1 // Auto-generated file. Do not edit!
2 // Template: src/f32-sigmoid/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_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x8(size_t n,const float * x,float * y,const void * params)18 void xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x8(
19 size_t n,
20 const float* x,
21 float* y,
22 const void* params) XNN_DISABLE_TSAN
23 {
24 assert(n % sizeof(float) == 0);
25
26 const float32x4_t vmagic_bias = vmovq_n_f32(0x1.8000FEp23f);
27 const float32x4_t vminus_log2e = vmovq_n_f32(-0x1.715476p+0f);
28 const float32x4_t vln2 = vmovq_n_f32(0x1.62E43p-1f);
29 const float32x4_t vc5 = vmovq_n_f32(-0x1.0F9F9Cp-7f);
30 const float32x4_t vc4 = vmovq_n_f32(0x1.573A1Ap-5f);
31 const float32x4_t vc3 = vmovq_n_f32(-0x1.555A80p-3f);
32 const float32x4_t vc2 = vmovq_n_f32(0x1.FFFDC6p-2f);
33 const float32x4_t vc1 = vmovq_n_f32(-0x1.FFFFF6p-1f);
34 const float32x4_t vone = vmovq_n_f32(1.0f);
35 const float32x4_t vdenorm_cutoff = vmovq_n_f32(-0x1.5D589Ep+6f);
36
37 for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
38 const float32x4_t vx0123 = vld1q_f32(x); x += 4;
39 const float32x4_t vx4567 = vld1q_f32(x); x += 4;
40
41 const float32x4_t vz0123 = vabsq_f32(vx0123);
42 const float32x4_t vz4567 = vabsq_f32(vx4567);
43
44 float32x4_t vn0123 = vfmaq_f32(vmagic_bias, vz0123, vminus_log2e);
45 float32x4_t vn4567 = vfmaq_f32(vmagic_bias, vz4567, vminus_log2e);
46
47 const float32x4_t vs0123 = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn0123), 23));
48 const float32x4_t vs4567 = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn4567), 23));
49
50 vn0123 = vsubq_f32(vn0123, vmagic_bias);
51 vn4567 = vsubq_f32(vn4567, vmagic_bias);
52
53 float32x4_t vt0123 = vfmaq_f32(vz0123, vn0123, vln2);
54 float32x4_t vt4567 = vfmaq_f32(vz4567, vn4567, vln2);
55
56 float32x4_t vp0123 = vfmaq_f32(vc4, vc5, vt0123);
57 float32x4_t vp4567 = vfmaq_f32(vc4, vc5, vt4567);
58
59 vp0123 = vfmaq_f32(vc3, vp0123, vt0123);
60 vp4567 = vfmaq_f32(vc3, vp4567, vt4567);
61
62 vp0123 = vfmaq_f32(vc2, vp0123, vt0123);
63 vp4567 = vfmaq_f32(vc2, vp4567, vt4567);
64
65 vp0123 = vfmaq_f32(vc1, vp0123, vt0123);
66 vp4567 = vfmaq_f32(vc1, vp4567, vt4567);
67
68 vt0123 = vmulq_f32(vt0123, vs0123);
69 vt4567 = vmulq_f32(vt4567, vs4567);
70
71 const float32x4_t ve0123 = vfmaq_f32(vs0123, vp0123, vt0123);
72 const float32x4_t ve4567 = vfmaq_f32(vs4567, vp4567, vt4567);
73
74 const float32x4_t vd0123 = vaddq_f32(ve0123, vone);
75 const float32x4_t vd4567 = vaddq_f32(ve4567, vone);
76
77 float32x4_t vr0123 = vrecpeq_f32(vd0123);
78 float32x4_t vr4567 = vrecpeq_f32(vd4567);
79
80 vr0123 = vmulq_f32(vr0123, vrecpsq_f32(vr0123, vd0123));
81 vr4567 = vmulq_f32(vr4567, vrecpsq_f32(vr4567, vd4567));
82
83 vr0123 = vfmaq_f32(vr0123, vr0123, vfmsq_f32(vone, vr0123, vd0123));
84 vr4567 = vfmaq_f32(vr4567, vr4567, vfmsq_f32(vone, vr4567, vd4567));
85
86 float32x4_t vf0123 = vmulq_f32(ve0123, vr0123);
