/* Autocorrelations for a warped frequency axis */
void silk_warped_autocorrelation_FIX(
opus_int32 *corr, /* O Result [order + 1] */
opus_int *scale, /* O Scaling of the correlation vector */
const opus_int16 *input, /* I Input data to correlate */
const opus_int warping_Q16, /* I Warping coefficient */
const opus_int length, /* I Length of input */
const opus_int order /* I Correlation order (even) */
)
{
opus_int n, i, lsh;
opus_int32 tmp1_QS, tmp2_QS;
opus_int32 state_QS[ MAX_SHAPE_LPC_ORDER + 1 ] = { 0 };
opus_int64 corr_QC[ MAX_SHAPE_LPC_ORDER + 1 ] = { 0 };
/* Order must be even */
silk_assert( ( order & 1 ) == 0 );
silk_assert( 2 * QS - QC >= 0 );
/* Loop over samples */
for( n = 0; n < length; n++ ) {
tmp1_QS = silk_LSHIFT32( (opus_int32)input[ n ], QS );
/* Loop over allpass sections */
for( i = 0; i < order; i += 2 ) {
/* Output of allpass section */
tmp2_QS = silk_SMLAWB( state_QS[ i ], state_QS[ i + 1 ] - tmp1_QS, warping_Q16 );
state_QS[ i ] = tmp1_QS;
corr_QC[ i ] += silk_RSHIFT64( silk_SMULL( tmp1_QS, state_QS[ 0 ] ), 2 * QS - QC );
/* Output of allpass section */
tmp1_QS = silk_SMLAWB( state_QS[ i + 1 ], state_QS[ i + 2 ] - tmp2_QS, warping_Q16 );
state_QS[ i + 1 ] = tmp2_QS;
corr_QC[ i + 1 ] += silk_RSHIFT64( silk_SMULL( tmp2_QS, state_QS[ 0 ] ), 2 * QS - QC );
}
state_QS[ order ] = tmp1_QS;
corr_QC[ order ] += silk_RSHIFT64( silk_SMULL( tmp1_QS, state_QS[ 0 ] ), 2 * QS - QC );
}
lsh = silk_CLZ64( corr_QC[ 0 ] ) - 35;
lsh = silk_LIMIT( lsh, -12 - QC, 30 - QC );
*scale = -( QC + lsh );
silk_assert( *scale >= -30 && *scale <= 12 );
if( lsh >= 0 ) {
for( i = 0; i < order + 1; i++ ) {
corr[ i ] = (opus_int32)silk_CHECK_FIT32( silk_LSHIFT64( corr_QC[ i ], lsh ) );
}
} else {
for( i = 0; i < order + 1; i++ ) {
corr[ i ] = (opus_int32)silk_CHECK_FIT32( silk_RSHIFT64( corr_QC[ i ], -lsh ) );
}
}
silk_assert( corr_QC[ 0 ] >= 0 ); /* If breaking, decrease QC*/
}
/* Resample with a 2nd order AR filter followed by FIR interpolation */
void silk_resampler_private_down_FIR(
void *SS, /* I/O Resampler state */
opus_int16 out[], /* O Output signal */
const opus_int16 in[], /* I Input signal */
opus_int32 inLen /* I Number of input samples */
)
{
silk_resampler_state_struct *S = (silk_resampler_state_struct *)SS;
opus_int32 nSamplesIn;
opus_int32 max_index_Q16, index_increment_Q16;
VARDECL( opus_int32, buf );
const opus_int16 *FIR_Coefs;
SAVE_STACK;
ALLOC( buf, S->batchSize + S->FIR_Order, opus_int32 );
/* Copy buffered samples to start of buffer */
silk_memcpy( buf, S->sFIR.i32, S->FIR_Order * sizeof( opus_int32 ) );
FIR_Coefs = &S->Coefs[ 2 ];
/* Iterate over blocks of frameSizeIn input samples */
index_increment_Q16 = S->invRatio_Q16;
while( 1 ) {
nSamplesIn = silk_min( inLen, S->batchSize );
/* Second-order AR filter (output in Q8) */
silk_resampler_private_AR2( S->sIIR, &buf[ S->FIR_Order ], in, S->Coefs, nSamplesIn );
max_index_Q16 = silk_LSHIFT32( nSamplesIn, 16 );
/* Interpolate filtered signal */
out = silk_resampler_private_down_FIR_INTERPOL( out, buf, FIR_Coefs, S->FIR_Order,
S->FIR_Fracs, max_index_Q16, index_increment_Q16 );
in += nSamplesIn;
inLen -= nSamplesIn;
if( inLen > 1 ) {
/* More iterations to do; copy last part of filtered signal to beginning of buffer */
silk_memcpy( buf, &buf[ nSamplesIn ], S->FIR_Order * sizeof( opus_int32 ) );
} else {
break;
}
}
/* Copy last part of filtered signal to the state for the next call */
silk_memcpy( S->sFIR.i32, &buf[ nSamplesIn ], S->FIR_Order * sizeof( opus_int32 ) );
RESTORE_STACK;
}
/* Upsample using a combination of allpass-based 2x upsampling and FIR interpolation */
void silk_resampler_private_IIR_FIR(
void *SS, /* I/O Resampler state */
opus_int16 out[], /* O Output signal */
const opus_int16 in[], /* I Input signal */
opus_int32 inLen /* I Number of input samples */
)
{
silk_resampler_state_struct *S = (silk_resampler_state_struct *)SS;
opus_int32 nSamplesIn;
opus_int32 max_index_Q16, index_increment_Q16;
/* VARDECL( opus_int16, buf );
SAVE_STACK; */
/* ALLOC( buf, 2 * S->batchSize + RESAMPLER_ORDER_FIR_12, opus_int16 ); */
/* worst case = 2*16*10+8 = 328 * 2 = 656bytes */
opus_int16 buf[2 * S->batchSize + RESAMPLER_ORDER_FIR_12];
/* Copy buffered samples to start of buffer */
silk_memcpy( buf, S->sFIR.i16, RESAMPLER_ORDER_FIR_12 * sizeof( opus_int16 ) );
/* Iterate over blocks of frameSizeIn input samples */
index_increment_Q16 = S->invRatio_Q16;
while( 1 ) {
nSamplesIn = silk_min( inLen, S->batchSize );
/* Upsample 2x */
silk_resampler_private_up2_HQ( S->sIIR, &buf[ RESAMPLER_ORDER_FIR_12 ], in, nSamplesIn );
max_index_Q16 = silk_LSHIFT32( nSamplesIn, 16 + 1 ); /* + 1 because 2x upsampling */
out = silk_resampler_private_IIR_FIR_INTERPOL( out, buf, max_index_Q16, index_increment_Q16 );
in += nSamplesIn;
inLen -= nSamplesIn;
if( inLen > 0 ) {
/* More iterations to do; copy last part of filtered signal to beginning of buffer */
silk_memcpy( buf, &buf[ nSamplesIn << 1 ], RESAMPLER_ORDER_FIR_12 * sizeof( opus_int16 ) );
} else {
break;
}
}
/* Copy last part of filtered signal to the state for the next call */
silk_memcpy( S->sFIR.i16, &buf[ nSamplesIn << 1 ], RESAMPLER_ORDER_FIR_12 * sizeof( opus_int16 ) );
/* RESTORE_STACK; */
}
/* For input in Q12 domain */
opus_int32 silk_LPC_inverse_pred_gain_c( /* O Returns inverse prediction gain in energy domain, Q30 */
const opus_int16 *A_Q12, /* I Prediction coefficients, Q12 [order] */
const opus_int order /* I Prediction order */
)
{
opus_int k;
opus_int32 Atmp_QA[ SILK_MAX_ORDER_LPC ];
opus_int32 DC_resp = 0;
/* Increase Q domain of the AR coefficients */
for( k = 0; k < order; k++ ) {
DC_resp += (opus_int32)A_Q12[ k ];
Atmp_QA[ k ] = silk_LSHIFT32( (opus_int32)A_Q12[ k ], QA - 12 );
}
/* If the DC is unstable, we don't even need to do the full calculations */
if( DC_resp >= 4096 ) {
return 0;
}
return LPC_inverse_pred_gain_QA_c( Atmp_QA, order );
}
/* For input in Q12 domain */
int32_t silk_LPC_inverse_pred_gain( /* O Returns inverse prediction gain in energy domain, Q30 */
const int16_t * A_Q12, /* I Prediction coefficients, Q12 [order] */
const int order /* I Prediction order */
)
{
int k;
int32_t Atmp_QA[2][SILK_MAX_ORDER_LPC];
int32_t *Anew_QA;
int32_t DC_resp = 0;
Anew_QA = Atmp_QA[order & 1];
/* Increase Q domain of the AR coefficients */
for (k = 0; k < order; k++) {
DC_resp += (int32_t) A_Q12[k];
Anew_QA[k] = silk_LSHIFT32((int32_t) A_Q12[k], QA - 12);
}
/* If the DC is unstable, we don't even need to do the full calculations */
if (DC_resp >= 4096) {
return 0;
}
return LPC_inverse_pred_gain_QA(Atmp_QA, order);
}
