static FLOAT8
calc_sfb_noise(const FLOAT8 * xr, const FLOAT8 * xr34, unsigned int bw, int sf)
{
unsigned int j;
fi_union fi;
FLOAT8 temp;
FLOAT8 xfsf = 0.0;
FLOAT8 const sfpow = POW20(sf); /*pow(2.0,sf/4.0); */
FLOAT8 const sfpow34 = IPOW20(sf); /*pow(sfpow,-3.0/4.0); */
for (j = 0; j < bw; ++j) {
temp = xr34[j] * sfpow34;
if (temp > IXMAX_VAL)
return -1;
#ifdef TAKEHIRO_IEEE754_HACK
temp += MAGIC_FLOAT;
fi.f = temp;
fi.f = temp + (adj43asm - MAGIC_INT)[fi.i];
fi.i -= MAGIC_INT;
#else
XRPOW_FTOI(temp, fi.i);
XRPOW_FTOI(temp + QUANTFAC(fi.i), fi.i);
#endif
temp = fabs(xr[j]) - pow43[fi.i] * sfpow;
xfsf += temp * temp;
}
return xfsf;
}
inline static void
k_34_4(DOUBLEX x[4], int l3[4])
{
#ifdef TAKEHIRO_IEEE754_HACK
fi_union fi[4];
assert(x[0] <= IXMAX_VAL && x[1] <= IXMAX_VAL && x[2] <= IXMAX_VAL && x[3] <= IXMAX_VAL);
x[0] += MAGIC_FLOAT;
fi[0].f = x[0];
x[1] += MAGIC_FLOAT;
fi[1].f = x[1];
x[2] += MAGIC_FLOAT;
fi[2].f = x[2];
x[3] += MAGIC_FLOAT;
fi[3].f = x[3];
fi[0].f = x[0] + adj43asm[fi[0].i - MAGIC_INT];
fi[1].f = x[1] + adj43asm[fi[1].i - MAGIC_INT];
fi[2].f = x[2] + adj43asm[fi[2].i - MAGIC_INT];
fi[3].f = x[3] + adj43asm[fi[3].i - MAGIC_INT];
l3[0] = fi[0].i - MAGIC_INT;
l3[1] = fi[1].i - MAGIC_INT;
l3[2] = fi[2].i - MAGIC_INT;
l3[3] = fi[3].i - MAGIC_INT;
#else
assert(x[0] <= IXMAX_VAL && x[1] <= IXMAX_VAL && x[2] <= IXMAX_VAL && x[3] <= IXMAX_VAL);
XRPOW_FTOI(x[0], l3[0]);
XRPOW_FTOI(x[1], l3[1]);
XRPOW_FTOI(x[2], l3[2]);
XRPOW_FTOI(x[3], l3[3]);
x[0] += QUANTFAC(l3[0]);
x[1] += QUANTFAC(l3[1]);
x[2] += QUANTFAC(l3[2]);
x[3] += QUANTFAC(l3[3]);
XRPOW_FTOI(x[0], l3[0]);
XRPOW_FTOI(x[1], l3[1]);
XRPOW_FTOI(x[2], l3[2]);
XRPOW_FTOI(x[3], l3[3]);
#endif
}
void quantize_xrpow(const FLOAT8 *xr, int *ix, FLOAT8 istep) {
/* quantize on xr^(3/4) instead of xr */
/* from [email protected]. Reported to be 2x faster than
the above code (when not using ASM) on PowerPC */
int j;
for ( j = 576/8; j > 0; --j) {
FLOAT8 x1, x2, x3, x4, x5, x6, x7, x8;
int rx1, rx2, rx3, rx4, rx5, rx6, rx7, rx8;
x1 = *xr++ * istep;
x2 = *xr++ * istep;
XRPOW_FTOI(x1, rx1);
x3 = *xr++ * istep;
XRPOW_FTOI(x2, rx2);
x4 = *xr++ * istep;
XRPOW_FTOI(x3, rx3);
x5 = *xr++ * istep;
XRPOW_FTOI(x4, rx4);
x6 = *xr++ * istep;
XRPOW_FTOI(x5, rx5);
x7 = *xr++ * istep;
XRPOW_FTOI(x6, rx6);
x8 = *xr++ * istep;
XRPOW_FTOI(x7, rx7);
x1 += QUANTFAC(rx1);
XRPOW_FTOI(x8, rx8);
x2 += QUANTFAC(rx2);
XRPOW_FTOI(x1,*ix++);
x3 += QUANTFAC(rx3);
XRPOW_FTOI(x2,*ix++);
x4 += QUANTFAC(rx4);
XRPOW_FTOI(x3,*ix++);
x5 += QUANTFAC(rx5);
XRPOW_FTOI(x4,*ix++);
x6 += QUANTFAC(rx6);
XRPOW_FTOI(x5,*ix++);
x7 += QUANTFAC(rx7);
XRPOW_FTOI(x6,*ix++);
x8 += QUANTFAC(rx8);
XRPOW_FTOI(x7,*ix++);
XRPOW_FTOI(x8,*ix++);
}
}
static FLOAT8
calc_sfb_noise_mq(const FLOAT8 * xr, const FLOAT8 * xr34, int bw, int sf, int mq, FLOAT8 * scratch)
{
int j, k;
fi_union fi;
FLOAT8 temp;
FLOAT8 xfsfm = 0.0, xfsf = 0.0;
FLOAT8 const sfpow = POW20(sf); /*pow(2.0,sf/4.0); */
FLOAT8 const sfpow34 = IPOW20(sf); /*pow(sfpow,-3.0/4.0); */
