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 // TODO: ResamplingQueue should probably take the full device spec.
240 // (It's only used for console output, though.)
241 device->resampling_queue.reset(new ResamplingQueue(
242 device_spec.index, device->capture_frequency, OUTPUT_FREQUENCY, device->interesting_channels.size(),
243 global_flags.audio_queue_length_ms * 0.001));
247 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)
249 AudioDevice *device = find_audio_device(device_spec);
251 unique_lock<timed_mutex> lock(audio_mutex, defer_lock);
252 if (!lock.try_lock_for(chrono::milliseconds(10))) {
255 if (device->resampling_queue == nullptr) {
256 // No buses use this device; throw it away.
260 unsigned num_channels = device->interesting_channels.size();
261 assert(num_channels > 0);
263 // Convert the audio to fp32.
264 unique_ptr<float[]> audio(new float[num_samples * num_channels]);
265 unsigned channel_index = 0;
266 for (auto channel_it = device->interesting_channels.cbegin(); channel_it != device->interesting_channels.end(); ++channel_it, ++channel_index) {
267 switch (audio_format.bits_per_sample) {
269 assert(num_samples == 0);
272 convert_fixed16_to_fp32(audio.get(), channel_index, num_channels, data, *channel_it, audio_format.num_channels, num_samples);
275 convert_fixed24_to_fp32(audio.get(), channel_index, num_channels, data, *channel_it, audio_format.num_channels, num_samples);
278 convert_fixed32_to_fp32(audio.get(), channel_index, num_channels, data, *channel_it, audio_format.num_channels, num_samples);
281 fprintf(stderr, "Cannot handle audio with %u bits per sample\n", audio_format.bits_per_sample);
286 // If we changed frequency since last frame, we'll need to reset the resampler.
287 if (audio_format.sample_rate != device->capture_frequency) {
288 device->capture_frequency = audio_format.sample_rate;
289 reset_resampler_mutex_held(device_spec);
293 device->resampling_queue->add_input_samples(frame_time, audio.get(), num_samples, ResamplingQueue::ADJUST_RATE);
297 bool AudioMixer::add_silence(DeviceSpec device_spec, unsigned samples_per_frame, unsigned num_frames, int64_t frame_length)
299 AudioDevice *device = find_audio_device(device_spec);
301 unique_lock<timed_mutex> lock(audio_mutex, defer_lock);
302 if (!lock.try_lock_for(chrono::milliseconds(10))) {
305 if (device->resampling_queue == nullptr) {
306 // No buses use this device; throw it away.
310 unsigned num_channels = device->interesting_channels.size();
311 assert(num_channels > 0);
313 vector<float> silence(samples_per_frame * num_channels, 0.0f);
314 for (unsigned i = 0; i < num_frames; ++i) {
315 device->resampling_queue->add_input_samples(steady_clock::now(), silence.data(), samples_per_frame, ResamplingQueue::DO_NOT_ADJUST_RATE);
320 bool AudioMixer::silence_card(DeviceSpec device_spec, bool silence)
322 AudioDevice *device = find_audio_device(device_spec);
324 unique_lock<timed_mutex> lock(audio_mutex, defer_lock);
325 if (!lock.try_lock_for(chrono::milliseconds(10))) {
329 if (device->silenced && !silence) {
330 reset_resampler_mutex_held(device_spec);
332 device->silenced = silence;
336 AudioMixer::BusSettings AudioMixer::get_default_bus_settings()
338 BusSettings settings;
339 settings.fader_volume_db = 0.0f;
340 settings.muted = false;
341 settings.locut_enabled = global_flags.locut_enabled;
342 for (unsigned band_index = 0; band_index < NUM_EQ_BANDS; ++band_index) {
343 settings.eq_level_db[band_index] = 0.0f;
345 settings.gain_staging_db = global_flags.initial_gain_staging_db;
346 settings.level_compressor_enabled = global_flags.gain_staging_auto;
347 settings.compressor_threshold_dbfs = ref_level_dbfs - 12.0f; // -12 dB.
