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_x16(size_t n,const float * x,float * y,const void * params)18 void xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x16(
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 >= 16 * sizeof(float); n -= 16 * 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 const float32x4_t vxCDEF = vld1q_f32(x); x += 4;
43
44 const float32x4_t vz0123 = vabsq_f32(vx0123);
45 const float32x4_t vz4567 = vabsq_f32(vx4567);
46 const float32x4_t vz89AB = vabsq_f32(vx89AB);
47 const float32x4_t vzCDEF = vabsq_f32(vxCDEF);
48
49 float32x4_t vn0123 = vmlaq_f32(vmagic_bias, vz0123, vminus_log2e);
50 float32x4_t vn4567 = vmlaq_f32(vmagic_bias, vz4567, vminus_log2e);
51 float32x4_t vn89AB = vmlaq_f32(vmagic_bias, vz89AB, vminus_log2e);
52 float32x4_t vnCDEF = vmlaq_f32(vmagic_bias, vzCDEF, vminus_log2e);
53
54 const float32x4_t vs0123 = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn0123), 23));
55 const float32x4_t vs4567 = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn4567), 23));
56 const float32x4_t vs89AB = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn89AB), 23));
57 const float32x4_t vsCDEF = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vnCDEF), 23));
58
59 vn0123 = vsubq_f32(vn0123, vmagic_bias);
60 vn4567 = vsubq_f32(vn4567, vmagic_bias);
61 vn89AB = vsubq_f32(vn89AB, vmagic_bias);
62 vnCDEF = vsubq_f32(vnCDEF, vmagic_bias);
63
64 float32x4_t vt0123 = vmlaq_f32(vz0123, vn0123, vln2_hi);
65 float32x4_t vt4567 = vmlaq_f32(vz4567, vn4567, vln2_hi);
66 float32x4_t vt89AB = vmlaq_f32(vz89AB, vn89AB, vln2_hi);
67 float32x4_t vtCDEF = vmlaq_f32(vzCDEF, vnCDEF, vln2_hi);
68
69 vt0123 = vmlaq_f32(vt0123, vn0123, vln2_lo);
70 vt4567 = vmlaq_f32(vt4567, vn4567, vln2_lo);
71 vt89AB = vmlaq_f32(vt89AB, vn89AB, vln2_lo);
72 vtCDEF = vmlaq_f32(vtCDEF, vnCDEF, vln2_lo);
73
74 float32x4_t vp0123 = vmlaq_f32(vc4, vc5, vt0123);
75 float32x4_t vp4567 = vmlaq_f32(vc4, vc5, vt4567);
76 float32x4_t vp89AB = vmlaq_f32(vc4, vc5, vt89AB);
77 float32x4_t vpCDEF = vmlaq_f32(vc4, vc5, vtCDEF);
78
79 vp0123 = vmlaq_f32(vc3, vp0123, vt0123);
80 vp4567 = vmlaq_f32(vc3, vp4567, vt4567);
81 vp89AB = vmlaq_f32(vc3, vp89AB, vt89AB);
82 vpCDEF = vmlaq_f32(vc3, vpCDEF, vtCDEF);
83
84 vp0123 = vmlaq_f32(vc2, vp0123, vt0123);
85 vp4567 = vmlaq_f32(vc2, vp4567, vt4567);
86 vp89AB = vmlaq_f32(vc2, vp89AB, vt89AB);
87 vpCDEF = vmlaq_f32(vc2, vpCDEF, vtCDEF);
88
89 vp0123 = vmlaq_f32(vc1, vp0123, vt0123);
90 vp4567 = vmlaq_f32(vc1, vp4567, vt4567);
91 vp89AB = vmlaq_f32(vc1, vp89AB, vt89AB);
