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                 if (frame_callback != nullptr) {
 442                         frame_callback(-1, AVRational{1, TIMEBASE}, -1, AVRational{1, TIMEBASE}, timecode++,
 443                                 video_frame, /*video_offset=*/0, video_format,
 444                                 FrameAllocator::Frame(), /*audio_offset=*/0, AudioFormat());
 445                 }
 446                 last_frame_was_connected = false;
 447         }
 448
 449         if (play_once) {
 450                 disconnected = true;
 451                 if (card_disconnected_callback != nullptr) {
 452                         card_disconnected_callback();
 453                 }
 454         }
 455 }
 456
 457 template<AVHWDeviceType type>
 458 AVPixelFormat get_hw_format(AVCodecContext *ctx, const AVPixelFormat *fmt)
 459 {
 460         for (const AVPixelFormat *fmt_ptr = fmt; *fmt_ptr != -1; ++fmt_ptr) {
 461                 for (int i = 0;; ++i) {  // Termination condition inside loop.
 462                         const AVCodecHWConfig *config = avcodec_get_hw_config(ctx->codec, i);
 463                         if (config == nullptr) {  // End of list.
 464                                 fprintf(stderr, "Decoder %s does not support device.\n", ctx->codec->name);
 465                                 break;
 466                         }
 467                         if (config->methods & AV_CODEC_HW_CONFIG_METHOD_HW_DEVICE_CTX &&
 468                             config->device_type == type &&
 469                             config->pix_fmt == *fmt_ptr) {
 470                                 return config->pix_fmt;
 471                         }
 472                 }
 473         }
 474
 475         // We found no VA-API formats, so take the best software format.
 476         return fmt[0];
 477 }
 478
 479 bool FFmpegCapture::play_video(const string &pathname)
 480 {
 481         // Note: Call before open, not after; otherwise, there's a race.
 482         // (There is now, too, but it tips the correct way. We could use fstat()
 483         // if we had the file descriptor.)
 484         timespec last_modified;
 485         struct stat buf;
 486         if (stat(pathname.c_str(), &buf) != 0) {
 487                 // Probably some sort of protocol, so can't stat.
 488                 last_modified.tv_sec = -1;
 489         } else {
 490                 last_modified = buf.st_mtim;
 491         }
 492         last_colorspace = static_cast<AVColorSpace>(-1);
 493         last_chroma_location = static_cast<AVChromaLocation>(-1);
 494
 495         AVFormatContextWithCloser format_ctx;
 496         if (srt_sock == -1) {
 497                 // Regular file.
 498                 format_ctx = avformat_open_input_unique(pathname.c_str(), /*fmt=*/nullptr,
 499                         /*options=*/nullptr,
 500                         AVIOInterruptCB{ &FFmpegCapture::interrupt_cb_thunk, this });
 501         } else {
 502 #ifdef HAVE_SRT
 503                 // SRT socket, already opened.
 504                 const AVInputFormat *mpegts_fmt = av_find_input_format("mpegts");
 505                 format_ctx = avformat_open_input_unique(&FFmpegCapture::read_srt_thunk, this,
 506                         mpegts_fmt, /*options=*/nullptr,
 507                         AVIOInterruptCB{ &FFmpegCapture::interrupt_cb_thunk, this });
 508 #else
 509                 assert(false);
 510 #endif
 511         }
 512         if (format_ctx == nullptr) {
 513                 fprintf(stderr, "%s: Error opening file\n", pathname.c_str());
 514                 return false;
 515         }
 516
 517         if (avformat_find_stream_info(format_ctx.get(), nullptr) < 0) {
 518                 fprintf(stderr, "%s: Error finding stream info\n", pathname.c_str());
 519                 return false;
 520         }
 521
 522         int video_stream_index = find_stream_index(format_ctx.get(), AVMEDIA_TYPE_VIDEO);
 523         if (video_stream_index == -1) {
 524                 fprintf(stderr, "%s: No video stream found\n", pathname.c_str());
 525                 return false;
 526         }
 527
 528         int audio_stream_index = find_stream_index(format_ctx.get(), AVMEDIA_TYPE_AUDIO);
 529         int subtitle_stream_index = find_stream_index(format_ctx.get(), AVMEDIA_TYPE_SUBTITLE);
 530         has_last_subtitle = false;
 531
 532         // Open video decoder.
