git.sesse.net Git - nageru/blob - 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 "context.h"
  10 #include "flags.h"
  11 #include "flow.h"
  12 #include "httpd.h"
  13 #include "jpeg_frame_view.h"
  14 #include "movit/util.h"
  15 #include "mux.h"
  16 #include "player.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 namespace {
  30
  31 string read_file(const string &filename)
  32 {
  33         FILE *fp = fopen(filename.c_str(), "rb");
  34         if (fp == nullptr) {
  35                 perror(filename.c_str());
  36                 return "";
  37         }
  38
  39         fseek(fp, 0, SEEK_END);
  40         long len = ftell(fp);
  41         rewind(fp);
  42
  43         string ret;
  44         ret.resize(len);
  45         fread(&ret[0], len, 1, fp);
  46         fclose(fp);
  47         return ret;
  48 }
  49
  50 }  // namespace
  51
  52 struct VectorDestinationManager {
  53         jpeg_destination_mgr pub;
  54         std::vector<uint8_t> dest;
  55
  56         VectorDestinationManager()
  57         {
  58                 pub.init_destination = init_destination_thunk;
  59                 pub.empty_output_buffer = empty_output_buffer_thunk;
  60                 pub.term_destination = term_destination_thunk;
  61         }
  62
  63         static void init_destination_thunk(j_compress_ptr ptr)
  64         {
  65                 ((VectorDestinationManager *)(ptr->dest))->init_destination();
  66         }
  67
  68         inline void init_destination()
  69         {
  70                 make_room(0);
  71         }
  72
  73         static boolean empty_output_buffer_thunk(j_compress_ptr ptr)
  74         {
  75                 return ((VectorDestinationManager *)(ptr->dest))->empty_output_buffer();
  76         }
  77
  78         inline bool empty_output_buffer()
  79         {
  80                 make_room(dest.size());  // Should ignore pub.free_in_buffer!
  81                 return true;
  82         }
  83
  84         inline void make_room(size_t bytes_used)
  85         {
  86                 dest.resize(bytes_used + 4096);
  87                 dest.resize(dest.capacity());
  88                 pub.next_output_byte = dest.data() + bytes_used;
  89                 pub.free_in_buffer = dest.size() - bytes_used;
  90         }
  91
  92         static void term_destination_thunk(j_compress_ptr ptr)
  93         {
  94                 ((VectorDestinationManager *)(ptr->dest))->term_destination();
  95         }
  96
  97         inline void term_destination()
  98         {
  99                 dest.resize(dest.size() - pub.free_in_buffer);
 100         }
 101 };
 102 static_assert(std::is_standard_layout<VectorDestinationManager>::value, "");
 103
 104 vector<uint8_t> encode_jpeg(const uint8_t *y_data, const uint8_t *cb_data, const uint8_t *cr_data, unsigned width, unsigned height)
 105 {
 106         VectorDestinationManager dest;
 107
 108         jpeg_compress_struct cinfo;
 109         jpeg_error_mgr jerr;
 110         cinfo.err = jpeg_std_error(&jerr);
 111         jpeg_create_compress(&cinfo);
 112
 113         cinfo.dest = (jpeg_destination_mgr *)&dest;
 114         cinfo.input_components = 3;
 115         cinfo.in_color_space = JCS_RGB;
 116         jpeg_set_defaults(&cinfo);
 117         constexpr int quality = 90;
 118         jpeg_set_quality(&cinfo, quality, /*force_baseline=*/false);
 119
 120         cinfo.image_width = width;
 121         cinfo.image_height = height;
 122         cinfo.raw_data_in = true;
 123         jpeg_set_colorspace(&cinfo, JCS_YCbCr);
 124         cinfo.comp_info[0].h_samp_factor = 2;
 125         cinfo.comp_info[0].v_samp_factor = 1;
 126         cinfo.comp_info[1].h_samp_factor = 1;
 127         cinfo.comp_info[1].v_samp_factor = 1;
 128         cinfo.comp_info[2].h_samp_factor = 1;
 129         cinfo.comp_info[2].v_samp_factor = 1;
 130         cinfo.CCIR601_sampling = true;  // Seems to be mostly ignored by libjpeg, though.
