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

   1 #include "audio_mixer.h"
   2
   3 #include <assert.h>
   4 #include <bmusb/bmusb.h>
   5 #include <endian.h>
   6 #include <math.h>
   7 #ifdef __SSE2__
   8 #include <immintrin.h>
   9 #endif
  10 #include <stdbool.h>
  11 #include <stdio.h>
  12 #include <stdlib.h>
  13 #include <string.h>
  14 #include <algorithm>
  15 #include <chrono>
  16 #include <cmath>
  17 #include <cstddef>
  18 #include <limits>
  19 #include <utility>
  20
  21 #include "db.h"
  22 #include "flags.h"
  23 #include "metrics.h"
  24 #include "state.pb.h"
  25 #include "timebase.h"
  26
  27 using namespace bmusb;
  28 using namespace std;
  29 using namespace std::chrono;
  30 using namespace std::placeholders;
  31
  32 namespace {
  33
  34 // TODO: If these prove to be a bottleneck, they can be SSSE3-optimized
  35 // (usually including multiple channels at a time).
  36
  37 void convert_fixed16_to_fp32(float *dst, size_t out_channel, size_t out_num_channels,
  38                              const uint8_t *src, size_t in_channel, size_t in_num_channels,
  39                              size_t num_samples)
  40 {
  41         assert(in_channel < in_num_channels);
  42         assert(out_channel < out_num_channels);
  43         src += in_channel * 2;
  44         dst += out_channel;
  45
  46         for (size_t i = 0; i < num_samples; ++i) {
  47                 int16_t s = le16toh(*(int16_t *)src);
  48                 *dst = s * (1.0f / 32768.0f);
  49
  50                 src += 2 * in_num_channels;
  51                 dst += out_num_channels;
  52         }
  53 }
  54
  55 void convert_fixed24_to_fp32(float *dst, size_t out_channel, size_t out_num_channels,
  56                              const uint8_t *src, size_t in_channel, size_t in_num_channels,
  57                              size_t num_samples)
  58 {
  59         assert(in_channel < in_num_channels);
  60         assert(out_channel < out_num_channels);
  61         src += in_channel * 3;
  62         dst += out_channel;
  63
  64         for (size_t i = 0; i < num_samples; ++i) {
  65                 uint32_t s1 = src[0];
  66                 uint32_t s2 = src[1];
  67                 uint32_t s3 = src[2];
  68                 uint32_t s = s1 | (s1 << 8) | (s2 << 16) | (s3 << 24);
  69                 *dst = int(s) * (1.0f / 2147483648.0f);
  70
  71                 src += 3 * in_num_channels;
  72                 dst += out_num_channels;
  73         }
  74 }
  75
  76 void convert_fixed32_to_fp32(float *dst, size_t out_channel, size_t out_num_channels,
  77                              const uint8_t *src, size_t in_channel, size_t in_num_channels,
  78                              size_t num_samples)
  79 {
  80         assert(in_channel < in_num_channels);
  81         assert(out_channel < out_num_channels);
  82         src += in_channel * 4;
  83         dst += out_channel;
  84
  85         for (size_t i = 0; i < num_samples; ++i) {
  86                 int32_t s = le32toh(*(int32_t *)src);
  87                 *dst = s * (1.0f / 2147483648.0f);
  88
  89                 src += 4 * in_num_channels;
  90                 dst += out_num_channels;
  91         }
  92 }
  93
  94 float find_peak_plain(const float *samples, size_t num_samples) __attribute__((unused));
  95
  96 float find_peak_plain(const float *samples, size_t num_samples)
  97 {
  98         float m = fabs(samples[0]);
  99         for (size_t i = 1; i < num_samples; ++i) {
 100                 m = max(m, fabs(samples[i]));
 101         }
 102         return m;
 103 }
 104
 105 #ifdef __SSE__
 106 static inline float horizontal_max(__m128 m)
 107 {
 108         __m128 tmp = _mm_shuffle_ps(m, m, _MM_SHUFFLE(1, 0, 3, 2));
 109         m = _mm_max_ps(m, tmp);
 110         tmp = _mm_shuffle_ps(m, m, _MM_SHUFFLE(2, 3, 0, 1));
 111         m = _mm_max_ps(m, tmp);
 112         return _mm_cvtss_f32(m);
 113 }
 114
 115 float find_peak(const float *samples, size_t num_samples)
 116 {
 117         const __m128 abs_mask = _mm_castsi128_ps(_mm_set1_epi32(0x7fffffffu));
 118         __m128 m = _mm_setzero_ps();
 119         for (size_t i = 0; i < (num_samples & ~3); i += 4) {
 120                 __m128 x = _mm_loadu_ps(samples + i);
 121                 x = _mm_and_ps(x, abs_mask);
 122                 m = _mm_max_ps(m, x);
 123         }
 124         float result = horizontal_max(m);
 125
 126         for (size_t i = (num_samples & ~3); i < num_samples; ++i) {
 127                 result = max(result, fabs(samples[i]));
 128         }
 129
 130 #if 0
 131         // Self-test. We should be bit-exact the same.
 132         float reference_result = find_peak_plain(samples, num_samples);
 133         if (result != reference_result) {
 134                 fprintf(stderr, "Error: Peak is %f [%f %f %f %f]; should be %f.\n",
 135                         result,
 136                         _mm_cvtss_f32(_mm_shuffle_ps(m, m, _MM_SHUFFLE(0, 0, 0, 0))),
 137                         _mm_cvtss_f32(_mm_shuffle_ps(m, m, _MM_SHUFFLE(1, 1, 1, 1))),
 138                         _mm_cvtss_f32(_mm_shuffle_ps(m, m, _MM_SHUFFLE(2, 2, 2, 2))),
 139                         _mm_cvtss_f32(_mm_shuffle_ps(m, m, _MM_SHUFFLE(3, 3, 3, 3))),
 140                         reference_result);
 141                 abort();
 142         }
 143 #endif
 144         return result;
 145 }
 146 #else
 147 float find_peak(const float *samples, size_t num_samples)
 148 {
 149         return find_peak_plain(samples, num_samples);
 150 }
 151 #endif
 152
 153 void deinterleave_samples(const vector<float> &in, vector<float> *out_l, vector<float> *out_r)
 154 {
 155         size_t num_samples = in.size() / 2;
 156         out_l->resize(num_samples);
 157         out_r->resize(num_samples);
 158
 159         const float *inptr = in.data();
 160         float *lptr = &(*out_l)[0];
 161         float *rptr = &(*out_r)[0];
 162         for (size_t i = 0; i < num_samples; ++i) {
 163                 *lptr++ = *inptr++;
 164                 *rptr++ = *inptr++;
 165         }
 166 }
 167
 168 }  // namespace
 169
 170 AudioMixer::AudioMixer(unsigned num_capture_cards, unsigned num_ffmpeg_inputs)
 171         : num_capture_cards(num_capture_cards),
 172           num_ffmpeg_inputs(num_ffmpeg_inputs),
 173           ffmpeg_inputs(new AudioDevice[num_ffmpeg_inputs]),
 174           limiter(OUTPUT_FREQUENCY),
 175           correlation(OUTPUT_FREQUENCY)
 176 {
 177         for (unsigned bus_index = 0; bus_index < MAX_BUSES; ++bus_index) {
 178                 locut[bus_index].init(FILTER_HPF, 2);
 179                 eq[bus_index][EQ_BAND_BASS].init(FILTER_LOW_SHELF, 1);
 180                 // Note: EQ_BAND_MID isn't used (see comments in apply_eq()).
