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                 unique_ptr<InterpolatedFrameResources> resource(new InterpolatedFrameResources);
 185                 resource->owner = this;
 186                 resource->input_tex = input_tex[i];
 187                 resource->gray_tex = gray_tex[i];
 188                 resource->fade_y_output_tex = fade_y_output_tex[i];
 189                 resource->fade_cbcr_output_tex = fade_cbcr_output_tex[i];
 190                 resource->cb_tex = cb_tex[i];
 191                 resource->cr_tex = cr_tex[i];
 192                 glCreateFramebuffers(2, resource->input_fbos);
 193                 check_error();
 194                 glCreateFramebuffers(1, &resource->fade_fbo);
 195                 check_error();
 196
 197                 glNamedFramebufferTextureLayer(resource->input_fbos[0], GL_COLOR_ATTACHMENT0, input_tex[i], 0, 0);
 198                 check_error();
 199                 glNamedFramebufferTextureLayer(resource->input_fbos[0], GL_COLOR_ATTACHMENT1, gray_tex[i], 0, 0);
 200                 check_error();
 201                 glNamedFramebufferTextureLayer(resource->input_fbos[1], GL_COLOR_ATTACHMENT0, input_tex[i], 0, 1);
 202                 check_error();
 203                 glNamedFramebufferTextureLayer(resource->input_fbos[1], GL_COLOR_ATTACHMENT1, gray_tex[i], 0, 1);
 204                 check_error();
 205                 glNamedFramebufferTexture(resource->fade_fbo, GL_COLOR_ATTACHMENT0, fade_y_output_tex[i], 0);
 206                 check_error();
 207                 glNamedFramebufferTexture(resource->fade_fbo, GL_COLOR_ATTACHMENT1, fade_cbcr_output_tex[i], 0);
 208                 check_error();
 209
 210                 GLuint bufs[] = { GL_COLOR_ATTACHMENT0, GL_COLOR_ATTACHMENT1 };
 211                 glNamedFramebufferDrawBuffers(resource->input_fbos[0], 2, bufs);
 212                 check_error();
 213                 glNamedFramebufferDrawBuffers(resource->input_fbos[1], 2, bufs);
 214                 check_error();
 215                 glNamedFramebufferDrawBuffers(resource->fade_fbo, 2, bufs);
 216                 check_error();
 217
 218                 glCreateBuffers(1, &resource->pbo);
 219                 check_error();
 220                 glNamedBufferStorage(resource->pbo, width * height * 4, nullptr, GL_MAP_READ_BIT | GL_MAP_PERSISTENT_BIT);
 221                 check_error();
 222                 resource->pbo_contents = glMapNamedBufferRange(resource->pbo, 0, width * height * 4, GL_MAP_READ_BIT | GL_MAP_PERSISTENT_BIT);
 223                 interpolate_resources.push_back(move(resource));
 224         }
 225
 226         check_error();
 227
 228         OperatingPoint op;
 229         if (global_flags.interpolation_quality == 1) {
 230                 op = operating_point1;
 231         } else if (global_flags.interpolation_quality == 2) {
 232                 op = operating_point2;
 233         } else if (global_flags.interpolation_quality == 3) {
 234                 op = operating_point3;
 235         } else if (global_flags.interpolation_quality == 4) {
 236                 op = operating_point4;
 237         } else {
 238                 assert(false);
 239         }
 240
 241         compute_flow.reset(new DISComputeFlow(width, height, op));
 242         interpolate.reset(new Interpolate(op, /*split_ycbcr_output=*/true));
 243         interpolate_no_split.reset(new Interpolate(op, /*split_ycbcr_output=*/false));
 244         chroma_subsampler.reset(new ChromaSubsampler);
 245         check_error();
 246
 247         // The “last frame” is initially black.
