// Auto-generated file. Do not edit!
//   Template: src/f32-qs8-vcvt/sse.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/common.h"
#include "xnnpack/unaligned.h"
#include "xnnpack/vcvt.h"


void xnn_f32_qs8_vcvt_ukernel__sse2_u8(
    size_t batch,
    const float* input,
    int8_t* output,
    const struct xnn_f32_qs8_cvt_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
{
  assert(batch != 0);
  assert(batch % sizeof(float) == 0);
  assert(input != NULL);
  assert(output != NULL);

  const __m128 vscale = _mm_set1_ps(params->scalar.scale);
  const __m128 voutput_max_less_zero_point = _mm_set1_ps((float) ((int32_t) 127 - (int32_t) params->scalar.output_zero_point));
  const __m128i voutput_zero_point = _mm_set1_epi16(params->scalar.output_zero_point);
  XNN_FORCE_REALIZATION(vscale);
  XNN_FORCE_REALIZATION(voutput_max_less_zero_point);
  XNN_FORCE_REALIZATION(voutput_zero_point);

  for (; batch >= 8 * sizeof(float); batch -= 8 * sizeof(float)) {
    __m128 vx_lo = _mm_loadu_ps(input);
    __m128 vx_hi = _mm_loadu_ps(input + 4);
    input += 8;

    vx_lo = _mm_mul_ps(vx_lo, vscale);
    vx_hi = _mm_mul_ps(vx_hi, vscale);

    vx_lo = _mm_min_ps(vx_lo, voutput_max_less_zero_point);
    vx_hi = _mm_min_ps(vx_hi, voutput_max_less_zero_point);

    const __m128i vy_lo = _mm_cvtps_epi32(vx_lo);
    const __m128i vy_hi = _mm_cvtps_epi32(vx_hi);

    __m128i vy = _mm_packs_epi32(vy_lo, vy_hi);
    vy = _mm_adds_epi16(vy, voutput_zero_point);
    vy = _mm_packs_epi16(vy, vy);

    _mm_storel_epi64((__m128i*) output, vy);
    output += 8;
  }
  if XNN_UNLIKELY(batch != 0) {
    __m128 vx_lo = _mm_loadu_ps(input);
    const float* x_hi = (const float*) ((uintptr_t) input + (batch & (4 * sizeof(float))));
    __m128 vx_hi = _mm_loadu_ps(x_hi);

    vx_lo = _mm_mul_ps(vx_lo, vscale);
    vx_hi = _mm_mul_ps(vx_hi, vscale);

    vx_lo = _mm_min_ps(vx_lo, voutput_max_less_zero_point);
    vx_hi = _mm_min_ps(vx_hi, voutput_max_less_zero_point);

    const __m128i vy_lo = _mm_cvtps_epi32(vx_lo);
    const __m128i vy_hi = _mm_cvtps_epi32(vx_hi);

    __m128i vy = _mm_packs_epi32(vy_lo, vy_hi);
    vy = _mm_adds_epi16(vy, voutput_zero_point);
    vy = _mm_packs_epi16(vy, vy);

    if (batch & (4 * sizeof(float))) {
      unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vy));
      output += 4;
      vy = _mm_srli_epi64(vy, 32);
    }
    {
      uint32_t vy_lo = (uint32_t) _mm_cvtsi128_si32(vy);
      if (batch & (2 * sizeof(float))) {
        unaligned_store_u16(output, (uint16_t) vy_lo);
        output += 2;
        vy_lo >>= 16;
      }
      if (batch & (1 * sizeof(float))) {
        *output = (int8_t) vy_lo;
      }
    }
  }
}