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


void xnn_f32_vlrelu_ukernel__wasm_x${BATCH_TILE}(
    size_t n,
    const float* x,
    float* y,
    const union xnn_f32_lrelu_params params[restrict XNN_MIN_ELEMENTS(1)])
{
  assert(n != 0);
  assert(n % sizeof(float) == 0);

  const float vslope = params->scalar.slope;
  const float vzero = 0.0f;

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

      $for N in range(BATCH_TILE):
        const float vnegx${ABC[N]} = __builtin_wasm_min_f32(vx${ABC[N]}, vzero);

      $for N in range(BATCH_TILE):
        float vacc${ABC[N]} = vnegx${ABC[N]} * vslope;
        const float vposx${ABC[N]} = __builtin_wasm_max_f32(vx${ABC[N]}, vzero);

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

      $for N in range(BATCH_TILE):
        y[${N}] = vacc${ABC[N]};
      y += ${BATCH_TILE};
    }
    if XNN_UNLIKELY(n != 0) {
      $if BATCH_TILE > 2:
        do {
          const float vx = *x++;
          const float vnegx = __builtin_wasm_min_f32(vx, vzero);
          float vacc = vnegx * vslope;
          const float vposx = __builtin_wasm_max_f32(vx, vzero);
          vacc += vposx;
          *y++ = vacc;
          n -= sizeof(float);
        } while (n != 0);
      $else:
        const float vx = *x;
        const float vnegx = __builtin_wasm_min_f32(vx, vzero);
        float vacc = vnegx * vslope;
        const float vposx = __builtin_wasm_max_f32(vx, vzero);
        vacc += vposx;
        *y = vacc;
    }
  $else:
    do {
      const float vx = *x++;
      const float vnegx = __builtin_wasm_min_f32(vx, vzero);
      float vacc = vnegx * vslope;
      const float vposx = __builtin_wasm_max_f32(vx, vzero);
      vacc += vposx;
      *y++ = vacc;
      n -= sizeof(float);
    } while (n != 0);
}