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         settings.stereo_width = 1.0f;
 343         for (unsigned band_index = 0; band_index < NUM_EQ_BANDS; ++band_index) {
 344                 settings.eq_level_db[band_index] = 0.0f;
 345         }
 346         settings.gain_staging_db = global_flags.initial_gain_staging_db;
 347         settings.level_compressor_enabled = global_flags.gain_staging_auto;
 348         settings.compressor_threshold_dbfs = ref_level_dbfs - 12.0f;  // -12 dB.
 349         settings.compressor_enabled = global_flags.compressor_enabled;
 350         return settings;
 351 }
 352
 353 AudioMixer::BusSettings AudioMixer::get_bus_settings(unsigned bus_index) const
 354 {
 355         lock_guard<timed_mutex> lock(audio_mutex);
 356         BusSettings settings;
 357         settings.fader_volume_db = fader_volume_db[bus_index];
 358         settings.muted = mute[bus_index];
 359         settings.locut_enabled = locut_enabled[bus_index];
 360         settings.stereo_width = stereo_width[bus_index];
 361         for (unsigned band_index = 0; band_index < NUM_EQ_BANDS; ++band_index) {
 362                 settings.eq_level_db[band_index] = eq_level_db[bus_index][band_index];
 363         }
 364         settings.gain_staging_db = gain_staging_db[bus_index];
 365         settings.level_compressor_enabled = level_compressor_enabled[bus_index];
 366         settings.compressor_threshold_dbfs = compressor_threshold_dbfs[bus_index];
 367         settings.compressor_enabled = compressor_enabled[bus_index];
 368         return settings;
 369 }
 370
 371 void AudioMixer::set_bus_settings(unsigned bus_index, const AudioMixer::BusSettings &settings)
 372 {
 373         lock_guard<timed_mutex> lock(audio_mutex);
 374         fader_volume_db[bus_index] = settings.fader_volume_db;
 375         mute[bus_index] = settings.muted;
 376         locut_enabled[bus_index] = settings.locut_enabled;
 377         stereo_width[bus_index] = settings.stereo_width;
 378         for (unsigned band_index = 0; band_index < NUM_EQ_BANDS; ++band_index) {
 379                 eq_level_db[bus_index][band_index] = settings.eq_level_db[band_index];
 380         }
 381         gain_staging_db[bus_index] = settings.gain_staging_db;
 382         last_gain_staging_db[bus_index] = gain_staging_db[bus_index];
 383         level_compressor_enabled[bus_index] = settings.level_compressor_enabled;
 384         compressor_threshold_dbfs[bus_index] = settings.compressor_threshold_dbfs;
 385         compressor_enabled[bus_index] = settings.compressor_enabled;
 386 }
 387
 388 AudioMixer::AudioDevice *AudioMixer::find_audio_device(DeviceSpec device)
 389 {
 390         switch (device.type) {
 391         case InputSourceType::CAPTURE_CARD:
 392                 return &video_cards[device.index];
 393         case InputSourceType::ALSA_INPUT:
 394                 return &alsa_inputs[device.index];
 395         case InputSourceType::FFMPEG_VIDEO_INPUT:
 396                 return &ffmpeg_inputs[device.index];
 397         case InputSourceType::SILENCE:
 398         default:
 399                 assert(false);
 400         }
 401         return nullptr;
 402 }
 403
 404 // Get a pointer to the given channel from the given device.
 405 // The channel must be picked out earlier and resampled.
 406 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)
 407 {
 408         static float zero = 0.0f;
 409         if (source_channel == -1 || device_spec.type == InputSourceType::SILENCE) {
 410                 *srcptr = &zero;
 411                 *stride = 0;
 412                 return;
 413         }
 414         AudioDevice *device = find_audio_device(device_spec);
 415         assert(device->interesting_channels.count(source_channel) != 0);
 416         unsigned channel_index = 0;
 417         for (int channel : device->interesting_channels) {
 418                 if (channel == source_channel) break;
 419                 ++channel_index;
 420         }
 421         assert(channel_index < device->interesting_channels.size());
 422         const auto it = samples_card.find(device_spec);
 423         assert(it != samples_card.end());
 424         *srcptr = &(it->second)[channel_index];
 425         *stride = device->interesting_channels.size();
 426 }
 427
 428 // TODO: Can be SSSE3-optimized if need be.
 429 void AudioMixer::fill_audio_bus(const map<DeviceSpec, vector<float>> &samples_card, const InputMapping::Bus &bus, unsigned num_samples, float stereo_width, float *output)
 430 {
 431         if (bus.device.type == InputSourceType::SILENCE) {
 432                 memset(output, 0, num_samples * 2 * sizeof(*output));
 433         } else {
 434                 assert(bus.device.type == InputSourceType::CAPTURE_CARD ||
 435                        bus.device.type == InputSourceType::ALSA_INPUT ||
 436                        bus.device.type == InputSourceType::FFMPEG_VIDEO_INPUT);
 437                 const float *lsrc, *rsrc;
 438                 unsigned lstride, rstride;
 439                 float *dptr = output;
 440                 find_sample_src_from_device(samples_card, bus.device, bus.source_channel[0], &lsrc, &lstride);
 441                 find_sample_src_from_device(samples_card, bus.device, bus.source_channel[1], &rsrc, &rstride);
 442
 443                 // Apply stereo width settings. Set stereo width w to a 0..1 range instead of
 444                 // -1..1, since it makes for much easier calculations (so 0.5 = completely mono).
