git.sesse.net Git - nageru/blob - nageru/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 "decibel.h"
  22 #include "flags.h"
  23 #include "shared/metrics.h"
  24 #include "state.pb.h"
  25 #include "shared/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_fixed16_to_fixed32(int32_t *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 * 2;
  62         dst += out_channel;
  63
  64         for (size_t i = 0; i < num_samples; ++i) {
  65                 uint32_t s = uint32_t(uint16_t(le16toh(*(int16_t *)src))) << 16;
  66
  67                 // Keep the sign bit in place, repeat the other 15 bits as far as they go.
  68                 *dst = s | ((s & 0x7fffffff) >> 15) | ((s & 0x7fffffff) >> 30);
  69
  70                 src += 2 * in_num_channels;
  71                 dst += out_num_channels;
  72         }
  73 }
  74
  75 void convert_fixed24_to_fp32(float *dst, size_t out_channel, size_t out_num_channels,
  76                              const uint8_t *src, size_t in_channel, size_t in_num_channels,
  77                              size_t num_samples)
  78 {
  79         assert(in_channel < in_num_channels);
  80         assert(out_channel < out_num_channels);
  81         src += in_channel * 3;
  82         dst += out_channel;
  83
  84         for (size_t i = 0; i < num_samples; ++i) {
  85                 uint32_t s1 = src[0];
  86                 uint32_t s2 = src[1];
  87                 uint32_t s3 = src[2];
  88                 uint32_t s = (s1 << 8) | (s2 << 16) | (s3 << 24);  // Note: The bottom eight bits are zero; s3 includes the sign bit.
  89                 *dst = int(s) * (1.0f / (256.0f * 8388608.0f));  // 256 for signed down-shift by 8, then 2^23 for the actual conversion.
  90
  91                 src += 3 * in_num_channels;
  92                 dst += out_num_channels;
  93         }
  94 }
  95
  96 void convert_fixed24_to_fixed32(int32_t *dst, size_t out_channel, size_t out_num_channels,
  97                                 const uint8_t *src, size_t in_channel, size_t in_num_channels,
  98                                 size_t num_samples)
  99 {
 100         assert(in_channel < in_num_channels);
 101         assert(out_channel < out_num_channels);
 102         src += in_channel * 3;
 103         dst += out_channel;
 104
 105         for (size_t i = 0; i < num_samples; ++i) {
 106                 uint32_t s1 = src[0];
 107                 uint32_t s2 = src[1];
 108                 uint32_t s3 = src[2];
 109                 uint32_t s = (s1 << 8) | (s2 << 16) | (s3 << 24);
 110
 111                 // Keep the sign bit in place, repeat the other 23 bits as far as they go.
 112                 *dst = s | ((s & 0x7fffffff) >> 23);
 113
 114                 src += 3 * in_num_channels;
 115                 dst += out_num_channels;
 116         }
 117 }
 118
 119 void convert_fixed32_to_fp32(float *dst, size_t out_channel, size_t out_num_channels,
 120                              const uint8_t *src, size_t in_channel, size_t in_num_channels,
 121                              size_t num_samples)
 122 {
 123         assert(in_channel < in_num_channels);
 124         assert(out_channel < out_num_channels);
 125         src += in_channel * 4;
 126         dst += out_channel;
 127
 128         for (size_t i = 0; i < num_samples; ++i) {
 129                 int32_t s = le32toh(*(int32_t *)src);
 130                 *dst = s * (1.0f / 2147483648.0f);
 131
 132                 src += 4 * in_num_channels;
 133                 dst += out_num_channels;
 134         }
 135 }
 136
 137 // Basically just a reinterleave.
 138 void convert_fixed32_to_fixed32(int32_t *dst, size_t out_channel, size_t out_num_channels,
 139                                 const uint8_t *src, size_t in_channel, size_t in_num_channels,
 140                                 size_t num_samples)
 141 {
 142         assert(in_channel < in_num_channels);
 143         assert(out_channel < out_num_channels);
 144         src += in_channel * 4;
 145         dst += out_channel;
 146
 147         for (size_t i = 0; i < num_samples; ++i) {
 148                 int32_t s = le32toh(*(int32_t *)src);
 149                 *dst = s;
 150
 151                 src += 4 * in_num_channels;
 152                 dst += out_num_channels;
 153         }
 154 }
 155
 156 float find_peak_plain(const float *samples, size_t num_samples) __attribute__((unused));
 157
 158 float find_peak_plain(const float *samples, size_t num_samples)
 159 {
 160         float m = fabs(samples[0]);
 161         for (size_t i = 1; i < num_samples; ++i) {
 162                 m = max(m, fabs(samples[i]));
 163         }
 164         return m;
 165 }
 166
 167 #ifdef __SSE__
 168 static inline float horizontal_max(__m128 m)
 169 {
 170         __m128 tmp = _mm_shuffle_ps(m, m, _MM_SHUFFLE(1, 0, 3, 2));
 171         m = _mm_max_ps(m, tmp);
 172         tmp = _mm_shuffle_ps(m, m, _MM_SHUFFLE(2, 3, 0, 1));
 173         m = _mm_max_ps(m, tmp);
 174         return _mm_cvtss_f32(m);
 175 }
 176
 177 float find_peak(const float *samples, size_t num_samples)
 178 {
 179         const __m128 abs_mask = _mm_castsi128_ps(_mm_set1_epi32(0x7fffffffu));
 180         __m128 m = _mm_setzero_ps();
 181         for (size_t i = 0; i < (num_samples & ~3); i += 4) {
 182                 __m128 x = _mm_loadu_ps(samples + i);
 183                 x = _mm_and_ps(x, abs_mask);
 184                 m = _mm_max_ps(m, x);
 185         }
 186         float result = horizontal_max(m);
 187
 188         for (size_t i = (num_samples & ~3); i < num_samples; ++i) {
 189                 result = max(result, fabs(samples[i]));
 190         }
 191
 192 #if 0
 193         // Self-test. We should be bit-exact the same.
 194         float reference_result = find_peak_plain(samples, num_samples);
 195         if (result != reference_result) {
 196                 fprintf(stderr, "Error: Peak is %f [%f %f %f %f]; should be %f.\n",
 197                         result,
 198                         _mm_cvtss_f32(_mm_shuffle_ps(m, m, _MM_SHUFFLE(0, 0, 0, 0))),
 199                         _mm_cvtss_f32(_mm_shuffle_ps(m, m, _MM_SHUFFLE(1, 1, 1, 1))),
 200                         _mm_cvtss_f32(_mm_shuffle_ps(m, m, _MM_SHUFFLE(2, 2, 2, 2))),
 201                         _mm_cvtss_f32(_mm_shuffle_ps(m, m, _MM_SHUFFLE(3, 3, 3, 3))),
 202                         reference_result);
 203                 abort();
 204         }
 205 #endif
 206         return result;
 207 }
 208 #else
 209 float find_peak(const float *samples, size_t num_samples)
 210 {
 211         return find_peak_plain(samples, num_samples);
 212 }
 213 #endif
 214
 215 void deinterleave_samples(const vector<float> &in, vector<float> *out_l, vector<float> *out_r)
 216 {
 217         size_t num_samples = in.size() / 2;
 218         out_l->resize(num_samples);
 219         out_r->resize(num_samples);
 220
 221         const float *inptr = in.data();
 222         float *lptr = &(*out_l)[0];
 223         float *rptr = &(*out_r)[0];
 224         for (size_t i = 0; i < num_samples; ++i) {
 225                 *lptr++ = *inptr++;
 226                 *rptr++ = *inptr++;
 227         }
 228 }
 229
 230 }  // namespace
 231
 232 AudioMixer::AudioMixer(unsigned num_capture_cards, unsigned num_ffmpeg_inputs)
 233         : num_capture_cards(num_capture_cards),
 234           num_ffmpeg_inputs(num_ffmpeg_inputs),
 235           ffmpeg_inputs(new AudioDevice[num_ffmpeg_inputs]),
 236           limiter(OUTPUT_FREQUENCY),
 237           correlation(OUTPUT_FREQUENCY)
 238 {
 239         for (unsigned bus_index = 0; bus_index < MAX_BUSES; ++bus_index) {
 240                 locut[bus_index].init(FILTER_HPF, 2);
 241                 eq[bus_index][EQ_BAND_BASS].init(FILTER_LOW_SHELF, 1);
 242                 // Note: EQ_BAND_MID isn't used (see comments in apply_eq()).
