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 "delay_analyzer.h"
  23 #include "flags.h"
  24 #include "shared/metrics.h"
  25 #include "state.pb.h"
  26 #include "shared/timebase.h"
  27
  28 using namespace bmusb;
  29 using namespace std;
  30 using namespace std::chrono;
  31 using namespace std::placeholders;
  32
  33 namespace {
  34
  35 // TODO: If these prove to be a bottleneck, they can be SSSE3-optimized
  36 // (usually including multiple channels at a time).
  37
  38 void convert_fixed16_to_fp32(float *dst, size_t out_channel, size_t out_num_channels,
  39                              const uint8_t *src, size_t in_channel, size_t in_num_channels,
  40                              size_t num_samples)
  41 {
  42         assert(in_channel < in_num_channels);
  43         assert(out_channel < out_num_channels);
  44         src += in_channel * 2;
  45         dst += out_channel;
  46
  47         for (size_t i = 0; i < num_samples; ++i) {
  48                 int16_t s = le16toh(*(int16_t *)src);
  49                 *dst = s * (1.0f / 32768.0f);
  50
  51                 src += 2 * in_num_channels;
  52                 dst += out_num_channels;
  53         }
  54 }
  55
  56 void convert_fixed16_to_fixed32(int32_t *dst, size_t out_channel, size_t out_num_channels,
  57                                 const uint8_t *src, size_t in_channel, size_t in_num_channels,
  58                                 size_t num_samples)
  59 {
  60         assert(in_channel < in_num_channels);
  61         assert(out_channel < out_num_channels);
  62         src += in_channel * 2;
  63         dst += out_channel;
  64
  65         for (size_t i = 0; i < num_samples; ++i) {
  66                 uint32_t s = uint32_t(uint16_t(le16toh(*(int16_t *)src))) << 16;
  67
  68                 // Keep the sign bit in place, repeat the other 15 bits as far as they go.
  69                 *dst = s | ((s & 0x7fffffff) >> 15) | ((s & 0x7fffffff) >> 30);
  70
  71                 src += 2 * in_num_channels;
  72                 dst += out_num_channels;
  73         }
  74 }
  75
  76 void convert_fixed24_to_fp32(float *dst, size_t out_channel, size_t out_num_channels,
  77                              const uint8_t *src, size_t in_channel, size_t in_num_channels,
  78                              size_t num_samples)
  79 {
  80         assert(in_channel < in_num_channels);
  81         assert(out_channel < out_num_channels);
  82         src += in_channel * 3;
  83         dst += out_channel;
  84
  85         for (size_t i = 0; i < num_samples; ++i) {
  86                 uint32_t s1 = src[0];
  87                 uint32_t s2 = src[1];
  88                 uint32_t s3 = src[2];
  89                 uint32_t s = (s1 << 8) | (s2 << 16) | (s3 << 24);  // Note: The bottom eight bits are zero; s3 includes the sign bit.
  90                 *dst = int(s) * (1.0f / (256.0f * 8388608.0f));  // 256 for signed down-shift by 8, then 2^23 for the actual conversion.
  91
  92                 src += 3 * in_num_channels;
  93                 dst += out_num_channels;
  94         }
  95 }
  96
  97 void convert_fixed24_to_fixed32(int32_t *dst, size_t out_channel, size_t out_num_channels,
  98                                 const uint8_t *src, size_t in_channel, size_t in_num_channels,
  99                                 size_t num_samples)
 100 {
 101         assert(in_channel < in_num_channels);
 102         assert(out_channel < out_num_channels);
 103         src += in_channel * 3;
 104         dst += out_channel;
 105
 106         for (size_t i = 0; i < num_samples; ++i) {
 107                 uint32_t s1 = src[0];
 108                 uint32_t s2 = src[1];
 109                 uint32_t s3 = src[2];
 110                 uint32_t s = (s1 << 8) | (s2 << 16) | (s3 << 24);
 111
 112                 // Keep the sign bit in place, repeat the other 23 bits as far as they go.
 113                 *dst = s | ((s & 0x7fffffff) >> 23);
 114
 115                 src += 3 * in_num_channels;
 116                 dst += out_num_channels;
 117         }
 118 }
 119
 120 void convert_fixed32_to_fp32(float *dst, size_t out_channel, size_t out_num_channels,
 121                              const uint8_t *src, size_t in_channel, size_t in_num_channels,
 122                              size_t num_samples)
 123 {
 124         assert(in_channel < in_num_channels);
 125         assert(out_channel < out_num_channels);
 126         src += in_channel * 4;
 127         dst += out_channel;
 128
 129         for (size_t i = 0; i < num_samples; ++i) {
 130                 int32_t s = le32toh(*(int32_t *)src);
 131                 *dst = s * (1.0f / 2147483648.0f);
 132
 133                 src += 4 * in_num_channels;
 134                 dst += out_num_channels;
 135         }
 136 }
 137
 138 // Basically just a reinterleave.
 139 void convert_fixed32_to_fixed32(int32_t *dst, size_t out_channel, size_t out_num_channels,
 140                                 const uint8_t *src, size_t in_channel, size_t in_num_channels,
 141                                 size_t num_samples)
 142 {
 143         assert(in_channel < in_num_channels);
 144         assert(out_channel < out_num_channels);
 145         src += in_channel * 4;
 146         dst += out_channel;
 147
 148         for (size_t i = 0; i < num_samples; ++i) {
 149                 int32_t s = le32toh(*(int32_t *)src);
 150                 *dst = s;
 151
 152                 src += 4 * in_num_channels;
 153                 dst += out_num_channels;
 154         }
 155 }
 156
 157 float find_peak_plain(const float *samples, size_t num_samples) __attribute__((unused));
 158
 159 float find_peak_plain(const float *samples, size_t num_samples)
 160 {
 161         float m = fabs(samples[0]);
 162         for (size_t i = 1; i < num_samples; ++i) {
 163                 m = max(m, fabs(samples[i]));
 164         }
 165         return m;
 166 }
 167
 168 #ifdef __SSE__
 169 static inline float horizontal_max(__m128 m)
 170 {
 171         __m128 tmp = _mm_shuffle_ps(m, m, _MM_SHUFFLE(1, 0, 3, 2));
 172         m = _mm_max_ps(m, tmp);
 173         tmp = _mm_shuffle_ps(m, m, _MM_SHUFFLE(2, 3, 0, 1));
 174         m = _mm_max_ps(m, tmp);
 175         return _mm_cvtss_f32(m);
 176 }
 177
 178 float find_peak(const float *samples, size_t num_samples)
 179 {
 180         const __m128 abs_mask = _mm_castsi128_ps(_mm_set1_epi32(0x7fffffffu));
 181         __m128 m = _mm_setzero_ps();
 182         for (size_t i = 0; i < (num_samples & ~3); i += 4) {
 183                 __m128 x = _mm_loadu_ps(samples + i);
 184                 x = _mm_and_ps(x, abs_mask);
 185                 m = _mm_max_ps(m, x);
 186         }
 187         float result = horizontal_max(m);
 188
 189         for (size_t i = (num_samples & ~3); i < num_samples; ++i) {
 190                 result = max(result, fabs(samples[i]));
 191         }
 192
 193 #if 0
 194         // Self-test. We should be bit-exact the same.
