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 "pbo_pool.h"
  15 #include "player.h"
  16 #include "shared/context.h"
  17 #include "shared/httpd.h"
  18 #include "shared/metrics.h"
  19 #include "shared/shared_defs.h"
  20 #include "shared/mux.h"
  21 #include "util.h"
  22 #include "ycbcr_converter.h"
  23
  24 #include <epoxy/glx.h>
  25 #include <jpeglib.h>
  26 #include <unistd.h>
  27
  28 using namespace movit;
  29 using namespace std;
  30 using namespace std::chrono;
  31
  32 namespace {
  33
  34 once_flag video_metrics_inited;
  35 Summary metric_jpeg_encode_time_seconds;
  36 Summary metric_fade_latency_seconds;
  37 Summary metric_interpolation_latency_seconds;
  38 Summary metric_fade_fence_wait_time_seconds;
  39 Summary metric_interpolation_fence_wait_time_seconds;
  40
  41 void wait_for_upload(shared_ptr<Frame> &frame)
  42 {
  43         if (frame->uploaded_interpolation != nullptr) {
  44                 glWaitSync(frame->uploaded_interpolation.get(), /*flags=*/0, GL_TIMEOUT_IGNORED);
  45                 frame->uploaded_interpolation.reset();
  46         }
  47 }
  48
  49 }  // namespace
  50
  51 extern HTTPD *global_httpd;
  52
  53 struct VectorDestinationManager {
  54         jpeg_destination_mgr pub;
  55         string dest;
  56
  57         VectorDestinationManager()
  58         {
  59                 pub.init_destination = init_destination_thunk;
  60                 pub.empty_output_buffer = empty_output_buffer_thunk;
  61                 pub.term_destination = term_destination_thunk;
  62         }
  63
  64         static void init_destination_thunk(j_compress_ptr ptr)
  65         {
  66                 ((VectorDestinationManager *)(ptr->dest))->init_destination();
  67         }
  68
  69         inline void init_destination()
  70         {
  71                 make_room(0);
  72         }
  73
  74         static boolean empty_output_buffer_thunk(j_compress_ptr ptr)
  75         {
  76                 return ((VectorDestinationManager *)(ptr->dest))->empty_output_buffer();
  77         }
  78
  79         inline bool empty_output_buffer()
  80         {
  81                 make_room(dest.size());  // Should ignore pub.free_in_buffer!
  82                 return true;
  83         }
  84
  85         inline void make_room(size_t bytes_used)
  86         {
  87                 dest.resize(bytes_used + 4096);
  88                 dest.resize(dest.capacity());
  89                 pub.next_output_byte = (uint8_t *)dest.data() + bytes_used;
  90                 pub.free_in_buffer = dest.size() - bytes_used;
  91         }
  92
  93         static void term_destination_thunk(j_compress_ptr ptr)
  94         {
  95                 ((VectorDestinationManager *)(ptr->dest))->term_destination();
  96         }
  97
  98         inline void term_destination()
  99         {
 100                 dest.resize(dest.size() - pub.free_in_buffer);
 101         }
 102 };
 103 static_assert(std::is_standard_layout<VectorDestinationManager>::value, "");
 104
 105 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)
 106 {
 107         steady_clock::time_point start = steady_clock::now();
 108         VectorDestinationManager dest;
 109
 110         jpeg_compress_struct cinfo;
 111         jpeg_error_mgr jerr;
 112         cinfo.err = jpeg_std_error(&jerr);
 113         jpeg_create_compress(&cinfo);
 114
 115         cinfo.dest = (jpeg_destination_mgr *)&dest;
 116         cinfo.input_components = 3;
 117         cinfo.in_color_space = JCS_RGB;
 118         jpeg_set_defaults(&cinfo);
 119         constexpr int quality = 90;
 120         jpeg_set_quality(&cinfo, quality, /*force_baseline=*/false);
 121
 122         cinfo.image_width = width;
 123         cinfo.image_height = height;
 124         cinfo.raw_data_in = true;
 125         jpeg_set_colorspace(&cinfo, JCS_YCbCr);
 126         cinfo.comp_info[0].h_samp_factor = 2;
 127         cinfo.comp_info[0].v_samp_factor = 1;
 128         cinfo.comp_info[1].h_samp_factor = 1;
 129         cinfo.comp_info[1].v_samp_factor = 1;
 130         cinfo.comp_info[2].h_samp_factor = 1;
 131         cinfo.comp_info[2].v_samp_factor = 1;
 132         cinfo.CCIR601_sampling = true;  // Seems to be mostly ignored by libjpeg, though.
 133         jpeg_start_compress(&cinfo, true);
 134
 135         // This comment marker is private to FFmpeg. It signals limited Y'CbCr range
 136         // (and nothing else).
