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 "video_encoder.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 }  // namespace
  96
  97 void QueueLengthPolicy::update_policy(int queue_length)
  98 {
  99         if (queue_length < 0) {  // Starvation.
 100                 if (been_at_safe_point_since_last_starvation && safe_queue_length < 5) {
 101                         ++safe_queue_length;
 102                         fprintf(stderr, "Card %u: Starvation, increasing safe limit to %u frames\n",
 103                                 card_index, safe_queue_length);
 104                 }
 105                 frames_with_at_least_one = 0;
 106                 been_at_safe_point_since_last_starvation = false;
 107                 return;
 108         }
 109         if (queue_length > 0) {
 110                 if (queue_length >= int(safe_queue_length)) {
 111                         been_at_safe_point_since_last_starvation = true;
 112                 }
 113                 if (++frames_with_at_least_one >= 1000 && safe_queue_length > 0) {
 114                         --safe_queue_length;
 115                         fprintf(stderr, "Card %u: Spare frames for more than 1000 frames, reducing safe limit to %u frames\n",
 116                                 card_index, safe_queue_length);
 117                         frames_with_at_least_one = 0;
 118                 }
 119         } else {
 120                 frames_with_at_least_one = 0;
 121         }
 122 }
 123
 124 Mixer::Mixer(const QSurfaceFormat &format, unsigned num_cards)
 125         : httpd(),
 126           num_cards(num_cards),
 127           mixer_surface(create_surface(format)),
 128           h264_encoder_surface(create_surface(format)),
 129           correlation(OUTPUT_FREQUENCY),
 130           level_compressor(OUTPUT_FREQUENCY),
 131           limiter(OUTPUT_FREQUENCY),
 132           compressor(OUTPUT_FREQUENCY)
 133 {
 134         CHECK(init_movit(MOVIT_SHADER_DIR, MOVIT_DEBUG_OFF));
 135         check_error();
 136
 137         // Since we allow non-bouncing 4:2:2 YCbCrInputs, effective subpixel precision
 138         // will be halved when sampling them, and we need to compensate here.
 139         movit_texel_subpixel_precision /= 2.0;
 140
 141         resource_pool.reset(new ResourcePool);
 142         theme.reset(new Theme(global_flags.theme_filename.c_str(), resource_pool.get(), num_cards));
 143         for (unsigned i = 0; i < NUM_OUTPUTS; ++i) {
 144                 output_channel[i].parent = this;
 145                 output_channel[i].channel = i;
 146         }
 147
 148         ImageFormat inout_format;
 149         inout_format.color_space = COLORSPACE_sRGB;
 150         inout_format.gamma_curve = GAMMA_sRGB;
 151
 152         // Display chain; shows the live output produced by the main chain (its RGBA version).
 153         display_chain.reset(new EffectChain(WIDTH, HEIGHT, resource_pool.get()));
 154         check_error();
 155         display_input = new FlatInput(inout_format, FORMAT_RGB, GL_UNSIGNED_BYTE, WIDTH, HEIGHT);  // FIXME: GL_UNSIGNED_BYTE is really wrong.
 156         display_chain->add_input(display_input);
 157         display_chain->add_output(inout_format, OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED);
 158         display_chain->set_dither_bits(0);  // Don't bother.
 159         display_chain->finalize();
 160
 161         video_encoder.reset(new VideoEncoder(resource_pool.get(), h264_encoder_surface, global_flags.va_display, WIDTH, HEIGHT, &httpd));
 162
 163         // Start listening for clients only once VideoEncoder has written its header, if any.
 164         httpd.start(9095);
 165
 166         // First try initializing the PCI devices, then USB, until we have the desired number of cards.
 167         unsigned num_pci_devices = 0, num_usb_devices = 0;
 168         unsigned card_index = 0;
 169
 170         IDeckLinkIterator *decklink_iterator = CreateDeckLinkIteratorInstance();
 171         if (decklink_iterator != nullptr) {
 172                 for ( ; card_index < num_cards; ++card_index) {
 173                         IDeckLink *decklink;
 174                         if (decklink_iterator->Next(&decklink) != S_OK) {
 175                                 break;
 176                         }
 177
 178                         configure_card(card_index, format, new DeckLinkCapture(decklink, card_index));
 179                         ++num_pci_devices;
 180                 }
 181                 decklink_iterator->Release();
 182                 fprintf(stderr, "Found %d DeckLink PCI card(s).\n", num_pci_devices);
 183         } else {
 184                 fprintf(stderr, "DeckLink drivers not found. Probing for USB cards only.\n");
 185         }
 186         for ( ; card_index < num_cards; ++card_index) {
 187                 configure_card(card_index, format, new BMUSBCapture(card_index - num_pci_devices));
 188                 ++num_usb_devices;
 189         }
 190
 191         if (num_usb_devices > 0) {
 192                 BMUSBCapture::start_bm_thread();
 193         }
 194
 195         for (card_index = 0; card_index < num_cards; ++card_index) {
 196                 cards[card_index].queue_length_policy.reset(card_index);
 197                 cards[card_index].capture->start_bm_capture();
 198         }
 199
 200         // Set up stuff for NV12 conversion.
 201
 202         // Cb/Cr shader.
 203         string cbcr_vert_shader =
 204                 "#version 130 \n"
 205                 " \n"
 206                 "in vec2 position; \n"
 207                 "in vec2 texcoord; \n"
 208                 "out vec2 tc0; \n"
 209                 "uniform vec2 foo_chroma_offset_0; \n"
 210                 " \n"
 211                 "void main() \n"
 212                 "{ \n"
 213                 "    // The result of glOrtho(0.0, 1.0, 0.0, 1.0, 0.0, 1.0) is: \n"
 214                 "    // \n"
 215                 "    //   2.000  0.000  0.000 -1.000 \n"
 216                 "    //   0.000  2.000  0.000 -1.000 \n"
 217                 "    //   0.000  0.000 -2.000 -1.000 \n"
 218                 "    //   0.000  0.000  0.000  1.000 \n"
 219                 "    gl_Position = vec4(2.0 * position.x - 1.0, 2.0 * position.y - 1.0, -1.0, 1.0); \n"
 220                 "    vec2 flipped_tc = texcoord; \n"
 221                 "    tc0 = flipped_tc + foo_chroma_offset_0; \n"
 222                 "} \n";
 223         string cbcr_frag_shader =
 224                 "#version 130 \n"
 225                 "in vec2 tc0; \n"
 226                 "uniform sampler2D cbcr_tex; \n"
 227                 "out vec4 FragColor; \n"
 228                 "void main() { \n"
 229                 "    FragColor = texture(cbcr_tex, tc0); \n"
 230                 "} \n";
 231         vector<string> frag_shader_outputs;
 232         cbcr_program_num = resource_pool->compile_glsl_program(cbcr_vert_shader, cbcr_frag_shader, frag_shader_outputs);
 233
 234         float vertices[] = {
 235                 0.0f, 2.0f,
 236                 0.0f, 0.0f,
 237                 2.0f, 0.0f
 238         };
 239         cbcr_vbo = generate_vbo(2, GL_FLOAT, sizeof(vertices), vertices);
 240         cbcr_position_attribute_index = glGetAttribLocation(cbcr_program_num, "position");
 241         cbcr_texcoord_attribute_index = glGetAttribLocation(cbcr_program_num, "texcoord");
 242
 243         r128.init(2, OUTPUT_FREQUENCY);
 244         r128.integr_start();
 245
 246         locut.init(FILTER_HPF, 2);
 247
 248         // If --flat-audio is given, turn off everything that messes with the sound,
 249         // except the final makeup gain.
