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

   1 #include "audio_mixer.h"
   2
   3 #include <assert.h>
   4 #include <bmusb/bmusb.h>
   5 #include <endian.h>
   6 #include <math.h>
   7 #ifdef __SSE2__
   8 #include <immintrin.h>
   9 #endif
  10 #include <stdbool.h>
  11 #include <stdio.h>
  12 #include <stdlib.h>
  13 #include <string.h>
  14 #include <algorithm>
  15 #include <chrono>
  16 #include <cmath>
  17 #include <cstddef>
  18 #include <limits>
  19 #include <utility>
  20
  21 #include "decibel.h"
  22 #include "flags.h"
  23 #include "shared/metrics.h"
  24 #include "state.pb.h"
  25 #include "shared/timebase.h"
  26
  27 using namespace bmusb;
  28 using namespace std;
  29 using namespace std::chrono;
  30 using namespace std::placeholders;
  31
  32 namespace {
  33
  34 // TODO: If these prove to be a bottleneck, they can be SSSE3-optimized
  35 // (usually including multiple channels at a time).
  36
  37 void convert_fixed16_to_fp32(float *dst, size_t out_channel, size_t out_num_channels,
  38                              const uint8_t *src, size_t in_channel, size_t in_num_channels,
  39                              size_t num_samples)
  40 {
  41         assert(in_channel < in_num_channels);
  42         assert(out_channel < out_num_channels);
  43         src += in_channel * 2;
  44         dst += out_channel;
  45
  46         for (size_t i = 0; i < num_samples; ++i) {
  47                 int16_t s = le16toh(*(int16_t *)src);
  48                 *dst = s * (1.0f / 32768.0f);
  49
  50                 src += 2 * in_num_channels;
  51                 dst += out_num_channels;
  52         }
  53 }
  54
  55 void convert_fixed16_to_fixed32(int32_t *dst, size_t out_channel, size_t out_num_channels,
  56                                 const uint8_t *src, size_t in_channel, size_t in_num_channels,
  57                                 size_t num_samples)
  58 {
  59         assert(in_channel < in_num_channels);
  60         assert(out_channel < out_num_channels);
  61         src += in_channel * 2;
  62         dst += out_channel;
  63
  64         for (size_t i = 0; i < num_samples; ++i) {
  65                 uint32_t s = uint32_t(uint16_t(le16toh(*(int16_t *)src))) << 16;
  66
  67                 // Keep the sign bit in place, repeat the other 15 bits as far as they go.
  68                 *dst = s | ((s & 0x7fffffff) >> 15) | ((s & 0x7fffffff) >> 30);
  69
  70                 src += 2 * in_num_channels;
  71                 dst += out_num_channels;
  72         }
  73 }
  74
  75 void convert_fixed24_to_fp32(float *dst, size_t out_channel, size_t out_num_channels,
  76                              const uint8_t *src, size_t in_channel, size_t in_num_channels,
  77                              size_t num_samples)
  78 {
  79         assert(in_channel < in_num_channels);
  80         assert(out_channel < out_num_channels);
  81         src += in_channel * 3;
  82         dst += out_channel;
  83
  84         for (size_t i = 0; i < num_samples; ++i) {
  85                 uint32_t s1 = src[0];
  86                 uint32_t s2 = src[1];
  87                 uint32_t s3 = src[2];
  88                 uint32_t s = (s1 << 8) | (s2 << 16) | (s3 << 24);  // Note: The bottom eight bits are zero; s3 includes the sign bit.
  89                 *dst = int(s) * (1.0f / (256.0f * 8388608.0f));  // 256 for signed down-shift by 8, then 2^23 for the actual conversion.
  90
  91                 src += 3 * in_num_channels;
  92                 dst += out_num_channels;
  93         }
  94 }
  95
  96 void convert_fixed24_to_fixed32(int32_t *dst, size_t out_channel, size_t out_num_channels,
  97                                 const uint8_t *src, size_t in_channel, size_t in_num_channels,
  98                                 size_t num_samples)
  99 {
 100         assert(in_channel < in_num_channels);
 101         assert(out_channel < out_num_channels);
 102         src += in_channel * 3;
 103         dst += out_channel;
 104
 105         for (size_t i = 0; i < num_samples; ++i) {
 106                 uint32_t s1 = src[0];
 107                 uint32_t s2 = src[1];
 108                 uint32_t s3 = src[2];
 109                 uint32_t s = (s1 << 8) | (s2 << 16) | (s3 << 24);
 110
 111                 // Keep the sign bit in place, repeat the other 23 bits as far as they go.
 112                 *dst = s | ((s & 0x7fffffff) >> 23);
 113
 114                 src += 3 * in_num_channels;
 115                 dst += out_num_channels;
 116         }
 117 }
 118
 119 void convert_fixed32_to_fp32(float *dst, size_t out_channel, size_t out_num_channels,
 120                              const uint8_t *src, size_t in_channel, size_t in_num_channels,
 121                              size_t num_samples)
 122 {
 123         assert(in_channel < in_num_channels);
 124         assert(out_channel < out_num_channels);
 125         src += in_channel * 4;
 126         dst += out_channel;
 127
 128         for (size_t i = 0; i < num_samples; ++i) {
 129                 int32_t s = le32toh(*(int32_t *)src);
 130                 *dst = s * (1.0f / 2147483648.0f);
 131
 132                 src += 4 * in_num_channels;
 133                 dst += out_num_channels;
 134         }
 135 }
 136
 137 // Basically just a reinterleave.
 138 void convert_fixed32_to_fixed32(int32_t *dst, size_t out_channel, size_t out_num_channels,
 139                                 const uint8_t *src, size_t in_channel, size_t in_num_channels,
 140                                 size_t num_samples)
 141 {
 142         assert(in_channel < in_num_channels);
 143         assert(out_channel < out_num_channels);
 144         src += in_channel * 4;
 145         dst += out_channel;
 146
 147         for (size_t i = 0; i < num_samples; ++i) {
 148                 int32_t s = le32toh(*(int32_t *)src);
 149                 *dst = s;
 150
 151                 src += 4 * in_num_channels;
 152                 dst += out_num_channels;
 153         }
 154 }
 155
 156 float find_peak_plain(const float *samples, size_t num_samples) __attribute__((unused));
 157
 158 float find_peak_plain(const float *samples, size_t num_samples)
 159 {
 160         float m = fabs(samples[0]);
 161         for (size_t i = 1; i < num_samples; ++i) {
 162                 m = max(m, fabs(samples[i]));
 163         }
 164         return m;
 165 }
 166
 167 #ifdef __SSE__
 168 static inline float horizontal_max(__m128 m)
 169 {
 170         __m128 tmp = _mm_shuffle_ps(m, m, _MM_SHUFFLE(1, 0, 3, 2));
 171         m = _mm_max_ps(m, tmp);
 172         tmp = _mm_shuffle_ps(m, m, _MM_SHUFFLE(2, 3, 0, 1));
 173         m = _mm_max_ps(m, tmp);
 174         return _mm_cvtss_f32(m);
 175 }
 176
 177 float find_peak(const float *samples, size_t num_samples)
 178 {
 179         const __m128 abs_mask = _mm_castsi128_ps(_mm_set1_epi32(0x7fffffffu));
 180         __m128 m = _mm_setzero_ps();
 181         for (size_t i = 0; i < (num_samples & ~3); i += 4) {
 182                 __m128 x = _mm_loadu_ps(samples + i);
 183                 x = _mm_and_ps(x, abs_mask);
 184                 m = _mm_max_ps(m, x);
 185         }
 186         float result = horizontal_max(m);
 187
 188         for (size_t i = (num_samples & ~3); i < num_samples; ++i) {
 189                 result = max(result, fabs(samples[i]));
 190         }
 191
 192 #if 0
 193         // Self-test. We should be bit-exact the same.
