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 #include <sys/time.h>
  31 #include <sys/resource.h>
  32
  33 #include "bmusb/bmusb.h"
  34 #include "context.h"
  35 #include "decklink_capture.h"
  36 #include "defs.h"
  37 #include "fake_capture.h"
  38 #include "flags.h"
  39 #include "video_encoder.h"
  40 #include "pbo_frame_allocator.h"
  41 #include "ref_counted_gl_sync.h"
  42 #include "timebase.h"
  43
  44 class QOpenGLContext;
  45
  46 using namespace movit;
  47 using namespace std;
  48 using namespace std::placeholders;
  49
  50 Mixer *global_mixer = nullptr;
  51 bool uses_mlock = false;
  52
  53 namespace {
  54
  55 void convert_fixed24_to_fp32(float *dst, size_t out_channels, const uint8_t *src, size_t in_channels, size_t num_samples)
  56 {
  57         assert(in_channels >= out_channels);
  58         for (size_t i = 0; i < num_samples; ++i) {
  59                 for (size_t j = 0; j < out_channels; ++j) {
  60                         uint32_t s1 = *src++;
  61                         uint32_t s2 = *src++;
  62                         uint32_t s3 = *src++;
  63                         uint32_t s = s1 | (s1 << 8) | (s2 << 16) | (s3 << 24);
  64                         dst[i * out_channels + j] = int(s) * (1.0f / 4294967296.0f);
  65                 }
  66                 src += 3 * (in_channels - out_channels);
  67         }
  68 }
  69
  70 void convert_fixed32_to_fp32(float *dst, size_t out_channels, const uint8_t *src, size_t in_channels, size_t num_samples)
  71 {
  72         assert(in_channels >= out_channels);
  73         for (size_t i = 0; i < num_samples; ++i) {
  74                 for (size_t j = 0; j < out_channels; ++j) {
  75                         // Note: Assumes little-endian.
  76                         int32_t s = *(int32_t *)src;
  77                         dst[i * out_channels + j] = s * (1.0f / 4294967296.0f);
  78                         src += 4;
  79                 }
  80                 src += 4 * (in_channels - out_channels);
  81         }
  82 }
  83
  84 void insert_new_frame(RefCountedFrame frame, unsigned field_num, bool interlaced, unsigned card_index, InputState *input_state)
  85 {
  86         if (interlaced) {
  87                 for (unsigned frame_num = FRAME_HISTORY_LENGTH; frame_num --> 1; ) {  // :-)
  88                         input_state->buffered_frames[card_index][frame_num] =
  89                                 input_state->buffered_frames[card_index][frame_num - 1];
  90                 }
  91                 input_state->buffered_frames[card_index][0] = { frame, field_num };
  92         } else {
  93                 for (unsigned frame_num = 0; frame_num < FRAME_HISTORY_LENGTH; ++frame_num) {
  94                         input_state->buffered_frames[card_index][frame_num] = { frame, field_num };
  95                 }
  96         }
  97 }
  98
  99 }  // namespace
 100
 101 void QueueLengthPolicy::update_policy(int queue_length)
 102 {
 103         if (queue_length < 0) {  // Starvation.
 104                 if (been_at_safe_point_since_last_starvation && safe_queue_length < 5) {
 105                         ++safe_queue_length;
 106                         fprintf(stderr, "Card %u: Starvation, increasing safe limit to %u frames\n",
 107                                 card_index, safe_queue_length);
 108                 }
 109                 frames_with_at_least_one = 0;
 110                 been_at_safe_point_since_last_starvation = false;
 111                 return;
 112         }
 113         if (queue_length > 0) {
 114                 if (queue_length >= int(safe_queue_length)) {
 115                         been_at_safe_point_since_last_starvation = true;
 116                 }
 117                 if (++frames_with_at_least_one >= 1000 && safe_queue_length > 0) {
 118                         --safe_queue_length;
 119                         fprintf(stderr, "Card %u: Spare frames for more than 1000 frames, reducing safe limit to %u frames\n",
 120                                 card_index, safe_queue_length);
 121                         frames_with_at_least_one = 0;
 122                 }
 123         } else {
 124                 frames_with_at_least_one = 0;
 125         }
 126 }
 127
 128 Mixer::Mixer(const QSurfaceFormat &format, unsigned num_cards)
 129         : httpd(),
 130           num_cards(num_cards),
 131           mixer_surface(create_surface(format)),
 132           h264_encoder_surface(create_surface(format)),
 133           correlation(OUTPUT_FREQUENCY),
 134           level_compressor(OUTPUT_FREQUENCY),
 135           limiter(OUTPUT_FREQUENCY),
 136           compressor(OUTPUT_FREQUENCY)
 137 {
 138         CHECK(init_movit(MOVIT_SHADER_DIR, MOVIT_DEBUG_OFF));
 139         check_error();
 140
 141         // Since we allow non-bouncing 4:2:2 YCbCrInputs, effective subpixel precision
 142         // will be halved when sampling them, and we need to compensate here.
 143         movit_texel_subpixel_precision /= 2.0;
 144
 145         resource_pool.reset(new ResourcePool);
 146         theme.reset(new Theme(global_flags.theme_filename.c_str(), resource_pool.get(), num_cards));
 147         for (unsigned i = 0; i < NUM_OUTPUTS; ++i) {
 148                 output_channel[i].parent = this;
 149                 output_channel[i].channel = i;
 150         }
 151
 152         ImageFormat inout_format;
 153         inout_format.color_space = COLORSPACE_sRGB;
 154         inout_format.gamma_curve = GAMMA_sRGB;
 155
 156         // Display chain; shows the live output produced by the main chain (its RGBA version).
 157         display_chain.reset(new EffectChain(WIDTH, HEIGHT, resource_pool.get()));
 158         check_error();
 159         display_input = new FlatInput(inout_format, FORMAT_RGB, GL_UNSIGNED_BYTE, WIDTH, HEIGHT);  // FIXME: GL_UNSIGNED_BYTE is really wrong.
 160         display_chain->add_input(display_input);
 161         display_chain->add_output(inout_format, OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED);
 162         display_chain->set_dither_bits(0);  // Don't bother.
 163         display_chain->finalize();
 164
 165         video_encoder.reset(new VideoEncoder(resource_pool.get(), h264_encoder_surface, global_flags.va_display, WIDTH, HEIGHT, &httpd));
 166
 167         // Start listening for clients only once VideoEncoder has written its header, if any.
 168         httpd.start(9095);
 169
 170         // First try initializing the fake devices, then PCI devices, then USB,
 171         // until we have the desired number of cards.
 172         unsigned num_pci_devices = 0, num_usb_devices = 0;
 173         unsigned card_index = 0;
 174
 175         assert(global_flags.num_fake_cards >= 0);  // Enforced in flags.cpp.
 176         unsigned num_fake_cards = global_flags.num_fake_cards;
 177
 178         assert(num_fake_cards <= num_cards);  // Enforced in flags.cpp.