 181                 eq[bus_index][EQ_BAND_TREBLE].init(FILTER_HIGH_SHELF, 1);
 182                 compressor[bus_index].reset(new StereoCompressor(OUTPUT_FREQUENCY));
 183                 level_compressor[bus_index].reset(new StereoCompressor(OUTPUT_FREQUENCY));
 184
 185                 set_bus_settings(bus_index, get_default_bus_settings());
 186         }
 187         set_limiter_enabled(global_flags.limiter_enabled);
 188         set_final_makeup_gain_auto(global_flags.final_makeup_gain_auto);
 189
 190         r128.init(2, OUTPUT_FREQUENCY);
 191         r128.integr_start();
 192
 193         // hlen=16 is pretty low quality, but we use quite a bit of CPU otherwise,
 194         // and there's a limit to how important the peak meter is.
 195         peak_resampler.setup(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY * 4, /*num_channels=*/2, /*hlen=*/16, /*frel=*/1.0);
 196
 197         global_audio_mixer = this;
 198         alsa_pool.init();
 199
 200         if (!global_flags.input_mapping_filename.empty()) {
 201                 // Must happen after ALSAPool is initialized, as it needs to know the card list.
 202                 current_mapping_mode = MappingMode::MULTICHANNEL;
 203                 InputMapping new_input_mapping;
 204                 if (!load_input_mapping_from_file(get_devices(),
 205                                                   global_flags.input_mapping_filename,
 206                                                   &new_input_mapping)) {
 207                         fprintf(stderr, "Failed to load input mapping from '%s', exiting.\n",
 208                                 global_flags.input_mapping_filename.c_str());
 209                         exit(1);
 210                 }
 211                 set_input_mapping(new_input_mapping);
 212         } else {
 213                 set_simple_input(/*card_index=*/0);
 214                 if (global_flags.multichannel_mapping_mode) {
 215                         current_mapping_mode = MappingMode::MULTICHANNEL;
 216                 }
 217         }
 218
 219         global_metrics.add("audio_loudness_short_lufs", &metric_audio_loudness_short_lufs, Metrics::TYPE_GAUGE);
 220         global_metrics.add("audio_loudness_integrated_lufs", &metric_audio_loudness_integrated_lufs, Metrics::TYPE_GAUGE);
 221         global_metrics.add("audio_loudness_range_low_lufs", &metric_audio_loudness_range_low_lufs, Metrics::TYPE_GAUGE);
 222         global_metrics.add("audio_loudness_range_high_lufs", &metric_audio_loudness_range_high_lufs, Metrics::TYPE_GAUGE);
 223         global_metrics.add("audio_peak_dbfs", &metric_audio_peak_dbfs, Metrics::TYPE_GAUGE);
 224         global_metrics.add("audio_final_makeup_gain_db", &metric_audio_final_makeup_gain_db, Metrics::TYPE_GAUGE);
 225         global_metrics.add("audio_correlation", &metric_audio_correlation, Metrics::TYPE_GAUGE);
 226 }
 227
 228 void AudioMixer::reset_resampler(DeviceSpec device_spec)
 229 {
 230         lock_guard<timed_mutex> lock(audio_mutex);
 231         reset_resampler_mutex_held(device_spec);
 232 }
 233
 234 void AudioMixer::reset_resampler_mutex_held(DeviceSpec device_spec)
 235 {
 236         AudioDevice *device = find_audio_device(device_spec);
 237
 238         if (device->interesting_channels.empty()) {
 239                 device->resampling_queue.reset();
 240         } else {
 241                 device->resampling_queue.reset(new ResamplingQueue(
 242                         device_spec, device->capture_frequency, OUTPUT_FREQUENCY, device->interesting_channels.size(),
 243                         global_flags.audio_queue_length_ms * 0.001));
 244         }
 245 }
 246
 247 bool AudioMixer::add_audio(DeviceSpec device_spec, const uint8_t *data, unsigned num_samples, AudioFormat audio_format, int64_t frame_length, steady_clock::time_point frame_time)
 248 {
 249         AudioDevice *device = find_audio_device(device_spec);
 250
 251         unique_lock<timed_mutex> lock(audio_mutex, defer_lock);
 252         if (!lock.try_lock_for(chrono::milliseconds(10))) {
 253                 return false;
 254         }
 255         if (device->resampling_queue == nullptr) {
 256                 // No buses use this device; throw it away.
 257                 return true;
 258         }
 259
 260         unsigned num_channels = device->interesting_channels.size();
 261         assert(num_channels > 0);
 262
 263         // Convert the audio to fp32.
 264         unique_ptr<float[]> audio(new float[num_samples * num_channels]);
 265         unsigned channel_index = 0;
 266         for (auto channel_it = device->interesting_channels.cbegin(); channel_it != device->interesting_channels.end(); ++channel_it, ++channel_index) {
 267                 switch (audio_format.bits_per_sample) {
 268                 case 0:
 269                         assert(num_samples == 0);
 270                         break;
 271                 case 16:
 272                         convert_fixed16_to_fp32(audio.get(), channel_index, num_channels, data, *channel_it, audio_format.num_channels, num_samples);
 273                         break;
 274                 case 24:
 275                         convert_fixed24_to_fp32(audio.get(), channel_index, num_channels, data, *channel_it, audio_format.num_channels, num_samples);
 276                         break;
 277                 case 32:
 278                         convert_fixed32_to_fp32(audio.get(), channel_index, num_channels, data, *channel_it, audio_format.num_channels, num_samples);
 279                         break;
 280                 default:
 281                         fprintf(stderr, "Cannot handle audio with %u bits per sample\n", audio_format.bits_per_sample);
 282                         assert(false);
 283                 }
 284         }
 285
 286         // If we changed frequency since last frame, we'll need to reset the resampler.
 287         if (audio_format.sample_rate != device->capture_frequency) {
 288                 device->capture_frequency = audio_format.sample_rate;
 289                 reset_resampler_mutex_held(device_spec);
 290         }
 291
 292         // Now add it.
 293         device->resampling_queue->add_input_samples(frame_time, audio.get(), num_samples, ResamplingQueue::ADJUST_RATE);
 294         return true;
 295 }
 296
 297 bool AudioMixer::add_silence(DeviceSpec device_spec, unsigned samples_per_frame, unsigned num_frames, int64_t frame_length)
 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         if (device->resampling_queue == nullptr) {
 306                 // No buses use this device; throw it away.
