#include <stdio.h>
#include <endian.h>
#include <cmath>
+#ifdef __SSE__
+#include <immintrin.h>
+#endif
#include "db.h"
#include "flags.h"
}
}
-float find_peak(const float *samples, size_t num_samples)
+float find_peak_plain(const float *samples, size_t num_samples) __attribute__((unused));
+
+float find_peak_plain(const float *samples, size_t num_samples)
{
float m = fabs(samples[0]);
for (size_t i = 1; i < num_samples; ++i) {
return m;
}
+#ifdef __SSE__
+static inline float horizontal_max(__m128 m)
+{
+ __m128 tmp = _mm_shuffle_ps(m, m, _MM_SHUFFLE(1, 0, 3, 2));
+ m = _mm_max_ps(m, tmp);
+ tmp = _mm_shuffle_ps(m, m, _MM_SHUFFLE(2, 3, 0, 1));
+ m = _mm_max_ps(m, tmp);
+ return _mm_cvtss_f32(m);
+}
+
+float find_peak(const float *samples, size_t num_samples)
+{
+ const __m128 abs_mask = _mm_castsi128_ps(_mm_set1_epi32(0x7fffffffu));
+ __m128 m = _mm_setzero_ps();
+ for (size_t i = 0; i < (num_samples & ~3); i += 4) {
+ __m128 x = _mm_loadu_ps(samples + i);
+ x = _mm_and_ps(x, abs_mask);
+ m = _mm_max_ps(m, x);
+ }
+ float result = horizontal_max(m);
+
+ for (size_t i = (num_samples & ~3); i < num_samples; ++i) {
+ result = max(result, fabs(samples[i]));
+ }
+
+#if 0
+ // Self-test. We should be bit-exact the same.
+ float reference_result = find_peak_plain(samples, num_samples);
+ if (result != reference_result) {
+ fprintf(stderr, "Error: Peak is %f [%f %f %f %f]; should be %f.\n",
+ result,
+ _mm_cvtss_f32(_mm_shuffle_ps(m, m, _MM_SHUFFLE(0, 0, 0, 0))),
+ _mm_cvtss_f32(_mm_shuffle_ps(m, m, _MM_SHUFFLE(1, 1, 1, 1))),
+ _mm_cvtss_f32(_mm_shuffle_ps(m, m, _MM_SHUFFLE(2, 2, 2, 2))),
+ _mm_cvtss_f32(_mm_shuffle_ps(m, m, _MM_SHUFFLE(3, 3, 3, 3))),
+ reference_result);
+ abort();
+ }
+#endif
+ return result;
+}
+#else
+float find_peak(const float *samples, size_t num_samples)
+{
+ return find_peak_plain(samples, num_samples);
+}
+#endif
+
void deinterleave_samples(const vector<float> &in, vector<float> *out_l, vector<float> *out_r)
{
size_t num_samples = in.size() / 2;
for (unsigned bus_index = 0; bus_index < MAX_BUSES; ++bus_index) {
locut[bus_index].init(FILTER_HPF, 2);
locut_enabled[bus_index] = global_flags.locut_enabled;
+ eq[bus_index][EQ_BAND_BASS].init(FILTER_LOW_SHELF, 1);
+ // Note: EQ_BAND_MID isn't used (see comments in apply_eq()).
+ eq[bus_index][EQ_BAND_TREBLE].init(FILTER_HIGH_SHELF, 1);
+
gain_staging_db[bus_index] = global_flags.initial_gain_staging_db;
compressor[bus_index].reset(new StereoCompressor(OUTPUT_FREQUENCY));
compressor_threshold_dbfs[bus_index] = ref_level_dbfs - 12.0f; // -12 dB.
assert(num_channels > 0);
// Convert the audio to fp32.
