static OPUS_INLINE opus_int silk_setup_LBRR(
    silk_encoder_state          *psEncC,            /* I/O                      */
    const opus_int32            TargetRate_bps      /* I                        */
)
{
    opus_int   LBRR_in_previous_packet, ret = SILK_NO_ERROR;
    opus_int32 LBRR_rate_thres_bps;

    LBRR_in_previous_packet = psEncC->LBRR_enabled;
    psEncC->LBRR_enabled = 0;
    if( psEncC->useInBandFEC && psEncC->PacketLoss_perc > 0 ) {
        if( psEncC->fs_kHz == 8 ) {
            LBRR_rate_thres_bps = LBRR_NB_MIN_RATE_BPS;
        } else if( psEncC->fs_kHz == 12 ) {
            LBRR_rate_thres_bps = LBRR_MB_MIN_RATE_BPS;
        } else {
            LBRR_rate_thres_bps = LBRR_WB_MIN_RATE_BPS;
        }
        LBRR_rate_thres_bps = silk_SMULWB( silk_MUL( LBRR_rate_thres_bps, 125 - silk_min( psEncC->PacketLoss_perc, 25 ) ), SILK_FIX_CONST( 0.01, 16 ) );

        if( TargetRate_bps > LBRR_rate_thres_bps ) {
            /* Set gain increase for coding LBRR excitation */
            if( LBRR_in_previous_packet == 0 ) {
                /* Previous packet did not have LBRR, and was therefore coded at a higher bitrate */
                psEncC->LBRR_GainIncreases = 7;
            } else {
                psEncC->LBRR_GainIncreases = silk_max_int( 7 - silk_SMULWB( (opus_int32)psEncC->PacketLoss_perc, SILK_FIX_CONST( 0.4, 16 ) ), 2 );
            }
            psEncC->LBRR_enabled = 1;
        }
    }

    return ret;
}

/* Find least-squares prediction gain for one signal based on another and quantize it */
int32_t silk_stereo_find_predictor(	/* O    Returns predictor in Q13                    */
					     int32_t * ratio_Q14,	/* O    Ratio of residual and mid energies          */
					     const int16_t x[],	/* I    Basis signal                                */
					     const int16_t y[],	/* I    Target signal                               */
					     int32_t mid_res_amp_Q0[],	/* I/O  Smoothed mid, residual norms                */
					     int length,	/* I    Number of samples                           */
					     int smooth_coef_Q16	/* I    Smoothing coefficient                       */
    )
{
	int scale, scale1, scale2;
	int32_t nrgx, nrgy, corr, pred_Q13, pred2_Q10;

	/* Find  predictor */
	silk_sum_sqr_shift(&nrgx, &scale1, x, length);
	silk_sum_sqr_shift(&nrgy, &scale2, y, length);
	scale = silk_max_int(scale1, scale2);
	scale = scale + (scale & 1);	/* make even */
	nrgy = silk_RSHIFT32(nrgy, scale - scale2);
	nrgx = silk_RSHIFT32(nrgx, scale - scale1);
	nrgx = silk_max_int(nrgx, 1);
	corr = silk_inner_prod_aligned_scale(x, y, scale, length);
	pred_Q13 = silk_DIV32_varQ(corr, nrgx, 13);
	pred_Q13 = silk_LIMIT(pred_Q13, -(1 << 14), 1 << 14);
	pred2_Q10 = silk_SMULWB(pred_Q13, pred_Q13);

	/* Faster update for signals with large prediction parameters */
	smooth_coef_Q16 =
	    (int) silk_max_int(smooth_coef_Q16, silk_abs(pred2_Q10));

	/* Smoothed mid and residual norms */
	assert(smooth_coef_Q16 < 32768);
	scale = silk_RSHIFT(scale, 1);
	mid_res_amp_Q0[0] =
	    silk_SMLAWB(mid_res_amp_Q0[0],
			silk_LSHIFT(silk_SQRT_APPROX(nrgx),
				    scale) - mid_res_amp_Q0[0],
			smooth_coef_Q16);
	/* Residual energy = nrgy - 2 * pred * corr + pred^2 * nrgx */
	nrgy = silk_SUB_LSHIFT32(nrgy, silk_SMULWB(corr, pred_Q13), 3 + 1);
	nrgy = silk_ADD_LSHIFT32(nrgy, silk_SMULWB(nrgx, pred2_Q10), 6);
	mid_res_amp_Q0[1] =
	    silk_SMLAWB(mid_res_amp_Q0[1],
			silk_LSHIFT(silk_SQRT_APPROX(nrgy),
				    scale) - mid_res_amp_Q0[1],
			smooth_coef_Q16);

