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 <cmath>
   8
   9 #include "db.h"
  10 #include "flags.h"
  11 #include "timebase.h"
  12
  13 using namespace bmusb;
  14 using namespace std;
  15
  16 namespace {
  17
  18 // TODO: If these prove to be a bottleneck, they can be SSSE3-optimized.
  19
  20 void convert_fixed24_to_fp32(float *dst, size_t out_channels, const uint8_t *src, size_t in_channels, size_t num_samples)
  21 {
  22         assert(in_channels >= out_channels);
  23         for (size_t i = 0; i < num_samples; ++i) {
  24                 for (size_t j = 0; j < out_channels; ++j) {
  25                         uint32_t s1 = *src++;
  26                         uint32_t s2 = *src++;
  27                         uint32_t s3 = *src++;
  28                         uint32_t s = s1 | (s1 << 8) | (s2 << 16) | (s3 << 24);
  29                         dst[i * out_channels + j] = int(s) * (1.0f / 2147483648.0f);
  30                 }
  31                 src += 3 * (in_channels - out_channels);
  32         }
  33 }
  34
  35 void convert_fixed32_to_fp32(float *dst, size_t out_channels, const uint8_t *src, size_t in_channels, size_t num_samples)
  36 {
  37         assert(in_channels >= out_channels);
  38         for (size_t i = 0; i < num_samples; ++i) {
  39                 for (size_t j = 0; j < out_channels; ++j) {
  40                         int32_t s = le32toh(*(int32_t *)src);
  41                         dst[i * out_channels + j] = s * (1.0f / 2147483648.0f);
  42                         src += 4;
  43                 }
  44                 src += 4 * (in_channels - out_channels);
  45         }
  46 }
  47
  48 }  // namespace
  49
  50 AudioMixer::AudioMixer(unsigned num_cards)
  51         : num_cards(num_cards),
  52           level_compressor(OUTPUT_FREQUENCY),
  53           limiter(OUTPUT_FREQUENCY),
  54           compressor(OUTPUT_FREQUENCY)
  55 {
  56         locut.init(FILTER_HPF, 2);
  57
  58         set_locut_enabled(global_flags.locut_enabled);
  59         set_gain_staging_db(global_flags.initial_gain_staging_db);
  60         set_gain_staging_auto(global_flags.gain_staging_auto);
  61         set_compressor_enabled(global_flags.compressor_enabled);
  62         set_limiter_enabled(global_flags.limiter_enabled);
  63         set_final_makeup_gain_auto(global_flags.final_makeup_gain_auto);
  64
  65         // Generate a very simple, default input mapping.
  66         InputMapping::Input input;
  67         input.name = "Main";
  68         input.input_source_type = InputSourceType::CAPTURE_CARD;
  69         input.input_source_index = 0;
  70         input.source_channel[0] = 0;
  71         input.source_channel[1] = 1;
  72
  73         InputMapping new_input_mapping;
  74         new_input_mapping.inputs.push_back(input);
  75         set_input_mapping(new_input_mapping);
  76 }
  77
  78 void AudioMixer::reset_card(unsigned card_index)
  79 {
  80         lock_guard<mutex> lock(audio_mutex);
  81         reset_card_mutex_held(card_index);
  82 }
  83
  84 void AudioMixer::reset_card_mutex_held(unsigned card_index)
  85 {
  86         CaptureCard *card = &cards[card_index];
  87         if (card->interesting_channels.empty()) {
  88                 card->resampling_queue.reset();
  89         } else {
  90                 card->resampling_queue.reset(new ResamplingQueue(card_index, OUTPUT_FREQUENCY, OUTPUT_FREQUENCY, card->interesting_channels.size()));
  91         }
  92         card->next_local_pts = 0;
  93 }
  94
  95 void AudioMixer::add_audio(unsigned card_index, const uint8_t *data, unsigned num_samples, AudioFormat audio_format, int64_t frame_length)
  96 {
  97         lock_guard<mutex> lock(audio_mutex);
  98         CaptureCard *card = &cards[card_index];
  99
 100         if (card->resampling_queue == nullptr) {
 101                 // No inputs use this card; throw it away.
 102                 return;
 103         }
 104
 105         unsigned num_channels = card->interesting_channels.size();
 106         assert(num_channels > 0);
 107
 108         // Convert the audio to stereo fp32.
