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                 locut_enabled[bus_index] = global_flags.locut_enabled;
 171                 eq[bus_index][EQ_BAND_BASS].init(FILTER_LOW_SHELF, 1);
 172                 // Note: EQ_BAND_MID isn't used (see comments in apply_eq()).
 173                 eq[bus_index][EQ_BAND_TREBLE].init(FILTER_HIGH_SHELF, 1);
 174
 175                 gain_staging_db[bus_index] = global_flags.initial_gain_staging_db;
 176                 compressor[bus_index].reset(new StereoCompressor(OUTPUT_FREQUENCY));
 177                 compressor_threshold_dbfs[bus_index] = ref_level_dbfs - 12.0f;  // -12 dB.
 178                 compressor_enabled[bus_index] = global_flags.compressor_enabled;
 179                 level_compressor[bus_index].reset(new StereoCompressor(OUTPUT_FREQUENCY));
 180                 level_compressor_enabled[bus_index] = global_flags.gain_staging_auto;
 181         }
 182         set_limiter_enabled(global_flags.limiter_enabled);
 183         set_final_makeup_gain_auto(global_flags.final_makeup_gain_auto);
 184
 185         // Generate a very simple, default input mapping.
 186         InputMapping::Bus input;
 187         input.name = "Main";
 188         input.device.type = InputSourceType::CAPTURE_CARD;
 189         input.device.index = 0;
 190         input.source_channel[0] = 0;
 191         input.source_channel[1] = 1;
 192
 193         InputMapping new_input_mapping;
 194         new_input_mapping.buses.push_back(input);
 195         set_input_mapping(new_input_mapping);
 196
 197         alsa_pool.init();
 198
 199         r128.init(2, OUTPUT_FREQUENCY);
 200         r128.integr_start();
 201
 202         // hlen=16 is pretty low quality, but we use quite a bit of CPU otherwise,
 203         // and there's a limit to how important the peak meter is.
 204         peak_resampler.setup(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY * 4, /*num_channels=*/2, /*hlen=*/16, /*frel=*/1.0);
 205 }
 206
 207 void AudioMixer::reset_resampler(DeviceSpec device_spec)
 208 {
 209         lock_guard<timed_mutex> lock(audio_mutex);
 210         reset_resampler_mutex_held(device_spec);
 211 }
 212
 213 void AudioMixer::reset_resampler_mutex_held(DeviceSpec device_spec)
 214 {
 215         AudioDevice *device = find_audio_device(device_spec);
 216
 217         if (device->interesting_channels.empty()) {
 218                 device->resampling_queue.reset();
 219         } else {
 220                 // TODO: ResamplingQueue should probably take the full device spec.
 221                 // (It's only used for console output, though.)
 222                 device->resampling_queue.reset(new ResamplingQueue(device_spec.index, device->capture_frequency, OUTPUT_FREQUENCY, device->interesting_channels.size()));
 223         }
 224         device->next_local_pts = 0;
 225 }
 226
 227 bool AudioMixer::add_audio(DeviceSpec device_spec, const uint8_t *data, unsigned num_samples, AudioFormat audio_format, int64_t frame_length)
 228 {
 229         AudioDevice *device = find_audio_device(device_spec);
 230
 231         unique_lock<timed_mutex> lock(audio_mutex, defer_lock);
 232         if (!lock.try_lock_for(chrono::milliseconds(10))) {
 233                 return false;
 234         }
 235         if (device->resampling_queue == nullptr) {
 236                 // No buses use this device; throw it away.
 237                 return true;
 238         }
 239
 240         unsigned num_channels = device->interesting_channels.size();
 241         assert(num_channels > 0);
 242
 243         // Convert the audio to fp32.