 194         float reference_result = find_peak_plain(samples, num_samples);
 195         if (result != reference_result) {
 196                 fprintf(stderr, "Error: Peak is %f [%f %f %f %f]; should be %f.\n",
 197                         result,
 198                         _mm_cvtss_f32(_mm_shuffle_ps(m, m, _MM_SHUFFLE(0, 0, 0, 0))),
 199                         _mm_cvtss_f32(_mm_shuffle_ps(m, m, _MM_SHUFFLE(1, 1, 1, 1))),
 200                         _mm_cvtss_f32(_mm_shuffle_ps(m, m, _MM_SHUFFLE(2, 2, 2, 2))),
 201                         _mm_cvtss_f32(_mm_shuffle_ps(m, m, _MM_SHUFFLE(3, 3, 3, 3))),
 202                         reference_result);
 203                 abort();
 204         }
 205 #endif
 206         return result;
 207 }
 208 #else
 209 float find_peak(const float *samples, size_t num_samples)
 210 {
 211         return find_peak_plain(samples, num_samples);
 212 }
 213 #endif
 214
 215 void deinterleave_samples(const vector<float> &in, vector<float> *out_l, vector<float> *out_r)
 216 {
 217         size_t num_samples = in.size() / 2;
 218         out_l->resize(num_samples);
 219         out_r->resize(num_samples);
 220
 221         const float *inptr = in.data();
 222         float *lptr = &(*out_l)[0];
 223         float *rptr = &(*out_r)[0];
 224         for (size_t i = 0; i < num_samples; ++i) {
 225                 *lptr++ = *inptr++;
 226                 *rptr++ = *inptr++;
 227         }
 228 }
 229
 230 }  // namespace
 231
 232 AudioMixer::AudioMixer(unsigned num_capture_cards, unsigned num_ffmpeg_inputs)
 233         : num_capture_cards(num_capture_cards),
 234           num_ffmpeg_inputs(num_ffmpeg_inputs),
 235           ffmpeg_inputs(new AudioDevice[num_ffmpeg_inputs]),
 236           limiter(OUTPUT_FREQUENCY),
 237           correlation(OUTPUT_FREQUENCY)
 238 {
 239         for (unsigned bus_index = 0; bus_index < MAX_BUSES; ++bus_index) {
 240                 locut[bus_index].init(FILTER_HPF, 2);
 241                 eq[bus_index][EQ_BAND_BASS].init(FILTER_LOW_SHELF, 1);
 242                 // Note: EQ_BAND_MID isn't used (see comments in apply_eq()).
 243                 eq[bus_index][EQ_BAND_TREBLE].init(FILTER_HIGH_SHELF, 1);
 244                 compressor[bus_index].reset(new StereoCompressor(OUTPUT_FREQUENCY));
 245                 level_compressor[bus_index].reset(new StereoCompressor(OUTPUT_FREQUENCY));
 246
 247                 set_bus_settings(bus_index, get_default_bus_settings());
 248         }
 249         set_limiter_enabled(global_flags.limiter_enabled);
 250         set_final_makeup_gain_auto(global_flags.final_makeup_gain_auto);
 251
 252         r128.init(2, OUTPUT_FREQUENCY);
 253         r128.integr_start();
 254
 255         // hlen=16 is pretty low quality, but we use quite a bit of CPU otherwise,
 256         // and there's a limit to how important the peak meter is.
 257         peak_resampler.setup(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY * 4, /*num_channels=*/2, /*hlen=*/16, /*frel=*/1.0);
 258
 259         global_audio_mixer = this;
 260         alsa_pool.init();
 261
 262         if (!global_flags.input_mapping_filename.empty()) {
 263                 // Must happen after ALSAPool is initialized, as it needs to know the card list.
 264                 current_mapping_mode = MappingMode::MULTICHANNEL;
 265                 InputMapping new_input_mapping;
 266                 if (!load_input_mapping_from_file(get_devices(),
 267                                                   global_flags.input_mapping_filename,
 268                                                   &new_input_mapping)) {
 269                         fprintf(stderr, "Failed to load input mapping from '%s', exiting.\n",
 270                                 global_flags.input_mapping_filename.c_str());
 271                         abort();
 272                 }
 273                 set_input_mapping(new_input_mapping);
 274         } else {
 275                 set_simple_input(/*card_index=*/0);
 276                 if (global_flags.multichannel_mapping_mode) {
 277                         current_mapping_mode = MappingMode::MULTICHANNEL;
 278                 }
 279         }
 280
 281         global_metrics.add("audio_loudness_short_lufs", &metric_audio_loudness_short_lufs, Metrics::TYPE_GAUGE);
 282         global_metrics.add("audio_loudness_integrated_lufs", &metric_audio_loudness_integrated_lufs, Metrics::TYPE_GAUGE);
 283         global_metrics.add("audio_loudness_range_low_lufs", &metric_audio_loudness_range_low_lufs, Metrics::TYPE_GAUGE);
 284         global_metrics.add("audio_loudness_range_high_lufs", &metric_audio_loudness_range_high_lufs, Metrics::TYPE_GAUGE);
 285         global_metrics.add("audio_peak_dbfs", &metric_audio_peak_dbfs, Metrics::TYPE_GAUGE);
 286         global_metrics.add("audio_final_makeup_gain_db", &metric_audio_final_makeup_gain_db, Metrics::TYPE_GAUGE);
 287         global_metrics.add("audio_correlation", &metric_audio_correlation, Metrics::TYPE_GAUGE);
 288 }
 289
 290 void AudioMixer::reset_resampler(DeviceSpec device_spec)
 291 {
 292         lock_guard<timed_mutex> lock(audio_mutex);
 293         reset_resampler_mutex_held(device_spec);
 294 }
 295
 296 void AudioMixer::reset_resampler_mutex_held(DeviceSpec device_spec)
 297 {
 298         AudioDevice *device = find_audio_device(device_spec);
 299
 300         if (device->interesting_channels.empty()) {
 301                 device->resampling_queue.reset();
 302         } else {
 303                 // Make sure we never get negative delay. Even 1 ms is probably way less than we
 304                 // could ever hope to actually have; this is just a failsafe.
 305                 double delay_ms = max(global_flags.audio_queue_length_ms + device->extra_delay_ms, 1.0);
 306
 307                 device->resampling_queue.reset(new ResamplingQueue(
 308                         device_spec, device->capture_frequency, OUTPUT_FREQUENCY, device->interesting_channels.size(),
 309                         delay_ms * 0.001));
 310         }
 311 }
 312
 313 bool AudioMixer::add_audio(DeviceSpec device_spec, const uint8_t *data, unsigned num_samples, AudioFormat audio_format, steady_clock::time_point frame_time)
 314 {
 315         AudioDevice *device = find_audio_device(device_spec);
 316
 317         unique_lock<timed_mutex> lock(audio_mutex, defer_lock);
 318         if (!lock.try_lock_for(chrono::milliseconds(10))) {
 319                 return false;
 320         }
 321         if (device->resampling_queue == nullptr) {
 322                 // No buses use this device; throw it away.
 323                 return true;
 324         }
 325
 326         unsigned num_channels = device->interesting_channels.size();
 327         assert(num_channels > 0);
 328
 329         // Convert the audio to fp32.
 330         unique_ptr<float[]> audio(new float[num_samples * num_channels]);
 331         unsigned channel_index = 0;
 332         for (auto channel_it = device->interesting_channels.cbegin(); channel_it != device->interesting_channels.end(); ++channel_it, ++channel_index) {
 333                 convert_audio_to_fp32(audio.get(), channel_index, num_channels, data, *channel_it, audio_format, num_samples);
 334         }
 335
 336         // If we changed frequency since last frame, we'll need to reset the resampler.
 337         if (audio_format.sample_rate != device->capture_frequency) {
 338                 device->capture_frequency = audio_format.sample_rate;
 339                 reset_resampler_mutex_held(device_spec);
 340         }
 341
 342         // Now add it.
 343         device->resampling_queue->add_input_samples(frame_time, audio.get(), num_samples, ResamplingQueue::ADJUST_RATE);
 344         return true;
 345 }
 346
 347 // Converts all channels.
