xref: /aosp_15_r20/external/XNNPACK/src/f16-igemm/gen/1x8-minmax-avx2-broadcast.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)

// Auto-generated file. Do not edit!
//   Template: src/f16-igemm/avx2-broadcast.c.in
//   Generator: tools/xngen
//
// Copyright 2022 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>


void xnn_f16_igemm_minmax_ukernel_1x8__avx2_broadcast(
    size_t mr,
    size_t nc,
    size_t kc,
    size_t ks,
    const void**restrict a,
    const void*restrict w,
    void*restrict c,
    size_t cm_stride,
    size_t cn_stride,
    size_t a_offset,
    const void* zero,
    const union xnn_f16_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
{
  assert(mr != 0);
  assert(mr <= 1);
  assert(nc != 0);
  assert(kc != 0);
  assert(kc % sizeof(uint16_t) == 0);
  assert(ks != 0);
  assert(ks % (1 * sizeof(void*)) == 0);
  assert(a_offset % sizeof(uint16_t) == 0);
  assert(a != NULL);
  assert(w != NULL);
  assert(c != NULL);

  uint16_t* c0 = c;

  do {
    __m256 vacc0x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) w));
    w = (const uint16_t*) w + 8;

    size_t p = ks;
    do {
      const uint16_t* restrict a0 = (const uint16_t*) a[0];
      assert(a0 != NULL);
      if XNN_UNPREDICTABLE(a0 != zero) {
        a0 = (const uint16_t*) ((uintptr_t) a0 + a_offset);
      }
      a += 1;

      size_t k = kc;
      do {
        const __m256 vb01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) w));
        w = (const uint16_t*) w + 8;

        const __m256 va0 = _mm256_cvtph_ps(_mm_set1_epi16((short) *a0));
        a0 += 1;

        vacc0x01234567 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(va0, vb01234567, vacc0x01234567), _MM_FROUND_NO_EXC));

        k -= sizeof(uint16_t);
      } while (k != 0);
      p -= 1 * sizeof(void*);
    } while (p != 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) {
      _mm_storeu_si128((__m128i*) c0, _mm256_cvtps_ph(vacc0x01234567, _MM_FROUND_NO_EXC));
      c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);

      a = (const void**restrict) ((uintptr_t) a - ks);
      nc -= 8;
    } else {
      __m128i vh0x01234567 = _mm256_cvtps_ph(vacc0x01234567, _MM_FROUND_NO_EXC);
      if (nc & 4) {
        _mm_storel_epi64((__m128i*) c0, vh0x01234567);

        vh0x01234567 = _mm_unpackhi_epi64(vh0x01234567, vh0x01234567);

        c0 += 4;
      }
      if (nc & 2) {
        _mm_storeu_si32(c0, vh0x01234567);

        vh0x01234567 = _mm_srli_epi64(vh0x01234567, 32);

        c0 += 2;
      }
      if (nc & 1) {
        *c0 = _mm_extract_epi16(vh0x01234567, 0);
      }

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