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__neon_rr2_p5_nr2recps_x12(size_t n,const float * x,float * y,const void * params)18 void xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x12(
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_hi = vmovq_n_f32(0x1.62E400p-1f);
29 const float32x4_t vln2_lo = vmovq_n_f32(0x1.7F7D1Cp-20f);
30 const float32x4_t vc5 = vmovq_n_f32(-0x1.0F9F9Cp-7f);
31 const float32x4_t vc4 = vmovq_n_f32(0x1.573A1Ap-5f);
32 const float32x4_t vc3 = vmovq_n_f32(-0x1.555A80p-3f);
33 const float32x4_t vc2 = vmovq_n_f32(0x1.FFFDC6p-2f);
34 const float32x4_t vc1 = vmovq_n_f32(-0x1.FFFFF6p-1f);
35 const float32x4_t vone = vmovq_n_f32(1.0f);
36 const float32x4_t vdenorm_cutoff = vmovq_n_f32(-0x1.5D589Ep+6f);
37
38 for (; n >= 12 * sizeof(float); n -= 12 * sizeof(float)) {
39 const float32x4_t vx0123 = vld1q_f32(x); x += 4;
40 const float32x4_t vx4567 = vld1q_f32(x); x += 4;
41 const float32x4_t vx89AB = vld1q_f32(x); x += 4;
42
43 const float32x4_t vz0123 = vabsq_f32(vx0123);
44 const float32x4_t vz4567 = vabsq_f32(vx4567);
45 const float32x4_t vz89AB = vabsq_f32(vx89AB);
46
47 float32x4_t vn0123 = vmlaq_f32(vmagic_bias, vz0123, vminus_log2e);
48 float32x4_t vn4567 = vmlaq_f32(vmagic_bias, vz4567, vminus_log2e);
49 float32x4_t vn89AB = vmlaq_f32(vmagic_bias, vz89AB, vminus_log2e);
50
51 const float32x4_t vs0123 = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn0123), 23));
52 const float32x4_t vs4567 = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn4567), 23));
53 const float32x4_t vs89AB = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn89AB), 23));
54
55 vn0123 = vsubq_f32(vn0123, vmagic_bias);
56 vn4567 = vsubq_f32(vn4567, vmagic_bias);
57 vn89AB = vsubq_f32(vn89AB, vmagic_bias);
58
59 float32x4_t vt0123 = vmlaq_f32(vz0123, vn0123, vln2_hi);
60 float32x4_t vt4567 = vmlaq_f32(vz4567, vn4567, vln2_hi);
61 float32x4_t vt89AB = vmlaq_f32(vz89AB, vn89AB, vln2_hi);
62
63 vt0123 = vmlaq_f32(vt0123, vn0123, vln2_lo);
64 vt4567 = vmlaq_f32(vt4567, vn4567, vln2_lo);
65 vt89AB = vmlaq_f32(vt89AB, vn89AB, vln2_lo);
66
67 float32x4_t vp0123 = vmlaq_f32(vc4, vc5, vt0123);
68 float32x4_t vp4567 = vmlaq_f32(vc4, vc5, vt4567);
69 float32x4_t vp89AB = vmlaq_f32(vc4, vc5, vt89AB);
70
71 vp0123 = vmlaq_f32(vc3, vp0123, vt0123);
72 vp4567 = vmlaq_f32(vc3, vp4567, vt4567);
73 vp89AB = vmlaq_f32(vc3, vp89AB, vt89AB);
74
75 vp0123 = vmlaq_f32(vc2, vp0123, vt0123);
76 vp4567 = vmlaq_f32(vc2, vp4567, vt4567);
77 vp89AB = vmlaq_f32(vc2, vp89AB, vt89AB);
78
79 vp0123 = vmlaq_f32(vc1, vp0123, vt0123);
80 vp4567 = vmlaq_f32(vc1, vp4567, vt4567);
81 vp89AB = vmlaq_f32(vc1, vp89AB, vt89AB);
82
83 vt0123 = vmulq_f32(vt0123, vs0123);
84 vt4567 = vmulq_f32(vt4567, vs4567);
85 vt89AB = vmulq_f32(vt89AB, vs89AB);
86
87 const float32x4_t ve0123 = vmlaq_f32(vs0123, vp0123, vt0123);
88 const float32x4_t ve4567 = vmlaq_f32(vs4567, vp4567, vt4567);
89 const float32x4_t ve89AB = vmlaq_f32(vs89AB, vp89AB, vt89AB);
90
91 const float32x4_t vd0123 = vaddq_f32(ve0123, vone);
92 const float32x4_t vd4567 = vaddq_f32(ve4567, vone);
93 const float32x4_t vd89AB = vaddq_f32(ve89AB, vone);
94
95 float32x4_t vr0123 = vrecpeq_f32(vd0123);
96 float32x4_t vr4567 = vrecpeq_f32(vd4567);
97 float32x4_t vr89AB = vrecpeq_f32(vd89AB);
98
99 vr0123 = vmulq_f32(vr0123, vrecpsq_f32(vr0123, vd0123));
100 vr4567 = vmulq_f32(vr4567, vrecpsq_f32(vr4567, vd4567));
101 vr89AB = vmulq_f32(vr89AB, vrecpsq_f32(vr89AB, vd89AB));
102
103 vr0123 = vmulq_f32(vr0123, vrecpsq_f32(vr0123, vd0123));
