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_x24(size_t n,const float * x,float * y,const void * params)18 void xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x24(
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 >= 24 * sizeof(float); n -= 24 * 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 const float32x4_t vxKLMN = vld1q_f32(x); x += 4;
44
45 const float32x4_t vz0123 = vabsq_f32(vx0123);
46 const float32x4_t vz4567 = vabsq_f32(vx4567);
47 const float32x4_t vz89AB = vabsq_f32(vx89AB);
48 const float32x4_t vzCDEF = vabsq_f32(vxCDEF);
49 const float32x4_t vzGHIJ = vabsq_f32(vxGHIJ);
50 const float32x4_t vzKLMN = vabsq_f32(vxKLMN);
51
52 float32x4_t vn0123 = vfmaq_f32(vmagic_bias, vz0123, vminus_log2e);
53 float32x4_t vn4567 = vfmaq_f32(vmagic_bias, vz4567, vminus_log2e);
54 float32x4_t vn89AB = vfmaq_f32(vmagic_bias, vz89AB, vminus_log2e);
55 float32x4_t vnCDEF = vfmaq_f32(vmagic_bias, vzCDEF, vminus_log2e);
56 float32x4_t vnGHIJ = vfmaq_f32(vmagic_bias, vzGHIJ, vminus_log2e);
57 float32x4_t vnKLMN = vfmaq_f32(vmagic_bias, vzKLMN, vminus_log2e);
58
59 const float32x4_t vs0123 = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn0123), 23));
60 const float32x4_t vs4567 = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn4567), 23));
61 const float32x4_t vs89AB = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn89AB), 23));
62 const float32x4_t vsCDEF = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vnCDEF), 23));
63 const float32x4_t vsGHIJ = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vnGHIJ), 23));
64 const float32x4_t vsKLMN = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vnKLMN), 23));
65
66 vn0123 = vsubq_f32(vn0123, vmagic_bias);
67 vn4567 = vsubq_f32(vn4567, vmagic_bias);
68 vn89AB = vsubq_f32(vn89AB, vmagic_bias);
69 vnCDEF = vsubq_f32(vnCDEF, vmagic_bias);
70 vnGHIJ = vsubq_f32(vnGHIJ, vmagic_bias);
71 vnKLMN = vsubq_f32(vnKLMN, vmagic_bias);
72
73 float32x4_t vt0123 = vfmaq_f32(vz0123, vn0123, vln2);
74 float32x4_t vt4567 = vfmaq_f32(vz4567, vn4567, vln2);
75 float32x4_t vt89AB = vfmaq_f32(vz89AB, vn89AB, vln2);
76 float32x4_t vtCDEF = vfmaq_f32(vzCDEF, vnCDEF, vln2);
77 float32x4_t vtGHIJ = vfmaq_f32(vzGHIJ, vnGHIJ, vln2);
78 float32x4_t vtKLMN = vfmaq_f32(vzKLMN, vnKLMN, vln2);
79
80 float32x4_t vp0123 = vfmaq_f32(vc4, vc5, vt0123);
81 float32x4_t vp4567 = vfmaq_f32(vc4, vc5, vt4567);
82 float32x4_t vp89AB = vfmaq_f32(vc4, vc5, vt89AB);
83 float32x4_t vpCDEF = vfmaq_f32(vc4, vc5, vtCDEF);
84 float32x4_t vpGHIJ = vfmaq_f32(vc4, vc5, vtGHIJ);
85 float32x4_t vpKLMN = vfmaq_f32(vc4, vc5, vtKLMN);
86
87 vp0123 = vfmaq_f32(vc3, vp0123, vt0123);
88 vp4567 = vfmaq_f32(vc3, vp4567, vt4567);
89 vp89AB = vfmaq_f32(vc3, vp89AB, vt89AB);
90 vpCDEF = vfmaq_f32(vc3, vpCDEF, vtCDEF);
91 vpGHIJ = vfmaq_f32(vc3, vpGHIJ, vtGHIJ);
92 vpKLMN = vfmaq_f32(vc3, vpKLMN, vtKLMN);
93
94 vp0123 = vfmaq_f32(vc2, vp0123, vt0123);
95 vp4567 = vfmaq_f32(vc2, vp4567, vt4567);
96 vp89AB = vfmaq_f32(vc2, vp89AB, vt89AB);
97 vpCDEF = vfmaq_f32(vc2, vpCDEF, vtCDEF);
98 vpGHIJ = vfmaq_f32(vc2, vpGHIJ, vtGHIJ);
99 vpKLMN = vfmaq_f32(vc2, vpKLMN, vtKLMN);
100
101 vp0123 = vfmaq_f32(vc1, vp0123, vt0123);
102 vp4567 = vfmaq_f32(vc1, vp4567, vt4567);
103 vp89AB = vfmaq_f32(vc1, vp89AB, vt89AB);