 243                 eq[bus_index][EQ_BAND_TREBLE].init(FILTER_HIGH_SHELF, 1);
 244                 compressor[bus_index].reset(new StereoCompressor(OUTPUT_FREQUENCY));
 245                 level_compressor[bus_index].reset(new StereoCompressor(OUTPUT_FREQUENCY));
 246
 247                 set_bus_settings(bus_index, get_default_bus_settings());
 248         }
 249         set_limiter_enabled(global_flags.limiter_enabled);
 250         set_final_makeup_gain_auto(global_flags.final_makeup_gain_auto);
 251
 252         r128.init(2, OUTPUT_FREQUENCY);
 253         r128.integr_start();
 254
 255         // hlen=16 is pretty low quality, but we use quite a bit of CPU otherwise,
 256         // and there's a limit to how important the peak meter is.
 257         peak_resampler.setup(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY * 4, /*num_channels=*/2, /*hlen=*/16, /*frel=*/1.0);
 258
 259         global_audio_mixer = this;
 260         alsa_pool.init();
 261
 262         if (!global_flags.input_mapping_filename.empty()) {
 263                 // Must happen after ALSAPool is initialized, as it needs to know the card list.
 264                 current_mapping_mode = MappingMode::MULTICHANNEL;
 265                 InputMapping new_input_mapping;
 266                 if (!load_input_mapping_from_file(get_devices(),
 267                                                   global_flags.input_mapping_filename,
 268                                                   &new_input_mapping)) {
 269                         fprintf(stderr, "Failed to load input mapping from '%s', exiting.\n",
 270                                 global_flags.input_mapping_filename.c_str());
 271                         abort();
 272                 }
 273                 set_input_mapping(new_input_mapping);
 274         } else {
 275                 set_simple_input(/*card_index=*/0);
 276                 if (global_flags.multichannel_mapping_mode) {
 277                         current_mapping_mode = MappingMode::MULTICHANNEL;
 278                 }
 279         }
 280
 281         global_metrics.add("audio_loudness_short_lufs", &metric_audio_loudness_short_lufs, Metrics::TYPE_GAUGE);
 282         global_metrics.add("audio_loudness_integrated_lufs", &metric_audio_loudness_integrated_lufs, Metrics::TYPE_GAUGE);
 283         global_metrics.add("audio_loudness_range_low_lufs", &metric_audio_loudness_range_low_lufs, Metrics::TYPE_GAUGE);
 284         global_metrics.add("audio_loudness_range_high_lufs", &metric_audio_loudness_range_high_lufs, Metrics::TYPE_GAUGE);
 285         global_metrics.add("audio_peak_dbfs", &metric_audio_peak_dbfs, Metrics::TYPE_GAUGE);
 286         global_metrics.add("audio_final_makeup_gain_db", &metric_audio_final_makeup_gain_db, Metrics::TYPE_GAUGE);
 287         global_metrics.add("audio_correlation", &metric_audio_correlation, Metrics::TYPE_GAUGE);
 288 }
 289
 290 void AudioMixer::reset_resampler(DeviceSpec device_spec)
 291 {
 292         lock_guard<timed_mutex> lock(audio_mutex);
 293         reset_resampler_mutex_held(device_spec);
 294 }
 295
 296 void AudioMixer::reset_resampler_mutex_held(DeviceSpec device_spec)
 297 {
 298         AudioDevice *device = find_audio_device(device_spec);
 299
 300         if (device->interesting_channels.empty()) {
 301                 device->resampling_queue.reset();
 302         } else {
 303                 // Make sure we never get negative delay. Even 1 ms is probably way less than we
 304                 // could ever hope to actually have; this is just a failsafe.
 305                 double delay_ms = max(global_flags.audio_queue_length_ms + device->extra_delay_ms, 1.0);
 306
 307                 device->resampling_queue.reset(new ResamplingQueue(
 308                         device_spec, device->capture_frequency, OUTPUT_FREQUENCY, device->interesting_channels.size(),
 309                         delay_ms * 0.001));
 310         }
 311 }
 312
 313 bool AudioMixer::add_audio(DeviceSpec device_spec, const uint8_t *data, unsigned num_samples, AudioFormat audio_format, steady_clock::time_point frame_time)
 314 {
 315         AudioDevice *device = find_audio_device(device_spec);
 316
 317         unique_lock<timed_mutex> lock(audio_mutex, defer_lock);
 318         if (!lock.try_lock_for(chrono::milliseconds(10))) {
 319                 return false;
 320         }
 321         if (device->resampling_queue == nullptr) {
 322                 // No buses use this device; throw it away.
 323                 return true;
 324         }
 325
 326         unsigned num_channels = device->interesting_channels.size();
 327         assert(num_channels > 0);
 328
 329         // Convert the audio to fp32.
 330         unique_ptr<float[]> audio(new float[num_samples * num_channels]);
 331         unsigned channel_index = 0;
 332         for (auto channel_it = device->interesting_channels.cbegin(); channel_it != device->interesting_channels.end(); ++channel_it, ++channel_index) {
 333                 switch (audio_format.bits_per_sample) {
 334                 case 0:
 335                         assert(num_samples == 0);
 336                         break;
 337                 case 16:
 338                         convert_fixed16_to_fp32(audio.get(), channel_index, num_channels, data, *channel_it, audio_format.num_channels, num_samples);
 339                         break;
 340                 case 24:
 341                         convert_fixed24_to_fp32(audio.get(), channel_index, num_channels, data, *channel_it, audio_format.num_channels, num_samples);
 342                         break;
 343                 case 32:
 344                         convert_fixed32_to_fp32(audio.get(), channel_index, num_channels, data, *channel_it, audio_format.num_channels, num_samples);
 345                         break;
 346                 default:
 347                         fprintf(stderr, "Cannot handle audio with %u bits per sample\n", audio_format.bits_per_sample);
 348                         assert(false);
 349                 }
 350         }
 351
 352         // If we changed frequency since last frame, we'll need to reset the resampler.
 353         if (audio_format.sample_rate != device->capture_frequency) {
 354                 device->capture_frequency = audio_format.sample_rate;
 355                 reset_resampler_mutex_held(device_spec);
 356         }
 357
 358         // Now add it.
