Optimize NSQ_del_dec() for AVX2

The optimization is bit-exact with C function.

This optimization speeds up SILK encoder (floating point) as following:

AMD Zen:
Complexity 0-5 :      0%
Complexity 6-7 : 3 -  7%
Complexity 8-10: 8 - 15%

Intel Skylake:
Complexity 0-5 :       0%
Complexity 6-7 : 14 - 18%
Complexity 8-10: 17 - 22%

Adapted by Jean-Marc Valin

5 files changed, 1105 insertions, 2 deletions

diff --git a/Makefile.am b/Makefile.am
index 046d3069..a3161e82 100644
--- a/Makefile.am
+++ b/Makefile.am
@@ -70,6 +70,7 @@ LPCNET_SOURCES += $(DNN_SOURCES_SSE4_1)
 endif
 endif
 if HAVE_AVX2
+SILK_SOURCES += $(SILK_SOURCES_AVX2)
 CELT_SOURCES += $(CELT_SOURCES_AVX2)
 if ENABLE_DEEP_PLC
 LPCNET_SOURCES += $(DNN_SOURCES_AVX2)
@@ -425,6 +426,7 @@ endif
 
 if HAVE_AVX2
 AVX2_OBJ = $(CELT_SOURCES_AVX2:.c=.lo) \
+           $(SILK_SOURCES_AVX2:.c=.lo) \
            $(DNN_SOURCES_AVX2:.c=.lo)
 $(AVX2_OBJ): CFLAGS += $(OPUS_X86_AVX2_CFLAGS)
 endif
diff --git a/silk/x86/NSQ_del_dec_avx2.c b/silk/x86/NSQ_del_dec_avx2.c
new file mode 100644
index 00000000..ab81e7fe
--- /dev/null
+++ b/silk/x86/NSQ_del_dec_avx2.c
@@ -0,0 +1,1072 @@
+/***********************************************************************
+Copyright (c) 2021 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.
+- Neither the name of Internet Society, IETF or IETF Trust, nor the
+names of specific contributors, may be used to endorse or promote
+products derived from this software without specific prior written
+permission.
+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
+
+#ifdef OPUS_CHECK_ASM
+#include <string.h>
+#endif
+
+#include <immintrin.h>
+
+#include "main.h"
+#include "stack_alloc.h"
+#include "NSQ.h"
+
+/* Returns TRUE if all assumptions met */
+static OPUS_INLINE int verify_assumptions(const silk_encoder_state *psEncC)
+{
+    /* This optimization is based on these assumptions        */
+    /* These assumptions are fundamental and hence assert are */
+    /* used. Should any assert triggers, we have to re-visit  */
+    /* all related code to make sure it still functions the   */
+    /* same as the C implementation.                          */
+    silk_assert(MAX_DEL_DEC_STATES  <= 4      &&
+                MAX_FRAME_LENGTH     % 4 == 0 &&
+                MAX_SUB_FRAME_LENGTH % 4 == 0 &&
+                LTP_MEM_LENGTH_MS    % 4 == 0 );
+    silk_assert(psEncC->fs_kHz ==  8 ||
+                psEncC->fs_kHz == 12 ||
+                psEncC->fs_kHz == 16 );
+    silk_assert(psEncC->nb_subfr <= MAX_NB_SUBFR &&
+                psEncC->nb_subfr > 0             );
+    silk_assert(psEncC->nStatesDelayedDecision <= MAX_DEL_DEC_STATES &&
+                psEncC->nStatesDelayedDecision > 0                   );
+    silk_assert(psEncC->ltp_mem_length == psEncC->fs_kHz * LTP_MEM_LENGTH_MS);
+
+    /* Regressions were observed on certain AMD Zen CPUs when      */
+    /* nStatesDelayedDecision is 1 or 2. Ideally we should detect  */
+    /* these CPUs and enable this optimization on others; however, */
+    /* there is no good way to do so under current OPUS framework. */
+    return psEncC->nStatesDelayedDecision == 3 ||
+           psEncC->nStatesDelayedDecision == 4;
+}
+
+/* Intrinsics not defined on MSVC */
+#ifdef _MSC_VER
+#include <Intsafe.h>
+#define __m128i_u __m128i
+static inline int __builtin_sadd_overflow(opus_int32 a, opus_int32 b, opus_int32* res)
+{
+    *res = a+b;
+    return (*res ^ a) & (*res ^ b) & 0x80000000;
+}
+static inline int __builtin_ctz(unsigned int x)
+{
+    DWORD res = 0;
+    return _BitScanForward(&res, x) ? res : 32;
+}
+#endif
+
+/*
+ * GCC implemented _mm_loadu_si32() since GCC 11; HOWEVER, there is a bug!
+ * https://gcc.gnu.org/bugzilla/show_bug.cgi?id=99754
+ */
+#if defined(__GNUC__) && !defined(__clang__)
+#define _mm_loadu_si32 WORKAROUND_mm_loadu_si32
+static inline __m128i WORKAROUND_mm_loadu_si32(void const* mem_addr)
+{
+  return _mm_set_epi32(0, 0, 0, *(int32_t*)mem_addr);
+}
+#endif
+
+static OPUS_INLINE __m128i silk_cvtepi64_epi32_high(__m256i num)
+{
+    return _mm256_castsi256_si128(_mm256_permutevar8x32_epi32(num, _mm256_set_epi32(0, 0, 0, 0, 7, 5, 3, 1)));
+}
+
+static OPUS_INLINE opus_int16 silk_sat16(opus_int32 num)
+{
+    num = num > silk_int16_MAX ? silk_int16_MAX : num;
+    num = num < silk_int16_MIN ? silk_int16_MIN : num;
+    return num;
+}
+
+static OPUS_INLINE opus_int32 silk_sar_round_32(opus_int32 a, int bits)
+{
+    silk_assert(bits > 0 && bits < 31);
+    a += 1 << (bits-1);
+    return a >> bits;
+}
+
+static OPUS_INLINE opus_int64 silk_sar_round_smulww(opus_int32 a, opus_int32 b, int bits)
+{
+    silk_assert(bits > 0 && bits < 63);
+    opus_int64 t = ((opus_int64)a) * ((opus_int64)b);
+    bits += 16;
+    t += 1ull << (bits-1);
+    return t >> bits;
+}
+
+static OPUS_INLINE opus_int32 silk_add_sat32(opus_int32 a, opus_int32 b)
+{
+    opus_int32 sum;
+    if (__builtin_sadd_overflow(a, b, &sum))
+    {
+        return a >= 0 ? silk_int32_MAX : silk_int32_MIN;
+    }
+    return sum;
+}
+
+static OPUS_INLINE __m128i silk_mm_srai_round_epi32(__m128i a, int bits)
+{
+    silk_assert(bits > 0 && bits < 31);
+    return _mm_srai_epi32(_mm_add_epi32(a, _mm_set1_epi32(1 << (bits - 1))), bits);
+}
+
+/* add/subtract with output saturated */
+static OPUS_INLINE __m128i silk_mm_add_sat_epi32(__m128i a, __m128i b)
+{
+    __m128i r = _mm_add_epi32(a, b);
+    __m128i OF = _mm_and_si128(_mm_xor_si128(a, r), _mm_xor_si128(b, r));           /* OF = (sum ^ a) & (sum ^ b)   */
+    __m128i SAT = _mm_add_epi32(_mm_srli_epi32(a, 31), _mm_set1_epi32(0x7FFFFFFF)); /* SAT = (a >> 31) + 0x7FFFFFFF */
+    return _mm_blendv_epi8(r, SAT, _mm_srai_epi32(OF, 31));
+}
+static OPUS_INLINE __m128i silk_mm_sub_sat_epi32(__m128i a, __m128i b)
+{
+    __m128i r = _mm_sub_epi32(a, b);