87 float32x4_t vf4567 = vmulq_f32(ve4567, vr4567);
88
89 vf0123 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf0123), vcagtq_f32(vx0123, vdenorm_cutoff)));
90 vf4567 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf4567), vcagtq_f32(vx4567, vdenorm_cutoff)));
91
92 const uint32x4_t vm0123 = vcltq_f32(vx0123, vmovq_n_f32(0.0f));
93 const uint32x4_t vm4567 = vcltq_f32(vx4567, vmovq_n_f32(0.0f));
94
95 vf0123 = vbslq_f32(vm0123, vf0123, vsubq_f32(vone, vf0123));
96 vf4567 = vbslq_f32(vm4567, vf4567, vsubq_f32(vone, vf4567));
97
98 vst1q_f32(y, vf0123); y += 4;
99 vst1q_f32(y, vf4567); y += 4;
100 }
101 for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
102 const float32x4_t vx = vld1q_f32(x); x += 4;
103
104 const float32x4_t vz = vabsq_f32(vx);
105
106 float32x4_t vn = vfmaq_f32(vmagic_bias, vz, vminus_log2e);
107 const float32x4_t vs = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn), 23));
108 vn = vsubq_f32(vn, vmagic_bias);
109 float32x4_t vt = vfmaq_f32(vz, vn, vln2);
110
111 float32x4_t vp = vfmaq_f32(vc4, vc5, vt);
112 vp = vfmaq_f32(vc3, vp, vt);
113 vp = vfmaq_f32(vc2, vp, vt);
114 vp = vfmaq_f32(vc1, vp, vt);
115
116 vt = vmulq_f32(vt, vs);
117 const float32x4_t ve = vfmaq_f32(vs, vp, vt);
118 const float32x4_t vd = vaddq_f32(ve, vone);
119
120 float32x4_t vr = vrecpeq_f32(vd);
121 vr = vmulq_f32(vr, vrecpsq_f32(vr, vd));
122 vr = vfmaq_f32(vr, vr, vfmsq_f32(vone, vr, vd));
123
124 float32x4_t vf = vmulq_f32(ve, vr);
125 vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcagtq_f32(vx, vdenorm_cutoff)));
126 const uint32x4_t vm = vcltq_f32(vx, vmovq_n_f32(0.0f));
127 vf = vbslq_f32(vm, vf, vsubq_f32(vone, vf));
128
129 vst1q_f32(y, vf); y += 4;
130 }
131 if XNN_UNLIKELY(n != 0) {
132 const float32x4_t vx = vld1q_f32(x);
133
134 const float32x4_t vz = vabsq_f32(vx);
135
136 float32x4_t vn = vfmaq_f32(vmagic_bias, vz, vminus_log2e);
137 const float32x4_t vs = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn), 23));
138 vn = vsubq_f32(vn, vmagic_bias);
139 float32x4_t vt = vfmaq_f32(vz, vn, vln2);
140
141 float32x4_t vp = vfmaq_f32(vc4, vc5, vt);
142 vp = vfmaq_f32(vc3, vp, vt);
143 vp = vfmaq_f32(vc2, vp, vt);
144 vp = vfmaq_f32(vc1, vp, vt);
145
146 vt = vmulq_f32(vt, vs);
147 const float32x4_t ve = vfmaq_f32(vs, vp, vt);
148 const float32x4_t vd = vaddq_f32(ve, vone);
149
150 float32x4_t vr = vrecpeq_f32(vd);
151 vr = vmulq_f32(vr, vrecpsq_f32(vr, vd));
152 vr = vfmaq_f32(vr, vr, vfmsq_f32(vone, vr, vd));
153
154 float32x4_t vf = vmulq_f32(ve, vr);
155 vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcagtq_f32(vx, vdenorm_cutoff)));
156 const uint32x4_t vm = vcltq_f32(vx, vmovq_n_f32(0.0f));
157 vf = vbslq_f32(vm, vf, vsubq_f32(vone, vf));
158
159 float32x2_t vf_lo = vget_low_f32(vf);
160 if (n & (2 * sizeof(float))) {
161 vst1_f32(y, vf_lo); y += 2;
162 vf_lo = vget_high_f32(vf);
163 }
164 if (n & (1 * sizeof(float))) {
165 vst1_lane_f32(y, vf_lo, 0);
166 }
167 }
168 }
169