 445                 // Then, what we want is
 446                 //
 447                 //   L' = wL + (1-w)R = R + w(L-R)
 448                 //   R' = wR + (1-w)L = L + w(R-L)
 449                 //
 450                 // This can be further simplified calculation-wise by defining the weighted
 451                 // difference signal D = w(R-L), so that:
 452                 //
 453                 //   L' = R - D
 454                 //   R' = L + D
 455                 float w = 0.5f * stereo_width + 0.5f;
 456                 if (bus.source_channel[0] == bus.source_channel[1]) {
 457                         // Mono anyway, so no need to bother.
 458                         w = 1.0f;
 459                 } else if (fabs(w) < 1e-3) {
 460                         // Perfect inverse.
 461                         swap(lsrc, rsrc);
 462                         swap(lstride, rstride);
 463                         w = 1.0f;
 464                 }
 465                 if (fabs(w - 1.0f) < 1e-3) {
 466                         // No calculations needed for stereo_width = 1.
 467                         for (unsigned i = 0; i < num_samples; ++i) {
 468                                 *dptr++ = *lsrc;
 469                                 *dptr++ = *rsrc;
 470                                 lsrc += lstride;
 471                                 rsrc += rstride;
 472                         }
 473                 } else {
 474                         // General case.
 475                         for (unsigned i = 0; i < num_samples; ++i) {
 476                                 float left = *lsrc, right = *rsrc;
 477                                 float diff = w * (right - left);
 478                                 *dptr++ = right - diff;
 479                                 *dptr++ = left + diff;
 480                                 lsrc += lstride;
 481                                 rsrc += rstride;
 482                         }
 483                 }
 484         }
 485 }
 486
 487 vector<DeviceSpec> AudioMixer::get_active_devices() const
 488 {
 489         vector<DeviceSpec> ret;
 490         for (unsigned card_index = 0; card_index < MAX_VIDEO_CARDS; ++card_index) {
 491                 const DeviceSpec device_spec{InputSourceType::CAPTURE_CARD, card_index};
 492                 if (!find_audio_device(device_spec)->interesting_channels.empty()) {
 493                         ret.push_back(device_spec);
 494                 }
 495         }
 496         for (unsigned card_index = 0; card_index < MAX_ALSA_CARDS; ++card_index) {
 497                 const DeviceSpec device_spec{InputSourceType::ALSA_INPUT, card_index};
 498                 if (!find_audio_device(device_spec)->interesting_channels.empty()) {
 499                         ret.push_back(device_spec);
 500                 }
 501         }
 502         for (unsigned card_index = 0; card_index < num_ffmpeg_inputs; ++card_index) {
 503                 const DeviceSpec device_spec{InputSourceType::FFMPEG_VIDEO_INPUT, card_index};
 504                 if (!find_audio_device(device_spec)->interesting_channels.empty()) {
 505                         ret.push_back(device_spec);
 506                 }
 507         }
 508         return ret;
 509 }
 510
 511 namespace {
 512
 513 void apply_gain(float db, float last_db, vector<float> *samples)
 514 {
 515         if (fabs(db - last_db) < 1e-3) {
 516                 // Constant over this frame.
 517                 const float gain = from_db(db);
 518                 for (size_t i = 0; i < samples->size(); ++i) {
 519                         (*samples)[i] *= gain;
 520                 }
 521         } else {
 522                 // We need to do a fade.
 523                 unsigned num_samples = samples->size() / 2;
 524                 float gain = from_db(last_db);
 525                 const float gain_inc = pow(from_db(db - last_db), 1.0 / num_samples);
 526                 for (size_t i = 0; i < num_samples; ++i) {
 527                         (*samples)[i * 2 + 0] *= gain;
 528                         (*samples)[i * 2 + 1] *= gain;
 529                         gain *= gain_inc;
 530                 }
 531         }
 532 }
 533
 534 }  // namespace
 535
 536 vector<float> AudioMixer::get_output(steady_clock::time_point ts, unsigned num_samples, ResamplingQueue::RateAdjustmentPolicy rate_adjustment_policy)
 537 {
 538         map<DeviceSpec, vector<float>> samples_card;
 539         vector<float> samples_bus;
 540
 541         lock_guard<timed_mutex> lock(audio_mutex);
 542
 543         // Pick out all the interesting channels from all the cards.
 544         for (const DeviceSpec &device_spec : get_active_devices()) {
 545                 AudioDevice *device = find_audio_device(device_spec);
 546                 samples_card[device_spec].resize(num_samples * device->interesting_channels.size());
 547                 if (device->silenced) {
 548                         memset(&samples_card[device_spec][0], 0, samples_card[device_spec].size() * sizeof(float));
 549                 } else {
 550                         device->resampling_queue->get_output_samples(
 551                                 ts,
 552                                 &samples_card[device_spec][0],
 553                                 num_samples,
 554                                 rate_adjustment_policy);
 555                 }
 556         }
 557
 558         vector<float> samples_out, left, right;
 559         samples_out.resize(num_samples * 2);
 560         samples_bus.resize(num_samples * 2);
 561         for (unsigned bus_index = 0; bus_index < input_mapping.buses.size(); ++bus_index) {
 562                 fill_audio_bus(samples_card, input_mapping.buses[bus_index], num_samples, stereo_width[bus_index], &samples_bus[0]);
 563                 apply_eq(bus_index, &samples_bus);
 564
 565                 {
 566                         lock_guard<mutex> lock(compressor_mutex);
 567
 568                         // Apply a level compressor to get the general level right.
