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         vector<float> audio;
 268         audio.resize(num_samples * num_channels);
 269         unsigned channel_index = 0;
 270         for (auto channel_it = device->interesting_channels.cbegin(); channel_it != device->interesting_channels.end(); ++channel_it, ++channel_index) {
 271                 switch (audio_format.bits_per_sample) {
 272                 case 0:
 273                         assert(num_samples == 0);
 274                         break;
 275                 case 16:
 276                         convert_fixed16_to_fp32(&audio[0], channel_index, num_channels, data, *channel_it, audio_format.num_channels, num_samples);
 277                         break;
 278                 case 24:
 279                         convert_fixed24_to_fp32(&audio[0], channel_index, num_channels, data, *channel_it, audio_format.num_channels, num_samples);
 280                         break;
 281                 case 32:
 282                         convert_fixed32_to_fp32(&audio[0], channel_index, num_channels, data, *channel_it, audio_format.num_channels, num_samples);
 283                         break;
 284                 default:
 285                         fprintf(stderr, "Cannot handle audio with %u bits per sample\n", audio_format.bits_per_sample);
 286                         assert(false);
 287                 }
 288         }
 289
 290         // Now add it.
 291         int64_t local_pts = device->next_local_pts;
 292         device->resampling_queue->add_input_samples(local_pts / double(TIMEBASE), audio.data(), num_samples);
 293         device->next_local_pts = local_pts + frame_length;
 294         return true;
 295 }
 296
 297 bool AudioMixer::add_silence(DeviceSpec device_spec, unsigned samples_per_frame, unsigned num_frames, int64_t frame_length)
 298 {
 299         AudioDevice *device = find_audio_device(device_spec);
 300
 301         unique_lock<timed_mutex> lock(audio_mutex, defer_lock);
 302         if (!lock.try_lock_for(chrono::milliseconds(10))) {
 303                 return false;
 304         }
 305         if (device->resampling_queue == nullptr) {
 306                 // No buses use this device; throw it away.
 307                 return true;
 308         }
 309
 310         unsigned num_channels = device->interesting_channels.size();
 311         assert(num_channels > 0);
 312
 313         vector<float> silence(samples_per_frame * num_channels, 0.0f);
 314         for (unsigned i = 0; i < num_frames; ++i) {
 315                 device->resampling_queue->add_input_samples(device->next_local_pts / double(TIMEBASE), silence.data(), samples_per_frame);
 316                 // Note that if the format changed in the meantime, we have
 317                 // no way of detecting that; we just have to assume the frame length
 318                 // is always the same.
 319                 device->next_local_pts += frame_length;
 320         }
 321         return true;
 322 }
 323
 324 AudioMixer::AudioDevice *AudioMixer::find_audio_device(DeviceSpec device)
 325 {
 326         switch (device.type) {
 327         case InputSourceType::CAPTURE_CARD:
 328                 return &video_cards[device.index];
 329         case InputSourceType::ALSA_INPUT:
 330                 return &alsa_inputs[device.index];
 331         case InputSourceType::SILENCE:
 332         default:
 333                 assert(false);
 334         }
 335         return nullptr;
 336 }
 337
 338 // Get a pointer to the given channel from the given device.
 339 // The channel must be picked out earlier and resampled.
 340 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)
 341 {
 342         static float zero = 0.0f;
 343         if (source_channel == -1 || device_spec.type == InputSourceType::SILENCE) {
 344                 *srcptr = &zero;
 345                 *stride = 0;
 346                 return;
 347         }
 348         AudioDevice *device = find_audio_device(device_spec);
 349         assert(device->interesting_channels.count(source_channel) != 0);
 350         unsigned channel_index = 0;
 351         for (int channel : device->interesting_channels) {
 352                 if (channel == source_channel) break;
 353                 ++channel_index;
 354         }
 355         assert(channel_index < device->interesting_channels.size());
 356         const auto it = samples_card.find(device_spec);
 357         assert(it != samples_card.end());
 358         *srcptr = &(it->second)[channel_index];
 359         *stride = device->interesting_channels.size();
 360 }
 361
 362 // TODO: Can be SSSE3-optimized if need be.
