git.sesse.net Git - nageru/blob - mixer.cpp

   1 #undef Success
   2
   3 #include "mixer.h"
   4
   5 #include <assert.h>
   6 #include <epoxy/egl.h>
   7 #include <movit/effect_chain.h>
   8 #include <movit/effect_util.h>
   9 #include <movit/flat_input.h>
  10 #include <movit/image_format.h>
  11 #include <movit/init.h>
  12 #include <movit/resource_pool.h>
  13 #include <movit/util.h>
  14 #include <stdint.h>
  15 #include <stdio.h>
  16 #include <stdlib.h>
  17 #include <sys/time.h>
  18 #include <time.h>
  19 #include <algorithm>
  20 #include <cmath>
  21 #include <condition_variable>
  22 #include <cstddef>
  23 #include <memory>
  24 #include <mutex>
  25 #include <string>
  26 #include <thread>
  27 #include <utility>
  28 #include <vector>
  29 #include <arpa/inet.h>
  30
  31 #include "bmusb/bmusb.h"
  32 #include "context.h"
  33 #include "decklink_capture.h"
  34 #include "defs.h"
  35 #include "flags.h"
  36 #include "h264encode.h"
  37 #include "pbo_frame_allocator.h"
  38 #include "ref_counted_gl_sync.h"
  39 #include "timebase.h"
  40
  41 class QOpenGLContext;
  42
  43 using namespace movit;
  44 using namespace std;
  45 using namespace std::placeholders;
  46
  47 Mixer *global_mixer = nullptr;
  48
  49 namespace {
  50
  51 void convert_fixed24_to_fp32(float *dst, size_t out_channels, const uint8_t *src, size_t in_channels, size_t num_samples)
  52 {
  53         assert(in_channels >= out_channels);
  54         for (size_t i = 0; i < num_samples; ++i) {
  55                 for (size_t j = 0; j < out_channels; ++j) {
  56                         uint32_t s1 = *src++;
  57                         uint32_t s2 = *src++;
  58                         uint32_t s3 = *src++;
  59                         uint32_t s = s1 | (s1 << 8) | (s2 << 16) | (s3 << 24);
  60                         dst[i * out_channels + j] = int(s) * (1.0f / 4294967296.0f);
  61                 }
  62                 src += 3 * (in_channels - out_channels);
  63         }
  64 }
  65
  66 void convert_fixed32_to_fp32(float *dst, size_t out_channels, const uint8_t *src, size_t in_channels, size_t num_samples)
  67 {
  68         assert(in_channels >= out_channels);
  69         for (size_t i = 0; i < num_samples; ++i) {
  70                 for (size_t j = 0; j < out_channels; ++j) {
  71                         // Note: Assumes little-endian.
  72                         int32_t s = *(int32_t *)src;
  73                         dst[i * out_channels + j] = s * (1.0f / 4294967296.0f);
  74                         src += 4;
  75                 }
  76                 src += 4 * (in_channels - out_channels);
  77         }
  78 }
  79
  80 void insert_new_frame(RefCountedFrame frame, unsigned field_num, bool interlaced, unsigned card_index, InputState *input_state)
  81 {
  82         if (interlaced) {
  83                 for (unsigned frame_num = FRAME_HISTORY_LENGTH; frame_num --> 1; ) {  // :-)
  84                         input_state->buffered_frames[card_index][frame_num] =
  85                                 input_state->buffered_frames[card_index][frame_num - 1];
  86                 }
  87                 input_state->buffered_frames[card_index][0] = { frame, field_num };
  88         } else {
  89                 for (unsigned frame_num = 0; frame_num < FRAME_HISTORY_LENGTH; ++frame_num) {
  90                         input_state->buffered_frames[card_index][frame_num] = { frame, field_num };
  91                 }
  92         }
  93 }
  94
  95 string generate_local_dump_filename(int frame)
  96 {
  97         time_t now = time(NULL);
  98         tm now_tm;
  99         localtime_r(&now, &now_tm);
 100
 101         char timestamp[256];
 102         strftime(timestamp, sizeof(timestamp), "%F-%T%z", &now_tm);
 103
 104         // Use the frame number to disambiguate between two cuts starting
 105         // on the same second.
 106         char filename[256];
 107         snprintf(filename, sizeof(filename), "%s%s-f%02d%s",
 108                 LOCAL_DUMP_PREFIX, timestamp, frame % 100, LOCAL_DUMP_SUFFIX);
 109         return filename;
 110 }
 111
 112 }  // namespace
 113
 114 Mixer::Mixer(const QSurfaceFormat &format, unsigned num_cards)
 115         : httpd(WIDTH, HEIGHT),
 116           num_cards(num_cards),
 117           mixer_surface(create_surface(format)),
 118           h264_encoder_surface(create_surface(format)),
 119           correlation(OUTPUT_FREQUENCY),
 120           level_compressor(OUTPUT_FREQUENCY),
 121           limiter(OUTPUT_FREQUENCY),
 122           compressor(OUTPUT_FREQUENCY)
 123 {
 124         httpd.open_output_file(generate_local_dump_filename(/*frame=*/0).c_str());
 125         httpd.start(9095);
 126
 127         CHECK(init_movit(MOVIT_SHADER_DIR, MOVIT_DEBUG_OFF));
 128         check_error();
 129
 130         // Since we allow non-bouncing 4:2:2 YCbCrInputs, effective subpixel precision
 131         // will be halved when sampling them, and we need to compensate here.
 132         movit_texel_subpixel_precision /= 2.0;
 133
 134         resource_pool.reset(new ResourcePool);
 135         theme.reset(new Theme("theme.lua", resource_pool.get(), num_cards));
 136         for (unsigned i = 0; i < NUM_OUTPUTS; ++i) {
 137                 output_channel[i].parent = this;
 138         }
 139
 140         ImageFormat inout_format;
 141         inout_format.color_space = COLORSPACE_sRGB;
 142         inout_format.gamma_curve = GAMMA_sRGB;
 143
 144         // Display chain; shows the live output produced by the main chain (its RGBA version).
 145         display_chain.reset(new EffectChain(WIDTH, HEIGHT, resource_pool.get()));
 146         check_error();
 147         display_input = new FlatInput(inout_format, FORMAT_RGB, GL_UNSIGNED_BYTE, WIDTH, HEIGHT);  // FIXME: GL_UNSIGNED_BYTE is really wrong.