 131         jpeg_start_compress(&cinfo, true);
 132
 133         JSAMPROW yptr[8], cbptr[8], crptr[8];
 134         JSAMPARRAY data[3] = { yptr, cbptr, crptr };
 135         for (unsigned y = 0; y < height; y += 8) {
 136                 for (unsigned yy = 0; yy < 8; ++yy) {
 137                         yptr[yy] = const_cast<JSAMPROW>(&y_data[(y + yy) * width]);
 138                         cbptr[yy] = const_cast<JSAMPROW>(&cb_data[(y + yy) * width / 2]);
 139                         crptr[yy] = const_cast<JSAMPROW>(&cr_data[(y + yy) * width / 2]);
 140                 }
 141
 142                 jpeg_write_raw_data(&cinfo, data, /*num_lines=*/8);
 143         }
 144
 145         jpeg_finish_compress(&cinfo);
 146         jpeg_destroy_compress(&cinfo);
 147
 148         return move(dest.dest);
 149 }
 150
 151 VideoStream::VideoStream()
 152 {
 153         ycbcr_converter.reset(new YCbCrConverter(YCbCrConverter::OUTPUT_TO_DUAL_YCBCR, /*resource_pool=*/nullptr));
 154         ycbcr_semiplanar_converter.reset(new YCbCrConverter(YCbCrConverter::OUTPUT_TO_SEMIPLANAR, /*resource_pool=*/nullptr));
 155
 156         GLuint input_tex[num_interpolate_slots], gray_tex[num_interpolate_slots];
 157         GLuint fade_y_output_tex[num_interpolate_slots], fade_cbcr_output_tex[num_interpolate_slots];
 158         GLuint cb_tex[num_interpolate_slots], cr_tex[num_interpolate_slots];
 159
 160         glCreateTextures(GL_TEXTURE_2D_ARRAY, num_interpolate_slots, input_tex);
 161         glCreateTextures(GL_TEXTURE_2D_ARRAY, num_interpolate_slots, gray_tex);
 162         glCreateTextures(GL_TEXTURE_2D, num_interpolate_slots, fade_y_output_tex);
 163         glCreateTextures(GL_TEXTURE_2D, num_interpolate_slots, fade_cbcr_output_tex);
 164         glCreateTextures(GL_TEXTURE_2D, num_interpolate_slots, cb_tex);
 165         glCreateTextures(GL_TEXTURE_2D, num_interpolate_slots, cr_tex);
 166         check_error();
 167
 168         constexpr size_t width = 1280, height = 720;  // FIXME: adjustable width, height
 169         int levels = find_num_levels(width, height);
 170         for (size_t i = 0; i < num_interpolate_slots; ++i) {
 171                 glTextureStorage3D(input_tex[i], levels, GL_RGBA8, width, height, 2);
 172                 check_error();
 173                 glTextureStorage3D(gray_tex[i], levels, GL_R8, width, height, 2);
 174                 check_error();
 175                 glTextureStorage2D(fade_y_output_tex[i], 1, GL_R8, width, height);
 176                 check_error();
 177                 glTextureStorage2D(fade_cbcr_output_tex[i], 1, GL_RG8, width, height);
 178                 check_error();
 179                 glTextureStorage2D(cb_tex[i], 1, GL_R8, width / 2, height);
 180                 check_error();
 181                 glTextureStorage2D(cr_tex[i], 1, GL_R8, width / 2, height);
 182                 check_error();
 183
 184                 InterpolatedFrameResources resource;
 185                 resource.input_tex = input_tex[i];
 186                 resource.gray_tex = gray_tex[i];
 187                 resource.fade_y_output_tex = fade_y_output_tex[i];
 188                 resource.fade_cbcr_output_tex = fade_cbcr_output_tex[i];
 189                 resource.cb_tex = cb_tex[i];
 190                 resource.cr_tex = cr_tex[i];
 191                 glCreateFramebuffers(2, resource.input_fbos);
 192                 check_error();
 193                 glCreateFramebuffers(1, &resource.fade_fbo);
 194                 check_error();
 195
 196                 glNamedFramebufferTextureLayer(resource.input_fbos[0], GL_COLOR_ATTACHMENT0, input_tex[i], 0, 0);
 197                 check_error();
