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 #include <libavutil/channel_layout.h>
   7 }
   8
   9 #include "chroma_subsampler.h"
  10 #include "exif_parser.h"
  11 #include "flags.h"
  12 #include "flow.h"
  13 #include "jpeg_frame_view.h"
  14 #include "movit/util.h"
  15 #include "pbo_pool.h"
  16 #include "player.h"
  17 #include "shared/context.h"
  18 #include "shared/httpd.h"
  19 #include "shared/metrics.h"
  20 #include "shared/shared_defs.h"
  21 #include "shared/mux.h"
  22 #include "util.h"
  23 #include "ycbcr_converter.h"
  24
  25 #include <epoxy/glx.h>
  26 #include <jpeglib.h>
  27 #include <unistd.h>
  28
  29 using namespace movit;
  30 using namespace std;
  31 using namespace std::chrono;
  32
  33 namespace {
  34
  35 once_flag video_metrics_inited;
  36 Summary metric_jpeg_encode_time_seconds;
  37 Summary metric_fade_latency_seconds;
  38 Summary metric_interpolation_latency_seconds;
  39 Summary metric_fade_fence_wait_time_seconds;
  40 Summary metric_interpolation_fence_wait_time_seconds;
  41
  42 void wait_for_upload(shared_ptr<Frame> &frame)
  43 {
  44         if (frame->uploaded_interpolation != nullptr) {
  45                 glWaitSync(frame->uploaded_interpolation.get(), /*flags=*/0, GL_TIMEOUT_IGNORED);
  46                 frame->uploaded_interpolation.reset();
  47         }
  48 }
  49
  50 }  // namespace
  51
  52 extern HTTPD *global_httpd;
  53
  54 struct VectorDestinationManager {
  55         jpeg_destination_mgr pub;
  56         string dest;
  57
  58         VectorDestinationManager()
  59         {
  60                 pub.init_destination = init_destination_thunk;
  61                 pub.empty_output_buffer = empty_output_buffer_thunk;
  62                 pub.term_destination = term_destination_thunk;
  63         }
  64
  65         static void init_destination_thunk(j_compress_ptr ptr)
  66         {
  67                 ((VectorDestinationManager *)(ptr->dest))->init_destination();
  68         }
  69
  70         inline void init_destination()
  71         {
  72                 make_room(0);
  73         }
  74
  75         static boolean empty_output_buffer_thunk(j_compress_ptr ptr)
  76         {
  77                 return ((VectorDestinationManager *)(ptr->dest))->empty_output_buffer();
  78         }
  79
  80         inline bool empty_output_buffer()
  81         {
  82                 make_room(dest.size());  // Should ignore pub.free_in_buffer!
  83                 return true;
  84         }
  85
  86         inline void make_room(size_t bytes_used)
  87         {
  88                 dest.resize(bytes_used + 4096);
  89                 dest.resize(dest.capacity());
  90                 pub.next_output_byte = (uint8_t *)dest.data() + bytes_used;
  91                 pub.free_in_buffer = dest.size() - bytes_used;
  92         }
  93
  94         static void term_destination_thunk(j_compress_ptr ptr)
  95         {
  96                 ((VectorDestinationManager *)(ptr->dest))->term_destination();
  97         }
  98
  99         inline void term_destination()
 100         {
 101                 dest.resize(dest.size() - pub.free_in_buffer);
 102         }
 103 };
 104 static_assert(std::is_standard_layout<VectorDestinationManager>::value, "");
 105
 106 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)
 107 {
 108         steady_clock::time_point start = steady_clock::now();
 109         VectorDestinationManager dest;
 110
 111         jpeg_compress_struct cinfo;
 112         jpeg_error_mgr jerr;
 113         cinfo.err = jpeg_std_error(&jerr);
 114         jpeg_create_compress(&cinfo);
 115
 116         cinfo.dest = (jpeg_destination_mgr *)&dest;
 117         cinfo.input_components = 3;
 118         cinfo.in_color_space = JCS_RGB;
 119         jpeg_set_defaults(&cinfo);
 120         constexpr int quality = 90;
 121         jpeg_set_quality(&cinfo, quality, /*force_baseline=*/false);
 122
 123         cinfo.image_width = width;
 124         cinfo.image_height = height;
 125         cinfo.raw_data_in = true;
 126         jpeg_set_colorspace(&cinfo, JCS_YCbCr);
 127         cinfo.comp_info[0].h_samp_factor = 2;
 128         cinfo.comp_info[0].v_samp_factor = 1;
 129         cinfo.comp_info[1].h_samp_factor = 1;
 130         cinfo.comp_info[1].v_samp_factor = 1;
 131         cinfo.comp_info[2].h_samp_factor = 1;
 132         cinfo.comp_info[2].v_samp_factor = 1;
 133         cinfo.CCIR601_sampling = true;  // Seems to be mostly ignored by libjpeg, though.
 134         jpeg_start_compress(&cinfo, true);
 135
 136         // This comment marker is private to FFmpeg. It signals limited Y'CbCr range
 137         // (and nothing else).