 533         const AVCodecParameters *video_codecpar = format_ctx->streams[video_stream_index]->codecpar;
 534         const AVCodec *video_codec = avcodec_find_decoder(video_codecpar->codec_id);
 535
 536         video_timebase = format_ctx->streams[video_stream_index]->time_base;
 537         AVCodecContextWithDeleter video_codec_ctx = avcodec_alloc_context3_unique(nullptr);
 538         if (avcodec_parameters_to_context(video_codec_ctx.get(), video_codecpar) < 0) {
 539                 fprintf(stderr, "%s: Cannot fill video codec parameters\n", pathname.c_str());
 540                 return false;
 541         }
 542         if (video_codec == nullptr) {
 543                 fprintf(stderr, "%s: Cannot find video decoder\n", pathname.c_str());
 544                 return false;
 545         }
 546
 547         // Seemingly, it's not too easy to make something that just initializes
 548         // “whatever goes”, so we don't get CUDA or VULKAN or whatever here
 549         // without enumerating through several different types.
 550         // VA-API and VDPAU will do for now. We prioritize VDPAU for the
 551         // simple reason that there's a VA-API-via-VDPAU emulation for NVidia
 552         // cards that seems to work, but just hangs when trying to transfer the frame.
 553         //
 554         // Note that we don't actually check codec support beforehand,
 555         // so if you have a low-end VDPAU device but a high-end VA-API device,
 556         // you lose out on the extra codec support from the latter.
 557         AVBufferRef *hw_device_ctx = nullptr;
 558         if (av_hwdevice_ctx_create(&hw_device_ctx, AV_HWDEVICE_TYPE_VDPAU, nullptr, nullptr, 0) >= 0) {
 559                 video_codec_ctx->hw_device_ctx = av_buffer_ref(hw_device_ctx);
 560                 video_codec_ctx->get_format = get_hw_format<AV_HWDEVICE_TYPE_VDPAU>;
 561         } else if (av_hwdevice_ctx_create(&hw_device_ctx, AV_HWDEVICE_TYPE_VAAPI, nullptr, nullptr, 0) >= 0) {
 562                 video_codec_ctx->hw_device_ctx = av_buffer_ref(hw_device_ctx);
 563                 video_codec_ctx->get_format = get_hw_format<AV_HWDEVICE_TYPE_VAAPI>;
 564         } else {
 565                 fprintf(stderr, "Failed to initialize VA-API or VDPAU for FFmpeg acceleration. Decoding video in software.\n");
 566         }
 567
 568         if (avcodec_open2(video_codec_ctx.get(), video_codec, nullptr) < 0) {
 569                 fprintf(stderr, "%s: Cannot open video decoder\n", pathname.c_str());
 570                 return false;
 571         }
 572         unique_ptr<AVCodecContext, decltype(avcodec_close)*> video_codec_ctx_cleanup(
 573                 video_codec_ctx.get(), avcodec_close);
 574
 575         // Used in decode_ycbcr_format().
 576         is_mjpeg = video_codecpar->codec_id == AV_CODEC_ID_MJPEG;
 577
 578         // Open audio decoder, if we have audio.
 579         AVCodecContextWithDeleter audio_codec_ctx;
 580         if (audio_stream_index != -1) {
 581                 audio_codec_ctx = avcodec_alloc_context3_unique(nullptr);
 582                 const AVCodecParameters *audio_codecpar = format_ctx->streams[audio_stream_index]->codecpar;
 583                 audio_timebase = format_ctx->streams[audio_stream_index]->time_base;
 584                 if (avcodec_parameters_to_context(audio_codec_ctx.get(), audio_codecpar) < 0) {
 585                         fprintf(stderr, "%s: Cannot fill audio codec parameters\n", pathname.c_str());
 586                         return false;
 587                 }
 588                 const AVCodec *audio_codec = avcodec_find_decoder(audio_codecpar->codec_id);
 589                 if (audio_codec == nullptr) {
 590                         fprintf(stderr, "%s: Cannot find audio decoder\n", pathname.c_str());
 591                         return false;
 592                 }
 593                 if (avcodec_open2(audio_codec_ctx.get(), audio_codec, nullptr) < 0) {
 594                         fprintf(stderr, "%s: Cannot open audio decoder\n", pathname.c_str());
 595                         return false;
 596                 }
 597         }
 598         unique_ptr<AVCodecContext, decltype(avcodec_close)*> audio_codec_ctx_cleanup(
 599                 audio_codec_ctx.get(), avcodec_close);
 600
 601         internal_rewind();
 602
 603         // Main loop.