static OPUS_INLINE void silk_PLC_conceal(
silk_decoder_state *psDec, /* I/O Decoder state */
silk_decoder_control *psDecCtrl, /* I/O Decoder control */
opus_int16 frame[], /* O LPC residual signal */
int arch /* I Run-time architecture */
)
{
opus_int i, j, k;
opus_int lag, idx, sLTP_buf_idx, shift1, shift2;
opus_int32 rand_seed, harm_Gain_Q15, rand_Gain_Q15, inv_gain_Q30;
opus_int32 energy1, energy2, *rand_ptr, *pred_lag_ptr;
opus_int32 LPC_pred_Q10, LTP_pred_Q12;
opus_int16 rand_scale_Q14;
opus_int16 *B_Q14;
opus_int32 *sLPC_Q14_ptr;
opus_int16 A_Q12[ MAX_LPC_ORDER ];
#ifdef SMALL_FOOTPRINT
opus_int16 *sLTP;
#else
VARDECL( opus_int16, sLTP );
#endif
VARDECL( opus_int32, sLTP_Q14 );
silk_PLC_struct *psPLC = &psDec->sPLC;
opus_int32 prevGain_Q10[2];
SAVE_STACK;
ALLOC( sLTP_Q14, psDec->ltp_mem_length + psDec->frame_length, opus_int32 );
#ifdef SMALL_FOOTPRINT
/* Ugly hack that breaks aliasing rules to save stack: put sLTP at the very end of sLTP_Q14. */
sLTP = ((opus_int16*)&sLTP_Q14[psDec->ltp_mem_length + psDec->frame_length])-psDec->ltp_mem_length;
#else
ALLOC( sLTP, psDec->ltp_mem_length, opus_int16 );
#endif
prevGain_Q10[0] = silk_RSHIFT( psPLC->prevGain_Q16[ 0 ], 6);
prevGain_Q10[1] = silk_RSHIFT( psPLC->prevGain_Q16[ 1 ], 6);
if( psDec->first_frame_after_reset ) {
silk_memset( psPLC->prevLPC_Q12, 0, sizeof( psPLC->prevLPC_Q12 ) );
}
silk_PLC_energy(&energy1, &shift1, &energy2, &shift2, psDec->exc_Q14, prevGain_Q10, psDec->subfr_length, psDec->nb_subfr);
if( silk_RSHIFT( energy1, shift2 ) < silk_RSHIFT( energy2, shift1 ) ) {
/* First sub-frame has lowest energy */
rand_ptr = &psDec->exc_Q14[ silk_max_int( 0, ( psPLC->nb_subfr - 1 ) * psPLC->subfr_length - RAND_BUF_SIZE ) ];
} else {
/* Second sub-frame has lowest energy */
rand_ptr = &psDec->exc_Q14[ silk_max_int( 0, psPLC->nb_subfr * psPLC->subfr_length - RAND_BUF_SIZE ) ];
}
/* Set up Gain to random noise component */
B_Q14 = psPLC->LTPCoef_Q14;
rand_scale_Q14 = psPLC->randScale_Q14;
/* Set up attenuation gains */
harm_Gain_Q15 = HARM_ATT_Q15[ silk_min_int( NB_ATT - 1, psDec->lossCnt ) ];
if( psDec->prevSignalType == TYPE_VOICED ) {
rand_Gain_Q15 = PLC_RAND_ATTENUATE_V_Q15[ silk_min_int( NB_ATT - 1, psDec->lossCnt ) ];
} else {
rand_Gain_Q15 = PLC_RAND_ATTENUATE_UV_Q15[ silk_min_int( NB_ATT - 1, psDec->lossCnt ) ];
}
/* LPC concealment. Apply BWE to previous LPC */
silk_bwexpander( psPLC->prevLPC_Q12, psDec->LPC_order, SILK_FIX_CONST( BWE_COEF, 16 ) );
/* Preload LPC coeficients to array on stack. Gives small performance gain */
silk_memcpy( A_Q12, psPLC->prevLPC_Q12, psDec->LPC_order * sizeof( opus_int16 ) );
/* First Lost frame */
if( psDec->lossCnt == 0 ) {
rand_scale_Q14 = 1 << 14;
/* Reduce random noise Gain for voiced frames */
if( psDec->prevSignalType == TYPE_VOICED ) {
for( i = 0; i < LTP_ORDER; i++ ) {
rand_scale_Q14 -= B_Q14[ i ];
}
rand_scale_Q14 = silk_max_16( 3277, rand_scale_Q14 ); /* 0.2 */
rand_scale_Q14 = (opus_int16)silk_RSHIFT( silk_SMULBB( rand_scale_Q14, psPLC->prevLTP_scale_Q14 ), 14 );
} else {
/* Reduce random noise for unvoiced frames with high LPC gain */
opus_int32 invGain_Q30, down_scale_Q30;
invGain_Q30 = silk_LPC_inverse_pred_gain( psPLC->prevLPC_Q12, psDec->LPC_order, arch );
down_scale_Q30 = silk_min_32( silk_RSHIFT( (opus_int32)1 << 30, LOG2_INV_LPC_GAIN_HIGH_THRES ), invGain_Q30 );
down_scale_Q30 = silk_max_32( silk_RSHIFT( (opus_int32)1 << 30, LOG2_INV_LPC_GAIN_LOW_THRES ), down_scale_Q30 );
down_scale_Q30 = silk_LSHIFT( down_scale_Q30, LOG2_INV_LPC_GAIN_HIGH_THRES );
rand_Gain_Q15 = silk_RSHIFT( silk_SMULWB( down_scale_Q30, rand_Gain_Q15 ), 14 );
}
}
rand_seed = psPLC->rand_seed;
lag = silk_RSHIFT_ROUND( psPLC->pitchL_Q8, 8 );
sLTP_buf_idx = psDec->ltp_mem_length;
/* Rewhiten LTP state */
idx = psDec->ltp_mem_length - lag - psDec->LPC_order - LTP_ORDER / 2;
silk_assert( idx > 0 );
silk_LPC_analysis_filter( &sLTP[ idx ], &psDec->outBuf[ idx ], A_Q12, psDec->ltp_mem_length - idx, psDec->LPC_order, arch );
/* Scale LTP state */
inv_gain_Q30 = silk_INVERSE32_varQ( psPLC->prevGain_Q16[ 1 ], 46 );
inv_gain_Q30 = silk_min( inv_gain_Q30, silk_int32_MAX >> 1 );
for( i = idx + psDec->LPC_order; i < psDec->ltp_mem_length; i++ ) {
sLTP_Q14[ i ] = silk_SMULWB( inv_gain_Q30, sLTP[ i ] );
}
/***************************/
/* LTP synthesis filtering */
/***************************/
for( k = 0; k < psDec->nb_subfr; k++ ) {
/* Set up pointer */
pred_lag_ptr = &sLTP_Q14[ sLTP_buf_idx - lag + LTP_ORDER / 2 ];
for( i = 0; i < psDec->subfr_length; i++ ) {
/* Unrolled loop */
/* Avoids introducing a bias because silk_SMLAWB() always rounds to -inf */
LTP_pred_Q12 = 2;
LTP_pred_Q12 = silk_SMLAWB( LTP_pred_Q12, pred_lag_ptr[ 0 ], B_Q14[ 0 ] );
LTP_pred_Q12 = silk_SMLAWB( LTP_pred_Q12, pred_lag_ptr[ -1 ], B_Q14[ 1 ] );
LTP_pred_Q12 = silk_SMLAWB( LTP_pred_Q12, pred_lag_ptr[ -2 ], B_Q14[ 2 ] );
LTP_pred_Q12 = silk_SMLAWB( LTP_pred_Q12, pred_lag_ptr[ -3 ], B_Q14[ 3 ] );
LTP_pred_Q12 = silk_SMLAWB( LTP_pred_Q12, pred_lag_ptr[ -4 ], B_Q14[ 4 ] );
pred_lag_ptr++;
/* Generate LPC excitation */
rand_seed = silk_RAND( rand_seed );
idx = silk_RSHIFT( rand_seed, 25 ) & RAND_BUF_MASK;
sLTP_Q14[ sLTP_buf_idx ] = silk_LSHIFT32( silk_SMLAWB( LTP_pred_Q12, rand_ptr[ idx ], rand_scale_Q14 ), 2 );
sLTP_buf_idx++;
}
/* Gradually reduce LTP gain */
for( j = 0; j < LTP_ORDER; j++ ) {
B_Q14[ j ] = silk_RSHIFT( silk_SMULBB( harm_Gain_Q15, B_Q14[ j ] ), 15 );
}
if ( psDec->indices.signalType != TYPE_NO_VOICE_ACTIVITY ) {
/* Gradually reduce excitation gain */
rand_scale_Q14 = silk_RSHIFT( silk_SMULBB( rand_scale_Q14, rand_Gain_Q15 ), 15 );
}
/* Slowly increase pitch lag */
psPLC->pitchL_Q8 = silk_SMLAWB( psPLC->pitchL_Q8, psPLC->pitchL_Q8, PITCH_DRIFT_FAC_Q16 );
psPLC->pitchL_Q8 = silk_min_32( psPLC->pitchL_Q8, silk_LSHIFT( silk_SMULBB( MAX_PITCH_LAG_MS, psDec->fs_kHz ), 8 ) );
lag = silk_RSHIFT_ROUND( psPLC->pitchL_Q8, 8 );
}
/***************************/
/* LPC synthesis filtering */
/***************************/
sLPC_Q14_ptr = &sLTP_Q14[ psDec->ltp_mem_length - MAX_LPC_ORDER ];
/* Copy LPC state */
silk_memcpy( sLPC_Q14_ptr, psDec->sLPC_Q14_buf, MAX_LPC_ORDER * sizeof( opus_int32 ) );
silk_assert( psDec->LPC_order >= 10 ); /* check that unrolling works */
for( i = 0; i < psDec->frame_length; i++ ) {
/* partly unrolled */
/* Avoids introducing a bias because silk_SMLAWB() always rounds to -inf */
LPC_pred_Q10 = silk_RSHIFT( psDec->LPC_order, 1 );
LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14_ptr[ MAX_LPC_ORDER + i - 1 ], A_Q12[ 0 ] );
LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14_ptr[ MAX_LPC_ORDER + i - 2 ], A_Q12[ 1 ] );
LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14_ptr[ MAX_LPC_ORDER + i - 3 ], A_Q12[ 2 ] );
LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14_ptr[ MAX_LPC_ORDER + i - 4 ], A_Q12[ 3 ] );
LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14_ptr[ MAX_LPC_ORDER + i - 5 ], A_Q12[ 4 ] );
LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14_ptr[ MAX_LPC_ORDER + i - 6 ], A_Q12[ 5 ] );
LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14_ptr[ MAX_LPC_ORDER + i - 7 ], A_Q12[ 6 ] );
LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14_ptr[ MAX_LPC_ORDER + i - 8 ], A_Q12[ 7 ] );
LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14_ptr[ MAX_LPC_ORDER + i - 9 ], A_Q12[ 8 ] );
LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14_ptr[ MAX_LPC_ORDER + i - 10 ], A_Q12[ 9 ] );
for( j = 10; j < psDec->LPC_order; j++ ) {
LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14_ptr[ MAX_LPC_ORDER + i - j - 1 ], A_Q12[ j ] );
}
/* Add prediction to LPC excitation */
sLPC_Q14_ptr[ MAX_LPC_ORDER + i ] = silk_ADD_SAT32( sLPC_Q14_ptr[ MAX_LPC_ORDER + i ],
silk_LSHIFT_SAT32( LPC_pred_Q10, 4 ));
/* Scale with Gain */
frame[ i ] = (opus_int16)silk_SAT16( silk_SAT16( silk_RSHIFT_ROUND( silk_SMULWW( sLPC_Q14_ptr[ MAX_LPC_ORDER + i ], prevGain_Q10[ 1 ] ), 8 ) ) );
}
/* Save LPC state */
silk_memcpy( psDec->sLPC_Q14_buf, &sLPC_Q14_ptr[ psDec->frame_length ], MAX_LPC_ORDER * sizeof( opus_int32 ) );
/**************************************/
/* Update states */
/**************************************/
psPLC->rand_seed = rand_seed;
psPLC->randScale_Q14 = rand_scale_Q14;
for( i = 0; i < MAX_NB_SUBFR; i++ ) {
psDecCtrl->pitchL[ i ] = lag;
}
RESTORE_STACK;
}
static inline void silk_PLC_conceal(
silk_decoder_state *psDec, /* I/O Decoder state */
silk_decoder_control *psDecCtrl, /* I/O Decoder control */
opus_int16 frame[] /* O LPC residual signal */
)
{
opus_int i, j, k;
opus_int lag, idx, sLTP_buf_idx, shift1, shift2;
opus_int32 rand_seed, harm_Gain_Q15, rand_Gain_Q15, inv_gain_Q16, inv_gain_Q30;
opus_int32 energy1, energy2, *rand_ptr, *pred_lag_ptr;
opus_int32 LPC_exc_Q14, LPC_pred_Q10, LTP_pred_Q12;
opus_int16 rand_scale_Q14;
opus_int16 *B_Q14, *exc_buf_ptr;
opus_int32 *sLPC_Q14_ptr;
opus_int16 exc_buf[ 2 * MAX_SUB_FRAME_LENGTH ];
opus_int16 A_Q12[ MAX_LPC_ORDER ];
opus_int16 sLTP[ MAX_FRAME_LENGTH ];
opus_int32 sLTP_Q14[ 2 * MAX_FRAME_LENGTH ];
silk_PLC_struct *psPLC = &psDec->sPLC;
if (psDec->first_frame_after_reset)
silk_memset(psPLC->prevLPC_Q12, 0, MAX_LPC_ORDER*sizeof(psPLC->prevLPC_Q12[ 0 ]));
/* Find random noise component */
/* Scale previous excitation signal */
exc_buf_ptr = exc_buf;
for( k = 0; k < 2; k++ ) {
for( i = 0; i < psPLC->subfr_length; i++ ) {
exc_buf_ptr[ i ] = ( opus_int16 )silk_RSHIFT(
silk_SMULWW( psDec->exc_Q10[ i + ( k + psPLC->nb_subfr - 2 ) * psPLC->subfr_length ], psPLC->prevGain_Q16[ k ] ), 10 );
}
exc_buf_ptr += psPLC->subfr_length;
}
/* Find the subframe with lowest energy of the last two and use that as random noise generator */
silk_sum_sqr_shift( &energy1, &shift1, exc_buf, psPLC->subfr_length );
silk_sum_sqr_shift( &energy2, &shift2, &exc_buf[ psPLC->subfr_length ], psPLC->subfr_length );
if( silk_RSHIFT( energy1, shift2 ) < silk_RSHIFT( energy2, shift1 ) ) {
/* First sub-frame has lowest energy */
rand_ptr = &psDec->exc_Q10[ silk_max_int( 0, ( psPLC->nb_subfr - 1 ) * psPLC->subfr_length - RAND_BUF_SIZE ) ];
} else {
/* Second sub-frame has lowest energy */
rand_ptr = &psDec->exc_Q10[ silk_max_int( 0, psPLC->nb_subfr * psPLC->subfr_length - RAND_BUF_SIZE ) ];
}
/* Setup Gain to random noise component */
B_Q14 = psPLC->LTPCoef_Q14;
rand_scale_Q14 = psPLC->randScale_Q14;
/* Setup attenuation gains */
harm_Gain_Q15 = HARM_ATT_Q15[ silk_min_int( NB_ATT - 1, psDec->lossCnt ) ];
if( psDec->prevSignalType == TYPE_VOICED ) {
rand_Gain_Q15 = PLC_RAND_ATTENUATE_V_Q15[ silk_min_int( NB_ATT - 1, psDec->lossCnt ) ];
} else {
rand_Gain_Q15 = PLC_RAND_ATTENUATE_UV_Q15[ silk_min_int( NB_ATT - 1, psDec->lossCnt ) ];
}
/* LPC concealment. Apply BWE to previous LPC */
silk_bwexpander( psPLC->prevLPC_Q12, psDec->LPC_order, SILK_FIX_CONST( BWE_COEF, 16 ) );
/* Preload LPC coeficients to array on stack. Gives small performance gain */
silk_memcpy( A_Q12, psPLC->prevLPC_Q12, psDec->LPC_order * sizeof( opus_int16 ) );
/* First Lost frame */
if( psDec->lossCnt == 0 ) {
rand_scale_Q14 = 1 << 14;
/* Reduce random noise Gain for voiced frames */
if( psDec->prevSignalType == TYPE_VOICED ) {
for( i = 0; i < LTP_ORDER; i++ ) {
rand_scale_Q14 -= B_Q14[ i ];
}
rand_scale_Q14 = silk_max_16( 3277, rand_scale_Q14 ); /* 0.2 */
rand_scale_Q14 = ( opus_int16 )silk_RSHIFT( silk_SMULBB( rand_scale_Q14, psPLC->prevLTP_scale_Q14 ), 14 );
} else {
/* Reduce random noise for unvoiced frames with high LPC gain */
opus_int32 invGain_Q30, down_scale_Q30;
silk_LPC_inverse_pred_gain( &invGain_Q30, psPLC->prevLPC_Q12, psDec->LPC_order );
down_scale_Q30 = silk_min_32( silk_RSHIFT( 1 << 30, LOG2_INV_LPC_GAIN_HIGH_THRES ), invGain_Q30 );
down_scale_Q30 = silk_max_32( silk_RSHIFT( 1 << 30, LOG2_INV_LPC_GAIN_LOW_THRES ), down_scale_Q30 );
down_scale_Q30 = silk_LSHIFT( down_scale_Q30, LOG2_INV_LPC_GAIN_HIGH_THRES );
rand_Gain_Q15 = silk_RSHIFT( silk_SMULWB( down_scale_Q30, rand_Gain_Q15 ), 14 );
}
}
rand_seed = psPLC->rand_seed;
lag = silk_RSHIFT_ROUND( psPLC->pitchL_Q8, 8 );
sLTP_buf_idx = psDec->ltp_mem_length;
/* Rewhiten LTP state */
idx = psDec->ltp_mem_length - lag - psDec->LPC_order - LTP_ORDER / 2;
silk_assert( idx > 0 );
silk_LPC_analysis_filter( &sLTP[ idx ], &psDec->outBuf[ idx ], A_Q12, psDec->ltp_mem_length - idx, psDec->LPC_order );
/* Scale LTP state */
inv_gain_Q16 = silk_INVERSE32_varQ( psPLC->prevGain_Q16[ 1 ], 32 );
inv_gain_Q16 = silk_min( inv_gain_Q16, silk_int16_MAX );
inv_gain_Q30 = silk_LSHIFT( inv_gain_Q16, 14 );
for( i = idx + psDec->LPC_order; i < psDec->ltp_mem_length; i++ ) {
sLTP_Q14[ i ] = silk_SMULWB( inv_gain_Q30, sLTP[ i ] );
}
/***************************/
/* LTP synthesis filtering */
/***************************/
for( k = 0; k < psDec->nb_subfr; k++ ) {
/* Setup pointer */
pred_lag_ptr = &sLTP_Q14[ sLTP_buf_idx - lag + LTP_ORDER / 2 ];
for( i = 0; i < psDec->subfr_length; i++ ) {
/* Unrolled loop */
LTP_pred_Q12 = silk_SMULWB( pred_lag_ptr[ 0 ], B_Q14[ 0 ] );
LTP_pred_Q12 = silk_SMLAWB( LTP_pred_Q12, pred_lag_ptr[ -1 ], B_Q14[ 1 ] );
LTP_pred_Q12 = silk_SMLAWB( LTP_pred_Q12, pred_lag_ptr[ -2 ], B_Q14[ 2 ] );
LTP_pred_Q12 = silk_SMLAWB( LTP_pred_Q12, pred_lag_ptr[ -3 ], B_Q14[ 3 ] );