 250         if (global_flags.flat_audio) {
 251                 set_locut_enabled(false);
 252                 set_gain_staging_auto(false);
 253                 set_limiter_enabled(false);
 254                 set_compressor_enabled(false);
 255         }
 256
 257         // hlen=16 is pretty low quality, but we use quite a bit of CPU otherwise,
 258         // and there's a limit to how important the peak meter is.
 259         peak_resampler.setup(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY * 4, /*num_channels=*/2, /*hlen=*/16, /*frel=*/1.0);
 260
 261         if (global_flags.enable_alsa_output) {
 262                 alsa.reset(new ALSAOutput(OUTPUT_FREQUENCY, /*num_channels=*/2));
 263         }
 264 }
 265
 266 Mixer::~Mixer()
 267 {
 268         resource_pool->release_glsl_program(cbcr_program_num);
 269         glDeleteBuffers(1, &cbcr_vbo);
 270         BMUSBCapture::stop_bm_thread();
 271
 272         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 273                 {
 274                         unique_lock<mutex> lock(bmusb_mutex);
 275                         cards[card_index].should_quit = true;  // Unblock thread.
 276                         cards[card_index].new_frames_changed.notify_all();
 277                 }
 278                 cards[card_index].capture->stop_dequeue_thread();
 279         }
 280
 281         video_encoder.reset(nullptr);
 282 }
 283
 284 void Mixer::configure_card(unsigned card_index, const QSurfaceFormat &format, CaptureInterface *capture)
 285 {
 286         printf("Configuring card %d...\n", card_index);
 287
 288         CaptureCard *card = &cards[card_index];
 289         card->capture = capture;
 290         card->capture->set_frame_callback(bind(&Mixer::bm_frame, this, card_index, _1, _2, _3, _4, _5, _6, _7));
 291         card->frame_allocator.reset(new PBOFrameAllocator(8 << 20, WIDTH, HEIGHT));  // 8 MB.
 292         card->capture->set_video_frame_allocator(card->frame_allocator.get());
 293         card->surface = create_surface(format);
 294         card->resampling_queue.reset(new ResamplingQueue(card_index, OUTPUT_FREQUENCY, OUTPUT_FREQUENCY, 2));
 295         card->capture->configure_card();
 296 }
 297
 298
 299 namespace {
 300
 301 int unwrap_timecode(uint16_t current_wrapped, int last)
 302 {
 303         uint16_t last_wrapped = last & 0xffff;
 304         if (current_wrapped > last_wrapped) {
 305                 return (last & ~0xffff) | current_wrapped;
 306         } else {
 307                 return 0x10000 + ((last & ~0xffff) | current_wrapped);
 308         }
 309 }
 310
 311 float find_peak(const float *samples, size_t num_samples)
 312 {
 313         float m = fabs(samples[0]);
 314         for (size_t i = 1; i < num_samples; ++i) {
 315                 m = max(m, fabs(samples[i]));
 316         }
 317         return m;
 318 }
 319
 320 void deinterleave_samples(const vector<float> &in, vector<float> *out_l, vector<float> *out_r)
 321 {
 322         size_t num_samples = in.size() / 2;
 323         out_l->resize(num_samples);
 324         out_r->resize(num_samples);
 325
 326         const float *inptr = in.data();
 327         float *lptr = &(*out_l)[0];
 328         float *rptr = &(*out_r)[0];
 329         for (size_t i = 0; i < num_samples; ++i) {
 330                 *lptr++ = *inptr++;
 331                 *rptr++ = *inptr++;
 332         }
 333 }
 334
 335 }  // namespace
 336
 337 void Mixer::bm_frame(unsigned card_index, uint16_t timecode,
 338                      FrameAllocator::Frame video_frame, size_t video_offset, VideoFormat video_format,
 339                      FrameAllocator::Frame audio_frame, size_t audio_offset, AudioFormat audio_format)
 340 {
 341         CaptureCard *card = &cards[card_index];
 342
 343         if (is_mode_scanning[card_index]) {
 344                 if (video_format.has_signal) {
 345                         // Found a stable signal, so stop scanning.
 346                         is_mode_scanning[card_index] = false;
 347                 } else {
 348                         static constexpr double switch_time_s = 0.5;  // Should be enough time for the signal to stabilize.
 349                         timespec now;
 350                         clock_gettime(CLOCK_MONOTONIC, &now);
 351                         double sec_since_last_switch = (now.tv_sec - last_mode_scan_change[card_index].tv_sec) +
 352                                 1e-9 * (now.tv_nsec - last_mode_scan_change[card_index].tv_nsec);
 353                         if (sec_since_last_switch > switch_time_s) {
 354                                 // It isn't this mode; try the next one.
 355                                 mode_scanlist_index[card_index]++;
 356                                 mode_scanlist_index[card_index] %= mode_scanlist[card_index].size();
 357                                 cards[card_index].capture->set_video_mode(mode_scanlist[card_index][mode_scanlist_index[card_index]]);
 358                                 last_mode_scan_change[card_index] = now;
 359                         }
 360                 }
 361         }
 362
 363         int64_t frame_length = int64_t(TIMEBASE) * video_format.frame_rate_den / video_format.frame_rate_nom;
 364
 365         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;
 366         if (num_samples > OUTPUT_FREQUENCY / 10) {
 367                 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",
 368                         card_index, int(audio_frame.len), int(audio_offset),
 369                         timecode, int(video_frame.len), int(video_offset), video_format.id);
 370                 if (video_frame.owner) {
 371                         video_frame.owner->release_frame(video_frame);
 372                 }
 373                 if (audio_frame.owner) {
 374                         audio_frame.owner->release_frame(audio_frame);
 375                 }
 376                 return;
 377         }
 378
 379         int64_t local_pts = card->next_local_pts;
 380         int dropped_frames = 0;
 381         if (card->last_timecode != -1) {
 382                 dropped_frames = unwrap_timecode(timecode, card->last_timecode) - card->last_timecode - 1;
 383         }
 384
 385         // Convert the audio to stereo fp32 and add it.
