opus_int silk_InitEncoder(                              /* O    Returns error code                              */
    void                            *encState,          /* I/O  State                                           */
    int                              arch,              /* I    Run-time architecture                           */
    silk_EncControlStruct           *encStatus          /* O    Encoder Status                                  */
)
{
    silk_encoder *psEnc;
    opus_int n, ret = SILK_NO_ERROR;

    psEnc = (silk_encoder *)encState;

    /* Reset encoder */
    silk_memset( psEnc, 0, sizeof( silk_encoder ) );
    for( n = 0; n < ENCODER_NUM_CHANNELS; n++ ) {
        if( ret += silk_init_encoder( &psEnc->state_Fxx[ n ], arch ) ) {
            silk_assert( 0 );
        }
    }

    psEnc->nChannelsAPI = 1;
    psEnc->nChannelsInternal = 1;

    /* Read control structure */
    if( ret += silk_QueryEncoder( encState, encStatus ) ) {
        silk_assert( 0 );
    }

    return ret;
}

/* Compute quantization errors for an LPC_order element input vector for a VQ codebook */
void silk_NLSF_VQ(
    opus_int32                  err_Q26[],                      /* O    Quantization errors [K]                     */
    const opus_int16            in_Q15[],                       /* I    Input vectors to be quantized [LPC_order]   */
    const opus_uint8            pCB_Q8[],                       /* I    Codebook vectors [K*LPC_order]              */
    const opus_int              K,                              /* I    Number of codebook vectors                  */
    const opus_int              LPC_order                       /* I    Number of LPCs                              */
)
{
    opus_int        i, m;
    opus_int32      diff_Q15, sum_error_Q30, sum_error_Q26;

    silk_assert( LPC_order <= 16 );
    silk_assert( ( LPC_order & 1 ) == 0 );

    /* Loop over codebook */
    for( i = 0; i < K; i++ ) {
        sum_error_Q26 = 0;
        for( m = 0; m < LPC_order; m += 2 ) {
            /* Compute weighted squared quantization error for index m */
            diff_Q15 = silk_SUB_LSHIFT32( in_Q15[ m ], (opus_int32)*pCB_Q8++, 7 ); /* range: [ -32767 : 32767 ]*/
            sum_error_Q30 = silk_SMULBB( diff_Q15, diff_Q15 );

            /* Compute weighted squared quantization error for index m + 1 */
            diff_Q15 = silk_SUB_LSHIFT32( in_Q15[m + 1], (opus_int32)*pCB_Q8++, 7 ); /* range: [ -32767 : 32767 ]*/
            sum_error_Q30 = silk_SMLABB( sum_error_Q30, diff_Q15, diff_Q15 );

            sum_error_Q26 = silk_ADD_RSHIFT32( sum_error_Q26, sum_error_Q30, 4 );

            silk_assert( sum_error_Q26 >= 0 );
            silk_assert( sum_error_Q30 >= 0 );
        }
        err_Q26[ i ] = sum_error_Q26;
    }
}

static OPUS_INLINE void silk_LDL_FLP(
    silk_float          *A,         /* I/O  Pointer to Symetric Square Matrix                               */
    opus_int            M,          /* I    Size of Matrix                                                  */
    silk_float          *L,         /* I/O  Pointer to Square Upper triangular Matrix                       */
    silk_float          *Dinv       /* I/O  Pointer to vector holding the inverse diagonal elements of D    */
)
{
    opus_int i, j, k, loop_count, err = 1;
    silk_float *ptr1, *ptr2;
    double temp, diag_min_value;
    silk_float v[ MAX_MATRIX_SIZE ] = { 0 }, D[ MAX_MATRIX_SIZE ]; /* temp arrays*/

    silk_assert(M <= MAX_MATRIX_SIZE);

    diag_min_value = FIND_LTP_COND_FAC * 0.5f * (A[ 0 ] + A[ M * M - 1 ]);
    for(loop_count = 0; loop_count < M && err == 1; loop_count++) {
        err = 0;
        for(j = 0; j < M; j++) {
            ptr1 = matrix_adr(L, j, 0, M);
            temp = matrix_ptr(A, j, j, M); /* element in row j column j*/
            for(i = 0; i < j; i++) {
                v[ i ] = ptr1[ i ] * D[ i ];
                temp  -= ptr1[ i ] * v[ i ];
            }
            if(temp < diag_min_value) {
                /* Badly conditioned matrix: add white noise and run again */
                temp = (loop_count + 1) * diag_min_value - temp;
                for(i = 0; i < M; i++) {
                    matrix_ptr(A, i, i, M) += (silk_float)temp;
                }
                err = 1;
                break;
            }
            D[ j ]    = (silk_float)temp;
            Dinv[ j ] = (silk_float)(1.0f / temp);
            matrix_ptr(L, j, j, M) = 1.0f;

            ptr1 = matrix_adr(A, j, 0, M);
            ptr2 = matrix_adr(L, j + 1, 0, M);
            for(i = j + 1; i < M; i++) {
                temp = 0.0;
                for(k = 0; k < j; k++) {
                    temp += ptr2[ k ] * v[ k ];
                }
                matrix_ptr(L, i, j, M) = (silk_float)((ptr1[ i ] - temp) * Dinv[ j ]);
                ptr2 += M; /* go to next column*/
            }
        }
    }
    silk_assert(err == 0);
}

