {
 279                 description = "SRT stream";
 280         } else {
 281                 description = stream_id;
 282         }
 283         play_as_fast_as_possible = true;
 284         play_once = true;
 285         last_frame = steady_clock::now();
 286 }
 287 #endif
 288
 289 FFmpegCapture::~FFmpegCapture()
 290 {
 291         if (has_dequeue_callbacks) {
 292                 dequeue_cleanup_callback();
 293         }
 294         swr_free(&resampler);
 295 #ifdef HAVE_SRT
 296         if (srt_sock != -1) {
 297                 srt_close(srt_sock);
 298         }
 299 #endif
 300 }
 301
 302 void FFmpegCapture::configure_card()
 303 {
 304         if (video_frame_allocator == nullptr) {
 305                 owned_video_frame_allocator.reset(new MallocFrameAllocator(FRAME_SIZE, NUM_QUEUED_VIDEO_FRAMES));
 306                 set_video_frame_allocator(owned_video_frame_allocator.get());
 307         }
 308         if (audio_frame_allocator == nullptr) {
 309                 // Audio can come out in pretty large chunks, so increase from the default 1 MB.
 310                 owned_audio_frame_allocator.reset(new MallocFrameAllocator(1 << 20, NUM_QUEUED_AUDIO_FRAMES));
 311                 set_audio_frame_allocator(owned_audio_frame_allocator.get());
 312         }
 313 }
 314
 315 void FFmpegCapture::start_bm_capture()
 316 {
 317         if (running) {
 318                 return;
 319         }
 320         running = true;
 321         producer_thread_should_quit.unquit();
 322         producer_thread = thread(&FFmpegCapture::producer_thread_func, this);
 323 }
 324
 325 void FFmpegCapture::stop_dequeue_thread()
 326 {
 327         if (!running) {
 328                 return;
 329         }
 330         running = false;
 331         producer_thread_should_quit.quit();
 332         producer_thread.join();
 333 }
 334
 335 std::map<uint32_t, VideoMode> FFmpegCapture::get_available_video_modes() const
 336 {
 337         // Note: This will never really be shown in the UI.
 338         VideoMode mode;
 339
 340         char buf[256];
 341         snprintf(buf, sizeof(buf), "%ux%u", sws_last_width, sws_last_height);
 342         mode.name = buf;
 343
 344         mode.autodetect = false;
 345         mode.width = sws_last_width;
 346         mode.height = sws_last_height;
 347         mode.frame_rate_num = 60;
 348         mode.frame_rate_den = 1;
 349         mode.interlaced = false;
 350
 351         return {{ 0, mode }};
 352 }
 353
 354 void FFmpegCapture::producer_thread_func()
 355 {
 356         char thread_name[16];
 357         snprintf(thread_name, sizeof(thread_name), "FFmpeg_C_%d", card_index);
 358         pthread_setname_np(pthread_self(), thread_name);
 359
 360         while (!producer_thread_should_quit.should_quit()) {
 361                 string filename_copy;
 362                 {
 363                         lock_guard<mutex> lock(filename_mu);
 364                         filename_copy = filename;
 365                 }
 366
 367                 string pathname;
 368                 if (srt_sock == -1) {
 369                         pathname = search_for_file(filename_copy);
 370                 } else {
 371                         pathname = description;
 372                 }
 373                 if (pathname.empty()) {
 374                         send_disconnected_frame();
 375                         if (play_once) {
 376                                 break;
 377                         }
 378                         producer_thread_should_quit.sleep_for(seconds(1));
 379                         fprintf(stderr, "%s not found, sleeping one second and trying again...\n", filename_copy.c_str());
 380                         continue;
 381                 }
 382                 should_interrupt = false;
 383                 if (!play_video(pathname)) {
 384                         // Error.
 385                         send_disconnected_frame();
 386                         if (play_once) {
 387                                 break;
 388                         }
 389                         fprintf(stderr, "Error when playing %s, sleeping one second and trying again...\n", pathname.c_str());
 390                         producer_thread_should_quit.sleep_for(seconds(1));
 391                         continue;
 392                 }
 393
 394                 if (play_once) {
 395                         send_disconnected_frame();
 396                         break;
 397                 }
 398
 399                 // Probably just EOF, will exit the loop above on next test.
 400         }
 401
 402         if (has_dequeue_callbacks) {
 403                 dequeue_cleanup_callback();
 404                 has_dequeue_callbacks = false;
 405         }
 406 }
 407
 408 void FFmpegCapture::send_disconnected_frame()
 409 {
 410         // Send an empty frame to signal that we have no signal anymore.