 386         vector<float> audio;
 387         audio.resize(num_samples * 2);
 388         switch (audio_format.bits_per_sample) {
 389         case 0:
 390                 assert(num_samples == 0);
 391                 break;
 392         case 24:
 393                 convert_fixed24_to_fp32(&audio[0], 2, audio_frame.data + audio_offset, audio_format.num_channels, num_samples);
 394                 break;
 395         case 32:
 396                 convert_fixed32_to_fp32(&audio[0], 2, audio_frame.data + audio_offset, audio_format.num_channels, num_samples);
 397                 break;
 398         default:
 399                 fprintf(stderr, "Cannot handle audio with %u bits per sample\n", audio_format.bits_per_sample);
 400                 assert(false);
 401         }
 402
 403         // Add the audio.
 404         {
 405                 unique_lock<mutex> lock(card->audio_mutex);
 406
 407                 // Number of samples per frame if we need to insert silence.
 408                 // (Could be nonintegral, but resampling will save us then.)
 409                 int silence_samples = OUTPUT_FREQUENCY * video_format.frame_rate_den / video_format.frame_rate_nom;
 410
 411                 if (dropped_frames > MAX_FPS * 2) {
 412                         fprintf(stderr, "Card %d lost more than two seconds (or time code jumping around; from 0x%04x to 0x%04x), resetting resampler\n",
 413                                 card_index, card->last_timecode, timecode);
 414                         card->resampling_queue.reset(new ResamplingQueue(card_index, OUTPUT_FREQUENCY, OUTPUT_FREQUENCY, 2));
 415                         dropped_frames = 0;
 416                 } else if (dropped_frames > 0) {
 417                         // Insert silence as needed.
 418                         fprintf(stderr, "Card %d dropped %d frame(s) (before timecode 0x%04x), inserting silence.\n",
 419                                 card_index, dropped_frames, timecode);
 420                         vector<float> silence(silence_samples * 2, 0.0f);
 421                         for (int i = 0; i < dropped_frames; ++i) {
 422                                 card->resampling_queue->add_input_samples(local_pts / double(TIMEBASE), silence.data(), silence_samples);
 423                                 // Note that if the format changed in the meantime, we have
 424                                 // no way of detecting that; we just have to assume the frame length
 425                                 // is always the same.
 426                                 local_pts += frame_length;
 427                         }
 428                 }
 429                 if (num_samples == 0) {
 430                         audio.resize(silence_samples * 2);
 431                         num_samples = silence_samples;
 432                 }
 433                 card->resampling_queue->add_input_samples(local_pts / double(TIMEBASE), audio.data(), num_samples);
 434                 card->next_local_pts = local_pts + frame_length;
 435         }
 436
 437         card->last_timecode = timecode;
 438
 439         // Done with the audio, so release it.
 440         if (audio_frame.owner) {
 441                 audio_frame.owner->release_frame(audio_frame);
 442         }
 443
 444         size_t expected_length = video_format.width * (video_format.height + video_format.extra_lines_top + video_format.extra_lines_bottom) * 2;
 445         if (video_frame.len - video_offset == 0 ||
 446             video_frame.len - video_offset != expected_length) {
 447                 if (video_frame.len != 0) {
 448                         printf("Card %d: Dropping video frame with wrong length (%ld; expected %ld)\n",
 449                                 card_index, video_frame.len - video_offset, expected_length);
 450                 }
 451                 if (video_frame.owner) {
 452                         video_frame.owner->release_frame(video_frame);
 453                 }
 454
 455                 // Still send on the information that we _had_ a frame, even though it's corrupted,
 456                 // so that pts can go up accordingly.
 457                 {
 458                         unique_lock<mutex> lock(bmusb_mutex);
 459                         CaptureCard::NewFrame new_frame;
 460                         new_frame.frame = RefCountedFrame(FrameAllocator::Frame());
 461                         new_frame.length = frame_length;
 462                         new_frame.interlaced = false;
 463                         new_frame.dropped_frames = dropped_frames;
 464                         card->new_frames.push(move(new_frame));
 465                         card->new_frames_changed.notify_all();
 466                 }
 467                 return;
 468         }
 469
 470         PBOFrameAllocator::Userdata *userdata = (PBOFrameAllocator::Userdata *)video_frame.userdata;
 471
 472         unsigned num_fields = video_format.interlaced ? 2 : 1;
 473         timespec frame_upload_start;
 474         if (video_format.interlaced) {
 475                 // Send the two fields along as separate frames; the other side will need to add
 476                 // a deinterlacer to actually get this right.
 477                 assert(video_format.height % 2 == 0);
 478                 video_format.height /= 2;
 479                 assert(frame_length % 2 == 0);
 480                 frame_length /= 2;
 481                 num_fields = 2;
 482                 clock_gettime(CLOCK_MONOTONIC, &frame_upload_start);
 483         }
 484         userdata->last_interlaced = video_format.interlaced;
 485         userdata->last_has_signal = video_format.has_signal;
 486         userdata->last_frame_rate_nom = video_format.frame_rate_nom;
 487         userdata->last_frame_rate_den = video_format.frame_rate_den;
 488         RefCountedFrame frame(video_frame);
 489
 490         // Upload the textures.
 491         size_t cbcr_width = video_format.width / 2;
 492         size_t cbcr_offset = video_offset / 2;
 493         size_t y_offset = video_frame.size / 2 + video_offset / 2;
 494
 495         for (unsigned field = 0; field < num_fields; ++field) {
 496                 // Put the actual texture upload in a lambda that is executed in the main thread.
 497                 // It is entirely possible to do this in the same thread (and it might even be
 498                 // faster, depending on the GPU and driver), but it appears to be trickling
 499                 // driver bugs very easily.
 500                 //
 501                 // Note that this means we must hold on to the actual frame data in <userdata>
 502                 // until the upload command is run, but we hold on to <frame> much longer than that
 503                 // (in fact, all the way until we no longer use the texture in rendering).
