Exemplo n.º 1
0
/* Uses SMULL(), available on armv4                                     */
opus_int32 silk_schur64(                            /* O    returns residual energy                                     */
    opus_int32                  rc_Q16[],           /* O    Reflection coefficients [order] Q16                         */
    const opus_int32            c[],                /* I    Correlations [order+1]                                      */
    opus_int32                  order               /* I    Prediction order                                            */
)
{
    opus_int   k, n;
    opus_int32 C[ SILK_MAX_ORDER_LPC + 1 ][ 2 ];
    opus_int32 Ctmp1_Q30, Ctmp2_Q30, rc_tmp_Q31;

    silk_assert( order==6||order==8||order==10||order==12||order==14||order==16 );

    /* Check for invalid input */
    if( c[ 0 ] <= 0 ) {
        silk_memset( rc_Q16, 0, order * sizeof( opus_int32 ) );
        return 0;
    }

    for( k = 0; k < order + 1; k++ ) {
        C[ k ][ 0 ] = C[ k ][ 1 ] = c[ k ];
    }

    for( k = 0; k < order; k++ ) {
        /* Check that we won't be getting an unstable rc, otherwise stop here. */
        if (silk_abs_int32(C[ k + 1 ][ 0 ]) >= C[ 0 ][ 1 ]) {
           if ( C[ k + 1 ][ 0 ] > 0 ) {
              rc_Q16[ k ] = -SILK_FIX_CONST( .99f, 16 );
           } else {
              rc_Q16[ k ] = SILK_FIX_CONST( .99f, 16 );
           }
           k++;
           break;
        }

        /* Get reflection coefficient: divide two Q30 values and get result in Q31 */
        rc_tmp_Q31 = silk_DIV32_varQ( -C[ k + 1 ][ 0 ], C[ 0 ][ 1 ], 31 );

        /* Save the output */
        rc_Q16[ k ] = silk_RSHIFT_ROUND( rc_tmp_Q31, 15 );

        /* Update correlations */
        for( n = 0; n < order - k; n++ ) {
            Ctmp1_Q30 = C[ n + k + 1 ][ 0 ];
            Ctmp2_Q30 = C[ n ][ 1 ];

            /* Multiply and add the highest int32 */
            C[ n + k + 1 ][ 0 ] = Ctmp1_Q30 + silk_SMMUL( silk_LSHIFT( Ctmp2_Q30, 1 ), rc_tmp_Q31 );
            C[ n ][ 1 ]         = Ctmp2_Q30 + silk_SMMUL( silk_LSHIFT( Ctmp1_Q30, 1 ), rc_tmp_Q31 );
        }
    }

    for(; k < order; k++ ) {
       rc_Q16[ k ] = 0;
    }

    return silk_max_32( 1, C[ 0 ][ 1 ] );
}
/* Processing of gains */
void silk_process_gains_FIX(
    silk_encoder_state_FIX          *psEnc,                                 /* I/O  Encoder state                                                               */
    silk_encoder_control_FIX        *psEncCtrl,                             /* I/O  Encoder control                                                             */
    opus_int                        condCoding                              /* I    The type of conditional coding to use                                       */
)
{
    silk_shape_state_FIX *psShapeSt = &psEnc->sShape;
    opus_int     k;
    opus_int32   s_Q16, InvMaxSqrVal_Q16, gain, gain_squared, ResNrg, ResNrgPart, quant_offset_Q10;

    /* Gain reduction when LTP coding gain is high */
    if( psEnc->sCmn.indices.signalType == TYPE_VOICED ) {
        /*s = -0.5f * silk_sigmoid( 0.25f * ( psEncCtrl->LTPredCodGain - 12.0f ) ); */
        s_Q16 = -silk_sigm_Q15( silk_RSHIFT_ROUND( psEncCtrl->LTPredCodGain_Q7 - SILK_FIX_CONST( 12.0, 7 ), 4 ) );
        for( k = 0; k < psEnc->sCmn.nb_subfr; k++ ) {
            psEncCtrl->Gains_Q16[ k ] = silk_SMLAWB( psEncCtrl->Gains_Q16[ k ], psEncCtrl->Gains_Q16[ k ], s_Q16 );
        }
    }

    /* Limit the quantized signal */
    /* InvMaxSqrVal = pow( 2.0f, 0.33f * ( 21.0f - SNR_dB ) ) / subfr_length; */
    InvMaxSqrVal_Q16 = silk_DIV32_16( silk_log2lin(
        silk_SMULWB( SILK_FIX_CONST( 21 + 16 / 0.33, 7 ) - psEnc->sCmn.SNR_dB_Q7, SILK_FIX_CONST( 0.33, 16 ) ) ), psEnc->sCmn.subfr_length );

    for( k = 0; k < psEnc->sCmn.nb_subfr; k++ ) {
        /* Soft limit on ratio residual energy and squared gains */
        ResNrg     = psEncCtrl->ResNrg[ k ];
        ResNrgPart = silk_SMULWW( ResNrg, InvMaxSqrVal_Q16 );
        if( psEncCtrl->ResNrgQ[ k ] > 0 ) {
            ResNrgPart = silk_RSHIFT_ROUND( ResNrgPart, psEncCtrl->ResNrgQ[ k ] );
        } else {
            if( ResNrgPart >= silk_RSHIFT( silk_int32_MAX, -psEncCtrl->ResNrgQ[ k ] ) ) {
                ResNrgPart = silk_int32_MAX;
            } else {
                ResNrgPart = silk_LSHIFT( ResNrgPart, -psEncCtrl->ResNrgQ[ k ] );
            }
        }
        gain = psEncCtrl->Gains_Q16[ k ];
        gain_squared = silk_ADD_SAT32( ResNrgPart, silk_SMMUL( gain, gain ) );
        if( gain_squared < silk_int16_MAX ) {
            /* recalculate with higher precision */
            gain_squared = silk_SMLAWW( silk_LSHIFT( ResNrgPart, 16 ), gain, gain );
            silk_assert( gain_squared > 0 );
            gain = silk_SQRT_APPROX( gain_squared );                    /* Q8   */
            gain = silk_min( gain, silk_int32_MAX >> 8 );
            psEncCtrl->Gains_Q16[ k ] = silk_LSHIFT_SAT32( gain, 8 );   /* Q16  */
        } else {
Exemplo n.º 3
0
static OPUS_INLINE void silk_LS_divide_Q16_FIX(
    opus_int32          T[],        /* I/O  Numenator vector                                            */
    inv_D_t             *inv_D,     /* I    1 / D vector                                                */
    opus_int            M           /* I    dimension                                                   */
)
{
    opus_int   i;
    opus_int32 tmp_32;
    opus_int32 one_div_diag_Q36, one_div_diag_Q48;

    for( i = 0; i < M; i++ ) {
        one_div_diag_Q36 = inv_D[ i ].Q36_part;
        one_div_diag_Q48 = inv_D[ i ].Q48_part;

        tmp_32 = T[ i ];
        T[ i ] = silk_ADD32( silk_SMMUL( tmp_32, one_div_diag_Q48 ), silk_RSHIFT( silk_SMULWW( tmp_32, one_div_diag_Q36 ), 4 ) );
    }
}
Exemplo n.º 4
0
static OPUS_INLINE void silk_LDL_factorize_FIX(
    opus_int32          *A,         /* I/O Pointer to Symetric Square Matrix                            */
    opus_int            M,          /* I   Size of Matrix                                               */
    opus_int32          *L_Q16,     /* I/O Pointer to Square Upper triangular Matrix                    */
    inv_D_t             *inv_D      /* I/O Pointer to vector holding inverted diagonal elements of D    */
)
{
    opus_int   i, j, k, status, loop_count;
    const opus_int32 *ptr1, *ptr2;
    opus_int32 diag_min_value, tmp_32, err;
    opus_int32 v_Q0[ MAX_MATRIX_SIZE ], D_Q0[ MAX_MATRIX_SIZE ];
    opus_int32 one_div_diag_Q36, one_div_diag_Q40, one_div_diag_Q48;

    silk_assert( M <= MAX_MATRIX_SIZE );

    status = 1;
    diag_min_value = silk_max_32( silk_SMMUL( silk_ADD_SAT32( A[ 0 ], A[ silk_SMULBB( M, M ) - 1 ] ), SILK_FIX_CONST( FIND_LTP_COND_FAC, 31 ) ), 1 << 9 );
    for( loop_count = 0; loop_count < M && status == 1; loop_count++ ) {
        status = 0;
        for( j = 0; j < M; j++ ) {
            ptr1 = matrix_adr( L_Q16, j, 0, M );
            tmp_32 = 0;
            for( i = 0; i < j; i++ ) {
                v_Q0[ i ] = silk_SMULWW(         D_Q0[ i ], ptr1[ i ] ); /* Q0 */
                tmp_32    = silk_SMLAWW( tmp_32, v_Q0[ i ], ptr1[ i ] ); /* Q0 */
            }
            tmp_32 = silk_SUB32( matrix_ptr( A, j, j, M ), tmp_32 );

            if( tmp_32 < diag_min_value ) {
                tmp_32 = silk_SUB32( silk_SMULBB( loop_count + 1, diag_min_value ), tmp_32 );
                /* Matrix not positive semi-definite, or ill conditioned */
                for( i = 0; i < M; i++ ) {
                    matrix_ptr( A, i, i, M ) = silk_ADD32( matrix_ptr( A, i, i, M ), tmp_32 );
                }
                status = 1;
                break;
            }
            D_Q0[ j ] = tmp_32;                         /* always < max(Correlation) */

