git.sesse.net Git - nageru/blob - 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 "db.h"
  22 #include "flags.h"
  23 #include "state.pb.h"
  24 #include "timebase.h"
  25
  26 using namespace bmusb;
  27 using namespace std;
  28 using namespace std::chrono;
  29 using namespace std::placeholders;
  30
  31 namespace {
  32
  33 // TODO: If these prove to be a bottleneck, they can be SSSE3-optimized
  34 // (usually including multiple channels at a time).
  35
  36 void convert_fixed16_to_fp32(float *dst, size_t out_channel, size_t out_num_channels,
  37                              const uint8_t *src, size_t in_channel, size_t in_num_channels,
  38                              size_t num_samples)
  39 {
  40         assert(in_channel < in_num_channels);
  41         assert(out_channel < out_num_channels);
  42         src += in_channel * 2;
  43         dst += out_channel;
  44
  45         for (size_t i = 0; i < num_samples; ++i) {
  46                 int16_t s = le16toh(*(int16_t *)src);
  47                 *dst = s * (1.0f / 32768.0f);
  48
  49                 src += 2 * in_num_channels;
  50                 dst += out_num_channels;
  51         }
  52 }
  53
  54 void convert_fixed24_to_fp32(float *dst, size_t out_channel, size_t out_num_channels,
  55                              const uint8_t *src, size_t in_channel, size_t in_num_channels,
  56                              size_t num_samples)
  57 {
  58         assert(in_channel < in_num_channels);
  59         assert(out_channel < out_num_channels);
  60         src += in_channel * 3;
  61         dst += out_channel;
  62
  63         for (size_t i = 0; i < num_samples; ++i) {
  64                 uint32_t s1 = src[0];
  65                 uint32_t s2 = src[1];
  66                 uint32_t s3 = src[2];
  67                 uint32_t s = s1 | (s1 << 8) | (s2 << 16) | (s3 << 24);
  68                 *dst = int(s) * (1.0f / 2147483648.0f);
  69
  70                 src += 3 * in_num_channels;
  71                 dst += out_num_channels;
  72         }
  73 }
  74
  75 void convert_fixed32_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 * 4;
  82         dst += out_channel;
  83
  84         for (size_t i = 0; i < num_samples; ++i) {
  85                 int32_t s = le32toh(*(int32_t *)src);
  86                 *dst = s * (1.0f / 2147483648.0f);
  87
  88                 src += 4 * in_num_channels;
  89                 dst += out_num_channels;
  90         }
  91 }
  92
  93 float find_peak_plain(const float *samples, size_t num_samples) __attribute__((unused));
  94
  95 float find_peak_plain(const float *samples, size_t num_samples)
  96 {
  97         float m = fabs(samples[0]);
  98         for (size_t i = 1; i < num_samples; ++i) {
  99                 m = max(m, fabs(samples[i]));
 100         }
 101         return m;
 102 }
 103
 104 #ifdef __SSE__
 105 static inline float horizontal_max(__m128 m)
 106 {
 107         __m128 tmp = _mm_shuffle_ps(m, m, _MM_SHUFFLE(1, 0, 3, 2));
 108         m = _mm_max_ps(m, tmp);
 109         tmp = _mm_shuffle_ps(m, m, _MM_SHUFFLE(2, 3, 0, 1));
 110         m = _mm_max_ps(m, tmp);
 111         return _mm_cvtss_f32(m);
 112 }
 113
 114 float find_peak(const float *samples, size_t num_samples)
 115 {
 116         const __m128 abs_mask = _mm_castsi128_ps(_mm_set1_epi32(0x7fffffffu));
 117         __m128 m = _mm_setzero_ps();
 118         for (size_t i = 0; i < (num_samples & ~3); i += 4) {
 119                 __m128 x = _mm_loadu_ps(samples + i);
 120                 x = _mm_and_ps(x, abs_mask);
 121                 m = _mm_max_ps(m, x);
 122         }
 123         float result = horizontal_max(m);
 124
 125         for (size_t i = (num_samples & ~3); i < num_samples; ++i) {
 126                 result = max(result, fabs(samples[i]));
 127         }
 128
 129 #if 0
 130         // Self-test. We should be bit-exact the same.
 131         float reference_result = find_peak_plain(samples, num_samples);
 132         if (result != reference_result) {
 133                 fprintf(stderr, "Error: Peak is %f [%f %f %f %f]; should be %f.\n",
 134                         result,
 135                         _mm_cvtss_f32(_mm_shuffle_ps(m, m, _MM_SHUFFLE(0, 0, 0, 0))),
 136                         _mm_cvtss_f32(_mm_shuffle_ps(m, m, _MM_SHUFFLE(1, 1, 1, 1))),
 137                         _mm_cvtss_f32(_mm_shuffle_ps(m, m, _MM_SHUFFLE(2, 2, 2, 2))),
 138                         _mm_cvtss_f32(_mm_shuffle_ps(m, m, _MM_SHUFFLE(3, 3, 3, 3))),
 139                         reference_result);
 140                 abort();
 141         }
 142 #endif
 143         return result;
 144 }
 145 #else
 146 float find_peak(const float *samples, size_t num_samples)
 147 {
 148         return find_peak_plain(samples, num_samples);
 149 }
 150 #endif
 151
 152 void deinterleave_samples(const vector<float> &in, vector<float> *out_l, vector<float> *out_r)
 153 {
 154         size_t num_samples = in.size() / 2;
 155         out_l->resize(num_samples);
 156         out_r->resize(num_samples);
 157
 158         const float *inptr = in.data();
 159         float *lptr = &(*out_l)[0];
 160         float *rptr = &(*out_r)[0];
 161         for (size_t i = 0; i < num_samples; ++i) {
 162                 *lptr++ = *inptr++;
 163                 *rptr++ = *inptr++;
 164         }
 165 }
 166
 167 }  // namespace
 168
 169 AudioMixer::AudioMixer(unsigned num_cards)
 170         : num_cards(num_cards),
 171           limiter(OUTPUT_FREQUENCY),
 172           correlation(OUTPUT_FREQUENCY)
 173 {
 174         for (unsigned bus_index = 0; bus_index < MAX_BUSES; ++bus_index) {
 175                 locut[bus_index].init(FILTER_HPF, 2);
 176                 eq[bus_index][EQ_BAND_BASS].init(FILTER_LOW_SHELF, 1);
 177                 // Note: EQ_BAND_MID isn't used (see comments in apply_eq()).