 138         jpeg_write_marker(&cinfo, JPEG_COM, (const JOCTET *)"CS=ITU601", strlen("CS=ITU601"));
 139
 140         if (!exif_data.empty()) {
 141                 jpeg_write_marker(&cinfo, JPEG_APP0 + 1, (const JOCTET *)exif_data.data(), exif_data.size());
 142         }
 143
 144         JSAMPROW yptr[8], cbptr[8], crptr[8];
 145         JSAMPARRAY data[3] = { yptr, cbptr, crptr };
 146         for (unsigned y = 0; y < height; y += 8) {
 147                 for (unsigned yy = 0; yy < 8; ++yy) {
 148                         yptr[yy] = const_cast<JSAMPROW>(&y_data[(y + yy) * width]);
 149                         cbptr[yy] = const_cast<JSAMPROW>(&cb_data[(y + yy) * width / 2]);
 150                         crptr[yy] = const_cast<JSAMPROW>(&cr_data[(y + yy) * width / 2]);
 151                 }
 152
 153                 jpeg_write_raw_data(&cinfo, data, /*num_lines=*/8);
 154         }
 155
 156         jpeg_finish_compress(&cinfo);
 157         jpeg_destroy_compress(&cinfo);
 158
 159         steady_clock::time_point stop = steady_clock::now();
 160         metric_jpeg_encode_time_seconds.count_event(duration<double>(stop - start).count());
 161
 162         return move(dest.dest);
 163 }
 164
 165 string encode_jpeg_from_pbo(void *contents, unsigned width, unsigned height, const string exif_data)
 166 {
 167         unsigned chroma_width = width / 2;
 168
 169         const uint8_t *y = (const uint8_t *)contents;
 170         const uint8_t *cb = (const uint8_t *)contents + width * height;
 171         const uint8_t *cr = (const uint8_t *)contents + width * height + chroma_width * height;
 172         return encode_jpeg(y, cb, cr, width, height, move(exif_data));
 173 }
 174
 175 VideoStream::VideoStream(AVFormatContext *file_avctx)
 176         : avctx(file_avctx), output_fast_forward(file_avctx != nullptr)
 177 {
 178         call_once(video_metrics_inited, [] {
 179                 vector<double> quantiles{ 0.01, 0.1, 0.25, 0.5, 0.75, 0.9, 0.99 };
 180                 metric_jpeg_encode_time_seconds.init(quantiles, 60.0);
 181                 global_metrics.add("jpeg_encode_time_seconds", &metric_jpeg_encode_time_seconds);
 182                 metric_fade_fence_wait_time_seconds.init(quantiles, 60.0);
 183                 global_metrics.add("fade_fence_wait_time_seconds", &metric_fade_fence_wait_time_seconds);
 184                 metric_interpolation_fence_wait_time_seconds.init(quantiles, 60.0);
 185                 global_metrics.add("interpolation_fence_wait_time_seconds", &metric_interpolation_fence_wait_time_seconds);
 186                 metric_fade_latency_seconds.init(quantiles, 60.0);
 187                 global_metrics.add("fade_latency_seconds", &metric_fade_latency_seconds);
 188                 metric_interpolation_latency_seconds.init(quantiles, 60.0);
 189                 global_metrics.add("interpolation_latency_seconds", &metric_interpolation_latency_seconds);
 190         });
 191
 192         ycbcr_converter.reset(new YCbCrConverter(YCbCrConverter::OUTPUT_TO_DUAL_YCBCR, /*resource_pool=*/nullptr));
 193         ycbcr_semiplanar_converter.reset(new YCbCrConverter(YCbCrConverter::OUTPUT_TO_SEMIPLANAR, /*resource_pool=*/nullptr));
 194
 195         GLuint input_tex[num_interpolate_slots], gray_tex[num_interpolate_slots];
 196         GLuint fade_y_output_tex[num_interpolate_slots], fade_cbcr_output_tex[num_interpolate_slots];
 197         GLuint cb_tex[num_interpolate_slots], cr_tex[num_interpolate_slots];
 198
 199         glCreateTextures(GL_TEXTURE_2D_ARRAY, num_interpolate_slots, input_tex);
 200         glCreateTextures(GL_TEXTURE_2D_ARRAY, num_interpolate_slots, gray_tex);
 201         glCreateTextures(GL_TEXTURE_2D, num_interpolate_slots, fade_y_output_tex);
 202         glCreateTextures(GL_TEXTURE_2D, num_interpolate_slots, fade_cbcr_output_tex);
 203         glCreateTextures(GL_TEXTURE_2D, num_interpolate_slots, cb_tex);
 204         glCreateTextures(GL_TEXTURE_2D, num_interpolate_slots, cr_tex);
 205         check_error();
 206
 207         size_t width = global_flags.width, height = global_flags.height;
 208         int levels = find_num_levels(width, height);
 209         for (size_t i = 0; i < num_interpolate_slots; ++i) {
 210                 glTextureStorage3D(input_tex[i], levels, GL_RGBA8, width, height, 2);
 211                 check_error();
 212                 glTextureStorage3D(gray_tex[i], levels, GL_R8, width, height, 2);
 213                 check_error();
 214                 glTextureStorage2D(fade_y_output_tex[i], 1, GL_R8, width, height);
 215                 check_error();
 216                 glTextureStorage2D(fade_cbcr_output_tex[i], 1, GL_RG8, width, height);
 217                 check_error();
 218                 glTextureStorage2D(cb_tex[i], 1, GL_R8, width / 2, height);
 219                 check_error();
 220                 glTextureStorage2D(cr_tex[i], 1, GL_R8, width / 2, height);
 221                 check_error();
 222
 223                 unique_ptr<InterpolatedFrameResources> resource(new InterpolatedFrameResources);
 224                 resource->owner = this;
 225                 resource->input_tex = input_tex[i];
 226                 resource->gray_tex = gray_tex[i];
 227                 resource->fade_y_output_tex = fade_y_output_tex[i];
 228                 resource->fade_cbcr_output_tex = fade_cbcr_output_tex[i];
 229                 resource->cb_tex = cb_tex[i];
 230                 resource->cr_tex = cr_tex[i];
 231                 glCreateFramebuffers(2, resource->input_fbos);
 232                 check_error();
