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/vunary.h>
15
16 #include <fp16/bitcasts.h>
17
18
19 extern XNN_INTERNAL const uint32_t xnn_table_exp2minus_k_over_16[16];
20
xnn_f32_velu_ukernel__wasm_rr2_lut16_p3_x1(size_t n,const float * x,float * y,const union xnn_f32_elu_params params[restrict XNN_MIN_ELEMENTS (1)])21 void xnn_f32_velu_ukernel__wasm_rr2_lut16_p3_x1(
22 size_t n,
23 const float* x,
24 float* y,
25 const union xnn_f32_elu_params params[restrict XNN_MIN_ELEMENTS(1)])
26 {
27 assert(n % sizeof(float) == 0);
28
29 const float vprescale = params->scalar.prescale;
30 const float valpha = params->scalar.alpha;
31 const float vbeta = params->scalar.beta;
32
33 const float vmagic_bias = 0x1.800000p19f;
34 const float vlog2e = 0x1.715476p+0f;
35 const uint32_t vindex_mask = UINT32_C(0xF);
36 const float vsat_cutoff = -0x1.154246p+4f;
37 const float vminus_ln2_hi = -0x1.62E400p-1f;
38 const float vminus_ln2_lo = -0x1.7F7D1Cp-20f;
39 const float vc3 = 0x1.55561Cp-3f;
40 const float vc2 = 0x1.0001ECp-1f;
41 const float vone = 1.0f;
42
43 do {
44 float vx = *x++;
45
46 const float vz = __builtin_wasm_min_f32(__builtin_wasm_max_f32(vx * vprescale, vsat_cutoff), 0.0f);
47
48 float vn = vz * vlog2e + vmagic_bias;
49 const uint32_t ven = fp32_to_bits(vn) << 19;
50 const uint32_t vidx = fp32_to_bits(vn) & vindex_mask;
51 vn -= vmagic_bias;
52
53 float vt = vn * vminus_ln2_hi + vz;
54 float vs = fp32_from_bits(xnn_table_exp2minus_k_over_16[vidx] + ven);
55
56 vt = vn * vminus_ln2_lo + vt;
57
58 float vp = vc3 * vt + vc2;
59 vp *= vt;
60
61 vt *= vs;
62 vs -= vone;
63 vp = vp * vt + vt;
64 const float ve = (vp + vs) * valpha;
65
66 float vy = __builtin_wasm_max_f32(vx * vbeta, 0.0f);
67 vy += __builtin_wasm_min_f32(ve, 0.0f);
68
69 *y++ = vy;
70
71 n -= sizeof(float);
72 } while (n != 0);
73 }
74