c-gaborator/cgaborator.cpp

268 lines
8.6 KiB
C++

#include <iostream>
#include "include/cgaborator.h"
#include "gaborator.h"
#include <cmath>
#include <memory>
#ifdef __AVX2__
#include <immintrin.h>
#endif
class Gaborator {
public:
Gaborator(int block_size, double sampleRate, int bandsPerOctave, double minimumFrequency, double referenceFrequency, double maximumFrequency, int stepSize) :
parameters(bandsPerOctave, minimumFrequency / sampleRate, referenceFrequency / sampleRate, 1.0, 1e-5),
analyzer(parameters),
coefs(analyzer),
frequencyBinTimeStepSize(stepSize),
sample_rate((int) sampleRate),
blockSize(block_size)
{
//converts frequency (ff_max) in hertz to the number of bands above the min frequency
//the ceil is used to end up at a full band
int interesting_bands = ceil(bandsPerOctave * log(maximumFrequency/minimumFrequency)/log(2.0f));
//since bands are ordered from high to low we are only interested in lower bands:
//fs/2.0 is the nyquist frequency
int total_bands = ceil(bandsPerOctave * log(sampleRate/2.0/minimumFrequency)/log(2.0f));
latency = (int64_t) ceil(analyzer.analysis_support());
min_band = total_bands - interesting_bands;
t_in = 0;
int max_band = analyzer.bandpass_bands_end();
std::vector<float> bandcenterCache(max_band);
for(int i = 0 ; i < max_band ; i++){
if(i < min_band){
bandcenterCache[i] = -1;
}else{
bandcenterCache[i] = analyzer.band_ff(i) * sample_rate;
}
if (bandcenterCache[i] > 0) {
++numberOfBandsCache;
if (firstBandCache == -1) {
firstBandCache = i;
}
}
}
coefficientSize = (latency + 2*blockSize) / frequencyBinTimeStepSize;
//Allocate ring buffer and members in a contiguous array
coefficients = static_cast<float *>(calloc(coefficientSize * numberOfBandsCache, sizeof(float)));
assert(t_in == 0);
}
float* gaborTransform(float* audio_block, int64_t audio_block_length, size_t* return_size, size_t* slice_size) {
resultCache.clear();
if (audio_block == nullptr || audio_block_length == 0) { //finish
finish();
} else {
analyze(audio_block, audio_block_length);
}
if(!resultCache.empty()){
*return_size = resultCache.size();
*slice_size = numberOfBandsCache;
return resultCache.data();
} else{
*return_size = 0;
*slice_size = 0;
return nullptr;
}
}
int64_t analysisSupport() const {
return latency;
}
int numberOfBands() const {
return numberOfBandsCache;
}
~Gaborator() {
free(coefficients);
}
private:
void analyze(float* audio_block, int64_t audio_block_length){
analyzer.analyze(audio_block, t_in, t_in + audio_block_length, coefs);
int64_t st0 = t_in - latency;
int64_t st1 = t_in - latency + audio_block_length;
gaborApplySlice(st0, st1);
t_in += audio_block_length;
int64_t t_out = t_in - latency;
forget_before(analyzer, coefs, t_out - audio_block_length);
}
void finish(){
int64_t st0 = t_in - latency;
int64_t st1 = t_in;
//flush all till latency spot
gaborApplySlice(st0, st1);
//flush remaining
for (int i = 1; i < coefficientSize; ++i) {
float* currentCoefficient = &coefficients[((mostRecentCoefficentIndex + i) % coefficientSize) * numberOfBandsCache];
resultCache.insert(resultCache.end(), currentCoefficient, currentCoefficient + numberOfBandsCache);
// fill the oldest with zeros, but only the first round
if(i <= coefficientSize) {
std::fill(currentCoefficient, currentCoefficient + numberOfBandsCache, 0);
}
}
}
inline void gaborApplySlice(int64_t st0, int64_t st1) {
//Adjust start to match gaborProcessEntry requirements
if((st0 / frequencyBinTimeStepSize) <= 0){
st0 = frequencyBinTimeStepSize;
}
//Skip if nothing to process, the first results have a negative audio sample index
if(st0 > st1){
return;
}
int b0 = min_band;
int b1 = numberOfBandsCache + firstBandCache;
/*
Following code is equivalent, but it has been inlined for performance
gaborator::process([&](int band, int64_t audioSampleIndex, std::complex<float>& coef) {
gaborProcessEntry(band, audioSampleIndex, coef);
}, b0, b1, st0, st1, coefs);
*/
