1 #include "audio_mixer.h"
4 #include <bmusb/bmusb.h>
27 using namespace bmusb;
29 using namespace std::chrono;
30 using namespace std::placeholders;
34 // TODO: If these prove to be a bottleneck, they can be SSSE3-optimized
35 // (usually including multiple channels at a time).
37 void convert_fixed16_to_fp32(float *dst, size_t out_channel, size_t out_num_channels,
38 const uint8_t *src, size_t in_channel, size_t in_num_channels,
41 assert(in_channel < in_num_channels);
42 assert(out_channel < out_num_channels);
43 src += in_channel * 2;
46 for (size_t i = 0; i < num_samples; ++i) {
47 int16_t s = le16toh(*(int16_t *)src);
48 *dst = s * (1.0f / 32768.0f);
50 src += 2 * in_num_channels;
51 dst += out_num_channels;
55 void convert_fixed24_to_fp32(float *dst, size_t out_channel, size_t out_num_channels,
56 const uint8_t *src, size_t in_channel, size_t in_num_channels,
59 assert(in_channel < in_num_channels);
60 assert(out_channel < out_num_channels);
61 src += in_channel * 3;
64 for (size_t i = 0; i < num_samples; ++i) {
68 uint32_t s = s1 | (s1 << 8) | (s2 << 16) | (s3 << 24);
69 *dst = int(s) * (1.0f / 2147483648.0f);
71 src += 3 * in_num_channels;
72 dst += out_num_channels;
76 void convert_fixed32_to_fp32(float *dst, size_t out_channel, size_t out_num_channels,
77 const uint8_t *src, size_t in_channel, size_t in_num_channels,
80 assert(in_channel < in_num_channels);
81 assert(out_channel < out_num_channels);
82 src += in_channel * 4;
85 for (size_t i = 0; i < num_samples; ++i) {
86 int32_t s = le32toh(*(int32_t *)src);
87 *dst = s * (1.0f / 2147483648.0f);
89 src += 4 * in_num_channels;
90 dst += out_num_channels;
94 float find_peak_plain(const float *samples, size_t num_samples) __attribute__((unused));
96 float find_peak_plain(const float *samples, size_t num_samples)
98 float m = fabs(samples[0]);
99 for (size_t i = 1; i < num_samples; ++i) {
100 m = max(m, fabs(samples[i]));
106 static inline float horizontal_max(__m128 m)
108 __m128 tmp = _mm_shuffle_ps(m, m, _MM_SHUFFLE(1, 0, 3, 2));
109 m = _mm_max_ps(m, tmp);
110 tmp = _mm_shuffle_ps(m, m, _MM_SHUFFLE(2, 3, 0, 1));
111 m = _mm_max_ps(m, tmp);
112 return _mm_cvtss_f32(m);
115 float find_peak(const float *samples, size_t num_samples)
117 const __m128 abs_mask = _mm_castsi128_ps(_mm_set1_epi32(0x7fffffffu));
118 __m128 m = _mm_setzero_ps();
119 for (size_t i = 0; i < (num_samples & ~3); i += 4) {
120 __m128 x = _mm_loadu_ps(samples + i);
121 x = _mm_and_ps(x, abs_mask);
122 m = _mm_max_ps(m, x);
124 float result = horizontal_max(m);
126 for (size_t i = (num_samples & ~3); i < num_samples; ++i) {
127 result = max(result, fabs(samples[i]));
131 // Self-test. We should be bit-exact the same.
132 float reference_result = find_peak_plain(samples, num_samples);
133 if (result != reference_result) {
134 fprintf(stderr, "Error: Peak is %f [%f %f %f %f]; should be %f.\n",
136 _mm_cvtss_f32(_mm_shuffle_ps(m, m, _MM_SHUFFLE(0, 0, 0, 0))),
137 _mm_cvtss_f32(_mm_shuffle_ps(m, m, _MM_SHUFFLE(1, 1, 1, 1))),
138 _mm_cvtss_f32(_mm_shuffle_ps(m, m, _MM_SHUFFLE(2, 2, 2, 2))),
139 _mm_cvtss_f32(_mm_shuffle_ps(m, m, _MM_SHUFFLE(3, 3, 3, 3))),
147 float find_peak(const float *samples, size_t num_samples)
149 return find_peak_plain(samples, num_samples);
153 void deinterleave_samples(const vector<float> &in, vector<float> *out_l, vector<float> *out_r)
155 size_t num_samples = in.size() / 2;
156 out_l->resize(num_samples);
157 out_r->resize(num_samples);
159 const float *inptr = in.data();
160 float *lptr = &(*out_l)[0];
161 float *rptr = &(*out_r)[0];
162 for (size_t i = 0; i < num_samples; ++i) {
170 AudioMixer::AudioMixer(unsigned num_cards)
171 : num_cards(num_cards),
172 limiter(OUTPUT_FREQUENCY),
173 correlation(OUTPUT_FREQUENCY)
175 for (unsigned bus_index = 0; bus_index < MAX_BUSES; ++bus_index) {
176 locut[bus_index].init(FILTER_HPF, 2);
177 eq[bus_index][EQ_BAND_BASS].init(FILTER_LOW_SHELF, 1);
178 // Note: EQ_BAND_MID isn't used (see comments in apply_eq()).
179 eq[bus_index][EQ_BAND_TREBLE].init(FILTER_HIGH_SHELF, 1);
180 compressor[bus_index].reset(new StereoCompressor(OUTPUT_FREQUENCY));
181 level_compressor[bus_index].reset(new StereoCompressor(OUTPUT_FREQUENCY));
183 set_bus_settings(bus_index, get_default_bus_settings());
185 set_limiter_enabled(global_flags.limiter_enabled);
186 set_final_makeup_gain_auto(global_flags.final_makeup_gain_auto);
188 r128.init(2, OUTPUT_FREQUENCY);
191 // hlen=16 is pretty low quality, but we use quite a bit of CPU otherwise,
192 // and there's a limit to how important the peak meter is.