/* Autocorrelations for a warped frequency axis */
void silk_warped_autocorrelation_FIX(
          opus_int32                *corr,                                  /* O    Result [order + 1]                                                          */
          opus_int                  *scale,                                 /* O    Scaling of the correlation vector                                           */
    const opus_int16                *input,                                 /* I    Input data to correlate                                                     */
    const opus_int                  warping_Q16,                            /* I    Warping coefficient                                                         */
    const opus_int                  length,                                 /* I    Length of input                                                             */
    const opus_int                  order                                   /* I    Correlation order (even)                                                    */
)
{
    opus_int   n, i, lsh;
    opus_int32 tmp1_QS, tmp2_QS;
    opus_int32 state_QS[ MAX_SHAPE_LPC_ORDER + 1 ] = { 0 };
    opus_int64 corr_QC[  MAX_SHAPE_LPC_ORDER + 1 ] = { 0 };

    /* Order must be even */
    silk_assert( ( order & 1 ) == 0 );
    silk_assert( 2 * QS - QC >= 0 );

    /* Loop over samples */
    for( n = 0; n < length; n++ ) {
        tmp1_QS = silk_LSHIFT32( (opus_int32)input[ n ], QS );
        /* Loop over allpass sections */
        for( i = 0; i < order; i += 2 ) {
            /* Output of allpass section */
            tmp2_QS = silk_SMLAWB( state_QS[ i ], state_QS[ i + 1 ] - tmp1_QS, warping_Q16 );
            state_QS[ i ]  = tmp1_QS;
            corr_QC[  i ] += silk_RSHIFT64( silk_SMULL( tmp1_QS, state_QS[ 0 ] ), 2 * QS - QC );
            /* Output of allpass section */
            tmp1_QS = silk_SMLAWB( state_QS[ i + 1 ], state_QS[ i + 2 ] - tmp2_QS, warping_Q16 );
            state_QS[ i + 1 ]  = tmp2_QS;
            corr_QC[  i + 1 ] += silk_RSHIFT64( silk_SMULL( tmp2_QS, state_QS[ 0 ] ), 2 * QS - QC );
        }
        state_QS[ order ] = tmp1_QS;
        corr_QC[  order ] += silk_RSHIFT64( silk_SMULL( tmp1_QS, state_QS[ 0 ] ), 2 * QS - QC );
    }

    lsh = silk_CLZ64( corr_QC[ 0 ] ) - 35;
    lsh = silk_LIMIT( lsh, -12 - QC, 30 - QC );
    *scale = -( QC + lsh );
    silk_assert( *scale >= -30 && *scale <= 12 );
    if( lsh >= 0 ) {
        for( i = 0; i < order + 1; i++ ) {
            corr[ i ] = (opus_int32)silk_CHECK_FIT32( silk_LSHIFT64( corr_QC[ i ], lsh ) );
        }
    } else {
        for( i = 0; i < order + 1; i++ ) {
            corr[ i ] = (opus_int32)silk_CHECK_FIT32( silk_RSHIFT64( corr_QC[ i ], -lsh ) );
        }
    }
    silk_assert( corr_QC[ 0 ] >= 0 ); /* If breaking, decrease QC*/
}

/* Residual energy: nrg = wxx - 2 * wXx * c + c' * wXX * c */
silk_float silk_residual_energy_covar_FLP(                              /* O    Weighted residual energy                    */
    const silk_float                *c,                                 /* I    Filter coefficients                         */
    silk_float                      *wXX,                               /* I/O  Weighted correlation matrix, reg. out       */
    const silk_float                *wXx,                               /* I    Weighted correlation vector                 */
    const silk_float                wxx,                                /* I    Weighted correlation value                  */
    const opus_int                  D                                   /* I    Dimension                                   */
)
{
    opus_int   i, j, k;
    silk_float tmp, nrg = 0.0f, regularization;

    /* Safety checks */
    silk_assert( D >= 0 );

    regularization = REGULARIZATION_FACTOR * ( wXX[ 0 ] + wXX[ D * D - 1 ] );
    for( k = 0; k < MAX_ITERATIONS_RESIDUAL_NRG; k++ ) {
        nrg = wxx;

        tmp = 0.0f;
        for( i = 0; i < D; i++ ) {
            tmp += wXx[ i ] * c[ i ];
        }
        nrg -= 2.0f * tmp;

