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         global_audio_mixer = this;
 174
 175         for (unsigned bus_index = 0; bus_index < MAX_BUSES; ++bus_index) {
 176                 locut[bus_index].init(FILTER_HPF, 2);
 177                 eq[bus_index][EQ_BAND_BASS].init(FILTER_LOW_SHELF, 1);
 178                 // Note: EQ_BAND_MID isn't used (see comments in apply_eq()).
 179                 eq[bus_index][EQ_BAND_TREBLE].init(FILTER_HIGH_SHELF, 1);
 180                 compressor[bus_index].reset(new StereoCompressor(OUTPUT_FREQUENCY));
 181                 level_compressor[bus_index].reset(new StereoCompressor(OUTPUT_FREQUENCY));
 182
 183                 set_bus_settings(bus_index, get_default_bus_settings());
 184         }
 185         set_limiter_enabled(global_flags.limiter_enabled);
 186         set_final_makeup_gain_auto(global_flags.final_makeup_gain_auto);
 187         alsa_pool.init();
 188
 189         if (!global_flags.input_mapping_filename.empty()) {
 190                 current_mapping_mode = MappingMode::MULTICHANNEL;
 191                 InputMapping new_input_mapping;
 192                 if (!load_input_mapping_from_file(get_devices(),
 193                                                   global_flags.input_mapping_filename,
 194                                                   &new_input_mapping)) {
 195                         fprintf(stderr, "Failed to load input mapping from '%s', exiting.\n",
 196                                 global_flags.input_mapping_filename.c_str());
 197                         exit(1);
 198                 }
 199                 set_input_mapping(new_input_mapping);
 200         } else {
 201                 set_simple_input(/*card_index=*/0);
 202                 if (global_flags.multichannel_mapping_mode) {
 203                         current_mapping_mode = MappingMode::MULTICHANNEL;
 204                 }
 205         }
 206
 207         r128.init(2, OUTPUT_FREQUENCY);
 208         r128.integr_start();
 209
 210         // hlen=16 is pretty low quality, but we use quite a bit of CPU otherwise,
 211         // and there's a limit to how important the peak meter is.
 212         peak_resampler.setup(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY * 4, /*num_channels=*/2, /*hlen=*/16, /*frel=*/1.0);
 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(device_spec.index, device->capture_frequency, OUTPUT_FREQUENCY, device->interesting_channels.size()));
 231         }
 232         device->next_local_pts = 0;
 233 }
 234
 235 bool AudioMixer::add_audio(DeviceSpec device_spec, const uint8_t *data, unsigned num_samples, AudioFormat audio_format, int64_t frame_length)
 236 {
 237         AudioDevice *device = find_audio_device(device_spec);
 238
 239         unique_lock<timed_mutex> lock(audio_mutex, defer_lock);
 240         if (!lock.try_lock_for(chrono::milliseconds(10))) {
 241                 return false;
 242         }
 243         if (device->resampling_queue == nullptr) {
 244                 // No buses use this device; throw it away.
 245                 return true;
 246         }
 247
 248         unsigned num_channels = device->interesting_channels.size();
 249         assert(num_channels > 0);
 250
 251         // Convert the audio to fp32.
 252         unique_ptr<float[]> audio(new float[num_samples * num_channels]);
 253         unsigned channel_index = 0;
 254         for (auto channel_it = device->interesting_channels.cbegin(); channel_it != device->interesting_channels.end(); ++channel_it, ++channel_index) {
 255                 switch (audio_format.bits_per_sample) {
 256                 case 0:
 257                         assert(num_samples == 0);
 258                         break;
 259                 case 16:
 260                         convert_fixed16_to_fp32(audio.get(), channel_index, num_channels, data, *channel_it, audio_format.num_channels, num_samples);
 261                         break;
 262                 case 24:
 263                         convert_fixed24_to_fp32(audio.get(), channel_index, num_channels, data, *channel_it, audio_format.num_channels, num_samples);
 264                         break;
 265                 case 32:
 266                         convert_fixed32_to_fp32(audio.get(), channel_index, num_channels, data, *channel_it, audio_format.num_channels, num_samples);
 267                         break;
 268                 default:
 269                         fprintf(stderr, "Cannot handle audio with %u bits per sample\n", audio_format.bits_per_sample);
 270                         assert(false);
 271                 }
 272         }
 273
 274         // Now add it.
 275         int64_t local_pts = device->next_local_pts;
 276         device->resampling_queue->add_input_samples(local_pts / double(TIMEBASE), audio.get(), num_samples);
 277         device->next_local_pts = local_pts + frame_length;
 278         return true;
 279 }
 280
 281 bool AudioMixer::add_silence(DeviceSpec device_spec, unsigned samples_per_frame, unsigned num_frames, int64_t frame_length)
 282 {
 283         AudioDevice *device = find_audio_device(device_spec);
 284
 285         unique_lock<timed_mutex> lock(audio_mutex, defer_lock);
 286         if (!lock.try_lock_for(chrono::milliseconds(10))) {
 287                 return false;
 288         }
 289         if (device->resampling_queue == nullptr) {
 290                 // No buses use this device; throw it away.
