// 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 >= 1
$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
#include <assert.h>
#include <math.h>

#include <arm_fp16.h>

#include "xnnpack/common.h"
#include "xnnpack/intrinsics-polyfill.h"
#include "xnnpack/vunary.h"


void xnn_f16_vsqrt_ukernel__fp16arith_sqrt_u${BATCH_TILE}(
    size_t batch,
    const xnn_float16* input,
    xnn_float16* output,
    const struct xnn_f16_default_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
{
  assert(batch != 0);
  assert(batch % sizeof(float16_t) == 0);
  assert(input != NULL);
  assert(output != NULL);

  const float16_t* i = (const float16_t*) input;
  float16_t* o = (float16_t*) output;
  $if BATCH_TILE > 1:
    for (; batch >= ${BATCH_TILE} * sizeof(float16_t); batch -= ${BATCH_TILE} * sizeof(float16_t)) {
      $for N in range(BATCH_TILE):
        float16_t vacc${ABC[N]} = *i++;

      $for N in range(BATCH_TILE):
        vacc${ABC[N]} = vsqrth_f16(vacc${ABC[N]});

      $for N in range(BATCH_TILE):
        *o++ = vacc${ABC[N]};
    }
    if XNN_UNLIKELY(batch != 0) {
      $if BATCH_TILE > 2:
        do {
          float16_t vacc = *i++;
          vacc = vsqrth_f16(vacc);
          *o++ = vacc;
          batch -= sizeof(float16_t);
        } while (batch != 0);
      $else:
        float16_t vacc = *i;
        vacc = vsqrth_f16(vacc);
        *o = vacc;
    }
  $else:
    do {
      float16_t vacc = *i++;
      vacc = vsqrth_f16(vacc);
      *o++ = vacc;
      batch -= sizeof(float16_t);
    } while (batch != 0);
}
