Add m-esr52 at 52.6.0

7 files changed, 2541 insertions, 0 deletions

diff --git a/media/libopus/silk/x86/NSQ_del_dec_sse.c b/media/libopus/silk/x86/NSQ_del_dec_sse.c
new file mode 100644
index 000000000..21d4a8bc1
--- /dev/null
+++ b/media/libopus/silk/x86/NSQ_del_dec_sse.c
@@ -0,0 +1,857 @@
+/* Copyright (c) 2014, Cisco Systems, INC
+   Written by XiangMingZhu WeiZhou MinPeng YanWang
+
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions
+   are met:
+
+   - Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+
+   - Redistributions in binary form must reproduce the above copyright
+   notice, this list of conditions and the following disclaimer in the
+   documentation and/or other materials provided with the distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
+   OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+   EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+   PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+   PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+   LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+   NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+   SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+
+#ifdef HAVE_CONFIG_H
+#include "config.h"
+#endif
+
+#include <xmmintrin.h>
+#include <emmintrin.h>
+#include <smmintrin.h>
+#include "main.h"
+#include "celt/x86/x86cpu.h"
+
+#include "stack_alloc.h"
+
+typedef struct {
+    opus_int32 sLPC_Q14[ MAX_SUB_FRAME_LENGTH + NSQ_LPC_BUF_LENGTH ];
+    opus_int32 RandState[ DECISION_DELAY ];
+    opus_int32 Q_Q10[     DECISION_DELAY ];
+    opus_int32 Xq_Q14[    DECISION_DELAY ];
+    opus_int32 Pred_Q15[  DECISION_DELAY ];
+    opus_int32 Shape_Q14[ DECISION_DELAY ];
+    opus_int32 sAR2_Q14[ MAX_SHAPE_LPC_ORDER ];
+    opus_int32 LF_AR_Q14;
+    opus_int32 Seed;
+    opus_int32 SeedInit;
+    opus_int32 RD_Q10;
+} NSQ_del_dec_struct;
+
+typedef struct {
+    opus_int32 Q_Q10;
+    opus_int32 RD_Q10;
+    opus_int32 xq_Q14;
+    opus_int32 LF_AR_Q14;
+    opus_int32 sLTP_shp_Q14;
+    opus_int32 LPC_exc_Q14;
+} NSQ_sample_struct;
+
+typedef NSQ_sample_struct  NSQ_sample_pair[ 2 ];
+
+static OPUS_INLINE void silk_nsq_del_dec_scale_states_sse4_1(
+    const silk_encoder_state *psEncC,               /* I    Encoder State                       */
+    silk_nsq_state      *NSQ,                       /* I/O  NSQ state                           */
+    NSQ_del_dec_struct  psDelDec[],                 /* I/O  Delayed decision states             */
+    const opus_int32    x_Q3[],                     /* I    Input in Q3                         */
+    opus_int32          x_sc_Q10[],                 /* O    Input scaled with 1/Gain in Q10     */
+    const opus_int16    sLTP[],                     /* I    Re-whitened LTP state in Q0         */
+    opus_int32          sLTP_Q15[],                 /* O    LTP state matching scaled input     */
+    opus_int            subfr,                      /* I    Subframe number                     */
+    opus_int            nStatesDelayedDecision,     /* I    Number of del dec states            */
+    const opus_int      LTP_scale_Q14,              /* I    LTP state scaling                   */
+    const opus_int32    Gains_Q16[ MAX_NB_SUBFR ],  /* I                                        */
+    const opus_int      pitchL[ MAX_NB_SUBFR ],     /* I    Pitch lag                           */
+    const opus_int      signal_type,                /* I    Signal type                         */
+    const opus_int      decisionDelay               /* I    Decision delay                      */
+);
+
+/******************************************/
+/* Noise shape quantizer for one subframe */
+/******************************************/
+static OPUS_INLINE void silk_noise_shape_quantizer_del_dec_sse4_1(
+    silk_nsq_state      *NSQ,                   /* I/O  NSQ state                           */
+    NSQ_del_dec_struct  psDelDec[],             /* I/O  Delayed decision states             */
+    opus_int            signalType,             /* I    Signal type                         */
+    const opus_int32    x_Q10[],                /* I                                        */
+    opus_int8           pulses[],               /* O                                        */
+    opus_int16          xq[],                   /* O                                        */
+    opus_int32          sLTP_Q15[],             /* I/O  LTP filter state                    */
+    opus_int32          delayedGain_Q10[],      /* I/O  Gain delay buffer                   */
+    const opus_int16    a_Q12[],                /* I    Short term prediction coefs         */
+    const opus_int16    b_Q14[],                /* I    Long term prediction coefs          */
+    const opus_int16    AR_shp_Q13[],           /* I    Noise shaping coefs                 */
+    opus_int            lag,                    /* I    Pitch lag                           */
+    opus_int32          HarmShapeFIRPacked_Q14, /* I                                        */
+    opus_int            Tilt_Q14,               /* I    Spectral tilt                       */
+    opus_int32          LF_shp_Q14,             /* I                                        */
+    opus_int32          Gain_Q16,               /* I                                        */
+    opus_int            Lambda_Q10,             /* I                                        */
+    opus_int            offset_Q10,             /* I                                        */
+    opus_int            length,                 /* I    Input length                        */
+    opus_int            subfr,                  /* I    Subframe number                     */
+    opus_int            shapingLPCOrder,        /* I    Shaping LPC filter order            */
+    opus_int            predictLPCOrder,        /* I    Prediction filter order             */
+    opus_int            warping_Q16,            /* I                                        */
+    opus_int            nStatesDelayedDecision, /* I    Number of states in decision tree   */
+    opus_int            *smpl_buf_idx,          /* I    Index to newest samples in buffers  */
+    opus_int            decisionDelay           /* I                                        */
+);
+
+void silk_NSQ_del_dec_sse4_1(
+    const silk_encoder_state    *psEncC,                                    /* I/O  Encoder State                   */
+    silk_nsq_state              *NSQ,                                       /* I/O  NSQ state                       */
+    SideInfoIndices             *psIndices,                                 /* I/O  Quantization Indices            */
+    const opus_int32            x_Q3[],                                     /* I    Prefiltered input signal        */
+    opus_int8                   pulses[],                                   /* O    Quantized pulse signal          */
+    const opus_int16            PredCoef_Q12[ 2 * MAX_LPC_ORDER ],          /* I    Short term prediction coefs     */
+    const opus_int16            LTPCoef_Q14[ LTP_ORDER * MAX_NB_SUBFR ],    /* I    Long term prediction coefs      */
+    const opus_int16            AR2_Q13[ MAX_NB_SUBFR * MAX_SHAPE_LPC_ORDER ], /* I Noise shaping coefs             */
+    const opus_int              HarmShapeGain_Q14[ MAX_NB_SUBFR ],          /* I    Long term shaping coefs         */
+    const opus_int              Tilt_Q14[ MAX_NB_SUBFR ],                   /* I    Spectral tilt                   */
+    const opus_int32            LF_shp_Q14[ MAX_NB_SUBFR ],                 /* I    Low frequency shaping coefs     */
+    const opus_int32            Gains_Q16[ MAX_NB_SUBFR ],                  /* I    Quantization step sizes         */
+    const opus_int              pitchL[ MAX_NB_SUBFR ],                     /* I    Pitch lags                      */
+    const opus_int              Lambda_Q10,                                 /* I    Rate/distortion tradeoff        */
+    const opus_int              LTP_scale_Q14                               /* I    LTP state scaling               */
+)
+{
+    opus_int            i, k, lag, start_idx, LSF_interpolation_flag, Winner_ind, subfr;
+    opus_int            last_smple_idx, smpl_buf_idx, decisionDelay;
+    const opus_int16    *A_Q12, *B_Q14, *AR_shp_Q13;
+    opus_int16          *pxq;
+    VARDECL( opus_int32, sLTP_Q15 );
+    VARDECL( opus_int16, sLTP );
+    opus_int32          HarmShapeFIRPacked_Q14;
+    opus_int            offset_Q10;
+    opus_int32          RDmin_Q10, Gain_Q10;
+    VARDECL( opus_int32, x_sc_Q10 );
+    VARDECL( opus_int32, delayedGain_Q10 );
+    VARDECL( NSQ_del_dec_struct, psDelDec );
+    NSQ_del_dec_struct  *psDD;
+    SAVE_STACK;
+
+    /* Set unvoiced lag to the previous one, overwrite later for voiced */
+    lag = NSQ->lagPrev;
+
+    silk_assert( NSQ->prev_gain_Q16 != 0 );
+
+    /* Initialize delayed decision states */
+    ALLOC( psDelDec, psEncC->nStatesDelayedDecision, NSQ_del_dec_struct );
+    silk_memset( psDelDec, 0, psEncC->nStatesDelayedDecision * sizeof( NSQ_del_dec_struct ) );
+    for( k = 0; k < psEncC->nStatesDelayedDecision; k++ ) {
+        psDD                 = &psDelDec[ k ];
+        psDD->Seed           = ( k + psIndices->Seed ) & 3;
+        psDD->SeedInit       = psDD->Seed;
+        psDD->RD_Q10         = 0;
+        psDD->LF_AR_Q14      = NSQ->sLF_AR_shp_Q14;
+        psDD->Shape_Q14[ 0 ] = NSQ->sLTP_shp_Q14[ psEncC->ltp_mem_length - 1 ];
+        silk_memcpy( psDD->sLPC_Q14, NSQ->sLPC_Q14, NSQ_LPC_BUF_LENGTH * sizeof( opus_int32 ) );
+        silk_memcpy( psDD->sAR2_Q14, NSQ->sAR2_Q14, sizeof( NSQ->sAR2_Q14 ) );
+    }
+
+    offset_Q10   = silk_Quantization_Offsets_Q10[ psIndices->signalType >> 1 ][ psIndices->quantOffsetType ];
+    smpl_buf_idx = 0; /* index of oldest samples */
+
+    decisionDelay = silk_min_int( DECISION_DELAY, psEncC->subfr_length );
+
+    /* For voiced frames limit the decision delay to lower than the pitch lag */
+    if( psIndices->signalType == TYPE_VOICED ) {
+        for( k = 0; k < psEncC->nb_subfr; k++ ) {
+            decisionDelay = silk_min_int( decisionDelay, pitchL[ k ] - LTP_ORDER / 2 - 1 );
+        }
+    } else {
+        if( lag > 0 ) {
+            decisionDelay = silk_min_int( decisionDelay, lag - LTP_ORDER / 2 - 1 );
+        }
+    }
+
+    if( psIndices->NLSFInterpCoef_Q2 == 4 ) {
+        LSF_interpolation_flag = 0;
+    } else {
+        LSF_interpolation_flag = 1;
+    }
+
+    ALLOC( sLTP_Q15,
+           psEncC->ltp_mem_length + psEncC->frame_length, opus_int32 );
+    ALLOC( sLTP, psEncC->ltp_mem_length + psEncC->frame_length, opus_int16 );
+    ALLOC( x_sc_Q10, psEncC->subfr_length, opus_int32 );
+    ALLOC( delayedGain_Q10, DECISION_DELAY, opus_int32 );
+    /* Set up pointers to start of sub frame */
+    pxq                   = &NSQ->xq[ psEncC->ltp_mem_length ];
+    NSQ->sLTP_shp_buf_idx = psEncC->ltp_mem_length;
+    NSQ->sLTP_buf_idx     = psEncC->ltp_mem_length;
+    subfr = 0;
+    for( k = 0; k < psEncC->nb_subfr; k++ ) {
+        A_Q12      = &PredCoef_Q12[ ( ( k >> 1 ) | ( 1 - LSF_interpolation_flag ) ) * MAX_LPC_ORDER ];
+        B_Q14      = &LTPCoef_Q14[ k * LTP_ORDER           ];
+        AR_shp_Q13 = &AR2_Q13[     k * MAX_SHAPE_LPC_ORDER ];
+
+        /* Noise shape parameters */
+        silk_assert( HarmShapeGain_Q14[ k ] >= 0 );
+        HarmShapeFIRPacked_Q14  =                          silk_RSHIFT( HarmShapeGain_Q14[ k ], 2 );
+        HarmShapeFIRPacked_Q14 |= silk_LSHIFT( (opus_int32)silk_RSHIFT( HarmShapeGain_Q14[ k ], 1 ), 16 );
+
+        NSQ->rewhite_flag = 0;
+        if( psIndices->signalType == TYPE_VOICED ) {
+            /* Voiced */
+            lag = pitchL[ k ];
+
+            /* Re-whitening */
+            if( ( k & ( 3 - silk_LSHIFT( LSF_interpolation_flag, 1 ) ) ) == 0 ) {
+                if( k == 2 ) {
+                    /* RESET DELAYED DECISIONS */
+                    /* Find winner */
+                    RDmin_Q10 = psDelDec[ 0 ].RD_Q10;
+                    Winner_ind = 0;
+                    for( i = 1; i < psEncC->nStatesDelayedDecision; i++ ) {
+                        if( psDelDec[ i ].RD_Q10 < RDmin_Q10 ) {
+                            RDmin_Q10 = psDelDec[ i ].RD_Q10;
+                            Winner_ind = i;
+                        }
+                    }
+                    for( i = 0; i < psEncC->nStatesDelayedDecision; i++ ) {
+                        if( i != Winner_ind ) {
+                            psDelDec[ i ].RD_Q10 += ( silk_int32_MAX >> 4 );
+                            silk_assert( psDelDec[ i ].RD_Q10 >= 0 );
+                        }
+                    }
+
+                    /* Copy final part of signals from winner state to output and long-term filter states */
+                    psDD = &psDelDec[ Winner_ind ];
+                    last_smple_idx = smpl_buf_idx + decisionDelay;
+                    for( i = 0; i < decisionDelay; i++ ) {
+                        last_smple_idx = ( last_smple_idx - 1 ) & DECISION_DELAY_MASK;
+                        pulses[   i - decisionDelay ] = (opus_int8)silk_RSHIFT_ROUND( psDD->Q_Q10[ last_smple_idx ], 10 );
+                        pxq[ i - decisionDelay ] = (opus_int16)silk_SAT16( silk_RSHIFT_ROUND(
+                            silk_SMULWW( psDD->Xq_Q14[ last_smple_idx ], Gains_Q16[ 1 ] ), 14 ) );
+                        NSQ->sLTP_shp_Q14[ NSQ->sLTP_shp_buf_idx - decisionDelay + i ] = psDD->Shape_Q14[ last_smple_idx ];
+                    }
+
+                    subfr = 0;
+                }
+
+                /* Rewhiten with new A coefs */
+                start_idx = psEncC->ltp_mem_length - lag - psEncC->predictLPCOrder - LTP_ORDER / 2;
+                silk_assert( start_idx > 0 );
+
+                silk_LPC_analysis_filter( &sLTP[ start_idx ], &NSQ->xq[ start_idx + k * psEncC->subfr_length ],
+                    A_Q12, psEncC->ltp_mem_length - start_idx, psEncC->predictLPCOrder, psEncC->arch );
+
+                NSQ->sLTP_buf_idx = psEncC->ltp_mem_length;
+                NSQ->rewhite_flag = 1;
+            }
+        }
+
+        silk_nsq_del_dec_scale_states_sse4_1( psEncC, NSQ, psDelDec, x_Q3, x_sc_Q10, sLTP, sLTP_Q15, k,
+            psEncC->nStatesDelayedDecision, LTP_scale_Q14, Gains_Q16, pitchL, psIndices->signalType, decisionDelay );
+
+        silk_noise_shape_quantizer_del_dec_sse4_1( NSQ, psDelDec, psIndices->signalType, x_sc_Q10, pulses, pxq, sLTP_Q15,
+            delayedGain_Q10, A_Q12, B_Q14, AR_shp_Q13, lag, HarmShapeFIRPacked_Q14, Tilt_Q14[ k ], LF_shp_Q14[ k ],
+            Gains_Q16[ k ], Lambda_Q10, offset_Q10, psEncC->subfr_length, subfr++, psEncC->shapingLPCOrder,
+            psEncC->predictLPCOrder, psEncC->warping_Q16, psEncC->nStatesDelayedDecision, &smpl_buf_idx, decisionDelay );
+
+        x_Q3   += psEncC->subfr_length;
+        pulses += psEncC->subfr_length;
+        pxq    += psEncC->subfr_length;
+    }
+
+    /* Find winner */
+    RDmin_Q10 = psDelDec[ 0 ].RD_Q10;
+    Winner_ind = 0;
+    for( k = 1; k < psEncC->nStatesDelayedDecision; k++ ) {
+        if( psDelDec[ k ].RD_Q10 < RDmin_Q10 ) {
+            RDmin_Q10 = psDelDec[ k ].RD_Q10;
+            Winner_ind = k;
+        }
+    }
+
+    /* Copy final part of signals from winner state to output and long-term filter states */
+    psDD = &psDelDec[ Winner_ind ];
+    psIndices->Seed = psDD->SeedInit;
+    last_smple_idx = smpl_buf_idx + decisionDelay;
+    Gain_Q10 = silk_RSHIFT32( Gains_Q16[ psEncC->nb_subfr - 1 ], 6 );
+    for( i = 0; i < decisionDelay; i++ ) {
+        last_smple_idx = ( last_smple_idx - 1 ) & DECISION_DELAY_MASK;
+        pulses[   i - decisionDelay ] = (opus_int8)silk_RSHIFT_ROUND( psDD->Q_Q10[ last_smple_idx ], 10 );
+        pxq[ i - decisionDelay ] = (opus_int16)silk_SAT16( silk_RSHIFT_ROUND(
+            silk_SMULWW( psDD->Xq_Q14[ last_smple_idx ], Gain_Q10 ), 8 ) );
+        NSQ->sLTP_shp_Q14[ NSQ->sLTP_shp_buf_idx - decisionDelay + i ] = psDD->Shape_Q14[ last_smple_idx ];
+    }
+    silk_memcpy( NSQ->sLPC_Q14, &psDD->sLPC_Q14[ psEncC->subfr_length ], NSQ_LPC_BUF_LENGTH * sizeof( opus_int32 ) );
