1 // Auto-generated file. Do not edit!
2 // Template: src/f32-hswish/neon.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2019 Google LLC
6 //
7 // This source code is licensed under the BSD-style license found in the
8 // LICENSE file in the root directory of this source tree.
9
10 #include <assert.h>
11
12 #include <arm_neon.h>
13
14 #include <xnnpack/common.h>
15 #include <xnnpack/vbinary.h>
16
17
xnn_f32_hswish_ukernel__neonfma_x4(size_t n,const float * x,float * y,const union xnn_f32_hswish_params params[restrict static1])18 void xnn_f32_hswish_ukernel__neonfma_x4(
19 size_t n,
20 const float* x,
21 float* y,
22 const union xnn_f32_hswish_params params[restrict static 1])
23 {
24 assert(n != 0);
25 assert(n % sizeof(float) == 0);
26
27 const float32x4_t vsixth = vld1q_dup_f32(¶ms->scalar.sixth);
28 const float32x4_t vhalf = vld1q_dup_f32(¶ms->scalar.half);
29 const float32x4_t vone = vld1q_dup_f32(¶ms->scalar.one);
30 const float32x4_t vzero = vdupq_n_f32(0.0f);
31
32 for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
33 const float32x4_t vx0123 = vld1q_f32(x); x += 4;
34
35 float32x4_t vacc0123 = vfmaq_f32(vhalf, vx0123, vsixth);
36
37 vacc0123 = vmaxq_f32(vacc0123, vzero);
38
39 vacc0123 = vminq_f32(vacc0123, vone);
40
41 vacc0123 = vmulq_f32(vacc0123, vx0123);
42
43 vst1q_f32(y, vacc0123); y += 4;
44 }
45 for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
46 const float32x4_t vx0123 = vld1q_f32(x); x += 4;
47 float32x4_t vacc0123 = vfmaq_f32(vhalf, vx0123, vsixth);
48 vacc0123 = vmaxq_f32(vacc0123, vzero);
49 vacc0123 = vminq_f32(vacc0123, vone);
50 vacc0123 = vmulq_f32(vacc0123, vx0123);
51 vst1q_f32(y, vacc0123); y += 4;
52 }
53 if XNN_UNLIKELY(n != 0) {
54 const float32x4_t vx0123 = vld1q_f32(x);
55 float32x4_t vacc0123 = vfmaq_f32(vhalf, vx0123, vsixth);
56 vacc0123 = vmaxq_f32(vacc0123, vzero);
57 vacc0123 = vminq_f32(vacc0123, vone);
58 vacc0123 = vmulq_f32(vacc0123, vx0123);
59
60 float32x2_t vacc01 = vget_low_f32(vacc0123);
61 if (n & (2 * sizeof(float))) {
62 vst1_f32(y, vacc01); y += 2;
63 vacc01 = vget_high_f32(vacc0123);
64 }
65 if (n & (1 * sizeof(float))) {
66 vst1_lane_f32(y, vacc01, 0);
67 }
68 }
69 }
70