 137         jpeg_write_marker(&cinfo, JPEG_COM, (const JOCTET *)"CS=ITU601", strlen("CS=ITU601"));
 138
 139         if (!exif_data.empty()) {
 140                 jpeg_write_marker(&cinfo, JPEG_APP0 + 1, (const JOCTET *)exif_data.data(), exif_data.size());
 141         }
 142
 143         JSAMPROW yptr[8], cbptr[8], crptr[8];
 144         JSAMPARRAY data[3] = { yptr, cbptr, crptr };
 145         for (unsigned y = 0; y < height; y += 8) {
 146                 for (unsigned yy = 0; yy < 8; ++yy) {
 147                         yptr[yy] = const_cast<JSAMPROW>(&y_data[(y + yy) * width]);
 148                         cbptr[yy] = const_cast<JSAMPROW>(&cb_data[(y + yy) * width / 2]);
 149                         crptr[yy] = const_cast<JSAMPROW>(&cr_data[(y + yy) * width / 2]);
 150                 }
 151
 152                 jpeg_write_raw_data(&cinfo, data, /*num_lines=*/8);
 153         }
 154
 155         jpeg_finish_compress(&cinfo);
 156         jpeg_destroy_compress(&cinfo);
 157
 158         steady_clock::time_point stop = steady_clock::now();
 159         metric_jpeg_encode_time_seconds.count_event(duration<double>(stop - start).count());
 160
 161         return move(dest.dest);
 162 }
 163
 164 string encode_jpeg_from_pbo(void *contents, unsigned width, unsigned height, const string exif_data)
 165 {
 166         unsigned chroma_width = width / 2;
 167
 168         const uint8_t *y = (const uint8_t *)contents;
 169         const uint8_t *cb = (const uint8_t *)contents + width * height;
 170         const uint8_t *cr = (const uint8_t *)contents + width * height + chroma_width * height;
 171         return encode_jpeg(y, cb, cr, width, height, move(exif_data));
 172 }
 173
 174 VideoStream::VideoStream(AVFormatContext *file_avctx)
 175         : avctx(file_avctx), output_fast_forward(file_avctx != nullptr)
 176 {
 177         call_once(video_metrics_inited, [] {
 178                 vector<double> quantiles{ 0.01, 0.1, 0.25, 0.5, 0.75, 0.9, 0.99 };
 179                 metric_jpeg_encode_time_seconds.init(quantiles, 60.0);
 180                 global_metrics.add("jpeg_encode_time_seconds", &metric_jpeg_encode_time_seconds);
 181                 metric_fade_fence_wait_time_seconds.init(quantiles, 60.0);
 182                 global_metrics.add("fade_fence_wait_time_seconds", &metric_fade_fence_wait_time_seconds);
 183                 metric_interpolation_fence_wait_time_seconds.init(quantiles, 60.0);
 184                 global_metrics.add("interpolation_fence_wait_time_seconds", &metric_interpolation_fence_wait_time_seconds);
 185                 metric_fade_latency_seconds.init(quantiles, 60.0);
 186                 global_metrics.add("fade_latency_seconds", &metric_fade_latency_seconds);
 187                 metric_interpolation_latency_seconds.init(quantiles, 60.0);
 188                 global_metrics.add("interpolation_latency_seconds", &metric_interpolation_latency_seconds);
 189         });
 190
 191         ycbcr_converter.reset(new YCbCrConverter(YCbCrConverter::OUTPUT_TO_DUAL_YCBCR, /*resource_pool=*/nullptr));
 192         ycbcr_semiplanar_converter.reset(new YCbCrConverter(YCbCrConverter::OUTPUT_TO_SEMIPLANAR, /*resource_pool=*/nullptr));
 193
 194         GLuint input_tex[num_interpolate_slots], gray_tex[num_interpolate_slots];
 195         GLuint fade_y_output_tex[num_interpolate_slots], fade_cbcr_output_tex[num_interpolate_slots];
 196         GLuint cb_tex[num_interpolate_slots], cr_tex[num_interpolate_slots];
 197
 198         glCreateTextures(GL_TEXTURE_2D_ARRAY, num_interpolate_slots, input_tex);
 199         glCreateTextures(GL_TEXTURE_2D_ARRAY, num_interpolate_slots, gray_tex);
 200         glCreateTextures(GL_TEXTURE_2D, num_interpolate_slots, fade_y_output_tex);
 201         glCreateTextures(GL_TEXTURE_2D, num_interpolate_slots, fade_cbcr_output_tex);
 202         glCreateTextures(GL_TEXTURE_2D, num_interpolate_slots, cb_tex);
 203         glCreateTextures(GL_TEXTURE_2D, num_interpolate_slots, cr_tex);
 204         check_error();
 205
 206         size_t width = global_flags.width, height = global_flags.height;
 207         int levels = find_num_levels(width, height);
 208         for (size_t i = 0; i < num_interpolate_slots; ++i) {
 209                 glTextureStorage3D(input_tex[i], levels, GL_RGBA8, width, height, 2);
 210                 check_error();
 211                 glTextureStorage3D(gray_tex[i], levels, GL_R8, width, height, 2);
 212                 check_error();
 213                 glTextureStorage2D(fade_y_output_tex[i], 1, GL_R8, width, height);
 214                 check_error();
 215                 glTextureStorage2D(fade_cbcr_output_tex[i], 1, GL_RG8, width, height);
 216                 check_error();
 217                 glTextureStorage2D(cb_tex[i], 1, GL_R8, width / 2, height);
 218                 check_error();
 219                 glTextureStorage2D(cr_tex[i], 1, GL_R8, width / 2, height);
 220                 check_error();
 221
 222                 unique_ptr<InterpolatedFrameResources> resource(new InterpolatedFrameResources);
 223                 resource->owner = this;
 224                 resource->input_tex = input_tex[i];
 225                 resource->gray_tex = gray_tex[i];
 226                 resource->fade_y_output_tex = fade_y_output_tex[i];
 227                 resource->fade_cbcr_output_tex = fade_cbcr_output_tex[i];
 228                 resource->cb_tex = cb_tex[i];
 229                 resource->cr_tex = cr_tex[i];
 230                 glCreateFramebuffers(2, resource->input_fbos);
 231                 check_error();
