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

   1 #include "audio_mixer.h"
   2
   3 #include <assert.h>
   4 #include <endian.h>
   5 #include <bmusb/bmusb.h>
   6 #include <stdio.h>
   7 #include <endian.h>
   8 #include <cmath>
   9 #include <limits>
  10 #ifdef __SSE__
  11 #include <immintrin.h>
  12 #endif
  13
  14 #include "db.h"
  15 #include "flags.h"
  16 #include "mixer.h"
  17 #include "state.pb.h"
  18 #include "timebase.h"
  19
  20 using namespace bmusb;
  21 using namespace std;
  22 using namespace std::placeholders;
  23
  24 namespace {
  25
  26 // TODO: If these prove to be a bottleneck, they can be SSSE3-optimized
  27 // (usually including multiple channels at a time).
  28
  29 void convert_fixed16_to_fp32(float *dst, size_t out_channel, size_t out_num_channels,
  30                              const uint8_t *src, size_t in_channel, size_t in_num_channels,
  31                              size_t num_samples)
  32 {
  33         assert(in_channel < in_num_channels);
  34         assert(out_channel < out_num_channels);
  35         src += in_channel * 2;
  36         dst += out_channel;
  37
  38         for (size_t i = 0; i < num_samples; ++i) {
  39                 int16_t s = le16toh(*(int16_t *)src);
  40                 *dst = s * (1.0f / 32768.0f);
  41
  42                 src += 2 * in_num_channels;
  43                 dst += out_num_channels;
  44         }
  45 }
  46
  47 void convert_fixed24_to_fp32(float *dst, size_t out_channel, size_t out_num_channels,
  48                              const uint8_t *src, size_t in_channel, size_t in_num_channels,
  49                              size_t num_samples)
  50 {
  51         assert(in_channel < in_num_channels);
  52         assert(out_channel < out_num_channels);
  53         src += in_channel * 3;
  54         dst += out_channel;
  55
  56         for (size_t i = 0; i < num_samples; ++i) {
  57                 uint32_t s1 = src[0];
  58                 uint32_t s2 = src[1];
  59                 uint32_t s3 = src[2];
  60                 uint32_t s = s1 | (s1 << 8) | (s2 << 16) | (s3 << 24);
  61                 *dst = int(s) * (1.0f / 2147483648.0f);
  62
  63                 src += 3 * in_num_channels;
  64                 dst += out_num_channels;
  65         }
  66 }
  67
  68 void convert_fixed32_to_fp32(float *dst, size_t out_channel, size_t out_num_channels,
  69                              const uint8_t *src, size_t in_channel, size_t in_num_channels,
  70                              size_t num_samples)
  71 {
  72         assert(in_channel < in_num_channels);
  73         assert(out_channel < out_num_channels);
  74         src += in_channel * 4;
  75         dst += out_channel;
  76
  77         for (size_t i = 0; i < num_samples; ++i) {
  78                 int32_t s = le32toh(*(int32_t *)src);
  79                 *dst = s * (1.0f / 2147483648.0f);
  80
  81                 src += 4 * in_num_channels;
  82                 dst += out_num_channels;
  83         }
  84 }
  85
  86 float find_peak_plain(const float *samples, size_t num_samples) __attribute__((unused));
  87
  88 float find_peak_plain(const float *samples, size_t num_samples)
  89 {
  90         float m = fabs(samples[0]);
  91         for (size_t i = 1; i < num_samples; ++i) {
  92                 m = max(m, fabs(samples[i]));
  93         }
  94         return m;
  95 }
  96
  97 #ifdef __SSE__
  98 static inline float horizontal_max(__m128 m)
  99 {
 100         __m128 tmp = _mm_shuffle_ps(m, m, _MM_SHUFFLE(1, 0, 3, 2));
 101         m = _mm_max_ps(m, tmp);
 102         tmp = _mm_shuffle_ps(m, m, _MM_SHUFFLE(2, 3, 0, 1));
 103         m = _mm_max_ps(m, tmp);
 104         return _mm_cvtss_f32(m);
 105 }
 106
 107 float find_peak(const float *samples, size_t num_samples)
 108 {
 109         const __m128 abs_mask = _mm_castsi128_ps(_mm_set1_epi32(0x7fffffffu));
 110         __m128 m = _mm_setzero_ps();
 111         for (size_t i = 0; i < (num_samples & ~3); i += 4) {
 112                 __m128 x = _mm_loadu_ps(samples + i);
 113                 x = _mm_and_ps(x, abs_mask);
 114                 m = _mm_max_ps(m, x);
 115         }
 116         float result = horizontal_max(m);
 117
 118         for (size_t i = (num_samples & ~3); i < num_samples; ++i) {
 119                 result = max(result, fabs(samples[i]));
 120         }
 121
 122 #if 0
 123         // Self-test. We should be bit-exact the same.
