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
  10 #include "db.h"
  11 #include "flags.h"
  12 #include "timebase.h"
  13
  14 using namespace bmusb;
  15 using namespace std;
  16 using namespace std::placeholders;
  17
  18 namespace {
  19
  20 // TODO: If these prove to be a bottleneck, they can be SSSE3-optimized
  21 // (usually including multiple channels at a time).
  22
  23 void convert_fixed16_to_fp32(float *dst, size_t out_channel, size_t out_num_channels,
  24                              const uint8_t *src, size_t in_channel, size_t in_num_channels,
  25                              size_t num_samples)
  26 {
  27         assert(in_channel < in_num_channels);
  28         assert(out_channel < out_num_channels);
  29         src += in_channel * 2;
  30         dst += out_channel;
  31
  32         for (size_t i = 0; i < num_samples; ++i) {
  33                 int16_t s = le16toh(*(int16_t *)src);
  34                 *dst = s * (1.0f / 32768.0f);
  35
  36                 src += 2 * in_num_channels;
  37                 dst += out_num_channels;
  38         }
  39 }
  40
  41 void convert_fixed24_to_fp32(float *dst, size_t out_channel, size_t out_num_channels,
  42                              const uint8_t *src, size_t in_channel, size_t in_num_channels,
  43                              size_t num_samples)
  44 {
  45         assert(in_channel < in_num_channels);
  46         assert(out_channel < out_num_channels);
  47         src += in_channel * 3;
  48         dst += out_channel;
  49
  50         for (size_t i = 0; i < num_samples; ++i) {
  51                 uint32_t s1 = src[0];
  52                 uint32_t s2 = src[1];
  53                 uint32_t s3 = src[2];
  54                 uint32_t s = s1 | (s1 << 8) | (s2 << 16) | (s3 << 24);
  55                 *dst = int(s) * (1.0f / 2147483648.0f);
  56
  57                 src += 3 * in_num_channels;
  58                 dst += out_num_channels;
  59         }
  60 }
  61
  62 void convert_fixed32_to_fp32(float *dst, size_t out_channel, size_t out_num_channels,
  63                              const uint8_t *src, size_t in_channel, size_t in_num_channels,
  64                              size_t num_samples)
  65 {
  66         assert(in_channel < in_num_channels);
  67         assert(out_channel < out_num_channels);
  68         src += in_channel * 4;
  69         dst += out_channel;
  70
  71         for (size_t i = 0; i < num_samples; ++i) {
  72                 int32_t s = le32toh(*(int32_t *)src);
  73                 *dst = s * (1.0f / 2147483648.0f);
  74
  75                 src += 4 * in_num_channels;
  76                 dst += out_num_channels;
  77         }
  78 }
  79
  80 }  // namespace
  81
  82 AudioMixer::AudioMixer(unsigned num_cards)
  83         : num_cards(num_cards),
  84           level_compressor(OUTPUT_FREQUENCY),
  85           limiter(OUTPUT_FREQUENCY),
  86           compressor(OUTPUT_FREQUENCY)
  87 {
  88         locut.init(FILTER_HPF, 2);
  89
  90         set_locut_enabled(global_flags.locut_enabled);
  91         set_gain_staging_db(global_flags.initial_gain_staging_db);
  92         set_gain_staging_auto(global_flags.gain_staging_auto);
  93         set_compressor_enabled(global_flags.compressor_enabled);
  94         set_limiter_enabled(global_flags.limiter_enabled);
  95         set_final_makeup_gain_auto(global_flags.final_makeup_gain_auto);
  96
  97         // Generate a very simple, default input mapping.
  98         InputMapping::Bus input;
  99         input.name = "Main";
 100         input.device.type = InputSourceType::CAPTURE_CARD;
 101         input.device.index = 0;
 102         input.source_channel[0] = 0;
 103         input.source_channel[1] = 1;
 104
 105         InputMapping new_input_mapping;
 106         new_input_mapping.buses.push_back(input);
 107         set_input_mapping(new_input_mapping);
 108
 109         // Look for ALSA cards.
