static int32
get_scores_8b_feat_3(s2_semi_mgau_t * s, int i,
int16 *senone_scores, uint8 *senone_active,
int32 n_senone_active)
{
int32 j, l;
uint8 *pid_cw0, *pid_cw1, *pid_cw2;
pid_cw0 = s->mixw[i][s->f[i][0].codeword];
pid_cw1 = s->mixw[i][s->f[i][1].codeword];
pid_cw2 = s->mixw[i][s->f[i][2].codeword];
for (l = j = 0; j < n_senone_active; j++) {
int sen = senone_active[j] + l;
int32 tmp = pid_cw0[sen] + s->f[i][0].score;
tmp = fast_logmath_add(s->lmath_8b, tmp,
pid_cw1[sen] + s->f[i][1].score);
tmp = fast_logmath_add(s->lmath_8b, tmp,
pid_cw2[sen] + s->f[i][2].score);
senone_scores[sen] += tmp;
l = sen;
}
return 0;
}
static int32
get_scores_8b_feat_3(s2_semi_mgau_t * s, int i,
int32 *senone_scores, uint8 *senone_active)
{
int32 j;
uint8 *pid_cw0, *pid_cw1, *pid_cw2;
pid_cw0 = s->mixw[i][s->f[i][0].codeword];
pid_cw1 = s->mixw[i][s->f[i][1].codeword];
pid_cw2 = s->mixw[i][s->f[i][2].codeword];
for (j = 0; j < s->n_sen; j++) {
int32 tmp;
if (!senone_active[j]) continue;
tmp = pid_cw0[j] + s->f[i][0].score;
tmp = fast_logmath_add(s->lmath_8b, tmp,
pid_cw1[j] + s->f[i][1].score);
tmp = fast_logmath_add(s->lmath_8b, tmp,
pid_cw2[j] + s->f[i][2].score);
senone_scores[j] -= (tmp << SENSCR_SHIFT);
}
return 0;
}
static int32
get_scores_4b_feat_6(s2_semi_mgau_t * s, int i,
int16 *senone_scores, uint8 *senone_active,
int32 n_senone_active)
{
int32 j, l;
uint8 *pid_cw0, *pid_cw1, *pid_cw2, *pid_cw3, *pid_cw4, *pid_cw5;
uint8 w_den[6][16];
/* Precompute scaled densities. */
for (j = 0; j < 16; ++j) {
w_den[0][j] = s->mixw_cb[j] + s->f[i][0].score;
w_den[1][j] = s->mixw_cb[j] + s->f[i][1].score;
w_den[2][j] = s->mixw_cb[j] + s->f[i][2].score;
w_den[3][j] = s->mixw_cb[j] + s->f[i][3].score;
w_den[4][j] = s->mixw_cb[j] + s->f[i][4].score;
w_den[5][j] = s->mixw_cb[j] + s->f[i][5].score;
}
pid_cw0 = s->mixw[i][s->f[i][0].codeword];
pid_cw1 = s->mixw[i][s->f[i][1].codeword];
pid_cw2 = s->mixw[i][s->f[i][2].codeword];
pid_cw3 = s->mixw[i][s->f[i][3].codeword];
pid_cw4 = s->mixw[i][s->f[i][4].codeword];
pid_cw5 = s->mixw[i][s->f[i][5].codeword];
for (l = j = 0; j < n_senone_active; j++) {
int n = senone_active[j] + l;
int tmp, cw;
if (n & 1) {
cw = pid_cw0[n/2] >> 4;
tmp = w_den[0][cw];
cw = pid_cw1[n/2] >> 4;
tmp = fast_logmath_add(s->lmath_8b, tmp, w_den[1][cw]);
cw = pid_cw2[n/2] >> 4;
tmp = fast_logmath_add(s->lmath_8b, tmp, w_den[2][cw]);
cw = pid_cw3[n/2] >> 4;
tmp = fast_logmath_add(s->lmath_8b, tmp, w_den[3][cw]);
cw = pid_cw4[n/2] >> 4;