 248         unique_ptr<uint8_t[]> y(new uint8_t[1280 * 720]);
 249         unique_ptr<uint8_t[]> cb_or_cr(new uint8_t[640 * 720]);
 250         memset(y.get(), 16, 1280 * 720);
 251         memset(cb_or_cr.get(), 128, 640 * 720);
 252         last_frame = encode_jpeg(y.get(), cb_or_cr.get(), cb_or_cr.get(), 1280, 720);
 253 }
 254
 255 VideoStream::~VideoStream() {}
 256
 257 void VideoStream::start()
 258 {
 259         AVFormatContext *avctx = avformat_alloc_context();
 260         avctx->oformat = av_guess_format("nut", nullptr, nullptr);
 261
 262         uint8_t *buf = (uint8_t *)av_malloc(MUX_BUFFER_SIZE);
 263         avctx->pb = avio_alloc_context(buf, MUX_BUFFER_SIZE, 1, this, nullptr, nullptr, nullptr);
 264         avctx->pb->write_data_type = &VideoStream::write_packet2_thunk;
 265         avctx->pb->ignore_boundary_point = 1;
 266
 267         Mux::Codec video_codec = Mux::CODEC_MJPEG;
 268
 269         avctx->flags = AVFMT_FLAG_CUSTOM_IO;
 270
 271         string video_extradata;
 272
 273         constexpr int width = 1280, height = 720;  // Doesn't matter for MJPEG.
 274         stream_mux.reset(new Mux(avctx, width, height, video_codec, video_extradata, /*audio_codec_parameters=*/nullptr, COARSE_TIMEBASE,
 275                 /*write_callback=*/nullptr, Mux::WRITE_FOREGROUND, {}));
 276
 277
 278         encode_thread = thread(&VideoStream::encode_thread_func, this);
 279 }
 280
 281 void VideoStream::stop()
 282 {
 283         encode_thread.join();
 284 }
 285
 286 void VideoStream::clear_queue()
 287 {
 288         deque<QueuedFrame> q;
 289
 290         {
 291                 unique_lock<mutex> lock(queue_lock);
 292                 q = move(frame_queue);
 293         }
 294
 295         // These are not RAII-ed, unfortunately, so we'll need to clean them ourselves.
 296         // Note that release_texture() is thread-safe.
 297         for (const QueuedFrame &qf : q) {
 298                 if (qf.type == QueuedFrame::INTERPOLATED ||
 299                     qf.type == QueuedFrame::FADED_INTERPOLATED) {
 300                         compute_flow->release_texture(qf.flow_tex);
 301                 }
 302                 if (qf.type == QueuedFrame::INTERPOLATED) {
 303                         interpolate->release_texture(qf.output_tex);
 304                         interpolate->release_texture(qf.cbcr_tex);
 305                 }
 306         }
 307
 308         // Destroy q outside the mutex, as that would be a double-lock.
 309 }
 310
 311 void VideoStream::schedule_original_frame(steady_clock::time_point local_pts,
 312                                           int64_t output_pts, function<void()> &&display_func,
 313                                           QueueSpotHolder &&queue_spot_holder,
 314                                           unsigned stream_idx, int64_t input_pts)
 315 {
 316         fprintf(stderr, "output_pts=%ld  original      input_pts=%ld\n", output_pts, input_pts);
 317
 318         // Preload the file from disk, so that the encoder thread does not get stalled.
 319         // TODO: Consider sending it through the queue instead.
