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 "timebase.h"
  16
  17 using namespace bmusb;
  18 using namespace std;
  19 using namespace std::placeholders;
  20
  21 namespace {
  22
  23 // TODO: If these prove to be a bottleneck, they can be SSSE3-optimized
  24 // (usually including multiple channels at a time).
  25
  26 void convert_fixed16_to_fp32(float *dst, size_t out_channel, size_t out_num_channels,
  27                              const uint8_t *src, size_t in_channel, size_t in_num_channels,
  28                              size_t num_samples)
  29 {
  30         assert(in_channel < in_num_channels);
  31         assert(out_channel < out_num_channels);
  32         src += in_channel * 2;
  33         dst += out_channel;
  34
  35         for (size_t i = 0; i < num_samples; ++i) {
  36                 int16_t s = le16toh(*(int16_t *)src);
  37                 *dst = s * (1.0f / 32768.0f);
  38
  39                 src += 2 * in_num_channels;
  40                 dst += out_num_channels;
  41         }
  42 }
  43
  44 void convert_fixed24_to_fp32(float *dst, size_t out_channel, size_t out_num_channels,
  45                              const uint8_t *src, size_t in_channel, size_t in_num_channels,
  46                              size_t num_samples)
  47 {
  48         assert(in_channel < in_num_channels);
  49         assert(out_channel < out_num_channels);
  50         src += in_channel * 3;
  51         dst += out_channel;
  52
  53         for (size_t i = 0; i < num_samples; ++i) {
  54                 uint32_t s1 = src[0];
  55                 uint32_t s2 = src[1];
  56                 uint32_t s3 = src[2];
  57                 uint32_t s = s1 | (s1 << 8) | (s2 << 16) | (s3 << 24);
  58                 *dst = int(s) * (1.0f / 2147483648.0f);
  59
  60                 src += 3 * in_num_channels;
  61                 dst += out_num_channels;
  62         }
  63 }
  64
  65 void convert_fixed32_to_fp32(float *dst, size_t out_channel, size_t out_num_channels,
  66                              const uint8_t *src, size_t in_channel, size_t in_num_channels,
  67                              size_t num_samples)
  68 {
  69         assert(in_channel < in_num_channels);
  70         assert(out_channel < out_num_channels);
  71         src += in_channel * 4;
  72         dst += out_channel;
  73
  74         for (size_t i = 0; i < num_samples; ++i) {
  75                 int32_t s = le32toh(*(int32_t *)src);
  76                 *dst = s * (1.0f / 2147483648.0f);
  77
  78                 src += 4 * in_num_channels;
  79                 dst += out_num_channels;
  80         }
  81 }
  82
  83 float find_peak_plain(const float *samples, size_t num_samples) __attribute__((unused));
  84
  85 float find_peak_plain(const float *samples, size_t num_samples)
  86 {
  87         float m = fabs(samples[0]);
  88         for (size_t i = 1; i < num_samples; ++i) {
  89                 m = max(m, fabs(samples[i]));
  90         }
  91         return m;
  92 }
  93
  94 #ifdef __SSE__
  95 static inline float horizontal_max(__m128 m)
  96 {
  97         __m128 tmp = _mm_shuffle_ps(m, m, _MM_SHUFFLE(1, 0, 3, 2));
  98         m = _mm_max_ps(m, tmp);
  99         tmp = _mm_shuffle_ps(m, m, _MM_SHUFFLE(2, 3, 0, 1));
 100         m = _mm_max_ps(m, tmp);
 101         return _mm_cvtss_f32(m);
 102 }
 103
 104 float find_peak(const float *samples, size_t num_samples)
 105 {
 106         const __m128 abs_mask = _mm_castsi128_ps(_mm_set1_epi32(0x7fffffffu));
 107         __m128 m = _mm_setzero_ps();
 108         for (size_t i = 0; i < (num_samples & ~3); i += 4) {
 109                 __m128 x = _mm_loadu_ps(samples + i);
 110                 x = _mm_and_ps(x, abs_mask);
 111                 m = _mm_max_ps(m, x);
 112         }
 113         float result = horizontal_max(m);
 114
 115         for (size_t i = (num_samples & ~3); i < num_samples; ++i) {
 116                 result = max(result, fabs(samples[i]));
 117         }
 118
 119 #if 0
 120         // Self-test. We should be bit-exact the same.
