git.sesse.net Git - nageru/blob - futatabi/video_stream.cpp

   1 #include "video_stream.h"
   2
   3 extern "C" {
   4 #include <libavformat/avformat.h>
   5 #include <libavformat/avio.h>
   6 }
   7
   8 #include "chroma_subsampler.h"
   9 #include "exif_parser.h"
  10 #include "flags.h"
  11 #include "flow.h"
  12 #include "jpeg_frame_view.h"
  13 #include "movit/util.h"
  14 #include "player.h"
  15 #include "shared/context.h"
  16 #include "shared/httpd.h"
  17 #include "shared/metrics.h"
  18 #include "shared/shared_defs.h"
  19 #include "shared/mux.h"
  20 #include "util.h"
  21 #include "ycbcr_converter.h"
  22
  23 #include <epoxy/glx.h>
  24 #include <jpeglib.h>
  25 #include <unistd.h>
  26
  27 using namespace movit;
  28 using namespace std;
  29 using namespace std::chrono;
  30
  31 namespace {
  32
  33 once_flag video_metrics_inited;
  34 Summary metric_jpeg_encode_time_seconds;
  35 Summary metric_fade_latency_seconds;
  36 Summary metric_interpolation_latency_seconds;
  37 Summary metric_fade_fence_wait_time_seconds;
  38 Summary metric_interpolation_fence_wait_time_seconds;
  39
  40 }  // namespace
  41
  42 extern HTTPD *global_httpd;
  43
  44 struct VectorDestinationManager {
  45         jpeg_destination_mgr pub;
  46         string dest;
  47
  48         VectorDestinationManager()
  49         {
  50                 pub.init_destination = init_destination_thunk;
  51                 pub.empty_output_buffer = empty_output_buffer_thunk;
  52                 pub.term_destination = term_destination_thunk;
  53         }
  54
  55         static void init_destination_thunk(j_compress_ptr ptr)
  56         {
  57                 ((VectorDestinationManager *)(ptr->dest))->init_destination();
  58         }
  59
  60         inline void init_destination()
  61         {
  62                 make_room(0);
  63         }
  64
  65         static boolean empty_output_buffer_thunk(j_compress_ptr ptr)
  66         {
  67                 return ((VectorDestinationManager *)(ptr->dest))->empty_output_buffer();
  68         }
  69
  70         inline bool empty_output_buffer()
  71         {
  72                 make_room(dest.size());  // Should ignore pub.free_in_buffer!
  73                 return true;
  74         }
  75
  76         inline void make_room(size_t bytes_used)
  77         {
  78                 dest.resize(bytes_used + 4096);
  79                 dest.resize(dest.capacity());
  80                 pub.next_output_byte = (uint8_t *)dest.data() + bytes_used;
  81                 pub.free_in_buffer = dest.size() - bytes_used;
  82         }
  83
  84         static void term_destination_thunk(j_compress_ptr ptr)
  85         {
  86                 ((VectorDestinationManager *)(ptr->dest))->term_destination();
  87         }
  88
  89         inline void term_destination()
  90         {
  91                 dest.resize(dest.size() - pub.free_in_buffer);
  92         }
  93 };
  94 static_assert(std::is_standard_layout<VectorDestinationManager>::value, "");
  95
  96 string encode_jpeg(const uint8_t *y_data, const uint8_t *cb_data, const uint8_t *cr_data, unsigned width, unsigned height, const string exif_data)
  97 {
  98         steady_clock::time_point start = steady_clock::now();
  99         VectorDestinationManager dest;
 100
 101         jpeg_compress_struct cinfo;
 102         jpeg_error_mgr jerr;
 103         cinfo.err = jpeg_std_error(&jerr);
 104         jpeg_create_compress(&cinfo);
 105
 106         cinfo.dest = (jpeg_destination_mgr *)&dest;
 107         cinfo.input_components = 3;
 108         cinfo.in_color_space = JCS_RGB;
 109         jpeg_set_defaults(&cinfo);
 110         constexpr int quality = 90;
 111         jpeg_set_quality(&cinfo, quality, /*force_baseline=*/false);
 112
 113         cinfo.image_width = width;
 114         cinfo.image_height = height;
 115         cinfo.raw_data_in = true;
 116         jpeg_set_colorspace(&cinfo, JCS_YCbCr);
 117         cinfo.comp_info[0].h_samp_factor = 2;
 118         cinfo.comp_info[0].v_samp_factor = 1;
 119         cinfo.comp_info[1].h_samp_factor = 1;
 120         cinfo.comp_info[1].v_samp_factor = 1;
 121         cinfo.comp_info[2].h_samp_factor = 1;
 122         cinfo.comp_info[2].v_samp_factor = 1;
 123         cinfo.CCIR601_sampling = true;  // Seems to be mostly ignored by libjpeg, though.
 124         jpeg_start_compress(&cinfo, true);
 125
 126         // This comment marker is private to FFmpeg. It signals limited Y'CbCr range
 127         // (and nothing else).