 569                         // Basically, if it's over about -40 dBFS, we squeeze it down to that level
 570                         // (or more precisely, near it, since we don't use infinite ratio),
 571                         // then apply a makeup gain to get it to -14 dBFS. -14 dBFS is, of course,
 572                         // entirely arbitrary, but from practical tests with speech, it seems to
 573                         // put ut around -23 LUFS, so it's a reasonable starting point for later use.
 574                         if (level_compressor_enabled[bus_index]) {
 575                                 float threshold = 0.01f;   // -40 dBFS.
 576                                 float ratio = 20.0f;
 577                                 float attack_time = 0.5f;
 578                                 float release_time = 20.0f;
 579                                 float makeup_gain = from_db(ref_level_dbfs - (-40.0f));  // +26 dB.
 580                                 level_compressor[bus_index]->process(samples_bus.data(), samples_bus.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 581                                 gain_staging_db[bus_index] = to_db(level_compressor[bus_index]->get_attenuation() * makeup_gain);
 582                         } else {
 583                                 // Just apply the gain we already had.
 584                                 float db = gain_staging_db[bus_index];
 585                                 float last_db = last_gain_staging_db[bus_index];
 586                                 apply_gain(db, last_db, &samples_bus);
 587                         }
 588                         last_gain_staging_db[bus_index] = gain_staging_db[bus_index];
 589
 590 #if 0
 591                         printf("level=%f (%+5.2f dBFS) attenuation=%f (%+5.2f dB) end_result=%+5.2f dB\n",
 592                                 level_compressor.get_level(), to_db(level_compressor.get_level()),
 593                                 level_compressor.get_attenuation(), to_db(level_compressor.get_attenuation()),
 594                                 to_db(level_compressor.get_level() * level_compressor.get_attenuation() * makeup_gain));
 595 #endif
 596
 597                         // The real compressor.
 598                         if (compressor_enabled[bus_index]) {
 599                                 float threshold = from_db(compressor_threshold_dbfs[bus_index]);
 600                                 float ratio = 20.0f;
 601                                 float attack_time = 0.005f;
 602                                 float release_time = 0.040f;
 603                                 float makeup_gain = 2.0f;  // +6 dB.
 604                                 compressor[bus_index]->process(samples_bus.data(), samples_bus.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 605                 //              compressor_att = compressor.get_attenuation();
 606                         }
 607                 }
 608
 609                 add_bus_to_master(bus_index, samples_bus, &samples_out);
 610                 deinterleave_samples(samples_bus, &left, &right);
 611                 measure_bus_levels(bus_index, left, right);
 612         }
 613
 614         {
 615                 lock_guard<mutex> lock(compressor_mutex);
 616
 617                 // Finally a limiter at -4 dB (so, -10 dBFS) to take out the worst peaks only.
 618                 // Note that since ratio is not infinite, we could go slightly higher than this.
 619                 if (limiter_enabled) {
 620                         float threshold = from_db(limiter_threshold_dbfs);
 621                         float ratio = 30.0f;
 622                         float attack_time = 0.0f;  // Instant.
 623                         float release_time = 0.020f;
 624                         float makeup_gain = 1.0f;  // 0 dB.
 625                         limiter.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 626         //              limiter_att = limiter.get_attenuation();
 627                 }
 628
 629         //      printf("limiter=%+5.1f  compressor=%+5.1f\n", to_db(limiter_att), to_db(compressor_att));
 630         }
 631
 632         // At this point, we are most likely close to +0 LU (at least if the
 633         // faders sum to 0 dB and the compressors are on), but all of our
 634         // measurements have been on raw sample values, not R128 values.
 635         // So we have a final makeup gain to get us to +0 LU; the gain
 636         // adjustments required should be relatively small, and also, the
 637         // offset shouldn't change much (only if the type of audio changes
 638         // significantly). Thus, we shoot for updating this value basically
 639         // “whenever we process buffers”, since the R128 calculation isn't exactly
 640         // something we get out per-sample.
 641         //
 642         // Note that there's a feedback loop here, so we choose a very slow filter
 643         // (half-time of 30 seconds).
 644         double target_loudness_factor, alpha;
 645         double loudness_lu = r128.loudness_M() - ref_level_lufs;
 646         target_loudness_factor = final_makeup_gain * from_db(-loudness_lu);
 647
 648         // If we're outside +/- 5 LU (after correction), we don't count it as
 649         // a normal signal (probably silence) and don't change the
 650         // correction factor; just apply what we already have.
 651         if (fabs(loudness_lu) >= 5.0 || !final_makeup_gain_auto) {
 652                 alpha = 0.0;
 653         } else {
 654                 // Formula adapted from
 655                 // https://en.wikipedia.org/wiki/Low-pass_filter#Simple_infinite_impulse_response_filter.
 656                 const double half_time_s = 30.0;
 657                 const double fc_mul_2pi_delta_t = 1.0 / (half_time_s * OUTPUT_FREQUENCY);
 658                 alpha = fc_mul_2pi_delta_t / (fc_mul_2pi_delta_t + 1.0);
 659         }
 660
 661         {
 662                 lock_guard<mutex> lock(compressor_mutex);
 663                 double m = final_makeup_gain;
 664                 for (size_t i = 0; i < samples_out.size(); i += 2) {
 665                         samples_out[i + 0] *= m;
 666                         samples_out[i + 1] *= m;
 667                         m += (target_loudness_factor - m) * alpha;
 668                 }
 669                 final_makeup_gain = m;
 670         }
 671
 672         update_meters(samples_out);
 673
 674         return samples_out;
 675 }
 676
 677 namespace {
 678
 679 void apply_filter_fade(StereoFilter *filter, float *data, unsigned num_samples, float cutoff_hz, float db, float last_db)
 680 {
 681         // A granularity of 32 samples is an okay tradeoff between speed and
 682         // smoothness; recalculating the filters is pretty expensive, so it's
 683         // good that we don't do this all the time.