104 vr4567 = vmulq_f32(vr4567, vrecpsq_f32(vr4567, vd4567));
105 vr89AB = vmulq_f32(vr89AB, vrecpsq_f32(vr89AB, vd89AB));
106
107 float32x4_t vf0123 = vmulq_f32(ve0123, vr0123);
108 float32x4_t vf4567 = vmulq_f32(ve4567, vr4567);
109 float32x4_t vf89AB = vmulq_f32(ve89AB, vr89AB);
110
111 vf0123 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf0123), vcagtq_f32(vx0123, vdenorm_cutoff)));
112 vf4567 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf4567), vcagtq_f32(vx4567, vdenorm_cutoff)));
113 vf89AB = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf89AB), vcagtq_f32(vx89AB, vdenorm_cutoff)));
114
115 const uint32x4_t vm0123 = vcltq_f32(vx0123, vmovq_n_f32(0.0f));
116 const uint32x4_t vm4567 = vcltq_f32(vx4567, vmovq_n_f32(0.0f));
117 const uint32x4_t vm89AB = vcltq_f32(vx89AB, vmovq_n_f32(0.0f));
118
119 vf0123 = vbslq_f32(vm0123, vf0123, vsubq_f32(vone, vf0123));
120 vf4567 = vbslq_f32(vm4567, vf4567, vsubq_f32(vone, vf4567));
121 vf89AB = vbslq_f32(vm89AB, vf89AB, vsubq_f32(vone, vf89AB));
122
123 vst1q_f32(y, vf0123); y += 4;
124 vst1q_f32(y, vf4567); y += 4;
125 vst1q_f32(y, vf89AB); y += 4;
126 }
127 for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
128 const float32x4_t vx = vld1q_f32(x); x += 4;
129
130 const float32x4_t vz = vabsq_f32(vx);
131
132 float32x4_t vn = vmlaq_f32(vmagic_bias, vz, vminus_log2e);
133 const float32x4_t vs = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn), 23));
134 vn = vsubq_f32(vn, vmagic_bias);
135 float32x4_t vt = vmlaq_f32(vz, vn, vln2_hi);
136 vt = vmlaq_f32(vt, vn, vln2_lo);
137
138 float32x4_t vp = vmlaq_f32(vc4, vc5, vt);
139 vp = vmlaq_f32(vc3, vp, vt);
140 vp = vmlaq_f32(vc2, vp, vt);
141 vp = vmlaq_f32(vc1, vp, vt);
142
143 vt = vmulq_f32(vt, vs);
144 const float32x4_t ve = vmlaq_f32(vs, vp, vt);
145 const float32x4_t vd = vaddq_f32(ve, vone);
146
147 float32x4_t vr = vrecpeq_f32(vd);
148 vr = vmulq_f32(vr, vrecpsq_f32(vr, vd));
149 vr = vmulq_f32(vr, vrecpsq_f32(vr, vd));
150
151 float32x4_t vf = vmulq_f32(ve, vr);
152 vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcagtq_f32(vx, vdenorm_cutoff)));
153 const uint32x4_t vm = vcltq_f32(vx, vmovq_n_f32(0.0f));
154 vf = vbslq_f32(vm, vf, vsubq_f32(vone, vf));
155
156 vst1q_f32(y, vf); y += 4;
157 }
158 if XNN_UNLIKELY(n != 0) {
159 const float32x4_t vx = vld1q_f32(x);
160
161 const float32x4_t vz = vabsq_f32(vx);
162
163 float32x4_t vn = vmlaq_f32(vmagic_bias, vz, vminus_log2e);
164 const float32x4_t vs = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn), 23));
165 vn = vsubq_f32(vn, vmagic_bias);
166 float32x4_t vt = vmlaq_f32(vz, vn, vln2_hi);
167 vt = vmlaq_f32(vt, vn, vln2_lo);
168
169 float32x4_t vp = vmlaq_f32(vc4, vc5, vt);
170 vp = vmlaq_f32(vc3, vp, vt);
171 vp = vmlaq_f32(vc2, vp, vt);
172 vp = vmlaq_f32(vc1, vp, vt);
173
174 vt = vmulq_f32(vt, vs);
175 const float32x4_t ve = vmlaq_f32(vs, vp, vt);
176 const float32x4_t vd = vaddq_f32(ve, vone);
177
178 float32x4_t vr = vrecpeq_f32(vd);
179 vr = vmulq_f32(vr, vrecpsq_f32(vr, vd));
180 vr = vmulq_f32(vr, vrecpsq_f32(vr, vd));
181
182 float32x4_t vf = vmulq_f32(ve, vr);
183 vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcagtq_f32(vx, vdenorm_cutoff)));
184 const uint32x4_t vm = vcltq_f32(vx, vmovq_n_f32(0.0f));
185 vf = vbslq_f32(vm, vf, vsubq_f32(vone, vf));
186
187 float32x2_t vf_lo = vget_low_f32(vf);
188 if (n & (2 * sizeof(float))) {
189 vst1_f32(y, vf_lo); y += 2;
190 vf_lo = vget_high_f32(vf);
191 }
192 if (n & (1 * sizeof(float))) {
193 vst1_lane_f32(y, vf_lo, 0);
194 }
195 }
196 }
197