// 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 <arm_neon.h>

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


void xnn_f16_f32_vcvt_ukernel__neonfp16_x${BATCH_TILE}(
    size_t n,
    const void* input,
    float* output,
    const union xnn_f16_f32_cvt_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
{
  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 > 8:
    for (; n >= ${BATCH_TILE} * sizeof(uint16_t); n -= ${BATCH_TILE} * sizeof(uint16_t)) {
      $for N in range(SIMD_TILE):
        const float16x8_t vh${N} = vreinterpretq_f16_u16(vld1q_u16(i)); i += 8;

      $for N in range(SIMD_TILE):
        const float32x4_t vf${2*N} = vcvt_f32_f16(vget_low_f16(vh${N}));
        const float32x4_t vf${2*N+1} = vcvt_f32_f16(vget_high_f16(vh${N}));

      $for N in range(2*SIMD_TILE):
        vst1q_f32(output, vf${N}); output += 4;
    }
  for (; n >= 8 * sizeof(uint16_t); n -= 8 * sizeof(uint16_t)) {
    const float16x8_t vh = vreinterpretq_f16_u16(vld1q_u16(i)); i += 8;

    const float32x4_t vf_lo = vcvt_f32_f16(vget_low_f16(vh));
    const float32x4_t vf_hi = vcvt_f32_f16(vget_high_f16(vh));

    vst1q_f32(output, vf_lo); output += 4;
    vst1q_f32(output, vf_hi); output += 4;
  }
  if XNN_UNLIKELY(n != 0) {
    assert(n >= 1 * sizeof(uint16_t));
    assert(n <= 7 * sizeof(uint16_t));
    const float16x8_t vh = vreinterpretq_f16_u16(vld1q_u16(i)); i += 8;

    float32x4_t vf = vcvt_f32_f16(vget_low_f16(vh));
    if (n & (4 * sizeof(uint16_t))) {
      vst1q_f32(output, vf); output += 4;
      vf = vcvt_f32_f16(vget_high_f16(vh));
    }
    float32x2_t vf_lo = vget_low_f32(vf);
    if (n & (2 * sizeof(uint16_t))) {
      vst1_f32(output, vf_lo); output += 2;
      vf_lo = vget_high_f32(vf);
    }
    if (n & (1 * sizeof(uint16_t))) {
      vst1_lane_f32(output, vf_lo, 0);
    }
  }
}