+    silk_memcpy( NSQ->sAR2_Q14, psDD->sAR2_Q14, sizeof( psDD->sAR2_Q14 ) );
+
+    /* Update states */
+    NSQ->sLF_AR_shp_Q14 = psDD->LF_AR_Q14;
+    NSQ->lagPrev        = pitchL[ psEncC->nb_subfr - 1 ];
+
+    /* Save quantized speech signal */
+    /* DEBUG_STORE_DATA( enc.pcm, &NSQ->xq[psEncC->ltp_mem_length], psEncC->frame_length * sizeof( opus_int16 ) ) */
+    silk_memmove( NSQ->xq,           &NSQ->xq[           psEncC->frame_length ], psEncC->ltp_mem_length * sizeof( opus_int16 ) );
+    silk_memmove( NSQ->sLTP_shp_Q14, &NSQ->sLTP_shp_Q14[ psEncC->frame_length ], psEncC->ltp_mem_length * sizeof( opus_int32 ) );
+    RESTORE_STACK;
+}
+
+/******************************************/
+/* Noise shape quantizer for one subframe */
+/******************************************/
+static OPUS_INLINE void silk_noise_shape_quantizer_del_dec_sse4_1(
+    silk_nsq_state      *NSQ,                   /* I/O  NSQ state                           */
+    NSQ_del_dec_struct  psDelDec[],             /* I/O  Delayed decision states             */
+    opus_int            signalType,             /* I    Signal type                         */
+    const opus_int32    x_Q10[],                /* I                                        */
+    opus_int8           pulses[],               /* O                                        */
+    opus_int16          xq[],                   /* O                                        */
+    opus_int32          sLTP_Q15[],             /* I/O  LTP filter state                    */
+    opus_int32          delayedGain_Q10[],      /* I/O  Gain delay buffer                   */
+    const opus_int16    a_Q12[],                /* I    Short term prediction coefs         */
+    const opus_int16    b_Q14[],                /* I    Long term prediction coefs          */
+    const opus_int16    AR_shp_Q13[],           /* I    Noise shaping coefs                 */
+    opus_int            lag,                    /* I    Pitch lag                           */
+    opus_int32          HarmShapeFIRPacked_Q14, /* I                                        */
+    opus_int            Tilt_Q14,               /* I    Spectral tilt                       */
+    opus_int32          LF_shp_Q14,             /* I                                        */
+    opus_int32          Gain_Q16,               /* I                                        */
+    opus_int            Lambda_Q10,             /* I                                        */
+    opus_int            offset_Q10,             /* I                                        */
+    opus_int            length,                 /* I    Input length                        */
+    opus_int            subfr,                  /* I    Subframe number                     */
+    opus_int            shapingLPCOrder,        /* I    Shaping LPC filter order            */
+    opus_int            predictLPCOrder,        /* I    Prediction filter order             */
+    opus_int            warping_Q16,            /* I                                        */
+    opus_int            nStatesDelayedDecision, /* I    Number of states in decision tree   */
+    opus_int            *smpl_buf_idx,          /* I    Index to newest samples in buffers  */
+    opus_int            decisionDelay           /* I                                        */
+)
+{
+    opus_int     i, j, k, Winner_ind, RDmin_ind, RDmax_ind, last_smple_idx;
+    opus_int32   Winner_rand_state;
+    opus_int32   LTP_pred_Q14, LPC_pred_Q14, n_AR_Q14, n_LTP_Q14;
+    opus_int32   n_LF_Q14, r_Q10, rr_Q10, rd1_Q10, rd2_Q10, RDmin_Q10, RDmax_Q10;
+    opus_int32   q1_Q0, q1_Q10, q2_Q10, exc_Q14, LPC_exc_Q14, xq_Q14, Gain_Q10;
+    opus_int32   tmp1, tmp2, sLF_AR_shp_Q14;
+    opus_int32   *pred_lag_ptr, *shp_lag_ptr, *psLPC_Q14;
+    VARDECL( NSQ_sample_pair, psSampleState );
+    NSQ_del_dec_struct *psDD;
+    NSQ_sample_struct  *psSS;
+
+    __m128i a_Q12_0123, a_Q12_4567, a_Q12_89AB, a_Q12_CDEF;
+    __m128i b_Q12_0123, b_sr_Q12_0123;
+    SAVE_STACK;
+
+    silk_assert( nStatesDelayedDecision > 0 );
+    ALLOC( psSampleState, nStatesDelayedDecision, NSQ_sample_pair );
+
+    shp_lag_ptr  = &NSQ->sLTP_shp_Q14[ NSQ->sLTP_shp_buf_idx - lag + HARM_SHAPE_FIR_TAPS / 2 ];
+    pred_lag_ptr = &sLTP_Q15[ NSQ->sLTP_buf_idx - lag + LTP_ORDER / 2 ];
+    Gain_Q10     = silk_RSHIFT( Gain_Q16, 6 );
+
+    a_Q12_0123 = OP_CVTEPI16_EPI32_M64( a_Q12 );
+    a_Q12_4567 = OP_CVTEPI16_EPI32_M64( a_Q12 + 4 );
+
+    if( opus_likely( predictLPCOrder == 16 ) ) {
+        a_Q12_89AB = OP_CVTEPI16_EPI32_M64( a_Q12 + 8 );
+        a_Q12_CDEF = OP_CVTEPI16_EPI32_M64( a_Q12 + 12 );
+    }
+
+    if( signalType == TYPE_VOICED ){
+        b_Q12_0123 = OP_CVTEPI16_EPI32_M64( b_Q14 );
+        b_sr_Q12_0123 = _mm_shuffle_epi32( b_Q12_0123, _MM_SHUFFLE( 0, 3, 2, 1 ) ); /* equal shift right 4 bytes */
+    }
+    for( i = 0; i < length; i++ ) {
+        /* Perform common calculations used in all states */
+
+        /* Long-term prediction */
+        if( signalType == TYPE_VOICED ) {
+            /* Unrolled loop */
+            /* Avoids introducing a bias because silk_SMLAWB() always rounds to -inf */
+            LTP_pred_Q14 = 2;
+            {
+                __m128i tmpa, tmpb, pred_lag_ptr_tmp;
+                pred_lag_ptr_tmp    = _mm_loadu_si128( (__m128i *)(&pred_lag_ptr[ -3 ] ) );
+                pred_lag_ptr_tmp    = _mm_shuffle_epi32( pred_lag_ptr_tmp, 0x1B );
+                tmpa                = _mm_mul_epi32( pred_lag_ptr_tmp, b_Q12_0123 );
+                tmpa                = _mm_srli_si128( tmpa, 2 );
+
+                pred_lag_ptr_tmp = _mm_shuffle_epi32( pred_lag_ptr_tmp, _MM_SHUFFLE( 0, 3, 2, 1 ) );/* equal shift right 4 bytes */
+                pred_lag_ptr_tmp    = _mm_mul_epi32( pred_lag_ptr_tmp, b_sr_Q12_0123 );
+                pred_lag_ptr_tmp    = _mm_srli_si128( pred_lag_ptr_tmp, 2 );
+                pred_lag_ptr_tmp    = _mm_add_epi32( pred_lag_ptr_tmp, tmpa );
+
+                tmpb = _mm_shuffle_epi32( pred_lag_ptr_tmp, _MM_SHUFFLE( 0, 0, 3, 2 ) );/* equal shift right 8 bytes */
+                pred_lag_ptr_tmp    = _mm_add_epi32( pred_lag_ptr_tmp, tmpb );
+                LTP_pred_Q14        += _mm_cvtsi128_si32( pred_lag_ptr_tmp );
+
+                LTP_pred_Q14 = silk_SMLAWB( LTP_pred_Q14, pred_lag_ptr[ -4 ], b_Q14[ 4 ] );
+                LTP_pred_Q14 = silk_LSHIFT( LTP_pred_Q14, 1 );                          /* Q13 -> Q14 */
+                pred_lag_ptr++;
+            }
+        } else {
+            LTP_pred_Q14 = 0;
+        }
+
+        /* Long-term shaping */
+        if( lag > 0 ) {
+            /* Symmetric, packed FIR coefficients */
+            n_LTP_Q14 = silk_SMULWB( silk_ADD32( shp_lag_ptr[ 0 ], shp_lag_ptr[ -2 ] ), HarmShapeFIRPacked_Q14 );
+            n_LTP_Q14 = silk_SMLAWT( n_LTP_Q14, shp_lag_ptr[ -1 ],                      HarmShapeFIRPacked_Q14 );
+            n_LTP_Q14 = silk_SUB_LSHIFT32( LTP_pred_Q14, n_LTP_Q14, 2 );            /* Q12 -> Q14 */
+            shp_lag_ptr++;
+        } else {
+            n_LTP_Q14 = 0;
+        }
+        {
+            __m128i tmpa, tmpb, psLPC_Q14_tmp, a_Q12_tmp;
+
+            for( k = 0; k < nStatesDelayedDecision; k++ ) {
+                /* Delayed decision state */
+                psDD = &psDelDec[ k ];
+
+                /* Sample state */
+                psSS = psSampleState[ k ];
+
+                /* Generate dither */
+                psDD->Seed = silk_RAND( psDD->Seed );
+
+                /* Pointer used in short term prediction and shaping */
+                psLPC_Q14 = &psDD->sLPC_Q14[ NSQ_LPC_BUF_LENGTH - 1 + i ];
+                /* Short-term prediction */
+                silk_assert( predictLPCOrder == 10 || predictLPCOrder == 16 );
+                /* Avoids introducing a bias because silk_SMLAWB() always rounds to -inf */
+                LPC_pred_Q14 = silk_RSHIFT( predictLPCOrder, 1 );
+
+                tmpb = _mm_setzero_si128();
+
+                /* step 1 */
+                psLPC_Q14_tmp   = _mm_loadu_si128( (__m128i *)(&psLPC_Q14[ -3 ] ) ); /* -3, -2 , -1, 0 */
+                psLPC_Q14_tmp   = _mm_shuffle_epi32( psLPC_Q14_tmp, 0x1B );      /* 0, -1, -2, -3 */
+                tmpa            = _mm_mul_epi32( psLPC_Q14_tmp, a_Q12_0123 );    /* 0, -1, -2, -3 * 0123 -> 0*0, 2*-2 */
+
+                tmpa            = _mm_srli_epi64( tmpa, 16 );
+                tmpb            = _mm_add_epi32( tmpb, tmpa );
+
+                psLPC_Q14_tmp = _mm_shuffle_epi32( psLPC_Q14_tmp, _MM_SHUFFLE( 0, 3, 2, 1 ) ); /* equal shift right 4 bytes */
+                a_Q12_tmp = _mm_shuffle_epi32( a_Q12_0123, _MM_SHUFFLE(0, 3, 2, 1 ) ); /* equal shift right 4 bytes */
+                psLPC_Q14_tmp   = _mm_mul_epi32( psLPC_Q14_tmp, a_Q12_tmp ); /* 1*-1, 3*-3 */
+                psLPC_Q14_tmp   = _mm_srli_epi64( psLPC_Q14_tmp, 16 );
+                tmpb            = _mm_add_epi32( tmpb, psLPC_Q14_tmp );
+
+                /* step 2 */
+                psLPC_Q14_tmp   = _mm_loadu_si128( (__m128i *)(&psLPC_Q14[ -7 ] ) );
+                psLPC_Q14_tmp   = _mm_shuffle_epi32( psLPC_Q14_tmp, 0x1B );
+                tmpa            = _mm_mul_epi32( psLPC_Q14_tmp, a_Q12_4567 );
+                tmpa            = _mm_srli_epi64( tmpa, 16 );
+                tmpb            = _mm_add_epi32( tmpb, tmpa );
+
+                psLPC_Q14_tmp = _mm_shuffle_epi32( psLPC_Q14_tmp, _MM_SHUFFLE( 0, 3, 2, 1 ) ); /* equal shift right 4 bytes */
+                a_Q12_tmp = _mm_shuffle_epi32( a_Q12_4567, _MM_SHUFFLE(0, 3, 2, 1 ) ); /* equal shift right 4 bytes */
+                psLPC_Q14_tmp   = _mm_mul_epi32( psLPC_Q14_tmp, a_Q12_tmp );
+                psLPC_Q14_tmp   = _mm_srli_epi64( psLPC_Q14_tmp, 16 );
+                tmpb            = _mm_add_epi32( tmpb, psLPC_Q14_tmp );
+
+                if ( opus_likely( predictLPCOrder == 16 ) )
+                {
+                    /* step 3 */
+                    psLPC_Q14_tmp   = _mm_loadu_si128( (__m128i *)(&psLPC_Q14[ -11 ] ) );
+                    psLPC_Q14_tmp   = _mm_shuffle_epi32( psLPC_Q14_tmp, 0x1B );
+                    tmpa            = _mm_mul_epi32( psLPC_Q14_tmp, a_Q12_89AB );
+                    tmpa            = _mm_srli_epi64( tmpa, 16 );
+                    tmpb            = _mm_add_epi32( tmpb, tmpa );
+
+                    psLPC_Q14_tmp = _mm_shuffle_epi32( psLPC_Q14_tmp, _MM_SHUFFLE( 0, 3, 2, 1 ) ); /* equal shift right 4 bytes */
+                    a_Q12_tmp = _mm_shuffle_epi32( a_Q12_89AB, _MM_SHUFFLE(0, 3, 2, 1 ) );/* equal shift right 4 bytes */
+                    psLPC_Q14_tmp   = _mm_mul_epi32( psLPC_Q14_tmp, a_Q12_tmp );
+                    psLPC_Q14_tmp   = _mm_srli_epi64( psLPC_Q14_tmp, 16 );
+                    tmpb            = _mm_add_epi32( tmpb, psLPC_Q14_tmp );
+
+                    /* setp 4 */
+                    psLPC_Q14_tmp   = _mm_loadu_si128( (__m128i *)(&psLPC_Q14[ -15 ] ) );
+                    psLPC_Q14_tmp   = _mm_shuffle_epi32( psLPC_Q14_tmp, 0x1B );
+                    tmpa            = _mm_mul_epi32( psLPC_Q14_tmp, a_Q12_CDEF );
+                    tmpa            = _mm_srli_epi64( tmpa, 16 );
+                    tmpb            = _mm_add_epi32( tmpb, tmpa );
+
+                    psLPC_Q14_tmp = _mm_shuffle_epi32( psLPC_Q14_tmp, _MM_SHUFFLE( 0, 3, 2, 1 ) ); /* equal shift right 4 bytes */
+                    a_Q12_tmp = _mm_shuffle_epi32( a_Q12_CDEF, _MM_SHUFFLE(0, 3, 2, 1 ) ); /* equal shift right 4 bytes */
+                    psLPC_Q14_tmp   = _mm_mul_epi32( psLPC_Q14_tmp, a_Q12_tmp );
+                    psLPC_Q14_tmp   = _mm_srli_epi64( psLPC_Q14_tmp, 16 );
+                    tmpb            = _mm_add_epi32( tmpb, psLPC_Q14_tmp );
+
+                    /* add at last */
+                    /* equal shift right 8 bytes*/
+                    tmpa            = _mm_shuffle_epi32( tmpb, _MM_SHUFFLE( 0, 0, 3, 2 ) );
+                    tmpb            = _mm_add_epi32( tmpb, tmpa );
+                    LPC_pred_Q14    += _mm_cvtsi128_si32( tmpb );
+                }
+                else
+                {
+                    /* add at last */
+                    tmpa            = _mm_shuffle_epi32( tmpb, _MM_SHUFFLE( 0, 0, 3, 2 ) ); /* equal shift right 8 bytes*/
+                    tmpb            = _mm_add_epi32( tmpb, tmpa );
+                    LPC_pred_Q14    += _mm_cvtsi128_si32( tmpb );
+
+                    LPC_pred_Q14 = silk_SMLAWB( LPC_pred_Q14, psLPC_Q14[ -8 ], a_Q12[ 8 ] );
+                    LPC_pred_Q14 = silk_SMLAWB( LPC_pred_Q14, psLPC_Q14[ -9 ], a_Q12[ 9 ] );
+                }
+
+                LPC_pred_Q14 = silk_LSHIFT( LPC_pred_Q14, 4 ); /* Q10 -> Q14 */
+
+                /* Noise shape feedback */
+                silk_assert( ( shapingLPCOrder & 1 ) == 0 );   /* check that order is even */
+                /* Output of lowpass section */
+                tmp2 = silk_SMLAWB( psLPC_Q14[ 0 ], psDD->sAR2_Q14[ 0 ], warping_Q16 );
+                /* Output of allpass section */
+                tmp1 = silk_SMLAWB( psDD->sAR2_Q14[ 0 ], psDD->sAR2_Q14[ 1 ] - tmp2, warping_Q16 );
+                psDD->sAR2_Q14[ 0 ] = tmp2;
+                n_AR_Q14 = silk_RSHIFT( shapingLPCOrder, 1 );
+                n_AR_Q14 = silk_SMLAWB( n_AR_Q14, tmp2, AR_shp_Q13[ 0 ] );
+                /* Loop over allpass sections */
+                for( j = 2; j < shapingLPCOrder; j += 2 ) {
+                    /* Output of allpass section */
+                    tmp2 = silk_SMLAWB( psDD->sAR2_Q14[ j - 1 ], psDD->sAR2_Q14[ j + 0 ] - tmp1, warping_Q16 );
+                    psDD->sAR2_Q14[ j - 1 ] = tmp1;
+                    n_AR_Q14 = silk_SMLAWB( n_AR_Q14, tmp1, AR_shp_Q13[ j - 1 ] );
+                    /* Output of allpass section */
+                    tmp1 = silk_SMLAWB( psDD->sAR2_Q14[ j + 0 ], psDD->sAR2_Q14[ j + 1 ] - tmp2, warping_Q16 );
+                    psDD->sAR2_Q14[ j + 0 ] = tmp2;
+                    n_AR_Q14 = silk_SMLAWB( n_AR_Q14, tmp2, AR_shp_Q13[ j ] );
+                }
+                psDD->sAR2_Q14[ shapingLPCOrder - 1 ] = tmp1;
+                n_AR_Q14 = silk_SMLAWB( n_AR_Q14, tmp1, AR_shp_Q13[ shapingLPCOrder - 1 ] );
+
+                n_AR_Q14 = silk_LSHIFT( n_AR_Q14, 1 );                                      /* Q11 -> Q12 */
+                n_AR_Q14 = silk_SMLAWB( n_AR_Q14, psDD->LF_AR_Q14, Tilt_Q14 );              /* Q12 */
+                n_AR_Q14 = silk_LSHIFT( n_AR_Q14, 2 );                                      /* Q12 -> Q14 */
+
+                n_LF_Q14 = silk_SMULWB( psDD->Shape_Q14[ *smpl_buf_idx ], LF_shp_Q14 );     /* Q12 */
+                n_LF_Q14 = silk_SMLAWT( n_LF_Q14, psDD->LF_AR_Q14, LF_shp_Q14 );            /* Q12 */
+                n_LF_Q14 = silk_LSHIFT( n_LF_Q14, 2 );                                      /* Q12 -> Q14 */
+
+                /* Input minus prediction plus noise feedback                       */
+                /* r = x[ i ] - LTP_pred - LPC_pred + n_AR + n_Tilt + n_LF + n_LTP  */
+                tmp1 = silk_ADD32( n_AR_Q14, n_LF_Q14 );                                    /* Q14 */
+                tmp2 = silk_ADD32( n_LTP_Q14, LPC_pred_Q14 );                               /* Q13 */
+                tmp1 = silk_SUB32( tmp2, tmp1 );                                            /* Q13 */
+                tmp1 = silk_RSHIFT_ROUND( tmp1, 4 );                                        /* Q10 */
+
+                r_Q10 = silk_SUB32( x_Q10[ i ], tmp1 );                                     /* residual error Q10 */
+
+                /* Flip sign depending on dither */
+                if ( psDD->Seed < 0 ) {
+                    r_Q10 = -r_Q10;
+                }
+                r_Q10 = silk_LIMIT_32( r_Q10, -(31 << 10), 30 << 10 );
+
+                /* Find two quantization level candidates and measure their rate-distortion */
+                q1_Q10 = silk_SUB32( r_Q10, offset_Q10 );
+                q1_Q0 = silk_RSHIFT( q1_Q10, 10 );
+                if( q1_Q0 > 0 ) {
+                    q1_Q10  = silk_SUB32( silk_LSHIFT( q1_Q0, 10 ), QUANT_LEVEL_ADJUST_Q10 );
