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