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_capture_cards, unsigned num_ffmpeg_inputs)
171 : num_capture_cards(num_capture_cards),
172 num_ffmpeg_inputs(num_ffmpeg_inputs),
173 ffmpeg_inputs(new AudioDevice[num_ffmpeg_inputs]),
174 limiter(OUTPUT_FREQUENCY),
175 correlation(OUTPUT_FREQUENCY)
177 for (unsigned bus_index = 0; bus_index < MAX_BUSES; ++bus_index) {
178 locut[bus_index].init(FILTER_HPF, 2);
179 eq[bus_index][EQ_BAND_BASS].init(FILTER_LOW_SHELF, 1);
180 // Note: EQ_BAND_MID isn't used (see comments in apply_eq()).
181 eq[bus_index][EQ_BAND_TREBLE].init(FILTER_HIGH_SHELF, 1);
182 compressor[bus_index].reset(new StereoCompressor(OUTPUT_FREQUENCY));
183 level_compressor[bus_index].reset(new StereoCompressor(OUTPUT_FREQUENCY));
185 set_bus_settings(bus_index, get_default_bus_settings());
187 set_limiter_enabled(global_flags.limiter_enabled);
188 set_final_makeup_gain_auto(global_flags.final_makeup_gain_auto);
190 r128.init(2, OUTPUT_FREQUENCY);
193 // hlen=16 is pretty low quality, but we use quite a bit of CPU otherwise,
194 // and there's a limit to how important the peak meter is.
195 peak_resampler.setup(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY * 4, /*num_channels=*/2, /*hlen=*/16, /*frel=*/1.0);
197 global_audio_mixer = this;
200 if (!global_flags.input_mapping_filename.empty()) {
201 // Must happen after ALSAPool is initialized, as it needs to know the card list.
202 current_mapping_mode = MappingMode::MULTICHANNEL;
203 InputMapping new_input_mapping;
204 if (!load_input_mapping_from_file(get_devices(),
205 global_flags.input_mapping_filename,
206 &new_input_mapping)) {
207 fprintf(stderr, "Failed to load input mapping from '%s', exiting.\n",
208 global_flags.input_mapping_filename.c_str());
211 set_input_mapping(new_input_mapping);
213 set_simple_input(/*card_index=*/0);
214 if (global_flags.multichannel_mapping_mode) {
215 current_mapping_mode = MappingMode::MULTICHANNEL;
219 global_metrics.add("audio_loudness_short_lufs", &metric_audio_loudness_short_lufs, Metrics::TYPE_GAUGE);
220 global_metrics.add("audio_loudness_integrated_lufs", &metric_audio_loudness_integrated_lufs, Metrics::TYPE_GAUGE);
221 global_metrics.add("audio_loudness_range_low_lufs", &metric_audio_loudness_range_low_lufs, Metrics::TYPE_GAUGE);
222 global_metrics.add("audio_loudness_range_high_lufs", &metric_audio_loudness_range_high_lufs, Metrics::TYPE_GAUGE);
223 global_metrics.add("audio_peak_dbfs", &metric_audio_peak_dbfs, Metrics::TYPE_GAUGE);
224 global_metrics.add("audio_final_makeup_gain_db", &metric_audio_final_makeup_gain_db, Metrics::TYPE_GAUGE);
225 global_metrics.add("audio_correlation", &metric_audio_correlation, Metrics::TYPE_GAUGE);
228 void AudioMixer::reset_resampler(DeviceSpec device_spec)
230 lock_guard<timed_mutex> lock(audio_mutex);
231 reset_resampler_mutex_held(device_spec);
234 void AudioMixer::reset_resampler_mutex_held(DeviceSpec device_spec)
236 AudioDevice *device = find_audio_device(device_spec);
238 if (device->interesting_channels.empty()) {
239 device->resampling_queue.reset();
241 device->resampling_queue.reset(new ResamplingQueue(
242 device_spec, 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 settings.stereo_width = 1.0f;
343 for (unsigned band_index = 0; band_index < NUM_EQ_BANDS; ++band_index) {
344 settings.eq_level_db[band_index] = 0.0f;
346 settings.gain_staging_db = global_flags.initial_gain_staging_db;
347 settings.level_compressor_enabled = global_flags.gain_staging_auto;
348 settings.compressor_threshold_dbfs = ref_level_dbfs - 12.0f; // -12 dB.