            /* two-step division */
            one_div_diag_Q36 = silk_INVERSE32_varQ( tmp_32, 36 );                    /* Q36 */
            one_div_diag_Q40 = silk_LSHIFT( one_div_diag_Q36, 4 );                   /* Q40 */
            err = silk_SUB32( (opus_int32)1 << 24, silk_SMULWW( tmp_32, one_div_diag_Q40 ) );     /* Q24 */
            one_div_diag_Q48 = silk_SMULWW( err, one_div_diag_Q40 );                 /* Q48 */

            /* Save 1/Ds */
            inv_D[ j ].Q36_part = one_div_diag_Q36;
            inv_D[ j ].Q48_part = one_div_diag_Q48;

            matrix_ptr( L_Q16, j, j, M ) = 65536; /* 1.0 in Q16 */
            ptr1 = matrix_adr( A, j, 0, M );
            ptr2 = matrix_adr( L_Q16, j + 1, 0, M );
            for( i = j + 1; i < M; i++ ) {
                tmp_32 = 0;
                for( k = 0; k < j; k++ ) {
                    tmp_32 = silk_SMLAWW( tmp_32, v_Q0[ k ], ptr2[ k ] ); /* Q0 */
                }
                tmp_32 = silk_SUB32( ptr1[ i ], tmp_32 ); /* always < max(Correlation) */

                /* tmp_32 / D_Q0[j] : Divide to Q16 */
                matrix_ptr( L_Q16, i, j, M ) = silk_ADD32( silk_SMMUL( tmp_32, one_div_diag_Q48 ),
                    silk_RSHIFT( silk_SMULWW( tmp_32, one_div_diag_Q36 ), 4 ) );

                /* go to next column */
                ptr2 += M;
            }
        }
    }

    silk_assert( status == 0 );
}
Exemplo n.º 5
0
/* test if LPC coefficients are stable (all poles within unit circle)   */
static opus_int32 LPC_inverse_pred_gain_QA_c(               /* O   Returns inverse prediction gain in energy domain, Q30    */
    opus_int32           A_QA[ SILK_MAX_ORDER_LPC ],        /* I   Prediction coefficients                                  */
    const opus_int       order                              /* I   Prediction order                                         */
)
{
    opus_int   k, n, mult2Q;
    opus_int32 invGain_Q30, rc_Q31, rc_mult1_Q30, rc_mult2, tmp1, tmp2;

    invGain_Q30 = SILK_FIX_CONST( 1, 30 );
    for( k = order - 1; k > 0; k-- ) {
        /* Check for stability */
        if( ( A_QA[ k ] > A_LIMIT ) || ( A_QA[ k ] < -A_LIMIT ) ) {
            return 0;
        }

        /* Set RC equal to negated AR coef */
        rc_Q31 = -silk_LSHIFT( A_QA[ k ], 31 - QA );

        /* rc_mult1_Q30 range: [ 1 : 2^30 ] */
        rc_mult1_Q30 = silk_SUB32( SILK_FIX_CONST( 1, 30 ), silk_SMMUL( rc_Q31, rc_Q31 ) );
        silk_assert( rc_mult1_Q30 > ( 1 << 15 ) );                   /* reduce A_LIMIT if fails */
        silk_assert( rc_mult1_Q30 <= ( 1 << 30 ) );

        /* Update inverse gain */
        /* invGain_Q30 range: [ 0 : 2^30 ] */
        invGain_Q30 = silk_LSHIFT( silk_SMMUL( invGain_Q30, rc_mult1_Q30 ), 2 );
        silk_assert( invGain_Q30 >= 0           );
        silk_assert( invGain_Q30 <= ( 1 << 30 ) );
        if( invGain_Q30 < SILK_FIX_CONST( 1.0f / MAX_PREDICTION_POWER_GAIN, 30 ) ) {
            return 0;
        }

        /* rc_mult2 range: [ 2^30 : silk_int32_MAX ] */
        mult2Q = 32 - silk_CLZ32( silk_abs( rc_mult1_Q30 ) );
        rc_mult2 = silk_INVERSE32_varQ( rc_mult1_Q30, mult2Q + 30 );

        /* Update AR coefficient */
        for( n = 0; n < (k + 1) >> 1; n++ ) {
            opus_int64 tmp64;
            tmp1 = A_QA[ n ];
            tmp2 = A_QA[ k - n - 1 ];
            tmp64 = silk_RSHIFT_ROUND64( silk_SMULL( silk_SUB_SAT32(tmp1,
                  MUL32_FRAC_Q( tmp2, rc_Q31, 31 ) ), rc_mult2 ), mult2Q);
            if( tmp64 > silk_int32_MAX || tmp64 < silk_int32_MIN ) {
               return 0;
            }
            A_QA[ n ] = ( opus_int32 )tmp64;
            tmp64 = silk_RSHIFT_ROUND64( silk_SMULL( silk_SUB_SAT32(tmp2,
                  MUL32_FRAC_Q( tmp1, rc_Q31, 31 ) ), rc_mult2), mult2Q);
            if( tmp64 > silk_int32_MAX || tmp64 < silk_int32_MIN ) {
               return 0;
            }
            A_QA[ k - n - 1 ] = ( opus_int32 )tmp64;
        }
    }

    /* Check for stability */
    if( ( A_QA[ k ] > A_LIMIT ) || ( A_QA[ k ] < -A_LIMIT ) ) {
        return 0;
    }

    /* Set RC equal to negated AR coef */
    rc_Q31 = -silk_LSHIFT( A_QA[ 0 ], 31 - QA );

    /* Range: [ 1 : 2^30 ] */
    rc_mult1_Q30 = silk_SUB32( SILK_FIX_CONST( 1, 30 ), silk_SMMUL( rc_Q31, rc_Q31 ) );

    /* Update inverse gain */
    /* Range: [ 0 : 2^30 ] */
    invGain_Q30 = silk_LSHIFT( silk_SMMUL( invGain_Q30, rc_mult1_Q30 ), 2 );
    silk_assert( invGain_Q30 >= 0           );
    silk_assert( invGain_Q30 <= ( 1 << 30 ) );
    if( invGain_Q30 < SILK_FIX_CONST( 1.0f / MAX_PREDICTION_POWER_GAIN, 30 ) ) {
        return 0;
    }

    return invGain_Q30;
}
Exemplo n.º 6
0
/* test if LPC coefficients are stable (all poles within unit circle)   */
static opus_int32 LPC_inverse_pred_gain_QA(                 /* O   Returns inverse prediction gain in energy domain, Q30    */
    opus_int32           A_QA[ 2 ][ SILK_MAX_ORDER_LPC ],   /* I   Prediction coefficients                                  */
    const opus_int       order                              /* I   Prediction order                                         */
)
{
    opus_int   k, n, mult2Q;
    opus_int32 invGain_Q30, rc_Q31, rc_mult1_Q30, rc_mult2, tmp_QA;
    opus_int32 *Aold_QA, *Anew_QA;

    Anew_QA = A_QA[ order & 1 ];

    invGain_Q30 = (opus_int32)1 << 30;
    for( k = order - 1; k > 0; k-- ) {
        /* Check for stability */
        if( ( Anew_QA[ k ] > A_LIMIT ) || ( Anew_QA[ k ] < -A_LIMIT ) ) {
            return 0;
        }

        /* Set RC equal to negated AR coef */
        rc_Q31 = -silk_LSHIFT( Anew_QA[ k ], 31 - QA );

        /* rc_mult1_Q30 range: [ 1 : 2^30 ] */
        rc_mult1_Q30 = ( (opus_int32)1 << 30 ) - silk_SMMUL( rc_Q31, rc_Q31 );
        silk_assert( rc_mult1_Q30 > ( 1 << 15 ) );                   /* reduce A_LIMIT if fails */
        silk_assert( rc_mult1_Q30 <= ( 1 << 30 ) );

        /* rc_mult2 range: [ 2^30 : silk_int32_MAX ] */
        mult2Q = 32 - silk_CLZ32( silk_abs( rc_mult1_Q30 ) );
        rc_mult2 = silk_INVERSE32_varQ( rc_mult1_Q30, mult2Q + 30 );

        /* Update inverse gain */
        /* invGain_Q30 range: [ 0 : 2^30 ] */
        invGain_Q30 = silk_LSHIFT( silk_SMMUL( invGain_Q30, rc_mult1_Q30 ), 2 );
        silk_assert( invGain_Q30 >= 0           );
        silk_assert( invGain_Q30 <= ( 1 << 30 ) );