 348 vector<int32_t> convert_audio_to_fixed32(const uint8_t *data, unsigned num_samples, bmusb::AudioFormat audio_format, unsigned num_channels)
 349 {
 350         vector<int32_t> audio;
 351
 352         if (num_channels > audio_format.num_channels) {
 353                 audio.resize(num_samples * num_channels, 0);
 354         } else {
 355                 audio.resize(num_samples * num_channels);
 356         }
 357         for (unsigned channel_index = 0; channel_index < num_channels && channel_index < audio_format.num_channels; ++channel_index) {
 358                 switch (audio_format.bits_per_sample) {
 359                 case 0:
 360                         assert(num_samples == 0);
 361                         break;
 362                 case 16:
 363                         convert_fixed16_to_fixed32(&audio[0], channel_index, num_channels, data, channel_index, audio_format.num_channels, num_samples);
 364                         break;
 365                 case 24:
 366                         convert_fixed24_to_fixed32(&audio[0], channel_index, num_channels, data, channel_index, audio_format.num_channels, num_samples);
 367                         break;
 368                 case 32:
 369                         convert_fixed32_to_fixed32(&audio[0], channel_index, num_channels, data, channel_index, audio_format.num_channels, num_samples);
 370                         break;
 371                 default:
 372                         fprintf(stderr, "Cannot handle audio with %u bits per sample\n", audio_format.bits_per_sample);
 373                         assert(false);
 374                 }
 375         }
 376
 377         return audio;
 378 }
 379
 380 // Converts only one channel.
 381 void convert_audio_to_fp32(float *dst, size_t out_channel, size_t out_num_channels,
 382                            const uint8_t *src, size_t in_channel, bmusb::AudioFormat in_audio_format,
 383                            size_t num_samples)
 384 {
 385         switch (in_audio_format.bits_per_sample) {
 386         case 0:
 387                 assert(num_samples == 0);
 388                 break;
 389         case 16:
 390                 convert_fixed16_to_fp32(dst, out_channel, out_num_channels, src, in_channel, in_audio_format.num_channels, num_samples);
 391                 break;
 392         case 24:
 393                 convert_fixed24_to_fp32(dst, out_channel, out_num_channels, src, in_channel, in_audio_format.num_channels, num_samples);
 394                 break;
 395         case 32:
 396                 convert_fixed32_to_fp32(dst, out_channel, out_num_channels, src, in_channel, in_audio_format.num_channels, num_samples);
 397                 break;
 398         default:
 399                 fprintf(stderr, "Cannot handle audio with %u bits per sample\n", in_audio_format.bits_per_sample);
 400                 assert(false);
 401         }
 402 }
 403
 404 bool AudioMixer::add_silence(DeviceSpec device_spec, unsigned samples_per_frame, unsigned num_frames)
 405 {
 406         AudioDevice *device = find_audio_device(device_spec);
 407
 408         unique_lock<timed_mutex> lock(audio_mutex, defer_lock);
 409         if (!lock.try_lock_for(chrono::milliseconds(10))) {
 410                 return false;
 411         }
 412         if (device->resampling_queue == nullptr) {
 413                 // No buses use this device; throw it away.
 414                 return true;
 415         }
 416
 417         unsigned num_channels = device->interesting_channels.size();
 418         assert(num_channels > 0);
 419
 420         vector<float> silence(samples_per_frame * num_channels, 0.0f);
 421         for (unsigned i = 0; i < num_frames; ++i) {
 422                 device->resampling_queue->add_input_samples(steady_clock::now(), silence.data(), samples_per_frame, ResamplingQueue::DO_NOT_ADJUST_RATE);
 423         }
 424         return true;
 425 }
 426
 427 bool AudioMixer::silence_card(DeviceSpec device_spec, bool silence)
 428 {
 429         AudioDevice *device = find_audio_device(device_spec);
 430
 431         unique_lock<timed_mutex> lock(audio_mutex, defer_lock);
 432         if (!lock.try_lock_for(chrono::milliseconds(10))) {
 433                 return false;
 434         }
 435
 436         if (device->silenced && !silence) {
 437                 reset_resampler_mutex_held(device_spec);
 438         }
 439         device->silenced = silence;
 440         return true;
 441 }
 442
 443 AudioMixer::BusSettings AudioMixer::get_default_bus_settings()
 444 {
 445         BusSettings settings;
 446         settings.fader_volume_db = 0.0f;
 447         settings.muted = false;
 448         settings.locut_enabled = global_flags.locut_enabled;
 449         settings.stereo_width = 1.0f;
 450         for (unsigned band_index = 0; band_index < NUM_EQ_BANDS; ++band_index) {
 451                 settings.eq_level_db[band_index] = 0.0f;
 452         }
 453         settings.gain_staging_db = global_flags.initial_gain_staging_db;
 454         settings.level_compressor_enabled = global_flags.gain_staging_auto;
 455         settings.compressor_threshold_dbfs = ref_level_dbfs - 12.0f;  // -12 dB.
 456         settings.compressor_enabled = global_flags.compressor_enabled;
 457         return settings;
 458 }
 459
 460 AudioMixer::BusSettings AudioMixer::get_bus_settings(unsigned bus_index) const
 461 {
 462         lock_guard<timed_mutex> lock(audio_mutex);
 463         BusSettings settings;
 464         settings.fader_volume_db = fader_volume_db[bus_index];
 465         settings.muted = mute[bus_index];
 466         settings.locut_enabled = locut_enabled[bus_index];
 467         settings.stereo_width = stereo_width[bus_index];
 468         for (unsigned band_index = 0; band_index < NUM_EQ_BANDS; ++band_index) {
 469                 settings.eq_level_db[band_index] = eq_level_db[bus_index][band_index];
 470         }
 471         settings.gain_staging_db = gain_staging_db[bus_index];
 472         settings.level_compressor_enabled = level_compressor_enabled[bus_index];
 473         settings.compressor_threshold_dbfs = compressor_threshold_dbfs[bus_index];
 474         settings.compressor_enabled = compressor_enabled[bus_index];
 475         return settings;
 476 }
 477
 478 void AudioMixer::set_bus_settings(unsigned bus_index, const AudioMixer::BusSettings &settings)
 479 {
 480         lock_guard<timed_mutex> lock(audio_mutex);
 481         fader_volume_db[bus_index] = settings.fader_volume_db;
 482         mute[bus_index] = settings.muted;
 483         locut_enabled[bus_index] = settings.locut_enabled;
 484         stereo_width[bus_index] = settings.stereo_width;
 485         for (unsigned band_index = 0; band_index < NUM_EQ_BANDS; ++band_index) {
 486                 eq_level_db[bus_index][band_index] = settings.eq_level_db[band_index];
 487         }
 488         gain_staging_db[bus_index] = settings.gain_staging_db;
 489         last_gain_staging_db[bus_index] = gain_staging_db[bus_index];
 490         level_compressor_enabled[bus_index] = settings.level_compressor_enabled;
 491         compressor_threshold_dbfs[bus_index] = settings.compressor_threshold_dbfs;
 492         compressor_enabled[bus_index] = settings.compressor_enabled;
 493 }
 494
 495 AudioMixer::AudioDevice *AudioMixer::find_audio_device(DeviceSpec device)
 496 {
 497         switch (device.type) {
 498         case InputSourceType::CAPTURE_CARD:
 499                 return &video_cards[device.index];
 500         case InputSourceType::ALSA_INPUT:
 501                 return &alsa_inputs[device.index];
 502         case InputSourceType::FFMPEG_VIDEO_INPUT:
 503                 return &ffmpeg_inputs[device.index];
 504         case InputSourceType::SILENCE:
 505         default:
 506                 assert(false);
 507         }
 508         return nullptr;
 509 }
 510
 511 // Get a pointer to the given channel from the given device.
 512 // The channel must be picked out earlier and resampled.