 604         bool first_frame = true;
 605         int consecutive_errors = 0;
 606         while (!producer_thread_should_quit.should_quit()) {
 607                 if (process_queued_commands(format_ctx.get(), pathname, last_modified, /*rewound=*/nullptr)) {
 608                         return true;
 609                 }
 610                 if (should_interrupt.load()) {
 611                         // Check as a failsafe, so that we don't need to rely on avio if we don't have to.
 612                         return false;
 613                 }
 614                 UniqueFrame audio_frame = audio_frame_allocator->alloc_frame();
 615                 AudioFormat audio_format;
 616
 617                 int64_t audio_pts;
 618                 bool error;
 619                 AVFrameWithDeleter frame = decode_frame(format_ctx.get(), video_codec_ctx.get(), audio_codec_ctx.get(),
 620                         pathname, video_stream_index, audio_stream_index, subtitle_stream_index, audio_frame.get(), &audio_format, &audio_pts, &error);
 621                 if (error) {
 622                         if (++consecutive_errors >= 100) {
 623                                 fprintf(stderr, "More than 100 consecutive video frames, aborting playback.\n");
 624                                 return false;
 625                         } else {
 626                                 continue;
 627                         }
 628                 } else {
 629                         consecutive_errors = 0;
 630                 }
 631                 if (frame == nullptr) {
 632                         // EOF. Loop back to the start if we can.
 633                         if (format_ctx->pb != nullptr && format_ctx->pb->seekable == 0) {
 634                                 // Not seekable (but seemingly, sometimes av_seek_frame() would return 0 anyway,
 635                                 // so don't try).
 636                                 return true;
 637                         }
 638                         if (av_seek_frame(format_ctx.get(), /*stream_index=*/-1, /*timestamp=*/0, /*flags=*/0) < 0) {
 639                                 fprintf(stderr, "%s: Rewind failed, not looping.\n", pathname.c_str());
 640                                 return true;
 641                         }
 642                         if (video_codec_ctx != nullptr) {
 643                                 avcodec_flush_buffers(video_codec_ctx.get());
 644                         }
 645                         if (audio_codec_ctx != nullptr) {
 646                                 avcodec_flush_buffers(audio_codec_ctx.get());
 647                         }
 648                         // If the file has changed since last time, return to get it reloaded.
 649                         // Note that depending on how you move the file into place, you might
 650                         // end up corrupting the one you're already playing, so this path
 651                         // might not trigger.
 652                         if (changed_since(pathname, last_modified)) {
 653                                 return true;
 654                         }
 655                         internal_rewind();
 656                         continue;
 657                 }
 658
 659                 VideoFormat video_format = construct_video_format(frame.get(), video_timebase);
 660                 if (video_format.frame_rate_nom == 0 || video_format.frame_rate_den == 0) {
 661                         // Invalid frame rate; try constructing it from the previous frame length.
 662                         // (This is especially important if we are the master card, for SRT,
 663                         // since it affects audio. Not all senders have good timebases
 664                         // (e.g., Larix rounds first to timebase 1000 and then multiplies by
 665                         // 90 from there, it seems), but it's much better to have an oscillating
 666                         // value than just locking at 60.
 667                         if (last_pts != 0 && frame->pts > last_pts) {
 668                                 int64_t pts_diff = frame->pts - last_pts;
 669                                 video_format.frame_rate_nom = video_timebase.den;
 670                                 video_format.frame_rate_den = video_timebase.num * pts_diff;
 671                         } else {
 672                                 video_format.frame_rate_nom = 60;
 673                                 video_format.frame_rate_den = 1;
 674                         }
 675                 }
 676                 UniqueFrame video_frame = make_video_frame(frame.get(), pathname, &error);
 677                 if (error) {
 678                         return false;
 679                 }
 680
 681                 for ( ;; ) {
 682                         if (last_pts == 0 && pts_origin == 0) {
 683                                 pts_origin = frame->pts;
 684                         }
 685                         steady_clock::time_point now = steady_clock::now();
 686                         if (play_as_fast_as_possible) {
 687                                 video_frame->received_timestamp = now;
 688                                 audio_frame->received_timestamp = now;
 689                                 next_frame_start = now;
 690                         } else {
 691                                 next_frame_start = compute_frame_start(frame->pts, pts_origin, video_timebase, start, rate);
 692                                 if (first_frame && last_frame_was_connected) {
 693                                         // If reconnect took more than one second, this is probably a live feed,
 694                                         // and we should reset the resampler. (Or the rate is really, really low,
 695                                         // in which case a reset on the first frame is fine anyway.)