 195         float reference_result = find_peak_plain(samples, num_samples);
 196         if (result != reference_result) {
 197                 fprintf(stderr, "Error: Peak is %f [%f %f %f %f]; should be %f.\n",
 198                         result,
 199                         _mm_cvtss_f32(_mm_shuffle_ps(m, m, _MM_SHUFFLE(0, 0, 0, 0))),
 200                         _mm_cvtss_f32(_mm_shuffle_ps(m, m, _MM_SHUFFLE(1, 1, 1, 1))),
 201                         _mm_cvtss_f32(_mm_shuffle_ps(m, m, _MM_SHUFFLE(2, 2, 2, 2))),
 202                         _mm_cvtss_f32(_mm_shuffle_ps(m, m, _MM_SHUFFLE(3, 3, 3, 3))),
 203                         reference_result);
 204                 abort();
 205         }
 206 #endif
 207         return result;
 208 }
 209 #else
 210 float find_peak(const float *samples, size_t num_samples)
 211 {
 212         return find_peak_plain(samples, num_samples);
 213 }
 214 #endif
 215
 216 void deinterleave_samples(const vector<float> &in, vector<float> *out_l, vector<float> *out_r)
 217 {
 218         size_t num_samples = in.size() / 2;
 219         out_l->resize(num_samples);
 220         out_r->resize(num_samples);
 221
 222         const float *inptr = in.data();
 223         float *lptr = &(*out_l)[0];
 224         float *rptr = &(*out_r)[0];
 225         for (size_t i = 0; i < num_samples; ++i) {
 226                 *lptr++ = *inptr++;
 227                 *rptr++ = *inptr++;
 228         }
 229 }
 230
 231 }  // namespace
 232
 233 AudioMixer::AudioMixer(unsigned num_capture_cards, unsigned num_ffmpeg_inputs)
 234         : num_capture_cards(num_capture_cards),
 235           num_ffmpeg_inputs(num_ffmpeg_inputs),
 236           ffmpeg_inputs(new AudioDevice[num_ffmpeg_inputs]),
 237           limiter(OUTPUT_FREQUENCY),
 238           correlation(OUTPUT_FREQUENCY)
 239 {
 240         for (unsigned bus_index = 0; bus_index < MAX_BUSES; ++bus_index) {
 241                 locut[bus_index].init(FILTER_HPF, 2);
 242                 eq[bus_index][EQ_BAND_BASS].init(FILTER_LOW_SHELF, 1);
 243                 // Note: EQ_BAND_MID isn't used (see comments in apply_eq()).
 244                 eq[bus_index][EQ_BAND_TREBLE].init(FILTER_HIGH_SHELF, 1);
 245                 compressor[bus_index].reset(new StereoCompressor(OUTPUT_FREQUENCY));
 246                 level_compressor[bus_index].reset(new StereoCompressor(OUTPUT_FREQUENCY));
 247
 248                 set_bus_settings(bus_index, get_default_bus_settings());
 249         }
 250         set_limiter_enabled(global_flags.limiter_enabled);
 251         set_final_makeup_gain_auto(global_flags.final_makeup_gain_auto);
 252
 253         r128.init(2, OUTPUT_FREQUENCY);
 254         r128.integr_start();
 255
 256         // hlen=16 is pretty low quality, but we use quite a bit of CPU otherwise,
 257         // and there's a limit to how important the peak meter is.
 258         peak_resampler.setup(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY * 4, /*num_channels=*/2, /*hlen=*/16, /*frel=*/1.0);
 259
 260         global_audio_mixer = this;
 261         alsa_pool.init();
 262
 263         if (!global_flags.input_mapping_filename.empty()) {
 264                 // Must happen after ALSAPool is initialized, as it needs to know the card list.
 265                 current_mapping_mode = MappingMode::MULTICHANNEL;
 266                 InputMapping new_input_mapping;
 267                 if (!load_input_mapping_from_file(get_devices(HOLD_ALSA_DEVICES),
 268                                                   global_flags.input_mapping_filename,
 269                                                   &new_input_mapping)) {
 270                         fprintf(stderr, "Failed to load input mapping from '%s', exiting.\n",
 271                                 global_flags.input_mapping_filename.c_str());
 272                         abort();
 273                 }
 274                 set_input_mapping(new_input_mapping);
 275         } else {
 276                 set_simple_input(/*card_index=*/0);
 277                 if (global_flags.multichannel_mapping_mode) {
 278                         current_mapping_mode = MappingMode::MULTICHANNEL;
 279                 }
 280         }
 281
 282         global_metrics.add("audio_loudness_short_lufs", &metric_audio_loudness_short_lufs, Metrics::TYPE_GAUGE);
 283         global_metrics.add("audio_loudness_integrated_lufs", &metric_audio_loudness_integrated_lufs, Metrics::TYPE_GAUGE);
 284         global_metrics.add("audio_loudness_range_low_lufs", &metric_audio_loudness_range_low_lufs, Metrics::TYPE_GAUGE);
 285         global_metrics.add("audio_loudness_range_high_lufs", &metric_audio_loudness_range_high_lufs, Metrics::TYPE_GAUGE);
 286         global_metrics.add("audio_peak_dbfs", &metric_audio_peak_dbfs, Metrics::TYPE_GAUGE);
 287         global_metrics.add("audio_final_makeup_gain_db", &metric_audio_final_makeup_gain_db, Metrics::TYPE_GAUGE);
 288         global_metrics.add("audio_correlation", &metric_audio_correlation, Metrics::TYPE_GAUGE);
 289 }
 290
 291 void AudioMixer::reset_resampler(DeviceSpec device_spec)
 292 {
 293         lock_guard<timed_mutex> lock(audio_mutex);
 294         reset_resampler_mutex_held(device_spec);
 295 }
 296
 297 void AudioMixer::reset_resampler_mutex_held(DeviceSpec device_spec)
 298 {
 299         AudioDevice *device = find_audio_device(device_spec);
 300
 301         if (device->interesting_channels.empty()) {
 302                 device->resampling_queue.reset();
 303         } else {
 304                 // Make sure we never get negative delay. Even 1 ms is probably way less than we
 305                 // could ever hope to actually have; this is just a failsafe.
 306                 double delay_ms = max(global_flags.audio_queue_length_ms + device->extra_delay_ms, 1.0);
 307
 308                 device->resampling_queue.reset(new ResamplingQueue(
 309                         device_spec, device->capture_frequency, OUTPUT_FREQUENCY, device->interesting_channels.size(),
 310                         delay_ms * 0.001));
 311         }
 312 }
 313
 314 bool AudioMixer::add_audio(DeviceSpec device_spec, const uint8_t *data, unsigned num_samples, AudioFormat audio_format, steady_clock::time_point frame_time)
 315 {
 316         if (delay_analyzer != nullptr && delay_analyzer->is_grabbing()) {
 317                 delay_analyzer->add_audio(device_spec, data, num_samples, audio_format, frame_time);
 318         }
 319
 320         AudioDevice *device = find_audio_device(device_spec);
 321
 322         unique_lock<timed_mutex> lock(audio_mutex, defer_lock);
 323         if (!lock.try_lock_for(chrono::milliseconds(10))) {
 324                 return false;
 325         }
 326         if (device->resampling_queue == nullptr) {
 327                 // No buses use this device; throw it away.
 328                 return true;
 329         }
 330
 331         unsigned num_channels = device->interesting_channels.size();
 332         assert(num_channels > 0);
 333
 334         // Convert the audio to fp32.
 335         unique_ptr<float[]> audio(new float[num_samples * num_channels]);
 336         unsigned channel_index = 0;
 337         for (auto channel_it = device->interesting_channels.cbegin(); channel_it != device->interesting_channels.end(); ++channel_it, ++channel_index) {
 338                 convert_audio_to_fp32(audio.get(), channel_index, num_channels, data, *channel_it, audio_format, num_samples);
 339         }
 340
 341         // If we changed frequency since last frame, we'll need to reset the resampler.
 342         if (audio_format.sample_rate != device->capture_frequency) {
 343                 device->capture_frequency = audio_format.sample_rate;
 344                 reset_resampler_mutex_held(device_spec);
 345         }
 346
 347         // Now add it.
 348         device->resampling_queue->add_input_samples(frame_time, audio.get(), num_samples, ResamplingQueue::ADJUST_RATE);
 349         return true;
 350 }
 351
 352 // Converts all channels.
