/* Step up function, converts reflection coefficients to prediction coefficients */
void silk_k2a(
opus_int32 *A_Q24, /* O Prediction coefficients [order] Q24 */
const opus_int16 *rc_Q15, /* I Reflection coefficients [order] Q15 */
const opus_int32 order /* I Prediction order */
)
{
opus_int k, n;
opus_int32 Atmp[ SILK_MAX_ORDER_LPC ];
for( k = 0; k < order; k++ ) {
for( n = 0; n < k; n++ ) {
Atmp[ n ] = A_Q24[ n ];
}
for( n = 0; n < k; n++ ) {
A_Q24[ n ] = silk_SMLAWB( A_Q24[ n ], silk_LSHIFT( Atmp[ k - n - 1 ], 1 ), rc_Q15[ k ] );
}
A_Q24[ k ] = -silk_LSHIFT( (opus_int32)rc_Q15[ k ], 9 );
}
}
void silk_warped_LPC_analysis_filter_FIX(
opus_int32 state[], /* I/O State [order + 1] */
opus_int32 res_Q2[], /* O Residual signal [length] */
const opus_int16 coef_Q13[], /* I Coefficients [order] */
const opus_int16 input[], /* I Input signal [length] */
const opus_int16 lambda_Q16, /* I Warping factor */
const opus_int length, /* I Length of input signal */
const opus_int order /* I Filter order (even) */
) {
opus_int n, i;
opus_int32 acc_Q11, tmp1, tmp2;
/* Order must be even */
silk_assert((order & 1) == 0);
for (n = 0; n < length; n++) {
/* Output of lowpass section */
tmp2 = silk_SMLAWB(state[0], state[1], lambda_Q16);
state[0] = silk_LSHIFT(input[n], 14);
/* Output of allpass section */
tmp1 = silk_SMLAWB(state[1], state[2] - tmp2, lambda_Q16);
state[1] = tmp2;
acc_Q11 = silk_RSHIFT(order, 1);
acc_Q11 = silk_SMLAWB(acc_Q11, tmp2, coef_Q13[0]);
/* Loop over allpass sections */
for (i = 2; i < order; i += 2) {
/* Output of allpass section */
tmp2 = silk_SMLAWB(state[i], state[i + 1] - tmp1, lambda_Q16);
state[i] = tmp1;
acc_Q11 = silk_SMLAWB(acc_Q11, tmp1, coef_Q13[i - 1]);
/* Output of allpass section */
tmp1 = silk_SMLAWB(state[i + 1], state[i + 2] - tmp2, lambda_Q16);
state[i + 1] = tmp2;
acc_Q11 = silk_SMLAWB(acc_Q11, tmp2, coef_Q13[i]);
}
state[order] = tmp1;
acc_Q11 = silk_SMLAWB(acc_Q11, tmp1, coef_Q13[order - 1]);
res_Q2[n] = silk_LSHIFT((opus_int32) input[n], 2) - silk_RSHIFT_ROUND(acc_Q11, 9);
}
}
/* Predictive dequantizer for NLSF residuals */
static OPUS_INLINE void silk_NLSF_residual_dequant( /* O Returns RD value in Q30 */
opus_int16 x_Q10[], /* O Output [ order ] */
const opus_int8 indices[], /* I Quantization indices [ order ] */
const opus_uint8 pred_coef_Q8[], /* I Backward predictor coefs [ order ] */
const opus_int quant_step_size_Q16, /* I Quantization step size */
const opus_int16 order /* I Number of input values */
)
{
opus_int i, out_Q10, pred_Q10;
out_Q10 = 0;
for( i = order-1; i >= 0; i-- ) {
pred_Q10 = silk_RSHIFT( silk_SMULBB( out_Q10, (opus_int16)pred_coef_Q8[ i ] ), 8 );
out_Q10 = silk_LSHIFT( indices[ i ], 10 );
if( out_Q10 > 0 ) {
out_Q10 = silk_SUB16( out_Q10, SILK_FIX_CONST( NLSF_QUANT_LEVEL_ADJ, 10 ) );
} else if( out_Q10 < 0 ) {
out_Q10 = silk_ADD16( out_Q10, SILK_FIX_CONST( NLSF_QUANT_LEVEL_ADJ, 10 ) );
}
out_Q10 = silk_SMLAWB( pred_Q10, (opus_int32)out_Q10, quant_step_size_Q16 );
x_Q10[ i ] = out_Q10;
}
}
/* Split signal into two decimated bands using first-order allpass filters */
void silk_ana_filt_bank_1(
const opus_int16 *in, /* I Input signal [N] */
opus_int32 *S, /* I/O State vector [2] */
opus_int16 *outL, /* O Low band [N/2] */
opus_int16 *outH, /* O High band [N/2] */
const opus_int32 N /* I Number of input samples */
)
{
opus_int k, N2 = silk_RSHIFT( N, 1 );
opus_int32 in32, X, Y, out_1, out_2;
/* Internal variables and state are in Q10 format */
for( k = 0; k < N2; k++ ) {
/* Convert to Q10 */
in32 = silk_LSHIFT( (opus_int32)in[ 2 * k ], 10 );
/* All-pass section for even input sample */
Y = silk_SUB32( in32, S[ 0 ] );
X = silk_SMLAWB( Y, Y, A_fb1_21 );
out_1 = silk_ADD32( S[ 0 ], X );
S[ 0 ] = silk_ADD32( in32, X );
/* Convert to Q10 */
in32 = silk_LSHIFT( (opus_int32)in[ 2 * k + 1 ], 10 );
/* All-pass section for odd input sample, and add to output of previous section */
Y = silk_SUB32( in32, S[ 1 ] );
X = silk_SMULWB( Y, A_fb1_20 );
out_2 = silk_ADD32( S[ 1 ], X );
S[ 1 ] = silk_ADD32( in32, X );
/* Add/subtract, convert back to int16 and store to output */
outL[ k ] = (opus_int16)silk_SAT16( silk_RSHIFT_ROUND( silk_ADD32( out_2, out_1 ), 11 ) );
outH[ k ] = (opus_int16)silk_SAT16( silk_RSHIFT_ROUND( silk_SUB32( out_2, out_1 ), 11 ) );
}
}
/* Downsample by a factor 2/3, low quality */
void silk_resampler_down2_3(
opus_int32 *S, /* I/O State vector [ 6 ] */
opus_int16 *out, /* O Output signal [ floor(2*inLen/3) ] */
const opus_int16 *in, /* I Input signal [ inLen ] */
opus_int32 inLen /* I Number of input samples */
)
{
opus_int32 nSamplesIn, counter, res_Q6;
VARDECL( opus_int32, buf );
opus_int32 *buf_ptr;
SAVE_STACK;
ALLOC( buf, RESAMPLER_MAX_BATCH_SIZE_IN + ORDER_FIR, opus_int32 );
/* Copy buffered samples to start of buffer */
silk_memcpy( buf, S, ORDER_FIR * sizeof( opus_int32 ) );
/* Iterate over blocks of frameSizeIn input samples */
while( 1 ) {
nSamplesIn = silk_min( inLen, RESAMPLER_MAX_BATCH_SIZE_IN );
/* Second-order AR filter (output in Q8) */
silk_resampler_private_AR2( &S[ ORDER_FIR ], &buf[ ORDER_FIR ], in,
silk_Resampler_2_3_COEFS_LQ, nSamplesIn );
/* Interpolate filtered signal */
buf_ptr = buf;
counter = nSamplesIn;
while( counter > 2 ) {
/* Inner product */
res_Q6 = silk_SMULWB( buf_ptr[ 0 ], silk_Resampler_2_3_COEFS_LQ[ 2 ] );
res_Q6 = silk_SMLAWB( res_Q6, buf_ptr[ 1 ], silk_Resampler_2_3_COEFS_LQ[ 3 ] );
res_Q6 = silk_SMLAWB( res_Q6, buf_ptr[ 2 ], silk_Resampler_2_3_COEFS_LQ[ 5 ] );
res_Q6 = silk_SMLAWB( res_Q6, buf_ptr[ 3 ], silk_Resampler_2_3_COEFS_LQ[ 4 ] );
/* Scale down, saturate and store in output array */
*out++ = (opus_int16)silk_SAT16( silk_RSHIFT_ROUND( res_Q6, 6 ) );
res_Q6 = silk_SMULWB( buf_ptr[ 1 ], silk_Resampler_2_3_COEFS_LQ[ 4 ] );
res_Q6 = silk_SMLAWB( res_Q6, buf_ptr[ 2 ], silk_Resampler_2_3_COEFS_LQ[ 5 ] );
res_Q6 = silk_SMLAWB( res_Q6, buf_ptr[ 3 ], silk_Resampler_2_3_COEFS_LQ[ 3 ] );
res_Q6 = silk_SMLAWB( res_Q6, buf_ptr[ 4 ], silk_Resampler_2_3_COEFS_LQ[ 2 ] );
/* Scale down, saturate and store in output array */
*out++ = (opus_int16)silk_SAT16( silk_RSHIFT_ROUND( res_Q6, 6 ) );
buf_ptr += 3;
counter -= 3;
}
in += nSamplesIn;
inLen -= nSamplesIn;
if( inLen > 0 ) {
/* More iterations to do; copy last part of filtered signal to beginning of buffer */
silk_memcpy( buf, &buf[ nSamplesIn ], ORDER_FIR * sizeof( opus_int32 ) );
} else {
break;
}
}
/* Copy last part of filtered signal to the state for the next call */
silk_memcpy( S, &buf[ nSamplesIn ], ORDER_FIR * sizeof( opus_int32 ) );
RESTORE_STACK;
}
/* notch filter just above Nyquist. */
void silk_resampler_private_up2_HQ(
opus_int32 *S, /* I/O Resampler state [ 6 ] */
opus_int16 *out, /* O Output signal [ 2 * len ] */
const opus_int16 *in, /* I Input signal [ len ] */
opus_int32 len /* I Number of input samples */
)
{
opus_int32 k;
opus_int32 in32, out32_1, out32_2, Y, X;
silk_assert( silk_resampler_up2_hq_0[ 0 ] > 0 );
silk_assert( silk_resampler_up2_hq_0[ 1 ] > 0 );
silk_assert( silk_resampler_up2_hq_0[ 2 ] < 0 );
silk_assert( silk_resampler_up2_hq_1[ 0 ] > 0 );
silk_assert( silk_resampler_up2_hq_1[ 1 ] > 0 );
silk_assert( silk_resampler_up2_hq_1[ 2 ] < 0 );
/* Internal variables and state are in Q10 format */
for( k = 0; k < len; k++ ) {
/* Convert to Q10 */
in32 = silk_LSHIFT( (opus_int32)in[ k ], 10 );
/* First all-pass section for even output sample */
Y = silk_SUB32( in32, S[ 0 ] );
X = silk_SMULWB( Y, silk_resampler_up2_hq_0[ 0 ] );
out32_1 = silk_ADD32( S[ 0 ], X );
S[ 0 ] = silk_ADD32( in32, X );
/* Second all-pass section for even output sample */
Y = silk_SUB32( out32_1, S[ 1 ] );
X = silk_SMULWB( Y, silk_resampler_up2_hq_0[ 1 ] );
out32_2 = silk_ADD32( S[ 1 ], X );
S[ 1 ] = silk_ADD32( out32_1, X );
/* Third all-pass section for even output sample */
Y = silk_SUB32( out32_2, S[ 2 ] );
X = silk_SMLAWB( Y, Y, silk_resampler_up2_hq_0[ 2 ] );
out32_1 = silk_ADD32( S[ 2 ], X );
S[ 2 ] = silk_ADD32( out32_2, X );
/* Apply gain in Q15, convert back to int16 and store to output */
out[ 2 * k ] = (opus_int16)silk_SAT16( silk_RSHIFT_ROUND( out32_1, 10 ) );
/* First all-pass section for odd output sample */
Y = silk_SUB32( in32, S[ 3 ] );
X = silk_SMULWB( Y, silk_resampler_up2_hq_1[ 0 ] );
out32_1 = silk_ADD32( S[ 3 ], X );
S[ 3 ] = silk_ADD32( in32, X );
/* Second all-pass section for odd output sample */
Y = silk_SUB32( out32_1, S[ 4 ] );
X = silk_SMULWB( Y, silk_resampler_up2_hq_1[ 1 ] );
out32_2 = silk_ADD32( S[ 4 ], X );
S[ 4 ] = silk_ADD32( out32_1, X );
/* Third all-pass section for odd output sample */
Y = silk_SUB32( out32_2, S[ 5 ] );
X = silk_SMLAWB( Y, Y, silk_resampler_up2_hq_1[ 2 ] );
out32_1 = silk_ADD32( S[ 5 ], X );
S[ 5 ] = silk_ADD32( out32_2, X );
/* Apply gain in Q15, convert back to int16 and store to output */
out[ 2 * k + 1 ] = (opus_int16)silk_SAT16( silk_RSHIFT_ROUND( out32_1, 10 ) );
}
}
/* Entropy constrained matrix-weighted VQ, hard-coded to 5-element vectors, for a single input data vector */
void silk_VQ_WMat_EC(
opus_int8 *ind, /* O index of best codebook vector */
opus_int32 *rate_dist_Q14, /* O best weighted quant error + mu * rate */
opus_int *gain_Q7, /* O sum of absolute LTP coefficients */
const opus_int16 *in_Q14, /* I input vector to be quantized */
const opus_int32 *W_Q18, /* I weighting matrix */
const opus_int8 *cb_Q7, /* I codebook */
const opus_uint8 *cb_gain_Q7, /* I codebook effective gain */
const opus_uint8 *cl_Q5, /* I code length for each codebook vector */
const opus_int mu_Q9, /* I tradeoff betw. weighted error and rate */
const opus_int32 max_gain_Q7, /* I maximum sum of absolute LTP coefficients */
opus_int L /* I number of vectors in codebook */
)
{
opus_int k, gain_tmp_Q7;
const opus_int8 *cb_row_Q7;
opus_int16 diff_Q14[ 5 ];
opus_int32 sum1_Q14, sum2_Q16;
/* Loop over codebook */
*rate_dist_Q14 = silk_int32_MAX;
cb_row_Q7 = cb_Q7;
for( k = 0; k < L; k++ ) {
gain_tmp_Q7 = cb_gain_Q7[k];
diff_Q14[ 0 ] = in_Q14[ 0 ] - silk_LSHIFT( cb_row_Q7[ 0 ], 7 );
diff_Q14[ 1 ] = in_Q14[ 1 ] - silk_LSHIFT( cb_row_Q7[ 1 ], 7 );
diff_Q14[ 2 ] = in_Q14[ 2 ] - silk_LSHIFT( cb_row_Q7[ 2 ], 7 );
diff_Q14[ 3 ] = in_Q14[ 3 ] - silk_LSHIFT( cb_row_Q7[ 3 ], 7 );
diff_Q14[ 4 ] = in_Q14[ 4 ] - silk_LSHIFT( cb_row_Q7[ 4 ], 7 );
/* Weighted rate */
sum1_Q14 = silk_SMULBB( mu_Q9, cl_Q5[ k ] );
/* Penalty for too large gain */
sum1_Q14 = silk_ADD_LSHIFT32( sum1_Q14, silk_max( silk_SUB32( gain_tmp_Q7, max_gain_Q7 ), 0 ), 10 );
silk_assert( sum1_Q14 >= 0 );
/* first row of W_Q18 */
sum2_Q16 = silk_SMULWB( W_Q18[ 1 ], diff_Q14[ 1 ] );
sum2_Q16 = silk_SMLAWB( sum2_Q16, W_Q18[ 2 ], diff_Q14[ 2 ] );
sum2_Q16 = silk_SMLAWB( sum2_Q16, W_Q18[ 3 ], diff_Q14[ 3 ] );
sum2_Q16 = silk_SMLAWB( sum2_Q16, W_Q18[ 4 ], diff_Q14[ 4 ] );
sum2_Q16 = silk_LSHIFT( sum2_Q16, 1 );
sum2_Q16 = silk_SMLAWB( sum2_Q16, W_Q18[ 0 ], diff_Q14[ 0 ] );
sum1_Q14 = silk_SMLAWB( sum1_Q14, sum2_Q16, diff_Q14[ 0 ] );
/* second row of W_Q18 */
sum2_Q16 = silk_SMULWB( W_Q18[ 7 ], diff_Q14[ 2 ] );
