/* Find least-squares prediction gain for one signal based on another and quantize it */
int32_t silk_stereo_find_predictor( /* O Returns predictor in Q13 */
int32_t * ratio_Q14, /* O Ratio of residual and mid energies */
const int16_t x[], /* I Basis signal */
const int16_t y[], /* I Target signal */
int32_t mid_res_amp_Q0[], /* I/O Smoothed mid, residual norms */
int length, /* I Number of samples */
int smooth_coef_Q16 /* I Smoothing coefficient */
)
{
int scale, scale1, scale2;
int32_t nrgx, nrgy, corr, pred_Q13, pred2_Q10;
/* Find predictor */
silk_sum_sqr_shift(&nrgx, &scale1, x, length);
silk_sum_sqr_shift(&nrgy, &scale2, y, length);
scale = silk_max_int(scale1, scale2);
scale = scale + (scale & 1); /* make even */
nrgy = silk_RSHIFT32(nrgy, scale - scale2);
nrgx = silk_RSHIFT32(nrgx, scale - scale1);
nrgx = silk_max_int(nrgx, 1);
corr = silk_inner_prod_aligned_scale(x, y, scale, length);
pred_Q13 = silk_DIV32_varQ(corr, nrgx, 13);
pred_Q13 = silk_LIMIT(pred_Q13, -(1 << 14), 1 << 14);
pred2_Q10 = silk_SMULWB(pred_Q13, pred_Q13);
/* Faster update for signals with large prediction parameters */
smooth_coef_Q16 =
(int) silk_max_int(smooth_coef_Q16, silk_abs(pred2_Q10));
/* Smoothed mid and residual norms */
assert(smooth_coef_Q16 < 32768);
scale = silk_RSHIFT(scale, 1);
mid_res_amp_Q0[0] =
silk_SMLAWB(mid_res_amp_Q0[0],
silk_LSHIFT(silk_SQRT_APPROX(nrgx),
scale) - mid_res_amp_Q0[0],
smooth_coef_Q16);
/* Residual energy = nrgy - 2 * pred * corr + pred^2 * nrgx */
nrgy = silk_SUB_LSHIFT32(nrgy, silk_SMULWB(corr, pred_Q13), 3 + 1);
nrgy = silk_ADD_LSHIFT32(nrgy, silk_SMULWB(nrgx, pred2_Q10), 6);
mid_res_amp_Q0[1] =
silk_SMLAWB(mid_res_amp_Q0[1],
silk_LSHIFT(silk_SQRT_APPROX(nrgy),
scale) - mid_res_amp_Q0[1],
smooth_coef_Q16);
/* Ratio of smoothed residual and mid norms */
*ratio_Q14 =
silk_DIV32_varQ(mid_res_amp_Q0[1], silk_max(mid_res_amp_Q0[0], 1),
14);
*ratio_Q14 = silk_LIMIT(*ratio_Q14, 0, 32767);
return pred_Q13;
}
/* Calculates correlation vector X'*t */
void silk_corrVector_FIX(
const opus_int16 *x, /* I x vector [L + order - 1] used to form data matrix X */
const opus_int16 *t, /* I Target vector [L] */
const opus_int L, /* I Length of vectors */
const opus_int order, /* I Max lag for correlation */
opus_int32 *Xt, /* O Pointer to X'*t correlation vector [order] */
const opus_int rshifts /* I Right shifts of correlations */
)
{
opus_int lag, i;
const opus_int16 *ptr1, *ptr2;
opus_int32 inner_prod;
ptr1 = &x[ order - 1 ]; /* Points to first sample of column 0 of X: X[:,0] */
ptr2 = t;
/* Calculate X'*t */
if( rshifts > 0 ) {
/* Right shifting used */
for( lag = 0; lag < order; lag++ ) {
inner_prod = 0;
for( i = 0; i < L; i++ ) {
inner_prod += silk_RSHIFT32( silk_SMULBB( ptr1[ i ], ptr2[i] ), rshifts );
}
Xt[ lag ] = inner_prod; /* X[:,lag]'*t */
ptr1--; /* Go to next column of X */
}
} else {
silk_assert( rshifts == 0 );
for( lag = 0; lag < order; lag++ ) {
Xt[ lag ] = silk_inner_prod_aligned( ptr1, ptr2, L ); /* X[:,lag]'*t */
ptr1--; /* Go to next column of X */
}
}
}
/* For input in Q24 domain */
opus_int32 silk_LPC_inverse_pred_gain_Q24( /* O Returns inverse prediction gain in energy domain, Q30 */
const opus_int32 *A_Q24, /* I Prediction coefficients [order] */
const opus_int order /* I Prediction order */
)
{
opus_int k;
opus_int32 Atmp_QA[ SILK_MAX_ORDER_LPC ];
/* Increase Q domain of the AR coefficients */
for( k = 0; k < order; k++ ) {
Atmp_QA[ k ] = silk_RSHIFT32( A_Q24[ k ], 24 - QA );
}
return LPC_inverse_pred_gain_QA( Atmp_QA, order );
}
/* Convert int32 coefficients to int16 coefs and make sure there's no wrap-around */
void silk_LPC_fit(
opus_int16 *a_QOUT, /* O Output signal */
opus_int32 *a_QIN, /* I/O Input signal */
const opus_int QOUT, /* I Input Q domain */
const opus_int QIN, /* I Input Q domain */
const opus_int d /* I Filter order */
)
{
opus_int i, k, idx = 0;
opus_int32 maxabs, absval, chirp_Q16;
/* Limit the maximum absolute value of the prediction coefficients, so that they'll fit in int16 */
for( i = 0; i < 10; i++ ) {
/* Find maximum absolute value and its index */
maxabs = 0;
for( k = 0; k < d; k++ ) {
absval = silk_abs( a_QIN[k] );
if( absval > maxabs ) {
maxabs = absval;
idx = k;
}
}
maxabs = silk_RSHIFT_ROUND( maxabs, QIN - QOUT );
if( maxabs > silk_int16_MAX ) {
/* Reduce magnitude of prediction coefficients */
maxabs = silk_min( maxabs, 163838 ); /* ( silk_int32_MAX >> 14 ) + silk_int16_MAX = 163838 */
chirp_Q16 = SILK_FIX_CONST( 0.999, 16 ) - silk_DIV32( silk_LSHIFT( maxabs - silk_int16_MAX, 14 ),
silk_RSHIFT32( silk_MUL( maxabs, idx + 1), 2 ) );
silk_bwexpander_32( a_QIN, d, chirp_Q16 );
} else {
break;
}
}
if( i == 10 ) {
/* Reached the last iteration, clip the coefficients */
for( k = 0; k < d; k++ ) {
a_QOUT[ k ] = (opus_int16)silk_SAT16( silk_RSHIFT_ROUND( a_QIN[ k ], QIN - QOUT ) );
a_QIN[ k ] = silk_LSHIFT( (opus_int32)a_QOUT[ k ], QIN - QOUT );
}
} else {
for( k = 0; k < d; k++ ) {
a_QOUT[ k ] = (opus_int16)silk_RSHIFT_ROUND( a_QIN[ k ], QIN - QOUT );
}
}
}
/* For input in Q24 domain */
int32_t silk_LPC_inverse_pred_gain_Q24( /* O Returns inverse prediction gain in energy domain, Q30 */
const int32_t * A_Q24, /* I Prediction coefficients [order] */
const int order /* I Prediction order */
)
{
int k;
int32_t Atmp_QA[2][SILK_MAX_ORDER_LPC];
int32_t *Anew_QA;
memzero(Atmp_QA, (2 * SILK_MAX_ORDER_LPC) * sizeof(int32_t));
Anew_QA = Atmp_QA[order & 1];
/* Increase Q domain of the AR coefficients */
for (k = 0; k < order; k++) {
Anew_QA[k] = silk_RSHIFT32(A_Q24[k], 24 - QA);
}
return LPC_inverse_pred_gain_QA(Atmp_QA, order);
}
/* Calculates correlation matrix X'*X */
void silk_corrMatrix_FIX(
const opus_int16 *x, /* I x vector [L + order - 1] used to form data matrix X */
const opus_int L, /* I Length of vectors */
const opus_int order, /* I Max lag for correlation */
const opus_int head_room, /* I Desired headroom */
opus_int32 *XX, /* O Pointer to X'*X correlation matrix [ order x order ] */
opus_int *rshifts /* I/O Right shifts of correlations */
)
{
opus_int i, j, lag, rshifts_local, head_room_rshifts;
opus_int32 energy;
const opus_int16 *ptr1, *ptr2;
/* Calculate energy to find shift used to fit in 32 bits */
silk_sum_sqr_shift( &energy, &rshifts_local, x, L + order - 1 );
/* Add shifts to get the desired head room */
head_room_rshifts = silk_max( head_room - silk_CLZ32( energy ), 0 );
energy = silk_RSHIFT32( energy, head_room_rshifts );
rshifts_local += head_room_rshifts;
/* Calculate energy of first column (0) of X: X[:,0]'*X[:,0] */
/* Remove contribution of first order - 1 samples */
for( i = 0; i < order - 1; i++ ) {
energy -= silk_RSHIFT32( silk_SMULBB( x[ i ], x[ i ] ), rshifts_local );
}
if( rshifts_local < *rshifts ) {
/* Adjust energy */
energy = silk_RSHIFT32( energy, *rshifts - rshifts_local );
rshifts_local = *rshifts;
}
/* Calculate energy of remaining columns of X: X[:,j]'*X[:,j] */
/* Fill out the diagonal of the correlation matrix */
matrix_ptr( XX, 0, 0, order ) = energy;
ptr1 = &x[ order - 1 ]; /* First sample of column 0 of X */
for( j = 1; j < order; j++ ) {
energy = silk_SUB32( energy, silk_RSHIFT32( silk_SMULBB( ptr1[ L - j ], ptr1[ L - j ] ), rshifts_local ) );
energy = silk_ADD32( energy, silk_RSHIFT32( silk_SMULBB( ptr1[ -j ], ptr1[ -j ] ), rshifts_local ) );
matrix_ptr( XX, j, j, order ) = energy;
}
ptr2 = &x[ order - 2 ]; /* First sample of column 1 of X */
/* Calculate the remaining elements of the correlation matrix */
if( rshifts_local > 0 ) {
/* Right shifting used */
for( lag = 1; lag < order; lag++ ) {
/* Inner product of column 0 and column lag: X[:,0]'*X[:,lag] */
energy = 0;
for( i = 0; i < L; i++ ) {
energy += silk_RSHIFT32( silk_SMULBB( ptr1[ i ], ptr2[i] ), rshifts_local );
}
/* Calculate remaining off diagonal: X[:,j]'*X[:,j + lag] */
matrix_ptr( XX, lag, 0, order ) = energy;
matrix_ptr( XX, 0, lag, order ) = energy;
for( j = 1; j < ( order - lag ); j++ ) {
