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::placeholders;
  29
  30 namespace {
  31
  32 // TODO: If these prove to be a bottleneck, they can be SSSE3-optimized
  33 // (usually including multiple channels at a time).
  34
  35 void convert_fixed16_to_fp32(float *dst, size_t out_channel, size_t out_num_channels,
  36                              const uint8_t *src, size_t in_channel, size_t in_num_channels,
  37                              size_t num_samples)
  38 {
  39         assert(in_channel < in_num_channels);
  40         assert(out_channel < out_num_channels);
  41         src += in_channel * 2;
  42         dst += out_channel;
  43
  44         for (size_t i = 0; i < num_samples; ++i) {
  45                 int16_t s = le16toh(*(int16_t *)src);
  46                 *dst = s * (1.0f / 32768.0f);
  47
  48                 src += 2 * in_num_channels;
  49                 dst += out_num_channels;
  50         }
  51 }
  52
  53 void convert_fixed24_to_fp32(float *dst, size_t out_channel, size_t out_num_channels,
  54                              const uint8_t *src, size_t in_channel, size_t in_num_channels,
  55                              size_t num_samples)
  56 {
  57         assert(in_channel < in_num_channels);
  58         assert(out_channel < out_num_channels);
  59         src += in_channel * 3;
  60         dst += out_channel;
  61
  62         for (size_t i = 0; i < num_samples; ++i) {
  63                 uint32_t s1 = src[0];
  64                 uint32_t s2 = src[1];
  65                 uint32_t s3 = src[2];
  66                 uint32_t s = s1 | (s1 << 8) | (s2 << 16) | (s3 << 24);
  67                 *dst = int(s) * (1.0f / 2147483648.0f);
  68
  69                 src += 3 * in_num_channels;
  70                 dst += out_num_channels;
  71         }
  72 }
  73
  74 void convert_fixed32_to_fp32(float *dst, size_t out_channel, size_t out_num_channels,
  75                              const uint8_t *src, size_t in_channel, size_t in_num_channels,
  76                              size_t num_samples)
  77 {
  78         assert(in_channel < in_num_channels);
  79         assert(out_channel < out_num_channels);
  80         src += in_channel * 4;
  81         dst += out_channel;
  82
  83         for (size_t i = 0; i < num_samples; ++i) {
  84                 int32_t s = le32toh(*(int32_t *)src);
  85                 *dst = s * (1.0f / 2147483648.0f);
  86
  87                 src += 4 * in_num_channels;
  88                 dst += out_num_channels;
  89         }
  90 }
  91
  92 float find_peak_plain(const float *samples, size_t num_samples) __attribute__((unused));
  93
  94 float find_peak_plain(const float *samples, size_t num_samples)
  95 {
  96         float m = fabs(samples[0]);
  97         for (size_t i = 1; i < num_samples; ++i) {
  98                 m = max(m, fabs(samples[i]));
  99         }
 100         return m;
 101 }
 102
 103 #ifdef __SSE__
 104 static inline float horizontal_max(__m128 m)
 105 {
 106         __m128 tmp = _mm_shuffle_ps(m, m, _MM_SHUFFLE(1, 0, 3, 2));
 107         m = _mm_max_ps(m, tmp);
 108         tmp = _mm_shuffle_ps(m, m, _MM_SHUFFLE(2, 3, 0, 1));
 109         m = _mm_max_ps(m, tmp);
 110         return _mm_cvtss_f32(m);
 111 }
 112
 113 float find_peak(const float *samples, size_t num_samples)
 114 {
 115         const __m128 abs_mask = _mm_castsi128_ps(_mm_set1_epi32(0x7fffffffu));
 116         __m128 m = _mm_setzero_ps();
 117         for (size_t i = 0; i < (num_samples & ~3); i += 4) {
 118                 __m128 x = _mm_loadu_ps(samples + i);
 119                 x = _mm_and_ps(x, abs_mask);
 120                 m = _mm_max_ps(m, x);
 121         }
 122         float result = horizontal_max(m);
 123
 124         for (size_t i = (num_samples & ~3); i < num_samples; ++i) {
 125                 result = max(result, fabs(samples[i]));
 126         }
 127
 128 #if 0
 129         // Self-test. We should be bit-exact the same.
 130         float reference_result = find_peak_plain(samples, num_samples);
 131         if (result != reference_result) {
 132                 fprintf(stderr, "Error: Peak is %f [%f %f %f %f]; should be %f.\n",
 133                         result,
 134                         _mm_cvtss_f32(_mm_shuffle_ps(m, m, _MM_SHUFFLE(0, 0, 0, 0))),
 135                         _mm_cvtss_f32(_mm_shuffle_ps(m, m, _MM_SHUFFLE(1, 1, 1, 1))),
 136                         _mm_cvtss_f32(_mm_shuffle_ps(m, m, _MM_SHUFFLE(2, 2, 2, 2))),
 137                         _mm_cvtss_f32(_mm_shuffle_ps(m, m, _MM_SHUFFLE(3, 3, 3, 3))),
 138                         reference_result);
 139                 abort();
 140         }
 141 #endif
 142         return result;
 143 }
 144 #else
 145 float find_peak(const float *samples, size_t num_samples)
 146 {
 147         return find_peak_plain(samples, num_samples);
 148 }
 149 #endif
 150
 151 void deinterleave_samples(const vector<float> &in, vector<float> *out_l, vector<float> *out_r)
 152 {
 153         size_t num_samples = in.size() / 2;
 154         out_l->resize(num_samples);
 155         out_r->resize(num_samples);
 156
 157         const float *inptr = in.data();
 158         float *lptr = &(*out_l)[0];
 159         float *rptr = &(*out_r)[0];
 160         for (size_t i = 0; i < num_samples; ++i) {
 161                 *lptr++ = *inptr++;
 162                 *rptr++ = *inptr++;
 163         }
 164 }
 165
 166 }  // namespace
 167
 168 AudioMixer::AudioMixer(unsigned num_cards)
 169         : num_cards(num_cards),
 170           limiter(OUTPUT_FREQUENCY),
 171           correlation(OUTPUT_FREQUENCY)
 172 {
 173         for (unsigned bus_index = 0; bus_index < MAX_BUSES; ++bus_index) {
 174                 locut[bus_index].init(FILTER_HPF, 2);
 175                 eq[bus_index][EQ_BAND_BASS].init(FILTER_LOW_SHELF, 1);
 176                 // Note: EQ_BAND_MID isn't used (see comments in apply_eq()).