193 peak_resampler.setup(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY * 4, /*num_channels=*/2, /*hlen=*/16, /*frel=*/1.0);
195 global_audio_mixer = this;
198 if (!global_flags.input_mapping_filename.empty()) {
199 // Must happen after ALSAPool is initialized, as it needs to know the card list.
200 current_mapping_mode = MappingMode::MULTICHANNEL;
201 InputMapping new_input_mapping;
202 if (!load_input_mapping_from_file(get_devices(),
203 global_flags.input_mapping_filename,
204 &new_input_mapping)) {
205 fprintf(stderr, "Failed to load input mapping from '%s', exiting.\n",
206 global_flags.input_mapping_filename.c_str());
209 set_input_mapping(new_input_mapping);
211 set_simple_input(/*card_index=*/0);
212 if (global_flags.multichannel_mapping_mode) {
213 current_mapping_mode = MappingMode::MULTICHANNEL;
217 global_metrics.add("audio_loudness_short_lufs", &metric_audio_loudness_short_lufs, Metrics::TYPE_GAUGE);
218 global_metrics.add("audio_loudness_integrated_lufs", &metric_audio_loudness_integrated_lufs, Metrics::TYPE_GAUGE);
219 global_metrics.add("audio_loudness_range_low_lufs", &metric_audio_loudness_range_low_lufs, Metrics::TYPE_GAUGE);
220 global_metrics.add("audio_loudness_range_high_lufs", &metric_audio_loudness_range_high_lufs, Metrics::TYPE_GAUGE);
221 global_metrics.add("audio_peak_dbfs", &metric_audio_peak_dbfs, Metrics::TYPE_GAUGE);
222 global_metrics.add("audio_final_makeup_gain_db", &metric_audio_final_makeup_gain_db, Metrics::TYPE_GAUGE);
223 global_metrics.add("audio_correlation", &metric_audio_correlation, Metrics::TYPE_GAUGE);
226 void AudioMixer::reset_resampler(DeviceSpec device_spec)
228 lock_guard<timed_mutex> lock(audio_mutex);
229 reset_resampler_mutex_held(device_spec);
232 void AudioMixer::reset_resampler_mutex_held(DeviceSpec device_spec)
234 AudioDevice *device = find_audio_device(device_spec);
236 if (device->interesting_channels.empty()) {
237 device->resampling_queue.reset();
239 device->resampling_queue.reset(new ResamplingQueue(
240 device_spec, device->capture_frequency, OUTPUT_FREQUENCY, device->interesting_channels.size(),
241 global_flags.audio_queue_length_ms * 0.001));
245 bool AudioMixer::add_audio(DeviceSpec device_spec, const uint8_t *data, unsigned num_samples, AudioFormat audio_format, int64_t frame_length, steady_clock::time_point frame_time)
247 AudioDevice *device = find_audio_device(device_spec);
249 unique_lock<timed_mutex> lock(audio_mutex, defer_lock);
250 if (!lock.try_lock_for(chrono::milliseconds(10))) {
253 if (device->resampling_queue == nullptr) {
254 // No buses use this device; throw it away.
258 unsigned num_channels = device->interesting_channels.size();
259 assert(num_channels > 0);
261 // Convert the audio to fp32.
262 unique_ptr<float[]> audio(new float[num_samples * num_channels]);
263 unsigned channel_index = 0;
264 for (auto channel_it = device->interesting_channels.cbegin(); channel_it != device->interesting_channels.end(); ++channel_it, ++channel_index) {
265 switch (audio_format.bits_per_sample) {
267 assert(num_samples == 0);
270 convert_fixed16_to_fp32(audio.get(), channel_index, num_channels, data, *channel_it, audio_format.num_channels, num_samples);
273 convert_fixed24_to_fp32(audio.get(), channel_index, num_channels, data, *channel_it, audio_format.num_channels, num_samples);
276 convert_fixed32_to_fp32(audio.get(), channel_index, num_channels, data, *channel_it, audio_format.num_channels, num_samples);
279 fprintf(stderr, "Cannot handle audio with %u bits per sample\n", audio_format.bits_per_sample);
284 // If we changed frequency since last frame, we'll need to reset the resampler.
285 if (audio_format.sample_rate != device->capture_frequency) {
286 device->capture_frequency = audio_format.sample_rate;
287 reset_resampler_mutex_held(device_spec);
291 device->resampling_queue->add_input_samples(frame_time, audio.get(), num_samples, ResamplingQueue::ADJUST_RATE);
295 bool AudioMixer::add_silence(DeviceSpec device_spec, unsigned samples_per_frame, unsigned num_frames, int64_t frame_length)
297 AudioDevice *device = find_audio_device(device_spec);
299 unique_lock<timed_mutex> lock(audio_mutex, defer_lock);
300 if (!lock.try_lock_for(chrono::milliseconds(10))) {
303 if (device->resampling_queue == nullptr) {
304 // No buses use this device; throw it away.