        /* compute c' * wXX * c, assuming wXX is symmetric */
        for( i = 0; i < D; i++ ) {
            tmp = 0.0f;
            for( j = i + 1; j < D; j++ ) {
                tmp += matrix_c_ptr( wXX, i, j, D ) * c[ j ];
            }
            nrg += c[ i ] * ( 2.0f * tmp + matrix_c_ptr( wXX, i, i, D ) * c[ i ] );
        }
        if( nrg > 0 ) {
            break;
        } else {
            /* Add white noise */
            for( i = 0; i < D; i++ ) {
                matrix_c_ptr( wXX, i, i, D ) +=  regularization;
            }
            /* Increase noise for next run */
            regularization *= 2.0f;
        }
    }
    if( k == MAX_ITERATIONS_RESIDUAL_NRG ) {
        silk_assert( nrg == 0 );
        nrg = 1.0f;
    }

    return nrg;
}

/* sum of squares of a silk_float array, with result as double */
double silk_energy_FLP(
    const silk_float    *data,
    opus_int            dataSize
)
{
    opus_int  i, dataSize4;
    double   result;

    /* 4x unrolled loop */
    result = 0.0;
    dataSize4 = dataSize & 0xFFFC;
    for( i = 0; i < dataSize4; i += 4 ) {
        result += data[ i + 0 ] * (double)data[ i + 0 ] +
                  data[ i + 1 ] * (double)data[ i + 1 ] +
                  data[ i + 2 ] * (double)data[ i + 2 ] +
                  data[ i + 3 ] * (double)data[ i + 3 ];
    }

    /* add any remaining products */
    for( ; i < dataSize; i++ ) {
        result += data[ i ] * (double)data[ i ];
    }

    silk_assert( result >= 0.0 );
    return result;
}

/* Shell decoder, operates on one shell code frame of 16 pulses */
void silk_shell_decoder(
    opus_int                    *pulses0,                       /* O    data: nonnegative pulse amplitudes          */
    ec_dec                      *psRangeDec,                    /* I/O  Compressor data structure                   */
    const opus_int              pulses4                         /* I    number of pulses per pulse-subframe         */
)
{
    opus_int pulses3[ 2 ], pulses2[ 4 ], pulses1[ 8 ];

    /* this function operates on one shell code frame of 16 pulses */
    silk_assert( SHELL_CODEC_FRAME_LENGTH == 16 );

    decode_split( &pulses3[  0 ], &pulses3[  1 ], psRangeDec, pulses4,      silk_shell_code_table3 );

    decode_split( &pulses2[  0 ], &pulses2[  1 ], psRangeDec, pulses3[ 0 ], silk_shell_code_table2 );

    decode_split( &pulses1[  0 ], &pulses1[  1 ], psRangeDec, pulses2[ 0 ], silk_shell_code_table1 );
    decode_split( &pulses0[  0 ], &pulses0[  1 ], psRangeDec, pulses1[ 0 ], silk_shell_code_table0 );
    decode_split( &pulses0[  2 ], &pulses0[  3 ], psRangeDec, pulses1[ 1 ], silk_shell_code_table0 );

    decode_split( &pulses1[  2 ], &pulses1[  3 ], psRangeDec, pulses2[ 1 ], silk_shell_code_table1 );
    decode_split( &pulses0[  4 ], &pulses0[  5 ], psRangeDec, pulses1[ 2 ], silk_shell_code_table0 );
    decode_split( &pulses0[  6 ], &pulses0[  7 ], psRangeDec, pulses1[ 3 ], silk_shell_code_table0 );

    decode_split( &pulses2[  2 ], &pulses2[  3 ], psRangeDec, pulses3[ 1 ], silk_shell_code_table2 );

    decode_split( &pulses1[  4 ], &pulses1[  5 ], psRangeDec, pulses2[ 2 ], silk_shell_code_table1 );
    decode_split( &pulses0[  8 ], &pulses0[  9 ], psRangeDec, pulses1[ 4 ], silk_shell_code_table0 );
    decode_split( &pulses0[ 10 ], &pulses0[ 11 ], psRangeDec, pulses1[ 5 ], silk_shell_code_table0 );

    decode_split( &pulses1[  6 ], &pulses1[  7 ], psRangeDec, pulses2[ 3 ], silk_shell_code_table1 );
    decode_split( &pulses0[ 12 ], &pulses0[ 13 ], psRangeDec, pulses1[ 6 ], silk_shell_code_table0 );
    decode_split( &pulses0[ 14 ], &pulses0[ 15 ], psRangeDec, pulses1[ 7 ], silk_shell_code_table0 );
}

