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

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