	/* Ratio of smoothed residual and mid norms */
	*ratio_Q14 =
	    silk_DIV32_varQ(mid_res_amp_Q0[1], silk_max(mid_res_amp_Q0[0], 1),
			    14);
	*ratio_Q14 = silk_LIMIT(*ratio_Q14, 0, 32767);

	return pred_Q13;
}

/* Gain scalar quantization with hysteresis, uniform on log scale */
void silk_gains_quant(
    opus_int8                   ind[ MAX_NB_SUBFR ],            /* O    gain indices                                */
    opus_int32                  gain_Q16[ MAX_NB_SUBFR ],       /* I/O  gains (quantized out)                       */
    opus_int8                   *prev_ind,                      /* I/O  last index in previous frame                */
    const opus_int              conditional,                    /* I    first gain is delta coded if 1              */
    const opus_int              nb_subfr                        /* I    number of subframes                         */
)
{
    opus_int k, double_step_size_threshold;

    for( k = 0; k < nb_subfr; k++ ) {
        /* Convert to log scale, scale, floor() */
        ind[ k ] = silk_SMULWB( SCALE_Q16, silk_lin2log( gain_Q16[ k ] ) - OFFSET );

        /* Round towards previous quantized gain (hysteresis) */
        if( ind[ k ] < *prev_ind ) {
            ind[ k ]++;
        }
        ind[ k ] = silk_LIMIT_int( ind[ k ], 0, N_LEVELS_QGAIN - 1 );

        /* Compute delta indices and limit */
        if( k == 0 && conditional == 0 ) {
            /* Full index */
            ind[ k ] = silk_LIMIT_int( ind[ k ], *prev_ind + MIN_DELTA_GAIN_QUANT, N_LEVELS_QGAIN - 1 );
            *prev_ind = ind[ k ];
        } else {
            /* Delta index */
            ind[ k ] = ind[ k ] - *prev_ind;

            /* Double the quantization step size for large gain increases, so that the max gain level can be reached */
            double_step_size_threshold = 2 * MAX_DELTA_GAIN_QUANT - N_LEVELS_QGAIN + *prev_ind;
            if( ind[ k ] > double_step_size_threshold ) {
                ind[ k ] = double_step_size_threshold + silk_RSHIFT( ind[ k ] - double_step_size_threshold + 1, 1 );
            }

            ind[ k ] = silk_LIMIT_int( ind[ k ], MIN_DELTA_GAIN_QUANT, MAX_DELTA_GAIN_QUANT );

            /* Accumulate deltas */
            if( ind[ k ] > double_step_size_threshold ) {
                *prev_ind += silk_LSHIFT( ind[ k ], 1 ) - double_step_size_threshold;
                *prev_ind = silk_min_int( *prev_ind, N_LEVELS_QGAIN - 1 );
            } else {
                *prev_ind += ind[ k ];
            }

            /* Shift to make non-negative */
            ind[ k ] -= MIN_DELTA_GAIN_QUANT;
        }

        /* Scale and convert to linear scale */
        gain_Q16[ k ] = silk_log2lin( silk_min_32( silk_SMULWB( INV_SCALE_Q16, *prev_ind ) + OFFSET, 3967 ) ); /* 3967 = 31 in Q7 */
    }
}

/* Gains scalar dequantization, uniform on log scale */
void silk_gains_dequant(
    opus_int32                  gain_Q16[ MAX_NB_SUBFR ],       /* O    quantized gains                             */
    const opus_int8             ind[ MAX_NB_SUBFR ],            /* I    gain indices                                */
    opus_int8                   *prev_ind,                      /* I/O  last index in previous frame                */
    const opus_int              conditional,                    /* I    first gain is delta coded if 1              */
    const opus_int              nb_subfr                        /* I    number of subframes                          */
)
{
    opus_int   k, ind_tmp, double_step_size_threshold;

    for( k = 0; k < nb_subfr; k++ ) {
        if( k == 0 && conditional == 0 ) {
            /* Gain index is not allowed to go down more than 16 steps (~21.8 dB) */
            *prev_ind = silk_max_int( ind[ k ], *prev_ind - 16 );
        } else {
            /* Delta index */
            ind_tmp = ind[ k ] + MIN_DELTA_GAIN_QUANT;

            /* Accumulate deltas */
            double_step_size_threshold = 2 * MAX_DELTA_GAIN_QUANT - N_LEVELS_QGAIN + *prev_ind;
            if( ind_tmp > double_step_size_threshold ) {
                *prev_ind += silk_LSHIFT( ind_tmp, 1 ) - double_step_size_threshold;
            } else {
                *prev_ind += ind_tmp;
            }
        }
        *prev_ind = silk_LIMIT_int( *prev_ind, 0, N_LEVELS_QGAIN - 1 );

