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_x24(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_x24(
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 >= 24 * sizeof(float); n -= 24 * 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 const float32x4_t vxGHIJ = vld1q_f32(x); x += 4;
44 const float32x4_t vxKLMN = vld1q_f32(x); x += 4;
45
46 const float32x4_t vz0123 = vabsq_f32(vx0123);
47 const float32x4_t vz4567 = vabsq_f32(vx4567);
48 const float32x4_t vz89AB = vabsq_f32(vx89AB);
49 const float32x4_t vzCDEF = vabsq_f32(vxCDEF);
50 const float32x4_t vzGHIJ = vabsq_f32(vxGHIJ);
51 const float32x4_t vzKLMN = vabsq_f32(vxKLMN);
52
53 float32x4_t vn0123 = vmlaq_f32(vmagic_bias, vz0123, vminus_log2e);
54 float32x4_t vn4567 = vmlaq_f32(vmagic_bias, vz4567, vminus_log2e);
55 float32x4_t vn89AB = vmlaq_f32(vmagic_bias, vz89AB, vminus_log2e);
56 float32x4_t vnCDEF = vmlaq_f32(vmagic_bias, vzCDEF, vminus_log2e);
57 float32x4_t vnGHIJ = vmlaq_f32(vmagic_bias, vzGHIJ, vminus_log2e);
58 float32x4_t vnKLMN = vmlaq_f32(vmagic_bias, vzKLMN, vminus_log2e);
59
60 const float32x4_t vs0123 = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn0123), 23));
61 const float32x4_t vs4567 = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn4567), 23));
62 const float32x4_t vs89AB = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn89AB), 23));
63 const float32x4_t vsCDEF = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vnCDEF), 23));
64 const float32x4_t vsGHIJ = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vnGHIJ), 23));
65 const float32x4_t vsKLMN = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vnKLMN), 23));
66
67 vn0123 = vsubq_f32(vn0123, vmagic_bias);
68 vn4567 = vsubq_f32(vn4567, vmagic_bias);
69 vn89AB = vsubq_f32(vn89AB, vmagic_bias);
70 vnCDEF = vsubq_f32(vnCDEF, vmagic_bias);
71 vnGHIJ = vsubq_f32(vnGHIJ, vmagic_bias);
72 vnKLMN = vsubq_f32(vnKLMN, vmagic_bias);
73
74 float32x4_t vt0123 = vmlaq_f32(vz0123, vn0123, vln2_hi);
75 float32x4_t vt4567 = vmlaq_f32(vz4567, vn4567, vln2_hi);
76 float32x4_t vt89AB = vmlaq_f32(vz89AB, vn89AB, vln2_hi);
77 float32x4_t vtCDEF = vmlaq_f32(vzCDEF, vnCDEF, vln2_hi);
78 float32x4_t vtGHIJ = vmlaq_f32(vzGHIJ, vnGHIJ, vln2_hi);
79 float32x4_t vtKLMN = vmlaq_f32(vzKLMN, vnKLMN, vln2_hi);
80
81 vt0123 = vmlaq_f32(vt0123, vn0123, vln2_lo);
82 vt4567 = vmlaq_f32(vt4567, vn4567, vln2_lo);
83 vt89AB = vmlaq_f32(vt89AB, vn89AB, vln2_lo);
84 vtCDEF = vmlaq_f32(vtCDEF, vnCDEF, vln2_lo);
85 vtGHIJ = vmlaq_f32(vtGHIJ, vnGHIJ, vln2_lo);
86 vtKLMN = vmlaq_f32(vtKLMN, vnKLMN, vln2_lo);
87
88 float32x4_t vp0123 = vmlaq_f32(vc4, vc5, vt0123);
89 float32x4_t vp4567 = vmlaq_f32(vc4, vc5, vt4567);
90 float32x4_t vp89AB = vmlaq_f32(vc4, vc5, vt89AB);
91 float32x4_t vpCDEF = vmlaq_f32(vc4, vc5, vtCDEF);
92 float32x4_t vpGHIJ = vmlaq_f32(vc4, vc5, vtGHIJ);
93 float32x4_t vpKLMN = vmlaq_f32(vc4, vc5, vtKLMN);
94
95 vp0123 = vmlaq_f32(vc3, vp0123, vt0123);
96 vp4567 = vmlaq_f32(vc3, vp4567, vt4567);
97 vp89AB = vmlaq_f32(vc3, vp89AB, vt89AB);
98 vpCDEF = vmlaq_f32(vc3, vpCDEF, vtCDEF);
99 vpGHIJ = vmlaq_f32(vc3, vpGHIJ, vtGHIJ);
100 vpKLMN = vmlaq_f32(vc3, vpKLMN, vtKLMN);
101
102 vp0123 = vmlaq_f32(vc2, vp0123, vt0123);
103 vp4567 = vmlaq_f32(vc2, vp4567, vt4567);