349 settings.compressor_enabled = global_flags.compressor_enabled;
353 AudioMixer::BusSettings AudioMixer::get_bus_settings(unsigned bus_index) const
355 lock_guard<timed_mutex> lock(audio_mutex);
356 BusSettings settings;
357 settings.fader_volume_db = fader_volume_db[bus_index];
358 settings.muted = mute[bus_index];
359 settings.locut_enabled = locut_enabled[bus_index];
360 settings.stereo_width = stereo_width[bus_index];
361 for (unsigned band_index = 0; band_index < NUM_EQ_BANDS; ++band_index) {
362 settings.eq_level_db[band_index] = eq_level_db[bus_index][band_index];
364 settings.gain_staging_db = gain_staging_db[bus_index];
365 settings.level_compressor_enabled = level_compressor_enabled[bus_index];
366 settings.compressor_threshold_dbfs = compressor_threshold_dbfs[bus_index];
367 settings.compressor_enabled = compressor_enabled[bus_index];
371 void AudioMixer::set_bus_settings(unsigned bus_index, const AudioMixer::BusSettings &settings)
373 lock_guard<timed_mutex> lock(audio_mutex);
374 fader_volume_db[bus_index] = settings.fader_volume_db;
375 mute[bus_index] = settings.muted;
376 locut_enabled[bus_index] = settings.locut_enabled;
377 stereo_width[bus_index] = settings.stereo_width;
378 for (unsigned band_index = 0; band_index < NUM_EQ_BANDS; ++band_index) {
379 eq_level_db[bus_index][band_index] = settings.eq_level_db[band_index];
381 gain_staging_db[bus_index] = settings.gain_staging_db;
382 last_gain_staging_db[bus_index] = gain_staging_db[bus_index];
383 level_compressor_enabled[bus_index] = settings.level_compressor_enabled;
384 compressor_threshold_dbfs[bus_index] = settings.compressor_threshold_dbfs;
385 compressor_enabled[bus_index] = settings.compressor_enabled;
388 AudioMixer::AudioDevice *AudioMixer::find_audio_device(DeviceSpec device)
390 switch (device.type) {
391 case InputSourceType::CAPTURE_CARD:
392 return &video_cards[device.index];
393 case InputSourceType::ALSA_INPUT:
394 return &alsa_inputs[device.index];
395 case InputSourceType::FFMPEG_VIDEO_INPUT:
396 return &ffmpeg_inputs[device.index];
397 case InputSourceType::SILENCE:
404 // Get a pointer to the given channel from the given device.
405 // The channel must be picked out earlier and resampled.
406 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)
408 static float zero = 0.0f;
409 if (source_channel == -1 || device_spec.type == InputSourceType::SILENCE) {
414 AudioDevice *device = find_audio_device(device_spec);
415 assert(device->interesting_channels.count(source_channel) != 0);
416 unsigned channel_index = 0;
417 for (int channel : device->interesting_channels) {
418 if (channel == source_channel) break;
421 assert(channel_index < device->interesting_channels.size());
422 const auto it = samples_card.find(device_spec);
423 assert(it != samples_card.end());
424 *srcptr = &(it->second)[channel_index];
425 *stride = device->interesting_channels.size();
428 // TODO: Can be SSSE3-optimized if need be.
429 void AudioMixer::fill_audio_bus(const map<DeviceSpec, vector<float>> &samples_card, const InputMapping::Bus &bus, unsigned num_samples, float stereo_width, float *output)
431 if (bus.device.type == InputSourceType::SILENCE) {
432 memset(output, 0, num_samples * 2 * sizeof(*output));
434 assert(bus.device.type == InputSourceType::CAPTURE_CARD ||
435 bus.device.type == InputSourceType::ALSA_INPUT ||
436 bus.device.type == InputSourceType::FFMPEG_VIDEO_INPUT);
437 const float *lsrc, *rsrc;
438 unsigned lstride, rstride;
439 float *dptr = output;
440 find_sample_src_from_device(samples_card, bus.device, bus.source_channel[0], &lsrc, &lstride);
441 find_sample_src_from_device(samples_card, bus.device, bus.source_channel[1], &rsrc, &rstride);
443 // Apply stereo width settings. Set stereo width w to a 0..1 range instead of
444 // -1..1, since it makes for much easier calculations (so 0.5 = completely mono).
445 // Then, what we want is
447 // L' = wL + (1-w)R = R + w(L-R)
448 // R' = wR + (1-w)L = L + w(R-L)
450 // This can be further simplified calculation-wise by defining the weighted
451 // difference signal D = w(R-L), so that:
455 float w = 0.5f * stereo_width + 0.5f;
456 if (fabs(w) < 1e-3) {
459 swap(lstride, rstride);
462 if (fabs(w - 1.0f) < 1e-3) {
463 // No calculations needed for stereo_width = 1.
464 for (unsigned i = 0; i < num_samples; ++i) {
472 for (unsigned i = 0; i < num_samples; ++i) {
473 float left = *lsrc, right = *rsrc;
474 float diff = w * (right - left);
475 *dptr++ = right - diff;
476 *dptr++ = left + diff;
484 vector<DeviceSpec> AudioMixer::get_active_devices() const
486 vector<DeviceSpec> ret;
487 for (unsigned card_index = 0; card_index < MAX_VIDEO_CARDS; ++card_index) {
488 const DeviceSpec device_spec{InputSourceType::CAPTURE_CARD, card_index};
489 if (!find_audio_device(device_spec)->interesting_channels.empty()) {
490 ret.push_back(device_spec);
493 for (unsigned card_index = 0; card_index < MAX_ALSA_CARDS; ++card_index) {
494 const DeviceSpec device_spec{InputSourceType::ALSA_INPUT, card_index};
495 if (!find_audio_device(device_spec)->interesting_channels.empty()) {
496 ret.push_back(device_spec);
499 for (unsigned card_index = 0; card_index < num_ffmpeg_inputs; ++card_index) {
500 const DeviceSpec device_spec{InputSourceType::FFMPEG_VIDEO_INPUT, card_index};
501 if (!find_audio_device(device_spec)->interesting_channels.empty()) {
502 ret.push_back(device_spec);
510 void apply_gain(float db, float last_db, vector<float> *samples)
512 if (fabs(db - last_db) < 1e-3) {
513 // Constant over this frame.