LTP_pred_Q12 = silk_SMLAWB( LTP_pred_Q12, pred_lag_ptr[ -4 ], B_Q14[ 4 ] );
pred_lag_ptr++;
/* Generate LPC excitation */
rand_seed = silk_RAND( rand_seed );
idx = silk_RSHIFT( rand_seed, 25 ) & RAND_BUF_MASK;
LPC_exc_Q14 = silk_LSHIFT32( silk_SMULWB( rand_ptr[ idx ], rand_scale_Q14 ), 6 ); /* Random noise part */
LPC_exc_Q14 = silk_ADD32( LPC_exc_Q14, silk_LSHIFT32( LTP_pred_Q12, 2 ) ); /* Harmonic part */
sLTP_Q14[ sLTP_buf_idx ] = LPC_exc_Q14;
sLTP_buf_idx++;
}
/* Gradually reduce LTP gain */
for( j = 0; j < LTP_ORDER; j++ ) {
B_Q14[ j ] = silk_RSHIFT( silk_SMULBB( harm_Gain_Q15, B_Q14[ j ] ), 15 );
}
/* Gradually reduce excitation gain */
rand_scale_Q14 = silk_RSHIFT( silk_SMULBB( rand_scale_Q14, rand_Gain_Q15 ), 15 );
/* Slowly increase pitch lag */
psPLC->pitchL_Q8 = silk_SMLAWB( psPLC->pitchL_Q8, psPLC->pitchL_Q8, PITCH_DRIFT_FAC_Q16 );
psPLC->pitchL_Q8 = silk_min_32( psPLC->pitchL_Q8, silk_LSHIFT( silk_SMULBB( MAX_PITCH_LAG_MS, psDec->fs_kHz ), 8 ) );
lag = silk_RSHIFT_ROUND( psPLC->pitchL_Q8, 8 );
}
/***************************/
/* LPC synthesis filtering */
/***************************/
sLPC_Q14_ptr = &sLTP_Q14[ psDec->ltp_mem_length - MAX_LPC_ORDER ];
/* Copy LPC state */
silk_memcpy( sLPC_Q14_ptr, psDec->sLPC_Q14_buf, MAX_LPC_ORDER * sizeof( opus_int32 ) );
silk_assert( psDec->LPC_order >= 10 ); /* check that unrolling works */
for( i = 0; i < psDec->frame_length; i++ ) {
/* partly unrolled */
LPC_pred_Q10 = silk_SMULWB( sLPC_Q14_ptr[ MAX_LPC_ORDER + i - 1 ], A_Q12[ 0 ] );
LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14_ptr[ MAX_LPC_ORDER + i - 2 ], A_Q12[ 1 ] );
LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14_ptr[ MAX_LPC_ORDER + i - 3 ], A_Q12[ 2 ] );
LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14_ptr[ MAX_LPC_ORDER + i - 4 ], A_Q12[ 3 ] );
LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14_ptr[ MAX_LPC_ORDER + i - 5 ], A_Q12[ 4 ] );
LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14_ptr[ MAX_LPC_ORDER + i - 6 ], A_Q12[ 5 ] );
LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14_ptr[ MAX_LPC_ORDER + i - 7 ], A_Q12[ 6 ] );
LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14_ptr[ MAX_LPC_ORDER + i - 8 ], A_Q12[ 7 ] );
LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14_ptr[ MAX_LPC_ORDER + i - 9 ], A_Q12[ 8 ] );
LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14_ptr[ MAX_LPC_ORDER + i - 10 ], A_Q12[ 9 ] );
for( j = 10; j < psDec->LPC_order; j++ ) {
LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14_ptr[ MAX_LPC_ORDER + i - j - 1 ], A_Q12[ j ] );
}
/* Add prediction to LPC excitation */
sLPC_Q14_ptr[ MAX_LPC_ORDER + i ] = silk_ADD_LSHIFT32( sLPC_Q14_ptr[ MAX_LPC_ORDER + i ], LPC_pred_Q10, 4 );
/* Scale with Gain */
frame[ i ] = ( opus_int16 )silk_SAT16( silk_RSHIFT_ROUND( silk_SMULWW( sLPC_Q14_ptr[ MAX_LPC_ORDER + i ], psPLC->prevGain_Q16[ 1 ] ), 14 ) );
}
/* Save LPC state */
silk_memcpy( psDec->sLPC_Q14_buf, &sLPC_Q14_ptr[ psDec->frame_length ], MAX_LPC_ORDER * sizeof( opus_int32 ) );
/**************************************/
/* Update states */
/**************************************/
psPLC->rand_seed = rand_seed;
psPLC->randScale_Q14 = rand_scale_Q14;
for( i = 0; i < MAX_NB_SUBFR; i++ ) {
psDecCtrl->pitchL[ i ] = lag;
}
}
/* Compute reflection coefficients from input signal */
void silk_burg_modified(
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_NSQ_wrapper_FLP(
silk_encoder_state_FLP *psEnc, /* I/O Encoder state FLP */
silk_encoder_control_FLP *psEncCtrl, /* I/O Encoder control FLP */
SideInfoIndices *psIndices, /* I/O Quantization indices */
silk_nsq_state *psNSQ, /* I/O Noise Shaping Quantzation state */
opus_int8 pulses[], /* O Quantized pulse signal */
const silk_float x[] /* I Prefiltered input signal */
)
{
opus_int i, j;
opus_int32 x_Q3[ MAX_FRAME_LENGTH ];
opus_int32 Gains_Q16[ MAX_NB_SUBFR ];
silk_DWORD_ALIGN opus_int16 PredCoef_Q12[ 2 ][ MAX_LPC_ORDER ];
opus_int16 LTPCoef_Q14[ LTP_ORDER * MAX_NB_SUBFR ];
opus_int LTP_scale_Q14;
/* Noise shaping parameters */
opus_int16 AR2_Q13[ MAX_NB_SUBFR * MAX_SHAPE_LPC_ORDER ];
opus_int32 LF_shp_Q14[ MAX_NB_SUBFR ]; /* Packs two int16 coefficients per int32 value */
opus_int Lambda_Q10;
opus_int Tilt_Q14[ MAX_NB_SUBFR ];
opus_int HarmShapeGain_Q14[ MAX_NB_SUBFR ];
/* Convert control struct to fix control struct */
/* Noise shape parameters */
for( i = 0; i < psEnc->sCmn.nb_subfr; i++ ) {
for( j = 0; j < psEnc->sCmn.shapingLPCOrder; j++ ) {
AR2_Q13[ i * MAX_SHAPE_LPC_ORDER + j ] = silk_float2int( psEncCtrl->AR2[ i * MAX_SHAPE_LPC_ORDER + j ] * 8192.0f );
}
}
for( i = 0; i < psEnc->sCmn.nb_subfr; i++ ) {
LF_shp_Q14[ i ] = silk_LSHIFT32( silk_float2int( psEncCtrl->LF_AR_shp[ i ] * 16384.0f ), 16 ) |
(opus_uint16)silk_float2int( psEncCtrl->LF_MA_shp[ i ] * 16384.0f );
Tilt_Q14[ i ] = (opus_int)silk_float2int( psEncCtrl->Tilt[ i ] * 16384.0f );
HarmShapeGain_Q14[ i ] = (opus_int)silk_float2int( psEncCtrl->HarmShapeGain[ i ] * 16384.0f );
}
Lambda_Q10 = ( opus_int )silk_float2int( psEncCtrl->Lambda * 1024.0f );
/* prediction and coding parameters */
for( i = 0; i < psEnc->sCmn.nb_subfr * LTP_ORDER; i++ ) {
LTPCoef_Q14[ i ] = (opus_int16)silk_float2int( psEncCtrl->LTPCoef[ i ] * 16384.0f );
}
for( j = 0; j < 2; j++ ) {
for( i = 0; i < psEnc->sCmn.predictLPCOrder; i++ ) {
PredCoef_Q12[ j ][ i ] = (opus_int16)silk_float2int( psEncCtrl->PredCoef[ j ][ i ] * 4096.0f );
}
}
for( i = 0; i < psEnc->sCmn.nb_subfr; i++ ) {
Gains_Q16[ i ] = silk_float2int( psEncCtrl->Gains[ i ] * 65536.0f );
silk_assert( Gains_Q16[ i ] > 0 );
}
if( psIndices->signalType == TYPE_VOICED ) {
LTP_scale_Q14 = silk_LTPScales_table_Q14[ psIndices->LTP_scaleIndex ];
} else {
LTP_scale_Q14 = 0;
}
/* Convert input to fix */
for( i = 0; i < psEnc->sCmn.frame_length; i++ ) {
x_Q3[ i ] = silk_float2int( 8.0f * x[ i ] );
}
/* Call NSQ */
if( psEnc->sCmn.nStatesDelayedDecision > 1 || psEnc->sCmn.warping_Q16 > 0 ) {
silk_NSQ_del_dec( &psEnc->sCmn, psNSQ, psIndices, x_Q3, pulses, PredCoef_Q12[ 0 ], LTPCoef_Q14,
AR2_Q13, HarmShapeGain_Q14, Tilt_Q14, LF_shp_Q14, Gains_Q16, psEncCtrl->pitchL, Lambda_Q10, LTP_scale_Q14 );
} else {
silk_NSQ( &psEnc->sCmn, psNSQ, psIndices, x_Q3, pulses, PredCoef_Q12[ 0 ], LTPCoef_Q14,
AR2_Q13, HarmShapeGain_Q14, Tilt_Q14, LF_shp_Q14, Gains_Q16, psEncCtrl->pitchL, Lambda_Q10, LTP_scale_Q14 );
}
}
/* Initialize/reset the resampler state for a given pair of input/output sampling rates */
opus_int silk_resampler_init(
silk_resampler_state_struct *S, /* I/O Resampler state */
opus_int32 Fs_Hz_in, /* I Input sampling rate (Hz) */
