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 void convert_fixed24_to_fp32(float *dst, size_t out_channels, const uint8_t *src, size_t in_channels, size_t num_samples)
  19 {
  20         assert(in_channels >= out_channels);
  21         for (size_t i = 0; i < num_samples; ++i) {
  22                 for (size_t j = 0; j < out_channels; ++j) {
  23                         uint32_t s1 = *src++;
  24                         uint32_t s2 = *src++;
  25                         uint32_t s3 = *src++;
  26                         uint32_t s = s1 | (s1 << 8) | (s2 << 16) | (s3 << 24);
  27                         dst[i * out_channels + j] = int(s) * (1.0f / 2147483648.0f);
  28                 }
  29                 src += 3 * (in_channels - out_channels);
  30         }
  31 }
  32
  33 void convert_fixed32_to_fp32(float *dst, size_t out_channels, const uint8_t *src, size_t in_channels, size_t num_samples)
  34 {
  35         assert(in_channels >= out_channels);
  36         for (size_t i = 0; i < num_samples; ++i) {
  37                 for (size_t j = 0; j < out_channels; ++j) {
  38                         int32_t s = le32toh(*(int32_t *)src);
  39                         dst[i * out_channels + j] = s * (1.0f / 2147483648.0f);
  40                         src += 4;
  41                 }
  42                 src += 4 * (in_channels - out_channels);
  43         }
  44 }
  45
  46 }  // namespace
  47
  48 AudioMixer::AudioMixer(unsigned num_cards)
  49         : num_cards(num_cards),
  50           level_compressor(OUTPUT_FREQUENCY),
  51           limiter(OUTPUT_FREQUENCY),
  52           compressor(OUTPUT_FREQUENCY)
  53 {
  54         locut.init(FILTER_HPF, 2);
  55
  56         set_locut_enabled(global_flags.locut_enabled);
  57         set_gain_staging_db(global_flags.initial_gain_staging_db);
  58         set_gain_staging_auto(global_flags.gain_staging_auto);
  59         set_compressor_enabled(global_flags.compressor_enabled);
  60         set_limiter_enabled(global_flags.limiter_enabled);
  61         set_final_makeup_gain_auto(global_flags.final_makeup_gain_auto);
  62
  63         // Generate a very simple, default input mapping.
  64         InputMapping::Input input;
  65         input.name = "Main";
  66         input.input_source_type = InputSourceType::CAPTURE_CARD;
  67         input.input_source_index = 0;
  68         input.source_channel[0] = 0;
  69         input.source_channel[1] = 1;
  70
  71         input_mapping.inputs.push_back(input);
  72 }
  73
  74 void AudioMixer::reset_card(unsigned card_index)
  75 {
  76         CaptureCard *card = &cards[card_index];
  77
  78         unique_lock<mutex> lock(card->audio_mutex);
  79         card->resampling_queue.reset(new ResamplingQueue(card_index, OUTPUT_FREQUENCY, OUTPUT_FREQUENCY, 2));
  80         card->next_local_pts = 0;
  81 }
  82
  83 void AudioMixer::add_audio(unsigned card_index, const uint8_t *data, unsigned num_samples, AudioFormat audio_format, int64_t frame_length)
  84 {
  85         CaptureCard *card = &cards[card_index];
  86
  87         // Convert the audio to stereo fp32.
  88         vector<float> audio;
  89         audio.resize(num_samples * 2);
  90         switch (audio_format.bits_per_sample) {
  91         case 0:
  92                 assert(num_samples == 0);
  93                 break;
  94         case 24:
  95                 convert_fixed24_to_fp32(&audio[0], 2, data, audio_format.num_channels, num_samples);
  96                 break;
  97         case 32:
  98                 convert_fixed32_to_fp32(&audio[0], 2, data, audio_format.num_channels, num_samples);
  99                 break;
 100         default:
 101                 fprintf(stderr, "Cannot handle audio with %u bits per sample\n", audio_format.bits_per_sample);
 102                 assert(false);
 103         }
 104
 105         // Now add it.