 121         float reference_result = find_peak_plain(samples, num_samples);
 122         if (result != reference_result) {
 123                 fprintf(stderr, "Error: Peak is %f [%f %f %f %f]; should be %f.\n",
 124                         result,
 125                         _mm_cvtss_f32(_mm_shuffle_ps(m, m, _MM_SHUFFLE(0, 0, 0, 0))),
 126                         _mm_cvtss_f32(_mm_shuffle_ps(m, m, _MM_SHUFFLE(1, 1, 1, 1))),
 127                         _mm_cvtss_f32(_mm_shuffle_ps(m, m, _MM_SHUFFLE(2, 2, 2, 2))),
 128                         _mm_cvtss_f32(_mm_shuffle_ps(m, m, _MM_SHUFFLE(3, 3, 3, 3))),
 129                         reference_result);
 130                 abort();
 131         }
 132 #endif
 133         return result;
 134 }
 135 #else
 136 float find_peak(const float *samples, size_t num_samples)
 137 {
 138         return find_peak_plain(samples, num_samples);
 139 }
 140 #endif
 141
 142 void deinterleave_samples(const vector<float> &in, vector<float> *out_l, vector<float> *out_r)
 143 {
 144         size_t num_samples = in.size() / 2;
 145         out_l->resize(num_samples);
 146         out_r->resize(num_samples);
 147
 148         const float *inptr = in.data();
 149         float *lptr = &(*out_l)[0];
 150         float *rptr = &(*out_r)[0];
 151         for (size_t i = 0; i < num_samples; ++i) {
 152                 *lptr++ = *inptr++;
 153                 *rptr++ = *inptr++;
 154         }
 155 }
 156
 157 }  // namespace
 158
 159 AudioMixer::AudioMixer(unsigned num_cards)
 160         : num_cards(num_cards),
 161           limiter(OUTPUT_FREQUENCY),
 162           correlation(OUTPUT_FREQUENCY)
 163 {
 164         for (unsigned bus_index = 0; bus_index < MAX_BUSES; ++bus_index) {
 165                 locut[bus_index].init(FILTER_HPF, 2);
 166                 locut_enabled[bus_index] = global_flags.locut_enabled;
 167                 gain_staging_db[bus_index] = global_flags.initial_gain_staging_db;
 168                 compressor[bus_index].reset(new StereoCompressor(OUTPUT_FREQUENCY));
 169                 compressor_threshold_dbfs[bus_index] = ref_level_dbfs - 12.0f;  // -12 dB.
 170                 compressor_enabled[bus_index] = global_flags.compressor_enabled;
 171                 level_compressor[bus_index].reset(new StereoCompressor(OUTPUT_FREQUENCY));
 172                 level_compressor_enabled[bus_index] = global_flags.gain_staging_auto;
 173         }
 174         set_limiter_enabled(global_flags.limiter_enabled);
 175         set_final_makeup_gain_auto(global_flags.final_makeup_gain_auto);
 176
 177         // Generate a very simple, default input mapping.
 178         InputMapping::Bus input;
 179         input.name = "Main";
 180         input.device.type = InputSourceType::CAPTURE_CARD;
 181         input.device.index = 0;
 182         input.source_channel[0] = 0;
 183         input.source_channel[1] = 1;
 184
 185         InputMapping new_input_mapping;
 186         new_input_mapping.buses.push_back(input);
 187         set_input_mapping(new_input_mapping);
 188
 189         // Look for ALSA cards.
 190         available_alsa_cards = ALSAInput::enumerate_devices();
 191
 192         r128.init(2, OUTPUT_FREQUENCY);
 193         r128.integr_start();
 194
 195         // hlen=16 is pretty low quality, but we use quite a bit of CPU otherwise,
 196         // and there's a limit to how important the peak meter is.
 197         peak_resampler.setup(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY * 4, /*num_channels=*/2, /*hlen=*/16, /*frel=*/1.0);
 198 }
 199
 200 AudioMixer::~AudioMixer()
 201 {
 202         for (unsigned card_index = 0; card_index < available_alsa_cards.size(); ++card_index) {
 203                 const AudioDevice &device = alsa_inputs[card_index];
 204                 if (device.alsa_device != nullptr) {
 205                         device.alsa_device->stop_capture_thread();
 206                 }
 207         }
 208 }
 209
 210
 211 void AudioMixer::reset_resampler(DeviceSpec device_spec)
 212 {
 213         lock_guard<timed_mutex> lock(audio_mutex);
 214         reset_resampler_mutex_held(device_spec);
 215 }
 216
 217 void AudioMixer::reset_resampler_mutex_held(DeviceSpec device_spec)
 218 {
 219         AudioDevice *device = find_audio_device(device_spec);
 220
 221         if (device->interesting_channels.empty()) {
 222                 device->resampling_queue.reset();
 223         } else {
 224                 // TODO: ResamplingQueue should probably take the full device spec.
