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 level_compressor(OUTPUT_FREQUENCY),
109 limiter(OUTPUT_FREQUENCY),
110 compressor(OUTPUT_FREQUENCY),
111 correlation(OUTPUT_FREQUENCY)
113 for (unsigned bus_index = 0; bus_index < MAX_BUSES; ++bus_index) {
114 locut[bus_index].init(FILTER_HPF, 2);
115 locut_enabled[bus_index] = global_flags.locut_enabled;
117 set_gain_staging_db(global_flags.initial_gain_staging_db);
118 set_gain_staging_auto(global_flags.gain_staging_auto);
119 set_compressor_enabled(global_flags.compressor_enabled);
120 set_limiter_enabled(global_flags.limiter_enabled);
121 set_final_makeup_gain_auto(global_flags.final_makeup_gain_auto);
123 // Generate a very simple, default input mapping.
124 InputMapping::Bus input;
126 input.device.type = InputSourceType::CAPTURE_CARD;
127 input.device.index = 0;
128 input.source_channel[0] = 0;
129 input.source_channel[1] = 1;
131 InputMapping new_input_mapping;
132 new_input_mapping.buses.push_back(input);
133 set_input_mapping(new_input_mapping);
135 // Look for ALSA cards.
136 available_alsa_cards = ALSAInput::enumerate_devices();
138 r128.init(2, OUTPUT_FREQUENCY);
141 // hlen=16 is pretty low quality, but we use quite a bit of CPU otherwise,
142 // and there's a limit to how important the peak meter is.
143 peak_resampler.setup(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY * 4, /*num_channels=*/2, /*hlen=*/16, /*frel=*/1.0);
146 AudioMixer::~AudioMixer()
148 for (unsigned card_index = 0; card_index < available_alsa_cards.size(); ++card_index) {
149 const AudioDevice &device = alsa_inputs[card_index];
150 if (device.alsa_device != nullptr) {
151 device.alsa_device->stop_capture_thread();
157 void AudioMixer::reset_resampler(DeviceSpec device_spec)
159 lock_guard<timed_mutex> lock(audio_mutex);
160 reset_resampler_mutex_held(device_spec);
163 void AudioMixer::reset_resampler_mutex_held(DeviceSpec device_spec)
165 AudioDevice *device = find_audio_device(device_spec);
167 if (device->interesting_channels.empty()) {
168 device->resampling_queue.reset();
170 // TODO: ResamplingQueue should probably take the full device spec.
171 // (It's only used for console output, though.)
172 device->resampling_queue.reset(new ResamplingQueue(device_spec.index, device->capture_frequency, OUTPUT_FREQUENCY, device->interesting_channels.size()));
174 device->next_local_pts = 0;
177 void AudioMixer::reset_alsa_mutex_held(DeviceSpec device_spec)
179 assert(device_spec.type == InputSourceType::ALSA_INPUT);
180 unsigned card_index = device_spec.index;
181 AudioDevice *device = find_audio_device(device_spec);
183 if (device->alsa_device != nullptr) {
184 device->alsa_device->stop_capture_thread();
186 if (device->interesting_channels.empty()) {
187 device->alsa_device.reset();
189 const ALSAInput::Device &alsa_dev = available_alsa_cards[card_index];
190 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)));
191 device->capture_frequency = device->alsa_device->get_sample_rate();
192 device->alsa_device->start_capture_thread();
196 bool AudioMixer::add_audio(DeviceSpec device_spec, const uint8_t *data, unsigned num_samples, AudioFormat audio_format, int64_t frame_length)
198 AudioDevice *device = find_audio_device(device_spec);
200 unique_lock<timed_mutex> lock(audio_mutex, defer_lock);
201 if (!lock.try_lock_for(chrono::milliseconds(10))) {
204 if (device->resampling_queue == nullptr) {
205 // No buses use this device; throw it away.
209 unsigned num_channels = device->interesting_channels.size();
210 assert(num_channels > 0);
212 // Convert the audio to fp32.
