// clang-format off

// Auto-generated file. Do not edit!

//   Template: src/qs8-gemm/MRx8c8-avxvnni.c.in

//   Generator: tools/xngen

//

// Copyright 2024 Google LLC

//

// This source code is licensed under the BSD-style license found in the

// LICENSE file in the root directory of this source tree.

#include <assert.h>

#include <stddef.h>

#include <stdint.h>

#include <immintrin.h>

#include "src/xnnpack/common.h"

#include "src/xnnpack/gemm.h"

#include "src/xnnpack/intrinsics-polyfill.h"

#include "src/xnnpack/math.h"

#include "src/xnnpack/microparams.h"

#include "src/xnnpack/prefetch.h"

#include "src/xnnpack/unaligned.h"

void xnn_qs8_qc4w_gemm_minmax_fp32_ukernel_7x8c8__avx2_madd_prfm(

    size_t mr,

    size_t nc,

    size_t kc,

    const int8_t* restrict a,

    size_t a_stride,

    const void* restrict w,

    int8_t* restrict c,

    size_t cm_stride,

    size_t cn_stride,

    const union xnn_qs8_qc8w_conv_minmax_params* restrict params) XNN_OOB_READS

{

  assert(mr != 0);

  assert(mr <= 7);

  assert(nc != 0);

  assert(kc != 0);

  assert(kc % sizeof(int8_t) == 0);

  assert(a != NULL);

  assert(w != NULL);

  assert(c != NULL);

  kc = round_up_po2(kc, 8 * sizeof(int8_t));

  const int8_t* a0 = a;

  int8_t* c0 = c;

  const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);

  int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);

  if XNN_UNPREDICTABLE(mr < 2) {

    a1 = a0;

    c1 = c0;

  }

  const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);

  int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);

  if XNN_UNPREDICTABLE(mr <= 2) {

    a2 = a1;

    c2 = c1;

  }

  const int8_t* a3 = (const int8_t*) ((uintptr_t) a2 + a_stride);

  int8_t* c3 = (int8_t*) ((uintptr_t) c2 + cm_stride);

  if XNN_UNPREDICTABLE(mr < 4) {

    a3 = a2;

    c3 = c2;

  }

  const int8_t* a4 = (const int8_t*) ((uintptr_t) a3 + a_stride);

  int8_t* c4 = (int8_t*) ((uintptr_t) c3 + cm_stride);

  if XNN_UNPREDICTABLE(mr <= 4) {

    a4 = a3;

    c4 = c3;

  }

  const int8_t* a5 = (const int8_t*) ((uintptr_t) a4 + a_stride);

  int8_t* c5 = (int8_t*) ((uintptr_t) c4 + cm_stride);

  if XNN_UNPREDICTABLE(mr < 6) {

    a5 = a4;

    c5 = c4;

  }

  const int8_t* a6 = (const int8_t*) ((uintptr_t) a5 + a_stride);

  int8_t* c6 = (int8_t*) ((uintptr_t) c5 + cm_stride);

  if XNN_UNPREDICTABLE(mr <= 6) {

    a6 = a5;

    c6 = c5;

  }

  const __m256i vsign_mask = _mm256_set1_epi8(0x80);

  XNN_FORCE_REALIZATION(vsign_mask);

  const __m256 voutput_max_less_zero_point = _mm256_set1_ps((int32_t) params->fp32_scalar.output_max - (int32_t) params->fp32_scalar.output_zero_point);

  const __m256i voutput_zero_point = _mm256_set1_epi32(params->fp32_scalar.output_zero_point);

  const __m128i voutput_min = _mm_set1_epi8(params->fp32_scalar.output_min);

  // XNN_FORCE_REALIZATION(voutput_max_less_zero_point);

  // XNN_FORCE_REALIZATION(voutput_zero_point);

  // XNN_FORCE_REALIZATION(voutput_min);

  const __m256i vmask = _mm256_set1_epi8(0x0F);