348 settings.compressor_enabled = global_flags.compressor_enabled;
352 AudioMixer::BusSettings AudioMixer::get_bus_settings(unsigned bus_index) const
354 lock_guard<timed_mutex> lock(audio_mutex);
355 BusSettings settings;
356 settings.fader_volume_db = fader_volume_db[bus_index];
357 settings.muted = mute[bus_index];
358 settings.locut_enabled = locut_enabled[bus_index];
359 for (unsigned band_index = 0; band_index < NUM_EQ_BANDS; ++band_index) {
360 settings.eq_level_db[band_index] = eq_level_db[bus_index][band_index];
362 settings.gain_staging_db = gain_staging_db[bus_index];
363 settings.level_compressor_enabled = level_compressor_enabled[bus_index];
364 settings.compressor_threshold_dbfs = compressor_threshold_dbfs[bus_index];
365 settings.compressor_enabled = compressor_enabled[bus_index];
369 void AudioMixer::set_bus_settings(unsigned bus_index, const AudioMixer::BusSettings &settings)
371 lock_guard<timed_mutex> lock(audio_mutex);
372 fader_volume_db[bus_index] = settings.fader_volume_db;
373 mute[bus_index] = settings.muted;
374 locut_enabled[bus_index] = settings.locut_enabled;
375 for (unsigned band_index = 0; band_index < NUM_EQ_BANDS; ++band_index) {
376 eq_level_db[bus_index][band_index] = settings.eq_level_db[band_index];
378 gain_staging_db[bus_index] = settings.gain_staging_db;
379 last_gain_staging_db[bus_index] = gain_staging_db[bus_index];
380 level_compressor_enabled[bus_index] = settings.level_compressor_enabled;
381 compressor_threshold_dbfs[bus_index] = settings.compressor_threshold_dbfs;
382 compressor_enabled[bus_index] = settings.compressor_enabled;
385 AudioMixer::AudioDevice *AudioMixer::find_audio_device(DeviceSpec device)
387 switch (device.type) {
388 case InputSourceType::CAPTURE_CARD:
389 return &video_cards[device.index];
390 case InputSourceType::ALSA_INPUT:
391 return &alsa_inputs[device.index];
392 case InputSourceType::SILENCE:
399 // Get a pointer to the given channel from the given device.
400 // The channel must be picked out earlier and resampled.
401 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)
403 static float zero = 0.0f;
404 if (source_channel == -1 || device_spec.type == InputSourceType::SILENCE) {
409 AudioDevice *device = find_audio_device(device_spec);
410 assert(device->interesting_channels.count(source_channel) != 0);
411 unsigned channel_index = 0;
412 for (int channel : device->interesting_channels) {
413 if (channel == source_channel) break;
416 assert(channel_index < device->interesting_channels.size());
417 const auto it = samples_card.find(device_spec);
418 assert(it != samples_card.end());
419 *srcptr = &(it->second)[channel_index];
420 *stride = device->interesting_channels.size();
423 // TODO: Can be SSSE3-optimized if need be.
424 void AudioMixer::fill_audio_bus(const map<DeviceSpec, vector<float>> &samples_card, const InputMapping::Bus &bus, unsigned num_samples, float *output)
426 if (bus.device.type == InputSourceType::SILENCE) {
427 memset(output, 0, num_samples * 2 * sizeof(*output));
429 assert(bus.device.type == InputSourceType::CAPTURE_CARD ||
430 bus.device.type == InputSourceType::ALSA_INPUT);
431 const float *lsrc, *rsrc;
432 unsigned lstride, rstride;
433 float *dptr = output;
434 find_sample_src_from_device(samples_card, bus.device, bus.source_channel[0], &lsrc, &lstride);
435 find_sample_src_from_device(samples_card, bus.device, bus.source_channel[1], &rsrc, &rstride);
436 for (unsigned i = 0; i < num_samples; ++i) {
445 vector<DeviceSpec> AudioMixer::get_active_devices() const
447 vector<DeviceSpec> ret;
448 for (unsigned card_index = 0; card_index < MAX_VIDEO_CARDS; ++card_index) {
449 const DeviceSpec device_spec{InputSourceType::CAPTURE_CARD, card_index};
450 if (!find_audio_device(device_spec)->interesting_channels.empty()) {
451 ret.push_back(device_spec);
454 for (unsigned card_index = 0; card_index < MAX_ALSA_CARDS; ++card_index) {
455 const DeviceSpec device_spec{InputSourceType::ALSA_INPUT, card_index};
456 if (!find_audio_device(device_spec)->interesting_channels.empty()) {
457 ret.push_back(device_spec);
465 void apply_gain(float db, float last_db, vector<float> *samples)
467 if (fabs(db - last_db) < 1e-3) {
468 // Constant over this frame.
469 const float gain = from_db(db);
470 for (size_t i = 0; i < samples->size(); ++i) {
471 (*samples)[i] *= gain;
474 // We need to do a fade.