for (j = 0; j < bw; ++j) {
temp = xr34[j] * sfpow34;
if (temp > IXMAX_VAL)
return -1;
#ifdef TAKEHIRO_IEEE754_HACK
temp += MAGIC_FLOAT;
fi.f = temp;
fi.f = temp + (adj43asm - MAGIC_INT)[fi.i];
fi.i -= MAGIC_INT;
#else
XRPOW_FTOI(temp, fi.i);
XRPOW_FTOI(temp + QUANTFAC(fi.i), fi.i);
#endif
temp = fabs(xr[j]) - pow43[fi.i] * sfpow;
temp *= temp;
scratch[j] = temp;
if (xfsfm < temp)
xfsfm = temp;
xfsf += temp;
}
if (mq == 1)
return bw * select_kth(scratch, bw, bw * 13 / 16);
xfsf /= bw;
for (k = 1, j = 0; j < bw; ++j) {
if (scratch[j] > xfsf) {
xfsfm += scratch[j];
++k;
}
}
return xfsfm / k * bw;
}
static void Quantize(const double *xr, int *ix, double istep)
{
int j;
for (j = FRAME_LEN/8; j > 0; --j) {
double x1, x2, x3, x4, x5, x6, x7, x8;
int rx1, rx2, rx3, rx4, rx5, rx6, rx7, rx8;
x1 = *xr++ * istep;
x2 = *xr++ * istep;
XRPOW_FTOI(x1, rx1);
x3 = *xr++ * istep;
XRPOW_FTOI(x2, rx2);
x4 = *xr++ * istep;
XRPOW_FTOI(x3, rx3);
x5 = *xr++ * istep;
XRPOW_FTOI(x4, rx4);
x6 = *xr++ * istep;
XRPOW_FTOI(x5, rx5);
x7 = *xr++ * istep;
XRPOW_FTOI(x6, rx6);
x8 = *xr++ * istep;
XRPOW_FTOI(x7, rx7);
x1 += QUANTFAC(rx1);
XRPOW_FTOI(x8, rx8);
x2 += QUANTFAC(rx2);
XRPOW_FTOI(x1,*ix++);
x3 += QUANTFAC(rx3);
XRPOW_FTOI(x2,*ix++);
x4 += QUANTFAC(rx4);
XRPOW_FTOI(x3,*ix++);
x5 += QUANTFAC(rx5);
XRPOW_FTOI(x4,*ix++);
x6 += QUANTFAC(rx6);
XRPOW_FTOI(x5,*ix++);
x7 += QUANTFAC(rx7);
XRPOW_FTOI(x6,*ix++);
x8 += QUANTFAC(rx8);
XRPOW_FTOI(x7,*ix++);
XRPOW_FTOI(x8,*ix++);
}
}
static FLOAT8
calc_sfb_noise_ave(const FLOAT8 * xr, const FLOAT8 * xr34, int bw, int sf)
{
FLOAT8 xp;
FLOAT8 xe;
FLOAT8 xm;
#ifdef TAKEHIRO_IEEE754_HACK
FLOAT8 x0;
unsigned int index = 0;
#endif
int xx[3], j;
fi_union *fi = (fi_union *) xx;
FLOAT8 xfsf_eq = 0.0, xfsf_p1 = 0.0, xfsf_m1 = 0.0;
FLOAT8 const sfpow_eq = POW20(sf); /*pow(2.0,sf/4.0); */
FLOAT8 const sfpow_m1 = sfpow_eq * facm1;
FLOAT8 const sfpow_p1 = sfpow_eq * facp1;
FLOAT8 const sfpow34_eq = IPOW20(sf); /*pow(sfpow,-3.0/4.0); */
FLOAT8 const sfpow34_m1 = sfpow34_eq * fac34m1;
FLOAT8 const sfpow34_p1 = sfpow34_eq * fac34p1;
#ifdef TAKEHIRO_IEEE754_HACK
/*
* loop unrolled into "Duff's Device". Robert Hegemann
*/
j = (bw + 3) / 4;
switch (bw % 4) {
default:
case 0:
do {
DUFFBLOCK();
case 3:
DUFFBLOCK();
case 2:
DUFFBLOCK();
case 1:
DUFFBLOCK();
} while (--j);
}
#else
for (j = 0; j < bw; ++j) {
xm = xr34[j] * sfpow34_m1;
if (xm > IXMAX_VAL)
return -1;
XRPOW_FTOI(xm, fi[0].i);
XRPOW_FTOI(xm + QUANTFAC(fi[0].i), fi[0].i);
xm = fabs(xr[j]) - pow43[fi[0].i] * sfpow_m1;
xm *= xm;
xe = xr34[j] * sfpow34_eq;
XRPOW_FTOI(xe, fi[0].i);
XRPOW_FTOI(xe + QUANTFAC(fi[0].i), fi[0].i);
xe = fabs(xr[j]) - pow43[fi[0].i] * sfpow_eq;
xe *= xe;
xp = xr34[j] * sfpow34_p1;
XRPOW_FTOI(xp, fi[0].i);
XRPOW_FTOI(xp + QUANTFAC(fi[0].i), fi[0].i);
xp = fabs(xr[j]) - pow43[fi[0].i] * sfpow_p1;
xp *= xp;
xfsf_m1 += xm;
xfsf_eq += xe;
xfsf_p1 += xp;
}
#endif
if (xfsf_eq < xfsf_p1)
xfsf_eq = xfsf_p1;
if (xfsf_eq < xfsf_m1)
xfsf_eq = xfsf_m1;
return xfsf_eq;
}
