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
  26 extern HTTPD *global_httpd;
  27
  28 namespace {
  29
  30 string read_file(const string &filename)
  31 {
  32         FILE *fp = fopen(filename.c_str(), "rb");
  33         if (fp == nullptr) {
  34                 perror(filename.c_str());
  35                 return "";
  36         }
  37
  38         fseek(fp, 0, SEEK_END);
  39         long len = ftell(fp);
  40         rewind(fp);
  41
  42         string ret;
  43         ret.resize(len);
  44         fread(&ret[0], len, 1, fp);
  45         fclose(fp);
  46         return ret;
  47 }
  48
  49 }  // namespace
  50
  51 struct VectorDestinationManager {
  52         jpeg_destination_mgr pub;
  53         std::vector<uint8_t> dest;
  54
  55         VectorDestinationManager()
  56         {
  57                 pub.init_destination = init_destination_thunk;
  58                 pub.empty_output_buffer = empty_output_buffer_thunk;
  59                 pub.term_destination = term_destination_thunk;
  60         }
  61
  62         static void init_destination_thunk(j_compress_ptr ptr)
  63         {
  64                 ((VectorDestinationManager *)(ptr->dest))->init_destination();
  65         }
  66
  67         inline void init_destination()
  68         {
  69                 make_room(0);
  70         }
  71
  72         static boolean empty_output_buffer_thunk(j_compress_ptr ptr)
  73         {
  74                 return ((VectorDestinationManager *)(ptr->dest))->empty_output_buffer();
  75         }
  76
  77         inline bool empty_output_buffer()
  78         {
  79                 make_room(dest.size());  // Should ignore pub.free_in_buffer!
  80                 return true;
  81         }
  82
  83         inline void make_room(size_t bytes_used)
  84         {
  85                 dest.resize(bytes_used + 4096);
  86                 dest.resize(dest.capacity());
  87                 pub.next_output_byte = dest.data() + bytes_used;
  88                 pub.free_in_buffer = dest.size() - bytes_used;
  89         }
  90
  91         static void term_destination_thunk(j_compress_ptr ptr)
  92         {
  93                 ((VectorDestinationManager *)(ptr->dest))->term_destination();
  94         }
  95
  96         inline void term_destination()
  97         {
  98                 dest.resize(dest.size() - pub.free_in_buffer);
  99         }
 100 };
 101 static_assert(std::is_standard_layout<VectorDestinationManager>::value, "");
 102
 103 vector<uint8_t> encode_jpeg(const uint8_t *y_data, const uint8_t *cb_data, const uint8_t *cr_data, unsigned width, unsigned height)
 104 {
 105         VectorDestinationManager dest;
 106
 107         jpeg_compress_struct cinfo;
 108         jpeg_error_mgr jerr;
 109         cinfo.err = jpeg_std_error(&jerr);
 110         jpeg_create_compress(&cinfo);
 111
 112         cinfo.dest = (jpeg_destination_mgr *)&dest;
 113         cinfo.input_components = 3;
 114         cinfo.in_color_space = JCS_RGB;
 115         jpeg_set_defaults(&cinfo);
 116         constexpr int quality = 90;
 117         jpeg_set_quality(&cinfo, quality, /*force_baseline=*/false);
 118
 119         cinfo.image_width = width;
 120         cinfo.image_height = height;
 121         cinfo.raw_data_in = true;
 122         jpeg_set_colorspace(&cinfo, JCS_YCbCr);
 123         cinfo.comp_info[0].h_samp_factor = 2;
 124         cinfo.comp_info[0].v_samp_factor = 1;
 125         cinfo.comp_info[1].h_samp_factor = 1;
 126         cinfo.comp_info[1].v_samp_factor = 1;
 127         cinfo.comp_info[2].h_samp_factor = 1;
 128         cinfo.comp_info[2].v_samp_factor = 1;
 129         cinfo.CCIR601_sampling = true;  // Seems to be mostly ignored by libjpeg, though.
