1 #include "audio_mixer.h"
5 #include <bmusb/bmusb.h>
10 #include <immintrin.h>
17 using namespace bmusb;
19 using namespace std::placeholders;
23 // TODO: If these prove to be a bottleneck, they can be SSSE3-optimized
24 // (usually including multiple channels at a time).
26 void convert_fixed16_to_fp32(float *dst, size_t out_channel, size_t out_num_channels,
27 const uint8_t *src, size_t in_channel, size_t in_num_channels,
30 assert(in_channel < in_num_channels);
31 assert(out_channel < out_num_channels);
32 src += in_channel * 2;
35 for (size_t i = 0; i < num_samples; ++i) {
36 int16_t s = le16toh(*(int16_t *)src);
37 *dst = s * (1.0f / 32768.0f);
39 src += 2 * in_num_channels;
40 dst += out_num_channels;
44 void convert_fixed24_to_fp32(float *dst, size_t out_channel, size_t out_num_channels,
45 const uint8_t *src, size_t in_channel, size_t in_num_channels,
48 assert(in_channel < in_num_channels);
49 assert(out_channel < out_num_channels);
50 src += in_channel * 3;
53 for (size_t i = 0; i < num_samples; ++i) {
57 uint32_t s = s1 | (s1 << 8) | (s2 << 16) | (s3 << 24);
58 *dst = int(s) * (1.0f / 2147483648.0f);
60 src += 3 * in_num_channels;
61 dst += out_num_channels;
65 void convert_fixed32_to_fp32(float *dst, size_t out_channel, size_t out_num_channels,
66 const uint8_t *src, size_t in_channel, size_t in_num_channels,
69 assert(in_channel < in_num_channels);
70 assert(out_channel < out_num_channels);
71 src += in_channel * 4;
74 for (size_t i = 0; i < num_samples; ++i) {
75 int32_t s = le32toh(*(int32_t *)src);
76 *dst = s * (1.0f / 2147483648.0f);
78 src += 4 * in_num_channels;
79 dst += out_num_channels;
83 float find_peak_plain(const float *samples, size_t num_samples) __attribute__((unused));
85 float find_peak_plain(const float *samples, size_t num_samples)
87 float m = fabs(samples[0]);
88 for (size_t i = 1; i < num_samples; ++i) {
89 m = max(m, fabs(samples[i]));
95 static inline float horizontal_max(__m128 m)
97 __m128 tmp = _mm_shuffle_ps(m, m, _MM_SHUFFLE(1, 0, 3, 2));
98 m = _mm_max_ps(m, tmp);
99 tmp = _mm_shuffle_ps(m, m, _MM_SHUFFLE(2, 3, 0, 1));
100 m = _mm_max_ps(m, tmp);
101 return _mm_cvtss_f32(m);
104 float find_peak(const float *samples, size_t num_samples)
106 const __m128 abs_mask = _mm_castsi128_ps(_mm_set1_epi32(0x7fffffffu));
107 __m128 m = _mm_setzero_ps();
108 for (size_t i = 0; i < (num_samples & ~3); i += 4) {
109 __m128 x = _mm_loadu_ps(samples + i);
110 x = _mm_and_ps(x, abs_mask);
111 m = _mm_max_ps(m, x);
113 float result = horizontal_max(m);
115 for (size_t i = (num_samples & ~3); i < num_samples; ++i) {
116 result = max(result, fabs(samples[i]));
120 // Self-test. We should be bit-exact the same.
