xref: /aosp_15_r20/external/XNNPACK/src/f32-vhswish/gen/vhswish-scalar-x4.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1*4bdc9457SAndroid Build Coastguard Worker // Auto-generated file. Do not edit!
2*4bdc9457SAndroid Build Coastguard Worker //   Template: src/f32-vhswish/scalar.c.in
3*4bdc9457SAndroid Build Coastguard Worker //   Generator: tools/xngen
4*4bdc9457SAndroid Build Coastguard Worker //
5*4bdc9457SAndroid Build Coastguard Worker // Copyright 2019 Google LLC
6*4bdc9457SAndroid Build Coastguard Worker //
7*4bdc9457SAndroid Build Coastguard Worker // This source code is licensed under the BSD-style license found in the
8*4bdc9457SAndroid Build Coastguard Worker // LICENSE file in the root directory of this source tree.
9*4bdc9457SAndroid Build Coastguard Worker 
10*4bdc9457SAndroid Build Coastguard Worker #include <assert.h>
11*4bdc9457SAndroid Build Coastguard Worker 
12*4bdc9457SAndroid Build Coastguard Worker #include <xnnpack/common.h>
13*4bdc9457SAndroid Build Coastguard Worker #include <xnnpack/math.h>
14*4bdc9457SAndroid Build Coastguard Worker #include <xnnpack/vunary.h>
15*4bdc9457SAndroid Build Coastguard Worker 
16*4bdc9457SAndroid Build Coastguard Worker 
xnn_f32_vhswish_ukernel__scalar_x4(size_t n,const float * x,float * y,const union xnn_f32_hswish_params params[restrict XNN_MIN_ELEMENTS (1)])17*4bdc9457SAndroid Build Coastguard Worker void xnn_f32_vhswish_ukernel__scalar_x4(
18*4bdc9457SAndroid Build Coastguard Worker     size_t n,
19*4bdc9457SAndroid Build Coastguard Worker     const float* x,
20*4bdc9457SAndroid Build Coastguard Worker     float* y,
21*4bdc9457SAndroid Build Coastguard Worker     const union xnn_f32_hswish_params params[restrict XNN_MIN_ELEMENTS(1)])
22*4bdc9457SAndroid Build Coastguard Worker {
23*4bdc9457SAndroid Build Coastguard Worker   assert(n != 0);
24*4bdc9457SAndroid Build Coastguard Worker   assert(n % sizeof(float) == 0);
25*4bdc9457SAndroid Build Coastguard Worker 
26*4bdc9457SAndroid Build Coastguard Worker   const float vsixth = params->scalar.sixth;
27*4bdc9457SAndroid Build Coastguard Worker   const float vthree = params->scalar.three;
28*4bdc9457SAndroid Build Coastguard Worker   const float vsix = params->scalar.six;
29*4bdc9457SAndroid Build Coastguard Worker   const float vzero = 0.0f;
30*4bdc9457SAndroid Build Coastguard Worker   assert(vthree == 3.0f);
31*4bdc9457SAndroid Build Coastguard Worker   assert(vsix == 6.0f);
32*4bdc9457SAndroid Build Coastguard Worker 
33*4bdc9457SAndroid Build Coastguard Worker   for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
34*4bdc9457SAndroid Build Coastguard Worker     float vx0 = x[0];
35*4bdc9457SAndroid Build Coastguard Worker     float vx1 = x[1];
36*4bdc9457SAndroid Build Coastguard Worker     float vx2 = x[2];
37*4bdc9457SAndroid Build Coastguard Worker     float vx3 = x[3];
38*4bdc9457SAndroid Build Coastguard Worker     x += 4;
39*4bdc9457SAndroid Build Coastguard Worker 
40*4bdc9457SAndroid Build Coastguard Worker     float vacc0 = vx0 + vthree;
41*4bdc9457SAndroid Build Coastguard Worker     vx0 *= vsixth;
42*4bdc9457SAndroid Build Coastguard Worker     float vacc1 = vx1 + vthree;
43*4bdc9457SAndroid Build Coastguard Worker     vx1 *= vsixth;
44*4bdc9457SAndroid Build Coastguard Worker     float vacc2 = vx2 + vthree;
45*4bdc9457SAndroid Build Coastguard Worker     vx2 *= vsixth;
46*4bdc9457SAndroid Build Coastguard Worker     float vacc3 = vx3 + vthree;
47*4bdc9457SAndroid Build Coastguard Worker     vx3 *= vsixth;
48*4bdc9457SAndroid Build Coastguard Worker 
49*4bdc9457SAndroid Build Coastguard Worker     vacc0 = math_max_f32(vacc0, vzero);
50*4bdc9457SAndroid Build Coastguard Worker     vacc1 = math_max_f32(vacc1, vzero);
51*4bdc9457SAndroid Build Coastguard Worker     vacc2 = math_max_f32(vacc2, vzero);
52*4bdc9457SAndroid Build Coastguard Worker     vacc3 = math_max_f32(vacc3, vzero);
53*4bdc9457SAndroid Build Coastguard Worker 
54*4bdc9457SAndroid Build Coastguard Worker     vacc0 = math_min_f32(vacc0, vsix);
55*4bdc9457SAndroid Build Coastguard Worker     vacc1 = math_min_f32(vacc1, vsix);
56*4bdc9457SAndroid Build Coastguard Worker     vacc2 = math_min_f32(vacc2, vsix);
57*4bdc9457SAndroid Build Coastguard Worker     vacc3 = math_min_f32(vacc3, vsix);
58*4bdc9457SAndroid Build Coastguard Worker 
59*4bdc9457SAndroid Build Coastguard Worker     vacc0 *= vx0;
60*4bdc9457SAndroid Build Coastguard Worker     vacc1 *= vx1;
61*4bdc9457SAndroid Build Coastguard Worker     vacc2 *= vx2;
62*4bdc9457SAndroid Build Coastguard Worker     vacc3 *= vx3;
63*4bdc9457SAndroid Build Coastguard Worker 
64*4bdc9457SAndroid Build Coastguard Worker     y[0] = vacc0;
65*4bdc9457SAndroid Build Coastguard Worker     y[1] = vacc1;
66*4bdc9457SAndroid Build Coastguard Worker     y[2] = vacc2;
67*4bdc9457SAndroid Build Coastguard Worker     y[3] = vacc3;
68*4bdc9457SAndroid Build Coastguard Worker     y += 4;
69*4bdc9457SAndroid Build Coastguard Worker   }
70*4bdc9457SAndroid Build Coastguard Worker   if XNN_UNLIKELY(n != 0) {
71*4bdc9457SAndroid Build Coastguard Worker     do {
72*4bdc9457SAndroid Build Coastguard Worker       float vx = *x++;
73*4bdc9457SAndroid Build Coastguard Worker       float vacc = vx + vthree;
74*4bdc9457SAndroid Build Coastguard Worker       vx *= vsixth;
75*4bdc9457SAndroid Build Coastguard Worker       vacc = math_max_f32(vacc, vzero);
76*4bdc9457SAndroid Build Coastguard Worker       vacc = math_min_f32(vacc, vsix);
77*4bdc9457SAndroid Build Coastguard Worker       vacc *= vx;
78*4bdc9457SAndroid Build Coastguard Worker       *y++ = vacc;
79*4bdc9457SAndroid Build Coastguard Worker       n -= sizeof(float);
80*4bdc9457SAndroid Build Coastguard Worker     } while (n != 0);
81*4bdc9457SAndroid Build Coastguard Worker   }
82*4bdc9457SAndroid Build Coastguard Worker }
83