92 vpCDEF = vmlaq_f32(vc1, vpCDEF, vtCDEF);
93
94 vt0123 = vmulq_f32(vt0123, vs0123);
95 vt4567 = vmulq_f32(vt4567, vs4567);
96 vt89AB = vmulq_f32(vt89AB, vs89AB);
97 vtCDEF = vmulq_f32(vtCDEF, vsCDEF);
98
99 const float32x4_t ve0123 = vmlaq_f32(vs0123, vp0123, vt0123);
100 const float32x4_t ve4567 = vmlaq_f32(vs4567, vp4567, vt4567);
101 const float32x4_t ve89AB = vmlaq_f32(vs89AB, vp89AB, vt89AB);
102 const float32x4_t veCDEF = vmlaq_f32(vsCDEF, vpCDEF, vtCDEF);
103
104 const float32x4_t vd0123 = vaddq_f32(ve0123, vone);
105 const float32x4_t vd4567 = vaddq_f32(ve4567, vone);
106 const float32x4_t vd89AB = vaddq_f32(ve89AB, vone);
107 const float32x4_t vdCDEF = vaddq_f32(veCDEF, vone);
108
109 float32x4_t vr0123 = vrecpeq_f32(vd0123);
110 float32x4_t vr4567 = vrecpeq_f32(vd4567);
111 float32x4_t vr89AB = vrecpeq_f32(vd89AB);
112 float32x4_t vrCDEF = vrecpeq_f32(vdCDEF);
113
114 vr0123 = vmulq_f32(vr0123, vrecpsq_f32(vr0123, vd0123));
115 vr4567 = vmulq_f32(vr4567, vrecpsq_f32(vr4567, vd4567));
116 vr89AB = vmulq_f32(vr89AB, vrecpsq_f32(vr89AB, vd89AB));
117 vrCDEF = vmulq_f32(vrCDEF, vrecpsq_f32(vrCDEF, vdCDEF));
118
119 vr0123 = vmulq_f32(vr0123, vrecpsq_f32(vr0123, vd0123));
120 vr4567 = vmulq_f32(vr4567, vrecpsq_f32(vr4567, vd4567));
121 vr89AB = vmulq_f32(vr89AB, vrecpsq_f32(vr89AB, vd89AB));
122 vrCDEF = vmulq_f32(vrCDEF, vrecpsq_f32(vrCDEF, vdCDEF));
123
124 float32x4_t vf0123 = vmulq_f32(ve0123, vr0123);
125 float32x4_t vf4567 = vmulq_f32(ve4567, vr4567);
126 float32x4_t vf89AB = vmulq_f32(ve89AB, vr89AB);
127 float32x4_t vfCDEF = vmulq_f32(veCDEF, vrCDEF);
128
129 vf0123 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf0123), vcagtq_f32(vx0123, vdenorm_cutoff)));
130 vf4567 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf4567), vcagtq_f32(vx4567, vdenorm_cutoff)));
131 vf89AB = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf89AB), vcagtq_f32(vx89AB, vdenorm_cutoff)));
132 vfCDEF = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vfCDEF), vcagtq_f32(vxCDEF, vdenorm_cutoff)));
133
134 const uint32x4_t vm0123 = vcltq_f32(vx0123, vmovq_n_f32(0.0f));
135 const uint32x4_t vm4567 = vcltq_f32(vx4567, vmovq_n_f32(0.0f));
136 const uint32x4_t vm89AB = vcltq_f32(vx89AB, vmovq_n_f32(0.0f));
137 const uint32x4_t vmCDEF = vcltq_f32(vxCDEF, vmovq_n_f32(0.0f));
138
139 vf0123 = vbslq_f32(vm0123, vf0123, vsubq_f32(vone, vf0123));
140 vf4567 = vbslq_f32(vm4567, vf4567, vsubq_f32(vone, vf4567));
141 vf89AB = vbslq_f32(vm89AB, vf89AB, vsubq_f32(vone, vf89AB));
142 vfCDEF = vbslq_f32(vmCDEF, vfCDEF, vsubq_f32(vone, vfCDEF));
