// 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_vlrelu_ukernel__neon_u${BATCH_TILE}(
    size_t batch,
    const float* input,
    float* output,
    const struct xnn_f32_lrelu_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
{
  assert(batch != 0);
  assert(batch % sizeof(float) == 0);
  assert(input != NULL);
  assert(output != NULL);

  const float32x4_t vslope = vld1q_dup_f32(&params->scalar.slope);

  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;

    $for N in range(0, BATCH_TILE, 4):
      float32x4_t vacc${ABC[N:N+4]} = vmulq_f32(vx${ABC[N:N+4]}, vslope);
      const uint32x4_t vmask${ABC[N:N+4]} = vcltq_s32(vreinterpretq_s32_f32(vx${ABC[N:N+4]}), vmovq_n_s32(0));

    $for N in range(0, BATCH_TILE, 4):
      vacc${ABC[N:N+4]} = vbslq_f32(vmask${ABC[N:N+4]}, vacc${ABC[N:N+4]}, vx${ABC[N:N+4]});

    $for N in range(0, BATCH_TILE, 4):
      vst1q_f32(output, vacc${ABC[N:N+4]}); output += 4;
  }
  $if BATCH_TILE > 4:
    for (; batch >= 4 * sizeof(float); batch -= 4 * sizeof(float)) {
      const float32x4_t vx = vld1q_f32(input); input += 4;
      float32x4_t vacc = vmulq_f32(vx, vslope);
      const uint32x4_t vmask = vcltq_s32(vreinterpretq_s32_f32(vx), vmovq_n_s32(0));
      vacc = vbslq_f32(vmask, vacc, vx);
      vst1q_f32(output, vacc); output += 4;
    }
  if XNN_UNLIKELY(batch != 0) {
    const float32x4_t vx = vld1q_f32(input);
    float32x4_t vacc = vmulq_f32(vx, vslope);
    const uint32x4_t vmask = vcltq_s32(vreinterpretq_s32_f32(vx), vmovq_n_s32(0));
    vacc = vbslq_f32(vmask, vacc, vx);

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