514 const float gain = from_db(db);
515 for (size_t i = 0; i < samples->size(); ++i) {
516 (*samples)[i] *= gain;
519 // We need to do a fade.
520 unsigned num_samples = samples->size() / 2;
521 float gain = from_db(last_db);
522 const float gain_inc = pow(from_db(db - last_db), 1.0 / num_samples);
523 for (size_t i = 0; i < num_samples; ++i) {
524 (*samples)[i * 2 + 0] *= gain;
525 (*samples)[i * 2 + 1] *= gain;
533 vector<float> AudioMixer::get_output(steady_clock::time_point ts, unsigned num_samples, ResamplingQueue::RateAdjustmentPolicy rate_adjustment_policy)
535 map<DeviceSpec, vector<float>> samples_card;
536 vector<float> samples_bus;
538 lock_guard<timed_mutex> lock(audio_mutex);
540 // Pick out all the interesting channels from all the cards.
541 for (const DeviceSpec &device_spec : get_active_devices()) {
542 AudioDevice *device = find_audio_device(device_spec);
543 samples_card[device_spec].resize(num_samples * device->interesting_channels.size());
544 if (device->silenced) {
545 memset(&samples_card[device_spec][0], 0, samples_card[device_spec].size() * sizeof(float));
547 device->resampling_queue->get_output_samples(
549 &samples_card[device_spec][0],
551 rate_adjustment_policy);
555 vector<float> samples_out, left, right;
556 samples_out.resize(num_samples * 2);
557 samples_bus.resize(num_samples * 2);
558 for (unsigned bus_index = 0; bus_index < input_mapping.buses.size(); ++bus_index) {
559 fill_audio_bus(samples_card, input_mapping.buses[bus_index], num_samples, stereo_width[bus_index], &samples_bus[0]);
560 apply_eq(bus_index, &samples_bus);
563 lock_guard<mutex> lock(compressor_mutex);
565 // Apply a level compressor to get the general level right.
566 // Basically, if it's over about -40 dBFS, we squeeze it down to that level
567 // (or more precisely, near it, since we don't use infinite ratio),
568 // then apply a makeup gain to get it to -14 dBFS. -14 dBFS is, of course,
569 // entirely arbitrary, but from practical tests with speech, it seems to
570 // put ut around -23 LUFS, so it's a reasonable starting point for later use.
571 if (level_compressor_enabled[bus_index]) {
572 float threshold = 0.01f; // -40 dBFS.
574 float attack_time = 0.5f;
575 float release_time = 20.0f;
576 float makeup_gain = from_db(ref_level_dbfs - (-40.0f)); // +26 dB.
577 level_compressor[bus_index]->process(samples_bus.data(), samples_bus.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
578 gain_staging_db[bus_index] = to_db(level_compressor[bus_index]->get_attenuation() * makeup_gain);
580 // Just apply the gain we already had.
581 float db = gain_staging_db[bus_index];
582 float last_db = last_gain_staging_db[bus_index];
583 apply_gain(db, last_db, &samples_bus);
585 last_gain_staging_db[bus_index] = gain_staging_db[bus_index];
588 printf("level=%f (%+5.2f dBFS) attenuation=%f (%+5.2f dB) end_result=%+5.2f dB\n",
589 level_compressor.get_level(), to_db(level_compressor.get_level()),
590 level_compressor.get_attenuation(), to_db(level_compressor.get_attenuation()),
591 to_db(level_compressor.get_level() * level_compressor.get_attenuation() * makeup_gain));
594 // The real compressor.
595 if (compressor_enabled[bus_index]) {
596 float threshold = from_db(compressor_threshold_dbfs[bus_index]);
598 float attack_time = 0.005f;
599 float release_time = 0.040f;
600 float makeup_gain = 2.0f; // +6 dB.
601 compressor[bus_index]->process(samples_bus.data(), samples_bus.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
602 // compressor_att = compressor.get_attenuation();
606 add_bus_to_master(bus_index, samples_bus, &samples_out);
607 deinterleave_samples(samples_bus, &left, &right);
608 measure_bus_levels(bus_index, left, right);
612 lock_guard<mutex> lock(compressor_mutex);
614 // Finally a limiter at -4 dB (so, -10 dBFS) to take out the worst peaks only.
615 // Note that since ratio is not infinite, we could go slightly higher than this.
616 if (limiter_enabled) {
617 float threshold = from_db(limiter_threshold_dbfs);
619 float attack_time = 0.0f; // Instant.
620 float release_time = 0.020f;
621 float makeup_gain = 1.0f; // 0 dB.
622 limiter.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
623 // limiter_att = limiter.get_attenuation();
626 // printf("limiter=%+5.1f compressor=%+5.1f\n", to_db(limiter_att), to_db(compressor_att));
629 // At this point, we are most likely close to +0 LU (at least if the
630 // faders sum to 0 dB and the compressors are on), but all of our
631 // measurements have been on raw sample values, not R128 values.
