[opus] [PATCH 8/8] Optimize silk_NSQ_del_dec() for ARM NEON

Linfeng Zhang linfengz at google.com
Tue Aug 23 16:10:37 UTC 2016


Created corresponding unit test, and the optimization is bit exact with C
function.

This optimization speeds up SILK encoder on NEON as following.

Fixed-point:
Complexity 0-5:  0%
Complexity 6-7:  6%
Complexity 8-9: 10%
Complexity  10:  8%

Got similar results on floating-point.
---
 silk/NSQ_del_dec.c                              |    6 +-
 silk/SigProc_FIX.h                              |    4 +-
 silk/arm/NSQ_del_dec_arm.h                      |   88 ++
 silk/arm/NSQ_del_dec_neon_intr.c                | 1125 +++++++++++++++++++++++
 silk/arm/arm_silk_map.c                         |   23 +
 silk/main.h                                     |    6 +-
 silk/mips/NSQ_del_dec_mipsr1.h                  |    2 +-
 silk/tests/test_unit_optimization_NSQ_del_dec.c |  142 +++
 silk/x86/NSQ_del_dec_sse.c                      |    6 +-
 silk/x86/main_sse.h                             |    4 +-
 silk/x86/x86_silk_map.c                         |    2 +-
 silk_sources.mk                                 |    1 +
 tests/test_unit_optimization.c                  |    2 +
 13 files changed, 1399 insertions(+), 12 deletions(-)
 create mode 100644 silk/arm/NSQ_del_dec_arm.h
 create mode 100644 silk/arm/NSQ_del_dec_neon_intr.c
 create mode 100644 silk/tests/test_unit_optimization_NSQ_del_dec.c

diff --git a/silk/NSQ_del_dec.c b/silk/NSQ_del_dec.c
index 3495613..20640a4 100644
--- a/silk/NSQ_del_dec.c
+++ b/silk/NSQ_del_dec.c
@@ -109,13 +109,13 @@ static OPUS_INLINE void silk_noise_shape_quantizer_del_dec(
     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            *smpl_buf_idx,          /* I/O  Index to newest samples in buffers  */
     opus_int            decisionDelay,          /* I                                        */
     int                 arch                    /* I                                        */
 );
 
 void silk_NSQ_del_dec_c(
-    const silk_encoder_state    *psEncC,                                    /* I/O  Encoder State                   */
+    const silk_encoder_state    *psEncC,                                    /* I    Encoder State                   */
     silk_nsq_state              *NSQ,                                       /* I/O  NSQ state                       */
     SideInfoIndices             *psIndices,                                 /* I/O  Quantization Indices            */
     const opus_int16            x16[],                                        /* I    Input                           */
@@ -341,7 +341,7 @@ static OPUS_INLINE void silk_noise_shape_quantizer_del_dec(
     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            *smpl_buf_idx,          /* I/O  Index to newest samples in buffers  */
     opus_int            decisionDelay,          /* I                                        */
     int                 arch                    /* I                                        */
 )
diff --git a/silk/SigProc_FIX.h b/silk/SigProc_FIX.h
index 0e619d0..4f0a09e 100644
--- a/silk/SigProc_FIX.h
+++ b/silk/SigProc_FIX.h
@@ -599,7 +599,9 @@ static OPUS_INLINE opus_int64 silk_max_64(opus_int64 a, opus_int64 b)
 /* Make sure to store the result as the seed for the next call (also in between     */
 /* frames), otherwise result won't be random at all. When only using some of the    */
 /* bits, take the most significant bits by right-shifting.                          */
-#define silk_RAND(seed)                     (silk_MLA_ovflw(907633515, (seed), 196314165))
+#define RAND_MULTIPLIER                     196314165
+#define RAND_INCREMENT                      907633515
+#define silk_RAND(seed)                     (silk_MLA_ovflw((RAND_INCREMENT), (seed), (RAND_MULTIPLIER)))
 
