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

   1 #include "audio_mixer.h"
   2
   3 #include <assert.h>
   4 #include <endian.h>
   5 #include <bmusb/bmusb.h>
   6 #include <stdio.h>
   7 #include <endian.h>
   8 #include <cmath>
   9 #ifdef __SSE__
  10 #include <immintrin.h>
  11 #endif
  12
  13 #include "db.h"
  14 #include "flags.h"
  15 #include "mixer.h"
  16 #include "state.pb.h"
  17 #include "timebase.h"
  18
  19 using namespace bmusb;
  20 using namespace std;
  21 using namespace std::placeholders;
  22
  23 namespace {
  24
  25 // TODO: If these prove to be a bottleneck, they can be SSSE3-optimized
  26 // (usually including multiple channels at a time).
  27
  28 void convert_fixed16_to_fp32(float *dst, size_t out_channel, size_t out_num_channels,
  29                              const uint8_t *src, size_t in_channel, size_t in_num_channels,
  30                              size_t num_samples)
  31 {
  32         assert(in_channel < in_num_channels);
  33         assert(out_channel < out_num_channels);
  34         src += in_channel * 2;
  35         dst += out_channel;
  36
  37         for (size_t i = 0; i < num_samples; ++i) {
  38                 int16_t s = le16toh(*(int16_t *)src);
  39                 *dst = s * (1.0f / 32768.0f);
  40
  41                 src += 2 * in_num_channels;
  42                 dst += out_num_channels;
  43         }
  44 }
  45
  46 void convert_fixed24_to_fp32(float *dst, size_t out_channel, size_t out_num_channels,
  47                              const uint8_t *src, size_t in_channel, size_t in_num_channels,
  48                              size_t num_samples)
  49 {
  50         assert(in_channel < in_num_channels);
  51         assert(out_channel < out_num_channels);
  52         src += in_channel * 3;
  53         dst += out_channel;
  54
  55         for (size_t i = 0; i < num_samples; ++i) {
  56                 uint32_t s1 = src[0];
  57                 uint32_t s2 = src[1];
  58                 uint32_t s3 = src[2];
  59                 uint32_t s = s1 | (s1 << 8) | (s2 << 16) | (s3 << 24);
  60                 *dst = int(s) * (1.0f / 2147483648.0f);
  61
  62                 src += 3 * in_num_channels;
  63                 dst += out_num_channels;
  64         }
  65 }
  66
  67 void convert_fixed32_to_fp32(float *dst, size_t out_channel, size_t out_num_channels,
  68                              const uint8_t *src, size_t in_channel, size_t in_num_channels,
  69                              size_t num_samples)
  70 {
  71         assert(in_channel < in_num_channels);
  72         assert(out_channel < out_num_channels);
  73         src += in_channel * 4;
  74         dst += out_channel;
  75
  76         for (size_t i = 0; i < num_samples; ++i) {
  77                 int32_t s = le32toh(*(int32_t *)src);
  78                 *dst = s * (1.0f / 2147483648.0f);
  79
  80                 src += 4 * in_num_channels;
  81                 dst += out_num_channels;
  82         }
  83 }
  84
  85 float find_peak_plain(const float *samples, size_t num_samples) __attribute__((unused));
  86
  87 float find_peak_plain(const float *samples, size_t num_samples)
  88 {
  89         float m = fabs(samples[0]);
  90         for (size_t i = 1; i < num_samples; ++i) {
  91                 m = max(m, fabs(samples[i]));
  92         }
  93         return m;
  94 }
  95
  96 #ifdef __SSE__
  97 static inline float horizontal_max(__m128 m)
  98 {
  99         __m128 tmp = _mm_shuffle_ps(m, m, _MM_SHUFFLE(1, 0, 3, 2));
 100         m = _mm_max_ps(m, tmp);
 101         tmp = _mm_shuffle_ps(m, m, _MM_SHUFFLE(2, 3, 0, 1));
 102         m = _mm_max_ps(m, tmp);
 103         return _mm_cvtss_f32(m);
 104 }
 105
 106 float find_peak(const float *samples, size_t num_samples)
 107 {
 108         const __m128 abs_mask = _mm_castsi128_ps(_mm_set1_epi32(0x7fffffffu));
 109         __m128 m = _mm_setzero_ps();
 110         for (size_t i = 0; i < (num_samples & ~3); i += 4) {
 111                 __m128 x = _mm_loadu_ps(samples + i);
 112                 x = _mm_and_ps(x, abs_mask);
 113                 m = _mm_max_ps(m, x);
 114         }
 115         float result = horizontal_max(m);
 116
 117         for (size_t i = (num_samples & ~3); i < num_samples; ++i) {
 118                 result = max(result, fabs(samples[i]));
 119         }
 120
 121 #if 0
 122         // Self-test. We should be bit-exact the same.