 232                 glCreateFramebuffers(1, &resource->fade_fbo);
 233                 check_error();
 234
 235                 glNamedFramebufferTextureLayer(resource->input_fbos[0], GL_COLOR_ATTACHMENT0, input_tex[i], 0, 0);
 236                 check_error();
 237                 glNamedFramebufferTextureLayer(resource->input_fbos[0], GL_COLOR_ATTACHMENT1, gray_tex[i], 0, 0);
 238                 check_error();
 239                 glNamedFramebufferTextureLayer(resource->input_fbos[1], GL_COLOR_ATTACHMENT0, input_tex[i], 0, 1);
 240                 check_error();
 241                 glNamedFramebufferTextureLayer(resource->input_fbos[1], GL_COLOR_ATTACHMENT1, gray_tex[i], 0, 1);
 242                 check_error();
 243                 glNamedFramebufferTexture(resource->fade_fbo, GL_COLOR_ATTACHMENT0, fade_y_output_tex[i], 0);
 244                 check_error();
 245                 glNamedFramebufferTexture(resource->fade_fbo, GL_COLOR_ATTACHMENT1, fade_cbcr_output_tex[i], 0);
 246                 check_error();
 247
 248                 GLuint bufs[] = { GL_COLOR_ATTACHMENT0, GL_COLOR_ATTACHMENT1 };
 249                 glNamedFramebufferDrawBuffers(resource->input_fbos[0], 2, bufs);
 250                 check_error();
 251                 glNamedFramebufferDrawBuffers(resource->input_fbos[1], 2, bufs);
 252                 check_error();
 253                 glNamedFramebufferDrawBuffers(resource->fade_fbo, 2, bufs);
 254                 check_error();
 255
 256                 glCreateBuffers(1, &resource->pbo);
 257                 check_error();
 258                 glNamedBufferStorage(resource->pbo, width * height * 4, nullptr, GL_MAP_READ_BIT | GL_MAP_PERSISTENT_BIT);
 259                 check_error();
 260                 resource->pbo_contents = glMapNamedBufferRange(resource->pbo, 0, width * height * 4, GL_MAP_READ_BIT | GL_MAP_PERSISTENT_BIT);
 261                 interpolate_resources.push_back(move(resource));
 262         }
 263
 264         check_error();
 265
 266         OperatingPoint op;
 267         if (global_flags.interpolation_quality == 0 ||
 268             global_flags.interpolation_quality == 1) {
 269                 op = operating_point1;
 270         } else if (global_flags.interpolation_quality == 2) {
 271                 op = operating_point2;
 272         } else if (global_flags.interpolation_quality == 3) {
 273                 op = operating_point3;
 274         } else if (global_flags.interpolation_quality == 4) {
 275                 op = operating_point4;
 276         } else {
 277                 // Quality 0 will be changed to 1 in flags.cpp.
 278                 assert(false);
 279         }
 280
 281         compute_flow.reset(new DISComputeFlow(width, height, op));
 282         interpolate.reset(new Interpolate(op, /*split_ycbcr_output=*/true));
 283         interpolate_no_split.reset(new Interpolate(op, /*split_ycbcr_output=*/false));
 284         chroma_subsampler.reset(new ChromaSubsampler);
 285         check_error();
 286
 287         // The “last frame” is initially black.
 288         unique_ptr<uint8_t[]> y(new uint8_t[global_flags.width * global_flags.height]);
 289         unique_ptr<uint8_t[]> cb_or_cr(new uint8_t[(global_flags.width / 2) * global_flags.height]);
 290         memset(y.get(), 16, global_flags.width * global_flags.height);
 291         memset(cb_or_cr.get(), 128, (global_flags.width / 2) * global_flags.height);
 292         last_frame = encode_jpeg(y.get(), cb_or_cr.get(), cb_or_cr.get(), global_flags.width, global_flags.height, /*exif_data=*/"");
 293
 294         if (file_avctx != nullptr) {
 295                 with_subtitles = Mux::WITHOUT_SUBTITLES;
 296         } else {
 297                 with_subtitles = Mux::WITH_SUBTITLES;
 298         }
 299 }
 300
 301 VideoStream::~VideoStream()
 302 {
 303         if (last_flow_tex != 0) {
 304                 compute_flow->release_texture(last_flow_tex);
 305         }
 306
 307         for (const unique_ptr<InterpolatedFrameResources> &resource : interpolate_resources) {
 308                 glUnmapNamedBuffer(resource->pbo);
 309                 check_error();
 310                 glDeleteBuffers(1, &resource->pbo);
 311                 check_error();
 312                 glDeleteFramebuffers(2, resource->input_fbos);
 313                 check_error();
 314                 glDeleteFramebuffers(1, &resource->fade_fbo);
 315                 check_error();
 316                 glDeleteTextures(1, &resource->input_tex);
 317                 check_error();
 318                 glDeleteTextures(1, &resource->gray_tex);
 319                 check_error();
 320                 glDeleteTextures(1, &resource->fade_y_output_tex);
 321                 check_error();
 322                 glDeleteTextures(1, &resource->fade_cbcr_output_tex);
 323                 check_error();
 324                 glDeleteTextures(1, &resource->cb_tex);
 325                 check_error();
 326                 glDeleteTextures(1, &resource->cr_tex);
 327                 check_error();
 328         }
 329         assert(interpolate_resources.size() == num_interpolate_slots);
 330 }
 331
 332 void VideoStream::start()
 333 {
 334         if (avctx == nullptr) {
 335                 avctx = avformat_alloc_context();
 336
 337                 // We use Matroska, because it's pretty much the only mux where FFmpeg
 338                 // allows writing chroma location to override JFIF's default center placement.
 339                 // (Note that at the time of writing, however, FFmpeg does not correctly
 340                 // _read_ this information!)