632 // So we have a final makeup gain to get us to +0 LU; the gain
633 // adjustments required should be relatively small, and also, the
634 // offset shouldn't change much (only if the type of audio changes
635 // significantly). Thus, we shoot for updating this value basically
636 // “whenever we process buffers”, since the R128 calculation isn't exactly
637 // something we get out per-sample.
639 // Note that there's a feedback loop here, so we choose a very slow filter
640 // (half-time of 30 seconds).
641 double target_loudness_factor, alpha;
642 double loudness_lu = r128.loudness_M() - ref_level_lufs;
643 target_loudness_factor = final_makeup_gain * from_db(-loudness_lu);
645 // If we're outside +/- 5 LU (after correction), we don't count it as
646 // a normal signal (probably silence) and don't change the
647 // correction factor; just apply what we already have.
648 if (fabs(loudness_lu) >= 5.0 || !final_makeup_gain_auto) {
651 // Formula adapted from
652 // https://en.wikipedia.org/wiki/Low-pass_filter#Simple_infinite_impulse_response_filter.
653 const double half_time_s = 30.0;
654 const double fc_mul_2pi_delta_t = 1.0 / (half_time_s * OUTPUT_FREQUENCY);
655 alpha = fc_mul_2pi_delta_t / (fc_mul_2pi_delta_t + 1.0);
659 lock_guard<mutex> lock(compressor_mutex);
660 double m = final_makeup_gain;
661 for (size_t i = 0; i < samples_out.size(); i += 2) {
662 samples_out[i + 0] *= m;
663 samples_out[i + 1] *= m;
664 m += (target_loudness_factor - m) * alpha;
666 final_makeup_gain = m;
669 update_meters(samples_out);
676 void apply_filter_fade(StereoFilter *filter, float *data, unsigned num_samples, float cutoff_hz, float db, float last_db)
678 // A granularity of 32 samples is an okay tradeoff between speed and
679 // smoothness; recalculating the filters is pretty expensive, so it's
680 // good that we don't do this all the time.
681 static constexpr unsigned filter_granularity_samples = 32;
683 const float cutoff_linear = cutoff_hz * 2.0 * M_PI / OUTPUT_FREQUENCY;
684 if (fabs(db - last_db) < 1e-3) {
685 // Constant over this frame.
686 if (fabs(db) > 0.01f) {
687 filter->render(data, num_samples, cutoff_linear, 0.5f, db / 40.0f);
690 // We need to do a fade. (Rounding up avoids division by zero.)
691 unsigned num_blocks = (num_samples + filter_granularity_samples - 1) / filter_granularity_samples;
692 const float inc_db_norm = (db - last_db) / 40.0f / num_blocks;
693 float db_norm = db / 40.0f;
694 for (size_t i = 0; i < num_samples; i += filter_granularity_samples) {
695 size_t samples_this_block = std::min<size_t>(num_samples - i, filter_granularity_samples);
696 filter->render(data + i * 2, samples_this_block, cutoff_linear, 0.5f, db_norm);
697 db_norm += inc_db_norm;
704 void AudioMixer::apply_eq(unsigned bus_index, vector<float> *samples_bus)
706 constexpr float bass_freq_hz = 200.0f;
707 constexpr float treble_freq_hz = 4700.0f;
709 // Cut away everything under 120 Hz (or whatever the cutoff is);
710 // we don't need it for voice, and it will reduce headroom
711 // and confuse the compressor. (In particular, any hums at 50 or 60 Hz
712 // should be dampened.)
713 if (locut_enabled[bus_index]) {
714 locut[bus_index].render(samples_bus->data(), samples_bus->size() / 2, locut_cutoff_hz * 2.0 * M_PI / OUTPUT_FREQUENCY, 0.5f);
717 // Apply the rest of the EQ. Since we only have a simple three-band EQ,
718 // we can implement it with two shelf filters. We use a simple gain to
719 // set the mid-level filter, and then offset the low and high bands
720 // from that if we need to. (We could perhaps have folded the gain into
721 // the next part, but it's so cheap that the trouble isn't worth it.)
723 // If any part of the EQ has changed appreciably since last frame,
724 // we fade smoothly during the course of this frame.
725 const float bass_db = eq_level_db[bus_index][EQ_BAND_BASS];
726 const float mid_db = eq_level_db[bus_index][EQ_BAND_MID];
727 const float treble_db = eq_level_db[bus_index][EQ_BAND_TREBLE];
729 const float last_bass_db = last_eq_level_db[bus_index][EQ_BAND_BASS];
730 const float last_mid_db = last_eq_level_db[bus_index][EQ_BAND_MID];
731 const float last_treble_db = last_eq_level_db[bus_index][EQ_BAND_TREBLE];
733 assert(samples_bus->size() % 2 == 0);
734 const unsigned num_samples = samples_bus->size() / 2;
736 apply_gain(mid_db, last_mid_db, samples_bus);
738 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);
739 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);
741 last_eq_level_db[bus_index][EQ_BAND_BASS] = bass_db;
742 last_eq_level_db[bus_index][EQ_BAND_MID] = mid_db;
743 last_eq_level_db[bus_index][EQ_BAND_TREBLE] = treble_db;
746 void AudioMixer::add_bus_to_master(unsigned bus_index, const vector<float> &samples_bus, vector<float> *samples_out)
748 assert(samples_bus.size() == samples_out->size());
749 assert(samples_bus.size() % 2 == 0);
750 unsigned num_samples = samples_bus.size() / 2;
751 const float new_volume_db = mute[bus_index] ? -90.0f : fader_volume_db[bus_index].load();
752 if (fabs(new_volume_db - last_fader_volume_db[bus_index]) > 1e-3) {
753 // The volume has changed; do a fade over the course of this frame.