 696                                         if (duration<double>(next_frame_start - last_frame).count() >= 1.0) {
 697                                                 last_frame_was_connected = false;
 698                                         }
 699                                 }
 700                                 video_frame->received_timestamp = next_frame_start;
 701
 702                                 // The easiest way to get all the rate conversions etc. right is to move the
 703                                 // audio PTS into the video PTS timebase and go from there. (We'll get some
 704                                 // rounding issues, but they should not be a big problem.)
 705                                 int64_t audio_pts_as_video_pts = av_rescale_q(audio_pts, audio_timebase, video_timebase);
 706                                 audio_frame->received_timestamp = compute_frame_start(audio_pts_as_video_pts, pts_origin, video_timebase, start, rate);
 707
 708                                 if (audio_frame->len != 0) {
 709                                         // The received timestamps in Nageru are measured after we've just received the frame.
 710                                         // However, pts (especially audio pts) is at the _beginning_ of the frame.
 711                                         // If we have locked audio, the distinction doesn't really matter, as pts is
 712                                         // on a relative scale and a fixed offset is fine. But if we don't, we will have
 713                                         // a different number of samples each time, which will cause huge audio jitter
 714                                         // and throw off the resampler.
 715                                         //
 716                                         // In a sense, we should have compensated by adding the frame and audio lengths
 717                                         // to video_frame->received_timestamp and audio_frame->received_timestamp respectively,
 718                                         // but that would mean extra waiting in sleep_until(). All we need is that they
 719                                         // are correct relative to each other, though (and to the other frames we send),
 720                                         // so just align the end of the audio frame, and we're fine.
 721                                         size_t num_samples = (audio_frame->len * 8) / audio_format.bits_per_sample / audio_format.num_channels;
 722                                         double offset = double(num_samples) / OUTPUT_FREQUENCY -
 723                                                 double(video_format.frame_rate_den) / video_format.frame_rate_nom;
 724                                         audio_frame->received_timestamp += duration_cast<steady_clock::duration>(duration<double>(offset));
 725                                 }
 726
 727                                 if (duration<double>(now - next_frame_start).count() >= 0.1) {
 728                                         // If we don't have enough CPU to keep up, or if we have a live stream
 729                                         // where the initial origin was somehow wrong, we could be behind indefinitely.
 730                                         // In particular, this will give the audio resampler problems as it tries
 731                                         // to speed up to reduce the delay, hitting the low end of the buffer every time.
 732                                         fprintf(stderr, "%s: Playback %.0f ms behind, resetting time scale\n",
 733                                                 pathname.c_str(),
 734                                                 1e3 * duration<double>(now - next_frame_start).count());
 735                                         pts_origin = frame->pts;
 736                                         start = next_frame_start = now;
 737                                         timecode += TYPICAL_FPS * 2 + 1;
 738                                 }
 739                         }
 740                         bool finished_wakeup;
 741                         if (play_as_fast_as_possible) {
 742                                 finished_wakeup = !producer_thread_should_quit.should_quit();
 743                         } else {
 744                                 finished_wakeup = producer_thread_should_quit.sleep_until(next_frame_start);
 745                         }
 746                         if (finished_wakeup) {
 747                                 if (audio_frame->len > 0) {
 748                                         assert(audio_pts != -1);
 749                                 }
 750                                 if (!last_frame_was_connected) {
 751                                         // We're recovering from an error (or really slow load, see above).
 752                                         // Make sure to get the audio resampler reset. (This is a hack;
 753                                         // ideally, the frame callback should just accept a way to signal
 754                                         // audio discontinuity.)
 755                                         timecode += TYPICAL_FPS * 2 + 1;
 756                                 }
 757                                 last_neutral_color = get_neutral_color(frame->metadata);
 758                                 if (frame_callback != nullptr) {
 759                                         frame_callback(frame->pts, video_timebase, audio_pts, audio_timebase, timecode++,
 760                                                 video_frame.get_and_release(), 0, video_format,
 761                                                 audio_frame.get_and_release(), 0, audio_format);
 762                                 }
 763                                 first_frame = false;
 764                                 last_frame = steady_clock::now();
 765                                 last_frame_was_connected = true;
 766                                 break;
 767                         } else {
 768                                 if (producer_thread_should_quit.should_quit()) break;
 769
 770                                 bool rewound = false;
 771                                 if (process_queued_commands(format_ctx.get(), pathname, last_modified, &rewound)) {
 772                                         return true;
 773                                 }
 774                                 // If we just rewound, drop this frame on the floor and be done.
