git.sesse.net Git - nageru/blob - 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 "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 "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)
 138 {
 139         YCbCrFormat format;
 140         AVColorSpace colorspace = av_frame_get_colorspace(frame);
 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         format.cr_x_position = format.cb_x_position;
 202         format.cr_y_position = format.cb_y_position;
 203         return format;
 204 }
 205
 206 }  // namespace
 207
 208 FFmpegCapture::FFmpegCapture(const string &filename, unsigned width, unsigned height)
 209         : filename(filename), width(width), height(height), video_timebase{1, 1}
 210 {
 211         description = "Video: " + filename;
 212
 213         last_frame = steady_clock::now();
 214
 215         avformat_network_init();  // In case someone wants this.
 216 }
 217
 218 FFmpegCapture::~FFmpegCapture()
 219 {
 220         if (has_dequeue_callbacks) {
 221                 dequeue_cleanup_callback();
 222         }
 223         avresample_free(&resampler);
 224 }
 225
 226 void FFmpegCapture::configure_card()
 227 {
 228         if (video_frame_allocator == nullptr) {
 229                 owned_video_frame_allocator.reset(new MallocFrameAllocator(FRAME_SIZE, NUM_QUEUED_VIDEO_FRAMES));
 230                 set_video_frame_allocator(owned_video_frame_allocator.get());
 231         }
 232         if (audio_frame_allocator == nullptr) {
 233                 // Audio can come out in pretty large chunks, so increase from the default 1 MB.
 234                 owned_audio_frame_allocator.reset(new MallocFrameAllocator(1 << 20, NUM_QUEUED_AUDIO_FRAMES));
 235                 set_audio_frame_allocator(owned_audio_frame_allocator.get());
 236         }
 237 }
 238
 239 void FFmpegCapture::start_bm_capture()
 240 {
 241         if (running) {
 242                 return;
 243         }
 244         running = true;
 245         producer_thread_should_quit.unquit();
 246         producer_thread = thread(&FFmpegCapture::producer_thread_func, this);
 247 }
 248
 249 void FFmpegCapture::stop_dequeue_thread()
 250 {
 251         if (!running) {
 252                 return;
 253         }
 254         running = false;
 255         producer_thread_should_quit.quit();
 256         producer_thread.join();
 257 }
 258
 259 std::map<uint32_t, VideoMode> FFmpegCapture::get_available_video_modes() const
 260 {
 261         // Note: This will never really be shown in the UI.
 262         VideoMode mode;
 263
 264         char buf[256];
 265         snprintf(buf, sizeof(buf), "%ux%u", width, height);
 266         mode.name = buf;
 267
 268         mode.autodetect = false;
 269         mode.width = width;
 270         mode.height = height;
 271         mode.frame_rate_num = 60;
 272         mode.frame_rate_den = 1;
 273         mode.interlaced = false;
 274
 275         return {{ 0, mode }};
 276 }
 277
 278 void FFmpegCapture::producer_thread_func()
 279 {
 280         char thread_name[16];
 281         snprintf(thread_name, sizeof(thread_name), "FFmpeg_C_%d", card_index);
 282         pthread_setname_np(pthread_self(), thread_name);
 283
 284         while (!producer_thread_should_quit.should_quit()) {
 285                 string filename_copy;
 286                 {
 287                         lock_guard<mutex> lock(filename_mu);
 288                         filename_copy = filename;
 289                 }
 290
 291                 string pathname = search_for_file(filename_copy);
 292                 if (pathname.empty()) {
 293                         fprintf(stderr, "%s not found, sleeping one second and trying again...\n", filename_copy.c_str());
 294                         send_disconnected_frame();
 295                         producer_thread_should_quit.sleep_for(seconds(1));
 296                         continue;
 297                 }
 298                 should_interrupt = false;
 299                 if (!play_video(pathname)) {
 300                         // Error.
 301                         fprintf(stderr, "Error when playing %s, sleeping one second and trying again...\n", pathname.c_str());
 302                         send_disconnected_frame();
 303                         producer_thread_should_quit.sleep_for(seconds(1));
 304                         continue;
 305                 }
 306
 307                 // Probably just EOF, will exit the loop above on next test.
 308         }
 309
 310         if (has_dequeue_callbacks) {
 311                 dequeue_cleanup_callback();
 312                 has_dequeue_callbacks = false;
 313         }
 314 }
 315
 316 void FFmpegCapture::send_disconnected_frame()
 317 {
 318         // Send an empty frame to signal that we have no signal anymore.