 225                 // (It's only used for console output, though.)
 226                 device->resampling_queue.reset(new ResamplingQueue(device_spec.index, device->capture_frequency, OUTPUT_FREQUENCY, device->interesting_channels.size()));
 227         }
 228         device->next_local_pts = 0;
 229 }
 230
 231 void AudioMixer::reset_alsa_mutex_held(DeviceSpec device_spec)
 232 {
 233         assert(device_spec.type == InputSourceType::ALSA_INPUT);
 234         unsigned card_index = device_spec.index;
 235         AudioDevice *device = find_audio_device(device_spec);
 236
 237         if (device->alsa_device != nullptr) {
 238                 device->alsa_device->stop_capture_thread();
 239         }
 240         if (device->interesting_channels.empty()) {
 241                 device->alsa_device.reset();
 242         } else {
 243                 const ALSAInput::Device &alsa_dev = available_alsa_cards[card_index];
 244                 device->alsa_device.reset(new ALSAInput(alsa_dev.address.c_str(), OUTPUT_FREQUENCY, alsa_dev.num_channels, bind(&AudioMixer::add_audio, this, device_spec, _1, _2, _3, _4)));
 245                 device->capture_frequency = device->alsa_device->get_sample_rate();
 246                 device->alsa_device->start_capture_thread();
 247         }
 248 }
 249
 250 bool AudioMixer::add_audio(DeviceSpec device_spec, const uint8_t *data, unsigned num_samples, AudioFormat audio_format, int64_t frame_length)
 251 {
 252         AudioDevice *device = find_audio_device(device_spec);
 253
 254         unique_lock<timed_mutex> lock(audio_mutex, defer_lock);
 255         if (!lock.try_lock_for(chrono::milliseconds(10))) {
 256                 return false;
 257         }
 258         if (device->resampling_queue == nullptr) {
 259                 // No buses use this device; throw it away.
 260                 return true;
 261         }
 262
 263         unsigned num_channels = device->interesting_channels.size();
 264         assert(num_channels > 0);
 265
 266         // Convert the audio to fp32.
 267         unique_ptr<float[]> audio(new float[num_samples * num_channels]);
 268         unsigned channel_index = 0;
 269         for (auto channel_it = device->interesting_channels.cbegin(); channel_it != device->interesting_channels.end(); ++channel_it, ++channel_index) {
 270                 switch (audio_format.bits_per_sample) {
 271                 case 0:
 272                         assert(num_samples == 0);
 273                         break;
 274                 case 16:
 275                         convert_fixed16_to_fp32(audio.get(), channel_index, num_channels, data, *channel_it, audio_format.num_channels, num_samples);
 276                         break;
 277                 case 24:
 278                         convert_fixed24_to_fp32(audio.get(), channel_index, num_channels, data, *channel_it, audio_format.num_channels, num_samples);
 279                         break;
 280                 case 32:
 281                         convert_fixed32_to_fp32(audio.get(), channel_index, num_channels, data, *channel_it, audio_format.num_channels, num_samples);
 282                         break;
 283                 default:
 284                         fprintf(stderr, "Cannot handle audio with %u bits per sample\n", audio_format.bits_per_sample);
 285                         assert(false);
 286                 }
 287         }
 288
 289         // Now add it.
 290         int64_t local_pts = device->next_local_pts;
 291         device->resampling_queue->add_input_samples(local_pts / double(TIMEBASE), audio.get(), num_samples);
 292         device->next_local_pts = local_pts + frame_length;
 293         return true;
 294 }
 295
 296 bool AudioMixer::add_silence(DeviceSpec device_spec, unsigned samples_per_frame, unsigned num_frames, int64_t frame_length)
 297 {
 298         AudioDevice *device = find_audio_device(device_spec);
 299
 300         unique_lock<timed_mutex> lock(audio_mutex, defer_lock);
 301         if (!lock.try_lock_for(chrono::milliseconds(10))) {
 302                 return false;
 303         }
 304         if (device->resampling_queue == nullptr) {
 305                 // No buses use this device; throw it away.
 306                 return true;
 307         }
 308
 309         unsigned num_channels = device->interesting_channels.size();
 310         assert(num_channels > 0);
 311
 312         vector<float> silence(samples_per_frame * num_channels, 0.0f);
 313         for (unsigned i = 0; i < num_frames; ++i) {
 314                 device->resampling_queue->add_input_samples(device->next_local_pts / double(TIMEBASE), silence.data(), samples_per_frame);
 315                 // Note that if the format changed in the meantime, we have
 316                 // no way of detecting that; we just have to assume the frame length
 317                 // is always the same.