475 unsigned num_samples = samples->size() / 2;
476 float gain = from_db(last_db);
477 const float gain_inc = pow(from_db(db - last_db), 1.0 / num_samples);
478 for (size_t i = 0; i < num_samples; ++i) {
479 (*samples)[i * 2 + 0] *= gain;
480 (*samples)[i * 2 + 1] *= gain;
488 vector<float> AudioMixer::get_output(steady_clock::time_point ts, unsigned num_samples, ResamplingQueue::RateAdjustmentPolicy rate_adjustment_policy)
490 map<DeviceSpec, vector<float>> samples_card;
491 vector<float> samples_bus;
493 lock_guard<timed_mutex> lock(audio_mutex);
495 // Pick out all the interesting channels from all the cards.
496 for (const DeviceSpec &device_spec : get_active_devices()) {
497 AudioDevice *device = find_audio_device(device_spec);
498 samples_card[device_spec].resize(num_samples * device->interesting_channels.size());
499 if (device->silenced) {
500 memset(&samples_card[device_spec][0], 0, samples_card[device_spec].size() * sizeof(float));
502 device->resampling_queue->get_output_samples(
504 &samples_card[device_spec][0],
506 rate_adjustment_policy);
510 vector<float> samples_out, left, right;
511 samples_out.resize(num_samples * 2);
512 samples_bus.resize(num_samples * 2);
513 for (unsigned bus_index = 0; bus_index < input_mapping.buses.size(); ++bus_index) {
514 fill_audio_bus(samples_card, input_mapping.buses[bus_index], num_samples, &samples_bus[0]);
515 apply_eq(bus_index, &samples_bus);
518 lock_guard<mutex> lock(compressor_mutex);
520 // Apply a level compressor to get the general level right.
521 // Basically, if it's over about -40 dBFS, we squeeze it down to that level
522 // (or more precisely, near it, since we don't use infinite ratio),
523 // then apply a makeup gain to get it to -14 dBFS. -14 dBFS is, of course,
524 // entirely arbitrary, but from practical tests with speech, it seems to
525 // put ut around -23 LUFS, so it's a reasonable starting point for later use.
526 if (level_compressor_enabled[bus_index]) {
527 float threshold = 0.01f; // -40 dBFS.
529 float attack_time = 0.5f;
530 float release_time = 20.0f;
531 float makeup_gain = from_db(ref_level_dbfs - (-40.0f)); // +26 dB.
532 level_compressor[bus_index]->process(samples_bus.data(), samples_bus.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
533 gain_staging_db[bus_index] = to_db(level_compressor[bus_index]->get_attenuation() * makeup_gain);
535 // Just apply the gain we already had.
536 float db = gain_staging_db[bus_index];
537 float last_db = last_gain_staging_db[bus_index];
538 apply_gain(db, last_db, &samples_bus);
540 last_gain_staging_db[bus_index] = gain_staging_db[bus_index];
543 printf("level=%f (%+5.2f dBFS) attenuation=%f (%+5.2f dB) end_result=%+5.2f dB\n",
544 level_compressor.get_level(), to_db(level_compressor.get_level()),
545 level_compressor.get_attenuation(), to_db(level_compressor.get_attenuation()),
546 to_db(level_compressor.get_level() * level_compressor.get_attenuation() * makeup_gain));
549 // The real compressor.
550 if (compressor_enabled[bus_index]) {
551 float threshold = from_db(compressor_threshold_dbfs[bus_index]);
553 float attack_time = 0.005f;
554 float release_time = 0.040f;
555 float makeup_gain = 2.0f; // +6 dB.
556 compressor[bus_index]->process(samples_bus.data(), samples_bus.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
557 // compressor_att = compressor.get_attenuation();
561 add_bus_to_master(bus_index, samples_bus, &samples_out);
562 deinterleave_samples(samples_bus, &left, &right);
563 measure_bus_levels(bus_index, left, right);
567 lock_guard<mutex> lock(compressor_mutex);
569 // Finally a limiter at -4 dB (so, -10 dBFS) to take out the worst peaks only.
570 // Note that since ratio is not infinite, we could go slightly higher than this.
571 if (limiter_enabled) {
572 float threshold = from_db(limiter_threshold_dbfs);
574 float attack_time = 0.0f; // Instant.
575 float release_time = 0.020f;
576 float makeup_gain = 1.0f; // 0 dB.
577 limiter.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
578 // limiter_att = limiter.get_attenuation();
581 // printf("limiter=%+5.1f compressor=%+5.1f\n", to_db(limiter_att), to_db(compressor_att));
584 // At this point, we are most likely close to +0 LU (at least if the
585 // faders sum to 0 dB and the compressors are on), but all of our
586 // measurements have been on raw sample values, not R128 values.
587 // So we have a final makeup gain to get us to +0 LU; the gain
588 // adjustments required should be relatively small, and also, the
589 // offset shouldn't change much (only if the type of audio changes
590 // significantly). Thus, we shoot for updating this value basically
591 // “whenever we process buffers”, since the R128 calculation isn't exactly
592 // something we get out per-sample.
