1 #include "audio_mixer.h"
5 #include <bmusb/bmusb.h>
14 using namespace bmusb;
19 // TODO: If these prove to be a bottleneck, they can be SSSE3-optimized
20 // (usually including multiple channels at a time).
22 void convert_fixed24_to_fp32(float *dst, size_t out_channel, size_t out_num_channels,
23 const uint8_t *src, size_t in_channel, size_t in_num_channels,
26 assert(in_channel < in_num_channels);
27 assert(out_channel < out_num_channels);
28 src += in_channel * 3;
31 for (size_t i = 0; i < num_samples; ++i) {
35 uint32_t s = s1 | (s1 << 8) | (s2 << 16) | (s3 << 24);
36 *dst = int(s) * (1.0f / 2147483648.0f);
38 src += 3 * in_num_channels;
39 dst += out_num_channels;
43 void convert_fixed32_to_fp32(float *dst, size_t out_channel, size_t out_num_channels,
44 const uint8_t *src, size_t in_channel, size_t in_num_channels,
47 assert(in_channel < in_num_channels);
48 assert(out_channel < out_num_channels);
49 src += in_channel * 4;
52 for (size_t i = 0; i < num_samples; ++i) {
53 int32_t s = le32toh(*(int32_t *)src);
54 *dst = s * (1.0f / 2147483648.0f);
56 src += 4 * in_num_channels;
57 dst += out_num_channels;
63 AudioMixer::AudioMixer(unsigned num_cards)
64 : num_cards(num_cards),
65 level_compressor(OUTPUT_FREQUENCY),
66 limiter(OUTPUT_FREQUENCY),
67 compressor(OUTPUT_FREQUENCY)
69 locut.init(FILTER_HPF, 2);
71 set_locut_enabled(global_flags.locut_enabled);
72 set_gain_staging_db(global_flags.initial_gain_staging_db);
73 set_gain_staging_auto(global_flags.gain_staging_auto);
74 set_compressor_enabled(global_flags.compressor_enabled);
75 set_limiter_enabled(global_flags.limiter_enabled);
76 set_final_makeup_gain_auto(global_flags.final_makeup_gain_auto);
78 // Generate a very simple, default input mapping.
79 InputMapping::Bus input;
81 input.device.type = InputSourceType::CAPTURE_CARD;
82 input.device.index = 0;
83 input.source_channel[0] = 0;
84 input.source_channel[1] = 1;
86 InputMapping new_input_mapping;
87 new_input_mapping.buses.push_back(input);
88 set_input_mapping(new_input_mapping);
91 void AudioMixer::reset_device(DeviceSpec device_spec)
93 lock_guard<mutex> lock(audio_mutex);
94 reset_device_mutex_held(device_spec);
97 void AudioMixer::reset_device_mutex_held(DeviceSpec device_spec)
99 AudioDevice *device = find_audio_device(device_spec);
100 if (device->interesting_channels.empty()) {
101 device->resampling_queue.reset();
103 // TODO: ResamplingQueue should probably take the full device spec.
104 // (It's only used for console output, though.)
105 device->resampling_queue.reset(new ResamplingQueue(device_spec.index, OUTPUT_FREQUENCY, OUTPUT_FREQUENCY, device->interesting_channels.size()));
107 device->next_local_pts = 0;
110 void AudioMixer::add_audio(DeviceSpec device_spec, const uint8_t *data, unsigned num_samples, AudioFormat audio_format, int64_t frame_length)
112 AudioDevice *device = find_audio_device(device_spec);
114 lock_guard<mutex> lock(audio_mutex);
115 if (device->resampling_queue == nullptr) {
116 // No buses use this device; throw it away.
120 unsigned num_channels = device->interesting_channels.size();
121 assert(num_channels > 0);
123 // Convert the audio to stereo fp32.
125 audio.resize(num_samples * num_channels);
126 unsigned channel_index = 0;
127 for (auto channel_it = device->interesting_channels.cbegin(); channel_it != device->interesting_channels.end(); ++channel_it, ++channel_index) {
128 switch (audio_format.bits_per_sample) {
130 assert(num_samples == 0);
133 convert_fixed24_to_fp32(&audio[0], channel_index, num_channels, data, *channel_it, audio_format.num_channels, num_samples);
136 convert_fixed32_to_fp32(&audio[0], channel_index, num_channels, data, *channel_it, audio_format.num_channels, num_samples);
139 fprintf(stderr, "Cannot handle audio with %u bits per sample\n", audio_format.bits_per_sample);
145 int64_t local_pts = device->next_local_pts;
146 device->resampling_queue->add_input_samples(local_pts / double(TIMEBASE), audio.data(), num_samples);
147 device->next_local_pts = local_pts + frame_length;
150 void AudioMixer::add_silence(DeviceSpec device_spec, unsigned samples_per_frame, unsigned num_frames, int64_t frame_length)
152 AudioDevice *device = find_audio_device(device_spec);
154 lock_guard<mutex> lock(audio_mutex);
155 if (device->resampling_queue == nullptr) {
156 // No buses use this device; throw it away.