 106         {
 107                 unique_lock<mutex> lock(card->audio_mutex);
 108
 109                 int64_t local_pts = card->next_local_pts;
 110                 card->resampling_queue->add_input_samples(local_pts / double(TIMEBASE), audio.data(), num_samples);
 111                 card->next_local_pts = local_pts + frame_length;
 112         }
 113 }
 114
 115 void AudioMixer::add_silence(unsigned card_index, unsigned samples_per_frame, unsigned num_frames, int64_t frame_length)
 116 {
 117         CaptureCard *card = &cards[card_index];
 118         unique_lock<mutex> lock(card->audio_mutex);
 119
 120         vector<float> silence(samples_per_frame * 2, 0.0f);
 121         for (unsigned i = 0; i < num_frames; ++i) {
 122                 card->resampling_queue->add_input_samples(card->next_local_pts / double(TIMEBASE), silence.data(), samples_per_frame);
 123                 // Note that if the format changed in the meantime, we have
 124                 // no way of detecting that; we just have to assume the frame length
 125                 // is always the same.
 126                 card->next_local_pts += frame_length;
 127         }
 128 }
 129
 130 vector<float> AudioMixer::get_output(double pts, unsigned num_samples, ResamplingQueue::RateAdjustmentPolicy rate_adjustment_policy)
 131 {
 132         vector<float> samples_card;
 133         vector<float> samples_out;
 134         samples_out.resize(num_samples * 2);
 135
 136         // TODO: Allow more flexible input mapping.
 137         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 138                 samples_card.resize(num_samples * 2);
 139                 {
 140                         unique_lock<mutex> lock(cards[card_index].audio_mutex);
 141                         cards[card_index].resampling_queue->get_output_samples(
 142                                 pts,
 143                                 &samples_card[0],
 144                                 num_samples,
 145                                 rate_adjustment_policy);
 146                 }
 147
 148                 float volume = from_db(cards[card_index].fader_volume_db);
 149                 if (card_index == 0) {
 150                         for (unsigned i = 0; i < num_samples * 2; ++i) {
 151                                 samples_out[i] = samples_card[i] * volume;
 152                         }
 153                 } else {
 154                         for (unsigned i = 0; i < num_samples * 2; ++i) {
 155                                 samples_out[i] += samples_card[i] * volume;
 156                         }
 157                 }
 158         }
 159
 160         // Cut away everything under 120 Hz (or whatever the cutoff is);
 161         // we don't need it for voice, and it will reduce headroom
 162         // and confuse the compressor. (In particular, any hums at 50 or 60 Hz
 163         // should be dampened.)
 164         if (locut_enabled) {
 165                 locut.render(samples_out.data(), samples_out.size() / 2, locut_cutoff_hz * 2.0 * M_PI / OUTPUT_FREQUENCY, 0.5f);
 166         }
 167
 168         {
 169                 unique_lock<mutex> lock(compressor_mutex);
 170
 171                 // Apply a level compressor to get the general level right.
 172                 // Basically, if it's over about -40 dBFS, we squeeze it down to that level
 173                 // (or more precisely, near it, since we don't use infinite ratio),
 174                 // then apply a makeup gain to get it to -14 dBFS. -14 dBFS is, of course,
 175                 // entirely arbitrary, but from practical tests with speech, it seems to
 176                 // put ut around -23 LUFS, so it's a reasonable starting point for later use.
 177                 {
 178                         if (level_compressor_enabled) {
 179                                 float threshold = 0.01f;   // -40 dBFS.
 180                                 float ratio = 20.0f;
 181                                 float attack_time = 0.5f;
 182                                 float release_time = 20.0f;
 183                                 float makeup_gain = from_db(ref_level_dbfs - (-40.0f));  // +26 dB.
 184                                 level_compressor.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 185                                 gain_staging_db = to_db(level_compressor.get_attenuation() * makeup_gain);
 186                         } else {
 187                                 // Just apply the gain we already had.
 188                                 float g = from_db(gain_staging_db);
 189                                 for (size_t i = 0; i < samples_out.size(); ++i) {
 190                                         samples_out[i] *= g;
 191                                 }
 192                         }
 193                 }
 194
 195         #if 0
 196                 printf("level=%f (%+5.2f dBFS) attenuation=%f (%+5.2f dB) end_result=%+5.2f dB\n",
 197                         level_compressor.get_level(), to_db(level_compressor.get_level()),
 198                         level_compressor.get_attenuation(), to_db(level_compressor.get_attenuation()),
 199                         to_db(level_compressor.get_level() * level_compressor.get_attenuation() * makeup_gain));
 200         #endif
 201
 202         //      float limiter_att, compressor_att;
 203
 204                 // The real compressor.
