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

   1 #include "audio_mixer.h"
   2
   3 #include <assert.h>
   4 #include <endian.h>
   5 #include <bmusb/bmusb.h>
   6 #include <stdio.h>
   7 #include <endian.h>
   8 #include <cmath>
   9 #ifdef __SSE__
  10 #include <immintrin.h>
  11 #endif
  12
  13 #include "db.h"
  14 #include "flags.h"
  15 #include "mixer.h"
  16 #include "state.pb.h"
  17 #include "timebase.h"
  18
  19 using namespace bmusb;
  20 using namespace std;
  21 using namespace std::placeholders;
  22
  23 namespace {
  24
  25 // TODO: If these prove to be a bottleneck, they can be SSSE3-optimized
  26 // (usually including multiple channels at a time).
  27
  28 void convert_fixed16_to_fp32(float *dst, size_t out_channel, size_t out_num_channels,
  29                              const uint8_t *src, size_t in_channel, size_t in_num_channels,
  30                              size_t num_samples)
  31 {
  32         assert(in_channel < in_num_channels);
  33         assert(out_channel < out_num_channels);
  34         src += in_channel * 2;
  35         dst += out_channel;
  36
  37         for (size_t i = 0; i < num_samples; ++i) {
  38                 int16_t s = le16toh(*(int16_t *)src);
  39                 *dst = s * (1.0f / 32768.0f);
  40
  41                 src += 2 * in_num_channels;
  42                 dst += out_num_channels;
  43         }
  44 }
  45
  46 void convert_fixed24_to_fp32(float *dst, size_t out_channel, size_t out_num_channels,
  47                              const uint8_t *src, size_t in_channel, size_t in_num_channels,
  48                              size_t num_samples)
  49 {
  50         assert(in_channel < in_num_channels);
  51         assert(out_channel < out_num_channels);
  52         src += in_channel * 3;
  53         dst += out_channel;
  54
  55         for (size_t i = 0; i < num_samples; ++i) {
  56                 uint32_t s1 = src[0];
  57                 uint32_t s2 = src[1];
  58                 uint32_t s3 = src[2];
  59                 uint32_t s = s1 | (s1 << 8) | (s2 << 16) | (s3 << 24);
  60                 *dst = int(s) * (1.0f / 2147483648.0f);
  61
  62                 src += 3 * in_num_channels;
  63                 dst += out_num_channels;
  64         }
  65 }
  66
  67 void convert_fixed32_to_fp32(float *dst, size_t out_channel, size_t out_num_channels,
  68                              const uint8_t *src, size_t in_channel, size_t in_num_channels,
  69                              size_t num_samples)
  70 {
  71         assert(in_channel < in_num_channels);
  72         assert(out_channel < out_num_channels);
  73         src += in_channel * 4;
  74         dst += out_channel;
  75
  76         for (size_t i = 0; i < num_samples; ++i) {
  77                 int32_t s = le32toh(*(int32_t *)src);
  78                 *dst = s * (1.0f / 2147483648.0f);
  79
  80                 src += 4 * in_num_channels;
  81                 dst += out_num_channels;
  82         }
  83 }
  84
  85 float find_peak_plain(const float *samples, size_t num_samples) __attribute__((unused));
  86
  87 float find_peak_plain(const float *samples, size_t num_samples)
  88 {
  89         float m = fabs(samples[0]);
  90         for (size_t i = 1; i < num_samples; ++i) {
  91                 m = max(m, fabs(samples[i]));
  92         }
  93         return m;
  94 }
  95
  96 #ifdef __SSE__
  97 static inline float horizontal_max(__m128 m)
  98 {
  99         __m128 tmp = _mm_shuffle_ps(m, m, _MM_SHUFFLE(1, 0, 3, 2));
 100         m = _mm_max_ps(m, tmp);
 101         tmp = _mm_shuffle_ps(m, m, _MM_SHUFFLE(2, 3, 0, 1));
 102         m = _mm_max_ps(m, tmp);
 103         return _mm_cvtss_f32(m);
 104 }
 105
 106 float find_peak(const float *samples, size_t num_samples)
 107 {
 108         const __m128 abs_mask = _mm_castsi128_ps(_mm_set1_epi32(0x7fffffffu));
 109         __m128 m = _mm_setzero_ps();
 110         for (size_t i = 0; i < (num_samples & ~3); i += 4) {
 111                 __m128 x = _mm_loadu_ps(samples + i);
 112                 x = _mm_and_ps(x, abs_mask);
 113                 m = _mm_max_ps(m, x);
 114         }
 115         float result = horizontal_max(m);
 116
 117         for (size_t i = (num_samples & ~3); i < num_samples; ++i) {
 118                 result = max(result, fabs(samples[i]));
 119         }
 120
 121 #if 0
 122         // Self-test. We should be bit-exact the same.