 148         display_chain->add_input(display_input);
 149         display_chain->add_output(inout_format, OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED);
 150         display_chain->set_dither_bits(0);  // Don't bother.
 151         display_chain->finalize();
 152
 153         h264_encoder.reset(new H264Encoder(h264_encoder_surface, global_flags.va_display, WIDTH, HEIGHT, &httpd));
 154
 155         // First try initializing the PCI devices, then USB, until we have the desired number of cards.
 156         unsigned num_pci_devices = 0, num_usb_devices = 0;
 157         unsigned card_index = 0;
 158
 159         IDeckLinkIterator *decklink_iterator = CreateDeckLinkIteratorInstance();
 160         if (decklink_iterator != nullptr) {
 161                 for ( ; card_index < num_cards; ++card_index) {
 162                         IDeckLink *decklink;
 163                         if (decklink_iterator->Next(&decklink) != S_OK) {
 164                                 break;
 165                         }
 166
 167                         configure_card(card_index, format, new DeckLinkCapture(decklink, card_index));
 168                         ++num_pci_devices;
 169                 }
 170                 decklink_iterator->Release();
 171                 fprintf(stderr, "Found %d DeckLink PCI card(s).\n", num_pci_devices);
 172         } else {
 173                 fprintf(stderr, "DeckLink drivers not found. Probing for USB cards only.\n");
 174         }
 175         for ( ; card_index < num_cards; ++card_index) {
 176                 configure_card(card_index, format, new BMUSBCapture(card_index - num_pci_devices));
 177                 ++num_usb_devices;
 178         }
 179
 180         if (num_usb_devices > 0) {
 181                 BMUSBCapture::start_bm_thread();
 182         }
 183
 184         for (card_index = 0; card_index < num_cards; ++card_index) {
 185                 cards[card_index].capture->start_bm_capture();
 186         }
 187
 188         // Set up stuff for NV12 conversion.
 189
 190         // Cb/Cr shader.
 191         string cbcr_vert_shader =
 192                 "#version 130 \n"
 193                 " \n"
 194                 "in vec2 position; \n"
 195                 "in vec2 texcoord; \n"
 196                 "out vec2 tc0; \n"
 197                 "uniform vec2 foo_chroma_offset_0; \n"
 198                 " \n"
 199                 "void main() \n"
 200                 "{ \n"
 201                 "    // The result of glOrtho(0.0, 1.0, 0.0, 1.0, 0.0, 1.0) is: \n"
 202                 "    // \n"
 203                 "    //   2.000  0.000  0.000 -1.000 \n"
 204                 "    //   0.000  2.000  0.000 -1.000 \n"
 205                 "    //   0.000  0.000 -2.000 -1.000 \n"
 206                 "    //   0.000  0.000  0.000  1.000 \n"
 207                 "    gl_Position = vec4(2.0 * position.x - 1.0, 2.0 * position.y - 1.0, -1.0, 1.0); \n"
 208                 "    vec2 flipped_tc = texcoord; \n"
 209                 "    tc0 = flipped_tc + foo_chroma_offset_0; \n"
 210                 "} \n";
 211         string cbcr_frag_shader =
 212                 "#version 130 \n"
 213                 "in vec2 tc0; \n"
 214                 "uniform sampler2D cbcr_tex; \n"
 215                 "out vec4 FragColor; \n"
 216                 "void main() { \n"
 217                 "    FragColor = texture(cbcr_tex, tc0); \n"
 218                 "} \n";
 219         vector<string> frag_shader_outputs;
 220         cbcr_program_num = resource_pool->compile_glsl_program(cbcr_vert_shader, cbcr_frag_shader, frag_shader_outputs);
 221
 222         float vertices[] = {
 223                 0.0f, 2.0f,
 224                 0.0f, 0.0f,
 225                 2.0f, 0.0f
 226         };
 227         cbcr_vbo = generate_vbo(2, GL_FLOAT, sizeof(vertices), vertices);
 228         cbcr_position_attribute_index = glGetAttribLocation(cbcr_program_num, "position");
 229         cbcr_texcoord_attribute_index = glGetAttribLocation(cbcr_program_num, "texcoord");
 230
 231         r128.init(2, OUTPUT_FREQUENCY);
 232         r128.integr_start();
 233
 234         locut.init(FILTER_HPF, 2);
 235
 236         // hlen=16 is pretty low quality, but we use quite a bit of CPU otherwise,
 237         // and there's a limit to how important the peak meter is.
 238         peak_resampler.setup(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY * 4, /*num_channels=*/2, /*hlen=*/16, /*frel=*/1.0);
 239
 240         alsa.reset(new ALSAOutput(OUTPUT_FREQUENCY, /*num_channels=*/2));
 241 }
 242
 243 Mixer::~Mixer()
 244 {
 245         resource_pool->release_glsl_program(cbcr_program_num);
 246         glDeleteBuffers(1, &cbcr_vbo);
 247         BMUSBCapture::stop_bm_thread();
 248
 249         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 250                 {
 251                         unique_lock<mutex> lock(bmusb_mutex);
 252                         cards[card_index].should_quit = true;  // Unblock thread.
 253                         cards[card_index].new_data_ready_changed.notify_all();
 254                 }
 255                 cards[card_index].capture->stop_dequeue_thread();
 256         }
 257
 258         h264_encoder.reset(nullptr);
 259 }
 260
 261 void Mixer::configure_card(unsigned card_index, const QSurfaceFormat &format, CaptureInterface *capture)
 262 {
 263         printf("Configuring card %d...\n", card_index);
 264
 265         CaptureCard *card = &cards[card_index];
 266         card->capture = capture;
 267         card->capture->set_frame_callback(bind(&Mixer::bm_frame, this, card_index, _1, _2, _3, _4, _5, _6, _7));
 268         card->frame_allocator.reset(new PBOFrameAllocator(8 << 20, WIDTH, HEIGHT));  // 8 MB.