 233                 glCreateFramebuffers(1, &resource->fade_fbo);
 234                 check_error();
 235
 236                 glNamedFramebufferTextureLayer(resource->input_fbos[0], GL_COLOR_ATTACHMENT0, input_tex[i], 0, 0);
 237                 check_error();
 238                 glNamedFramebufferTextureLayer(resource->input_fbos[0], GL_COLOR_ATTACHMENT1, gray_tex[i], 0, 0);
 239                 check_error();
 240                 glNamedFramebufferTextureLayer(resource->input_fbos[1], GL_COLOR_ATTACHMENT0, input_tex[i], 0, 1);
 241                 check_error();
 242                 glNamedFramebufferTextureLayer(resource->input_fbos[1], GL_COLOR_ATTACHMENT1, gray_tex[i], 0, 1);
 243                 check_error();
 244                 glNamedFramebufferTexture(resource->fade_fbo, GL_COLOR_ATTACHMENT0, fade_y_output_tex[i], 0);
 245                 check_error();
 246                 glNamedFramebufferTexture(resource->fade_fbo, GL_COLOR_ATTACHMENT1, fade_cbcr_output_tex[i], 0);
 247                 check_error();
 248
 249                 GLuint bufs[] = { GL_COLOR_ATTACHMENT0, GL_COLOR_ATTACHMENT1 };
 250                 glNamedFramebufferDrawBuffers(resource->input_fbos[0], 2, bufs);
 251                 check_error();
 252                 glNamedFramebufferDrawBuffers(resource->input_fbos[1], 2, bufs);
 253                 check_error();
 254                 glNamedFramebufferDrawBuffers(resource->fade_fbo, 2, bufs);
 255                 check_error();
 256
 257                 glCreateBuffers(1, &resource->pbo);
 258                 check_error();
 259                 glNamedBufferStorage(resource->pbo, width * height * 4, nullptr, GL_MAP_READ_BIT | GL_MAP_PERSISTENT_BIT);
 260                 check_error();
 261                 resource->pbo_contents = glMapNamedBufferRange(resource->pbo, 0, width * height * 4, GL_MAP_READ_BIT | GL_MAP_PERSISTENT_BIT);
 262                 interpolate_resources.push_back(move(resource));
 263         }
 264
 265         check_error();
 266
 267         OperatingPoint op;
 268         if (global_flags.interpolation_quality == 0 ||
 269             global_flags.interpolation_quality == 1) {
 270                 op = operating_point1;
 271         } else if (global_flags.interpolation_quality == 2) {
 272                 op = operating_point2;
 273         } else if (global_flags.interpolation_quality == 3) {
 274                 op = operating_point3;
 275         } else if (global_flags.interpolation_quality == 4) {
 276                 op = operating_point4;
 277         } else {
 278                 // Quality 0 will be changed to 1 in flags.cpp.
 279                 assert(false);
 280         }
 281
 282         compute_flow.reset(new DISComputeFlow(width, height, op));
 283         interpolate.reset(new Interpolate(op, /*split_ycbcr_output=*/true));
 284         interpolate_no_split.reset(new Interpolate(op, /*split_ycbcr_output=*/false));
 285         chroma_subsampler.reset(new ChromaSubsampler);
 286         check_error();
 287
 288         // The “last frame” is initially black.
 289         unique_ptr<uint8_t[]> y(new uint8_t[global_flags.width * global_flags.height]);
 290         unique_ptr<uint8_t[]> cb_or_cr(new uint8_t[(global_flags.width / 2) * global_flags.height]);
 291         memset(y.get(), 16, global_flags.width * global_flags.height);
 292         memset(cb_or_cr.get(), 128, (global_flags.width / 2) * global_flags.height);
 293         last_frame = encode_jpeg(y.get(), cb_or_cr.get(), cb_or_cr.get(), global_flags.width, global_flags.height, /*exif_data=*/"");
 294
 295         if (file_avctx != nullptr) {
 296                 with_subtitles = Mux::WITHOUT_SUBTITLES;
 297         } else {
 298                 with_subtitles = Mux::WITH_SUBTITLES;
 299         }
 300 }
 301
 302 VideoStream::~VideoStream()
 303 {
 304         if (last_flow_tex != 0) {
 305                 compute_flow->release_texture(last_flow_tex);
 306         }
 307
 308         for (const unique_ptr<InterpolatedFrameResources> &resource : interpolate_resources) {
 309                 glUnmapNamedBuffer(resource->pbo);
 310                 check_error();
 311                 glDeleteBuffers(1, &resource->pbo);
 312                 check_error();
 313                 glDeleteFramebuffers(2, resource->input_fbos);
 314                 check_error();
 315                 glDeleteFramebuffers(1, &resource->fade_fbo);
 316                 check_error();
 317                 glDeleteTextures(1, &resource->input_tex);
 318                 check_error();
 319                 glDeleteTextures(1, &resource->gray_tex);
 320                 check_error();
 321                 glDeleteTextures(1, &resource->fade_y_output_tex);
 322                 check_error();
 323                 glDeleteTextures(1, &resource->fade_cbcr_output_tex);
 324                 check_error();
 325                 glDeleteTextures(1, &resource->cb_tex);
 326                 check_error();
 327                 glDeleteTextures(1, &resource->cr_tex);
 328                 check_error();
 329         }
 330         assert(interpolate_resources.size() == num_interpolate_slots);
 331 }
 332
 333 void VideoStream::start()
 334 {
 335         if (avctx == nullptr) {
 336                 avctx = avformat_alloc_context();
 337
 338                 // We use Matroska, because it's pretty much the only mux where FFmpeg
 339                 // allows writing chroma location to override JFIF's default center placement.
 340                 // (Note that at the time of writing, however, FFmpeg does not correctly
 341                 // _read_ this information!)