 130         jpeg_start_compress(&cinfo, true);
 131
 132         JSAMPROW yptr[8], cbptr[8], crptr[8];
 133         JSAMPARRAY data[3] = { yptr, cbptr, crptr };
 134         for (unsigned y = 0; y < height; y += 8) {
 135                 for (unsigned yy = 0; yy < 8; ++yy) {
 136                         yptr[yy] = const_cast<JSAMPROW>(&y_data[(y + yy) * width]);
 137                         cbptr[yy] = const_cast<JSAMPROW>(&cb_data[(y + yy) * width / 2]);
 138                         crptr[yy] = const_cast<JSAMPROW>(&cr_data[(y + yy) * width / 2]);
 139                 }
 140
 141                 jpeg_write_raw_data(&cinfo, data, /*num_lines=*/8);
 142         }
 143
 144         jpeg_finish_compress(&cinfo);
 145         jpeg_destroy_compress(&cinfo);
 146
 147         return move(dest.dest);
 148 }
 149
 150 VideoStream::VideoStream()
 151 {
 152         ycbcr_converter.reset(new YCbCrConverter(YCbCrConverter::OUTPUT_TO_DUAL_YCBCR, /*resource_pool=*/nullptr));
 153         ycbcr_semiplanar_converter.reset(new YCbCrConverter(YCbCrConverter::OUTPUT_TO_SEMIPLANAR, /*resource_pool=*/nullptr));
 154
 155         GLuint input_tex[num_interpolate_slots], gray_tex[num_interpolate_slots];
 156         GLuint fade_y_output_tex[num_interpolate_slots], fade_cbcr_output_tex[num_interpolate_slots];
 157         GLuint cb_tex[num_interpolate_slots], cr_tex[num_interpolate_slots];
 158
 159         glCreateTextures(GL_TEXTURE_2D_ARRAY, 10, input_tex);
 160         glCreateTextures(GL_TEXTURE_2D_ARRAY, 10, gray_tex);
 161         glCreateTextures(GL_TEXTURE_2D, 10, fade_y_output_tex);
 162         glCreateTextures(GL_TEXTURE_2D, 10, fade_cbcr_output_tex);
 163         glCreateTextures(GL_TEXTURE_2D, 10, cb_tex);
 164         glCreateTextures(GL_TEXTURE_2D, 10, cr_tex);
 165         check_error();
 166
 167         constexpr size_t width = 1280, height = 720;  // FIXME: adjustable width, height
 168         int levels = find_num_levels(width, height);
 169         for (size_t i = 0; i < num_interpolate_slots; ++i) {
 170                 glTextureStorage3D(input_tex[i], levels, GL_RGBA8, width, height, 2);
 171                 check_error();
 172                 glTextureStorage3D(gray_tex[i], levels, GL_R8, width, height, 2);
 173                 check_error();
 174                 glTextureStorage2D(fade_y_output_tex[i], 1, GL_R8, width, height);
 175                 check_error();
 176                 glTextureStorage2D(fade_cbcr_output_tex[i], 1, GL_RG8, width, height);
 177                 check_error();
 178                 glTextureStorage2D(cb_tex[i], 1, GL_R8, width / 2, height);
 179                 check_error();
 180                 glTextureStorage2D(cr_tex[i], 1, GL_R8, width / 2, height);
 181                 check_error();
 182
 183                 InterpolatedFrameResources resource;
 184                 resource.input_tex = input_tex[i];
 185                 resource.gray_tex = gray_tex[i];
 186                 resource.fade_y_output_tex = fade_y_output_tex[i];
 187                 resource.fade_cbcr_output_tex = fade_cbcr_output_tex[i];
 188                 resource.cb_tex = cb_tex[i];
 189                 resource.cr_tex = cr_tex[i];
 190                 glCreateFramebuffers(2, resource.input_fbos);
 191                 check_error();
 192                 glCreateFramebuffers(1, &resource.fade_fbo);
 193                 check_error();
 194
 195                 glNamedFramebufferTextureLayer(resource.input_fbos[0], GL_COLOR_ATTACHMENT0, input_tex[i], 0, 0);
 196                 check_error();
 197                 glNamedFramebufferTextureLayer(resource.input_fbos[0], GL_COLOR_ATTACHMENT1, gray_tex[i], 0, 0);
