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<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 void 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         lock_guard<mutex> lock(audio_mutex);
 167         if (device->resampling_queue == nullptr) {
 168                 // No buses use this device; throw it away.
 169                 return;
 170         }
 171
 172         unsigned num_channels = device->interesting_channels.size();
 173         assert(num_channels > 0);
 174
 175         // Convert the audio to fp32.
 176         vector<float> audio;
 177         audio.resize(num_samples * num_channels);
 178         unsigned channel_index = 0;
 179         for (auto channel_it = device->interesting_channels.cbegin(); channel_it != device->interesting_channels.end(); ++channel_it, ++channel_index) {
 180                 switch (audio_format.bits_per_sample) {
 181                 case 0:
 182                         assert(num_samples == 0);
 183                         break;
 184                 case 16:
 185                         convert_fixed16_to_fp32(&audio[0], channel_index, num_channels, data, *channel_it, audio_format.num_channels, num_samples);
 186                         break;
 187                 case 24:
 188                         convert_fixed24_to_fp32(&audio[0], channel_index, num_channels, data, *channel_it, audio_format.num_channels, num_samples);
 189                         break;
 190                 case 32:
 191                         convert_fixed32_to_fp32(&audio[0], channel_index, num_channels, data, *channel_it, audio_format.num_channels, num_samples);
 192                         break;
 193                 default:
 194                         fprintf(stderr, "Cannot handle audio with %u bits per sample\n", audio_format.bits_per_sample);
 195                         assert(false);
 196                 }
 197         }
 198
 199         // Now add it.
 200         int64_t local_pts = device->next_local_pts;
 201         device->resampling_queue->add_input_samples(local_pts / double(TIMEBASE), audio.data(), num_samples);
 202         device->next_local_pts = local_pts + frame_length;
 203 }
 204
 205 void AudioMixer::add_silence(DeviceSpec device_spec, unsigned samples_per_frame, unsigned num_frames, int64_t frame_length)
 206 {
 207         AudioDevice *device = find_audio_device(device_spec);
 208
 209         lock_guard<mutex> lock(audio_mutex);
 210         if (device->resampling_queue == nullptr) {
 211                 // No buses use this device; throw it away.
 212                 return;
 213         }
 214
 215         unsigned num_channels = device->interesting_channels.size();
 216         assert(num_channels > 0);
 217
 218         vector<float> silence(samples_per_frame * num_channels, 0.0f);
 219         for (unsigned i = 0; i < num_frames; ++i) {
 220                 device->resampling_queue->add_input_samples(device->next_local_pts / double(TIMEBASE), silence.data(), samples_per_frame);
 221                 // Note that if the format changed in the meantime, we have
 222                 // no way of detecting that; we just have to assume the frame length
 223                 // is always the same.
 224                 device->next_local_pts += frame_length;
 225         }
 226 }
 227
 228 AudioMixer::AudioDevice *AudioMixer::find_audio_device(DeviceSpec device)
 229 {
 230         switch (device.type) {
 231         case InputSourceType::CAPTURE_CARD:
 232                 return &video_cards[device.index];
 233         case InputSourceType::ALSA_INPUT:
 234                 return &alsa_inputs[device.index];
 235         case InputSourceType::SILENCE:
 236         default:
 237                 assert(false);
 238         }
 239         return nullptr;
 240 }
 241
 242 // Get a pointer to the given channel from the given device.
 243 // The channel must be picked out earlier and resampled.
 244 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)
 245 {
 246         static float zero = 0.0f;
 247         if (source_channel == -1 || device_spec.type == InputSourceType::SILENCE) {
 248                 *srcptr = &zero;
 249                 *stride = 0;
 250                 return;
 251         }
 252         AudioDevice *device = find_audio_device(device_spec);
 253         assert(device->interesting_channels.count(source_channel) != 0);
 254         unsigned channel_index = 0;
 255         for (int channel : device->interesting_channels) {
 256                 if (channel == source_channel) break;
 257                 ++channel_index;
 258         }
 259         assert(channel_index < device->interesting_channels.size());
 260         const auto it = samples_card.find(device_spec);
 261         assert(it != samples_card.end());
 262         *srcptr = &(it->second)[channel_index];
 263         *stride = device->interesting_channels.size();
 264 }
 265
 266 // TODO: Can be SSSE3-optimized if need be.
