1 #include "audio_mixer.h"
5 #include <bmusb/bmusb.h>
14 using namespace bmusb;
16 using namespace std::placeholders;
20 // TODO: If these prove to be a bottleneck, they can be SSSE3-optimized
21 // (usually including multiple channels at a time).
23 void convert_fixed16_to_fp32(float *dst, size_t out_channel, size_t out_num_channels,
24 const uint8_t *src, size_t in_channel, size_t in_num_channels,
27 assert(in_channel < in_num_channels);
28 assert(out_channel < out_num_channels);
29 src += in_channel * 2;
32 for (size_t i = 0; i < num_samples; ++i) {
33 int16_t s = le16toh(*(int16_t *)src);
34 *dst = s * (1.0f / 32768.0f);
36 src += 2 * in_num_channels;
37 dst += out_num_channels;
41 void convert_fixed24_to_fp32(float *dst, size_t out_channel, size_t out_num_channels,
42 const uint8_t *src, size_t in_channel, size_t in_num_channels,
45 assert(in_channel < in_num_channels);
46 assert(out_channel < out_num_channels);
47 src += in_channel * 3;
50 for (size_t i = 0; i < num_samples; ++i) {
54 uint32_t s = s1 | (s1 << 8) | (s2 << 16) | (s3 << 24);
55 *dst = int(s) * (1.0f / 2147483648.0f);
57 src += 3 * in_num_channels;
58 dst += out_num_channels;
62 void convert_fixed32_to_fp32(float *dst, size_t out_channel, size_t out_num_channels,
63 const uint8_t *src, size_t in_channel, size_t in_num_channels,
66 assert(in_channel < in_num_channels);
67 assert(out_channel < out_num_channels);
68 src += in_channel * 4;
71 for (size_t i = 0; i < num_samples; ++i) {
72 int32_t s = le32toh(*(int32_t *)src);
73 *dst = s * (1.0f / 2147483648.0f);
75 src += 4 * in_num_channels;
76 dst += out_num_channels;
80 float find_peak(const float *samples, size_t num_samples)
82 float m = fabs(samples[0]);
83 for (size_t i = 1; i < num_samples; ++i) {
84 m = max(m, fabs(samples[i]));
89 void deinterleave_samples(const vector<float> &in, vector<float> *out_l, vector<float> *out_r)
91 size_t num_samples = in.size() / 2;
92 out_l->resize(num_samples);
93 out_r->resize(num_samples);
95 const float *inptr = in.data();
96 float *lptr = &(*out_l)[0];
97 float *rptr = &(*out_r)[0];
98 for (size_t i = 0; i < num_samples; ++i) {
106 AudioMixer::AudioMixer(unsigned num_cards)
107 : num_cards(num_cards),
108 limiter(OUTPUT_FREQUENCY),
109 correlation(OUTPUT_FREQUENCY)
111 for (unsigned bus_index = 0; bus_index < MAX_BUSES; ++bus_index) {
112 locut[bus_index].init(FILTER_HPF, 2);
113 locut_enabled[bus_index] = global_flags.locut_enabled;
114 gain_staging_db[bus_index] = global_flags.initial_gain_staging_db;
115 compressor[bus_index].reset(new StereoCompressor(OUTPUT_FREQUENCY));
116 compressor_threshold_dbfs[bus_index] = ref_level_dbfs - 12.0f; // -12 dB.
117 compressor_enabled[bus_index] = global_flags.compressor_enabled;
118 level_compressor[bus_index].reset(new StereoCompressor(OUTPUT_FREQUENCY));
119 level_compressor_enabled[bus_index] = global_flags.gain_staging_auto;
121 set_limiter_enabled(global_flags.limiter_enabled);
122 set_final_makeup_gain_auto(global_flags.final_makeup_gain_auto);
124 // Generate a very simple, default input mapping.