 504                 auto upload_func = [field, video_format, y_offset, cbcr_offset, cbcr_width, userdata]() {
 505                         unsigned field_start_line = (field == 1) ? video_format.second_field_start : video_format.extra_lines_top + field * (video_format.height + 22);
 506
 507                         if (userdata->tex_y[field] == 0 ||
 508                             userdata->tex_cbcr[field] == 0 ||
 509                             video_format.width != userdata->last_width[field] ||
 510                             video_format.height != userdata->last_height[field]) {
 511                                 // We changed resolution since last use of this texture, so we need to create
 512                                 // a new object. Note that this each card has its own PBOFrameAllocator,
 513                                 // we don't need to worry about these flip-flopping between resolutions.
 514                                 glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr[field]);
 515                                 check_error();
 516                                 glTexImage2D(GL_TEXTURE_2D, 0, GL_RG8, cbcr_width, video_format.height, 0, GL_RG, GL_UNSIGNED_BYTE, nullptr);
 517                                 check_error();
 518                                 glBindTexture(GL_TEXTURE_2D, userdata->tex_y[field]);
 519                                 check_error();
 520                                 glTexImage2D(GL_TEXTURE_2D, 0, GL_R8, video_format.width, video_format.height, 0, GL_RED, GL_UNSIGNED_BYTE, nullptr);
 521                                 check_error();
 522                                 userdata->last_width[field] = video_format.width;
 523                                 userdata->last_height[field] = video_format.height;
 524                         }
 525
 526                         GLuint pbo = userdata->pbo;
 527                         check_error();
 528                         glBindBuffer(GL_PIXEL_UNPACK_BUFFER, pbo);
 529                         check_error();
 530
 531                         size_t field_y_start = y_offset + video_format.width * field_start_line;
 532                         size_t field_cbcr_start = cbcr_offset + cbcr_width * field_start_line * sizeof(uint16_t);
 533
 534                         if (global_flags.flush_pbos) {
 535                                 glFlushMappedBufferRange(GL_PIXEL_UNPACK_BUFFER, field_y_start, video_format.width * video_format.height);
 536                                 check_error();
 537                                 glFlushMappedBufferRange(GL_PIXEL_UNPACK_BUFFER, field_cbcr_start, cbcr_width * video_format.height * sizeof(uint16_t));
 538                                 check_error();
 539                         }
 540
 541                         glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr[field]);
 542                         check_error();
 543                         glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, cbcr_width, video_format.height, GL_RG, GL_UNSIGNED_BYTE, BUFFER_OFFSET(field_cbcr_start));
 544                         check_error();
 545                         glBindTexture(GL_TEXTURE_2D, userdata->tex_y[field]);
 546                         check_error();
 547                         glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, video_format.width, video_format.height, GL_RED, GL_UNSIGNED_BYTE, BUFFER_OFFSET(field_y_start));
 548                         check_error();
 549                         glBindTexture(GL_TEXTURE_2D, 0);
 550                         check_error();
 551                         glBindBuffer(GL_PIXEL_UNPACK_BUFFER, 0);
 552                         check_error();
 553                 };
 554
 555                 if (field == 1) {
 556                         // Don't upload the second field as fast as we can; wait until
 557                         // the field time has approximately passed. (Otherwise, we could
 558                         // get timing jitter against the other sources, and possibly also
 559                         // against the video display, although the latter is not as critical.)
 560                         // This requires our system clock to be reasonably close to the
 561                         // video clock, but that's not an unreasonable assumption.
 562                         timespec second_field_start;
 563                         second_field_start.tv_nsec = frame_upload_start.tv_nsec +
 564                                 frame_length * 1000000000 / TIMEBASE;
 565                         second_field_start.tv_sec = frame_upload_start.tv_sec +
 566                                 second_field_start.tv_nsec / 1000000000;
 567                         second_field_start.tv_nsec %= 1000000000;
 568
 569                         while (clock_nanosleep(CLOCK_MONOTONIC, TIMER_ABSTIME,
 570                                                &second_field_start, nullptr) == -1 &&
 571                                errno == EINTR) ;
 572                 }
 573
 574                 {
 575                         unique_lock<mutex> lock(bmusb_mutex);
 576                         CaptureCard::NewFrame new_frame;
 577                         new_frame.frame = frame;
 578                         new_frame.length = frame_length;
 579                         new_frame.field = field;
 580                         new_frame.interlaced = video_format.interlaced;
 581                         new_frame.upload_func = upload_func;
 582                         new_frame.dropped_frames = dropped_frames;
 583                         card->new_frames.push(move(new_frame));
 584                         card->new_frames_changed.notify_all();
 585                 }
 586         }
 587 }
 588
 589 void Mixer::thread_func()
 590 {
 591         eglBindAPI(EGL_OPENGL_API);
 592         QOpenGLContext *context = create_context(mixer_surface);
 593         if (!make_current(context, mixer_surface)) {
 594                 printf("oops\n");
 595                 exit(1);
 596         }
 597
 598         struct timespec start, now;
 599         clock_gettime(CLOCK_MONOTONIC, &start);
 600
 601         int frame = 0;
 602         int stats_dropped_frames = 0;
 603
 604         while (!should_quit) {
 605                 CaptureCard::NewFrame new_frames[MAX_CARDS];
 606                 bool has_new_frame[MAX_CARDS] = { false };
 607                 int num_samples[MAX_CARDS] = { 0 };
 608
 609                 // TODO: Add a timeout.
 610                 unsigned master_card_index = theme->map_signal(master_clock_channel);
 611                 assert(master_card_index < num_cards);
 612
 613                 get_one_frame_from_each_card(master_card_index, new_frames, has_new_frame, num_samples);
 614                 schedule_audio_resampling_tasks(new_frames[master_card_index].dropped_frames, num_samples[master_card_index], new_frames[master_card_index].length);
 615                 stats_dropped_frames += new_frames[master_card_index].dropped_frames;
 616                 send_audio_level_callback();
 617
 618                 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 619                         if (card_index == master_card_index || !has_new_frame[card_index]) {
 620                                 continue;
 621                         }
 622                         if (new_frames[card_index].frame->len == 0) {
 623                                 ++new_frames[card_index].dropped_frames;
 624                         }
 625                         if (new_frames[card_index].dropped_frames > 0) {
 626                                 printf("Card %u dropped %d frames before this\n",
 627                                         card_index, int(new_frames[card_index].dropped_frames));
 628                         }
 629                 }
 630
 631                 // If the first card is reporting a corrupted or otherwise dropped frame,
 632                 // just increase the pts (skipping over this frame) and don't try to compute anything new.