  XNN_FORCE_REALIZATION(vmask);

  do {

    __m256i vacc0x0123 = _mm256_cvtepu32_epi64(_mm_load_si128((const __m128i*) w));

    __m256i vacc0x4567 = _mm256_cvtepu32_epi64(_mm_load_si128((const __m128i*) ((const int32_t*) w + 4)));

    __m256i vacc1x0123 = vacc0x0123;

    __m256i vacc1x4567 = vacc0x4567;

    __m256i vacc2x0123 = vacc0x0123;

    __m256i vacc2x4567 = vacc0x4567;

    __m256i vacc3x0123 = vacc0x0123;

    __m256i vacc3x4567 = vacc0x4567;

    __m256i vacc4x0123 = vacc0x0123;

    __m256i vacc4x4567 = vacc0x4567;

    __m256i vacc5x0123 = vacc0x0123;

    __m256i vacc5x4567 = vacc0x4567;

    __m256i vacc6x0123 = vacc0x0123;

    __m256i vacc6x4567 = vacc0x4567;

    w = (const int32_t*) w + 8;

    size_t k = kc;

    while (k >= 16 * sizeof(int8_t)) {

      const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);

      const __m256i va0x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0 + 8)), vsign_mask);

      a0 += 16;

      const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);

      const __m256i va1x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1 + 8)), vsign_mask);

      a1 += 16;

      const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);

      const __m256i va2x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2 + 8)), vsign_mask);

      a2 += 16;

      const __m256i va3x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3)), vsign_mask);

      const __m256i va3x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3 + 8)), vsign_mask);

      a3 += 16;

      const __m256i va4x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4)), vsign_mask);

      const __m256i va4x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4 + 8)), vsign_mask);

      a4 += 16;

      const __m256i va5x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5)), vsign_mask);

      const __m256i va5x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5 + 8)), vsign_mask);

      a5 += 16;

      const __m256i va6x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a6)), vsign_mask);

      const __m256i va6x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a6 + 8)), vsign_mask);

      a6 += 16;

      const __m256i vbb01234567x01234567 = _mm256_load_si256(w);

      const __m256i vbb89ABCDEFx01234567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));

      const __m256i vbs01234567x4567 = _mm256_srli_epi32(vbb01234567x01234567, 4);

      const __m256i vbs89ABCDEFx4567 = _mm256_srli_epi32(vbb89ABCDEFx01234567, 4);

      const __m256i vb01234567x0123 = _mm256_and_si256(vbb01234567x01234567, vmask);

      const __m256i vb89ABCDEFx0123 = _mm256_and_si256(vbb89ABCDEFx01234567, vmask);

      const __m256i vb01234567x4567 = _mm256_and_si256(vbs01234567x4567, vmask);

      const __m256i vb89ABCDEFx4567 = _mm256_and_si256(vbs89ABCDEFx4567, vmask);

      vacc0x0123 = _mm256_dpbusd_epi32_madd(vacc0x0123, va0x01234567, vb01234567x0123);

      vacc0x4567 = _mm256_dpbusd_epi32_madd(vacc0x4567, va0x01234567, vb89ABCDEFx0123);

      vacc1x0123 = _mm256_dpbusd_epi32_madd(vacc1x0123, va1x01234567, vb01234567x0123);

      vacc1x4567 = _mm256_dpbusd_epi32_madd(vacc1x4567, va1x01234567, vb89ABCDEFx0123);

      vacc2x0123 = _mm256_dpbusd_epi32_madd(vacc2x0123, va2x01234567, vb01234567x0123);

      vacc2x4567 = _mm256_dpbusd_epi32_madd(vacc2x4567, va2x01234567, vb89ABCDEFx0123);

      vacc3x0123 = _mm256_dpbusd_epi32_madd(vacc3x0123, va3x01234567, vb01234567x0123);