 179         for ( ; card_index < num_fake_cards; ++card_index) {
 180                 configure_card(card_index, format, new FakeCapture(card_index));
 181         }
 182
 183         if (global_flags.num_fake_cards > 0) {
 184                 fprintf(stderr, "Initialized %d fake cards.\n", global_flags.num_fake_cards);
 185         }
 186
 187         if (card_index < num_cards) {
 188                 IDeckLinkIterator *decklink_iterator = CreateDeckLinkIteratorInstance();
 189                 if (decklink_iterator != nullptr) {
 190                         for ( ; card_index < num_cards; ++card_index) {
 191                                 IDeckLink *decklink;
 192                                 if (decklink_iterator->Next(&decklink) != S_OK) {
 193                                         break;
 194                                 }
 195
 196                                 configure_card(card_index, format, new DeckLinkCapture(decklink, card_index - num_fake_cards));
 197                                 ++num_pci_devices;
 198                         }
 199                         decklink_iterator->Release();
 200                         fprintf(stderr, "Found %d DeckLink PCI card(s).\n", num_pci_devices);
 201                 } else {
 202                         fprintf(stderr, "DeckLink drivers not found. Probing for USB cards only.\n");
 203                 }
 204         }
 205         for ( ; card_index < num_cards; ++card_index) {
 206                 configure_card(card_index, format, new BMUSBCapture(card_index - num_pci_devices - num_fake_cards));
 207                 ++num_usb_devices;
 208         }
 209
 210         if (num_usb_devices > 0) {
 211                 has_bmusb_thread = true;
 212                 BMUSBCapture::start_bm_thread();
 213         }
 214
 215         for (card_index = 0; card_index < num_cards; ++card_index) {
 216                 cards[card_index].queue_length_policy.reset(card_index);
 217                 cards[card_index].capture->start_bm_capture();
 218         }
 219
 220         // Set up stuff for NV12 conversion.
 221
 222         // Cb/Cr shader.
 223         string cbcr_vert_shader =
 224                 "#version 130 \n"
 225                 " \n"
 226                 "in vec2 position; \n"
 227                 "in vec2 texcoord; \n"
 228                 "out vec2 tc0; \n"
 229                 "uniform vec2 foo_chroma_offset_0; \n"
 230                 " \n"
 231                 "void main() \n"
 232                 "{ \n"
 233                 "    // The result of glOrtho(0.0, 1.0, 0.0, 1.0, 0.0, 1.0) is: \n"
 234                 "    // \n"
 235                 "    //   2.000  0.000  0.000 -1.000 \n"
 236                 "    //   0.000  2.000  0.000 -1.000 \n"
 237                 "    //   0.000  0.000 -2.000 -1.000 \n"
 238                 "    //   0.000  0.000  0.000  1.000 \n"
 239                 "    gl_Position = vec4(2.0 * position.x - 1.0, 2.0 * position.y - 1.0, -1.0, 1.0); \n"
 240                 "    vec2 flipped_tc = texcoord; \n"
 241                 "    tc0 = flipped_tc + foo_chroma_offset_0; \n"
 242                 "} \n";
 243         string cbcr_frag_shader =
 244                 "#version 130 \n"
 245                 "in vec2 tc0; \n"
 246                 "uniform sampler2D cbcr_tex; \n"
 247                 "out vec4 FragColor; \n"
 248                 "void main() { \n"
 249                 "    FragColor = texture(cbcr_tex, tc0); \n"
 250                 "} \n";
 251         vector<string> frag_shader_outputs;
 252         cbcr_program_num = resource_pool->compile_glsl_program(cbcr_vert_shader, cbcr_frag_shader, frag_shader_outputs);
 253
 254         float vertices[] = {
 255                 0.0f, 2.0f,
 256                 0.0f, 0.0f,
 257                 2.0f, 0.0f
 258         };
 259         cbcr_vbo = generate_vbo(2, GL_FLOAT, sizeof(vertices), vertices);
 260         cbcr_position_attribute_index = glGetAttribLocation(cbcr_program_num, "position");
 261         cbcr_texcoord_attribute_index = glGetAttribLocation(cbcr_program_num, "texcoord");
 262
 263         r128.init(2, OUTPUT_FREQUENCY);
 264         r128.integr_start();
 265
 266         locut.init(FILTER_HPF, 2);
 267
 268         // If --flat-audio is given, turn off everything that messes with the sound,
 269         // except the final makeup gain.
 270         if (global_flags.flat_audio) {
 271                 set_locut_enabled(false);
 272                 set_gain_staging_auto(false);
 273                 set_limiter_enabled(false);
 274                 set_compressor_enabled(false);
 275         }
 276
 277         // hlen=16 is pretty low quality, but we use quite a bit of CPU otherwise,
 278         // and there's a limit to how important the peak meter is.
 279         peak_resampler.setup(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY * 4, /*num_channels=*/2, /*hlen=*/16, /*frel=*/1.0);
 280
 281         if (global_flags.enable_alsa_output) {
 282                 alsa.reset(new ALSAOutput(OUTPUT_FREQUENCY, /*num_channels=*/2));
 283         }
 284 }
 285
 286 Mixer::~Mixer()
 287 {
 288         resource_pool->release_glsl_program(cbcr_program_num);
 289         glDeleteBuffers(1, &cbcr_vbo);
 290         if (has_bmusb_thread) {
 291                 BMUSBCapture::stop_bm_thread();
 292         }
 293
 294         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 295                 {
 296                         unique_lock<mutex> lock(bmusb_mutex);
 297                         cards[card_index].should_quit = true;  // Unblock thread.
 298                         cards[card_index].new_frames_changed.notify_all();
 299                 }
 300                 cards[card_index].capture->stop_dequeue_thread();
 301         }
 302
 303         video_encoder.reset(nullptr);
 304 }
 305
 306 void Mixer::configure_card(unsigned card_index, const QSurfaceFormat &format, CaptureInterface *capture)
 307 {
 308         printf("Configuring card %d...\n", card_index);
 309
 310         CaptureCard *card = &cards[card_index];
 311         card->capture = capture;
 312         card->capture->set_frame_callback(bind(&Mixer::bm_frame, this, card_index, _1, _2, _3, _4, _5, _6, _7));
 313         card->frame_allocator.reset(new PBOFrameAllocator(8 << 20, WIDTH, HEIGHT));  // 8 MB.