+                    q1_Q10  = silk_ADD32( q1_Q10, offset_Q10 );
+                    q2_Q10  = silk_ADD32( q1_Q10, 1024 );
+                    rd1_Q10 = silk_SMULBB( q1_Q10, Lambda_Q10 );
+                    rd2_Q10 = silk_SMULBB( q2_Q10, Lambda_Q10 );
+                } else if( q1_Q0 == 0 ) {
+                    q1_Q10  = offset_Q10;
+                    q2_Q10  = silk_ADD32( q1_Q10, 1024 - QUANT_LEVEL_ADJUST_Q10 );
+                    rd1_Q10 = silk_SMULBB( q1_Q10, Lambda_Q10 );
+                    rd2_Q10 = silk_SMULBB( q2_Q10, Lambda_Q10 );
+                } else if( q1_Q0 == -1 ) {
+                    q2_Q10  = offset_Q10;
+                    q1_Q10  = silk_SUB32( q2_Q10, 1024 - QUANT_LEVEL_ADJUST_Q10 );
+                    rd1_Q10 = silk_SMULBB( -q1_Q10, Lambda_Q10 );
+                    rd2_Q10 = silk_SMULBB(  q2_Q10, Lambda_Q10 );
+                } else {            /* q1_Q0 < -1 */
+                    q1_Q10  = silk_ADD32( silk_LSHIFT( q1_Q0, 10 ), QUANT_LEVEL_ADJUST_Q10 );
+                    q1_Q10  = silk_ADD32( q1_Q10, offset_Q10 );
+                    q2_Q10  = silk_ADD32( q1_Q10, 1024 );
+                    rd1_Q10 = silk_SMULBB( -q1_Q10, Lambda_Q10 );
+                    rd2_Q10 = silk_SMULBB( -q2_Q10, Lambda_Q10 );
+                }
+                rr_Q10  = silk_SUB32( r_Q10, q1_Q10 );
+                rd1_Q10 = silk_RSHIFT( silk_SMLABB( rd1_Q10, rr_Q10, rr_Q10 ), 10 );
+                rr_Q10  = silk_SUB32( r_Q10, q2_Q10 );
+                rd2_Q10 = silk_RSHIFT( silk_SMLABB( rd2_Q10, rr_Q10, rr_Q10 ), 10 );
+
+                if( rd1_Q10 < rd2_Q10 ) {
+                    psSS[ 0 ].RD_Q10 = silk_ADD32( psDD->RD_Q10, rd1_Q10 );
+                    psSS[ 1 ].RD_Q10 = silk_ADD32( psDD->RD_Q10, rd2_Q10 );
+                    psSS[ 0 ].Q_Q10  = q1_Q10;
+                    psSS[ 1 ].Q_Q10  = q2_Q10;
+                } else {
+                    psSS[ 0 ].RD_Q10 = silk_ADD32( psDD->RD_Q10, rd2_Q10 );
+                    psSS[ 1 ].RD_Q10 = silk_ADD32( psDD->RD_Q10, rd1_Q10 );
+                    psSS[ 0 ].Q_Q10  = q2_Q10;
+                    psSS[ 1 ].Q_Q10  = q1_Q10;
+                }
+
+                /* Update states for best quantization */
+
+                /* Quantized excitation */
+                exc_Q14 = silk_LSHIFT32( psSS[ 0 ].Q_Q10, 4 );
+                if ( psDD->Seed < 0 ) {
+                    exc_Q14 = -exc_Q14;
+                }
+
+                /* Add predictions */
+                LPC_exc_Q14 = silk_ADD32( exc_Q14, LTP_pred_Q14 );
+                xq_Q14      = silk_ADD32( LPC_exc_Q14, LPC_pred_Q14 );
+
+                /* Update states */
+                sLF_AR_shp_Q14         = silk_SUB32( xq_Q14, n_AR_Q14 );
+                psSS[ 0 ].sLTP_shp_Q14 = silk_SUB32( sLF_AR_shp_Q14, n_LF_Q14 );
+                psSS[ 0 ].LF_AR_Q14    = sLF_AR_shp_Q14;
+                psSS[ 0 ].LPC_exc_Q14  = LPC_exc_Q14;
+                psSS[ 0 ].xq_Q14       = xq_Q14;
+
+                /* Update states for second best quantization */
+
+                /* Quantized excitation */
+                exc_Q14 = silk_LSHIFT32( psSS[ 1 ].Q_Q10, 4 );
+                if ( psDD->Seed < 0 ) {
+                    exc_Q14 = -exc_Q14;
+                }
+
+
+                /* Add predictions */
+                LPC_exc_Q14 = silk_ADD32( exc_Q14, LTP_pred_Q14 );
+                xq_Q14      = silk_ADD32( LPC_exc_Q14, LPC_pred_Q14 );
+
+                /* Update states */
+                sLF_AR_shp_Q14         = silk_SUB32( xq_Q14, n_AR_Q14 );
+                psSS[ 1 ].sLTP_shp_Q14 = silk_SUB32( sLF_AR_shp_Q14, n_LF_Q14 );
+                psSS[ 1 ].LF_AR_Q14    = sLF_AR_shp_Q14;
+                psSS[ 1 ].LPC_exc_Q14  = LPC_exc_Q14;
+                psSS[ 1 ].xq_Q14       = xq_Q14;
+            }
+        }
+        *smpl_buf_idx  = ( *smpl_buf_idx - 1 ) & DECISION_DELAY_MASK;                   /* Index to newest samples              */
+        last_smple_idx = ( *smpl_buf_idx + decisionDelay ) & DECISION_DELAY_MASK;       /* Index to decisionDelay old samples   */
+
+        /* Find winner */
+        RDmin_Q10 = psSampleState[ 0 ][ 0 ].RD_Q10;
+        Winner_ind = 0;
+        for( k = 1; k < nStatesDelayedDecision; k++ ) {
+            if( psSampleState[ k ][ 0 ].RD_Q10 < RDmin_Q10 ) {
+                RDmin_Q10  = psSampleState[ k ][ 0 ].RD_Q10;
+                Winner_ind = k;
+            }
+        }
+
+        /* Increase RD values of expired states */
+        Winner_rand_state = psDelDec[ Winner_ind ].RandState[ last_smple_idx ];
+        for( k = 0; k < nStatesDelayedDecision; k++ ) {
+            if( psDelDec[ k ].RandState[ last_smple_idx ] != Winner_rand_state ) {
+                psSampleState[ k ][ 0 ].RD_Q10 = silk_ADD32( psSampleState[ k ][ 0 ].RD_Q10, silk_int32_MAX >> 4 );
+                psSampleState[ k ][ 1 ].RD_Q10 = silk_ADD32( psSampleState[ k ][ 1 ].RD_Q10, silk_int32_MAX >> 4 );
+                silk_assert( psSampleState[ k ][ 0 ].RD_Q10 >= 0 );
+            }
+        }
+
+        /* Find worst in first set and best in second set */
+        RDmax_Q10  = psSampleState[ 0 ][ 0 ].RD_Q10;
+        RDmin_Q10  = psSampleState[ 0 ][ 1 ].RD_Q10;
+        RDmax_ind = 0;
+        RDmin_ind = 0;
+        for( k = 1; k < nStatesDelayedDecision; k++ ) {
+            /* find worst in first set */
+            if( psSampleState[ k ][ 0 ].RD_Q10 > RDmax_Q10 ) {
+                RDmax_Q10  = psSampleState[ k ][ 0 ].RD_Q10;
+                RDmax_ind = k;
+            }
+            /* find best in second set */
+            if( psSampleState[ k ][ 1 ].RD_Q10 < RDmin_Q10 ) {
+                RDmin_Q10  = psSampleState[ k ][ 1 ].RD_Q10;
+                RDmin_ind = k;
+            }
+        }
+
+        /* Replace a state if best from second set outperforms worst in first set */
+        if( RDmin_Q10 < RDmax_Q10 ) {
+            silk_memcpy( ( (opus_int32 *)&psDelDec[ RDmax_ind ] ) + i,
+                         ( (opus_int32 *)&psDelDec[ RDmin_ind ] ) + i, sizeof( NSQ_del_dec_struct ) - i * sizeof( opus_int32) );
+            silk_memcpy( &psSampleState[ RDmax_ind ][ 0 ], &psSampleState[ RDmin_ind ][ 1 ], sizeof( NSQ_sample_struct ) );
+        }
+
+        /* Write samples from winner to output and long-term filter states */
+        psDD = &psDelDec[ Winner_ind ];
+        if( subfr > 0 || i >= decisionDelay ) {
+            pulses[  i - decisionDelay ] = (opus_int8)silk_RSHIFT_ROUND( psDD->Q_Q10[ last_smple_idx ], 10 );
+            xq[ i - decisionDelay ] = (opus_int16)silk_SAT16( silk_RSHIFT_ROUND(
+                silk_SMULWW( psDD->Xq_Q14[ last_smple_idx ], delayedGain_Q10[ last_smple_idx ] ), 8 ) );
+            NSQ->sLTP_shp_Q14[ NSQ->sLTP_shp_buf_idx - decisionDelay ] = psDD->Shape_Q14[ last_smple_idx ];
+            sLTP_Q15[          NSQ->sLTP_buf_idx     - decisionDelay ] = psDD->Pred_Q15[  last_smple_idx ];
+        }
+        NSQ->sLTP_shp_buf_idx++;
+        NSQ->sLTP_buf_idx++;
+
+        /* Update states */
+        for( k = 0; k < nStatesDelayedDecision; k++ ) {
+            psDD                                     = &psDelDec[ k ];
+            psSS                                     = &psSampleState[ k ][ 0 ];
+            psDD->LF_AR_Q14                          = psSS->LF_AR_Q14;
+            psDD->sLPC_Q14[ NSQ_LPC_BUF_LENGTH + i ] = psSS->xq_Q14;
+            psDD->Xq_Q14[    *smpl_buf_idx ]         = psSS->xq_Q14;
+            psDD->Q_Q10[     *smpl_buf_idx ]         = psSS->Q_Q10;
+            psDD->Pred_Q15[  *smpl_buf_idx ]         = silk_LSHIFT32( psSS->LPC_exc_Q14, 1 );
+            psDD->Shape_Q14[ *smpl_buf_idx ]         = psSS->sLTP_shp_Q14;
+            psDD->Seed                               = silk_ADD32_ovflw( psDD->Seed, silk_RSHIFT_ROUND( psSS->Q_Q10, 10 ) );
+            psDD->RandState[ *smpl_buf_idx ]         = psDD->Seed;
+            psDD->RD_Q10                             = psSS->RD_Q10;
+        }
+        delayedGain_Q10[     *smpl_buf_idx ]         = Gain_Q10;
+    }
+    /* Update LPC states */
+    for( k = 0; k < nStatesDelayedDecision; k++ ) {
+        psDD = &psDelDec[ k ];
+        silk_memcpy( psDD->sLPC_Q14, &psDD->sLPC_Q14[ length ], NSQ_LPC_BUF_LENGTH * sizeof( opus_int32 ) );
+    }
+    RESTORE_STACK;
+}
+
+static OPUS_INLINE void silk_nsq_del_dec_scale_states_sse4_1(
+    const silk_encoder_state *psEncC,               /* I    Encoder State                       */
+    silk_nsq_state      *NSQ,                       /* I/O  NSQ state                           */
+    NSQ_del_dec_struct  psDelDec[],                 /* I/O  Delayed decision states             */
+    const opus_int32    x_Q3[],                     /* I    Input in Q3                         */
+    opus_int32          x_sc_Q10[],                 /* O    Input scaled with 1/Gain in Q10     */
+    const opus_int16    sLTP[],                     /* I    Re-whitened LTP state in Q0         */
+    opus_int32          sLTP_Q15[],                 /* O    LTP state matching scaled input     */
+    opus_int            subfr,                      /* I    Subframe number                     */
+    opus_int            nStatesDelayedDecision,     /* I    Number of del dec states            */
+    const opus_int      LTP_scale_Q14,              /* I    LTP state scaling                   */
+    const opus_int32    Gains_Q16[ MAX_NB_SUBFR ],  /* I                                        */
+    const opus_int      pitchL[ MAX_NB_SUBFR ],     /* I    Pitch lag                           */
+    const opus_int      signal_type,                /* I    Signal type                         */
+    const opus_int      decisionDelay               /* I    Decision delay                      */
+)
+{
+    opus_int            i, k, lag;
+    opus_int32          gain_adj_Q16, inv_gain_Q31, inv_gain_Q23;
+    NSQ_del_dec_struct  *psDD;
+    __m128i xmm_inv_gain_Q23, xmm_x_Q3_x2x0, xmm_x_Q3_x3x1;
+
+    lag          = pitchL[ subfr ];
+    inv_gain_Q31 = silk_INVERSE32_varQ( silk_max( Gains_Q16[ subfr ], 1 ), 47 );
+
+    silk_assert( inv_gain_Q31 != 0 );
+
+    /* Calculate gain adjustment factor */
+    if( Gains_Q16[ subfr ] != NSQ->prev_gain_Q16 ) {
+        gain_adj_Q16 =  silk_DIV32_varQ( NSQ->prev_gain_Q16, Gains_Q16[ subfr ], 16 );
+    } else {
+        gain_adj_Q16 = (opus_int32)1 << 16;
+    }
+
+    /* Scale input */
+    inv_gain_Q23 = silk_RSHIFT_ROUND( inv_gain_Q31, 8 );
+
+    /* prepare inv_gain_Q23 in packed 4 32-bits */
+    xmm_inv_gain_Q23 = _mm_set1_epi32(inv_gain_Q23);
+
+    for( i = 0; i < psEncC->subfr_length - 3; i += 4 ) {
+        xmm_x_Q3_x2x0 = _mm_loadu_si128( (__m128i *)(&(x_Q3[ i ] ) ) );
+        /* equal shift right 4 bytes*/
+        xmm_x_Q3_x3x1 = _mm_shuffle_epi32( xmm_x_Q3_x2x0, _MM_SHUFFLE( 0, 3, 2, 1 ) );
+
+        xmm_x_Q3_x2x0 = _mm_mul_epi32( xmm_x_Q3_x2x0, xmm_inv_gain_Q23 );
+        xmm_x_Q3_x3x1 = _mm_mul_epi32( xmm_x_Q3_x3x1, xmm_inv_gain_Q23 );
+
+        xmm_x_Q3_x2x0 = _mm_srli_epi64( xmm_x_Q3_x2x0, 16 );
+        xmm_x_Q3_x3x1 = _mm_slli_epi64( xmm_x_Q3_x3x1, 16 );
+
+        xmm_x_Q3_x2x0 = _mm_blend_epi16( xmm_x_Q3_x2x0, xmm_x_Q3_x3x1, 0xCC );
+
+        _mm_storeu_si128( (__m128i *)(&(x_sc_Q10[ i ])), xmm_x_Q3_x2x0 );
+    }
+
+    for( ; i < psEncC->subfr_length; i++ ) {
+        x_sc_Q10[ i ] = silk_SMULWW( x_Q3[ i ], inv_gain_Q23 );
+    }
+
+    /* Save inverse gain */
+    NSQ->prev_gain_Q16 = Gains_Q16[ subfr ];
+
+    /* After rewhitening the LTP state is un-scaled, so scale with inv_gain_Q16 */
+    if( NSQ->rewhite_flag ) {
+        if( subfr == 0 ) {
+            /* Do LTP downscaling */
+            inv_gain_Q31 = silk_LSHIFT( silk_SMULWB( inv_gain_Q31, LTP_scale_Q14 ), 2 );
+        }
+        for( i = NSQ->sLTP_buf_idx - lag - LTP_ORDER / 2; i < NSQ->sLTP_buf_idx; i++ ) {
+            silk_assert( i < MAX_FRAME_LENGTH );
+            sLTP_Q15[ i ] = silk_SMULWB( inv_gain_Q31, sLTP[ i ] );
+        }
+    }
+
+    /* Adjust for changing gain */
+    if( gain_adj_Q16 != (opus_int32)1 << 16 ) {
+        /* Scale long-term shaping state */
+        {
+            __m128i xmm_gain_adj_Q16, xmm_sLTP_shp_Q14_x2x0, xmm_sLTP_shp_Q14_x3x1;
+
+            /* prepare gain_adj_Q16 in packed 4 32-bits */
+            xmm_gain_adj_Q16 = _mm_set1_epi32( gain_adj_Q16 );
+
+            for( i = NSQ->sLTP_shp_buf_idx - psEncC->ltp_mem_length; i < NSQ->sLTP_shp_buf_idx - 3; i += 4 )
+            {
+                xmm_sLTP_shp_Q14_x2x0 = _mm_loadu_si128( (__m128i *)(&(NSQ->sLTP_shp_Q14[ i ] ) ) );
+                /* equal shift right 4 bytes*/
+                xmm_sLTP_shp_Q14_x3x1 = _mm_shuffle_epi32( xmm_sLTP_shp_Q14_x2x0, _MM_SHUFFLE( 0, 3, 2, 1 ) );
+
+                xmm_sLTP_shp_Q14_x2x0 = _mm_mul_epi32( xmm_sLTP_shp_Q14_x2x0, xmm_gain_adj_Q16 );
+                xmm_sLTP_shp_Q14_x3x1 = _mm_mul_epi32( xmm_sLTP_shp_Q14_x3x1, xmm_gain_adj_Q16 );
+
+                xmm_sLTP_shp_Q14_x2x0 = _mm_srli_epi64( xmm_sLTP_shp_Q14_x2x0, 16 );
+                xmm_sLTP_shp_Q14_x3x1 = _mm_slli_epi64( xmm_sLTP_shp_Q14_x3x1, 16 );
+
+                xmm_sLTP_shp_Q14_x2x0 = _mm_blend_epi16( xmm_sLTP_shp_Q14_x2x0, xmm_sLTP_shp_Q14_x3x1, 0xCC );
+
+                _mm_storeu_si128( (__m128i *)(&(NSQ->sLTP_shp_Q14[ i ] ) ), xmm_sLTP_shp_Q14_x2x0 );
+            }
+
+            for( ; i < NSQ->sLTP_shp_buf_idx; i++ ) {
+                NSQ->sLTP_shp_Q14[ i ] = silk_SMULWW( gain_adj_Q16, NSQ->sLTP_shp_Q14[ i ] );
+            }
+
+            /* Scale long-term prediction state */
+            if( signal_type == TYPE_VOICED && NSQ->rewhite_flag == 0 ) {
+                for( i = NSQ->sLTP_buf_idx - lag - LTP_ORDER / 2; i < NSQ->sLTP_buf_idx - decisionDelay; i++ ) {
+                    sLTP_Q15[ i ] = silk_SMULWW( gain_adj_Q16, sLTP_Q15[ i ] );
+                }
+            }
+
+            for( k = 0; k < nStatesDelayedDecision; k++ ) {
+                psDD = &psDelDec[ k ];
+
+                /* Scale scalar states */
+                psDD->LF_AR_Q14 = silk_SMULWW( gain_adj_Q16, psDD->LF_AR_Q14 );
+
+                /* Scale short-term prediction and shaping states */
+                for( i = 0; i < NSQ_LPC_BUF_LENGTH; i++ ) {
+                    psDD->sLPC_Q14[ i ] = silk_SMULWW( gain_adj_Q16, psDD->sLPC_Q14[ i ] );
+                }
+                for( i = 0; i < MAX_SHAPE_LPC_ORDER; i++ ) {
+                    psDD->sAR2_Q14[ i ] = silk_SMULWW( gain_adj_Q16, psDD->sAR2_Q14[ i ] );
+                }
+                for( i = 0; i < DECISION_DELAY; i++ ) {
+                    psDD->Pred_Q15[  i ] = silk_SMULWW( gain_adj_Q16, psDD->Pred_Q15[  i ] );
+                    psDD->Shape_Q14[ i ] = silk_SMULWW( gain_adj_Q16, psDD->Shape_Q14[ i ] );
+                }
+            }
+        }
+    }
+}
diff --git a/media/libopus/silk/x86/NSQ_sse.c b/media/libopus/silk/x86/NSQ_sse.c
new file mode 100644
index 000000000..bb3c5f195
--- /dev/null
+++ b/media/libopus/silk/x86/NSQ_sse.c
@@ -0,0 +1,720 @@
+/* Copyright (c) 2014, Cisco Systems, INC
+   Written by XiangMingZhu WeiZhou MinPeng YanWang
+
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions
+   are met:
+
+   - Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+
+   - Redistributions in binary form must reproduce the above copyright
+   notice, this list of conditions and the following disclaimer in the
+   documentation and/or other materials provided with the distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
+   OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+   EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+   PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+   PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+   LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+   NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+   SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+
+#ifdef HAVE_CONFIG_H
+#include "config.h"
+#endif
+
+#include <xmmintrin.h>
+#include <emmintrin.h>
+#include <smmintrin.h>
+#include "main.h"
+#include "celt/x86/x86cpu.h"
+#include "stack_alloc.h"
+
+static OPUS_INLINE void silk_nsq_scale_states_sse4_1(
+    const silk_encoder_state *psEncC,           /* I    Encoder State                   */