125 InputMapping::Bus input;
127 input.device.type = InputSourceType::CAPTURE_CARD;
128 input.device.index = 0;
129 input.source_channel[0] = 0;
130 input.source_channel[1] = 1;
132 InputMapping new_input_mapping;
133 new_input_mapping.buses.push_back(input);
134 set_input_mapping(new_input_mapping);
136 // Look for ALSA cards.
137 available_alsa_cards = ALSAInput::enumerate_devices();
139 r128.init(2, OUTPUT_FREQUENCY);
142 // hlen=16 is pretty low quality, but we use quite a bit of CPU otherwise,
143 // and there's a limit to how important the peak meter is.
144 peak_resampler.setup(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY * 4, /*num_channels=*/2, /*hlen=*/16, /*frel=*/1.0);
147 AudioMixer::~AudioMixer()
149 for (unsigned card_index = 0; card_index < available_alsa_cards.size(); ++card_index) {
150 const AudioDevice &device = alsa_inputs[card_index];
151 if (device.alsa_device != nullptr) {
152 device.alsa_device->stop_capture_thread();
158 void AudioMixer::reset_resampler(DeviceSpec device_spec)
160 lock_guard<timed_mutex> lock(audio_mutex);
161 reset_resampler_mutex_held(device_spec);
164 void AudioMixer::reset_resampler_mutex_held(DeviceSpec device_spec)
166 AudioDevice *device = find_audio_device(device_spec);
168 if (device->interesting_channels.empty()) {
169 device->resampling_queue.reset();
171 // TODO: ResamplingQueue should probably take the full device spec.
172 // (It's only used for console output, though.)
173 device->resampling_queue.reset(new ResamplingQueue(device_spec.index, device->capture_frequency, OUTPUT_FREQUENCY, device->interesting_channels.size()));
175 device->next_local_pts = 0;
178 void AudioMixer::reset_alsa_mutex_held(DeviceSpec device_spec)
180 assert(device_spec.type == InputSourceType::ALSA_INPUT);
181 unsigned card_index = device_spec.index;
182 AudioDevice *device = find_audio_device(device_spec);
184 if (device->alsa_device != nullptr) {
185 device->alsa_device->stop_capture_thread();
187 if (device->interesting_channels.empty()) {
188 device->alsa_device.reset();
190 const ALSAInput::Device &alsa_dev = available_alsa_cards[card_index];
191 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)));
192 device->capture_frequency = device->alsa_device->get_sample_rate();
193 device->alsa_device->start_capture_thread();
197 bool AudioMixer::add_audio(DeviceSpec device_spec, const uint8_t *data, unsigned num_samples, AudioFormat audio_format, int64_t frame_length)
199 AudioDevice *device = find_audio_device(device_spec);
201 unique_lock<timed_mutex> lock(audio_mutex, defer_lock);
202 if (!lock.try_lock_for(chrono::milliseconds(10))) {
205 if (device->resampling_queue == nullptr) {
206 // No buses use this device; throw it away.
210 unsigned num_channels = device->interesting_channels.size();
211 assert(num_channels > 0);
213 // Convert the audio to fp32.
215 audio.resize(num_samples * num_channels);
216 unsigned channel_index = 0;
217 for (auto channel_it = device->interesting_channels.cbegin(); channel_it != device->interesting_channels.end(); ++channel_it, ++channel_index) {
218 switch (audio_format.bits_per_sample) {
220 assert(num_samples == 0);
223 convert_fixed16_to_fp32(&audio[0], channel_index, num_channels, data, *channel_it, audio_format.num_channels, num_samples);
226 convert_fixed24_to_fp32(&audio[0], channel_index, num_channels, data, *channel_it, audio_format.num_channels, num_samples);
229 convert_fixed32_to_fp32(&audio[0], channel_index, num_channels, data, *channel_it, audio_format.num_channels, num_samples);
232 fprintf(stderr, "Cannot handle audio with %u bits per sample\n", audio_format.bits_per_sample);
238 int64_t local_pts = device->next_local_pts;
239 device->resampling_queue->add_input_samples(local_pts / double(TIMEBASE), audio.data(), num_samples);
240 device->next_local_pts = local_pts + frame_length;
244 bool AudioMixer::add_silence(DeviceSpec device_spec, unsigned samples_per_frame, unsigned num_frames, int64_t frame_length)
246 AudioDevice *device = find_audio_device(device_spec);
248 unique_lock<timed_mutex> lock(audio_mutex, defer_lock);
249 if (!lock.try_lock_for(chrono::milliseconds(10))) {
252 if (device->resampling_queue == nullptr) {
253 // No buses use this device; throw it away.