 307                 return true;
 308         }
 309
 310         unsigned num_channels = device->interesting_channels.size();
 311         assert(num_channels > 0);
 312
 313         vector<float> silence(samples_per_frame * num_channels, 0.0f);
 314         for (unsigned i = 0; i < num_frames; ++i) {
 315                 device->resampling_queue->add_input_samples(steady_clock::now(), silence.data(), samples_per_frame, ResamplingQueue::DO_NOT_ADJUST_RATE);
 316         }
 317         return true;
 318 }
 319
 320 bool AudioMixer::silence_card(DeviceSpec device_spec, bool silence)
 321 {
 322         AudioDevice *device = find_audio_device(device_spec);
 323
 324         unique_lock<timed_mutex> lock(audio_mutex, defer_lock);
 325         if (!lock.try_lock_for(chrono::milliseconds(10))) {
 326                 return false;
 327         }
 328
 329         if (device->silenced && !silence) {
 330                 reset_resampler_mutex_held(device_spec);
 331         }
 332         device->silenced = silence;
 333         return true;
 334 }
 335
 336 AudioMixer::BusSettings AudioMixer::get_default_bus_settings()
 337 {
 338         BusSettings settings;
 339         settings.fader_volume_db = 0.0f;
 340         settings.muted = false;
 341         settings.locut_enabled = global_flags.locut_enabled;
 342         for (unsigned band_index = 0; band_index < NUM_EQ_BANDS; ++band_index) {
 343                 settings.eq_level_db[band_index] = 0.0f;
 344         }
 345         settings.gain_staging_db = global_flags.initial_gain_staging_db;
 346         settings.level_compressor_enabled = global_flags.gain_staging_auto;
 347         settings.compressor_threshold_dbfs = ref_level_dbfs - 12.0f;  // -12 dB.
 348         settings.compressor_enabled = global_flags.compressor_enabled;
 349         return settings;
 350 }
 351
 352 AudioMixer::BusSettings AudioMixer::get_bus_settings(unsigned bus_index) const
 353 {
 354         lock_guard<timed_mutex> lock(audio_mutex);
 355         BusSettings settings;
 356         settings.fader_volume_db = fader_volume_db[bus_index];
 357         settings.muted = mute[bus_index];
 358         settings.locut_enabled = locut_enabled[bus_index];
 359         for (unsigned band_index = 0; band_index < NUM_EQ_BANDS; ++band_index) {
 360                 settings.eq_level_db[band_index] = eq_level_db[bus_index][band_index];
 361         }
 362         settings.gain_staging_db = gain_staging_db[bus_index];
 363         settings.level_compressor_enabled = level_compressor_enabled[bus_index];
 364         settings.compressor_threshold_dbfs = compressor_threshold_dbfs[bus_index];
 365         settings.compressor_enabled = compressor_enabled[bus_index];
 366         return settings;
 367 }
 368
 369 void AudioMixer::set_bus_settings(unsigned bus_index, const AudioMixer::BusSettings &settings)
 370 {
 371         lock_guard<timed_mutex> lock(audio_mutex);
 372         fader_volume_db[bus_index] = settings.fader_volume_db;
 373         mute[bus_index] = settings.muted;
 374         locut_enabled[bus_index] = settings.locut_enabled;
 375         for (unsigned band_index = 0; band_index < NUM_EQ_BANDS; ++band_index) {
 376                 eq_level_db[bus_index][band_index] = settings.eq_level_db[band_index];
 377         }
 378         gain_staging_db[bus_index] = settings.gain_staging_db;
 379         last_gain_staging_db[bus_index] = gain_staging_db[bus_index];
 380         level_compressor_enabled[bus_index] = settings.level_compressor_enabled;
 381         compressor_threshold_dbfs[bus_index] = settings.compressor_threshold_dbfs;
 382         compressor_enabled[bus_index] = settings.compressor_enabled;
 383 }
 384
 385 AudioMixer::AudioDevice *AudioMixer::find_audio_device(DeviceSpec device)
 386 {
 387         switch (device.type) {
 388         case InputSourceType::CAPTURE_CARD:
 389                 return &video_cards[device.index];
 390         case InputSourceType::ALSA_INPUT:
 391                 return &alsa_inputs[device.index];
 392         case InputSourceType::FFMPEG_VIDEO_INPUT:
 393                 return &ffmpeg_inputs[device.index];
 394         case InputSourceType::SILENCE:
 395         default:
 396                 assert(false);
 397         }
 398         return nullptr;
 399 }
 400
 401 // Get a pointer to the given channel from the given device.
 402 // The channel must be picked out earlier and resampled.
 403 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)
 404 {
 405         static float zero = 0.0f;
 406         if (source_channel == -1 || device_spec.type == InputSourceType::SILENCE) {
 407                 *srcptr = &zero;
 408                 *stride = 0;
 409                 return;
 410         }
 411         AudioDevice *device = find_audio_device(device_spec);
 412         assert(device->interesting_channels.count(source_channel) != 0);
 413         unsigned channel_index = 0;
 414         for (int channel : device->interesting_channels) {
 415                 if (channel == source_channel) break;
 416                 ++channel_index;
 417         }
 418         assert(channel_index < device->interesting_channels.size());
 419         const auto it = samples_card.find(device_spec);
 420         assert(it != samples_card.end());
 421         *srcptr = &(it->second)[channel_index];
 422         *stride = device->interesting_channels.size();
 423 }
 424
 425 // TODO: Can be SSSE3-optimized if need be.
 426 void AudioMixer::fill_audio_bus(const map<DeviceSpec, vector<float>> &samples_card, const InputMapping::Bus &bus, unsigned num_samples, float *output)
 427 {
 428         if (bus.device.type == InputSourceType::SILENCE) {
 429                 memset(output, 0, num_samples * 2 * sizeof(*output));
 430         } else {
 431                 assert(bus.device.type == InputSourceType::CAPTURE_CARD ||
 432                        bus.device.type == InputSourceType::ALSA_INPUT ||
 433                        bus.device.type == InputSourceType::FFMPEG_VIDEO_INPUT);
 434                 const float *lsrc, *rsrc;
 435                 unsigned lstride, rstride;
 436                 float *dptr = output;
 437                 find_sample_src_from_device(samples_card, bus.device, bus.source_channel[0], &lsrc, &lstride);
 438                 find_sample_src_from_device(samples_card, bus.device, bus.source_channel[1], &rsrc, &rstride);
 439                 for (unsigned i = 0; i < num_samples; ++i) {
 440                         *dptr++ = *lsrc;
 441                         *dptr++ = *rsrc;
 442                         lsrc += lstride;
 443                         rsrc += rstride;
 444                 }
 445         }
 446 }
 447
 448 vector<DeviceSpec> AudioMixer::get_active_devices() const
 449 {
 450         vector<DeviceSpec> ret;
 451         for (unsigned card_index = 0; card_index < MAX_VIDEO_CARDS; ++card_index) {
 452                 const DeviceSpec device_spec{InputSourceType::CAPTURE_CARD, card_index};
 453                 if (!find_audio_device(device_spec)->interesting_channels.empty()) {
 454                         ret.push_back(device_spec);
 455                 }
 456         }
 457         for (unsigned card_index = 0; card_index < MAX_ALSA_CARDS; ++card_index) {
 458                 const DeviceSpec device_spec{InputSourceType::ALSA_INPUT, card_index};
 459                 if (!find_audio_device(device_spec)->interesting_channels.empty()) {
 460                         ret.push_back(device_spec);
 461                 }
 462         }
 463         for (unsigned card_index = 0; card_index < num_ffmpeg_inputs; ++card_index) {
 464                 const DeviceSpec device_spec{InputSourceType::FFMPEG_VIDEO_INPUT, card_index};
 465                 if (!find_audio_device(device_spec)->interesting_channels.empty()) {
 466                         ret.push_back(device_spec);
 467                 }
 468         }
 469         return ret;
 470 }
 471
 472 namespace {
 473
 474 void apply_gain(float db, float last_db, vector<float> *samples)
 475 {
 476         if (fabs(db - last_db) < 1e-3) {
 477                 // Constant over this frame.