/* Calculates correlation vector X'*t */
void silk_corrVector_FIX(
    const opus_int16                *x,                                     /* I    x vector [L + order - 1] used to form data matrix X                         */
    const opus_int16                *t,                                     /* I    Target vector [L]                                                           */
    const opus_int                  L,                                      /* I    Length of vectors                                                           */
    const opus_int                  order,                                  /* I    Max lag for correlation                                                     */
    opus_int32                      *Xt,                                    /* O    Pointer to X'*t correlation vector [order]                                  */
    const opus_int                  rshifts                                 /* I    Right shifts of correlations                                                */
)
{
    opus_int         lag, i;
    const opus_int16 *ptr1, *ptr2;
    opus_int32       inner_prod;

    ptr1 = &x[ order - 1 ]; /* Points to first sample of column 0 of X: X[:,0] */
    ptr2 = t;
    /* Calculate X'*t */
    if( rshifts > 0 ) {
        /* Right shifting used */
        for( lag = 0; lag < order; lag++ ) {
            inner_prod = 0;
            for( i = 0; i < L; i++ ) {
                inner_prod += silk_RSHIFT32( silk_SMULBB( ptr1[ i ], ptr2[i] ), rshifts );
            }
            Xt[ lag ] = inner_prod; /* X[:,lag]'*t */
            ptr1--; /* Go to next column of X */
        }
    } else {
        silk_assert( rshifts == 0 );
        for( lag = 0; lag < order; lag++ ) {
            Xt[ lag ] = silk_inner_prod_aligned( ptr1, ptr2, L ); /* X[:,lag]'*t */
            ptr1--; /* Go to next column of X */
        }
    }
}

/* Interpolate two vectors */
void silk_interpolate(
    opus_int16                  xi[ MAX_LPC_ORDER ],            /* O    interpolated vector                         */
    const opus_int16            x0[ MAX_LPC_ORDER ],            /* I    first vector                                */
    const opus_int16            x1[ MAX_LPC_ORDER ],            /* I    second vector                               */
    const opus_int              ifact_Q2,                       /* I    interp. factor, weight on 2nd vector        */
    const opus_int              d                               /* I    number of parameters                        */
)
{
    opus_int i;

    silk_assert( ifact_Q2 >= 0 );
    silk_assert( ifact_Q2 <= 4 );

    for( i = 0; i < d; i++ ) {
        xi[ i ] = (opus_int16)silk_ADD_RSHIFT( x0[ i ], silk_SMULBB( x1[ i ] - x0[ i ], ifact_Q2 ), 2 );
    }
}

void silk_LPC_analysis_filter(
    opus_int16                  *out,               /* O    Output signal                                               */
    const opus_int16            *in,                /* I    Input signal                                                */
    const opus_int16            *B,                 /* I    MA prediction coefficients, Q12 [order]                     */
    const opus_int32            len,                /* I    Signal length                                               */
    const opus_int32            d                   /* I    Filter order                                                */
)
{
    opus_int         ix, j;
    opus_int32       out32_Q12, out32;
    const opus_int16 *in_ptr;

    silk_assert( d >= 6 );
    silk_assert( (d & 1) == 0 );
    silk_assert( d <= len );

    for( ix = d; ix < len; ix++ ) {
        in_ptr = &in[ ix - 1 ];

        out32_Q12 = silk_SMULBB( in_ptr[  0 ], B[ 0 ] );
        /* Allowing wrap around so that two wraps can cancel each other. The rare
           cases where the result wraps around can only be triggered by invalid streams*/
        out32_Q12 = silk_SMLABB_ovflw( out32_Q12, in_ptr[ -1 ], B[ 1 ] );
        out32_Q12 = silk_SMLABB_ovflw( out32_Q12, in_ptr[ -2 ], B[ 2 ] );
        out32_Q12 = silk_SMLABB_ovflw( out32_Q12, in_ptr[ -3 ], B[ 3 ] );
        out32_Q12 = silk_SMLABB_ovflw( out32_Q12, in_ptr[ -4 ], B[ 4 ] );
        out32_Q12 = silk_SMLABB_ovflw( out32_Q12, in_ptr[ -5 ], B[ 5 ] );
        for( j = 6; j < d; j += 2 ) {
            out32_Q12 = silk_SMLABB_ovflw( out32_Q12, in_ptr[ -j     ], B[ j     ] );
            out32_Q12 = silk_SMLABB_ovflw( out32_Q12, in_ptr[ -j - 1 ], B[ j + 1 ] );
        }

        /* Subtract prediction */
        out32_Q12 = silk_SUB32_ovflw( silk_LSHIFT( (opus_int32)in_ptr[ 1 ], 12 ), out32_Q12 );

        /* Scale to Q0 */
        out32 = silk_RSHIFT_ROUND( out32_Q12, 12 );

        /* Saturate output */
        out[ ix ] = (opus_int16)silk_SAT16( out32 );
    }

    /* Set first d output samples to zero */
    silk_memset( out, 0, d * sizeof( opus_int16 ) );
}