 198                 check_error();
 199                 glNamedFramebufferTextureLayer(resource.input_fbos[1], GL_COLOR_ATTACHMENT0, input_tex[i], 0, 1);
 200                 check_error();
 201                 glNamedFramebufferTextureLayer(resource.input_fbos[1], GL_COLOR_ATTACHMENT1, gray_tex[i], 0, 1);
 202                 check_error();
 203                 glNamedFramebufferTexture(resource.fade_fbo, GL_COLOR_ATTACHMENT0, fade_y_output_tex[i], 0);
 204                 check_error();
 205                 glNamedFramebufferTexture(resource.fade_fbo, GL_COLOR_ATTACHMENT1, fade_cbcr_output_tex[i], 0);
 206                 check_error();
 207
 208                 GLuint bufs[] = { GL_COLOR_ATTACHMENT0, GL_COLOR_ATTACHMENT1 };
 209                 glNamedFramebufferDrawBuffers(resource.input_fbos[0], 2, bufs);
 210                 check_error();
 211                 glNamedFramebufferDrawBuffers(resource.input_fbos[1], 2, bufs);
 212                 check_error();
 213                 glNamedFramebufferDrawBuffers(resource.fade_fbo, 2, bufs);
 214                 check_error();
 215
 216                 glCreateBuffers(1, &resource.pbo);
 217                 check_error();
 218                 glNamedBufferStorage(resource.pbo, width * height * 4, nullptr, GL_MAP_READ_BIT | GL_MAP_PERSISTENT_BIT);
 219                 check_error();
 220                 resource.pbo_contents = glMapNamedBufferRange(resource.pbo, 0, width * height * 4, GL_MAP_READ_BIT | GL_MAP_PERSISTENT_BIT);
 221                 interpolate_resources.push_back(resource);
 222         }
 223
 224         check_error();
 225
 226         OperatingPoint op;
 227         if (global_flags.interpolation_quality == 1) {
 228                 op = operating_point1;
 229         } else if (global_flags.interpolation_quality == 2) {
 230                 op = operating_point2;
 231         } else if (global_flags.interpolation_quality == 3) {
 232                 op = operating_point3;
 233         } else if (global_flags.interpolation_quality == 4) {
 234                 op = operating_point4;
 235         } else {
 236                 assert(false);
 237         }
 238
 239         compute_flow.reset(new DISComputeFlow(width, height, op));
 240         interpolate.reset(new Interpolate(op, /*split_ycbcr_output=*/true));
 241         interpolate_no_split.reset(new Interpolate(op, /*split_ycbcr_output=*/false));
 242         chroma_subsampler.reset(new ChromaSubsampler);
 243         check_error();
 244
 245         // The “last frame” is initially black.
 246         unique_ptr<uint8_t[]> y(new uint8_t[1280 * 720]);
 247         unique_ptr<uint8_t[]> cb_or_cr(new uint8_t[640 * 720]);
 248         memset(y.get(), 16, 1280 * 720);
 249         memset(cb_or_cr.get(), 128, 640 * 720);
 250         last_frame = encode_jpeg(y.get(), cb_or_cr.get(), cb_or_cr.get(), 1280, 720);
 251 }
 252
 253 VideoStream::~VideoStream() {}
 254
 255 void VideoStream::start()
 256 {
 257         AVFormatContext *avctx = avformat_alloc_context();
 258         avctx->oformat = av_guess_format("nut", nullptr, nullptr);
 259
 260         uint8_t *buf = (uint8_t *)av_malloc(MUX_BUFFER_SIZE);
 261         avctx->pb = avio_alloc_context(buf, MUX_BUFFER_SIZE, 1, this, nullptr, nullptr, nullptr);
 262         avctx->pb->write_data_type = &VideoStream::write_packet2_thunk;
 263         avctx->pb->ignore_boundary_point = 1;
 264
 265         Mux::Codec video_codec = Mux::CODEC_MJPEG;
 266
 267         avctx->flags = AVFMT_FLAG_CUSTOM_IO;
 268
 269         string video_extradata;
 270
 271         constexpr int width = 1280, height = 720;  // Doesn't matter for MJPEG.