 178                 eq[bus_index][EQ_BAND_TREBLE].init(FILTER_HIGH_SHELF, 1);
 179                 compressor[bus_index].reset(new StereoCompressor(OUTPUT_FREQUENCY));
 180                 level_compressor[bus_index].reset(new StereoCompressor(OUTPUT_FREQUENCY));
 181
 182                 set_bus_settings(bus_index, get_default_bus_settings());
 183         }
 184         set_limiter_enabled(global_flags.limiter_enabled);
 185         set_final_makeup_gain_auto(global_flags.final_makeup_gain_auto);
 186
 187         r128.init(2, OUTPUT_FREQUENCY);
 188         r128.integr_start();
 189
 190         // hlen=16 is pretty low quality, but we use quite a bit of CPU otherwise,
 191         // and there's a limit to how important the peak meter is.
 192         peak_resampler.setup(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY * 4, /*num_channels=*/2, /*hlen=*/16, /*frel=*/1.0);
 193
 194         global_audio_mixer = this;
 195         alsa_pool.init();
 196
 197         if (!global_flags.input_mapping_filename.empty()) {
 198                 // Must happen after ALSAPool is initialized, as it needs to know the card list.
 199                 current_mapping_mode = MappingMode::MULTICHANNEL;
 200                 InputMapping new_input_mapping;
 201                 if (!load_input_mapping_from_file(get_devices(),
 202                                                   global_flags.input_mapping_filename,
 203                                                   &new_input_mapping)) {
 204                         fprintf(stderr, "Failed to load input mapping from '%s', exiting.\n",
 205                                 global_flags.input_mapping_filename.c_str());
 206                         exit(1);
 207                 }
 208                 set_input_mapping(new_input_mapping);
 209         } else {
 210                 set_simple_input(/*card_index=*/0);
 211                 if (global_flags.multichannel_mapping_mode) {
 212                         current_mapping_mode = MappingMode::MULTICHANNEL;
 213                 }
 214         }
 215 }
 216
 217 void AudioMixer::reset_resampler(DeviceSpec device_spec)
 218 {
 219         lock_guard<timed_mutex> lock(audio_mutex);
 220         reset_resampler_mutex_held(device_spec);
 221 }
 222
 223 void AudioMixer::reset_resampler_mutex_held(DeviceSpec device_spec)
 224 {
 225         AudioDevice *device = find_audio_device(device_spec);
 226
 227         if (device->interesting_channels.empty()) {
 228                 device->resampling_queue.reset();
 229         } else {
 230                 // TODO: ResamplingQueue should probably take the full device spec.
 231                 // (It's only used for console output, though.)
 232                 device->resampling_queue.reset(new ResamplingQueue(
 233                         device_spec.index, device->capture_frequency, OUTPUT_FREQUENCY, device->interesting_channels.size(),
 234                         global_flags.audio_queue_length_ms * 0.001));
 235         }
 236 }
 237
 238 bool AudioMixer::add_audio(DeviceSpec device_spec, const uint8_t *data, unsigned num_samples, AudioFormat audio_format, int64_t frame_length, steady_clock::time_point frame_time)
 239 {
 240         AudioDevice *device = find_audio_device(device_spec);
 241
 242         unique_lock<timed_mutex> lock(audio_mutex, defer_lock);
 243         if (!lock.try_lock_for(chrono::milliseconds(10))) {
 244                 return false;
 245         }
 246         if (device->resampling_queue == nullptr) {
 247                 // No buses use this device; throw it away.
 248                 return true;
 249         }
 250
 251         unsigned num_channels = device->interesting_channels.size();
 252         assert(num_channels > 0);
 253
 254         // Convert the audio to fp32.
 255         unique_ptr<float[]> audio(new float[num_samples * num_channels]);
 256         unsigned channel_index = 0;
 257         for (auto channel_it = device->interesting_channels.cbegin(); channel_it != device->interesting_channels.end(); ++channel_it, ++channel_index) {
 258                 switch (audio_format.bits_per_sample) {
 259                 case 0:
 260                         assert(num_samples == 0);
 261                         break;
 262                 case 16:
 263                         convert_fixed16_to_fp32(audio.get(), channel_index, num_channels, data, *channel_it, audio_format.num_channels, num_samples);
 264                         break;
 265                 case 24:
 266                         convert_fixed24_to_fp32(audio.get(), channel_index, num_channels, data, *channel_it, audio_format.num_channels, num_samples);
 267                         break;
 268                 case 32:
 269                         convert_fixed32_to_fp32(audio.get(), channel_index, num_channels, data, *channel_it, audio_format.num_channels, num_samples);
 270                         break;
 271                 default:
 272                         fprintf(stderr, "Cannot handle audio with %u bits per sample\n", audio_format.bits_per_sample);
 273                         assert(false);
 274                 }
 275         }
 276
 277         // Now add it.
