// 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/intrinsics-polyfill.h>
#include <xnnpack/vcvt.h>


$XINT8_T = {"QS8": "int8_t", "QU8": "uint8_t"}[DATATYPE]
$_MM256_CVTEPX8_EPI32 = {"QS8": "_mm256_cvtepi8_epi32", "QU8": "_mm256_cvtepu8_epi32"}[DATATYPE]
void xnn_${DATATYPE.lower()}_f32_vcvt_ukernel__avx2_x${BATCH_TILE}(
    size_t n,
    const ${XINT8_T}* x,
    float* y,
    const union xnn_${DATATYPE.lower()}_f32_cvt_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
{
  assert(n != 0);
  assert(n % sizeof(${XINT8_T}) == 0);
  assert(x != NULL);
  assert(y != NULL);

  const __m256i vminus_zero_point = _mm256_load_si256((const __m256i*) params->avx.minus_zero_point);
  const __m256 vscale = _mm256_load_ps(params->avx.scale);
  for (; n >= ${BATCH_TILE} * sizeof(${XINT8_T}); n -= ${BATCH_TILE} * sizeof(${XINT8_T})) {
    __m256i vx${ABC[0:8]} = ${_MM256_CVTEPX8_EPI32}(_mm_loadl_epi64((const __m128i*) x));
    $for N in range(8, BATCH_TILE, 8):
      __m256i vx${ABC[N:N+8]} = ${_MM256_CVTEPX8_EPI32}(_mm_loadl_epi64((const __m128i*) (x + ${N})));
    x += ${BATCH_TILE};

    $for N in range(0, BATCH_TILE, 8):
      vx${ABC[N:N+8]} = _mm256_add_epi32(vx${ABC[N:N+8]}, vminus_zero_point);

    $for N in range(0, BATCH_TILE, 8):
      __m256 vy${ABC[N:N+8]} = _mm256_cvtepi32_ps(vx${ABC[N:N+8]});

    $for N in range(0, BATCH_TILE, 8):
      vy${ABC[N:N+8]} = _mm256_mul_ps(vy${ABC[N:N+8]}, vscale);

    _mm256_storeu_ps(y, vy${ABC[0:8]});
    $for N in range(8, BATCH_TILE, 8):
      _mm256_storeu_ps(y + ${N}, vy${ABC[N:N+8]});
    y += ${BATCH_TILE};
  }
  for (; n >= 8 * sizeof(${XINT8_T}); n -= 8 * sizeof(${XINT8_T})) {
    __m256i vx = ${_MM256_CVTEPX8_EPI32}(_mm_loadl_epi64((const __m128i*) x));
    vx = _mm256_add_epi32(vx, vminus_zero_point);
    x += 8;

    __m256 vy = _mm256_cvtepi32_ps(vx);
    vy = _mm256_mul_ps(vy, vscale);

    _mm256_storeu_ps(y, vy);
    y += 8;
  }
  if XNN_UNLIKELY(n != 0) {
    assert(n >= 1 * sizeof(${XINT8_T}));
    assert(n <= 7 * sizeof(${XINT8_T}));

    __m256i vx = ${_MM256_CVTEPX8_EPI32}(_mm_loadl_epi64((const __m128i*) x));
    vx = _mm256_add_epi32(vx, vminus_zero_point);

    __m256 vy = _mm256_cvtepi32_ps(vx);
    vy = _mm256_mul_ps(vy, vscale);

    __m128 vy_lo = _mm256_castps256_ps128(vy);
    if (n & (4 * sizeof(${XINT8_T}))) {
      _mm_storeu_ps(y, vy_lo);
      vy_lo = _mm256_extractf128_ps(vy, 1);
      y += 4;
    }
    if (n & (2 * sizeof(${XINT8_T}))) {
      _mm_storel_pi((__m64*) y, vy_lo);
      vy_lo = _mm_movehl_ps(vy_lo, vy_lo);
      y += 2;
    }
    if (n & (1 * sizeof(${XINT8_T}))) {
      _mm_store_ss(y, vy_lo);
    }
  }
}