 342                 avctx->oformat = av_guess_format("matroska", nullptr, nullptr);
 343
 344                 uint8_t *buf = (uint8_t *)av_malloc(MUX_BUFFER_SIZE);
 345                 avctx->pb = avio_alloc_context(buf, MUX_BUFFER_SIZE, 1, this, nullptr, nullptr, nullptr);
 346                 avctx->pb->write_data_type = &VideoStream::write_packet2_thunk;
 347                 avctx->pb->ignore_boundary_point = 1;
 348
 349                 avctx->flags = AVFMT_FLAG_CUSTOM_IO;
 350         }
 351
 352         AVCodecParameters *audio_codecpar = avcodec_parameters_alloc();
 353
 354         audio_codecpar->codec_type = AVMEDIA_TYPE_AUDIO;
 355         audio_codecpar->codec_id = AV_CODEC_ID_PCM_S32LE;
 356         audio_codecpar->ch_layout.order = AV_CHANNEL_ORDER_NATIVE;
 357         audio_codecpar->ch_layout.nb_channels = 2;
 358         audio_codecpar->ch_layout.u.mask = AV_CH_LAYOUT_STEREO;
 359         audio_codecpar->sample_rate = OUTPUT_FREQUENCY;
 360
 361         size_t width = global_flags.width, height = global_flags.height;  // Doesn't matter for MJPEG.
 362         mux.reset(new Mux(avctx, width, height, Mux::CODEC_MJPEG, /*video_extradata=*/"", audio_codecpar,
 363                           AVCOL_SPC_BT709, COARSE_TIMEBASE, /*write_callback=*/nullptr, Mux::WRITE_FOREGROUND, {}, with_subtitles));
 364
 365         avcodec_parameters_free(&audio_codecpar);
 366         encode_thread = thread(&VideoStream::encode_thread_func, this);
 367 }
 368
 369 void VideoStream::stop()
 370 {
 371         should_quit = true;
 372         queue_changed.notify_all();
 373         clear_queue();
 374         encode_thread.join();
 375 }
 376
 377 void VideoStream::clear_queue()
 378 {
 379         deque<QueuedFrame> q;
 380
 381         {
 382                 lock_guard<mutex> lock(queue_lock);
 383                 q = move(frame_queue);
 384         }
 385
 386         // These are not RAII-ed, unfortunately, so we'll need to clean them ourselves.
 387         // Note that release_texture() is thread-safe.
 388         for (const QueuedFrame &qf : q) {
 389                 if (qf.type == QueuedFrame::INTERPOLATED ||
 390                     qf.type == QueuedFrame::FADED_INTERPOLATED) {
 391                         if (qf.flow_tex != 0) {
 392                                 compute_flow->release_texture(qf.flow_tex);
 393                         }
 394                 }
 395                 if (qf.type == QueuedFrame::INTERPOLATED) {
 396                         interpolate->release_texture(qf.output_tex);
 397                         interpolate->release_texture(qf.cbcr_tex);
 398                 }
 399         }
 400
 401         // Destroy q outside the mutex, as that would be a double-lock.
 402 }
 403
 404 void VideoStream::schedule_original_frame(steady_clock::time_point local_pts,
 405                                           int64_t output_pts, function<void()> &&display_func,
 406                                           QueueSpotHolder &&queue_spot_holder,
 407                                           FrameOnDisk frame, const string &subtitle, bool include_audio)
 408 {
 409         fprintf(stderr, "output_pts=%" PRId64 "  original      input_pts=%" PRId64 "\n", output_pts, frame.pts);
 410
 411         QueuedFrame qf;
 412         qf.local_pts = local_pts;
 413         qf.type = QueuedFrame::ORIGINAL;
 414         qf.output_pts = output_pts;
 415         qf.display_func = move(display_func);
 416         qf.queue_spot_holder = move(queue_spot_holder);
 417         qf.subtitle = subtitle;
 418         FrameReader::Frame read_frame = frame_reader.read_frame(frame, /*read_video=*/true, include_audio);
 419         qf.encoded_jpeg.reset(new string(move(read_frame.video)));
 420         qf.audio = move(read_frame.audio);
 421
 422         lock_guard<mutex> lock(queue_lock);
 423         frame_queue.push_back(move(qf));
 424         queue_changed.notify_all();
 425 }
 426
 427 void VideoStream::schedule_faded_frame(steady_clock::time_point local_pts, int64_t output_pts,
 428                                        function<void()> &&display_func,
 429                                        QueueSpotHolder &&queue_spot_holder,
 430                                        FrameOnDisk frame1_spec, FrameOnDisk frame2_spec,
 431                                        float fade_alpha, const string &subtitle)
 432 {
 433         fprintf(stderr, "output_pts=%" PRId64 "  faded         input_pts=%" PRId64 ",%" PRId64 "  fade_alpha=%.2f\n", output_pts, frame1_spec.pts, frame2_spec.pts, fade_alpha);
 434
 435         // Get the temporary OpenGL resources we need for doing the fade.
 436         // (We share these with interpolated frames, which is slightly
 437         // overkill, but there's no need to waste resources on keeping
 438         // separate pools around.)
