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