 124         float reference_result = find_peak_plain(samples, num_samples);
 125         if (result != reference_result) {
 126                 fprintf(stderr, "Error: Peak is %f [%f %f %f %f]; should be %f.\n",
 127                         result,
 128                         _mm_cvtss_f32(_mm_shuffle_ps(m, m, _MM_SHUFFLE(0, 0, 0, 0))),
 129                         _mm_cvtss_f32(_mm_shuffle_ps(m, m, _MM_SHUFFLE(1, 1, 1, 1))),
 130                         _mm_cvtss_f32(_mm_shuffle_ps(m, m, _MM_SHUFFLE(2, 2, 2, 2))),
 131                         _mm_cvtss_f32(_mm_shuffle_ps(m, m, _MM_SHUFFLE(3, 3, 3, 3))),
 132                         reference_result);
 133                 abort();
 134         }
 135 #endif
 136         return result;
 137 }
 138 #else
 139 float find_peak(const float *samples, size_t num_samples)
 140 {
 141         return find_peak_plain(samples, num_samples);
 142 }
 143 #endif
 144
 145 void deinterleave_samples(const vector<float> &in, vector<float> *out_l, vector<float> *out_r)
 146 {
 147         size_t num_samples = in.size() / 2;
 148         out_l->resize(num_samples);
 149         out_r->resize(num_samples);
 150
 151         const float *inptr = in.data();
 152         float *lptr = &(*out_l)[0];
 153         float *rptr = &(*out_r)[0];
 154         for (size_t i = 0; i < num_samples; ++i) {
 155                 *lptr++ = *inptr++;
 156                 *rptr++ = *inptr++;
 157         }
 158 }
 159
 160 }  // namespace
 161
 162 AudioMixer::AudioMixer(unsigned num_cards)
 163         : num_cards(num_cards),
 164           limiter(OUTPUT_FREQUENCY),
 165           correlation(OUTPUT_FREQUENCY)
 166 {
 167         global_audio_mixer = this;
 168
 169         for (unsigned bus_index = 0; bus_index < MAX_BUSES; ++bus_index) {
 170                 locut[bus_index].init(FILTER_HPF, 2);
 171                 eq[bus_index][EQ_BAND_BASS].init(FILTER_LOW_SHELF, 1);
 172                 // Note: EQ_BAND_MID isn't used (see comments in apply_eq()).
 173                 eq[bus_index][EQ_BAND_TREBLE].init(FILTER_HIGH_SHELF, 1);
 174                 compressor[bus_index].reset(new StereoCompressor(OUTPUT_FREQUENCY));
 175                 level_compressor[bus_index].reset(new StereoCompressor(OUTPUT_FREQUENCY));
 176
 177                 set_bus_settings(bus_index, get_default_bus_settings());
 178         }
 179         set_limiter_enabled(global_flags.limiter_enabled);
 180         set_final_makeup_gain_auto(global_flags.final_makeup_gain_auto);
 181         alsa_pool.init();
 182
 183         if (!global_flags.input_mapping_filename.empty()) {
 184                 current_mapping_mode = MappingMode::MULTICHANNEL;
 185                 InputMapping new_input_mapping;
 186                 if (!load_input_mapping_from_file(get_devices(),
 187                                                   global_flags.input_mapping_filename,
 188                                                   &new_input_mapping)) {
 189                         fprintf(stderr, "Failed to load input mapping from '%s', exiting.\n",
 190                                 global_flags.input_mapping_filename.c_str());
 191                         exit(1);
 192                 }
 193                 set_input_mapping(new_input_mapping);
 194         } else {
 195                 set_simple_input(/*card_index=*/0);
 196                 if (global_flags.multichannel_mapping_mode) {
 197                         current_mapping_mode = MappingMode::MULTICHANNEL;
 198                 }
 199         }
 200
 201         r128.init(2, OUTPUT_FREQUENCY);
 202         r128.integr_start();
 203
 204         // hlen=16 is pretty low quality, but we use quite a bit of CPU otherwise,
 205         // and there's a limit to how important the peak meter is.
 206         peak_resampler.setup(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY * 4, /*num_channels=*/2, /*hlen=*/16, /*frel=*/1.0);
 207 }
 208
 209 void AudioMixer::reset_resampler(DeviceSpec device_spec)
 210 {
 211         lock_guard<timed_mutex> lock(audio_mutex);
 212         reset_resampler_mutex_held(device_spec);
 213 }
 214
 215 void AudioMixer::reset_resampler_mutex_held(DeviceSpec device_spec)
 216 {
 217         AudioDevice *device = find_audio_device(device_spec);
 218
 219         if (device->interesting_channels.empty()) {
 220                 device->resampling_queue.reset();
 221         } else {
 222                 // TODO: ResamplingQueue should probably take the full device spec.
 223                 // (It's only used for console output, though.)
 224                 device->resampling_queue.reset(new ResamplingQueue(device_spec.index, device->capture_frequency, OUTPUT_FREQUENCY, device->interesting_channels.size()));
 225         }
 226         device->next_local_pts = 0;
 227 }
 228
 229 bool AudioMixer::add_audio(DeviceSpec device_spec, const uint8_t *data, unsigned num_samples, AudioFormat audio_format, int64_t frame_length)
 230 {
 231         AudioDevice *device = find_audio_device(device_spec);
 232
 233         unique_lock<timed_mutex> lock(audio_mutex, defer_lock);
 234         if (!lock.try_lock_for(chrono::milliseconds(10))) {
 235                 return false;
 236         }
 237         if (device->resampling_queue == nullptr) {
 238                 // No buses use this device; throw it away.
 239                 return true;
 240         }
 241
 242         unsigned num_channels = device->interesting_channels.size();
 243         assert(num_channels > 0);
 244
 245         // Convert the audio to fp32.
 246         unique_ptr<float[]> audio(new float[num_samples * num_channels]);
 247         unsigned channel_index = 0;
 248         for (auto channel_it = device->interesting_channels.cbegin(); channel_it != device->interesting_channels.end(); ++channel_it, ++channel_index) {
 249                 switch (audio_format.bits_per_sample) {
 250                 case 0:
 251                         assert(num_samples == 0);
 252                         break;
 253                 case 16:
 254                         convert_fixed16_to_fp32(audio.get(), channel_index, num_channels, data, *channel_it, audio_format.num_channels, num_samples);
 255                         break;
 256                 case 24:
 257                         convert_fixed24_to_fp32(audio.get(), channel_index, num_channels, data, *channel_it, audio_format.num_channels, num_samples);
 258                         break;
 259                 case 32:
 260                         convert_fixed32_to_fp32(audio.get(), channel_index, num_channels, data, *channel_it, audio_format.num_channels, num_samples);
 261                         break;
 262                 default:
 263                         fprintf(stderr, "Cannot handle audio with %u bits per sample\n", audio_format.bits_per_sample);
 264                         assert(false);
 265                 }
 266         }
 267
 268         // Now add it.
