// 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_f32_f16_vcvt_ukernel__neonfp16_x${BATCH_TILE}(
    size_t batch,
    const float* input,
    void* output,
    const union xnn_f32_f16_cvt_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
{
  assert(batch != 0);
  assert(batch % sizeof(float) == 0);
  assert(input != NULL);
  assert(output != NULL);

  uint16_t* o = (uint16_t*) output;
  for (; batch >= ${BATCH_TILE} * sizeof(float); batch -= ${BATCH_TILE} * sizeof(float)) {
    $for N in range(2*SIMD_TILE):
      const float32x4_t vf${N} = vld1q_f32(input); input += 4;

    $for N in range(SIMD_TILE):
      const uint16x8_t vh${N} = vreinterpretq_u16_f16(vcombine_f16(vcvt_f16_f32(vf${2*N}), vcvt_f16_f32(vf${2*N+1})));

    $for N in range(SIMD_TILE):
      vst1q_u16(o, vh${N}); o += 8;
  }
  for (; batch >= 4 * sizeof(float); batch -= 4 * sizeof(float)) {
    const float32x4_t vf = vld1q_f32(input); input += 4;

    const uint16x4_t vh = vreinterpret_u16_f16(vcvt_f16_f32(vf));

    vst1_u16(o, vh); o += 4;
  }
  if XNN_UNLIKELY(batch != 0) {
    assert(batch % sizeof(float) == 0);
    assert(batch >= 1 * sizeof(float));
    assert(batch <= 3 * sizeof(float));
    const float32x4_t vf = vld1q_f32(input);

    uint16x4_t vh = vreinterpret_u16_f16(vcvt_f16_f32(vf));

    if (batch & (2 * sizeof(float))) {
      vst1_lane_u32((void*) o, vreinterpret_u32_u16(vh), 0); o += 2;
      vh = vext_u16(vh, vh, 2);
    }
    if (batch & (1 * sizeof(float))) {
      vst1_lane_u16(o, vh, 0);
    }
  }
}