ref: 877b1975b60e2aee2a75dcc4cfc0e66b3405a416
parent: 2aaa0fee93a4a507d30c18b6b7ace9be07401db7
author: Jean-Marc Valin <[email protected]>
date: Fri Feb 29 11:40:39 EST 2008
fixed-point: converted intra prediction and folding, unb0rked mixed-precision
--- a/libcelt/_kiss_fft_guts.h
+++ b/libcelt/_kiss_fft_guts.h
@@ -149,6 +149,8 @@
#else /* not FIXED_POINT*/
+#define EXT32(a) (a)
+
# define S_MUL(a,b) ( (a)*(b) )
#define C_MUL(m,a,b) \
do{ (m).r = (a).r*(b).r - (a).i*(b).i;\--- a/libcelt/kiss_fft.c
+++ b/libcelt/kiss_fft.c
@@ -23,6 +23,7 @@
#include "_kiss_fft_guts.h"
#include "arch.h"
#include "os_support.h"
+#include "mathops.h"
/* The guts header contains all the multiplication and addition macros that are defined for
fixed or floating point complex numbers. It also delares the kf_ internal functions.
--- a/libcelt/mdct.c
+++ b/libcelt/mdct.c
@@ -51,6 +51,7 @@
#include <math.h>
#include "os_support.h"
#include "_kiss_fft_guts.h"
+#include "mathops.h"
#ifndef M_PI
#define M_PI 3.141592653
--- a/libcelt/vq.c
+++ b/libcelt/vq.c
@@ -314,7 +314,7 @@
#if 0
if (0) {- float err=0;
+ celt_word32_t err=0;
for (i=0;i<N;i++)
err += (x[i]-nbest[0]->gain*y[0][i])*(x[i]-nbest[0]->gain*y[0][i]);
/*if (N<=10)
@@ -324,7 +324,7 @@
if (0) {for (i=0;i<N;i++)
x[i] = p[i]+nbest[0]->gain*y[0][i];
- float E=1e-15;
+ celt_word32_t E=1e-15;
int ABS = 0;
for (i=0;i<N;i++)
ABS += abs(iy[0][i]);
@@ -367,10 +367,10 @@
{int i,j;
int best=0;
- float best_score=0;
- float s = 1;
+ celt_word32_t best_score=0;
+ celt_word16_t s = 1;
int sign;
- float E;
+ celt_word32_t E;
float pred_gain;
int max_pos = N0-N/B;
if (max_pos > 32)
@@ -379,14 +379,14 @@
for (i=0;i<max_pos*B;i+=B)
{int j;
- float xy=0, yy=0;
+ celt_word32_t xy=0, yy=0;
float score;
for (j=0;j<N;j++)
{- xy += 1.f*x[j]*Y[i+N-j-1];
- yy += 1.f*Y[i+N-j-1]*Y[i+N-j-1];
+ xy = MAC16_16(xy, x[j], Y[i+N-j-1]);
+ yy = MAC16_16(yy, Y[i+N-j-1], Y[i+N-j-1]);
}
- score = xy*xy/(.001+yy);
+ score = 1.f*xy*xy/(.001+yy);
if (score > best_score)
{best_score = score;
@@ -414,11 +414,11 @@
for (j=0;j<N;j++)
{P[j] = s*Y[best+N-j-1];
- E += NORM_SCALING_1*NORM_SCALING_1*P[j]*P[j];
+ E = MAC16_16(E, P[j],P[j]);
}
- E = pred_gain/sqrt(E);
+ pred_gain = NORM_SCALING*pred_gain/sqrt(E);
for (j=0;j<N;j++)
- P[j] *= E;
+ P[j] *= pred_gain;
if (K>0)
{for (j=0;j<N;j++)
@@ -436,9 +436,9 @@
{int j;
int sign;
- float s;
+ celt_word16_t s;
int best;
- float E;
+ celt_word32_t E;
float pred_gain;
int max_pos = N0-N/B;
if (max_pos > 32)
@@ -461,11 +461,11 @@
for (j=0;j<N;j++)
{P[j] = s*Y[best+N-j-1];
- E += NORM_SCALING_1*NORM_SCALING_1*P[j]*P[j];
+ E = MAC16_16(E, P[j],P[j]);
}
- E = pred_gain/sqrt(E);
+ pred_gain = NORM_SCALING*pred_gain/sqrt(E);
for (j=0;j<N;j++)
- P[j] *= E;
+ P[j] *= pred_gain;
if (K==0)
{for (j=0;j<N;j++)
@@ -476,7 +476,8 @@
void intra_fold(celt_norm_t *x, int N, celt_norm_t *Y, celt_norm_t *P, int B, int N0, int Nmax)
{int i, j;
- float E;
+ celt_word32_t E;
+ float g;
E = 1e-10;
if (N0 >= Nmax/2)
@@ -486,7 +487,7 @@
for (j=0;j<N/B;j++)
{P[j*B+i] = Y[(Nmax-N0-j-1)*B+i];
- E += NORM_SCALING_1*NORM_SCALING_1*P[j*B+i]*P[j*B+i];
+ E += P[j*B+i]*P[j*B+i];
}
}
} else {@@ -493,12 +494,12 @@
for (j=0;j<N;j++)
{P[j] = Y[j];
- E += NORM_SCALING_1*NORM_SCALING_1*P[j]*P[j];
+ E = MAC16_16(E, P[j],P[j]);
}
}
- E = 1.f/sqrt(E);
+ g = NORM_SCALING/sqrt(E);
for (j=0;j<N;j++)
- P[j] *= E;
+ P[j] *= g;
for (j=0;j<N;j++)
x[j] = P[j];
}