214 audio.resize(num_samples * num_channels);
215 unsigned channel_index = 0;
216 for (auto channel_it = device->interesting_channels.cbegin(); channel_it != device->interesting_channels.end(); ++channel_it, ++channel_index) {
217 switch (audio_format.bits_per_sample) {
219 assert(num_samples == 0);
222 convert_fixed16_to_fp32(&audio[0], channel_index, num_channels, data, *channel_it, audio_format.num_channels, num_samples);
225 convert_fixed24_to_fp32(&audio[0], channel_index, num_channels, data, *channel_it, audio_format.num_channels, num_samples);
228 convert_fixed32_to_fp32(&audio[0], channel_index, num_channels, data, *channel_it, audio_format.num_channels, num_samples);
231 fprintf(stderr, "Cannot handle audio with %u bits per sample\n", audio_format.bits_per_sample);
237 int64_t local_pts = device->next_local_pts;
238 device->resampling_queue->add_input_samples(local_pts / double(TIMEBASE), audio.data(), num_samples);
239 device->next_local_pts = local_pts + frame_length;
243 bool AudioMixer::add_silence(DeviceSpec device_spec, unsigned samples_per_frame, unsigned num_frames, int64_t frame_length)
245 AudioDevice *device = find_audio_device(device_spec);
247 unique_lock<timed_mutex> lock(audio_mutex, defer_lock);
248 if (!lock.try_lock_for(chrono::milliseconds(10))) {
251 if (device->resampling_queue == nullptr) {
252 // No buses use this device; throw it away.
256 unsigned num_channels = device->interesting_channels.size();
257 assert(num_channels > 0);
259 vector<float> silence(samples_per_frame * num_channels, 0.0f);
260 for (unsigned i = 0; i < num_frames; ++i) {
261 device->resampling_queue->add_input_samples(device->next_local_pts / double(TIMEBASE), silence.data(), samples_per_frame);
262 // Note that if the format changed in the meantime, we have
263 // no way of detecting that; we just have to assume the frame length
264 // is always the same.
265 device->next_local_pts += frame_length;
270 AudioMixer::AudioDevice *AudioMixer::find_audio_device(DeviceSpec device)
272 switch (device.type) {
273 case InputSourceType::CAPTURE_CARD:
274 return &video_cards[device.index];
275 case InputSourceType::ALSA_INPUT:
276 return &alsa_inputs[device.index];
277 case InputSourceType::SILENCE:
284 // Get a pointer to the given channel from the given device.
285 // The channel must be picked out earlier and resampled.
286 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)
288 static float zero = 0.0f;
289 if (source_channel == -1 || device_spec.type == InputSourceType::SILENCE) {
294 AudioDevice *device = find_audio_device(device_spec);
295 assert(device->interesting_channels.count(source_channel) != 0);
296 unsigned channel_index = 0;
297 for (int channel : device->interesting_channels) {
298 if (channel == source_channel) break;
301 assert(channel_index < device->interesting_channels.size());
302 const auto it = samples_card.find(device_spec);
303 assert(it != samples_card.end());
304 *srcptr = &(it->second)[channel_index];
305 *stride = device->interesting_channels.size();
308 // TODO: Can be SSSE3-optimized if need be.
309 void AudioMixer::fill_audio_bus(const map<DeviceSpec, vector<float>> &samples_card, const InputMapping::Bus &bus, unsigned num_samples, float *output)
311 if (bus.device.type == InputSourceType::SILENCE) {
312 memset(output, 0, num_samples * sizeof(*output));
314 assert(bus.device.type == InputSourceType::CAPTURE_CARD ||
315 bus.device.type == InputSourceType::ALSA_INPUT);
316 const float *lsrc, *rsrc;
317 unsigned lstride, rstride;
318 float *dptr = output;
319 find_sample_src_from_device(samples_card, bus.device, bus.source_channel[0], &lsrc, &lstride);
320 find_sample_src_from_device(samples_card, bus.device, bus.source_channel[1], &rsrc, &rstride);
321 for (unsigned i = 0; i < num_samples; ++i) {
330 vector<float> AudioMixer::get_output(double pts, unsigned num_samples, ResamplingQueue::RateAdjustmentPolicy rate_adjustment_policy)
332 map<DeviceSpec, vector<float>> samples_card;
333 vector<float> samples_bus;
335 lock_guard<timed_mutex> lock(audio_mutex);
337 // Pick out all the interesting channels from all the cards.
338 // TODO: If the card has been hotswapped, the number of channels
339 // might have changed; if so, we need to do some sort of remapping
341 for (const auto &spec_and_info : get_devices_mutex_held()) {
342 const DeviceSpec &device_spec = spec_and_info.first;
343 AudioDevice *device = find_audio_device(device_spec);
344 if (!device->interesting_channels.empty()) {
345 samples_card[device_spec].resize(num_samples * device->interesting_channels.size());
346 device->resampling_queue->get_output_samples(
348 &samples_card[device_spec][0],
350 rate_adjustment_policy);
354 // TODO: Move lo-cut etc. into each bus.
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);
369 // TODO: We should measure post-fader.
370 deinterleave_samples(samples_bus, &left, &right);
371 measure_bus_levels(bus_index, left, right);
373 float volume = from_db(fader_volume_db[bus_index]);
374 if (bus_index == 0) {
375 for (unsigned i = 0; i < num_samples * 2; ++i) {
376 samples_out[i] = samples_bus[i] * volume;
379 for (unsigned i = 0; i < num_samples * 2; ++i) {
380 samples_out[i] += samples_bus[i] * volume;
386 lock_guard<mutex> lock(compressor_mutex);
388 // Apply a level compressor to get the general level right.