 319         FrameAllocator::Frame video_frame = video_frame_allocator->alloc_frame();
 320         if (video_frame.data) {
 321                 VideoFormat video_format;
 322                 video_format.width = width;
 323                 video_format.height = height;
 324                 video_format.frame_rate_nom = 60;
 325                 video_format.frame_rate_den = 1;
 326                 video_format.is_connected = false;
 327                 if (pixel_format == bmusb::PixelFormat_8BitBGRA) {
 328                         video_format.stride = width * 4;
 329                         video_frame.len = width * height * 4;
 330                         memset(video_frame.data, 0, video_frame.len);
 331                 } else {
 332                         video_format.stride = width;
 333                         current_frame_ycbcr_format.luma_coefficients = YCBCR_REC_709;
 334                         current_frame_ycbcr_format.full_range = true;
 335                         current_frame_ycbcr_format.num_levels = 256;
 336                         current_frame_ycbcr_format.chroma_subsampling_x = 2;
 337                         current_frame_ycbcr_format.chroma_subsampling_y = 2;
 338                         current_frame_ycbcr_format.cb_x_position = 0.0f;
 339                         current_frame_ycbcr_format.cb_y_position = 0.0f;
 340                         current_frame_ycbcr_format.cr_x_position = 0.0f;
 341                         current_frame_ycbcr_format.cr_y_position = 0.0f;
 342                         video_frame.len = width * height * 2;
 343                         memset(video_frame.data, 0, width * height);
 344                         memset(video_frame.data + width * height, 128, width * height);  // Valid for both NV12 and planar.
 345                 }
 346
 347                 frame_callback(-1, AVRational{1, TIMEBASE}, -1, AVRational{1, TIMEBASE}, timecode++,
 348                         video_frame, /*video_offset=*/0, video_format,
 349                         FrameAllocator::Frame(), /*audio_offset=*/0, AudioFormat());
 350                 last_frame_was_connected = false;
 351         }
 352 }
 353
 354 bool FFmpegCapture::play_video(const string &pathname)
 355 {
 356         // Note: Call before open, not after; otherwise, there's a race.
 357         // (There is now, too, but it tips the correct way. We could use fstat()
 358         // if we had the file descriptor.)
 359         timespec last_modified;
 360         struct stat buf;
 361         if (stat(pathname.c_str(), &buf) != 0) {
 362                 // Probably some sort of protocol, so can't stat.
 363                 last_modified.tv_sec = -1;
 364         } else {
 365                 last_modified = buf.st_mtim;
 366         }
 367
 368         AVDictionary *opts = nullptr;
 369         av_dict_set(&opts, "fflags", "nobuffer", 0);
 370
 371         auto format_ctx = avformat_open_input_unique(pathname.c_str(), nullptr, &opts, AVIOInterruptCB{ &FFmpegCapture::interrupt_cb_thunk, this });
 372         if (format_ctx == nullptr) {
 373                 fprintf(stderr, "%s: Error opening file\n", pathname.c_str());
 374                 return false;
 375         }
 376
 377         if (avformat_find_stream_info(format_ctx.get(), nullptr) < 0) {
 378                 fprintf(stderr, "%s: Error finding stream info\n", pathname.c_str());
 379                 return false;
 380         }
 381
 382         int video_stream_index = find_stream_index(format_ctx.get(), AVMEDIA_TYPE_VIDEO);
 383         int audio_stream_index = find_stream_index(format_ctx.get(), AVMEDIA_TYPE_AUDIO);
 384
 385         if (video_stream_index == -1 && audio_stream_index == -1) {
 386                 fprintf(stderr, "%s: No audio nor video stream found\n", pathname.c_str());
 387                 return false;
 388         }
 389         if (video_stream_index == -1) {
 390                 fprintf(stderr, "%s: No video stream found, assuming audio-only.\n", pathname.c_str());
 391         }
 392         const bool audio_only_stream = (video_stream_index == -1);
 393
 394         // Open video decoder, if we have video.