 411         FrameAllocator::Frame video_frame = video_frame_allocator->alloc_frame();
 412         size_t frame_width = width == 0 ? global_flags.width : width;
 413         size_t frame_height = height == 0 ? global_flags.height : height;
 414         if (video_frame.data) {
 415                 VideoFormat video_format;
 416                 video_format.width = frame_width;
 417                 video_format.height = frame_height;
 418                 video_format.frame_rate_nom = 60;
 419                 video_format.frame_rate_den = 1;
 420                 video_format.is_connected = false;
 421                 if (pixel_format == bmusb::PixelFormat_8BitBGRA) {
 422                         video_format.stride = frame_width * 4;
 423                         video_frame.len = frame_width * frame_height * 4;
 424                         memset(video_frame.data, 0, video_frame.len);
 425                 } else {
 426                         video_format.stride = frame_width;
 427                         current_frame_ycbcr_format.luma_coefficients = YCBCR_REC_709;
 428                         current_frame_ycbcr_format.full_range = true;
 429                         current_frame_ycbcr_format.num_levels = 256;
 430                         current_frame_ycbcr_format.chroma_subsampling_x = 2;
 431                         current_frame_ycbcr_format.chroma_subsampling_y = 2;
 432                         current_frame_ycbcr_format.cb_x_position = 0.0f;
 433                         current_frame_ycbcr_format.cb_y_position = 0.0f;
 434                         current_frame_ycbcr_format.cr_x_position = 0.0f;
 435                         current_frame_ycbcr_format.cr_y_position = 0.0f;
 436                         video_frame.len = frame_width * frame_height * 2;
 437                         memset(video_frame.data, 0, frame_width * frame_height);
 438                         memset(video_frame.data + frame_width * frame_height, 128, frame_width * frame_height);  // Valid for both NV12 and planar.
 439                 }
 440
 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());
 444                 last_frame_was_connected = false;
 445         }
 446
 447         if (play_once) {
 448                 disconnected = true;
 449                 if (card_disconnected_callback != nullptr) {
 450                         card_disconnected_callback();
 451                 }
 452         }
 453 }
 454
 455 AVPixelFormat get_vaapi_hw_format(AVCodecContext *ctx, const AVPixelFormat *fmt)
 456 {
 457         for (const AVPixelFormat *fmt_ptr = fmt; *fmt_ptr != -1; ++fmt_ptr) {
 458                 for (int i = 0;; ++i) {  // Termination condition inside loop.
 459                         const AVCodecHWConfig *config = avcodec_get_hw_config(ctx->codec, i);
 460                         if (config == nullptr) {  // End of list.
 461                                 fprintf(stderr, "Decoder %s does not support device.\n", ctx->codec->name);
 462                                 break;
 463                         }
 464                         if (config->methods & AV_CODEC_HW_CONFIG_METHOD_HW_DEVICE_CTX &&
 465                             config->device_type == AV_HWDEVICE_TYPE_VAAPI &&
 466                             config->pix_fmt == *fmt_ptr) {
 467                                 return config->pix_fmt;
 468                         }
 469                 }
 470         }
 471
 472         // We found no VA-API formats, so take the best software format.
 473         return fmt[0];
 474 }
 475
 476 bool FFmpegCapture::play_video(const string &pathname)
 477 {
 478         // Note: Call before open, not after; otherwise, there's a race.
 479         // (There is now, too, but it tips the correct way. We could use fstat()
 480         // if we had the file descriptor.)
 481         timespec last_modified;
 482         struct stat buf;
 483         if (stat(pathname.c_str(), &buf) != 0) {
 484                 // Probably some sort of protocol, so can't stat.
 485                 last_modified.tv_sec = -1;
 486         } else {
 487                 last_modified = buf.st_mtim;
 488         }
 489         last_colorspace = static_cast<AVColorSpace>(-1);
 490         last_chroma_location = static_cast<AVChromaLocation>(-1);
 491
 492         AVFormatContextWithCloser format_ctx;
 493         if (srt_sock == -1) {
 494                 // Regular file.
 495                 format_ctx = avformat_open_input_unique(pathname.c_str(), /*fmt=*/nullptr,
 496                         /*options=*/nullptr,
 497                         AVIOInterruptCB{ &FFmpegCapture::interrupt_cb_thunk, this });
 498         } else {
 499 #ifdef HAVE_SRT
 500                 // SRT socket, already opened.
 501                 AVInputFormat *mpegts_fmt = av_find_input_format("mpegts");
 502                 format_ctx = avformat_open_input_unique(&FFmpegCapture::read_srt_thunk, this,
 503                         mpegts_fmt, /*options=*/nullptr,
 504                         AVIOInterruptCB{ &FFmpegCapture::interrupt_cb_thunk, this });
 505 #else
 506                 assert(false);
 507 #endif
 508         }
 509         if (format_ctx == nullptr) {
 510                 fprintf(stderr, "%s: Error opening file\n", pathname.c_str());
 511                 return false;
 512         }
 513
 514         if (avformat_find_stream_info(format_ctx.get(), nullptr) < 0) {
 515                 fprintf(stderr, "%s: Error finding stream info\n", pathname.c_str());
 516                 return false;
 517         }
 518
 519         int video_stream_index = find_stream_index(format_ctx.get(), AVMEDIA_TYPE_VIDEO);
 520         if (video_stream_index == -1) {
 521                 fprintf(stderr, "%s: No video stream found\n", pathname.c_str());
 522                 return false;
 523         }
 524
 525         int audio_stream_index = find_stream_index(format_ctx.get(), AVMEDIA_TYPE_AUDIO);
 526         int subtitle_stream_index = find_stream_index(format_ctx.get(), AVMEDIA_TYPE_SUBTITLE);
 527         has_last_subtitle = false;
 528
 529         // Open video decoder.