308 unsigned num_channels = device->interesting_channels.size();
309 assert(num_channels > 0);
311 vector<float> silence(samples_per_frame * num_channels, 0.0f);
312 for (unsigned i = 0; i < num_frames; ++i) {
313 device->resampling_queue->add_input_samples(steady_clock::now(), silence.data(), samples_per_frame, ResamplingQueue::DO_NOT_ADJUST_RATE);
318 bool AudioMixer::silence_card(DeviceSpec device_spec, bool silence)
320 AudioDevice *device = find_audio_device(device_spec);
322 unique_lock<timed_mutex> lock(audio_mutex, defer_lock);
323 if (!lock.try_lock_for(chrono::milliseconds(10))) {
327 if (device->silenced && !silence) {
328 reset_resampler_mutex_held(device_spec);
330 device->silenced = silence;
334 AudioMixer::BusSettings AudioMixer::get_default_bus_settings()
336 BusSettings settings;
337 settings.fader_volume_db = 0.0f;
338 settings.muted = false;
339 settings.locut_enabled = global_flags.locut_enabled;
340 for (unsigned band_index = 0; band_index < NUM_EQ_BANDS; ++band_index) {
341 settings.eq_level_db[band_index] = 0.0f;
343 settings.gain_staging_db = global_flags.initial_gain_staging_db;
344 settings.level_compressor_enabled = global_flags.gain_staging_auto;
345 settings.compressor_threshold_dbfs = ref_level_dbfs - 12.0f; // -12 dB.
346 settings.compressor_enabled = global_flags.compressor_enabled;
350 AudioMixer::BusSettings AudioMixer::get_bus_settings(unsigned bus_index) const
352 lock_guard<timed_mutex> lock(audio_mutex);
353 BusSettings settings;
354 settings.fader_volume_db = fader_volume_db[bus_index];
355 settings.muted = mute[bus_index];
356 settings.locut_enabled = locut_enabled[bus_index];
357 for (unsigned band_index = 0; band_index < NUM_EQ_BANDS; ++band_index) {
358 settings.eq_level_db[band_index] = eq_level_db[bus_index][band_index];
360 settings.gain_staging_db = gain_staging_db[bus_index];
361 settings.level_compressor_enabled = level_compressor_enabled[bus_index];
362 settings.compressor_threshold_dbfs = compressor_threshold_dbfs[bus_index];
363 settings.compressor_enabled = compressor_enabled[bus_index];
367 void AudioMixer::set_bus_settings(unsigned bus_index, const AudioMixer::BusSettings &settings)
369 lock_guard<timed_mutex> lock(audio_mutex);
370 fader_volume_db[bus_index] = settings.fader_volume_db;
371 mute[bus_index] = settings.muted;
372 locut_enabled[bus_index] = settings.locut_enabled;
373 for (unsigned band_index = 0; band_index < NUM_EQ_BANDS; ++band_index) {
374 eq_level_db[bus_index][band_index] = settings.eq_level_db[band_index];
376 gain_staging_db[bus_index] = settings.gain_staging_db;
377 last_gain_staging_db[bus_index] = gain_staging_db[bus_index];
378 level_compressor_enabled[bus_index] = settings.level_compressor_enabled;
379 compressor_threshold_dbfs[bus_index] = settings.compressor_threshold_dbfs;
380 compressor_enabled[bus_index] = settings.compressor_enabled;
383 AudioMixer::AudioDevice *AudioMixer::find_audio_device(DeviceSpec device)
385 switch (device.type) {
386 case InputSourceType::CAPTURE_CARD:
387 return &video_cards[device.index];
388 case InputSourceType::ALSA_INPUT:
389 return &alsa_inputs[device.index];
390 case InputSourceType::SILENCE:
397 // Get a pointer to the given channel from the given device.
398 // The channel must be picked out earlier and resampled.
399 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)
401 static float zero = 0.0f;
402 if (source_channel == -1 || device_spec.type == InputSourceType::SILENCE) {
407 AudioDevice *device = find_audio_device(device_spec);
408 assert(device->interesting_channels.count(source_channel) != 0);
409 unsigned channel_index = 0;
410 for (int channel : device->interesting_channels) {
411 if (channel == source_channel) break;
414 assert(channel_index < device->interesting_channels.size());
415 const auto it = samples_card.find(device_spec);
416 assert(it != samples_card.end());
417 *srcptr = &(it->second)[channel_index];
418 *stride = device->interesting_channels.size();
421 // TODO: Can be SSSE3-optimized if need be.
422 void AudioMixer::fill_audio_bus(const map<DeviceSpec, vector<float>> &samples_card, const InputMapping::Bus &bus, unsigned num_samples, float *output)
424 if (bus.device.type == InputSourceType::SILENCE) {
425 memset(output, 0, num_samples * 2 * sizeof(*output));
427 assert(bus.device.type == InputSourceType::CAPTURE_CARD ||
428 bus.device.type == InputSourceType::ALSA_INPUT);
429 const float *lsrc, *rsrc;
430 unsigned lstride, rstride;
431 float *dptr = output;
432 find_sample_src_from_device(samples_card, bus.device, bus.source_channel[0], &lsrc, &lstride);
433 find_sample_src_from_device(samples_card, bus.device, bus.source_channel[1], &rsrc, &rstride);
434 for (unsigned i = 0; i < num_samples; ++i) {
443 vector<DeviceSpec> AudioMixer::get_active_devices() const
445 vector<DeviceSpec> ret;
446 for (unsigned card_index = 0; card_index < MAX_VIDEO_CARDS; ++card_index) {
447 const DeviceSpec device_spec{InputSourceType::CAPTURE_CARD, card_index};
448 if (!find_audio_device(device_spec)->interesting_channels.empty()) {
449 ret.push_back(device_spec);
452 for (unsigned card_index = 0; card_index < MAX_ALSA_CARDS; ++card_index) {
453 const DeviceSpec device_spec{InputSourceType::ALSA_INPUT, card_index};
454 if (!find_audio_device(device_spec)->interesting_channels.empty()) {
455 ret.push_back(device_spec);
463 void apply_gain(float db, float last_db, vector<float> *samples)
465 if (fabs(db - last_db) < 1e-3) {
466 // Constant over this frame.
