1 // Auto-generated file. Do not edit!
2 // Template: src/f32-vsigmoid/neon-p5.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
12 #include <arm_neon.h>
13
14 #include <xnnpack/common.h>
15 #include <xnnpack/vunary.h>
16
17
xnn_f32_vsigmoid_ukernel__neon_rr2_p5_nr2recps_x4(size_t n,const float * x,float * y,const union xnn_f32_sigmoid_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_f32_vsigmoid_ukernel__neon_rr2_p5_nr2recps_x4(
19 size_t n,
20 const float* x,
21 float* y,
22 const union xnn_f32_sigmoid_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
23 {
24 assert(n % sizeof(float) == 0);
25
26 const float32x4_t vmagic_bias = vld1q_dup_f32(¶ms->neon_rr2_p5.magic_bias);
27 const float32x4_t vminus_log2e = vld1q_dup_f32(¶ms->neon_rr2_p5.minus_log2e);
28 const float32x4_t vln2_hi = vld1q_dup_f32(¶ms->neon_rr2_p5.ln2_hi);
29 const float32x4_t vln2_lo = vld1q_dup_f32(¶ms->neon_rr2_p5.ln2_lo);
30 const float32x4_t vc5 = vld1q_dup_f32(¶ms->neon_rr2_p5.c5);
31 const float32x4_t vc4 = vld1q_dup_f32(¶ms->neon_rr2_p5.c4);
32 const float32x4_t vc3 = vld1q_dup_f32(¶ms->neon_rr2_p5.c3);
33 const float32x4_t vc2 = vld1q_dup_f32(¶ms->neon_rr2_p5.c2);
34 const float32x4_t vc1 = vld1q_dup_f32(¶ms->neon_rr2_p5.c1);
35 const float32x4_t vone = vmovq_n_f32(1.0f);
36 const float32x4_t vdenorm_cutoff = vld1q_dup_f32(¶ms->neon_rr2_p5.denorm_cutoff);
37
38 for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
39 const float32x4_t vx = vld1q_f32(x); x += 4;
40
41 const float32x4_t vz = vabsq_f32(vx);
42
43 float32x4_t vn = vmlaq_f32(vmagic_bias, vz, vminus_log2e);
44 const float32x4_t vs = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn), 23));
45 vn = vsubq_f32(vn, vmagic_bias);
46 float32x4_t vt = vmlaq_f32(vz, vn, vln2_hi);
47 vt = vmlaq_f32(vt, vn, vln2_lo);
48
49 float32x4_t vp = vmlaq_f32(vc4, vc5, vt);
50 vp = vmlaq_f32(vc3, vp, vt);
51 vp = vmlaq_f32(vc2, vp, vt);
52 vp = vmlaq_f32(vc1, vp, vt);
53
54 vt = vmulq_f32(vt, vs);
55 const float32x4_t ve = vmlaq_f32(vs, vp, vt);
56 const float32x4_t vd = vaddq_f32(ve, vone);
57
58 float32x4_t vr = vrecpeq_f32(vd);
59 vr = vmulq_f32(vr, vrecpsq_f32(vr, vd));
60 vr = vmulq_f32(vr, vrecpsq_f32(vr, vd));
61
62 float32x4_t vf = vmulq_f32(ve, vr);
63 vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcagtq_f32(vx, vdenorm_cutoff)));
64 const uint32x4_t vm = vcltq_f32(vx, vmovq_n_f32(0.0f));
65 vf = vbslq_f32(vm, vf, vsubq_f32(vone, vf));
66
67 vst1q_f32(y, vf); y += 4;
68 }
69 if XNN_UNLIKELY(n != 0) {
70 const float32x4_t vx = vld1q_f32(x);
71
72 const float32x4_t vz = vabsq_f32(vx);
73
74 float32x4_t vn = vmlaq_f32(vmagic_bias, vz, vminus_log2e);
75 const float32x4_t vs = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn), 23));
76 vn = vsubq_f32(vn, vmagic_bias);
77 float32x4_t vt = vmlaq_f32(vz, vn, vln2_hi);
78 vt = vmlaq_f32(vt, vn, vln2_lo);
79
80 float32x4_t vp = vmlaq_f32(vc4, vc5, vt);
81 vp = vmlaq_f32(vc3, vp, vt);
82 vp = vmlaq_f32(vc2, vp, vt);
83 vp = vmlaq_f32(vc1, vp, vt);
84
85 vt = vmulq_f32(vt, vs);
86 const float32x4_t ve = vmlaq_f32(vs, vp, vt);
87 const float32x4_t vd = vaddq_f32(ve, vone);
88
89 float32x4_t vr = vrecpeq_f32(vd);
90 vr = vmulq_f32(vr, vrecpsq_f32(vr, vd));
91 vr = vmulq_f32(vr, vrecpsq_f32(vr, vd));
92
93 float32x4_t vf = vmulq_f32(ve, vr);
94 vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcagtq_f32(vx, vdenorm_cutoff)));
95 const uint32x4_t vm = vcltq_f32(vx, vmovq_n_f32(0.0f));
96 vf = vbslq_f32(vm, vf, vsubq_f32(vone, vf));
97
98 float32x2_t vf_lo = vget_low_f32(vf);
99 if (n & (2 * sizeof(float))) {
100 vst1_f32(y, vf_lo); y += 2;
101 vf_lo = vget_high_f32(vf);
102 }
103 if (n & (1 * sizeof(float))) {
104 vst1_lane_f32(y, vf_lo, 0);
105 }
106 }
107 }
108