 530         const AVCodecParameters *video_codecpar = format_ctx->streams[video_stream_index]->codecpar;
 531         AVCodec *video_codec = avcodec_find_decoder(video_codecpar->codec_id);
 532
 533         video_timebase = format_ctx->streams[video_stream_index]->time_base;
 534         AVCodecContextWithDeleter video_codec_ctx = avcodec_alloc_context3_unique(nullptr);
 535         if (avcodec_parameters_to_context(video_codec_ctx.get(), video_codecpar) < 0) {
 536                 fprintf(stderr, "%s: Cannot fill video codec parameters\n", pathname.c_str());
 537                 return false;
 538         }
 539         if (video_codec == nullptr) {
 540                 fprintf(stderr, "%s: Cannot find video decoder\n", pathname.c_str());
 541                 return false;
 542         }
 543
 544         // Seemingly, it's not too easy to make something that just initializes
 545         // “whatever goes”, so we don't get VDPAU or CUDA here without enumerating
 546         // through several different types. VA-API will do for now.
 547         AVBufferRef *hw_device_ctx = nullptr;
 548         if (av_hwdevice_ctx_create(&hw_device_ctx, AV_HWDEVICE_TYPE_VAAPI, nullptr, nullptr, 0) < 0) {
 549                 fprintf(stderr, "Failed to initialize VA-API for FFmpeg acceleration. Decoding video in software.\n");
 550         } else {
 551                 video_codec_ctx->hw_device_ctx = av_buffer_ref(hw_device_ctx);
 552                 video_codec_ctx->get_format = get_vaapi_hw_format;
 553         }
 554
 555         if (avcodec_open2(video_codec_ctx.get(), video_codec, nullptr) < 0) {
 556                 fprintf(stderr, "%s: Cannot open video decoder\n", pathname.c_str());
 557                 return false;
 558         }
 559         unique_ptr<AVCodecContext, decltype(avcodec_close)*> video_codec_ctx_cleanup(
 560                 video_codec_ctx.get(), avcodec_close);
 561
 562         // Used in decode_ycbcr_format().
 563         is_mjpeg = video_codecpar->codec_id == AV_CODEC_ID_MJPEG;
 564
 565         // Open audio decoder, if we have audio.
 566         AVCodecContextWithDeleter audio_codec_ctx;
 567         if (audio_stream_index != -1) {
 568                 audio_codec_ctx = avcodec_alloc_context3_unique(nullptr);
 569                 const AVCodecParameters *audio_codecpar = format_ctx->streams[audio_stream_index]->codecpar;
 570                 audio_timebase = format_ctx->streams[audio_stream_index]->time_base;
 571                 if (avcodec_parameters_to_context(audio_codec_ctx.get(), audio_codecpar) < 0) {
 572                         fprintf(stderr, "%s: Cannot fill audio codec parameters\n", pathname.c_str());
 573                         return false;
 574                 }
 575                 AVCodec *audio_codec = avcodec_find_decoder(audio_codecpar->codec_id);
 576                 if (audio_codec == nullptr) {
 577                         fprintf(stderr, "%s: Cannot find audio decoder\n", pathname.c_str());
 578                         return false;
 579                 }
 580                 if (avcodec_open2(audio_codec_ctx.get(), audio_codec, nullptr) < 0) {
 581                         fprintf(stderr, "%s: Cannot open audio decoder\n", pathname.c_str());
 582                         return false;
 583                 }
 584         }
 585         unique_ptr<AVCodecContext, decltype(avcodec_close)*> audio_codec_ctx_cleanup(
 586                 audio_codec_ctx.get(), avcodec_close);
 587
 588         internal_rewind();
 589
 590         // Main loop.
 591         bool first_frame = true;
 592         while (!producer_thread_should_quit.should_quit()) {
 593                 if (process_queued_commands(format_ctx.get(), pathname, last_modified, /*rewound=*/nullptr)) {
 594                         return true;
 595                 }
 596                 if (should_interrupt.load()) {
 597                         // Check as a failsafe, so that we don't need to rely on avio if we don't have to.
 598                         return false;
 599                 }
 600                 UniqueFrame audio_frame = audio_frame_allocator->alloc_frame();
 601                 AudioFormat audio_format;
 602
 603                 int64_t audio_pts;
 604                 bool error;
 605                 AVFrameWithDeleter frame = decode_frame(format_ctx.get(), video_codec_ctx.get(), audio_codec_ctx.get(),
 606                         pathname, video_stream_index, audio_stream_index, subtitle_stream_index, audio_frame.get(), &audio_format, &audio_pts, &error);
 607                 if (error) {
 608                         return false;
 609                 }
 610                 if (frame == nullptr) {
 611                         // EOF. Loop back to the start if we can.