 123         float reference_result = find_peak_plain(samples, num_samples);
 124         if (result != reference_result) {
 125                 fprintf(stderr, "Error: Peak is %f [%f %f %f %f]; should be %f.\n",
 126                         result,
 127                         _mm_cvtss_f32(_mm_shuffle_ps(m, m, _MM_SHUFFLE(0, 0, 0, 0))),
 128                         _mm_cvtss_f32(_mm_shuffle_ps(m, m, _MM_SHUFFLE(1, 1, 1, 1))),
 129                         _mm_cvtss_f32(_mm_shuffle_ps(m, m, _MM_SHUFFLE(2, 2, 2, 2))),
 130                         _mm_cvtss_f32(_mm_shuffle_ps(m, m, _MM_SHUFFLE(3, 3, 3, 3))),
 131                         reference_result);
 132                 abort();
 133         }
 134 #endif
 135         return result;
 136 }
 137 #else
 138 float find_peak(const float *samples, size_t num_samples)
 139 {
 140         return find_peak_plain(samples, num_samples);
 141 }
 142 #endif
 143
 144 void deinterleave_samples(const vector<float> &in, vector<float> *out_l, vector<float> *out_r)
 145 {
 146         size_t num_samples = in.size() / 2;
 147         out_l->resize(num_samples);
 148         out_r->resize(num_samples);
 149
 150         const float *inptr = in.data();
 151         float *lptr = &(*out_l)[0];
 152         float *rptr = &(*out_r)[0];
 153         for (size_t i = 0; i < num_samples; ++i) {
 154                 *lptr++ = *inptr++;
 155                 *rptr++ = *inptr++;
 156         }
 157 }
 158
 159 }  // namespace
 160
 161 AudioMixer::AudioMixer(unsigned num_cards)
 162         : num_cards(num_cards),
 163           limiter(OUTPUT_FREQUENCY),
 164           correlation(OUTPUT_FREQUENCY)
 165 {
 166         global_audio_mixer = this;
 167
 168         for (unsigned bus_index = 0; bus_index < MAX_BUSES; ++bus_index) {
 169                 locut[bus_index].init(FILTER_HPF, 2);
 170                 eq[bus_index][EQ_BAND_BASS].init(FILTER_LOW_SHELF, 1);
 171                 // Note: EQ_BAND_MID isn't used (see comments in apply_eq()).
 172                 eq[bus_index][EQ_BAND_TREBLE].init(FILTER_HIGH_SHELF, 1);
 173                 compressor[bus_index].reset(new StereoCompressor(OUTPUT_FREQUENCY));
 174                 level_compressor[bus_index].reset(new StereoCompressor(OUTPUT_FREQUENCY));
 175
 176                 set_bus_settings(bus_index, get_default_bus_settings());
 177         }
 178         set_limiter_enabled(global_flags.limiter_enabled);
 179         set_final_makeup_gain_auto(global_flags.final_makeup_gain_auto);
 180         alsa_pool.init();
 181
 182         InputMapping new_input_mapping;
 183         if (!global_flags.input_mapping_filename.empty()) {
 184                 if (!load_input_mapping_from_file(get_devices(),
 185                                                   global_flags.input_mapping_filename,
 186                                                   &new_input_mapping)) {
 187                         fprintf(stderr, "Failed to load input mapping from '%s', exiting.\n",
 188                                 global_flags.input_mapping_filename.c_str());
 189                         exit(1);
 190                 }
 191         } else {
 192                 // Generate a very simple, default input mapping.
 193                 InputMapping::Bus input;
 194                 input.name = "Main";
 195                 input.device.type = InputSourceType::CAPTURE_CARD;
 196                 input.device.index = 0;
 197                 input.source_channel[0] = 0;
 198                 input.source_channel[1] = 1;
 199
 200                 new_input_mapping.buses.push_back(input);
 201         }
 202         set_input_mapping(new_input_mapping);
 203
 204         r128.init(2, OUTPUT_FREQUENCY);
 205         r128.integr_start();
 206
 207         // hlen=16 is pretty low quality, but we use quite a bit of CPU otherwise,
 208         // and there's a limit to how important the peak meter is.
 209         peak_resampler.setup(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY * 4, /*num_channels=*/2, /*hlen=*/16, /*frel=*/1.0);
 210 }
 211
 212 void AudioMixer::reset_resampler(DeviceSpec device_spec)
 213 {
 214         lock_guard<timed_mutex> lock(audio_mutex);
 215         reset_resampler_mutex_held(device_spec);
 216 }
 217
 218 void AudioMixer::reset_resampler_mutex_held(DeviceSpec device_spec)
 219 {
 220         AudioDevice *device = find_audio_device(device_spec);
 221
 222         if (device->interesting_channels.empty()) {
 223                 device->resampling_queue.reset();
 224         } else {
 225                 // TODO: ResamplingQueue should probably take the full device spec.
 226                 // (It's only used for console output, though.)
 227                 device->resampling_queue.reset(new ResamplingQueue(device_spec.index, device->capture_frequency, OUTPUT_FREQUENCY, device->interesting_channels.size()));
 228         }
 229         device->next_local_pts = 0;
 230 }
 231
 232 bool AudioMixer::add_audio(DeviceSpec device_spec, const uint8_t *data, unsigned num_samples, AudioFormat audio_format, int64_t frame_length)
 233 {
 234         AudioDevice *device = find_audio_device(device_spec);
 235
 236         unique_lock<timed_mutex> lock(audio_mutex, defer_lock);
 237         if (!lock.try_lock_for(chrono::milliseconds(10))) {
 238                 return false;
 239         }
 240         if (device->resampling_queue == nullptr) {
 241                 // No buses use this device; throw it away.
 242                 return true;
 243         }
 244
 245         unsigned num_channels = device->interesting_channels.size();
 246         assert(num_channels > 0);
 247
 248         // Convert the audio to fp32.
 249         unique_ptr<float[]> audio(new float[num_samples * num_channels]);
 250         unsigned channel_index = 0;
 251         for (auto channel_it = device->interesting_channels.cbegin(); channel_it != device->interesting_channels.end(); ++channel_it, ++channel_index) {
 252                 switch (audio_format.bits_per_sample) {
 253                 case 0:
 254                         assert(num_samples == 0);
 255                         break;
 256                 case 16:
 257                         convert_fixed16_to_fp32(audio.get(), channel_index, num_channels, data, *channel_it, audio_format.num_channels, num_samples);
 258                         break;
 259                 case 24:
 260                         convert_fixed24_to_fp32(audio.get(), channel_index, num_channels, data, *channel_it, audio_format.num_channels, num_samples);
 261                         break;
 262                 case 32:
 263                         convert_fixed32_to_fp32(audio.get(), channel_index, num_channels, data, *channel_it, audio_format.num_channels, num_samples);
 264                         break;
 265                 default:
 266                         fprintf(stderr, "Cannot handle audio with %u bits per sample\n", audio_format.bits_per_sample);
 267                         assert(false);
 268                 }
 269         }
 270
 271         // Now add it.