        /* Scale and convert to linear scale */
        gain_Q16[ k ] = silk_log2lin( silk_min_32( silk_SMULWB( INV_SCALE_Q16, *prev_ind ) + OFFSET, 3967 ) ); /* 3967 = 31 in Q7 */
    }
}

static OPUS_INLINE opus_int16 *silk_resampler_private_IIR_FIR_INTERPOL(
    opus_int16  *out,
    opus_int16  *buf,
    opus_int32  max_index_Q16,
    opus_int32  index_increment_Q16
)
{
    opus_int32 index_Q16, res_Q15;
    opus_int16 *buf_ptr;
    opus_int32 table_index;

    /* Interpolate upsampled signal and store in output array */
    for( index_Q16 = 0; index_Q16 < max_index_Q16; index_Q16 += index_increment_Q16 ) {
        table_index = silk_SMULWB( index_Q16 & 0xFFFF, 12 );
        buf_ptr = &buf[ index_Q16 >> 16 ];

        res_Q15 = silk_SMULBB(          buf_ptr[ 0 ], silk_resampler_frac_FIR_12[      table_index ][ 0 ] );
        res_Q15 = silk_SMLABB( res_Q15, buf_ptr[ 1 ], silk_resampler_frac_FIR_12[      table_index ][ 1 ] );
        res_Q15 = silk_SMLABB( res_Q15, buf_ptr[ 2 ], silk_resampler_frac_FIR_12[      table_index ][ 2 ] );
        res_Q15 = silk_SMLABB( res_Q15, buf_ptr[ 3 ], silk_resampler_frac_FIR_12[      table_index ][ 3 ] );
        res_Q15 = silk_SMLABB( res_Q15, buf_ptr[ 4 ], silk_resampler_frac_FIR_12[ 11 - table_index ][ 3 ] );
        res_Q15 = silk_SMLABB( res_Q15, buf_ptr[ 5 ], silk_resampler_frac_FIR_12[ 11 - table_index ][ 2 ] );
        res_Q15 = silk_SMLABB( res_Q15, buf_ptr[ 6 ], silk_resampler_frac_FIR_12[ 11 - table_index ][ 1 ] );
        res_Q15 = silk_SMLABB( res_Q15, buf_ptr[ 7 ], silk_resampler_frac_FIR_12[ 11 - table_index ][ 0 ] );
        *out++ = (opus_int16)silk_SAT16( silk_RSHIFT_ROUND( res_Q15, 15 ) );
    }
    return out;
}

/* Decode mid/side predictors */
void silk_stereo_decode_pred(
    ec_dec                      *psRangeDec,                    /* I/O  Compressor data structure                   */
    opus_int32                  pred_Q13[]                      /* O    Predictors                                  */
)
{
    opus_int   n, ix[ 2 ][ 3 ];
    opus_int32 low_Q13, step_Q13;

    /* Entropy decoding */
    n = ec_dec_icdf( psRangeDec, silk_stereo_pred_joint_iCDF, 8 );
    ix[ 0 ][ 2 ] = silk_DIV32_16( n, 5 );
    ix[ 1 ][ 2 ] = n - 5 * ix[ 0 ][ 2 ];
    for( n = 0; n < 2; n++ ) {
        ix[ n ][ 0 ] = ec_dec_icdf( psRangeDec, silk_uniform3_iCDF, 8 );
        ix[ n ][ 1 ] = ec_dec_icdf( psRangeDec, silk_uniform5_iCDF, 8 );
    }

    /* Dequantize */
    for( n = 0; n < 2; n++ ) {
        ix[ n ][ 0 ] += 3 * ix[ n ][ 2 ];
        low_Q13 = silk_stereo_pred_quant_Q13[ ix[ n ][ 0 ] ];
        step_Q13 = silk_SMULWB( silk_stereo_pred_quant_Q13[ ix[ n ][ 0 ] + 1 ] - low_Q13,
            SILK_FIX_CONST( 0.5 / STEREO_QUANT_SUB_STEPS, 16 ) );
        pred_Q13[ n ] = silk_SMLABB( low_Q13, step_Q13, 2 * ix[ n ][ 1 ] + 1 );
    }

    /* Subtract second from first predictor (helps when actually applying these) */
    pred_Q13[ 0 ] -= pred_Q13[ 1 ];
}

