Linear Predictor: Rearrange things somewhat
The original didn't really handle backwards versus forwards differently, as far as the predictor coefficients should have been, as they probably should have been reversed for a different direction window. This didn't fix my problem, though, but did possibly expose something else to mess with. Signed-off-by: Christopher Snowhill <kode54@gmail.com>CQTexperiment
Christopher Snowhill
parent
3274bc9fe7
commit
51e8223078
|
|
@ -19,186 +19,175 @@
|
|||
#include <stdbool.h>
|
||||
#include "lpc.h"
|
||||
|
||||
static void apply_window(float * const data, const size_t data_len);
|
||||
static void compute_autocorr(const float * const data, const size_t data_len, double * const autoc, const int m);
|
||||
static int compute_lpc(const double * const autoc, double * const lpc, const int lpc_order);
|
||||
static void lpc_extrapolate_data(float * const data0, const size_t data_len, const size_t extra, const double * const lpc, const int order, const bool invdir);
|
||||
|
||||
void lpc_extrapolate2(float * const data, const size_t data_len, const int nch, const int lpc_order, const size_t extra_bkwd, const size_t extra_fwd, void ** extrapolate_buffer, size_t * extrapolate_buffer_size)
|
||||
{
|
||||
const size_t tdata0_size = sizeof(float) * (extra_bkwd + data_len + extra_fwd);
|
||||
const size_t autoc_size = sizeof(double) * (lpc_order + 1);
|
||||
const size_t lpc_size = sizeof(double) * lpc_order;
|
||||
|
||||
const size_t new_size = tdata0_size + autoc_size + lpc_size;
|
||||
|
||||
if (new_size > *extrapolate_buffer_size)
|
||||
{
|
||||
*extrapolate_buffer = realloc(*extrapolate_buffer, new_size);
|
||||
*extrapolate_buffer_size = new_size;
|
||||
}
|
||||
|
||||
float* tdata0 = (float*)(*extrapolate_buffer); // for 1 channel only
|
||||
float* const tdata = tdata0 + extra_bkwd; // for 1 channel only
|
||||
|
||||
double* autoc = (double*)(*extrapolate_buffer + tdata0_size);
|
||||
double* lpc = (double*)(*extrapolate_buffer + tdata0_size + autoc_size);
|
||||
int max_order;
|
||||
|
||||
for(int c = 0; c < nch; c++)
|
||||
{
|
||||
for (int i = -(int)extra_bkwd; i < (int)(data_len+extra_fwd); i++) { tdata[i] = 0; } // should be removed after debugging etc
|
||||
|
||||
for (int i = 0; i < (int)data_len; i++)
|
||||
tdata[i] = data[i*nch + c];
|
||||
|
||||
apply_window(tdata, data_len);
|
||||
compute_autocorr(tdata, data_len, autoc, lpc_order);
|
||||
max_order = compute_lpc(autoc, lpc, lpc_order);
|
||||
|
||||
// restore after apply_window
|
||||
for (int i = 0; i < (int)data_len; i++)
|
||||
tdata[i] = data[i*nch + c];
|
||||
|
||||
if (extra_fwd)
|
||||
{
|
||||
lpc_extrapolate_data(tdata, data_len, extra_fwd, lpc, max_order, false);
|
||||
for (size_t i = data_len; i < (data_len+extra_fwd); i++)
|
||||
data[i*nch + c] = tdata[i];
|
||||
}
|
||||
if (extra_bkwd)
|
||||
{
|
||||
lpc_extrapolate_data(tdata, data_len, extra_bkwd, lpc, max_order, true);