 353 vector<int32_t> convert_audio_to_fixed32(const uint8_t *data, unsigned num_samples, bmusb::AudioFormat audio_format, unsigned num_channels)
 354 {
 355         vector<int32_t> audio;
 356
 357         if (num_channels > audio_format.num_channels) {
 358                 audio.resize(num_samples * num_channels, 0);
 359         } else {
 360                 audio.resize(num_samples * num_channels);
 361         }
 362         for (unsigned channel_index = 0; channel_index < num_channels && channel_index < audio_format.num_channels; ++channel_index) {
 363                 switch (audio_format.bits_per_sample) {
 364                 case 0:
 365                         assert(num_samples == 0);
 366                         break;
 367                 case 16:
 368                         convert_fixed16_to_fixed32(&audio[0], channel_index, num_channels, data, channel_index, audio_format.num_channels, num_samples);
 369                         break;
 370                 case 24:
 371                         convert_fixed24_to_fixed32(&audio[0], channel_index, num_channels, data, channel_index, audio_format.num_channels, num_samples);
 372                         break;
 373                 case 32:
 374                         convert_fixed32_to_fixed32(&audio[0], channel_index, num_channels, data, channel_index, audio_format.num_channels, num_samples);
 375                         break;
 376                 default:
 377                         fprintf(stderr, "Cannot handle audio with %u bits per sample\n", audio_format.bits_per_sample);
 378                         assert(false);
 379                 }
 380         }
 381
 382         return audio;
 383 }
 384
 385 // Converts only one channel.
 386 void convert_audio_to_fp32(float *dst, size_t out_channel, size_t out_num_channels,
 387                            const uint8_t *src, size_t in_channel, bmusb::AudioFormat in_audio_format,
 388                            size_t num_samples)
 389 {
 390         switch (in_audio_format.bits_per_sample) {
 391         case 0:
 392                 assert(num_samples == 0);
 393                 break;
 394         case 16:
 395                 convert_fixed16_to_fp32(dst, out_channel, out_num_channels, src, in_channel, in_audio_format.num_channels, num_samples);
 396                 break;
 397         case 24:
 398                 convert_fixed24_to_fp32(dst, out_channel, out_num_channels, src, in_channel, in_audio_format.num_channels, num_samples);
 399                 break;
 400         case 32:
 401                 convert_fixed32_to_fp32(dst, out_channel, out_num_channels, src, in_channel, in_audio_format.num_channels, num_samples);
 402                 break;
 403         default:
 404                 fprintf(stderr, "Cannot handle audio with %u bits per sample\n", in_audio_format.bits_per_sample);
 405                 assert(false);
 406         }
 407 }
 408
 409 bool AudioMixer::add_silence(DeviceSpec device_spec, unsigned samples_per_frame, unsigned num_frames)
 410 {
 411         AudioDevice *device = find_audio_device(device_spec);
 412
 413         unique_lock<timed_mutex> lock(audio_mutex, defer_lock);
 414         if (!lock.try_lock_for(chrono::milliseconds(10))) {
 415                 return false;
 416         }
 417         if (device->resampling_queue == nullptr) {
 418                 // No buses use this device; throw it away.
 419                 return true;
 420         }
 421
 422         unsigned num_channels = device->interesting_channels.size();
 423         assert(num_channels > 0);
 424
 425         vector<float> silence(samples_per_frame * num_channels, 0.0f);
 426         for (unsigned i = 0; i < num_frames; ++i) {
 427                 device->resampling_queue->add_input_samples(steady_clock::now(), silence.data(), samples_per_frame, ResamplingQueue::DO_NOT_ADJUST_RATE);
 428         }
 429         return true;
 430 }
 431
 432 bool AudioMixer::silence_card(DeviceSpec device_spec, bool silence)
 433 {
 434         AudioDevice *device = find_audio_device(device_spec);
 435
 436         unique_lock<timed_mutex> lock(audio_mutex, defer_lock);
 437         if (!lock.try_lock_for(chrono::milliseconds(10))) {
 438                 return false;
 439         }
 440
 441         if (device->silenced && !silence) {
 442                 reset_resampler_mutex_held(device_spec);
 443         }
 444         device->silenced = silence;
 445         return true;
 446 }
 447
 448 AudioMixer::BusSettings AudioMixer::get_default_bus_settings()
 449 {
 450         BusSettings settings;
 451         settings.fader_volume_db = 0.0f;
 452         settings.muted = false;
 453         settings.locut_enabled = global_flags.locut_enabled;
 454         settings.stereo_width = 1.0f;
 455         for (unsigned band_index = 0; band_index < NUM_EQ_BANDS; ++band_index) {
 456                 settings.eq_level_db[band_index] = 0.0f;
 457         }
 458         settings.gain_staging_db = global_flags.initial_gain_staging_db;
 459         settings.level_compressor_enabled = global_flags.gain_staging_auto;
 460         settings.compressor_threshold_dbfs = ref_level_dbfs - 12.0f;  // -12 dB.
 461         settings.compressor_enabled = global_flags.compressor_enabled;
 462         return settings;
 463 }
 464
 465 AudioMixer::BusSettings AudioMixer::get_bus_settings(unsigned bus_index) const
 466 {
 467         lock_guard<timed_mutex> lock(audio_mutex);
 468         BusSettings settings;
 469         settings.fader_volume_db = fader_volume_db[bus_index];
 470         settings.muted = mute[bus_index];
 471         settings.locut_enabled = locut_enabled[bus_index];
 472         settings.stereo_width = stereo_width[bus_index];
 473         for (unsigned band_index = 0; band_index < NUM_EQ_BANDS; ++band_index) {
 474                 settings.eq_level_db[band_index] = eq_level_db[bus_index][band_index];
 475         }
 476         settings.gain_staging_db = gain_staging_db[bus_index];
 477         settings.level_compressor_enabled = level_compressor_enabled[bus_index];
 478         settings.compressor_threshold_dbfs = compressor_threshold_dbfs[bus_index];
 479         settings.compressor_enabled = compressor_enabled[bus_index];
 480         return settings;
 481 }
 482
 483 void AudioMixer::set_bus_settings(unsigned bus_index, const AudioMixer::BusSettings &settings)
 484 {
 485         lock_guard<timed_mutex> lock(audio_mutex);
 486         fader_volume_db[bus_index] = settings.fader_volume_db;
 487         mute[bus_index] = settings.muted;
 488         locut_enabled[bus_index] = settings.locut_enabled;
 489         stereo_width[bus_index] = settings.stereo_width;
 490         for (unsigned band_index = 0; band_index < NUM_EQ_BANDS; ++band_index) {
 491                 eq_level_db[bus_index][band_index] = settings.eq_level_db[band_index];
 492         }
 493         gain_staging_db[bus_index] = settings.gain_staging_db;
 494         last_gain_staging_db[bus_index] = gain_staging_db[bus_index];
 495         level_compressor_enabled[bus_index] = settings.level_compressor_enabled;
 496         compressor_threshold_dbfs[bus_index] = settings.compressor_threshold_dbfs;
 497         compressor_enabled[bus_index] = settings.compressor_enabled;
 498 }
 499
 500 AudioMixer::AudioDevice *AudioMixer::find_audio_device(DeviceSpec device)
 501 {
 502         switch (device.type) {
 503         case InputSourceType::CAPTURE_CARD:
 504                 return &video_cards[device.index];
 505         case InputSourceType::ALSA_INPUT:
 506                 return &alsa_inputs[device.index];
 507         case InputSourceType::FFMPEG_VIDEO_INPUT:
 508                 return &ffmpeg_inputs[device.index];
 509         case InputSourceType::SILENCE:
 510         default:
 511                 assert(false);
 512         }
 513         return nullptr;
 514 }
 515
 516 // Get a pointer to the given channel from the given device.
 517 // The channel must be picked out earlier and resampled.
 518 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)
 519 {
 520         static float zero = 0.0f;
 521         if (source_channel == -1 || device_spec.type == InputSourceType::SILENCE) {
 522                 *srcptr = &zero;
 523                 *stride = 0;
 524                 return;
 525         }
 526         AudioDevice *device = find_audio_device(device_spec);
 527         assert(device->interesting_channels.count(source_channel) != 0);
 528         unsigned channel_index = 0;
 529         for (int channel : device->interesting_channels) {
 530                 if (channel == source_channel) break;
 531                 ++channel_index;
 532         }
 533         assert(channel_index < device->interesting_channels.size());
 534         const auto it = samples_card.find(device_spec);
 535         assert(it != samples_card.end());
 536         *srcptr = &(it->second)[channel_index];
 537         *stride = device->interesting_channels.size();
 538 }
 539
 540 // TODO: Can be SSSE3-optimized if need be.