energy = silk_SUB32( energy, silk_RSHIFT32( silk_SMULBB( ptr1[ L - j ], ptr2[ L - j ] ), rshifts_local ) );
energy = silk_ADD32( energy, silk_RSHIFT32( silk_SMULBB( ptr1[ -j ], ptr2[ -j ] ), rshifts_local ) );
matrix_ptr( XX, lag + j, j, order ) = energy;
matrix_ptr( XX, j, lag + j, order ) = energy;
}
ptr2--; /* Update pointer to first sample of next column (lag) in X */
}
} else {
for( lag = 1; lag < order; lag++ ) {
/* Inner product of column 0 and column lag: X[:,0]'*X[:,lag] */
energy = silk_inner_prod_aligned( ptr1, ptr2, L );
matrix_ptr( XX, lag, 0, order ) = energy;
matrix_ptr( XX, 0, lag, order ) = energy;
/* Calculate remaining off diagonal: X[:,j]'*X[:,j + lag] */
for( j = 1; j < ( order - lag ); j++ ) {
energy = silk_SUB32( energy, silk_SMULBB( ptr1[ L - j ], ptr2[ L - j ] ) );
energy = silk_SMLABB( energy, ptr1[ -j ], ptr2[ -j ] );
matrix_ptr( XX, lag + j, j, order ) = energy;
matrix_ptr( XX, j, lag + j, order ) = energy;
}
ptr2--;/* Update pointer to first sample of next column (lag) in X */
}
}
*rshifts = rshifts_local;
}
/* Compute reflection coefficients from input signal */
void silk_burg_modified(
opus_int32 *res_nrg, /* O Residual energy */
opus_int *res_nrg_Q, /* O Residual energy Q value */
opus_int32 A_Q16[], /* O Prediction coefficients (length order) */
const opus_int16 x[], /* I Input signal, length: nb_subfr * ( D + subfr_length ) */
const opus_int subfr_length, /* I Input signal subframe length (incl. D preceeding samples) */
const opus_int nb_subfr, /* I Number of subframes stacked in x */
const opus_int32 WhiteNoiseFrac_Q32, /* I Fraction added to zero-lag autocorrelation */
const opus_int D /* I Order */
)
{
opus_int k, n, s, lz, rshifts, rshifts_extra;
opus_int32 C0, num, nrg, rc_Q31, Atmp_QA, Atmp1, tmp1, tmp2, x1, x2;
const opus_int16 *x_ptr;
opus_int32 C_first_row[ SILK_MAX_ORDER_LPC ];
opus_int32 C_last_row[ SILK_MAX_ORDER_LPC ];
opus_int32 Af_QA[ SILK_MAX_ORDER_LPC ];
opus_int32 CAf[ SILK_MAX_ORDER_LPC + 1 ];
opus_int32 CAb[ SILK_MAX_ORDER_LPC + 1 ];
silk_assert( subfr_length * nb_subfr <= MAX_FRAME_SIZE );
silk_assert( nb_subfr <= MAX_NB_SUBFR );
/* Compute autocorrelations, added over subframes */
silk_sum_sqr_shift( &C0, &rshifts, x, nb_subfr * subfr_length );
if( rshifts > MAX_RSHIFTS ) {
C0 = silk_LSHIFT32( C0, rshifts - MAX_RSHIFTS );
silk_assert( C0 > 0 );
rshifts = MAX_RSHIFTS;
} else {
lz = silk_CLZ32( C0 ) - 1;
rshifts_extra = N_BITS_HEAD_ROOM - lz;
if( rshifts_extra > 0 ) {
rshifts_extra = silk_min( rshifts_extra, MAX_RSHIFTS - rshifts );
C0 = silk_RSHIFT32( C0, rshifts_extra );
} else {
rshifts_extra = silk_max( rshifts_extra, MIN_RSHIFTS - rshifts );
C0 = silk_LSHIFT32( C0, -rshifts_extra );
}
rshifts += rshifts_extra;
}
silk_memset( C_first_row, 0, SILK_MAX_ORDER_LPC * sizeof( opus_int32 ) );
if( rshifts > 0 ) {
for( s = 0; s < nb_subfr; s++ ) {
x_ptr = x + s * subfr_length;
for( n = 1; n < D + 1; n++ ) {
C_first_row[ n - 1 ] += (opus_int32)silk_RSHIFT64(
silk_inner_prod16_aligned_64( x_ptr, x_ptr + n, subfr_length - n ), rshifts );
}
}
} else {
for( s = 0; s < nb_subfr; s++ ) {
x_ptr = x + s * subfr_length;
for( n = 1; n < D + 1; n++ ) {
C_first_row[ n - 1 ] += silk_LSHIFT32(
silk_inner_prod_aligned( x_ptr, x_ptr + n, subfr_length - n ), -rshifts );
}
}
}
silk_memcpy( C_last_row, C_first_row, SILK_MAX_ORDER_LPC * sizeof( opus_int32 ) );
/* Initialize */
CAb[ 0 ] = CAf[ 0 ] = C0 + silk_SMMUL( WhiteNoiseFrac_Q32, C0 ) + 1; /* Q(-rshifts)*/
for( n = 0; n < D; n++ ) {
/* Update first row of correlation matrix (without first element) */
/* Update last row of correlation matrix (without last element, stored in reversed order) */
/* Update C * Af */
/* Update C * flipud(Af) (stored in reversed order) */
if( rshifts > -2 ) {
for( s = 0; s < nb_subfr; s++ ) {
x_ptr = x + s * subfr_length;
x1 = -silk_LSHIFT32( (opus_int32)x_ptr[ n ], 16 - rshifts ); /* Q(16-rshifts)*/
x2 = -silk_LSHIFT32( (opus_int32)x_ptr[ subfr_length - n - 1 ], 16 - rshifts ); /* Q(16-rshifts)*/
tmp1 = silk_LSHIFT32( (opus_int32)x_ptr[ n ], QA - 16 ); /* Q(QA-16)*/
tmp2 = silk_LSHIFT32( (opus_int32)x_ptr[ subfr_length - n - 1 ], QA - 16 ); /* Q(QA-16)*/
for( k = 0; k < n; k++ ) {
C_first_row[ k ] = silk_SMLAWB( C_first_row[ k ], x1, x_ptr[ n - k - 1 ] ); /* Q( -rshifts )*/
C_last_row[ k ] = silk_SMLAWB( C_last_row[ k ], x2, x_ptr[ subfr_length - n + k ] ); /* Q( -rshifts )*/
Atmp_QA = Af_QA[ k ];
tmp1 = silk_SMLAWB( tmp1, Atmp_QA, x_ptr[ n - k - 1 ] ); /* Q(QA-16)*/
tmp2 = silk_SMLAWB( tmp2, Atmp_QA, x_ptr[ subfr_length - n + k ] ); /* Q(QA-16)*/
}
tmp1 = silk_LSHIFT32( -tmp1, 32 - QA - rshifts ); /* Q(16-rshifts)*/
tmp2 = silk_LSHIFT32( -tmp2, 32 - QA - rshifts ); /* Q(16-rshifts)*/
for( k = 0; k <= n; k++ ) {
CAf[ k ] = silk_SMLAWB( CAf[ k ], tmp1, x_ptr[ n - k ] ); /* Q( -rshift )*/
CAb[ k ] = silk_SMLAWB( CAb[ k ], tmp2, x_ptr[ subfr_length - n + k - 1 ] ); /* Q( -rshift )*/
}
}
} else {
for( s = 0; s < nb_subfr; s++ ) {
x_ptr = x + s * subfr_length;
x1 = -silk_LSHIFT32( (opus_int32)x_ptr[ n ], -rshifts ); /* Q( -rshifts )*/
x2 = -silk_LSHIFT32( (opus_int32)x_ptr[ subfr_length - n - 1 ], -rshifts ); /* Q( -rshifts )*/
tmp1 = silk_LSHIFT32( (opus_int32)x_ptr[ n ], 17 ); /* Q17*/
tmp2 = silk_LSHIFT32( (opus_int32)x_ptr[ subfr_length - n - 1 ], 17 ); /* Q17*/
for( k = 0; k < n; k++ ) {
C_first_row[ k ] = silk_MLA( C_first_row[ k ], x1, x_ptr[ n - k - 1 ] ); /* Q( -rshifts )*/
C_last_row[ k ] = silk_MLA( C_last_row[ k ], x2, x_ptr[ subfr_length - n + k ] ); /* Q( -rshifts )*/
Atmp1 = silk_RSHIFT_ROUND( Af_QA[ k ], QA - 17 ); /* Q17*/
tmp1 = silk_MLA( tmp1, x_ptr[ n - k - 1 ], Atmp1 ); /* Q17*/
tmp2 = silk_MLA( tmp2, x_ptr[ subfr_length - n + k ], Atmp1 ); /* Q17*/
}
tmp1 = -tmp1; /* Q17*/
tmp2 = -tmp2; /* Q17*/
for( k = 0; k <= n; k++ ) {
CAf[ k ] = silk_SMLAWW( CAf[ k ], tmp1,
silk_LSHIFT32( (opus_int32)x_ptr[ n - k ], -rshifts - 1 ) ); /* Q( -rshift )*/
CAb[ k ] = silk_SMLAWW( CAb[ k ], tmp2,
silk_LSHIFT32( (opus_int32)x_ptr[ subfr_length - n + k - 1 ], -rshifts - 1 ) ); /* Q( -rshift )*/
}
}
}
/* Calculate nominator and denominator for the next order reflection (parcor) coefficient */
tmp1 = C_first_row[ n ]; /* Q( -rshifts )*/
tmp2 = C_last_row[ n ]; /* Q( -rshifts )*/
num = 0; /* Q( -rshifts )*/
nrg = silk_ADD32( CAb[ 0 ], CAf[ 0 ] ); /* Q( 1-rshifts )*/
for( k = 0; k < n; k++ ) {
Atmp_QA = Af_QA[ k ];
lz = silk_CLZ32( silk_abs( Atmp_QA ) ) - 1;
lz = silk_min( 32 - QA, lz );
Atmp1 = silk_LSHIFT32( Atmp_QA, lz ); /* Q( QA + lz )*/
tmp1 = silk_ADD_LSHIFT32( tmp1, silk_SMMUL( C_last_row[ n - k - 1 ], Atmp1 ), 32 - QA - lz ); /* Q( -rshifts )*/
tmp2 = silk_ADD_LSHIFT32( tmp2, silk_SMMUL( C_first_row[ n - k - 1 ], Atmp1 ), 32 - QA - lz ); /* Q( -rshifts )*/
num = silk_ADD_LSHIFT32( num, silk_SMMUL( CAb[ n - k ], Atmp1 ), 32 - QA - lz ); /* Q( -rshifts )*/
nrg = silk_ADD_LSHIFT32( nrg, silk_SMMUL( silk_ADD32( CAb[ k + 1 ], CAf[ k + 1 ] ),
Atmp1 ), 32 - QA - lz ); /* Q( 1-rshifts )*/
}
CAf[ n + 1 ] = tmp1; /* Q( -rshifts )*/
CAb[ n + 1 ] = tmp2; /* Q( -rshifts )*/
num = silk_ADD32( num, tmp2 ); /* Q( -rshifts )*/
num = silk_LSHIFT32( -num, 1 ); /* Q( 1-rshifts )*/
/* Calculate the next order reflection (parcor) coefficient */
if( silk_abs( num ) < nrg ) {
rc_Q31 = silk_DIV32_varQ( num, nrg, 31 );
} else {
/* Negative energy or ratio too high; set remaining coefficients to zero and exit loop */
silk_memset( &Af_QA[ n ], 0, ( D - n ) * sizeof( opus_int32 ) );
silk_assert( 0 );
break;
}
/* Update the AR coefficients */
for( k = 0; k < (n + 1) >> 1; k++ ) {
tmp1 = Af_QA[ k ]; /* QA*/
tmp2 = Af_QA[ n - k - 1 ]; /* QA*/
Af_QA[ k ] = silk_ADD_LSHIFT32( tmp1, silk_SMMUL( tmp2, rc_Q31 ), 1 ); /* QA*/
Af_QA[ n - k - 1 ] = silk_ADD_LSHIFT32( tmp2, silk_SMMUL( tmp1, rc_Q31 ), 1 ); /* QA*/
}
Af_QA[ n ] = silk_RSHIFT32( rc_Q31, 31 - QA ); /* QA*/
/* Update C * Af and C * Ab */
for( k = 0; k <= n + 1; k++ ) {
tmp1 = CAf[ k ]; /* Q( -rshifts )*/
tmp2 = CAb[ n - k + 1 ]; /* Q( -rshifts )*/