 177                 eq[bus_index][EQ_BAND_TREBLE].init(FILTER_HIGH_SHELF, 1);
 178                 compressor[bus_index].reset(new StereoCompressor(OUTPUT_FREQUENCY));
 179                 level_compressor[bus_index].reset(new StereoCompressor(OUTPUT_FREQUENCY));
 180
 181                 set_bus_settings(bus_index, get_default_bus_settings());
 182         }
 183         set_limiter_enabled(global_flags.limiter_enabled);
 184         set_final_makeup_gain_auto(global_flags.final_makeup_gain_auto);
 185
 186         if (!global_flags.input_mapping_filename.empty()) {
 187                 current_mapping_mode = MappingMode::MULTICHANNEL;
 188                 InputMapping new_input_mapping;
 189                 if (!load_input_mapping_from_file(get_devices(),
 190                                                   global_flags.input_mapping_filename,
 191                                                   &new_input_mapping)) {
 192                         fprintf(stderr, "Failed to load input mapping from '%s', exiting.\n",
 193                                 global_flags.input_mapping_filename.c_str());
 194                         exit(1);
 195                 }
 196                 set_input_mapping(new_input_mapping);
 197         } else {
 198                 set_simple_input(/*card_index=*/0);
 199                 if (global_flags.multichannel_mapping_mode) {
 200                         current_mapping_mode = MappingMode::MULTICHANNEL;
 201                 }
 202         }
 203
 204         r128.init(2, OUTPUT_FREQUENCY);
 205         r128.integr_start();
 206
 207         // hlen=16 is pretty low quality, but we use quite a bit of CPU otherwise,
 208         // and there's a limit to how important the peak meter is.
 209         peak_resampler.setup(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY * 4, /*num_channels=*/2, /*hlen=*/16, /*frel=*/1.0);
 210
 211         global_audio_mixer = this;
 212         alsa_pool.init();
 213 }
 214
 215 void AudioMixer::reset_resampler(DeviceSpec device_spec)
 216 {
 217         lock_guard<timed_mutex> lock(audio_mutex);
 218         reset_resampler_mutex_held(device_spec);
 219 }
 220
 221 void AudioMixer::reset_resampler_mutex_held(DeviceSpec device_spec)
 222 {
 223         AudioDevice *device = find_audio_device(device_spec);
 224
 225         if (device->interesting_channels.empty()) {
 226                 device->resampling_queue.reset();
 227         } else {
 228                 // TODO: ResamplingQueue should probably take the full device spec.
 229                 // (It's only used for console output, though.)
 230                 device->resampling_queue.reset(new ResamplingQueue(
 231                         device_spec.index, device->capture_frequency, OUTPUT_FREQUENCY, device->interesting_channels.size(),
 232                         global_flags.audio_queue_length_ms * 0.001));
 233         }
 234         device->next_local_pts = 0;
 235 }
 236
 237 bool AudioMixer::add_audio(DeviceSpec device_spec, const uint8_t *data, unsigned num_samples, AudioFormat audio_format, int64_t frame_length)
 238 {
 239         AudioDevice *device = find_audio_device(device_spec);
 240
 241         unique_lock<timed_mutex> lock(audio_mutex, defer_lock);
 242         if (!lock.try_lock_for(chrono::milliseconds(10))) {
 243                 return false;
 244         }
 245         if (device->resampling_queue == nullptr) {
 246                 // No buses use this device; throw it away.
 247                 return true;
 248         }
 249
 250         unsigned num_channels = device->interesting_channels.size();
 251         assert(num_channels > 0);
 252
 253         // Convert the audio to fp32.
 254         unique_ptr<float[]> audio(new float[num_samples * num_channels]);
 255         unsigned channel_index = 0;
 256         for (auto channel_it = device->interesting_channels.cbegin(); channel_it != device->interesting_channels.end(); ++channel_it, ++channel_index) {
 257                 switch (audio_format.bits_per_sample) {
 258                 case 0:
 259                         assert(num_samples == 0);
 260                         break;
 261                 case 16:
 262                         convert_fixed16_to_fp32(audio.get(), channel_index, num_channels, data, *channel_it, audio_format.num_channels, num_samples);
 263                         break;
 264                 case 24:
 265                         convert_fixed24_to_fp32(audio.get(), channel_index, num_channels, data, *channel_it, audio_format.num_channels, num_samples);
 266                         break;
 267                 case 32:
 268                         convert_fixed32_to_fp32(audio.get(), channel_index, num_channels, data, *channel_it, audio_format.num_channels, num_samples);
 269                         break;
 270                 default:
 271                         fprintf(stderr, "Cannot handle audio with %u bits per sample\n", audio_format.bits_per_sample);
 272                         assert(false);
 273                 }
 274         }
 275
 276         // Now add it.
 277         int64_t local_pts = device->next_local_pts;
 278         device->resampling_queue->add_input_samples(local_pts / double(TIMEBASE), audio.get(), num_samples);
 279         device->next_local_pts = local_pts + frame_length;
 280         return true;
 281 }
 282
 283 bool AudioMixer::add_silence(DeviceSpec device_spec, unsigned samples_per_frame, unsigned num_frames, int64_t frame_length)
 284 {
 285         AudioDevice *device = find_audio_device(device_spec);
 286
 287         unique_lock<timed_mutex> lock(audio_mutex, defer_lock);
 288         if (!lock.try_lock_for(chrono::milliseconds(10))) {
 289                 return false;
 290         }
 291         if (device->resampling_queue == nullptr) {
 292                 // No buses use this device; throw it away.