 /*  Add some multiplication functions that can be easily mapped to ARM. */
 
diff --git a/silk/arm/NSQ_del_dec_arm.h b/silk/arm/NSQ_del_dec_arm.h
new file mode 100644
index 0000000..93581e1
--- /dev/null
+++ b/silk/arm/NSQ_del_dec_arm.h
@@ -0,0 +1,88 @@
+/* Copyright (c) 2016 Google Inc. */
+/*
+   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(NSQ_DEL_DEC_ARM_H)
+# define NSQ_DEL_DEC_ARM_H
+
+# include "celt/arm/armcpu.h"
+
+# if (defined(OPUS_ARM_MAY_HAVE_NEON_INTR))
+void silk_NSQ_del_dec_neon(
+    const silk_encoder_state    *psEncC,                                    /* I    Encoder State                   */
+    silk_nsq_state              *NSQ,                                       /* I/O  NSQ state                       */
+    SideInfoIndices             *psIndices,                                 /* I/O  Quantization Indices            */
+    const opus_int16            x16[],                                      /* I    Input                           */
+    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            AR_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               */
+);
+# endif
+
+#  if !defined(OPUS_HAVE_RTCD)
+#   define OVERRIDE_silk_NSQ_del_dec      (1)
+#   define silk_NSQ_del_dec(psEncC, NSQ, psIndices, x16, pulses, PredCoef_Q12, LTPCoef_Q14, AR_Q13, HarmShapeGain_Q14, Tilt_Q14, LF_shp_Q14, Gains_Q16, pitchL, Lambda_Q10, LTP_scale_Q14, arch) \
+    ((void)(arch),PRESUME_NEON(silk_NSQ_del_dec)(psEncC, NSQ, psIndices, x16, pulses, PredCoef_Q12, LTPCoef_Q14, AR_Q13, HarmShapeGain_Q14, Tilt_Q14, LF_shp_Q14, Gains_Q16, pitchL, Lambda_Q10, LTP_scale_Q14))
+#  endif
+
+#  if !defined(OVERRIDE_silk_NSQ_del_dec)
+/*Is run-time CPU detection enabled on this platform?*/
+#   if defined(OPUS_HAVE_RTCD) && (defined(OPUS_ARM_MAY_HAVE_NEON_INTR) && !defined(OPUS_ARM_PRESUME_NEON_INTR))
+extern void (*const SILK_NSQ_DEL_DEC_IMPL[OPUS_ARCHMASK+1])(
+        const silk_encoder_state    *psEncC,                                    /* I    Encoder State                   */
+        silk_nsq_state              *NSQ,                                       /* I/O  NSQ state                       */
+        SideInfoIndices             *psIndices,                                 /* I/O  Quantization Indices            */
+        const opus_int16            x16[],                                      /* I    Input                           */
+        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            AR_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 OVERRIDE_silk_NSQ_del_dec      (1)
+#    define silk_NSQ_del_dec(psEncC, NSQ, psIndices, x16, pulses, PredCoef_Q12, LTPCoef_Q14, AR_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, x16, pulses, PredCoef_Q12, LTPCoef_Q14, AR_Q13, HarmShapeGain_Q14, Tilt_Q14, LF_shp_Q14, Gains_Q16, pitchL, Lambda_Q10, LTP_scale_Q14))
+#   elif defined(OPUS_ARM_PRESUME_NEON_INTR)
+#    define OVERRIDE_silk_NSQ_del_dec      (1)
+#    define silk_NSQ_del_dec(psEncC, NSQ, psIndices, x16, pulses, PredCoef_Q12, LTPCoef_Q14, AR_Q13, HarmShapeGain_Q14, Tilt_Q14, LF_shp_Q14, Gains_Q16, pitchL, Lambda_Q10, LTP_scale_Q14, arch) \
+    ((void)(arch),silk_NSQ_del_dec_neon(psEncC, NSQ, psIndices, x16, pulses, PredCoef_Q12, LTPCoef_Q14, AR_Q13, HarmShapeGain_Q14, Tilt_Q14, LF_shp_Q14, Gains_Q16, pitchL, Lambda_Q10, LTP_scale_Q14))
+#   endif
+#  endif
+
+#endif /* end NSQ_DEL_DEC_ARM_H */
diff --git a/silk/arm/NSQ_del_dec_neon_intr.c b/silk/arm/NSQ_del_dec_neon_intr.c
new file mode 100644
index 0000000..c541dcd
--- /dev/null
+++ b/silk/arm/NSQ_del_dec_neon_intr.c
@@ -0,0 +1,1125 @@
+/* Copyright (c) 2016 Google Inc. */
+/**
+   @file NSQ_del_dec_neon_intr.c
+   @brief ARM Neon Intrinsic optimizations for silk NSQ_del_dec functions
+ */
+
+/*
+   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 <arm_neon.h>
+#include "main.h"
+#include "stack_alloc.h"
+
+typedef struct {
+    opus_int32 sLPC_Q14[ MAX_SUB_FRAME_LENGTH + NSQ_LPC_BUF_LENGTH ][ MAX_DEL_DEC_STATES ];
+    opus_int32 RandState[ DECISION_DELAY ][ MAX_DEL_DEC_STATES ];
+    opus_int32 Q_Q10[     DECISION_DELAY ][ MAX_DEL_DEC_STATES ];
+    opus_int32 Xq_Q14[    DECISION_DELAY ][ MAX_DEL_DEC_STATES ];
+    opus_int32 Pred_Q15[  DECISION_DELAY ][ MAX_DEL_DEC_STATES ];
+    opus_int32 Shape_Q14[ DECISION_DELAY ][ MAX_DEL_DEC_STATES ];
+    opus_int32 sAR2_Q14[ MAX_SHAPE_LPC_ORDER ][ MAX_DEL_DEC_STATES ];
+    opus_int32 LF_AR_Q14[ MAX_DEL_DEC_STATES ];
+    opus_int32 Diff_Q14[ MAX_DEL_DEC_STATES ];
+    opus_int32 Seed[ MAX_DEL_DEC_STATES ];
+    opus_int32 SeedInit[ MAX_DEL_DEC_STATES ];
+    opus_int32 RD_Q10[ MAX_DEL_DEC_STATES ];
+} NSQ_del_decs_struct;
+
+typedef struct {
+    opus_int32 Q_Q10[ MAX_DEL_DEC_STATES ];
+    opus_int32 RD_Q10[ MAX_DEL_DEC_STATES ];
+    opus_int32 xq_Q14[ MAX_DEL_DEC_STATES ];
+    opus_int32 LF_AR_Q14[ MAX_DEL_DEC_STATES ];
+    opus_int32 Diff_Q14[ MAX_DEL_DEC_STATES ];
+    opus_int32 sLTP_shp_Q14[ MAX_DEL_DEC_STATES ];
+    opus_int32 LPC_exc_Q14[ MAX_DEL_DEC_STATES ];
+} NSQ_samples_struct;
+
+static OPUS_INLINE void silk_nsq_del_dec_scale_states(
+    const silk_encoder_state *psEncC,               /* I    Encoder State                       */
+    silk_nsq_state      *NSQ,                       /* I/O  NSQ state                           */
+    NSQ_del_decs_struct psDelDec[],                 /* I/O  Delayed decision states             */
+    const opus_int16    x16[],                      /* I    Input                               */
+    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                     */
+    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(
+    silk_nsq_state      *NSQ,                   /* I/O  NSQ state                           */
+    NSQ_del_decs_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/O  Index to newest samples in buffers  */
+    opus_int            decisionDelay           /* I                                        */
+);
+
+/* This table records ((index + DECISION_DELAY - 1) % DECISION_DELAY) to avoid the division operation when DECISION_DELAY is not a power of 2.
+ * This table must be updated if DECISION_DELAY changes.
+ */
+static const opus_int8 next_smpl_buf_idx_table[ DECISION_DELAY ] =
+{
+        39,  0,  1,  2,  3,  4,  5,  6,  7,  8,
+        9,  10, 11, 12, 13, 14, 15, 16, 17, 18,
+        19, 20, 21, 22, 23, 24, 25, 26, 27, 28,
+        29, 30, 31, 32, 33, 34, 35, 36, 37, 38
+};
+
+/* This table records (index % DECISION_DELAY) to avoid the division operation when DECISION_DELAY is not a power of 2.
+ * This table must be updated if DECISION_DELAY changes.
+ */
+static const opus_int8 mode_DECISION_DELAY_table[ 3 * DECISION_DELAY ] =
+{
+        /* Repeat 2 times to guarantee no index out of bounds. */
+        0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39,
+        0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39,
+        0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39
+};
+
+static OPUS_INLINE void copy_winner_state_kernel(
+    const NSQ_del_decs_struct *psDelDec,
+    const opus_int            offset,
+    const opus_int            last_smple_idx,
+    const opus_int            Winner_ind,
+    const int32x2_t           gain_lo_s32x2,
+    const int32x2_t           gain_hi_s32x2,
+    const int32x4_t           shift_s32x4,
+    int32x4_t                 t0_s32x4,
+    int32x4_t                 t1_s32x4,
+    opus_int8                 *pulses,
+    opus_int16                *pxq,
+    silk_nsq_state            *NSQ
+)
+{
+    int16x8_t t_s16x8;
+    int32x4_t o0_s32x4, o1_s32x4;
+
+    t0_s32x4 = vld1q_lane_s32( &psDelDec->Q_Q10[ last_smple_idx - 0 ][ Winner_ind ], t0_s32x4, 0 );
+    t0_s32x4 = vld1q_lane_s32( &psDelDec->Q_Q10[ last_smple_idx - 1 ][ Winner_ind ], t0_s32x4, 1 );
+    t0_s32x4 = vld1q_lane_s32( &psDelDec->Q_Q10[ last_smple_idx - 2 ][ Winner_ind ], t0_s32x4, 2 );
+    t0_s32x4 = vld1q_lane_s32( &psDelDec->Q_Q10[ last_smple_idx - 3 ][ Winner_ind ], t0_s32x4, 3 );
+    t1_s32x4 = vld1q_lane_s32( &psDelDec->Q_Q10[ last_smple_idx - 4 ][ Winner_ind ], t1_s32x4, 0 );
+    t1_s32x4 = vld1q_lane_s32( &psDelDec->Q_Q10[ last_smple_idx - 5 ][ Winner_ind ], t1_s32x4, 1 );
+    t1_s32x4 = vld1q_lane_s32( &psDelDec->Q_Q10[ last_smple_idx - 6 ][ Winner_ind ], t1_s32x4, 2 );
+    t1_s32x4 = vld1q_lane_s32( &psDelDec->Q_Q10[ last_smple_idx - 7 ][ Winner_ind ], t1_s32x4, 3 );
+    t_s16x8  = vcombine_s16( vrshrn_n_s32( t0_s32x4, 10 ), vrshrn_n_s32( t1_s32x4, 10 ) );
+    vst1_s8( &pulses[ offset ], vmovn_s16( t_s16x8 ) );
+
+    t0_s32x4 = vld1q_lane_s32( &psDelDec->Xq_Q14[ last_smple_idx - 0 ][ Winner_ind ], t0_s32x4, 0 );
+    t0_s32x4 = vld1q_lane_s32( &psDelDec->Xq_Q14[ last_smple_idx - 1 ][ Winner_ind ], t0_s32x4, 1 );
+    t0_s32x4 = vld1q_lane_s32( &psDelDec->Xq_Q14[ last_smple_idx - 2 ][ Winner_ind ], t0_s32x4, 2 );
+    t0_s32x4 = vld1q_lane_s32( &psDelDec->Xq_Q14[ last_smple_idx - 3 ][ Winner_ind ], t0_s32x4, 3 );
+    t1_s32x4 = vld1q_lane_s32( &psDelDec->Xq_Q14[ last_smple_idx - 4 ][ Winner_ind ], t1_s32x4, 0 );
+    t1_s32x4 = vld1q_lane_s32( &psDelDec->Xq_Q14[ last_smple_idx - 5 ][ Winner_ind ], t1_s32x4, 1 );
+    t1_s32x4 = vld1q_lane_s32( &psDelDec->Xq_Q14[ last_smple_idx - 6 ][ Winner_ind ], t1_s32x4, 2 );
+    t1_s32x4 = vld1q_lane_s32( &psDelDec->Xq_Q14[ last_smple_idx - 7 ][ Winner_ind ], t1_s32x4, 3 );
+    o0_s32x4 = vqdmulhq_lane_s32( t0_s32x4, gain_lo_s32x2, 0 );
+    o1_s32x4 = vqdmulhq_lane_s32( t1_s32x4, gain_lo_s32x2, 0 );
+    o0_s32x4 = vmlaq_lane_s32( o0_s32x4, t0_s32x4, gain_hi_s32x2, 0 );
+    o1_s32x4 = vmlaq_lane_s32( o1_s32x4, t1_s32x4, gain_hi_s32x2, 0 );
+    o0_s32x4 = vrshlq_s32( o0_s32x4, shift_s32x4 );
+    o1_s32x4 = vrshlq_s32( o1_s32x4, shift_s32x4 );
+    vst1_s16( &pxq[ offset + 0 ], vqmovn_s32( o0_s32x4 ) );
+    vst1_s16( &pxq[ offset + 4 ], vqmovn_s32( o1_s32x4 ) );
+
+    t0_s32x4 = vld1q_lane_s32( &psDelDec->Shape_Q14[ last_smple_idx - 0 ][ Winner_ind ], t0_s32x4, 0 );
+    t0_s32x4 = vld1q_lane_s32( &psDelDec->Shape_Q14[ last_smple_idx - 1 ][ Winner_ind ], t0_s32x4, 1 );
+    t0_s32x4 = vld1q_lane_s32( &psDelDec->Shape_Q14[ last_smple_idx - 2 ][ Winner_ind ], t0_s32x4, 2 );
+    t0_s32x4 = vld1q_lane_s32( &psDelDec->Shape_Q14[ last_smple_idx - 3 ][ Winner_ind ], t0_s32x4, 3 );
+    t1_s32x4 = vld1q_lane_s32( &psDelDec->Shape_Q14[ last_smple_idx - 4 ][ Winner_ind ], t1_s32x4, 0 );
+    t1_s32x4 = vld1q_lane_s32( &psDelDec->Shape_Q14[ last_smple_idx - 5 ][ Winner_ind ], t1_s32x4, 1 );
+    t1_s32x4 = vld1q_lane_s32( &psDelDec->Shape_Q14[ last_smple_idx - 6 ][ Winner_ind ], t1_s32x4, 2 );
+    t1_s32x4 = vld1q_lane_s32( &psDelDec->Shape_Q14[ last_smple_idx - 7 ][ Winner_ind ], t1_s32x4, 3 );
+    vst1q_s32( &NSQ->sLTP_shp_Q14[ NSQ->sLTP_shp_buf_idx + offset + 0 ], t0_s32x4 );
+    vst1q_s32( &NSQ->sLTP_shp_Q14[ NSQ->sLTP_shp_buf_idx + offset + 4 ], t1_s32x4 );
+}
+
+static OPUS_INLINE void copy_winner_state(
+    const NSQ_del_decs_struct *psDelDec,
+    const opus_int            decisionDelay,
+    const opus_int            smpl_buf_idx,
+    const opus_int            Winner_ind,
+    const opus_int32          gain,
+    const opus_int32          shift,
+    opus_int8                 *pulses,
+    opus_int16                *pxq,
+    silk_nsq_state            *NSQ
+)
+{
+    opus_int        i, last_smple_idx;
+    const int32x2_t gain_lo_s32x2 = vdup_n_s32( ( gain & 0x0000FFFF ) << 15 );
+    const int32x2_t gain_hi_s32x2 = vdup_n_s32( gain >> 16 );
+    const int32x4_t shift_s32x4   = vdupq_n_s32( -shift );
+    int32x4_t       t0_s32x4, t1_s32x4;
+
+    t0_s32x4 = t1_s32x4 = vdupq_n_s32( 0 ); /* initialization */
+    last_smple_idx = smpl_buf_idx + decisionDelay;
+    last_smple_idx = mode_DECISION_DELAY_table[ last_smple_idx - 1 + DECISION_DELAY ];
+
+    for( i = 0; ( i < ( decisionDelay - 7 ) ) && ( last_smple_idx >= 7 ); i += 8, last_smple_idx -= 8 ) {
+        copy_winner_state_kernel( psDelDec, i - decisionDelay, last_smple_idx, Winner_ind, gain_lo_s32x2, gain_hi_s32x2, shift_s32x4, t0_s32x4, t1_s32x4, pulses, pxq, NSQ );
+    }
+    for( ; ( i < decisionDelay ) && ( last_smple_idx >= 0 ); i++, last_smple_idx-- ) {
+        pulses[ i - decisionDelay ] = (opus_int8)silk_RSHIFT_ROUND( psDelDec->Q_Q10[ last_smple_idx ][ Winner_ind ], 10 );
+        pxq[ i - decisionDelay ] = (opus_int16)silk_SAT16( silk_RSHIFT_ROUND( silk_SMULWW( psDelDec->Xq_Q14[ last_smple_idx ][ Winner_ind ], gain ), shift ) );
+        NSQ->sLTP_shp_Q14[ NSQ->sLTP_shp_buf_idx - decisionDelay + i ] = psDelDec->Shape_Q14[ last_smple_idx ][ Winner_ind ];
+    }
+
+    last_smple_idx += DECISION_DELAY;
+    for( ; i < ( decisionDelay - 7 ); i++, last_smple_idx-- ) {
+        copy_winner_state_kernel( psDelDec, i - decisionDelay, last_smple_idx, Winner_ind, gain_lo_s32x2, gain_hi_s32x2, shift_s32x4, t0_s32x4, t1_s32x4, pulses, pxq, NSQ );
+    }
+    for( ; i < decisionDelay; i++, last_smple_idx-- ) {
+        pulses[ i - decisionDelay ] = (opus_int8)silk_RSHIFT_ROUND( psDelDec->Q_Q10[ last_smple_idx ][ Winner_ind ], 10 );
+        pxq[ i - decisionDelay ] = (opus_int16)silk_SAT16( silk_RSHIFT_ROUND( silk_SMULWW( psDelDec->Xq_Q14[ last_smple_idx ][ Winner_ind ], gain ), shift ) );
+        NSQ->sLTP_shp_Q14[ NSQ->sLTP_shp_buf_idx - decisionDelay + i ] = psDelDec->Shape_Q14[ last_smple_idx ][ Winner_ind ];
+    }
+}
+
+void silk_NSQ_del_dec_neon(
+    const silk_encoder_state    *psEncC,                                    /* I    Encoder State                   */
+    silk_nsq_state              *NSQ,                                       /* I/O  NSQ state                       */
+    SideInfoIndices             *psIndices,                                 /* I/O  Quantization Indices            */
+    const opus_int16            x16[],                                      /* I    Input                           */
+    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            AR_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               */
+)
+{
+    /* The optimization parallelizes the different delay decision states. */
