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