 359         device->resampling_queue->add_input_samples(frame_time, audio.get(), num_samples, ResamplingQueue::ADJUST_RATE);
 360         return true;
 361 }
 362
 363 vector<int32_t> convert_audio_to_fixed32(const uint8_t *data, unsigned num_samples, bmusb::AudioFormat audio_format, unsigned num_channels)
 364 {
 365         vector<int32_t> audio;
 366
 367         if (num_channels > audio_format.num_channels) {
 368                 audio.resize(num_samples * num_channels, 0);
 369         } else {
 370                 audio.resize(num_samples * num_channels);
 371         }
 372         for (unsigned channel_index = 0; channel_index < num_channels && channel_index < audio_format.num_channels; ++channel_index) {
 373                 switch (audio_format.bits_per_sample) {
 374                 case 0:
 375                         assert(num_samples == 0);
 376                         break;
 377                 case 16:
 378                         convert_fixed16_to_fixed32(&audio[0], channel_index, num_channels, data, channel_index, audio_format.num_channels, num_samples);
 379                         break;
 380                 case 24:
 381                         convert_fixed24_to_fixed32(&audio[0], channel_index, num_channels, data, channel_index, audio_format.num_channels, num_samples);
 382                         break;
 383                 case 32:
 384                         convert_fixed32_to_fixed32(&audio[0], channel_index, num_channels, data, channel_index, audio_format.num_channels, num_samples);
 385                         break;
 386                 default:
 387                         fprintf(stderr, "Cannot handle audio with %u bits per sample\n", audio_format.bits_per_sample);
 388                         assert(false);
 389                 }
 390         }
 391
 392         return audio;
 393 }
 394
 395 bool AudioMixer::add_silence(DeviceSpec device_spec, unsigned samples_per_frame, unsigned num_frames)
 396 {
 397         AudioDevice *device = find_audio_device(device_spec);
 398
 399         unique_lock<timed_mutex> lock(audio_mutex, defer_lock);
 400         if (!lock.try_lock_for(chrono::milliseconds(10))) {
 401                 return false;
 402         }
 403         if (device->resampling_queue == nullptr) {
 404                 // No buses use this device; throw it away.
 405                 return true;
 406         }
 407
 408         unsigned num_channels = device->interesting_channels.size();
 409         assert(num_channels > 0);
 410
 411         vector<float> silence(samples_per_frame * num_channels, 0.0f);
 412         for (unsigned i = 0; i < num_frames; ++i) {
 413                 device->resampling_queue->add_input_samples(steady_clock::now(), silence.data(), samples_per_frame, ResamplingQueue::DO_NOT_ADJUST_RATE);
 414         }
 415         return true;
 416 }
 417
 418 bool AudioMixer::silence_card(DeviceSpec device_spec, bool silence)
 419 {
 420         AudioDevice *device = find_audio_device(device_spec);
 421
 422         unique_lock<timed_mutex> lock(audio_mutex, defer_lock);
 423         if (!lock.try_lock_for(chrono::milliseconds(10))) {
 424                 return false;
 425         }
 426
 427         if (device->silenced && !silence) {
 428                 reset_resampler_mutex_held(device_spec);
 429         }
 430         device->silenced = silence;
 431         return true;
 432 }
 433
 434 AudioMixer::BusSettings AudioMixer::get_default_bus_settings()
 435 {
 436         BusSettings settings;
 437         settings.fader_volume_db = 0.0f;
 438         settings.muted = false;
 439         settings.locut_enabled = global_flags.locut_enabled;
 440         settings.stereo_width = 1.0f;
 441         for (unsigned band_index = 0; band_index < NUM_EQ_BANDS; ++band_index) {
 442                 settings.eq_level_db[band_index] = 0.0f;
 443         }
 444         settings.gain_staging_db = global_flags.initial_gain_staging_db;
 445         settings.level_compressor_enabled = global_flags.gain_staging_auto;
 446         settings.compressor_threshold_dbfs = ref_level_dbfs - 12.0f;  // -12 dB.
 447         settings.compressor_enabled = global_flags.compressor_enabled;
 448         return settings;
 449 }
 450
 451 AudioMixer::BusSettings AudioMixer::get_bus_settings(unsigned bus_index) const
 452 {
 453         lock_guard<timed_mutex> lock(audio_mutex);
 454         BusSettings settings;
 455         settings.fader_volume_db = fader_volume_db[bus_index];
 456         settings.muted = mute[bus_index];
 457         settings.locut_enabled = locut_enabled[bus_index];
 458         settings.stereo_width = stereo_width[bus_index];
 459         for (unsigned band_index = 0; band_index < NUM_EQ_BANDS; ++band_index) {
 460                 settings.eq_level_db[band_index] = eq_level_db[bus_index][band_index];
 461         }
 462         settings.gain_staging_db = gain_staging_db[bus_index];
 463         settings.level_compressor_enabled = level_compressor_enabled[bus_index];
 464         settings.compressor_threshold_dbfs = compressor_threshold_dbfs[bus_index];
 465         settings.compressor_enabled = compressor_enabled[bus_index];
 466         return settings;
 467 }
 468
 469 void AudioMixer::set_bus_settings(unsigned bus_index, const AudioMixer::BusSettings &settings)
 470 {
 471         lock_guard<timed_mutex> lock(audio_mutex);
 472         fader_volume_db[bus_index] = settings.fader_volume_db;
 473         mute[bus_index] = settings.muted;
 474         locut_enabled[bus_index] = settings.locut_enabled;
 475         stereo_width[bus_index] = settings.stereo_width;
 476         for (unsigned band_index = 0; band_index < NUM_EQ_BANDS; ++band_index) {
 477                 eq_level_db[bus_index][band_index] = settings.eq_level_db[band_index];
 478         }
 479         gain_staging_db[bus_index] = settings.gain_staging_db;
 480         last_gain_staging_db[bus_index] = gain_staging_db[bus_index];
 481         level_compressor_enabled[bus_index] = settings.level_compressor_enabled;
 482         compressor_threshold_dbfs[bus_index] = settings.compressor_threshold_dbfs;
 483         compressor_enabled[bus_index] = settings.compressor_enabled;
 484 }
 485
 486 AudioMixer::AudioDevice *AudioMixer::find_audio_device(DeviceSpec device)
 487 {
 488         switch (device.type) {
 489         case InputSourceType::CAPTURE_CARD:
 490                 return &video_cards[device.index];
 491         case InputSourceType::ALSA_INPUT:
 492                 return &alsa_inputs[device.index];
 493         case InputSourceType::FFMPEG_VIDEO_INPUT:
 494                 return &ffmpeg_inputs[device.index];
 495         case InputSourceType::SILENCE:
 496         default:
 497                 assert(false);
 498         }
 499         return nullptr;
 500 }
 501
 502 // Get a pointer to the given channel from the given device.
 503 // The channel must be picked out earlier and resampled.
 504 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)
 505 {
 506         static float zero = 0.0f;
 507         if (source_channel == -1 || device_spec.type == InputSourceType::SILENCE) {
 508                 *srcptr = &zero;
 509                 *stride = 0;
 510                 return;
 511         }
 512         AudioDevice *device = find_audio_device(device_spec);
 513         assert(device->interesting_channels.count(source_channel) != 0);
 514         unsigned channel_index = 0;
 515         for (int channel : device->interesting_channels) {
 516                 if (channel == source_channel) break;
 517                 ++channel_index;
 518         }
 519         assert(channel_index < device->interesting_channels.size());
 520         const auto it = samples_card.find(device_spec);
 521         assert(it != samples_card.end());
 522         *srcptr = &(it->second)[channel_index];
 523         *stride = device->interesting_channels.size();
 524 }
 525
 526 // TODO: Can be SSSE3-optimized if need be.
 527 void AudioMixer::fill_audio_bus(const map<DeviceSpec, vector<float>> &samples_card, const InputMapping::Bus &bus, unsigned num_samples, float stereo_width, float *output)
 528 {
 529         if (bus.device.type == InputSourceType::SILENCE) {
 530                 memset(output, 0, num_samples * 2 * sizeof(*output));
 531         } else {
 532                 assert(bus.device.type == InputSourceType::CAPTURE_CARD ||
 533                        bus.device.type == InputSourceType::ALSA_INPUT ||
 534                        bus.device.type == InputSourceType::FFMPEG_VIDEO_INPUT);
 535                 const float *lsrc, *rsrc;
 536                 unsigned lstride, rstride;
 537                 float *dptr = output;
 538                 find_sample_src_from_device(samples_card, bus.device, bus.source_channel[0], &lsrc, &lstride);
 539                 find_sample_src_from_device(samples_card, bus.device, bus.source_channel[1], &rsrc, &rstride);
 540
 541                 // Apply stereo width settings. Set stereo width w to a 0..1 range instead of
 542                 // -1..1, since it makes for much easier calculations (so 0.5 = completely mono).