 272         int64_t local_pts = device->next_local_pts;
 273         device->resampling_queue->add_input_samples(local_pts / double(TIMEBASE), audio.get(), num_samples);
 274         device->next_local_pts = local_pts + frame_length;
 275         return true;
 276 }
 277
 278 bool AudioMixer::add_silence(DeviceSpec device_spec, unsigned samples_per_frame, unsigned num_frames, int64_t frame_length)
 279 {
 280         AudioDevice *device = find_audio_device(device_spec);
 281
 282         unique_lock<timed_mutex> lock(audio_mutex, defer_lock);
 283         if (!lock.try_lock_for(chrono::milliseconds(10))) {
 284                 return false;
 285         }
 286         if (device->resampling_queue == nullptr) {
 287                 // No buses use this device; throw it away.
 288                 return true;
 289         }
 290
 291         unsigned num_channels = device->interesting_channels.size();
 292         assert(num_channels > 0);
 293
 294         vector<float> silence(samples_per_frame * num_channels, 0.0f);
 295         for (unsigned i = 0; i < num_frames; ++i) {
 296                 device->resampling_queue->add_input_samples(device->next_local_pts / double(TIMEBASE), silence.data(), samples_per_frame);
 297                 // Note that if the format changed in the meantime, we have
 298                 // no way of detecting that; we just have to assume the frame length
 299                 // is always the same.
 300                 device->next_local_pts += frame_length;
 301         }
 302         return true;
 303 }
 304
 305 bool AudioMixer::silence_card(DeviceSpec device_spec, bool silence)
 306 {
 307         AudioDevice *device = find_audio_device(device_spec);
 308
 309         unique_lock<timed_mutex> lock(audio_mutex, defer_lock);
 310         if (!lock.try_lock_for(chrono::milliseconds(10))) {
 311                 return false;
 312         }
 313
 314         if (device->silenced && !silence) {
 315                 reset_resampler_mutex_held(device_spec);
 316         }
 317         device->silenced = silence;
 318         return true;
 319 }
 320
 321 AudioMixer::BusSettings AudioMixer::get_default_bus_settings()
 322 {
 323         BusSettings settings;
 324         settings.fader_volume_db = 0.0f;
 325         settings.locut_enabled = global_flags.locut_enabled;
 326         for (unsigned band_index = 0; band_index < NUM_EQ_BANDS; ++band_index) {
 327                 settings.eq_level_db[band_index] = 0.0f;
 328         }
 329         settings.gain_staging_db = global_flags.initial_gain_staging_db;
 330         settings.level_compressor_enabled = global_flags.gain_staging_auto;
 331         settings.compressor_threshold_dbfs = ref_level_dbfs - 12.0f;  // -12 dB.
 332         settings.compressor_enabled = global_flags.compressor_enabled;
 333         return settings;
 334 }
 335
 336 AudioMixer::BusSettings AudioMixer::get_bus_settings(unsigned bus_index) const
 337 {
 338         lock_guard<timed_mutex> lock(audio_mutex);
 339         BusSettings settings;
 340         settings.fader_volume_db = fader_volume_db[bus_index];
 341         settings.locut_enabled = locut_enabled[bus_index];
 342         for (unsigned band_index = 0; band_index < NUM_EQ_BANDS; ++band_index) {
 343                 settings.eq_level_db[band_index] = eq_level_db[bus_index][band_index];
 344         }
 345         settings.gain_staging_db = gain_staging_db[bus_index];
 346         settings.level_compressor_enabled = level_compressor_enabled[bus_index];
 347         settings.compressor_threshold_dbfs = compressor_threshold_dbfs[bus_index];
 348         settings.compressor_enabled = compressor_enabled[bus_index];
 349         return settings;
 350 }
 351
 352 void AudioMixer::set_bus_settings(unsigned bus_index, const AudioMixer::BusSettings &settings)
 353 {
 354         lock_guard<timed_mutex> lock(audio_mutex);
 355         fader_volume_db[bus_index] = settings.fader_volume_db;
 356         locut_enabled[bus_index] = settings.locut_enabled;
 357         for (unsigned band_index = 0; band_index < NUM_EQ_BANDS; ++band_index) {
 358                 eq_level_db[bus_index][band_index] = settings.eq_level_db[band_index];
 359         }
 360         gain_staging_db[bus_index] = settings.gain_staging_db;
 361         level_compressor_enabled[bus_index] = settings.level_compressor_enabled;
 362         compressor_threshold_dbfs[bus_index] = settings.compressor_threshold_dbfs;
 363         compressor_enabled[bus_index] = settings.compressor_enabled;
 364 }
 365
 366 AudioMixer::AudioDevice *AudioMixer::find_audio_device(DeviceSpec device)
 367 {
 368         switch (device.type) {
 369         case InputSourceType::CAPTURE_CARD:
 370                 return &video_cards[device.index];
 371         case InputSourceType::ALSA_INPUT:
 372                 return &alsa_inputs[device.index];
 373         case InputSourceType::SILENCE:
 374         default:
 375                 assert(false);
 376         }
 377         return nullptr;
 378 }
 379
 380 // Get a pointer to the given channel from the given device.
 381 // The channel must be picked out earlier and resampled.
