static inline spx_word32_t cheb_poly_eva(
spx_word16_t *coef, /* P or Q coefs in Q13 format */
spx_word16_t x, /* cos of freq (-1.0 to 1.0) in Q14 format */
int m, /* LPC order/2 */
char *stack
)
{
int i;
spx_word16_t b0, b1;
spx_word32_t sum;
/*Prevents overflows*/
if (x>16383)
x = 16383;
if (x<-16383)
x = -16383;
/* Initialise values */
b1=16384;
b0=x;
/* Evaluate Chebyshev series formulation usin g iterative approach */
sum = ADD32(EXTEND32(coef[m]), EXTEND32(MULT16_16_P14(coef[m-1],x)));
for(i=2;i<=m;i++)
{
spx_word16_t tmp=b0;
b0 = SUB16(MULT16_16_Q13(x,b0), b1);
b1 = tmp;
sum = ADD32(sum, EXTEND32(MULT16_16_P14(coef[m-i],b0)));
}
return sum;
}
void speex_decode_stereo_int(spx_int16_t *data, int frame_size, SpeexStereoState *_stereo)
{
int i;
spx_word32_t balance;
spx_word16_t e_left, e_right, e_ratio;
RealSpeexStereoState *stereo = (RealSpeexStereoState*)_stereo;
/* COMPATIBILITY_HACK(stereo); */
balance=stereo->balance;
e_ratio=stereo->e_ratio;
/* These two are Q14, with max value just below 2. */
e_right = DIV32(QCONST32(1., 22), spx_sqrt(MULT16_32_Q15(e_ratio, ADD32(QCONST32(1., 16), balance))));
e_left = SHR32(MULT16_16(spx_sqrt(balance), e_right), 8);
for (i=frame_size-1;i>=0;i--)
{
spx_int16_t tmp=data[i];
stereo->smooth_left = EXTRACT16(PSHR32(MAC16_16(MULT16_16(stereo->smooth_left, QCONST16(0.98, 15)), e_left, QCONST16(0.02, 15)), 15));
stereo->smooth_right = EXTRACT16(PSHR32(MAC16_16(MULT16_16(stereo->smooth_right, QCONST16(0.98, 15)), e_right, QCONST16(0.02, 15)), 15));
data[2*i] = (spx_int16_t)MULT16_16_P14(stereo->smooth_left, tmp);
data[2*i+1] = (spx_int16_t)MULT16_16_P14(stereo->smooth_right, tmp);
}
}
void lsp_interpolate(spx_lsp_t *old_lsp, spx_lsp_t *new_lsp, spx_lsp_t *interp_lsp, int len, int subframe, int nb_subframes)
{
int i;
spx_word16_t tmp = DIV32_16(SHL(1 + subframe,14),nb_subframes);
spx_word16_t tmp2 = 16384-tmp;
for (i=0;i<len;i++)
{
interp_lsp[i] = MULT16_16_P14(tmp2,old_lsp[i]) + MULT16_16_P14(tmp,new_lsp[i]);
}
}
static inline spx_word32_t cheb_poly_eva(spx_word32_t *coef,spx_word16_t x,int m,char *stack)
/* float coef[] coefficients of the polynomial to be evaluated */
/* float x the point where polynomial is to be evaluated */
/* int m order of the polynomial */
{
int i;
spx_word16_t *T;
spx_word32_t sum;
int m2=m>>1;
spx_word16_t *coefn;
/*Prevents overflows*/
if (x>16383)
x = 16383;
if (x<-16383)
x = -16383;
/* Allocate memory for Chebyshev series formulation */
T=PUSH(stack, m2+1, spx_word16_t);
coefn=PUSH(stack, m2+1, spx_word16_t);
for (i=0;i<m2+1;i++)
{
coefn[i] = coef[i];
/*printf ("%f ", coef[i]);*/
}
/*printf ("\n");*/
/* Initialise values */
T[0]=16384;
T[1]=x;
/* Evaluate Chebyshev series formulation using iterative approach */
/* Evaluate polynomial and return value also free memory space */
sum = ADD32(coefn[m2], MULT16_16_P14(coefn[m2-1],x));
/*x *= 2;*/
for(i=2;i<=m2;i++)
{
T[i] = MULT16_16_Q13(x,T[i-1]) - T[i-2];
sum = ADD32(sum, MULT16_16_P14(coefn[m2-i],T[i]));
/*printf ("%f ", sum);*/
}
/*printf ("\n");*/
return sum;
}