 633                 if (new_frames[master_card_index].frame->len == 0) {
 634                         ++stats_dropped_frames;
 635                         pts_int += new_frames[master_card_index].length;
 636                         continue;
 637                 }
 638
 639                 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 640                         if (!has_new_frame[card_index] || new_frames[card_index].frame->len == 0)
 641                                 continue;
 642
 643                         CaptureCard::NewFrame *new_frame = &new_frames[card_index];
 644                         assert(new_frame->frame != nullptr);
 645                         insert_new_frame(new_frame->frame, new_frame->field, new_frame->interlaced, card_index, &input_state);
 646                         check_error();
 647
 648                         // The new texture might need uploading before use.
 649                         if (new_frame->upload_func) {
 650                                 new_frame->upload_func();
 651                                 new_frame->upload_func = nullptr;
 652                         }
 653                 }
 654
 655                 int64_t duration = new_frames[master_card_index].length;
 656                 render_one_frame(duration);
 657                 ++frame;
 658                 pts_int += duration;
 659
 660                 clock_gettime(CLOCK_MONOTONIC, &now);
 661                 double elapsed = now.tv_sec - start.tv_sec +
 662                         1e-9 * (now.tv_nsec - start.tv_nsec);
 663                 if (frame % 100 == 0) {
 664                         printf("%d frames (%d dropped) in %.3f seconds = %.1f fps (%.1f ms/frame)\n",
 665                                 frame, stats_dropped_frames, elapsed, frame / elapsed,
 666                                 1e3 * elapsed / frame);
 667                 //      chain->print_phase_timing();
 668                 }
 669
 670                 if (should_cut.exchange(false)) {  // Test and clear.
 671                         video_encoder->do_cut(frame);
 672                 }
 673
 674 #if 0
 675                 // Reset every 100 frames, so that local variations in frame times
 676                 // (especially for the first few frames, when the shaders are
 677                 // compiled etc.) don't make it hard to measure for the entire
 678                 // remaining duration of the program.
 679                 if (frame == 10000) {
 680                         frame = 0;
 681                         start = now;
 682                 }
 683 #endif
 684                 check_error();
 685         }
 686
 687         resource_pool->clean_context();
 688 }
 689
 690 void Mixer::get_one_frame_from_each_card(unsigned master_card_index, CaptureCard::NewFrame new_frames[MAX_CARDS], bool has_new_frame[MAX_CARDS], int num_samples[MAX_CARDS])
 691 {
 692         // The first card is the master timer, so wait for it to have a new frame.
 693         unique_lock<mutex> lock(bmusb_mutex);
 694         cards[master_card_index].new_frames_changed.wait(lock, [this, master_card_index]{ return !cards[master_card_index].new_frames.empty(); });
 695
 696         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 697                 CaptureCard *card = &cards[card_index];
 698                 if (card->new_frames.empty()) {
 699                         assert(card_index != master_card_index);
 700                         card->queue_length_policy.update_policy(-1);
 701                         continue;
 702                 }
 703                 new_frames[card_index] = move(card->new_frames.front());
 704                 has_new_frame[card_index] = true;
 705                 card->new_frames.pop();
 706                 card->new_frames_changed.notify_all();
 707
 708                 int num_samples_times_timebase = OUTPUT_FREQUENCY * new_frames[card_index].length + card->fractional_samples;
 709                 num_samples[card_index] = num_samples_times_timebase / TIMEBASE;
 710                 card->fractional_samples = num_samples_times_timebase % TIMEBASE;
 711                 assert(num_samples[card_index] >= 0);
 712
 713                 if (card_index == master_card_index) {
 714                         // We don't use the queue length policy for the master card,
 715                         // but we will if it stops being the master. Thus, clear out
 716                         // the policy in case we switch in the future.
 717                         card->queue_length_policy.reset(card_index);
 718                 } else {
 719                         // If we have excess frames compared to the policy for this card,
 720                         // drop frames from the head.
 721                         card->queue_length_policy.update_policy(card->new_frames.size());
 722                         while (card->new_frames.size() > card->queue_length_policy.get_safe_queue_length()) {
 723                                 card->new_frames.pop();
 724                         }
 725                 }
 726         }
 727 }
 728
 729 void Mixer::schedule_audio_resampling_tasks(unsigned dropped_frames, int num_samples_per_frame, int length_per_frame)
 730 {
 731         // Resample the audio as needed, including from previously dropped frames.
 732         assert(num_cards > 0);
 733         for (unsigned frame_num = 0; frame_num < dropped_frames + 1; ++frame_num) {
 734                 {
 735                         // Signal to the audio thread to process this frame.
 736                         unique_lock<mutex> lock(audio_mutex);
 737                         audio_task_queue.push(AudioTask{pts_int, num_samples_per_frame});
 738                         audio_task_queue_changed.notify_one();
 739                 }
 740                 if (frame_num != dropped_frames) {
 741                         // For dropped frames, increase the pts. Note that if the format changed
 742                         // in the meantime, we have no way of detecting that; we just have to
 743                         // assume the frame length is always the same.
 744                         pts_int += length_per_frame;
 745                 }
 746         }
 747 }
 748
 749 void Mixer::render_one_frame(int64_t duration)
 750 {
 751         // Get the main chain from the theme, and set its state immediately.
 752         Theme::Chain theme_main_chain = theme->get_chain(0, pts(), WIDTH, HEIGHT, input_state);
 753         EffectChain *chain = theme_main_chain.chain;
 754         theme_main_chain.setup_chain();
 755         //theme_main_chain.chain->enable_phase_timing(true);
 756
 757         GLuint y_tex, cbcr_tex;
 758         bool got_frame = video_encoder->begin_frame(&y_tex, &cbcr_tex);
 759         assert(got_frame);
 760
 761         // Render main chain.
 762         GLuint cbcr_full_tex = resource_pool->create_2d_texture(GL_RG8, WIDTH, HEIGHT);
 763         GLuint rgba_tex = resource_pool->create_2d_texture(GL_RGB565, WIDTH, HEIGHT);  // Saves texture bandwidth, although dithering gets messed up.