 541 void AudioMixer::fill_audio_bus(const map<DeviceSpec, vector<float>> &samples_card, const InputMapping::Bus &bus, unsigned num_samples, float stereo_width, float *output)
 542 {
 543         if (bus.device.type == InputSourceType::SILENCE) {
 544                 memset(output, 0, num_samples * 2 * sizeof(*output));
 545         } else {
 546                 assert(bus.device.type == InputSourceType::CAPTURE_CARD ||
 547                        bus.device.type == InputSourceType::ALSA_INPUT ||
 548                        bus.device.type == InputSourceType::FFMPEG_VIDEO_INPUT);
 549                 const float *lsrc, *rsrc;
 550                 unsigned lstride, rstride;
 551                 float *dptr = output;
 552                 find_sample_src_from_device(samples_card, bus.device, bus.source_channel[0], &lsrc, &lstride);
 553                 find_sample_src_from_device(samples_card, bus.device, bus.source_channel[1], &rsrc, &rstride);
 554
 555                 // Apply stereo width settings. Set stereo width w to a 0..1 range instead of
 556                 // -1..1, since it makes for much easier calculations (so 0.5 = completely mono).
 557                 // Then, what we want is
 558                 //
 559                 //   L' = wL + (1-w)R = R + w(L-R)
 560                 //   R' = wR + (1-w)L = L + w(R-L)
 561                 //
 562                 // This can be further simplified calculation-wise by defining the weighted
 563                 // difference signal D = w(R-L), so that:
 564                 //
 565                 //   L' = R - D
 566                 //   R' = L + D
 567                 float w = 0.5f * stereo_width + 0.5f;
 568                 if (bus.source_channel[0] == bus.source_channel[1]) {
 569                         // Mono anyway, so no need to bother.
 570                         w = 1.0f;
 571                 } else if (fabs(w) < 1e-3) {
 572                         // Perfect inverse.
 573                         swap(lsrc, rsrc);
 574                         swap(lstride, rstride);
 575                         w = 1.0f;
 576                 }
 577                 if (fabs(w - 1.0f) < 1e-3) {
 578                         // No calculations needed for stereo_width = 1.
 579                         for (unsigned i = 0; i < num_samples; ++i) {
 580                                 *dptr++ = *lsrc;
 581                                 *dptr++ = *rsrc;
 582                                 lsrc += lstride;
 583                                 rsrc += rstride;
 584                         }
 585                 } else {
 586                         // General case.
 587                         for (unsigned i = 0; i < num_samples; ++i) {
 588                                 float left = *lsrc, right = *rsrc;
 589                                 float diff = w * (right - left);
 590                                 *dptr++ = right - diff;
 591                                 *dptr++ = left + diff;
 592                                 lsrc += lstride;
 593                                 rsrc += rstride;
 594                         }
 595                 }
 596         }
 597 }
 598
 599 vector<DeviceSpec> AudioMixer::get_active_devices() const
 600 {
 601         vector<DeviceSpec> ret;
 602         for (unsigned card_index = 0; card_index < MAX_VIDEO_CARDS; ++card_index) {
 603                 const DeviceSpec device_spec{InputSourceType::CAPTURE_CARD, card_index};
 604                 if (!find_audio_device(device_spec)->interesting_channels.empty()) {
 605                         ret.push_back(device_spec);
 606                 }
 607         }
 608         for (unsigned card_index = 0; card_index < MAX_ALSA_CARDS; ++card_index) {
 609                 const DeviceSpec device_spec{InputSourceType::ALSA_INPUT, card_index};
 610                 if (!find_audio_device(device_spec)->interesting_channels.empty()) {
 611                         ret.push_back(device_spec);
 612                 }
 613         }
 614         for (unsigned card_index = 0; card_index < num_ffmpeg_inputs; ++card_index) {
 615                 const DeviceSpec device_spec{InputSourceType::FFMPEG_VIDEO_INPUT, card_index};
 616                 if (!find_audio_device(device_spec)->interesting_channels.empty()) {
 617                         ret.push_back(device_spec);
 618                 }
 619         }
 620         return ret;
 621 }
 622
 623 namespace {
 624
 625 void apply_gain(float db, float last_db, vector<float> *samples)
 626 {
 627         if (fabs(db - last_db) < 1e-3) {
 628                 // Constant over this frame.
 629                 const float gain = from_db(db);
 630                 for (size_t i = 0; i < samples->size(); ++i) {
 631                         (*samples)[i] *= gain;
 632                 }
 633         } else {
 634                 // We need to do a fade.
 635                 unsigned num_samples = samples->size() / 2;
 636                 float gain = from_db(last_db);
 637                 const float gain_inc = pow(from_db(db - last_db), 1.0 / num_samples);
 638                 for (size_t i = 0; i < num_samples; ++i) {
 639                         (*samples)[i * 2 + 0] *= gain;
 640                         (*samples)[i * 2 + 1] *= gain;
 641                         gain *= gain_inc;
 642                 }
 643         }
 644 }
 645
 646 }  // namespace
 647
 648 vector<float> AudioMixer::get_output(steady_clock::time_point ts, unsigned num_samples, ResamplingQueue::RateAdjustmentPolicy rate_adjustment_policy)
 649 {
 650         map<DeviceSpec, vector<float>> samples_card;
 651         vector<float> samples_bus;
 652
 653         lock_guard<timed_mutex> lock(audio_mutex);
 654
 655         // Pick out all the interesting channels from all the cards.
 656         for (const DeviceSpec &device_spec : get_active_devices()) {
 657                 AudioDevice *device = find_audio_device(device_spec);
 658                 samples_card[device_spec].resize(num_samples * device->interesting_channels.size());
 659                 if (device->silenced) {
 660                         memset(&samples_card[device_spec][0], 0, samples_card[device_spec].size() * sizeof(float));
 661                 } else {
 662                         device->resampling_queue->get_output_samples(
 663                                 ts,
 664                                 &samples_card[device_spec][0],
 665                                 num_samples,
 666                                 rate_adjustment_policy);
 667                 }
 668         }
 669
 670         vector<float> samples_out, left, right;
 671         samples_out.resize(num_samples * 2);
 672         samples_bus.resize(num_samples * 2);
 673         for (unsigned bus_index = 0; bus_index < input_mapping.buses.size(); ++bus_index) {
 674                 fill_audio_bus(samples_card, input_mapping.buses[bus_index], num_samples, stereo_width[bus_index], &samples_bus[0]);
 675                 apply_eq(bus_index, &samples_bus);
 676
 677                 {
 678                         lock_guard<mutex> lock(compressor_mutex);
 679
 680                         // Apply a level compressor to get the general level right.
 681                         // Basically, if it's over about -40 dBFS, we squeeze it down to that level
 682                         // (or more precisely, near it, since we don't use infinite ratio),
 683                         // then apply a makeup gain to get it to -14 dBFS. -14 dBFS is, of course,
 684                         // entirely arbitrary, but from practical tests with speech, it seems to
 685                         // put ut around -23 LUFS, so it's a reasonable starting point for later use.
 686                         if (level_compressor_enabled[bus_index]) {
 687                                 float threshold = 0.01f;   // -40 dBFS.
 688                                 float ratio = 20.0f;
 689                                 float attack_time = 0.5f;
 690                                 float release_time = 20.0f;
 691                                 float makeup_gain = from_db(ref_level_dbfs - (-40.0f));  // +26 dB.
 692                                 level_compressor[bus_index]->process(samples_bus.data(), samples_bus.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 693                                 gain_staging_db[bus_index] = to_db(level_compressor[bus_index]->get_attenuation() * makeup_gain);
 694                         } else {
 695                                 // Just apply the gain we already had.