 244         unique_ptr<float[]> audio(new float[num_samples * num_channels]);
 245         unsigned channel_index = 0;
 246         for (auto channel_it = device->interesting_channels.cbegin(); channel_it != device->interesting_channels.end(); ++channel_it, ++channel_index) {
 247                 switch (audio_format.bits_per_sample) {
 248                 case 0:
 249                         assert(num_samples == 0);
 250                         break;
 251                 case 16:
 252                         convert_fixed16_to_fp32(audio.get(), channel_index, num_channels, data, *channel_it, audio_format.num_channels, num_samples);
 253                         break;
 254                 case 24:
 255                         convert_fixed24_to_fp32(audio.get(), channel_index, num_channels, data, *channel_it, audio_format.num_channels, num_samples);
 256                         break;
 257                 case 32:
 258                         convert_fixed32_to_fp32(audio.get(), channel_index, num_channels, data, *channel_it, audio_format.num_channels, num_samples);
 259                         break;
 260                 default:
 261                         fprintf(stderr, "Cannot handle audio with %u bits per sample\n", audio_format.bits_per_sample);
 262                         assert(false);
 263                 }
 264         }
 265
 266         // Now add it.
 267         int64_t local_pts = device->next_local_pts;
 268         device->resampling_queue->add_input_samples(local_pts / double(TIMEBASE), audio.get(), num_samples);
 269         device->next_local_pts = local_pts + frame_length;
 270         return true;
 271 }
 272
 273 bool AudioMixer::add_silence(DeviceSpec device_spec, unsigned samples_per_frame, unsigned num_frames, int64_t frame_length)
 274 {
 275         AudioDevice *device = find_audio_device(device_spec);
 276
 277         unique_lock<timed_mutex> lock(audio_mutex, defer_lock);
 278         if (!lock.try_lock_for(chrono::milliseconds(10))) {
 279                 return false;
 280         }
 281         if (device->resampling_queue == nullptr) {
 282                 // No buses use this device; throw it away.
 283                 return true;
 284         }
 285
 286         unsigned num_channels = device->interesting_channels.size();
 287         assert(num_channels > 0);
 288
 289         vector<float> silence(samples_per_frame * num_channels, 0.0f);
 290         for (unsigned i = 0; i < num_frames; ++i) {
 291                 device->resampling_queue->add_input_samples(device->next_local_pts / double(TIMEBASE), silence.data(), samples_per_frame);
 292                 // Note that if the format changed in the meantime, we have
 293                 // no way of detecting that; we just have to assume the frame length
 294                 // is always the same.
 295                 device->next_local_pts += frame_length;
 296         }
 297         return true;
 298 }
 299
 300 bool AudioMixer::silence_card(DeviceSpec device_spec, bool silence)
 301 {
 302         AudioDevice *device = find_audio_device(device_spec);
 303
 304         unique_lock<timed_mutex> lock(audio_mutex, defer_lock);
 305         if (!lock.try_lock_for(chrono::milliseconds(10))) {
 306                 return false;
 307         }
 308
 309         if (device->silenced && !silence) {
 310                 reset_resampler_mutex_held(device_spec);
 311         }
 312         device->silenced = silence;
 313         return true;
 314 }
 315
 316 AudioMixer::AudioDevice *AudioMixer::find_audio_device(DeviceSpec device)
 317 {
 318         switch (device.type) {
 319         case InputSourceType::CAPTURE_CARD:
 320                 return &video_cards[device.index];
 321         case InputSourceType::ALSA_INPUT:
 322                 return &alsa_inputs[device.index];
 323         case InputSourceType::SILENCE:
 324         default:
 325                 assert(false);
 326         }
 327         return nullptr;
 328 }
 329
 330 // Get a pointer to the given channel from the given device.
 331 // The channel must be picked out earlier and resampled.
 332 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)
 333 {
 334         static float zero = 0.0f;
 335         if (source_channel == -1 || device_spec.type == InputSourceType::SILENCE) {
 336                 *srcptr = &zero;
 337                 *stride = 0;
 338                 return;
 339         }
 340         AudioDevice *device = find_audio_device(device_spec);
 341         assert(device->interesting_channels.count(source_channel) != 0);
 342         unsigned channel_index = 0;
 343         for (int channel : device->interesting_channels) {
 344                 if (channel == source_channel) break;
 345                 ++channel_index;
 346         }
 347         assert(channel_index < device->interesting_channels.size());
 348         const auto it = samples_card.find(device_spec);
 349         assert(it != samples_card.end());
 350         *srcptr = &(it->second)[channel_index];
 351         *stride = device->interesting_channels.size();
 352 }
 353
 354 // TODO: Can be SSSE3-optimized if need be.