594 // Note that there's a feedback loop here, so we choose a very slow filter
595 // (half-time of 30 seconds).
596 double target_loudness_factor, alpha;
597 double loudness_lu = r128.loudness_M() - ref_level_lufs;
598 target_loudness_factor = final_makeup_gain * from_db(-loudness_lu);
600 // If we're outside +/- 5 LU (after correction), we don't count it as
601 // a normal signal (probably silence) and don't change the
602 // correction factor; just apply what we already have.
603 if (fabs(loudness_lu) >= 5.0 || !final_makeup_gain_auto) {
606 // Formula adapted from
607 // https://en.wikipedia.org/wiki/Low-pass_filter#Simple_infinite_impulse_response_filter.
608 const double half_time_s = 30.0;
609 const double fc_mul_2pi_delta_t = 1.0 / (half_time_s * OUTPUT_FREQUENCY);
610 alpha = fc_mul_2pi_delta_t / (fc_mul_2pi_delta_t + 1.0);
614 lock_guard<mutex> lock(compressor_mutex);
615 double m = final_makeup_gain;
616 for (size_t i = 0; i < samples_out.size(); i += 2) {
617 samples_out[i + 0] *= m;
618 samples_out[i + 1] *= m;
619 m += (target_loudness_factor - m) * alpha;
621 final_makeup_gain = m;
624 update_meters(samples_out);
631 void apply_filter_fade(StereoFilter *filter, float *data, unsigned num_samples, float cutoff_hz, float db, float last_db)
633 // A granularity of 32 samples is an okay tradeoff between speed and
634 // smoothness; recalculating the filters is pretty expensive, so it's
635 // good that we don't do this all the time.
636 static constexpr unsigned filter_granularity_samples = 32;
638 const float cutoff_linear = cutoff_hz * 2.0 * M_PI / OUTPUT_FREQUENCY;
639 if (fabs(db - last_db) < 1e-3) {
640 // Constant over this frame.
641 if (fabs(db) > 0.01f) {
642 filter->render(data, num_samples, cutoff_linear, 0.5f, db / 40.0f);
645 // We need to do a fade. (Rounding up avoids division by zero.)
646 unsigned num_blocks = (num_samples + filter_granularity_samples - 1) / filter_granularity_samples;
647 const float inc_db_norm = (db - last_db) / 40.0f / num_blocks;
648 float db_norm = db / 40.0f;
649 for (size_t i = 0; i < num_samples; i += filter_granularity_samples) {
650 size_t samples_this_block = std::min<size_t>(num_samples - i, filter_granularity_samples);
651 filter->render(data + i * 2, samples_this_block, cutoff_linear, 0.5f, db_norm);
652 db_norm += inc_db_norm;
659 void AudioMixer::apply_eq(unsigned bus_index, vector<float> *samples_bus)
661 constexpr float bass_freq_hz = 200.0f;
662 constexpr float treble_freq_hz = 4700.0f;
664 // Cut away everything under 120 Hz (or whatever the cutoff is);
665 // we don't need it for voice, and it will reduce headroom
666 // and confuse the compressor. (In particular, any hums at 50 or 60 Hz
667 // should be dampened.)
668 if (locut_enabled[bus_index]) {
669 locut[bus_index].render(samples_bus->data(), samples_bus->size() / 2, locut_cutoff_hz * 2.0 * M_PI / OUTPUT_FREQUENCY, 0.5f);
672 // Apply the rest of the EQ. Since we only have a simple three-band EQ,
673 // we can implement it with two shelf filters. We use a simple gain to
674 // set the mid-level filter, and then offset the low and high bands
675 // from that if we need to. (We could perhaps have folded the gain into
676 // the next part, but it's so cheap that the trouble isn't worth it.)
678 // If any part of the EQ has changed appreciably since last frame,
679 // we fade smoothly during the course of this frame.
