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         for (unsigned bus_index = 0; bus_index < MAX_BUSES; ++bus_index) {
 166                 locut[bus_index].init(FILTER_HPF, 2);
 167                 locut_enabled[bus_index] = global_flags.locut_enabled;
 168                 eq[bus_index][EQ_BAND_BASS].init(FILTER_LOW_SHELF, 1);
 169                 // Note: EQ_BAND_MID isn't used (see comments in apply_eq()).
 170                 eq[bus_index][EQ_BAND_TREBLE].init(FILTER_HIGH_SHELF, 1);
 171
 172                 gain_staging_db[bus_index] = global_flags.initial_gain_staging_db;
 173                 compressor[bus_index].reset(new StereoCompressor(OUTPUT_FREQUENCY));
 174                 compressor_threshold_dbfs[bus_index] = ref_level_dbfs - 12.0f;  // -12 dB.
 175                 compressor_enabled[bus_index] = global_flags.compressor_enabled;
 176                 level_compressor[bus_index].reset(new StereoCompressor(OUTPUT_FREQUENCY));
 177                 level_compressor_enabled[bus_index] = global_flags.gain_staging_auto;
 178         }
 179         set_limiter_enabled(global_flags.limiter_enabled);
 180         set_final_makeup_gain_auto(global_flags.final_makeup_gain_auto);
 181
 182         // Generate a very simple, default input mapping.
 183         InputMapping::Bus input;
 184         input.name = "Main";
 185         input.device.type = InputSourceType::CAPTURE_CARD;
 186         input.device.index = 0;
 187         input.source_channel[0] = 0;
 188         input.source_channel[1] = 1;
 189
 190         InputMapping new_input_mapping;
 191         new_input_mapping.buses.push_back(input);
 192         set_input_mapping(new_input_mapping);
 193
 194         alsa_pool.init();
 195
 196         r128.init(2, OUTPUT_FREQUENCY);
 197         r128.integr_start();
 198
 199         // hlen=16 is pretty low quality, but we use quite a bit of CPU otherwise,
 200         // and there's a limit to how important the peak meter is.
 201         peak_resampler.setup(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY * 4, /*num_channels=*/2, /*hlen=*/16, /*frel=*/1.0);
 202 }
 203
 204 void AudioMixer::reset_resampler(DeviceSpec device_spec)
 205 {
 206         lock_guard<timed_mutex> lock(audio_mutex);
 207         reset_resampler_mutex_held(device_spec);
 208 }
 209
 210 void AudioMixer::reset_resampler_mutex_held(DeviceSpec device_spec)
 211 {
 212         AudioDevice *device = find_audio_device(device_spec);
 213
 214         if (device->interesting_channels.empty()) {
 215                 device->resampling_queue.reset();
 216         } else {
 217                 // TODO: ResamplingQueue should probably take the full device spec.
 218                 // (It's only used for console output, though.)
 219                 device->resampling_queue.reset(new ResamplingQueue(device_spec.index, device->capture_frequency, OUTPUT_FREQUENCY, device->interesting_channels.size()));
 220         }
 221         device->next_local_pts = 0;
 222 }
 223
 224 bool AudioMixer::add_audio(DeviceSpec device_spec, const uint8_t *data, unsigned num_samples, AudioFormat audio_format, int64_t frame_length)
 225 {
 226         AudioDevice *device = find_audio_device(device_spec);
 227
 228         unique_lock<timed_mutex> lock(audio_mutex, defer_lock);
 229         if (!lock.try_lock_for(chrono::milliseconds(10))) {
 230                 return false;
 231         }
 232         if (device->resampling_queue == nullptr) {
 233                 // No buses use this device; throw it away.
 234                 return true;
 235         }
 236
 237         unsigned num_channels = device->interesting_channels.size();
 238         assert(num_channels > 0);
 239
 240         // Convert the audio to fp32.