 128         jpeg_write_marker(&cinfo, JPEG_COM, (const JOCTET *)"CS=ITU601", strlen("CS=ITU601"));
 129
 130         if (!exif_data.empty()) {
 131                 jpeg_write_marker(&cinfo, JPEG_APP0 + 1, (const JOCTET *)exif_data.data(), exif_data.size());
 132         }
 133
 134         JSAMPROW yptr[8], cbptr[8], crptr[8];
 135         JSAMPARRAY data[3] = { yptr, cbptr, crptr };
 136         for (unsigned y = 0; y < height; y += 8) {
 137                 for (unsigned yy = 0; yy < 8; ++yy) {
 138                         yptr[yy] = const_cast<JSAMPROW>(&y_data[(y + yy) * width]);
 139                         cbptr[yy] = const_cast<JSAMPROW>(&cb_data[(y + yy) * width / 2]);
 140                         crptr[yy] = const_cast<JSAMPROW>(&cr_data[(y + yy) * width / 2]);
 141                 }
 142
 143                 jpeg_write_raw_data(&cinfo, data, /*num_lines=*/8);
 144         }
 145
 146         jpeg_finish_compress(&cinfo);
 147         jpeg_destroy_compress(&cinfo);
 148
 149         steady_clock::time_point stop = steady_clock::now();
 150         metric_jpeg_encode_time_seconds.count_event(duration<double>(stop - start).count());
 151
 152         return move(dest.dest);
 153 }
 154
 155 VideoStream::VideoStream(AVFormatContext *file_avctx)
 156         : avctx(file_avctx), output_fast_forward(file_avctx != nullptr)
 157 {
 158         call_once(video_metrics_inited, [] {
 159                 vector<double> quantiles{ 0.01, 0.1, 0.25, 0.5, 0.75, 0.9, 0.99 };
 160                 metric_jpeg_encode_time_seconds.init(quantiles, 60.0);
 161                 global_metrics.add("jpeg_encode_time_seconds", &metric_jpeg_encode_time_seconds);
 162                 metric_fade_fence_wait_time_seconds.init(quantiles, 60.0);
 163                 global_metrics.add("fade_fence_wait_time_seconds", &metric_fade_fence_wait_time_seconds);
 164                 metric_interpolation_fence_wait_time_seconds.init(quantiles, 60.0);
 165                 global_metrics.add("interpolation_fence_wait_time_seconds", &metric_interpolation_fence_wait_time_seconds);
 166                 metric_fade_latency_seconds.init(quantiles, 60.0);
 167                 global_metrics.add("fade_latency_seconds", &metric_fade_latency_seconds);
 168                 metric_interpolation_latency_seconds.init(quantiles, 60.0);
 169                 global_metrics.add("interpolation_latency_seconds", &metric_interpolation_latency_seconds);
 170         });
 171
 172         ycbcr_converter.reset(new YCbCrConverter(YCbCrConverter::OUTPUT_TO_DUAL_YCBCR, /*resource_pool=*/nullptr));
 173         ycbcr_semiplanar_converter.reset(new YCbCrConverter(YCbCrConverter::OUTPUT_TO_SEMIPLANAR, /*resource_pool=*/nullptr));
 174
 175         GLuint input_tex[num_interpolate_slots], gray_tex[num_interpolate_slots];
 176         GLuint fade_y_output_tex[num_interpolate_slots], fade_cbcr_output_tex[num_interpolate_slots];
 177         GLuint cb_tex[num_interpolate_slots], cr_tex[num_interpolate_slots];
 178
 179         glCreateTextures(GL_TEXTURE_2D_ARRAY, num_interpolate_slots, input_tex);
 180         glCreateTextures(GL_TEXTURE_2D_ARRAY, num_interpolate_slots, gray_tex);
 181         glCreateTextures(GL_TEXTURE_2D, num_interpolate_slots, fade_y_output_tex);
 182         glCreateTextures(GL_TEXTURE_2D, num_interpolate_slots, fade_cbcr_output_tex);
 183         glCreateTextures(GL_TEXTURE_2D, num_interpolate_slots, cb_tex);
 184         glCreateTextures(GL_TEXTURE_2D, num_interpolate_slots, cr_tex);
 185         check_error();
 186
 187         size_t width = global_flags.width, height = global_flags.height;
 188         int levels = find_num_levels(width, height);
 189         for (size_t i = 0; i < num_interpolate_slots; ++i) {
 190                 glTextureStorage3D(input_tex[i], levels, GL_RGBA8, width, height, 2);
 191                 check_error();
 192                 glTextureStorage3D(gray_tex[i], levels, GL_R8, width, height, 2);
 193                 check_error();
 194                 glTextureStorage2D(fade_y_output_tex[i], 1, GL_R8, width, height);
 195                 check_error();
 196                 glTextureStorage2D(fade_cbcr_output_tex[i], 1, GL_RG8, width, height);
 197                 check_error();
 198                 glTextureStorage2D(cb_tex[i], 1, GL_R8, width / 2, height);
 199                 check_error();
 200                 glTextureStorage2D(cr_tex[i], 1, GL_R8, width / 2, height);
 201                 check_error();
 202
 203                 unique_ptr<InterpolatedFrameResources> resource(new InterpolatedFrameResources);
 204                 resource->owner = this;
 205                 resource->input_tex = input_tex[i];
 206                 resource->gray_tex = gray_tex[i];
 207                 resource->fade_y_output_tex = fade_y_output_tex[i];
 208                 resource->fade_cbcr_output_tex = fade_cbcr_output_tex[i];
 209                 resource->cb_tex = cb_tex[i];
 210                 resource->cr_tex = cr_tex[i];
 211                 glCreateFramebuffers(2, resource->input_fbos);
 212                 check_error();
 213                 glCreateFramebuffers(1, &resource->fade_fbo);
 214                 check_error();
 215
 216                 glNamedFramebufferTextureLayer(resource->input_fbos[0], GL_COLOR_ATTACHMENT0, input_tex[i], 0, 0);
 217                 check_error();
 218                 glNamedFramebufferTextureLayer(resource->input_fbos[0], GL_COLOR_ATTACHMENT1, gray_tex[i], 0, 0);
 219                 check_error();
 220                 glNamedFramebufferTextureLayer(resource->input_fbos[1], GL_COLOR_ATTACHMENT0, input_tex[i], 0, 1);
 221                 check_error();
 222                 glNamedFramebufferTextureLayer(resource->input_fbos[1], GL_COLOR_ATTACHMENT1, gray_tex[i], 0, 1);
 223                 check_error();