 110         available_alsa_cards = ALSAInput::enumerate_devices();
 111 }
 112
 113 AudioMixer::~AudioMixer()
 114 {
 115         for (unsigned card_index = 0; card_index < available_alsa_cards.size(); ++card_index) {
 116                 const AudioDevice &device = alsa_inputs[card_index];
 117                 if (device.alsa_device != nullptr) {
 118                         device.alsa_device->stop_capture_thread();
 119                 }
 120         }
 121 }
 122
 123
 124 void AudioMixer::reset_resampler(DeviceSpec device_spec)
 125 {
 126         lock_guard<timed_mutex> lock(audio_mutex);
 127         reset_resampler_mutex_held(device_spec);
 128 }
 129
 130 void AudioMixer::reset_resampler_mutex_held(DeviceSpec device_spec)
 131 {
 132         AudioDevice *device = find_audio_device(device_spec);
 133
 134         if (device->interesting_channels.empty()) {
 135                 device->resampling_queue.reset();
 136         } else {
 137                 // TODO: ResamplingQueue should probably take the full device spec.
 138                 // (It's only used for console output, though.)
 139                 device->resampling_queue.reset(new ResamplingQueue(device_spec.index, device->capture_frequency, OUTPUT_FREQUENCY, device->interesting_channels.size()));
 140         }
 141         device->next_local_pts = 0;
 142 }
 143
 144 void AudioMixer::reset_alsa_mutex_held(DeviceSpec device_spec)
 145 {
 146         assert(device_spec.type == InputSourceType::ALSA_INPUT);
 147         unsigned card_index = device_spec.index;
 148         AudioDevice *device = find_audio_device(device_spec);
 149
 150         if (device->alsa_device != nullptr) {
 151                 device->alsa_device->stop_capture_thread();
 152         }
 153         if (device->interesting_channels.empty()) {
 154                 device->alsa_device.reset();
 155         } else {
 156                 device->alsa_device.reset(new ALSAInput(available_alsa_cards[card_index].address.c_str(), OUTPUT_FREQUENCY, 2, bind(&AudioMixer::add_audio, this, device_spec, _1, _2, _3, _4)));
 157                 device->capture_frequency = device->alsa_device->get_sample_rate();
 158                 device->alsa_device->start_capture_thread();
 159         }
 160 }
 161
 162 bool AudioMixer::add_audio(DeviceSpec device_spec, const uint8_t *data, unsigned num_samples, AudioFormat audio_format, int64_t frame_length)
 163 {
 164         AudioDevice *device = find_audio_device(device_spec);
 165
 166         unique_lock<timed_mutex> lock(audio_mutex, defer_lock);
 167         if (!lock.try_lock_for(chrono::milliseconds(10))) {
 168                 return false;
 169         }
 170         if (device->resampling_queue == nullptr) {
 171                 // No buses use this device; throw it away.
 172                 return true;
 173         }
 174
 175         unsigned num_channels = device->interesting_channels.size();
 176         assert(num_channels > 0);
 177
 178         // Convert the audio to fp32.
 179         vector<float> audio;
 180         audio.resize(num_samples * num_channels);
 181         unsigned channel_index = 0;
 182         for (auto channel_it = device->interesting_channels.cbegin(); channel_it != device->interesting_channels.end(); ++channel_it, ++channel_index) {
 183                 switch (audio_format.bits_per_sample) {
 184                 case 0:
 185                         assert(num_samples == 0);
 186                         break;
 187                 case 16:
 188                         convert_fixed16_to_fp32(&audio[0], channel_index, num_channels, data, *channel_it, audio_format.num_channels, num_samples);
 189                         break;
 190                 case 24:
 191                         convert_fixed24_to_fp32(&audio[0], channel_index, num_channels, data, *channel_it, audio_format.num_channels, num_samples);
 192                         break;
 193                 case 32:
 194                         convert_fixed32_to_fp32(&audio[0], channel_index, num_channels, data, *channel_it, audio_format.num_channels, num_samples);
 195                         break;
 196                 default:
 197                         fprintf(stderr, "Cannot handle audio with %u bits per sample\n", audio_format.bits_per_sample);
 198                         assert(false);
 199                 }
 200         }
 201
 202         // Now add it.