 764         GLuint fbo = resource_pool->create_fbo(y_tex, cbcr_full_tex, rgba_tex);
 765         check_error();
 766         chain->render_to_fbo(fbo, WIDTH, HEIGHT);
 767         resource_pool->release_fbo(fbo);
 768
 769         subsample_chroma(cbcr_full_tex, cbcr_tex);
 770         resource_pool->release_2d_texture(cbcr_full_tex);
 771
 772         // Set the right state for rgba_tex.
 773         glBindFramebuffer(GL_FRAMEBUFFER, 0);
 774         glBindTexture(GL_TEXTURE_2D, rgba_tex);
 775         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
 776         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
 777         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
 778
 779         const int64_t av_delay = TIMEBASE / 10;  // Corresponds to the fixed delay in resampling_queue.h. TODO: Make less hard-coded.
 780         RefCountedGLsync fence = video_encoder->end_frame(pts_int + av_delay, duration, theme_main_chain.input_frames);
 781
 782         // The live frame just shows the RGBA texture we just rendered.
 783         // It owns rgba_tex now.
 784         DisplayFrame live_frame;
 785         live_frame.chain = display_chain.get();
 786         live_frame.setup_chain = [this, rgba_tex]{
 787                 display_input->set_texture_num(rgba_tex);
 788         };
 789         live_frame.ready_fence = fence;
 790         live_frame.input_frames = {};
 791         live_frame.temp_textures = { rgba_tex };
 792         output_channel[OUTPUT_LIVE].output_frame(live_frame);
 793
 794         // Set up preview and any additional channels.
 795         for (int i = 1; i < theme->get_num_channels() + 2; ++i) {
 796                 DisplayFrame display_frame;
 797                 Theme::Chain chain = theme->get_chain(i, pts(), WIDTH, HEIGHT, input_state);  // FIXME: dimensions
 798                 display_frame.chain = chain.chain;
 799                 display_frame.setup_chain = chain.setup_chain;
 800                 display_frame.ready_fence = fence;
 801                 display_frame.input_frames = chain.input_frames;
 802                 display_frame.temp_textures = {};
 803                 output_channel[i].output_frame(display_frame);
 804         }
 805 }
 806
 807 void Mixer::send_audio_level_callback()
 808 {
 809         if (audio_level_callback == nullptr) {
 810                 return;
 811         }
 812
 813         unique_lock<mutex> lock(compressor_mutex);
 814         double loudness_s = r128.loudness_S();
 815         double loudness_i = r128.integrated();
 816         double loudness_range_low = r128.range_min();
 817         double loudness_range_high = r128.range_max();
 818
 819         audio_level_callback(loudness_s, 20.0 * log10(peak),
 820                 loudness_i, loudness_range_low, loudness_range_high,
 821                 gain_staging_db, 20.0 * log10(final_makeup_gain),
 822                 correlation.get_correlation());
 823 }
 824
 825 void Mixer::audio_thread_func()
 826 {
 827         while (!should_quit) {
 828                 AudioTask task;
 829
 830                 {
 831                         unique_lock<mutex> lock(audio_mutex);
 832                         audio_task_queue_changed.wait(lock, [this]{ return should_quit || !audio_task_queue.empty(); });
 833                         if (should_quit) {
 834                                 return;
 835                         }
 836                         task = audio_task_queue.front();
 837                         audio_task_queue.pop();
 838                 }
 839
 840                 process_audio_one_frame(task.pts_int, task.num_samples);
 841         }
 842 }
 843
 844 void Mixer::process_audio_one_frame(int64_t frame_pts_int, int num_samples)
 845 {
 846         vector<float> samples_card;
 847         vector<float> samples_out;
 848
 849         // TODO: Allow mixing audio from several sources.
 850         unsigned selected_audio_card = theme->map_signal(audio_source_channel);
 851         assert(selected_audio_card < num_cards);
 852
 853         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 854                 samples_card.resize(num_samples * 2);
 855                 {
 856                         unique_lock<mutex> lock(cards[card_index].audio_mutex);
 857                         cards[card_index].resampling_queue->get_output_samples(double(frame_pts_int) / TIMEBASE, &samples_card[0], num_samples);
 858                 }
 859                 if (card_index == selected_audio_card) {
 860                         samples_out = move(samples_card);
 861                 }
 862         }
 863
 864         // Cut away everything under 120 Hz (or whatever the cutoff is);
 865         // we don't need it for voice, and it will reduce headroom
 866         // and confuse the compressor. (In particular, any hums at 50 or 60 Hz
 867         // should be dampened.)
 868         if (locut_enabled) {
 869                 locut.render(samples_out.data(), samples_out.size() / 2, locut_cutoff_hz * 2.0 * M_PI / OUTPUT_FREQUENCY, 0.5f);
 870         }
 871
 872         // Apply a level compressor to get the general level right.
 873         // Basically, if it's over about -40 dBFS, we squeeze it down to that level
 874         // (or more precisely, near it, since we don't use infinite ratio),
 875         // then apply a makeup gain to get it to -14 dBFS. -14 dBFS is, of course,
 876         // entirely arbitrary, but from practical tests with speech, it seems to
 877         // put ut around -23 LUFS, so it's a reasonable starting point for later use.
 878         {
 879                 unique_lock<mutex> lock(compressor_mutex);
 880                 if (level_compressor_enabled) {
 881                         float threshold = 0.01f;   // -40 dBFS.
 882                         float ratio = 20.0f;
 883                         float attack_time = 0.5f;
 884                         float release_time = 20.0f;
 885                         float makeup_gain = pow(10.0f, (ref_level_dbfs - (-40.0f)) / 20.0f);  // +26 dB.
 886                         level_compressor.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 887                         gain_staging_db = 20.0 * log10(level_compressor.get_attenuation() * makeup_gain);
 888                 } else {
 889                         // Just apply the gain we already had.
 890                         float g = pow(10.0f, gain_staging_db / 20.0f);
 891                         for (size_t i = 0; i < samples_out.size(); ++i) {
 892                                 samples_out[i] *= g;
 893                         }
 894                 }
 895         }
 896
 897 #if 0
 898         printf("level=%f (%+5.2f dBFS) attenuation=%f (%+5.2f dB) end_result=%+5.2f dB\n",
 899                 level_compressor.get_level(), 20.0 * log10(level_compressor.get_level()),
 900                 level_compressor.get_attenuation(), 20.0 * log10(level_compressor.get_attenuation()),
 901                 20.0 * log10(level_compressor.get_level() * level_compressor.get_attenuation() * makeup_gain));
 902 #endif
 903
 904 //      float limiter_att, compressor_att;
 905
 906         // The real compressor.