 293                 return true;
 294         }
 295
 296         unsigned num_channels = device->interesting_channels.size();
 297         assert(num_channels > 0);
 298
 299         vector<float> silence(samples_per_frame * num_channels, 0.0f);
 300         for (unsigned i = 0; i < num_frames; ++i) {
 301                 device->resampling_queue->add_input_samples(device->next_local_pts / double(TIMEBASE), silence.data(), samples_per_frame);
 302                 // Note that if the format changed in the meantime, we have
 303                 // no way of detecting that; we just have to assume the frame length
 304                 // is always the same.
 305                 device->next_local_pts += frame_length;
 306         }
 307         return true;
 308 }
 309
 310 bool AudioMixer::silence_card(DeviceSpec device_spec, bool silence)
 311 {
 312         AudioDevice *device = find_audio_device(device_spec);
 313
 314         unique_lock<timed_mutex> lock(audio_mutex, defer_lock);
 315         if (!lock.try_lock_for(chrono::milliseconds(10))) {
 316                 return false;
 317         }
 318
 319         if (device->silenced && !silence) {
 320                 reset_resampler_mutex_held(device_spec);
 321         }
 322         device->silenced = silence;
 323         return true;
 324 }
 325
 326 AudioMixer::BusSettings AudioMixer::get_default_bus_settings()
 327 {
 328         BusSettings settings;
 329         settings.fader_volume_db = 0.0f;
 330         settings.muted = false;
 331         settings.locut_enabled = global_flags.locut_enabled;
 332         for (unsigned band_index = 0; band_index < NUM_EQ_BANDS; ++band_index) {
 333                 settings.eq_level_db[band_index] = 0.0f;
 334         }
 335         settings.gain_staging_db = global_flags.initial_gain_staging_db;
 336         settings.level_compressor_enabled = global_flags.gain_staging_auto;
 337         settings.compressor_threshold_dbfs = ref_level_dbfs - 12.0f;  // -12 dB.
 338         settings.compressor_enabled = global_flags.compressor_enabled;
 339         return settings;
 340 }
 341
 342 AudioMixer::BusSettings AudioMixer::get_bus_settings(unsigned bus_index) const
 343 {
 344         lock_guard<timed_mutex> lock(audio_mutex);
 345         BusSettings settings;
 346         settings.fader_volume_db = fader_volume_db[bus_index];
 347         settings.muted = mute[bus_index];
 348         settings.locut_enabled = locut_enabled[bus_index];
 349         for (unsigned band_index = 0; band_index < NUM_EQ_BANDS; ++band_index) {
 350                 settings.eq_level_db[band_index] = eq_level_db[bus_index][band_index];
 351         }
 352         settings.gain_staging_db = gain_staging_db[bus_index];
 353         settings.level_compressor_enabled = level_compressor_enabled[bus_index];
 354         settings.compressor_threshold_dbfs = compressor_threshold_dbfs[bus_index];
 355         settings.compressor_enabled = compressor_enabled[bus_index];
 356         return settings;
 357 }
 358
 359 void AudioMixer::set_bus_settings(unsigned bus_index, const AudioMixer::BusSettings &settings)
 360 {
 361         lock_guard<timed_mutex> lock(audio_mutex);
 362         fader_volume_db[bus_index] = settings.fader_volume_db;
 363         mute[bus_index] = settings.muted;
 364         locut_enabled[bus_index] = settings.locut_enabled;
 365         for (unsigned band_index = 0; band_index < NUM_EQ_BANDS; ++band_index) {
 366                 eq_level_db[bus_index][band_index] = settings.eq_level_db[band_index];
 367         }
 368         gain_staging_db[bus_index] = settings.gain_staging_db;
 369         last_gain_staging_db[bus_index] = gain_staging_db[bus_index];
 370         level_compressor_enabled[bus_index] = settings.level_compressor_enabled;
 371         compressor_threshold_dbfs[bus_index] = settings.compressor_threshold_dbfs;
 372         compressor_enabled[bus_index] = settings.compressor_enabled;
 373 }
 374
 375 AudioMixer::AudioDevice *AudioMixer::find_audio_device(DeviceSpec device)
 376 {
 377         switch (device.type) {
 378         case InputSourceType::CAPTURE_CARD:
 379                 return &video_cards[device.index];
 380         case InputSourceType::ALSA_INPUT:
 381                 return &alsa_inputs[device.index];
 382         case InputSourceType::SILENCE:
 383         default:
 384                 assert(false);
 385         }
 386         return nullptr;
 387 }
 388
 389 // Get a pointer to the given channel from the given device.
 390 // The channel must be picked out earlier and resampled.
 391 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)
 392 {
 393         static float zero = 0.0f;
 394         if (source_channel == -1 || device_spec.type == InputSourceType::SILENCE) {
 395                 *srcptr = &zero;
 396                 *stride = 0;
 397                 return;
 398         }
 399         AudioDevice *device = find_audio_device(device_spec);
 400         assert(device->interesting_channels.count(source_channel) != 0);
 401         unsigned channel_index = 0;
 402         for (int channel : device->interesting_channels) {
 403                 if (channel == source_channel) break;
 404                 ++channel_index;
 405         }
 406         assert(channel_index < device->interesting_channels.size());
 407         const auto it = samples_card.find(device_spec);
 408         assert(it != samples_card.end());
 409         *srcptr = &(it->second)[channel_index];
 410         *stride = device->interesting_channels.size();
 411 }
 412
 413 // TODO: Can be SSSE3-optimized if need be.