 269         int64_t local_pts = device->next_local_pts;
 270         device->resampling_queue->add_input_samples(local_pts / double(TIMEBASE), audio.get(), num_samples);
 271         device->next_local_pts = local_pts + frame_length;
 272         return true;
 273 }
 274
 275 bool AudioMixer::add_silence(DeviceSpec device_spec, unsigned samples_per_frame, unsigned num_frames, int64_t frame_length)
 276 {
 277         AudioDevice *device = find_audio_device(device_spec);
 278
 279         unique_lock<timed_mutex> lock(audio_mutex, defer_lock);
 280         if (!lock.try_lock_for(chrono::milliseconds(10))) {
 281                 return false;
 282         }
 283         if (device->resampling_queue == nullptr) {
 284                 // No buses use this device; throw it away.
 285                 return true;
 286         }
 287
 288         unsigned num_channels = device->interesting_channels.size();
 289         assert(num_channels > 0);
 290
 291         vector<float> silence(samples_per_frame * num_channels, 0.0f);
 292         for (unsigned i = 0; i < num_frames; ++i) {
 293                 device->resampling_queue->add_input_samples(device->next_local_pts / double(TIMEBASE), silence.data(), samples_per_frame);
 294                 // Note that if the format changed in the meantime, we have
 295                 // no way of detecting that; we just have to assume the frame length
 296                 // is always the same.
 297                 device->next_local_pts += frame_length;
 298         }
 299         return true;
 300 }
 301
 302 bool AudioMixer::silence_card(DeviceSpec device_spec, bool silence)
 303 {
 304         AudioDevice *device = find_audio_device(device_spec);
 305
 306         unique_lock<timed_mutex> lock(audio_mutex, defer_lock);
 307         if (!lock.try_lock_for(chrono::milliseconds(10))) {
 308                 return false;
 309         }
 310
 311         if (device->silenced && !silence) {
 312                 reset_resampler_mutex_held(device_spec);
 313         }
 314         device->silenced = silence;
 315         return true;
 316 }
 317
 318 AudioMixer::BusSettings AudioMixer::get_default_bus_settings()
 319 {
 320         BusSettings settings;
 321         settings.fader_volume_db = 0.0f;
 322         settings.muted = false;
 323         settings.locut_enabled = global_flags.locut_enabled;
 324         for (unsigned band_index = 0; band_index < NUM_EQ_BANDS; ++band_index) {
 325                 settings.eq_level_db[band_index] = 0.0f;
 326         }
 327         settings.gain_staging_db = global_flags.initial_gain_staging_db;
 328         settings.level_compressor_enabled = global_flags.gain_staging_auto;
 329         settings.compressor_threshold_dbfs = ref_level_dbfs - 12.0f;  // -12 dB.
 330         settings.compressor_enabled = global_flags.compressor_enabled;
 331         return settings;
 332 }
 333
 334 AudioMixer::BusSettings AudioMixer::get_bus_settings(unsigned bus_index) const
 335 {
 336         lock_guard<timed_mutex> lock(audio_mutex);
 337         BusSettings settings;
 338         settings.fader_volume_db = fader_volume_db[bus_index];
 339         settings.muted = mute[bus_index];
 340         settings.locut_enabled = locut_enabled[bus_index];
 341         for (unsigned band_index = 0; band_index < NUM_EQ_BANDS; ++band_index) {
 342                 settings.eq_level_db[band_index] = eq_level_db[bus_index][band_index];
 343         }
 344         settings.gain_staging_db = gain_staging_db[bus_index];
 345         settings.level_compressor_enabled = level_compressor_enabled[bus_index];
 346         settings.compressor_threshold_dbfs = compressor_threshold_dbfs[bus_index];
 347         settings.compressor_enabled = compressor_enabled[bus_index];
 348         return settings;
 349 }
 350
 351 void AudioMixer::set_bus_settings(unsigned bus_index, const AudioMixer::BusSettings &settings)
 352 {
 353         lock_guard<timed_mutex> lock(audio_mutex);
 354         fader_volume_db[bus_index] = settings.fader_volume_db;
 355         mute[bus_index] = settings.muted;
 356         locut_enabled[bus_index] = settings.locut_enabled;
 357         for (unsigned band_index = 0; band_index < NUM_EQ_BANDS; ++band_index) {
 358                 eq_level_db[bus_index][band_index] = settings.eq_level_db[band_index];
 359         }
 360         gain_staging_db[bus_index] = settings.gain_staging_db;
 361         last_gain_staging_db[bus_index] = gain_staging_db[bus_index];
 362         level_compressor_enabled[bus_index] = settings.level_compressor_enabled;
 363         compressor_threshold_dbfs[bus_index] = settings.compressor_threshold_dbfs;
 364         compressor_enabled[bus_index] = settings.compressor_enabled;
 365 }
 366
 367 AudioMixer::AudioDevice *AudioMixer::find_audio_device(DeviceSpec device)
 368 {
 369         switch (device.type) {
 370         case InputSourceType::CAPTURE_CARD:
 371                 return &video_cards[device.index];
 372         case InputSourceType::ALSA_INPUT:
 373                 return &alsa_inputs[device.index];
 374         case InputSourceType::SILENCE:
 375         default:
 376                 assert(false);
 377         }
 378         return nullptr;
 379 }
 380
 381 // Get a pointer to the given channel from the given device.
 382 // The channel must be picked out earlier and resampled.
 383 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)
 384 {
 385         static float zero = 0.0f;
 386         if (source_channel == -1 || device_spec.type == InputSourceType::SILENCE) {
 387                 *srcptr = &zero;
 388                 *stride = 0;
 389                 return;
 390         }
 391         AudioDevice *device = find_audio_device(device_spec);
 392         assert(device->interesting_channels.count(source_channel) != 0);
 393         unsigned channel_index = 0;
 394         for (int channel : device->interesting_channels) {
 395                 if (channel == source_channel) break;
 396                 ++channel_index;
 397         }
 398         assert(channel_index < device->interesting_channels.size());
 399         const auto it = samples_card.find(device_spec);
 400         assert(it != samples_card.end());
 401         *srcptr = &(it->second)[channel_index];
 402         *stride = device->interesting_channels.size();
 403 }
 404
 405 // TODO: Can be SSSE3-optimized if need be.