 123         float reference_result = find_peak_plain(samples, num_samples);
 124         if (result != reference_result) {
 125                 fprintf(stderr, "Error: Peak is %f [%f %f %f %f]; should be %f.\n",
 126                         result,
 127                         _mm_cvtss_f32(_mm_shuffle_ps(m, m, _MM_SHUFFLE(0, 0, 0, 0))),
 128                         _mm_cvtss_f32(_mm_shuffle_ps(m, m, _MM_SHUFFLE(1, 1, 1, 1))),
 129                         _mm_cvtss_f32(_mm_shuffle_ps(m, m, _MM_SHUFFLE(2, 2, 2, 2))),
 130                         _mm_cvtss_f32(_mm_shuffle_ps(m, m, _MM_SHUFFLE(3, 3, 3, 3))),
 131                         reference_result);
 132                 abort();
 133         }
 134 #endif
 135         return result;
 136 }
 137 #else
 138 float find_peak(const float *samples, size_t num_samples)
 139 {
 140         return find_peak_plain(samples, num_samples);
 141 }
 142 #endif
 143
 144 void deinterleave_samples(const vector<float> &in, vector<float> *out_l, vector<float> *out_r)
 145 {
 146         size_t num_samples = in.size() / 2;
 147         out_l->resize(num_samples);
 148         out_r->resize(num_samples);
 149
 150         const float *inptr = in.data();
 151         float *lptr = &(*out_l)[0];
 152         float *rptr = &(*out_r)[0];
 153         for (size_t i = 0; i < num_samples; ++i) {
 154                 *lptr++ = *inptr++;
 155                 *rptr++ = *inptr++;
 156         }
 157 }
 158
 159 }  // namespace
 160
 161 AudioMixer::AudioMixer(unsigned num_cards)
 162         : num_cards(num_cards),
 163           limiter(OUTPUT_FREQUENCY),
 164           correlation(OUTPUT_FREQUENCY)
 165 {
 166         global_audio_mixer = this;
 167
 168         for (unsigned bus_index = 0; bus_index < MAX_BUSES; ++bus_index) {
 169                 locut[bus_index].init(FILTER_HPF, 2);
 170                 locut_enabled[bus_index] = global_flags.locut_enabled;
 171                 eq[bus_index][EQ_BAND_BASS].init(FILTER_LOW_SHELF, 1);
 172                 // Note: EQ_BAND_MID isn't used (see comments in apply_eq()).
 173                 eq[bus_index][EQ_BAND_TREBLE].init(FILTER_HIGH_SHELF, 1);
 174
 175                 gain_staging_db[bus_index] = global_flags.initial_gain_staging_db;
 176                 compressor[bus_index].reset(new StereoCompressor(OUTPUT_FREQUENCY));
 177                 compressor_threshold_dbfs[bus_index] = ref_level_dbfs - 12.0f;  // -12 dB.
 178                 compressor_enabled[bus_index] = global_flags.compressor_enabled;
 179                 level_compressor[bus_index].reset(new StereoCompressor(OUTPUT_FREQUENCY));
 180                 level_compressor_enabled[bus_index] = global_flags.gain_staging_auto;
 181         }
 182         set_limiter_enabled(global_flags.limiter_enabled);
 183         set_final_makeup_gain_auto(global_flags.final_makeup_gain_auto);
 184         alsa_pool.init();
 185
 186         InputMapping new_input_mapping;
 187         if (!global_flags.input_mapping_filename.empty()) {
 188                 if (!load_input_mapping_from_file(get_devices(),
 189                                                   global_flags.input_mapping_filename,
 190                                                   &new_input_mapping)) {
 191                         fprintf(stderr, "Failed to load input mapping from '%s', exiting.\n",
 192                                 global_flags.input_mapping_filename.c_str());
 193                         exit(1);
 194                 }
 195         } else {
 196                 // Generate a very simple, default input mapping.
 197                 InputMapping::Bus input;
 198                 input.name = "Main";
 199                 input.device.type = InputSourceType::CAPTURE_CARD;
 200                 input.device.index = 0;
 201                 input.source_channel[0] = 0;
 202                 input.source_channel[1] = 1;
 203
 204                 new_input_mapping.buses.push_back(input);
 205         }
 206         set_input_mapping(new_input_mapping);
 207
 208         r128.init(2, OUTPUT_FREQUENCY);
 209         r128.integr_start();
 210
 211         // hlen=16 is pretty low quality, but we use quite a bit of CPU otherwise,
 212         // and there's a limit to how important the peak meter is.
 213         peak_resampler.setup(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY * 4, /*num_channels=*/2, /*hlen=*/16, /*frel=*/1.0);
 214 }
 215
 216 void AudioMixer::reset_resampler(DeviceSpec device_spec)
 217 {
 218         lock_guard<timed_mutex> lock(audio_mutex);
 219         reset_resampler_mutex_held(device_spec);
 220 }
 221
 222 void AudioMixer::reset_resampler_mutex_held(DeviceSpec device_spec)
 223 {
 224         AudioDevice *device = find_audio_device(device_spec);
 225
 226         if (device->interesting_channels.empty()) {
 227                 device->resampling_queue.reset();
 228         } else {
 229                 // TODO: ResamplingQueue should probably take the full device spec.
 230                 // (It's only used for console output, though.)