 314         card->capture->set_video_frame_allocator(card->frame_allocator.get());
 315         card->surface = create_surface(format);
 316         card->resampling_queue.reset(new ResamplingQueue(card_index, OUTPUT_FREQUENCY, OUTPUT_FREQUENCY, 2));
 317         card->capture->configure_card();
 318 }
 319
 320
 321 namespace {
 322
 323 int unwrap_timecode(uint16_t current_wrapped, int last)
 324 {
 325         uint16_t last_wrapped = last & 0xffff;
 326         if (current_wrapped > last_wrapped) {
 327                 return (last & ~0xffff) | current_wrapped;
 328         } else {
 329                 return 0x10000 + ((last & ~0xffff) | current_wrapped);
 330         }
 331 }
 332
 333 float find_peak(const float *samples, size_t num_samples)
 334 {
 335         float m = fabs(samples[0]);
 336         for (size_t i = 1; i < num_samples; ++i) {
 337                 m = max(m, fabs(samples[i]));
 338         }
 339         return m;
 340 }
 341
 342 void deinterleave_samples(const vector<float> &in, vector<float> *out_l, vector<float> *out_r)
 343 {
 344         size_t num_samples = in.size() / 2;
 345         out_l->resize(num_samples);
 346         out_r->resize(num_samples);
 347
 348         const float *inptr = in.data();
 349         float *lptr = &(*out_l)[0];
 350         float *rptr = &(*out_r)[0];
 351         for (size_t i = 0; i < num_samples; ++i) {
 352                 *lptr++ = *inptr++;
 353                 *rptr++ = *inptr++;
 354         }
 355 }
 356
 357 }  // namespace
 358
 359 void Mixer::bm_frame(unsigned card_index, uint16_t timecode,
 360                      FrameAllocator::Frame video_frame, size_t video_offset, VideoFormat video_format,
 361                      FrameAllocator::Frame audio_frame, size_t audio_offset, AudioFormat audio_format)
 362 {
 363         CaptureCard *card = &cards[card_index];
 364
 365         if (is_mode_scanning[card_index]) {
 366                 if (video_format.has_signal) {
 367                         // Found a stable signal, so stop scanning.
 368                         is_mode_scanning[card_index] = false;
 369                 } else {
 370                         static constexpr double switch_time_s = 0.5;  // Should be enough time for the signal to stabilize.
 371                         timespec now;
 372                         clock_gettime(CLOCK_MONOTONIC, &now);
 373                         double sec_since_last_switch = (now.tv_sec - last_mode_scan_change[card_index].tv_sec) +
 374                                 1e-9 * (now.tv_nsec - last_mode_scan_change[card_index].tv_nsec);
 375                         if (sec_since_last_switch > switch_time_s) {
 376                                 // It isn't this mode; try the next one.
 377                                 mode_scanlist_index[card_index]++;
 378                                 mode_scanlist_index[card_index] %= mode_scanlist[card_index].size();
 379                                 cards[card_index].capture->set_video_mode(mode_scanlist[card_index][mode_scanlist_index[card_index]]);
 380                                 last_mode_scan_change[card_index] = now;
 381                         }
 382                 }
 383         }
 384
 385         int64_t frame_length = int64_t(TIMEBASE) * video_format.frame_rate_den / video_format.frame_rate_nom;
 386         assert(frame_length > 0);
 387
 388         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;
 389         if (num_samples > OUTPUT_FREQUENCY / 10) {
 390                 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",
 391                         card_index, int(audio_frame.len), int(audio_offset),
 392                         timecode, int(video_frame.len), int(video_offset), video_format.id);
 393                 if (video_frame.owner) {
 394                         video_frame.owner->release_frame(video_frame);
 395                 }
 396                 if (audio_frame.owner) {
 397                         audio_frame.owner->release_frame(audio_frame);
 398                 }
 399                 return;
 400         }
 401
 402         int64_t local_pts = card->next_local_pts;
 403         int dropped_frames = 0;
 404         if (card->last_timecode != -1) {
 405                 dropped_frames = unwrap_timecode(timecode, card->last_timecode) - card->last_timecode - 1;
 406         }
 407
 408         // Convert the audio to stereo fp32 and add it.
 409         vector<float> audio;
 410         audio.resize(num_samples * 2);
 411         switch (audio_format.bits_per_sample) {
 412         case 0:
 413                 assert(num_samples == 0);
 414                 break;
 415         case 24:
 416                 convert_fixed24_to_fp32(&audio[0], 2, audio_frame.data + audio_offset, audio_format.num_channels, num_samples);
 417                 break;
 418         case 32:
 419                 convert_fixed32_to_fp32(&audio[0], 2, audio_frame.data + audio_offset, audio_format.num_channels, num_samples);
 420                 break;
 421         default:
 422                 fprintf(stderr, "Cannot handle audio with %u bits per sample\n", audio_format.bits_per_sample);
 423                 assert(false);
 424         }
 425
 426         // Add the audio.
 427         {
 428                 unique_lock<mutex> lock(card->audio_mutex);
 429
 430                 // Number of samples per frame if we need to insert silence.
 431                 // (Could be nonintegral, but resampling will save us then.)
 432                 int silence_samples = OUTPUT_FREQUENCY * video_format.frame_rate_den / video_format.frame_rate_nom;
 433
 434                 if (dropped_frames > MAX_FPS * 2) {
 435                         fprintf(stderr, "Card %d lost more than two seconds (or time code jumping around; from 0x%04x to 0x%04x), resetting resampler\n",
 436                                 card_index, card->last_timecode, timecode);
 437                         card->resampling_queue.reset(new ResamplingQueue(card_index, OUTPUT_FREQUENCY, OUTPUT_FREQUENCY, 2));
 438                         dropped_frames = 0;
 439                 } else if (dropped_frames > 0) {
 440                         // Insert silence as needed.
 441                         fprintf(stderr, "Card %d dropped %d frame(s) (before timecode 0x%04x), inserting silence.\n",
 442                                 card_index, dropped_frames, timecode);
 443                         vector<float> silence(silence_samples * 2, 0.0f);
 444                         for (int i = 0; i < dropped_frames; ++i) {
 445                                 card->resampling_queue->add_input_samples(local_pts / double(TIMEBASE), silence.data(), silence_samples);
 446                                 // Note that if the format changed in the meantime, we have
 447                                 // no way of detecting that; we just have to assume the frame length
 448                                 // is always the same.
 449                                 local_pts += frame_length;
 450                         }
 451                 }
 452                 if (num_samples == 0) {
 453                         audio.resize(silence_samples * 2);
 454                         num_samples = silence_samples;
 455                 }
 456                 card->resampling_queue->add_input_samples(local_pts / double(TIMEBASE), audio.data(), num_samples);
 457                 card->next_local_pts = local_pts + frame_length;
 458         }
 459
 460         card->last_timecode = timecode;
 461
 462         // Done with the audio, so release it.
 463         if (audio_frame.owner) {
 464                 audio_frame.owner->release_frame(audio_frame);
 465         }
 466
 467         size_t expected_length = video_format.width * (video_format.height + video_format.extra_lines_top + video_format.extra_lines_bottom) * 2;
 468         if (video_frame.len - video_offset == 0 ||
 469             video_frame.len - video_offset != expected_length) {
 470                 if (video_frame.len != 0) {
 471                         printf("Card %d: Dropping video frame with wrong length (%ld; expected %ld)\n",
 472                                 card_index, video_frame.len - video_offset, expected_length);
 473                 }
 474                 if (video_frame.owner) {
 475                         video_frame.owner->release_frame(video_frame);
 476                 }
 477
 478                 // Still send on the information that we _had_ a frame, even though it's corrupted,
 479                 // so that pts can go up accordingly.
