• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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_x8(size_t n,const float * x,float * y,const union xnn_f32_hswish_params params[restrict static1])18 void xnn_f32_hswish_ukernel__neonfma_x8(
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(&params->scalar.sixth);
28   const float32x4_t vhalf = vld1q_dup_f32(&params->scalar.half);
29   const float32x4_t vone = vld1q_dup_f32(&params->scalar.one);
30   const float32x4_t vzero = vdupq_n_f32(0.0f);
31 
32   for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
33     const float32x4_t vx0123 = vld1q_f32(x); x += 4;
34     const float32x4_t vx4567 = vld1q_f32(x); x += 4;
35 
36     float32x4_t vacc0123 = vfmaq_f32(vhalf, vx0123, vsixth);
37     float32x4_t vacc4567 = vfmaq_f32(vhalf, vx4567, vsixth);
38 
39     vacc0123 = vmaxq_f32(vacc0123, vzero);
40     vacc4567 = vmaxq_f32(vacc4567, vzero);
41 
42     vacc0123 = vminq_f32(vacc0123, vone);
43     vacc4567 = vminq_f32(vacc4567, vone);
44 
45     vacc0123 = vmulq_f32(vacc0123, vx0123);
46     vacc4567 = vmulq_f32(vacc4567, vx4567);
47 
48     vst1q_f32(y, vacc0123); y += 4;
49     vst1q_f32(y, vacc4567); y += 4;
50   }
51   for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
52     const float32x4_t vx0123 = vld1q_f32(x); x += 4;
53     float32x4_t vacc0123 = vfmaq_f32(vhalf, vx0123, vsixth);
54     vacc0123 = vmaxq_f32(vacc0123, vzero);
55     vacc0123 = vminq_f32(vacc0123, vone);
56     vacc0123 = vmulq_f32(vacc0123, vx0123);
57     vst1q_f32(y, vacc0123); y += 4;
58   }
59   if XNN_UNLIKELY(n != 0) {
60     const float32x4_t vx0123 = vld1q_f32(x);
61     float32x4_t vacc0123 = vfmaq_f32(vhalf, vx0123, vsixth);
62     vacc0123 = vmaxq_f32(vacc0123, vzero);
63     vacc0123 = vminq_f32(vacc0123, vone);
64     vacc0123 = vmulq_f32(vacc0123, vx0123);
65 
66     float32x2_t vacc01 = vget_low_f32(vacc0123);
67     if (n & (2 * sizeof(float))) {
68       vst1_f32(y, vacc01); y += 2;
69       vacc01 = vget_high_f32(vacc0123);
70     }
71     if (n & (1 * sizeof(float))) {
72       vst1_lane_f32(y, vacc01, 0);
73     }
74   }
75 }
76