 341                 avctx->oformat = av_guess_format("matroska", nullptr, nullptr);
 342
 343                 uint8_t *buf = (uint8_t *)av_malloc(MUX_BUFFER_SIZE);
 344                 avctx->pb = avio_alloc_context(buf, MUX_BUFFER_SIZE, 1, this, nullptr, nullptr, nullptr);
 345                 avctx->pb->write_data_type = &VideoStream::write_packet2_thunk;
 346                 avctx->pb->ignore_boundary_point = 1;
 347
 348                 avctx->flags = AVFMT_FLAG_CUSTOM_IO;
 349         }
 350
 351         AVCodecParameters *audio_codecpar = avcodec_parameters_alloc();
 352
 353         audio_codecpar->codec_type = AVMEDIA_TYPE_AUDIO;
 354         audio_codecpar->codec_id = AV_CODEC_ID_PCM_S32LE;
 355         audio_codecpar->channel_layout = AV_CH_LAYOUT_STEREO;
 356         audio_codecpar->channels = 2;
 357         audio_codecpar->sample_rate = OUTPUT_FREQUENCY;
 358
 359         size_t width = global_flags.width, height = global_flags.height;  // Doesn't matter for MJPEG.
 360         mux.reset(new Mux(avctx, width, height, Mux::CODEC_MJPEG, /*video_extradata=*/"", audio_codecpar,
 361                           AVCOL_SPC_BT709, COARSE_TIMEBASE, /*write_callback=*/nullptr, Mux::WRITE_FOREGROUND, {}, with_subtitles));
 362
 363         avcodec_parameters_free(&audio_codecpar);
 364         encode_thread = thread(&VideoStream::encode_thread_func, this);
 365 }
 366
 367 void VideoStream::stop()
 368 {
 369         should_quit = true;
 370         queue_changed.notify_all();
 371         clear_queue();
 372         encode_thread.join();
 373 }
 374
 375 void VideoStream::clear_queue()
 376 {
 377         deque<QueuedFrame> q;
 378
 379         {
 380                 lock_guard<mutex> lock(queue_lock);
 381                 q = move(frame_queue);
 382         }
 383
 384         // These are not RAII-ed, unfortunately, so we'll need to clean them ourselves.
 385         // Note that release_texture() is thread-safe.
 386         for (const QueuedFrame &qf : q) {
 387                 if (qf.type == QueuedFrame::INTERPOLATED ||
 388                     qf.type == QueuedFrame::FADED_INTERPOLATED) {
 389                         if (qf.flow_tex != 0) {
 390                                 compute_flow->release_texture(qf.flow_tex);
 391                         }
 392                 }
 393                 if (qf.type == QueuedFrame::INTERPOLATED) {
 394                         interpolate->release_texture(qf.output_tex);
 395                         interpolate->release_texture(qf.cbcr_tex);
 396                 }
 397         }
 398
 399         // Destroy q outside the mutex, as that would be a double-lock.
 400 }
 401
 402 void VideoStream::schedule_original_frame(steady_clock::time_point local_pts,
 403                                           int64_t output_pts, function<void()> &&display_func,
 404                                           QueueSpotHolder &&queue_spot_holder,
 405                                           FrameOnDisk frame, const string &subtitle, bool include_audio)
 406 {
 407         fprintf(stderr, "output_pts=%" PRId64 "  original      input_pts=%" PRId64 "\n", output_pts, frame.pts);
 408
 409         QueuedFrame qf;
 410         qf.local_pts = local_pts;
 411         qf.type = QueuedFrame::ORIGINAL;
 412         qf.output_pts = output_pts;
 413         qf.display_func = move(display_func);
 414         qf.queue_spot_holder = move(queue_spot_holder);
 415         qf.subtitle = subtitle;
 416         FrameReader::Frame read_frame = frame_reader.read_frame(frame, /*read_video=*/true, include_audio);
 417         qf.encoded_jpeg.reset(new string(move(read_frame.video)));
 418         qf.audio = move(read_frame.audio);
 419
 420         lock_guard<mutex> lock(queue_lock);
 421         frame_queue.push_back(move(qf));
 422         queue_changed.notify_all();
 423 }
 424
 425 void VideoStream::schedule_faded_frame(steady_clock::time_point local_pts, int64_t output_pts,
 426                                        function<void()> &&display_func,
 427                                        QueueSpotHolder &&queue_spot_holder,
 428                                        FrameOnDisk frame1_spec, FrameOnDisk frame2_spec,
 429                                        float fade_alpha, const string &subtitle)
 430 {
 431         fprintf(stderr, "output_pts=%" PRId64 "  faded         input_pts=%" PRId64 ",%" PRId64 "  fade_alpha=%.2f\n", output_pts, frame1_spec.pts, frame2_spec.pts, fade_alpha);
 432
 433         // Get the temporary OpenGL resources we need for doing the fade.
 434         // (We share these with interpolated frames, which is slightly
 435         // overkill, but there's no need to waste resources on keeping
 436         // separate pools around.)
 437         BorrowedInterpolatedFrameResources resources;
 438         {
 439                 lock_guard<mutex> lock(queue_lock);
 440                 if (interpolate_resources.empty()) {
 441                         fprintf(stderr, "WARNING: Too many interpolated frames already in transit; dropping one.\n");
 442                         return;
 443                 }
 444                 resources = BorrowedInterpolatedFrameResources(interpolate_resources.front().release());
 445                 interpolate_resources.pop_front();
 446         }
 447
 448         bool did_decode;
 449
 450         shared_ptr<Frame> frame1 = decode_jpeg_with_cache(frame1_spec, DECODE_IF_NOT_IN_CACHE, &frame_reader, &did_decode);
 451         shared_ptr<Frame> frame2 = decode_jpeg_with_cache(frame2_spec, DECODE_IF_NOT_IN_CACHE, &frame_reader, &did_decode);
 452         wait_for_upload(frame1);
 453         wait_for_upload(frame2);
 454
 455         ycbcr_semiplanar_converter->prepare_chain_for_fade(frame1, frame2, fade_alpha)->render_to_fbo(resources->fade_fbo, global_flags.width, global_flags.height);
 456
 457         QueuedFrame qf;
 458         qf.local_pts = local_pts;
 459         qf.type = QueuedFrame::FADED;
 460         qf.output_pts = output_pts;
 461         qf.frame1 = frame1_spec;
 462         qf.display_func = move(display_func);
 463         qf.queue_spot_holder = move(queue_spot_holder);
 464         qf.subtitle = subtitle;
 465
 466         qf.secondary_frame = frame2_spec;
 467
 468         // Subsample and split Cb/Cr.