 612                         if (format_ctx->pb != nullptr && format_ctx->pb->seekable == 0) {
 613                                 // Not seekable (but seemingly, sometimes av_seek_frame() would return 0 anyway,
 614                                 // so don't try).
 615                                 return true;
 616                         }
 617                         if (av_seek_frame(format_ctx.get(), /*stream_index=*/-1, /*timestamp=*/0, /*flags=*/0) < 0) {
 618                                 fprintf(stderr, "%s: Rewind failed, not looping.\n", pathname.c_str());
 619                                 return true;
 620                         }
 621                         if (video_codec_ctx != nullptr) {
 622                                 avcodec_flush_buffers(video_codec_ctx.get());
 623                         }
 624                         if (audio_codec_ctx != nullptr) {
 625                                 avcodec_flush_buffers(audio_codec_ctx.get());
 626                         }
 627                         // If the file has changed since last time, return to get it reloaded.
 628                         // Note that depending on how you move the file into place, you might
 629                         // end up corrupting the one you're already playing, so this path
 630                         // might not trigger.
 631                         if (changed_since(pathname, last_modified)) {
 632                                 return true;
 633                         }
 634                         internal_rewind();
 635                         continue;
 636                 }
 637
 638                 VideoFormat video_format = construct_video_format(frame.get(), video_timebase);
 639                 if (video_format.frame_rate_nom == 0 || video_format.frame_rate_den == 0) {
 640                         // Invalid frame rate; try constructing it from the previous frame length.
 641                         // (This is especially important if we are the master card, for SRT,
 642                         // since it affects audio. Not all senders have good timebases
 643                         // (e.g., Larix rounds first to timebase 1000 and then multiplies by
 644                         // 90 from there, it seems), but it's much better to have an oscillating
 645                         // value than just locking at 60.
 646                         if (last_pts != 0 && frame->pts > last_pts) {
 647                                 int64_t pts_diff = frame->pts - last_pts;
 648                                 video_format.frame_rate_nom = video_timebase.den;
 649                                 video_format.frame_rate_den = video_timebase.num * pts_diff;
 650                         } else {
 651                                 video_format.frame_rate_nom = 60;
 652                                 video_format.frame_rate_den = 1;
 653                         }
 654                 }
 655                 UniqueFrame video_frame = make_video_frame(frame.get(), pathname, &error);
 656                 if (error) {
 657                         return false;
 658                 }
 659
 660                 for ( ;; ) {
 661                         if (last_pts == 0 && pts_origin == 0) {
 662                                 pts_origin = frame->pts;
 663                         }
 664                         steady_clock::time_point now = steady_clock::now();
 665                         if (play_as_fast_as_possible) {
 666                                 video_frame->received_timestamp = now;
 667                                 audio_frame->received_timestamp = now;
 668                                 next_frame_start = now;
 669                         } else {
 670                                 next_frame_start = compute_frame_start(frame->pts, pts_origin, video_timebase, start, rate);
 671                                 if (first_frame && last_frame_was_connected) {
 672                                         // If reconnect took more than one second, this is probably a live feed,
 673                                         // and we should reset the resampler. (Or the rate is really, really low,
 674                                         // in which case a reset on the first frame is fine anyway.)
 675                                         if (duration<double>(next_frame_start - last_frame).count() >= 1.0) {
 676                                                 last_frame_was_connected = false;
 677                                         }
 678                                 }
 679                                 video_frame->received_timestamp = next_frame_start;
 680
 681                                 // The easiest way to get all the rate conversions etc. right is to move the
 682                                 // audio PTS into the video PTS timebase and go from there. (We'll get some
 683                                 // rounding issues, but they should not be a big problem.)
 684                                 int64_t audio_pts_as_video_pts = av_rescale_q(audio_pts, audio_timebase, video_timebase);
 685                                 audio_frame->received_timestamp = compute_frame_start(audio_pts_as_video_pts, pts_origin, video_timebase, start, rate);
 686
 687                                 if (audio_frame->len != 0) {
 688                                         // The received timestamps in Nageru are measured after we've just received the frame.
 689                                         // However, pts (especially audio pts) is at the _beginning_ of the frame.
 690                                         // If we have locked audio, the distinction doesn't really matter, as pts is
 691                                         // on a relative scale and a fixed offset is fine. But if we don't, we will have
 692                                         // a different number of samples each time, which will cause huge audio jitter
 693                                         // and throw off the resampler.
 694                                         //
 695                                         // In a sense, we should have compensated by adding the frame and audio lengths
 696                                         // to video_frame->received_timestamp and audio_frame->received_timestamp respectively,
 697                                         // but that would mean extra waiting in sleep_until(). All we need is that they
 698                                         // are correct relative to each other, though (and to the other frames we send),
 699                                         // so just align the end of the audio frame, and we're fine.