 439         BorrowedInterpolatedFrameResources resources;
 440         {
 441                 lock_guard<mutex> lock(queue_lock);
 442                 if (interpolate_resources.empty()) {
 443                         fprintf(stderr, "WARNING: Too many interpolated frames already in transit; dropping one.\n");
 444                         return;
 445                 }
 446                 resources = BorrowedInterpolatedFrameResources(interpolate_resources.front().release());
 447                 interpolate_resources.pop_front();
 448         }
 449
 450         bool did_decode;
 451
 452         shared_ptr<Frame> frame1 = decode_jpeg_with_cache(frame1_spec, DECODE_IF_NOT_IN_CACHE, &frame_reader, &did_decode);
 453         shared_ptr<Frame> frame2 = decode_jpeg_with_cache(frame2_spec, DECODE_IF_NOT_IN_CACHE, &frame_reader, &did_decode);
 454         wait_for_upload(frame1);
 455         wait_for_upload(frame2);
 456
 457         ycbcr_semiplanar_converter->prepare_chain_for_fade(frame1, frame2, fade_alpha)->render_to_fbo(resources->fade_fbo, global_flags.width, global_flags.height);
 458
 459         QueuedFrame qf;
 460         qf.local_pts = local_pts;
 461         qf.type = QueuedFrame::FADED;
 462         qf.output_pts = output_pts;
 463         qf.frame1 = frame1_spec;
 464         qf.display_func = move(display_func);
 465         qf.queue_spot_holder = move(queue_spot_holder);
 466         qf.subtitle = subtitle;
 467
 468         qf.secondary_frame = frame2_spec;
 469
 470         // Subsample and split Cb/Cr.
 471         chroma_subsampler->subsample_chroma(resources->fade_cbcr_output_tex, global_flags.width, global_flags.height, resources->cb_tex, resources->cr_tex);
 472
 473         // Read it down (asynchronously) to the CPU.
 474         glPixelStorei(GL_PACK_ROW_LENGTH, 0);
 475         glBindBuffer(GL_PIXEL_PACK_BUFFER, resources->pbo);
 476         check_error();
 477         glGetTextureImage(resources->fade_y_output_tex, 0, GL_RED, GL_UNSIGNED_BYTE, global_flags.width * global_flags.height * 4, BUFFER_OFFSET(0));
 478         check_error();
 479         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));
 480         check_error();
 481         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));
 482         check_error();
 483         glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
 484
 485         // Set a fence we can wait for to make sure the CPU sees the read.
 486         glMemoryBarrier(GL_CLIENT_MAPPED_BUFFER_BARRIER_BIT);
 487         check_error();
 488         qf.fence_created = steady_clock::now();
 489         qf.fence = RefCountedGLsync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
 490         check_error();
 491         qf.resources = move(resources);
 492         qf.local_pts = local_pts;
 493
 494         lock_guard<mutex> lock(queue_lock);
 495         frame_queue.push_back(move(qf));
 496         queue_changed.notify_all();
 497 }
 498
 499 void VideoStream::schedule_interpolated_frame(steady_clock::time_point local_pts,
 500                                               int64_t output_pts, function<void(shared_ptr<Frame>)> &&display_func,
 501                                               QueueSpotHolder &&queue_spot_holder,
 502                                               FrameOnDisk frame1, FrameOnDisk frame2,
 503                                               float alpha, FrameOnDisk secondary_frame, float fade_alpha, const string &subtitle,
 504                                               bool play_audio)
 505 {
 506         if (secondary_frame.pts != -1) {
 507                 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);
 508         } else {
 509                 fprintf(stderr, "output_pts=%" PRId64 "  interpolated  input_pts1=%" PRId64 " input_pts2=%" PRId64 " alpha=%.3f\n", output_pts, frame1.pts, frame2.pts, alpha);
 510         }
 511
 512         // Get the temporary OpenGL resources we need for doing the interpolation.
 513         BorrowedInterpolatedFrameResources resources;
 514         {
 515                 lock_guard<mutex> lock(queue_lock);
 516                 if (interpolate_resources.empty()) {
 517                         fprintf(stderr, "WARNING: Too many interpolated frames already in transit; dropping one.\n");
 518                         return;
 519                 }
 520                 resources = BorrowedInterpolatedFrameResources(interpolate_resources.front().release());
 521                 interpolate_resources.pop_front();
 522         }
 523
 524         QueuedFrame qf;
 525         qf.type = (secondary_frame.pts == -1) ? QueuedFrame::INTERPOLATED : QueuedFrame::FADED_INTERPOLATED;
 526         qf.output_pts = output_pts;
 527         qf.display_decoded_func = move(display_func);
 528         qf.queue_spot_holder = move(queue_spot_holder);
 529         qf.local_pts = local_pts;
 530         qf.subtitle = subtitle;
 531
 532         if (play_audio) {
 533                 qf.audio = frame_reader.read_frame(frame1, /*read_video=*/false, /*read_audio=*/true).audio;
 534         }
 535
 536         check_error();
 537
 538         // Convert frame0 and frame1 to OpenGL textures.
 539         for (size_t frame_no = 0; frame_no < 2; ++frame_no) {
 540                 FrameOnDisk frame_spec = frame_no == 1 ? frame2 : frame1;
 541                 bool did_decode;
 542                 shared_ptr<Frame> frame = decode_jpeg_with_cache(frame_spec, DECODE_IF_NOT_IN_CACHE, &frame_reader, &did_decode);
 543                 wait_for_upload(frame);
 544                 ycbcr_converter->prepare_chain_for_conversion(frame)->render_to_fbo(resources->input_fbos[frame_no], global_flags.width, global_flags.height);
 545                 if (frame_no == 1) {
 546                         qf.exif_data = frame->exif_data;  // Use the white point from the last frame.
 547                 }
 548         }
 549
 550         glGenerateTextureMipmap(resources->input_tex);
 551         check_error();
 552         glGenerateTextureMipmap(resources->gray_tex);
 553         check_error();
 554
 555         GLuint flow_tex;
 556         if (last_flow_tex != 0 && frame1 == last_frame1 && frame2 == last_frame2) {
 557                 // Reuse the flow from previous computation. This frequently happens
 558                 // if we slow down by more than 2x, so that there are multiple interpolated
 559                 // frames between each original.