 395         AVCodecContextWithDeleter video_codec_ctx;
 396         if (video_stream_index != -1) {
 397                 const AVCodecParameters *video_codecpar = format_ctx->streams[video_stream_index]->codecpar;
 398                 AVCodec *video_codec = avcodec_find_decoder(video_codecpar->codec_id);
 399                 video_timebase = format_ctx->streams[video_stream_index]->time_base;
 400                 video_codec_ctx = avcodec_alloc_context3_unique(nullptr);
 401                 if (avcodec_parameters_to_context(video_codec_ctx.get(), video_codecpar) < 0) {
 402                         fprintf(stderr, "%s: Cannot fill video codec parameters\n", pathname.c_str());
 403                         return false;
 404                 }
 405                 if (video_codec == nullptr) {
 406                         fprintf(stderr, "%s: Cannot find video decoder\n", pathname.c_str());
 407                         return false;
 408                 }
 409                 if (avcodec_open2(video_codec_ctx.get(), video_codec, nullptr) < 0) {
 410                         fprintf(stderr, "%s: Cannot open video decoder\n", pathname.c_str());
 411                         return false;
 412                 }
 413         }
 414         unique_ptr<AVCodecContext, decltype(avcodec_close)*> video_codec_ctx_cleanup(
 415                 video_codec_ctx.get(), avcodec_close);
 416
 417         // Open audio decoder, if we have audio.
 418         AVCodecContextWithDeleter audio_codec_ctx;
 419         if (audio_stream_index != -1) {
 420                 audio_codec_ctx = avcodec_alloc_context3_unique(nullptr);
 421                 const AVCodecParameters *audio_codecpar = format_ctx->streams[audio_stream_index]->codecpar;
 422                 audio_timebase = format_ctx->streams[audio_stream_index]->time_base;
 423                 if (avcodec_parameters_to_context(audio_codec_ctx.get(), audio_codecpar) < 0) {
 424                         fprintf(stderr, "%s: Cannot fill audio codec parameters\n", pathname.c_str());
 425                         return false;
 426                 }
 427                 AVCodec *audio_codec = avcodec_find_decoder(audio_codecpar->codec_id);
 428                 if (audio_codec == nullptr) {
 429                         fprintf(stderr, "%s: Cannot find audio decoder\n", pathname.c_str());
 430                         return false;
 431                 }
 432                 if (avcodec_open2(audio_codec_ctx.get(), audio_codec, nullptr) < 0) {
 433                         fprintf(stderr, "%s: Cannot open audio decoder\n", pathname.c_str());
 434                         return false;
 435                 }
 436         }
 437         unique_ptr<AVCodecContext, decltype(avcodec_close)*> audio_codec_ctx_cleanup(
 438                 audio_codec_ctx.get(), avcodec_close);
 439
 440         internal_rewind();
 441
 442         // Main loop.
 443         bool first_frame = true;
 444         while (!producer_thread_should_quit.should_quit()) {
 445                 if (process_queued_commands(format_ctx.get(), pathname, last_modified, /*rewound=*/nullptr)) {
 446                         return true;
 447                 }
 448                 UniqueFrame audio_frame = audio_frame_allocator->alloc_frame();
 449                 AudioFormat audio_format;
 450
 451                 int64_t audio_pts;
 452                 bool error;
 453                 AVFrameWithDeleter frame = decode_frame(format_ctx.get(), video_codec_ctx.get(), audio_codec_ctx.get(),
 454                         pathname, video_stream_index, audio_stream_index, audio_frame.get(), &audio_format, &audio_pts, &error);
 455                 if (error) {
 456                         return false;
 457                 }
 458                 if (frame == nullptr && !(audio_only_stream && audio_frame->len > 0)) {
 459                         // EOF. Loop back to the start if we can.
 460                         if (av_seek_frame(format_ctx.get(), /*stream_index=*/-1, /*timestamp=*/0, /*flags=*/0) < 0) {
 461                                 fprintf(stderr, "%s: Rewind failed, not looping.\n", pathname.c_str());
 462                                 return true;
 463                         }
 464                         if (video_codec_ctx != nullptr) {
 465                                 avcodec_flush_buffers(video_codec_ctx.get());
 466                         }
 467                         if (audio_codec_ctx != nullptr) {
 468                                 avcodec_flush_buffers(audio_codec_ctx.get());
 469                         }
 470                         // If the file has changed since last time, return to get it reloaded.
 471                         // Note that depending on how you move the file into place, you might
 472                         // end up corrupting the one you're already playing, so this path
 473                         // might not trigger.
 474                         if (changed_since(pathname, last_modified)) {
 475                                 return true;
 476                         }
 477                         internal_rewind();
 478                         continue;
 479                 }
 480
 481                 VideoFormat video_format;
 482                 UniqueFrame video_frame;
 483                 if (!audio_only_stream) {
 484                         video_format = construct_video_format(frame.get(), video_timebase);
 485                         video_frame = make_video_frame(frame.get(), pathname, &error);
 486                         if (error) {
 487                                 return false;
 488                         }
 489                 }
 490
 491                 int64_t frame_pts = audio_only_stream ? audio_pts : frame->pts;
 492                 AVRational timebase = audio_only_stream ? audio_timebase : video_timebase;
 493                 for ( ;; ) {  // Try sending the frame in a loop as long as we get interrupted (then break).