+    if( psEncC->nStatesDelayedDecision <= 2 ) {
+        /* When the number of delay decision states is less than 3, there are penalties using the optimization based on ( MAX_DEL_DEC_STATES == 4 ).
+         * In this case C function is called instead.
+         * When the number of delay decision states is 2, it's better to specialize another structure NSQ_del_dec2_struct. (Low priority)
+         */
+        silk_NSQ_del_dec_c( psEncC, NSQ, psIndices, x16, pulses, PredCoef_Q12, LTPCoef_Q14, AR_Q13, HarmShapeGain_Q14,
+                Tilt_Q14, LF_shp_Q14, Gains_Q16, pitchL, Lambda_Q10, LTP_scale_Q14 );
+    }
+    else {
+        opus_int            i, k, lag, start_idx, LSF_interpolation_flag, Winner_ind, subfr;
+        opus_int            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_decs_struct, psDelDec );
+        int32x4_t           t_s32x4;
+        SAVE_STACK;
+
+        /* Set unvoiced lag to the previous one, overwrite later for voiced */
+        lag = NSQ->lagPrev;
+
+        silk_assert( NSQ->prev_gain_Q16 != 0 );
+        silk_assert( MAX_DEL_DEC_STATES == 4 );
+        silk_assert( DECISION_DELAY == 40 ); /* Table next_smpl_buf_idx_table[] and mode_DECISION_DELAY_table[] is hard coded. */
+
+        /* Initialize delayed decision states */
+        ALLOC( psDelDec, 1, NSQ_del_decs_struct );
+        /* Only RandState and RD_Q10 need to be initialized to 0. */
+        silk_memset( psDelDec->RandState, 0, sizeof( psDelDec->RandState ) );
+        vst1q_s32( psDelDec->RD_Q10, vld1q_dup_s32( psDelDec->RD_Q10 ) );
+
+        for( k = 0; k < psEncC->nStatesDelayedDecision; k++ ) {
+            psDelDec->SeedInit[ k ] = psDelDec->Seed[ k ] = ( k + psIndices->Seed ) & 3;
+        }
+        vst1q_s32( psDelDec->LF_AR_Q14,      vld1q_dup_s32( &NSQ->sLF_AR_shp_Q14 ) );
+        vst1q_s32( psDelDec->Diff_Q14,       vld1q_dup_s32( &NSQ->sDiff_shp_Q14 ) );
+        vst1q_s32( psDelDec->Shape_Q14[ 0 ], vld1q_dup_s32( &NSQ->sLTP_shp_Q14[ psEncC->ltp_mem_length - 1 ] ) );
+        for( i = 0; i < NSQ_LPC_BUF_LENGTH; i++ ) {
+            vst1q_s32( psDelDec->sLPC_Q14[ i ], vld1q_dup_s32( &NSQ->sLPC_Q14[ i ] ) );
+        }
+        for( i = 0; i < (opus_int)( sizeof( NSQ->sAR2_Q14 ) / sizeof( NSQ->sAR2_Q14[ 0 ] ) ); i++ ) {
+            vst1q_s32( psDelDec->sAR2_Q14[ i ], vld1q_dup_s32( &NSQ->sAR2_Q14[ i ] ) );
+        }
+
+        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 ) {
+            opus_int pitch_min = pitchL[ 0 ];
+            for( k = 1; k < psEncC->nb_subfr; k++ ) {
+                pitch_min = silk_min_int( pitch_min, pitchL[ k ] );
+            }
+            decisionDelay = silk_min_int( decisionDelay, pitch_min - 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 = &AR_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 */
+                        int32x4_t RD_Q10_s32x4;
+                        RDmin_Q10 = psDelDec->RD_Q10[ 0 ];
+                        Winner_ind = 0;
+                        for( i = 1; i < psEncC->nStatesDelayedDecision; i++ ) {
+                            if( psDelDec->RD_Q10[ i ] < RDmin_Q10 ) {
+                                RDmin_Q10 = psDelDec->RD_Q10[ i ];
+                                Winner_ind = i;
+                            }
+                        }
+                        psDelDec->RD_Q10[ Winner_ind ] -= ( silk_int32_MAX >> 4 );
+                        RD_Q10_s32x4 = vld1q_s32( psDelDec->RD_Q10 );
+                        RD_Q10_s32x4 = vaddq_s32( RD_Q10_s32x4, vdupq_n_s32( silk_int32_MAX >> 4 ) );
+                        vst1q_s32( psDelDec->RD_Q10, RD_Q10_s32x4 );
+
+                        /* Copy final part of signals from winner state to output and long-term filter states */
+                        copy_winner_state( psDelDec, decisionDelay, smpl_buf_idx, Winner_ind, Gains_Q16[ 1 ], 14, pulses, pxq, NSQ );
+
+                        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( psEncC, NSQ, psDelDec, x16, x_sc_Q10, sLTP, sLTP_Q15, k,
+                LTP_scale_Q14, Gains_Q16, pitchL, psIndices->signalType, decisionDelay );
+
+            silk_noise_shape_quantizer_del_dec( 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 );
+
+            x16    += psEncC->subfr_length;
+            pulses += psEncC->subfr_length;
+            pxq    += psEncC->subfr_length;
+        }
+
+        /* Find winner */
+        RDmin_Q10 = psDelDec->RD_Q10[ 0 ];
+        Winner_ind = 0;
+        for( k = 1; k < psEncC->nStatesDelayedDecision; k++ ) {
+            if( psDelDec->RD_Q10[ k ] < RDmin_Q10 ) {
+                RDmin_Q10 = psDelDec->RD_Q10[ k ];
+                Winner_ind = k;
+            }
+        }
+
+        /* Copy final part of signals from winner state to output and long-term filter states */
+        psIndices->Seed = psDelDec->SeedInit[ Winner_ind ];
+        Gain_Q10 = silk_RSHIFT32( Gains_Q16[ psEncC->nb_subfr - 1 ], 6 );
+        copy_winner_state( psDelDec, decisionDelay, smpl_buf_idx, Winner_ind, Gain_Q10, 8, pulses, pxq, NSQ );
+
+        silk_assert( !( NSQ_LPC_BUF_LENGTH % 4 ) );
+        t_s32x4 = vdupq_n_s32( 0 ); /* initialization */
+        for( i = 0; i < NSQ_LPC_BUF_LENGTH; i += 4 ) {
+            t_s32x4 = vld1q_lane_s32( &psDelDec->sLPC_Q14[ i + 0 ][ Winner_ind ], t_s32x4, 0 );
+            t_s32x4 = vld1q_lane_s32( &psDelDec->sLPC_Q14[ i + 1 ][ Winner_ind ], t_s32x4, 1 );
+            t_s32x4 = vld1q_lane_s32( &psDelDec->sLPC_Q14[ i + 2 ][ Winner_ind ], t_s32x4, 2 );
+            t_s32x4 = vld1q_lane_s32( &psDelDec->sLPC_Q14[ i + 3 ][ Winner_ind ], t_s32x4, 3 );
+            vst1q_s32( &NSQ->sLPC_Q14[ i ], t_s32x4 );
+        }
+
+        silk_assert( !( ( sizeof( NSQ->sAR2_Q14 ) / sizeof( NSQ->sAR2_Q14[ 0 ] ) ) % 4 ) );
+        for( i = 0; i < (opus_int)( sizeof( NSQ->sAR2_Q14 ) / sizeof( NSQ->sAR2_Q14[ 0 ] ) ); i += 4 ) {
+            t_s32x4 = vld1q_lane_s32( &psDelDec->sAR2_Q14[ i + 0 ][ Winner_ind ], t_s32x4, 0 );
+            t_s32x4 = vld1q_lane_s32( &psDelDec->sAR2_Q14[ i + 1 ][ Winner_ind ], t_s32x4, 1 );
+            t_s32x4 = vld1q_lane_s32( &psDelDec->sAR2_Q14[ i + 2 ][ Winner_ind ], t_s32x4, 2 );
+            t_s32x4 = vld1q_lane_s32( &psDelDec->sAR2_Q14[ i + 3 ][ Winner_ind ], t_s32x4, 3 );
+            vst1q_s32( &NSQ->sAR2_Q14[ i ], t_s32x4 );
+        }
+
+        /* Update states */
+        NSQ->sLF_AR_shp_Q14 = psDelDec->LF_AR_Q14[ Winner_ind ];
+        NSQ->sDiff_shp_Q14  = psDelDec->Diff_Q14[ Winner_ind ];
+        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 int32x4_t silk_SMLAWB_lane_neon(
+        const int32x4_t out_s32x4,
+        const int32x4_t in_s32x4,
+        const int32x2_t coef_s32x2,
+        const opus_int  lane
+)
+{
+    return vaddq_s32( out_s32x4, vqdmulhq_lane_s32( in_s32x4, coef_s32x2, lane ) );
+}
+
+static OPUS_INLINE void silk_noise_shape_quantizer_del_dec(
+    silk_nsq_state      *NSQ,                   /* I/O  NSQ state                           */
+    NSQ_del_decs_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/O  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, n_LTP_Q14;
+    opus_int32       RDmin_Q10, RDmax_Q10;
+    opus_int32       Gain_Q10;
+    opus_int32       *pred_lag_ptr, *shp_lag_ptr;
+    opus_int32       a_Q12_arch[ MAX_LPC_ORDER ];
+    const int32x2_t  warping_Q16_s32x2 = vdup_n_s32( ( warping_Q16 << 16 ) >> 1 );
+    const opus_int32 LF_shp_Q29 = ( LF_shp_Q14 << 16 ) >> 1;
+    opus_int32       AR_shp_Q28[ MAX_SHAPE_LPC_ORDER ];
+    const uint32x4_t rand_multiplier_u32x4 = vdupq_n_u32( RAND_MULTIPLIER );
+    const uint32x4_t rand_increment_u32x4  = vdupq_n_u32( RAND_INCREMENT );
+
+    VARDECL( NSQ_samples_struct, psSampleState );
+    SAVE_STACK;
+
+    silk_assert( nStatesDelayedDecision > 0 );
+    silk_assert( ( shapingLPCOrder & 1 ) == 0 );   /* check that order is even */
+    ALLOC( psSampleState, 2, NSQ_samples_struct );
+
+    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 );
+
+    silk_assert( MAX_SHAPE_LPC_ORDER == 24 );
+    {
+        const int16x8_t t0_s16x8 = vld1q_s16( AR_shp_Q13 +  0 );
+        const int16x8_t t1_s16x8 = vld1q_s16( AR_shp_Q13 +  8 );
+        const int16x8_t t2_s16x8 = vld1q_s16( AR_shp_Q13 + 16 );
+        vst1q_s32( AR_shp_Q28 +  0, vshll_n_s16( vget_low_s16 ( t0_s16x8 ), 15 ) );
+        vst1q_s32( AR_shp_Q28 +  4, vshll_n_s16( vget_high_s16( t0_s16x8 ), 15 ) );
+        vst1q_s32( AR_shp_Q28 +  8, vshll_n_s16( vget_low_s16 ( t1_s16x8 ), 15 ) );
+        vst1q_s32( AR_shp_Q28 + 12, vshll_n_s16( vget_high_s16( t1_s16x8 ), 15 ) );
+        vst1q_s32( AR_shp_Q28 + 16, vshll_n_s16( vget_low_s16 ( t2_s16x8 ), 15 ) );
+        vst1q_s32( AR_shp_Q28 + 20, vshll_n_s16( vget_high_s16( t2_s16x8 ), 15 ) );
+    }
+
+    /* silk_short_prediction_create_arch_coef( a_Q12_arch, a_Q12, predictLPCOrder ); */
+    silk_assert( MAX_LPC_ORDER == 16 );
+    if( predictLPCOrder == MAX_LPC_ORDER ) {
+        int16x8_t t0_s16x8, t1_s16x8;
+        t0_s16x8 = vld1q_s16( a_Q12 + 0 );                                          /* 7 6 5 4  3 2 1 0 */
+        t1_s16x8 = vld1q_s16( a_Q12 + 8 );                                          /* F E D C  B A 9 8 */
+        t0_s16x8 = vrev64q_s16( t0_s16x8 );                                         /* 4 5 6 7  0 1 2 3 */
+        t1_s16x8 = vrev64q_s16( t1_s16x8 );                                         /* C D E F  8 9 A B */
+        vst1q_s32( a_Q12_arch +  0, vshll_n_s16( vget_high_s16( t1_s16x8 ), 15 ) ); /* C D E F          */
+        vst1q_s32( a_Q12_arch +  4, vshll_n_s16( vget_low_s16 ( t1_s16x8 ), 15 ) ); /* 8 9 A B          */
+        vst1q_s32( a_Q12_arch +  8, vshll_n_s16( vget_high_s16( t0_s16x8 ), 15 ) ); /* 4 5 6 7          */
+        vst1q_s32( a_Q12_arch + 12, vshll_n_s16( vget_low_s16 ( t0_s16x8 ), 15 ) ); /* 0 1 2 3          */
+    }
+    else {
+        int16x8_t t_s16x8;
+        int16x4_t t_s16x4;
+        int32x4_t t_s32x4;
+        silk_assert( predictLPCOrder == MIN_LPC_ORDER );
+        t_s16x8 = vld1q_s16( a_Q12 + 0 );                                           /* 7 6 5 4  3 2 1 0 */
+        t_s16x4 = vld1_s16 ( a_Q12 + 8 );                                           /* B A 9 8          */
+        t_s16x8 = vrev64q_s16( t_s16x8 );                                           /* 4 5 6 7  0 1 2 3 */
+        t_s16x4 = vrev64_s16( t_s16x4 );                                            /* 8 9 A B          */
+        t_s32x4 = vshll_n_s16( t_s16x4, 15 );
+        t_s32x4 = vcombine_s32( vdup_n_s32( 0 ), vget_high_s32( t_s32x4 ) );        /* 8 9 zero zero    */
+        vst1q_s32( a_Q12_arch +  0, vdupq_n_s32( 0 ) );                             /* zero zero zero zero */
+        vst1q_s32( a_Q12_arch +  4, t_s32x4 );                                      /* 8 9 zero zero    */
+        vst1q_s32( a_Q12_arch +  8, vshll_n_s16( vget_high_s16( t_s16x8 ), 15 ) );  /* 4 5 6 7          */
+        vst1q_s32( a_Q12_arch + 12, vshll_n_s16( vget_low_s16 ( t_s16x8 ), 15 ) );  /* 0 1 2 3          */
+    }
+
+    for( i = 0; i < length; i++ ) {
+        int32x4_t Seed_s32x4, LPC_pred_Q14_s32x4;
+        int32x4_t sign_s32x4, tmp1_s32x4, tmp2_s32x4;
+        int32x4_t n_AR_Q14_s32x4, n_LF_Q14_s32x4;
+        int32x2_t AR_shp_Q28_s32x2;
+        int16x4_t r_Q10_s16x4, rr_Q10_s16x4;
+
+        /* 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;
+            LTP_pred_Q14 = silk_SMLAWB( LTP_pred_Q14, pred_lag_ptr[  0 ], b_Q14[ 0 ] );
+            LTP_pred_Q14 = silk_SMLAWB( LTP_pred_Q14, pred_lag_ptr[ -1 ], b_Q14[ 1 ] );
+            LTP_pred_Q14 = silk_SMLAWB( LTP_pred_Q14, pred_lag_ptr[ -2 ], b_Q14[ 2 ] );
+            LTP_pred_Q14 = silk_SMLAWB( LTP_pred_Q14, pred_lag_ptr[ -3 ], b_Q14[ 3 ] );
+            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;
+        }
+
+        /* Generate dither */
+        Seed_s32x4 = vld1q_s32( psDelDec->Seed );
+        Seed_s32x4 = vreinterpretq_s32_u32( vmlaq_u32( rand_increment_u32x4, vreinterpretq_u32_s32( Seed_s32x4 ), rand_multiplier_u32x4 ) );
+        vst1q_s32( psDelDec->Seed, Seed_s32x4 );
+
+        /* Short-term prediction */
+        {
+            const opus_int32 *buf32            = psDelDec->sLPC_Q14[ NSQ_LPC_BUF_LENGTH - 16 + i ];
+            const int32x4_t  a_Q12_arch0_s32x4 = vld1q_s32( a_Q12_arch +  0 );
+            const int32x4_t  a_Q12_arch1_s32x4 = vld1q_s32( a_Q12_arch +  4 );
+            const int32x4_t  a_Q12_arch2_s32x4 = vld1q_s32( a_Q12_arch +  8 );
+            const int32x4_t  a_Q12_arch3_s32x4 = vld1q_s32( a_Q12_arch + 12 );
+            LPC_pred_Q14_s32x4 = vdupq_n_s32( silk_RSHIFT( predictLPCOrder, 1 ) );
+            LPC_pred_Q14_s32x4 = silk_SMLAWB_lane_neon( LPC_pred_Q14_s32x4, vld1q_s32( buf32 +  0 * MAX_DEL_DEC_STATES ), vget_low_s32 ( a_Q12_arch0_s32x4 ), 0 );
+            LPC_pred_Q14_s32x4 = silk_SMLAWB_lane_neon( LPC_pred_Q14_s32x4, vld1q_s32( buf32 +  1 * MAX_DEL_DEC_STATES ), vget_low_s32 ( a_Q12_arch0_s32x4 ), 1 );
+            LPC_pred_Q14_s32x4 = silk_SMLAWB_lane_neon( LPC_pred_Q14_s32x4, vld1q_s32( buf32 +  2 * MAX_DEL_DEC_STATES ), vget_high_s32( a_Q12_arch0_s32x4 ), 0 );
+            LPC_pred_Q14_s32x4 = silk_SMLAWB_lane_neon( LPC_pred_Q14_s32x4, vld1q_s32( buf32 +  3 * MAX_DEL_DEC_STATES ), vget_high_s32( a_Q12_arch0_s32x4 ), 1 );
+            LPC_pred_Q14_s32x4 = silk_SMLAWB_lane_neon( LPC_pred_Q14_s32x4, vld1q_s32( buf32 +  4 * MAX_DEL_DEC_STATES ), vget_low_s32 ( a_Q12_arch1_s32x4 ), 0 );
+            LPC_pred_Q14_s32x4 = silk_SMLAWB_lane_neon( LPC_pred_Q14_s32x4, vld1q_s32( buf32 +  5 * MAX_DEL_DEC_STATES ), vget_low_s32 ( a_Q12_arch1_s32x4 ), 1 );
+            LPC_pred_Q14_s32x4 = silk_SMLAWB_lane_neon( LPC_pred_Q14_s32x4, vld1q_s32( buf32 +  6 * MAX_DEL_DEC_STATES ), vget_high_s32( a_Q12_arch1_s32x4 ), 0 );
+            LPC_pred_Q14_s32x4 = silk_SMLAWB_lane_neon( LPC_pred_Q14_s32x4, vld1q_s32( buf32 +  7 * MAX_DEL_DEC_STATES ), vget_high_s32( a_Q12_arch1_s32x4 ), 1 );
+            LPC_pred_Q14_s32x4 = silk_SMLAWB_lane_neon( LPC_pred_Q14_s32x4, vld1q_s32( buf32 +  8 * MAX_DEL_DEC_STATES ), vget_low_s32 ( a_Q12_arch2_s32x4 ), 0 );