 278         device->resampling_queue->add_input_samples(frame_time, audio.get(), num_samples, ResamplingQueue::ADJUST_RATE);
 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(steady_clock::now(), silence.data(), samples_per_frame, ResamplingQueue::DO_NOT_ADJUST_RATE);
 301         }
 302         return true;
 303 }
 304
 305 bool AudioMixer::silence_card(DeviceSpec device_spec, bool silence)
 306 {
 307         AudioDevice *device = find_audio_device(device_spec);
 308
 309         unique_lock<timed_mutex> lock(audio_mutex, defer_lock);
 310         if (!lock.try_lock_for(chrono::milliseconds(10))) {
 311                 return false;
 312         }
 313
 314         if (device->silenced && !silence) {
 315                 reset_resampler_mutex_held(device_spec);
 316         }
 317         device->silenced = silence;
 318         return true;
 319 }
 320
 321 AudioMixer::BusSettings AudioMixer::get_default_bus_settings()
 322 {
 323         BusSettings settings;
 324         settings.fader_volume_db = 0.0f;
 325         settings.muted = false;
 326         settings.locut_enabled = global_flags.locut_enabled;
 327         for (unsigned band_index = 0; band_index < NUM_EQ_BANDS; ++band_index) {
 328                 settings.eq_level_db[band_index] = 0.0f;
 329         }
 330         settings.gain_staging_db = global_flags.initial_gain_staging_db;
 331         settings.level_compressor_enabled = global_flags.gain_staging_auto;
 332         settings.compressor_threshold_dbfs = ref_level_dbfs - 12.0f;  // -12 dB.
 333         settings.compressor_enabled = global_flags.compressor_enabled;
 334         return settings;
 335 }
 336
 337 AudioMixer::BusSettings AudioMixer::get_bus_settings(unsigned bus_index) const
 338 {
 339         lock_guard<timed_mutex> lock(audio_mutex);
 340         BusSettings settings;
 341         settings.fader_volume_db = fader_volume_db[bus_index];
 342         settings.muted = mute[bus_index];
 343         settings.locut_enabled = locut_enabled[bus_index];
 344         for (unsigned band_index = 0; band_index < NUM_EQ_BANDS; ++band_index) {
 345                 settings.eq_level_db[band_index] = eq_level_db[bus_index][band_index];
 346         }
 347         settings.gain_staging_db = gain_staging_db[bus_index];
 348         settings.level_compressor_enabled = level_compressor_enabled[bus_index];
 349         settings.compressor_threshold_dbfs = compressor_threshold_dbfs[bus_index];
 350         settings.compressor_enabled = compressor_enabled[bus_index];
 351         return settings;
 352 }
 353
 354 void AudioMixer::set_bus_settings(unsigned bus_index, const AudioMixer::BusSettings &settings)
 355 {
 356         lock_guard<timed_mutex> lock(audio_mutex);
 357         fader_volume_db[bus_index] = settings.fader_volume_db;
 358         mute[bus_index] = settings.muted;
 359         locut_enabled[bus_index] = settings.locut_enabled;
 360         for (unsigned band_index = 0; band_index < NUM_EQ_BANDS; ++band_index) {
 361                 eq_level_db[bus_index][band_index] = settings.eq_level_db[band_index];
 362         }
 363         gain_staging_db[bus_index] = settings.gain_staging_db;
 364         last_gain_staging_db[bus_index] = gain_staging_db[bus_index];
 365         level_compressor_enabled[bus_index] = settings.level_compressor_enabled;
 366         compressor_threshold_dbfs[bus_index] = settings.compressor_threshold_dbfs;
 367         compressor_enabled[bus_index] = settings.compressor_enabled;
 368 }
 369
 370 AudioMixer::AudioDevice *AudioMixer::find_audio_device(DeviceSpec device)
 371 {
 372         switch (device.type) {
 373         case InputSourceType::CAPTURE_CARD:
 374                 return &video_cards[device.index];
 375         case InputSourceType::ALSA_INPUT:
 376                 return &alsa_inputs[device.index];
 377         case InputSourceType::SILENCE:
 378         default:
 379                 assert(false);
 380         }
 381         return nullptr;
 382 }
 383
 384 // Get a pointer to the given channel from the given device.
 385 // The channel must be picked out earlier and resampled.
 386 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)
 387 {
 388         static float zero = 0.0f;
 389         if (source_channel == -1 || device_spec.type == InputSourceType::SILENCE) {
 390                 *srcptr = &zero;
 391                 *stride = 0;
 392                 return;
 393         }
 394         AudioDevice *device = find_audio_device(device_spec);
 395         assert(device->interesting_channels.count(source_channel) != 0);
 396         unsigned channel_index = 0;
 397         for (int channel : device->interesting_channels) {
 398                 if (channel == source_channel) break;
 399                 ++channel_index;
 400         }
 401         assert(channel_index < device->interesting_channels.size());
 402         const auto it = samples_card.find(device_spec);
 403         assert(it != samples_card.end());
 404         *srcptr = &(it->second)[channel_index];
 405         *stride = device->interesting_channels.size();
 406 }
 407
 408 // TODO: Can be SSSE3-optimized if need be.