 684         static constexpr unsigned filter_granularity_samples = 32;
 685
 686         const float cutoff_linear = cutoff_hz * 2.0 * M_PI / OUTPUT_FREQUENCY;
 687         if (fabs(db - last_db) < 1e-3) {
 688                 // Constant over this frame.
 689                 if (fabs(db) > 0.01f) {
 690                         filter->render(data, num_samples, cutoff_linear, 0.5f, db / 40.0f);
 691                 }
 692         } else {
 693                 // We need to do a fade. (Rounding up avoids division by zero.)
 694                 unsigned num_blocks = (num_samples + filter_granularity_samples - 1) / filter_granularity_samples;
 695                 const float inc_db_norm = (db - last_db) / 40.0f / num_blocks;
 696                 float db_norm = db / 40.0f;
 697                 for (size_t i = 0; i < num_samples; i += filter_granularity_samples) {
 698                         size_t samples_this_block = std::min<size_t>(num_samples - i, filter_granularity_samples);
 699                         filter->render(data + i * 2, samples_this_block, cutoff_linear, 0.5f, db_norm);
 700                         db_norm += inc_db_norm;
 701                 }
 702         }
 703 }
 704
 705 }  // namespace
 706
 707 void AudioMixer::apply_eq(unsigned bus_index, vector<float> *samples_bus)
 708 {
 709         constexpr float bass_freq_hz = 200.0f;
 710         constexpr float treble_freq_hz = 4700.0f;
 711
 712         // Cut away everything under 120 Hz (or whatever the cutoff is);
 713         // we don't need it for voice, and it will reduce headroom
 714         // and confuse the compressor. (In particular, any hums at 50 or 60 Hz
 715         // should be dampened.)
 716         if (locut_enabled[bus_index]) {
 717                 locut[bus_index].render(samples_bus->data(), samples_bus->size() / 2, locut_cutoff_hz * 2.0 * M_PI / OUTPUT_FREQUENCY, 0.5f);
 718         }
 719
 720         // Apply the rest of the EQ. Since we only have a simple three-band EQ,
 721         // we can implement it with two shelf filters. We use a simple gain to
 722         // set the mid-level filter, and then offset the low and high bands
 723         // from that if we need to. (We could perhaps have folded the gain into
 724         // the next part, but it's so cheap that the trouble isn't worth it.)
 725         //
 726         // If any part of the EQ has changed appreciably since last frame,
 727         // we fade smoothly during the course of this frame.
 728         const float bass_db = eq_level_db[bus_index][EQ_BAND_BASS];
 729         const float mid_db = eq_level_db[bus_index][EQ_BAND_MID];
 730         const float treble_db = eq_level_db[bus_index][EQ_BAND_TREBLE];
 731
 732         const float last_bass_db = last_eq_level_db[bus_index][EQ_BAND_BASS];
 733         const float last_mid_db = last_eq_level_db[bus_index][EQ_BAND_MID];
 734         const float last_treble_db = last_eq_level_db[bus_index][EQ_BAND_TREBLE];
 735
 736         assert(samples_bus->size() % 2 == 0);
 737         const unsigned num_samples = samples_bus->size() / 2;
 738
 739         apply_gain(mid_db, last_mid_db, samples_bus);
 740
 741         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);
 742         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);
 743
 744         last_eq_level_db[bus_index][EQ_BAND_BASS] = bass_db;
 745         last_eq_level_db[bus_index][EQ_BAND_MID] = mid_db;
 746         last_eq_level_db[bus_index][EQ_BAND_TREBLE] = treble_db;
 747 }
 748
 749 void AudioMixer::add_bus_to_master(unsigned bus_index, const vector<float> &samples_bus, vector<float> *samples_out)
 750 {
 751         assert(samples_bus.size() == samples_out->size());
 752         assert(samples_bus.size() % 2 == 0);
 753         unsigned num_samples = samples_bus.size() / 2;
 754         const float new_volume_db = mute[bus_index] ? -90.0f : fader_volume_db[bus_index].load();
 755         if (fabs(new_volume_db - last_fader_volume_db[bus_index]) > 1e-3) {
 756                 // The volume has changed; do a fade over the course of this frame.
 757                 // (We might have some numerical issues here, but it seems to sound OK.)
 758                 // For the purpose of fading here, the silence floor is set to -90 dB
 759                 // (the fader only goes to -84).
