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         JSAMPROW yptr[8], cbptr[8], crptr[8];
 111         JSAMPARRAY data[3] = { yptr, cbptr, crptr };
 112         for (unsigned y = 0; y < height; y += 8) {
 113                 for (unsigned yy = 0; yy < 8; ++yy) {
 114                         yptr[yy] = const_cast<JSAMPROW>(&y_data[(y + yy) * width]);
 115                         cbptr[yy] = const_cast<JSAMPROW>(&cb_data[(y + yy) * width / 2]);
 116                         crptr[yy] = const_cast<JSAMPROW>(&cr_data[(y + yy) * width / 2]);
 117                 }
 118
 119                 jpeg_write_raw_data(&cinfo, data, /*num_lines=*/8);
 120         }
 121
 122         jpeg_finish_compress(&cinfo);
 123         jpeg_destroy_compress(&cinfo);
 124
 125         return move(dest.dest);
 126 }
 127
 128 VideoStream::VideoStream()
 129 {
 130         ycbcr_converter.reset(new YCbCrConverter(YCbCrConverter::OUTPUT_TO_DUAL_YCBCR, /*resource_pool=*/nullptr));
 131         ycbcr_semiplanar_converter.reset(new YCbCrConverter(YCbCrConverter::OUTPUT_TO_SEMIPLANAR, /*resource_pool=*/nullptr));
 132
 133         GLuint input_tex[num_interpolate_slots], gray_tex[num_interpolate_slots];
 134         GLuint fade_y_output_tex[num_interpolate_slots], fade_cbcr_output_tex[num_interpolate_slots];
 135         GLuint cb_tex[num_interpolate_slots], cr_tex[num_interpolate_slots];
 136
 137         glCreateTextures(GL_TEXTURE_2D_ARRAY, num_interpolate_slots, input_tex);
 138         glCreateTextures(GL_TEXTURE_2D_ARRAY, num_interpolate_slots, gray_tex);
 139         glCreateTextures(GL_TEXTURE_2D, num_interpolate_slots, fade_y_output_tex);
 140         glCreateTextures(GL_TEXTURE_2D, num_interpolate_slots, fade_cbcr_output_tex);
 141         glCreateTextures(GL_TEXTURE_2D, num_interpolate_slots, cb_tex);
 142         glCreateTextures(GL_TEXTURE_2D, num_interpolate_slots, cr_tex);
 143         check_error();
 144
 145         constexpr size_t width = 1280, height = 720;  // FIXME: adjustable width, height
 146         int levels = find_num_levels(width, height);
 147         for (size_t i = 0; i < num_interpolate_slots; ++i) {
 148                 glTextureStorage3D(input_tex[i], levels, GL_RGBA8, width, height, 2);
 149                 check_error();
 150                 glTextureStorage3D(gray_tex[i], levels, GL_R8, width, height, 2);
 151                 check_error();
 152                 glTextureStorage2D(fade_y_output_tex[i], 1, GL_R8, width, height);
 153                 check_error();
 154                 glTextureStorage2D(fade_cbcr_output_tex[i], 1, GL_RG8, width, height);
 155                 check_error();
 156                 glTextureStorage2D(cb_tex[i], 1, GL_R8, width / 2, height);
 157                 check_error();
 158                 glTextureStorage2D(cr_tex[i], 1, GL_R8, width / 2, height);
 159                 check_error();
 160
 161                 unique_ptr<InterpolatedFrameResources> resource(new InterpolatedFrameResources);
 162                 resource->owner = this;
 163                 resource->input_tex = input_tex[i];
 164                 resource->gray_tex = gray_tex[i];
 165                 resource->fade_y_output_tex = fade_y_output_tex[i];
 166                 resource->fade_cbcr_output_tex = fade_cbcr_output_tex[i];
 167                 resource->cb_tex = cb_tex[i];
 168                 resource->cr_tex = cr_tex[i];
 169                 glCreateFramebuffers(2, resource->input_fbos);
 170                 check_error();
 171                 glCreateFramebuffers(1, &resource->fade_fbo);
 172                 check_error();
 173
 174                 glNamedFramebufferTextureLayer(resource->input_fbos[0], GL_COLOR_ATTACHMENT0, input_tex[i], 0, 0);
 175                 check_error();
