src/f32-hswish/neon.c.in

// Copyright 2019 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/hswish.h>


void xnn_f32_hswish_ukernel__neon_x${BATCH_TILE}(
    size_t n,
    const float* x,
    float* y,
    const union xnn_f32_hswish_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
{
  assert(n != 0);
  assert(n % sizeof(float) == 0);

  const float32x4_t vsixth = vld1q_dup_f32(&params->scalar.sixth);
  const float32x4_t vthree = vld1q_dup_f32(&params->scalar.three);
  const int32x4_t vsix = vreinterpretq_s32_f32(vld1q_dup_f32(&params->scalar.six));
  const int32x4_t vzero = vdupq_n_s32(0);

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

      $for N in range(0, BATCH_TILE, 4):
        float32x4_t vacc${ABC[N:N+4]} = vaddq_f32(vx${ABC[N:N+4]}, vthree);
        vx${ABC[N:N+4]} = vmulq_f32(vx${ABC[N:N+4]}, vsixth);

      $for N in range(0, BATCH_TILE, 4):
        vacc${ABC[N:N+4]} = vreinterpretq_f32_s32(vmaxq_s32(vreinterpretq_s32_f32(vacc${ABC[N:N+4]}), vzero));

      $for N in range(0, BATCH_TILE, 4):
        vacc${ABC[N:N+4]} = vreinterpretq_f32_s32(vminq_s32(vreinterpretq_s32_f32(vacc${ABC[N:N+4]}), vsix));

      $for N in range(0, BATCH_TILE, 4):
        vacc${ABC[N:N+4]} = vmulq_f32(vacc${ABC[N:N+4]}, vx${ABC[N:N+4]});

      $for N in range(0, BATCH_TILE, 4):
        vst1q_f32(y, vacc${ABC[N:N+4]}); y += 4;
    }
  for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
    float32x4_t vx = vld1q_f32(x); x += 4;
    float32x4_t vacc = vaddq_f32(vx, vthree);
    vx = vmulq_f32(vx, vsixth);
    vacc = vreinterpretq_f32_s32(vmaxq_s32(vreinterpretq_s32_f32(vacc), vzero));
    vacc = vreinterpretq_f32_s32(vminq_s32(vreinterpretq_s32_f32(vacc), vsix));
    vacc = vmulq_f32(vacc, vx);
    vst1q_f32(y, vacc); y += 4;
  }
  if XNN_UNLIKELY(n != 0) {
    float32x4_t vx = vld1q_f32(x);
    float32x4_t vacc = vaddq_f32(vx, vthree);
    vx = vmulq_f32(vx, vsixth);
    vacc = vreinterpretq_f32_s32(vmaxq_s32(vreinterpretq_s32_f32(vacc), vzero));
    vacc = vreinterpretq_f32_s32(vminq_s32(vreinterpretq_s32_f32(vacc), vsix));
    vacc = vmulq_f32(vacc, vx);

    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);
    }
  }
}