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