// 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 % 4 == 0
$assert BATCH_TILE >= 4
$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
#include <assert.h>
#include <math.h>

#include <arm_neon.h>

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


void xnn_f32_vsqrt_ukernel__aarch64_neon_sqrt_u${BATCH_TILE}(
    size_t batch,
    const float* input,
    float* output,
    const struct xnn_f32_default_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
{
  assert(batch != 0);
  assert(batch % sizeof(float) == 0);
  assert(input != NULL);
  assert(output != NULL);

  $if BATCH_TILE > 4:
    for (; batch >= ${BATCH_TILE} * sizeof(float); batch -= ${BATCH_TILE} * sizeof(float)) {
      $for N in range(0, BATCH_TILE, 4):
        const float32x4_t vx${ABC[N:N+4]} = vld1q_f32(input); input += 4;

      $for N in range(0, BATCH_TILE, 4):
        const float32x4_t vy${ABC[N:N+4]} = vsqrtq_f32(vx${ABC[N:N+4]});

      $for N in range(0, BATCH_TILE, 4):
        vst1q_f32(output, vy${ABC[N:N+4]}); output += 4;
    }
  for (; batch >= 4 * sizeof(float); batch -= 4 * sizeof(float)) {
    const float32x4_t vx = vld1q_f32(input); input += 4;
    const float32x4_t vy = vsqrtq_f32(vx);
    vst1q_f32(output, vy); output += 4;
  }
  if XNN_UNLIKELY(batch != 0) {
    const float32x4_t vx = vld1q_f32(input);

    float32x2_t vy_lo = vsqrt_f32(vget_low_f32(vx));
    const float32x2_t vx_hi = vget_high_f32(vx);
    if (batch & (2 * sizeof(float))) {
      vst1_f32(output, vy_lo); output += 2;
      vy_lo = vsqrt_f32(vx_hi);
    }
    if (batch & (1 * sizeof(float))) {
      vst1_lane_f32(output, vy_lo, 0);
    }
  }
}