 513 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)
 514 {
 515         static float zero = 0.0f;
 516         if (source_channel == -1 || device_spec.type == InputSourceType::SILENCE) {
 517                 *srcptr = &zero;
 518                 *stride = 0;
 519                 return;
 520         }
 521         AudioDevice *device = find_audio_device(device_spec);
 522         assert(device->interesting_channels.count(source_channel) != 0);
 523         unsigned channel_index = 0;
 524         for (int channel : device->interesting_channels) {
 525                 if (channel == source_channel) break;
 526                 ++channel_index;
 527         }
 528         assert(channel_index < device->interesting_channels.size());
 529         const auto it = samples_card.find(device_spec);
 530         assert(it != samples_card.end());
 531         *srcptr = &(it->second)[channel_index];
 532         *stride = device->interesting_channels.size();
 533 }
 534
 535 // TODO: Can be SSSE3-optimized if need be.
 536 void AudioMixer::fill_audio_bus(const map<DeviceSpec, vector<float>> &samples_card, const InputMapping::Bus &bus, unsigned num_samples, float stereo_width, float *output)
 537 {
 538         if (bus.device.type == InputSourceType::SILENCE) {
 539                 memset(output, 0, num_samples * 2 * sizeof(*output));
 540         } else {
 541                 assert(bus.device.type == InputSourceType::CAPTURE_CARD ||
 542                        bus.device.type == InputSourceType::ALSA_INPUT ||
 543                        bus.device.type == InputSourceType::FFMPEG_VIDEO_INPUT);
 544                 const float *lsrc, *rsrc;
 545                 unsigned lstride, rstride;
 546                 float *dptr = output;
 547                 find_sample_src_from_device(samples_card, bus.device, bus.source_channel[0], &lsrc, &lstride);
 548                 find_sample_src_from_device(samples_card, bus.device, bus.source_channel[1], &rsrc, &rstride);
 549
 550                 // Apply stereo width settings. Set stereo width w to a 0..1 range instead of
 551                 // -1..1, since it makes for much easier calculations (so 0.5 = completely mono).
 552                 // Then, what we want is
 553                 //
 554                 //   L' = wL + (1-w)R = R + w(L-R)
 555                 //   R' = wR + (1-w)L = L + w(R-L)
 556                 //
 557                 // This can be further simplified calculation-wise by defining the weighted
 558                 // difference signal D = w(R-L), so that:
 559                 //
 560                 //   L' = R - D
 561                 //   R' = L + D
 562                 float w = 0.5f * stereo_width + 0.5f;
 563                 if (bus.source_channel[0] == bus.source_channel[1]) {
 564                         // Mono anyway, so no need to bother.
 565                         w = 1.0f;
 566                 } else if (fabs(w) < 1e-3) {
 567                         // Perfect inverse.
 568                         swap(lsrc, rsrc);
 569                         swap(lstride, rstride);
 570                         w = 1.0f;
 571                 }
 572                 if (fabs(w - 1.0f) < 1e-3) {
 573                         // No calculations needed for stereo_width = 1.
 574                         for (unsigned i = 0; i < num_samples; ++i) {
 575                                 *dptr++ = *lsrc;
 576                                 *dptr++ = *rsrc;
 577                                 lsrc += lstride;
 578                                 rsrc += rstride;
 579                         }
 580                 } else {
 581                         // General case.
 582                         for (unsigned i = 0; i < num_samples; ++i) {
 583                                 float left = *lsrc, right = *rsrc;
 584                                 float diff = w * (right - left);
 585                                 *dptr++ = right - diff;
 586                                 *dptr++ = left + diff;
 587                                 lsrc += lstride;
 588                                 rsrc += rstride;
 589                         }
 590                 }
 591         }
 592 }
 593
 594 vector<DeviceSpec> AudioMixer::get_active_devices() const
 595 {
 596         vector<DeviceSpec> ret;
 597         for (unsigned card_index = 0; card_index < MAX_VIDEO_CARDS; ++card_index) {
 598                 const DeviceSpec device_spec{InputSourceType::CAPTURE_CARD, card_index};
 599                 if (!find_audio_device(device_spec)->interesting_channels.empty()) {
 600                         ret.push_back(device_spec);
 601                 }
 602         }
 603         for (unsigned card_index = 0; card_index < MAX_ALSA_CARDS; ++card_index) {
 604                 const DeviceSpec device_spec{InputSourceType::ALSA_INPUT, card_index};
 605                 if (!find_audio_device(device_spec)->interesting_channels.empty()) {
 606                         ret.push_back(device_spec);
 607                 }
 608         }
 609         for (unsigned card_index = 0; card_index < num_ffmpeg_inputs; ++card_index) {
 610                 const DeviceSpec device_spec{InputSourceType::FFMPEG_VIDEO_INPUT, card_index};
 611                 if (!find_audio_device(device_spec)->interesting_channels.empty()) {
 612                         ret.push_back(device_spec);
 613                 }
 614         }
 615         return ret;
 616 }
 617
 618 namespace {
 619
 620 void apply_gain(float db, float last_db, vector<float> *samples)
 621 {
 622         if (fabs(db - last_db) < 1e-3) {
 623                 // Constant over this frame.
 624                 const float gain = from_db(db);
 625                 for (size_t i = 0; i < samples->size(); ++i) {
 626                         (*samples)[i] *= gain;
 627                 }
 628         } else {
 629                 // We need to do a fade.
 630                 unsigned num_samples = samples->size() / 2;
 631                 float gain = from_db(last_db);
 632                 const float gain_inc = pow(from_db(db - last_db), 1.0 / num_samples);
 633                 for (size_t i = 0; i < num_samples; ++i) {
 634                         (*samples)[i * 2 + 0] *= gain;
 635                         (*samples)[i * 2 + 1] *= gain;
 636                         gain *= gain_inc;
 637                 }
 638         }
 639 }
 640
 641 }  // namespace
 642
 643 vector<float> AudioMixer::get_output(steady_clock::time_point ts, unsigned num_samples, ResamplingQueue::RateAdjustmentPolicy rate_adjustment_policy)
 644 {
 645         map<DeviceSpec, vector<float>> samples_card;
 646         vector<float> samples_bus;
 647
 648         lock_guard<timed_mutex> lock(audio_mutex);
 649
 650         // Pick out all the interesting channels from all the cards.
 651         for (const DeviceSpec &device_spec : get_active_devices()) {
 652                 AudioDevice *device = find_audio_device(device_spec);
 653                 samples_card[device_spec].resize(num_samples * device->interesting_channels.size());
 654                 if (device->silenced) {
 655                         memset(&samples_card[device_spec][0], 0, samples_card[device_spec].size() * sizeof(float));
 656                 } else {
 657                         device->resampling_queue->get_output_samples(
 658                                 ts,
 659                                 &samples_card[device_spec][0],
 660                                 num_samples,
 661                                 rate_adjustment_policy);
 662                 }
 663         }
 664
 665         vector<float> samples_out, left, right;
 666         samples_out.resize(num_samples * 2);
 667         samples_bus.resize(num_samples * 2);
 668         for (unsigned bus_index = 0; bus_index < input_mapping.buses.size(); ++bus_index) {
 669                 fill_audio_bus(samples_card, input_mapping.buses[bus_index], num_samples, stereo_width[bus_index], &samples_bus[0]);
 670                 apply_eq(bus_index, &samples_bus);
 671
 672                 {
 673                         lock_guard<mutex> lock(compressor_mutex);
 674
 675                         // Apply a level compressor to get the general level right.
 676                         // Basically, if it's over about -40 dBFS, we squeeze it down to that level
 677                         // (or more precisely, near it, since we don't use infinite ratio),
 678                         // then apply a makeup gain to get it to -14 dBFS. -14 dBFS is, of course,
 679                         // entirely arbitrary, but from practical tests with speech, it seems to
 680                         // put ut around -23 LUFS, so it's a reasonable starting point for later use.
 681                         if (level_compressor_enabled[bus_index]) {
 682                                 float threshold = 0.01f;   // -40 dBFS.
 683                                 float ratio = 20.0f;
 684                                 float attack_time = 0.5f;
 685                                 float release_time = 20.0f;
 686                                 float makeup_gain = from_db(ref_level_dbfs - (-40.0f));  // +26 dB.
 687                                 level_compressor[bus_index]->process(samples_bus.data(), samples_bus.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 688                                 gain_staging_db[bus_index] = to_db(level_compressor[bus_index]->get_attenuation() * makeup_gain);
 689                         } else {
 690                                 // Just apply the gain we already had.