104 vp89AB = vmlaq_f32(vc2, vp89AB, vt89AB);
105 vpCDEF = vmlaq_f32(vc2, vpCDEF, vtCDEF);
106 vpGHIJ = vmlaq_f32(vc2, vpGHIJ, vtGHIJ);
107 vpKLMN = vmlaq_f32(vc2, vpKLMN, vtKLMN);
108
109 vp0123 = vmlaq_f32(vc1, vp0123, vt0123);
110 vp4567 = vmlaq_f32(vc1, vp4567, vt4567);
111 vp89AB = vmlaq_f32(vc1, vp89AB, vt89AB);
112 vpCDEF = vmlaq_f32(vc1, vpCDEF, vtCDEF);
113 vpGHIJ = vmlaq_f32(vc1, vpGHIJ, vtGHIJ);
114 vpKLMN = vmlaq_f32(vc1, vpKLMN, vtKLMN);
115
116 vt0123 = vmulq_f32(vt0123, vs0123);
117 vt4567 = vmulq_f32(vt4567, vs4567);
118 vt89AB = vmulq_f32(vt89AB, vs89AB);
119 vtCDEF = vmulq_f32(vtCDEF, vsCDEF);
120 vtGHIJ = vmulq_f32(vtGHIJ, vsGHIJ);
121 vtKLMN = vmulq_f32(vtKLMN, vsKLMN);
122
123 const float32x4_t ve0123 = vmlaq_f32(vs0123, vp0123, vt0123);
124 const float32x4_t ve4567 = vmlaq_f32(vs4567, vp4567, vt4567);
125 const float32x4_t ve89AB = vmlaq_f32(vs89AB, vp89AB, vt89AB);
126 const float32x4_t veCDEF = vmlaq_f32(vsCDEF, vpCDEF, vtCDEF);
127 const float32x4_t veGHIJ = vmlaq_f32(vsGHIJ, vpGHIJ, vtGHIJ);
128 const float32x4_t veKLMN = vmlaq_f32(vsKLMN, vpKLMN, vtKLMN);
129
130 const float32x4_t vd0123 = vaddq_f32(ve0123, vone);
131 const float32x4_t vd4567 = vaddq_f32(ve4567, vone);
132 const float32x4_t vd89AB = vaddq_f32(ve89AB, vone);
133 const float32x4_t vdCDEF = vaddq_f32(veCDEF, vone);
134 const float32x4_t vdGHIJ = vaddq_f32(veGHIJ, vone);
135 const float32x4_t vdKLMN = vaddq_f32(veKLMN, vone);
136
137 float32x4_t vr0123 = vrecpeq_f32(vd0123);
138 float32x4_t vr4567 = vrecpeq_f32(vd4567);
139 float32x4_t vr89AB = vrecpeq_f32(vd89AB);
140 float32x4_t vrCDEF = vrecpeq_f32(vdCDEF);
141 float32x4_t vrGHIJ = vrecpeq_f32(vdGHIJ);
142 float32x4_t vrKLMN = vrecpeq_f32(vdKLMN);
143
144 vr0123 = vmulq_f32(vr0123, vrecpsq_f32(vr0123, vd0123));
145 vr4567 = vmulq_f32(vr4567, vrecpsq_f32(vr4567, vd4567));
146 vr89AB = vmulq_f32(vr89AB, vrecpsq_f32(vr89AB, vd89AB));
147 vrCDEF = vmulq_f32(vrCDEF, vrecpsq_f32(vrCDEF, vdCDEF));
148 vrGHIJ = vmulq_f32(vrGHIJ, vrecpsq_f32(vrGHIJ, vdGHIJ));
149 vrKLMN = vmulq_f32(vrKLMN, vrecpsq_f32(vrKLMN, vdKLMN));
150
151 vr0123 = vmulq_f32(vr0123, vrecpsq_f32(vr0123, vd0123));
152 vr4567 = vmulq_f32(vr4567, vrecpsq_f32(vr4567, vd4567));
153 vr89AB = vmulq_f32(vr89AB, vrecpsq_f32(vr89AB, vd89AB));
154 vrCDEF = vmulq_f32(vrCDEF, vrecpsq_f32(vrCDEF, vdCDEF));
155 vrGHIJ = vmulq_f32(vrGHIJ, vrecpsq_f32(vrGHIJ, vdGHIJ));
156 vrKLMN = vmulq_f32(vrKLMN, vrecpsq_f32(vrKLMN, vdKLMN));
157
158 float32x4_t vf0123 = vmulq_f32(ve0123, vr0123);
159 float32x4_t vf4567 = vmulq_f32(ve4567, vr4567);
160 float32x4_t vf89AB = vmulq_f32(ve89AB, vr89AB);
161 float32x4_t vfCDEF = vmulq_f32(veCDEF, vrCDEF);
162 float32x4_t vfGHIJ = vmulq_f32(veGHIJ, vrGHIJ);
163 float32x4_t vfKLMN = vmulq_f32(veKLMN, vrKLMN);
164
165 vf0123 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf0123), vcagtq_f32(vx0123, vdenorm_cutoff)));
166 vf4567 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf4567), vcagtq_f32(vx4567, vdenorm_cutoff)));
167 vf89AB = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf89AB), vcagtq_f32(vx89AB, vdenorm_cutoff)));
168 vfCDEF = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vfCDEF), vcagtq_f32(vxCDEF, vdenorm_cutoff)));