121 float reference_result = find_peak_plain(samples, num_samples);
122 if (result != reference_result) {
123 fprintf(stderr, "Error: Peak is %f [%f %f %f %f]; should be %f.\n",
125 _mm_cvtss_f32(_mm_shuffle_ps(m, m, _MM_SHUFFLE(0, 0, 0, 0))),
126 _mm_cvtss_f32(_mm_shuffle_ps(m, m, _MM_SHUFFLE(1, 1, 1, 1))),
127 _mm_cvtss_f32(_mm_shuffle_ps(m, m, _MM_SHUFFLE(2, 2, 2, 2))),
128 _mm_cvtss_f32(_mm_shuffle_ps(m, m, _MM_SHUFFLE(3, 3, 3, 3))),
136 float find_peak(const float *samples, size_t num_samples)
138 return find_peak_plain(samples, num_samples);
142 void deinterleave_samples(const vector<float> &in, vector<float> *out_l, vector<float> *out_r)
144 size_t num_samples = in.size() / 2;
145 out_l->resize(num_samples);
146 out_r->resize(num_samples);
148 const float *inptr = in.data();
149 float *lptr = &(*out_l)[0];
150 float *rptr = &(*out_r)[0];
151 for (size_t i = 0; i < num_samples; ++i) {
159 AudioMixer::AudioMixer(unsigned num_cards)
160 : num_cards(num_cards),
161 limiter(OUTPUT_FREQUENCY),
162 correlation(OUTPUT_FREQUENCY)
164 for (unsigned bus_index = 0; bus_index < MAX_BUSES; ++bus_index) {
165 locut[bus_index].init(FILTER_HPF, 2);
166 locut_enabled[bus_index] = global_flags.locut_enabled;
167 eq[bus_index][EQ_BAND_BASS].init(FILTER_LOW_SHELF, 1);
168 // Note: EQ_BAND_MID isn't used (see comments in apply_eq()).
169 eq[bus_index][EQ_BAND_TREBLE].init(FILTER_HIGH_SHELF, 1);
171 gain_staging_db[bus_index] = global_flags.initial_gain_staging_db;
172 compressor[bus_index].reset(new StereoCompressor(OUTPUT_FREQUENCY));
173 compressor_threshold_dbfs[bus_index] = ref_level_dbfs - 12.0f; // -12 dB.
174 compressor_enabled[bus_index] = global_flags.compressor_enabled;
175 level_compressor[bus_index].reset(new StereoCompressor(OUTPUT_FREQUENCY));
176 level_compressor_enabled[bus_index] = global_flags.gain_staging_auto;
178 set_limiter_enabled(global_flags.limiter_enabled);
179 set_final_makeup_gain_auto(global_flags.final_makeup_gain_auto);
181 // Generate a very simple, default input mapping.
182 InputMapping::Bus input;
184 input.device.type = InputSourceType::CAPTURE_CARD;
185 input.device.index = 0;
186 input.source_channel[0] = 0;
187 input.source_channel[1] = 1;
189 InputMapping new_input_mapping;
190 new_input_mapping.buses.push_back(input);
191 set_input_mapping(new_input_mapping);
193 // Look for ALSA cards.
194 available_alsa_cards = ALSAInput::enumerate_devices();
196 r128.init(2, OUTPUT_FREQUENCY);
199 // hlen=16 is pretty low quality, but we use quite a bit of CPU otherwise,
200 // and there's a limit to how important the peak meter is.
201 peak_resampler.setup(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY * 4, /*num_channels=*/2, /*hlen=*/16, /*frel=*/1.0);
204 AudioMixer::~AudioMixer()
206 for (unsigned card_index = 0; card_index < available_alsa_cards.size(); ++card_index) {
207 const AudioDevice &device = alsa_inputs[card_index];
208 if (device.alsa_device != nullptr) {
209 device.alsa_device->stop_capture_thread();
215 void AudioMixer::reset_resampler(DeviceSpec device_spec)
217 lock_guard<timed_mutex> lock(audio_mutex);
218 reset_resampler_mutex_held(device_spec);
221 void AudioMixer::reset_resampler_mutex_held(DeviceSpec device_spec)
223 AudioDevice *device = find_audio_device(device_spec);
225 if (device->interesting_channels.empty()) {
226 device->resampling_queue.reset();
228 // TODO: ResamplingQueue should probably take the full device spec.
229 // (It's only used for console output, though.)