+    __m128i OF = _mm_andnot_si128(_mm_xor_si128(b, r), _mm_xor_si128(a, r));        /* OF = (sum ^ a) & (sum ^ ~b) = (sum ^ a) & ~(sum ^ b) */
+    __m128i SAT = _mm_add_epi32(_mm_srli_epi32(a, 31), _mm_set1_epi32(0x7FFFFFFF)); /* SAT = (a >> 31) + 0x7FFFFFFF                         */
+    return _mm_blendv_epi8(r, SAT, _mm_srai_epi32(OF, 31));
+}
+static OPUS_INLINE __m256i silk_mm256_sub_sat_epi32(__m256i a, __m256i b)
+{
+    __m256i r = _mm256_sub_epi32(a, b);
+    __m256i OF = _mm256_andnot_si256(_mm256_xor_si256(b, r), _mm256_xor_si256(a, r));        /* OF = (sum ^ a) & (sum ^ ~b) = (sum ^ a) & ~(sum ^ b) */
+    __m256i SAT = _mm256_add_epi32(_mm256_srli_epi32(a, 31), _mm256_set1_epi32(0x7FFFFFFF)); /* SAT = (a >> 31) + 0x7FFFFFFF                         */
+    return _mm256_blendv_epi8(r, SAT, _mm256_srai_epi32(OF, 31));
+}
+
+static OPUS_INLINE __m128i silk_mm_limit_epi32(__m128i num, opus_int32 limit1, opus_int32 limit2)
+{
+    opus_int32 lo = limit1 < limit2 ? limit1 : limit2;
+    opus_int32 hi = limit1 > limit2 ? limit1 : limit2;
+
+    num = _mm_min_epi32(num, _mm_set1_epi32(hi));
+    num = _mm_max_epi32(num, _mm_set1_epi32(lo));
+    return num;
+}
+
+/* cond < 0 ? -num : num */
+static OPUS_INLINE __m128i silk_mm_sign_epi32(__m128i num, __m128i cond)
+{
+    return _mm_sign_epi32(num, _mm_or_si128(cond, _mm_set1_epi32(1)));
+}
+static OPUS_INLINE __m256i silk_mm256_sign_epi32(__m256i num, __m256i cond)
+{
+    return _mm256_sign_epi32(num, _mm256_or_si256(cond, _mm256_set1_epi32(1)));
+}
+
+/* (a32 * b32) >> 16 */
+static OPUS_INLINE __m128i silk_mm_smulww_epi32(__m128i a, opus_int32 b)
+{
+    return silk_cvtepi64_epi32_high(_mm256_slli_epi64(_mm256_mul_epi32(_mm256_cvtepi32_epi64(a), _mm256_set1_epi32(b)), 16));
+}
+
+/* (a32 * (opus_int32)((opus_int16)(b32))) >> 16 output have to be 32bit int */
+static OPUS_INLINE __m128i silk_mm_smulwb_epi32(__m128i a, opus_int32 b)
+{
+    return silk_cvtepi64_epi32_high(_mm256_mul_epi32(_mm256_cvtepi32_epi64(a), _mm256_set1_epi32(b << 16)));
+}
+
+/* (opus_int32)((opus_int16)(a3))) * (opus_int32)((opus_int16)(b32)) output have to be 32bit int */
+static OPUS_INLINE __m256i silk_mm256_smulbb_epi32(__m256i a, __m256i b)
+{
+    __m256i msk = _mm256_set_epi8(
+        0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 13, 12, 9, 8, 5, 4, 1, 0,
+        0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 13, 12, 9, 8, 5, 4, 1, 0);
+    __m256i lo = _mm256_mullo_epi16(a, b);
+    __m256i hi = _mm256_mulhi_epi16(a, b);
+    lo = _mm256_shuffle_epi8(lo, msk);
+    hi = _mm256_shuffle_epi8(hi, msk);
+    return _mm256_unpacklo_epi16(lo, hi);
+}
+
+static OPUS_INLINE __m256i silk_mm256_reverse_epi32(__m256i v)
+{
+    v = _mm256_shuffle_epi32(v, 0x1B);
+    v = _mm256_permute4x64_epi64(v, 0x4E);
+    return v;
+}
+
+static OPUS_INLINE opus_int32 silk_mm256_hsum_epi32(__m256i v)
+{
+    __m128i sum = _mm_add_epi32(_mm256_extracti128_si256(v, 1), _mm256_extracti128_si256(v, 0));
+    sum = _mm_add_epi32(sum, _mm_shuffle_epi32(sum, 0x4E));
+    sum = _mm_add_epi32(sum, _mm_shuffle_epi32(sum, 0xB1));
+    return _mm_cvtsi128_si32(sum);
+}
+
+static OPUS_INLINE __m128i silk_mm_hmin_epi32(__m128i num)
+{
+    num = _mm_min_epi32(num, _mm_shuffle_epi32(num, 0x4E)); /* 0123 -> 2301 */
+    num = _mm_min_epi32(num, _mm_shuffle_epi32(num, 0xB1)); /* 0123 -> 1032 */
+    return num;
+}
+
+static OPUS_INLINE __m128i silk_mm_hmax_epi32(__m128i num)
+{
+    num = _mm_max_epi32(num, _mm_shuffle_epi32(num, 0x4E)); /* 0123 -> 2310 */
+    num = _mm_max_epi32(num, _mm_shuffle_epi32(num, 0xB1)); /* 0123 -> 1032 */
+    return num;
+}
+
+static OPUS_INLINE __m128i silk_mm_mask_hmin_epi32(__m128i num, __m128i mask)
+{
+    num = _mm_blendv_epi8(num, _mm_set1_epi32(silk_int32_MAX), mask);
+    return silk_mm_hmin_epi32(num);
+}
+
+static OPUS_INLINE __m128i silk_mm_mask_hmax_epi32(__m128i num, __m128i mask)
+{
+    num = _mm_blendv_epi8(num, _mm_set1_epi32(silk_int32_MIN), mask);
+    return silk_mm_hmax_epi32(num);
+}
+
+static OPUS_INLINE __m128i silk_mm256_rand_epi32(__m128i seed)
+{
+    seed = _mm_mullo_epi32(seed, _mm_set1_epi32(RAND_MULTIPLIER));
+    seed = _mm_add_epi32(seed, _mm_set1_epi32(RAND_INCREMENT));
+    return seed;
+}
+
+static OPUS_INLINE opus_int32 silk_index_of_first_equal_epi32(__m128i a, __m128i b)
+{
+    unsigned int mask = _mm_movemask_epi8(_mm_cmpeq_epi32(a, b)) & 0x1111;
+    silk_assert(mask != 0);
+    return __builtin_ctz(mask) >> 2;
+}
+
+static __m128i silk_index_to_selector(opus_int32 index)
+{
+    silk_assert(index < 4);
+    index <<= 2;
+    return _mm_set_epi8(
+        index + 3, index + 2, index + 1, index + 0,
+        index + 3, index + 2, index + 1, index + 0,
+        index + 3, index + 2, index + 1, index + 0,
+        index + 3, index + 2, index + 1, index + 0);
+}
+
+static opus_int32 silk_select_winner(__m128i num, __m128i selector)
+{
+    return _mm_cvtsi128_si32(_mm_shuffle_epi8(num, selector));
+}
+
+typedef struct
+{
+    __m128i RandState;
+    __m128i Q_Q10;
+    __m128i Xq_Q14;
+    __m128i Pred_Q15;
+    __m128i Shape_Q14;
+} NSQ_del_dec_sample_struct;
+
+typedef struct
+{
+    __m128i sLPC_Q14[MAX_SUB_FRAME_LENGTH + NSQ_LPC_BUF_LENGTH];
+    __m128i LF_AR_Q14;
+    __m128i Seed;
+    __m128i SeedInit;
+    __m128i RD_Q10;
+    __m128i Diff_Q14;
+    __m128i sAR2_Q14[MAX_SHAPE_LPC_ORDER];
+    NSQ_del_dec_sample_struct Samples[DECISION_DELAY];
+} NSQ_del_dec_struct;
+
+static OPUS_INLINE void silk_nsq_del_dec_scale_states_avx2(
+    const silk_encoder_state *psEncC,          /* I    Encoder State                   */
+    silk_nsq_state *NSQ,                       /* I/O  NSQ state                       */
+    NSQ_del_dec_struct *psDelDec,              /* I/O  Delayed decision states         */
+    const opus_int16 x16[],                    /* I    Input                           */