 696                                 float db = gain_staging_db[bus_index];
 697                                 float last_db = last_gain_staging_db[bus_index];
 698                                 apply_gain(db, last_db, &samples_bus);
 699                         }
 700                         last_gain_staging_db[bus_index] = gain_staging_db[bus_index];
 701
 702 #if 0
 703                         printf("level=%f (%+5.2f dBFS) attenuation=%f (%+5.2f dB) end_result=%+5.2f dB\n",
 704                                 level_compressor.get_level(), to_db(level_compressor.get_level()),
 705                                 level_compressor.get_attenuation(), to_db(level_compressor.get_attenuation()),
 706                                 to_db(level_compressor.get_level() * level_compressor.get_attenuation() * makeup_gain));
 707 #endif
 708
 709                         // The real compressor.
 710                         if (compressor_enabled[bus_index]) {
 711                                 float threshold = from_db(compressor_threshold_dbfs[bus_index]);
 712                                 float ratio = 20.0f;
 713                                 float attack_time = 0.005f;
 714                                 float release_time = 0.040f;
 715                                 float makeup_gain = 2.0f;  // +6 dB.
 716                                 compressor[bus_index]->process(samples_bus.data(), samples_bus.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 717                 //              compressor_att = compressor.get_attenuation();
 718                         }
 719                 }
 720
 721                 add_bus_to_master(bus_index, samples_bus, &samples_out);
 722                 deinterleave_samples(samples_bus, &left, &right);
 723                 measure_bus_levels(bus_index, left, right);
 724         }
 725
 726         {
 727                 lock_guard<mutex> lock(compressor_mutex);
 728
 729                 // Finally a limiter at -4 dB (so, -10 dBFS) to take out the worst peaks only.
 730                 // Note that since ratio is not infinite, we could go slightly higher than this.
 731                 if (limiter_enabled) {
 732                         float threshold = from_db(limiter_threshold_dbfs);
 733                         float ratio = 30.0f;
 734                         float attack_time = 0.0f;  // Instant.
 735                         float release_time = 0.020f;
 736                         float makeup_gain = 1.0f;  // 0 dB.
 737                         limiter.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 738         //              limiter_att = limiter.get_attenuation();
 739                 }
 740
 741         //      printf("limiter=%+5.1f  compressor=%+5.1f\n", to_db(limiter_att), to_db(compressor_att));
 742         }
 743
 744         // At this point, we are most likely close to +0 LU (at least if the
 745         // faders sum to 0 dB and the compressors are on), but all of our
 746         // measurements have been on raw sample values, not R128 values.
 747         // So we have a final makeup gain to get us to +0 LU; the gain
 748         // adjustments required should be relatively small, and also, the
 749         // offset shouldn't change much (only if the type of audio changes
 750         // significantly). Thus, we shoot for updating this value basically
 751         // “whenever we process buffers”, since the R128 calculation isn't exactly
 752         // something we get out per-sample.
 753         //
 754         // Note that there's a feedback loop here, so we choose a very slow filter
 755         // (half-time of 30 seconds).
 756         double target_loudness_factor, alpha;
 757         double loudness_lu = r128.loudness_M() - ref_level_lufs;
 758         target_loudness_factor = final_makeup_gain * from_db(-loudness_lu);
 759
 760         // If we're outside +/- 5 LU (after correction), we don't count it as
 761         // a normal signal (probably silence) and don't change the
 762         // correction factor; just apply what we already have.
 763         if (fabs(loudness_lu) >= 5.0 || !final_makeup_gain_auto) {
 764                 alpha = 0.0;
 765         } else {
 766                 // Formula adapted from
 767                 // https://en.wikipedia.org/wiki/Low-pass_filter#Simple_infinite_impulse_response_filter.
 768                 const double half_time_s = 30.0;
 769                 const double fc_mul_2pi_delta_t = 1.0 / (half_time_s * OUTPUT_FREQUENCY);
 770                 alpha = fc_mul_2pi_delta_t / (fc_mul_2pi_delta_t + 1.0);
 771         }
 772
 773         {
 774                 lock_guard<mutex> lock(compressor_mutex);
 775                 double m = final_makeup_gain;
 776                 for (size_t i = 0; i < samples_out.size(); i += 2) {
 777                         samples_out[i + 0] *= m;
 778                         samples_out[i + 1] *= m;
 779                         m += (target_loudness_factor - m) * alpha;
 780                 }
 781                 final_makeup_gain = m;
 782         }
 783
 784         update_meters(samples_out);
 785
 786         return samples_out;
 787 }
 788
 789 namespace {
 790
 791 void apply_filter_fade(StereoFilter *filter, float *data, unsigned num_samples, float cutoff_hz, float db, float last_db)
 792 {
 793         // A granularity of 32 samples is an okay tradeoff between speed and
 794         // smoothness; recalculating the filters is pretty expensive, so it's
 795         // good that we don't do this all the time.
 796         static constexpr unsigned filter_granularity_samples = 32;
 797
 798         const float cutoff_linear = cutoff_hz * 2.0 * M_PI / OUTPUT_FREQUENCY;
 799         if (fabs(db - last_db) < 1e-3) {
 800                 // Constant over this frame.
 801                 if (fabs(db) > 0.01f) {
 802                         filter->render(data, num_samples, cutoff_linear, 0.5f, db / 40.0f);
 803                 }
 804         } else {
 805                 // We need to do a fade. (Rounding up avoids division by zero.)
 806                 unsigned num_blocks = (num_samples + filter_granularity_samples - 1) / filter_granularity_samples;
 807                 const float inc_db_norm = (db - last_db) / 40.0f / num_blocks;
 808                 float db_norm = db / 40.0f;
 809                 for (size_t i = 0; i < num_samples; i += filter_granularity_samples) {
 810                         size_t samples_this_block = std::min<size_t>(num_samples - i, filter_granularity_samples);
 811                         filter->render(data + i * 2, samples_this_block, cutoff_linear, 0.5f, db_norm);
 812                         db_norm += inc_db_norm;
 813                 }
 814         }
 815 }
 816
 817 }  // namespace
 818
 819 void AudioMixer::apply_eq(unsigned bus_index, vector<float> *samples_bus)
 820 {
 821         constexpr float bass_freq_hz = 200.0f;
 822         constexpr float treble_freq_hz = 4700.0f;
 823
 824         // Cut away everything under 120 Hz (or whatever the cutoff is);
 825         // we don't need it for voice, and it will reduce headroom
 826         // and confuse the compressor. (In particular, any hums at 50 or 60 Hz
 827         // should be dampened.)
 828         if (locut_enabled[bus_index]) {
 829                 locut[bus_index].render(samples_bus->data(), samples_bus->size() / 2, locut_cutoff_hz * 2.0 * M_PI / OUTPUT_FREQUENCY, 0.5f);
 830         }
 831
 832         // Apply the rest of the EQ. Since we only have a simple three-band EQ,
 833         // we can implement it with two shelf filters. We use a simple gain to
 834         // set the mid-level filter, and then offset the low and high bands
 835         // from that if we need to. (We could perhaps have folded the gain into
 836         // the next part, but it's so cheap that the trouble isn't worth it.)
 837         //
 838         // If any part of the EQ has changed appreciably since last frame,
 839         // we fade smoothly during the course of this frame.
 840         const float bass_db = eq_level_db[bus_index][EQ_BAND_BASS];
 841         const float mid_db = eq_level_db[bus_index][EQ_BAND_MID];
 842         const float treble_db = eq_level_db[bus_index][EQ_BAND_TREBLE];
 843
 844         const float last_bass_db = last_eq_level_db[bus_index][EQ_BAND_BASS];
 845         const float last_mid_db = last_eq_level_db[bus_index][EQ_BAND_MID];
 846         const float last_treble_db = last_eq_level_db[bus_index][EQ_BAND_TREBLE];
 847
 848         assert(samples_bus->size() % 2 == 0);
 849         const unsigned num_samples = samples_bus->size() / 2;
 850
 851         apply_gain(mid_db, last_mid_db, samples_bus);
 852
 853         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);
 854         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);
 855
 856         last_eq_level_db[bus_index][EQ_BAND_BASS] = bass_db;
 857         last_eq_level_db[bus_index][EQ_BAND_MID] = mid_db;
 858         last_eq_level_db[bus_index][EQ_BAND_TREBLE] = treble_db;
 859 }
 860
 861 void AudioMixer::add_bus_to_master(unsigned bus_index, const vector<float> &samples_bus, vector<float> *samples_out)
 862 {
 863         assert(samples_bus.size() == samples_out->size());
 864         assert(samples_bus.size() % 2 == 0);
 865         unsigned num_samples = samples_bus.size() / 2;
 866         const float new_volume_db = mute[bus_index] ? -90.0f : fader_volume_db[bus_index].load();
 867         if (fabs(new_volume_db - last_fader_volume_db[bus_index]) > 1e-3) {
 868                 // The volume has changed; do a fade over the course of this frame.