sum2_Q16 = silk_SMLAWB( sum2_Q16, W_Q18[ 8 ], diff_Q14[ 3 ] );
sum2_Q16 = silk_SMLAWB( sum2_Q16, W_Q18[ 9 ], diff_Q14[ 4 ] );
sum2_Q16 = silk_LSHIFT( sum2_Q16, 1 );
sum2_Q16 = silk_SMLAWB( sum2_Q16, W_Q18[ 6 ], diff_Q14[ 1 ] );
sum1_Q14 = silk_SMLAWB( sum1_Q14, sum2_Q16, diff_Q14[ 1 ] );
/* third row of W_Q18 */
sum2_Q16 = silk_SMULWB( W_Q18[ 13 ], diff_Q14[ 3 ] );
sum2_Q16 = silk_SMLAWB( sum2_Q16, W_Q18[ 14 ], diff_Q14[ 4 ] );
sum2_Q16 = silk_LSHIFT( sum2_Q16, 1 );
sum2_Q16 = silk_SMLAWB( sum2_Q16, W_Q18[ 12 ], diff_Q14[ 2 ] );
sum1_Q14 = silk_SMLAWB( sum1_Q14, sum2_Q16, diff_Q14[ 2 ] );
/* fourth row of W_Q18 */
sum2_Q16 = silk_SMULWB( W_Q18[ 19 ], diff_Q14[ 4 ] );
sum2_Q16 = silk_LSHIFT( sum2_Q16, 1 );
sum2_Q16 = silk_SMLAWB( sum2_Q16, W_Q18[ 18 ], diff_Q14[ 3 ] );
sum1_Q14 = silk_SMLAWB( sum1_Q14, sum2_Q16, diff_Q14[ 3 ] );
/* last row of W_Q18 */
sum2_Q16 = silk_SMULWB( W_Q18[ 24 ], diff_Q14[ 4 ] );
sum1_Q14 = silk_SMLAWB( sum1_Q14, sum2_Q16, diff_Q14[ 4 ] );
silk_assert( sum1_Q14 >= 0 );
/* find best */
if( sum1_Q14 < *rate_dist_Q14 ) {
*rate_dist_Q14 = sum1_Q14;
*ind = (opus_int8)k;
*gain_Q7 = gain_tmp_Q7;
}
/* Go to next cbk vector */
cb_row_Q7 += LTP_ORDER;
}
}
/* Residual energy: nrg = wxx - 2 * wXx * c + c' * wXX * c */
opus_int32 silk_residual_energy16_covar_FIX(
const opus_int16 *c, /* I Prediction vector */
const opus_int32 *wXX, /* I Correlation matrix */
const opus_int32 *wXx, /* I Correlation vector */
opus_int32 wxx, /* I Signal energy */
opus_int D, /* I Dimension */
opus_int cQ /* I Q value for c vector 0 - 15 */
)
{
opus_int i, j, lshifts, Qxtra;
opus_int32 c_max, w_max, tmp, tmp2, nrg;
opus_int cn[ MAX_MATRIX_SIZE ];
const opus_int32 *pRow;
/* Safety checks */
silk_assert( D >= 0 );
silk_assert( D <= 16 );
silk_assert( cQ > 0 );
silk_assert( cQ < 16 );
lshifts = 16 - cQ;
Qxtra = lshifts;
c_max = 0;
for( i = 0; i < D; i++ ) {
c_max = silk_max_32( c_max, silk_abs( (opus_int32)c[ i ] ) );
}
Qxtra = silk_min_int( Qxtra, silk_CLZ32( c_max ) - 17 );
w_max = silk_max_32( wXX[ 0 ], wXX[ D * D - 1 ] );
Qxtra = silk_min_int( Qxtra, silk_CLZ32( silk_MUL( D, silk_RSHIFT( silk_SMULWB( w_max, c_max ), 4 ) ) ) - 5 );
Qxtra = silk_max_int( Qxtra, 0 );
for( i = 0; i < D; i++ ) {
cn[ i ] = silk_LSHIFT( ( opus_int )c[ i ], Qxtra );
silk_assert( silk_abs(cn[i]) <= ( silk_int16_MAX + 1 ) ); /* Check that silk_SMLAWB can be used */
}
lshifts -= Qxtra;
/* Compute wxx - 2 * wXx * c */
tmp = 0;
for( i = 0; i < D; i++ ) {
tmp = silk_SMLAWB( tmp, wXx[ i ], cn[ i ] );
}
nrg = silk_RSHIFT( wxx, 1 + lshifts ) - tmp; /* Q: -lshifts - 1 */
/* Add c' * wXX * c, assuming wXX is symmetric */
tmp2 = 0;
for( i = 0; i < D; i++ ) {
tmp = 0;
pRow = &wXX[ i * D ];
for( j = i + 1; j < D; j++ ) {
tmp = silk_SMLAWB( tmp, pRow[ j ], cn[ j ] );
}
tmp = silk_SMLAWB( tmp, silk_RSHIFT( pRow[ i ], 1 ), cn[ i ] );
tmp2 = silk_SMLAWB( tmp2, tmp, cn[ i ] );
}
nrg = silk_ADD_LSHIFT32( nrg, tmp2, lshifts ); /* Q: -lshifts - 1 */
/* Keep one bit free always, because we add them for LSF interpolation */
if( nrg < 1 ) {
nrg = 1;
} else if( nrg > silk_RSHIFT( silk_int32_MAX, lshifts + 2 ) ) {
nrg = silk_int32_MAX >> 1;
} else {
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;
}
}
static inline opus_int16 *silk_resampler_private_down_FIR_INTERPOL(
opus_int16 *out,
opus_int32 *buf,
const opus_int16 *FIR_Coefs,
opus_int FIR_Order,
opus_int FIR_Fracs,
opus_int32 max_index_Q16,
opus_int32 index_increment_Q16
)
{
opus_int32 index_Q16, res_Q6;
opus_int32 *buf_ptr;
opus_int32 interpol_ind;
const opus_int16 *interpol_ptr;
switch( FIR_Order ) {
case RESAMPLER_DOWN_ORDER_FIR0:
for( index_Q16 = 0; index_Q16 < max_index_Q16; index_Q16 += index_increment_Q16 ) {
/* Integer part gives pointer to buffered input */
buf_ptr = buf + silk_RSHIFT( index_Q16, 16 );
/* Fractional part gives interpolation coefficients */
interpol_ind = silk_SMULWB( index_Q16 & 0xFFFF, FIR_Fracs );
/* Inner product */
interpol_ptr = &FIR_Coefs[ RESAMPLER_DOWN_ORDER_FIR0 / 2 * interpol_ind ];
res_Q6 = silk_SMULWB( buf_ptr[ 0 ], interpol_ptr[ 0 ] );
res_Q6 = silk_SMLAWB( res_Q6, buf_ptr[ 1 ], interpol_ptr[ 1 ] );
res_Q6 = silk_SMLAWB( res_Q6, buf_ptr[ 2 ], interpol_ptr[ 2 ] );
res_Q6 = silk_SMLAWB( res_Q6, buf_ptr[ 3 ], interpol_ptr[ 3 ] );
res_Q6 = silk_SMLAWB( res_Q6, buf_ptr[ 4 ], interpol_ptr[ 4 ] );
res_Q6 = silk_SMLAWB( res_Q6, buf_ptr[ 5 ], interpol_ptr[ 5 ] );
res_Q6 = silk_SMLAWB( res_Q6, buf_ptr[ 6 ], interpol_ptr[ 6 ] );
res_Q6 = silk_SMLAWB( res_Q6, buf_ptr[ 7 ], interpol_ptr[ 7 ] );
res_Q6 = silk_SMLAWB( res_Q6, buf_ptr[ 8 ], interpol_ptr[ 8 ] );
interpol_ptr = &FIR_Coefs[ RESAMPLER_DOWN_ORDER_FIR0 / 2 * ( FIR_Fracs - 1 - interpol_ind ) ];
res_Q6 = silk_SMLAWB( res_Q6, buf_ptr[ 17 ], interpol_ptr[ 0 ] );
res_Q6 = silk_SMLAWB( res_Q6, buf_ptr[ 16 ], interpol_ptr[ 1 ] );
res_Q6 = silk_SMLAWB( res_Q6, buf_ptr[ 15 ], interpol_ptr[ 2 ] );
res_Q6 = silk_SMLAWB( res_Q6, buf_ptr[ 14 ], interpol_ptr[ 3 ] );
res_Q6 = silk_SMLAWB( res_Q6, buf_ptr[ 13 ], interpol_ptr[ 4 ] );
res_Q6 = silk_SMLAWB( res_Q6, buf_ptr[ 12 ], interpol_ptr[ 5 ] );
res_Q6 = silk_SMLAWB( res_Q6, buf_ptr[ 11 ], interpol_ptr[ 6 ] );
res_Q6 = silk_SMLAWB( res_Q6, buf_ptr[ 10 ], interpol_ptr[ 7 ] );
res_Q6 = silk_SMLAWB( res_Q6, buf_ptr[ 9 ], interpol_ptr[ 8 ] );
/* Scale down, saturate and store in output array */
*out++ = (opus_int16)silk_SAT16( silk_RSHIFT_ROUND( res_Q6, 6 ) );
}
break;
case RESAMPLER_DOWN_ORDER_FIR1:
for( index_Q16 = 0; index_Q16 < max_index_Q16; index_Q16 += index_increment_Q16 ) {
/* Integer part gives pointer to buffered input */
buf_ptr = buf + silk_RSHIFT( index_Q16, 16 );
/* Inner product */
res_Q6 = silk_SMULWB( silk_ADD32( buf_ptr[ 0 ], buf_ptr[ 23 ] ), FIR_Coefs[ 0 ] );
res_Q6 = silk_SMLAWB( res_Q6, silk_ADD32( buf_ptr[ 1 ], buf_ptr[ 22 ] ), FIR_Coefs[ 1 ] );
res_Q6 = silk_SMLAWB( res_Q6, silk_ADD32( buf_ptr[ 2 ], buf_ptr[ 21 ] ), FIR_Coefs[ 2 ] );
res_Q6 = silk_SMLAWB( res_Q6, silk_ADD32( buf_ptr[ 3 ], buf_ptr[ 20 ] ), FIR_Coefs[ 3 ] );
res_Q6 = silk_SMLAWB( res_Q6, silk_ADD32( buf_ptr[ 4 ], buf_ptr[ 19 ] ), FIR_Coefs[ 4 ] );
res_Q6 = silk_SMLAWB( res_Q6, silk_ADD32( buf_ptr[ 5 ], buf_ptr[ 18 ] ), FIR_Coefs[ 5 ] );
res_Q6 = silk_SMLAWB( res_Q6, silk_ADD32( buf_ptr[ 6 ], buf_ptr[ 17 ] ), FIR_Coefs[ 6 ] );
res_Q6 = silk_SMLAWB( res_Q6, silk_ADD32( buf_ptr[ 7 ], buf_ptr[ 16 ] ), FIR_Coefs[ 7 ] );
res_Q6 = silk_SMLAWB( res_Q6, silk_ADD32( buf_ptr[ 8 ], buf_ptr[ 15 ] ), FIR_Coefs[ 8 ] );
res_Q6 = silk_SMLAWB( res_Q6, silk_ADD32( buf_ptr[ 9 ], buf_ptr[ 14 ] ), FIR_Coefs[ 9 ] );
res_Q6 = silk_SMLAWB( res_Q6, silk_ADD32( buf_ptr[ 10 ], buf_ptr[ 13 ] ), FIR_Coefs[ 10 ] );
res_Q6 = silk_SMLAWB( res_Q6, silk_ADD32( buf_ptr[ 11 ], buf_ptr[ 12 ] ), FIR_Coefs[ 11 ] );
/* Scale down, saturate and store in output array */
*out++ = (opus_int16)silk_SAT16( silk_RSHIFT_ROUND( res_Q6, 6 ) );
}
break;
case RESAMPLER_DOWN_ORDER_FIR2:
for( index_Q16 = 0; index_Q16 < max_index_Q16; index_Q16 += index_increment_Q16 ) {
/* Integer part gives pointer to buffered input */
buf_ptr = buf + silk_RSHIFT( index_Q16, 16 );
/* Inner product */
res_Q6 = silk_SMULWB( silk_ADD32( buf_ptr[ 0 ], buf_ptr[ 35 ] ), FIR_Coefs[ 0 ] );
res_Q6 = silk_SMLAWB( res_Q6, silk_ADD32( buf_ptr[ 1 ], buf_ptr[ 34 ] ), FIR_Coefs[ 1 ] );
res_Q6 = silk_SMLAWB( res_Q6, silk_ADD32( buf_ptr[ 2 ], buf_ptr[ 33 ] ), FIR_Coefs[ 2 ] );
res_Q6 = silk_SMLAWB( res_Q6, silk_ADD32( buf_ptr[ 3 ], buf_ptr[ 32 ] ), FIR_Coefs[ 3 ] );
res_Q6 = silk_SMLAWB( res_Q6, silk_ADD32( buf_ptr[ 4 ], buf_ptr[ 31 ] ), FIR_Coefs[ 4 ] );
res_Q6 = silk_SMLAWB( res_Q6, silk_ADD32( buf_ptr[ 5 ], buf_ptr[ 30 ] ), FIR_Coefs[ 5 ] );
res_Q6 = silk_SMLAWB( res_Q6, silk_ADD32( buf_ptr[ 6 ], buf_ptr[ 29 ] ), FIR_Coefs[ 6 ] );
res_Q6 = silk_SMLAWB( res_Q6, silk_ADD32( buf_ptr[ 7 ], buf_ptr[ 28 ] ), FIR_Coefs[ 7 ] );
res_Q6 = silk_SMLAWB( res_Q6, silk_ADD32( buf_ptr[ 8 ], buf_ptr[ 27 ] ), FIR_Coefs[ 8 ] );
res_Q6 = silk_SMLAWB( res_Q6, silk_ADD32( buf_ptr[ 9 ], buf_ptr[ 26 ] ), FIR_Coefs[ 9 ] );
res_Q6 = silk_SMLAWB( res_Q6, silk_ADD32( buf_ptr[ 10 ], buf_ptr[ 25 ] ), FIR_Coefs[ 10 ] );
res_Q6 = silk_SMLAWB( res_Q6, silk_ADD32( buf_ptr[ 11 ], buf_ptr[ 24 ] ), FIR_Coefs[ 11 ] );
res_Q6 = silk_SMLAWB( res_Q6, silk_ADD32( buf_ptr[ 12 ], buf_ptr[ 23 ] ), FIR_Coefs[ 12 ] );
res_Q6 = silk_SMLAWB( res_Q6, silk_ADD32( buf_ptr[ 13 ], buf_ptr[ 22 ] ), FIR_Coefs[ 13 ] );
res_Q6 = silk_SMLAWB( res_Q6, silk_ADD32( buf_ptr[ 14 ], buf_ptr[ 21 ] ), FIR_Coefs[ 14 ] );
res_Q6 = silk_SMLAWB( res_Q6, silk_ADD32( buf_ptr[ 15 ], buf_ptr[ 20 ] ), FIR_Coefs[ 15 ] );
res_Q6 = silk_SMLAWB( res_Q6, silk_ADD32( buf_ptr[ 16 ], buf_ptr[ 19 ] ), FIR_Coefs[ 16 ] );
res_Q6 = silk_SMLAWB( res_Q6, silk_ADD32( buf_ptr[ 17 ], buf_ptr[ 18 ] ), FIR_Coefs[ 17 ] );
/* Scale down, saturate and store in output array */
*out++ = (opus_int16)silk_SAT16( silk_RSHIFT_ROUND( res_Q6, 6 ) );
}
break;
default:
silk_assert( 0 );
}
return out;
}
/* amplitude of monic warped coefficients by using bandwidth expansion on the true coefficients */
static inline void limit_warped_coefs(
opus_int32 *coefs_syn_Q24,
opus_int32 *coefs_ana_Q24,
opus_int lambda_Q16,
opus_int32 limit_Q24,
opus_int order
) {
opus_int i, iter, ind = 0;
opus_int32 tmp, maxabs_Q24, chirp_Q16, gain_syn_Q16, gain_ana_Q16;
opus_int32 nom_Q16, den_Q24;
/* Convert to monic coefficients */
lambda_Q16 = -lambda_Q16;
for( i = order - 1; i > 0; i-- ) {
coefs_syn_Q24[ i - 1 ] = silk_SMLAWB( coefs_syn_Q24[ i - 1 ], coefs_syn_Q24[ i ], lambda_Q16 );
coefs_ana_Q24[ i - 1 ] = silk_SMLAWB( coefs_ana_Q24[ i - 1 ], coefs_ana_Q24[ i ], lambda_Q16 );
}
lambda_Q16 = -lambda_Q16;
nom_Q16 = silk_SMLAWB( SILK_FIX_CONST( 1.0, 16 ), -lambda_Q16, lambda_Q16 );
den_Q24 = silk_SMLAWB( SILK_FIX_CONST( 1.0, 24 ), coefs_syn_Q24[ 0 ], lambda_Q16 );
gain_syn_Q16 = silk_DIV32_varQ( nom_Q16, den_Q24, 24 );
den_Q24 = silk_SMLAWB( SILK_FIX_CONST( 1.0, 24 ), coefs_ana_Q24[ 0 ], lambda_Q16 );
gain_ana_Q16 = silk_DIV32_varQ( nom_Q16, den_Q24, 24 );
for( i = 0; i < order; i++ ) {
coefs_syn_Q24[ i ] = silk_SMULWW( gain_syn_Q16, coefs_syn_Q24[ i ] );
coefs_ana_Q24[ i ] = silk_SMULWW( gain_ana_Q16, coefs_ana_Q24[ i ] );
}
for( iter = 0; iter < 10; iter++ ) {
/* Find maximum absolute value */
maxabs_Q24 = -1;
for( i = 0; i < order; i++ ) {
tmp = silk_max( silk_abs_int32( coefs_syn_Q24[ i ] ), silk_abs_int32( coefs_ana_Q24[ i ] ) );
if( tmp > maxabs_Q24 ) {
maxabs_Q24 = tmp;
ind = i;
}
}
if( maxabs_Q24 <= limit_Q24 ) {
/* Coefficients are within range - done */
return;
}
/* Convert back to true warped coefficients */
for( i = 1; i < order; i++ ) {
coefs_syn_Q24[ i - 1 ] = silk_SMLAWB( coefs_syn_Q24[ i - 1 ], coefs_syn_Q24[ i ], lambda_Q16 );
coefs_ana_Q24[ i - 1 ] = silk_SMLAWB( coefs_ana_Q24[ i - 1 ], coefs_ana_Q24[ i ], lambda_Q16 );
}
gain_syn_Q16 = silk_INVERSE32_varQ( gain_syn_Q16, 32 );
gain_ana_Q16 = silk_INVERSE32_varQ( gain_ana_Q16, 32 );
for( i = 0; i < order; i++ ) {
coefs_syn_Q24[ i ] = silk_SMULWW( gain_syn_Q16, coefs_syn_Q24[ i ] );
coefs_ana_Q24[ i ] = silk_SMULWW( gain_ana_Q16, coefs_ana_Q24[ i ] );
