1 // Auto-generated file. Do not edit!
2 // Template: src/f32-velu/scalar-rr2-lut16-p3.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
18 extern XNN_INTERNAL const uint32_t xnn_table_exp2minus_k_over_16[16];
19
xnn_f32_velu_ukernel__wasm_rr2_lut16_p3_x5(size_t n,const float * x,float * y,const union xnn_f32_elu_params params[restrict XNN_MIN_ELEMENTS (1)])20 void xnn_f32_velu_ukernel__wasm_rr2_lut16_p3_x5(
21 size_t n,
22 const float* x,
23 float* y,
24 const union xnn_f32_elu_params params[restrict XNN_MIN_ELEMENTS(1)])
25 {
26 assert(n % sizeof(float) == 0);
27
28 const float vprescale = params->scalar_rr2_lut16_p3.prescale;
29 const float valpha = params->scalar_rr2_lut16_p3.alpha;
30 const float vbeta = params->scalar_rr2_lut16_p3.beta;
31 const float vmagic_bias = params->scalar_rr2_lut16_p3.magic_bias;
32 const float vlog2e = params->scalar_rr2_lut16_p3.log2e;
33 const uint32_t vindex_mask = UINT32_C(0xF);
34 const float vsat_cutoff = params->scalar_rr2_lut16_p3.sat_cutoff;
35 const float vminus_ln2_hi = params->scalar_rr2_lut16_p3.minus_ln2_hi;
36 const float vminus_ln2_lo = params->scalar_rr2_lut16_p3.minus_ln2_lo;
37 const float vc3 = params->scalar_rr2_lut16_p3.c3;
38 const float vc2 = params->scalar_rr2_lut16_p3.c2;
39 const float vone = params->scalar_rr2_lut16_p3.one;
40
41 for (; n >= 5 * sizeof(float); n -= 5 * sizeof(float)) {
42 float vx0 = x[0];
43 float vx1 = x[1];
44 float vx2 = x[2];
45 float vx3 = x[3];
46 float vx4 = x[4];
47 x += 5;
48
49 const float vz0 = __builtin_wasm_min_f32(__builtin_wasm_max_f32(vx0 * vprescale, vsat_cutoff), 0.0f);
50 const float vz1 = __builtin_wasm_min_f32(__builtin_wasm_max_f32(vx1 * vprescale, vsat_cutoff), 0.0f);
51 const float vz2 = __builtin_wasm_min_f32(__builtin_wasm_max_f32(vx2 * vprescale, vsat_cutoff), 0.0f);
52 const float vz3 = __builtin_wasm_min_f32(__builtin_wasm_max_f32(vx3 * vprescale, vsat_cutoff), 0.0f);
53 const float vz4 = __builtin_wasm_min_f32(__builtin_wasm_max_f32(vx4 * 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 float vn4 = vz4 * vlog2e + vmagic_bias;
60
61 const uint32_t ven0 = float_as_uint32(vn0) << 19;
62 const uint32_t vidx0 = float_as_uint32(vn0) & vindex_mask;
63 vn0 -= vmagic_bias;
64 const uint32_t ven1 = float_as_uint32(vn1) << 19;
65 const uint32_t vidx1 = float_as_uint32(vn1) & vindex_mask;
66 vn1 -= vmagic_bias;
67 const uint32_t ven2 = float_as_uint32(vn2) << 19;
68 const uint32_t vidx2 = float_as_uint32(vn2) & vindex_mask;
69 vn2 -= vmagic_bias;
70 const uint32_t ven3 = float_as_uint32(vn3) << 19;
71 const uint32_t vidx3 = float_as_uint32(vn3) & vindex_mask;
72 vn3 -= vmagic_bias;
73 const uint32_t ven4 = float_as_uint32(vn4) << 19;
74 const uint32_t vidx4 = float_as_uint32(vn4) & vindex_mask;
75 vn4 -= vmagic_bias;
76
77 float vt0 = vn0 * vminus_ln2_hi + vz0;
78 float vs0 = uint32_as_float(xnn_table_exp2minus_k_over_16[vidx0] + ven0);