 382 void AudioMixer::find_sample_src_from_device(const map<DeviceSpec, vector<float>> &samples_card, DeviceSpec device_spec, int source_channel, const float **srcptr, unsigned *stride)
 383 {
 384         static float zero = 0.0f;
 385         if (source_channel == -1 || device_spec.type == InputSourceType::SILENCE) {
 386                 *srcptr = &zero;
 387                 *stride = 0;
 388                 return;
 389         }
 390         AudioDevice *device = find_audio_device(device_spec);
 391         assert(device->interesting_channels.count(source_channel) != 0);
 392         unsigned channel_index = 0;
 393         for (int channel : device->interesting_channels) {
 394                 if (channel == source_channel) break;
 395                 ++channel_index;
 396         }
 397         assert(channel_index < device->interesting_channels.size());
 398         const auto it = samples_card.find(device_spec);
 399         assert(it != samples_card.end());
 400         *srcptr = &(it->second)[channel_index];
 401         *stride = device->interesting_channels.size();
 402 }
 403
 404 // TODO: Can be SSSE3-optimized if need be.
 405 void AudioMixer::fill_audio_bus(const map<DeviceSpec, vector<float>> &samples_card, const InputMapping::Bus &bus, unsigned num_samples, float *output)
 406 {
 407         if (bus.device.type == InputSourceType::SILENCE) {
 408                 memset(output, 0, num_samples * sizeof(*output));
 409         } else {
 410                 assert(bus.device.type == InputSourceType::CAPTURE_CARD ||
 411                        bus.device.type == InputSourceType::ALSA_INPUT);
 412                 const float *lsrc, *rsrc;
 413                 unsigned lstride, rstride;
 414                 float *dptr = output;
 415                 find_sample_src_from_device(samples_card, bus.device, bus.source_channel[0], &lsrc, &lstride);
 416                 find_sample_src_from_device(samples_card, bus.device, bus.source_channel[1], &rsrc, &rstride);
 417                 for (unsigned i = 0; i < num_samples; ++i) {
 418                         *dptr++ = *lsrc;
 419                         *dptr++ = *rsrc;
 420                         lsrc += lstride;
 421                         rsrc += rstride;
 422                 }
 423         }
 424 }
 425
 426 vector<DeviceSpec> AudioMixer::get_active_devices() const
 427 {
 428         vector<DeviceSpec> ret;
 429         for (unsigned card_index = 0; card_index < MAX_VIDEO_CARDS; ++card_index) {
 430                 const DeviceSpec device_spec{InputSourceType::CAPTURE_CARD, card_index};
 431                 if (!find_audio_device(device_spec)->interesting_channels.empty()) {
 432                         ret.push_back(device_spec);
 433                 }
 434         }
 435         for (unsigned card_index = 0; card_index < MAX_ALSA_CARDS; ++card_index) {
 436                 const DeviceSpec device_spec{InputSourceType::ALSA_INPUT, card_index};
 437                 if (!find_audio_device(device_spec)->interesting_channels.empty()) {
 438                         ret.push_back(device_spec);
 439                 }
 440         }
 441         return ret;
 442 }
 443
 444 vector<float> AudioMixer::get_output(double pts, unsigned num_samples, ResamplingQueue::RateAdjustmentPolicy rate_adjustment_policy)
 445 {
 446         map<DeviceSpec, vector<float>> samples_card;
 447         vector<float> samples_bus;
 448
 449         lock_guard<timed_mutex> lock(audio_mutex);
 450
 451         // Pick out all the interesting channels from all the cards.
 452         for (const DeviceSpec &device_spec : get_active_devices()) {
 453                 AudioDevice *device = find_audio_device(device_spec);
 454                 samples_card[device_spec].resize(num_samples * device->interesting_channels.size());
 455                 if (device->silenced) {
 456                         memset(&samples_card[device_spec][0], 0, samples_card[device_spec].size() * sizeof(float));
 457                 } else {
 458                         device->resampling_queue->get_output_samples(
 459                                 pts,
 460                                 &samples_card[device_spec][0],
 461                                 num_samples,
 462                                 rate_adjustment_policy);
 463                 }
 464         }
 465
 466         vector<float> samples_out, left, right;
 467         samples_out.resize(num_samples * 2);
 468         samples_bus.resize(num_samples * 2);
 469         for (unsigned bus_index = 0; bus_index < input_mapping.buses.size(); ++bus_index) {
 470                 fill_audio_bus(samples_card, input_mapping.buses[bus_index], num_samples, &samples_bus[0]);
 471                 apply_eq(bus_index, &samples_bus);
 472
 473                 {
 474                         lock_guard<mutex> lock(compressor_mutex);
 475
 476                         // Apply a level compressor to get the general level right.
 477                         // Basically, if it's over about -40 dBFS, we squeeze it down to that level
 478                         // (or more precisely, near it, since we don't use infinite ratio),
 479                         // then apply a makeup gain to get it to -14 dBFS. -14 dBFS is, of course,
 480                         // entirely arbitrary, but from practical tests with speech, it seems to
 481                         // put ut around -23 LUFS, so it's a reasonable starting point for later use.
 482                         if (level_compressor_enabled[bus_index]) {
 483                                 float threshold = 0.01f;   // -40 dBFS.
 484                                 float ratio = 20.0f;
 485                                 float attack_time = 0.5f;
 486                                 float release_time = 20.0f;
 487                                 float makeup_gain = from_db(ref_level_dbfs - (-40.0f));  // +26 dB.
 488                                 level_compressor[bus_index]->process(samples_bus.data(), samples_bus.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 489                                 gain_staging_db[bus_index] = to_db(level_compressor[bus_index]->get_attenuation() * makeup_gain);
 490                         } else {
 491                                 // Just apply the gain we already had.