}
/* Apply bandwidth expansion */
chirp_Q16 = SILK_FIX_CONST( 0.99, 16 ) - silk_DIV32_varQ(
silk_SMULWB( maxabs_Q24 - limit_Q24, silk_SMLABB( SILK_FIX_CONST( 0.8, 10 ), SILK_FIX_CONST( 0.1, 10 ), iter ) ),
silk_MUL( maxabs_Q24, ind + 1 ), 22 );
silk_bwexpander_32( coefs_syn_Q24, order, chirp_Q16 );
silk_bwexpander_32( coefs_ana_Q24, order, chirp_Q16 );
/* Convert to monic warped coefficients */
lambda_Q16 = -lambda_Q16;
for( i = order - 1; i > 0; i-- ) {
coefs_syn_Q24[ i - 1 ] = silk_SMLAWB( coefs_syn_Q24[ i - 1 ], coefs_syn_Q24[ i ], lambda_Q16 );
coefs_ana_Q24[ i - 1 ] = silk_SMLAWB( coefs_ana_Q24[ i - 1 ], coefs_ana_Q24[ i ], lambda_Q16 );
}
lambda_Q16 = -lambda_Q16;
nom_Q16 = silk_SMLAWB( SILK_FIX_CONST( 1.0, 16 ), -lambda_Q16, lambda_Q16 );
den_Q24 = silk_SMLAWB( SILK_FIX_CONST( 1.0, 24 ), coefs_syn_Q24[ 0 ], lambda_Q16 );
gain_syn_Q16 = silk_DIV32_varQ( nom_Q16, den_Q24, 24 );
den_Q24 = silk_SMLAWB( SILK_FIX_CONST( 1.0, 24 ), coefs_ana_Q24[ 0 ], lambda_Q16 );
gain_ana_Q16 = silk_DIV32_varQ( nom_Q16, den_Q24, 24 );
for( i = 0; i < order; i++ ) {
coefs_syn_Q24[ i ] = silk_SMULWW( gain_syn_Q16, coefs_syn_Q24[ i ] );
coefs_ana_Q24[ i ] = silk_SMULWW( gain_ana_Q16, coefs_ana_Q24[ i ] );
}
}
silk_assert( 0 );
}
static inline void silk_PLC_conceal(
silk_decoder_state *psDec, /* I/O Decoder state */
silk_decoder_control *psDecCtrl, /* I/O Decoder control */
opus_int16 frame[] /* O LPC residual signal */
)
{
opus_int i, j, k;
opus_int lag, idx, sLTP_buf_idx, shift1, shift2;
opus_int32 rand_seed, harm_Gain_Q15, rand_Gain_Q15, inv_gain_Q16, inv_gain_Q30;
opus_int32 energy1, energy2, *rand_ptr, *pred_lag_ptr;
opus_int32 LPC_exc_Q14, LPC_pred_Q10, LTP_pred_Q12;
opus_int16 rand_scale_Q14;
opus_int16 *B_Q14, *exc_buf_ptr;
opus_int32 *sLPC_Q14_ptr;
opus_int16 exc_buf[ 2 * MAX_SUB_FRAME_LENGTH ];
opus_int16 A_Q12[ MAX_LPC_ORDER ];
opus_int16 sLTP[ MAX_FRAME_LENGTH ];
opus_int32 sLTP_Q14[ 2 * MAX_FRAME_LENGTH ];
silk_PLC_struct *psPLC = &psDec->sPLC;
if (psDec->first_frame_after_reset)
silk_memset(psPLC->prevLPC_Q12, 0, MAX_LPC_ORDER*sizeof(psPLC->prevLPC_Q12[ 0 ]));
/* Find random noise component */
/* Scale previous excitation signal */
exc_buf_ptr = exc_buf;
for( k = 0; k < 2; k++ ) {
for( i = 0; i < psPLC->subfr_length; i++ ) {
exc_buf_ptr[ i ] = ( opus_int16 )silk_RSHIFT(
silk_SMULWW( psDec->exc_Q10[ i + ( k + psPLC->nb_subfr - 2 ) * psPLC->subfr_length ], psPLC->prevGain_Q16[ k ] ), 10 );
}
exc_buf_ptr += psPLC->subfr_length;
}
/* Find the subframe with lowest energy of the last two and use that as random noise generator */
silk_sum_sqr_shift( &energy1, &shift1, exc_buf, psPLC->subfr_length );
silk_sum_sqr_shift( &energy2, &shift2, &exc_buf[ psPLC->subfr_length ], psPLC->subfr_length );
if( silk_RSHIFT( energy1, shift2 ) < silk_RSHIFT( energy2, shift1 ) ) {
/* First sub-frame has lowest energy */
rand_ptr = &psDec->exc_Q10[ silk_max_int( 0, ( psPLC->nb_subfr - 1 ) * psPLC->subfr_length - RAND_BUF_SIZE ) ];
} else {
/* Second sub-frame has lowest energy */
rand_ptr = &psDec->exc_Q10[ silk_max_int( 0, psPLC->nb_subfr * psPLC->subfr_length - RAND_BUF_SIZE ) ];
}
/* Setup Gain to random noise component */
B_Q14 = psPLC->LTPCoef_Q14;
rand_scale_Q14 = psPLC->randScale_Q14;
/* Setup attenuation gains */
harm_Gain_Q15 = HARM_ATT_Q15[ silk_min_int( NB_ATT - 1, psDec->lossCnt ) ];
if( psDec->prevSignalType == TYPE_VOICED ) {
rand_Gain_Q15 = PLC_RAND_ATTENUATE_V_Q15[ silk_min_int( NB_ATT - 1, psDec->lossCnt ) ];
} else {
rand_Gain_Q15 = PLC_RAND_ATTENUATE_UV_Q15[ silk_min_int( NB_ATT - 1, psDec->lossCnt ) ];
}
/* LPC concealment. Apply BWE to previous LPC */
silk_bwexpander( psPLC->prevLPC_Q12, psDec->LPC_order, SILK_FIX_CONST( BWE_COEF, 16 ) );
/* Preload LPC coeficients to array on stack. Gives small performance gain */
silk_memcpy( A_Q12, psPLC->prevLPC_Q12, psDec->LPC_order * sizeof( opus_int16 ) );
/* First Lost frame */
if( psDec->lossCnt == 0 ) {
rand_scale_Q14 = 1 << 14;
/* Reduce random noise Gain for voiced frames */
if( psDec->prevSignalType == TYPE_VOICED ) {
for( i = 0; i < LTP_ORDER; i++ ) {
rand_scale_Q14 -= B_Q14[ i ];
}
rand_scale_Q14 = silk_max_16( 3277, rand_scale_Q14 ); /* 0.2 */
rand_scale_Q14 = ( opus_int16 )silk_RSHIFT( silk_SMULBB( rand_scale_Q14, psPLC->prevLTP_scale_Q14 ), 14 );
} else {
/* Reduce random noise for unvoiced frames with high LPC gain */
opus_int32 invGain_Q30, down_scale_Q30;
silk_LPC_inverse_pred_gain( &invGain_Q30, psPLC->prevLPC_Q12, psDec->LPC_order );
down_scale_Q30 = silk_min_32( silk_RSHIFT( 1 << 30, LOG2_INV_LPC_GAIN_HIGH_THRES ), invGain_Q30 );
down_scale_Q30 = silk_max_32( silk_RSHIFT( 1 << 30, LOG2_INV_LPC_GAIN_LOW_THRES ), down_scale_Q30 );
down_scale_Q30 = silk_LSHIFT( down_scale_Q30, LOG2_INV_LPC_GAIN_HIGH_THRES );
rand_Gain_Q15 = silk_RSHIFT( silk_SMULWB( down_scale_Q30, rand_Gain_Q15 ), 14 );
}
}
rand_seed = psPLC->rand_seed;
lag = silk_RSHIFT_ROUND( psPLC->pitchL_Q8, 8 );
sLTP_buf_idx = psDec->ltp_mem_length;
/* Rewhiten LTP state */
idx = psDec->ltp_mem_length - lag - psDec->LPC_order - LTP_ORDER / 2;
silk_assert( idx > 0 );
silk_LPC_analysis_filter( &sLTP[ idx ], &psDec->outBuf[ idx ], A_Q12, psDec->ltp_mem_length - idx, psDec->LPC_order );
/* Scale LTP state */
inv_gain_Q16 = silk_INVERSE32_varQ( psPLC->prevGain_Q16[ 1 ], 32 );
inv_gain_Q16 = silk_min( inv_gain_Q16, silk_int16_MAX );
inv_gain_Q30 = silk_LSHIFT( inv_gain_Q16, 14 );
for( i = idx + psDec->LPC_order; i < psDec->ltp_mem_length; i++ ) {
sLTP_Q14[ i ] = silk_SMULWB( inv_gain_Q30, sLTP[ i ] );
}
/***************************/
/* LTP synthesis filtering */
/***************************/
for( k = 0; k < psDec->nb_subfr; k++ ) {
/* Setup pointer */
pred_lag_ptr = &sLTP_Q14[ sLTP_buf_idx - lag + LTP_ORDER / 2 ];
for( i = 0; i < psDec->subfr_length; i++ ) {
/* Unrolled loop */
LTP_pred_Q12 = silk_SMULWB( pred_lag_ptr[ 0 ], B_Q14[ 0 ] );
LTP_pred_Q12 = silk_SMLAWB( LTP_pred_Q12, pred_lag_ptr[ -1 ], B_Q14[ 1 ] );
LTP_pred_Q12 = silk_SMLAWB( LTP_pred_Q12, pred_lag_ptr[ -2 ], B_Q14[ 2 ] );
LTP_pred_Q12 = silk_SMLAWB( LTP_pred_Q12, pred_lag_ptr[ -3 ], B_Q14[ 3 ] );
LTP_pred_Q12 = silk_SMLAWB( LTP_pred_Q12, pred_lag_ptr[ -4 ], B_Q14[ 4 ] );
pred_lag_ptr++;
/* Generate LPC excitation */
rand_seed = silk_RAND( rand_seed );
idx = silk_RSHIFT( rand_seed, 25 ) & RAND_BUF_MASK;
LPC_exc_Q14 = silk_LSHIFT32( silk_SMULWB( rand_ptr[ idx ], rand_scale_Q14 ), 6 ); /* Random noise part */
LPC_exc_Q14 = silk_ADD32( LPC_exc_Q14, silk_LSHIFT32( LTP_pred_Q12, 2 ) ); /* Harmonic part */
sLTP_Q14[ sLTP_buf_idx ] = LPC_exc_Q14;
sLTP_buf_idx++;
}
/* Gradually reduce LTP gain */
for( j = 0; j < LTP_ORDER; j++ ) {
B_Q14[ j ] = silk_RSHIFT( silk_SMULBB( harm_Gain_Q15, B_Q14[ j ] ), 15 );
}
/* Gradually reduce excitation gain */
rand_scale_Q14 = silk_RSHIFT( silk_SMULBB( rand_scale_Q14, rand_Gain_Q15 ), 15 );
/* Slowly increase pitch lag */
psPLC->pitchL_Q8 = silk_SMLAWB( psPLC->pitchL_Q8, psPLC->pitchL_Q8, PITCH_DRIFT_FAC_Q16 );
psPLC->pitchL_Q8 = silk_min_32( psPLC->pitchL_Q8, silk_LSHIFT( silk_SMULBB( MAX_PITCH_LAG_MS, psDec->fs_kHz ), 8 ) );
lag = silk_RSHIFT_ROUND( psPLC->pitchL_Q8, 8 );
}
/***************************/
/* LPC synthesis filtering */
/***************************/
sLPC_Q14_ptr = &sLTP_Q14[ psDec->ltp_mem_length - MAX_LPC_ORDER ];
/* Copy LPC state */
silk_memcpy( sLPC_Q14_ptr, psDec->sLPC_Q14_buf, MAX_LPC_ORDER * sizeof( opus_int32 ) );
silk_assert( psDec->LPC_order >= 10 ); /* check that unrolling works */
for( i = 0; i < psDec->frame_length; i++ ) {
/* partly unrolled */
LPC_pred_Q10 = silk_SMULWB( sLPC_Q14_ptr[ MAX_LPC_ORDER + i - 1 ], A_Q12[ 0 ] );
LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14_ptr[ MAX_LPC_ORDER + i - 2 ], A_Q12[ 1 ] );
LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14_ptr[ MAX_LPC_ORDER + i - 3 ], A_Q12[ 2 ] );
LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14_ptr[ MAX_LPC_ORDER + i - 4 ], A_Q12[ 3 ] );
LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14_ptr[ MAX_LPC_ORDER + i - 5 ], A_Q12[ 4 ] );
LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14_ptr[ MAX_LPC_ORDER + i - 6 ], A_Q12[ 5 ] );
LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14_ptr[ MAX_LPC_ORDER + i - 7 ], A_Q12[ 6 ] );
LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14_ptr[ MAX_LPC_ORDER + i - 8 ], A_Q12[ 7 ] );
LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14_ptr[ MAX_LPC_ORDER + i - 9 ], A_Q12[ 8 ] );
LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14_ptr[ MAX_LPC_ORDER + i - 10 ], A_Q12[ 9 ] );
for( j = 10; j < psDec->LPC_order; j++ ) {
LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14_ptr[ MAX_LPC_ORDER + i - j - 1 ], A_Q12[ j ] );
}
/* Add prediction to LPC excitation */
sLPC_Q14_ptr[ MAX_LPC_ORDER + i ] = silk_ADD_LSHIFT32( sLPC_Q14_ptr[ MAX_LPC_ORDER + i ], LPC_pred_Q10, 4 );
/* Scale with Gain */
frame[ i ] = ( opus_int16 )silk_SAT16( silk_RSHIFT_ROUND( silk_SMULWW( sLPC_Q14_ptr[ MAX_LPC_ORDER + i ], psPLC->prevGain_Q16[ 1 ] ), 14 ) );
}
/* Save LPC state */
silk_memcpy( psDec->sLPC_Q14_buf, &sLPC_Q14_ptr[ psDec->frame_length ], MAX_LPC_ORDER * sizeof( opus_int32 ) );
/**************************************/
/* Update states */
/**************************************/
psPLC->rand_seed = rand_seed;
psPLC->randScale_Q14 = rand_scale_Q14;
for( i = 0; i < MAX_NB_SUBFR; i++ ) {
psDecCtrl->pitchL[ i ] = lag;
}
}
/* 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;
}
}
void silk_find_pred_coefs_FIX(
silk_encoder_state_FIX *psEnc, /* I/O encoder state */
silk_encoder_control_FIX *psEncCtrl, /* I/O encoder control */
const opus_int16 res_pitch[], /* I Residual from pitch analysis */
const opus_int16 x[], /* I Speech signal */
opus_int condCoding /* I The type of conditional coding to use */
)
{
opus_int i;
opus_int32 invGains_Q16[ MAX_NB_SUBFR ], local_gains[ MAX_NB_SUBFR ], Wght_Q15[ MAX_NB_SUBFR ];
opus_int16 NLSF_Q15[ MAX_LPC_ORDER ];
const opus_int16 *x_ptr;
opus_int16 *x_pre_ptr;
VARDECL( opus_int16, LPC_in_pre );
opus_int32 tmp, min_gain_Q16, minInvGain_Q30;
opus_int LTP_corrs_rshift[ MAX_NB_SUBFR ];
SAVE_STACK;
/* weighting for weighted least squares */
min_gain_Q16 = silk_int32_MAX >> 6;
for( i = 0; i < psEnc->sCmn.nb_subfr; i++ ) {
min_gain_Q16 = silk_min( min_gain_Q16, psEncCtrl->Gains_Q16[ i ] );
}
for( i = 0; i < psEnc->sCmn.nb_subfr; i++ ) {
/* Divide to Q16 */
silk_assert( psEncCtrl->Gains_Q16[ i ] > 0 );
/* Invert and normalize gains, and ensure that maximum invGains_Q16 is within range of a 16 bit int */
invGains_Q16[ i ] = silk_DIV32_varQ( min_gain_Q16, psEncCtrl->Gains_Q16[ i ], 16 - 2 );
/* Ensure Wght_Q15 a minimum value 1 */
invGains_Q16[ i ] = silk_max( invGains_Q16[ i ], 363 );
/* Square the inverted gains */
silk_assert( invGains_Q16[ i ] == silk_SAT16( invGains_Q16[ i ] ) );
tmp = silk_SMULWB( invGains_Q16[ i ], invGains_Q16[ i ] );
Wght_Q15[ i ] = silk_RSHIFT( tmp, 1 );
/* Invert the inverted and normalized gains */
local_gains[ i ] = silk_DIV32( ( (opus_int32)1 << 16 ), invGains_Q16[ i ] );