 269         card->capture->set_video_frame_allocator(card->frame_allocator.get());
 270         card->surface = create_surface(format);
 271         card->capture->set_dequeue_thread_callbacks(
 272                 [card]{
 273                         eglBindAPI(EGL_OPENGL_API);
 274                         card->context = create_context(card->surface);
 275                         if (!make_current(card->context, card->surface)) {
 276                                 printf("failed to create bmusb context\n");
 277                                 exit(1);
 278                         }
 279                 },
 280                 [this]{
 281                         resource_pool->clean_context();
 282                 });
 283         card->resampling_queue.reset(new ResamplingQueue(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY, 2));
 284         card->capture->configure_card();
 285 }
 286
 287
 288 namespace {
 289
 290 int unwrap_timecode(uint16_t current_wrapped, int last)
 291 {
 292         uint16_t last_wrapped = last & 0xffff;
 293         if (current_wrapped > last_wrapped) {
 294                 return (last & ~0xffff) | current_wrapped;
 295         } else {
 296                 return 0x10000 + ((last & ~0xffff) | current_wrapped);
 297         }
 298 }
 299
 300 float find_peak(const float *samples, size_t num_samples)
 301 {
 302         float m = fabs(samples[0]);
 303         for (size_t i = 1; i < num_samples; ++i) {
 304                 m = max(m, fabs(samples[i]));
 305         }
 306         return m;
 307 }
 308
 309 void deinterleave_samples(const vector<float> &in, vector<float> *out_l, vector<float> *out_r)
 310 {
 311         size_t num_samples = in.size() / 2;
 312         out_l->resize(num_samples);
 313         out_r->resize(num_samples);
 314
 315         const float *inptr = in.data();
 316         float *lptr = &(*out_l)[0];
 317         float *rptr = &(*out_r)[0];
 318         for (size_t i = 0; i < num_samples; ++i) {
 319                 *lptr++ = *inptr++;
 320                 *rptr++ = *inptr++;
 321         }
 322 }
 323
 324 }  // namespace
 325
 326 void Mixer::bm_frame(unsigned card_index, uint16_t timecode,
 327                      FrameAllocator::Frame video_frame, size_t video_offset, VideoFormat video_format,
 328                      FrameAllocator::Frame audio_frame, size_t audio_offset, AudioFormat audio_format)
 329 {
 330         CaptureCard *card = &cards[card_index];
 331
 332         int64_t frame_length = int64_t(TIMEBASE * video_format.frame_rate_den) / video_format.frame_rate_nom;
 333
 334         size_t num_samples = (audio_frame.len >= audio_offset) ? (audio_frame.len - audio_offset) / audio_format.num_channels / (audio_format.bits_per_sample / 8) : 0;
 335         if (num_samples > OUTPUT_FREQUENCY / 10) {
 336                 printf("Card %d: Dropping frame with implausible audio length (len=%d, offset=%d) [timecode=0x%04x video_len=%d video_offset=%d video_format=%x)\n",
 337                         card_index, int(audio_frame.len), int(audio_offset),
 338                         timecode, int(video_frame.len), int(video_offset), video_format.id);
 339                 if (video_frame.owner) {
 340                         video_frame.owner->release_frame(video_frame);
 341                 }
 342                 if (audio_frame.owner) {
 343                         audio_frame.owner->release_frame(audio_frame);
 344                 }
 345                 return;
 346         }
 347
 348         int64_t local_pts = card->next_local_pts;
 349         int dropped_frames = 0;
 350         if (card->last_timecode != -1) {
 351                 dropped_frames = unwrap_timecode(timecode, card->last_timecode) - card->last_timecode - 1;
 352         }
 353
 354         // Convert the audio to stereo fp32 and add it.
 355         vector<float> audio;
 356         audio.resize(num_samples * 2);
 357         switch (audio_format.bits_per_sample) {
 358         case 24:
 359                 convert_fixed24_to_fp32(&audio[0], 2, audio_frame.data + audio_offset, audio_format.num_channels, num_samples);
 360                 break;
 361         case 32:
 362                 convert_fixed32_to_fp32(&audio[0], 2, audio_frame.data + audio_offset, audio_format.num_channels, num_samples);
 363                 break;
 364         default:
 365                 fprintf(stderr, "Cannot handle audio with %u bits per sample\n", audio_format.bits_per_sample);
 366                 assert(false);
 367         }
 368
 369         // Add the audio.
 370         {
 371                 unique_lock<mutex> lock(card->audio_mutex);
 372
 373                 // Number of samples per frame if we need to insert silence.
 374                 // (Could be nonintegral, but resampling will save us then.)
 375                 int silence_samples = OUTPUT_FREQUENCY * video_format.frame_rate_den / video_format.frame_rate_nom;
 376
 377                 if (dropped_frames > MAX_FPS * 2) {
 378                         fprintf(stderr, "Card %d lost more than two seconds (or time code jumping around; from 0x%04x to 0x%04x), resetting resampler\n",
 379                                 card_index, card->last_timecode, timecode);
 380                         card->resampling_queue.reset(new ResamplingQueue(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY, 2));
 381                         dropped_frames = 0;
 382                 } else if (dropped_frames > 0) {
 383                         // Insert silence as needed.
 384                         fprintf(stderr, "Card %d dropped %d frame(s) (before timecode 0x%04x), inserting silence.\n",
 385                                 card_index, dropped_frames, timecode);
 386                         vector<float> silence(silence_samples * 2, 0.0f);
 387                         for (int i = 0; i < dropped_frames; ++i) {
 388                                 card->resampling_queue->add_input_samples(local_pts / double(TIMEBASE), silence.data(), silence_samples);
 389                                 // Note that if the format changed in the meantime, we have
 390                                 // no way of detecting that; we just have to assume the frame length
 391                                 // is always the same.
 392                                 local_pts += frame_length;
 393                         }
 394                 }
 395                 if (num_samples == 0) {
 396                         audio.resize(silence_samples * 2);
 397                         num_samples = silence_samples;
 398                 }
 399                 card->resampling_queue->add_input_samples(local_pts / double(TIMEBASE), audio.data(), num_samples);
 400                 card->next_local_pts = local_pts + frame_length;
 401         }
 402
 403         card->last_timecode = timecode;
 404
 405         // Done with the audio, so release it.
 406         if (audio_frame.owner) {
 407                 audio_frame.owner->release_frame(audio_frame);
 408         }
 409
 410         {
 411                 // Wait until the previous frame was consumed.