680 const float bass_db = eq_level_db[bus_index][EQ_BAND_BASS];
681 const float mid_db = eq_level_db[bus_index][EQ_BAND_MID];
682 const float treble_db = eq_level_db[bus_index][EQ_BAND_TREBLE];
684 const float last_bass_db = last_eq_level_db[bus_index][EQ_BAND_BASS];
685 const float last_mid_db = last_eq_level_db[bus_index][EQ_BAND_MID];
686 const float last_treble_db = last_eq_level_db[bus_index][EQ_BAND_TREBLE];
688 assert(samples_bus->size() % 2 == 0);
689 const unsigned num_samples = samples_bus->size() / 2;
691 apply_gain(mid_db, last_mid_db, samples_bus);
693 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);
694 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);
696 last_eq_level_db[bus_index][EQ_BAND_BASS] = bass_db;
697 last_eq_level_db[bus_index][EQ_BAND_MID] = mid_db;
698 last_eq_level_db[bus_index][EQ_BAND_TREBLE] = treble_db;
701 void AudioMixer::add_bus_to_master(unsigned bus_index, const vector<float> &samples_bus, vector<float> *samples_out)
703 assert(samples_bus.size() == samples_out->size());
704 assert(samples_bus.size() % 2 == 0);
705 unsigned num_samples = samples_bus.size() / 2;
706 const float new_volume_db = mute[bus_index] ? -90.0f : fader_volume_db[bus_index].load();
707 if (fabs(new_volume_db - last_fader_volume_db[bus_index]) > 1e-3) {
708 // The volume has changed; do a fade over the course of this frame.
709 // (We might have some numerical issues here, but it seems to sound OK.)
710 // For the purpose of fading here, the silence floor is set to -90 dB
711 // (the fader only goes to -84).
712 float old_volume = from_db(max<float>(last_fader_volume_db[bus_index], -90.0f));
713 float volume = from_db(max<float>(new_volume_db, -90.0f));
715 float volume_inc = pow(volume / old_volume, 1.0 / num_samples);
717 if (bus_index == 0) {
718 for (unsigned i = 0; i < num_samples; ++i) {
719 (*samples_out)[i * 2 + 0] = samples_bus[i * 2 + 0] * volume;
720 (*samples_out)[i * 2 + 1] = samples_bus[i * 2 + 1] * volume;
721 volume *= volume_inc;
724 for (unsigned i = 0; i < num_samples; ++i) {
725 (*samples_out)[i * 2 + 0] += samples_bus[i * 2 + 0] * volume;
726 (*samples_out)[i * 2 + 1] += samples_bus[i * 2 + 1] * volume;
727 volume *= volume_inc;
730 } else if (new_volume_db > -90.0f) {
731 float volume = from_db(new_volume_db);
732 if (bus_index == 0) {
733 for (unsigned i = 0; i < num_samples; ++i) {
734 (*samples_out)[i * 2 + 0] = samples_bus[i * 2 + 0] * volume;
735 (*samples_out)[i * 2 + 1] = samples_bus[i * 2 + 1] * volume;
738 for (unsigned i = 0; i < num_samples; ++i) {
739 (*samples_out)[i * 2 + 0] += samples_bus[i * 2 + 0] * volume;
740 (*samples_out)[i * 2 + 1] += samples_bus[i * 2 + 1] * volume;
745 last_fader_volume_db[bus_index] = new_volume_db;
748 void AudioMixer::measure_bus_levels(unsigned bus_index, const vector<float> &left, const vector<float> &right)
750 assert(left.size() == right.size());
751 const float volume = mute[bus_index] ? 0.0f : from_db(fader_volume_db[bus_index]);
752 const float peak_levels[2] = {
753 find_peak(left.data(), left.size()) * volume,
754 find_peak(right.data(), right.size()) * volume
756 for (unsigned channel = 0; channel < 2; ++channel) {
757 // Compute the current value, including hold and falloff.
758 // The constants are borrowed from zita-mu1 by Fons Adriaensen.
759 static constexpr float hold_sec = 0.5f;
760 static constexpr float falloff_db_sec = 15.0f; // dB/sec falloff after hold.
762 PeakHistory &history = peak_history[bus_index][channel];
763 history.historic_peak = max(history.historic_peak, peak_levels[channel]);
764 if (history.age_seconds < hold_sec) {
765 current_peak = history.last_peak;
767 current_peak = history.last_peak * from_db(-falloff_db_sec * (history.age_seconds - hold_sec));
770 // See if we have a new peak to replace the old (possibly falling) one.
771 if (peak_levels[channel] > current_peak) {
772 history.last_peak = peak_levels[channel];
773 history.age_seconds = 0.0f; // Not 100% correct, but more than good enough given our frame sizes.
774 current_peak = peak_levels[channel];
776 history.age_seconds += float(left.size()) / OUTPUT_FREQUENCY;
778 history.current_level = peak_levels[channel];
779 history.current_peak = current_peak;
783 void AudioMixer::update_meters(const vector<float> &samples)
785 // Upsample 4x to find interpolated peak.
786 peak_resampler.inp_data = const_cast<float *>(samples.data());
787 peak_resampler.inp_count = samples.size() / 2;
789 vector<float> interpolated_samples;
790 interpolated_samples.resize(samples.size());
792 lock_guard<mutex> lock(audio_measure_mutex);
794 while (peak_resampler.inp_count > 0) { // About four iterations.