 869                 // (We might have some numerical issues here, but it seems to sound OK.)
 870                 // For the purpose of fading here, the silence floor is set to -90 dB
 871                 // (the fader only goes to -84).
 872                 float old_volume = from_db(max<float>(last_fader_volume_db[bus_index], -90.0f));
 873                 float volume = from_db(max<float>(new_volume_db, -90.0f));
 874
 875                 float volume_inc = pow(volume / old_volume, 1.0 / num_samples);
 876                 volume = old_volume;
 877                 if (bus_index == 0) {
 878                         for (unsigned i = 0; i < num_samples; ++i) {
 879                                 (*samples_out)[i * 2 + 0] = samples_bus[i * 2 + 0] * volume;
 880                                 (*samples_out)[i * 2 + 1] = samples_bus[i * 2 + 1] * volume;
 881                                 volume *= volume_inc;
 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                                 volume *= volume_inc;
 888                         }
 889                 }
 890         } else if (new_volume_db > -90.0f) {
 891                 float volume = from_db(new_volume_db);
 892                 if (bus_index == 0) {
 893                         for (unsigned i = 0; i < num_samples; ++i) {
 894                                 (*samples_out)[i * 2 + 0] = samples_bus[i * 2 + 0] * volume;
 895                                 (*samples_out)[i * 2 + 1] = samples_bus[i * 2 + 1] * volume;
 896                         }
 897                 } else {
 898                         for (unsigned i = 0; i < num_samples; ++i) {
 899                                 (*samples_out)[i * 2 + 0] += samples_bus[i * 2 + 0] * volume;
 900                                 (*samples_out)[i * 2 + 1] += samples_bus[i * 2 + 1] * volume;
 901                         }
 902                 }
 903         }
 904
 905         last_fader_volume_db[bus_index] = new_volume_db;
 906 }
 907
 908 void AudioMixer::measure_bus_levels(unsigned bus_index, const vector<float> &left, const vector<float> &right)
 909 {
 910         assert(left.size() == right.size());
 911         const float volume = mute[bus_index] ? 0.0f : from_db(fader_volume_db[bus_index]);
 912         const float peak_levels[2] = {
 913                 find_peak(left.data(), left.size()) * volume,
 914                 find_peak(right.data(), right.size()) * volume
 915         };
 916         for (unsigned channel = 0; channel < 2; ++channel) {
 917                 // Compute the current value, including hold and falloff.
 918                 // The constants are borrowed from zita-mu1 by Fons Adriaensen.
 919                 static constexpr float hold_sec = 0.5f;
 920                 static constexpr float falloff_db_sec = 15.0f;  // dB/sec falloff after hold.
 921                 float current_peak;
 922                 PeakHistory &history = peak_history[bus_index][channel];
 923                 history.historic_peak = max(history.historic_peak, peak_levels[channel]);
 924                 if (history.age_seconds < hold_sec) {
 925                         current_peak = history.last_peak;
 926                 } else {
 927                         current_peak = history.last_peak * from_db(-falloff_db_sec * (history.age_seconds - hold_sec));
 928                 }
 929
 930                 // See if we have a new peak to replace the old (possibly falling) one.
 931                 if (peak_levels[channel] > current_peak) {
 932                         history.last_peak = peak_levels[channel];
 933                         history.age_seconds = 0.0f;  // Not 100% correct, but more than good enough given our frame sizes.
 934                         current_peak = peak_levels[channel];
 935                 } else {
 936                         history.age_seconds += float(left.size()) / OUTPUT_FREQUENCY;
 937                 }
 938                 history.current_level = peak_levels[channel];
 939                 history.current_peak = current_peak;
 940         }
 941 }
 942
 943 void AudioMixer::update_meters(const vector<float> &samples)
 944 {
 945         // Upsample 4x to find interpolated peak.
 946         peak_resampler.inp_data = const_cast<float *>(samples.data());
 947         peak_resampler.inp_count = samples.size() / 2;
 948
 949         vector<float> interpolated_samples;
 950         interpolated_samples.resize(samples.size());
 951         {
 952                 lock_guard<mutex> lock(audio_measure_mutex);
 953
 954                 while (peak_resampler.inp_count > 0) {  // About four iterations.
 955                         peak_resampler.out_data = &interpolated_samples[0];
 956                         peak_resampler.out_count = interpolated_samples.size() / 2;
 957                         peak_resampler.process();
 958                         size_t out_stereo_samples = interpolated_samples.size() / 2 - peak_resampler.out_count;
 959                         peak = max<float>(peak, find_peak(interpolated_samples.data(), out_stereo_samples * 2));
 960                         peak_resampler.out_data = nullptr;
 961                 }
 962         }
 963
 964         // Find R128 levels and L/R correlation.
 965         vector<float> left, right;
 966         deinterleave_samples(samples, &left, &right);
 967         float *ptrs[] = { left.data(), right.data() };
 968         {
 969                 lock_guard<mutex> lock(audio_measure_mutex);
 970                 r128.process(left.size(), ptrs);
 971                 correlation.process_samples(samples);
 972         }
 973
 974         send_audio_level_callback();
 975 }
 976
 977 void AudioMixer::reset_meters()
 978 {
 979         lock_guard<mutex> lock(audio_measure_mutex);
 980         peak_resampler.reset();
 981         peak = 0.0f;
 982         r128.reset();
 983         r128.integr_start();
 984         correlation.reset();
 985 }
 986
 987 void AudioMixer::send_audio_level_callback()
 988 {
 989         if (audio_level_callback == nullptr) {
 990                 return;
 991         }
 992
 993         lock_guard<mutex> lock(audio_measure_mutex);
 994         double loudness_s = r128.loudness_S();
 995         double loudness_i = r128.integrated();
 996         double loudness_range_low = r128.range_min();
 997         double loudness_range_high = r128.range_max();
 998
 999         metric_audio_loudness_short_lufs = loudness_s;
1000         metric_audio_loudness_integrated_lufs = loudness_i;
1001         metric_audio_loudness_range_low_lufs = loudness_range_low;
1002         metric_audio_loudness_range_high_lufs = loudness_range_high;
1003         metric_audio_peak_dbfs = to_db(peak);
1004         metric_audio_final_makeup_gain_db = to_db(final_makeup_gain);
1005         metric_audio_correlation = correlation.get_correlation();
1006
1007         vector<BusLevel> bus_levels;
1008         bus_levels.resize(input_mapping.buses.size());
1009         {
1010                 lock_guard<mutex> lock(compressor_mutex);
1011                 for (unsigned bus_index = 0; bus_index < bus_levels.size(); ++bus_index) {
1012                         BusLevel &levels = bus_levels[bus_index];
1013                         BusMetrics &metrics = bus_metrics[bus_index];
1014
1015                         levels.current_level_dbfs[0] = metrics.current_level_dbfs[0] = to_db(peak_history[bus_index][0].current_level);
1016                         levels.current_level_dbfs[1] = metrics.current_level_dbfs[1] = to_db(peak_history[bus_index][1].current_level);
1017                         levels.peak_level_dbfs[0] = metrics.peak_level_dbfs[0] = to_db(peak_history[bus_index][0].current_peak);
1018                         levels.peak_level_dbfs[1] = metrics.peak_level_dbfs[1] = to_db(peak_history[bus_index][1].current_peak);
1019                         levels.historic_peak_dbfs = metrics.historic_peak_dbfs = to_db(
1020                                 max(peak_history[bus_index][0].historic_peak,