opus_int32 Fs_Hz_out, /* I Output sampling rate (Hz) */
opus_int forEnc /* I If 1: encoder; if 0: decoder */
)
{
opus_int up2x;
/* Clear state */
silk_memset( S, 0, sizeof( silk_resampler_state_struct ) );
/* Input checking */
if( forEnc ) {
if( ( Fs_Hz_in != 8000 && Fs_Hz_in != 12000 && Fs_Hz_in != 16000 && Fs_Hz_in != 24000 && Fs_Hz_in != 48000 ) ||
( Fs_Hz_out != 8000 && Fs_Hz_out != 12000 && Fs_Hz_out != 16000 ) ) {
silk_assert( 0 );
return -1;
}
S->inputDelay = delay_matrix_enc[ rateID( Fs_Hz_in ) ][ rateID( Fs_Hz_out ) ];
} else {
if( ( Fs_Hz_in != 8000 && Fs_Hz_in != 12000 && Fs_Hz_in != 16000 ) ||
( Fs_Hz_out != 8000 && Fs_Hz_out != 12000 && Fs_Hz_out != 16000 && Fs_Hz_out != 24000 && Fs_Hz_out != 48000 ) ) {
silk_assert( 0 );
return -1;
}
S->inputDelay = delay_matrix_dec[ rateID( Fs_Hz_in ) ][ rateID( Fs_Hz_out ) ];
}
S->Fs_in_kHz = silk_DIV32_16( Fs_Hz_in, 1000 );
S->Fs_out_kHz = silk_DIV32_16( Fs_Hz_out, 1000 );
/* Number of samples processed per batch */
S->batchSize = S->Fs_in_kHz * RESAMPLER_MAX_BATCH_SIZE_MS;
/* Find resampler with the right sampling ratio */
up2x = 0;
if( Fs_Hz_out > Fs_Hz_in ) {
/* Upsample */
if( Fs_Hz_out == silk_MUL( Fs_Hz_in, 2 ) ) { /* Fs_out : Fs_in = 2 : 1 */
/* Special case: directly use 2x upsampler */
S->resampler_function = USE_silk_resampler_private_up2_HQ_wrapper;
} else {
/* Default resampler */
S->resampler_function = USE_silk_resampler_private_IIR_FIR;
up2x = 1;
}
} else if ( Fs_Hz_out < Fs_Hz_in ) {
/* Downsample */
S->resampler_function = USE_silk_resampler_private_down_FIR;
if( silk_MUL( Fs_Hz_out, 4 ) == silk_MUL( Fs_Hz_in, 3 ) ) { /* Fs_out : Fs_in = 3 : 4 */
S->FIR_Fracs = 3;
S->FIR_Order = RESAMPLER_DOWN_ORDER_FIR0;
S->Coefs = silk_Resampler_3_4_COEFS;
} else if( silk_MUL( Fs_Hz_out, 3 ) == silk_MUL( Fs_Hz_in, 2 ) ) { /* Fs_out : Fs_in = 2 : 3 */
S->FIR_Fracs = 2;
S->FIR_Order = RESAMPLER_DOWN_ORDER_FIR0;
S->Coefs = silk_Resampler_2_3_COEFS;
} else if( silk_MUL( Fs_Hz_out, 2 ) == Fs_Hz_in ) { /* Fs_out : Fs_in = 1 : 2 */
S->FIR_Fracs = 1;
S->FIR_Order = RESAMPLER_DOWN_ORDER_FIR1;
S->Coefs = silk_Resampler_1_2_COEFS;
} else if( silk_MUL( Fs_Hz_out, 3 ) == Fs_Hz_in ) { /* Fs_out : Fs_in = 1 : 3 */
S->FIR_Fracs = 1;
S->FIR_Order = RESAMPLER_DOWN_ORDER_FIR2;
S->Coefs = silk_Resampler_1_3_COEFS;
} else if( silk_MUL( Fs_Hz_out, 4 ) == Fs_Hz_in ) { /* Fs_out : Fs_in = 1 : 4 */
S->FIR_Fracs = 1;
S->FIR_Order = RESAMPLER_DOWN_ORDER_FIR2;
S->Coefs = silk_Resampler_1_4_COEFS;
} else if( silk_MUL( Fs_Hz_out, 6 ) == Fs_Hz_in ) { /* Fs_out : Fs_in = 1 : 6 */
S->FIR_Fracs = 1;
S->FIR_Order = RESAMPLER_DOWN_ORDER_FIR2;
S->Coefs = silk_Resampler_1_6_COEFS;
} else {
/* None available */
silk_assert( 0 );
return -1;
}
} else {
/* Input and output sampling rates are equal: copy */
S->resampler_function = USE_silk_resampler_copy;
}
/* Ratio of input/output samples */
S->invRatio_Q16 = silk_LSHIFT32( silk_DIV32( silk_LSHIFT32( Fs_Hz_in, 14 + up2x ), Fs_Hz_out ), 2 );
/* Make sure the ratio is rounded up */
while( silk_SMULWW( S->invRatio_Q16, Fs_Hz_out ) < silk_LSHIFT32( Fs_Hz_in, up2x ) ) {
S->invRatio_Q16++;
}
return 0;
}
/* Compute reflection coefficients from input signal */
void silk_burg_modified_sse4_1(
opus_int32 *res_nrg, /* O Residual energy */
opus_int *res_nrg_Q, /* O Residual energy Q value */
opus_int32 A_Q16[], /* O Prediction coefficients (length order) */
const opus_int16 x[], /* I Input signal, length: nb_subfr * (D + subfr_length) */
const opus_int32 minInvGain_Q30, /* I Inverse of max prediction gain */
const opus_int subfr_length, /* I Input signal subframe length (incl. D preceding samples) */
const opus_int nb_subfr, /* I Number of subframes stacked in x */
const opus_int D, /* I Order */
int arch /* I Run-time architecture */
)
{
opus_int k, n, s, lz, rshifts, rshifts_extra, reached_max_gain;
opus_int32 C0, num, nrg, rc_Q31, invGain_Q30, Atmp_QA, Atmp1, tmp1, tmp2, x1, x2;
const opus_int16 *x_ptr;
opus_int32 C_first_row[ SILK_MAX_ORDER_LPC ];
opus_int32 C_last_row[ SILK_MAX_ORDER_LPC ];
opus_int32 Af_QA[ SILK_MAX_ORDER_LPC ];
opus_int32 CAf[ SILK_MAX_ORDER_LPC + 1 ];
opus_int32 CAb[ SILK_MAX_ORDER_LPC + 1 ];
opus_int32 xcorr[ SILK_MAX_ORDER_LPC ];
__m128i FIRST_3210, LAST_3210, ATMP_3210, TMP1_3210, TMP2_3210, T1_3210, T2_3210, PTR_3210, SUBFR_3210, X1_3210, X2_3210;
__m128i CONST1 = _mm_set1_epi32(1);
silk_assert(subfr_length * nb_subfr <= MAX_FRAME_SIZE);
/* Compute autocorrelations, added over subframes */
silk_sum_sqr_shift(&C0, &rshifts, x, nb_subfr * subfr_length);
if(rshifts > MAX_RSHIFTS) {
C0 = silk_LSHIFT32(C0, rshifts - MAX_RSHIFTS);
silk_assert(C0 > 0);
rshifts = MAX_RSHIFTS;
} else {
lz = silk_CLZ32(C0) - 1;
rshifts_extra = N_BITS_HEAD_ROOM - lz;
if(rshifts_extra > 0) {
rshifts_extra = silk_min(rshifts_extra, MAX_RSHIFTS - rshifts);
C0 = silk_RSHIFT32(C0, rshifts_extra);
} else {
rshifts_extra = silk_max(rshifts_extra, MIN_RSHIFTS - rshifts);
C0 = silk_LSHIFT32(C0, -rshifts_extra);
}
rshifts += rshifts_extra;
}
CAb[ 0 ] = CAf[ 0 ] = C0 + silk_SMMUL(SILK_FIX_CONST(FIND_LPC_COND_FAC, 32), C0) + 1; /* Q(-rshifts) */
silk_memset(C_first_row, 0, SILK_MAX_ORDER_LPC * sizeof(opus_int32));
if(rshifts > 0) {
for(s = 0; s < nb_subfr; s++) {
x_ptr = x + s * subfr_length;
for(n = 1; n < D + 1; n++) {
C_first_row[ n - 1 ] += (opus_int32)silk_RSHIFT64(
silk_inner_prod16_aligned_64(x_ptr, x_ptr + n, subfr_length - n, arch), rshifts);
}
}
} else {
for(s = 0; s < nb_subfr; s++) {
int i;
opus_int32 d;
x_ptr = x + s * subfr_length;
celt_pitch_xcorr(x_ptr, x_ptr + 1, xcorr, subfr_length - D, D, arch);
for(n = 1; n < D + 1; n++) {
for (i = n + subfr_length - D, d = 0; i < subfr_length; i++)
d = MAC16_16(d, x_ptr[ i ], x_ptr[ i - n ]);
xcorr[ n - 1 ] += d;
}
for(n = 1; n < D + 1; n++) {
C_first_row[ n - 1 ] += silk_LSHIFT32(xcorr[ n - 1 ], -rshifts);
}
}
}
silk_memcpy(C_last_row, C_first_row, SILK_MAX_ORDER_LPC * sizeof(opus_int32));
/* Initialize */
CAb[ 0 ] = CAf[ 0 ] = C0 + silk_SMMUL(SILK_FIX_CONST(FIND_LPC_COND_FAC, 32), C0) + 1; /* Q(-rshifts) */
invGain_Q30 = (opus_int32)1 << 30;
reached_max_gain = 0;
for(n = 0; n < D; n++) {
/* Update first row of correlation matrix (without first element) */
/* Update last row of correlation matrix (without last element, stored in reversed order) */
/* Update C * Af */
/* Update C * flipud(Af) (stored in reversed order) */
if(rshifts > -2) {
for(s = 0; s < nb_subfr; s++) {
x_ptr = x + s * subfr_length;
x1 = -silk_LSHIFT32((opus_int32)x_ptr[ n ], 16 - rshifts); /* Q(16-rshifts) */
x2 = -silk_LSHIFT32((opus_int32)x_ptr[ subfr_length - n - 1 ], 16 - rshifts); /* Q(16-rshifts) */