+    opus_int32 x_sc_Q10[MAX_SUB_FRAME_LENGTH], /* 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                  */
+);
+
+/*******************************************/
+/* LPC analysis filter                     */
+/* NB! State is kept internally and the    */
+/* filter always starts with zero state    */
+/* first d output samples are set to zero  */
+/*******************************************/
+static OPUS_INLINE void silk_LPC_analysis_filter_avx2(
+    opus_int16                  *out,               /* O    Output signal                           */
+    const opus_int16            *in,                /* I    Input signal                            */
+    const opus_int16            *B,                 /* I    MA prediction coefficients, Q12 [order] */
+    const opus_int32            len,                /* I    Signal length                           */
+    const opus_int32            order               /* I    Filter order                            */
+);
+
+/******************************************/
+/* Noise shape quantizer for one subframe */
+/******************************************/
+static OPUS_INLINE void silk_noise_shape_quantizer_del_dec_avx2(
+    silk_nsq_state *NSQ,                        /* I/O  NSQ state                          */
+    NSQ_del_dec_struct psDelDec[],              /* I/O  Delayed decision states            */
+    opus_int signalType,                        /* I    Signal type                        */
+    const opus_int32 x_Q10[],                   /* I                                       */
+    opus_int8 pulses[],                         /* O                                       */
+    opus_int16 xq[],                            /* O                                       */
+    opus_int32 sLTP_Q15[],                      /* I/O  LTP filter state                   */
+    opus_int32 delayedGain_Q10[DECISION_DELAY], /* 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                                       */
+    __m128i MaskDelDec,                         /* I    Mask of states in decision tree    */
+    opus_int *smpl_buf_idx,                     /* I/O  Index to newest samples in buffers */
+    opus_int decisionDelay                      /* I                                       */
+);
+
+void silk_NSQ_del_dec_avx2(
+    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_int32 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           */
+)
+{
+#ifdef OPUS_CHECK_ASM
+    silk_nsq_state NSQ_c;
+    SideInfoIndices psIndices_c;
+    opus_int8 pulses_c[MAX_FRAME_LENGTH];
+    const opus_int8 *const pulses_a = pulses;
+
+    silk_memcpy(&NSQ_c, NSQ, sizeof(NSQ_c));
+    silk_memcpy(&psIndices_c, psIndices, sizeof(psIndices_c));
+    silk_memcpy(pulses_c, pulses, sizeof(pulses_c));
+    silk_NSQ_del_dec_c(psEncC, &NSQ_c, &psIndices_c, x16, pulses_c, PredCoef_Q12, LTPCoef_Q14, AR_Q13, HarmShapeGain_Q14, Tilt_Q14, LF_shp_Q14, Gains_Q16,
+                       pitchL, Lambda_Q10, LTP_scale_Q14);
+#endif
+
+    if (!verify_assumptions(psEncC))
+    {
+        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);
+        return;
+    }
+
+    opus_int i, k, lag, start_idx, LSF_interpolation_flag, Winner_ind, subfr;
+    opus_int last_smple_idx, smpl_buf_idx, decisionDelay;
+    const opus_int16 *A_Q12, *B_Q14, *AR_shp_Q13;
+    opus_int16 *pxq;
+    VARDECL(opus_int32, sLTP_Q15);
+    VARDECL(opus_int16, sLTP);
+    opus_int32 HarmShapeFIRPacked_Q14;
+    opus_int offset_Q10;
+    opus_int32 Gain_Q10;
+    opus_int32 x_sc_Q10[MAX_SUB_FRAME_LENGTH];
+    opus_int32 delayedGain_Q10[DECISION_DELAY];
+    NSQ_del_dec_struct psDelDec = {0};
+    NSQ_del_dec_sample_struct *psSample;
+    __m128i RDmin_Q10, MaskDelDec, Winner_selector;
+    SAVE_STACK;
+
+    MaskDelDec = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(0xFFFFFF00ul << ((psEncC->nStatesDelayedDecision - 1) << 3)));
+
+    /* Set unvoiced lag to the previous one, overwrite later for voiced */
+    lag = NSQ->lagPrev;
+
+    silk_assert(NSQ->prev_gain_Q16 != 0);
+    psDelDec.Seed = _mm_and_si128(
+        _mm_add_epi32(_mm_set_epi32(3, 2, 1, 0), _mm_set1_epi32(psIndices->Seed)),
+        _mm_set1_epi32(3));
+    psDelDec.SeedInit = psDelDec.Seed;
+    psDelDec.RD_Q10 = _mm_setzero_si128();
+    psDelDec.LF_AR_Q14 = _mm_set1_epi32(NSQ->sLF_AR_shp_Q14);
+    psDelDec.Diff_Q14 = _mm_set1_epi32(NSQ->sDiff_shp_Q14);
+    psDelDec.Samples[0].Shape_Q14 = _mm_set1_epi32(NSQ->sLTP_shp_Q14[psEncC->ltp_mem_length - 1]);
+    for (i = 0; i < NSQ_LPC_BUF_LENGTH; i++)
+    {
+        psDelDec.sLPC_Q14[i] = _mm_set1_epi32(NSQ->sLPC_Q14[i]);
+    }
+    for (i = 0; i < MAX_SHAPE_LPC_ORDER; i++)
+    {
+        psDelDec.sAR2_Q14[i] = _mm_set1_epi32(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)
+    {
+        for (k = 0; k < psEncC->nb_subfr; k++)
+        {
+            decisionDelay = silk_min_int(decisionDelay, pitchL[k] - LTP_ORDER / 2 - 1);
+        }
+    }
+    else
+    {
+        if (lag > 0)
+        {
+            decisionDelay = silk_min_int(decisionDelay, lag - LTP_ORDER / 2 - 1);
+        }
+    }
+
+    if (psIndices->NLSFInterpCoef_Q2 == 4)
+    {
+        LSF_interpolation_flag = 0;
+    }
+    else
+    {
+        LSF_interpolation_flag = 1;
+    }
+
+    ALLOC(sLTP_Q15, psEncC->ltp_mem_length + psEncC->frame_length, opus_int32);
+    ALLOC(sLTP, psEncC->ltp_mem_length + psEncC->frame_length, opus_int16);
+    /* 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 = HarmShapeGain_Q14[k] >> 2;
+        HarmShapeFIRPacked_Q14 |= ((opus_int32)(HarmShapeGain_Q14[k] >> 1)) << 16;
+
+        NSQ->rewhite_flag = 0;
+        if (psIndices->signalType == TYPE_VOICED)
+        {
+            /* Voiced */
+            lag = pitchL[k];
+
+            /* Re-whitening */
+            if ((k & (3 ^ (LSF_interpolation_flag << 1))) == 0)