/* High-pass filter with cutoff frequency adaptation based on pitch lag statistics */
void silk_HP_variable_cutoff(silk_encoder_state_Fxx state_Fxx[]	/* I/O  Encoder states                              */
    )
{
	int quality_Q15;
	int32_t pitch_freq_Hz_Q16, pitch_freq_log_Q7, delta_freq_Q7;
	silk_encoder_state *psEncC1 = &state_Fxx[0].sCmn;

	/* Adaptive cutoff frequency: estimate low end of pitch frequency range */
	if (psEncC1->prevSignalType == TYPE_VOICED) {
		/* difference, in log domain */
		pitch_freq_Hz_Q16 =
		    silk_DIV32_16(silk_LSHIFT
				  (silk_MUL(psEncC1->fs_kHz, 1000), 16),
				  psEncC1->prevLag);
		pitch_freq_log_Q7 = silk_lin2log(pitch_freq_Hz_Q16) - (16 << 7);

		/* adjustment based on quality */
		quality_Q15 = psEncC1->input_quality_bands_Q15[0];
		pitch_freq_log_Q7 =
		    silk_SMLAWB(pitch_freq_log_Q7,
				silk_SMULWB(silk_LSHIFT(-quality_Q15, 2),
					    quality_Q15),
				pitch_freq_log_Q7 -
				(silk_lin2log
				 (SILK_FIX_CONST(VARIABLE_HP_MIN_CUTOFF_HZ, 16))
				 - (16 << 7)));

		/* delta_freq = pitch_freq_log - psEnc->variable_HP_smth1; */
		delta_freq_Q7 =
		    pitch_freq_log_Q7 -
		    silk_RSHIFT(psEncC1->variable_HP_smth1_Q15, 8);
		if (delta_freq_Q7 < 0) {
			/* less smoothing for decreasing pitch frequency, to track something close to the minimum */
			delta_freq_Q7 = silk_MUL(delta_freq_Q7, 3);
		}

		/* limit delta, to reduce impact of outliers in pitch estimation */
		delta_freq_Q7 =
		    silk_LIMIT_32(delta_freq_Q7,
				  -SILK_FIX_CONST(VARIABLE_HP_MAX_DELTA_FREQ,
						  7),
				  SILK_FIX_CONST(VARIABLE_HP_MAX_DELTA_FREQ,
						 7));

		/* update smoother */
		psEncC1->variable_HP_smth1_Q15 =
		    silk_SMLAWB(psEncC1->variable_HP_smth1_Q15,
				silk_SMULBB(psEncC1->speech_activity_Q8,
					    delta_freq_Q7),
				SILK_FIX_CONST(VARIABLE_HP_SMTH_COEF1, 16));

		/* limit frequency range */
		psEncC1->variable_HP_smth1_Q15 =
		    silk_LIMIT_32(psEncC1->variable_HP_smth1_Q15,
				  silk_LSHIFT(silk_lin2log
					      (VARIABLE_HP_MIN_CUTOFF_HZ), 8),
				  silk_LSHIFT(silk_lin2log
					      (VARIABLE_HP_MAX_CUTOFF_HZ), 8));
	}
}

/* Glues concealed frames with new good received frames */
void silk_PLC_glue_frames(
    silk_decoder_state                  *psDec,             /* I/O decoder state        */
    opus_int16                          frame[],            /* I/O signal               */
    opus_int                            length              /* I length of signal       */
)
{
    opus_int   i, energy_shift;
    opus_int32 energy;
    silk_PLC_struct *psPLC;
    psPLC = &psDec->sPLC;

    if( psDec->lossCnt ) {
        /* Calculate energy in concealed residual */
        silk_sum_sqr_shift( &psPLC->conc_energy, &psPLC->conc_energy_shift, frame, length );

        psPLC->last_frame_lost = 1;
    } else {
        if( psDec->sPLC.last_frame_lost ) {
            /* Calculate residual in decoded signal if last frame was lost */
            silk_sum_sqr_shift( &energy, &energy_shift, frame, length );

            /* Normalize energies */
            if( energy_shift > psPLC->conc_energy_shift ) {
                psPLC->conc_energy = silk_RSHIFT( psPLC->conc_energy, energy_shift - psPLC->conc_energy_shift );
            } else if( energy_shift < psPLC->conc_energy_shift ) {
                energy = silk_RSHIFT( energy, psPLC->conc_energy_shift - energy_shift );
            }

            /* Fade in the energy difference */
            if( energy > psPLC->conc_energy ) {
                opus_int32 frac_Q24, LZ;
                opus_int32 gain_Q16, slope_Q16;