 492                                 float g = from_db(gain_staging_db[bus_index]);
 493                                 for (size_t i = 0; i < samples_bus.size(); ++i) {
 494                                         samples_bus[i] *= g;
 495                                 }
 496                         }
 497
 498 #if 0
 499                         printf("level=%f (%+5.2f dBFS) attenuation=%f (%+5.2f dB) end_result=%+5.2f dB\n",
 500                                 level_compressor.get_level(), to_db(level_compressor.get_level()),
 501                                 level_compressor.get_attenuation(), to_db(level_compressor.get_attenuation()),
 502                                 to_db(level_compressor.get_level() * level_compressor.get_attenuation() * makeup_gain));
 503 #endif
 504
 505                         // The real compressor.
 506                         if (compressor_enabled[bus_index]) {
 507                                 float threshold = from_db(compressor_threshold_dbfs[bus_index]);
 508                                 float ratio = 20.0f;
 509                                 float attack_time = 0.005f;
 510                                 float release_time = 0.040f;
 511                                 float makeup_gain = 2.0f;  // +6 dB.
 512                                 compressor[bus_index]->process(samples_bus.data(), samples_bus.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 513                 //              compressor_att = compressor.get_attenuation();
 514                         }
 515                 }
 516
 517                 add_bus_to_master(bus_index, samples_bus, &samples_out);
 518                 deinterleave_samples(samples_bus, &left, &right);
 519                 measure_bus_levels(bus_index, left, right);
 520         }
 521
 522         {
 523                 lock_guard<mutex> lock(compressor_mutex);
 524
 525                 // Finally a limiter at -4 dB (so, -10 dBFS) to take out the worst peaks only.
 526                 // Note that since ratio is not infinite, we could go slightly higher than this.
 527                 if (limiter_enabled) {
 528                         float threshold = from_db(limiter_threshold_dbfs);
 529                         float ratio = 30.0f;
 530                         float attack_time = 0.0f;  // Instant.
 531                         float release_time = 0.020f;
 532                         float makeup_gain = 1.0f;  // 0 dB.
 533                         limiter.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 534         //              limiter_att = limiter.get_attenuation();
 535                 }
 536
 537         //      printf("limiter=%+5.1f  compressor=%+5.1f\n", to_db(limiter_att), to_db(compressor_att));
 538         }
 539
 540         // At this point, we are most likely close to +0 LU (at least if the
 541         // faders sum to 0 dB and the compressors are on), but all of our
 542         // measurements have been on raw sample values, not R128 values.
 543         // So we have a final makeup gain to get us to +0 LU; the gain
 544         // adjustments required should be relatively small, and also, the
 545         // offset shouldn't change much (only if the type of audio changes
 546         // significantly). Thus, we shoot for updating this value basically
 547         // “whenever we process buffers”, since the R128 calculation isn't exactly
 548         // something we get out per-sample.
 549         //
 550         // Note that there's a feedback loop here, so we choose a very slow filter
 551         // (half-time of 30 seconds).
 552         double target_loudness_factor, alpha;
 553         double loudness_lu = r128.loudness_M() - ref_level_lufs;
 554         double current_makeup_lu = to_db(final_makeup_gain);
 555         target_loudness_factor = final_makeup_gain * from_db(-loudness_lu);
 556
 557         // If we're outside +/- 5 LU uncorrected, we don't count it as
 558         // a normal signal (probably silence) and don't change the
 559         // correction factor; just apply what we already have.
 560         if (fabs(loudness_lu - current_makeup_lu) >= 5.0 || !final_makeup_gain_auto) {
 561                 alpha = 0.0;
 562         } else {
 563                 // Formula adapted from
 564                 // https://en.wikipedia.org/wiki/Low-pass_filter#Simple_infinite_impulse_response_filter.
 565                 const double half_time_s = 30.0;
 566                 const double fc_mul_2pi_delta_t = 1.0 / (half_time_s * OUTPUT_FREQUENCY);
 567                 alpha = fc_mul_2pi_delta_t / (fc_mul_2pi_delta_t + 1.0);
 568         }
 569
 570         {
 571                 lock_guard<mutex> lock(compressor_mutex);
 572                 double m = final_makeup_gain;
 573                 for (size_t i = 0; i < samples_out.size(); i += 2) {
 574                         samples_out[i + 0] *= m;
 575                         samples_out[i + 1] *= m;
 576                         m += (target_loudness_factor - m) * alpha;
 577                 }
 578                 final_makeup_gain = m;
 579         }
 580
 581         update_meters(samples_out);
 582
 583         return samples_out;
 584 }
 585
 586 namespace {
 587
 588 void apply_filter_fade(StereoFilter *filter, float *data, unsigned num_samples, float cutoff_hz, float db, float last_db)
 589 {
 590         // A granularity of 32 samples is an okay tradeoff between speed and
 591         // smoothness; recalculating the filters is pretty expensive, so it's
 592         // good that we don't do this all the time.
 593         static constexpr unsigned filter_granularity_samples = 32;
 594
 595         const float cutoff_linear = cutoff_hz * 2.0 * M_PI / OUTPUT_FREQUENCY;
 596         if (fabs(db - last_db) < 1e-3) {
 597                 // Constant over this frame.
 598                 if (fabs(db) > 0.01f) {
 599                         filter->render(data, num_samples, cutoff_linear, 0.5f, db / 40.0f);
 600                 }
 601         } else {
 602                 // We need to do a fade. (Rounding up avoids division by zero.)