 231                 device->resampling_queue.reset(new ResamplingQueue(device_spec.index, device->capture_frequency, OUTPUT_FREQUENCY, device->interesting_channels.size()));
 232         }
 233         device->next_local_pts = 0;
 234 }
 235
 236 bool AudioMixer::add_audio(DeviceSpec device_spec, const uint8_t *data, unsigned num_samples, AudioFormat audio_format, int64_t frame_length)
 237 {
 238         AudioDevice *device = find_audio_device(device_spec);
 239
 240         unique_lock<timed_mutex> lock(audio_mutex, defer_lock);
 241         if (!lock.try_lock_for(chrono::milliseconds(10))) {
 242                 return false;
 243         }
 244         if (device->resampling_queue == nullptr) {
 245                 // No buses use this device; throw it away.
 246                 return true;
 247         }
 248
 249         unsigned num_channels = device->interesting_channels.size();
 250         assert(num_channels > 0);
 251
 252         // Convert the audio to fp32.
 253         unique_ptr<float[]> audio(new float[num_samples * num_channels]);
 254         unsigned channel_index = 0;
 255         for (auto channel_it = device->interesting_channels.cbegin(); channel_it != device->interesting_channels.end(); ++channel_it, ++channel_index) {
 256                 switch (audio_format.bits_per_sample) {
 257                 case 0:
 258                         assert(num_samples == 0);
 259                         break;
 260                 case 16:
 261                         convert_fixed16_to_fp32(audio.get(), channel_index, num_channels, data, *channel_it, audio_format.num_channels, num_samples);
 262                         break;
 263                 case 24:
 264                         convert_fixed24_to_fp32(audio.get(), channel_index, num_channels, data, *channel_it, audio_format.num_channels, num_samples);
 265                         break;
 266                 case 32:
 267                         convert_fixed32_to_fp32(audio.get(), channel_index, num_channels, data, *channel_it, audio_format.num_channels, num_samples);
 268                         break;
 269                 default:
 270                         fprintf(stderr, "Cannot handle audio with %u bits per sample\n", audio_format.bits_per_sample);
 271                         assert(false);
 272                 }
 273         }
 274
 275         // Now add it.
 276         int64_t local_pts = device->next_local_pts;
 277         device->resampling_queue->add_input_samples(local_pts / double(TIMEBASE), audio.get(), num_samples);
 278         device->next_local_pts = local_pts + frame_length;
 279         return true;
 280 }
 281
 282 bool AudioMixer::add_silence(DeviceSpec device_spec, unsigned samples_per_frame, unsigned num_frames, int64_t frame_length)
 283 {
 284         AudioDevice *device = find_audio_device(device_spec);
 285
 286         unique_lock<timed_mutex> lock(audio_mutex, defer_lock);
 287         if (!lock.try_lock_for(chrono::milliseconds(10))) {
 288                 return false;
 289         }
 290         if (device->resampling_queue == nullptr) {
 291                 // No buses use this device; throw it away.
 292                 return true;
 293         }
 294
 295         unsigned num_channels = device->interesting_channels.size();
 296         assert(num_channels > 0);
 297
 298         vector<float> silence(samples_per_frame * num_channels, 0.0f);
 299         for (unsigned i = 0; i < num_frames; ++i) {
 300                 device->resampling_queue->add_input_samples(device->next_local_pts / double(TIMEBASE), silence.data(), samples_per_frame);
 301                 // Note that if the format changed in the meantime, we have
 302                 // no way of detecting that; we just have to assume the frame length
 303                 // is always the same.
 304                 device->next_local_pts += frame_length;
 305         }
 306         return true;
 307 }
 308
 309 bool AudioMixer::silence_card(DeviceSpec device_spec, bool silence)
 310 {
 311         AudioDevice *device = find_audio_device(device_spec);
 312
 313         unique_lock<timed_mutex> lock(audio_mutex, defer_lock);
 314         if (!lock.try_lock_for(chrono::milliseconds(10))) {
 315                 return false;
 316         }
 317
 318         if (device->silenced && !silence) {
 319                 reset_resampler_mutex_held(device_spec);
 320         }
 321         device->silenced = silence;
 322         return true;
 323 }
 324
 325 AudioMixer::AudioDevice *AudioMixer::find_audio_device(DeviceSpec device)
 326 {
 327         switch (device.type) {
 328         case InputSourceType::CAPTURE_CARD:
 329                 return &video_cards[device.index];
 330         case InputSourceType::ALSA_INPUT:
 331                 return &alsa_inputs[device.index];
 332         case InputSourceType::SILENCE:
 333         default:
 334                 assert(false);
 335         }
 336         return nullptr;
 337 }
 338
 339 // Get a pointer to the given channel from the given device.
 340 // The channel must be picked out earlier and resampled.
 341 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)
 342 {
 343         static float zero = 0.0f;
 344         if (source_channel == -1 || device_spec.type == InputSourceType::SILENCE) {
 345                 *srcptr = &zero;
 346                 *stride = 0;
 347                 return;
 348         }
 349         AudioDevice *device = find_audio_device(device_spec);
 350         assert(device->interesting_channels.count(source_channel) != 0);
 351         unsigned channel_index = 0;
 352         for (int channel : device->interesting_channels) {
 353                 if (channel == source_channel) break;
 354                 ++channel_index;
 355         }
 356         assert(channel_index < device->interesting_channels.size());
 357         const auto it = samples_card.find(device_spec);
 358         assert(it != samples_card.end());
 359         *srcptr = &(it->second)[channel_index];
 360         *stride = device->interesting_channels.size();
 361 }
 362
 363 // TODO: Can be SSSE3-optimized if need be.
