// 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 <immintrin.h>

#include "xnnpack/gemm.h"
#include "xnnpack/intrinsics-polyfill.h"
#include "xnnpack/math.h"
#include "xnnpack/unaligned.h"
#include "xnnpack/prefetch.h"


void xnn_qd8_f32_qc8w_gemm_minmax_ukernel_7x8c8__avx256vnni_prfm(
    size_t mr,
    size_t nc,
    size_t kc,
    const int8_t* restrict a,
    size_t a_stride,
    const void* restrict w,
    float* restrict c,
    size_t cm_stride,
    size_t cn_stride,
    const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
    const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) 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;
  float* c0 = c;
  const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
  float* c1 = (float*) ((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);
  float* c2 = (float*) ((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);
  float* c3 = (float*) ((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);
  float* c4 = (float*) ((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);
  float* c5 = (float*) ((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);
  float* c6 = (float*) ((uintptr_t) c5 + cm_stride);
  if XNN_UNPREDICTABLE(mr <= 6) {
    a6 = a5;
    c6 = c5;
  }

  const __m256i vinput_zero_point0 = _mm256_set1_epi32((int) quantization_params[0].zero_point);
  const __m256i vinput_zero_point1 = _mm256_set1_epi32((int) quantization_params[1].zero_point);
  const __m256i vinput_zero_point2 = _mm256_set1_epi32((int) quantization_params[2].zero_point);
  const __m256i vinput_zero_point3 = _mm256_set1_epi32((int) quantization_params[3].zero_point);
  const __m256i vinput_zero_point4 = _mm256_set1_epi32((int) quantization_params[4].zero_point);
  const __m256i vinput_zero_point5 = _mm256_set1_epi32((int) quantization_params[5].zero_point);
  const __m256i vinput_zero_point6 = _mm256_set1_epi32((int) quantization_params[6].zero_point);
  const __m256 voutput_min = _mm256_set1_ps(params->scalar.min);
  const __m256 voutput_max = _mm256_set1_ps(params->scalar.max);
  // XNN_FORCE_REALIZATION(voutput_min);
  // XNN_FORCE_REALIZATION(voutput_max);
  do {
    const __m256i vksum01234567 = _mm256_load_si256(w);
    __m256i vsum0x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point0);
    __m256i vacc0x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum0x01234567, 0));
    __m256i vacc0x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum0x01234567, 1));
    __m256i vsum1x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point1);
    __m256i vacc1x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum1x01234567, 0));
    __m256i vacc1x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum1x01234567, 1));
    __m256i vsum2x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point2);
    __m256i vacc2x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum2x01234567, 0));
    __m256i vacc2x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum2x01234567, 1));
    __m256i vsum3x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point3);
    __m256i vacc3x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum3x01234567, 0));
    __m256i vacc3x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum3x01234567, 1));
    __m256i vsum4x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point4);
    __m256i vacc4x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum4x01234567, 0));
    __m256i vacc4x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum4x01234567, 1));
    __m256i vsum5x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point5);
    __m256i vacc5x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum5x01234567, 0));
    __m256i vacc5x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum5x01234567, 1));
    __m256i vsum6x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point6);
    __m256i vacc6x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum6x01234567, 0));
    __m256i vacc6x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum6x01234567, 1));
    w = (const int32_t*) w + 8;

    size_t k = kc;
    while (k >= 16 * sizeof(int8_t)) {
      const __m256i va0x01234567 = _mm256_set1_epi64x((int64_t) unaligned_load_u64(a0));
      const __m256i va0x89ABCDEF = _mm256_set1_epi64x((int64_t) unaligned_load_u64(a0 + 8));
      a0 += 16;
      const __m256i va1x01234567 = _mm256_set1_epi64x((int64_t) unaligned_load_u64(a1));
      const __m256i va1x89ABCDEF = _mm256_set1_epi64x((int64_t) unaligned_load_u64(a1 + 8));
      a1 += 16;
      const __m256i va2x01234567 = _mm256_set1_epi64x((int64_t) unaligned_load_u64(a2));
      const __m256i va2x89ABCDEF = _mm256_set1_epi64x((int64_t) unaligned_load_u64(a2 + 8));
      a2 += 16;
      const __m256i va3x01234567 = _mm256_set1_epi64x((int64_t) unaligned_load_u64(a3));
      const __m256i va3x89ABCDEF = _mm256_set1_epi64x((int64_t) unaligned_load_u64(a3 + 8));
      a3 += 16;
      const __m256i va4x01234567 = _mm256_set1_epi64x((int64_t) unaligned_load_u64(a4));
      const __m256i va4x89ABCDEF = _mm256_set1_epi64x((int64_t) unaligned_load_u64(a4 + 8));
      a4 += 16;
      const __m256i va5x01234567 = _mm256_set1_epi64x((int64_t) unaligned_load_u64(a5));
      const __m256i va5x89ABCDEF = _mm256_set1_epi64x((int64_t) unaligned_load_u64(a5 + 8));
      a5 += 16;
      const __m256i va6x01234567 = _mm256_set1_epi64x((int64_t) unaligned_load_u64(a6));
      const __m256i va6x89ABCDEF = _mm256_set1_epi64x((int64_t) unaligned_load_u64(a6 + 8));
      a6 += 16;

