void silk_NLSF_decode(
opus_int16 *pNLSF_Q15, /* O Quantized NLSF vector [ LPC_ORDER ] */
opus_int8 *NLSFIndices, /* I Codebook path vector [ LPC_ORDER + 1 ] */
const silk_NLSF_CB_struct *psNLSF_CB /* I Codebook object */
)
{
opus_int i;
opus_uint8 pred_Q8[ MAX_LPC_ORDER ];
opus_int16 ec_ix[ MAX_LPC_ORDER ];
opus_int16 res_Q10[ MAX_LPC_ORDER ];
opus_int32 NLSF_Q15_tmp;
const opus_uint8 *pCB_element;
const opus_int16 *pCB_Wght_Q9;
/* Unpack entropy table indices and predictor for current CB1 index */
silk_NLSF_unpack( ec_ix, pred_Q8, psNLSF_CB, NLSFIndices[ 0 ] );
/* Predictive residual dequantizer */
silk_NLSF_residual_dequant( res_Q10, &NLSFIndices[ 1 ], pred_Q8, psNLSF_CB->quantStepSize_Q16, psNLSF_CB->order );
/* Apply inverse square-rooted weights to first stage and add to output */
pCB_element = &psNLSF_CB->CB1_NLSF_Q8[ NLSFIndices[ 0 ] * psNLSF_CB->order ];
pCB_Wght_Q9 = &psNLSF_CB->CB1_Wght_Q9[ NLSFIndices[ 0 ] * psNLSF_CB->order ];
for( i = 0; i < psNLSF_CB->order; i++ ) {
NLSF_Q15_tmp = silk_ADD_LSHIFT32( silk_DIV32_16( silk_LSHIFT( (opus_int32)res_Q10[ i ], 14 ), pCB_Wght_Q9[ i ] ), (opus_int16)pCB_element[ i ], 7 );
pNLSF_Q15[ i ] = (opus_int16)silk_LIMIT( NLSF_Q15_tmp, 0, 32767 );
}
/* NLSF stabilization */
silk_NLSF_stabilize( pNLSF_Q15, psNLSF_CB->deltaMin_Q15, psNLSF_CB->order );
}
/* Find least-squares prediction gain for one signal based on another and quantize it */
int32_t silk_stereo_find_predictor( /* O Returns predictor in Q13 */
int32_t * ratio_Q14, /* O Ratio of residual and mid energies */
const int16_t x[], /* I Basis signal */
const int16_t y[], /* I Target signal */
int32_t mid_res_amp_Q0[], /* I/O Smoothed mid, residual norms */
int length, /* I Number of samples */
int smooth_coef_Q16 /* I Smoothing coefficient */
)
{
int scale, scale1, scale2;
int32_t nrgx, nrgy, corr, pred_Q13, pred2_Q10;
/* Find predictor */
silk_sum_sqr_shift(&nrgx, &scale1, x, length);
silk_sum_sqr_shift(&nrgy, &scale2, y, length);
scale = silk_max_int(scale1, scale2);
scale = scale + (scale & 1); /* make even */
nrgy = silk_RSHIFT32(nrgy, scale - scale2);
nrgx = silk_RSHIFT32(nrgx, scale - scale1);
nrgx = silk_max_int(nrgx, 1);
corr = silk_inner_prod_aligned_scale(x, y, scale, length);
pred_Q13 = silk_DIV32_varQ(corr, nrgx, 13);
pred_Q13 = silk_LIMIT(pred_Q13, -(1 << 14), 1 << 14);
pred2_Q10 = silk_SMULWB(pred_Q13, pred_Q13);
/* Faster update for signals with large prediction parameters */
smooth_coef_Q16 =
(int) silk_max_int(smooth_coef_Q16, silk_abs(pred2_Q10));
/* Smoothed mid and residual norms */
assert(smooth_coef_Q16 < 32768);
scale = silk_RSHIFT(scale, 1);
mid_res_amp_Q0[0] =
silk_SMLAWB(mid_res_amp_Q0[0],
silk_LSHIFT(silk_SQRT_APPROX(nrgx),
scale) - mid_res_amp_Q0[0],
smooth_coef_Q16);
/* Residual energy = nrgy - 2 * pred * corr + pred^2 * nrgx */
nrgy = silk_SUB_LSHIFT32(nrgy, silk_SMULWB(corr, pred_Q13), 3 + 1);
nrgy = silk_ADD_LSHIFT32(nrgy, silk_SMULWB(nrgx, pred2_Q10), 6);
mid_res_amp_Q0[1] =
silk_SMLAWB(mid_res_amp_Q0[1],
silk_LSHIFT(silk_SQRT_APPROX(nrgy),
scale) - mid_res_amp_Q0[1],
smooth_coef_Q16);
/* Ratio of smoothed residual and mid norms */
*ratio_Q14 =
silk_DIV32_varQ(mid_res_amp_Q0[1], silk_max(mid_res_amp_Q0[0], 1),
14);
*ratio_Q14 = silk_LIMIT(*ratio_Q14, 0, 32767);
return pred_Q13;
}
/* Compute unique identifier of gain indices vector */
opus_int32 silk_gains_ID( /* O returns unique identifier of gains */
const opus_int8 ind[ MAX_NB_SUBFR ], /* I gain indices */
const opus_int nb_subfr /* I number of subframes */
)
{
opus_int k;
opus_int32 gainsID;
gainsID = 0;
for( k = 0; k < nb_subfr; k++ ) {
gainsID = silk_ADD_LSHIFT32( ind[ k ], gainsID, 8 );
}
return gainsID;
}
/* Second order AR filter with single delay elements */
void silk_resampler_private_AR2(int32_t S[], /* I/O State vector [ 2 ] */
int32_t out_Q8[], /* O Output signal */
const int16_t in[], /* I Input signal */
const int16_t A_Q14[], /* I AR coefficients, Q14 */
int32_t len /* I Signal length */
)
{
int32_t k;
int32_t out32;
for (k = 0; k < len; k++) {
out32 = silk_ADD_LSHIFT32(S[0], (int32_t) in[k], 8);
out_Q8[k] = out32;
out32 = silk_LSHIFT(out32, 2);
S[0] = silk_SMLAWB(S[1], out32, A_Q14[0]);
S[1] = silk_SMULWB(out32, A_Q14[1]);
}
}
/* Second order AR filter with single delay elements */
void silk_resampler_private_AR2(
opus_int32 S[], /* I/O State vector [ 2 ] */
