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/vunary.h>
15
16 #include <fp16/bitcasts.h>
17
18
xnn_f32_velu_ukernel__wasm_rr2_p6_x6(size_t n,const float * x,float * y,const union xnn_f32_elu_params params[restrict XNN_MIN_ELEMENTS (1)])19 void xnn_f32_velu_ukernel__wasm_rr2_p6_x6(
20 size_t n,
21 const float* x,
22 float* y,
23 const union xnn_f32_elu_params params[restrict XNN_MIN_ELEMENTS(1)])
24 {
25 assert(n % sizeof(float) == 0);
26
27 const float vprescale = params->scalar.prescale;
28 const float valpha = params->scalar.alpha;
29 const float vbeta = params->scalar.beta;
30
31 const float vmagic_bias = 0x1.8000FEp23f;
32 const float vlog2e = 0x1.715476p+0f;
33 const float vsat_cutoff = -0x1.154246p+4f;
34 const float vminus_ln2_hi = -0x1.62E440p-1f;
35 const float vminus_ln2_lo = 0x1.0105C6p-21f;
36 const float vc6 = 0x1.6b7338p-10f;
37 const float vc5 = 0x1.12278Ep-7f;
38 const float vc4 = 0x1.555716p-5f;
39 const float vc3 = 0x1.5554B0p-3f;
40 const float vc2 = 0x1.FFFFFEp-2f;
41 const float vone = 1.0f;
42
43 for (; n >= 6 * sizeof(float); n -= 6 * sizeof(float)) {
44 float vx0 = x[0];
45 float vx1 = x[1];
46 float vx2 = x[2];
47 float vx3 = x[3];
48 float vx4 = x[4];
49 float vx5 = x[5];
50 x += 6;
51
52 const float vz0 = __builtin_wasm_min_f32(__builtin_wasm_max_f32(vx0 * vprescale, vsat_cutoff), 0.0f);
53 const float vz1 = __builtin_wasm_min_f32(__builtin_wasm_max_f32(vx1 * vprescale, vsat_cutoff), 0.0f);
54 const float vz2 = __builtin_wasm_min_f32(__builtin_wasm_max_f32(vx2 * vprescale, vsat_cutoff), 0.0f);
55 const float vz3 = __builtin_wasm_min_f32(__builtin_wasm_max_f32(vx3 * vprescale, vsat_cutoff), 0.0f);
56 const float vz4 = __builtin_wasm_min_f32(__builtin_wasm_max_f32(vx4 * vprescale, vsat_cutoff), 0.0f);
57 const float vz5 = __builtin_wasm_min_f32(__builtin_wasm_max_f32(vx5 * vprescale, vsat_cutoff), 0.0f);
58
59 float vn0 = vz0 * vlog2e + vmagic_bias;
60 float vn1 = vz1 * vlog2e + vmagic_bias;
61 float vn2 = vz2 * vlog2e + vmagic_bias;
62 float vn3 = vz3 * vlog2e + vmagic_bias;
63 float vn4 = vz4 * vlog2e + vmagic_bias;
64 float vn5 = vz5 * vlog2e + vmagic_bias;
65
66 float vs0 = fp32_from_bits(fp32_to_bits(vn0) << 23);
67 vn0 -= vmagic_bias;
68 float vs1 = fp32_from_bits(fp32_to_bits(vn1) << 23);
69 vn1 -= vmagic_bias;
70 float vs2 = fp32_from_bits(fp32_to_bits(vn2) << 23);
71 vn2 -= vmagic_bias;
72 float vs3 = fp32_from_bits(fp32_to_bits(vn3) << 23);
73 vn3 -= vmagic_bias;
74 float vs4 = fp32_from_bits(fp32_to_bits(vn4) << 23);
75 vn4 -= vmagic_bias;
76 float vs5 = fp32_from_bits(fp32_to_bits(vn5) << 23);
77 vn5 -= vmagic_bias;
78
79 float vt0 = vn0 * vminus_ln2_hi + vz0;
80 float vt1 = vn1 * vminus_ln2_hi + vz1;
81 float vt2 = vn2 * vminus_ln2_hi + vz2;
82 float vt3 = vn3 * vminus_ln2_hi + vz3;
83 float vt4 = vn4 * vminus_ln2_hi + vz4;
84 float vt5 = vn5 * vminus_ln2_hi + vz5;
85
86 vt0 = vn0 * vminus_ln2_lo + vt0;
87 vt1 = vn1 * vminus_ln2_lo + vt1;
88 vt2 = vn2 * vminus_ln2_lo + vt2;
89 vt3 = vn3 * vminus_ln2_lo + vt3;
