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 locut.init(FILTER_HPF, 2);
115 set_locut_enabled(global_flags.locut_enabled);
116 set_gain_staging_db(global_flags.initial_gain_staging_db);
117 set_gain_staging_auto(global_flags.gain_staging_auto);
118 set_compressor_enabled(global_flags.compressor_enabled);
119 set_limiter_enabled(global_flags.limiter_enabled);
120 set_final_makeup_gain_auto(global_flags.final_makeup_gain_auto);
122 // Generate a very simple, default input mapping.
123 InputMapping::Bus input;
125 input.device.type = InputSourceType::CAPTURE_CARD;
126 input.device.index = 0;
127 input.source_channel[0] = 0;
128 input.source_channel[1] = 1;
130 InputMapping new_input_mapping;
131 new_input_mapping.buses.push_back(input);
132 set_input_mapping(new_input_mapping);
134 // Look for ALSA cards.
135 available_alsa_cards = ALSAInput::enumerate_devices();
137 r128.init(2, OUTPUT_FREQUENCY);
140 // hlen=16 is pretty low quality, but we use quite a bit of CPU otherwise,
141 // and there's a limit to how important the peak meter is.
142 peak_resampler.setup(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY * 4, /*num_channels=*/2, /*hlen=*/16, /*frel=*/1.0);
145 AudioMixer::~AudioMixer()
147 for (unsigned card_index = 0; card_index < available_alsa_cards.size(); ++card_index) {
148 const AudioDevice &device = alsa_inputs[card_index];
149 if (device.alsa_device != nullptr) {
150 device.alsa_device->stop_capture_thread();
156 void AudioMixer::reset_resampler(DeviceSpec device_spec)
158 lock_guard<timed_mutex> lock(audio_mutex);
159 reset_resampler_mutex_held(device_spec);
162 void AudioMixer::reset_resampler_mutex_held(DeviceSpec device_spec)
164 AudioDevice *device = find_audio_device(device_spec);
166 if (device->interesting_channels.empty()) {
167 device->resampling_queue.reset();
169 // TODO: ResamplingQueue should probably take the full device spec.
170 // (It's only used for console output, though.)
171 device->resampling_queue.reset(new ResamplingQueue(device_spec.index, device->capture_frequency, OUTPUT_FREQUENCY, device->interesting_channels.size()));
173 device->next_local_pts = 0;
176 void AudioMixer::reset_alsa_mutex_held(DeviceSpec device_spec)
178 assert(device_spec.type == InputSourceType::ALSA_INPUT);
179 unsigned card_index = device_spec.index;
180 AudioDevice *device = find_audio_device(device_spec);
182 if (device->alsa_device != nullptr) {
183 device->alsa_device->stop_capture_thread();
185 if (device->interesting_channels.empty()) {
186 device->alsa_device.reset();
188 device->alsa_device.reset(new ALSAInput(available_alsa_cards[card_index].address.c_str(), OUTPUT_FREQUENCY, 2, bind(&AudioMixer::add_audio, this, device_spec, _1, _2, _3, _4)));
189 device->capture_frequency = device->alsa_device->get_sample_rate();
190 device->alsa_device->start_capture_thread();
194 bool AudioMixer::add_audio(DeviceSpec device_spec, const uint8_t *data, unsigned num_samples, AudioFormat audio_format, int64_t frame_length)
196 AudioDevice *device = find_audio_device(device_spec);
198 unique_lock<timed_mutex> lock(audio_mutex, defer_lock);
199 if (!lock.try_lock_for(chrono::milliseconds(10))) {
202 if (device->resampling_queue == nullptr) {
203 // No buses use this device; throw it away.
207 unsigned num_channels = device->interesting_channels.size();
208 assert(num_channels > 0);
210 // Convert the audio to fp32.
212 audio.resize(num_samples * num_channels);
213 unsigned channel_index = 0;
214 for (auto channel_it = device->interesting_channels.cbegin(); channel_it != device->interesting_channels.end(); ++channel_it, ++channel_index) {
215 switch (audio_format.bits_per_sample) {
217 assert(num_samples == 0);
220 convert_fixed16_to_fp32(&audio[0], channel_index, num_channels, data, *channel_it, audio_format.num_channels, num_samples);
223 convert_fixed24_to_fp32(&audio[0], channel_index, num_channels, data, *channel_it, audio_format.num_channels, num_samples);
226 convert_fixed32_to_fp32(&audio[0], channel_index, num_channels, data, *channel_it, audio_format.num_channels, num_samples);
229 fprintf(stderr, "Cannot handle audio with %u bits per sample\n", audio_format.bits_per_sample);
235 int64_t local_pts = device->next_local_pts;
236 device->resampling_queue->add_input_samples(local_pts / double(TIMEBASE), audio.data(), num_samples);
237 device->next_local_pts = local_pts + frame_length;
241 bool AudioMixer::add_silence(DeviceSpec device_spec, unsigned samples_per_frame, unsigned num_frames, int64_t frame_length)
243 AudioDevice *device = find_audio_device(device_spec);
245 unique_lock<timed_mutex> lock(audio_mutex, defer_lock);
246 if (!lock.try_lock_for(chrono::milliseconds(10))) {
249 if (device->resampling_queue == nullptr) {
250 // No buses use this device; throw it away.
