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_fixed24_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 * 3;
  29         dst += out_channel;
  30
  31         for (size_t i = 0; i < num_samples; ++i) {
  32                 uint32_t s1 = src[0];
  33                 uint32_t s2 = src[1];
  34                 uint32_t s3 = src[2];
  35                 uint32_t s = s1 | (s1 << 8) | (s2 << 16) | (s3 << 24);
  36                 *dst = int(s) * (1.0f / 2147483648.0f);
  37
  38                 src += 3 * in_num_channels;
  39                 dst += out_num_channels;
  40         }
  41 }
  42
  43 void convert_fixed32_to_fp32(float *dst, size_t out_channel, size_t out_num_channels,
  44                              const uint8_t *src, size_t in_channel, size_t in_num_channels,
  45                              size_t num_samples)
  46 {
  47         assert(in_channel < in_num_channels);
  48         assert(out_channel < out_num_channels);
  49         src += in_channel * 4;
  50         dst += out_channel;
  51
  52         for (size_t i = 0; i < num_samples; ++i) {
  53                 int32_t s = le32toh(*(int32_t *)src);
  54                 *dst = s * (1.0f / 2147483648.0f);
  55
  56                 src += 4 * in_num_channels;
  57                 dst += out_num_channels;
  58         }
  59 }
  60
  61 }  // namespace
  62
  63 AudioMixer::AudioMixer(unsigned num_cards)
  64         : num_cards(num_cards),
  65           level_compressor(OUTPUT_FREQUENCY),
  66           limiter(OUTPUT_FREQUENCY),
  67           compressor(OUTPUT_FREQUENCY)
  68 {
  69         locut.init(FILTER_HPF, 2);
  70
  71         set_locut_enabled(global_flags.locut_enabled);
  72         set_gain_staging_db(global_flags.initial_gain_staging_db);
  73         set_gain_staging_auto(global_flags.gain_staging_auto);
  74         set_compressor_enabled(global_flags.compressor_enabled);
  75         set_limiter_enabled(global_flags.limiter_enabled);
  76         set_final_makeup_gain_auto(global_flags.final_makeup_gain_auto);
  77
  78         // Generate a very simple, default input mapping.
  79         InputMapping::Bus input;
  80         input.name = "Main";
  81         input.input_source_type = InputSourceType::CAPTURE_CARD;
  82         input.input_source_index = 0;
  83         input.source_channel[0] = 0;
  84         input.source_channel[1] = 1;
  85
  86         InputMapping new_input_mapping;
  87         new_input_mapping.buses.push_back(input);
  88         set_input_mapping(new_input_mapping);
  89 }
  90
  91 void AudioMixer::reset_card(unsigned card_index)
  92 {
  93         lock_guard<mutex> lock(audio_mutex);
  94         reset_card_mutex_held(card_index);
  95 }
  96
  97 void AudioMixer::reset_card_mutex_held(unsigned card_index)
  98 {
  99         CaptureCard *card = &cards[card_index];
 100         if (card->interesting_channels.empty()) {
 101                 card->resampling_queue.reset();
 102         } else {
 103                 card->resampling_queue.reset(new ResamplingQueue(card_index, OUTPUT_FREQUENCY, OUTPUT_FREQUENCY, card->interesting_channels.size()));
 104         }
 105         card->next_local_pts = 0;
 106 }
 107
 108 void AudioMixer::add_audio(unsigned card_index, const uint8_t *data, unsigned num_samples, AudioFormat audio_format, int64_t frame_length)
 109 {
 110         lock_guard<mutex> lock(audio_mutex);
 111         CaptureCard *card = &cards[card_index];
 112
 113         if (card->resampling_queue == nullptr) {
 114                 // No buses use this card; throw it away.
 115                 return;
 116         }
 117
 118         unsigned num_channels = card->interesting_channels.size();
 119         assert(num_channels > 0);
 120
 121         // Convert the audio to stereo fp32.