 109         // FIXME: Pick out the right channels; this takes the first ones.
 110         vector<float> audio;
 111         audio.resize(num_samples * num_channels);
 112         switch (audio_format.bits_per_sample) {
 113         case 0:
 114                 assert(num_samples == 0);
 115                 break;
 116         case 24:
 117                 convert_fixed24_to_fp32(&audio[0], num_channels, data, audio_format.num_channels, num_samples);
 118                 break;
 119         case 32:
 120                 convert_fixed32_to_fp32(&audio[0], num_channels, data, audio_format.num_channels, num_samples);
 121                 break;
 122         default:
 123                 fprintf(stderr, "Cannot handle audio with %u bits per sample\n", audio_format.bits_per_sample);
 124                 assert(false);
 125         }
 126
 127         // Now add it.
 128         int64_t local_pts = card->next_local_pts;
 129         card->resampling_queue->add_input_samples(local_pts / double(TIMEBASE), audio.data(), num_samples);
 130         card->next_local_pts = local_pts + frame_length;
 131 }
 132
 133 void AudioMixer::add_silence(unsigned card_index, unsigned samples_per_frame, unsigned num_frames, int64_t frame_length)
 134 {
 135         CaptureCard *card = &cards[card_index];
 136         lock_guard<mutex> lock(audio_mutex);
 137
 138         if (card->resampling_queue == nullptr) {
 139                 // No inputs use this card; throw it away.
 140                 return;
 141         }
 142
 143         unsigned num_channels = card->interesting_channels.size();
 144         assert(num_channels > 0);
 145
 146         vector<float> silence(samples_per_frame * num_channels, 0.0f);
 147         for (unsigned i = 0; i < num_frames; ++i) {
 148                 card->resampling_queue->add_input_samples(card->next_local_pts / double(TIMEBASE), silence.data(), samples_per_frame);
 149                 // Note that if the format changed in the meantime, we have
 150                 // no way of detecting that; we just have to assume the frame length
 151                 // is always the same.
 152                 card->next_local_pts += frame_length;
 153         }
 154 }
 155
 156 void AudioMixer::find_sample_src_from_capture_card(const vector<float> *samples_card, unsigned card_index, int source_channel, const float **srcptr, unsigned *stride)
 157 {
 158         static float zero = 0.0f;
 159         if (source_channel == -1) {
 160                 *srcptr = &zero;
 161                 *stride = 0;
 162                 return;
 163         }
 164         // FIXME: map back through the interesting_channels squeeze map instead of using source_channel
 165         // directly, which will be wrong (and might even overrun).
 166         *srcptr = &samples_card[card_index][source_channel];
 167         *stride = cards[card_index].interesting_channels.size();
 168 }
 169
 170 vector<float> AudioMixer::get_output(double pts, unsigned num_samples, ResamplingQueue::RateAdjustmentPolicy rate_adjustment_policy)
 171 {
 172         vector<float> samples_card[MAX_CARDS];
 173         vector<float> samples_bus;
 174
 175         lock_guard<mutex> lock(audio_mutex);
 176
 177         // Pick out all the interesting channels from all the cards.
 178         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 179                 CaptureCard *card = &cards[card_index];
 180                 if (!card->interesting_channels.empty()) {
 181                         samples_card[card_index].resize(num_samples * card->interesting_channels.size());
 182                         card->resampling_queue->get_output_samples(
 183                                 pts,
 184                                 &samples_card[card_index][0],
 185                                 num_samples,
 186                                 rate_adjustment_policy);
 187                 }
 188         }
 189
 190         // TODO: Move lo-cut etc. into each bus.
 191         vector<float> samples_out;
 192         samples_out.resize(num_samples * 2);
 193         samples_bus.resize(num_samples * 2);
 194         for (unsigned input_index = 0; input_index < input_mapping.inputs.size(); ++input_index) {
 195                 const InputMapping::Input &input = input_mapping.inputs[input_index];
 196                 if (input.input_source_type == InputSourceType::SILENCE) {
 197                         memset(&samples_bus[0], 0, samples_bus.size() * sizeof(samples_bus[0]));
 198                 } else {
 199                         // TODO: Move this into its own function. Can be SSSE3-optimized if need be.