 412                 unique_lock<mutex> lock(bmusb_mutex);
 413                 card->new_data_ready_changed.wait(lock, [card]{ return !card->new_data_ready || card->should_quit; });
 414                 if (card->should_quit) return;
 415         }
 416
 417         size_t expected_length = video_format.width * (video_format.height + video_format.extra_lines_top + video_format.extra_lines_bottom) * 2;
 418         if (video_frame.len - video_offset == 0 ||
 419             video_frame.len - video_offset != expected_length) {
 420                 if (video_frame.len != 0) {
 421                         printf("Card %d: Dropping video frame with wrong length (%ld; expected %ld)\n",
 422                                 card_index, video_frame.len - video_offset, expected_length);
 423                 }
 424                 if (video_frame.owner) {
 425                         video_frame.owner->release_frame(video_frame);
 426                 }
 427
 428                 // Still send on the information that we _had_ a frame, even though it's corrupted,
 429                 // so that pts can go up accordingly.
 430                 {
 431                         unique_lock<mutex> lock(bmusb_mutex);
 432                         card->new_data_ready = true;
 433                         card->new_frame = RefCountedFrame(FrameAllocator::Frame());
 434                         card->new_frame_length = frame_length;
 435                         card->new_frame_interlaced = false;
 436                         card->new_data_ready_fence = nullptr;
 437                         card->dropped_frames = dropped_frames;
 438                         card->new_data_ready_changed.notify_all();
 439                 }
 440                 return;
 441         }
 442
 443         PBOFrameAllocator::Userdata *userdata = (PBOFrameAllocator::Userdata *)video_frame.userdata;
 444
 445         unsigned num_fields = video_format.interlaced ? 2 : 1;
 446         timespec frame_upload_start;
 447         if (video_format.interlaced) {
 448                 // Send the two fields along as separate frames; the other side will need to add
 449                 // a deinterlacer to actually get this right.
 450                 assert(video_format.height % 2 == 0);
 451                 video_format.height /= 2;
 452                 assert(frame_length % 2 == 0);
 453                 frame_length /= 2;
 454                 num_fields = 2;
 455                 clock_gettime(CLOCK_MONOTONIC, &frame_upload_start);
 456         }
 457         userdata->last_interlaced = video_format.interlaced;
 458         userdata->last_has_signal = video_format.has_signal;
 459         userdata->last_frame_rate_nom = video_format.frame_rate_nom;
 460         userdata->last_frame_rate_den = video_format.frame_rate_den;
 461         RefCountedFrame new_frame(video_frame);
 462
 463         // Upload the textures.
 464         size_t cbcr_width = video_format.width / 2;
 465         size_t cbcr_offset = video_offset / 2;
 466         size_t y_offset = video_frame.size / 2 + video_offset / 2;
 467
 468         for (unsigned field = 0; field < num_fields; ++field) {
 469                 unsigned field_start_line = (field == 1) ? video_format.second_field_start : video_format.extra_lines_top + field * (video_format.height + 22);
 470
 471                 if (userdata->tex_y[field] == 0 ||
 472                     userdata->tex_cbcr[field] == 0 ||
 473                     video_format.width != userdata->last_width[field] ||
 474                     video_format.height != userdata->last_height[field]) {
 475                         // We changed resolution since last use of this texture, so we need to create
 476                         // a new object. Note that this each card has its own PBOFrameAllocator,
 477                         // we don't need to worry about these flip-flopping between resolutions.
 478                         glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr[field]);
 479                         check_error();
 480                         glTexImage2D(GL_TEXTURE_2D, 0, GL_RG8, cbcr_width, video_format.height, 0, GL_RG, GL_UNSIGNED_BYTE, nullptr);
 481                         check_error();
 482                         glBindTexture(GL_TEXTURE_2D, userdata->tex_y[field]);
 483                         check_error();
 484                         glTexImage2D(GL_TEXTURE_2D, 0, GL_R8, video_format.width, video_format.height, 0, GL_RED, GL_UNSIGNED_BYTE, nullptr);
 485                         check_error();
 486                         userdata->last_width[field] = video_format.width;
 487                         userdata->last_height[field] = video_format.height;
 488                 }
 489
 490                 GLuint pbo = userdata->pbo;
 491                 check_error();
 492                 glBindBuffer(GL_PIXEL_UNPACK_BUFFER_ARB, pbo);
 493                 check_error();
 494                 glMemoryBarrier(GL_CLIENT_MAPPED_BUFFER_BARRIER_BIT);
 495                 check_error();
 496
 497                 glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr[field]);
 498                 check_error();
 499                 glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, cbcr_width, video_format.height, GL_RG, GL_UNSIGNED_BYTE, BUFFER_OFFSET(cbcr_offset + cbcr_width * field_start_line * sizeof(uint16_t)));
 500                 check_error();
 501                 glBindTexture(GL_TEXTURE_2D, userdata->tex_y[field]);
 502                 check_error();
 503                 glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, video_format.width, video_format.height, GL_RED, GL_UNSIGNED_BYTE, BUFFER_OFFSET(y_offset + video_format.width * field_start_line));
 504                 check_error();
 505                 glBindTexture(GL_TEXTURE_2D, 0);
 506                 check_error();
 507                 glBindBuffer(GL_PIXEL_UNPACK_BUFFER_ARB, 0);
 508                 check_error();
 509                 GLsync fence = glFenceSync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
 510                 check_error();
 511                 assert(fence != nullptr);
 512
 513                 if (field == 1) {
 514                         // Don't upload the second field as fast as we can; wait until
 515                         // the field time has approximately passed. (Otherwise, we could
 516                         // get timing jitter against the other sources, and possibly also
 517                         // against the video display, although the latter is not as critical.)
 518                         // This requires our system clock to be reasonably close to the
 519                         // video clock, but that's not an unreasonable assumption.