 224                 glNamedFramebufferTexture(resource->fade_fbo, GL_COLOR_ATTACHMENT0, fade_y_output_tex[i], 0);
 225                 check_error();
 226                 glNamedFramebufferTexture(resource->fade_fbo, GL_COLOR_ATTACHMENT1, fade_cbcr_output_tex[i], 0);
 227                 check_error();
 228
 229                 GLuint bufs[] = { GL_COLOR_ATTACHMENT0, GL_COLOR_ATTACHMENT1 };
 230                 glNamedFramebufferDrawBuffers(resource->input_fbos[0], 2, bufs);
 231                 check_error();
 232                 glNamedFramebufferDrawBuffers(resource->input_fbos[1], 2, bufs);
 233                 check_error();
 234                 glNamedFramebufferDrawBuffers(resource->fade_fbo, 2, bufs);
 235                 check_error();
 236
 237                 glCreateBuffers(1, &resource->pbo);
 238                 check_error();
 239                 glNamedBufferStorage(resource->pbo, width * height * 4, nullptr, GL_MAP_READ_BIT | GL_MAP_PERSISTENT_BIT);
 240                 check_error();
 241                 resource->pbo_contents = glMapNamedBufferRange(resource->pbo, 0, width * height * 4, GL_MAP_READ_BIT | GL_MAP_PERSISTENT_BIT);
 242                 interpolate_resources.push_back(move(resource));
 243         }
 244
 245         check_error();
 246
 247         OperatingPoint op;
 248         if (global_flags.interpolation_quality == 0 ||
 249             global_flags.interpolation_quality == 1) {
 250                 op = operating_point1;
 251         } else if (global_flags.interpolation_quality == 2) {
 252                 op = operating_point2;
 253         } else if (global_flags.interpolation_quality == 3) {
 254                 op = operating_point3;
 255         } else if (global_flags.interpolation_quality == 4) {
 256                 op = operating_point4;
 257         } else {
 258                 // Quality 0 will be changed to 1 in flags.cpp.
 259                 assert(false);
 260         }
 261
 262         compute_flow.reset(new DISComputeFlow(width, height, op));
 263         interpolate.reset(new Interpolate(op, /*split_ycbcr_output=*/true));
 264         interpolate_no_split.reset(new Interpolate(op, /*split_ycbcr_output=*/false));
 265         chroma_subsampler.reset(new ChromaSubsampler);
 266         check_error();
 267
 268         // The “last frame” is initially black.
 269         unique_ptr<uint8_t[]> y(new uint8_t[global_flags.width * global_flags.height]);
 270         unique_ptr<uint8_t[]> cb_or_cr(new uint8_t[(global_flags.width / 2) * global_flags.height]);
 271         memset(y.get(), 16, global_flags.width * global_flags.height);
 272         memset(cb_or_cr.get(), 128, (global_flags.width / 2) * global_flags.height);
 273         last_frame = encode_jpeg(y.get(), cb_or_cr.get(), cb_or_cr.get(), global_flags.width, global_flags.height, /*exif_data=*/"");
 274
 275         if (file_avctx != nullptr) {
 276                 with_subtitles = Mux::WITHOUT_SUBTITLES;
 277         } else {
 278                 with_subtitles = Mux::WITH_SUBTITLES;
 279         }
 280 }
 281
 282 VideoStream::~VideoStream()
 283 {
 284         if (last_flow_tex != 0) {
 285                 compute_flow->release_texture(last_flow_tex);
 286         }
 287
 288         for (const unique_ptr<InterpolatedFrameResources> &resource : interpolate_resources) {
 289                 glUnmapNamedBuffer(resource->pbo);
 290                 check_error();
 291                 glDeleteBuffers(1, &resource->pbo);
 292                 check_error();
 293                 glDeleteFramebuffers(2, resource->input_fbos);
 294                 check_error();
 295                 glDeleteFramebuffers(1, &resource->fade_fbo);
 296                 check_error();
 297                 glDeleteTextures(1, &resource->input_tex);
 298                 check_error();
 299                 glDeleteTextures(1, &resource->gray_tex);
 300                 check_error();
 301                 glDeleteTextures(1, &resource->fade_y_output_tex);
 302                 check_error();
 303                 glDeleteTextures(1, &resource->fade_cbcr_output_tex);
 304                 check_error();
 305                 glDeleteTextures(1, &resource->cb_tex);
 306                 check_error();
 307                 glDeleteTextures(1, &resource->cr_tex);
 308                 check_error();
 309         }
 310         assert(interpolate_resources.size() == num_interpolate_slots);
 311 }
 312
 313 void VideoStream::start()
 314 {
 315         if (avctx == nullptr) {
 316                 avctx = avformat_alloc_context();
 317
 318                 // We use Matroska, because it's pretty much the only mux where FFmpeg
 319                 // allows writing chroma location to override JFIF's default center placement.
 320                 // (Note that at the time of writing, however, FFmpeg does not correctly
 321                 // _read_ this information!)
 322                 avctx->oformat = av_guess_format("matroska", nullptr, nullptr);
 323
 324                 uint8_t *buf = (uint8_t *)av_malloc(MUX_BUFFER_SIZE);
 325                 avctx->pb = avio_alloc_context(buf, MUX_BUFFER_SIZE, 1, this, nullptr, nullptr, nullptr);
 326                 avctx->pb->write_data_type = &VideoStream::write_packet2_thunk;
 327                 avctx->pb->ignore_boundary_point = 1;
 328
 329                 avctx->flags = AVFMT_FLAG_CUSTOM_IO;
 330         }
 331
 332         AVCodecParameters *audio_codecpar = avcodec_parameters_alloc();
 333
 334         audio_codecpar->codec_type = AVMEDIA_TYPE_AUDIO;
 335         audio_codecpar->codec_id = AV_CODEC_ID_PCM_S32LE;
 336         audio_codecpar->channel_layout = AV_CH_LAYOUT_STEREO;
 337         audio_codecpar->channels = 2;
 338         audio_codecpar->sample_rate = OUTPUT_FREQUENCY;
 339
 340         size_t width = global_flags.width, height = global_flags.height;  // Doesn't matter for MJPEG.