 176                 glNamedFramebufferTextureLayer(resource->input_fbos[0], GL_COLOR_ATTACHMENT1, gray_tex[i], 0, 0);
 177                 check_error();
 178                 glNamedFramebufferTextureLayer(resource->input_fbos[1], GL_COLOR_ATTACHMENT0, input_tex[i], 0, 1);
 179                 check_error();
 180                 glNamedFramebufferTextureLayer(resource->input_fbos[1], GL_COLOR_ATTACHMENT1, gray_tex[i], 0, 1);
 181                 check_error();
 182                 glNamedFramebufferTexture(resource->fade_fbo, GL_COLOR_ATTACHMENT0, fade_y_output_tex[i], 0);
 183                 check_error();
 184                 glNamedFramebufferTexture(resource->fade_fbo, GL_COLOR_ATTACHMENT1, fade_cbcr_output_tex[i], 0);
 185                 check_error();
 186
 187                 GLuint bufs[] = { GL_COLOR_ATTACHMENT0, GL_COLOR_ATTACHMENT1 };
 188                 glNamedFramebufferDrawBuffers(resource->input_fbos[0], 2, bufs);
 189                 check_error();
 190                 glNamedFramebufferDrawBuffers(resource->input_fbos[1], 2, bufs);
 191                 check_error();
 192                 glNamedFramebufferDrawBuffers(resource->fade_fbo, 2, bufs);
 193                 check_error();
 194
 195                 glCreateBuffers(1, &resource->pbo);
 196                 check_error();
 197                 glNamedBufferStorage(resource->pbo, width * height * 4, nullptr, GL_MAP_READ_BIT | GL_MAP_PERSISTENT_BIT);
 198                 check_error();
 199                 resource->pbo_contents = glMapNamedBufferRange(resource->pbo, 0, width * height * 4, GL_MAP_READ_BIT | GL_MAP_PERSISTENT_BIT);
 200                 interpolate_resources.push_back(move(resource));
 201         }
 202
 203         check_error();
 204
 205         OperatingPoint op;
 206         if (global_flags.interpolation_quality == 1) {
 207                 op = operating_point1;
 208         } else if (global_flags.interpolation_quality == 2) {
 209                 op = operating_point2;
 210         } else if (global_flags.interpolation_quality == 3) {
 211                 op = operating_point3;
 212         } else if (global_flags.interpolation_quality == 4) {
 213                 op = operating_point4;
 214         } else {
 215                 assert(false);
 216         }
 217
 218         compute_flow.reset(new DISComputeFlow(width, height, op));
 219         interpolate.reset(new Interpolate(op, /*split_ycbcr_output=*/true));
 220         interpolate_no_split.reset(new Interpolate(op, /*split_ycbcr_output=*/false));
 221         chroma_subsampler.reset(new ChromaSubsampler);
 222         check_error();
 223
 224         // The “last frame” is initially black.
 225         unique_ptr<uint8_t[]> y(new uint8_t[1280 * 720]);
 226         unique_ptr<uint8_t[]> cb_or_cr(new uint8_t[640 * 720]);
 227         memset(y.get(), 16, 1280 * 720);
 228         memset(cb_or_cr.get(), 128, 640 * 720);
 229         last_frame = encode_jpeg(y.get(), cb_or_cr.get(), cb_or_cr.get(), 1280, 720);
 230 }
 231
 232 VideoStream::~VideoStream() {}
 233
 234 void VideoStream::start()
 235 {
 236         AVFormatContext *avctx = avformat_alloc_context();
 237         avctx->oformat = av_guess_format("nut", nullptr, nullptr);
 238
 239         uint8_t *buf = (uint8_t *)av_malloc(MUX_BUFFER_SIZE);
 240         avctx->pb = avio_alloc_context(buf, MUX_BUFFER_SIZE, 1, this, nullptr, nullptr, nullptr);
 241         avctx->pb->write_data_type = &VideoStream::write_packet2_thunk;
 242         avctx->pb->ignore_boundary_point = 1;
 243
 244         Mux::Codec video_codec = Mux::CODEC_MJPEG;
 245
 246         avctx->flags = AVFMT_FLAG_CUSTOM_IO;
 247
 248         string video_extradata;
 249
 250         constexpr int width = 1280, height = 720;  // Doesn't matter for MJPEG.