+            LPC_pred_Q14_s32x4 = silk_SMLAWB_lane_neon( LPC_pred_Q14_s32x4, vld1q_s32( buf32 +  9 * MAX_DEL_DEC_STATES ), vget_low_s32 ( a_Q12_arch2_s32x4 ), 1 );
+            LPC_pred_Q14_s32x4 = silk_SMLAWB_lane_neon( LPC_pred_Q14_s32x4, vld1q_s32( buf32 + 10 * MAX_DEL_DEC_STATES ), vget_high_s32( a_Q12_arch2_s32x4 ), 0 );
+            LPC_pred_Q14_s32x4 = silk_SMLAWB_lane_neon( LPC_pred_Q14_s32x4, vld1q_s32( buf32 + 11 * MAX_DEL_DEC_STATES ), vget_high_s32( a_Q12_arch2_s32x4 ), 1 );
+            LPC_pred_Q14_s32x4 = silk_SMLAWB_lane_neon( LPC_pred_Q14_s32x4, vld1q_s32( buf32 + 12 * MAX_DEL_DEC_STATES ), vget_low_s32 ( a_Q12_arch3_s32x4 ), 0 );
+            LPC_pred_Q14_s32x4 = silk_SMLAWB_lane_neon( LPC_pred_Q14_s32x4, vld1q_s32( buf32 + 13 * MAX_DEL_DEC_STATES ), vget_low_s32 ( a_Q12_arch3_s32x4 ), 1 );
+            LPC_pred_Q14_s32x4 = silk_SMLAWB_lane_neon( LPC_pred_Q14_s32x4, vld1q_s32( buf32 + 14 * MAX_DEL_DEC_STATES ), vget_high_s32( a_Q12_arch3_s32x4 ), 0 );
+            LPC_pred_Q14_s32x4 = silk_SMLAWB_lane_neon( LPC_pred_Q14_s32x4, vld1q_s32( buf32 + 15 * MAX_DEL_DEC_STATES ), vget_high_s32( a_Q12_arch3_s32x4 ), 1 );
+            LPC_pred_Q14_s32x4 = vshlq_n_s32( LPC_pred_Q14_s32x4, 4 );                              /* Q10 -> Q14 */
+        }
+
+        /* Noise shape feedback */
+        /* Output of lowpass section */
+        tmp2_s32x4 = silk_SMLAWB_lane_neon( vld1q_s32( psDelDec->Diff_Q14 ), vld1q_s32( psDelDec->sAR2_Q14[ 0 ] ), warping_Q16_s32x2, 0 );
+        /* Output of allpass section */
+        tmp1_s32x4 = vsubq_s32( vld1q_s32( psDelDec->sAR2_Q14[ 1 ] ), tmp2_s32x4 );
+        tmp1_s32x4 = silk_SMLAWB_lane_neon( vld1q_s32( psDelDec->sAR2_Q14[ 0 ] ), tmp1_s32x4, warping_Q16_s32x2, 0 );
+        vst1q_s32( psDelDec->sAR2_Q14[ 0 ], tmp2_s32x4 );
+        AR_shp_Q28_s32x2 = vld1_s32( AR_shp_Q28 );
+        n_AR_Q14_s32x4 = vaddq_s32( vdupq_n_s32( silk_RSHIFT( shapingLPCOrder, 1 ) ), vqdmulhq_lane_s32( tmp2_s32x4, AR_shp_Q28_s32x2, 0 ) );
+
+        /* Loop over allpass sections */
+        for( j = 2; j < shapingLPCOrder; j += 2 ) {
+            /* Output of allpass section */
+            tmp2_s32x4 = vsubq_s32( vld1q_s32( psDelDec->sAR2_Q14[ j + 0 ] ), tmp1_s32x4 );
+            tmp2_s32x4 = silk_SMLAWB_lane_neon( vld1q_s32( psDelDec->sAR2_Q14[ j - 1 ] ), tmp2_s32x4, warping_Q16_s32x2, 0 );
+            vst1q_s32( psDelDec->sAR2_Q14[ j - 1 ], tmp1_s32x4 );
+            n_AR_Q14_s32x4 = vaddq_s32( n_AR_Q14_s32x4, vqdmulhq_lane_s32( tmp1_s32x4, AR_shp_Q28_s32x2, 1 ) );
+            /* Output of allpass section */
+            tmp1_s32x4 = vsubq_s32( vld1q_s32( psDelDec->sAR2_Q14[ j + 1 ] ), tmp2_s32x4 );
+            tmp1_s32x4 = silk_SMLAWB_lane_neon( vld1q_s32( psDelDec->sAR2_Q14[ j + 0 ] ), tmp1_s32x4, warping_Q16_s32x2, 0 );
+            vst1q_s32( psDelDec->sAR2_Q14[ j + 0 ], tmp2_s32x4 );
+            AR_shp_Q28_s32x2 = vld1_s32( &AR_shp_Q28[ j ] );
+            n_AR_Q14_s32x4 = vaddq_s32( n_AR_Q14_s32x4, vqdmulhq_lane_s32( tmp2_s32x4, AR_shp_Q28_s32x2, 0 ) );
+        }
+        vst1q_s32( psDelDec->sAR2_Q14[ shapingLPCOrder - 1 ], tmp1_s32x4 );
+        n_AR_Q14_s32x4 = vaddq_s32( n_AR_Q14_s32x4, vqdmulhq_lane_s32( tmp1_s32x4, AR_shp_Q28_s32x2, 1 ) );
+        n_AR_Q14_s32x4 = vshlq_n_s32( n_AR_Q14_s32x4, 1 );                                                                            /* Q11 -> Q12 */
+        n_AR_Q14_s32x4 = vaddq_s32( n_AR_Q14_s32x4, vqdmulhq_n_s32( vld1q_s32( psDelDec->LF_AR_Q14 ), ( Tilt_Q14 << 16 ) >> 1 ) );    /* Q12 */
+        n_AR_Q14_s32x4 = vshlq_n_s32( n_AR_Q14_s32x4, 2 );                                                                            /* Q12 -> Q14 */
+        n_LF_Q14_s32x4 = vqdmulhq_n_s32( vld1q_s32( psDelDec->Shape_Q14[ *smpl_buf_idx ] ), LF_shp_Q29 );                             /* Q12 */
+        n_LF_Q14_s32x4 = vaddq_s32( n_LF_Q14_s32x4, vqdmulhq_n_s32( vld1q_s32( psDelDec->LF_AR_Q14 ), ( LF_shp_Q14 >> 16 ) << 15 ) ); /* Q12 */
+        n_LF_Q14_s32x4 = vshlq_n_s32( n_LF_Q14_s32x4, 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_s32x4 = vaddq_s32( n_AR_Q14_s32x4, n_LF_Q14_s32x4 );                                                                     /* Q14 */
+        tmp2_s32x4 = vaddq_s32( vdupq_n_s32( n_LTP_Q14 ), LPC_pred_Q14_s32x4 );                                                       /* Q13 */
+        tmp1_s32x4 = vsubq_s32( tmp2_s32x4, tmp1_s32x4 );                                                                             /* Q13 */
+        tmp1_s32x4 = vrshrq_n_s32( tmp1_s32x4, 4 );                                                                                   /* Q10 */
+        tmp1_s32x4 = vsubq_s32( vdupq_n_s32( x_Q10[ i ] ), tmp1_s32x4 );                                                              /* residual error Q10 */
+
+        /* Flip sign depending on dither */
+        sign_s32x4 = vreinterpretq_s32_u32( vcltq_s32( Seed_s32x4, vdupq_n_s32( 0 ) ) );
+        tmp1_s32x4 = veorq_s32( tmp1_s32x4, sign_s32x4 );
+        tmp1_s32x4 = vsubq_s32( tmp1_s32x4, sign_s32x4 );
+        tmp1_s32x4 = vmaxq_s32( tmp1_s32x4, vdupq_n_s32( -( 31 << 10 ) ) );
+        tmp1_s32x4 = vminq_s32( tmp1_s32x4, vdupq_n_s32( 30 << 10 ) );
+        r_Q10_s16x4 = vmovn_s32( tmp1_s32x4 );
+
+        /* Find two quantization level candidates and measure their rate-distortion */
+        {
+            int16x4_t  q1_Q10_s16x4 = vsub_s16( r_Q10_s16x4, vdup_n_s16( offset_Q10 ) );
+            int16x4_t  q1_Q0_s16x4  = vshr_n_s16( q1_Q10_s16x4, 10 );
+            int16x4_t  q2_Q10_s16x4;
+            int32x4_t  rd1_Q10_s32x4, rd2_Q10_s32x4;
+            uint32x4_t t_u32x4;
+
+            if( Lambda_Q10 > 2048 ) {
+                /* For aggressive RDO, the bias becomes more than one pulse. */
+                const int rdo_offset = Lambda_Q10/2 - 512;
+                const uint16x4_t greaterThanRdo   = vcgt_s16( q1_Q10_s16x4, vdup_n_s16( rdo_offset ) );
+                const uint16x4_t lessThanMinusRdo = vclt_s16( q1_Q10_s16x4, vdup_n_s16( -rdo_offset ) );
+                /* If Lambda_Q10 > 32767, then q1_Q0, q1_Q10 and q2_Q10 must change to 32-bit. */
+                silk_assert( Lambda_Q10 <= 32767 );
+
+                q1_Q0_s16x4 = vreinterpret_s16_u16( vclt_s16( q1_Q10_s16x4, vdup_n_s16( 0 ) ) );
+                q1_Q0_s16x4 = vbsl_s16( greaterThanRdo,   vsub_s16( q1_Q10_s16x4, vdup_n_s16( rdo_offset ) ), q1_Q0_s16x4 );
+                q1_Q0_s16x4 = vbsl_s16( lessThanMinusRdo, vadd_s16( q1_Q10_s16x4, vdup_n_s16( rdo_offset ) ), q1_Q0_s16x4 );
+                q1_Q0_s16x4 = vshr_n_s16( q1_Q0_s16x4, 10 );
+            }
+            {
+                const uint16x4_t equal0_u16x4         = vceq_s16( q1_Q0_s16x4, vdup_n_s16( 0 ) );
+                const uint16x4_t equalMinus1_u16x4    = vceq_s16( q1_Q0_s16x4, vdup_n_s16( -1 ) );
+                const uint16x4_t lessThanMinus1_u16x4 = vclt_s16( q1_Q0_s16x4, vdup_n_s16( -1 ) );
+                int16x4_t        tmp1_s16x4, tmp2_s16x4;
+
+                q1_Q10_s16x4 = vshl_n_s16( q1_Q0_s16x4, 10 );
+                tmp1_s16x4   = vadd_s16( q1_Q10_s16x4, vdup_n_s16( offset_Q10 - QUANT_LEVEL_ADJUST_Q10 ) );
+                q1_Q10_s16x4 = vadd_s16( q1_Q10_s16x4, vdup_n_s16( offset_Q10 + QUANT_LEVEL_ADJUST_Q10 ) );
+                q1_Q10_s16x4 = vbsl_s16( lessThanMinus1_u16x4, q1_Q10_s16x4,                                                 tmp1_s16x4 );
+                q1_Q10_s16x4 = vbsl_s16( equal0_u16x4,         vdup_n_s16( offset_Q10 ),                                     q1_Q10_s16x4 );
+                q1_Q10_s16x4 = vbsl_s16( equalMinus1_u16x4,    vdup_n_s16( offset_Q10 - ( 1024 - QUANT_LEVEL_ADJUST_Q10 ) ), q1_Q10_s16x4 );
+                q2_Q10_s16x4 = vadd_s16( q1_Q10_s16x4, vdup_n_s16( 1024 ) );
+                q2_Q10_s16x4 = vbsl_s16( equal0_u16x4,         vdup_n_s16( offset_Q10 + 1024 - QUANT_LEVEL_ADJUST_Q10 ),     q2_Q10_s16x4 );
+                q2_Q10_s16x4 = vbsl_s16( equalMinus1_u16x4,    vdup_n_s16( offset_Q10 ),                                     q2_Q10_s16x4 );
+                tmp1_s16x4   = q1_Q10_s16x4;
+                tmp2_s16x4   = q2_Q10_s16x4;
+                tmp1_s16x4 = vbsl_s16( vorr_u16( equalMinus1_u16x4, lessThanMinus1_u16x4 ), vneg_s16( tmp1_s16x4 ),          tmp1_s16x4 );
+                tmp2_s16x4 = vbsl_s16( lessThanMinus1_u16x4, vneg_s16( tmp2_s16x4 ),                                         tmp2_s16x4 );
+                rd1_Q10_s32x4 = vmull_s16( tmp1_s16x4, vdup_n_s16( Lambda_Q10 ) );
+                rd2_Q10_s32x4 = vmull_s16( tmp2_s16x4, vdup_n_s16( Lambda_Q10 ) );
+            }
+
+            rr_Q10_s16x4  = vsub_s16( r_Q10_s16x4, q1_Q10_s16x4 );
+            rd1_Q10_s32x4 = vmlal_s16( rd1_Q10_s32x4, rr_Q10_s16x4, rr_Q10_s16x4 );
+            rd1_Q10_s32x4 = vshrq_n_s32( rd1_Q10_s32x4, 10 );
+
+            rr_Q10_s16x4  = vsub_s16( r_Q10_s16x4, q2_Q10_s16x4 );
+            rd2_Q10_s32x4 = vmlal_s16( rd2_Q10_s32x4, rr_Q10_s16x4, rr_Q10_s16x4 );
+            rd2_Q10_s32x4 = vshrq_n_s32( rd2_Q10_s32x4, 10 );
+
+            tmp2_s32x4 = vld1q_s32( psDelDec->RD_Q10 );
+            tmp1_s32x4 = vaddq_s32( tmp2_s32x4, vminq_s32( rd1_Q10_s32x4, rd2_Q10_s32x4 ) );
+            tmp2_s32x4 = vaddq_s32( tmp2_s32x4, vmaxq_s32( rd1_Q10_s32x4, rd2_Q10_s32x4 ) );
+            vst1q_s32( psSampleState[ 0 ].RD_Q10, tmp1_s32x4 );
+            vst1q_s32( psSampleState[ 1 ].RD_Q10, tmp2_s32x4 );
+            t_u32x4    = vcltq_s32( rd1_Q10_s32x4, rd2_Q10_s32x4 );
+            tmp1_s32x4 = vbslq_s32( t_u32x4, vmovl_s16( q1_Q10_s16x4 ), vmovl_s16( q2_Q10_s16x4 ) );
+            tmp2_s32x4 = vbslq_s32( t_u32x4, vmovl_s16( q2_Q10_s16x4 ), vmovl_s16( q1_Q10_s16x4 ) );
+            vst1q_s32( psSampleState[ 0 ].Q_Q10, tmp1_s32x4 );
+            vst1q_s32( psSampleState[ 1 ].Q_Q10, tmp2_s32x4 );
+        }
+
+        {
+            /* Update states for best quantization */
+            int32x4_t exc_Q14_s32x4, LPC_exc_Q14_s32x4, xq_Q14_s32x4, sLF_AR_shp_Q14_s32x4;
+
+            /* Quantized excitation */
+            exc_Q14_s32x4 = vshlq_n_s32( tmp1_s32x4, 4 );
+            exc_Q14_s32x4 = veorq_s32( exc_Q14_s32x4, sign_s32x4 );
+            exc_Q14_s32x4 = vsubq_s32( exc_Q14_s32x4, sign_s32x4 );
+
+            /* Add predictions */
+            LPC_exc_Q14_s32x4 = vaddq_s32( exc_Q14_s32x4, vdupq_n_s32( LTP_pred_Q14 ) );
+            xq_Q14_s32x4      = vaddq_s32( LPC_exc_Q14_s32x4, LPC_pred_Q14_s32x4 );
+
+            /* Update states */
+            tmp1_s32x4 = vsubq_s32( xq_Q14_s32x4, vshlq_n_s32( vdupq_n_s32( x_Q10[ i ] ), 4 ) );
+            vst1q_s32( psSampleState[ 0 ].Diff_Q14, tmp1_s32x4 );
+            sLF_AR_shp_Q14_s32x4 = vsubq_s32( tmp1_s32x4, n_AR_Q14_s32x4 );
+            vst1q_s32( psSampleState[ 0 ].sLTP_shp_Q14, vsubq_s32( sLF_AR_shp_Q14_s32x4, n_LF_Q14_s32x4 ) );
+            vst1q_s32( psSampleState[ 0 ].LF_AR_Q14,    sLF_AR_shp_Q14_s32x4 );
+            vst1q_s32( psSampleState[ 0 ].LPC_exc_Q14,  LPC_exc_Q14_s32x4 );
+            vst1q_s32( psSampleState[ 0 ].xq_Q14,       xq_Q14_s32x4 );
+
+            /* Quantized excitation */
+            exc_Q14_s32x4 = vshlq_n_s32( tmp2_s32x4, 4 );
+            exc_Q14_s32x4 = veorq_s32( exc_Q14_s32x4, sign_s32x4 );
+            exc_Q14_s32x4 = vsubq_s32( exc_Q14_s32x4, sign_s32x4 );
+
+            /* Add predictions */
+            LPC_exc_Q14_s32x4 = vaddq_s32( exc_Q14_s32x4, vdupq_n_s32( LTP_pred_Q14 ) );
+            xq_Q14_s32x4      = vaddq_s32( LPC_exc_Q14_s32x4, LPC_pred_Q14_s32x4 );
+
+            /* Update states */
+            tmp1_s32x4 = vsubq_s32( xq_Q14_s32x4, vshlq_n_s32( vdupq_n_s32( x_Q10[ i ] ), 4 ) );
+            vst1q_s32( psSampleState[ 1 ].Diff_Q14, tmp1_s32x4 );
+            sLF_AR_shp_Q14_s32x4 = vsubq_s32( tmp1_s32x4, n_AR_Q14_s32x4 );
+            vst1q_s32( psSampleState[ 1 ].sLTP_shp_Q14, vsubq_s32( sLF_AR_shp_Q14_s32x4, n_LF_Q14_s32x4 ) );
+            vst1q_s32( psSampleState[ 1 ].LF_AR_Q14,    sLF_AR_shp_Q14_s32x4 );
+            vst1q_s32( psSampleState[ 1 ].LPC_exc_Q14,  LPC_exc_Q14_s32x4 );
+            vst1q_s32( psSampleState[ 1 ].xq_Q14,       xq_Q14_s32x4 );
+        }
+
+        *smpl_buf_idx  = next_smpl_buf_idx_table[ *smpl_buf_idx ];
+        last_smple_idx = mode_DECISION_DELAY_table[ *smpl_buf_idx + decisionDelay + DECISION_DELAY ];
+
+        /* Find winner */
+        RDmin_Q10  = psSampleState[ 0 ].RD_Q10[ 0 ];
+        Winner_ind = 0;
+        for( k = 1; k < nStatesDelayedDecision; k++ ) {
+            if( psSampleState[ 0 ].RD_Q10[ k ] < RDmin_Q10 ) {
+                RDmin_Q10  = psSampleState[ 0 ].RD_Q10[ k ];
+                Winner_ind = k;
+            }
+        }
+
+        /* Increase RD values of expired states */
+        {
+            uint32x4_t t_u32x4;
+            Winner_rand_state = psDelDec->RandState[ last_smple_idx ][ Winner_ind ];
+            t_u32x4           = vceqq_s32( vld1q_s32( psDelDec->RandState[ last_smple_idx ] ), vdupq_n_s32( Winner_rand_state ) );
+            t_u32x4           = vmvnq_u32( t_u32x4 );
+            t_u32x4           = vshrq_n_u32( t_u32x4, 5 );
+            tmp1_s32x4        = vld1q_s32( psSampleState[ 0 ].RD_Q10 );
+            tmp2_s32x4        = vld1q_s32( psSampleState[ 1 ].RD_Q10 );
+            tmp1_s32x4        = vaddq_s32( tmp1_s32x4, vreinterpretq_s32_u32( t_u32x4 ) );
+            tmp2_s32x4        = vaddq_s32( tmp2_s32x4, vreinterpretq_s32_u32( t_u32x4 ) );
+            vst1q_s32( psSampleState[ 0 ].RD_Q10, tmp1_s32x4 );
+            vst1q_s32( psSampleState[ 1 ].RD_Q10, tmp2_s32x4 );
+
+            /* Find worst in first set and best in second set */
+            RDmax_Q10 = psSampleState[ 0 ].RD_Q10[ 0 ];
+            RDmin_Q10 = psSampleState[ 1 ].RD_Q10[ 0 ];
+            RDmax_ind = 0;
+            RDmin_ind = 0;
+            for( k = 1; k < nStatesDelayedDecision; k++ ) {
+                /* find worst in first set */
+                if( psSampleState[ 0 ].RD_Q10[ k ] > RDmax_Q10 ) {
+                    RDmax_Q10 = psSampleState[ 0 ].RD_Q10[ k ];