 355 void AudioMixer::fill_audio_bus(const map<DeviceSpec, vector<float>> &samples_card, const InputMapping::Bus &bus, unsigned num_samples, float *output)
 356 {
 357         if (bus.device.type == InputSourceType::SILENCE) {
 358                 memset(output, 0, num_samples * sizeof(*output));
 359         } else {
 360                 assert(bus.device.type == InputSourceType::CAPTURE_CARD ||
 361                        bus.device.type == InputSourceType::ALSA_INPUT);
 362                 const float *lsrc, *rsrc;
 363                 unsigned lstride, rstride;
 364                 float *dptr = output;
 365                 find_sample_src_from_device(samples_card, bus.device, bus.source_channel[0], &lsrc, &lstride);
 366                 find_sample_src_from_device(samples_card, bus.device, bus.source_channel[1], &rsrc, &rstride);
 367                 for (unsigned i = 0; i < num_samples; ++i) {
 368                         *dptr++ = *lsrc;
 369                         *dptr++ = *rsrc;
 370                         lsrc += lstride;
 371                         rsrc += rstride;
 372                 }
 373         }
 374 }
 375
 376 vector<DeviceSpec> AudioMixer::get_active_devices() const
 377 {
 378         vector<DeviceSpec> ret;
 379         for (unsigned card_index = 0; card_index < MAX_VIDEO_CARDS; ++card_index) {
 380                 const DeviceSpec device_spec{InputSourceType::CAPTURE_CARD, card_index};
 381                 if (!find_audio_device(device_spec)->interesting_channels.empty()) {
 382                         ret.push_back(device_spec);
 383                 }
 384         }
 385         for (unsigned card_index = 0; card_index < MAX_ALSA_CARDS; ++card_index) {
 386                 const DeviceSpec device_spec{InputSourceType::ALSA_INPUT, card_index};
 387                 if (!find_audio_device(device_spec)->interesting_channels.empty()) {
 388                         ret.push_back(device_spec);
 389                 }
 390         }
 391         return ret;
 392 }
 393
 394 vector<float> AudioMixer::get_output(double pts, unsigned num_samples, ResamplingQueue::RateAdjustmentPolicy rate_adjustment_policy)
 395 {
 396         map<DeviceSpec, vector<float>> samples_card;
 397         vector<float> samples_bus;
 398
 399         lock_guard<timed_mutex> lock(audio_mutex);
 400
 401         // Pick out all the interesting channels from all the cards.
 402         for (const DeviceSpec &device_spec : get_active_devices()) {
 403                 AudioDevice *device = find_audio_device(device_spec);
 404                 samples_card[device_spec].resize(num_samples * device->interesting_channels.size());
 405                 if (device->silenced) {
 406                         memset(&samples_card[device_spec][0], 0, samples_card[device_spec].size() * sizeof(float));
 407                 } else {
 408                         device->resampling_queue->get_output_samples(
 409                                 pts,
 410                                 &samples_card[device_spec][0],
 411                                 num_samples,
 412                                 rate_adjustment_policy);
 413                 }
 414         }
 415
 416         vector<float> samples_out, left, right;
 417         samples_out.resize(num_samples * 2);
 418         samples_bus.resize(num_samples * 2);
 419         for (unsigned bus_index = 0; bus_index < input_mapping.buses.size(); ++bus_index) {
 420                 fill_audio_bus(samples_card, input_mapping.buses[bus_index], num_samples, &samples_bus[0]);
 421                 apply_eq(bus_index, &samples_bus);
 422
 423                 {
 424                         lock_guard<mutex> lock(compressor_mutex);
 425
 426                         // Apply a level compressor to get the general level right.
 427                         // Basically, if it's over about -40 dBFS, we squeeze it down to that level
 428                         // (or more precisely, near it, since we don't use infinite ratio),
 429                         // then apply a makeup gain to get it to -14 dBFS. -14 dBFS is, of course,
 430                         // entirely arbitrary, but from practical tests with speech, it seems to
 431                         // put ut around -23 LUFS, so it's a reasonable starting point for later use.
 432                         if (level_compressor_enabled[bus_index]) {
 433                                 float threshold = 0.01f;   // -40 dBFS.
 434                                 float ratio = 20.0f;
 435                                 float attack_time = 0.5f;
 436                                 float release_time = 20.0f;
 437                                 float makeup_gain = from_db(ref_level_dbfs - (-40.0f));  // +26 dB.