 251         stream_mux.reset(new Mux(avctx, width, height, video_codec, video_extradata, /*audio_codec_parameters=*/nullptr,
 252                 AVCOL_SPC_BT709, Mux::WITHOUT_AUDIO,
 253                 COARSE_TIMEBASE, /*write_callback=*/nullptr, Mux::WRITE_FOREGROUND, {}));
 254
 255
 256         encode_thread = thread(&VideoStream::encode_thread_func, this);
 257 }
 258
 259 void VideoStream::stop()
 260 {
 261         encode_thread.join();
 262 }
 263
 264 void VideoStream::clear_queue()
 265 {
 266         deque<QueuedFrame> q;
 267
 268         {
 269                 unique_lock<mutex> lock(queue_lock);
 270                 q = move(frame_queue);
 271         }
 272
 273         // These are not RAII-ed, unfortunately, so we'll need to clean them ourselves.
 274         // Note that release_texture() is thread-safe.
 275         for (const QueuedFrame &qf : q) {
 276                 if (qf.type == QueuedFrame::INTERPOLATED ||
 277                     qf.type == QueuedFrame::FADED_INTERPOLATED) {
 278                         compute_flow->release_texture(qf.flow_tex);
 279                 }
 280                 if (qf.type == QueuedFrame::INTERPOLATED) {
 281                         interpolate->release_texture(qf.output_tex);
 282                         interpolate->release_texture(qf.cbcr_tex);
 283                 }
 284         }
 285
 286         // Destroy q outside the mutex, as that would be a double-lock.
 287 }
 288
 289 void VideoStream::schedule_original_frame(steady_clock::time_point local_pts,
 290                                           int64_t output_pts, function<void()> &&display_func,
 291                                           QueueSpotHolder &&queue_spot_holder,
 292                                           FrameOnDisk frame)
 293 {
 294         fprintf(stderr, "output_pts=%ld  original      input_pts=%ld\n", output_pts, frame.pts);
 295
 296         // Preload the file from disk, so that the encoder thread does not get stalled.
 297         // TODO: Consider sending it through the queue instead.
 298         (void)frame_reader.read_frame(frame);
 299
 300         QueuedFrame qf;
 301         qf.local_pts = local_pts;
 302         qf.type = QueuedFrame::ORIGINAL;
 303         qf.output_pts = output_pts;
 304         qf.frame1 = frame;
 305         qf.display_func = move(display_func);
 306         qf.queue_spot_holder = move(queue_spot_holder);
 307
 308         unique_lock<mutex> lock(queue_lock);
 309         frame_queue.push_back(move(qf));
 310         queue_changed.notify_all();
 311 }
 312
 313 void VideoStream::schedule_faded_frame(steady_clock::time_point local_pts, int64_t output_pts,
 314                                        function<void()> &&display_func,
 315                                        QueueSpotHolder &&queue_spot_holder,
 316                                        FrameOnDisk frame1_spec, FrameOnDisk frame2_spec,
 317                                        float fade_alpha)
 318 {
 319         fprintf(stderr, "output_pts=%ld  faded         input_pts=%ld,%ld  fade_alpha=%.2f\n", output_pts, frame1_spec.pts, frame2_spec.pts, fade_alpha);
 320
 321         // Get the temporary OpenGL resources we need for doing the fade.
 322         // (We share these with interpolated frames, which is slightly
 323         // overkill, but there's no need to waste resources on keeping
 324         // separate pools around.)