+    silk_nsq_state      *NSQ,                   /* I/O  NSQ state                       */
+    const opus_int32    x_Q3[],                 /* I    input in Q3                     */
+    opus_int32          x_sc_Q10[],             /* O    input scaled with 1/Gain        */
+    const opus_int16    sLTP[],                 /* I    re-whitened LTP state in Q0     */
+    opus_int32          sLTP_Q15[],             /* O    LTP state matching scaled input */
+    opus_int            subfr,                  /* I    subframe number                 */
+    const opus_int      LTP_scale_Q14,          /* I                                    */
+    const opus_int32    Gains_Q16[ MAX_NB_SUBFR ], /* I                                 */
+    const opus_int      pitchL[ MAX_NB_SUBFR ], /* I    Pitch lag                       */
+    const opus_int      signal_type             /* I    Signal type                     */
+);
+
+static OPUS_INLINE void silk_noise_shape_quantizer_10_16_sse4_1(
+    silk_nsq_state      *NSQ,                   /* I/O  NSQ state                       */
+    opus_int            signalType,             /* I    Signal type                     */
+    const opus_int32    x_sc_Q10[],             /* I                                    */
+    opus_int8           pulses[],               /* O                                    */
+    opus_int16          xq[],                   /* O                                    */
+    opus_int32          sLTP_Q15[],             /* I/O  LTP state                       */
+    const opus_int16    a_Q12[],                /* I    Short term prediction coefs     */
+    const opus_int16    b_Q14[],                /* I    Long term prediction coefs      */
+    const opus_int16    AR_shp_Q13[],           /* I    Noise shaping AR coefs          */
+    opus_int            lag,                    /* I    Pitch lag                       */
+    opus_int32          HarmShapeFIRPacked_Q14, /* I                                    */
+    opus_int            Tilt_Q14,               /* I    Spectral tilt                   */
+    opus_int32          LF_shp_Q14,             /* I                                    */
+    opus_int32          Gain_Q16,               /* I                                    */
+    opus_int            offset_Q10,             /* I                                    */
+    opus_int            length,                 /* I    Input length                    */
+    opus_int32          table[][4]              /* I                                    */
+);
+
+void silk_NSQ_sse4_1(
+    const silk_encoder_state    *psEncC,                                    /* I/O  Encoder State                   */
+    silk_nsq_state              *NSQ,                                       /* I/O  NSQ state                       */
+    SideInfoIndices             *psIndices,                                 /* I/O  Quantization Indices            */
+    const opus_int32            x_Q3[],                                     /* I    Prefiltered input signal        */
+    opus_int8                   pulses[],                                   /* O    Quantized pulse signal          */
+    const opus_int16            PredCoef_Q12[ 2 * MAX_LPC_ORDER ],          /* I    Short term prediction coefs     */
+    const opus_int16            LTPCoef_Q14[ LTP_ORDER * MAX_NB_SUBFR ],    /* I    Long term prediction coefs      */
+    const opus_int16            AR2_Q13[ MAX_NB_SUBFR * MAX_SHAPE_LPC_ORDER ], /* I Noise shaping coefs             */
+    const opus_int              HarmShapeGain_Q14[ MAX_NB_SUBFR ],          /* I    Long term shaping coefs         */
+    const opus_int              Tilt_Q14[ MAX_NB_SUBFR ],                   /* I    Spectral tilt                   */
+    const opus_int32            LF_shp_Q14[ MAX_NB_SUBFR ],                 /* I    Low frequency shaping coefs     */
+    const opus_int32            Gains_Q16[ MAX_NB_SUBFR ],                  /* I    Quantization step sizes         */
+    const opus_int              pitchL[ MAX_NB_SUBFR ],                     /* I    Pitch lags                      */
+    const opus_int              Lambda_Q10,                                 /* I    Rate/distortion tradeoff        */
+    const opus_int              LTP_scale_Q14                               /* I    LTP state scaling               */
+)
+{
+    opus_int            k, lag, start_idx, LSF_interpolation_flag;
+    const opus_int16    *A_Q12, *B_Q14, *AR_shp_Q13;
+    opus_int16          *pxq;
+    VARDECL( opus_int32, sLTP_Q15 );
+    VARDECL( opus_int16, sLTP );
+    opus_int32          HarmShapeFIRPacked_Q14;
+    opus_int            offset_Q10;
+    VARDECL( opus_int32, x_sc_Q10 );
+
+    opus_int32   table[ 64 ][ 4 ];
+    opus_int32   tmp1;
+    opus_int32   q1_Q10, q2_Q10, rd1_Q20, rd2_Q20;
+
+    SAVE_STACK;
+
+    NSQ->rand_seed = psIndices->Seed;
+
+    /* Set unvoiced lag to the previous one, overwrite later for voiced */
+    lag = NSQ->lagPrev;
+
+    silk_assert( NSQ->prev_gain_Q16 != 0 );
+
+    offset_Q10 = silk_Quantization_Offsets_Q10[ psIndices->signalType >> 1 ][ psIndices->quantOffsetType ];
+
+    /* 0 */
+    q1_Q10  = offset_Q10;
+    q2_Q10  = offset_Q10 + ( 1024 - QUANT_LEVEL_ADJUST_Q10 );
+    rd1_Q20 = q1_Q10 * Lambda_Q10;
+    rd2_Q20 = q2_Q10 * Lambda_Q10;
+
+    table[ 32 ][ 0 ] = q1_Q10;
+    table[ 32 ][ 1 ] = q2_Q10;
+    table[ 32 ][ 2 ] = 2 * (q1_Q10 - q2_Q10);
+    table[ 32 ][ 3 ] = (rd1_Q20 - rd2_Q20) + (q1_Q10 * q1_Q10 - q2_Q10 * q2_Q10);
+
+    /* -1 */
+    q1_Q10  = offset_Q10 - ( 1024 - QUANT_LEVEL_ADJUST_Q10 );
+    q2_Q10  = offset_Q10;
+    rd1_Q20 = - q1_Q10 * Lambda_Q10;
+    rd2_Q20 = q2_Q10 * Lambda_Q10;
+
+    table[ 31 ][ 0 ] = q1_Q10;
+    table[ 31 ][ 1 ] = q2_Q10;
+    table[ 31 ][ 2 ] = 2 * (q1_Q10 - q2_Q10);
+    table[ 31 ][ 3 ] = (rd1_Q20 - rd2_Q20) + (q1_Q10 * q1_Q10 - q2_Q10 * q2_Q10);
+
+    /* > 0 */
+    for (k = 1; k <= 31; k++)
+    {
+        tmp1 = offset_Q10 + silk_LSHIFT( k, 10 );
+
+        q1_Q10  = tmp1 - QUANT_LEVEL_ADJUST_Q10;
+        q2_Q10  = tmp1 - QUANT_LEVEL_ADJUST_Q10 + 1024;
+        rd1_Q20 = q1_Q10 * Lambda_Q10;
+        rd2_Q20 = q2_Q10 * Lambda_Q10;
+
+        table[ 32 + k ][ 0 ] = q1_Q10;
+        table[ 32 + k ][ 1 ] = q2_Q10;
+        table[ 32 + k ][ 2 ] = 2 * (q1_Q10 - q2_Q10);
+        table[ 32 + k ][ 3 ] = (rd1_Q20 - rd2_Q20) + (q1_Q10 * q1_Q10 - q2_Q10 * q2_Q10);
+    }
+
+    /* < -1 */
+    for (k = -32; k <= -2; k++)
+    {
+        tmp1 = offset_Q10 + silk_LSHIFT( k, 10 );
+
+        q1_Q10  = tmp1 + QUANT_LEVEL_ADJUST_Q10;
+        q2_Q10  = tmp1 + QUANT_LEVEL_ADJUST_Q10 + 1024;
+        rd1_Q20 = - q1_Q10 * Lambda_Q10;
+        rd2_Q20 = - q2_Q10 * Lambda_Q10;
+
+        table[ 32 + k ][ 0 ] = q1_Q10;
+        table[ 32 + k ][ 1 ] = q2_Q10;
+        table[ 32 + k ][ 2 ] = 2 * (q1_Q10 - q2_Q10);
+        table[ 32 + k ][ 3 ] = (rd1_Q20 - rd2_Q20) + (q1_Q10 * q1_Q10 - q2_Q10 * q2_Q10);
+    }
+
+    if( psIndices->NLSFInterpCoef_Q2 == 4 ) {
+        LSF_interpolation_flag = 0;
+    } else {
+        LSF_interpolation_flag = 1;
+    }
+
+    ALLOC( sLTP_Q15,
+           psEncC->ltp_mem_length + psEncC->frame_length, opus_int32 );
+    ALLOC( sLTP, psEncC->ltp_mem_length + psEncC->frame_length, opus_int16 );
+    ALLOC( x_sc_Q10, psEncC->subfr_length, opus_int32 );
+    /* Set up pointers to start of sub frame */
+    NSQ->sLTP_shp_buf_idx = psEncC->ltp_mem_length;
+    NSQ->sLTP_buf_idx     = psEncC->ltp_mem_length;
+    pxq                   = &NSQ->xq[ psEncC->ltp_mem_length ];
+    for( k = 0; k < psEncC->nb_subfr; k++ ) {
+        A_Q12      = &PredCoef_Q12[ (( k >> 1 ) | ( 1 - LSF_interpolation_flag )) * MAX_LPC_ORDER ];
+        B_Q14      = &LTPCoef_Q14[ k * LTP_ORDER ];
+        AR_shp_Q13 = &AR2_Q13[     k * MAX_SHAPE_LPC_ORDER ];
+
+        /* Noise shape parameters */
+        silk_assert( HarmShapeGain_Q14[ k ] >= 0 );
+        HarmShapeFIRPacked_Q14  =                          silk_RSHIFT( HarmShapeGain_Q14[ k ], 2 );
+        HarmShapeFIRPacked_Q14 |= silk_LSHIFT( (opus_int32)silk_RSHIFT( HarmShapeGain_Q14[ k ], 1 ), 16 );
+
+        NSQ->rewhite_flag = 0;
+        if( psIndices->signalType == TYPE_VOICED ) {
+            /* Voiced */
+            lag = pitchL[ k ];
+
+            /* Re-whitening */
+            if( ( k & ( 3 - silk_LSHIFT( LSF_interpolation_flag, 1 ) ) ) == 0 ) {
+                /* Rewhiten with new A coefs */
+                start_idx = psEncC->ltp_mem_length - lag - psEncC->predictLPCOrder - LTP_ORDER / 2;
+                silk_assert( start_idx > 0 );
+
+                silk_LPC_analysis_filter( &sLTP[ start_idx ], &NSQ->xq[ start_idx + k * psEncC->subfr_length ],
+                    A_Q12, psEncC->ltp_mem_length - start_idx, psEncC->predictLPCOrder, psEncC->arch );
+
+                NSQ->rewhite_flag = 1;
+                NSQ->sLTP_buf_idx = psEncC->ltp_mem_length;
+            }
+        }
+
+        silk_nsq_scale_states_sse4_1( psEncC, NSQ, x_Q3, x_sc_Q10, sLTP, sLTP_Q15, k, LTP_scale_Q14, Gains_Q16, pitchL, psIndices->signalType );
+
+        if ( opus_likely( ( 10 == psEncC->shapingLPCOrder ) && ( 16 == psEncC->predictLPCOrder) ) )
+        {
+            silk_noise_shape_quantizer_10_16_sse4_1( NSQ, psIndices->signalType, x_sc_Q10, pulses, pxq, sLTP_Q15, A_Q12, B_Q14,
+                AR_shp_Q13, lag, HarmShapeFIRPacked_Q14, Tilt_Q14[ k ], LF_shp_Q14[ k ], Gains_Q16[ k ],
+                offset_Q10, psEncC->subfr_length, &(table[32]) );
+        }
+        else
+        {
+            silk_noise_shape_quantizer( NSQ, psIndices->signalType, x_sc_Q10, pulses, pxq, sLTP_Q15, A_Q12, B_Q14,
+                AR_shp_Q13, lag, HarmShapeFIRPacked_Q14, Tilt_Q14[ k ], LF_shp_Q14[ k ], Gains_Q16[ k ], Lambda_Q10,
+                offset_Q10, psEncC->subfr_length, psEncC->shapingLPCOrder, psEncC->predictLPCOrder, psEncC->arch );
+        }
+
+        x_Q3   += psEncC->subfr_length;
+        pulses += psEncC->subfr_length;
+        pxq    += psEncC->subfr_length;
+    }
+
+    /* Update lagPrev for next frame */
+    NSQ->lagPrev = pitchL[ psEncC->nb_subfr - 1 ];
+
+    /* Save quantized speech and noise shaping signals */
+    /* DEBUG_STORE_DATA( enc.pcm, &NSQ->xq[ psEncC->ltp_mem_length ], psEncC->frame_length * sizeof( opus_int16 ) ) */
+    silk_memmove( NSQ->xq,           &NSQ->xq[           psEncC->frame_length ], psEncC->ltp_mem_length * sizeof( opus_int16 ) );
+    silk_memmove( NSQ->sLTP_shp_Q14, &NSQ->sLTP_shp_Q14[ psEncC->frame_length ], psEncC->ltp_mem_length * sizeof( opus_int32 ) );
+    RESTORE_STACK;
+}
+
+/***********************************/
+/* silk_noise_shape_quantizer_10_16  */
+/***********************************/
+static OPUS_INLINE void silk_noise_shape_quantizer_10_16_sse4_1(
+    silk_nsq_state      *NSQ,                   /* I/O  NSQ state                       */
+    opus_int            signalType,             /* I    Signal type                     */
+    const opus_int32    x_sc_Q10[],             /* I                                    */
+    opus_int8           pulses[],               /* O                                    */
+    opus_int16          xq[],                   /* O                                    */
+    opus_int32          sLTP_Q15[],             /* I/O  LTP state                       */
+    const opus_int16    a_Q12[],                /* I    Short term prediction coefs     */
+    const opus_int16    b_Q14[],                /* I    Long term prediction coefs      */
+    const opus_int16    AR_shp_Q13[],           /* I    Noise shaping AR coefs          */
+    opus_int            lag,                    /* I    Pitch lag                       */
+    opus_int32          HarmShapeFIRPacked_Q14, /* I                                    */
+    opus_int            Tilt_Q14,               /* I    Spectral tilt                   */
+    opus_int32          LF_shp_Q14,             /* I                                    */
+    opus_int32          Gain_Q16,               /* I                                    */
+    opus_int            offset_Q10,             /* I                                    */
+    opus_int            length,                 /* I    Input length                    */
+    opus_int32          table[][4]              /* I                                    */
+)
+{
+    opus_int     i;
+    opus_int32   LTP_pred_Q13, LPC_pred_Q10, n_AR_Q12, n_LTP_Q13;
+    opus_int32   n_LF_Q12, r_Q10, q1_Q0, q1_Q10, q2_Q10;
+    opus_int32   exc_Q14, LPC_exc_Q14, xq_Q14, Gain_Q10;
+    opus_int32   tmp1, tmp2, sLF_AR_shp_Q14;
+    opus_int32   *psLPC_Q14, *shp_lag_ptr, *pred_lag_ptr;
+
+    __m128i xmm_tempa, xmm_tempb;
+
+    __m128i xmm_one;
+
+    __m128i psLPC_Q14_hi_01234567, psLPC_Q14_hi_89ABCDEF;
+    __m128i psLPC_Q14_lo_01234567, psLPC_Q14_lo_89ABCDEF;
+    __m128i a_Q12_01234567,        a_Q12_89ABCDEF;
+
+    __m128i sAR2_Q14_hi_76543210, sAR2_Q14_lo_76543210;
+    __m128i AR_shp_Q13_76543210;
+
+    shp_lag_ptr  = &NSQ->sLTP_shp_Q14[ NSQ->sLTP_shp_buf_idx - lag + HARM_SHAPE_FIR_TAPS / 2 ];
+    pred_lag_ptr = &sLTP_Q15[ NSQ->sLTP_buf_idx - lag + LTP_ORDER / 2 ];
+    Gain_Q10     = silk_RSHIFT( Gain_Q16, 6 );
+
+    /* Set up short term AR state */
+    psLPC_Q14 = &NSQ->sLPC_Q14[ NSQ_LPC_BUF_LENGTH - 1 ];
+
+    sLF_AR_shp_Q14 = NSQ->sLF_AR_shp_Q14;
+    xq_Q14         = psLPC_Q14[ 0 ];
+    LTP_pred_Q13   = 0;
+
+    /* load a_Q12 */
+    xmm_one = _mm_set_epi8( 1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14 );
+
+    /* load a_Q12[0] - a_Q12[7] */
+    a_Q12_01234567 = _mm_loadu_si128( (__m128i *)(&a_Q12[ 0 ] ) );
+    /* load a_Q12[ 8 ] - a_Q12[ 15 ] */
+    a_Q12_89ABCDEF = _mm_loadu_si128( (__m128i *)(&a_Q12[ 8 ] ) );
+
+    a_Q12_01234567 = _mm_shuffle_epi8( a_Q12_01234567, xmm_one );
+    a_Q12_89ABCDEF = _mm_shuffle_epi8( a_Q12_89ABCDEF, xmm_one );
+
+    /* load AR_shp_Q13 */
+    AR_shp_Q13_76543210 = _mm_loadu_si128( (__m128i *)(&AR_shp_Q13[0] ) );
+
+    /* load psLPC_Q14 */
+    xmm_one = _mm_set_epi8(15, 14, 11, 10, 7, 6, 3, 2, 13, 12, 9, 8, 5, 4, 1, 0 );
+
+    xmm_tempa = _mm_loadu_si128( (__m128i *)(&psLPC_Q14[-16]) );
+    xmm_tempb = _mm_loadu_si128( (__m128i *)(&psLPC_Q14[-12]) );
+
+    xmm_tempa = _mm_shuffle_epi8( xmm_tempa, xmm_one );
+    xmm_tempb = _mm_shuffle_epi8( xmm_tempb, xmm_one );
+
+    psLPC_Q14_hi_89ABCDEF = _mm_unpackhi_epi64( xmm_tempa, xmm_tempb );
+    psLPC_Q14_lo_89ABCDEF = _mm_unpacklo_epi64( xmm_tempa, xmm_tempb );
+
+    xmm_tempa = _mm_loadu_si128( (__m128i *)(&psLPC_Q14[ -8 ]) );
+    xmm_tempb = _mm_loadu_si128( (__m128i *)(&psLPC_Q14[ -4 ]) );
+
+    xmm_tempa = _mm_shuffle_epi8( xmm_tempa, xmm_one );
+    xmm_tempb = _mm_shuffle_epi8( xmm_tempb, xmm_one );
+
+    psLPC_Q14_hi_01234567 = _mm_unpackhi_epi64( xmm_tempa, xmm_tempb );
+    psLPC_Q14_lo_01234567 = _mm_unpacklo_epi64( xmm_tempa, xmm_tempb );
+
+    /* load sAR2_Q14 */
+    xmm_tempa = _mm_loadu_si128( (__m128i *)(&(NSQ->sAR2_Q14[ 0 ]) ) );
+    xmm_tempb = _mm_loadu_si128( (__m128i *)(&(NSQ->sAR2_Q14[ 4 ]) ) );
+
+    xmm_tempa = _mm_shuffle_epi8( xmm_tempa, xmm_one );
+    xmm_tempb = _mm_shuffle_epi8( xmm_tempb, xmm_one );
+
+    sAR2_Q14_hi_76543210 = _mm_unpackhi_epi64( xmm_tempa, xmm_tempb );
+    sAR2_Q14_lo_76543210 = _mm_unpacklo_epi64( xmm_tempa, xmm_tempb );
+
+    /* prepare 1 in 8 * 16bit */
+    xmm_one = _mm_set1_epi16(1);
+
+    for( i = 0; i < length; i++ )
+    {
+        /* Short-term prediction */
+        __m128i xmm_hi_07, xmm_hi_8F, xmm_lo_07, xmm_lo_8F;
+
+        /* Avoids introducing a bias because silk_SMLAWB() always rounds to -inf */
+        LPC_pred_Q10 = 8; /* silk_RSHIFT( predictLPCOrder, 1 ); */
+
+        /* shift psLPC_Q14 */
+        psLPC_Q14_hi_89ABCDEF = _mm_alignr_epi8( psLPC_Q14_hi_01234567, psLPC_Q14_hi_89ABCDEF, 2 );
+        psLPC_Q14_lo_89ABCDEF = _mm_alignr_epi8( psLPC_Q14_lo_01234567, psLPC_Q14_lo_89ABCDEF, 2 );
+
+        psLPC_Q14_hi_01234567 = _mm_srli_si128( psLPC_Q14_hi_01234567, 2 );
+        psLPC_Q14_lo_01234567 = _mm_srli_si128( psLPC_Q14_lo_01234567, 2 );
+
+        psLPC_Q14_hi_01234567 = _mm_insert_epi16( psLPC_Q14_hi_01234567, (xq_Q14 >> 16), 7 );