 318                 device->next_local_pts += frame_length;
 319         }
 320         return true;
 321 }
 322
 323 AudioMixer::AudioDevice *AudioMixer::find_audio_device(DeviceSpec device)
 324 {
 325         switch (device.type) {
 326         case InputSourceType::CAPTURE_CARD:
 327                 return &video_cards[device.index];
 328         case InputSourceType::ALSA_INPUT:
 329                 return &alsa_inputs[device.index];
 330         case InputSourceType::SILENCE:
 331         default:
 332                 assert(false);
 333         }
 334         return nullptr;
 335 }
 336
 337 // Get a pointer to the given channel from the given device.
 338 // The channel must be picked out earlier and resampled.
 339 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)
 340 {
 341         static float zero = 0.0f;
 342         if (source_channel == -1 || device_spec.type == InputSourceType::SILENCE) {
 343                 *srcptr = &zero;
 344                 *stride = 0;
 345                 return;
 346         }
 347         AudioDevice *device = find_audio_device(device_spec);
 348         assert(device->interesting_channels.count(source_channel) != 0);
 349         unsigned channel_index = 0;
 350         for (int channel : device->interesting_channels) {
 351                 if (channel == source_channel) break;
 352                 ++channel_index;
 353         }
 354         assert(channel_index < device->interesting_channels.size());
 355         const auto it = samples_card.find(device_spec);
 356         assert(it != samples_card.end());
 357         *srcptr = &(it->second)[channel_index];
 358         *stride = device->interesting_channels.size();
 359 }
 360
 361 // TODO: Can be SSSE3-optimized if need be.
 362 void AudioMixer::fill_audio_bus(const map<DeviceSpec, vector<float>> &samples_card, const InputMapping::Bus &bus, unsigned num_samples, float *output)
 363 {
 364         if (bus.device.type == InputSourceType::SILENCE) {
 365                 memset(output, 0, num_samples * sizeof(*output));
 366         } else {
 367                 assert(bus.device.type == InputSourceType::CAPTURE_CARD ||
 368                        bus.device.type == InputSourceType::ALSA_INPUT);
 369                 const float *lsrc, *rsrc;
 370                 unsigned lstride, rstride;
 371                 float *dptr = output;
 372                 find_sample_src_from_device(samples_card, bus.device, bus.source_channel[0], &lsrc, &lstride);
 373                 find_sample_src_from_device(samples_card, bus.device, bus.source_channel[1], &rsrc, &rstride);
 374                 for (unsigned i = 0; i < num_samples; ++i) {
 375                         *dptr++ = *lsrc;
 376                         *dptr++ = *rsrc;
 377                         lsrc += lstride;
 378                         rsrc += rstride;
 379                 }
 380         }
 381 }
 382
 383 vector<float> AudioMixer::get_output(double pts, unsigned num_samples, ResamplingQueue::RateAdjustmentPolicy rate_adjustment_policy)
 384 {
 385         map<DeviceSpec, vector<float>> samples_card;
 386         vector<float> samples_bus;
 387
 388         lock_guard<timed_mutex> lock(audio_mutex);
 389
 390         // Pick out all the interesting channels from all the cards.
 391         // TODO: If the card has been hotswapped, the number of channels
 392         // might have changed; if so, we need to do some sort of remapping
 393         // to silence.
 394         for (const auto &spec_and_info : get_devices_mutex_held()) {
 395                 const DeviceSpec &device_spec = spec_and_info.first;
 396                 AudioDevice *device = find_audio_device(device_spec);
 397                 if (!device->interesting_channels.empty()) {
 398                         samples_card[device_spec].resize(num_samples * device->interesting_channels.size());
 399                         device->resampling_queue->get_output_samples(
 400                                 pts,
 401                                 &samples_card[device_spec][0],
 402                                 num_samples,
 403                                 rate_adjustment_policy);
 404                 }
 405         }
 406
 407         vector<float> samples_out, left, right;
 408         samples_out.resize(num_samples * 2);
 409         samples_bus.resize(num_samples * 2);
 410         for (unsigned bus_index = 0; bus_index < input_mapping.buses.size(); ++bus_index) {
 411                 fill_audio_bus(samples_card, input_mapping.buses[bus_index], num_samples, &samples_bus[0]);
 412
 413                 // Cut away everything under 120 Hz (or whatever the cutoff is);
 414                 // we don't need it for voice, and it will reduce headroom
 415                 // and confuse the compressor. (In particular, any hums at 50 or 60 Hz
 416                 // should be dampened.)