 320         (void)read_file(filename_for_frame(stream_idx, input_pts));
 321
 322         QueuedFrame qf;
 323         qf.local_pts = local_pts;
 324         qf.type = QueuedFrame::ORIGINAL;
 325         qf.output_pts = output_pts;
 326         qf.stream_idx = stream_idx;
 327         qf.input_first_pts = input_pts;
 328         qf.display_func = move(display_func);
 329         qf.queue_spot_holder = move(queue_spot_holder);
 330
 331         unique_lock<mutex> lock(queue_lock);
 332         frame_queue.push_back(move(qf));
 333         queue_changed.notify_all();
 334 }
 335
 336 void VideoStream::schedule_faded_frame(steady_clock::time_point local_pts, int64_t output_pts,
 337                                        function<void()> &&display_func,
 338                                        QueueSpotHolder &&queue_spot_holder,
 339                                        unsigned stream_idx, int64_t input_pts, int secondary_stream_idx,
 340                                        int64_t secondary_input_pts, float fade_alpha)
 341 {
 342         fprintf(stderr, "output_pts=%ld  faded         input_pts=%ld,%ld  fade_alpha=%.2f\n", output_pts, input_pts, secondary_input_pts, fade_alpha);
 343
 344         // Get the temporary OpenGL resources we need for doing the fade.
 345         // (We share these with interpolated frames, which is slightly
 346         // overkill, but there's no need to waste resources on keeping
 347         // separate pools around.)
 348         BorrowedInterpolatedFrameResources resources;
 349         {
 350                 unique_lock<mutex> lock(queue_lock);
 351                 if (interpolate_resources.empty()) {
 352                         fprintf(stderr, "WARNING: Too many interpolated frames already in transit; dropping one.\n");
 353                         return;
 354                 }
 355                 resources = BorrowedInterpolatedFrameResources(interpolate_resources.front().release());
 356                 interpolate_resources.pop_front();
 357         }
 358
 359         bool did_decode;
 360
 361         JPEGID jpeg_id1;
 362         jpeg_id1.stream_idx = stream_idx;
 363         jpeg_id1.pts = input_pts;
 364         shared_ptr<Frame> frame1 = decode_jpeg_with_cache(jpeg_id1, DECODE_IF_NOT_IN_CACHE, &did_decode);
 365
 366         JPEGID jpeg_id2;
 367         jpeg_id2.stream_idx = secondary_stream_idx;
 368         jpeg_id2.pts = secondary_input_pts;
 369         shared_ptr<Frame> frame2 = decode_jpeg_with_cache(jpeg_id2, DECODE_IF_NOT_IN_CACHE, &did_decode);
 370
 371         ycbcr_semiplanar_converter->prepare_chain_for_fade(frame1, frame2, fade_alpha)->render_to_fbo(resources->fade_fbo, 1280, 720);
 372
 373         QueuedFrame qf;
 374         qf.local_pts = local_pts;
 375         qf.type = QueuedFrame::FADED;
 376         qf.output_pts = output_pts;
 377         qf.stream_idx = stream_idx;
 378         qf.input_first_pts = input_pts;
 379         qf.display_func = move(display_func);
 380         qf.queue_spot_holder = move(queue_spot_holder);
 381
 382         qf.secondary_stream_idx = secondary_stream_idx;
 383         qf.secondary_input_pts = secondary_input_pts;
 384
 385         // Subsample and split Cb/Cr.
 386         chroma_subsampler->subsample_chroma(resources->fade_cbcr_output_tex, 1280, 720, resources->cb_tex, resources->cr_tex);
 387
 388         // Read it down (asynchronously) to the CPU.
 389         glPixelStorei(GL_PACK_ROW_LENGTH, 0);
 390         glBindBuffer(GL_PIXEL_PACK_BUFFER, resources->pbo);
 391         check_error();
 392         glGetTextureImage(resources->fade_y_output_tex, 0, GL_RED, GL_UNSIGNED_BYTE, 1280 * 720 * 4, BUFFER_OFFSET(0));
 393         check_error();
 394         glGetTextureImage(resources->cb_tex, 0, GL_RED, GL_UNSIGNED_BYTE, 1280 * 720 * 3, BUFFER_OFFSET(1280 * 720));
 395         check_error();
 396         glGetTextureImage(resources->cr_tex, 0, GL_RED, GL_UNSIGNED_BYTE, 1280 * 720 * 3 - 640 * 720, BUFFER_OFFSET(1280 * 720 + 640 * 720));
 397         check_error();
 398         glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
 399
 400         // Set a fence we can wait for to make sure the CPU sees the read.