 409 void AudioMixer::fill_audio_bus(const map<DeviceSpec, vector<float>> &samples_card, const InputMapping::Bus &bus, unsigned num_samples, float *output)
 410 {
 411         if (bus.device.type == InputSourceType::SILENCE) {
 412                 memset(output, 0, num_samples * 2 * sizeof(*output));
 413         } else {
 414                 assert(bus.device.type == InputSourceType::CAPTURE_CARD ||
 415                        bus.device.type == InputSourceType::ALSA_INPUT);
 416                 const float *lsrc, *rsrc;
 417                 unsigned lstride, rstride;
 418                 float *dptr = output;
 419                 find_sample_src_from_device(samples_card, bus.device, bus.source_channel[0], &lsrc, &lstride);
 420                 find_sample_src_from_device(samples_card, bus.device, bus.source_channel[1], &rsrc, &rstride);
 421                 for (unsigned i = 0; i < num_samples; ++i) {
 422                         *dptr++ = *lsrc;
 423                         *dptr++ = *rsrc;
 424                         lsrc += lstride;
 425                         rsrc += rstride;
 426                 }
 427         }
 428 }
 429
 430 vector<DeviceSpec> AudioMixer::get_active_devices() const
 431 {
 432         vector<DeviceSpec> ret;
 433         for (unsigned card_index = 0; card_index < MAX_VIDEO_CARDS; ++card_index) {
 434                 const DeviceSpec device_spec{InputSourceType::CAPTURE_CARD, card_index};
 435                 if (!find_audio_device(device_spec)->interesting_channels.empty()) {
 436                         ret.push_back(device_spec);
 437                 }
 438         }
 439         for (unsigned card_index = 0; card_index < MAX_ALSA_CARDS; ++card_index) {
 440                 const DeviceSpec device_spec{InputSourceType::ALSA_INPUT, card_index};
 441                 if (!find_audio_device(device_spec)->interesting_channels.empty()) {
 442                         ret.push_back(device_spec);
 443                 }
 444         }
 445         return ret;
 446 }
 447
 448 namespace {
 449
 450 void apply_gain(float db, float last_db, vector<float> *samples)
 451 {
 452         if (fabs(db - last_db) < 1e-3) {
 453                 // Constant over this frame.
 454                 const float gain = from_db(db);
 455                 for (size_t i = 0; i < samples->size(); ++i) {
 456                         (*samples)[i] *= gain;
 457                 }
 458         } else {
 459                 // We need to do a fade.
 460                 unsigned num_samples = samples->size() / 2;
 461                 float gain = from_db(last_db);
 462                 const float gain_inc = pow(from_db(db - last_db), 1.0 / num_samples);
 463                 for (size_t i = 0; i < num_samples; ++i) {
 464                         (*samples)[i * 2 + 0] *= gain;
 465                         (*samples)[i * 2 + 1] *= gain;
 466                         gain *= gain_inc;
 467                 }
 468         }
 469 }
 470
 471 }  // namespace
 472
 473 vector<float> AudioMixer::get_output(steady_clock::time_point ts, unsigned num_samples, ResamplingQueue::RateAdjustmentPolicy rate_adjustment_policy)
 474 {
 475         map<DeviceSpec, vector<float>> samples_card;
 476         vector<float> samples_bus;
 477
 478         lock_guard<timed_mutex> lock(audio_mutex);
 479
 480         // Pick out all the interesting channels from all the cards.
 481         for (const DeviceSpec &device_spec : get_active_devices()) {
 482                 AudioDevice *device = find_audio_device(device_spec);
 483                 samples_card[device_spec].resize(num_samples * device->interesting_channels.size());
 484                 if (device->silenced) {
 485                         memset(&samples_card[device_spec][0], 0, samples_card[device_spec].size() * sizeof(float));
 486                 } else {
 487                         device->resampling_queue->get_output_samples(
 488                                 ts,
 489                                 &samples_card[device_spec][0],
 490                                 num_samples,
 491                                 rate_adjustment_policy);
 492                 }
 493         }
 494
 495         vector<float> samples_out, left, right;
 496         samples_out.resize(num_samples * 2);
 497         samples_bus.resize(num_samples * 2);
 498         for (unsigned bus_index = 0; bus_index < input_mapping.buses.size(); ++bus_index) {
 499                 fill_audio_bus(samples_card, input_mapping.buses[bus_index], num_samples, &samples_bus[0]);
 500                 apply_eq(bus_index, &samples_bus);
 501
 502                 {
 503                         lock_guard<mutex> lock(compressor_mutex);
 504
 505                         // Apply a level compressor to get the general level right.
 506                         // Basically, if it's over about -40 dBFS, we squeeze it down to that level
 507                         // (or more precisely, near it, since we don't use infinite ratio),
 508                         // then apply a makeup gain to get it to -14 dBFS. -14 dBFS is, of course,
 509                         // entirely arbitrary, but from practical tests with speech, it seems to
 510                         // put ut around -23 LUFS, so it's a reasonable starting point for later use.
 511                         if (level_compressor_enabled[bus_index]) {
 512                                 float threshold = 0.01f;   // -40 dBFS.
 513                                 float ratio = 20.0f;
 514                                 float attack_time = 0.5f;
 515                                 float release_time = 20.0f;
 516                                 float makeup_gain = from_db(ref_level_dbfs - (-40.0f));  // +26 dB.
 517                                 level_compressor[bus_index]->process(samples_bus.data(), samples_bus.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 518                                 gain_staging_db[bus_index] = to_db(level_compressor[bus_index]->get_attenuation() * makeup_gain);
 519                         } else {
 520                                 // Just apply the gain we already had.
 521                                 float db = gain_staging_db[bus_index];
 522                                 float last_db = last_gain_staging_db[bus_index];
 523                                 apply_gain(db, last_db, &samples_bus);
 524                         }
 525                         last_gain_staging_db[bus_index] = gain_staging_db[bus_index];
 526
 527 #if 0
 528                         printf("level=%f (%+5.2f dBFS) attenuation=%f (%+5.2f dB) end_result=%+5.2f dB\n",
 529                                 level_compressor.get_level(), to_db(level_compressor.get_level()),
 530                                 level_compressor.get_attenuation(), to_db(level_compressor.get_attenuation()),
 531                                 to_db(level_compressor.get_level() * level_compressor.get_attenuation() * makeup_gain));
 532 #endif
 533
 534                         // The real compressor.
 535                         if (compressor_enabled[bus_index]) {
 536                                 float threshold = from_db(compressor_threshold_dbfs[bus_index]);
 537                                 float ratio = 20.0f;
 538                                 float attack_time = 0.005f;
 539                                 float release_time = 0.040f;
 540                                 float makeup_gain = 2.0f;  // +6 dB.