 603                 unsigned num_blocks = (num_samples + filter_granularity_samples - 1) / filter_granularity_samples;
 604                 const float inc_db_norm = (db - last_db) / 40.0f / num_blocks;
 605                 float db_norm = db / 40.0f;
 606                 for (size_t i = 0; i < num_samples; i += filter_granularity_samples) {
 607                         size_t samples_this_block = std::min<size_t>(num_samples - i, filter_granularity_samples);
 608                         filter->render(data + i * 2, samples_this_block, cutoff_linear, 0.5f, db_norm);
 609                         db_norm += inc_db_norm;
 610                 }
 611         }
 612 }
 613
 614 }  // namespace
 615
 616 void AudioMixer::apply_eq(unsigned bus_index, vector<float> *samples_bus)
 617 {
 618         constexpr float bass_freq_hz = 200.0f;
 619         constexpr float treble_freq_hz = 4700.0f;
 620
 621         // Cut away everything under 120 Hz (or whatever the cutoff is);
 622         // we don't need it for voice, and it will reduce headroom
 623         // and confuse the compressor. (In particular, any hums at 50 or 60 Hz
 624         // should be dampened.)
 625         if (locut_enabled[bus_index]) {
 626                 locut[bus_index].render(samples_bus->data(), samples_bus->size() / 2, locut_cutoff_hz * 2.0 * M_PI / OUTPUT_FREQUENCY, 0.5f);
 627         }
 628
 629         // Apply the rest of the EQ. Since we only have a simple three-band EQ,
 630         // we can implement it with two shelf filters. We use a simple gain to
 631         // set the mid-level filter, and then offset the low and high bands
 632         // from that if we need to. (We could perhaps have folded the gain into
 633         // the next part, but it's so cheap that the trouble isn't worth it.)
 634         //
 635         // If any part of the EQ has changed appreciably since last frame,
 636         // we fade smoothly during the course of this frame.
 637         const float bass_db = eq_level_db[bus_index][EQ_BAND_BASS];
 638         const float mid_db = eq_level_db[bus_index][EQ_BAND_MID];
 639         const float treble_db = eq_level_db[bus_index][EQ_BAND_TREBLE];
 640
 641         const float last_bass_db = last_eq_level_db[bus_index][EQ_BAND_BASS];
 642         const float last_mid_db = last_eq_level_db[bus_index][EQ_BAND_MID];
 643         const float last_treble_db = last_eq_level_db[bus_index][EQ_BAND_TREBLE];
 644
 645         assert(samples_bus->size() % 2 == 0);
 646         const unsigned num_samples = samples_bus->size() / 2;
 647
 648         if (fabs(mid_db - last_mid_db) < 1e-3) {
 649                 // Constant over this frame.
 650                 const float gain = from_db(mid_db);
 651                 for (size_t i = 0; i < samples_bus->size(); ++i) {
 652                         (*samples_bus)[i] *= gain;
 653                 }
 654         } else {
 655                 // We need to do a fade.
 656                 float gain = from_db(last_mid_db);
 657                 const float gain_inc = pow(from_db(mid_db - last_mid_db), 1.0 / num_samples);
 658                 for (size_t i = 0; i < num_samples; ++i) {
 659                         (*samples_bus)[i * 2 + 0] *= gain;
 660                         (*samples_bus)[i * 2 + 1] *= gain;
 661                         gain *= gain_inc;
 662                 }
 663         }
 664
 665         apply_filter_fade(&eq[bus_index][EQ_BAND_BASS], samples_bus->data(), num_samples, bass_freq_hz, bass_db - mid_db, last_bass_db - last_mid_db);
 666         apply_filter_fade(&eq[bus_index][EQ_BAND_TREBLE], samples_bus->data(), num_samples, treble_freq_hz, treble_db - mid_db, last_treble_db - last_mid_db);
 667
 668         last_eq_level_db[bus_index][EQ_BAND_BASS] = bass_db;
 669         last_eq_level_db[bus_index][EQ_BAND_MID] = mid_db;
 670         last_eq_level_db[bus_index][EQ_BAND_TREBLE] = treble_db;
 671 }
 672
 673 void AudioMixer::add_bus_to_master(unsigned bus_index, const vector<float> &samples_bus, vector<float> *samples_out)
 674 {
 675         assert(samples_bus.size() == samples_out->size());
 676         assert(samples_bus.size() % 2 == 0);
 677         unsigned num_samples = samples_bus.size() / 2;
 678         if (fabs(fader_volume_db[bus_index] - last_fader_volume_db[bus_index]) > 1e-3) {
 679                 // The volume has changed; do a fade over the course of this frame.
 680                 // (We might have some numerical issues here, but it seems to sound OK.)
 681                 // For the purpose of fading here, the silence floor is set to -90 dB
 682                 // (the fader only goes to -84).
 683                 float old_volume = from_db(max<float>(last_fader_volume_db[bus_index], -90.0f));
 684                 float volume = from_db(max<float>(fader_volume_db[bus_index], -90.0f));
 685
 686                 float volume_inc = pow(volume / old_volume, 1.0 / num_samples);
 687                 volume = old_volume;
 688                 if (bus_index == 0) {
 689                         for (unsigned i = 0; i < num_samples; ++i) {
 690                                 (*samples_out)[i * 2 + 0] = samples_bus[i * 2 + 0] * volume;
 691                                 (*samples_out)[i * 2 + 1] = samples_bus[i * 2 + 1] * volume;
 692                                 volume *= volume_inc;
 693                         }
 694                 } else {
 695                         for (unsigned i = 0; i < num_samples; ++i) {
 696                                 (*samples_out)[i * 2 + 0] += samples_bus[i * 2 + 0] * volume;
 697                                 (*samples_out)[i * 2 + 1] += samples_bus[i * 2 + 1] * volume;
 698                                 volume *= volume_inc;
 699                         }
 700                 }
 701         } else {
 702                 float volume = from_db(fader_volume_db[bus_index]);
 703                 if (bus_index == 0) {
 704                         for (unsigned i = 0; i < num_samples; ++i) {
 705                                 (*samples_out)[i * 2 + 0] = samples_bus[i * 2 + 0] * volume;
 706                                 (*samples_out)[i * 2 + 1] = samples_bus[i * 2 + 1] * volume;
 707                         }
 708                 } else {
 709                         for (unsigned i = 0; i < num_samples; ++i) {
 710                                 (*samples_out)[i * 2 + 0] += samples_bus[i * 2 + 0] * volume;
 711                                 (*samples_out)[i * 2 + 1] += samples_bus[i * 2 + 1] * volume;
 712                         }
 713                 }
 714         }
 715
 716         last_fader_volume_db[bus_index] = fader_volume_db[bus_index];
 717 }
 718
 719 void AudioMixer::measure_bus_levels(unsigned bus_index, const vector<float> &left, const vector<float> &right)
 720 {
 721         assert(left.size() == right.size());
 722         const float volume = from_db(fader_volume_db[bus_index]);
 723         const float peak_levels[2] = {
 724                 find_peak(left.data(), left.size()) * volume,
 725                 find_peak(right.data(), right.size()) * volume
 726         };
 727         for (unsigned channel = 0; channel < 2; ++channel) {
 728                 // Compute the current value, including hold and falloff.