 272         stream_mux.reset(new Mux(avctx, width, height, video_codec, video_extradata, /*audio_codec_parameters=*/nullptr, COARSE_TIMEBASE,
 273                 /*write_callback=*/nullptr, Mux::WRITE_FOREGROUND, {}));
 274
 275
 276         encode_thread = thread(&VideoStream::encode_thread_func, this);
 277 }
 278
 279 void VideoStream::stop()
 280 {
 281         encode_thread.join();
 282 }
 283
 284 void VideoStream::schedule_original_frame(int64_t output_pts, function<void()> &&display_func, unsigned stream_idx, int64_t input_pts)
 285 {
 286         fprintf(stderr, "output_pts=%ld  original      input_pts=%ld\n", output_pts, input_pts);
 287
 288         QueuedFrame qf;
 289         qf.type = QueuedFrame::ORIGINAL;
 290         qf.output_pts = output_pts;
 291         qf.stream_idx = stream_idx;
 292         qf.input_first_pts = input_pts;
 293         qf.display_func = move(display_func);
 294
 295         unique_lock<mutex> lock(queue_lock);
 296         frame_queue.push_back(qf);
 297         queue_nonempty.notify_all();
 298 }
 299
 300 void VideoStream::schedule_faded_frame(int64_t output_pts, function<void()> &&display_func, unsigned stream_idx, int64_t input_pts, int secondary_stream_idx, int64_t secondary_input_pts, float fade_alpha)
 301 {
 302         fprintf(stderr, "output_pts=%ld  faded         input_pts=%ld,%ld  fade_alpha=%.2f\n", output_pts, input_pts, secondary_input_pts, fade_alpha);
 303
 304         // Get the temporary OpenGL resources we need for doing the fade.
 305         // (We share these with interpolated frames, which is slightly
 306         // overkill, but there's no need to waste resources on keeping
 307         // separate pools around.)
 308         InterpolatedFrameResources resources;
 309         {
 310                 unique_lock<mutex> lock(queue_lock);
 311                 if (interpolate_resources.empty()) {
 312                         fprintf(stderr, "WARNING: Too many interpolated frames already in transit; dropping one.\n");
 313                         return;
 314                 }
 315                 resources = interpolate_resources.front();
 316                 interpolate_resources.pop_front();
 317         }
 318
 319         bool did_decode;
 320
 321         JPEGID jpeg_id1;
 322         jpeg_id1.stream_idx = stream_idx;
 323         jpeg_id1.pts = input_pts;
 324         jpeg_id1.interpolated = false;
 325         shared_ptr<Frame> frame1 = decode_jpeg_with_cache(jpeg_id1, DECODE_IF_NOT_IN_CACHE, &did_decode);
 326
 327         JPEGID jpeg_id2;
 328         jpeg_id2.stream_idx = secondary_stream_idx;
 329         jpeg_id2.pts = secondary_input_pts;
 330         jpeg_id2.interpolated = false;
 331         shared_ptr<Frame> frame2 = decode_jpeg_with_cache(jpeg_id2, DECODE_IF_NOT_IN_CACHE, &did_decode);
 332
 333         ycbcr_semiplanar_converter->prepare_chain_for_fade(frame1, frame2, fade_alpha)->render_to_fbo(resources.fade_fbo, 1280, 720);
 334
 335         QueuedFrame qf;
 336         qf.type = QueuedFrame::FADED;
 337         qf.output_pts = output_pts;
 338         qf.stream_idx = stream_idx;
 339         qf.resources = resources;
 340         qf.input_first_pts = input_pts;
 341         qf.display_func = move(display_func);
 342
 343         qf.secondary_stream_idx = secondary_stream_idx;
 344         qf.secondary_input_pts = secondary_input_pts;
 345
 346         // Subsample and split Cb/Cr.
 347         chroma_subsampler->subsample_chroma(resources.fade_cbcr_output_tex, 1280, 720, resources.cb_tex, resources.cr_tex);
 348
 349         // Read it down (asynchronously) to the CPU.
 350         glPixelStorei(GL_PACK_ROW_LENGTH, 0);
 351         glBindBuffer(GL_PIXEL_PACK_BUFFER, resources.pbo);
 352         check_error();
 353         glGetTextureImage(resources.fade_y_output_tex, 0, GL_RED, GL_UNSIGNED_BYTE, 1280 * 720 * 4, BUFFER_OFFSET(0));
 354         check_error();
 355         glGetTextureImage(resources.cb_tex, 0, GL_RED, GL_UNSIGNED_BYTE, 1280 * 720 * 3, BUFFER_OFFSET(1280 * 720));
 356         check_error();
 357         glGetTextureImage(resources.cr_tex, 0, GL_RED, GL_UNSIGNED_BYTE, 1280 * 720 * 3 - 640 * 720, BUFFER_OFFSET(1280 * 720 + 640 * 720));
 358         check_error();
 359         glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
 360
 361         // Set a fence we can wait for to make sure the CPU sees the read.