467 const float gain = from_db(db);
468 for (size_t i = 0; i < samples->size(); ++i) {
469 (*samples)[i] *= gain;
472 // We need to do a fade.
473 unsigned num_samples = samples->size() / 2;
474 float gain = from_db(last_db);
475 const float gain_inc = pow(from_db(db - last_db), 1.0 / num_samples);
476 for (size_t i = 0; i < num_samples; ++i) {
477 (*samples)[i * 2 + 0] *= gain;
478 (*samples)[i * 2 + 1] *= gain;
486 vector<float> AudioMixer::get_output(steady_clock::time_point ts, unsigned num_samples, ResamplingQueue::RateAdjustmentPolicy rate_adjustment_policy)
488 map<DeviceSpec, vector<float>> samples_card;
489 vector<float> samples_bus;
491 lock_guard<timed_mutex> lock(audio_mutex);
493 // Pick out all the interesting channels from all the cards.
494 for (const DeviceSpec &device_spec : get_active_devices()) {
495 AudioDevice *device = find_audio_device(device_spec);
496 samples_card[device_spec].resize(num_samples * device->interesting_channels.size());
497 if (device->silenced) {
498 memset(&samples_card[device_spec][0], 0, samples_card[device_spec].size() * sizeof(float));
500 device->resampling_queue->get_output_samples(
502 &samples_card[device_spec][0],
504 rate_adjustment_policy);
508 vector<float> samples_out, left, right;
509 samples_out.resize(num_samples * 2);
510 samples_bus.resize(num_samples * 2);
511 for (unsigned bus_index = 0; bus_index < input_mapping.buses.size(); ++bus_index) {
512 fill_audio_bus(samples_card, input_mapping.buses[bus_index], num_samples, &samples_bus[0]);
513 apply_eq(bus_index, &samples_bus);
516 lock_guard<mutex> lock(compressor_mutex);
518 // Apply a level compressor to get the general level right.
519 // Basically, if it's over about -40 dBFS, we squeeze it down to that level
520 // (or more precisely, near it, since we don't use infinite ratio),
521 // then apply a makeup gain to get it to -14 dBFS. -14 dBFS is, of course,
522 // entirely arbitrary, but from practical tests with speech, it seems to
523 // put ut around -23 LUFS, so it's a reasonable starting point for later use.
524 if (level_compressor_enabled[bus_index]) {
525 float threshold = 0.01f; // -40 dBFS.
527 float attack_time = 0.5f;
528 float release_time = 20.0f;
529 float makeup_gain = from_db(ref_level_dbfs - (-40.0f)); // +26 dB.
530 level_compressor[bus_index]->process(samples_bus.data(), samples_bus.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
531 gain_staging_db[bus_index] = to_db(level_compressor[bus_index]->get_attenuation() * makeup_gain);
533 // Just apply the gain we already had.
534 float db = gain_staging_db[bus_index];
535 float last_db = last_gain_staging_db[bus_index];
536 apply_gain(db, last_db, &samples_bus);
538 last_gain_staging_db[bus_index] = gain_staging_db[bus_index];
541 printf("level=%f (%+5.2f dBFS) attenuation=%f (%+5.2f dB) end_result=%+5.2f dB\n",
542 level_compressor.get_level(), to_db(level_compressor.get_level()),
543 level_compressor.get_attenuation(), to_db(level_compressor.get_attenuation()),
544 to_db(level_compressor.get_level() * level_compressor.get_attenuation() * makeup_gain));
547 // The real compressor.
548 if (compressor_enabled[bus_index]) {
549 float threshold = from_db(compressor_threshold_dbfs[bus_index]);
551 float attack_time = 0.005f;
552 float release_time = 0.040f;
553 float makeup_gain = 2.0f; // +6 dB.
554 compressor[bus_index]->process(samples_bus.data(), samples_bus.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
555 // compressor_att = compressor.get_attenuation();
559 add_bus_to_master(bus_index, samples_bus, &samples_out);
560 deinterleave_samples(samples_bus, &left, &right);
561 measure_bus_levels(bus_index, left, right);
565 lock_guard<mutex> lock(compressor_mutex);
567 // Finally a limiter at -4 dB (so, -10 dBFS) to take out the worst peaks only.
568 // Note that since ratio is not infinite, we could go slightly higher than this.
569 if (limiter_enabled) {
570 float threshold = from_db(limiter_threshold_dbfs);
572 float attack_time = 0.0f; // Instant.
573 float release_time = 0.020f;
574 float makeup_gain = 1.0f; // 0 dB.
575 limiter.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
576 // limiter_att = limiter.get_attenuation();
579 // printf("limiter=%+5.1f compressor=%+5.1f\n", to_db(limiter_att), to_db(compressor_att));
582 // At this point, we are most likely close to +0 LU (at least if the
583 // faders sum to 0 dB and the compressors are on), but all of our
584 // measurements have been on raw sample values, not R128 values.
585 // So we have a final makeup gain to get us to +0 LU; the gain
586 // adjustments required should be relatively small, and also, the
587 // offset shouldn't change much (only if the type of audio changes
588 // significantly). Thus, we shoot for updating this value basically
589 // “whenever we process buffers”, since the R128 calculation isn't exactly
590 // something we get out per-sample.
592 // Note that there's a feedback loop here, so we choose a very slow filter
593 // (half-time of 30 seconds).
594 double target_loudness_factor, alpha;
595 double loudness_lu = r128.loudness_M() - ref_level_lufs;
596 target_loudness_factor = final_makeup_gain * from_db(-loudness_lu);
598 // If we're outside +/- 5 LU (after correction), we don't count it as
599 // a normal signal (probably silence) and don't change the
600 // correction factor; just apply what we already have.