 478                 const float gain = from_db(db);
 479                 for (size_t i = 0; i < samples->size(); ++i) {
 480                         (*samples)[i] *= gain;
 481                 }
 482         } else {
 483                 // We need to do a fade.
 484                 unsigned num_samples = samples->size() / 2;
 485                 float gain = from_db(last_db);
 486                 const float gain_inc = pow(from_db(db - last_db), 1.0 / num_samples);
 487                 for (size_t i = 0; i < num_samples; ++i) {
 488                         (*samples)[i * 2 + 0] *= gain;
 489                         (*samples)[i * 2 + 1] *= gain;
 490                         gain *= gain_inc;
 491                 }
 492         }
 493 }
 494
 495 }  // namespace
 496
 497 vector<float> AudioMixer::get_output(steady_clock::time_point ts, unsigned num_samples, ResamplingQueue::RateAdjustmentPolicy rate_adjustment_policy)
 498 {
 499         map<DeviceSpec, vector<float>> samples_card;
 500         vector<float> samples_bus;
 501
 502         lock_guard<timed_mutex> lock(audio_mutex);
 503
 504         // Pick out all the interesting channels from all the cards.
 505         for (const DeviceSpec &device_spec : get_active_devices()) {
 506                 AudioDevice *device = find_audio_device(device_spec);
 507                 samples_card[device_spec].resize(num_samples * device->interesting_channels.size());
 508                 if (device->silenced) {
 509                         memset(&samples_card[device_spec][0], 0, samples_card[device_spec].size() * sizeof(float));
 510                 } else {
 511                         device->resampling_queue->get_output_samples(
 512                                 ts,
 513                                 &samples_card[device_spec][0],
 514                                 num_samples,
 515                                 rate_adjustment_policy);
 516                 }
 517         }
 518
 519         vector<float> samples_out, left, right;
 520         samples_out.resize(num_samples * 2);
 521         samples_bus.resize(num_samples * 2);
 522         for (unsigned bus_index = 0; bus_index < input_mapping.buses.size(); ++bus_index) {
 523                 fill_audio_bus(samples_card, input_mapping.buses[bus_index], num_samples, &samples_bus[0]);
 524                 apply_eq(bus_index, &samples_bus);
 525
 526                 {
 527                         lock_guard<mutex> lock(compressor_mutex);
 528
 529                         // Apply a level compressor to get the general level right.
 530                         // Basically, if it's over about -40 dBFS, we squeeze it down to that level
 531                         // (or more precisely, near it, since we don't use infinite ratio),
 532                         // then apply a makeup gain to get it to -14 dBFS. -14 dBFS is, of course,
 533                         // entirely arbitrary, but from practical tests with speech, it seems to
 534                         // put ut around -23 LUFS, so it's a reasonable starting point for later use.
 535                         if (level_compressor_enabled[bus_index]) {
 536                                 float threshold = 0.01f;   // -40 dBFS.
 537                                 float ratio = 20.0f;
 538                                 float attack_time = 0.5f;
 539                                 float release_time = 20.0f;
 540                                 float makeup_gain = from_db(ref_level_dbfs - (-40.0f));  // +26 dB.
 541                                 level_compressor[bus_index]->process(samples_bus.data(), samples_bus.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 542                                 gain_staging_db[bus_index] = to_db(level_compressor[bus_index]->get_attenuation() * makeup_gain);
 543                         } else {
 544                                 // Just apply the gain we already had.
 545                                 float db = gain_staging_db[bus_index];
 546                                 float last_db = last_gain_staging_db[bus_index];
 547                                 apply_gain(db, last_db, &samples_bus);
 548                         }
 549                         last_gain_staging_db[bus_index] = gain_staging_db[bus_index];
 550
 551 #if 0
 552                         printf("level=%f (%+5.2f dBFS) attenuation=%f (%+5.2f dB) end_result=%+5.2f dB\n",
 553                                 level_compressor.get_level(), to_db(level_compressor.get_level()),
 554                                 level_compressor.get_attenuation(), to_db(level_compressor.get_attenuation()),
 555                                 to_db(level_compressor.get_level() * level_compressor.get_attenuation() * makeup_gain));
 556 #endif
 557
 558                         // The real compressor.
 559                         if (compressor_enabled[bus_index]) {
 560                                 float threshold = from_db(compressor_threshold_dbfs[bus_index]);
 561                                 float ratio = 20.0f;
 562                                 float attack_time = 0.005f;
 563                                 float release_time = 0.040f;
 564                                 float makeup_gain = 2.0f;  // +6 dB.
 565                                 compressor[bus_index]->process(samples_bus.data(), samples_bus.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 566                 //              compressor_att = compressor.get_attenuation();
 567                         }
 568                 }
 569
 570                 add_bus_to_master(bus_index, samples_bus, &samples_out);
 571                 deinterleave_samples(samples_bus, &left, &right);
 572                 measure_bus_levels(bus_index, left, right);
 573         }
 574
 575         {
 576                 lock_guard<mutex> lock(compressor_mutex);
 577
 578                 // Finally a limiter at -4 dB (so, -10 dBFS) to take out the worst peaks only.
 579                 // Note that since ratio is not infinite, we could go slightly higher than this.
 580                 if (limiter_enabled) {
 581                         float threshold = from_db(limiter_threshold_dbfs);
 582                         float ratio = 30.0f;
 583                         float attack_time = 0.0f;  // Instant.
 584                         float release_time = 0.020f;
 585                         float makeup_gain = 1.0f;  // 0 dB.
 586                         limiter.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 587         //              limiter_att = limiter.get_attenuation();
 588                 }
 589
 590         //      printf("limiter=%+5.1f  compressor=%+5.1f\n", to_db(limiter_att), to_db(compressor_att));
 591         }
 592
 593         // At this point, we are most likely close to +0 LU (at least if the
 594         // faders sum to 0 dB and the compressors are on), but all of our
 595         // measurements have been on raw sample values, not R128 values.
 596         // So we have a final makeup gain to get us to +0 LU; the gain
 597         // adjustments required should be relatively small, and also, the
 598         // offset shouldn't change much (only if the type of audio changes
 599         // significantly). Thus, we shoot for updating this value basically
 600         // “whenever we process buffers”, since the R128 calculation isn't exactly
 601         // something we get out per-sample.
 602         //
 603         // Note that there's a feedback loop here, so we choose a very slow filter
 604         // (half-time of 30 seconds).
 605         double target_loudness_factor, alpha;
 606         double loudness_lu = r128.loudness_M() - ref_level_lufs;
 607         target_loudness_factor = final_makeup_gain * from_db(-loudness_lu);
 608
 609         // If we're outside +/- 5 LU (after correction), we don't count it as
 610         // a normal signal (probably silence) and don't change the
 611         // correction factor; just apply what we already have.