254 unsigned num_channels = device->interesting_channels.size();
255 assert(num_channels > 0);
257 vector<float> silence(samples_per_frame * num_channels, 0.0f);
258 for (unsigned i = 0; i < num_frames; ++i) {
259 device->resampling_queue->add_input_samples(device->next_local_pts / double(TIMEBASE), silence.data(), samples_per_frame);
260 // Note that if the format changed in the meantime, we have
261 // no way of detecting that; we just have to assume the frame length
262 // is always the same.
263 device->next_local_pts += frame_length;
268 AudioMixer::AudioDevice *AudioMixer::find_audio_device(DeviceSpec device)
270 switch (device.type) {
271 case InputSourceType::CAPTURE_CARD:
272 return &video_cards[device.index];
273 case InputSourceType::ALSA_INPUT:
274 return &alsa_inputs[device.index];
275 case InputSourceType::SILENCE:
282 // Get a pointer to the given channel from the given device.
283 // The channel must be picked out earlier and resampled.
284 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)
286 static float zero = 0.0f;
287 if (source_channel == -1 || device_spec.type == InputSourceType::SILENCE) {
292 AudioDevice *device = find_audio_device(device_spec);
293 assert(device->interesting_channels.count(source_channel) != 0);
294 unsigned channel_index = 0;
295 for (int channel : device->interesting_channels) {
296 if (channel == source_channel) break;
299 assert(channel_index < device->interesting_channels.size());
300 const auto it = samples_card.find(device_spec);
301 assert(it != samples_card.end());
302 *srcptr = &(it->second)[channel_index];
303 *stride = device->interesting_channels.size();
306 // TODO: Can be SSSE3-optimized if need be.
307 void AudioMixer::fill_audio_bus(const map<DeviceSpec, vector<float>> &samples_card, const InputMapping::Bus &bus, unsigned num_samples, float *output)
309 if (bus.device.type == InputSourceType::SILENCE) {
310 memset(output, 0, num_samples * sizeof(*output));
312 assert(bus.device.type == InputSourceType::CAPTURE_CARD ||
313 bus.device.type == InputSourceType::ALSA_INPUT);
314 const float *lsrc, *rsrc;
315 unsigned lstride, rstride;
316 float *dptr = output;
317 find_sample_src_from_device(samples_card, bus.device, bus.source_channel[0], &lsrc, &lstride);
318 find_sample_src_from_device(samples_card, bus.device, bus.source_channel[1], &rsrc, &rstride);
319 for (unsigned i = 0; i < num_samples; ++i) {
328 vector<float> AudioMixer::get_output(double pts, unsigned num_samples, ResamplingQueue::RateAdjustmentPolicy rate_adjustment_policy)
330 map<DeviceSpec, vector<float>> samples_card;
331 vector<float> samples_bus;
333 lock_guard<timed_mutex> lock(audio_mutex);
335 // Pick out all the interesting channels from all the cards.
336 // TODO: If the card has been hotswapped, the number of channels
337 // might have changed; if so, we need to do some sort of remapping
339 for (const auto &spec_and_info : get_devices_mutex_held()) {
340 const DeviceSpec &device_spec = spec_and_info.first;
341 AudioDevice *device = find_audio_device(device_spec);
342 if (!device->interesting_channels.empty()) {
343 samples_card[device_spec].resize(num_samples * device->interesting_channels.size());
344 device->resampling_queue->get_output_samples(
346 &samples_card[device_spec][0],
348 rate_adjustment_policy);
352 // TODO: Move lo-cut etc. into each bus.
353 vector<float> samples_out;
354 samples_out.resize(num_samples * 2);
355 samples_bus.resize(num_samples * 2);
356 for (unsigned bus_index = 0; bus_index < input_mapping.buses.size(); ++bus_index) {
357 fill_audio_bus(samples_card, input_mapping.buses[bus_index], num_samples, &samples_bus[0]);
359 float volume = from_db(fader_volume_db[bus_index]);
360 if (bus_index == 0) {
361 for (unsigned i = 0; i < num_samples * 2; ++i) {
362 samples_out[i] = samples_bus[i] * volume;
365 for (unsigned i = 0; i < num_samples * 2; ++i) {
366 samples_out[i] += samples_bus[i] * volume;
371 // Cut away everything under 120 Hz (or whatever the cutoff is);
372 // we don't need it for voice, and it will reduce headroom
373 // and confuse the compressor. (In particular, any hums at 50 or 60 Hz
374 // should be dampened.)