+            {
+                if (k == 2)
+                {
+                    /* RESET DELAYED DECISIONS */
+                    /* Find winner */
+                    RDmin_Q10 = silk_mm_mask_hmin_epi32(psDelDec.RD_Q10, MaskDelDec);
+                    Winner_ind = silk_index_of_first_equal_epi32(RDmin_Q10, psDelDec.RD_Q10);
+                    Winner_selector = silk_index_to_selector(Winner_ind);
+                    psDelDec.RD_Q10 = _mm_add_epi32(
+                        psDelDec.RD_Q10,
+                        _mm_blendv_epi8(
+                            _mm_set1_epi32(silk_int32_MAX >> 4),
+                            _mm_setzero_si128(),
+                            _mm_cvtepi8_epi32(_mm_cvtsi32_si128(0xFF << (Winner_ind << 3)))));
+
+                    /* Copy final part of signals from winner state to output and long-term filter states */
+                    last_smple_idx = smpl_buf_idx + decisionDelay;
+                    for (i = 0; i < decisionDelay; i++)
+                    {
+                        last_smple_idx = (last_smple_idx + DECISION_DELAY - 1) % DECISION_DELAY;
+                        psSample = &psDelDec.Samples[last_smple_idx];
+                        pulses[i - decisionDelay] =
+                            (opus_int8)silk_sar_round_32(silk_select_winner(psSample->Q_Q10, Winner_selector), 10);
+                        pxq[i - decisionDelay] =
+                            silk_sat16((opus_int32)silk_sar_round_smulww(silk_select_winner(psSample->Xq_Q14, Winner_selector), Gains_Q16[1], 14));
+                        NSQ->sLTP_shp_Q14[NSQ->sLTP_shp_buf_idx - decisionDelay + i] =
+                            silk_select_winner(psSample->Shape_Q14, Winner_selector);
+                    }
+
+                    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_avx2(&sLTP[start_idx], &NSQ->xq[start_idx + k * psEncC->subfr_length],
+                                              A_Q12, psEncC->ltp_mem_length - start_idx, psEncC->predictLPCOrder);
+
+                NSQ->sLTP_buf_idx = psEncC->ltp_mem_length;
+                NSQ->rewhite_flag = 1;
+            }
+        }
+
+        silk_nsq_del_dec_scale_states_avx2(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_avx2(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, MaskDelDec, &smpl_buf_idx, decisionDelay);
+
+        x16 += psEncC->subfr_length;
+        pulses += psEncC->subfr_length;
+        pxq += psEncC->subfr_length;
+    }
+
+    /* Find winner */
+    RDmin_Q10 = silk_mm_mask_hmin_epi32(psDelDec.RD_Q10, MaskDelDec);
+    Winner_selector = silk_index_to_selector(silk_index_of_first_equal_epi32(RDmin_Q10, psDelDec.RD_Q10));
+
+    /* Copy final part of signals from winner state to output and long-term filter states */
+    psIndices->Seed = silk_select_winner(psDelDec.SeedInit, Winner_selector);
+    last_smple_idx = smpl_buf_idx + decisionDelay;
+    Gain_Q10 = Gains_Q16[psEncC->nb_subfr - 1] >> 6;
+    for (i = 0; i < decisionDelay; i++)
+    {
+        last_smple_idx = (last_smple_idx + DECISION_DELAY - 1) % DECISION_DELAY;
+        psSample = &psDelDec.Samples[last_smple_idx];
+
+        pulses[i - decisionDelay] =
+            (opus_int8)silk_sar_round_32(silk_select_winner(psSample->Q_Q10, Winner_selector), 10);
+        pxq[i - decisionDelay] =
+            silk_sat16((opus_int32)silk_sar_round_smulww(silk_select_winner(psSample->Xq_Q14, Winner_selector), Gain_Q10, 8));
+        NSQ->sLTP_shp_Q14[NSQ->sLTP_shp_buf_idx - decisionDelay + i] =
+            silk_select_winner(psSample->Shape_Q14, Winner_selector);
+    }
+    for (i = 0; i < NSQ_LPC_BUF_LENGTH; i++)
+    {
+        NSQ->sLPC_Q14[i] = silk_select_winner(psDelDec.sLPC_Q14[i], Winner_selector);
+    }
+    for (i = 0; i < MAX_SHAPE_LPC_ORDER; i++)
+    {
+        NSQ->sAR2_Q14[i] = silk_select_winner(psDelDec.sAR2_Q14[i], Winner_selector);
+    }
+
+    /* Update states */
+    NSQ->sLF_AR_shp_Q14 = silk_select_winner(psDelDec.LF_AR_Q14, Winner_selector);
+    NSQ->sDiff_shp_Q14 = silk_select_winner(psDelDec.Diff_Q14, Winner_selector);
+    NSQ->lagPrev = pitchL[psEncC->nb_subfr - 1];
+
+    /* Save quantized speech signal */
+    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;
+
+#ifdef OPUS_CHECK_ASM
+    silk_assert(!memcmp(&NSQ_c, NSQ, sizeof(NSQ_c)));
+    silk_assert(!memcmp(&psIndices_c, psIndices, sizeof(psIndices_c)));
+    silk_assert(!memcmp(pulses_c, pulses_a, sizeof(pulses_c)));
+#endif
+}
+
+static OPUS_INLINE __m128i silk_noise_shape_quantizer_short_prediction_x4(const __m128i *buf32, const opus_int16 *coef16, opus_int order)
+{
+    __m256i out;
+    silk_assert(order == 10 || order == 16);
+
+    /* Avoids introducing a bias because silk_SMLAWB() always rounds to -inf */
+    out = _mm256_set1_epi32(order >> 1);
+    out = _mm256_add_epi32(out, _mm256_mul_epi32(_mm256_cvtepi32_epi64(buf32[-0]), _mm256_set1_epi32(coef16[0] << 16))); /* High DWORD */
+    out = _mm256_add_epi32(out, _mm256_mul_epi32(_mm256_cvtepi32_epi64(buf32[-1]), _mm256_set1_epi32(coef16[1] << 16))); /* High DWORD */
+    out = _mm256_add_epi32(out, _mm256_mul_epi32(_mm256_cvtepi32_epi64(buf32[-2]), _mm256_set1_epi32(coef16[2] << 16))); /* High DWORD */
+    out = _mm256_add_epi32(out, _mm256_mul_epi32(_mm256_cvtepi32_epi64(buf32[-3]), _mm256_set1_epi32(coef16[3] << 16))); /* High DWORD */
+    out = _mm256_add_epi32(out, _mm256_mul_epi32(_mm256_cvtepi32_epi64(buf32[-4]), _mm256_set1_epi32(coef16[4] << 16))); /* High DWORD */
+    out = _mm256_add_epi32(out, _mm256_mul_epi32(_mm256_cvtepi32_epi64(buf32[-5]), _mm256_set1_epi32(coef16[5] << 16))); /* High DWORD */
+    out = _mm256_add_epi32(out, _mm256_mul_epi32(_mm256_cvtepi32_epi64(buf32[-6]), _mm256_set1_epi32(coef16[6] << 16))); /* High DWORD */