/* uses SMLAWB(), requiring armv5E and higher.                          */
opus_int32 silk_schur(                              /* O    Returns residual energy                                     */
    opus_int16                  *rc_Q15,            /* O    reflection coefficients [order] Q15                         */
    const opus_int32            *c,                 /* I    correlations [order+1]                                      */
    const opus_int32            order               /* I    prediction order                                            */
)
{
    opus_int        k, n, lz;
    opus_int32    C[ SILK_MAX_ORDER_LPC + 1 ][ 2 ];
    opus_int32    Ctmp1, Ctmp2, rc_tmp_Q15;

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

    /* Get number of leading zeros */
    lz = silk_CLZ32( c[ 0 ] );

    /* Copy correlations and adjust level to Q30 */
    if( lz < 2 ) {
        /* lz must be 1, so shift one to the right */
        for( k = 0; k < order + 1; k++ ) {
            C[ k ][ 0 ] = C[ k ][ 1 ] = silk_RSHIFT( c[ k ], 1 );
        }
    } else if( lz > 2 ) {
        /* Shift to the left */
        lz -= 2;
        for( k = 0; k < order + 1; k++ ) {
            C[ k ][ 0 ] = C[ k ][ 1 ] = silk_LSHIFT( c[ k ], lz );
        }
    } else {
        /* No need to shift */
        for( k = 0; k < order + 1; k++ ) {
            C[ k ][ 0 ] = C[ k ][ 1 ] = c[ k ];
        }
    }

    for( k = 0; k < order; k++ ) {

        /* Get reflection coefficient */
        rc_tmp_Q15 = -silk_DIV32_16( C[ k + 1 ][ 0 ], silk_max_32( silk_RSHIFT( C[ 0 ][ 1 ], 15 ), 1 ) );

        /* Clip (shouldn't happen for properly conditioned inputs) */
        rc_tmp_Q15 = silk_SAT16( rc_tmp_Q15 );

        /* Store */
        rc_Q15[ k ] = (opus_int16)rc_tmp_Q15;

        /* Update correlations */
        for( n = 0; n < order - k; n++ ) {
            Ctmp1 = C[ n + k + 1 ][ 0 ];
            Ctmp2 = C[ n ][ 1 ];
            C[ n + k + 1 ][ 0 ] = silk_SMLAWB( Ctmp1, silk_LSHIFT( Ctmp2, 1 ), rc_tmp_Q15 );
            C[ n ][ 1 ]         = silk_SMLAWB( Ctmp2, silk_LSHIFT( Ctmp1, 1 ), rc_tmp_Q15 );
        }
    }

    /* return residual energy */
    return C[ 0 ][ 1 ];
}

/* Entropy code the mid/side quantization indices */
void silk_stereo_encode_pred(
    ec_enc                      *psRangeEnc,                    /* I/O  Compressor data structure                   */
    opus_int8                   ix[ 2 ][ 3 ]                    /* I    Quantization indices                        */
)
{
    opus_int   n;

    /* Entropy coding */
    n = 5 * ix[ 0 ][ 2 ] + ix[ 1 ][ 2 ];
    silk_assert( n < 25 );
    ec_enc_icdf( psRangeEnc, n, silk_stereo_pred_joint_iCDF, 8 );
    for( n = 0; n < 2; n++ ) {
        silk_assert( ix[ n ][ 0 ] < 3 );
        silk_assert( ix[ n ][ 1 ] < STEREO_QUANT_SUB_STEPS );
        ec_enc_icdf( psRangeEnc, ix[ n ][ 0 ], silk_uniform3_iCDF, 8 );
        ec_enc_icdf( psRangeEnc, ix[ n ][ 1 ], silk_uniform5_iCDF, 8 );
    }
}

/* 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 ] );
}

/*  2 -> sine window from pi/2 to pi    */
void silk_apply_sine_window_FLP(
    silk_float                      px_win[],                           /* O    Pointer to windowed signal                  */
    const silk_float                px[],                               /* I    Pointer to input signal                     */
    const opus_int                  win_type,                           /* I    Selects a window type                       */
    const opus_int                  length                              /* I    Window length, multiple of 4                */
)
{
    opus_int   k;
    silk_float freq, c, S0, S1;

    silk_assert( win_type == 1 || win_type == 2 );

    /* Length must be multiple of 4 */
    silk_assert( ( length & 3 ) == 0 );

    freq = PI / ( length + 1 );

    /* Approximation of 2 * cos(f) */
    c = 2.0f - freq * freq;

    /* Initialize state */
    if( win_type < 2 ) {
        /* Start from 0 */
        S0 = 0.0f;
        /* Approximation of sin(f) */
        S1 = freq;
    } else {
        /* Start from 1 */
        S0 = 1.0f;
        /* Approximation of cos(f) */
        S1 = 0.5f * c;
    }