 241         unique_ptr<float[]> audio(new float[num_samples * num_channels]);
 242         unsigned channel_index = 0;
 243         for (auto channel_it = device->interesting_channels.cbegin(); channel_it != device->interesting_channels.end(); ++channel_it, ++channel_index) {
 244                 switch (audio_format.bits_per_sample) {
 245                 case 0:
 246                         assert(num_samples == 0);
 247                         break;
 248                 case 16:
 249                         convert_fixed16_to_fp32(audio.get(), channel_index, num_channels, data, *channel_it, audio_format.num_channels, num_samples);
 250                         break;
 251                 case 24:
 252                         convert_fixed24_to_fp32(audio.get(), channel_index, num_channels, data, *channel_it, audio_format.num_channels, num_samples);
 253                         break;
 254                 case 32:
 255                         convert_fixed32_to_fp32(audio.get(), channel_index, num_channels, data, *channel_it, audio_format.num_channels, num_samples);
 256                         break;
 257                 default:
 258                         fprintf(stderr, "Cannot handle audio with %u bits per sample\n", audio_format.bits_per_sample);
 259                         assert(false);
 260                 }
 261         }
 262
 263         // Now add it.
 264         int64_t local_pts = device->next_local_pts;
 265         device->resampling_queue->add_input_samples(local_pts / double(TIMEBASE), audio.get(), num_samples);
 266         device->next_local_pts = local_pts + frame_length;
 267         return true;
 268 }
 269
 270 bool AudioMixer::add_silence(DeviceSpec device_spec, unsigned samples_per_frame, unsigned num_frames, int64_t frame_length)
 271 {
 272         AudioDevice *device = find_audio_device(device_spec);
 273
 274         unique_lock<timed_mutex> lock(audio_mutex, defer_lock);
 275         if (!lock.try_lock_for(chrono::milliseconds(10))) {
 276                 return false;
 277         }
 278         if (device->resampling_queue == nullptr) {
 279                 // No buses use this device; throw it away.
 280                 return true;
 281         }
 282
 283         unsigned num_channels = device->interesting_channels.size();
 284         assert(num_channels > 0);
 285
 286         vector<float> silence(samples_per_frame * num_channels, 0.0f);
 287         for (unsigned i = 0; i < num_frames; ++i) {
 288                 device->resampling_queue->add_input_samples(device->next_local_pts / double(TIMEBASE), silence.data(), samples_per_frame);
 289                 // Note that if the format changed in the meantime, we have
 290                 // no way of detecting that; we just have to assume the frame length
 291                 // is always the same.
 292                 device->next_local_pts += frame_length;
 293         }
 294         return true;
 295 }
 296
 297 bool AudioMixer::silence_card(DeviceSpec device_spec, bool silence)
 298 {
 299         AudioDevice *device = find_audio_device(device_spec);
 300
 301         unique_lock<timed_mutex> lock(audio_mutex, defer_lock);
 302         if (!lock.try_lock_for(chrono::milliseconds(10))) {
 303                 return false;
 304         }
 305
 306         if (device->silenced && !silence) {
 307                 reset_resampler_mutex_held(device_spec);
 308         }
 309         device->silenced = silence;
 310         return true;
 311 }
 312
 313 AudioMixer::AudioDevice *AudioMixer::find_audio_device(DeviceSpec device)
 314 {
 315         switch (device.type) {
 316         case InputSourceType::CAPTURE_CARD:
 317                 return &video_cards[device.index];
 318         case InputSourceType::ALSA_INPUT:
 319                 return &alsa_inputs[device.index];
 320         case InputSourceType::SILENCE:
 321         default:
 322                 assert(false);
 323         }
 324         return nullptr;
 325 }
 326
 327 // Get a pointer to the given channel from the given device.
 328 // The channel must be picked out earlier and resampled.
 329 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)
 330 {
 331         static float zero = 0.0f;
 332         if (source_channel == -1 || device_spec.type == InputSourceType::SILENCE) {
 333                 *srcptr = &zero;
 334                 *stride = 0;
 335                 return;
 336         }
 337         AudioDevice *device = find_audio_device(device_spec);
 338         assert(device->interesting_channels.count(source_channel) != 0);
 339         unsigned channel_index = 0;
 340         for (int channel : device->interesting_channels) {
 341                 if (channel == source_channel) break;
 342                 ++channel_index;
 343         }
 344         assert(channel_index < device->interesting_channels.size());
 345         const auto it = samples_card.find(device_spec);
 346         assert(it != samples_card.end());
 347         *srcptr = &(it->second)[channel_index];
 348         *stride = device->interesting_channels.size();
 349 }
 350
 351 // TODO: Can be SSSE3-optimized if need be.