 122         vector<float> audio;
 123         audio.resize(num_samples * num_channels);
 124         unsigned channel_index = 0;
 125         for (auto channel_it = card->interesting_channels.cbegin(); channel_it != card->interesting_channels.end(); ++channel_it, ++channel_index) {
 126                 switch (audio_format.bits_per_sample) {
 127                 case 0:
 128                         assert(num_samples == 0);
 129                         break;
 130                 case 24:
 131                         convert_fixed24_to_fp32(&audio[0], channel_index, num_channels, data, *channel_it, audio_format.num_channels, num_samples);
 132                         break;
 133                 case 32:
 134                         convert_fixed32_to_fp32(&audio[0], channel_index, num_channels, data, *channel_it, audio_format.num_channels, num_samples);
 135                         break;
 136                 default:
 137                         fprintf(stderr, "Cannot handle audio with %u bits per sample\n", audio_format.bits_per_sample);
 138                         assert(false);
 139                 }
 140         }
 141
 142         // Now add it.
 143         int64_t local_pts = card->next_local_pts;
 144         card->resampling_queue->add_input_samples(local_pts / double(TIMEBASE), audio.data(), num_samples);
 145         card->next_local_pts = local_pts + frame_length;
 146 }
 147
 148 void AudioMixer::add_silence(unsigned card_index, unsigned samples_per_frame, unsigned num_frames, int64_t frame_length)
 149 {
 150         CaptureCard *card = &cards[card_index];
 151         lock_guard<mutex> lock(audio_mutex);
 152
 153         if (card->resampling_queue == nullptr) {
 154                 // No buses use this card; throw it away.
 155                 return;
 156         }
 157
 158         unsigned num_channels = card->interesting_channels.size();
 159         assert(num_channels > 0);
 160
 161         vector<float> silence(samples_per_frame * num_channels, 0.0f);
 162         for (unsigned i = 0; i < num_frames; ++i) {
 163                 card->resampling_queue->add_input_samples(card->next_local_pts / double(TIMEBASE), silence.data(), samples_per_frame);
 164                 // Note that if the format changed in the meantime, we have
 165                 // no way of detecting that; we just have to assume the frame length
 166                 // is always the same.
 167                 card->next_local_pts += frame_length;
 168         }
 169 }
 170
 171 void AudioMixer::find_sample_src_from_capture_card(const vector<float> *samples_card, unsigned card_index, int source_channel, const float **srcptr, unsigned *stride)
 172 {
 173         static float zero = 0.0f;
 174         if (source_channel == -1) {
 175                 *srcptr = &zero;
 176                 *stride = 0;
 177                 return;
 178         }
 179         CaptureCard *card = &cards[card_index];
 180         unsigned channel_index = 0;
 181         for (int channel : card->interesting_channels) {
 182                 if (channel == source_channel) break;
 183                 ++channel_index;
 184         }
 185         assert(channel_index < card->interesting_channels.size());
 186         *srcptr = &samples_card[card_index][channel_index];
 187         *stride = card->interesting_channels.size();
 188 }
 189
 190 vector<float> AudioMixer::get_output(double pts, unsigned num_samples, ResamplingQueue::RateAdjustmentPolicy rate_adjustment_policy)
 191 {
 192         vector<float> samples_card[MAX_CARDS];
 193         vector<float> samples_bus;
 194
 195         lock_guard<mutex> lock(audio_mutex);
 196
 197         // Pick out all the interesting channels from all the cards.
 198         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 199                 CaptureCard *card = &cards[card_index];
 200                 if (!card->interesting_channels.empty()) {
 201                         samples_card[card_index].resize(num_samples * card->interesting_channels.size());
 202                         card->resampling_queue->get_output_samples(
 203                                 pts,
 204                                 &samples_card[card_index][0],
 205                                 num_samples,
 206                                 rate_adjustment_policy);
 207                 }
 208         }
 209
 210         // TODO: Move lo-cut etc. into each bus.
 211         vector<float> samples_out;
 212         samples_out.resize(num_samples * 2);
 213         samples_bus.resize(num_samples * 2);
 214         for (unsigned bus_index = 0; bus_index < input_mapping.buses.size(); ++bus_index) {
 215                 const InputMapping::Bus &input = input_mapping.buses[bus_index];
 216                 if (input.input_source_type == InputSourceType::SILENCE) {
 217                         memset(&samples_bus[0], 0, samples_bus.size() * sizeof(samples_bus[0]));
 218                 } else {
 219                         // TODO: Move this into its own function. Can be SSSE3-optimized if need be.