160 unsigned num_channels = device->interesting_channels.size();
161 assert(num_channels > 0);
163 vector<float> silence(samples_per_frame * num_channels, 0.0f);
164 for (unsigned i = 0; i < num_frames; ++i) {
165 device->resampling_queue->add_input_samples(device->next_local_pts / double(TIMEBASE), silence.data(), samples_per_frame);
166 // Note that if the format changed in the meantime, we have
167 // no way of detecting that; we just have to assume the frame length
168 // is always the same.
169 device->next_local_pts += frame_length;
173 AudioMixer::AudioDevice *AudioMixer::find_audio_device(DeviceSpec device)
175 switch (device.type) {
176 case InputSourceType::CAPTURE_CARD:
177 return &cards[device.index];
179 case InputSourceType::SILENCE:
186 void AudioMixer::find_sample_src_from_device(const vector<float> *samples_card, DeviceSpec device_spec, int source_channel, const float **srcptr, unsigned *stride)
188 static float zero = 0.0f;
189 if (source_channel == -1 || device_spec.type == InputSourceType::SILENCE) {
194 AudioDevice *device = find_audio_device(device_spec);
195 unsigned channel_index = 0;
196 for (int channel : device->interesting_channels) {
197 if (channel == source_channel) break;
200 assert(channel_index < device->interesting_channels.size());
201 *srcptr = &samples_card[device_spec.index][channel_index];
202 *stride = device->interesting_channels.size();
205 // TODO: Can be SSSE3-optimized if need be.
206 void AudioMixer::fill_audio_bus(const vector<float> *samples_card, const InputMapping::Bus &bus, unsigned num_samples, float *output)
208 if (bus.device.type == InputSourceType::SILENCE) {
209 memset(output, 0, num_samples * sizeof(*output));
211 assert(bus.device.type == InputSourceType::CAPTURE_CARD);
212 const float *lsrc, *rsrc;
213 unsigned lstride, rstride;
214 float *dptr = output;
215 find_sample_src_from_device(samples_card, bus.device, bus.source_channel[0], &lsrc, &lstride);
216 find_sample_src_from_device(samples_card, bus.device, bus.source_channel[1], &rsrc, &rstride);
217 for (unsigned i = 0; i < num_samples; ++i) {
226 vector<float> AudioMixer::get_output(double pts, unsigned num_samples, ResamplingQueue::RateAdjustmentPolicy rate_adjustment_policy)
228 vector<float> samples_card[MAX_CARDS]; // TODO: Needs room for other kinds of capture cards.
229 vector<float> samples_bus;
231 lock_guard<mutex> lock(audio_mutex);
233 // Pick out all the interesting channels from all the cards.
234 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
235 AudioDevice *device = &cards[card_index];
236 if (!device->interesting_channels.empty()) {
237 samples_card[card_index].resize(num_samples * device->interesting_channels.size());
238 device->resampling_queue->get_output_samples(
240 &samples_card[card_index][0],
242 rate_adjustment_policy);
246 // TODO: Move lo-cut etc. into each bus.
247 vector<float> samples_out;
248 samples_out.resize(num_samples * 2);
249 samples_bus.resize(num_samples * 2);
250 for (unsigned bus_index = 0; bus_index < input_mapping.buses.size(); ++bus_index) {
251 fill_audio_bus(samples_card, input_mapping.buses[bus_index], num_samples, &samples_bus[0]);
253 float volume = from_db(fader_volume_db[bus_index]);
254 if (bus_index == 0) {
255 for (unsigned i = 0; i < num_samples * 2; ++i) {
256 samples_out[i] = samples_bus[i] * volume;
259 for (unsigned i = 0; i < num_samples * 2; ++i) {
260 samples_out[i] += samples_bus[i] * volume;
265 // Cut away everything under 120 Hz (or whatever the cutoff is);
266 // we don't need it for voice, and it will reduce headroom
267 // and confuse the compressor. (In particular, any hums at 50 or 60 Hz
268 // should be dampened.)
270 locut.render(samples_out.data(), samples_out.size() / 2, locut_cutoff_hz * 2.0 * M_PI / OUTPUT_FREQUENCY, 0.5f);
274 lock_guard<mutex> lock(compressor_mutex);
276 // Apply a level compressor to get the general level right.
277 // Basically, if it's over about -40 dBFS, we squeeze it down to that level
278 // (or more precisely, near it, since we don't use infinite ratio),
279 // then apply a makeup gain to get it to -14 dBFS. -14 dBFS is, of course,
280 // entirely arbitrary, but from practical tests with speech, it seems to
281 // put ut around -23 LUFS, so it's a reasonable starting point for later use.
283 if (level_compressor_enabled) {
284 float threshold = 0.01f; // -40 dBFS.
286 float attack_time = 0.5f;
287 float release_time = 20.0f;
288 float makeup_gain = from_db(ref_level_dbfs - (-40.0f)); // +26 dB.
289 level_compressor.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
290 gain_staging_db = to_db(level_compressor.get_attenuation() * makeup_gain);
292 // Just apply the gain we already had.