601 if (fabs(loudness_lu) >= 5.0 || !final_makeup_gain_auto) {
604 // Formula adapted from
605 // https://en.wikipedia.org/wiki/Low-pass_filter#Simple_infinite_impulse_response_filter.
606 const double half_time_s = 30.0;
607 const double fc_mul_2pi_delta_t = 1.0 / (half_time_s * OUTPUT_FREQUENCY);
608 alpha = fc_mul_2pi_delta_t / (fc_mul_2pi_delta_t + 1.0);
612 lock_guard<mutex> lock(compressor_mutex);
613 double m = final_makeup_gain;
614 for (size_t i = 0; i < samples_out.size(); i += 2) {
615 samples_out[i + 0] *= m;
616 samples_out[i + 1] *= m;
617 m += (target_loudness_factor - m) * alpha;
619 final_makeup_gain = m;
622 update_meters(samples_out);
629 void apply_filter_fade(StereoFilter *filter, float *data, unsigned num_samples, float cutoff_hz, float db, float last_db)
631 // A granularity of 32 samples is an okay tradeoff between speed and
632 // smoothness; recalculating the filters is pretty expensive, so it's
633 // good that we don't do this all the time.
634 static constexpr unsigned filter_granularity_samples = 32;
636 const float cutoff_linear = cutoff_hz * 2.0 * M_PI / OUTPUT_FREQUENCY;
637 if (fabs(db - last_db) < 1e-3) {
638 // Constant over this frame.
639 if (fabs(db) > 0.01f) {
640 filter->render(data, num_samples, cutoff_linear, 0.5f, db / 40.0f);
643 // We need to do a fade. (Rounding up avoids division by zero.)
644 unsigned num_blocks = (num_samples + filter_granularity_samples - 1) / filter_granularity_samples;
645 const float inc_db_norm = (db - last_db) / 40.0f / num_blocks;
646 float db_norm = db / 40.0f;
647 for (size_t i = 0; i < num_samples; i += filter_granularity_samples) {
648 size_t samples_this_block = std::min<size_t>(num_samples - i, filter_granularity_samples);
649 filter->render(data + i * 2, samples_this_block, cutoff_linear, 0.5f, db_norm);
650 db_norm += inc_db_norm;
657 void AudioMixer::apply_eq(unsigned bus_index, vector<float> *samples_bus)
659 constexpr float bass_freq_hz = 200.0f;
660 constexpr float treble_freq_hz = 4700.0f;
662 // Cut away everything under 120 Hz (or whatever the cutoff is);
663 // we don't need it for voice, and it will reduce headroom
664 // and confuse the compressor. (In particular, any hums at 50 or 60 Hz
665 // should be dampened.)
666 if (locut_enabled[bus_index]) {
667 locut[bus_index].render(samples_bus->data(), samples_bus->size() / 2, locut_cutoff_hz * 2.0 * M_PI / OUTPUT_FREQUENCY, 0.5f);
670 // Apply the rest of the EQ. Since we only have a simple three-band EQ,
671 // we can implement it with two shelf filters. We use a simple gain to
672 // set the mid-level filter, and then offset the low and high bands
673 // from that if we need to. (We could perhaps have folded the gain into
674 // the next part, but it's so cheap that the trouble isn't worth it.)
676 // If any part of the EQ has changed appreciably since last frame,
677 // we fade smoothly during the course of this frame.
678 const float bass_db = eq_level_db[bus_index][EQ_BAND_BASS];
679 const float mid_db = eq_level_db[bus_index][EQ_BAND_MID];
680 const float treble_db = eq_level_db[bus_index][EQ_BAND_TREBLE];
682 const float last_bass_db = last_eq_level_db[bus_index][EQ_BAND_BASS];
683 const float last_mid_db = last_eq_level_db[bus_index][EQ_BAND_MID];
684 const float last_treble_db = last_eq_level_db[bus_index][EQ_BAND_TREBLE];
686 assert(samples_bus->size() % 2 == 0);
687 const unsigned num_samples = samples_bus->size() / 2;
689 apply_gain(mid_db, last_mid_db, samples_bus);
691 apply_filter_fade(&eq[bus_index][EQ_BAND_BASS], samples_bus->data(), num_samples, bass_freq_hz, bass_db - mid_db, last_bass_db - last_mid_db);
692 apply_filter_fade(&eq[bus_index][EQ_BAND_TREBLE], samples_bus->data(), num_samples, treble_freq_hz, treble_db - mid_db, last_treble_db - last_mid_db);
694 last_eq_level_db[bus_index][EQ_BAND_BASS] = bass_db;
695 last_eq_level_db[bus_index][EQ_BAND_MID] = mid_db;
696 last_eq_level_db[bus_index][EQ_BAND_TREBLE] = treble_db;
699 void AudioMixer::add_bus_to_master(unsigned bus_index, const vector<float> &samples_bus, vector<float> *samples_out)
701 assert(samples_bus.size() == samples_out->size());
702 assert(samples_bus.size() % 2 == 0);
703 unsigned num_samples = samples_bus.size() / 2;
704 const float new_volume_db = mute[bus_index] ? -90.0f : fader_volume_db[bus_index].load();
705 if (fabs(new_volume_db - last_fader_volume_db[bus_index]) > 1e-3) {
706 // The volume has changed; do a fade over the course of this frame.
707 // (We might have some numerical issues here, but it seems to sound OK.)
708 // For the purpose of fading here, the silence floor is set to -90 dB
709 // (the fader only goes to -84).