      vacc3x4567 = _mm256_dpbusd_epi32_madd(vacc3x4567, va3x01234567, vb89ABCDEFx0123);

      vacc4x0123 = _mm256_dpbusd_epi32_madd(vacc4x0123, va4x01234567, vb01234567x0123);

      vacc4x4567 = _mm256_dpbusd_epi32_madd(vacc4x4567, va4x01234567, vb89ABCDEFx0123);

      vacc5x0123 = _mm256_dpbusd_epi32_madd(vacc5x0123, va5x01234567, vb01234567x0123);

      vacc5x4567 = _mm256_dpbusd_epi32_madd(vacc5x4567, va5x01234567, vb89ABCDEFx0123);

      vacc6x0123 = _mm256_dpbusd_epi32_madd(vacc6x0123, va6x01234567, vb01234567x0123);

      vacc6x4567 = _mm256_dpbusd_epi32_madd(vacc6x4567, va6x01234567, vb89ABCDEFx0123);

      xnn_prefetch_to_l1((const int8_t*) w + 960);

      vacc0x0123 = _mm256_dpbusd_epi32_madd(vacc0x0123, va0x89ABCDEF, vb01234567x4567);

      vacc0x4567 = _mm256_dpbusd_epi32_madd(vacc0x4567, va0x89ABCDEF, vb89ABCDEFx4567);

      vacc1x0123 = _mm256_dpbusd_epi32_madd(vacc1x0123, va1x89ABCDEF, vb01234567x4567);

      vacc1x4567 = _mm256_dpbusd_epi32_madd(vacc1x4567, va1x89ABCDEF, vb89ABCDEFx4567);

      vacc2x0123 = _mm256_dpbusd_epi32_madd(vacc2x0123, va2x89ABCDEF, vb01234567x4567);

      vacc2x4567 = _mm256_dpbusd_epi32_madd(vacc2x4567, va2x89ABCDEF, vb89ABCDEFx4567);

      vacc3x0123 = _mm256_dpbusd_epi32_madd(vacc3x0123, va3x89ABCDEF, vb01234567x4567);

      vacc3x4567 = _mm256_dpbusd_epi32_madd(vacc3x4567, va3x89ABCDEF, vb89ABCDEFx4567);

      vacc4x0123 = _mm256_dpbusd_epi32_madd(vacc4x0123, va4x89ABCDEF, vb01234567x4567);

      vacc4x4567 = _mm256_dpbusd_epi32_madd(vacc4x4567, va4x89ABCDEF, vb89ABCDEFx4567);

      vacc5x0123 = _mm256_dpbusd_epi32_madd(vacc5x0123, va5x89ABCDEF, vb01234567x4567);

      vacc5x4567 = _mm256_dpbusd_epi32_madd(vacc5x4567, va5x89ABCDEF, vb89ABCDEFx4567);

      vacc6x0123 = _mm256_dpbusd_epi32_madd(vacc6x0123, va6x89ABCDEF, vb01234567x4567);

      vacc6x4567 = _mm256_dpbusd_epi32_madd(vacc6x4567, va6x89ABCDEF, vb89ABCDEFx4567);

      w = (const int8_t*) w + 64;

      k -= 16 * sizeof(int8_t);

    }

    if (k != 0) {

      const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);

      a0 += 8;

      const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);

      a1 += 8;

      const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);

      a2 += 8;

      const __m256i va3x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3)), vsign_mask);

      a3 += 8;

      const __m256i va4x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4)), vsign_mask);

      a4 += 8;

      const __m256i va5x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5)), vsign_mask);

      a5 += 8;

      const __m256i va6x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a6)), vsign_mask);

      a6 += 8;

      const __m256i vbb01234567x01234567 = _mm256_load_si256(w);

      const __m256i vbb89ABCDEFx01234567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));

      const __m256i vb01234567x0123 = _mm256_and_si256(vbb01234567x01234567, vmask);