 364 void AudioMixer::fill_audio_bus(const map<DeviceSpec, vector<float>> &samples_card, const InputMapping::Bus &bus, unsigned num_samples, float *output)
 365 {
 366         if (bus.device.type == InputSourceType::SILENCE) {
 367                 memset(output, 0, num_samples * sizeof(*output));
 368         } else {
 369                 assert(bus.device.type == InputSourceType::CAPTURE_CARD ||
 370                        bus.device.type == InputSourceType::ALSA_INPUT);
 371                 const float *lsrc, *rsrc;
 372                 unsigned lstride, rstride;
 373                 float *dptr = output;
 374                 find_sample_src_from_device(samples_card, bus.device, bus.source_channel[0], &lsrc, &lstride);
 375                 find_sample_src_from_device(samples_card, bus.device, bus.source_channel[1], &rsrc, &rstride);
 376                 for (unsigned i = 0; i < num_samples; ++i) {
 377                         *dptr++ = *lsrc;
 378                         *dptr++ = *rsrc;
 379                         lsrc += lstride;
 380                         rsrc += rstride;
 381                 }
 382         }
 383 }
 384
 385 vector<DeviceSpec> AudioMixer::get_active_devices() const
 386 {
 387         vector<DeviceSpec> ret;
 388         for (unsigned card_index = 0; card_index < MAX_VIDEO_CARDS; ++card_index) {
 389                 const DeviceSpec device_spec{InputSourceType::CAPTURE_CARD, card_index};
 390                 if (!find_audio_device(device_spec)->interesting_channels.empty()) {
 391                         ret.push_back(device_spec);
 392                 }
 393         }
 394         for (unsigned card_index = 0; card_index < MAX_ALSA_CARDS; ++card_index) {
 395                 const DeviceSpec device_spec{InputSourceType::ALSA_INPUT, card_index};
 396                 if (!find_audio_device(device_spec)->interesting_channels.empty()) {
 397                         ret.push_back(device_spec);
 398                 }
 399         }
 400         return ret;
 401 }
 402
 403 vector<float> AudioMixer::get_output(double pts, unsigned num_samples, ResamplingQueue::RateAdjustmentPolicy rate_adjustment_policy)
 404 {
 405         map<DeviceSpec, vector<float>> samples_card;
 406         vector<float> samples_bus;
 407
 408         lock_guard<timed_mutex> lock(audio_mutex);
 409
 410         // Pick out all the interesting channels from all the cards.
 411         for (const DeviceSpec &device_spec : get_active_devices()) {
 412                 AudioDevice *device = find_audio_device(device_spec);
 413                 samples_card[device_spec].resize(num_samples * device->interesting_channels.size());
 414                 if (device->silenced) {
 415                         memset(&samples_card[device_spec][0], 0, samples_card[device_spec].size() * sizeof(float));
 416                 } else {
 417                         device->resampling_queue->get_output_samples(
 418                                 pts,
 419                                 &samples_card[device_spec][0],
 420                                 num_samples,
 421                                 rate_adjustment_policy);
 422                 }
 423         }
 424
 425         vector<float> samples_out, left, right;
 426         samples_out.resize(num_samples * 2);
 427         samples_bus.resize(num_samples * 2);
 428         for (unsigned bus_index = 0; bus_index < input_mapping.buses.size(); ++bus_index) {
 429                 fill_audio_bus(samples_card, input_mapping.buses[bus_index], num_samples, &samples_bus[0]);
 430                 apply_eq(bus_index, &samples_bus);
 431
 432                 {
 433                         lock_guard<mutex> lock(compressor_mutex);
 434
 435                         // Apply a level compressor to get the general level right.
 436                         // Basically, if it's over about -40 dBFS, we squeeze it down to that level
 437                         // (or more precisely, near it, since we don't use infinite ratio),
 438                         // then apply a makeup gain to get it to -14 dBFS. -14 dBFS is, of course,
 439                         // entirely arbitrary, but from practical tests with speech, it seems to
 440                         // put ut around -23 LUFS, so it's a reasonable starting point for later use.
 441                         if (level_compressor_enabled[bus_index]) {
 442                                 float threshold = 0.01f;   // -40 dBFS.
 443                                 float ratio = 20.0f;
 444                                 float attack_time = 0.5f;
 445                                 float release_time = 20.0f;
 446                                 float makeup_gain = from_db(ref_level_dbfs - (-40.0f));  // +26 dB.
 447                                 level_compressor[bus_index]->process(samples_bus.data(), samples_bus.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 448                                 gain_staging_db[bus_index] = to_db(level_compressor[bus_index]->get_attenuation() * makeup_gain);
 449                         } else {
 450                                 // Just apply the gain we already had.