|
||||
for (int i = -(int)extra_bkwd; i < 0; i++)
|
||||
data[i*nch + c] = tdata[i];
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
static void apply_window(float * const data, const size_t data_len)
|
||||
{
|
||||
static void apply_window(float *const data, const size_t data_len) {
|
||||
#if 0
|
||||
if (0) // subtract the mean
|
||||
{
|
||||
double mean = 0;
|
||||
for(int i = 0; i < (int)data_len; i++)
|
||||
mean += data[i];
|
||||
mean /= data_len;
|
||||
if (0) // subtract the mean
|
||||
{
|
||||
double mean = 0;
|
||||
for(int i = 0; i < (int)data_len; i++)
|
||||
mean += data[i];
|
||||
mean /= data_len;
|
||||
|
||||
for(int i = 0; i < (int)data_len; i++)
|
||||
data[i] -= (float)mean;
|
||||
}
|
||||
for(int i = 0; i < (int)data_len; i++)
|
||||
data[i] -= (float)mean;
|
||||
}
|
||||
#endif
|
||||
|
||||
if (1) // Welch window
|
||||
{
|
||||
const float n2 = (data_len+1)/2.0f;
|
||||
for(int i = 0; i < (int)data_len; i++)
|
||||
{
|
||||
float k = (i+1-n2)/n2;
|
||||
data[i] *= 1.0f - k*k;
|
||||
}
|
||||
}
|
||||
if(1) // Welch window
|
||||
{
|
||||
const float n2 = (data_len + 1) / 2.0f;
|
||||
for(int i = 0; i < (int)data_len; i++) {
|
||||
float k = (i + 1 - n2) / n2;
|
||||
data[data_len - 1 - i] *= 1.0f - k * k;
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
static void compute_autocorr(const float * const data, const size_t data_len, double * const autoc, const int m)
|
||||
{
|
||||
int i, j;
|
||||
static float vorbis_lpc_from_data(float *data, float *lpci, int n, int m, double *aut, double *lpc) {
|
||||
double error;
|
||||
double epsilon;
|
||||
int i, j;
|
||||
|
||||
j = m + 1;
|
||||
// for(j = 0; j <= m; j++)
|
||||
while(j--)
|
||||
{
|
||||
double d = 0;
|
||||
for(i = j; i < (int)data_len; i++)
|
||||
d += (double)data[i] * data[i-j];
|
||||
/* autocorrelation, p+1 lag coefficients */
|
||||
j = m + 1;
|
||||
while(j--) {
|
||||
double d = 0; /* double needed for accumulator depth */
|
||||
for(i = j; i < n; i++) d += (double)data[i] * data[i - j];
|
||||
aut[j] = d;
|
||||
}
|
||||
|
||||
autoc[j] = d;
|
||||
}
|
||||
/* Generate lpc coefficients from autocorr values */
|
||||
|
||||
/* set our noise floor to about -100dB */
|
||||
error = aut[0] * (1. + 1e-10);
|
||||
epsilon = 1e-9 * aut[0] + 1e-10;
|
||||
|
||||
for(i = 0; i < m; i++) {
|
||||
double r = -aut[i + 1];
|
||||
|
||||
if(error < epsilon) {
|
||||
memset(lpc + i, 0, (m - i) * sizeof(*lpc));
|
||||
goto done;
|
||||
}
|
||||
|
||||
/* Sum up this iteration's reflection coefficient; note that in
|
||||
Vorbis we don't save it. If anyone wants to recycle this code
|
||||
and needs reflection coefficients, save the results of 'r' from
|
||||
each iteration. */
|
||||
|
||||
for(j = 0; j < i; j++) r -= lpc[j] * aut[i - j];
|
||||
r /= error;
|
||||
|
||||
/* Update LPC coefficients and total error */
|
||||
|
||||
lpc[i] = r;
|
||||
for(j = 0; j < i / 2; j++) {