 352 void AudioMixer::fill_audio_bus(const map<DeviceSpec, vector<float>> &samples_card, const InputMapping::Bus &bus, unsigned num_samples, float *output)
 353 {
 354         if (bus.device.type == InputSourceType::SILENCE) {
 355                 memset(output, 0, num_samples * sizeof(*output));
 356         } else {
 357                 assert(bus.device.type == InputSourceType::CAPTURE_CARD ||
 358                        bus.device.type == InputSourceType::ALSA_INPUT);
 359                 const float *lsrc, *rsrc;
 360                 unsigned lstride, rstride;
 361                 float *dptr = output;
 362                 find_sample_src_from_device(samples_card, bus.device, bus.source_channel[0], &lsrc, &lstride);
 363                 find_sample_src_from_device(samples_card, bus.device, bus.source_channel[1], &rsrc, &rstride);
 364                 for (unsigned i = 0; i < num_samples; ++i) {
 365                         *dptr++ = *lsrc;
 366                         *dptr++ = *rsrc;
 367                         lsrc += lstride;
 368                         rsrc += rstride;
 369                 }
 370         }
 371 }
 372
 373 vector<DeviceSpec> AudioMixer::get_active_devices() const
 374 {
 375         vector<DeviceSpec> ret;
 376         for (unsigned card_index = 0; card_index < MAX_VIDEO_CARDS; ++card_index) {
 377                 const DeviceSpec device_spec{InputSourceType::CAPTURE_CARD, card_index};
 378                 if (!find_audio_device(device_spec)->interesting_channels.empty()) {
 379                         ret.push_back(device_spec);
 380                 }
 381         }
 382         for (unsigned card_index = 0; card_index < MAX_ALSA_CARDS; ++card_index) {
 383                 const DeviceSpec device_spec{InputSourceType::ALSA_INPUT, card_index};
 384                 if (!find_audio_device(device_spec)->interesting_channels.empty()) {
 385                         ret.push_back(device_spec);
 386                 }
 387         }
 388         return ret;
 389 }
 390
 391 vector<float> AudioMixer::get_output(double pts, unsigned num_samples, ResamplingQueue::RateAdjustmentPolicy rate_adjustment_policy)
 392 {
 393         map<DeviceSpec, vector<float>> samples_card;
 394         vector<float> samples_bus;
 395
 396         lock_guard<timed_mutex> lock(audio_mutex);
 397
 398         // Pick out all the interesting channels from all the cards.
 399         for (const DeviceSpec &device_spec : get_active_devices()) {
 400                 AudioDevice *device = find_audio_device(device_spec);
 401                 samples_card[device_spec].resize(num_samples * device->interesting_channels.size());
 402                 if (device->silenced) {
 403                         memset(&samples_card[device_spec][0], 0, samples_card[device_spec].size() * sizeof(float));
 404                 } else {
 405                         device->resampling_queue->get_output_samples(
 406                                 pts,
 407                                 &samples_card[device_spec][0],
 408                                 num_samples,
 409                                 rate_adjustment_policy);
 410                 }
 411         }
 412
 413         vector<float> samples_out, left, right;
 414         samples_out.resize(num_samples * 2);
 415         samples_bus.resize(num_samples * 2);
 416         for (unsigned bus_index = 0; bus_index < input_mapping.buses.size(); ++bus_index) {
 417                 fill_audio_bus(samples_card, input_mapping.buses[bus_index], num_samples, &samples_bus[0]);
 418                 apply_eq(bus_index, &samples_bus);
 419
 420                 {
 421                         lock_guard<mutex> lock(compressor_mutex);
 422
 423                         // Apply a level compressor to get the general level right.
 424                         // Basically, if it's over about -40 dBFS, we squeeze it down to that level
 425                         // (or more precisely, near it, since we don't use infinite ratio),
 426                         // then apply a makeup gain to get it to -14 dBFS. -14 dBFS is, of course,
 427                         // entirely arbitrary, but from practical tests with speech, it seems to
 428                         // put ut around -23 LUFS, so it's a reasonable starting point for later use.
