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