389 // Basically, if it's over about -40 dBFS, we squeeze it down to that level
390 // (or more precisely, near it, since we don't use infinite ratio),
391 // then apply a makeup gain to get it to -14 dBFS. -14 dBFS is, of course,
392 // entirely arbitrary, but from practical tests with speech, it seems to
393 // put ut around -23 LUFS, so it's a reasonable starting point for later use.
395 if (level_compressor_enabled) {
396 float threshold = 0.01f; // -40 dBFS.
398 float attack_time = 0.5f;
399 float release_time = 20.0f;
400 float makeup_gain = from_db(ref_level_dbfs - (-40.0f)); // +26 dB.
401 level_compressor.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
402 gain_staging_db = to_db(level_compressor.get_attenuation() * makeup_gain);
404 // Just apply the gain we already had.
405 float g = from_db(gain_staging_db);
406 for (size_t i = 0; i < samples_out.size(); ++i) {
413 printf("level=%f (%+5.2f dBFS) attenuation=%f (%+5.2f dB) end_result=%+5.2f dB\n",
414 level_compressor.get_level(), to_db(level_compressor.get_level()),
415 level_compressor.get_attenuation(), to_db(level_compressor.get_attenuation()),
416 to_db(level_compressor.get_level() * level_compressor.get_attenuation() * makeup_gain));
419 // float limiter_att, compressor_att;
421 // The real compressor.
422 if (compressor_enabled) {
423 float threshold = from_db(compressor_threshold_dbfs);
425 float attack_time = 0.005f;
426 float release_time = 0.040f;
427 float makeup_gain = 2.0f; // +6 dB.
428 compressor.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
429 // compressor_att = compressor.get_attenuation();
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, but all of our
448 // measurements have been on raw sample values, not R128 values.
449 // So we have a final makeup gain to get us to +0 LU; the gain
450 // adjustments required should be relatively small, and also, the
451 // offset shouldn't change much (only if the type of audio changes
452 // significantly). Thus, we shoot for updating this value basically
453 // “whenever we process buffers”, since the R128 calculation isn't exactly
454 // something we get out per-sample.
456 // Note that there's a feedback loop here, so we choose a very slow filter
457 // (half-time of 30 seconds).
458 double target_loudness_factor, alpha;
459 double loudness_lu = r128.loudness_M() - ref_level_lufs;
460 double current_makeup_lu = to_db(final_makeup_gain);
461 target_loudness_factor = final_makeup_gain * from_db(-loudness_lu);
463 // If we're outside +/- 5 LU uncorrected, we don't count it as
464 // a normal signal (probably silence) and don't change the
465 // correction factor; just apply what we already have.
466 if (fabs(loudness_lu - current_makeup_lu) >= 5.0 || !final_makeup_gain_auto) {
469 // Formula adapted from
470 // https://en.wikipedia.org/wiki/Low-pass_filter#Simple_infinite_impulse_response_filter.
471 const double half_time_s = 30.0;
472 const double fc_mul_2pi_delta_t = 1.0 / (half_time_s * OUTPUT_FREQUENCY);
473 alpha = fc_mul_2pi_delta_t / (fc_mul_2pi_delta_t + 1.0);
477 lock_guard<mutex> lock(compressor_mutex);
478 double m = final_makeup_gain;
479 for (size_t i = 0; i < samples_out.size(); i += 2) {
480 samples_out[i + 0] *= m;
481 samples_out[i + 1] *= m;
482 m += (target_loudness_factor - m) * alpha;
484 final_makeup_gain = m;
487 update_meters(samples_out);
492 void AudioMixer::measure_bus_levels(unsigned bus_index, const vector<float> &left, const vector<float> &right)
494 const float *ptrs[] = { left.data(), right.data() };
496 lock_guard<mutex> lock(audio_measure_mutex);
497 bus_r128[bus_index]->process(left.size(), const_cast<float **>(ptrs));
501 void AudioMixer::update_meters(const vector<float> &samples)
503 // Upsample 4x to find interpolated peak.
504 peak_resampler.inp_data = const_cast<float *>(samples.data());
505 peak_resampler.inp_count = samples.size() / 2;
507 vector<float> interpolated_samples;
508 interpolated_samples.resize(samples.size());
510 lock_guard<mutex> lock(audio_measure_mutex);
512 while (peak_resampler.inp_count > 0) { // About four iterations.