 291                 return true;
 292         }
 293
 294         unsigned num_channels = device->interesting_channels.size();
 295         assert(num_channels > 0);
 296
 297         vector<float> silence(samples_per_frame * num_channels, 0.0f);
 298         for (unsigned i = 0; i < num_frames; ++i) {
 299                 device->resampling_queue->add_input_samples(device->next_local_pts / double(TIMEBASE), silence.data(), samples_per_frame);
 300                 // Note that if the format changed in the meantime, we have
 301                 // no way of detecting that; we just have to assume the frame length
 302                 // is always the same.
 303                 device->next_local_pts += frame_length;
 304         }
 305         return true;
 306 }
 307
 308 bool AudioMixer::silence_card(DeviceSpec device_spec, bool silence)
 309 {
 310         AudioDevice *device = find_audio_device(device_spec);
 311
 312         unique_lock<timed_mutex> lock(audio_mutex, defer_lock);
 313         if (!lock.try_lock_for(chrono::milliseconds(10))) {
 314                 return false;
 315         }
 316
 317         if (device->silenced && !silence) {
 318                 reset_resampler_mutex_held(device_spec);
 319         }
 320         device->silenced = silence;
 321         return true;
 322 }
 323
 324 AudioMixer::BusSettings AudioMixer::get_default_bus_settings()
 325 {
 326         BusSettings settings;
 327         settings.fader_volume_db = 0.0f;
 328         settings.muted = false;
 329         settings.locut_enabled = global_flags.locut_enabled;
 330         for (unsigned band_index = 0; band_index < NUM_EQ_BANDS; ++band_index) {
 331                 settings.eq_level_db[band_index] = 0.0f;
 332         }
 333         settings.gain_staging_db = global_flags.initial_gain_staging_db;
 334         settings.level_compressor_enabled = global_flags.gain_staging_auto;
 335         settings.compressor_threshold_dbfs = ref_level_dbfs - 12.0f;  // -12 dB.
 336         settings.compressor_enabled = global_flags.compressor_enabled;
 337         return settings;
 338 }
 339
 340 AudioMixer::BusSettings AudioMixer::get_bus_settings(unsigned bus_index) const
 341 {
 342         lock_guard<timed_mutex> lock(audio_mutex);
 343         BusSettings settings;
 344         settings.fader_volume_db = fader_volume_db[bus_index];
 345         settings.muted = mute[bus_index];
 346         settings.locut_enabled = locut_enabled[bus_index];
 347         for (unsigned band_index = 0; band_index < NUM_EQ_BANDS; ++band_index) {
 348                 settings.eq_level_db[band_index] = eq_level_db[bus_index][band_index];
 349         }
 350         settings.gain_staging_db = gain_staging_db[bus_index];
 351         settings.level_compressor_enabled = level_compressor_enabled[bus_index];
 352         settings.compressor_threshold_dbfs = compressor_threshold_dbfs[bus_index];
 353         settings.compressor_enabled = compressor_enabled[bus_index];
 354         return settings;
 355 }
 356
 357 void AudioMixer::set_bus_settings(unsigned bus_index, const AudioMixer::BusSettings &settings)
 358 {
 359         lock_guard<timed_mutex> lock(audio_mutex);
 360         fader_volume_db[bus_index] = settings.fader_volume_db;
 361         mute[bus_index] = settings.muted;
 362         locut_enabled[bus_index] = settings.locut_enabled;
 363         for (unsigned band_index = 0; band_index < NUM_EQ_BANDS; ++band_index) {
 364                 eq_level_db[bus_index][band_index] = settings.eq_level_db[band_index];
 365         }
 366         gain_staging_db[bus_index] = settings.gain_staging_db;
 367         last_gain_staging_db[bus_index] = gain_staging_db[bus_index];
 368         level_compressor_enabled[bus_index] = settings.level_compressor_enabled;
 369         compressor_threshold_dbfs[bus_index] = settings.compressor_threshold_dbfs;
 370         compressor_enabled[bus_index] = settings.compressor_enabled;
 371 }
 372
 373 AudioMixer::AudioDevice *AudioMixer::find_audio_device(DeviceSpec device)
 374 {
 375         switch (device.type) {
 376         case InputSourceType::CAPTURE_CARD:
 377                 return &video_cards[device.index];
 378         case InputSourceType::ALSA_INPUT:
 379                 return &alsa_inputs[device.index];
 380         case InputSourceType::SILENCE:
 381         default:
 382                 assert(false);
 383         }
 384         return nullptr;
 385 }
 386
 387 // Get a pointer to the given channel from the given device.
 388 // The channel must be picked out earlier and resampled.
