/* Compute the amplitude (sqrt energy) in each of the bands */
void compute_band_energies(const CELTMode *m, const celt_sig *X, celt_ener *bandE, int end, int C, int M)
{
int i, c, N;
const opus_int16 *eBands = m->eBands;
N = M*m->shortMdctSize;
c=0; do {
for (i=0;i<end;i++)
{
int j;
opus_val32 maxval=0;
opus_val32 sum = 0;
j=M*eBands[i]; do {
maxval = MAX32(maxval, X[j+c*N]);
maxval = MAX32(maxval, -X[j+c*N]);
} while (++j<M*eBands[i+1]);
if (maxval > 0)
{
int shift = celt_ilog2(maxval)-10;
j=M*eBands[i]; do {
sum = MAC16_16(sum, EXTRACT16(VSHR32(X[j+c*N],shift)),
EXTRACT16(VSHR32(X[j+c*N],shift)));
} while (++j<M*eBands[i+1]);
/* We're adding one here to make damn sure we never end up with a pitch vector that's
larger than unity norm */
bandE[i+c*m->nbEBands] = EPSILON+VSHR32(EXTEND32(celt_sqrt(sum)),-shift);
} else {
bandE[i+c*m->nbEBands] = EPSILON;
}
/*printf ("%f ", bandE[i+c*m->nbEBands]);*/
}
} while (++c<C);
/*printf ("\n");*/
}
/*
*
* Allocates all necessary storage space for the fft and ifft.
* The return value is a contiguous block of memory. As such,
* It can be freed with free().
* */
kiss_fft_state *opus_fft_alloc_twiddles(int nfft,void * mem,size_t * lenmem,
const kiss_fft_state *base, int arch)
{
kiss_fft_state *st=NULL;
size_t memneeded = sizeof(struct kiss_fft_state); /* twiddle factors*/
if ( lenmem==NULL ) {
st = ( kiss_fft_state*)KISS_FFT_MALLOC( memneeded );
}else{
if (mem != NULL && *lenmem >= memneeded)
st = (kiss_fft_state*)mem;
*lenmem = memneeded;
}
if (st) {
opus_int16 *bitrev;
kiss_twiddle_cpx *twiddles;
st->nfft=nfft;
#ifdef FIXED_POINT
st->scale_shift = celt_ilog2(st->nfft);
if (st->nfft == 1<<st->scale_shift)
st->scale = Q15ONE;
else
st->scale = (1073741824+st->nfft/2)/st->nfft>>(15-st->scale_shift);
#else
st->scale = 1.f/nfft;
#endif
if (base != NULL)
{
st->twiddles = base->twiddles;
st->shift = 0;
while (st->shift < 32 && nfft<<st->shift != base->nfft)
st->shift++;
if (st->shift>=32)
goto fail;
} else {
st->twiddles = twiddles = (kiss_twiddle_cpx*)KISS_FFT_MALLOC(sizeof(kiss_twiddle_cpx)*nfft);
compute_twiddles(twiddles, nfft);
st->shift = -1;
}
if (!kf_factor(nfft,st->factors))
{
goto fail;
}
/* bitrev */
st->bitrev = bitrev = (opus_int16*)KISS_FFT_MALLOC(sizeof(opus_int16)*nfft);
if (st->bitrev==NULL)
goto fail;
compute_bitrev_table(0, bitrev, 1,1, st->factors,st);
/* Initialize architecture specific fft parameters */
if (opus_fft_alloc_arch(st, arch))
goto fail;
}
return st;
fail:
opus_fft_free(st, arch);
return NULL;
}
static void find_best_pitch(opus_val32 *xcorr, opus_val16 *y, int len,
int max_pitch, int *best_pitch
#ifdef FIXED_POINT
, int yshift, opus_val32 maxcorr
#endif
)
{
int i, j;
opus_val32 Syy=1;
opus_val16 best_num[2];
opus_val32 best_den[2];
#ifdef FIXED_POINT
int xshift;
xshift = celt_ilog2(maxcorr)-14;
#endif
best_num[0] = -1;
best_num[1] = -1;
best_den[0] = 0;
best_den[1] = 0;
best_pitch[0] = 0;
best_pitch[1] = 1;
for (j=0;j<len;j++)
Syy = ADD32(Syy, SHR32(MULT16_16(y[j],y[j]), yshift));
for (i=0;i<max_pitch;i++)
{
if (xcorr[i]>0)
{
opus_val16 num;
opus_val32 xcorr16;