 203         int64_t local_pts = device->next_local_pts;
 204         device->resampling_queue->add_input_samples(local_pts / double(TIMEBASE), audio.data(), num_samples);
 205         device->next_local_pts = local_pts + frame_length;
 206         return true;
 207 }
 208
 209 bool AudioMixer::add_silence(DeviceSpec device_spec, unsigned samples_per_frame, unsigned num_frames, int64_t frame_length)
 210 {
 211         AudioDevice *device = find_audio_device(device_spec);
 212
 213         unique_lock<timed_mutex> lock(audio_mutex, defer_lock);
 214         if (!lock.try_lock_for(chrono::milliseconds(10))) {
 215                 return false;
 216         }
 217         if (device->resampling_queue == nullptr) {
 218                 // No buses use this device; throw it away.
 219                 return true;
 220         }
 221
 222         unsigned num_channels = device->interesting_channels.size();
 223         assert(num_channels > 0);
 224
 225         vector<float> silence(samples_per_frame * num_channels, 0.0f);
 226         for (unsigned i = 0; i < num_frames; ++i) {
 227                 device->resampling_queue->add_input_samples(device->next_local_pts / double(TIMEBASE), silence.data(), samples_per_frame);
 228                 // Note that if the format changed in the meantime, we have
 229                 // no way of detecting that; we just have to assume the frame length
 230                 // is always the same.
 231                 device->next_local_pts += frame_length;
 232         }
 233         return true;
 234 }
 235
 236 AudioMixer::AudioDevice *AudioMixer::find_audio_device(DeviceSpec device)
 237 {
 238         switch (device.type) {
 239         case InputSourceType::CAPTURE_CARD:
 240                 return &video_cards[device.index];
 241         case InputSourceType::ALSA_INPUT:
 242                 return &alsa_inputs[device.index];
 243         case InputSourceType::SILENCE:
 244         default:
 245                 assert(false);
 246         }
 247         return nullptr;
 248 }
 249
 250 // Get a pointer to the given channel from the given device.
 251 // The channel must be picked out earlier and resampled.
 252 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)
 253 {
 254         static float zero = 0.0f;
 255         if (source_channel == -1 || device_spec.type == InputSourceType::SILENCE) {
 256                 *srcptr = &zero;
 257                 *stride = 0;
 258                 return;
 259         }
 260         AudioDevice *device = find_audio_device(device_spec);
 261         assert(device->interesting_channels.count(source_channel) != 0);
 262         unsigned channel_index = 0;
 263         for (int channel : device->interesting_channels) {
 264                 if (channel == source_channel) break;
 265                 ++channel_index;
 266         }
 267         assert(channel_index < device->interesting_channels.size());
 268         const auto it = samples_card.find(device_spec);
 269         assert(it != samples_card.end());
 270         *srcptr = &(it->second)[channel_index];
 271         *stride = device->interesting_channels.size();
 272 }
 273
 274 // TODO: Can be SSSE3-optimized if need be.