 341         mux.reset(new Mux(avctx, width, height, Mux::CODEC_MJPEG, /*video_extradata=*/"", audio_codecpar,
 342                           AVCOL_SPC_BT709, COARSE_TIMEBASE, /*write_callback=*/nullptr, Mux::WRITE_FOREGROUND, {}, with_subtitles));
 343
 344         avcodec_parameters_free(&audio_codecpar);
 345         encode_thread = thread(&VideoStream::encode_thread_func, this);
 346 }
 347
 348 void VideoStream::stop()
 349 {
 350         should_quit = true;
 351         queue_changed.notify_all();
 352         clear_queue();
 353         encode_thread.join();
 354 }
 355
 356 void VideoStream::clear_queue()
 357 {
 358         deque<QueuedFrame> q;
 359
 360         {
 361                 lock_guard<mutex> lock(queue_lock);
 362                 q = move(frame_queue);
 363         }
 364
 365         // These are not RAII-ed, unfortunately, so we'll need to clean them ourselves.
 366         // Note that release_texture() is thread-safe.
 367         for (const QueuedFrame &qf : q) {
 368                 if (qf.type == QueuedFrame::INTERPOLATED ||
 369                     qf.type == QueuedFrame::FADED_INTERPOLATED) {
 370                         if (qf.flow_tex != 0) {
 371                                 compute_flow->release_texture(qf.flow_tex);
 372                         }
 373                 }
 374                 if (qf.type == QueuedFrame::INTERPOLATED) {
 375                         interpolate->release_texture(qf.output_tex);
 376                         interpolate->release_texture(qf.cbcr_tex);
 377                 }
 378         }
 379
 380         // Destroy q outside the mutex, as that would be a double-lock.
 381 }
 382
 383 void VideoStream::schedule_original_frame(steady_clock::time_point local_pts,
 384                                           int64_t output_pts, function<void()> &&display_func,
 385                                           QueueSpotHolder &&queue_spot_holder,
 386                                           FrameOnDisk frame, const string &subtitle, bool include_audio)
 387 {
 388         fprintf(stderr, "output_pts=%" PRId64 "  original      input_pts=%" PRId64 "\n", output_pts, frame.pts);
 389
 390         QueuedFrame qf;
 391         qf.local_pts = local_pts;
 392         qf.type = QueuedFrame::ORIGINAL;
 393         qf.output_pts = output_pts;
 394         qf.display_func = move(display_func);
 395         qf.queue_spot_holder = move(queue_spot_holder);
 396         qf.subtitle = subtitle;
 397         FrameReader::Frame read_frame = frame_reader.read_frame(frame, /*read_video=*/true, include_audio);
 398         qf.encoded_jpeg.reset(new string(move(read_frame.video)));
 399         qf.audio = move(read_frame.audio);
 400
 401         lock_guard<mutex> lock(queue_lock);
 402         frame_queue.push_back(move(qf));
 403         queue_changed.notify_all();
 404 }
 405
 406 void VideoStream::schedule_faded_frame(steady_clock::time_point local_pts, int64_t output_pts,
 407                                        function<void()> &&display_func,
 408                                        QueueSpotHolder &&queue_spot_holder,
 409                                        FrameOnDisk frame1_spec, FrameOnDisk frame2_spec,
 410                                        float fade_alpha, const string &subtitle)
 411 {
 412         fprintf(stderr, "output_pts=%" PRId64 "  faded         input_pts=%" PRId64 ",%" PRId64 "  fade_alpha=%.2f\n", output_pts, frame1_spec.pts, frame2_spec.pts, fade_alpha);
 413
 414         // Get the temporary OpenGL resources we need for doing the fade.
 415         // (We share these with interpolated frames, which is slightly
 416         // overkill, but there's no need to waste resources on keeping
 417         // separate pools around.)
 418         BorrowedInterpolatedFrameResources resources;
 419         {
 420                 lock_guard<mutex> lock(queue_lock);
 421                 if (interpolate_resources.empty()) {
 422                         fprintf(stderr, "WARNING: Too many interpolated frames already in transit; dropping one.\n");
 423                         return;
 424                 }
 425                 resources = BorrowedInterpolatedFrameResources(interpolate_resources.front().release());
 426                 interpolate_resources.pop_front();
 427         }
 428
 429         bool did_decode;
 430
 431         shared_ptr<Frame> frame1 = decode_jpeg_with_cache(frame1_spec, DECODE_IF_NOT_IN_CACHE, &frame_reader, &did_decode);
 432         shared_ptr<Frame> frame2 = decode_jpeg_with_cache(frame2_spec, DECODE_IF_NOT_IN_CACHE, &frame_reader, &did_decode);
 433
 434         ycbcr_semiplanar_converter->prepare_chain_for_fade(frame1, frame2, fade_alpha)->render_to_fbo(resources->fade_fbo, global_flags.width, global_flags.height);
 435
 436         QueuedFrame qf;
 437         qf.local_pts = local_pts;
 438         qf.type = QueuedFrame::FADED;
 439         qf.output_pts = output_pts;
 440         qf.frame1 = frame1_spec;
 441         qf.display_func = move(display_func);
 442         qf.queue_spot_holder = move(queue_spot_holder);
 443         qf.subtitle = subtitle;
 444
 445         qf.secondary_frame = frame2_spec;
 446
 447         // Subsample and split Cb/Cr.
 448         chroma_subsampler->subsample_chroma(resources->fade_cbcr_output_tex, global_flags.width, global_flags.height, resources->cb_tex, resources->cr_tex);
 449
 450         // Read it down (asynchronously) to the CPU.