CAf[ k ] = silk_ADD_LSHIFT32( tmp1, silk_SMMUL( tmp2, rc_Q31 ), 1 ); /* Q( -rshifts )*/
CAb[ n - k + 1 ] = silk_ADD_LSHIFT32( tmp2, silk_SMMUL( tmp1, rc_Q31 ), 1 ); /* Q( -rshifts )*/
}
}
/* Return residual energy */
nrg = CAf[ 0 ]; /* Q( -rshifts )*/
tmp1 = 1 << 16; /* Q16*/
for( k = 0; k < D; k++ ) {
Atmp1 = silk_RSHIFT_ROUND( Af_QA[ k ], QA - 16 ); /* Q16*/
nrg = silk_SMLAWW( nrg, CAf[ k + 1 ], Atmp1 ); /* Q( -rshifts )*/
tmp1 = silk_SMLAWW( tmp1, Atmp1, Atmp1 ); /* Q16*/
A_Q16[ k ] = -Atmp1;
}
*res_nrg = silk_SMLAWW( nrg, silk_SMMUL( WhiteNoiseFrac_Q32, C0 ), -tmp1 ); /* Q( -rshifts )*/
*res_nrg_Q = -rshifts;
}
/* compute whitening filter coefficients from normalized line spectral frequencies */
void silk_NLSF2A(
opus_int16 *a_Q12, /* O monic whitening filter coefficients in Q12, [ d ] */
const opus_int16 *NLSF, /* I normalized line spectral frequencies in Q15, [ d ] */
const opus_int d /* I filter order (should be even) */
)
{
/* This ordering was found to maximize quality. It improves numerical accuracy of
silk_NLSF2A_find_poly() compared to "standard" ordering. */
static const unsigned char ordering16[16] = {
0, 15, 8, 7, 4, 11, 12, 3, 2, 13, 10, 5, 6, 9, 14, 1
};
static const unsigned char ordering10[10] = {
0, 9, 6, 3, 4, 5, 8, 1, 2, 7
};
const unsigned char *ordering;
opus_int k, i, dd;
opus_int32 cos_LSF_QA[ SILK_MAX_ORDER_LPC ];
opus_int32 P[ SILK_MAX_ORDER_LPC / 2 + 1 ], Q[ SILK_MAX_ORDER_LPC / 2 + 1 ];
opus_int32 Ptmp, Qtmp, f_int, f_frac, cos_val, delta;
opus_int32 a32_QA1[ SILK_MAX_ORDER_LPC ];
opus_int32 maxabs, absval, idx=0, sc_Q16;
silk_assert( LSF_COS_TAB_SZ_FIX == 128 );
silk_assert( d==10||d==16 );
/* convert LSFs to 2*cos(LSF), using piecewise linear curve from table */
ordering = d == 16 ? ordering16 : ordering10;
for( k = 0; k < d; k++ ) {
silk_assert(NLSF[k] >= 0 );
/* f_int on a scale 0-127 (rounded down) */
f_int = silk_RSHIFT( NLSF[k], 15 - 7 );
/* f_frac, range: 0..255 */
f_frac = NLSF[k] - silk_LSHIFT( f_int, 15 - 7 );
silk_assert(f_int >= 0);
silk_assert(f_int < LSF_COS_TAB_SZ_FIX );
/* Read start and end value from table */
cos_val = silk_LSFCosTab_FIX_Q12[ f_int ]; /* Q12 */
delta = silk_LSFCosTab_FIX_Q12[ f_int + 1 ] - cos_val; /* Q12, with a range of 0..200 */
/* Linear interpolation */
cos_LSF_QA[ordering[k]] = silk_RSHIFT_ROUND( silk_LSHIFT( cos_val, 8 ) + silk_MUL( delta, f_frac ), 20 - QA ); /* QA */
}
dd = silk_RSHIFT( d, 1 );
/* generate even and odd polynomials using convolution */
silk_NLSF2A_find_poly( P, &cos_LSF_QA[ 0 ], dd );
silk_NLSF2A_find_poly( Q, &cos_LSF_QA[ 1 ], dd );
/* convert even and odd polynomials to opus_int32 Q12 filter coefs */
for( k = 0; k < dd; k++ ) {
Ptmp = P[ k+1 ] + P[ k ];
Qtmp = Q[ k+1 ] - Q[ k ];
/* the Ptmp and Qtmp values at this stage need to fit in int32 */
a32_QA1[ k ] = -Qtmp - Ptmp; /* QA+1 */
a32_QA1[ d-k-1 ] = Qtmp - Ptmp; /* QA+1 */
}
/* Limit the maximum absolute value of the prediction coefficients, so that they'll fit in int16 */
for( i = 0; i < 10; i++ ) {
/* Find maximum absolute value and its index */
maxabs = 0;
for( k = 0; k < d; k++ ) {
absval = silk_abs( a32_QA1[k] );
if( absval > maxabs ) {
maxabs = absval;
idx = k;
}
}
maxabs = silk_RSHIFT_ROUND( maxabs, QA + 1 - 12 ); /* QA+1 -> Q12 */
if( maxabs > silk_int16_MAX ) {
/* Reduce magnitude of prediction coefficients */
maxabs = silk_min( maxabs, 163838 ); /* ( silk_int32_MAX >> 14 ) + silk_int16_MAX = 163838 */
sc_Q16 = SILK_FIX_CONST( 0.999, 16 ) - silk_DIV32( silk_LSHIFT( maxabs - silk_int16_MAX, 14 ),
silk_RSHIFT32( silk_MUL( maxabs, idx + 1), 2 ) );
silk_bwexpander_32( a32_QA1, d, sc_Q16 );
} else {
break;
}
}
if( i == 10 ) {
/* Reached the last iteration, clip the coefficients */
for( k = 0; k < d; k++ ) {
a_Q12[ k ] = (opus_int16)silk_SAT16( silk_RSHIFT_ROUND( a32_QA1[ k ], QA + 1 - 12 ) ); /* QA+1 -> Q12 */
a32_QA1[ k ] = silk_LSHIFT( (opus_int32)a_Q12[ k ], QA + 1 - 12 );
}
} else {
for( k = 0; k < d; k++ ) {
a_Q12[ k ] = (opus_int16)silk_RSHIFT_ROUND( a32_QA1[ k ], QA + 1 - 12 ); /* QA+1 -> Q12 */
}
}
for( i = 0; i < MAX_LPC_STABILIZE_ITERATIONS; i++ ) {
if( silk_LPC_inverse_pred_gain( a_Q12, d ) < SILK_FIX_CONST( 1.0 / MAX_PREDICTION_POWER_GAIN, 30 ) ) {
/* Prediction coefficients are (too close to) unstable; apply bandwidth expansion */
/* on the unscaled coefficients, convert to Q12 and measure again */
silk_bwexpander_32( a32_QA1, d, 65536 - silk_LSHIFT( 2, i ) );
for( k = 0; k < d; k++ ) {
a_Q12[ k ] = (opus_int16)silk_RSHIFT_ROUND( a32_QA1[ k ], QA + 1 - 12 ); /* QA+1 -> Q12 */
}
} else {
break;
}
}
}
/* Compute reflection coefficients from input signal */
void silk_burg_modified_sse4_1(
opus_int32 *res_nrg, /* O Residual energy */
opus_int *res_nrg_Q, /* O Residual energy Q value */
opus_int32 A_Q16[], /* O Prediction coefficients (length order) */
const opus_int16 x[], /* I Input signal, length: nb_subfr * (D + subfr_length) */
const opus_int32 minInvGain_Q30, /* I Inverse of max prediction gain */
const opus_int subfr_length, /* I Input signal subframe length (incl. D preceding samples) */
const opus_int nb_subfr, /* I Number of subframes stacked in x */
const opus_int D, /* I Order */
int arch /* I Run-time architecture */
)
{
opus_int k, n, s, lz, rshifts, rshifts_extra, reached_max_gain;
opus_int32 C0, num, nrg, rc_Q31, invGain_Q30, Atmp_QA, Atmp1, tmp1, tmp2, x1, x2;
const opus_int16 *x_ptr;
opus_int32 C_first_row[ SILK_MAX_ORDER_LPC ];
opus_int32 C_last_row[ SILK_MAX_ORDER_LPC ];
opus_int32 Af_QA[ SILK_MAX_ORDER_LPC ];
opus_int32 CAf[ SILK_MAX_ORDER_LPC + 1 ];
opus_int32 CAb[ SILK_MAX_ORDER_LPC + 1 ];
opus_int32 xcorr[ SILK_MAX_ORDER_LPC ];
__m128i FIRST_3210, LAST_3210, ATMP_3210, TMP1_3210, TMP2_3210, T1_3210, T2_3210, PTR_3210, SUBFR_3210, X1_3210, X2_3210;
__m128i CONST1 = _mm_set1_epi32(1);
silk_assert(subfr_length * nb_subfr <= MAX_FRAME_SIZE);
/* Compute autocorrelations, added over subframes */
silk_sum_sqr_shift(&C0, &rshifts, x, nb_subfr * subfr_length);
if(rshifts > MAX_RSHIFTS) {
C0 = silk_LSHIFT32(C0, rshifts - MAX_RSHIFTS);
silk_assert(C0 > 0);
rshifts = MAX_RSHIFTS;
} else {
lz = silk_CLZ32(C0) - 1;
rshifts_extra = N_BITS_HEAD_ROOM - lz;
if(rshifts_extra > 0) {
rshifts_extra = silk_min(rshifts_extra, MAX_RSHIFTS - rshifts);
C0 = silk_RSHIFT32(C0, rshifts_extra);
} else {
rshifts_extra = silk_max(rshifts_extra, MIN_RSHIFTS - rshifts);
C0 = silk_LSHIFT32(C0, -rshifts_extra);
}
rshifts += rshifts_extra;
}
CAb[ 0 ] = CAf[ 0 ] = C0 + silk_SMMUL(SILK_FIX_CONST(FIND_LPC_COND_FAC, 32), C0) + 1; /* Q(-rshifts) */
silk_memset(C_first_row, 0, SILK_MAX_ORDER_LPC * sizeof(opus_int32));
if(rshifts > 0) {
for(s = 0; s < nb_subfr; s++) {
x_ptr = x + s * subfr_length;
for(n = 1; n < D + 1; n++) {
C_first_row[ n - 1 ] += (opus_int32)silk_RSHIFT64(
silk_inner_prod16_aligned_64(x_ptr, x_ptr + n, subfr_length - n, arch), rshifts);
}
}
} else {
for(s = 0; s < nb_subfr; s++) {
int i;
opus_int32 d;
x_ptr = x + s * subfr_length;
celt_pitch_xcorr(x_ptr, x_ptr + 1, xcorr, subfr_length - D, D, arch);
for(n = 1; n < D + 1; n++) {
for (i = n + subfr_length - D, d = 0; i < subfr_length; i++)
d = MAC16_16(d, x_ptr[ i ], x_ptr[ i - n ]);
xcorr[ n - 1 ] += d;
}
for(n = 1; n < D + 1; n++) {
C_first_row[ n - 1 ] += silk_LSHIFT32(xcorr[ n - 1 ], -rshifts);
}
}
}
silk_memcpy(C_last_row, C_first_row, SILK_MAX_ORDER_LPC * sizeof(opus_int32));
/* Initialize */
CAb[ 0 ] = CAf[ 0 ] = C0 + silk_SMMUL(SILK_FIX_CONST(FIND_LPC_COND_FAC, 32), C0) + 1; /* Q(-rshifts) */
invGain_Q30 = (opus_int32)1 << 30;
reached_max_gain = 0;
for(n = 0; n < D; n++) {
/* Update first row of correlation matrix (without first element) */
/* Update last row of correlation matrix (without last element, stored in reversed order) */
/* Update C * Af */
/* Update C * flipud(Af) (stored in reversed order) */
if(rshifts > -2) {
for(s = 0; s < nb_subfr; s++) {
x_ptr = x + s * subfr_length;
x1 = -silk_LSHIFT32((opus_int32)x_ptr[ n ], 16 - rshifts); /* Q(16-rshifts) */