257 unsigned num_channels = device->interesting_channels.size();
258 assert(num_channels > 0);
260 vector<float> silence(samples_per_frame * num_channels, 0.0f);
261 for (unsigned i = 0; i < num_frames; ++i) {
262 device->resampling_queue->add_input_samples(device->next_local_pts / double(TIMEBASE), silence.data(), samples_per_frame);
263 // Note that if the format changed in the meantime, we have
264 // no way of detecting that; we just have to assume the frame length
265 // is always the same.
266 device->next_local_pts += frame_length;
271 AudioMixer::AudioDevice *AudioMixer::find_audio_device(DeviceSpec device)
273 switch (device.type) {
274 case InputSourceType::CAPTURE_CARD:
275 return &video_cards[device.index];
276 case InputSourceType::ALSA_INPUT:
277 return &alsa_inputs[device.index];
278 case InputSourceType::SILENCE:
285 // Get a pointer to the given channel from the given device.
286 // The channel must be picked out earlier and resampled.
287 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)
289 static float zero = 0.0f;
290 if (source_channel == -1 || device_spec.type == InputSourceType::SILENCE) {
295 AudioDevice *device = find_audio_device(device_spec);
296 assert(device->interesting_channels.count(source_channel) != 0);
297 unsigned channel_index = 0;
298 for (int channel : device->interesting_channels) {
299 if (channel == source_channel) break;
302 assert(channel_index < device->interesting_channels.size());
303 const auto it = samples_card.find(device_spec);
304 assert(it != samples_card.end());
305 *srcptr = &(it->second)[channel_index];
306 *stride = device->interesting_channels.size();
309 // TODO: Can be SSSE3-optimized if need be.
310 void AudioMixer::fill_audio_bus(const map<DeviceSpec, vector<float>> &samples_card, const InputMapping::Bus &bus, unsigned num_samples, float *output)
312 if (bus.device.type == InputSourceType::SILENCE) {
313 memset(output, 0, num_samples * sizeof(*output));
315 assert(bus.device.type == InputSourceType::CAPTURE_CARD ||
316 bus.device.type == InputSourceType::ALSA_INPUT);
317 const float *lsrc, *rsrc;
318 unsigned lstride, rstride;
319 float *dptr = output;
320 find_sample_src_from_device(samples_card, bus.device, bus.source_channel[0], &lsrc, &lstride);
321 find_sample_src_from_device(samples_card, bus.device, bus.source_channel[1], &rsrc, &rstride);
322 for (unsigned i = 0; i < num_samples; ++i) {
331 vector<float> AudioMixer::get_output(double pts, unsigned num_samples, ResamplingQueue::RateAdjustmentPolicy rate_adjustment_policy)
333 map<DeviceSpec, vector<float>> samples_card;
334 vector<float> samples_bus;
336 lock_guard<timed_mutex> lock(audio_mutex);
338 // Pick out all the interesting channels from all the cards.