 198                 glNamedFramebufferTextureLayer(resource.input_fbos[0], GL_COLOR_ATTACHMENT1, gray_tex[i], 0, 0);
 199                 check_error();
 200                 glNamedFramebufferTextureLayer(resource.input_fbos[1], GL_COLOR_ATTACHMENT0, input_tex[i], 0, 1);
 201                 check_error();
 202                 glNamedFramebufferTextureLayer(resource.input_fbos[1], GL_COLOR_ATTACHMENT1, gray_tex[i], 0, 1);
 203                 check_error();
 204                 glNamedFramebufferTexture(resource.fade_fbo, GL_COLOR_ATTACHMENT0, fade_y_output_tex[i], 0);
 205                 check_error();
 206                 glNamedFramebufferTexture(resource.fade_fbo, GL_COLOR_ATTACHMENT1, fade_cbcr_output_tex[i], 0);
 207                 check_error();
 208
 209                 GLuint bufs[] = { GL_COLOR_ATTACHMENT0, GL_COLOR_ATTACHMENT1 };
 210                 glNamedFramebufferDrawBuffers(resource.input_fbos[0], 2, bufs);
 211                 check_error();
 212                 glNamedFramebufferDrawBuffers(resource.input_fbos[1], 2, bufs);
 213                 check_error();
 214                 glNamedFramebufferDrawBuffers(resource.fade_fbo, 2, bufs);
 215                 check_error();
 216
 217                 glCreateBuffers(1, &resource.pbo);
 218                 check_error();
 219                 glNamedBufferStorage(resource.pbo, width * height * 4, nullptr, GL_MAP_READ_BIT | GL_MAP_PERSISTENT_BIT);
 220                 check_error();
 221                 resource.pbo_contents = glMapNamedBufferRange(resource.pbo, 0, width * height * 4, GL_MAP_READ_BIT | GL_MAP_PERSISTENT_BIT);
 222                 interpolate_resources.push_back(resource);
 223         }
 224
 225         check_error();
 226
 227         OperatingPoint op;
 228         if (global_flags.interpolation_quality == 1) {
 229                 op = operating_point1;
 230         } else if (global_flags.interpolation_quality == 2) {
 231                 op = operating_point2;
 232         } else if (global_flags.interpolation_quality == 3) {
 233                 op = operating_point3;
 234         } else if (global_flags.interpolation_quality == 4) {
 235                 op = operating_point4;
 236         } else {
 237                 assert(false);
 238         }
 239
 240         compute_flow.reset(new DISComputeFlow(width, height, op));
 241         interpolate.reset(new Interpolate(op, /*split_ycbcr_output=*/true));
 242         interpolate_no_split.reset(new Interpolate(op, /*split_ycbcr_output=*/false));
 243         chroma_subsampler.reset(new ChromaSubsampler);
 244         check_error();
 245
 246         // The “last frame” is initially black.
 247         unique_ptr<uint8_t[]> y(new uint8_t[1280 * 720]);
 248         unique_ptr<uint8_t[]> cb_or_cr(new uint8_t[640 * 720]);
 249         memset(y.get(), 16, 1280 * 720);
 250         memset(cb_or_cr.get(), 128, 640 * 720);
 251         last_frame = encode_jpeg(y.get(), cb_or_cr.get(), cb_or_cr.get(), 1280, 720);
 252 }
 253
 254 VideoStream::~VideoStream() {}
 255
 256 void VideoStream::start()
 257 {
 258         AVFormatContext *avctx = avformat_alloc_context();
 259         avctx->oformat = av_guess_format("nut", nullptr, nullptr);
 260
 261         uint8_t *buf = (uint8_t *)av_malloc(MUX_BUFFER_SIZE);
 262         avctx->pb = avio_alloc_context(buf, MUX_BUFFER_SIZE, 1, this, nullptr, nullptr, nullptr);
 263         avctx->pb->write_data_type = &VideoStream::write_packet2_thunk;
 264         avctx->pb->ignore_boundary_point = 1;
 265
 266         Mux::Codec video_codec = Mux::CODEC_MJPEG;
 267
 268         avctx->flags = AVFMT_FLAG_CUSTOM_IO;
 269
 270         string video_extradata;
 271
 272         constexpr int width = 1280, height = 720;  // Doesn't matter for MJPEG.