513 peak_resampler.out_data = &interpolated_samples[0];
514 peak_resampler.out_count = interpolated_samples.size() / 2;
515 peak_resampler.process();
516 size_t out_stereo_samples = interpolated_samples.size() / 2 - peak_resampler.out_count;
517 peak = max<float>(peak, find_peak(interpolated_samples.data(), out_stereo_samples * 2));
518 peak_resampler.out_data = nullptr;
522 // Find R128 levels and L/R correlation.
523 vector<float> left, right;
524 deinterleave_samples(samples, &left, &right);
525 float *ptrs[] = { left.data(), right.data() };
527 lock_guard<mutex> lock(audio_measure_mutex);
528 r128.process(left.size(), ptrs);
529 correlation.process_samples(samples);
532 send_audio_level_callback();
535 void AudioMixer::reset_meters()
537 lock_guard<mutex> lock(audio_measure_mutex);
538 peak_resampler.reset();
545 void AudioMixer::send_audio_level_callback()
547 if (audio_level_callback == nullptr) {
551 lock_guard<mutex> lock(audio_measure_mutex);
552 double loudness_s = r128.loudness_S();
553 double loudness_i = r128.integrated();
554 double loudness_range_low = r128.range_min();
555 double loudness_range_high = r128.range_max();
557 vector<float> bus_loudness;
558 bus_loudness.resize(input_mapping.buses.size());
559 for (unsigned bus_index = 0; bus_index < bus_r128.size(); ++bus_index) {
560 bus_loudness[bus_index] = bus_r128[bus_index]->loudness_S();
563 audio_level_callback(loudness_s, to_db(peak), bus_loudness,
564 loudness_i, loudness_range_low, loudness_range_high,
566 to_db(final_makeup_gain),
567 correlation.get_correlation());
570 map<DeviceSpec, DeviceInfo> AudioMixer::get_devices() const
572 lock_guard<timed_mutex> lock(audio_mutex);
573 return get_devices_mutex_held();
576 map<DeviceSpec, DeviceInfo> AudioMixer::get_devices_mutex_held() const
578 map<DeviceSpec, DeviceInfo> devices;
579 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
580 const DeviceSpec spec{ InputSourceType::CAPTURE_CARD, card_index };
581 const AudioDevice *device = &video_cards[card_index];
583 info.name = device->name;
584 info.num_channels = 8; // FIXME: This is wrong for fake cards.
585 devices.insert(make_pair(spec, info));
587 for (unsigned card_index = 0; card_index < available_alsa_cards.size(); ++card_index) {
588 const DeviceSpec spec{ InputSourceType::ALSA_INPUT, card_index };
589 const ALSAInput::Device &device = available_alsa_cards[card_index];
591 info.name = device.name + " (" + device.info + ")";
592 info.num_channels = device.num_channels;
593 devices.insert(make_pair(spec, info));
598 void AudioMixer::set_name(DeviceSpec device_spec, const string &name)
600 AudioDevice *device = find_audio_device(device_spec);
602 lock_guard<timed_mutex> lock(audio_mutex);
606 void AudioMixer::set_input_mapping(const InputMapping &new_input_mapping)
608 lock_guard<timed_mutex> lock(audio_mutex);
610 map<DeviceSpec, set<unsigned>> interesting_channels;
611 for (const InputMapping::Bus &bus : new_input_mapping.buses) {
612 if (bus.device.type == InputSourceType::CAPTURE_CARD ||
613 bus.device.type == InputSourceType::ALSA_INPUT) {
614 for (unsigned channel = 0; channel < 2; ++channel) {
615 if (bus.source_channel[channel] != -1) {
616 interesting_channels[bus.device].insert(bus.source_channel[channel]);
622 // Reset resamplers for all cards that don't have the exact same state as before.
623 for (const auto &spec_and_info : get_devices_mutex_held()) {
624 const DeviceSpec &device_spec = spec_and_info.first;
625 AudioDevice *device = find_audio_device(device_spec);
626 if (device->interesting_channels != interesting_channels[device_spec]) {
627 device->interesting_channels = interesting_channels[device_spec];
628 if (device_spec.type == InputSourceType::ALSA_INPUT) {
629 reset_alsa_mutex_held(device_spec);
631 reset_resampler_mutex_held(device_spec);
636 lock_guard<mutex> lock(audio_measure_mutex);
637 bus_r128.resize(new_input_mapping.buses.size());
638 for (unsigned bus_index = 0; bus_index < bus_r128.size(); ++bus_index) {
639 if (bus_r128[bus_index] == nullptr) {
640 bus_r128[bus_index].reset(new Ebu_r128_proc);
642 bus_r128[bus_index]->init(2, OUTPUT_FREQUENCY);
646 input_mapping = new_input_mapping;
649 InputMapping AudioMixer::get_input_mapping() const
651 lock_guard<timed_mutex> lock(audio_mutex);
652 return input_mapping;