opus_int32 out_Q8[], /* O Output signal */
const opus_int16 in[], /* I Input signal */
const opus_int16 A_Q14[], /* I AR coefficients, Q14 */
opus_int32 len /* I Signal length */
)
{
opus_int32 k;
opus_int32 out32;
for(k = 0; k < len; k++) {
out32 = silk_ADD_LSHIFT32(S[ 0 ], (opus_int32)in[ k ], 8);
out_Q8[ k ] = out32;
out32 = silk_LSHIFT(out32, 2);
S[ 0 ] = silk_SMLAWB(S[ 1 ], out32, A_Q14[ 0 ]);
S[ 1 ] = silk_SMULWB(out32, A_Q14[ 1 ]);
}
}
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;
}
}
/* 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 {
/* 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;
}
}
/* Compute reflection coefficients from input signal */
void silk_burg_modified(
opus_int32 *res_nrg, /* O Residual energy */
opus_int *res_nrg_Q, /* O Residual energy Q value */
opus_int32 A_Q16[], /* O Prediction coefficients (length order) */
const opus_int16 x[], /* I Input signal, length: nb_subfr * ( D + subfr_length ) */
const opus_int subfr_length, /* I Input signal subframe length (incl. D preceeding samples) */
const opus_int nb_subfr, /* I Number of subframes stacked in x */
const opus_int32 WhiteNoiseFrac_Q32, /* I Fraction added to zero-lag autocorrelation */
const opus_int D /* I Order */
)
{
opus_int k, n, s, lz, rshifts, rshifts_extra;
opus_int32 C0, num, nrg, rc_Q31, Atmp_QA, Atmp1, tmp1, tmp2, x1, x2;
const opus_int16 *x_ptr;
opus_int32 C_first_row[ SILK_MAX_ORDER_LPC ];
opus_int32 C_last_row[ SILK_MAX_ORDER_LPC ];
opus_int32 Af_QA[ SILK_MAX_ORDER_LPC ];
opus_int32 CAf[ SILK_MAX_ORDER_LPC + 1 ];
opus_int32 CAb[ SILK_MAX_ORDER_LPC + 1 ];
silk_assert( subfr_length * nb_subfr <= MAX_FRAME_SIZE );
silk_assert( nb_subfr <= MAX_NB_SUBFR );
/* Compute autocorrelations, added over subframes */
silk_sum_sqr_shift( &C0, &rshifts, x, nb_subfr * subfr_length );
if( rshifts > MAX_RSHIFTS ) {
C0 = silk_LSHIFT32( C0, rshifts - MAX_RSHIFTS );
silk_assert( C0 > 0 );
rshifts = MAX_RSHIFTS;
} else {
lz = silk_CLZ32( C0 ) - 1;
rshifts_extra = N_BITS_HEAD_ROOM - lz;
if( rshifts_extra > 0 ) {
rshifts_extra = silk_min( rshifts_extra, MAX_RSHIFTS - rshifts );
C0 = silk_RSHIFT32( C0, rshifts_extra );
} else {
rshifts_extra = silk_max( rshifts_extra, MIN_RSHIFTS - rshifts );
C0 = silk_LSHIFT32( C0, -rshifts_extra );
}
rshifts += rshifts_extra;
}
silk_memset( C_first_row, 0, SILK_MAX_ORDER_LPC * sizeof( opus_int32 ) );
if( rshifts > 0 ) {
for( s = 0; s < nb_subfr; s++ ) {
x_ptr = x + s * subfr_length;
for( n = 1; n < D + 1; n++ ) {
C_first_row[ n - 1 ] += (opus_int32)silk_RSHIFT64(
silk_inner_prod16_aligned_64( x_ptr, x_ptr + n, subfr_length - n ), rshifts );
}
}
} else {
for( s = 0; s < nb_subfr; s++ ) {
x_ptr = x + s * subfr_length;
for( n = 1; n < D + 1; n++ ) {
C_first_row[ n - 1 ] += silk_LSHIFT32(
silk_inner_prod_aligned( x_ptr, x_ptr + n, subfr_length - n ), -rshifts );
}
}
}
silk_memcpy( C_last_row, C_first_row, SILK_MAX_ORDER_LPC * sizeof( opus_int32 ) );
/* Initialize */
CAb[ 0 ] = CAf[ 0 ] = C0 + silk_SMMUL( WhiteNoiseFrac_Q32, C0 ) + 1; /* Q(-rshifts)*/
for( n = 0; n < D; n++ ) {
/* Update first row of correlation matrix (without first element) */
/* Update last row of correlation matrix (without last element, stored in reversed order) */
/* Update C * Af */
/* Update C * flipud(Af) (stored in reversed order) */
if( rshifts > -2 ) {
for( s = 0; s < nb_subfr; s++ ) {
x_ptr = x + s * subfr_length;
x1 = -silk_LSHIFT32( (opus_int32)x_ptr[ n ], 16 - rshifts ); /* Q(16-rshifts)*/
x2 = -silk_LSHIFT32( (opus_int32)x_ptr[ subfr_length - n - 1 ], 16 - rshifts ); /* Q(16-rshifts)*/
tmp1 = silk_LSHIFT32( (opus_int32)x_ptr[ n ], QA - 16 ); /* Q(QA-16)*/
tmp2 = silk_LSHIFT32( (opus_int32)x_ptr[ subfr_length - n - 1 ], QA - 16 ); /* Q(QA-16)*/
for( k = 0; k < n; k++ ) {
C_first_row[ k ] = silk_SMLAWB( C_first_row[ k ], x1, x_ptr[ n - k - 1 ] ); /* Q( -rshifts )*/
C_last_row[ k ] = silk_SMLAWB( C_last_row[ k ], x2, x_ptr[ subfr_length - n + k ] ); /* Q( -rshifts )*/
Atmp_QA = Af_QA[ k ];
tmp1 = silk_SMLAWB( tmp1, Atmp_QA, x_ptr[ n - k - 1 ] ); /* Q(QA-16)*/
tmp2 = silk_SMLAWB( tmp2, Atmp_QA, x_ptr[ subfr_length - n + k ] ); /* Q(QA-16)*/
}
tmp1 = silk_LSHIFT32( -tmp1, 32 - QA - rshifts ); /* Q(16-rshifts)*/
tmp2 = silk_LSHIFT32( -tmp2, 32 - QA - rshifts ); /* Q(16-rshifts)*/
for( k = 0; k <= n; k++ ) {