}
ALLOC( LPC_in_pre,
psEnc->sCmn.nb_subfr * psEnc->sCmn.predictLPCOrder
+ psEnc->sCmn.frame_length, opus_int16 );
if( psEnc->sCmn.indices.signalType == TYPE_VOICED ) {
VARDECL( opus_int32, WLTP );
/**********/
/* VOICED */
/**********/
silk_assert( psEnc->sCmn.ltp_mem_length - psEnc->sCmn.predictLPCOrder >= psEncCtrl->pitchL[ 0 ] + LTP_ORDER / 2 );
ALLOC( WLTP, psEnc->sCmn.nb_subfr * LTP_ORDER * LTP_ORDER, opus_int32 );
/* LTP analysis */
silk_find_LTP_FIX( psEncCtrl->LTPCoef_Q14, WLTP, &psEncCtrl->LTPredCodGain_Q7,
res_pitch, psEncCtrl->pitchL, Wght_Q15, psEnc->sCmn.subfr_length,
psEnc->sCmn.nb_subfr, psEnc->sCmn.ltp_mem_length, LTP_corrs_rshift );
/* Quantize LTP gain parameters */
silk_quant_LTP_gains( psEncCtrl->LTPCoef_Q14, psEnc->sCmn.indices.LTPIndex, &psEnc->sCmn.indices.PERIndex,
&psEnc->sCmn.sum_log_gain_Q7, WLTP, psEnc->sCmn.mu_LTP_Q9, psEnc->sCmn.LTPQuantLowComplexity, psEnc->sCmn.nb_subfr);
/* Control LTP scaling */
silk_LTP_scale_ctrl_FIX( psEnc, psEncCtrl, condCoding );
/* Create LTP residual */
silk_LTP_analysis_filter_FIX( LPC_in_pre, x - psEnc->sCmn.predictLPCOrder, psEncCtrl->LTPCoef_Q14,
psEncCtrl->pitchL, invGains_Q16, psEnc->sCmn.subfr_length, psEnc->sCmn.nb_subfr, psEnc->sCmn.predictLPCOrder );
} else {
/************/
/* UNVOICED */
/************/
/* Create signal with prepended subframes, scaled by inverse gains */
x_ptr = x - psEnc->sCmn.predictLPCOrder;
x_pre_ptr = LPC_in_pre;
for( i = 0; i < psEnc->sCmn.nb_subfr; i++ ) {
silk_scale_copy_vector16( x_pre_ptr, x_ptr, invGains_Q16[ i ],
psEnc->sCmn.subfr_length + psEnc->sCmn.predictLPCOrder );
x_pre_ptr += psEnc->sCmn.subfr_length + psEnc->sCmn.predictLPCOrder;
x_ptr += psEnc->sCmn.subfr_length;
}
silk_memset( psEncCtrl->LTPCoef_Q14, 0, psEnc->sCmn.nb_subfr * LTP_ORDER * sizeof( opus_int16 ) );
psEncCtrl->LTPredCodGain_Q7 = 0;
psEnc->sCmn.sum_log_gain_Q7 = 0;
}
/* Limit on total predictive coding gain */
if( psEnc->sCmn.first_frame_after_reset ) {
minInvGain_Q30 = SILK_FIX_CONST( 1.0f / MAX_PREDICTION_POWER_GAIN_AFTER_RESET, 30 );
} else {
minInvGain_Q30 = silk_log2lin( silk_SMLAWB( 16 << 7, (opus_int32)psEncCtrl->LTPredCodGain_Q7, SILK_FIX_CONST( 1.0 / 3, 16 ) ) ); /* Q16 */
minInvGain_Q30 = silk_DIV32_varQ( minInvGain_Q30,
silk_SMULWW( SILK_FIX_CONST( MAX_PREDICTION_POWER_GAIN, 0 ),
silk_SMLAWB( SILK_FIX_CONST( 0.25, 18 ), SILK_FIX_CONST( 0.75, 18 ), psEncCtrl->coding_quality_Q14 ) ), 14 );
}
/* LPC_in_pre contains the LTP-filtered input for voiced, and the unfiltered input for unvoiced */
silk_find_LPC_FIX( &psEnc->sCmn, NLSF_Q15, LPC_in_pre, minInvGain_Q30 );
/* Quantize LSFs */
silk_process_NLSFs( &psEnc->sCmn, psEncCtrl->PredCoef_Q12, NLSF_Q15, psEnc->sCmn.prev_NLSFq_Q15 );
/* Calculate residual energy using quantized LPC coefficients */
silk_residual_energy_FIX( psEncCtrl->ResNrg, psEncCtrl->ResNrgQ, LPC_in_pre, psEncCtrl->PredCoef_Q12, local_gains,
psEnc->sCmn.subfr_length, psEnc->sCmn.nb_subfr, psEnc->sCmn.predictLPCOrder );
/* Copy to prediction struct for use in next frame for interpolation */
silk_memcpy( psEnc->sCmn.prev_NLSFq_Q15, NLSF_Q15, sizeof( psEnc->sCmn.prev_NLSFq_Q15 ) );
RESTORE_STACK;
}
/* 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;
}
/* Entropy constrained matrix-weighted VQ, hard-coded to 5-element vectors, for a single input data vector */
void silk_VQ_WMat_EC_c(
opus_int8 *ind, /* O index of best codebook vector */
opus_int32 *res_nrg_Q15, /* O best residual energy */
opus_int32 *rate_dist_Q8, /* O best total bitrate */
opus_int *gain_Q7, /* O sum of absolute LTP coefficients */
const opus_int32 *XX_Q17, /* I correlation matrix */
const opus_int32 *xX_Q17, /* I correlation vector */
const opus_int8 *cb_Q7, /* I codebook */
const opus_uint8 *cb_gain_Q7, /* I codebook effective gain */
const opus_uint8 *cl_Q5, /* I code length for each codebook vector */
const opus_int subfr_len, /* I number of samples per subframe */
const opus_int32 max_gain_Q7, /* I maximum sum of absolute LTP coefficients */
const opus_int L /* I number of vectors in codebook */
)
{
opus_int k, gain_tmp_Q7;
const opus_int8 *cb_row_Q7;
opus_int32 neg_xX_Q24[ 5 ];
opus_int32 sum1_Q15, sum2_Q24;
opus_int32 bits_res_Q8, bits_tot_Q8;
/* Negate and convert to new Q domain */
neg_xX_Q24[ 0 ] = -silk_LSHIFT32( xX_Q17[ 0 ], 7 );
neg_xX_Q24[ 1 ] = -silk_LSHIFT32( xX_Q17[ 1 ], 7 );
neg_xX_Q24[ 2 ] = -silk_LSHIFT32( xX_Q17[ 2 ], 7 );
neg_xX_Q24[ 3 ] = -silk_LSHIFT32( xX_Q17[ 3 ], 7 );
neg_xX_Q24[ 4 ] = -silk_LSHIFT32( xX_Q17[ 4 ], 7 );
/* Loop over codebook */
*rate_dist_Q8 = silk_int32_MAX;
*res_nrg_Q15 = silk_int32_MAX;
cb_row_Q7 = cb_Q7;
/* In things go really bad, at least *ind is set to something safe. */
*ind = 0;
for( k = 0; k < L; k++ ) {
opus_int32 penalty;
gain_tmp_Q7 = cb_gain_Q7[k];
/* Weighted rate */
/* Quantization error: 1 - 2 * xX * cb + cb' * XX * cb */
sum1_Q15 = SILK_FIX_CONST( 1.001, 15 );
/* Penalty for too large gain */
penalty = silk_LSHIFT32( silk_max( silk_SUB32( gain_tmp_Q7, max_gain_Q7 ), 0 ), 11 );
/* first row of XX_Q17 */
sum2_Q24 = silk_MLA( neg_xX_Q24[ 0 ], XX_Q17[ 1 ], cb_row_Q7[ 1 ] );
sum2_Q24 = silk_MLA( sum2_Q24, XX_Q17[ 2 ], cb_row_Q7[ 2 ] );
sum2_Q24 = silk_MLA( sum2_Q24, XX_Q17[ 3 ], cb_row_Q7[ 3 ] );
sum2_Q24 = silk_MLA( sum2_Q24, XX_Q17[ 4 ], cb_row_Q7[ 4 ] );
sum2_Q24 = silk_LSHIFT32( sum2_Q24, 1 );
sum2_Q24 = silk_MLA( sum2_Q24, XX_Q17[ 0 ], cb_row_Q7[ 0 ] );
sum1_Q15 = silk_SMLAWB( sum1_Q15, sum2_Q24, cb_row_Q7[ 0 ] );
/* second row of XX_Q17 */
sum2_Q24 = silk_MLA( neg_xX_Q24[ 1 ], XX_Q17[ 7 ], cb_row_Q7[ 2 ] );
sum2_Q24 = silk_MLA( sum2_Q24, XX_Q17[ 8 ], cb_row_Q7[ 3 ] );
sum2_Q24 = silk_MLA( sum2_Q24, XX_Q17[ 9 ], cb_row_Q7[ 4 ] );
sum2_Q24 = silk_LSHIFT32( sum2_Q24, 1 );
sum2_Q24 = silk_MLA( sum2_Q24, XX_Q17[ 6 ], cb_row_Q7[ 1 ] );
sum1_Q15 = silk_SMLAWB( sum1_Q15, sum2_Q24, cb_row_Q7[ 1 ] );
/* third row of XX_Q17 */
sum2_Q24 = silk_MLA( neg_xX_Q24[ 2 ], XX_Q17[ 13 ], cb_row_Q7[ 3 ] );
sum2_Q24 = silk_MLA( sum2_Q24, XX_Q17[ 14 ], cb_row_Q7[ 4 ] );
sum2_Q24 = silk_LSHIFT32( sum2_Q24, 1 );
sum2_Q24 = silk_MLA( sum2_Q24, XX_Q17[ 12 ], cb_row_Q7[ 2 ] );
sum1_Q15 = silk_SMLAWB( sum1_Q15, sum2_Q24, cb_row_Q7[ 2 ] );
/* fourth row of XX_Q17 */
sum2_Q24 = silk_MLA( neg_xX_Q24[ 3 ], XX_Q17[ 19 ], cb_row_Q7[ 4 ] );
sum2_Q24 = silk_LSHIFT32( sum2_Q24, 1 );
sum2_Q24 = silk_MLA( sum2_Q24, XX_Q17[ 18 ], cb_row_Q7[ 3 ] );
sum1_Q15 = silk_SMLAWB( sum1_Q15, sum2_Q24, cb_row_Q7[ 3 ] );
/* last row of XX_Q17 */
sum2_Q24 = silk_LSHIFT32( neg_xX_Q24[ 4 ], 1 );
sum2_Q24 = silk_MLA( sum2_Q24, XX_Q17[ 24 ], cb_row_Q7[ 4 ] );
sum1_Q15 = silk_SMLAWB( sum1_Q15, sum2_Q24, cb_row_Q7[ 4 ] );
/* find best */
if( sum1_Q15 >= 0 ) {
/* Translate residual energy to bits using high-rate assumption (6 dB ==> 1 bit/sample) */
bits_res_Q8 = silk_SMULBB( subfr_len, silk_lin2log( sum1_Q15 + penalty) - (15 << 7) );
/* In the following line we reduce the codelength component by half ("-1"); seems to slghtly improve quality */
bits_tot_Q8 = silk_ADD_LSHIFT32( bits_res_Q8, cl_Q5[ k ], 3-1 );
if( bits_tot_Q8 <= *rate_dist_Q8 ) {
*rate_dist_Q8 = bits_tot_Q8;
*res_nrg_Q15 = sum1_Q15 + penalty;
*ind = (opus_int8)k;
*gain_Q7 = gain_tmp_Q7;
}
}
/* Go to next cbk vector */
cb_row_Q7 += LTP_ORDER;
}
}
void silk_find_LTP_FIX(
opus_int16 b_Q14[ MAX_NB_SUBFR * LTP_ORDER ], /* O LTP coefs */
opus_int32 WLTP[ MAX_NB_SUBFR * LTP_ORDER * LTP_ORDER ], /* O Weight for LTP quantization */
opus_int *LTPredCodGain_Q7, /* O LTP coding gain */
const opus_int16 r_lpc[], /* I residual signal after LPC signal + state for first 10 ms */
const opus_int lag[ MAX_NB_SUBFR ], /* I LTP lags */
const opus_int32 Wght_Q15[ MAX_NB_SUBFR ], /* I weights */
const opus_int subfr_length, /* I subframe length */
const opus_int nb_subfr, /* I number of subframes */
const opus_int mem_offset, /* I number of samples in LTP memory */
opus_int corr_rshifts[ MAX_NB_SUBFR ] /* O right shifts applied to correlations */
)
{
opus_int i, k, lshift;
const opus_int16 *r_ptr, *lag_ptr;
opus_int16 *b_Q14_ptr;
opus_int32 regu;
opus_int32 *WLTP_ptr;
opus_int32 b_Q16[ LTP_ORDER ], delta_b_Q14[ LTP_ORDER ], d_Q14[ MAX_NB_SUBFR ], nrg[ MAX_NB_SUBFR ], g_Q26;
opus_int32 w[ MAX_NB_SUBFR ], WLTP_max, max_abs_d_Q14, max_w_bits;
opus_int32 temp32, denom32;
opus_int extra_shifts;
opus_int rr_shifts, maxRshifts, maxRshifts_wxtra, LZs;
opus_int32 LPC_res_nrg, LPC_LTP_res_nrg, div_Q16;
opus_int32 Rr[ LTP_ORDER ], rr[ MAX_NB_SUBFR ];
opus_int32 wd, m_Q12;
b_Q14_ptr = b_Q14;
WLTP_ptr = WLTP;
r_ptr = &r_lpc[ mem_offset ];
for( k = 0; k < nb_subfr; k++ ) {
lag_ptr = r_ptr - ( lag[ k ] + LTP_ORDER / 2 );
silk_sum_sqr_shift( &rr[ k ], &rr_shifts, r_ptr, subfr_length ); /* rr[ k ] in Q( -rr_shifts ) */
/* Assure headroom */
LZs = silk_CLZ32( rr[k] );
if( LZs < LTP_CORRS_HEAD_ROOM ) {
rr[ k ] = silk_RSHIFT_ROUND( rr[ k ], LTP_CORRS_HEAD_ROOM - LZs );
rr_shifts += ( LTP_CORRS_HEAD_ROOM - LZs );
}
corr_rshifts[ k ] = rr_shifts;
silk_corrMatrix_FIX( lag_ptr, subfr_length, LTP_ORDER, LTP_CORRS_HEAD_ROOM, WLTP_ptr, &corr_rshifts[ k ] ); /* WLTP_fix_ptr in Q( -corr_rshifts[ k ] ) */
/* The correlation vector always has lower max abs value than rr and/or RR so head room is assured */
silk_corrVector_FIX( lag_ptr, r_ptr, subfr_length, LTP_ORDER, Rr, corr_rshifts[ k ] ); /* Rr_fix_ptr in Q( -corr_rshifts[ k ] ) */
if( corr_rshifts[ k ] > rr_shifts ) {
rr[ k ] = silk_RSHIFT( rr[ k ], corr_rshifts[ k ] - rr_shifts ); /* rr[ k ] in Q( -corr_rshifts[ k ] ) */
}
silk_assert( rr[ k ] >= 0 );
regu = 1;
regu = silk_SMLAWB( regu, rr[ k ], SILK_FIX_CONST( LTP_DAMPING/3, 16 ) );
regu = silk_SMLAWB( regu, matrix_ptr( WLTP_ptr, 0, 0, LTP_ORDER ), SILK_FIX_CONST( LTP_DAMPING/3, 16 ) );
regu = silk_SMLAWB( regu, matrix_ptr( WLTP_ptr, LTP_ORDER-1, LTP_ORDER-1, LTP_ORDER ), SILK_FIX_CONST( LTP_DAMPING/3, 16 ) );
silk_regularize_correlations_FIX( WLTP_ptr, &rr[k], regu, LTP_ORDER );
silk_solve_LDL_FIX( WLTP_ptr, LTP_ORDER, Rr, b_Q16 ); /* WLTP_fix_ptr and Rr_fix_ptr both in Q(-corr_rshifts[k]) */
/* Limit and store in Q14 */
silk_fit_LTP( b_Q16, b_Q14_ptr );
/* Calculate residual energy */
nrg[ k ] = silk_residual_energy16_covar_FIX( b_Q14_ptr, WLTP_ptr, Rr, rr[ k ], LTP_ORDER, 14 ); /* nrg_fix in Q( -corr_rshifts[ k ] ) */
/* temp = Wght[ k ] / ( nrg[ k ] * Wght[ k ] + 0.01f * subfr_length ); */
extra_shifts = silk_min_int( corr_rshifts[ k ], LTP_CORRS_HEAD_ROOM );
denom32 = silk_LSHIFT_SAT32( silk_SMULWB( nrg[ k ], Wght_Q15[ k ] ), 1 + extra_shifts ) + /* Q( -corr_rshifts[ k ] + extra_shifts ) */
silk_RSHIFT( silk_SMULWB( subfr_length, 655 ), corr_rshifts[ k ] - extra_shifts ); /* Q( -corr_rshifts[ k ] + extra_shifts ) */
denom32 = silk_max( denom32, 1 );
silk_assert( ((opus_int64)Wght_Q15[ k ] << 16 ) < silk_int32_MAX ); /* Wght always < 0.5 in Q0 */
temp32 = silk_DIV32( silk_LSHIFT( (opus_int32)Wght_Q15[ k ], 16 ), denom32 ); /* Q( 15 + 16 + corr_rshifts[k] - extra_shifts ) */
temp32 = silk_RSHIFT( temp32, 31 + corr_rshifts[ k ] - extra_shifts - 26 ); /* Q26 */
