xref: /aosp_15_r20/external/XNNPACK/src/f32-gemm/gen/1x8-minmax-fma3-broadcast.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)

// Auto-generated file. Do not edit!
//   Template: src/f32-gemm/avx-broadcast.c.in
//   Generator: tools/xngen
//
// Copyright 2019 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>


void xnn_f32_gemm_minmax_ukernel_1x8__fma3_broadcast(
    size_t mr,
    size_t nc,
    size_t kc,
    const float*restrict a,
    size_t a_stride,
    const float*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)])
{
  assert(mr != 0);
  assert(mr <= 1);
  assert(nc != 0);
  assert(kc != 0);
  assert(kc % sizeof(float) == 0);
  assert(a != NULL);
  assert(w != NULL);
  assert(c != NULL);

  const float* a0 = a;
  float* c0 = c;

  do {
    __m256 vacc0x01234567 = _mm256_load_ps(w + 0);
    w += 8;

    size_t k = kc;
    do {
      const __m256 va0 = _mm256_broadcast_ss(a0);
      a0 += 1;

      const __m256 vb01234567 = _mm256_load_ps(w);
      w += 8;

      vacc0x01234567 = _mm256_fmadd_ps(va0, vb01234567, vacc0x01234567);

      k -= sizeof(float);
    } while (k != 0);

    const __m256 vmin = _mm256_load_ps(params->avx.min);
    vacc0x01234567 = _mm256_max_ps(vacc0x01234567, vmin);

    const __m256 vmax = _mm256_load_ps(params->avx.max);
    vacc0x01234567 = _mm256_min_ps(vacc0x01234567, vmax);

    if XNN_LIKELY(nc >= 8) {
      _mm256_storeu_ps(c0, vacc0x01234567);
      c0 = (float*) ((uintptr_t) c0 + cn_stride);

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

      nc -= 8;
    } else {
      __m128 vacc0x0123 = _mm256_castps256_ps128(vacc0x01234567);
      if (nc & 4) {
        _mm_storeu_ps(c0, vacc0x0123);

        vacc0x0123 = _mm256_extractf128_ps(vacc0x01234567, 1);

        c0 += 4;
      }
      if (nc & 2) {
        _mm_storel_pi((__m64*) c0, vacc0x0123);

        vacc0x0123 = _mm_movehl_ps(vacc0x0123, vacc0x0123);

        c0 += 2;
      }
      if (nc & 1) {
        _mm_store_ss(c0, vacc0x0123);
      }

      nc = 0;
    }
  } while (nc != 0);
}