 480                 {
 481                         unique_lock<mutex> lock(bmusb_mutex);
 482                         CaptureCard::NewFrame new_frame;
 483                         new_frame.frame = RefCountedFrame(FrameAllocator::Frame());
 484                         new_frame.length = frame_length;
 485                         new_frame.interlaced = false;
 486                         new_frame.dropped_frames = dropped_frames;
 487                         card->new_frames.push(move(new_frame));
 488                         card->new_frames_changed.notify_all();
 489                 }
 490                 return;
 491         }
 492
 493         PBOFrameAllocator::Userdata *userdata = (PBOFrameAllocator::Userdata *)video_frame.userdata;
 494
 495         unsigned num_fields = video_format.interlaced ? 2 : 1;
 496         timespec frame_upload_start;
 497         if (video_format.interlaced) {
 498                 // Send the two fields along as separate frames; the other side will need to add
 499                 // a deinterlacer to actually get this right.
 500                 assert(video_format.height % 2 == 0);
 501                 video_format.height /= 2;
 502                 assert(frame_length % 2 == 0);
 503                 frame_length /= 2;
 504                 num_fields = 2;
 505                 clock_gettime(CLOCK_MONOTONIC, &frame_upload_start);
 506         }
 507         userdata->last_interlaced = video_format.interlaced;
 508         userdata->last_has_signal = video_format.has_signal;
 509         userdata->last_is_connected = video_format.is_connected;
 510         userdata->last_frame_rate_nom = video_format.frame_rate_nom;
 511         userdata->last_frame_rate_den = video_format.frame_rate_den;
 512         RefCountedFrame frame(video_frame);
 513
 514         // Upload the textures.
 515         size_t cbcr_width = video_format.width / 2;
 516         size_t cbcr_offset = video_offset / 2;
 517         size_t y_offset = video_frame.size / 2 + video_offset / 2;
 518
 519         for (unsigned field = 0; field < num_fields; ++field) {
 520                 // Put the actual texture upload in a lambda that is executed in the main thread.
 521                 // It is entirely possible to do this in the same thread (and it might even be
 522                 // faster, depending on the GPU and driver), but it appears to be trickling
 523                 // driver bugs very easily.
 524                 //
 525                 // Note that this means we must hold on to the actual frame data in <userdata>
 526                 // until the upload command is run, but we hold on to <frame> much longer than that
 527                 // (in fact, all the way until we no longer use the texture in rendering).
 528                 auto upload_func = [field, video_format, y_offset, cbcr_offset, cbcr_width, userdata]() {
 529                         unsigned field_start_line = (field == 1) ? video_format.second_field_start : video_format.extra_lines_top + field * (video_format.height + 22);
 530
 531                         if (userdata->tex_y[field] == 0 ||
 532                             userdata->tex_cbcr[field] == 0 ||
 533                             video_format.width != userdata->last_width[field] ||
 534                             video_format.height != userdata->last_height[field]) {
 535                                 // We changed resolution since last use of this texture, so we need to create
 536                                 // a new object. Note that this each card has its own PBOFrameAllocator,
 537                                 // we don't need to worry about these flip-flopping between resolutions.
 538                                 glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr[field]);
 539                                 check_error();
 540                                 glTexImage2D(GL_TEXTURE_2D, 0, GL_RG8, cbcr_width, video_format.height, 0, GL_RG, GL_UNSIGNED_BYTE, nullptr);
 541                                 check_error();
 542                                 glBindTexture(GL_TEXTURE_2D, userdata->tex_y[field]);
 543                                 check_error();
 544                                 glTexImage2D(GL_TEXTURE_2D, 0, GL_R8, video_format.width, video_format.height, 0, GL_RED, GL_UNSIGNED_BYTE, nullptr);
 545                                 check_error();
 546                                 userdata->last_width[field] = video_format.width;
 547                                 userdata->last_height[field] = video_format.height;
 548                         }
 549
 550                         GLuint pbo = userdata->pbo;
 551                         check_error();
 552                         glBindBuffer(GL_PIXEL_UNPACK_BUFFER, pbo);
 553                         check_error();
 554
 555                         size_t field_y_start = y_offset + video_format.width * field_start_line;
 556                         size_t field_cbcr_start = cbcr_offset + cbcr_width * field_start_line * sizeof(uint16_t);
 557
 558                         if (global_flags.flush_pbos) {
 559                                 glFlushMappedBufferRange(GL_PIXEL_UNPACK_BUFFER, field_y_start, video_format.width * video_format.height);
 560                                 check_error();
 561                                 glFlushMappedBufferRange(GL_PIXEL_UNPACK_BUFFER, field_cbcr_start, cbcr_width * video_format.height * sizeof(uint16_t));
 562                                 check_error();
 563                         }
 564
 565                         glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr[field]);
 566                         check_error();
 567                         glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, cbcr_width, video_format.height, GL_RG, GL_UNSIGNED_BYTE, BUFFER_OFFSET(field_cbcr_start));
 568                         check_error();
 569                         glBindTexture(GL_TEXTURE_2D, userdata->tex_y[field]);
 570                         check_error();
 571                         glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, video_format.width, video_format.height, GL_RED, GL_UNSIGNED_BYTE, BUFFER_OFFSET(field_y_start));
 572                         check_error();
 573                         glBindTexture(GL_TEXTURE_2D, 0);
 574                         check_error();
 575                         glBindBuffer(GL_PIXEL_UNPACK_BUFFER, 0);
 576                         check_error();
 577                 };
 578
 579                 if (field == 1) {
 580                         // Don't upload the second field as fast as we can; wait until
 581                         // the field time has approximately passed. (Otherwise, we could
 582                         // get timing jitter against the other sources, and possibly also
 583                         // against the video display, although the latter is not as critical.)
 584                         // This requires our system clock to be reasonably close to the
 585                         // video clock, but that's not an unreasonable assumption.
 586                         timespec second_field_start;
 587                         second_field_start.tv_nsec = frame_upload_start.tv_nsec +
 588                                 frame_length * 1000000000 / TIMEBASE;
 589                         second_field_start.tv_sec = frame_upload_start.tv_sec +
 590                                 second_field_start.tv_nsec / 1000000000;
 591                         second_field_start.tv_nsec %= 1000000000;
 592
 593                         while (clock_nanosleep(CLOCK_MONOTONIC, TIMER_ABSTIME,
 594                                                &second_field_start, nullptr) == -1 &&
 595                                errno == EINTR) ;
 596                 }
 597
 598                 {
 599                         unique_lock<mutex> lock(bmusb_mutex);
 600                         CaptureCard::NewFrame new_frame;
 601                         new_frame.frame = frame;
 602                         new_frame.length = frame_length;
 603                         new_frame.field = field;
 604                         new_frame.interlaced = video_format.interlaced;
 605                         new_frame.upload_func = upload_func;
 606                         new_frame.dropped_frames = dropped_frames;
 607                         card->new_frames.push(move(new_frame));
 608                         card->new_frames_changed.notify_all();
 609                 }
 610         }
 611 }
 612
 613 void Mixer::thread_func()
 614 {
 615         eglBindAPI(EGL_OPENGL_API);
 616         QOpenGLContext *context = create_context(mixer_surface);
 617         if (!make_current(context, mixer_surface)) {
 618                 printf("oops\n");
 619                 exit(1);
 620         }
 621
 622         struct timespec start, now;
 623         clock_gettime(CLOCK_MONOTONIC, &start);
 624
 625         int frame = 0;
 626         int stats_dropped_frames = 0;
 627
 628         while (!should_quit) {
 629                 CaptureCard::NewFrame new_frames[MAX_CARDS];
 630                 bool has_new_frame[MAX_CARDS] = { false };
 631                 int num_samples[MAX_CARDS] = { 0 };
 632
 633                 // TODO: Add a timeout.