 451         glPixelStorei(GL_PACK_ROW_LENGTH, 0);
 452         glBindBuffer(GL_PIXEL_PACK_BUFFER, resources->pbo);
 453         check_error();
 454         glGetTextureImage(resources->fade_y_output_tex, 0, GL_RED, GL_UNSIGNED_BYTE, global_flags.width * global_flags.height * 4, BUFFER_OFFSET(0));
 455         check_error();
 456         glGetTextureImage(resources->cb_tex, 0, GL_RED, GL_UNSIGNED_BYTE, global_flags.width * global_flags.height * 3, BUFFER_OFFSET(global_flags.width * global_flags.height));
 457         check_error();
 458         glGetTextureImage(resources->cr_tex, 0, GL_RED, GL_UNSIGNED_BYTE, global_flags.width * global_flags.height * 3 - (global_flags.width / 2) * global_flags.height, BUFFER_OFFSET(global_flags.width * global_flags.height + (global_flags.width / 2) * global_flags.height));
 459         check_error();
 460         glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
 461
 462         // Set a fence we can wait for to make sure the CPU sees the read.
 463         glMemoryBarrier(GL_CLIENT_MAPPED_BUFFER_BARRIER_BIT);
 464         check_error();
 465         qf.fence_created = steady_clock::now();
 466         qf.fence = RefCountedGLsync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
 467         check_error();
 468         qf.resources = move(resources);
 469         qf.local_pts = local_pts;
 470
 471         lock_guard<mutex> lock(queue_lock);
 472         frame_queue.push_back(move(qf));
 473         queue_changed.notify_all();
 474 }
 475
 476 void VideoStream::schedule_interpolated_frame(steady_clock::time_point local_pts,
 477                                               int64_t output_pts, function<void(shared_ptr<Frame>)> &&display_func,
 478                                               QueueSpotHolder &&queue_spot_holder,
 479                                               FrameOnDisk frame1, FrameOnDisk frame2,
 480                                               float alpha, FrameOnDisk secondary_frame, float fade_alpha, const string &subtitle,
 481                                               bool play_audio)
 482 {
 483         if (secondary_frame.pts != -1) {
 484                 fprintf(stderr, "output_pts=%" PRId64 "  interpolated  input_pts1=%" PRId64 " input_pts2=%" PRId64 " alpha=%.3f  secondary_pts=%" PRId64 "  fade_alpha=%.2f\n", output_pts, frame1.pts, frame2.pts, alpha, secondary_frame.pts, fade_alpha);
 485         } else {
 486                 fprintf(stderr, "output_pts=%" PRId64 "  interpolated  input_pts1=%" PRId64 " input_pts2=%" PRId64 " alpha=%.3f\n", output_pts, frame1.pts, frame2.pts, alpha);
 487         }
 488
 489         // Get the temporary OpenGL resources we need for doing the interpolation.
 490         BorrowedInterpolatedFrameResources resources;
 491         {
 492                 lock_guard<mutex> lock(queue_lock);
 493                 if (interpolate_resources.empty()) {
 494                         fprintf(stderr, "WARNING: Too many interpolated frames already in transit; dropping one.\n");
 495                         return;
 496                 }
 497                 resources = BorrowedInterpolatedFrameResources(interpolate_resources.front().release());
 498                 interpolate_resources.pop_front();
 499         }
 500
 501         QueuedFrame qf;
 502         qf.type = (secondary_frame.pts == -1) ? QueuedFrame::INTERPOLATED : QueuedFrame::FADED_INTERPOLATED;
 503         qf.output_pts = output_pts;
 504         qf.display_decoded_func = move(display_func);
 505         qf.queue_spot_holder = move(queue_spot_holder);
 506         qf.local_pts = local_pts;
 507         qf.subtitle = subtitle;
 508
 509         if (play_audio) {
 510                 qf.audio = frame_reader.read_frame(frame1, /*read_video=*/false, /*read_audio=*/true).audio;
 511         }
 512
 513         check_error();
 514
 515         // Convert frame0 and frame1 to OpenGL textures.
 516         for (size_t frame_no = 0; frame_no < 2; ++frame_no) {
 517                 FrameOnDisk frame_spec = frame_no == 1 ? frame2 : frame1;
 518                 bool did_decode;
 519                 shared_ptr<Frame> frame = decode_jpeg_with_cache(frame_spec, DECODE_IF_NOT_IN_CACHE, &frame_reader, &did_decode);
 520                 ycbcr_converter->prepare_chain_for_conversion(frame)->render_to_fbo(resources->input_fbos[frame_no], global_flags.width, global_flags.height);
 521                 if (frame_no == 1) {
 522                         qf.exif_data = frame->exif_data;  // Use the white point from the last frame.
 523                 }
 524         }
 525
 526         glGenerateTextureMipmap(resources->input_tex);
 527         check_error();
 528         glGenerateTextureMipmap(resources->gray_tex);
 529         check_error();
 530
 531         GLuint flow_tex;
 532         if (last_flow_tex != 0 && frame1 == last_frame1 && frame2 == last_frame2) {
 533                 // Reuse the flow from previous computation. This frequently happens
 534                 // if we slow down by more than 2x, so that there are multiple interpolated
 535                 // frames between each original.
 536                 flow_tex = last_flow_tex;
 537                 qf.flow_tex = 0;
 538         } else {
 539                 // Cache miss, so release last_flow_tex.
 540                 qf.flow_tex = last_flow_tex;
 541
 542                 // Compute the flow.
 543                 flow_tex = compute_flow->exec(resources->gray_tex, DISComputeFlow::FORWARD_AND_BACKWARD, DISComputeFlow::DO_NOT_RESIZE_FLOW);
 544                 check_error();
 545
 546                 // Store the flow texture for possible reuse next frame.
 547                 last_flow_tex = flow_tex;
 548                 last_frame1 = frame1;
 549                 last_frame2 = frame2;
 550         }
 551
 552         if (secondary_frame.pts != -1) {
 553                 // Fade. First kick off the interpolation.
