xref: /aosp_15_r20/external/XNNPACK/src/f32-vlrelu/scalar.c.in (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)

// 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__scalar_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;

  $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};

      $for N in range(BATCH_TILE):
        float vacc${ABC[N]} = vx${ABC[N]} * vslope;

      $for N in range(BATCH_TILE):
        vacc${ABC[N]} = XNN_UNPREDICTABLE(vx${ABC[N]} < 0.0f) ? vacc${ABC[N]} : vx${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++;
          float vacc = vx * vslope;
          vacc = XNN_UNPREDICTABLE(vx < 0.0f) ? vacc : vx;
          *y++ = vacc;
          n -= sizeof(float);
        } while (n != 0);
      $else:
        const float vx = *x;
        float vacc = vx * vslope;
        vacc = XNN_UNPREDICTABLE(vx < 0.0f) ? vacc : vx;
        *y = vacc;
    }
  $else:
    do {
      const float vx = *x++;
      float vacc = vx * vslope;
      vacc = XNN_UNPREDICTABLE(vx < 0.0f) ? vacc : vx;
      *y++ = vacc;
      n -= sizeof(float);
    } while (n != 0);
}