 429                         if (level_compressor_enabled[bus_index]) {
 430                                 float threshold = 0.01f;   // -40 dBFS.
 431                                 float ratio = 20.0f;
 432                                 float attack_time = 0.5f;
 433                                 float release_time = 20.0f;
 434                                 float makeup_gain = from_db(ref_level_dbfs - (-40.0f));  // +26 dB.
 435                                 level_compressor[bus_index]->process(samples_bus.data(), samples_bus.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 436                                 gain_staging_db[bus_index] = to_db(level_compressor[bus_index]->get_attenuation() * makeup_gain);
 437                         } else {
 438                                 // Just apply the gain we already had.
 439                                 float g = from_db(gain_staging_db[bus_index]);
 440                                 for (size_t i = 0; i < samples_bus.size(); ++i) {
 441                                         samples_bus[i] *= g;
 442                                 }
 443                         }
 444
 445 #if 0
 446                         printf("level=%f (%+5.2f dBFS) attenuation=%f (%+5.2f dB) end_result=%+5.2f dB\n",
 447                                 level_compressor.get_level(), to_db(level_compressor.get_level()),
 448                                 level_compressor.get_attenuation(), to_db(level_compressor.get_attenuation()),
 449                                 to_db(level_compressor.get_level() * level_compressor.get_attenuation() * makeup_gain));
 450 #endif
 451
 452                         // The real compressor.
 453                         if (compressor_enabled[bus_index]) {
 454                                 float threshold = from_db(compressor_threshold_dbfs[bus_index]);
 455                                 float ratio = 20.0f;
 456                                 float attack_time = 0.005f;
 457                                 float release_time = 0.040f;
 458                                 float makeup_gain = 2.0f;  // +6 dB.
 459                                 compressor[bus_index]->process(samples_bus.data(), samples_bus.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 460                 //              compressor_att = compressor.get_attenuation();
 461                         }
 462                 }
 463
 464                 add_bus_to_master(bus_index, samples_bus, &samples_out);
 465                 deinterleave_samples(samples_bus, &left, &right);
 466                 measure_bus_levels(bus_index, left, right);
 467         }
 468
 469         {
 470                 lock_guard<mutex> lock(compressor_mutex);
 471
 472                 // Finally a limiter at -4 dB (so, -10 dBFS) to take out the worst peaks only.
 473                 // Note that since ratio is not infinite, we could go slightly higher than this.
 474                 if (limiter_enabled) {
 475                         float threshold = from_db(limiter_threshold_dbfs);
 476                         float ratio = 30.0f;
 477                         float attack_time = 0.0f;  // Instant.
 478                         float release_time = 0.020f;
 479                         float makeup_gain = 1.0f;  // 0 dB.
 480                         limiter.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 481         //              limiter_att = limiter.get_attenuation();
 482                 }
 483
 484         //      printf("limiter=%+5.1f  compressor=%+5.1f\n", to_db(limiter_att), to_db(compressor_att));
 485         }
 486
 487         // At this point, we are most likely close to +0 LU (at least if the
 488         // faders sum to 0 dB and the compressors are on), but all of our
 489         // measurements have been on raw sample values, not R128 values.
 490         // So we have a final makeup gain to get us to +0 LU; the gain
 491         // adjustments required should be relatively small, and also, the
 492         // offset shouldn't change much (only if the type of audio changes
 493         // significantly). Thus, we shoot for updating this value basically
 494         // “whenever we process buffers”, since the R128 calculation isn't exactly
 495         // something we get out per-sample.
 496         //
 497         // Note that there's a feedback loop here, so we choose a very slow filter
 498         // (half-time of 30 seconds).
 499         double target_loudness_factor, alpha;
 500         double loudness_lu = r128.loudness_M() - ref_level_lufs;
 501         double current_makeup_lu = to_db(final_makeup_gain);
 502         target_loudness_factor = final_makeup_gain * from_db(-loudness_lu);
 503
 504         // If we're outside +/- 5 LU uncorrected, we don't count it as
 505         // a normal signal (probably silence) and don't change the
 506         // correction factor; just apply what we already have.