 273         stream_mux.reset(new Mux(avctx, width, height, video_codec, video_extradata, /*audio_codec_parameters=*/nullptr, COARSE_TIMEBASE,
 274                 /*write_callback=*/nullptr, Mux::WRITE_FOREGROUND, {}));
 275
 276
 277         encode_thread = thread(&VideoStream::encode_thread_func, this);
 278 }
 279
 280 void VideoStream::stop()
 281 {
 282         encode_thread.join();
 283 }
 284
 285 void VideoStream::clear_queue()
 286 {
 287         unique_lock<mutex> lock(queue_lock);
 288         frame_queue.clear();
 289 }
 290
 291 void VideoStream::schedule_original_frame(steady_clock::time_point local_pts,
 292                                           int64_t output_pts, function<void()> &&display_func,
 293                                           QueueSpotHolder &&queue_spot_holder,
 294                                           unsigned stream_idx, int64_t input_pts)
 295 {
 296         fprintf(stderr, "output_pts=%ld  original      input_pts=%ld\n", output_pts, input_pts);
 297
 298         // Preload the file from disk, so that the encoder thread does not get stalled.
 299         // TODO: Consider sending it through the queue instead.
 300         (void)read_file(filename_for_frame(stream_idx, input_pts));
 301
 302         QueuedFrame qf;
 303         qf.local_pts = local_pts;
 304         qf.type = QueuedFrame::ORIGINAL;
 305         qf.output_pts = output_pts;
 306         qf.stream_idx = stream_idx;
 307         qf.input_first_pts = input_pts;
 308         qf.display_func = move(display_func);
 309         qf.queue_spot_holder = move(queue_spot_holder);
 310
 311         unique_lock<mutex> lock(queue_lock);
 312         frame_queue.push_back(move(qf));
 313         queue_changed.notify_all();
 314 }
 315
 316 void VideoStream::schedule_faded_frame(steady_clock::time_point local_pts, int64_t output_pts,
 317                                        function<void()> &&display_func,
 318                                        QueueSpotHolder &&queue_spot_holder,
 319                                        unsigned stream_idx, int64_t input_pts, int secondary_stream_idx,
 320                                        int64_t secondary_input_pts, float fade_alpha)
 321 {
 322         fprintf(stderr, "output_pts=%ld  faded         input_pts=%ld,%ld  fade_alpha=%.2f\n", output_pts, input_pts, secondary_input_pts, fade_alpha);
 323
 324         // Get the temporary OpenGL resources we need for doing the fade.
 325         // (We share these with interpolated frames, which is slightly
 326         // overkill, but there's no need to waste resources on keeping
 327         // separate pools around.)
 328         InterpolatedFrameResources resources;
 329         {
 330                 unique_lock<mutex> lock(queue_lock);
 331                 if (interpolate_resources.empty()) {
 332                         fprintf(stderr, "WARNING: Too many interpolated frames already in transit; dropping one.\n");
 333                         return;
 334                 }
 335                 resources = interpolate_resources.front();
 336                 interpolate_resources.pop_front();
 337         }
 338
 339         bool did_decode;
 340
 341         JPEGID jpeg_id1;
 342         jpeg_id1.stream_idx = stream_idx;
 343         jpeg_id1.pts = input_pts;
 344         jpeg_id1.interpolated = false;
 345         shared_ptr<Frame> frame1 = decode_jpeg_with_cache(jpeg_id1, DECODE_IF_NOT_IN_CACHE, &did_decode);
 346
 347         JPEGID jpeg_id2;
 348         jpeg_id2.stream_idx = secondary_stream_idx;
 349         jpeg_id2.pts = secondary_input_pts;
 350         jpeg_id2.interpolated = false;
 351         shared_ptr<Frame> frame2 = decode_jpeg_with_cache(jpeg_id2, DECODE_IF_NOT_IN_CACHE, &did_decode);
 352
 353         ycbcr_semiplanar_converter->prepare_chain_for_fade(frame1, frame2, fade_alpha)->render_to_fbo(resources.fade_fbo, 1280, 720);
 354
 355         QueuedFrame qf;
 356         qf.local_pts = local_pts;
 357         qf.type = QueuedFrame::FADED;
 358         qf.output_pts = output_pts;
 359         qf.stream_idx = stream_idx;
 360         qf.resources = resources;
 361         qf.input_first_pts = input_pts;
 362         qf.display_func = move(display_func);
 363         qf.queue_spot_holder = move(queue_spot_holder);
 364
 365         qf.secondary_stream_idx = secondary_stream_idx;
 366         qf.secondary_input_pts = secondary_input_pts;
 367
 368         // Subsample and split Cb/Cr.
 369         chroma_subsampler->subsample_chroma(resources.fade_cbcr_output_tex, 1280, 720, resources.cb_tex, resources.cr_tex);
 370
 371         // Read it down (asynchronously) to the CPU.
