void split_cb_shape_sign_unquant(
spx_sig_t *exc,
const void *par, /* non-overlapping codebook */
int nsf, /* number of samples in subframe */
SpeexBits *bits,
char *stack,
spx_int32_t *seed
)
{
int i,j;
VARDECL(int *ind);
VARDECL(int *signs);
const signed char *shape_cb;
//int shape_cb_size;
int subvect_size, nb_subvect;
const split_cb_params *params;
int have_sign;
params = (const 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;
ALLOC(ind, nb_subvect, int);
ALLOC(signs, nb_subvect, int);
/* Decode codewords and gains */
for (i=0;i<nb_subvect;i++)
{
if (have_sign)
signs[i] = speex_bits_unpack_unsigned(bits, 1);
else
signs[i] = 0;
ind[i] = speex_bits_unpack_unsigned(bits, params->shape_bits);
}
/* Compute decoded excitation */
for (i=0;i<nb_subvect;i++)
{
spx_word16_t s=1;
if (signs[i])
s=-1;
#ifdef FIXED_POINT
if (s==1)
{
for (j=0;j<subvect_size;j++)
exc[subvect_size*i+j]=SHL32(EXTEND32(shape_cb[ind[i]*subvect_size+j]),SIG_SHIFT-5);
} else {
for (j=0;j<subvect_size;j++)
exc[subvect_size*i+j]=NEG32(SHL32(EXTEND32(shape_cb[ind[i]*subvect_size+j]),SIG_SHIFT-5));
}
#else
for (j=0;j<subvect_size;j++)
exc[subvect_size*i+j]+=s*0.03125*shape_cb[ind[i]*subvect_size+j];
#endif
}
}
EXPORT int speex_echo_ctl(SpeexEchoState *st, int request, void *ptr)
{
switch(request)
{
case SPEEX_ECHO_GET_FRAME_SIZE:
(*(int*)ptr) = st->frame_size;
break;
case SPEEX_ECHO_SET_SAMPLING_RATE:
st->sampling_rate = (*(int*)ptr);
st->spec_average = DIV32_16(SHL32(EXTEND32(st->frame_size), 15), st->sampling_rate);
#ifdef FIXED_POINT
st->beta0 = DIV32_16(SHL32(EXTEND32(st->frame_size), 16), st->sampling_rate);
st->beta_max = DIV32_16(SHL32(EXTEND32(st->frame_size), 14), st->sampling_rate);
#else
st->beta0 = (2.0f*st->frame_size)/st->sampling_rate;
st->beta_max = (.5f*st->frame_size)/st->sampling_rate;
#endif
if (st->sampling_rate<12000)
st->notch_radius = QCONST16(.9, 15);
else if (st->sampling_rate<24000)
st->notch_radius = QCONST16(.982, 15);
else
st->notch_radius = QCONST16(.992, 15);
break;
case SPEEX_ECHO_GET_SAMPLING_RATE:
(*(int*)ptr) = st->sampling_rate;
break;
case SPEEX_ECHO_GET_IMPULSE_RESPONSE_SIZE:
/*FIXME: Implement this for multiple channels */
*((spx_int32_t *)ptr) = st->M * st->frame_size;
break;
case SPEEX_ECHO_GET_IMPULSE_RESPONSE:
{
int M = st->M, N = st->window_size, n = st->frame_size, i, j;
spx_int32_t *filt = (spx_int32_t *) ptr;
for(j=0;j<M;j++)
{
/*FIXME: Implement this for multiple channels */
#ifdef FIXED_POINT
for (i=0;i<N;i++)
st->wtmp2[i] = EXTRACT16(PSHR32(st->W[j*N+i],16+NORMALIZE_SCALEDOWN));
spx_ifft(st->fft_table, st->wtmp2, st->wtmp);
#else
spx_ifft(st->fft_table, &st->W[j*N], st->wtmp);
#endif
for(i=0;i<n;i++)
filt[j*n+i] = PSHR32(MULT16_16(32767,st->wtmp[i]), WEIGHT_SHIFT-NORMALIZE_SCALEDOWN);
}
}
break;
default:
speex_warning_int("Unknown speex_echo_ctl request: ", request);
return -1;
}
return 0;
}
static void kf_bfly2(
kiss_fft_cpx * Fout,
const size_t fstride,
const kiss_fft_cfg st,
int m,
int N,
int mm
)
{
kiss_fft_cpx * Fout2;
kiss_fft_cpx * tw1;
kiss_fft_cpx t;
if (!st->inverse) {
int i,j;
kiss_fft_cpx * Fout_beg = Fout;
for (i=0;i<N;i++)
{
Fout = Fout_beg + i*mm;
Fout2 = Fout + m;
tw1 = st->twiddles;
for(j=0;j<m;j++)
{
/* Almost the same as the code path below, except that we divide the input by two
(while keeping the best accuracy possible) */
ms_word32_t tr, ti;
tr = SHR32(SUB32(MULT16_16(Fout2->r , tw1->r),MULT16_16(Fout2->i , tw1->i)), 1);
ti = SHR32(ADD32(MULT16_16(Fout2->i , tw1->r),MULT16_16(Fout2->r , tw1->i)), 1);
tw1 += fstride;
Fout2->r = PSHR32(SUB32(SHL32(EXTEND32(Fout->r), 14), tr), 15);
Fout2->i = PSHR32(SUB32(SHL32(EXTEND32(Fout->i), 14), ti), 15);
Fout->r = PSHR32(ADD32(SHL32(EXTEND32(Fout->r), 14), tr), 15);
Fout->i = PSHR32(ADD32(SHL32(EXTEND32(Fout->i), 14), ti), 15);
++Fout2;
++Fout;
}
}
} else {
int i,j;
kiss_fft_cpx * Fout_beg = Fout;
for (i=0;i<N;i++)
{
Fout = Fout_beg + i*mm;
Fout2 = Fout + m;
tw1 = st->twiddles;
for(j=0;j<m;j++)
{
C_MUL (t, *Fout2 , *tw1);
tw1 += fstride;
C_SUB( *Fout2 , *Fout , t );
C_ADDTO( *Fout , t );
++Fout2;
++Fout;
}
}
}
}
static
void bmw_small_f2(uint32_t* h, uint32_t* q, const void* m){
uint32_t xl=0, xh;
uint8_t i;
for(i=16;i<24;++i){
xl ^= q[i];
}
xh = xl;
for(i=24;i<32;++i){
xh ^= q[i];
}
#if DEBUG
cli_putstr("\r\n XL = ");
cli_hexdump_rev(&xl, 4);
cli_putstr("\r\n XH = ");
cli_hexdump_rev(&xh, 4);
#endif
memcpy(h, m, 16*4);
h[0] ^= SHL32(xh, 5) ^ SHR32(q[16], 5);
h[1] ^= SHR32(xh, 7) ^ SHL32(q[17], 8);
h[2] ^= SHR32(xh, 5) ^ SHL32(q[18], 5);
h[3] ^= SHR32(xh, 1) ^ SHL32(q[19], 5);
h[4] ^= SHR32(xh, 3) ^ q[20];
h[5] ^= SHL32(xh, 6) ^ SHR32(q[21], 6);
h[6] ^= SHR32(xh, 4) ^ SHL32(q[22], 6);
h[7] ^= SHR32(xh,11) ^ SHL32(q[23], 2);
for(i=0; i<8; ++i){
h[i] += xl ^ q[24+i] ^ q[i];
}
for(i=0; i<8; ++i){
h[8+i] ^= xh ^ q[24+i];
h[8+i] += ROTL32(h[(4+i)%8],i+9);
}
/*
h[ 8] += SHL32(xl, 8) ^ q[23] ^ q[ 8];
h[ 9] += SHR32(xl, 6) ^ q[16] ^ q[ 9];
h[10] += SHL32(xl, 6) ^ q[17] ^ q[10];
h[11] += SHL32(xl, 4) ^ q[18] ^ q[11];
h[12] += SHR32(xl, 3) ^ q[19] ^ q[12];