754 // (We might have some numerical issues here, but it seems to sound OK.)
755 // For the purpose of fading here, the silence floor is set to -90 dB
756 // (the fader only goes to -84).
757 float old_volume = from_db(max<float>(last_fader_volume_db[bus_index], -90.0f));
758 float volume = from_db(max<float>(new_volume_db, -90.0f));
760 float volume_inc = pow(volume / old_volume, 1.0 / num_samples);
762 if (bus_index == 0) {
763 for (unsigned i = 0; i < num_samples; ++i) {
764 (*samples_out)[i * 2 + 0] = samples_bus[i * 2 + 0] * volume;
765 (*samples_out)[i * 2 + 1] = samples_bus[i * 2 + 1] * volume;
766 volume *= volume_inc;
769 for (unsigned i = 0; i < num_samples; ++i) {
770 (*samples_out)[i * 2 + 0] += samples_bus[i * 2 + 0] * volume;
771 (*samples_out)[i * 2 + 1] += samples_bus[i * 2 + 1] * volume;
772 volume *= volume_inc;
775 } else if (new_volume_db > -90.0f) {
776 float volume = from_db(new_volume_db);
777 if (bus_index == 0) {
778 for (unsigned i = 0; i < num_samples; ++i) {
779 (*samples_out)[i * 2 + 0] = samples_bus[i * 2 + 0] * volume;
780 (*samples_out)[i * 2 + 1] = samples_bus[i * 2 + 1] * volume;
783 for (unsigned i = 0; i < num_samples; ++i) {
784 (*samples_out)[i * 2 + 0] += samples_bus[i * 2 + 0] * volume;
785 (*samples_out)[i * 2 + 1] += samples_bus[i * 2 + 1] * volume;
790 last_fader_volume_db[bus_index] = new_volume_db;
793 void AudioMixer::measure_bus_levels(unsigned bus_index, const vector<float> &left, const vector<float> &right)
795 assert(left.size() == right.size());
796 const float volume = mute[bus_index] ? 0.0f : from_db(fader_volume_db[bus_index]);
797 const float peak_levels[2] = {
798 find_peak(left.data(), left.size()) * volume,
799 find_peak(right.data(), right.size()) * volume
801 for (unsigned channel = 0; channel < 2; ++channel) {
802 // Compute the current value, including hold and falloff.
803 // The constants are borrowed from zita-mu1 by Fons Adriaensen.
804 static constexpr float hold_sec = 0.5f;
805 static constexpr float falloff_db_sec = 15.0f; // dB/sec falloff after hold.
807 PeakHistory &history = peak_history[bus_index][channel];
808 history.historic_peak = max(history.historic_peak, peak_levels[channel]);
809 if (history.age_seconds < hold_sec) {
810 current_peak = history.last_peak;
812 current_peak = history.last_peak * from_db(-falloff_db_sec * (history.age_seconds - hold_sec));
815 // See if we have a new peak to replace the old (possibly falling) one.
816 if (peak_levels[channel] > current_peak) {
817 history.last_peak = peak_levels[channel];
818 history.age_seconds = 0.0f; // Not 100% correct, but more than good enough given our frame sizes.
819 current_peak = peak_levels[channel];
821 history.age_seconds += float(left.size()) / OUTPUT_FREQUENCY;
823 history.current_level = peak_levels[channel];
824 history.current_peak = current_peak;
828 void AudioMixer::update_meters(const vector<float> &samples)
830 // Upsample 4x to find interpolated peak.
831 peak_resampler.inp_data = const_cast<float *>(samples.data());
832 peak_resampler.inp_count = samples.size() / 2;
834 vector<float> interpolated_samples;
835 interpolated_samples.resize(samples.size());
837 lock_guard<mutex> lock(audio_measure_mutex);
839 while (peak_resampler.inp_count > 0) { // About four iterations.
840 peak_resampler.out_data = &interpolated_samples[0];
841 peak_resampler.out_count = interpolated_samples.size() / 2;
842 peak_resampler.process();
843 size_t out_stereo_samples = interpolated_samples.size() / 2 - peak_resampler.out_count;
844 peak = max<float>(peak, find_peak(interpolated_samples.data(), out_stereo_samples * 2));
845 peak_resampler.out_data = nullptr;
849 // Find R128 levels and L/R correlation.