 372         glPixelStorei(GL_PACK_ROW_LENGTH, 0);
 373         glBindBuffer(GL_PIXEL_PACK_BUFFER, resources.pbo);
 374         check_error();
 375         glGetTextureImage(resources.fade_y_output_tex, 0, GL_RED, GL_UNSIGNED_BYTE, 1280 * 720 * 4, BUFFER_OFFSET(0));
 376         check_error();
 377         glGetTextureImage(resources.cb_tex, 0, GL_RED, GL_UNSIGNED_BYTE, 1280 * 720 * 3, BUFFER_OFFSET(1280 * 720));
 378         check_error();
 379         glGetTextureImage(resources.cr_tex, 0, GL_RED, GL_UNSIGNED_BYTE, 1280 * 720 * 3 - 640 * 720, BUFFER_OFFSET(1280 * 720 + 640 * 720));
 380         check_error();
 381         glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
 382
 383         // Set a fence we can wait for to make sure the CPU sees the read.
 384         glMemoryBarrier(GL_CLIENT_MAPPED_BUFFER_BARRIER_BIT);
 385         check_error();
 386         qf.fence = RefCountedGLsync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
 387         check_error();
 388
 389         unique_lock<mutex> lock(queue_lock);
 390         frame_queue.push_back(move(qf));
 391         queue_changed.notify_all();
 392 }
 393
 394 void VideoStream::schedule_interpolated_frame(steady_clock::time_point local_pts,
 395                                               int64_t output_pts, function<void()> &&display_func,
 396                                               QueueSpotHolder &&queue_spot_holder,
 397                                               unsigned stream_idx, int64_t input_first_pts,
 398                                               int64_t input_second_pts, float alpha,
 399                                               int secondary_stream_idx, int64_t secondary_input_pts,
 400                                               float fade_alpha)
 401 {
 402         if (secondary_stream_idx != -1) {
 403                 fprintf(stderr, "output_pts=%ld  interpolated  input_pts1=%ld input_pts2=%ld alpha=%.3f  secondary_pts=%ld  fade_alpha=%.2f\n", output_pts, input_first_pts, input_second_pts, alpha, secondary_input_pts, fade_alpha);
 404         } else {
 405                 fprintf(stderr, "output_pts=%ld  interpolated  input_pts1=%ld input_pts2=%ld alpha=%.3f\n", output_pts, input_first_pts, input_second_pts, alpha);
 406         }
 407
 408         JPEGID id;
 409         if (secondary_stream_idx == -1) {
 410                 id = JPEGID{ stream_idx, output_pts, /*interpolated=*/true };
 411         } else {
 412                 id = create_jpegid_for_interpolated_fade(stream_idx, output_pts, secondary_stream_idx, secondary_input_pts);
 413         }
 414
 415         // Get the temporary OpenGL resources we need for doing the interpolation.
 416         InterpolatedFrameResources resources;
 417         {
 418                 unique_lock<mutex> lock(queue_lock);
 419                 if (interpolate_resources.empty()) {
 420                         fprintf(stderr, "WARNING: Too many interpolated frames already in transit; dropping one.\n");
 421                         return;
 422                 }
 423                 resources = interpolate_resources.front();
 424                 interpolate_resources.pop_front();
 425         }
 426
 427         QueuedFrame qf;
 428         qf.type = (secondary_stream_idx == -1) ? QueuedFrame::INTERPOLATED : QueuedFrame::FADED_INTERPOLATED;
 429         qf.output_pts = output_pts;
 430         qf.stream_idx = stream_idx;
 431         qf.resources = resources;
 432         qf.id = id;
 433         qf.display_func = move(display_func);
 434         qf.queue_spot_holder = move(queue_spot_holder);
 435
 436         check_error();
 437
 438         // Convert frame0 and frame1 to OpenGL textures.
 439         for (size_t frame_no = 0; frame_no < 2; ++frame_no) {
 440                 JPEGID jpeg_id;
 441                 jpeg_id.stream_idx = stream_idx;
 442                 jpeg_id.pts = frame_no == 1 ? input_second_pts : input_first_pts;
 443                 jpeg_id.interpolated = false;
 444                 bool did_decode;
 445                 shared_ptr<Frame> frame = decode_jpeg_with_cache(jpeg_id, DECODE_IF_NOT_IN_CACHE, &did_decode);
 446                 ycbcr_converter->prepare_chain_for_conversion(frame)->render_to_fbo(resources.input_fbos[frame_no], 1280, 720);
 447         }
 448
 449         glGenerateTextureMipmap(resources.input_tex);
 450         check_error();
 451         glGenerateTextureMipmap(resources.gray_tex);
 452         check_error();
 453
 454         // Compute the interpolated frame.
