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