// 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 // 8
#include <assert.h>

#include <immintrin.h>

#include <xnnpack/common.h>
#include <xnnpack/vcvt.h>


void xnn_f16_f32_vcvt_ukernel__f16c_x${BATCH_TILE}(
    size_t batch,
    const void* input,
    float* output,
    const union xnn_f16_f32_cvt_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
{
  assert(batch != 0);
  assert(batch % sizeof(uint16_t) == 0);
  assert(input != NULL);
  assert(output != NULL);

  const uint16_t* i = (const uint16_t*) input;
  $if BATCH_TILE > 8:
    for (; batch >= ${BATCH_TILE} * sizeof(uint16_t); batch -= ${BATCH_TILE} * sizeof(uint16_t)) {
      const __m256 vacc0 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i));
      $for N in range(1, SIMD_TILE):
        const __m256 vacc${N} = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i + ${N * 8})));
      i += ${BATCH_TILE};

      _mm256_storeu_ps(output, vacc0);
      $for N in range(1, SIMD_TILE):
        _mm256_storeu_ps(output + ${N * 8}, vacc${N});
      output += ${BATCH_TILE};
    }
  for (; batch >= 8 * sizeof(uint16_t); batch -= 8 * sizeof(uint16_t)) {
    const __m256 vacc = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i));
    i += 8;

    _mm256_storeu_ps(output, vacc);
    output += 8;
  }
  if XNN_UNLIKELY(batch != 0) {
    assert(batch >= 1 * sizeof(uint16_t));
    assert(batch <= 7 * sizeof(uint16_t));
    const __m256 vacc = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i));

    __m128 vacc_lo = _mm256_castps256_ps128(vacc);
    if (batch & (4 * sizeof(uint16_t))) {
      _mm_storeu_ps(output, vacc_lo);
      vacc_lo = _mm256_extractf128_ps(vacc, 1);
      output += 4;
    }
    if (batch & (2 * sizeof(uint16_t))) {
      _mm_storel_pi((__m64*) output, vacc_lo);
      vacc_lo = _mm_movehl_ps(vacc_lo, vacc_lo);
      output += 2;
    }
    if (batch & (1 * sizeof(uint16_t))) {
      _mm_store_ss(output, vacc_lo);
    }
  }
}