        /* Swap pointers */
        Aold_QA = Anew_QA;
        Anew_QA = A_QA[ k & 1 ];

        /* Update AR coefficient */
        for( n = 0; n < k; n++ ) {
            tmp_QA = Aold_QA[ n ] - MUL32_FRAC_Q( Aold_QA[ k - n - 1 ], rc_Q31, 31 );
            Anew_QA[ n ] = MUL32_FRAC_Q( tmp_QA, rc_mult2 , mult2Q );
        }
    }

    /* Check for stability */
    if( ( Anew_QA[ 0 ] > A_LIMIT ) || ( Anew_QA[ 0 ] < -A_LIMIT ) ) {
        return 0;
    }

    /* Set RC equal to negated AR coef */
    rc_Q31 = -silk_LSHIFT( Anew_QA[ 0 ], 31 - QA );

    /* Range: [ 1 : 2^30 ] */
    rc_mult1_Q30 = ( (opus_int32)1 << 30 ) - silk_SMMUL( rc_Q31, rc_Q31 );

    /* Update inverse gain */
    /* Range: [ 0 : 2^30 ] */
    invGain_Q30 = silk_LSHIFT( silk_SMMUL( invGain_Q30, rc_mult1_Q30 ), 2 );
    silk_assert( invGain_Q30 >= 0     );
    silk_assert( invGain_Q30 <= 1<<30 );

    return invGain_Q30;
}
Exemplo n.º 7
0
/* Compute reflection coefficients from input signal */
void silk_burg_modified(
    opus_int32                  *res_nrg,           /* O    Residual energy                                             */
    opus_int                    *res_nrg_Q,         /* O    Residual energy Q value                                     */
    opus_int32                  A_Q16[],            /* O    Prediction coefficients (length order)                      */
    const opus_int16            x[],                /* I    Input signal, length: nb_subfr * ( D + subfr_length )       */
    const opus_int              subfr_length,       /* I    Input signal subframe length (incl. D preceeding samples)   */
    const opus_int              nb_subfr,           /* I    Number of subframes stacked in x                            */
    const opus_int32            WhiteNoiseFrac_Q32, /* I    Fraction added to zero-lag autocorrelation                  */
    const opus_int              D                   /* I    Order                                                       */
)
{
    opus_int         k, n, s, lz, rshifts, rshifts_extra;
    opus_int32       C0, num, nrg, rc_Q31, Atmp_QA, Atmp1, tmp1, tmp2, x1, x2;
    const opus_int16 *x_ptr;

    opus_int32       C_first_row[ SILK_MAX_ORDER_LPC ];
    opus_int32       C_last_row[  SILK_MAX_ORDER_LPC ];
    opus_int32       Af_QA[       SILK_MAX_ORDER_LPC ];

    opus_int32       CAf[ SILK_MAX_ORDER_LPC + 1 ];
    opus_int32       CAb[ SILK_MAX_ORDER_LPC + 1 ];

    silk_assert( subfr_length * nb_subfr <= MAX_FRAME_SIZE );
    silk_assert( nb_subfr <= MAX_NB_SUBFR );


    /* Compute autocorrelations, added over subframes */
    silk_sum_sqr_shift( &C0, &rshifts, x, nb_subfr * subfr_length );
    if( rshifts > MAX_RSHIFTS ) {
        C0 = silk_LSHIFT32( C0, rshifts - MAX_RSHIFTS );
        silk_assert( C0 > 0 );
        rshifts = MAX_RSHIFTS;
    } else {
        lz = silk_CLZ32( C0 ) - 1;
        rshifts_extra = N_BITS_HEAD_ROOM - lz;
        if( rshifts_extra > 0 ) {
            rshifts_extra = silk_min( rshifts_extra, MAX_RSHIFTS - rshifts );
            C0 = silk_RSHIFT32( C0, rshifts_extra );
        } else {
            rshifts_extra = silk_max( rshifts_extra, MIN_RSHIFTS - rshifts );
            C0 = silk_LSHIFT32( C0, -rshifts_extra );
        }
        rshifts += rshifts_extra;
    }
    silk_memset( C_first_row, 0, SILK_MAX_ORDER_LPC * sizeof( opus_int32 ) );
    if( rshifts > 0 ) {
        for( s = 0; s < nb_subfr; s++ ) {
            x_ptr = x + s * subfr_length;
            for( n = 1; n < D + 1; n++ ) {
                C_first_row[ n - 1 ] += (opus_int32)silk_RSHIFT64(
                    silk_inner_prod16_aligned_64( x_ptr, x_ptr + n, subfr_length - n ), rshifts );
            }
        }
    } else {
        for( s = 0; s < nb_subfr; s++ ) {
            x_ptr = x + s * subfr_length;
            for( n = 1; n < D + 1; n++ ) {
                C_first_row[ n - 1 ] += silk_LSHIFT32(
                    silk_inner_prod_aligned( x_ptr, x_ptr + n, subfr_length - n ), -rshifts );
            }
        }
    }
    silk_memcpy( C_last_row, C_first_row, SILK_MAX_ORDER_LPC * sizeof( opus_int32 ) );

    /* Initialize */
    CAb[ 0 ] = CAf[ 0 ] = C0 + silk_SMMUL( WhiteNoiseFrac_Q32, C0 ) + 1;                                /* Q(-rshifts)*/

    for( n = 0; n < D; n++ ) {
        /* Update first row of correlation matrix (without first element) */
        /* Update last row of correlation matrix (without last element, stored in reversed order) */
        /* Update C * Af */
        /* Update C * flipud(Af) (stored in reversed order) */
        if( rshifts > -2 ) {
            for( s = 0; s < nb_subfr; s++ ) {
                x_ptr = x + s * subfr_length;
                x1  = -silk_LSHIFT32( (opus_int32)x_ptr[ n ],                    16 - rshifts );        /* Q(16-rshifts)*/
                x2  = -silk_LSHIFT32( (opus_int32)x_ptr[ subfr_length - n - 1 ], 16 - rshifts );        /* Q(16-rshifts)*/
                tmp1 = silk_LSHIFT32( (opus_int32)x_ptr[ n ],                    QA - 16 );             /* Q(QA-16)*/
                tmp2 = silk_LSHIFT32( (opus_int32)x_ptr[ subfr_length - n - 1 ], QA - 16 );             /* Q(QA-16)*/
                for( k = 0; k < n; k++ ) {
                    C_first_row[ k ] = silk_SMLAWB( C_first_row[ k ], x1, x_ptr[ n - k - 1 ]            ); /* Q( -rshifts )*/
                    C_last_row[ k ]  = silk_SMLAWB( C_last_row[ k ],  x2, x_ptr[ subfr_length - n + k ] ); /* Q( -rshifts )*/
                    Atmp_QA = Af_QA[ k ];
                    tmp1 = silk_SMLAWB( tmp1, Atmp_QA, x_ptr[ n - k - 1 ]            );                 /* Q(QA-16)*/
                    tmp2 = silk_SMLAWB( tmp2, Atmp_QA, x_ptr[ subfr_length - n + k ] );                 /* Q(QA-16)*/
                }
                tmp1 = silk_LSHIFT32( -tmp1, 32 - QA - rshifts );                                       /* Q(16-rshifts)*/
                tmp2 = silk_LSHIFT32( -tmp2, 32 - QA - rshifts );                                       /* Q(16-rshifts)*/
                for( k = 0; k <= n; k++ ) {
                    CAf[ k ] = silk_SMLAWB( CAf[ k ], tmp1, x_ptr[ n - k ]                    );        /* Q( -rshift )*/
                    CAb[ k ] = silk_SMLAWB( CAb[ k ], tmp2, x_ptr[ subfr_length - n + k - 1 ] );        /* Q( -rshift )*/
                }
            }
        } else {
            for( s = 0; s < nb_subfr; s++ ) {
                x_ptr = x + s * subfr_length;
                x1  = -silk_LSHIFT32( (opus_int32)x_ptr[ n ],                    -rshifts );            /* Q( -rshifts )*/
                x2  = -silk_LSHIFT32( (opus_int32)x_ptr[ subfr_length - n - 1 ], -rshifts );            /* Q( -rshifts )*/
                tmp1 = silk_LSHIFT32( (opus_int32)x_ptr[ n ],                    17 );                  /* Q17*/
                tmp2 = silk_LSHIFT32( (opus_int32)x_ptr[ subfr_length - n - 1 ], 17 );                  /* Q17*/
                for( k = 0; k < n; k++ ) {
                    C_first_row[ k ] = silk_MLA( C_first_row[ k ], x1, x_ptr[ n - k - 1 ]            ); /* Q( -rshifts )*/
                    C_last_row[ k ]  = silk_MLA( C_last_row[ k ],  x2, x_ptr[ subfr_length - n + k ] ); /* Q( -rshifts )*/
                    Atmp1 = silk_RSHIFT_ROUND( Af_QA[ k ], QA - 17 );                                   /* Q17*/
                    tmp1 = silk_MLA( tmp1, x_ptr[ n - k - 1 ],            Atmp1 );                      /* Q17*/
                    tmp2 = silk_MLA( tmp2, x_ptr[ subfr_length - n + k ], Atmp1 );                      /* Q17*/
                }
                tmp1 = -tmp1;                                                                           /* Q17*/
                tmp2 = -tmp2;                                                                           /* Q17*/
                for( k = 0; k <= n; k++ ) {
                    CAf[ k ] = silk_SMLAWW( CAf[ k ], tmp1,
                        silk_LSHIFT32( (opus_int32)x_ptr[ n - k ], -rshifts - 1 ) );                    /* Q( -rshift )*/
                    CAb[ k ] = silk_SMLAWW( CAb[ k ], tmp2,
                        silk_LSHIFT32( (opus_int32)x_ptr[ subfr_length - n + k - 1 ], -rshifts - 1 ) ); /* Q( -rshift )*/
                }
            }
        }