 267 void AudioMixer::fill_audio_bus(const map<DeviceSpec, vector<float>> &samples_card, const InputMapping::Bus &bus, unsigned num_samples, float *output)
 268 {
 269         if (bus.device.type == InputSourceType::SILENCE) {
 270                 memset(output, 0, num_samples * sizeof(*output));
 271         } else {
 272                 assert(bus.device.type == InputSourceType::CAPTURE_CARD ||
 273                        bus.device.type == InputSourceType::ALSA_INPUT);
 274                 const float *lsrc, *rsrc;
 275                 unsigned lstride, rstride;
 276                 float *dptr = output;
 277                 find_sample_src_from_device(samples_card, bus.device, bus.source_channel[0], &lsrc, &lstride);
 278                 find_sample_src_from_device(samples_card, bus.device, bus.source_channel[1], &rsrc, &rstride);
 279                 for (unsigned i = 0; i < num_samples; ++i) {
 280                         *dptr++ = *lsrc;
 281                         *dptr++ = *rsrc;
 282                         lsrc += lstride;
 283                         rsrc += rstride;
 284                 }
 285         }
 286 }
 287
 288 vector<float> AudioMixer::get_output(double pts, unsigned num_samples, ResamplingQueue::RateAdjustmentPolicy rate_adjustment_policy)
 289 {
 290         map<DeviceSpec, vector<float>> samples_card;
 291         vector<float> samples_bus;
 292
 293         lock_guard<mutex> lock(audio_mutex);
 294
 295         // Pick out all the interesting channels from all the cards.
 296         // TODO: If the card has been hotswapped, the number of channels
 297         // might have changed; if so, we need to do some sort of remapping
 298         // to silence.
 299         for (const auto &spec_and_info : get_devices_mutex_held()) {
 300                 const DeviceSpec &device_spec = spec_and_info.first;
 301                 AudioDevice *device = find_audio_device(device_spec);
 302                 if (!device->interesting_channels.empty()) {
 303                         samples_card[device_spec].resize(num_samples * device->interesting_channels.size());
 304                         device->resampling_queue->get_output_samples(
 305                                 pts,
 306                                 &samples_card[device_spec][0],
 307                                 num_samples,
 308                                 rate_adjustment_policy);
 309                 }
 310         }
 311
 312         // TODO: Move lo-cut etc. into each bus.
 313         vector<float> samples_out;
 314         samples_out.resize(num_samples * 2);
 315         samples_bus.resize(num_samples * 2);
 316         for (unsigned bus_index = 0; bus_index < input_mapping.buses.size(); ++bus_index) {
 317                 fill_audio_bus(samples_card, input_mapping.buses[bus_index], num_samples, &samples_bus[0]);
 318
 319                 float volume = from_db(fader_volume_db[bus_index]);
 320                 if (bus_index == 0) {
 321                         for (unsigned i = 0; i < num_samples * 2; ++i) {
 322                                 samples_out[i] = samples_bus[i] * volume;
 323                         }
 324                 } else {
 325                         for (unsigned i = 0; i < num_samples * 2; ++i) {
 326                                 samples_out[i] += samples_bus[i] * volume;
 327                         }
 328                 }
 329         }
 330
 331         // Cut away everything under 120 Hz (or whatever the cutoff is);
 332         // we don't need it for voice, and it will reduce headroom
 333         // and confuse the compressor. (In particular, any hums at 50 or 60 Hz
 334         // should be dampened.)
 335         if (locut_enabled) {
 336                 locut.render(samples_out.data(), samples_out.size() / 2, locut_cutoff_hz * 2.0 * M_PI / OUTPUT_FREQUENCY, 0.5f);
 337         }
 338
 339         {
 340                 lock_guard<mutex> lock(compressor_mutex);
 341
 342                 // Apply a level compressor to get the general level right.
