static inline opus_val16 tansig_approx(opus_val32 _x)
{ /* Q19 */
int i;
opus_val16 xx; /* Q11 */
/*double x, y; */
opus_val16 dy, yy; /* Q14 */
/*x = 1.9073e-06*_x; */
if (_x >= QCONST32(8, 19))
return QCONST32(1., 14);
if (_x <= -QCONST32(8, 19))
return -QCONST32(1., 14);
xx = EXTRACT16(SHR32(_x, 8));
/*i = lrint(25*x); */
i = SHR32(ADD32(1024, MULT16_16(25, xx)), 11);
/*x -= .04*i; */
xx -= EXTRACT16(SHR32(MULT16_16(20972, i), 8));
/*x = xx*(1./2048); */
/*y = tansig_table[250+i]; */
yy = tansig_table[250 + i];
/*y = yy*(1./16384); */
dy = 16384 - MULT16_16_Q14(yy, yy);
yy = yy + MULT16_16_Q14(MULT16_16_Q11(xx, dy),
(16384 - MULT16_16_Q11(yy, xx)));
return yy;
}
void split_cb_search_shape_sign(
spx_sig_t target[], /* target vector */
spx_coef_t ak[], /* LPCs for this subframe */
spx_coef_t awk1[], /* Weighted LPCs for this subframe */
spx_coef_t awk2[], /* Weighted LPCs for this subframe */
const void *par, /* Codebook/search parameters*/
int p, /* number of LPC coeffs */
int nsf, /* number of samples in subframe */
spx_sig_t *exc,
spx_sig_t *r,
SpeexBits *bits,
char *stack,
int complexity
)
{
int i,j,k,m,n,q;
spx_word16_t *resp;
#ifdef _USE_SSE
__m128 *resp2;
__m128 *E;
#else
spx_word16_t *resp2;
spx_word32_t *E;
#endif
spx_word16_t *t;
spx_sig_t *e, *r2;
spx_word16_t *tmp;
spx_word32_t *ndist, *odist;
int *itmp;
spx_word16_t **ot, **nt;
int **nind, **oind;
int *ind;
const signed char *shape_cb;
int shape_cb_size, subvect_size, nb_subvect;
split_cb_params *params;
int N=2;
int *best_index;
spx_word32_t *best_dist;
int have_sign;
N=complexity;
if (N>10)
N=10;
ot=PUSH(stack, N, spx_word16_t*);
nt=PUSH(stack, N, spx_word16_t*);
oind=PUSH(stack, N, int*);
nind=PUSH(stack, N, int*);
params = (split_cb_params *) par;
subvect_size = params->subvect_size;
nb_subvect = params->nb_subvect;
shape_cb_size = 1<<params->shape_bits;
shape_cb = params->shape_cb;
have_sign = params->have_sign;
resp = PUSH(stack, shape_cb_size*subvect_size, spx_word16_t);
#ifdef _USE_SSE
resp2 = PUSH(stack, (shape_cb_size*subvect_size)>>2, __m128);
E = PUSH(stack, shape_cb_size>>2, __m128);
#else
resp2 = resp;
E = PUSH(stack, shape_cb_size, spx_word32_t);
#endif
t = PUSH(stack, nsf, spx_word16_t);
e = PUSH(stack, nsf, spx_sig_t);
r2 = PUSH(stack, nsf, spx_sig_t);
ind = PUSH(stack, nb_subvect, int);
tmp = PUSH(stack, 2*N*nsf, spx_word16_t);
for (i=0;i<N;i++)
{
ot[i]=tmp;
tmp += nsf;
nt[i]=tmp;
tmp += nsf;
}
best_index = PUSH(stack, N, int);
best_dist = PUSH(stack, N, spx_word32_t);
ndist = PUSH(stack, N, spx_word32_t);
odist = PUSH(stack, N, spx_word32_t);
itmp = PUSH(stack, 2*N*nb_subvect, int);
for (i=0;i<N;i++)
{
nind[i]=itmp;
itmp+=nb_subvect;
oind[i]=itmp;
itmp+=nb_subvect;
for (j=0;j<nb_subvect;j++)
nind[i][j]=oind[i][j]=-1;
}
/* FIXME: make that adaptive? */
for (i=0;i<nsf;i++)
t[i]=SHR(target[i],6);
for (j=0;j<N;j++)
for (i=0;i<nsf;i++)
ot[j][i]=t[i];
/*for (i=0;i<nsf;i++)
printf ("%d\n", (int)t[i]);*/
/* Pre-compute codewords response and energy */
compute_weighted_codebook(shape_cb, r, resp, resp2, E, shape_cb_size, subvect_size, stack);
for (j=0;j<N;j++)
odist[j]=0;
/*For all subvectors*/
for (i=0;i<nb_subvect;i++)
{
/*"erase" nbest list*/