 612         if (fabs(loudness_lu) >= 5.0 || !final_makeup_gain_auto) {
 613                 alpha = 0.0;
 614         } else {
 615                 // Formula adapted from
 616                 // https://en.wikipedia.org/wiki/Low-pass_filter#Simple_infinite_impulse_response_filter.
 617                 const double half_time_s = 30.0;
 618                 const double fc_mul_2pi_delta_t = 1.0 / (half_time_s * OUTPUT_FREQUENCY);
 619                 alpha = fc_mul_2pi_delta_t / (fc_mul_2pi_delta_t + 1.0);
 620         }
 621
 622         {
 623                 lock_guard<mutex> lock(compressor_mutex);
 624                 double m = final_makeup_gain;
 625                 for (size_t i = 0; i < samples_out.size(); i += 2) {
 626                         samples_out[i + 0] *= m;
 627                         samples_out[i + 1] *= m;
 628                         m += (target_loudness_factor - m) * alpha;
 629                 }
 630                 final_makeup_gain = m;
 631         }
 632
 633         update_meters(samples_out);
 634
 635         return samples_out;
 636 }
 637
 638 namespace {
 639
 640 void apply_filter_fade(StereoFilter *filter, float *data, unsigned num_samples, float cutoff_hz, float db, float last_db)
 641 {
 642         // A granularity of 32 samples is an okay tradeoff between speed and
 643         // smoothness; recalculating the filters is pretty expensive, so it's
 644         // good that we don't do this all the time.
 645         static constexpr unsigned filter_granularity_samples = 32;
 646
 647         const float cutoff_linear = cutoff_hz * 2.0 * M_PI / OUTPUT_FREQUENCY;
 648         if (fabs(db - last_db) < 1e-3) {
 649                 // Constant over this frame.
 650                 if (fabs(db) > 0.01f) {
 651                         filter->render(data, num_samples, cutoff_linear, 0.5f, db / 40.0f);
 652                 }
 653         } else {
 654                 // We need to do a fade. (Rounding up avoids division by zero.)
 655                 unsigned num_blocks = (num_samples + filter_granularity_samples - 1) / filter_granularity_samples;
 656                 const float inc_db_norm = (db - last_db) / 40.0f / num_blocks;
 657                 float db_norm = db / 40.0f;
 658                 for (size_t i = 0; i < num_samples; i += filter_granularity_samples) {
 659                         size_t samples_this_block = std::min<size_t>(num_samples - i, filter_granularity_samples);
 660                         filter->render(data + i * 2, samples_this_block, cutoff_linear, 0.5f, db_norm);
 661                         db_norm += inc_db_norm;
 662                 }
 663         }
 664 }
 665
 666 }  // namespace
 667
 668 void AudioMixer::apply_eq(unsigned bus_index, vector<float> *samples_bus)
 669 {
 670         constexpr float bass_freq_hz = 200.0f;
 671         constexpr float treble_freq_hz = 4700.0f;
 672
 673         // Cut away everything under 120 Hz (or whatever the cutoff is);
 674         // we don't need it for voice, and it will reduce headroom
 675         // and confuse the compressor. (In particular, any hums at 50 or 60 Hz
 676         // should be dampened.)
 677         if (locut_enabled[bus_index]) {
 678                 locut[bus_index].render(samples_bus->data(), samples_bus->size() / 2, locut_cutoff_hz * 2.0 * M_PI / OUTPUT_FREQUENCY, 0.5f);
 679         }
 680
 681         // Apply the rest of the EQ. Since we only have a simple three-band EQ,
 682         // we can implement it with two shelf filters. We use a simple gain to
 683         // set the mid-level filter, and then offset the low and high bands
 684         // from that if we need to. (We could perhaps have folded the gain into
 685         // the next part, but it's so cheap that the trouble isn't worth it.)
 686         //
 687         // If any part of the EQ has changed appreciably since last frame,
 688         // we fade smoothly during the course of this frame.
 689         const float bass_db = eq_level_db[bus_index][EQ_BAND_BASS];
 690         const float mid_db = eq_level_db[bus_index][EQ_BAND_MID];
 691         const float treble_db = eq_level_db[bus_index][EQ_BAND_TREBLE];
 692
 693         const float last_bass_db = last_eq_level_db[bus_index][EQ_BAND_BASS];
 694         const float last_mid_db = last_eq_level_db[bus_index][EQ_BAND_MID];
 695         const float last_treble_db = last_eq_level_db[bus_index][EQ_BAND_TREBLE];
 696
 697         assert(samples_bus->size() % 2 == 0);
 698         const unsigned num_samples = samples_bus->size() / 2;
 699
 700         apply_gain(mid_db, last_mid_db, samples_bus);
 701
 702         apply_filter_fade(&eq[bus_index][EQ_BAND_BASS], samples_bus->data(), num_samples, bass_freq_hz, bass_db - mid_db, last_bass_db - last_mid_db);
 703         apply_filter_fade(&eq[bus_index][EQ_BAND_TREBLE], samples_bus->data(), num_samples, treble_freq_hz, treble_db - mid_db, last_treble_db - last_mid_db);
 704
 705         last_eq_level_db[bus_index][EQ_BAND_BASS] = bass_db;
 706         last_eq_level_db[bus_index][EQ_BAND_MID] = mid_db;
 707         last_eq_level_db[bus_index][EQ_BAND_TREBLE] = treble_db;
 708 }
 709
 710 void AudioMixer::add_bus_to_master(unsigned bus_index, const vector<float> &samples_bus, vector<float> *samples_out)
 711 {
 712         assert(samples_bus.size() == samples_out->size());
 713         assert(samples_bus.size() % 2 == 0);
 714         unsigned num_samples = samples_bus.size() / 2;
 715         const float new_volume_db = mute[bus_index] ? -90.0f : fader_volume_db[bus_index].load();
 716         if (fabs(new_volume_db - last_fader_volume_db[bus_index]) > 1e-3) {
 717                 // The volume has changed; do a fade over the course of this frame.
 718                 // (We might have some numerical issues here, but it seems to sound OK.)
 719                 // For the purpose of fading here, the silence floor is set to -90 dB
 720                 // (the fader only goes to -84).