 275 void AudioMixer::fill_audio_bus(const map<DeviceSpec, vector<float>> &samples_card, const InputMapping::Bus &bus, unsigned num_samples, float *output)
 276 {
 277         if (bus.device.type == InputSourceType::SILENCE) {
 278                 memset(output, 0, num_samples * sizeof(*output));
 279         } else {
 280                 assert(bus.device.type == InputSourceType::CAPTURE_CARD ||
 281                        bus.device.type == InputSourceType::ALSA_INPUT);
 282                 const float *lsrc, *rsrc;
 283                 unsigned lstride, rstride;
 284                 float *dptr = output;
 285                 find_sample_src_from_device(samples_card, bus.device, bus.source_channel[0], &lsrc, &lstride);
 286                 find_sample_src_from_device(samples_card, bus.device, bus.source_channel[1], &rsrc, &rstride);
 287                 for (unsigned i = 0; i < num_samples; ++i) {
 288                         *dptr++ = *lsrc;
 289                         *dptr++ = *rsrc;
 290                         lsrc += lstride;
 291                         rsrc += rstride;
 292                 }
 293         }
 294 }
 295
 296 vector<float> AudioMixer::get_output(double pts, unsigned num_samples, ResamplingQueue::RateAdjustmentPolicy rate_adjustment_policy)
 297 {
 298         map<DeviceSpec, vector<float>> samples_card;
 299         vector<float> samples_bus;
 300
 301         lock_guard<timed_mutex> lock(audio_mutex);
 302
 303         // Pick out all the interesting channels from all the cards.
 304         // TODO: If the card has been hotswapped, the number of channels
 305         // might have changed; if so, we need to do some sort of remapping
 306         // to silence.
 307         for (const auto &spec_and_info : get_devices_mutex_held()) {
 308                 const DeviceSpec &device_spec = spec_and_info.first;
 309                 AudioDevice *device = find_audio_device(device_spec);
 310                 if (!device->interesting_channels.empty()) {
 311                         samples_card[device_spec].resize(num_samples * device->interesting_channels.size());
 312                         device->resampling_queue->get_output_samples(
 313                                 pts,
 314                                 &samples_card[device_spec][0],
 315                                 num_samples,
 316                                 rate_adjustment_policy);
 317                 }
 318         }
 319
 320         // TODO: Move lo-cut etc. into each bus.
 321         vector<float> samples_out;
 322         samples_out.resize(num_samples * 2);
 323         samples_bus.resize(num_samples * 2);
 324         for (unsigned bus_index = 0; bus_index < input_mapping.buses.size(); ++bus_index) {
 325                 fill_audio_bus(samples_card, input_mapping.buses[bus_index], num_samples, &samples_bus[0]);
 326
 327                 float volume = from_db(fader_volume_db[bus_index]);
 328                 if (bus_index == 0) {
 329                         for (unsigned i = 0; i < num_samples * 2; ++i) {
 330                                 samples_out[i] = samples_bus[i] * volume;
 331                         }
 332                 } else {
 333                         for (unsigned i = 0; i < num_samples * 2; ++i) {
 334                                 samples_out[i] += samples_bus[i] * volume;
 335                         }
 336                 }
 337         }
 338
 339         // Cut away everything under 120 Hz (or whatever the cutoff is);
 340         // we don't need it for voice, and it will reduce headroom
 341         // and confuse the compressor. (In particular, any hums at 50 or 60 Hz
 342         // should be dampened.)
 343         if (locut_enabled) {
 344                 locut.render(samples_out.data(), samples_out.size() / 2, locut_cutoff_hz * 2.0 * M_PI / OUTPUT_FREQUENCY, 0.5f);
 345         }
 346
 347         {
 348                 lock_guard<mutex> lock(compressor_mutex);
 349
 350                 // Apply a level compressor to get the general level right.
 351                 // Basically, if it's over about -40 dBFS, we squeeze it down to that level
 352                 // (or more precisely, near it, since we don't use infinite ratio),
 353                 // then apply a makeup gain to get it to -14 dBFS. -14 dBFS is, of course,
 354                 // entirely arbitrary, but from practical tests with speech, it seems to
 355                 // put ut around -23 LUFS, so it's a reasonable starting point for later use.
 356                 {
 357                         if (level_compressor_enabled) {
 358                                 float threshold = 0.01f;   // -40 dBFS.