+    out = _mm256_add_epi32(out, _mm256_mul_epi32(_mm256_cvtepi32_epi64(buf32[-7]), _mm256_set1_epi32(coef16[7] << 16))); /* High DWORD */
+    out = _mm256_add_epi32(out, _mm256_mul_epi32(_mm256_cvtepi32_epi64(buf32[-8]), _mm256_set1_epi32(coef16[8] << 16))); /* High DWORD */
+    out = _mm256_add_epi32(out, _mm256_mul_epi32(_mm256_cvtepi32_epi64(buf32[-9]), _mm256_set1_epi32(coef16[9] << 16))); /* High DWORD */
+
+    if (order == 16)
+    {
+        out = _mm256_add_epi32(out, _mm256_mul_epi32(_mm256_cvtepi32_epi64(buf32[-10]), _mm256_set1_epi32(coef16[10] << 16))); /* High DWORD */
+        out = _mm256_add_epi32(out, _mm256_mul_epi32(_mm256_cvtepi32_epi64(buf32[-11]), _mm256_set1_epi32(coef16[11] << 16))); /* High DWORD */
+        out = _mm256_add_epi32(out, _mm256_mul_epi32(_mm256_cvtepi32_epi64(buf32[-12]), _mm256_set1_epi32(coef16[12] << 16))); /* High DWORD */
+        out = _mm256_add_epi32(out, _mm256_mul_epi32(_mm256_cvtepi32_epi64(buf32[-13]), _mm256_set1_epi32(coef16[13] << 16))); /* High DWORD */
+        out = _mm256_add_epi32(out, _mm256_mul_epi32(_mm256_cvtepi32_epi64(buf32[-14]), _mm256_set1_epi32(coef16[14] << 16))); /* High DWORD */
+        out = _mm256_add_epi32(out, _mm256_mul_epi32(_mm256_cvtepi32_epi64(buf32[-15]), _mm256_set1_epi32(coef16[15] << 16))); /* High DWORD */
+    }
+    return silk_cvtepi64_epi32_high(out);
+}
+
+/******************************************/
+/* Noise shape quantizer for one subframe */
+/******************************************/
+static OPUS_INLINE void silk_noise_shape_quantizer_del_dec_avx2(
+    silk_nsq_state *NSQ,                        /* I/O  NSQ state                          */
+    NSQ_del_dec_struct *psDelDec,               /* I/O  Delayed decision states            */
+    opus_int signalType,                        /* I    Signal type                        */
+    const opus_int32 x_Q10[],                   /* I                                       */
+    opus_int8 pulses[],                         /* O                                       */
+    opus_int16 xq[],                            /* O                                       */
+    opus_int32 sLTP_Q15[],                      /* I/O  LTP filter state                   */
+    opus_int32 delayedGain_Q10[DECISION_DELAY], /* 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                                       */
+    __m128i MaskDelDec,                         /* I    Mask of states in decision tree    */
+    opus_int *smpl_buf_idx,                     /* I/O  Index to newest samples in buffers */
+    opus_int decisionDelay                      /* I                                       */
+)
+{
+    opus_int32 *shp_lag_ptr = &NSQ->sLTP_shp_Q14[NSQ->sLTP_shp_buf_idx - lag + HARM_SHAPE_FIR_TAPS / 2];
+    opus_int32 *pred_lag_ptr = &sLTP_Q15[NSQ->sLTP_buf_idx - lag + LTP_ORDER / 2];
+    opus_int32 Gain_Q10 = Gain_Q16 >> 6;
+
+    for (int i = 0; i < length; i++)
+    {
+        /* Perform common calculations used in all states */
+        /* NSQ_sample_struct */
+        /* Low  128 bits => 1st set */
+        /* High 128 bits => 2nd set */
+        __m256i SS_Q_Q10;
+        __m256i SS_RD_Q10;
+        __m256i SS_xq_Q14;
+        __m256i SS_LF_AR_Q14;
+        __m256i SS_Diff_Q14;
+        __m256i SS_sLTP_shp_Q14;
+        __m256i SS_LPC_exc_Q14;
+        __m256i exc_Q14;
+        __m256i q_Q10, rr_Q10, rd_Q10;
+        __m256i mask;
+        __m128i LPC_pred_Q14, n_AR_Q14;
+        __m128i RDmin_Q10, RDmax_Q10;
+        __m128i n_LF_Q14;
+        __m128i r_Q10, q1_Q0, q1_Q10, q2_Q10;
+        __m128i Winner_rand_state, Winner_selector;
+        __m128i tmp0, tmp1;
+        NSQ_del_dec_sample_struct *psLastSample, *psSample;
+        opus_int32 RDmin_ind, RDmax_ind, last_smple_idx;
+        opus_int32 LTP_pred_Q14, n_LTP_Q14;
+
+        /* 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_SMULWB(pred_lag_ptr[-0], b_Q14[0]);
+            LTP_pred_Q14 += silk_SMULWB(pred_lag_ptr[-1], b_Q14[1]);
+            LTP_pred_Q14 += silk_SMULWB(pred_lag_ptr[-2], b_Q14[2]);
+            LTP_pred_Q14 += silk_SMULWB(pred_lag_ptr[-3], b_Q14[3]);
+            LTP_pred_Q14 += silk_SMULWB(pred_lag_ptr[-4], b_Q14[4]);
+            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_add_sat32(shp_lag_ptr[0], shp_lag_ptr[-2]);
+            n_LTP_Q14 = silk_SMULWB(n_LTP_Q14, HarmShapeFIRPacked_Q14);
+            n_LTP_Q14 = n_LTP_Q14 + silk_SMULWT(shp_lag_ptr[-1], HarmShapeFIRPacked_Q14);
+            n_LTP_Q14 = LTP_pred_Q14 - (n_LTP_Q14 << 2); /* Q12 -> Q14 */
+            shp_lag_ptr++;
+        }
+        else
+        {
+            n_LTP_Q14 = 0;
+        }
+
+        /* BEGIN Updating Delayed Decision States */
+
+        /* Generate dither */
+        psDelDec->Seed = silk_mm256_rand_epi32(psDelDec->Seed);
+
+        /* Short-term prediction */
+        LPC_pred_Q14 = silk_noise_shape_quantizer_short_prediction_x4(&psDelDec->sLPC_Q14[NSQ_LPC_BUF_LENGTH - 1 + i], a_Q12, predictLPCOrder);
+        LPC_pred_Q14 = _mm_slli_epi32(LPC_pred_Q14, 4); /* Q10 -> Q14 */
+
+        /* Noise shape feedback */
+        silk_assert(shapingLPCOrder > 0);
+        silk_assert((shapingLPCOrder & 1) == 0); /* check that order is even */
+        /* Output of lowpass section */
+        tmp0 = _mm_add_epi32(psDelDec->Diff_Q14, silk_mm_smulwb_epi32(psDelDec->sAR2_Q14[0], warping_Q16));
+        n_AR_Q14 = _mm_set1_epi32(shapingLPCOrder >> 1);
+        for (int j = 0; j < shapingLPCOrder - 1; j++)
+        {
+            /* Output of allpass section */
+            tmp1 = psDelDec->sAR2_Q14[j];
+            psDelDec->sAR2_Q14[j] = tmp0;
+            n_AR_Q14 = _mm_add_epi32(n_AR_Q14, silk_mm_smulwb_epi32(tmp0, AR_shp_Q13[j]));
+            tmp0 = _mm_add_epi32(tmp1, silk_mm_smulwb_epi32(_mm_sub_epi32(psDelDec->sAR2_Q14[j + 1], tmp0), warping_Q16));
+        }
+        psDelDec->sAR2_Q14[shapingLPCOrder - 1] = tmp0;