 543                 // Then, what we want is
 544                 //
 545                 //   L' = wL + (1-w)R = R + w(L-R)
 546                 //   R' = wR + (1-w)L = L + w(R-L)
 547                 //
 548                 // This can be further simplified calculation-wise by defining the weighted
 549                 // difference signal D = w(R-L), so that:
 550                 //
 551                 //   L' = R - D
 552                 //   R' = L + D
 553                 float w = 0.5f * stereo_width + 0.5f;
 554                 if (bus.source_channel[0] == bus.source_channel[1]) {
 555                         // Mono anyway, so no need to bother.
 556                         w = 1.0f;
 557                 } else if (fabs(w) < 1e-3) {
 558                         // Perfect inverse.
 559                         swap(lsrc, rsrc);
 560                         swap(lstride, rstride);
 561                         w = 1.0f;
 562                 }
 563                 if (fabs(w - 1.0f) < 1e-3) {
 564                         // No calculations needed for stereo_width = 1.
 565                         for (unsigned i = 0; i < num_samples; ++i) {
 566                                 *dptr++ = *lsrc;
 567                                 *dptr++ = *rsrc;
 568                                 lsrc += lstride;
 569                                 rsrc += rstride;
 570                         }
 571                 } else {
 572                         // General case.
 573                         for (unsigned i = 0; i < num_samples; ++i) {
 574                                 float left = *lsrc, right = *rsrc;
 575                                 float diff = w * (right - left);
 576                                 *dptr++ = right - diff;
 577                                 *dptr++ = left + diff;
 578                                 lsrc += lstride;
 579                                 rsrc += rstride;
 580                         }
 581                 }
 582         }
 583 }
 584
 585 vector<DeviceSpec> AudioMixer::get_active_devices() const
 586 {
 587         vector<DeviceSpec> ret;
 588         for (unsigned card_index = 0; card_index < MAX_VIDEO_CARDS; ++card_index) {
 589                 const DeviceSpec device_spec{InputSourceType::CAPTURE_CARD, card_index};
 590                 if (!find_audio_device(device_spec)->interesting_channels.empty()) {
 591                         ret.push_back(device_spec);
 592                 }
 593         }
 594         for (unsigned card_index = 0; card_index < MAX_ALSA_CARDS; ++card_index) {
 595                 const DeviceSpec device_spec{InputSourceType::ALSA_INPUT, card_index};
 596                 if (!find_audio_device(device_spec)->interesting_channels.empty()) {
 597                         ret.push_back(device_spec);
 598                 }
 599         }
 600         for (unsigned card_index = 0; card_index < num_ffmpeg_inputs; ++card_index) {
 601                 const DeviceSpec device_spec{InputSourceType::FFMPEG_VIDEO_INPUT, card_index};
 602                 if (!find_audio_device(device_spec)->interesting_channels.empty()) {
 603                         ret.push_back(device_spec);
 604                 }
 605         }
 606         return ret;
 607 }
 608
 609 namespace {
 610
 611 void apply_gain(float db, float last_db, vector<float> *samples)
 612 {
 613         if (fabs(db - last_db) < 1e-3) {
 614                 // Constant over this frame.
 615                 const float gain = from_db(db);
 616                 for (size_t i = 0; i < samples->size(); ++i) {
 617                         (*samples)[i] *= gain;
 618                 }
 619         } else {
 620                 // We need to do a fade.
 621                 unsigned num_samples = samples->size() / 2;
 622                 float gain = from_db(last_db);
 623                 const float gain_inc = pow(from_db(db - last_db), 1.0 / num_samples);
 624                 for (size_t i = 0; i < num_samples; ++i) {
 625                         (*samples)[i * 2 + 0] *= gain;
 626                         (*samples)[i * 2 + 1] *= gain;
 627                         gain *= gain_inc;
 628                 }
 629         }
 630 }
 631
 632 }  // namespace
 633
 634 vector<float> AudioMixer::get_output(steady_clock::time_point ts, unsigned num_samples, ResamplingQueue::RateAdjustmentPolicy rate_adjustment_policy)
 635 {
 636         map<DeviceSpec, vector<float>> samples_card;
 637         vector<float> samples_bus;
 638
 639         lock_guard<timed_mutex> lock(audio_mutex);
 640
 641         // Pick out all the interesting channels from all the cards.
 642         for (const DeviceSpec &device_spec : get_active_devices()) {
 643                 AudioDevice *device = find_audio_device(device_spec);
 644                 samples_card[device_spec].resize(num_samples * device->interesting_channels.size());
 645                 if (device->silenced) {
 646                         memset(&samples_card[device_spec][0], 0, samples_card[device_spec].size() * sizeof(float));
 647                 } else {
 648                         device->resampling_queue->get_output_samples(
 649                                 ts,
 650                                 &samples_card[device_spec][0],
 651                                 num_samples,
 652                                 rate_adjustment_policy);
 653                 }
 654         }
 655
 656         vector<float> samples_out, left, right;
 657         samples_out.resize(num_samples * 2);
 658         samples_bus.resize(num_samples * 2);
 659         for (unsigned bus_index = 0; bus_index < input_mapping.buses.size(); ++bus_index) {
 660                 fill_audio_bus(samples_card, input_mapping.buses[bus_index], num_samples, stereo_width[bus_index], &samples_bus[0]);
 661                 apply_eq(bus_index, &samples_bus);
 662
 663                 {
 664                         lock_guard<mutex> lock(compressor_mutex);
 665
 666                         // Apply a level compressor to get the general level right.
 667                         // Basically, if it's over about -40 dBFS, we squeeze it down to that level
 668                         // (or more precisely, near it, since we don't use infinite ratio),
 669                         // then apply a makeup gain to get it to -14 dBFS. -14 dBFS is, of course,
 670                         // entirely arbitrary, but from practical tests with speech, it seems to
 671                         // put ut around -23 LUFS, so it's a reasonable starting point for later use.
 672                         if (level_compressor_enabled[bus_index]) {
 673                                 float threshold = 0.01f;   // -40 dBFS.
 674                                 float ratio = 20.0f;
 675                                 float attack_time = 0.5f;
 676                                 float release_time = 20.0f;
 677                                 float makeup_gain = from_db(ref_level_dbfs - (-40.0f));  // +26 dB.
 678                                 level_compressor[bus_index]->process(samples_bus.data(), samples_bus.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 679                                 gain_staging_db[bus_index] = to_db(level_compressor[bus_index]->get_attenuation() * makeup_gain);
 680                         } else {
 681                                 // Just apply the gain we already had.
 682                                 float db = gain_staging_db[bus_index];
 683                                 float last_db = last_gain_staging_db[bus_index];
 684                                 apply_gain(db, last_db, &samples_bus);
 685                         }
 686                         last_gain_staging_db[bus_index] = gain_staging_db[bus_index];
 687
 688 #if 0
 689                         printf("level=%f (%+5.2f dBFS) attenuation=%f (%+5.2f dB) end_result=%+5.2f dB\n",
 690                                 level_compressor.get_level(), to_db(level_compressor.get_level()),
 691                                 level_compressor.get_attenuation(), to_db(level_compressor.get_attenuation()),
 692                                 to_db(level_compressor.get_level() * level_compressor.get_attenuation() * makeup_gain));
 693 #endif
 694
 695                         // The real compressor.
 696                         if (compressor_enabled[bus_index]) {
 697                                 float threshold = from_db(compressor_threshold_dbfs[bus_index]);
 698                                 float ratio = 20.0f;
 699                                 float attack_time = 0.005f;
 700                                 float release_time = 0.040f;
 701                                 float makeup_gain = 2.0f;  // +6 dB.