 634                 unsigned master_card_index = theme->map_signal(master_clock_channel);
 635                 assert(master_card_index < num_cards);
 636
 637                 get_one_frame_from_each_card(master_card_index, new_frames, has_new_frame, num_samples);
 638                 schedule_audio_resampling_tasks(new_frames[master_card_index].dropped_frames, num_samples[master_card_index], new_frames[master_card_index].length);
 639                 stats_dropped_frames += new_frames[master_card_index].dropped_frames;
 640                 send_audio_level_callback();
 641
 642                 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 643                         if (card_index == master_card_index || !has_new_frame[card_index]) {
 644                                 continue;
 645                         }
 646                         if (new_frames[card_index].frame->len == 0) {
 647                                 ++new_frames[card_index].dropped_frames;
 648                         }
 649                         if (new_frames[card_index].dropped_frames > 0) {
 650                                 printf("Card %u dropped %d frames before this\n",
 651                                         card_index, int(new_frames[card_index].dropped_frames));
 652                         }
 653                 }
 654
 655                 // If the first card is reporting a corrupted or otherwise dropped frame,
 656                 // just increase the pts (skipping over this frame) and don't try to compute anything new.
 657                 if (new_frames[master_card_index].frame->len == 0) {
 658                         ++stats_dropped_frames;
 659                         pts_int += new_frames[master_card_index].length;
 660                         continue;
 661                 }
 662
 663                 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 664                         if (!has_new_frame[card_index] || new_frames[card_index].frame->len == 0)
 665                                 continue;
 666
 667                         CaptureCard::NewFrame *new_frame = &new_frames[card_index];
 668                         assert(new_frame->frame != nullptr);
 669                         insert_new_frame(new_frame->frame, new_frame->field, new_frame->interlaced, card_index, &input_state);
 670                         check_error();
 671
 672                         // The new texture might need uploading before use.
 673                         if (new_frame->upload_func) {
 674                                 new_frame->upload_func();
 675                                 new_frame->upload_func = nullptr;
 676                         }
 677                 }
 678
 679                 int64_t duration = new_frames[master_card_index].length;
 680                 render_one_frame(duration);
 681                 ++frame;
 682                 pts_int += duration;
 683
 684                 clock_gettime(CLOCK_MONOTONIC, &now);
 685                 double elapsed = now.tv_sec - start.tv_sec +
 686                         1e-9 * (now.tv_nsec - start.tv_nsec);
 687                 if (frame % 100 == 0) {
 688                 // check our memory usage, to see if we are close to our mlockall()
 689                 // limit (if at all set).
 690                 rusage used;
 691                 if (getrusage(RUSAGE_SELF, &used) == -1) {
 692                         perror("getrusage(RUSAGE_SELF)");
 693                         assert(false);
 694                 }
 695
 696                 rlimit limit;
 697                 if (getrlimit(RLIMIT_MEMLOCK, &limit) == -1) {
 698                         perror("getrlimit(RLIMIT_MEMLOCK)");
 699                         assert(false);
 700                 }
 701
 702                         printf("%d frames (%d dropped) in %.3f seconds = %.1f fps (%.1f ms/frame)",
 703                                 frame, stats_dropped_frames, elapsed, frame / elapsed,
 704                                 1e3 * elapsed / frame);
 705                 //      chain->print_phase_timing();
 706
 707                         if (uses_mlock) {
 708                                 // Check our memory usage, to see if we are close to our mlockall()
 709                                 // limit (if at all set).
 710                                 rusage used;
 711                                 if (getrusage(RUSAGE_SELF, &used) == -1) {
 712                                         perror("getrusage(RUSAGE_SELF)");
 713                                         assert(false);
 714                                 }
 715
 716                                 rlimit limit;
 717                                 if (getrlimit(RLIMIT_MEMLOCK, &limit) == -1) {
 718                                         perror("getrlimit(RLIMIT_MEMLOCK)");
 719                                         assert(false);
 720                                 }
 721
 722                                 printf(", using %ld / %ld MB lockable memory (%.1f%%)",
 723                                         long(used.ru_maxrss / 1024),
 724                                         long(limit.rlim_cur / 1048576),
 725                                         float(100.0 * (used.ru_maxrss * 1024.0) / limit.rlim_cur));
 726                         }
 727
 728                         printf("\n");
 729                 }
 730
 731
 732                 if (should_cut.exchange(false)) {  // Test and clear.
 733                         video_encoder->do_cut(frame);
 734                 }
 735
 736 #if 0
 737                 // Reset every 100 frames, so that local variations in frame times
 738                 // (especially for the first few frames, when the shaders are
 739                 // compiled etc.) don't make it hard to measure for the entire
 740                 // remaining duration of the program.
 741                 if (frame == 10000) {
 742                         frame = 0;
 743                         start = now;
 744                 }
 745 #endif
 746                 check_error();
 747         }
 748
 749         resource_pool->clean_context();
 750 }
 751
 752 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])
 753 {
 754         // The first card is the master timer, so wait for it to have a new frame.
 755         unique_lock<mutex> lock(bmusb_mutex);
 756         cards[master_card_index].new_frames_changed.wait(lock, [this, master_card_index]{ return !cards[master_card_index].new_frames.empty(); });
 757
 758         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 759                 CaptureCard *card = &cards[card_index];
 760                 if (card->new_frames.empty()) {
 761                         assert(card_index != master_card_index);
 762                         card->queue_length_policy.update_policy(-1);
 763                         continue;
 764                 }
 765                 new_frames[card_index] = move(card->new_frames.front());
 766                 has_new_frame[card_index] = true;
 767                 card->new_frames.pop();
 768                 card->new_frames_changed.notify_all();
 769
 770                 int num_samples_times_timebase = OUTPUT_FREQUENCY * new_frames[card_index].length + card->fractional_samples;
 771                 num_samples[card_index] = num_samples_times_timebase / TIMEBASE;
 772                 card->fractional_samples = num_samples_times_timebase % TIMEBASE;
 773                 assert(num_samples[card_index] >= 0);
 774
 775                 if (card_index == master_card_index) {
 776                         // We don't use the queue length policy for the master card,
 777                         // but we will if it stops being the master. Thus, clear out
 778                         // the policy in case we switch in the future.
 779                         card->queue_length_policy.reset(card_index);
 780                 } else {
 781                         // If we have excess frames compared to the policy for this card,
 782                         // drop frames from the head.
 783                         card->queue_length_policy.update_policy(card->new_frames.size());
 784                         while (card->new_frames.size() > card->queue_length_policy.get_safe_queue_length()) {
 785                                 card->new_frames.pop();
 786                         }
 787                 }
 788         }
 789 }
 790
 791 void Mixer::schedule_audio_resampling_tasks(unsigned dropped_frames, int num_samples_per_frame, int length_per_frame)
 792 {
 793         // Resample the audio as needed, including from previously dropped frames.