 520                         timespec second_field_start;
 521                         second_field_start.tv_nsec = frame_upload_start.tv_nsec +
 522                                 frame_length * 1000000000 / TIMEBASE;
 523                         second_field_start.tv_sec = frame_upload_start.tv_sec +
 524                                 second_field_start.tv_nsec / 1000000000;
 525                         second_field_start.tv_nsec %= 1000000000;
 526
 527                         while (clock_nanosleep(CLOCK_MONOTONIC, TIMER_ABSTIME,
 528                                                &second_field_start, nullptr) == -1 &&
 529                                errno == EINTR) ;
 530                 }
 531
 532                 {
 533                         unique_lock<mutex> lock(bmusb_mutex);
 534                         card->new_data_ready = true;
 535                         card->new_frame = new_frame;
 536                         card->new_frame_length = frame_length;
 537                         card->new_frame_field = field;
 538                         card->new_frame_interlaced = video_format.interlaced;
 539                         card->new_data_ready_fence = fence;
 540                         card->dropped_frames = dropped_frames;
 541                         card->new_data_ready_changed.notify_all();
 542
 543                         if (field != num_fields - 1) {
 544                                 // Wait until the previous frame was consumed.
 545                                 card->new_data_ready_changed.wait(lock, [card]{ return !card->new_data_ready || card->should_quit; });
 546                                 if (card->should_quit) return;
 547                         }
 548                 }
 549         }
 550 }
 551
 552 void Mixer::thread_func()
 553 {
 554         eglBindAPI(EGL_OPENGL_API);
 555         QOpenGLContext *context = create_context(mixer_surface);
 556         if (!make_current(context, mixer_surface)) {
 557                 printf("oops\n");
 558                 exit(1);
 559         }
 560
 561         struct timespec start, now;
 562         clock_gettime(CLOCK_MONOTONIC, &start);
 563
 564         int frame = 0;
 565         int stats_dropped_frames = 0;
 566
 567         while (!should_quit) {
 568                 CaptureCard card_copy[MAX_CARDS];
 569                 int num_samples[MAX_CARDS];
 570
 571                 {
 572                         unique_lock<mutex> lock(bmusb_mutex);
 573
 574                         // The first card is the master timer, so wait for it to have a new frame.
 575                         // TODO: Make configurable, and with a timeout.
 576                         cards[0].new_data_ready_changed.wait(lock, [this]{ return cards[0].new_data_ready; });
 577
 578                         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 579                                 CaptureCard *card = &cards[card_index];
 580                                 card_copy[card_index].new_data_ready = card->new_data_ready;
 581                                 card_copy[card_index].new_frame = card->new_frame;
 582                                 card_copy[card_index].new_frame_length = card->new_frame_length;
 583                                 card_copy[card_index].new_frame_field = card->new_frame_field;
 584                                 card_copy[card_index].new_frame_interlaced = card->new_frame_interlaced;
 585                                 card_copy[card_index].new_data_ready_fence = card->new_data_ready_fence;
 586                                 card_copy[card_index].dropped_frames = card->dropped_frames;
 587                                 card->new_data_ready = false;
 588                                 card->new_data_ready_changed.notify_all();
 589
 590                                 int num_samples_times_timebase = OUTPUT_FREQUENCY * card->new_frame_length + card->fractional_samples;
 591                                 num_samples[card_index] = num_samples_times_timebase / TIMEBASE;
 592                                 card->fractional_samples = num_samples_times_timebase % TIMEBASE;
 593                                 assert(num_samples[card_index] >= 0);
 594                         }
 595                 }
 596
 597                 // Resample the audio as needed, including from previously dropped frames.
 598                 assert(num_cards > 0);
 599                 for (unsigned frame_num = 0; frame_num < card_copy[0].dropped_frames + 1; ++frame_num) {
 600                         {
 601                                 // Signal to the audio thread to process this frame.
 602                                 unique_lock<mutex> lock(audio_mutex);
 603                                 audio_task_queue.push(AudioTask{pts_int, num_samples[0]});
 604                                 audio_task_queue_changed.notify_one();
 605                         }
 606                         if (frame_num != card_copy[0].dropped_frames) {
 607                                 // For dropped frames, increase the pts. Note that if the format changed
 608                                 // in the meantime, we have no way of detecting that; we just have to
 609                                 // assume the frame length is always the same.
 610                                 ++stats_dropped_frames;
 611                                 pts_int += card_copy[0].new_frame_length;
 612                         }
 613                 }
 614
 615                 if (audio_level_callback != nullptr) {
 616                         unique_lock<mutex> lock(compressor_mutex);
 617                         double loudness_s = r128.loudness_S();
 618                         double loudness_i = r128.integrated();
 619                         double loudness_range_low = r128.range_min();
 620                         double loudness_range_high = r128.range_max();
 621
 622                         audio_level_callback(loudness_s, 20.0 * log10(peak),
 623                                              loudness_i, loudness_range_low, loudness_range_high,
 624                                              gain_staging_db, 20.0 * log10(final_makeup_gain),
 625                                              correlation.get_correlation());
 626                 }
 627
 628                 for (unsigned card_index = 1; card_index < num_cards; ++card_index) {
 629                         if (card_copy[card_index].new_data_ready && card_copy[card_index].new_frame->len == 0) {
 630                                 ++card_copy[card_index].dropped_frames;
 631                         }
 632                         if (card_copy[card_index].dropped_frames > 0) {
 633                                 printf("Card %u dropped %d frames before this\n",
 634                                         card_index, int(card_copy[card_index].dropped_frames));
 635                         }
 636                 }
 637
 638                 // If the first card is reporting a corrupted or otherwise dropped frame,
 639                 // just increase the pts (skipping over this frame) and don't try to compute anything new.
 640                 if (card_copy[0].new_frame->len == 0) {
 641                         ++stats_dropped_frames;
 642                         pts_int += card_copy[0].new_frame_length;
 643                         continue;
 644                 }
 645
 646                 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 647                         CaptureCard *card = &card_copy[card_index];
 648                         if (!card->new_data_ready || card->new_frame->len == 0)
 649                                 continue;
 650
 651                         assert(card->new_frame != nullptr);
 652                         insert_new_frame(card->new_frame, card->new_frame_field, card->new_frame_interlaced, card_index, &input_state);
 653                         check_error();
 654
 655                         // The new texture might still be uploaded,
 656                         // tell the GPU to wait until it's there.
 657                         if (card->new_data_ready_fence) {
 658                                 glWaitSync(card->new_data_ready_fence, /*flags=*/0, GL_TIMEOUT_IGNORED);
 659                                 check_error();
 660                                 glDeleteSync(card->new_data_ready_fence);
 661                                 check_error();
 662                         }
 663                 }
 664
 665                 // Get the main chain from the theme, and set its state immediately.