104 vpCDEF = vfmaq_f32(vc1, vpCDEF, vtCDEF);
105 vpGHIJ = vfmaq_f32(vc1, vpGHIJ, vtGHIJ);
106 vpKLMN = vfmaq_f32(vc1, vpKLMN, vtKLMN);
107
108 vt0123 = vmulq_f32(vt0123, vs0123);
109 vt4567 = vmulq_f32(vt4567, vs4567);
110 vt89AB = vmulq_f32(vt89AB, vs89AB);
111 vtCDEF = vmulq_f32(vtCDEF, vsCDEF);
112 vtGHIJ = vmulq_f32(vtGHIJ, vsGHIJ);
113 vtKLMN = vmulq_f32(vtKLMN, vsKLMN);
114
115 const float32x4_t ve0123 = vfmaq_f32(vs0123, vp0123, vt0123);
116 const float32x4_t ve4567 = vfmaq_f32(vs4567, vp4567, vt4567);
117 const float32x4_t ve89AB = vfmaq_f32(vs89AB, vp89AB, vt89AB);
118 const float32x4_t veCDEF = vfmaq_f32(vsCDEF, vpCDEF, vtCDEF);
119 const float32x4_t veGHIJ = vfmaq_f32(vsGHIJ, vpGHIJ, vtGHIJ);
120 const float32x4_t veKLMN = vfmaq_f32(vsKLMN, vpKLMN, vtKLMN);
121
122 const float32x4_t vd0123 = vaddq_f32(ve0123, vone);
123 const float32x4_t vd4567 = vaddq_f32(ve4567, vone);
124 const float32x4_t vd89AB = vaddq_f32(ve89AB, vone);
125 const float32x4_t vdCDEF = vaddq_f32(veCDEF, vone);
126 const float32x4_t vdGHIJ = vaddq_f32(veGHIJ, vone);
127 const float32x4_t vdKLMN = vaddq_f32(veKLMN, vone);
128
129 float32x4_t vr0123 = vrecpeq_f32(vd0123);
130 float32x4_t vr4567 = vrecpeq_f32(vd4567);
131 float32x4_t vr89AB = vrecpeq_f32(vd89AB);
132 float32x4_t vrCDEF = vrecpeq_f32(vdCDEF);
133 float32x4_t vrGHIJ = vrecpeq_f32(vdGHIJ);
134 float32x4_t vrKLMN = vrecpeq_f32(vdKLMN);
135
136 vr0123 = vmulq_f32(vr0123, vrecpsq_f32(vr0123, vd0123));
137 vr4567 = vmulq_f32(vr4567, vrecpsq_f32(vr4567, vd4567));
138 vr89AB = vmulq_f32(vr89AB, vrecpsq_f32(vr89AB, vd89AB));
139 vrCDEF = vmulq_f32(vrCDEF, vrecpsq_f32(vrCDEF, vdCDEF));
140 vrGHIJ = vmulq_f32(vrGHIJ, vrecpsq_f32(vrGHIJ, vdGHIJ));
141 vrKLMN = vmulq_f32(vrKLMN, vrecpsq_f32(vrKLMN, vdKLMN));
142
143 vr0123 = vfmaq_f32(vr0123, vr0123, vfmsq_f32(vone, vr0123, vd0123));
144 vr4567 = vfmaq_f32(vr4567, vr4567, vfmsq_f32(vone, vr4567, vd4567));
145 vr89AB = vfmaq_f32(vr89AB, vr89AB, vfmsq_f32(vone, vr89AB, vd89AB));
146 vrCDEF = vfmaq_f32(vrCDEF, vrCDEF, vfmsq_f32(vone, vrCDEF, vdCDEF));
147 vrGHIJ = vfmaq_f32(vrGHIJ, vrGHIJ, vfmsq_f32(vone, vrGHIJ, vdGHIJ));
148 vrKLMN = vfmaq_f32(vrKLMN, vrKLMN, vfmsq_f32(vone, vrKLMN, vdKLMN));
149
150 float32x4_t vf0123 = vmulq_f32(ve0123, vr0123);
151 float32x4_t vf4567 = vmulq_f32(ve4567, vr4567);
152 float32x4_t vf89AB = vmulq_f32(ve89AB, vr89AB);
153 float32x4_t vfCDEF = vmulq_f32(veCDEF, vrCDEF);
154 float32x4_t vfGHIJ = vmulq_f32(veGHIJ, vrGHIJ);
155 float32x4_t vfKLMN = vmulq_f32(veKLMN, vrKLMN);
156
157 vf0123 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf0123), vcagtq_f32(vx0123, vdenorm_cutoff)));
158 vf4567 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf4567), vcagtq_f32(vx4567, vdenorm_cutoff)));
159 vf89AB = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf89AB), vcagtq_f32(vx89AB, vdenorm_cutoff)));
160 vfCDEF = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vfCDEF), vcagtq_f32(vxCDEF, vdenorm_cutoff)));
161 vfGHIJ = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vfGHIJ), vcagtq_f32(vxGHIJ, vdenorm_cutoff)));
162 vfKLMN = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vfKLMN), vcagtq_f32(vxKLMN, vdenorm_cutoff)));
163
164 const uint32x4_t vm0123 = vcltq_f32(vx0123, vmovq_n_f32(0.0f));