 417                 if (locut_enabled[bus_index]) {
 418                         locut[bus_index].render(samples_bus.data(), samples_bus.size() / 2, locut_cutoff_hz * 2.0 * M_PI / OUTPUT_FREQUENCY, 0.5f);
 419                 }
 420
 421                 {
 422                         lock_guard<mutex> lock(compressor_mutex);
 423
 424                         // Apply a level compressor to get the general level right.
 425                         // Basically, if it's over about -40 dBFS, we squeeze it down to that level
 426                         // (or more precisely, near it, since we don't use infinite ratio),
 427                         // then apply a makeup gain to get it to -14 dBFS. -14 dBFS is, of course,
 428                         // entirely arbitrary, but from practical tests with speech, it seems to
 429                         // put ut around -23 LUFS, so it's a reasonable starting point for later use.
 430                         if (level_compressor_enabled[bus_index]) {
 431                                 float threshold = 0.01f;   // -40 dBFS.
 432                                 float ratio = 20.0f;
 433                                 float attack_time = 0.5f;
 434                                 float release_time = 20.0f;
 435                                 float makeup_gain = from_db(ref_level_dbfs - (-40.0f));  // +26 dB.
 436                                 level_compressor[bus_index]->process(samples_bus.data(), samples_bus.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 437                                 gain_staging_db[bus_index] = to_db(level_compressor[bus_index]->get_attenuation() * makeup_gain);
 438                         } else {
 439                                 // Just apply the gain we already had.
 440                                 float g = from_db(gain_staging_db[bus_index]);
 441                                 for (size_t i = 0; i < samples_bus.size(); ++i) {
 442                                         samples_bus[i] *= g;
 443                                 }
 444                         }
 445
 446 #if 0
 447                         printf("level=%f (%+5.2f dBFS) attenuation=%f (%+5.2f dB) end_result=%+5.2f dB\n",
 448                                 level_compressor.get_level(), to_db(level_compressor.get_level()),
 449                                 level_compressor.get_attenuation(), to_db(level_compressor.get_attenuation()),
 450                                 to_db(level_compressor.get_level() * level_compressor.get_attenuation() * makeup_gain));
 451 #endif
 452
 453                         // The real compressor.
 454                         if (compressor_enabled[bus_index]) {
 455                                 float threshold = from_db(compressor_threshold_dbfs[bus_index]);
 456                                 float ratio = 20.0f;
 457                                 float attack_time = 0.005f;
 458                                 float release_time = 0.040f;
 459                                 float makeup_gain = 2.0f;  // +6 dB.
 460                                 compressor[bus_index]->process(samples_bus.data(), samples_bus.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 461                 //              compressor_att = compressor.get_attenuation();
 462                         }
 463                 }
 464
 465                 add_bus_to_master(bus_index, samples_bus, &samples_out);
 466                 deinterleave_samples(samples_bus, &left, &right);
 467                 measure_bus_levels(bus_index, left, right);
 468         }
 469
 470         {
 471                 lock_guard<mutex> lock(compressor_mutex);
 472
 473                 // Finally a limiter at -4 dB (so, -10 dBFS) to take out the worst peaks only.
 474                 // Note that since ratio is not infinite, we could go slightly higher than this.
 475                 if (limiter_enabled) {
 476                         float threshold = from_db(limiter_threshold_dbfs);
 477                         float ratio = 30.0f;
 478                         float attack_time = 0.0f;  // Instant.
 479                         float release_time = 0.020f;
 480                         float makeup_gain = 1.0f;  // 0 dB.
 481                         limiter.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 482         //              limiter_att = limiter.get_attenuation();
 483                 }
 484
 485         //      printf("limiter=%+5.1f  compressor=%+5.1f\n", to_db(limiter_att), to_db(compressor_att));
 486         }
 487
 488         // At this point, we are most likely close to +0 LU (at least if the
 489         // faders sum to 0 dB and the compressors are on), but all of our
 490         // measurements have been on raw sample values, not R128 values.
 491         // So we have a final makeup gain to get us to +0 LU; the gain
 492         // adjustments required should be relatively small, and also, the
 493         // offset shouldn't change much (only if the type of audio changes
 494         // significantly). Thus, we shoot for updating this value basically
 495         // “whenever we process buffers”, since the R128 calculation isn't exactly
 496         // something we get out per-sample.
 497         //
 498         // Note that there's a feedback loop here, so we choose a very slow filter
 499         // (half-time of 30 seconds).