 554                 tie(qf.output_tex, ignore) = interpolate_no_split->exec(resources->input_tex, resources->gray_tex, flow_tex, global_flags.width, global_flags.height, alpha);
 555                 check_error();
 556
 557                 // Now decode the image we are fading against.
 558                 bool did_decode;
 559                 shared_ptr<Frame> frame2 = decode_jpeg_with_cache(secondary_frame, DECODE_IF_NOT_IN_CACHE, &frame_reader, &did_decode);
 560
 561                 // Then fade against it, putting it into the fade Y' and CbCr textures.
 562                 RGBTriplet neutral_color = get_neutral_color(qf.exif_data);
 563                 ycbcr_semiplanar_converter->prepare_chain_for_fade_from_texture(qf.output_tex, neutral_color, global_flags.width, global_flags.height, frame2, fade_alpha)->render_to_fbo(resources->fade_fbo, global_flags.width, global_flags.height);
 564
 565                 // Subsample and split Cb/Cr.
 566                 chroma_subsampler->subsample_chroma(resources->fade_cbcr_output_tex, global_flags.width, global_flags.height, resources->cb_tex, resources->cr_tex);
 567
 568                 interpolate_no_split->release_texture(qf.output_tex);
 569
 570                 // We already applied the white balance, so don't have the client redo it.
 571                 qf.exif_data.clear();
 572         } else {
 573                 tie(qf.output_tex, qf.cbcr_tex) = interpolate->exec(resources->input_tex, resources->gray_tex, flow_tex, global_flags.width, global_flags.height, alpha);
 574                 check_error();
 575
 576                 // Subsample and split Cb/Cr.
 577                 chroma_subsampler->subsample_chroma(qf.cbcr_tex, global_flags.width, global_flags.height, resources->cb_tex, resources->cr_tex);
 578         }
 579
 580         // We could have released qf.flow_tex here, but to make sure we don't cause a stall
 581         // when trying to reuse it for the next frame, we can just as well hold on to it
 582         // and release it only when the readback is done.
 583         //
 584         // TODO: This is maybe less relevant now that qf.flow_tex contains the texture we used
 585         // _last_ frame, not this one.
 586
 587         // Read it down (asynchronously) to the CPU.
 588         glPixelStorei(GL_PACK_ROW_LENGTH, 0);
 589         glBindBuffer(GL_PIXEL_PACK_BUFFER, resources->pbo);
 590         check_error();
 591         if (secondary_frame.pts != -1) {
 592                 glGetTextureImage(resources->fade_y_output_tex, 0, GL_RED, GL_UNSIGNED_BYTE, global_flags.width * global_flags.height * 4, BUFFER_OFFSET(0));
 593         } else {
 594                 glGetTextureImage(qf.output_tex, 0, GL_RED, GL_UNSIGNED_BYTE, global_flags.width * global_flags.height * 4, BUFFER_OFFSET(0));
 595         }
 596         check_error();
 597         glGetTextureImage(resources->cb_tex, 0, GL_RED, GL_UNSIGNED_BYTE, global_flags.width * global_flags.height * 3, BUFFER_OFFSET(global_flags.width * global_flags.height));
 598         check_error();
 599         glGetTextureImage(resources->cr_tex, 0, GL_RED, GL_UNSIGNED_BYTE, global_flags.width * global_flags.height * 3 - (global_flags.width / 2) * global_flags.height, BUFFER_OFFSET(global_flags.width * global_flags.height + (global_flags.width / 2) * global_flags.height));
 600         check_error();
 601         glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
 602
 603         // Set a fence we can wait for to make sure the CPU sees the read.
 604         glMemoryBarrier(GL_CLIENT_MAPPED_BUFFER_BARRIER_BIT);
 605         check_error();
 606         qf.fence_created = steady_clock::now();
 607         qf.fence = RefCountedGLsync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
 608         check_error();
 609         qf.resources = move(resources);
 610
 611         lock_guard<mutex> lock(queue_lock);
 612         frame_queue.push_back(move(qf));
 613         queue_changed.notify_all();
 614 }
 615
 616 void VideoStream::schedule_refresh_frame(steady_clock::time_point local_pts,
 617                                          int64_t output_pts, function<void()> &&display_func,
 618                                          QueueSpotHolder &&queue_spot_holder, const string &subtitle)
 619 {
 620         QueuedFrame qf;
 621         qf.type = QueuedFrame::REFRESH;
 622         qf.output_pts = output_pts;
 623         qf.display_func = move(display_func);
 624         qf.queue_spot_holder = move(queue_spot_holder);
 625         qf.subtitle = subtitle;
 626
 627         lock_guard<mutex> lock(queue_lock);
 628         frame_queue.push_back(move(qf));
 629         queue_changed.notify_all();
 630 }
 631
 632 void VideoStream::schedule_silence(steady_clock::time_point local_pts, int64_t output_pts,
 633                                    int64_t length_pts, QueueSpotHolder &&queue_spot_holder)
 634 {
 635         QueuedFrame qf;
 636         qf.type = QueuedFrame::SILENCE;
 637         qf.output_pts = output_pts;
 638         qf.queue_spot_holder = move(queue_spot_holder);
 639         qf.silence_length_pts = length_pts;
 640
 641         lock_guard<mutex> lock(queue_lock);
 642         frame_queue.push_back(move(qf));
 643         queue_changed.notify_all();
 644 }
 645
 646 namespace {
 647
 648 shared_ptr<Frame> frame_from_pbo(void *contents, size_t width, size_t height)
 649 {
 650         size_t chroma_width = width / 2;
 651
 652         const uint8_t *y = (const uint8_t *)contents;
 653         const uint8_t *cb = (const uint8_t *)contents + width * height;
 654         const uint8_t *cr = (const uint8_t *)contents + width * height + chroma_width * height;
 655
 656         shared_ptr<Frame> frame(new Frame);
 657         frame->y.reset(new uint8_t[width * height]);
 658         frame->cb.reset(new uint8_t[chroma_width * height]);
 659         frame->cr.reset(new uint8_t[chroma_width * height]);
 660         for (unsigned yy = 0; yy < height; ++yy) {
 661                 memcpy(frame->y.get() + width * yy, y + width * yy, width);
 662                 memcpy(frame->cb.get() + chroma_width * yy, cb + chroma_width * yy, chroma_width);
 663                 memcpy(frame->cr.get() + chroma_width * yy, cr + chroma_width * yy, chroma_width);
 664         }
 665         frame->is_semiplanar = false;
 666         frame->width = width;
 667         frame->height = height;
 668         frame->chroma_subsampling_x = 2;
 669         frame->chroma_subsampling_y = 1;
 670         frame->pitch_y = width;
 671         frame->pitch_chroma = chroma_width;
 672         return frame;
 673 }
 674
 675 }  // namespace
 676
 677 void VideoStream::encode_thread_func()
 678 {
 679         pthread_setname_np(pthread_self(), "VideoStream");
 680         QSurface *surface = create_surface();
 681         QOpenGLContext *context = create_context(surface);
 682         bool ok = make_current(context, surface);
 683         if (!ok) {
 684                 fprintf(stderr, "Video stream couldn't get an OpenGL context\n");
 685                 abort();
 686         }
 687
 688         while (!should_quit) {
 689                 QueuedFrame qf;
 690                 {
 691                         unique_lock<mutex> lock(queue_lock);
 692
 693                         // Wait until we have a frame to play.