 414 void AudioMixer::fill_audio_bus(const map<DeviceSpec, vector<float>> &samples_card, const InputMapping::Bus &bus, unsigned num_samples, float *output)
 415 {
 416         if (bus.device.type == InputSourceType::SILENCE) {
 417                 memset(output, 0, num_samples * 2 * sizeof(*output));
 418         } else {
 419                 assert(bus.device.type == InputSourceType::CAPTURE_CARD ||
 420                        bus.device.type == InputSourceType::ALSA_INPUT);
 421                 const float *lsrc, *rsrc;
 422                 unsigned lstride, rstride;
 423                 float *dptr = output;
 424                 find_sample_src_from_device(samples_card, bus.device, bus.source_channel[0], &lsrc, &lstride);
 425                 find_sample_src_from_device(samples_card, bus.device, bus.source_channel[1], &rsrc, &rstride);
 426                 for (unsigned i = 0; i < num_samples; ++i) {
 427                         *dptr++ = *lsrc;
 428                         *dptr++ = *rsrc;
 429                         lsrc += lstride;
 430                         rsrc += rstride;
 431                 }
 432         }
 433 }
 434
 435 vector<DeviceSpec> AudioMixer::get_active_devices() const
 436 {
 437         vector<DeviceSpec> ret;
 438         for (unsigned card_index = 0; card_index < MAX_VIDEO_CARDS; ++card_index) {
 439                 const DeviceSpec device_spec{InputSourceType::CAPTURE_CARD, card_index};
 440                 if (!find_audio_device(device_spec)->interesting_channels.empty()) {
 441                         ret.push_back(device_spec);
 442                 }
 443         }
 444         for (unsigned card_index = 0; card_index < MAX_ALSA_CARDS; ++card_index) {
 445                 const DeviceSpec device_spec{InputSourceType::ALSA_INPUT, card_index};
 446                 if (!find_audio_device(device_spec)->interesting_channels.empty()) {
 447                         ret.push_back(device_spec);
 448                 }
 449         }
 450         return ret;
 451 }
 452
 453 namespace {
 454
 455 void apply_gain(float db, float last_db, vector<float> *samples)
 456 {
 457         if (fabs(db - last_db) < 1e-3) {
 458                 // Constant over this frame.
 459                 const float gain = from_db(db);
 460                 for (size_t i = 0; i < samples->size(); ++i) {
 461                         (*samples)[i] *= gain;
 462                 }
 463         } else {
 464                 // We need to do a fade.
 465                 unsigned num_samples = samples->size() / 2;
 466                 float gain = from_db(last_db);
 467                 const float gain_inc = pow(from_db(db - last_db), 1.0 / num_samples);
 468                 for (size_t i = 0; i < num_samples; ++i) {
 469                         (*samples)[i * 2 + 0] *= gain;
 470                         (*samples)[i * 2 + 1] *= gain;
 471                         gain *= gain_inc;
 472                 }
 473         }
 474 }
 475
 476 }  // namespace
 477
 478 vector<float> AudioMixer::get_output(double pts, unsigned num_samples, ResamplingQueue::RateAdjustmentPolicy rate_adjustment_policy)
 479 {
 480         map<DeviceSpec, vector<float>> samples_card;
 481         vector<float> samples_bus;
 482
 483         lock_guard<timed_mutex> lock(audio_mutex);
 484
 485         // Pick out all the interesting channels from all the cards.
 486         for (const DeviceSpec &device_spec : get_active_devices()) {
 487                 AudioDevice *device = find_audio_device(device_spec);
 488                 samples_card[device_spec].resize(num_samples * device->interesting_channels.size());
 489                 if (device->silenced) {
 490                         memset(&samples_card[device_spec][0], 0, samples_card[device_spec].size() * sizeof(float));
 491                 } else {
 492                         device->resampling_queue->get_output_samples(
 493                                 pts,
 494                                 &samples_card[device_spec][0],
 495                                 num_samples,
 496                                 rate_adjustment_policy);
 497                 }
 498         }
 499
 500         vector<float> samples_out, left, right;
 501         samples_out.resize(num_samples * 2);
 502         samples_bus.resize(num_samples * 2);
 503         for (unsigned bus_index = 0; bus_index < input_mapping.buses.size(); ++bus_index) {
 504                 fill_audio_bus(samples_card, input_mapping.buses[bus_index], num_samples, &samples_bus[0]);
 505                 apply_eq(bus_index, &samples_bus);
 506
 507                 {
 508                         lock_guard<mutex> lock(compressor_mutex);
 509
 510                         // Apply a level compressor to get the general level right.
 511                         // Basically, if it's over about -40 dBFS, we squeeze it down to that level
 512                         // (or more precisely, near it, since we don't use infinite ratio),
 513                         // then apply a makeup gain to get it to -14 dBFS. -14 dBFS is, of course,
 514                         // entirely arbitrary, but from practical tests with speech, it seems to
 515                         // put ut around -23 LUFS, so it's a reasonable starting point for later use.
 516                         if (level_compressor_enabled[bus_index]) {
 517                                 float threshold = 0.01f;   // -40 dBFS.
 518                                 float ratio = 20.0f;
 519                                 float attack_time = 0.5f;
 520                                 float release_time = 20.0f;
 521                                 float makeup_gain = from_db(ref_level_dbfs - (-40.0f));  // +26 dB.
 522                                 level_compressor[bus_index]->process(samples_bus.data(), samples_bus.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 523                                 gain_staging_db[bus_index] = to_db(level_compressor[bus_index]->get_attenuation() * makeup_gain);
 524                         } else {
 525                                 // Just apply the gain we already had.
 526                                 float db = gain_staging_db[bus_index];
 527                                 float last_db = last_gain_staging_db[bus_index];
 528                                 apply_gain(db, last_db, &samples_bus);
 529                         }
 530                         last_gain_staging_db[bus_index] = gain_staging_db[bus_index];
 531
 532 #if 0
 533                         printf("level=%f (%+5.2f dBFS) attenuation=%f (%+5.2f dB) end_result=%+5.2f dB\n",
 534                                 level_compressor.get_level(), to_db(level_compressor.get_level()),
 535                                 level_compressor.get_attenuation(), to_db(level_compressor.get_attenuation()),
 536                                 to_db(level_compressor.get_level() * level_compressor.get_attenuation() * makeup_gain));
 537 #endif
 538
 539                         // The real compressor.
