• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // Auto-generated file. Do not edit!
2 //   Template: src/f32-sigmoid/scalar-lut2048-p1-div.c.in
3 //   Generator: tools/xngen
4 //
5 // Copyright 2019 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 // Note redefine as uint32[] to avoid redundant bitcasts.
20 extern XNN_INTERNAL const uint32_t xnn_table_exp2minus_k_over_2048[2048];
21 
xnn_f32_sigmoid_ukernel__scalar_lut2048_p1_div_x1(size_t n,const float * x,float * y,const void * params)22 void xnn_f32_sigmoid_ukernel__scalar_lut2048_p1_div_x1(
23     size_t n,
24     const float* x,
25     float* y,
26     const void* params)
27 {
28   assert(n % sizeof(float) == 0);
29 
30   const float vmagic_bias = 0x1.800000p12f;
31   const float vminus_log2e = -0x1.715476p0f;
32   const uint32_t vindex_mask = UINT32_C(0x7FF);
33   const float vln2_hi = 0x1.600000p-1f;
34   const float vln2_lo = 0x1.7217F8p-8f;
35   const float vc1 = -0x1.FFFFFEp-1f;
36   const float vone = 1.0f;
37   const float vdenorm_cutoff = 0x1.5D589Ep+6f;
38 
39   do {
40     const float vx = *x++;
41 
42     const float vz = fabsf(vx);
43 
44     float vn = vz * vminus_log2e + vmagic_bias;
45     const uint32_t ve = fp32_to_bits(vn) << 12;
46     const uint32_t vidx = fp32_to_bits(vn) & vindex_mask;
47     const float vs = fp32_from_bits(xnn_table_exp2minus_k_over_2048[vidx] + ve);
48     vn -= vmagic_bias;
49 
50     float vt = vn * vln2_hi + vz;
51     vt = vn * vln2_lo + vt;
52 
53     const float vp = vt * vc1;
54     const float vy = vp * vs + vs;
55 
56     const float vd = vy + vone;
57     float vf = vy / vd;
58     if XNN_UNPREDICTABLE(vz > vdenorm_cutoff) {
59       vf = 0.0f;
60     }
61     if XNN_UNPREDICTABLE(vx > 0.0f) {
62       vf = vone - vf;
63     }
64 
65     *y++ = vf;
66 
67     n -= sizeof(float);
68   } while (n != 0);
69 }
70