tmp1 = silk_LSHIFT32((opus_int32)x_ptr[ n ], QA - 16); /* Q(QA-16) */
tmp2 = silk_LSHIFT32((opus_int32)x_ptr[ subfr_length - n - 1 ], QA - 16); /* Q(QA-16) */
for(k = 0; k < n; k++) {
C_first_row[ k ] = silk_SMLAWB(C_first_row[ k ], x1, x_ptr[ n - k - 1 ] ); /* Q(-rshifts) */
C_last_row[ k ] = silk_SMLAWB(C_last_row[ k ], x2, x_ptr[ subfr_length - n + k ]); /* Q(-rshifts) */
Atmp_QA = Af_QA[ k ];
tmp1 = silk_SMLAWB(tmp1, Atmp_QA, x_ptr[ n - k - 1 ] ); /* Q(QA-16) */
tmp2 = silk_SMLAWB(tmp2, Atmp_QA, x_ptr[ subfr_length - n + k ]); /* Q(QA-16) */
}
tmp1 = silk_LSHIFT32(-tmp1, 32 - QA - rshifts); /* Q(16-rshifts) */
tmp2 = silk_LSHIFT32(-tmp2, 32 - QA - rshifts); /* Q(16-rshifts) */
for(k = 0; k <= n; k++) {
CAf[ k ] = silk_SMLAWB(CAf[ k ], tmp1, x_ptr[ n - k ] ); /* Q(-rshift) */
CAb[ k ] = silk_SMLAWB(CAb[ k ], tmp2, x_ptr[ subfr_length - n + k - 1 ]); /* Q(-rshift) */
}
}
} else {
for(s = 0; s < nb_subfr; s++) {
x_ptr = x + s * subfr_length;
x1 = -silk_LSHIFT32((opus_int32)x_ptr[ n ], -rshifts); /* Q(-rshifts) */
x2 = -silk_LSHIFT32((opus_int32)x_ptr[ subfr_length - n - 1 ], -rshifts); /* Q(-rshifts) */
tmp1 = silk_LSHIFT32((opus_int32)x_ptr[ n ], 17); /* Q17 */
tmp2 = silk_LSHIFT32((opus_int32)x_ptr[ subfr_length - n - 1 ], 17); /* Q17 */
X1_3210 = _mm_set1_epi32(x1);
X2_3210 = _mm_set1_epi32(x2);
TMP1_3210 = _mm_setzero_si128();
TMP2_3210 = _mm_setzero_si128();
for(k = 0; k < n - 3; k += 4) {
PTR_3210 = OP_CVTEPI16_EPI32_M64(&x_ptr[ n - k - 1 - 3 ]);
SUBFR_3210 = OP_CVTEPI16_EPI32_M64(&x_ptr[ subfr_length - n + k ]);
FIRST_3210 = _mm_loadu_si128((__m128i *)&C_first_row[ k ]);
PTR_3210 = _mm_shuffle_epi32(PTR_3210, _MM_SHUFFLE(0, 1, 2, 3));
LAST_3210 = _mm_loadu_si128((__m128i *)&C_last_row[ k ]);
ATMP_3210 = _mm_loadu_si128((__m128i *)&Af_QA[ k ]);
T1_3210 = _mm_mullo_epi32(PTR_3210, X1_3210);
T2_3210 = _mm_mullo_epi32(SUBFR_3210, X2_3210);
ATMP_3210 = _mm_srai_epi32(ATMP_3210, 7);
ATMP_3210 = _mm_add_epi32(ATMP_3210, CONST1);
ATMP_3210 = _mm_srai_epi32(ATMP_3210, 1);
FIRST_3210 = _mm_add_epi32(FIRST_3210, T1_3210);
LAST_3210 = _mm_add_epi32(LAST_3210, T2_3210);
PTR_3210 = _mm_mullo_epi32(ATMP_3210, PTR_3210);
SUBFR_3210 = _mm_mullo_epi32(ATMP_3210, SUBFR_3210);
_mm_storeu_si128((__m128i *)&C_first_row[ k ], FIRST_3210);
_mm_storeu_si128((__m128i *)&C_last_row[ k ], LAST_3210);
TMP1_3210 = _mm_add_epi32(TMP1_3210, PTR_3210);
TMP2_3210 = _mm_add_epi32(TMP2_3210, SUBFR_3210);
}
TMP1_3210 = _mm_add_epi32(TMP1_3210, _mm_unpackhi_epi64(TMP1_3210, TMP1_3210));
TMP2_3210 = _mm_add_epi32(TMP2_3210, _mm_unpackhi_epi64(TMP2_3210, TMP2_3210));
TMP1_3210 = _mm_add_epi32(TMP1_3210, _mm_shufflelo_epi16(TMP1_3210, 0x0E));
TMP2_3210 = _mm_add_epi32(TMP2_3210, _mm_shufflelo_epi16(TMP2_3210, 0x0E));
tmp1 += _mm_cvtsi128_si32(TMP1_3210);
tmp2 += _mm_cvtsi128_si32(TMP2_3210);
for(; k < n; k++) {
C_first_row[ k ] = silk_MLA(C_first_row[ k ], x1, x_ptr[ n - k - 1 ] ); /* Q(-rshifts) */
C_last_row[ k ] = silk_MLA(C_last_row[ k ], x2, x_ptr[ subfr_length - n + k ]); /* Q(-rshifts) */
Atmp1 = silk_RSHIFT_ROUND(Af_QA[ k ], QA - 17); /* Q17 */
tmp1 = silk_MLA(tmp1, x_ptr[ n - k - 1 ], Atmp1); /* Q17 */
tmp2 = silk_MLA(tmp2, x_ptr[ subfr_length - n + k ], Atmp1); /* Q17 */
}
tmp1 = -tmp1; /* Q17 */
tmp2 = -tmp2; /* Q17 */
{
__m128i xmm_tmp1, xmm_tmp2;
__m128i xmm_x_ptr_n_k_x2x0, xmm_x_ptr_n_k_x3x1;
__m128i xmm_x_ptr_sub_x2x0, xmm_x_ptr_sub_x3x1;
xmm_tmp1 = _mm_set1_epi32(tmp1);
xmm_tmp2 = _mm_set1_epi32(tmp2);
for(k = 0; k <= n - 3; k += 4) {
xmm_x_ptr_n_k_x2x0 = OP_CVTEPI16_EPI32_M64(&x_ptr[ n - k - 3 ]);
xmm_x_ptr_sub_x2x0 = OP_CVTEPI16_EPI32_M64(&x_ptr[ subfr_length - n + k - 1 ]);
xmm_x_ptr_n_k_x2x0 = _mm_shuffle_epi32(xmm_x_ptr_n_k_x2x0, _MM_SHUFFLE(0, 1, 2, 3));
xmm_x_ptr_n_k_x2x0 = _mm_slli_epi32(xmm_x_ptr_n_k_x2x0, -rshifts - 1);
xmm_x_ptr_sub_x2x0 = _mm_slli_epi32(xmm_x_ptr_sub_x2x0, -rshifts - 1);
/* equal shift right 4 bytes, xmm_x_ptr_n_k_x3x1 = _mm_srli_si128(xmm_x_ptr_n_k_x2x0, 4)*/
xmm_x_ptr_n_k_x3x1 = _mm_shuffle_epi32(xmm_x_ptr_n_k_x2x0, _MM_SHUFFLE(0, 3, 2, 1));
xmm_x_ptr_sub_x3x1 = _mm_shuffle_epi32(xmm_x_ptr_sub_x2x0, _MM_SHUFFLE(0, 3, 2, 1));
xmm_x_ptr_n_k_x2x0 = _mm_mul_epi32(xmm_x_ptr_n_k_x2x0, xmm_tmp1);
xmm_x_ptr_n_k_x3x1 = _mm_mul_epi32(xmm_x_ptr_n_k_x3x1, xmm_tmp1);
xmm_x_ptr_sub_x2x0 = _mm_mul_epi32(xmm_x_ptr_sub_x2x0, xmm_tmp2);
xmm_x_ptr_sub_x3x1 = _mm_mul_epi32(xmm_x_ptr_sub_x3x1, xmm_tmp2);
xmm_x_ptr_n_k_x2x0 = _mm_srli_epi64(xmm_x_ptr_n_k_x2x0, 16);
xmm_x_ptr_n_k_x3x1 = _mm_slli_epi64(xmm_x_ptr_n_k_x3x1, 16);
xmm_x_ptr_sub_x2x0 = _mm_srli_epi64(xmm_x_ptr_sub_x2x0, 16);
xmm_x_ptr_sub_x3x1 = _mm_slli_epi64(xmm_x_ptr_sub_x3x1, 16);
xmm_x_ptr_n_k_x2x0 = _mm_blend_epi16(xmm_x_ptr_n_k_x2x0, xmm_x_ptr_n_k_x3x1, 0xCC);
xmm_x_ptr_sub_x2x0 = _mm_blend_epi16(xmm_x_ptr_sub_x2x0, xmm_x_ptr_sub_x3x1, 0xCC);
X1_3210 = _mm_loadu_si128((__m128i *)&CAf[ k ]);
PTR_3210 = _mm_loadu_si128((__m128i *)&CAb[ k ]);
X1_3210 = _mm_add_epi32(X1_3210, xmm_x_ptr_n_k_x2x0);
PTR_3210 = _mm_add_epi32(PTR_3210, xmm_x_ptr_sub_x2x0);
_mm_storeu_si128((__m128i *)&CAf[ k ], X1_3210);
_mm_storeu_si128((__m128i *)&CAb[ k ], PTR_3210);
}
for(; k <= n; k++) {
CAf[ k ] = silk_SMLAWW(CAf[ k ], tmp1,
silk_LSHIFT32((opus_int32)x_ptr[ n - k ], -rshifts - 1)); /* Q(-rshift) */
CAb[ k ] = silk_SMLAWW(CAb[ k ], tmp2,
silk_LSHIFT32((opus_int32)x_ptr[ subfr_length - n + k - 1 ], -rshifts - 1)); /* Q(-rshift) */
}
}
}
}
/* Calculate nominator and denominator for the next order reflection (parcor) coefficient */
tmp1 = C_first_row[ n ]; /* Q(-rshifts) */
tmp2 = C_last_row[ n ]; /* Q(-rshifts) */
num = 0; /* Q(-rshifts) */
nrg = silk_ADD32(CAb[ 0 ], CAf[ 0 ]); /* Q(1-rshifts) */
for(k = 0; k < n; k++) {
Atmp_QA = Af_QA[ k ];
lz = silk_CLZ32(silk_abs(Atmp_QA)) - 1;
lz = silk_min(32 - QA, lz);
Atmp1 = silk_LSHIFT32(Atmp_QA, lz); /* Q(QA + lz) */
tmp1 = silk_ADD_LSHIFT32(tmp1, silk_SMMUL(C_last_row[ n - k - 1 ], Atmp1), 32 - QA - lz); /* Q(-rshifts) */
tmp2 = silk_ADD_LSHIFT32(tmp2, silk_SMMUL(C_first_row[ n - k - 1 ], Atmp1), 32 - QA - lz); /* Q(-rshifts) */
num = silk_ADD_LSHIFT32(num, silk_SMMUL(CAb[ n - k ], Atmp1), 32 - QA - lz); /* Q(-rshifts) */
nrg = silk_ADD_LSHIFT32(nrg, silk_SMMUL(silk_ADD32(CAb[ k + 1 ], CAf[ k + 1 ]),
Atmp1), 32 - QA - lz); /* Q(1-rshifts) */
}
CAf[ n + 1 ] = tmp1; /* Q(-rshifts) */
CAb[ n + 1 ] = tmp2; /* Q(-rshifts) */
num = silk_ADD32(num, tmp2); /* Q(-rshifts) */
num = silk_LSHIFT32(-num, 1); /* Q(1-rshifts) */
/* Calculate the next order reflection (parcor) coefficient */
if(silk_abs(num) < nrg) {
