// 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_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 __m256i vzero_point = _mm256_set1_epi32(params->scalar.zero_point);
  const __m256 vscale = _mm256_set1_ps(params->scalar.scale);
  XNN_FORCE_REALIZATION(vzero_point);
  XNN_FORCE_REALIZATION(vscale);
  for (; batch >= ${BATCH_TILE} * sizeof(${XINT8_T}); batch -= ${BATCH_TILE} * sizeof(${XINT8_T})) {
    __m256i vx${ABC[0:8]} = ${_MM256_CVTEPX8_EPI32}(_mm_loadl_epi64((const __m128i*) input));
    $for N in range(8, BATCH_TILE, 8):
      __m256i vx${ABC[N:N+8]} = ${_MM256_CVTEPX8_EPI32}(_mm_loadl_epi64((const __m128i*) (input + ${N})));
    input += ${BATCH_TILE};

    $for N in range(0, BATCH_TILE, 8):
      vx${ABC[N:N+8]} = _mm256_sub_epi32(vx${ABC[N:N+8]}, vzero_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(output, vy${ABC[0:8]});
    $for N in range(8, BATCH_TILE, 8):
      _mm256_storeu_ps(output + ${N}, vy${ABC[N:N+8]});
    output += ${BATCH_TILE};
  }
  for (; batch >= 8 * sizeof(${XINT8_T}); batch -= 8 * sizeof(${XINT8_T})) {
    __m256i vx = ${_MM256_CVTEPX8_EPI32}(_mm_loadl_epi64((const __m128i*) input));
    vx = _mm256_sub_epi32(vx, vzero_point);
    input += 8;

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

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

    __m256i vx = ${_MM256_CVTEPX8_EPI32}(_mm_loadl_epi64((const __m128i*) input));
    vx = _mm256_sub_epi32(vx, vzero_point);

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

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