• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // Auto-generated file. Do not edit!
2 //   Template: src/f32-hswish/scalar.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 <xnnpack/common.h>
13 #include <xnnpack/math.h>
14 #include <xnnpack/hswish.h>
15 
16 
xnn_f32_hswish_ukernel__wasm_x4(size_t n,const float * x,float * y,const union xnn_f32_hswish_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_f32_hswish_ukernel__wasm_x4(
18     size_t n,
19     const float* x,
20     float* y,
21     const union xnn_f32_hswish_params params[restrict XNN_MIN_ELEMENTS(1)])
22 {
23   assert(n != 0);
24   assert(n % sizeof(float) == 0);
25 
26   const float vsixth = params->scalar.sixth;
27   const float vthree = params->scalar.three;
28   const float vsix = params->scalar.six;
29   const float vzero = 0.0f;
30   assert(vthree == 3.0f);
31   assert(vsix == 6.0f);
32 
33   for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
34     float vx0 = x[0];
35     float vx1 = x[1];
36     float vx2 = x[2];
37     float vx3 = x[3];
38     x += 4;
39 
40     float vacc0 = vx0 + vthree;
41     vx0 *= vsixth;
42     float vacc1 = vx1 + vthree;
43     vx1 *= vsixth;
44     float vacc2 = vx2 + vthree;
45     vx2 *= vsixth;
46     float vacc3 = vx3 + vthree;
47     vx3 *= vsixth;
48 
49     vacc0 = __builtin_wasm_max_f32(vacc0, vzero);
50     vacc1 = __builtin_wasm_max_f32(vacc1, vzero);
51     vacc2 = __builtin_wasm_max_f32(vacc2, vzero);
52     vacc3 = __builtin_wasm_max_f32(vacc3, vzero);
53 
54     vacc0 = __builtin_wasm_min_f32(vacc0, vsix);
55     vacc1 = __builtin_wasm_min_f32(vacc1, vsix);
56     vacc2 = __builtin_wasm_min_f32(vacc2, vsix);
57     vacc3 = __builtin_wasm_min_f32(vacc3, vsix);
58 
59     vacc0 *= vx0;
60     vacc1 *= vx1;
61     vacc2 *= vx2;
62     vacc3 *= vx3;
63 
64     y[0] = vacc0;
65     y[1] = vacc1;
66     y[2] = vacc2;
67     y[3] = vacc3;
68     y += 4;
69   }
70   if XNN_UNLIKELY(n != 0) {
71     do {
72       float vx = *x++;
73       float vacc = vx + vthree;
74       vx *= vsixth;
75       vacc = __builtin_wasm_max_f32(vacc, vzero);
76       vacc = __builtin_wasm_min_f32(vacc, vsix);
77       vacc *= vx;
78       *y++ = vacc;
79       n -= sizeof(float);
80     } while (n != 0);
81   }
82 }
83