 541                                 compressor[bus_index]->process(samples_bus.data(), samples_bus.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 542                 //              compressor_att = compressor.get_attenuation();
 543                         }
 544                 }
 545
 546                 add_bus_to_master(bus_index, samples_bus, &samples_out);
 547                 deinterleave_samples(samples_bus, &left, &right);
 548                 measure_bus_levels(bus_index, left, right);
 549         }
 550
 551         {
 552                 lock_guard<mutex> lock(compressor_mutex);
 553
 554                 // Finally a limiter at -4 dB (so, -10 dBFS) to take out the worst peaks only.
 555                 // Note that since ratio is not infinite, we could go slightly higher than this.
 556                 if (limiter_enabled) {
 557                         float threshold = from_db(limiter_threshold_dbfs);
 558                         float ratio = 30.0f;
 559                         float attack_time = 0.0f;  // Instant.
 560                         float release_time = 0.020f;
 561                         float makeup_gain = 1.0f;  // 0 dB.
 562                         limiter.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 563         //              limiter_att = limiter.get_attenuation();
 564                 }
 565
 566         //      printf("limiter=%+5.1f  compressor=%+5.1f\n", to_db(limiter_att), to_db(compressor_att));
 567         }
 568
 569         // At this point, we are most likely close to +0 LU (at least if the
 570         // faders sum to 0 dB and the compressors are on), but all of our
 571         // measurements have been on raw sample values, not R128 values.
 572         // So we have a final makeup gain to get us to +0 LU; the gain
 573         // adjustments required should be relatively small, and also, the
 574         // offset shouldn't change much (only if the type of audio changes
 575         // significantly). Thus, we shoot for updating this value basically
 576         // “whenever we process buffers”, since the R128 calculation isn't exactly
 577         // something we get out per-sample.
 578         //
 579         // Note that there's a feedback loop here, so we choose a very slow filter
 580         // (half-time of 30 seconds).
 581         double target_loudness_factor, alpha;
 582         double loudness_lu = r128.loudness_M() - ref_level_lufs;
 583         target_loudness_factor = final_makeup_gain * from_db(-loudness_lu);
 584
 585         // If we're outside +/- 5 LU (after correction), we don't count it as
 586         // a normal signal (probably silence) and don't change the
 587         // correction factor; just apply what we already have.
 588         if (fabs(loudness_lu) >= 5.0 || !final_makeup_gain_auto) {
 589                 alpha = 0.0;
 590         } else {
 591                 // Formula adapted from
 592                 // https://en.wikipedia.org/wiki/Low-pass_filter#Simple_infinite_impulse_response_filter.
 593                 const double half_time_s = 30.0;
 594                 const double fc_mul_2pi_delta_t = 1.0 / (half_time_s * OUTPUT_FREQUENCY);
 595                 alpha = fc_mul_2pi_delta_t / (fc_mul_2pi_delta_t + 1.0);
 596         }
 597
 598         {
 599                 lock_guard<mutex> lock(compressor_mutex);
 600                 double m = final_makeup_gain;
 601                 for (size_t i = 0; i < samples_out.size(); i += 2) {
 602                         samples_out[i + 0] *= m;
 603                         samples_out[i + 1] *= m;
 604                         m += (target_loudness_factor - m) * alpha;
 605                 }
 606                 final_makeup_gain = m;
 607         }
 608
 609         update_meters(samples_out);
 610
 611         return samples_out;
 612 }
 613
 614 namespace {
 615
 616 void apply_filter_fade(StereoFilter *filter, float *data, unsigned num_samples, float cutoff_hz, float db, float last_db)
 617 {
 618         // A granularity of 32 samples is an okay tradeoff between speed and
 619         // smoothness; recalculating the filters is pretty expensive, so it's
 620         // good that we don't do this all the time.
 621         static constexpr unsigned filter_granularity_samples = 32;
 622
 623         const float cutoff_linear = cutoff_hz * 2.0 * M_PI / OUTPUT_FREQUENCY;
 624         if (fabs(db - last_db) < 1e-3) {
 625                 // Constant over this frame.
 626                 if (fabs(db) > 0.01f) {
 627                         filter->render(data, num_samples, cutoff_linear, 0.5f, db / 40.0f);
 628                 }
 629         } else {
 630                 // We need to do a fade. (Rounding up avoids division by zero.)
 631                 unsigned num_blocks = (num_samples + filter_granularity_samples - 1) / filter_granularity_samples;
 632                 const float inc_db_norm = (db - last_db) / 40.0f / num_blocks;
 633                 float db_norm = db / 40.0f;
 634                 for (size_t i = 0; i < num_samples; i += filter_granularity_samples) {
 635                         size_t samples_this_block = std::min<size_t>(num_samples - i, filter_granularity_samples);
 636                         filter->render(data + i * 2, samples_this_block, cutoff_linear, 0.5f, db_norm);
 637                         db_norm += inc_db_norm;
 638                 }
 639         }
 640 }
 641
 642 }  // namespace
 643
 644 void AudioMixer::apply_eq(unsigned bus_index, vector<float> *samples_bus)
 645 {
 646         constexpr float bass_freq_hz = 200.0f;
 647         constexpr float treble_freq_hz = 4700.0f;
 648
 649         // Cut away everything under 120 Hz (or whatever the cutoff is);
 650         // we don't need it for voice, and it will reduce headroom
 651         // and confuse the compressor. (In particular, any hums at 50 or 60 Hz
 652         // should be dampened.)