1021                                     peak_history[bus_index][1].historic_peak));
1022                         levels.gain_staging_db = metrics.gain_staging_db = gain_staging_db[bus_index];
1023                         if (compressor_enabled[bus_index]) {
1024                                 levels.compressor_attenuation_db = metrics.compressor_attenuation_db = -to_db(compressor[bus_index]->get_attenuation());
1025                         } else {
1026                                 levels.compressor_attenuation_db = 0.0;
1027                                 metrics.compressor_attenuation_db = 0.0 / 0.0;
1028                         }
1029                 }
1030         }
1031
1032         audio_level_callback(loudness_s, to_db(peak), bus_levels,
1033                 loudness_i, loudness_range_low, loudness_range_high,
1034                 to_db(final_makeup_gain),
1035                 correlation.get_correlation());
1036 }
1037
1038 map<DeviceSpec, DeviceInfo> AudioMixer::get_devices(HoldDevices hold_devices)
1039 {
1040         lock_guard<timed_mutex> lock(audio_mutex);
1041
1042         map<DeviceSpec, DeviceInfo> devices;
1043         for (unsigned card_index = 0; card_index < num_capture_cards; ++card_index) {
1044                 const DeviceSpec spec{ InputSourceType::CAPTURE_CARD, card_index };
1045                 const AudioDevice *device = &video_cards[card_index];
1046                 DeviceInfo info;
1047                 info.display_name = device->display_name;
1048                 info.num_channels = 8;
1049                 devices.insert(make_pair(spec, info));
1050         }
1051         vector<ALSAPool::Device> available_alsa_devices = alsa_pool.get_devices(hold_devices);
1052         for (unsigned card_index = 0; card_index < available_alsa_devices.size(); ++card_index) {
1053                 const DeviceSpec spec{ InputSourceType::ALSA_INPUT, card_index };
1054                 const ALSAPool::Device &device = available_alsa_devices[card_index];
1055                 DeviceInfo info;
1056                 info.display_name = device.display_name();
1057                 info.num_channels = device.num_channels;
1058                 info.alsa_name = device.name;
1059                 info.alsa_info = device.info;
1060                 info.alsa_address = device.address;
1061                 devices.insert(make_pair(spec, info));
1062         }
1063         for (unsigned card_index = 0; card_index < num_ffmpeg_inputs; ++card_index) {
1064                 const DeviceSpec spec{ InputSourceType::FFMPEG_VIDEO_INPUT, card_index };
1065                 const AudioDevice *device = &ffmpeg_inputs[card_index];
1066                 DeviceInfo info;
1067                 info.display_name = device->display_name;
1068                 info.num_channels = 2;
1069                 devices.insert(make_pair(spec, info));
1070         }
1071         return devices;
1072 }
1073
1074 void AudioMixer::set_display_name(DeviceSpec device_spec, const string &name)
1075 {
1076         AudioDevice *device = find_audio_device(device_spec);
1077
1078         lock_guard<timed_mutex> lock(audio_mutex);
1079         device->display_name = name;
1080 }
1081
1082 void AudioMixer::serialize_device(DeviceSpec device_spec, DeviceSpecProto *device_spec_proto)
1083 {
1084         lock_guard<timed_mutex> lock(audio_mutex);
1085         switch (device_spec.type) {
1086                 case InputSourceType::SILENCE:
1087                         device_spec_proto->set_type(DeviceSpecProto::SILENCE);
1088                         break;
1089                 case InputSourceType::CAPTURE_CARD:
1090                         device_spec_proto->set_type(DeviceSpecProto::CAPTURE_CARD);
1091                         device_spec_proto->set_index(device_spec.index);
1092                         device_spec_proto->set_display_name(video_cards[device_spec.index].display_name);
1093                         break;
1094                 case InputSourceType::ALSA_INPUT:
1095                         alsa_pool.serialize_device(device_spec.index, device_spec_proto);
1096                         break;
1097                 case InputSourceType::FFMPEG_VIDEO_INPUT:
1098                         device_spec_proto->set_type(DeviceSpecProto::FFMPEG_VIDEO_INPUT);
1099                         device_spec_proto->set_index(device_spec.index);
1100                         device_spec_proto->set_display_name(ffmpeg_inputs[device_spec.index].display_name);
1101                         break;
1102         }
1103 }
1104
1105 void AudioMixer::set_simple_input(unsigned card_index)
1106 {
1107         assert(card_index < num_capture_cards + num_ffmpeg_inputs);
1108         InputMapping new_input_mapping;
1109         InputMapping::Bus input;
1110         input.name = "Main";
1111         if (card_index >= num_capture_cards) {
1112                 input.device = DeviceSpec{InputSourceType::FFMPEG_VIDEO_INPUT, card_index - num_capture_cards};
1113         } else {
1114                 input.device = DeviceSpec{InputSourceType::CAPTURE_CARD, card_index};
1115         }
1116         input.source_channel[0] = 0;
1117         input.source_channel[1] = 1;
1118
1119         new_input_mapping.buses.push_back(input);
1120
1121         // NOTE: Delay is implicitly at 0.0 ms, since none has been set in the mapping.
1122
1123         lock_guard<timed_mutex> lock(audio_mutex);
1124         current_mapping_mode = MappingMode::SIMPLE;
1125         set_input_mapping_lock_held(new_input_mapping);
1126         fader_volume_db[0] = 0.0f;
1127 }
1128
1129 unsigned AudioMixer::get_simple_input() const
1130 {
1131         lock_guard<timed_mutex> lock(audio_mutex);
1132         if (input_mapping.buses.size() == 1 &&
1133             input_mapping.buses[0].device.type == InputSourceType::CAPTURE_CARD &&
1134             input_mapping.buses[0].source_channel[0] == 0 &&
1135             input_mapping.buses[0].source_channel[1] == 1) {
1136                 return input_mapping.buses[0].device.index;
1137         } else if (input_mapping.buses.size() == 1 &&
1138                    input_mapping.buses[0].device.type == InputSourceType::FFMPEG_VIDEO_INPUT &&
1139                    input_mapping.buses[0].source_channel[0] == 0 &&
1140                    input_mapping.buses[0].source_channel[1] == 1) {
1141                 return input_mapping.buses[0].device.index + num_capture_cards;
1142         } else {
1143                 return numeric_limits<unsigned>::max();
1144         }
1145 }
1146
1147 void AudioMixer::set_input_mapping(const InputMapping &new_input_mapping)
1148 {
1149         lock_guard<timed_mutex> lock(audio_mutex);
1150         set_input_mapping_lock_held(new_input_mapping);
1151         current_mapping_mode = MappingMode::MULTICHANNEL;
1152 }
1153
1154 AudioMixer::MappingMode AudioMixer::get_mapping_mode() const
1155 {
1156         lock_guard<timed_mutex> lock(audio_mutex);
1157         return current_mapping_mode;
1158 }
1159
1160 void AudioMixer::set_input_mapping_lock_held(const InputMapping &new_input_mapping)
1161 {
1162         map<DeviceSpec, set<unsigned>> interesting_channels;
1163         for (const InputMapping::Bus &bus : new_input_mapping.buses) {
1164                 if (bus.device.type == InputSourceType::CAPTURE_CARD ||
1165                     bus.device.type == InputSourceType::ALSA_INPUT ||
1166                     bus.device.type == InputSourceType::FFMPEG_VIDEO_INPUT) {
1167                         for (unsigned channel = 0; channel < 2; ++channel) {
1168                                 if (bus.source_channel[channel] != -1) {
1169                                         interesting_channels[bus.device].insert(bus.source_channel[channel]);
1170                                 }
1171                         }
1172                 } else {
1173                         assert(bus.device.type == InputSourceType::SILENCE);
1174                 }
1175         }
1176
1177         // Kill all the old metrics, and set up new ones.