+        n_AR_Q14 = _mm_add_epi32(n_AR_Q14, silk_mm_smulwb_epi32(tmp0, AR_shp_Q13[shapingLPCOrder - 1]));
+
+        n_AR_Q14 = _mm_slli_epi32(n_AR_Q14, 1);                                                  /* Q11 -> Q12 */
+        n_AR_Q14 = _mm_add_epi32(n_AR_Q14, silk_mm_smulwb_epi32(psDelDec->LF_AR_Q14, Tilt_Q14)); /* Q12 */
+        n_AR_Q14 = _mm_slli_epi32(n_AR_Q14, 2);                                                  /* Q12 -> Q14 */
+
+        tmp0 = silk_mm_smulwb_epi32(psDelDec->Samples[*smpl_buf_idx].Shape_Q14, LF_shp_Q14); /* Q12 */
+        tmp1 = silk_mm_smulwb_epi32(psDelDec->LF_AR_Q14, LF_shp_Q14 >> 16);                  /* Q12 */
+        n_LF_Q14 = _mm_add_epi32(tmp0, tmp1);                                                /* Q12 */
+        n_LF_Q14 = _mm_slli_epi32(n_LF_Q14, 2);                                              /* Q12 -> Q14 */
+
+        /* Input minus prediction plus noise feedback                       */
+        /* r = x[ i ] - LTP_pred - LPC_pred + n_AR + n_Tilt + n_LF + n_LTP  */
+        tmp0 = silk_mm_add_sat_epi32(n_AR_Q14, n_LF_Q14);              /* Q14 */
+        tmp1 = _mm_add_epi32(_mm_set1_epi32(n_LTP_Q14), LPC_pred_Q14); /* Q13 */
+        tmp0 = silk_mm_sub_sat_epi32(tmp1, tmp0);                      /* Q13 */
+        tmp0 = silk_mm_srai_round_epi32(tmp0, 4);                      /* Q10 */
+
+        r_Q10 = _mm_sub_epi32(_mm_set1_epi32(x_Q10[i]), tmp0); /* residual error Q10 */
+
+        /* Flip sign depending on dither */
+        r_Q10 = silk_mm_sign_epi32(r_Q10, psDelDec->Seed);
+        r_Q10 = silk_mm_limit_epi32(r_Q10, -(31 << 10), 30 << 10);
+
+        /* Find two quantization level candidates and measure their rate-distortion */
+        q1_Q10 = _mm_sub_epi32(r_Q10, _mm_set1_epi32(offset_Q10));
+        q1_Q0 = _mm_srai_epi32(q1_Q10, 10);
+        if (Lambda_Q10 > 2048)
+        {
+            /* For aggressive RDO, the bias becomes more than one pulse. */
+            tmp0 = _mm_sub_epi32(_mm_abs_epi32(q1_Q10), _mm_set1_epi32(Lambda_Q10 / 2 - 512)); /* rdo_offset */
+            q1_Q0 = _mm_srai_epi32(q1_Q10, 31);
+            tmp1 = _mm_cmpgt_epi32(tmp0, _mm_setzero_si128());
+            tmp0 = _mm_srai_epi32(silk_mm_sign_epi32(tmp0, q1_Q10), 10);
+            q1_Q0 = _mm_blendv_epi8(q1_Q0, tmp0, tmp1);
+        }
+
+        tmp0 = _mm_sign_epi32(_mm_set1_epi32(QUANT_LEVEL_ADJUST_Q10), q1_Q0);
+        q1_Q10 = _mm_sub_epi32(_mm_slli_epi32(q1_Q0, 10), tmp0);
+        q1_Q10 = _mm_add_epi32(q1_Q10, _mm_set1_epi32(offset_Q10));
+
+        /* check if q1_Q0 is 0 or -1 */
+        tmp0 = _mm_add_epi32(_mm_srli_epi32(q1_Q0, 31), q1_Q0);
+        tmp1 = _mm_cmpeq_epi32(tmp0, _mm_setzero_si128());
+        tmp0 = _mm_blendv_epi8(_mm_set1_epi32(1024), _mm_set1_epi32(1024 - QUANT_LEVEL_ADJUST_Q10), tmp1);
+        q2_Q10 = _mm_add_epi32(q1_Q10, tmp0);
+        q_Q10 = _mm256_set_m128i(q2_Q10, q1_Q10);
+
+        rr_Q10 = _mm256_sub_epi32(_mm256_broadcastsi128_si256(r_Q10), q_Q10);
+        rd_Q10 = _mm256_abs_epi32(q_Q10);
+        rr_Q10 = silk_mm256_smulbb_epi32(rr_Q10, rr_Q10);
+        rd_Q10 = silk_mm256_smulbb_epi32(rd_Q10, _mm256_set1_epi32(Lambda_Q10));
+        rd_Q10 = _mm256_add_epi32(rd_Q10, rr_Q10);
+        rd_Q10 = _mm256_srai_epi32(rd_Q10, 10);
+
+        mask = _mm256_broadcastsi128_si256(_mm_cmplt_epi32(_mm256_extracti128_si256(rd_Q10, 0), _mm256_extracti128_si256(rd_Q10, 1)));
+        SS_RD_Q10 = _mm256_add_epi32(
+            _mm256_broadcastsi128_si256(psDelDec->RD_Q10),
+            _mm256_blendv_epi8(
+                _mm256_permute2x128_si256(rd_Q10, rd_Q10, 0x1),
+                rd_Q10,
+                mask));
+        SS_Q_Q10 = _mm256_blendv_epi8(
+            _mm256_permute2x128_si256(q_Q10, q_Q10, 0x1),
+            q_Q10,
+            mask);
+
+        /* Update states for best and second best quantization */
+
+        /* Quantized excitation */
+        exc_Q14 = silk_mm256_sign_epi32(_mm256_slli_epi32(SS_Q_Q10, 4), _mm256_broadcastsi128_si256(psDelDec->Seed));
+
+        /* Add predictions */
+        exc_Q14 = _mm256_add_epi32(exc_Q14, _mm256_set1_epi32(LTP_pred_Q14));
+        SS_LPC_exc_Q14 = _mm256_slli_epi32(exc_Q14, 1);
+        SS_xq_Q14 = _mm256_add_epi32(exc_Q14, _mm256_broadcastsi128_si256(LPC_pred_Q14));
+
+        /* Update states */
+        SS_Diff_Q14 = _mm256_sub_epi32(SS_xq_Q14, _mm256_set1_epi32(x_Q10[i] << 4));
+        SS_LF_AR_Q14 = _mm256_sub_epi32(SS_Diff_Q14, _mm256_broadcastsi128_si256(n_AR_Q14));
+        SS_sLTP_shp_Q14 = silk_mm256_sub_sat_epi32(SS_LF_AR_Q14, _mm256_broadcastsi128_si256(n_LF_Q14));
+
+        /* END Updating Delayed Decision States */
+
+        *smpl_buf_idx = (*smpl_buf_idx + DECISION_DELAY - 1) % DECISION_DELAY;
+        last_smple_idx = (*smpl_buf_idx + decisionDelay) % DECISION_DELAY;
+        psLastSample = &psDelDec->Samples[last_smple_idx];
+
+        /* Find winner */
+        RDmin_Q10 = silk_mm_mask_hmin_epi32(_mm256_castsi256_si128(SS_RD_Q10), MaskDelDec);
+        Winner_selector = silk_index_to_selector(silk_index_of_first_equal_epi32(RDmin_Q10, _mm256_castsi256_si128(SS_RD_Q10)));
+
+        /* Increase RD values of expired states */
+        Winner_rand_state = _mm_shuffle_epi8(psLastSample->RandState, Winner_selector);
+
+        SS_RD_Q10 = _mm256_blendv_epi8(
+            _mm256_add_epi32(SS_RD_Q10, _mm256_set1_epi32(silk_int32_MAX >> 4)),
+            SS_RD_Q10,
+            _mm256_broadcastsi128_si256(_mm_cmpeq_epi32(psLastSample->RandState, Winner_rand_state)));
+
+        /* find worst in first set */
+        RDmax_Q10 = silk_mm_mask_hmax_epi32(_mm256_extracti128_si256(SS_RD_Q10, 0), MaskDelDec);
+        /* find best in second set */
+        RDmin_Q10 = silk_mm_mask_hmin_epi32(_mm256_extracti128_si256(SS_RD_Q10, 1), MaskDelDec);
+