 406 void AudioMixer::fill_audio_bus(const map<DeviceSpec, vector<float>> &samples_card, const InputMapping::Bus &bus, unsigned num_samples, float *output)
 407 {
 408         if (bus.device.type == InputSourceType::SILENCE) {
 409                 memset(output, 0, num_samples * 2 * sizeof(*output));
 410         } else {
 411                 assert(bus.device.type == InputSourceType::CAPTURE_CARD ||
 412                        bus.device.type == InputSourceType::ALSA_INPUT);
 413                 const float *lsrc, *rsrc;
 414                 unsigned lstride, rstride;
 415                 float *dptr = output;
 416                 find_sample_src_from_device(samples_card, bus.device, bus.source_channel[0], &lsrc, &lstride);
 417                 find_sample_src_from_device(samples_card, bus.device, bus.source_channel[1], &rsrc, &rstride);
 418                 for (unsigned i = 0; i < num_samples; ++i) {
 419                         *dptr++ = *lsrc;
 420                         *dptr++ = *rsrc;
 421                         lsrc += lstride;
 422                         rsrc += rstride;
 423                 }
 424         }
 425 }
 426
 427 vector<DeviceSpec> AudioMixer::get_active_devices() const
 428 {
 429         vector<DeviceSpec> ret;
 430         for (unsigned card_index = 0; card_index < MAX_VIDEO_CARDS; ++card_index) {
 431                 const DeviceSpec device_spec{InputSourceType::CAPTURE_CARD, card_index};
 432                 if (!find_audio_device(device_spec)->interesting_channels.empty()) {
 433                         ret.push_back(device_spec);
 434                 }
 435         }
 436         for (unsigned card_index = 0; card_index < MAX_ALSA_CARDS; ++card_index) {
 437                 const DeviceSpec device_spec{InputSourceType::ALSA_INPUT, card_index};
 438                 if (!find_audio_device(device_spec)->interesting_channels.empty()) {
 439                         ret.push_back(device_spec);
 440                 }
 441         }
 442         return ret;
 443 }
 444
 445 namespace {
 446
 447 void apply_gain(float db, float last_db, vector<float> *samples)
 448 {
 449         if (fabs(db - last_db) < 1e-3) {
 450                 // Constant over this frame.
 451                 const float gain = from_db(db);
 452                 for (size_t i = 0; i < samples->size(); ++i) {
 453                         (*samples)[i] *= gain;
 454                 }
 455         } else {
 456                 // We need to do a fade.
 457                 unsigned num_samples = samples->size() / 2;
 458                 float gain = from_db(last_db);
 459                 const float gain_inc = pow(from_db(db - last_db), 1.0 / num_samples);
 460                 for (size_t i = 0; i < num_samples; ++i) {
 461                         (*samples)[i * 2 + 0] *= gain;
 462                         (*samples)[i * 2 + 1] *= gain;
 463                         gain *= gain_inc;
 464                 }
 465         }
 466 }
 467
 468 }  // namespace
 469
 470 vector<float> AudioMixer::get_output(double pts, unsigned num_samples, ResamplingQueue::RateAdjustmentPolicy rate_adjustment_policy)
 471 {
 472         map<DeviceSpec, vector<float>> samples_card;
 473         vector<float> samples_bus;
 474
 475         lock_guard<timed_mutex> lock(audio_mutex);
 476
 477         // Pick out all the interesting channels from all the cards.
 478         for (const DeviceSpec &device_spec : get_active_devices()) {
 479                 AudioDevice *device = find_audio_device(device_spec);
 480                 samples_card[device_spec].resize(num_samples * device->interesting_channels.size());
 481                 if (device->silenced) {
 482                         memset(&samples_card[device_spec][0], 0, samples_card[device_spec].size() * sizeof(float));
 483                 } else {
 484                         device->resampling_queue->get_output_samples(
 485                                 pts,
 486                                 &samples_card[device_spec][0],
 487                                 num_samples,
 488                                 rate_adjustment_policy);
 489                 }
 490         }
 491
 492         vector<float> samples_out, left, right;
 493         samples_out.resize(num_samples * 2);
 494         samples_bus.resize(num_samples * 2);
 495         for (unsigned bus_index = 0; bus_index < input_mapping.buses.size(); ++bus_index) {
 496                 fill_audio_bus(samples_card, input_mapping.buses[bus_index], num_samples, &samples_bus[0]);
 497                 apply_eq(bus_index, &samples_bus);
 498
 499                 {
 500                         lock_guard<mutex> lock(compressor_mutex);
 501
 502                         // Apply a level compressor to get the general level right.
 503                         // Basically, if it's over about -40 dBFS, we squeeze it down to that level
 504                         // (or more precisely, near it, since we don't use infinite ratio),
 505                         // then apply a makeup gain to get it to -14 dBFS. -14 dBFS is, of course,
 506                         // entirely arbitrary, but from practical tests with speech, it seems to
 507                         // put ut around -23 LUFS, so it's a reasonable starting point for later use.
 508                         if (level_compressor_enabled[bus_index]) {
 509                                 float threshold = 0.01f;   // -40 dBFS.
 510                                 float ratio = 20.0f;
 511                                 float attack_time = 0.5f;
 512                                 float release_time = 20.0f;
 513                                 float makeup_gain = from_db(ref_level_dbfs - (-40.0f));  // +26 dB.
 514                                 level_compressor[bus_index]->process(samples_bus.data(), samples_bus.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 515                                 gain_staging_db[bus_index] = to_db(level_compressor[bus_index]->get_attenuation() * makeup_gain);
 516                         } else {
 517                                 // Just apply the gain we already had.