1178         for (unsigned bus_index = 0; bus_index < input_mapping.buses.size(); ++bus_index) {
1179                 BusMetrics &metrics = bus_metrics[bus_index];
1180
1181                 vector<pair<string, string>> labels_left = metrics.labels;
1182                 labels_left.emplace_back("channel", "left");
1183                 vector<pair<string, string>> labels_right = metrics.labels;
1184                 labels_right.emplace_back("channel", "right");
1185
1186                 global_metrics.remove("bus_current_level_dbfs", labels_left);
1187                 global_metrics.remove("bus_current_level_dbfs", labels_right);
1188                 global_metrics.remove("bus_peak_level_dbfs", labels_left);
1189                 global_metrics.remove("bus_peak_level_dbfs", labels_right);
1190                 global_metrics.remove("bus_historic_peak_dbfs", metrics.labels);
1191                 global_metrics.remove("bus_gain_staging_db", metrics.labels);
1192                 global_metrics.remove("bus_compressor_attenuation_db", metrics.labels);
1193         }
1194         bus_metrics.reset(new BusMetrics[new_input_mapping.buses.size()]);
1195         for (unsigned bus_index = 0; bus_index < new_input_mapping.buses.size(); ++bus_index) {
1196                 const InputMapping::Bus &bus = new_input_mapping.buses[bus_index];
1197                 BusMetrics &metrics = bus_metrics[bus_index];
1198
1199                 char bus_index_str[16], source_index_str[16], source_channels_str[64];
1200                 snprintf(bus_index_str, sizeof(bus_index_str), "%u", bus_index);
1201                 snprintf(source_index_str, sizeof(source_index_str), "%u", bus.device.index);
1202                 snprintf(source_channels_str, sizeof(source_channels_str), "%d:%d", bus.source_channel[0], bus.source_channel[1]);
1203
1204                 vector<pair<string, string>> labels;
1205                 metrics.labels.emplace_back("index", bus_index_str);
1206                 metrics.labels.emplace_back("name", bus.name);
1207                 if (bus.device.type == InputSourceType::SILENCE) {
1208                         metrics.labels.emplace_back("source_type", "silence");
1209                 } else if (bus.device.type == InputSourceType::CAPTURE_CARD) {
1210                         metrics.labels.emplace_back("source_type", "capture_card");
1211                 } else if (bus.device.type == InputSourceType::ALSA_INPUT) {
1212                         metrics.labels.emplace_back("source_type", "alsa_input");
1213                 } else if (bus.device.type == InputSourceType::FFMPEG_VIDEO_INPUT) {
1214                         metrics.labels.emplace_back("source_type", "ffmpeg_video_input");
1215                 } else {
1216                         assert(false);
1217                 }
1218                 metrics.labels.emplace_back("source_index", source_index_str);
1219                 metrics.labels.emplace_back("source_channels", source_channels_str);
1220
1221                 vector<pair<string, string>> labels_left = metrics.labels;
1222                 labels_left.emplace_back("channel", "left");
1223                 vector<pair<string, string>> labels_right = metrics.labels;
1224                 labels_right.emplace_back("channel", "right");
1225
1226                 global_metrics.add("bus_current_level_dbfs", labels_left, &metrics.current_level_dbfs[0], Metrics::TYPE_GAUGE);
1227                 global_metrics.add("bus_current_level_dbfs", labels_right, &metrics.current_level_dbfs[1], Metrics::TYPE_GAUGE);
1228                 global_metrics.add("bus_peak_level_dbfs", labels_left, &metrics.peak_level_dbfs[0], Metrics::TYPE_GAUGE);
1229                 global_metrics.add("bus_peak_level_dbfs", labels_right, &metrics.peak_level_dbfs[1], Metrics::TYPE_GAUGE);
1230                 global_metrics.add("bus_historic_peak_dbfs", metrics.labels, &metrics.historic_peak_dbfs, Metrics::TYPE_GAUGE);
1231                 global_metrics.add("bus_gain_staging_db", metrics.labels, &metrics.gain_staging_db, Metrics::TYPE_GAUGE);
1232                 global_metrics.add("bus_compressor_attenuation_db", metrics.labels, &metrics.compressor_attenuation_db, Metrics::TYPE_GAUGE);
1233         }
1234
1235         // Reset resamplers for all cards that don't have the exact same state as before.
1236         map<DeviceSpec, double> new_extra_delay_ms = new_input_mapping.extra_delay_ms;  // Convenience so we can use [].
1237         for (unsigned card_index = 0; card_index < MAX_VIDEO_CARDS; ++card_index) {
1238                 const DeviceSpec device_spec{InputSourceType::CAPTURE_CARD, card_index};
1239                 AudioDevice *device = find_audio_device(device_spec);
1240                 if (device->interesting_channels != interesting_channels[device_spec] ||
1241                     device->extra_delay_ms != new_extra_delay_ms[device_spec]) {
1242                         device->interesting_channels = interesting_channels[device_spec];
1243                         device->extra_delay_ms = new_extra_delay_ms[device_spec];
1244                         reset_resampler_mutex_held(device_spec);
1245                 }
1246         }
1247         for (unsigned card_index = 0; card_index < MAX_ALSA_CARDS; ++card_index) {
1248                 const DeviceSpec device_spec{InputSourceType::ALSA_INPUT, card_index};
1249                 AudioDevice *device = find_audio_device(device_spec);
1250                 if (interesting_channels[device_spec].empty()) {
1251                         alsa_pool.release_device(card_index);
1252                 } else {
1253                         alsa_pool.hold_device(card_index);
1254                 }
1255                 if (device->interesting_channels != interesting_channels[device_spec] ||
1256                     device->extra_delay_ms != new_extra_delay_ms[device_spec]) {
1257                         device->interesting_channels = interesting_channels[device_spec];
1258                         device->extra_delay_ms = new_extra_delay_ms[device_spec];
1259                         alsa_pool.reset_device(device_spec.index);
1260                         reset_resampler_mutex_held(device_spec);
1261                 }
1262         }
1263         for (unsigned card_index = 0; card_index < num_ffmpeg_inputs; ++card_index) {
1264                 const DeviceSpec device_spec{InputSourceType::FFMPEG_VIDEO_INPUT, card_index};
1265                 AudioDevice *device = find_audio_device(device_spec);
1266                 if (device->interesting_channels != interesting_channels[device_spec] ||
1267                     device->extra_delay_ms != new_extra_delay_ms[device_spec]) {
1268                         device->interesting_channels = interesting_channels[device_spec];
1269                         device->extra_delay_ms = new_extra_delay_ms[device_spec];
1270                         reset_resampler_mutex_held(device_spec);
1271                 }
1272         }
1273
1274         input_mapping = new_input_mapping;
1275 }
1276
1277 InputMapping AudioMixer::get_input_mapping() const
1278 {
1279         lock_guard<timed_mutex> lock(audio_mutex);
1280         return input_mapping;
1281 }
1282
1283 unsigned AudioMixer::num_buses() const
1284 {
1285         lock_guard<timed_mutex> lock(audio_mutex);
1286         return input_mapping.buses.size();
1287 }
1288
1289 void AudioMixer::reset_peak(unsigned bus_index)
1290 {
1291         lock_guard<timed_mutex> lock(audio_mutex);
1292         for (unsigned channel = 0; channel < 2; ++channel) {
1293                 PeakHistory &history = peak_history[bus_index][channel];
1294                 history.current_level = 0.0f;
1295                 history.historic_peak = 0.0f;
1296                 history.current_peak = 0.0f;
1297                 history.last_peak = 0.0f;
1298                 history.age_seconds = 0.0f;
1299         }
1300 }
1301
1302 bool AudioMixer::is_mono(unsigned bus_index)
1303 {
1304         lock_guard<timed_mutex> lock(audio_mutex);
1305         const InputMapping::Bus &bus = input_mapping.buses[bus_index];
1306         if (bus.device.type == InputSourceType::SILENCE) {
1307                 return true;
1308         } else {
1309                 assert(bus.device.type == InputSourceType::CAPTURE_CARD ||
1310                        bus.device.type == InputSourceType::ALSA_INPUT ||
1311                        bus.device.type == InputSourceType::FFMPEG_VIDEO_INPUT);
1312                 return bus.source_channel[0] == bus.source_channel[1];
1313         }
1314 }
1315
1316 AudioMixer *global_audio_mixer = nullptr;