/* Convert input to a linear scale */ opus_int32 silk_log2lin( const opus_int32 inLog_Q7 /* I input on log scale */ ) { opus_int32 out, frac_Q7; if( inLog_Q7 < 0 ) { return 0; } out = silk_LSHIFT( 1, silk_RSHIFT( inLog_Q7, 7 ) ); frac_Q7 = inLog_Q7 & 0x7F; if( inLog_Q7 < 2048 ) { /* Piece-wise parabolic approximation */ out = silk_ADD_RSHIFT32( out, silk_MUL( out, silk_SMLAWB( frac_Q7, silk_SMULBB( frac_Q7, 128 - frac_Q7 ), -174 ) ), 7 ); } else { /* Piece-wise parabolic approximation */ out = silk_MLA( out, silk_RSHIFT( out, 7 ), silk_SMLAWB( frac_Q7, silk_SMULBB( frac_Q7, 128 - frac_Q7 ), -174 ) ); } return out; }
void silk_prefilter_FIX( silk_encoder_state_FIX *psEnc, /* I/O Encoder state */ const silk_encoder_control_FIX *psEncCtrl, /* I Encoder control */ opus_int32 xw_Q3[], /* O Weighted signal */ const opus_int16 x[] /* I Speech signal */ ) { silk_prefilter_state_FIX *P = &psEnc->sPrefilt; opus_int j, k, lag; opus_int32 tmp_32; const opus_int16 *AR1_shp_Q13; const opus_int16 *px; opus_int32 *pxw_Q3; opus_int HarmShapeGain_Q12, Tilt_Q14; opus_int32 HarmShapeFIRPacked_Q12, LF_shp_Q14; VARDECL( opus_int32, x_filt_Q12 ); VARDECL( opus_int32, st_res_Q2 ); opus_int16 B_Q10[ 2 ]; SAVE_STACK; /* Set up pointers */ px = x; pxw_Q3 = xw_Q3; lag = P->lagPrev; ALLOC( x_filt_Q12, psEnc->sCmn.subfr_length, opus_int32 ); ALLOC( st_res_Q2, psEnc->sCmn.subfr_length, opus_int32 ); for( k = 0; k < psEnc->sCmn.nb_subfr; k++ ) { /* Update Variables that change per sub frame */ if( psEnc->sCmn.indices.signalType == TYPE_VOICED ) { lag = psEncCtrl->pitchL[ k ]; } /* Noise shape parameters */ HarmShapeGain_Q12 = silk_SMULWB( (opus_int32)psEncCtrl->HarmShapeGain_Q14[ k ], 16384 - psEncCtrl->HarmBoost_Q14[ k ] ); silk_assert( HarmShapeGain_Q12 >= 0 ); HarmShapeFIRPacked_Q12 = silk_RSHIFT( HarmShapeGain_Q12, 2 ); HarmShapeFIRPacked_Q12 |= silk_LSHIFT( (opus_int32)silk_RSHIFT( HarmShapeGain_Q12, 1 ), 16 ); Tilt_Q14 = psEncCtrl->Tilt_Q14[ k ]; LF_shp_Q14 = psEncCtrl->LF_shp_Q14[ k ]; AR1_shp_Q13 = &psEncCtrl->AR1_Q13[ k * MAX_SHAPE_LPC_ORDER ]; /* Short term FIR filtering*/ silk_warped_LPC_analysis_filter_FIX( P->sAR_shp, st_res_Q2, AR1_shp_Q13, px, psEnc->sCmn.warping_Q16, psEnc->sCmn.subfr_length, psEnc->sCmn.shapingLPCOrder ); /* Reduce (mainly) low frequencies during harmonic emphasis */ B_Q10[ 0 ] = silk_RSHIFT_ROUND( psEncCtrl->GainsPre_Q14[ k ], 4 ); tmp_32 = silk_SMLABB( SILK_FIX_CONST( INPUT_TILT, 26 ), psEncCtrl->HarmBoost_Q14[ k ], HarmShapeGain_Q12 ); /* Q26 */ tmp_32 = silk_SMLABB( tmp_32, psEncCtrl->coding_quality_Q14, SILK_FIX_CONST( HIGH_RATE_INPUT_TILT, 12 ) ); /* Q26 */ tmp_32 = silk_SMULWB( tmp_32, -psEncCtrl->GainsPre_Q14[ k ] ); /* Q24 */ tmp_32 = silk_RSHIFT_ROUND( tmp_32, 14 ); /* Q10 */ B_Q10[ 1 ]= silk_SAT16( tmp_32 ); x_filt_Q12[ 0 ] = silk_MLA( silk_MUL( st_res_Q2[ 0 ], B_Q10[ 0 ] ), P->sHarmHP_Q2, B_Q10[ 1 ] ); for( j = 1; j < psEnc->sCmn.subfr_length; j++ ) { x_filt_Q12[ j ] = silk_MLA( silk_MUL( st_res_Q2[ j ], B_Q10[ 0 ] ), st_res_Q2[ j - 1 ], B_Q10[ 1 ] ); } P->sHarmHP_Q2 = st_res_Q2[ psEnc->sCmn.subfr_length - 1 ]; silk_prefilt_FIX( P, x_filt_Q12, pxw_Q3, HarmShapeFIRPacked_Q12, Tilt_Q14, LF_shp_Q14, lag, psEnc->sCmn.subfr_length ); px += psEnc->sCmn.subfr_length; pxw_Q3 += psEnc->sCmn.subfr_length; } P->lagPrev = psEncCtrl->pitchL[ psEnc->sCmn.nb_subfr - 1 ]; 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; }
/* Delayed-decision quantizer for NLSF residuals */ opus_int32 silk_NLSF_del_dec_quant( /* O Returns RD value in Q25 */ opus_int8 indices[], /* O Quantization indices [ order ] */ const opus_int16 x_Q10[], /* I Input [ order ] */ const opus_int16 w_Q5[], /* I Weights [ order ] */ const opus_uint8 pred_coef_Q8[], /* I Backward predictor coefs [ order ] */ const opus_int16 ec_ix[], /* I Indices to entropy coding tables [ order ] */ const opus_uint8 ec_rates_Q5[], /* I Rates [] */ const opus_int quant_step_size_Q16, /* I Quantization step size */ const opus_int16 inv_quant_step_size_Q6, /* I Inverse quantization step size */ const opus_int32 mu_Q20, /* I R/D tradeoff */ const opus_int16 order /* I Number of input values */ ) { opus_int i, j, nStates, ind_tmp, ind_min_max, ind_max_min, in_Q10, res_Q10; opus_int pred_Q10, diff_Q10, out0_Q10, out1_Q10, rate0_Q5, rate1_Q5; opus_int32 RD_tmp_Q25, min_Q25, min_max_Q25, max_min_Q25, pred_coef_Q16; opus_int ind_sort[ NLSF_QUANT_DEL_DEC_STATES ]; opus_int8 ind[ NLSF_QUANT_DEL_DEC_STATES ][ MAX_LPC_ORDER ]; opus_int16 prev_out_Q10[ 2 * NLSF_QUANT_DEL_DEC_STATES ]; opus_int32 RD_Q25[ 2 * NLSF_QUANT_DEL_DEC_STATES ]; opus_int32 RD_min_Q25[ NLSF_QUANT_DEL_DEC_STATES ]; opus_int32 RD_max_Q25[ NLSF_QUANT_DEL_DEC_STATES ]; const opus_uint8 *rates_Q5; silk_assert( (NLSF_QUANT_DEL_DEC_STATES & (NLSF_QUANT_DEL_DEC_STATES-1)) == 0 ); /* must be power of two */ nStates = 1; RD_Q25[ 0 ] = 0; prev_out_Q10[ 0 ] = 0; for( i = order - 1; ; i-- ) { rates_Q5 = &ec_rates_Q5[ ec_ix[ i ] ]; pred_coef_Q16 = silk_LSHIFT( (opus_int32)pred_coef_Q8[ i ], 8 ); in_Q10 = x_Q10[ i ]; for( j = 0; j < nStates; j++ ) { pred_Q10 = silk_SMULWB( pred_coef_Q16, prev_out_Q10[ j ] ); res_Q10 = silk_SUB16( in_Q10, pred_Q10 ); ind_tmp = silk_SMULWB( inv_quant_step_size_Q6, res_Q10 ); ind_tmp = silk_LIMIT( ind_tmp, -NLSF_QUANT_MAX_AMPLITUDE_EXT, NLSF_QUANT_MAX_AMPLITUDE_EXT-1 ); ind[ j ][ i ] = (opus_int8)ind_tmp; /* compute outputs for ind_tmp and ind_tmp + 1 */ out0_Q10 = silk_LSHIFT( ind_tmp, 10 ); out1_Q10 = silk_ADD16( out0_Q10, 1024 ); if( ind_tmp > 0 ) { out0_Q10 = silk_SUB16( out0_Q10, SILK_FIX_CONST( NLSF_QUANT_LEVEL_ADJ, 10 ) ); out1_Q10 = silk_SUB16( out1_Q10, SILK_FIX_CONST( NLSF_QUANT_LEVEL_ADJ, 10 ) ); } else if( ind_tmp == 0 ) { out1_Q10 = silk_SUB16( out1_Q10, SILK_FIX_CONST( NLSF_QUANT_LEVEL_ADJ, 10 ) ); } else if( ind_tmp == -1 ) { out0_Q10 = silk_ADD16( out0_Q10, SILK_FIX_CONST( NLSF_QUANT_LEVEL_ADJ, 10 ) ); } else { out0_Q10 = silk_ADD16( out0_Q10, SILK_FIX_CONST( NLSF_QUANT_LEVEL_ADJ, 10 ) ); out1_Q10 = silk_ADD16( out1_Q10, SILK_FIX_CONST( NLSF_QUANT_LEVEL_ADJ, 10 ) ); } out0_Q10 = silk_SMULWB( out0_Q10, quant_step_size_Q16 ); out1_Q10 = silk_SMULWB( out1_Q10, quant_step_size_Q16 ); out0_Q10 = silk_ADD16( out0_Q10, pred_Q10 ); out1_Q10 = silk_ADD16( out1_Q10, pred_Q10 ); prev_out_Q10[ j ] = out0_Q10; prev_out_Q10[ j + nStates ] = out1_Q10; /* compute RD for ind_tmp and ind_tmp + 1 */ if( ind_tmp + 1 >= NLSF_QUANT_MAX_AMPLITUDE ) { if( ind_tmp + 1 == NLSF_QUANT_MAX_AMPLITUDE ) { rate0_Q5 = rates_Q5[ ind_tmp + NLSF_QUANT_MAX_AMPLITUDE ]; rate1_Q5 = 280; } else { rate0_Q5 = silk_SMLABB( 280 - 43 * NLSF_QUANT_MAX_AMPLITUDE, 43, ind_tmp ); rate1_Q5 = silk_ADD16( rate0_Q5, 43 ); } } else if( ind_tmp <= -NLSF_QUANT_MAX_AMPLITUDE ) { if( ind_tmp == -NLSF_QUANT_MAX_AMPLITUDE ) { rate0_Q5 = 280; rate1_Q5 = rates_Q5[ ind_tmp + 1 + NLSF_QUANT_MAX_AMPLITUDE ]; } else { rate0_Q5 = silk_SMLABB( 280 - 43 * NLSF_QUANT_MAX_AMPLITUDE, -43, ind_tmp ); rate1_Q5 = silk_SUB16( rate0_Q5, 43 ); } } else { rate0_Q5 = rates_Q5[ ind_tmp + NLSF_QUANT_MAX_AMPLITUDE ]; rate1_Q5 = rates_Q5[ ind_tmp + 1 + NLSF_QUANT_MAX_AMPLITUDE ]; } RD_tmp_Q25 = RD_Q25[ j ]; diff_Q10 = silk_SUB16( in_Q10, out0_Q10 ); RD_Q25[ j ] = silk_SMLABB( silk_MLA( RD_tmp_Q25, silk_SMULBB( diff_Q10, diff_Q10 ), w_Q5[ i ] ), mu_Q20, rate0_Q5 ); diff_Q10 = silk_SUB16( in_Q10, out1_Q10 ); RD_Q25[ j + nStates ] = silk_SMLABB( silk_MLA( RD_tmp_Q25, silk_SMULBB( diff_Q10, diff_Q10 ), w_Q5[ i ] ), mu_Q20, rate1_Q5 ); } if( nStates < NLSF_QUANT_DEL_DEC_STATES ) { /* double number of states and copy */ for( j = 0; j < nStates; j++ ) { ind[ j + nStates ][ i ] = ind[ j ][ i ] + 1; } nStates = silk_LSHIFT( nStates, 1 ); for( j = nStates; j < NLSF_QUANT_DEL_DEC_STATES; j++ ) { ind[ j ][ i ] = ind[ j - nStates ][ i ]; } } else if( i > 0 ) { /* sort lower and upper half of RD_Q25, pairwise */ for( j = 0; j < NLSF_QUANT_DEL_DEC_STATES; j++ ) { if( RD_Q25[ j ] > RD_Q25[ j + NLSF_QUANT_DEL_DEC_STATES ] ) { RD_max_Q25[ j ] = RD_Q25[ j ]; RD_min_Q25[ j ] = RD_Q25[ j + NLSF_QUANT_DEL_DEC_STATES ]; RD_Q25[ j ] = RD_min_Q25[ j ]; RD_Q25[ j + NLSF_QUANT_DEL_DEC_STATES ] = RD_max_Q25[ j ]; /* swap prev_out values */ out0_Q10 = prev_out_Q10[ j ]; prev_out_Q10[ j ] = prev_out_Q10[ j + NLSF_QUANT_DEL_DEC_STATES ]; prev_out_Q10[ j + NLSF_QUANT_DEL_DEC_STATES ] = out0_Q10; ind_sort[ j ] = j + NLSF_QUANT_DEL_DEC_STATES; } else { RD_min_Q25[ j ] = RD_Q25[ j ]; RD_max_Q25[ j ] = RD_Q25[ j + NLSF_QUANT_DEL_DEC_STATES ]; ind_sort[ j ] = j; } } /* compare the highest RD values of the winning half with the lowest one in the losing half, and copy if necessary */ /* afterwards ind_sort[] will contain the indices of the NLSF_QUANT_DEL_DEC_STATES winning RD values */ while( 1 ) { min_max_Q25 = silk_int32_MAX; max_min_Q25 = 0; ind_min_max = 0; ind_max_min = 0; for( j = 0; j < NLSF_QUANT_DEL_DEC_STATES; j++ ) { if( min_max_Q25 > RD_max_Q25[ j ] ) { min_max_Q25 = RD_max_Q25[ j ]; ind_min_max = j; } if( max_min_Q25 < RD_min_Q25[ j ] ) { max_min_Q25 = RD_min_Q25[ j ]; ind_max_min = j; } } if( min_max_Q25 >= max_min_Q25 ) { break; } /* copy ind_min_max to ind_max_min */ ind_sort[ ind_max_min ] = ind_sort[ ind_min_max ] ^ NLSF_QUANT_DEL_DEC_STATES; RD_Q25[ ind_max_min ] = RD_Q25[ ind_min_max + NLSF_QUANT_DEL_DEC_STATES ]; prev_out_Q10[ ind_max_min ] = prev_out_Q10[ ind_min_max + NLSF_QUANT_DEL_DEC_STATES ]; RD_min_Q25[ ind_max_min ] = 0; RD_max_Q25[ ind_min_max ] = silk_int32_MAX; silk_memcpy( ind[ ind_max_min ], ind[ ind_min_max ], MAX_LPC_ORDER * sizeof( opus_int8 ) ); } /* increment index if it comes from the upper half */ for( j = 0; j < NLSF_QUANT_DEL_DEC_STATES; j++ ) { ind[ j ][ i ] += silk_RSHIFT( ind_sort[ j ], NLSF_QUANT_DEL_DEC_STATES_LOG2 ); } } else { /* i == 0 */ break; } } /* last sample: find winner, copy indices and return RD value */ ind_tmp = 0; min_Q25 = silk_int32_MAX; for( j = 0; j < 2 * NLSF_QUANT_DEL_DEC_STATES; j++ ) { if( min_Q25 > RD_Q25[ j ] ) { min_Q25 = RD_Q25[ j ]; ind_tmp = j; } } for( j = 0; j < order; j++ ) { indices[ j ] = ind[ ind_tmp & ( NLSF_QUANT_DEL_DEC_STATES - 1 ) ][ j ]; silk_assert( indices[ j ] >= -NLSF_QUANT_MAX_AMPLITUDE_EXT ); silk_assert( indices[ j ] <= NLSF_QUANT_MAX_AMPLITUDE_EXT ); } indices[ 0 ] += silk_RSHIFT( ind_tmp, NLSF_QUANT_DEL_DEC_STATES_LOG2 ); silk_assert( indices[ 0 ] <= NLSF_QUANT_MAX_AMPLITUDE_EXT ); silk_assert( min_Q25 >= 0 ); return min_Q25; }
/* Compute reflection coefficients from input signal */ void silk_burg_modified_sse4_1( opus_int32 *res_nrg, /* O Residual energy */ opus_int *res_nrg_Q, /* O Residual energy Q value */ opus_int32 A_Q16[], /* O Prediction coefficients (length order) */ const opus_int16 x[], /* I Input signal, length: nb_subfr * (D + subfr_length) */ const opus_int32 minInvGain_Q30, /* I Inverse of max prediction gain */ const opus_int subfr_length, /* I Input signal subframe length (incl. D preceding samples) */ const opus_int nb_subfr, /* I Number of subframes stacked in x */ const opus_int D, /* I Order */ int arch /* I Run-time architecture */ ) { opus_int k, n, s, lz, rshifts, rshifts_extra, reached_max_gain; opus_int32 C0, num, nrg, rc_Q31, invGain_Q30, Atmp_QA, Atmp1, tmp1, tmp2, x1, x2; const opus_int16 *x_ptr; opus_int32 C_first_row[ SILK_MAX_ORDER_LPC ]; opus_int32 C_last_row[ SILK_MAX_ORDER_LPC ]; opus_int32 Af_QA[ SILK_MAX_ORDER_LPC ]; opus_int32 CAf[ SILK_MAX_ORDER_LPC + 1 ]; opus_int32 CAb[ SILK_MAX_ORDER_LPC + 1 ]; opus_int32 xcorr[ SILK_MAX_ORDER_LPC ]; __m128i FIRST_3210, LAST_3210, ATMP_3210, TMP1_3210, TMP2_3210, T1_3210, T2_3210, PTR_3210, SUBFR_3210, X1_3210, X2_3210; __m128i CONST1 = _mm_set1_epi32(1); silk_assert(subfr_length * nb_subfr <= MAX_FRAME_SIZE); /* Compute autocorrelations, added over subframes */ silk_sum_sqr_shift(&C0, &rshifts, x, nb_subfr * subfr_length); if(rshifts > MAX_RSHIFTS) { C0 = silk_LSHIFT32(C0, rshifts - MAX_RSHIFTS); silk_assert(C0 > 0); rshifts = MAX_RSHIFTS; } else { lz = silk_CLZ32(C0) - 1; rshifts_extra = N_BITS_HEAD_ROOM - lz; if(rshifts_extra > 0) { rshifts_extra = silk_min(rshifts_extra, MAX_RSHIFTS - rshifts); C0 = silk_RSHIFT32(C0, rshifts_extra); } else { rshifts_extra = silk_max(rshifts_extra, MIN_RSHIFTS - rshifts); C0 = silk_LSHIFT32(C0, -rshifts_extra); } rshifts += rshifts_extra; } CAb[ 0 ] = CAf[ 0 ] = C0 + silk_SMMUL(SILK_FIX_CONST(FIND_LPC_COND_FAC, 32), C0) + 1; /* Q(-rshifts) */ silk_memset(C_first_row, 0, SILK_MAX_ORDER_LPC * sizeof(opus_int32)); if(rshifts > 0) { for(s = 0; s < nb_subfr; s++) { x_ptr = x + s * subfr_length; for(n = 1; n < D + 1; n++) { C_first_row[ n - 1 ] += (opus_int32)silk_RSHIFT64( silk_inner_prod16_aligned_64(x_ptr, x_ptr + n, subfr_length - n, arch), rshifts); } } } else { for(s = 0; s < nb_subfr; s++) { int i; opus_int32 d; x_ptr = x + s * subfr_length; celt_pitch_xcorr(x_ptr, x_ptr + 1, xcorr, subfr_length - D, D, arch); for(n = 1; n < D + 1; n++) { for (i = n + subfr_length - D, d = 0; i < subfr_length; i++) d = MAC16_16(d, x_ptr[ i ], x_ptr[ i - n ]); xcorr[ n - 1 ] += d; } for(n = 1; n < D + 1; n++) { C_first_row[ n - 1 ] += silk_LSHIFT32(xcorr[ n - 1 ], -rshifts); } } } silk_memcpy(C_last_row, C_first_row, SILK_MAX_ORDER_LPC * sizeof(opus_int32)); /* Initialize */ CAb[ 0 ] = CAf[ 0 ] = C0 + silk_SMMUL(SILK_FIX_CONST(FIND_LPC_COND_FAC, 32), C0) + 1; /* Q(-rshifts) */ invGain_Q30 = (opus_int32)1 << 30; reached_max_gain = 0; for(n = 0; n < D; n++) { /* Update first row of correlation matrix (without first element) */ /* Update last row of correlation matrix (without last element, stored in reversed order) */ /* Update C * Af */ /* Update C * flipud(Af) (stored in reversed order) */ if(rshifts > -2) { for(s = 0; s < nb_subfr; s++) { x_ptr = x + s * subfr_length; x1 = -silk_LSHIFT32((opus_int32)x_ptr[ n ], 16 - rshifts); /* Q(16-rshifts) */ x2 = -silk_LSHIFT32((opus_int32)x_ptr[ subfr_length - n - 1 ], 16 - rshifts); /* Q(16-rshifts) */ tmp1 = silk_LSHIFT32((opus_int32)x_ptr[ n ], QA - 16); /* Q(QA-16) */ tmp2 = silk_LSHIFT32((opus_int32)x_ptr[ subfr_length - n - 1 ], QA - 16); /* Q(QA-16) */ for(k = 0; k < n; k++) { C_first_row[ k ] = silk_SMLAWB(C_first_row[ k ], x1, x_ptr[ n - k - 1 ] ); /* Q(-rshifts) */ C_last_row[ k ] = silk_SMLAWB(C_last_row[ k ], x2, x_ptr[ subfr_length - n + k ]); /* Q(-rshifts) */ Atmp_QA = Af_QA[ k ]; tmp1 = silk_SMLAWB(tmp1, Atmp_QA, x_ptr[ n - k - 1 ] ); /* Q(QA-16) */ tmp2 = silk_SMLAWB(tmp2, Atmp_QA, x_ptr[ subfr_length - n + k ]); /* Q(QA-16) */ } tmp1 = silk_LSHIFT32(-tmp1, 32 - QA - rshifts); /* Q(16-rshifts) */ tmp2 = silk_LSHIFT32(-tmp2, 32 - QA - rshifts); /* Q(16-rshifts) */ for(k = 0; k <= n; k++) { CAf[ k ] = silk_SMLAWB(CAf[ k ], tmp1, x_ptr[ n - k ] ); /* Q(-rshift) */ CAb[ k ] = silk_SMLAWB(CAb[ k ], tmp2, x_ptr[ subfr_length - n + k - 1 ]); /* Q(-rshift) */ } } } else { for(s = 0; s < nb_subfr; s++) { x_ptr = x + s * subfr_length; x1 = -silk_LSHIFT32((opus_int32)x_ptr[ n ], -rshifts); /* Q(-rshifts) */ x2 = -silk_LSHIFT32((opus_int32)x_ptr[ subfr_length - n - 1 ], -rshifts); /* Q(-rshifts) */ tmp1 = silk_LSHIFT32((opus_int32)x_ptr[ n ], 17); /* Q17 */ tmp2 = silk_LSHIFT32((opus_int32)x_ptr[ subfr_length - n - 1 ], 17); /* Q17 */ X1_3210 = _mm_set1_epi32(x1); X2_3210 = _mm_set1_epi32(x2); TMP1_3210 = _mm_setzero_si128(); TMP2_3210 = _mm_setzero_si128(); for(k = 0; k < n - 3; k += 4) { PTR_3210 = OP_CVTEPI16_EPI32_M64(&x_ptr[ n - k - 1 - 3 ]); SUBFR_3210 = OP_CVTEPI16_EPI32_M64(&x_ptr[ subfr_length - n + k ]); FIRST_3210 = _mm_loadu_si128((__m128i *)&C_first_row[ k ]); PTR_3210 = _mm_shuffle_epi32(PTR_3210, _MM_SHUFFLE(0, 1, 2, 3)); LAST_3210 = _mm_loadu_si128((__m128i *)&C_last_row[ k ]); ATMP_3210 = _mm_loadu_si128((__m128i *)&Af_QA[ k ]); T1_3210 = _mm_mullo_epi32(PTR_3210, X1_3210); T2_3210 = _mm_mullo_epi32(SUBFR_3210, X2_3210); ATMP_3210 = _mm_srai_epi32(ATMP_3210, 7); ATMP_3210 = _mm_add_epi32(ATMP_3210, CONST1); ATMP_3210 = _mm_srai_epi32(ATMP_3210, 1); FIRST_3210 = _mm_add_epi32(FIRST_3210, T1_3210); LAST_3210 = _mm_add_epi32(LAST_3210, T2_3210); PTR_3210 = _mm_mullo_epi32(ATMP_3210, PTR_3210); SUBFR_3210 = _mm_mullo_epi32(ATMP_3210, SUBFR_3210); _mm_storeu_si128((__m128i *)&C_first_row[ k ], FIRST_3210); _mm_storeu_si128((__m128i *)&C_last_row[ k ], LAST_3210); TMP1_3210 = _mm_add_epi32(TMP1_3210, PTR_3210); TMP2_3210 = _mm_add_epi32(TMP2_3210, SUBFR_3210); } TMP1_3210 = _mm_add_epi32(TMP1_3210, _mm_unpackhi_epi64(TMP1_3210, TMP1_3210)); TMP2_3210 = _mm_add_epi32(TMP2_3210, _mm_unpackhi_epi64(TMP2_3210, TMP2_3210)); TMP1_3210 = _mm_add_epi32(TMP1_3210, _mm_shufflelo_epi16(TMP1_3210, 0x0E)); TMP2_3210 = _mm_add_epi32(TMP2_3210, _mm_shufflelo_epi16(TMP2_3210, 0x0E)); tmp1 += _mm_cvtsi128_si32(TMP1_3210); tmp2 += _mm_cvtsi128_si32(TMP2_3210); for(; k < n; k++) { C_first_row[ k ] = silk_MLA(C_first_row[ k ], x1, x_ptr[ n - k - 1 ] ); /* Q(-rshifts) */ C_last_row[ k ] = silk_MLA(C_last_row[ k ], x2, x_ptr[ subfr_length - n + k ]); /* Q(-rshifts) */ Atmp1 = silk_RSHIFT_ROUND(Af_QA[ k ], QA - 17); /* Q17 */ tmp1 = silk_MLA(tmp1, x_ptr[ n - k - 1 ], Atmp1); /* Q17 */ tmp2 = silk_MLA(tmp2, x_ptr[ subfr_length - n + k ], Atmp1); /* Q17 */ } tmp1 = -tmp1; /* Q17 */ tmp2 = -tmp2; /* Q17 */ { __m128i xmm_tmp1, xmm_tmp2; __m128i xmm_x_ptr_n_k_x2x0, xmm_x_ptr_n_k_x3x1; __m128i xmm_x_ptr_sub_x2x0, xmm_x_ptr_sub_x3x1; xmm_tmp1 = _mm_set1_epi32(tmp1); xmm_tmp2 = _mm_set1_epi32(tmp2); for(k = 0; k <= n - 3; k += 4) { xmm_x_ptr_n_k_x2x0 = OP_CVTEPI16_EPI32_M64(&x_ptr[ n - k - 3 ]); xmm_x_ptr_sub_x2x0 = OP_CVTEPI16_EPI32_M64(&x_ptr[ subfr_length - n + k - 1 ]); xmm_x_ptr_n_k_x2x0 = _mm_shuffle_epi32(xmm_x_ptr_n_k_x2x0, _MM_SHUFFLE(0, 1, 2, 3)); xmm_x_ptr_n_k_x2x0 = _mm_slli_epi32(xmm_x_ptr_n_k_x2x0, -rshifts - 1); xmm_x_ptr_sub_x2x0 = _mm_slli_epi32(xmm_x_ptr_sub_x2x0, -rshifts - 1); /* equal shift right 4 bytes, xmm_x_ptr_n_k_x3x1 = _mm_srli_si128(xmm_x_ptr_n_k_x2x0, 4)*/ xmm_x_ptr_n_k_x3x1 = _mm_shuffle_epi32(xmm_x_ptr_n_k_x2x0, _MM_SHUFFLE(0, 3, 2, 1)); xmm_x_ptr_sub_x3x1 = _mm_shuffle_epi32(xmm_x_ptr_sub_x2x0, _MM_SHUFFLE(0, 3, 2, 1)); xmm_x_ptr_n_k_x2x0 = _mm_mul_epi32(xmm_x_ptr_n_k_x2x0, xmm_tmp1); xmm_x_ptr_n_k_x3x1 = _mm_mul_epi32(xmm_x_ptr_n_k_x3x1, xmm_tmp1); xmm_x_ptr_sub_x2x0 = _mm_mul_epi32(xmm_x_ptr_sub_x2x0, xmm_tmp2); xmm_x_ptr_sub_x3x1 = _mm_mul_epi32(xmm_x_ptr_sub_x3x1, xmm_tmp2); xmm_x_ptr_n_k_x2x0 = _mm_srli_epi64(xmm_x_ptr_n_k_x2x0, 16); xmm_x_ptr_n_k_x3x1 = _mm_slli_epi64(xmm_x_ptr_n_k_x3x1, 16); xmm_x_ptr_sub_x2x0 = _mm_srli_epi64(xmm_x_ptr_sub_x2x0, 16); xmm_x_ptr_sub_x3x1 = _mm_slli_epi64(xmm_x_ptr_sub_x3x1, 16); xmm_x_ptr_n_k_x2x0 = _mm_blend_epi16(xmm_x_ptr_n_k_x2x0, xmm_x_ptr_n_k_x3x1, 0xCC); xmm_x_ptr_sub_x2x0 = _mm_blend_epi16(xmm_x_ptr_sub_x2x0, xmm_x_ptr_sub_x3x1, 0xCC); X1_3210 = _mm_loadu_si128((__m128i *)&CAf[ k ]); PTR_3210 = _mm_loadu_si128((__m128i *)&CAb[ k ]); X1_3210 = _mm_add_epi32(X1_3210, xmm_x_ptr_n_k_x2x0); PTR_3210 = _mm_add_epi32(PTR_3210, xmm_x_ptr_sub_x2x0); _mm_storeu_si128((__m128i *)&CAf[ k ], X1_3210); _mm_storeu_si128((__m128i *)&CAb[ k ], PTR_3210); } for(; k <= n; k++) { CAf[ k ] = silk_SMLAWW(CAf[ k ], tmp1, silk_LSHIFT32((opus_int32)x_ptr[ n - k ], -rshifts - 1)); /* Q(-rshift) */ CAb[ k ] = silk_SMLAWW(CAb[ k ], tmp2, silk_LSHIFT32((opus_int32)x_ptr[ subfr_length - n + k - 1 ], -rshifts - 1)); /* Q(-rshift) */ } } } } /* Calculate nominator and denominator for the next order reflection (parcor) coefficient */ tmp1 = C_first_row[ n ]; /* Q(-rshifts) */ tmp2 = C_last_row[ n ]; /* Q(-rshifts) */ num = 0; /* Q(-rshifts) */ nrg = silk_ADD32(CAb[ 0 ], CAf[ 0 ]); /* Q(1-rshifts) */ for(k = 0; k < n; k++) { Atmp_QA = Af_QA[ k ]; lz = silk_CLZ32(silk_abs(Atmp_QA)) - 1; lz = silk_min(32 - QA, lz); Atmp1 = silk_LSHIFT32(Atmp_QA, lz); /* Q(QA + lz) */ tmp1 = silk_ADD_LSHIFT32(tmp1, silk_SMMUL(C_last_row[ n - k - 1 ], Atmp1), 32 - QA - lz); /* Q(-rshifts) */ tmp2 = silk_ADD_LSHIFT32(tmp2, silk_SMMUL(C_first_row[ n - k - 1 ], Atmp1), 32 - QA - lz); /* Q(-rshifts) */ num = silk_ADD_LSHIFT32(num, silk_SMMUL(CAb[ n - k ], Atmp1), 32 - QA - lz); /* Q(-rshifts) */ nrg = silk_ADD_LSHIFT32(nrg, silk_SMMUL(silk_ADD32(CAb[ k + 1 ], CAf[ k + 1 ]), Atmp1), 32 - QA - lz); /* Q(1-rshifts) */ } CAf[ n + 1 ] = tmp1; /* Q(-rshifts) */ CAb[ n + 1 ] = tmp2; /* Q(-rshifts) */ num = silk_ADD32(num, tmp2); /* Q(-rshifts) */ num = silk_LSHIFT32(-num, 1); /* Q(1-rshifts) */ /* Calculate the next order reflection (parcor) coefficient */ if(silk_abs(num) < nrg) { rc_Q31 = silk_DIV32_varQ(num, nrg, 31); } else { rc_Q31 = (num > 0) ? silk_int32_MAX : silk_int32_MIN; } /* Update inverse prediction gain */ tmp1 = ((opus_int32)1 << 30) - silk_SMMUL(rc_Q31, rc_Q31); tmp1 = silk_LSHIFT(silk_SMMUL(invGain_Q30, tmp1), 2); if(tmp1 <= minInvGain_Q30) { /* Max prediction gain exceeded; set reflection coefficient such that max prediction gain is exactly hit */ tmp2 = ((opus_int32)1 << 30) - silk_DIV32_varQ(minInvGain_Q30, invGain_Q30, 30); /* Q30 */ rc_Q31 = silk_SQRT_APPROX(tmp2); /* Q15 */ /* Newton-Raphson iteration */ rc_Q31 = silk_RSHIFT32(rc_Q31 + silk_DIV32(tmp2, rc_Q31), 1); /* Q15 */ rc_Q31 = silk_LSHIFT32(rc_Q31, 16); /* Q31 */ if(num < 0) { /* Ensure adjusted reflection coefficients has the original sign */ rc_Q31 = -rc_Q31; } invGain_Q30 = minInvGain_Q30; reached_max_gain = 1; } else { invGain_Q30 = tmp1; } /* Update the AR coefficients */ for(k = 0; k < (n + 1) >> 1; k++) { tmp1 = Af_QA[ k ]; /* QA */ tmp2 = Af_QA[ n - k - 1 ]; /* QA */ Af_QA[ k ] = silk_ADD_LSHIFT32(tmp1, silk_SMMUL(tmp2, rc_Q31), 1); /* QA */ Af_QA[ n - k - 1 ] = silk_ADD_LSHIFT32(tmp2, silk_SMMUL(tmp1, rc_Q31), 1); /* QA */ } Af_QA[ n ] = silk_RSHIFT32(rc_Q31, 31 - QA); /* QA */ if(reached_max_gain) { /* Reached max prediction gain; set remaining coefficients to zero and exit loop */ for(k = n + 1; k < D; k++) { Af_QA[ k ] = 0; } break; } /* Update C * Af and C * Ab */ for(k = 0; k <= n + 1; k++) { tmp1 = CAf[ k ]; /* Q(-rshifts) */ tmp2 = CAb[ n - k + 1 ]; /* Q(-rshifts) */ CAf[ k ] = silk_ADD_LSHIFT32(tmp1, silk_SMMUL(tmp2, rc_Q31), 1); /* Q(-rshifts) */ CAb[ n - k + 1 ] = silk_ADD_LSHIFT32(tmp2, silk_SMMUL(tmp1, rc_Q31), 1); /* Q(-rshifts) */ } } if(reached_max_gain) { for(k = 0; k < D; k++) { /* Scale coefficients */ A_Q16[ k ] = -silk_RSHIFT_ROUND(Af_QA[ k ], QA - 16); } /* Subtract energy of preceding samples from C0 */ if(rshifts > 0) { for(s = 0; s < nb_subfr; s++) { x_ptr = x + s * subfr_length; C0 -= (opus_int32)silk_RSHIFT64(silk_inner_prod16_aligned_64(x_ptr, x_ptr, D, arch), rshifts); } } else { for(s = 0; s < nb_subfr; s++) { x_ptr = x + s * subfr_length; C0 -= silk_LSHIFT32(silk_inner_prod_aligned(x_ptr, x_ptr, D, arch), -rshifts); } } /* Approximate residual energy */ *res_nrg = silk_LSHIFT(silk_SMMUL(invGain_Q30, C0), 2); *res_nrg_Q = -rshifts; } else { /* Return residual energy */ nrg = CAf[ 0 ]; /* Q(-rshifts) */ tmp1 = (opus_int32)1 << 16; /* Q16 */ for(k = 0; k < D; k++) { Atmp1 = silk_RSHIFT_ROUND(Af_QA[ k ], QA - 16); /* Q16 */ nrg = silk_SMLAWW(nrg, CAf[ k + 1 ], Atmp1); /* Q(-rshifts) */ tmp1 = silk_SMLAWW(tmp1, Atmp1, Atmp1); /* Q16 */ A_Q16[ k ] = -Atmp1; } *res_nrg = silk_SMLAWW(nrg, silk_SMMUL(SILK_FIX_CONST(FIND_LPC_COND_FAC, 32), C0), -tmp1);/* Q(-rshifts) */ *res_nrg_Q = -rshifts; } }