 691                                 float db = gain_staging_db[bus_index];
 692                                 float last_db = last_gain_staging_db[bus_index];
 693                                 apply_gain(db, last_db, &samples_bus);
 694                         }
 695                         last_gain_staging_db[bus_index] = gain_staging_db[bus_index];
 696
 697 #if 0
 698                         printf("level=%f (%+5.2f dBFS) attenuation=%f (%+5.2f dB) end_result=%+5.2f dB\n",
 699                                 level_compressor.get_level(), to_db(level_compressor.get_level()),
 700                                 level_compressor.get_attenuation(), to_db(level_compressor.get_attenuation()),
 701                                 to_db(level_compressor.get_level() * level_compressor.get_attenuation() * makeup_gain));
 702 #endif
 703
 704                         // The real compressor.
 705                         if (compressor_enabled[bus_index]) {
 706                                 float threshold = from_db(compressor_threshold_dbfs[bus_index]);
 707                                 float ratio = 20.0f;
 708                                 float attack_time = 0.005f;
 709                                 float release_time = 0.040f;
 710                                 float makeup_gain = 2.0f;  // +6 dB.
 711                                 compressor[bus_index]->process(samples_bus.data(), samples_bus.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 712                 //              compressor_att = compressor.get_attenuation();
 713                         }
 714                 }
 715
 716                 add_bus_to_master(bus_index, samples_bus, &samples_out);
 717                 deinterleave_samples(samples_bus, &left, &right);
 718                 measure_bus_levels(bus_index, left, right);
 719         }
 720
 721         {
 722                 lock_guard<mutex> lock(compressor_mutex);
 723
 724                 // Finally a limiter at -4 dB (so, -10 dBFS) to take out the worst peaks only.
 725                 // Note that since ratio is not infinite, we could go slightly higher than this.
 726                 if (limiter_enabled) {
 727                         float threshold = from_db(limiter_threshold_dbfs);
 728                         float ratio = 30.0f;
 729                         float attack_time = 0.0f;  // Instant.
 730                         float release_time = 0.020f;
 731                         float makeup_gain = 1.0f;  // 0 dB.
 732                         limiter.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 733         //              limiter_att = limiter.get_attenuation();
 734                 }
 735
 736         //      printf("limiter=%+5.1f  compressor=%+5.1f\n", to_db(limiter_att), to_db(compressor_att));
 737         }
 738
 739         // At this point, we are most likely close to +0 LU (at least if the
 740         // faders sum to 0 dB and the compressors are on), but all of our
 741         // measurements have been on raw sample values, not R128 values.
 742         // So we have a final makeup gain to get us to +0 LU; the gain
 743         // adjustments required should be relatively small, and also, the
 744         // offset shouldn't change much (only if the type of audio changes
 745         // significantly). Thus, we shoot for updating this value basically
 746         // “whenever we process buffers”, since the R128 calculation isn't exactly
 747         // something we get out per-sample.
 748         //
 749         // Note that there's a feedback loop here, so we choose a very slow filter
 750         // (half-time of 30 seconds).
 751         double target_loudness_factor, alpha;
 752         double loudness_lu = r128.loudness_M() - ref_level_lufs;
 753         target_loudness_factor = final_makeup_gain * from_db(-loudness_lu);
 754
 755         // If we're outside +/- 5 LU (after correction), we don't count it as
 756         // a normal signal (probably silence) and don't change the
 757         // correction factor; just apply what we already have.
 758         if (fabs(loudness_lu) >= 5.0 || !final_makeup_gain_auto) {
 759                 alpha = 0.0;
 760         } else {
 761                 // Formula adapted from
 762                 // https://en.wikipedia.org/wiki/Low-pass_filter#Simple_infinite_impulse_response_filter.
 763                 const double half_time_s = 30.0;
 764                 const double fc_mul_2pi_delta_t = 1.0 / (half_time_s * OUTPUT_FREQUENCY);
 765                 alpha = fc_mul_2pi_delta_t / (fc_mul_2pi_delta_t + 1.0);
 766         }
 767
 768         {
 769                 lock_guard<mutex> lock(compressor_mutex);
 770                 double m = final_makeup_gain;
 771                 for (size_t i = 0; i < samples_out.size(); i += 2) {
 772                         samples_out[i + 0] *= m;
 773                         samples_out[i + 1] *= m;
 774                         m += (target_loudness_factor - m) * alpha;
 775                 }
 776                 final_makeup_gain = m;
 777         }
 778
 779         update_meters(samples_out);
 780
 781         return samples_out;
 782 }
 783
 784 namespace {
 785
 786 void apply_filter_fade(StereoFilter *filter, float *data, unsigned num_samples, float cutoff_hz, float db, float last_db)
 787 {
 788         // A granularity of 32 samples is an okay tradeoff between speed and
 789         // smoothness; recalculating the filters is pretty expensive, so it's
 790         // good that we don't do this all the time.
 791         static constexpr unsigned filter_granularity_samples = 32;
 792
 793         const float cutoff_linear = cutoff_hz * 2.0 * M_PI / OUTPUT_FREQUENCY;
 794         if (fabs(db - last_db) < 1e-3) {
 795                 // Constant over this frame.
 796                 if (fabs(db) > 0.01f) {
 797                         filter->render(data, num_samples, cutoff_linear, 0.5f, db / 40.0f);
 798                 }
 799         } else {
 800                 // We need to do a fade. (Rounding up avoids division by zero.)
 801                 unsigned num_blocks = (num_samples + filter_granularity_samples - 1) / filter_granularity_samples;
 802                 const float inc_db_norm = (db - last_db) / 40.0f / num_blocks;
 803                 float db_norm = db / 40.0f;
 804                 for (size_t i = 0; i < num_samples; i += filter_granularity_samples) {
 805                         size_t samples_this_block = std::min<size_t>(num_samples - i, filter_granularity_samples);
 806                         filter->render(data + i * 2, samples_this_block, cutoff_linear, 0.5f, db_norm);
 807                         db_norm += inc_db_norm;
 808                 }
 809         }
 810 }
 811
 812 }  // namespace
 813
 814 void AudioMixer::apply_eq(unsigned bus_index, vector<float> *samples_bus)
 815 {
 816         constexpr float bass_freq_hz = 200.0f;
 817         constexpr float treble_freq_hz = 4700.0f;
 818
 819         // Cut away everything under 120 Hz (or whatever the cutoff is);
 820         // we don't need it for voice, and it will reduce headroom
 821         // and confuse the compressor. (In particular, any hums at 50 or 60 Hz
 822         // should be dampened.)
 823         if (locut_enabled[bus_index]) {
 824                 locut[bus_index].render(samples_bus->data(), samples_bus->size() / 2, locut_cutoff_hz * 2.0 * M_PI / OUTPUT_FREQUENCY, 0.5f);
 825         }
 826
 827         // Apply the rest of the EQ. Since we only have a simple three-band EQ,
 828         // we can implement it with two shelf filters. We use a simple gain to
 829         // set the mid-level filter, and then offset the low and high bands
 830         // from that if we need to. (We could perhaps have folded the gain into
 831         // the next part, but it's so cheap that the trouble isn't worth it.)
 832         //
 833         // If any part of the EQ has changed appreciably since last frame,
 834         // we fade smoothly during the course of this frame.