 359                                 float ratio = 20.0f;
 360                                 float attack_time = 0.5f;
 361                                 float release_time = 20.0f;
 362                                 float makeup_gain = from_db(ref_level_dbfs - (-40.0f));  // +26 dB.
 363                                 level_compressor.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 364                                 gain_staging_db = to_db(level_compressor.get_attenuation() * makeup_gain);
 365                         } else {
 366                                 // Just apply the gain we already had.
 367                                 float g = from_db(gain_staging_db);
 368                                 for (size_t i = 0; i < samples_out.size(); ++i) {
 369                                         samples_out[i] *= g;
 370                                 }
 371                         }
 372                 }
 373
 374         #if 0
 375                 printf("level=%f (%+5.2f dBFS) attenuation=%f (%+5.2f dB) end_result=%+5.2f dB\n",
 376                         level_compressor.get_level(), to_db(level_compressor.get_level()),
 377                         level_compressor.get_attenuation(), to_db(level_compressor.get_attenuation()),
 378                         to_db(level_compressor.get_level() * level_compressor.get_attenuation() * makeup_gain));
 379         #endif
 380
 381         //      float limiter_att, compressor_att;
 382
 383                 // The real compressor.
 384                 if (compressor_enabled) {
 385                         float threshold = from_db(compressor_threshold_dbfs);
 386                         float ratio = 20.0f;
 387                         float attack_time = 0.005f;
 388                         float release_time = 0.040f;
 389                         float makeup_gain = 2.0f;  // +6 dB.
 390                         compressor.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 391         //              compressor_att = compressor.get_attenuation();
 392                 }
 393
 394                 // Finally a limiter at -4 dB (so, -10 dBFS) to take out the worst peaks only.
 395                 // Note that since ratio is not infinite, we could go slightly higher than this.
 396                 if (limiter_enabled) {
 397                         float threshold = from_db(limiter_threshold_dbfs);
 398                         float ratio = 30.0f;
 399                         float attack_time = 0.0f;  // Instant.
 400                         float release_time = 0.020f;
 401                         float makeup_gain = 1.0f;  // 0 dB.
 402                         limiter.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 403         //              limiter_att = limiter.get_attenuation();
 404                 }
 405
 406         //      printf("limiter=%+5.1f  compressor=%+5.1f\n", to_db(limiter_att), to_db(compressor_att));
 407         }
 408
 409         // At this point, we are most likely close to +0 LU, but all of our
 410         // measurements have been on raw sample values, not R128 values.
 411         // So we have a final makeup gain to get us to +0 LU; the gain
 412         // adjustments required should be relatively small, and also, the
 413         // offset shouldn't change much (only if the type of audio changes
 414         // significantly). Thus, we shoot for updating this value basically
 415         // “whenever we process buffers”, since the R128 calculation isn't exactly
 416         // something we get out per-sample.
 417         //
 418         // Note that there's a feedback loop here, so we choose a very slow filter
 419         // (half-time of 30 seconds).
 420         double target_loudness_factor, alpha;
 421         double loudness_lu = loudness_lufs - ref_level_lufs;
 422         double current_makeup_lu = to_db(final_makeup_gain);
 423         target_loudness_factor = final_makeup_gain * from_db(-loudness_lu);
 424
 425         // If we're outside +/- 5 LU uncorrected, we don't count it as
 426         // a normal signal (probably silence) and don't change the
 427         // correction factor; just apply what we already have.
 428         if (fabs(loudness_lu - current_makeup_lu) >= 5.0 || !final_makeup_gain_auto) {
 429                 alpha = 0.0;
 430         } else {
 431                 // Formula adapted from
 432                 // https://en.wikipedia.org/wiki/Low-pass_filter#Simple_infinite_impulse_response_filter.