x2 = -silk_LSHIFT32((opus_int32)x_ptr[ subfr_length - n - 1 ], 16 - rshifts); /* Q(16-rshifts) */
tmp1 = silk_LSHIFT32((opus_int32)x_ptr[ n ], QA - 16); /* Q(QA-16) */
tmp2 = silk_LSHIFT32((opus_int32)x_ptr[ subfr_length - n - 1 ], QA - 16); /* Q(QA-16) */
for(k = 0; k < n; k++) {
C_first_row[ k ] = silk_SMLAWB(C_first_row[ k ], x1, x_ptr[ n - k - 1 ] ); /* Q(-rshifts) */
C_last_row[ k ] = silk_SMLAWB(C_last_row[ k ], x2, x_ptr[ subfr_length - n + k ]); /* Q(-rshifts) */
Atmp_QA = Af_QA[ k ];
tmp1 = silk_SMLAWB(tmp1, Atmp_QA, x_ptr[ n - k - 1 ] ); /* Q(QA-16) */
tmp2 = silk_SMLAWB(tmp2, Atmp_QA, x_ptr[ subfr_length - n + k ]); /* Q(QA-16) */
}
tmp1 = silk_LSHIFT32(-tmp1, 32 - QA - rshifts); /* Q(16-rshifts) */
tmp2 = silk_LSHIFT32(-tmp2, 32 - QA - rshifts); /* Q(16-rshifts) */
for(k = 0; k <= n; k++) {
CAf[ k ] = silk_SMLAWB(CAf[ k ], tmp1, x_ptr[ n - k ] ); /* Q(-rshift) */
CAb[ k ] = silk_SMLAWB(CAb[ k ], tmp2, x_ptr[ subfr_length - n + k - 1 ]); /* Q(-rshift) */
}
}
} else {
for(s = 0; s < nb_subfr; s++) {
x_ptr = x + s * subfr_length;
x1 = -silk_LSHIFT32((opus_int32)x_ptr[ n ], -rshifts); /* Q(-rshifts) */
x2 = -silk_LSHIFT32((opus_int32)x_ptr[ subfr_length - n - 1 ], -rshifts); /* Q(-rshifts) */
tmp1 = silk_LSHIFT32((opus_int32)x_ptr[ n ], 17); /* Q17 */
tmp2 = silk_LSHIFT32((opus_int32)x_ptr[ subfr_length - n - 1 ], 17); /* Q17 */
X1_3210 = _mm_set1_epi32(x1);
X2_3210 = _mm_set1_epi32(x2);
TMP1_3210 = _mm_setzero_si128();
TMP2_3210 = _mm_setzero_si128();
for(k = 0; k < n - 3; k += 4) {
PTR_3210 = OP_CVTEPI16_EPI32_M64(&x_ptr[ n - k - 1 - 3 ]);
SUBFR_3210 = OP_CVTEPI16_EPI32_M64(&x_ptr[ subfr_length - n + k ]);
FIRST_3210 = _mm_loadu_si128((__m128i *)&C_first_row[ k ]);
PTR_3210 = _mm_shuffle_epi32(PTR_3210, _MM_SHUFFLE(0, 1, 2, 3));
LAST_3210 = _mm_loadu_si128((__m128i *)&C_last_row[ k ]);
ATMP_3210 = _mm_loadu_si128((__m128i *)&Af_QA[ k ]);
T1_3210 = _mm_mullo_epi32(PTR_3210, X1_3210);
T2_3210 = _mm_mullo_epi32(SUBFR_3210, X2_3210);
ATMP_3210 = _mm_srai_epi32(ATMP_3210, 7);
ATMP_3210 = _mm_add_epi32(ATMP_3210, CONST1);
ATMP_3210 = _mm_srai_epi32(ATMP_3210, 1);
FIRST_3210 = _mm_add_epi32(FIRST_3210, T1_3210);
LAST_3210 = _mm_add_epi32(LAST_3210, T2_3210);
PTR_3210 = _mm_mullo_epi32(ATMP_3210, PTR_3210);
SUBFR_3210 = _mm_mullo_epi32(ATMP_3210, SUBFR_3210);
_mm_storeu_si128((__m128i *)&C_first_row[ k ], FIRST_3210);
_mm_storeu_si128((__m128i *)&C_last_row[ k ], LAST_3210);
TMP1_3210 = _mm_add_epi32(TMP1_3210, PTR_3210);
TMP2_3210 = _mm_add_epi32(TMP2_3210, SUBFR_3210);
}
TMP1_3210 = _mm_add_epi32(TMP1_3210, _mm_unpackhi_epi64(TMP1_3210, TMP1_3210));
TMP2_3210 = _mm_add_epi32(TMP2_3210, _mm_unpackhi_epi64(TMP2_3210, TMP2_3210));
TMP1_3210 = _mm_add_epi32(TMP1_3210, _mm_shufflelo_epi16(TMP1_3210, 0x0E));
TMP2_3210 = _mm_add_epi32(TMP2_3210, _mm_shufflelo_epi16(TMP2_3210, 0x0E));
tmp1 += _mm_cvtsi128_si32(TMP1_3210);
tmp2 += _mm_cvtsi128_si32(TMP2_3210);
for(; k < n; k++) {
C_first_row[ k ] = silk_MLA(C_first_row[ k ], x1, x_ptr[ n - k - 1 ] ); /* Q(-rshifts) */
C_last_row[ k ] = silk_MLA(C_last_row[ k ], x2, x_ptr[ subfr_length - n + k ]); /* Q(-rshifts) */
Atmp1 = silk_RSHIFT_ROUND(Af_QA[ k ], QA - 17); /* Q17 */
tmp1 = silk_MLA(tmp1, x_ptr[ n - k - 1 ], Atmp1); /* Q17 */
tmp2 = silk_MLA(tmp2, x_ptr[ subfr_length - n + k ], Atmp1); /* Q17 */
}
tmp1 = -tmp1; /* Q17 */
tmp2 = -tmp2; /* Q17 */
{
__m128i xmm_tmp1, xmm_tmp2;
__m128i xmm_x_ptr_n_k_x2x0, xmm_x_ptr_n_k_x3x1;
__m128i xmm_x_ptr_sub_x2x0, xmm_x_ptr_sub_x3x1;
xmm_tmp1 = _mm_set1_epi32(tmp1);
xmm_tmp2 = _mm_set1_epi32(tmp2);
for(k = 0; k <= n - 3; k += 4) {
xmm_x_ptr_n_k_x2x0 = OP_CVTEPI16_EPI32_M64(&x_ptr[ n - k - 3 ]);
xmm_x_ptr_sub_x2x0 = OP_CVTEPI16_EPI32_M64(&x_ptr[ subfr_length - n + k - 1 ]);
xmm_x_ptr_n_k_x2x0 = _mm_shuffle_epi32(xmm_x_ptr_n_k_x2x0, _MM_SHUFFLE(0, 1, 2, 3));
xmm_x_ptr_n_k_x2x0 = _mm_slli_epi32(xmm_x_ptr_n_k_x2x0, -rshifts - 1);
xmm_x_ptr_sub_x2x0 = _mm_slli_epi32(xmm_x_ptr_sub_x2x0, -rshifts - 1);
/* equal shift right 4 bytes, xmm_x_ptr_n_k_x3x1 = _mm_srli_si128(xmm_x_ptr_n_k_x2x0, 4)*/
xmm_x_ptr_n_k_x3x1 = _mm_shuffle_epi32(xmm_x_ptr_n_k_x2x0, _MM_SHUFFLE(0, 3, 2, 1));
xmm_x_ptr_sub_x3x1 = _mm_shuffle_epi32(xmm_x_ptr_sub_x2x0, _MM_SHUFFLE(0, 3, 2, 1));
xmm_x_ptr_n_k_x2x0 = _mm_mul_epi32(xmm_x_ptr_n_k_x2x0, xmm_tmp1);
xmm_x_ptr_n_k_x3x1 = _mm_mul_epi32(xmm_x_ptr_n_k_x3x1, xmm_tmp1);
xmm_x_ptr_sub_x2x0 = _mm_mul_epi32(xmm_x_ptr_sub_x2x0, xmm_tmp2);
xmm_x_ptr_sub_x3x1 = _mm_mul_epi32(xmm_x_ptr_sub_x3x1, xmm_tmp2);
xmm_x_ptr_n_k_x2x0 = _mm_srli_epi64(xmm_x_ptr_n_k_x2x0, 16);
xmm_x_ptr_n_k_x3x1 = _mm_slli_epi64(xmm_x_ptr_n_k_x3x1, 16);
xmm_x_ptr_sub_x2x0 = _mm_srli_epi64(xmm_x_ptr_sub_x2x0, 16);
xmm_x_ptr_sub_x3x1 = _mm_slli_epi64(xmm_x_ptr_sub_x3x1, 16);
xmm_x_ptr_n_k_x2x0 = _mm_blend_epi16(xmm_x_ptr_n_k_x2x0, xmm_x_ptr_n_k_x3x1, 0xCC);
xmm_x_ptr_sub_x2x0 = _mm_blend_epi16(xmm_x_ptr_sub_x2x0, xmm_x_ptr_sub_x3x1, 0xCC);
X1_3210 = _mm_loadu_si128((__m128i *)&CAf[ k ]);
PTR_3210 = _mm_loadu_si128((__m128i *)&CAb[ k ]);
X1_3210 = _mm_add_epi32(X1_3210, xmm_x_ptr_n_k_x2x0);
PTR_3210 = _mm_add_epi32(PTR_3210, xmm_x_ptr_sub_x2x0);
_mm_storeu_si128((__m128i *)&CAf[ k ], X1_3210);
_mm_storeu_si128((__m128i *)&CAb[ k ], PTR_3210);
}
for(; k <= n; k++) {
CAf[ k ] = silk_SMLAWW(CAf[ k ], tmp1,
silk_LSHIFT32((opus_int32)x_ptr[ n - k ], -rshifts - 1)); /* Q(-rshift) */
CAb[ k ] = silk_SMLAWW(CAb[ k ], tmp2,
silk_LSHIFT32((opus_int32)x_ptr[ subfr_length - n + k - 1 ], -rshifts - 1)); /* Q(-rshift) */
}
}
}
}
/* Calculate nominator and denominator for the next order reflection (parcor) coefficient */
tmp1 = C_first_row[ n ]; /* Q(-rshifts) */
tmp2 = C_last_row[ n ]; /* Q(-rshifts) */
num = 0; /* Q(-rshifts) */
nrg = silk_ADD32(CAb[ 0 ], CAf[ 0 ]); /* Q(1-rshifts) */
for(k = 0; k < n; k++) {
Atmp_QA = Af_QA[ k ];
lz = silk_CLZ32(silk_abs(Atmp_QA)) - 1;
lz = silk_min(32 - QA, lz);
Atmp1 = silk_LSHIFT32(Atmp_QA, lz); /* Q(QA + lz) */
tmp1 = silk_ADD_LSHIFT32(tmp1, silk_SMMUL(C_last_row[ n - k - 1 ], Atmp1), 32 - QA - lz); /* Q(-rshifts) */
tmp2 = silk_ADD_LSHIFT32(tmp2, silk_SMMUL(C_first_row[ n - k - 1 ], Atmp1), 32 - QA - lz); /* Q(-rshifts) */
num = silk_ADD_LSHIFT32(num, silk_SMMUL(CAb[ n - k ], Atmp1), 32 - QA - lz); /* Q(-rshifts) */
nrg = silk_ADD_LSHIFT32(nrg, silk_SMMUL(silk_ADD32(CAb[ k + 1 ], CAf[ k + 1 ]),
Atmp1), 32 - QA - lz); /* Q(1-rshifts) */
}
CAf[ n + 1 ] = tmp1; /* Q(-rshifts) */
CAb[ n + 1 ] = tmp2; /* Q(-rshifts) */
num = silk_ADD32(num, tmp2); /* Q(-rshifts) */
num = silk_LSHIFT32(-num, 1); /* Q(1-rshifts) */
/* Calculate the next order reflection (parcor) coefficient */
if(silk_abs(num) < nrg) {
rc_Q31 = silk_DIV32_varQ(num, nrg, 31);
} else {
rc_Q31 = (num > 0) ? silk_int32_MAX : silk_int32_MIN;
}
/* Update inverse prediction gain */
tmp1 = ((opus_int32)1 << 30) - silk_SMMUL(rc_Q31, rc_Q31);
tmp1 = silk_LSHIFT(silk_SMMUL(invGain_Q30, tmp1), 2);
if(tmp1 <= minInvGain_Q30) {
/* Max prediction gain exceeded; set reflection coefficient such that max prediction gain is exactly hit */
tmp2 = ((opus_int32)1 << 30) - silk_DIV32_varQ(minInvGain_Q30, invGain_Q30, 30); /* Q30 */
rc_Q31 = silk_SQRT_APPROX(tmp2); /* Q15 */
/* Newton-Raphson iteration */
rc_Q31 = silk_RSHIFT32(rc_Q31 + silk_DIV32(tmp2, rc_Q31), 1); /* Q15 */
rc_Q31 = silk_LSHIFT32(rc_Q31, 16); /* Q31 */
if(num < 0) {
/* Ensure adjusted reflection coefficients has the original sign */
rc_Q31 = -rc_Q31;
}
invGain_Q30 = minInvGain_Q30;
reached_max_gain = 1;
} else {
invGain_Q30 = tmp1;
}
/* Update the AR coefficients */
for(k = 0; k < (n + 1) >> 1; k++) {
tmp1 = Af_QA[ k ]; /* QA */
tmp2 = Af_QA[ n - k - 1 ]; /* QA */
Af_QA[ k ] = silk_ADD_LSHIFT32(tmp1, silk_SMMUL(tmp2, rc_Q31), 1); /* QA */
Af_QA[ n - k - 1 ] = silk_ADD_LSHIFT32(tmp2, silk_SMMUL(tmp1, rc_Q31), 1); /* QA */
}
Af_QA[ n ] = silk_RSHIFT32(rc_Q31, 31 - QA); /* QA */
if(reached_max_gain) {
/* Reached max prediction gain; set remaining coefficients to zero and exit loop */
for(k = n + 1; k < D; k++) {
Af_QA[ k ] = 0;
}
break;
}