        /* Calculate nominator and denominator for the next order reflection (parcor) coefficient */
        tmp1 = C_first_row[ n ];                                                                        /* Q( -rshifts )*/
        tmp2 = C_last_row[ n ];                                                                         /* Q( -rshifts )*/
        num  = 0;                                                                                       /* Q( -rshifts )*/
        nrg  = silk_ADD32( CAb[ 0 ], CAf[ 0 ] );                                                        /* Q( 1-rshifts )*/
        for( k = 0; k < n; k++ ) {
            Atmp_QA = Af_QA[ k ];
            lz = silk_CLZ32( silk_abs( Atmp_QA ) ) - 1;
            lz = silk_min( 32 - QA, lz );
            Atmp1 = silk_LSHIFT32( Atmp_QA, lz );                                                       /* Q( QA + lz )*/

            tmp1 = silk_ADD_LSHIFT32( tmp1, silk_SMMUL( C_last_row[  n - k - 1 ], Atmp1 ), 32 - QA - lz );  /* Q( -rshifts )*/
            tmp2 = silk_ADD_LSHIFT32( tmp2, silk_SMMUL( C_first_row[ n - k - 1 ], Atmp1 ), 32 - QA - lz );  /* Q( -rshifts )*/
            num  = silk_ADD_LSHIFT32( num,  silk_SMMUL( CAb[ n - k ],             Atmp1 ), 32 - QA - lz );  /* Q( -rshifts )*/
            nrg  = silk_ADD_LSHIFT32( nrg,  silk_SMMUL( silk_ADD32( CAb[ k + 1 ], CAf[ k + 1 ] ),
                                                                                Atmp1 ), 32 - QA - lz );    /* Q( 1-rshifts )*/
        }
        CAf[ n + 1 ] = tmp1;                                                                            /* Q( -rshifts )*/
        CAb[ n + 1 ] = tmp2;                                                                            /* Q( -rshifts )*/
        num = silk_ADD32( num, tmp2 );                                                                  /* Q( -rshifts )*/
        num = silk_LSHIFT32( -num, 1 );                                                                 /* Q( 1-rshifts )*/

        /* Calculate the next order reflection (parcor) coefficient */
        if( silk_abs( num ) < nrg ) {
            rc_Q31 = silk_DIV32_varQ( num, nrg, 31 );
        } else {
            /* Negative energy or ratio too high; set remaining coefficients to zero and exit loop */
            silk_memset( &Af_QA[ n ], 0, ( D - n ) * sizeof( opus_int32 ) );
            silk_assert( 0 );
            break;
        }

        /* Update the AR coefficients */
        for( k = 0; k < (n + 1) >> 1; k++ ) {
            tmp1 = Af_QA[ k ];                                                                  /* QA*/
            tmp2 = Af_QA[ n - k - 1 ];                                                          /* QA*/
            Af_QA[ k ]         = silk_ADD_LSHIFT32( tmp1, silk_SMMUL( tmp2, rc_Q31 ), 1 );      /* QA*/
            Af_QA[ n - k - 1 ] = silk_ADD_LSHIFT32( tmp2, silk_SMMUL( tmp1, rc_Q31 ), 1 );      /* QA*/
        }
        Af_QA[ n ] = silk_RSHIFT32( rc_Q31, 31 - QA );                                          /* QA*/

        /* Update C * Af and C * Ab */
        for( k = 0; k <= n + 1; k++ ) {
            tmp1 = CAf[ k ];                                                                    /* Q( -rshifts )*/
            tmp2 = CAb[ n - k + 1 ];                                                            /* Q( -rshifts )*/
            CAf[ k ]         = silk_ADD_LSHIFT32( tmp1, silk_SMMUL( tmp2, rc_Q31 ), 1 );        /* Q( -rshifts )*/
            CAb[ n - k + 1 ] = silk_ADD_LSHIFT32( tmp2, silk_SMMUL( tmp1, rc_Q31 ), 1 );        /* Q( -rshifts )*/
        }
    }

    /* Return residual energy */
    nrg  = CAf[ 0 ];                                                                            /* Q( -rshifts )*/
    tmp1 = 1 << 16;                                                                             /* Q16*/
    for( k = 0; k < D; k++ ) {
        Atmp1 = silk_RSHIFT_ROUND( Af_QA[ k ], QA - 16 );                                       /* Q16*/
        nrg  = silk_SMLAWW( nrg, CAf[ k + 1 ], Atmp1 );                                         /* Q( -rshifts )*/
        tmp1 = silk_SMLAWW( tmp1, Atmp1, Atmp1 );                                               /* Q16*/
        A_Q16[ k ] = -Atmp1;
    }
    *res_nrg = silk_SMLAWW( nrg, silk_SMMUL( WhiteNoiseFrac_Q32, C0 ), -tmp1 );                 /* Q( -rshifts )*/
    *res_nrg_Q = -rshifts;
}
/* Compute reflection coefficients from input signal */
void silk_burg_modified_sse4_1(
    opus_int32                  *res_nrg,           /* O    Residual energy                                             */
    opus_int                    *res_nrg_Q,         /* O    Residual energy Q value                                     */
    opus_int32                  A_Q16[],            /* O    Prediction coefficients (length order)                      */
    const opus_int16            x[],                /* I    Input signal, length: nb_subfr * (D + subfr_length)       */
    const opus_int32            minInvGain_Q30,     /* I    Inverse of max prediction gain                              */
    const opus_int              subfr_length,       /* I    Input signal subframe length (incl. D preceding samples)    */
    const opus_int              nb_subfr,           /* I    Number of subframes stacked in x                            */
    const opus_int              D,                  /* I    Order                                                       */
    int                         arch                /* I    Run-time architecture                                       */
)
{
    opus_int         k, n, s, lz, rshifts, rshifts_extra, reached_max_gain;
    opus_int32       C0, num, nrg, rc_Q31, invGain_Q30, Atmp_QA, Atmp1, tmp1, tmp2, x1, x2;
    const opus_int16 *x_ptr;
    opus_int32       C_first_row[ SILK_MAX_ORDER_LPC ];
    opus_int32       C_last_row[  SILK_MAX_ORDER_LPC ];
    opus_int32       Af_QA[       SILK_MAX_ORDER_LPC ];
    opus_int32       CAf[ SILK_MAX_ORDER_LPC + 1 ];
    opus_int32       CAb[ SILK_MAX_ORDER_LPC + 1 ];
    opus_int32       xcorr[ SILK_MAX_ORDER_LPC ];

    __m128i FIRST_3210, LAST_3210, ATMP_3210, TMP1_3210, TMP2_3210, T1_3210, T2_3210, PTR_3210, SUBFR_3210, X1_3210, X2_3210;
    __m128i CONST1 = _mm_set1_epi32(1);

    silk_assert(subfr_length * nb_subfr <= MAX_FRAME_SIZE);