+                    RDmax_ind = k;
+                }
+                /* find best in second set */
+                if( psSampleState[ 1 ].RD_Q10[ k ] < RDmin_Q10 ) {
+                    RDmin_Q10 = psSampleState[ 1 ].RD_Q10[ k ];
+                    RDmin_ind = k;
+                }
+            }
+        }
+
+        /* Replace a state if best from second set outperforms worst in first set */
+        if( RDmin_Q10 < RDmax_Q10 ) {
+            /* Only ( predictLPCOrder - 1 ) of sLPC_Q14 buffer need to be updated, though the first several
+             * useless sLPC_Q14[] will be different comparing with C when predictLPCOrder < NSQ_LPC_BUF_LENGTH.
+             * Here just update constant ( NSQ_LPC_BUF_LENGTH - 1 ) for simplicity.
+             */
+            opus_int32       *dst = (opus_int32 *)psDelDec + ( i + 1 ) * MAX_DEL_DEC_STATES + RDmax_ind;
+            const opus_int32 *src = dst + RDmin_ind - RDmax_ind;
+            for( j = 0; j < NSQ_LPC_BUF_LENGTH - 1; j++ ) {
+                dst[ 4 * j ] = src[ 4 * j ];
+            }
+            dst = (opus_int32 *)psDelDec->RandState + RDmax_ind;
+            src = dst + RDmin_ind - RDmax_ind;
+            for( j = 0; j < (int)( ( sizeof( NSQ_del_decs_struct ) - sizeof( ( (NSQ_del_decs_struct *) 0 )->sLPC_Q14 ) ) / ( MAX_DEL_DEC_STATES * sizeof(opus_int32) ) - 9 ); j += 10 ) {
+                dst[ 4 * ( j + 0 ) ] = src[ 4 * ( j + 0 ) ];
+                dst[ 4 * ( j + 1 ) ] = src[ 4 * ( j + 1 ) ];
+                dst[ 4 * ( j + 2 ) ] = src[ 4 * ( j + 2 ) ];
+                dst[ 4 * ( j + 3 ) ] = src[ 4 * ( j + 3 ) ];
+                dst[ 4 * ( j + 4 ) ] = src[ 4 * ( j + 4 ) ];
+                dst[ 4 * ( j + 5 ) ] = src[ 4 * ( j + 5 ) ];
+                dst[ 4 * ( j + 6 ) ] = src[ 4 * ( j + 6 ) ];
+                dst[ 4 * ( j + 7 ) ] = src[ 4 * ( j + 7 ) ];
+                dst[ 4 * ( j + 8 ) ] = src[ 4 * ( j + 8 ) ];
+                dst[ 4 * ( j + 9 ) ] = src[ 4 * ( j + 9 ) ];
+            }
+            for( ; j < (int)( ( sizeof( NSQ_del_decs_struct ) - sizeof( ( (NSQ_del_decs_struct *) 0 )->sLPC_Q14 ) ) / ( MAX_DEL_DEC_STATES * sizeof(opus_int32) ) ); j++ ) {
+                dst[ 4 * j ] = src[ 4 * j ];
+            }
+            psSampleState[ 0 ].Q_Q10[        RDmax_ind ] = psSampleState[ 1 ].Q_Q10[        RDmin_ind ];
+            psSampleState[ 0 ].RD_Q10[       RDmax_ind ] = psSampleState[ 1 ].RD_Q10[       RDmin_ind ];
+            psSampleState[ 0 ].xq_Q14[       RDmax_ind ] = psSampleState[ 1 ].xq_Q14[       RDmin_ind ];
+            psSampleState[ 0 ].LF_AR_Q14[    RDmax_ind ] = psSampleState[ 1 ].LF_AR_Q14[    RDmin_ind ];
+            psSampleState[ 0 ].Diff_Q14[     RDmax_ind ] = psSampleState[ 1 ].Diff_Q14[     RDmin_ind ];
+            psSampleState[ 0 ].sLTP_shp_Q14[ RDmax_ind ] = psSampleState[ 1 ].sLTP_shp_Q14[ RDmin_ind ];
+            psSampleState[ 0 ].LPC_exc_Q14[  RDmax_ind ] = psSampleState[ 1 ].LPC_exc_Q14[  RDmin_ind ];
+        }
+
+        /* Write samples from winner to output and long-term filter states */
+        if( subfr > 0 || i >= decisionDelay ) {
+            pulses[ i - decisionDelay ] = (opus_int8)silk_RSHIFT_ROUND( psDelDec->Q_Q10[ last_smple_idx ][ Winner_ind ], 10 );
+            xq[ i - decisionDelay ] = (opus_int16)silk_SAT16( silk_RSHIFT_ROUND(
+                silk_SMULWW( psDelDec->Xq_Q14[ last_smple_idx ][ Winner_ind ], delayedGain_Q10[ last_smple_idx ] ), 8 ) );
+            NSQ->sLTP_shp_Q14[ NSQ->sLTP_shp_buf_idx - decisionDelay ] = psDelDec->Shape_Q14[ last_smple_idx ][ Winner_ind ];
+            sLTP_Q15[          NSQ->sLTP_buf_idx     - decisionDelay ] = psDelDec->Pred_Q15[  last_smple_idx ][ Winner_ind ];
+        }
+        NSQ->sLTP_shp_buf_idx++;
+        NSQ->sLTP_buf_idx++;
+
+        /* Update states */
+        vst1q_s32( psDelDec->LF_AR_Q14,                          vld1q_s32( psSampleState[ 0 ].LF_AR_Q14 ) );
+        vst1q_s32( psDelDec->Diff_Q14,                           vld1q_s32( psSampleState[ 0 ].Diff_Q14 ) );
+        vst1q_s32( psDelDec->sLPC_Q14[ NSQ_LPC_BUF_LENGTH + i ], vld1q_s32( psSampleState[ 0 ].xq_Q14 ) );
+        vst1q_s32( psDelDec->Xq_Q14[    *smpl_buf_idx ],         vld1q_s32( psSampleState[ 0 ].xq_Q14 ) );
+        tmp1_s32x4 = vld1q_s32( psSampleState[ 0 ].Q_Q10 );
+        vst1q_s32( psDelDec->Q_Q10[     *smpl_buf_idx ], tmp1_s32x4 );
+        vst1q_s32( psDelDec->Pred_Q15[  *smpl_buf_idx ], vshlq_n_s32( vld1q_s32( psSampleState[ 0 ].LPC_exc_Q14 ), 1 ) );
+        vst1q_s32( psDelDec->Shape_Q14[ *smpl_buf_idx ], vld1q_s32( psSampleState[ 0 ].sLTP_shp_Q14 ) );
+        tmp1_s32x4 = vrshrq_n_s32( tmp1_s32x4, 10 );
+        tmp1_s32x4 = vaddq_s32( vld1q_s32( psDelDec->Seed ), tmp1_s32x4 );
+        vst1q_s32( psDelDec->Seed,                           tmp1_s32x4 );
+        vst1q_s32( psDelDec->RandState[ *smpl_buf_idx ],     tmp1_s32x4 );
+        vst1q_s32( psDelDec->RD_Q10,                         vld1q_s32( psSampleState[ 0 ].RD_Q10 ) );
+        delayedGain_Q10[ *smpl_buf_idx ] = Gain_Q10;
+    }
+    /* Update LPC states */
+    silk_memcpy( psDelDec->sLPC_Q14[ 0 ], psDelDec->sLPC_Q14[ length ], MAX_DEL_DEC_STATES * NSQ_LPC_BUF_LENGTH * sizeof(opus_int32) );
+
+    RESTORE_STACK;
+}
+
+static OPUS_INLINE void silk_SMULWB_8_neon(
+    const opus_int16 *a,
+    const int32x2_t  b,
+    opus_int32       *o
+)
+{
+    const int16x8_t a_s16x8 = vld1q_s16( a );
+    int32x4_t       o0_s32x4, o1_s32x4;
+
+    o0_s32x4 = vshll_n_s16( vget_low_s16 ( a_s16x8 ), 15 );
+    o1_s32x4 = vshll_n_s16( vget_high_s16( a_s16x8 ), 15 );
+    o0_s32x4 = vqdmulhq_lane_s32( o0_s32x4, b, 0 );
+    o1_s32x4 = vqdmulhq_lane_s32( o1_s32x4, b, 0 );
+    vst1q_s32( o,     o0_s32x4 );
+    vst1q_s32( o + 4, o1_s32x4 );
+}
+
+/* Only works when ( b >= -65536 ) && ( b < 65536 ). */
+static OPUS_INLINE void silk_SMULWW_small_b_4_neon(
+    const opus_int32 *a,
+    const int32x2_t  b_s32x2,
+    opus_int32       *o
+)
+{
+    int32x4_t o_s32x4;
+
+    o_s32x4 = vld1q_s32( a );
+    o_s32x4 = vqdmulhq_lane_s32( o_s32x4, b_s32x2, 0 );
+    vst1q_s32( o, o_s32x4 );
+}
+
+/* Only works when ( b >= -65536 ) && ( b < 65536 ). */
+static OPUS_INLINE void silk_SMULWW_small_b_8_neon(
+    const opus_int32 *a,
+    const int32x2_t  b_s32x2,
+    opus_int32       *o
+)
+{
+    int32x4_t o0_s32x4, o1_s32x4;
+
+    o0_s32x4 = vld1q_s32( a );
+    o1_s32x4 = vld1q_s32( a + 4 );
+    o0_s32x4 = vqdmulhq_lane_s32( o0_s32x4, b_s32x2, 0 );
+    o1_s32x4 = vqdmulhq_lane_s32( o1_s32x4, b_s32x2, 0 );
+    vst1q_s32( o,     o0_s32x4 );
+    vst1q_s32( o + 4, o1_s32x4 );
+}
+
+static OPUS_INLINE void silk_SMULWW_4_neon(
+    const opus_int32 *a,
+    const int32x2_t  b_s32x2,
+    opus_int32       *o
+)
+{
+    int32x4_t a_s32x4, o_s32x4;
+
+    a_s32x4 = vld1q_s32( a );
+    o_s32x4 = vqdmulhq_lane_s32( a_s32x4, b_s32x2, 0 );
+    o_s32x4 = vmlaq_lane_s32( o_s32x4, a_s32x4, b_s32x2, 1 );
+    vst1q_s32( o, o_s32x4 );
+}
+
+static OPUS_INLINE void silk_SMULWW_8_neon(
+    const opus_int32 *a,
+    const int32x2_t  b_s32x2,
+    opus_int32       *o
+)
+{
+    int32x4_t a0_s32x4, a1_s32x4, o0_s32x4, o1_s32x4;
+
+    a0_s32x4 = vld1q_s32( a );
+    a1_s32x4 = vld1q_s32( a + 4 );
+    o0_s32x4 = vqdmulhq_lane_s32( a0_s32x4, b_s32x2, 0 );
+    o1_s32x4 = vqdmulhq_lane_s32( a1_s32x4, b_s32x2, 0 );
+    o0_s32x4 = vmlaq_lane_s32( o0_s32x4, a0_s32x4, b_s32x2, 1 );
+    o1_s32x4 = vmlaq_lane_s32( o1_s32x4, a1_s32x4, b_s32x2, 1 );
+    vst1q_s32( o,     o0_s32x4 );
+    vst1q_s32( o + 4, o1_s32x4 );
+}
+
+static OPUS_INLINE void silk_SMULWW_loop_neon(
+    const opus_int16 *a,
+    const opus_int32 b,
+    opus_int32       *o,
+    const opus_int   loop_num
+)
+{
+    opus_int i;
+    int32x2_t b_s32x2;
+
+    b_s32x2 = vdup_n_s32( b );
+    for( i = 0; i < loop_num - 7; i += 8 ) {
+        silk_SMULWB_8_neon( a + i, b_s32x2, o + i );
+    }
+    for( ; i < loop_num; i++ ) {
+        o[ i ] = silk_SMULWW( a[ i ], b );
+    }
+}
+
+static OPUS_INLINE void silk_nsq_del_dec_scale_states(
+    const silk_encoder_state *psEncC,               /* I    Encoder State                       */
+    silk_nsq_state      *NSQ,                       /* I/O  NSQ state                           */
+    NSQ_del_decs_struct psDelDec[],                 /* I/O  Delayed decision states             */
+    const opus_int16    x16[],                      /* I    Input                               */
+    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                     */
+    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, lag;
+    opus_int32          gain_adj_Q16, inv_gain_Q31, inv_gain_Q26;
+
+    lag          = pitchL[ subfr ];
+    inv_gain_Q31 = silk_INVERSE32_varQ( silk_max( Gains_Q16[ subfr ], 1 ), 47 );
+    silk_assert( inv_gain_Q31 != 0 );
+
+    /* Scale input */
+    inv_gain_Q26 = silk_RSHIFT_ROUND( inv_gain_Q31, 5 );
+    silk_SMULWW_loop_neon( x16, inv_gain_Q26, x_sc_Q10, psEncC->subfr_length );
+
+    /* 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 );
+        }
+        silk_SMULWW_loop_neon( sLTP + NSQ->sLTP_buf_idx - lag - LTP_ORDER / 2, inv_gain_Q31, sLTP_Q15 + NSQ->sLTP_buf_idx - lag - LTP_ORDER / 2, lag + LTP_ORDER / 2 );
+    }
+
+    /* Adjust for changing gain */
+    if( Gains_Q16[ subfr ] != NSQ->prev_gain_Q16 ) {
+        int32x2_t gain_adj_Q16_s32x2;
+        gain_adj_Q16 =  silk_DIV32_varQ( NSQ->prev_gain_Q16, Gains_Q16[ subfr ], 16 );
+
+        /* Scale long-term shaping state */
+        silk_assert( !( ( MAX_DEL_DEC_STATES * NSQ_LPC_BUF_LENGTH )  & 7 ) );
+        silk_assert( !( ( MAX_DEL_DEC_STATES * MAX_SHAPE_LPC_ORDER ) & 7 ) );
+        silk_assert( !( ( MAX_DEL_DEC_STATES * DECISION_DELAY )      & 7 ) );
+        if( ( gain_adj_Q16 >= -65536 ) && ( gain_adj_Q16 < 65536 ) ) {
+            gain_adj_Q16_s32x2 = vdup_n_s32( gain_adj_Q16 << 15 );
+            for( i = NSQ->sLTP_shp_buf_idx - psEncC->ltp_mem_length; i < NSQ->sLTP_shp_buf_idx - 7; i += 8 ) {
+                silk_SMULWW_small_b_8_neon( NSQ->sLTP_shp_Q14 + i, gain_adj_Q16_s32x2, NSQ->sLTP_shp_Q14 + i );
+            }
+            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 - 7; i += 8 ) {
+                    silk_SMULWW_small_b_8_neon( sLTP_Q15 + i, gain_adj_Q16_s32x2, sLTP_Q15 + i );
+                }
+                for( ; i < NSQ->sLTP_buf_idx - decisionDelay; i++ ) {
+                    sLTP_Q15[ i ] = silk_SMULWW( gain_adj_Q16, sLTP_Q15[ i ] );
+                }
+            }
+
+            /* Scale scalar states */
+            silk_SMULWW_small_b_4_neon( psDelDec->LF_AR_Q14, gain_adj_Q16_s32x2, psDelDec->LF_AR_Q14 );
+            silk_SMULWW_small_b_4_neon( psDelDec->Diff_Q14, gain_adj_Q16_s32x2, psDelDec->Diff_Q14 );
+
+            /* Scale short-term prediction and shaping states */
+            opus_int32 *state0, *state1;
+            state0 = psDelDec->sLPC_Q14[ 0 ];
+            for( i = 0; i < MAX_DEL_DEC_STATES * NSQ_LPC_BUF_LENGTH; i += 8 ) {
+                silk_SMULWW_small_b_8_neon( state0 + i, gain_adj_Q16_s32x2, state0 + i );
+            }
+            state0 = psDelDec->sAR2_Q14[ 0 ];
+            for( i = 0; i < MAX_DEL_DEC_STATES * MAX_SHAPE_LPC_ORDER; i += 8 ) {
+                silk_SMULWW_small_b_8_neon( state0 + i, gain_adj_Q16_s32x2, state0 + i );
+            }
+            state0 = psDelDec->Pred_Q15[ 0 ];
+            state1 = psDelDec->Shape_Q14[ 0 ];
+            for( i = 0; i < MAX_DEL_DEC_STATES * DECISION_DELAY; i += 8 ) {
+                silk_SMULWW_small_b_8_neon( state0 + i, gain_adj_Q16_s32x2, state0 + i );
+                silk_SMULWW_small_b_8_neon( state1 + i, gain_adj_Q16_s32x2, state1 + i );
+            }
+        }
+        else {
+            gain_adj_Q16_s32x2 = vdup_n_s32( ( gain_adj_Q16 & 0x0000FFFF ) << 15 );
+            gain_adj_Q16_s32x2 = vset_lane_s32( gain_adj_Q16 >> 16, gain_adj_Q16_s32x2, 1 );
+            for( i = NSQ->sLTP_shp_buf_idx - psEncC->ltp_mem_length; i < NSQ->sLTP_shp_buf_idx - 7; i += 8 ) {
+                silk_SMULWW_8_neon( NSQ->sLTP_shp_Q14 + i, gain_adj_Q16_s32x2, NSQ->sLTP_shp_Q14 + i );
+            }
+            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 - 7; i += 8 ) {
+                    silk_SMULWW_8_neon( sLTP_Q15 + i, gain_adj_Q16_s32x2, sLTP_Q15 + i );
+                }
+                for( ; i < NSQ->sLTP_buf_idx - decisionDelay; i++ ) {
+                    sLTP_Q15[ i ] = silk_SMULWW( gain_adj_Q16, sLTP_Q15[ i ] );
+                }
+            }
+
+            /* Scale scalar states */
+            silk_SMULWW_4_neon( psDelDec->LF_AR_Q14, gain_adj_Q16_s32x2, psDelDec->LF_AR_Q14 );
+            silk_SMULWW_4_neon( psDelDec->Diff_Q14,  gain_adj_Q16_s32x2, psDelDec->Diff_Q14 );
+
+            /* Scale short-term prediction and shaping states */
+            opus_int32 *state0, *state1;
+            state0 = psDelDec->sLPC_Q14[ 0 ];
+            for( i = 0; i < MAX_DEL_DEC_STATES * NSQ_LPC_BUF_LENGTH; i += 8 ) {
+                silk_SMULWW_8_neon( state0 + i, gain_adj_Q16_s32x2, state0 + i );
+            }
+            state0 = psDelDec->sAR2_Q14[ 0 ];
+            for( i = 0; i < MAX_DEL_DEC_STATES * MAX_SHAPE_LPC_ORDER; i += 8 ) {
+                silk_SMULWW_8_neon( state0 + i, gain_adj_Q16_s32x2, state0 + i );
+            }
+            state0 = psDelDec->Pred_Q15[ 0 ];
+            state1 = psDelDec->Shape_Q14[ 0 ];
+            for( i = 0; i < MAX_DEL_DEC_STATES * DECISION_DELAY; i += 8 ) {
+                silk_SMULWW_8_neon( state0 + i, gain_adj_Q16_s32x2, state0 + i );
+                silk_SMULWW_8_neon( state1 + i, gain_adj_Q16_s32x2, state1 + i );
+            }
+        }
+
+        /* Save inverse gain */
+        NSQ->prev_gain_Q16 = Gains_Q16[ subfr ];
+    }
+}
diff --git a/silk/arm/arm_silk_map.c b/silk/arm/arm_silk_map.c
index b1783c7..8deaf99 100644
--- a/silk/arm/arm_silk_map.c
+++ b/silk/arm/arm_silk_map.c
@@ -60,6 +60,29 @@ opus_int32 (*const SILK_LPC_INVERSE_PRED_GAIN_IMPL[OPUS_ARCHMASK + 1])( /* O   R
       MAY_HAVE_NEON(silk_LPC_inverse_pred_gain), /* Neon */
 };
 