 775                                 if (rewound) {
 776                                         break;
 777                                 }
 778                                 // OK, we didn't, so probably a rate change. Recalculate next_frame_start,
 779                                 // but if it's now in the past, we'll reset the origin, so that we don't
 780                                 // generate a huge backlog of frames that we need to run through quickly.
 781                                 next_frame_start = compute_frame_start(frame->pts, pts_origin, video_timebase, start, rate);
 782                                 steady_clock::time_point now = steady_clock::now();
 783                                 if (next_frame_start < now) {
 784                                         pts_origin = frame->pts;
 785                                         start = next_frame_start = now;
 786                                 }
 787                         }
 788                 }
 789                 last_pts = frame->pts;
 790         }
 791         return true;
 792 }
 793
 794 void FFmpegCapture::internal_rewind()
 795 {
 796         pts_origin = last_pts = 0;
 797         start = next_frame_start = steady_clock::now();
 798 }
 799
 800 bool FFmpegCapture::process_queued_commands(AVFormatContext *format_ctx, const std::string &pathname, timespec last_modified, bool *rewound)
 801 {
 802         // Process any queued commands from other threads.
 803         vector<QueuedCommand> commands;
 804         {
 805                 lock_guard<mutex> lock(queue_mu);
 806                 swap(commands, command_queue);
 807         }
 808         for (const QueuedCommand &cmd : commands) {
 809                 switch (cmd.command) {
 810                 case QueuedCommand::REWIND:
 811                         if (av_seek_frame(format_ctx, /*stream_index=*/-1, /*timestamp=*/0, /*flags=*/0) < 0) {
 812                                 fprintf(stderr, "%s: Rewind failed, stopping play.\n", pathname.c_str());
 813                         }
 814                         // If the file has changed since last time, return to get it reloaded.
 815                         // Note that depending on how you move the file into place, you might
 816                         // end up corrupting the one you're already playing, so this path
 817                         // might not trigger.
 818                         if (changed_since(pathname, last_modified)) {
 819                                 return true;
 820                         }
 821                         internal_rewind();
 822                         if (rewound != nullptr) {
 823                                 *rewound = true;
 824                         }
 825                         break;
 826
 827                 case QueuedCommand::CHANGE_RATE:
 828                         // Change the origin to the last played frame.
 829                         start = compute_frame_start(last_pts, pts_origin, video_timebase, start, rate);
 830                         pts_origin = last_pts;
 831                         rate = cmd.new_rate;
 832                         play_as_fast_as_possible = (rate >= 10.0);
 833                         break;
 834                 }
 835         }
 836         return false;
 837 }
 838
 839 namespace {
 840
 841 }  // namespace
 842
 843 AVFrameWithDeleter FFmpegCapture::decode_frame(AVFormatContext *format_ctx, AVCodecContext *video_codec_ctx, AVCodecContext *audio_codec_ctx,
 844         const std::string &pathname, int video_stream_index, int audio_stream_index, int subtitle_stream_index,
 845         FrameAllocator::Frame *audio_frame, AudioFormat *audio_format, int64_t *audio_pts, bool *error)
 846 {
 847         *error = false;
 848
 849         // Read packets until we have a frame or there are none left.
 850         bool frame_finished = false;
 851         AVFrameWithDeleter audio_avframe = av_frame_alloc_unique();
 852         AVFrameWithDeleter video_avframe = av_frame_alloc_unique();
 853         bool eof = false;
 854         *audio_pts = -1;
 855         bool has_audio = false;
 856         do {
 857                 AVPacket pkt;
 858                 unique_ptr<AVPacket, decltype(av_packet_unref)*> pkt_cleanup(
 859                         &pkt, av_packet_unref);
 860                 av_init_packet(&pkt);
 861                 pkt.data = nullptr;
 862                 pkt.size = 0;
 863                 if (av_read_frame(format_ctx, &pkt) == 0) {
 864                         if (pkt.stream_index == audio_stream_index && audio_callback != nullptr) {
 865                                 audio_callback(&pkt, format_ctx->streams[audio_stream_index]->time_base);
 866                         }
 867                         if (pkt.stream_index == video_stream_index && video_callback != nullptr) {
 868                                 video_callback(&pkt, format_ctx->streams[video_stream_index]->time_base);
 869                         }
 870                         if (pkt.stream_index == video_stream_index && global_flags.transcode_video) {
 871                                 if (avcodec_send_packet(video_codec_ctx, &pkt) < 0) {
 872                                         fprintf(stderr, "%s: Cannot send packet to video codec.\n", pathname.c_str());
 873                                         *error = true;
 874                                         return AVFrameWithDeleter(nullptr);
 875                                 }
 876                         } else if (pkt.stream_index == audio_stream_index && global_flags.transcode_audio) {
 877                                 has_audio = true;
 878                                 if (avcodec_send_packet(audio_codec_ctx, &pkt) < 0) {
 879                                         fprintf(stderr, "%s: Cannot send packet to audio codec.\n", pathname.c_str());
 880                                         *error = true;
 881                                         return AVFrameWithDeleter(nullptr);
 882                                 }
 883                         } else if (pkt.stream_index == subtitle_stream_index) {
 884                                 last_subtitle = string(reinterpret_cast<const char *>(pkt.data), pkt.size);
 885                                 has_last_subtitle = true;
 886                         }
 887                 } else {
 888                         eof = true;  // Or error, but ignore that for the time being.
