#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;
AudioMixer::AudioMixer(unsigned num_cards)
: num_cards(num_cards),
- level_compressor(OUTPUT_FREQUENCY),
limiter(OUTPUT_FREQUENCY),
- compressor(OUTPUT_FREQUENCY),
correlation(OUTPUT_FREQUENCY)
{
- locut.init(FILTER_HPF, 2);
-
- set_locut_enabled(global_flags.locut_enabled);
- set_gain_staging_db(global_flags.initial_gain_staging_db);
- set_gain_staging_auto(global_flags.gain_staging_auto);
- set_compressor_enabled(global_flags.compressor_enabled);
+ 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;
+ 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.
+ compressor_enabled[bus_index] = global_flags.compressor_enabled;
+ level_compressor[bus_index].reset(new StereoCompressor(OUTPUT_FREQUENCY));
+ level_compressor_enabled[bus_index] = global_flags.gain_staging_auto;
+ }
set_limiter_enabled(global_flags.limiter_enabled);
set_final_makeup_gain_auto(global_flags.final_makeup_gain_auto);
if (device->interesting_channels.empty()) {
device->alsa_device.reset();
} else {
- 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)));
+ const ALSAInput::Device &alsa_dev = available_alsa_cards[card_index];
+ 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)));
device->capture_frequency = device->alsa_device->get_sample_rate();
device->alsa_device->start_capture_thread();
}
}
}
- // TODO: Move lo-cut etc. into each bus.
vector<float> samples_out, left, right;
samples_out.resize(num_samples * 2);
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);
+ }
+
+ {
+ lock_guard<mutex> lock(compressor_mutex);
+
+ // Apply a level compressor to get the general level right.
+ // Basically, if it's over about -40 dBFS, we squeeze it down to that level
+ // (or more precisely, near it, since we don't use infinite ratio),
+ // then apply a makeup gain to get it to -14 dBFS. -14 dBFS is, of course,
+ // entirely arbitrary, but from practical tests with speech, it seems to
+ // put ut around -23 LUFS, so it's a reasonable starting point for later use.
+ if (level_compressor_enabled[bus_index]) {
+ float threshold = 0.01f; // -40 dBFS.
+ float ratio = 20.0f;
+ float attack_time = 0.5f;
+ float release_time = 20.0f;
+ float makeup_gain = from_db(ref_level_dbfs - (-40.0f)); // +26 dB.
+ level_compressor[bus_index]->process(samples_bus.data(), samples_bus.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
+ gain_staging_db[bus_index] = to_db(level_compressor[bus_index]->get_attenuation() * makeup_gain);
+ } else {
+ // Just apply the gain we already had.
+ float g = from_db(gain_staging_db[bus_index]);
+ for (size_t i = 0; i < samples_bus.size(); ++i) {
+ samples_bus[i] *= g;
+ }
+ }
+
+#if 0
+ printf("level=%f (%+5.2f dBFS) attenuation=%f (%+5.2f dB) end_result=%+5.2f dB\n",
+ level_compressor.get_level(), to_db(level_compressor.get_level()),
+ level_compressor.get_attenuation(), to_db(level_compressor.get_attenuation()),
+ to_db(level_compressor.get_level() * level_compressor.get_attenuation() * makeup_gain));
+#endif
+
+ // The real compressor.
+ if (compressor_enabled[bus_index]) {
+ float threshold = from_db(compressor_threshold_dbfs[bus_index]);
+ float ratio = 20.0f;
+ float attack_time = 0.005f;
+ float release_time = 0.040f;
+ float makeup_gain = 2.0f; // +6 dB.
+ compressor[bus_index]->process(samples_bus.data(), samples_bus.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
+ // compressor_att = compressor.get_attenuation();
+ }
+ }
+
// TODO: We should measure post-fader.
deinterleave_samples(samples_bus, &left, &right);
measure_bus_levels(bus_index, left, right);
}
}
- // 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) {
- locut.render(samples_out.data(), samples_out.size() / 2, locut_cutoff_hz * 2.0 * M_PI / OUTPUT_FREQUENCY, 0.5f);
- }
-
{
lock_guard<mutex> lock(compressor_mutex);
- // Apply a level compressor to get the general level right.
- // Basically, if it's over about -40 dBFS, we squeeze it down to that level
- // (or more precisely, near it, since we don't use infinite ratio),
- // then apply a makeup gain to get it to -14 dBFS. -14 dBFS is, of course,
- // entirely arbitrary, but from practical tests with speech, it seems to
- // put ut around -23 LUFS, so it's a reasonable starting point for later use.
- {
- if (level_compressor_enabled) {
- float threshold = 0.01f; // -40 dBFS.
- float ratio = 20.0f;
- float attack_time = 0.5f;
- float release_time = 20.0f;
- float makeup_gain = from_db(ref_level_dbfs - (-40.0f)); // +26 dB.
- level_compressor.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
- gain_staging_db = to_db(level_compressor.get_attenuation() * makeup_gain);
- } else {
- // Just apply the gain we already had.
- float g = from_db(gain_staging_db);
- for (size_t i = 0; i < samples_out.size(); ++i) {
- samples_out[i] *= g;
- }
- }
- }
-
- #if 0
- printf("level=%f (%+5.2f dBFS) attenuation=%f (%+5.2f dB) end_result=%+5.2f dB\n",
- level_compressor.get_level(), to_db(level_compressor.get_level()),
- level_compressor.get_attenuation(), to_db(level_compressor.get_attenuation()),
- to_db(level_compressor.get_level() * level_compressor.get_attenuation() * makeup_gain));
- #endif
-
- // float limiter_att, compressor_att;
-
- // The real compressor.
- if (compressor_enabled) {
- float threshold = from_db(compressor_threshold_dbfs);
- float ratio = 20.0f;
- float attack_time = 0.005f;
- float release_time = 0.040f;
- float makeup_gain = 2.0f; // +6 dB.
- compressor.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
- // compressor_att = compressor.get_attenuation();
- }
-
// Finally a limiter at -4 dB (so, -10 dBFS) to take out the worst peaks only.
// Note that since ratio is not infinite, we could go slightly higher than this.
if (limiter_enabled) {
// printf("limiter=%+5.1f compressor=%+5.1f\n", to_db(limiter_att), to_db(compressor_att));
}
- // At this point, we are most likely close to +0 LU, but all of our
+ // At this point, we are most likely close to +0 LU (at least if the
+ // faders sum to 0 dB and the compressors are on), but all of our
// measurements have been on raw sample values, not R128 values.
// So we have a final makeup gain to get us to +0 LU; the gain
// adjustments required should be relatively small, and also, the
double loudness_range_low = r128.range_min();
double loudness_range_high = r128.range_max();
- vector<float> bus_loudness;
- bus_loudness.resize(input_mapping.buses.size());
- for (unsigned bus_index = 0; bus_index < bus_r128.size(); ++bus_index) {
- bus_loudness[bus_index] = bus_r128[bus_index]->loudness_S();
+ vector<BusLevel> bus_levels;
+ bus_levels.resize(input_mapping.buses.size());
+ {
+ lock_guard<mutex> lock(compressor_mutex);
+ 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];
+ 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_loudness,
+ audio_level_callback(loudness_s, to_db(peak), bus_levels,
loudness_i, loudness_range_low, loudness_range_high,
- gain_staging_db,
to_db(final_makeup_gain),
correlation.get_correlation());
}
projects / nageru / blobdiff