 702                                 compressor[bus_index]->process(samples_bus.data(), samples_bus.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 703                 //              compressor_att = compressor.get_attenuation();
 704                         }
 705                 }
 706
 707                 add_bus_to_master(bus_index, samples_bus, &samples_out);
 708                 deinterleave_samples(samples_bus, &left, &right);
 709                 measure_bus_levels(bus_index, left, right);
 710         }
 711
 712         {
 713                 lock_guard<mutex> lock(compressor_mutex);
 714
 715                 // Finally a limiter at -4 dB (so, -10 dBFS) to take out the worst peaks only.
 716                 // Note that since ratio is not infinite, we could go slightly higher than this.
 717                 if (limiter_enabled) {
 718                         float threshold = from_db(limiter_threshold_dbfs);
 719                         float ratio = 30.0f;
 720                         float attack_time = 0.0f;  // Instant.
 721                         float release_time = 0.020f;
 722                         float makeup_gain = 1.0f;  // 0 dB.
 723                         limiter.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 724         //              limiter_att = limiter.get_attenuation();
 725                 }
 726
 727         //      printf("limiter=%+5.1f  compressor=%+5.1f\n", to_db(limiter_att), to_db(compressor_att));
 728         }
 729
 730         // At this point, we are most likely close to +0 LU (at least if the
 731         // faders sum to 0 dB and the compressors are on), but all of our
 732         // measurements have been on raw sample values, not R128 values.
 733         // So we have a final makeup gain to get us to +0 LU; the gain
 734         // adjustments required should be relatively small, and also, the
 735         // offset shouldn't change much (only if the type of audio changes
 736         // significantly). Thus, we shoot for updating this value basically
 737         // “whenever we process buffers”, since the R128 calculation isn't exactly
 738         // something we get out per-sample.
 739         //
 740         // Note that there's a feedback loop here, so we choose a very slow filter
 741         // (half-time of 30 seconds).
 742         double target_loudness_factor, alpha;
 743         double loudness_lu = r128.loudness_M() - ref_level_lufs;
 744         target_loudness_factor = final_makeup_gain * from_db(-loudness_lu);
 745
 746         // If we're outside +/- 5 LU (after correction), we don't count it as
 747         // a normal signal (probably silence) and don't change the
 748         // correction factor; just apply what we already have.
 749         if (fabs(loudness_lu) >= 5.0 || !final_makeup_gain_auto) {
 750                 alpha = 0.0;
 751         } else {
 752                 // Formula adapted from
 753                 // https://en.wikipedia.org/wiki/Low-pass_filter#Simple_infinite_impulse_response_filter.
 754                 const double half_time_s = 30.0;
 755                 const double fc_mul_2pi_delta_t = 1.0 / (half_time_s * OUTPUT_FREQUENCY);
 756                 alpha = fc_mul_2pi_delta_t / (fc_mul_2pi_delta_t + 1.0);
 757         }
 758
 759         {
 760                 lock_guard<mutex> lock(compressor_mutex);
 761                 double m = final_makeup_gain;
 762                 for (size_t i = 0; i < samples_out.size(); i += 2) {
 763                         samples_out[i + 0] *= m;
 764                         samples_out[i + 1] *= m;
 765                         m += (target_loudness_factor - m) * alpha;
 766                 }
 767                 final_makeup_gain = m;
 768         }
 769
 770         update_meters(samples_out);
 771
 772         return samples_out;
 773 }
 774
 775 namespace {
 776
 777 void apply_filter_fade(StereoFilter *filter, float *data, unsigned num_samples, float cutoff_hz, float db, float last_db)
 778 {
 779         // A granularity of 32 samples is an okay tradeoff between speed and
 780         // smoothness; recalculating the filters is pretty expensive, so it's
 781         // good that we don't do this all the time.
 782         static constexpr unsigned filter_granularity_samples = 32;
 783
 784         const float cutoff_linear = cutoff_hz * 2.0 * M_PI / OUTPUT_FREQUENCY;
 785         if (fabs(db - last_db) < 1e-3) {
 786                 // Constant over this frame.
 787                 if (fabs(db) > 0.01f) {
 788                         filter->render(data, num_samples, cutoff_linear, 0.5f, db / 40.0f);
 789                 }
 790         } else {
 791                 // We need to do a fade. (Rounding up avoids division by zero.)
 792                 unsigned num_blocks = (num_samples + filter_granularity_samples - 1) / filter_granularity_samples;
 793                 const float inc_db_norm = (db - last_db) / 40.0f / num_blocks;
 794                 float db_norm = db / 40.0f;
 795                 for (size_t i = 0; i < num_samples; i += filter_granularity_samples) {
 796                         size_t samples_this_block = std::min<size_t>(num_samples - i, filter_granularity_samples);
 797                         filter->render(data + i * 2, samples_this_block, cutoff_linear, 0.5f, db_norm);
 798                         db_norm += inc_db_norm;
 799                 }
 800         }
 801 }
 802
 803 }  // namespace
 804
 805 void AudioMixer::apply_eq(unsigned bus_index, vector<float> *samples_bus)
 806 {
 807         constexpr float bass_freq_hz = 200.0f;
 808         constexpr float treble_freq_hz = 4700.0f;
 809
 810         // Cut away everything under 120 Hz (or whatever the cutoff is);
 811         // we don't need it for voice, and it will reduce headroom
 812         // and confuse the compressor. (In particular, any hums at 50 or 60 Hz
 813         // should be dampened.)
 814         if (locut_enabled[bus_index]) {
 815                 locut[bus_index].render(samples_bus->data(), samples_bus->size() / 2, locut_cutoff_hz * 2.0 * M_PI / OUTPUT_FREQUENCY, 0.5f);
 816         }
 817
 818         // Apply the rest of the EQ. Since we only have a simple three-band EQ,
 819         // we can implement it with two shelf filters. We use a simple gain to
 820         // set the mid-level filter, and then offset the low and high bands
 821         // from that if we need to. (We could perhaps have folded the gain into
 822         // the next part, but it's so cheap that the trouble isn't worth it.)
 823         //
 824         // If any part of the EQ has changed appreciably since last frame,
 825         // we fade smoothly during the course of this frame.
 826         const float bass_db = eq_level_db[bus_index][EQ_BAND_BASS];
 827         const float mid_db = eq_level_db[bus_index][EQ_BAND_MID];
 828         const float treble_db = eq_level_db[bus_index][EQ_BAND_TREBLE];
 829
 830         const float last_bass_db = last_eq_level_db[bus_index][EQ_BAND_BASS];
 831         const float last_mid_db = last_eq_level_db[bus_index][EQ_BAND_MID];
 832         const float last_treble_db = last_eq_level_db[bus_index][EQ_BAND_TREBLE];
 833
 834         assert(samples_bus->size() % 2 == 0);
 835         const unsigned num_samples = samples_bus->size() / 2;
 836
 837         apply_gain(mid_db, last_mid_db, samples_bus);
 838
 839         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);
 840         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);
 841
 842         last_eq_level_db[bus_index][EQ_BAND_BASS] = bass_db;
 843         last_eq_level_db[bus_index][EQ_BAND_MID] = mid_db;
 844         last_eq_level_db[bus_index][EQ_BAND_TREBLE] = treble_db;
 845 }
 846
 847 void AudioMixer::add_bus_to_master(unsigned bus_index, const vector<float> &samples_bus, vector<float> *samples_out)
 848 {
 849         assert(samples_bus.size() == samples_out->size());
 850         assert(samples_bus.size() % 2 == 0);
 851         unsigned num_samples = samples_bus.size() / 2;
 852         const float new_volume_db = mute[bus_index] ? -90.0f : fader_volume_db[bus_index].load();
 853         if (fabs(new_volume_db - last_fader_volume_db[bus_index]) > 1e-3) {
 854                 // The volume has changed; do a fade over the course of this frame.
 855                 // (We might have some numerical issues here, but it seems to sound OK.)
 856                 // For the purpose of fading here, the silence floor is set to -90 dB
 857                 // (the fader only goes to -84).