+        /* Replace a state if best from second set outperforms worst in first set */
+        tmp0 = _mm_cmplt_epi32(RDmin_Q10, RDmax_Q10);
+        if (!_mm_test_all_zeros(tmp0, tmp0))
+        {
+            RDmax_ind = silk_index_of_first_equal_epi32(RDmax_Q10, _mm256_extracti128_si256(SS_RD_Q10, 0));
+            RDmin_ind = silk_index_of_first_equal_epi32(RDmin_Q10, _mm256_extracti128_si256(SS_RD_Q10, 1));
+            tmp1 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(0xFF << (RDmax_ind << 3)));
+            tmp0 = _mm_blendv_epi8(
+                _mm_set_epi8(0xF, 0xE, 0xD, 0xC, 0xB, 0xA, 0x9, 0x8, 0x7, 0x6, 0x5, 0x4, 0x3, 0x2, 0x1, 0x0),
+                silk_index_to_selector(RDmin_ind),
+                tmp1);
+            for (int t = i; t < MAX_SUB_FRAME_LENGTH + NSQ_LPC_BUF_LENGTH; t++)
+            {
+                psDelDec->sLPC_Q14[t] = _mm_shuffle_epi8(psDelDec->sLPC_Q14[t], tmp0);
+            }
+            psDelDec->Seed = _mm_shuffle_epi8(psDelDec->Seed, tmp0);
+            psDelDec->SeedInit = _mm_shuffle_epi8(psDelDec->SeedInit, tmp0);
+            for (int t = 0; t < MAX_SHAPE_LPC_ORDER; t++)
+            {
+                psDelDec->sAR2_Q14[t] = _mm_shuffle_epi8(psDelDec->sAR2_Q14[t], tmp0);
+            }
+            for (int t = 0; t < DECISION_DELAY; t++)
+            {
+                psDelDec->Samples[t].RandState = _mm_shuffle_epi8(psDelDec->Samples[t].RandState, tmp0);
+                psDelDec->Samples[t].Q_Q10 = _mm_shuffle_epi8(psDelDec->Samples[t].Q_Q10, tmp0);
+                psDelDec->Samples[t].Xq_Q14 = _mm_shuffle_epi8(psDelDec->Samples[t].Xq_Q14, tmp0);
+                psDelDec->Samples[t].Pred_Q15 = _mm_shuffle_epi8(psDelDec->Samples[t].Pred_Q15, tmp0);
+                psDelDec->Samples[t].Shape_Q14 = _mm_shuffle_epi8(psDelDec->Samples[t].Shape_Q14, tmp0);
+            }
+            mask = _mm256_castsi128_si256(_mm_blendv_epi8(_mm_set_epi32(0x3, 0x2, 0x1, 0x0), _mm_set1_epi32(RDmin_ind + 4), tmp1));
+            SS_Q_Q10 = _mm256_permutevar8x32_epi32(SS_Q_Q10, mask);
+            SS_RD_Q10 = _mm256_permutevar8x32_epi32(SS_RD_Q10, mask);
+            SS_xq_Q14 = _mm256_permutevar8x32_epi32(SS_xq_Q14, mask);
+            SS_LF_AR_Q14 = _mm256_permutevar8x32_epi32(SS_LF_AR_Q14, mask);
+            SS_Diff_Q14 = _mm256_permutevar8x32_epi32(SS_Diff_Q14, mask);
+            SS_sLTP_shp_Q14 = _mm256_permutevar8x32_epi32(SS_sLTP_shp_Q14, mask);
+            SS_LPC_exc_Q14 = _mm256_permutevar8x32_epi32(SS_LPC_exc_Q14, mask);
+        }
+
+        /* Write samples from winner to output and long-term filter states */
+        if (subfr > 0 || i >= decisionDelay)
+        {
+            pulses[i - decisionDelay] =
+                (opus_int8)silk_sar_round_32(silk_select_winner(psLastSample->Q_Q10, Winner_selector), 10);
+            xq[i - decisionDelay] =
+                silk_sat16((opus_int32)silk_sar_round_smulww(silk_select_winner(psLastSample->Xq_Q14, Winner_selector), delayedGain_Q10[last_smple_idx], 8));
+            NSQ->sLTP_shp_Q14[NSQ->sLTP_shp_buf_idx - decisionDelay] =
+                silk_select_winner(psLastSample->Shape_Q14, Winner_selector);
+            sLTP_Q15[NSQ->sLTP_buf_idx - decisionDelay] =
+                silk_select_winner(psLastSample->Pred_Q15, Winner_selector);
+        }
+        NSQ->sLTP_shp_buf_idx++;
+        NSQ->sLTP_buf_idx++;
+
+        /* Update states */
+        psSample = &psDelDec->Samples[*smpl_buf_idx];
+        psDelDec->Seed = _mm_add_epi32(psDelDec->Seed, silk_mm_srai_round_epi32(_mm256_castsi256_si128(SS_Q_Q10), 10));
+        psDelDec->LF_AR_Q14 = _mm256_castsi256_si128(SS_LF_AR_Q14);
+        psDelDec->Diff_Q14 = _mm256_castsi256_si128(SS_Diff_Q14);
+        psDelDec->sLPC_Q14[i + NSQ_LPC_BUF_LENGTH] = _mm256_castsi256_si128(SS_xq_Q14);
+        psDelDec->RD_Q10 = _mm256_castsi256_si128(SS_RD_Q10);
+        psSample->Xq_Q14 = _mm256_castsi256_si128(SS_xq_Q14);
+        psSample->Q_Q10 = _mm256_castsi256_si128(SS_Q_Q10);
+        psSample->Pred_Q15 = _mm256_castsi256_si128(SS_LPC_exc_Q14);
+        psSample->Shape_Q14 = _mm256_castsi256_si128(SS_sLTP_shp_Q14);
+        psSample->RandState = psDelDec->Seed;
+        delayedGain_Q10[*smpl_buf_idx] = Gain_Q10;
+    }
+    /* Update LPC states */
+    for (int i = 0; i < NSQ_LPC_BUF_LENGTH; i++)
+    {
+        psDelDec->sLPC_Q14[i] = (&psDelDec->sLPC_Q14[length])[i];
+    }
+}
+
+static OPUS_INLINE void silk_nsq_del_dec_scale_states_avx2(
+    const silk_encoder_state *psEncC,          /* I    Encoder State                   */
+    silk_nsq_state *NSQ,                       /* I/O  NSQ state                       */
+    NSQ_del_dec_struct *psDelDec,              /* I/O  Delayed decision states         */
+    const opus_int16 x16[],                    /* I    Input                           */
+    opus_int32 x_sc_Q10[MAX_SUB_FRAME_LENGTH], /* 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 lag;
+    opus_int32 gain_adj_Q16, inv_gain_Q31, inv_gain_Q26;
+    NSQ_del_dec_sample_struct *psSample;
+
+    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_sar_round_32(inv_gain_Q31, 5);
+    for (int i = 0; i < psEncC->subfr_length; i+=4)
+    {
+        __m256i x = _mm256_cvtepi16_epi64(_mm_loadu_si64(&x16[i]));
+        x = _mm256_slli_epi64(_mm256_mul_epi32(x, _mm256_set1_epi32(inv_gain_Q26)), 16);
+        _mm_storeu_si128((__m128i_u*)&x_sc_Q10[i], silk_cvtepi64_epi32_high(x));
+    }
+
+    /* 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_SMULWB(inv_gain_Q31, LTP_scale_Q14) << 2;
+        }
+        for (int i = NSQ->sLTP_buf_idx - lag - LTP_ORDER / 2; i < NSQ->sLTP_buf_idx; i++)
+        {
+            silk_assert(i < MAX_FRAME_LENGTH);
+            sLTP_Q15[i] = silk_SMULWB(inv_gain_Q31, sLTP[i]);
+        }
+    }
+
+    /* Adjust for changing gain */
+    if (Gains_Q16[subfr] != NSQ->prev_gain_Q16)
+    {