710 float old_volume = from_db(max<float>(last_fader_volume_db[bus_index], -90.0f));
711 float volume = from_db(max<float>(new_volume_db, -90.0f));
713 float volume_inc = pow(volume / old_volume, 1.0 / num_samples);
715 if (bus_index == 0) {
716 for (unsigned i = 0; i < num_samples; ++i) {
717 (*samples_out)[i * 2 + 0] = samples_bus[i * 2 + 0] * volume;
718 (*samples_out)[i * 2 + 1] = samples_bus[i * 2 + 1] * volume;
719 volume *= volume_inc;
722 for (unsigned i = 0; i < num_samples; ++i) {
723 (*samples_out)[i * 2 + 0] += samples_bus[i * 2 + 0] * volume;
724 (*samples_out)[i * 2 + 1] += samples_bus[i * 2 + 1] * volume;
725 volume *= volume_inc;
728 } else if (new_volume_db > -90.0f) {
729 float volume = from_db(new_volume_db);
730 if (bus_index == 0) {
731 for (unsigned i = 0; i < num_samples; ++i) {
732 (*samples_out)[i * 2 + 0] = samples_bus[i * 2 + 0] * volume;
733 (*samples_out)[i * 2 + 1] = samples_bus[i * 2 + 1] * volume;
736 for (unsigned i = 0; i < num_samples; ++i) {
737 (*samples_out)[i * 2 + 0] += samples_bus[i * 2 + 0] * volume;
738 (*samples_out)[i * 2 + 1] += samples_bus[i * 2 + 1] * volume;
743 last_fader_volume_db[bus_index] = new_volume_db;
746 void AudioMixer::measure_bus_levels(unsigned bus_index, const vector<float> &left, const vector<float> &right)
748 assert(left.size() == right.size());
749 const float volume = mute[bus_index] ? 0.0f : from_db(fader_volume_db[bus_index]);
750 const float peak_levels[2] = {
751 find_peak(left.data(), left.size()) * volume,
752 find_peak(right.data(), right.size()) * volume
754 for (unsigned channel = 0; channel < 2; ++channel) {
755 // Compute the current value, including hold and falloff.
756 // The constants are borrowed from zita-mu1 by Fons Adriaensen.
757 static constexpr float hold_sec = 0.5f;
758 static constexpr float falloff_db_sec = 15.0f; // dB/sec falloff after hold.
760 PeakHistory &history = peak_history[bus_index][channel];
761 history.historic_peak = max(history.historic_peak, peak_levels[channel]);
762 if (history.age_seconds < hold_sec) {
763 current_peak = history.last_peak;
765 current_peak = history.last_peak * from_db(-falloff_db_sec * (history.age_seconds - hold_sec));
768 // See if we have a new peak to replace the old (possibly falling) one.
769 if (peak_levels[channel] > current_peak) {
770 history.last_peak = peak_levels[channel];
771 history.age_seconds = 0.0f; // Not 100% correct, but more than good enough given our frame sizes.
772 current_peak = peak_levels[channel];
774 history.age_seconds += float(left.size()) / OUTPUT_FREQUENCY;
776 history.current_level = peak_levels[channel];
777 history.current_peak = current_peak;
781 void AudioMixer::update_meters(const vector<float> &samples)
783 // Upsample 4x to find interpolated peak.
784 peak_resampler.inp_data = const_cast<float *>(samples.data());
785 peak_resampler.inp_count = samples.size() / 2;
787 vector<float> interpolated_samples;
788 interpolated_samples.resize(samples.size());
790 lock_guard<mutex> lock(audio_measure_mutex);
792 while (peak_resampler.inp_count > 0) { // About four iterations.
793 peak_resampler.out_data = &interpolated_samples[0];
794 peak_resampler.out_count = interpolated_samples.size() / 2;
795 peak_resampler.process();
796 size_t out_stereo_samples = interpolated_samples.size() / 2 - peak_resampler.out_count;
797 peak = max<float>(peak, find_peak(interpolated_samples.data(), out_stereo_samples * 2));
798 peak_resampler.out_data = nullptr;
802 // Find R128 levels and L/R correlation.
803 vector<float> left, right;
804 deinterleave_samples(samples, &left, &right);
805 float *ptrs[] = { left.data(), right.data() };
807 lock_guard<mutex> lock(audio_measure_mutex);
808 r128.process(left.size(), ptrs);
809 correlation.process_samples(samples);
812 send_audio_level_callback();
815 void AudioMixer::reset_meters()
817 lock_guard<mutex> lock(audio_measure_mutex);
818 peak_resampler.reset();
825 void AudioMixer::send_audio_level_callback()
827 if (audio_level_callback == nullptr) {
831 lock_guard<mutex> lock(audio_measure_mutex);
832 double loudness_s = r128.loudness_S();
833 double loudness_i = r128.integrated();
834 double loudness_range_low = r128.range_min();
835 double loudness_range_high = r128.range_max();
837 metric_audio_loudness_short_lufs = loudness_s;
838 metric_audio_loudness_integrated_lufs = loudness_i;
839 metric_audio_loudness_range_low_lufs = loudness_range_low;
840 metric_audio_loudness_range_high_lufs = loudness_range_high;
841 metric_audio_peak_dbfs = to_db(peak);
842 metric_audio_final_makeup_gain_db = to_db(final_makeup_gain);
843 metric_audio_correlation = correlation.get_correlation();
845 vector<BusLevel> bus_levels;
846 bus_levels.resize(input_mapping.buses.size());
848 lock_guard<mutex> lock(compressor_mutex);
849 for (unsigned bus_index = 0; bus_index < bus_levels.size(); ++bus_index) {
850 BusLevel &levels = bus_levels[bus_index];