795 peak_resampler.out_data = &interpolated_samples[0];
796 peak_resampler.out_count = interpolated_samples.size() / 2;
797 peak_resampler.process();
798 size_t out_stereo_samples = interpolated_samples.size() / 2 - peak_resampler.out_count;
799 peak = max<float>(peak, find_peak(interpolated_samples.data(), out_stereo_samples * 2));
800 peak_resampler.out_data = nullptr;
804 // Find R128 levels and L/R correlation.
805 vector<float> left, right;
806 deinterleave_samples(samples, &left, &right);
807 float *ptrs[] = { left.data(), right.data() };
809 lock_guard<mutex> lock(audio_measure_mutex);
810 r128.process(left.size(), ptrs);
811 correlation.process_samples(samples);
814 send_audio_level_callback();
817 void AudioMixer::reset_meters()
819 lock_guard<mutex> lock(audio_measure_mutex);
820 peak_resampler.reset();
827 void AudioMixer::send_audio_level_callback()
829 if (audio_level_callback == nullptr) {
833 lock_guard<mutex> lock(audio_measure_mutex);
834 double loudness_s = r128.loudness_S();
835 double loudness_i = r128.integrated();
836 double loudness_range_low = r128.range_min();
837 double loudness_range_high = r128.range_max();
839 metric_audio_loudness_short_lufs = loudness_s;
840 metric_audio_loudness_integrated_lufs = loudness_i;
841 metric_audio_loudness_range_low_lufs = loudness_range_low;
842 metric_audio_loudness_range_high_lufs = loudness_range_high;
843 metric_audio_peak_dbfs = to_db(peak);
844 metric_audio_final_makeup_gain_db = to_db(final_makeup_gain);
845 metric_audio_correlation = correlation.get_correlation();
847 vector<BusLevel> bus_levels;
848 bus_levels.resize(input_mapping.buses.size());
850 lock_guard<mutex> lock(compressor_mutex);
851 for (unsigned bus_index = 0; bus_index < bus_levels.size(); ++bus_index) {
852 BusLevel &levels = bus_levels[bus_index];
853 BusMetrics &metrics = bus_metrics[bus_index];
855 levels.current_level_dbfs[0] = metrics.current_level_dbfs[0] = to_db(peak_history[bus_index][0].current_level);
856 levels.current_level_dbfs[1] = metrics.current_level_dbfs[1] = to_db(peak_history[bus_index][1].current_level);
857 levels.peak_level_dbfs[0] = metrics.peak_level_dbfs[0] = to_db(peak_history[bus_index][0].current_peak);
858 levels.peak_level_dbfs[1] = metrics.peak_level_dbfs[1] = to_db(peak_history[bus_index][1].current_peak);
859 levels.historic_peak_dbfs = metrics.historic_peak_dbfs = to_db(
860 max(peak_history[bus_index][0].historic_peak,
861 peak_history[bus_index][1].historic_peak));
862 levels.gain_staging_db = metrics.gain_staging_db = gain_staging_db[bus_index];
863 if (compressor_enabled[bus_index]) {
864 levels.compressor_attenuation_db = metrics.compressor_attenuation_db = -to_db(compressor[bus_index]->get_attenuation());
866 levels.compressor_attenuation_db = 0.0;
867 metrics.compressor_attenuation_db = 0.0 / 0.0;
872 audio_level_callback(loudness_s, to_db(peak), bus_levels,
873 loudness_i, loudness_range_low, loudness_range_high,
874 to_db(final_makeup_gain),
875 correlation.get_correlation());
878 map<DeviceSpec, DeviceInfo> AudioMixer::get_devices()
880 lock_guard<timed_mutex> lock(audio_mutex);
882 map<DeviceSpec, DeviceInfo> devices;
883 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
884 const DeviceSpec spec{ InputSourceType::CAPTURE_CARD, card_index };
885 const AudioDevice *device = &video_cards[card_index];