 494                         if (last_pts == 0 && pts_origin == 0) {
 495                                 pts_origin = frame_pts;
 496                         }
 497                         next_frame_start = compute_frame_start(frame_pts, pts_origin, timebase, start, rate);
 498                         if (audio_only_stream) {
 499                                 audio_frame->received_timestamp = next_frame_start;
 500                         } else {
 501                                 if (first_frame && last_frame_was_connected) {
 502                                         // If reconnect took more than one second, this is probably a live feed,
 503                                         // and we should reset the resampler. (Or the rate is really, really low,
 504                                         // in which case a reset on the first frame is fine anyway.)
 505                                         if (duration<double>(next_frame_start - last_frame).count() >= 1.0) {
 506                                                 last_frame_was_connected = false;
 507                                         }
 508                                 }
 509                                 video_frame->received_timestamp = next_frame_start;
 510
 511                                 // The easiest way to get all the rate conversions etc. right is to move the
 512                                 // audio PTS into the video PTS timebase and go from there. (We'll get some
 513                                 // rounding issues, but they should not be a big problem.)
 514                                 int64_t audio_pts_as_video_pts = av_rescale_q(audio_pts, audio_timebase, video_timebase);
 515                                 audio_frame->received_timestamp = compute_frame_start(audio_pts_as_video_pts, pts_origin, video_timebase, start, rate);
 516
 517                                 if (audio_frame->len != 0) {
 518                                         // The received timestamps in Nageru are measured after we've just received the frame.
 519                                         // However, pts (especially audio pts) is at the _beginning_ of the frame.
 520                                         // If we have locked audio, the distinction doesn't really matter, as pts is
 521                                         // on a relative scale and a fixed offset is fine. But if we don't, we will have
 522                                         // a different number of samples each time, which will cause huge audio jitter
 523                                         // and throw off the resampler.
 524                                         //
 525                                         // In a sense, we should have compensated by adding the frame and audio lengths
 526                                         // to video_frame->received_timestamp and audio_frame->received_timestamp respectively,
 527                                         // but that would mean extra waiting in sleep_until(). All we need is that they
 528                                         // are correct relative to each other, though (and to the other frames we send),
 529                                         // so just align the end of the audio frame, and we're fine.
 530                                         size_t num_samples = (audio_frame->len * 8) / audio_format.bits_per_sample / audio_format.num_channels;
 531                                         double offset = double(num_samples) / OUTPUT_FREQUENCY -
 532                                                 double(video_format.frame_rate_den) / video_format.frame_rate_nom;
 533                                         audio_frame->received_timestamp += duration_cast<steady_clock::duration>(duration<double>(offset));
 534                                 }
 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,
 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                         }
 682                 } else {
 683                         eof = true;  // Or error, but ignore that for the time being.
 684                 }
 685
 686                 // Decode audio, if any.
 687                 if (has_audio) {
 688                         for ( ;; ) {
 689                                 int err = avcodec_receive_frame(audio_codec_ctx, audio_avframe.get());
 690                                 if (err == 0) {
 691                                         if (*audio_pts == -1) {
 692                                                 *audio_pts = audio_avframe->pts;
 693                                         }
 694                                         convert_audio(audio_avframe.get(), audio_frame, audio_format);
 695                                 } else if (err == AVERROR(EAGAIN)) {
 696                                         break;
 697                                 } else {
 698                                         fprintf(stderr, "%s: Cannot receive frame from audio codec.\n", pathname.c_str());
 699                                         *error = true;
 700                                         return AVFrameWithDeleter(nullptr);
 701                                 }
 702                         }
 703                 }
 704
 705                 if (video_codec_ctx != nullptr) {
 706                         // Decode video, if we have a frame.