 518                                 float db = gain_staging_db[bus_index];
 519                                 float last_db = last_gain_staging_db[bus_index];
 520                                 apply_gain(db, last_db, &samples_bus);
 521                         }
 522                         last_gain_staging_db[bus_index] = gain_staging_db[bus_index];
 523
 524 #if 0
 525                         printf("level=%f (%+5.2f dBFS) attenuation=%f (%+5.2f dB) end_result=%+5.2f dB\n",
 526                                 level_compressor.get_level(), to_db(level_compressor.get_level()),
 527                                 level_compressor.get_attenuation(), to_db(level_compressor.get_attenuation()),
 528                                 to_db(level_compressor.get_level() * level_compressor.get_attenuation() * makeup_gain));
 529 #endif
 530
 531                         // The real compressor.
 532                         if (compressor_enabled[bus_index]) {
 533                                 float threshold = from_db(compressor_threshold_dbfs[bus_index]);
 534                                 float ratio = 20.0f;
 535                                 float attack_time = 0.005f;
 536                                 float release_time = 0.040f;
 537                                 float makeup_gain = 2.0f;  // +6 dB.
 538                                 compressor[bus_index]->process(samples_bus.data(), samples_bus.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 539                 //              compressor_att = compressor.get_attenuation();
 540                         }
 541                 }
 542
 543                 add_bus_to_master(bus_index, samples_bus, &samples_out);
 544                 deinterleave_samples(samples_bus, &left, &right);
 545                 measure_bus_levels(bus_index, left, right);
 546         }
 547
 548         {
 549                 lock_guard<mutex> lock(compressor_mutex);
 550
 551                 // Finally a limiter at -4 dB (so, -10 dBFS) to take out the worst peaks only.
 552                 // Note that since ratio is not infinite, we could go slightly higher than this.
 553                 if (limiter_enabled) {
 554                         float threshold = from_db(limiter_threshold_dbfs);
 555                         float ratio = 30.0f;
 556                         float attack_time = 0.0f;  // Instant.
 557                         float release_time = 0.020f;
 558                         float makeup_gain = 1.0f;  // 0 dB.
 559                         limiter.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 560         //              limiter_att = limiter.get_attenuation();
 561                 }
 562
 563         //      printf("limiter=%+5.1f  compressor=%+5.1f\n", to_db(limiter_att), to_db(compressor_att));
 564         }
 565
 566         // At this point, we are most likely close to +0 LU (at least if the
 567         // faders sum to 0 dB and the compressors are on), but all of our
 568         // measurements have been on raw sample values, not R128 values.
 569         // So we have a final makeup gain to get us to +0 LU; the gain
 570         // adjustments required should be relatively small, and also, the
 571         // offset shouldn't change much (only if the type of audio changes
 572         // significantly). Thus, we shoot for updating this value basically
 573         // “whenever we process buffers”, since the R128 calculation isn't exactly
 574         // something we get out per-sample.
 575         //
 576         // Note that there's a feedback loop here, so we choose a very slow filter
 577         // (half-time of 30 seconds).
 578         double target_loudness_factor, alpha;
 579         double loudness_lu = r128.loudness_M() - ref_level_lufs;
 580         double current_makeup_lu = to_db(final_makeup_gain);
 581         target_loudness_factor = final_makeup_gain * from_db(-loudness_lu);
 582
 583         // If we're outside +/- 5 LU uncorrected, we don't count it as
 584         // a normal signal (probably silence) and don't change the
 585         // correction factor; just apply what we already have.
 586         if (fabs(loudness_lu - current_makeup_lu) >= 5.0 || !final_makeup_gain_auto) {
 587                 alpha = 0.0;
 588         } else {
 589                 // Formula adapted from
 590                 // https://en.wikipedia.org/wiki/Low-pass_filter#Simple_infinite_impulse_response_filter.
 591                 const double half_time_s = 30.0;
 592                 const double fc_mul_2pi_delta_t = 1.0 / (half_time_s * OUTPUT_FREQUENCY);
 593                 alpha = fc_mul_2pi_delta_t / (fc_mul_2pi_delta_t + 1.0);
 594         }
 595
 596         {
 597                 lock_guard<mutex> lock(compressor_mutex);
 598                 double m = final_makeup_gain;
 599                 for (size_t i = 0; i < samples_out.size(); i += 2) {
 600                         samples_out[i + 0] *= m;
 601                         samples_out[i + 1] *= m;
 602                         m += (target_loudness_factor - m) * alpha;
 603                 }
 604                 final_makeup_gain = m;
 605         }
 606
 607         update_meters(samples_out);
 608
 609         return samples_out;
 610 }
 611
 612 namespace {
 613
 614 void apply_filter_fade(StereoFilter *filter, float *data, unsigned num_samples, float cutoff_hz, float db, float last_db)
 615 {
 616         // A granularity of 32 samples is an okay tradeoff between speed and
 617         // smoothness; recalculating the filters is pretty expensive, so it's
 618         // good that we don't do this all the time.
 619         static constexpr unsigned filter_granularity_samples = 32;
 620
 621         const float cutoff_linear = cutoff_hz * 2.0 * M_PI / OUTPUT_FREQUENCY;
 622         if (fabs(db - last_db) < 1e-3) {
 623                 // Constant over this frame.
 624                 if (fabs(db) > 0.01f) {
 625                         filter->render(data, num_samples, cutoff_linear, 0.5f, db / 40.0f);
 626                 }
 627         } else {
 628                 // We need to do a fade. (Rounding up avoids division by zero.)
 629                 unsigned num_blocks = (num_samples + filter_granularity_samples - 1) / filter_granularity_samples;
 630                 const float inc_db_norm = (db - last_db) / 40.0f / num_blocks;
 631                 float db_norm = db / 40.0f;
 632                 for (size_t i = 0; i < num_samples; i += filter_granularity_samples) {
 633                         size_t samples_this_block = std::min<size_t>(num_samples - i, filter_granularity_samples);
 634                         filter->render(data + i * 2, samples_this_block, cutoff_linear, 0.5f, db_norm);
 635                         db_norm += inc_db_norm;
 636                 }
 637         }
 638 }
 639
 640 }  // namespace
 641
 642 void AudioMixer::apply_eq(unsigned bus_index, vector<float> *samples_bus)
 643 {
 644         constexpr float bass_freq_hz = 200.0f;
 645         constexpr float treble_freq_hz = 4700.0f;
 646
 647         // Cut away everything under 120 Hz (or whatever the cutoff is);
 648         // we don't need it for voice, and it will reduce headroom
 649         // and confuse the compressor. (In particular, any hums at 50 or 60 Hz
 650         // should be dampened.)