 343                 // Basically, if it's over about -40 dBFS, we squeeze it down to that level
 344                 // (or more precisely, near it, since we don't use infinite ratio),
 345                 // then apply a makeup gain to get it to -14 dBFS. -14 dBFS is, of course,
 346                 // entirely arbitrary, but from practical tests with speech, it seems to
 347                 // put ut around -23 LUFS, so it's a reasonable starting point for later use.
 348                 {
 349                         if (level_compressor_enabled) {
 350                                 float threshold = 0.01f;   // -40 dBFS.
 351                                 float ratio = 20.0f;
 352                                 float attack_time = 0.5f;
 353                                 float release_time = 20.0f;
 354                                 float makeup_gain = from_db(ref_level_dbfs - (-40.0f));  // +26 dB.
 355                                 level_compressor.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 356                                 gain_staging_db = to_db(level_compressor.get_attenuation() * makeup_gain);
 357                         } else {
 358                                 // Just apply the gain we already had.
 359                                 float g = from_db(gain_staging_db);
 360                                 for (size_t i = 0; i < samples_out.size(); ++i) {
 361                                         samples_out[i] *= g;
 362                                 }
 363                         }
 364                 }
 365
 366         #if 0
 367                 printf("level=%f (%+5.2f dBFS) attenuation=%f (%+5.2f dB) end_result=%+5.2f dB\n",
 368                         level_compressor.get_level(), to_db(level_compressor.get_level()),
 369                         level_compressor.get_attenuation(), to_db(level_compressor.get_attenuation()),
 370                         to_db(level_compressor.get_level() * level_compressor.get_attenuation() * makeup_gain));
 371         #endif
 372
 373         //      float limiter_att, compressor_att;
 374
 375                 // The real compressor.
 376                 if (compressor_enabled) {
 377                         float threshold = from_db(compressor_threshold_dbfs);
 378                         float ratio = 20.0f;
 379                         float attack_time = 0.005f;
 380                         float release_time = 0.040f;
 381                         float makeup_gain = 2.0f;  // +6 dB.
 382                         compressor.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 383         //              compressor_att = compressor.get_attenuation();
 384                 }
 385
 386                 // Finally a limiter at -4 dB (so, -10 dBFS) to take out the worst peaks only.
 387                 // Note that since ratio is not infinite, we could go slightly higher than this.
 388                 if (limiter_enabled) {
 389                         float threshold = from_db(limiter_threshold_dbfs);
 390                         float ratio = 30.0f;
 391                         float attack_time = 0.0f;  // Instant.
 392                         float release_time = 0.020f;
 393                         float makeup_gain = 1.0f;  // 0 dB.
 394                         limiter.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 395         //              limiter_att = limiter.get_attenuation();
 396                 }
 397
 398         //      printf("limiter=%+5.1f  compressor=%+5.1f\n", to_db(limiter_att), to_db(compressor_att));
 399         }
 400
 401         // At this point, we are most likely close to +0 LU, but all of our
 402         // measurements have been on raw sample values, not R128 values.
 403         // So we have a final makeup gain to get us to +0 LU; the gain
 404         // adjustments required should be relatively small, and also, the
 405         // offset shouldn't change much (only if the type of audio changes
 406         // significantly). Thus, we shoot for updating this value basically
 407         // “whenever we process buffers”, since the R128 calculation isn't exactly
 408         // something we get out per-sample.
 409         //
 410         // Note that there's a feedback loop here, so we choose a very slow filter
 411         // (half-time of 30 seconds).
 412         double target_loudness_factor, alpha;
 413         double loudness_lu = loudness_lufs - ref_level_lufs;
 414         double current_makeup_lu = to_db(final_makeup_gain);
 415         target_loudness_factor = final_makeup_gain * from_db(-loudness_lu);
 416
 417         // If we're outside +/- 5 LU uncorrected, we don't count it as
 418         // a normal signal (probably silence) and don't change the
 419         // correction factor; just apply what we already have.
 420         if (fabs(loudness_lu - current_makeup_lu) >= 5.0 || !final_makeup_gain_auto) {
 421                 alpha = 0.0;
 422         } else {
 423                 // Formula adapted from
 424                 // https://en.wikipedia.org/wiki/Low-pass_filter#Simple_infinite_impulse_response_filter.