+        gain_adj_Q16 = silk_DIV32_varQ(NSQ->prev_gain_Q16, Gains_Q16[subfr], 16);
+
+        /* Scale long-term shaping state */
+        for (int i = NSQ->sLTP_shp_buf_idx - psEncC->ltp_mem_length; i < NSQ->sLTP_shp_buf_idx; i+=4)
+        {
+            __m128i_u* p = (__m128i_u*)&NSQ->sLTP_shp_Q14[i];
+            *p = silk_mm_smulww_epi32(*p, gain_adj_Q16);
+        }
+
+        /* Scale long-term prediction state */
+        if (signal_type == TYPE_VOICED && NSQ->rewhite_flag == 0)
+        {
+            for (int i = NSQ->sLTP_buf_idx - lag - LTP_ORDER / 2; i < NSQ->sLTP_buf_idx - decisionDelay; i++)
+            {
+                sLTP_Q15[i] = ((opus_int64)sLTP_Q15[i]) * ((opus_int64)gain_adj_Q16) >> 16;
+            }
+        }
+
+        /* Scale scalar states */
+        psDelDec->LF_AR_Q14 = silk_mm_smulww_epi32(psDelDec->LF_AR_Q14, gain_adj_Q16);
+        psDelDec->Diff_Q14 = silk_mm_smulww_epi32(psDelDec->Diff_Q14, gain_adj_Q16);
+
+        /* Scale short-term prediction and shaping states */
+        for (int i = 0; i < NSQ_LPC_BUF_LENGTH; i++)
+        {
+            psDelDec->sLPC_Q14[i] = silk_mm_smulww_epi32(psDelDec->sLPC_Q14[i], gain_adj_Q16);
+        }
+        for (int i = 0; i < DECISION_DELAY; i++)
+        {
+            psSample = &psDelDec->Samples[i];
+            psSample->Pred_Q15 = silk_mm_smulww_epi32(psSample->Pred_Q15, gain_adj_Q16);
+            psSample->Shape_Q14 = silk_mm_smulww_epi32(psSample->Shape_Q14, gain_adj_Q16);
+        }
+        for (int i = 0; i < MAX_SHAPE_LPC_ORDER; i++)
+        {
+            psDelDec->sAR2_Q14[i] = silk_mm_smulww_epi32(psDelDec->sAR2_Q14[i], gain_adj_Q16);
+        }
+
+        /* Save inverse gain */
+        NSQ->prev_gain_Q16 = Gains_Q16[subfr];
+    }
+}
+
+static OPUS_INLINE void silk_LPC_analysis_filter_avx2(
+    opus_int16                  *out,               /* O    Output signal                           */
+    const opus_int16            *in,                /* I    Input signal                            */
+    const opus_int16            *B,                 /* I    MA prediction coefficients, Q12 [order] */
+    const opus_int32            len,                /* I    Signal length                           */
+    const opus_int32            order               /* I    Filter order                            */
+)
+{
+    opus_int32       out32_Q12, out32;
+    silk_assert(order == 10 || order == 16);
+
+    for(int i = order; i < len; i++ )
+    {
+        const opus_int16 *in_ptr = &in[ i ];
+        /* Allowing wrap around so that two wraps can cancel each other. The rare
+           cases where the result wraps around can only be triggered by invalid streams*/
+
+        __m256i in_v = _mm256_cvtepi16_epi32(_mm_loadu_si128((__m128i_u*)&in_ptr[-8]));
+        __m256i B_v  = _mm256_cvtepi16_epi32(_mm_loadu_si128((__m128i_u*)&      B[0]));
+        __m256i sum = _mm256_mullo_epi32(in_v, silk_mm256_reverse_epi32(B_v));
+        if (order > 10)
+        {
+            in_v = _mm256_cvtepi16_epi32(_mm_loadu_si128((__m128i_u*)&in_ptr[-16]));
+            B_v  = _mm256_cvtepi16_epi32(_mm_loadu_si128((__m128i_u*)&B       [8]));
+            B_v  = silk_mm256_reverse_epi32(B_v);
+        }
+        else
+        {
+            in_v = _mm256_cvtepi16_epi32(_mm_loadu_si32(&in_ptr[-10]));
+            B_v  = _mm256_cvtepi16_epi32(_mm_loadu_si32(&B       [8]));
+            B_v  = _mm256_shuffle_epi32(B_v, 0x01);
+        }
+        sum = _mm256_add_epi32(sum, _mm256_mullo_epi32(in_v, B_v));
+
+        out32_Q12 = silk_mm256_hsum_epi32(sum);
+
+        /* Subtract prediction */
+        out32_Q12 = (((opus_int32)*in_ptr) << 12 ) - out32_Q12;
+
+        /* Scale to Q0 */
+        out32 = silk_sar_round_32(out32_Q12, 12);
+
+        /* Saturate output */
+        out[ i ] = silk_sat16(out32);
+    }
+
+    /* Set first d output samples to zero */
+    silk_memset( out, 0, order * sizeof( opus_int16 ) );
+}
diff --git a/silk/x86/main_sse.h b/silk/x86/main_sse.h
index 1b6fa3f4..7a8add27 100644
--- a/silk/x86/main_sse.h
+++ b/silk/x86/main_sse.h
@@ -154,7 +154,33 @@ void silk_NSQ_del_dec_sse4_1(
     const opus_int              LTP_scale_Q14                                 /* I    LTP state scaling               */
 );
 
-#  if defined OPUS_X86_PRESUME_SSE4_1
+void silk_NSQ_del_dec_avx2(
+    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_int32 pitchL[MAX_NB_SUBFR],                       /* I    Pitch lags                  */
+    const opus_int Lambda_Q10,                                   /* I    Rate/distortion tradeoff    */
+    const opus_int LTP_scale_Q14                                 /* I    LTP state scaling           */
+);
+
+#  if defined (OPUS_X86_PRESUME_AVX2)
+
+#   define OVERRIDE_silk_NSQ_del_dec
+#   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_avx2(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_X86_PRESUME_SSE4_1)
 
 #   define OVERRIDE_silk_NSQ_del_dec
 #   define silk_NSQ_del_dec(psEncC, NSQ, psIndices, x16, pulses, PredCoef_Q12, LTPCoef_Q14, AR_Q13, \
diff --git a/silk/x86/x86_silk_map.c b/silk/x86/x86_silk_map.c
index 81ac2a01..dfec2614 100644
--- a/silk/x86/x86_silk_map.c
+++ b/silk/x86/x86_silk_map.c
@@ -132,7 +132,7 @@ void (*const SILK_NSQ_DEL_DEC_IMPL[ OPUS_ARCHMASK + 1 ] )(
   silk_NSQ_del_dec_c,
   silk_NSQ_del_dec_c,
   MAY_HAVE_SSE4_1( silk_NSQ_del_dec ), /* sse4.1 */
-  MAY_HAVE_SSE4_1( silk_NSQ_del_dec )  /* avx */
+  MAY_HAVE_AVX2( silk_NSQ_del_dec )  /* avx */
 };
 
 #if defined(FIXED_POINT)
diff --git a/silk_sources.mk b/silk_sources.mk
index 3df24816..b1144694 100644
--- a/silk_sources.mk
+++ b/silk_sources.mk
@@ -86,6 +86,9 @@ silk/x86/NSQ_del_dec_sse4_1.c \
 silk/x86/VAD_sse4_1.c \
 silk/x86/VQ_WMat_EC_sse4_1.c
 
+SILK_SOURCES_AVX2 =  \
+silk/x86/NSQ_del_dec_avx2.c
+
 SILK_SOURCES_ARM_RTCD = \
 silk/arm/arm_silk_map.c