339 // TODO: If the card has been hotswapped, the number of channels
340 // might have changed; if so, we need to do some sort of remapping
342 for (const auto &spec_and_info : get_devices_mutex_held()) {
343 const DeviceSpec &device_spec = spec_and_info.first;
344 AudioDevice *device = find_audio_device(device_spec);
345 if (!device->interesting_channels.empty()) {
346 samples_card[device_spec].resize(num_samples * device->interesting_channels.size());
347 device->resampling_queue->get_output_samples(
349 &samples_card[device_spec][0],
351 rate_adjustment_policy);
355 vector<float> samples_out, left, right;
356 samples_out.resize(num_samples * 2);
357 samples_bus.resize(num_samples * 2);
358 for (unsigned bus_index = 0; bus_index < input_mapping.buses.size(); ++bus_index) {
359 fill_audio_bus(samples_card, input_mapping.buses[bus_index], num_samples, &samples_bus[0]);
361 // Cut away everything under 120 Hz (or whatever the cutoff is);
362 // we don't need it for voice, and it will reduce headroom
363 // and confuse the compressor. (In particular, any hums at 50 or 60 Hz
364 // should be dampened.)
365 if (locut_enabled[bus_index]) {
366 locut[bus_index].render(samples_bus.data(), samples_bus.size() / 2, locut_cutoff_hz * 2.0 * M_PI / OUTPUT_FREQUENCY, 0.5f);
370 lock_guard<mutex> lock(compressor_mutex);
372 // Apply a level compressor to get the general level right.
373 // Basically, if it's over about -40 dBFS, we squeeze it down to that level
374 // (or more precisely, near it, since we don't use infinite ratio),
375 // then apply a makeup gain to get it to -14 dBFS. -14 dBFS is, of course,
376 // entirely arbitrary, but from practical tests with speech, it seems to
377 // put ut around -23 LUFS, so it's a reasonable starting point for later use.
378 if (level_compressor_enabled[bus_index]) {
379 float threshold = 0.01f; // -40 dBFS.
381 float attack_time = 0.5f;
382 float release_time = 20.0f;
383 float makeup_gain = from_db(ref_level_dbfs - (-40.0f)); // +26 dB.
384 level_compressor[bus_index]->process(samples_bus.data(), samples_bus.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
385 gain_staging_db[bus_index] = to_db(level_compressor[bus_index]->get_attenuation() * makeup_gain);
387 // Just apply the gain we already had.
388 float g = from_db(gain_staging_db[bus_index]);
389 for (size_t i = 0; i < samples_bus.size(); ++i) {
395 printf("level=%f (%+5.2f dBFS) attenuation=%f (%+5.2f dB) end_result=%+5.2f dB\n",
396 level_compressor.get_level(), to_db(level_compressor.get_level()),
397 level_compressor.get_attenuation(), to_db(level_compressor.get_attenuation()),
398 to_db(level_compressor.get_level() * level_compressor.get_attenuation() * makeup_gain));
401 // The real compressor.
402 if (compressor_enabled[bus_index]) {
403 float threshold = from_db(compressor_threshold_dbfs[bus_index]);
405 float attack_time = 0.005f;
406 float release_time = 0.040f;
407 float makeup_gain = 2.0f; // +6 dB.
408 compressor[bus_index]->process(samples_bus.data(), samples_bus.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
409 // compressor_att = compressor.get_attenuation();
413 // TODO: We should measure post-fader.
414 deinterleave_samples(samples_bus, &left, &right);
415 measure_bus_levels(bus_index, left, right);
417 float volume = from_db(fader_volume_db[bus_index]);
418 if (bus_index == 0) {
419 for (unsigned i = 0; i < num_samples * 2; ++i) {
420 samples_out[i] = samples_bus[i] * volume;
423 for (unsigned i = 0; i < num_samples * 2; ++i) {
424 samples_out[i] += samples_bus[i] * volume;
430 lock_guard<mutex> lock(compressor_mutex);
432 // Finally a limiter at -4 dB (so, -10 dBFS) to take out the worst peaks only.
433 // Note that since ratio is not infinite, we could go slightly higher than this.
434 if (limiter_enabled) {
435 float threshold = from_db(limiter_threshold_dbfs);
437 float attack_time = 0.0f; // Instant.
438 float release_time = 0.020f;
439 float makeup_gain = 1.0f; // 0 dB.