850 vector<float> left, right;
851 deinterleave_samples(samples, &left, &right);
852 float *ptrs[] = { left.data(), right.data() };
854 lock_guard<mutex> lock(audio_measure_mutex);
855 r128.process(left.size(), ptrs);
856 correlation.process_samples(samples);
859 send_audio_level_callback();
862 void AudioMixer::reset_meters()
864 lock_guard<mutex> lock(audio_measure_mutex);
865 peak_resampler.reset();
872 void AudioMixer::send_audio_level_callback()
874 if (audio_level_callback == nullptr) {
878 lock_guard<mutex> lock(audio_measure_mutex);
879 double loudness_s = r128.loudness_S();
880 double loudness_i = r128.integrated();
881 double loudness_range_low = r128.range_min();
882 double loudness_range_high = r128.range_max();
884 metric_audio_loudness_short_lufs = loudness_s;
885 metric_audio_loudness_integrated_lufs = loudness_i;
886 metric_audio_loudness_range_low_lufs = loudness_range_low;
887 metric_audio_loudness_range_high_lufs = loudness_range_high;
888 metric_audio_peak_dbfs = to_db(peak);
889 metric_audio_final_makeup_gain_db = to_db(final_makeup_gain);
890 metric_audio_correlation = correlation.get_correlation();
892 vector<BusLevel> bus_levels;
893 bus_levels.resize(input_mapping.buses.size());
895 lock_guard<mutex> lock(compressor_mutex);
896 for (unsigned bus_index = 0; bus_index < bus_levels.size(); ++bus_index) {
897 BusLevel &levels = bus_levels[bus_index];
898 BusMetrics &metrics = bus_metrics[bus_index];
900 levels.current_level_dbfs[0] = metrics.current_level_dbfs[0] = to_db(peak_history[bus_index][0].current_level);
901 levels.current_level_dbfs[1] = metrics.current_level_dbfs[1] = to_db(peak_history[bus_index][1].current_level);
902 levels.peak_level_dbfs[0] = metrics.peak_level_dbfs[0] = to_db(peak_history[bus_index][0].current_peak);
903 levels.peak_level_dbfs[1] = metrics.peak_level_dbfs[1] = to_db(peak_history[bus_index][1].current_peak);
904 levels.historic_peak_dbfs = metrics.historic_peak_dbfs = to_db(
905 max(peak_history[bus_index][0].historic_peak,
906 peak_history[bus_index][1].historic_peak));
907 levels.gain_staging_db = metrics.gain_staging_db = gain_staging_db[bus_index];
908 if (compressor_enabled[bus_index]) {
909 levels.compressor_attenuation_db = metrics.compressor_attenuation_db = -to_db(compressor[bus_index]->get_attenuation());
911 levels.compressor_attenuation_db = 0.0;
912 metrics.compressor_attenuation_db = 0.0 / 0.0;
917 audio_level_callback(loudness_s, to_db(peak), bus_levels,
918 loudness_i, loudness_range_low, loudness_range_high,
919 to_db(final_makeup_gain),
920 correlation.get_correlation());
923 map<DeviceSpec, DeviceInfo> AudioMixer::get_devices()
925 lock_guard<timed_mutex> lock(audio_mutex);
927 map<DeviceSpec, DeviceInfo> devices;
928 for (unsigned card_index = 0; card_index < num_capture_cards; ++card_index) {
929 const DeviceSpec spec{ InputSourceType::CAPTURE_CARD, card_index };
930 const AudioDevice *device = &video_cards[card_index];
932 info.display_name = device->display_name;
933 info.num_channels = 8;
934 devices.insert(make_pair(spec, info));
936 vector<ALSAPool::Device> available_alsa_devices = alsa_pool.get_devices();
937 for (unsigned card_index = 0; card_index < available_alsa_devices.size(); ++card_index) {
938 const DeviceSpec spec{ InputSourceType::ALSA_INPUT, card_index };
939 const ALSAPool::Device &device = available_alsa_devices[card_index];
941 info.display_name = device.display_name();
942 info.num_channels = device.num_channels;
943 info.alsa_name = device.name;
944 info.alsa_info = device.info;
945 info.alsa_address = device.address;
946 devices.insert(make_pair(spec, info));
948 for (unsigned card_index = 0; card_index < num_ffmpeg_inputs; ++card_index) {
949 const DeviceSpec spec{ InputSourceType::FFMPEG_VIDEO_INPUT, card_index };