CAf[ k ] = silk_SMLAWB( CAf[ k ], tmp1, x_ptr[ n - k ] ); /* Q( -rshift )*/
CAb[ k ] = silk_SMLAWB( CAb[ k ], tmp2, x_ptr[ subfr_length - n + k - 1 ] ); /* Q( -rshift )*/
}
}
} else {
for( s = 0; s < nb_subfr; s++ ) {
x_ptr = x + s * subfr_length;
x1 = -silk_LSHIFT32( (opus_int32)x_ptr[ n ], -rshifts ); /* Q( -rshifts )*/
x2 = -silk_LSHIFT32( (opus_int32)x_ptr[ subfr_length - n - 1 ], -rshifts ); /* Q( -rshifts )*/
tmp1 = silk_LSHIFT32( (opus_int32)x_ptr[ n ], 17 ); /* Q17*/
tmp2 = silk_LSHIFT32( (opus_int32)x_ptr[ subfr_length - n - 1 ], 17 ); /* Q17*/
for( k = 0; k < n; k++ ) {
C_first_row[ k ] = silk_MLA( C_first_row[ k ], x1, x_ptr[ n - k - 1 ] ); /* Q( -rshifts )*/
C_last_row[ k ] = silk_MLA( C_last_row[ k ], x2, x_ptr[ subfr_length - n + k ] ); /* Q( -rshifts )*/
Atmp1 = silk_RSHIFT_ROUND( Af_QA[ k ], QA - 17 ); /* Q17*/
tmp1 = silk_MLA( tmp1, x_ptr[ n - k - 1 ], Atmp1 ); /* Q17*/
tmp2 = silk_MLA( tmp2, x_ptr[ subfr_length - n + k ], Atmp1 ); /* Q17*/
}
tmp1 = -tmp1; /* Q17*/
tmp2 = -tmp2; /* Q17*/
for( k = 0; k <= n; k++ ) {
CAf[ k ] = silk_SMLAWW( CAf[ k ], tmp1,
silk_LSHIFT32( (opus_int32)x_ptr[ n - k ], -rshifts - 1 ) ); /* Q( -rshift )*/
CAb[ k ] = silk_SMLAWW( CAb[ k ], tmp2,
silk_LSHIFT32( (opus_int32)x_ptr[ subfr_length - n + k - 1 ], -rshifts - 1 ) ); /* Q( -rshift )*/
}
}
}
/* Calculate nominator and denominator for the next order reflection (parcor) coefficient */
tmp1 = C_first_row[ n ]; /* Q( -rshifts )*/
tmp2 = C_last_row[ n ]; /* Q( -rshifts )*/
num = 0; /* Q( -rshifts )*/
nrg = silk_ADD32( CAb[ 0 ], CAf[ 0 ] ); /* Q( 1-rshifts )*/
for( k = 0; k < n; k++ ) {
Atmp_QA = Af_QA[ k ];
lz = silk_CLZ32( silk_abs( Atmp_QA ) ) - 1;
lz = silk_min( 32 - QA, lz );
Atmp1 = silk_LSHIFT32( Atmp_QA, lz ); /* Q( QA + lz )*/
tmp1 = silk_ADD_LSHIFT32( tmp1, silk_SMMUL( C_last_row[ n - k - 1 ], Atmp1 ), 32 - QA - lz ); /* Q( -rshifts )*/
tmp2 = silk_ADD_LSHIFT32( tmp2, silk_SMMUL( C_first_row[ n - k - 1 ], Atmp1 ), 32 - QA - lz ); /* Q( -rshifts )*/
num = silk_ADD_LSHIFT32( num, silk_SMMUL( CAb[ n - k ], Atmp1 ), 32 - QA - lz ); /* Q( -rshifts )*/
nrg = silk_ADD_LSHIFT32( nrg, silk_SMMUL( silk_ADD32( CAb[ k + 1 ], CAf[ k + 1 ] ),
Atmp1 ), 32 - QA - lz ); /* Q( 1-rshifts )*/
}
CAf[ n + 1 ] = tmp1; /* Q( -rshifts )*/
CAb[ n + 1 ] = tmp2; /* Q( -rshifts )*/
num = silk_ADD32( num, tmp2 ); /* Q( -rshifts )*/
num = silk_LSHIFT32( -num, 1 ); /* Q( 1-rshifts )*/
/* Calculate the next order reflection (parcor) coefficient */
if( silk_abs( num ) < nrg ) {
rc_Q31 = silk_DIV32_varQ( num, nrg, 31 );
} else {
/* Negative energy or ratio too high; set remaining coefficients to zero and exit loop */
silk_memset( &Af_QA[ n ], 0, ( D - n ) * sizeof( opus_int32 ) );
silk_assert( 0 );
break;
}
/* Update the AR coefficients */
for( k = 0; k < (n + 1) >> 1; k++ ) {
tmp1 = Af_QA[ k ]; /* QA*/
tmp2 = Af_QA[ n - k - 1 ]; /* QA*/
Af_QA[ k ] = silk_ADD_LSHIFT32( tmp1, silk_SMMUL( tmp2, rc_Q31 ), 1 ); /* QA*/
Af_QA[ n - k - 1 ] = silk_ADD_LSHIFT32( tmp2, silk_SMMUL( tmp1, rc_Q31 ), 1 ); /* QA*/
}
Af_QA[ n ] = silk_RSHIFT32( rc_Q31, 31 - QA ); /* QA*/
/* Update C * Af and C * Ab */
for( k = 0; k <= n + 1; k++ ) {
tmp1 = CAf[ k ]; /* Q( -rshifts )*/
tmp2 = CAb[ n - k + 1 ]; /* Q( -rshifts )*/
CAf[ k ] = silk_ADD_LSHIFT32( tmp1, silk_SMMUL( tmp2, rc_Q31 ), 1 ); /* Q( -rshifts )*/
CAb[ n - k + 1 ] = silk_ADD_LSHIFT32( tmp2, silk_SMMUL( tmp1, rc_Q31 ), 1 ); /* Q( -rshifts )*/
}
}
/* Return residual energy */
nrg = CAf[ 0 ]; /* Q( -rshifts )*/
tmp1 = 1 << 16; /* Q16*/
for( k = 0; k < D; k++ ) {
Atmp1 = silk_RSHIFT_ROUND( Af_QA[ k ], QA - 16 ); /* Q16*/
nrg = silk_SMLAWW( nrg, CAf[ k + 1 ], Atmp1 ); /* Q( -rshifts )*/
tmp1 = silk_SMLAWW( tmp1, Atmp1, Atmp1 ); /* Q16*/
A_Q16[ k ] = -Atmp1;
}
*res_nrg = silk_SMLAWW( nrg, silk_SMMUL( WhiteNoiseFrac_Q32, C0 ), -tmp1 ); /* Q( -rshifts )*/
*res_nrg_Q = -rshifts;
}
/* Compute reflection coefficients from input signal */
void silk_burg_modified_sse4_1(
opus_int32 *res_nrg, /* O Residual energy */
opus_int *res_nrg_Q, /* O Residual energy Q value */
opus_int32 A_Q16[], /* O Prediction coefficients (length order) */
const opus_int16 x[], /* I Input signal, length: nb_subfr * (D + subfr_length) */
const opus_int32 minInvGain_Q30, /* I Inverse of max prediction gain */
const opus_int subfr_length, /* I Input signal subframe length (incl. D preceding samples) */