 362         glMemoryBarrier(GL_CLIENT_MAPPED_BUFFER_BARRIER_BIT);
 363         check_error();
 364         qf.fence = RefCountedGLsync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
 365         check_error();
 366
 367         unique_lock<mutex> lock(queue_lock);
 368         frame_queue.push_back(qf);
 369         queue_nonempty.notify_all();
 370 }
 371
 372 void VideoStream::schedule_interpolated_frame(int64_t output_pts, function<void()> &&display_func, unsigned stream_idx, int64_t input_first_pts, int64_t input_second_pts, float alpha, int secondary_stream_idx, int64_t secondary_input_pts, float fade_alpha)
 373 {
 374         if (secondary_stream_idx != -1) {
 375                 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);
 376         } else {
 377                 fprintf(stderr, "output_pts=%ld  interpolated  input_pts1=%ld input_pts2=%ld alpha=%.3f\n", output_pts, input_first_pts, input_second_pts, alpha);
 378         }
 379
 380         JPEGID id;
 381         if (secondary_stream_idx == -1) {
 382                 id = JPEGID{ stream_idx, output_pts, /*interpolated=*/true };
 383         } else {
 384                 id = create_jpegid_for_interpolated_fade(stream_idx, output_pts, secondary_stream_idx, secondary_input_pts);
 385         }
 386
 387         // Get the temporary OpenGL resources we need for doing the interpolation.
 388         InterpolatedFrameResources resources;
 389         {
 390                 unique_lock<mutex> lock(queue_lock);
 391                 if (interpolate_resources.empty()) {
 392                         fprintf(stderr, "WARNING: Too many interpolated frames already in transit; dropping one.\n");
 393                         return;
 394                 }
 395                 resources = interpolate_resources.front();
 396                 interpolate_resources.pop_front();
 397         }
 398
 399         QueuedFrame qf;
 400         qf.type = (secondary_stream_idx == -1) ? QueuedFrame::INTERPOLATED : QueuedFrame::FADED_INTERPOLATED;
 401         qf.output_pts = output_pts;
 402         qf.stream_idx = stream_idx;
 403         qf.resources = resources;
 404         qf.id = id;
 405         qf.display_func = move(display_func);
 406
 407         check_error();
 408
 409         // Convert frame0 and frame1 to OpenGL textures.
 410         for (size_t frame_no = 0; frame_no < 2; ++frame_no) {
 411                 JPEGID jpeg_id;
 412                 jpeg_id.stream_idx = stream_idx;
 413                 jpeg_id.pts = frame_no == 1 ? input_second_pts : input_first_pts;
 414                 jpeg_id.interpolated = false;
 415                 bool did_decode;
 416                 shared_ptr<Frame> frame = decode_jpeg_with_cache(jpeg_id, DECODE_IF_NOT_IN_CACHE, &did_decode);
 417                 ycbcr_converter->prepare_chain_for_conversion(frame)->render_to_fbo(resources.input_fbos[frame_no], 1280, 720);
 418         }
 419
 420         glGenerateTextureMipmap(resources.input_tex);
 421         check_error();
 422         glGenerateTextureMipmap(resources.gray_tex);
 423         check_error();
 424
 425         // Compute the interpolated frame.
 426         qf.flow_tex = compute_flow->exec(resources.gray_tex, DISComputeFlow::FORWARD_AND_BACKWARD, DISComputeFlow::DO_NOT_RESIZE_FLOW);
 427         check_error();
 428
 429         if (secondary_stream_idx != -1) {
 430                 // Fade. First kick off the interpolation.
 431                 tie(qf.output_tex, ignore) = interpolate_no_split->exec(resources.input_tex, resources.gray_tex, qf.flow_tex, 1280, 720, alpha);
 432                 check_error();
 433
 434                 // Now decode the image we are fading against.
 435                 JPEGID jpeg_id;
 436                 jpeg_id.stream_idx = secondary_stream_idx;
 437                 jpeg_id.pts = secondary_input_pts;
 438                 jpeg_id.interpolated = false;
 439                 bool did_decode;
 440                 shared_ptr<Frame> frame2 = decode_jpeg_with_cache(jpeg_id, DECODE_IF_NOT_IN_CACHE, &did_decode);
 441
 442                 // Then fade against it, putting it into the fade Y' and CbCr textures.