440 limiter.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
441 // limiter_att = limiter.get_attenuation();
444 // printf("limiter=%+5.1f compressor=%+5.1f\n", to_db(limiter_att), to_db(compressor_att));
447 // At this point, we are most likely close to +0 LU (at least if the
448 // faders sum to 0 dB and the compressors are on), but all of our
449 // measurements have been on raw sample values, not R128 values.
450 // So we have a final makeup gain to get us to +0 LU; the gain
451 // adjustments required should be relatively small, and also, the
452 // offset shouldn't change much (only if the type of audio changes
453 // significantly). Thus, we shoot for updating this value basically
454 // “whenever we process buffers”, since the R128 calculation isn't exactly
455 // something we get out per-sample.
457 // Note that there's a feedback loop here, so we choose a very slow filter
458 // (half-time of 30 seconds).
459 double target_loudness_factor, alpha;
460 double loudness_lu = r128.loudness_M() - ref_level_lufs;
461 double current_makeup_lu = to_db(final_makeup_gain);
462 target_loudness_factor = final_makeup_gain * from_db(-loudness_lu);
464 // If we're outside +/- 5 LU uncorrected, we don't count it as
465 // a normal signal (probably silence) and don't change the
466 // correction factor; just apply what we already have.
467 if (fabs(loudness_lu - current_makeup_lu) >= 5.0 || !final_makeup_gain_auto) {
470 // Formula adapted from
471 // https://en.wikipedia.org/wiki/Low-pass_filter#Simple_infinite_impulse_response_filter.
472 const double half_time_s = 30.0;
473 const double fc_mul_2pi_delta_t = 1.0 / (half_time_s * OUTPUT_FREQUENCY);
474 alpha = fc_mul_2pi_delta_t / (fc_mul_2pi_delta_t + 1.0);
478 lock_guard<mutex> lock(compressor_mutex);
479 double m = final_makeup_gain;
480 for (size_t i = 0; i < samples_out.size(); i += 2) {
481 samples_out[i + 0] *= m;
482 samples_out[i + 1] *= m;
483 m += (target_loudness_factor - m) * alpha;
485 final_makeup_gain = m;
488 update_meters(samples_out);
493 void AudioMixer::measure_bus_levels(unsigned bus_index, const vector<float> &left, const vector<float> &right)
495 const float *ptrs[] = { left.data(), right.data() };
497 lock_guard<mutex> lock(audio_measure_mutex);
498 bus_r128[bus_index]->process(left.size(), const_cast<float **>(ptrs));
502 void AudioMixer::update_meters(const vector<float> &samples)
504 // Upsample 4x to find interpolated peak.
505 peak_resampler.inp_data = const_cast<float *>(samples.data());
506 peak_resampler.inp_count = samples.size() / 2;
508 vector<float> interpolated_samples;
509 interpolated_samples.resize(samples.size());
511 lock_guard<mutex> lock(audio_measure_mutex);
513 while (peak_resampler.inp_count > 0) { // About four iterations.
514 peak_resampler.out_data = &interpolated_samples[0];
515 peak_resampler.out_count = interpolated_samples.size() / 2;
516 peak_resampler.process();
517 size_t out_stereo_samples = interpolated_samples.size() / 2 - peak_resampler.out_count;
518 peak = max<float>(peak, find_peak(interpolated_samples.data(), out_stereo_samples * 2));
519 peak_resampler.out_data = nullptr;
523 // Find R128 levels and L/R correlation.