void lsp_interpolate(spx_lsp_t *old_lsp, spx_lsp_t *new_lsp, spx_lsp_t *lsp, int len, int subframe, int nb_subframes, spx_word16_t margin)
{
int i;
spx_word16_t m = margin;
spx_word16_t m2 = 25736-margin;
spx_word16_t tmp = DIV32_16(SHL32(EXTEND32(1 + subframe),14),nb_subframes);
spx_word16_t tmp2 = 16384-tmp;
for (i=0;i<len;i++)
lsp[i] = MULT16_16_P14(tmp2,old_lsp[i]) + MULT16_16_P14(tmp,new_lsp[i]);
/* Enforce margin to sure the LSPs are stable*/
if (lsp[0]<m)
lsp[0]=m;
if (lsp[len-1]>m2)
lsp[len-1]=m2;
for (i=1;i<len-1;i++)
{
if (lsp[i]<lsp[i-1]+m)
lsp[i]=lsp[i-1]+m;
if (lsp[i]>lsp[i+1]-m)
lsp[i]= SHR16(lsp[i],1) + SHR16(lsp[i+1]-m,1);
}
}
void lsp_interpolate(spx_lsp_t *old_lsp, spx_lsp_t *new_lsp, spx_lsp_t *interp_lsp, int len, int subframe, int nb_subframes)
{
#ifndef FIXED_LPC_SIZE
int i;
#endif
spx_word16_t tmp = DIV32_16(SHL32(EXTEND32(1 + subframe),14),nb_subframes);
spx_word16_t tmp2 = 16384-tmp;
#ifndef FIXED_LPC_SIZE
for (i=0;i<len;i++)
{
interp_lsp[i] = MULT16_16_P14(tmp2,old_lsp[i]) + MULT16_16_P14(tmp,new_lsp[i]);
}
#else
interp_lsp[0] = MULT16_16_P14(tmp2,old_lsp[0]) + MULT16_16_P14(tmp,new_lsp[0]);
interp_lsp[1] = MULT16_16_P14(tmp2,old_lsp[1]) + MULT16_16_P14(tmp,new_lsp[1]);
interp_lsp[2] = MULT16_16_P14(tmp2,old_lsp[2]) + MULT16_16_P14(tmp,new_lsp[2]);
interp_lsp[3] = MULT16_16_P14(tmp2,old_lsp[3]) + MULT16_16_P14(tmp,new_lsp[3]);
interp_lsp[4] = MULT16_16_P14(tmp2,old_lsp[4]) + MULT16_16_P14(tmp,new_lsp[4]);
interp_lsp[5] = MULT16_16_P14(tmp2,old_lsp[5]) + MULT16_16_P14(tmp,new_lsp[5]);
interp_lsp[6] = MULT16_16_P14(tmp2,old_lsp[6]) + MULT16_16_P14(tmp,new_lsp[6]);
interp_lsp[7] = MULT16_16_P14(tmp2,old_lsp[7]) + MULT16_16_P14(tmp,new_lsp[7]);
interp_lsp[8] = MULT16_16_P14(tmp2,old_lsp[8]) + MULT16_16_P14(tmp,new_lsp[8]);
interp_lsp[9] = MULT16_16_P14(tmp2,old_lsp[9]) + MULT16_16_P14(tmp,new_lsp[9]);
#endif
}
/** Finds the best quantized 3-tap pitch predictor by analysis by synthesis */
static spx_word32_t pitch_gain_search_3tap(
const spx_word16_t target[], /* Target vector */
const spx_coef_t ak[], /* LPCs for this subframe */
const spx_coef_t awk1[], /* Weighted LPCs #1 for this subframe */
const spx_coef_t awk2[], /* Weighted LPCs #2 for this subframe */
spx_sig_t exc[], /* Excitation */
const signed char* gain_cdbk,
int gain_cdbk_size,
int pitch, /* Pitch value */
int p, /* Number of LPC coeffs */
int nsf, /* Number of samples in subframe */
SpeexBits* bits,
char* stack,
const spx_word16_t* exc2,
const spx_word16_t* r,
spx_word16_t* new_target,
int* cdbk_index,
int plc_tuning,
spx_word32_t cumul_gain,
int scaledown
) {
int i, j;
VARDECL(spx_word16_t * tmp1);
VARDECL(spx_word16_t * e);
spx_word16_t* x[3];
spx_word32_t corr[3];
spx_word32_t A[3][3];
spx_word16_t gain[3];
spx_word32_t err;
spx_word16_t max_gain = 128;
int best_cdbk = 0;
ALLOC(tmp1, 3 * nsf, spx_word16_t);
ALLOC(e, nsf, spx_word16_t);
if (cumul_gain > 262144)
max_gain = 31;
x[0] = tmp1;
x[1] = tmp1 + nsf;
x[2] = tmp1 + 2 * nsf;
for (j = 0; j < nsf; j++)
new_target[j] = target[j];
{
VARDECL(spx_mem_t * mm);
int pp = pitch - 1;
ALLOC(mm, p, spx_mem_t);
for (j = 0; j < nsf; j++) {
if (j - pp < 0)