 700                                         size_t num_samples = (audio_frame->len * 8) / audio_format.bits_per_sample / audio_format.num_channels;
 701                                         double offset = double(num_samples) / OUTPUT_FREQUENCY -
 702                                                 double(video_format.frame_rate_den) / video_format.frame_rate_nom;
 703                                         audio_frame->received_timestamp += duration_cast<steady_clock::duration>(duration<double>(offset));
 704                                 }
 705
 706                                 if (duration<double>(now - next_frame_start).count() >= 0.1) {
 707                                         // If we don't have enough CPU to keep up, or if we have a live stream
 708                                         // where the initial origin was somehow wrong, we could be behind indefinitely.
 709                                         // In particular, this will give the audio resampler problems as it tries
 710                                         // to speed up to reduce the delay, hitting the low end of the buffer every time.
 711                                         fprintf(stderr, "%s: Playback %.0f ms behind, resetting time scale\n",
 712                                                 pathname.c_str(),
 713                                                 1e3 * duration<double>(now - next_frame_start).count());
 714                                         pts_origin = frame->pts;
 715                                         start = next_frame_start = now;
 716                                         timecode += MAX_FPS * 2 + 1;
 717                                 }
 718                         }
 719                         bool finished_wakeup;
 720                         if (play_as_fast_as_possible) {
 721                                 finished_wakeup = !producer_thread_should_quit.should_quit();
 722                         } else {
 723                                 finished_wakeup = producer_thread_should_quit.sleep_until(next_frame_start);
 724                         }
 725                         if (finished_wakeup) {
 726                                 if (audio_frame->len > 0) {
 727                                         assert(audio_pts != -1);
 728                                 }
 729                                 if (!last_frame_was_connected) {
 730                                         // We're recovering from an error (or really slow load, see above).
 731                                         // Make sure to get the audio resampler reset. (This is a hack;
 732                                         // ideally, the frame callback should just accept a way to signal
 733                                         // audio discontinuity.)
 734                                         timecode += MAX_FPS * 2 + 1;
 735                                 }
 736                                 last_neutral_color = get_neutral_color(frame->metadata);
 737                                 frame_callback(frame->pts, video_timebase, audio_pts, audio_timebase, timecode++,
 738                                         video_frame.get_and_release(), 0, video_format,
 739                                         audio_frame.get_and_release(), 0, audio_format);
 740                                 first_frame = false;
 741                                 last_frame = steady_clock::now();
 742                                 last_frame_was_connected = true;
 743                                 break;
 744                         } else {
 745                                 if (producer_thread_should_quit.should_quit()) break;
 746
 747                                 bool rewound = false;
 748                                 if (process_queued_commands(format_ctx.get(), pathname, last_modified, &rewound)) {
 749                                         return true;
 750                                 }
 751                                 // If we just rewound, drop this frame on the floor and be done.
 752                                 if (rewound) {
 753                                         break;
 754                                 }
 755                                 // OK, we didn't, so probably a rate change. Recalculate next_frame_start,
 756                                 // but if it's now in the past, we'll reset the origin, so that we don't
 757                                 // generate a huge backlog of frames that we need to run through quickly.
 758                                 next_frame_start = compute_frame_start(frame->pts, pts_origin, video_timebase, start, rate);
 759                                 steady_clock::time_point now = steady_clock::now();
 760                                 if (next_frame_start < now) {
 761                                         pts_origin = frame->pts;
 762                                         start = next_frame_start = now;
 763                                 }
 764                         }
 765                 }
 766                 last_pts = frame->pts;
 767         }
 768         return true;
 769 }
 770
 771 void FFmpegCapture::internal_rewind()
 772 {
 773         pts_origin = last_pts = 0;
 774         start = next_frame_start = steady_clock::now();
 775 }
 776
 777 bool FFmpegCapture::process_queued_commands(AVFormatContext *format_ctx, const std::string &pathname, timespec last_modified, bool *rewound)
 778 {
 779         // Process any queued commands from other threads.
 780         vector<QueuedCommand> commands;
 781         {
 782                 lock_guard<mutex> lock(queue_mu);
 783                 swap(commands, command_queue);
 784         }
 785         for (const QueuedCommand &cmd : commands) {
 786                 switch (cmd.command) {
 787                 case QueuedCommand::REWIND:
 788                         if (av_seek_frame(format_ctx, /*stream_index=*/-1, /*timestamp=*/0, /*flags=*/0) < 0) {
 789                                 fprintf(stderr, "%s: Rewind failed, stopping play.\n", pathname.c_str());
 790                         }
 791                         // If the file has changed since last time, return to get it reloaded.
 792                         // Note that depending on how you move the file into place, you might
 793                         // end up corrupting the one you're already playing, so this path
 794                         // might not trigger.
 795                         if (changed_since(pathname, last_modified)) {
 796                                 return true;
 797                         }
 798                         internal_rewind();
 799                         if (rewound != nullptr) {
 800                                 *rewound = true;
 801                         }
 802                         break;
 803
 804                 case QueuedCommand::CHANGE_RATE:
 805                         // Change the origin to the last played frame.