 760                 float old_volume = from_db(max<float>(last_fader_volume_db[bus_index], -90.0f));
 761                 float volume = from_db(max<float>(new_volume_db, -90.0f));
 762
 763                 float volume_inc = pow(volume / old_volume, 1.0 / num_samples);
 764                 volume = old_volume;
 765                 if (bus_index == 0) {
 766                         for (unsigned i = 0; i < num_samples; ++i) {
 767                                 (*samples_out)[i * 2 + 0] = samples_bus[i * 2 + 0] * volume;
 768                                 (*samples_out)[i * 2 + 1] = samples_bus[i * 2 + 1] * volume;
 769                                 volume *= volume_inc;
 770                         }
 771                 } else {
 772                         for (unsigned i = 0; i < num_samples; ++i) {
 773                                 (*samples_out)[i * 2 + 0] += samples_bus[i * 2 + 0] * volume;
 774                                 (*samples_out)[i * 2 + 1] += samples_bus[i * 2 + 1] * volume;
 775                                 volume *= volume_inc;
 776                         }
 777                 }
 778         } else if (new_volume_db > -90.0f) {
 779                 float volume = from_db(new_volume_db);
 780                 if (bus_index == 0) {
 781                         for (unsigned i = 0; i < num_samples; ++i) {
 782                                 (*samples_out)[i * 2 + 0] = samples_bus[i * 2 + 0] * volume;
 783                                 (*samples_out)[i * 2 + 1] = samples_bus[i * 2 + 1] * volume;
 784                         }
 785                 } else {
 786                         for (unsigned i = 0; i < num_samples; ++i) {
 787                                 (*samples_out)[i * 2 + 0] += samples_bus[i * 2 + 0] * volume;
 788                                 (*samples_out)[i * 2 + 1] += samples_bus[i * 2 + 1] * volume;
 789                         }
 790                 }
 791         }
 792
 793         last_fader_volume_db[bus_index] = new_volume_db;
 794 }
 795
 796 void AudioMixer::measure_bus_levels(unsigned bus_index, const vector<float> &left, const vector<float> &right)
 797 {
 798         assert(left.size() == right.size());
 799         const float volume = mute[bus_index] ? 0.0f : from_db(fader_volume_db[bus_index]);
 800         const float peak_levels[2] = {
 801                 find_peak(left.data(), left.size()) * volume,
 802                 find_peak(right.data(), right.size()) * volume
 803         };
 804         for (unsigned channel = 0; channel < 2; ++channel) {
 805                 // Compute the current value, including hold and falloff.
 806                 // The constants are borrowed from zita-mu1 by Fons Adriaensen.
 807                 static constexpr float hold_sec = 0.5f;
 808                 static constexpr float falloff_db_sec = 15.0f;  // dB/sec falloff after hold.
 809                 float current_peak;
 810                 PeakHistory &history = peak_history[bus_index][channel];
 811                 history.historic_peak = max(history.historic_peak, peak_levels[channel]);
 812                 if (history.age_seconds < hold_sec) {
 813                         current_peak = history.last_peak;
 814                 } else {
 815                         current_peak = history.last_peak * from_db(-falloff_db_sec * (history.age_seconds - hold_sec));
 816                 }
 817
 818                 // See if we have a new peak to replace the old (possibly falling) one.
 819                 if (peak_levels[channel] > current_peak) {
 820                         history.last_peak = peak_levels[channel];
 821                         history.age_seconds = 0.0f;  // Not 100% correct, but more than good enough given our frame sizes.
 822                         current_peak = peak_levels[channel];
 823                 } else {
 824                         history.age_seconds += float(left.size()) / OUTPUT_FREQUENCY;
 825                 }
 826                 history.current_level = peak_levels[channel];
 827                 history.current_peak = current_peak;
 828         }
 829 }
 830
 831 void AudioMixer::update_meters(const vector<float> &samples)
 832 {
 833         // Upsample 4x to find interpolated peak.
 834         peak_resampler.inp_data = const_cast<float *>(samples.data());
 835         peak_resampler.inp_count = samples.size() / 2;
 836
 837         vector<float> interpolated_samples;
 838         interpolated_samples.resize(samples.size());
 839         {
 840                 lock_guard<mutex> lock(audio_measure_mutex);
 841
 842                 while (peak_resampler.inp_count > 0) {  // About four iterations.
 843                         peak_resampler.out_data = &interpolated_samples[0];
 844                         peak_resampler.out_count = interpolated_samples.size() / 2;
 845                         peak_resampler.process();
 846                         size_t out_stereo_samples = interpolated_samples.size() / 2 - peak_resampler.out_count;
 847                         peak = max<float>(peak, find_peak(interpolated_samples.data(), out_stereo_samples * 2));
 848                         peak_resampler.out_data = nullptr;
 849                 }
 850         }
 851
 852         // Find R128 levels and L/R correlation.