 389 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)
 390 {
 391         static float zero = 0.0f;
 392         if (source_channel == -1 || device_spec.type == InputSourceType::SILENCE) {
 393                 *srcptr = &zero;
 394                 *stride = 0;
 395                 return;
 396         }
 397         AudioDevice *device = find_audio_device(device_spec);
 398         assert(device->interesting_channels.count(source_channel) != 0);
 399         unsigned channel_index = 0;
 400         for (int channel : device->interesting_channels) {
 401                 if (channel == source_channel) break;
 402                 ++channel_index;
 403         }
 404         assert(channel_index < device->interesting_channels.size());
 405         const auto it = samples_card.find(device_spec);
 406         assert(it != samples_card.end());
 407         *srcptr = &(it->second)[channel_index];
 408         *stride = device->interesting_channels.size();
 409 }
 410
 411 // TODO: Can be SSSE3-optimized if need be.
 412 void AudioMixer::fill_audio_bus(const map<DeviceSpec, vector<float>> &samples_card, const InputMapping::Bus &bus, unsigned num_samples, float *output)
 413 {
 414         if (bus.device.type == InputSourceType::SILENCE) {
 415                 memset(output, 0, num_samples * 2 * sizeof(*output));
 416         } else {
 417                 assert(bus.device.type == InputSourceType::CAPTURE_CARD ||
 418                        bus.device.type == InputSourceType::ALSA_INPUT);
 419                 const float *lsrc, *rsrc;
 420                 unsigned lstride, rstride;
 421                 float *dptr = output;
 422                 find_sample_src_from_device(samples_card, bus.device, bus.source_channel[0], &lsrc, &lstride);
 423                 find_sample_src_from_device(samples_card, bus.device, bus.source_channel[1], &rsrc, &rstride);
 424                 for (unsigned i = 0; i < num_samples; ++i) {
 425                         *dptr++ = *lsrc;
 426                         *dptr++ = *rsrc;
 427                         lsrc += lstride;
 428                         rsrc += rstride;
 429                 }
 430         }
 431 }
 432
 433 vector<DeviceSpec> AudioMixer::get_active_devices() const
 434 {
 435         vector<DeviceSpec> ret;
 436         for (unsigned card_index = 0; card_index < MAX_VIDEO_CARDS; ++card_index) {
 437                 const DeviceSpec device_spec{InputSourceType::CAPTURE_CARD, card_index};
 438                 if (!find_audio_device(device_spec)->interesting_channels.empty()) {
 439                         ret.push_back(device_spec);
 440                 }
 441         }
 442         for (unsigned card_index = 0; card_index < MAX_ALSA_CARDS; ++card_index) {
 443                 const DeviceSpec device_spec{InputSourceType::ALSA_INPUT, card_index};
 444                 if (!find_audio_device(device_spec)->interesting_channels.empty()) {
 445                         ret.push_back(device_spec);
 446                 }
 447         }
 448         return ret;
 449 }
 450
 451 namespace {
 452
 453 void apply_gain(float db, float last_db, vector<float> *samples)
 454 {
 455         if (fabs(db - last_db) < 1e-3) {
 456                 // Constant over this frame.
 457                 const float gain = from_db(db);
 458                 for (size_t i = 0; i < samples->size(); ++i) {
 459                         (*samples)[i] *= gain;
 460                 }
 461         } else {
 462                 // We need to do a fade.
 463                 unsigned num_samples = samples->size() / 2;
 464                 float gain = from_db(last_db);
 465                 const float gain_inc = pow(from_db(db - last_db), 1.0 / num_samples);
 466                 for (size_t i = 0; i < num_samples; ++i) {
 467                         (*samples)[i * 2 + 0] *= gain;
 468                         (*samples)[i * 2 + 1] *= gain;
 469                         gain *= gain_inc;
 470                 }
 471         }
 472 }
 473
 474 }  // namespace
 475
 476 vector<float> AudioMixer::get_output(double pts, unsigned num_samples, ResamplingQueue::RateAdjustmentPolicy rate_adjustment_policy)
 477 {
 478         map<DeviceSpec, vector<float>> samples_card;
 479         vector<float> samples_bus;
 480
 481         lock_guard<timed_mutex> lock(audio_mutex);
 482
 483         // Pick out all the interesting channels from all the cards.
 484         for (const DeviceSpec &device_spec : get_active_devices()) {
 485                 AudioDevice *device = find_audio_device(device_spec);
 486                 samples_card[device_spec].resize(num_samples * device->interesting_channels.size());
 487                 if (device->silenced) {
 488                         memset(&samples_card[device_spec][0], 0, samples_card[device_spec].size() * sizeof(float));
 489                 } else {
 490                         device->resampling_queue->get_output_samples(
 491                                 pts,
 492                                 &samples_card[device_spec][0],
 493                                 num_samples,
 494                                 rate_adjustment_policy);
 495                 }
 496         }
 497
 498         vector<float> samples_out, left, right;
 499         samples_out.resize(num_samples * 2);
 500         samples_bus.resize(num_samples * 2);
 501         for (unsigned bus_index = 0; bus_index < input_mapping.buses.size(); ++bus_index) {
 502                 fill_audio_bus(samples_card, input_mapping.buses[bus_index], num_samples, &samples_bus[0]);
 503                 apply_eq(bus_index, &samples_bus);
 504
 505                 {
 506                         lock_guard<mutex> lock(compressor_mutex);
 507
 508                         // Apply a level compressor to get the general level right.