 500         double target_loudness_factor, alpha;
 501         double loudness_lu = r128.loudness_M() - ref_level_lufs;
 502         double current_makeup_lu = to_db(final_makeup_gain);
 503         target_loudness_factor = final_makeup_gain * from_db(-loudness_lu);
 504
 505         // If we're outside +/- 5 LU uncorrected, we don't count it as
 506         // a normal signal (probably silence) and don't change the
 507         // correction factor; just apply what we already have.
 508         if (fabs(loudness_lu - current_makeup_lu) >= 5.0 || !final_makeup_gain_auto) {
 509                 alpha = 0.0;
 510         } else {
 511                 // Formula adapted from
 512                 // https://en.wikipedia.org/wiki/Low-pass_filter#Simple_infinite_impulse_response_filter.
 513                 const double half_time_s = 30.0;
 514                 const double fc_mul_2pi_delta_t = 1.0 / (half_time_s * OUTPUT_FREQUENCY);
 515                 alpha = fc_mul_2pi_delta_t / (fc_mul_2pi_delta_t + 1.0);
 516         }
 517
 518         {
 519                 lock_guard<mutex> lock(compressor_mutex);
 520                 double m = final_makeup_gain;
 521                 for (size_t i = 0; i < samples_out.size(); i += 2) {
 522                         samples_out[i + 0] *= m;
 523                         samples_out[i + 1] *= m;
 524                         m += (target_loudness_factor - m) * alpha;
 525                 }
 526                 final_makeup_gain = m;
 527         }
 528
 529         update_meters(samples_out);
 530
 531         return samples_out;
 532 }
 533
 534 void AudioMixer::add_bus_to_master(unsigned bus_index, const vector<float> &samples_bus, vector<float> *samples_out)
 535 {
 536         assert(samples_bus.size() == samples_out->size());
 537         assert(samples_bus.size() % 2 == 0);
 538         unsigned num_samples = samples_bus.size() / 2;
 539         if (fabs(fader_volume_db[bus_index] - last_fader_volume_db[bus_index]) > 1e-3) {
 540                 // The volume has changed; do a fade over the course of this frame.
 541                 // (We might have some numerical issues here, but it seems to sound OK.)
 542                 // For the purpose of fading here, the silence floor is set to -90 dB
 543                 // (the fader only goes to -84).
 544                 float old_volume = from_db(max<float>(last_fader_volume_db[bus_index], -90.0f));
 545                 float volume = from_db(max<float>(fader_volume_db[bus_index], -90.0f));
 546
 547                 float volume_inc = pow(volume / old_volume, 1.0 / num_samples);
 548                 volume = old_volume;
 549                 if (bus_index == 0) {
 550                         for (unsigned i = 0; i < num_samples; ++i) {
 551                                 (*samples_out)[i * 2 + 0] = samples_bus[i * 2 + 0] * volume;
 552                                 (*samples_out)[i * 2 + 1] = samples_bus[i * 2 + 1] * volume;
 553                                 volume *= volume_inc;
 554                         }
 555                 } else {
 556                         for (unsigned i = 0; i < num_samples; ++i) {
 557                                 (*samples_out)[i * 2 + 0] += samples_bus[i * 2 + 0] * volume;
 558                                 (*samples_out)[i * 2 + 1] += samples_bus[i * 2 + 1] * volume;
 559                                 volume *= volume_inc;
 560                         }
 561                 }
 562         } else {
 563                 float volume = from_db(fader_volume_db[bus_index]);
 564                 if (bus_index == 0) {
 565                         for (unsigned i = 0; i < num_samples; ++i) {
 566                                 (*samples_out)[i * 2 + 0] = samples_bus[i * 2 + 0] * volume;
 567                                 (*samples_out)[i * 2 + 1] = samples_bus[i * 2 + 1] * volume;
 568                         }
 569                 } else {
 570                         for (unsigned i = 0; i < num_samples; ++i) {
 571                                 (*samples_out)[i * 2 + 0] += samples_bus[i * 2 + 0] * volume;
 572                                 (*samples_out)[i * 2 + 1] += samples_bus[i * 2 + 1] * volume;
 573                         }
 574                 }
 575         }
 576
 577         last_fader_volume_db[bus_index] = fader_volume_db[bus_index];
 578 }
 579
 580 void AudioMixer::measure_bus_levels(unsigned bus_index, const vector<float> &left, const vector<float> &right)
 581 {
 582         assert(left.size() == right.size());
 583         const float volume = from_db(fader_volume_db[bus_index]);
 584         const float peak_levels[2] = {
 585                 find_peak(left.data(), left.size()) * volume,
 586                 find_peak(right.data(), right.size()) * volume
 587         };
 588         for (unsigned channel = 0; channel < 2; ++channel) {
 589                 // Compute the current value, including hold and falloff.
