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
  64 void AudioMixer::reset_card(unsigned card_index)
  65 {
  66         CaptureCard *card = &cards[card_index];
  67
  68         unique_lock<mutex> lock(card->audio_mutex);
  69         card->resampling_queue.reset(new ResamplingQueue(card_index, OUTPUT_FREQUENCY, OUTPUT_FREQUENCY, 2));
  70         card->next_local_pts = 0;
  71 }
  72
  73 void AudioMixer::add_audio(unsigned card_index, const uint8_t *data, unsigned num_samples, AudioFormat audio_format, int64_t frame_length)
  74 {
  75         CaptureCard *card = &cards[card_index];
  76
  77         // Convert the audio to stereo fp32.
  78         vector<float> audio;
  79         audio.resize(num_samples * 2);
  80         switch (audio_format.bits_per_sample) {
  81         case 0:
  82                 assert(num_samples == 0);
  83                 break;
  84         case 24:
  85                 convert_fixed24_to_fp32(&audio[0], 2, data, audio_format.num_channels, num_samples);
  86                 break;
  87         case 32:
  88                 convert_fixed32_to_fp32(&audio[0], 2, data, audio_format.num_channels, num_samples);
  89                 break;
  90         default:
  91                 fprintf(stderr, "Cannot handle audio with %u bits per sample\n", audio_format.bits_per_sample);
  92                 assert(false);
  93         }
  94
  95         // Now add it.
  96         {
  97                 unique_lock<mutex> lock(card->audio_mutex);
  98
  99                 int64_t local_pts = card->next_local_pts;
 100                 card->resampling_queue->add_input_samples(local_pts / double(TIMEBASE), audio.data(), num_samples);
 101                 card->next_local_pts = local_pts + frame_length;
 102         }
 103 }
 104
 105 void AudioMixer::add_silence(unsigned card_index, unsigned samples_per_frame, unsigned num_frames, int64_t frame_length)
 106 {
 107         CaptureCard *card = &cards[card_index];
 108         unique_lock<mutex> lock(card->audio_mutex);
 109
 110         vector<float> silence(samples_per_frame * 2, 0.0f);
 111         for (unsigned i = 0; i < num_frames; ++i) {
 112                 card->resampling_queue->add_input_samples(card->next_local_pts / double(TIMEBASE), silence.data(), samples_per_frame);
 113                 // Note that if the format changed in the meantime, we have
 114                 // no way of detecting that; we just have to assume the frame length
 115                 // is always the same.
 116                 card->next_local_pts += frame_length;
 117         }
 118 }
 119
 120 vector<float> AudioMixer::get_output(double pts, unsigned num_samples, ResamplingQueue::RateAdjustmentPolicy rate_adjustment_policy)
 121 {
 122         vector<float> samples_card;
 123         vector<float> samples_out;
 124         samples_out.resize(num_samples * 2);
 125
 126         // TODO: Allow more flexible input mapping.
 127         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 128                 samples_card.resize(num_samples * 2);
 129                 {
 130                         unique_lock<mutex> lock(cards[card_index].audio_mutex);
 131                         cards[card_index].resampling_queue->get_output_samples(
 132                                 pts,
 133                                 &samples_card[0],
 134                                 num_samples,
 135                                 rate_adjustment_policy);
 136                 }
 137                 if (card_index == 0) {
 138                         for (unsigned i = 0; i < num_samples * 2; ++i) {
 139                                 samples_out[i] = samples_card[i];
 140                         }
 141                 } else {
 142                         for (unsigned i = 0; i < num_samples * 2; ++i) {
 143                                 samples_out[i] += samples_card[i];
 144                         }
 145                 }
 146         }
 147
 148         // Cut away everything under 120 Hz (or whatever the cutoff is);
 149         // we don't need it for voice, and it will reduce headroom
 150         // and confuse the compressor. (In particular, any hums at 50 or 60 Hz
 151         // should be dampened.)
 152         if (locut_enabled) {
 153                 locut.render(samples_out.data(), samples_out.size() / 2, locut_cutoff_hz * 2.0 * M_PI / OUTPUT_FREQUENCY, 0.5f);
 154         }
 155
 156         {
 157                 unique_lock<mutex> lock(compressor_mutex);
 158
 159                 // Apply a level compressor to get the general level right.
 160                 // Basically, if it's over about -40 dBFS, we squeeze it down to that level
 161                 // (or more precisely, near it, since we don't use infinite ratio),
 162                 // then apply a makeup gain to get it to -14 dBFS. -14 dBFS is, of course,
 163                 // entirely arbitrary, but from practical tests with speech, it seems to
 164                 // put ut around -23 LUFS, so it's a reasonable starting point for later use.