376 locut.render(samples_out.data(), samples_out.size() / 2, locut_cutoff_hz * 2.0 * M_PI / OUTPUT_FREQUENCY, 0.5f);
380 lock_guard<mutex> lock(compressor_mutex);
382 // Apply a level compressor to get the general level right.
383 // Basically, if it's over about -40 dBFS, we squeeze it down to that level
384 // (or more precisely, near it, since we don't use infinite ratio),
385 // then apply a makeup gain to get it to -14 dBFS. -14 dBFS is, of course,
386 // entirely arbitrary, but from practical tests with speech, it seems to
387 // put ut around -23 LUFS, so it's a reasonable starting point for later use.
389 if (level_compressor_enabled) {
390 float threshold = 0.01f; // -40 dBFS.
392 float attack_time = 0.5f;
393 float release_time = 20.0f;
394 float makeup_gain = from_db(ref_level_dbfs - (-40.0f)); // +26 dB.
395 level_compressor.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
396 gain_staging_db = to_db(level_compressor.get_attenuation() * makeup_gain);
398 // Just apply the gain we already had.
399 float g = from_db(gain_staging_db);
400 for (size_t i = 0; i < samples_out.size(); ++i) {
407 printf("level=%f (%+5.2f dBFS) attenuation=%f (%+5.2f dB) end_result=%+5.2f dB\n",
408 level_compressor.get_level(), to_db(level_compressor.get_level()),
409 level_compressor.get_attenuation(), to_db(level_compressor.get_attenuation()),
410 to_db(level_compressor.get_level() * level_compressor.get_attenuation() * makeup_gain));
413 // float limiter_att, compressor_att;
415 // The real compressor.
416 if (compressor_enabled) {
417 float threshold = from_db(compressor_threshold_dbfs);
419 float attack_time = 0.005f;
420 float release_time = 0.040f;
421 float makeup_gain = 2.0f; // +6 dB.
422 compressor.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
423 // compressor_att = compressor.get_attenuation();
426 // Finally a limiter at -4 dB (so, -10 dBFS) to take out the worst peaks only.
427 // Note that since ratio is not infinite, we could go slightly higher than this.
428 if (limiter_enabled) {
429 float threshold = from_db(limiter_threshold_dbfs);
431 float attack_time = 0.0f; // Instant.
432 float release_time = 0.020f;
433 float makeup_gain = 1.0f; // 0 dB.
434 limiter.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
435 // limiter_att = limiter.get_attenuation();
438 // printf("limiter=%+5.1f compressor=%+5.1f\n", to_db(limiter_att), to_db(compressor_att));
441 // At this point, we are most likely close to +0 LU, but all of our
442 // measurements have been on raw sample values, not R128 values.
443 // So we have a final makeup gain to get us to +0 LU; the gain
444 // adjustments required should be relatively small, and also, the
445 // offset shouldn't change much (only if the type of audio changes
446 // significantly). Thus, we shoot for updating this value basically
447 // “whenever we process buffers”, since the R128 calculation isn't exactly
448 // something we get out per-sample.
450 // Note that there's a feedback loop here, so we choose a very slow filter
451 // (half-time of 30 seconds).
452 double target_loudness_factor, alpha;
453 double loudness_lu = r128.loudness_M() - ref_level_lufs;
454 double current_makeup_lu = to_db(final_makeup_gain);
455 target_loudness_factor = final_makeup_gain * from_db(-loudness_lu);
457 // If we're outside +/- 5 LU uncorrected, we don't count it as
458 // a normal signal (probably silence) and don't change the
459 // correction factor; just apply what we already have.
460 if (fabs(loudness_lu - current_makeup_lu) >= 5.0 || !final_makeup_gain_auto) {
463 // Formula adapted from
464 // https://en.wikipedia.org/wiki/Low-pass_filter#Simple_infinite_impulse_response_filter.
465 const double half_time_s = 30.0;
466 const double fc_mul_2pi_delta_t = 1.0 / (half_time_s * OUTPUT_FREQUENCY);
467 alpha = fc_mul_2pi_delta_t / (fc_mul_2pi_delta_t + 1.0);
471 lock_guard<mutex> lock(compressor_mutex);
472 double m = final_makeup_gain;
473 for (size_t i = 0; i < samples_out.size(); i += 2) {
474 samples_out[i + 0] *= m;
475 samples_out[i + 1] *= m;
476 m += (target_loudness_factor - m) * alpha;
478 final_makeup_gain = m;
481 update_meters(samples_out);
486 void AudioMixer::update_meters(const vector<float> &samples)
488 // Upsample 4x to find interpolated peak.