 509                         // Basically, if it's over about -40 dBFS, we squeeze it down to that level
 510                         // (or more precisely, near it, since we don't use infinite ratio),
 511                         // then apply a makeup gain to get it to -14 dBFS. -14 dBFS is, of course,
 512                         // entirely arbitrary, but from practical tests with speech, it seems to
 513                         // put ut around -23 LUFS, so it's a reasonable starting point for later use.
 514                         if (level_compressor_enabled[bus_index]) {
 515                                 float threshold = 0.01f;   // -40 dBFS.
 516                                 float ratio = 20.0f;
 517                                 float attack_time = 0.5f;
 518                                 float release_time = 20.0f;
 519                                 float makeup_gain = from_db(ref_level_dbfs - (-40.0f));  // +26 dB.
 520                                 level_compressor[bus_index]->process(samples_bus.data(), samples_bus.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 521                                 gain_staging_db[bus_index] = to_db(level_compressor[bus_index]->get_attenuation() * makeup_gain);
 522                         } else {
 523                                 // Just apply the gain we already had.
 524                                 float db = gain_staging_db[bus_index];
 525                                 float last_db = last_gain_staging_db[bus_index];
 526                                 apply_gain(db, last_db, &samples_bus);
 527                         }
 528                         last_gain_staging_db[bus_index] = gain_staging_db[bus_index];
 529
 530 #if 0
 531                         printf("level=%f (%+5.2f dBFS) attenuation=%f (%+5.2f dB) end_result=%+5.2f dB\n",
 532                                 level_compressor.get_level(), to_db(level_compressor.get_level()),
 533                                 level_compressor.get_attenuation(), to_db(level_compressor.get_attenuation()),
 534                                 to_db(level_compressor.get_level() * level_compressor.get_attenuation() * makeup_gain));
 535 #endif
 536
 537                         // The real compressor.
 538                         if (compressor_enabled[bus_index]) {
 539                                 float threshold = from_db(compressor_threshold_dbfs[bus_index]);
 540                                 float ratio = 20.0f;
 541                                 float attack_time = 0.005f;
 542                                 float release_time = 0.040f;
 543                                 float makeup_gain = 2.0f;  // +6 dB.
 544                                 compressor[bus_index]->process(samples_bus.data(), samples_bus.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 545                 //              compressor_att = compressor.get_attenuation();
 546                         }
 547                 }
 548
 549                 add_bus_to_master(bus_index, samples_bus, &samples_out);
 550                 deinterleave_samples(samples_bus, &left, &right);
 551                 measure_bus_levels(bus_index, left, right);
 552         }
 553
 554         {
 555                 lock_guard<mutex> lock(compressor_mutex);
 556
 557                 // Finally a limiter at -4 dB (so, -10 dBFS) to take out the worst peaks only.
 558                 // Note that since ratio is not infinite, we could go slightly higher than this.
 559                 if (limiter_enabled) {
 560                         float threshold = from_db(limiter_threshold_dbfs);
 561                         float ratio = 30.0f;
 562                         float attack_time = 0.0f;  // Instant.
 563                         float release_time = 0.020f;
 564                         float makeup_gain = 1.0f;  // 0 dB.
 565                         limiter.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 566         //              limiter_att = limiter.get_attenuation();
 567                 }
 568
 569         //      printf("limiter=%+5.1f  compressor=%+5.1f\n", to_db(limiter_att), to_db(compressor_att));
 570         }
 571
 572         // At this point, we are most likely close to +0 LU (at least if the
 573         // faders sum to 0 dB and the compressors are on), but all of our
 574         // measurements have been on raw sample values, not R128 values.
 575         // So we have a final makeup gain to get us to +0 LU; the gain
 576         // adjustments required should be relatively small, and also, the
 577         // offset shouldn't change much (only if the type of audio changes
 578         // significantly). Thus, we shoot for updating this value basically
 579         // “whenever we process buffers”, since the R128 calculation isn't exactly
 580         // something we get out per-sample.
 581         //
 582         // Note that there's a feedback loop here, so we choose a very slow filter
 583         // (half-time of 30 seconds).
 584         double target_loudness_factor, alpha;
 585         double loudness_lu = r128.loudness_M() - ref_level_lufs;
 586         double current_makeup_lu = to_db(final_makeup_gain);
 587         target_loudness_factor = final_makeup_gain * from_db(-loudness_lu);
 588
 589         // If we're outside +/- 5 LU uncorrected, we don't count it as
 590         // a normal signal (probably silence) and don't change the
 591         // correction factor; just apply what we already have.
 592         if (fabs(loudness_lu - current_makeup_lu) >= 5.0 || !final_makeup_gain_auto) {
 593                 alpha = 0.0;
 594         } else {
 595                 // Formula adapted from
 596                 // https://en.wikipedia.org/wiki/Low-pass_filter#Simple_infinite_impulse_response_filter.