 455         qf.flow_tex = compute_flow->exec(resources.gray_tex, DISComputeFlow::FORWARD_AND_BACKWARD, DISComputeFlow::DO_NOT_RESIZE_FLOW);
 456         check_error();
 457
 458         if (secondary_stream_idx != -1) {
 459                 // Fade. First kick off the interpolation.
 460                 tie(qf.output_tex, ignore) = interpolate_no_split->exec(resources.input_tex, resources.gray_tex, qf.flow_tex, 1280, 720, alpha);
 461                 check_error();
 462
 463                 // Now decode the image we are fading against.
 464                 JPEGID jpeg_id;
 465                 jpeg_id.stream_idx = secondary_stream_idx;
 466                 jpeg_id.pts = secondary_input_pts;
 467                 jpeg_id.interpolated = false;
 468                 bool did_decode;
 469                 shared_ptr<Frame> frame2 = decode_jpeg_with_cache(jpeg_id, DECODE_IF_NOT_IN_CACHE, &did_decode);
 470
 471                 // Then fade against it, putting it into the fade Y' and CbCr textures.
 472                 ycbcr_semiplanar_converter->prepare_chain_for_fade_from_texture(qf.output_tex, frame2, fade_alpha)->render_to_fbo(resources.fade_fbo, 1280, 720);
 473
 474                 // Subsample and split Cb/Cr.
 475                 chroma_subsampler->subsample_chroma(resources.fade_cbcr_output_tex, 1280, 720, resources.cb_tex, resources.cr_tex);
 476
 477                 interpolate_no_split->release_texture(qf.output_tex);
 478         } else {
 479                 tie(qf.output_tex, qf.cbcr_tex) = interpolate->exec(resources.input_tex, resources.gray_tex, qf.flow_tex, 1280, 720, alpha);
 480                 check_error();
 481
 482                 // Subsample and split Cb/Cr.
 483                 chroma_subsampler->subsample_chroma(qf.cbcr_tex, 1280, 720, resources.cb_tex, resources.cr_tex);
 484         }
 485
 486         // We could have released qf.flow_tex here, but to make sure we don't cause a stall
 487         // when trying to reuse it for the next frame, we can just as well hold on to it
 488         // and release it only when the readback is done.
 489
 490         // Read it down (asynchronously) to the CPU.
 491         glPixelStorei(GL_PACK_ROW_LENGTH, 0);
 492         glBindBuffer(GL_PIXEL_PACK_BUFFER, resources.pbo);
 493         check_error();
 494         if (secondary_stream_idx != -1) {
 495                 glGetTextureImage(resources.fade_y_output_tex, 0, GL_RED, GL_UNSIGNED_BYTE, 1280 * 720 * 4, BUFFER_OFFSET(0));
 496         } else {
 497                 glGetTextureImage(qf.output_tex, 0, GL_RED, GL_UNSIGNED_BYTE, 1280 * 720 * 4, BUFFER_OFFSET(0));
 498         }
 499         check_error();
 500         glGetTextureImage(resources.cb_tex, 0, GL_RED, GL_UNSIGNED_BYTE, 1280 * 720 * 3, BUFFER_OFFSET(1280 * 720));
 501         check_error();
 502         glGetTextureImage(resources.cr_tex, 0, GL_RED, GL_UNSIGNED_BYTE, 1280 * 720 * 3 - 640 * 720, BUFFER_OFFSET(1280 * 720 + 640 * 720));
 503         check_error();
 504         glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
 505
 506         // Set a fence we can wait for to make sure the CPU sees the read.