 806                         start = compute_frame_start(last_pts, pts_origin, video_timebase, start, rate);
 807                         pts_origin = last_pts;
 808                         rate = cmd.new_rate;
 809                         play_as_fast_as_possible = (rate >= 10.0);
 810                         break;
 811                 }
 812         }
 813         return false;
 814 }
 815
 816 namespace {
 817
 818 }  // namespace
 819
 820 AVFrameWithDeleter FFmpegCapture::decode_frame(AVFormatContext *format_ctx, AVCodecContext *video_codec_ctx, AVCodecContext *audio_codec_ctx,
 821         const std::string &pathname, int video_stream_index, int audio_stream_index, int subtitle_stream_index,
 822         FrameAllocator::Frame *audio_frame, AudioFormat *audio_format, int64_t *audio_pts, bool *error)
 823 {
 824         *error = false;
 825
 826         // Read packets until we have a frame or there are none left.
 827         bool frame_finished = false;
 828         AVFrameWithDeleter audio_avframe = av_frame_alloc_unique();
 829         AVFrameWithDeleter video_avframe = av_frame_alloc_unique();
 830         bool eof = false;
 831         *audio_pts = -1;
 832         bool has_audio = false;
 833         do {
 834                 AVPacket pkt;
 835                 unique_ptr<AVPacket, decltype(av_packet_unref)*> pkt_cleanup(
 836                         &pkt, av_packet_unref);
 837                 av_init_packet(&pkt);
 838                 pkt.data = nullptr;
 839                 pkt.size = 0;
 840                 if (av_read_frame(format_ctx, &pkt) == 0) {
 841                         if (pkt.stream_index == audio_stream_index && audio_callback != nullptr) {
 842                                 audio_callback(&pkt, format_ctx->streams[audio_stream_index]->time_base);
 843                         }
 844                         if (pkt.stream_index == video_stream_index) {
 845                                 if (avcodec_send_packet(video_codec_ctx, &pkt) < 0) {
 846                                         fprintf(stderr, "%s: Cannot send packet to video codec.\n", pathname.c_str());
 847                                         *error = true;
 848                                         return AVFrameWithDeleter(nullptr);
 849                                 }
 850                         } else if (pkt.stream_index == audio_stream_index) {
 851                                 has_audio = true;
 852                                 if (avcodec_send_packet(audio_codec_ctx, &pkt) < 0) {
 853                                         fprintf(stderr, "%s: Cannot send packet to audio codec.\n", pathname.c_str());
 854                                         *error = true;
 855                                         return AVFrameWithDeleter(nullptr);
 856                                 }
 857                         } else if (pkt.stream_index == subtitle_stream_index) {
 858                                 last_subtitle = string(reinterpret_cast<const char *>(pkt.data), pkt.size);
 859                                 has_last_subtitle = true;
 860                         }
 861                 } else {
 862                         eof = true;  // Or error, but ignore that for the time being.
 863                 }
 864
 865                 // Decode audio, if any.
 866                 if (has_audio) {
 867                         for ( ;; ) {
 868                                 int err = avcodec_receive_frame(audio_codec_ctx, audio_avframe.get());
 869                                 if (err == 0) {
 870                                         if (*audio_pts == -1) {
 871                                                 *audio_pts = audio_avframe->pts;
 872                                         }
 873                                         convert_audio(audio_avframe.get(), audio_frame, audio_format);
 874                                 } else if (err == AVERROR(EAGAIN)) {
 875                                         break;
 876                                 } else {
 877                                         fprintf(stderr, "%s: Cannot receive frame from audio codec.\n", pathname.c_str());
 878                                         *error = true;
 879                                         return AVFrameWithDeleter(nullptr);
 880                                 }
 881                         }
 882                 }
 883
 884                 // Decode video, if we have a frame.
 885                 int err = avcodec_receive_frame(video_codec_ctx, video_avframe.get());
 886                 if (err == 0) {
 887                         if (video_avframe->format == AV_PIX_FMT_VAAPI) {
 888                                 // Get the frame down to the CPU. (TODO: See if we can keep it
 889                                 // on the GPU all the way, since it will be going up again later.
 890                                 // However, this only works if the OpenGL GPU is the same one.)
 891                                 AVFrameWithDeleter sw_frame = av_frame_alloc_unique();
 892                                 int err = av_hwframe_transfer_data(sw_frame.get(), video_avframe.get(), 0);
 893                                 if (err != 0) {
 894                                         fprintf(stderr, "%s: Cannot transfer hardware video frame to software.\n", pathname.c_str());
 895                                         *error = true;
 896                                         return AVFrameWithDeleter(nullptr);
 897                                 }
 898                                 video_avframe = move(sw_frame);
 899                         }
 900                         frame_finished = true;
 901                         break;
 902                 } else if (err != AVERROR(EAGAIN)) {
 903                         fprintf(stderr, "%s: Cannot receive frame from video codec.\n", pathname.c_str());
 904                         *error = true;
 905                         return AVFrameWithDeleter(nullptr);
 906                 }
 907         } while (!eof);
 908
 909         if (frame_finished)
 910                 return video_avframe;
 911         else
 912                 return AVFrameWithDeleter(nullptr);
 913 }
 914
 915 void FFmpegCapture::convert_audio(const AVFrame *audio_avframe, FrameAllocator::Frame *audio_frame, AudioFormat *audio_format)
 916 {
 917         // Decide on a format. If there already is one in this audio frame,
 918         // we're pretty much forced to use it. If not, we try to find an exact match.