/* Limit temp such that the below scaling never wraps around */
WLTP_max = 0;
for( i = 0; i < LTP_ORDER * LTP_ORDER; i++ ) {
WLTP_max = silk_max( WLTP_ptr[ i ], WLTP_max );
}
lshift = silk_CLZ32( WLTP_max ) - 1 - 3; /* keep 3 bits free for vq_nearest_neighbor_fix */
silk_assert( 26 - 18 + lshift >= 0 );
if( 26 - 18 + lshift < 31 ) {
temp32 = silk_min_32( temp32, silk_LSHIFT( (opus_int32)1, 26 - 18 + lshift ) );
}
silk_scale_vector32_Q26_lshift_18( WLTP_ptr, temp32, LTP_ORDER * LTP_ORDER ); /* WLTP_ptr in Q( 18 - corr_rshifts[ k ] ) */
w[ k ] = matrix_ptr( WLTP_ptr, LTP_ORDER/2, LTP_ORDER/2, LTP_ORDER ); /* w in Q( 18 - corr_rshifts[ k ] ) */
silk_assert( w[k] >= 0 );
r_ptr += subfr_length;
b_Q14_ptr += LTP_ORDER;
WLTP_ptr += LTP_ORDER * LTP_ORDER;
}
maxRshifts = 0;
for( k = 0; k < nb_subfr; k++ ) {
maxRshifts = silk_max_int( corr_rshifts[ k ], maxRshifts );
}
/* Compute LTP coding gain */
if( LTPredCodGain_Q7 != NULL ) {
LPC_LTP_res_nrg = 0;
LPC_res_nrg = 0;
silk_assert( LTP_CORRS_HEAD_ROOM >= 2 ); /* Check that no overflow will happen when adding */
for( k = 0; k < nb_subfr; k++ ) {
LPC_res_nrg = silk_ADD32( LPC_res_nrg, silk_RSHIFT( silk_ADD32( silk_SMULWB( rr[ k ], Wght_Q15[ k ] ), 1 ), 1 + ( maxRshifts - corr_rshifts[ k ] ) ) ); /* Q( -maxRshifts ) */
LPC_LTP_res_nrg = silk_ADD32( LPC_LTP_res_nrg, silk_RSHIFT( silk_ADD32( silk_SMULWB( nrg[ k ], Wght_Q15[ k ] ), 1 ), 1 + ( maxRshifts - corr_rshifts[ k ] ) ) ); /* Q( -maxRshifts ) */
}
LPC_LTP_res_nrg = silk_max( LPC_LTP_res_nrg, 1 ); /* avoid division by zero */
div_Q16 = silk_DIV32_varQ( LPC_res_nrg, LPC_LTP_res_nrg, 16 );
*LTPredCodGain_Q7 = ( opus_int )silk_SMULBB( 3, silk_lin2log( div_Q16 ) - ( 16 << 7 ) );
silk_assert( *LTPredCodGain_Q7 == ( opus_int )silk_SAT16( silk_MUL( 3, silk_lin2log( div_Q16 ) - ( 16 << 7 ) ) ) );
}
/* smoothing */
/* d = sum( B, 1 ); */
b_Q14_ptr = b_Q14;
for( k = 0; k < nb_subfr; k++ ) {
d_Q14[ k ] = 0;
for( i = 0; i < LTP_ORDER; i++ ) {
d_Q14[ k ] += b_Q14_ptr[ i ];
}
b_Q14_ptr += LTP_ORDER;
}
/* m = ( w * d' ) / ( sum( w ) + 1e-3 ); */
/* Find maximum absolute value of d_Q14 and the bits used by w in Q0 */
max_abs_d_Q14 = 0;
max_w_bits = 0;
for( k = 0; k < nb_subfr; k++ ) {
max_abs_d_Q14 = silk_max_32( max_abs_d_Q14, silk_abs( d_Q14[ k ] ) );
/* w[ k ] is in Q( 18 - corr_rshifts[ k ] ) */
/* Find bits needed in Q( 18 - maxRshifts ) */
max_w_bits = silk_max_32( max_w_bits, 32 - silk_CLZ32( w[ k ] ) + corr_rshifts[ k ] - maxRshifts );
}
/* max_abs_d_Q14 = (5 << 15); worst case, i.e. LTP_ORDER * -silk_int16_MIN */
silk_assert( max_abs_d_Q14 <= ( 5 << 15 ) );
/* How many bits is needed for w*d' in Q( 18 - maxRshifts ) in the worst case, of all d_Q14's being equal to max_abs_d_Q14 */
extra_shifts = max_w_bits + 32 - silk_CLZ32( max_abs_d_Q14 ) - 14;
/* Subtract what we got available; bits in output var plus maxRshifts */
extra_shifts -= ( 32 - 1 - 2 + maxRshifts ); /* Keep sign bit free as well as 2 bits for accumulation */
extra_shifts = silk_max_int( extra_shifts, 0 );
maxRshifts_wxtra = maxRshifts + extra_shifts;
temp32 = silk_RSHIFT( 262, maxRshifts + extra_shifts ) + 1; /* 1e-3f in Q( 18 - (maxRshifts + extra_shifts) ) */
wd = 0;
for( k = 0; k < nb_subfr; k++ ) {
/* w has at least 2 bits of headroom so no overflow should happen */
temp32 = silk_ADD32( temp32, silk_RSHIFT( w[ k ], maxRshifts_wxtra - corr_rshifts[ k ] ) ); /* Q( 18 - maxRshifts_wxtra ) */
wd = silk_ADD32( wd, silk_LSHIFT( silk_SMULWW( silk_RSHIFT( w[ k ], maxRshifts_wxtra - corr_rshifts[ k ] ), d_Q14[ k ] ), 2 ) ); /* Q( 18 - maxRshifts_wxtra ) */
}
m_Q12 = silk_DIV32_varQ( wd, temp32, 12 );
b_Q14_ptr = b_Q14;
for( k = 0; k < nb_subfr; k++ ) {
/* w_fix[ k ] from Q( 18 - corr_rshifts[ k ] ) to Q( 16 ) */
if( 2 - corr_rshifts[k] > 0 ) {
temp32 = silk_RSHIFT( w[ k ], 2 - corr_rshifts[ k ] );
} else {
temp32 = silk_LSHIFT_SAT32( w[ k ], corr_rshifts[ k ] - 2 );
}
g_Q26 = silk_MUL(
silk_DIV32(
SILK_FIX_CONST( LTP_SMOOTHING, 26 ),
silk_RSHIFT( SILK_FIX_CONST( LTP_SMOOTHING, 26 ), 10 ) + temp32 ), /* Q10 */
silk_LSHIFT_SAT32( silk_SUB_SAT32( (opus_int32)m_Q12, silk_RSHIFT( d_Q14[ k ], 2 ) ), 4 ) ); /* Q16 */
temp32 = 0;
for( i = 0; i < LTP_ORDER; i++ ) {
delta_b_Q14[ i ] = silk_max_16( b_Q14_ptr[ i ], 1638 ); /* 1638_Q14 = 0.1_Q0 */
temp32 += delta_b_Q14[ i ]; /* Q14 */
}
temp32 = silk_DIV32( g_Q26, temp32 ); /* Q14 -> Q12 */
for( i = 0; i < LTP_ORDER; i++ ) {
b_Q14_ptr[ i ] = silk_LIMIT_32( (opus_int32)b_Q14_ptr[ i ] + silk_SMULWB( silk_LSHIFT_SAT32( temp32, 4 ), delta_b_Q14[ i ] ), -16000, 28000 );
}
b_Q14_ptr += LTP_ORDER;
}
}
opus_int silk_VAD_GetSA_Q8( /* O Return value, 0 if success */
silk_encoder_state *psEncC, /* I/O Encoder state */
const opus_int16 pIn[] /* I PCM input */
)
{
opus_int SA_Q15, pSNR_dB_Q7, input_tilt;
opus_int decimated_framelength1, decimated_framelength2;
opus_int decimated_framelength;
opus_int dec_subframe_length, dec_subframe_offset, SNR_Q7, i, b, s;
opus_int32 sumSquared, smooth_coef_Q16;
opus_int16 HPstateTmp;
VARDECL( opus_int16, X );
opus_int32 Xnrg[ VAD_N_BANDS ];
opus_int32 NrgToNoiseRatio_Q8[ VAD_N_BANDS ];
opus_int32 speech_nrg, x_tmp;
opus_int X_offset[ VAD_N_BANDS ];
opus_int ret = 0;
silk_VAD_state *psSilk_VAD = &psEncC->sVAD;
SAVE_STACK;
/* Safety checks */
silk_assert( VAD_N_BANDS == 4 );
silk_assert( MAX_FRAME_LENGTH >= psEncC->frame_length );
silk_assert( psEncC->frame_length <= 512 );
silk_assert( psEncC->frame_length == 8 * silk_RSHIFT( psEncC->frame_length, 3 ) );
/***********************/
/* Filter and Decimate */
/***********************/
decimated_framelength1 = silk_RSHIFT( psEncC->frame_length, 1 );
decimated_framelength2 = silk_RSHIFT( psEncC->frame_length, 2 );
decimated_framelength = silk_RSHIFT( psEncC->frame_length, 3 );
/* Decimate into 4 bands:
0 L 3L L 3L 5L
- -- - -- --
8 8 2 4 4
[0-1 kHz| temp. |1-2 kHz| 2-4 kHz | 4-8 kHz |
They're arranged to allow the minimal ( frame_length / 4 ) extra
scratch space during the downsampling process */
X_offset[ 0 ] = 0;
X_offset[ 1 ] = decimated_framelength + decimated_framelength2;
X_offset[ 2 ] = X_offset[ 1 ] + decimated_framelength;
X_offset[ 3 ] = X_offset[ 2 ] + decimated_framelength2;
ALLOC( X, X_offset[ 3 ] + decimated_framelength1, opus_int16 );
/* 0-8 kHz to 0-4 kHz and 4-8 kHz */
silk_ana_filt_bank_1( pIn, &psSilk_VAD->AnaState[ 0 ],
X, &X[ X_offset[ 3 ] ], psEncC->frame_length );
/* 0-4 kHz to 0-2 kHz and 2-4 kHz */
silk_ana_filt_bank_1( X, &psSilk_VAD->AnaState1[ 0 ],
X, &X[ X_offset[ 2 ] ], decimated_framelength1 );
/* 0-2 kHz to 0-1 kHz and 1-2 kHz */
silk_ana_filt_bank_1( X, &psSilk_VAD->AnaState2[ 0 ],
X, &X[ X_offset[ 1 ] ], decimated_framelength2 );
/*********************************************/
/* HP filter on lowest band (differentiator) */
/*********************************************/
X[ decimated_framelength - 1 ] = silk_RSHIFT( X[ decimated_framelength - 1 ], 1 );
HPstateTmp = X[ decimated_framelength - 1 ];
for( i = decimated_framelength - 1; i > 0; i-- ) {
X[ i - 1 ] = silk_RSHIFT( X[ i - 1 ], 1 );
X[ i ] -= X[ i - 1 ];
}
X[ 0 ] -= psSilk_VAD->HPstate;
psSilk_VAD->HPstate = HPstateTmp;
/*************************************/
/* Calculate the energy in each band */
/*************************************/
for( b = 0; b < VAD_N_BANDS; b++ ) {
/* Find the decimated framelength in the non-uniformly divided bands */
decimated_framelength = silk_RSHIFT( psEncC->frame_length, silk_min_int( VAD_N_BANDS - b, VAD_N_BANDS - 1 ) );
/* Split length into subframe lengths */
dec_subframe_length = silk_RSHIFT( decimated_framelength, VAD_INTERNAL_SUBFRAMES_LOG2 );
dec_subframe_offset = 0;
/* Compute energy per sub-frame */
/* initialize with summed energy of last subframe */
Xnrg[ b ] = psSilk_VAD->XnrgSubfr[ b ];
for( s = 0; s < VAD_INTERNAL_SUBFRAMES; s++ ) {
sumSquared = 0;
for( i = 0; i < dec_subframe_length; i++ ) {
/* The energy will be less than dec_subframe_length * ( silk_int16_MIN / 8 ) ^ 2. */
/* Therefore we can accumulate with no risk of overflow (unless dec_subframe_length > 128) */
x_tmp = silk_RSHIFT(
X[ X_offset[ b ] + i + dec_subframe_offset ], 3 );
sumSquared = silk_SMLABB( sumSquared, x_tmp, x_tmp );
/* Safety check */
silk_assert( sumSquared >= 0 );
}
/* Add/saturate summed energy of current subframe */
if( s < VAD_INTERNAL_SUBFRAMES - 1 ) {
Xnrg[ b ] = silk_ADD_POS_SAT32( Xnrg[ b ], sumSquared );
} else {
/* Look-ahead subframe */
Xnrg[ b ] = silk_ADD_POS_SAT32( Xnrg[ b ], silk_RSHIFT( sumSquared, 1 ) );
}
dec_subframe_offset += dec_subframe_length;
}
psSilk_VAD->XnrgSubfr[ b ] = sumSquared;
}
/********************/
/* Noise estimation */
/********************/
silk_VAD_GetNoiseLevels( &Xnrg[ 0 ], psSilk_VAD );
/***********************************************/
/* Signal-plus-noise to noise ratio estimation */
/***********************************************/
sumSquared = 0;
input_tilt = 0;
for( b = 0; b < VAD_N_BANDS; b++ ) {
speech_nrg = Xnrg[ b ] - psSilk_VAD->NL[ b ];
if( speech_nrg > 0 ) {
/* Divide, with sufficient resolution */
if( ( Xnrg[ b ] & 0xFF800000 ) == 0 ) {
NrgToNoiseRatio_Q8[ b ] = silk_DIV32( silk_LSHIFT( Xnrg[ b ], 8 ), psSilk_VAD->NL[ b ] + 1 );
} else {
NrgToNoiseRatio_Q8[ b ] = silk_DIV32( Xnrg[ b ], silk_RSHIFT( psSilk_VAD->NL[ b ], 8 ) + 1 );
}
/* Convert to log domain */
SNR_Q7 = silk_lin2log( NrgToNoiseRatio_Q8[ b ] ) - 8 * 128;
/* Sum-of-squares */
sumSquared = silk_SMLABB( sumSquared, SNR_Q7, SNR_Q7 ); /* Q14 */
/* Tilt measure */
if( speech_nrg < ( (opus_int32)1 << 20 ) ) {
/* Scale down SNR value for small subband speech energies */
SNR_Q7 = silk_SMULWB( silk_LSHIFT( silk_SQRT_APPROX( speech_nrg ), 6 ), SNR_Q7 );
}
input_tilt = silk_SMLAWB( input_tilt, tiltWeights[ b ], SNR_Q7 );
} else {
NrgToNoiseRatio_Q8[ b ] = 256;
}
}
/* Mean-of-squares */
sumSquared = silk_DIV32_16( sumSquared, VAD_N_BANDS ); /* Q14 */
/* Root-mean-square approximation, scale to dBs, and write to output pointer */
pSNR_dB_Q7 = (opus_int16)( 3 * silk_SQRT_APPROX( sumSquared ) ); /* Q7 */
/*********************************/
/* Speech Probability Estimation */
/*********************************/
SA_Q15 = silk_sigm_Q15( silk_SMULWB( VAD_SNR_FACTOR_Q16, pSNR_dB_Q7 ) - VAD_NEGATIVE_OFFSET_Q5 );
/**************************/
/* Frequency Tilt Measure */
/**************************/
psEncC->input_tilt_Q15 = silk_LSHIFT( silk_sigm_Q15( input_tilt ) - 16384, 1 );
/**************************************************/
/* Scale the sigmoid output based on power levels */
/**************************************************/
speech_nrg = 0;
for( b = 0; b < VAD_N_BANDS; b++ ) {
/* Accumulate signal-without-noise energies, higher frequency bands have more weight */
speech_nrg += ( b + 1 ) * silk_RSHIFT( Xnrg[ b ] - psSilk_VAD->NL[ b ], 4 );
}
/* Power scaling */
if( speech_nrg <= 0 ) {
SA_Q15 = silk_RSHIFT( SA_Q15, 1 );
} else if( speech_nrg < 32768 ) {
if( psEncC->frame_length == 10 * psEncC->fs_kHz ) {
speech_nrg = silk_LSHIFT_SAT32( speech_nrg, 16 );
} else {