 694                         queue_changed.wait(lock, [this] {
 695                                 return !frame_queue.empty() || should_quit;
 696                         });
 697                         if (should_quit) {
 698                                 break;
 699                         }
 700                         steady_clock::time_point frame_start = frame_queue.front().local_pts;
 701
 702                         // Now sleep until the frame is supposed to start (the usual case),
 703                         // _or_ clear_queue() happened.
 704                         bool aborted;
 705                         if (output_fast_forward) {
 706                                 aborted = frame_queue.empty() || frame_queue.front().local_pts != frame_start;
 707                         } else {
 708                                 aborted = queue_changed.wait_until(lock, frame_start, [this, frame_start] {
 709                                         return frame_queue.empty() || frame_queue.front().local_pts != frame_start;
 710                                 });
 711                         }
 712                         if (aborted) {
 713                                 // clear_queue() happened, so don't play this frame after all.
 714                                 continue;
 715                         }
 716                         qf = move(frame_queue.front());
 717                         frame_queue.pop_front();
 718                 }
 719
 720                 // Hack: We mux the subtitle packet one time unit before the actual frame,
 721                 // so that Nageru is sure to get it first.
 722                 if (!qf.subtitle.empty() && with_subtitles == Mux::WITH_SUBTITLES) {
 723                         AVPacket pkt;
 724                         av_init_packet(&pkt);
 725                         pkt.stream_index = mux->get_subtitle_stream_idx();
 726                         assert(pkt.stream_index != -1);
 727                         pkt.data = (uint8_t *)qf.subtitle.data();
 728                         pkt.size = qf.subtitle.size();
 729                         pkt.flags = 0;
 730                         pkt.duration = lrint(TIMEBASE / global_flags.output_framerate);  // Doesn't really matter for Nageru.
 731                         mux->add_packet(pkt, qf.output_pts - 1, qf.output_pts - 1);
 732                 }
 733
 734                 if (qf.type == QueuedFrame::ORIGINAL) {
 735                         // Send the JPEG frame on, unchanged.
 736                         string jpeg = move(*qf.encoded_jpeg);
 737                         AVPacket pkt;
 738                         av_init_packet(&pkt);
 739                         pkt.stream_index = 0;
 740                         pkt.data = (uint8_t *)jpeg.data();
 741                         pkt.size = jpeg.size();
 742                         pkt.flags = AV_PKT_FLAG_KEY;
 743                         mux->add_packet(pkt, qf.output_pts, qf.output_pts);
 744                         last_frame = move(jpeg);
 745
 746                         add_audio_or_silence(qf);
 747                 } else if (qf.type == QueuedFrame::FADED) {
 748                         steady_clock::time_point start = steady_clock::now();
 749                         glClientWaitSync(qf.fence.get(), /*flags=*/0, GL_TIMEOUT_IGNORED);
 750                         steady_clock::time_point stop = steady_clock::now();
 751                         metric_fade_fence_wait_time_seconds.count_event(duration<double>(stop - start).count());
 752                         metric_fade_latency_seconds.count_event(duration<double>(stop - qf.fence_created).count());
 753
 754                         shared_ptr<Frame> frame = frame_from_pbo(qf.resources->pbo_contents, global_flags.width, global_flags.height);
 755                         assert(frame->exif_data.empty());
 756
 757                         // Now JPEG encode it, and send it on to the stream.
 758                         string jpeg = encode_jpeg(frame->y.get(), frame->cb.get(), frame->cr.get(), global_flags.width, global_flags.height, /*exif_data=*/"");
 759
 760                         AVPacket pkt;
 761                         av_init_packet(&pkt);
 762                         pkt.stream_index = 0;
 763                         pkt.data = (uint8_t *)jpeg.data();
 764                         pkt.size = jpeg.size();
 765                         pkt.flags = AV_PKT_FLAG_KEY;
 766                         mux->add_packet(pkt, qf.output_pts, qf.output_pts);
 767                         last_frame = move(jpeg);
 768
 769                         add_audio_or_silence(qf);
 770                 } else if (qf.type == QueuedFrame::INTERPOLATED || qf.type == QueuedFrame::FADED_INTERPOLATED) {
 771                         steady_clock::time_point start = steady_clock::now();
 772                         glClientWaitSync(qf.fence.get(), /*flags=*/0, GL_TIMEOUT_IGNORED);
 773                         steady_clock::time_point stop = steady_clock::now();
 774                         metric_interpolation_fence_wait_time_seconds.count_event(duration<double>(stop - start).count());
 775                         metric_interpolation_latency_seconds.count_event(duration<double>(stop - qf.fence_created).count());
 776
 777                         // Send it on to display.