851 BusMetrics &metrics = bus_metrics[bus_index];
853 levels.current_level_dbfs[0] = metrics.current_level_dbfs[0] = to_db(peak_history[bus_index][0].current_level);
854 levels.current_level_dbfs[1] = metrics.current_level_dbfs[1] = to_db(peak_history[bus_index][1].current_level);
855 levels.peak_level_dbfs[0] = metrics.peak_level_dbfs[0] = to_db(peak_history[bus_index][0].current_peak);
856 levels.peak_level_dbfs[1] = metrics.peak_level_dbfs[1] = to_db(peak_history[bus_index][1].current_peak);
857 levels.historic_peak_dbfs = metrics.historic_peak_dbfs = to_db(
858 max(peak_history[bus_index][0].historic_peak,
859 peak_history[bus_index][1].historic_peak));
860 levels.gain_staging_db = metrics.gain_staging_db = gain_staging_db[bus_index];
861 if (compressor_enabled[bus_index]) {
862 levels.compressor_attenuation_db = metrics.compressor_attenuation_db = -to_db(compressor[bus_index]->get_attenuation());
864 levels.compressor_attenuation_db = 0.0;
865 metrics.compressor_attenuation_db = 0.0 / 0.0;
870 audio_level_callback(loudness_s, to_db(peak), bus_levels,
871 loudness_i, loudness_range_low, loudness_range_high,
872 to_db(final_makeup_gain),
873 correlation.get_correlation());
876 map<DeviceSpec, DeviceInfo> AudioMixer::get_devices()
878 lock_guard<timed_mutex> lock(audio_mutex);
880 map<DeviceSpec, DeviceInfo> devices;
881 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
882 const DeviceSpec spec{ InputSourceType::CAPTURE_CARD, card_index };
883 const AudioDevice *device = &video_cards[card_index];
885 info.display_name = device->display_name;
886 info.num_channels = 8;
887 devices.insert(make_pair(spec, info));
889 vector<ALSAPool::Device> available_alsa_devices = alsa_pool.get_devices();
890 for (unsigned card_index = 0; card_index < available_alsa_devices.size(); ++card_index) {
891 const DeviceSpec spec{ InputSourceType::ALSA_INPUT, card_index };
892 const ALSAPool::Device &device = available_alsa_devices[card_index];
894 info.display_name = device.display_name();
895 info.num_channels = device.num_channels;
896 info.alsa_name = device.name;
897 info.alsa_info = device.info;
898 info.alsa_address = device.address;
899 devices.insert(make_pair(spec, info));
904 void AudioMixer::set_display_name(DeviceSpec device_spec, const string &name)
906 AudioDevice *device = find_audio_device(device_spec);
908 lock_guard<timed_mutex> lock(audio_mutex);
909 device->display_name = name;
912 void AudioMixer::serialize_device(DeviceSpec device_spec, DeviceSpecProto *device_spec_proto)
914 lock_guard<timed_mutex> lock(audio_mutex);
915 switch (device_spec.type) {
916 case InputSourceType::SILENCE:
917 device_spec_proto->set_type(DeviceSpecProto::SILENCE);
919 case InputSourceType::CAPTURE_CARD:
920 device_spec_proto->set_type(DeviceSpecProto::CAPTURE_CARD);
921 device_spec_proto->set_index(device_spec.index);
922 device_spec_proto->set_display_name(video_cards[device_spec.index].display_name);
924 case InputSourceType::ALSA_INPUT:
925 alsa_pool.serialize_device(device_spec.index, device_spec_proto);
930 void AudioMixer::set_simple_input(unsigned card_index)
932 InputMapping new_input_mapping;
933 InputMapping::Bus input;
935 input.device.type = InputSourceType::CAPTURE_CARD;
936 input.device.index = card_index;
937 input.source_channel[0] = 0;
938 input.source_channel[1] = 1;
940 new_input_mapping.buses.push_back(input);
942 lock_guard<timed_mutex> lock(audio_mutex);
943 current_mapping_mode = MappingMode::SIMPLE;
944 set_input_mapping_lock_held(new_input_mapping);
945 fader_volume_db[0] = 0.0f;
948 unsigned AudioMixer::get_simple_input() const
950 lock_guard<timed_mutex> lock(audio_mutex);
951 if (input_mapping.buses.size() == 1 &&
952 input_mapping.buses[0].device.type == InputSourceType::CAPTURE_CARD &&
953 input_mapping.buses[0].source_channel[0] == 0 &&
954 input_mapping.buses[0].source_channel[1] == 1) {
955 return input_mapping.buses[0].device.index;
957 return numeric_limits<unsigned>::max();
961 void AudioMixer::set_input_mapping(const InputMapping &new_input_mapping)
963 lock_guard<timed_mutex> lock(audio_mutex);
964 set_input_mapping_lock_held(new_input_mapping);
965 current_mapping_mode = MappingMode::MULTICHANNEL;
968 AudioMixer::MappingMode AudioMixer::get_mapping_mode() const
970 lock_guard<timed_mutex> lock(audio_mutex);
971 return current_mapping_mode;
974 void AudioMixer::set_input_mapping_lock_held(const InputMapping &new_input_mapping)
976 map<DeviceSpec, set<unsigned>> interesting_channels;
977 for (const InputMapping::Bus &bus : new_input_mapping.buses) {
978 if (bus.device.type == InputSourceType::CAPTURE_CARD ||
979 bus.device.type == InputSourceType::ALSA_INPUT) {
980 for (unsigned channel = 0; channel < 2; ++channel) {
981 if (bus.source_channel[channel] != -1) {
982 interesting_channels[bus.device].insert(bus.source_channel[channel]);
988 // Kill all the old metrics, and set up new ones.