90 vt4 = vn4 * vminus_ln2_lo + vt4;
91 vt5 = vn5 * vminus_ln2_lo + vt5;
92
93
94 float vp0 = vc6 * vt0 + vc5;
95 float vp1 = vc6 * vt1 + vc5;
96 float vp2 = vc6 * vt2 + vc5;
97 float vp3 = vc6 * vt3 + vc5;
98 float vp4 = vc6 * vt4 + vc5;
99 float vp5 = vc6 * vt5 + vc5;
100
101 vp0 = vp0 * vt0 + vc4;
102 vp1 = vp1 * vt1 + vc4;
103 vp2 = vp2 * vt2 + vc4;
104 vp3 = vp3 * vt3 + vc4;
105 vp4 = vp4 * vt4 + vc4;
106 vp5 = vp5 * vt5 + vc4;
107
108 vp0 = vp0 * vt0 + vc3;
109 vp1 = vp1 * vt1 + vc3;
110 vp2 = vp2 * vt2 + vc3;
111 vp3 = vp3 * vt3 + vc3;
112 vp4 = vp4 * vt4 + vc3;
113 vp5 = vp5 * vt5 + vc3;
114
115 vp0 = vp0 * vt0 + vc2;
116 vp1 = vp1 * vt1 + vc2;
117 vp2 = vp2 * vt2 + vc2;
118 vp3 = vp3 * vt3 + vc2;
119 vp4 = vp4 * vt4 + vc2;
120 vp5 = vp5 * vt5 + vc2;
121
122 vp0 *= vt0;
123 vp1 *= vt1;
124 vp2 *= vt2;
125 vp3 *= vt3;
126 vp4 *= vt4;
127 vp5 *= vt5;
128
129 vt0 *= vs0;
130 vs0 -= vone;
131 vt1 *= vs1;
132 vs1 -= vone;
133 vt2 *= vs2;
134 vs2 -= vone;
135 vt3 *= vs3;
136 vs3 -= vone;
137 vt4 *= vs4;
138 vs4 -= vone;
139 vt5 *= vs5;
140 vs5 -= vone;
141
142 vp0 = vp0 * vt0 + vt0;
143 vp1 = vp1 * vt1 + vt1;
144 vp2 = vp2 * vt2 + vt2;
145 vp3 = vp3 * vt3 + vt3;
146 vp4 = vp4 * vt4 + vt4;
147 vp5 = vp5 * vt5 + vt5;
148
149 const float ve0 = (vp0 + vs0) * valpha;
150 float vy0 = __builtin_wasm_max_f32(vx0 * vbeta, 0.0f);
151 const float ve1 = (vp1 + vs1) * valpha;
152 float vy1 = __builtin_wasm_max_f32(vx1 * vbeta, 0.0f);
153 const float ve2 = (vp2 + vs2) * valpha;
154 float vy2 = __builtin_wasm_max_f32(vx2 * vbeta, 0.0f);
155 const float ve3 = (vp3 + vs3) * valpha;
156 float vy3 = __builtin_wasm_max_f32(vx3 * vbeta, 0.0f);
157 const float ve4 = (vp4 + vs4) * valpha;
158 float vy4 = __builtin_wasm_max_f32(vx4 * vbeta, 0.0f);
159 const float ve5 = (vp5 + vs5) * valpha;
160 float vy5 = __builtin_wasm_max_f32(vx5 * vbeta, 0.0f);
161
162 vy0 += __builtin_wasm_min_f32(ve0, 0.0f);
163 vy1 += __builtin_wasm_min_f32(ve1, 0.0f);
164 vy2 += __builtin_wasm_min_f32(ve2, 0.0f);
165 vy3 += __builtin_wasm_min_f32(ve3, 0.0f);
166 vy4 += __builtin_wasm_min_f32(ve4, 0.0f);
167 vy5 += __builtin_wasm_min_f32(ve5, 0.0f);
168
169 y[0] = vy0;
170 y[1] = vy1;
171 y[2] = vy2;
172 y[3] = vy3;
173 y[4] = vy4;
174 y[5] = vy5;
175 y += 6;
176 }
177 if XNN_UNLIKELY(n != 0) {
178 do {
179 float vx = *x++;
180
181 const float vz = __builtin_wasm_min_f32(__builtin_wasm_max_f32(vx * vprescale, vsat_cutoff), 0.0f);
182
183 float vn = vz * vlog2e + vmagic_bias;
184 float vs = fp32_from_bits(fp32_to_bits(vn) << 23);
185 vn -= vmagic_bias;
186
187 float vt = vn * vminus_ln2_hi + vz;
188 vt = vn * vminus_ln2_lo + vt;
189
190
191 float vp = vc6 * vt + vc5;
192 vp = vp * vt + vc4;
193 vp = vp * vt + vc3;
194 vp = vp * vt + vc2;
195 vp *= vt;
196
197 vt *= vs;
198 vs -= vone;
199 vp = vp * vt + vt;
200 const float ve = (vp + vs) * valpha;
201
202 float vy = __builtin_wasm_max_f32(vx * vbeta, 0.0f);
203 vy += __builtin_wasm_min_f32(ve, 0.0f);
204
205 *y++ = vy;
206
207 n -= sizeof(float);
208 } while (n != 0);
209 }
210 }
211