 778                         shared_ptr<Frame> frame = frame_from_pbo(qf.resources->pbo_contents, global_flags.width, global_flags.height);
 779                         if (qf.display_decoded_func != nullptr) {
 780                                 qf.display_decoded_func(frame);
 781                         }
 782
 783                         // Now JPEG encode it, and send it on to the stream.
 784                         string jpeg = encode_jpeg(frame->y.get(), frame->cb.get(), frame->cr.get(), global_flags.width, global_flags.height, move(qf.exif_data));
 785                         if (qf.flow_tex != 0) {
 786                                 compute_flow->release_texture(qf.flow_tex);
 787                         }
 788                         if (qf.type != QueuedFrame::FADED_INTERPOLATED) {
 789                                 interpolate->release_texture(qf.output_tex);
 790                                 interpolate->release_texture(qf.cbcr_tex);
 791                         }
 792
 793                         AVPacket pkt;
 794                         av_init_packet(&pkt);
 795                         pkt.stream_index = 0;
 796                         pkt.data = (uint8_t *)jpeg.data();
 797                         pkt.size = jpeg.size();
 798                         pkt.flags = AV_PKT_FLAG_KEY;
 799                         mux->add_packet(pkt, qf.output_pts, qf.output_pts);
 800                         last_frame = move(jpeg);
 801
 802                         add_audio_or_silence(qf);
 803                 } else if (qf.type == QueuedFrame::REFRESH) {
 804                         AVPacket pkt;
 805                         av_init_packet(&pkt);
 806                         pkt.stream_index = 0;
 807                         pkt.data = (uint8_t *)last_frame.data();
 808                         pkt.size = last_frame.size();
 809                         pkt.flags = AV_PKT_FLAG_KEY;
 810                         mux->add_packet(pkt, qf.output_pts, qf.output_pts);
 811
 812                         add_audio_or_silence(qf);  // Definitely silence.
 813                 } else if (qf.type == QueuedFrame::SILENCE) {
 814                         add_silence(qf.output_pts, qf.silence_length_pts);
 815                 } else {
 816                         assert(false);
 817                 }
 818                 if (qf.display_func != nullptr) {
 819                         qf.display_func();
 820                 }
 821         }
 822 }
 823
 824 int VideoStream::write_packet2_thunk(void *opaque, uint8_t *buf, int buf_size, AVIODataMarkerType type, int64_t time)
 825 {
 826         VideoStream *video_stream = (VideoStream *)opaque;
 827         return video_stream->write_packet2(buf, buf_size, type, time);
 828 }
 829
 830 int VideoStream::write_packet2(uint8_t *buf, int buf_size, AVIODataMarkerType type, int64_t time)
 831 {
 832         if (type == AVIO_DATA_MARKER_SYNC_POINT || type == AVIO_DATA_MARKER_BOUNDARY_POINT) {
 833                 seen_sync_markers = true;
 834         } else if (type == AVIO_DATA_MARKER_UNKNOWN && !seen_sync_markers) {
 835                 // We don't know if this is a keyframe or not (the muxer could
 836                 // avoid marking it), so we just have to make the best of it.
 837                 type = AVIO_DATA_MARKER_SYNC_POINT;
 838         }
 839
 840         HTTPD::StreamID stream_id{ HTTPD::MAIN_STREAM, 0 };
 841         if (type == AVIO_DATA_MARKER_HEADER) {
 842                 stream_mux_header.append((char *)buf, buf_size);
 843                 global_httpd->set_header(stream_id, stream_mux_header);
 844         } else {
 845                 global_httpd->add_data(stream_id, (char *)buf, buf_size, type == AVIO_DATA_MARKER_SYNC_POINT, time, AVRational{ AV_TIME_BASE, 1 });
 846         }
 847         return buf_size;
 848 }
 849
 850 void VideoStream::add_silence(int64_t pts, int64_t length_pts)
 851 {
 852         // At 59.94, this will never quite add up (even discounting refresh frames,
 853         // which have unpredictable length), but hopefully, the player in the other
 854         // end should be able to stretch silence easily enough.
 855         long num_samples = lrint(length_pts * double(OUTPUT_FREQUENCY) / double(TIMEBASE)) * 2;
 856         uint8_t *zero = (uint8_t *)calloc(num_samples, sizeof(int32_t));
 857
 858         AVPacket pkt;
 859         av_init_packet(&pkt);
 860         pkt.stream_index = 1;
 861         pkt.data = zero;
 862         pkt.size = num_samples * sizeof(int32_t);
 863         pkt.flags = AV_PKT_FLAG_KEY;
 864         mux->add_packet(pkt, pts, pts);
 865
 866         free(zero);
 867 }
 868
 869 void VideoStream::add_audio_or_silence(const QueuedFrame &qf)
 870 {
 871         if (qf.audio.empty()) {
 872                 int64_t frame_length = lrint(double(TIMEBASE) / global_flags.output_framerate);
 873                 add_silence(qf.output_pts, frame_length);
 874         } else {
 875                 AVPacket pkt;
 876                 av_init_packet(&pkt);
 877                 pkt.stream_index = 1;
 878                 pkt.data = (uint8_t *)qf.audio.data();
 879                 pkt.size = qf.audio.size();
 880                 pkt.flags = AV_PKT_FLAG_KEY;
 881                 mux->add_packet(pkt, qf.output_pts, qf.output_pts);
 882         }
 883 }