 721                 float old_volume = from_db(max<float>(last_fader_volume_db[bus_index], -90.0f));
 722                 float volume = from_db(max<float>(new_volume_db, -90.0f));
 723
 724                 float volume_inc = pow(volume / old_volume, 1.0 / num_samples);
 725                 volume = old_volume;
 726                 if (bus_index == 0) {
 727                         for (unsigned i = 0; i < num_samples; ++i) {
 728                                 (*samples_out)[i * 2 + 0] = samples_bus[i * 2 + 0] * volume;
 729                                 (*samples_out)[i * 2 + 1] = samples_bus[i * 2 + 1] * volume;
 730                                 volume *= volume_inc;
 731                         }
 732                 } else {
 733                         for (unsigned i = 0; i < num_samples; ++i) {
 734                                 (*samples_out)[i * 2 + 0] += samples_bus[i * 2 + 0] * volume;
 735                                 (*samples_out)[i * 2 + 1] += samples_bus[i * 2 + 1] * volume;
 736                                 volume *= volume_inc;
 737                         }
 738                 }
 739         } else if (new_volume_db > -90.0f) {
 740                 float volume = from_db(new_volume_db);
 741                 if (bus_index == 0) {
 742                         for (unsigned i = 0; i < num_samples; ++i) {
 743                                 (*samples_out)[i * 2 + 0] = samples_bus[i * 2 + 0] * volume;
 744                                 (*samples_out)[i * 2 + 1] = samples_bus[i * 2 + 1] * volume;
 745                         }
 746                 } else {
 747                         for (unsigned i = 0; i < num_samples; ++i) {
 748                                 (*samples_out)[i * 2 + 0] += samples_bus[i * 2 + 0] * volume;
 749                                 (*samples_out)[i * 2 + 1] += samples_bus[i * 2 + 1] * volume;
 750                         }
 751                 }
 752         }
 753
 754         last_fader_volume_db[bus_index] = new_volume_db;
 755 }
 756
 757 void AudioMixer::measure_bus_levels(unsigned bus_index, const vector<float> &left, const vector<float> &right)
 758 {
 759         assert(left.size() == right.size());
 760         const float volume = mute[bus_index] ? 0.0f : from_db(fader_volume_db[bus_index]);
 761         const float peak_levels[2] = {
 762                 find_peak(left.data(), left.size()) * volume,
 763                 find_peak(right.data(), right.size()) * volume
 764         };
 765         for (unsigned channel = 0; channel < 2; ++channel) {
 766                 // Compute the current value, including hold and falloff.
 767                 // The constants are borrowed from zita-mu1 by Fons Adriaensen.
 768                 static constexpr float hold_sec = 0.5f;
 769                 static constexpr float falloff_db_sec = 15.0f;  // dB/sec falloff after hold.
 770                 float current_peak;
 771                 PeakHistory &history = peak_history[bus_index][channel];
 772                 history.historic_peak = max(history.historic_peak, peak_levels[channel]);
 773                 if (history.age_seconds < hold_sec) {
 774                         current_peak = history.last_peak;
 775                 } else {
 776                         current_peak = history.last_peak * from_db(-falloff_db_sec * (history.age_seconds - hold_sec));
 777                 }
 778
 779                 // See if we have a new peak to replace the old (possibly falling) one.
 780                 if (peak_levels[channel] > current_peak) {
 781                         history.last_peak = peak_levels[channel];
 782                         history.age_seconds = 0.0f;  // Not 100% correct, but more than good enough given our frame sizes.
 783                         current_peak = peak_levels[channel];
 784                 } else {
 785                         history.age_seconds += float(left.size()) / OUTPUT_FREQUENCY;
 786                 }
 787                 history.current_level = peak_levels[channel];
 788                 history.current_peak = current_peak;
 789         }
 790 }
 791
 792 void AudioMixer::update_meters(const vector<float> &samples)
 793 {
 794         // Upsample 4x to find interpolated peak.
 795         peak_resampler.inp_data = const_cast<float *>(samples.data());
 796         peak_resampler.inp_count = samples.size() / 2;
 797
 798         vector<float> interpolated_samples;
 799         interpolated_samples.resize(samples.size());
 800         {
 801                 lock_guard<mutex> lock(audio_measure_mutex);
 802
 803                 while (peak_resampler.inp_count > 0) {  // About four iterations.
 804                         peak_resampler.out_data = &interpolated_samples[0];
 805                         peak_resampler.out_count = interpolated_samples.size() / 2;
 806                         peak_resampler.process();
 807                         size_t out_stereo_samples = interpolated_samples.size() / 2 - peak_resampler.out_count;
 808                         peak = max<float>(peak, find_peak(interpolated_samples.data(), out_stereo_samples * 2));
 809                         peak_resampler.out_data = nullptr;
 810                 }
 811         }
 812
 813         // Find R128 levels and L/R correlation.
 814         vector<float> left, right;
 815         deinterleave_samples(samples, &left, &right);
 816         float *ptrs[] = { left.data(), right.data() };
 817         {
 818                 lock_guard<mutex> lock(audio_measure_mutex);
 819                 r128.process(left.size(), ptrs);
 820                 correlation.process_samples(samples);
 821         }
 822
 823         send_audio_level_callback();
 824 }
 825
 826 void AudioMixer::reset_meters()
 827 {
 828         lock_guard<mutex> lock(audio_measure_mutex);
 829         peak_resampler.reset();
 830         peak = 0.0f;
 831         r128.reset();
 832         r128.integr_start();
 833         correlation.reset();
 834 }
 835
 836 void AudioMixer::send_audio_level_callback()
 837 {
 838         if (audio_level_callback == nullptr) {
 839                 return;
 840         }
 841
 842         lock_guard<mutex> lock(audio_measure_mutex);
 843         double loudness_s = r128.loudness_S();
 844         double loudness_i = r128.integrated();
 845         double loudness_range_low = r128.range_min();
 846         double loudness_range_high = r128.range_max();
 847
 848         metric_audio_loudness_short_lufs = loudness_s;
 849         metric_audio_loudness_integrated_lufs = loudness_i;
 850         metric_audio_loudness_range_low_lufs = loudness_range_low;
 851         metric_audio_loudness_range_high_lufs = loudness_range_high;
 852         metric_audio_peak_dbfs = to_db(peak);
 853         metric_audio_final_makeup_gain_db = to_db(final_makeup_gain);
 854         metric_audio_correlation = correlation.get_correlation();
 855
 856         vector<BusLevel> bus_levels;
 857         bus_levels.resize(input_mapping.buses.size());
 858         {
 859                 lock_guard<mutex> lock(compressor_mutex);
 860                 for (unsigned bus_index = 0; bus_index < bus_levels.size(); ++bus_index) {
 861                         BusLevel &levels = bus_levels[bus_index];
 862                         BusMetrics &metrics = bus_metrics[bus_index];
 863
 864                         levels.current_level_dbfs[0] = metrics.current_level_dbfs[0] = to_db(peak_history[bus_index][0].current_level);
 865                         levels.current_level_dbfs[1] = metrics.current_level_dbfs[1] = to_db(peak_history[bus_index][1].current_level);
 866                         levels.peak_level_dbfs[0] = metrics.peak_level_dbfs[0] = to_db(peak_history[bus_index][0].current_peak);
 867                         levels.peak_level_dbfs[1] = metrics.peak_level_dbfs[1] = to_db(peak_history[bus_index][1].current_peak);
 868                         levels.historic_peak_dbfs = metrics.historic_peak_dbfs = to_db(
