// Copyright 2022 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
$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
$assert OP in ["ABS", "NEG", "SQR"]
#include <assert.h>

#include <arm_neon.h>

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


$VOPQ_F16 = {
$  "ABS": lambda x: "vabsq_f16(%s)" % x,
$  "NEG": lambda x: "vnegq_f16(%s)" % x,
$  "SQR": lambda x: "vmulq_f16(%s, %s)" % (x, x),
$}[OP]
$PARAMS = {
$  "ABS": "xnn_f16_abs_params",
$  "NEG": "xnn_f16_neg_params",
$  "SQR": "xnn_f16_default_params",
$}[OP]
void xnn_f16_v${OP.lower()}_ukernel__neonfp16arith_x${BATCH_TILE}(
    size_t batch,
    const void* input,
    void* output,
    const union ${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;
  uint16_t* o = (uint16_t*) output;
  $if BATCH_TILE > 8:
    for (; batch >= ${BATCH_TILE} * sizeof(uint16_t); batch -= ${BATCH_TILE} * sizeof(uint16_t)) {
      $for N in range(SIMD_TILE):
        float16x8_t vacc${ABC[N]} = vreinterpretq_f16_u16(vld1q_u16(i)); i += 8;

      $for N in range(SIMD_TILE):
        vacc${ABC[N]} = ${VOPQ_F16("vacc" + ABC[N])};

      $for N in range(SIMD_TILE):
        vst1q_u16(o, vreinterpretq_u16_f16(vacc${ABC[N]})); o += 8;
    }
  for (; batch >= 8 * sizeof(uint16_t); batch -= 8 * sizeof(uint16_t)) {
    float16x8_t vacc = vreinterpretq_f16_u16(vld1q_u16(i)); i += 8;
    vacc = ${VOPQ_F16("vacc")};
    vst1q_u16(o, vreinterpretq_u16_f16(vacc)); o += 8;
  }
  if XNN_UNLIKELY(batch != 0) {
    float16x8_t vacc = vreinterpretq_f16_u16(vld1q_u16(i));
    vacc = ${VOPQ_F16("vacc")};
    float16x4_t vacc_lo = vget_low_f16(vacc);
    if (batch & (4 * sizeof(uint16_t))) {
      vst1_u16(o, vreinterpret_u16_f16(vacc_lo)); o += 4;
      vacc_lo = vget_high_f16(vacc);
    }
    if (batch & (2 * sizeof(uint16_t))) {
      vst1_lane_u32((void*) o, vreinterpret_u32_f16(vacc_lo), 0); o += 2;
      vacc_lo = vext_f16(vacc_lo, vacc_lo, 2);
    }
    if (batch & (1 * sizeof(uint16_t))) {
      vst1_lane_u16(o, vreinterpret_u16_f16(vacc_lo), 0);
    }
  }
}