- vector<float> audio;
- audio.resize(num_samples * num_channels);
+ unique_ptr<float[]> audio(new float[num_samples * num_channels]);
unsigned channel_index = 0;
for (auto channel_it = device->interesting_channels.cbegin(); channel_it != device->interesting_channels.end(); ++channel_it, ++channel_index) {
switch (audio_format.bits_per_sample) {
assert(num_samples == 0);
break;
case 16:
- convert_fixed16_to_fp32(&audio[0], channel_index, num_channels, data, *channel_it, audio_format.num_channels, num_samples);
+ convert_fixed16_to_fp32(audio.get(), channel_index, num_channels, data, *channel_it, audio_format.num_channels, num_samples);
break;
case 24:
- convert_fixed24_to_fp32(&audio[0], channel_index, num_channels, data, *channel_it, audio_format.num_channels, num_samples);
+ convert_fixed24_to_fp32(audio.get(), channel_index, num_channels, data, *channel_it, audio_format.num_channels, num_samples);
break;
case 32:
- convert_fixed32_to_fp32(&audio[0], channel_index, num_channels, data, *channel_it, audio_format.num_channels, num_samples);
+ convert_fixed32_to_fp32(audio.get(), channel_index, num_channels, data, *channel_it, audio_format.num_channels, num_samples);
break;
default:
fprintf(stderr, "Cannot handle audio with %u bits per sample\n", audio_format.bits_per_sample);
// Now add it.
int64_t local_pts = device->next_local_pts;
- device->resampling_queue->add_input_samples(local_pts / double(TIMEBASE), audio.data(), num_samples);
+ device->resampling_queue->add_input_samples(local_pts / double(TIMEBASE), audio.get(), num_samples);
device->next_local_pts = local_pts + frame_length;
return true;
}
samples_bus.resize(num_samples * 2);
for (unsigned bus_index = 0; bus_index < input_mapping.buses.size(); ++bus_index) {
fill_audio_bus(samples_card, input_mapping.buses[bus_index], num_samples, &samples_bus[0]);
-
- // Cut away everything under 120 Hz (or whatever the cutoff is);
- // we don't need it for voice, and it will reduce headroom
- // and confuse the compressor. (In particular, any hums at 50 or 60 Hz
- // should be dampened.)
- if (locut_enabled[bus_index]) {
- locut[bus_index].render(samples_bus.data(), samples_bus.size() / 2, locut_cutoff_hz * 2.0 * M_PI / OUTPUT_FREQUENCY, 0.5f);
- }
+ apply_eq(bus_index, &samples_bus);
{
lock_guard<mutex> lock(compressor_mutex);
}
}
- // TODO: We should measure post-fader.
+ add_bus_to_master(bus_index, samples_bus, &samples_out);
deinterleave_samples(samples_bus, &left, &right);
measure_bus_levels(bus_index, left, right);
-
- float volume = from_db(fader_volume_db[bus_index]);
- if (bus_index == 0) {
- for (unsigned i = 0; i < num_samples * 2; ++i) {
- samples_out[i] = samples_bus[i] * volume;
- }
- } else {
- for (unsigned i = 0; i < num_samples * 2; ++i) {
- samples_out[i] += samples_bus[i] * volume;
- }
- }
}
{
return samples_out;
}
+void AudioMixer::apply_eq(unsigned bus_index, vector<float> *samples_bus)
+{
+ constexpr float bass_freq_hz = 200.0f;
+ constexpr float treble_freq_hz = 4700.0f;
+
+ // Cut away everything under 120 Hz (or whatever the cutoff is);
+ // we don't need it for voice, and it will reduce headroom
+ // and confuse the compressor. (In particular, any hums at 50 or 60 Hz
+ // should be dampened.)
+ if (locut_enabled[bus_index]) {
+ locut[bus_index].render(samples_bus->data(), samples_bus->size() / 2, locut_cutoff_hz * 2.0 * M_PI / OUTPUT_FREQUENCY, 0.5f);
+ }
+
+ // Apply the rest of the EQ. Since we only have a simple three-band EQ,
+ // we can implement it with two shelf filters. We use a simple gain to
+ // set the mid-level filter, and then offset the low and high bands
+ // from that if we need to. (We could perhaps have folded the gain into
+ // the next part, but it's so cheap that the trouble isn't worth it.)
