src/qs8-vmulc/gen/minmax-fp32-sse2-mul16-ld64-x8.c

// Auto-generated file. Do not edit!
//   Template: src/qs8-vmulc/sse-mul16-ld64.c.in
//   Generator: tools/xngen
//
// Copyright 2021 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 <emmintrin.h>

#include <xnnpack/vmul.h>


void xnn_qs8_vmulc_minmax_fp32_ukernel__sse2_mul16_ld64_x8(
    size_t n,
    const int8_t* input_a,
    const int8_t* input_b,
    int8_t* output,
    const union xnn_qs8_mul_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN XNN_DISABLE_MSAN

{
  const __m128i va_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.a_zero_point);
  const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
  const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
  const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse2.output_min);
  const __m128i voutput_max = _mm_load_si128((const __m128i*) params->fp32_sse2.output_max);

  __m128i vxb = _mm_sub_epi16(
    _mm_shuffle_epi32(_mm_cvtsi32_si128(UINT32_C(0x00010001) * (uint32_t) (uint16_t) (int16_t) *input_b), 0),
    _mm_load_si128((const __m128i*) params->fp32_sse2.b_zero_point));
  for (; n >= 8 * sizeof(int8_t); n -= 8 * sizeof(int8_t)) {
    __m128i va01234567 = _mm_loadl_epi64((const __m128i*) input_a);
    input_a += 8;

    va01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(va01234567, va01234567), 8);

    const __m128i vxa01234567 = _mm_sub_epi16(va01234567, va_zero_point);

    const __m128i vprod01234567lo = _mm_mullo_epi16(vxa01234567, vxb);
    const __m128i vprod01234567hi = _mm_mulhi_epi16(vxa01234567, vxb);

    const __m128i vprod0123 = _mm_unpacklo_epi16(vprod01234567lo, vprod01234567hi);
    const __m128i vprod4567 = _mm_unpackhi_epi16(vprod01234567lo, vprod01234567hi);

    __m128 vfpacc0123 = _mm_cvtepi32_ps(vprod0123);
    __m128 vfpacc4567 = _mm_cvtepi32_ps(vprod4567);

    vfpacc0123 = _mm_mul_ps(vfpacc0123, vscale);
    vfpacc4567 = _mm_mul_ps(vfpacc4567, vscale);

    const __m128i vacc0123 = _mm_cvtps_epi32(vfpacc0123);
    const __m128i vacc4567 = _mm_cvtps_epi32(vfpacc4567);

    __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);

    vout01234567 = _mm_max_epi16(vout01234567, voutput_min);

    vout01234567 = _mm_min_epi16(vout01234567, voutput_max);

    __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);


    _mm_storel_epi64((__m128i*) output, vout0123456701234567);
    output += 8;
  }
  if XNN_UNLIKELY(n != 0) {
    {
      __m128i va01234567 = _mm_loadl_epi64((const __m128i*) input_a);

      va01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(va01234567, va01234567), 8);

      const __m128i vxa01234567 = _mm_sub_epi16(va01234567, va_zero_point);

      const __m128i vprod01234567lo = _mm_mullo_epi16(vxa01234567, vxb);
      const __m128i vprod01234567hi = _mm_mulhi_epi16(vxa01234567, vxb);

      const __m128i vprod0123 = _mm_unpacklo_epi16(vprod01234567lo, vprod01234567hi);
      const __m128i vprod4567 = _mm_unpackhi_epi16(vprod01234567lo, vprod01234567hi);

      __m128 vfpacc0123 = _mm_cvtepi32_ps(vprod0123);
      __m128 vfpacc4567 = _mm_cvtepi32_ps(vprod4567);

      vfpacc0123 = _mm_mul_ps(vfpacc0123, vscale);
      vfpacc4567 = _mm_mul_ps(vfpacc4567, vscale);

      const __m128i vacc0123 = _mm_cvtps_epi32(vfpacc0123);
      const __m128i vacc4567 = _mm_cvtps_epi32(vfpacc4567);

      __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
      vout01234567 = _mm_max_epi16(vout01234567, voutput_min);
      vout01234567 = _mm_min_epi16(vout01234567, voutput_max);

      __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);

      if (n & (4 * sizeof(int8_t))) {
        *((uint32_t*) output) = (uint32_t) _mm_cvtsi128_si32(vout0123456701234567);
        vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
        output += 4;
      }
      if (n & (2 * sizeof(int8_t))) {
        *((uint16_t*) output) = (uint16_t) _mm_cvtsi128_si32(vout0123456701234567);
        vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
        output += 2;
      }
      if (n & (1 * sizeof(int8_t))) {
        *output = (int8_t) _mm_cvtsi128_si32(vout0123456701234567);
      }
    }
  }
}