xcorr16 = EXTRACT16(VSHR32(xcorr[i], xshift));
#ifndef FIXED_POINT
/* Considering the range of xcorr16, this should avoid both underflows
and overflows (inf) when squaring xcorr16 */
xcorr16 *= 1e-12f;
#endif
num = MULT16_16_Q15(xcorr16,xcorr16);
if (MULT16_32_Q15(num,best_den[1]) > MULT16_32_Q15(best_num[1],Syy))
{
if (MULT16_32_Q15(num,best_den[0]) > MULT16_32_Q15(best_num[0],Syy))
{
best_num[1] = best_num[0];
best_den[1] = best_den[0];
best_pitch[1] = best_pitch[0];
best_num[0] = num;
best_den[0] = Syy;
best_pitch[0] = i;
} else {
best_num[1] = num;
best_den[1] = Syy;
best_pitch[1] = i;
}
}
}
Syy += SHR32(MULT16_16(y[i+len],y[i+len]),yshift) - SHR32(MULT16_16(y[i],y[i]),yshift);
Syy = MAX32(1, Syy);
}
}
static void find_best_pitch(opus_val32 *xcorr, opus_val16 *y, int len,
int max_pitch, int *best_pitch
#ifdef FIXED_POINT
, int yshift, opus_val32 maxcorr
#endif
)
{
int i, j;
opus_val32 Syy=1;
opus_val16 best_num[2];
opus_val32 best_den[2];
#ifdef FIXED_POINT
int xshift;
xshift = celt_ilog2(maxcorr)-14;
#endif
best_num[0] = -1;
best_num[1] = -1;
best_den[0] = 0;
best_den[1] = 0;
best_pitch[0] = 0;
best_pitch[1] = 1;
for (j=0;j<len;j++)
Syy = MAC16_16(Syy, y[j],y[j]);
for (i=0;i<max_pitch;i++)
{
if (xcorr[i]>0)
{
opus_val16 num;
opus_val32 xcorr16;
xcorr16 = EXTRACT16(VSHR32(xcorr[i], xshift));
num = MULT16_16_Q15(xcorr16,xcorr16);
if (MULT16_32_Q15(num,best_den[1]) > MULT16_32_Q15(best_num[1],Syy))
{
if (MULT16_32_Q15(num,best_den[0]) > MULT16_32_Q15(best_num[0],Syy))
{
best_num[1] = best_num[0];
best_den[1] = best_den[0];
best_pitch[1] = best_pitch[0];
best_num[0] = num;
best_den[0] = Syy;
best_pitch[0] = i;
} else {
best_num[1] = num;
best_den[1] = Syy;
best_pitch[1] = i;
}
}
}
Syy += SHR32(MULT16_16(y[i+len],y[i+len]),yshift) - SHR32(MULT16_16(y[i],y[i]),yshift);
Syy = MAX32(1, Syy);
}
}
void testilog2(void)
{
celt_word32 x;
for (x=1;x<=268435455;x+=127)
{
celt_word32 error = abs(celt_ilog2(x)-(int)floor(log2(x)));
if (error!=0)
{
print("celt_ilog2 failed: celt_ilog2(x)!=floor(log2(x)) (x = %d, error = %d)\n",x,error);
ret = 1;
}
}
}
void _celt_autocorr(
const opus_val16 *x, /* in: [0...n-1] samples x */
opus_val32 *ac, /* out: [0...lag-1] ac values */
const opus_val16 *window,
int overlap,
int lag,
int n
)
{
opus_val32 d;
int i;
VARDECL(opus_val16, xx);
SAVE_STACK;
ALLOC(xx, n, opus_val16);
celt_assert(n>0);
celt_assert(overlap>=0);
for (i=0;i<n;i++)
xx[i] = x[i];
for (i=0;i<overlap;i++)
{
xx[i] = MULT16_16_Q15(x[i],window[i]);
xx[n-i-1] = MULT16_16_Q15(x[n-i-1],window[i]);
}
#ifdef FIXED_POINT
{
opus_val32 ac0=0;
int shift;
for(i=0;i<n;i++)
ac0 += SHR32(MULT16_16(xx[i],xx[i]),9);
ac0 += 1+n;
shift = celt_ilog2(ac0)-30+10;
shift = (shift+1)/2;
for(i=0;i<n;i++)
xx[i] = VSHR32(xx[i], shift);
}
#endif
while (lag>=0)
{
for (i = lag, d = 0; i < n; i++)
d += xx[i] * xx[i-lag];
ac[lag] = d;
/*printf ("%f ", ac[lag]);*/
lag--;
}
/*printf ("\n");*/
ac[0] += 10;
RESTORE_STACK;
}
/* Compute the amplitude (sqrt energy) in each of the bands */
void compute_band_energies(const CELTMode *m, const celt_sig *X, celt_ener *bank, int _C)