 451                                 float g = from_db(gain_staging_db[bus_index]);
 452                                 for (size_t i = 0; i < samples_bus.size(); ++i) {
 453                                         samples_bus[i] *= g;
 454                                 }
 455                         }
 456
 457 #if 0
 458                         printf("level=%f (%+5.2f dBFS) attenuation=%f (%+5.2f dB) end_result=%+5.2f dB\n",
 459                                 level_compressor.get_level(), to_db(level_compressor.get_level()),
 460                                 level_compressor.get_attenuation(), to_db(level_compressor.get_attenuation()),
 461                                 to_db(level_compressor.get_level() * level_compressor.get_attenuation() * makeup_gain));
 462 #endif
 463
 464                         // The real compressor.
 465                         if (compressor_enabled[bus_index]) {
 466                                 float threshold = from_db(compressor_threshold_dbfs[bus_index]);
 467                                 float ratio = 20.0f;
 468                                 float attack_time = 0.005f;
 469                                 float release_time = 0.040f;
 470                                 float makeup_gain = 2.0f;  // +6 dB.
 471                                 compressor[bus_index]->process(samples_bus.data(), samples_bus.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 472                 //              compressor_att = compressor.get_attenuation();
 473                         }
 474                 }
 475
 476                 add_bus_to_master(bus_index, samples_bus, &samples_out);
 477                 deinterleave_samples(samples_bus, &left, &right);
 478                 measure_bus_levels(bus_index, left, right);
 479         }
 480
 481         {
 482                 lock_guard<mutex> lock(compressor_mutex);
 483
 484                 // Finally a limiter at -4 dB (so, -10 dBFS) to take out the worst peaks only.
 485                 // Note that since ratio is not infinite, we could go slightly higher than this.
 486                 if (limiter_enabled) {
 487                         float threshold = from_db(limiter_threshold_dbfs);
 488                         float ratio = 30.0f;
 489                         float attack_time = 0.0f;  // Instant.
 490                         float release_time = 0.020f;
 491                         float makeup_gain = 1.0f;  // 0 dB.
 492                         limiter.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 493         //              limiter_att = limiter.get_attenuation();
 494                 }
 495
 496         //      printf("limiter=%+5.1f  compressor=%+5.1f\n", to_db(limiter_att), to_db(compressor_att));
 497         }
 498
 499         // At this point, we are most likely close to +0 LU (at least if the
 500         // faders sum to 0 dB and the compressors are on), but all of our
 501         // measurements have been on raw sample values, not R128 values.
 502         // So we have a final makeup gain to get us to +0 LU; the gain
 503         // adjustments required should be relatively small, and also, the
 504         // offset shouldn't change much (only if the type of audio changes
 505         // significantly). Thus, we shoot for updating this value basically
 506         // “whenever we process buffers”, since the R128 calculation isn't exactly
 507         // something we get out per-sample.
 508         //
 509         // Note that there's a feedback loop here, so we choose a very slow filter
 510         // (half-time of 30 seconds).
 511         double target_loudness_factor, alpha;
 512         double loudness_lu = r128.loudness_M() - ref_level_lufs;
 513         double current_makeup_lu = to_db(final_makeup_gain);
 514         target_loudness_factor = final_makeup_gain * from_db(-loudness_lu);
 515
 516         // If we're outside +/- 5 LU uncorrected, we don't count it as
 517         // a normal signal (probably silence) and don't change the
 518         // correction factor; just apply what we already have.
 519         if (fabs(loudness_lu - current_makeup_lu) >= 5.0 || !final_makeup_gain_auto) {
 520                 alpha = 0.0;
 521         } else {
 522                 // Formula adapted from
 523                 // https://en.wikipedia.org/wiki/Low-pass_filter#Simple_infinite_impulse_response_filter.
 524                 const double half_time_s = 30.0;
 525                 const double fc_mul_2pi_delta_t = 1.0 / (half_time_s * OUTPUT_FREQUENCY);
 526                 alpha = fc_mul_2pi_delta_t / (fc_mul_2pi_delta_t + 1.0);
 527         }
 528
 529         {
 530                 lock_guard<mutex> lock(compressor_mutex);
 531                 double m = final_makeup_gain;
 532                 for (size_t i = 0; i < samples_out.size(); i += 2) {
 533                         samples_out[i + 0] *= m;
 534                         samples_out[i + 1] *= m;
 535                         m += (target_loudness_factor - m) * alpha;
 536                 }
 537                 final_makeup_gain = m;
 538         }
 539
 540         update_meters(samples_out);
 541
 542         return samples_out;
 543 }
 544
 545 void AudioMixer::apply_eq(unsigned bus_index, vector<float> *samples_bus)
 546 {
 547         constexpr float bass_freq_hz = 200.0f;
 548         constexpr float treble_freq_hz = 4700.0f;
 549
 550         // Cut away everything under 120 Hz (or whatever the cutoff is);
 551         // we don't need it for voice, and it will reduce headroom
 552         // and confuse the compressor. (In particular, any hums at 50 or 60 Hz
 553         // should be dampened.)