    /* Uses the recursive equation:   sin(n*f) = 2 * cos(f) * sin((n-1)*f) - sin((n-2)*f)   */
    /* 4 samples at a time */
    for( k = 0; k < length; k += 4 ) {
        px_win[ k + 0 ] = px[ k + 0 ] * 0.5f * ( S0 + S1 );
        px_win[ k + 1 ] = px[ k + 1 ] * S1;
        S0 = c * S1 - S0;
        px_win[ k + 2 ] = px[ k + 2 ] * 0.5f * ( S1 + S0 );
        px_win[ k + 3 ] = px[ k + 3 ] * S0;
        S1 = c * S0 - S1;
    }
}

/* Prefilter for finding Quantizer input signal */
static OPUS_INLINE void silk_prefilt_FIX(
        silk_prefilter_state_FIX *P,                         /* I/O  state                               */
        opus_int32 st_res_Q12[],               /* I    short term residual signal          */
        opus_int32 xw_Q3[],                    /* O    prefiltered signal                  */
        opus_int32 HarmShapeFIRPacked_Q12,     /* I    Harmonic shaping coeficients        */
        opus_int Tilt_Q14,                   /* I    Tilt shaping coeficient             */
        opus_int32 LF_shp_Q14,                 /* I    Low-frequancy shaping coeficients   */
        opus_int lag,                        /* I    Lag for harmonic shaping            */
        opus_int length                      /* I    Length of signals                   */
) {
    opus_int i, idx, LTP_shp_buf_idx;
    opus_int32 n_LTP_Q12, n_Tilt_Q10, n_LF_Q10;
    opus_int32 sLF_MA_shp_Q12, sLF_AR_shp_Q12;
    opus_int16 *LTP_shp_buf;

    /* To speed up use temp variables instead of using the struct */
    LTP_shp_buf = P->sLTP_shp;
    LTP_shp_buf_idx = P->sLTP_shp_buf_idx;
    sLF_AR_shp_Q12 = P->sLF_AR_shp_Q12;
    sLF_MA_shp_Q12 = P->sLF_MA_shp_Q12;

    for (i = 0; i < length; i++) {
        if (lag > 0) {
            /* unrolled loop */
            silk_assert(HARM_SHAPE_FIR_TAPS == 3);
            idx = lag + LTP_shp_buf_idx;
            n_LTP_Q12 = silk_SMULBB(LTP_shp_buf[(idx - HARM_SHAPE_FIR_TAPS / 2 - 1) & LTP_MASK],
                                    HarmShapeFIRPacked_Q12);
            n_LTP_Q12 = silk_SMLABT(n_LTP_Q12,
                                    LTP_shp_buf[(idx - HARM_SHAPE_FIR_TAPS / 2) & LTP_MASK],
                                    HarmShapeFIRPacked_Q12);
            n_LTP_Q12 = silk_SMLABB(n_LTP_Q12,
                                    LTP_shp_buf[(idx - HARM_SHAPE_FIR_TAPS / 2 + 1) & LTP_MASK],
                                    HarmShapeFIRPacked_Q12);
        } else {
            n_LTP_Q12 = 0;
        }

        n_Tilt_Q10 = silk_SMULWB(sLF_AR_shp_Q12, Tilt_Q14);
        n_LF_Q10 = silk_SMLAWB(silk_SMULWT(sLF_AR_shp_Q12, LF_shp_Q14), sLF_MA_shp_Q12, LF_shp_Q14);

        sLF_AR_shp_Q12 = silk_SUB32(st_res_Q12[i], silk_LSHIFT(n_Tilt_Q10, 2));
        sLF_MA_shp_Q12 = silk_SUB32(sLF_AR_shp_Q12, silk_LSHIFT(n_LF_Q10, 2));

        LTP_shp_buf_idx = (LTP_shp_buf_idx - 1) & LTP_MASK;
        LTP_shp_buf[LTP_shp_buf_idx] = (opus_int16) silk_SAT16(
                silk_RSHIFT_ROUND(sLF_MA_shp_Q12, 12));

        xw_Q3[i] = silk_RSHIFT_ROUND(silk_SUB32(sLF_MA_shp_Q12, n_LTP_Q12), 9);
    }

    /* Copy temp variable back to state */
    P->sLF_AR_shp_Q12 = sLF_AR_shp_Q12;
    P->sLF_MA_shp_Q12 = sLF_MA_shp_Q12;
    P->sLTP_shp_buf_idx = LTP_shp_buf_idx;
}

/* Input and output sampling rate are at most 48000 Hz  */
opus_int silk_resampler(
    silk_resampler_state_struct *S,                 /* I/O  Resampler state                                             */
    opus_int16                  out[],              /* O    Output signal                                               */
    const opus_int16            in[],               /* I    Input signal                                                */
    opus_int32                  inLen               /* I    Number of input samples                                     */
)
{
    opus_int nSamples;