 560                 flow_tex = last_flow_tex;
 561                 qf.flow_tex = 0;
 562         } else {
 563                 // Cache miss, so release last_flow_tex.
 564                 qf.flow_tex = last_flow_tex;
 565
 566                 // Compute the flow.
 567                 flow_tex = compute_flow->exec(resources->gray_tex, DISComputeFlow::FORWARD_AND_BACKWARD, DISComputeFlow::DO_NOT_RESIZE_FLOW);
 568                 check_error();
 569
 570                 // Store the flow texture for possible reuse next frame.
 571                 last_flow_tex = flow_tex;
 572                 last_frame1 = frame1;
 573                 last_frame2 = frame2;
 574         }
 575
 576         if (secondary_frame.pts != -1) {
 577                 // Fade. First kick off the interpolation.
 578                 tie(qf.output_tex, ignore) = interpolate_no_split->exec(resources->input_tex, resources->gray_tex, flow_tex, global_flags.width, global_flags.height, alpha);
 579                 check_error();
 580
 581                 // Now decode the image we are fading against.
 582                 bool did_decode;
 583                 shared_ptr<Frame> frame2 = decode_jpeg_with_cache(secondary_frame, DECODE_IF_NOT_IN_CACHE, &frame_reader, &did_decode);
 584                 wait_for_upload(frame2);
 585
 586                 // Then fade against it, putting it into the fade Y' and CbCr textures.
 587                 RGBTriplet neutral_color = get_neutral_color(qf.exif_data);
 588                 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);
 589
 590                 // Subsample and split Cb/Cr.
 591                 chroma_subsampler->subsample_chroma(resources->fade_cbcr_output_tex, global_flags.width, global_flags.height, resources->cb_tex, resources->cr_tex);
 592
 593                 interpolate_no_split->release_texture(qf.output_tex);
 594
 595                 // We already applied the white balance, so don't have the client redo it.
 596                 qf.exif_data.clear();
 597         } else {
 598                 tie(qf.output_tex, qf.cbcr_tex) = interpolate->exec(resources->input_tex, resources->gray_tex, flow_tex, global_flags.width, global_flags.height, alpha);
 599                 check_error();
 600
 601                 // Subsample and split Cb/Cr.
 602                 chroma_subsampler->subsample_chroma(qf.cbcr_tex, global_flags.width, global_flags.height, resources->cb_tex, resources->cr_tex);
 603         }
 604
 605         // We could have released qf.flow_tex here, but to make sure we don't cause a stall
 606         // when trying to reuse it for the next frame, we can just as well hold on to it
 607         // and release it only when the readback is done.
 608         //
 609         // TODO: This is maybe less relevant now that qf.flow_tex contains the texture we used
 610         // _last_ frame, not this one.
 611
 612         // Read it down (asynchronously) to the CPU.
 613         glPixelStorei(GL_PACK_ROW_LENGTH, 0);
 614         glBindBuffer(GL_PIXEL_PACK_BUFFER, resources->pbo);
 615         check_error();
 616         if (secondary_frame.pts != -1) {
 617                 glGetTextureImage(resources->fade_y_output_tex, 0, GL_RED, GL_UNSIGNED_BYTE, global_flags.width * global_flags.height * 4, BUFFER_OFFSET(0));
 618         } else {
 619                 glGetTextureImage(qf.output_tex, 0, GL_RED, GL_UNSIGNED_BYTE, global_flags.width * global_flags.height * 4, BUFFER_OFFSET(0));
 620         }
 621         check_error();
 622         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));
 623         check_error();
 624         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));
 625         check_error();
 626         glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
 627
 628         // Set a fence we can wait for to make sure the CPU sees the read.
 629         glMemoryBarrier(GL_CLIENT_MAPPED_BUFFER_BARRIER_BIT);
 630         check_error();
 631         qf.fence_created = steady_clock::now();
 632         qf.fence = RefCountedGLsync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
 633         check_error();
 634         qf.resources = move(resources);
 635
 636         lock_guard<mutex> lock(queue_lock);
 637         frame_queue.push_back(move(qf));
 638         queue_changed.notify_all();
 639 }
 640
 641 void VideoStream::schedule_refresh_frame(steady_clock::time_point local_pts,
 642                                          int64_t output_pts, function<void()> &&display_func,
 643                                          QueueSpotHolder &&queue_spot_holder, const string &subtitle)
 644 {
 645         QueuedFrame qf;
 646         qf.type = QueuedFrame::REFRESH;
 647         qf.output_pts = output_pts;
 648         qf.display_func = move(display_func);
 649         qf.queue_spot_holder = move(queue_spot_holder);
 650         qf.subtitle = subtitle;
 651
 652         lock_guard<mutex> lock(queue_lock);
 653         frame_queue.push_back(move(qf));
 654         queue_changed.notify_all();
 655 }
 656
 657 void VideoStream::schedule_silence(steady_clock::time_point local_pts, int64_t output_pts,
 658                                    int64_t length_pts, QueueSpotHolder &&queue_spot_holder)
 659 {
 660         QueuedFrame qf;
 661         qf.type = QueuedFrame::SILENCE;
 662         qf.output_pts = output_pts;
 663         qf.queue_spot_holder = move(queue_spot_holder);
 664         qf.silence_length_pts = length_pts;
 665
 666         lock_guard<mutex> lock(queue_lock);
 667         frame_queue.push_back(move(qf));
 668         queue_changed.notify_all();
 669 }
 670
 671 namespace {
 672
 673 RefCountedTexture clone_r8_texture(GLuint src_tex, unsigned width, unsigned height)
 674 {
 675         GLuint tex;
 676         glCreateTextures(GL_TEXTURE_2D, 1, &tex);
 677         check_error();
 678         glTextureStorage2D(tex, 1, GL_R8, width, height);
 679         check_error();
 680         glCopyImageSubData(src_tex, GL_TEXTURE_2D, 0, 0, 0, 0,
 681                            tex, GL_TEXTURE_2D, 0, 0, 0, 0,
 682                            width, height, 1);
 683         check_error();
 684         glTextureParameteri(tex, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
 685         check_error();
 686         glTextureParameteri(tex, GL_TEXTURE_MAG_FILTER, GL_LINEAR);
 687         check_error();
 688         glTextureParameteri(tex, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
 689         check_error();
 690         glTextureParameteri(tex, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
 691         check_error();
 692
 693         return RefCountedTexture(new GLuint(tex), TextureDeleter());
 694 }
 695
 696 }  // namespace
 697
 698 void VideoStream::encode_thread_func()
 699 {
 700         pthread_setname_np(pthread_self(), "VideoStream");
 701         QSurface *surface = create_surface();
 702         QOpenGLContext *context = create_context(surface);
 703         bool ok = make_current(context, surface);
 704         if (!ok) {
 705                 fprintf(stderr, "Video stream couldn't get an OpenGL context\n");
 706                 abort();
 707         }
 708
 709         init_pbo_pool();
 710
 711         while (!should_quit) {
 712                 QueuedFrame qf;
 713                 {
 714                         unique_lock<mutex> lock(queue_lock);
 715
 716                         // Wait until we have a frame to play.