 554         if (locut_enabled[bus_index]) {
 555                 locut[bus_index].render(samples_bus->data(), samples_bus->size() / 2, locut_cutoff_hz * 2.0 * M_PI / OUTPUT_FREQUENCY, 0.5f);
 556         }
 557
 558         // Apply the rest of the EQ. Since we only have a simple three-band EQ,
 559         // we can implement it with two shelf filters. We use a simple gain to
 560         // set the mid-level filter, and then offset the low and high bands
 561         // from that if we need to. (We could perhaps have folded the gain into
 562         // the next part, but it's so cheap that the trouble isn't worth it.)
 563         if (fabs(eq_level_db[bus_index][EQ_BAND_MID]) > 0.01f) {
 564                 float g = from_db(eq_level_db[bus_index][EQ_BAND_MID]);
 565                 for (size_t i = 0; i < samples_bus->size(); ++i) {
 566                         (*samples_bus)[i] *= g;
 567                 }
 568         }
 569
 570         float bass_adj_db = eq_level_db[bus_index][EQ_BAND_BASS] - eq_level_db[bus_index][EQ_BAND_MID];
 571         if (fabs(bass_adj_db) > 0.01f) {
 572                 eq[bus_index][EQ_BAND_BASS].render(samples_bus->data(), samples_bus->size() / 2,
 573                         bass_freq_hz * 2.0 * M_PI / OUTPUT_FREQUENCY, 0.5f, bass_adj_db / 40.0f);
 574         }
 575
 576         float treble_adj_db = eq_level_db[bus_index][EQ_BAND_TREBLE] - eq_level_db[bus_index][EQ_BAND_MID];
 577         if (fabs(treble_adj_db) > 0.01f) {
 578                 eq[bus_index][EQ_BAND_TREBLE].render(samples_bus->data(), samples_bus->size() / 2,
 579                         treble_freq_hz * 2.0 * M_PI / OUTPUT_FREQUENCY, 0.5f, treble_adj_db / 40.0f);
 580         }
 581 }
 582
 583 void AudioMixer::add_bus_to_master(unsigned bus_index, const vector<float> &samples_bus, vector<float> *samples_out)
 584 {
 585         assert(samples_bus.size() == samples_out->size());
 586         assert(samples_bus.size() % 2 == 0);
 587         unsigned num_samples = samples_bus.size() / 2;
 588         if (fabs(fader_volume_db[bus_index] - last_fader_volume_db[bus_index]) > 1e-3) {
 589                 // The volume has changed; do a fade over the course of this frame.
 590                 // (We might have some numerical issues here, but it seems to sound OK.)
 591                 // For the purpose of fading here, the silence floor is set to -90 dB
 592                 // (the fader only goes to -84).
 593                 float old_volume = from_db(max<float>(last_fader_volume_db[bus_index], -90.0f));
 594                 float volume = from_db(max<float>(fader_volume_db[bus_index], -90.0f));
 595
 596                 float volume_inc = pow(volume / old_volume, 1.0 / num_samples);
 597                 volume = old_volume;
 598                 if (bus_index == 0) {
 599                         for (unsigned i = 0; i < num_samples; ++i) {
 600                                 (*samples_out)[i * 2 + 0] = samples_bus[i * 2 + 0] * volume;
 601                                 (*samples_out)[i * 2 + 1] = samples_bus[i * 2 + 1] * volume;
 602                                 volume *= volume_inc;
 603                         }
 604                 } else {
 605                         for (unsigned i = 0; i < num_samples; ++i) {
 606                                 (*samples_out)[i * 2 + 0] += samples_bus[i * 2 + 0] * volume;
 607                                 (*samples_out)[i * 2 + 1] += samples_bus[i * 2 + 1] * volume;
 608                                 volume *= volume_inc;
 609                         }
 610                 }
 611         } else {
 612                 float volume = from_db(fader_volume_db[bus_index]);
 613                 if (bus_index == 0) {
 614                         for (unsigned i = 0; i < num_samples; ++i) {
 615                                 (*samples_out)[i * 2 + 0] = samples_bus[i * 2 + 0] * volume;
 616                                 (*samples_out)[i * 2 + 1] = samples_bus[i * 2 + 1] * volume;
 617                         }
 618                 } else {
 619                         for (unsigned i = 0; i < num_samples; ++i) {
 620                                 (*samples_out)[i * 2 + 0] += samples_bus[i * 2 + 0] * volume;
 621                                 (*samples_out)[i * 2 + 1] += samples_bus[i * 2 + 1] * volume;
 622                         }
 623                 }
 624         }
 625
 626         last_fader_volume_db[bus_index] = fader_volume_db[bus_index];
 627 }
 628
 629 void AudioMixer::measure_bus_levels(unsigned bus_index, const vector<float> &left, const vector<float> &right)
 630 {
 631         assert(left.size() == right.size());
 632         const float volume = from_db(fader_volume_db[bus_index]);
 633         const float peak_levels[2] = {
 634                 find_peak(left.data(), left.size()) * volume,
 635                 find_peak(right.data(), right.size()) * volume
 636         };
 637         for (unsigned channel = 0; channel < 2; ++channel) {
 638                 // Compute the current value, including hold and falloff.
 639                 // The constants are borrowed from zita-mu1 by Fons Adriaensen.
 640                 static constexpr float hold_sec = 0.5f;
 641                 static constexpr float falloff_db_sec = 15.0f;  // dB/sec falloff after hold.