      const __m256i vb89ABCDEFx0123 = _mm256_and_si256(vbb89ABCDEFx01234567, vmask);

      vacc0x0123 = _mm256_dpbusd_epi32_madd(vacc0x0123, va0x01234567, vb01234567x0123);

      vacc0x4567 = _mm256_dpbusd_epi32_madd(vacc0x4567, va0x01234567, vb89ABCDEFx0123);

      vacc1x0123 = _mm256_dpbusd_epi32_madd(vacc1x0123, va1x01234567, vb01234567x0123);

      vacc1x4567 = _mm256_dpbusd_epi32_madd(vacc1x4567, va1x01234567, vb89ABCDEFx0123);

      vacc2x0123 = _mm256_dpbusd_epi32_madd(vacc2x0123, va2x01234567, vb01234567x0123);

      vacc2x4567 = _mm256_dpbusd_epi32_madd(vacc2x4567, va2x01234567, vb89ABCDEFx0123);

      vacc3x0123 = _mm256_dpbusd_epi32_madd(vacc3x0123, va3x01234567, vb01234567x0123);

      vacc3x4567 = _mm256_dpbusd_epi32_madd(vacc3x4567, va3x01234567, vb89ABCDEFx0123);

      vacc4x0123 = _mm256_dpbusd_epi32_madd(vacc4x0123, va4x01234567, vb01234567x0123);

      vacc4x4567 = _mm256_dpbusd_epi32_madd(vacc4x4567, va4x01234567, vb89ABCDEFx0123);

      vacc5x0123 = _mm256_dpbusd_epi32_madd(vacc5x0123, va5x01234567, vb01234567x0123);

      vacc5x4567 = _mm256_dpbusd_epi32_madd(vacc5x4567, va5x01234567, vb89ABCDEFx0123);

      vacc6x0123 = _mm256_dpbusd_epi32_madd(vacc6x0123, va6x01234567, vb01234567x0123);

      vacc6x4567 = _mm256_dpbusd_epi32_madd(vacc6x4567, va6x01234567, vb89ABCDEFx0123);

      xnn_prefetch_to_l1((const int8_t*) w + 960);

      w = (const int8_t*) w + 64;

      k -= 8 * sizeof(int8_t);

    }

    // Add adjacent pairs

    const __m256i vsum0x02134657 = _mm256_hadd_epi32(vacc0x0123, vacc0x4567);

    __m256i vacc0x01234567 = _mm256_permute4x64_epi64(vsum0x02134657, _MM_SHUFFLE(3, 1, 2, 0));

    const __m256i vsum1x02134657 = _mm256_hadd_epi32(vacc1x0123, vacc1x4567);

    __m256i vacc1x01234567 = _mm256_permute4x64_epi64(vsum1x02134657, _MM_SHUFFLE(3, 1, 2, 0));

    const __m256i vsum2x02134657 = _mm256_hadd_epi32(vacc2x0123, vacc2x4567);

    __m256i vacc2x01234567 = _mm256_permute4x64_epi64(vsum2x02134657, _MM_SHUFFLE(3, 1, 2, 0));

    const __m256i vsum3x02134657 = _mm256_hadd_epi32(vacc3x0123, vacc3x4567);

    __m256i vacc3x01234567 = _mm256_permute4x64_epi64(vsum3x02134657, _MM_SHUFFLE(3, 1, 2, 0));

    const __m256i vsum4x02134657 = _mm256_hadd_epi32(vacc4x0123, vacc4x4567);

    __m256i vacc4x01234567 = _mm256_permute4x64_epi64(vsum4x02134657, _MM_SHUFFLE(3, 1, 2, 0));

    const __m256i vsum5x02134657 = _mm256_hadd_epi32(vacc5x0123, vacc5x4567);

    __m256i vacc5x01234567 = _mm256_permute4x64_epi64(vsum5x02134657, _MM_SHUFFLE(3, 1, 2, 0));

    const __m256i vsum6x02134657 = _mm256_hadd_epi32(vacc6x0123, vacc6x4567);