 325         BorrowedInterpolatedFrameResources resources;
 326         {
 327                 unique_lock<mutex> lock(queue_lock);
 328                 if (interpolate_resources.empty()) {
 329                         fprintf(stderr, "WARNING: Too many interpolated frames already in transit; dropping one.\n");
 330                         return;
 331                 }
 332                 resources = BorrowedInterpolatedFrameResources(interpolate_resources.front().release());
 333                 interpolate_resources.pop_front();
 334         }
 335
 336         bool did_decode;
 337
 338         shared_ptr<Frame> frame1 = decode_jpeg_with_cache(frame1_spec, DECODE_IF_NOT_IN_CACHE, &frame_reader, &did_decode);
 339         shared_ptr<Frame> frame2 = decode_jpeg_with_cache(frame2_spec, DECODE_IF_NOT_IN_CACHE, &frame_reader, &did_decode);
 340
 341         ycbcr_semiplanar_converter->prepare_chain_for_fade(frame1, frame2, fade_alpha)->render_to_fbo(resources->fade_fbo, 1280, 720);
 342
 343         QueuedFrame qf;
 344         qf.local_pts = local_pts;
 345         qf.type = QueuedFrame::FADED;
 346         qf.output_pts = output_pts;
 347         qf.frame1 = frame1_spec;
 348         qf.display_func = move(display_func);
 349         qf.queue_spot_holder = move(queue_spot_holder);
 350
 351         qf.secondary_frame = frame2_spec;
 352
 353         // Subsample and split Cb/Cr.
 354         chroma_subsampler->subsample_chroma(resources->fade_cbcr_output_tex, 1280, 720, resources->cb_tex, resources->cr_tex);
 355
 356         // Read it down (asynchronously) to the CPU.
 357         glPixelStorei(GL_PACK_ROW_LENGTH, 0);
 358         glBindBuffer(GL_PIXEL_PACK_BUFFER, resources->pbo);
 359         check_error();
 360         glGetTextureImage(resources->fade_y_output_tex, 0, GL_RED, GL_UNSIGNED_BYTE, 1280 * 720 * 4, BUFFER_OFFSET(0));
 361         check_error();
 362         glGetTextureImage(resources->cb_tex, 0, GL_RED, GL_UNSIGNED_BYTE, 1280 * 720 * 3, BUFFER_OFFSET(1280 * 720));
 363         check_error();
 364         glGetTextureImage(resources->cr_tex, 0, GL_RED, GL_UNSIGNED_BYTE, 1280 * 720 * 3 - 640 * 720, BUFFER_OFFSET(1280 * 720 + 640 * 720));
 365         check_error();
 366         glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
 367
 368         // Set a fence we can wait for to make sure the CPU sees the read.
 369         glMemoryBarrier(GL_CLIENT_MAPPED_BUFFER_BARRIER_BIT);
 370         check_error();
 371         qf.fence = RefCountedGLsync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
 372         check_error();
 373         qf.resources = move(resources);
 374         qf.local_pts = local_pts;
 375
 376         unique_lock<mutex> lock(queue_lock);
 377         frame_queue.push_back(move(qf));
 378         queue_changed.notify_all();
 379 }
 380
 381 void VideoStream::schedule_interpolated_frame(steady_clock::time_point local_pts,
 382                                               int64_t output_pts, function<void(shared_ptr<Frame>)> &&display_func,
 383                                               QueueSpotHolder &&queue_spot_holder,
 384                                               FrameOnDisk frame1, FrameOnDisk frame2,
 385                                               float alpha, FrameOnDisk secondary_frame, float fade_alpha)
 386 {
 387         if (secondary_frame.pts != -1) {
 388                 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);
 389         } else {
 390                 fprintf(stderr, "output_pts=%ld  interpolated  input_pts1=%ld input_pts2=%ld alpha=%.3f\n", output_pts, frame1.pts, frame2.pts, alpha);
 391         }
 392
 393         // Get the temporary OpenGL resources we need for doing the interpolation.
 394         BorrowedInterpolatedFrameResources resources;
 395         {
 396                 unique_lock<mutex> lock(queue_lock);
 397                 if (interpolate_resources.empty()) {
 398                         fprintf(stderr, "WARNING: Too many interpolated frames already in transit; dropping one.\n");
 399                         return;
 400                 }
 401                 resources = BorrowedInterpolatedFrameResources(interpolate_resources.front().release());
 402                 interpolate_resources.pop_front();
 403         }
 404
 405         QueuedFrame qf;
 406         qf.type = (secondary_frame.pts == -1) ? QueuedFrame::INTERPOLATED : QueuedFrame::FADED_INTERPOLATED;
 407         qf.output_pts = output_pts;
 408         qf.display_decoded_func = move(display_func);
 409         qf.queue_spot_holder = move(queue_spot_holder);
 410         qf.local_pts = local_pts;
 411
 412         check_error();
 413
 414         // Convert frame0 and frame1 to OpenGL textures.