const opus_int nb_subfr, /* I Number of subframes stacked in x */
const opus_int D, /* I Order */
int arch /* I Run-time architecture */
)
{
opus_int k, n, s, lz, rshifts, rshifts_extra, reached_max_gain;
opus_int32 C0, num, nrg, rc_Q31, invGain_Q30, Atmp_QA, Atmp1, tmp1, tmp2, x1, x2;
const opus_int16 *x_ptr;
opus_int32 C_first_row[ SILK_MAX_ORDER_LPC ];
opus_int32 C_last_row[ SILK_MAX_ORDER_LPC ];
opus_int32 Af_QA[ SILK_MAX_ORDER_LPC ];
opus_int32 CAf[ SILK_MAX_ORDER_LPC + 1 ];
opus_int32 CAb[ SILK_MAX_ORDER_LPC + 1 ];
opus_int32 xcorr[ SILK_MAX_ORDER_LPC ];
__m128i FIRST_3210, LAST_3210, ATMP_3210, TMP1_3210, TMP2_3210, T1_3210, T2_3210, PTR_3210, SUBFR_3210, X1_3210, X2_3210;
__m128i CONST1 = _mm_set1_epi32(1);
silk_assert(subfr_length * nb_subfr <= MAX_FRAME_SIZE);
/* Compute autocorrelations, added over subframes */
silk_sum_sqr_shift(&C0, &rshifts, x, nb_subfr * subfr_length);
if(rshifts > MAX_RSHIFTS) {
C0 = silk_LSHIFT32(C0, rshifts - MAX_RSHIFTS);
silk_assert(C0 > 0);
rshifts = MAX_RSHIFTS;
} else {
lz = silk_CLZ32(C0) - 1;
rshifts_extra = N_BITS_HEAD_ROOM - lz;
if(rshifts_extra > 0) {
rshifts_extra = silk_min(rshifts_extra, MAX_RSHIFTS - rshifts);
C0 = silk_RSHIFT32(C0, rshifts_extra);
} else {
rshifts_extra = silk_max(rshifts_extra, MIN_RSHIFTS - rshifts);
C0 = silk_LSHIFT32(C0, -rshifts_extra);
}
rshifts += rshifts_extra;
}
CAb[ 0 ] = CAf[ 0 ] = C0 + silk_SMMUL(SILK_FIX_CONST(FIND_LPC_COND_FAC, 32), C0) + 1; /* Q(-rshifts) */
silk_memset(C_first_row, 0, SILK_MAX_ORDER_LPC * sizeof(opus_int32));
if(rshifts > 0) {
for(s = 0; s < nb_subfr; s++) {
x_ptr = x + s * subfr_length;
for(n = 1; n < D + 1; n++) {
C_first_row[ n - 1 ] += (opus_int32)silk_RSHIFT64(
silk_inner_prod16_aligned_64(x_ptr, x_ptr + n, subfr_length - n, arch), rshifts);
}
}
} else {
for(s = 0; s < nb_subfr; s++) {
int i;
opus_int32 d;
x_ptr = x + s * subfr_length;
celt_pitch_xcorr(x_ptr, x_ptr + 1, xcorr, subfr_length - D, D, arch);
for(n = 1; n < D + 1; n++) {
for (i = n + subfr_length - D, d = 0; i < subfr_length; i++)
d = MAC16_16(d, x_ptr[ i ], x_ptr[ i - n ]);
xcorr[ n - 1 ] += d;
}
for(n = 1; n < D + 1; n++) {
C_first_row[ n - 1 ] += silk_LSHIFT32(xcorr[ n - 1 ], -rshifts);
}
}
}
silk_memcpy(C_last_row, C_first_row, SILK_MAX_ORDER_LPC * sizeof(opus_int32));
/* Initialize */
CAb[ 0 ] = CAf[ 0 ] = C0 + silk_SMMUL(SILK_FIX_CONST(FIND_LPC_COND_FAC, 32), C0) + 1; /* Q(-rshifts) */
invGain_Q30 = (opus_int32)1 << 30;
reached_max_gain = 0;
for(n = 0; n < D; n++) {
/* Update first row of correlation matrix (without first element) */
/* Update last row of correlation matrix (without last element, stored in reversed order) */
/* Update C * Af */
/* Update C * flipud(Af) (stored in reversed order) */
if(rshifts > -2) {
for(s = 0; s < nb_subfr; s++) {
x_ptr = x + s * subfr_length;
x1 = -silk_LSHIFT32((opus_int32)x_ptr[ n ], 16 - rshifts); /* Q(16-rshifts) */
x2 = -silk_LSHIFT32((opus_int32)x_ptr[ subfr_length - n - 1 ], 16 - rshifts); /* Q(16-rshifts) */
tmp1 = silk_LSHIFT32((opus_int32)x_ptr[ n ], QA - 16); /* Q(QA-16) */
tmp2 = silk_LSHIFT32((opus_int32)x_ptr[ subfr_length - n - 1 ], QA - 16); /* Q(QA-16) */
for(k = 0; k < n; k++) {
C_first_row[ k ] = silk_SMLAWB(C_first_row[ k ], x1, x_ptr[ n - k - 1 ] ); /* Q(-rshifts) */
C_last_row[ k ] = silk_SMLAWB(C_last_row[ k ], x2, x_ptr[ subfr_length - n + k ]); /* Q(-rshifts) */
Atmp_QA = Af_QA[ k ];
tmp1 = silk_SMLAWB(tmp1, Atmp_QA, x_ptr[ n - k - 1 ] ); /* Q(QA-16) */
tmp2 = silk_SMLAWB(tmp2, Atmp_QA, x_ptr[ subfr_length - n + k ]); /* Q(QA-16) */
}
tmp1 = silk_LSHIFT32(-tmp1, 32 - QA - rshifts); /* Q(16-rshifts) */
tmp2 = silk_LSHIFT32(-tmp2, 32 - QA - rshifts); /* Q(16-rshifts) */
for(k = 0; k <= n; k++) {
CAf[ k ] = silk_SMLAWB(CAf[ k ], tmp1, x_ptr[ n - k ] ); /* Q(-rshift) */
CAb[ k ] = silk_SMLAWB(CAb[ k ], tmp2, x_ptr[ subfr_length - n + k - 1 ]); /* Q(-rshift) */
}
}
} else {
for(s = 0; s < nb_subfr; s++) {
x_ptr = x + s * subfr_length;
x1 = -silk_LSHIFT32((opus_int32)x_ptr[ n ], -rshifts); /* Q(-rshifts) */
x2 = -silk_LSHIFT32((opus_int32)x_ptr[ subfr_length - n - 1 ], -rshifts); /* Q(-rshifts) */
tmp1 = silk_LSHIFT32((opus_int32)x_ptr[ n ], 17); /* Q17 */
tmp2 = silk_LSHIFT32((opus_int32)x_ptr[ subfr_length - n - 1 ], 17); /* Q17 */
X1_3210 = _mm_set1_epi32(x1);
X2_3210 = _mm_set1_epi32(x2);
