xref: /aosp_15_r20/external/XNNPACK/src/f32-vhswish/neon.c.in (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1// Copyright 2019 Google LLC
2//
3// This source code is licensed under the BSD-style license found in the
4// LICENSE file in the root directory of this source tree.
5
6$assert BATCH_TILE % 4 == 0
7$assert BATCH_TILE >= 4
8$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
9#include <assert.h>
10
11#include <arm_neon.h>
12
13#include <xnnpack/common.h>
14#include <xnnpack/vunary.h>
15
16
17void xnn_f32_vhswish_ukernel__neon_x${BATCH_TILE}(
18    size_t n,
19    const float* x,
20    float* y,
21    const union xnn_f32_hswish_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
22{
23  assert(n != 0);
24  assert(n % sizeof(float) == 0);
25
26  const float32x4_t vsixth = vld1q_dup_f32(&params->scalar.sixth);
27  const float32x4_t vthree = vld1q_dup_f32(&params->scalar.three);
28  const int32x4_t vsix = vreinterpretq_s32_f32(vld1q_dup_f32(&params->scalar.six));
29  const int32x4_t vzero = vdupq_n_s32(0);
30
31  $if BATCH_TILE > 4:
32    for (; n >= ${BATCH_TILE} * sizeof(float); n -= ${BATCH_TILE} * sizeof(float)) {
33      $for N in range(0, BATCH_TILE, 4):
34        float32x4_t vx${ABC[N:N+4]} = vld1q_f32(x); x += 4;
35
36      $for N in range(0, BATCH_TILE, 4):
37        float32x4_t vacc${ABC[N:N+4]} = vaddq_f32(vx${ABC[N:N+4]}, vthree);
38        vx${ABC[N:N+4]} = vmulq_f32(vx${ABC[N:N+4]}, vsixth);
39
40      $for N in range(0, BATCH_TILE, 4):
41        vacc${ABC[N:N+4]} = vreinterpretq_f32_s32(vmaxq_s32(vreinterpretq_s32_f32(vacc${ABC[N:N+4]}), vzero));
42
43      $for N in range(0, BATCH_TILE, 4):
44        vacc${ABC[N:N+4]} = vreinterpretq_f32_s32(vminq_s32(vreinterpretq_s32_f32(vacc${ABC[N:N+4]}), vsix));
45
46      $for N in range(0, BATCH_TILE, 4):
47        vacc${ABC[N:N+4]} = vmulq_f32(vacc${ABC[N:N+4]}, vx${ABC[N:N+4]});
48
49      $for N in range(0, BATCH_TILE, 4):
50        vst1q_f32(y, vacc${ABC[N:N+4]}); y += 4;
51    }
52  for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
53    float32x4_t vx = vld1q_f32(x); x += 4;
54    float32x4_t vacc = vaddq_f32(vx, vthree);
55    vx = vmulq_f32(vx, vsixth);
56    vacc = vreinterpretq_f32_s32(vmaxq_s32(vreinterpretq_s32_f32(vacc), vzero));
57    vacc = vreinterpretq_f32_s32(vminq_s32(vreinterpretq_s32_f32(vacc), vsix));
58    vacc = vmulq_f32(vacc, vx);
59    vst1q_f32(y, vacc); y += 4;
60  }
61  if XNN_UNLIKELY(n != 0) {
62    float32x4_t vx = vld1q_f32(x);
63    float32x4_t vacc = vaddq_f32(vx, vthree);
64    vx = vmulq_f32(vx, vsixth);
65    vacc = vreinterpretq_f32_s32(vmaxq_s32(vreinterpretq_s32_f32(vacc), vzero));
66    vacc = vreinterpretq_f32_s32(vminq_s32(vreinterpretq_s32_f32(vacc), vsix));
67    vacc = vmulq_f32(vacc, vx);
68
69    float32x2_t vacc_lo = vget_low_f32(vacc);
70    if (n & (2 * sizeof(float))) {
71      vst1_f32(y, vacc_lo); y += 2;
72      vacc_lo = vget_high_f32(vacc);
73    }
74    if (n & (1 * sizeof(float))) {
75      vst1_lane_f32(y, vacc_lo, 0);
76    }
77  }
78}
79