/* 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;
}
/* Uses SMULL(), available on armv4 */
opus_int32 silk_schur64( /* O returns residual energy */
opus_int32 rc_Q16[], /* O Reflection coefficients [order] Q16 */
const opus_int32 c[], /* I Correlations [order+1] */
opus_int32 order /* I Prediction order */
)
{
opus_int k, n;
opus_int32 C[ SILK_MAX_ORDER_LPC + 1 ][ 2 ];
opus_int32 Ctmp1_Q30, Ctmp2_Q30, rc_tmp_Q31;
silk_assert( order==6||order==8||order==10||order==12||order==14||order==16 );
/* Check for invalid input */
if( c[ 0 ] <= 0 ) {
silk_memset( rc_Q16, 0, order * sizeof( opus_int32 ) );
return 0;
}
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_Q16[ k ] = -SILK_FIX_CONST( .99f, 16 );
} else {
rc_Q16[ k ] = SILK_FIX_CONST( .99f, 16 );
}
k++;
break;
}
/* Get reflection coefficient: divide two Q30 values and get result in Q31 */
rc_tmp_Q31 = silk_DIV32_varQ( -C[ k + 1 ][ 0 ], C[ 0 ][ 1 ], 31 );
/* Save the output */
rc_Q16[ k ] = silk_RSHIFT_ROUND( rc_tmp_Q31, 15 );
/* Update correlations */
for( n = 0; n < order - k; n++ ) {
Ctmp1_Q30 = C[ n + k + 1 ][ 0 ];
Ctmp2_Q30 = C[ n ][ 1 ];
/* Multiply and add the highest int32 */
C[ n + k + 1 ][ 0 ] = Ctmp1_Q30 + silk_SMMUL( silk_LSHIFT( Ctmp2_Q30, 1 ), rc_tmp_Q31 );
C[ n ][ 1 ] = Ctmp2_Q30 + silk_SMMUL( silk_LSHIFT( Ctmp1_Q30, 1 ), rc_tmp_Q31 );
}
}
for(; k < order; k++ ) {
rc_Q16[ k ] = 0;
}
return silk_max_32( 1, C[ 0 ][ 1 ] );
}
/* 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 );
}
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;
}
}
/* 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;
}
/* 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;
}
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_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;
}
}
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;
}
}
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 WLTP[ MAX_NB_SUBFR * LTP_ORDER * LTP_ORDER ];
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, LPC_in_pre[ MAX_NB_SUBFR * MAX_LPC_ORDER + MAX_FRAME_LENGTH ];
opus_int32 tmp, min_gain_Q16;
opus_int LTP_corrs_rshift[ MAX_NB_SUBFR ];
/* 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( ( 1 << 16 ), invGains_Q16[ i ] );
}
if( psEnc->sCmn.indices.signalType == TYPE_VOICED ) {
/**********/
/* VOICED */
/**********/
silk_assert( psEnc->sCmn.ltp_mem_length - psEnc->sCmn.predictLPCOrder >= psEncCtrl->pitchL[ 0 ] + LTP_ORDER / 2 );
/* 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,
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;
}
/* LPC_in_pre contains the LTP-filtered input for voiced, and the unfiltered input for unvoiced */
silk_find_LPC_FIX( NLSF_Q15, &psEnc->sCmn.indices.NLSFInterpCoef_Q2, psEnc->sCmn.prev_NLSFq_Q15,
psEnc->sCmn.useInterpolatedNLSFs, psEnc->sCmn.first_frame_after_reset, psEnc->sCmn.predictLPCOrder,
LPC_in_pre, psEnc->sCmn.subfr_length + psEnc->sCmn.predictLPCOrder, psEnc->sCmn.nb_subfr );
/* 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 ) );
}
/* 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;
}
