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_x2(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_x2(
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 >= 2 * sizeof(float); n -= 2 * sizeof(float)) {
42 float vx0 = x[0];
43 float vx1 = x[1];
44 x += 2;
45
46 const float vz0 = __builtin_wasm_min_f32(__builtin_wasm_max_f32(vx0 * vprescale, vsat_cutoff), 0.0f);
47 const float vz1 = __builtin_wasm_min_f32(__builtin_wasm_max_f32(vx1 * vprescale, vsat_cutoff), 0.0f);
48
49 float vn0 = vz0 * vlog2e + vmagic_bias;
50 float vn1 = vz1 * vlog2e + vmagic_bias;
51
52 float vs0 = uint32_as_float(float_as_uint32(vn0) << 23);
53 vn0 -= vmagic_bias;
54 float vs1 = uint32_as_float(float_as_uint32(vn1) << 23);
55 vn1 -= vmagic_bias;
56
57 float vt0 = vn0 * vminus_ln2_hi + vz0;
58 float vt1 = vn1 * vminus_ln2_hi + vz1;
59
60 vt0 = vn0 * vminus_ln2_lo + vt0;
61 vt1 = vn1 * vminus_ln2_lo + vt1;
62
63
64 float vp0 = vc6 * vt0 + vc5;
65 float vp1 = vc6 * vt1 + vc5;
66
67 vp0 = vp0 * vt0 + vc4;
68 vp1 = vp1 * vt1 + vc4;
69
70 vp0 = vp0 * vt0 + vc3;
71 vp1 = vp1 * vt1 + vc3;
72
73 vp0 = vp0 * vt0 + vc2;
74 vp1 = vp1 * vt1 + vc2;
75
76 vp0 *= vt0;
77 vp1 *= vt1;
78
79 vt0 *= vs0;
80 vs0 -= vone;
81 vt1 *= vs1;
82 vs1 -= vone;
83
84 vp0 = vp0 * vt0 + vt0;
85 vp1 = vp1 * vt1 + vt1;
86
87 const float ve0 = (vp0 + vs0) * valpha;
88 float vy0 = __builtin_wasm_max_f32(vx0 * vbeta, 0.0f);
89 const float ve1 = (vp1 + vs1) * valpha;
90 float vy1 = __builtin_wasm_max_f32(vx1 * vbeta, 0.0f);
91
92 vy0 += __builtin_wasm_min_f32(ve0, 0.0f);
93 vy1 += __builtin_wasm_min_f32(ve1, 0.0f);
94
95 y[0] = vy0;
96 y[1] = vy1;
97 y += 2;
98 }
99 if XNN_UNLIKELY(n != 0) {
100 float vx = *x;
101
102 const float vz = __builtin_wasm_min_f32(__builtin_wasm_max_f32(vx * vprescale, vsat_cutoff), 0.0f);
103
104 float vn = vz * vlog2e + vmagic_bias;
105 float vs = uint32_as_float(float_as_uint32(vn) << 23);
106 vn -= vmagic_bias;
107
108 float vt = vn * vminus_ln2_hi + vz;
109 vt = vn * vminus_ln2_lo + vt;
110
111
112 float vp = vc6 * vt + vc5;
113 vp = vp * vt + vc4;
114 vp = vp * vt + vc3;
115 vp = vp * vt + vc2;
116 vp *= vt;
117
118 vt *= vs;
119 vs -= vone;
120 vp = vp * vt + vt;
121 const float ve = (vp + vs) * valpha;
122
123 float vy = __builtin_wasm_max_f32(vx * vbeta, 0.0f);
124 vy += __builtin_wasm_min_f32(ve, 0.0f);
125
126 *y = vy;
127 }
128 }
129