 794         assert(num_cards > 0);
 795         for (unsigned frame_num = 0; frame_num < dropped_frames + 1; ++frame_num) {
 796                 {
 797                         // Signal to the audio thread to process this frame.
 798                         unique_lock<mutex> lock(audio_mutex);
 799                         audio_task_queue.push(AudioTask{pts_int, num_samples_per_frame});
 800                         audio_task_queue_changed.notify_one();
 801                 }
 802                 if (frame_num != dropped_frames) {
 803                         // For dropped frames, increase the pts. Note that if the format changed
 804                         // in the meantime, we have no way of detecting that; we just have to
 805                         // assume the frame length is always the same.
 806                         pts_int += length_per_frame;
 807                 }
 808         }
 809 }
 810
 811 void Mixer::render_one_frame(int64_t duration)
 812 {
 813         // Get the main chain from the theme, and set its state immediately.
 814         Theme::Chain theme_main_chain = theme->get_chain(0, pts(), WIDTH, HEIGHT, input_state);
 815         EffectChain *chain = theme_main_chain.chain;
 816         theme_main_chain.setup_chain();
 817         //theme_main_chain.chain->enable_phase_timing(true);
 818
 819         GLuint y_tex, cbcr_tex;
 820         bool got_frame = video_encoder->begin_frame(&y_tex, &cbcr_tex);
 821         assert(got_frame);
 822
 823         // Render main chain.
 824         GLuint cbcr_full_tex = resource_pool->create_2d_texture(GL_RG8, WIDTH, HEIGHT);
 825         GLuint rgba_tex = resource_pool->create_2d_texture(GL_RGB565, WIDTH, HEIGHT);  // Saves texture bandwidth, although dithering gets messed up.
 826         GLuint fbo = resource_pool->create_fbo(y_tex, cbcr_full_tex, rgba_tex);
 827         check_error();
 828         chain->render_to_fbo(fbo, WIDTH, HEIGHT);
 829         resource_pool->release_fbo(fbo);
 830
 831         subsample_chroma(cbcr_full_tex, cbcr_tex);
 832         resource_pool->release_2d_texture(cbcr_full_tex);
 833
 834         // Set the right state for rgba_tex.
 835         glBindFramebuffer(GL_FRAMEBUFFER, 0);
 836         glBindTexture(GL_TEXTURE_2D, rgba_tex);
 837         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
 838         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
 839         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
 840
 841         const int64_t av_delay = TIMEBASE / 10;  // Corresponds to the fixed delay in resampling_queue.h. TODO: Make less hard-coded.
 842         RefCountedGLsync fence = video_encoder->end_frame(pts_int + av_delay, duration, theme_main_chain.input_frames);
 843
 844         // The live frame just shows the RGBA texture we just rendered.
 845         // It owns rgba_tex now.
 846         DisplayFrame live_frame;
 847         live_frame.chain = display_chain.get();
 848         live_frame.setup_chain = [this, rgba_tex]{
 849                 display_input->set_texture_num(rgba_tex);
 850         };
 851         live_frame.ready_fence = fence;
 852         live_frame.input_frames = {};
 853         live_frame.temp_textures = { rgba_tex };
 854         output_channel[OUTPUT_LIVE].output_frame(live_frame);
 855
 856         // Set up preview and any additional channels.
 857         for (int i = 1; i < theme->get_num_channels() + 2; ++i) {
 858                 DisplayFrame display_frame;
 859                 Theme::Chain chain = theme->get_chain(i, pts(), WIDTH, HEIGHT, input_state);  // FIXME: dimensions
 860                 display_frame.chain = chain.chain;
 861                 display_frame.setup_chain = chain.setup_chain;
 862                 display_frame.ready_fence = fence;
 863                 display_frame.input_frames = chain.input_frames;
 864                 display_frame.temp_textures = {};
 865                 output_channel[i].output_frame(display_frame);
 866         }
 867 }
 868
 869 void Mixer::send_audio_level_callback()
 870 {
 871         if (audio_level_callback == nullptr) {
 872                 return;
 873         }
 874
 875         unique_lock<mutex> lock(compressor_mutex);
 876         double loudness_s = r128.loudness_S();
 877         double loudness_i = r128.integrated();
 878         double loudness_range_low = r128.range_min();
 879         double loudness_range_high = r128.range_max();
 880
 881         audio_level_callback(loudness_s, 20.0 * log10(peak),
 882                 loudness_i, loudness_range_low, loudness_range_high,
 883                 gain_staging_db, 20.0 * log10(final_makeup_gain),
 884                 correlation.get_correlation());
 885 }
 886
 887 void Mixer::audio_thread_func()
 888 {
 889         while (!should_quit) {
 890                 AudioTask task;
 891
 892                 {
 893                         unique_lock<mutex> lock(audio_mutex);
 894                         audio_task_queue_changed.wait(lock, [this]{ return should_quit || !audio_task_queue.empty(); });
 895                         if (should_quit) {
 896                                 return;
 897                         }
 898                         task = audio_task_queue.front();
 899                         audio_task_queue.pop();
 900                 }
 901
 902                 process_audio_one_frame(task.pts_int, task.num_samples);
 903         }
 904 }
 905
 906 void Mixer::process_audio_one_frame(int64_t frame_pts_int, int num_samples)
 907 {
 908         vector<float> samples_card;
 909         vector<float> samples_out;
 910
 911         // TODO: Allow mixing audio from several sources.
 912         unsigned selected_audio_card = theme->map_signal(audio_source_channel);
 913         assert(selected_audio_card < num_cards);
 914
 915         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 916                 samples_card.resize(num_samples * 2);
 917                 {
 918                         unique_lock<mutex> lock(cards[card_index].audio_mutex);
 919                         cards[card_index].resampling_queue->get_output_samples(double(frame_pts_int) / TIMEBASE, &samples_card[0], num_samples);
 920                 }
 921                 if (card_index == selected_audio_card) {
 922                         samples_out = move(samples_card);
 923                 }
 924         }
 925
 926         // Cut away everything under 120 Hz (or whatever the cutoff is);
 927         // we don't need it for voice, and it will reduce headroom
 928         // and confuse the compressor. (In particular, any hums at 50 or 60 Hz
 929         // should be dampened.)
 930         if (locut_enabled) {
 931                 locut.render(samples_out.data(), samples_out.size() / 2, locut_cutoff_hz * 2.0 * M_PI / OUTPUT_FREQUENCY, 0.5f);
 932         }
 933
 934         // Apply a level compressor to get the general level right.
 935         // Basically, if it's over about -40 dBFS, we squeeze it down to that level
 936         // (or more precisely, near it, since we don't use infinite ratio),
 937         // then apply a makeup gain to get it to -14 dBFS. -14 dBFS is, of course,
 938         // entirely arbitrary, but from practical tests with speech, it seems to
 939         // put ut around -23 LUFS, so it's a reasonable starting point for later use.
 940         {
 941                 unique_lock<mutex> lock(compressor_mutex);
 942                 if (level_compressor_enabled) {
 943                         float threshold = 0.01f;   // -40 dBFS.
