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