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_resampler(DeviceSpec device_spec)
 110 {
 111         lock_guard<mutex> lock(audio_mutex);
 112         reset_resampler_mutex_held(device_spec);
 113 }
 114
 115 void AudioMixer::reset_resampler_mutex_held(DeviceSpec device_spec)
 116 {
 117         AudioDevice *device = find_audio_device(device_spec);
 118
 119         if (device->interesting_channels.empty()) {
 120                 device->resampling_queue.reset();
 121         } else {
 122                 // TODO: ResamplingQueue should probably take the full device spec.
 123                 // (It's only used for console output, though.)
 124                 device->resampling_queue.reset(new ResamplingQueue(device_spec.index, device->capture_frequency, OUTPUT_FREQUENCY, device->interesting_channels.size()));
 125         }
 126         device->next_local_pts = 0;
 127 }
 128
 129 void AudioMixer::add_audio(DeviceSpec device_spec, const uint8_t *data, unsigned num_samples, AudioFormat audio_format, int64_t frame_length)
 130 {
 131         AudioDevice *device = find_audio_device(device_spec);
 132
 133         lock_guard<mutex> lock(audio_mutex);
 134         if (device->resampling_queue == nullptr) {
 135                 // No buses use this device; throw it away.
 136                 return;
 137         }
 138
 139         unsigned num_channels = device->interesting_channels.size();
 140         assert(num_channels > 0);
 141
 142         // Convert the audio to fp32.
 143         vector<float> audio;
 144         audio.resize(num_samples * num_channels);
 145         unsigned channel_index = 0;
 146         for (auto channel_it = device->interesting_channels.cbegin(); channel_it != device->interesting_channels.end(); ++channel_it, ++channel_index) {
 147                 switch (audio_format.bits_per_sample) {
 148                 case 0:
 149                         assert(num_samples == 0);
 150                         break;
 151                 case 16:
 152                         convert_fixed16_to_fp32(&audio[0], channel_index, num_channels, data, *channel_it, audio_format.num_channels, num_samples);
 153                         break;
 154                 case 24:
 155                         convert_fixed24_to_fp32(&audio[0], channel_index, num_channels, data, *channel_it, audio_format.num_channels, num_samples);
 156                         break;
 157                 case 32:
 158                         convert_fixed32_to_fp32(&audio[0], channel_index, num_channels, data, *channel_it, audio_format.num_channels, num_samples);
 159                         break;
 160                 default:
 161                         fprintf(stderr, "Cannot handle audio with %u bits per sample\n", audio_format.bits_per_sample);
 162                         assert(false);
 163                 }
 164         }
 165
 166         // Now add it.
 167         int64_t local_pts = device->next_local_pts;
 168         device->resampling_queue->add_input_samples(local_pts / double(TIMEBASE), audio.data(), num_samples);
 169         device->next_local_pts = local_pts + frame_length;
 170 }
 171
 172 void AudioMixer::add_silence(DeviceSpec device_spec, unsigned samples_per_frame, unsigned num_frames, int64_t frame_length)
 173 {
 174         AudioDevice *device = find_audio_device(device_spec);
 175
 176         lock_guard<mutex> lock(audio_mutex);
 177         if (device->resampling_queue == nullptr) {
 178                 // No buses use this device; throw it away.
 179                 return;
 180         }
 181
 182         unsigned num_channels = device->interesting_channels.size();
 183         assert(num_channels > 0);
 184
 185         vector<float> silence(samples_per_frame * num_channels, 0.0f);
 186         for (unsigned i = 0; i < num_frames; ++i) {
 187                 device->resampling_queue->add_input_samples(device->next_local_pts / double(TIMEBASE), silence.data(), samples_per_frame);
 188                 // Note that if the format changed in the meantime, we have
 189                 // no way of detecting that; we just have to assume the frame length
 190                 // is always the same.