 651         if (locut_enabled[bus_index]) {
 652                 locut[bus_index].render(samples_bus->data(), samples_bus->size() / 2, locut_cutoff_hz * 2.0 * M_PI / OUTPUT_FREQUENCY, 0.5f);
 653         }
 654
 655         // Apply the rest of the EQ. Since we only have a simple three-band EQ,
 656         // we can implement it with two shelf filters. We use a simple gain to
 657         // set the mid-level filter, and then offset the low and high bands
 658         // from that if we need to. (We could perhaps have folded the gain into
 659         // the next part, but it's so cheap that the trouble isn't worth it.)
 660         //
 661         // If any part of the EQ has changed appreciably since last frame,
 662         // we fade smoothly during the course of this frame.
 663         const float bass_db = eq_level_db[bus_index][EQ_BAND_BASS];
 664         const float mid_db = eq_level_db[bus_index][EQ_BAND_MID];
 665         const float treble_db = eq_level_db[bus_index][EQ_BAND_TREBLE];
 666
 667         const float last_bass_db = last_eq_level_db[bus_index][EQ_BAND_BASS];
 668         const float last_mid_db = last_eq_level_db[bus_index][EQ_BAND_MID];
 669         const float last_treble_db = last_eq_level_db[bus_index][EQ_BAND_TREBLE];
 670
 671         assert(samples_bus->size() % 2 == 0);
 672         const unsigned num_samples = samples_bus->size() / 2;
 673
 674         apply_gain(mid_db, last_mid_db, samples_bus);
 675
 676         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);
 677         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);
 678
 679         last_eq_level_db[bus_index][EQ_BAND_BASS] = bass_db;
 680         last_eq_level_db[bus_index][EQ_BAND_MID] = mid_db;
 681         last_eq_level_db[bus_index][EQ_BAND_TREBLE] = treble_db;
 682 }
 683
 684 void AudioMixer::add_bus_to_master(unsigned bus_index, const vector<float> &samples_bus, vector<float> *samples_out)
 685 {
 686         assert(samples_bus.size() == samples_out->size());
 687         assert(samples_bus.size() % 2 == 0);
 688         unsigned num_samples = samples_bus.size() / 2;
 689         const float new_volume_db = mute[bus_index] ? -90.0f : fader_volume_db[bus_index].load();
 690         if (fabs(new_volume_db - last_fader_volume_db[bus_index]) > 1e-3) {
 691                 // The volume has changed; do a fade over the course of this frame.
 692                 // (We might have some numerical issues here, but it seems to sound OK.)
 693                 // For the purpose of fading here, the silence floor is set to -90 dB
 694                 // (the fader only goes to -84).
 695                 float old_volume = from_db(max<float>(last_fader_volume_db[bus_index], -90.0f));
 696                 float volume = from_db(max<float>(new_volume_db, -90.0f));
 697
 698                 float volume_inc = pow(volume / old_volume, 1.0 / num_samples);
 699                 volume = old_volume;
 700                 if (bus_index == 0) {
 701                         for (unsigned i = 0; i < num_samples; ++i) {
 702                                 (*samples_out)[i * 2 + 0] = samples_bus[i * 2 + 0] * volume;
 703                                 (*samples_out)[i * 2 + 1] = samples_bus[i * 2 + 1] * volume;
 704                                 volume *= volume_inc;
 705                         }
 706                 } else {
 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                 }
 713         } else if (new_volume_db > -90.0f) {
 714                 float volume = from_db(new_volume_db);
 715                 if (bus_index == 0) {
 716                         for (unsigned i = 0; i < num_samples; ++i) {
 717                                 (*samples_out)[i * 2 + 0] = samples_bus[i * 2 + 0] * volume;
 718                                 (*samples_out)[i * 2 + 1] = samples_bus[i * 2 + 1] * volume;
 719                         }
 720                 } else {
 721                         for (unsigned i = 0; i < num_samples; ++i) {
 722                                 (*samples_out)[i * 2 + 0] += samples_bus[i * 2 + 0] * volume;
 723                                 (*samples_out)[i * 2 + 1] += samples_bus[i * 2 + 1] * volume;
 724                         }
 725                 }
 726         }
 727
 728         last_fader_volume_db[bus_index] = new_volume_db;
 729 }
 730
 731 void AudioMixer::measure_bus_levels(unsigned bus_index, const vector<float> &left, const vector<float> &right)
 732 {
 733         assert(left.size() == right.size());
 734         const float volume = mute[bus_index] ? 0.0f : from_db(fader_volume_db[bus_index]);
 735         const float peak_levels[2] = {
 736                 find_peak(left.data(), left.size()) * volume,
 737                 find_peak(right.data(), right.size()) * volume
 738         };
 739         for (unsigned channel = 0; channel < 2; ++channel) {
 740                 // Compute the current value, including hold and falloff.
 741                 // The constants are borrowed from zita-mu1 by Fons Adriaensen.
 742                 static constexpr float hold_sec = 0.5f;
 743                 static constexpr float falloff_db_sec = 15.0f;  // dB/sec falloff after hold.
 744                 float current_peak;
 745                 PeakHistory &history = peak_history[bus_index][channel];
 746                 history.historic_peak = max(history.historic_peak, peak_levels[channel]);
 747                 if (history.age_seconds < hold_sec) {
 748                         current_peak = history.last_peak;
 749                 } else {
 750                         current_peak = history.last_peak * from_db(-falloff_db_sec * (history.age_seconds - hold_sec));
 751                 }
 752
 753                 // See if we have a new peak to replace the old (possibly falling) one.