speech_nrg = silk_LSHIFT_SAT32( speech_nrg, 15 );
}
/* square-root */
speech_nrg = silk_SQRT_APPROX( speech_nrg );
SA_Q15 = silk_SMULWB( 32768 + speech_nrg, SA_Q15 );
}
/* Copy the resulting speech activity in Q8 */
psEncC->speech_activity_Q8 = silk_min_int( silk_RSHIFT( SA_Q15, 7 ), silk_uint8_MAX );
/***********************************/
/* Energy Level and SNR estimation */
/***********************************/
/* Smoothing coefficient */
smooth_coef_Q16 = silk_SMULWB( VAD_SNR_SMOOTH_COEF_Q18, silk_SMULWB( (opus_int32)SA_Q15, SA_Q15 ) );
if( psEncC->frame_length == 10 * psEncC->fs_kHz ) {
smooth_coef_Q16 >>= 1;
}
/* Limit, stabilize, convert and quantize NLSFs */
void silk_process_NLSFs(
silk_encoder_state *psEncC, /* I/O Encoder state */
opus_int16 PredCoef_Q12[ 2 ][ MAX_LPC_ORDER ], /* O Prediction coefficients */
opus_int16 pNLSF_Q15[ MAX_LPC_ORDER ], /* I/O Normalized LSFs (quant out) (0 - (2^15-1)) */
const opus_int16 prev_NLSFq_Q15[ MAX_LPC_ORDER ] /* I Previous Normalized LSFs (0 - (2^15-1)) */
)
{
opus_int i, doInterpolate;
opus_int NLSF_mu_Q20;
opus_int32 i_sqr_Q15;
opus_int16 pNLSF0_temp_Q15[ MAX_LPC_ORDER ];
opus_int16 pNLSFW_QW[ MAX_LPC_ORDER ];
opus_int16 pNLSFW0_temp_QW[ MAX_LPC_ORDER ];
silk_assert(psEncC->speech_activity_Q8 >= 0);
silk_assert(psEncC->speech_activity_Q8 <= SILK_FIX_CONST(1.0, 8));
silk_assert(psEncC->useInterpolatedNLSFs == 1 || psEncC->indices.NLSFInterpCoef_Q2 == (1 << 2));
/***********************/
/* Calculate mu values */
/***********************/
/* NLSF_mu = 0.003 - 0.0015 * psEnc->speech_activity; */
NLSF_mu_Q20 = silk_SMLAWB(SILK_FIX_CONST(0.003, 20), SILK_FIX_CONST(-0.001, 28), psEncC->speech_activity_Q8);
if(psEncC->nb_subfr == 2) {
/* Multiply by 1.5 for 10 ms packets */
NLSF_mu_Q20 = silk_ADD_RSHIFT(NLSF_mu_Q20, NLSF_mu_Q20, 1);
}
silk_assert(NLSF_mu_Q20 > 0);
silk_assert(NLSF_mu_Q20 <= SILK_FIX_CONST(0.005, 20));
/* Calculate NLSF weights */
silk_NLSF_VQ_weights_laroia(pNLSFW_QW, pNLSF_Q15, psEncC->predictLPCOrder);
/* Update NLSF weights for interpolated NLSFs */
doInterpolate = (psEncC->useInterpolatedNLSFs == 1) && (psEncC->indices.NLSFInterpCoef_Q2 < 4);
if(doInterpolate) {
/* Calculate the interpolated NLSF vector for the first half */
silk_interpolate(pNLSF0_temp_Q15, prev_NLSFq_Q15, pNLSF_Q15,
psEncC->indices.NLSFInterpCoef_Q2, psEncC->predictLPCOrder);
/* Calculate first half NLSF weights for the interpolated NLSFs */
silk_NLSF_VQ_weights_laroia(pNLSFW0_temp_QW, pNLSF0_temp_Q15, psEncC->predictLPCOrder);
/* Update NLSF weights with contribution from first half */
i_sqr_Q15 = silk_LSHIFT(silk_SMULBB(psEncC->indices.NLSFInterpCoef_Q2, psEncC->indices.NLSFInterpCoef_Q2), 11);
for(i = 0; i < psEncC->predictLPCOrder; i++) {
pNLSFW_QW[ i ] = silk_SMLAWB(silk_RSHIFT(pNLSFW_QW[ i ], 1), (opus_int32)pNLSFW0_temp_QW[ i ], i_sqr_Q15);
silk_assert(pNLSFW_QW[ i ] >= 1);
}
}
silk_NLSF_encode(psEncC->indices.NLSFIndices, pNLSF_Q15, psEncC->psNLSF_CB, pNLSFW_QW,
NLSF_mu_Q20, psEncC->NLSF_MSVQ_Survivors, psEncC->indices.signalType);
/* Convert quantized NLSFs back to LPC coefficients */
silk_NLSF2A(PredCoef_Q12[ 1 ], pNLSF_Q15, psEncC->predictLPCOrder);
if(doInterpolate) {
/* Calculate the interpolated, quantized LSF vector for the first half */
silk_interpolate(pNLSF0_temp_Q15, prev_NLSFq_Q15, pNLSF_Q15,
psEncC->indices.NLSFInterpCoef_Q2, psEncC->predictLPCOrder);
/* Convert back to LPC coefficients */
silk_NLSF2A(PredCoef_Q12[ 0 ], pNLSF0_temp_Q15, psEncC->predictLPCOrder);
} else {
/* Copy LPC coefficients for first half from second half */
silk_memcpy(PredCoef_Q12[ 0 ], PredCoef_Q12[ 1 ], psEncC->predictLPCOrder * sizeof(opus_int16));
}
}
/* Convert Left/Right stereo signal to adaptive Mid/Side representation */
void silk_stereo_LR_to_MS(
stereo_enc_state *state, /* I/O State */
opus_int16 x1[], /* I/O Left input signal, becomes mid signal */
opus_int16 x2[], /* I/O Right input signal, becomes side signal */
opus_int8 ix[ 2 ][ 3 ], /* O Quantization indices */
opus_int8 *mid_only_flag, /* O Flag: only mid signal coded */
opus_int32 mid_side_rates_bps[], /* O Bitrates for mid and side signals */
opus_int32 total_rate_bps, /* I Total bitrate */
opus_int prev_speech_act_Q8, /* I Speech activity level in previous frame */
opus_int toMono, /* I Last frame before a stereo->mono transition */
opus_int fs_kHz, /* I Sample rate (kHz) */
opus_int frame_length /* I Number of samples */
)
{
opus_int n, is10msFrame, denom_Q16, delta0_Q13, delta1_Q13;
opus_int32 sum, diff, smooth_coef_Q16, pred_Q13[ 2 ], pred0_Q13, pred1_Q13;
opus_int32 LP_ratio_Q14, HP_ratio_Q14, frac_Q16, frac_3_Q16, min_mid_rate_bps, width_Q14, w_Q24, deltaw_Q24;
VARDECL( opus_int16, side );
VARDECL( opus_int16, LP_mid );
VARDECL( opus_int16, HP_mid );
VARDECL( opus_int16, LP_side );
VARDECL( opus_int16, HP_side );
opus_int16 *mid = &x1[ -2 ];
SAVE_STACK;
ALLOC( side, frame_length + 2, opus_int16 );
/* Convert to basic mid/side signals */
for( n = 0; n < frame_length + 2; n++ ) {
sum = x1[ n - 2 ] + (opus_int32)x2[ n - 2 ];
diff = x1[ n - 2 ] - (opus_int32)x2[ n - 2 ];
mid[ n ] = (opus_int16)silk_RSHIFT_ROUND( sum, 1 );
side[ n ] = (opus_int16)silk_SAT16( silk_RSHIFT_ROUND( diff, 1 ) );
}
/* Buffering */
silk_memcpy( mid, state->sMid, 2 * sizeof( opus_int16 ) );
silk_memcpy( side, state->sSide, 2 * sizeof( opus_int16 ) );
silk_memcpy( state->sMid, &mid[ frame_length ], 2 * sizeof( opus_int16 ) );
silk_memcpy( state->sSide, &side[ frame_length ], 2 * sizeof( opus_int16 ) );
/* LP and HP filter mid signal */
ALLOC( LP_mid, frame_length, opus_int16 );
ALLOC( HP_mid, frame_length, opus_int16 );
for( n = 0; n < frame_length; n++ ) {
sum = silk_RSHIFT_ROUND( silk_ADD_LSHIFT( mid[ n ] + mid[ n + 2 ], mid[ n + 1 ], 1 ), 2 );
LP_mid[ n ] = sum;
HP_mid[ n ] = mid[ n + 1 ] - sum;
}
/* LP and HP filter side signal */
ALLOC( LP_side, frame_length, opus_int16 );
ALLOC( HP_side, frame_length, opus_int16 );
for( n = 0; n < frame_length; n++ ) {
sum = silk_RSHIFT_ROUND( silk_ADD_LSHIFT( side[ n ] + side[ n + 2 ], side[ n + 1 ], 1 ), 2 );
LP_side[ n ] = sum;
HP_side[ n ] = side[ n + 1 ] - sum;
}
/* Find energies and predictors */
is10msFrame = frame_length == 10 * fs_kHz;
smooth_coef_Q16 = is10msFrame ?
SILK_FIX_CONST( STEREO_RATIO_SMOOTH_COEF / 2, 16 ) :
SILK_FIX_CONST( STEREO_RATIO_SMOOTH_COEF, 16 );
smooth_coef_Q16 = silk_SMULWB( silk_SMULBB( prev_speech_act_Q8, prev_speech_act_Q8 ), smooth_coef_Q16 );
pred_Q13[ 0 ] = silk_stereo_find_predictor( &LP_ratio_Q14, LP_mid, LP_side, &state->mid_side_amp_Q0[ 0 ], frame_length, smooth_coef_Q16 );
pred_Q13[ 1 ] = silk_stereo_find_predictor( &HP_ratio_Q14, HP_mid, HP_side, &state->mid_side_amp_Q0[ 2 ], frame_length, smooth_coef_Q16 );
/* Ratio of the norms of residual and mid signals */
frac_Q16 = silk_SMLABB( HP_ratio_Q14, LP_ratio_Q14, 3 );
frac_Q16 = silk_min( frac_Q16, SILK_FIX_CONST( 1, 16 ) );
/* Determine bitrate distribution between mid and side, and possibly reduce stereo width */
total_rate_bps -= is10msFrame ? 1200 : 600; /* Subtract approximate bitrate for coding stereo parameters */
if( total_rate_bps < 1 ) {
total_rate_bps = 1;
}
min_mid_rate_bps = silk_SMLABB( 2000, fs_kHz, 900 );
silk_assert( min_mid_rate_bps < 32767 );
/* Default bitrate distribution: 8 parts for Mid and (5+3*frac) parts for Side. so: mid_rate = ( 8 / ( 13 + 3 * frac ) ) * total_ rate */
frac_3_Q16 = silk_MUL( 3, frac_Q16 );
mid_side_rates_bps[ 0 ] = silk_DIV32_varQ( total_rate_bps, SILK_FIX_CONST( 8 + 5, 16 ) + frac_3_Q16, 16+3 );
/* If Mid bitrate below minimum, reduce stereo width */
if( mid_side_rates_bps[ 0 ] < min_mid_rate_bps ) {
mid_side_rates_bps[ 0 ] = min_mid_rate_bps;
mid_side_rates_bps[ 1 ] = total_rate_bps - mid_side_rates_bps[ 0 ];
/* width = 4 * ( 2 * side_rate - min_rate ) / ( ( 1 + 3 * frac ) * min_rate ) */
width_Q14 = silk_DIV32_varQ( silk_LSHIFT( mid_side_rates_bps[ 1 ], 1 ) - min_mid_rate_bps,
silk_SMULWB( SILK_FIX_CONST( 1, 16 ) + frac_3_Q16, min_mid_rate_bps ), 14+2 );
width_Q14 = silk_LIMIT( width_Q14, 0, SILK_FIX_CONST( 1, 14 ) );
} else {
mid_side_rates_bps[ 1 ] = total_rate_bps - mid_side_rates_bps[ 0 ];
width_Q14 = SILK_FIX_CONST( 1, 14 );
}
/* Smoother */
state->smth_width_Q14 = (opus_int16)silk_SMLAWB( state->smth_width_Q14, width_Q14 - state->smth_width_Q14, smooth_coef_Q16 );
/* At very low bitrates or for inputs that are nearly amplitude panned, switch to panned-mono coding */
*mid_only_flag = 0;
if( toMono ) {
/* Last frame before stereo->mono transition; collapse stereo width */
width_Q14 = 0;
pred_Q13[ 0 ] = 0;
pred_Q13[ 1 ] = 0;
silk_stereo_quant_pred( pred_Q13, ix );
} else if( state->width_prev_Q14 == 0 &&
( 8 * total_rate_bps < 13 * min_mid_rate_bps || silk_SMULWB( frac_Q16, state->smth_width_Q14 ) < SILK_FIX_CONST( 0.05, 14 ) ) )
{
/* Code as panned-mono; previous frame already had zero width */
/* Scale down and quantize predictors */
pred_Q13[ 0 ] = silk_RSHIFT( silk_SMULBB( state->smth_width_Q14, pred_Q13[ 0 ] ), 14 );
pred_Q13[ 1 ] = silk_RSHIFT( silk_SMULBB( state->smth_width_Q14, pred_Q13[ 1 ] ), 14 );
silk_stereo_quant_pred( pred_Q13, ix );
/* Collapse stereo width */
width_Q14 = 0;
pred_Q13[ 0 ] = 0;
pred_Q13[ 1 ] = 0;
mid_side_rates_bps[ 0 ] = total_rate_bps;
mid_side_rates_bps[ 1 ] = 0;
*mid_only_flag = 1;
} else if( state->width_prev_Q14 != 0 &&
( 8 * total_rate_bps < 11 * min_mid_rate_bps || silk_SMULWB( frac_Q16, state->smth_width_Q14 ) < SILK_FIX_CONST( 0.02, 14 ) ) )
{
/* Transition to zero-width stereo */
/* Scale down and quantize predictors */
pred_Q13[ 0 ] = silk_RSHIFT( silk_SMULBB( state->smth_width_Q14, pred_Q13[ 0 ] ), 14 );
pred_Q13[ 1 ] = silk_RSHIFT( silk_SMULBB( state->smth_width_Q14, pred_Q13[ 1 ] ), 14 );
silk_stereo_quant_pred( pred_Q13, ix );
/* Collapse stereo width */
width_Q14 = 0;
pred_Q13[ 0 ] = 0;
pred_Q13[ 1 ] = 0;
} else if( state->smth_width_Q14 > SILK_FIX_CONST( 0.95, 14 ) ) {
/* Full-width stereo coding */
silk_stereo_quant_pred( pred_Q13, ix );
width_Q14 = SILK_FIX_CONST( 1, 14 );
} else {
/* Reduced-width stereo coding; scale down and quantize predictors */
pred_Q13[ 0 ] = silk_RSHIFT( silk_SMULBB( state->smth_width_Q14, pred_Q13[ 0 ] ), 14 );
pred_Q13[ 1 ] = silk_RSHIFT( silk_SMULBB( state->smth_width_Q14, pred_Q13[ 1 ] ), 14 );
silk_stereo_quant_pred( pred_Q13, ix );
width_Q14 = state->smth_width_Q14;
}
/* Make sure to keep on encoding until the tapered output has been transmitted */
if( *mid_only_flag == 1 ) {
state->silent_side_len += frame_length - STEREO_INTERP_LEN_MS * fs_kHz;
if( state->silent_side_len < LA_SHAPE_MS * fs_kHz ) {
*mid_only_flag = 0;
} else {
/* Limit to avoid wrapping around */
state->silent_side_len = 10000;
}
} else {
state->silent_side_len = 0;
}
if( *mid_only_flag == 0 && mid_side_rates_bps[ 1 ] < 1 ) {
mid_side_rates_bps[ 1 ] = 1;
mid_side_rates_bps[ 0 ] = silk_max_int( 1, total_rate_bps - mid_side_rates_bps[ 1 ]);
}
/* Interpolate predictors and subtract prediction from side channel */
pred0_Q13 = -state->pred_prev_Q13[ 0 ];
pred1_Q13 = -state->pred_prev_Q13[ 1 ];