 919         // If that still doesn't work, we default to 32-bit signed chunked
 920         // (float would be nice, but there's really no way to signal that yet).
 921         AVSampleFormat dst_format;
 922         if (audio_format->bits_per_sample == 0) {
 923                 switch (audio_avframe->format) {
 924                 case AV_SAMPLE_FMT_S16:
 925                 case AV_SAMPLE_FMT_S16P:
 926                         audio_format->bits_per_sample = 16;
 927                         dst_format = AV_SAMPLE_FMT_S16;
 928                         break;
 929                 case AV_SAMPLE_FMT_S32:
 930                 case AV_SAMPLE_FMT_S32P:
 931                 default:
 932                         audio_format->bits_per_sample = 32;
 933                         dst_format = AV_SAMPLE_FMT_S32;
 934                         break;
 935                 }
 936         } else if (audio_format->bits_per_sample == 16) {
 937                 dst_format = AV_SAMPLE_FMT_S16;
 938         } else if (audio_format->bits_per_sample == 32) {
 939                 dst_format = AV_SAMPLE_FMT_S32;
 940         } else {
 941                 assert(false);
 942         }
 943         audio_format->num_channels = 2;
 944
 945         int64_t channel_layout = audio_avframe->channel_layout;
 946         if (channel_layout == 0) {
 947                 channel_layout = av_get_default_channel_layout(audio_avframe->channels);
 948         }
 949
 950         if (resampler == nullptr ||
 951             audio_avframe->format != last_src_format ||
 952             dst_format != last_dst_format ||
 953             channel_layout != last_channel_layout ||
 954             audio_avframe->sample_rate != last_sample_rate) {
 955                 swr_free(&resampler);
 956                 resampler = swr_alloc_set_opts(nullptr,
 957                                                /*out_ch_layout=*/AV_CH_LAYOUT_STEREO_DOWNMIX,
 958                                                /*out_sample_fmt=*/dst_format,
 959                                                /*out_sample_rate=*/OUTPUT_FREQUENCY,
 960                                                /*in_ch_layout=*/channel_layout,
 961                                                /*in_sample_fmt=*/AVSampleFormat(audio_avframe->format),
 962                                                /*in_sample_rate=*/audio_avframe->sample_rate,
 963                                                /*log_offset=*/0,
 964                                                /*log_ctx=*/nullptr);
 965
 966                 if (resampler == nullptr) {
 967                         fprintf(stderr, "Allocating resampler failed.\n");
 968                         abort();
 969                 }
 970
 971                 if (swr_init(resampler) < 0) {
 972                         fprintf(stderr, "Could not open resample context.\n");
 973                         abort();
 974                 }
 975
 976                 last_src_format = AVSampleFormat(audio_avframe->format);
 977                 last_dst_format = dst_format;
 978                 last_channel_layout = channel_layout;
 979                 last_sample_rate = audio_avframe->sample_rate;
 980         }
 981
 982         size_t bytes_per_sample = (audio_format->bits_per_sample / 8) * 2;
 983         size_t num_samples_room = (audio_frame->size - audio_frame->len) / bytes_per_sample;
 984
 985         uint8_t *data = audio_frame->data + audio_frame->len;
 986         int out_samples = swr_convert(resampler, &data, num_samples_room,
 987                 const_cast<const uint8_t **>(audio_avframe->data), audio_avframe->nb_samples);
 988         if (out_samples < 0) {
 989                 fprintf(stderr, "Audio conversion failed.\n");
 990                 abort();
 991         }
 992
 993         audio_frame->len += out_samples * bytes_per_sample;
 994 }
 995
 996 VideoFormat FFmpegCapture::construct_video_format(const AVFrame *frame, AVRational video_timebase)
 997 {
 998         VideoFormat video_format;
 999         video_format.width = frame_width(frame);
1000         video_format.height = frame_height(frame);
1001         if (pixel_format == bmusb::PixelFormat_8BitBGRA) {
1002                 video_format.stride = frame_width(frame) * 4;
1003         } else if (pixel_format == FFmpegCapture::PixelFormat_NV12) {
1004                 video_format.stride = frame_width(frame);
1005         } else {
1006                 assert(pixel_format == bmusb::PixelFormat_8BitYCbCrPlanar);
1007                 video_format.stride = frame_width(frame);
1008         }
1009         video_format.frame_rate_nom = video_timebase.den;
1010         video_format.frame_rate_den = frame->pkt_duration * video_timebase.num;
1011         video_format.has_signal = true;
1012         video_format.is_connected = true;