 165                 {
 166                         if (level_compressor_enabled) {
 167                                 float threshold = 0.01f;   // -40 dBFS.
 168                                 float ratio = 20.0f;
 169                                 float attack_time = 0.5f;
 170                                 float release_time = 20.0f;
 171                                 float makeup_gain = from_db(ref_level_dbfs - (-40.0f));  // +26 dB.
 172                                 level_compressor.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 173                                 gain_staging_db = to_db(level_compressor.get_attenuation() * makeup_gain);
 174                         } else {
 175                                 // Just apply the gain we already had.
 176                                 float g = from_db(gain_staging_db);
 177                                 for (size_t i = 0; i < samples_out.size(); ++i) {
 178                                         samples_out[i] *= g;
 179                                 }
 180                         }
 181                 }
 182
 183         #if 0
 184                 printf("level=%f (%+5.2f dBFS) attenuation=%f (%+5.2f dB) end_result=%+5.2f dB\n",
 185                         level_compressor.get_level(), to_db(level_compressor.get_level()),
 186                         level_compressor.get_attenuation(), to_db(level_compressor.get_attenuation()),
 187                         to_db(level_compressor.get_level() * level_compressor.get_attenuation() * makeup_gain));
 188         #endif
 189
 190         //      float limiter_att, compressor_att;
 191
 192                 // The real compressor.
 193                 if (compressor_enabled) {
 194                         float threshold = from_db(compressor_threshold_dbfs);
 195                         float ratio = 20.0f;
 196                         float attack_time = 0.005f;
 197                         float release_time = 0.040f;
 198                         float makeup_gain = 2.0f;  // +6 dB.
 199                         compressor.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 200         //              compressor_att = compressor.get_attenuation();
 201                 }
 202
 203                 // Finally a limiter at -4 dB (so, -10 dBFS) to take out the worst peaks only.
 204                 // Note that since ratio is not infinite, we could go slightly higher than this.
 205                 if (limiter_enabled) {
 206                         float threshold = from_db(limiter_threshold_dbfs);
 207                         float ratio = 30.0f;
 208                         float attack_time = 0.0f;  // Instant.
 209                         float release_time = 0.020f;
 210                         float makeup_gain = 1.0f;  // 0 dB.
 211                         limiter.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 212         //              limiter_att = limiter.get_attenuation();
 213                 }
 214
 215         //      printf("limiter=%+5.1f  compressor=%+5.1f\n", to_db(limiter_att), to_db(compressor_att));
 216         }
 217
 218         // At this point, we are most likely close to +0 LU, but all of our
 219         // measurements have been on raw sample values, not R128 values.
 220         // So we have a final makeup gain to get us to +0 LU; the gain
 221         // adjustments required should be relatively small, and also, the
 222         // offset shouldn't change much (only if the type of audio changes
 223         // significantly). Thus, we shoot for updating this value basically
 224         // “whenever we process buffers”, since the R128 calculation isn't exactly
 225         // something we get out per-sample.
 226         //
 227         // Note that there's a feedback loop here, so we choose a very slow filter
 228         // (half-time of 100 seconds).
 229         double target_loudness_factor, alpha;
 230         double loudness_lu = loudness_lufs - ref_level_lufs;
 231         double current_makeup_lu = to_db(final_makeup_gain);
 232         target_loudness_factor = from_db(-loudness_lu);
 233
 234         // If we're outside +/- 5 LU uncorrected, we don't count it as
 235         // a normal signal (probably silence) and don't change the
 236         // correction factor; just apply what we already have.
 237         if (fabs(loudness_lu - current_makeup_lu) >= 5.0 || !final_makeup_gain_auto) {
 238                 alpha = 0.0;
 239         } else {
 240                 // Formula adapted from
 241                 // https://en.wikipedia.org/wiki/Low-pass_filter#Simple_infinite_impulse_response_filter.
 242                 const double half_time_s = 100.0;
 243                 const double fc_mul_2pi_delta_t = 1.0 / (half_time_s * OUTPUT_FREQUENCY);
 244                 alpha = fc_mul_2pi_delta_t / (fc_mul_2pi_delta_t + 1.0);
 245         }
 246
 247         {
 248                 unique_lock<mutex> lock(compressor_mutex);
 249                 double m = final_makeup_gain;
 250                 for (size_t i = 0; i < samples_out.size(); i += 2) {
 251                         samples_out[i + 0] *= m;
 252                         samples_out[i + 1] *= m;
 253                         m += (target_loudness_factor - m) * alpha;
 254                 }
 255                 final_makeup_gain = m;
 256         }
 257
 258         return samples_out;
 259 }