230 device->resampling_queue.reset(new ResamplingQueue(device_spec.index, device->capture_frequency, OUTPUT_FREQUENCY, device->interesting_channels.size()));
232 device->next_local_pts = 0;
235 void AudioMixer::reset_alsa_mutex_held(DeviceSpec device_spec)
237 assert(device_spec.type == InputSourceType::ALSA_INPUT);
238 unsigned card_index = device_spec.index;
239 AudioDevice *device = find_audio_device(device_spec);
241 if (device->alsa_device != nullptr) {
242 device->alsa_device->stop_capture_thread();
244 if (device->interesting_channels.empty()) {
245 device->alsa_device.reset();
247 const ALSAInput::Device &alsa_dev = available_alsa_cards[card_index];
248 device->alsa_device.reset(new ALSAInput(alsa_dev.address.c_str(), OUTPUT_FREQUENCY, alsa_dev.num_channels, bind(&AudioMixer::add_audio, this, device_spec, _1, _2, _3, _4)));
249 device->capture_frequency = device->alsa_device->get_sample_rate();
250 device->alsa_device->start_capture_thread();
254 bool AudioMixer::add_audio(DeviceSpec device_spec, const uint8_t *data, unsigned num_samples, AudioFormat audio_format, int64_t frame_length)
256 AudioDevice *device = find_audio_device(device_spec);
258 unique_lock<timed_mutex> lock(audio_mutex, defer_lock);
259 if (!lock.try_lock_for(chrono::milliseconds(10))) {
262 if (device->resampling_queue == nullptr) {
263 // No buses use this device; throw it away.
267 unsigned num_channels = device->interesting_channels.size();
268 assert(num_channels > 0);
270 // Convert the audio to fp32.
271 unique_ptr<float[]> audio(new float[num_samples * num_channels]);
272 unsigned channel_index = 0;
273 for (auto channel_it = device->interesting_channels.cbegin(); channel_it != device->interesting_channels.end(); ++channel_it, ++channel_index) {
274 switch (audio_format.bits_per_sample) {
276 assert(num_samples == 0);
279 convert_fixed16_to_fp32(audio.get(), channel_index, num_channels, data, *channel_it, audio_format.num_channels, num_samples);
282 convert_fixed24_to_fp32(audio.get(), channel_index, num_channels, data, *channel_it, audio_format.num_channels, num_samples);
285 convert_fixed32_to_fp32(audio.get(), channel_index, num_channels, data, *channel_it, audio_format.num_channels, num_samples);
288 fprintf(stderr, "Cannot handle audio with %u bits per sample\n", audio_format.bits_per_sample);
294 int64_t local_pts = device->next_local_pts;
295 device->resampling_queue->add_input_samples(local_pts / double(TIMEBASE), audio.get(), num_samples);
296 device->next_local_pts = local_pts + frame_length;
300 bool AudioMixer::add_silence(DeviceSpec device_spec, unsigned samples_per_frame, unsigned num_frames, int64_t frame_length)
302 AudioDevice *device = find_audio_device(device_spec);
304 unique_lock<timed_mutex> lock(audio_mutex, defer_lock);
305 if (!lock.try_lock_for(chrono::milliseconds(10))) {
308 if (device->resampling_queue == nullptr) {
309 // No buses use this device; throw it away.
313 unsigned num_channels = device->interesting_channels.size();
314 assert(num_channels > 0);
316 vector<float> silence(samples_per_frame * num_channels, 0.0f);
317 for (unsigned i = 0; i < num_frames; ++i) {
318 device->resampling_queue->add_input_samples(device->next_local_pts / double(TIMEBASE), silence.data(), samples_per_frame);
319 // Note that if the format changed in the meantime, we have
320 // no way of detecting that; we just have to assume the frame length
321 // is always the same.
322 device->next_local_pts += frame_length;
327 AudioMixer::AudioDevice *AudioMixer::find_audio_device(DeviceSpec device)
329 switch (device.type) {
330 case InputSourceType::CAPTURE_CARD:
331 return &video_cards[device.index];
332 case InputSourceType::ALSA_INPUT:
333 return &alsa_inputs[device.index];
334 case InputSourceType::SILENCE:
341 // Get a pointer to the given channel from the given device.
