// Copyright 2020 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 "xnnpack/common.h"
#include "xnnpack/vunary.h"


void xnn_f32_vsqrt_ukernel__scalar_sqrt_u${BATCH_TILE}(
    size_t batch,
    const float* input,
    float* output,
    const struct xnn_f32_default_params params[restrict XNN_MIN_ELEMENTS(1)])
{
  assert(batch != 0);
  assert(batch % sizeof(float) == 0);
  assert(input != NULL);
  assert(output != NULL);

  $if BATCH_TILE > 1:
    for (; batch >= ${BATCH_TILE} * sizeof(float); batch -= ${BATCH_TILE} * sizeof(float)) {
      $for N in range(BATCH_TILE):
        const float vx${ABC[N]} = input[${N}];
      input += ${BATCH_TILE};

      $for N in range(BATCH_TILE):
        const float vy${ABC[N]} = sqrtf(vx${ABC[N]});

      $for N in range(BATCH_TILE):
        output[${N}] = vy${ABC[N]};
      output += ${BATCH_TILE};
    }
    if XNN_UNLIKELY(batch != 0) {
      $if BATCH_TILE > 2:
        do {
          const float vx = *input++;
          const float vy = sqrtf(vx);
          *output++ = vy;
          batch -= sizeof(float);
        } while (batch != 0);
      $else:
        const float vx = *input;
        const float vy = sqrtf(vx);
        *output = vy;
    }
  $else:
    for (; batch >= sizeof(float); batch -= sizeof(float)) {
      const float vx = *input++;
      const float vy = sqrtf(vx);
      *output++ = vy;
    }
}