165 const uint32x4_t vm4567 = vcltq_f32(vx4567, vmovq_n_f32(0.0f));
166 const uint32x4_t vm89AB = vcltq_f32(vx89AB, vmovq_n_f32(0.0f));
167 const uint32x4_t vmCDEF = vcltq_f32(vxCDEF, vmovq_n_f32(0.0f));
168 const uint32x4_t vmGHIJ = vcltq_f32(vxGHIJ, vmovq_n_f32(0.0f));
169 const uint32x4_t vmKLMN = vcltq_f32(vxKLMN, vmovq_n_f32(0.0f));
170
171 vf0123 = vbslq_f32(vm0123, vf0123, vsubq_f32(vone, vf0123));
172 vf4567 = vbslq_f32(vm4567, vf4567, vsubq_f32(vone, vf4567));
173 vf89AB = vbslq_f32(vm89AB, vf89AB, vsubq_f32(vone, vf89AB));
174 vfCDEF = vbslq_f32(vmCDEF, vfCDEF, vsubq_f32(vone, vfCDEF));
175 vfGHIJ = vbslq_f32(vmGHIJ, vfGHIJ, vsubq_f32(vone, vfGHIJ));
176 vfKLMN = vbslq_f32(vmKLMN, vfKLMN, vsubq_f32(vone, vfKLMN));
177
178 vst1q_f32(y, vf0123); y += 4;
179 vst1q_f32(y, vf4567); y += 4;
180 vst1q_f32(y, vf89AB); y += 4;
181 vst1q_f32(y, vfCDEF); y += 4;
182 vst1q_f32(y, vfGHIJ); y += 4;
183 vst1q_f32(y, vfKLMN); y += 4;
184 }
185 for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
186 const float32x4_t vx = vld1q_f32(x); x += 4;
187
188 const float32x4_t vz = vabsq_f32(vx);
189
190 float32x4_t vn = vfmaq_f32(vmagic_bias, vz, vminus_log2e);
191 const float32x4_t vs = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn), 23));
192 vn = vsubq_f32(vn, vmagic_bias);
193 float32x4_t vt = vfmaq_f32(vz, vn, vln2);
194
195 float32x4_t vp = vfmaq_f32(vc4, vc5, vt);
196 vp = vfmaq_f32(vc3, vp, vt);
197 vp = vfmaq_f32(vc2, vp, vt);
198 vp = vfmaq_f32(vc1, vp, vt);
199
200 vt = vmulq_f32(vt, vs);
201 const float32x4_t ve = vfmaq_f32(vs, vp, vt);
202 const float32x4_t vd = vaddq_f32(ve, vone);
203
204 float32x4_t vr = vrecpeq_f32(vd);
205 vr = vmulq_f32(vr, vrecpsq_f32(vr, vd));
206 vr = vfmaq_f32(vr, vr, vfmsq_f32(vone, vr, vd));
207
208 float32x4_t vf = vmulq_f32(ve, vr);
209 vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcagtq_f32(vx, vdenorm_cutoff)));
210 const uint32x4_t vm = vcltq_f32(vx, vmovq_n_f32(0.0f));
211 vf = vbslq_f32(vm, vf, vsubq_f32(vone, vf));
212
213 vst1q_f32(y, vf); y += 4;
214 }
215 if XNN_UNLIKELY(n != 0) {
216 const float32x4_t vx = vld1q_f32(x);
217
218 const float32x4_t vz = vabsq_f32(vx);
219
220 float32x4_t vn = vfmaq_f32(vmagic_bias, vz, vminus_log2e);
221 const float32x4_t vs = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn), 23));
222 vn = vsubq_f32(vn, vmagic_bias);
223 float32x4_t vt = vfmaq_f32(vz, vn, vln2);
224
225 float32x4_t vp = vfmaq_f32(vc4, vc5, vt);
226 vp = vfmaq_f32(vc3, vp, vt);
227 vp = vfmaq_f32(vc2, vp, vt);
228 vp = vfmaq_f32(vc1, vp, vt);
229
230 vt = vmulq_f32(vt, vs);
231 const float32x4_t ve = vfmaq_f32(vs, vp, vt);
232 const float32x4_t vd = vaddq_f32(ve, vone);
233
234 float32x4_t vr = vrecpeq_f32(vd);
235 vr = vmulq_f32(vr, vrecpsq_f32(vr, vd));
236 vr = vfmaq_f32(vr, vr, vfmsq_f32(vone, vr, vd));
237
238 float32x4_t vf = vmulq_f32(ve, vr);
239 vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcagtq_f32(vx, vdenorm_cutoff)));
240 const uint32x4_t vm = vcltq_f32(vx, vmovq_n_f32(0.0f));
241 vf = vbslq_f32(vm, vf, vsubq_f32(vone, vf));
242
243 float32x2_t vf_lo = vget_low_f32(vf);
244 if (n & (2 * sizeof(float))) {
245 vst1_f32(y, vf_lo); y += 2;
246 vf_lo = vget_high_f32(vf);
247 }
248 if (n & (1 * sizeof(float))) {
249 vst1_lane_f32(y, vf_lo, 0);
250 }
251 }
252 }
253