 469         chroma_subsampler->subsample_chroma(resources->fade_cbcr_output_tex, global_flags.width, global_flags.height, resources->cb_tex, resources->cr_tex);
 470
 471         // Read it down (asynchronously) to the CPU.
 472         glPixelStorei(GL_PACK_ROW_LENGTH, 0);
 473         glBindBuffer(GL_PIXEL_PACK_BUFFER, resources->pbo);
 474         check_error();
 475         glGetTextureImage(resources->fade_y_output_tex, 0, GL_RED, GL_UNSIGNED_BYTE, global_flags.width * global_flags.height * 4, BUFFER_OFFSET(0));
 476         check_error();
 477         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));
 478         check_error();
 479         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));
 480         check_error();
 481         glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
 482
 483         // Set a fence we can wait for to make sure the CPU sees the read.
 484         glMemoryBarrier(GL_CLIENT_MAPPED_BUFFER_BARRIER_BIT);
 485         check_error();
 486         qf.fence_created = steady_clock::now();
 487         qf.fence = RefCountedGLsync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
 488         check_error();
 489         qf.resources = move(resources);
 490         qf.local_pts = local_pts;
 491
 492         lock_guard<mutex> lock(queue_lock);
 493         frame_queue.push_back(move(qf));
 494         queue_changed.notify_all();
 495 }
 496
 497 void VideoStream::schedule_interpolated_frame(steady_clock::time_point local_pts,
 498                                               int64_t output_pts, function<void(shared_ptr<Frame>)> &&display_func,
 499                                               QueueSpotHolder &&queue_spot_holder,
 500                                               FrameOnDisk frame1, FrameOnDisk frame2,
 501                                               float alpha, FrameOnDisk secondary_frame, float fade_alpha, const string &subtitle,
 502                                               bool play_audio)
 503 {
 504         if (secondary_frame.pts != -1) {
 505                 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);
 506         } else {
 507                 fprintf(stderr, "output_pts=%" PRId64 "  interpolated  input_pts1=%" PRId64 " input_pts2=%" PRId64 " alpha=%.3f\n", output_pts, frame1.pts, frame2.pts, alpha);
 508         }
 509
 510         // Get the temporary OpenGL resources we need for doing the interpolation.
 511         BorrowedInterpolatedFrameResources resources;
 512         {
 513                 lock_guard<mutex> lock(queue_lock);
 514                 if (interpolate_resources.empty()) {
 515                         fprintf(stderr, "WARNING: Too many interpolated frames already in transit; dropping one.\n");
 516                         return;
 517                 }
 518                 resources = BorrowedInterpolatedFrameResources(interpolate_resources.front().release());
 519                 interpolate_resources.pop_front();
 520         }
 521
 522         QueuedFrame qf;
 523         qf.type = (secondary_frame.pts == -1) ? QueuedFrame::INTERPOLATED : QueuedFrame::FADED_INTERPOLATED;
 524         qf.output_pts = output_pts;
 525         qf.display_decoded_func = move(display_func);
 526         qf.queue_spot_holder = move(queue_spot_holder);
 527         qf.local_pts = local_pts;
 528         qf.subtitle = subtitle;
 529
 530         if (play_audio) {
 531                 qf.audio = frame_reader.read_frame(frame1, /*read_video=*/false, /*read_audio=*/true).audio;
 532         }
 533
 534         check_error();
 535
 536         // Convert frame0 and frame1 to OpenGL textures.
 537         for (size_t frame_no = 0; frame_no < 2; ++frame_no) {
 538                 FrameOnDisk frame_spec = frame_no == 1 ? frame2 : frame1;
 539                 bool did_decode;
 540                 shared_ptr<Frame> frame = decode_jpeg_with_cache(frame_spec, DECODE_IF_NOT_IN_CACHE, &frame_reader, &did_decode);
 541                 wait_for_upload(frame);
 542                 ycbcr_converter->prepare_chain_for_conversion(frame)->render_to_fbo(resources->input_fbos[frame_no], global_flags.width, global_flags.height);
 543                 if (frame_no == 1) {
 544                         qf.exif_data = frame->exif_data;  // Use the white point from the last frame.
 545                 }
 546         }
 547
 548         glGenerateTextureMipmap(resources->input_tex);
 549         check_error();
 550         glGenerateTextureMipmap(resources->gray_tex);
 551         check_error();
 552
 553         GLuint flow_tex;
 554         if (last_flow_tex != 0 && frame1 == last_frame1 && frame2 == last_frame2) {
 555                 // Reuse the flow from previous computation. This frequently happens
 556                 // if we slow down by more than 2x, so that there are multiple interpolated
 557                 // frames between each original.
 558                 flow_tex = last_flow_tex;
 559                 qf.flow_tex = 0;
 560         } else {
 561                 // Cache miss, so release last_flow_tex.
 562                 qf.flow_tex = last_flow_tex;
 563
 564                 // Compute the flow.
 565                 flow_tex = compute_flow->exec(resources->gray_tex, DISComputeFlow::FORWARD_AND_BACKWARD, DISComputeFlow::DO_NOT_RESIZE_FLOW);
 566                 check_error();
 567
 568                 // Store the flow texture for possible reuse next frame.