 200                         assert(input.input_source_type == InputSourceType::CAPTURE_CARD);
 201                         const float *lsrc, *rsrc;
 202                         unsigned lstride, rstride;
 203                         float *dptr = &samples_bus[0];
 204                         find_sample_src_from_capture_card(samples_card, input.input_source_index, input.source_channel[0], &lsrc, &lstride);
 205                         find_sample_src_from_capture_card(samples_card, input.input_source_index, input.source_channel[1], &rsrc, &rstride);
 206                         for (unsigned i = 0; i < num_samples; ++i) {
 207                                 *dptr++ = *lsrc;
 208                                 *dptr++ = *rsrc;
 209                                 lsrc += lstride;
 210                                 rsrc += rstride;
 211                         }
 212                 }
 213
 214                 float volume = from_db(fader_volume_db[input_index]);
 215                 if (input_index == 0) {
 216                         for (unsigned i = 0; i < num_samples * 2; ++i) {
 217                                 samples_out[i] = samples_bus[i] * volume;
 218                         }
 219                 } else {
 220                         for (unsigned i = 0; i < num_samples * 2; ++i) {
 221                                 samples_out[i] += samples_bus[i] * volume;
 222                         }
 223                 }
 224         }
 225
 226         // Cut away everything under 120 Hz (or whatever the cutoff is);
 227         // we don't need it for voice, and it will reduce headroom
 228         // and confuse the compressor. (In particular, any hums at 50 or 60 Hz
 229         // should be dampened.)
 230         if (locut_enabled) {
 231                 locut.render(samples_out.data(), samples_out.size() / 2, locut_cutoff_hz * 2.0 * M_PI / OUTPUT_FREQUENCY, 0.5f);
 232         }
 233
 234         {
 235                 lock_guard<mutex> lock(compressor_mutex);
 236
 237                 // Apply a level compressor to get the general level right.
 238                 // Basically, if it's over about -40 dBFS, we squeeze it down to that level
 239                 // (or more precisely, near it, since we don't use infinite ratio),
 240                 // then apply a makeup gain to get it to -14 dBFS. -14 dBFS is, of course,
 241                 // entirely arbitrary, but from practical tests with speech, it seems to
 242                 // put ut around -23 LUFS, so it's a reasonable starting point for later use.
 243                 {
 244                         if (level_compressor_enabled) {
 245                                 float threshold = 0.01f;   // -40 dBFS.
 246                                 float ratio = 20.0f;
 247                                 float attack_time = 0.5f;
 248                                 float release_time = 20.0f;
 249                                 float makeup_gain = from_db(ref_level_dbfs - (-40.0f));  // +26 dB.
 250                                 level_compressor.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 251                                 gain_staging_db = to_db(level_compressor.get_attenuation() * makeup_gain);
 252                         } else {
 253                                 // Just apply the gain we already had.
 254                                 float g = from_db(gain_staging_db);
 255                                 for (size_t i = 0; i < samples_out.size(); ++i) {
 256                                         samples_out[i] *= g;
 257                                 }
 258                         }
 259                 }
 260
 261         #if 0
 262                 printf("level=%f (%+5.2f dBFS) attenuation=%f (%+5.2f dB) end_result=%+5.2f dB\n",
 263                         level_compressor.get_level(), to_db(level_compressor.get_level()),
 264                         level_compressor.get_attenuation(), to_db(level_compressor.get_attenuation()),
 265                         to_db(level_compressor.get_level() * level_compressor.get_attenuation() * makeup_gain));
 266         #endif
 267
 268         //      float limiter_att, compressor_att;
 269
 270                 // The real compressor.
 271                 if (compressor_enabled) {
 272                         float threshold = from_db(compressor_threshold_dbfs);
 273                         float ratio = 20.0f;
 274                         float attack_time = 0.005f;
 275                         float release_time = 0.040f;
 276                         float makeup_gain = 2.0f;  // +6 dB.
 277                         compressor.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 278         //              compressor_att = compressor.get_attenuation();
 279                 }
 280
 281                 // Finally a limiter at -4 dB (so, -10 dBFS) to take out the worst peaks only.
 282                 // Note that since ratio is not infinite, we could go slightly higher than this.
 283                 if (limiter_enabled) {
 284                         float threshold = from_db(limiter_threshold_dbfs);
 285                         float ratio = 30.0f;
 286                         float attack_time = 0.0f;  // Instant.