for (j=0;j<N;j++)
ndist[j]=-2;
/*For all n-bests of previous subvector*/
for (j=0;j<N;j++)
{
spx_word16_t *x=ot[j]+subvect_size*i;
/*Find new n-best based on previous n-best j*/
if (have_sign)
vq_nbest_sign(x, resp2, subvect_size, shape_cb_size, E, N, best_index, best_dist, stack);
else
vq_nbest(x, resp2, subvect_size, shape_cb_size, E, N, best_index, best_dist, stack);
/*For all new n-bests*/
for (k=0;k<N;k++)
{
spx_word16_t *ct;
spx_word32_t err=0;
ct = ot[j];
/*update target*/
/*previous target*/
for (m=i*subvect_size;m<(i+1)*subvect_size;m++)
t[m]=ct[m];
/* New code: update only enough of the target to calculate error*/
{
int rind;
spx_word16_t *res;
spx_word16_t sign=1;
rind = best_index[k];
if (rind>=shape_cb_size)
{
sign=-1;
rind-=shape_cb_size;
}
res = resp+rind*subvect_size;
if (sign>0)
for (m=0;m<subvect_size;m++)
t[subvect_size*i+m] -= res[m];
else
for (m=0;m<subvect_size;m++)
t[subvect_size*i+m] += res[m];
}
/*compute error (distance)*/
err=odist[j];
for (m=i*subvect_size;m<(i+1)*subvect_size;m++)
err += t[m]*t[m];
/*update n-best list*/
if (err<ndist[N-1] || ndist[N-1]<-1)
{
/*previous target (we don't care what happened before*/
for (m=(i+1)*subvect_size;m<nsf;m++)
t[m]=ct[m];
/* New code: update the rest of the target only if it's worth it */
for (m=0;m<subvect_size;m++)
{
spx_word16_t g;
int rind;
spx_word16_t sign=1;
rind = best_index[k];
if (rind>=shape_cb_size)
{
sign=-1;
rind-=shape_cb_size;
}
q=subvect_size-m;
#ifdef FIXED_POINT
g=sign*shape_cb[rind*subvect_size+m];
for (n=subvect_size*(i+1);n<nsf;n++,q++)
t[n] = SUB32(t[n],MULT16_16_Q11(g,r[q]));
#else
g=sign*0.03125*shape_cb[rind*subvect_size+m];
for (n=subvect_size*(i+1);n<nsf;n++,q++)
t[n] = SUB32(t[n],g*r[q]);
#endif
}
for (m=0;m<N;m++)
{
if (err < ndist[m] || ndist[m]<-1)
{
for (n=N-1;n>m;n--)
{
for (q=(i+1)*subvect_size;q<nsf;q++)
nt[n][q]=nt[n-1][q];
for (q=0;q<nb_subvect;q++)
nind[n][q]=nind[n-1][q];
ndist[n]=ndist[n-1];
}
for (q=(i+1)*subvect_size;q<nsf;q++)
nt[m][q]=t[q];
for (q=0;q<nb_subvect;q++)
nind[m][q]=oind[j][q];
nind[m][i]=best_index[k];
ndist[m]=err;
break;
}
}
}
}
if (i==0)
break;
}
/*update old-new data*/
/* just swap pointers instead of a long copy */
{
spx_word16_t **tmp2;
tmp2=ot;
ot=nt;
nt=tmp2;
}
for (j=0;j<N;j++)
for (m=0;m<nb_subvect;m++)
oind[j][m]=nind[j][m];
for (j=0;j<N;j++)
odist[j]=ndist[j];
}
/*save indices*/
for (i=0;i<nb_subvect;i++)
{
ind[i]=nind[0][i];
speex_bits_pack(bits,ind[i],params->shape_bits+have_sign);
}
/* Put everything back together */
for (i=0;i<nb_subvect;i++)
{
int rind;
spx_word16_t sign=1;
rind = ind[i];
if (rind>=shape_cb_size)
{
sign=-1;
rind-=shape_cb_size;
}
#ifdef FIXED_POINT
if (sign==1)
{
for (j=0;j<subvect_size;j++)
e[subvect_size*i+j]=SHL((spx_word32_t)shape_cb[rind*subvect_size+j],SIG_SHIFT-5);
} else {
for (j=0;j<subvect_size;j++)
e[subvect_size*i+j]=-SHL((spx_word32_t)shape_cb[rind*subvect_size+j],SIG_SHIFT-5);
}
#else
for (j=0;j<subvect_size;j++)
e[subvect_size*i+j]=sign*0.03125*shape_cb[rind*subvect_size+j];
#endif
}
/* Update excitation */
for (j=0;j<nsf;j++)
exc[j]+=e[j];
/* Update target */
syn_percep_zero(e, ak, awk1, awk2, r2, nsf,p, stack);
for (j=0;j<nsf;j++)
target[j]-=r2[j];
}