/* Compute reflection coefficients from input signal */ void silk_burg_modified( opus_int32 *res_nrg, /* O Residual energy */ opus_int *res_nrg_Q, /* O Residual energy Q value */ opus_int32 A_Q16[], /* O Prediction coefficients (length order) */ const opus_int16 x[], /* I Input signal, length: nb_subfr * ( D + subfr_length ) */ const opus_int32 minInvGain_Q30, /* I Inverse of max prediction gain */ const opus_int subfr_length, /* I Input signal subframe length (incl. D preceding samples) */ const opus_int nb_subfr, /* I Number of subframes stacked in x */ const opus_int D /* I Order */ ) { opus_int k, n, s, lz, rshifts, rshifts_extra, reached_max_gain; opus_int32 C0, num, nrg, rc_Q31, invGain_Q30, Atmp_QA, Atmp1, tmp1, tmp2, x1, x2; const opus_int16 *x_ptr; opus_int32 C_first_row[ SILK_MAX_ORDER_LPC ]; opus_int32 C_last_row[ SILK_MAX_ORDER_LPC ]; opus_int32 Af_QA[ SILK_MAX_ORDER_LPC ]; opus_int32 CAf[ SILK_MAX_ORDER_LPC + 1 ]; opus_int32 CAb[ SILK_MAX_ORDER_LPC + 1 ]; silk_assert( subfr_length * nb_subfr <= MAX_FRAME_SIZE ); /* Compute autocorrelations, added over subframes */ silk_sum_sqr_shift( &C0, &rshifts, x, nb_subfr * subfr_length ); if( rshifts > MAX_RSHIFTS ) { C0 = silk_LSHIFT32( C0, rshifts - MAX_RSHIFTS ); silk_assert( C0 > 0 ); rshifts = MAX_RSHIFTS; } else { lz = silk_CLZ32( C0 ) - 1; rshifts_extra = N_BITS_HEAD_ROOM - lz; if( rshifts_extra > 0 ) { rshifts_extra = silk_min( rshifts_extra, MAX_RSHIFTS - rshifts ); C0 = silk_RSHIFT32( C0, rshifts_extra ); } else { rshifts_extra = silk_max( rshifts_extra, MIN_RSHIFTS - rshifts ); C0 = silk_LSHIFT32( C0, -rshifts_extra ); } rshifts += rshifts_extra; } CAb[ 0 ] = CAf[ 0 ] = C0 + silk_SMMUL( SILK_FIX_CONST( FIND_LPC_COND_FAC, 32 ), C0 ) + 1; /* Q(-rshifts) */ silk_memset( C_first_row, 0, SILK_MAX_ORDER_LPC * sizeof( opus_int32 ) ); if( rshifts > 0 ) { for( s = 0; s < nb_subfr; s++ ) { x_ptr = x + s * subfr_length; for( n = 1; n < D + 1; n++ ) { C_first_row[ n - 1 ] += (opus_int32)silk_RSHIFT64( silk_inner_prod16_aligned_64( x_ptr, x_ptr + n, subfr_length - n ), rshifts ); } } } else { for( s = 0; s < nb_subfr; s++ ) { x_ptr = x + s * subfr_length; for( n = 1; n < D + 1; n++ ) { C_first_row[ n - 1 ] += silk_LSHIFT32( silk_inner_prod_aligned( x_ptr, x_ptr + n, subfr_length - n ), -rshifts ); } } } silk_memcpy( C_last_row, C_first_row, SILK_MAX_ORDER_LPC * sizeof( opus_int32 ) ); /* Initialize */ CAb[ 0 ] = CAf[ 0 ] = C0 + silk_SMMUL( SILK_FIX_CONST( FIND_LPC_COND_FAC, 32 ), C0 ) + 1; /* Q(-rshifts) */ invGain_Q30 = (opus_int32)1 << 30; reached_max_gain = 0; for( n = 0; n < D; n++ ) { /* Update first row of correlation matrix (without first element) */ /* Update last row of correlation matrix (without last element, stored in reversed order) */ /* Update C * Af */ /* Update C * flipud(Af) (stored in reversed order) */ if( rshifts > -2 ) { for( s = 0; s < nb_subfr; s++ ) { x_ptr = x + s * subfr_length; x1 = -silk_LSHIFT32( (opus_int32)x_ptr[ n ], 16 - rshifts ); /* Q(16-rshifts) */ x2 = -silk_LSHIFT32( (opus_int32)x_ptr[ subfr_length - n - 1 ], 16 - rshifts ); /* Q(16-rshifts) */ tmp1 = silk_LSHIFT32( (opus_int32)x_ptr[ n ], QA - 16 ); /* Q(QA-16) */ tmp2 = silk_LSHIFT32( (opus_int32)x_ptr[ subfr_length - n - 1 ], QA - 16 ); /* Q(QA-16) */ for( k = 0; k < n; k++ ) { C_first_row[ k ] = silk_SMLAWB( C_first_row[ k ], x1, x_ptr[ n - k - 1 ] ); /* Q( -rshifts ) */ C_last_row[ k ] = silk_SMLAWB( C_last_row[ k ], x2, x_ptr[ subfr_length - n + k ] ); /* Q( -rshifts ) */ Atmp_QA = Af_QA[ k ]; tmp1 = silk_SMLAWB( tmp1, Atmp_QA, x_ptr[ n - k - 1 ] ); /* Q(QA-16) */ tmp2 = silk_SMLAWB( tmp2, Atmp_QA, x_ptr[ subfr_length - n + k ] ); /* Q(QA-16) */ } tmp1 = silk_LSHIFT32( -tmp1, 32 - QA - rshifts ); /* Q(16-rshifts) */ tmp2 = silk_LSHIFT32( -tmp2, 32 - QA - rshifts ); /* Q(16-rshifts) */ for( k = 0; k <= n; k++ ) { CAf[ k ] = silk_SMLAWB( CAf[ k ], tmp1, x_ptr[ n - k ] ); /* Q( -rshift ) */ CAb[ k ] = silk_SMLAWB( CAb[ k ], tmp2, x_ptr[ subfr_length - n + k - 1 ] ); /* Q( -rshift ) */ } } } else { for( s = 0; s < nb_subfr; s++ ) { x_ptr = x + s * subfr_length; x1 = -silk_LSHIFT32( (opus_int32)x_ptr[ n ], -rshifts ); /* Q( -rshifts ) */ x2 = -silk_LSHIFT32( (opus_int32)x_ptr[ subfr_length - n - 1 ], -rshifts ); /* Q( -rshifts ) */ tmp1 = silk_LSHIFT32( (opus_int32)x_ptr[ n ], 17 ); /* Q17 */ tmp2 = silk_LSHIFT32( (opus_int32)x_ptr[ subfr_length - n - 1 ], 17 ); /* Q17 */ for( k = 0; k < n; k++ ) { C_first_row[ k ] = silk_MLA( C_first_row[ k ], x1, x_ptr[ n - k - 1 ] ); /* Q( -rshifts ) */ C_last_row[ k ] = silk_MLA( C_last_row[ k ], x2, x_ptr[ subfr_length - n + k ] ); /* Q( -rshifts ) */ Atmp1 = silk_RSHIFT_ROUND( Af_QA[ k ], QA - 17 ); /* Q17 */ tmp1 = silk_MLA( tmp1, x_ptr[ n - k - 1 ], Atmp1 ); /* Q17 */ tmp2 = silk_MLA( tmp2, x_ptr[ subfr_length - n + k ], Atmp1 ); /* Q17 */ } tmp1 = -tmp1; /* Q17 */ tmp2 = -tmp2; /* Q17 */ for( k = 0; k <= n; k++ ) { CAf[ k ] = silk_SMLAWW( CAf[ k ], tmp1, silk_LSHIFT32( (opus_int32)x_ptr[ n - k ], -rshifts - 1 ) ); /* Q( -rshift ) */ CAb[ k ] = silk_SMLAWW( CAb[ k ], tmp2, silk_LSHIFT32( (opus_int32)x_ptr[ subfr_length - n + k - 1 ], -rshifts - 1 ) ); /* Q( -rshift ) */ } } } /* Calculate nominator and denominator for the next order reflection (parcor) coefficient */ tmp1 = C_first_row[ n ]; /* Q( -rshifts ) */ tmp2 = C_last_row[ n ]; /* Q( -rshifts ) */ num = 0; /* Q( -rshifts ) */ nrg = silk_ADD32( CAb[ 0 ], CAf[ 0 ] ); /* Q( 1-rshifts ) */ for( k = 0; k < n; k++ ) { Atmp_QA = Af_QA[ k ]; lz = silk_CLZ32( silk_abs( Atmp_QA ) ) - 1; lz = silk_min( 32 - QA, lz ); Atmp1 = silk_LSHIFT32( Atmp_QA, lz ); /* Q( QA + lz ) */ tmp1 = silk_ADD_LSHIFT32( tmp1, silk_SMMUL( C_last_row[ n - k - 1 ], Atmp1 ), 32 - QA - lz ); /* Q( -rshifts ) */ tmp2 = silk_ADD_LSHIFT32( tmp2, silk_SMMUL( C_first_row[ n - k - 1 ], Atmp1 ), 32 - QA - lz ); /* Q( -rshifts ) */ num = silk_ADD_LSHIFT32( num, silk_SMMUL( CAb[ n - k ], Atmp1 ), 32 - QA - lz ); /* Q( -rshifts ) */ nrg = silk_ADD_LSHIFT32( nrg, silk_SMMUL( silk_ADD32( CAb[ k + 1 ], CAf[ k + 1 ] ), Atmp1 ), 32 - QA - lz ); /* Q( 1-rshifts ) */ } CAf[ n + 1 ] = tmp1; /* Q( -rshifts ) */ CAb[ n + 1 ] = tmp2; /* Q( -rshifts ) */ num = silk_ADD32( num, tmp2 ); /* Q( -rshifts ) */ num = silk_LSHIFT32( -num, 1 ); /* Q( 1-rshifts ) */ /* Calculate the next order reflection (parcor) coefficient */ if( silk_abs( num ) < nrg ) { rc_Q31 = silk_DIV32_varQ( num, nrg, 31 ); } else { rc_Q31 = ( num > 0 ) ? silk_int32_MAX : silk_int32_MIN; } /* Update inverse prediction gain */ tmp1 = ( (opus_int32)1 << 30 ) - silk_SMMUL( rc_Q31, rc_Q31 ); tmp1 = silk_LSHIFT( silk_SMMUL( invGain_Q30, tmp1 ), 2 ); if( tmp1 <= minInvGain_Q30 ) { /* Max prediction gain exceeded; set reflection coefficient such that max prediction gain is exactly hit */ tmp2 = ( (opus_int32)1 << 30 ) - silk_DIV32_varQ( minInvGain_Q30, invGain_Q30, 30 ); /* Q30 */ rc_Q31 = silk_SQRT_APPROX( tmp2 ); /* Q15 */ /* Newton-Raphson iteration */ rc_Q31 = silk_RSHIFT32( rc_Q31 + silk_DIV32( tmp2, rc_Q31 ), 1 ); /* Q15 */ rc_Q31 = silk_LSHIFT32( rc_Q31, 16 ); /* Q31 */ if( num < 0 ) { /* Ensure adjusted reflection coefficients has the original sign */ rc_Q31 = -rc_Q31; } invGain_Q30 = minInvGain_Q30; reached_max_gain = 1; } else { invGain_Q30 = tmp1; } /* Update the AR coefficients */ for( k = 0; k < (n + 1) >> 1; k++ ) { tmp1 = Af_QA[ k ]; /* QA */ tmp2 = Af_QA[ n - k - 1 ]; /* QA */ Af_QA[ k ] = silk_ADD_LSHIFT32( tmp1, silk_SMMUL( tmp2, rc_Q31 ), 1 ); /* QA */ Af_QA[ n - k - 1 ] = silk_ADD_LSHIFT32( tmp2, silk_SMMUL( tmp1, rc_Q31 ), 1 ); /* QA */ } Af_QA[ n ] = silk_RSHIFT32( rc_Q31, 31 - QA ); /* QA */ if( reached_max_gain ) { /* Reached max prediction gain; set remaining coefficients to zero and exit loop */ for( k = n + 1; k < D; k++ ) { Af_QA[ k ] = 0; } break; } /* Update C * Af and C * Ab */ for( k = 0; k <= n + 1; k++ ) { tmp1 = CAf[ k ]; /* Q( -rshifts ) */ tmp2 = CAb[ n - k + 1 ]; /* Q( -rshifts ) */ CAf[ k ] = silk_ADD_LSHIFT32( tmp1, silk_SMMUL( tmp2, rc_Q31 ), 1 ); /* Q( -rshifts ) */ CAb[ n - k + 1 ] = silk_ADD_LSHIFT32( tmp2, silk_SMMUL( tmp1, rc_Q31 ), 1 ); /* Q( -rshifts ) */ } } if( reached_max_gain ) { for( k = 0; k < D; k++ ) { /* Scale coefficients */ A_Q16[ k ] = -silk_RSHIFT_ROUND( Af_QA[ k ], QA - 16 ); } /* Subtract energy of preceding samples from C0 */ if( rshifts > 0 ) { for( s = 0; s < nb_subfr; s++ ) { x_ptr = x + s * subfr_length; C0 -= (opus_int32)silk_RSHIFT64( silk_inner_prod16_aligned_64( x_ptr, x_ptr, D ), rshifts ); } } else { for( s = 0; s < nb_subfr; s++ ) { x_ptr = x + s * subfr_length; C0 -= silk_LSHIFT32( silk_inner_prod_aligned( x_ptr, x_ptr, D ), -rshifts ); } } /* Approximate residual energy */ *res_nrg = silk_LSHIFT( silk_SMMUL( invGain_Q30, C0 ), 2 ); *res_nrg_Q = -rshifts; } else { /* Return residual energy */ nrg = CAf[ 0 ]; /* Q( -rshifts ) */ tmp1 = (opus_int32)1 << 16; /* Q16 */ for( k = 0; k < D; k++ ) { Atmp1 = silk_RSHIFT_ROUND( Af_QA[ k ], QA - 16 ); /* Q16 */ nrg = silk_SMLAWW( nrg, CAf[ k + 1 ], Atmp1 ); /* Q( -rshifts ) */ tmp1 = silk_SMLAWW( tmp1, Atmp1, Atmp1 ); /* Q16 */ A_Q16[ k ] = -Atmp1; } *res_nrg = silk_SMLAWW( nrg, silk_SMMUL( FIND_LPC_COND_FAC, C0 ), -tmp1 ); /* Q( -rshifts ) */ *res_nrg_Q = -rshifts; } }
