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

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