TMP1_3210 = _mm_setzero_si128();
TMP2_3210 = _mm_setzero_si128();
for(k = 0; k < n - 3; k += 4) {
PTR_3210 = OP_CVTEPI16_EPI32_M64(&x_ptr[ n - k - 1 - 3 ]);
SUBFR_3210 = OP_CVTEPI16_EPI32_M64(&x_ptr[ subfr_length - n + k ]);
FIRST_3210 = _mm_loadu_si128((__m128i *)&C_first_row[ k ]);
PTR_3210 = _mm_shuffle_epi32(PTR_3210, _MM_SHUFFLE(0, 1, 2, 3));
LAST_3210 = _mm_loadu_si128((__m128i *)&C_last_row[ k ]);
ATMP_3210 = _mm_loadu_si128((__m128i *)&Af_QA[ k ]);
T1_3210 = _mm_mullo_epi32(PTR_3210, X1_3210);
T2_3210 = _mm_mullo_epi32(SUBFR_3210, X2_3210);
ATMP_3210 = _mm_srai_epi32(ATMP_3210, 7);
ATMP_3210 = _mm_add_epi32(ATMP_3210, CONST1);
ATMP_3210 = _mm_srai_epi32(ATMP_3210, 1);
FIRST_3210 = _mm_add_epi32(FIRST_3210, T1_3210);
LAST_3210 = _mm_add_epi32(LAST_3210, T2_3210);
PTR_3210 = _mm_mullo_epi32(ATMP_3210, PTR_3210);
SUBFR_3210 = _mm_mullo_epi32(ATMP_3210, SUBFR_3210);
_mm_storeu_si128((__m128i *)&C_first_row[ k ], FIRST_3210);
_mm_storeu_si128((__m128i *)&C_last_row[ k ], LAST_3210);
TMP1_3210 = _mm_add_epi32(TMP1_3210, PTR_3210);
TMP2_3210 = _mm_add_epi32(TMP2_3210, SUBFR_3210);
}
TMP1_3210 = _mm_add_epi32(TMP1_3210, _mm_unpackhi_epi64(TMP1_3210, TMP1_3210));
TMP2_3210 = _mm_add_epi32(TMP2_3210, _mm_unpackhi_epi64(TMP2_3210, TMP2_3210));
TMP1_3210 = _mm_add_epi32(TMP1_3210, _mm_shufflelo_epi16(TMP1_3210, 0x0E));
TMP2_3210 = _mm_add_epi32(TMP2_3210, _mm_shufflelo_epi16(TMP2_3210, 0x0E));
tmp1 += _mm_cvtsi128_si32(TMP1_3210);
tmp2 += _mm_cvtsi128_si32(TMP2_3210);
for(; k < n; k++) {
C_first_row[ k ] = silk_MLA(C_first_row[ k ], x1, x_ptr[ n - k - 1 ] ); /* Q(-rshifts) */
C_last_row[ k ] = silk_MLA(C_last_row[ k ], x2, x_ptr[ subfr_length - n + k ]); /* Q(-rshifts) */
Atmp1 = silk_RSHIFT_ROUND(Af_QA[ k ], QA - 17); /* Q17 */
tmp1 = silk_MLA(tmp1, x_ptr[ n - k - 1 ], Atmp1); /* Q17 */
tmp2 = silk_MLA(tmp2, x_ptr[ subfr_length - n + k ], Atmp1); /* Q17 */
}
tmp1 = -tmp1; /* Q17 */
tmp2 = -tmp2; /* Q17 */
{
__m128i xmm_tmp1, xmm_tmp2;
__m128i xmm_x_ptr_n_k_x2x0, xmm_x_ptr_n_k_x3x1;
__m128i xmm_x_ptr_sub_x2x0, xmm_x_ptr_sub_x3x1;
xmm_tmp1 = _mm_set1_epi32(tmp1);
xmm_tmp2 = _mm_set1_epi32(tmp2);
for(k = 0; k <= n - 3; k += 4) {
xmm_x_ptr_n_k_x2x0 = OP_CVTEPI16_EPI32_M64(&x_ptr[ n - k - 3 ]);
xmm_x_ptr_sub_x2x0 = OP_CVTEPI16_EPI32_M64(&x_ptr[ subfr_length - n + k - 1 ]);
xmm_x_ptr_n_k_x2x0 = _mm_shuffle_epi32(xmm_x_ptr_n_k_x2x0, _MM_SHUFFLE(0, 1, 2, 3));
xmm_x_ptr_n_k_x2x0 = _mm_slli_epi32(xmm_x_ptr_n_k_x2x0, -rshifts - 1);
xmm_x_ptr_sub_x2x0 = _mm_slli_epi32(xmm_x_ptr_sub_x2x0, -rshifts - 1);
/* equal shift right 4 bytes, xmm_x_ptr_n_k_x3x1 = _mm_srli_si128(xmm_x_ptr_n_k_x2x0, 4)*/
xmm_x_ptr_n_k_x3x1 = _mm_shuffle_epi32(xmm_x_ptr_n_k_x2x0, _MM_SHUFFLE(0, 3, 2, 1));
xmm_x_ptr_sub_x3x1 = _mm_shuffle_epi32(xmm_x_ptr_sub_x2x0, _MM_SHUFFLE(0, 3, 2, 1));
xmm_x_ptr_n_k_x2x0 = _mm_mul_epi32(xmm_x_ptr_n_k_x2x0, xmm_tmp1);
xmm_x_ptr_n_k_x3x1 = _mm_mul_epi32(xmm_x_ptr_n_k_x3x1, xmm_tmp1);
xmm_x_ptr_sub_x2x0 = _mm_mul_epi32(xmm_x_ptr_sub_x2x0, xmm_tmp2);
xmm_x_ptr_sub_x3x1 = _mm_mul_epi32(xmm_x_ptr_sub_x3x1, xmm_tmp2);
xmm_x_ptr_n_k_x2x0 = _mm_srli_epi64(xmm_x_ptr_n_k_x2x0, 16);
xmm_x_ptr_n_k_x3x1 = _mm_slli_epi64(xmm_x_ptr_n_k_x3x1, 16);
xmm_x_ptr_sub_x2x0 = _mm_srli_epi64(xmm_x_ptr_sub_x2x0, 16);
xmm_x_ptr_sub_x3x1 = _mm_slli_epi64(xmm_x_ptr_sub_x3x1, 16);
xmm_x_ptr_n_k_x2x0 = _mm_blend_epi16(xmm_x_ptr_n_k_x2x0, xmm_x_ptr_n_k_x3x1, 0xCC);
xmm_x_ptr_sub_x2x0 = _mm_blend_epi16(xmm_x_ptr_sub_x2x0, xmm_x_ptr_sub_x3x1, 0xCC);
X1_3210 = _mm_loadu_si128((__m128i *)&CAf[ k ]);
PTR_3210 = _mm_loadu_si128((__m128i *)&CAb[ k ]);
X1_3210 = _mm_add_epi32(X1_3210, xmm_x_ptr_n_k_x2x0);
PTR_3210 = _mm_add_epi32(PTR_3210, xmm_x_ptr_sub_x2x0);
_mm_storeu_si128((__m128i *)&CAf[ k ], X1_3210);
_mm_storeu_si128((__m128i *)&CAb[ k ], PTR_3210);
}
for(; k <= n; k++) {
CAf[ k ] = silk_SMLAWW(CAf[ k ], tmp1,
silk_LSHIFT32((opus_int32)x_ptr[ n - k ], -rshifts - 1)); /* Q(-rshift) */
CAb[ k ] = silk_SMLAWW(CAb[ k ], tmp2,
silk_LSHIFT32((opus_int32)x_ptr[ subfr_length - n + k - 1 ], -rshifts - 1)); /* Q(-rshift) */
}
}
}
}
/* Calculate nominator and denominator for the next order reflection (parcor) coefficient */
tmp1 = C_first_row[ n ]; /* Q(-rshifts) */
tmp2 = C_last_row[ n ]; /* Q(-rshifts) */
num = 0; /* Q(-rshifts) */
nrg = silk_ADD32(CAb[ 0 ], CAf[ 0 ]); /* Q(1-rshifts) */