 507         if (fabs(loudness_lu - current_makeup_lu) >= 5.0 || !final_makeup_gain_auto) {
 508                 alpha = 0.0;
 509         } else {
 510                 // Formula adapted from
 511                 // https://en.wikipedia.org/wiki/Low-pass_filter#Simple_infinite_impulse_response_filter.
 512                 const double half_time_s = 30.0;
 513                 const double fc_mul_2pi_delta_t = 1.0 / (half_time_s * OUTPUT_FREQUENCY);
 514                 alpha = fc_mul_2pi_delta_t / (fc_mul_2pi_delta_t + 1.0);
 515         }
 516
 517         {
 518                 lock_guard<mutex> lock(compressor_mutex);
 519                 double m = final_makeup_gain;
 520                 for (size_t i = 0; i < samples_out.size(); i += 2) {
 521                         samples_out[i + 0] *= m;
 522                         samples_out[i + 1] *= m;
 523                         m += (target_loudness_factor - m) * alpha;
 524                 }
 525                 final_makeup_gain = m;
 526         }
 527
 528         update_meters(samples_out);
 529
 530         return samples_out;
 531 }
 532
 533 void AudioMixer::apply_eq(unsigned bus_index, vector<float> *samples_bus)
 534 {
 535         constexpr float bass_freq_hz = 200.0f;
 536         constexpr float treble_freq_hz = 4700.0f;
 537
 538         // Cut away everything under 120 Hz (or whatever the cutoff is);
 539         // we don't need it for voice, and it will reduce headroom
 540         // and confuse the compressor. (In particular, any hums at 50 or 60 Hz
 541         // should be dampened.)
 542         if (locut_enabled[bus_index]) {
 543                 locut[bus_index].render(samples_bus->data(), samples_bus->size() / 2, locut_cutoff_hz * 2.0 * M_PI / OUTPUT_FREQUENCY, 0.5f);
 544         }
 545
 546         // Apply the rest of the EQ. Since we only have a simple three-band EQ,
 547         // we can implement it with two shelf filters. We use a simple gain to
 548         // set the mid-level filter, and then offset the low and high bands
 549         // from that if we need to. (We could perhaps have folded the gain into
 550         // the next part, but it's so cheap that the trouble isn't worth it.)
 551         if (fabs(eq_level_db[bus_index][EQ_BAND_MID]) > 0.01f) {
 552                 float g = from_db(eq_level_db[bus_index][EQ_BAND_MID]);
 553                 for (size_t i = 0; i < samples_bus->size(); ++i) {
 554                         (*samples_bus)[i] *= g;
 555                 }
 556         }
 557
 558         float bass_adj_db = eq_level_db[bus_index][EQ_BAND_BASS] - eq_level_db[bus_index][EQ_BAND_MID];
 559         if (fabs(bass_adj_db) > 0.01f) {
 560                 eq[bus_index][EQ_BAND_BASS].render(samples_bus->data(), samples_bus->size() / 2,
 561                         bass_freq_hz * 2.0 * M_PI / OUTPUT_FREQUENCY, 0.5f, bass_adj_db / 40.0f);
 562         }
 563
 564         float treble_adj_db = eq_level_db[bus_index][EQ_BAND_TREBLE] - eq_level_db[bus_index][EQ_BAND_MID];
 565         if (fabs(treble_adj_db) > 0.01f) {
 566                 eq[bus_index][EQ_BAND_TREBLE].render(samples_bus->data(), samples_bus->size() / 2,
 567                         treble_freq_hz * 2.0 * M_PI / OUTPUT_FREQUENCY, 0.5f, treble_adj_db / 40.0f);
 568         }
 569 }
 570
 571 void AudioMixer::add_bus_to_master(unsigned bus_index, const vector<float> &samples_bus, vector<float> *samples_out)
 572 {
 573         assert(samples_bus.size() == samples_out->size());
 574         assert(samples_bus.size() % 2 == 0);
 575         unsigned num_samples = samples_bus.size() / 2;
 576         if (fabs(fader_volume_db[bus_index] - last_fader_volume_db[bus_index]) > 1e-3) {
 577                 // The volume has changed; do a fade over the course of this frame.