+ if (fabs(eq_level_db[bus_index][EQ_BAND_MID]) > 0.01f) {
+ float g = from_db(eq_level_db[bus_index][EQ_BAND_MID]);
+ for (size_t i = 0; i < samples_bus->size(); ++i) {
+ (*samples_bus)[i] *= g;
+ }
+ }
+
+ float bass_adj_db = eq_level_db[bus_index][EQ_BAND_BASS] - eq_level_db[bus_index][EQ_BAND_MID];
+ if (fabs(bass_adj_db) > 0.01f) {
+ eq[bus_index][EQ_BAND_BASS].render(samples_bus->data(), samples_bus->size() / 2,
+ bass_freq_hz * 2.0 * M_PI / OUTPUT_FREQUENCY, 0.5f, bass_adj_db / 40.0f);
+ }
+
+ float treble_adj_db = eq_level_db[bus_index][EQ_BAND_TREBLE] - eq_level_db[bus_index][EQ_BAND_MID];
+ if (fabs(treble_adj_db) > 0.01f) {
+ eq[bus_index][EQ_BAND_TREBLE].render(samples_bus->data(), samples_bus->size() / 2,
+ treble_freq_hz * 2.0 * M_PI / OUTPUT_FREQUENCY, 0.5f, treble_adj_db / 40.0f);
+ }
+}
+
+void AudioMixer::add_bus_to_master(unsigned bus_index, const vector<float> &samples_bus, vector<float> *samples_out)
+{
+ assert(samples_bus.size() == samples_out->size());
+ assert(samples_bus.size() % 2 == 0);
+ unsigned num_samples = samples_bus.size() / 2;
+ if (fabs(fader_volume_db[bus_index] - last_fader_volume_db[bus_index]) > 1e-3) {
+ // The volume has changed; do a fade over the course of this frame.
+ // (We might have some numerical issues here, but it seems to sound OK.)
+ // For the purpose of fading here, the silence floor is set to -90 dB
+ // (the fader only goes to -84).
+ float old_volume = from_db(max<float>(last_fader_volume_db[bus_index], -90.0f));
+ float volume = from_db(max<float>(fader_volume_db[bus_index], -90.0f));
+
+ float volume_inc = pow(volume / old_volume, 1.0 / num_samples);
+ volume = old_volume;
+ if (bus_index == 0) {
+ for (unsigned i = 0; i < num_samples; ++i) {
+ (*samples_out)[i * 2 + 0] = samples_bus[i * 2 + 0] * volume;
+ (*samples_out)[i * 2 + 1] = samples_bus[i * 2 + 1] * volume;
+ volume *= volume_inc;
+ }
+ } else {
+ for (unsigned i = 0; i < num_samples; ++i) {
+ (*samples_out)[i * 2 + 0] += samples_bus[i * 2 + 0] * volume;
+ (*samples_out)[i * 2 + 1] += samples_bus[i * 2 + 1] * volume;
+ volume *= volume_inc;
+ }
+ }
+ } else {
+ float volume = from_db(fader_volume_db[bus_index]);
+ if (bus_index == 0) {
+ for (unsigned i = 0; i < num_samples; ++i) {
+ (*samples_out)[i * 2 + 0] = samples_bus[i * 2 + 0] * volume;
+ (*samples_out)[i * 2 + 1] = samples_bus[i * 2 + 1] * volume;
+ }
+ } else {
+ for (unsigned i = 0; i < num_samples; ++i) {
+ (*samples_out)[i * 2 + 0] += samples_bus[i * 2 + 0] * volume;
+ (*samples_out)[i * 2 + 1] += samples_bus[i * 2 + 1] * volume;
+ }
+ }
+ }
+
+ last_fader_volume_db[bus_index] = fader_volume_db[bus_index];
+}
+
void AudioMixer::measure_bus_levels(unsigned bus_index, const vector<float> &left, const vector<float> &right)
{
- const float *ptrs[] = { left.data(), right.data() };
- {
- lock_guard<mutex> lock(audio_measure_mutex);
- bus_r128[bus_index]->process(left.size(), const_cast<float **>(ptrs));
+ assert(left.size() == right.size());
+ const float volume = from_db(fader_volume_db[bus_index]);
+ const float peak_levels[2] = {
+ find_peak(left.data(), left.size()) * volume,
+ find_peak(right.data(), right.size()) * volume
+ };
+ for (unsigned channel = 0; channel < 2; ++channel) {
+ // Compute the current value, including hold and falloff.