 363 void AudioMixer::fill_audio_bus(const map<DeviceSpec, vector<float>> &samples_card, const InputMapping::Bus &bus, unsigned num_samples, float *output)
 364 {
 365         if (bus.device.type == InputSourceType::SILENCE) {
 366                 memset(output, 0, num_samples * sizeof(*output));
 367         } else {
 368                 assert(bus.device.type == InputSourceType::CAPTURE_CARD ||
 369                        bus.device.type == InputSourceType::ALSA_INPUT);
 370                 const float *lsrc, *rsrc;
 371                 unsigned lstride, rstride;
 372                 float *dptr = output;
 373                 find_sample_src_from_device(samples_card, bus.device, bus.source_channel[0], &lsrc, &lstride);
 374                 find_sample_src_from_device(samples_card, bus.device, bus.source_channel[1], &rsrc, &rstride);
 375                 for (unsigned i = 0; i < num_samples; ++i) {
 376                         *dptr++ = *lsrc;
 377                         *dptr++ = *rsrc;
 378                         lsrc += lstride;
 379                         rsrc += rstride;
 380                 }
 381         }
 382 }
 383
 384 vector<float> AudioMixer::get_output(double pts, unsigned num_samples, ResamplingQueue::RateAdjustmentPolicy rate_adjustment_policy)
 385 {
 386         map<DeviceSpec, vector<float>> samples_card;
 387         vector<float> samples_bus;
 388
 389         lock_guard<timed_mutex> lock(audio_mutex);
 390
 391         // Pick out all the interesting channels from all the cards.
 392         // TODO: If the card has been hotswapped, the number of channels
 393         // might have changed; if so, we need to do some sort of remapping
 394         // to silence.
 395         for (const auto &spec_and_info : get_devices_mutex_held()) {
 396                 const DeviceSpec &device_spec = spec_and_info.first;
 397                 AudioDevice *device = find_audio_device(device_spec);
 398                 if (!device->interesting_channels.empty()) {
 399                         samples_card[device_spec].resize(num_samples * device->interesting_channels.size());
 400                         device->resampling_queue->get_output_samples(
 401                                 pts,
 402                                 &samples_card[device_spec][0],
 403                                 num_samples,
 404                                 rate_adjustment_policy);
 405                 }
 406         }
 407
 408         vector<float> samples_out, left, right;
 409         samples_out.resize(num_samples * 2);
 410         samples_bus.resize(num_samples * 2);
 411         for (unsigned bus_index = 0; bus_index < input_mapping.buses.size(); ++bus_index) {
 412                 fill_audio_bus(samples_card, input_mapping.buses[bus_index], num_samples, &samples_bus[0]);
 413
 414                 // Cut away everything under 120 Hz (or whatever the cutoff is);
 415                 // we don't need it for voice, and it will reduce headroom
 416                 // and confuse the compressor. (In particular, any hums at 50 or 60 Hz
 417                 // should be dampened.)
 418                 if (locut_enabled[bus_index]) {
 419                         locut[bus_index].render(samples_bus.data(), samples_bus.size() / 2, locut_cutoff_hz * 2.0 * M_PI / OUTPUT_FREQUENCY, 0.5f);
 420                 }
 421
 422                 {
 423                         lock_guard<mutex> lock(compressor_mutex);
 424
 425                         // Apply a level compressor to get the general level right.
 426                         // Basically, if it's over about -40 dBFS, we squeeze it down to that level
 427                         // (or more precisely, near it, since we don't use infinite ratio),
 428                         // then apply a makeup gain to get it to -14 dBFS. -14 dBFS is, of course,
 429                         // entirely arbitrary, but from practical tests with speech, it seems to
 430                         // put ut around -23 LUFS, so it's a reasonable starting point for later use.
 431                         if (level_compressor_enabled[bus_index]) {
 432                                 float threshold = 0.01f;   // -40 dBFS.
 433                                 float ratio = 20.0f;
 434                                 float attack_time = 0.5f;
 435                                 float release_time = 20.0f;
 436                                 float makeup_gain = from_db(ref_level_dbfs - (-40.0f));  // +26 dB.
 437                                 level_compressor[bus_index]->process(samples_bus.data(), samples_bus.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 438                                 gain_staging_db[bus_index] = to_db(level_compressor[bus_index]->get_attenuation() * makeup_gain);
 439                         } else {
 440                                 // Just apply the gain we already had.