 287                         float release_time = 0.020f;
 288                         float makeup_gain = 1.0f;  // 0 dB.
 289                         limiter.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 290         //              limiter_att = limiter.get_attenuation();
 291                 }
 292
 293         //      printf("limiter=%+5.1f  compressor=%+5.1f\n", to_db(limiter_att), to_db(compressor_att));
 294         }
 295
 296         // At this point, we are most likely close to +0 LU, but all of our
 297         // measurements have been on raw sample values, not R128 values.
 298         // So we have a final makeup gain to get us to +0 LU; the gain
 299         // adjustments required should be relatively small, and also, the
 300         // offset shouldn't change much (only if the type of audio changes
 301         // significantly). Thus, we shoot for updating this value basically
 302         // “whenever we process buffers”, since the R128 calculation isn't exactly
 303         // something we get out per-sample.
 304         //
 305         // Note that there's a feedback loop here, so we choose a very slow filter
 306         // (half-time of 30 seconds).
 307         double target_loudness_factor, alpha;
 308         double loudness_lu = loudness_lufs - ref_level_lufs;
 309         double current_makeup_lu = to_db(final_makeup_gain);
 310         target_loudness_factor = final_makeup_gain * from_db(-loudness_lu);
 311
 312         // If we're outside +/- 5 LU uncorrected, we don't count it as
 313         // a normal signal (probably silence) and don't change the
 314         // correction factor; just apply what we already have.
 315         if (fabs(loudness_lu - current_makeup_lu) >= 5.0 || !final_makeup_gain_auto) {
 316                 alpha = 0.0;
 317         } else {
 318                 // Formula adapted from
 319                 // https://en.wikipedia.org/wiki/Low-pass_filter#Simple_infinite_impulse_response_filter.
 320                 const double half_time_s = 30.0;
 321                 const double fc_mul_2pi_delta_t = 1.0 / (half_time_s * OUTPUT_FREQUENCY);
 322                 alpha = fc_mul_2pi_delta_t / (fc_mul_2pi_delta_t + 1.0);
 323         }
 324
 325         {
 326                 lock_guard<mutex> lock(compressor_mutex);
 327                 double m = final_makeup_gain;
 328                 for (size_t i = 0; i < samples_out.size(); i += 2) {
 329                         samples_out[i + 0] *= m;
 330                         samples_out[i + 1] *= m;
 331                         m += (target_loudness_factor - m) * alpha;
 332                 }
 333                 final_makeup_gain = m;
 334         }
 335
 336         return samples_out;
 337 }
 338
 339 vector<string> AudioMixer::get_names() const
 340 {
 341         lock_guard<mutex> lock(audio_mutex);
 342         vector<string> names;
 343         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 344                 const CaptureCard *card = &cards[card_index];
 345                 names.push_back(card->name);
 346         }
 347         return names;
 348 }
 349
 350 void AudioMixer::set_name(unsigned card_index, const string &name)
 351 {
 352         lock_guard<mutex> lock(audio_mutex);
 353         CaptureCard *card = &cards[card_index];
 354         card->name = name;
 355 }
 356
 357 void AudioMixer::set_input_mapping(const InputMapping &new_input_mapping)
 358 {
 359         lock_guard<mutex> lock(audio_mutex);
 360
 361         map<unsigned, set<unsigned>> interesting_channels;
 362         for (const InputMapping::Input &input : new_input_mapping.inputs) {
 363                 if (input.input_source_type == InputSourceType::CAPTURE_CARD) {
 364                         for (unsigned channel = 0; channel < 2; ++channel) {
 365                                 if (input.source_channel[channel] != -1) {
 366                                         interesting_channels[input.input_source_index].insert(input.source_channel[channel]);
 367                                 }
 368                         }
 369                 }
 370         }
 371
 372         // Reset resamplers for all cards that don't have the exact same state as before.
 373         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 374                 CaptureCard *card = &cards[card_index];
 375                 if (card->interesting_channels != interesting_channels[card_index]) {
 376                         card->interesting_channels = interesting_channels[card_index];
 377                         reset_card_mutex_held(card_index);
 378                 }
 379         }
 380
 381         input_mapping = new_input_mapping;
 382 }
 383
 384 InputMapping AudioMixer::get_input_mapping() const
 385 {
 386         lock_guard<mutex> lock(audio_mutex);
 387         return input_mapping;
 388 }