    __m256i vacc6x01234567 = _mm256_permute4x64_epi64(vsum6x02134657, _MM_SHUFFLE(3, 1, 2, 0));

    __m256 vout0x01234567 = _mm256_cvtepi32_ps(vacc0x01234567);

    __m256 vout1x01234567 = _mm256_cvtepi32_ps(vacc1x01234567);

    __m256 vout2x01234567 = _mm256_cvtepi32_ps(vacc2x01234567);

    __m256 vout3x01234567 = _mm256_cvtepi32_ps(vacc3x01234567);

    __m256 vout4x01234567 = _mm256_cvtepi32_ps(vacc4x01234567);

    __m256 vout5x01234567 = _mm256_cvtepi32_ps(vacc5x01234567);

    __m256 vout6x01234567 = _mm256_cvtepi32_ps(vacc6x01234567);

    const __m256 vscale01234567 = _mm256_load_ps(w);

    w = (const float*) w + 8;

    vout0x01234567 = _mm256_mul_ps(vout0x01234567, vscale01234567);

    vout1x01234567 = _mm256_mul_ps(vout1x01234567, vscale01234567);

    vout2x01234567 = _mm256_mul_ps(vout2x01234567, vscale01234567);

    vout3x01234567 = _mm256_mul_ps(vout3x01234567, vscale01234567);

    vout4x01234567 = _mm256_mul_ps(vout4x01234567, vscale01234567);

    vout5x01234567 = _mm256_mul_ps(vout5x01234567, vscale01234567);

    vout6x01234567 = _mm256_mul_ps(vout6x01234567, vscale01234567);

    vout0x01234567 = _mm256_min_ps(vout0x01234567, voutput_max_less_zero_point);

    vout1x01234567 = _mm256_min_ps(vout1x01234567, voutput_max_less_zero_point);

    vout2x01234567 = _mm256_min_ps(vout2x01234567, voutput_max_less_zero_point);

    vout3x01234567 = _mm256_min_ps(vout3x01234567, voutput_max_less_zero_point);

    vout4x01234567 = _mm256_min_ps(vout4x01234567, voutput_max_less_zero_point);

    vout5x01234567 = _mm256_min_ps(vout5x01234567, voutput_max_less_zero_point);

    vout6x01234567 = _mm256_min_ps(vout6x01234567, voutput_max_less_zero_point);

    vacc0x01234567 = _mm256_cvtps_epi32(vout0x01234567);

    vacc1x01234567 = _mm256_cvtps_epi32(vout1x01234567);

    vacc2x01234567 = _mm256_cvtps_epi32(vout2x01234567);

    vacc3x01234567 = _mm256_cvtps_epi32(vout3x01234567);

    vacc4x01234567 = _mm256_cvtps_epi32(vout4x01234567);

    vacc5x01234567 = _mm256_cvtps_epi32(vout5x01234567);

    vacc6x01234567 = _mm256_cvtps_epi32(vout6x01234567);

    vacc0x01234567 = _mm256_add_epi32(vacc0x01234567, voutput_zero_point);

    vacc1x01234567 = _mm256_add_epi32(vacc1x01234567, voutput_zero_point);

    vacc2x01234567 = _mm256_add_epi32(vacc2x01234567, voutput_zero_point);

    vacc3x01234567 = _mm256_add_epi32(vacc3x01234567, voutput_zero_point);

    vacc4x01234567 = _mm256_add_epi32(vacc4x01234567, voutput_zero_point);

    vacc5x01234567 = _mm256_add_epi32(vacc5x01234567, voutput_zero_point);

    vacc6x01234567 = _mm256_add_epi32(vacc6x01234567, voutput_zero_point);

    vacc0x01234567 = _mm256_packs_epi32(vacc0x01234567, _mm256_castsi128_si256(_mm256_extracti128_si256(vacc0x01234567, 1)));