/* Update C * Af and C * Ab */
for(k = 0; k <= n + 1; k++) {
tmp1 = CAf[ k ]; /* Q(-rshifts) */
tmp2 = CAb[ n - k + 1 ]; /* Q(-rshifts) */
CAf[ k ] = silk_ADD_LSHIFT32(tmp1, silk_SMMUL(tmp2, rc_Q31), 1); /* Q(-rshifts) */
CAb[ n - k + 1 ] = silk_ADD_LSHIFT32(tmp2, silk_SMMUL(tmp1, rc_Q31), 1); /* Q(-rshifts) */
}
}
if(reached_max_gain) {
for(k = 0; k < D; k++) {
/* Scale coefficients */
A_Q16[ k ] = -silk_RSHIFT_ROUND(Af_QA[ k ], QA - 16);
}
/* Subtract energy of preceding samples from C0 */
if(rshifts > 0) {
for(s = 0; s < nb_subfr; s++) {
x_ptr = x + s * subfr_length;
C0 -= (opus_int32)silk_RSHIFT64(silk_inner_prod16_aligned_64(x_ptr, x_ptr, D, arch), rshifts);
}
} else {
for(s = 0; s < nb_subfr; s++) {
x_ptr = x + s * subfr_length;
C0 -= silk_LSHIFT32(silk_inner_prod_aligned(x_ptr, x_ptr, D, arch), -rshifts);
}
}
/* Approximate residual energy */
*res_nrg = silk_LSHIFT(silk_SMMUL(invGain_Q30, C0), 2);
*res_nrg_Q = -rshifts;
} else {
/* Return residual energy */
nrg = CAf[ 0 ]; /* Q(-rshifts) */
tmp1 = (opus_int32)1 << 16; /* Q16 */
for(k = 0; k < D; k++) {
Atmp1 = silk_RSHIFT_ROUND(Af_QA[ k ], QA - 16); /* Q16 */
nrg = silk_SMLAWW(nrg, CAf[ k + 1 ], Atmp1); /* Q(-rshifts) */
tmp1 = silk_SMLAWW(tmp1, Atmp1, Atmp1); /* Q16 */
A_Q16[ k ] = -Atmp1;
}
*res_nrg = silk_SMLAWW(nrg, silk_SMMUL(SILK_FIX_CONST(FIND_LPC_COND_FAC, 32), C0), -tmp1);/* Q(-rshifts) */
*res_nrg_Q = -rshifts;
}
}
/* Compute reflection coefficients from input signal */
void silk_burg_modified(
opus_int32 *res_nrg, /* O Residual energy */
opus_int *res_nrg_Q, /* O Residual energy Q value */
opus_int32 A_Q16[], /* O Prediction coefficients (length order) */
const opus_int16 x[], /* I Input signal, length: nb_subfr * ( D + subfr_length ) */
const opus_int32 minInvGain_Q30, /* I Inverse of max prediction gain */
const opus_int subfr_length, /* I Input signal subframe length (incl. D preceding samples) */
const opus_int nb_subfr, /* I Number of subframes stacked in x */
const opus_int D /* I Order */
)
{
opus_int k, n, s, lz, rshifts, rshifts_extra, reached_max_gain;
opus_int32 C0, num, nrg, rc_Q31, invGain_Q30, Atmp_QA, Atmp1, tmp1, tmp2, x1, x2;
const opus_int16 *x_ptr;
opus_int32 C_first_row[ SILK_MAX_ORDER_LPC ];
opus_int32 C_last_row[ SILK_MAX_ORDER_LPC ];
opus_int32 Af_QA[ SILK_MAX_ORDER_LPC ];
opus_int32 CAf[ SILK_MAX_ORDER_LPC + 1 ];
opus_int32 CAb[ SILK_MAX_ORDER_LPC + 1 ];
silk_assert( subfr_length * nb_subfr <= MAX_FRAME_SIZE );
/* Compute autocorrelations, added over subframes */
silk_sum_sqr_shift( &C0, &rshifts, x, nb_subfr * subfr_length );
if( rshifts > MAX_RSHIFTS ) {
C0 = silk_LSHIFT32( C0, rshifts - MAX_RSHIFTS );
silk_assert( C0 > 0 );
rshifts = MAX_RSHIFTS;
} else {
lz = silk_CLZ32( C0 ) - 1;
rshifts_extra = N_BITS_HEAD_ROOM - lz;
if( rshifts_extra > 0 ) {
rshifts_extra = silk_min( rshifts_extra, MAX_RSHIFTS - rshifts );
C0 = silk_RSHIFT32( C0, rshifts_extra );
} else {
rshifts_extra = silk_max( rshifts_extra, MIN_RSHIFTS - rshifts );
C0 = silk_LSHIFT32( C0, -rshifts_extra );
}
rshifts += rshifts_extra;
}
CAb[ 0 ] = CAf[ 0 ] = C0 + silk_SMMUL( SILK_FIX_CONST( FIND_LPC_COND_FAC, 32 ), C0 ) + 1; /* Q(-rshifts) */
silk_memset( C_first_row, 0, SILK_MAX_ORDER_LPC * sizeof( opus_int32 ) );
if( rshifts > 0 ) {
for( s = 0; s < nb_subfr; s++ ) {
x_ptr = x + s * subfr_length;
for( n = 1; n < D + 1; n++ ) {
C_first_row[ n - 1 ] += (opus_int32)silk_RSHIFT64(
silk_inner_prod16_aligned_64( x_ptr, x_ptr + n, subfr_length - n ), rshifts );
}
}
} else {
for( s = 0; s < nb_subfr; s++ ) {
x_ptr = x + s * subfr_length;
for( n = 1; n < D + 1; n++ ) {
C_first_row[ n - 1 ] += silk_LSHIFT32(
silk_inner_prod_aligned( x_ptr, x_ptr + n, subfr_length - n ), -rshifts );
}
}
}
silk_memcpy( C_last_row, C_first_row, SILK_MAX_ORDER_LPC * sizeof( opus_int32 ) );
/* Initialize */
CAb[ 0 ] = CAf[ 0 ] = C0 + silk_SMMUL( SILK_FIX_CONST( FIND_LPC_COND_FAC, 32 ), C0 ) + 1; /* Q(-rshifts) */
invGain_Q30 = (opus_int32)1 << 30;
reached_max_gain = 0;
for( n = 0; n < D; n++ ) {
/* Update first row of correlation matrix (without first element) */
/* Update last row of correlation matrix (without last element, stored in reversed order) */
/* Update C * Af */
/* Update C * flipud(Af) (stored in reversed order) */
if( rshifts > -2 ) {
for( s = 0; s < nb_subfr; s++ ) {
x_ptr = x + s * subfr_length;
x1 = -silk_LSHIFT32( (opus_int32)x_ptr[ n ], 16 - rshifts ); /* Q(16-rshifts) */
x2 = -silk_LSHIFT32( (opus_int32)x_ptr[ subfr_length - n - 1 ], 16 - rshifts ); /* Q(16-rshifts) */
tmp1 = silk_LSHIFT32( (opus_int32)x_ptr[ n ], QA - 16 ); /* Q(QA-16) */
tmp2 = silk_LSHIFT32( (opus_int32)x_ptr[ subfr_length - n - 1 ], QA - 16 ); /* Q(QA-16) */
for( k = 0; k < n; k++ ) {
C_first_row[ k ] = silk_SMLAWB( C_first_row[ k ], x1, x_ptr[ n - k - 1 ] ); /* Q( -rshifts ) */
C_last_row[ k ] = silk_SMLAWB( C_last_row[ k ], x2, x_ptr[ subfr_length - n + k ] ); /* Q( -rshifts ) */
Atmp_QA = Af_QA[ k ];
tmp1 = silk_SMLAWB( tmp1, Atmp_QA, x_ptr[ n - k - 1 ] ); /* Q(QA-16) */
tmp2 = silk_SMLAWB( tmp2, Atmp_QA, x_ptr[ subfr_length - n + k ] ); /* Q(QA-16) */
}
tmp1 = silk_LSHIFT32( -tmp1, 32 - QA - rshifts ); /* Q(16-rshifts) */
tmp2 = silk_LSHIFT32( -tmp2, 32 - QA - rshifts ); /* Q(16-rshifts) */
for( k = 0; k <= n; k++ ) {
CAf[ k ] = silk_SMLAWB( CAf[ k ], tmp1, x_ptr[ n - k ] ); /* Q( -rshift ) */
CAb[ k ] = silk_SMLAWB( CAb[ k ], tmp2, x_ptr[ subfr_length - n + k - 1 ] ); /* Q( -rshift ) */
}
}
} else {
for( s = 0; s < nb_subfr; s++ ) {
x_ptr = x + s * subfr_length;
x1 = -silk_LSHIFT32( (opus_int32)x_ptr[ n ], -rshifts ); /* Q( -rshifts ) */
x2 = -silk_LSHIFT32( (opus_int32)x_ptr[ subfr_length - n - 1 ], -rshifts ); /* Q( -rshifts ) */
tmp1 = silk_LSHIFT32( (opus_int32)x_ptr[ n ], 17 ); /* Q17 */
tmp2 = silk_LSHIFT32( (opus_int32)x_ptr[ subfr_length - n - 1 ], 17 ); /* Q17 */
for( k = 0; k < n; k++ ) {
C_first_row[ k ] = silk_MLA( C_first_row[ k ], x1, x_ptr[ n - k - 1 ] ); /* Q( -rshifts ) */
C_last_row[ k ] = silk_MLA( C_last_row[ k ], x2, x_ptr[ subfr_length - n + k ] ); /* Q( -rshifts ) */
Atmp1 = silk_RSHIFT_ROUND( Af_QA[ k ], QA - 17 ); /* Q17 */
tmp1 = silk_MLA( tmp1, x_ptr[ n - k - 1 ], Atmp1 ); /* Q17 */
tmp2 = silk_MLA( tmp2, x_ptr[ subfr_length - n + k ], Atmp1 ); /* Q17 */
}
tmp1 = -tmp1; /* Q17 */
tmp2 = -tmp2; /* Q17 */
for( k = 0; k <= n; k++ ) {
CAf[ k ] = silk_SMLAWW( CAf[ k ], tmp1,
silk_LSHIFT32( (opus_int32)x_ptr[ n - k ], -rshifts - 1 ) ); /* Q( -rshift ) */
CAb[ k ] = silk_SMLAWW( CAb[ k ], tmp2,
silk_LSHIFT32( (opus_int32)x_ptr[ subfr_length - n + k - 1 ], -rshifts - 1 ) ); /* Q( -rshift ) */
}
}
}
/* Calculate nominator and denominator for the next order reflection (parcor) coefficient */
tmp1 = C_first_row[ n ]; /* Q( -rshifts ) */
tmp2 = C_last_row[ n ]; /* Q( -rshifts ) */
num = 0; /* Q( -rshifts ) */
nrg = silk_ADD32( CAb[ 0 ], CAf[ 0 ] ); /* Q( 1-rshifts ) */
for( k = 0; k < n; k++ ) {
Atmp_QA = Af_QA[ k ];
lz = silk_CLZ32( silk_abs( Atmp_QA ) ) - 1;
lz = silk_min( 32 - QA, lz );
Atmp1 = silk_LSHIFT32( Atmp_QA, lz ); /* Q( QA + lz ) */
tmp1 = silk_ADD_LSHIFT32( tmp1, silk_SMMUL( C_last_row[ n - k - 1 ], Atmp1 ), 32 - QA - lz ); /* Q( -rshifts ) */
tmp2 = silk_ADD_LSHIFT32( tmp2, silk_SMMUL( C_first_row[ n - k - 1 ], Atmp1 ), 32 - QA - lz ); /* Q( -rshifts ) */
num = silk_ADD_LSHIFT32( num, silk_SMMUL( CAb[ n - k ], Atmp1 ), 32 - QA - lz ); /* Q( -rshifts ) */
nrg = silk_ADD_LSHIFT32( nrg, silk_SMMUL( silk_ADD32( CAb[ k + 1 ], CAf[ k + 1 ] ),
Atmp1 ), 32 - QA - lz ); /* Q( 1-rshifts ) */
}
CAf[ n + 1 ] = tmp1; /* Q( -rshifts ) */
CAb[ n + 1 ] = tmp2; /* Q( -rshifts ) */
num = silk_ADD32( num, tmp2 ); /* Q( -rshifts ) */
num = silk_LSHIFT32( -num, 1 ); /* Q( 1-rshifts ) */
/* Calculate the next order reflection (parcor) coefficient */
if( silk_abs( num ) < nrg ) {
rc_Q31 = silk_DIV32_varQ( num, nrg, 31 );
} else {
rc_Q31 = ( num > 0 ) ? silk_int32_MAX : silk_int32_MIN;
}
/* Update inverse prediction gain */
tmp1 = ( (opus_int32)1 << 30 ) - silk_SMMUL( rc_Q31, rc_Q31 );