 653         if (locut_enabled[bus_index]) {
 654                 locut[bus_index].render(samples_bus->data(), samples_bus->size() / 2, locut_cutoff_hz * 2.0 * M_PI / OUTPUT_FREQUENCY, 0.5f);
 655         }
 656
 657         // Apply the rest of the EQ. Since we only have a simple three-band EQ,
 658         // we can implement it with two shelf filters. We use a simple gain to
 659         // set the mid-level filter, and then offset the low and high bands
 660         // from that if we need to. (We could perhaps have folded the gain into
 661         // the next part, but it's so cheap that the trouble isn't worth it.)
 662         //
 663         // If any part of the EQ has changed appreciably since last frame,
 664         // we fade smoothly during the course of this frame.
 665         const float bass_db = eq_level_db[bus_index][EQ_BAND_BASS];
 666         const float mid_db = eq_level_db[bus_index][EQ_BAND_MID];
 667         const float treble_db = eq_level_db[bus_index][EQ_BAND_TREBLE];
 668
 669         const float last_bass_db = last_eq_level_db[bus_index][EQ_BAND_BASS];
 670         const float last_mid_db = last_eq_level_db[bus_index][EQ_BAND_MID];
 671         const float last_treble_db = last_eq_level_db[bus_index][EQ_BAND_TREBLE];
 672
 673         assert(samples_bus->size() % 2 == 0);
 674         const unsigned num_samples = samples_bus->size() / 2;
 675
 676         apply_gain(mid_db, last_mid_db, samples_bus);
 677
 678         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);
 679         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);
 680
 681         last_eq_level_db[bus_index][EQ_BAND_BASS] = bass_db;
 682         last_eq_level_db[bus_index][EQ_BAND_MID] = mid_db;
 683         last_eq_level_db[bus_index][EQ_BAND_TREBLE] = treble_db;
 684 }
 685
 686 void AudioMixer::add_bus_to_master(unsigned bus_index, const vector<float> &samples_bus, vector<float> *samples_out)
 687 {
 688         assert(samples_bus.size() == samples_out->size());
 689         assert(samples_bus.size() % 2 == 0);
 690         unsigned num_samples = samples_bus.size() / 2;
 691         const float new_volume_db = mute[bus_index] ? -90.0f : fader_volume_db[bus_index].load();
 692         if (fabs(new_volume_db - last_fader_volume_db[bus_index]) > 1e-3) {
 693                 // The volume has changed; do a fade over the course of this frame.
 694                 // (We might have some numerical issues here, but it seems to sound OK.)
 695                 // For the purpose of fading here, the silence floor is set to -90 dB
 696                 // (the fader only goes to -84).
 697                 float old_volume = from_db(max<float>(last_fader_volume_db[bus_index], -90.0f));
 698                 float volume = from_db(max<float>(new_volume_db, -90.0f));
 699
 700                 float volume_inc = pow(volume / old_volume, 1.0 / num_samples);
 701                 volume = old_volume;
 702                 if (bus_index == 0) {
 703                         for (unsigned i = 0; i < num_samples; ++i) {
 704                                 (*samples_out)[i * 2 + 0] = samples_bus[i * 2 + 0] * volume;
 705                                 (*samples_out)[i * 2 + 1] = samples_bus[i * 2 + 1] * volume;
 706                                 volume *= volume_inc;
 707                         }
 708                 } else {
 709                         for (unsigned i = 0; i < num_samples; ++i) {
 710                                 (*samples_out)[i * 2 + 0] += samples_bus[i * 2 + 0] * volume;
 711                                 (*samples_out)[i * 2 + 1] += samples_bus[i * 2 + 1] * volume;
 712                                 volume *= volume_inc;
 713                         }
 714                 }
 715         } else if (new_volume_db > -90.0f) {
 716                 float volume = from_db(new_volume_db);
 717                 if (bus_index == 0) {
 718                         for (unsigned i = 0; i < num_samples; ++i) {
 719                                 (*samples_out)[i * 2 + 0] = samples_bus[i * 2 + 0] * volume;
 720                                 (*samples_out)[i * 2 + 1] = samples_bus[i * 2 + 1] * volume;
 721                         }
 722                 } else {
 723                         for (unsigned i = 0; i < num_samples; ++i) {
 724                                 (*samples_out)[i * 2 + 0] += samples_bus[i * 2 + 0] * volume;
 725                                 (*samples_out)[i * 2 + 1] += samples_bus[i * 2 + 1] * volume;
 726                         }
 727                 }
 728         }
 729
 730         last_fader_volume_db[bus_index] = new_volume_db;
 731 }
 732
 733 void AudioMixer::measure_bus_levels(unsigned bus_index, const vector<float> &left, const vector<float> &right)
 734 {
 735         assert(left.size() == right.size());
 736         const float volume = mute[bus_index] ? 0.0f : from_db(fader_volume_db[bus_index]);
 737         const float peak_levels[2] = {
 738                 find_peak(left.data(), left.size()) * volume,
 739                 find_peak(right.data(), right.size()) * volume
 740         };
 741         for (unsigned channel = 0; channel < 2; ++channel) {
 742                 // Compute the current value, including hold and falloff.
 743                 // The constants are borrowed from zita-mu1 by Fons Adriaensen.
 744                 static constexpr float hold_sec = 0.5f;
 745                 static constexpr float falloff_db_sec = 15.0f;  // dB/sec falloff after hold.
 746                 float current_peak;
 747                 PeakHistory &history = peak_history[bus_index][channel];
 748                 history.historic_peak = max(history.historic_peak, peak_levels[channel]);
 749                 if (history.age_seconds < hold_sec) {
 750                         current_peak = history.last_peak;
 751                 } else {
 752                         current_peak = history.last_peak * from_db(-falloff_db_sec * (history.age_seconds - hold_sec));
 753                 }
 754
 755                 // See if we have a new peak to replace the old (possibly falling) one.
 756                 if (peak_levels[channel] > current_peak) {
 757                         history.last_peak = peak_levels[channel];
 758                         history.age_seconds = 0.0f;  // Not 100% correct, but more than good enough given our frame sizes.