 540                         if (compressor_enabled[bus_index]) {
 541                                 float threshold = from_db(compressor_threshold_dbfs[bus_index]);
 542                                 float ratio = 20.0f;
 543                                 float attack_time = 0.005f;
 544                                 float release_time = 0.040f;
 545                                 float makeup_gain = 2.0f;  // +6 dB.
 546                                 compressor[bus_index]->process(samples_bus.data(), samples_bus.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 547                 //              compressor_att = compressor.get_attenuation();
 548                         }
 549                 }
 550
 551                 add_bus_to_master(bus_index, samples_bus, &samples_out);
 552                 deinterleave_samples(samples_bus, &left, &right);
 553                 measure_bus_levels(bus_index, left, right);
 554         }
 555
 556         {
 557                 lock_guard<mutex> lock(compressor_mutex);
 558
 559                 // Finally a limiter at -4 dB (so, -10 dBFS) to take out the worst peaks only.
 560                 // Note that since ratio is not infinite, we could go slightly higher than this.
 561                 if (limiter_enabled) {
 562                         float threshold = from_db(limiter_threshold_dbfs);
 563                         float ratio = 30.0f;
 564                         float attack_time = 0.0f;  // Instant.
 565                         float release_time = 0.020f;
 566                         float makeup_gain = 1.0f;  // 0 dB.
 567                         limiter.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 568         //              limiter_att = limiter.get_attenuation();
 569                 }
 570
 571         //      printf("limiter=%+5.1f  compressor=%+5.1f\n", to_db(limiter_att), to_db(compressor_att));
 572         }
 573
 574         // At this point, we are most likely close to +0 LU (at least if the
 575         // faders sum to 0 dB and the compressors are on), but all of our
 576         // measurements have been on raw sample values, not R128 values.
 577         // So we have a final makeup gain to get us to +0 LU; the gain
 578         // adjustments required should be relatively small, and also, the
 579         // offset shouldn't change much (only if the type of audio changes
 580         // significantly). Thus, we shoot for updating this value basically
 581         // “whenever we process buffers”, since the R128 calculation isn't exactly
 582         // something we get out per-sample.
 583         //
 584         // Note that there's a feedback loop here, so we choose a very slow filter
 585         // (half-time of 30 seconds).
 586         double target_loudness_factor, alpha;
 587         double loudness_lu = r128.loudness_M() - ref_level_lufs;
 588         target_loudness_factor = final_makeup_gain * from_db(-loudness_lu);
 589
 590         // If we're outside +/- 5 LU (after correction), we don't count it as
 591         // a normal signal (probably silence) and don't change the
 592         // correction factor; just apply what we already have.
 593         if (fabs(loudness_lu) >= 5.0 || !final_makeup_gain_auto) {
 594                 alpha = 0.0;
 595         } else {
 596                 // Formula adapted from
 597                 // https://en.wikipedia.org/wiki/Low-pass_filter#Simple_infinite_impulse_response_filter.
 598                 const double half_time_s = 30.0;
 599                 const double fc_mul_2pi_delta_t = 1.0 / (half_time_s * OUTPUT_FREQUENCY);
 600                 alpha = fc_mul_2pi_delta_t / (fc_mul_2pi_delta_t + 1.0);
 601         }
 602
 603         {
 604                 lock_guard<mutex> lock(compressor_mutex);
 605                 double m = final_makeup_gain;
 606                 for (size_t i = 0; i < samples_out.size(); i += 2) {
 607                         samples_out[i + 0] *= m;
 608                         samples_out[i + 1] *= m;
 609                         m += (target_loudness_factor - m) * alpha;
 610                 }
 611                 final_makeup_gain = m;
 612         }
 613
 614         update_meters(samples_out);
 615
 616         return samples_out;
 617 }
 618
 619 namespace {
 620
 621 void apply_filter_fade(StereoFilter *filter, float *data, unsigned num_samples, float cutoff_hz, float db, float last_db)
 622 {
 623         // A granularity of 32 samples is an okay tradeoff between speed and
 624         // smoothness; recalculating the filters is pretty expensive, so it's
 625         // good that we don't do this all the time.
 626         static constexpr unsigned filter_granularity_samples = 32;
 627
 628         const float cutoff_linear = cutoff_hz * 2.0 * M_PI / OUTPUT_FREQUENCY;
 629         if (fabs(db - last_db) < 1e-3) {
 630                 // Constant over this frame.
 631                 if (fabs(db) > 0.01f) {
 632                         filter->render(data, num_samples, cutoff_linear, 0.5f, db / 40.0f);
 633                 }
 634         } else {
 635                 // We need to do a fade. (Rounding up avoids division by zero.)
 636                 unsigned num_blocks = (num_samples + filter_granularity_samples - 1) / filter_granularity_samples;
 637                 const float inc_db_norm = (db - last_db) / 40.0f / num_blocks;
 638                 float db_norm = db / 40.0f;
 639                 for (size_t i = 0; i < num_samples; i += filter_granularity_samples) {
 640                         size_t samples_this_block = std::min<size_t>(num_samples - i, filter_granularity_samples);
 641                         filter->render(data + i * 2, samples_this_block, cutoff_linear, 0.5f, db_norm);
 642                         db_norm += inc_db_norm;
 643                 }
 644         }
 645 }
 646
 647 }  // namespace
 648
 649 void AudioMixer::apply_eq(unsigned bus_index, vector<float> *samples_bus)
 650 {
 651         constexpr float bass_freq_hz = 200.0f;
 652         constexpr float treble_freq_hz = 4700.0f;
 653
 654         // Cut away everything under 120 Hz (or whatever the cutoff is);
 655         // we don't need it for voice, and it will reduce headroom
 656         // and confuse the compressor. (In particular, any hums at 50 or 60 Hz
 657         // should be dampened.)
