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

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