 759                         current_peak = peak_levels[channel];
 760                 } else {
 761                         history.age_seconds += float(left.size()) / OUTPUT_FREQUENCY;
 762                 }
 763                 history.current_level = peak_levels[channel];
 764                 history.current_peak = current_peak;
 765         }
 766 }
 767
 768 void AudioMixer::update_meters(const vector<float> &samples)
 769 {
 770         // Upsample 4x to find interpolated peak.
 771         peak_resampler.inp_data = const_cast<float *>(samples.data());
 772         peak_resampler.inp_count = samples.size() / 2;
 773
 774         vector<float> interpolated_samples;
 775         interpolated_samples.resize(samples.size());
 776         {
 777                 lock_guard<mutex> lock(audio_measure_mutex);
 778
 779                 while (peak_resampler.inp_count > 0) {  // About four iterations.
 780                         peak_resampler.out_data = &interpolated_samples[0];
 781                         peak_resampler.out_count = interpolated_samples.size() / 2;
 782                         peak_resampler.process();
 783                         size_t out_stereo_samples = interpolated_samples.size() / 2 - peak_resampler.out_count;
 784                         peak = max<float>(peak, find_peak(interpolated_samples.data(), out_stereo_samples * 2));
 785                         peak_resampler.out_data = nullptr;
 786                 }
 787         }
 788
 789         // Find R128 levels and L/R correlation.
 790         vector<float> left, right;
 791         deinterleave_samples(samples, &left, &right);
 792         float *ptrs[] = { left.data(), right.data() };
 793         {
 794                 lock_guard<mutex> lock(audio_measure_mutex);
 795                 r128.process(left.size(), ptrs);
 796                 correlation.process_samples(samples);
 797         }
 798
 799         send_audio_level_callback();
 800 }
 801
 802 void AudioMixer::reset_meters()
 803 {
 804         lock_guard<mutex> lock(audio_measure_mutex);
 805         peak_resampler.reset();
 806         peak = 0.0f;
 807         r128.reset();
 808         r128.integr_start();
 809         correlation.reset();
 810 }
 811
 812 void AudioMixer::send_audio_level_callback()
 813 {
 814         if (audio_level_callback == nullptr) {
 815                 return;
 816         }
 817
 818         lock_guard<mutex> lock(audio_measure_mutex);
 819         double loudness_s = r128.loudness_S();
 820         double loudness_i = r128.integrated();
 821         double loudness_range_low = r128.range_min();
 822         double loudness_range_high = r128.range_max();
 823
 824         vector<BusLevel> bus_levels;
 825         bus_levels.resize(input_mapping.buses.size());
 826         {
 827                 lock_guard<mutex> lock(compressor_mutex);
 828                 for (unsigned bus_index = 0; bus_index < bus_levels.size(); ++bus_index) {
 829                         bus_levels[bus_index].current_level_dbfs[0] = to_db(peak_history[bus_index][0].current_level);
 830                         bus_levels[bus_index].current_level_dbfs[1] = to_db(peak_history[bus_index][1].current_level);
 831                         bus_levels[bus_index].peak_level_dbfs[0] = to_db(peak_history[bus_index][0].current_peak);
 832                         bus_levels[bus_index].peak_level_dbfs[1] = to_db(peak_history[bus_index][1].current_peak);
 833                         bus_levels[bus_index].historic_peak_dbfs = to_db(
 834                                 max(peak_history[bus_index][0].historic_peak,
 835                                     peak_history[bus_index][1].historic_peak));
 836                         bus_levels[bus_index].gain_staging_db = gain_staging_db[bus_index];
 837                         if (compressor_enabled[bus_index]) {
 838                                 bus_levels[bus_index].compressor_attenuation_db = -to_db(compressor[bus_index]->get_attenuation());
 839                         } else {
 840                                 bus_levels[bus_index].compressor_attenuation_db = 0.0;
 841                         }
 842                 }
 843         }
 844
 845         audio_level_callback(loudness_s, to_db(peak), bus_levels,
 846                 loudness_i, loudness_range_low, loudness_range_high,
 847                 to_db(final_makeup_gain),
 848                 correlation.get_correlation());
 849 }
 850
 851 map<DeviceSpec, DeviceInfo> AudioMixer::get_devices()
 852 {
 853         lock_guard<timed_mutex> lock(audio_mutex);
 854
 855         map<DeviceSpec, DeviceInfo> devices;
 856         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 857                 const DeviceSpec spec{ InputSourceType::CAPTURE_CARD, card_index };
 858                 const AudioDevice *device = &video_cards[card_index];
 859                 DeviceInfo info;
 860                 info.display_name = device->display_name;
 861                 info.num_channels = 8;
 862                 devices.insert(make_pair(spec, info));
 863         }
 864         vector<ALSAPool::Device> available_alsa_devices = alsa_pool.get_devices();
 865         for (unsigned card_index = 0; card_index < available_alsa_devices.size(); ++card_index) {
 866                 const DeviceSpec spec{ InputSourceType::ALSA_INPUT, card_index };
 867                 const ALSAPool::Device &device = available_alsa_devices[card_index];
 868                 DeviceInfo info;
 869                 info.display_name = device.display_name();
 870                 info.num_channels = device.num_channels;
 871                 info.alsa_name = device.name;
 872                 info.alsa_info = device.info;
 873                 info.alsa_address = device.address;
 874                 devices.insert(make_pair(spec, info));
 875         }
 876         return devices;
 877 }
 878
 879 void AudioMixer::set_display_name(DeviceSpec device_spec, const string &name)
 880 {
 881         AudioDevice *device = find_audio_device(device_spec);
 882