 415         for (size_t frame_no = 0; frame_no < 2; ++frame_no) {
 416                 FrameOnDisk frame_spec = frame_no == 1 ? frame2 : frame1;
 417                 bool did_decode;
 418                 shared_ptr<Frame> frame = decode_jpeg_with_cache(frame_spec, DECODE_IF_NOT_IN_CACHE, &frame_reader, &did_decode);
 419                 ycbcr_converter->prepare_chain_for_conversion(frame)->render_to_fbo(resources->input_fbos[frame_no], 1280, 720);
 420         }
 421
 422         glGenerateTextureMipmap(resources->input_tex);
 423         check_error();
 424         glGenerateTextureMipmap(resources->gray_tex);
 425         check_error();
 426
 427         // Compute the interpolated frame.
 428         qf.flow_tex = compute_flow->exec(resources->gray_tex, DISComputeFlow::FORWARD_AND_BACKWARD, DISComputeFlow::DO_NOT_RESIZE_FLOW);
 429         check_error();
 430
 431         if (secondary_frame.pts != -1) {
 432                 // Fade. First kick off the interpolation.
 433                 tie(qf.output_tex, ignore) = interpolate_no_split->exec(resources->input_tex, resources->gray_tex, qf.flow_tex, 1280, 720, alpha);
 434                 check_error();
 435
 436                 // Now decode the image we are fading against.
 437                 bool did_decode;
 438                 shared_ptr<Frame> frame2 = decode_jpeg_with_cache(secondary_frame, DECODE_IF_NOT_IN_CACHE, &frame_reader, &did_decode);
 439
 440                 // Then fade against it, putting it into the fade Y' and CbCr textures.
 441                 ycbcr_semiplanar_converter->prepare_chain_for_fade_from_texture(qf.output_tex, frame2, fade_alpha)->render_to_fbo(resources->fade_fbo, 1280, 720);
 442
 443                 // Subsample and split Cb/Cr.
 444                 chroma_subsampler->subsample_chroma(resources->fade_cbcr_output_tex, 1280, 720, resources->cb_tex, resources->cr_tex);
 445
 446                 interpolate_no_split->release_texture(qf.output_tex);
 447         } else {
 448                 tie(qf.output_tex, qf.cbcr_tex) = interpolate->exec(resources->input_tex, resources->gray_tex, qf.flow_tex, 1280, 720, alpha);
 449                 check_error();
 450
 451                 // Subsample and split Cb/Cr.
 452                 chroma_subsampler->subsample_chroma(qf.cbcr_tex, 1280, 720, resources->cb_tex, resources->cr_tex);
 453         }
 454
 455         // We could have released qf.flow_tex here, but to make sure we don't cause a stall
 456         // when trying to reuse it for the next frame, we can just as well hold on to it
 457         // and release it only when the readback is done.
 458
 459         // Read it down (asynchronously) to the CPU.
 460         glPixelStorei(GL_PACK_ROW_LENGTH, 0);
 461         glBindBuffer(GL_PIXEL_PACK_BUFFER, resources->pbo);
 462         check_error();
 463         if (secondary_frame.pts != -1) {
 464                 glGetTextureImage(resources->fade_y_output_tex, 0, GL_RED, GL_UNSIGNED_BYTE, 1280 * 720 * 4, BUFFER_OFFSET(0));
 465         } else {
 466                 glGetTextureImage(qf.output_tex, 0, GL_RED, GL_UNSIGNED_BYTE, 1280 * 720 * 4, BUFFER_OFFSET(0));
 467         }
 468         check_error();
 469         glGetTextureImage(resources->cb_tex, 0, GL_RED, GL_UNSIGNED_BYTE, 1280 * 720 * 3, BUFFER_OFFSET(1280 * 720));
 470         check_error();
 471         glGetTextureImage(resources->cr_tex, 0, GL_RED, GL_UNSIGNED_BYTE, 1280 * 720 * 3 - 640 * 720, BUFFER_OFFSET(1280 * 720 + 640 * 720));
 472         check_error();
 473         glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
 474
 475         // Set a fence we can wait for to make sure the CPU sees the read.