524 vector<float> left, right;
525 deinterleave_samples(samples, &left, &right);
526 float *ptrs[] = { left.data(), right.data() };
528 lock_guard<mutex> lock(audio_measure_mutex);
529 r128.process(left.size(), ptrs);
530 correlation.process_samples(samples);
533 send_audio_level_callback();
536 void AudioMixer::reset_meters()
538 lock_guard<mutex> lock(audio_measure_mutex);
539 peak_resampler.reset();
546 void AudioMixer::send_audio_level_callback()
548 if (audio_level_callback == nullptr) {
552 lock_guard<mutex> lock(audio_measure_mutex);
553 double loudness_s = r128.loudness_S();
554 double loudness_i = r128.integrated();
555 double loudness_range_low = r128.range_min();
556 double loudness_range_high = r128.range_max();
558 vector<BusLevel> bus_levels;
559 bus_levels.resize(input_mapping.buses.size());
560 for (unsigned bus_index = 0; bus_index < bus_r128.size(); ++bus_index) {
561 bus_levels[bus_index].loudness_lufs = bus_r128[bus_index]->loudness_S();
562 bus_levels[bus_index].gain_staging_db = gain_staging_db[bus_index];
565 audio_level_callback(loudness_s, to_db(peak), bus_levels,
566 loudness_i, loudness_range_low, loudness_range_high,
567 to_db(final_makeup_gain),
568 correlation.get_correlation());
571 map<DeviceSpec, DeviceInfo> AudioMixer::get_devices() const
573 lock_guard<timed_mutex> lock(audio_mutex);
574 return get_devices_mutex_held();
577 map<DeviceSpec, DeviceInfo> AudioMixer::get_devices_mutex_held() const
579 map<DeviceSpec, DeviceInfo> devices;
580 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
581 const DeviceSpec spec{ InputSourceType::CAPTURE_CARD, card_index };
582 const AudioDevice *device = &video_cards[card_index];
584 info.name = device->name;
585 info.num_channels = 8; // FIXME: This is wrong for fake cards.
586 devices.insert(make_pair(spec, info));
588 for (unsigned card_index = 0; card_index < available_alsa_cards.size(); ++card_index) {
589 const DeviceSpec spec{ InputSourceType::ALSA_INPUT, card_index };
590 const ALSAInput::Device &device = available_alsa_cards[card_index];
592 info.name = device.name + " (" + device.info + ")";
593 info.num_channels = device.num_channels;
594 devices.insert(make_pair(spec, info));
599 void AudioMixer::set_name(DeviceSpec device_spec, const string &name)
601 AudioDevice *device = find_audio_device(device_spec);
603 lock_guard<timed_mutex> lock(audio_mutex);
607 void AudioMixer::set_input_mapping(const InputMapping &new_input_mapping)
609 lock_guard<timed_mutex> lock(audio_mutex);
611 map<DeviceSpec, set<unsigned>> interesting_channels;
612 for (const InputMapping::Bus &bus : new_input_mapping.buses) {
613 if (bus.device.type == InputSourceType::CAPTURE_CARD ||
614 bus.device.type == InputSourceType::ALSA_INPUT) {
615 for (unsigned channel = 0; channel < 2; ++channel) {
616 if (bus.source_channel[channel] != -1) {
617 interesting_channels[bus.device].insert(bus.source_channel[channel]);
623 // Reset resamplers for all cards that don't have the exact same state as before.
624 for (const auto &spec_and_info : get_devices_mutex_held()) {
625 const DeviceSpec &device_spec = spec_and_info.first;
626 AudioDevice *device = find_audio_device(device_spec);
627 if (device->interesting_channels != interesting_channels[device_spec]) {
628 device->interesting_channels = interesting_channels[device_spec];
629 if (device_spec.type == InputSourceType::ALSA_INPUT) {
630 reset_alsa_mutex_held(device_spec);
632 reset_resampler_mutex_held(device_spec);
637 lock_guard<mutex> lock(audio_measure_mutex);
638 bus_r128.resize(new_input_mapping.buses.size());
639 for (unsigned bus_index = 0; bus_index < bus_r128.size(); ++bus_index) {
640 if (bus_r128[bus_index] == nullptr) {
641 bus_r128[bus_index].reset(new Ebu_r128_proc);
643 bus_r128[bus_index]->init(2, OUTPUT_FREQUENCY);
647 input_mapping = new_input_mapping;
650 InputMapping AudioMixer::get_input_mapping() const
652 lock_guard<timed_mutex> lock(audio_mutex);
653 return input_mapping;