 858                 float old_volume = from_db(max<float>(last_fader_volume_db[bus_index], -90.0f));
 859                 float volume = from_db(max<float>(new_volume_db, -90.0f));
 860
 861                 float volume_inc = pow(volume / old_volume, 1.0 / num_samples);
 862                 volume = old_volume;
 863                 if (bus_index == 0) {
 864                         for (unsigned i = 0; i < num_samples; ++i) {
 865                                 (*samples_out)[i * 2 + 0] = samples_bus[i * 2 + 0] * volume;
 866                                 (*samples_out)[i * 2 + 1] = samples_bus[i * 2 + 1] * volume;
 867                                 volume *= volume_inc;
 868                         }
 869                 } else {
 870                         for (unsigned i = 0; i < num_samples; ++i) {
 871                                 (*samples_out)[i * 2 + 0] += samples_bus[i * 2 + 0] * volume;
 872                                 (*samples_out)[i * 2 + 1] += samples_bus[i * 2 + 1] * volume;
 873                                 volume *= volume_inc;
 874                         }
 875                 }
 876         } else if (new_volume_db > -90.0f) {
 877                 float volume = from_db(new_volume_db);
 878                 if (bus_index == 0) {
 879                         for (unsigned i = 0; i < num_samples; ++i) {
 880                                 (*samples_out)[i * 2 + 0] = samples_bus[i * 2 + 0] * volume;
 881                                 (*samples_out)[i * 2 + 1] = samples_bus[i * 2 + 1] * volume;
 882                         }
 883                 } else {
 884                         for (unsigned i = 0; i < num_samples; ++i) {
 885                                 (*samples_out)[i * 2 + 0] += samples_bus[i * 2 + 0] * volume;
 886                                 (*samples_out)[i * 2 + 1] += samples_bus[i * 2 + 1] * volume;
 887                         }
 888                 }
 889         }
 890
 891         last_fader_volume_db[bus_index] = new_volume_db;
 892 }
 893
 894 void AudioMixer::measure_bus_levels(unsigned bus_index, const vector<float> &left, const vector<float> &right)
 895 {
 896         assert(left.size() == right.size());
 897         const float volume = mute[bus_index] ? 0.0f : from_db(fader_volume_db[bus_index]);
 898         const float peak_levels[2] = {
 899                 find_peak(left.data(), left.size()) * volume,
 900                 find_peak(right.data(), right.size()) * volume
 901         };
 902         for (unsigned channel = 0; channel < 2; ++channel) {
 903                 // Compute the current value, including hold and falloff.
 904                 // The constants are borrowed from zita-mu1 by Fons Adriaensen.
 905                 static constexpr float hold_sec = 0.5f;
 906                 static constexpr float falloff_db_sec = 15.0f;  // dB/sec falloff after hold.
 907                 float current_peak;
 908                 PeakHistory &history = peak_history[bus_index][channel];
 909                 history.historic_peak = max(history.historic_peak, peak_levels[channel]);
 910                 if (history.age_seconds < hold_sec) {
 911                         current_peak = history.last_peak;
 912                 } else {
 913                         current_peak = history.last_peak * from_db(-falloff_db_sec * (history.age_seconds - hold_sec));
 914                 }
 915
 916                 // See if we have a new peak to replace the old (possibly falling) one.
 917                 if (peak_levels[channel] > current_peak) {
 918                         history.last_peak = peak_levels[channel];
 919                         history.age_seconds = 0.0f;  // Not 100% correct, but more than good enough given our frame sizes.
 920                         current_peak = peak_levels[channel];
 921                 } else {
 922                         history.age_seconds += float(left.size()) / OUTPUT_FREQUENCY;
 923                 }
 924                 history.current_level = peak_levels[channel];
 925                 history.current_peak = current_peak;
 926         }
 927 }
 928
 929 void AudioMixer::update_meters(const vector<float> &samples)
 930 {
 931         // Upsample 4x to find interpolated peak.
 932         peak_resampler.inp_data = const_cast<float *>(samples.data());
 933         peak_resampler.inp_count = samples.size() / 2;
 934
 935         vector<float> interpolated_samples;
 936         interpolated_samples.resize(samples.size());
 937         {
 938                 lock_guard<mutex> lock(audio_measure_mutex);
 939
 940                 while (peak_resampler.inp_count > 0) {  // About four iterations.
 941                         peak_resampler.out_data = &interpolated_samples[0];
 942                         peak_resampler.out_count = interpolated_samples.size() / 2;
 943                         peak_resampler.process();
 944                         size_t out_stereo_samples = interpolated_samples.size() / 2 - peak_resampler.out_count;
 945                         peak = max<float>(peak, find_peak(interpolated_samples.data(), out_stereo_samples * 2));
 946                         peak_resampler.out_data = nullptr;
 947                 }
 948         }
 949
 950         // Find R128 levels and L/R correlation.
 951         vector<float> left, right;
 952         deinterleave_samples(samples, &left, &right);
 953         float *ptrs[] = { left.data(), right.data() };
 954         {
 955                 lock_guard<mutex> lock(audio_measure_mutex);
 956                 r128.process(left.size(), ptrs);
 957                 correlation.process_samples(samples);
 958         }
 959
 960         send_audio_level_callback();
 961 }
 962
 963 void AudioMixer::reset_meters()
 964 {
 965         lock_guard<mutex> lock(audio_measure_mutex);
 966         peak_resampler.reset();
 967         peak = 0.0f;
 968         r128.reset();
 969         r128.integr_start();
 970         correlation.reset();
 971 }
 972
 973 void AudioMixer::send_audio_level_callback()
 974 {
 975         if (audio_level_callback == nullptr) {
 976                 return;
 977         }
 978
 979         lock_guard<mutex> lock(audio_measure_mutex);
 980         double loudness_s = r128.loudness_S();
 981         double loudness_i = r128.integrated();
 982         double loudness_range_low = r128.range_min();
 983         double loudness_range_high = r128.range_max();
 984
 985         metric_audio_loudness_short_lufs = loudness_s;
 986         metric_audio_loudness_integrated_lufs = loudness_i;
 987         metric_audio_loudness_range_low_lufs = loudness_range_low;
 988         metric_audio_loudness_range_high_lufs = loudness_range_high;
 989         metric_audio_peak_dbfs = to_db(peak);
 990         metric_audio_final_makeup_gain_db = to_db(final_makeup_gain);
 991         metric_audio_correlation = correlation.get_correlation();
 992
 993         vector<BusLevel> bus_levels;
 994         bus_levels.resize(input_mapping.buses.size());
 995         {
 996                 lock_guard<mutex> lock(compressor_mutex);
 997                 for (unsigned bus_index = 0; bus_index < bus_levels.size(); ++bus_index) {
 998                         BusLevel &levels = bus_levels[bus_index];
 999                         BusMetrics &metrics = bus_metrics[bus_index];
1000
1001                         levels.current_level_dbfs[0] = metrics.current_level_dbfs[0] = to_db(peak_history[bus_index][0].current_level);
1002                         levels.current_level_dbfs[1] = metrics.current_level_dbfs[1] = to_db(peak_history[bus_index][1].current_level);
1003                         levels.peak_level_dbfs[0] = metrics.peak_level_dbfs[0] = to_db(peak_history[bus_index][0].current_peak);
1004                         levels.peak_level_dbfs[1] = metrics.peak_level_dbfs[1] = to_db(peak_history[bus_index][1].current_peak);
1005                         levels.historic_peak_dbfs = metrics.historic_peak_dbfs = to_db(