      const __m256i vb01234567x0123 = _mm256_load_si256(w);
      const __m256i vb89ABCDEFx0123 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
      const __m256i vb01234567x4567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 64));
      const __m256i vb89ABCDEFx4567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 96));
      xnn_prefetch_to_l1((const int8_t*) w + 896);

      vacc0x0123 = _mm256_dpbusd_epi32(vacc0x0123, va0x01234567, vb01234567x0123);
      vacc0x4567 = _mm256_dpbusd_epi32(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
      vacc1x0123 = _mm256_dpbusd_epi32(vacc1x0123, va1x01234567, vb01234567x0123);
      vacc1x4567 = _mm256_dpbusd_epi32(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
      vacc2x0123 = _mm256_dpbusd_epi32(vacc2x0123, va2x01234567, vb01234567x0123);
      vacc2x4567 = _mm256_dpbusd_epi32(vacc2x4567, va2x01234567, vb89ABCDEFx0123);
      vacc3x0123 = _mm256_dpbusd_epi32(vacc3x0123, va3x01234567, vb01234567x0123);
      vacc3x4567 = _mm256_dpbusd_epi32(vacc3x4567, va3x01234567, vb89ABCDEFx0123);
      vacc4x0123 = _mm256_dpbusd_epi32(vacc4x0123, va4x01234567, vb01234567x0123);
      vacc4x4567 = _mm256_dpbusd_epi32(vacc4x4567, va4x01234567, vb89ABCDEFx0123);
      vacc5x0123 = _mm256_dpbusd_epi32(vacc5x0123, va5x01234567, vb01234567x0123);
      vacc5x4567 = _mm256_dpbusd_epi32(vacc5x4567, va5x01234567, vb89ABCDEFx0123);
      vacc6x0123 = _mm256_dpbusd_epi32(vacc6x0123, va6x01234567, vb01234567x0123);
      vacc6x4567 = _mm256_dpbusd_epi32(vacc6x4567, va6x01234567, vb89ABCDEFx0123);
      xnn_prefetch_to_l1((const int8_t*) w + 960);
      vacc0x0123 = _mm256_dpbusd_epi32(vacc0x0123, va0x89ABCDEF, vb01234567x4567);
      vacc0x4567 = _mm256_dpbusd_epi32(vacc0x4567, va0x89ABCDEF, vb89ABCDEFx4567);
      vacc1x0123 = _mm256_dpbusd_epi32(vacc1x0123, va1x89ABCDEF, vb01234567x4567);
      vacc1x4567 = _mm256_dpbusd_epi32(vacc1x4567, va1x89ABCDEF, vb89ABCDEFx4567);
      vacc2x0123 = _mm256_dpbusd_epi32(vacc2x0123, va2x89ABCDEF, vb01234567x4567);
      vacc2x4567 = _mm256_dpbusd_epi32(vacc2x4567, va2x89ABCDEF, vb89ABCDEFx4567);
      vacc3x0123 = _mm256_dpbusd_epi32(vacc3x0123, va3x89ABCDEF, vb01234567x4567);
      vacc3x4567 = _mm256_dpbusd_epi32(vacc3x4567, va3x89ABCDEF, vb89ABCDEFx4567);
      vacc4x0123 = _mm256_dpbusd_epi32(vacc4x0123, va4x89ABCDEF, vb01234567x4567);
      vacc4x4567 = _mm256_dpbusd_epi32(vacc4x4567, va4x89ABCDEF, vb89ABCDEFx4567);
      vacc5x0123 = _mm256_dpbusd_epi32(vacc5x0123, va5x89ABCDEF, vb01234567x4567);
      vacc5x4567 = _mm256_dpbusd_epi32(vacc5x4567, va5x89ABCDEF, vb89ABCDEFx4567);
      vacc6x0123 = _mm256_dpbusd_epi32(vacc6x0123, va6x89ABCDEF, vb01234567x4567);
      vacc6x4567 = _mm256_dpbusd_epi32(vacc6x4567, va6x89ABCDEF, vb89ABCDEFx4567);