 944                         float ratio = 20.0f;
 945                         float attack_time = 0.5f;
 946                         float release_time = 20.0f;
 947                         float makeup_gain = pow(10.0f, (ref_level_dbfs - (-40.0f)) / 20.0f);  // +26 dB.
 948                         level_compressor.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 949                         gain_staging_db = 20.0 * log10(level_compressor.get_attenuation() * makeup_gain);
 950                 } else {
 951                         // Just apply the gain we already had.
 952                         float g = pow(10.0f, gain_staging_db / 20.0f);
 953                         for (size_t i = 0; i < samples_out.size(); ++i) {
 954                                 samples_out[i] *= g;
 955                         }
 956                 }
 957         }
 958
 959 #if 0
 960         printf("level=%f (%+5.2f dBFS) attenuation=%f (%+5.2f dB) end_result=%+5.2f dB\n",
 961                 level_compressor.get_level(), 20.0 * log10(level_compressor.get_level()),
 962                 level_compressor.get_attenuation(), 20.0 * log10(level_compressor.get_attenuation()),
 963                 20.0 * log10(level_compressor.get_level() * level_compressor.get_attenuation() * makeup_gain));
 964 #endif
 965
 966 //      float limiter_att, compressor_att;
 967
 968         // The real compressor.
 969         if (compressor_enabled) {
 970                 float threshold = pow(10.0f, compressor_threshold_dbfs / 20.0f);
 971                 float ratio = 20.0f;
 972                 float attack_time = 0.005f;
 973                 float release_time = 0.040f;
 974                 float makeup_gain = 2.0f;  // +6 dB.
 975                 compressor.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 976 //              compressor_att = compressor.get_attenuation();
 977         }
 978
 979         // Finally a limiter at -4 dB (so, -10 dBFS) to take out the worst peaks only.
 980         // Note that since ratio is not infinite, we could go slightly higher than this.
 981         if (limiter_enabled) {
 982                 float threshold = pow(10.0f, limiter_threshold_dbfs / 20.0f);
 983                 float ratio = 30.0f;
 984                 float attack_time = 0.0f;  // Instant.
 985                 float release_time = 0.020f;
 986                 float makeup_gain = 1.0f;  // 0 dB.
 987                 limiter.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 988 //              limiter_att = limiter.get_attenuation();
 989         }
 990
 991 //      printf("limiter=%+5.1f  compressor=%+5.1f\n", 20.0*log10(limiter_att), 20.0*log10(compressor_att));
 992
 993         // Upsample 4x to find interpolated peak.
 994         peak_resampler.inp_data = samples_out.data();
 995         peak_resampler.inp_count = samples_out.size() / 2;
 996
 997         vector<float> interpolated_samples_out;
 998         interpolated_samples_out.resize(samples_out.size());
 999         while (peak_resampler.inp_count > 0) {  // About four iterations.
1000                 peak_resampler.out_data = &interpolated_samples_out[0];
1001                 peak_resampler.out_count = interpolated_samples_out.size() / 2;
1002                 peak_resampler.process();
1003                 size_t out_stereo_samples = interpolated_samples_out.size() / 2 - peak_resampler.out_count;
1004                 peak = max<float>(peak, find_peak(interpolated_samples_out.data(), out_stereo_samples * 2));
1005                 peak_resampler.out_data = nullptr;
1006         }
1007
1008         // At this point, we are most likely close to +0 LU, but all of our
1009         // measurements have been on raw sample values, not R128 values.
1010         // So we have a final makeup gain to get us to +0 LU; the gain
1011         // adjustments required should be relatively small, and also, the
1012         // offset shouldn't change much (only if the type of audio changes
1013         // significantly). Thus, we shoot for updating this value basically
1014         // “whenever we process buffers”, since the R128 calculation isn't exactly
1015         // something we get out per-sample.
1016         //
1017         // Note that there's a feedback loop here, so we choose a very slow filter
1018         // (half-time of 100 seconds).
1019         double target_loudness_factor, alpha;
1020         {
1021                 unique_lock<mutex> lock(compressor_mutex);
1022                 double loudness_lu = r128.loudness_M() - ref_level_lufs;
1023                 double current_makeup_lu = 20.0f * log10(final_makeup_gain);
1024                 target_loudness_factor = pow(10.0f, -loudness_lu / 20.0f);
1025
1026                 // If we're outside +/- 5 LU uncorrected, we don't count it as
1027                 // a normal signal (probably silence) and don't change the
1028                 // correction factor; just apply what we already have.
1029                 if (fabs(loudness_lu - current_makeup_lu) >= 5.0 || !final_makeup_gain_auto) {
1030                         alpha = 0.0;
1031                 } else {
1032                         // Formula adapted from
1033                         // https://en.wikipedia.org/wiki/Low-pass_filter#Simple_infinite_impulse_response_filter.
1034                         const double half_time_s = 100.0;
1035                         const double fc_mul_2pi_delta_t = 1.0 / (half_time_s * OUTPUT_FREQUENCY);
1036                         alpha = fc_mul_2pi_delta_t / (fc_mul_2pi_delta_t + 1.0);
1037                 }
1038
1039                 double m = final_makeup_gain;
1040                 for (size_t i = 0; i < samples_out.size(); i += 2) {
1041                         samples_out[i + 0] *= m;
1042                         samples_out[i + 1] *= m;
1043                         m += (target_loudness_factor - m) * alpha;
1044                 }
1045                 final_makeup_gain = m;
1046         }
1047
1048         // Find R128 levels and L/R correlation.
1049         vector<float> left, right;
1050         deinterleave_samples(samples_out, &left, &right);
1051         float *ptrs[] = { left.data(), right.data() };
1052         {
1053                 unique_lock<mutex> lock(compressor_mutex);
1054                 r128.process(left.size(), ptrs);
1055                 correlation.process_samples(samples_out);
1056         }
1057
1058         // Send the samples to the sound card.
1059         if (alsa) {
1060                 alsa->write(samples_out);
1061         }
1062
1063         // And finally add them to the output.
1064         video_encoder->add_audio(frame_pts_int, move(samples_out));
1065 }
1066
1067 void Mixer::subsample_chroma(GLuint src_tex, GLuint dst_tex)
1068 {
1069         GLuint vao;
1070         glGenVertexArrays(1, &vao);
1071         check_error();
1072
1073         glBindVertexArray(vao);
1074         check_error();
1075
1076         // Extract Cb/Cr.