    /* Compute autocorrelations, added over subframes */
    silk_sum_sqr_shift(&C0, &rshifts, x, nb_subfr * subfr_length);
    if(rshifts > MAX_RSHIFTS) {
        C0 = silk_LSHIFT32(C0, rshifts - MAX_RSHIFTS);
        silk_assert(C0 > 0);
        rshifts = MAX_RSHIFTS;
    } else {
        lz = silk_CLZ32(C0) - 1;
        rshifts_extra = N_BITS_HEAD_ROOM - lz;
        if(rshifts_extra > 0) {
            rshifts_extra = silk_min(rshifts_extra, MAX_RSHIFTS - rshifts);
            C0 = silk_RSHIFT32(C0, rshifts_extra);
        } else {
            rshifts_extra = silk_max(rshifts_extra, MIN_RSHIFTS - rshifts);
            C0 = silk_LSHIFT32(C0, -rshifts_extra);
        }
        rshifts += rshifts_extra;
    }
    CAb[ 0 ] = CAf[ 0 ] = C0 + silk_SMMUL(SILK_FIX_CONST(FIND_LPC_COND_FAC, 32), C0) + 1;                                /* Q(-rshifts) */
    silk_memset(C_first_row, 0, SILK_MAX_ORDER_LPC * sizeof(opus_int32));
    if(rshifts > 0) {
        for(s = 0; s < nb_subfr; s++) {
            x_ptr = x + s * subfr_length;
            for(n = 1; n < D + 1; n++) {
                C_first_row[ n - 1 ] += (opus_int32)silk_RSHIFT64(
                    silk_inner_prod16_aligned_64(x_ptr, x_ptr + n, subfr_length - n, arch), rshifts);
            }
        }
    } else {
        for(s = 0; s < nb_subfr; s++) {
            int i;
            opus_int32 d;
            x_ptr = x + s * subfr_length;
            celt_pitch_xcorr(x_ptr, x_ptr + 1, xcorr, subfr_length - D, D, arch);
            for(n = 1; n < D + 1; n++) {
               for (i = n + subfr_length - D, d = 0; i < subfr_length; i++)
                  d = MAC16_16(d, x_ptr[ i ], x_ptr[ i - n ]);
               xcorr[ n - 1 ] += d;
            }
            for(n = 1; n < D + 1; n++) {
                C_first_row[ n - 1 ] += silk_LSHIFT32(xcorr[ n - 1 ], -rshifts);
            }
        }
    }
    silk_memcpy(C_last_row, C_first_row, SILK_MAX_ORDER_LPC * sizeof(opus_int32));

    /* Initialize */
    CAb[ 0 ] = CAf[ 0 ] = C0 + silk_SMMUL(SILK_FIX_CONST(FIND_LPC_COND_FAC, 32), C0) + 1;                                /* Q(-rshifts) */

    invGain_Q30 = (opus_int32)1 << 30;
    reached_max_gain = 0;
    for(n = 0; n < D; n++) {
        /* Update first row of correlation matrix (without first element) */
        /* Update last row of correlation matrix (without last element, stored in reversed order) */
        /* Update C * Af */
        /* Update C * flipud(Af) (stored in reversed order) */
        if(rshifts > -2) {
            for(s = 0; s < nb_subfr; s++) {
                x_ptr = x + s * subfr_length;
                x1  = -silk_LSHIFT32((opus_int32)x_ptr[ n ],                    16 - rshifts);        /* Q(16-rshifts) */
                x2  = -silk_LSHIFT32((opus_int32)x_ptr[ subfr_length - n - 1 ], 16 - rshifts);        /* Q(16-rshifts) */
                tmp1 = silk_LSHIFT32((opus_int32)x_ptr[ n ],                    QA - 16);             /* Q(QA-16) */
                tmp2 = silk_LSHIFT32((opus_int32)x_ptr[ subfr_length - n - 1 ], QA - 16);             /* Q(QA-16) */
                for(k = 0; k < n; k++) {
                    C_first_row[ k ] = silk_SMLAWB(C_first_row[ k ], x1, x_ptr[ n - k - 1 ]           ); /* Q(-rshifts) */
                    C_last_row[ k ]  = silk_SMLAWB(C_last_row[ k ],  x2, x_ptr[ subfr_length - n + k ]); /* Q(-rshifts) */
                    Atmp_QA = Af_QA[ k ];
                    tmp1 = silk_SMLAWB(tmp1, Atmp_QA, x_ptr[ n - k - 1 ]           );                 /* Q(QA-16) */
                    tmp2 = silk_SMLAWB(tmp2, Atmp_QA, x_ptr[ subfr_length - n + k ]);                 /* Q(QA-16) */
                }
                tmp1 = silk_LSHIFT32(-tmp1, 32 - QA - rshifts);                                       /* Q(16-rshifts) */
                tmp2 = silk_LSHIFT32(-tmp2, 32 - QA - rshifts);                                       /* Q(16-rshifts) */
                for(k = 0; k <= n; k++) {
                    CAf[ k ] = silk_SMLAWB(CAf[ k ], tmp1, x_ptr[ n - k ]                   );        /* Q(-rshift) */
                    CAb[ k ] = silk_SMLAWB(CAb[ k ], tmp2, x_ptr[ subfr_length - n + k - 1 ]);        /* Q(-rshift) */
                }
            }
        } else {
            for(s = 0; s < nb_subfr; s++) {
                x_ptr = x + s * subfr_length;
                x1  = -silk_LSHIFT32((opus_int32)x_ptr[ n ],                    -rshifts);            /* Q(-rshifts) */
                x2  = -silk_LSHIFT32((opus_int32)x_ptr[ subfr_length - n - 1 ], -rshifts);            /* Q(-rshifts) */
                tmp1 = silk_LSHIFT32((opus_int32)x_ptr[ n ],                    17);                  /* Q17 */
                tmp2 = silk_LSHIFT32((opus_int32)x_ptr[ subfr_length - n - 1 ], 17);                  /* Q17 */

                X1_3210 = _mm_set1_epi32(x1);
                X2_3210 = _mm_set1_epi32(x2);
                TMP1_3210 = _mm_setzero_si128();
                TMP2_3210 = _mm_setzero_si128();
                for(k = 0; k < n - 3; k += 4) {
                    PTR_3210   = OP_CVTEPI16_EPI32_M64(&x_ptr[ n - k - 1 - 3 ]);
                    SUBFR_3210 = OP_CVTEPI16_EPI32_M64(&x_ptr[ subfr_length - n + k ]);
                    FIRST_3210 = _mm_loadu_si128((__m128i *)&C_first_row[ k ]);
                    PTR_3210   = _mm_shuffle_epi32(PTR_3210,  _MM_SHUFFLE(0, 1, 2, 3));
                    LAST_3210  = _mm_loadu_si128((__m128i *)&C_last_row[ k ]);
                    ATMP_3210  = _mm_loadu_si128((__m128i *)&Af_QA[ k ]);

                    T1_3210 = _mm_mullo_epi32(PTR_3210, X1_3210);
                    T2_3210 = _mm_mullo_epi32(SUBFR_3210, X2_3210);

                    ATMP_3210 = _mm_srai_epi32(ATMP_3210, 7);
                    ATMP_3210 = _mm_add_epi32(ATMP_3210, CONST1);
                    ATMP_3210 = _mm_srai_epi32(ATMP_3210, 1);

                    FIRST_3210 = _mm_add_epi32(FIRST_3210, T1_3210);
                    LAST_3210 = _mm_add_epi32(LAST_3210, T2_3210);

                    PTR_3210   = _mm_mullo_epi32(ATMP_3210, PTR_3210);
                    SUBFR_3210   = _mm_mullo_epi32(ATMP_3210, SUBFR_3210);

                    _mm_storeu_si128((__m128i *)&C_first_row[ k ], FIRST_3210);
                    _mm_storeu_si128((__m128i *)&C_last_row[ k ], LAST_3210);

                    TMP1_3210 = _mm_add_epi32(TMP1_3210, PTR_3210);
                    TMP2_3210 = _mm_add_epi32(TMP2_3210, SUBFR_3210);
                }

                TMP1_3210 = _mm_add_epi32(TMP1_3210, _mm_unpackhi_epi64(TMP1_3210, TMP1_3210));
                TMP2_3210 = _mm_add_epi32(TMP2_3210, _mm_unpackhi_epi64(TMP2_3210, TMP2_3210));
                TMP1_3210 = _mm_add_epi32(TMP1_3210, _mm_shufflelo_epi16(TMP1_3210, 0x0E));
                TMP2_3210 = _mm_add_epi32(TMP2_3210, _mm_shufflelo_epi16(TMP2_3210, 0x0E));

                tmp1 += _mm_cvtsi128_si32(TMP1_3210);
                tmp2 += _mm_cvtsi128_si32(TMP2_3210);

                for(; k < n; k++) {
                    C_first_row[ k ] = silk_MLA(C_first_row[ k ], x1, x_ptr[ n - k - 1 ]           ); /* Q(-rshifts) */
                    C_last_row[ k ]  = silk_MLA(C_last_row[ k ],  x2, x_ptr[ subfr_length - n + k ]); /* Q(-rshifts) */
                    Atmp1 = silk_RSHIFT_ROUND(Af_QA[ k ], QA - 17);                                   /* Q17 */
                    tmp1 = silk_MLA(tmp1, x_ptr[ n - k - 1 ],            Atmp1);                      /* Q17 */
                    tmp2 = silk_MLA(tmp2, x_ptr[ subfr_length - n + k ], Atmp1);                      /* Q17 */
                }

                tmp1 = -tmp1;                /* Q17 */
                tmp2 = -tmp2;                /* Q17 */

                {
                    __m128i xmm_tmp1, xmm_tmp2;
                    __m128i xmm_x_ptr_n_k_x2x0, xmm_x_ptr_n_k_x3x1;
                    __m128i xmm_x_ptr_sub_x2x0, xmm_x_ptr_sub_x3x1;

                    xmm_tmp1 = _mm_set1_epi32(tmp1);
                    xmm_tmp2 = _mm_set1_epi32(tmp2);

                    for(k = 0; k <= n - 3; k += 4) {
                        xmm_x_ptr_n_k_x2x0 = OP_CVTEPI16_EPI32_M64(&x_ptr[ n - k - 3 ]);
                        xmm_x_ptr_sub_x2x0 = OP_CVTEPI16_EPI32_M64(&x_ptr[ subfr_length - n + k - 1 ]);