 569                 last_flow_tex = flow_tex;
 570                 last_frame1 = frame1;
 571                 last_frame2 = frame2;
 572         }
 573
 574         if (secondary_frame.pts != -1) {
 575                 // Fade. First kick off the interpolation.
 576                 tie(qf.output_tex, ignore) = interpolate_no_split->exec(resources->input_tex, resources->gray_tex, flow_tex, global_flags.width, global_flags.height, alpha);
 577                 check_error();
 578
 579                 // Now decode the image we are fading against.
 580                 bool did_decode;
 581                 shared_ptr<Frame> frame2 = decode_jpeg_with_cache(secondary_frame, DECODE_IF_NOT_IN_CACHE, &frame_reader, &did_decode);
 582                 wait_for_upload(frame2);
 583
 584                 // Then fade against it, putting it into the fade Y' and CbCr textures.
 585                 RGBTriplet neutral_color = get_neutral_color(qf.exif_data);
 586                 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);
 587
 588                 // Subsample and split Cb/Cr.
 589                 chroma_subsampler->subsample_chroma(resources->fade_cbcr_output_tex, global_flags.width, global_flags.height, resources->cb_tex, resources->cr_tex);
 590
 591                 interpolate_no_split->release_texture(qf.output_tex);
 592
 593                 // We already applied the white balance, so don't have the client redo it.
 594                 qf.exif_data.clear();
 595         } else {
 596                 tie(qf.output_tex, qf.cbcr_tex) = interpolate->exec(resources->input_tex, resources->gray_tex, flow_tex, global_flags.width, global_flags.height, alpha);
 597                 check_error();
 598
 599                 // Subsample and split Cb/Cr.
 600                 chroma_subsampler->subsample_chroma(qf.cbcr_tex, global_flags.width, global_flags.height, resources->cb_tex, resources->cr_tex);
 601         }
 602
 603         // We could have released qf.flow_tex here, but to make sure we don't cause a stall
 604         // when trying to reuse it for the next frame, we can just as well hold on to it
 605         // and release it only when the readback is done.
 606         //
 607         // TODO: This is maybe less relevant now that qf.flow_tex contains the texture we used
 608         // _last_ frame, not this one.
 609
 610         // Read it down (asynchronously) to the CPU.
 611         glPixelStorei(GL_PACK_ROW_LENGTH, 0);
 612         glBindBuffer(GL_PIXEL_PACK_BUFFER, resources->pbo);
 613         check_error();
 614         if (secondary_frame.pts != -1) {
 615                 glGetTextureImage(resources->fade_y_output_tex, 0, GL_RED, GL_UNSIGNED_BYTE, global_flags.width * global_flags.height * 4, BUFFER_OFFSET(0));
 616         } else {
 617                 glGetTextureImage(qf.output_tex, 0, GL_RED, GL_UNSIGNED_BYTE, global_flags.width * global_flags.height * 4, BUFFER_OFFSET(0));
 618         }
 619         check_error();
 620         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));
 621         check_error();
 622         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));
 623         check_error();
 624         glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
 625
 626         // Set a fence we can wait for to make sure the CPU sees the read.
 627         glMemoryBarrier(GL_CLIENT_MAPPED_BUFFER_BARRIER_BIT);
 628         check_error();
 629         qf.fence_created = steady_clock::now();
 630         qf.fence = RefCountedGLsync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
 631         check_error();
 632         qf.resources = move(resources);
 633
 634         lock_guard<mutex> lock(queue_lock);
 635         frame_queue.push_back(move(qf));
 636         queue_changed.notify_all();
 637 }
 638
 639 void VideoStream::schedule_refresh_frame(steady_clock::time_point local_pts,
 640                                          int64_t output_pts, function<void()> &&display_func,
 641                                          QueueSpotHolder &&queue_spot_holder, const string &subtitle)
 642 {
 643         QueuedFrame qf;
 644         qf.type = QueuedFrame::REFRESH;
 645         qf.output_pts = output_pts;
 646         qf.display_func = move(display_func);
 647         qf.queue_spot_holder = move(queue_spot_holder);
 648         qf.subtitle = subtitle;
 649
 650         lock_guard<mutex> lock(queue_lock);
 651         frame_queue.push_back(move(qf));
 652         queue_changed.notify_all();
 653 }
 654
 655 void VideoStream::schedule_silence(steady_clock::time_point local_pts, int64_t output_pts,
 656                                    int64_t length_pts, QueueSpotHolder &&queue_spot_holder)
 657 {
 658         QueuedFrame qf;
 659         qf.type = QueuedFrame::SILENCE;
 660         qf.output_pts = output_pts;
 661         qf.queue_spot_holder = move(queue_spot_holder);
 662         qf.silence_length_pts = length_pts;
 663
 664         lock_guard<mutex> lock(queue_lock);
 665         frame_queue.push_back(move(qf));
 666         queue_changed.notify_all();
 667 }
 668
 669 namespace {
 670
 671 RefCountedTexture clone_r8_texture(GLuint src_tex, unsigned width, unsigned height)
 672 {
 673         GLuint tex;
 674         glCreateTextures(GL_TEXTURE_2D, 1, &tex);
 675         check_error();
 676         glTextureStorage2D(tex, 1, GL_R8, width, height);
 677         check_error();
 678         glCopyImageSubData(src_tex, GL_TEXTURE_2D, 0, 0, 0, 0,
 679                            tex, GL_TEXTURE_2D, 0, 0, 0, 0,
 680                            width, height, 1);
 681         check_error();
 682         glTextureParameteri(tex, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
 683         check_error();
 684         glTextureParameteri(tex, GL_TEXTURE_MAG_FILTER, GL_LINEAR);
 685         check_error();
 686         glTextureParameteri(tex, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
 687         check_error();
 688         glTextureParameteri(tex, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
 689         check_error();
 690
 691         return RefCountedTexture(new GLuint(tex), TextureDeleter());
 692 }
 693
 694 }  // namespace
 695
 696 void VideoStream::encode_thread_func()
 697 {
 698         pthread_setname_np(pthread_self(), "VideoStream");
 699         QSurface *surface = create_surface();
 700         QOpenGLContext *context = create_context(surface);
 701         bool ok = make_current(context, surface);
 702         if (!ok) {
 703                 fprintf(stderr, "Video stream couldn't get an OpenGL context\n");
 704                 abort();
 705         }
 706
 707         init_pbo_pool();
 708
 709         while (!should_quit) {
 710                 QueuedFrame qf;
 711                 {
 712                         unique_lock<mutex> lock(queue_lock);
 713
 714                         // Wait until we have a frame to play.