+        psLPC_Q14_lo_01234567 = _mm_insert_epi16( psLPC_Q14_lo_01234567, (xq_Q14),       7 );
+
+        /* high part, use pmaddwd, results in 4 32-bit */
+        xmm_hi_07 = _mm_madd_epi16( psLPC_Q14_hi_01234567, a_Q12_01234567 );
+        xmm_hi_8F = _mm_madd_epi16( psLPC_Q14_hi_89ABCDEF, a_Q12_89ABCDEF );
+
+        /* low part, use pmulhw, results in 8 16-bit, note we need simulate unsigned * signed, _mm_srai_epi16(psLPC_Q14_lo_01234567, 15) */
+        xmm_tempa = _mm_cmpgt_epi16( _mm_setzero_si128(), psLPC_Q14_lo_01234567 );
+        xmm_tempb = _mm_cmpgt_epi16( _mm_setzero_si128(), psLPC_Q14_lo_89ABCDEF );
+
+        xmm_tempa = _mm_and_si128( xmm_tempa, a_Q12_01234567 );
+        xmm_tempb = _mm_and_si128( xmm_tempb, a_Q12_89ABCDEF );
+
+        xmm_lo_07 = _mm_mulhi_epi16( psLPC_Q14_lo_01234567, a_Q12_01234567 );
+        xmm_lo_8F = _mm_mulhi_epi16( psLPC_Q14_lo_89ABCDEF, a_Q12_89ABCDEF );
+
+        xmm_lo_07 = _mm_add_epi16( xmm_lo_07, xmm_tempa );
+        xmm_lo_8F = _mm_add_epi16( xmm_lo_8F, xmm_tempb );
+
+        xmm_lo_07 = _mm_madd_epi16( xmm_lo_07, xmm_one );
+        xmm_lo_8F = _mm_madd_epi16( xmm_lo_8F, xmm_one );
+
+        /* accumulate */
+        xmm_hi_07 = _mm_add_epi32( xmm_hi_07, xmm_hi_8F );
+        xmm_lo_07 = _mm_add_epi32( xmm_lo_07, xmm_lo_8F );
+
+        xmm_hi_07 = _mm_add_epi32( xmm_hi_07, xmm_lo_07 );
+
+        xmm_hi_07 = _mm_add_epi32( xmm_hi_07, _mm_unpackhi_epi64(xmm_hi_07, xmm_hi_07 ) );
+        xmm_hi_07 = _mm_add_epi32( xmm_hi_07, _mm_shufflelo_epi16(xmm_hi_07, 0x0E ) );
+
+        LPC_pred_Q10 += _mm_cvtsi128_si32( xmm_hi_07 );
+
+        /* Long-term prediction */
+        if ( opus_likely( signalType == TYPE_VOICED ) ) {
+            /* Unrolled loop */
+            /* Avoids introducing a bias because silk_SMLAWB() always rounds to -inf */
+            LTP_pred_Q13 = 2;
+            {
+                __m128i b_Q14_3210, b_Q14_0123, pred_lag_ptr_0123;
+
+                b_Q14_3210 = OP_CVTEPI16_EPI32_M64( b_Q14 );
+                b_Q14_0123 = _mm_shuffle_epi32( b_Q14_3210, 0x1B );
+
+                /* loaded: [0] [-1] [-2] [-3] */
+                pred_lag_ptr_0123 = _mm_loadu_si128( (__m128i *)(&pred_lag_ptr[ -3 ] ) );
+                /* shuffle to [-3] [-2] [-1] [0] and to new xmm */
+                xmm_tempa = _mm_shuffle_epi32( pred_lag_ptr_0123, 0x1B );
+                /*64-bit multiply, a[2] * b[-2], a[0] * b[0] */
+                xmm_tempa = _mm_mul_epi32( xmm_tempa, b_Q14_3210 );
+                /* right shift 2 bytes (16 bits), zero extended */
+                xmm_tempa = _mm_srli_si128( xmm_tempa, 2 );
+
+                /* a[1] * b[-1], a[3] * b[-3] */
+                pred_lag_ptr_0123 = _mm_mul_epi32( pred_lag_ptr_0123, b_Q14_0123 );
+                pred_lag_ptr_0123 = _mm_srli_si128( pred_lag_ptr_0123, 2 );
+
+                pred_lag_ptr_0123 = _mm_add_epi32( pred_lag_ptr_0123, xmm_tempa );
+                /* equal shift right 8 bytes*/
+                xmm_tempa = _mm_shuffle_epi32( pred_lag_ptr_0123, _MM_SHUFFLE( 0, 0, 3, 2 ) );
+                xmm_tempa = _mm_add_epi32( xmm_tempa, pred_lag_ptr_0123 );
+
+                LTP_pred_Q13 += _mm_cvtsi128_si32( xmm_tempa );
+
+                LTP_pred_Q13 = silk_SMLAWB( LTP_pred_Q13, pred_lag_ptr[ -4 ], b_Q14[ 4 ] );
+                pred_lag_ptr++;
+            }
+        }
+
+        /* Noise shape feedback */
+        NSQ->sAR2_Q14[ 9 ] = NSQ->sAR2_Q14[ 8 ];
+        NSQ->sAR2_Q14[ 8 ] = _mm_cvtsi128_si32( _mm_srli_si128(_mm_unpackhi_epi16( sAR2_Q14_lo_76543210, sAR2_Q14_hi_76543210 ), 12 ) );
+
+        sAR2_Q14_hi_76543210 = _mm_slli_si128( sAR2_Q14_hi_76543210, 2 );
+        sAR2_Q14_lo_76543210 = _mm_slli_si128( sAR2_Q14_lo_76543210, 2 );
+
+        sAR2_Q14_hi_76543210 = _mm_insert_epi16( sAR2_Q14_hi_76543210, (xq_Q14 >> 16), 0 );
+        sAR2_Q14_lo_76543210 = _mm_insert_epi16( sAR2_Q14_lo_76543210, (xq_Q14),       0 );
+
+        /* high part, use pmaddwd, results in 4 32-bit */
+        xmm_hi_07 = _mm_madd_epi16( sAR2_Q14_hi_76543210, AR_shp_Q13_76543210 );
+
+        /* low part, use pmulhw, results in 8 16-bit, note we need simulate unsigned * signed,_mm_srai_epi16(sAR2_Q14_lo_76543210, 15) */
+        xmm_tempa = _mm_cmpgt_epi16( _mm_setzero_si128(), sAR2_Q14_lo_76543210 );
+        xmm_tempa = _mm_and_si128( xmm_tempa, AR_shp_Q13_76543210 );
+
+        xmm_lo_07 = _mm_mulhi_epi16( sAR2_Q14_lo_76543210, AR_shp_Q13_76543210 );
+        xmm_lo_07 = _mm_add_epi16( xmm_lo_07, xmm_tempa );
+
+        xmm_lo_07 = _mm_madd_epi16( xmm_lo_07, xmm_one );
+
+        /* accumulate */
+        xmm_hi_07 = _mm_add_epi32( xmm_hi_07, xmm_lo_07 );
+
+        xmm_hi_07 = _mm_add_epi32( xmm_hi_07, _mm_unpackhi_epi64(xmm_hi_07, xmm_hi_07 ) );
+        xmm_hi_07 = _mm_add_epi32( xmm_hi_07, _mm_shufflelo_epi16(xmm_hi_07, 0x0E ) );
+
+        n_AR_Q12 = 5 + _mm_cvtsi128_si32( xmm_hi_07 );
+
+        n_AR_Q12 = silk_SMLAWB( n_AR_Q12, NSQ->sAR2_Q14[ 8 ], AR_shp_Q13[ 8 ] );
+        n_AR_Q12 = silk_SMLAWB( n_AR_Q12, NSQ->sAR2_Q14[ 9 ], AR_shp_Q13[ 9 ] );
+
+        n_AR_Q12 = silk_LSHIFT32( n_AR_Q12, 1 );                                /* Q11 -> Q12 */
+        n_AR_Q12 = silk_SMLAWB( n_AR_Q12, sLF_AR_shp_Q14, Tilt_Q14 );
+
+        n_LF_Q12 = silk_SMULWB( NSQ->sLTP_shp_Q14[ NSQ->sLTP_shp_buf_idx - 1 ], LF_shp_Q14 );
+        n_LF_Q12 = silk_SMLAWT( n_LF_Q12, sLF_AR_shp_Q14, LF_shp_Q14 );
+
+        silk_assert( lag > 0 || signalType != TYPE_VOICED );
+
+        /* Combine prediction and noise shaping signals */
+        tmp1 = silk_SUB32( silk_LSHIFT32( LPC_pred_Q10, 2 ), n_AR_Q12 );        /* Q12 */
+        tmp1 = silk_SUB32( tmp1, n_LF_Q12 );                                    /* Q12 */
+        if( lag > 0 ) {
+            /* Symmetric, packed FIR coefficients */
+            n_LTP_Q13 = silk_SMULWB( silk_ADD32( shp_lag_ptr[ 0 ], shp_lag_ptr[ -2 ] ), HarmShapeFIRPacked_Q14 );
+            n_LTP_Q13 = silk_SMLAWT( n_LTP_Q13, shp_lag_ptr[ -1 ],                      HarmShapeFIRPacked_Q14 );
+            n_LTP_Q13 = silk_LSHIFT( n_LTP_Q13, 1 );
+            shp_lag_ptr++;
+
+            tmp2 = silk_SUB32( LTP_pred_Q13, n_LTP_Q13 );                       /* Q13 */
+            tmp1 = silk_ADD_LSHIFT32( tmp2, tmp1, 1 );                          /* Q13 */
+            tmp1 = silk_RSHIFT_ROUND( tmp1, 3 );                                /* Q10 */
+        } else {
+            tmp1 = silk_RSHIFT_ROUND( tmp1, 2 );                                /* Q10 */
+        }
+
+        r_Q10 = silk_SUB32( x_sc_Q10[ i ], tmp1 );                              /* residual error Q10 */
+
+        /* Generate dither */
+        NSQ->rand_seed = silk_RAND( NSQ->rand_seed );
+
+        /* Flip sign depending on dither */
+        tmp2 = -r_Q10;
+        if ( NSQ->rand_seed < 0 ) r_Q10 = tmp2;
+
+        r_Q10 = silk_LIMIT_32( r_Q10, -(31 << 10), 30 << 10 );
+
+        /* Find two quantization level candidates and measure their rate-distortion */
+        q1_Q10 = silk_SUB32( r_Q10, offset_Q10 );
+        q1_Q0 = silk_RSHIFT( q1_Q10, 10 );
+
+        q1_Q10 = table[q1_Q0][0];
+        q2_Q10 = table[q1_Q0][1];
+
+        if (r_Q10 * table[q1_Q0][2] - table[q1_Q0][3] < 0)
+        {
+            q1_Q10 = q2_Q10;
+        }
+
+        pulses[ i ] = (opus_int8)silk_RSHIFT_ROUND( q1_Q10, 10 );
+
+        /* Excitation */
+        exc_Q14 = silk_LSHIFT( q1_Q10, 4 );
+
+        tmp2 = -exc_Q14;
+        if ( NSQ->rand_seed < 0 ) exc_Q14 = tmp2;
+
+        /* Add predictions */
+        LPC_exc_Q14 = silk_ADD_LSHIFT32( exc_Q14, LTP_pred_Q13, 1 );
+        xq_Q14      = silk_ADD_LSHIFT32( LPC_exc_Q14, LPC_pred_Q10, 4 );
+
+        /* Update states */
+        psLPC_Q14++;
+        *psLPC_Q14 = xq_Q14;
+        sLF_AR_shp_Q14 = silk_SUB_LSHIFT32( xq_Q14, n_AR_Q12, 2 );
+
+        NSQ->sLTP_shp_Q14[ NSQ->sLTP_shp_buf_idx ] = silk_SUB_LSHIFT32( sLF_AR_shp_Q14, n_LF_Q12, 2 );
+        sLTP_Q15[ NSQ->sLTP_buf_idx ] = silk_LSHIFT( LPC_exc_Q14, 1 );
+        NSQ->sLTP_shp_buf_idx++;
+        NSQ->sLTP_buf_idx++;
+
+        /* Make dither dependent on quantized signal */
+        NSQ->rand_seed = silk_ADD32_ovflw( NSQ->rand_seed, pulses[ i ] );
+    }
+
+    NSQ->sLF_AR_shp_Q14 = sLF_AR_shp_Q14;
+
+    /* Scale XQ back to normal level before saving */
+    psLPC_Q14 = &NSQ->sLPC_Q14[ NSQ_LPC_BUF_LENGTH ];
+
+    /* write back sAR2_Q14 */
+    xmm_tempa = _mm_unpackhi_epi16( sAR2_Q14_lo_76543210, sAR2_Q14_hi_76543210 );
+    xmm_tempb = _mm_unpacklo_epi16( sAR2_Q14_lo_76543210, sAR2_Q14_hi_76543210 );
+    _mm_storeu_si128( (__m128i *)(&NSQ->sAR2_Q14[ 4 ]), xmm_tempa );
+    _mm_storeu_si128( (__m128i *)(&NSQ->sAR2_Q14[ 0 ]), xmm_tempb );
+
+    /* xq[ i ] = (opus_int16)silk_SAT16( silk_RSHIFT_ROUND( silk_SMULWW( psLPC_Q14[ i ], Gain_Q10 ), 8 ) ); */
+    {
+        __m128i xmm_Gain_Q10;
+        __m128i xmm_xq_Q14_3210, xmm_xq_Q14_x3x1, xmm_xq_Q14_7654, xmm_xq_Q14_x7x5;
+
+        /* prepare (1 << 7) in packed 4 32-bits */
+        xmm_tempa = _mm_set1_epi32( (1 << 7) );
+
+        /* prepare Gain_Q10 in packed 4 32-bits */
+        xmm_Gain_Q10 = _mm_set1_epi32( Gain_Q10 );
+
+        /* process xq */
+        for (i = 0; i < length - 7; i += 8)
+        {
+            xmm_xq_Q14_3210 = _mm_loadu_si128( (__m128i *)(&(psLPC_Q14[ i + 0 ] ) ) );
+            xmm_xq_Q14_7654 = _mm_loadu_si128( (__m128i *)(&(psLPC_Q14[ i + 4 ] ) ) );
+
+            /* equal shift right 4 bytes*/
+            xmm_xq_Q14_x3x1 = _mm_shuffle_epi32( xmm_xq_Q14_3210, _MM_SHUFFLE( 0, 3, 2, 1 ) );
+            /* equal shift right 4 bytes*/
+            xmm_xq_Q14_x7x5 = _mm_shuffle_epi32( xmm_xq_Q14_7654, _MM_SHUFFLE( 0, 3, 2, 1 ) );
+
+            xmm_xq_Q14_3210 = _mm_mul_epi32( xmm_xq_Q14_3210, xmm_Gain_Q10 );
+            xmm_xq_Q14_x3x1 = _mm_mul_epi32( xmm_xq_Q14_x3x1, xmm_Gain_Q10 );
+            xmm_xq_Q14_7654 = _mm_mul_epi32( xmm_xq_Q14_7654, xmm_Gain_Q10 );
+            xmm_xq_Q14_x7x5 = _mm_mul_epi32( xmm_xq_Q14_x7x5, xmm_Gain_Q10 );
+
+            xmm_xq_Q14_3210 = _mm_srli_epi64( xmm_xq_Q14_3210, 16 );
+            xmm_xq_Q14_x3x1 = _mm_slli_epi64( xmm_xq_Q14_x3x1, 16 );
+            xmm_xq_Q14_7654 = _mm_srli_epi64( xmm_xq_Q14_7654, 16 );
+            xmm_xq_Q14_x7x5 = _mm_slli_epi64( xmm_xq_Q14_x7x5, 16 );
+
+            xmm_xq_Q14_3210 = _mm_blend_epi16( xmm_xq_Q14_3210, xmm_xq_Q14_x3x1, 0xCC );
+            xmm_xq_Q14_7654 = _mm_blend_epi16( xmm_xq_Q14_7654, xmm_xq_Q14_x7x5, 0xCC );
+
+            /* silk_RSHIFT_ROUND(xq, 8) */
+            xmm_xq_Q14_3210 = _mm_add_epi32( xmm_xq_Q14_3210, xmm_tempa );
+            xmm_xq_Q14_7654 = _mm_add_epi32( xmm_xq_Q14_7654, xmm_tempa );
+
+            xmm_xq_Q14_3210 = _mm_srai_epi32( xmm_xq_Q14_3210, 8 );
+            xmm_xq_Q14_7654 = _mm_srai_epi32( xmm_xq_Q14_7654, 8 );
+
+            /* silk_SAT16 */
+            xmm_xq_Q14_3210 = _mm_packs_epi32( xmm_xq_Q14_3210, xmm_xq_Q14_7654 );
+
+            /* save to xq */
+            _mm_storeu_si128( (__m128i *)(&xq[ i ] ), xmm_xq_Q14_3210 );
+        }
+    }
+    for ( ; i < length; i++)
+    {
+        xq[i] = (opus_int16)silk_SAT16( silk_RSHIFT_ROUND( silk_SMULWW( psLPC_Q14[ i ], Gain_Q10 ), 8 ) );
+    }
+
+    /* Update LPC synth buffer */
+    silk_memcpy( NSQ->sLPC_Q14, &NSQ->sLPC_Q14[ length ], NSQ_LPC_BUF_LENGTH * sizeof( opus_int32 ) );
+}
+
+static OPUS_INLINE void silk_nsq_scale_states_sse4_1(
+    const silk_encoder_state *psEncC,           /* I    Encoder State                   */
+    silk_nsq_state      *NSQ,                   /* I/O  NSQ state                       */
+    const opus_int32    x_Q3[],                 /* I    input in Q3                     */
+    opus_int32          x_sc_Q10[],             /* O    input scaled with 1/Gain        */
+    const opus_int16    sLTP[],                 /* I    re-whitened LTP state in Q0     */
+    opus_int32          sLTP_Q15[],             /* O    LTP state matching scaled input */
+    opus_int            subfr,                  /* I    subframe number                 */
+    const opus_int      LTP_scale_Q14,          /* I                                    */
+    const opus_int32    Gains_Q16[ MAX_NB_SUBFR ], /* I                                 */
+    const opus_int      pitchL[ MAX_NB_SUBFR ], /* I    Pitch lag                       */
+    const opus_int      signal_type             /* I    Signal type                     */
+)
+{
+    opus_int   i, lag;
+    opus_int32 gain_adj_Q16, inv_gain_Q31, inv_gain_Q23;
+    __m128i xmm_inv_gain_Q23, xmm_x_Q3_x2x0, xmm_x_Q3_x3x1;
+
+    lag          = pitchL[ subfr ];
+    inv_gain_Q31 = silk_INVERSE32_varQ( silk_max( Gains_Q16[ subfr ], 1 ), 47 );
+    silk_assert( inv_gain_Q31 != 0 );
+
+    /* Calculate gain adjustment factor */
+    if( Gains_Q16[ subfr ] != NSQ->prev_gain_Q16 ) {
+        gain_adj_Q16 =  silk_DIV32_varQ( NSQ->prev_gain_Q16, Gains_Q16[ subfr ], 16 );
+    } else {
+        gain_adj_Q16 = (opus_int32)1 << 16;
+    }
+
+    /* Scale input */
+    inv_gain_Q23 = silk_RSHIFT_ROUND( inv_gain_Q31, 8 );
+
+    /* prepare inv_gain_Q23 in packed 4 32-bits */
+    xmm_inv_gain_Q23 = _mm_set1_epi32(inv_gain_Q23);
+
+    for( i = 0; i < psEncC->subfr_length - 3; i += 4 ) {
+        xmm_x_Q3_x2x0 = _mm_loadu_si128( (__m128i *)(&(x_Q3[ i ] ) ) );
+
+        /* equal shift right 4 bytes*/
+        xmm_x_Q3_x3x1 = _mm_shuffle_epi32( xmm_x_Q3_x2x0, _MM_SHUFFLE( 0, 3, 2, 1 ) );
+
+        xmm_x_Q3_x2x0 = _mm_mul_epi32( xmm_x_Q3_x2x0, xmm_inv_gain_Q23 );
+        xmm_x_Q3_x3x1 = _mm_mul_epi32( xmm_x_Q3_x3x1, xmm_inv_gain_Q23 );
+
+        xmm_x_Q3_x2x0 = _mm_srli_epi64( xmm_x_Q3_x2x0, 16 );
+        xmm_x_Q3_x3x1 = _mm_slli_epi64( xmm_x_Q3_x3x1, 16 );
+
+        xmm_x_Q3_x2x0 = _mm_blend_epi16( xmm_x_Q3_x2x0, xmm_x_Q3_x3x1, 0xCC );
+
+        _mm_storeu_si128( (__m128i *)(&(x_sc_Q10[ i ] ) ), xmm_x_Q3_x2x0 );
+    }
+
+    for( ; i < psEncC->subfr_length; i++ ) {
+        x_sc_Q10[ i ] = silk_SMULWW( x_Q3[ i ], inv_gain_Q23 );
+    }
+
+    /* Save inverse gain */
+    NSQ->prev_gain_Q16 = Gains_Q16[ subfr ];
+
+    /* After rewhitening the LTP state is un-scaled, so scale with inv_gain_Q16 */
+    if( NSQ->rewhite_flag ) {
+        if( subfr == 0 ) {
+            /* Do LTP downscaling */
+            inv_gain_Q31 = silk_LSHIFT( silk_SMULWB( inv_gain_Q31, LTP_scale_Q14 ), 2 );
+        }
+        for( i = NSQ->sLTP_buf_idx - lag - LTP_ORDER / 2; i < NSQ->sLTP_buf_idx; i++ ) {
+            silk_assert( i < MAX_FRAME_LENGTH );
+            sLTP_Q15[ i ] = silk_SMULWB( inv_gain_Q31, sLTP[ i ] );
+        }
+    }
+
+    /* Adjust for changing gain */
+    if( gain_adj_Q16 != (opus_int32)1 << 16 ) {
+        /* Scale long-term shaping state */
+        __m128i xmm_gain_adj_Q16, xmm_sLTP_shp_Q14_x2x0, xmm_sLTP_shp_Q14_x3x1;
+
+        /* prepare gain_adj_Q16 in packed 4 32-bits */
+        xmm_gain_adj_Q16 = _mm_set1_epi32(gain_adj_Q16);
+
+        for( i = NSQ->sLTP_shp_buf_idx - psEncC->ltp_mem_length; i < NSQ->sLTP_shp_buf_idx - 3; i += 4 )
+        {
+            xmm_sLTP_shp_Q14_x2x0 = _mm_loadu_si128( (__m128i *)(&(NSQ->sLTP_shp_Q14[ i ] ) ) );
+            /* equal shift right 4 bytes*/
+            xmm_sLTP_shp_Q14_x3x1 = _mm_shuffle_epi32( xmm_sLTP_shp_Q14_x2x0, _MM_SHUFFLE( 0, 3, 2, 1 ) );
+
+            xmm_sLTP_shp_Q14_x2x0 = _mm_mul_epi32( xmm_sLTP_shp_Q14_x2x0, xmm_gain_adj_Q16 );
+            xmm_sLTP_shp_Q14_x3x1 = _mm_mul_epi32( xmm_sLTP_shp_Q14_x3x1, xmm_gain_adj_Q16 );
+
+            xmm_sLTP_shp_Q14_x2x0 = _mm_srli_epi64( xmm_sLTP_shp_Q14_x2x0, 16 );
+            xmm_sLTP_shp_Q14_x3x1 = _mm_slli_epi64( xmm_sLTP_shp_Q14_x3x1, 16 );
+
+            xmm_sLTP_shp_Q14_x2x0 = _mm_blend_epi16( xmm_sLTP_shp_Q14_x2x0, xmm_sLTP_shp_Q14_x3x1, 0xCC );
+
+            _mm_storeu_si128( (__m128i *)(&(NSQ->sLTP_shp_Q14[ i ] ) ), xmm_sLTP_shp_Q14_x2x0 );
+        }
+
+        for( ; i < NSQ->sLTP_shp_buf_idx; i++ ) {
+            NSQ->sLTP_shp_Q14[ i ] = silk_SMULWW( gain_adj_Q16, NSQ->sLTP_shp_Q14[ i ] );
+        }
+
+        /* Scale long-term prediction state */
+        if( signal_type == TYPE_VOICED && NSQ->rewhite_flag == 0 ) {
+            for( i = NSQ->sLTP_buf_idx - lag - LTP_ORDER / 2; i < NSQ->sLTP_buf_idx; i++ ) {