                        xmm_x_ptr_n_k_x2x0 = _mm_shuffle_epi32(xmm_x_ptr_n_k_x2x0, _MM_SHUFFLE(0, 1, 2, 3));

                        xmm_x_ptr_n_k_x2x0 = _mm_slli_epi32(xmm_x_ptr_n_k_x2x0, -rshifts - 1);
                        xmm_x_ptr_sub_x2x0 = _mm_slli_epi32(xmm_x_ptr_sub_x2x0, -rshifts - 1);

                        /* equal shift right 4 bytes, xmm_x_ptr_n_k_x3x1 = _mm_srli_si128(xmm_x_ptr_n_k_x2x0, 4)*/
                        xmm_x_ptr_n_k_x3x1 = _mm_shuffle_epi32(xmm_x_ptr_n_k_x2x0, _MM_SHUFFLE(0, 3, 2, 1));
                        xmm_x_ptr_sub_x3x1 = _mm_shuffle_epi32(xmm_x_ptr_sub_x2x0, _MM_SHUFFLE(0, 3, 2, 1));

                        xmm_x_ptr_n_k_x2x0 = _mm_mul_epi32(xmm_x_ptr_n_k_x2x0, xmm_tmp1);
                        xmm_x_ptr_n_k_x3x1 = _mm_mul_epi32(xmm_x_ptr_n_k_x3x1, xmm_tmp1);
                        xmm_x_ptr_sub_x2x0 = _mm_mul_epi32(xmm_x_ptr_sub_x2x0, xmm_tmp2);
                        xmm_x_ptr_sub_x3x1 = _mm_mul_epi32(xmm_x_ptr_sub_x3x1, xmm_tmp2);

                        xmm_x_ptr_n_k_x2x0 = _mm_srli_epi64(xmm_x_ptr_n_k_x2x0, 16);
                        xmm_x_ptr_n_k_x3x1 = _mm_slli_epi64(xmm_x_ptr_n_k_x3x1, 16);
                        xmm_x_ptr_sub_x2x0 = _mm_srli_epi64(xmm_x_ptr_sub_x2x0, 16);
                        xmm_x_ptr_sub_x3x1 = _mm_slli_epi64(xmm_x_ptr_sub_x3x1, 16);

                        xmm_x_ptr_n_k_x2x0 = _mm_blend_epi16(xmm_x_ptr_n_k_x2x0, xmm_x_ptr_n_k_x3x1, 0xCC);
                        xmm_x_ptr_sub_x2x0 = _mm_blend_epi16(xmm_x_ptr_sub_x2x0, xmm_x_ptr_sub_x3x1, 0xCC);

                        X1_3210  = _mm_loadu_si128((__m128i *)&CAf[ k ]);
                        PTR_3210 = _mm_loadu_si128((__m128i *)&CAb[ k ]);

                        X1_3210  = _mm_add_epi32(X1_3210, xmm_x_ptr_n_k_x2x0);
                        PTR_3210 = _mm_add_epi32(PTR_3210, xmm_x_ptr_sub_x2x0);

                        _mm_storeu_si128((__m128i *)&CAf[ k ], X1_3210);
                        _mm_storeu_si128((__m128i *)&CAb[ k ], PTR_3210);
                    }

                    for(; k <= n; k++) {
                        CAf[ k ] = silk_SMLAWW(CAf[ k ], tmp1,
                            silk_LSHIFT32((opus_int32)x_ptr[ n - k ], -rshifts - 1));                    /* Q(-rshift) */
                        CAb[ k ] = silk_SMLAWW(CAb[ k ], tmp2,
                            silk_LSHIFT32((opus_int32)x_ptr[ subfr_length - n + k - 1 ], -rshifts - 1)); /* Q(-rshift) */
                    }
                }
            }
        }

        /* Calculate nominator and denominator for the next order reflection (parcor) coefficient */
        tmp1 = C_first_row[ n ];                                                                        /* Q(-rshifts) */
        tmp2 = C_last_row[ n ];                                                                         /* Q(-rshifts) */
        num  = 0;                                                                                       /* Q(-rshifts) */
        nrg  = silk_ADD32(CAb[ 0 ], CAf[ 0 ]);                                                        /* Q(1-rshifts) */
        for(k = 0; k < n; k++) {
            Atmp_QA = Af_QA[ k ];
            lz = silk_CLZ32(silk_abs(Atmp_QA)) - 1;
            lz = silk_min(32 - QA, lz);
            Atmp1 = silk_LSHIFT32(Atmp_QA, lz);                                                       /* Q(QA + lz) */

            tmp1 = silk_ADD_LSHIFT32(tmp1, silk_SMMUL(C_last_row[  n - k - 1 ], Atmp1), 32 - QA - lz);  /* Q(-rshifts) */
            tmp2 = silk_ADD_LSHIFT32(tmp2, silk_SMMUL(C_first_row[ n - k - 1 ], Atmp1), 32 - QA - lz);  /* Q(-rshifts) */
            num  = silk_ADD_LSHIFT32(num,  silk_SMMUL(CAb[ n - k ],             Atmp1), 32 - QA - lz);  /* Q(-rshifts) */
            nrg  = silk_ADD_LSHIFT32(nrg,  silk_SMMUL(silk_ADD32(CAb[ k + 1 ], CAf[ k + 1 ]),
                                                                                Atmp1), 32 - QA - lz);    /* Q(1-rshifts) */
        }
        CAf[ n + 1 ] = tmp1;                                                                            /* Q(-rshifts) */
        CAb[ n + 1 ] = tmp2;                                                                            /* Q(-rshifts) */
        num = silk_ADD32(num, tmp2);                                                                  /* Q(-rshifts) */
        num = silk_LSHIFT32(-num, 1);                                                                 /* Q(1-rshifts) */

        /* Calculate the next order reflection (parcor) coefficient */
        if(silk_abs(num) < nrg) {
            rc_Q31 = silk_DIV32_varQ(num, nrg, 31);
        } else {
            rc_Q31 = (num > 0) ? silk_int32_MAX : silk_int32_MIN;
        }

        /* Update inverse prediction gain */
        tmp1 = ((opus_int32)1 << 30) - silk_SMMUL(rc_Q31, rc_Q31);
        tmp1 = silk_LSHIFT(silk_SMMUL(invGain_Q30, tmp1), 2);
        if(tmp1 <= minInvGain_Q30) {
            /* Max prediction gain exceeded; set reflection coefficient such that max prediction gain is exactly hit */
            tmp2 = ((opus_int32)1 << 30) - silk_DIV32_varQ(minInvGain_Q30, invGain_Q30, 30);            /* Q30 */
            rc_Q31 = silk_SQRT_APPROX(tmp2);                                                  /* Q15 */
            /* Newton-Raphson iteration */
            rc_Q31 = silk_RSHIFT32(rc_Q31 + silk_DIV32(tmp2, rc_Q31), 1);                   /* Q15 */
            rc_Q31 = silk_LSHIFT32(rc_Q31, 16);                                               /* Q31 */
            if(num < 0) {
                /* Ensure adjusted reflection coefficients has the original sign */
                rc_Q31 = -rc_Q31;
            }
            invGain_Q30 = minInvGain_Q30;
            reached_max_gain = 1;
        } else {
            invGain_Q30 = tmp1;
        }

        /* Update the AR coefficients */
        for(k = 0; k < (n + 1) >> 1; k++) {
            tmp1 = Af_QA[ k ];                                                                  /* QA */
            tmp2 = Af_QA[ n - k - 1 ];                                                          /* QA */
            Af_QA[ k ]         = silk_ADD_LSHIFT32(tmp1, silk_SMMUL(tmp2, rc_Q31), 1);      /* QA */
            Af_QA[ n - k - 1 ] = silk_ADD_LSHIFT32(tmp2, silk_SMMUL(tmp1, rc_Q31), 1);      /* QA */
        }
        Af_QA[ n ] = silk_RSHIFT32(rc_Q31, 31 - QA);                                          /* QA */

        if(reached_max_gain) {
            /* Reached max prediction gain; set remaining coefficients to zero and exit loop */
            for(k = n + 1; k < D; k++) {
                Af_QA[ k ] = 0;
            }
            break;
        }