/* Entropy constrained matrix-weighted VQ, hard-coded to 5-element vectors, for a single input data vector */ void silk_VQ_WMat_EC_c( opus_int8 *ind, /* O index of best codebook vector */ opus_int32 *res_nrg_Q15, /* O best residual energy */ opus_int32 *rate_dist_Q8, /* O best total bitrate */ opus_int *gain_Q7, /* O sum of absolute LTP coefficients */ const opus_int32 *XX_Q17, /* I correlation matrix */ const opus_int32 *xX_Q17, /* I correlation vector */ const opus_int8 *cb_Q7, /* I codebook */ const opus_uint8 *cb_gain_Q7, /* I codebook effective gain */ const opus_uint8 *cl_Q5, /* I code length for each codebook vector */ const opus_int subfr_len, /* I number of samples per subframe */ const opus_int32 max_gain_Q7, /* I maximum sum of absolute LTP coefficients */ const opus_int L /* I number of vectors in codebook */ ) { opus_int k, gain_tmp_Q7; const opus_int8 *cb_row_Q7; opus_int32 neg_xX_Q24[ 5 ]; opus_int32 sum1_Q15, sum2_Q24; opus_int32 bits_res_Q8, bits_tot_Q8; /* Negate and convert to new Q domain */ neg_xX_Q24[ 0 ] = -silk_LSHIFT32( xX_Q17[ 0 ], 7 ); neg_xX_Q24[ 1 ] = -silk_LSHIFT32( xX_Q17[ 1 ], 7 ); neg_xX_Q24[ 2 ] = -silk_LSHIFT32( xX_Q17[ 2 ], 7 ); neg_xX_Q24[ 3 ] = -silk_LSHIFT32( xX_Q17[ 3 ], 7 ); neg_xX_Q24[ 4 ] = -silk_LSHIFT32( xX_Q17[ 4 ], 7 ); /* Loop over codebook */ *rate_dist_Q8 = silk_int32_MAX; *res_nrg_Q15 = silk_int32_MAX; cb_row_Q7 = cb_Q7; /* In things go really bad, at least *ind is set to something safe. */ *ind = 0; for( k = 0; k < L; k++ ) { opus_int32 penalty; gain_tmp_Q7 = cb_gain_Q7[k]; /* Weighted rate */ /* Quantization error: 1 - 2 * xX * cb + cb' * XX * cb */ sum1_Q15 = SILK_FIX_CONST( 1.001, 15 ); /* Penalty for too large gain */ penalty = silk_LSHIFT32( silk_max( silk_SUB32( gain_tmp_Q7, max_gain_Q7 ), 0 ), 11 ); /* first row of XX_Q17 */ sum2_Q24 = silk_MLA( neg_xX_Q24[ 0 ], XX_Q17[ 1 ], cb_row_Q7[ 1 ] ); sum2_Q24 = silk_MLA( sum2_Q24, XX_Q17[ 2 ], cb_row_Q7[ 2 ] ); sum2_Q24 = silk_MLA( sum2_Q24, XX_Q17[ 3 ], cb_row_Q7[ 3 ] ); sum2_Q24 = silk_MLA( sum2_Q24, XX_Q17[ 4 ], cb_row_Q7[ 4 ] ); sum2_Q24 = silk_LSHIFT32( sum2_Q24, 1 ); sum2_Q24 = silk_MLA( sum2_Q24, XX_Q17[ 0 ], cb_row_Q7[ 0 ] ); sum1_Q15 = silk_SMLAWB( sum1_Q15, sum2_Q24, cb_row_Q7[ 0 ] ); /* second row of XX_Q17 */ sum2_Q24 = silk_MLA( neg_xX_Q24[ 1 ], XX_Q17[ 7 ], cb_row_Q7[ 2 ] ); sum2_Q24 = silk_MLA( sum2_Q24, XX_Q17[ 8 ], cb_row_Q7[ 3 ] ); sum2_Q24 = silk_MLA( sum2_Q24, XX_Q17[ 9 ], cb_row_Q7[ 4 ] ); sum2_Q24 = silk_LSHIFT32( sum2_Q24, 1 ); sum2_Q24 = silk_MLA( sum2_Q24, XX_Q17[ 6 ], cb_row_Q7[ 1 ] ); sum1_Q15 = silk_SMLAWB( sum1_Q15, sum2_Q24, cb_row_Q7[ 1 ] ); /* third row of XX_Q17 */ sum2_Q24 = silk_MLA( neg_xX_Q24[ 2 ], XX_Q17[ 13 ], cb_row_Q7[ 3 ] ); sum2_Q24 = silk_MLA( sum2_Q24, XX_Q17[ 14 ], cb_row_Q7[ 4 ] ); sum2_Q24 = silk_LSHIFT32( sum2_Q24, 1 ); sum2_Q24 = silk_MLA( sum2_Q24, XX_Q17[ 12 ], cb_row_Q7[ 2 ] ); sum1_Q15 = silk_SMLAWB( sum1_Q15, sum2_Q24, cb_row_Q7[ 2 ] ); /* fourth row of XX_Q17 */ sum2_Q24 = silk_MLA( neg_xX_Q24[ 3 ], XX_Q17[ 19 ], cb_row_Q7[ 4 ] ); sum2_Q24 = silk_LSHIFT32( sum2_Q24, 1 ); sum2_Q24 = silk_MLA( sum2_Q24, XX_Q17[ 18 ], cb_row_Q7[ 3 ] ); sum1_Q15 = silk_SMLAWB( sum1_Q15, sum2_Q24, cb_row_Q7[ 3 ] ); /* last row of XX_Q17 */ sum2_Q24 = silk_LSHIFT32( neg_xX_Q24[ 4 ], 1 ); sum2_Q24 = silk_MLA( sum2_Q24, XX_Q17[ 24 ], cb_row_Q7[ 4 ] ); sum1_Q15 = silk_SMLAWB( sum1_Q15, sum2_Q24, cb_row_Q7[ 4 ] ); /* find best */ if( sum1_Q15 >= 0 ) { /* Translate residual energy to bits using high-rate assumption (6 dB ==> 1 bit/sample) */ bits_res_Q8 = silk_SMULBB( subfr_len, silk_lin2log( sum1_Q15 + penalty) - (15 << 7) ); /* In the following line we reduce the codelength component by half ("-1"); seems to slghtly improve quality */ bits_tot_Q8 = silk_ADD_LSHIFT32( bits_res_Q8, cl_Q5[ k ], 3-1 ); if( bits_tot_Q8 <= *rate_dist_Q8 ) { *rate_dist_Q8 = bits_tot_Q8; *res_nrg_Q15 = sum1_Q15 + penalty; *ind = (opus_int8)k; *gain_Q7 = gain_tmp_Q7; } } /* Go to next cbk vector */ cb_row_Q7 += LTP_ORDER; } }