|
||||
double tmp = lpc[j];
|
||||
|
||||
lpc[j] += r * lpc[i - 1 - j];
|
||||
lpc[i - 1 - j] += r * tmp;
|
||||
}
|
||||
if(i & 1) lpc[j] += lpc[j] * r;
|
||||
|
||||
error *= 1. - r * r;
|
||||
}
|
||||
|
||||
done:
|
||||
|
||||
/* slightly damp the filter */
|
||||
{
|
||||
double g = .99;
|
||||
double damp = g;
|
||||
for(j = 0; j < m; j++) {
|
||||
lpc[j] *= damp;
|
||||
damp *= g;
|
||||
}
|
||||
}
|
||||
|
||||
for(j = 0; j < m; j++) lpci[j] = (float)lpc[j];
|
||||
|
||||
/* we need the error value to know how big an impulse to hit the
|
||||
filter with later */
|
||||
|
||||
return error;
|
||||
}
|
||||
|
||||
static int compute_lpc(const double * const autoc, double * const lpc, const int lpc_order)
|
||||
{
|
||||
int i, j;
|
||||
double error, epsilon;
|
||||
int max_order = lpc_order;
|
||||
static void vorbis_lpc_predict(float *coeff, float *prime, int m, float *data, long n, float *work) {
|
||||
/* in: coeff[0...m-1] LPC coefficients
|
||||
prime[0...m-1] initial values (allocated size of n+m-1)
|
||||
out: data[0...n-1] data samples */
|
||||
|
||||
error = autoc[0] * (1.+1e-10);
|
||||
epsilon = 1e-9*autoc[0] + 1e-10;
|
||||
long i, j, o, p;
|
||||
float y;
|
||||
|
||||
for(i = 0; i < lpc_order; i++)
|
||||
{
|
||||
if (error < epsilon)
|
||||
{
|
||||
memset(&lpc[i], 0, (lpc_order-i)*sizeof(lpc[0]));
|
||||
max_order = i; break;
|
||||
}
|
||||
if(!prime)
|
||||
for(i = 0; i < m; i++)
|
||||
work[i] = 0.f;
|
||||
else
|
||||
for(i = 0; i < m; i++)
|
||||
work[i] = prime[i];
|
||||
|
||||
double r = -autoc[i+1];
|
||||
for(j = 0; j < i; j++)
|
||||
r -= lpc[j] * autoc[i-j];
|
||||
r /= error;
|
||||
for(i = 0; i < n; i++) {
|
||||
y = 0;
|
||||
o = i;
|
||||
p = m;
|
||||
for(j = 0; j < m; j++)
|
||||
y -= work[o++] * coeff[--p];
|
||||
|
||||
lpc[i] = r;
|
||||
for(j = 0; j < i/2; j++)
|
||||
{
|
||||
double tmp = lpc[j];
|
||||
lpc[j ] += r * lpc[i-1-j];
|
||||
lpc[i-1-j] += r * tmp;
|
||||
}
|
||||
if (i&1)
|
||||
lpc[j] += lpc[j]*r;
|
||||
|
||||
error *= 1.0 - r*r;
|
||||
}
|
||||
|
||||
if (1) /* slightly damp the filter */
|
||||
{
|
||||
const double g = 0.999;
|
||||
double damp = g;
|
||||
for(j = 0; j < max_order; j++)
|
||||
{
|
||||
lpc[j] *= damp;
|
||||
damp *= g;
|
||||
}
|
||||
}
|
||||
|
||||
if (max_order == 0) /* in case the signal is constant AND we subtract the mean in apply_window() */
|
||||
{
|
||||
max_order = 1;
|
||||
lpc[0] = -1;
|
||||
}
|
||||
|
||||
return max_order;
|
||||
data[i] = work[o] = y;
|
||||
}
|
||||
}
|
||||
|
||||
static void lpc_extrapolate_data(float * const data0, const size_t data_len, const size_t extra, const double * const lpc, const int order, const bool invdir)
|
||||
{
|
||||
int i, j;
|
||||
if (invdir == false)
|
||||
{
|
||||
float* data = data0 + data_len - order;
|
||||
for(i = 0; i < (int)extra; i++)
|
||||
{
|
||||
float sum = 0;
|
||||
for(j = 0; j < order; j++)
|
||||
sum -= data[i+j] * (float)lpc[order-1-j];
