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__neonfma_rr1_p5_nr2recps_x20(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__neonfma_rr1_p5_nr2recps_x20(
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->neonfma_rr1_p5.magic_bias);
27 const float32x4_t vminus_log2e = vld1q_dup_f32(¶ms->neonfma_rr1_p5.minus_log2e);
28 const float32x4_t vln2 = vld1q_dup_f32(¶ms->neonfma_rr1_p5.ln2);
29 const float32x4_t vc5 = vld1q_dup_f32(¶ms->neonfma_rr1_p5.c5);
30 const float32x4_t vc4 = vld1q_dup_f32(¶ms->neonfma_rr1_p5.c4);
31 const float32x4_t vc3 = vld1q_dup_f32(¶ms->neonfma_rr1_p5.c3);
32 const float32x4_t vc2 = vld1q_dup_f32(¶ms->neonfma_rr1_p5.c2);
33 const float32x4_t vc1 = vld1q_dup_f32(¶ms->neonfma_rr1_p5.c1);
34 const float32x4_t vone = vmovq_n_f32(1.0f);
35 const float32x4_t vdenorm_cutoff = vld1q_dup_f32(¶ms->neonfma_rr1_p5.denorm_cutoff);
36
37 for (; n >= 20 * sizeof(float); n -= 20 * sizeof(float)) {
38 const float32x4_t vx0123 = vld1q_f32(x); x += 4;
39 const float32x4_t vx4567 = vld1q_f32(x); x += 4;
40 const float32x4_t vx89AB = vld1q_f32(x); x += 4;
41 const float32x4_t vxCDEF = vld1q_f32(x); x += 4;
42 const float32x4_t vxGHIJ = 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 const float32x4_t vzGHIJ = vabsq_f32(vxGHIJ);
49
50 float32x4_t vn0123 = vfmaq_f32(vmagic_bias, vz0123, vminus_log2e);
51 float32x4_t vn4567 = vfmaq_f32(vmagic_bias, vz4567, vminus_log2e);
52 float32x4_t vn89AB = vfmaq_f32(vmagic_bias, vz89AB, vminus_log2e);
53 float32x4_t vnCDEF = vfmaq_f32(vmagic_bias, vzCDEF, vminus_log2e);
54 float32x4_t vnGHIJ = vfmaq_f32(vmagic_bias, vzGHIJ, vminus_log2e);
55
56 const float32x4_t vs0123 = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn0123), 23));
57 const float32x4_t vs4567 = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn4567), 23));
58 const float32x4_t vs89AB = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn89AB), 23));
59 const float32x4_t vsCDEF = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vnCDEF), 23));
60 const float32x4_t vsGHIJ = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vnGHIJ), 23));
61
62 vn0123 = vsubq_f32(vn0123, vmagic_bias);
63 vn4567 = vsubq_f32(vn4567, vmagic_bias);
64 vn89AB = vsubq_f32(vn89AB, vmagic_bias);
65 vnCDEF = vsubq_f32(vnCDEF, vmagic_bias);
66 vnGHIJ = vsubq_f32(vnGHIJ, vmagic_bias);
67
68 float32x4_t vt0123 = vfmaq_f32(vz0123, vn0123, vln2);
69 float32x4_t vt4567 = vfmaq_f32(vz4567, vn4567, vln2);
70 float32x4_t vt89AB = vfmaq_f32(vz89AB, vn89AB, vln2);
71 float32x4_t vtCDEF = vfmaq_f32(vzCDEF, vnCDEF, vln2);
72 float32x4_t vtGHIJ = vfmaq_f32(vzGHIJ, vnGHIJ, vln2);
73
74 float32x4_t vp0123 = vfmaq_f32(vc4, vc5, vt0123);
75 float32x4_t vp4567 = vfmaq_f32(vc4, vc5, vt4567);
76 float32x4_t vp89AB = vfmaq_f32(vc4, vc5, vt89AB);
77 float32x4_t vpCDEF = vfmaq_f32(vc4, vc5, vtCDEF);
78 float32x4_t vpGHIJ = vfmaq_f32(vc4, vc5, vtGHIJ);
79
80 vp0123 = vfmaq_f32(vc3, vp0123, vt0123);
81 vp4567 = vfmaq_f32(vc3, vp4567, vt4567);
82 vp89AB = vfmaq_f32(vc3, vp89AB, vt89AB);
83 vpCDEF = vfmaq_f32(vc3, vpCDEF, vtCDEF);
