// 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.

$assert BATCH_TILE % 8 == 0
$assert BATCH_TILE >= 8
$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
#include <assert.h>

#include <immintrin.h>

#include "xnnpack/common.h"
#include "xnnpack/unaligned.h"
#include "xnnpack/vcvt.h"


$XINT8_T = {"QS8": "int8_t", "QU8": "uint8_t"}[DATATYPE]
$_MM_CVTEPX8_EPI32 = {"QS8": "_mm_cvtepi8_epi32", "QU8": "_mm_cvtepu8_epi32"}[DATATYPE]
void xnn_${DATATYPE.lower()}_f32_vcvt_ukernel__sse41_u${BATCH_TILE}(
    size_t batch,
    const ${XINT8_T}* input,
    float* output,
    const struct xnn_${DATATYPE.lower()}_f32_cvt_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
{
  assert(batch != 0);
  assert(batch % sizeof(${XINT8_T}) == 0);
  assert(input != NULL);
  assert(output != NULL);

  const __m128i vzero_point = _mm_set1_epi32(params->scalar.zero_point);
  const __m128 vscale = _mm_set1_ps(params->scalar.scale);
  XNN_FORCE_REALIZATION(vzero_point);
  XNN_FORCE_REALIZATION(vscale);
  $if BATCH_TILE > 4:
    for (; batch >= ${BATCH_TILE} * sizeof(${XINT8_T}); batch -= ${BATCH_TILE} * sizeof(${XINT8_T})) {
      __m128i vx${ABC[0:4]} = ${_MM_CVTEPX8_EPI32}(_mm_cvtsi32_si128((int) unaligned_load_s32(input)));
      $for N in range(4, BATCH_TILE, 4):
        __m128i vx${ABC[N:N+4]} = ${_MM_CVTEPX8_EPI32}(_mm_cvtsi32_si128((int) unaligned_load_s32(input + ${N})));
      input += ${BATCH_TILE};

      $for N in range(0, BATCH_TILE, 4):
        vx${ABC[N:N+4]} = _mm_sub_epi32(vx${ABC[N:N+4]}, vzero_point);

      $for N in range(0, BATCH_TILE, 4):
        __m128 vy${ABC[N:N+4]} = _mm_cvtepi32_ps(vx${ABC[N:N+4]});

      $for N in range(0, BATCH_TILE, 4):
        vy${ABC[N:N+4]} = _mm_mul_ps(vy${ABC[N:N+4]}, vscale);

      _mm_storeu_ps(output, vy${ABC[0:4]});
      $for N in range(4, BATCH_TILE, 4):
        _mm_storeu_ps(output + ${N}, vy${ABC[N:N+4]});
      output += ${BATCH_TILE};
    }
  for (; batch >= 4 * sizeof(${XINT8_T}); batch -= 4 * sizeof(${XINT8_T})) {
    __m128i vx = ${_MM_CVTEPX8_EPI32}(_mm_cvtsi32_si128((int) unaligned_load_s32(input)));
    vx = _mm_sub_epi32(vx, vzero_point);
    input += 4;

    __m128 vy = _mm_cvtepi32_ps(vx);
    vy = _mm_mul_ps(vy, vscale);

    _mm_storeu_ps(output, vy);
    output += 4;
  }
  if XNN_UNLIKELY(batch != 0) {
    assert(batch >= 1 * sizeof(${XINT8_T}));
    assert(batch <= 3 * sizeof(${XINT8_T}));

    __m128i vx = ${_MM_CVTEPX8_EPI32}(_mm_cvtsi32_si128((int) unaligned_load_s32(input)));
    vx = _mm_sub_epi32(vx, vzero_point);

    __m128 vy = _mm_cvtepi32_ps(vx);
    vy = _mm_mul_ps(vy, vscale);

    if (batch & (2 * sizeof(${XINT8_T}))) {
      _mm_storel_pi((__m64*) output, vy);
      vy = _mm_movehl_ps(vy, vy);
      output += 2;
    }
    if (batch & (1 * sizeof(${XINT8_T}))) {
      _mm_store_ss(output, vy);
    }
  }
}