h[13] += SHR32(xl, 4) ^ q[20] ^ q[13];
h[14] += SHR32(xl, 7) ^ q[21] ^ q[14];
h[15] += SHR32(xl, 2) ^ q[22] ^ q[15];
*/
memxor(q+9, q+16, 7*4);
q[8] ^= q[23];
h[ 8] += SHL32(xl, 8) ^ q[ 8];
h[ 9] += SHR32(xl, 6) ^ q[ 9];
h[10] += SHL32(xl, 6) ^ q[10];
h[11] += SHL32(xl, 4) ^ q[11];
h[12] += SHR32(xl, 3) ^ q[12];
h[13] += SHR32(xl, 4) ^ q[13];
h[14] += SHR32(xl, 7) ^ q[14];
h[15] += SHR32(xl, 2) ^ q[15];
}
void clt_mdct_init(mdct_lookup *l,int N)
{
int i;
int N2;
l->n = N;
N2 = N>>1;
l->kfft = cpx32_fft_alloc(N>>2);
#ifndef ENABLE_TI_DSPLIB55
if (l->kfft==NULL)
return;
#endif
l->trig = (kiss_twiddle_scalar*)celt_alloc(N2*sizeof(kiss_twiddle_scalar));
if (l->trig==NULL)
return;
/* We have enough points that sine isn't necessary */
#if defined(FIXED_POINT)
#if defined(DOUBLE_PRECISION) & !defined(MIXED_PRECISION)
for (i=0;i<N2;i++)
l->trig[i] = SAMP_MAX*cos(2*M_PI*(i+1./8.)/N);
#else
for (i=0;i<N2;i++)
l->trig[i] = TRIG_UPSCALE*celt_cos_norm(DIV32(ADD32(SHL32(EXTEND32(i),17),16386),N));
#endif
#else
for (i=0;i<N2;i++)
l->trig[i] = cos(2*M_PI*(i+1./8.)/N);
#endif
}
void noise_codebook_quant(
spx_word16_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_word16_t *r,
SpeexBits *bits,
char *stack,
int complexity,
int update_target
)
{
int i;
VARDECL(spx_word16_t *tmp);
ALLOC(tmp, nsf, spx_word16_t);
residue_percep_zero16(target, ak, awk1, awk2, tmp, nsf, p, stack);
for (i=0;i<nsf;i++)
exc[i]+=SHL32(EXTEND32(tmp[i]),8);
SPEEX_MEMSET(target, 0, nsf);
}
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
}
static void celt_fir5(opus_val16 *x,
const opus_val16 *num,
int N)
{
int i;
opus_val16 num0, num1, num2, num3, num4;
opus_val32 mem0, mem1, mem2, mem3, mem4;
num0=num[0];
num1=num[1];
num2=num[2];
num3=num[3];
num4=num[4];
mem0=0;
mem1=0;
mem2=0;
mem3=0;
mem4=0;
for (i=0;i<N;i++)
{
opus_val32 sum = SHL32(EXTEND32(x[i]), SIG_SHIFT);
sum = MAC16_16(sum,num0,mem0);
sum = MAC16_16(sum,num1,mem1);
sum = MAC16_16(sum,num2,mem2);
sum = MAC16_16(sum,num3,mem3);
sum = MAC16_16(sum,num4,mem4);
mem4 = mem3;
mem3 = mem2;
mem2 = mem1;
mem1 = mem0;
mem0 = x[i];
x[i] = ROUND16(sum, SIG_SHIFT);
}
}
static int resampler_basic_interpolate_single(SpeexResamplerState *st, spx_uint32_t channel_index, const spx_word16_t *in, spx_uint32_t *in_len, spx_word16_t *out, spx_uint32_t *out_len)
{
const int N = st->filt_len;
int out_sample = 0;
int last_sample = st->last_sample[channel_index];
spx_uint32_t samp_frac_num = st->samp_frac_num[channel_index];
const int out_stride = st->out_stride;
const int int_advance = st->int_advance;
const int frac_advance = st->frac_advance;
const spx_uint32_t den_rate = st->den_rate;
spx_word32_t sum;
while (!(last_sample >= (spx_int32_t)*in_len || out_sample >= (spx_int32_t)*out_len))
{
const spx_word16_t *iptr = & in[last_sample];
const int offset = samp_frac_num*st->oversample/st->den_rate;
#ifdef FIXED_POINT
const spx_word16_t frac = PDIV32(SHL32((samp_frac_num*st->oversample) % st->den_rate,15),st->den_rate);
#else
const spx_word16_t frac = ((float)((samp_frac_num*st->oversample) % st->den_rate))/st->den_rate;
#endif
spx_word16_t interp[4];
#ifndef OVERRIDE_INTERPOLATE_PRODUCT_SINGLE
int j;
spx_word32_t accum[4] = {0,0,0,0};
for(j=0;j<N;j++) {
const spx_word16_t curr_in=iptr[j];
accum[0] += MULT16_16(curr_in,st->sinc_table[4+(j+1)*st->oversample-offset-2]);
accum[1] += MULT16_16(curr_in,st->sinc_table[4+(j+1)*st->oversample-offset-1]);
accum[2] += MULT16_16(curr_in,st->sinc_table[4+(j+1)*st->oversample-offset]);
accum[3] += MULT16_16(curr_in,st->sinc_table[4+(j+1)*st->oversample-offset+1]);
}
cubic_coef(frac, interp);
sum = MULT16_32_Q15(interp[0],SHR32(accum[0], 1)) + MULT16_32_Q15(interp[1],SHR32(accum[1], 1)) + MULT16_32_Q15(interp[2],SHR32(accum[2], 1)) + MULT16_32_Q15(interp[3],SHR32(accum[3], 1));
sum = SATURATE32PSHR(sum, 15, 32767);
#else
cubic_coef(frac, interp);
sum = interpolate_product_single(iptr, st->sinc_table + st->oversample + 4 - offset - 2, N, st->oversample, interp);
#endif
out[out_stride * out_sample++] = sum;
last_sample += int_advance;
samp_frac_num += frac_advance;
if (samp_frac_num >= den_rate)
{
samp_frac_num -= den_rate;
last_sample++;
}
}
st->last_sample[channel_index] = last_sample;
st->samp_frac_num[channel_index] = samp_frac_num;
return out_sample;
}
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(SHL32(EXTEND32(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]);
}
}
void celt_fir_c(
const opus_val16 *_x,
const opus_val16 *num,
opus_val16 *_y,
int N,
int ord,
opus_val16 *mem,
int arch)
{
int i,j;
VARDECL(opus_val16, rnum);
VARDECL(opus_val16, x);
SAVE_STACK;
ALLOC(rnum, ord, opus_val16);
ALLOC(x, N+ord, opus_val16);
for(i=0;i<ord;i++)
rnum[i] = num[ord-i-1];
for(i=0;i<ord;i++)
x[i] = mem[ord-i-1];
for (i=0;i<N;i++)
x[i+ord]=_x[i];
for(i=0;i<ord;i++)
mem[i] = _x[N-i-1];
#ifdef SMALL_FOOTPRINT
(void)arch;
for (i=0;i<N;i++)
{
opus_val32 sum = SHL32(EXTEND32(_x[i]), SIG_SHIFT);
for (j=0;j<ord;j++)
{
sum = MAC16_16(sum,rnum[j],x[i+j]);
}
_y[i] = SATURATE16(PSHR32(sum, SIG_SHIFT));
}
#else