 578                 // (We might have some numerical issues here, but it seems to sound OK.)
 579                 // For the purpose of fading here, the silence floor is set to -90 dB
 580                 // (the fader only goes to -84).
 581                 float old_volume = from_db(max<float>(last_fader_volume_db[bus_index], -90.0f));
 582                 float volume = from_db(max<float>(fader_volume_db[bus_index], -90.0f));
 583
 584                 float volume_inc = pow(volume / old_volume, 1.0 / num_samples);
 585                 volume = old_volume;
 586                 if (bus_index == 0) {
 587                         for (unsigned i = 0; i < num_samples; ++i) {
 588                                 (*samples_out)[i * 2 + 0] = samples_bus[i * 2 + 0] * volume;
 589                                 (*samples_out)[i * 2 + 1] = samples_bus[i * 2 + 1] * volume;
 590                                 volume *= volume_inc;
 591                         }
 592                 } else {
 593                         for (unsigned i = 0; i < num_samples; ++i) {
 594                                 (*samples_out)[i * 2 + 0] += samples_bus[i * 2 + 0] * volume;
 595                                 (*samples_out)[i * 2 + 1] += samples_bus[i * 2 + 1] * volume;
 596                                 volume *= volume_inc;
 597                         }
 598                 }
 599         } else {
 600                 float volume = from_db(fader_volume_db[bus_index]);
 601                 if (bus_index == 0) {
 602                         for (unsigned i = 0; i < num_samples; ++i) {
 603                                 (*samples_out)[i * 2 + 0] = samples_bus[i * 2 + 0] * volume;
 604                                 (*samples_out)[i * 2 + 1] = samples_bus[i * 2 + 1] * volume;
 605                         }
 606                 } else {
 607                         for (unsigned i = 0; i < num_samples; ++i) {
 608                                 (*samples_out)[i * 2 + 0] += samples_bus[i * 2 + 0] * volume;
 609                                 (*samples_out)[i * 2 + 1] += samples_bus[i * 2 + 1] * volume;
 610                         }
 611                 }
 612         }
 613
 614         last_fader_volume_db[bus_index] = fader_volume_db[bus_index];
 615 }
 616
 617 void AudioMixer::measure_bus_levels(unsigned bus_index, const vector<float> &left, const vector<float> &right)
 618 {
 619         assert(left.size() == right.size());
 620         const float volume = from_db(fader_volume_db[bus_index]);
 621         const float peak_levels[2] = {
 622                 find_peak(left.data(), left.size()) * volume,
 623                 find_peak(right.data(), right.size()) * volume
 624         };
 625         for (unsigned channel = 0; channel < 2; ++channel) {
 626                 // Compute the current value, including hold and falloff.
 627                 // The constants are borrowed from zita-mu1 by Fons Adriaensen.
 628                 static constexpr float hold_sec = 0.5f;
 629                 static constexpr float falloff_db_sec = 15.0f;  // dB/sec falloff after hold.
 630                 float current_peak;
 631                 PeakHistory &history = peak_history[bus_index][channel];
 632                 history.historic_peak = max(history.historic_peak, peak_levels[channel]);
 633                 if (history.age_seconds < hold_sec) {
 634                         current_peak = history.last_peak;
 635                 } else {
 636                         current_peak = history.last_peak * from_db(-falloff_db_sec * (history.age_seconds - hold_sec));
 637                 }
 638
 639                 // See if we have a new peak to replace the old (possibly falling) one.
 640                 if (peak_levels[channel] > current_peak) {
 641                         history.last_peak = peak_levels[channel];
 642                         history.age_seconds = 0.0f;  // Not 100% correct, but more than good enough given our frame sizes.
 643                         current_peak = peak_levels[channel];
 644                 } else {
 645                         history.age_seconds += float(left.size()) / OUTPUT_FREQUENCY;
 646                 }
 647                 history.current_level = peak_levels[channel];
 648                 history.current_peak = current_peak;
 649         }
 650 }
 651
 652 void AudioMixer::update_meters(const vector<float> &samples)
 653 {
 654         // Upsample 4x to find interpolated peak.
