// Auto-generated file. Do not edit!
//   Template: src/qs8-igemm/MRx8c8-avx2.c.in
//   Generator: tools/xngen
//
// Copyright 2020 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/igemm.h"
#include "xnnpack/intrinsics-polyfill.h"
#include "xnnpack/math.h"
#include "xnnpack/unaligned.h"


void xnn_qd8_f32_qc8w_igemm_minmax_ukernel_1x8c8__avx256skx(
    size_t mr,
    size_t nc,
    size_t kc,
    size_t ks,
    const int8_t** restrict a,
    const void* restrict w,
    float* restrict c,
    size_t cm_stride,
    size_t cn_stride,
    size_t a_offset,
    const int8_t* zero,
    const int8_t* zero_data,
    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 <= 1);
  assert(nc != 0);
  assert(kc != 0);
  assert(ks != 0);
  assert(ks % (1 * sizeof(void*)) == 0);
  assert(a_offset % sizeof(int8_t) == 0);
  assert(a != NULL);
  assert(w != NULL);
  assert(c != NULL);

  kc = round_up_po2(kc, 8 * sizeof(int8_t));
  float* c0 = c;

  const __m256i vinput_zero_point = _mm256_set1_epi32((int) quantization_params->zero_point);
  const __m256 vinput_scale = _mm256_broadcast_ss(&quantization_params->inv_scale);
  const __m256 vmin = _mm256_set1_ps(params->scalar.min);
  const __m256 vmax = _mm256_set1_ps(params->scalar.max);
  // XNN_FORCE_REALIZATION(vinput_zero_point);
  // XNN_FORCE_REALIZATION(vinput_scale);
  // XNN_FORCE_REALIZATION(vmin);
  // XNN_FORCE_REALIZATION(vmax);

  do {
    const __m128i vinit0 = _mm_cvtsi32_si128(((const int*) w)[0]);
    const __m128i vinit1 = _mm_cvtsi32_si128(((const int*) w)[1]);
    const __m256i vinit01 = _mm256_inserti128_si256(_mm256_castsi128_si256(vinit0), vinit1, 1);
    const __m128i vinit2 = _mm_cvtsi32_si128(((const int*) w)[2]);
    const __m128i vinit3 = _mm_cvtsi32_si128(((const int*) w)[3]);
    const __m256i vinit23 = _mm256_inserti128_si256(_mm256_castsi128_si256(vinit2), vinit3, 1);
    const __m128i vinit4 = _mm_cvtsi32_si128(((const int*) w)[4]);
    const __m128i vinit5 = _mm_cvtsi32_si128(((const int*) w)[5]);
    const __m256i vinit45 = _mm256_inserti128_si256(_mm256_castsi128_si256(vinit4), vinit5, 1);
    const __m128i vinit6 = _mm_cvtsi32_si128(((const int*) w)[6]);
    const __m128i vinit7 = _mm_cvtsi32_si128(((const int*) w)[7]);
    const __m256i vinit67 = _mm256_inserti128_si256(_mm256_castsi128_si256(vinit6), vinit7, 1);
    __m256i vacc0x01 = _mm256_mullo_epi32(vinit01, vinput_zero_point);
    __m256i vacc0x23 = _mm256_mullo_epi32(vinit23, vinput_zero_point);
    __m256i vacc0x45 = _mm256_mullo_epi32(vinit45, vinput_zero_point);
    __m256i vacc0x67 = _mm256_mullo_epi32(vinit67, vinput_zero_point);
    w = (const int32_t*) w + 8;

    size_t p = ks;
    do {
      const int8_t* restrict a0 = a[0];
      if XNN_UNPREDICTABLE(a0 != zero) {
        a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
      } else {
        a0 = zero_data;
      }
      a += 1;

      size_t k = 0;
      while (k < kc) {
        const __m128i va0 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a0));
        const __m256i vxa0 = _mm256_cvtepi8_epi16(va0);
        a0 += 8;

        const __m128i vb01 = _mm_load_si128((const __m128i*) w);
        const __m256i vxb01 = _mm256_cvtepi8_epi16(vb01);

        vacc0x01 = _mm256_add_epi32(vacc0x01, _mm256_madd_epi16(vxa0, vxb01));
        const __m128i vb23 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 16));
        const __m256i vxb23 = _mm256_cvtepi8_epi16(vb23);

        vacc0x23 = _mm256_add_epi32(vacc0x23, _mm256_madd_epi16(vxa0, vxb23));
        const __m128i vb45 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 32));
        const __m256i vxb45 = _mm256_cvtepi8_epi16(vb45);

        vacc0x45 = _mm256_add_epi32(vacc0x45, _mm256_madd_epi16(vxa0, vxb45));
        const __m128i vb67 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 48));
        const __m256i vxb67 = _mm256_cvtepi8_epi16(vb67);

        vacc0x67 = _mm256_add_epi32(vacc0x67, _mm256_madd_epi16(vxa0, vxb67));

        w = (const void*) ((const int8_t*) w + 64);
        k += 8 * sizeof(int8_t);
      }
      p -= 1 * sizeof(void*);
    } while (p != 0);

    const __m256i vacc0x0213 = _mm256_hadd_epi32(vacc0x01, vacc0x23);
    const __m256i vacc0x4657 = _mm256_hadd_epi32(vacc0x45, vacc0x67);

    const __m256i vacc0x02461357 = _mm256_hadd_epi32(vacc0x0213, vacc0x4657);

    const __m256i vpermute_mask = _mm256_set_epi32(7, 3, 6, 2, 5, 1, 4, 0);
    __m256i vacc0x01234567 = _mm256_permutevar8x32_epi32(vacc0x02461357, vpermute_mask);

    __m256 vout0x01234567 = _mm256_cvtepi32_ps(vacc0x01234567);

    vout0x01234567 = _mm256_mul_ps(vout0x01234567, vinput_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);

    vout0x01234567 = _mm256_max_ps(vout0x01234567, vmin);

    vout0x01234567 = _mm256_min_ps(vout0x01234567, vmax);

    if XNN_LIKELY(nc >= 8) {
      _mm256_storeu_ps(c0, vout0x01234567);
      c0 = (float*) ((uintptr_t) c0 + cn_stride);

      a = (const int8_t**restrict) ((uintptr_t) a - ks);
      nc -= 8;
    } else {
      // Prepare mask for valid 32-bit elements (depends on nc).
      const __mmask8 vmask = _cvtu32_mask8((UINT32_C(1) << nc) - 1);
      _mm256_mask_storeu_ps(c0, vmask, vout0x01234567);
      nc = 0;
    }
  } while (nc != 0);
}