 425                 const double half_time_s = 30.0;
 426                 const double fc_mul_2pi_delta_t = 1.0 / (half_time_s * OUTPUT_FREQUENCY);
 427                 alpha = fc_mul_2pi_delta_t / (fc_mul_2pi_delta_t + 1.0);
 428         }
 429
 430         {
 431                 lock_guard<mutex> lock(compressor_mutex);
 432                 double m = final_makeup_gain;
 433                 for (size_t i = 0; i < samples_out.size(); i += 2) {
 434                         samples_out[i + 0] *= m;
 435                         samples_out[i + 1] *= m;
 436                         m += (target_loudness_factor - m) * alpha;
 437                 }
 438                 final_makeup_gain = m;
 439         }
 440
 441         return samples_out;
 442 }
 443
 444 map<DeviceSpec, DeviceInfo> AudioMixer::get_devices() const
 445 {
 446         lock_guard<mutex> lock(audio_mutex);
 447         return get_devices_mutex_held();
 448 }
 449
 450 map<DeviceSpec, DeviceInfo> AudioMixer::get_devices_mutex_held() const
 451 {
 452         map<DeviceSpec, DeviceInfo> devices;
 453         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 454                 const DeviceSpec spec{ InputSourceType::CAPTURE_CARD, card_index };
 455                 const AudioDevice *device = &video_cards[card_index];
 456                 DeviceInfo info;
 457                 info.name = device->name;
 458                 info.num_channels = 8;  // FIXME: This is wrong for fake cards.
 459                 devices.insert(make_pair(spec, info));
 460         }
 461         for (unsigned card_index = 0; card_index < available_alsa_cards.size(); ++card_index) {
 462                 const DeviceSpec spec{ InputSourceType::ALSA_INPUT, card_index };
 463                 const ALSAInput::Device &device = available_alsa_cards[card_index];
 464                 DeviceInfo info;
 465                 info.name = device.name + " (" + device.info + ")";
 466                 info.num_channels = device.num_channels;
 467                 devices.insert(make_pair(spec, info));
 468         }
 469         return devices;
 470 }
 471
 472 void AudioMixer::set_name(DeviceSpec device_spec, const string &name)
 473 {
 474         AudioDevice *device = find_audio_device(device_spec);
 475
 476         lock_guard<mutex> lock(audio_mutex);
 477         device->name = name;
 478 }
 479
 480 void AudioMixer::set_input_mapping(const InputMapping &new_input_mapping)
 481 {
 482         lock_guard<mutex> lock(audio_mutex);
 483
 484         map<DeviceSpec, set<unsigned>> interesting_channels;
 485         for (const InputMapping::Bus &bus : new_input_mapping.buses) {
 486                 if (bus.device.type == InputSourceType::CAPTURE_CARD ||
 487                     bus.device.type == InputSourceType::ALSA_INPUT) {
 488                         for (unsigned channel = 0; channel < 2; ++channel) {
 489                                 if (bus.source_channel[channel] != -1) {
 490                                         interesting_channels[bus.device].insert(bus.source_channel[channel]);
 491                                 }
 492                         }
 493                 }
 494         }
 495
 496         // Reset resamplers for all cards that don't have the exact same state as before.
 497         for (const auto &spec_and_info : get_devices_mutex_held()) {
 498                 const DeviceSpec &device_spec = spec_and_info.first;
 499                 AudioDevice *device = find_audio_device(device_spec);
 500                 if (device->interesting_channels != interesting_channels[device_spec]) {
 501                         device->interesting_channels = interesting_channels[device_spec];
 502                         if (device_spec.type == InputSourceType::ALSA_INPUT) {
 503                                 reset_alsa_mutex_held(device_spec);
 504                         }
 505                         reset_resampler_mutex_held(device_spec);
 506                 }
 507         }
 508
 509         input_mapping = new_input_mapping;
 510 }
 511
 512 InputMapping AudioMixer::get_input_mapping() const
 513 {
 514         lock_guard<mutex> lock(audio_mutex);
 515         return input_mapping;
 516 }