gaborator::apply_to_slice(false, [&](int band, int64_t sampleIndex, int time_step, unsigned len, const std::complex<float> *p0) {
//process magnitudes beforehand for easier auto-vectorization
if(magnitudeCache.size() < len){
magnitudeCache.resize(len);
}
#ifdef __AVX2__
int64_t i;
for (i = 0; i < (((int64_t)len) - 7); i += 8) {
// load 8 complex values (--> 16 floats overall) into two SIMD registers
__m256 inLo = _mm256_loadu_ps(reinterpret_cast<const float *> (p0 + i ));
__m256 inHi = _mm256_loadu_ps(reinterpret_cast<const float *> (p0 + i + 4));
// separates the real and imaginary part, however values are in a wrong order
__m256 re = _mm256_shuffle_ps (inLo, inHi, _MM_SHUFFLE(2, 0, 2, 0));
__m256 im = _mm256_shuffle_ps (inLo, inHi, _MM_SHUFFLE(3, 1, 3, 1));
// do the heavy work on the unordered vectors
__m256 abs = _mm256_sqrt_ps(_mm256_add_ps(_mm256_mul_ps(re, re), _mm256_mul_ps(im, im)));
// reorder values prior to storing
__m256d ordered = _mm256_permute4x64_pd (_mm256_castps_pd(abs), _MM_SHUFFLE(3, 1, 2, 0));
_mm256_storeu_ps(magnitudeCache.data() + i, _mm256_castpd_ps(ordered));
}
for (int64_t j = i; j < len; j++) {
#else
for (int64_t j = 0; j < len; j++) {
#endif
magnitudeCache[j] = std::abs(p0[j]);
}
int bandIndex = band - firstBandCache;
for (unsigned int j = 0; j < len; j++) {
gaborProcessEntry(bandIndex, (sampleIndex + time_step * j) / frequencyBinTimeStepSize, magnitudeCache[j]);
}
}, b0, b1, st0, st1, coefs);
}
inline void gaborProcessEntry(int bandIndex, int64_t coefficientIndex, float coefficient) {
float* currentCoefficient = &coefficients[(coefficientIndex % coefficientSize) * numberOfBandsCache];
// If a new index is reached, save the old (fixed) coefficients in the history
// Fill the array with zeros to get the max
if (coefficientIndex > mostRecentCoefficentIndex && coefficientIndex > coefficientSize) {
// keep the new maximum
mostRecentCoefficentIndex = coefficientIndex;
// "copy" the oldest data to the history
// the slice can be reused thanks to the oldest being filled with zeros just after
resultCache.insert(resultCache.end(), currentCoefficient, currentCoefficient + numberOfBandsCache);
// fill the oldest with zeros
std::fill(currentCoefficient, currentCoefficient + numberOfBandsCache, 0);
}
// due to reduction in precision (from audio sample accuracy to steps) multiple
// magnitudes could be placed in the same stepIndex, bandIndex pair.
// We take the maximum magnitudes value.
currentCoefficient[bandIndex] = std::max(currentCoefficient[bandIndex], coefficient);
}
private:
std::vector<float> resultCache;
//circular buffer with current coefficients
float* coefficients = nullptr;
int firstBandCache = -1;
int numberOfBandsCache = 0;
//The index of the most recent coefficient (in steps)
int64_t mostRecentCoefficentIndex = 0;
const int blockSize;
std::vector<float> magnitudeCache;
const int64_t frequencyBinTimeStepSize;
int64_t t_in;
int min_band;
const int sample_rate;
int64_t latency;
int64_t coefficientSize;
private:
const gaborator::parameters parameters;
gaborator::analyzer<float> analyzer;
gaborator::coefs<float> coefs;
};
uintptr_t gaborator_initialize(int blockSize, double sampleRate, int bandsPerOctave, double minimumFrequency, double referenceFrequency, double maximumFrequency, int stepSize){
return reinterpret_cast<uintptr_t>(new Gaborator(blockSize, sampleRate, bandsPerOctave, minimumFrequency, referenceFrequency,
maximumFrequency, stepSize));
}
int64_t gaborator_analysis_support(uintptr_t ptr) {
return reinterpret_cast<Gaborator*>(ptr)->analysisSupport();
}
int gaborator_number_of_bands(uintptr_t ptr) {
return reinterpret_cast<Gaborator*>(ptr)->numberOfBands();
}
float* gaborator_transform(uintptr_t ptr, float* audio_block, int64_t audio_block_length, size_t* return_size, size_t* slice_size){
return reinterpret_cast<Gaborator*>(ptr)->gaborTransform(audio_block, audio_block_length, return_size, slice_size);
}
void gaborator_release(uintptr_t ptr) {
auto g = reinterpret_cast<Gaborator*>(ptr);
delete g;
}