+                sLTP_Q15[ i ] = silk_SMULWW( gain_adj_Q16, sLTP_Q15[ i ] );
+            }
+        }
+
+        NSQ->sLF_AR_shp_Q14 = silk_SMULWW( gain_adj_Q16, NSQ->sLF_AR_shp_Q14 );
+
+        /* Scale short-term prediction and shaping states */
+        for( i = 0; i < NSQ_LPC_BUF_LENGTH; i++ ) {
+            NSQ->sLPC_Q14[ i ] = silk_SMULWW( gain_adj_Q16, NSQ->sLPC_Q14[ i ] );
+        }
+        for( i = 0; i < MAX_SHAPE_LPC_ORDER; i++ ) {
+            NSQ->sAR2_Q14[ i ] = silk_SMULWW( gain_adj_Q16, NSQ->sAR2_Q14[ i ] );
+        }
+    }
+}
diff --git a/media/libopus/silk/x86/SigProc_FIX_sse.h b/media/libopus/silk/x86/SigProc_FIX_sse.h
new file mode 100644
index 000000000..61efa8da4
--- /dev/null
+++ b/media/libopus/silk/x86/SigProc_FIX_sse.h
@@ -0,0 +1,94 @@
+/* Copyright (c) 2014, Cisco Systems, INC
+   Written by XiangMingZhu WeiZhou MinPeng YanWang
+
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions
+   are met:
+
+   - Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+
+   - Redistributions in binary form must reproduce the above copyright
+   notice, this list of conditions and the following disclaimer in the
+   documentation and/or other materials provided with the distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
+   OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+   EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+   PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+   PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+   LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+   NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+   SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+
+#ifndef SIGPROC_FIX_SSE_H
+#define SIGPROC_FIX_SSE_H
+
+#ifdef HAVE_CONFIG_H
+#include "config.h"
+#endif
+
+#if defined(OPUS_X86_MAY_HAVE_SSE4_1)
+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                                       */
+);
+
+#if defined(OPUS_X86_PRESUME_SSE4_1)
+#define silk_burg_modified(res_nrg, res_nrg_Q, A_Q16, x, minInvGain_Q30, subfr_length, nb_subfr, D, arch) \
+    ((void)(arch), silk_burg_modified_sse4_1(res_nrg, res_nrg_Q, A_Q16, x, minInvGain_Q30, subfr_length, nb_subfr, D, arch))
+
+#else
+
+extern void (*const SILK_BURG_MODIFIED_IMPL[OPUS_ARCHMASK + 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                                       */);
+
+#  define silk_burg_modified(res_nrg, res_nrg_Q, A_Q16, x, minInvGain_Q30, subfr_length, nb_subfr, D, arch) \
+    ((*SILK_BURG_MODIFIED_IMPL[(arch) & OPUS_ARCHMASK])(res_nrg, res_nrg_Q, A_Q16, x, minInvGain_Q30, subfr_length, nb_subfr, D, arch))
+
+#endif
+
+opus_int64 silk_inner_prod16_aligned_64_sse4_1(
+    const opus_int16 *inVec1,
+    const opus_int16 *inVec2,
+    const opus_int   len
+);
+
+
+#if defined(OPUS_X86_PRESUME_SSE4_1)
+
+#define silk_inner_prod16_aligned_64(inVec1, inVec2, len, arch) \
+    ((void)(arch),silk_inner_prod16_aligned_64_sse4_1(inVec1, inVec2, len))
+
+#else
+
+extern opus_int64 (*const SILK_INNER_PROD16_ALIGNED_64_IMPL[OPUS_ARCHMASK + 1])(
+                    const opus_int16 *inVec1,
+                    const opus_int16 *inVec2,
+                    const opus_int   len);
+
+#  define silk_inner_prod16_aligned_64(inVec1, inVec2, len, arch) \
+    ((*SILK_INNER_PROD16_ALIGNED_64_IMPL[(arch) & OPUS_ARCHMASK])(inVec1, inVec2, len))
+
+#endif
+#endif
+#endif
diff --git a/media/libopus/silk/x86/VAD_sse.c b/media/libopus/silk/x86/VAD_sse.c
new file mode 100644
index 000000000..4e90f4410
--- /dev/null
+++ b/media/libopus/silk/x86/VAD_sse.c
@@ -0,0 +1,277 @@
+/* Copyright (c) 2014, Cisco Systems, INC
+   Written by XiangMingZhu WeiZhou MinPeng YanWang
+
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions
+   are met:
+
+   - Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+
+   - Redistributions in binary form must reproduce the above copyright
+   notice, this list of conditions and the following disclaimer in the
+   documentation and/or other materials provided with the distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
+   OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+   EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+   PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+   PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+   LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+   NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+   SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+
+#ifdef HAVE_CONFIG_H
+#include "config.h"
+#endif
+
+#include <xmmintrin.h>
+#include <emmintrin.h>
+#include <smmintrin.h>
+
+#include "main.h"
+#include "stack_alloc.h"
+
+/* Weighting factors for tilt measure */
+static const opus_int32 tiltWeights[ VAD_N_BANDS ] = { 30000, 6000, -12000, -12000 };
+
+/***************************************/
+/* Get the speech activity level in Q8 */
+/***************************************/
+opus_int silk_VAD_GetSA_Q8_sse4_1(                  /* O    Return value, 0 if success                  */
+    silk_encoder_state          *psEncC,            /* I/O  Encoder state                               */
+    const opus_int16            pIn[]               /* I    PCM input                                   */
+)
+{
+    opus_int   SA_Q15, pSNR_dB_Q7, input_tilt;
+    opus_int   decimated_framelength1, decimated_framelength2;
+    opus_int   decimated_framelength;
+    opus_int   dec_subframe_length, dec_subframe_offset, SNR_Q7, i, b, s;
+    opus_int32 sumSquared, smooth_coef_Q16;
+    opus_int16 HPstateTmp;
+    VARDECL( opus_int16, X );
+    opus_int32 Xnrg[ VAD_N_BANDS ];
+    opus_int32 NrgToNoiseRatio_Q8[ VAD_N_BANDS ];
+    opus_int32 speech_nrg, x_tmp;
+    opus_int   X_offset[ VAD_N_BANDS ];
+    opus_int   ret = 0;
+    silk_VAD_state *psSilk_VAD = &psEncC->sVAD;
+
+    SAVE_STACK;
+
+    /* Safety checks */
+    silk_assert( VAD_N_BANDS == 4 );
+    silk_assert( MAX_FRAME_LENGTH >= psEncC->frame_length );
+    silk_assert( psEncC->frame_length <= 512 );
+    silk_assert( psEncC->frame_length == 8 * silk_RSHIFT( psEncC->frame_length, 3 ) );
+
+    /***********************/
+    /* Filter and Decimate */
+    /***********************/
+    decimated_framelength1 = silk_RSHIFT( psEncC->frame_length, 1 );
+    decimated_framelength2 = silk_RSHIFT( psEncC->frame_length, 2 );
+    decimated_framelength = silk_RSHIFT( psEncC->frame_length, 3 );
+    /* Decimate into 4 bands:
+       0       L      3L       L              3L                             5L
+               -      --       -              --                             --
+               8       8       2               4                              4
+
+       [0-1 kHz| temp. |1-2 kHz|    2-4 kHz    |            4-8 kHz           |
+
+       They're arranged to allow the minimal ( frame_length / 4 ) extra
+       scratch space during the downsampling process */
+    X_offset[ 0 ] = 0;
+    X_offset[ 1 ] = decimated_framelength + decimated_framelength2;
+    X_offset[ 2 ] = X_offset[ 1 ] + decimated_framelength;
+    X_offset[ 3 ] = X_offset[ 2 ] + decimated_framelength2;
+    ALLOC( X, X_offset[ 3 ] + decimated_framelength1, opus_int16 );
+
+    /* 0-8 kHz to 0-4 kHz and 4-8 kHz */
+    silk_ana_filt_bank_1( pIn, &psSilk_VAD->AnaState[  0 ],
+        X, &X[ X_offset[ 3 ] ], psEncC->frame_length );
+
+    /* 0-4 kHz to 0-2 kHz and 2-4 kHz */
+    silk_ana_filt_bank_1( X, &psSilk_VAD->AnaState1[ 0 ],
+        X, &X[ X_offset[ 2 ] ], decimated_framelength1 );
+
+    /* 0-2 kHz to 0-1 kHz and 1-2 kHz */
+    silk_ana_filt_bank_1( X, &psSilk_VAD->AnaState2[ 0 ],
+        X, &X[ X_offset[ 1 ] ], decimated_framelength2 );
+
+    /*********************************************/
+    /* HP filter on lowest band (differentiator) */
+    /*********************************************/
+    X[ decimated_framelength - 1 ] = silk_RSHIFT( X[ decimated_framelength - 1 ], 1 );
+    HPstateTmp = X[ decimated_framelength - 1 ];
+    for( i = decimated_framelength - 1; i > 0; i-- ) {
+        X[ i - 1 ]  = silk_RSHIFT( X[ i - 1 ], 1 );
+        X[ i ]     -= X[ i - 1 ];
+    }
+    X[ 0 ] -= psSilk_VAD->HPstate;
+    psSilk_VAD->HPstate = HPstateTmp;
+
+    /*************************************/
+    /* Calculate the energy in each band */
+    /*************************************/
+    for( b = 0; b < VAD_N_BANDS; b++ ) {
+        /* Find the decimated framelength in the non-uniformly divided bands */
+        decimated_framelength = silk_RSHIFT( psEncC->frame_length, silk_min_int( VAD_N_BANDS - b, VAD_N_BANDS - 1 ) );
+
+        /* Split length into subframe lengths */
+        dec_subframe_length = silk_RSHIFT( decimated_framelength, VAD_INTERNAL_SUBFRAMES_LOG2 );
+        dec_subframe_offset = 0;
+
+        /* Compute energy per sub-frame */
+        /* initialize with summed energy of last subframe */
+        Xnrg[ b ] = psSilk_VAD->XnrgSubfr[ b ];
+        for( s = 0; s < VAD_INTERNAL_SUBFRAMES; s++ ) {
+            __m128i xmm_X, xmm_acc;
+            sumSquared = 0;
+
+            xmm_acc = _mm_setzero_si128();
+
+            for( i = 0; i < dec_subframe_length - 7; i += 8 )
+            {
+                xmm_X   = _mm_loadu_si128( (__m128i *)&(X[ X_offset[ b ] + i + dec_subframe_offset ] ) );
+                xmm_X   = _mm_srai_epi16( xmm_X, 3 );
+                xmm_X   = _mm_madd_epi16( xmm_X, xmm_X );
+                xmm_acc = _mm_add_epi32( xmm_acc, xmm_X );
+            }
+
+            xmm_acc = _mm_add_epi32( xmm_acc, _mm_unpackhi_epi64( xmm_acc, xmm_acc ) );
+            xmm_acc = _mm_add_epi32( xmm_acc, _mm_shufflelo_epi16( xmm_acc, 0x0E ) );
+
+            sumSquared += _mm_cvtsi128_si32( xmm_acc );
+
+            for( ; i < dec_subframe_length; i++ ) {
+                /* The energy will be less than dec_subframe_length * ( silk_int16_MIN / 8 ) ^ 2.            */
+                /* Therefore we can accumulate with no risk of overflow (unless dec_subframe_length > 128)  */
+                x_tmp = silk_RSHIFT(
+                    X[ X_offset[ b ] + i + dec_subframe_offset ], 3 );
+                sumSquared = silk_SMLABB( sumSquared, x_tmp, x_tmp );
+
+                /* Safety check */
+                silk_assert( sumSquared >= 0 );
+            }
+
+            /* Add/saturate summed energy of current subframe */
+            if( s < VAD_INTERNAL_SUBFRAMES - 1 ) {
+                Xnrg[ b ] = silk_ADD_POS_SAT32( Xnrg[ b ], sumSquared );
+            } else {
+                /* Look-ahead subframe */
+                Xnrg[ b ] = silk_ADD_POS_SAT32( Xnrg[ b ], silk_RSHIFT( sumSquared, 1 ) );
+            }
+
+            dec_subframe_offset += dec_subframe_length;
+        }
+        psSilk_VAD->XnrgSubfr[ b ] = sumSquared;
+    }
+
+    /********************/
+    /* Noise estimation */
+    /********************/
+    silk_VAD_GetNoiseLevels( &Xnrg[ 0 ], psSilk_VAD );
+
+    /***********************************************/
+    /* Signal-plus-noise to noise ratio estimation */
+    /***********************************************/
+    sumSquared = 0;
+    input_tilt = 0;
+    for( b = 0; b < VAD_N_BANDS; b++ ) {
+        speech_nrg = Xnrg[ b ] - psSilk_VAD->NL[ b ];
+        if( speech_nrg > 0 ) {
+            /* Divide, with sufficient resolution */
+            if( ( Xnrg[ b ] & 0xFF800000 ) == 0 ) {
+                NrgToNoiseRatio_Q8[ b ] = silk_DIV32( silk_LSHIFT( Xnrg[ b ], 8 ), psSilk_VAD->NL[ b ] + 1 );
+            } else {
+                NrgToNoiseRatio_Q8[ b ] = silk_DIV32( Xnrg[ b ], silk_RSHIFT( psSilk_VAD->NL[ b ], 8 ) + 1 );
+            }
+
+            /* Convert to log domain */
+            SNR_Q7 = silk_lin2log( NrgToNoiseRatio_Q8[ b ] ) - 8 * 128;
+
+            /* Sum-of-squares */
+            sumSquared = silk_SMLABB( sumSquared, SNR_Q7, SNR_Q7 );          /* Q14 */
+
+            /* Tilt measure */
+            if( speech_nrg < ( (opus_int32)1 << 20 ) ) {
+                /* Scale down SNR value for small subband speech energies */
+                SNR_Q7 = silk_SMULWB( silk_LSHIFT( silk_SQRT_APPROX( speech_nrg ), 6 ), SNR_Q7 );
+            }
+            input_tilt = silk_SMLAWB( input_tilt, tiltWeights[ b ], SNR_Q7 );
+        } else {
+            NrgToNoiseRatio_Q8[ b ] = 256;
+        }
+    }
+
+    /* Mean-of-squares */
+    sumSquared = silk_DIV32_16( sumSquared, VAD_N_BANDS ); /* Q14 */
+
+    /* Root-mean-square approximation, scale to dBs, and write to output pointer */
+    pSNR_dB_Q7 = (opus_int16)( 3 * silk_SQRT_APPROX( sumSquared ) ); /* Q7 */
+
+    /*********************************/
+    /* Speech Probability Estimation */
+    /*********************************/
+    SA_Q15 = silk_sigm_Q15( silk_SMULWB( VAD_SNR_FACTOR_Q16, pSNR_dB_Q7 ) - VAD_NEGATIVE_OFFSET_Q5 );
+
+    /**************************/
+    /* Frequency Tilt Measure */
+    /**************************/
+    psEncC->input_tilt_Q15 = silk_LSHIFT( silk_sigm_Q15( input_tilt ) - 16384, 1 );
+
+    /**************************************************/
+    /* Scale the sigmoid output based on power levels */
+    /**************************************************/
+    speech_nrg = 0;
+    for( b = 0; b < VAD_N_BANDS; b++ ) {
+        /* Accumulate signal-without-noise energies, higher frequency bands have more weight */
+        speech_nrg += ( b + 1 ) * silk_RSHIFT( Xnrg[ b ] - psSilk_VAD->NL[ b ], 4 );
+    }
+
+    /* Power scaling */
+    if( speech_nrg <= 0 ) {
+        SA_Q15 = silk_RSHIFT( SA_Q15, 1 );
+    } else if( speech_nrg < 32768 ) {
+        if( psEncC->frame_length == 10 * psEncC->fs_kHz ) {
+            speech_nrg = silk_LSHIFT_SAT32( speech_nrg, 16 );
+        } else {
+            speech_nrg = silk_LSHIFT_SAT32( speech_nrg, 15 );
+        }
+
+        /* square-root */
+        speech_nrg = silk_SQRT_APPROX( speech_nrg );
+        SA_Q15 = silk_SMULWB( 32768 + speech_nrg, SA_Q15 );
+    }
+
+    /* Copy the resulting speech activity in Q8 */
+    psEncC->speech_activity_Q8 = silk_min_int( silk_RSHIFT( SA_Q15, 7 ), silk_uint8_MAX );
+
+    /***********************************/
+    /* Energy Level and SNR estimation */
+    /***********************************/
+    /* Smoothing coefficient */
+    smooth_coef_Q16 = silk_SMULWB( VAD_SNR_SMOOTH_COEF_Q18, silk_SMULWB( (opus_int32)SA_Q15, SA_Q15 ) );
+
+    if( psEncC->frame_length == 10 * psEncC->fs_kHz ) {
+        smooth_coef_Q16 >>= 1;
+    }
+
+    for( b = 0; b < VAD_N_BANDS; b++ ) {
+        /* compute smoothed energy-to-noise ratio per band */
+        psSilk_VAD->NrgRatioSmth_Q8[ b ] = silk_SMLAWB( psSilk_VAD->NrgRatioSmth_Q8[ b ],
+            NrgToNoiseRatio_Q8[ b ] - psSilk_VAD->NrgRatioSmth_Q8[ b ], smooth_coef_Q16 );
+
+        /* signal to noise ratio in dB per band */
+        SNR_Q7 = 3 * ( silk_lin2log( psSilk_VAD->NrgRatioSmth_Q8[b] ) - 8 * 128 );
+        /* quality = sigmoid( 0.25 * ( SNR_dB - 16 ) ); */
+        psEncC->input_quality_bands_Q15[ b ] = silk_sigm_Q15( silk_RSHIFT( SNR_Q7 - 16 * 128, 4 ) );
+    }
+
+    RESTORE_STACK;
+    return( ret );
+}
diff --git a/media/libopus/silk/x86/VQ_WMat_EC_sse.c b/media/libopus/silk/x86/VQ_WMat_EC_sse.c
new file mode 100644
index 000000000..74d6c6d0e
--- /dev/null
+++ b/media/libopus/silk/x86/VQ_WMat_EC_sse.c
@@ -0,0 +1,142 @@
+/* Copyright (c) 2014, Cisco Systems, INC
+   Written by XiangMingZhu WeiZhou MinPeng YanWang
+
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions
+   are met:
+
+   - Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+
+   - Redistributions in binary form must reproduce the above copyright
+   notice, this list of conditions and the following disclaimer in the
+   documentation and/or other materials provided with the distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
+   OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+   EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+   PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+   PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+   LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+   NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+   SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+
+#ifdef HAVE_CONFIG_H