342 // The channel must be picked out earlier and resampled.
343 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)
345 static float zero = 0.0f;
346 if (source_channel == -1 || device_spec.type == InputSourceType::SILENCE) {
351 AudioDevice *device = find_audio_device(device_spec);
352 assert(device->interesting_channels.count(source_channel) != 0);
353 unsigned channel_index = 0;
354 for (int channel : device->interesting_channels) {
355 if (channel == source_channel) break;
358 assert(channel_index < device->interesting_channels.size());
359 const auto it = samples_card.find(device_spec);
360 assert(it != samples_card.end());
361 *srcptr = &(it->second)[channel_index];
362 *stride = device->interesting_channels.size();
365 // TODO: Can be SSSE3-optimized if need be.
366 void AudioMixer::fill_audio_bus(const map<DeviceSpec, vector<float>> &samples_card, const InputMapping::Bus &bus, unsigned num_samples, float *output)
368 if (bus.device.type == InputSourceType::SILENCE) {
369 memset(output, 0, num_samples * sizeof(*output));
371 assert(bus.device.type == InputSourceType::CAPTURE_CARD ||
372 bus.device.type == InputSourceType::ALSA_INPUT);
373 const float *lsrc, *rsrc;
374 unsigned lstride, rstride;
375 float *dptr = output;
376 find_sample_src_from_device(samples_card, bus.device, bus.source_channel[0], &lsrc, &lstride);
377 find_sample_src_from_device(samples_card, bus.device, bus.source_channel[1], &rsrc, &rstride);
378 for (unsigned i = 0; i < num_samples; ++i) {
387 vector<float> AudioMixer::get_output(double pts, unsigned num_samples, ResamplingQueue::RateAdjustmentPolicy rate_adjustment_policy)
389 map<DeviceSpec, vector<float>> samples_card;
390 vector<float> samples_bus;
392 lock_guard<timed_mutex> lock(audio_mutex);
394 // Pick out all the interesting channels from all the cards.
395 // TODO: If the card has been hotswapped, the number of channels
396 // might have changed; if so, we need to do some sort of remapping
398 for (const auto &spec_and_info : get_devices_mutex_held()) {
399 const DeviceSpec &device_spec = spec_and_info.first;
400 AudioDevice *device = find_audio_device(device_spec);
401 if (!device->interesting_channels.empty()) {
402 samples_card[device_spec].resize(num_samples * device->interesting_channels.size());
403 device->resampling_queue->get_output_samples(
405 &samples_card[device_spec][0],
407 rate_adjustment_policy);
411 vector<float> samples_out, left, right;
412 samples_out.resize(num_samples * 2);
413 samples_bus.resize(num_samples * 2);
414 for (unsigned bus_index = 0; bus_index < input_mapping.buses.size(); ++bus_index) {
415 fill_audio_bus(samples_card, input_mapping.buses[bus_index], num_samples, &samples_bus[0]);
416 apply_eq(bus_index, &samples_bus);
419 lock_guard<mutex> lock(compressor_mutex);
421 // Apply a level compressor to get the general level right.
422 // Basically, if it's over about -40 dBFS, we squeeze it down to that level
423 // (or more precisely, near it, since we don't use infinite ratio),
424 // then apply a makeup gain to get it to -14 dBFS. -14 dBFS is, of course,
425 // entirely arbitrary, but from practical tests with speech, it seems to
426 // put ut around -23 LUFS, so it's a reasonable starting point for later use.
427 if (level_compressor_enabled[bus_index]) {
428 float threshold = 0.01f; // -40 dBFS.
430 float attack_time = 0.5f;
431 float release_time = 20.0f;
432 float makeup_gain = from_db(ref_level_dbfs - (-40.0f)); // +26 dB.
433 level_compressor[bus_index]->process(samples_bus.data(), samples_bus.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
434 gain_staging_db[bus_index] = to_db(level_compressor[bus_index]->get_attenuation() * makeup_gain);
436 // Just apply the gain we already had.