 401         glMemoryBarrier(GL_CLIENT_MAPPED_BUFFER_BARRIER_BIT);
 402         check_error();
 403         qf.fence = RefCountedGLsync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
 404         check_error();
 405         qf.resources = move(resources);
 406         qf.local_pts = local_pts;
 407
 408         unique_lock<mutex> lock(queue_lock);
 409         frame_queue.push_back(move(qf));
 410         queue_changed.notify_all();
 411 }
 412
 413 void VideoStream::schedule_interpolated_frame(steady_clock::time_point local_pts,
 414                                               int64_t output_pts, function<void(shared_ptr<Frame>)> &&display_func,
 415                                               QueueSpotHolder &&queue_spot_holder,
 416                                               unsigned stream_idx, int64_t input_first_pts,
 417                                               int64_t input_second_pts, float alpha,
 418                                               int secondary_stream_idx, int64_t secondary_input_pts,
 419                                               float fade_alpha)
 420 {
 421         if (secondary_stream_idx != -1) {
 422                 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);
 423         } else {
 424                 fprintf(stderr, "output_pts=%ld  interpolated  input_pts1=%ld input_pts2=%ld alpha=%.3f\n", output_pts, input_first_pts, input_second_pts, alpha);
 425         }
 426
 427         // Get the temporary OpenGL resources we need for doing the interpolation.
 428         BorrowedInterpolatedFrameResources resources;
 429         {
 430                 unique_lock<mutex> lock(queue_lock);
 431                 if (interpolate_resources.empty()) {
 432                         fprintf(stderr, "WARNING: Too many interpolated frames already in transit; dropping one.\n");
 433                         return;
 434                 }
 435                 resources = BorrowedInterpolatedFrameResources(interpolate_resources.front().release());
 436                 interpolate_resources.pop_front();
 437         }
 438
 439         QueuedFrame qf;
 440         qf.type = (secondary_stream_idx == -1) ? QueuedFrame::INTERPOLATED : QueuedFrame::FADED_INTERPOLATED;
 441         qf.output_pts = output_pts;
 442         qf.stream_idx = stream_idx;
 443         qf.display_decoded_func = move(display_func);
 444         qf.queue_spot_holder = move(queue_spot_holder);
 445         qf.local_pts = local_pts;
 446
 447         check_error();
 448
 449         // Convert frame0 and frame1 to OpenGL textures.
 450         for (size_t frame_no = 0; frame_no < 2; ++frame_no) {
 451                 JPEGID jpeg_id;
 452                 jpeg_id.stream_idx = stream_idx;
 453                 jpeg_id.pts = frame_no == 1 ? input_second_pts : input_first_pts;
 454                 bool did_decode;
 455                 shared_ptr<Frame> frame = decode_jpeg_with_cache(jpeg_id, DECODE_IF_NOT_IN_CACHE, &did_decode);
 456                 ycbcr_converter->prepare_chain_for_conversion(frame)->render_to_fbo(resources->input_fbos[frame_no], 1280, 720);
 457         }
 458
 459         glGenerateTextureMipmap(resources->input_tex);
 460         check_error();
 461         glGenerateTextureMipmap(resources->gray_tex);
 462         check_error();
 463
 464         // Compute the interpolated frame.
 465         qf.flow_tex = compute_flow->exec(resources->gray_tex, DISComputeFlow::FORWARD_AND_BACKWARD, DISComputeFlow::DO_NOT_RESIZE_FLOW);
 466         check_error();
 467
 468         if (secondary_stream_idx != -1) {
 469                 // Fade. First kick off the interpolation.
 470                 tie(qf.output_tex, ignore) = interpolate_no_split->exec(resources->input_tex, resources->gray_tex, qf.flow_tex, 1280, 720, alpha);
 471                 check_error();
 472
 473                 // Now decode the image we are fading against.