 438                                 level_compressor[bus_index]->process(samples_bus.data(), samples_bus.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 439                                 gain_staging_db[bus_index] = to_db(level_compressor[bus_index]->get_attenuation() * makeup_gain);
 440                         } else {
 441                                 // Just apply the gain we already had.
 442                                 float g = from_db(gain_staging_db[bus_index]);
 443                                 for (size_t i = 0; i < samples_bus.size(); ++i) {
 444                                         samples_bus[i] *= g;
 445                                 }
 446                         }
 447
 448 #if 0
 449                         printf("level=%f (%+5.2f dBFS) attenuation=%f (%+5.2f dB) end_result=%+5.2f dB\n",
 450                                 level_compressor.get_level(), to_db(level_compressor.get_level()),
 451                                 level_compressor.get_attenuation(), to_db(level_compressor.get_attenuation()),
 452                                 to_db(level_compressor.get_level() * level_compressor.get_attenuation() * makeup_gain));
 453 #endif
 454
 455                         // The real compressor.
 456                         if (compressor_enabled[bus_index]) {
 457                                 float threshold = from_db(compressor_threshold_dbfs[bus_index]);
 458                                 float ratio = 20.0f;
 459                                 float attack_time = 0.005f;
 460                                 float release_time = 0.040f;
 461                                 float makeup_gain = 2.0f;  // +6 dB.
 462                                 compressor[bus_index]->process(samples_bus.data(), samples_bus.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 463                 //              compressor_att = compressor.get_attenuation();
 464                         }
 465                 }
 466
 467                 add_bus_to_master(bus_index, samples_bus, &samples_out);
 468                 deinterleave_samples(samples_bus, &left, &right);
 469                 measure_bus_levels(bus_index, left, right);
 470         }
 471
 472         {
 473                 lock_guard<mutex> lock(compressor_mutex);
 474
 475                 // Finally a limiter at -4 dB (so, -10 dBFS) to take out the worst peaks only.
 476                 // Note that since ratio is not infinite, we could go slightly higher than this.
 477                 if (limiter_enabled) {
 478                         float threshold = from_db(limiter_threshold_dbfs);
 479                         float ratio = 30.0f;
 480                         float attack_time = 0.0f;  // Instant.
 481                         float release_time = 0.020f;
 482                         float makeup_gain = 1.0f;  // 0 dB.
 483                         limiter.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 484         //              limiter_att = limiter.get_attenuation();
 485                 }
 486
 487         //      printf("limiter=%+5.1f  compressor=%+5.1f\n", to_db(limiter_att), to_db(compressor_att));
 488         }
 489
 490         // At this point, we are most likely close to +0 LU (at least if the
 491         // faders sum to 0 dB and the compressors are on), but all of our
 492         // measurements have been on raw sample values, not R128 values.
 493         // So we have a final makeup gain to get us to +0 LU; the gain
 494         // adjustments required should be relatively small, and also, the
 495         // offset shouldn't change much (only if the type of audio changes
 496         // significantly). Thus, we shoot for updating this value basically
 497         // “whenever we process buffers”, since the R128 calculation isn't exactly
 498         // something we get out per-sample.
 499         //
 500         // Note that there's a feedback loop here, so we choose a very slow filter
 501         // (half-time of 30 seconds).
 502         double target_loudness_factor, alpha;
 503         double loudness_lu = r128.loudness_M() - ref_level_lufs;
 504         double current_makeup_lu = to_db(final_makeup_gain);
 505         target_loudness_factor = final_makeup_gain * from_db(-loudness_lu);
 506
 507         // If we're outside +/- 5 LU uncorrected, we don't count it as
 508         // a normal signal (probably silence) and don't change the
 509         // correction factor; just apply what we already have.
 510         if (fabs(loudness_lu - current_makeup_lu) >= 5.0 || !final_makeup_gain_auto) {
 511                 alpha = 0.0;
 512         } else {
 513                 // Formula adapted from
 514                 // https://en.wikipedia.org/wiki/Low-pass_filter#Simple_infinite_impulse_response_filter.
