src/f16-hswish/neonfp16arith.c.in

// 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 % 8 == 0
$assert BATCH_TILE >= 8
$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
#include <assert.h>

#include <arm_neon.h>

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


void xnn_f16_hswish_ukernel__neonfp16arith_x${BATCH_TILE}(
    size_t n,
    const void* restrict x_ptr,
    void* restrict y_ptr,
    const struct xnn_f16_hswish_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
{
  assert(n != 0);
  assert(n % sizeof(__fp16) == 0);

  const __fp16* x = (const __fp16*) x_ptr;
  __fp16* y = (__fp16*) y_ptr;

  const float16x8_t vsixth = vld1q_dup_f16(&params->sixth);
  const float16x8_t vthree = vld1q_dup_f16(&params->three);
  const int16x8_t vsix = vreinterpretq_s16_f16(vld1q_dup_f16(&params->six));
  const int16x8_t vzero = vdupq_n_s16(0);

  $if BATCH_TILE > 8:
    for (; n >= ${BATCH_TILE} * sizeof(__fp16); n -= ${BATCH_TILE} * sizeof(__fp16)) {
      $for N in range(0, BATCH_TILE, 8):
        float16x8_t vx${ABC[N:N+8]} = vld1q_f16(x); x += 8;

      $for N in range(0, BATCH_TILE, 8):
        float16x8_t vacc${ABC[N:N+8]} = vaddq_f16(vx${ABC[N:N+8]}, vthree);
        vx${ABC[N:N+8]} = vmulq_f16(vx${ABC[N:N+8]}, vsixth);

      $for N in range(0, BATCH_TILE, 8):
        vacc${ABC[N:N+8]} = vreinterpretq_f16_s16(vmaxq_s16(vreinterpretq_s16_f16(vacc${ABC[N:N+8]}), vzero));

      $for N in range(0, BATCH_TILE, 8):
        vacc${ABC[N:N+8]} = vreinterpretq_f16_s16(vminq_s16(vreinterpretq_s16_f16(vacc${ABC[N:N+8]}), vsix));

      $for N in range(0, BATCH_TILE, 8):
        vacc${ABC[N:N+8]} = vmulq_f16(vacc${ABC[N:N+8]}, vx${ABC[N:N+8]});

      $for N in range(0, BATCH_TILE, 8):
        vst1q_f16(y, vacc${ABC[N:N+8]}); y += 8;
    }
  for (; n >= 8 * sizeof(__fp16); n -= 8 * sizeof(__fp16)) {
    float16x8_t vx = vld1q_f16(x); x += 8;
    float16x8_t vacc = vaddq_f16(vx, vthree);
    vx = vmulq_f16(vx, vsixth);
    vacc = vreinterpretq_f16_s16(vmaxq_s16(vreinterpretq_s16_f16(vacc), vzero));
    vacc = vreinterpretq_f16_s16(vminq_s16(vreinterpretq_s16_f16(vacc), vsix));
    vacc = vmulq_f16(vacc, vx);
    vst1q_f16(y, vacc); y += 8;
  }
  if XNN_UNLIKELY(n != 0) {
    float16x8_t vx = vld1q_f16(x);
    float16x8_t vacc = vaddq_f16(vx, vthree);
    vx = vmulq_f16(vx, vsixth);
    vacc = vreinterpretq_f16_s16(vmaxq_s16(vreinterpretq_s16_f16(vacc), vzero));
    vacc = vreinterpretq_f16_s16(vminq_s16(vreinterpretq_s16_f16(vacc), vsix));
    vacc = vmulq_f16(vacc, vx);

    float16x4_t vacc_lo = vget_low_f16(vacc);
    if (n & (4 * sizeof(__fp16))) {
      vst1_f16(y, vacc_lo); y += 4;
      vacc_lo = vget_high_f16(vacc);
    }
    if (n & (2 * sizeof(__fp16))) {
      vst1_lane_u32(__builtin_assume_aligned(y, 1), vreinterpret_u32_f16(vacc_lo), 0); y += 2;
      vacc_lo = vext_f16(vacc_lo, vacc_lo, 2);
    }
    if (n & (1 * sizeof(__fp16))) {
      vst1_lane_f16(y, vacc_lo, 0);
    }
  }
}