 443                 ycbcr_semiplanar_converter->prepare_chain_for_fade_from_texture(qf.output_tex, frame2, fade_alpha)->render_to_fbo(resources.fade_fbo, 1280, 720);
 444
 445                 // Subsample and split Cb/Cr.
 446                 chroma_subsampler->subsample_chroma(resources.fade_cbcr_output_tex, 1280, 720, resources.cb_tex, resources.cr_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_stream_idx != -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
 481         unique_lock<mutex> lock(queue_lock);
 482         frame_queue.push_back(qf);
 483         queue_nonempty.notify_all();
 484 }
 485
 486 void VideoStream::schedule_refresh_frame(int64_t output_pts, function<void()> &&display_func)
 487 {
 488         QueuedFrame qf;
 489         qf.type = QueuedFrame::REFRESH;
 490         qf.output_pts = output_pts;
 491         qf.display_func = move(display_func);
 492
 493         unique_lock<mutex> lock(queue_lock);
 494         frame_queue.push_back(qf);
 495         queue_nonempty.notify_all();
 496 }
 497
 498 namespace {
 499
 500 shared_ptr<Frame> frame_from_pbo(void *contents, size_t width, size_t height)
 501 {
 502         size_t chroma_width = width / 2;
 503
 504         const uint8_t *y = (const uint8_t *)contents;
 505         const uint8_t *cb = (const uint8_t *)contents + width * height;
 506         const uint8_t *cr = (const uint8_t *)contents + width * height + chroma_width * height;
 507
 508         shared_ptr<Frame> frame(new Frame);
 509         frame->y.reset(new uint8_t[width * height]);
 510         frame->cb.reset(new uint8_t[chroma_width * height]);
 511         frame->cr.reset(new uint8_t[chroma_width * height]);
 512         for (unsigned yy = 0; yy < height; ++yy) {
 513                 memcpy(frame->y.get() + width * yy, y + width * yy, width);
 514                 memcpy(frame->cb.get() + chroma_width * yy, cb + chroma_width * yy, chroma_width);
 515                 memcpy(frame->cr.get() + chroma_width * yy, cr + chroma_width * yy, chroma_width);
 516         }
 517         frame->is_semiplanar = false;
 518         frame->width = width;
 519         frame->height = height;
 520         frame->chroma_subsampling_x = 2;
 521         frame->chroma_subsampling_y = 1;
 522         frame->pitch_y = width;
 523         frame->pitch_chroma = chroma_width;
 524         return frame;
 525 }
 526
 527 }  // namespace
 528
 529 void VideoStream::encode_thread_func()
 530 {
 531         pthread_setname_np(pthread_self(), "VideoStream");
 532         QSurface *surface = create_surface();
 533         QOpenGLContext *context = create_context(surface);
 534         bool ok = make_current(context, surface);
 535         if (!ok) {
 536                 fprintf(stderr, "Video stream couldn't get an OpenGL context\n");
 537                 exit(1);
 538         }
 539
 540         for ( ;; ) {
 541                 QueuedFrame qf;
 542                 {
 543                         unique_lock<mutex> lock(queue_lock);
 544                         queue_nonempty.wait(lock, [this]{
 545                                 return !frame_queue.empty();
 546                         });
 547                         qf = frame_queue.front();
 548                         frame_queue.pop_front();
 549                 }
 550
 551                 if (qf.type == QueuedFrame::ORIGINAL) {
 552                         // Send the JPEG frame on, unchanged.
 553                         string jpeg = read_file(filename_for_frame(qf.stream_idx, qf.input_first_pts));
 554                         AVPacket pkt;
 555                         av_init_packet(&pkt);
 556                         pkt.stream_index = 0;
 557                         pkt.data = (uint8_t *)jpeg.data();
 558                         pkt.size = jpeg.size();
 559                         stream_mux->add_packet(pkt, qf.output_pts, qf.output_pts);
 560
 561                         last_frame.assign(&jpeg[0], &jpeg[0] + jpeg.size());
 562                 } else if (qf.type == QueuedFrame::FADED) {
 563                         glClientWaitSync(qf.fence.get(), /*flags=*/0, GL_TIMEOUT_IGNORED);
 564
 565                         shared_ptr<Frame> frame = frame_from_pbo(qf.resources.pbo_contents, 1280, 720);
 566
 567                         // Now JPEG encode it, and send it on to the stream.