169 vfGHIJ = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vfGHIJ), vcagtq_f32(vxGHIJ, vdenorm_cutoff)));
170 vfKLMN = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vfKLMN), vcagtq_f32(vxKLMN, vdenorm_cutoff)));
171
172 const uint32x4_t vm0123 = vcltq_f32(vx0123, vmovq_n_f32(0.0f));
173 const uint32x4_t vm4567 = vcltq_f32(vx4567, vmovq_n_f32(0.0f));
174 const uint32x4_t vm89AB = vcltq_f32(vx89AB, vmovq_n_f32(0.0f));
175 const uint32x4_t vmCDEF = vcltq_f32(vxCDEF, vmovq_n_f32(0.0f));
176 const uint32x4_t vmGHIJ = vcltq_f32(vxGHIJ, vmovq_n_f32(0.0f));
177 const uint32x4_t vmKLMN = vcltq_f32(vxKLMN, vmovq_n_f32(0.0f));
178
179 vf0123 = vbslq_f32(vm0123, vf0123, vsubq_f32(vone, vf0123));
180 vf4567 = vbslq_f32(vm4567, vf4567, vsubq_f32(vone, vf4567));
181 vf89AB = vbslq_f32(vm89AB, vf89AB, vsubq_f32(vone, vf89AB));
182 vfCDEF = vbslq_f32(vmCDEF, vfCDEF, vsubq_f32(vone, vfCDEF));
183 vfGHIJ = vbslq_f32(vmGHIJ, vfGHIJ, vsubq_f32(vone, vfGHIJ));
184 vfKLMN = vbslq_f32(vmKLMN, vfKLMN, vsubq_f32(vone, vfKLMN));
185
186 vst1q_f32(y, vf0123); y += 4;
187 vst1q_f32(y, vf4567); y += 4;
188 vst1q_f32(y, vf89AB); y += 4;
189 vst1q_f32(y, vfCDEF); y += 4;
190 vst1q_f32(y, vfGHIJ); y += 4;
191 vst1q_f32(y, vfKLMN); y += 4;
192 }
193 for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
194 const float32x4_t vx = vld1q_f32(x); x += 4;
195
196 const float32x4_t vz = vabsq_f32(vx);
197
198 float32x4_t vn = vmlaq_f32(vmagic_bias, vz, vminus_log2e);
199 const float32x4_t vs = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn), 23));
200 vn = vsubq_f32(vn, vmagic_bias);
201 float32x4_t vt = vmlaq_f32(vz, vn, vln2_hi);
202 vt = vmlaq_f32(vt, vn, vln2_lo);
203
204 float32x4_t vp = vmlaq_f32(vc4, vc5, vt);
205 vp = vmlaq_f32(vc3, vp, vt);
206 vp = vmlaq_f32(vc2, vp, vt);
207 vp = vmlaq_f32(vc1, vp, vt);
208
209 vt = vmulq_f32(vt, vs);
210 const float32x4_t ve = vmlaq_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 vst1q_f32(y, vf); y += 4;
223 }
224 if XNN_UNLIKELY(n != 0) {
225 const float32x4_t vx = vld1q_f32(x);
226
227 const float32x4_t vz = vabsq_f32(vx);
228
229 float32x4_t vn = vmlaq_f32(vmagic_bias, vz, vminus_log2e);
230 const float32x4_t vs = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn), 23));
231 vn = vsubq_f32(vn, vmagic_bias);
232 float32x4_t vt = vmlaq_f32(vz, vn, vln2_hi);
233 vt = vmlaq_f32(vt, vn, vln2_lo);
234
235 float32x4_t vp = vmlaq_f32(vc4, vc5, vt);
236 vp = vmlaq_f32(vc3, vp, vt);
237 vp = vmlaq_f32(vc2, vp, vt);
238 vp = vmlaq_f32(vc1, vp, vt);
239
240 vt = vmulq_f32(vt, vs);
241 const float32x4_t ve = vmlaq_f32(vs, vp, vt);
242 const float32x4_t vd = vaddq_f32(ve, vone);
243
244 float32x4_t vr = vrecpeq_f32(vd);
245 vr = vmulq_f32(vr, vrecpsq_f32(vr, vd));
246 vr = vmulq_f32(vr, vrecpsq_f32(vr, vd));
247
248 float32x4_t vf = vmulq_f32(ve, vr);
249 vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcagtq_f32(vx, vdenorm_cutoff)));
250 const uint32x4_t vm = vcltq_f32(vx, vmovq_n_f32(0.0f));
251 vf = vbslq_f32(vm, vf, vsubq_f32(vone, vf));
252
253 float32x2_t vf_lo = vget_low_f32(vf);
254 if (n & (2 * sizeof(float))) {
255 vst1_f32(y, vf_lo); y += 2;
256 vf_lo = vget_high_f32(vf);
257 }
258 if (n & (1 * sizeof(float))) {
259 vst1_lane_f32(y, vf_lo, 0);
260 }
261 }
262 }
263