 666                 Theme::Chain theme_main_chain = theme->get_chain(0, pts(), WIDTH, HEIGHT, input_state);
 667                 EffectChain *chain = theme_main_chain.chain;
 668                 theme_main_chain.setup_chain();
 669                 //theme_main_chain.chain->enable_phase_timing(true);
 670
 671                 GLuint y_tex, cbcr_tex;
 672                 bool got_frame = h264_encoder->begin_frame(&y_tex, &cbcr_tex);
 673                 assert(got_frame);
 674
 675                 // Render main chain.
 676                 GLuint cbcr_full_tex = resource_pool->create_2d_texture(GL_RG8, WIDTH, HEIGHT);
 677                 GLuint rgba_tex = resource_pool->create_2d_texture(GL_RGB565, WIDTH, HEIGHT);  // Saves texture bandwidth, although dithering gets messed up.
 678                 GLuint fbo = resource_pool->create_fbo(y_tex, cbcr_full_tex, rgba_tex);
 679                 check_error();
 680                 chain->render_to_fbo(fbo, WIDTH, HEIGHT);
 681                 resource_pool->release_fbo(fbo);
 682
 683                 subsample_chroma(cbcr_full_tex, cbcr_tex);
 684                 resource_pool->release_2d_texture(cbcr_full_tex);
 685
 686                 // Set the right state for rgba_tex.
 687                 glBindFramebuffer(GL_FRAMEBUFFER, 0);
 688                 glBindTexture(GL_TEXTURE_2D, rgba_tex);
 689                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
 690                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
 691                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
 692
 693                 RefCountedGLsync fence(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
 694                 check_error();
 695
 696                 const int64_t av_delay = TIMEBASE / 10;  // Corresponds to the fixed delay in resampling_queue.h. TODO: Make less hard-coded.
 697                 h264_encoder->end_frame(fence, pts_int + av_delay, theme_main_chain.input_frames);
 698                 ++frame;
 699                 pts_int += card_copy[0].new_frame_length;
 700
 701                 // The live frame just shows the RGBA texture we just rendered.
 702                 // It owns rgba_tex now.
 703                 DisplayFrame live_frame;
 704                 live_frame.chain = display_chain.get();
 705                 live_frame.setup_chain = [this, rgba_tex]{
 706                         display_input->set_texture_num(rgba_tex);
 707                 };
 708                 live_frame.ready_fence = fence;
 709                 live_frame.input_frames = {};
 710                 live_frame.temp_textures = { rgba_tex };
 711                 output_channel[OUTPUT_LIVE].output_frame(live_frame);
 712
 713                 // Set up preview and any additional channels.
 714                 for (int i = 1; i < theme->get_num_channels() + 2; ++i) {
 715                         DisplayFrame display_frame;
 716                         Theme::Chain chain = theme->get_chain(i, pts(), WIDTH, HEIGHT, input_state);  // FIXME: dimensions
 717                         display_frame.chain = chain.chain;
 718                         display_frame.setup_chain = chain.setup_chain;
 719                         display_frame.ready_fence = fence;
 720                         display_frame.input_frames = chain.input_frames;
 721                         display_frame.temp_textures = {};
 722                         output_channel[i].output_frame(display_frame);
 723                 }
 724
 725                 clock_gettime(CLOCK_MONOTONIC, &now);
 726                 double elapsed = now.tv_sec - start.tv_sec +
 727                         1e-9 * (now.tv_nsec - start.tv_nsec);
 728                 if (frame % 100 == 0) {
 729                         printf("%d frames (%d dropped) in %.3f seconds = %.1f fps (%.1f ms/frame)\n",
 730                                 frame, stats_dropped_frames, elapsed, frame / elapsed,
 731                                 1e3 * elapsed / frame);
 732                 //      chain->print_phase_timing();
 733                 }
 734
 735                 if (should_cut.exchange(false)) {  // Test and clear.
 736                         string filename = generate_local_dump_filename(frame);
 737                         printf("Starting new recording: %s\n", filename.c_str());
 738                         h264_encoder->shutdown();
 739                         httpd.close_output_file();
 740                         httpd.open_output_file(filename.c_str());
 741                         h264_encoder.reset(new H264Encoder(h264_encoder_surface, global_flags.va_display, WIDTH, HEIGHT, &httpd));
 742                 }
 743
 744 #if 0
 745                 // Reset every 100 frames, so that local variations in frame times
 746                 // (especially for the first few frames, when the shaders are
 747                 // compiled etc.) don't make it hard to measure for the entire
 748                 // remaining duration of the program.
 749                 if (frame == 10000) {
 750                         frame = 0;
 751                         start = now;
 752                 }
 753 #endif
 754                 check_error();
 755         }
 756
 757         resource_pool->clean_context();
 758 }
 759
 760 void Mixer::audio_thread_func()
 761 {
 762         while (!should_quit) {
 763                 AudioTask task;
 764
 765                 {
 766                         unique_lock<mutex> lock(audio_mutex);
 767                         audio_task_queue_changed.wait(lock, [this]{ return !audio_task_queue.empty(); });
 768                         task = audio_task_queue.front();
 769                         audio_task_queue.pop();
 770                 }
 771
 772                 process_audio_one_frame(task.pts_int, task.num_samples);
 773         }
 774 }
 775
 776 void Mixer::process_audio_one_frame(int64_t frame_pts_int, int num_samples)
 777 {
 778         vector<float> samples_card;
 779         vector<float> samples_out;
 780
 781         // TODO: Allow mixing audio from several sources.
 782         unsigned selected_audio_card = theme->map_signal(audio_source_channel);
 783         assert(selected_audio_card < num_cards);
 784
 785         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 786                 samples_card.resize(num_samples * 2);
 787                 {
 788                         unique_lock<mutex> lock(cards[card_index].audio_mutex);
 789                         if (!cards[card_index].resampling_queue->get_output_samples(double(frame_pts_int) / TIMEBASE, &samples_card[0], num_samples)) {
 790                                 printf("Card %d reported previous underrun.\n", card_index);
 791                         }
 792                 }
 793                 if (card_index == selected_audio_card) {
 794                         samples_out = move(samples_card);
 795                 }
 796         }
 797
 798         // Cut away everything under 120 Hz (or whatever the cutoff is);
 799         // we don't need it for voice, and it will reduce headroom
 800         // and confuse the compressor. (In particular, any hums at 50 or 60 Hz
 801         // should be dampened.)