 597                 const double half_time_s = 30.0;
 598                 const double fc_mul_2pi_delta_t = 1.0 / (half_time_s * OUTPUT_FREQUENCY);
 599                 alpha = fc_mul_2pi_delta_t / (fc_mul_2pi_delta_t + 1.0);
 600         }
 601
 602         {
 603                 lock_guard<mutex> lock(compressor_mutex);
 604                 double m = final_makeup_gain;
 605                 for (size_t i = 0; i < samples_out.size(); i += 2) {
 606                         samples_out[i + 0] *= m;
 607                         samples_out[i + 1] *= m;
 608                         m += (target_loudness_factor - m) * alpha;
 609                 }
 610                 final_makeup_gain = m;
 611         }
 612
 613         update_meters(samples_out);
 614
 615         return samples_out;
 616 }
 617
 618 namespace {
 619
 620 void apply_filter_fade(StereoFilter *filter, float *data, unsigned num_samples, float cutoff_hz, float db, float last_db)
 621 {
 622         // A granularity of 32 samples is an okay tradeoff between speed and
 623         // smoothness; recalculating the filters is pretty expensive, so it's
 624         // good that we don't do this all the time.
 625         static constexpr unsigned filter_granularity_samples = 32;
 626
 627         const float cutoff_linear = cutoff_hz * 2.0 * M_PI / OUTPUT_FREQUENCY;
 628         if (fabs(db - last_db) < 1e-3) {
 629                 // Constant over this frame.
 630                 if (fabs(db) > 0.01f) {
 631                         filter->render(data, num_samples, cutoff_linear, 0.5f, db / 40.0f);
 632                 }
 633         } else {
 634                 // We need to do a fade. (Rounding up avoids division by zero.)
 635                 unsigned num_blocks = (num_samples + filter_granularity_samples - 1) / filter_granularity_samples;
 636                 const float inc_db_norm = (db - last_db) / 40.0f / num_blocks;
 637                 float db_norm = db / 40.0f;
 638                 for (size_t i = 0; i < num_samples; i += filter_granularity_samples) {
 639                         size_t samples_this_block = std::min<size_t>(num_samples - i, filter_granularity_samples);
 640                         filter->render(data + i * 2, samples_this_block, cutoff_linear, 0.5f, db_norm);
 641                         db_norm += inc_db_norm;
 642                 }
 643         }
 644 }
 645
 646 }  // namespace
 647
 648 void AudioMixer::apply_eq(unsigned bus_index, vector<float> *samples_bus)
 649 {
 650         constexpr float bass_freq_hz = 200.0f;
 651         constexpr float treble_freq_hz = 4700.0f;
 652
 653         // Cut away everything under 120 Hz (or whatever the cutoff is);
 654         // we don't need it for voice, and it will reduce headroom
 655         // and confuse the compressor. (In particular, any hums at 50 or 60 Hz
 656         // should be dampened.)
 657         if (locut_enabled[bus_index]) {
 658                 locut[bus_index].render(samples_bus->data(), samples_bus->size() / 2, locut_cutoff_hz * 2.0 * M_PI / OUTPUT_FREQUENCY, 0.5f);
 659         }
 660
 661         // Apply the rest of the EQ. Since we only have a simple three-band EQ,
 662         // we can implement it with two shelf filters. We use a simple gain to
 663         // set the mid-level filter, and then offset the low and high bands
 664         // from that if we need to. (We could perhaps have folded the gain into
 665         // the next part, but it's so cheap that the trouble isn't worth it.)
 666         //
 667         // If any part of the EQ has changed appreciably since last frame,
 668         // we fade smoothly during the course of this frame.
 669         const float bass_db = eq_level_db[bus_index][EQ_BAND_BASS];
 670         const float mid_db = eq_level_db[bus_index][EQ_BAND_MID];
 671         const float treble_db = eq_level_db[bus_index][EQ_BAND_TREBLE];
 672
 673         const float last_bass_db = last_eq_level_db[bus_index][EQ_BAND_BASS];
 674         const float last_mid_db = last_eq_level_db[bus_index][EQ_BAND_MID];
 675         const float last_treble_db = last_eq_level_db[bus_index][EQ_BAND_TREBLE];
 676
 677         assert(samples_bus->size() % 2 == 0);
 678         const unsigned num_samples = samples_bus->size() / 2;
 679
 680         apply_gain(mid_db, last_mid_db, samples_bus);
 681
 682         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);
 683         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);
 684
 685         last_eq_level_db[bus_index][EQ_BAND_BASS] = bass_db;
 686         last_eq_level_db[bus_index][EQ_BAND_MID] = mid_db;
 687         last_eq_level_db[bus_index][EQ_BAND_TREBLE] = treble_db;
 688 }
 689
 690 void AudioMixer::add_bus_to_master(unsigned bus_index, const vector<float> &samples_bus, vector<float> *samples_out)
 691 {
 692         assert(samples_bus.size() == samples_out->size());
 693         assert(samples_bus.size() % 2 == 0);
 694         unsigned num_samples = samples_bus.size() / 2;
 695         const float new_volume_db = mute[bus_index] ? -90.0f : fader_volume_db[bus_index].load();
 696         if (fabs(new_volume_db - last_fader_volume_db[bus_index]) > 1e-3) {
 697                 // The volume has changed; do a fade over the course of this frame.