        /* Update C * Af and C * Ab */
        for(k = 0; k <= n + 1; k++) {
            tmp1 = CAf[ k ];                                                                    /* Q(-rshifts) */
            tmp2 = CAb[ n - k + 1 ];                                                            /* Q(-rshifts) */
            CAf[ k ]         = silk_ADD_LSHIFT32(tmp1, silk_SMMUL(tmp2, rc_Q31), 1);        /* Q(-rshifts) */
            CAb[ n - k + 1 ] = silk_ADD_LSHIFT32(tmp2, silk_SMMUL(tmp1, rc_Q31), 1);        /* Q(-rshifts) */
        }
    }

    if(reached_max_gain) {
        for(k = 0; k < D; k++) {
            /* Scale coefficients */
            A_Q16[ k ] = -silk_RSHIFT_ROUND(Af_QA[ k ], QA - 16);
        }
        /* Subtract energy of preceding samples from C0 */
        if(rshifts > 0) {
            for(s = 0; s < nb_subfr; s++) {
                x_ptr = x + s * subfr_length;
                C0 -= (opus_int32)silk_RSHIFT64(silk_inner_prod16_aligned_64(x_ptr, x_ptr, D, arch), rshifts);
            }
        } else {
            for(s = 0; s < nb_subfr; s++) {
                x_ptr = x + s * subfr_length;
                C0 -= silk_LSHIFT32(silk_inner_prod_aligned(x_ptr, x_ptr, D, arch), -rshifts);
            }
        }
        /* Approximate residual energy */
        *res_nrg = silk_LSHIFT(silk_SMMUL(invGain_Q30, C0), 2);
        *res_nrg_Q = -rshifts;
    } else {
        /* Return residual energy */
        nrg  = CAf[ 0 ];                                                                            /* Q(-rshifts) */
        tmp1 = (opus_int32)1 << 16;                                                                             /* Q16 */
        for(k = 0; k < D; k++) {
            Atmp1 = silk_RSHIFT_ROUND(Af_QA[ k ], QA - 16);                                       /* Q16 */
            nrg  = silk_SMLAWW(nrg, CAf[ k + 1 ], Atmp1);                                         /* Q(-rshifts) */
            tmp1 = silk_SMLAWW(tmp1, Atmp1, Atmp1);                                               /* Q16 */
            A_Q16[ k ] = -Atmp1;
        }
        *res_nrg = silk_SMLAWW(nrg, silk_SMMUL(SILK_FIX_CONST(FIND_LPC_COND_FAC, 32), C0), -tmp1);/* Q(-rshifts) */
        *res_nrg_Q = -rshifts;
    }
}
/* Compute reflection coefficients from input signal */
void silk_burg_modified(
    opus_int32                  *res_nrg,           /* O    Residual energy                                             */
    opus_int                    *res_nrg_Q,         /* O    Residual energy Q value                                     */
    opus_int32                  A_Q16[],            /* O    Prediction coefficients (length order)                      */
    const opus_int16            x[],                /* I    Input signal, length: nb_subfr * ( D + subfr_length )       */
    const opus_int32            minInvGain_Q30,     /* I    Inverse of max prediction gain                              */
    const opus_int              subfr_length,       /* I    Input signal subframe length (incl. D preceding samples)    */
    const opus_int              nb_subfr,           /* I    Number of subframes stacked in x                            */
    const opus_int              D                   /* I    Order                                                       */
)
{
    opus_int         k, n, s, lz, rshifts, rshifts_extra, reached_max_gain;
    opus_int32       C0, num, nrg, rc_Q31, invGain_Q30, Atmp_QA, Atmp1, tmp1, tmp2, x1, x2;
    const opus_int16 *x_ptr;
    opus_int32       C_first_row[ SILK_MAX_ORDER_LPC ];
    opus_int32       C_last_row[  SILK_MAX_ORDER_LPC ];
    opus_int32       Af_QA[       SILK_MAX_ORDER_LPC ];
    opus_int32       CAf[ SILK_MAX_ORDER_LPC + 1 ];
    opus_int32       CAb[ SILK_MAX_ORDER_LPC + 1 ];

    silk_assert( subfr_length * nb_subfr <= MAX_FRAME_SIZE );

    /* Compute autocorrelations, added over subframes */
    silk_sum_sqr_shift( &C0, &rshifts, x, nb_subfr * subfr_length );
    if( rshifts > MAX_RSHIFTS ) {
        C0 = silk_LSHIFT32( C0, rshifts - MAX_RSHIFTS );
        silk_assert( C0 > 0 );
        rshifts = MAX_RSHIFTS;
    } else {
        lz = silk_CLZ32( C0 ) - 1;
        rshifts_extra = N_BITS_HEAD_ROOM - lz;
        if( rshifts_extra > 0 ) {
            rshifts_extra = silk_min( rshifts_extra, MAX_RSHIFTS - rshifts );
            C0 = silk_RSHIFT32( C0, rshifts_extra );
        } else {
            rshifts_extra = silk_max( rshifts_extra, MIN_RSHIFTS - rshifts );
            C0 = silk_LSHIFT32( C0, -rshifts_extra );
        }
        rshifts += rshifts_extra;
    }
    CAb[ 0 ] = CAf[ 0 ] = C0 + silk_SMMUL( SILK_FIX_CONST( FIND_LPC_COND_FAC, 32 ), C0 ) + 1;                                /* Q(-rshifts) */
    silk_memset( C_first_row, 0, SILK_MAX_ORDER_LPC * sizeof( opus_int32 ) );
    if( rshifts > 0 ) {
        for( s = 0; s < nb_subfr; s++ ) {
            x_ptr = x + s * subfr_length;
            for( n = 1; n < D + 1; n++ ) {
                C_first_row[ n - 1 ] += (opus_int32)silk_RSHIFT64(
                    silk_inner_prod16_aligned_64( x_ptr, x_ptr + n, subfr_length - n ), rshifts );
            }
        }
    } else {
        for( s = 0; s < nb_subfr; s++ ) {
            x_ptr = x + s * subfr_length;
            for( n = 1; n < D + 1; n++ ) {
                C_first_row[ n - 1 ] += silk_LSHIFT32(
                    silk_inner_prod_aligned( x_ptr, x_ptr + n, subfr_length - n ), -rshifts );
            }
        }
    }
    silk_memcpy( C_last_row, C_first_row, SILK_MAX_ORDER_LPC * sizeof( opus_int32 ) );

    /* Initialize */
    CAb[ 0 ] = CAf[ 0 ] = C0 + silk_SMMUL( SILK_FIX_CONST( FIND_LPC_COND_FAC, 32 ), C0 ) + 1;                                /* Q(-rshifts) */

    invGain_Q30 = (opus_int32)1 << 30;
    reached_max_gain = 0;
    for( n = 0; n < D; n++ ) {
        /* Update first row of correlation matrix (without first element) */
        /* Update last row of correlation matrix (without last element, stored in reversed order) */
        /* Update C * Af */
        /* Update C * flipud(Af) (stored in reversed order) */
        if( rshifts > -2 ) {
            for( s = 0; s < nb_subfr; s++ ) {
                x_ptr = x + s * subfr_length;
                x1  = -silk_LSHIFT32( (opus_int32)x_ptr[ n ],                    16 - rshifts );        /* Q(16-rshifts) */
                x2  = -silk_LSHIFT32( (opus_int32)x_ptr[ subfr_length - n - 1 ], 16 - rshifts );        /* Q(16-rshifts) */
                tmp1 = silk_LSHIFT32( (opus_int32)x_ptr[ n ],                    QA - 16 );             /* Q(QA-16) */
                tmp2 = silk_LSHIFT32( (opus_int32)x_ptr[ subfr_length - n - 1 ], QA - 16 );             /* Q(QA-16) */
                for( k = 0; k < n; k++ ) {
                    C_first_row[ k ] = silk_SMLAWB( C_first_row[ k ], x1, x_ptr[ n - k - 1 ]            ); /* Q( -rshifts ) */
                    C_last_row[ k ]  = silk_SMLAWB( C_last_row[ k ],  x2, x_ptr[ subfr_length - n + k ] ); /* Q( -rshifts ) */
                    Atmp_QA = Af_QA[ k ];
                    tmp1 = silk_SMLAWB( tmp1, Atmp_QA, x_ptr[ n - k - 1 ]            );                 /* Q(QA-16) */
                    tmp2 = silk_SMLAWB( tmp2, Atmp_QA, x_ptr[ subfr_length - n + k ] );                 /* Q(QA-16) */
                }
                tmp1 = silk_LSHIFT32( -tmp1, 32 - QA - rshifts );                                       /* Q(16-rshifts) */
                tmp2 = silk_LSHIFT32( -tmp2, 32 - QA - rshifts );                                       /* Q(16-rshifts) */
                for( k = 0; k <= n; k++ ) {
                    CAf[ k ] = silk_SMLAWB( CAf[ k ], tmp1, x_ptr[ n - k ]                    );        /* Q( -rshift ) */
                    CAb[ k ] = silk_SMLAWB( CAb[ k ], tmp2, x_ptr[ subfr_length - n + k - 1 ] );        /* Q( -rshift ) */
                }
            }
        } else {
            for( s = 0; s < nb_subfr; s++ ) {
                x_ptr = x + s * subfr_length;
                x1  = -silk_LSHIFT32( (opus_int32)x_ptr[ n ],                    -rshifts );            /* Q( -rshifts ) */
                x2  = -silk_LSHIFT32( (opus_int32)x_ptr[ subfr_length - n - 1 ], -rshifts );            /* Q( -rshifts ) */
                tmp1 = silk_LSHIFT32( (opus_int32)x_ptr[ n ],                    17 );                  /* Q17 */
                tmp2 = silk_LSHIFT32( (opus_int32)x_ptr[ subfr_length - n - 1 ], 17 );                  /* Q17 */
                for( k = 0; k < n; k++ ) {
                    C_first_row[ k ] = silk_MLA( C_first_row[ k ], x1, x_ptr[ n - k - 1 ]            ); /* Q( -rshifts ) */
                    C_last_row[ k ]  = silk_MLA( C_last_row[ k ],  x2, x_ptr[ subfr_length - n + k ] ); /* Q( -rshifts ) */
                    Atmp1 = silk_RSHIFT_ROUND( Af_QA[ k ], QA - 17 );                                   /* Q17 */
                    tmp1 = silk_MLA( tmp1, x_ptr[ n - k - 1 ],            Atmp1 );                      /* Q17 */
                    tmp2 = silk_MLA( tmp2, x_ptr[ subfr_length - n + k ], Atmp1 );                      /* Q17 */
                }
                tmp1 = -tmp1;                                                                           /* Q17 */
                tmp2 = -tmp2;                                                                           /* Q17 */
                for( k = 0; k <= n; k++ ) {
                    CAf[ k ] = silk_SMLAWW( CAf[ k ], tmp1,
                        silk_LSHIFT32( (opus_int32)x_ptr[ n - k ], -rshifts - 1 ) );                    /* Q( -rshift ) */
                    CAb[ k ] = silk_SMLAWW( CAb[ k ], tmp2,
                        silk_LSHIFT32( (opus_int32)x_ptr[ subfr_length - n + k - 1 ], -rshifts - 1 ) ); /* Q( -rshift ) */
                }
            }
        }

