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