1077         GLuint fbo = resource_pool->create_fbo(dst_tex);
1078         glBindFramebuffer(GL_FRAMEBUFFER, fbo);
1079         glViewport(0, 0, WIDTH/2, HEIGHT/2);
1080         check_error();
1081
1082         glUseProgram(cbcr_program_num);
1083         check_error();
1084
1085         glActiveTexture(GL_TEXTURE0);
1086         check_error();
1087         glBindTexture(GL_TEXTURE_2D, src_tex);
1088         check_error();
1089         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
1090         check_error();
1091         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
1092         check_error();
1093         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
1094         check_error();
1095
1096         float chroma_offset_0[] = { -0.5f / WIDTH, 0.0f };
1097         set_uniform_vec2(cbcr_program_num, "foo", "chroma_offset_0", chroma_offset_0);
1098
1099         glBindBuffer(GL_ARRAY_BUFFER, cbcr_vbo);
1100         check_error();
1101
1102         for (GLint attr_index : { cbcr_position_attribute_index, cbcr_texcoord_attribute_index }) {
1103                 glEnableVertexAttribArray(attr_index);
1104                 check_error();
1105                 glVertexAttribPointer(attr_index, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
1106                 check_error();
1107         }
1108
1109         glDrawArrays(GL_TRIANGLES, 0, 3);
1110         check_error();
1111
1112         for (GLint attr_index : { cbcr_position_attribute_index, cbcr_texcoord_attribute_index }) {
1113                 glDisableVertexAttribArray(attr_index);
1114                 check_error();
1115         }
1116
1117         glUseProgram(0);
1118         check_error();
1119         glBindFramebuffer(GL_FRAMEBUFFER, 0);
1120         check_error();
1121
1122         resource_pool->release_fbo(fbo);
1123         glDeleteVertexArrays(1, &vao);
1124 }
1125
1126 void Mixer::release_display_frame(DisplayFrame *frame)
1127 {
1128         for (GLuint texnum : frame->temp_textures) {
1129                 resource_pool->release_2d_texture(texnum);
1130         }
1131         frame->temp_textures.clear();
1132         frame->ready_fence.reset();
1133         frame->input_frames.clear();
1134 }
1135
1136 void Mixer::start()
1137 {
1138         mixer_thread = thread(&Mixer::thread_func, this);
1139         audio_thread = thread(&Mixer::audio_thread_func, this);
1140 }
1141
1142 void Mixer::quit()
1143 {
1144         should_quit = true;
1145         audio_task_queue_changed.notify_one();
1146         mixer_thread.join();
1147         audio_thread.join();
1148 }
1149
1150 void Mixer::transition_clicked(int transition_num)
1151 {
1152         theme->transition_clicked(transition_num, pts());
1153 }
1154
1155 void Mixer::channel_clicked(int preview_num)
1156 {
1157         theme->channel_clicked(preview_num);
1158 }
1159
1160 void Mixer::reset_meters()
1161 {
1162         peak_resampler.reset();
1163         peak = 0.0f;
1164         r128.reset();
1165         r128.integr_start();
1166         correlation.reset();
1167 }
1168
1169 void Mixer::start_mode_scanning(unsigned card_index)
1170 {
1171         assert(card_index < num_cards);
1172         if (is_mode_scanning[card_index]) {
1173                 return;
1174         }
1175         is_mode_scanning[card_index] = true;
1176         mode_scanlist[card_index].clear();
1177         for (const auto &mode : cards[card_index].capture->get_available_video_modes()) {
1178                 mode_scanlist[card_index].push_back(mode.first);
1179         }
1180         assert(!mode_scanlist[card_index].empty());
1181         mode_scanlist_index[card_index] = 0;
1182         cards[card_index].capture->set_video_mode(mode_scanlist[card_index][0]);
1183         clock_gettime(CLOCK_MONOTONIC, &last_mode_scan_change[card_index]);
1184 }
1185
1186 Mixer::OutputChannel::~OutputChannel()
1187 {
1188         if (has_current_frame) {
1189                 parent->release_display_frame(&current_frame);
1190         }
1191         if (has_ready_frame) {
1192                 parent->release_display_frame(&ready_frame);
1193         }
1194 }
1195
1196 void Mixer::OutputChannel::output_frame(DisplayFrame frame)
1197 {
1198         // Store this frame for display. Remove the ready frame if any
1199         // (it was seemingly never used).
1200         {
1201                 unique_lock<mutex> lock(frame_mutex);
1202                 if (has_ready_frame) {
1203                         parent->release_display_frame(&ready_frame);
1204                 }
1205                 ready_frame = frame;
1206                 has_ready_frame = true;
1207         }
1208
1209         if (new_frame_ready_callback) {
1210                 new_frame_ready_callback();
1211         }
1212
1213         // Reduce the number of callbacks by filtering duplicates. The reason
1214         // why we bother doing this is that Qt seemingly can get into a state
1215         // where its builds up an essentially unbounded queue of signals,
1216         // consuming more and more memory, and there's no good way of collapsing
1217         // user-defined signals or limiting the length of the queue.
1218         if (transition_names_updated_callback) {
1219                 vector<string> transition_names = global_mixer->get_transition_names();
1220                 bool changed = false;
1221                 if (transition_names.size() != last_transition_names.size()) {
1222                         changed = true;
1223                 } else {
1224                         for (unsigned i = 0; i < transition_names.size(); ++i) {
1225                                 if (transition_names[i] != last_transition_names[i]) {
1226                                         changed = true;
1227                                         break;
1228                                 }
1229                         }
1230                 }
1231                 if (changed) {
1232                         transition_names_updated_callback(transition_names);
1233                         last_transition_names = transition_names;
1234                 }
1235         }
1236         if (name_updated_callback) {
1237                 string name = global_mixer->get_channel_name(channel);
1238                 if (name != last_name) {
1239                         name_updated_callback(name);
1240                         last_name = name;
1241                 }
1242         }
1243         if (color_updated_callback) {
1244                 string color = global_mixer->get_channel_color(channel);
1245                 if (color != last_color) {
1246                         color_updated_callback(color);
1247                         last_color = color;
1248                 }
1249         }
1250 }
1251
1252 bool Mixer::OutputChannel::get_display_frame(DisplayFrame *frame)
1253 {
1254         unique_lock<mutex> lock(frame_mutex);
1255         if (!has_current_frame && !has_ready_frame) {
1256                 return false;
1257         }
1258
1259         if (has_current_frame && has_ready_frame) {
1260                 // We have a new ready frame. Toss the current one.
1261                 parent->release_display_frame(&current_frame);
1262                 has_current_frame = false;
1263         }
1264         if (has_ready_frame) {
1265                 assert(!has_current_frame);
1266                 current_frame = ready_frame;
1267                 ready_frame.ready_fence.reset();  // Drop the refcount.
1268                 ready_frame.input_frames.clear();  // Drop the refcounts.
1269                 has_current_frame = true;
1270                 has_ready_frame = false;
1271         }
1272
1273         *frame = current_frame;
1274         return true;
1275 }
1276
1277 void Mixer::OutputChannel::set_frame_ready_callback(Mixer::new_frame_ready_callback_t callback)
1278 {
1279         new_frame_ready_callback = callback;
1280 }
1281
1282 void Mixer::OutputChannel::set_transition_names_updated_callback(Mixer::transition_names_updated_callback_t callback)
1283 {
1284         transition_names_updated_callback = callback;
1285 }
1286
1287 void Mixer::OutputChannel::set_name_updated_callback(Mixer::name_updated_callback_t callback)
1288 {
1289         name_updated_callback = callback;
1290 }
1291
1292 void Mixer::OutputChannel::set_color_updated_callback(Mixer::color_updated_callback_t callback)
1293 {
1294         color_updated_callback = callback;
1295 }
1296
1297 mutex RefCountedGLsync::fence_lock;