 707                         int err = avcodec_receive_frame(video_codec_ctx, video_avframe.get());
 708                         if (err == 0) {
 709                                 frame_finished = true;
 710                                 break;
 711                         } else if (err != AVERROR(EAGAIN)) {
 712                                 fprintf(stderr, "%s: Cannot receive frame from video codec.\n", pathname.c_str());
 713                                 *error = true;
 714                                 return AVFrameWithDeleter(nullptr);
 715                         }
 716                 } else {
 717                         return AVFrameWithDeleter(nullptr);
 718                 }
 719         } while (!eof);
 720
 721         if (frame_finished)
 722                 return video_avframe;
 723         else
 724                 return AVFrameWithDeleter(nullptr);
 725 }
 726
 727 void FFmpegCapture::convert_audio(const AVFrame *audio_avframe, FrameAllocator::Frame *audio_frame, AudioFormat *audio_format)
 728 {
 729         // Decide on a format. If there already is one in this audio frame,
 730         // we're pretty much forced to use it. If not, we try to find an exact match.
 731         // If that still doesn't work, we default to 32-bit signed chunked
 732         // (float would be nice, but there's really no way to signal that yet).
 733         AVSampleFormat dst_format;
 734         if (audio_format->bits_per_sample == 0) {
 735                 switch (audio_avframe->format) {
 736                 case AV_SAMPLE_FMT_S16:
 737                 case AV_SAMPLE_FMT_S16P:
 738                         audio_format->bits_per_sample = 16;
 739                         dst_format = AV_SAMPLE_FMT_S16;
 740                         break;
 741                 case AV_SAMPLE_FMT_S32:
 742                 case AV_SAMPLE_FMT_S32P:
 743                 default:
 744                         audio_format->bits_per_sample = 32;
 745                         dst_format = AV_SAMPLE_FMT_S32;
 746                         break;
 747                 }
 748         } else if (audio_format->bits_per_sample == 16) {
 749                 dst_format = AV_SAMPLE_FMT_S16;
 750         } else if (audio_format->bits_per_sample == 32) {
 751                 dst_format = AV_SAMPLE_FMT_S32;
 752         } else {
 753                 assert(false);
 754         }
 755         audio_format->num_channels = 2;
 756
 757         int64_t channel_layout = audio_avframe->channel_layout;
 758         if (channel_layout == 0) {
 759                 channel_layout = av_get_default_channel_layout(audio_avframe->channels);
 760         }
 761
 762         if (resampler == nullptr ||
 763             audio_avframe->format != last_src_format ||
 764             dst_format != last_dst_format ||
 765             channel_layout != last_channel_layout ||
 766             av_frame_get_sample_rate(audio_avframe) != last_sample_rate) {
 767                 avresample_free(&resampler);
 768                 resampler = avresample_alloc_context();
 769                 if (resampler == nullptr) {
 770                         fprintf(stderr, "Allocating resampler failed.\n");
 771                         exit(1);
 772                 }
 773
 774                 av_opt_set_int(resampler, "in_channel_layout",  channel_layout,                             0);
 775                 av_opt_set_int(resampler, "out_channel_layout", AV_CH_LAYOUT_STEREO_DOWNMIX,                0);
 776                 av_opt_set_int(resampler, "in_sample_rate",     av_frame_get_sample_rate(audio_avframe),    0);
 777                 av_opt_set_int(resampler, "out_sample_rate",    OUTPUT_FREQUENCY,                           0);
 778                 av_opt_set_int(resampler, "in_sample_fmt",      audio_avframe->format,                      0);
 779                 av_opt_set_int(resampler, "out_sample_fmt",     dst_format,                                 0);
 780
 781                 if (avresample_open(resampler) < 0) {
 782                         fprintf(stderr, "Could not open resample context.\n");
 783                         exit(1);
 784                 }
 785
 786                 last_src_format = AVSampleFormat(audio_avframe->format);
 787                 last_dst_format = dst_format;
 788                 last_channel_layout = channel_layout;
 789                 last_sample_rate = av_frame_get_sample_rate(audio_avframe);
 790         }
 791
 792         size_t bytes_per_sample = (audio_format->bits_per_sample / 8) * 2;
 793         size_t num_samples_room = (audio_frame->size - audio_frame->len) / bytes_per_sample;
 794
 795         uint8_t *data = audio_frame->data + audio_frame->len;
 796         int out_samples = avresample_convert(resampler, &data, 0, num_samples_room,
 797                 const_cast<uint8_t **>(audio_avframe->data), audio_avframe->linesize[0], audio_avframe->nb_samples);
 798         if (out_samples < 0) {
 799                 fprintf(stderr, "Audio conversion failed.\n");
 800                 exit(1);
 801         }
 802
 803         audio_frame->len += out_samples * bytes_per_sample;
 804 }
 805
 806 VideoFormat FFmpegCapture::construct_video_format(const AVFrame *frame, AVRational video_timebase)
 807 {
 808         VideoFormat video_format;
 809         video_format.width = width;
 810         video_format.height = height;
 811         if (pixel_format == bmusb::PixelFormat_8BitBGRA) {
 812                 video_format.stride = width * 4;
 813         } else if (pixel_format == FFmpegCapture::PixelFormat_NV12) {
 814                 video_format.stride = width;
 815         } else {
 816                 assert(pixel_format == bmusb::PixelFormat_8BitYCbCrPlanar);
 817                 video_format.stride = width;
 818         }
 819         video_format.frame_rate_nom = video_timebase.den;
 820         video_format.frame_rate_den = av_frame_get_pkt_duration(frame) * video_timebase.num;
 821         if (video_format.frame_rate_nom == 0 || video_format.frame_rate_den == 0) {
 822                 // Invalid frame rate.