989 for (unsigned bus_index = 0; bus_index < input_mapping.buses.size(); ++bus_index) {
990 BusMetrics &metrics = bus_metrics[bus_index];
992 vector<pair<string, string>> labels_left = metrics.labels;
993 labels_left.emplace_back("channel", "left");
994 vector<pair<string, string>> labels_right = metrics.labels;
995 labels_right.emplace_back("channel", "right");
997 global_metrics.remove("bus_current_level_dbfs", labels_left);
998 global_metrics.remove("bus_current_level_dbfs", labels_right);
999 global_metrics.remove("bus_peak_level_dbfs", labels_left);
1000 global_metrics.remove("bus_peak_level_dbfs", labels_right);
1001 global_metrics.remove("bus_historic_peak_dbfs", metrics.labels);
1002 global_metrics.remove("bus_gain_staging_db", metrics.labels);
1003 global_metrics.remove("bus_compressor_attenuation_db", metrics.labels);
1005 bus_metrics.reset(new BusMetrics[new_input_mapping.buses.size()]);
1006 for (unsigned bus_index = 0; bus_index < new_input_mapping.buses.size(); ++bus_index) {
1007 const InputMapping::Bus &bus = new_input_mapping.buses[bus_index];
1008 BusMetrics &metrics = bus_metrics[bus_index];
1010 char bus_index_str[16], source_index_str[16], source_channels_str[64];
1011 snprintf(bus_index_str, sizeof(bus_index_str), "%u", bus_index);
1012 snprintf(source_index_str, sizeof(source_index_str), "%u", bus.device.index);
1013 snprintf(source_channels_str, sizeof(source_channels_str), "%d:%d", bus.source_channel[0], bus.source_channel[1]);
1015 vector<pair<string, string>> labels;
1016 metrics.labels.emplace_back("index", bus_index_str);
1017 metrics.labels.emplace_back("name", bus.name);
1018 if (bus.device.type == InputSourceType::SILENCE) {
1019 metrics.labels.emplace_back("source_type", "silence");
1020 } else if (bus.device.type == InputSourceType::CAPTURE_CARD) {
1021 metrics.labels.emplace_back("source_type", "capture_card");
1022 } else if (bus.device.type == InputSourceType::ALSA_INPUT) {
1023 metrics.labels.emplace_back("source_type", "alsa_input");
1027 metrics.labels.emplace_back("source_index", source_index_str);
1028 metrics.labels.emplace_back("source_channels", source_channels_str);
1030 vector<pair<string, string>> labels_left = metrics.labels;
1031 labels_left.emplace_back("channel", "left");
1032 vector<pair<string, string>> labels_right = metrics.labels;
1033 labels_right.emplace_back("channel", "right");
1035 global_metrics.add("bus_current_level_dbfs", labels_left, &metrics.current_level_dbfs[0], Metrics::TYPE_GAUGE);
1036 global_metrics.add("bus_current_level_dbfs", labels_right, &metrics.current_level_dbfs[1], Metrics::TYPE_GAUGE);
1037 global_metrics.add("bus_peak_level_dbfs", labels_left, &metrics.peak_level_dbfs[0], Metrics::TYPE_GAUGE);
1038 global_metrics.add("bus_peak_level_dbfs", labels_right, &metrics.peak_level_dbfs[1], Metrics::TYPE_GAUGE);
1039 global_metrics.add("bus_historic_peak_dbfs", metrics.labels, &metrics.historic_peak_dbfs, Metrics::TYPE_GAUGE);
1040 global_metrics.add("bus_gain_staging_db", metrics.labels, &metrics.gain_staging_db, Metrics::TYPE_GAUGE);
1041 global_metrics.add("bus_compressor_attenuation_db", metrics.labels, &metrics.compressor_attenuation_db, Metrics::TYPE_GAUGE);
1044 // Reset resamplers for all cards that don't have the exact same state as before.
1045 for (unsigned card_index = 0; card_index < MAX_VIDEO_CARDS; ++card_index) {
1046 const DeviceSpec device_spec{InputSourceType::CAPTURE_CARD, card_index};
1047 AudioDevice *device = find_audio_device(device_spec);
1048 if (device->interesting_channels != interesting_channels[device_spec]) {
1049 device->interesting_channels = interesting_channels[device_spec];
1050 reset_resampler_mutex_held(device_spec);
1053 for (unsigned card_index = 0; card_index < MAX_ALSA_CARDS; ++card_index) {
1054 const DeviceSpec device_spec{InputSourceType::ALSA_INPUT, card_index};
1055 AudioDevice *device = find_audio_device(device_spec);
1056 if (interesting_channels[device_spec].empty()) {
1057 alsa_pool.release_device(card_index);
1059 alsa_pool.hold_device(card_index);
1061 if (device->interesting_channels != interesting_channels[device_spec]) {
1062 device->interesting_channels = interesting_channels[device_spec];
1063 alsa_pool.reset_device(device_spec.index);
1064 reset_resampler_mutex_held(device_spec);
1068 input_mapping = new_input_mapping;
1071 InputMapping AudioMixer::get_input_mapping() const
1073 lock_guard<timed_mutex> lock(audio_mutex);
1074 return input_mapping;
1077 unsigned AudioMixer::num_buses() const
1079 lock_guard<timed_mutex> lock(audio_mutex);
1080 return input_mapping.buses.size();
1083 void AudioMixer::reset_peak(unsigned bus_index)
1085 lock_guard<timed_mutex> lock(audio_mutex);
1086 for (unsigned channel = 0; channel < 2; ++channel) {
1087 PeakHistory &history = peak_history[bus_index][channel];
1088 history.current_level = 0.0f;
1089 history.historic_peak = 0.0f;
1090 history.current_peak = 0.0f;
1091 history.last_peak = 0.0f;
1092 history.age_seconds = 0.0f;
1096 AudioMixer *global_audio_mixer = nullptr;