 642                 float current_peak;
 643                 PeakHistory &history = peak_history[bus_index][channel];
 644                 history.historic_peak = max(history.historic_peak, peak_levels[channel]);
 645                 if (history.age_seconds < hold_sec) {
 646                         current_peak = history.last_peak;
 647                 } else {
 648                         current_peak = history.last_peak * from_db(-falloff_db_sec * (history.age_seconds - hold_sec));
 649                 }
 650
 651                 // See if we have a new peak to replace the old (possibly falling) one.
 652                 if (peak_levels[channel] > current_peak) {
 653                         history.last_peak = peak_levels[channel];
 654                         history.age_seconds = 0.0f;  // Not 100% correct, but more than good enough given our frame sizes.
 655                         current_peak = peak_levels[channel];
 656                 } else {
 657                         history.age_seconds += float(left.size()) / OUTPUT_FREQUENCY;
 658                 }
 659                 history.current_level = peak_levels[channel];
 660                 history.current_peak = current_peak;
 661         }
 662 }
 663
 664 void AudioMixer::update_meters(const vector<float> &samples)
 665 {
 666         // Upsample 4x to find interpolated peak.
 667         peak_resampler.inp_data = const_cast<float *>(samples.data());
 668         peak_resampler.inp_count = samples.size() / 2;
 669
 670         vector<float> interpolated_samples;
 671         interpolated_samples.resize(samples.size());
 672         {
 673                 lock_guard<mutex> lock(audio_measure_mutex);
 674
 675                 while (peak_resampler.inp_count > 0) {  // About four iterations.
 676                         peak_resampler.out_data = &interpolated_samples[0];
 677                         peak_resampler.out_count = interpolated_samples.size() / 2;
 678                         peak_resampler.process();
 679                         size_t out_stereo_samples = interpolated_samples.size() / 2 - peak_resampler.out_count;
 680                         peak = max<float>(peak, find_peak(interpolated_samples.data(), out_stereo_samples * 2));
 681                         peak_resampler.out_data = nullptr;
 682                 }
 683         }
 684
 685         // Find R128 levels and L/R correlation.
 686         vector<float> left, right;
 687         deinterleave_samples(samples, &left, &right);
 688         float *ptrs[] = { left.data(), right.data() };
 689         {
 690                 lock_guard<mutex> lock(audio_measure_mutex);
 691                 r128.process(left.size(), ptrs);
 692                 correlation.process_samples(samples);
 693         }
 694
 695         send_audio_level_callback();
 696 }
 697
 698 void AudioMixer::reset_meters()
 699 {
 700         lock_guard<mutex> lock(audio_measure_mutex);
 701         peak_resampler.reset();
 702         peak = 0.0f;
 703         r128.reset();
 704         r128.integr_start();
 705         correlation.reset();
 706 }
 707
 708 void AudioMixer::send_audio_level_callback()
 709 {
 710         if (audio_level_callback == nullptr) {
 711                 return;
 712         }
 713
 714         lock_guard<mutex> lock(audio_measure_mutex);
 715         double loudness_s = r128.loudness_S();
 716         double loudness_i = r128.integrated();
 717         double loudness_range_low = r128.range_min();
 718         double loudness_range_high = r128.range_max();
 719
 720         vector<BusLevel> bus_levels;
 721         bus_levels.resize(input_mapping.buses.size());
 722         {
 723                 lock_guard<mutex> lock(compressor_mutex);
 724                 for (unsigned bus_index = 0; bus_index < bus_levels.size(); ++bus_index) {
 725                         bus_levels[bus_index].current_level_dbfs[0] = to_db(peak_history[bus_index][0].current_level);
 726                         bus_levels[bus_index].current_level_dbfs[1] = to_db(peak_history[bus_index][1].current_level);
 727                         bus_levels[bus_index].peak_level_dbfs[0] = to_db(peak_history[bus_index][0].current_peak);
 728                         bus_levels[bus_index].peak_level_dbfs[1] = to_db(peak_history[bus_index][1].current_peak);
 729                         bus_levels[bus_index].historic_peak_dbfs = to_db(
 730                                 max(peak_history[bus_index][0].historic_peak,
 731                                     peak_history[bus_index][1].historic_peak));
 732                         bus_levels[bus_index].gain_staging_db = gain_staging_db[bus_index];
 733                         if (compressor_enabled[bus_index]) {
 734                                 bus_levels[bus_index].compressor_attenuation_db = -to_db(compressor[bus_index]->get_attenuation());
 735                         } else {
 736                                 bus_levels[bus_index].compressor_attenuation_db = 0.0;
 737                         }
 738                 }
 739         }
 740
 741         audio_level_callback(loudness_s, to_db(peak), bus_levels,
 742                 loudness_i, loudness_range_low, loudness_range_high,
 743                 to_db(final_makeup_gain),
 744                 correlation.get_correlation());
 745 }
 746
 747 map<DeviceSpec, DeviceInfo> AudioMixer::get_devices()
 748 {
 749         lock_guard<timed_mutex> lock(audio_mutex);
 750
 751         map<DeviceSpec, DeviceInfo> devices;
 752         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 753                 const DeviceSpec spec{ InputSourceType::CAPTURE_CARD, card_index };