for (i=0;i<N-3;i+=4)
{
opus_val32 sum[4]={0,0,0,0};
xcorr_kernel(rnum, x+i, sum, ord, arch);
_y[i ] = SATURATE16(ADD32(EXTEND32(_x[i ]), PSHR32(sum[0], SIG_SHIFT)));
_y[i+1] = SATURATE16(ADD32(EXTEND32(_x[i+1]), PSHR32(sum[1], SIG_SHIFT)));
_y[i+2] = SATURATE16(ADD32(EXTEND32(_x[i+2]), PSHR32(sum[2], SIG_SHIFT)));
_y[i+3] = SATURATE16(ADD32(EXTEND32(_x[i+3]), PSHR32(sum[3], SIG_SHIFT)));
}
for (;i<N;i++)
{
opus_val32 sum = 0;
for (j=0;j<ord;j++)
sum = MAC16_16(sum,rnum[j],x[i+j]);
_y[i] = SATURATE16(ADD32(EXTEND32(_x[i]), PSHR32(sum, SIG_SHIFT)));
}
#endif
RESTORE_STACK;
}
void mlp_process(const MLP * m, const opus_val16 * in, opus_val16 * out)
{
int j;
opus_val16 hidden[MAX_NEURONS];
const opus_val16 *W = m->weights;
/* Copy to tmp_in */
for (j = 0; j < m->topo[1]; j++) {
int k;
opus_val32 sum = SHL32(EXTEND32(*W++), 8);
for (k = 0; k < m->topo[0]; k++)
sum = MAC16_16(sum, in[k], *W++);
hidden[j] = tansig_approx(sum);
}
for (j = 0; j < m->topo[2]; j++) {
int k;
opus_val32 sum = SHL32(EXTEND32(*W++), 14);
for (k = 0; k < m->topo[1]; k++)
sum = MAC16_16(sum, hidden[k], *W++);
out[j] = tansig_approx(EXTRACT16(PSHR32(sum, 17)));
}
}
void _celt_lpc(
opus_val16 *_lpc, /* out: [0...p-1] LPC coefficients */
const opus_val32 *ac, /* in: [0...p] autocorrelation values */
int p
)
{
int i, j;
opus_val32 r;
opus_val32 error = ac[0];
#ifdef OPUS_FIXED_POINT
opus_val32 lpc[LPC_ORDER];
#else
float *lpc = _lpc;
#endif
for (i = 0; i < p; i++)
lpc[i] = 0;
if (ac[0] != 0)
{
for (i = 0; i < p; i++) {
/* Sum up this iteration's reflection coefficient */
opus_val32 rr = 0;
for (j = 0; j < i; j++)
rr += MULT32_32_Q31(lpc[j],ac[i - j]);
rr += SHR32(ac[i + 1],3);
r = -frac_div32(SHL32(rr,3), error);
/* Update LPC coefficients and total error */
lpc[i] = SHR32(r,3);
for (j = 0; j < (i+1)>>1; j++)
{
opus_val32 tmp1, tmp2;
tmp1 = lpc[j];
tmp2 = lpc[i-1-j];
lpc[j] = tmp1 + MULT32_32_Q31(r,tmp2);
lpc[i-1-j] = tmp2 + MULT32_32_Q31(r,tmp1);
}
error = error - MULT32_32_Q31(MULT32_32_Q31(r,r),error);
/* Bail out once we get 30 dB gain */
#ifdef OPUS_FIXED_POINT
if (error<SHR32(ac[0],10))
break;
#else
if (error<.001f*ac[0])
break;
#endif
}
}
#ifdef OPUS_FIXED_POINT
for (i=0;i<p;i++)
_lpc[i] = ROUND16(lpc[i],16);
#endif
}
void noise_codebook_unquant(
spx_sig_t *exc,
const void *par, /* non-overlapping codebook */
int nsf, /* number of samples in subframe */
SpeexBits *bits,
char *stack,
spx_int32_t *seed
)
{
int i;
/* FIXME: This is bad, but I don't think the function ever gets called anyway */
for (i=0;i<nsf;i++)
exc[i]=SHL32(EXTEND32(speex_rand(1, seed)),SIG_SHIFT);
}
static inline void filter_dc_notch16(const spx_int16_t *in, spx_word16_t radius, spx_word16_t *out, int len, spx_mem_t *mem, int stride)
{
int i;
spx_word16_t den2;
#ifdef FIXED_POINT
den2 = MULT16_16_Q15(radius,radius) + MULT16_16_Q15(QCONST16(.7,15),MULT16_16_Q15(32767-radius,32767-radius));
#else
den2 = radius*radius + .7*(1-radius)*(1-radius);
#endif
/*printf ("%d %d %d %d %d %d\n", num[0], num[1], num[2], den[0], den[1], den[2]);*/
for (i=0;i<len;i++)
{
spx_word16_t vin = in[i*stride];
spx_word32_t vout = mem[0] + SHL32(EXTEND32(vin),15);
#ifdef FIXED_POINT
mem[0] = mem[1] + SHL32(SHL32(-EXTEND32(vin),15) + MULT16_32_Q15(radius,vout),1);
#else
mem[0] = mem[1] + 2*(-vin + radius*vout);
#endif
mem[1] = SHL32(EXTEND32(vin),15) - MULT16_32_Q15(den2,vout);
out[i] = SATURATE32(PSHR32(MULT16_32_Q15(radius,vout),15),32767);
}
}
static void compute_twiddles(kiss_twiddle_cpx *twiddles, int nfft)
{
int i;
#ifdef FIXED_POINT
for (i=0;i<nfft;++i) {
opus_val32 phase = -i;
kf_cexp2(twiddles+i, DIV32(SHL32(phase,17),nfft));
}
#else
for (i=0;i<nfft;++i) {
const double pi=3.14159265358979323846264338327;
double phase = ( -2*pi /nfft ) * i;
kf_cexp(twiddles+i, phase );
}
#endif
}
/* returns minimum mean square error */
spx_word32_t _spx_lpc(
spx_coef_t *lpc, /* out: [0...p-1] LPC coefficients */
const spx_word16_t *ac, /* in: [0...p] autocorrelation values */
int p
)
{
int i, j;
spx_word16_t r;
spx_word16_t error = ac[0];
if (ac[0] == 0)
{
for (i = 0; i < p; i++)
lpc[i] = 0;
return 0;
}
for (i = 0; i < p; i++) {
/* Sum up this iteration's reflection coefficient */
spx_word32_t rr = NEG32(SHL32(EXTEND32(ac[i + 1]),13));
for (j = 0; j < i; j++)
rr = SUB32(rr,MULT16_16(lpc[j],ac[i - j]));
#ifdef FIXED_POINT
r = DIV32_16(rr+PSHR32(error,1),ADD16(error,8));
#else
r = rr/(error+.003*ac[0]);
#endif
/* Update LPC coefficients and total error */
lpc[i] = r;
for (j = 0; j < i>>1; j++)
{
spx_word16_t tmp = lpc[j];
lpc[j] = MAC16_16_P13(lpc[j],r,lpc[i-1-j]);
lpc[i-1-j] = MAC16_16_P13(lpc[i-1-j],r,tmp);
}
if (i & 1)
lpc[j] = MAC16_16_P13(lpc[j],lpc[j],r);
error = SUB16(error,MULT16_16_Q13(r,MULT16_16_Q13(error,r)));
}
return error;
}
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);
}
}
/*
*
* User-callable function to allocate all necessary storage space for the fft.
*
* The return value is a contiguous block of memory, allocated with malloc. As such,
* It can be freed with free(), rather than a kiss_fft-specific function.