 835         const float bass_db = eq_level_db[bus_index][EQ_BAND_BASS];
 836         const float mid_db = eq_level_db[bus_index][EQ_BAND_MID];
 837         const float treble_db = eq_level_db[bus_index][EQ_BAND_TREBLE];
 838
 839         const float last_bass_db = last_eq_level_db[bus_index][EQ_BAND_BASS];
 840         const float last_mid_db = last_eq_level_db[bus_index][EQ_BAND_MID];
 841         const float last_treble_db = last_eq_level_db[bus_index][EQ_BAND_TREBLE];
 842
 843         assert(samples_bus->size() % 2 == 0);
 844         const unsigned num_samples = samples_bus->size() / 2;
 845
 846         apply_gain(mid_db, last_mid_db, samples_bus);
 847
 848         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);
 849         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);
 850
 851         last_eq_level_db[bus_index][EQ_BAND_BASS] = bass_db;
 852         last_eq_level_db[bus_index][EQ_BAND_MID] = mid_db;
 853         last_eq_level_db[bus_index][EQ_BAND_TREBLE] = treble_db;
 854 }
 855
 856 void AudioMixer::add_bus_to_master(unsigned bus_index, const vector<float> &samples_bus, vector<float> *samples_out)
 857 {
 858         assert(samples_bus.size() == samples_out->size());
 859         assert(samples_bus.size() % 2 == 0);
 860         unsigned num_samples = samples_bus.size() / 2;
 861         const float new_volume_db = mute[bus_index] ? -90.0f : fader_volume_db[bus_index].load();
 862         if (fabs(new_volume_db - last_fader_volume_db[bus_index]) > 1e-3) {
 863                 // The volume has changed; do a fade over the course of this frame.
 864                 // (We might have some numerical issues here, but it seems to sound OK.)
 865                 // For the purpose of fading here, the silence floor is set to -90 dB
 866                 // (the fader only goes to -84).
 867                 float old_volume = from_db(max<float>(last_fader_volume_db[bus_index], -90.0f));
 868                 float volume = from_db(max<float>(new_volume_db, -90.0f));
 869
 870                 float volume_inc = pow(volume / old_volume, 1.0 / num_samples);
 871                 volume = old_volume;
 872                 if (bus_index == 0) {
 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                                 volume *= volume_inc;
 877                         }
 878                 } else {
 879                         for (unsigned i = 0; i < num_samples; ++i) {
 880                                 (*samples_out)[i * 2 + 0] += samples_bus[i * 2 + 0] * volume;
 881                                 (*samples_out)[i * 2 + 1] += samples_bus[i * 2 + 1] * volume;
 882                                 volume *= volume_inc;
 883                         }
 884                 }
 885         } else if (new_volume_db > -90.0f) {
 886                 float volume = from_db(new_volume_db);
 887                 if (bus_index == 0) {
 888                         for (unsigned i = 0; i < num_samples; ++i) {
 889                                 (*samples_out)[i * 2 + 0] = samples_bus[i * 2 + 0] * volume;
 890                                 (*samples_out)[i * 2 + 1] = samples_bus[i * 2 + 1] * volume;
 891                         }
 892                 } else {
 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                 }
 898         }
 899
 900         last_fader_volume_db[bus_index] = new_volume_db;
 901 }
 902
 903 void AudioMixer::measure_bus_levels(unsigned bus_index, const vector<float> &left, const vector<float> &right)
 904 {
 905         assert(left.size() == right.size());
 906         const float volume = mute[bus_index] ? 0.0f : from_db(fader_volume_db[bus_index]);
 907         const float peak_levels[2] = {
 908                 find_peak(left.data(), left.size()) * volume,
 909                 find_peak(right.data(), right.size()) * volume
 910         };
 911         for (unsigned channel = 0; channel < 2; ++channel) {
 912                 // Compute the current value, including hold and falloff.
 913                 // The constants are borrowed from zita-mu1 by Fons Adriaensen.
 914                 static constexpr float hold_sec = 0.5f;
 915                 static constexpr float falloff_db_sec = 15.0f;  // dB/sec falloff after hold.
 916                 float current_peak;
 917                 PeakHistory &history = peak_history[bus_index][channel];
 918                 history.historic_peak = max(history.historic_peak, peak_levels[channel]);
 919                 if (history.age_seconds < hold_sec) {
 920                         current_peak = history.last_peak;
 921                 } else {
 922                         current_peak = history.last_peak * from_db(-falloff_db_sec * (history.age_seconds - hold_sec));
 923                 }
 924
 925                 // See if we have a new peak to replace the old (possibly falling) one.
 926                 if (peak_levels[channel] > current_peak) {
 927                         history.last_peak = peak_levels[channel];
 928                         history.age_seconds = 0.0f;  // Not 100% correct, but more than good enough given our frame sizes.
 929                         current_peak = peak_levels[channel];
 930                 } else {
 931                         history.age_seconds += float(left.size()) / OUTPUT_FREQUENCY;
 932                 }
 933                 history.current_level = peak_levels[channel];
 934                 history.current_peak = current_peak;
 935         }
 936 }
 937
 938 void AudioMixer::update_meters(const vector<float> &samples)
 939 {
 940         // Upsample 4x to find interpolated peak.
 941         peak_resampler.inp_data = const_cast<float *>(samples.data());
 942         peak_resampler.inp_count = samples.size() / 2;
 943
 944         vector<float> interpolated_samples;
 945         interpolated_samples.resize(samples.size());
 946         {
 947                 lock_guard<mutex> lock(audio_measure_mutex);
 948
 949                 while (peak_resampler.inp_count > 0) {  // About four iterations.
 950                         peak_resampler.out_data = &interpolated_samples[0];
 951                         peak_resampler.out_count = interpolated_samples.size() / 2;
 952                         peak_resampler.process();
 953                         size_t out_stereo_samples = interpolated_samples.size() / 2 - peak_resampler.out_count;
 954                         peak = max<float>(peak, find_peak(interpolated_samples.data(), out_stereo_samples * 2));
 955                         peak_resampler.out_data = nullptr;
 956                 }
 957         }
 958
 959         // Find R128 levels and L/R correlation.
 960         vector<float> left, right;
 961         deinterleave_samples(samples, &left, &right);
 962         float *ptrs[] = { left.data(), right.data() };
 963         {
 964                 lock_guard<mutex> lock(audio_measure_mutex);
 965                 r128.process(left.size(), ptrs);
 966                 correlation.process_samples(samples);
 967         }
 968
 969         send_audio_level_callback();
 970 }
 971
 972 void AudioMixer::reset_meters()
 973 {
 974         lock_guard<mutex> lock(audio_measure_mutex);
 975         peak_resampler.reset();
 976         peak = 0.0f;
 977         r128.reset();
 978         r128.integr_start();
 979         correlation.reset();
 980 }
 981
 982 void AudioMixer::send_audio_level_callback()
 983 {
 984         if (audio_level_callback == nullptr) {
 985                 return;
 986         }
 987
 988         lock_guard<mutex> lock(audio_measure_mutex);
 989         double loudness_s = r128.loudness_S();
 990         double loudness_i = r128.integrated();
 991         double loudness_range_low = r128.range_min();
 992         double loudness_range_high = r128.range_max();
 993
 994         metric_audio_loudness_short_lufs = loudness_s;
 995         metric_audio_loudness_integrated_lufs = loudness_i;
 996         metric_audio_loudness_range_low_lufs = loudness_range_low;
 997         metric_audio_loudness_range_high_lufs = loudness_range_high;
 998         metric_audio_peak_dbfs = to_db(peak);
 999         metric_audio_final_makeup_gain_db = to_db(final_makeup_gain);
1000         metric_audio_correlation = correlation.get_correlation();
1001
1002         vector<BusLevel> bus_levels;
1003         bus_levels.resize(input_mapping.buses.size());
1004         {
1005                 lock_guard<mutex> lock(compressor_mutex);
1006                 for (unsigned bus_index = 0; bus_index < bus_levels.size(); ++bus_index) {
1007                         BusLevel &levels = bus_levels[bus_index];