489 peak_resampler.inp_data = const_cast<float *>(samples.data());
490 peak_resampler.inp_count = samples.size() / 2;
492 vector<float> interpolated_samples;
493 interpolated_samples.resize(samples.size());
495 unique_lock<mutex> lock(audio_measure_mutex);
497 while (peak_resampler.inp_count > 0) { // About four iterations.
498 peak_resampler.out_data = &interpolated_samples[0];
499 peak_resampler.out_count = interpolated_samples.size() / 2;
500 peak_resampler.process();
501 size_t out_stereo_samples = interpolated_samples.size() / 2 - peak_resampler.out_count;
502 peak = max<float>(peak, find_peak(interpolated_samples.data(), out_stereo_samples * 2));
503 peak_resampler.out_data = nullptr;
507 // Find R128 levels and L/R correlation.
508 vector<float> left, right;
509 deinterleave_samples(samples, &left, &right);
510 float *ptrs[] = { left.data(), right.data() };
512 unique_lock<mutex> lock(audio_measure_mutex);
513 r128.process(left.size(), ptrs);
514 correlation.process_samples(samples);
517 send_audio_level_callback();
520 void AudioMixer::reset_meters()
522 unique_lock<mutex> lock(audio_measure_mutex);
523 peak_resampler.reset();
530 void AudioMixer::send_audio_level_callback()
532 if (audio_level_callback == nullptr) {
536 unique_lock<mutex> lock(audio_measure_mutex);
537 double loudness_s = r128.loudness_S();
538 double loudness_i = r128.integrated();
539 double loudness_range_low = r128.range_min();
540 double loudness_range_high = r128.range_max();
542 audio_level_callback(loudness_s, to_db(peak),
543 loudness_i, loudness_range_low, loudness_range_high,
545 to_db(final_makeup_gain),
546 correlation.get_correlation());
549 map<DeviceSpec, DeviceInfo> AudioMixer::get_devices() const
551 lock_guard<timed_mutex> lock(audio_mutex);
552 return get_devices_mutex_held();
555 map<DeviceSpec, DeviceInfo> AudioMixer::get_devices_mutex_held() const
557 map<DeviceSpec, DeviceInfo> devices;
558 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
559 const DeviceSpec spec{ InputSourceType::CAPTURE_CARD, card_index };
560 const AudioDevice *device = &video_cards[card_index];
562 info.name = device->name;
563 info.num_channels = 8; // FIXME: This is wrong for fake cards.
564 devices.insert(make_pair(spec, info));
566 for (unsigned card_index = 0; card_index < available_alsa_cards.size(); ++card_index) {
567 const DeviceSpec spec{ InputSourceType::ALSA_INPUT, card_index };
568 const ALSAInput::Device &device = available_alsa_cards[card_index];
570 info.name = device.name + " (" + device.info + ")";
571 info.num_channels = device.num_channels;
572 devices.insert(make_pair(spec, info));
577 void AudioMixer::set_name(DeviceSpec device_spec, const string &name)
579 AudioDevice *device = find_audio_device(device_spec);
581 lock_guard<timed_mutex> lock(audio_mutex);
585 void AudioMixer::set_input_mapping(const InputMapping &new_input_mapping)
587 lock_guard<timed_mutex> lock(audio_mutex);
589 map<DeviceSpec, set<unsigned>> interesting_channels;
590 for (const InputMapping::Bus &bus : new_input_mapping.buses) {
591 if (bus.device.type == InputSourceType::CAPTURE_CARD ||
592 bus.device.type == InputSourceType::ALSA_INPUT) {
593 for (unsigned channel = 0; channel < 2; ++channel) {
594 if (bus.source_channel[channel] != -1) {
595 interesting_channels[bus.device].insert(bus.source_channel[channel]);
601 // Reset resamplers for all cards that don't have the exact same state as before.
602 for (const auto &spec_and_info : get_devices_mutex_held()) {
603 const DeviceSpec &device_spec = spec_and_info.first;
604 AudioDevice *device = find_audio_device(device_spec);
605 if (device->interesting_channels != interesting_channels[device_spec]) {
606 device->interesting_channels = interesting_channels[device_spec];
607 if (device_spec.type == InputSourceType::ALSA_INPUT) {
608 reset_alsa_mutex_held(device_spec);
610 reset_resampler_mutex_held(device_spec);
614 input_mapping = new_input_mapping;
617 InputMapping AudioMixer::get_input_mapping() const
619 lock_guard<timed_mutex> lock(audio_mutex);
620 return input_mapping;