rc_Q31 = silk_DIV32_varQ(num, nrg, 31);
} else {
rc_Q31 = (num > 0) ? silk_int32_MAX : silk_int32_MIN;
}
/* Update inverse prediction gain */
tmp1 = ((opus_int32)1 << 30) - silk_SMMUL(rc_Q31, rc_Q31);
tmp1 = silk_LSHIFT(silk_SMMUL(invGain_Q30, tmp1), 2);
if(tmp1 <= minInvGain_Q30) {
/* Max prediction gain exceeded; set reflection coefficient such that max prediction gain is exactly hit */
tmp2 = ((opus_int32)1 << 30) - silk_DIV32_varQ(minInvGain_Q30, invGain_Q30, 30); /* Q30 */
rc_Q31 = silk_SQRT_APPROX(tmp2); /* Q15 */
/* Newton-Raphson iteration */
rc_Q31 = silk_RSHIFT32(rc_Q31 + silk_DIV32(tmp2, rc_Q31), 1); /* Q15 */
rc_Q31 = silk_LSHIFT32(rc_Q31, 16); /* Q31 */
if(num < 0) {
/* Ensure adjusted reflection coefficients has the original sign */
rc_Q31 = -rc_Q31;
}
invGain_Q30 = minInvGain_Q30;
reached_max_gain = 1;
} else {
invGain_Q30 = tmp1;
}
/* Update the AR coefficients */
for(k = 0; k < (n + 1) >> 1; k++) {
tmp1 = Af_QA[ k ]; /* QA */
tmp2 = Af_QA[ n - k - 1 ]; /* QA */
Af_QA[ k ] = silk_ADD_LSHIFT32(tmp1, silk_SMMUL(tmp2, rc_Q31), 1); /* QA */
Af_QA[ n - k - 1 ] = silk_ADD_LSHIFT32(tmp2, silk_SMMUL(tmp1, rc_Q31), 1); /* QA */
}
Af_QA[ n ] = silk_RSHIFT32(rc_Q31, 31 - QA); /* QA */
if(reached_max_gain) {
/* Reached max prediction gain; set remaining coefficients to zero and exit loop */
for(k = n + 1; k < D; k++) {
Af_QA[ k ] = 0;
}
break;
}
/* Update C * Af and C * Ab */
for(k = 0; k <= n + 1; k++) {
tmp1 = CAf[ k ]; /* Q(-rshifts) */
tmp2 = CAb[ n - k + 1 ]; /* Q(-rshifts) */
CAf[ k ] = silk_ADD_LSHIFT32(tmp1, silk_SMMUL(tmp2, rc_Q31), 1); /* Q(-rshifts) */
CAb[ n - k + 1 ] = silk_ADD_LSHIFT32(tmp2, silk_SMMUL(tmp1, rc_Q31), 1); /* Q(-rshifts) */
}
}
if(reached_max_gain) {
for(k = 0; k < D; k++) {
/* Scale coefficients */
A_Q16[ k ] = -silk_RSHIFT_ROUND(Af_QA[ k ], QA - 16);
}
/* Subtract energy of preceding samples from C0 */
if(rshifts > 0) {
for(s = 0; s < nb_subfr; s++) {
x_ptr = x + s * subfr_length;
C0 -= (opus_int32)silk_RSHIFT64(silk_inner_prod16_aligned_64(x_ptr, x_ptr, D, arch), rshifts);
}
} else {
for(s = 0; s < nb_subfr; s++) {
x_ptr = x + s * subfr_length;
C0 -= silk_LSHIFT32(silk_inner_prod_aligned(x_ptr, x_ptr, D, arch), -rshifts);
}
}
/* Approximate residual energy */
*res_nrg = silk_LSHIFT(silk_SMMUL(invGain_Q30, C0), 2);
*res_nrg_Q = -rshifts;
} else {
/* Return residual energy */
nrg = CAf[ 0 ]; /* Q(-rshifts) */
tmp1 = (opus_int32)1 << 16; /* Q16 */
for(k = 0; k < D; k++) {
Atmp1 = silk_RSHIFT_ROUND(Af_QA[ k ], QA - 16); /* Q16 */
nrg = silk_SMLAWW(nrg, CAf[ k + 1 ], Atmp1); /* Q(-rshifts) */
tmp1 = silk_SMLAWW(tmp1, Atmp1, Atmp1); /* Q16 */
A_Q16[ k ] = -Atmp1;
}
*res_nrg = silk_SMLAWW(nrg, silk_SMMUL(SILK_FIX_CONST(FIND_LPC_COND_FAC, 32), C0), -tmp1);/* Q(-rshifts) */
*res_nrg_Q = -rshifts;
}
}
void silk_warped_autocorrelation_FIX_neon(
opus_int32 *corr, /* O Result [order + 1] */
opus_int *scale, /* O Scaling of the correlation vector */
const opus_int16 *input, /* I Input data to correlate */
const opus_int warping_Q16, /* I Warping coefficient */
const opus_int length, /* I Length of input */
const opus_int order /* I Correlation order (even) */
)
{
if( ( MAX_SHAPE_LPC_ORDER > 24 ) || ( order < 6 ) ) {
silk_warped_autocorrelation_FIX_c( corr, scale, input, warping_Q16, length, order );
} else {
opus_int n, i, lsh;
opus_int64 corr_QC[ MAX_SHAPE_LPC_ORDER + 1 ] = { 0 }; /* In reverse order */
opus_int64 corr_QC_orderT;
int64x2_t lsh_s64x2;
const opus_int orderT = ( order + 3 ) & ~3;
opus_int64 *corr_QCT;
opus_int32 *input_QS;
VARDECL( opus_int32, input_QST );
VARDECL( opus_int32, state );
SAVE_STACK;
/* Order must be even */
silk_assert( ( order & 1 ) == 0 );
silk_assert( 2 * QS - QC >= 0 );
ALLOC( input_QST, length + 2 * MAX_SHAPE_LPC_ORDER, opus_int32 );
input_QS = input_QST;
/* input_QS has zero paddings in the beginning and end. */
vst1q_s32( input_QS, vdupq_n_s32( 0 ) );
input_QS += 4;
vst1q_s32( input_QS, vdupq_n_s32( 0 ) );
input_QS += 4;
vst1q_s32( input_QS, vdupq_n_s32( 0 ) );
input_QS += 4;
vst1q_s32( input_QS, vdupq_n_s32( 0 ) );
input_QS += 4;
vst1q_s32( input_QS, vdupq_n_s32( 0 ) );
input_QS += 4;
vst1q_s32( input_QS, vdupq_n_s32( 0 ) );
input_QS += 4;
/* Loop over samples */
for( n = 0; n < length - 7; n += 8, input_QS += 8 ) {
const int16x8_t t0_s16x4 = vld1q_s16( input + n );
vst1q_s32( input_QS + 0, vshll_n_s16( vget_low_s16( t0_s16x4 ), QS ) );
vst1q_s32( input_QS + 4, vshll_n_s16( vget_high_s16( t0_s16x4 ), QS ) );
}
for( ; n < length; n++, input_QS++ ) {
input_QS[ 0 ] = silk_LSHIFT32( (opus_int32)input[ n ], QS );
}
vst1q_s32( input_QS, vdupq_n_s32( 0 ) );
input_QS += 4;
vst1q_s32( input_QS, vdupq_n_s32( 0 ) );
input_QS += 4;
vst1q_s32( input_QS, vdupq_n_s32( 0 ) );
input_QS += 4;
vst1q_s32( input_QS, vdupq_n_s32( 0 ) );
input_QS += 4;
vst1q_s32( input_QS, vdupq_n_s32( 0 ) );
input_QS += 4;
vst1q_s32( input_QS, vdupq_n_s32( 0 ) );
input_QS = input_QST + MAX_SHAPE_LPC_ORDER - orderT;
/* The following loop runs ( length + order ) times, with ( order ) extra epilogues. */
/* The zero paddings in input_QS guarantee corr_QC's correctness even with the extra epilogues. */
/* The values of state_QS will be polluted by the extra epilogues, however they are temporary values. */
/* Keep the C code here to help understand the intrinsics optimization. */
/*
{
opus_int32 state_QS[ 2 ][ MAX_SHAPE_LPC_ORDER + 1 ] = { 0 };
opus_int32 *state_QST[ 3 ];
state_QST[ 0 ] = state_QS[ 0 ];
state_QST[ 1 ] = state_QS[ 1 ];
for( n = 0; n < length + order; n++, input_QS++ ) {
state_QST[ 0 ][ orderT ] = input_QS[ orderT ];
for( i = 0; i < orderT; i++ ) {
corr_QC[ i ] += silk_RSHIFT64( silk_SMULL( state_QST[ 0 ][ i ], input_QS[ i ] ), 2 * QS - QC );
state_QST[ 1 ][ i ] = silk_SMLAWB( state_QST[ 1 ][ i + 1 ], state_QST[ 0 ][ i ] - state_QST[ 0 ][ i + 1 ], warping_Q16 );
}
state_QST[ 2 ] = state_QST[ 0 ];
state_QST[ 0 ] = state_QST[ 1 ];
state_QST[ 1 ] = state_QST[ 2 ];
}
}
*/
{
const int32x4_t warping_Q16_s32x4 = vdupq_n_s32( warping_Q16 << 15 );
const opus_int32 *in = input_QS + orderT;
opus_int o = orderT;
int32x4_t state_QS_s32x4[ 3 ][ 2 ];
ALLOC( state, length + orderT, opus_int32 );
state_QS_s32x4[ 2 ][ 1 ] = vdupq_n_s32( 0 );
/* Calculate 8 taps of all inputs in each loop. */
do {
state_QS_s32x4[ 0 ][ 0 ] = state_QS_s32x4[ 0 ][ 1 ] =
state_QS_s32x4[ 1 ][ 0 ] = state_QS_s32x4[ 1 ][ 1 ] = vdupq_n_s32( 0 );
n = 0;
do {
calc_corr( input_QS + n, corr_QC, o - 8, state_QS_s32x4[ 0 ][ 0 ] );
calc_corr( input_QS + n, corr_QC, o - 4, state_QS_s32x4[ 0 ][ 1 ] );
state_QS_s32x4[ 2 ][ 1 ] = vld1q_s32( in + n );
vst1q_lane_s32( state + n, state_QS_s32x4[ 0 ][ 0 ], 0 );