* */
kiss_fft_cfg kiss_fft_alloc(int nfft,int inverse_fft,void * mem,size_t * lenmem )
{
kiss_fft_cfg st=NULL;
size_t memneeded = sizeof(struct kiss_fft_state)
+ sizeof(kiss_fft_cpx)*(nfft-1); /* twiddle factors*/
if ( lenmem==NULL ) {
st = ( kiss_fft_cfg)KISS_FFT_MALLOC( memneeded );
}else{
if (mem != NULL && *lenmem >= memneeded)
st = (kiss_fft_cfg)mem;
*lenmem = memneeded;
}
if (st) {
int i;
st->nfft=nfft;
st->inverse = inverse_fft;
#ifdef MS_FIXED_POINT
for (i=0;i<nfft;++i) {
ms_word32_t phase = i;
if (!st->inverse)
phase = -phase;
kf_cexp2(st->twiddles+i, DIV32(SHL32(phase,17),nfft));
}
#else
for (i=0;i<nfft;++i) {
const double pi=3.14159265358979323846264338327;
double phase = ( -2*pi /nfft ) * i;
if (st->inverse)
phase *= -1;
kf_cexp(st->twiddles+i, phase );
}
#endif
kf_factor(nfft,st->factors);
}
return st;
}
void celt_fir(const opus_val16 *x,
const opus_val16 *num,
opus_val16 *y,
int N,
int ord,
opus_val16 *mem)
{
int i,j;
for (i=0;i<N;i++)
{
opus_val32 sum = SHL32(EXTEND32(x[i]), SIG_SHIFT);
for (j=0;j<ord;j++)
{
sum += MULT16_16(num[j],mem[j]);
}
for (j=ord-1;j>=1;j--)
{
mem[j]=mem[j-1];
}
mem[0] = x[i];
y[i] = ROUND16(sum, SIG_SHIFT);
}
}
void pitch_unquant_3tap(
spx_word16_t exc[], /* Input excitation */
spx_word32_t exc_out[], /* Output excitation */
int start, /* Smallest pitch value allowed */
int end, /* Largest pitch value allowed */
spx_word16_t pitch_coef, /* Voicing (pitch) coefficient */
const void* par,
int nsf, /* Number of samples in subframe */
int* pitch_val,
spx_word16_t* gain_val,
SpeexBits* bits,
char* stack,
int count_lost,
int subframe_offset,
spx_word16_t last_pitch_gain,
int cdbk_offset
) {
int i;
int pitch;
int gain_index;
spx_word16_t gain[3];
const signed char* gain_cdbk;
int gain_cdbk_size;
const ltp_params* params;
params = (const ltp_params*) par;
gain_cdbk_size = 1 << params->gain_bits;
gain_cdbk = params->gain_cdbk + 4 * gain_cdbk_size * cdbk_offset;
pitch = speex_bits_unpack_unsigned(bits, params->pitch_bits);
pitch += start;
gain_index = speex_bits_unpack_unsigned(bits, params->gain_bits);
/*printf ("decode pitch: %d %d\n", pitch, gain_index);*/
#ifdef FIXED_POINT
gain[0] = ADD16(32, (spx_word16_t)gain_cdbk[gain_index * 4]);
gain[1] = ADD16(32, (spx_word16_t)gain_cdbk[gain_index * 4 + 1]);
gain[2] = ADD16(32, (spx_word16_t)gain_cdbk[gain_index * 4 + 2]);
#else
gain[0] = 0.015625 * gain_cdbk[gain_index * 4] + .5;
gain[1] = 0.015625 * gain_cdbk[gain_index * 4 + 1] + .5;
gain[2] = 0.015625 * gain_cdbk[gain_index * 4 + 2] + .5;
#endif
if (count_lost && pitch > subframe_offset) {
spx_word16_t gain_sum;
if (1) {
#ifdef FIXED_POINT
spx_word16_t tmp = count_lost < 4 ? last_pitch_gain : SHR16(last_pitch_gain, 1);
if (tmp > 62)
tmp = 62;
#else
spx_word16_t tmp = count_lost < 4 ? last_pitch_gain : 0.5 * last_pitch_gain;
if (tmp > .95)
tmp = .95;
#endif
gain_sum = gain_3tap_to_1tap(gain);
if (gain_sum > tmp) {
spx_word16_t fact = DIV32_16(SHL32(EXTEND32(tmp), 14), gain_sum);
for (i = 0; i < 3; i++)
gain[i] = MULT16_16_Q14(fact, gain[i]);
}
}
}
*pitch_val = pitch;
gain_val[0] = gain[0];
gain_val[1] = gain[1];
gain_val[2] = gain[2];
gain[0] = SHL16(gain[0], 7);
gain[1] = SHL16(gain[1], 7);
gain[2] = SHL16(gain[2], 7);
SPEEX_MEMSET(exc_out, 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_out[j] = MAC16_16(exc_out[j], gain[2 - i], exc[j - pp]);
tmp3 = nsf;
if (tmp3 > pp + pitch)
tmp3 = pp + pitch;
for (j = tmp1; j < tmp3; j++)
exc_out[j] = MAC16_16(exc_out[j], gain[2 - i], exc[j - pp - pitch]);
}
/*for (i=0;i<nsf;i++)
exc[i]=PSHR32(exc32[i],13);*/
}
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;
}
void split_cb_search_shape_sign(
spx_word16_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_word16_t *r,
SpeexBits *bits,
char *stack,
int complexity,
int update_target
)
{
int i,j,m,q;
const signed char *shape_cb;
int shape_cb_size = 32, subvect_size = 10;
int best_index;
spx_word32_t best_dist;
spx_word16_t resp[320];
spx_word16_t *resp2 = resp;
spx_word32_t E[32];
spx_word16_t t[40];
spx_sig_t e[40];
shape_cb=exc_10_32_table;
/* FIXME: Do we still need to copy the target? */
SPEEX_COPY(t, target, nsf);
//compute_weighted_codebook
{
int i, k;
spx_word16_t shape[10];
for (i=0;i<shape_cb_size;i++)
{
spx_word16_t *res;
res = resp+i*subvect_size;
for (k=0;k<subvect_size;k++)
shape[k] = (spx_word16_t)shape_cb[i*subvect_size+k];
E[i]=0;
/* Compute codeword response using convolution with impulse response */
{
spx_word32_t resj;
spx_word16_t res16;
// 0
resj = MULT16_16(shape[0],r[0]);
res16 = EXTRACT16(SHR32(resj, 13));
// Compute codeword energy
E[i]=MAC16_16(E[i],res16,res16);
res[0] = res16;
//++++++++++++++++++++++++++
// 1
resj = MULT16_16(shape[0],r[1]);
resj = MAC16_16(resj,shape[1],r[0]);
res16 = EXTRACT16(SHR32(resj, 13));
// Compute codeword energy
E[i]=MAC16_16(E[i],res16,res16);
res[1] = res16;
//++++++++++++++++++++++++++
// 2
resj = MULT16_16(shape[0],r[2]);
resj = MAC16_16(resj,shape[1],r[1]);
resj = MAC16_16(resj,shape[2],r[0]);
res16 = EXTRACT16(SHR32(resj, 13));
// Compute codeword energy
E[i]=MAC16_16(E[i],res16,res16);
res[2] = res16;
//++++++++++++++++++++++++++
// 3
resj = MULT16_16(shape[0],r[3]);
resj = MAC16_16(resj,shape[1],r[2]);
resj = MAC16_16(resj,shape[2],r[1]);
resj = MAC16_16(resj,shape[3],r[0]);
res16 = EXTRACT16(SHR32(resj, 13));
// Compute codeword energy
E[i]=MAC16_16(E[i],res16,res16);
res[3] = res16;
//++++++++++++++++++++++++++
// 4
resj = MULT16_16(shape[0],r[4]);
resj = MAC16_16(resj,shape[1],r[3]);
resj = MAC16_16(resj,shape[2],r[2]);
resj = MAC16_16(resj,shape[3],r[1]);
resj = MAC16_16(resj,shape[4],r[0]);
res16 = EXTRACT16(SHR32(resj, 13));
// Compute codeword energy
E[i]=MAC16_16(E[i],res16,res16);
res[4] = res16;
//++++++++++++++++++++++++++
// 5
resj = MULT16_16(shape[0],r[5]);
resj = MAC16_16(resj,shape[1],r[4]);
resj = MAC16_16(resj,shape[2],r[3]);
resj = MAC16_16(resj,shape[3],r[2]);
resj = MAC16_16(resj,shape[4],r[1]);
resj = MAC16_16(resj,shape[5],r[0]);
res16 = EXTRACT16(SHR32(resj, 13));
// Compute codeword energy
E[i]=MAC16_16(E[i],res16,res16);
res[5] = res16;
//++++++++++++++++++++++++++
// 6
resj = MULT16_16(shape[0],r[6]);
resj = MAC16_16(resj,shape[1],r[5]);
resj = MAC16_16(resj,shape[2],r[4]);
resj = MAC16_16(resj,shape[3],r[3]);
resj = MAC16_16(resj,shape[4],r[2]);
resj = MAC16_16(resj,shape[5],r[1]);
resj = MAC16_16(resj,shape[6],r[0]);
res16 = EXTRACT16(SHR32(resj, 13));
// Compute codeword energy
E[i]=MAC16_16(E[i],res16,res16);
res[6] = res16;
//++++++++++++++++++++++++++
// 7
resj = MULT16_16(shape[0],r[7]);
resj = MAC16_16(resj,shape[1],r[6]);
resj = MAC16_16(resj,shape[2],r[5]);