437 float g = from_db(gain_staging_db[bus_index]);
438 for (size_t i = 0; i < samples_bus.size(); ++i) {
444 printf("level=%f (%+5.2f dBFS) attenuation=%f (%+5.2f dB) end_result=%+5.2f dB\n",
445 level_compressor.get_level(), to_db(level_compressor.get_level()),
446 level_compressor.get_attenuation(), to_db(level_compressor.get_attenuation()),
447 to_db(level_compressor.get_level() * level_compressor.get_attenuation() * makeup_gain));
450 // The real compressor.
451 if (compressor_enabled[bus_index]) {
452 float threshold = from_db(compressor_threshold_dbfs[bus_index]);
454 float attack_time = 0.005f;
455 float release_time = 0.040f;
456 float makeup_gain = 2.0f; // +6 dB.
457 compressor[bus_index]->process(samples_bus.data(), samples_bus.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
458 // compressor_att = compressor.get_attenuation();
462 add_bus_to_master(bus_index, samples_bus, &samples_out);
463 deinterleave_samples(samples_bus, &left, &right);
464 measure_bus_levels(bus_index, left, right);
468 lock_guard<mutex> lock(compressor_mutex);
470 // Finally a limiter at -4 dB (so, -10 dBFS) to take out the worst peaks only.
471 // Note that since ratio is not infinite, we could go slightly higher than this.
472 if (limiter_enabled) {
473 float threshold = from_db(limiter_threshold_dbfs);
475 float attack_time = 0.0f; // Instant.
476 float release_time = 0.020f;
477 float makeup_gain = 1.0f; // 0 dB.
478 limiter.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
479 // limiter_att = limiter.get_attenuation();
482 // printf("limiter=%+5.1f compressor=%+5.1f\n", to_db(limiter_att), to_db(compressor_att));
485 // At this point, we are most likely close to +0 LU (at least if the
486 // faders sum to 0 dB and the compressors are on), but all of our
487 // measurements have been on raw sample values, not R128 values.
488 // So we have a final makeup gain to get us to +0 LU; the gain
489 // adjustments required should be relatively small, and also, the
490 // offset shouldn't change much (only if the type of audio changes
491 // significantly). Thus, we shoot for updating this value basically
492 // “whenever we process buffers”, since the R128 calculation isn't exactly
493 // something we get out per-sample.
495 // Note that there's a feedback loop here, so we choose a very slow filter
496 // (half-time of 30 seconds).
497 double target_loudness_factor, alpha;
498 double loudness_lu = r128.loudness_M() - ref_level_lufs;
499 double current_makeup_lu = to_db(final_makeup_gain);
500 target_loudness_factor = final_makeup_gain * from_db(-loudness_lu);
502 // If we're outside +/- 5 LU uncorrected, we don't count it as
503 // a normal signal (probably silence) and don't change the
504 // correction factor; just apply what we already have.
505 if (fabs(loudness_lu - current_makeup_lu) >= 5.0 || !final_makeup_gain_auto) {
508 // Formula adapted from
509 // https://en.wikipedia.org/wiki/Low-pass_filter#Simple_infinite_impulse_response_filter.
510 const double half_time_s = 30.0;
511 const double fc_mul_2pi_delta_t = 1.0 / (half_time_s * OUTPUT_FREQUENCY);
512 alpha = fc_mul_2pi_delta_t / (fc_mul_2pi_delta_t + 1.0);
516 lock_guard<mutex> lock(compressor_mutex);
517 double m = final_makeup_gain;
518 for (size_t i = 0; i < samples_out.size(); i += 2) {
519 samples_out[i + 0] *= m;
520 samples_out[i + 1] *= m;
521 m += (target_loudness_factor - m) * alpha;
523 final_makeup_gain = m;
526 update_meters(samples_out);
531 void AudioMixer::apply_eq(unsigned bus_index, vector<float> *samples_bus)
533 constexpr float bass_freq_hz = 200.0f;
534 constexpr float treble_freq_hz = 4700.0f;
536 // Cut away everything under 120 Hz (or whatever the cutoff is);
537 // we don't need it for voice, and it will reduce headroom
538 // and confuse the compressor. (In particular, any hums at 50 or 60 Hz
539 // should be dampened.)
