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