 476         glMemoryBarrier(GL_CLIENT_MAPPED_BUFFER_BARRIER_BIT);
 477         check_error();
 478         qf.fence = RefCountedGLsync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
 479         check_error();
 480         qf.resources = move(resources);
 481
 482         unique_lock<mutex> lock(queue_lock);
 483         frame_queue.push_back(move(qf));
 484         queue_changed.notify_all();
 485 }
 486
 487 void VideoStream::schedule_refresh_frame(steady_clock::time_point local_pts,
 488                                          int64_t output_pts, function<void()> &&display_func,
 489                                          QueueSpotHolder &&queue_spot_holder)
 490 {
 491         QueuedFrame qf;
 492         qf.type = QueuedFrame::REFRESH;
 493         qf.output_pts = output_pts;
 494         qf.display_func = move(display_func);
 495         qf.queue_spot_holder = move(queue_spot_holder);
 496
 497         unique_lock<mutex> lock(queue_lock);
 498         frame_queue.push_back(move(qf));
 499         queue_changed.notify_all();
 500 }
 501
 502 namespace {
 503
 504 shared_ptr<Frame> frame_from_pbo(void *contents, size_t width, size_t height)
 505 {
 506         size_t chroma_width = width / 2;
 507
 508         const uint8_t *y = (const uint8_t *)contents;
 509         const uint8_t *cb = (const uint8_t *)contents + width * height;
 510         const uint8_t *cr = (const uint8_t *)contents + width * height + chroma_width * height;
 511
 512         shared_ptr<Frame> frame(new Frame);
 513         frame->y.reset(new uint8_t[width * height]);
 514         frame->cb.reset(new uint8_t[chroma_width * height]);
 515         frame->cr.reset(new uint8_t[chroma_width * height]);
 516         for (unsigned yy = 0; yy < height; ++yy) {
 517                 memcpy(frame->y.get() + width * yy, y + width * yy, width);
 518                 memcpy(frame->cb.get() + chroma_width * yy, cb + chroma_width * yy, chroma_width);
 519                 memcpy(frame->cr.get() + chroma_width * yy, cr + chroma_width * yy, chroma_width);
 520         }
 521         frame->is_semiplanar = false;
 522         frame->width = width;
 523         frame->height = height;
 524         frame->chroma_subsampling_x = 2;
 525         frame->chroma_subsampling_y = 1;
 526         frame->pitch_y = width;
 527         frame->pitch_chroma = chroma_width;
 528         return frame;
 529 }
 530
 531 }  // namespace
 532
 533 void VideoStream::encode_thread_func()
 534 {
 535         pthread_setname_np(pthread_self(), "VideoStream");
 536         QSurface *surface = create_surface();
 537         QOpenGLContext *context = create_context(surface);
 538         bool ok = make_current(context, surface);
 539         if (!ok) {
 540                 fprintf(stderr, "Video stream couldn't get an OpenGL context\n");
 541                 exit(1);
 542         }
 543
 544         for ( ;; ) {
 545                 QueuedFrame qf;
 546                 {
 547                         unique_lock<mutex> lock(queue_lock);
 548
 549                         // Wait until we have a frame to play.
 550                         queue_changed.wait(lock, [this]{
 551                                 return !frame_queue.empty();
 552                         });
 553                         steady_clock::time_point frame_start = frame_queue.front().local_pts;
 554
 555                         // Now sleep until the frame is supposed to start (the usual case),
 556                         // _or_ clear_queue() happened.
 557                         bool aborted = queue_changed.wait_until(lock, frame_start, [this, frame_start]{
 558                                 return frame_queue.empty() || frame_queue.front().local_pts != frame_start;
 559                         });
 560                         if (aborted) {
 561                                 // clear_queue() happened, so don't play this frame after all.
 562                                 continue;
 563                         }
 564                         qf = move(frame_queue.front());
 565                         frame_queue.pop_front();
 566                 }
 567
 568                 if (qf.type == QueuedFrame::ORIGINAL) {
 569                         // Send the JPEG frame on, unchanged.