 889                 }
 890
 891                 // Decode audio, if any.
 892                 if (has_audio) {
 893                         for ( ;; ) {
 894                                 int err = avcodec_receive_frame(audio_codec_ctx, audio_avframe.get());
 895                                 if (err == 0) {
 896                                         if (*audio_pts == -1) {
 897                                                 *audio_pts = audio_avframe->pts;
 898                                         }
 899                                         convert_audio(audio_avframe.get(), audio_frame, audio_format);
 900                                 } else if (err == AVERROR(EAGAIN)) {
 901                                         break;
 902                                 } else {
 903                                         fprintf(stderr, "%s: Cannot receive frame from audio codec.\n", pathname.c_str());
 904                                         *error = true;
 905                                         return AVFrameWithDeleter(nullptr);
 906                                 }
 907                         }
 908                 }
 909
 910                 // Decode video, if we have a frame.
 911                 int err = avcodec_receive_frame(video_codec_ctx, video_avframe.get());
 912                 if (err == 0) {
 913                         if (video_avframe->format == AV_PIX_FMT_VAAPI ||
 914                             video_avframe->format == AV_PIX_FMT_VDPAU) {
 915                                 // Get the frame down to the CPU. (TODO: See if we can keep it
 916                                 // on the GPU all the way, since it will be going up again later.
 917                                 // However, this only works if the OpenGL GPU is the same one.)
 918                                 AVFrameWithDeleter sw_frame = av_frame_alloc_unique();
 919                                 int err = av_hwframe_transfer_data(sw_frame.get(), video_avframe.get(), 0);
 920                                 if (err != 0) {
 921                                         fprintf(stderr, "%s: Cannot transfer hardware video frame to software.\n", pathname.c_str());
 922                                         *error = true;
 923                                         return AVFrameWithDeleter(nullptr);
 924                                 }
 925                                 sw_frame->pts = video_avframe->pts;
 926                                 sw_frame->pkt_duration = video_avframe->pkt_duration;
 927                                 video_avframe = move(sw_frame);
 928                         }
 929                         frame_finished = true;
 930                         break;
 931                 } else if (err != AVERROR(EAGAIN)) {
 932                         fprintf(stderr, "%s: Cannot receive frame from video codec.\n", pathname.c_str());
 933                         *error = true;
 934                         return AVFrameWithDeleter(nullptr);
 935                 }
 936         } while (!eof);
 937
 938         if (frame_finished)
 939                 return video_avframe;
 940         else
 941                 return AVFrameWithDeleter(nullptr);
 942 }
 943
 944 void FFmpegCapture::convert_audio(const AVFrame *audio_avframe, FrameAllocator::Frame *audio_frame, AudioFormat *audio_format)
 945 {
 946         // Decide on a format. If there already is one in this audio frame,
 947         // we're pretty much forced to use it. If not, we try to find an exact match.
 948         // If that still doesn't work, we default to 32-bit signed chunked
 949         // (float would be nice, but there's really no way to signal that yet).