 205                 if (compressor_enabled) {
 206                         float threshold = from_db(compressor_threshold_dbfs);
 207                         float ratio = 20.0f;
 208                         float attack_time = 0.005f;
 209                         float release_time = 0.040f;
 210                         float makeup_gain = 2.0f;  // +6 dB.
 211                         compressor.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 212         //              compressor_att = compressor.get_attenuation();
 213                 }
 214
 215                 // Finally a limiter at -4 dB (so, -10 dBFS) to take out the worst peaks only.
 216                 // Note that since ratio is not infinite, we could go slightly higher than this.
 217                 if (limiter_enabled) {
 218                         float threshold = from_db(limiter_threshold_dbfs);
 219                         float ratio = 30.0f;
 220                         float attack_time = 0.0f;  // Instant.
 221                         float release_time = 0.020f;
 222                         float makeup_gain = 1.0f;  // 0 dB.
 223                         limiter.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 224         //              limiter_att = limiter.get_attenuation();
 225                 }
 226
 227         //      printf("limiter=%+5.1f  compressor=%+5.1f\n", to_db(limiter_att), to_db(compressor_att));
 228         }
 229
 230         // At this point, we are most likely close to +0 LU, but all of our
 231         // measurements have been on raw sample values, not R128 values.
 232         // So we have a final makeup gain to get us to +0 LU; the gain
 233         // adjustments required should be relatively small, and also, the
 234         // offset shouldn't change much (only if the type of audio changes
 235         // significantly). Thus, we shoot for updating this value basically
 236         // “whenever we process buffers”, since the R128 calculation isn't exactly
 237         // something we get out per-sample.
 238         //
 239         // Note that there's a feedback loop here, so we choose a very slow filter
 240         // (half-time of 30 seconds).
 241         double target_loudness_factor, alpha;
 242         double loudness_lu = loudness_lufs - ref_level_lufs;
 243         double current_makeup_lu = to_db(final_makeup_gain);
 244         target_loudness_factor = final_makeup_gain * from_db(-loudness_lu);
 245
 246         // If we're outside +/- 5 LU uncorrected, we don't count it as
 247         // a normal signal (probably silence) and don't change the
 248         // correction factor; just apply what we already have.
 249         if (fabs(loudness_lu - current_makeup_lu) >= 5.0 || !final_makeup_gain_auto) {
 250                 alpha = 0.0;
 251         } else {
 252                 // Formula adapted from
 253                 // https://en.wikipedia.org/wiki/Low-pass_filter#Simple_infinite_impulse_response_filter.
 254                 const double half_time_s = 30.0;
 255                 const double fc_mul_2pi_delta_t = 1.0 / (half_time_s * OUTPUT_FREQUENCY);
 256                 alpha = fc_mul_2pi_delta_t / (fc_mul_2pi_delta_t + 1.0);
 257         }
 258
 259         {
 260                 unique_lock<mutex> lock(compressor_mutex);
 261                 double m = final_makeup_gain;
 262                 for (size_t i = 0; i < samples_out.size(); i += 2) {
 263                         samples_out[i + 0] *= m;
 264                         samples_out[i + 1] *= m;
 265                         m += (target_loudness_factor - m) * alpha;
 266                 }
 267                 final_makeup_gain = m;
 268         }
 269
 270         return samples_out;
 271 }
 272
 273 vector<string> AudioMixer::get_names() const
 274 {
 275         vector<string> names;
 276         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 277                 const CaptureCard *card = &cards[card_index];
 278                 unique_lock<mutex> lock(card->audio_mutex);
 279                 names.push_back(card->name);
 280         }
 281         return names;
 282 }
 283
 284 void AudioMixer::set_name(unsigned card_index, const string &name)
 285 {
 286         CaptureCard *card = &cards[card_index];
 287         unique_lock<mutex> lock(card->audio_mutex);
 288         card->name = name;
 289 }
 290
 291 void AudioMixer::set_input_mapping(const InputMapping &input_mapping)
 292 {
 293         lock_guard<mutex> lock(mapping_mutex);
 294         this->input_mapping = input_mapping;
 295 }
 296
 297 InputMapping AudioMixer::get_input_mapping() const
 298 {
 299         lock_guard<mutex> lock(mapping_mutex);
 300         return input_mapping;
 301 }