 754                 if (peak_levels[channel] > current_peak) {
 755                         history.last_peak = peak_levels[channel];
 756                         history.age_seconds = 0.0f;  // Not 100% correct, but more than good enough given our frame sizes.
 757                         current_peak = peak_levels[channel];
 758                 } else {
 759                         history.age_seconds += float(left.size()) / OUTPUT_FREQUENCY;
 760                 }
 761                 history.current_level = peak_levels[channel];
 762                 history.current_peak = current_peak;
 763         }
 764 }
 765
 766 void AudioMixer::update_meters(const vector<float> &samples)
 767 {
 768         // Upsample 4x to find interpolated peak.
 769         peak_resampler.inp_data = const_cast<float *>(samples.data());
 770         peak_resampler.inp_count = samples.size() / 2;
 771
 772         vector<float> interpolated_samples;
 773         interpolated_samples.resize(samples.size());
 774         {
 775                 lock_guard<mutex> lock(audio_measure_mutex);
 776
 777                 while (peak_resampler.inp_count > 0) {  // About four iterations.
 778                         peak_resampler.out_data = &interpolated_samples[0];
 779                         peak_resampler.out_count = interpolated_samples.size() / 2;
 780                         peak_resampler.process();
 781                         size_t out_stereo_samples = interpolated_samples.size() / 2 - peak_resampler.out_count;
 782                         peak = max<float>(peak, find_peak(interpolated_samples.data(), out_stereo_samples * 2));
 783                         peak_resampler.out_data = nullptr;
 784                 }
 785         }
 786
 787         // Find R128 levels and L/R correlation.
 788         vector<float> left, right;
 789         deinterleave_samples(samples, &left, &right);
 790         float *ptrs[] = { left.data(), right.data() };
 791         {
 792                 lock_guard<mutex> lock(audio_measure_mutex);
 793                 r128.process(left.size(), ptrs);
 794                 correlation.process_samples(samples);
 795         }
 796
 797         send_audio_level_callback();
 798 }
 799
 800 void AudioMixer::reset_meters()
 801 {
 802         lock_guard<mutex> lock(audio_measure_mutex);
 803         peak_resampler.reset();
 804         peak = 0.0f;
 805         r128.reset();
 806         r128.integr_start();
 807         correlation.reset();
 808 }
 809
 810 void AudioMixer::send_audio_level_callback()
 811 {
 812         if (audio_level_callback == nullptr) {
 813                 return;
 814         }
 815
 816         lock_guard<mutex> lock(audio_measure_mutex);
 817         double loudness_s = r128.loudness_S();
 818         double loudness_i = r128.integrated();
 819         double loudness_range_low = r128.range_min();
 820         double loudness_range_high = r128.range_max();
 821
 822         vector<BusLevel> bus_levels;
 823         bus_levels.resize(input_mapping.buses.size());
 824         {
 825                 lock_guard<mutex> lock(compressor_mutex);
 826                 for (unsigned bus_index = 0; bus_index < bus_levels.size(); ++bus_index) {
 827                         bus_levels[bus_index].current_level_dbfs[0] = to_db(peak_history[bus_index][0].current_level);
 828                         bus_levels[bus_index].current_level_dbfs[1] = to_db(peak_history[bus_index][1].current_level);
 829                         bus_levels[bus_index].peak_level_dbfs[0] = to_db(peak_history[bus_index][0].current_peak);
 830                         bus_levels[bus_index].peak_level_dbfs[1] = to_db(peak_history[bus_index][1].current_peak);
 831                         bus_levels[bus_index].historic_peak_dbfs = to_db(
 832                                 max(peak_history[bus_index][0].historic_peak,
 833                                     peak_history[bus_index][1].historic_peak));
 834                         bus_levels[bus_index].gain_staging_db = gain_staging_db[bus_index];
 835                         if (compressor_enabled[bus_index]) {
 836                                 bus_levels[bus_index].compressor_attenuation_db = -to_db(compressor[bus_index]->get_attenuation());
 837                         } else {
 838                                 bus_levels[bus_index].compressor_attenuation_db = 0.0;
 839                         }
 840                 }
 841         }
 842
 843         audio_level_callback(loudness_s, to_db(peak), bus_levels,
 844                 loudness_i, loudness_range_low, loudness_range_high,
 845                 to_db(final_makeup_gain),
 846                 correlation.get_correlation());
 847 }
 848
 849 map<DeviceSpec, DeviceInfo> AudioMixer::get_devices()
 850 {
 851         lock_guard<timed_mutex> lock(audio_mutex);
 852
 853         map<DeviceSpec, DeviceInfo> devices;
 854         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 855                 const DeviceSpec spec{ InputSourceType::CAPTURE_CARD, card_index };
 856                 const AudioDevice *device = &video_cards[card_index];
 857                 DeviceInfo info;
 858                 info.display_name = device->display_name;
 859                 info.num_channels = 8;
 860                 devices.insert(make_pair(spec, info));
 861         }
 862         vector<ALSAPool::Device> available_alsa_devices = alsa_pool.get_devices();
 863         for (unsigned card_index = 0; card_index < available_alsa_devices.size(); ++card_index) {
 864                 const DeviceSpec spec{ InputSourceType::ALSA_INPUT, card_index };
 865                 const ALSAPool::Device &device = available_alsa_devices[card_index];
 866                 DeviceInfo info;
 867                 info.display_name = device.display_name();
 868                 info.num_channels = device.num_channels;
 869                 info.alsa_name = device.name;
 870                 info.alsa_info = device.info;
 871                 info.alsa_address = device.address;
 872                 devices.insert(make_pair(spec, info));
 873         }
 874         return devices;
 875 }
 876