 507         glMemoryBarrier(GL_CLIENT_MAPPED_BUFFER_BARRIER_BIT);
 508         check_error();
 509         qf.fence = RefCountedGLsync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
 510         check_error();
 511
 512         unique_lock<mutex> lock(queue_lock);
 513         frame_queue.push_back(move(qf));
 514         queue_changed.notify_all();
 515 }
 516
 517 void VideoStream::schedule_refresh_frame(steady_clock::time_point local_pts,
 518                                          int64_t output_pts, function<void()> &&display_func,
 519                                          QueueSpotHolder &&queue_spot_holder)
 520 {
 521         QueuedFrame qf;
 522         qf.type = QueuedFrame::REFRESH;
 523         qf.output_pts = output_pts;
 524         qf.display_func = move(display_func);
 525         qf.queue_spot_holder = move(queue_spot_holder);
 526
 527         unique_lock<mutex> lock(queue_lock);
 528         frame_queue.push_back(move(qf));
 529         queue_changed.notify_all();
 530 }
 531
 532 namespace {
 533
 534 shared_ptr<Frame> frame_from_pbo(void *contents, size_t width, size_t height)
 535 {
 536         size_t chroma_width = width / 2;
 537
 538         const uint8_t *y = (const uint8_t *)contents;
 539         const uint8_t *cb = (const uint8_t *)contents + width * height;
 540         const uint8_t *cr = (const uint8_t *)contents + width * height + chroma_width * height;
 541
 542         shared_ptr<Frame> frame(new Frame);
 543         frame->y.reset(new uint8_t[width * height]);
 544         frame->cb.reset(new uint8_t[chroma_width * height]);
 545         frame->cr.reset(new uint8_t[chroma_width * height]);
 546         for (unsigned yy = 0; yy < height; ++yy) {
 547                 memcpy(frame->y.get() + width * yy, y + width * yy, width);
 548                 memcpy(frame->cb.get() + chroma_width * yy, cb + chroma_width * yy, chroma_width);
 549                 memcpy(frame->cr.get() + chroma_width * yy, cr + chroma_width * yy, chroma_width);
 550         }
 551         frame->is_semiplanar = false;
 552         frame->width = width;
 553         frame->height = height;
 554         frame->chroma_subsampling_x = 2;
 555         frame->chroma_subsampling_y = 1;
 556         frame->pitch_y = width;
 557         frame->pitch_chroma = chroma_width;
 558         return frame;
 559 }
 560
 561 }  // namespace
 562
 563 void VideoStream::encode_thread_func()
 564 {
 565         pthread_setname_np(pthread_self(), "VideoStream");
 566         QSurface *surface = create_surface();
 567         QOpenGLContext *context = create_context(surface);
 568         bool ok = make_current(context, surface);
 569         if (!ok) {
 570                 fprintf(stderr, "Video stream couldn't get an OpenGL context\n");
 571                 exit(1);
 572         }
 573
 574         for ( ;; ) {
 575                 QueuedFrame qf;
 576                 {
 577                         unique_lock<mutex> lock(queue_lock);
 578
 579                         // Wait until we have a frame to play.
 580                         queue_changed.wait(lock, [this]{
 581                                 return !frame_queue.empty();
 582                         });
 583                         steady_clock::time_point frame_start = frame_queue.front().local_pts;
 584
 585                         // Now sleep until the frame is supposed to start (the usual case),
 586                         // _or_ clear_queue() happened.
 587                         bool aborted = queue_changed.wait_until(lock, frame_start, [this, frame_start]{
 588                                 return frame_queue.empty() || frame_queue.front().local_pts != frame_start;
 589                         });
 590                         if (aborted) {
 591                                 // clear_queue() happened, so don't play this frame after all.
 592                                 continue;
 593                         }
 594                         qf = move(frame_queue.front());
 595                         frame_queue.pop_front();
 596                 }
 597
 598                 if (qf.type == QueuedFrame::ORIGINAL) {
 599                         // Send the JPEG frame on, unchanged.
 600                         string jpeg = read_file(filename_for_frame(qf.stream_idx, qf.input_first_pts));
 601                         AVPacket pkt;
 602                         av_init_packet(&pkt);
 603                         pkt.stream_index = 0;
 604                         pkt.data = (uint8_t *)jpeg.data();
 605                         pkt.size = jpeg.size();
 606                         stream_mux->add_packet(pkt, qf.output_pts, qf.output_pts);
 607
 608                         last_frame.assign(&jpeg[0], &jpeg[0] + jpeg.size());
 609                 } else if (qf.type == QueuedFrame::FADED) {
 610                         glClientWaitSync(qf.fence.get(), /*flags=*/0, GL_TIMEOUT_IGNORED);
 611
 612                         shared_ptr<Frame> frame = frame_from_pbo(qf.resources.pbo_contents, 1280, 720);
 613
 614                         // Now JPEG encode it, and send it on to the stream.