1008                         BusMetrics &metrics = bus_metrics[bus_index];
1009
1010                         levels.current_level_dbfs[0] = metrics.current_level_dbfs[0] = to_db(peak_history[bus_index][0].current_level);
1011                         levels.current_level_dbfs[1] = metrics.current_level_dbfs[1] = to_db(peak_history[bus_index][1].current_level);
1012                         levels.peak_level_dbfs[0] = metrics.peak_level_dbfs[0] = to_db(peak_history[bus_index][0].current_peak);
1013                         levels.peak_level_dbfs[1] = metrics.peak_level_dbfs[1] = to_db(peak_history[bus_index][1].current_peak);
1014                         levels.historic_peak_dbfs = metrics.historic_peak_dbfs = to_db(
1015                                 max(peak_history[bus_index][0].historic_peak,
1016                                     peak_history[bus_index][1].historic_peak));
1017                         levels.gain_staging_db = metrics.gain_staging_db = gain_staging_db[bus_index];
1018                         if (compressor_enabled[bus_index]) {
1019                                 levels.compressor_attenuation_db = metrics.compressor_attenuation_db = -to_db(compressor[bus_index]->get_attenuation());
1020                         } else {
1021                                 levels.compressor_attenuation_db = 0.0;
1022                                 metrics.compressor_attenuation_db = 0.0 / 0.0;
1023                         }
1024                 }
1025         }
1026
1027         audio_level_callback(loudness_s, to_db(peak), bus_levels,
1028                 loudness_i, loudness_range_low, loudness_range_high,
1029                 to_db(final_makeup_gain),
1030                 correlation.get_correlation());
1031 }
1032
1033 map<DeviceSpec, DeviceInfo> AudioMixer::get_devices()
1034 {
1035         lock_guard<timed_mutex> lock(audio_mutex);
1036
1037         map<DeviceSpec, DeviceInfo> devices;
1038         for (unsigned card_index = 0; card_index < num_capture_cards; ++card_index) {
1039                 const DeviceSpec spec{ InputSourceType::CAPTURE_CARD, card_index };
1040                 const AudioDevice *device = &video_cards[card_index];
1041                 DeviceInfo info;
1042                 info.display_name = device->display_name;
1043                 info.num_channels = 8;
1044                 devices.insert(make_pair(spec, info));
1045         }
1046         vector<ALSAPool::Device> available_alsa_devices = alsa_pool.get_devices();
1047         for (unsigned card_index = 0; card_index < available_alsa_devices.size(); ++card_index) {
1048                 const DeviceSpec spec{ InputSourceType::ALSA_INPUT, card_index };
1049                 const ALSAPool::Device &device = available_alsa_devices[card_index];
1050                 DeviceInfo info;
1051                 info.display_name = device.display_name();
1052                 info.num_channels = device.num_channels;
1053                 info.alsa_name = device.name;
1054                 info.alsa_info = device.info;
1055                 info.alsa_address = device.address;
1056                 devices.insert(make_pair(spec, info));
1057         }
1058         for (unsigned card_index = 0; card_index < num_ffmpeg_inputs; ++card_index) {
1059                 const DeviceSpec spec{ InputSourceType::FFMPEG_VIDEO_INPUT, card_index };
1060                 const AudioDevice *device = &ffmpeg_inputs[card_index];
1061                 DeviceInfo info;
1062                 info.display_name = device->display_name;
1063                 info.num_channels = 2;
1064                 devices.insert(make_pair(spec, info));
1065         }
1066         return devices;
1067 }
1068
1069 void AudioMixer::set_display_name(DeviceSpec device_spec, const string &name)
1070 {
1071         AudioDevice *device = find_audio_device(device_spec);
1072
1073         lock_guard<timed_mutex> lock(audio_mutex);
1074         device->display_name = name;
1075 }
1076
1077 void AudioMixer::serialize_device(DeviceSpec device_spec, DeviceSpecProto *device_spec_proto)
1078 {
1079         lock_guard<timed_mutex> lock(audio_mutex);
1080         switch (device_spec.type) {
1081                 case InputSourceType::SILENCE:
1082                         device_spec_proto->set_type(DeviceSpecProto::SILENCE);
1083                         break;
1084                 case InputSourceType::CAPTURE_CARD:
1085                         device_spec_proto->set_type(DeviceSpecProto::CAPTURE_CARD);
1086                         device_spec_proto->set_index(device_spec.index);
1087                         device_spec_proto->set_display_name(video_cards[device_spec.index].display_name);
1088                         break;
1089                 case InputSourceType::ALSA_INPUT:
1090                         alsa_pool.serialize_device(device_spec.index, device_spec_proto);
1091                         break;
1092                 case InputSourceType::FFMPEG_VIDEO_INPUT:
1093                         device_spec_proto->set_type(DeviceSpecProto::FFMPEG_VIDEO_INPUT);
1094                         device_spec_proto->set_index(device_spec.index);
1095                         device_spec_proto->set_display_name(ffmpeg_inputs[device_spec.index].display_name);
1096                         break;
1097         }
1098 }
1099
1100 void AudioMixer::set_simple_input(unsigned card_index)
1101 {
1102         assert(card_index < num_capture_cards + num_ffmpeg_inputs);
1103         InputMapping new_input_mapping;
1104         InputMapping::Bus input;
1105         input.name = "Main";
1106         if (card_index >= num_capture_cards) {
1107                 input.device = DeviceSpec{InputSourceType::FFMPEG_VIDEO_INPUT, card_index - num_capture_cards};
1108         } else {
1109                 input.device = DeviceSpec{InputSourceType::CAPTURE_CARD, card_index};
1110         }
1111         input.source_channel[0] = 0;
1112         input.source_channel[1] = 1;
1113
1114         new_input_mapping.buses.push_back(input);
1115
1116         lock_guard<timed_mutex> lock(audio_mutex);
1117         current_mapping_mode = MappingMode::SIMPLE;
1118         set_input_mapping_lock_held(new_input_mapping);
1119         fader_volume_db[0] = 0.0f;
1120 }
1121
1122 unsigned AudioMixer::get_simple_input() const
1123 {
1124         lock_guard<timed_mutex> lock(audio_mutex);
1125         if (input_mapping.buses.size() == 1 &&
1126             input_mapping.buses[0].device.type == InputSourceType::CAPTURE_CARD &&
1127             input_mapping.buses[0].source_channel[0] == 0 &&
1128             input_mapping.buses[0].source_channel[1] == 1) {
1129                 return input_mapping.buses[0].device.index;
1130         } else if (input_mapping.buses.size() == 1 &&
1131                    input_mapping.buses[0].device.type == InputSourceType::FFMPEG_VIDEO_INPUT &&
1132                    input_mapping.buses[0].source_channel[0] == 0 &&
1133                    input_mapping.buses[0].source_channel[1] == 1) {
1134                 return input_mapping.buses[0].device.index + num_capture_cards;
1135         } else {
1136                 return numeric_limits<unsigned>::max();
1137         }
1138 }
1139
1140 void AudioMixer::set_input_mapping(const InputMapping &new_input_mapping)
1141 {
1142         lock_guard<timed_mutex> lock(audio_mutex);
1143         set_input_mapping_lock_held(new_input_mapping);
1144         current_mapping_mode = MappingMode::MULTICHANNEL;
1145 }
1146
1147 AudioMixer::MappingMode AudioMixer::get_mapping_mode() const
1148 {
1149         lock_guard<timed_mutex> lock(audio_mutex);
1150         return current_mapping_mode;
1151 }
1152
1153 void AudioMixer::set_input_mapping_lock_held(const InputMapping &new_input_mapping)
1154 {
1155         map<DeviceSpec, set<unsigned>> interesting_channels;
1156         for (const InputMapping::Bus &bus : new_input_mapping.buses) {
1157                 if (bus.device.type == InputSourceType::CAPTURE_CARD ||
1158                     bus.device.type == InputSourceType::ALSA_INPUT ||
1159                     bus.device.type == InputSourceType::FFMPEG_VIDEO_INPUT) {
1160                         for (unsigned channel = 0; channel < 2; ++channel) {
1161                                 if (bus.source_channel[channel] != -1) {
1162                                         interesting_channels[bus.device].insert(bus.source_channel[channel]);
1163                                 }
1164                         }
1165                 } else {
1166                         assert(bus.device.type == InputSourceType::SILENCE);
1167                 }
1168         }
1169
1170         // Kill all the old metrics, and set up new ones.