 590                 // The constants are borrowed from zita-mu1 by Fons Adriaensen.
 591                 static constexpr float hold_sec = 0.5f;
 592                 static constexpr float falloff_db_sec = 15.0f;  // dB/sec falloff after hold.
 593                 float current_peak;
 594                 PeakHistory &history = peak_history[bus_index][channel];
 595                 history.historic_peak = max(history.historic_peak, peak_levels[channel]);
 596                 if (history.age_seconds < hold_sec) {
 597                         current_peak = history.last_peak;
 598                 } else {
 599                         current_peak = history.last_peak * from_db(-falloff_db_sec * (history.age_seconds - hold_sec));
 600                 }
 601
 602                 // See if we have a new peak to replace the old (possibly falling) one.
 603                 if (peak_levels[channel] > current_peak) {
 604                         history.last_peak = peak_levels[channel];
 605                         history.age_seconds = 0.0f;  // Not 100% correct, but more than good enough given our frame sizes.
 606                         current_peak = peak_levels[channel];
 607                 } else {
 608                         history.age_seconds += float(left.size()) / OUTPUT_FREQUENCY;
 609                 }
 610                 history.current_level = peak_levels[channel];
 611                 history.current_peak = current_peak;
 612         }
 613 }
 614
 615 void AudioMixer::update_meters(const vector<float> &samples)
 616 {
 617         // Upsample 4x to find interpolated peak.
 618         peak_resampler.inp_data = const_cast<float *>(samples.data());
 619         peak_resampler.inp_count = samples.size() / 2;
 620
 621         vector<float> interpolated_samples;
 622         interpolated_samples.resize(samples.size());
 623         {
 624                 lock_guard<mutex> lock(audio_measure_mutex);
 625
 626                 while (peak_resampler.inp_count > 0) {  // About four iterations.
 627                         peak_resampler.out_data = &interpolated_samples[0];
 628                         peak_resampler.out_count = interpolated_samples.size() / 2;
 629                         peak_resampler.process();
 630                         size_t out_stereo_samples = interpolated_samples.size() / 2 - peak_resampler.out_count;
 631                         peak = max<float>(peak, find_peak(interpolated_samples.data(), out_stereo_samples * 2));
 632                         peak_resampler.out_data = nullptr;
 633                 }
 634         }
 635
 636         // Find R128 levels and L/R correlation.
 637         vector<float> left, right;
 638         deinterleave_samples(samples, &left, &right);
 639         float *ptrs[] = { left.data(), right.data() };
 640         {
 641                 lock_guard<mutex> lock(audio_measure_mutex);
 642                 r128.process(left.size(), ptrs);
 643                 correlation.process_samples(samples);
 644         }
 645
 646         send_audio_level_callback();
 647 }
 648
 649 void AudioMixer::reset_meters()
 650 {
 651         lock_guard<mutex> lock(audio_measure_mutex);
 652         peak_resampler.reset();
 653         peak = 0.0f;
 654         r128.reset();
 655         r128.integr_start();
 656         correlation.reset();
 657 }
 658
 659 void AudioMixer::send_audio_level_callback()
 660 {
 661         if (audio_level_callback == nullptr) {
 662                 return;
 663         }
 664
 665         lock_guard<mutex> lock(audio_measure_mutex);
 666         double loudness_s = r128.loudness_S();
 667         double loudness_i = r128.integrated();
 668         double loudness_range_low = r128.range_min();
 669         double loudness_range_high = r128.range_max();
 670
 671         vector<BusLevel> bus_levels;
 672         bus_levels.resize(input_mapping.buses.size());
 673         {
 674                 lock_guard<mutex> lock(compressor_mutex);
 675                 for (unsigned bus_index = 0; bus_index < bus_levels.size(); ++bus_index) {
 676                         bus_levels[bus_index].current_level_dbfs[0] = to_db(peak_history[bus_index][0].current_level);
 677                         bus_levels[bus_index].current_level_dbfs[1] = to_db(peak_history[bus_index][1].current_level);
 678                         bus_levels[bus_index].peak_level_dbfs[0] = to_db(peak_history[bus_index][0].current_peak);
 679                         bus_levels[bus_index].peak_level_dbfs[1] = to_db(peak_history[bus_index][1].current_peak);
 680                         bus_levels[bus_index].historic_peak_dbfs = to_db(
 681                                 max(peak_history[bus_index][0].historic_peak,
 682                                     peak_history[bus_index][1].historic_peak));