143
144 vst1q_f32(y, vf0123); y += 4;
145 vst1q_f32(y, vf4567); y += 4;
146 vst1q_f32(y, vf89AB); y += 4;
147 vst1q_f32(y, vfCDEF); y += 4;
148 }
149 for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
150 const float32x4_t vx = vld1q_f32(x); x += 4;
151
152 const float32x4_t vz = vabsq_f32(vx);
153
154 float32x4_t vn = vmlaq_f32(vmagic_bias, vz, vminus_log2e);
155 const float32x4_t vs = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn), 23));
156 vn = vsubq_f32(vn, vmagic_bias);
157 float32x4_t vt = vmlaq_f32(vz, vn, vln2_hi);
158 vt = vmlaq_f32(vt, vn, vln2_lo);
159
160 float32x4_t vp = vmlaq_f32(vc4, vc5, vt);
161 vp = vmlaq_f32(vc3, vp, vt);
162 vp = vmlaq_f32(vc2, vp, vt);
163 vp = vmlaq_f32(vc1, vp, vt);
164
165 vt = vmulq_f32(vt, vs);
166 const float32x4_t ve = vmlaq_f32(vs, vp, vt);
167 const float32x4_t vd = vaddq_f32(ve, vone);
168
169 float32x4_t vr = vrecpeq_f32(vd);
170 vr = vmulq_f32(vr, vrecpsq_f32(vr, vd));
171 vr = vmulq_f32(vr, vrecpsq_f32(vr, vd));
172
173 float32x4_t vf = vmulq_f32(ve, vr);
174 vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcagtq_f32(vx, vdenorm_cutoff)));
175 const uint32x4_t vm = vcltq_f32(vx, vmovq_n_f32(0.0f));
176 vf = vbslq_f32(vm, vf, vsubq_f32(vone, vf));
177
178 vst1q_f32(y, vf); y += 4;
179 }
180 if XNN_UNLIKELY(n != 0) {
181 const float32x4_t vx = vld1q_f32(x);
182
183 const float32x4_t vz = vabsq_f32(vx);
184
185 float32x4_t vn = vmlaq_f32(vmagic_bias, vz, vminus_log2e);
186 const float32x4_t vs = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn), 23));
187 vn = vsubq_f32(vn, vmagic_bias);
188 float32x4_t vt = vmlaq_f32(vz, vn, vln2_hi);
189 vt = vmlaq_f32(vt, vn, vln2_lo);
190
191 float32x4_t vp = vmlaq_f32(vc4, vc5, vt);
192 vp = vmlaq_f32(vc3, vp, vt);
193 vp = vmlaq_f32(vc2, vp, vt);
194 vp = vmlaq_f32(vc1, vp, vt);
195
196 vt = vmulq_f32(vt, vs);
197 const float32x4_t ve = vmlaq_f32(vs, vp, vt);
198 const float32x4_t vd = vaddq_f32(ve, vone);
199
200 float32x4_t vr = vrecpeq_f32(vd);
201 vr = vmulq_f32(vr, vrecpsq_f32(vr, vd));
202 vr = vmulq_f32(vr, vrecpsq_f32(vr, vd));
203
204 float32x4_t vf = vmulq_f32(ve, vr);
205 vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcagtq_f32(vx, vdenorm_cutoff)));
206 const uint32x4_t vm = vcltq_f32(vx, vmovq_n_f32(0.0f));
207 vf = vbslq_f32(vm, vf, vsubq_f32(vone, vf));
208
209 float32x2_t vf_lo = vget_low_f32(vf);
210 if (n & (2 * sizeof(float))) {
211 vst1_f32(y, vf_lo); y += 2;
212 vf_lo = vget_high_f32(vf);
213 }
214 if (n & (1 * sizeof(float))) {
215 vst1_lane_f32(y, vf_lo, 0);
216 }
217 }
218 }
219