resj = MAC16_16(resj,shape[3],r[4]);
resj = MAC16_16(resj,shape[4],r[3]);
resj = MAC16_16(resj,shape[5],r[2]);
resj = MAC16_16(resj,shape[6],r[1]);
resj = MAC16_16(resj,shape[7],r[0]);
res16 = EXTRACT16(SHR32(resj, 13));
// Compute codeword energy
E[i]=MAC16_16(E[i],res16,res16);
res[7] = res16;
//++++++++++++++++++++++++++
// 8
resj = MULT16_16(shape[0],r[8]);
resj = MAC16_16(resj,shape[1],r[7]);
resj = MAC16_16(resj,shape[2],r[6]);
resj = MAC16_16(resj,shape[3],r[5]);
resj = MAC16_16(resj,shape[4],r[4]);
resj = MAC16_16(resj,shape[5],r[3]);
resj = MAC16_16(resj,shape[6],r[2]);
resj = MAC16_16(resj,shape[7],r[1]);
resj = MAC16_16(resj,shape[8],r[0]);
res16 = EXTRACT16(SHR32(resj, 13));
// Compute codeword energy
E[i]=MAC16_16(E[i],res16,res16);
res[8] = res16;
//++++++++++++++++++++++++++
// 9
resj = MULT16_16(shape[0],r[9]);
resj = MAC16_16(resj,shape[1],r[8]);
resj = MAC16_16(resj,shape[2],r[7]);
resj = MAC16_16(resj,shape[3],r[6]);
resj = MAC16_16(resj,shape[4],r[5]);
resj = MAC16_16(resj,shape[5],r[4]);
resj = MAC16_16(resj,shape[6],r[3]);
resj = MAC16_16(resj,shape[7],r[2]);
resj = MAC16_16(resj,shape[8],r[1]);
resj = MAC16_16(resj,shape[9],r[0]);
res16 = EXTRACT16(SHR32(resj, 13));
// Compute codeword energy
E[i]=MAC16_16(E[i],res16,res16);
res[9] = res16;
//++++++++++++++++++++++++++
}
}
}
for (i=0;i<4;i++)
{
spx_word16_t *x=t+subvect_size*i;
/*Find new n-best based on previous n-best j*/
vq_nbest(x, resp2, subvect_size, shape_cb_size, E, 1, &best_index, &best_dist, stack);
speex_bits_pack(bits,best_index,5);
{
int rind;
spx_word16_t *res;
spx_word16_t sign=1;
rind = best_index;
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] = SUB16(t[subvect_size*i+m], res[m]);
else
for (m=0;m<subvect_size;m++)
t[subvect_size*i+m] = ADD16(t[subvect_size*i+m], res[m]);
if (sign==1)
{
for (j=0;j<subvect_size;j++)
e[subvect_size*i+j]=SHL32(EXTEND32(shape_cb[rind*subvect_size+j]),SIG_SHIFT-5);
} else {
for (j=0;j<subvect_size;j++)
e[subvect_size*i+j]=NEG32(SHL32(EXTEND32(shape_cb[rind*subvect_size+j]),SIG_SHIFT-5));
}
}
for (m=0;m<subvect_size;m++)
{
spx_word16_t g;
int rind;
spx_word16_t sign=1;
rind = best_index;
if (rind>=shape_cb_size)
{
sign=-1;
rind-=shape_cb_size;
}
q=subvect_size-m;
g=sign*shape_cb[rind*subvect_size+m];
target_update(t+subvect_size*(i+1), g, r+q, nsf-subvect_size*(i+1));
}
}
/* Update excitation */
/* FIXME: We could update the excitation directly above */
for (j=0;j<nsf;j++)
exc[j]=ADD32(exc[j],e[j]);
}
/** 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;
}
void open_loop_nbest_pitch(spx_word16_t* sw, int start, int end, int len, int* pitch, spx_word16_t* gain, int N, char* stack) {
int i, j, k;
VARDECL(spx_word32_t * best_score);
VARDECL(spx_word32_t * best_ener);
spx_word32_t e0;
VARDECL(spx_word32_t * corr);
#ifdef FIXED_POINT
/* In fixed-point, we need only one (temporary) array of 32-bit values and two (corr16, ener16)
arrays for (normalized) 16-bit values */
VARDECL(spx_word16_t * corr16);
VARDECL(spx_word16_t * ener16);
spx_word32_t* energy;
int cshift = 0, eshift = 0;
int scaledown = 0;
ALLOC(corr16, end - start + 1, spx_word16_t);
ALLOC(ener16, end - start + 1, spx_word16_t);
ALLOC(corr, end - start + 1, spx_word32_t);
energy = corr;
#else
/* In floating-point, we need to float arrays and no normalized copies */
VARDECL(spx_word32_t * energy);
spx_word16_t* corr16;
spx_word16_t* ener16;
ALLOC(energy, end - start + 2, spx_word32_t);
ALLOC(corr, end - start + 1, spx_word32_t);
corr16 = corr;
ener16 = energy;
#endif
ALLOC(best_score, N, spx_word32_t);
ALLOC(best_ener, N, spx_word32_t);
for (i = 0; i < N; i++) {
best_score[i] = -1;
best_ener[i] = 0;
pitch[i] = start;
}
#ifdef FIXED_POINT
for (i = -end; i < len; i++) {
if (ABS16(sw[i]) > 16383) {
scaledown = 1;
break;
}
}
/* If the weighted input is close to saturation, then we scale it down */
if (scaledown) {
for (i = -end; i < len; i++) {
sw[i] = SHR16(sw[i], 1);
}
}
#endif
energy[0] = inner_prod(sw - start, sw - start, len);
e0 = inner_prod(sw, sw, len);
for (i = start; i < end; i++) {
/* Update energy for next pitch*/
energy[i - start + 1] = SUB32(ADD32(energy[i - start], SHR32(MULT16_16(sw[-i - 1], sw[-i - 1]), 6)), SHR32(MULT16_16(sw[-i + len - 1], sw[-i + len - 1]), 6));
if (energy[i - start + 1] < 0)
energy[i - start + 1] = 0;
}
#ifdef FIXED_POINT
eshift = normalize16(energy, ener16, 32766, end - start + 1);
#endif
/* In fixed-point, this actually overrites the energy array (aliased to corr) */
pitch_xcorr(sw, sw - end, corr, len, end - start + 1, stack);
#ifdef FIXED_POINT
/* Normalize to 180 so we can square it and it still fits in 16 bits */
cshift = normalize16(corr, corr16, 180, end - start + 1);
/* If we scaled weighted input down, we need to scale it up again (OK, so we've just lost the LSB, who cares?) */
if (scaledown) {
for (i = -end; i < len; i++) {
sw[i] = SHL16(sw[i], 1);
}
}
#endif
/* Search for the best pitch prediction gain */
for (i = start; i <= end; i++) {
spx_word16_t tmp = MULT16_16_16(corr16[i - start], corr16[i - start]);
/* Instead of dividing the tmp by the energy, we multiply on the other side */
if (MULT16_16(tmp, best_ener[N - 1]) > MULT16_16(best_score[N - 1], ADD16(1, ener16[i - start]))) {
/* We can safely put it last and then check */
best_score[N - 1] = tmp;
best_ener[N - 1] = ener16[i - start] + 1;
pitch[N - 1] = i;
/* Check if it comes in front of others */
for (j = 0; j < N - 1; j++) {
if (MULT16_16(tmp, best_ener[j]) > MULT16_16(best_score[j], ADD16(1, ener16[i - start]))) {
for (k = N - 1; k > j; k--) {
best_score[k] = best_score[k - 1];
best_ener[k] = best_ener[k - 1];
pitch[k] = pitch[k - 1];
}
best_score[j] = tmp;
best_ener[j] = ener16[i - start] + 1;
pitch[j] = i;
break;
}
}
}
}
/* Compute open-loop gain if necessary */
if (gain) {
for (j = 0; j < N; j++) {
spx_word16_t g;
i = pitch[j];
g = DIV32(SHL32(EXTEND32(corr16[i - start]), cshift), 10 + SHR32(MULT16_16(spx_sqrt(e0), spx_sqrt(SHL32(EXTEND32(ener16[i - start]), eshift))), 6));
/* FIXME: g = max(g,corr/energy) */
if (g < 0)
g = 0;
gain[j] = g;
}
}
}
void lsp_to_lpc(const spx_lsp_t *freq,spx_coef_t *ak,int lpcrdr, char *stack)
/* float *freq array of LSP frequencies in the x domain */
/* float *ak array of LPC coefficients */
/* int lpcrdr order of LPC coefficients */
{
int i,j;
spx_word32_t xout1,xout2,xin;
spx_word32_t mult, a;
VARDECL(spx_word16_t *freqn);
VARDECL(spx_word32_t **xp);
VARDECL(spx_word32_t *xpmem);
VARDECL(spx_word32_t **xq);
VARDECL(spx_word32_t *xqmem);
int m = lpcrdr>>1;
/*
Reconstruct P(z) and Q(z) by cascading second order polynomials
in form 1 - 2cos(w)z(-1) + z(-2), where w is the LSP frequency.