 474                 JPEGID jpeg_id;
 475                 jpeg_id.stream_idx = secondary_stream_idx;
 476                 jpeg_id.pts = secondary_input_pts;
 477                 bool did_decode;
 478                 shared_ptr<Frame> frame2 = decode_jpeg_with_cache(jpeg_id, DECODE_IF_NOT_IN_CACHE, &did_decode);
 479
 480                 // Then fade against it, putting it into the fade Y' and CbCr textures.
 481                 ycbcr_semiplanar_converter->prepare_chain_for_fade_from_texture(qf.output_tex, frame2, fade_alpha)->render_to_fbo(resources->fade_fbo, 1280, 720);
 482
 483                 // Subsample and split Cb/Cr.
 484                 chroma_subsampler->subsample_chroma(resources->fade_cbcr_output_tex, 1280, 720, resources->cb_tex, resources->cr_tex);
 485
 486                 interpolate_no_split->release_texture(qf.output_tex);
 487         } else {
 488                 tie(qf.output_tex, qf.cbcr_tex) = interpolate->exec(resources->input_tex, resources->gray_tex, qf.flow_tex, 1280, 720, alpha);
 489                 check_error();
 490
 491                 // Subsample and split Cb/Cr.
 492                 chroma_subsampler->subsample_chroma(qf.cbcr_tex, 1280, 720, resources->cb_tex, resources->cr_tex);
 493         }
 494
 495         // We could have released qf.flow_tex here, but to make sure we don't cause a stall
 496         // when trying to reuse it for the next frame, we can just as well hold on to it
 497         // and release it only when the readback is done.
 498
 499         // Read it down (asynchronously) to the CPU.
 500         glPixelStorei(GL_PACK_ROW_LENGTH, 0);
 501         glBindBuffer(GL_PIXEL_PACK_BUFFER, resources->pbo);
 502         check_error();
 503         if (secondary_stream_idx != -1) {
 504                 glGetTextureImage(resources->fade_y_output_tex, 0, GL_RED, GL_UNSIGNED_BYTE, 1280 * 720 * 4, BUFFER_OFFSET(0));
 505         } else {
 506                 glGetTextureImage(qf.output_tex, 0, GL_RED, GL_UNSIGNED_BYTE, 1280 * 720 * 4, BUFFER_OFFSET(0));
 507         }
 508         check_error();
 509         glGetTextureImage(resources->cb_tex, 0, GL_RED, GL_UNSIGNED_BYTE, 1280 * 720 * 3, BUFFER_OFFSET(1280 * 720));
 510         check_error();
 511         glGetTextureImage(resources->cr_tex, 0, GL_RED, GL_UNSIGNED_BYTE, 1280 * 720 * 3 - 640 * 720, BUFFER_OFFSET(1280 * 720 + 640 * 720));
 512         check_error();
 513         glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
 514
 515         // Set a fence we can wait for to make sure the CPU sees the read.