 615                         vector<uint8_t> jpeg = encode_jpeg(frame->y.get(), frame->cb.get(), frame->cr.get(), 1280, 720);
 616
 617                         AVPacket pkt;
 618                         av_init_packet(&pkt);
 619                         pkt.stream_index = 0;
 620                         pkt.data = (uint8_t *)jpeg.data();
 621                         pkt.size = jpeg.size();
 622                         stream_mux->add_packet(pkt, qf.output_pts, qf.output_pts);
 623                         last_frame = move(jpeg);
 624
 625                         // Put the frame resources back.
 626                         unique_lock<mutex> lock(queue_lock);
 627                         interpolate_resources.push_back(qf.resources);
 628                 } else if (qf.type == QueuedFrame::INTERPOLATED || qf.type == QueuedFrame::FADED_INTERPOLATED) {
 629                         glClientWaitSync(qf.fence.get(), /*flags=*/0, GL_TIMEOUT_IGNORED);
 630
 631                         // Send a copy of the frame on to display.
 632                         shared_ptr<Frame> frame = frame_from_pbo(qf.resources.pbo_contents, 1280, 720);
 633                         JPEGFrameView::insert_interpolated_frame(qf.id, frame);
 634
 635                         // Now JPEG encode it, and send it on to the stream.
 636                         vector<uint8_t> jpeg = encode_jpeg(frame->y.get(), frame->cb.get(), frame->cr.get(), 1280, 720);
 637                         compute_flow->release_texture(qf.flow_tex);
 638                         if (qf.type != QueuedFrame::FADED_INTERPOLATED) {
 639                                 interpolate->release_texture(qf.output_tex);
 640                                 interpolate->release_texture(qf.cbcr_tex);
 641                         }
 642
 643                         AVPacket pkt;
 644                         av_init_packet(&pkt);
 645                         pkt.stream_index = 0;
 646                         pkt.data = (uint8_t *)jpeg.data();
 647                         pkt.size = jpeg.size();
 648                         stream_mux->add_packet(pkt, qf.output_pts, qf.output_pts);
 649                         last_frame = move(jpeg);
 650
 651                         // Put the frame resources back.
 652                         unique_lock<mutex> lock(queue_lock);
 653                         interpolate_resources.push_back(qf.resources);
 654                 } else if (qf.type == QueuedFrame::REFRESH) {
 655                         AVPacket pkt;
 656                         av_init_packet(&pkt);
 657                         pkt.stream_index = 0;
 658                         pkt.data = (uint8_t *)last_frame.data();
 659                         pkt.size = last_frame.size();
 660                         stream_mux->add_packet(pkt, qf.output_pts, qf.output_pts);
 661                 } else {
 662                         assert(false);
 663                 }
 664                 if (qf.display_func != nullptr) {
 665                         qf.display_func();
 666                 }
 667         }
 668 }
 669
 670 int VideoStream::write_packet2_thunk(void *opaque, uint8_t *buf, int buf_size, AVIODataMarkerType type, int64_t time)
 671 {
 672         VideoStream *video_stream = (VideoStream *)opaque;
 673         return video_stream->write_packet2(buf, buf_size, type, time);
 674 }
 675
 676 int VideoStream::write_packet2(uint8_t *buf, int buf_size, AVIODataMarkerType type, int64_t time)
 677 {
 678         if (type == AVIO_DATA_MARKER_SYNC_POINT || type == AVIO_DATA_MARKER_BOUNDARY_POINT) {
 679                 seen_sync_markers = true;
 680         } else if (type == AVIO_DATA_MARKER_UNKNOWN && !seen_sync_markers) {
 681                 // We don't know if this is a keyframe or not (the muxer could
 682                 // avoid marking it), so we just have to make the best of it.
 683                 type = AVIO_DATA_MARKER_SYNC_POINT;
 684         }
 685
 686         if (type == AVIO_DATA_MARKER_HEADER) {
 687                 stream_mux_header.append((char *)buf, buf_size);
 688                 global_httpd->set_header(stream_mux_header);
 689         } else {
 690                 global_httpd->add_data((char *)buf, buf_size, type == AVIO_DATA_MARKER_SYNC_POINT, time, AVRational{ AV_TIME_BASE, 1 });
 691         }
 692         return buf_size;
 693 }