1171         for (unsigned bus_index = 0; bus_index < input_mapping.buses.size(); ++bus_index) {
1172                 BusMetrics &metrics = bus_metrics[bus_index];
1173
1174                 vector<pair<string, string>> labels_left = metrics.labels;
1175                 labels_left.emplace_back("channel", "left");
1176                 vector<pair<string, string>> labels_right = metrics.labels;
1177                 labels_right.emplace_back("channel", "right");
1178
1179                 global_metrics.remove("bus_current_level_dbfs", labels_left);
1180                 global_metrics.remove("bus_current_level_dbfs", labels_right);
1181                 global_metrics.remove("bus_peak_level_dbfs", labels_left);
1182                 global_metrics.remove("bus_peak_level_dbfs", labels_right);
1183                 global_metrics.remove("bus_historic_peak_dbfs", metrics.labels);
1184                 global_metrics.remove("bus_gain_staging_db", metrics.labels);
1185                 global_metrics.remove("bus_compressor_attenuation_db", metrics.labels);
1186         }
1187         bus_metrics.reset(new BusMetrics[new_input_mapping.buses.size()]);
1188         for (unsigned bus_index = 0; bus_index < new_input_mapping.buses.size(); ++bus_index) {
1189                 const InputMapping::Bus &bus = new_input_mapping.buses[bus_index];
1190                 BusMetrics &metrics = bus_metrics[bus_index];
1191
1192                 char bus_index_str[16], source_index_str[16], source_channels_str[64];
1193                 snprintf(bus_index_str, sizeof(bus_index_str), "%u", bus_index);
1194                 snprintf(source_index_str, sizeof(source_index_str), "%u", bus.device.index);
1195                 snprintf(source_channels_str, sizeof(source_channels_str), "%d:%d", bus.source_channel[0], bus.source_channel[1]);
1196
1197                 vector<pair<string, string>> labels;
1198                 metrics.labels.emplace_back("index", bus_index_str);
1199                 metrics.labels.emplace_back("name", bus.name);
1200                 if (bus.device.type == InputSourceType::SILENCE) {
1201                         metrics.labels.emplace_back("source_type", "silence");
1202                 } else if (bus.device.type == InputSourceType::CAPTURE_CARD) {
1203                         metrics.labels.emplace_back("source_type", "capture_card");
1204                 } else if (bus.device.type == InputSourceType::ALSA_INPUT) {
1205                         metrics.labels.emplace_back("source_type", "alsa_input");
1206                 } else if (bus.device.type == InputSourceType::FFMPEG_VIDEO_INPUT) {
1207                         metrics.labels.emplace_back("source_type", "ffmpeg_video_input");
1208                 } else {
1209                         assert(false);
1210                 }
1211                 metrics.labels.emplace_back("source_index", source_index_str);
1212                 metrics.labels.emplace_back("source_channels", source_channels_str);
1213
1214                 vector<pair<string, string>> labels_left = metrics.labels;
1215                 labels_left.emplace_back("channel", "left");
1216                 vector<pair<string, string>> labels_right = metrics.labels;
1217                 labels_right.emplace_back("channel", "right");
1218
1219                 global_metrics.add("bus_current_level_dbfs", labels_left, &metrics.current_level_dbfs[0], Metrics::TYPE_GAUGE);
1220                 global_metrics.add("bus_current_level_dbfs", labels_right, &metrics.current_level_dbfs[1], Metrics::TYPE_GAUGE);
1221                 global_metrics.add("bus_peak_level_dbfs", labels_left, &metrics.peak_level_dbfs[0], Metrics::TYPE_GAUGE);
1222                 global_metrics.add("bus_peak_level_dbfs", labels_right, &metrics.peak_level_dbfs[1], Metrics::TYPE_GAUGE);
1223                 global_metrics.add("bus_historic_peak_dbfs", metrics.labels, &metrics.historic_peak_dbfs, Metrics::TYPE_GAUGE);
1224                 global_metrics.add("bus_gain_staging_db", metrics.labels, &metrics.gain_staging_db, Metrics::TYPE_GAUGE);
1225                 global_metrics.add("bus_compressor_attenuation_db", metrics.labels, &metrics.compressor_attenuation_db, Metrics::TYPE_GAUGE);
1226         }
1227
1228         // Reset resamplers for all cards that don't have the exact same state as before.
1229         map<DeviceSpec, double> new_extra_delay_ms = new_input_mapping.extra_delay_ms;  // Convenience so we can use [].
1230         for (unsigned card_index = 0; card_index < MAX_VIDEO_CARDS; ++card_index) {
1231                 const DeviceSpec device_spec{InputSourceType::CAPTURE_CARD, card_index};
1232                 AudioDevice *device = find_audio_device(device_spec);
1233                 if (device->interesting_channels != interesting_channels[device_spec] ||
1234                     device->extra_delay_ms != new_extra_delay_ms[device_spec]) {
1235                         device->interesting_channels = interesting_channels[device_spec];
1236                         device->extra_delay_ms = new_extra_delay_ms[device_spec];
1237                         reset_resampler_mutex_held(device_spec);
1238                 }
1239         }
1240         for (unsigned card_index = 0; card_index < MAX_ALSA_CARDS; ++card_index) {
1241                 const DeviceSpec device_spec{InputSourceType::ALSA_INPUT, card_index};
1242                 AudioDevice *device = find_audio_device(device_spec);
1243                 if (interesting_channels[device_spec].empty()) {
1244                         alsa_pool.release_device(card_index);
1245                 } else {
1246                         alsa_pool.hold_device(card_index);
1247                 }
1248                 if (device->interesting_channels != interesting_channels[device_spec] ||
1249                     device->extra_delay_ms != new_extra_delay_ms[device_spec]) {
1250                         device->interesting_channels = interesting_channels[device_spec];
1251                         device->extra_delay_ms = new_extra_delay_ms[device_spec];
1252                         alsa_pool.reset_device(device_spec.index);
1253                         reset_resampler_mutex_held(device_spec);
1254                 }
1255         }
1256         for (unsigned card_index = 0; card_index < num_ffmpeg_inputs; ++card_index) {
1257                 const DeviceSpec device_spec{InputSourceType::FFMPEG_VIDEO_INPUT, card_index};
1258                 AudioDevice *device = find_audio_device(device_spec);
1259                 if (device->interesting_channels != interesting_channels[device_spec] ||
1260                     device->extra_delay_ms != new_extra_delay_ms[device_spec]) {
1261                         device->interesting_channels = interesting_channels[device_spec];
1262                         device->extra_delay_ms = new_extra_delay_ms[device_spec];
1263                         reset_resampler_mutex_held(device_spec);
1264                 }
1265         }
1266
1267         input_mapping = new_input_mapping;
1268 }
1269
1270 InputMapping AudioMixer::get_input_mapping() const
1271 {
1272         lock_guard<timed_mutex> lock(audio_mutex);
1273         return input_mapping;
1274 }
1275
1276 unsigned AudioMixer::num_buses() const
1277 {
1278         lock_guard<timed_mutex> lock(audio_mutex);
1279         return input_mapping.buses.size();
1280 }
1281
1282 void AudioMixer::reset_peak(unsigned bus_index)
1283 {
1284         lock_guard<timed_mutex> lock(audio_mutex);
1285         for (unsigned channel = 0; channel < 2; ++channel) {
1286                 PeakHistory &history = peak_history[bus_index][channel];
1287                 history.current_level = 0.0f;
1288                 history.historic_peak = 0.0f;
1289                 history.current_peak = 0.0f;
1290                 history.last_peak = 0.0f;
1291                 history.age_seconds = 0.0f;
1292         }
1293 }
1294
1295 bool AudioMixer::is_mono(unsigned bus_index)
1296 {
1297         lock_guard<timed_mutex> lock(audio_mutex);
1298         const InputMapping::Bus &bus = input_mapping.buses[bus_index];
1299         if (bus.device.type == InputSourceType::SILENCE) {
1300                 return true;
1301         } else {
1302                 assert(bus.device.type == InputSourceType::CAPTURE_CARD ||
1303                        bus.device.type == InputSourceType::ALSA_INPUT ||
1304                        bus.device.type == InputSourceType::FFMPEG_VIDEO_INPUT);
1305                 return bus.source_channel[0] == bus.source_channel[1];
1306         }
1307 }
1308
1309 AudioMixer *global_audio_mixer = nullptr;