 655         peak_resampler.inp_data = const_cast<float *>(samples.data());
 656         peak_resampler.inp_count = samples.size() / 2;
 657
 658         vector<float> interpolated_samples;
 659         interpolated_samples.resize(samples.size());
 660         {
 661                 lock_guard<mutex> lock(audio_measure_mutex);
 662
 663                 while (peak_resampler.inp_count > 0) {  // About four iterations.
 664                         peak_resampler.out_data = &interpolated_samples[0];
 665                         peak_resampler.out_count = interpolated_samples.size() / 2;
 666                         peak_resampler.process();
 667                         size_t out_stereo_samples = interpolated_samples.size() / 2 - peak_resampler.out_count;
 668                         peak = max<float>(peak, find_peak(interpolated_samples.data(), out_stereo_samples * 2));
 669                         peak_resampler.out_data = nullptr;
 670                 }
 671         }
 672
 673         // Find R128 levels and L/R correlation.
 674         vector<float> left, right;
 675         deinterleave_samples(samples, &left, &right);
 676         float *ptrs[] = { left.data(), right.data() };
 677         {
 678                 lock_guard<mutex> lock(audio_measure_mutex);
 679                 r128.process(left.size(), ptrs);
 680                 correlation.process_samples(samples);
 681         }
 682
 683         send_audio_level_callback();
 684 }
 685
 686 void AudioMixer::reset_meters()
 687 {
 688         lock_guard<mutex> lock(audio_measure_mutex);
 689         peak_resampler.reset();
 690         peak = 0.0f;
 691         r128.reset();
 692         r128.integr_start();
 693         correlation.reset();
 694 }
 695
 696 void AudioMixer::send_audio_level_callback()
 697 {
 698         if (audio_level_callback == nullptr) {
 699                 return;
 700         }
 701
 702         lock_guard<mutex> lock(audio_measure_mutex);
 703         double loudness_s = r128.loudness_S();
 704         double loudness_i = r128.integrated();
 705         double loudness_range_low = r128.range_min();
 706         double loudness_range_high = r128.range_max();
 707
 708         vector<BusLevel> bus_levels;
 709         bus_levels.resize(input_mapping.buses.size());
 710         {
 711                 lock_guard<mutex> lock(compressor_mutex);
 712                 for (unsigned bus_index = 0; bus_index < bus_levels.size(); ++bus_index) {
 713                         bus_levels[bus_index].current_level_dbfs[0] = to_db(peak_history[bus_index][0].current_level);
 714                         bus_levels[bus_index].current_level_dbfs[1] = to_db(peak_history[bus_index][1].current_level);
 715                         bus_levels[bus_index].peak_level_dbfs[0] = to_db(peak_history[bus_index][0].current_peak);
 716                         bus_levels[bus_index].peak_level_dbfs[1] = to_db(peak_history[bus_index][1].current_peak);
 717                         bus_levels[bus_index].historic_peak_dbfs = to_db(
 718                                 max(peak_history[bus_index][0].historic_peak,
 719                                     peak_history[bus_index][1].historic_peak));
 720                         bus_levels[bus_index].gain_staging_db = gain_staging_db[bus_index];
 721                         if (compressor_enabled[bus_index]) {
 722                                 bus_levels[bus_index].compressor_attenuation_db = -to_db(compressor[bus_index]->get_attenuation());
 723                         } else {
 724                                 bus_levels[bus_index].compressor_attenuation_db = 0.0;
 725                         }
 726                 }
 727         }
 728
 729         audio_level_callback(loudness_s, to_db(peak), bus_levels,
 730                 loudness_i, loudness_range_low, loudness_range_high,
 731                 to_db(final_makeup_gain),
 732                 correlation.get_correlation());
 733 }
 734
 735 map<DeviceSpec, DeviceInfo> AudioMixer::get_devices()
 736 {
 737         lock_guard<timed_mutex> lock(audio_mutex);
 738
 739         map<DeviceSpec, DeviceInfo> devices;
 740         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 741                 const DeviceSpec spec{ InputSourceType::CAPTURE_CARD, card_index };