 441                                 float g = from_db(gain_staging_db[bus_index]);
 442                                 for (size_t i = 0; i < samples_bus.size(); ++i) {
 443                                         samples_bus[i] *= g;
 444                                 }
 445                         }
 446
 447 #if 0
 448                         printf("level=%f (%+5.2f dBFS) attenuation=%f (%+5.2f dB) end_result=%+5.2f dB\n",
 449                                 level_compressor.get_level(), to_db(level_compressor.get_level()),
 450                                 level_compressor.get_attenuation(), to_db(level_compressor.get_attenuation()),
 451                                 to_db(level_compressor.get_level() * level_compressor.get_attenuation() * makeup_gain));
 452 #endif
 453
 454                         // The real compressor.
 455                         if (compressor_enabled[bus_index]) {
 456                                 float threshold = from_db(compressor_threshold_dbfs[bus_index]);
 457                                 float ratio = 20.0f;
 458                                 float attack_time = 0.005f;
 459                                 float release_time = 0.040f;
 460                                 float makeup_gain = 2.0f;  // +6 dB.
 461                                 compressor[bus_index]->process(samples_bus.data(), samples_bus.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 462                 //              compressor_att = compressor.get_attenuation();
 463                         }
 464                 }
 465
 466                 // TODO: We should measure post-fader.
 467                 deinterleave_samples(samples_bus, &left, &right);
 468                 measure_bus_levels(bus_index, left, right);
 469
 470                 float volume = from_db(fader_volume_db[bus_index]);
 471                 if (bus_index == 0) {
 472                         for (unsigned i = 0; i < num_samples * 2; ++i) {
 473                                 samples_out[i] = samples_bus[i] * volume;
 474                         }
 475                 } else {
 476                         for (unsigned i = 0; i < num_samples * 2; ++i) {
 477                                 samples_out[i] += samples_bus[i] * volume;
 478                         }
 479                 }
 480         }
 481
 482         {
 483                 lock_guard<mutex> lock(compressor_mutex);
 484
 485                 // Finally a limiter at -4 dB (so, -10 dBFS) to take out the worst peaks only.
 486                 // Note that since ratio is not infinite, we could go slightly higher than this.
 487                 if (limiter_enabled) {
 488                         float threshold = from_db(limiter_threshold_dbfs);
 489                         float ratio = 30.0f;
 490                         float attack_time = 0.0f;  // Instant.
 491                         float release_time = 0.020f;
 492                         float makeup_gain = 1.0f;  // 0 dB.
 493                         limiter.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 494         //              limiter_att = limiter.get_attenuation();
 495                 }
 496
 497         //      printf("limiter=%+5.1f  compressor=%+5.1f\n", to_db(limiter_att), to_db(compressor_att));
 498         }
 499
 500         // At this point, we are most likely close to +0 LU (at least if the
 501         // faders sum to 0 dB and the compressors are on), but all of our
 502         // measurements have been on raw sample values, not R128 values.
 503         // So we have a final makeup gain to get us to +0 LU; the gain
 504         // adjustments required should be relatively small, and also, the
 505         // offset shouldn't change much (only if the type of audio changes
 506         // significantly). Thus, we shoot for updating this value basically
 507         // “whenever we process buffers”, since the R128 calculation isn't exactly
 508         // something we get out per-sample.
 509         //
 510         // Note that there's a feedback loop here, so we choose a very slow filter
 511         // (half-time of 30 seconds).
 512         double target_loudness_factor, alpha;
 513         double loudness_lu = r128.loudness_M() - ref_level_lufs;
 514         double current_makeup_lu = to_db(final_makeup_gain);
 515         target_loudness_factor = final_makeup_gain * from_db(-loudness_lu);
 516
 517         // If we're outside +/- 5 LU uncorrected, we don't count it as
 518         // a normal signal (probably silence) and don't change the
 519         // correction factor; just apply what we already have.
 520         if (fabs(loudness_lu - current_makeup_lu) >= 5.0 || !final_makeup_gain_auto) {
 521                 alpha = 0.0;
 522         } else {
 523                 // Formula adapted from
 524                 // https://en.wikipedia.org/wiki/Low-pass_filter#Simple_infinite_impulse_response_filter.