 877 void AudioMixer::set_display_name(DeviceSpec device_spec, const string &name)
 878 {
 879         AudioDevice *device = find_audio_device(device_spec);
 880
 881         lock_guard<timed_mutex> lock(audio_mutex);
 882         device->display_name = name;
 883 }
 884
 885 void AudioMixer::serialize_device(DeviceSpec device_spec, DeviceSpecProto *device_spec_proto)
 886 {
 887         lock_guard<timed_mutex> lock(audio_mutex);
 888         switch (device_spec.type) {
 889                 case InputSourceType::SILENCE:
 890                         device_spec_proto->set_type(DeviceSpecProto::SILENCE);
 891                         break;
 892                 case InputSourceType::CAPTURE_CARD:
 893                         device_spec_proto->set_type(DeviceSpecProto::CAPTURE_CARD);
 894                         device_spec_proto->set_index(device_spec.index);
 895                         device_spec_proto->set_display_name(video_cards[device_spec.index].display_name);
 896                         break;
 897                 case InputSourceType::ALSA_INPUT:
 898                         alsa_pool.serialize_device(device_spec.index, device_spec_proto);
 899                         break;
 900         }
 901 }
 902
 903 void AudioMixer::set_simple_input(unsigned card_index)
 904 {
 905         InputMapping new_input_mapping;
 906         InputMapping::Bus input;
 907         input.name = "Main";
 908         input.device.type = InputSourceType::CAPTURE_CARD;
 909         input.device.index = card_index;
 910         input.source_channel[0] = 0;
 911         input.source_channel[1] = 1;
 912
 913         new_input_mapping.buses.push_back(input);
 914
 915         lock_guard<timed_mutex> lock(audio_mutex);
 916         current_mapping_mode = MappingMode::SIMPLE;
 917         set_input_mapping_lock_held(new_input_mapping);
 918         fader_volume_db[0] = 0.0f;
 919 }
 920
 921 unsigned AudioMixer::get_simple_input() const
 922 {
 923         lock_guard<timed_mutex> lock(audio_mutex);
 924         if (input_mapping.buses.size() == 1 &&
 925             input_mapping.buses[0].device.type == InputSourceType::CAPTURE_CARD &&
 926             input_mapping.buses[0].source_channel[0] == 0 &&
 927             input_mapping.buses[0].source_channel[1] == 1) {
 928                 return input_mapping.buses[0].device.index;
 929         } else {
 930                 return numeric_limits<unsigned>::max();
 931         }
 932 }
 933
 934 void AudioMixer::set_input_mapping(const InputMapping &new_input_mapping)
 935 {
 936         lock_guard<timed_mutex> lock(audio_mutex);
 937         set_input_mapping_lock_held(new_input_mapping);
 938         current_mapping_mode = MappingMode::MULTICHANNEL;
 939 }
 940
 941 AudioMixer::MappingMode AudioMixer::get_mapping_mode() const
 942 {
 943         lock_guard<timed_mutex> lock(audio_mutex);
 944         return current_mapping_mode;
 945 }
 946
 947 void AudioMixer::set_input_mapping_lock_held(const InputMapping &new_input_mapping)
 948 {
 949         map<DeviceSpec, set<unsigned>> interesting_channels;
 950         for (const InputMapping::Bus &bus : new_input_mapping.buses) {
 951                 if (bus.device.type == InputSourceType::CAPTURE_CARD ||
 952                     bus.device.type == InputSourceType::ALSA_INPUT) {
 953                         for (unsigned channel = 0; channel < 2; ++channel) {
 954                                 if (bus.source_channel[channel] != -1) {
 955                                         interesting_channels[bus.device].insert(bus.source_channel[channel]);
 956                                 }
 957                         }
 958                 }
 959         }
 960
 961         // Reset resamplers for all cards that don't have the exact same state as before.
 962         for (unsigned card_index = 0; card_index < MAX_VIDEO_CARDS; ++card_index) {
 963                 const DeviceSpec device_spec{InputSourceType::CAPTURE_CARD, card_index};
 964                 AudioDevice *device = find_audio_device(device_spec);
 965                 if (device->interesting_channels != interesting_channels[device_spec]) {
 966                         device->interesting_channels = interesting_channels[device_spec];
 967                         reset_resampler_mutex_held(device_spec);
 968                 }
 969         }
 970         for (unsigned card_index = 0; card_index < MAX_ALSA_CARDS; ++card_index) {
 971                 const DeviceSpec device_spec{InputSourceType::ALSA_INPUT, card_index};
 972                 AudioDevice *device = find_audio_device(device_spec);
 973                 if (interesting_channels[device_spec].empty()) {
 974                         alsa_pool.release_device(card_index);
 975                 } else {
 976                         alsa_pool.hold_device(card_index);
 977                 }
 978                 if (device->interesting_channels != interesting_channels[device_spec]) {
 979                         device->interesting_channels = interesting_channels[device_spec];
 980                         alsa_pool.reset_device(device_spec.index);
 981                         reset_resampler_mutex_held(device_spec);
 982                 }
 983         }
 984
 985         input_mapping = new_input_mapping;
 986 }
 987
 988 InputMapping AudioMixer::get_input_mapping() const
 989 {
 990         lock_guard<timed_mutex> lock(audio_mutex);
 991         return input_mapping;
 992 }
 993
 994 unsigned AudioMixer::num_buses() const
 995 {
 996         lock_guard<timed_mutex> lock(audio_mutex);
 997         return input_mapping.buses.size();
 998 }
 999
1000 void AudioMixer::reset_peak(unsigned bus_index)
1001 {
1002         lock_guard<timed_mutex> lock(audio_mutex);
1003         for (unsigned channel = 0; channel < 2; ++channel) {
1004                 PeakHistory &history = peak_history[bus_index][channel];
1005                 history.current_level = 0.0f;
1006                 history.historic_peak = 0.0f;
1007                 history.current_peak = 0.0f;
1008                 history.last_peak = 0.0f;
1009                 history.age_seconds = 0.0f;
1010         }
1011 }
1012
1013 AudioMixer *global_audio_mixer = nullptr;