• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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_x4(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_x4(
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 >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
44     float vx0 = x[0];
45     float vx1 = x[1];
46     float vx2 = x[2];
47     float vx3 = x[3];
48     x += 4;
49 
50     const float vz0 = __builtin_wasm_min_f32(__builtin_wasm_max_f32(vx0 * vprescale, vsat_cutoff), 0.0f);
51     const float vz1 = __builtin_wasm_min_f32(__builtin_wasm_max_f32(vx1 * vprescale, vsat_cutoff), 0.0f);
52     const float vz2 = __builtin_wasm_min_f32(__builtin_wasm_max_f32(vx2 * vprescale, vsat_cutoff), 0.0f);
53     const float vz3 = __builtin_wasm_min_f32(__builtin_wasm_max_f32(vx3 * vprescale, vsat_cutoff), 0.0f);
54 
55     float vn0 = vz0 * vlog2e + vmagic_bias;
56     float vn1 = vz1 * vlog2e + vmagic_bias;
57     float vn2 = vz2 * vlog2e + vmagic_bias;
58     float vn3 = vz3 * vlog2e + vmagic_bias;
59 
60     float vs0 = fp32_from_bits(fp32_to_bits(vn0) << 23);
61     vn0 -= vmagic_bias;
62     float vs1 = fp32_from_bits(fp32_to_bits(vn1) << 23);
63     vn1 -= vmagic_bias;
64     float vs2 = fp32_from_bits(fp32_to_bits(vn2) << 23);
65     vn2 -= vmagic_bias;
66     float vs3 = fp32_from_bits(fp32_to_bits(vn3) << 23);
67     vn3 -= vmagic_bias;
68 
69     float vt0 = vn0 * vminus_ln2_hi + vz0;
70     float vt1 = vn1 * vminus_ln2_hi + vz1;
71     float vt2 = vn2 * vminus_ln2_hi + vz2;
72     float vt3 = vn3 * vminus_ln2_hi + vz3;
73 
74     vt0 = vn0 * vminus_ln2_lo + vt0;
75     vt1 = vn1 * vminus_ln2_lo + vt1;
76     vt2 = vn2 * vminus_ln2_lo + vt2;
77     vt3 = vn3 * vminus_ln2_lo + vt3;
78 
79 
80     float vp0 = vc6 * vt0 + vc5;
81     float vp1 = vc6 * vt1 + vc5;
82     float vp2 = vc6 * vt2 + vc5;
83     float vp3 = vc6 * vt3 + vc5;
84 
85     vp0 = vp0 * vt0 + vc4;
86     vp1 = vp1 * vt1 + vc4;
87     vp2 = vp2 * vt2 + vc4;
88     vp3 = vp3 * vt3 + vc4;
89 
90     vp0 = vp0 * vt0 + vc3;
91     vp1 = vp1 * vt1 + vc3;
92     vp2 = vp2 * vt2 + vc3;
93     vp3 = vp3 * vt3 + vc3;
94 
95     vp0 = vp0 * vt0 + vc2;
96     vp1 = vp1 * vt1 + vc2;
97     vp2 = vp2 * vt2 + vc2;
98     vp3 = vp3 * vt3 + vc2;
99 
100     vp0 *= vt0;
101     vp1 *= vt1;
102     vp2 *= vt2;
103     vp3 *= vt3;
104 
105     vt0 *= vs0;
106     vs0 -= vone;
107     vt1 *= vs1;
108     vs1 -= vone;
109     vt2 *= vs2;
110     vs2 -= vone;
111     vt3 *= vs3;
112     vs3 -= vone;
113 
114     vp0 = vp0 * vt0 + vt0;
115     vp1 = vp1 * vt1 + vt1;
116     vp2 = vp2 * vt2 + vt2;
117     vp3 = vp3 * vt3 + vt3;
118 
119     const float ve0 = (vp0 + vs0) * valpha;
120     float vy0 = __builtin_wasm_max_f32(vx0 * vbeta, 0.0f);
121     const float ve1 = (vp1 + vs1) * valpha;
122     float vy1 = __builtin_wasm_max_f32(vx1 * vbeta, 0.0f);
123     const float ve2 = (vp2 + vs2) * valpha;
124     float vy2 = __builtin_wasm_max_f32(vx2 * vbeta, 0.0f);
125     const float ve3 = (vp3 + vs3) * valpha;
126     float vy3 = __builtin_wasm_max_f32(vx3 * vbeta, 0.0f);
127 
128     vy0 += __builtin_wasm_min_f32(ve0, 0.0f);
129     vy1 += __builtin_wasm_min_f32(ve1, 0.0f);
130     vy2 += __builtin_wasm_min_f32(ve2, 0.0f);
131     vy3 += __builtin_wasm_min_f32(ve3, 0.0f);
132 
133     y[0] = vy0;
134     y[1] = vy1;
135     y[2] = vy2;
136     y[3] = vy3;
137     y += 4;
138   }
139   if XNN_UNLIKELY(n != 0) {
140     do {
141       float vx = *x++;
142 
143       const float vz = __builtin_wasm_min_f32(__builtin_wasm_max_f32(vx * vprescale, vsat_cutoff), 0.0f);
144 
145       float vn = vz * vlog2e + vmagic_bias;
146       float vs = fp32_from_bits(fp32_to_bits(vn) << 23);
147       vn -= vmagic_bias;
148 
149       float vt = vn * vminus_ln2_hi + vz;
150       vt = vn * vminus_ln2_lo + vt;
151 
152 
153       float vp = vc6 * vt + vc5;
154       vp = vp * vt + vc4;
155       vp = vp * vt + vc3;
156       vp = vp * vt + vc2;
157       vp *= vt;
158 
159       vt *= vs;
160       vs -= vone;
161       vp = vp * vt + vt;
162       const float ve = (vp + vs) * valpha;
163 
164       float vy = __builtin_wasm_max_f32(vx * vbeta, 0.0f);
165       vy += __builtin_wasm_min_f32(ve, 0.0f);
166 
167       *y++ = vy;
168 
169       n -= sizeof(float);
170     } while (n != 0);
171   }
172 }
173