 698                 // (We might have some numerical issues here, but it seems to sound OK.)
 699                 // For the purpose of fading here, the silence floor is set to -90 dB
 700                 // (the fader only goes to -84).
 701                 float old_volume = from_db(max<float>(last_fader_volume_db[bus_index], -90.0f));
 702                 float volume = from_db(max<float>(new_volume_db, -90.0f));
 703
 704                 float volume_inc = pow(volume / old_volume, 1.0 / num_samples);
 705                 volume = old_volume;
 706                 if (bus_index == 0) {
 707                         for (unsigned i = 0; i < num_samples; ++i) {
 708                                 (*samples_out)[i * 2 + 0] = samples_bus[i * 2 + 0] * volume;
 709                                 (*samples_out)[i * 2 + 1] = samples_bus[i * 2 + 1] * volume;
 710                                 volume *= volume_inc;
 711                         }
 712                 } else {
 713                         for (unsigned i = 0; i < num_samples; ++i) {
 714                                 (*samples_out)[i * 2 + 0] += samples_bus[i * 2 + 0] * volume;
 715                                 (*samples_out)[i * 2 + 1] += samples_bus[i * 2 + 1] * volume;
 716                                 volume *= volume_inc;
 717                         }
 718                 }
 719         } else if (new_volume_db > -90.0f) {
 720                 float volume = from_db(new_volume_db);
 721                 if (bus_index == 0) {
 722                         for (unsigned i = 0; i < num_samples; ++i) {
 723                                 (*samples_out)[i * 2 + 0] = samples_bus[i * 2 + 0] * volume;
 724                                 (*samples_out)[i * 2 + 1] = samples_bus[i * 2 + 1] * volume;
 725                         }
 726                 } else {
 727                         for (unsigned i = 0; i < num_samples; ++i) {
 728                                 (*samples_out)[i * 2 + 0] += samples_bus[i * 2 + 0] * volume;
 729                                 (*samples_out)[i * 2 + 1] += samples_bus[i * 2 + 1] * volume;
 730                         }
 731                 }
 732         }
 733
 734         last_fader_volume_db[bus_index] = new_volume_db;
 735 }
 736
 737 void AudioMixer::measure_bus_levels(unsigned bus_index, const vector<float> &left, const vector<float> &right)
 738 {
 739         assert(left.size() == right.size());
 740         const float volume = mute[bus_index] ? 0.0f : from_db(fader_volume_db[bus_index]);
 741         const float peak_levels[2] = {
 742                 find_peak(left.data(), left.size()) * volume,
 743                 find_peak(right.data(), right.size()) * volume
 744         };
 745         for (unsigned channel = 0; channel < 2; ++channel) {
 746                 // Compute the current value, including hold and falloff.
 747                 // The constants are borrowed from zita-mu1 by Fons Adriaensen.
 748                 static constexpr float hold_sec = 0.5f;
 749                 static constexpr float falloff_db_sec = 15.0f;  // dB/sec falloff after hold.
 750                 float current_peak;
 751                 PeakHistory &history = peak_history[bus_index][channel];
 752                 history.historic_peak = max(history.historic_peak, peak_levels[channel]);
 753                 if (history.age_seconds < hold_sec) {
 754                         current_peak = history.last_peak;
 755                 } else {
 756                         current_peak = history.last_peak * from_db(-falloff_db_sec * (history.age_seconds - hold_sec));
 757                 }
 758
 759                 // See if we have a new peak to replace the old (possibly falling) one.
 760                 if (peak_levels[channel] > current_peak) {
 761                         history.last_peak = peak_levels[channel];
 762                         history.age_seconds = 0.0f;  // Not 100% correct, but more than good enough given our frame sizes.
 763                         current_peak = peak_levels[channel];
 764                 } else {
 765                         history.age_seconds += float(left.size()) / OUTPUT_FREQUENCY;
 766                 }
 767                 history.current_level = peak_levels[channel];
 768                 history.current_peak = current_peak;
 769         }
 770 }
 771
 772 void AudioMixer::update_meters(const vector<float> &samples)
 773 {
 774         // Upsample 4x to find interpolated peak.
 775         peak_resampler.inp_data = const_cast<float *>(samples.data());
 776         peak_resampler.inp_count = samples.size() / 2;
 777
 778         vector<float> interpolated_samples;
 779         interpolated_samples.resize(samples.size());
 780         {
 781                 lock_guard<mutex> lock(audio_measure_mutex);
 782
 783                 while (peak_resampler.inp_count > 0) {  // About four iterations.
 784                         peak_resampler.out_data = &interpolated_samples[0];
 785                         peak_resampler.out_count = interpolated_samples.size() / 2;
 786                         peak_resampler.process();
 787                         size_t out_stereo_samples = interpolated_samples.size() / 2 - peak_resampler.out_count;
 788                         peak = max<float>(peak, find_peak(interpolated_samples.data(), out_stereo_samples * 2));
 789                         peak_resampler.out_data = nullptr;
 790                 }
 791         }
 792
 793         // Find R128 levels and L/R correlation.
