xref: /aosp_15_r20/external/XNNPACK/src/f32-vsigmoid/gen/vsigmoid-scalar-rr2-p5-div-x2.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1 // Auto-generated file. Do not edit!
2 //   Template: src/f32-vsigmoid/scalar-rr2-p5-div.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 #include <math.h>
12 
13 #include <xnnpack/common.h>
14 #include <xnnpack/math.h>
15 #include <xnnpack/vunary.h>
16 
17 
xnn_f32_vsigmoid_ukernel__scalar_rr2_p5_div_x2(size_t n,const float * x,float * y,const union xnn_f32_sigmoid_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_f32_vsigmoid_ukernel__scalar_rr2_p5_div_x2(
19     size_t n,
20     const float* x,
21     float* y,
22     const union xnn_f32_sigmoid_params params[restrict XNN_MIN_ELEMENTS(1)])
23 {
24   assert(n % sizeof(float) == 0);
25 
26   const float vmagic_bias = params->scalar_rr2_p5.magic_bias;
27   const float vminus_log2e = params->scalar_rr2_p5.minus_log2e;
28   const float vln2_hi = params->scalar_rr2_p5.ln2_hi;
29   const float vln2_lo = params->scalar_rr2_p5.ln2_lo;
30   const float vc5 = params->scalar_rr2_p5.c5;
31   const float vc4 = params->scalar_rr2_p5.c4;
32   const float vc3 = params->scalar_rr2_p5.c3;
33   const float vc2 = params->scalar_rr2_p5.c2;
34   const float vc1 = params->scalar_rr2_p5.c1;
35   const float vone = params->scalar_rr2_p5.one;
36   const float vdenorm_cutoff = params->scalar_rr2_p5.denorm_cutoff;
37 
38   for (; n >= 2 * sizeof(float); n -= 2 * sizeof(float)) {
39     const float vx0 = x[0];
40     const float vx1 = x[1];
41     x += 2;
42 
43     const float vz0 = fabsf(vx0);
44     const float vz1 = fabsf(vx1);
45 
46     float vn0 = vz0 * vminus_log2e + vmagic_bias;
47     float vn1 = vz1 * vminus_log2e + vmagic_bias;
48 
49     const float vs0 = uint32_as_float(float_as_uint32(vn0) << 23);
50     const float vs1 = uint32_as_float(float_as_uint32(vn1) << 23);
51 
52     vn0 -= vmagic_bias;
53     vn1 -= vmagic_bias;
54 
55     float vt0 = vn0 * vln2_hi + vz0;
56     float vt1 = vn1 * vln2_hi + vz1;
57 
58     vt0 = vn0 * vln2_lo + vt0;
59     vt1 = vn1 * vln2_lo + vt1;
60 
61     float vp0 = vt0 * vc5 + vc4;
62     float vp1 = vt1 * vc5 + vc4;
63 
64     vp0 = vt0 * vp0 + vc3;
65     vp1 = vt1 * vp1 + vc3;
66 
67     vp0 = vt0 * vp0 + vc2;
68     vp1 = vt1 * vp1 + vc2;
69 
70     vp0 = vt0 * vp0 + vc1;
71     vp1 = vt1 * vp1 + vc1;
72 
73     vt0 *= vs0;
74     vt1 *= vs1;
75 
76     const float ve0 = vt0 * vp0 + vs0;
77     const float ve1 = vt1 * vp1 + vs1;
78 
79     const float vd0 = ve0 + vone;
80     const float vd1 = ve1 + vone;
81 
82     float vf0 = ve0 / vd0;
83     float vf1 = ve1 / vd1;
84 
85     if XNN_UNPREDICTABLE(vz0 > vdenorm_cutoff) {
86       vf0 = 0.0f;
87     }
88     if XNN_UNPREDICTABLE(vz1 > vdenorm_cutoff) {
89       vf1 = 0.0f;
90     }
91 
92     if XNN_UNPREDICTABLE(vx0 > 0.0f) {
93       vf0 = vone - vf0;
94     }
95     if XNN_UNPREDICTABLE(vx1 > 0.0f) {
96       vf1 = vone - vf1;
97     }
98 
99     y[0] = vf0;
100     y[1] = vf1;
101     y += 2;
102   }
103   if XNN_UNLIKELY(n != 0) {
104     const float vx = *x;
105 
106     const float vz = fabsf(vx);
107 
108     float vn = vz * vminus_log2e + vmagic_bias;
109     const float vs = uint32_as_float(float_as_uint32(vn) << 23);
110     vn -= vmagic_bias;
111 
112     float vt = vn * vln2_hi + vz;
113     vt = vn * vln2_lo + vt;
114 
115     float vp = vt * vc5 + vc4;
116     vp = vt * vp + vc3;
117     vp = vt * vp + vc2;
118     vp = vt * vp + vc1;
119 
120     vt *= vs;
121     const float ve = vt * vp + vs;
122     const float vd = ve + vone;
123 
124     float vf = ve / vd;
125     if XNN_UNPREDICTABLE(vz > vdenorm_cutoff) {
126       vf = 0.0f;
127     }
128     if XNN_UNPREDICTABLE(vx > 0.0f) {
129       vf = vone - vf;
130     }
131 
132     *y = vf;
133   }
134 }
135