tmp1 = silk_LSHIFT( silk_SMMUL( invGain_Q30, tmp1 ), 2 );
if( tmp1 <= minInvGain_Q30 ) {
/* Max prediction gain exceeded; set reflection coefficient such that max prediction gain is exactly hit */
tmp2 = ( (opus_int32)1 << 30 ) - silk_DIV32_varQ( minInvGain_Q30, invGain_Q30, 30 ); /* Q30 */
rc_Q31 = silk_SQRT_APPROX( tmp2 ); /* Q15 */
/* Newton-Raphson iteration */
rc_Q31 = silk_RSHIFT32( rc_Q31 + silk_DIV32( tmp2, rc_Q31 ), 1 ); /* Q15 */
rc_Q31 = silk_LSHIFT32( rc_Q31, 16 ); /* Q31 */
if( num < 0 ) {
/* Ensure adjusted reflection coefficients has the original sign */
rc_Q31 = -rc_Q31;
}
invGain_Q30 = minInvGain_Q30;
reached_max_gain = 1;
} else {
invGain_Q30 = tmp1;
}
/* Update the AR coefficients */
for( k = 0; k < (n + 1) >> 1; k++ ) {
tmp1 = Af_QA[ k ]; /* QA */
tmp2 = Af_QA[ n - k - 1 ]; /* QA */
Af_QA[ k ] = silk_ADD_LSHIFT32( tmp1, silk_SMMUL( tmp2, rc_Q31 ), 1 ); /* QA */
Af_QA[ n - k - 1 ] = silk_ADD_LSHIFT32( tmp2, silk_SMMUL( tmp1, rc_Q31 ), 1 ); /* QA */
}
Af_QA[ n ] = silk_RSHIFT32( rc_Q31, 31 - QA ); /* QA */
if( reached_max_gain ) {
/* Reached max prediction gain; set remaining coefficients to zero and exit loop */
for( k = n + 1; k < D; k++ ) {
Af_QA[ k ] = 0;
}
break;
}
/* Update C * Af and C * Ab */
for( k = 0; k <= n + 1; k++ ) {
tmp1 = CAf[ k ]; /* Q( -rshifts ) */
tmp2 = CAb[ n - k + 1 ]; /* Q( -rshifts ) */
CAf[ k ] = silk_ADD_LSHIFT32( tmp1, silk_SMMUL( tmp2, rc_Q31 ), 1 ); /* Q( -rshifts ) */
CAb[ n - k + 1 ] = silk_ADD_LSHIFT32( tmp2, silk_SMMUL( tmp1, rc_Q31 ), 1 ); /* Q( -rshifts ) */
}
}
if( reached_max_gain ) {
for( k = 0; k < D; k++ ) {
/* Scale coefficients */
A_Q16[ k ] = -silk_RSHIFT_ROUND( Af_QA[ k ], QA - 16 );
}
/* Subtract energy of preceding samples from C0 */
if( rshifts > 0 ) {
for( s = 0; s < nb_subfr; s++ ) {
x_ptr = x + s * subfr_length;
C0 -= (opus_int32)silk_RSHIFT64( silk_inner_prod16_aligned_64( x_ptr, x_ptr, D ), rshifts );
}
} else {
for( s = 0; s < nb_subfr; s++ ) {
x_ptr = x + s * subfr_length;
C0 -= silk_LSHIFT32( silk_inner_prod_aligned( x_ptr, x_ptr, D ), -rshifts );
}
}
/* Approximate residual energy */
*res_nrg = silk_LSHIFT( silk_SMMUL( invGain_Q30, C0 ), 2 );
*res_nrg_Q = -rshifts;
} else {
/* Return residual energy */
nrg = CAf[ 0 ]; /* Q( -rshifts ) */
tmp1 = (opus_int32)1 << 16; /* Q16 */
for( k = 0; k < D; k++ ) {
Atmp1 = silk_RSHIFT_ROUND( Af_QA[ k ], QA - 16 ); /* Q16 */
nrg = silk_SMLAWW( nrg, CAf[ k + 1 ], Atmp1 ); /* Q( -rshifts ) */
tmp1 = silk_SMLAWW( tmp1, Atmp1, Atmp1 ); /* Q16 */
A_Q16[ k ] = -Atmp1;
}
*res_nrg = silk_SMLAWW( nrg, silk_SMMUL( FIND_LPC_COND_FAC, C0 ), -tmp1 ); /* Q( -rshifts ) */
*res_nrg_Q = -rshifts;
}
}
/* encControl->payloadSize_ms is set to */
opus_int silk_Encode( /* O Returns error code */
void *encState, /* I/O State */
silk_EncControlStruct *encControl, /* I Control status */
const opus_int16 *samplesIn, /* I Speech sample input vector */
opus_int nSamplesIn, /* I Number of samples in input vector */
ec_enc *psRangeEnc, /* I/O Compressor data structure */
opus_int32 *nBytesOut, /* I/O Number of bytes in payload (input: Max bytes) */
const opus_int prefillFlag /* I Flag to indicate prefilling buffers no coding */
)
{
opus_int n, i, nBits, flags, tmp_payloadSize_ms = 0, tmp_complexity = 0, ret = 0;
opus_int nSamplesToBuffer, nSamplesToBufferMax, nBlocksOf10ms;
opus_int nSamplesFromInput = 0, nSamplesFromInputMax;
opus_int speech_act_thr_for_switch_Q8;
opus_int32 TargetRate_bps, MStargetRates_bps[ 2 ], channelRate_bps, LBRR_symbol, sum;
silk_encoder *psEnc = ( silk_encoder * )encState;
VARDECL( opus_int16, buf );
opus_int transition, curr_block, tot_blocks;
SAVE_STACK;
if (encControl->reducedDependency)
{
psEnc->state_Fxx[0].sCmn.first_frame_after_reset = 1;
psEnc->state_Fxx[1].sCmn.first_frame_after_reset = 1;
}
psEnc->state_Fxx[ 0 ].sCmn.nFramesEncoded = psEnc->state_Fxx[ 1 ].sCmn.nFramesEncoded = 0;
/* Check values in encoder control structure */
if( ( ret = check_control_input( encControl ) ) != 0 ) {
silk_assert( 0 );
RESTORE_STACK;
return ret;
}
encControl->switchReady = 0;
if( encControl->nChannelsInternal > psEnc->nChannelsInternal ) {
/* Mono -> Stereo transition: init state of second channel and stereo state */
ret += silk_init_encoder( &psEnc->state_Fxx[ 1 ], psEnc->state_Fxx[ 0 ].sCmn.arch );
silk_memset( psEnc->sStereo.pred_prev_Q13, 0, sizeof( psEnc->sStereo.pred_prev_Q13 ) );
silk_memset( psEnc->sStereo.sSide, 0, sizeof( psEnc->sStereo.sSide ) );
psEnc->sStereo.mid_side_amp_Q0[ 0 ] = 0;
psEnc->sStereo.mid_side_amp_Q0[ 1 ] = 1;
psEnc->sStereo.mid_side_amp_Q0[ 2 ] = 0;
psEnc->sStereo.mid_side_amp_Q0[ 3 ] = 1;
psEnc->sStereo.width_prev_Q14 = 0;
psEnc->sStereo.smth_width_Q14 = SILK_FIX_CONST( 1, 14 );
if( psEnc->nChannelsAPI == 2 ) {
silk_memcpy( &psEnc->state_Fxx[ 1 ].sCmn.resampler_state, &psEnc->state_Fxx[ 0 ].sCmn.resampler_state, sizeof( silk_resampler_state_struct ) );
silk_memcpy( &psEnc->state_Fxx[ 1 ].sCmn.In_HP_State, &psEnc->state_Fxx[ 0 ].sCmn.In_HP_State, sizeof( psEnc->state_Fxx[ 1 ].sCmn.In_HP_State ) );
}
}
transition = (encControl->payloadSize_ms != psEnc->state_Fxx[ 0 ].sCmn.PacketSize_ms) || (psEnc->nChannelsInternal != encControl->nChannelsInternal);
psEnc->nChannelsAPI = encControl->nChannelsAPI;
psEnc->nChannelsInternal = encControl->nChannelsInternal;
nBlocksOf10ms = silk_DIV32( 100 * nSamplesIn, encControl->API_sampleRate );
tot_blocks = ( nBlocksOf10ms > 1 ) ? nBlocksOf10ms >> 1 : 1;
curr_block = 0;
if( prefillFlag ) {
/* Only accept input length of 10 ms */
if( nBlocksOf10ms != 1 ) {
silk_assert( 0 );
RESTORE_STACK;
return SILK_ENC_INPUT_INVALID_NO_OF_SAMPLES;
}
/* Reset Encoder */
for( n = 0; n < encControl->nChannelsInternal; n++ ) {
ret = silk_init_encoder( &psEnc->state_Fxx[ n ], psEnc->state_Fxx[ n ].sCmn.arch );
silk_assert( !ret );
}
tmp_payloadSize_ms = encControl->payloadSize_ms;
encControl->payloadSize_ms = 10;
tmp_complexity = encControl->complexity;
encControl->complexity = 0;
for( n = 0; n < encControl->nChannelsInternal; n++ ) {
psEnc->state_Fxx[ n ].sCmn.controlled_since_last_payload = 0;
psEnc->state_Fxx[ n ].sCmn.prefillFlag = 1;
}
} else {
/* Only accept input lengths that are a multiple of 10 ms */
if( nBlocksOf10ms * encControl->API_sampleRate != 100 * nSamplesIn || nSamplesIn < 0 ) {
silk_assert( 0 );
RESTORE_STACK;
return SILK_ENC_INPUT_INVALID_NO_OF_SAMPLES;
}
/* Make sure no more than one packet can be produced */
if( 1000 * (opus_int32)nSamplesIn > encControl->payloadSize_ms * encControl->API_sampleRate ) {
silk_assert( 0 );
RESTORE_STACK;
return SILK_ENC_INPUT_INVALID_NO_OF_SAMPLES;
}
}
TargetRate_bps = silk_RSHIFT32( encControl->bitRate, encControl->nChannelsInternal - 1 );
for( n = 0; n < encControl->nChannelsInternal; n++ ) {
/* Force the side channel to the same rate as the mid */
opus_int force_fs_kHz = (n==1) ? psEnc->state_Fxx[0].sCmn.fs_kHz : 0;
if( ( ret = silk_control_encoder( &psEnc->state_Fxx[ n ], encControl, TargetRate_bps, psEnc->allowBandwidthSwitch, n, force_fs_kHz ) ) != 0 ) {
silk_assert( 0 );
RESTORE_STACK;
return ret;
}
if( psEnc->state_Fxx[n].sCmn.first_frame_after_reset || transition ) {
for( i = 0; i < psEnc->state_Fxx[ 0 ].sCmn.nFramesPerPacket; i++ ) {
psEnc->state_Fxx[ n ].sCmn.LBRR_flags[ i ] = 0;
}
}
psEnc->state_Fxx[ n ].sCmn.inDTX = psEnc->state_Fxx[ n ].sCmn.useDTX;
}
silk_assert( encControl->nChannelsInternal == 1 || psEnc->state_Fxx[ 0 ].sCmn.fs_kHz == psEnc->state_Fxx[ 1 ].sCmn.fs_kHz );
/* Input buffering/resampling and encoding */
nSamplesToBufferMax =
10 * nBlocksOf10ms * psEnc->state_Fxx[ 0 ].sCmn.fs_kHz;
nSamplesFromInputMax =
silk_DIV32_16( nSamplesToBufferMax *
psEnc->state_Fxx[ 0 ].sCmn.API_fs_Hz,
psEnc->state_Fxx[ 0 ].sCmn.fs_kHz * 1000 );
ALLOC( buf, nSamplesFromInputMax, opus_int16 );
while( 1 ) {
nSamplesToBuffer = psEnc->state_Fxx[ 0 ].sCmn.frame_length - psEnc->state_Fxx[ 0 ].sCmn.inputBufIx;
nSamplesToBuffer = silk_min( nSamplesToBuffer, nSamplesToBufferMax );