for(k = 0; k < n; k++) {
Atmp_QA = Af_QA[ k ];
lz = silk_CLZ32(silk_abs(Atmp_QA)) - 1;
lz = silk_min(32 - QA, lz);
Atmp1 = silk_LSHIFT32(Atmp_QA, lz); /* Q(QA + lz) */
tmp1 = silk_ADD_LSHIFT32(tmp1, silk_SMMUL(C_last_row[ n - k - 1 ], Atmp1), 32 - QA - lz); /* Q(-rshifts) */
tmp2 = silk_ADD_LSHIFT32(tmp2, silk_SMMUL(C_first_row[ n - k - 1 ], Atmp1), 32 - QA - lz); /* Q(-rshifts) */
num = silk_ADD_LSHIFT32(num, silk_SMMUL(CAb[ n - k ], Atmp1), 32 - QA - lz); /* Q(-rshifts) */
nrg = silk_ADD_LSHIFT32(nrg, silk_SMMUL(silk_ADD32(CAb[ k + 1 ], CAf[ k + 1 ]),
Atmp1), 32 - QA - lz); /* Q(1-rshifts) */
}
CAf[ n + 1 ] = tmp1; /* Q(-rshifts) */
CAb[ n + 1 ] = tmp2; /* Q(-rshifts) */
num = silk_ADD32(num, tmp2); /* Q(-rshifts) */
num = silk_LSHIFT32(-num, 1); /* Q(1-rshifts) */
/* Calculate the next order reflection (parcor) coefficient */
if(silk_abs(num) < nrg) {
rc_Q31 = silk_DIV32_varQ(num, nrg, 31);
} else {
rc_Q31 = (num > 0) ? silk_int32_MAX : silk_int32_MIN;
}
/* Update inverse prediction gain */
tmp1 = ((opus_int32)1 << 30) - silk_SMMUL(rc_Q31, rc_Q31);
tmp1 = silk_LSHIFT(silk_SMMUL(invGain_Q30, tmp1), 2);
if(tmp1 <= minInvGain_Q30) {
/* Max prediction gain exceeded; set reflection coefficient such that max prediction gain is exactly hit */
tmp2 = ((opus_int32)1 << 30) - silk_DIV32_varQ(minInvGain_Q30, invGain_Q30, 30); /* Q30 */
rc_Q31 = silk_SQRT_APPROX(tmp2); /* Q15 */
/* Newton-Raphson iteration */
rc_Q31 = silk_RSHIFT32(rc_Q31 + silk_DIV32(tmp2, rc_Q31), 1); /* Q15 */
rc_Q31 = silk_LSHIFT32(rc_Q31, 16); /* Q31 */
if(num < 0) {
/* Ensure adjusted reflection coefficients has the original sign */
rc_Q31 = -rc_Q31;
}
invGain_Q30 = minInvGain_Q30;
reached_max_gain = 1;
} else {
invGain_Q30 = tmp1;
}
/* Update the AR coefficients */
for(k = 0; k < (n + 1) >> 1; k++) {
tmp1 = Af_QA[ k ]; /* QA */
tmp2 = Af_QA[ n - k - 1 ]; /* QA */
Af_QA[ k ] = silk_ADD_LSHIFT32(tmp1, silk_SMMUL(tmp2, rc_Q31), 1); /* QA */
Af_QA[ n - k - 1 ] = silk_ADD_LSHIFT32(tmp2, silk_SMMUL(tmp1, rc_Q31), 1); /* QA */
}
Af_QA[ n ] = silk_RSHIFT32(rc_Q31, 31 - QA); /* QA */
if(reached_max_gain) {
/* Reached max prediction gain; set remaining coefficients to zero and exit loop */
for(k = n + 1; k < D; k++) {
Af_QA[ k ] = 0;
}
break;
}
/* Update C * Af and C * Ab */
for(k = 0; k <= n + 1; k++) {
tmp1 = CAf[ k ]; /* Q(-rshifts) */
tmp2 = CAb[ n - k + 1 ]; /* Q(-rshifts) */
CAf[ k ] = silk_ADD_LSHIFT32(tmp1, silk_SMMUL(tmp2, rc_Q31), 1); /* Q(-rshifts) */
CAb[ n - k + 1 ] = silk_ADD_LSHIFT32(tmp2, silk_SMMUL(tmp1, rc_Q31), 1); /* Q(-rshifts) */
}
}
if(reached_max_gain) {
for(k = 0; k < D; k++) {
/* Scale coefficients */
A_Q16[ k ] = -silk_RSHIFT_ROUND(Af_QA[ k ], QA - 16);
}
/* Subtract energy of preceding samples from C0 */
if(rshifts > 0) {
for(s = 0; s < nb_subfr; s++) {
x_ptr = x + s * subfr_length;
C0 -= (opus_int32)silk_RSHIFT64(silk_inner_prod16_aligned_64(x_ptr, x_ptr, D, arch), rshifts);
}
} else {
for(s = 0; s < nb_subfr; s++) {
x_ptr = x + s * subfr_length;
C0 -= silk_LSHIFT32(silk_inner_prod_aligned(x_ptr, x_ptr, D, arch), -rshifts);
}
}
/* Approximate residual energy */
*res_nrg = silk_LSHIFT(silk_SMMUL(invGain_Q30, C0), 2);
*res_nrg_Q = -rshifts;
} else {
/* Return residual energy */
nrg = CAf[ 0 ]; /* Q(-rshifts) */
tmp1 = (opus_int32)1 << 16; /* Q16 */
for(k = 0; k < D; k++) {
Atmp1 = silk_RSHIFT_ROUND(Af_QA[ k ], QA - 16); /* Q16 */
nrg = silk_SMLAWW(nrg, CAf[ k + 1 ], Atmp1); /* Q(-rshifts) */
tmp1 = silk_SMLAWW(tmp1, Atmp1, Atmp1); /* Q16 */
A_Q16[ k ] = -Atmp1;
}
*res_nrg = silk_SMLAWW(nrg, silk_SMMUL(SILK_FIX_CONST(FIND_LPC_COND_FAC, 32), C0), -tmp1);/* Q(-rshifts) */
*res_nrg_Q = -rshifts;
}
}
/* Compute reflection coefficients from input signal */
void silk_burg_modified(
opus_int32 *res_nrg, /* O Residual energy */
opus_int *res_nrg_Q, /* O Residual energy Q value */
opus_int32 A_Q16[], /* O Prediction coefficients (length order) */
const opus_int16 x[], /* I Input signal, length: nb_subfr * ( D + subfr_length ) */
const opus_int32 minInvGain_Q30, /* I Inverse of max prediction gain */
const opus_int subfr_length, /* I Input signal subframe length (incl. D preceding samples) */
const opus_int nb_subfr, /* I Number of subframes stacked in x */
const opus_int D /* I Order */
)
{
opus_int k, n, s, lz, rshifts, rshifts_extra, reached_max_gain;