 715                         queue_changed.wait(lock, [this] {
 716                                 return !frame_queue.empty() || should_quit;
 717                         });
 718                         if (should_quit) {
 719                                 break;
 720                         }
 721                         steady_clock::time_point frame_start = frame_queue.front().local_pts;
 722
 723                         // Now sleep until the frame is supposed to start (the usual case),
 724                         // _or_ clear_queue() happened.
 725                         bool aborted;
 726                         if (output_fast_forward) {
 727                                 aborted = frame_queue.empty() || frame_queue.front().local_pts != frame_start;
 728                         } else {
 729                                 aborted = queue_changed.wait_until(lock, frame_start, [this, frame_start] {
 730                                         return frame_queue.empty() || frame_queue.front().local_pts != frame_start;
 731                                 });
 732                         }
 733                         if (aborted) {
 734                                 // clear_queue() happened, so don't play this frame after all.
 735                                 continue;
 736                         }
 737                         qf = move(frame_queue.front());
 738                         frame_queue.pop_front();
 739                 }
 740
 741                 // Hack: We mux the subtitle packet one time unit before the actual frame,
 742                 // so that Nageru is sure to get it first.
 743                 if (!qf.subtitle.empty() && with_subtitles == Mux::WITH_SUBTITLES) {
 744                         AVPacket pkt;
 745                         av_init_packet(&pkt);
 746                         pkt.stream_index = mux->get_subtitle_stream_idx();
 747                         assert(pkt.stream_index != -1);
 748                         pkt.data = (uint8_t *)qf.subtitle.data();
 749                         pkt.size = qf.subtitle.size();
 750                         pkt.flags = 0;
 751                         pkt.duration = lrint(TIMEBASE / global_flags.output_framerate);  // Doesn't really matter for Nageru.
 752                         mux->add_packet(pkt, qf.output_pts - 1, qf.output_pts - 1);
 753                 }
 754
 755                 if (qf.type == QueuedFrame::ORIGINAL) {
 756                         // Send the JPEG frame on, unchanged.
 757                         string jpeg = move(*qf.encoded_jpeg);
 758                         AVPacket pkt;
 759                         av_init_packet(&pkt);
 760                         pkt.stream_index = 0;
 761                         pkt.data = (uint8_t *)jpeg.data();
 762                         pkt.size = jpeg.size();
 763                         pkt.flags = AV_PKT_FLAG_KEY;
 764                         mux->add_packet(pkt, qf.output_pts, qf.output_pts);
 765                         last_frame = move(jpeg);
 766
 767                         add_audio_or_silence(qf);
 768                 } else if (qf.type == QueuedFrame::FADED) {
 769                         steady_clock::time_point start = steady_clock::now();
 770                         glClientWaitSync(qf.fence.get(), /*flags=*/0, GL_TIMEOUT_IGNORED);
 771                         steady_clock::time_point stop = steady_clock::now();
 772                         metric_fade_fence_wait_time_seconds.count_event(duration<double>(stop - start).count());
 773                         metric_fade_latency_seconds.count_event(duration<double>(stop - qf.fence_created).count());
 774
 775                         // Now JPEG encode it, and send it on to the stream.
 776                         string jpeg = encode_jpeg_from_pbo(qf.resources->pbo_contents, global_flags.width, global_flags.height, /*exif_data=*/"");
 777
 778                         AVPacket pkt;
 779                         av_init_packet(&pkt);
 780                         pkt.stream_index = 0;
 781                         pkt.data = (uint8_t *)jpeg.data();
 782                         pkt.size = jpeg.size();
 783                         pkt.flags = AV_PKT_FLAG_KEY;
 784                         mux->add_packet(pkt, qf.output_pts, qf.output_pts);
 785                         last_frame = move(jpeg);
 786
 787                         add_audio_or_silence(qf);
 788                 } else if (qf.type == QueuedFrame::INTERPOLATED || qf.type == QueuedFrame::FADED_INTERPOLATED) {
 789                         steady_clock::time_point start = steady_clock::now();
 790                         glClientWaitSync(qf.fence.get(), /*flags=*/0, GL_TIMEOUT_IGNORED);
 791                         steady_clock::time_point stop = steady_clock::now();
 792                         metric_interpolation_fence_wait_time_seconds.count_event(duration<double>(stop - start).count());
 793                         metric_interpolation_latency_seconds.count_event(duration<double>(stop - qf.fence_created).count());
 794
 795                         // Send it on to display.