 525                 const double half_time_s = 30.0;
 526                 const double fc_mul_2pi_delta_t = 1.0 / (half_time_s * OUTPUT_FREQUENCY);
 527                 alpha = fc_mul_2pi_delta_t / (fc_mul_2pi_delta_t + 1.0);
 528         }
 529
 530         {
 531                 lock_guard<mutex> lock(compressor_mutex);
 532                 double m = final_makeup_gain;
 533                 for (size_t i = 0; i < samples_out.size(); i += 2) {
 534                         samples_out[i + 0] *= m;
 535                         samples_out[i + 1] *= m;
 536                         m += (target_loudness_factor - m) * alpha;
 537                 }
 538                 final_makeup_gain = m;
 539         }
 540
 541         update_meters(samples_out);
 542
 543         return samples_out;
 544 }
 545
 546 void AudioMixer::measure_bus_levels(unsigned bus_index, const vector<float> &left, const vector<float> &right)
 547 {
 548         const float *ptrs[] = { left.data(), right.data() };
 549         {
 550                 lock_guard<mutex> lock(audio_measure_mutex);
 551                 bus_r128[bus_index]->process(left.size(), const_cast<float **>(ptrs));
 552         }
 553 }
 554
 555 void AudioMixer::update_meters(const vector<float> &samples)
 556 {
 557         // Upsample 4x to find interpolated peak.
 558         peak_resampler.inp_data = const_cast<float *>(samples.data());
 559         peak_resampler.inp_count = samples.size() / 2;
 560
 561         vector<float> interpolated_samples;
 562         interpolated_samples.resize(samples.size());
 563         {
 564                 lock_guard<mutex> lock(audio_measure_mutex);
 565
 566                 while (peak_resampler.inp_count > 0) {  // About four iterations.
 567                         peak_resampler.out_data = &interpolated_samples[0];
 568                         peak_resampler.out_count = interpolated_samples.size() / 2;
 569                         peak_resampler.process();
 570                         size_t out_stereo_samples = interpolated_samples.size() / 2 - peak_resampler.out_count;
 571                         peak = max<float>(peak, find_peak(interpolated_samples.data(), out_stereo_samples * 2));
 572                         peak_resampler.out_data = nullptr;
 573                 }
 574         }
 575
 576         // Find R128 levels and L/R correlation.
 577         vector<float> left, right;
 578         deinterleave_samples(samples, &left, &right);
 579         float *ptrs[] = { left.data(), right.data() };
 580         {
 581                 lock_guard<mutex> lock(audio_measure_mutex);
 582                 r128.process(left.size(), ptrs);
 583                 correlation.process_samples(samples);
 584         }
 585
 586         send_audio_level_callback();
 587 }
 588
 589 void AudioMixer::reset_meters()
 590 {
 591         lock_guard<mutex> lock(audio_measure_mutex);
 592         peak_resampler.reset();
 593         peak = 0.0f;
 594         r128.reset();
 595         r128.integr_start();
 596         correlation.reset();
 597 }
 598
 599 void AudioMixer::send_audio_level_callback()
 600 {
 601         if (audio_level_callback == nullptr) {
 602                 return;
 603         }
 604
 605         lock_guard<mutex> lock(audio_measure_mutex);
 606         double loudness_s = r128.loudness_S();
 607         double loudness_i = r128.integrated();
 608         double loudness_range_low = r128.range_min();
 609         double loudness_range_high = r128.range_max();
 610
 611         vector<BusLevel> bus_levels;
 612         bus_levels.resize(input_mapping.buses.size());
 613         {
 614                 lock_guard<mutex> lock(compressor_mutex);
 615                 for (unsigned bus_index = 0; bus_index < bus_r128.size(); ++bus_index) {
 616                         bus_levels[bus_index].loudness_lufs = bus_r128[bus_index]->loudness_S();
 617                         bus_levels[bus_index].gain_staging_db = gain_staging_db[bus_index];
 618                         if (compressor_enabled[bus_index]) {
 619                                 bus_levels[bus_index].compressor_attenuation_db = -to_db(compressor[bus_index]->get_attenuation());
 620                         } else {
 621                                 bus_levels[bus_index].compressor_attenuation_db = 0.0;
 622                         }
 623                 }
 624         }
 625
 626         audio_level_callback(loudness_s, to_db(peak), bus_levels,
 627                 loudness_i, loudness_range_low, loudness_range_high,
 628                 to_db(final_makeup_gain),
 629                 correlation.get_correlation());
 630 }
 631
 632 map<DeviceSpec, DeviceInfo> AudioMixer::get_devices() const
 633 {
 634         lock_guard<timed_mutex> lock(audio_mutex);
 635         return get_devices_mutex_held();
 636 }
 637
 638 map<DeviceSpec, DeviceInfo> AudioMixer::get_devices_mutex_held() const
 639 {
 640         map<DeviceSpec, DeviceInfo> devices;
 641         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 642                 const DeviceSpec spec{ InputSourceType::CAPTURE_CARD, card_index };
 643                 const AudioDevice *device = &video_cards[card_index];
 644                 DeviceInfo info;
 645                 info.name = device->name;
 646                 info.num_channels = 8;  // FIXME: This is wrong for fake cards.