+void  (*const SILK_NSQ_DEL_DEC_IMPL[OPUS_ARCHMASK + 1])(
+        const silk_encoder_state    *psEncC,                                    /* I    Encoder State                   */
+        silk_nsq_state              *NSQ,                                       /* I/O  NSQ state                       */
+        SideInfoIndices             *psIndices,                                 /* I/O  Quantization Indices            */
+        const opus_int16            x16[],                                      /* I    Input                           */
+        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            AR_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,              /* ARMv4 */
+      silk_NSQ_del_dec_c,              /* EDSP */
+      silk_NSQ_del_dec_c,              /* Media */
+      MAY_HAVE_NEON(silk_NSQ_del_dec), /* Neon */
+};
+
 /*There is no table for silk_noise_shape_quantizer_short_prediction because the
    NEON version takes different parameters than the C version.
   Instead RTCD is done via if statements at the call sites.
diff --git a/silk/main.h b/silk/main.h
index 13d4241..73c94fc 100644
--- a/silk/main.h
+++ b/silk/main.h
@@ -42,6 +42,10 @@ POSSIBILITY OF SUCH DAMAGE.
 #include "x86/main_sse.h"
 #endif
 
+#if (defined(OPUS_ARM_ASM) || defined(OPUS_ARM_MAY_HAVE_NEON_INTR))
+#include "arm/NSQ_del_dec_arm.h"
+#endif
+
 /* Convert Left/Right stereo signal to adaptive Mid/Side representation */
 void silk_stereo_LR_to_MS(
     stereo_enc_state            *state,                         /* I/O  State                                       */
@@ -269,7 +273,7 @@ void silk_NSQ_c(
 