    /* Need at least 1 ms of input data */
    silk_assert( inLen >= S->Fs_in_kHz );
    /* Delay can't exceed the 1 ms of buffering */
    silk_assert( S->inputDelay <= S->Fs_in_kHz );

    nSamples = S->Fs_in_kHz - S->inputDelay;

    /* Copy to delay buffer */
    silk_memcpy( &S->delayBuf[ S->inputDelay ], in, nSamples * sizeof( opus_int16 ) );

    switch( S->resampler_function ) {
        case USE_silk_resampler_private_up2_HQ_wrapper:
            silk_resampler_private_up2_HQ_wrapper( S, out, S->delayBuf, S->Fs_in_kHz );
            silk_resampler_private_up2_HQ_wrapper( S, &out[ S->Fs_out_kHz ], &in[ nSamples ], inLen - S->Fs_in_kHz );
            break;
        case USE_silk_resampler_private_IIR_FIR:
            silk_resampler_private_IIR_FIR( S, out, S->delayBuf, S->Fs_in_kHz );
            silk_resampler_private_IIR_FIR( S, &out[ S->Fs_out_kHz ], &in[ nSamples ], inLen - S->Fs_in_kHz );
            break;
        case USE_silk_resampler_private_down_FIR:
            silk_resampler_private_down_FIR( S, out, S->delayBuf, S->Fs_in_kHz );
            silk_resampler_private_down_FIR( S, &out[ S->Fs_out_kHz ], &in[ nSamples ], inLen - S->Fs_in_kHz );
            break;
        default:
            silk_memcpy( out, S->delayBuf, S->Fs_in_kHz * sizeof( opus_int16 ) );
            silk_memcpy( &out[ S->Fs_out_kHz ], &in[ nSamples ], ( inLen - S->Fs_in_kHz ) * sizeof( opus_int16 ) );
    }

    /* Copy to delay buffer */
    silk_memcpy( S->delayBuf, &in[ inLen - S->inputDelay ], S->inputDelay * sizeof( opus_int16 ) );

    return 0;
}

/* Helper function, interpolates the filter taps */
static OPUS_INLINE void silk_LP_interpolate_filter_taps(
    opus_int32           B_Q28[ TRANSITION_NB ],
    opus_int32           A_Q28[ TRANSITION_NA ],
    const opus_int       ind,
    const opus_int32     fac_Q16
)
{
    opus_int nb, na;

    if( ind < TRANSITION_INT_NUM - 1 ) {
        if( fac_Q16 > 0 ) {
            if( fac_Q16 < 32768 ) { /* fac_Q16 is in range of a 16-bit int */
                /* Piece-wise linear interpolation of B and A */
                for( nb = 0; nb < TRANSITION_NB; nb++ ) {
                    B_Q28[ nb ] = silk_SMLAWB(
                        silk_Transition_LP_B_Q28[ ind     ][ nb ],
                        silk_Transition_LP_B_Q28[ ind + 1 ][ nb ] -
                        silk_Transition_LP_B_Q28[ ind     ][ nb ],
                        fac_Q16 );
                }
                for( na = 0; na < TRANSITION_NA; na++ ) {
                    A_Q28[ na ] = silk_SMLAWB(
                        silk_Transition_LP_A_Q28[ ind     ][ na ],
                        silk_Transition_LP_A_Q28[ ind + 1 ][ na ] -
                        silk_Transition_LP_A_Q28[ ind     ][ na ],
                        fac_Q16 );
                }
            } else { /* ( fac_Q16 - ( 1 << 16 ) ) is in range of a 16-bit int */
                silk_assert( fac_Q16 - ( 1 << 16 ) == silk_SAT16( fac_Q16 - ( 1 << 16 ) ) );
                /* Piece-wise linear interpolation of B and A */
                for( nb = 0; nb < TRANSITION_NB; nb++ ) {
                    B_Q28[ nb ] = silk_SMLAWB(
                        silk_Transition_LP_B_Q28[ ind + 1 ][ nb ],
                        silk_Transition_LP_B_Q28[ ind + 1 ][ nb ] -
                        silk_Transition_LP_B_Q28[ ind     ][ nb ],
                        fac_Q16 - ( (opus_int32)1 << 16 ) );
                }
                for( na = 0; na < TRANSITION_NA; na++ ) {
                    A_Q28[ na ] = silk_SMLAWB(
                        silk_Transition_LP_A_Q28[ ind + 1 ][ na ],
                        silk_Transition_LP_A_Q28[ ind + 1 ][ na ] -
                        silk_Transition_LP_A_Q28[ ind     ][ na ],
                        fac_Q16 - ( (opus_int32)1 << 16 ) );
                }
            }
        } else {
            silk_memcpy( B_Q28, silk_Transition_LP_B_Q28[ ind ], TRANSITION_NB * sizeof( opus_int32 ) );
            silk_memcpy( A_Q28, silk_Transition_LP_A_Q28[ ind ], TRANSITION_NA * sizeof( opus_int32 ) );
        }
    } else {
        silk_memcpy( B_Q28, silk_Transition_LP_B_Q28[ TRANSITION_INT_NUM - 1 ], TRANSITION_NB * sizeof( opus_int32 ) );
        silk_memcpy( A_Q28, silk_Transition_LP_A_Q28[ TRANSITION_INT_NUM - 1 ], TRANSITION_NA * sizeof( opus_int32 ) );
    }
}