tmp = fast_logmath_add(s->lmath_8b, tmp, w_den[4][cw]);
cw = pid_cw5[n/2] >> 4;
tmp = fast_logmath_add(s->lmath_8b, tmp, w_den[5][cw]);
}
else {
static int32
get_scores_8b_feat_all(s2_semi_mgau_t * s, int i, int topn, int16 *senone_scores)
{
int32 j, k;
for (j = 0; j < s->n_sen; j++) {
uint8 *pid_cw;
int32 tmp;
pid_cw = s->mixw[i][s->f[i][0].codeword];
tmp = pid_cw[j] + s->f[i][0].score;
for (k = 1; k < topn; ++k) {
pid_cw = s->mixw[i][s->f[i][k].codeword];
tmp = fast_logmath_add(s->lmath_8b, tmp,
pid_cw[j] + s->f[i][k].score);
}
senone_scores[j] += tmp;
}
return 0;
}
static int32
get_scores_8b_feat_any(s2_semi_mgau_t * s, int i, int topn,
int32 *senone_scores, uint8 *senone_active)
{
int32 j, k;
for (j = 0; j < s->n_sen; j++) {
uint8 *pid_cw;
int32 tmp;
if (!senone_active[j]) continue;
pid_cw = s->mixw[i][s->f[i][0].codeword];
tmp = pid_cw[j] + s->f[i][0].score;
for (k = 1; k < topn; ++k) {
pid_cw = s->mixw[i][s->f[i][k].codeword];
tmp = fast_logmath_add(s->lmath_8b, tmp,
pid_cw[j] + s->f[i][k].score);
}
senone_scores[j] -= (tmp << SENSCR_SHIFT);
}
return 0;
}
static int32
get_scores_8b_feat_any(s2_semi_mgau_t * s, int i, int topn,
int16 *senone_scores, uint8 *senone_active,
int32 n_senone_active)
{
int32 j, k, l;
for (l = j = 0; j < n_senone_active; j++) {
int sen = senone_active[j] + l;
uint8 *pid_cw;
int32 tmp;
pid_cw = s->mixw[i][s->f[i][0].codeword];
tmp = pid_cw[sen] + s->f[i][0].score;
for (k = 1; k < topn; ++k) {
pid_cw = s->mixw[i][s->f[i][k].codeword];
tmp = fast_logmath_add(s->lmath_8b, tmp,
pid_cw[sen] + s->f[i][k].score);
}
senone_scores[sen] += tmp;
l = sen;
}
return 0;
}
cw = pid_cw1[n/2] >> 4;
tmp = fast_logmath_add(s->lmath_8b, tmp, w_den[1][cw]);
cw = pid_cw2[n/2] >> 4;
tmp = fast_logmath_add(s->lmath_8b, tmp, w_den[2][cw]);
cw = pid_cw3[n/2] >> 4;
tmp = fast_logmath_add(s->lmath_8b, tmp, w_den[3][cw]);
cw = pid_cw4[n/2] >> 4;
tmp = fast_logmath_add(s->lmath_8b, tmp, w_den[4][cw]);
cw = pid_cw5[n/2] >> 4;
tmp = fast_logmath_add(s->lmath_8b, tmp, w_den[5][cw]);
}
else {
cw = pid_cw0[n/2] & 0x0f;
tmp = w_den[0][cw];
cw = pid_cw1[n/2] & 0x0f;
tmp = fast_logmath_add(s->lmath_8b, tmp, w_den[1][cw]);
cw = pid_cw2[n/2] & 0x0f;
tmp = fast_logmath_add(s->lmath_8b, tmp, w_den[2][cw]);
cw = pid_cw3[n/2] & 0x0f;
tmp = fast_logmath_add(s->lmath_8b, tmp, w_den[3][cw]);
cw = pid_cw4[n/2] & 0x0f;
tmp = fast_logmath_add(s->lmath_8b, tmp, w_den[4][cw]);
cw = pid_cw5[n/2] & 0x0f;
tmp = fast_logmath_add(s->lmath_8b, tmp, w_den[5][cw]);
}
senone_scores[n] += tmp;
l = n;
}
return 0;
}