 853         vector<float> left, right;
 854         deinterleave_samples(samples, &left, &right);
 855         float *ptrs[] = { left.data(), right.data() };
 856         {
 857                 lock_guard<mutex> lock(audio_measure_mutex);
 858                 r128.process(left.size(), ptrs);
 859                 correlation.process_samples(samples);
 860         }
 861
 862         send_audio_level_callback();
 863 }
 864
 865 void AudioMixer::reset_meters()
 866 {
 867         lock_guard<mutex> lock(audio_measure_mutex);
 868         peak_resampler.reset();
 869         peak = 0.0f;
 870         r128.reset();
 871         r128.integr_start();
 872         correlation.reset();
 873 }
 874
 875 void AudioMixer::send_audio_level_callback()
 876 {
 877         if (audio_level_callback == nullptr) {
 878                 return;
 879         }
 880
 881         lock_guard<mutex> lock(audio_measure_mutex);
 882         double loudness_s = r128.loudness_S();
 883         double loudness_i = r128.integrated();
 884         double loudness_range_low = r128.range_min();
 885         double loudness_range_high = r128.range_max();
 886
 887         metric_audio_loudness_short_lufs = loudness_s;
 888         metric_audio_loudness_integrated_lufs = loudness_i;
 889         metric_audio_loudness_range_low_lufs = loudness_range_low;
 890         metric_audio_loudness_range_high_lufs = loudness_range_high;
 891         metric_audio_peak_dbfs = to_db(peak);
 892         metric_audio_final_makeup_gain_db = to_db(final_makeup_gain);
 893         metric_audio_correlation = correlation.get_correlation();
 894
 895         vector<BusLevel> bus_levels;
 896         bus_levels.resize(input_mapping.buses.size());
 897         {
 898                 lock_guard<mutex> lock(compressor_mutex);
 899                 for (unsigned bus_index = 0; bus_index < bus_levels.size(); ++bus_index) {
 900                         BusLevel &levels = bus_levels[bus_index];
 901                         BusMetrics &metrics = bus_metrics[bus_index];
 902
 903                         levels.current_level_dbfs[0] = metrics.current_level_dbfs[0] = to_db(peak_history[bus_index][0].current_level);
 904                         levels.current_level_dbfs[1] = metrics.current_level_dbfs[1] = to_db(peak_history[bus_index][1].current_level);
 905                         levels.peak_level_dbfs[0] = metrics.peak_level_dbfs[0] = to_db(peak_history[bus_index][0].current_peak);
 906                         levels.peak_level_dbfs[1] = metrics.peak_level_dbfs[1] = to_db(peak_history[bus_index][1].current_peak);
 907                         levels.historic_peak_dbfs = metrics.historic_peak_dbfs = to_db(
 908                                 max(peak_history[bus_index][0].historic_peak,
 909                                     peak_history[bus_index][1].historic_peak));
 910                         levels.gain_staging_db = metrics.gain_staging_db = gain_staging_db[bus_index];
 911                         if (compressor_enabled[bus_index]) {
 912                                 levels.compressor_attenuation_db = metrics.compressor_attenuation_db = -to_db(compressor[bus_index]->get_attenuation());
 913                         } else {
 914                                 levels.compressor_attenuation_db = 0.0;
 915                                 metrics.compressor_attenuation_db = 0.0 / 0.0;
 916                         }
 917                 }
 918         }
 919
 920         audio_level_callback(loudness_s, to_db(peak), bus_levels,
 921                 loudness_i, loudness_range_low, loudness_range_high,
 922                 to_db(final_makeup_gain),
 923                 correlation.get_correlation());
 924 }
 925
 926 map<DeviceSpec, DeviceInfo> AudioMixer::get_devices()
 927 {
 928         lock_guard<timed_mutex> lock(audio_mutex);
 929
 930         map<DeviceSpec, DeviceInfo> devices;
 931         for (unsigned card_index = 0; card_index < num_capture_cards; ++card_index) {
 932                 const DeviceSpec spec{ InputSourceType::CAPTURE_CARD, card_index };
 933                 const AudioDevice *device = &video_cards[card_index];
 934                 DeviceInfo info;
 935                 info.display_name = device->display_name;
 936                 info.num_channels = 8;
 937                 devices.insert(make_pair(spec, info));
 938         }
 939         vector<ALSAPool::Device> available_alsa_devices = alsa_pool.get_devices();
 940         for (unsigned card_index = 0; card_index < available_alsa_devices.size(); ++card_index) {
 941                 const DeviceSpec spec{ InputSourceType::ALSA_INPUT, card_index };
 942                 const ALSAPool::Device &device = available_alsa_devices[card_index];
 943                 DeviceInfo info;
 944                 info.display_name = device.display_name();
 945                 info.num_channels = device.num_channels;
 946                 info.alsa_name = device.name;
 947                 info.alsa_info = device.info;
 948                 info.alsa_address = device.address;
 949                 devices.insert(make_pair(spec, info));
 950         }
 951         for (unsigned card_index = 0; card_index < num_ffmpeg_inputs; ++card_index) {
 952                 const DeviceSpec spec{ InputSourceType::FFMPEG_VIDEO_INPUT, card_index };
 953                 const AudioDevice *device = &ffmpeg_inputs[card_index];
 954                 DeviceInfo info;
 955                 info.display_name = device->display_name;
 956                 info.num_channels = 2;
 957                 devices.insert(make_pair(spec, info));
 958         }
 959         return devices;
 960 }
 961
 962 void AudioMixer::set_display_name(DeviceSpec device_spec, const string &name)
 963 {
 964         AudioDevice *device = find_audio_device(device_spec);
 965
 966         lock_guard<timed_mutex> lock(audio_mutex);
 967         device->display_name = name;
 968 }
 969
 970 void AudioMixer::serialize_device(DeviceSpec device_spec, DeviceSpecProto *device_spec_proto)
 971 {
 972         lock_guard<timed_mutex> lock(audio_mutex);
 973         switch (device_spec.type) {
 974                 case InputSourceType::SILENCE:
 975                         device_spec_proto->set_type(DeviceSpecProto::SILENCE);
 976                         break;
 977                 case InputSourceType::CAPTURE_CARD:
 978                         device_spec_proto->set_type(DeviceSpecProto::CAPTURE_CARD);
 979                         device_spec_proto->set_index(device_spec.index);
 980                         device_spec_proto->set_display_name(video_cards[device_spec.index].display_name);
 981                         break;
 982                 case InputSourceType::ALSA_INPUT:
 983                         alsa_pool.serialize_device(device_spec.index, device_spec_proto);
 984                         break;
 985                 case InputSourceType::FFMPEG_VIDEO_INPUT:
 986                         device_spec_proto->set_type(DeviceSpecProto::FFMPEG_VIDEO_INPUT);
 987                         device_spec_proto->set_index(device_spec.index);
 988                         device_spec_proto->set_display_name(ffmpeg_inputs[device_spec.index].display_name);
 989                         break;
 990         }
 991 }
 992
 993 void AudioMixer::set_simple_input(unsigned card_index)
 994 {
 995         assert(card_index < num_capture_cards + num_ffmpeg_inputs);
 996         InputMapping new_input_mapping;
 997         InputMapping::Bus input;
 998         input.name = "Main";
 999         if (card_index >= num_capture_cards) {
1000                 input.device = DeviceSpec{InputSourceType::FFMPEG_VIDEO_INPUT, card_index - num_capture_cards};
1001         } else {
1002                 input.device = DeviceSpec{InputSourceType::CAPTURE_CARD, card_index};
1003         }
1004         input.source_channel[0] = 0;
1005         input.source_channel[1] = 1;
1006
1007         new_input_mapping.buses.push_back(input);
1008
1009         lock_guard<timed_mutex> lock(audio_mutex);
1010         current_mapping_mode = MappingMode::SIMPLE;
1011         set_input_mapping_lock_held(new_input_mapping);
1012         fader_volume_db[0] = 0.0f;
1013 }
1014
1015 unsigned AudioMixer::get_simple_input() const
1016 {
1017         lock_guard<timed_mutex> lock(audio_mutex);
1018         if (input_mapping.buses.size() == 1 &&
1019             input_mapping.buses[0].device.type == InputSourceType::CAPTURE_CARD &&
1020             input_mapping.buses[0].source_channel[0] == 0 &&
1021             input_mapping.buses[0].source_channel[1] == 1) {
1022                 return input_mapping.buses[0].device.index;
1023         } else if (input_mapping.buses.size() == 1 &&
1024                    input_mapping.buses[0].device.type == InputSourceType::FFMPEG_VIDEO_INPUT &&
1025                    input_mapping.buses[0].source_channel[0] == 0 &&
1026                    input_mapping.buses[0].source_channel[1] == 1) {
1027                 return input_mapping.buses[0].device.index + num_capture_cards;
1028         } else {
1029                 return numeric_limits<unsigned>::max();
1030         }
1031 }
1032
1033 void AudioMixer::set_input_mapping(const InputMapping &new_input_mapping)
1034 {
1035         lock_guard<timed_mutex> lock(audio_mutex);
1036         set_input_mapping_lock_held(new_input_mapping);
1037         current_mapping_mode = MappingMode::MULTICHANNEL;
1038 }
1039
1040 AudioMixer::MappingMode AudioMixer::get_mapping_mode() const
1041 {
1042         lock_guard<timed_mutex> lock(audio_mutex);
1043         return current_mapping_mode;
1044 }
1045
1046 void AudioMixer::set_input_mapping_lock_held(const InputMapping &new_input_mapping)
1047 {
1048         map<DeviceSpec, set<unsigned>> interesting_channels;
1049         for (const InputMapping::Bus &bus : new_input_mapping.buses) {
1050                 if (bus.device.type == InputSourceType::CAPTURE_CARD ||
1051                     bus.device.type == InputSourceType::ALSA_INPUT ||
1052                     bus.device.type == InputSourceType::FFMPEG_VIDEO_INPUT) {
1053                         for (unsigned channel = 0; channel < 2; ++channel) {
1054                                 if (bus.source_channel[channel] != -1) {
1055                                         interesting_channels[bus.device].insert(bus.source_channel[channel]);
1056                                 }
1057                         }
1058                 } else {
1059                         assert(bus.device.type == InputSourceType::SILENCE);
1060                 }
1061         }
1062
1063         // Kill all the old metrics, and set up new ones.
1064         for (unsigned bus_index = 0; bus_index < input_mapping.buses.size(); ++bus_index) {
1065                 BusMetrics &metrics = bus_metrics[bus_index];
1066
1067                 vector<pair<string, string>> labels_left = metrics.labels;
1068                 labels_left.emplace_back("channel", "left");
1069                 vector<pair<string, string>> labels_right = metrics.labels;
1070                 labels_right.emplace_back("channel", "right");
1071
1072                 global_metrics.remove("bus_current_level_dbfs", labels_left);
1073                 global_metrics.remove("bus_current_level_dbfs", labels_right);
1074                 global_metrics.remove("bus_peak_level_dbfs", labels_left);
1075                 global_metrics.remove("bus_peak_level_dbfs", labels_right);
1076                 global_metrics.remove("bus_historic_peak_dbfs", metrics.labels);
1077                 global_metrics.remove("bus_gain_staging_db", metrics.labels);
1078                 global_metrics.remove("bus_compressor_attenuation_db", metrics.labels);
1079         }
1080         bus_metrics.reset(new BusMetrics[new_input_mapping.buses.size()]);
1081         for (unsigned bus_index = 0; bus_index < new_input_mapping.buses.size(); ++bus_index) {
1082                 const InputMapping::Bus &bus = new_input_mapping.buses[bus_index];
1083                 BusMetrics &metrics = bus_metrics[bus_index];
1084
1085                 char bus_index_str[16], source_index_str[16], source_channels_str[64];
1086                 snprintf(bus_index_str, sizeof(bus_index_str), "%u", bus_index);
1087                 snprintf(source_index_str, sizeof(source_index_str), "%u", bus.device.index);