 683                         bus_levels[bus_index].gain_staging_db = gain_staging_db[bus_index];
 684                         if (compressor_enabled[bus_index]) {
 685                                 bus_levels[bus_index].compressor_attenuation_db = -to_db(compressor[bus_index]->get_attenuation());
 686                         } else {
 687                                 bus_levels[bus_index].compressor_attenuation_db = 0.0;
 688                         }
 689                 }
 690         }
 691
 692         audio_level_callback(loudness_s, to_db(peak), bus_levels,
 693                 loudness_i, loudness_range_low, loudness_range_high,
 694                 to_db(final_makeup_gain),
 695                 correlation.get_correlation());
 696 }
 697
 698 map<DeviceSpec, DeviceInfo> AudioMixer::get_devices() const
 699 {
 700         lock_guard<timed_mutex> lock(audio_mutex);
 701         return get_devices_mutex_held();
 702 }
 703
 704 map<DeviceSpec, DeviceInfo> AudioMixer::get_devices_mutex_held() const
 705 {
 706         map<DeviceSpec, DeviceInfo> devices;
 707         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 708                 const DeviceSpec spec{ InputSourceType::CAPTURE_CARD, card_index };
 709                 const AudioDevice *device = &video_cards[card_index];
 710                 DeviceInfo info;
 711                 info.name = device->name;
 712                 info.num_channels = 8;  // FIXME: This is wrong for fake cards.
 713                 devices.insert(make_pair(spec, info));
 714         }
 715         for (unsigned card_index = 0; card_index < available_alsa_cards.size(); ++card_index) {
 716                 const DeviceSpec spec{ InputSourceType::ALSA_INPUT, card_index };
 717                 const ALSAInput::Device &device = available_alsa_cards[card_index];
 718                 DeviceInfo info;
 719                 info.name = device.name + " (" + device.info + ")";
 720                 info.num_channels = device.num_channels;
 721                 devices.insert(make_pair(spec, info));
 722         }
 723         return devices;
 724 }
 725
 726 void AudioMixer::set_name(DeviceSpec device_spec, const string &name)
 727 {
 728         AudioDevice *device = find_audio_device(device_spec);
 729
 730         lock_guard<timed_mutex> lock(audio_mutex);
 731         device->name = name;
 732 }
 733
 734 void AudioMixer::set_input_mapping(const InputMapping &new_input_mapping)
 735 {
 736         lock_guard<timed_mutex> lock(audio_mutex);
 737
 738         map<DeviceSpec, set<unsigned>> interesting_channels;
 739         for (const InputMapping::Bus &bus : new_input_mapping.buses) {
 740                 if (bus.device.type == InputSourceType::CAPTURE_CARD ||
 741                     bus.device.type == InputSourceType::ALSA_INPUT) {
 742                         for (unsigned channel = 0; channel < 2; ++channel) {
 743                                 if (bus.source_channel[channel] != -1) {
 744                                         interesting_channels[bus.device].insert(bus.source_channel[channel]);
 745                                 }
 746                         }
 747                 }
 748         }
 749
 750         // Reset resamplers for all cards that don't have the exact same state as before.
 751         for (const auto &spec_and_info : get_devices_mutex_held()) {
 752                 const DeviceSpec &device_spec = spec_and_info.first;
 753                 AudioDevice *device = find_audio_device(device_spec);
 754                 if (device->interesting_channels != interesting_channels[device_spec]) {
 755                         device->interesting_channels = interesting_channels[device_spec];
 756                         if (device_spec.type == InputSourceType::ALSA_INPUT) {
 757                                 reset_alsa_mutex_held(device_spec);
 758                         }
 759                         reset_resampler_mutex_held(device_spec);
 760                 }
 761         }
 762
 763         input_mapping = new_input_mapping;
 764 }
 765
 766 InputMapping AudioMixer::get_input_mapping() const
 767 {
 768         lock_guard<timed_mutex> lock(audio_mutex);
 769         return input_mapping;
 770 }
 771
 772 void AudioMixer::reset_peak(unsigned bus_index)
 773 {
 774         lock_guard<timed_mutex> lock(audio_mutex);
 775         for (unsigned channel = 0; channel < 2; ++channel) {
 776                 PeakHistory &history = peak_history[bus_index][channel];
 777                 history.current_level = 0.0f;
 778                 history.historic_peak = 0.0f;
 779                 history.current_peak = 0.0f;
 780                 history.last_peak = 0.0f;
 781                 history.age_seconds = 0.0f;
 782         }
 783 }