 823                 video_format.frame_rate_nom = 60;
 824                 video_format.frame_rate_den = 1;
 825         }
 826         video_format.has_signal = true;
 827         video_format.is_connected = true;
 828         return video_format;
 829 }
 830
 831 UniqueFrame FFmpegCapture::make_video_frame(const AVFrame *frame, const string &pathname, bool *error)
 832 {
 833         *error = false;
 834
 835         UniqueFrame video_frame(video_frame_allocator->alloc_frame());
 836         if (video_frame->data == nullptr) {
 837                 return video_frame;
 838         }
 839
 840         if (sws_ctx == nullptr ||
 841             sws_last_width != frame->width ||
 842             sws_last_height != frame->height ||
 843             sws_last_src_format != frame->format) {
 844                 sws_dst_format = decide_dst_format(AVPixelFormat(frame->format), pixel_format);
 845                 sws_ctx.reset(
 846                         sws_getContext(frame->width, frame->height, AVPixelFormat(frame->format),
 847                                 width, height, sws_dst_format,
 848                                 SWS_BICUBIC, nullptr, nullptr, nullptr));
 849                 sws_last_width = frame->width;
 850                 sws_last_height = frame->height;
 851                 sws_last_src_format = frame->format;
 852         }
 853         if (sws_ctx == nullptr) {
 854                 fprintf(stderr, "%s: Could not create scaler context\n", pathname.c_str());
 855                 *error = true;
 856                 return video_frame;
 857         }
 858
 859         uint8_t *pic_data[4] = { nullptr, nullptr, nullptr, nullptr };
 860         int linesizes[4] = { 0, 0, 0, 0 };
 861         if (pixel_format == bmusb::PixelFormat_8BitBGRA) {
 862                 pic_data[0] = video_frame->data;
 863                 linesizes[0] = width * 4;
 864                 video_frame->len = (width * 4) * height;
 865         } else if (pixel_format == PixelFormat_NV12) {
 866                 pic_data[0] = video_frame->data;
 867                 linesizes[0] = width;
 868
 869                 pic_data[1] = pic_data[0] + width * height;
 870                 linesizes[1] = width;
 871
 872                 video_frame->len = (width * 2) * height;
 873
 874                 const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(sws_dst_format);
 875                 current_frame_ycbcr_format = decode_ycbcr_format(desc, frame);
 876         } else {
 877                 assert(pixel_format == bmusb::PixelFormat_8BitYCbCrPlanar);
 878                 const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(sws_dst_format);
 879
 880                 int chroma_width = AV_CEIL_RSHIFT(int(width), desc->log2_chroma_w);
 881                 int chroma_height = AV_CEIL_RSHIFT(int(height), desc->log2_chroma_h);
 882
 883                 pic_data[0] = video_frame->data;
 884                 linesizes[0] = width;
 885
 886                 pic_data[1] = pic_data[0] + width * height;
 887                 linesizes[1] = chroma_width;
 888
 889                 pic_data[2] = pic_data[1] + chroma_width * chroma_height;
 890                 linesizes[2] = chroma_width;
 891
 892                 video_frame->len = width * height + 2 * chroma_width * chroma_height;
 893
 894                 current_frame_ycbcr_format = decode_ycbcr_format(desc, frame);
 895         }
 896         sws_scale(sws_ctx.get(), frame->data, frame->linesize, 0, frame->height, pic_data, linesizes);
 897
 898         return video_frame;
 899 }
 900
 901 int FFmpegCapture::interrupt_cb_thunk(void *unique)
 902 {
 903         return reinterpret_cast<FFmpegCapture *>(unique)->interrupt_cb();
 904 }
 905
 906 int FFmpegCapture::interrupt_cb()
 907 {
 908         return should_interrupt.load();
 909 }