    __m128i voutb0x01234567 = _mm256_castsi256_si128(_mm256_packs_epi16(vacc0x01234567, vacc0x01234567));

    vacc1x01234567 = _mm256_packs_epi32(vacc1x01234567, _mm256_castsi128_si256(_mm256_extracti128_si256(vacc1x01234567, 1)));

    __m128i voutb1x01234567 = _mm256_castsi256_si128(_mm256_packs_epi16(vacc1x01234567, vacc1x01234567));

    vacc2x01234567 = _mm256_packs_epi32(vacc2x01234567, _mm256_castsi128_si256(_mm256_extracti128_si256(vacc2x01234567, 1)));

    __m128i voutb2x01234567 = _mm256_castsi256_si128(_mm256_packs_epi16(vacc2x01234567, vacc2x01234567));

    vacc3x01234567 = _mm256_packs_epi32(vacc3x01234567, _mm256_castsi128_si256(_mm256_extracti128_si256(vacc3x01234567, 1)));

    __m128i voutb3x01234567 = _mm256_castsi256_si128(_mm256_packs_epi16(vacc3x01234567, vacc3x01234567));

    vacc4x01234567 = _mm256_packs_epi32(vacc4x01234567, _mm256_castsi128_si256(_mm256_extracti128_si256(vacc4x01234567, 1)));

    __m128i voutb4x01234567 = _mm256_castsi256_si128(_mm256_packs_epi16(vacc4x01234567, vacc4x01234567));

    vacc5x01234567 = _mm256_packs_epi32(vacc5x01234567, _mm256_castsi128_si256(_mm256_extracti128_si256(vacc5x01234567, 1)));

    __m128i voutb5x01234567 = _mm256_castsi256_si128(_mm256_packs_epi16(vacc5x01234567, vacc5x01234567));

    vacc6x01234567 = _mm256_packs_epi32(vacc6x01234567, _mm256_castsi128_si256(_mm256_extracti128_si256(vacc6x01234567, 1)));

    __m128i voutb6x01234567 = _mm256_castsi256_si128(_mm256_packs_epi16(vacc6x01234567, vacc6x01234567));

    voutb0x01234567 = _mm_max_epi8(voutb0x01234567, voutput_min);

    voutb1x01234567 = _mm_max_epi8(voutb1x01234567, voutput_min);

    voutb2x01234567 = _mm_max_epi8(voutb2x01234567, voutput_min);

    voutb3x01234567 = _mm_max_epi8(voutb3x01234567, voutput_min);

    voutb4x01234567 = _mm_max_epi8(voutb4x01234567, voutput_min);

    voutb5x01234567 = _mm_max_epi8(voutb5x01234567, voutput_min);

    voutb6x01234567 = _mm_max_epi8(voutb6x01234567, voutput_min);

    if (nc >= 8) {

      _mm_storel_epi64((__m128i*) c0, voutb0x01234567);

      c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);

      a0 = (const int8_t*) ((uintptr_t) a0 - kc);

      _mm_storel_epi64((__m128i*) c1, voutb1x01234567);

      c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);

      a1 = (const int8_t*) ((uintptr_t) a1 - kc);

      _mm_storel_epi64((__m128i*) c2, voutb2x01234567);

      c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);

      a2 = (const int8_t*) ((uintptr_t) a2 - kc);

      _mm_storel_epi64((__m128i*) c3, voutb3x01234567);

      c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);

      a3 = (const int8_t*) ((uintptr_t) a3 - kc);

      _mm_storel_epi64((__m128i*) c4, voutb4x01234567);

      c4 = (int8_t*) ((uintptr_t) c4 + cn_stride);

      a4 = (const int8_t*) ((uintptr_t) a4 - kc);

      _mm_storel_epi64((__m128i*) c5, voutb5x01234567);

      c5 = (int8_t*) ((uintptr_t) c5 + cn_stride);

      a5 = (const int8_t*) ((uintptr_t) a5 - kc);