 191                 device->next_local_pts += frame_length;
 192         }
 193 }
 194
 195 AudioMixer::AudioDevice *AudioMixer::find_audio_device(DeviceSpec device)
 196 {
 197         switch (device.type) {
 198         case InputSourceType::CAPTURE_CARD:
 199                 return &video_cards[device.index];
 200         case InputSourceType::SILENCE:
 201         default:
 202                 assert(false);
 203         }
 204         return nullptr;
 205 }
 206
 207 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)
 208 {
 209         static float zero = 0.0f;
 210         if (source_channel == -1 || device_spec.type == InputSourceType::SILENCE) {
 211                 *srcptr = &zero;
 212                 *stride = 0;
 213                 return;
 214         }
 215         AudioDevice *device = find_audio_device(device_spec);
 216         unsigned channel_index = 0;
 217         for (int channel : device->interesting_channels) {
 218                 if (channel == source_channel) break;
 219                 ++channel_index;
 220         }
 221         assert(channel_index < device->interesting_channels.size());
 222         const auto it = samples_card.find(device_spec);
 223         assert(it != samples_card.end());
 224         *srcptr = &(it->second)[channel_index];
 225         *stride = device->interesting_channels.size();
 226 }
 227
 228 // TODO: Can be SSSE3-optimized if need be.
 229 void AudioMixer::fill_audio_bus(const map<DeviceSpec, vector<float>> &samples_card, const InputMapping::Bus &bus, unsigned num_samples, float *output)
 230 {
 231         if (bus.device.type == InputSourceType::SILENCE) {
 232                 memset(output, 0, num_samples * sizeof(*output));
 233         } else {
 234                 assert(bus.device.type == InputSourceType::CAPTURE_CARD);
 235                 const float *lsrc, *rsrc;
 236                 unsigned lstride, rstride;
 237                 float *dptr = output;
 238                 find_sample_src_from_device(samples_card, bus.device, bus.source_channel[0], &lsrc, &lstride);
 239                 find_sample_src_from_device(samples_card, bus.device, bus.source_channel[1], &rsrc, &rstride);
 240                 for (unsigned i = 0; i < num_samples; ++i) {
 241                         *dptr++ = *lsrc;
 242                         *dptr++ = *rsrc;
 243                         lsrc += lstride;
 244                         rsrc += rstride;
 245                 }
 246         }
 247 }
 248
 249 vector<float> AudioMixer::get_output(double pts, unsigned num_samples, ResamplingQueue::RateAdjustmentPolicy rate_adjustment_policy)
 250 {
 251         map<DeviceSpec, vector<float>> samples_card;
 252         vector<float> samples_bus;
 253
 254         lock_guard<mutex> lock(audio_mutex);
 255
 256         // Pick out all the interesting channels from all the cards.
 257         // TODO: If the card has been hotswapped, the number of channels
 258         // might have changed; if so, we need to do some sort of remapping
 259         // to silence.
 260         for (const auto &spec_and_info : get_devices_mutex_held()) {
 261                 const DeviceSpec &device_spec = spec_and_info.first;
 262                 AudioDevice *device = find_audio_device(device_spec);
 263                 if (!device->interesting_channels.empty()) {
 264                         samples_card[device_spec].resize(num_samples * device->interesting_channels.size());
 265                         device->resampling_queue->get_output_samples(
 266                                 pts,
 267                                 &samples_card[device_spec][0],
 268                                 num_samples,
 269                                 rate_adjustment_policy);
 270                 }
 271         }
 272
 273         // TODO: Move lo-cut etc. into each bus.
 274         vector<float> samples_out;
 275         samples_out.resize(num_samples * 2);
 276         samples_bus.resize(num_samples * 2);
 277         for (unsigned bus_index = 0; bus_index < input_mapping.buses.size(); ++bus_index) {
 278                 fill_audio_bus(samples_card, input_mapping.buses[bus_index], num_samples, &samples_bus[0]);
 279
 280                 float volume = from_db(fader_volume_db[bus_index]);
 281                 if (bus_index == 0) {
 282                         for (unsigned i = 0; i < num_samples * 2; ++i) {
 283                                 samples_out[i] = samples_bus[i] * volume;
 284                         }
 285                 } else {
 286                         for (unsigned i = 0; i < num_samples * 2; ++i) {
 287                                 samples_out[i] += samples_bus[i] * volume;
 288                         }
 289                 }
 290         }
 291
 292         // Cut away everything under 120 Hz (or whatever the cutoff is);
 293         // we don't need it for voice, and it will reduce headroom
 294         // and confuse the compressor. (In particular, any hums at 50 or 60 Hz
 295         // should be dampened.)