 647                 devices.insert(make_pair(spec, info));
 648         }
 649         for (unsigned card_index = 0; card_index < available_alsa_cards.size(); ++card_index) {
 650                 const DeviceSpec spec{ InputSourceType::ALSA_INPUT, card_index };
 651                 const ALSAInput::Device &device = available_alsa_cards[card_index];
 652                 DeviceInfo info;
 653                 info.name = device.name + " (" + device.info + ")";
 654                 info.num_channels = device.num_channels;
 655                 devices.insert(make_pair(spec, info));
 656         }
 657         return devices;
 658 }
 659
 660 void AudioMixer::set_name(DeviceSpec device_spec, const string &name)
 661 {
 662         AudioDevice *device = find_audio_device(device_spec);
 663
 664         lock_guard<timed_mutex> lock(audio_mutex);
 665         device->name = name;
 666 }
 667
 668 void AudioMixer::set_input_mapping(const InputMapping &new_input_mapping)
 669 {
 670         lock_guard<timed_mutex> lock(audio_mutex);
 671
 672         map<DeviceSpec, set<unsigned>> interesting_channels;
 673         for (const InputMapping::Bus &bus : new_input_mapping.buses) {
 674                 if (bus.device.type == InputSourceType::CAPTURE_CARD ||
 675                     bus.device.type == InputSourceType::ALSA_INPUT) {
 676                         for (unsigned channel = 0; channel < 2; ++channel) {
 677                                 if (bus.source_channel[channel] != -1) {
 678                                         interesting_channels[bus.device].insert(bus.source_channel[channel]);
 679                                 }
 680                         }
 681                 }
 682         }
 683
 684         // Reset resamplers for all cards that don't have the exact same state as before.
 685         for (const auto &spec_and_info : get_devices_mutex_held()) {
 686                 const DeviceSpec &device_spec = spec_and_info.first;
 687                 AudioDevice *device = find_audio_device(device_spec);
 688                 if (device->interesting_channels != interesting_channels[device_spec]) {
 689                         device->interesting_channels = interesting_channels[device_spec];
 690                         if (device_spec.type == InputSourceType::ALSA_INPUT) {
 691                                 reset_alsa_mutex_held(device_spec);
 692                         }
 693                         reset_resampler_mutex_held(device_spec);
 694                 }
 695         }
 696
 697         {
 698                 lock_guard<mutex> lock(audio_measure_mutex);
 699                 bus_r128.resize(new_input_mapping.buses.size());
 700                 for (unsigned bus_index = 0; bus_index < bus_r128.size(); ++bus_index) {
 701                         if (bus_r128[bus_index] == nullptr) {
 702                                 bus_r128[bus_index].reset(new Ebu_r128_proc);
 703                         }
 704                         bus_r128[bus_index]->init(2, OUTPUT_FREQUENCY);
 705                 }
 706         }
 707
 708         input_mapping = new_input_mapping;
 709 }
 710
 711 InputMapping AudioMixer::get_input_mapping() const
 712 {
 713         lock_guard<timed_mutex> lock(audio_mutex);
 714         return input_mapping;
 715 }