 717                         queue_changed.wait(lock, [this] {
 718                                 return !frame_queue.empty() || should_quit;
 719                         });
 720                         if (should_quit) {
 721                                 break;
 722                         }
 723                         steady_clock::time_point frame_start = frame_queue.front().local_pts;
 724
 725                         // Now sleep until the frame is supposed to start (the usual case),
 726                         // _or_ clear_queue() happened.
 727                         bool aborted;
 728                         if (output_fast_forward) {
 729                                 aborted = frame_queue.empty() || frame_queue.front().local_pts != frame_start;
 730                         } else {
 731                                 aborted = queue_changed.wait_until(lock, frame_start, [this, frame_start] {
 732                                         return frame_queue.empty() || frame_queue.front().local_pts != frame_start;
 733                                 });
 734                         }
 735                         if (aborted) {
 736                                 // clear_queue() happened, so don't play this frame after all.
 737                                 continue;
 738                         }
 739                         qf = move(frame_queue.front());
 740                         frame_queue.pop_front();
 741                 }
 742
 743                 // Hack: We mux the subtitle packet one time unit before the actual frame,
 744                 // so that Nageru is sure to get it first.
 745                 if (!qf.subtitle.empty() && with_subtitles == Mux::WITH_SUBTITLES) {
 746                         AVPacket pkt;
 747                         av_init_packet(&pkt);
 748                         pkt.stream_index = mux->get_subtitle_stream_idx();
 749                         assert(pkt.stream_index != -1);
 750                         pkt.data = (uint8_t *)qf.subtitle.data();
 751                         pkt.size = qf.subtitle.size();
 752                         pkt.flags = 0;
 753                         pkt.duration = lrint(TIMEBASE / global_flags.output_framerate);  // Doesn't really matter for Nageru.
 754                         mux->add_packet(pkt, qf.output_pts - 1, qf.output_pts - 1);
 755                 }
 756
 757                 if (qf.type == QueuedFrame::ORIGINAL) {
 758                         // Send the JPEG frame on, unchanged.
 759                         string jpeg = move(*qf.encoded_jpeg);
 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::FADED) {
 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_fade_fence_wait_time_seconds.count_event(duration<double>(stop - start).count());
 775                         metric_fade_latency_seconds.count_event(duration<double>(stop - qf.fence_created).count());
 776
 777                         // Now JPEG encode it, and send it on to the stream.
 778                         string jpeg = encode_jpeg_from_pbo(qf.resources->pbo_contents, global_flags.width, global_flags.height, /*exif_data=*/"");
 779
 780                         AVPacket pkt;
 781                         av_init_packet(&pkt);
 782                         pkt.stream_index = 0;
 783                         pkt.data = (uint8_t *)jpeg.data();
 784                         pkt.size = jpeg.size();
 785                         pkt.flags = AV_PKT_FLAG_KEY;
 786                         mux->add_packet(pkt, qf.output_pts, qf.output_pts);
 787                         last_frame = move(jpeg);
 788
 789                         add_audio_or_silence(qf);
 790                 } else if (qf.type == QueuedFrame::INTERPOLATED || qf.type == QueuedFrame::FADED_INTERPOLATED) {
 791                         steady_clock::time_point start = steady_clock::now();
 792                         glClientWaitSync(qf.fence.get(), /*flags=*/0, GL_TIMEOUT_IGNORED);
 793                         steady_clock::time_point stop = steady_clock::now();
 794                         metric_interpolation_fence_wait_time_seconds.count_event(duration<double>(stop - start).count());
 795                         metric_interpolation_latency_seconds.count_event(duration<double>(stop - qf.fence_created).count());
 796
 797                         // Send it on to display.