 /* Noise shaping using delayed decision */
 void silk_NSQ_del_dec_c(
-    const silk_encoder_state    *psEncC,                                    /* I/O  Encoder State                   */
+    const silk_encoder_state    *psEncC,                                    /* I    Encoder State                   */
     silk_nsq_state              *NSQ,                                       /* I/O  NSQ state                       */
     SideInfoIndices             *psIndices,                                 /* I/O  Quantization Indices            */
     const opus_int16            x16[],                                      /* I    Input                           */
diff --git a/silk/mips/NSQ_del_dec_mipsr1.h b/silk/mips/NSQ_del_dec_mipsr1.h
index 3ca6464..cd70713 100644
--- a/silk/mips/NSQ_del_dec_mipsr1.h
+++ b/silk/mips/NSQ_del_dec_mipsr1.h
@@ -61,7 +61,7 @@ static inline void silk_noise_shape_quantizer_del_dec(
     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            *smpl_buf_idx,          /* I/O  Index to newest samples in buffers  */
     opus_int            decisionDelay,          /* I                                        */
     int                 arch                    /* I                                        */
 )
diff --git a/silk/tests/test_unit_optimization_NSQ_del_dec.c b/silk/tests/test_unit_optimization_NSQ_del_dec.c
new file mode 100644
index 0000000..572b06a
--- /dev/null
+++ b/silk/tests/test_unit_optimization_NSQ_del_dec.c
@@ -0,0 +1,142 @@
+/* Copyright (c) 2016 Google Inc. */
+/*
+   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
+
+#define SKIP_CONFIG_H
+
+#ifndef CUSTOM_MODES
+#define CUSTOM_MODES
+#endif
+
+#include <stdio.h>
+#include "main_FIX.h"
+#include "celt/_kiss_fft_guts.h"
+#include "silk/NSQ_del_dec.c"
+
+#define MIN_nStatesDelayedDecision 1
+
+static OPUS_INLINE void init_buffer(void* buffer, int size)
+{
+    char* tmp = (char*)buffer;
+    for(int i = 0; i < size; i++)
+    {
+        tmp[i] = rand();
+    }
+}
+
+static int test_silk_NSQ_del_dec(int arch)
+{
+    int                result = 0;
+    silk_encoder_state psEncC;
+    silk_nsq_state     NSQ_org, NSQ_opt;
+    SideInfoIndices    psIndices_org, psIndices_opt;
+    opus_int16         x16[MAX_FRAME_LENGTH];
+    opus_int8          pulses_org[MAX_FRAME_LENGTH], pulses_opt[MAX_FRAME_LENGTH];
+    opus_int16         PredCoef_Q12[ 2 * MAX_LPC_ORDER ];
+    opus_int16         LTPCoef_Q14[ LTP_ORDER * MAX_NB_SUBFR ];
+    opus_int16         AR_Q13[ MAX_NB_SUBFR * MAX_SHAPE_LPC_ORDER ];
+    opus_int           HarmShapeGain_Q14[ MAX_NB_SUBFR ];
+    opus_int           Tilt_Q14[ MAX_NB_SUBFR ];
+    opus_int32         LF_shp_Q14[ MAX_NB_SUBFR ];
+    opus_int32         Gains_Q16[ MAX_NB_SUBFR ];
+    opus_int           pitchL[ MAX_NB_SUBFR ];
+    opus_int           Lambda_Q10;
+    opus_int           LTP_scale_Q14;
+    opus_int           subfr_length;
+    opus_int           shapingLPCOrder;
+    opus_int           nStatesDelayedDecision;
+
+    printf("%50s", "silk_NSQ_del_dec() ...");
+    for( subfr_length = DECISION_DELAY; subfr_length <= MAX_SUB_FRAME_LENGTH; subfr_length++ )
+    {
+        for(nStatesDelayedDecision = MIN_nStatesDelayedDecision; nStatesDelayedDecision <= MAX_DEL_DEC_STATES; nStatesDelayedDecision++ )
+        {
+            for( shapingLPCOrder = 12; shapingLPCOrder <= MAX_SHAPE_LPC_ORDER; shapingLPCOrder += 2 ) // shapingLPCOrder must be even.
+            {
+                init_buffer(&psEncC,           sizeof(psEncC));
+                init_buffer(&NSQ_org,          sizeof(NSQ_org));
+                init_buffer(&psIndices_org,    sizeof(psIndices_org));
+                init_buffer(pulses_org,        sizeof(pulses_org));
+                init_buffer(x16,               sizeof(x16));
+                init_buffer(PredCoef_Q12,      sizeof(PredCoef_Q12));
+                init_buffer(LTPCoef_Q14,       sizeof(LTPCoef_Q14));
+                init_buffer(AR_Q13,            sizeof(AR_Q13));
+                init_buffer(HarmShapeGain_Q14, sizeof(HarmShapeGain_Q14));
+                init_buffer(Tilt_Q14,          sizeof(Tilt_Q14));
+                init_buffer(LF_shp_Q14,        sizeof(LF_shp_Q14));
+                init_buffer(Gains_Q16,         sizeof(Gains_Q16));
+
+                psEncC.subfr_length           = subfr_length;
+                psEncC.nStatesDelayedDecision = nStatesDelayedDecision;
+                psEncC.shapingLPCOrder        = shapingLPCOrder;
+                pitchL[0]                     = rand() % 289;
+                pitchL[0]                     = MAX(pitchL[0], 80); // Restrict to value range [80, 288]
+                for( int i = 1; i < MAX_NB_SUBFR; i++)
+                {
+                    // The following sub frame pitchL cannot have big difference from pitchL[0]. Otherwise sLTP_Q15[] in silk_nsq_del_dec_scale_states() will access uninitialized values.
+                    pitchL[i]  = pitchL[0];
+                    pitchL[i] += rand() % 6;
+                    pitchL[i] -= rand() % 6;
+                }
+                NSQ_org.lagPrev                 = rand() % 289;
+                NSQ_org.lagPrev                 = MAX( NSQ_org.lagPrev, 80);
+                Lambda_Q10                      = rand() % 32768;
+                LTP_scale_Q14                   = rand();
+                psEncC.predictLPCOrder          = (rand() & 1) ? MIN_LPC_ORDER : MAX_LPC_ORDER;
+                psEncC.warping_Q16              = rand() % 32767;
+                psEncC.arch                     = arch;
+                psEncC.nb_subfr                 = 4;
+                psEncC.frame_length             = MAX_FRAME_LENGTH; //psEncC.nb_subfr * subfr_length;
+                psEncC.ltp_mem_length           = psEncC.frame_length;//LTP_MEM_LENGTH_MS * 8; //???
+                psIndices_org.signalType        = rand() % (TYPE_VOICED + 1);
+                psIndices_org.quantOffsetType   = rand() & 1;
+                psIndices_org.NLSFInterpCoef_Q2 = rand() & 4;
+
+                memcpy(&NSQ_opt,       &NSQ_org,       sizeof(NSQ_org));
+                memcpy(&psIndices_opt, &psIndices_org, sizeof(psIndices_org));
+                memcpy(pulses_opt,     pulses_org,     sizeof(pulses_org));
+
+                silk_NSQ_del_dec_c(&psEncC, &NSQ_org, &psIndices_org, x16, pulses_org, PredCoef_Q12, LTPCoef_Q14, AR_Q13, HarmShapeGain_Q14, Tilt_Q14, LF_shp_Q14, Gains_Q16, pitchL, Lambda_Q10, LTP_scale_Q14);
+                silk_NSQ_del_dec  (&psEncC, &NSQ_opt, &psIndices_opt, x16, pulses_opt, PredCoef_Q12, LTPCoef_Q14, AR_Q13, HarmShapeGain_Q14, Tilt_Q14, LF_shp_Q14, Gains_Q16, pitchL, Lambda_Q10, LTP_scale_Q14, arch);
+
+                if (memcmp(&NSQ_opt,       &NSQ_org,       sizeof(NSQ_org)))       { printf("NSQ_org   different!\n"); result = -1; }
+                if (memcmp(&psIndices_opt, &psIndices_org, sizeof(psIndices_org))) { printf("psIndices different!\n"); result = -1; }
+                if (memcmp(pulses_opt,     pulses_org,     sizeof(pulses_org)))    { printf("pulses    different!\n"); result = -1; }
+                if (result)
+                {
+                    printf("subfr_length=%3d, nStatesDelayedDecision=%d, shapingLPCOrder=%2d, psEncC.predictLPCOrder=%d failed!\n", subfr_length, nStatesDelayedDecision, shapingLPCOrder, psEncC.predictLPCOrder);
+                    return result;
+                }
+            }
+        }
+    }
+
+    printf(" passed!\n");
+    return result;
+}
diff --git a/silk/x86/NSQ_del_dec_sse.c b/silk/x86/NSQ_del_dec_sse.c
index a6f84e1..29dd872 100644
--- a/silk/x86/NSQ_del_dec_sse.c
+++ b/silk/x86/NSQ_del_dec_sse.c
@@ -107,12 +107,12 @@ static OPUS_INLINE void silk_noise_shape_quantizer_del_dec_sse4_1(
     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            *smpl_buf_idx,          /* I/O  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                   */
+    const silk_encoder_state    *psEncC,                                    /* I    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        */
@@ -335,7 +335,7 @@ static OPUS_INLINE void silk_noise_shape_quantizer_del_dec_sse4_1(
     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            *smpl_buf_idx,          /* I/O  Index to newest samples in buffers  */
     opus_int            decisionDelay           /* I                                        */
 )
 {
diff --git a/silk/x86/main_sse.h b/silk/x86/main_sse.h
index a221f31..42a6c70 100644
--- a/silk/x86/main_sse.h
+++ b/silk/x86/main_sse.h
@@ -140,7 +140,7 @@ extern void (*const SILK_NSQ_IMPL[OPUS_ARCHMASK + 1])(
 #  define OVERRIDE_silk_NSQ_del_dec
 