      _mm_storel_epi64((__m128i*) c6, voutb6x01234567);

      c6 = (int8_t*) ((uintptr_t) c6 + cn_stride);

      a6 = (const int8_t*) ((uintptr_t) a6 - kc);

      nc -= 8;

    } else {

      if (nc & 4) {

        _mm_storeu_si32(c0, voutb0x01234567);

        c0 += 4;

        _mm_storeu_si32(c1, voutb1x01234567);

        c1 += 4;

        _mm_storeu_si32(c2, voutb2x01234567);

        c2 += 4;

        _mm_storeu_si32(c3, voutb3x01234567);

        c3 += 4;

        _mm_storeu_si32(c4, voutb4x01234567);

        c4 += 4;

        _mm_storeu_si32(c5, voutb5x01234567);

        c5 += 4;

        _mm_storeu_si32(c6, voutb6x01234567);

        c6 += 4;

        voutb0x01234567 = _mm_srli_epi64(voutb0x01234567, 32);

        voutb1x01234567 = _mm_srli_epi64(voutb1x01234567, 32);

        voutb2x01234567 = _mm_srli_epi64(voutb2x01234567, 32);

        voutb3x01234567 = _mm_srli_epi64(voutb3x01234567, 32);

        voutb4x01234567 = _mm_srli_epi64(voutb4x01234567, 32);

        voutb5x01234567 = _mm_srli_epi64(voutb5x01234567, 32);

        voutb6x01234567 = _mm_srli_epi64(voutb6x01234567, 32);

      }

      if (nc & 2) {

        unaligned_store_u16(c0, (uint16_t) _mm_extract_epi16(voutb0x01234567, 0));

        c0 += 2;

        unaligned_store_u16(c1, (uint16_t) _mm_extract_epi16(voutb1x01234567, 0));

        c1 += 2;

        unaligned_store_u16(c2, (uint16_t) _mm_extract_epi16(voutb2x01234567, 0));

        c2 += 2;

        unaligned_store_u16(c3, (uint16_t) _mm_extract_epi16(voutb3x01234567, 0));

        c3 += 2;

        unaligned_store_u16(c4, (uint16_t) _mm_extract_epi16(voutb4x01234567, 0));

        c4 += 2;

        unaligned_store_u16(c5, (uint16_t) _mm_extract_epi16(voutb5x01234567, 0));

        c5 += 2;

        unaligned_store_u16(c6, (uint16_t) _mm_extract_epi16(voutb6x01234567, 0));

        c6 += 2;

        voutb0x01234567 = _mm_srli_epi32(voutb0x01234567, 16);

        voutb1x01234567 = _mm_srli_epi32(voutb1x01234567, 16);

        voutb2x01234567 = _mm_srli_epi32(voutb2x01234567, 16);

        voutb3x01234567 = _mm_srli_epi32(voutb3x01234567, 16);

        voutb4x01234567 = _mm_srli_epi32(voutb4x01234567, 16);

        voutb5x01234567 = _mm_srli_epi32(voutb5x01234567, 16);

        voutb6x01234567 = _mm_srli_epi32(voutb6x01234567, 16);

      }

      if (nc & 1) {

        *c0 = (int8_t) _mm_extract_epi8(voutb0x01234567, 0);

        *c1 = (int8_t) _mm_extract_epi8(voutb1x01234567, 0);

        *c2 = (int8_t) _mm_extract_epi8(voutb2x01234567, 0);

        *c3 = (int8_t) _mm_extract_epi8(voutb3x01234567, 0);

        *c4 = (int8_t) _mm_extract_epi8(voutb4x01234567, 0);

        *c5 = (int8_t) _mm_extract_epi8(voutb5x01234567, 0);

        *c6 = (int8_t) _mm_extract_epi8(voutb6x01234567, 0);

      }

      nc = 0;

    }

  } while (nc != 0);

}