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