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