opus_int32 C0, num, nrg, rc_Q31, invGain_Q30, Atmp_QA, Atmp1, tmp1, tmp2, x1, x2;
const opus_int16 *x_ptr;
opus_int32 C_first_row[ SILK_MAX_ORDER_LPC ];
opus_int32 C_last_row[ SILK_MAX_ORDER_LPC ];
opus_int32 Af_QA[ SILK_MAX_ORDER_LPC ];
opus_int32 CAf[ SILK_MAX_ORDER_LPC + 1 ];
opus_int32 CAb[ SILK_MAX_ORDER_LPC + 1 ];
silk_assert( subfr_length * nb_subfr <= MAX_FRAME_SIZE );
/* Compute autocorrelations, added over subframes */
silk_sum_sqr_shift( &C0, &rshifts, x, nb_subfr * subfr_length );
if( rshifts > MAX_RSHIFTS ) {
C0 = silk_LSHIFT32( C0, rshifts - MAX_RSHIFTS );
silk_assert( C0 > 0 );
rshifts = MAX_RSHIFTS;
} else {
lz = silk_CLZ32( C0 ) - 1;
rshifts_extra = N_BITS_HEAD_ROOM - lz;
if( rshifts_extra > 0 ) {
rshifts_extra = silk_min( rshifts_extra, MAX_RSHIFTS - rshifts );
C0 = silk_RSHIFT32( C0, rshifts_extra );
} else {
rshifts_extra = silk_max( rshifts_extra, MIN_RSHIFTS - rshifts );
C0 = silk_LSHIFT32( C0, -rshifts_extra );
}
rshifts += rshifts_extra;
}
CAb[ 0 ] = CAf[ 0 ] = C0 + silk_SMMUL( SILK_FIX_CONST( FIND_LPC_COND_FAC, 32 ), C0 ) + 1; /* Q(-rshifts) */
silk_memset( C_first_row, 0, SILK_MAX_ORDER_LPC * sizeof( opus_int32 ) );
if( rshifts > 0 ) {
for( s = 0; s < nb_subfr; s++ ) {
x_ptr = x + s * subfr_length;
for( n = 1; n < D + 1; n++ ) {
C_first_row[ n - 1 ] += (opus_int32)silk_RSHIFT64(
silk_inner_prod16_aligned_64( x_ptr, x_ptr + n, subfr_length - n ), rshifts );
}
}
} else {
for( s = 0; s < nb_subfr; s++ ) {
x_ptr = x + s * subfr_length;
for( n = 1; n < D + 1; n++ ) {
C_first_row[ n - 1 ] += silk_LSHIFT32(
silk_inner_prod_aligned( x_ptr, x_ptr + n, subfr_length - n ), -rshifts );
}
}
}
silk_memcpy( C_last_row, C_first_row, SILK_MAX_ORDER_LPC * sizeof( opus_int32 ) );
/* Initialize */
CAb[ 0 ] = CAf[ 0 ] = C0 + silk_SMMUL( SILK_FIX_CONST( FIND_LPC_COND_FAC, 32 ), C0 ) + 1; /* Q(-rshifts) */
invGain_Q30 = (opus_int32)1 << 30;
reached_max_gain = 0;
for( n = 0; n < D; n++ ) {
/* Update first row of correlation matrix (without first element) */
/* Update last row of correlation matrix (without last element, stored in reversed order) */
/* Update C * Af */
/* Update C * flipud(Af) (stored in reversed order) */
if( rshifts > -2 ) {
for( s = 0; s < nb_subfr; s++ ) {
x_ptr = x + s * subfr_length;
x1 = -silk_LSHIFT32( (opus_int32)x_ptr[ n ], 16 - rshifts ); /* Q(16-rshifts) */
x2 = -silk_LSHIFT32( (opus_int32)x_ptr[ subfr_length - n - 1 ], 16 - rshifts ); /* Q(16-rshifts) */
tmp1 = silk_LSHIFT32( (opus_int32)x_ptr[ n ], QA - 16 ); /* Q(QA-16) */
tmp2 = silk_LSHIFT32( (opus_int32)x_ptr[ subfr_length - n - 1 ], QA - 16 ); /* Q(QA-16) */
for( k = 0; k < n; k++ ) {
C_first_row[ k ] = silk_SMLAWB( C_first_row[ k ], x1, x_ptr[ n - k - 1 ] ); /* Q( -rshifts ) */
C_last_row[ k ] = silk_SMLAWB( C_last_row[ k ], x2, x_ptr[ subfr_length - n + k ] ); /* Q( -rshifts ) */
Atmp_QA = Af_QA[ k ];
tmp1 = silk_SMLAWB( tmp1, Atmp_QA, x_ptr[ n - k - 1 ] ); /* Q(QA-16) */
tmp2 = silk_SMLAWB( tmp2, Atmp_QA, x_ptr[ subfr_length - n + k ] ); /* Q(QA-16) */
}
tmp1 = silk_LSHIFT32( -tmp1, 32 - QA - rshifts ); /* Q(16-rshifts) */
tmp2 = silk_LSHIFT32( -tmp2, 32 - QA - rshifts ); /* Q(16-rshifts) */
for( k = 0; k <= n; k++ ) {
CAf[ k ] = silk_SMLAWB( CAf[ k ], tmp1, x_ptr[ n - k ] ); /* Q( -rshift ) */
CAb[ k ] = silk_SMLAWB( CAb[ k ], tmp2, x_ptr[ subfr_length - n + k - 1 ] ); /* Q( -rshift ) */
}
}
} else {
for( s = 0; s < nb_subfr; s++ ) {
x_ptr = x + s * subfr_length;
x1 = -silk_LSHIFT32( (opus_int32)x_ptr[ n ], -rshifts ); /* Q( -rshifts ) */
x2 = -silk_LSHIFT32( (opus_int32)x_ptr[ subfr_length - n - 1 ], -rshifts ); /* Q( -rshifts ) */
tmp1 = silk_LSHIFT32( (opus_int32)x_ptr[ n ], 17 ); /* Q17 */
tmp2 = silk_LSHIFT32( (opus_int32)x_ptr[ subfr_length - n - 1 ], 17 ); /* Q17 */
for( k = 0; k < n; k++ ) {
C_first_row[ k ] = silk_MLA( C_first_row[ k ], x1, x_ptr[ n - k - 1 ] ); /* Q( -rshifts ) */
C_last_row[ k ] = silk_MLA( C_last_row[ k ], x2, x_ptr[ subfr_length - n + k ] ); /* Q( -rshifts ) */
Atmp1 = silk_RSHIFT_ROUND( Af_QA[ k ], QA - 17 ); /* Q17 */
tmp1 = silk_MLA( tmp1, x_ptr[ n - k - 1 ], Atmp1 ); /* Q17 */
tmp2 = silk_MLA( tmp2, x_ptr[ subfr_length - n + k ], Atmp1 ); /* Q17 */
}
tmp1 = -tmp1; /* Q17 */
tmp2 = -tmp2; /* Q17 */
for( k = 0; k <= n; k++ ) {
CAf[ k ] = silk_SMLAWW( CAf[ k ], tmp1,
silk_LSHIFT32( (opus_int32)x_ptr[ n - k ], -rshifts - 1 ) ); /* Q( -rshift ) */