 220                         assert(input.input_source_type == InputSourceType::CAPTURE_CARD);
 221                         const float *lsrc, *rsrc;
 222                         unsigned lstride, rstride;
 223                         float *dptr = &samples_bus[0];
 224                         find_sample_src_from_capture_card(samples_card, input.input_source_index, input.source_channel[0], &lsrc, &lstride);
 225                         find_sample_src_from_capture_card(samples_card, input.input_source_index, input.source_channel[1], &rsrc, &rstride);
 226                         for (unsigned i = 0; i < num_samples; ++i) {
 227                                 *dptr++ = *lsrc;
 228                                 *dptr++ = *rsrc;
 229                                 lsrc += lstride;
 230                                 rsrc += rstride;
 231                         }
 232                 }
 233
 234                 float volume = from_db(fader_volume_db[bus_index]);
 235                 if (bus_index == 0) {
 236                         for (unsigned i = 0; i < num_samples * 2; ++i) {
 237                                 samples_out[i] = samples_bus[i] * volume;
 238                         }
 239                 } else {
 240                         for (unsigned i = 0; i < num_samples * 2; ++i) {
 241                                 samples_out[i] += samples_bus[i] * volume;
 242                         }
 243                 }
 244         }
 245
 246         // Cut away everything under 120 Hz (or whatever the cutoff is);
 247         // we don't need it for voice, and it will reduce headroom
 248         // and confuse the compressor. (In particular, any hums at 50 or 60 Hz
 249         // should be dampened.)
 250         if (locut_enabled) {
 251                 locut.render(samples_out.data(), samples_out.size() / 2, locut_cutoff_hz * 2.0 * M_PI / OUTPUT_FREQUENCY, 0.5f);
 252         }
 253
 254         {
 255                 lock_guard<mutex> lock(compressor_mutex);
 256
 257                 // Apply a level compressor to get the general level right.
 258                 // Basically, if it's over about -40 dBFS, we squeeze it down to that level
 259                 // (or more precisely, near it, since we don't use infinite ratio),
 260                 // then apply a makeup gain to get it to -14 dBFS. -14 dBFS is, of course,
 261                 // entirely arbitrary, but from practical tests with speech, it seems to
 262                 // put ut around -23 LUFS, so it's a reasonable starting point for later use.
 263                 {
 264                         if (level_compressor_enabled) {
 265                                 float threshold = 0.01f;   // -40 dBFS.
 266                                 float ratio = 20.0f;
 267                                 float attack_time = 0.5f;
 268                                 float release_time = 20.0f;
 269                                 float makeup_gain = from_db(ref_level_dbfs - (-40.0f));  // +26 dB.
 270                                 level_compressor.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 271                                 gain_staging_db = to_db(level_compressor.get_attenuation() * makeup_gain);
 272                         } else {
 273                                 // Just apply the gain we already had.
 274                                 float g = from_db(gain_staging_db);
 275                                 for (size_t i = 0; i < samples_out.size(); ++i) {
 276                                         samples_out[i] *= g;
 277                                 }
 278                         }
 279                 }
 280
 281         #if 0
 282                 printf("level=%f (%+5.2f dBFS) attenuation=%f (%+5.2f dB) end_result=%+5.2f dB\n",
 283                         level_compressor.get_level(), to_db(level_compressor.get_level()),
 284                         level_compressor.get_attenuation(), to_db(level_compressor.get_attenuation()),
 285                         to_db(level_compressor.get_level() * level_compressor.get_attenuation() * makeup_gain));
 286         #endif
 287
 288         //      float limiter_att, compressor_att;
 289
 290                 // The real compressor.
 291                 if (compressor_enabled) {
 292                         float threshold = from_db(compressor_threshold_dbfs);
 293                         float ratio = 20.0f;
 294                         float attack_time = 0.005f;
 295                         float release_time = 0.040f;
 296                         float makeup_gain = 2.0f;  // +6 dB.
 297                         compressor.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 298         //              compressor_att = compressor.get_attenuation();
 299                 }
 300
 301                 // Finally a limiter at -4 dB (so, -10 dBFS) to take out the worst peaks only.
 302                 // Note that since ratio is not infinite, we could go slightly higher than this.
 303                 if (limiter_enabled) {
 304                         float threshold = from_db(limiter_threshold_dbfs);
 305                         float ratio = 30.0f;
 306                         float attack_time = 0.0f;  // Instant.
 307                         float release_time = 0.020f;
 308                         float makeup_gain = 1.0f;  // 0 dB.