 433                 const double half_time_s = 30.0;
 434                 const double fc_mul_2pi_delta_t = 1.0 / (half_time_s * OUTPUT_FREQUENCY);
 435                 alpha = fc_mul_2pi_delta_t / (fc_mul_2pi_delta_t + 1.0);
 436         }
 437
 438         {
 439                 lock_guard<mutex> lock(compressor_mutex);
 440                 double m = final_makeup_gain;
 441                 for (size_t i = 0; i < samples_out.size(); i += 2) {
 442                         samples_out[i + 0] *= m;
 443                         samples_out[i + 1] *= m;
 444                         m += (target_loudness_factor - m) * alpha;
 445                 }
 446                 final_makeup_gain = m;
 447         }
 448
 449         return samples_out;
 450 }
 451
 452 map<DeviceSpec, DeviceInfo> AudioMixer::get_devices() const
 453 {
 454         lock_guard<timed_mutex> lock(audio_mutex);
 455         return get_devices_mutex_held();
 456 }
 457
 458 map<DeviceSpec, DeviceInfo> AudioMixer::get_devices_mutex_held() const
 459 {
 460         map<DeviceSpec, DeviceInfo> devices;
 461         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 462                 const DeviceSpec spec{ InputSourceType::CAPTURE_CARD, card_index };
 463                 const AudioDevice *device = &video_cards[card_index];
 464                 DeviceInfo info;
 465                 info.name = device->name;
 466                 info.num_channels = 8;  // FIXME: This is wrong for fake cards.
 467                 devices.insert(make_pair(spec, info));
 468         }
 469         for (unsigned card_index = 0; card_index < available_alsa_cards.size(); ++card_index) {
 470                 const DeviceSpec spec{ InputSourceType::ALSA_INPUT, card_index };
 471                 const ALSAInput::Device &device = available_alsa_cards[card_index];
 472                 DeviceInfo info;
 473                 info.name = device.name + " (" + device.info + ")";
 474                 info.num_channels = device.num_channels;
 475                 devices.insert(make_pair(spec, info));
 476         }
 477         return devices;
 478 }
 479
 480 void AudioMixer::set_name(DeviceSpec device_spec, const string &name)
 481 {
 482         AudioDevice *device = find_audio_device(device_spec);
 483
 484         lock_guard<timed_mutex> lock(audio_mutex);
 485         device->name = name;
 486 }
 487
 488 void AudioMixer::set_input_mapping(const InputMapping &new_input_mapping)
 489 {
 490         lock_guard<timed_mutex> lock(audio_mutex);
 491
 492         map<DeviceSpec, set<unsigned>> interesting_channels;
 493         for (const InputMapping::Bus &bus : new_input_mapping.buses) {
 494                 if (bus.device.type == InputSourceType::CAPTURE_CARD ||
 495                     bus.device.type == InputSourceType::ALSA_INPUT) {
 496                         for (unsigned channel = 0; channel < 2; ++channel) {
 497                                 if (bus.source_channel[channel] != -1) {
 498                                         interesting_channels[bus.device].insert(bus.source_channel[channel]);
 499                                 }
 500                         }
 501                 }
 502         }
 503
 504         // Reset resamplers for all cards that don't have the exact same state as before.
 505         for (const auto &spec_and_info : get_devices_mutex_held()) {
 506                 const DeviceSpec &device_spec = spec_and_info.first;
 507                 AudioDevice *device = find_audio_device(device_spec);
 508                 if (device->interesting_channels != interesting_channels[device_spec]) {
 509                         device->interesting_channels = interesting_channels[device_spec];
 510                         if (device_spec.type == InputSourceType::ALSA_INPUT) {
 511                                 reset_alsa_mutex_held(device_spec);
 512                         }
 513                         reset_resampler_mutex_held(device_spec);
 514                 }
 515         }
 516
 517         input_mapping = new_input_mapping;
 518 }
 519
 520 InputMapping AudioMixer::get_input_mapping() const
 521 {
 522         lock_guard<timed_mutex> lock(audio_mutex);
 523         return input_mapping;
 524 }