79 float vt1 = vn1 * vminus_ln2_hi + vz1;
80 float vs1 = uint32_as_float(xnn_table_exp2minus_k_over_16[vidx1] + ven1);
81 float vt2 = vn2 * vminus_ln2_hi + vz2;
82 float vs2 = uint32_as_float(xnn_table_exp2minus_k_over_16[vidx2] + ven2);
83 float vt3 = vn3 * vminus_ln2_hi + vz3;
84 float vs3 = uint32_as_float(xnn_table_exp2minus_k_over_16[vidx3] + ven3);
85 float vt4 = vn4 * vminus_ln2_hi + vz4;
86 float vs4 = uint32_as_float(xnn_table_exp2minus_k_over_16[vidx4] + ven4);
87
88 vt0 = vn0 * vminus_ln2_lo + vt0;
89 vt1 = vn1 * vminus_ln2_lo + vt1;
90 vt2 = vn2 * vminus_ln2_lo + vt2;
91 vt3 = vn3 * vminus_ln2_lo + vt3;
92 vt4 = vn4 * vminus_ln2_lo + vt4;
93
94 float vp0 = vc3 * vt0 + vc2;
95 float vp1 = vc3 * vt1 + vc2;
96 float vp2 = vc3 * vt2 + vc2;
97 float vp3 = vc3 * vt3 + vc2;
98 float vp4 = vc3 * vt4 + vc2;
99
100 vp0 *= vt0;
101 vp1 *= vt1;
102 vp2 *= vt2;
103 vp3 *= vt3;
104 vp4 *= vt4;
105
106 vt0 *= vs0;
107 vs0 -= vone;
108 vt1 *= vs1;
109 vs1 -= vone;
110 vt2 *= vs2;
111 vs2 -= vone;
112 vt3 *= vs3;
113 vs3 -= vone;
114 vt4 *= vs4;
115 vs4 -= vone;
116
117 vp0 = vp0 * vt0 + vt0;
118 vp1 = vp1 * vt1 + vt1;
119 vp2 = vp2 * vt2 + vt2;
120 vp3 = vp3 * vt3 + vt3;
121 vp4 = vp4 * vt4 + vt4;
122
123 const float ve0 = (vp0 + vs0) * valpha;
124 float vy0 = __builtin_wasm_max_f32(vx0 * vbeta, 0.0f);
125 const float ve1 = (vp1 + vs1) * valpha;
126 float vy1 = __builtin_wasm_max_f32(vx1 * vbeta, 0.0f);
127 const float ve2 = (vp2 + vs2) * valpha;
128 float vy2 = __builtin_wasm_max_f32(vx2 * vbeta, 0.0f);
129 const float ve3 = (vp3 + vs3) * valpha;
130 float vy3 = __builtin_wasm_max_f32(vx3 * vbeta, 0.0f);
131 const float ve4 = (vp4 + vs4) * valpha;
132 float vy4 = __builtin_wasm_max_f32(vx4 * vbeta, 0.0f);
133
134 vy0 += __builtin_wasm_min_f32(ve0, 0.0f);
135 vy1 += __builtin_wasm_min_f32(ve1, 0.0f);
136 vy2 += __builtin_wasm_min_f32(ve2, 0.0f);
137 vy3 += __builtin_wasm_min_f32(ve3, 0.0f);
138 vy4 += __builtin_wasm_min_f32(ve4, 0.0f);
139
140 y[0] = vy0;
141 y[1] = vy1;
142 y[2] = vy2;
143 y[3] = vy3;
144 y[4] = vy4;
145 y += 5;
146 }
147 if XNN_UNLIKELY(n != 0) {
148 do {
149 float vx = *x++;
150
151 const float vz = __builtin_wasm_min_f32(__builtin_wasm_max_f32(vx * vprescale, vsat_cutoff), 0.0f);
152
153 float vn = vz * vlog2e + vmagic_bias;
154 const uint32_t ven = float_as_uint32(vn) << 19;
155 const uint32_t vidx = float_as_uint32(vn) & vindex_mask;
156 vn -= vmagic_bias;
157
158 float vt = vn * vminus_ln2_hi + vz;
159 float vs = uint32_as_float(xnn_table_exp2minus_k_over_16[vidx] + ven);
160
161 vt = vn * vminus_ln2_lo + vt;
162
163 float vp = vc3 * vt + vc2;
164 vp *= vt;
165
166 vt *= vs;
167 vs -= vone;
168 vp = vp * vt + vt;
169 const float ve = (vp + vs) * valpha;
170
171 float vy = __builtin_wasm_max_f32(vx * vbeta, 0.0f);
172 vy += __builtin_wasm_min_f32(ve, 0.0f);
173
174 *y++ = vy;
175
176 n -= sizeof(float);
177 } while (n != 0);
178 }
179 }
180