 658         if (locut_enabled[bus_index]) {
 659                 locut[bus_index].render(samples_bus->data(), samples_bus->size() / 2, locut_cutoff_hz * 2.0 * M_PI / OUTPUT_FREQUENCY, 0.5f);
 660         }
 661
 662         // Apply the rest of the EQ. Since we only have a simple three-band EQ,
 663         // we can implement it with two shelf filters. We use a simple gain to
 664         // set the mid-level filter, and then offset the low and high bands
 665         // from that if we need to. (We could perhaps have folded the gain into
 666         // the next part, but it's so cheap that the trouble isn't worth it.)
 667         //
 668         // If any part of the EQ has changed appreciably since last frame,
 669         // we fade smoothly during the course of this frame.
 670         const float bass_db = eq_level_db[bus_index][EQ_BAND_BASS];
 671         const float mid_db = eq_level_db[bus_index][EQ_BAND_MID];
 672         const float treble_db = eq_level_db[bus_index][EQ_BAND_TREBLE];
 673
 674         const float last_bass_db = last_eq_level_db[bus_index][EQ_BAND_BASS];
 675         const float last_mid_db = last_eq_level_db[bus_index][EQ_BAND_MID];
 676         const float last_treble_db = last_eq_level_db[bus_index][EQ_BAND_TREBLE];
 677
 678         assert(samples_bus->size() % 2 == 0);
 679         const unsigned num_samples = samples_bus->size() / 2;
 680
 681         apply_gain(mid_db, last_mid_db, samples_bus);
 682
 683         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);
 684         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);
 685
 686         last_eq_level_db[bus_index][EQ_BAND_BASS] = bass_db;
 687         last_eq_level_db[bus_index][EQ_BAND_MID] = mid_db;
 688         last_eq_level_db[bus_index][EQ_BAND_TREBLE] = treble_db;
 689 }
 690
 691 void AudioMixer::add_bus_to_master(unsigned bus_index, const vector<float> &samples_bus, vector<float> *samples_out)
 692 {
 693         assert(samples_bus.size() == samples_out->size());
 694         assert(samples_bus.size() % 2 == 0);
 695         unsigned num_samples = samples_bus.size() / 2;
 696         const float new_volume_db = mute[bus_index] ? -90.0f : fader_volume_db[bus_index].load();
 697         if (fabs(new_volume_db - last_fader_volume_db[bus_index]) > 1e-3) {
 698                 // The volume has changed; do a fade over the course of this frame.
 699                 // (We might have some numerical issues here, but it seems to sound OK.)
 700                 // For the purpose of fading here, the silence floor is set to -90 dB
 701                 // (the fader only goes to -84).
 702                 float old_volume = from_db(max<float>(last_fader_volume_db[bus_index], -90.0f));
 703                 float volume = from_db(max<float>(new_volume_db, -90.0f));
 704
 705                 float volume_inc = pow(volume / old_volume, 1.0 / num_samples);
 706                 volume = old_volume;
 707                 if (bus_index == 0) {
 708                         for (unsigned i = 0; i < num_samples; ++i) {
 709                                 (*samples_out)[i * 2 + 0] = samples_bus[i * 2 + 0] * volume;
 710                                 (*samples_out)[i * 2 + 1] = samples_bus[i * 2 + 1] * volume;
 711                                 volume *= volume_inc;
 712                         }
 713                 } else {
 714                         for (unsigned i = 0; i < num_samples; ++i) {
 715                                 (*samples_out)[i * 2 + 0] += samples_bus[i * 2 + 0] * volume;
 716                                 (*samples_out)[i * 2 + 1] += samples_bus[i * 2 + 1] * volume;
 717                                 volume *= volume_inc;
 718                         }
 719                 }
 720         } else if (new_volume_db > -90.0f) {
 721                 float volume = from_db(new_volume_db);
 722                 if (bus_index == 0) {
 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                 } else {
 728                         for (unsigned i = 0; i < num_samples; ++i) {
 729                                 (*samples_out)[i * 2 + 0] += samples_bus[i * 2 + 0] * volume;
 730                                 (*samples_out)[i * 2 + 1] += samples_bus[i * 2 + 1] * volume;
 731                         }
 732                 }
 733         }
 734
 735         last_fader_volume_db[bus_index] = new_volume_db;
 736 }
 737
 738 void AudioMixer::measure_bus_levels(unsigned bus_index, const vector<float> &left, const vector<float> &right)
 739 {
 740         assert(left.size() == right.size());
 741         const float volume = mute[bus_index] ? 0.0f : from_db(fader_volume_db[bus_index]);
 742         const float peak_levels[2] = {
 743                 find_peak(left.data(), left.size()) * volume,
 744                 find_peak(right.data(), right.size()) * volume
 745         };
 746         for (unsigned channel = 0; channel < 2; ++channel) {
 747                 // Compute the current value, including hold and falloff.
 748                 // The constants are borrowed from zita-mu1 by Fons Adriaensen.
 749                 static constexpr float hold_sec = 0.5f;
 750                 static constexpr float falloff_db_sec = 15.0f;  // dB/sec falloff after hold.
 751                 float current_peak;
 752                 PeakHistory &history = peak_history[bus_index][channel];
 753                 history.historic_peak = max(history.historic_peak, peak_levels[channel]);
 754                 if (history.age_seconds < hold_sec) {
 755                         current_peak = history.last_peak;
 756                 } else {
 757                         current_peak = history.last_peak * from_db(-falloff_db_sec * (history.age_seconds - hold_sec));
 758                 }
 759
 760                 // See if we have a new peak to replace the old (possibly falling) one.