                LZ = silk_CLZ32( psPLC->conc_energy );
                LZ = LZ - 1;
                psPLC->conc_energy = silk_LSHIFT( psPLC->conc_energy, LZ );
                energy = silk_RSHIFT( energy, silk_max_32( 24 - LZ, 0 ) );

                frac_Q24 = silk_DIV32( psPLC->conc_energy, silk_max( energy, 1 ) );

                gain_Q16 = silk_LSHIFT( silk_SQRT_APPROX( frac_Q24 ), 4 );
                slope_Q16 = silk_DIV32_16( ( (opus_int32)1 << 16 ) - gain_Q16, length );
                /* Make slope 4x steeper to avoid missing onsets after DTX */
                slope_Q16 = silk_LSHIFT( slope_Q16, 2 );

                for( i = 0; i < length; i++ ) {
                    frame[ i ] = silk_SMULWB( gain_Q16, frame[ i ] );
                    gain_Q16 += slope_Q16;
                    if( gain_Q16 > (opus_int32)1 << 16 ) {
                        break;
                    }
                }
            }
        }
        psPLC->last_frame_lost = 0;
    }
}

/* 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;
}

/* Second order ARMA filter, alternative implementation */
void silk_biquad_alt(const int16_t * in,	/* I     input signal                                               */
                     const int32_t * B_Q28,	/* I     MA coefficients [3]                                        */
                     const int32_t * A_Q28,	/* I     AR coefficients [2]                                        */
                     int32_t * S,	/* I/O   State vector [2]                                           */
                     int16_t * out,	/* O     output signal                                              */
                     const int32_t len,	/* I     signal length (must be even)                               */
                     int stride	/* I     Operate on interleaved signal if > 1                       */
                    )
{
    /* DIRECT FORM II TRANSPOSED (uses 2 element state vector) */
    int k;
    int32_t inval, A0_U_Q28, A0_L_Q28, A1_U_Q28, A1_L_Q28, out32_Q14;

    /* Negate A_Q28 values and split in two parts */
    A0_L_Q28 = (-A_Q28[0]) & 0x00003FFF;	/* lower part */
    A0_U_Q28 = silk_RSHIFT(-A_Q28[0], 14);	/* upper part */
    A1_L_Q28 = (-A_Q28[1]) & 0x00003FFF;	/* lower part */
    A1_U_Q28 = silk_RSHIFT(-A_Q28[1], 14);	/* upper part */

    for (k = 0; k < len; k++) {
        /* S[ 0 ], S[ 1 ]: Q12 */
        inval = in[k * stride];
        out32_Q14 = silk_LSHIFT(silk_SMLAWB(S[0], B_Q28[0], inval), 2);

        S[0] =
            S[1] + silk_RSHIFT_ROUND(silk_SMULWB(out32_Q14, A0_L_Q28),
                                     14);
        S[0] = silk_SMLAWB(S[0], out32_Q14, A0_U_Q28);
        S[0] = silk_SMLAWB(S[0], B_Q28[1], inval);

        S[1] = silk_RSHIFT_ROUND(silk_SMULWB(out32_Q14, A1_L_Q28), 14);
        S[1] = silk_SMLAWB(S[1], out32_Q14, A1_U_Q28);
        S[1] = silk_SMLAWB(S[1], B_Q28[2], inval);

        /* Scale back to Q0 and saturate */
        out[k * stride] =
            (int16_t)
            silk_SAT16(silk_RSHIFT(out32_Q14 + (1 << 14) - 1, 14));
    }
}

/* Copy and multiply a vector by a constant */
void silk_scale_copy_vector16(
    opus_int16                  *data_out,
    const opus_int16            *data_in,
    opus_int32                  gain_Q16,           /* I    Gain in Q16                                                 */
    const opus_int              dataSize            /* I    Length                                                      */
)
{
    opus_int  i;
    opus_int32 tmp32;

    for(i = 0; i < dataSize; i++) {
        tmp32 = silk_SMULWB(gain_Q16, data_in[ i ]);
        data_out[ i ] = (opus_int16)silk_CHECK_FIT16(tmp32);
    }
}

