// 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 % 16 == 0
$assert BATCH_TILE >= 16
$SIMD_TILE = BATCH_TILE // 16
#include <assert.h>

#include <immintrin.h>

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


void xnn_f16_f32_vcvt_ukernel__avx512skx_x${BATCH_TILE}(
    size_t n,
    const void* input,
    float* output,
    const union xnn_f16_f32_cvt_params params[restrict XNN_MIN_ELEMENTS(1)])
{
  assert(n != 0);
  assert(n % sizeof(uint16_t) == 0);
  assert(input != NULL);
  assert(output != NULL);

  const uint16_t* i = (const uint16_t*) input;
  $if BATCH_TILE > 16:
    for (; n >= ${BATCH_TILE} * sizeof(uint16_t); n -= ${BATCH_TILE} * sizeof(uint16_t)) {
      const __m512 vacc0 = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i*) i));
      $for N in range(1, SIMD_TILE):
        const __m512 vacc${N} = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i*) (i + ${N * 16})));
      i += ${BATCH_TILE};

      _mm512_storeu_ps(output, vacc0);
      $for N in range(1, SIMD_TILE):
        _mm512_storeu_ps(output + ${N * 16}, vacc${N});
      output += ${BATCH_TILE};
    }
  for (; n >= 16 * sizeof(uint16_t); n -= 16 * sizeof(uint16_t)) {
    const __m512 vacc = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i*) i));
    i += 16;

    _mm512_storeu_ps(output, vacc);
    output += 16;
  }
  if XNN_UNLIKELY(n != 0) {
    assert(n >= 1 * sizeof(uint16_t));
    assert(n <= 15 * sizeof(uint16_t));

    // Prepare mask for valid 32-bit elements (depends on n).
    n >>= 1 /* log2(sizeof(uint16_t)) */;
    const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << n) - UINT32_C(1)));

    const __m512 vacc = _mm512_cvtph_ps(_mm256_maskz_loadu_epi16(vmask, i));

    _mm512_mask_storeu_ps(output, vmask, vacc);
  }
}