 761                 if (peak_levels[channel] > current_peak) {
 762                         history.last_peak = peak_levels[channel];
 763                         history.age_seconds = 0.0f;  // Not 100% correct, but more than good enough given our frame sizes.
 764                         current_peak = peak_levels[channel];
 765                 } else {
 766                         history.age_seconds += float(left.size()) / OUTPUT_FREQUENCY;
 767                 }
 768                 history.current_level = peak_levels[channel];
 769                 history.current_peak = current_peak;
 770         }
 771 }
 772
 773 void AudioMixer::update_meters(const vector<float> &samples)
 774 {
 775         // Upsample 4x to find interpolated peak.
 776         peak_resampler.inp_data = const_cast<float *>(samples.data());
 777         peak_resampler.inp_count = samples.size() / 2;
 778
 779         vector<float> interpolated_samples;
 780         interpolated_samples.resize(samples.size());
 781         {
 782                 lock_guard<mutex> lock(audio_measure_mutex);
 783
 784                 while (peak_resampler.inp_count > 0) {  // About four iterations.
 785                         peak_resampler.out_data = &interpolated_samples[0];
 786                         peak_resampler.out_count = interpolated_samples.size() / 2;
 787                         peak_resampler.process();
 788                         size_t out_stereo_samples = interpolated_samples.size() / 2 - peak_resampler.out_count;
 789                         peak = max<float>(peak, find_peak(interpolated_samples.data(), out_stereo_samples * 2));
 790                         peak_resampler.out_data = nullptr;
 791                 }
 792         }
 793
 794         // Find R128 levels and L/R correlation.
 795         vector<float> left, right;
 796         deinterleave_samples(samples, &left, &right);
 797         float *ptrs[] = { left.data(), right.data() };
 798         {
 799                 lock_guard<mutex> lock(audio_measure_mutex);
 800                 r128.process(left.size(), ptrs);
 801                 correlation.process_samples(samples);
 802         }
 803
 804         send_audio_level_callback();
 805 }
 806
 807 void AudioMixer::reset_meters()
 808 {
 809         lock_guard<mutex> lock(audio_measure_mutex);
 810         peak_resampler.reset();
 811         peak = 0.0f;
 812         r128.reset();
 813         r128.integr_start();
 814         correlation.reset();
 815 }
 816
 817 void AudioMixer::send_audio_level_callback()
 818 {
 819         if (audio_level_callback == nullptr) {
 820                 return;
 821         }
 822
 823         lock_guard<mutex> lock(audio_measure_mutex);
 824         double loudness_s = r128.loudness_S();
 825         double loudness_i = r128.integrated();
 826         double loudness_range_low = r128.range_min();
 827         double loudness_range_high = r128.range_max();
 828
 829         vector<BusLevel> bus_levels;
 830         bus_levels.resize(input_mapping.buses.size());
 831         {
 832                 lock_guard<mutex> lock(compressor_mutex);
 833                 for (unsigned bus_index = 0; bus_index < bus_levels.size(); ++bus_index) {
 834                         bus_levels[bus_index].current_level_dbfs[0] = to_db(peak_history[bus_index][0].current_level);
 835                         bus_levels[bus_index].current_level_dbfs[1] = to_db(peak_history[bus_index][1].current_level);
 836                         bus_levels[bus_index].peak_level_dbfs[0] = to_db(peak_history[bus_index][0].current_peak);
 837                         bus_levels[bus_index].peak_level_dbfs[1] = to_db(peak_history[bus_index][1].current_peak);
 838                         bus_levels[bus_index].historic_peak_dbfs = to_db(
 839                                 max(peak_history[bus_index][0].historic_peak,
 840                                     peak_history[bus_index][1].historic_peak));
 841                         bus_levels[bus_index].gain_staging_db = gain_staging_db[bus_index];
 842                         if (compressor_enabled[bus_index]) {
 843                                 bus_levels[bus_index].compressor_attenuation_db = -to_db(compressor[bus_index]->get_attenuation());
 844                         } else {
 845                                 bus_levels[bus_index].compressor_attenuation_db = 0.0;
 846                         }
 847                 }
 848         }
 849
 850         audio_level_callback(loudness_s, to_db(peak), bus_levels,
 851                 loudness_i, loudness_range_low, loudness_range_high,
 852                 to_db(final_makeup_gain),
 853                 correlation.get_correlation());
 854 }
 855
 856 map<DeviceSpec, DeviceInfo> AudioMixer::get_devices()
 857 {
 858         lock_guard<timed_mutex> lock(audio_mutex);
 859
 860         map<DeviceSpec, DeviceInfo> devices;
 861         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 862                 const DeviceSpec spec{ InputSourceType::CAPTURE_CARD, card_index };
 863                 const AudioDevice *device = &video_cards[card_index];
 864                 DeviceInfo info;
 865                 info.display_name = device->display_name;
 866                 info.num_channels = 8;
 867                 devices.insert(make_pair(spec, info));
 868         }
 869         vector<ALSAPool::Device> available_alsa_devices = alsa_pool.get_devices();
 870         for (unsigned card_index = 0; card_index < available_alsa_devices.size(); ++card_index) {
 871                 const DeviceSpec spec{ InputSourceType::ALSA_INPUT, card_index };
 872                 const ALSAPool::Device &device = available_alsa_devices[card_index];
 873                 DeviceInfo info;
 874                 info.display_name = device.display_name();
 875                 info.num_channels = device.num_channels;
 876                 info.alsa_name = device.name;
 877                 info.alsa_info = device.info;
 878                 info.alsa_address = device.address;
 879                 devices.insert(make_pair(spec, info));
 880         }
 881         return devices;
 882 }
 883