 296         if (locut_enabled) {
 297                 locut.render(samples_out.data(), samples_out.size() / 2, locut_cutoff_hz * 2.0 * M_PI / OUTPUT_FREQUENCY, 0.5f);
 298         }
 299
 300         {
 301                 lock_guard<mutex> lock(compressor_mutex);
 302
 303                 // Apply a level compressor to get the general level right.
 304                 // Basically, if it's over about -40 dBFS, we squeeze it down to that level
 305                 // (or more precisely, near it, since we don't use infinite ratio),
 306                 // then apply a makeup gain to get it to -14 dBFS. -14 dBFS is, of course,
 307                 // entirely arbitrary, but from practical tests with speech, it seems to
 308                 // put ut around -23 LUFS, so it's a reasonable starting point for later use.
 309                 {
 310                         if (level_compressor_enabled) {
 311                                 float threshold = 0.01f;   // -40 dBFS.
 312                                 float ratio = 20.0f;
 313                                 float attack_time = 0.5f;
 314                                 float release_time = 20.0f;
 315                                 float makeup_gain = from_db(ref_level_dbfs - (-40.0f));  // +26 dB.
 316                                 level_compressor.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 317                                 gain_staging_db = to_db(level_compressor.get_attenuation() * makeup_gain);
 318                         } else {
 319                                 // Just apply the gain we already had.
 320                                 float g = from_db(gain_staging_db);
 321                                 for (size_t i = 0; i < samples_out.size(); ++i) {
 322                                         samples_out[i] *= g;
 323                                 }
 324                         }
 325                 }
 326
 327         #if 0
 328                 printf("level=%f (%+5.2f dBFS) attenuation=%f (%+5.2f dB) end_result=%+5.2f dB\n",
 329                         level_compressor.get_level(), to_db(level_compressor.get_level()),
 330                         level_compressor.get_attenuation(), to_db(level_compressor.get_attenuation()),
 331                         to_db(level_compressor.get_level() * level_compressor.get_attenuation() * makeup_gain));
 332         #endif
 333
 334         //      float limiter_att, compressor_att;
 335
 336                 // The real compressor.
 337                 if (compressor_enabled) {
 338                         float threshold = from_db(compressor_threshold_dbfs);
 339                         float ratio = 20.0f;
 340                         float attack_time = 0.005f;
 341                         float release_time = 0.040f;
 342                         float makeup_gain = 2.0f;  // +6 dB.
 343                         compressor.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 344         //              compressor_att = compressor.get_attenuation();
 345                 }
 346
 347                 // Finally a limiter at -4 dB (so, -10 dBFS) to take out the worst peaks only.
 348                 // Note that since ratio is not infinite, we could go slightly higher than this.
 349                 if (limiter_enabled) {
 350                         float threshold = from_db(limiter_threshold_dbfs);
 351                         float ratio = 30.0f;
 352                         float attack_time = 0.0f;  // Instant.
 353                         float release_time = 0.020f;
 354                         float makeup_gain = 1.0f;  // 0 dB.
 355                         limiter.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 356         //              limiter_att = limiter.get_attenuation();
 357                 }
 358
 359         //      printf("limiter=%+5.1f  compressor=%+5.1f\n", to_db(limiter_att), to_db(compressor_att));
 360         }
 361
 362         // At this point, we are most likely close to +0 LU, but all of our
 363         // measurements have been on raw sample values, not R128 values.
 364         // So we have a final makeup gain to get us to +0 LU; the gain
 365         // adjustments required should be relatively small, and also, the
 366         // offset shouldn't change much (only if the type of audio changes
 367         // significantly). Thus, we shoot for updating this value basically
 368         // “whenever we process buffers”, since the R128 calculation isn't exactly
 369         // something we get out per-sample.
 370         //
 371         // Note that there's a feedback loop here, so we choose a very slow filter
 372         // (half-time of 30 seconds).