+#include "config.h"
+#endif
+
+#include <xmmintrin.h>
+#include <emmintrin.h>
+#include <smmintrin.h>
+#include "main.h"
+#include "celt/x86/x86cpu.h"
+
+/* Entropy constrained matrix-weighted VQ, hard-coded to 5-element vectors, for a single input data vector */
+void silk_VQ_WMat_EC_sse4_1(
+    opus_int8                   *ind,                           /* O    index of best codebook vector               */
+    opus_int32                  *rate_dist_Q14,                 /* O    best weighted quant error + mu * rate       */
+    opus_int                    *gain_Q7,                       /* O    sum of absolute LTP coefficients            */
+    const opus_int16            *in_Q14,                        /* I    input vector to be quantized                */
+    const opus_int32            *W_Q18,                         /* I    weighting matrix                            */
+    const opus_int8             *cb_Q7,                         /* I    codebook                                    */
+    const opus_uint8            *cb_gain_Q7,                    /* I    codebook effective gain                     */
+    const opus_uint8            *cl_Q5,                         /* I    code length for each codebook vector        */
+    const opus_int              mu_Q9,                          /* I    tradeoff betw. weighted error and rate      */
+    const opus_int32            max_gain_Q7,                    /* I    maximum sum of absolute LTP coefficients    */
+    opus_int                    L                               /* I    number of vectors in codebook               */
+)
+{
+    opus_int   k, gain_tmp_Q7;
+    const opus_int8 *cb_row_Q7;
+    opus_int16 diff_Q14[ 5 ];
+    opus_int32 sum1_Q14, sum2_Q16;
+
+    __m128i C_tmp1, C_tmp2, C_tmp3, C_tmp4, C_tmp5;
+    /* Loop over codebook */
+    *rate_dist_Q14 = silk_int32_MAX;
+    cb_row_Q7 = cb_Q7;
+    for( k = 0; k < L; k++ ) {
+        gain_tmp_Q7 = cb_gain_Q7[k];
+
+        diff_Q14[ 0 ] = in_Q14[ 0 ] - silk_LSHIFT( cb_row_Q7[ 0 ], 7 );
+
+        C_tmp1 = OP_CVTEPI16_EPI32_M64( &in_Q14[ 1 ] );
+        C_tmp2 = OP_CVTEPI8_EPI32_M32( &cb_row_Q7[ 1 ] );
+        C_tmp2 = _mm_slli_epi32( C_tmp2, 7 );
+        C_tmp1 = _mm_sub_epi32( C_tmp1, C_tmp2 );
+
+        diff_Q14[ 1 ] = _mm_extract_epi16( C_tmp1, 0 );
+        diff_Q14[ 2 ] = _mm_extract_epi16( C_tmp1, 2 );
+        diff_Q14[ 3 ] = _mm_extract_epi16( C_tmp1, 4 );
+        diff_Q14[ 4 ] = _mm_extract_epi16( C_tmp1, 6 );
+
+        /* Weighted rate */
+        sum1_Q14 = silk_SMULBB( mu_Q9, cl_Q5[ k ] );
+
+        /* Penalty for too large gain */
+        sum1_Q14 = silk_ADD_LSHIFT32( sum1_Q14, silk_max( silk_SUB32( gain_tmp_Q7, max_gain_Q7 ), 0 ), 10 );
+
+        silk_assert( sum1_Q14 >= 0 );
+
+        /* first row of W_Q18 */
+        C_tmp3 = _mm_loadu_si128( (__m128i *)(&W_Q18[ 1 ] ) );
+        C_tmp4 = _mm_mul_epi32( C_tmp3, C_tmp1 );
+        C_tmp4 = _mm_srli_si128( C_tmp4, 2 );
+
+        C_tmp1 = _mm_shuffle_epi32( C_tmp1, _MM_SHUFFLE( 0, 3, 2, 1 ) ); /* shift right 4 bytes */
+        C_tmp3 = _mm_shuffle_epi32( C_tmp3, _MM_SHUFFLE( 0, 3, 2, 1 ) ); /* shift right 4 bytes */
+
+        C_tmp5 = _mm_mul_epi32( C_tmp3, C_tmp1 );
+        C_tmp5 = _mm_srli_si128( C_tmp5, 2 );
+
+        C_tmp5 = _mm_add_epi32( C_tmp4, C_tmp5 );
+        C_tmp5 = _mm_slli_epi32( C_tmp5, 1 );
+
+        C_tmp5 = _mm_add_epi32( C_tmp5, _mm_shuffle_epi32( C_tmp5, _MM_SHUFFLE( 0, 0, 0, 2 ) ) );
+        sum2_Q16 = _mm_cvtsi128_si32( C_tmp5 );
+
+        sum2_Q16 = silk_SMLAWB( sum2_Q16, W_Q18[  0 ], diff_Q14[ 0 ] );
+        sum1_Q14 = silk_SMLAWB( sum1_Q14, sum2_Q16,    diff_Q14[ 0 ] );
+
+        /* second row of W_Q18 */
+        sum2_Q16 = silk_SMULWB(           W_Q18[  7 ], diff_Q14[ 2 ] );
+        sum2_Q16 = silk_SMLAWB( sum2_Q16, W_Q18[  8 ], diff_Q14[ 3 ] );
+        sum2_Q16 = silk_SMLAWB( sum2_Q16, W_Q18[  9 ], diff_Q14[ 4 ] );
+        sum2_Q16 = silk_LSHIFT( sum2_Q16, 1 );
+        sum2_Q16 = silk_SMLAWB( sum2_Q16, W_Q18[  6 ], diff_Q14[ 1 ] );
+        sum1_Q14 = silk_SMLAWB( sum1_Q14, sum2_Q16,    diff_Q14[ 1 ] );
+
+        /* third row of W_Q18 */
+        sum2_Q16 = silk_SMULWB(           W_Q18[ 13 ], diff_Q14[ 3 ] );
+        sum2_Q16 = silk_SMLAWB( sum2_Q16, W_Q18[ 14 ], diff_Q14[ 4 ] );
+        sum2_Q16 = silk_LSHIFT( sum2_Q16, 1 );
+        sum2_Q16 = silk_SMLAWB( sum2_Q16, W_Q18[ 12 ], diff_Q14[ 2 ] );
+        sum1_Q14 = silk_SMLAWB( sum1_Q14, sum2_Q16,    diff_Q14[ 2 ] );
+
+        /* fourth row of W_Q18 */
+        sum2_Q16 = silk_SMULWB(           W_Q18[ 19 ], diff_Q14[ 4 ] );
+        sum2_Q16 = silk_LSHIFT( sum2_Q16, 1 );
+        sum2_Q16 = silk_SMLAWB( sum2_Q16, W_Q18[ 18 ], diff_Q14[ 3 ] );
+        sum1_Q14 = silk_SMLAWB( sum1_Q14, sum2_Q16,    diff_Q14[ 3 ] );
+
+        /* last row of W_Q18 */
+        sum2_Q16 = silk_SMULWB(           W_Q18[ 24 ], diff_Q14[ 4 ] );
+        sum1_Q14 = silk_SMLAWB( sum1_Q14, sum2_Q16,    diff_Q14[ 4 ] );
+
+        silk_assert( sum1_Q14 >= 0 );
+
+        /* find best */
+        if( sum1_Q14 < *rate_dist_Q14 ) {
+            *rate_dist_Q14 = sum1_Q14;
+            *ind = (opus_int8)k;
+            *gain_Q7 = gain_tmp_Q7;
+        }
+
+        /* Go to next cbk vector */
+        cb_row_Q7 += LTP_ORDER;
+    }
+}
diff --git a/media/libopus/silk/x86/main_sse.h b/media/libopus/silk/x86/main_sse.h
new file mode 100644
index 000000000..d8d61310e
--- /dev/null
+++ b/media/libopus/silk/x86/main_sse.h
@@ -0,0 +1,277 @@
+/* Copyright (c) 2014, Cisco Systems, INC
+   Written by XiangMingZhu WeiZhou MinPeng YanWang
+
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions
+   are met:
+
+   - Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+
+   - Redistributions in binary form must reproduce the above copyright
+   notice, this list of conditions and the following disclaimer in the
+   documentation and/or other materials provided with the distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
+   OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+   EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+   PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+   PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+   LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+   NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+   SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+
+#ifndef MAIN_SSE_H
+#define MAIN_SSE_H
+
+#ifdef HAVE_CONFIG_H
+#include "config.h"
+#endif
+
+# if defined(OPUS_X86_MAY_HAVE_SSE4_1)
+
+#  define OVERRIDE_silk_VQ_WMat_EC
+
+void silk_VQ_WMat_EC_sse4_1(
+    opus_int8                   *ind,                           /* O    index of best codebook vector               */
+    opus_int32                  *rate_dist_Q14,                 /* O    best weighted quant error + mu * rate       */
+    opus_int                    *gain_Q7,                       /* O    sum of absolute LTP coefficients            */
+    const opus_int16            *in_Q14,                        /* I    input vector to be quantized                */
+    const opus_int32            *W_Q18,                         /* I    weighting matrix                            */
+    const opus_int8             *cb_Q7,                         /* I    codebook                                    */
+    const opus_uint8            *cb_gain_Q7,                    /* I    codebook effective gain                     */
+    const opus_uint8            *cl_Q5,                         /* I    code length for each codebook vector        */
+    const opus_int              mu_Q9,                          /* I    tradeoff betw. weighted error and rate      */
+    const opus_int32            max_gain_Q7,                    /* I    maximum sum of absolute LTP coefficients    */
+    opus_int                    L                               /* I    number of vectors in codebook               */
+);
+
+#if defined OPUS_X86_PRESUME_SSE4_1
+
+#define silk_VQ_WMat_EC(ind, rate_dist_Q14, gain_Q7, in_Q14, W_Q18, cb_Q7, cb_gain_Q7, cl_Q5, \
+                          mu_Q9, max_gain_Q7, L, arch) \
+    ((void)(arch),silk_VQ_WMat_EC_sse4_1(ind, rate_dist_Q14, gain_Q7, in_Q14, W_Q18, cb_Q7, cb_gain_Q7, cl_Q5, \
+                          mu_Q9, max_gain_Q7, L))
+
+#else
+
+extern void (*const SILK_VQ_WMAT_EC_IMPL[OPUS_ARCHMASK + 1])(
+    opus_int8                   *ind,                           /* O    index of best codebook vector               */
+    opus_int32                  *rate_dist_Q14,                 /* O    best weighted quant error + mu * rate       */
+    opus_int                    *gain_Q7,                       /* O    sum of absolute LTP coefficients            */
+    const opus_int16            *in_Q14,                        /* I    input vector to be quantized                */
+    const opus_int32            *W_Q18,                         /* I    weighting matrix                            */
+    const opus_int8             *cb_Q7,                         /* I    codebook                                    */
+    const opus_uint8            *cb_gain_Q7,                    /* I    codebook effective gain                     */
+    const opus_uint8            *cl_Q5,                         /* I    code length for each codebook vector        */
+    const opus_int              mu_Q9,                          /* I    tradeoff betw. weighted error and rate      */
+    const opus_int32            max_gain_Q7,                    /* I    maximum sum of absolute LTP coefficients    */
+    opus_int                    L                               /* I    number of vectors in codebook               */
+);
+
+#  define silk_VQ_WMat_EC(ind, rate_dist_Q14, gain_Q7, in_Q14, W_Q18, cb_Q7, cb_gain_Q7, cl_Q5, \
+                          mu_Q9, max_gain_Q7, L, arch) \
+    ((*SILK_VQ_WMAT_EC_IMPL[(arch) & OPUS_ARCHMASK])(ind, rate_dist_Q14, gain_Q7, in_Q14, W_Q18, cb_Q7, cb_gain_Q7, cl_Q5, \
+                          mu_Q9, max_gain_Q7, L))
+
+#endif
+
+#  define OVERRIDE_silk_NSQ
+
+void silk_NSQ_sse4_1(
+    const silk_encoder_state    *psEncC,                                    /* I/O  Encoder State                   */
+    silk_nsq_state              *NSQ,                                       /* I/O  NSQ state                       */
+    SideInfoIndices             *psIndices,                                 /* I/O  Quantization Indices            */
+    const opus_int32            x_Q3[],                                     /* I    Prefiltered input signal        */
+    opus_int8                   pulses[],                                   /* O    Quantized pulse signal          */
+    const opus_int16            PredCoef_Q12[ 2 * MAX_LPC_ORDER ],          /* I    Short term prediction coefs     */
+    const opus_int16            LTPCoef_Q14[ LTP_ORDER * MAX_NB_SUBFR ],    /* I    Long term prediction coefs      */
+    const opus_int16            AR2_Q13[ MAX_NB_SUBFR * MAX_SHAPE_LPC_ORDER ], /* I Noise shaping coefs             */
+    const opus_int              HarmShapeGain_Q14[ MAX_NB_SUBFR ],          /* I    Long term shaping coefs         */
+    const opus_int              Tilt_Q14[ MAX_NB_SUBFR ],                   /* I    Spectral tilt                   */
+    const opus_int32            LF_shp_Q14[ MAX_NB_SUBFR ],                 /* I    Low frequency shaping coefs     */
+    const opus_int32            Gains_Q16[ MAX_NB_SUBFR ],                  /* I    Quantization step sizes         */
+    const opus_int              pitchL[ MAX_NB_SUBFR ],                     /* I    Pitch lags                      */
+    const opus_int              Lambda_Q10,                                 /* I    Rate/distortion tradeoff        */
+    const opus_int              LTP_scale_Q14                               /* I    LTP state scaling               */
+);
+
+#if defined OPUS_X86_PRESUME_SSE4_1
+
+#define silk_NSQ(psEncC, NSQ, psIndices, x_Q3, pulses, PredCoef_Q12, LTPCoef_Q14, AR2_Q13, \
+                   HarmShapeGain_Q14, Tilt_Q14, LF_shp_Q14, Gains_Q16, pitchL, Lambda_Q10, LTP_scale_Q14, arch) \
+    ((void)(arch),silk_NSQ_sse4_1(psEncC, NSQ, psIndices, x_Q3, pulses, PredCoef_Q12, LTPCoef_Q14, AR2_Q13, \
+                   HarmShapeGain_Q14, Tilt_Q14, LF_shp_Q14, Gains_Q16, pitchL, Lambda_Q10, LTP_scale_Q14))
+
+#else
+
+extern void (*const SILK_NSQ_IMPL[OPUS_ARCHMASK + 1])(
+    const silk_encoder_state    *psEncC,                                    /* I/O  Encoder State                   */
+    silk_nsq_state              *NSQ,                                       /* I/O  NSQ state                       */
+    SideInfoIndices             *psIndices,                                 /* I/O  Quantization Indices            */
+    const opus_int32            x_Q3[],                                     /* I    Prefiltered input signal        */
+    opus_int8                   pulses[],                                   /* O    Quantized pulse signal          */
+    const opus_int16            PredCoef_Q12[ 2 * MAX_LPC_ORDER ],          /* I    Short term prediction coefs     */
+    const opus_int16            LTPCoef_Q14[ LTP_ORDER * MAX_NB_SUBFR ],    /* I    Long term prediction coefs      */
+    const opus_int16            AR2_Q13[ MAX_NB_SUBFR * MAX_SHAPE_LPC_ORDER ], /* I Noise shaping coefs             */
+    const opus_int              HarmShapeGain_Q14[ MAX_NB_SUBFR ],          /* I    Long term shaping coefs         */
+    const opus_int              Tilt_Q14[ MAX_NB_SUBFR ],                   /* I    Spectral tilt                   */
+    const opus_int32            LF_shp_Q14[ MAX_NB_SUBFR ],                 /* I    Low frequency shaping coefs     */
+    const opus_int32            Gains_Q16[ MAX_NB_SUBFR ],                  /* I    Quantization step sizes         */
+    const opus_int              pitchL[ MAX_NB_SUBFR ],                     /* I    Pitch lags                      */
+    const opus_int              Lambda_Q10,                                 /* I    Rate/distortion tradeoff        */
+    const opus_int              LTP_scale_Q14                               /* I    LTP state scaling               */
+);
+
+#  define silk_NSQ(psEncC, NSQ, psIndices, x_Q3, pulses, PredCoef_Q12, LTPCoef_Q14, AR2_Q13, \
+                   HarmShapeGain_Q14, Tilt_Q14, LF_shp_Q14, Gains_Q16, pitchL, Lambda_Q10, LTP_scale_Q14, arch) \
+    ((*SILK_NSQ_IMPL[(arch) & OPUS_ARCHMASK])(psEncC, NSQ, psIndices, x_Q3, pulses, PredCoef_Q12, LTPCoef_Q14, AR2_Q13, \
+                   HarmShapeGain_Q14, Tilt_Q14, LF_shp_Q14, Gains_Q16, pitchL, Lambda_Q10, LTP_scale_Q14))
+
+#endif
+
+#  define OVERRIDE_silk_NSQ_del_dec
+
+void silk_NSQ_del_dec_sse4_1(
+    const silk_encoder_state    *psEncC,                                    /* I/O  Encoder State                   */
+    silk_nsq_state              *NSQ,                                       /* I/O  NSQ state                       */
+    SideInfoIndices             *psIndices,                                 /* I/O  Quantization Indices            */
+    const opus_int32            x_Q3[],                                     /* I    Prefiltered input signal        */
+    opus_int8                   pulses[],                                   /* O    Quantized pulse signal          */
+    const opus_int16            PredCoef_Q12[ 2 * MAX_LPC_ORDER ],          /* I    Short term prediction coefs     */
+    const opus_int16            LTPCoef_Q14[ LTP_ORDER * MAX_NB_SUBFR ],    /* I    Long term prediction coefs      */
+    const opus_int16            AR2_Q13[ MAX_NB_SUBFR * MAX_SHAPE_LPC_ORDER ], /* I Noise shaping coefs             */
+    const opus_int              HarmShapeGain_Q14[ MAX_NB_SUBFR ],          /* I    Long term shaping coefs         */
+    const opus_int              Tilt_Q14[ MAX_NB_SUBFR ],                   /* I    Spectral tilt                   */
+    const opus_int32            LF_shp_Q14[ MAX_NB_SUBFR ],                 /* I    Low frequency shaping coefs     */
+    const opus_int32            Gains_Q16[ MAX_NB_SUBFR ],                  /* I    Quantization step sizes         */