void silk_LPC_analysis_filter_FLP(
    silk_float                      r_LPC[],                            /* O    LPC residual signal                         */
    const silk_float                PredCoef[],                         /* I    LPC coefficients                            */
    const silk_float                s[],                                /* I    Input signal                                */
    const opus_int                  length,                             /* I    Length of input signal                      */
    const opus_int                  Order                               /* I    LPC order                                   */
)
{
    silk_assert( Order <= length );

    switch( Order ) {
        case 6:
            silk_LPC_analysis_filter6_FLP(  r_LPC, PredCoef, s, length );
        break;

        case 8:
            silk_LPC_analysis_filter8_FLP(  r_LPC, PredCoef, s, length );
        break;

        case 10:
            silk_LPC_analysis_filter10_FLP( r_LPC, PredCoef, s, length );
        break;

        case 12:
            silk_LPC_analysis_filter12_FLP( r_LPC, PredCoef, s, length );
        break;

        case 16:
            silk_LPC_analysis_filter16_FLP( r_LPC, PredCoef, s, length );
        break;

        default:
            silk_assert( 0 );
        break;
    }

    /* Set first Order output samples to zero */
    silk_memset( r_LPC, 0, Order * sizeof( silk_float ) );
}

/*
* Prefilter for finding Quantizer input signal
*/
static OPUS_INLINE void silk_prefilt_FLP(
    silk_prefilter_state_FLP    *P,                 /* I/O state */
    silk_float                  st_res[],           /* I */
    silk_float                  xw[],               /* O */
    silk_float                  *HarmShapeFIR,      /* I */
    silk_float                  Tilt,               /* I */
    silk_float                  LF_MA_shp,          /* I */
    silk_float                  LF_AR_shp,          /* I */
    opus_int                    lag,                /* I */
    opus_int                    length              /* I */
)
{
    opus_int   i;
    opus_int   idx, LTP_shp_buf_idx;
    silk_float n_Tilt, n_LF, n_LTP;
    silk_float sLF_AR_shp, sLF_MA_shp;
    silk_float *LTP_shp_buf;

    /* To speed up use temp variables instead of using the struct */
    LTP_shp_buf     = P->sLTP_shp;
    LTP_shp_buf_idx = P->sLTP_shp_buf_idx;
    sLF_AR_shp      = P->sLF_AR_shp;
    sLF_MA_shp      = P->sLF_MA_shp;

    for( i = 0; i < length; i++ ) {
        if( lag > 0 ) {
            silk_assert( HARM_SHAPE_FIR_TAPS == 3 );
            idx = lag + LTP_shp_buf_idx;
            n_LTP  = LTP_shp_buf[ ( idx - HARM_SHAPE_FIR_TAPS / 2 - 1) & LTP_MASK ] * HarmShapeFIR[ 0 ];
            n_LTP += LTP_shp_buf[ ( idx - HARM_SHAPE_FIR_TAPS / 2    ) & LTP_MASK ] * HarmShapeFIR[ 1 ];
            n_LTP += LTP_shp_buf[ ( idx - HARM_SHAPE_FIR_TAPS / 2 + 1) & LTP_MASK ] * HarmShapeFIR[ 2 ];
        } else {
            n_LTP = 0;
        }

        n_Tilt = sLF_AR_shp * Tilt;
        n_LF   = sLF_AR_shp * LF_AR_shp + sLF_MA_shp * LF_MA_shp;

        sLF_AR_shp = st_res[ i ] - n_Tilt;
        sLF_MA_shp = sLF_AR_shp - n_LF;

        LTP_shp_buf_idx = ( LTP_shp_buf_idx - 1 ) & LTP_MASK;
        LTP_shp_buf[ LTP_shp_buf_idx ] = sLF_MA_shp;

        xw[ i ] = sLF_MA_shp - n_LTP;
    }
    /* Copy temp variable back to state */
    P->sLF_AR_shp       = sLF_AR_shp;
    P->sLF_MA_shp       = sLF_MA_shp;
    P->sLTP_shp_buf_idx = LTP_shp_buf_idx;
}

/* Deactivate by setting psEncC->mode = 0;                  */
void silk_LP_variable_cutoff(
    silk_LP_state               *psLP,                          /* I/O  LP filter state                             */
    opus_int16                  *frame,                         /* I/O  Low-pass filtered output signal             */
    const opus_int              frame_length