      w = (const int8_t*) w + 128;
      k -= 16 * sizeof(int8_t);
    }

    if (k != 0) {
      const __m256i va0x01234567 = _mm256_set1_epi64x((int64_t) unaligned_load_u64(a0));
      a0 += 8;
      const __m256i va1x01234567 = _mm256_set1_epi64x((int64_t) unaligned_load_u64(a1));
      a1 += 8;
      const __m256i va2x01234567 = _mm256_set1_epi64x((int64_t) unaligned_load_u64(a2));
      a2 += 8;
      const __m256i va3x01234567 = _mm256_set1_epi64x((int64_t) unaligned_load_u64(a3));
      a3 += 8;
      const __m256i va4x01234567 = _mm256_set1_epi64x((int64_t) unaligned_load_u64(a4));
      a4 += 8;
      const __m256i va5x01234567 = _mm256_set1_epi64x((int64_t) unaligned_load_u64(a5));
      a5 += 8;
      const __m256i va6x01234567 = _mm256_set1_epi64x((int64_t) unaligned_load_u64(a6));
      a6 += 8;

      const __m256i vb01234567x0123 = _mm256_load_si256(w);
      const __m256i vb89ABCDEFx0123 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));

      vacc0x0123 = _mm256_dpbusd_epi32(vacc0x0123, va0x01234567, vb01234567x0123);
      vacc0x4567 = _mm256_dpbusd_epi32(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
      vacc1x0123 = _mm256_dpbusd_epi32(vacc1x0123, va1x01234567, vb01234567x0123);
      vacc1x4567 = _mm256_dpbusd_epi32(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
      vacc2x0123 = _mm256_dpbusd_epi32(vacc2x0123, va2x01234567, vb01234567x0123);
      vacc2x4567 = _mm256_dpbusd_epi32(vacc2x4567, va2x01234567, vb89ABCDEFx0123);
      vacc3x0123 = _mm256_dpbusd_epi32(vacc3x0123, va3x01234567, vb01234567x0123);
      vacc3x4567 = _mm256_dpbusd_epi32(vacc3x4567, va3x01234567, vb89ABCDEFx0123);
      vacc4x0123 = _mm256_dpbusd_epi32(vacc4x0123, va4x01234567, vb01234567x0123);
      vacc4x4567 = _mm256_dpbusd_epi32(vacc4x4567, va4x01234567, vb89ABCDEFx0123);
      vacc5x0123 = _mm256_dpbusd_epi32(vacc5x0123, va5x01234567, vb01234567x0123);
      vacc5x4567 = _mm256_dpbusd_epi32(vacc5x4567, va5x01234567, vb89ABCDEFx0123);
      vacc6x0123 = _mm256_dpbusd_epi32(vacc6x0123, va6x01234567, vb01234567x0123);
      vacc6x4567 = _mm256_dpbusd_epi32(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);

    vout0x01234567 = _mm256_mul_ps(vout0x01234567, _mm256_set1_ps(quantization_params[0].inv_scale));
    vout1x01234567 = _mm256_mul_ps(vout1x01234567, _mm256_set1_ps(quantization_params[1].inv_scale));
    vout2x01234567 = _mm256_mul_ps(vout2x01234567, _mm256_set1_ps(quantization_params[2].inv_scale));
    vout3x01234567 = _mm256_mul_ps(vout3x01234567, _mm256_set1_ps(quantization_params[3].inv_scale));
    vout4x01234567 = _mm256_mul_ps(vout4x01234567, _mm256_set1_ps(quantization_params[4].inv_scale));
    vout5x01234567 = _mm256_mul_ps(vout5x01234567, _mm256_set1_ps(quantization_params[5].inv_scale));
    vout6x01234567 = _mm256_mul_ps(vout6x01234567, _mm256_set1_ps(quantization_params[6].inv_scale));

    const __m256 vfilter_output_scale01234567 = _mm256_load_ps((const float*) w);
    const __m256 vbias01234567 = _mm256_load_ps((const float*) w + 8);
    w = (const float*) w + 16;