        /* Calculate nominator and denominator for the next order reflection (parcor) coefficient */
        tmp1 = C_first_row[ n ];                                                                        /* Q( -rshifts ) */
        tmp2 = C_last_row[ n ];                                                                         /* Q( -rshifts ) */
        num  = 0;                                                                                       /* Q( -rshifts ) */
        nrg  = silk_ADD32( CAb[ 0 ], CAf[ 0 ] );                                                        /* Q( 1-rshifts ) */
        for( k = 0; k < n; k++ ) {
            Atmp_QA = Af_QA[ k ];
            lz = silk_CLZ32( silk_abs( Atmp_QA ) ) - 1;
            lz = silk_min( 32 - QA, lz );
            Atmp1 = silk_LSHIFT32( Atmp_QA, lz );                                                       /* Q( QA + lz ) */

            tmp1 = silk_ADD_LSHIFT32( tmp1, silk_SMMUL( C_last_row[  n - k - 1 ], Atmp1 ), 32 - QA - lz );  /* Q( -rshifts ) */
            tmp2 = silk_ADD_LSHIFT32( tmp2, silk_SMMUL( C_first_row[ n - k - 1 ], Atmp1 ), 32 - QA - lz );  /* Q( -rshifts ) */
            num  = silk_ADD_LSHIFT32( num,  silk_SMMUL( CAb[ n - k ],             Atmp1 ), 32 - QA - lz );  /* Q( -rshifts ) */
            nrg  = silk_ADD_LSHIFT32( nrg,  silk_SMMUL( silk_ADD32( CAb[ k + 1 ], CAf[ k + 1 ] ),
                                                                                Atmp1 ), 32 - QA - lz );    /* Q( 1-rshifts ) */
        }
        CAf[ n + 1 ] = tmp1;                                                                            /* Q( -rshifts ) */
        CAb[ n + 1 ] = tmp2;                                                                            /* Q( -rshifts ) */
        num = silk_ADD32( num, tmp2 );                                                                  /* Q( -rshifts ) */
        num = silk_LSHIFT32( -num, 1 );                                                                 /* Q( 1-rshifts ) */

        /* Calculate the next order reflection (parcor) coefficient */
        if( silk_abs( num ) < nrg ) {
            rc_Q31 = silk_DIV32_varQ( num, nrg, 31 );
        } else {
            rc_Q31 = ( num > 0 ) ? silk_int32_MAX : silk_int32_MIN;
        }

        /* Update inverse prediction gain */
        tmp1 = ( (opus_int32)1 << 30 ) - silk_SMMUL( rc_Q31, rc_Q31 );
        tmp1 = silk_LSHIFT( silk_SMMUL( invGain_Q30, tmp1 ), 2 );
        if( tmp1 <= minInvGain_Q30 ) {
            /* Max prediction gain exceeded; set reflection coefficient such that max prediction gain is exactly hit */
            tmp2 = ( (opus_int32)1 << 30 ) - silk_DIV32_varQ( minInvGain_Q30, invGain_Q30, 30 );            /* Q30 */
            rc_Q31 = silk_SQRT_APPROX( tmp2 );                                                  /* Q15 */
            /* Newton-Raphson iteration */
            rc_Q31 = silk_RSHIFT32( rc_Q31 + silk_DIV32( tmp2, rc_Q31 ), 1 );                   /* Q15 */
            rc_Q31 = silk_LSHIFT32( rc_Q31, 16 );                                               /* Q31 */
            if( num < 0 ) {
                /* Ensure adjusted reflection coefficients has the original sign */
                rc_Q31 = -rc_Q31;
            }
            invGain_Q30 = minInvGain_Q30;
            reached_max_gain = 1;
        } else {
            invGain_Q30 = tmp1;
        }

        /* Update the AR coefficients */
        for( k = 0; k < (n + 1) >> 1; k++ ) {
            tmp1 = Af_QA[ k ];                                                                  /* QA */
            tmp2 = Af_QA[ n - k - 1 ];                                                          /* QA */
            Af_QA[ k ]         = silk_ADD_LSHIFT32( tmp1, silk_SMMUL( tmp2, rc_Q31 ), 1 );      /* QA */
            Af_QA[ n - k - 1 ] = silk_ADD_LSHIFT32( tmp2, silk_SMMUL( tmp1, rc_Q31 ), 1 );      /* QA */
        }
        Af_QA[ n ] = silk_RSHIFT32( rc_Q31, 31 - QA );                                          /* QA */

        if( reached_max_gain ) {
            /* Reached max prediction gain; set remaining coefficients to zero and exit loop */
            for( k = n + 1; k < D; k++ ) {
                Af_QA[ k ] = 0;
            }
            break;
        }

        /* Update C * Af and C * Ab */
        for( k = 0; k <= n + 1; k++ ) {
            tmp1 = CAf[ k ];                                                                    /* Q( -rshifts ) */
            tmp2 = CAb[ n - k + 1 ];                                                            /* Q( -rshifts ) */
            CAf[ k ]         = silk_ADD_LSHIFT32( tmp1, silk_SMMUL( tmp2, rc_Q31 ), 1 );        /* Q( -rshifts ) */
            CAb[ n - k + 1 ] = silk_ADD_LSHIFT32( tmp2, silk_SMMUL( tmp1, rc_Q31 ), 1 );        /* Q( -rshifts ) */
        }
    }

    if( reached_max_gain ) {
        for( k = 0; k < D; k++ ) {
            /* Scale coefficients */
            A_Q16[ k ] = -silk_RSHIFT_ROUND( Af_QA[ k ], QA - 16 );
        }
        /* Subtract energy of preceding samples from C0 */
        if( rshifts > 0 ) {
            for( s = 0; s < nb_subfr; s++ ) {
                x_ptr = x + s * subfr_length;
                C0 -= (opus_int32)silk_RSHIFT64( silk_inner_prod16_aligned_64( x_ptr, x_ptr, D ), rshifts );
            }
        } else {
            for( s = 0; s < nb_subfr; s++ ) {
                x_ptr = x + s * subfr_length;
                C0 -= silk_LSHIFT32( silk_inner_prod_aligned( x_ptr, x_ptr, D ), -rshifts );
            }
        }
        /* Approximate residual energy */
        *res_nrg = silk_LSHIFT( silk_SMMUL( invGain_Q30, C0 ), 2 );
        *res_nrg_Q = -rshifts;
    } else {
        /* Return residual energy */
        nrg  = CAf[ 0 ];                                                                            /* Q( -rshifts ) */
        tmp1 = (opus_int32)1 << 16;                                                                             /* Q16 */
        for( k = 0; k < D; k++ ) {
            Atmp1 = silk_RSHIFT_ROUND( Af_QA[ k ], QA - 16 );                                       /* Q16 */
            nrg  = silk_SMLAWW( nrg, CAf[ k + 1 ], Atmp1 );                                         /* Q( -rshifts ) */
            tmp1 = silk_SMLAWW( tmp1, Atmp1, Atmp1 );                                               /* Q16 */
            A_Q16[ k ] = -Atmp1;
        }
        *res_nrg = silk_SMLAWW( nrg, silk_SMMUL( FIND_LPC_COND_FAC, C0 ), -tmp1 );                  /* Q( -rshifts ) */
        *res_nrg_Q = -rshifts;
    }   
}