84 vpGHIJ = vfmaq_f32(vc3, vpGHIJ, vtGHIJ);
85
86 vp0123 = vfmaq_f32(vc2, vp0123, vt0123);
87 vp4567 = vfmaq_f32(vc2, vp4567, vt4567);
88 vp89AB = vfmaq_f32(vc2, vp89AB, vt89AB);
89 vpCDEF = vfmaq_f32(vc2, vpCDEF, vtCDEF);
90 vpGHIJ = vfmaq_f32(vc2, vpGHIJ, vtGHIJ);
91
92 vp0123 = vfmaq_f32(vc1, vp0123, vt0123);
93 vp4567 = vfmaq_f32(vc1, vp4567, vt4567);
94 vp89AB = vfmaq_f32(vc1, vp89AB, vt89AB);
95 vpCDEF = vfmaq_f32(vc1, vpCDEF, vtCDEF);
96 vpGHIJ = vfmaq_f32(vc1, vpGHIJ, vtGHIJ);
97
98 vt0123 = vmulq_f32(vt0123, vs0123);
99 vt4567 = vmulq_f32(vt4567, vs4567);
100 vt89AB = vmulq_f32(vt89AB, vs89AB);
101 vtCDEF = vmulq_f32(vtCDEF, vsCDEF);
102 vtGHIJ = vmulq_f32(vtGHIJ, vsGHIJ);
103
104 const float32x4_t ve0123 = vfmaq_f32(vs0123, vp0123, vt0123);
105 const float32x4_t ve4567 = vfmaq_f32(vs4567, vp4567, vt4567);
106 const float32x4_t ve89AB = vfmaq_f32(vs89AB, vp89AB, vt89AB);
107 const float32x4_t veCDEF = vfmaq_f32(vsCDEF, vpCDEF, vtCDEF);
108 const float32x4_t veGHIJ = vfmaq_f32(vsGHIJ, vpGHIJ, vtGHIJ);
109
110 const float32x4_t vd0123 = vaddq_f32(ve0123, vone);
111 const float32x4_t vd4567 = vaddq_f32(ve4567, vone);
112 const float32x4_t vd89AB = vaddq_f32(ve89AB, vone);
113 const float32x4_t vdCDEF = vaddq_f32(veCDEF, vone);
114 const float32x4_t vdGHIJ = vaddq_f32(veGHIJ, vone);
115
116 float32x4_t vr0123 = vrecpeq_f32(vd0123);
117 float32x4_t vr4567 = vrecpeq_f32(vd4567);
118 float32x4_t vr89AB = vrecpeq_f32(vd89AB);
119 float32x4_t vrCDEF = vrecpeq_f32(vdCDEF);
120 float32x4_t vrGHIJ = vrecpeq_f32(vdGHIJ);
121
122 vr0123 = vmulq_f32(vr0123, vrecpsq_f32(vr0123, vd0123));
123 vr4567 = vmulq_f32(vr4567, vrecpsq_f32(vr4567, vd4567));
124 vr89AB = vmulq_f32(vr89AB, vrecpsq_f32(vr89AB, vd89AB));
125 vrCDEF = vmulq_f32(vrCDEF, vrecpsq_f32(vrCDEF, vdCDEF));
126 vrGHIJ = vmulq_f32(vrGHIJ, vrecpsq_f32(vrGHIJ, vdGHIJ));
127
128 vr0123 = vmulq_f32(vr0123, vrecpsq_f32(vr0123, vd0123));
129 vr4567 = vmulq_f32(vr4567, vrecpsq_f32(vr4567, vd4567));
130 vr89AB = vmulq_f32(vr89AB, vrecpsq_f32(vr89AB, vd89AB));
131 vrCDEF = vmulq_f32(vrCDEF, vrecpsq_f32(vrCDEF, vdCDEF));
132 vrGHIJ = vmulq_f32(vrGHIJ, vrecpsq_f32(vrGHIJ, vdGHIJ));
133
134 float32x4_t vf0123 = vmulq_f32(ve0123, vr0123);
135 float32x4_t vf4567 = vmulq_f32(ve4567, vr4567);
136 float32x4_t vf89AB = vmulq_f32(ve89AB, vr89AB);
137 float32x4_t vfCDEF = vmulq_f32(veCDEF, vrCDEF);
138 float32x4_t vfGHIJ = vmulq_f32(veGHIJ, vrGHIJ);
139
140 vf0123 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf0123), vcagtq_f32(vx0123, vdenorm_cutoff)));
141 vf4567 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf4567), vcagtq_f32(vx4567, vdenorm_cutoff)));
142 vf89AB = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf89AB), vcagtq_f32(vx89AB, vdenorm_cutoff)));
143 vfCDEF = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vfCDEF), vcagtq_f32(vxCDEF, vdenorm_cutoff)));
144 vfGHIJ = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vfGHIJ), vcagtq_f32(vxGHIJ, vdenorm_cutoff)));
145
146 const uint32x4_t vm0123 = vcltq_f32(vx0123, vmovq_n_f32(0.0f));
147 const uint32x4_t vm4567 = vcltq_f32(vx4567, vmovq_n_f32(0.0f));