 570                         string jpeg = frame_reader.read_frame(qf.frame1);
 571                         AVPacket pkt;
 572                         av_init_packet(&pkt);
 573                         pkt.stream_index = 0;
 574                         pkt.data = (uint8_t *)jpeg.data();
 575                         pkt.size = jpeg.size();
 576                         stream_mux->add_packet(pkt, qf.output_pts, qf.output_pts);
 577
 578                         last_frame.assign(&jpeg[0], &jpeg[0] + jpeg.size());
 579                 } else if (qf.type == QueuedFrame::FADED) {
 580                         glClientWaitSync(qf.fence.get(), /*flags=*/0, GL_TIMEOUT_IGNORED);
 581
 582                         shared_ptr<Frame> frame = frame_from_pbo(qf.resources->pbo_contents, 1280, 720);
 583
 584                         // Now JPEG encode it, and send it on to the stream.
 585                         vector<uint8_t> jpeg = encode_jpeg(frame->y.get(), frame->cb.get(), frame->cr.get(), 1280, 720);
 586
 587                         AVPacket pkt;
 588                         av_init_packet(&pkt);
 589                         pkt.stream_index = 0;
 590                         pkt.data = (uint8_t *)jpeg.data();
 591                         pkt.size = jpeg.size();
 592                         stream_mux->add_packet(pkt, qf.output_pts, qf.output_pts);
 593                         last_frame = move(jpeg);
 594                 } else if (qf.type == QueuedFrame::INTERPOLATED || qf.type == QueuedFrame::FADED_INTERPOLATED) {
 595                         glClientWaitSync(qf.fence.get(), /*flags=*/0, GL_TIMEOUT_IGNORED);
 596
 597                         // Send it on to display.
 598                         shared_ptr<Frame> frame = frame_from_pbo(qf.resources->pbo_contents, 1280, 720);
 599                         if (qf.display_decoded_func != nullptr) {
 600                                 qf.display_decoded_func(frame);
 601                         }
 602
 603                         // Now JPEG encode it, and send it on to the stream.
 604                         vector<uint8_t> jpeg = encode_jpeg(frame->y.get(), frame->cb.get(), frame->cr.get(), 1280, 720);
 605                         compute_flow->release_texture(qf.flow_tex);
 606                         if (qf.type != QueuedFrame::FADED_INTERPOLATED) {
 607                                 interpolate->release_texture(qf.output_tex);
 608                                 interpolate->release_texture(qf.cbcr_tex);
 609                         }
 610
 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                         last_frame = move(jpeg);
 618                 } else if (qf.type == QueuedFrame::REFRESH) {
 619                         AVPacket pkt;
 620                         av_init_packet(&pkt);
 621                         pkt.stream_index = 0;
 622                         pkt.data = (uint8_t *)last_frame.data();
 623                         pkt.size = last_frame.size();
 624                         stream_mux->add_packet(pkt, qf.output_pts, qf.output_pts);
 625                 } else {
 626                         assert(false);
 627                 }
 628                 if (qf.display_func != nullptr) {
 629                         qf.display_func();
 630                 }
 631         }
 632 }
 633
 634 int VideoStream::write_packet2_thunk(void *opaque, uint8_t *buf, int buf_size, AVIODataMarkerType type, int64_t time)
 635 {
 636         VideoStream *video_stream = (VideoStream *)opaque;
 637         return video_stream->write_packet2(buf, buf_size, type, time);
 638 }
 639
 640 int VideoStream::write_packet2(uint8_t *buf, int buf_size, AVIODataMarkerType type, int64_t time)
 641 {
 642         if (type == AVIO_DATA_MARKER_SYNC_POINT || type == AVIO_DATA_MARKER_BOUNDARY_POINT) {
 643                 seen_sync_markers = true;
 644         } else if (type == AVIO_DATA_MARKER_UNKNOWN && !seen_sync_markers) {
 645                 // We don't know if this is a keyframe or not (the muxer could
 646                 // avoid marking it), so we just have to make the best of it.
 647                 type = AVIO_DATA_MARKER_SYNC_POINT;
 648         }
 649
 650         if (type == AVIO_DATA_MARKER_HEADER) {
 651                 stream_mux_header.append((char *)buf, buf_size);
 652                 global_httpd->set_header(HTTPD::MAIN_STREAM, stream_mux_header);
 653         } else {
 654                 global_httpd->add_data(HTTPD::MAIN_STREAM, (char *)buf, buf_size, type == AVIO_DATA_MARKER_SYNC_POINT, time, AVRational{ AV_TIME_BASE, 1 });
 655         }
 656         return buf_size;
 657 }