540 if (locut_enabled[bus_index]) {
541 locut[bus_index].render(samples_bus->data(), samples_bus->size() / 2, locut_cutoff_hz * 2.0 * M_PI / OUTPUT_FREQUENCY, 0.5f);
544 // Apply the rest of the EQ. Since we only have a simple three-band EQ,
545 // we can implement it with two shelf filters. We use a simple gain to
546 // set the mid-level filter, and then offset the low and high bands
547 // from that if we need to. (We could perhaps have folded the gain into
548 // the next part, but it's so cheap that the trouble isn't worth it.)
549 if (fabs(eq_level_db[bus_index][EQ_BAND_MID]) > 0.01f) {
550 float g = from_db(eq_level_db[bus_index][EQ_BAND_MID]);
551 for (size_t i = 0; i < samples_bus->size(); ++i) {
552 (*samples_bus)[i] *= g;
556 float bass_adj_db = eq_level_db[bus_index][EQ_BAND_BASS] - eq_level_db[bus_index][EQ_BAND_MID];
557 if (fabs(bass_adj_db) > 0.01f) {
558 eq[bus_index][EQ_BAND_BASS].render(samples_bus->data(), samples_bus->size() / 2,
559 bass_freq_hz * 2.0 * M_PI / OUTPUT_FREQUENCY, 0.5f, bass_adj_db / 40.0f);
562 float treble_adj_db = eq_level_db[bus_index][EQ_BAND_TREBLE] - eq_level_db[bus_index][EQ_BAND_MID];
563 if (fabs(treble_adj_db) > 0.01f) {
564 eq[bus_index][EQ_BAND_TREBLE].render(samples_bus->data(), samples_bus->size() / 2,
565 treble_freq_hz * 2.0 * M_PI / OUTPUT_FREQUENCY, 0.5f, treble_adj_db / 40.0f);
569 void AudioMixer::add_bus_to_master(unsigned bus_index, const vector<float> &samples_bus, vector<float> *samples_out)
571 assert(samples_bus.size() == samples_out->size());
572 assert(samples_bus.size() % 2 == 0);
573 unsigned num_samples = samples_bus.size() / 2;
574 if (fabs(fader_volume_db[bus_index] - last_fader_volume_db[bus_index]) > 1e-3) {
575 // The volume has changed; do a fade over the course of this frame.
576 // (We might have some numerical issues here, but it seems to sound OK.)
577 // For the purpose of fading here, the silence floor is set to -90 dB
578 // (the fader only goes to -84).
579 float old_volume = from_db(max<float>(last_fader_volume_db[bus_index], -90.0f));
580 float volume = from_db(max<float>(fader_volume_db[bus_index], -90.0f));
582 float volume_inc = pow(volume / old_volume, 1.0 / num_samples);
584 if (bus_index == 0) {
585 for (unsigned i = 0; i < num_samples; ++i) {
586 (*samples_out)[i * 2 + 0] = samples_bus[i * 2 + 0] * volume;
587 (*samples_out)[i * 2 + 1] = samples_bus[i * 2 + 1] * volume;
588 volume *= volume_inc;
591 for (unsigned i = 0; i < num_samples; ++i) {
592 (*samples_out)[i * 2 + 0] += samples_bus[i * 2 + 0] * volume;
593 (*samples_out)[i * 2 + 1] += samples_bus[i * 2 + 1] * volume;
594 volume *= volume_inc;
598 float volume = from_db(fader_volume_db[bus_index]);
599 if (bus_index == 0) {
600 for (unsigned i = 0; i < num_samples; ++i) {
601 (*samples_out)[i * 2 + 0] = samples_bus[i * 2 + 0] * volume;
602 (*samples_out)[i * 2 + 1] = samples_bus[i * 2 + 1] * volume;
605 for (unsigned i = 0; i < num_samples; ++i) {
606 (*samples_out)[i * 2 + 0] += samples_bus[i * 2 + 0] * volume;
607 (*samples_out)[i * 2 + 1] += samples_bus[i * 2 + 1] * volume;
612 last_fader_volume_db[bus_index] = fader_volume_db[bus_index];
615 void AudioMixer::measure_bus_levels(unsigned bus_index, const vector<float> &left, const vector<float> &right)
617 assert(left.size() == right.size());
618 const float volume = from_db(fader_volume_db[bus_index]);
619 const float peak_levels[2] = {
620 find_peak(left.data(), left.size()) * volume,
621 find_peak(right.data(), right.size()) * volume
623 for (unsigned channel = 0; channel < 2; ++channel) {
624 // Compute the current value, including hold and falloff.