w_Q24 = silk_LSHIFT( state->width_prev_Q14, 10 );
denom_Q16 = silk_DIV32_16( (opus_int32)1 << 16, STEREO_INTERP_LEN_MS * fs_kHz );
delta0_Q13 = -silk_RSHIFT_ROUND( silk_SMULBB( pred_Q13[ 0 ] - state->pred_prev_Q13[ 0 ], denom_Q16 ), 16 );
delta1_Q13 = -silk_RSHIFT_ROUND( silk_SMULBB( pred_Q13[ 1 ] - state->pred_prev_Q13[ 1 ], denom_Q16 ), 16 );
deltaw_Q24 = silk_LSHIFT( silk_SMULWB( width_Q14 - state->width_prev_Q14, denom_Q16 ), 10 );
for( n = 0; n < STEREO_INTERP_LEN_MS * fs_kHz; n++ ) {
pred0_Q13 += delta0_Q13;
pred1_Q13 += delta1_Q13;
w_Q24 += deltaw_Q24;
sum = silk_LSHIFT( silk_ADD_LSHIFT( mid[ n ] + mid[ n + 2 ], mid[ n + 1 ], 1 ), 9 ); /* Q11 */
sum = silk_SMLAWB( silk_SMULWB( w_Q24, side[ n + 1 ] ), sum, pred0_Q13 ); /* Q8 */
sum = silk_SMLAWB( sum, silk_LSHIFT( (opus_int32)mid[ n + 1 ], 11 ), pred1_Q13 ); /* Q8 */
x2[ n - 1 ] = (opus_int16)silk_SAT16( silk_RSHIFT_ROUND( sum, 8 ) );
}
pred0_Q13 = -pred_Q13[ 0 ];
pred1_Q13 = -pred_Q13[ 1 ];
w_Q24 = silk_LSHIFT( width_Q14, 10 );
for( n = STEREO_INTERP_LEN_MS * fs_kHz; n < frame_length; n++ ) {
sum = silk_LSHIFT( silk_ADD_LSHIFT( mid[ n ] + mid[ n + 2 ], mid[ n + 1 ], 1 ), 9 ); /* Q11 */
sum = silk_SMLAWB( silk_SMULWB( w_Q24, side[ n + 1 ] ), sum, pred0_Q13 ); /* Q8 */
sum = silk_SMLAWB( sum, silk_LSHIFT( (opus_int32)mid[ n + 1 ], 11 ), pred1_Q13 ); /* Q8 */
x2[ n - 1 ] = (opus_int16)silk_SAT16( silk_RSHIFT_ROUND( sum, 8 ) );
}
state->pred_prev_Q13[ 0 ] = (opus_int16)pred_Q13[ 0 ];
state->pred_prev_Q13[ 1 ] = (opus_int16)pred_Q13[ 1 ];
state->width_prev_Q14 = (opus_int16)width_Q14;
RESTORE_STACK;
}
/* 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;
}
/* Entropy constrained matrix-weighted VQ, hard-coded to 5-element vectors, for a single input data vector */
void silk_VQ_WMat_EC_sse4_1(
opus_int8 *ind, /* O index of best codebook vector */
opus_int32 *rate_dist_Q14, /* O best weighted quant error + mu * rate */
opus_int *gain_Q7, /* O sum of absolute LTP coefficients */
const opus_int16 *in_Q14, /* I input vector to be quantized */
const opus_int32 *W_Q18, /* I weighting matrix */
const opus_int8 *cb_Q7, /* I codebook */
const opus_uint8 *cb_gain_Q7, /* I codebook effective gain */
const opus_uint8 *cl_Q5, /* I code length for each codebook vector */
const opus_int mu_Q9, /* I tradeoff betw. weighted error and rate */
const opus_int32 max_gain_Q7, /* I maximum sum of absolute LTP coefficients */
opus_int L /* I number of vectors in codebook */
)
{
opus_int k, gain_tmp_Q7;
const opus_int8 *cb_row_Q7;
opus_int16 diff_Q14[ 5 ];
opus_int32 sum1_Q14, sum2_Q16;
__m128i C_tmp1, C_tmp2, C_tmp3, C_tmp4, C_tmp5;
/* Loop over codebook */
*rate_dist_Q14 = silk_int32_MAX;
cb_row_Q7 = cb_Q7;
for( k = 0; k < L; k++ ) {
gain_tmp_Q7 = cb_gain_Q7[k];
diff_Q14[ 0 ] = in_Q14[ 0 ] - silk_LSHIFT( cb_row_Q7[ 0 ], 7 );
C_tmp1 = OP_CVTEPI16_EPI32_M64( &in_Q14[ 1 ] );
C_tmp2 = OP_CVTEPI8_EPI32_M32( &cb_row_Q7[ 1 ] );
C_tmp2 = _mm_slli_epi32( C_tmp2, 7 );
C_tmp1 = _mm_sub_epi32( C_tmp1, C_tmp2 );
diff_Q14[ 1 ] = _mm_extract_epi16( C_tmp1, 0 );
diff_Q14[ 2 ] = _mm_extract_epi16( C_tmp1, 2 );
diff_Q14[ 3 ] = _mm_extract_epi16( C_tmp1, 4 );
diff_Q14[ 4 ] = _mm_extract_epi16( C_tmp1, 6 );
/* Weighted rate */
sum1_Q14 = silk_SMULBB( mu_Q9, cl_Q5[ k ] );
/* Penalty for too large gain */
sum1_Q14 = silk_ADD_LSHIFT32( sum1_Q14, silk_max( silk_SUB32( gain_tmp_Q7, max_gain_Q7 ), 0 ), 10 );
silk_assert( sum1_Q14 >= 0 );
/* first row of W_Q18 */
C_tmp3 = _mm_loadu_si128( (__m128i *)(&W_Q18[ 1 ] ) );
C_tmp4 = _mm_mul_epi32( C_tmp3, C_tmp1 );
C_tmp4 = _mm_srli_si128( C_tmp4, 2 );
C_tmp1 = _mm_shuffle_epi32( C_tmp1, _MM_SHUFFLE( 0, 3, 2, 1 ) ); /* shift right 4 bytes */
C_tmp3 = _mm_shuffle_epi32( C_tmp3, _MM_SHUFFLE( 0, 3, 2, 1 ) ); /* shift right 4 bytes */
C_tmp5 = _mm_mul_epi32( C_tmp3, C_tmp1 );
C_tmp5 = _mm_srli_si128( C_tmp5, 2 );
C_tmp5 = _mm_add_epi32( C_tmp4, C_tmp5 );
C_tmp5 = _mm_slli_epi32( C_tmp5, 1 );
C_tmp5 = _mm_add_epi32( C_tmp5, _mm_shuffle_epi32( C_tmp5, _MM_SHUFFLE( 0, 0, 0, 2 ) ) );
sum2_Q16 = _mm_cvtsi128_si32( C_tmp5 );
sum2_Q16 = silk_SMLAWB( sum2_Q16, W_Q18[ 0 ], diff_Q14[ 0 ] );
sum1_Q14 = silk_SMLAWB( sum1_Q14, sum2_Q16, diff_Q14[ 0 ] );
/* second row of W_Q18 */
sum2_Q16 = silk_SMULWB( W_Q18[ 7 ], diff_Q14[ 2 ] );
sum2_Q16 = silk_SMLAWB( sum2_Q16, W_Q18[ 8 ], diff_Q14[ 3 ] );
sum2_Q16 = silk_SMLAWB( sum2_Q16, W_Q18[ 9 ], diff_Q14[ 4 ] );
sum2_Q16 = silk_LSHIFT( sum2_Q16, 1 );
sum2_Q16 = silk_SMLAWB( sum2_Q16, W_Q18[ 6 ], diff_Q14[ 1 ] );
sum1_Q14 = silk_SMLAWB( sum1_Q14, sum2_Q16, diff_Q14[ 1 ] );
/* third row of W_Q18 */
sum2_Q16 = silk_SMULWB( W_Q18[ 13 ], diff_Q14[ 3 ] );
sum2_Q16 = silk_SMLAWB( sum2_Q16, W_Q18[ 14 ], diff_Q14[ 4 ] );
sum2_Q16 = silk_LSHIFT( sum2_Q16, 1 );
sum2_Q16 = silk_SMLAWB( sum2_Q16, W_Q18[ 12 ], diff_Q14[ 2 ] );
sum1_Q14 = silk_SMLAWB( sum1_Q14, sum2_Q16, diff_Q14[ 2 ] );
/* fourth row of W_Q18 */
sum2_Q16 = silk_SMULWB( W_Q18[ 19 ], diff_Q14[ 4 ] );
sum2_Q16 = silk_LSHIFT( sum2_Q16, 1 );
sum2_Q16 = silk_SMLAWB( sum2_Q16, W_Q18[ 18 ], diff_Q14[ 3 ] );
sum1_Q14 = silk_SMLAWB( sum1_Q14, sum2_Q16, diff_Q14[ 3 ] );
/* last row of W_Q18 */
sum2_Q16 = silk_SMULWB( W_Q18[ 24 ], diff_Q14[ 4 ] );
sum1_Q14 = silk_SMLAWB( sum1_Q14, sum2_Q16, diff_Q14[ 4 ] );
silk_assert( sum1_Q14 >= 0 );
/* find best */
if( sum1_Q14 < *rate_dist_Q14 ) {
*rate_dist_Q14 = sum1_Q14;
*ind = (opus_int8)k;
*gain_Q7 = gain_tmp_Q7;
}
/* Go to next cbk vector */
cb_row_Q7 += LTP_ORDER;
}
}
/* uses SMLAWB(), requiring armv5E and higher. */
opus_int32 silk_schur( /* O Returns residual energy */
opus_int16 *rc_Q15, /* O reflection coefficients [order] Q15 */
const opus_int32 *c, /* I correlations [order+1] */
const opus_int32 order /* I prediction order */
)
{
opus_int k, n, lz;
opus_int32 C[ SILK_MAX_ORDER_LPC + 1 ][ 2 ];
opus_int32 Ctmp1, Ctmp2, rc_tmp_Q15;
silk_assert( order==6||order==8||order==10||order==12||order==14||order==16 );
/* Get number of leading zeros */
lz = silk_CLZ32( c[ 0 ] );
/* Copy correlations and adjust level to Q30 */
if( lz < 2 ) {
/* lz must be 1, so shift one to the right */
for( k = 0; k < order + 1; k++ ) {
C[ k ][ 0 ] = C[ k ][ 1 ] = silk_RSHIFT( c[ k ], 1 );
}
} else if( lz > 2 ) {
/* Shift to the left */
lz -= 2;
for( k = 0; k < order + 1; k++ ) {
C[ k ][ 0 ] = C[ k ][ 1 ] = silk_LSHIFT( c[ k ], lz );
}
} else {
/* No need to shift */
for( k = 0; k < order + 1; k++ ) {
C[ k ][ 0 ] = C[ k ][ 1 ] = c[ k ];
}
}
for( k = 0; k < order; k++ ) {
/* Check that we won't be getting an unstable rc, otherwise stop here. */
if (silk_abs_int32(C[ k + 1 ][ 0 ]) >= C[ 0 ][ 1 ]) {
if ( C[ k + 1 ][ 0 ] > 0 ) {
rc_Q15[ k ] = -SILK_FIX_CONST( .99f, 15 );
} else {
rc_Q15[ k ] = SILK_FIX_CONST( .99f, 15 );
}
k++;
break;
}
/* Get reflection coefficient */
rc_tmp_Q15 = -silk_DIV32_16( C[ k + 1 ][ 0 ], silk_max_32( silk_RSHIFT( C[ 0 ][ 1 ], 15 ), 1 ) );
/* Clip (shouldn't happen for properly conditioned inputs) */
rc_tmp_Q15 = silk_SAT16( rc_tmp_Q15 );
/* Store */
rc_Q15[ k ] = (opus_int16)rc_tmp_Q15;
/* Update correlations */
for( n = 0; n < order - k; n++ ) {
Ctmp1 = C[ n + k + 1 ][ 0 ];
Ctmp2 = C[ n ][ 1 ];
C[ n + k + 1 ][ 0 ] = silk_SMLAWB( Ctmp1, silk_LSHIFT( Ctmp2, 1 ), rc_tmp_Q15 );
C[ n ][ 1 ] = silk_SMLAWB( Ctmp2, silk_LSHIFT( Ctmp1, 1 ), rc_tmp_Q15 );
}
}
for(; k < order; k++ ) {
rc_Q15[ k ] = 0;
}
/* return residual energy */
return silk_max_32( 1, C[ 0 ][ 1 ] );
}
/* Find pitch lags */
void silk_find_pitch_lags_FIX(
silk_encoder_state_FIX *psEnc, /* I/O encoder state */
silk_encoder_control_FIX *psEncCtrl, /* I/O encoder control */
opus_int16 res[], /* O residual */
const opus_int16 x[], /* I Speech signal */
int arch /* I Run-time architecture */
)
{
opus_int buf_len, i, scale;
opus_int32 thrhld_Q13, res_nrg;
const opus_int16 *x_buf, *x_buf_ptr;
VARDECL( opus_int16, Wsig );
opus_int16 *Wsig_ptr;
opus_int32 auto_corr[ MAX_FIND_PITCH_LPC_ORDER + 1 ];
opus_int16 rc_Q15[ MAX_FIND_PITCH_LPC_ORDER ];
opus_int32 A_Q24[ MAX_FIND_PITCH_LPC_ORDER ];
opus_int16 A_Q12[ MAX_FIND_PITCH_LPC_ORDER ];
SAVE_STACK;
/******************************************/
/* Set up buffer lengths etc based on Fs */
/******************************************/
buf_len = psEnc->sCmn.la_pitch + psEnc->sCmn.frame_length + psEnc->sCmn.ltp_mem_length;
/* Safety check */
silk_assert( buf_len >= psEnc->sCmn.pitch_LPC_win_length );
x_buf = x - psEnc->sCmn.ltp_mem_length;
/*************************************/
/* Estimate LPC AR coefficients */
/*************************************/
/* Calculate windowed signal */
ALLOC( Wsig, psEnc->sCmn.pitch_LPC_win_length, opus_int16 );
/* First LA_LTP samples */
x_buf_ptr = x_buf + buf_len - psEnc->sCmn.pitch_LPC_win_length;
Wsig_ptr = Wsig;
silk_apply_sine_window( Wsig_ptr, x_buf_ptr, 1, psEnc->sCmn.la_pitch );
/* Middle un - windowed samples */
Wsig_ptr += psEnc->sCmn.la_pitch;
x_buf_ptr += psEnc->sCmn.la_pitch;
silk_memcpy( Wsig_ptr, x_buf_ptr, ( psEnc->sCmn.pitch_LPC_win_length - silk_LSHIFT( psEnc->sCmn.la_pitch, 1 ) ) * sizeof( opus_int16 ) );
/* Last LA_LTP samples */
Wsig_ptr += psEnc->sCmn.pitch_LPC_win_length - silk_LSHIFT( psEnc->sCmn.la_pitch, 1 );
x_buf_ptr += psEnc->sCmn.pitch_LPC_win_length - silk_LSHIFT( psEnc->sCmn.la_pitch, 1 );
silk_apply_sine_window( Wsig_ptr, x_buf_ptr, 2, psEnc->sCmn.la_pitch );
/* Calculate autocorrelation sequence */
silk_autocorr( auto_corr, &scale, Wsig, psEnc->sCmn.pitch_LPC_win_length, psEnc->sCmn.pitchEstimationLPCOrder + 1, arch );
/* Add white noise, as fraction of energy */
auto_corr[ 0 ] = silk_SMLAWB( auto_corr[ 0 ], auto_corr[ 0 ], SILK_FIX_CONST( FIND_PITCH_WHITE_NOISE_FRACTION, 16 ) ) + 1;
/* Calculate the reflection coefficients using schur */
res_nrg = silk_schur( rc_Q15, auto_corr, psEnc->sCmn.pitchEstimationLPCOrder );
/* Prediction gain */
psEncCtrl->predGain_Q16 = silk_DIV32_varQ( auto_corr[ 0 ], silk_max_int( res_nrg, 1 ), 16 );
/* Convert reflection coefficients to prediction coefficients */
silk_k2a( A_Q24, rc_Q15, psEnc->sCmn.pitchEstimationLPCOrder );
/* Convert From 32 bit Q24 to 16 bit Q12 coefs */
for( i = 0; i < psEnc->sCmn.pitchEstimationLPCOrder; i++ ) {
A_Q12[ i ] = (opus_int16)silk_SAT16( silk_RSHIFT( A_Q24[ i ], 12 ) );
}
/* Do BWE */
silk_bwexpander( A_Q12, psEnc->sCmn.pitchEstimationLPCOrder, SILK_FIX_CONST( FIND_PITCH_BANDWIDTH_EXPANSION, 16 ) );
/*****************************************/
/* LPC analysis filtering */
/*****************************************/
silk_LPC_analysis_filter( res, x_buf, A_Q12, buf_len, psEnc->sCmn.pitchEstimationLPCOrder );
if( psEnc->sCmn.indices.signalType != TYPE_NO_VOICE_ACTIVITY && psEnc->sCmn.first_frame_after_reset == 0 ) {
/* Threshold for pitch estimator */
thrhld_Q13 = SILK_FIX_CONST( 0.6, 13 );
thrhld_Q13 = silk_SMLABB( thrhld_Q13, SILK_FIX_CONST( -0.004, 13 ), psEnc->sCmn.pitchEstimationLPCOrder );
thrhld_Q13 = silk_SMLAWB( thrhld_Q13, SILK_FIX_CONST( -0.1, 21 ), psEnc->sCmn.speech_activity_Q8 );
thrhld_Q13 = silk_SMLABB( thrhld_Q13, SILK_FIX_CONST( -0.15, 13 ), silk_RSHIFT( psEnc->sCmn.prevSignalType, 1 ) );