 void silk_NSQ_del_dec_sse4_1(
-    const silk_encoder_state    *psEncC,                                    /* I/O  Encoder State                   */
+    const silk_encoder_state    *psEncC,                                    /* I    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        */
@@ -167,7 +167,7 @@ void silk_NSQ_del_dec_sse4_1(
 #else
 
 extern void (*const SILK_NSQ_DEL_DEC_IMPL[OPUS_ARCHMASK + 1])(
-    const silk_encoder_state    *psEncC,                                    /* I/O  Encoder State                   */
+    const silk_encoder_state    *psEncC,                                    /* I    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        */
diff --git a/silk/x86/x86_silk_map.c b/silk/x86/x86_silk_map.c
index 6a1d75c..e69da60 100644
--- a/silk/x86/x86_silk_map.c
+++ b/silk/x86/x86_silk_map.c
@@ -116,7 +116,7 @@ void (*const SILK_VQ_WMAT_EC_IMPL[ OPUS_ARCHMASK + 1 ] )(
 
 #if 0 /* FIXME: SSE disabled until the NSQ code gets updated. */
 void (*const SILK_NSQ_DEL_DEC_IMPL[ OPUS_ARCHMASK + 1 ] )(
-    const silk_encoder_state    *psEncC,                                    /* I/O  Encoder State                   */
+    const silk_encoder_state    *psEncC,                                    /* I    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        */
diff --git a/silk_sources.mk b/silk_sources.mk
index d2d5b35..9dcfe83 100644
--- a/silk_sources.mk
+++ b/silk_sources.mk
@@ -87,6 +87,7 @@ SILK_SOURCES_ARM_NEON_INTR = \
 silk/arm/arm_silk_map.c \
 silk/arm/LPC_analysis_filter_neon_intr.c \
 silk/arm/LPC_inv_pred_gain_neon_intr.c \
+silk/arm/NSQ_del_dec_neon_intr.c \
 silk/arm/NSQ_neon.c
 
 SILK_SOURCES_FIXED = \
diff --git a/tests/test_unit_optimization.c b/tests/test_unit_optimization.c
index 55425c4..19fda42 100644
--- a/tests/test_unit_optimization.c
+++ b/tests/test_unit_optimization.c
@@ -47,6 +47,7 @@
 
 # include "silk/tests/test_unit_optimization_LPC_analysis_filter.c"
 # include "silk/tests/test_unit_optimization_LPC_inv_pred_gain.c"
+# include "silk/tests/test_unit_optimization_NSQ_del_dec.c"
 
 #define NUM_UNIT_TEST_LOOP 10
 
@@ -68,6 +69,7 @@ int main(void)
 #endif /* FIXED_POINT */
       result |= test_silk_LPC_analysis_filter(arch);
       result |= test_silk_LPC_inverse_pred_gain(arch);
+      result |= test_silk_NSQ_del_dec(arch);
    }
    return result;
 }
-- 
2.8.0.rc3.226.g39d4020



More information about the opus mailing list