887 info.display_name = device->display_name;
888 info.num_channels = 8;
889 devices.insert(make_pair(spec, info));
891 vector<ALSAPool::Device> available_alsa_devices = alsa_pool.get_devices();
892 for (unsigned card_index = 0; card_index < available_alsa_devices.size(); ++card_index) {
893 const DeviceSpec spec{ InputSourceType::ALSA_INPUT, card_index };
894 const ALSAPool::Device &device = available_alsa_devices[card_index];
896 info.display_name = device.display_name();
897 info.num_channels = device.num_channels;
898 info.alsa_name = device.name;
899 info.alsa_info = device.info;
900 info.alsa_address = device.address;
901 devices.insert(make_pair(spec, info));
906 void AudioMixer::set_display_name(DeviceSpec device_spec, const string &name)
908 AudioDevice *device = find_audio_device(device_spec);
910 lock_guard<timed_mutex> lock(audio_mutex);
911 device->display_name = name;
914 void AudioMixer::serialize_device(DeviceSpec device_spec, DeviceSpecProto *device_spec_proto)
916 lock_guard<timed_mutex> lock(audio_mutex);
917 switch (device_spec.type) {
918 case InputSourceType::SILENCE:
919 device_spec_proto->set_type(DeviceSpecProto::SILENCE);
921 case InputSourceType::CAPTURE_CARD:
922 device_spec_proto->set_type(DeviceSpecProto::CAPTURE_CARD);
923 device_spec_proto->set_index(device_spec.index);
924 device_spec_proto->set_display_name(video_cards[device_spec.index].display_name);
926 case InputSourceType::ALSA_INPUT:
927 alsa_pool.serialize_device(device_spec.index, device_spec_proto);
932 void AudioMixer::set_simple_input(unsigned card_index)
934 InputMapping new_input_mapping;
935 InputMapping::Bus input;
937 input.device.type = InputSourceType::CAPTURE_CARD;
938 input.device.index = card_index;
939 input.source_channel[0] = 0;
940 input.source_channel[1] = 1;
942 new_input_mapping.buses.push_back(input);
944 lock_guard<timed_mutex> lock(audio_mutex);
945 current_mapping_mode = MappingMode::SIMPLE;
946 set_input_mapping_lock_held(new_input_mapping);
947 fader_volume_db[0] = 0.0f;
950 unsigned AudioMixer::get_simple_input() const
952 lock_guard<timed_mutex> lock(audio_mutex);
953 if (input_mapping.buses.size() == 1 &&
954 input_mapping.buses[0].device.type == InputSourceType::CAPTURE_CARD &&
955 input_mapping.buses[0].source_channel[0] == 0 &&
956 input_mapping.buses[0].source_channel[1] == 1) {
957 return input_mapping.buses[0].device.index;
959 return numeric_limits<unsigned>::max();
963 void AudioMixer::set_input_mapping(const InputMapping &new_input_mapping)
965 lock_guard<timed_mutex> lock(audio_mutex);
966 set_input_mapping_lock_held(new_input_mapping);
967 current_mapping_mode = MappingMode::MULTICHANNEL;
970 AudioMixer::MappingMode AudioMixer::get_mapping_mode() const
972 lock_guard<timed_mutex> lock(audio_mutex);
973 return current_mapping_mode;
976 void AudioMixer::set_input_mapping_lock_held(const InputMapping &new_input_mapping)
978 map<DeviceSpec, set<unsigned>> interesting_channels;
979 for (const InputMapping::Bus &bus : new_input_mapping.buses) {
980 if (bus.device.type == InputSourceType::CAPTURE_CARD ||
981 bus.device.type == InputSourceType::ALSA_INPUT) {
982 for (unsigned channel = 0; channel < 2; ++channel) {
983 if (bus.source_channel[channel] != -1) {
984 interesting_channels[bus.device].insert(bus.source_channel[channel]);
990 // Kill all the old metrics, and set up new ones.