 754                 const AudioDevice *device = &video_cards[card_index];
 755                 DeviceInfo info;
 756                 info.display_name = device->display_name;
 757                 info.num_channels = 8;
 758                 devices.insert(make_pair(spec, info));
 759         }
 760         vector<ALSAPool::Device> available_alsa_devices = alsa_pool.get_devices();
 761         for (unsigned card_index = 0; card_index < available_alsa_devices.size(); ++card_index) {
 762                 const DeviceSpec spec{ InputSourceType::ALSA_INPUT, card_index };
 763                 const ALSAPool::Device &device = available_alsa_devices[card_index];
 764                 DeviceInfo info;
 765                 info.display_name = device.display_name();
 766                 info.num_channels = device.num_channels;
 767                 info.alsa_name = device.name;
 768                 info.alsa_info = device.info;
 769                 info.alsa_address = device.address;
 770                 devices.insert(make_pair(spec, info));
 771         }
 772         return devices;
 773 }
 774
 775 void AudioMixer::set_display_name(DeviceSpec device_spec, const string &name)
 776 {
 777         AudioDevice *device = find_audio_device(device_spec);
 778
 779         lock_guard<timed_mutex> lock(audio_mutex);
 780         device->display_name = name;
 781 }
 782
 783 void AudioMixer::serialize_device(DeviceSpec device_spec, DeviceSpecProto *device_spec_proto)
 784 {
 785         lock_guard<timed_mutex> lock(audio_mutex);
 786         switch (device_spec.type) {
 787                 case InputSourceType::SILENCE:
 788                         device_spec_proto->set_type(DeviceSpecProto::SILENCE);
 789                         break;
 790                 case InputSourceType::CAPTURE_CARD:
 791                         device_spec_proto->set_type(DeviceSpecProto::CAPTURE_CARD);
 792                         device_spec_proto->set_index(device_spec.index);
 793                         device_spec_proto->set_display_name(video_cards[device_spec.index].display_name);
 794                         break;
 795                 case InputSourceType::ALSA_INPUT:
 796                         alsa_pool.serialize_device(device_spec.index, device_spec_proto);
 797                         break;
 798         }
 799 }
 800
 801 void AudioMixer::set_input_mapping(const InputMapping &new_input_mapping)
 802 {
 803         lock_guard<timed_mutex> lock(audio_mutex);
 804
 805         map<DeviceSpec, set<unsigned>> interesting_channels;
 806         for (const InputMapping::Bus &bus : new_input_mapping.buses) {
 807                 if (bus.device.type == InputSourceType::CAPTURE_CARD ||
 808                     bus.device.type == InputSourceType::ALSA_INPUT) {
 809                         for (unsigned channel = 0; channel < 2; ++channel) {
 810                                 if (bus.source_channel[channel] != -1) {
 811                                         interesting_channels[bus.device].insert(bus.source_channel[channel]);
 812                                 }
 813                         }
 814                 }
 815         }
 816
 817         // Reset resamplers for all cards that don't have the exact same state as before.
 818         for (unsigned card_index = 0; card_index < MAX_VIDEO_CARDS; ++card_index) {
 819                 const DeviceSpec device_spec{InputSourceType::CAPTURE_CARD, card_index};
 820                 AudioDevice *device = find_audio_device(device_spec);
 821                 if (device->interesting_channels != interesting_channels[device_spec]) {
 822                         device->interesting_channels = interesting_channels[device_spec];
 823                         reset_resampler_mutex_held(device_spec);
 824                 }
 825         }
 826         for (unsigned card_index = 0; card_index < MAX_ALSA_CARDS; ++card_index) {
 827                 const DeviceSpec device_spec{InputSourceType::ALSA_INPUT, card_index};
 828                 AudioDevice *device = find_audio_device(device_spec);
 829                 if (interesting_channels[device_spec].empty()) {
 830                         alsa_pool.release_device(card_index);
 831                 } else {
 832                         alsa_pool.hold_device(card_index);
 833                 }
 834                 if (device->interesting_channels != interesting_channels[device_spec]) {
 835                         device->interesting_channels = interesting_channels[device_spec];
 836                         alsa_pool.reset_device(device_spec.index);
 837                         reset_resampler_mutex_held(device_spec);
 838                 }
 839         }
 840
 841         input_mapping = new_input_mapping;
 842 }
 843
 844 InputMapping AudioMixer::get_input_mapping() const
 845 {
 846         lock_guard<timed_mutex> lock(audio_mutex);
 847         return input_mapping;
 848 }
 849
 850 void AudioMixer::reset_peak(unsigned bus_index)
 851 {
 852         lock_guard<timed_mutex> lock(audio_mutex);
 853         for (unsigned channel = 0; channel < 2; ++channel) {
 854                 PeakHistory &history = peak_history[bus_index][channel];
 855                 history.current_level = 0.0f;
 856                 history.historic_peak = 0.0f;
 857                 history.current_peak = 0.0f;
 858                 history.last_peak = 0.0f;
 859                 history.age_seconds = 0.0f;
 860         }
 861 }
 862
 863 AudioMixer *global_audio_mixer = nullptr;