In the time domain this is:
y(n) = x(n) - 2cos(w)x(n-1) + x(n-2)
This is what the ALLOCS below are trying to do:
int xp[m+1][lpcrdr+1+2]; // P matrix in QIMP
int xq[m+1][lpcrdr+1+2]; // Q matrix in QIMP
These matrices store the output of each stage on each row. The
final (m-th) row has the output of the final (m-th) cascaded
2nd order filter. The first row is the impulse input to the
system (not written as it is known).
The version below takes advantage of the fact that a lot of the
outputs are zero or known, for example if we put an inpulse
into the first section the "clock" it 10 times only the first 3
outputs samples are non-zero (it's an FIR filter).
*/
ALLOC(xp, (m+1), spx_word32_t*);
ALLOC(xpmem, (m+1)*(lpcrdr+1+2), spx_word32_t);
ALLOC(xq, (m+1), spx_word32_t*);
ALLOC(xqmem, (m+1)*(lpcrdr+1+2), spx_word32_t);
for(i=0; i<=m; i++) {
xp[i] = xpmem + i*(lpcrdr+1+2);
xq[i] = xqmem + i*(lpcrdr+1+2);
}
/* work out 2cos terms in Q14 */
ALLOC(freqn, lpcrdr, spx_word16_t);
for (i=0;i<lpcrdr;i++)
freqn[i] = ANGLE2X(freq[i]);
#define QIMP 21 /* scaling for impulse */
xin = SHL32(EXTEND32(1), (QIMP-1)); /* 0.5 in QIMP format */
/* first col and last non-zero values of each row are trivial */
for(i=0;i<=m;i++) {
xp[i][1] = 0;
xp[i][2] = xin;
xp[i][2+2*i] = xin;
xq[i][1] = 0;
xq[i][2] = xin;
xq[i][2+2*i] = xin;
}
/* 2nd row (first output row) is trivial */
xp[1][3] = -MULT16_32_Q14(freqn[0],xp[0][2]);
xq[1][3] = -MULT16_32_Q14(freqn[1],xq[0][2]);
xout1 = xout2 = 0;
/* now generate remaining rows */
for(i=1;i<m;i++) {
for(j=1;j<2*(i+1)-1;j++) {
mult = MULT16_32_Q14(freqn[2*i],xp[i][j+1]);
xp[i+1][j+2] = ADD32(SUB32(xp[i][j+2], mult), xp[i][j]);
mult = MULT16_32_Q14(freqn[2*i+1],xq[i][j+1]);
xq[i+1][j+2] = ADD32(SUB32(xq[i][j+2], mult), xq[i][j]);
}
/* for last col xp[i][j+2] = xq[i][j+2] = 0 */
mult = MULT16_32_Q14(freqn[2*i],xp[i][j+1]);
xp[i+1][j+2] = SUB32(xp[i][j], mult);
mult = MULT16_32_Q14(freqn[2*i+1],xq[i][j+1]);
xq[i+1][j+2] = SUB32(xq[i][j], mult);
}
/* process last row to extra a{k} */
for(j=1;j<=lpcrdr;j++) {
int shift = QIMP-13;
/* final filter sections */
a = PSHR32(xp[m][j+2] + xout1 + xq[m][j+2] - xout2, shift);
xout1 = xp[m][j+2];
xout2 = xq[m][j+2];
/* hard limit ak's to +/- 32767 */
if (a < -32767) a = -32767;
if (a > 32767) a = 32767;
ak[j-1] = (short)a;
}
}
FilterBank *filterbank_new(int banks, spx_word32_t sampling, int len, int type)
{
FilterBank *bank;
spx_word32_t df;
spx_word32_t max_mel, mel_interval;
int i;
int id1;
int id2;
df = DIV32(SHL32(sampling,15),MULT16_16(2,len));
max_mel = toBARK(EXTRACT16(sampling/2));
mel_interval = PDIV32(max_mel,banks-1);
bank = (FilterBank*)speex_alloc(sizeof(FilterBank));
bank->nb_banks = banks;
bank->len = len;
bank->bank_left = (int*)speex_alloc(len*sizeof(int));
bank->bank_right = (int*)speex_alloc(len*sizeof(int));
bank->filter_left = (spx_word16_t*)speex_alloc(len*sizeof(spx_word16_t));
bank->filter_right = (spx_word16_t*)speex_alloc(len*sizeof(spx_word16_t));
/* Think I can safely disable normalisation that for fixed-point (and probably float as well) */
#ifndef FIXED_POINT
bank->scaling = (float*)speex_alloc(banks*sizeof(float));
#endif
for (i=0;i<len;i++)
{
spx_word16_t curr_freq;
spx_word32_t mel;
spx_word16_t val;
curr_freq = EXTRACT16(MULT16_32_P15(i,df));
mel = toBARK(curr_freq);
if (mel > max_mel)
break;
#ifdef FIXED_POINT
id1 = DIV32(mel,mel_interval);
#else
id1 = (int)(floor(mel/mel_interval));
#endif
if (id1>banks-2)
{
id1 = banks-2;
val = Q15_ONE;
} else {
val = DIV32_16(mel - id1*mel_interval,EXTRACT16(PSHR32(mel_interval,15)));
}
id2 = id1+1;
bank->bank_left[i] = id1;
bank->filter_left[i] = SUB16(Q15_ONE,val);
bank->bank_right[i] = id2;
bank->filter_right[i] = val;
}
/* Think I can safely disable normalisation for fixed-point (and probably float as well) */
#ifndef FIXED_POINT
for (i=0;i<bank->nb_banks;i++)
bank->scaling[i] = 0;
for (i=0;i<bank->len;i++)
{
int id = bank->bank_left[i];
bank->scaling[id] += bank->filter_left[i];
id = bank->bank_right[i];
bank->scaling[id] += bank->filter_right[i];
}
for (i=0;i<bank->nb_banks;i++)
bank->scaling[i] = Q15_ONE/(bank->scaling[i]);
#endif
return bank;
}
void split_cb_search_shape_sign(
spx_word16_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_word16_t *r,
SpeexBits *bits,
char *stack,
int complexity,
int update_target
)
{
int i,j,k,m,n,q;
VARDECL(spx_word16_t *resp);
#ifdef _USE_SSE
VARDECL(__m128 *resp2);
VARDECL(__m128 *E);
#else
spx_word16_t *resp2;
VARDECL(spx_word32_t *E);
#endif
VARDECL(spx_word16_t *t);
VARDECL(spx_sig_t *e);
VARDECL(spx_word16_t *tmp);
VARDECL(spx_word32_t *ndist);
VARDECL(spx_word32_t *odist);
VARDECL(int *itmp);
VARDECL(spx_word16_t **ot2);
VARDECL(spx_word16_t **nt2);
spx_word16_t **ot, **nt;
VARDECL(int **nind);
VARDECL(int **oind);
VARDECL(int *ind);
const signed char *shape_cb;
int shape_cb_size, subvect_size, nb_subvect;
const split_cb_params *params;
int N=2;
VARDECL(int *best_index);
VARDECL(spx_word32_t *best_dist);
VARDECL(int *best_nind);
VARDECL(int *best_ntarget);
int have_sign;
N=complexity;
if (N>10)
N=10;
/* Complexity isn't as important for the codebooks as it is for the pitch */
N=(2*N)/3;
if (N<1)
N=1;
if (N==1)
{
split_cb_search_shape_sign_N1(target,ak,awk1,awk2,par,p,nsf,exc,r,bits,stack,update_target);
return;
}
ALLOC(ot2, N, spx_word16_t*);
ALLOC(nt2, N, spx_word16_t*);
ALLOC(oind, N, int*);
ALLOC(nind, N, int*);
params = (const 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;
ALLOC(resp, shape_cb_size*subvect_size, spx_word16_t);
#ifdef _USE_SSE
ALLOC(resp2, (shape_cb_size*subvect_size)>>2, __m128);
ALLOC(E, shape_cb_size>>2, __m128);
#else
resp2 = resp;
ALLOC(E, shape_cb_size, spx_word32_t);
#endif
ALLOC(t, nsf, spx_word16_t);
ALLOC(e, nsf, spx_sig_t);
ALLOC(ind, nb_subvect, int);
ALLOC(tmp, 2*N*nsf, spx_word16_t);
for (i=0;i<N;i++)
{
ot2[i]=tmp+2*i*nsf;
nt2[i]=tmp+(2*i+1)*nsf;
}
ot=ot2;
nt=nt2;
ALLOC(best_index, N, int);
ALLOC(best_dist, N, spx_word32_t);
ALLOC(best_nind, N, int);
ALLOC(best_ntarget, N, int);
ALLOC(ndist, N, spx_word32_t);
ALLOC(odist, N, spx_word32_t);