1013         return video_format;
1014 }
1015
1016 UniqueFrame FFmpegCapture::make_video_frame(const AVFrame *frame, const string &pathname, bool *error)
1017 {
1018         *error = false;
1019
1020         UniqueFrame video_frame(video_frame_allocator->alloc_frame());
1021         if (video_frame->data == nullptr) {
1022                 return video_frame;
1023         }
1024
1025         if (sws_ctx == nullptr ||
1026             sws_last_width != frame->width ||
1027             sws_last_height != frame->height ||
1028             sws_last_src_format != frame->format) {
1029                 sws_dst_format = decide_dst_format(AVPixelFormat(frame->format), pixel_format);
1030                 sws_ctx.reset(
1031                         sws_getContext(frame->width, frame->height, AVPixelFormat(frame->format),
1032                                 frame_width(frame), frame_height(frame), sws_dst_format,
1033                                 SWS_BICUBIC, nullptr, nullptr, nullptr));
1034                 sws_last_width = frame->width;
1035                 sws_last_height = frame->height;
1036                 sws_last_src_format = frame->format;
1037         }
1038         if (sws_ctx == nullptr) {
1039                 fprintf(stderr, "%s: Could not create scaler context\n", pathname.c_str());
1040                 *error = true;
1041                 return video_frame;
1042         }
1043
1044         uint8_t *pic_data[4] = { nullptr, nullptr, nullptr, nullptr };
1045         int linesizes[4] = { 0, 0, 0, 0 };
1046         if (pixel_format == bmusb::PixelFormat_8BitBGRA) {
1047                 pic_data[0] = video_frame->data;
1048                 linesizes[0] = frame_width(frame) * 4;
1049                 video_frame->len = (frame_width(frame) * 4) * frame_height(frame);
1050         } else if (pixel_format == PixelFormat_NV12) {
1051                 pic_data[0] = video_frame->data;
1052                 linesizes[0] = frame_width(frame);
1053
1054                 pic_data[1] = pic_data[0] + frame_width(frame) * frame_height(frame);
1055                 linesizes[1] = frame_width(frame);
1056
1057                 video_frame->len = (frame_width(frame) * 2) * frame_height(frame);
1058
1059                 const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(sws_dst_format);
1060                 current_frame_ycbcr_format = decode_ycbcr_format(desc, frame, is_mjpeg, &last_colorspace, &last_chroma_location);
1061         } else {
1062                 assert(pixel_format == bmusb::PixelFormat_8BitYCbCrPlanar);
1063                 const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(sws_dst_format);
1064
1065                 int chroma_width = AV_CEIL_RSHIFT(int(frame_width(frame)), desc->log2_chroma_w);
1066                 int chroma_height = AV_CEIL_RSHIFT(int(frame_height(frame)), desc->log2_chroma_h);
1067
1068                 pic_data[0] = video_frame->data;
1069                 linesizes[0] = frame_width(frame);
1070
1071                 pic_data[1] = pic_data[0] + frame_width(frame) * frame_height(frame);
1072                 linesizes[1] = chroma_width;
1073
1074                 pic_data[2] = pic_data[1] + chroma_width * chroma_height;
1075                 linesizes[2] = chroma_width;
1076
1077                 video_frame->len = frame_width(frame) * frame_height(frame) + 2 * chroma_width * chroma_height;
1078
1079                 current_frame_ycbcr_format = decode_ycbcr_format(desc, frame, is_mjpeg, &last_colorspace, &last_chroma_location);
1080         }
1081         sws_scale(sws_ctx.get(), frame->data, frame->linesize, 0, frame->height, pic_data, linesizes);
1082
1083         return video_frame;
1084 }
1085
1086 int FFmpegCapture::interrupt_cb_thunk(void *opaque)
1087 {
1088         return reinterpret_cast<FFmpegCapture *>(opaque)->interrupt_cb();
1089 }
1090
1091 int FFmpegCapture::interrupt_cb()
1092 {
1093         return should_interrupt.load();
1094 }
1095
1096 unsigned FFmpegCapture::frame_width(const AVFrame *frame) const
1097 {
1098         if (width == 0) {
1099                 return frame->width;
1100         } else {
1101                 return width;
1102         }
1103 }
1104
1105 unsigned FFmpegCapture::frame_height(const AVFrame *frame) const
1106 {
1107         if (height == 0) {
1108                 return frame->height;
1109         } else {
1110                 return width;
1111         }
1112 }
1113
1114 #ifdef HAVE_SRT
1115 int FFmpegCapture::read_srt_thunk(void *opaque, uint8_t *buf, int buf_size)
1116 {
1117         return reinterpret_cast<FFmpegCapture *>(opaque)->read_srt(buf, buf_size);
1118 }
1119
1120 int FFmpegCapture::read_srt(uint8_t *buf, int buf_size)
1121 {
1122         SRT_MSGCTRL mc = srt_msgctrl_default;
1123         return srt_recvmsg2(srt_sock, reinterpret_cast<char *>(buf), buf_size, &mc);
1124 }
1125 #endif