 309                         limiter.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 310         //              limiter_att = limiter.get_attenuation();
 311                 }
 312
 313         //      printf("limiter=%+5.1f  compressor=%+5.1f\n", to_db(limiter_att), to_db(compressor_att));
 314         }
 315
 316         // At this point, we are most likely close to +0 LU, but all of our
 317         // measurements have been on raw sample values, not R128 values.
 318         // So we have a final makeup gain to get us to +0 LU; the gain
 319         // adjustments required should be relatively small, and also, the
 320         // offset shouldn't change much (only if the type of audio changes
 321         // significantly). Thus, we shoot for updating this value basically
 322         // “whenever we process buffers”, since the R128 calculation isn't exactly
 323         // something we get out per-sample.
 324         //
 325         // Note that there's a feedback loop here, so we choose a very slow filter
 326         // (half-time of 30 seconds).
 327         double target_loudness_factor, alpha;
 328         double loudness_lu = loudness_lufs - ref_level_lufs;
 329         double current_makeup_lu = to_db(final_makeup_gain);
 330         target_loudness_factor = final_makeup_gain * from_db(-loudness_lu);
 331
 332         // If we're outside +/- 5 LU uncorrected, we don't count it as
 333         // a normal signal (probably silence) and don't change the
 334         // correction factor; just apply what we already have.
 335         if (fabs(loudness_lu - current_makeup_lu) >= 5.0 || !final_makeup_gain_auto) {
 336                 alpha = 0.0;
 337         } else {
 338                 // Formula adapted from
 339                 // https://en.wikipedia.org/wiki/Low-pass_filter#Simple_infinite_impulse_response_filter.
 340                 const double half_time_s = 30.0;
 341                 const double fc_mul_2pi_delta_t = 1.0 / (half_time_s * OUTPUT_FREQUENCY);
 342                 alpha = fc_mul_2pi_delta_t / (fc_mul_2pi_delta_t + 1.0);
 343         }
 344
 345         {
 346                 lock_guard<mutex> lock(compressor_mutex);
 347                 double m = final_makeup_gain;
 348                 for (size_t i = 0; i < samples_out.size(); i += 2) {
 349                         samples_out[i + 0] *= m;
 350                         samples_out[i + 1] *= m;
 351                         m += (target_loudness_factor - m) * alpha;
 352                 }
 353                 final_makeup_gain = m;
 354         }
 355
 356         return samples_out;
 357 }
 358
 359 vector<string> AudioMixer::get_names() const
 360 {
 361         lock_guard<mutex> lock(audio_mutex);
 362         vector<string> names;
 363         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 364                 const CaptureCard *card = &cards[card_index];
 365                 names.push_back(card->name);
 366         }
 367         return names;
 368 }
 369
 370 void AudioMixer::set_name(unsigned card_index, const string &name)
 371 {
 372         lock_guard<mutex> lock(audio_mutex);
 373         CaptureCard *card = &cards[card_index];
 374         card->name = name;
 375 }
 376
 377 void AudioMixer::set_input_mapping(const InputMapping &new_input_mapping)
 378 {
 379         lock_guard<mutex> lock(audio_mutex);
 380
 381         map<unsigned, set<unsigned>> interesting_channels;
 382         for (const InputMapping::Bus &bus : new_input_mapping.buses) {
 383                 if (bus.input_source_type == InputSourceType::CAPTURE_CARD) {
 384                         for (unsigned channel = 0; channel < 2; ++channel) {
 385                                 if (bus.source_channel[channel] != -1) {
 386                                         interesting_channels[bus.input_source_index].insert(bus.source_channel[channel]);
 387                                 }
 388                         }
 389                 }
 390         }
 391
 392         // Reset resamplers for all cards that don't have the exact same state as before.
 393         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 394                 CaptureCard *card = &cards[card_index];
 395                 if (card->interesting_channels != interesting_channels[card_index]) {
 396                         card->interesting_channels = interesting_channels[card_index];
 397                         reset_card_mutex_held(card_index);
 398                 }
 399         }
 400
 401         input_mapping = new_input_mapping;
 402 }
 403
 404 InputMapping AudioMixer::get_input_mapping() const
 405 {
 406         lock_guard<mutex> lock(audio_mutex);
 407         return input_mapping;
 408 }