 883         lock_guard<timed_mutex> lock(audio_mutex);
 884         device->display_name = name;
 885 }
 886
 887 void AudioMixer::serialize_device(DeviceSpec device_spec, DeviceSpecProto *device_spec_proto)
 888 {
 889         lock_guard<timed_mutex> lock(audio_mutex);
 890         switch (device_spec.type) {
 891                 case InputSourceType::SILENCE:
 892                         device_spec_proto->set_type(DeviceSpecProto::SILENCE);
 893                         break;
 894                 case InputSourceType::CAPTURE_CARD:
 895                         device_spec_proto->set_type(DeviceSpecProto::CAPTURE_CARD);
 896                         device_spec_proto->set_index(device_spec.index);
 897                         device_spec_proto->set_display_name(video_cards[device_spec.index].display_name);
 898                         break;
 899                 case InputSourceType::ALSA_INPUT:
 900                         alsa_pool.serialize_device(device_spec.index, device_spec_proto);
 901                         break;
 902         }
 903 }
 904
 905 void AudioMixer::set_simple_input(unsigned card_index)
 906 {
 907         InputMapping new_input_mapping;
 908         InputMapping::Bus input;
 909         input.name = "Main";
 910         input.device.type = InputSourceType::CAPTURE_CARD;
 911         input.device.index = card_index;
 912         input.source_channel[0] = 0;
 913         input.source_channel[1] = 1;
 914
 915         new_input_mapping.buses.push_back(input);
 916
 917         lock_guard<timed_mutex> lock(audio_mutex);
 918         current_mapping_mode = MappingMode::SIMPLE;
 919         set_input_mapping_lock_held(new_input_mapping);
 920         fader_volume_db[0] = 0.0f;
 921 }
 922
 923 unsigned AudioMixer::get_simple_input() const
 924 {
 925         lock_guard<timed_mutex> lock(audio_mutex);
 926         if (input_mapping.buses.size() == 1 &&
 927             input_mapping.buses[0].device.type == InputSourceType::CAPTURE_CARD &&
 928             input_mapping.buses[0].source_channel[0] == 0 &&
 929             input_mapping.buses[0].source_channel[1] == 1) {
 930                 return input_mapping.buses[0].device.index;
 931         } else {
 932                 return numeric_limits<unsigned>::max();
 933         }
 934 }
 935
 936 void AudioMixer::set_input_mapping(const InputMapping &new_input_mapping)
 937 {
 938         lock_guard<timed_mutex> lock(audio_mutex);
 939         set_input_mapping_lock_held(new_input_mapping);
 940         current_mapping_mode = MappingMode::MULTICHANNEL;
 941 }
 942
 943 AudioMixer::MappingMode AudioMixer::get_mapping_mode() const
 944 {
 945         lock_guard<timed_mutex> lock(audio_mutex);
 946         return current_mapping_mode;
 947 }
 948
 949 void AudioMixer::set_input_mapping_lock_held(const InputMapping &new_input_mapping)
 950 {
 951         map<DeviceSpec, set<unsigned>> interesting_channels;
 952         for (const InputMapping::Bus &bus : new_input_mapping.buses) {
 953                 if (bus.device.type == InputSourceType::CAPTURE_CARD ||
 954                     bus.device.type == InputSourceType::ALSA_INPUT) {
 955                         for (unsigned channel = 0; channel < 2; ++channel) {
 956                                 if (bus.source_channel[channel] != -1) {
 957                                         interesting_channels[bus.device].insert(bus.source_channel[channel]);
 958                                 }
 959                         }
 960                 }
 961         }
 962
 963         // Reset resamplers for all cards that don't have the exact same state as before.
 964         for (unsigned card_index = 0; card_index < MAX_VIDEO_CARDS; ++card_index) {
 965                 const DeviceSpec device_spec{InputSourceType::CAPTURE_CARD, card_index};
 966                 AudioDevice *device = find_audio_device(device_spec);
 967                 if (device->interesting_channels != interesting_channels[device_spec]) {
 968                         device->interesting_channels = interesting_channels[device_spec];
 969                         reset_resampler_mutex_held(device_spec);
 970                 }
 971         }
 972         for (unsigned card_index = 0; card_index < MAX_ALSA_CARDS; ++card_index) {
 973                 const DeviceSpec device_spec{InputSourceType::ALSA_INPUT, card_index};
 974                 AudioDevice *device = find_audio_device(device_spec);
 975                 if (interesting_channels[device_spec].empty()) {
 976                         alsa_pool.release_device(card_index);
 977                 } else {
 978                         alsa_pool.hold_device(card_index);
 979                 }
 980                 if (device->interesting_channels != interesting_channels[device_spec]) {
 981                         device->interesting_channels = interesting_channels[device_spec];
 982                         alsa_pool.reset_device(device_spec.index);
 983                         reset_resampler_mutex_held(device_spec);
 984                 }
 985         }
 986
 987         input_mapping = new_input_mapping;
 988 }
 989
 990 InputMapping AudioMixer::get_input_mapping() const
 991 {
 992         lock_guard<timed_mutex> lock(audio_mutex);
 993         return input_mapping;
 994 }
 995
 996 unsigned AudioMixer::num_buses() const
 997 {
 998         lock_guard<timed_mutex> lock(audio_mutex);
 999         return input_mapping.buses.size();
1000 }
1001
1002 void AudioMixer::reset_peak(unsigned bus_index)
1003 {
1004         lock_guard<timed_mutex> lock(audio_mutex);
1005         for (unsigned channel = 0; channel < 2; ++channel) {
1006                 PeakHistory &history = peak_history[bus_index][channel];
1007                 history.current_level = 0.0f;
1008                 history.historic_peak = 0.0f;
1009                 history.current_peak = 0.0f;
1010                 history.last_peak = 0.0f;
1011                 history.age_seconds = 0.0f;
1012         }
1013 }
1014
1015 AudioMixer *global_audio_mixer = nullptr;