 907         if (compressor_enabled) {
 908                 float threshold = pow(10.0f, compressor_threshold_dbfs / 20.0f);
 909                 float ratio = 20.0f;
 910                 float attack_time = 0.005f;
 911                 float release_time = 0.040f;
 912                 float makeup_gain = 2.0f;  // +6 dB.
 913                 compressor.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 914 //              compressor_att = compressor.get_attenuation();
 915         }
 916
 917         // Finally a limiter at -4 dB (so, -10 dBFS) to take out the worst peaks only.
 918         // Note that since ratio is not infinite, we could go slightly higher than this.
 919         if (limiter_enabled) {
 920                 float threshold = pow(10.0f, limiter_threshold_dbfs / 20.0f);
 921                 float ratio = 30.0f;
 922                 float attack_time = 0.0f;  // Instant.
 923                 float release_time = 0.020f;
 924                 float makeup_gain = 1.0f;  // 0 dB.
 925                 limiter.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 926 //              limiter_att = limiter.get_attenuation();
 927         }
 928
 929 //      printf("limiter=%+5.1f  compressor=%+5.1f\n", 20.0*log10(limiter_att), 20.0*log10(compressor_att));
 930
 931         // Upsample 4x to find interpolated peak.
 932         peak_resampler.inp_data = samples_out.data();
 933         peak_resampler.inp_count = samples_out.size() / 2;
 934
 935         vector<float> interpolated_samples_out;
 936         interpolated_samples_out.resize(samples_out.size());
 937         while (peak_resampler.inp_count > 0) {  // About four iterations.
 938                 peak_resampler.out_data = &interpolated_samples_out[0];
 939                 peak_resampler.out_count = interpolated_samples_out.size() / 2;
 940                 peak_resampler.process();
 941                 size_t out_stereo_samples = interpolated_samples_out.size() / 2 - peak_resampler.out_count;
 942                 peak = max<float>(peak, find_peak(interpolated_samples_out.data(), out_stereo_samples * 2));
 943                 peak_resampler.out_data = nullptr;
 944         }
 945
 946         // At this point, we are most likely close to +0 LU, but all of our
 947         // measurements have been on raw sample values, not R128 values.
 948         // So we have a final makeup gain to get us to +0 LU; the gain
 949         // adjustments required should be relatively small, and also, the
 950         // offset shouldn't change much (only if the type of audio changes
 951         // significantly). Thus, we shoot for updating this value basically
 952         // “whenever we process buffers”, since the R128 calculation isn't exactly
 953         // something we get out per-sample.
 954         //
 955         // Note that there's a feedback loop here, so we choose a very slow filter
 956         // (half-time of 100 seconds).
 957         double target_loudness_factor, alpha;
 958         {
 959                 unique_lock<mutex> lock(compressor_mutex);
 960                 double loudness_lu = r128.loudness_M() - ref_level_lufs;
 961                 double current_makeup_lu = 20.0f * log10(final_makeup_gain);
 962                 target_loudness_factor = pow(10.0f, -loudness_lu / 20.0f);
 963
 964                 // If we're outside +/- 5 LU uncorrected, we don't count it as
 965                 // a normal signal (probably silence) and don't change the
 966                 // correction factor; just apply what we already have.
 967                 if (fabs(loudness_lu - current_makeup_lu) >= 5.0 || !final_makeup_gain_auto) {
 968                         alpha = 0.0;
 969                 } else {
 970                         // Formula adapted from
 971                         // https://en.wikipedia.org/wiki/Low-pass_filter#Simple_infinite_impulse_response_filter.
 972                         const double half_time_s = 100.0;
 973                         const double fc_mul_2pi_delta_t = 1.0 / (half_time_s * OUTPUT_FREQUENCY);
 974                         alpha = fc_mul_2pi_delta_t / (fc_mul_2pi_delta_t + 1.0);
 975                 }
 976
 977                 double m = final_makeup_gain;
 978                 for (size_t i = 0; i < samples_out.size(); i += 2) {
 979                         samples_out[i + 0] *= m;
 980                         samples_out[i + 1] *= m;
 981                         m += (target_loudness_factor - m) * alpha;
 982                 }
 983                 final_makeup_gain = m;
 984         }
 985
 986         // Find R128 levels and L/R correlation.
 987         vector<float> left, right;
 988         deinterleave_samples(samples_out, &left, &right);
 989         float *ptrs[] = { left.data(), right.data() };
 990         {
 991                 unique_lock<mutex> lock(compressor_mutex);
 992                 r128.process(left.size(), ptrs);
 993                 correlation.process_samples(samples_out);
 994         }
 995
 996         // Send the samples to the sound card.
 997         if (alsa) {
 998                 alsa->write(samples_out);
 999         }
1000
1001         // And finally add them to the output.
1002         video_encoder->add_audio(frame_pts_int, move(samples_out));
1003 }
1004
1005 void Mixer::subsample_chroma(GLuint src_tex, GLuint dst_tex)
1006 {
1007         GLuint vao;
1008         glGenVertexArrays(1, &vao);
1009         check_error();
1010
1011         glBindVertexArray(vao);
1012         check_error();
1013
1014         // Extract Cb/Cr.