 796                         if (qf.display_decoded_func != nullptr) {
 797                                 shared_ptr<Frame> frame(new Frame);
 798                                 if (qf.type == QueuedFrame::FADED_INTERPOLATED) {
 799                                         frame->y = clone_r8_texture(qf.resources->fade_y_output_tex, global_flags.width, global_flags.height);
 800                                 } else {
 801                                         frame->y = clone_r8_texture(qf.output_tex, global_flags.width, global_flags.height);
 802                                 }
 803                                 frame->cb = clone_r8_texture(qf.resources->cb_tex, global_flags.width / 2, global_flags.height);
 804                                 frame->cr = clone_r8_texture(qf.resources->cr_tex, global_flags.width / 2, global_flags.height);
 805                                 frame->width = global_flags.width;
 806                                 frame->height = global_flags.height;
 807                                 frame->chroma_subsampling_x = 2;
 808                                 frame->chroma_subsampling_y = 1;
 809                                 frame->uploaded_ui_thread = RefCountedGLsync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
 810                                 qf.display_decoded_func(move(frame));
 811                         }
 812
 813                         // Now JPEG encode it, and send it on to the stream.
 814                         string jpeg = encode_jpeg_from_pbo(qf.resources->pbo_contents, global_flags.width, global_flags.height, move(qf.exif_data));
 815                         if (qf.flow_tex != 0) {
 816                                 compute_flow->release_texture(qf.flow_tex);
 817                         }
 818                         if (qf.type != QueuedFrame::FADED_INTERPOLATED) {
 819                                 interpolate->release_texture(qf.output_tex);
 820                                 interpolate->release_texture(qf.cbcr_tex);
 821                         }
 822
 823                         AVPacket pkt;
 824                         av_init_packet(&pkt);
 825                         pkt.stream_index = 0;
 826                         pkt.data = (uint8_t *)jpeg.data();
 827                         pkt.size = jpeg.size();
 828                         pkt.flags = AV_PKT_FLAG_KEY;
 829                         mux->add_packet(pkt, qf.output_pts, qf.output_pts);
 830                         last_frame = move(jpeg);
 831
 832                         add_audio_or_silence(qf);
 833                 } else if (qf.type == QueuedFrame::REFRESH) {
 834                         AVPacket pkt;
 835                         av_init_packet(&pkt);
 836                         pkt.stream_index = 0;
 837                         pkt.data = (uint8_t *)last_frame.data();
 838                         pkt.size = last_frame.size();
 839                         pkt.flags = AV_PKT_FLAG_KEY;
 840                         mux->add_packet(pkt, qf.output_pts, qf.output_pts);
 841
 842                         add_audio_or_silence(qf);  // Definitely silence.
 843                 } else if (qf.type == QueuedFrame::SILENCE) {
 844                         add_silence(qf.output_pts, qf.silence_length_pts);
 845                 } else {
 846                         assert(false);
 847                 }
 848                 if (qf.display_func != nullptr) {
 849                         qf.display_func();
 850                 }
 851         }
 852 }
 853
 854 int VideoStream::write_packet2_thunk(void *opaque, uint8_t *buf, int buf_size, AVIODataMarkerType type, int64_t time)
 855 {
 856         VideoStream *video_stream = (VideoStream *)opaque;
 857         return video_stream->write_packet2(buf, buf_size, type, time);
 858 }
 859
 860 int VideoStream::write_packet2(uint8_t *buf, int buf_size, AVIODataMarkerType type, int64_t time)
 861 {
 862         if (type == AVIO_DATA_MARKER_SYNC_POINT || type == AVIO_DATA_MARKER_BOUNDARY_POINT) {
 863                 seen_sync_markers = true;
 864         } else if (type == AVIO_DATA_MARKER_UNKNOWN && !seen_sync_markers) {
 865                 // We don't know if this is a keyframe or not (the muxer could
 866                 // avoid marking it), so we just have to make the best of it.
 867                 type = AVIO_DATA_MARKER_SYNC_POINT;
 868         }
 869
 870         HTTPD::StreamID stream_id{ HTTPD::MAIN_STREAM, 0 };
 871         if (type == AVIO_DATA_MARKER_HEADER) {
 872                 stream_mux_header.append((char *)buf, buf_size);
 873                 global_httpd->set_header(stream_id, stream_mux_header);
 874         } else {
 875                 global_httpd->add_data(stream_id, (char *)buf, buf_size, type == AVIO_DATA_MARKER_SYNC_POINT, time, AVRational{ AV_TIME_BASE, 1 });
 876         }
 877         return buf_size;
 878 }
 879
 880 void VideoStream::add_silence(int64_t pts, int64_t length_pts)
 881 {
 882         // At 59.94, this will never quite add up (even discounting refresh frames,
 883         // which have unpredictable length), but hopefully, the player in the other
 884         // end should be able to stretch silence easily enough.
 885         long num_samples = lrint(length_pts * double(OUTPUT_FREQUENCY) / double(TIMEBASE)) * 2;
 886         uint8_t *zero = (uint8_t *)calloc(num_samples, sizeof(int32_t));
 887
 888         AVPacket pkt;
 889         av_init_packet(&pkt);
 890         pkt.stream_index = 1;
 891         pkt.data = zero;
 892         pkt.size = num_samples * sizeof(int32_t);
 893         pkt.flags = AV_PKT_FLAG_KEY;
 894         mux->add_packet(pkt, pts, pts);
 895
 896         free(zero);
 897 }
 898
 899 void VideoStream::add_audio_or_silence(const QueuedFrame &qf)
 900 {
 901         if (qf.audio.empty()) {
 902                 int64_t frame_length = lrint(double(TIMEBASE) / global_flags.output_framerate);
 903                 add_silence(qf.output_pts, frame_length);
 904         } else {
 905                 AVPacket pkt;
 906                 av_init_packet(&pkt);
 907                 pkt.stream_index = 1;
 908                 pkt.data = (uint8_t *)qf.audio.data();
 909                 pkt.size = qf.audio.size();
 910                 pkt.flags = AV_PKT_FLAG_KEY;
 911                 mux->add_packet(pkt, qf.output_pts, qf.output_pts);
 912         }
 913 }