 950         AVSampleFormat dst_format;
 951         if (audio_format->bits_per_sample == 0) {
 952                 switch (audio_avframe->format) {
 953                 case AV_SAMPLE_FMT_S16:
 954                 case AV_SAMPLE_FMT_S16P:
 955                         audio_format->bits_per_sample = 16;
 956                         dst_format = AV_SAMPLE_FMT_S16;
 957                         break;
 958                 case AV_SAMPLE_FMT_S32:
 959                 case AV_SAMPLE_FMT_S32P:
 960                 default:
 961                         audio_format->bits_per_sample = 32;
 962                         dst_format = AV_SAMPLE_FMT_S32;
 963                         break;
 964                 }
 965         } else if (audio_format->bits_per_sample == 16) {
 966                 dst_format = AV_SAMPLE_FMT_S16;
 967         } else if (audio_format->bits_per_sample == 32) {
 968                 dst_format = AV_SAMPLE_FMT_S32;
 969         } else {
 970                 assert(false);
 971         }
 972         audio_format->num_channels = 2;
 973
 974         int64_t channel_layout = audio_avframe->channel_layout;
 975         if (channel_layout == 0) {
 976                 channel_layout = av_get_default_channel_layout(audio_avframe->channels);
 977         }
 978
 979         if (resampler == nullptr ||
 980             audio_avframe->format != last_src_format ||
 981             dst_format != last_dst_format ||
 982             channel_layout != last_channel_layout ||
 983             audio_avframe->sample_rate != last_sample_rate) {
 984                 swr_free(&resampler);
 985                 resampler = swr_alloc_set_opts(nullptr,
 986                                                /*out_ch_layout=*/AV_CH_LAYOUT_STEREO_DOWNMIX,
 987                                                /*out_sample_fmt=*/dst_format,
 988                                                /*out_sample_rate=*/OUTPUT_FREQUENCY,
 989                                                /*in_ch_layout=*/channel_layout,
 990                                                /*in_sample_fmt=*/AVSampleFormat(audio_avframe->format),
 991                                                /*in_sample_rate=*/audio_avframe->sample_rate,
 992                                                /*log_offset=*/0,
 993                                                /*log_ctx=*/nullptr);
 994
 995                 if (resampler == nullptr) {
 996                         fprintf(stderr, "Allocating resampler failed.\n");
 997                         abort();
 998                 }
 999
1000                 if (swr_init(resampler) < 0) {
1001                         fprintf(stderr, "Could not open resample context.\n");
1002                         abort();
1003                 }
1004
1005                 last_src_format = AVSampleFormat(audio_avframe->format);
1006                 last_dst_format = dst_format;
1007                 last_channel_layout = channel_layout;
1008                 last_sample_rate = audio_avframe->sample_rate;
1009         }
1010
1011         size_t bytes_per_sample = (audio_format->bits_per_sample / 8) * 2;
1012         size_t num_samples_room = (audio_frame->size - audio_frame->len) / bytes_per_sample;
1013
1014         uint8_t *data = audio_frame->data + audio_frame->len;
1015         int out_samples = swr_convert(resampler, &data, num_samples_room,
1016                 const_cast<const uint8_t **>(audio_avframe->data), audio_avframe->nb_samples);
1017         if (out_samples < 0) {
1018                 fprintf(stderr, "Audio conversion failed.\n");
1019                 abort();
1020         }
1021
1022         audio_frame->len += out_samples * bytes_per_sample;
1023 }
1024
1025 VideoFormat FFmpegCapture::construct_video_format(const AVFrame *frame, AVRational video_timebase)
1026 {
1027         VideoFormat video_format;
1028         video_format.width = frame_width(frame);
1029         video_format.height = frame_height(frame);
1030         if (pixel_format == bmusb::PixelFormat_8BitBGRA) {
1031                 video_format.stride = frame_width(frame) * 4;
1032         } else if (pixel_format == FFmpegCapture::PixelFormat_NV12) {
1033                 video_format.stride = frame_width(frame);
1034         } else {
1035                 assert(pixel_format == bmusb::PixelFormat_8BitYCbCrPlanar);
1036                 video_format.stride = frame_width(frame);
1037         }
1038         video_format.frame_rate_nom = video_timebase.den;
1039         video_format.frame_rate_den = frame->pkt_duration * video_timebase.num;
1040         video_format.has_signal = true;
1041         video_format.is_connected = true;
1042         return video_format;
1043 }
1044