 869                                 max(peak_history[bus_index][0].historic_peak,
 870                                     peak_history[bus_index][1].historic_peak));
 871                         levels.gain_staging_db = metrics.gain_staging_db = gain_staging_db[bus_index];
 872                         if (compressor_enabled[bus_index]) {
 873                                 levels.compressor_attenuation_db = metrics.compressor_attenuation_db = -to_db(compressor[bus_index]->get_attenuation());
 874                         } else {
 875                                 levels.compressor_attenuation_db = 0.0;
 876                                 metrics.compressor_attenuation_db = 0.0 / 0.0;
 877                         }
 878                 }
 879         }
 880
 881         audio_level_callback(loudness_s, to_db(peak), bus_levels,
 882                 loudness_i, loudness_range_low, loudness_range_high,
 883                 to_db(final_makeup_gain),
 884                 correlation.get_correlation());
 885 }
 886
 887 map<DeviceSpec, DeviceInfo> AudioMixer::get_devices()
 888 {
 889         lock_guard<timed_mutex> lock(audio_mutex);
 890
 891         map<DeviceSpec, DeviceInfo> devices;
 892         for (unsigned card_index = 0; card_index < num_capture_cards; ++card_index) {
 893                 const DeviceSpec spec{ InputSourceType::CAPTURE_CARD, card_index };
 894                 const AudioDevice *device = &video_cards[card_index];
 895                 DeviceInfo info;
 896                 info.display_name = device->display_name;
 897                 info.num_channels = 8;
 898                 devices.insert(make_pair(spec, info));
 899         }
 900         vector<ALSAPool::Device> available_alsa_devices = alsa_pool.get_devices();
 901         for (unsigned card_index = 0; card_index < available_alsa_devices.size(); ++card_index) {
 902                 const DeviceSpec spec{ InputSourceType::ALSA_INPUT, card_index };
 903                 const ALSAPool::Device &device = available_alsa_devices[card_index];
 904                 DeviceInfo info;
 905                 info.display_name = device.display_name();
 906                 info.num_channels = device.num_channels;
 907                 info.alsa_name = device.name;
 908                 info.alsa_info = device.info;
 909                 info.alsa_address = device.address;
 910                 devices.insert(make_pair(spec, info));
 911         }
 912         for (unsigned card_index = 0; card_index < num_ffmpeg_inputs; ++card_index) {
 913                 const DeviceSpec spec{ InputSourceType::FFMPEG_VIDEO_INPUT, card_index };
 914                 const AudioDevice *device = &ffmpeg_inputs[card_index];
 915                 DeviceInfo info;
 916                 info.display_name = device->display_name;
 917                 info.num_channels = 2;
 918                 devices.insert(make_pair(spec, info));
 919         }
 920         return devices;
 921 }
 922
 923 void AudioMixer::set_display_name(DeviceSpec device_spec, const string &name)
 924 {
 925         AudioDevice *device = find_audio_device(device_spec);
 926
 927         lock_guard<timed_mutex> lock(audio_mutex);
 928         device->display_name = name;
 929 }
 930
 931 void AudioMixer::serialize_device(DeviceSpec device_spec, DeviceSpecProto *device_spec_proto)
 932 {
 933         lock_guard<timed_mutex> lock(audio_mutex);
 934         switch (device_spec.type) {
 935                 case InputSourceType::SILENCE:
 936                         device_spec_proto->set_type(DeviceSpecProto::SILENCE);
 937                         break;
 938                 case InputSourceType::CAPTURE_CARD:
 939                         device_spec_proto->set_type(DeviceSpecProto::CAPTURE_CARD);
 940                         device_spec_proto->set_index(device_spec.index);
 941                         device_spec_proto->set_display_name(video_cards[device_spec.index].display_name);
 942                         break;
 943                 case InputSourceType::ALSA_INPUT:
 944                         alsa_pool.serialize_device(device_spec.index, device_spec_proto);
 945                         break;
 946                 case InputSourceType::FFMPEG_VIDEO_INPUT:
 947                         device_spec_proto->set_type(DeviceSpecProto::FFMPEG_VIDEO_INPUT);
 948                         device_spec_proto->set_index(device_spec.index);
 949                         device_spec_proto->set_display_name(ffmpeg_inputs[device_spec.index].display_name);
 950                         break;
 951         }
 952 }
 953
 954 void AudioMixer::set_simple_input(unsigned card_index)
 955 {
 956         assert(card_index < num_capture_cards + num_ffmpeg_inputs);
 957         InputMapping new_input_mapping;
 958         InputMapping::Bus input;
 959         input.name = "Main";
 960         if (card_index >= num_capture_cards) {
 961                 input.device = DeviceSpec{InputSourceType::FFMPEG_VIDEO_INPUT, card_index - num_capture_cards};
 962         } else {
 963                 input.device = DeviceSpec{InputSourceType::CAPTURE_CARD, card_index};
 964         }
 965         input.source_channel[0] = 0;
 966         input.source_channel[1] = 1;
 967
 968         new_input_mapping.buses.push_back(input);
 969
 970         lock_guard<timed_mutex> lock(audio_mutex);
 971         current_mapping_mode = MappingMode::SIMPLE;
 972         set_input_mapping_lock_held(new_input_mapping);
 973         fader_volume_db[0] = 0.0f;
 974 }
 975
 976 unsigned AudioMixer::get_simple_input() const
 977 {
 978         lock_guard<timed_mutex> lock(audio_mutex);
 979         if (input_mapping.buses.size() == 1 &&
 980             input_mapping.buses[0].device.type == InputSourceType::CAPTURE_CARD &&
 981             input_mapping.buses[0].source_channel[0] == 0 &&
 982             input_mapping.buses[0].source_channel[1] == 1) {
 983                 return input_mapping.buses[0].device.index;
 984         } else if (input_mapping.buses.size() == 1 &&
 985                    input_mapping.buses[0].device.type == InputSourceType::FFMPEG_VIDEO_INPUT &&
 986                    input_mapping.buses[0].source_channel[0] == 0 &&
 987                    input_mapping.buses[0].source_channel[1] == 1) {
 988                 return input_mapping.buses[0].device.index + num_capture_cards;
 989         } else {
 990                 return numeric_limits<unsigned>::max();
 991         }
 992 }
 993
 994 void AudioMixer::set_input_mapping(const InputMapping &new_input_mapping)
 995 {
 996         lock_guard<timed_mutex> lock(audio_mutex);
 997         set_input_mapping_lock_held(new_input_mapping);
 998         current_mapping_mode = MappingMode::MULTICHANNEL;
 999 }
1000
1001 AudioMixer::MappingMode AudioMixer::get_mapping_mode() const
1002 {
1003         lock_guard<timed_mutex> lock(audio_mutex);
1004         return current_mapping_mode;
1005 }
1006
1007 void AudioMixer::set_input_mapping_lock_held(const InputMapping &new_input_mapping)
1008 {
1009         map<DeviceSpec, set<unsigned>> interesting_channels;
1010         for (const InputMapping::Bus &bus : new_input_mapping.buses) {
1011                 if (bus.device.type == InputSourceType::CAPTURE_CARD ||
1012                     bus.device.type == InputSourceType::ALSA_INPUT ||
1013                     bus.device.type == InputSourceType::FFMPEG_VIDEO_INPUT) {
1014                         for (unsigned channel = 0; channel < 2; ++channel) {
1015                                 if (bus.source_channel[channel] != -1) {
1016                                         interesting_channels[bus.device].insert(bus.source_channel[channel]);
1017                                 }
1018                         }
1019                 } else {
1020                         assert(bus.device.type == InputSourceType::SILENCE);
1021                 }
1022         }
1023
1024         // Kill all the old metrics, and set up new ones.