 742                 const AudioDevice *device = &video_cards[card_index];
 743                 DeviceInfo info;
 744                 info.name = device->name;
 745                 info.num_channels = 8;
 746                 devices.insert(make_pair(spec, info));
 747         }
 748         vector<ALSAPool::Device> available_alsa_devices = alsa_pool.get_devices();
 749         for (unsigned card_index = 0; card_index < available_alsa_devices.size(); ++card_index) {
 750                 const DeviceSpec spec{ InputSourceType::ALSA_INPUT, card_index };
 751                 const ALSAPool::Device &device = available_alsa_devices[card_index];
 752                 DeviceInfo info;
 753                 info.name = device.name + " (" + device.info + ")";
 754                 info.num_channels = device.num_channels;
 755                 devices.insert(make_pair(spec, info));
 756         }
 757         return devices;
 758 }
 759
 760 void AudioMixer::set_name(DeviceSpec device_spec, const string &name)
 761 {
 762         AudioDevice *device = find_audio_device(device_spec);
 763
 764         lock_guard<timed_mutex> lock(audio_mutex);
 765         device->name = name;
 766 }
 767
 768 void AudioMixer::set_input_mapping(const InputMapping &new_input_mapping)
 769 {
 770         lock_guard<timed_mutex> lock(audio_mutex);
 771
 772         map<DeviceSpec, set<unsigned>> interesting_channels;
 773         for (const InputMapping::Bus &bus : new_input_mapping.buses) {
 774                 if (bus.device.type == InputSourceType::CAPTURE_CARD ||
 775                     bus.device.type == InputSourceType::ALSA_INPUT) {
 776                         for (unsigned channel = 0; channel < 2; ++channel) {
 777                                 if (bus.source_channel[channel] != -1) {
 778                                         interesting_channels[bus.device].insert(bus.source_channel[channel]);
 779                                 }
 780                         }
 781                 }
 782         }
 783
 784         // Reset resamplers for all cards that don't have the exact same state as before.
 785         for (unsigned card_index = 0; card_index < MAX_VIDEO_CARDS; ++card_index) {
 786                 const DeviceSpec device_spec{InputSourceType::CAPTURE_CARD, card_index};
 787                 AudioDevice *device = find_audio_device(device_spec);
 788                 if (device->interesting_channels != interesting_channels[device_spec]) {
 789                         device->interesting_channels = interesting_channels[device_spec];
 790                         reset_resampler_mutex_held(device_spec);
 791                 }
 792         }
 793         for (unsigned card_index = 0; card_index < MAX_ALSA_CARDS; ++card_index) {
 794                 const DeviceSpec device_spec{InputSourceType::ALSA_INPUT, card_index};
 795                 AudioDevice *device = find_audio_device(device_spec);
 796                 if (interesting_channels[device_spec].empty()) {
 797                         alsa_pool.release_device(card_index);
 798                 } else {
 799                         alsa_pool.hold_device(card_index);
 800                 }
 801                 if (device->interesting_channels != interesting_channels[device_spec]) {
 802                         device->interesting_channels = interesting_channels[device_spec];
 803                         alsa_pool.reset_device(device_spec.index);
 804                         reset_resampler_mutex_held(device_spec);
 805                 }
 806         }
 807
 808         input_mapping = new_input_mapping;
 809 }
 810
 811 InputMapping AudioMixer::get_input_mapping() const
 812 {
 813         lock_guard<timed_mutex> lock(audio_mutex);
 814         return input_mapping;
 815 }
 816
 817 void AudioMixer::reset_peak(unsigned bus_index)
 818 {
 819         lock_guard<timed_mutex> lock(audio_mutex);
 820         for (unsigned channel = 0; channel < 2; ++channel) {
 821                 PeakHistory &history = peak_history[bus_index][channel];
 822                 history.current_level = 0.0f;
 823                 history.historic_peak = 0.0f;
 824                 history.current_peak = 0.0f;
 825                 history.last_peak = 0.0f;
 826                 history.age_seconds = 0.0f;
 827         }
 828 }
 829
 830 AudioMixer *global_audio_mixer = nullptr;