 515                 const double half_time_s = 30.0;
 516                 const double fc_mul_2pi_delta_t = 1.0 / (half_time_s * OUTPUT_FREQUENCY);
 517                 alpha = fc_mul_2pi_delta_t / (fc_mul_2pi_delta_t + 1.0);
 518         }
 519
 520         {
 521                 lock_guard<mutex> lock(compressor_mutex);
 522                 double m = final_makeup_gain;
 523                 for (size_t i = 0; i < samples_out.size(); i += 2) {
 524                         samples_out[i + 0] *= m;
 525                         samples_out[i + 1] *= m;
 526                         m += (target_loudness_factor - m) * alpha;
 527                 }
 528                 final_makeup_gain = m;
 529         }
 530
 531         update_meters(samples_out);
 532
 533         return samples_out;
 534 }
 535
 536 void AudioMixer::apply_eq(unsigned bus_index, vector<float> *samples_bus)
 537 {
 538         constexpr float bass_freq_hz = 200.0f;
 539         constexpr float treble_freq_hz = 4700.0f;
 540
 541         // Cut away everything under 120 Hz (or whatever the cutoff is);
 542         // we don't need it for voice, and it will reduce headroom
 543         // and confuse the compressor. (In particular, any hums at 50 or 60 Hz
 544         // should be dampened.)
 545         if (locut_enabled[bus_index]) {
 546                 locut[bus_index].render(samples_bus->data(), samples_bus->size() / 2, locut_cutoff_hz * 2.0 * M_PI / OUTPUT_FREQUENCY, 0.5f);
 547         }
 548
 549         // Apply the rest of the EQ. Since we only have a simple three-band EQ,
 550         // we can implement it with two shelf filters. We use a simple gain to
 551         // set the mid-level filter, and then offset the low and high bands
 552         // from that if we need to. (We could perhaps have folded the gain into
 553         // the next part, but it's so cheap that the trouble isn't worth it.)
 554         if (fabs(eq_level_db[bus_index][EQ_BAND_MID]) > 0.01f) {
 555                 float g = from_db(eq_level_db[bus_index][EQ_BAND_MID]);
 556                 for (size_t i = 0; i < samples_bus->size(); ++i) {
 557                         (*samples_bus)[i] *= g;
 558                 }
 559         }
 560
 561         float bass_adj_db = eq_level_db[bus_index][EQ_BAND_BASS] - eq_level_db[bus_index][EQ_BAND_MID];
 562         if (fabs(bass_adj_db) > 0.01f) {
 563                 eq[bus_index][EQ_BAND_BASS].render(samples_bus->data(), samples_bus->size() / 2,
 564                         bass_freq_hz * 2.0 * M_PI / OUTPUT_FREQUENCY, 0.5f, bass_adj_db / 40.0f);
 565         }
 566
 567         float treble_adj_db = eq_level_db[bus_index][EQ_BAND_TREBLE] - eq_level_db[bus_index][EQ_BAND_MID];
 568         if (fabs(treble_adj_db) > 0.01f) {
 569                 eq[bus_index][EQ_BAND_TREBLE].render(samples_bus->data(), samples_bus->size() / 2,
 570                         treble_freq_hz * 2.0 * M_PI / OUTPUT_FREQUENCY, 0.5f, treble_adj_db / 40.0f);
 571         }
 572 }
 573
 574 void AudioMixer::add_bus_to_master(unsigned bus_index, const vector<float> &samples_bus, vector<float> *samples_out)
 575 {
 576         assert(samples_bus.size() == samples_out->size());
 577         assert(samples_bus.size() % 2 == 0);
 578         unsigned num_samples = samples_bus.size() / 2;
 579         if (fabs(fader_volume_db[bus_index] - last_fader_volume_db[bus_index]) > 1e-3) {
 580                 // The volume has changed; do a fade over the course of this frame.
 581                 // (We might have some numerical issues here, but it seems to sound OK.)
 582                 // For the purpose of fading here, the silence floor is set to -90 dB
 583                 // (the fader only goes to -84).