thrhld_Q13 = silk_SMLAWB( thrhld_Q13, SILK_FIX_CONST( -0.1, 14 ), psEnc->sCmn.input_tilt_Q15 );
thrhld_Q13 = silk_SAT16( thrhld_Q13 );
/*****************************************/
/* Call pitch estimator */
/*****************************************/
if( silk_pitch_analysis_core( res, psEncCtrl->pitchL, &psEnc->sCmn.indices.lagIndex, &psEnc->sCmn.indices.contourIndex,
&psEnc->LTPCorr_Q15, psEnc->sCmn.prevLag, psEnc->sCmn.pitchEstimationThreshold_Q16,
(opus_int)thrhld_Q13, psEnc->sCmn.fs_kHz, psEnc->sCmn.pitchEstimationComplexity, psEnc->sCmn.nb_subfr,
psEnc->sCmn.arch) == 0 )
{
psEnc->sCmn.indices.signalType = TYPE_VOICED;
} else {
psEnc->sCmn.indices.signalType = TYPE_UNVOICED;
}
} else {
silk_memset( psEncCtrl->pitchL, 0, sizeof( psEncCtrl->pitchL ) );
psEnc->sCmn.indices.lagIndex = 0;
psEnc->sCmn.indices.contourIndex = 0;
psEnc->LTPCorr_Q15 = 0;
}
RESTORE_STACK;
}
static OPUS_INLINE void silk_PLC_conceal(
silk_decoder_state *psDec, /* I/O Decoder state */
silk_decoder_control *psDecCtrl, /* I/O Decoder control */
opus_int16 frame[], /* O LPC residual signal */
int arch /* I Run-time architecture */
)
{
opus_int i, j, k;
opus_int lag, idx, sLTP_buf_idx, shift1, shift2;
opus_int32 rand_seed, harm_Gain_Q15, rand_Gain_Q15, inv_gain_Q30;
opus_int32 energy1, energy2, *rand_ptr, *pred_lag_ptr;
opus_int32 LPC_pred_Q10, LTP_pred_Q12;
opus_int16 rand_scale_Q14;
opus_int16 *B_Q14;
opus_int32 *sLPC_Q14_ptr;
opus_int16 A_Q12[ MAX_LPC_ORDER ];
#ifdef SMALL_FOOTPRINT
opus_int16 *sLTP;
#else
VARDECL( opus_int16, sLTP );
#endif
VARDECL( opus_int32, sLTP_Q14 );
silk_PLC_struct *psPLC = &psDec->sPLC;
opus_int32 prevGain_Q10[2];
SAVE_STACK;
ALLOC( sLTP_Q14, psDec->ltp_mem_length + psDec->frame_length, opus_int32 );
#ifdef SMALL_FOOTPRINT
/* Ugly hack that breaks aliasing rules to save stack: put sLTP at the very end of sLTP_Q14. */
sLTP = ((opus_int16*)&sLTP_Q14[psDec->ltp_mem_length + psDec->frame_length])-psDec->ltp_mem_length;
#else
ALLOC( sLTP, psDec->ltp_mem_length, opus_int16 );
#endif
prevGain_Q10[0] = silk_RSHIFT( psPLC->prevGain_Q16[ 0 ], 6);
prevGain_Q10[1] = silk_RSHIFT( psPLC->prevGain_Q16[ 1 ], 6);
if( psDec->first_frame_after_reset ) {
silk_memset( psPLC->prevLPC_Q12, 0, sizeof( psPLC->prevLPC_Q12 ) );
}
silk_PLC_energy(&energy1, &shift1, &energy2, &shift2, psDec->exc_Q14, prevGain_Q10, psDec->subfr_length, psDec->nb_subfr);
if( silk_RSHIFT( energy1, shift2 ) < silk_RSHIFT( energy2, shift1 ) ) {
/* First sub-frame has lowest energy */
rand_ptr = &psDec->exc_Q14[ silk_max_int( 0, ( psPLC->nb_subfr - 1 ) * psPLC->subfr_length - RAND_BUF_SIZE ) ];
} else {
/* Second sub-frame has lowest energy */
rand_ptr = &psDec->exc_Q14[ silk_max_int( 0, psPLC->nb_subfr * psPLC->subfr_length - RAND_BUF_SIZE ) ];
}
/* Set up Gain to random noise component */
B_Q14 = psPLC->LTPCoef_Q14;
rand_scale_Q14 = psPLC->randScale_Q14;
/* Set up attenuation gains */
harm_Gain_Q15 = HARM_ATT_Q15[ silk_min_int( NB_ATT - 1, psDec->lossCnt ) ];
if( psDec->prevSignalType == TYPE_VOICED ) {
rand_Gain_Q15 = PLC_RAND_ATTENUATE_V_Q15[ silk_min_int( NB_ATT - 1, psDec->lossCnt ) ];
} else {
rand_Gain_Q15 = PLC_RAND_ATTENUATE_UV_Q15[ silk_min_int( NB_ATT - 1, psDec->lossCnt ) ];
}
/* LPC concealment. Apply BWE to previous LPC */
silk_bwexpander( psPLC->prevLPC_Q12, psDec->LPC_order, SILK_FIX_CONST( BWE_COEF, 16 ) );
/* Preload LPC coeficients to array on stack. Gives small performance gain */
silk_memcpy( A_Q12, psPLC->prevLPC_Q12, psDec->LPC_order * sizeof( opus_int16 ) );
/* First Lost frame */
if( psDec->lossCnt == 0 ) {
rand_scale_Q14 = 1 << 14;
/* Reduce random noise Gain for voiced frames */
if( psDec->prevSignalType == TYPE_VOICED ) {
for( i = 0; i < LTP_ORDER; i++ ) {
rand_scale_Q14 -= B_Q14[ i ];
}
rand_scale_Q14 = silk_max_16( 3277, rand_scale_Q14 ); /* 0.2 */
rand_scale_Q14 = (opus_int16)silk_RSHIFT( silk_SMULBB( rand_scale_Q14, psPLC->prevLTP_scale_Q14 ), 14 );
} else {
/* Reduce random noise for unvoiced frames with high LPC gain */
opus_int32 invGain_Q30, down_scale_Q30;
invGain_Q30 = silk_LPC_inverse_pred_gain( psPLC->prevLPC_Q12, psDec->LPC_order, arch );
down_scale_Q30 = silk_min_32( silk_RSHIFT( (opus_int32)1 << 30, LOG2_INV_LPC_GAIN_HIGH_THRES ), invGain_Q30 );
down_scale_Q30 = silk_max_32( silk_RSHIFT( (opus_int32)1 << 30, LOG2_INV_LPC_GAIN_LOW_THRES ), down_scale_Q30 );
down_scale_Q30 = silk_LSHIFT( down_scale_Q30, LOG2_INV_LPC_GAIN_HIGH_THRES );
rand_Gain_Q15 = silk_RSHIFT( silk_SMULWB( down_scale_Q30, rand_Gain_Q15 ), 14 );
}
}
rand_seed = psPLC->rand_seed;
lag = silk_RSHIFT_ROUND( psPLC->pitchL_Q8, 8 );
sLTP_buf_idx = psDec->ltp_mem_length;
/* Rewhiten LTP state */
idx = psDec->ltp_mem_length - lag - psDec->LPC_order - LTP_ORDER / 2;
silk_assert( idx > 0 );
silk_LPC_analysis_filter( &sLTP[ idx ], &psDec->outBuf[ idx ], A_Q12, psDec->ltp_mem_length - idx, psDec->LPC_order, arch );
/* Scale LTP state */
inv_gain_Q30 = silk_INVERSE32_varQ( psPLC->prevGain_Q16[ 1 ], 46 );
inv_gain_Q30 = silk_min( inv_gain_Q30, silk_int32_MAX >> 1 );
for( i = idx + psDec->LPC_order; i < psDec->ltp_mem_length; i++ ) {
sLTP_Q14[ i ] = silk_SMULWB( inv_gain_Q30, sLTP[ i ] );
}
/***************************/
/* LTP synthesis filtering */
/***************************/
for( k = 0; k < psDec->nb_subfr; k++ ) {
/* Set up pointer */
pred_lag_ptr = &sLTP_Q14[ sLTP_buf_idx - lag + LTP_ORDER / 2 ];
for( i = 0; i < psDec->subfr_length; i++ ) {
/* Unrolled loop */
/* Avoids introducing a bias because silk_SMLAWB() always rounds to -inf */
LTP_pred_Q12 = 2;
LTP_pred_Q12 = silk_SMLAWB( LTP_pred_Q12, pred_lag_ptr[ 0 ], B_Q14[ 0 ] );
LTP_pred_Q12 = silk_SMLAWB( LTP_pred_Q12, pred_lag_ptr[ -1 ], B_Q14[ 1 ] );
LTP_pred_Q12 = silk_SMLAWB( LTP_pred_Q12, pred_lag_ptr[ -2 ], B_Q14[ 2 ] );
LTP_pred_Q12 = silk_SMLAWB( LTP_pred_Q12, pred_lag_ptr[ -3 ], B_Q14[ 3 ] );
LTP_pred_Q12 = silk_SMLAWB( LTP_pred_Q12, pred_lag_ptr[ -4 ], B_Q14[ 4 ] );
pred_lag_ptr++;
/* Generate LPC excitation */
rand_seed = silk_RAND( rand_seed );
idx = silk_RSHIFT( rand_seed, 25 ) & RAND_BUF_MASK;
sLTP_Q14[ sLTP_buf_idx ] = silk_LSHIFT32( silk_SMLAWB( LTP_pred_Q12, rand_ptr[ idx ], rand_scale_Q14 ), 2 );
sLTP_buf_idx++;
}
/* Gradually reduce LTP gain */
for( j = 0; j < LTP_ORDER; j++ ) {
B_Q14[ j ] = silk_RSHIFT( silk_SMULBB( harm_Gain_Q15, B_Q14[ j ] ), 15 );
}
if ( psDec->indices.signalType != TYPE_NO_VOICE_ACTIVITY ) {
/* Gradually reduce excitation gain */
rand_scale_Q14 = silk_RSHIFT( silk_SMULBB( rand_scale_Q14, rand_Gain_Q15 ), 15 );
}
/* Slowly increase pitch lag */
psPLC->pitchL_Q8 = silk_SMLAWB( psPLC->pitchL_Q8, psPLC->pitchL_Q8, PITCH_DRIFT_FAC_Q16 );
psPLC->pitchL_Q8 = silk_min_32( psPLC->pitchL_Q8, silk_LSHIFT( silk_SMULBB( MAX_PITCH_LAG_MS, psDec->fs_kHz ), 8 ) );
lag = silk_RSHIFT_ROUND( psPLC->pitchL_Q8, 8 );
}
/***************************/
/* LPC synthesis filtering */
/***************************/
sLPC_Q14_ptr = &sLTP_Q14[ psDec->ltp_mem_length - MAX_LPC_ORDER ];
/* Copy LPC state */
silk_memcpy( sLPC_Q14_ptr, psDec->sLPC_Q14_buf, MAX_LPC_ORDER * sizeof( opus_int32 ) );
silk_assert( psDec->LPC_order >= 10 ); /* check that unrolling works */
for( i = 0; i < psDec->frame_length; i++ ) {
/* partly unrolled */
/* Avoids introducing a bias because silk_SMLAWB() always rounds to -inf */
LPC_pred_Q10 = silk_RSHIFT( psDec->LPC_order, 1 );
LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14_ptr[ MAX_LPC_ORDER + i - 1 ], A_Q12[ 0 ] );
LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14_ptr[ MAX_LPC_ORDER + i - 2 ], A_Q12[ 1 ] );
LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14_ptr[ MAX_LPC_ORDER + i - 3 ], A_Q12[ 2 ] );
LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14_ptr[ MAX_LPC_ORDER + i - 4 ], A_Q12[ 3 ] );
LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14_ptr[ MAX_LPC_ORDER + i - 5 ], A_Q12[ 4 ] );
LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14_ptr[ MAX_LPC_ORDER + i - 6 ], A_Q12[ 5 ] );
LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14_ptr[ MAX_LPC_ORDER + i - 7 ], A_Q12[ 6 ] );
LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14_ptr[ MAX_LPC_ORDER + i - 8 ], A_Q12[ 7 ] );
LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14_ptr[ MAX_LPC_ORDER + i - 9 ], A_Q12[ 8 ] );
LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14_ptr[ MAX_LPC_ORDER + i - 10 ], A_Q12[ 9 ] );
for( j = 10; j < psDec->LPC_order; j++ ) {
LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14_ptr[ MAX_LPC_ORDER + i - j - 1 ], A_Q12[ j ] );
}
/* Add prediction to LPC excitation */
sLPC_Q14_ptr[ MAX_LPC_ORDER + i ] = silk_ADD_SAT32( sLPC_Q14_ptr[ MAX_LPC_ORDER + i ],
silk_LSHIFT_SAT32( LPC_pred_Q10, 4 ));
/* Scale with Gain */
frame[ i ] = (opus_int16)silk_SAT16( silk_SAT16( silk_RSHIFT_ROUND( silk_SMULWW( sLPC_Q14_ptr[ MAX_LPC_ORDER + i ], prevGain_Q10[ 1 ] ), 8 ) ) );
}
/* Save LPC state */
silk_memcpy( psDec->sLPC_Q14_buf, &sLPC_Q14_ptr[ psDec->frame_length ], MAX_LPC_ORDER * sizeof( opus_int32 ) );
/**************************************/
/* Update states */
/**************************************/
psPLC->rand_seed = rand_seed;
psPLC->randScale_Q14 = rand_scale_Q14;
for( i = 0; i < MAX_NB_SUBFR; i++ ) {
psDecCtrl->pitchL[ i ] = lag;
}
RESTORE_STACK;
}
/* Autocorrelations for a warped frequency axis */
void silk_warped_autocorrelation_FIX(int32_t * corr, /* O Result [order + 1] */
int * scale, /* O Scaling of the correlation vector */
const int16_t * input, /* I Input data to correlate */
const int warping_Q16, /* I Warping coefficient */
const int length, /* I Length of input */
const int order /* I Correlation order (even) */
)
{
int n, i, lsh;
int32_t tmp1_QS, tmp2_QS;
int32_t state_QS[MAX_SHAPE_LPC_ORDER + 1] = { 0 };
int64_t corr_QC[MAX_SHAPE_LPC_ORDER + 1] = { 0 };
/* Order must be even */
assert((order & 1) == 0);
assert(2 * QS - QC >= 0);
/* Loop over samples */
for (n = 0; n < length; n++) {
tmp1_QS = silk_LSHIFT32((int32_t) input[n], QS);
/* Loop over allpass sections */
for (i = 0; i < order; i += 2) {
/* Output of allpass section */
tmp2_QS =
silk_SMLAWB(state_QS[i], state_QS[i + 1] - tmp1_QS,
warping_Q16);
state_QS[i] = tmp1_QS;
corr_QC[i] +=
silk_RSHIFT64(silk_SMULL(tmp1_QS, state_QS[0]),
2 * QS - QC);
/* Output of allpass section */
tmp1_QS =
silk_SMLAWB(state_QS[i + 1],
state_QS[i + 2] - tmp2_QS, warping_Q16);
state_QS[i + 1] = tmp2_QS;
corr_QC[i + 1] +=
silk_RSHIFT64(silk_SMULL(tmp2_QS, state_QS[0]),
2 * QS - QC);
}
state_QS[order] = tmp1_QS;
corr_QC[order] +=
silk_RSHIFT64(silk_SMULL(tmp1_QS, state_QS[0]),
2 * QS - QC);
}
lsh = silk_CLZ64(corr_QC[0]) - 35;
lsh = silk_LIMIT(lsh, -12 - QC, 30 - QC);
*scale = -(QC + lsh);
assert(*scale >= -30 && *scale <= 12);
if (lsh >= 0) {
for (i = 0; i < order + 1; i++) {
corr[i] =
(int32_t)
silk_CHECK_FIT32(silk_LSHIFT64(corr_QC[i], lsh));
}
} else {
for (i = 0; i < order + 1; i++) {
corr[i] =
(int32_t)
silk_CHECK_FIT32(silk_RSHIFT64(corr_QC[i], -lsh));
}
}
assert(corr_QC[0] >= 0); /* If breaking, decrease QC */
}