1015         GLuint fbo = resource_pool->create_fbo(dst_tex);
1016         glBindFramebuffer(GL_FRAMEBUFFER, fbo);
1017         glViewport(0, 0, WIDTH/2, HEIGHT/2);
1018         check_error();
1019
1020         glUseProgram(cbcr_program_num);
1021         check_error();
1022
1023         glActiveTexture(GL_TEXTURE0);
1024         check_error();
1025         glBindTexture(GL_TEXTURE_2D, src_tex);
1026         check_error();
1027         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
1028         check_error();
1029         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
1030         check_error();
1031         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
1032         check_error();
1033
1034         float chroma_offset_0[] = { -0.5f / WIDTH, 0.0f };
1035         set_uniform_vec2(cbcr_program_num, "foo", "chroma_offset_0", chroma_offset_0);
1036
1037         glBindBuffer(GL_ARRAY_BUFFER, cbcr_vbo);
1038         check_error();
1039
1040         for (GLint attr_index : { cbcr_position_attribute_index, cbcr_texcoord_attribute_index }) {
1041                 glEnableVertexAttribArray(attr_index);
1042                 check_error();
1043                 glVertexAttribPointer(attr_index, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
1044                 check_error();
1045         }
1046
1047         glDrawArrays(GL_TRIANGLES, 0, 3);
1048         check_error();
1049
1050         for (GLint attr_index : { cbcr_position_attribute_index, cbcr_texcoord_attribute_index }) {
1051                 glDisableVertexAttribArray(attr_index);
1052                 check_error();
1053         }
1054
1055         glUseProgram(0);
1056         check_error();
1057         glBindFramebuffer(GL_FRAMEBUFFER, 0);
1058         check_error();
1059
1060         resource_pool->release_fbo(fbo);
1061         glDeleteVertexArrays(1, &vao);
1062 }
1063
1064 void Mixer::release_display_frame(DisplayFrame *frame)
1065 {
1066         for (GLuint texnum : frame->temp_textures) {
1067                 resource_pool->release_2d_texture(texnum);
1068         }
1069         frame->temp_textures.clear();
1070         frame->ready_fence.reset();
1071         frame->input_frames.clear();
1072 }
1073
1074 void Mixer::start()
1075 {
1076         mixer_thread = thread(&Mixer::thread_func, this);
1077         audio_thread = thread(&Mixer::audio_thread_func, this);
1078 }
1079
1080 void Mixer::quit()
1081 {
1082         should_quit = true;
1083         audio_task_queue_changed.notify_one();
1084         mixer_thread.join();
1085         audio_thread.join();
1086 }
1087
1088 void Mixer::transition_clicked(int transition_num)
1089 {
1090         theme->transition_clicked(transition_num, pts());
1091 }
1092
1093 void Mixer::channel_clicked(int preview_num)
1094 {
1095         theme->channel_clicked(preview_num);
1096 }
1097
1098 void Mixer::reset_meters()
1099 {
1100         peak_resampler.reset();
1101         peak = 0.0f;
1102         r128.reset();
1103         r128.integr_start();
1104         correlation.reset();
1105 }
1106
1107 void Mixer::start_mode_scanning(unsigned card_index)
1108 {
1109         assert(card_index < num_cards);
1110         if (is_mode_scanning[card_index]) {
1111                 return;
1112         }
1113         is_mode_scanning[card_index] = true;
1114         mode_scanlist[card_index].clear();
1115         for (const auto &mode : cards[card_index].capture->get_available_video_modes()) {
1116                 mode_scanlist[card_index].push_back(mode.first);
1117         }
1118         assert(!mode_scanlist[card_index].empty());
1119         mode_scanlist_index[card_index] = 0;
1120         cards[card_index].capture->set_video_mode(mode_scanlist[card_index][0]);
1121         clock_gettime(CLOCK_MONOTONIC, &last_mode_scan_change[card_index]);
1122 }
1123
1124 Mixer::OutputChannel::~OutputChannel()
1125 {
1126         if (has_current_frame) {
1127                 parent->release_display_frame(&current_frame);
1128         }
1129         if (has_ready_frame) {
1130                 parent->release_display_frame(&ready_frame);
1131         }
1132 }
1133
1134 void Mixer::OutputChannel::output_frame(DisplayFrame frame)
1135 {
1136         // Store this frame for display. Remove the ready frame if any
1137         // (it was seemingly never used).
1138         {
1139                 unique_lock<mutex> lock(frame_mutex);
1140                 if (has_ready_frame) {
1141                         parent->release_display_frame(&ready_frame);
1142                 }
1143                 ready_frame = frame;
1144                 has_ready_frame = true;
1145         }
1146
1147         if (new_frame_ready_callback) {
1148                 new_frame_ready_callback();
1149         }
1150
1151         // Reduce the number of callbacks by filtering duplicates. The reason
1152         // why we bother doing this is that Qt seemingly can get into a state
1153         // where its builds up an essentially unbounded queue of signals,
1154         // consuming more and more memory, and there's no good way of collapsing
1155         // user-defined signals or limiting the length of the queue.
1156         if (transition_names_updated_callback) {
1157                 vector<string> transition_names = global_mixer->get_transition_names();
1158                 bool changed = false;
1159                 if (transition_names.size() != last_transition_names.size()) {
1160                         changed = true;
1161                 } else {
1162                         for (unsigned i = 0; i < transition_names.size(); ++i) {
1163                                 if (transition_names[i] != last_transition_names[i]) {
1164                                         changed = true;
1165                                         break;
1166                                 }
1167                         }
1168                 }
1169                 if (changed) {
1170                         transition_names_updated_callback(transition_names);
1171                         last_transition_names = transition_names;
1172                 }
1173         }
1174         if (name_updated_callback) {
1175                 string name = global_mixer->get_channel_name(channel);
1176                 if (name != last_name) {
1177                         name_updated_callback(name);
1178                         last_name = name;
1179                 }
1180         }
1181         if (color_updated_callback) {
1182                 string color = global_mixer->get_channel_color(channel);
1183                 if (color != last_color) {
1184                         color_updated_callback(color);
1185                         last_color = color;
1186                 }
1187         }
1188 }
1189
1190 bool Mixer::OutputChannel::get_display_frame(DisplayFrame *frame)
1191 {
1192         unique_lock<mutex> lock(frame_mutex);
1193         if (!has_current_frame && !has_ready_frame) {
1194                 return false;
1195         }
1196
1197         if (has_current_frame && has_ready_frame) {
1198                 // We have a new ready frame. Toss the current one.
1199                 parent->release_display_frame(&current_frame);
1200                 has_current_frame = false;
1201         }
1202         if (has_ready_frame) {
1203                 assert(!has_current_frame);
1204                 current_frame = ready_frame;
1205                 ready_frame.ready_fence.reset();  // Drop the refcount.
1206                 ready_frame.input_frames.clear();  // Drop the refcounts.
1207                 has_current_frame = true;
1208                 has_ready_frame = false;
1209         }
1210
1211         *frame = current_frame;
1212         return true;
1213 }
1214
1215 void Mixer::OutputChannel::set_frame_ready_callback(Mixer::new_frame_ready_callback_t callback)
1216 {
1217         new_frame_ready_callback = callback;
1218 }
1219
1220 void Mixer::OutputChannel::set_transition_names_updated_callback(Mixer::transition_names_updated_callback_t callback)
1221 {
1222         transition_names_updated_callback = callback;
1223 }
1224
1225 void Mixer::OutputChannel::set_name_updated_callback(Mixer::name_updated_callback_t callback)
1226 {
1227         name_updated_callback = callback;
1228 }
1229
1230 void Mixer::OutputChannel::set_color_updated_callback(Mixer::color_updated_callback_t callback)
1231 {
1232         color_updated_callback = callback;
1233 }
1234
1235 mutex RefCountedGLsync::fence_lock;