 584                 float old_volume = from_db(max<float>(last_fader_volume_db[bus_index], -90.0f));
 585                 float volume = from_db(max<float>(fader_volume_db[bus_index], -90.0f));
 586
 587                 float volume_inc = pow(volume / old_volume, 1.0 / num_samples);
 588                 volume = old_volume;
 589                 if (bus_index == 0) {
 590                         for (unsigned i = 0; i < num_samples; ++i) {
 591                                 (*samples_out)[i * 2 + 0] = samples_bus[i * 2 + 0] * volume;
 592                                 (*samples_out)[i * 2 + 1] = samples_bus[i * 2 + 1] * volume;
 593                                 volume *= volume_inc;
 594                         }
 595                 } else {
 596                         for (unsigned i = 0; i < num_samples; ++i) {
 597                                 (*samples_out)[i * 2 + 0] += samples_bus[i * 2 + 0] * volume;
 598                                 (*samples_out)[i * 2 + 1] += samples_bus[i * 2 + 1] * volume;
 599                                 volume *= volume_inc;
 600                         }
 601                 }
 602         } else {
 603                 float volume = from_db(fader_volume_db[bus_index]);
 604                 if (bus_index == 0) {
 605                         for (unsigned i = 0; i < num_samples; ++i) {
 606                                 (*samples_out)[i * 2 + 0] = samples_bus[i * 2 + 0] * volume;
 607                                 (*samples_out)[i * 2 + 1] = samples_bus[i * 2 + 1] * volume;
 608                         }
 609                 } else {
 610                         for (unsigned i = 0; i < num_samples; ++i) {
 611                                 (*samples_out)[i * 2 + 0] += samples_bus[i * 2 + 0] * volume;
 612                                 (*samples_out)[i * 2 + 1] += samples_bus[i * 2 + 1] * volume;
 613                         }
 614                 }
 615         }
 616
 617         last_fader_volume_db[bus_index] = fader_volume_db[bus_index];
 618 }
 619
 620 void AudioMixer::measure_bus_levels(unsigned bus_index, const vector<float> &left, const vector<float> &right)
 621 {
 622         assert(left.size() == right.size());
 623         const float volume = from_db(fader_volume_db[bus_index]);
 624         const float peak_levels[2] = {
 625                 find_peak(left.data(), left.size()) * volume,
 626                 find_peak(right.data(), right.size()) * volume
 627         };
 628         for (unsigned channel = 0; channel < 2; ++channel) {
 629                 // Compute the current value, including hold and falloff.
 630                 // The constants are borrowed from zita-mu1 by Fons Adriaensen.
 631                 static constexpr float hold_sec = 0.5f;
 632                 static constexpr float falloff_db_sec = 15.0f;  // dB/sec falloff after hold.
 633                 float current_peak;
 634                 PeakHistory &history = peak_history[bus_index][channel];
 635                 history.historic_peak = max(history.historic_peak, peak_levels[channel]);
 636                 if (history.age_seconds < hold_sec) {
 637                         current_peak = history.last_peak;
 638                 } else {
 639                         current_peak = history.last_peak * from_db(-falloff_db_sec * (history.age_seconds - hold_sec));
 640                 }
 641
 642                 // See if we have a new peak to replace the old (possibly falling) one.
 643                 if (peak_levels[channel] > current_peak) {
 644                         history.last_peak = peak_levels[channel];
 645                         history.age_seconds = 0.0f;  // Not 100% correct, but more than good enough given our frame sizes.
 646                         current_peak = peak_levels[channel];
 647                 } else {
 648                         history.age_seconds += float(left.size()) / OUTPUT_FREQUENCY;
 649                 }
 650                 history.current_level = peak_levels[channel];
 651                 history.current_peak = current_peak;
 652         }
 653 }
 654
 655 void AudioMixer::update_meters(const vector<float> &samples)
 656 {
 657         // Upsample 4x to find interpolated peak.
 658         peak_resampler.inp_data = const_cast<float *>(samples.data());
 659         peak_resampler.inp_count = samples.size() / 2;
 660
 661         vector<float> interpolated_samples;
 662         interpolated_samples.resize(samples.size());
 663         {
 664                 lock_guard<mutex> lock(audio_measure_mutex);
 665
 666                 while (peak_resampler.inp_count > 0) {  // About four iterations.
 667                         peak_resampler.out_data = &interpolated_samples[0];
 668                         peak_resampler.out_count = interpolated_samples.size() / 2;
 669                         peak_resampler.process();
 670                         size_t out_stereo_samples = interpolated_samples.size() / 2 - peak_resampler.out_count;
 671                         peak = max<float>(peak, find_peak(interpolated_samples.data(), out_stereo_samples * 2));
 672                         peak_resampler.out_data = nullptr;
 673                 }
 674         }
 675
 676         // Find R128 levels and L/R correlation.