950 const AudioDevice *device = &ffmpeg_inputs[card_index];
952 info.display_name = device->display_name;
953 info.num_channels = 2;
954 devices.insert(make_pair(spec, info));
959 void AudioMixer::set_display_name(DeviceSpec device_spec, const string &name)
961 AudioDevice *device = find_audio_device(device_spec);
963 lock_guard<timed_mutex> lock(audio_mutex);
964 device->display_name = name;
967 void AudioMixer::serialize_device(DeviceSpec device_spec, DeviceSpecProto *device_spec_proto)
969 lock_guard<timed_mutex> lock(audio_mutex);
970 switch (device_spec.type) {
971 case InputSourceType::SILENCE:
972 device_spec_proto->set_type(DeviceSpecProto::SILENCE);
974 case InputSourceType::CAPTURE_CARD:
975 device_spec_proto->set_type(DeviceSpecProto::CAPTURE_CARD);
976 device_spec_proto->set_index(device_spec.index);
977 device_spec_proto->set_display_name(video_cards[device_spec.index].display_name);
979 case InputSourceType::ALSA_INPUT:
980 alsa_pool.serialize_device(device_spec.index, device_spec_proto);
982 case InputSourceType::FFMPEG_VIDEO_INPUT:
983 device_spec_proto->set_type(DeviceSpecProto::FFMPEG_VIDEO_INPUT);
984 device_spec_proto->set_index(device_spec.index);
985 device_spec_proto->set_display_name(ffmpeg_inputs[device_spec.index].display_name);
990 void AudioMixer::set_simple_input(unsigned card_index)
992 assert(card_index < num_capture_cards + num_ffmpeg_inputs);
993 InputMapping new_input_mapping;
994 InputMapping::Bus input;
996 if (card_index >= num_capture_cards) {
997 input.device = DeviceSpec{InputSourceType::FFMPEG_VIDEO_INPUT, card_index - num_capture_cards};
999 input.device = DeviceSpec{InputSourceType::CAPTURE_CARD, card_index};
1001 input.source_channel[0] = 0;
1002 input.source_channel[1] = 1;
1004 new_input_mapping.buses.push_back(input);
1006 lock_guard<timed_mutex> lock(audio_mutex);
1007 current_mapping_mode = MappingMode::SIMPLE;
1008 set_input_mapping_lock_held(new_input_mapping);
1009 fader_volume_db[0] = 0.0f;
1012 unsigned AudioMixer::get_simple_input() const
1014 lock_guard<timed_mutex> lock(audio_mutex);
1015 if (input_mapping.buses.size() == 1 &&
1016 input_mapping.buses[0].device.type == InputSourceType::CAPTURE_CARD &&
1017 input_mapping.buses[0].source_channel[0] == 0 &&
1018 input_mapping.buses[0].source_channel[1] == 1) {
1019 return input_mapping.buses[0].device.index;
1020 } else if (input_mapping.buses.size() == 1 &&
1021 input_mapping.buses[0].device.type == InputSourceType::FFMPEG_VIDEO_INPUT &&
1022 input_mapping.buses[0].source_channel[0] == 0 &&
1023 input_mapping.buses[0].source_channel[1] == 1) {
1024 return input_mapping.buses[0].device.index + num_capture_cards;
1026 return numeric_limits<unsigned>::max();
1030 void AudioMixer::set_input_mapping(const InputMapping &new_input_mapping)
1032 lock_guard<timed_mutex> lock(audio_mutex);
1033 set_input_mapping_lock_held(new_input_mapping);
1034 current_mapping_mode = MappingMode::MULTICHANNEL;
1037 AudioMixer::MappingMode AudioMixer::get_mapping_mode() const
1039 lock_guard<timed_mutex> lock(audio_mutex);
1040 return current_mapping_mode;
1043 void AudioMixer::set_input_mapping_lock_held(const InputMapping &new_input_mapping)
1045 map<DeviceSpec, set<unsigned>> interesting_channels;
1046 for (const InputMapping::Bus &bus : new_input_mapping.buses) {
1047 if (bus.device.type == InputSourceType::CAPTURE_CARD ||
1048 bus.device.type == InputSourceType::ALSA_INPUT ||
1049 bus.device.type == InputSourceType::FFMPEG_VIDEO_INPUT) {
1050 for (unsigned channel = 0; channel < 2; ++channel) {
1051 if (bus.source_channel[channel] != -1) {
1052 interesting_channels[bus.device].insert(bus.source_channel[channel]);
1056 assert(bus.device.type == InputSourceType::SILENCE);
1060 // Kill all the old metrics, and set up new ones.