+    const opus_int              pitchL[ MAX_NB_SUBFR ],                     /* I    Pitch lags                      */
+    const opus_int              Lambda_Q10,                                 /* I    Rate/distortion tradeoff        */
+    const opus_int              LTP_scale_Q14                               /* I    LTP state scaling               */
+);
+
+#if defined OPUS_X86_PRESUME_SSE4_1
+
+#define silk_NSQ_del_dec(psEncC, NSQ, psIndices, x_Q3, pulses, PredCoef_Q12, LTPCoef_Q14, AR2_Q13, \
+                           HarmShapeGain_Q14, Tilt_Q14, LF_shp_Q14, Gains_Q16, pitchL, Lambda_Q10, LTP_scale_Q14, arch) \
+    ((void)(arch),silk_NSQ_del_dec_sse4_1(psEncC, NSQ, psIndices, x_Q3, pulses, PredCoef_Q12, LTPCoef_Q14, AR2_Q13, \
+                           HarmShapeGain_Q14, Tilt_Q14, LF_shp_Q14, Gains_Q16, pitchL, Lambda_Q10, LTP_scale_Q14))
+
+#else
+
+extern void (*const SILK_NSQ_DEL_DEC_IMPL[OPUS_ARCHMASK + 1])(
+    const silk_encoder_state    *psEncC,                                    /* I/O  Encoder State                   */
+    silk_nsq_state              *NSQ,                                       /* I/O  NSQ state                       */
+    SideInfoIndices             *psIndices,                                 /* I/O  Quantization Indices            */
+    const opus_int32            x_Q3[],                                     /* I    Prefiltered input signal        */
+    opus_int8                   pulses[],                                   /* O    Quantized pulse signal          */
+    const opus_int16            PredCoef_Q12[ 2 * MAX_LPC_ORDER ],          /* I    Short term prediction coefs     */
+    const opus_int16            LTPCoef_Q14[ LTP_ORDER * MAX_NB_SUBFR ],    /* I    Long term prediction coefs      */
+    const opus_int16            AR2_Q13[ MAX_NB_SUBFR * MAX_SHAPE_LPC_ORDER ], /* I Noise shaping coefs             */
+    const opus_int              HarmShapeGain_Q14[ MAX_NB_SUBFR ],          /* I    Long term shaping coefs         */
+    const opus_int              Tilt_Q14[ MAX_NB_SUBFR ],                   /* I    Spectral tilt                   */
+    const opus_int32            LF_shp_Q14[ MAX_NB_SUBFR ],                 /* I    Low frequency shaping coefs     */
+    const opus_int32            Gains_Q16[ MAX_NB_SUBFR ],                  /* I    Quantization step sizes         */
+    const opus_int              pitchL[ MAX_NB_SUBFR ],                     /* I    Pitch lags                      */
+    const opus_int              Lambda_Q10,                                 /* I    Rate/distortion tradeoff        */
+    const opus_int              LTP_scale_Q14                               /* I    LTP state scaling               */
+);
+
+#  define silk_NSQ_del_dec(psEncC, NSQ, psIndices, x_Q3, pulses, PredCoef_Q12, LTPCoef_Q14, AR2_Q13, \
+                           HarmShapeGain_Q14, Tilt_Q14, LF_shp_Q14, Gains_Q16, pitchL, Lambda_Q10, LTP_scale_Q14, arch) \
+    ((*SILK_NSQ_DEL_DEC_IMPL[(arch) & OPUS_ARCHMASK])(psEncC, NSQ, psIndices, x_Q3, pulses, PredCoef_Q12, LTPCoef_Q14, AR2_Q13, \
+                           HarmShapeGain_Q14, Tilt_Q14, LF_shp_Q14, Gains_Q16, pitchL, Lambda_Q10, LTP_scale_Q14))
+
+#endif
+
+void silk_noise_shape_quantizer(
+    silk_nsq_state      *NSQ,                   /* I/O  NSQ state                       */
+    opus_int            signalType,             /* I    Signal type                     */
+    const opus_int32    x_sc_Q10[],             /* I                                    */
+    opus_int8           pulses[],               /* O                                    */
+    opus_int16          xq[],                   /* O                                    */
+    opus_int32          sLTP_Q15[],             /* I/O  LTP state                       */
+    const opus_int16    a_Q12[],                /* I    Short term prediction coefs     */
+    const opus_int16    b_Q14[],                /* I    Long term prediction coefs      */
+    const opus_int16    AR_shp_Q13[],           /* I    Noise shaping AR coefs          */
+    opus_int            lag,                    /* I    Pitch lag                       */
+    opus_int32          HarmShapeFIRPacked_Q14, /* I                                    */
+    opus_int            Tilt_Q14,               /* I    Spectral tilt                   */
+    opus_int32          LF_shp_Q14,             /* I                                    */
+    opus_int32          Gain_Q16,               /* I                                    */
+    opus_int            Lambda_Q10,             /* I                                    */
+    opus_int            offset_Q10,             /* I                                    */
+    opus_int            length,                 /* I    Input length                    */
+    opus_int            shapingLPCOrder,        /* I    Noise shaping AR filter order   */
+    opus_int            predictLPCOrder,        /* I    Prediction filter order         */
+    int                 arch                    /* I    Architecture                    */
+);
+
+/**************************/
+/* Noise level estimation */
+/**************************/
+void silk_VAD_GetNoiseLevels(
+    const opus_int32            pX[ VAD_N_BANDS ],  /* I    subband energies                            */
+    silk_VAD_state              *psSilk_VAD         /* I/O  Pointer to Silk VAD state                   */
+);
+
+#  define OVERRIDE_silk_VAD_GetSA_Q8
+
+opus_int silk_VAD_GetSA_Q8_sse4_1(
+    silk_encoder_state *psEnC,
+    const opus_int16   pIn[]
+);
+
+#if defined(OPUS_X86_PRESUME_SSE4_1)
+#define silk_VAD_GetSA_Q8(psEnC, pIn, arch) ((void)(arch),silk_VAD_GetSA_Q8_sse4_1(psEnC, pIn))
+
+#else
+
+#  define silk_VAD_GetSA_Q8(psEnC, pIn, arch) \
+     ((*SILK_VAD_GETSA_Q8_IMPL[(arch) & OPUS_ARCHMASK])(psEnC, pIn))
+
+extern opus_int (*const SILK_VAD_GETSA_Q8_IMPL[OPUS_ARCHMASK + 1])(
+     silk_encoder_state *psEnC,
+     const opus_int16   pIn[]);
+
+#  define OVERRIDE_silk_warped_LPC_analysis_filter_FIX
+
+#endif
+
+void silk_warped_LPC_analysis_filter_FIX_sse4_1(
+          opus_int32            state[],                    /* I/O  State [order + 1]                   */
+          opus_int32            res_Q2[],                   /* 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)                 */
+);
+
+#if defined(OPUS_X86_PRESUME_SSE4_1)
+#define silk_warped_LPC_analysis_filter_FIX(state, res_Q2, coef_Q13, input, lambda_Q16, length, order, arch) \
+    ((void)(arch),silk_warped_LPC_analysis_filter_FIX_c(state, res_Q2, coef_Q13, input, lambda_Q16, length, order))
+
+#else
+
+extern void (*const SILK_WARPED_LPC_ANALYSIS_FILTER_FIX_IMPL[OPUS_ARCHMASK + 1])(
+          opus_int32            state[],                    /* I/O  State [order + 1]                   */
+          opus_int32            res_Q2[],                   /* 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)                 */
+);
+
+#  define silk_warped_LPC_analysis_filter_FIX(state, res_Q2, coef_Q13, input, lambda_Q16, length, order, arch) \
+    ((*SILK_WARPED_LPC_ANALYSIS_FILTER_FIX_IMPL[(arch) & OPUS_ARCHMASK])(state, res_Q2, coef_Q13, input, lambda_Q16, length, order))
+
+#endif
+
+# endif
+#endif
diff --git a/media/libopus/silk/x86/x86_silk_map.c b/media/libopus/silk/x86/x86_silk_map.c
new file mode 100644
index 000000000..818841f2c
--- /dev/null
+++ b/media/libopus/silk/x86/x86_silk_map.c
@@ -0,0 +1,174 @@
+/* Copyright (c) 2014, Cisco Systems, INC
+   Written by XiangMingZhu WeiZhou MinPeng YanWang
+
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions
+   are met:
+
+   - Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+
+   - Redistributions in binary form must reproduce the above copyright
+   notice, this list of conditions and the following disclaimer in the
+   documentation and/or other materials provided with the distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
+   OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+   EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+   PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+   PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+   LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+   NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+   SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+
+#if defined(HAVE_CONFIG_H)
+#include "config.h"
+#endif
+
+#include "celt/x86/x86cpu.h"
+#include "structs.h"
+#include "SigProc_FIX.h"
+#include "pitch.h"
+#include "main.h"
+
+#if !defined(OPUS_X86_PRESUME_SSE4_1)
+
+#if defined(FIXED_POINT)
+
+#include "fixed/main_FIX.h"
+
+opus_int64 (*const SILK_INNER_PROD16_ALIGNED_64_IMPL[ OPUS_ARCHMASK + 1 ] )(
+    const opus_int16 *inVec1,
+    const opus_int16 *inVec2,
+    const opus_int   len
+) = {
+  silk_inner_prod16_aligned_64_c,                  /* non-sse */
+  silk_inner_prod16_aligned_64_c,
+  silk_inner_prod16_aligned_64_c,
+  MAY_HAVE_SSE4_1( silk_inner_prod16_aligned_64 ), /* sse4.1 */
+  MAY_HAVE_SSE4_1( silk_inner_prod16_aligned_64 )  /* avx */
+};
+
+#endif
+
+opus_int (*const SILK_VAD_GETSA_Q8_IMPL[ OPUS_ARCHMASK + 1 ] )(
+    silk_encoder_state *psEncC,
+    const opus_int16   pIn[]
+) = {
+  silk_VAD_GetSA_Q8_c,                  /* non-sse */
+  silk_VAD_GetSA_Q8_c,
+  silk_VAD_GetSA_Q8_c,
+  MAY_HAVE_SSE4_1( silk_VAD_GetSA_Q8 ), /* sse4.1 */
+  MAY_HAVE_SSE4_1( silk_VAD_GetSA_Q8 )  /* avx */
+};
+
+void (*const SILK_NSQ_IMPL[ OPUS_ARCHMASK + 1 ] )(
+    const silk_encoder_state    *psEncC,                                    /* I/O  Encoder State                   */
+    silk_nsq_state              *NSQ,                                       /* I/O  NSQ state                       */
+    SideInfoIndices             *psIndices,                                 /* I/O  Quantization Indices            */
+    const opus_int32            x_Q3[],                                     /* I    Prefiltered input signal        */
+    opus_int8                   pulses[],                                   /* O    Quantized pulse signal          */
+    const opus_int16            PredCoef_Q12[ 2 * MAX_LPC_ORDER ],          /* I    Short term prediction coefs     */
+    const opus_int16            LTPCoef_Q14[ LTP_ORDER * MAX_NB_SUBFR ],    /* I    Long term prediction coefs      */
+    const opus_int16            AR2_Q13[ MAX_NB_SUBFR * MAX_SHAPE_LPC_ORDER ], /* I Noise shaping coefs             */
+    const opus_int              HarmShapeGain_Q14[ MAX_NB_SUBFR ],          /* I    Long term shaping coefs         */
+    const opus_int              Tilt_Q14[ MAX_NB_SUBFR ],                   /* I    Spectral tilt                   */
+    const opus_int32            LF_shp_Q14[ MAX_NB_SUBFR ],                 /* I    Low frequency shaping coefs     */
+    const opus_int32            Gains_Q16[ MAX_NB_SUBFR ],                  /* I    Quantization step sizes         */
+    const opus_int              pitchL[ MAX_NB_SUBFR ],                     /* I    Pitch lags                      */
+    const opus_int              Lambda_Q10,                                 /* I    Rate/distortion tradeoff        */
+    const opus_int              LTP_scale_Q14                               /* I    LTP state scaling               */
+) = {
+  silk_NSQ_c,                  /* non-sse */
+  silk_NSQ_c,
+  silk_NSQ_c,
+  MAY_HAVE_SSE4_1( silk_NSQ ), /* sse4.1 */
+  MAY_HAVE_SSE4_1( silk_NSQ )  /* avx */
+};
+
+void (*const SILK_VQ_WMAT_EC_IMPL[ OPUS_ARCHMASK + 1 ] )(
+    opus_int8                   *ind,                           /* O    index of best codebook vector               */
+    opus_int32                  *rate_dist_Q14,                 /* O    best weighted quant error + mu * rate       */
+    opus_int                    *gain_Q7,                       /* O    sum of absolute LTP coefficients            */
+    const opus_int16            *in_Q14,                        /* I    input vector to be quantized                */
+    const opus_int32            *W_Q18,                         /* I    weighting matrix                            */
+    const opus_int8             *cb_Q7,                         /* I    codebook                                    */
+    const opus_uint8            *cb_gain_Q7,                    /* I    codebook effective gain                     */
+    const opus_uint8            *cl_Q5,                         /* I    code length for each codebook vector        */
+    const opus_int              mu_Q9,                          /* I    tradeoff betw. weighted error and rate      */
+    const opus_int32            max_gain_Q7,                    /* I    maximum sum of absolute LTP coefficients    */
+    opus_int                    L                               /* I    number of vectors in codebook               */
+) = {
+  silk_VQ_WMat_EC_c,                  /* non-sse */
+  silk_VQ_WMat_EC_c,
+  silk_VQ_WMat_EC_c,
+  MAY_HAVE_SSE4_1( silk_VQ_WMat_EC ), /* sse4.1 */
+  MAY_HAVE_SSE4_1( silk_VQ_WMat_EC )  /* avx */
+};
+
+void (*const SILK_NSQ_DEL_DEC_IMPL[ OPUS_ARCHMASK + 1 ] )(
+    const silk_encoder_state    *psEncC,                                    /* I/O  Encoder State                   */
+    silk_nsq_state              *NSQ,                                       /* I/O  NSQ state                       */
+    SideInfoIndices             *psIndices,                                 /* I/O  Quantization Indices            */
+    const opus_int32            x_Q3[],                                     /* I    Prefiltered input signal        */
+    opus_int8                   pulses[],                                   /* O    Quantized pulse signal          */
+    const opus_int16            PredCoef_Q12[ 2 * MAX_LPC_ORDER ],          /* I    Short term prediction coefs     */
+    const opus_int16            LTPCoef_Q14[ LTP_ORDER * MAX_NB_SUBFR ],    /* I    Long term prediction coefs      */
+    const opus_int16            AR2_Q13[ MAX_NB_SUBFR * MAX_SHAPE_LPC_ORDER ], /* I Noise shaping coefs             */
+    const opus_int              HarmShapeGain_Q14[ MAX_NB_SUBFR ],          /* I    Long term shaping coefs         */
+    const opus_int              Tilt_Q14[ MAX_NB_SUBFR ],                   /* I    Spectral tilt                   */
+    const opus_int32            LF_shp_Q14[ MAX_NB_SUBFR ],                 /* I    Low frequency shaping coefs     */
+    const opus_int32            Gains_Q16[ MAX_NB_SUBFR ],                  /* I    Quantization step sizes         */
+    const opus_int              pitchL[ MAX_NB_SUBFR ],                     /* I    Pitch lags                      */
+    const opus_int              Lambda_Q10,                                 /* I    Rate/distortion tradeoff        */
+    const opus_int              LTP_scale_Q14                               /* I    LTP state scaling               */
+) = {
+  silk_NSQ_del_dec_c,                  /* non-sse */
+  silk_NSQ_del_dec_c,
+  silk_NSQ_del_dec_c,
+  MAY_HAVE_SSE4_1( silk_NSQ_del_dec ), /* sse4.1 */
+  MAY_HAVE_SSE4_1( silk_NSQ_del_dec )  /* avx */
+};
+
+#if defined(FIXED_POINT)
+
+void (*const SILK_WARPED_LPC_ANALYSIS_FILTER_FIX_IMPL[ OPUS_ARCHMASK + 1 ] )(
+    opus_int32                  state[],                    /* I/O  State [order + 1]                   */
+    opus_int32                  res_Q2[],                   /* 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)                 */
+) = {
+  silk_warped_LPC_analysis_filter_FIX_c,                  /* non-sse */
+  silk_warped_LPC_analysis_filter_FIX_c,
+  silk_warped_LPC_analysis_filter_FIX_c,
+  MAY_HAVE_SSE4_1( silk_warped_LPC_analysis_filter_FIX ), /* sse4.1 */
+  MAY_HAVE_SSE4_1( silk_warped_LPC_analysis_filter_FIX )  /* avx */
+};
+
+void (*const SILK_BURG_MODIFIED_IMPL[ OPUS_ARCHMASK + 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                                       */
+) = {
+  silk_burg_modified_c,                  /* non-sse */
+  silk_burg_modified_c,
+  silk_burg_modified_c,
+  MAY_HAVE_SSE4_1( silk_burg_modified ), /* sse4.1 */
+  MAY_HAVE_SSE4_1( silk_burg_modified )  /* avx */
+};
+
+#endif
+#endif