e[j] = exc2[j - pp];
else if (j - pp - pitch < 0)
e[j] = exc2[j - pp - pitch];
else
e[j] = 0;
}
#ifdef FIXED_POINT
/* Scale target and excitation down if needed (avoiding overflow) */
if (scaledown) {
for (j = 0; j < nsf; j++)
e[j] = SHR16(e[j], 1);
for (j = 0; j < nsf; j++)
new_target[j] = SHR16(new_target[j], 1);
}
#endif
for (j = 0; j < p; j++)
mm[j] = 0;
iir_mem16(e, ak, e, nsf, p, mm, stack);
for (j = 0; j < p; j++)
mm[j] = 0;
filter_mem16(e, awk1, awk2, e, nsf, p, mm, stack);
for (j = 0; j < nsf; j++)
x[2][j] = e[j];
}
for (i = 1; i >= 0; i--) {
spx_word16_t e0 = exc2[-pitch - 1 + i];
#ifdef FIXED_POINT
/* Scale excitation down if needed (avoiding overflow) */
if (scaledown)
e0 = SHR16(e0, 1);
#endif
x[i][0] = MULT16_16_Q14(r[0], e0);
for (j = 0; j < nsf - 1; j++)
x[i][j + 1] = ADD32(x[i + 1][j], MULT16_16_P14(r[j + 1], e0));
}
for (i = 0; i < 3; i++)
corr[i] = inner_prod(x[i], new_target, nsf);
for (i = 0; i < 3; i++)
for (j = 0; j <= i; j++)
A[i][j] = A[j][i] = inner_prod(x[i], x[j], nsf);
{
spx_word32_t C[9];
#ifdef FIXED_POINT
spx_word16_t C16[9];
#else
spx_word16_t* C16 = C;
#endif
C[0] = corr[2];
C[1] = corr[1];
C[2] = corr[0];
C[3] = A[1][2];
C[4] = A[0][1];
C[5] = A[0][2];
C[6] = A[2][2];
C[7] = A[1][1];
C[8] = A[0][0];
/*plc_tuning *= 2;*/
if (plc_tuning < 2)
plc_tuning = 2;
if (plc_tuning > 30)
plc_tuning = 30;
#ifdef FIXED_POINT
C[0] = SHL32(C[0], 1);
C[1] = SHL32(C[1], 1);
C[2] = SHL32(C[2], 1);
C[3] = SHL32(C[3], 1);
C[4] = SHL32(C[4], 1);
C[5] = SHL32(C[5], 1);
C[6] = MAC16_32_Q15(C[6], MULT16_16_16(plc_tuning, 655), C[6]);
C[7] = MAC16_32_Q15(C[7], MULT16_16_16(plc_tuning, 655), C[7]);
C[8] = MAC16_32_Q15(C[8], MULT16_16_16(plc_tuning, 655), C[8]);
normalize16(C, C16, 32767, 9);
#else
C[6] *= .5 * (1 + .02 * plc_tuning);
C[7] *= .5 * (1 + .02 * plc_tuning);
C[8] *= .5 * (1 + .02 * plc_tuning);
#endif
best_cdbk = pitch_gain_search_3tap_vq(gain_cdbk, gain_cdbk_size, C16, max_gain);
#ifdef FIXED_POINT
gain[0] = ADD16(32, (spx_word16_t)gain_cdbk[best_cdbk * 4]);
gain[1] = ADD16(32, (spx_word16_t)gain_cdbk[best_cdbk * 4 + 1]);
gain[2] = ADD16(32, (spx_word16_t)gain_cdbk[best_cdbk * 4 + 2]);
/*printf ("%d %d %d %d\n",gain[0],gain[1],gain[2], best_cdbk);*/
#else
gain[0] = 0.015625 * gain_cdbk[best_cdbk * 4] + .5;
gain[1] = 0.015625 * gain_cdbk[best_cdbk * 4 + 1] + .5;
gain[2] = 0.015625 * gain_cdbk[best_cdbk * 4 + 2] + .5;
#endif
*cdbk_index = best_cdbk;
}
SPEEX_MEMSET(exc, 0, nsf);
for (i = 0; i < 3; i++) {
int j;
int tmp1, tmp3;
int pp = pitch + 1 - i;
tmp1 = nsf;
if (tmp1 > pp)
tmp1 = pp;
for (j = 0; j < tmp1; j++)
exc[j] = MAC16_16(exc[j], SHL16(gain[2 - i], 7), exc2[j - pp]);
tmp3 = nsf;
if (tmp3 > pp + pitch)
tmp3 = pp + pitch;
for (j = tmp1; j < tmp3; j++)
exc[j] = MAC16_16(exc[j], SHL16(gain[2 - i], 7), exc2[j - pp - pitch]);
}
for (i = 0; i < nsf; i++) {
spx_word32_t tmp = ADD32(ADD32(MULT16_16(gain[0], x[2][i]), MULT16_16(gain[1], x[1][i])),
MULT16_16(gain[2], x[0][i]));
new_target[i] = SUB16(new_target[i], EXTRACT16(PSHR32(tmp, 6)));
}
err = inner_prod(new_target, new_target, nsf);
return err;
}
