// 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 <arm_neon.h>

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


void xnn_f32_vclamp_ukernel__neon_x${BATCH_TILE}(
    size_t n,
    const float* x,
    float* y,
    const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
{
  assert(n != 0);
  assert(n % sizeof(float) == 0);
  assert(x != NULL);
  assert(y != NULL);

  const float32x4_t vy_min = vld1q_dup_f32(&params->scalar.min);
  const float32x4_t vy_max = vld1q_dup_f32(&params->scalar.max);

  for (; n >= ${BATCH_TILE} * sizeof(float); n -= ${BATCH_TILE} * sizeof(float)) {
    $for N in range(0, BATCH_TILE, 4):
      float32x4_t vacc${ABC[N:N+4]} = vld1q_f32(x); x += 4;

    $for N in range(0, BATCH_TILE, 4):
      vacc${ABC[N:N+4]} = vmaxq_f32(vacc${ABC[N:N+4]}, vy_min);

    $for N in range(0, BATCH_TILE, 4):
      vacc${ABC[N:N+4]} = vminq_f32(vacc${ABC[N:N+4]}, vy_max);

    $for N in range(0, BATCH_TILE, 4):
      vst1q_f32(y, vacc${ABC[N:N+4]}); y += 4;
  }
  $if BATCH_TILE > 4:
    for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
      float32x4_t vacc = vld1q_f32(x); x += 4;
      vacc = vmaxq_f32(vacc, vy_min);
      vacc = vminq_f32(vacc, vy_max);
      vst1q_f32(y, vacc); y += 4;
    }
  if XNN_UNLIKELY(n != 0) {
    float32x4_t vacc = vld1q_f32(x);
    vacc = vmaxq_f32(vacc, vy_min);
    vacc = vminq_f32(vacc, vy_max);

    float32x2_t vacc_lo = vget_low_f32(vacc);
    if (n & (2 * sizeof(float))) {
      vst1_f32(y, vacc_lo); y += 2;
      vacc_lo = vget_high_f32(vacc);
    }
    if (n & (1 * sizeof(float))) {
      vst1_lane_f32(y, vacc_lo, 0);
    }
  }
}