CAb[ k ] = silk_SMLAWW( CAb[ k ], tmp2,
silk_LSHIFT32( (opus_int32)x_ptr[ subfr_length - n + k - 1 ], -rshifts - 1 ) ); /* Q( -rshift ) */
}
}
}
/* Calculate nominator and denominator for the next order reflection (parcor) coefficient */
tmp1 = C_first_row[ n ]; /* Q( -rshifts ) */
tmp2 = C_last_row[ n ]; /* Q( -rshifts ) */
num = 0; /* Q( -rshifts ) */
nrg = silk_ADD32( CAb[ 0 ], CAf[ 0 ] ); /* Q( 1-rshifts ) */
for( k = 0; k < n; k++ ) {
Atmp_QA = Af_QA[ k ];
lz = silk_CLZ32( silk_abs( Atmp_QA ) ) - 1;
lz = silk_min( 32 - QA, lz );
Atmp1 = silk_LSHIFT32( Atmp_QA, lz ); /* Q( QA + lz ) */
tmp1 = silk_ADD_LSHIFT32( tmp1, silk_SMMUL( C_last_row[ n - k - 1 ], Atmp1 ), 32 - QA - lz ); /* Q( -rshifts ) */
tmp2 = silk_ADD_LSHIFT32( tmp2, silk_SMMUL( C_first_row[ n - k - 1 ], Atmp1 ), 32 - QA - lz ); /* Q( -rshifts ) */
num = silk_ADD_LSHIFT32( num, silk_SMMUL( CAb[ n - k ], Atmp1 ), 32 - QA - lz ); /* Q( -rshifts ) */
nrg = silk_ADD_LSHIFT32( nrg, silk_SMMUL( silk_ADD32( CAb[ k + 1 ], CAf[ k + 1 ] ),
Atmp1 ), 32 - QA - lz ); /* Q( 1-rshifts ) */
}
CAf[ n + 1 ] = tmp1; /* Q( -rshifts ) */
CAb[ n + 1 ] = tmp2; /* Q( -rshifts ) */
num = silk_ADD32( num, tmp2 ); /* Q( -rshifts ) */
num = silk_LSHIFT32( -num, 1 ); /* Q( 1-rshifts ) */
/* Calculate the next order reflection (parcor) coefficient */
if( silk_abs( num ) < nrg ) {
rc_Q31 = silk_DIV32_varQ( num, nrg, 31 );
} else {
rc_Q31 = ( num > 0 ) ? silk_int32_MAX : silk_int32_MIN;
}
/* Update inverse prediction gain */
tmp1 = ( (opus_int32)1 << 30 ) - silk_SMMUL( rc_Q31, rc_Q31 );
tmp1 = silk_LSHIFT( silk_SMMUL( invGain_Q30, tmp1 ), 2 );
if( tmp1 <= minInvGain_Q30 ) {
/* Max prediction gain exceeded; set reflection coefficient such that max prediction gain is exactly hit */
tmp2 = ( (opus_int32)1 << 30 ) - silk_DIV32_varQ( minInvGain_Q30, invGain_Q30, 30 ); /* Q30 */
rc_Q31 = silk_SQRT_APPROX( tmp2 ); /* Q15 */
/* Newton-Raphson iteration */
rc_Q31 = silk_RSHIFT32( rc_Q31 + silk_DIV32( tmp2, rc_Q31 ), 1 ); /* Q15 */
rc_Q31 = silk_LSHIFT32( rc_Q31, 16 ); /* Q31 */
if( num < 0 ) {
/* Ensure adjusted reflection coefficients has the original sign */
rc_Q31 = -rc_Q31;
}
invGain_Q30 = minInvGain_Q30;
reached_max_gain = 1;
} else {
invGain_Q30 = tmp1;
}
/* Update the AR coefficients */
for( k = 0; k < (n + 1) >> 1; k++ ) {
tmp1 = Af_QA[ k ]; /* QA */
tmp2 = Af_QA[ n - k - 1 ]; /* QA */
Af_QA[ k ] = silk_ADD_LSHIFT32( tmp1, silk_SMMUL( tmp2, rc_Q31 ), 1 ); /* QA */
Af_QA[ n - k - 1 ] = silk_ADD_LSHIFT32( tmp2, silk_SMMUL( tmp1, rc_Q31 ), 1 ); /* QA */
}
Af_QA[ n ] = silk_RSHIFT32( rc_Q31, 31 - QA ); /* QA */
if( reached_max_gain ) {
/* Reached max prediction gain; set remaining coefficients to zero and exit loop */
for( k = n + 1; k < D; k++ ) {
Af_QA[ k ] = 0;
}
break;
}
/* Update C * Af and C * Ab */
for( k = 0; k <= n + 1; k++ ) {
tmp1 = CAf[ k ]; /* Q( -rshifts ) */
tmp2 = CAb[ n - k + 1 ]; /* Q( -rshifts ) */
CAf[ k ] = silk_ADD_LSHIFT32( tmp1, silk_SMMUL( tmp2, rc_Q31 ), 1 ); /* Q( -rshifts ) */
CAb[ n - k + 1 ] = silk_ADD_LSHIFT32( tmp2, silk_SMMUL( tmp1, rc_Q31 ), 1 ); /* Q( -rshifts ) */
}
}
if( reached_max_gain ) {
for( k = 0; k < D; k++ ) {
/* Scale coefficients */
A_Q16[ k ] = -silk_RSHIFT_ROUND( Af_QA[ k ], QA - 16 );
}
/* Subtract energy of preceding samples from C0 */
if( rshifts > 0 ) {
for( s = 0; s < nb_subfr; s++ ) {
x_ptr = x + s * subfr_length;
C0 -= (opus_int32)silk_RSHIFT64( silk_inner_prod16_aligned_64( x_ptr, x_ptr, D ), rshifts );
}
} else {
for( s = 0; s < nb_subfr; s++ ) {
x_ptr = x + s * subfr_length;
C0 -= silk_LSHIFT32( silk_inner_prod_aligned( x_ptr, x_ptr, D ), -rshifts );
}
}
/* Approximate residual energy */
*res_nrg = silk_LSHIFT( silk_SMMUL( invGain_Q30, C0 ), 2 );
*res_nrg_Q = -rshifts;
} else {
/* Return residual energy */
nrg = CAf[ 0 ]; /* Q( -rshifts ) */
tmp1 = (opus_int32)1 << 16; /* Q16 */
for( k = 0; k < D; k++ ) {
Atmp1 = silk_RSHIFT_ROUND( Af_QA[ k ], QA - 16 ); /* Q16 */
nrg = silk_SMLAWW( nrg, CAf[ k + 1 ], Atmp1 ); /* Q( -rshifts ) */
tmp1 = silk_SMLAWW( tmp1, Atmp1, Atmp1 ); /* Q16 */
A_Q16[ k ] = -Atmp1;
}
*res_nrg = silk_SMLAWW( nrg, silk_SMMUL( FIND_LPC_COND_FAC, C0 ), -tmp1 ); /* Q( -rshifts ) */
*res_nrg_Q = -rshifts;
}
}
/* 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;
}
}
/* 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;
}
}