 516         glMemoryBarrier(GL_CLIENT_MAPPED_BUFFER_BARRIER_BIT);
 517         check_error();
 518         qf.fence = RefCountedGLsync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
 519         check_error();
 520         qf.resources = move(resources);
 521
 522         unique_lock<mutex> lock(queue_lock);
 523         frame_queue.push_back(move(qf));
 524         queue_changed.notify_all();
 525 }
 526
 527 void VideoStream::schedule_refresh_frame(steady_clock::time_point local_pts,
 528                                          int64_t output_pts, function<void()> &&display_func,
 529                                          QueueSpotHolder &&queue_spot_holder)
 530 {
 531         QueuedFrame qf;
 532         qf.type = QueuedFrame::REFRESH;
 533         qf.output_pts = output_pts;
 534         qf.display_func = move(display_func);
 535         qf.queue_spot_holder = move(queue_spot_holder);
 536
 537         unique_lock<mutex> lock(queue_lock);
 538         frame_queue.push_back(move(qf));
 539         queue_changed.notify_all();
 540 }
 541
 542 namespace {
 543
 544 shared_ptr<Frame> frame_from_pbo(void *contents, size_t width, size_t height)
 545 {
 546         size_t chroma_width = width / 2;
 547
 548         const uint8_t *y = (const uint8_t *)contents;
 549         const uint8_t *cb = (const uint8_t *)contents + width * height;
 550         const uint8_t *cr = (const uint8_t *)contents + width * height + chroma_width * height;
 551
 552         shared_ptr<Frame> frame(new Frame);
 553         frame->y.reset(new uint8_t[width * height]);
 554         frame->cb.reset(new uint8_t[chroma_width * height]);
 555         frame->cr.reset(new uint8_t[chroma_width * height]);
 556         for (unsigned yy = 0; yy < height; ++yy) {
 557                 memcpy(frame->y.get() + width * yy, y + width * yy, width);
 558                 memcpy(frame->cb.get() + chroma_width * yy, cb + chroma_width * yy, chroma_width);
 559                 memcpy(frame->cr.get() + chroma_width * yy, cr + chroma_width * yy, chroma_width);
 560         }
 561         frame->is_semiplanar = false;
 562         frame->width = width;
 563         frame->height = height;
 564         frame->chroma_subsampling_x = 2;
 565         frame->chroma_subsampling_y = 1;
 566         frame->pitch_y = width;
 567         frame->pitch_chroma = chroma_width;
 568         return frame;
 569 }
 570
 571 }  // namespace
 572
 573 void VideoStream::encode_thread_func()
 574 {
 575         pthread_setname_np(pthread_self(), "VideoStream");
 576         QSurface *surface = create_surface();
 577         QOpenGLContext *context = create_context(surface);
 578         bool ok = make_current(context, surface);
 579         if (!ok) {
 580                 fprintf(stderr, "Video stream couldn't get an OpenGL context\n");
 581                 exit(1);
 582         }
 583
 584         for ( ;; ) {
 585                 QueuedFrame qf;
 586                 {
 587                         unique_lock<mutex> lock(queue_lock);
 588
 589                         // Wait until we have a frame to play.
 590                         queue_changed.wait(lock, [this]{
 591                                 return !frame_queue.empty();
 592                         });
 593                         steady_clock::time_point frame_start = frame_queue.front().local_pts;
 594
 595                         // Now sleep until the frame is supposed to start (the usual case),
 596                         // _or_ clear_queue() happened.
 597                         bool aborted = queue_changed.wait_until(lock, frame_start, [this, frame_start]{
 598                                 return frame_queue.empty() || frame_queue.front().local_pts != frame_start;
 599                         });
 600                         if (aborted) {
 601                                 // clear_queue() happened, so don't play this frame after all.
 602                                 continue;
 603                         }
 604                         qf = move(frame_queue.front());
 605                         frame_queue.pop_front();
 606                 }
 607
 608                 if (qf.type == QueuedFrame::ORIGINAL) {
 609                         // Send the JPEG frame on, unchanged.
 610                         string jpeg = read_file(filename_for_frame(qf.stream_idx, qf.input_first_pts));
 611                         AVPacket pkt;
 612                         av_init_packet(&pkt);
 613                         pkt.stream_index = 0;
 614                         pkt.data = (uint8_t *)jpeg.data();
 615                         pkt.size = jpeg.size();
 616                         stream_mux->add_packet(pkt, qf.output_pts, qf.output_pts);
 617
 618                         last_frame.assign(&jpeg[0], &jpeg[0] + jpeg.size());
 619                 } else if (qf.type == QueuedFrame::FADED) {
 620                         glClientWaitSync(qf.fence.get(), /*flags=*/0, GL_TIMEOUT_IGNORED);
 621
 622                         shared_ptr<Frame> frame = frame_from_pbo(qf.resources->pbo_contents, 1280, 720);
 623
 624                         // Now JPEG encode it, and send it on to the stream.