nSamplesFromInput = silk_DIV32_16( nSamplesToBuffer * psEnc->state_Fxx[ 0 ].sCmn.API_fs_Hz, psEnc->state_Fxx[ 0 ].sCmn.fs_kHz * 1000 );
/* Resample and write to buffer */
if( encControl->nChannelsAPI == 2 && encControl->nChannelsInternal == 2 ) {
opus_int id = psEnc->state_Fxx[ 0 ].sCmn.nFramesEncoded;
for( n = 0; n < nSamplesFromInput; n++ ) {
buf[ n ] = samplesIn[ 2 * n ];
}
/* Making sure to start both resamplers from the same state when switching from mono to stereo */
if( psEnc->nPrevChannelsInternal == 1 && id==0 ) {
silk_memcpy( &psEnc->state_Fxx[ 1 ].sCmn.resampler_state, &psEnc->state_Fxx[ 0 ].sCmn.resampler_state, sizeof(psEnc->state_Fxx[ 1 ].sCmn.resampler_state));
}
ret += silk_resampler( &psEnc->state_Fxx[ 0 ].sCmn.resampler_state,
&psEnc->state_Fxx[ 0 ].sCmn.inputBuf[ psEnc->state_Fxx[ 0 ].sCmn.inputBufIx + 2 ], buf, nSamplesFromInput );
psEnc->state_Fxx[ 0 ].sCmn.inputBufIx += nSamplesToBuffer;
nSamplesToBuffer = psEnc->state_Fxx[ 1 ].sCmn.frame_length - psEnc->state_Fxx[ 1 ].sCmn.inputBufIx;
nSamplesToBuffer = silk_min( nSamplesToBuffer, 10 * nBlocksOf10ms * psEnc->state_Fxx[ 1 ].sCmn.fs_kHz );
for( n = 0; n < nSamplesFromInput; n++ ) {
buf[ n ] = samplesIn[ 2 * n + 1 ];
}
ret += silk_resampler( &psEnc->state_Fxx[ 1 ].sCmn.resampler_state,
&psEnc->state_Fxx[ 1 ].sCmn.inputBuf[ psEnc->state_Fxx[ 1 ].sCmn.inputBufIx + 2 ], buf, nSamplesFromInput );
psEnc->state_Fxx[ 1 ].sCmn.inputBufIx += nSamplesToBuffer;
} else if( encControl->nChannelsAPI == 2 && encControl->nChannelsInternal == 1 ) {
/* Combine left and right channels before resampling */
for( n = 0; n < nSamplesFromInput; n++ ) {
sum = samplesIn[ 2 * n ] + samplesIn[ 2 * n + 1 ];
buf[ n ] = (opus_int16)silk_RSHIFT_ROUND( sum, 1 );
}
ret += silk_resampler( &psEnc->state_Fxx[ 0 ].sCmn.resampler_state,
&psEnc->state_Fxx[ 0 ].sCmn.inputBuf[ psEnc->state_Fxx[ 0 ].sCmn.inputBufIx + 2 ], buf, nSamplesFromInput );
/* On the first mono frame, average the results for the two resampler states */
if( psEnc->nPrevChannelsInternal == 2 && psEnc->state_Fxx[ 0 ].sCmn.nFramesEncoded == 0 ) {
ret += silk_resampler( &psEnc->state_Fxx[ 1 ].sCmn.resampler_state,
&psEnc->state_Fxx[ 1 ].sCmn.inputBuf[ psEnc->state_Fxx[ 1 ].sCmn.inputBufIx + 2 ], buf, nSamplesFromInput );
for( n = 0; n < psEnc->state_Fxx[ 0 ].sCmn.frame_length; n++ ) {
psEnc->state_Fxx[ 0 ].sCmn.inputBuf[ psEnc->state_Fxx[ 0 ].sCmn.inputBufIx+n+2 ] =
silk_RSHIFT(psEnc->state_Fxx[ 0 ].sCmn.inputBuf[ psEnc->state_Fxx[ 0 ].sCmn.inputBufIx+n+2 ]
+ psEnc->state_Fxx[ 1 ].sCmn.inputBuf[ psEnc->state_Fxx[ 1 ].sCmn.inputBufIx+n+2 ], 1);
}
}
psEnc->state_Fxx[ 0 ].sCmn.inputBufIx += nSamplesToBuffer;
} else {
silk_assert( encControl->nChannelsAPI == 1 && encControl->nChannelsInternal == 1 );
silk_memcpy(buf, samplesIn, nSamplesFromInput*sizeof(opus_int16));
ret += silk_resampler( &psEnc->state_Fxx[ 0 ].sCmn.resampler_state,
&psEnc->state_Fxx[ 0 ].sCmn.inputBuf[ psEnc->state_Fxx[ 0 ].sCmn.inputBufIx + 2 ], buf, nSamplesFromInput );
psEnc->state_Fxx[ 0 ].sCmn.inputBufIx += nSamplesToBuffer;
}
samplesIn += nSamplesFromInput * encControl->nChannelsAPI;
nSamplesIn -= nSamplesFromInput;
/* Default */
psEnc->allowBandwidthSwitch = 0;
/* Silk encoder */
if( psEnc->state_Fxx[ 0 ].sCmn.inputBufIx >= psEnc->state_Fxx[ 0 ].sCmn.frame_length ) {
/* Enough data in input buffer, so encode */
silk_assert( psEnc->state_Fxx[ 0 ].sCmn.inputBufIx == psEnc->state_Fxx[ 0 ].sCmn.frame_length );
silk_assert( encControl->nChannelsInternal == 1 || psEnc->state_Fxx[ 1 ].sCmn.inputBufIx == psEnc->state_Fxx[ 1 ].sCmn.frame_length );
/* Deal with LBRR data */
if( psEnc->state_Fxx[ 0 ].sCmn.nFramesEncoded == 0 && !prefillFlag ) {
/* Create space at start of payload for VAD and FEC flags */
opus_uint8 iCDF[ 2 ] = { 0, 0 };
iCDF[ 0 ] = 256 - silk_RSHIFT( 256, ( psEnc->state_Fxx[ 0 ].sCmn.nFramesPerPacket + 1 ) * encControl->nChannelsInternal );
ec_enc_icdf( psRangeEnc, 0, iCDF, 8 );
/* Encode any LBRR data from previous packet */
/* Encode LBRR flags */
for( n = 0; n < encControl->nChannelsInternal; n++ ) {
LBRR_symbol = 0;
for( i = 0; i < psEnc->state_Fxx[ n ].sCmn.nFramesPerPacket; i++ ) {
LBRR_symbol |= silk_LSHIFT( psEnc->state_Fxx[ n ].sCmn.LBRR_flags[ i ], i );
}
psEnc->state_Fxx[ n ].sCmn.LBRR_flag = LBRR_symbol > 0 ? 1 : 0;
if( LBRR_symbol && psEnc->state_Fxx[ n ].sCmn.nFramesPerPacket > 1 ) {
ec_enc_icdf( psRangeEnc, LBRR_symbol - 1, silk_LBRR_flags_iCDF_ptr[ psEnc->state_Fxx[ n ].sCmn.nFramesPerPacket - 2 ], 8 );
}
}
/* Code LBRR indices and excitation signals */
for( i = 0; i < psEnc->state_Fxx[ 0 ].sCmn.nFramesPerPacket; i++ ) {
for( n = 0; n < encControl->nChannelsInternal; n++ ) {
if( psEnc->state_Fxx[ n ].sCmn.LBRR_flags[ i ] ) {
opus_int condCoding;
if( encControl->nChannelsInternal == 2 && n == 0 ) {
silk_stereo_encode_pred( psRangeEnc, psEnc->sStereo.predIx[ i ] );
/* For LBRR data there's no need to code the mid-only flag if the side-channel LBRR flag is set */
if( psEnc->state_Fxx[ 1 ].sCmn.LBRR_flags[ i ] == 0 ) {
silk_stereo_encode_mid_only( psRangeEnc, psEnc->sStereo.mid_only_flags[ i ] );
}
}
/* Use conditional coding if previous frame available */
if( i > 0 && psEnc->state_Fxx[ n ].sCmn.LBRR_flags[ i - 1 ] ) {
condCoding = CODE_CONDITIONALLY;
} else {
condCoding = CODE_INDEPENDENTLY;
}
silk_encode_indices( &psEnc->state_Fxx[ n ].sCmn, psRangeEnc, i, 1, condCoding );
silk_encode_pulses( psRangeEnc, psEnc->state_Fxx[ n ].sCmn.indices_LBRR[i].signalType, psEnc->state_Fxx[ n ].sCmn.indices_LBRR[i].quantOffsetType,
psEnc->state_Fxx[ n ].sCmn.pulses_LBRR[ i ], psEnc->state_Fxx[ n ].sCmn.frame_length );
}
}
}
/* Reset LBRR flags */
for( n = 0; n < encControl->nChannelsInternal; n++ ) {
silk_memset( psEnc->state_Fxx[ n ].sCmn.LBRR_flags, 0, sizeof( psEnc->state_Fxx[ n ].sCmn.LBRR_flags ) );
}
psEnc->nBitsUsedLBRR = ec_tell( psRangeEnc );
}
silk_HP_variable_cutoff( psEnc->state_Fxx );
/* Total target bits for packet */
nBits = silk_DIV32_16( silk_MUL( encControl->bitRate, encControl->payloadSize_ms ), 1000 );
/* Subtract bits used for LBRR */
if( !prefillFlag ) {
nBits -= psEnc->nBitsUsedLBRR;
}
/* Divide by number of uncoded frames left in packet */
nBits = silk_DIV32_16( nBits, psEnc->state_Fxx[ 0 ].sCmn.nFramesPerPacket );
/* Convert to bits/second */
if( encControl->payloadSize_ms == 10 ) {
TargetRate_bps = silk_SMULBB( nBits, 100 );
} else {
TargetRate_bps = silk_SMULBB( nBits, 50 );
}
/* Subtract fraction of bits in excess of target in previous frames and packets */
TargetRate_bps -= silk_DIV32_16( silk_MUL( psEnc->nBitsExceeded, 1000 ), BITRESERVOIR_DECAY_TIME_MS );
if( !prefillFlag && psEnc->state_Fxx[ 0 ].sCmn.nFramesEncoded > 0 ) {
/* Compare actual vs target bits so far in this packet */
opus_int32 bitsBalance = ec_tell( psRangeEnc ) - psEnc->nBitsUsedLBRR - nBits * psEnc->state_Fxx[ 0 ].sCmn.nFramesEncoded;
TargetRate_bps -= silk_DIV32_16( silk_MUL( bitsBalance, 1000 ), BITRESERVOIR_DECAY_TIME_MS );
}
/* Never exceed input bitrate */
TargetRate_bps = silk_LIMIT( TargetRate_bps, encControl->bitRate, 5000 );
/* Convert Left/Right to Mid/Side */
if( encControl->nChannelsInternal == 2 ) {
silk_stereo_LR_to_MS( &psEnc->sStereo, &psEnc->state_Fxx[ 0 ].sCmn.inputBuf[ 2 ], &psEnc->state_Fxx[ 1 ].sCmn.inputBuf[ 2 ],
psEnc->sStereo.predIx[ psEnc->state_Fxx[ 0 ].sCmn.nFramesEncoded ], &psEnc->sStereo.mid_only_flags[ psEnc->state_Fxx[ 0 ].sCmn.nFramesEncoded ],
MStargetRates_bps, TargetRate_bps, psEnc->state_Fxx[ 0 ].sCmn.speech_activity_Q8, encControl->toMono,
psEnc->state_Fxx[ 0 ].sCmn.fs_kHz, psEnc->state_Fxx[ 0 ].sCmn.frame_length );
if( psEnc->sStereo.mid_only_flags[ psEnc->state_Fxx[ 0 ].sCmn.nFramesEncoded ] == 0 ) {
/* Reset side channel encoder memory for first frame with side coding */
if( psEnc->prev_decode_only_middle == 1 ) {
silk_memset( &psEnc->state_Fxx[ 1 ].sShape, 0, sizeof( psEnc->state_Fxx[ 1 ].sShape ) );
silk_memset( &psEnc->state_Fxx[ 1 ].sPrefilt, 0, sizeof( psEnc->state_Fxx[ 1 ].sPrefilt ) );