293 float g = from_db(gain_staging_db);
294 for (size_t i = 0; i < samples_out.size(); ++i) {
301 printf("level=%f (%+5.2f dBFS) attenuation=%f (%+5.2f dB) end_result=%+5.2f dB\n",
302 level_compressor.get_level(), to_db(level_compressor.get_level()),
303 level_compressor.get_attenuation(), to_db(level_compressor.get_attenuation()),
304 to_db(level_compressor.get_level() * level_compressor.get_attenuation() * makeup_gain));
307 // float limiter_att, compressor_att;
309 // The real compressor.
310 if (compressor_enabled) {
311 float threshold = from_db(compressor_threshold_dbfs);
313 float attack_time = 0.005f;
314 float release_time = 0.040f;
315 float makeup_gain = 2.0f; // +6 dB.
316 compressor.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
317 // compressor_att = compressor.get_attenuation();
320 // Finally a limiter at -4 dB (so, -10 dBFS) to take out the worst peaks only.
321 // Note that since ratio is not infinite, we could go slightly higher than this.
322 if (limiter_enabled) {
323 float threshold = from_db(limiter_threshold_dbfs);
325 float attack_time = 0.0f; // Instant.
326 float release_time = 0.020f;
327 float makeup_gain = 1.0f; // 0 dB.
328 limiter.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
329 // limiter_att = limiter.get_attenuation();
332 // printf("limiter=%+5.1f compressor=%+5.1f\n", to_db(limiter_att), to_db(compressor_att));
335 // At this point, we are most likely close to +0 LU, but all of our
336 // measurements have been on raw sample values, not R128 values.
337 // So we have a final makeup gain to get us to +0 LU; the gain
338 // adjustments required should be relatively small, and also, the
339 // offset shouldn't change much (only if the type of audio changes
340 // significantly). Thus, we shoot for updating this value basically
341 // “whenever we process buffers”, since the R128 calculation isn't exactly
342 // something we get out per-sample.
344 // Note that there's a feedback loop here, so we choose a very slow filter
345 // (half-time of 30 seconds).
346 double target_loudness_factor, alpha;
347 double loudness_lu = loudness_lufs - ref_level_lufs;
348 double current_makeup_lu = to_db(final_makeup_gain);
349 target_loudness_factor = final_makeup_gain * from_db(-loudness_lu);
351 // If we're outside +/- 5 LU uncorrected, we don't count it as
352 // a normal signal (probably silence) and don't change the
353 // correction factor; just apply what we already have.
354 if (fabs(loudness_lu - current_makeup_lu) >= 5.0 || !final_makeup_gain_auto) {
357 // Formula adapted from
358 // https://en.wikipedia.org/wiki/Low-pass_filter#Simple_infinite_impulse_response_filter.
359 const double half_time_s = 30.0;
360 const double fc_mul_2pi_delta_t = 1.0 / (half_time_s * OUTPUT_FREQUENCY);
361 alpha = fc_mul_2pi_delta_t / (fc_mul_2pi_delta_t + 1.0);
365 lock_guard<mutex> lock(compressor_mutex);
366 double m = final_makeup_gain;
367 for (size_t i = 0; i < samples_out.size(); i += 2) {
368 samples_out[i + 0] *= m;
369 samples_out[i + 1] *= m;
370 m += (target_loudness_factor - m) * alpha;
372 final_makeup_gain = m;
378 vector<string> AudioMixer::get_names() const
380 lock_guard<mutex> lock(audio_mutex);
381 vector<string> names;
382 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
383 const AudioDevice *device = &cards[card_index];
384 names.push_back(device->name);
389 void AudioMixer::set_name(DeviceSpec device_spec, const string &name)
391 AudioDevice *device = find_audio_device(device_spec);
393 lock_guard<mutex> lock(audio_mutex);
397 void AudioMixer::set_input_mapping(const InputMapping &new_input_mapping)
399 lock_guard<mutex> lock(audio_mutex);
401 // FIXME: This needs to be keyed on DeviceSpec.
402 map<unsigned, set<unsigned>> interesting_channels;
403 for (const InputMapping::Bus &bus : new_input_mapping.buses) {
404 if (bus.device.type == InputSourceType::CAPTURE_CARD) {
405 for (unsigned channel = 0; channel < 2; ++channel) {
406 if (bus.source_channel[channel] != -1) {
407 interesting_channels[bus.device.index].insert(bus.source_channel[channel]);
413 // Reset resamplers for all cards that don't have the exact same state as before.
414 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
415 AudioDevice *device = &cards[card_index];
416 if (device->interesting_channels != interesting_channels[card_index]) {
417 device->interesting_channels = interesting_channels[card_index];
418 reset_device_mutex_held(DeviceSpec{InputSourceType::CAPTURE_CARD, card_index});
422 input_mapping = new_input_mapping;
425 InputMapping AudioMixer::get_input_mapping() const
427 lock_guard<mutex> lock(audio_mutex);
428 return input_mapping;