1025         for (unsigned bus_index = 0; bus_index < input_mapping.buses.size(); ++bus_index) {
1026                 BusMetrics &metrics = bus_metrics[bus_index];
1027
1028                 vector<pair<string, string>> labels_left = metrics.labels;
1029                 labels_left.emplace_back("channel", "left");
1030                 vector<pair<string, string>> labels_right = metrics.labels;
1031                 labels_right.emplace_back("channel", "right");
1032
1033                 global_metrics.remove("bus_current_level_dbfs", labels_left);
1034                 global_metrics.remove("bus_current_level_dbfs", labels_right);
1035                 global_metrics.remove("bus_peak_level_dbfs", labels_left);
1036                 global_metrics.remove("bus_peak_level_dbfs", labels_right);
1037                 global_metrics.remove("bus_historic_peak_dbfs", metrics.labels);
1038                 global_metrics.remove("bus_gain_staging_db", metrics.labels);
1039                 global_metrics.remove("bus_compressor_attenuation_db", metrics.labels);
1040         }
1041         bus_metrics.reset(new BusMetrics[new_input_mapping.buses.size()]);
1042         for (unsigned bus_index = 0; bus_index < new_input_mapping.buses.size(); ++bus_index) {
1043                 const InputMapping::Bus &bus = new_input_mapping.buses[bus_index];
1044                 BusMetrics &metrics = bus_metrics[bus_index];
1045
1046                 char bus_index_str[16], source_index_str[16], source_channels_str[64];
1047                 snprintf(bus_index_str, sizeof(bus_index_str), "%u", bus_index);
1048                 snprintf(source_index_str, sizeof(source_index_str), "%u", bus.device.index);
1049                 snprintf(source_channels_str, sizeof(source_channels_str), "%d:%d", bus.source_channel[0], bus.source_channel[1]);
1050
1051                 vector<pair<string, string>> labels;
1052                 metrics.labels.emplace_back("index", bus_index_str);
1053                 metrics.labels.emplace_back("name", bus.name);
1054                 if (bus.device.type == InputSourceType::SILENCE) {
1055                         metrics.labels.emplace_back("source_type", "silence");
1056                 } else if (bus.device.type == InputSourceType::CAPTURE_CARD) {
1057                         metrics.labels.emplace_back("source_type", "capture_card");
1058                 } else if (bus.device.type == InputSourceType::ALSA_INPUT) {
1059                         metrics.labels.emplace_back("source_type", "alsa_input");
1060                 } else if (bus.device.type == InputSourceType::FFMPEG_VIDEO_INPUT) {
1061                         metrics.labels.emplace_back("source_type", "ffmpeg_video_input");
1062                 } else {
1063                         assert(false);
1064                 }
1065                 metrics.labels.emplace_back("source_index", source_index_str);
1066                 metrics.labels.emplace_back("source_channels", source_channels_str);
1067
1068                 vector<pair<string, string>> labels_left = metrics.labels;
1069                 labels_left.emplace_back("channel", "left");
1070                 vector<pair<string, string>> labels_right = metrics.labels;
1071                 labels_right.emplace_back("channel", "right");
1072
1073                 global_metrics.add("bus_current_level_dbfs", labels_left, &metrics.current_level_dbfs[0], Metrics::TYPE_GAUGE);
1074                 global_metrics.add("bus_current_level_dbfs", labels_right, &metrics.current_level_dbfs[1], Metrics::TYPE_GAUGE);
1075                 global_metrics.add("bus_peak_level_dbfs", labels_left, &metrics.peak_level_dbfs[0], Metrics::TYPE_GAUGE);
1076                 global_metrics.add("bus_peak_level_dbfs", labels_right, &metrics.peak_level_dbfs[1], Metrics::TYPE_GAUGE);
1077                 global_metrics.add("bus_historic_peak_dbfs", metrics.labels, &metrics.historic_peak_dbfs, Metrics::TYPE_GAUGE);
1078                 global_metrics.add("bus_gain_staging_db", metrics.labels, &metrics.gain_staging_db, Metrics::TYPE_GAUGE);
1079                 global_metrics.add("bus_compressor_attenuation_db", metrics.labels, &metrics.compressor_attenuation_db, Metrics::TYPE_GAUGE);
1080         }
1081
1082         // Reset resamplers for all cards that don't have the exact same state as before.
1083         for (unsigned card_index = 0; card_index < MAX_VIDEO_CARDS; ++card_index) {
1084                 const DeviceSpec device_spec{InputSourceType::CAPTURE_CARD, card_index};
1085                 AudioDevice *device = find_audio_device(device_spec);
1086                 if (device->interesting_channels != interesting_channels[device_spec]) {
1087                         device->interesting_channels = interesting_channels[device_spec];
1088                         reset_resampler_mutex_held(device_spec);
1089                 }
1090         }
1091         for (unsigned card_index = 0; card_index < MAX_ALSA_CARDS; ++card_index) {
1092                 const DeviceSpec device_spec{InputSourceType::ALSA_INPUT, card_index};
1093                 AudioDevice *device = find_audio_device(device_spec);
1094                 if (interesting_channels[device_spec].empty()) {
1095                         alsa_pool.release_device(card_index);
1096                 } else {
1097                         alsa_pool.hold_device(card_index);
1098                 }
1099                 if (device->interesting_channels != interesting_channels[device_spec]) {
1100                         device->interesting_channels = interesting_channels[device_spec];
1101                         alsa_pool.reset_device(device_spec.index);
1102                         reset_resampler_mutex_held(device_spec);
1103                 }
1104         }
1105         for (unsigned card_index = 0; card_index < num_ffmpeg_inputs; ++card_index) {
1106                 const DeviceSpec device_spec{InputSourceType::FFMPEG_VIDEO_INPUT, card_index};
1107                 AudioDevice *device = find_audio_device(device_spec);
1108                 if (device->interesting_channels != interesting_channels[device_spec]) {
1109                         device->interesting_channels = interesting_channels[device_spec];
1110                         reset_resampler_mutex_held(device_spec);
1111                 }
1112         }
1113
1114         input_mapping = new_input_mapping;
1115 }
1116
1117 InputMapping AudioMixer::get_input_mapping() const
1118 {
1119         lock_guard<timed_mutex> lock(audio_mutex);
1120         return input_mapping;
1121 }
1122
1123 unsigned AudioMixer::num_buses() const
1124 {
1125         lock_guard<timed_mutex> lock(audio_mutex);
1126         return input_mapping.buses.size();
1127 }
1128
1129 void AudioMixer::reset_peak(unsigned bus_index)
1130 {
1131         lock_guard<timed_mutex> lock(audio_mutex);
1132         for (unsigned channel = 0; channel < 2; ++channel) {
1133                 PeakHistory &history = peak_history[bus_index][channel];
1134                 history.current_level = 0.0f;
1135                 history.historic_peak = 0.0f;
1136                 history.current_peak = 0.0f;
1137                 history.last_peak = 0.0f;
1138                 history.age_seconds = 0.0f;
1139         }
1140 }
1141
1142 AudioMixer *global_audio_mixer = nullptr;