ALLOC(itmp, 2*N*nb_subvect, int);
for (i=0;i<N;i++)
{
nind[i]=itmp+2*i*nb_subvect;
oind[i]=itmp+(2*i+1)*nb_subvect;
}
SPEEX_COPY(t, target, nsf);
for (j=0;j<N;j++)
SPEEX_COPY(&ot[j][0], t, nsf);
/* 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]=VERY_LARGE32;
/* This is not strictly necessary, but it provides an additonal safety
to prevent crashes in case something goes wrong in the previous
steps (e.g. NaNs) */
for (j=0;j<N;j++)
best_nind[j] = best_ntarget[j] = 0;
/*For all n-bests of previous subvector*/
for (j=0;j<N;j++)
{
spx_word16_t *x=ot[j]+subvect_size*i;
spx_word32_t tener = 0;
for (m=0;m<subvect_size;m++)
tener = MAC16_16(tener, x[m],x[m]);
#ifdef FIXED_POINT
tener = SHR32(tener,1);
#else
tener *= .5;
#endif
/*Find new n-best based on previous n-best j*/
#ifndef DISABLE_WIDEBAND
if (have_sign)
vq_nbest_sign(x, resp2, subvect_size, shape_cb_size, E, N, best_index, best_dist, stack);
else
#endif /* DISABLE_WIDEBAND */
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++)
{
/* Compute total distance (including previous sub-vectors */
spx_word32_t err = ADD32(ADD32(odist[j],best_dist[k]),tener);
/*update n-best list*/
if (err<ndist[N-1])
{
for (m=0;m<N;m++)
{
if (err < ndist[m])
{
for (n=N-1;n>m;n--)
{
ndist[n] = ndist[n-1];
best_nind[n] = best_nind[n-1];
best_ntarget[n] = best_ntarget[n-1];
}
/* n is equal to m here, so they're interchangeable */
ndist[m] = err;
best_nind[n] = best_index[k];
best_ntarget[n] = j;
break;
}
}
}
}
if (i==0)
break;
}
for (j=0;j<N;j++)
{
/*previous target (we don't care what happened before*/
for (m=(i+1)*subvect_size;m<nsf;m++)
nt[j][m]=ot[best_ntarget[j]][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_nind[j];
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];
#else
g=sign*0.03125*shape_cb[rind*subvect_size+m];
#endif
target_update(nt[j]+subvect_size*(i+1), g, r+q, nsf-subvect_size*(i+1));
}
for (q=0;q<nb_subvect;q++)
nind[j][q]=oind[best_ntarget[j]][q];
nind[j][i]=best_nind[j];
}
/*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]=SHL32(EXTEND32(shape_cb[rind*subvect_size+j]),SIG_SHIFT-5);
} else {
for (j=0;j<subvect_size;j++)
e[subvect_size*i+j]=NEG32(SHL32(EXTEND32(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]=ADD32(exc[j],e[j]);
/* Update target: only update target if necessary */
if (update_target)
{
VARDECL(spx_word16_t *r2);
ALLOC(r2, nsf, spx_word16_t);
for (j=0;j<nsf;j++)
r2[j] = EXTRACT16(PSHR32(e[j] ,6));
syn_percep_zero16(r2, ak, awk1, awk2, r2, nsf,p, stack);
for (j=0;j<nsf;j++)
target[j]=SUB16(target[j],PSHR16(r2[j],2));
}
}
static void split_cb_search_shape_sign_N1(
spx_word16_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_word16_t *r,
SpeexBits *bits,
char *stack,
int update_target
)
{
int i,j,m,q;
VARDECL(spx_word16_t *resp);
#ifdef _USE_SSE
VARDECL(__m128 *resp2);
VARDECL(__m128 *E);
#else
spx_word16_t *resp2;
VARDECL(spx_word32_t *E);
#endif
VARDECL(spx_word16_t *t);
VARDECL(spx_sig_t *e);
const signed char *shape_cb;
int shape_cb_size, subvect_size, nb_subvect;
const split_cb_params *params;
int best_index;
spx_word32_t best_dist;
int have_sign;
params = (const 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;
ALLOC(resp, shape_cb_size*subvect_size, spx_word16_t);
#ifdef _USE_SSE
ALLOC(resp2, (shape_cb_size*subvect_size)>>2, __m128);
ALLOC(E, shape_cb_size>>2, __m128);
#else
resp2 = resp;
ALLOC(E, shape_cb_size, spx_word32_t);
#endif
ALLOC(t, nsf, spx_word16_t);
ALLOC(e, nsf, spx_sig_t);
/* FIXME: Do we still need to copy the target? */
SPEEX_COPY(t, target, nsf);
compute_weighted_codebook(shape_cb, r, resp, resp2, E, shape_cb_size, subvect_size, stack);
for (i=0;i<nb_subvect;i++)
{
spx_word16_t *x=t+subvect_size*i;
/*Find new n-best based on previous n-best j*/
#ifndef DISABLE_WIDEBAND
if (have_sign)
vq_nbest_sign(x, resp2, subvect_size, shape_cb_size, E, 1, &best_index, &best_dist, stack);
else
#endif /* DISABLE_WIDEBAND */
vq_nbest(x, resp2, subvect_size, shape_cb_size, E, 1, &best_index, &best_dist, stack);
speex_bits_pack(bits,best_index,params->shape_bits+have_sign);
{
int rind;
spx_word16_t *res;
spx_word16_t sign=1;
rind = best_index;
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] = SUB16(t[subvect_size*i+m], res[m]);
else
for (m=0;m<subvect_size;m++)
t[subvect_size*i+m] = ADD16(t[subvect_size*i+m], res[m]);
#ifdef FIXED_POINT
if (sign==1)
{
for (j=0;j<subvect_size;j++)
e[subvect_size*i+j]=SHL32(EXTEND32(shape_cb[rind*subvect_size+j]),SIG_SHIFT-5);
} else {
for (j=0;j<subvect_size;j++)
e[subvect_size*i+j]=NEG32(SHL32(EXTEND32(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
}
for (m=0;m<subvect_size;m++)
{
spx_word16_t g;
int rind;
spx_word16_t sign=1;
rind = best_index;
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];
#else
g=sign*0.03125*shape_cb[rind*subvect_size+m];
#endif
target_update(t+subvect_size*(i+1), g, r+q, nsf-subvect_size*(i+1));
}
}
/* Update excitation */
/* FIXME: We could update the excitation directly above */
for (j=0;j<nsf;j++)
exc[j]=ADD32(exc[j],e[j]);
/* Update target: only update target if necessary */
if (update_target)
{
VARDECL(spx_word16_t *r2);
ALLOC(r2, nsf, spx_word16_t);
for (j=0;j<nsf;j++)
r2[j] = EXTRACT16(PSHR32(e[j] ,6));
syn_percep_zero16(r2, ak, awk1, awk2, r2, nsf,p, stack);
for (j=0;j<nsf;j++)
target[j]=SUB16(target[j],PSHR16(r2[j],2));
}
}
void lsp_to_lpc(spx_lsp_t *freq,spx_coef_t *ak,int lpcrdr, char *stack)
/* float *freq array of LSP frequencies in the x domain */
/* float *ak array of LPC coefficients */
/* int lpcrdr order of LPC coefficients */
{
int i,j;
spx_word32_t xout1,xout2,xin;
spx_word32_t mult, a;
VARDECL(spx_word32_t *xpmem);
VARDECL(spx_word32_t *xqmem);
#ifndef FIXED_LPC_SIZE
VARDECL(spx_word16_t *freqn);
VARDECL(spx_word32_t **xp);
VARDECL(spx_word32_t **xq);
#else
spx_word16_t freqn[FIXED_LPC_SIZE];
spx_word32_t *xp[(FIXED_LPC_SIZE/2)+1];
spx_word32_t *xq[(FIXED_LPC_SIZE/2)+1];
#endif
int m = lpcrdr>>1;
/*
Reconstruct P(z) and Q(z) by cascading second order polynomials
in form 1 - 2cos(w)z(-1) + z(-2), where w is the LSP frequency.