 625                         vector<uint8_t> jpeg = encode_jpeg(frame->y.get(), frame->cb.get(), frame->cr.get(), 1280, 720);
 626
 627                         AVPacket pkt;
 628                         av_init_packet(&pkt);
 629                         pkt.stream_index = 0;
 630                         pkt.data = (uint8_t *)jpeg.data();
 631                         pkt.size = jpeg.size();
 632                         stream_mux->add_packet(pkt, qf.output_pts, qf.output_pts);
 633                         last_frame = move(jpeg);
 634                 } else if (qf.type == QueuedFrame::INTERPOLATED || qf.type == QueuedFrame::FADED_INTERPOLATED) {
 635                         glClientWaitSync(qf.fence.get(), /*flags=*/0, GL_TIMEOUT_IGNORED);
 636
 637                         // Send it on to display.
 638                         shared_ptr<Frame> frame = frame_from_pbo(qf.resources->pbo_contents, 1280, 720);
 639                         if (qf.display_decoded_func != nullptr) {
 640                                 qf.display_decoded_func(frame);
 641                         }
 642
 643                         // Now JPEG encode it, and send it on to the stream.
 644                         vector<uint8_t> jpeg = encode_jpeg(frame->y.get(), frame->cb.get(), frame->cr.get(), 1280, 720);
 645                         compute_flow->release_texture(qf.flow_tex);
 646                         if (qf.type != QueuedFrame::FADED_INTERPOLATED) {
 647                                 interpolate->release_texture(qf.output_tex);
 648                                 interpolate->release_texture(qf.cbcr_tex);
 649                         }
 650
 651                         AVPacket pkt;
 652                         av_init_packet(&pkt);
 653                         pkt.stream_index = 0;
 654                         pkt.data = (uint8_t *)jpeg.data();
 655                         pkt.size = jpeg.size();
 656                         stream_mux->add_packet(pkt, qf.output_pts, qf.output_pts);
 657                         last_frame = move(jpeg);
 658                 } else if (qf.type == QueuedFrame::REFRESH) {
 659                         AVPacket pkt;
 660                         av_init_packet(&pkt);
 661                         pkt.stream_index = 0;
 662                         pkt.data = (uint8_t *)last_frame.data();
 663                         pkt.size = last_frame.size();
 664                         stream_mux->add_packet(pkt, qf.output_pts, qf.output_pts);
 665                 } else {
 666                         assert(false);
 667                 }
 668                 if (qf.display_func != nullptr) {
 669                         qf.display_func();
 670                 }
 671         }
 672 }
 673
 674 int VideoStream::write_packet2_thunk(void *opaque, uint8_t *buf, int buf_size, AVIODataMarkerType type, int64_t time)
 675 {
 676         VideoStream *video_stream = (VideoStream *)opaque;
 677         return video_stream->write_packet2(buf, buf_size, type, time);
 678 }
 679
 680 int VideoStream::write_packet2(uint8_t *buf, int buf_size, AVIODataMarkerType type, int64_t time)
 681 {
 682         if (type == AVIO_DATA_MARKER_SYNC_POINT || type == AVIO_DATA_MARKER_BOUNDARY_POINT) {
 683                 seen_sync_markers = true;
 684         } else if (type == AVIO_DATA_MARKER_UNKNOWN && !seen_sync_markers) {
 685                 // We don't know if this is a keyframe or not (the muxer could
 686                 // avoid marking it), so we just have to make the best of it.
 687                 type = AVIO_DATA_MARKER_SYNC_POINT;
 688         }
 689
 690         if (type == AVIO_DATA_MARKER_HEADER) {
 691                 stream_mux_header.append((char *)buf, buf_size);
 692                 global_httpd->set_header(stream_mux_header);
 693         } else {
 694                 global_httpd->add_data((char *)buf, buf_size, type == AVIO_DATA_MARKER_SYNC_POINT, time, AVRational{ AV_TIME_BASE, 1 });
 695         }
 696         return buf_size;
 697 }