 568                         vector<uint8_t> jpeg = encode_jpeg(frame->y.get(), frame->cb.get(), frame->cr.get(), 1280, 720);
 569
 570                         AVPacket pkt;
 571                         av_init_packet(&pkt);
 572                         pkt.stream_index = 0;
 573                         pkt.data = (uint8_t *)jpeg.data();
 574                         pkt.size = jpeg.size();
 575                         stream_mux->add_packet(pkt, qf.output_pts, qf.output_pts);
 576                         last_frame = move(jpeg);
 577
 578                         // Put the frame resources back.
 579                         unique_lock<mutex> lock(queue_lock);
 580                         interpolate_resources.push_back(qf.resources);
 581                 } else if (qf.type == QueuedFrame::INTERPOLATED || qf.type == QueuedFrame::FADED_INTERPOLATED) {
 582                         glClientWaitSync(qf.fence.get(), /*flags=*/0, GL_TIMEOUT_IGNORED);
 583
 584                         // Send a copy of the frame on to display.
 585                         shared_ptr<Frame> frame = frame_from_pbo(qf.resources.pbo_contents, 1280, 720);
 586                         JPEGFrameView::insert_interpolated_frame(qf.id, frame);
 587
 588                         // Now JPEG encode it, and send it on to the stream.
 589                         vector<uint8_t> jpeg = encode_jpeg(frame->y.get(), frame->cb.get(), frame->cr.get(), 1280, 720);
 590                         compute_flow->release_texture(qf.flow_tex);
 591                         if (qf.type != QueuedFrame::FADED_INTERPOLATED) {
 592                                 interpolate->release_texture(qf.output_tex);
 593                                 interpolate->release_texture(qf.cbcr_tex);
 594                         }
 595
 596                         AVPacket pkt;
 597                         av_init_packet(&pkt);
 598                         pkt.stream_index = 0;
 599                         pkt.data = (uint8_t *)jpeg.data();
 600                         pkt.size = jpeg.size();
 601                         stream_mux->add_packet(pkt, qf.output_pts, qf.output_pts);
 602                         last_frame = move(jpeg);
 603
 604                         // Put the frame resources back.
 605                         unique_lock<mutex> lock(queue_lock);
 606                         interpolate_resources.push_back(qf.resources);
 607                 } else if (qf.type == QueuedFrame::REFRESH) {
 608                         AVPacket pkt;
 609                         av_init_packet(&pkt);
 610                         pkt.stream_index = 0;
 611                         pkt.data = (uint8_t *)last_frame.data();
 612                         pkt.size = last_frame.size();
 613                         stream_mux->add_packet(pkt, qf.output_pts, qf.output_pts);
 614                 } else {
 615                         assert(false);
 616                 }
 617                 if (qf.display_func != nullptr) {
 618                         qf.display_func();
 619                 }
 620         }
 621 }
 622
 623 int VideoStream::write_packet2_thunk(void *opaque, uint8_t *buf, int buf_size, AVIODataMarkerType type, int64_t time)
 624 {
 625         VideoStream *video_stream = (VideoStream *)opaque;
 626         return video_stream->write_packet2(buf, buf_size, type, time);
 627 }
 628
 629 int VideoStream::write_packet2(uint8_t *buf, int buf_size, AVIODataMarkerType type, int64_t time)
 630 {
 631         if (type == AVIO_DATA_MARKER_SYNC_POINT || type == AVIO_DATA_MARKER_BOUNDARY_POINT) {
 632                 seen_sync_markers = true;
 633         } else if (type == AVIO_DATA_MARKER_UNKNOWN && !seen_sync_markers) {
 634                 // We don't know if this is a keyframe or not (the muxer could
 635                 // avoid marking it), so we just have to make the best of it.
 636                 type = AVIO_DATA_MARKER_SYNC_POINT;
 637         }
 638
 639         if (type == AVIO_DATA_MARKER_HEADER) {
 640                 stream_mux_header.append((char *)buf, buf_size);
 641                 global_httpd->set_header(stream_mux_header);
 642         } else {
 643                 global_httpd->add_data((char *)buf, buf_size, type == AVIO_DATA_MARKER_SYNC_POINT, time, AVRational{ AV_TIME_BASE, 1 });
 644         }
 645         return buf_size;
 646 }