625 // The constants are borrowed from zita-mu1 by Fons Adriaensen.
626 static constexpr float hold_sec = 0.5f;
627 static constexpr float falloff_db_sec = 15.0f; // dB/sec falloff after hold.
629 PeakHistory &history = peak_history[bus_index][channel];
630 history.historic_peak = max(history.historic_peak, peak_levels[channel]);
631 if (history.age_seconds < hold_sec) {
632 current_peak = history.last_peak;
634 current_peak = history.last_peak * from_db(-falloff_db_sec * (history.age_seconds - hold_sec));
637 // See if we have a new peak to replace the old (possibly falling) one.
638 if (peak_levels[channel] > current_peak) {
639 history.last_peak = peak_levels[channel];
640 history.age_seconds = 0.0f; // Not 100% correct, but more than good enough given our frame sizes.
641 current_peak = peak_levels[channel];
643 history.age_seconds += float(left.size()) / OUTPUT_FREQUENCY;
645 history.current_level = peak_levels[channel];
646 history.current_peak = current_peak;
650 void AudioMixer::update_meters(const vector<float> &samples)
652 // Upsample 4x to find interpolated peak.
653 peak_resampler.inp_data = const_cast<float *>(samples.data());
654 peak_resampler.inp_count = samples.size() / 2;
656 vector<float> interpolated_samples;
657 interpolated_samples.resize(samples.size());
659 lock_guard<mutex> lock(audio_measure_mutex);
661 while (peak_resampler.inp_count > 0) { // About four iterations.
662 peak_resampler.out_data = &interpolated_samples[0];
663 peak_resampler.out_count = interpolated_samples.size() / 2;
664 peak_resampler.process();
665 size_t out_stereo_samples = interpolated_samples.size() / 2 - peak_resampler.out_count;
666 peak = max<float>(peak, find_peak(interpolated_samples.data(), out_stereo_samples * 2));
667 peak_resampler.out_data = nullptr;
671 // Find R128 levels and L/R correlation.
672 vector<float> left, right;
673 deinterleave_samples(samples, &left, &right);
674 float *ptrs[] = { left.data(), right.data() };
676 lock_guard<mutex> lock(audio_measure_mutex);
677 r128.process(left.size(), ptrs);
678 correlation.process_samples(samples);
681 send_audio_level_callback();
684 void AudioMixer::reset_meters()
686 lock_guard<mutex> lock(audio_measure_mutex);
687 peak_resampler.reset();
694 void AudioMixer::send_audio_level_callback()
696 if (audio_level_callback == nullptr) {
700 lock_guard<mutex> lock(audio_measure_mutex);
701 double loudness_s = r128.loudness_S();
702 double loudness_i = r128.integrated();
703 double loudness_range_low = r128.range_min();
704 double loudness_range_high = r128.range_max();
706 vector<BusLevel> bus_levels;
707 bus_levels.resize(input_mapping.buses.size());
709 lock_guard<mutex> lock(compressor_mutex);
710 for (unsigned bus_index = 0; bus_index < bus_levels.size(); ++bus_index) {
711 bus_levels[bus_index].current_level_dbfs[0] = to_db(peak_history[bus_index][0].current_level);
712 bus_levels[bus_index].current_level_dbfs[1] = to_db(peak_history[bus_index][1].current_level);
713 bus_levels[bus_index].peak_level_dbfs[0] = to_db(peak_history[bus_index][0].current_peak);
714 bus_levels[bus_index].peak_level_dbfs[1] = to_db(peak_history[bus_index][1].current_peak);
715 bus_levels[bus_index].historic_peak_dbfs = to_db(
716 max(peak_history[bus_index][0].historic_peak,