In the time domain this is:
y(n) = x(n) - 2cos(w)x(n-1) + x(n-2)
This is what the ALLOCS below are trying to do:
int xp[m+1][lpcrdr+1+2]; // P matrix in QIMP
int xq[m+1][lpcrdr+1+2]; // Q matrix in QIMP
These matrices store the output of each stage on each row. The
final (m-th) row has the output of the final (m-th) cascaded
2nd order filter. The first row is the impulse input to the
system (not written as it is known).
The version below takes advantage of the fact that a lot of the
outputs are zero or known, for example if we put an inpulse
into the first section the "clock" it 10 times only the first 3
outputs samples are non-zero (it's an FIR filter).
*/
#ifndef FIXED_LPC_SIZE
ALLOC(xp, (m+1), spx_word32_t*);
#endif
ALLOC(xpmem, (m+1)*(lpcrdr+1+2), spx_word32_t);
#ifndef FIXED_LPC_SIZE
ALLOC(xq, (m+1), spx_word32_t*);
#endif
ALLOC(xqmem, (m+1)*(lpcrdr+1+2), spx_word32_t);
#ifndef FIXED_LPC_SIZE
for(i=0; i<=m; i++) {
xp[i] = xpmem + i*(lpcrdr+1+2);
xq[i] = xqmem + i*(lpcrdr+1+2);
}
#else
for(i=0; i<=m; i++) {
xp[i] = xpmem + i*(FIXED_LPC_SIZE+1+2);
xq[i] = xqmem + i*(FIXED_LPC_SIZE+1+2);
}
#endif
/* work out 2cos terms in Q14 */
#ifndef FIXED_LPC_SIZE
ALLOC(freqn, lpcrdr, spx_word16_t);
for (i=0;i<lpcrdr;i++)
freqn[i] = ANGLE2X(freq[i]);
#else
for (i=0;i<FIXED_LPC_SIZE;i++)
freqn[i] = ANGLE2X(freq[i]);
#endif
#define QIMP 21 /* scaling for impulse */
xin = SHL32(EXTEND32(1), (QIMP-1)); /* 0.5 in QIMP format */
/* first col and last non-zero values of each row are trivial */
for(i=0;i<=m;i++) {
xp[i][1] = 0;
xp[i][2] = xin;
xp[i][2+2*i] = xin;
xq[i][1] = 0;
xq[i][2] = xin;
xq[i][2+2*i] = xin;
}
/* 2nd row (first output row) is trivial */
xp[1][3] = -MULT16_32_Q14(freqn[0],xp[0][2]);
xq[1][3] = -MULT16_32_Q14(freqn[1],xq[0][2]);
xout1 = xout2 = 0;
/* now generate remaining rows */
for(i=1;i<m;i++) {
for(j=1;j<2*(i+1)-1;j++) {
mult = MULT16_32_Q14(freqn[2*i],xp[i][j+1]);
xp[i+1][j+2] = ADD32(SUB32(xp[i][j+2], mult), xp[i][j]);
mult = MULT16_32_Q14(freqn[2*i+1],xq[i][j+1]);
xq[i+1][j+2] = ADD32(SUB32(xq[i][j+2], mult), xq[i][j]);
}
/* for last col xp[i][j+2] = xq[i][j+2] = 0 */
mult = MULT16_32_Q14(freqn[2*i],xp[i][j+1]);
xp[i+1][j+2] = SUB32(xp[i][j], mult);
mult = MULT16_32_Q14(freqn[2*i+1],xq[i][j+1]);
xq[i+1][j+2] = SUB32(xq[i][j], mult);
}
/* process last row to extra a{k} */
#ifndef FIXED_LPC_SIZE
for(j=1;j<=lpcrdr;j++) {
#else
for(j=1;j<=FIXED_LPC_SIZE;j++) {
#endif
int shift = QIMP-13;
/* final filter sections */
a = PSHR32(xp[m][j+2] + xout1 + xq[m][j+2] - xout2, shift);
xout1 = xp[m][j+2];
xout2 = xq[m][j+2];
/* hard limit ak's to +/- 32767 */
if (a < -32767) a = -32767;
if (a > 32767) a = 32767;
ak[j-1] = (short)a;
}
}
#else
void lsp_to_lpc(spx_lsp_t *freq,spx_coef_t *ak,int lpcrdr, char *stack)
/* float *freq array of LSP frequencies in the x domain */
/* float *ak array of LPC coefficients */
/* int lpcrdr order of LPC coefficients */
{
int i,j;
float xout1,xout2,xin1,xin2;
VARDECL(float *Wp);
float *pw,*n1,*n2,*n3,*n4=NULL;
VARDECL(float *x_freq);
int m = lpcrdr>>1;
ALLOC(Wp, 4*m+2, float);
pw = Wp;
/* initialise contents of array */
for(i=0;i<=4*m+1;i++){ /* set contents of buffer to 0 */
*pw++ = 0.0;
}
/* Set pointers up */
pw = Wp;
xin1 = 1.0;
xin2 = 1.0;
ALLOC(x_freq, lpcrdr, float);
for (i=0;i<lpcrdr;i++)
x_freq[i] = ANGLE2X(freq[i]);
/* reconstruct P(z) and Q(z) by cascading second order
polynomials in form 1 - 2xz(-1) +z(-2), where x is the
LSP coefficient */
for(j=0;j<=lpcrdr;j++){
int i2=0;
for(i=0;i<m;i++,i2+=2){
n1 = pw+(i*4);
n2 = n1 + 1;
n3 = n2 + 1;
n4 = n3 + 1;
xout1 = xin1 - 2.f*x_freq[i2] * *n1 + *n2;
xout2 = xin2 - 2.f*x_freq[i2+1] * *n3 + *n4;
*n2 = *n1;
*n4 = *n3;
*n1 = xin1;
*n3 = xin2;
xin1 = xout1;
xin2 = xout2;
}
xout1 = xin1 + *(n4+1);
xout2 = xin2 - *(n4+2);
if (j>0)
ak[j-1] = (xout1 + xout2)*0.5f;
*(n4+1) = xin1;
*(n4+2) = xin2;
xin1 = 0.0;
xin2 = 0.0;
}
}