// 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
$SIMD_TILE = BATCH_TILE // 16
$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]
$_MM512_CVTEPX8_EPI32 = {"QS8": "_mm512_cvtepi8_epi32", "QU8": "_mm512_cvtepu8_epi32"}[DATATYPE]
void xnn_${DATATYPE.lower()}_f32_vcvt_ukernel__avx512skx_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 __m512i vzero_point = _mm512_set1_epi32(params->scalar.zero_point);
  const __m512 vscale = _mm512_set1_ps(params->scalar.scale);
  XNN_FORCE_REALIZATION(vzero_point);
  XNN_FORCE_REALIZATION(vscale);
  $if BATCH_TILE > 16:
    for (; batch >= ${BATCH_TILE} * sizeof(${XINT8_T}); batch -= ${BATCH_TILE} * sizeof(${XINT8_T})) {
      __m512i vx${ABC[0:16]} = ${_MM512_CVTEPX8_EPI32}(_mm_loadu_si128((const __m128i*) input));
      $for N in range(16, BATCH_TILE, 16):
        __m512i vx${ABC[N:N+16]} = ${_MM512_CVTEPX8_EPI32}(_mm_loadu_si128((const __m128i*) (input + ${N})));
      input += ${BATCH_TILE};

      $for N in range(0, BATCH_TILE, 16):
        vx${ABC[N:N+16]} = _mm512_sub_epi32(vx${ABC[N:N+16]}, vzero_point);

      $for N in range(0, BATCH_TILE, 16):
        __m512 vy${ABC[N:N+16]} = _mm512_cvtepi32_ps(vx${ABC[N:N+16]});

      $for N in range(0, BATCH_TILE, 16):
        vy${ABC[N:N+16]} = _mm512_mul_ps(vy${ABC[N:N+16]}, vscale);

      _mm512_storeu_ps(output, vy${ABC[0:16]});
      $for N in range(16, BATCH_TILE, 16):
        _mm512_storeu_ps(output + ${N}, vy${ABC[N:N+16]});
      output += ${BATCH_TILE};
    }
  for (; batch >= 16 * sizeof(${XINT8_T}); batch -= 16 * sizeof(${XINT8_T})) {
    __m512i vx = ${_MM512_CVTEPX8_EPI32}(_mm_loadu_si128((const __m128i*) input));
    vx = _mm512_sub_epi32(vx, vzero_point);
    input += 16;

    __m512 vy = _mm512_cvtepi32_ps(vx);
    vy = _mm512_mul_ps(vy, vscale);

    _mm512_storeu_ps(output, vy);
    output += 16;
  }
  if XNN_UNLIKELY(batch != 0) {
    assert(batch >= 1 * sizeof(${XINT8_T}));
    assert(batch <= 15 * sizeof(${XINT8_T}));

    // Prepare mask for valid elements (depends on batch).
    const __mmask16 vmask = _cvtu32_mask16((uint32_t) ((UINT32_C(1) << batch) - UINT32_C(1)));

    __m512i vx = ${_MM512_CVTEPX8_EPI32}(_mm_maskz_loadu_epi8(vmask, input));
    vx = _mm512_sub_epi32(vx, vzero_point);

    __m512 vy = _mm512_cvtepi32_ps(vx);
    vy = _mm512_mul_ps(vy, vscale);

    _mm512_mask_storeu_ps(output, vmask, vy);
  }
}