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

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