1006                                 max(peak_history[bus_index][0].historic_peak,
1007                                     peak_history[bus_index][1].historic_peak));
1008                         levels.gain_staging_db = metrics.gain_staging_db = gain_staging_db[bus_index];
1009                         if (compressor_enabled[bus_index]) {
1010                                 levels.compressor_attenuation_db = metrics.compressor_attenuation_db = -to_db(compressor[bus_index]->get_attenuation());
1011                         } else {
1012                                 levels.compressor_attenuation_db = 0.0;
1013                                 metrics.compressor_attenuation_db = 0.0 / 0.0;
1014                         }
1015                 }
1016         }
1017
1018         audio_level_callback(loudness_s, to_db(peak), bus_levels,
1019                 loudness_i, loudness_range_low, loudness_range_high,
1020                 to_db(final_makeup_gain),
1021                 correlation.get_correlation());
1022 }
1023
1024 map<DeviceSpec, DeviceInfo> AudioMixer::get_devices()
1025 {
1026         lock_guard<timed_mutex> lock(audio_mutex);
1027
1028         map<DeviceSpec, DeviceInfo> devices;
1029         for (unsigned card_index = 0; card_index < num_capture_cards; ++card_index) {
1030                 const DeviceSpec spec{ InputSourceType::CAPTURE_CARD, card_index };
1031                 const AudioDevice *device = &video_cards[card_index];
1032                 DeviceInfo info;
1033                 info.display_name = device->display_name;
1034                 info.num_channels = 8;
1035                 devices.insert(make_pair(spec, info));
1036         }
1037         vector<ALSAPool::Device> available_alsa_devices = alsa_pool.get_devices();
1038         for (unsigned card_index = 0; card_index < available_alsa_devices.size(); ++card_index) {
1039                 const DeviceSpec spec{ InputSourceType::ALSA_INPUT, card_index };
1040                 const ALSAPool::Device &device = available_alsa_devices[card_index];
1041                 DeviceInfo info;
1042                 info.display_name = device.display_name();
1043                 info.num_channels = device.num_channels;
1044                 info.alsa_name = device.name;
1045                 info.alsa_info = device.info;
1046                 info.alsa_address = device.address;
1047                 devices.insert(make_pair(spec, info));
1048         }
1049         for (unsigned card_index = 0; card_index < num_ffmpeg_inputs; ++card_index) {
1050                 const DeviceSpec spec{ InputSourceType::FFMPEG_VIDEO_INPUT, card_index };
1051                 const AudioDevice *device = &ffmpeg_inputs[card_index];
1052                 DeviceInfo info;
1053                 info.display_name = device->display_name;
1054                 info.num_channels = 2;
1055                 devices.insert(make_pair(spec, info));
1056         }
1057         return devices;
1058 }
1059
1060 void AudioMixer::set_display_name(DeviceSpec device_spec, const string &name)
1061 {
1062         AudioDevice *device = find_audio_device(device_spec);
1063
1064         lock_guard<timed_mutex> lock(audio_mutex);
1065         device->display_name = name;
1066 }
1067
1068 void AudioMixer::serialize_device(DeviceSpec device_spec, DeviceSpecProto *device_spec_proto)
1069 {
1070         lock_guard<timed_mutex> lock(audio_mutex);
1071         switch (device_spec.type) {
1072                 case InputSourceType::SILENCE:
1073                         device_spec_proto->set_type(DeviceSpecProto::SILENCE);
1074                         break;
1075                 case InputSourceType::CAPTURE_CARD:
1076                         device_spec_proto->set_type(DeviceSpecProto::CAPTURE_CARD);
1077                         device_spec_proto->set_index(device_spec.index);
1078                         device_spec_proto->set_display_name(video_cards[device_spec.index].display_name);
1079                         break;
1080                 case InputSourceType::ALSA_INPUT:
1081                         alsa_pool.serialize_device(device_spec.index, device_spec_proto);
1082                         break;
1083                 case InputSourceType::FFMPEG_VIDEO_INPUT:
1084                         device_spec_proto->set_type(DeviceSpecProto::FFMPEG_VIDEO_INPUT);
1085                         device_spec_proto->set_index(device_spec.index);
1086                         device_spec_proto->set_display_name(ffmpeg_inputs[device_spec.index].display_name);
1087                         break;
1088         }
1089 }
1090
1091 void AudioMixer::set_simple_input(unsigned card_index)
1092 {
1093         assert(card_index < num_capture_cards + num_ffmpeg_inputs);
1094         InputMapping new_input_mapping;
1095         InputMapping::Bus input;
1096         input.name = "Main";
1097         if (card_index >= num_capture_cards) {
1098                 input.device = DeviceSpec{InputSourceType::FFMPEG_VIDEO_INPUT, card_index - num_capture_cards};
1099         } else {
1100                 input.device = DeviceSpec{InputSourceType::CAPTURE_CARD, card_index};
1101         }
1102         input.source_channel[0] = 0;
1103         input.source_channel[1] = 1;
1104
1105         new_input_mapping.buses.push_back(input);
1106
1107         lock_guard<timed_mutex> lock(audio_mutex);
1108         current_mapping_mode = MappingMode::SIMPLE;
1109         set_input_mapping_lock_held(new_input_mapping);
1110         fader_volume_db[0] = 0.0f;
1111 }
1112
1113 unsigned AudioMixer::get_simple_input() const
1114 {
1115         lock_guard<timed_mutex> lock(audio_mutex);
1116         if (input_mapping.buses.size() == 1 &&
1117             input_mapping.buses[0].device.type == InputSourceType::CAPTURE_CARD &&
1118             input_mapping.buses[0].source_channel[0] == 0 &&
1119             input_mapping.buses[0].source_channel[1] == 1) {
1120                 return input_mapping.buses[0].device.index;
1121         } else if (input_mapping.buses.size() == 1 &&
1122                    input_mapping.buses[0].device.type == InputSourceType::FFMPEG_VIDEO_INPUT &&
1123                    input_mapping.buses[0].source_channel[0] == 0 &&
1124                    input_mapping.buses[0].source_channel[1] == 1) {
1125                 return input_mapping.buses[0].device.index + num_capture_cards;
1126         } else {
1127                 return numeric_limits<unsigned>::max();
1128         }
1129 }
1130
1131 void AudioMixer::set_input_mapping(const InputMapping &new_input_mapping)
1132 {
1133         lock_guard<timed_mutex> lock(audio_mutex);
1134         set_input_mapping_lock_held(new_input_mapping);
1135         current_mapping_mode = MappingMode::MULTICHANNEL;
1136 }
1137
1138 AudioMixer::MappingMode AudioMixer::get_mapping_mode() const
1139 {
1140         lock_guard<timed_mutex> lock(audio_mutex);
1141         return current_mapping_mode;
1142 }
1143
1144 void AudioMixer::set_input_mapping_lock_held(const InputMapping &new_input_mapping)
1145 {
1146         map<DeviceSpec, set<unsigned>> interesting_channels;
1147         for (const InputMapping::Bus &bus : new_input_mapping.buses) {
1148                 if (bus.device.type == InputSourceType::CAPTURE_CARD ||
1149                     bus.device.type == InputSourceType::ALSA_INPUT ||
1150                     bus.device.type == InputSourceType::FFMPEG_VIDEO_INPUT) {
1151                         for (unsigned channel = 0; channel < 2; ++channel) {
1152                                 if (bus.source_channel[channel] != -1) {
1153                                         interesting_channels[bus.device].insert(bus.source_channel[channel]);
1154                                 }
1155                         }
1156                 } else {
1157                         assert(bus.device.type == InputSourceType::SILENCE);
1158                 }
1159         }
1160
1161         // Kill all the old metrics, and set up new ones.