1061 for (unsigned bus_index = 0; bus_index < input_mapping.buses.size(); ++bus_index) {
1062 BusMetrics &metrics = bus_metrics[bus_index];
1064 vector<pair<string, string>> labels_left = metrics.labels;
1065 labels_left.emplace_back("channel", "left");
1066 vector<pair<string, string>> labels_right = metrics.labels;
1067 labels_right.emplace_back("channel", "right");
1069 global_metrics.remove("bus_current_level_dbfs", labels_left);
1070 global_metrics.remove("bus_current_level_dbfs", labels_right);
1071 global_metrics.remove("bus_peak_level_dbfs", labels_left);
1072 global_metrics.remove("bus_peak_level_dbfs", labels_right);
1073 global_metrics.remove("bus_historic_peak_dbfs", metrics.labels);
1074 global_metrics.remove("bus_gain_staging_db", metrics.labels);
1075 global_metrics.remove("bus_compressor_attenuation_db", metrics.labels);
1077 bus_metrics.reset(new BusMetrics[new_input_mapping.buses.size()]);
1078 for (unsigned bus_index = 0; bus_index < new_input_mapping.buses.size(); ++bus_index) {
1079 const InputMapping::Bus &bus = new_input_mapping.buses[bus_index];
1080 BusMetrics &metrics = bus_metrics[bus_index];
1082 char bus_index_str[16], source_index_str[16], source_channels_str[64];
1083 snprintf(bus_index_str, sizeof(bus_index_str), "%u", bus_index);
1084 snprintf(source_index_str, sizeof(source_index_str), "%u", bus.device.index);
1085 snprintf(source_channels_str, sizeof(source_channels_str), "%d:%d", bus.source_channel[0], bus.source_channel[1]);
1087 vector<pair<string, string>> labels;
1088 metrics.labels.emplace_back("index", bus_index_str);
1089 metrics.labels.emplace_back("name", bus.name);
1090 if (bus.device.type == InputSourceType::SILENCE) {
1091 metrics.labels.emplace_back("source_type", "silence");
1092 } else if (bus.device.type == InputSourceType::CAPTURE_CARD) {
1093 metrics.labels.emplace_back("source_type", "capture_card");
1094 } else if (bus.device.type == InputSourceType::ALSA_INPUT) {
1095 metrics.labels.emplace_back("source_type", "alsa_input");
1096 } else if (bus.device.type == InputSourceType::FFMPEG_VIDEO_INPUT) {
1097 metrics.labels.emplace_back("source_type", "ffmpeg_video_input");
1101 metrics.labels.emplace_back("source_index", source_index_str);
1102 metrics.labels.emplace_back("source_channels", source_channels_str);
1104 vector<pair<string, string>> labels_left = metrics.labels;
1105 labels_left.emplace_back("channel", "left");
1106 vector<pair<string, string>> labels_right = metrics.labels;
1107 labels_right.emplace_back("channel", "right");
1109 global_metrics.add("bus_current_level_dbfs", labels_left, &metrics.current_level_dbfs[0], Metrics::TYPE_GAUGE);
1110 global_metrics.add("bus_current_level_dbfs", labels_right, &metrics.current_level_dbfs[1], Metrics::TYPE_GAUGE);
1111 global_metrics.add("bus_peak_level_dbfs", labels_left, &metrics.peak_level_dbfs[0], Metrics::TYPE_GAUGE);
1112 global_metrics.add("bus_peak_level_dbfs", labels_right, &metrics.peak_level_dbfs[1], Metrics::TYPE_GAUGE);
1113 global_metrics.add("bus_historic_peak_dbfs", metrics.labels, &metrics.historic_peak_dbfs, Metrics::TYPE_GAUGE);
1114 global_metrics.add("bus_gain_staging_db", metrics.labels, &metrics.gain_staging_db, Metrics::TYPE_GAUGE);
1115 global_metrics.add("bus_compressor_attenuation_db", metrics.labels, &metrics.compressor_attenuation_db, Metrics::TYPE_GAUGE);
1118 // Reset resamplers for all cards that don't have the exact same state as before.
1119 for (unsigned card_index = 0; card_index < MAX_VIDEO_CARDS; ++card_index) {
1120 const DeviceSpec device_spec{InputSourceType::CAPTURE_CARD, card_index};
1121 AudioDevice *device = find_audio_device(device_spec);
1122 if (device->interesting_channels != interesting_channels[device_spec]) {
1123 device->interesting_channels = interesting_channels[device_spec];
1124 reset_resampler_mutex_held(device_spec);
1127 for (unsigned card_index = 0; card_index < MAX_ALSA_CARDS; ++card_index) {
1128 const DeviceSpec device_spec{InputSourceType::ALSA_INPUT, card_index};
1129 AudioDevice *device = find_audio_device(device_spec);
1130 if (interesting_channels[device_spec].empty()) {
1131 alsa_pool.release_device(card_index);
1133 alsa_pool.hold_device(card_index);
1135 if (device->interesting_channels != interesting_channels[device_spec]) {
1136 device->interesting_channels = interesting_channels[device_spec];
1137 alsa_pool.reset_device(device_spec.index);
1138 reset_resampler_mutex_held(device_spec);
1141 for (unsigned card_index = 0; card_index < num_ffmpeg_inputs; ++card_index) {
1142 const DeviceSpec device_spec{InputSourceType::FFMPEG_VIDEO_INPUT, card_index};
1143 AudioDevice *device = find_audio_device(device_spec);
1144 if (device->interesting_channels != interesting_channels[device_spec]) {
1145 device->interesting_channels = interesting_channels[device_spec];
1146 reset_resampler_mutex_held(device_spec);
1150 input_mapping = new_input_mapping;
1153 InputMapping AudioMixer::get_input_mapping() const
1155 lock_guard<timed_mutex> lock(audio_mutex);
1156 return input_mapping;
1159 unsigned AudioMixer::num_buses() const
1161 lock_guard<timed_mutex> lock(audio_mutex);
1162 return input_mapping.buses.size();
1165 void AudioMixer::reset_peak(unsigned bus_index)
1167 lock_guard<timed_mutex> lock(audio_mutex);
1168 for (unsigned channel = 0; channel < 2; ++channel) {
1169 PeakHistory &history = peak_history[bus_index][channel];
1170 history.current_level = 0.0f;
1171 history.historic_peak = 0.0f;
1172 history.current_peak = 0.0f;
1173 history.last_peak = 0.0f;
1174 history.age_seconds = 0.0f;
1178 AudioMixer *global_audio_mixer = nullptr;