 677         vector<float> left, right;
 678         deinterleave_samples(samples, &left, &right);
 679         float *ptrs[] = { left.data(), right.data() };
 680         {
 681                 lock_guard<mutex> lock(audio_measure_mutex);
 682                 r128.process(left.size(), ptrs);
 683                 correlation.process_samples(samples);
 684         }
 685
 686         send_audio_level_callback();
 687 }
 688
 689 void AudioMixer::reset_meters()
 690 {
 691         lock_guard<mutex> lock(audio_measure_mutex);
 692         peak_resampler.reset();
 693         peak = 0.0f;
 694         r128.reset();
 695         r128.integr_start();
 696         correlation.reset();
 697 }
 698
 699 void AudioMixer::send_audio_level_callback()
 700 {
 701         if (audio_level_callback == nullptr) {
 702                 return;
 703         }
 704
 705         lock_guard<mutex> lock(audio_measure_mutex);
 706         double loudness_s = r128.loudness_S();
 707         double loudness_i = r128.integrated();
 708         double loudness_range_low = r128.range_min();
 709         double loudness_range_high = r128.range_max();
 710
 711         vector<BusLevel> bus_levels;
 712         bus_levels.resize(input_mapping.buses.size());
 713         {
 714                 lock_guard<mutex> lock(compressor_mutex);
 715                 for (unsigned bus_index = 0; bus_index < bus_levels.size(); ++bus_index) {
 716                         bus_levels[bus_index].current_level_dbfs[0] = to_db(peak_history[bus_index][0].current_level);
 717                         bus_levels[bus_index].current_level_dbfs[1] = to_db(peak_history[bus_index][1].current_level);
 718                         bus_levels[bus_index].peak_level_dbfs[0] = to_db(peak_history[bus_index][0].current_peak);
 719                         bus_levels[bus_index].peak_level_dbfs[1] = to_db(peak_history[bus_index][1].current_peak);
 720                         bus_levels[bus_index].historic_peak_dbfs = to_db(
 721                                 max(peak_history[bus_index][0].historic_peak,
 722                                     peak_history[bus_index][1].historic_peak));
 723                         bus_levels[bus_index].gain_staging_db = gain_staging_db[bus_index];
 724                         if (compressor_enabled[bus_index]) {
 725                                 bus_levels[bus_index].compressor_attenuation_db = -to_db(compressor[bus_index]->get_attenuation());
 726                         } else {
 727                                 bus_levels[bus_index].compressor_attenuation_db = 0.0;
 728                         }
 729                 }
 730         }
 731
 732         audio_level_callback(loudness_s, to_db(peak), bus_levels,
 733                 loudness_i, loudness_range_low, loudness_range_high,
 734                 to_db(final_makeup_gain),
 735                 correlation.get_correlation());
 736 }
 737
 738 map<DeviceSpec, DeviceInfo> AudioMixer::get_devices()
 739 {
 740         lock_guard<timed_mutex> lock(audio_mutex);
 741
 742         map<DeviceSpec, DeviceInfo> devices;
 743         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 744                 const DeviceSpec spec{ InputSourceType::CAPTURE_CARD, card_index };
 745                 const AudioDevice *device = &video_cards[card_index];
 746                 DeviceInfo info;
 747                 info.display_name = device->display_name;
 748                 info.num_channels = 8;
 749                 devices.insert(make_pair(spec, info));
 750         }
 751         vector<ALSAPool::Device> available_alsa_devices = alsa_pool.get_devices();
 752         for (unsigned card_index = 0; card_index < available_alsa_devices.size(); ++card_index) {
 753                 const DeviceSpec spec{ InputSourceType::ALSA_INPUT, card_index };
 754                 const ALSAPool::Device &device = available_alsa_devices[card_index];
 755                 DeviceInfo info;
 756                 info.display_name = device.display_name();
 757                 info.num_channels = device.num_channels;
 758                 info.alsa_name = device.name;
 759                 info.alsa_info = device.info;
 760                 info.alsa_address = device.address;
 761                 devices.insert(make_pair(spec, info));
 762         }
 763         return devices;
 764 }
 765