 373         double target_loudness_factor, alpha;
 374         double loudness_lu = loudness_lufs - ref_level_lufs;
 375         double current_makeup_lu = to_db(final_makeup_gain);
 376         target_loudness_factor = final_makeup_gain * from_db(-loudness_lu);
 377
 378         // If we're outside +/- 5 LU uncorrected, we don't count it as
 379         // a normal signal (probably silence) and don't change the
 380         // correction factor; just apply what we already have.
 381         if (fabs(loudness_lu - current_makeup_lu) >= 5.0 || !final_makeup_gain_auto) {
 382                 alpha = 0.0;
 383         } else {
 384                 // Formula adapted from
 385                 // https://en.wikipedia.org/wiki/Low-pass_filter#Simple_infinite_impulse_response_filter.
 386                 const double half_time_s = 30.0;
 387                 const double fc_mul_2pi_delta_t = 1.0 / (half_time_s * OUTPUT_FREQUENCY);
 388                 alpha = fc_mul_2pi_delta_t / (fc_mul_2pi_delta_t + 1.0);
 389         }
 390
 391         {
 392                 lock_guard<mutex> lock(compressor_mutex);
 393                 double m = final_makeup_gain;
 394                 for (size_t i = 0; i < samples_out.size(); i += 2) {
 395                         samples_out[i + 0] *= m;
 396                         samples_out[i + 1] *= m;
 397                         m += (target_loudness_factor - m) * alpha;
 398                 }
 399                 final_makeup_gain = m;
 400         }
 401
 402         return samples_out;
 403 }
 404
 405 map<DeviceSpec, DeviceInfo> AudioMixer::get_devices() const
 406 {
 407         lock_guard<mutex> lock(audio_mutex);
 408         return get_devices_mutex_held();
 409 }
 410
 411 map<DeviceSpec, DeviceInfo> AudioMixer::get_devices_mutex_held() const
 412 {
 413         map<DeviceSpec, DeviceInfo> devices;
 414         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 415                 const DeviceSpec spec{ InputSourceType::CAPTURE_CARD, card_index };
 416                 const AudioDevice *device = &video_cards[card_index];
 417                 DeviceInfo info;
 418                 info.name = device->name;
 419                 info.num_channels = 8;  // FIXME: This is wrong for fake cards.
 420                 devices.insert(make_pair(spec, info));
 421         }
 422         return devices;
 423 }
 424
 425 void AudioMixer::set_name(DeviceSpec device_spec, const string &name)
 426 {
 427         AudioDevice *device = find_audio_device(device_spec);
 428
 429         lock_guard<mutex> lock(audio_mutex);
 430         device->name = name;
 431 }
 432
 433 void AudioMixer::set_input_mapping(const InputMapping &new_input_mapping)
 434 {
 435         lock_guard<mutex> lock(audio_mutex);
 436
 437         map<DeviceSpec, set<unsigned>> interesting_channels;
 438         for (const InputMapping::Bus &bus : new_input_mapping.buses) {
 439                 if (bus.device.type == InputSourceType::CAPTURE_CARD) {
 440                         for (unsigned channel = 0; channel < 2; ++channel) {
 441                                 if (bus.source_channel[channel] != -1) {
 442                                         interesting_channels[bus.device].insert(bus.source_channel[channel]);
 443                                 }
 444                         }
 445                 }
 446         }
 447
 448         // Reset resamplers for all cards that don't have the exact same state as before.
 449         for (const auto &spec_and_info : get_devices_mutex_held()) {
 450                 const DeviceSpec &device_spec = spec_and_info.first;
 451                 AudioDevice *device = find_audio_device(device_spec);
 452                 if (device->interesting_channels != interesting_channels[device_spec]) {
 453                         device->interesting_channels = interesting_channels[device_spec];
 454                         reset_resampler_mutex_held(device_spec);
 455                 }
 456         }
 457
 458         input_mapping = new_input_mapping;
 459 }
 460
 461 InputMapping AudioMixer::get_input_mapping() const
 462 {
 463         lock_guard<mutex> lock(audio_mutex);
 464         return input_mapping;
 465 }