 802         if (locut_enabled) {
 803                 locut.render(samples_out.data(), samples_out.size() / 2, locut_cutoff_hz * 2.0 * M_PI / OUTPUT_FREQUENCY, 0.5f);
 804         }
 805
 806         // Apply a level compressor to get the general level right.
 807         // Basically, if it's over about -40 dBFS, we squeeze it down to that level
 808         // (or more precisely, near it, since we don't use infinite ratio),
 809         // then apply a makeup gain to get it to -14 dBFS. -14 dBFS is, of course,
 810         // entirely arbitrary, but from practical tests with speech, it seems to
 811         // put ut around -23 LUFS, so it's a reasonable starting point for later use.
 812         {
 813                 unique_lock<mutex> lock(compressor_mutex);
 814                 if (level_compressor_enabled) {
 815                         float threshold = 0.01f;   // -40 dBFS.
 816                         float ratio = 20.0f;
 817                         float attack_time = 0.5f;
 818                         float release_time = 20.0f;
 819                         float makeup_gain = pow(10.0f, (ref_level_dbfs - (-40.0f)) / 20.0f);  // +26 dB.
 820                         level_compressor.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 821                         gain_staging_db = 20.0 * log10(level_compressor.get_attenuation() * makeup_gain);
 822                 } else {
 823                         // Just apply the gain we already had.
 824                         float g = pow(10.0f, gain_staging_db / 20.0f);
 825                         for (size_t i = 0; i < samples_out.size(); ++i) {
 826                                 samples_out[i] *= g;
 827                         }
 828                 }
 829         }
 830
 831 #if 0
 832         printf("level=%f (%+5.2f dBFS) attenuation=%f (%+5.2f dB) end_result=%+5.2f dB\n",
 833                 level_compressor.get_level(), 20.0 * log10(level_compressor.get_level()),
 834                 level_compressor.get_attenuation(), 20.0 * log10(level_compressor.get_attenuation()),
 835                 20.0 * log10(level_compressor.get_level() * level_compressor.get_attenuation() * makeup_gain));
 836 #endif
 837
 838 //      float limiter_att, compressor_att;
 839
 840         // The real compressor.
 841         if (compressor_enabled) {
 842                 float threshold = pow(10.0f, compressor_threshold_dbfs / 20.0f);
 843                 float ratio = 20.0f;
 844                 float attack_time = 0.005f;
 845                 float release_time = 0.040f;
 846                 float makeup_gain = 2.0f;  // +6 dB.
 847                 compressor.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 848 //              compressor_att = compressor.get_attenuation();
 849         }
 850
 851         // Finally a limiter at -4 dB (so, -10 dBFS) to take out the worst peaks only.
 852         // Note that since ratio is not infinite, we could go slightly higher than this.
 853         if (limiter_enabled) {
 854                 float threshold = pow(10.0f, limiter_threshold_dbfs / 20.0f);
 855                 float ratio = 30.0f;
 856                 float attack_time = 0.0f;  // Instant.
 857                 float release_time = 0.020f;
 858                 float makeup_gain = 1.0f;  // 0 dB.
 859                 limiter.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 860 //              limiter_att = limiter.get_attenuation();
 861         }
 862
 863 //      printf("limiter=%+5.1f  compressor=%+5.1f\n", 20.0*log10(limiter_att), 20.0*log10(compressor_att));
 864
 865         // Upsample 4x to find interpolated peak.
 866         peak_resampler.inp_data = samples_out.data();
 867         peak_resampler.inp_count = samples_out.size() / 2;
 868
 869         vector<float> interpolated_samples_out;
 870         interpolated_samples_out.resize(samples_out.size());
 871         while (peak_resampler.inp_count > 0) {  // About four iterations.
 872                 peak_resampler.out_data = &interpolated_samples_out[0];
 873                 peak_resampler.out_count = interpolated_samples_out.size() / 2;
 874                 peak_resampler.process();
 875                 size_t out_stereo_samples = interpolated_samples_out.size() / 2 - peak_resampler.out_count;
 876                 peak = max<float>(peak, find_peak(interpolated_samples_out.data(), out_stereo_samples * 2));
 877                 peak_resampler.out_data = nullptr;
 878         }
 879
 880         // At this point, we are most likely close to +0 LU, but all of our
 881         // measurements have been on raw sample values, not R128 values.
 882         // So we have a final makeup gain to get us to +0 LU; the gain
 883         // adjustments required should be relatively small, and also, the
 884         // offset shouldn't change much (only if the type of audio changes
 885         // significantly). Thus, we shoot for updating this value basically
 886         // “whenever we process buffers”, since the R128 calculation isn't exactly
 887         // something we get out per-sample.
 888         //
 889         // Note that there's a feedback loop here, so we choose a very slow filter
 890         // (half-time of 100 seconds).
 891         double target_loudness_factor, alpha;
 892         {
 893                 unique_lock<mutex> lock(compressor_mutex);
 894                 double loudness_lu = r128.loudness_M() - ref_level_lufs;
 895                 double current_makeup_lu = 20.0f * log10(final_makeup_gain);
 896                 target_loudness_factor = pow(10.0f, -loudness_lu / 20.0f);
 897
 898                 // If we're outside +/- 5 LU uncorrected, we don't count it as
 899                 // a normal signal (probably silence) and don't change the
 900                 // correction factor; just apply what we already have.
 901                 if (fabs(loudness_lu - current_makeup_lu) >= 5.0 || !final_makeup_gain_auto) {
 902                         alpha = 0.0;
 903                 } else {
 904                         // Formula adapted from
 905                         // https://en.wikipedia.org/wiki/Low-pass_filter#Simple_infinite_impulse_response_filter.
 906                         const double half_time_s = 100.0;
 907                         const double fc_mul_2pi_delta_t = 1.0 / (half_time_s * OUTPUT_FREQUENCY);
 908                         alpha = fc_mul_2pi_delta_t / (fc_mul_2pi_delta_t + 1.0);
 909                 }
 910
 911                 double m = final_makeup_gain;
 912                 for (size_t i = 0; i < samples_out.size(); i += 2) {
 913                         samples_out[i + 0] *= m;
 914                         samples_out[i + 1] *= m;
 915                         m += (target_loudness_factor - m) * alpha;
 916                 }
 917                 final_makeup_gain = m;
 918         }
 919
 920         // Find R128 levels and L/R correlation.