1045 UniqueFrame FFmpegCapture::make_video_frame(const AVFrame *frame, const string &pathname, bool *error)
1046 {
1047         *error = false;
1048
1049         UniqueFrame video_frame(video_frame_allocator->alloc_frame());
1050         if (video_frame->data == nullptr) {
1051                 return video_frame;
1052         }
1053
1054         if (sws_ctx == nullptr ||
1055             sws_last_width != frame->width ||
1056             sws_last_height != frame->height ||
1057             sws_last_src_format != frame->format) {
1058                 sws_dst_format = decide_dst_format(AVPixelFormat(frame->format), pixel_format);
1059                 sws_ctx.reset(
1060                         sws_getContext(frame->width, frame->height, AVPixelFormat(frame->format),
1061                                 frame_width(frame), frame_height(frame), sws_dst_format,
1062                                 SWS_BICUBIC, nullptr, nullptr, nullptr));
1063                 sws_last_width = frame->width;
1064                 sws_last_height = frame->height;
1065                 sws_last_src_format = frame->format;
1066         }
1067         if (sws_ctx == nullptr) {
1068                 fprintf(stderr, "%s: Could not create scaler context\n", pathname.c_str());
1069                 *error = true;
1070                 return video_frame;
1071         }
1072
1073         uint8_t *pic_data[4] = { nullptr, nullptr, nullptr, nullptr };
1074         int linesizes[4] = { 0, 0, 0, 0 };
1075         if (pixel_format == bmusb::PixelFormat_8BitBGRA) {
1076                 pic_data[0] = video_frame->data;
1077                 linesizes[0] = frame_width(frame) * 4;
1078                 video_frame->len = (frame_width(frame) * 4) * frame_height(frame);
1079         } else if (pixel_format == PixelFormat_NV12) {
1080                 pic_data[0] = video_frame->data;
1081                 linesizes[0] = frame_width(frame);
1082
1083                 pic_data[1] = pic_data[0] + frame_width(frame) * frame_height(frame);
1084                 linesizes[1] = frame_width(frame);
1085
1086                 video_frame->len = (frame_width(frame) * 2) * frame_height(frame);
1087
1088                 const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(sws_dst_format);
1089                 current_frame_ycbcr_format = decode_ycbcr_format(desc, frame, is_mjpeg, &last_colorspace, &last_chroma_location);
1090         } else {
1091                 assert(pixel_format == bmusb::PixelFormat_8BitYCbCrPlanar);
1092                 const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(sws_dst_format);
1093
1094                 int chroma_width = AV_CEIL_RSHIFT(int(frame_width(frame)), desc->log2_chroma_w);
1095                 int chroma_height = AV_CEIL_RSHIFT(int(frame_height(frame)), desc->log2_chroma_h);
1096
1097                 pic_data[0] = video_frame->data;
1098                 linesizes[0] = frame_width(frame);
1099
1100                 pic_data[1] = pic_data[0] + frame_width(frame) * frame_height(frame);
1101                 linesizes[1] = chroma_width;
1102
1103                 pic_data[2] = pic_data[1] + chroma_width * chroma_height;
1104                 linesizes[2] = chroma_width;
1105
1106                 video_frame->len = frame_width(frame) * frame_height(frame) + 2 * chroma_width * chroma_height;
1107
1108                 current_frame_ycbcr_format = decode_ycbcr_format(desc, frame, is_mjpeg, &last_colorspace, &last_chroma_location);
1109         }
1110         sws_scale(sws_ctx.get(), frame->data, frame->linesize, 0, frame->height, pic_data, linesizes);
1111
1112         return video_frame;
1113 }
1114
1115 int FFmpegCapture::interrupt_cb_thunk(void *opaque)
1116 {
1117         return reinterpret_cast<FFmpegCapture *>(opaque)->interrupt_cb();
1118 }
1119
1120 int FFmpegCapture::interrupt_cb()
1121 {
1122         return should_interrupt.load();
1123 }
1124
1125 unsigned FFmpegCapture::frame_width(const AVFrame *frame) const
1126 {
1127         if (width == 0) {
1128                 return frame->width;
1129         } else {
1130                 return width;
1131         }
1132 }
1133
1134 unsigned FFmpegCapture::frame_height(const AVFrame *frame) const
1135 {
1136         if (height == 0) {
1137                 return frame->height;
1138         } else {
1139                 return height;
1140         }
1141 }
1142
1143 #ifdef HAVE_SRT
1144 int FFmpegCapture::read_srt_thunk(void *opaque, uint8_t *buf, int buf_size)
1145 {
1146         return reinterpret_cast<FFmpegCapture *>(opaque)->read_srt(buf, buf_size);
1147 }
1148
1149 int FFmpegCapture::read_srt(uint8_t *buf, int buf_size)
1150 {
1151         SRT_MSGCTRL mc = srt_msgctrl_default;
1152         return srt_recvmsg2(srt_sock, reinterpret_cast<char *>(buf), buf_size, &mc);
1153 }
1154 #endif