/* 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 {

void silk_noise_shape_analysis_FIX(
    silk_encoder_state_FIX          *psEnc,                                 /* I/O  Encoder state FIX                                                           */
    silk_encoder_control_FIX        *psEncCtrl,                             /* I/O  Encoder control FIX                                                         */
    const opus_int16                *pitch_res,                             /* I    LPC residual from pitch analysis                                            */
    const opus_int16                *x,                                     /* I    Input signal [ frame_length + la_shape ]                                    */
    int                              arch                                   /* I    Run-time architecture                                                       */
)
{
    silk_shape_state_FIX *psShapeSt = &psEnc->sShape;
    opus_int     k, i, nSamples, Qnrg, b_Q14, warping_Q16, scale = 0;
    opus_int32   SNR_adj_dB_Q7, HarmBoost_Q16, HarmShapeGain_Q16, Tilt_Q16, tmp32;
    opus_int32   nrg, pre_nrg_Q30, log_energy_Q7, log_energy_prev_Q7, energy_variation_Q7;
    opus_int32   delta_Q16, BWExp1_Q16, BWExp2_Q16, gain_mult_Q16, gain_add_Q16, strength_Q16, b_Q8;
    opus_int32   auto_corr[     MAX_SHAPE_LPC_ORDER + 1 ];
    opus_int32   refl_coef_Q16[ MAX_SHAPE_LPC_ORDER ];
    opus_int32   AR1_Q24[       MAX_SHAPE_LPC_ORDER ];
    opus_int32   AR2_Q24[       MAX_SHAPE_LPC_ORDER ];
    VARDECL( opus_int16, x_windowed );
    const opus_int16 *x_ptr, *pitch_res_ptr;
    SAVE_STACK;

    /* Point to start of first LPC analysis block */
    x_ptr = x - psEnc->sCmn.la_shape;

    /****************/
    /* GAIN CONTROL */
    /****************/
    SNR_adj_dB_Q7 = psEnc->sCmn.SNR_dB_Q7;

    /* Input quality is the average of the quality in the lowest two VAD bands */
    psEncCtrl->input_quality_Q14 = ( opus_int )silk_RSHIFT( (opus_int32)psEnc->sCmn.input_quality_bands_Q15[ 0 ]
        + psEnc->sCmn.input_quality_bands_Q15[ 1 ], 2 );

    /* Coding quality level, between 0.0_Q0 and 1.0_Q0, but in Q14 */
    psEncCtrl->coding_quality_Q14 = silk_RSHIFT( silk_sigm_Q15( silk_RSHIFT_ROUND( SNR_adj_dB_Q7 -
        SILK_FIX_CONST( 20.0, 7 ), 4 ) ), 1 );

    /* Reduce coding SNR during low speech activity */
    if( psEnc->sCmn.useCBR == 0 ) {
        b_Q8 = SILK_FIX_CONST( 1.0, 8 ) - psEnc->sCmn.speech_activity_Q8;
        b_Q8 = silk_SMULWB( silk_LSHIFT( b_Q8, 8 ), b_Q8 );
        SNR_adj_dB_Q7 = silk_SMLAWB( SNR_adj_dB_Q7,
            silk_SMULBB( SILK_FIX_CONST( -BG_SNR_DECR_dB, 7 ) >> ( 4 + 1 ), b_Q8 ),                                       /* Q11*/
            silk_SMULWB( SILK_FIX_CONST( 1.0, 14 ) + psEncCtrl->input_quality_Q14, psEncCtrl->coding_quality_Q14 ) );     /* Q12*/
    }

/* Second order AR filter with single delay elements */
void silk_resampler_private_AR2(int32_t S[],	/* I/O  State vector [ 2 ]          */
				int32_t out_Q8[],	/* O    Output signal               */
				const int16_t in[],	/* I    Input signal                */
				const int16_t A_Q14[],	/* I    AR coefficients, Q14        */
				int32_t len	/* I    Signal length               */
    )
{
	int32_t k;
	int32_t out32;

	for (k = 0; k < len; k++) {
		out32 = silk_ADD_LSHIFT32(S[0], (int32_t) in[k], 8);
		out_Q8[k] = out32;
		out32 = silk_LSHIFT(out32, 2);
		S[0] = silk_SMLAWB(S[1], out32, A_Q14[0]);
		S[1] = silk_SMULWB(out32, A_Q14[1]);
	}
}

/* Second order AR filter with single delay elements */
void silk_resampler_private_AR2(
    opus_int32                      S[],            /* I/O  State vector [ 2 ]          */
    opus_int32                      out_Q8[],       /* O    Output signal               */
    const opus_int16                in[],           /* I    Input signal                */
    const opus_int16                A_Q14[],        /* I    AR coefficients, Q14        */
    opus_int32                      len             /* I    Signal length               */
)
{
    opus_int32    k;
    opus_int32    out32;

    for(k = 0; k < len; k++) {
        out32       = silk_ADD_LSHIFT32(S[ 0 ], (opus_int32)in[ k ], 8);
        out_Q8[ k ] = out32;
        out32       = silk_LSHIFT(out32, 2);
        S[ 0 ]      = silk_SMLAWB(S[ 1 ], out32, A_Q14[ 0 ]);
        S[ 1 ]      = silk_SMULWB(out32, A_Q14[ 1 ]);
    }
}