 729                 // The constants are borrowed from zita-mu1 by Fons Adriaensen.
 730                 static constexpr float hold_sec = 0.5f;
 731                 static constexpr float falloff_db_sec = 15.0f;  // dB/sec falloff after hold.
 732                 float current_peak;
 733                 PeakHistory &history = peak_history[bus_index][channel];
 734                 history.historic_peak = max(history.historic_peak, peak_levels[channel]);
 735                 if (history.age_seconds < hold_sec) {
 736                         current_peak = history.last_peak;
 737                 } else {
 738                         current_peak = history.last_peak * from_db(-falloff_db_sec * (history.age_seconds - hold_sec));
 739                 }
 740
 741                 // See if we have a new peak to replace the old (possibly falling) one.
 742                 if (peak_levels[channel] > current_peak) {
 743                         history.last_peak = peak_levels[channel];
 744                         history.age_seconds = 0.0f;  // Not 100% correct, but more than good enough given our frame sizes.
 745                         current_peak = peak_levels[channel];
 746                 } else {
 747                         history.age_seconds += float(left.size()) / OUTPUT_FREQUENCY;
 748                 }
 749                 history.current_level = peak_levels[channel];
 750                 history.current_peak = current_peak;
 751         }
 752 }
 753
 754 void AudioMixer::update_meters(const vector<float> &samples)
 755 {
 756         // Upsample 4x to find interpolated peak.
 757         peak_resampler.inp_data = const_cast<float *>(samples.data());
 758         peak_resampler.inp_count = samples.size() / 2;
 759
 760         vector<float> interpolated_samples;
 761         interpolated_samples.resize(samples.size());
 762         {
 763                 lock_guard<mutex> lock(audio_measure_mutex);
 764
 765                 while (peak_resampler.inp_count > 0) {  // About four iterations.
 766                         peak_resampler.out_data = &interpolated_samples[0];
 767                         peak_resampler.out_count = interpolated_samples.size() / 2;
 768                         peak_resampler.process();
 769                         size_t out_stereo_samples = interpolated_samples.size() / 2 - peak_resampler.out_count;
 770                         peak = max<float>(peak, find_peak(interpolated_samples.data(), out_stereo_samples * 2));
 771                         peak_resampler.out_data = nullptr;
 772                 }
 773         }
 774
 775         // Find R128 levels and L/R correlation.
 776         vector<float> left, right;
 777         deinterleave_samples(samples, &left, &right);
 778         float *ptrs[] = { left.data(), right.data() };
 779         {
 780                 lock_guard<mutex> lock(audio_measure_mutex);
 781                 r128.process(left.size(), ptrs);
 782                 correlation.process_samples(samples);
 783         }
 784
 785         send_audio_level_callback();
 786 }
 787
 788 void AudioMixer::reset_meters()
 789 {
 790         lock_guard<mutex> lock(audio_measure_mutex);
 791         peak_resampler.reset();
 792         peak = 0.0f;
 793         r128.reset();
 794         r128.integr_start();
 795         correlation.reset();
 796 }
 797
 798 void AudioMixer::send_audio_level_callback()
 799 {
 800         if (audio_level_callback == nullptr) {
 801                 return;
 802         }
 803
 804         lock_guard<mutex> lock(audio_measure_mutex);
 805         double loudness_s = r128.loudness_S();
 806         double loudness_i = r128.integrated();
 807         double loudness_range_low = r128.range_min();
 808         double loudness_range_high = r128.range_max();
 809
 810         vector<BusLevel> bus_levels;
 811         bus_levels.resize(input_mapping.buses.size());
 812         {
 813                 lock_guard<mutex> lock(compressor_mutex);
 814                 for (unsigned bus_index = 0; bus_index < bus_levels.size(); ++bus_index) {
 815                         bus_levels[bus_index].current_level_dbfs[0] = to_db(peak_history[bus_index][0].current_level);
 816                         bus_levels[bus_index].current_level_dbfs[1] = to_db(peak_history[bus_index][1].current_level);
 817                         bus_levels[bus_index].peak_level_dbfs[0] = to_db(peak_history[bus_index][0].current_peak);
 818                         bus_levels[bus_index].peak_level_dbfs[1] = to_db(peak_history[bus_index][1].current_peak);
 819                         bus_levels[bus_index].historic_peak_dbfs = to_db(
 820                                 max(peak_history[bus_index][0].historic_peak,
 821                                     peak_history[bus_index][1].historic_peak));
 822                         bus_levels[bus_index].gain_staging_db = gain_staging_db[bus_index];
 823                         if (compressor_enabled[bus_index]) {
 824                                 bus_levels[bus_index].compressor_attenuation_db = -to_db(compressor[bus_index]->get_attenuation());
 825                         } else {
 826                                 bus_levels[bus_index].compressor_attenuation_db = 0.0;
 827                         }
 828                 }
 829         }
 830
 831         audio_level_callback(loudness_s, to_db(peak), bus_levels,
 832                 loudness_i, loudness_range_low, loudness_range_high,
 833                 to_db(final_makeup_gain),
 834                 correlation.get_correlation());
 835 }
 836
 837 map<DeviceSpec, DeviceInfo> AudioMixer::get_devices()
 838 {
 839         lock_guard<timed_mutex> lock(audio_mutex);
 840
