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_x1(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_x1(
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 do {
42 float vx = *x++;
43
44 const float vz = __builtin_wasm_min_f32(__builtin_wasm_max_f32(vx * vprescale, vsat_cutoff), 0.0f);
45
46 float vn = vz * vlog2e + vmagic_bias;
47 const uint32_t ven = float_as_uint32(vn) << 19;
48 const uint32_t vidx = float_as_uint32(vn) & vindex_mask;
49 vn -= vmagic_bias;
50
51 float vt = vn * vminus_ln2_hi + vz;
52 float vs = uint32_as_float(xnn_table_exp2minus_k_over_16[vidx] + ven);
53
54 vt = vn * vminus_ln2_lo + vt;
55
56 float vp = vc3 * vt + vc2;
57 vp *= vt;
58
59 vt *= vs;
60 vs -= vone;
61 vp = vp * vt + vt;
62 const float ve = (vp + vs) * valpha;
63
64 float vy = __builtin_wasm_max_f32(vx * vbeta, 0.0f);
65 vy += __builtin_wasm_min_f32(ve, 0.0f);
66
67 *y++ = vy;
68
69 n -= sizeof(float);
70 } while (n != 0);
71 }
72