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