991 for (unsigned bus_index = 0; bus_index < input_mapping.buses.size(); ++bus_index) {
992 BusMetrics &metrics = bus_metrics[bus_index];
994 vector<pair<string, string>> labels_left = metrics.labels;
995 labels_left.emplace_back("channel", "left");
996 vector<pair<string, string>> labels_right = metrics.labels;
997 labels_right.emplace_back("channel", "right");
999 global_metrics.remove("bus_current_level_dbfs", labels_left);
1000 global_metrics.remove("bus_current_level_dbfs", labels_right);
1001 global_metrics.remove("bus_peak_level_dbfs", labels_left);
1002 global_metrics.remove("bus_peak_level_dbfs", labels_right);
1003 global_metrics.remove("bus_historic_peak_dbfs", metrics.labels);
1004 global_metrics.remove("bus_gain_staging_db", metrics.labels);
1005 global_metrics.remove("bus_compressor_attenuation_db", metrics.labels);
1007 bus_metrics.reset(new BusMetrics[new_input_mapping.buses.size()]);
1008 for (unsigned bus_index = 0; bus_index < new_input_mapping.buses.size(); ++bus_index) {
1009 const InputMapping::Bus &bus = new_input_mapping.buses[bus_index];
1010 BusMetrics &metrics = bus_metrics[bus_index];
1012 char bus_index_str[16], source_index_str[16], source_channels_str[64];
1013 snprintf(bus_index_str, sizeof(bus_index_str), "%u", bus_index);
1014 snprintf(source_index_str, sizeof(source_index_str), "%u", bus.device.index);
1015 snprintf(source_channels_str, sizeof(source_channels_str), "%d:%d", bus.source_channel[0], bus.source_channel[1]);
1017 vector<pair<string, string>> labels;
1018 metrics.labels.emplace_back("index", bus_index_str);
1019 metrics.labels.emplace_back("name", bus.name);
1020 if (bus.device.type == InputSourceType::SILENCE) {
1021 metrics.labels.emplace_back("source_type", "silence");
1022 } else if (bus.device.type == InputSourceType::CAPTURE_CARD) {
1023 metrics.labels.emplace_back("source_type", "capture_card");
1024 } else if (bus.device.type == InputSourceType::ALSA_INPUT) {
1025 metrics.labels.emplace_back("source_type", "alsa_input");
1029 metrics.labels.emplace_back("source_index", source_index_str);
1030 metrics.labels.emplace_back("source_channels", source_channels_str);
1032 vector<pair<string, string>> labels_left = metrics.labels;
1033 labels_left.emplace_back("channel", "left");
1034 vector<pair<string, string>> labels_right = metrics.labels;
1035 labels_right.emplace_back("channel", "right");
1037 global_metrics.add("bus_current_level_dbfs", labels_left, &metrics.current_level_dbfs[0], Metrics::TYPE_GAUGE);
1038 global_metrics.add("bus_current_level_dbfs", labels_right, &metrics.current_level_dbfs[1], Metrics::TYPE_GAUGE);
1039 global_metrics.add("bus_peak_level_dbfs", labels_left, &metrics.peak_level_dbfs[0], Metrics::TYPE_GAUGE);
1040 global_metrics.add("bus_peak_level_dbfs", labels_right, &metrics.peak_level_dbfs[1], Metrics::TYPE_GAUGE);
1041 global_metrics.add("bus_historic_peak_dbfs", metrics.labels, &metrics.historic_peak_dbfs, Metrics::TYPE_GAUGE);
1042 global_metrics.add("bus_gain_staging_db", metrics.labels, &metrics.gain_staging_db, Metrics::TYPE_GAUGE);
1043 global_metrics.add("bus_compressor_attenuation_db", metrics.labels, &metrics.compressor_attenuation_db, Metrics::TYPE_GAUGE);
1046 // Reset resamplers for all cards that don't have the exact same state as before.
1047 for (unsigned card_index = 0; card_index < MAX_VIDEO_CARDS; ++card_index) {
1048 const DeviceSpec device_spec{InputSourceType::CAPTURE_CARD, card_index};
1049 AudioDevice *device = find_audio_device(device_spec);
1050 if (device->interesting_channels != interesting_channels[device_spec]) {
1051 device->interesting_channels = interesting_channels[device_spec];
1052 reset_resampler_mutex_held(device_spec);
1055 for (unsigned card_index = 0; card_index < MAX_ALSA_CARDS; ++card_index) {
1056 const DeviceSpec device_spec{InputSourceType::ALSA_INPUT, card_index};
1057 AudioDevice *device = find_audio_device(device_spec);
1058 if (interesting_channels[device_spec].empty()) {
1059 alsa_pool.release_device(card_index);
1061 alsa_pool.hold_device(card_index);
1063 if (device->interesting_channels != interesting_channels[device_spec]) {
1064 device->interesting_channels = interesting_channels[device_spec];
1065 alsa_pool.reset_device(device_spec.index);
1066 reset_resampler_mutex_held(device_spec);
1070 input_mapping = new_input_mapping;
1073 InputMapping AudioMixer::get_input_mapping() const
1075 lock_guard<timed_mutex> lock(audio_mutex);
1076 return input_mapping;
1079 unsigned AudioMixer::num_buses() const
1081 lock_guard<timed_mutex> lock(audio_mutex);
1082 return input_mapping.buses.size();
1085 void AudioMixer::reset_peak(unsigned bus_index)
1087 lock_guard<timed_mutex> lock(audio_mutex);
1088 for (unsigned channel = 0; channel < 2; ++channel) {
1089 PeakHistory &history = peak_history[bus_index][channel];
1090 history.current_level = 0.0f;
1091 history.historic_peak = 0.0f;
1092 history.current_peak = 0.0f;
1093 history.last_peak = 0.0f;
1094 history.age_seconds = 0.0f;
1098 AudioMixer *global_audio_mixer = nullptr;