 841         map<DeviceSpec, DeviceInfo> devices;
 842         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 843                 const DeviceSpec spec{ InputSourceType::CAPTURE_CARD, card_index };
 844                 const AudioDevice *device = &video_cards[card_index];
 845                 DeviceInfo info;
 846                 info.display_name = device->display_name;
 847                 info.num_channels = 8;
 848                 devices.insert(make_pair(spec, info));
 849         }
 850         vector<ALSAPool::Device> available_alsa_devices = alsa_pool.get_devices();
 851         for (unsigned card_index = 0; card_index < available_alsa_devices.size(); ++card_index) {
 852                 const DeviceSpec spec{ InputSourceType::ALSA_INPUT, card_index };
 853                 const ALSAPool::Device &device = available_alsa_devices[card_index];
 854                 DeviceInfo info;
 855                 info.display_name = device.display_name();
 856                 info.num_channels = device.num_channels;
 857                 info.alsa_name = device.name;
 858                 info.alsa_info = device.info;
 859                 info.alsa_address = device.address;
 860                 devices.insert(make_pair(spec, info));
 861         }
 862         return devices;
 863 }
 864
 865 void AudioMixer::set_display_name(DeviceSpec device_spec, const string &name)
 866 {
 867         AudioDevice *device = find_audio_device(device_spec);
 868
 869         lock_guard<timed_mutex> lock(audio_mutex);
 870         device->display_name = name;
 871 }
 872
 873 void AudioMixer::serialize_device(DeviceSpec device_spec, DeviceSpecProto *device_spec_proto)
 874 {
 875         lock_guard<timed_mutex> lock(audio_mutex);
 876         switch (device_spec.type) {
 877                 case InputSourceType::SILENCE:
 878                         device_spec_proto->set_type(DeviceSpecProto::SILENCE);
 879                         break;
 880                 case InputSourceType::CAPTURE_CARD:
 881                         device_spec_proto->set_type(DeviceSpecProto::CAPTURE_CARD);
 882                         device_spec_proto->set_index(device_spec.index);
 883                         device_spec_proto->set_display_name(video_cards[device_spec.index].display_name);
 884                         break;
 885                 case InputSourceType::ALSA_INPUT:
 886                         alsa_pool.serialize_device(device_spec.index, device_spec_proto);
 887                         break;
 888         }
 889 }
 890
 891 void AudioMixer::set_input_mapping(const InputMapping &new_input_mapping)
 892 {
 893         lock_guard<timed_mutex> lock(audio_mutex);
 894
 895         map<DeviceSpec, set<unsigned>> interesting_channels;
 896         for (const InputMapping::Bus &bus : new_input_mapping.buses) {
 897                 if (bus.device.type == InputSourceType::CAPTURE_CARD ||
 898                     bus.device.type == InputSourceType::ALSA_INPUT) {
 899                         for (unsigned channel = 0; channel < 2; ++channel) {
 900                                 if (bus.source_channel[channel] != -1) {
 901                                         interesting_channels[bus.device].insert(bus.source_channel[channel]);
 902                                 }
 903                         }
 904                 }
 905         }
 906
 907         // Reset resamplers for all cards that don't have the exact same state as before.
 908         for (unsigned card_index = 0; card_index < MAX_VIDEO_CARDS; ++card_index) {
 909                 const DeviceSpec device_spec{InputSourceType::CAPTURE_CARD, card_index};
 910                 AudioDevice *device = find_audio_device(device_spec);
 911                 if (device->interesting_channels != interesting_channels[device_spec]) {
 912                         device->interesting_channels = interesting_channels[device_spec];
 913                         reset_resampler_mutex_held(device_spec);
 914                 }
 915         }
 916         for (unsigned card_index = 0; card_index < MAX_ALSA_CARDS; ++card_index) {
 917                 const DeviceSpec device_spec{InputSourceType::ALSA_INPUT, card_index};
 918                 AudioDevice *device = find_audio_device(device_spec);
 919                 if (interesting_channels[device_spec].empty()) {
 920                         alsa_pool.release_device(card_index);
 921                 } else {
 922                         alsa_pool.hold_device(card_index);
 923                 }
 924                 if (device->interesting_channels != interesting_channels[device_spec]) {
 925                         device->interesting_channels = interesting_channels[device_spec];
 926                         alsa_pool.reset_device(device_spec.index);
 927                         reset_resampler_mutex_held(device_spec);
 928                 }
 929         }
 930
 931         input_mapping = new_input_mapping;
 932 }
 933
 934 InputMapping AudioMixer::get_input_mapping() const
 935 {
 936         lock_guard<timed_mutex> lock(audio_mutex);
 937         return input_mapping;
 938 }
 939
 940 void AudioMixer::reset_peak(unsigned bus_index)
 941 {
 942         lock_guard<timed_mutex> lock(audio_mutex);
 943         for (unsigned channel = 0; channel < 2; ++channel) {
 944                 PeakHistory &history = peak_history[bus_index][channel];
 945                 history.current_level = 0.0f;
 946                 history.historic_peak = 0.0f;
 947                 history.current_peak = 0.0f;
 948                 history.last_peak = 0.0f;
 949                 history.age_seconds = 0.0f;
 950         }
 951 }
 952
 953 AudioMixer *global_audio_mixer = nullptr;