{
int i, c, N;
const celt_int16 *eBands = m->eBands;
const int C = CHANNELS(_C);
N = FRAMESIZE(m);
for (c=0;c<C;c++)
{
for (i=0;i<m->nbEBands;i++)
{
int j;
celt_word32 maxval=0;
celt_word32 sum = 0;
j=eBands[i]; do {
maxval = MAX32(maxval, X[j+c*N]);
maxval = MAX32(maxval, -X[j+c*N]);
} while (++j<eBands[i+1]);
if (maxval > 0)
{
int shift = celt_ilog2(maxval)-10;
j=eBands[i]; do {
sum = MAC16_16(sum, EXTRACT16(VSHR32(X[j+c*N],shift)),
EXTRACT16(VSHR32(X[j+c*N],shift)));
} while (++j<eBands[i+1]);
/* We're adding one here to make damn sure we never end up with a pitch vector that's
larger than unity norm */
bank[i+c*m->nbEBands] = EPSILON+VSHR32(EXTEND32(celt_sqrt(sum)),-shift);
} else {
bank[i+c*m->nbEBands] = EPSILON;
}
/*printf ("%f ", bank[i+c*m->nbEBands]);*/
}
}
/*printf ("\n");*/
}
int _celt_autocorr(
const opus_val16 *x, /* in: [0...n-1] samples x */
opus_val32 *ac, /* out: [0...lag-1] ac values */
const opus_val16 *window,
int overlap,
int lag,
int n,
int arch
)
{
opus_val32 d;
int i, k;
int fastN=n-lag;
int shift;
const opus_val16 *xptr;
VARDECL(opus_val16, xx);
SAVE_STACK;
ALLOC(xx, n, opus_val16);
celt_assert(n>0);
celt_assert(overlap>=0);
if (overlap == 0)
{
xptr = x;
} else {
for (i=0;i<n;i++)
xx[i] = x[i];
for (i=0;i<overlap;i++)
{
xx[i] = MULT16_16_Q15(x[i],window[i]);
xx[n-i-1] = MULT16_16_Q15(x[n-i-1],window[i]);
}
xptr = xx;
}
shift=0;
#ifdef OPUS_FIXED_POINT
{
opus_val32 ac0;
ac0 = 1+(n<<7);
if (n&1) ac0 += SHR32(MULT16_16(xptr[0],xptr[0]),9);
for(i=(n&1);i<n;i+=2)
{
ac0 += SHR32(MULT16_16(xptr[i],xptr[i]),9);
ac0 += SHR32(MULT16_16(xptr[i+1],xptr[i+1]),9);
}
shift = celt_ilog2(ac0)-30+10;
shift = (shift)/2;
if (shift>0)
{
for(i=0;i<n;i++)
xx[i] = PSHR32(xptr[i], shift);
xptr = xx;
} else
shift = 0;
}
#endif
celt_pitch_xcorr(xptr, xptr, ac, fastN, lag+1, arch);
for (k=0;k<=lag;k++)
{
for (i = k+fastN, d = 0; i < n; i++)
d = MAC16_16(d, xptr[i], xptr[i-k]);
ac[k] += d;
}
#ifdef OPUS_FIXED_POINT
shift = 2*shift;
if (shift<=0)
ac[0] += SHL32((opus_int32)1, -shift);
if (ac[0] < 268435456)
{
int shift2 = 29 - EC_ILOG(ac[0]);
for (i=0;i<=lag;i++)
ac[i] = SHL32(ac[i], shift2);
shift -= shift2;
} else if (ac[0] >= 536870912)
{
int shift2=1;
if (ac[0] >= 1073741824)
shift2++;
for (i=0;i<=lag;i++)
ac[i] = SHR32(ac[i], shift2);
shift += shift2;
}
#endif
RESTORE_STACK;
return shift;
}
/** Takes the pitch vector and the decoded residual vector, computes the gain
that will give ||p+g*y||=1 and mixes the residual with the pitch. */
static void mix_pitch_and_residual(int * restrict iy, celt_norm_t * restrict X, int N, int K, const celt_norm_t * restrict P)
{
int i;
celt_word32_t Ryp, Ryy, Rpp;
celt_word16_t ryp, ryy, rpp;
celt_word32_t g;
VARDECL(celt_norm_t, y);
#ifdef FIXED_POINT
int yshift;
#endif
SAVE_STACK;
#ifdef FIXED_POINT
yshift = 13-celt_ilog2(K);
#endif
ALLOC(y, N, celt_norm_t);
Rpp = 0;
i=0;
do {
Rpp = MAC16_16(Rpp,P[i],P[i]);
y[i] = SHL16(iy[i],yshift);
} while (++i < N);
Ryp = 0;
Ryy = 0;
/* If this doesn't generate a dual MAC (on supported archs), fire the compiler guy */
i=0;
do {