/* Calculation of LTP state scaling */
void silk_LTP_scale_ctrl_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                                       */
)
{
    opus_int round_loss;

    if( condCoding == CODE_INDEPENDENTLY ) {
        /* Only scale if first frame in packet */
        round_loss = psEnc->sCmn.PacketLoss_perc + psEnc->sCmn.nFramesPerPacket;
        psEnc->sCmn.indices.LTP_scaleIndex = (opus_int8)silk_LIMIT(
            silk_SMULWB( silk_SMULBB( round_loss, psEncCtrl->LTPredCodGain_Q7 ), SILK_FIX_CONST( 0.1, 9 ) ), 0, 2 );
    } else {
        /* Default is minimum scaling */
        psEnc->sCmn.indices.LTP_scaleIndex = 0;
    }
    psEncCtrl->LTP_scale_Q14 = silk_LTPScales_table_Q14[ psEnc->sCmn.indices.LTP_scaleIndex ];
}

void silk_warped_LPC_analysis_filter_FIX(
          opus_int32            state[],                    /* I/O  State [order + 1]                   */
          opus_int16            res[],                      /* O    Residual signal [length]            */
    const opus_int16            coef_Q13[],                 /* I    Coefficients [order]                */
    const opus_int16            input[],                    /* I    Input signal [length]               */
    const opus_int16            lambda_Q16,                 /* I    Warping factor                      */
    const opus_int              length,                     /* I    Length of input signal              */
    const opus_int              order                       /* I    Filter order (even)                 */
)
{
    opus_int     n, i;
    opus_int32   acc_Q11, tmp1, tmp2;

    /* Order must be even */
    silk_assert( ( order & 1 ) == 0 );

    for( n = 0; n < length; n++ ) {
        /* Output of lowpass section */
        tmp2 = silk_SMLAWB( state[ 0 ], state[ 1 ], lambda_Q16 );
        state[ 0 ] = silk_LSHIFT( input[ n ], 14 );
        /* Output of allpass section */
        tmp1 = silk_SMLAWB( state[ 1 ], state[ 2 ] - tmp2, lambda_Q16 );
        state[ 1 ] = tmp2;
        acc_Q11 = silk_SMULWB( tmp2, coef_Q13[ 0 ] );
        /* Loop over allpass sections */
        for( i = 2; i < order; i += 2 ) {
            /* Output of allpass section */
            tmp2 = silk_SMLAWB( state[ i ], state[ i + 1 ] - tmp1, lambda_Q16 );
            state[ i ] = tmp1;
            acc_Q11 = silk_SMLAWB( acc_Q11, tmp1, coef_Q13[ i - 1 ] );
            /* Output of allpass section */
            tmp1 = silk_SMLAWB( state[ i + 1 ], state[ i + 2 ] - tmp2, lambda_Q16 );
            state[ i + 1 ] = tmp2;
            acc_Q11 = silk_SMLAWB( acc_Q11, tmp2, coef_Q13[ i ] );
        }
        state[ order ] = tmp1;
        acc_Q11 = silk_SMLAWB( acc_Q11, tmp1, coef_Q13[ order - 1 ] );
        res[ n ] = ( opus_int16 )silk_SAT16( ( opus_int32 )input[ n ] - silk_RSHIFT_ROUND( acc_Q11, 11 ) );
    }
}

/* Quantize mid/side predictors */
void silk_stereo_quant_pred(
    opus_int32                  pred_Q13[],                     /* I/O  Predictors (out: quantized)                 */
    opus_int8                   ix[ 2 ][ 3 ]                    /* O    Quantization indices                        */
)
{
    opus_int   i, j, n;
    opus_int32 low_Q13, step_Q13, lvl_Q13, err_min_Q13, err_Q13, quant_pred_Q13 = 0;

    /* Quantize */
    for( n = 0; n < 2; n++ ) {
        /* Brute-force search over quantization levels */
        err_min_Q13 = silk_int32_MAX;
        for( i = 0; i < STEREO_QUANT_TAB_SIZE - 1; i++ ) {
            low_Q13 = silk_stereo_pred_quant_Q13[ i ];
            step_Q13 = silk_SMULWB( silk_stereo_pred_quant_Q13[ i + 1 ] - low_Q13,
                SILK_FIX_CONST( 0.5 / STEREO_QUANT_SUB_STEPS, 16 ) );
            for( j = 0; j < STEREO_QUANT_SUB_STEPS; j++ ) {
                lvl_Q13 = silk_SMLABB( low_Q13, step_Q13, 2 * j + 1 );
                err_Q13 = silk_abs( pred_Q13[ n ] - lvl_Q13 );
                if( err_Q13 < err_min_Q13 ) {
                    err_min_Q13 = err_Q13;
                    quant_pred_Q13 = lvl_Q13;
                    ix[ n ][ 0 ] = i;
                    ix[ n ][ 1 ] = j;
                } else {
                    /* Error increasing, so we're past the optimum */
                    goto done;
                }
            }
        }
        done:
        ix[ n ][ 2 ]  = silk_DIV32_16( ix[ n ][ 0 ], 3 );
        ix[ n ][ 0 ] -= ix[ n ][ 2 ] * 3;
        pred_Q13[ n ] = quant_pred_Q13;
    }