|
||||
void lpc_extrapolate2(float *const data, const size_t data_len, const int nch, const int lpc_order, const size_t extra_bkwd, const size_t extra_fwd, void **extrapolate_buffer, size_t *extrapolate_buffer_size) {
|
||||
const size_t tdata_size = sizeof(float) * (extra_bkwd + data_len + extra_fwd);
|
||||
const size_t aut_size = sizeof(double) * (lpc_order + 1);
|
||||
const size_t lpc_size = sizeof(double) * lpc_order;
|
||||
const size_t lpci_size = sizeof(float) * lpc_order;
|
||||
const size_t work_size = sizeof(float) * (extra_bkwd + lpc_order + extra_fwd);
|
||||
|
||||
if (sum > 10.f) sum = 10.f; else if (sum < -10.f) sum = -10.f; // should be removed after debugging etc
|
||||
data[order+i] = sum;
|
||||
}
|
||||
}
|
||||
else
|
||||
{
|
||||
float* data = data0 - 1 + order;
|
||||
for(i = 0; i < (int)extra; i++)
|
||||
{
|
||||
float sum = 0;
|
||||
for(j = 0; j < order; j++)
|
||||
sum -= data[-i-j] * (float)lpc[order-1-j];
|
||||
const size_t new_size = tdata_size + aut_size + lpc_size + lpci_size + work_size;
|
||||
|
||||
if (sum > 10.f) sum = 10.f; else if (sum < -10.f) sum = -10.f; // should be removed after debugging etc
|
||||
data[-order-i] = sum;
|
||||
}
|
||||
}
|
||||
if(new_size > *extrapolate_buffer_size) {
|
||||
*extrapolate_buffer = realloc(*extrapolate_buffer, new_size);
|
||||
*extrapolate_buffer_size = new_size;
|
||||
}
|
||||
|
||||
float *tdata = (float *)(*extrapolate_buffer); // for 1 channel only
|
||||
|
||||
double *aut = (double *)(*extrapolate_buffer + tdata_size);
|
||||
double *lpc = (double *)(*extrapolate_buffer + tdata_size + aut_size);
|
||||
float *lpci = (float *)(*extrapolate_buffer + tdata_size + aut_size + lpc_size);
|
||||
float *work = (float *)(*extrapolate_buffer + tdata_size + aut_size + lpc_size + lpci_size);
|
||||
|
||||
for(int c = 0; c < nch; c++) {
|
||||
if(extra_bkwd) {
|
||||
for(int i = 0; i < (int)data_len; i++)
|
||||
tdata[data_len - 1 - i] = data[i * nch + c];
|
||||
} else {
|
||||
for(int i = 0; i < (int)data_len; i++)
|
||||
tdata[i] = data[i * nch + c];
|
||||
}
|
||||
|
||||
apply_window(tdata, data_len);
|
||||
vorbis_lpc_from_data(tdata, lpci, (int)data_len, lpc_order, aut, lpc);
|
||||
|
||||
// restore after apply_window
|
||||
if(extra_bkwd) {
|
||||
for(int i = 0; i < (int)data_len; i++)
|
||||
tdata[data_len - 1 - i] = data[i * nch + c];
|
||||
} else {
|
||||
for(int i = 0; i < (int)data_len; i++)
|
||||
tdata[i] = data[i * nch + c];
|
||||
}
|
||||
|
||||
vorbis_lpc_predict(lpci, tdata + data_len - lpc_order, lpc_order, tdata + data_len, extra_fwd + extra_bkwd, work);
|
||||
|
||||
if(extra_bkwd) {
|
||||
for(int i = 0; i < extra_bkwd; i++)
|
||||
data[(-i - 1) * nch + c] = tdata[data_len + i];
|
||||
} else {
|
||||
for(int i = 0; i < extra_fwd; i++)
|
||||
data[(i + data_len) * nch + c] = tdata[data_len + i];
|
||||
}
|
||||
}
|
||||
}
|
||||
|
|
|
|||
Loading…
Reference in New Issue