 921         vector<float> left, right;
 922         deinterleave_samples(samples_out, &left, &right);
 923         float *ptrs[] = { left.data(), right.data() };
 924         {
 925                 unique_lock<mutex> lock(compressor_mutex);
 926                 r128.process(left.size(), ptrs);
 927                 correlation.process_samples(samples_out);
 928         }
 929
 930         // Send the samples to the sound card.
 931         if (alsa) {
 932                 alsa->write(samples_out);
 933         }
 934
 935         // And finally add them to the output.
 936         h264_encoder->add_audio(frame_pts_int, move(samples_out));
 937 }
 938
 939 void Mixer::subsample_chroma(GLuint src_tex, GLuint dst_tex)
 940 {
 941         GLuint vao;
 942         glGenVertexArrays(1, &vao);
 943         check_error();
 944
 945         glBindVertexArray(vao);
 946         check_error();
 947
 948         // Extract Cb/Cr.
 949         GLuint fbo = resource_pool->create_fbo(dst_tex);
 950         glBindFramebuffer(GL_FRAMEBUFFER, fbo);
 951         glViewport(0, 0, WIDTH/2, HEIGHT/2);
 952         check_error();
 953
 954         glUseProgram(cbcr_program_num);
 955         check_error();
 956
 957         glActiveTexture(GL_TEXTURE0);
 958         check_error();
 959         glBindTexture(GL_TEXTURE_2D, src_tex);
 960         check_error();
 961         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
 962         check_error();
 963         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
 964         check_error();
 965         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
 966         check_error();
 967
 968         float chroma_offset_0[] = { -0.5f / WIDTH, 0.0f };
 969         set_uniform_vec2(cbcr_program_num, "foo", "chroma_offset_0", chroma_offset_0);
 970
 971         glBindBuffer(GL_ARRAY_BUFFER, cbcr_vbo);
 972         check_error();
 973
 974         for (GLint attr_index : { cbcr_position_attribute_index, cbcr_texcoord_attribute_index }) {
 975                 glEnableVertexAttribArray(attr_index);
 976                 check_error();
 977                 glVertexAttribPointer(attr_index, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
 978                 check_error();
 979         }
 980
 981         glDrawArrays(GL_TRIANGLES, 0, 3);
 982         check_error();
 983
 984         for (GLint attr_index : { cbcr_position_attribute_index, cbcr_texcoord_attribute_index }) {
 985                 glDisableVertexAttribArray(attr_index);
 986                 check_error();
 987         }
 988
 989         glUseProgram(0);
 990         check_error();
 991         glBindFramebuffer(GL_FRAMEBUFFER, 0);
 992         check_error();
 993
 994         resource_pool->release_fbo(fbo);
 995         glDeleteVertexArrays(1, &vao);
 996 }
 997
 998 void Mixer::release_display_frame(DisplayFrame *frame)
 999 {
1000         for (GLuint texnum : frame->temp_textures) {
1001                 resource_pool->release_2d_texture(texnum);
1002         }
1003         frame->temp_textures.clear();
1004         frame->ready_fence.reset();
1005         frame->input_frames.clear();
1006 }
1007
1008 void Mixer::start()
1009 {
1010         mixer_thread = thread(&Mixer::thread_func, this);
1011         audio_thread = thread(&Mixer::audio_thread_func, this);
1012 }
1013
1014 void Mixer::quit()
1015 {
1016         should_quit = true;
1017         mixer_thread.join();
1018         audio_thread.join();
1019 }
1020
1021 void Mixer::transition_clicked(int transition_num)
1022 {
1023         theme->transition_clicked(transition_num, pts());
1024 }
1025
1026 void Mixer::channel_clicked(int preview_num)
1027 {
1028         theme->channel_clicked(preview_num);
1029 }
1030
1031 void Mixer::reset_meters()
1032 {
1033         peak_resampler.reset();
1034         peak = 0.0f;
1035         r128.reset();
1036         r128.integr_start();
1037         correlation.reset();
1038 }
1039
1040 Mixer::OutputChannel::~OutputChannel()
1041 {
1042         if (has_current_frame) {
1043                 parent->release_display_frame(&current_frame);
1044         }
1045         if (has_ready_frame) {
1046                 parent->release_display_frame(&ready_frame);
1047         }
1048 }
1049
1050 void Mixer::OutputChannel::output_frame(DisplayFrame frame)
1051 {
1052         // Store this frame for display. Remove the ready frame if any
1053         // (it was seemingly never used).
1054         {
1055                 unique_lock<mutex> lock(frame_mutex);
1056                 if (has_ready_frame) {
1057                         parent->release_display_frame(&ready_frame);
1058                 }
1059                 ready_frame = frame;
1060                 has_ready_frame = true;
1061         }
1062
1063         if (has_new_frame_ready_callback) {
1064                 new_frame_ready_callback();
1065         }
1066 }
1067
1068 bool Mixer::OutputChannel::get_display_frame(DisplayFrame *frame)
1069 {
1070         unique_lock<mutex> lock(frame_mutex);
1071         if (!has_current_frame && !has_ready_frame) {
1072                 return false;
1073         }
1074
1075         if (has_current_frame && has_ready_frame) {
1076                 // We have a new ready frame. Toss the current one.
1077                 parent->release_display_frame(&current_frame);
1078                 has_current_frame = false;
1079         }
1080         if (has_ready_frame) {
1081                 assert(!has_current_frame);
1082                 current_frame = ready_frame;
1083                 ready_frame.ready_fence.reset();  // Drop the refcount.
1084                 ready_frame.input_frames.clear();  // Drop the refcounts.
1085                 has_current_frame = true;
1086                 has_ready_frame = false;
1087         }
1088
1089         *frame = current_frame;
1090         return true;
1091 }
1092
1093 void Mixer::OutputChannel::set_frame_ready_callback(Mixer::new_frame_ready_callback_t callback)
1094 {
1095         new_frame_ready_callback = callback;
1096         has_new_frame_ready_callback = true;
1097 }