1088                 snprintf(source_channels_str, sizeof(source_channels_str), "%d:%d", bus.source_channel[0], bus.source_channel[1]);
1089
1090                 vector<pair<string, string>> labels;
1091                 metrics.labels.emplace_back("index", bus_index_str);
1092                 metrics.labels.emplace_back("name", bus.name);
1093                 if (bus.device.type == InputSourceType::SILENCE) {
1094                         metrics.labels.emplace_back("source_type", "silence");
1095                 } else if (bus.device.type == InputSourceType::CAPTURE_CARD) {
1096                         metrics.labels.emplace_back("source_type", "capture_card");
1097                 } else if (bus.device.type == InputSourceType::ALSA_INPUT) {
1098                         metrics.labels.emplace_back("source_type", "alsa_input");
1099                 } else if (bus.device.type == InputSourceType::FFMPEG_VIDEO_INPUT) {
1100                         metrics.labels.emplace_back("source_type", "ffmpeg_video_input");
1101                 } else {
1102                         assert(false);
1103                 }
1104                 metrics.labels.emplace_back("source_index", source_index_str);
1105                 metrics.labels.emplace_back("source_channels", source_channels_str);
1106
1107                 vector<pair<string, string>> labels_left = metrics.labels;
1108                 labels_left.emplace_back("channel", "left");
1109                 vector<pair<string, string>> labels_right = metrics.labels;
1110                 labels_right.emplace_back("channel", "right");
1111
1112                 global_metrics.add("bus_current_level_dbfs", labels_left, &metrics.current_level_dbfs[0], Metrics::TYPE_GAUGE);
1113                 global_metrics.add("bus_current_level_dbfs", labels_right, &metrics.current_level_dbfs[1], Metrics::TYPE_GAUGE);
1114                 global_metrics.add("bus_peak_level_dbfs", labels_left, &metrics.peak_level_dbfs[0], Metrics::TYPE_GAUGE);
1115                 global_metrics.add("bus_peak_level_dbfs", labels_right, &metrics.peak_level_dbfs[1], Metrics::TYPE_GAUGE);
1116                 global_metrics.add("bus_historic_peak_dbfs", metrics.labels, &metrics.historic_peak_dbfs, Metrics::TYPE_GAUGE);
1117                 global_metrics.add("bus_gain_staging_db", metrics.labels, &metrics.gain_staging_db, Metrics::TYPE_GAUGE);
1118                 global_metrics.add("bus_compressor_attenuation_db", metrics.labels, &metrics.compressor_attenuation_db, Metrics::TYPE_GAUGE);
1119         }
1120
1121         // Reset resamplers for all cards that don't have the exact same state as before.
1122         for (unsigned card_index = 0; card_index < MAX_VIDEO_CARDS; ++card_index) {
1123                 const DeviceSpec device_spec{InputSourceType::CAPTURE_CARD, card_index};
1124                 AudioDevice *device = find_audio_device(device_spec);
1125                 if (device->interesting_channels != interesting_channels[device_spec]) {
1126                         device->interesting_channels = interesting_channels[device_spec];
1127                         reset_resampler_mutex_held(device_spec);
1128                 }
1129         }
1130         for (unsigned card_index = 0; card_index < MAX_ALSA_CARDS; ++card_index) {
1131                 const DeviceSpec device_spec{InputSourceType::ALSA_INPUT, card_index};
1132                 AudioDevice *device = find_audio_device(device_spec);
1133                 if (interesting_channels[device_spec].empty()) {
1134                         alsa_pool.release_device(card_index);
1135                 } else {
1136                         alsa_pool.hold_device(card_index);
1137                 }
1138                 if (device->interesting_channels != interesting_channels[device_spec]) {
1139                         device->interesting_channels = interesting_channels[device_spec];
1140                         alsa_pool.reset_device(device_spec.index);
1141                         reset_resampler_mutex_held(device_spec);
1142                 }
1143         }
1144         for (unsigned card_index = 0; card_index < num_ffmpeg_inputs; ++card_index) {
1145                 const DeviceSpec device_spec{InputSourceType::FFMPEG_VIDEO_INPUT, card_index};
1146                 AudioDevice *device = find_audio_device(device_spec);
1147                 if (device->interesting_channels != interesting_channels[device_spec]) {
1148                         device->interesting_channels = interesting_channels[device_spec];
1149                         reset_resampler_mutex_held(device_spec);
1150                 }
1151         }
1152
1153         input_mapping = new_input_mapping;
1154 }
1155
1156 InputMapping AudioMixer::get_input_mapping() const
1157 {
1158         lock_guard<timed_mutex> lock(audio_mutex);
1159         return input_mapping;
1160 }
1161
1162 unsigned AudioMixer::num_buses() const
1163 {
1164         lock_guard<timed_mutex> lock(audio_mutex);
1165         return input_mapping.buses.size();
1166 }
1167
1168 void AudioMixer::reset_peak(unsigned bus_index)
1169 {
1170         lock_guard<timed_mutex> lock(audio_mutex);
1171         for (unsigned channel = 0; channel < 2; ++channel) {
1172                 PeakHistory &history = peak_history[bus_index][channel];
1173                 history.current_level = 0.0f;
1174                 history.historic_peak = 0.0f;
1175                 history.current_peak = 0.0f;
1176                 history.last_peak = 0.0f;
1177                 history.age_seconds = 0.0f;
1178         }
1179 }
1180
1181 bool AudioMixer::is_mono(unsigned bus_index)
1182 {
1183         lock_guard<timed_mutex> lock(audio_mutex);
1184         const InputMapping::Bus &bus = input_mapping.buses[bus_index];
1185         if (bus.device.type == InputSourceType::SILENCE) {
1186                 return true;
1187         } else {
1188                 assert(bus.device.type == InputSourceType::CAPTURE_CARD ||
1189                        bus.device.type == InputSourceType::ALSA_INPUT ||
1190                        bus.device.type == InputSourceType::FFMPEG_VIDEO_INPUT);
1191                 return bus.source_channel[0] == bus.source_channel[1];
1192         }
1193 }
1194
1195 AudioMixer *global_audio_mixer = nullptr;