silk_memset( &psEnc->state_Fxx[ 1 ].sCmn.sNSQ, 0, sizeof( psEnc->state_Fxx[ 1 ].sCmn.sNSQ ) );
silk_memset( psEnc->state_Fxx[ 1 ].sCmn.prev_NLSFq_Q15, 0, sizeof( psEnc->state_Fxx[ 1 ].sCmn.prev_NLSFq_Q15 ) );
silk_memset( &psEnc->state_Fxx[ 1 ].sCmn.sLP.In_LP_State, 0, sizeof( psEnc->state_Fxx[ 1 ].sCmn.sLP.In_LP_State ) );
psEnc->state_Fxx[ 1 ].sCmn.prevLag = 100;
psEnc->state_Fxx[ 1 ].sCmn.sNSQ.lagPrev = 100;
psEnc->state_Fxx[ 1 ].sShape.LastGainIndex = 10;
psEnc->state_Fxx[ 1 ].sCmn.prevSignalType = TYPE_NO_VOICE_ACTIVITY;
psEnc->state_Fxx[ 1 ].sCmn.sNSQ.prev_gain_Q16 = 65536;
psEnc->state_Fxx[ 1 ].sCmn.first_frame_after_reset = 1;
}
silk_encode_do_VAD_Fxx( &psEnc->state_Fxx[ 1 ] );
} else {
psEnc->state_Fxx[ 1 ].sCmn.VAD_flags[ psEnc->state_Fxx[ 0 ].sCmn.nFramesEncoded ] = 0;
}
if( !prefillFlag ) {
silk_stereo_encode_pred( psRangeEnc, psEnc->sStereo.predIx[ psEnc->state_Fxx[ 0 ].sCmn.nFramesEncoded ] );
if( psEnc->state_Fxx[ 1 ].sCmn.VAD_flags[ psEnc->state_Fxx[ 0 ].sCmn.nFramesEncoded ] == 0 ) {
silk_stereo_encode_mid_only( psRangeEnc, psEnc->sStereo.mid_only_flags[ psEnc->state_Fxx[ 0 ].sCmn.nFramesEncoded ] );
}
}
} else {
/* Buffering */
silk_memcpy( psEnc->state_Fxx[ 0 ].sCmn.inputBuf, psEnc->sStereo.sMid, 2 * sizeof( opus_int16 ) );
silk_memcpy( psEnc->sStereo.sMid, &psEnc->state_Fxx[ 0 ].sCmn.inputBuf[ psEnc->state_Fxx[ 0 ].sCmn.frame_length ], 2 * sizeof( opus_int16 ) );
}
silk_encode_do_VAD_Fxx( &psEnc->state_Fxx[ 0 ] );
/* Encode */
for( n = 0; n < encControl->nChannelsInternal; n++ ) {
opus_int maxBits, useCBR;
/* Handling rate constraints */
maxBits = encControl->maxBits;
if( tot_blocks == 2 && curr_block == 0 ) {
maxBits = maxBits * 3 / 5;
} else if( tot_blocks == 3 ) {
if( curr_block == 0 ) {
maxBits = maxBits * 2 / 5;
} else if( curr_block == 1 ) {
maxBits = maxBits * 3 / 4;
}
}
useCBR = encControl->useCBR && curr_block == tot_blocks - 1;
if( encControl->nChannelsInternal == 1 ) {
channelRate_bps = TargetRate_bps;
} else {
channelRate_bps = MStargetRates_bps[ n ];
if( n == 0 && MStargetRates_bps[ 1 ] > 0 ) {
useCBR = 0;
/* Give mid up to 1/2 of the max bits for that frame */
maxBits -= encControl->maxBits / ( tot_blocks * 2 );
}
}
if( channelRate_bps > 0 ) {
opus_int condCoding;
silk_control_SNR( &psEnc->state_Fxx[ n ].sCmn, channelRate_bps );
/* Use independent coding if no previous frame available */
if( psEnc->state_Fxx[ 0 ].sCmn.nFramesEncoded - n <= 0 ) {
condCoding = CODE_INDEPENDENTLY;
} else if( n > 0 && psEnc->prev_decode_only_middle ) {
/* If we skipped a side frame in this packet, we don't
need LTP scaling; the LTP state is well-defined. */
condCoding = CODE_INDEPENDENTLY_NO_LTP_SCALING;
} else {
condCoding = CODE_CONDITIONALLY;
}
if( ( ret = silk_encode_frame_Fxx( &psEnc->state_Fxx[ n ], nBytesOut, psRangeEnc, condCoding, maxBits, useCBR ) ) != 0 ) {
silk_assert( 0 );
}
}
psEnc->state_Fxx[ n ].sCmn.controlled_since_last_payload = 0;
psEnc->state_Fxx[ n ].sCmn.inputBufIx = 0;
psEnc->state_Fxx[ n ].sCmn.nFramesEncoded++;
}
psEnc->prev_decode_only_middle = psEnc->sStereo.mid_only_flags[ psEnc->state_Fxx[ 0 ].sCmn.nFramesEncoded - 1 ];
/* Insert VAD and FEC flags at beginning of bitstream */
if( *nBytesOut > 0 && psEnc->state_Fxx[ 0 ].sCmn.nFramesEncoded == psEnc->state_Fxx[ 0 ].sCmn.nFramesPerPacket) {
flags = 0;
for( n = 0; n < encControl->nChannelsInternal; n++ ) {
for( i = 0; i < psEnc->state_Fxx[ n ].sCmn.nFramesPerPacket; i++ ) {
flags = silk_LSHIFT( flags, 1 );
flags |= psEnc->state_Fxx[ n ].sCmn.VAD_flags[ i ];
}
flags = silk_LSHIFT( flags, 1 );
flags |= psEnc->state_Fxx[ n ].sCmn.LBRR_flag;
}
if( !prefillFlag ) {
ec_enc_patch_initial_bits( psRangeEnc, flags, ( psEnc->state_Fxx[ 0 ].sCmn.nFramesPerPacket + 1 ) * encControl->nChannelsInternal );
}
/* Return zero bytes if all channels DTXed */
if( psEnc->state_Fxx[ 0 ].sCmn.inDTX && ( encControl->nChannelsInternal == 1 || psEnc->state_Fxx[ 1 ].sCmn.inDTX ) ) {
*nBytesOut = 0;
}
psEnc->nBitsExceeded += *nBytesOut * 8;
psEnc->nBitsExceeded -= silk_DIV32_16( silk_MUL( encControl->bitRate, encControl->payloadSize_ms ), 1000 );
psEnc->nBitsExceeded = silk_LIMIT( psEnc->nBitsExceeded, 0, 10000 );
/* Update flag indicating if bandwidth switching is allowed */
speech_act_thr_for_switch_Q8 = (opus_int) silk_SMLAWB( SILK_FIX_CONST( SPEECH_ACTIVITY_DTX_THRES, 8 ),
SILK_FIX_CONST( ( 1 - SPEECH_ACTIVITY_DTX_THRES ) / MAX_BANDWIDTH_SWITCH_DELAY_MS, 16 + 8 ), psEnc->timeSinceSwitchAllowed_ms );
if( psEnc->state_Fxx[ 0 ].sCmn.speech_activity_Q8 < speech_act_thr_for_switch_Q8 ) {
psEnc->allowBandwidthSwitch = 1;
psEnc->timeSinceSwitchAllowed_ms = 0;
} else {
psEnc->allowBandwidthSwitch = 0;
psEnc->timeSinceSwitchAllowed_ms += encControl->payloadSize_ms;
}
}
if( nSamplesIn == 0 ) {
break;
}
} else {
break;
}
curr_block++;
}
psEnc->nPrevChannelsInternal = encControl->nChannelsInternal;
encControl->allowBandwidthSwitch = psEnc->allowBandwidthSwitch;
encControl->inWBmodeWithoutVariableLP = psEnc->state_Fxx[ 0 ].sCmn.fs_kHz == 16 && psEnc->state_Fxx[ 0 ].sCmn.sLP.mode == 0;
encControl->internalSampleRate = silk_SMULBB( psEnc->state_Fxx[ 0 ].sCmn.fs_kHz, 1000 );
encControl->stereoWidth_Q14 = encControl->toMono ? 0 : psEnc->sStereo.smth_width_Q14;
if( prefillFlag ) {
encControl->payloadSize_ms = tmp_payloadSize_ms;
encControl->complexity = tmp_complexity;
for( n = 0; n < encControl->nChannelsInternal; n++ ) {
psEnc->state_Fxx[ n ].sCmn.controlled_since_last_payload = 0;
psEnc->state_Fxx[ n ].sCmn.prefillFlag = 0;
}
}
RESTORE_STACK;
return ret;
}
void silk_find_LTP_FIX(
opus_int32 XXLTP_Q17[ MAX_NB_SUBFR * LTP_ORDER * LTP_ORDER ], /* O Correlation matrix */
opus_int32 xXLTP_Q17[ MAX_NB_SUBFR * LTP_ORDER ], /* O Correlation vector */
const opus_int16 r_ptr[], /* I Residual signal after LPC */
const opus_int lag[ MAX_NB_SUBFR ], /* I LTP lags */
const opus_int subfr_length, /* I Subframe length */
const opus_int nb_subfr, /* I Number of subframes */
int arch /* I Run-time architecture */
)
{
opus_int i, k, extra_shifts;
opus_int xx_shifts, xX_shifts, XX_shifts;
const opus_int16 *lag_ptr;
opus_int32 *XXLTP_Q17_ptr, *xXLTP_Q17_ptr;
opus_int32 xx, nrg, temp;
xXLTP_Q17_ptr = xXLTP_Q17;
XXLTP_Q17_ptr = XXLTP_Q17;
for( k = 0; k < nb_subfr; k++ ) {
lag_ptr = r_ptr - ( lag[ k ] + LTP_ORDER / 2 );
silk_sum_sqr_shift( &xx, &xx_shifts, r_ptr, subfr_length + LTP_ORDER ); /* xx in Q( -xx_shifts ) */
silk_corrMatrix_FIX( lag_ptr, subfr_length, LTP_ORDER, XXLTP_Q17_ptr, &nrg, &XX_shifts, arch ); /* XXLTP_Q17_ptr and nrg in Q( -XX_shifts ) */
extra_shifts = xx_shifts - XX_shifts;
if( extra_shifts > 0 ) {
/* Shift XX */
xX_shifts = xx_shifts;
for( i = 0; i < LTP_ORDER * LTP_ORDER; i++ ) {
XXLTP_Q17_ptr[ i ] = silk_RSHIFT32( XXLTP_Q17_ptr[ i ], extra_shifts ); /* Q( -xX_shifts ) */
}
nrg = silk_RSHIFT32( nrg, extra_shifts ); /* Q( -xX_shifts ) */
} else if( extra_shifts < 0 ) {
/* Shift xx */
xX_shifts = XX_shifts;
xx = silk_RSHIFT32( xx, -extra_shifts ); /* Q( -xX_shifts ) */
} else {
xX_shifts = xx_shifts;
}
silk_corrVector_FIX( lag_ptr, r_ptr, subfr_length, LTP_ORDER, xXLTP_Q17_ptr, xX_shifts, arch ); /* xXLTP_Q17_ptr in Q( -xX_shifts ) */
/* At this point all correlations are in Q(-xX_shifts) */
temp = silk_SMLAWB( 1, nrg, SILK_FIX_CONST( LTP_CORR_INV_MAX, 16 ) );
temp = silk_max( temp, xx );
TIC(div)
#if 0
for( i = 0; i < LTP_ORDER * LTP_ORDER; i++ ) {
XXLTP_Q17_ptr[ i ] = silk_DIV32_varQ( XXLTP_Q17_ptr[ i ], temp, 17 );
}
for( i = 0; i < LTP_ORDER; i++ ) {
xXLTP_Q17_ptr[ i ] = silk_DIV32_varQ( xXLTP_Q17_ptr[ i ], temp, 17 );
}
#else
for( i = 0; i < LTP_ORDER * LTP_ORDER; i++ ) {
XXLTP_Q17_ptr[ i ] = (opus_int32)( silk_LSHIFT64( (opus_int64)XXLTP_Q17_ptr[ i ], 17 ) / temp );
}
for( i = 0; i < LTP_ORDER; i++ ) {
xXLTP_Q17_ptr[ i ] = (opus_int32)( silk_LSHIFT64( (opus_int64)xXLTP_Q17_ptr[ i ], 17 ) / temp );
}
#endif
TOC(div)
r_ptr += subfr_length;
XXLTP_Q17_ptr += LTP_ORDER * LTP_ORDER;
xXLTP_Q17_ptr += LTP_ORDER;
}
}