148 const uint32x4_t vm89AB = vcltq_f32(vx89AB, vmovq_n_f32(0.0f));
149 const uint32x4_t vmCDEF = vcltq_f32(vxCDEF, vmovq_n_f32(0.0f));
150 const uint32x4_t vmGHIJ = vcltq_f32(vxGHIJ, vmovq_n_f32(0.0f));
151
152 vf0123 = vbslq_f32(vm0123, vf0123, vsubq_f32(vone, vf0123));
153 vf4567 = vbslq_f32(vm4567, vf4567, vsubq_f32(vone, vf4567));
154 vf89AB = vbslq_f32(vm89AB, vf89AB, vsubq_f32(vone, vf89AB));
155 vfCDEF = vbslq_f32(vmCDEF, vfCDEF, vsubq_f32(vone, vfCDEF));
156 vfGHIJ = vbslq_f32(vmGHIJ, vfGHIJ, vsubq_f32(vone, vfGHIJ));
157
158 vst1q_f32(y, vf0123); y += 4;
159 vst1q_f32(y, vf4567); y += 4;
160 vst1q_f32(y, vf89AB); y += 4;
161 vst1q_f32(y, vfCDEF); y += 4;
162 vst1q_f32(y, vfGHIJ); y += 4;
163 }
164 for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
165 const float32x4_t vx = vld1q_f32(x); x += 4;
166
167 const float32x4_t vz = vabsq_f32(vx);
168
169 float32x4_t vn = vfmaq_f32(vmagic_bias, vz, vminus_log2e);
170 const float32x4_t vs = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn), 23));
171 vn = vsubq_f32(vn, vmagic_bias);
172 float32x4_t vt = vfmaq_f32(vz, vn, vln2);
173
174 float32x4_t vp = vfmaq_f32(vc4, vc5, vt);
175 vp = vfmaq_f32(vc3, vp, vt);
176 vp = vfmaq_f32(vc2, vp, vt);
177 vp = vfmaq_f32(vc1, vp, vt);
178
179 vt = vmulq_f32(vt, vs);
180 const float32x4_t ve = vfmaq_f32(vs, vp, vt);
181 const float32x4_t vd = vaddq_f32(ve, vone);
182
183 float32x4_t vr = vrecpeq_f32(vd);
184 vr = vmulq_f32(vr, vrecpsq_f32(vr, vd));
185 vr = vmulq_f32(vr, vrecpsq_f32(vr, vd));
186
187 float32x4_t vf = vmulq_f32(ve, vr);
188 vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcagtq_f32(vx, vdenorm_cutoff)));
189 const uint32x4_t vm = vcltq_f32(vx, vmovq_n_f32(0.0f));
190 vf = vbslq_f32(vm, vf, vsubq_f32(vone, vf));
191
192 vst1q_f32(y, vf); y += 4;
193 }
194 if XNN_UNLIKELY(n != 0) {
195 const float32x4_t vx = vld1q_f32(x);
196
197 const float32x4_t vz = vabsq_f32(vx);
198
199 float32x4_t vn = vfmaq_f32(vmagic_bias, vz, vminus_log2e);
200 const float32x4_t vs = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn), 23));
201 vn = vsubq_f32(vn, vmagic_bias);
202 float32x4_t vt = vfmaq_f32(vz, vn, vln2);
203
204 float32x4_t vp = vfmaq_f32(vc4, vc5, vt);
205 vp = vfmaq_f32(vc3, vp, vt);
206 vp = vfmaq_f32(vc2, vp, vt);
207 vp = vfmaq_f32(vc1, vp, vt);
208
209 vt = vmulq_f32(vt, vs);
210 const float32x4_t ve = vfmaq_f32(vs, vp, vt);
211 const float32x4_t vd = vaddq_f32(ve, vone);
212
213 float32x4_t vr = vrecpeq_f32(vd);
214 vr = vmulq_f32(vr, vrecpsq_f32(vr, vd));
215 vr = vmulq_f32(vr, vrecpsq_f32(vr, vd));
216
217 float32x4_t vf = vmulq_f32(ve, vr);
218 vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcagtq_f32(vx, vdenorm_cutoff)));
219 const uint32x4_t vm = vcltq_f32(vx, vmovq_n_f32(0.0f));
220 vf = vbslq_f32(vm, vf, vsubq_f32(vone, vf));
221
222 float32x2_t vf_lo = vget_low_f32(vf);
223 if (n & (2 * sizeof(float))) {
224 vst1_f32(y, vf_lo); y += 2;
225 vf_lo = vget_high_f32(vf);
226 }
227 if (n & (1 * sizeof(float))) {
228 vst1_lane_f32(y, vf_lo, 0);
229 }
230 }
231 }
232