 794         vector<float> left, right;
 795         deinterleave_samples(samples, &left, &right);
 796         float *ptrs[] = { left.data(), right.data() };
 797         {
 798                 lock_guard<mutex> lock(audio_measure_mutex);
 799                 r128.process(left.size(), ptrs);
 800                 correlation.process_samples(samples);
 801         }
 802
 803         send_audio_level_callback();
 804 }
 805
 806 void AudioMixer::reset_meters()
 807 {
 808         lock_guard<mutex> lock(audio_measure_mutex);
 809         peak_resampler.reset();
 810         peak = 0.0f;
 811         r128.reset();
 812         r128.integr_start();
 813         correlation.reset();
 814 }
 815
 816 void AudioMixer::send_audio_level_callback()
 817 {
 818         if (audio_level_callback == nullptr) {
 819                 return;
 820         }
 821
 822         lock_guard<mutex> lock(audio_measure_mutex);
 823         double loudness_s = r128.loudness_S();
 824         double loudness_i = r128.integrated();
 825         double loudness_range_low = r128.range_min();
 826         double loudness_range_high = r128.range_max();
 827
 828         vector<BusLevel> bus_levels;
 829         bus_levels.resize(input_mapping.buses.size());
 830         {
 831                 lock_guard<mutex> lock(compressor_mutex);
 832                 for (unsigned bus_index = 0; bus_index < bus_levels.size(); ++bus_index) {
 833                         bus_levels[bus_index].current_level_dbfs[0] = to_db(peak_history[bus_index][0].current_level);
 834                         bus_levels[bus_index].current_level_dbfs[1] = to_db(peak_history[bus_index][1].current_level);
 835                         bus_levels[bus_index].peak_level_dbfs[0] = to_db(peak_history[bus_index][0].current_peak);
 836                         bus_levels[bus_index].peak_level_dbfs[1] = to_db(peak_history[bus_index][1].current_peak);
 837                         bus_levels[bus_index].historic_peak_dbfs = to_db(
 838                                 max(peak_history[bus_index][0].historic_peak,
 839                                     peak_history[bus_index][1].historic_peak));
 840                         bus_levels[bus_index].gain_staging_db = gain_staging_db[bus_index];
 841                         if (compressor_enabled[bus_index]) {
 842                                 bus_levels[bus_index].compressor_attenuation_db = -to_db(compressor[bus_index]->get_attenuation());
 843                         } else {
 844                                 bus_levels[bus_index].compressor_attenuation_db = 0.0;
 845                         }
 846                 }
 847         }
 848
 849         audio_level_callback(loudness_s, to_db(peak), bus_levels,
 850                 loudness_i, loudness_range_low, loudness_range_high,
 851                 to_db(final_makeup_gain),
 852                 correlation.get_correlation());
 853 }
 854
 855 map<DeviceSpec, DeviceInfo> AudioMixer::get_devices()
 856 {
 857         lock_guard<timed_mutex> lock(audio_mutex);
 858
 859         map<DeviceSpec, DeviceInfo> devices;
 860         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 861                 const DeviceSpec spec{ InputSourceType::CAPTURE_CARD, card_index };
 862                 const AudioDevice *device = &video_cards[card_index];
 863                 DeviceInfo info;
 864                 info.display_name = device->display_name;
 865                 info.num_channels = 8;
 866                 devices.insert(make_pair(spec, info));
 867         }
 868         vector<ALSAPool::Device> available_alsa_devices = alsa_pool.get_devices();
 869         for (unsigned card_index = 0; card_index < available_alsa_devices.size(); ++card_index) {
 870                 const DeviceSpec spec{ InputSourceType::ALSA_INPUT, card_index };
 871                 const ALSAPool::Device &device = available_alsa_devices[card_index];
 872                 DeviceInfo info;
 873                 info.display_name = device.display_name();
 874                 info.num_channels = device.num_channels;
 875                 info.alsa_name = device.name;
 876                 info.alsa_info = device.info;
 877                 info.alsa_address = device.address;
 878                 devices.insert(make_pair(spec, info));
 879         }
 880         return devices;
 881 }
 882
 883 void AudioMixer::set_display_name(DeviceSpec device_spec, const string &name)
 884 {
 885         AudioDevice *device = find_audio_device(device_spec);
 886
 887         lock_guard<timed_mutex> lock(audio_mutex);
 888         device->display_name = name;
 889 }
 890
 891 void AudioMixer::serialize_device(DeviceSpec device_spec, DeviceSpecProto *device_spec_proto)
 892 {
 893         lock_guard<timed_mutex> lock(audio_mutex);
 894         switch (device_spec.type) {
 895                 case InputSourceType::SILENCE:
 896                         device_spec_proto->set_type(DeviceSpecProto::SILENCE);
 897                         break;
 898                 case InputSourceType::CAPTURE_CARD:
 899                         device_spec_proto->set_type(DeviceSpecProto::CAPTURE_CARD);
 900                         device_spec_proto->set_index(device_spec.index);
 901                         device_spec_proto->set_display_name(video_cards[device_spec.index].display_name);