1162         for (unsigned bus_index = 0; bus_index < input_mapping.buses.size(); ++bus_index) {
1163                 BusMetrics &metrics = bus_metrics[bus_index];
1164
1165                 vector<pair<string, string>> labels_left = metrics.labels;
1166                 labels_left.emplace_back("channel", "left");
1167                 vector<pair<string, string>> labels_right = metrics.labels;
1168                 labels_right.emplace_back("channel", "right");
1169
1170                 global_metrics.remove("bus_current_level_dbfs", labels_left);
1171                 global_metrics.remove("bus_current_level_dbfs", labels_right);
1172                 global_metrics.remove("bus_peak_level_dbfs", labels_left);
1173                 global_metrics.remove("bus_peak_level_dbfs", labels_right);
1174                 global_metrics.remove("bus_historic_peak_dbfs", metrics.labels);
1175                 global_metrics.remove("bus_gain_staging_db", metrics.labels);
1176                 global_metrics.remove("bus_compressor_attenuation_db", metrics.labels);
1177         }
1178         bus_metrics.reset(new BusMetrics[new_input_mapping.buses.size()]);
1179         for (unsigned bus_index = 0; bus_index < new_input_mapping.buses.size(); ++bus_index) {
1180                 const InputMapping::Bus &bus = new_input_mapping.buses[bus_index];
1181                 BusMetrics &metrics = bus_metrics[bus_index];
1182
1183                 char bus_index_str[16], source_index_str[16], source_channels_str[64];
1184                 snprintf(bus_index_str, sizeof(bus_index_str), "%u", bus_index);
1185                 snprintf(source_index_str, sizeof(source_index_str), "%u", bus.device.index);
1186                 snprintf(source_channels_str, sizeof(source_channels_str), "%d:%d", bus.source_channel[0], bus.source_channel[1]);
1187
1188                 vector<pair<string, string>> labels;
1189                 metrics.labels.emplace_back("index", bus_index_str);
1190                 metrics.labels.emplace_back("name", bus.name);
1191                 if (bus.device.type == InputSourceType::SILENCE) {
1192                         metrics.labels.emplace_back("source_type", "silence");
1193                 } else if (bus.device.type == InputSourceType::CAPTURE_CARD) {
1194                         metrics.labels.emplace_back("source_type", "capture_card");
1195                 } else if (bus.device.type == InputSourceType::ALSA_INPUT) {
1196                         metrics.labels.emplace_back("source_type", "alsa_input");
1197                 } else if (bus.device.type == InputSourceType::FFMPEG_VIDEO_INPUT) {
1198                         metrics.labels.emplace_back("source_type", "ffmpeg_video_input");
1199                 } else {
1200                         assert(false);
1201                 }
1202                 metrics.labels.emplace_back("source_index", source_index_str);
1203                 metrics.labels.emplace_back("source_channels", source_channels_str);
1204
1205                 vector<pair<string, string>> labels_left = metrics.labels;
1206                 labels_left.emplace_back("channel", "left");
1207                 vector<pair<string, string>> labels_right = metrics.labels;
1208                 labels_right.emplace_back("channel", "right");
1209
1210                 global_metrics.add("bus_current_level_dbfs", labels_left, &metrics.current_level_dbfs[0], Metrics::TYPE_GAUGE);
1211                 global_metrics.add("bus_current_level_dbfs", labels_right, &metrics.current_level_dbfs[1], Metrics::TYPE_GAUGE);
1212                 global_metrics.add("bus_peak_level_dbfs", labels_left, &metrics.peak_level_dbfs[0], Metrics::TYPE_GAUGE);
1213                 global_metrics.add("bus_peak_level_dbfs", labels_right, &metrics.peak_level_dbfs[1], Metrics::TYPE_GAUGE);
1214                 global_metrics.add("bus_historic_peak_dbfs", metrics.labels, &metrics.historic_peak_dbfs, Metrics::TYPE_GAUGE);
1215                 global_metrics.add("bus_gain_staging_db", metrics.labels, &metrics.gain_staging_db, Metrics::TYPE_GAUGE);
1216                 global_metrics.add("bus_compressor_attenuation_db", metrics.labels, &metrics.compressor_attenuation_db, Metrics::TYPE_GAUGE);
1217         }
1218
1219         // Reset resamplers for all cards that don't have the exact same state as before.
1220         map<DeviceSpec, double> new_extra_delay_ms = new_input_mapping.extra_delay_ms;  // Convenience so we can use [].
1221         for (unsigned card_index = 0; card_index < MAX_VIDEO_CARDS; ++card_index) {
1222                 const DeviceSpec device_spec{InputSourceType::CAPTURE_CARD, card_index};
1223                 AudioDevice *device = find_audio_device(device_spec);
1224                 if (device->interesting_channels != interesting_channels[device_spec] ||
1225                     device->extra_delay_ms != new_extra_delay_ms[device_spec]) {
1226                         device->interesting_channels = interesting_channels[device_spec];
1227                         device->extra_delay_ms = new_extra_delay_ms[device_spec];
1228                         reset_resampler_mutex_held(device_spec);
1229                 }
1230         }
1231         for (unsigned card_index = 0; card_index < MAX_ALSA_CARDS; ++card_index) {
1232                 const DeviceSpec device_spec{InputSourceType::ALSA_INPUT, card_index};
1233                 AudioDevice *device = find_audio_device(device_spec);
1234                 if (interesting_channels[device_spec].empty()) {
1235                         alsa_pool.release_device(card_index);
1236                 } else {
1237                         alsa_pool.hold_device(card_index);
1238                 }
1239                 if (device->interesting_channels != interesting_channels[device_spec] ||
1240                     device->extra_delay_ms != new_extra_delay_ms[device_spec]) {
1241                         device->interesting_channels = interesting_channels[device_spec];
1242                         device->extra_delay_ms = new_extra_delay_ms[device_spec];
1243                         alsa_pool.reset_device(device_spec.index);
1244                         reset_resampler_mutex_held(device_spec);
1245                 }
1246         }
1247         for (unsigned card_index = 0; card_index < num_ffmpeg_inputs; ++card_index) {
1248                 const DeviceSpec device_spec{InputSourceType::FFMPEG_VIDEO_INPUT, card_index};
1249                 AudioDevice *device = find_audio_device(device_spec);
1250                 if (device->interesting_channels != interesting_channels[device_spec] ||
1251                     device->extra_delay_ms != new_extra_delay_ms[device_spec]) {
1252                         device->interesting_channels = interesting_channels[device_spec];
1253                         device->extra_delay_ms = new_extra_delay_ms[device_spec];
1254                         reset_resampler_mutex_held(device_spec);
1255                 }
1256         }
1257
1258         input_mapping = new_input_mapping;
1259 }
1260
1261 InputMapping AudioMixer::get_input_mapping() const
1262 {
1263         lock_guard<timed_mutex> lock(audio_mutex);
1264         return input_mapping;
1265 }
1266
1267 unsigned AudioMixer::num_buses() const
1268 {
1269         lock_guard<timed_mutex> lock(audio_mutex);
1270         return input_mapping.buses.size();
1271 }
1272
1273 void AudioMixer::reset_peak(unsigned bus_index)
1274 {
1275         lock_guard<timed_mutex> lock(audio_mutex);
1276         for (unsigned channel = 0; channel < 2; ++channel) {
1277                 PeakHistory &history = peak_history[bus_index][channel];
1278                 history.current_level = 0.0f;
1279                 history.historic_peak = 0.0f;
1280                 history.current_peak = 0.0f;
1281                 history.last_peak = 0.0f;
1282                 history.age_seconds = 0.0f;
1283         }
1284 }
1285
1286 bool AudioMixer::is_mono(unsigned bus_index)
1287 {
1288         lock_guard<timed_mutex> lock(audio_mutex);
1289         const InputMapping::Bus &bus = input_mapping.buses[bus_index];
1290         if (bus.device.type == InputSourceType::SILENCE) {
1291                 return true;
1292         } else {
1293                 assert(bus.device.type == InputSourceType::CAPTURE_CARD ||
1294                        bus.device.type == InputSourceType::ALSA_INPUT ||
1295                        bus.device.type == InputSourceType::FFMPEG_VIDEO_INPUT);
1296                 return bus.source_channel[0] == bus.source_channel[1];
1297         }
1298 }
1299
1300 AudioMixer *global_audio_mixer = nullptr;