 798                         if (qf.display_decoded_func != nullptr) {
 799                                 shared_ptr<Frame> frame(new Frame);
 800                                 if (qf.type == QueuedFrame::FADED_INTERPOLATED) {
 801                                         frame->y = clone_r8_texture(qf.resources->fade_y_output_tex, global_flags.width, global_flags.height);
 802                                 } else {
 803                                         frame->y = clone_r8_texture(qf.output_tex, global_flags.width, global_flags.height);
 804                                 }
 805                                 frame->cb = clone_r8_texture(qf.resources->cb_tex, global_flags.width / 2, global_flags.height);
 806                                 frame->cr = clone_r8_texture(qf.resources->cr_tex, global_flags.width / 2, global_flags.height);
 807                                 frame->width = global_flags.width;
 808                                 frame->height = global_flags.height;
 809                                 frame->chroma_subsampling_x = 2;
 810                                 frame->chroma_subsampling_y = 1;
 811                                 frame->uploaded_ui_thread = RefCountedGLsync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
 812                                 qf.display_decoded_func(move(frame));
 813                         }
 814
 815                         // Now JPEG encode it, and send it on to the stream.
 816                         string jpeg = encode_jpeg_from_pbo(qf.resources->pbo_contents, global_flags.width, global_flags.height, move(qf.exif_data));
 817                         if (qf.flow_tex != 0) {
 818                                 compute_flow->release_texture(qf.flow_tex);
 819                         }
 820                         if (qf.type != QueuedFrame::FADED_INTERPOLATED) {
 821                                 interpolate->release_texture(qf.output_tex);
 822                                 interpolate->release_texture(qf.cbcr_tex);
 823                         }
 824
 825                         AVPacket pkt;
 826                         av_init_packet(&pkt);
 827                         pkt.stream_index = 0;
 828                         pkt.data = (uint8_t *)jpeg.data();
 829                         pkt.size = jpeg.size();
 830                         pkt.flags = AV_PKT_FLAG_KEY;
 831                         mux->add_packet(pkt, qf.output_pts, qf.output_pts);
 832                         last_frame = move(jpeg);
 833
 834                         add_audio_or_silence(qf);
 835                 } else if (qf.type == QueuedFrame::REFRESH) {
 836                         AVPacket pkt;
 837                         av_init_packet(&pkt);
 838                         pkt.stream_index = 0;
 839                         pkt.data = (uint8_t *)last_frame.data();
 840                         pkt.size = last_frame.size();
 841                         pkt.flags = AV_PKT_FLAG_KEY;
 842                         mux->add_packet(pkt, qf.output_pts, qf.output_pts);
 843
 844                         add_audio_or_silence(qf);  // Definitely silence.
 845                 } else if (qf.type == QueuedFrame::SILENCE) {
 846                         add_silence(qf.output_pts, qf.silence_length_pts);
 847                 } else {
 848                         assert(false);
 849                 }
 850                 if (qf.display_func != nullptr) {
 851                         qf.display_func();
 852                 }
 853         }
 854 }
 855
 856 int VideoStream::write_packet2_thunk(void *opaque, uint8_t *buf, int buf_size, AVIODataMarkerType type, int64_t time)
 857 {
 858         VideoStream *video_stream = (VideoStream *)opaque;
 859         return video_stream->write_packet2(buf, buf_size, type, time);
 860 }
 861
 862 int VideoStream::write_packet2(uint8_t *buf, int buf_size, AVIODataMarkerType type, int64_t time)
 863 {
 864         if (type == AVIO_DATA_MARKER_SYNC_POINT || type == AVIO_DATA_MARKER_BOUNDARY_POINT) {
 865                 seen_sync_markers = true;
 866         } else if (type == AVIO_DATA_MARKER_UNKNOWN && !seen_sync_markers) {
 867                 // We don't know if this is a keyframe or not (the muxer could
 868                 // avoid marking it), so we just have to make the best of it.
 869                 type = AVIO_DATA_MARKER_SYNC_POINT;
 870         }
 871
 872         HTTPD::StreamID stream_id{ HTTPD::MAIN_STREAM, 0 };
 873         if (type == AVIO_DATA_MARKER_HEADER) {
 874                 stream_mux_header.append((char *)buf, buf_size);
 875                 global_httpd->set_header(stream_id, stream_mux_header);
 876         } else {
 877                 global_httpd->add_data(stream_id, (char *)buf, buf_size, type == AVIO_DATA_MARKER_SYNC_POINT, time, AVRational{ AV_TIME_BASE, 1 });
 878         }
 879         return buf_size;
 880 }
 881
 882 void VideoStream::add_silence(int64_t pts, int64_t length_pts)
 883 {
 884         // At 59.94, this will never quite add up (even discounting refresh frames,
 885         // which have unpredictable length), but hopefully, the player in the other
 886         // end should be able to stretch silence easily enough.
 887         long num_samples = lrint(length_pts * double(OUTPUT_FREQUENCY) / double(TIMEBASE)) * 2;
 888         uint8_t *zero = (uint8_t *)calloc(num_samples, sizeof(int32_t));
 889
 890         AVPacket pkt;
 891         av_init_packet(&pkt);
 892         pkt.stream_index = 1;
 893         pkt.data = zero;
 894         pkt.size = num_samples * sizeof(int32_t);
 895         pkt.flags = AV_PKT_FLAG_KEY;
 896         mux->add_packet(pkt, pts, pts);
 897
 898         free(zero);
 899 }
 900
 901 void VideoStream::add_audio_or_silence(const QueuedFrame &qf)
 902 {
 903         if (qf.audio.empty()) {
 904                 int64_t frame_length = lrint(double(TIMEBASE) / global_flags.output_framerate);
 905                 add_silence(qf.output_pts, frame_length);
 906         } else {
 907                 AVPacket pkt;
 908                 av_init_packet(&pkt);
 909                 pkt.stream_index = 1;
 910                 pkt.data = (uint8_t *)qf.audio.data();
 911                 pkt.size = qf.audio.size();
 912                 pkt.flags = AV_PKT_FLAG_KEY;
 913                 mux->add_packet(pkt, qf.output_pts, qf.output_pts);
 914         }
 915 }