state_QS_s32x4[ 2 ][ 0 ] = vextq_s32( state_QS_s32x4[ 0 ][ 0 ], state_QS_s32x4[ 0 ][ 1 ], 1 );
state_QS_s32x4[ 2 ][ 1 ] = vextq_s32( state_QS_s32x4[ 0 ][ 1 ], state_QS_s32x4[ 2 ][ 1 ], 1 );
state_QS_s32x4[ 0 ][ 0 ] = calc_state( state_QS_s32x4[ 0 ][ 0 ], state_QS_s32x4[ 2 ][ 0 ], state_QS_s32x4[ 1 ][ 0 ], warping_Q16_s32x4 );
state_QS_s32x4[ 0 ][ 1 ] = calc_state( state_QS_s32x4[ 0 ][ 1 ], state_QS_s32x4[ 2 ][ 1 ], state_QS_s32x4[ 1 ][ 1 ], warping_Q16_s32x4 );
state_QS_s32x4[ 1 ][ 0 ] = state_QS_s32x4[ 2 ][ 0 ];
state_QS_s32x4[ 1 ][ 1 ] = state_QS_s32x4[ 2 ][ 1 ];
} while( ++n < ( length + order ) );
in = state;
o -= 8;
} while( o > 4 );
if( o ) {
/* Calculate the last 4 taps of all inputs. */
opus_int32 *stateT = state;
silk_assert( o == 4 );
state_QS_s32x4[ 0 ][ 0 ] = state_QS_s32x4[ 1 ][ 0 ] = vdupq_n_s32( 0 );
n = length + order;
do {
calc_corr( input_QS, corr_QC, 0, state_QS_s32x4[ 0 ][ 0 ] );
state_QS_s32x4[ 2 ][ 0 ] = vld1q_s32( stateT );
vst1q_lane_s32( stateT, state_QS_s32x4[ 0 ][ 0 ], 0 );
state_QS_s32x4[ 2 ][ 0 ] = vextq_s32( state_QS_s32x4[ 0 ][ 0 ], state_QS_s32x4[ 2 ][ 0 ], 1 );
state_QS_s32x4[ 0 ][ 0 ] = calc_state( state_QS_s32x4[ 0 ][ 0 ], state_QS_s32x4[ 2 ][ 0 ], state_QS_s32x4[ 1 ][ 0 ], warping_Q16_s32x4 );
state_QS_s32x4[ 1 ][ 0 ] = state_QS_s32x4[ 2 ][ 0 ];
input_QS++;
stateT++;
} while( --n );
}
}
{
const opus_int16 *inputT = input;
int32x4_t t_s32x4;
int64x1_t t_s64x1;
int64x2_t t_s64x2 = vdupq_n_s64( 0 );
for( n = 0; n <= length - 8; n += 8 ) {
int16x8_t input_s16x8 = vld1q_s16( inputT );
t_s32x4 = vmull_s16( vget_low_s16( input_s16x8 ), vget_low_s16( input_s16x8 ) );
t_s32x4 = vmlal_s16( t_s32x4, vget_high_s16( input_s16x8 ), vget_high_s16( input_s16x8 ) );
t_s64x2 = vaddw_s32( t_s64x2, vget_low_s32( t_s32x4 ) );
t_s64x2 = vaddw_s32( t_s64x2, vget_high_s32( t_s32x4 ) );
inputT += 8;
}
t_s64x1 = vadd_s64( vget_low_s64( t_s64x2 ), vget_high_s64( t_s64x2 ) );
corr_QC_orderT = vget_lane_s64( t_s64x1, 0 );
for( ; n < length; n++ ) {
corr_QC_orderT += silk_SMULL( input[ n ], input[ n ] );
}
corr_QC_orderT = silk_LSHIFT64( corr_QC_orderT, QC );
corr_QC[ orderT ] = corr_QC_orderT;
}
corr_QCT = corr_QC + orderT - order;
lsh = silk_CLZ64( corr_QC_orderT ) - 35;
lsh = silk_LIMIT( lsh, -12 - QC, 30 - QC );
*scale = -( QC + lsh );
silk_assert( *scale >= -30 && *scale <= 12 );
lsh_s64x2 = vdupq_n_s64( lsh );
for( i = 0; i <= order - 3; i += 4 ) {
int32x4_t corr_s32x4;
int64x2_t corr_QC0_s64x2, corr_QC1_s64x2;
corr_QC0_s64x2 = vld1q_s64( corr_QCT + i );
corr_QC1_s64x2 = vld1q_s64( corr_QCT + i + 2 );
corr_QC0_s64x2 = vshlq_s64( corr_QC0_s64x2, lsh_s64x2 );
corr_QC1_s64x2 = vshlq_s64( corr_QC1_s64x2, lsh_s64x2 );
corr_s32x4 = vcombine_s32( vmovn_s64( corr_QC1_s64x2 ), vmovn_s64( corr_QC0_s64x2 ) );
corr_s32x4 = vrev64q_s32( corr_s32x4 );
vst1q_s32( corr + order - i - 3, corr_s32x4 );
}
if( lsh >= 0 ) {
for( ; i < order + 1; i++ ) {
corr[ order - i ] = (opus_int32)silk_CHECK_FIT32( silk_LSHIFT64( corr_QCT[ i ], lsh ) );
}
} else {
for( ; i < order + 1; i++ ) {
corr[ order - i ] = (opus_int32)silk_CHECK_FIT32( silk_RSHIFT64( corr_QCT[ i ], -lsh ) );
}
}
silk_assert( corr_QCT[ order ] >= 0 ); /* If breaking, decrease QC*/
RESTORE_STACK;
}
#ifdef OPUS_CHECK_ASM
{
opus_int32 corr_c[ MAX_SHAPE_LPC_ORDER + 1 ];
opus_int scale_c;
silk_warped_autocorrelation_FIX_c( corr_c, &scale_c, input, warping_Q16, length, order );
silk_assert( !memcmp( corr_c, corr, sizeof( corr_c[ 0 ] ) * ( order + 1 ) ) );
silk_assert( scale_c == *scale );
}
#endif
}
/* 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;
}
}
/* Initialize/reset the resampler state for a given pair of input/output sampling rates */
opus_int silk_resampler_init(
silk_resampler_state_struct *S, /* I/O Resampler state */
opus_int32 Fs_Hz_in, /* I Input sampling rate (Hz) */
opus_int32 Fs_Hz_out /* I Output sampling rate (Hz) */
)
{
opus_int32 up2 = 0, down2 = 0;
/* Clear state */
silk_memset( S, 0, sizeof( silk_resampler_state_struct ) );
/* Input checking */
if( ( Fs_Hz_in != 8000 && Fs_Hz_in != 12000 && Fs_Hz_in != 16000 && Fs_Hz_in != 24000 && Fs_Hz_in != 48000 ) ||
( Fs_Hz_out != 8000 && Fs_Hz_out != 12000 && Fs_Hz_out != 16000 && Fs_Hz_out != 24000 && Fs_Hz_out != 48000 ) ) {
silk_assert( 0 );
return -1;
}
/* Number of samples processed per batch */
S->batchSize = silk_DIV32_16( Fs_Hz_in, 100 );
/* Find resampler with the right sampling ratio */
if( Fs_Hz_out > Fs_Hz_in ) {
/* Upsample */
if( Fs_Hz_out == silk_MUL( Fs_Hz_in, 2 ) ) { /* Fs_out : Fs_in = 2 : 1 */
/* Special case: directly use 2x upsampler */
S->resampler_function = USE_silk_resampler_private_up2_HQ_wrapper;
} else {
/* Default resampler */
S->resampler_function = USE_silk_resampler_private_IIR_FIR;
up2 = 1;
}
} else if ( Fs_Hz_out < Fs_Hz_in ) {
/* Downsample */
if( silk_MUL( Fs_Hz_out, 4 ) == silk_MUL( Fs_Hz_in, 3 ) ) { /* Fs_out : Fs_in = 3 : 4 */
S->FIR_Fracs = 3;
S->Coefs = silk_Resampler_3_4_COEFS;
S->resampler_function = USE_silk_resampler_private_down_FIR;
} else if( silk_MUL( Fs_Hz_out, 3 ) == silk_MUL( Fs_Hz_in, 2 ) ) { /* Fs_out : Fs_in = 2 : 3 */
S->FIR_Fracs = 2;
S->Coefs = silk_Resampler_2_3_COEFS;
S->resampler_function = USE_silk_resampler_private_down_FIR;
} else if( silk_MUL( Fs_Hz_out, 2 ) == Fs_Hz_in ) { /* Fs_out : Fs_in = 1 : 2 */
S->FIR_Fracs = 1;
S->Coefs = silk_Resampler_1_2_COEFS;
S->resampler_function = USE_silk_resampler_private_down_FIR;
} else if( silk_MUL( Fs_Hz_out, 3 ) == Fs_Hz_in ) { /* Fs_out : Fs_in = 1 : 3 */
S->FIR_Fracs = 1;
S->Coefs = silk_Resampler_1_3_COEFS;
S->resampler_function = USE_silk_resampler_private_down_FIR;
} else if( silk_MUL( Fs_Hz_out, 4 ) == Fs_Hz_in ) { /* Fs_out : Fs_in = 1 : 4 */
S->FIR_Fracs = 1;
down2 = 1;
S->Coefs = silk_Resampler_1_2_COEFS;
S->resampler_function = USE_silk_resampler_private_down_FIR;
} else if( silk_MUL( Fs_Hz_out, 6 ) == Fs_Hz_in ) { /* Fs_out : Fs_in = 1 : 6 */
S->FIR_Fracs = 1;
down2 = 1;
S->Coefs = silk_Resampler_1_3_COEFS;
S->resampler_function = USE_silk_resampler_private_down_FIR;
} else {
/* None available */
silk_assert( 0 );
return -1;
}
} else {
/* Input and output sampling rates are equal: copy */
S->resampler_function = USE_silk_resampler_copy;
}
S->input2x = up2 | down2;
/* Ratio of input/output samples */
S->invRatio_Q16 = silk_LSHIFT32( silk_DIV32( silk_LSHIFT32( Fs_Hz_in, 14 + up2 - down2 ), Fs_Hz_out ), 2 );
/* Make sure the ratio is rounded up */
while( silk_SMULWW( S->invRatio_Q16, silk_LSHIFT32( Fs_Hz_out, down2 ) ) < silk_LSHIFT32( Fs_Hz_in, up2 ) ) {
S->invRatio_Q16++;
}
return 0;
}
