1 // Auto-generated file. Do not edit!
2 // Template: src/f32-velu/scalar-rr2-p6.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2020 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_velu_ukernel__wasm_rr2_p6_x3(size_t n,const float * x,float * y,const union xnn_f32_elu_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_f32_velu_ukernel__wasm_rr2_p6_x3(
19 size_t n,
20 const float* x,
21 float* y,
22 const union xnn_f32_elu_params params[restrict XNN_MIN_ELEMENTS(1)])
23 {
24 assert(n % sizeof(float) == 0);
25
26 const float vprescale = params->scalar_rr2_p6.prescale;
27 const float valpha = params->scalar_rr2_p6.alpha;
28 const float vbeta = params->scalar_rr2_p6.beta;
29 const float vmagic_bias = params->scalar_rr2_p6.magic_bias;
30 const float vlog2e = params->scalar_rr2_p6.log2e;
31 const float vsat_cutoff = params->scalar_rr2_p6.sat_cutoff;
32 const float vminus_ln2_hi = params->scalar_rr2_p6.minus_ln2_hi;
33 const float vminus_ln2_lo = params->scalar_rr2_p6.minus_ln2_lo;
34 const float vc6 = params->scalar_rr2_p6.c6;
35 const float vc5 = params->scalar_rr2_p6.c5;
36 const float vc4 = params->scalar_rr2_p6.c4;
37 const float vc3 = params->scalar_rr2_p6.c3;
38 const float vc2 = params->scalar_rr2_p6.c2;
39 const float vone = params->scalar_rr2_p6.one;
40
41 for (; n >= 3 * sizeof(float); n -= 3 * sizeof(float)) {
42 float vx0 = x[0];
43 float vx1 = x[1];
44 float vx2 = x[2];
45 x += 3;
46
47 const float vz0 = __builtin_wasm_min_f32(__builtin_wasm_max_f32(vx0 * vprescale, vsat_cutoff), 0.0f);
48 const float vz1 = __builtin_wasm_min_f32(__builtin_wasm_max_f32(vx1 * vprescale, vsat_cutoff), 0.0f);
49 const float vz2 = __builtin_wasm_min_f32(__builtin_wasm_max_f32(vx2 * vprescale, vsat_cutoff), 0.0f);
50
51 float vn0 = vz0 * vlog2e + vmagic_bias;
52 float vn1 = vz1 * vlog2e + vmagic_bias;
53 float vn2 = vz2 * vlog2e + vmagic_bias;
54
55 float vs0 = uint32_as_float(float_as_uint32(vn0) << 23);
56 vn0 -= vmagic_bias;
57 float vs1 = uint32_as_float(float_as_uint32(vn1) << 23);
58 vn1 -= vmagic_bias;
59 float vs2 = uint32_as_float(float_as_uint32(vn2) << 23);
60 vn2 -= vmagic_bias;
61
62 float vt0 = vn0 * vminus_ln2_hi + vz0;
63 float vt1 = vn1 * vminus_ln2_hi + vz1;
64 float vt2 = vn2 * vminus_ln2_hi + vz2;
65
66 vt0 = vn0 * vminus_ln2_lo + vt0;
67 vt1 = vn1 * vminus_ln2_lo + vt1;
68 vt2 = vn2 * vminus_ln2_lo + vt2;
69
70
71 float vp0 = vc6 * vt0 + vc5;
72 float vp1 = vc6 * vt1 + vc5;
73 float vp2 = vc6 * vt2 + vc5;
74
75 vp0 = vp0 * vt0 + vc4;
76 vp1 = vp1 * vt1 + vc4;
77 vp2 = vp2 * vt2 + vc4;
78
79 vp0 = vp0 * vt0 + vc3;
80 vp1 = vp1 * vt1 + vc3;
81 vp2 = vp2 * vt2 + vc3;
82
83 vp0 = vp0 * vt0 + vc2;
84 vp1 = vp1 * vt1 + vc2;
85 vp2 = vp2 * vt2 + vc2;
86
87 vp0 *= vt0;
88 vp1 *= vt1;
89 vp2 *= vt2;
90
91 vt0 *= vs0;
92 vs0 -= vone;
93 vt1 *= vs1;
94 vs1 -= vone;
95 vt2 *= vs2;
96 vs2 -= vone;
97
98 vp0 = vp0 * vt0 + vt0;
99 vp1 = vp1 * vt1 + vt1;
100 vp2 = vp2 * vt2 + vt2;
101
102 const float ve0 = (vp0 + vs0) * valpha;
103 float vy0 = __builtin_wasm_max_f32(vx0 * vbeta, 0.0f);
104 const float ve1 = (vp1 + vs1) * valpha;
105 float vy1 = __builtin_wasm_max_f32(vx1 * vbeta, 0.0f);
106 const float ve2 = (vp2 + vs2) * valpha;
107 float vy2 = __builtin_wasm_max_f32(vx2 * vbeta, 0.0f);
108
109 vy0 += __builtin_wasm_min_f32(ve0, 0.0f);
110 vy1 += __builtin_wasm_min_f32(ve1, 0.0f);
111 vy2 += __builtin_wasm_min_f32(ve2, 0.0f);
112
113 y[0] = vy0;
114 y[1] = vy1;
115 y[2] = vy2;
116 y += 3;
117 }
118 if XNN_UNLIKELY(n != 0) {
119 do {
120 float vx = *x++;
121
122 const float vz = __builtin_wasm_min_f32(__builtin_wasm_max_f32(vx * vprescale, vsat_cutoff), 0.0f);
123
124 float vn = vz * vlog2e + vmagic_bias;
125 float vs = uint32_as_float(float_as_uint32(vn) << 23);
126 vn -= vmagic_bias;
127
128 float vt = vn * vminus_ln2_hi + vz;
129 vt = vn * vminus_ln2_lo + vt;
130
131
132 float vp = vc6 * vt + vc5;
133 vp = vp * vt + vc4;
134 vp = vp * vt + vc3;
135 vp = vp * vt + vc2;
136 vp *= vt;
137
138 vt *= vs;
139 vs -= vone;
140 vp = vp * vt + vt;
141 const float ve = (vp + vs) * valpha;
142
143 float vy = __builtin_wasm_max_f32(vx * vbeta, 0.0f);
144 vy += __builtin_wasm_min_f32(ve, 0.0f);
145
146 *y++ = vy;
147
148 n -= sizeof(float);
149 } while (n != 0);
150 }
151 }
152