717 peak_history[bus_index][1].historic_peak));
718 bus_levels[bus_index].gain_staging_db = gain_staging_db[bus_index];
719 if (compressor_enabled[bus_index]) {
720 bus_levels[bus_index].compressor_attenuation_db = -to_db(compressor[bus_index]->get_attenuation());
722 bus_levels[bus_index].compressor_attenuation_db = 0.0;
727 audio_level_callback(loudness_s, to_db(peak), bus_levels,
728 loudness_i, loudness_range_low, loudness_range_high,
729 to_db(final_makeup_gain),
730 correlation.get_correlation());
733 map<DeviceSpec, DeviceInfo> AudioMixer::get_devices() const
735 lock_guard<timed_mutex> lock(audio_mutex);
736 return get_devices_mutex_held();
739 map<DeviceSpec, DeviceInfo> AudioMixer::get_devices_mutex_held() const
741 map<DeviceSpec, DeviceInfo> devices;
742 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
743 const DeviceSpec spec{ InputSourceType::CAPTURE_CARD, card_index };
744 const AudioDevice *device = &video_cards[card_index];
746 info.name = device->name;
747 info.num_channels = 8; // FIXME: This is wrong for fake cards.
748 devices.insert(make_pair(spec, info));
750 for (unsigned card_index = 0; card_index < available_alsa_cards.size(); ++card_index) {
751 const DeviceSpec spec{ InputSourceType::ALSA_INPUT, card_index };
752 const ALSAInput::Device &device = available_alsa_cards[card_index];
754 info.name = device.name + " (" + device.info + ")";
755 info.num_channels = device.num_channels;
756 devices.insert(make_pair(spec, info));
761 void AudioMixer::set_name(DeviceSpec device_spec, const string &name)
763 AudioDevice *device = find_audio_device(device_spec);
765 lock_guard<timed_mutex> lock(audio_mutex);
769 void AudioMixer::set_input_mapping(const InputMapping &new_input_mapping)
771 lock_guard<timed_mutex> lock(audio_mutex);
773 map<DeviceSpec, set<unsigned>> interesting_channels;
774 for (const InputMapping::Bus &bus : new_input_mapping.buses) {
775 if (bus.device.type == InputSourceType::CAPTURE_CARD ||
776 bus.device.type == InputSourceType::ALSA_INPUT) {
777 for (unsigned channel = 0; channel < 2; ++channel) {
778 if (bus.source_channel[channel] != -1) {
779 interesting_channels[bus.device].insert(bus.source_channel[channel]);
785 // Reset resamplers for all cards that don't have the exact same state as before.
786 for (const auto &spec_and_info : get_devices_mutex_held()) {
787 const DeviceSpec &device_spec = spec_and_info.first;
788 AudioDevice *device = find_audio_device(device_spec);
789 if (device->interesting_channels != interesting_channels[device_spec]) {
790 device->interesting_channels = interesting_channels[device_spec];
791 if (device_spec.type == InputSourceType::ALSA_INPUT) {
792 reset_alsa_mutex_held(device_spec);
794 reset_resampler_mutex_held(device_spec);
798 input_mapping = new_input_mapping;
801 InputMapping AudioMixer::get_input_mapping() const
803 lock_guard<timed_mutex> lock(audio_mutex);
804 return input_mapping;
807 void AudioMixer::reset_peak(unsigned bus_index)
809 lock_guard<timed_mutex> lock(audio_mutex);
810 for (unsigned channel = 0; channel < 2; ++channel) {
811 PeakHistory &history = peak_history[bus_index][channel];
812 history.current_level = 0.0f;
813 history.historic_peak = 0.0f;
814 history.current_peak = 0.0f;
815 history.last_peak = 0.0f;
816 history.age_seconds = 0.0f;