    /* Subtract second from first predictor (helps when actually applying these) */
    pred_Q13[ 0 ] -= pred_Q13[ 1 ];
}

/* Split signal into two decimated bands using first-order allpass filters */
void silk_ana_filt_bank_1(
    const opus_int16            *in,                /* I    Input signal [N]                                            */
    opus_int32                  *S,                 /* I/O  State vector [2]                                            */
    opus_int16                  *outL,              /* O    Low band [N/2]                                              */
    opus_int16                  *outH,              /* O    High band [N/2]                                             */
    const opus_int32            N                   /* I    Number of input samples                                     */
)
{
    opus_int      k, N2 = silk_RSHIFT( N, 1 );
    opus_int32    in32, X, Y, out_1, out_2;

    /* Internal variables and state are in Q10 format */
    for( k = 0; k < N2; k++ ) {
        /* Convert to Q10 */
        in32 = silk_LSHIFT( (opus_int32)in[ 2 * k ], 10 );

        /* All-pass section for even input sample */
        Y      = silk_SUB32( in32, S[ 0 ] );
        X      = silk_SMLAWB( Y, Y, A_fb1_21 );
        out_1  = silk_ADD32( S[ 0 ], X );
        S[ 0 ] = silk_ADD32( in32, X );

        /* Convert to Q10 */
        in32 = silk_LSHIFT( (opus_int32)in[ 2 * k + 1 ], 10 );

        /* All-pass section for odd input sample, and add to output of previous section */
        Y      = silk_SUB32( in32, S[ 1 ] );
        X      = silk_SMULWB( Y, A_fb1_20 );
        out_2  = silk_ADD32( S[ 1 ], X );
        S[ 1 ] = silk_ADD32( in32, X );

        /* Add/subtract, convert back to int16 and store to output */
        outL[ k ] = (opus_int16)silk_SAT16( silk_RSHIFT_ROUND( silk_ADD32( out_2, out_1 ), 11 ) );
        outH[ k ] = (opus_int16)silk_SAT16( silk_RSHIFT_ROUND( silk_SUB32( out_2, out_1 ), 11 ) );
    }
}

static OPUS_INLINE void silk_PLC_conceal(
    silk_decoder_state                  *psDec,             /* I/O Decoder state        */
    silk_decoder_control                *psDecCtrl,         /* I/O Decoder control      */
    opus_int16                          frame[],            /* O LPC residual signal    */
    int                                 arch                /* I Run-time architecture  */
)
{
    opus_int   i, j, k;
    opus_int   lag, idx, sLTP_buf_idx, shift1, shift2;
    opus_int32 rand_seed, harm_Gain_Q15, rand_Gain_Q15, inv_gain_Q30;
    opus_int32 energy1, energy2, *rand_ptr, *pred_lag_ptr;
    opus_int32 LPC_pred_Q10, LTP_pred_Q12;
    opus_int16 rand_scale_Q14;
    opus_int16 *B_Q14;
    opus_int32 *sLPC_Q14_ptr;
    opus_int16 A_Q12[ MAX_LPC_ORDER ];
#ifdef SMALL_FOOTPRINT
    opus_int16 *sLTP;
#else
    VARDECL( opus_int16, sLTP );
#endif
    VARDECL( opus_int32, sLTP_Q14 );
    silk_PLC_struct *psPLC = &psDec->sPLC;
    opus_int32 prevGain_Q10[2];
    SAVE_STACK;

    ALLOC( sLTP_Q14, psDec->ltp_mem_length + psDec->frame_length, opus_int32 );
#ifdef SMALL_FOOTPRINT
    /* Ugly hack that breaks aliasing rules to save stack: put sLTP at the very end of sLTP_Q14. */
    sLTP = ((opus_int16*)&sLTP_Q14[psDec->ltp_mem_length + psDec->frame_length])-psDec->ltp_mem_length;
#else
    ALLOC( sLTP, psDec->ltp_mem_length, opus_int16 );
#endif

    prevGain_Q10[0] = silk_RSHIFT( psPLC->pre