    vout0x01234567 = _mm256_fmadd_ps(vout0x01234567, vfilter_output_scale01234567, vbias01234567);
    vout1x01234567 = _mm256_fmadd_ps(vout1x01234567, vfilter_output_scale01234567, vbias01234567);
    vout2x01234567 = _mm256_fmadd_ps(vout2x01234567, vfilter_output_scale01234567, vbias01234567);
    vout3x01234567 = _mm256_fmadd_ps(vout3x01234567, vfilter_output_scale01234567, vbias01234567);
    vout4x01234567 = _mm256_fmadd_ps(vout4x01234567, vfilter_output_scale01234567, vbias01234567);
    vout5x01234567 = _mm256_fmadd_ps(vout5x01234567, vfilter_output_scale01234567, vbias01234567);
    vout6x01234567 = _mm256_fmadd_ps(vout6x01234567, vfilter_output_scale01234567, vbias01234567);

    vout0x01234567 = _mm256_max_ps(vout0x01234567, voutput_min);
    vout1x01234567 = _mm256_max_ps(vout1x01234567, voutput_min);
    vout2x01234567 = _mm256_max_ps(vout2x01234567, voutput_min);
    vout3x01234567 = _mm256_max_ps(vout3x01234567, voutput_min);
    vout4x01234567 = _mm256_max_ps(vout4x01234567, voutput_min);
    vout5x01234567 = _mm256_max_ps(vout5x01234567, voutput_min);
    vout6x01234567 = _mm256_max_ps(vout6x01234567, voutput_min);

    vout0x01234567 = _mm256_min_ps(vout0x01234567, voutput_max);
    vout1x01234567 = _mm256_min_ps(vout1x01234567, voutput_max);
    vout2x01234567 = _mm256_min_ps(vout2x01234567, voutput_max);
    vout3x01234567 = _mm256_min_ps(vout3x01234567, voutput_max);
    vout4x01234567 = _mm256_min_ps(vout4x01234567, voutput_max);
    vout5x01234567 = _mm256_min_ps(vout5x01234567, voutput_max);
    vout6x01234567 = _mm256_min_ps(vout6x01234567, voutput_max);

    if XNN_LIKELY(nc >= 8) {
      _mm256_storeu_ps(c0, vout0x01234567);
      a0 = (const int8_t*) ((uintptr_t) a0 - kc);
      c0 = (float*) ((uintptr_t) c0 + cn_stride);
      _mm256_storeu_ps(c1, vout1x01234567);
      a1 = (const int8_t*) ((uintptr_t) a1 - kc);
      c1 = (float*) ((uintptr_t) c1 + cn_stride);
      _mm256_storeu_ps(c2, vout2x01234567);
      a2 = (const int8_t*) ((uintptr_t) a2 - kc);
      c2 = (float*) ((uintptr_t) c2 + cn_stride);
      _mm256_storeu_ps(c3, vout3x01234567);
      a3 = (const int8_t*) ((uintptr_t) a3 - kc);
      c3 = (float*) ((uintptr_t) c3 + cn_stride);
      _mm256_storeu_ps(c4, vout4x01234567);
      a4 = (const int8_t*) ((uintptr_t) a4 - kc);
      c4 = (float*) ((uintptr_t) c4 + cn_stride);
      _mm256_storeu_ps(c5, vout5x01234567);
      a5 = (const int8_t*) ((uintptr_t) a5 - kc);
      c5 = (float*) ((uintptr_t) c5 + cn_stride);
      _mm256_storeu_ps(c6, vout6x01234567);
      a6 = (const int8_t*) ((uintptr_t) a6 - kc);
      c6 = (float*) ((uintptr_t) c6 + cn_stride);
      nc -= 8;
    } else {
      // Prepare mask for valid 32-bit elements (depends on nc).
      const __mmask16 vmask = _cvtu32_mask16((UINT32_C(1) << nc) - 1);
      _mm256_mask_storeu_ps(c0, vmask, vout0x01234567);
      _mm256_mask_storeu_ps(c1, vmask, vout1x01234567);
      _mm256_mask_storeu_ps(c2, vmask, vout2x01234567);
      _mm256_mask_storeu_ps(c3, vmask, vout3x01234567);
      _mm256_mask_storeu_ps(c4, vmask, vout4x01234567);
      _mm256_mask_storeu_ps(c5, vmask, vout5x01234567);
      _mm256_mask_storeu_ps(c6, vmask, vout6x01234567);
      nc = 0;
    }
  } while (nc != 0);
}