 766 void AudioMixer::set_display_name(DeviceSpec device_spec, const string &name)
 767 {
 768         AudioDevice *device = find_audio_device(device_spec);
 769
 770         lock_guard<timed_mutex> lock(audio_mutex);
 771         device->display_name = name;
 772 }
 773
 774 void AudioMixer::serialize_device(DeviceSpec device_spec, DeviceSpecProto *device_spec_proto)
 775 {
 776         lock_guard<timed_mutex> lock(audio_mutex);
 777         switch (device_spec.type) {
 778                 case InputSourceType::SILENCE:
 779                         device_spec_proto->set_type(DeviceSpecProto::SILENCE);
 780                         break;
 781                 case InputSourceType::CAPTURE_CARD:
 782                         device_spec_proto->set_type(DeviceSpecProto::CAPTURE_CARD);
 783                         device_spec_proto->set_index(device_spec.index);
 784                         device_spec_proto->set_display_name(video_cards[device_spec.index].display_name);
 785                         break;
 786                 case InputSourceType::ALSA_INPUT:
 787                         alsa_pool.serialize_device(device_spec.index, device_spec_proto);
 788                         break;
 789         }
 790 }
 791
 792 void AudioMixer::set_input_mapping(const InputMapping &new_input_mapping)
 793 {
 794         lock_guard<timed_mutex> lock(audio_mutex);
 795
 796         map<DeviceSpec, set<unsigned>> interesting_channels;
 797         for (const InputMapping::Bus &bus : new_input_mapping.buses) {
 798                 if (bus.device.type == InputSourceType::CAPTURE_CARD ||
 799                     bus.device.type == InputSourceType::ALSA_INPUT) {
 800                         for (unsigned channel = 0; channel < 2; ++channel) {
 801                                 if (bus.source_channel[channel] != -1) {
 802                                         interesting_channels[bus.device].insert(bus.source_channel[channel]);
 803                                 }
 804                         }
 805                 }
 806         }
 807
 808         // Reset resamplers for all cards that don't have the exact same state as before.
 809         for (unsigned card_index = 0; card_index < MAX_VIDEO_CARDS; ++card_index) {
 810                 const DeviceSpec device_spec{InputSourceType::CAPTURE_CARD, card_index};
 811                 AudioDevice *device = find_audio_device(device_spec);
 812                 if (device->interesting_channels != interesting_channels[device_spec]) {
 813                         device->interesting_channels = interesting_channels[device_spec];
 814                         reset_resampler_mutex_held(device_spec);
 815                 }
 816         }
 817         for (unsigned card_index = 0; card_index < MAX_ALSA_CARDS; ++card_index) {
 818                 const DeviceSpec device_spec{InputSourceType::ALSA_INPUT, card_index};
 819                 AudioDevice *device = find_audio_device(device_spec);
 820                 if (interesting_channels[device_spec].empty()) {
 821                         alsa_pool.release_device(card_index);
 822                 } else {
 823                         alsa_pool.hold_device(card_index);
 824                 }
 825                 if (device->interesting_channels != interesting_channels[device_spec]) {
 826                         device->interesting_channels = interesting_channels[device_spec];
 827                         alsa_pool.reset_device(device_spec.index);
 828                         reset_resampler_mutex_held(device_spec);
 829                 }
 830         }
 831
 832         input_mapping = new_input_mapping;
 833 }
 834
 835 InputMapping AudioMixer::get_input_mapping() const
 836 {
 837         lock_guard<timed_mutex> lock(audio_mutex);
 838         return input_mapping;
 839 }
 840
 841 void AudioMixer::reset_peak(unsigned bus_index)
 842 {
 843         lock_guard<timed_mutex> lock(audio_mutex);
 844         for (unsigned channel = 0; channel < 2; ++channel) {
 845                 PeakHistory &history = peak_history[bus_index][channel];
 846                 history.current_level = 0.0f;
 847                 history.historic_peak = 0.0f;
 848                 history.current_peak = 0.0f;
 849                 history.last_peak = 0.0f;
 850                 history.age_seconds = 0.0f;
 851         }
 852 }
 853
 854 AudioMixer *global_audio_mixer = nullptr;