 884 void AudioMixer::set_display_name(DeviceSpec device_spec, const string &name)
 885 {
 886         AudioDevice *device = find_audio_device(device_spec);
 887
 888         lock_guard<timed_mutex> lock(audio_mutex);
 889         device->display_name = name;
 890 }
 891
 892 void AudioMixer::serialize_device(DeviceSpec device_spec, DeviceSpecProto *device_spec_proto)
 893 {
 894         lock_guard<timed_mutex> lock(audio_mutex);
 895         switch (device_spec.type) {
 896                 case InputSourceType::SILENCE:
 897                         device_spec_proto->set_type(DeviceSpecProto::SILENCE);
 898                         break;
 899                 case InputSourceType::CAPTURE_CARD:
 900                         device_spec_proto->set_type(DeviceSpecProto::CAPTURE_CARD);
 901                         device_spec_proto->set_index(device_spec.index);
 902                         device_spec_proto->set_display_name(video_cards[device_spec.index].display_name);
 903                         break;
 904                 case InputSourceType::ALSA_INPUT:
 905                         alsa_pool.serialize_device(device_spec.index, device_spec_proto);
 906                         break;
 907         }
 908 }
 909
 910 void AudioMixer::set_simple_input(unsigned card_index)
 911 {
 912         InputMapping new_input_mapping;
 913         InputMapping::Bus input;
 914         input.name = "Main";
 915         input.device.type = InputSourceType::CAPTURE_CARD;
 916         input.device.index = card_index;
 917         input.source_channel[0] = 0;
 918         input.source_channel[1] = 1;
 919
 920         new_input_mapping.buses.push_back(input);
 921
 922         lock_guard<timed_mutex> lock(audio_mutex);
 923         current_mapping_mode = MappingMode::SIMPLE;
 924         set_input_mapping_lock_held(new_input_mapping);
 925         fader_volume_db[0] = 0.0f;
 926 }
 927
 928 unsigned AudioMixer::get_simple_input() const
 929 {
 930         lock_guard<timed_mutex> lock(audio_mutex);
 931         if (input_mapping.buses.size() == 1 &&
 932             input_mapping.buses[0].device.type == InputSourceType::CAPTURE_CARD &&
 933             input_mapping.buses[0].source_channel[0] == 0 &&
 934             input_mapping.buses[0].source_channel[1] == 1) {
 935                 return input_mapping.buses[0].device.index;
 936         } else {
 937                 return numeric_limits<unsigned>::max();
 938         }
 939 }
 940
 941 void AudioMixer::set_input_mapping(const InputMapping &new_input_mapping)
 942 {
 943         lock_guard<timed_mutex> lock(audio_mutex);
 944         set_input_mapping_lock_held(new_input_mapping);
 945         current_mapping_mode = MappingMode::MULTICHANNEL;
 946 }
 947
 948 AudioMixer::MappingMode AudioMixer::get_mapping_mode() const
 949 {
 950         lock_guard<timed_mutex> lock(audio_mutex);
 951         return current_mapping_mode;
 952 }
 953
 954 void AudioMixer::set_input_mapping_lock_held(const InputMapping &new_input_mapping)
 955 {
 956         map<DeviceSpec, set<unsigned>> interesting_channels;
 957         for (const InputMapping::Bus &bus : new_input_mapping.buses) {
 958                 if (bus.device.type == InputSourceType::CAPTURE_CARD ||
 959                     bus.device.type == InputSourceType::ALSA_INPUT) {
 960                         for (unsigned channel = 0; channel < 2; ++channel) {
 961                                 if (bus.source_channel[channel] != -1) {
 962                                         interesting_channels[bus.device].insert(bus.source_channel[channel]);
 963                                 }
 964                         }
 965                 }
 966         }
 967
 968         // Reset resamplers for all cards that don't have the exact same state as before.
 969         for (unsigned card_index = 0; card_index < MAX_VIDEO_CARDS; ++card_index) {
 970                 const DeviceSpec device_spec{InputSourceType::CAPTURE_CARD, card_index};
 971                 AudioDevice *device = find_audio_device(device_spec);
 972                 if (device->interesting_channels != interesting_channels[device_spec]) {
 973                         device->interesting_channels = interesting_channels[device_spec];
 974                         reset_resampler_mutex_held(device_spec);
 975                 }
 976         }
 977         for (unsigned card_index = 0; card_index < MAX_ALSA_CARDS; ++card_index) {
 978                 const DeviceSpec device_spec{InputSourceType::ALSA_INPUT, card_index};
 979                 AudioDevice *device = find_audio_device(device_spec);
 980                 if (interesting_channels[device_spec].empty()) {
 981                         alsa_pool.release_device(card_index);
 982                 } else {
 983                         alsa_pool.hold_device(card_index);
 984                 }
 985                 if (device->interesting_channels != interesting_channels[device_spec]) {
 986                         device->interesting_channels = interesting_channels[device_spec];
 987                         alsa_pool.reset_device(device_spec.index);
 988                         reset_resampler_mutex_held(device_spec);
 989                 }
 990         }
 991
 992         input_mapping = new_input_mapping;
 993 }
 994
 995 InputMapping AudioMixer::get_input_mapping() const
 996 {
 997         lock_guard<timed_mutex> lock(audio_mutex);
 998         return input_mapping;
 999 }
1000
1001 unsigned AudioMixer::num_buses() const
1002 {
1003         lock_guard<timed_mutex> lock(audio_mutex);
1004         return input_mapping.buses.size();
1005 }
1006
1007 void AudioMixer::reset_peak(unsigned bus_index)
1008 {
1009         lock_guard<timed_mutex> lock(audio_mutex);
1010         for (unsigned channel = 0; channel < 2; ++channel) {
1011                 PeakHistory &history = peak_history[bus_index][channel];
1012                 history.current_level = 0.0f;
1013                 history.historic_peak = 0.0f;
1014                 history.current_peak = 0.0f;
1015                 history.last_peak = 0.0f;
1016                 history.age_seconds = 0.0f;
1017         }
1018 }
1019
1020 AudioMixer *global_audio_mixer = nullptr;