+ // The constants are borrowed from zita-mu1 by Fons Adriaensen.
+ static constexpr float hold_sec = 0.5f;
+ static constexpr float falloff_db_sec = 15.0f; // dB/sec falloff after hold.
+ float current_peak;
+ PeakHistory &history = peak_history[bus_index][channel];
+ history.historic_peak = max(history.historic_peak, peak_levels[channel]);
+ if (history.age_seconds < hold_sec) {
+ current_peak = history.last_peak;
+ } else {
+ current_peak = history.last_peak * from_db(-falloff_db_sec * (history.age_seconds - hold_sec));
+ }
+
+ // See if we have a new peak to replace the old (possibly falling) one.
+ if (peak_levels[channel] > current_peak) {
+ history.last_peak = peak_levels[channel];
+ history.age_seconds = 0.0f; // Not 100% correct, but more than good enough given our frame sizes.
+ current_peak = peak_levels[channel];
+ } else {
+ history.age_seconds += float(left.size()) / OUTPUT_FREQUENCY;
+ }
+ history.current_level = peak_levels[channel];
+ history.current_peak = current_peak;
}
}
vector<BusLevel> bus_levels;
bus_levels.resize(input_mapping.buses.size());
- for (unsigned bus_index = 0; bus_index < bus_r128.size(); ++bus_index) {
- bus_levels[bus_index].loudness_lufs = bus_r128[bus_index]->loudness_S();
- bus_levels[bus_index].gain_staging_db = gain_staging_db[bus_index];
+ {
+ lock_guard<mutex> lock(compressor_mutex);
+ for (unsigned bus_index = 0; bus_index < bus_levels.size(); ++bus_index) {
+ bus_levels[bus_index].current_level_dbfs[0] = to_db(peak_history[bus_index][0].current_level);
+ bus_levels[bus_index].current_level_dbfs[1] = to_db(peak_history[bus_index][1].current_level);
+ bus_levels[bus_index].peak_level_dbfs[0] = to_db(peak_history[bus_index][0].current_peak);
+ bus_levels[bus_index].peak_level_dbfs[1] = to_db(peak_history[bus_index][1].current_peak);
+ bus_levels[bus_index].historic_peak_dbfs = to_db(
+ max(peak_history[bus_index][0].historic_peak,
+ peak_history[bus_index][1].historic_peak));
+ bus_levels[bus_index].gain_staging_db = gain_staging_db[bus_index];
+ if (compressor_enabled[bus_index]) {
+ bus_levels[bus_index].compressor_attenuation_db = -to_db(compressor[bus_index]->get_attenuation());
+ } else {
+ bus_levels[bus_index].compressor_attenuation_db = 0.0;
+ }
+ }
}
audio_level_callback(loudness_s, to_db(peak), bus_levels,
}
}
- {
- lock_guard<mutex> lock(audio_measure_mutex);
- bus_r128.resize(new_input_mapping.buses.size());
- for (unsigned bus_index = 0; bus_index < bus_r128.size(); ++bus_index) {
- if (bus_r128[bus_index] == nullptr) {
- bus_r128[bus_index].reset(new Ebu_r128_proc);
- }
- bus_r128[bus_index]->init(2, OUTPUT_FREQUENCY);
- }
- }
-
input_mapping = new_input_mapping;
}
lock_guard<timed_mutex> lock(audio_mutex);
return input_mapping;
}
+
+void AudioMixer::reset_peak(unsigned bus_index)
+{
+ lock_guard<timed_mutex> lock(audio_mutex);
+ for (unsigned channel = 0; channel < 2; ++channel) {
+ PeakHistory &history = peak_history[bus_index][channel];
+ history.current_level = 0.0f;
+ history.historic_peak = 0.0f;
+ history.current_peak = 0.0f;
+ history.last_peak = 0.0f;
+ history.age_seconds = 0.0f;
+ }
+}
projects / nageru / blobdiff