 902                         break;
 903                 case InputSourceType::ALSA_INPUT:
 904                         alsa_pool.serialize_device(device_spec.index, device_spec_proto);
 905                         break;
 906         }
 907 }
 908
 909 void AudioMixer::set_simple_input(unsigned card_index)
 910 {
 911         InputMapping new_input_mapping;
 912         InputMapping::Bus input;
 913         input.name = "Main";
 914         input.device.type = InputSourceType::CAPTURE_CARD;
 915         input.device.index = card_index;
 916         input.source_channel[0] = 0;
 917         input.source_channel[1] = 1;
 918
 919         new_input_mapping.buses.push_back(input);
 920
 921         lock_guard<timed_mutex> lock(audio_mutex);
 922         current_mapping_mode = MappingMode::SIMPLE;
 923         set_input_mapping_lock_held(new_input_mapping);
 924         fader_volume_db[0] = 0.0f;
 925 }
 926
 927 unsigned AudioMixer::get_simple_input() const
 928 {
 929         lock_guard<timed_mutex> lock(audio_mutex);
 930         if (input_mapping.buses.size() == 1 &&
 931             input_mapping.buses[0].device.type == InputSourceType::CAPTURE_CARD &&
 932             input_mapping.buses[0].source_channel[0] == 0 &&
 933             input_mapping.buses[0].source_channel[1] == 1) {
 934                 return input_mapping.buses[0].device.index;
 935         } else {
 936                 return numeric_limits<unsigned>::max();
 937         }
 938 }
 939
 940 void AudioMixer::set_input_mapping(const InputMapping &new_input_mapping)
 941 {
 942         lock_guard<timed_mutex> lock(audio_mutex);
 943         set_input_mapping_lock_held(new_input_mapping);
 944         current_mapping_mode = MappingMode::MULTICHANNEL;
 945 }
 946
 947 AudioMixer::MappingMode AudioMixer::get_mapping_mode() const
 948 {
 949         lock_guard<timed_mutex> lock(audio_mutex);
 950         return current_mapping_mode;
 951 }
 952
 953 void AudioMixer::set_input_mapping_lock_held(const InputMapping &new_input_mapping)
 954 {
 955         map<DeviceSpec, set<unsigned>> interesting_channels;
 956         for (const InputMapping::Bus &bus : new_input_mapping.buses) {
 957                 if (bus.device.type == InputSourceType::CAPTURE_CARD ||
 958                     bus.device.type == InputSourceType::ALSA_INPUT) {
 959                         for (unsigned channel = 0; channel < 2; ++channel) {
 960                                 if (bus.source_channel[channel] != -1) {
 961                                         interesting_channels[bus.device].insert(bus.source_channel[channel]);
 962                                 }
 963                         }
 964                 }
 965         }
 966
 967         // Reset resamplers for all cards that don't have the exact same state as before.
 968         for (unsigned card_index = 0; card_index < MAX_VIDEO_CARDS; ++card_index) {
 969                 const DeviceSpec device_spec{InputSourceType::CAPTURE_CARD, card_index};
 970                 AudioDevice *device = find_audio_device(device_spec);
 971                 if (device->interesting_channels != interesting_channels[device_spec]) {
 972                         device->interesting_channels = interesting_channels[device_spec];
 973                         reset_resampler_mutex_held(device_spec);
 974                 }
 975         }
 976         for (unsigned card_index = 0; card_index < MAX_ALSA_CARDS; ++card_index) {
 977                 const DeviceSpec device_spec{InputSourceType::ALSA_INPUT, card_index};
 978                 AudioDevice *device = find_audio_device(device_spec);
 979                 if (interesting_channels[device_spec].empty()) {
 980                         alsa_pool.release_device(card_index);
 981                 } else {
 982                         alsa_pool.hold_device(card_index);
 983                 }
 984                 if (device->interesting_channels != interesting_channels[device_spec]) {
 985                         device->interesting_channels = interesting_channels[device_spec];
 986                         alsa_pool.reset_device(device_spec.index);
 987                         reset_resampler_mutex_held(device_spec);
 988                 }
 989         }
 990
 991         input_mapping = new_input_mapping;
 992 }
 993
 994 InputMapping AudioMixer::get_input_mapping() const
 995 {
 996         lock_guard<timed_mutex> lock(audio_mutex);
 997         return input_mapping;
 998 }
 999
1000 unsigned AudioMixer::num_buses() const
1001 {
1002         lock_guard<timed_mutex> lock(audio_mutex);
1003         return input_mapping.buses.size();
1004 }
1005
1006 void AudioMixer::reset_peak(unsigned bus_index)
1007 {
1008         lock_guard<timed_mutex> lock(audio_mutex);
1009         for (unsigned channel = 0; channel < 2; ++channel) {
1010                 PeakHistory &history = peak_history[bus_index][channel];
1011                 history.current_level = 0.0f;
1012                 history.historic_peak = 0.0f;
1013                 history.current_peak = 0.0f;
1014                 history.last_peak = 0.0f;
1015                 history.age_seconds = 0.0f;
1016         }
1017 }
1018
1019 AudioMixer *global_audio_mixer = nullptr;