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