1 // Auto-generated file. Do not edit!
2 // Template: src/f16-vsigmoid/neonfp16arith.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2022 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_f16_vsigmoid_ukernel__neonfp16arith_rr2_p2_nr1fma_x56(size_t batch,const void * input,void * output,const union xnn_f16_sigmoid_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_f16_vsigmoid_ukernel__neonfp16arith_rr2_p2_nr1fma_x56(
19 size_t batch,
20 const void* input,
21 void* output,
22 const union xnn_f16_sigmoid_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
23 {
24 assert(batch % sizeof(__fp16) == 0);
25
26 const float16x8_t vmagic_bias = vreinterpretq_f16_u16(vld1q_dup_u16(¶ms->neonfp16arith_rr2_p2.magic_bias));
27 const float16x8_t vminus_log2e = vreinterpretq_f16_u16(vld1q_dup_u16(¶ms->neonfp16arith_rr2_p2.minus_log2e));
28 const float16x8_t vln2_hi = vreinterpretq_f16_u16(vld1q_dup_u16(¶ms->neonfp16arith_rr2_p2.ln2_hi));
29 const float16x8_t vln2_lo = vreinterpretq_f16_u16(vld1q_dup_u16(¶ms->neonfp16arith_rr2_p2.ln2_lo));
30 const float16x8_t vc2 = vreinterpretq_f16_u16(vld1q_dup_u16(¶ms->neonfp16arith_rr2_p2.c2));
31 const float16x8_t vc1 = vreinterpretq_f16_u16(vld1q_dup_u16(¶ms->neonfp16arith_rr2_p2.c1));
32 const float16x8_t vone = vmovq_n_f16(1.0f);
33 const float16x8_t vdenorm_cutoff = vreinterpretq_f16_u16(vld1q_dup_u16(¶ms->neonfp16arith_rr2_p2.denorm_cutoff));
34
35 const __fp16* i = (const __fp16*) input;
36 __fp16* o = (__fp16*) output;
37 for (; batch >= 56 * sizeof(__fp16); batch -= 56 * sizeof(__fp16)) {
38 const float16x8_t vx0 = vld1q_f16(i); i += 8;
39 const float16x8_t vx1 = vld1q_f16(i); i += 8;
40 const float16x8_t vx2 = vld1q_f16(i); i += 8;
41 const float16x8_t vx3 = vld1q_f16(i); i += 8;
42 const float16x8_t vx4 = vld1q_f16(i); i += 8;
43 const float16x8_t vx5 = vld1q_f16(i); i += 8;
44 const float16x8_t vx6 = vld1q_f16(i); i += 8;
45
46 const float16x8_t vz0 = vabsq_f16(vx0);
47 const float16x8_t vz1 = vabsq_f16(vx1);
48 const float16x8_t vz2 = vabsq_f16(vx2);
49 const float16x8_t vz3 = vabsq_f16(vx3);
50 const float16x8_t vz4 = vabsq_f16(vx4);
51 const float16x8_t vz5 = vabsq_f16(vx5);
52 const float16x8_t vz6 = vabsq_f16(vx6);
53
54 float16x8_t vn0 = vfmaq_f16(vmagic_bias, vz0, vminus_log2e);
55 float16x8_t vn1 = vfmaq_f16(vmagic_bias, vz1, vminus_log2e);
56 float16x8_t vn2 = vfmaq_f16(vmagic_bias, vz2, vminus_log2e);
57 float16x8_t vn3 = vfmaq_f16(vmagic_bias, vz3, vminus_log2e);
58 float16x8_t vn4 = vfmaq_f16(vmagic_bias, vz4, vminus_log2e);
59 float16x8_t vn5 = vfmaq_f16(vmagic_bias, vz5, vminus_log2e);
60 float16x8_t vn6 = vfmaq_f16(vmagic_bias, vz6, vminus_log2e);
61
62 const float16x8_t vs0 = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn0), 10));
63 const float16x8_t vs1 = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn1), 10));
64 const float16x8_t vs2 = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn2), 10));
65 const float16x8_t vs3 = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn3), 10));
66 const float16x8_t vs4 = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn4), 10));
67 const float16x8_t vs5 = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn5), 10));
68 const float16x8_t vs6 = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn6), 10));
69
70 vn0 = vsubq_f16(vn0, vmagic_bias);
71 vn1 = vsubq_f16(vn1, vmagic_bias);
72 vn2 = vsubq_f16(vn2, vmagic_bias);
73 vn3 = vsubq_f16(vn3, vmagic_bias);
74 vn4 = vsubq_f16(vn4, vmagic_bias);
75 vn5 = vsubq_f16(vn5, vmagic_bias);
76 vn6 = vsubq_f16(vn6, vmagic_bias);
77
78 float16x8_t vt0 = vfmaq_f16(vz0, vn0, vln2_hi);
79 float16x8_t vt1 = vfmaq_f16(vz1, vn1, vln2_hi);
80 float16x8_t vt2 = vfmaq_f16(vz2, vn2, vln2_hi);
81 float16x8_t vt3 = vfmaq_f16(vz3, vn3, vln2_hi);
82 float16x8_t vt4 = vfmaq_f16(vz4, vn4, vln2_hi);
83 float16x8_t vt5 = vfmaq_f16(vz5, vn5, vln2_hi);
84 float16x8_t vt6 = vfmaq_f16(vz6, vn6, vln2_hi);
85
86 vt0 = vfmaq_f16(vt0, vn0, vln2_lo);
87 vt1 = vfmaq_f16(vt1, vn1, vln2_lo);
88 vt2 = vfmaq_f16(vt2, vn2, vln2_lo);
89 vt3 = vfmaq_f16(vt3, vn3, vln2_lo);
90 vt4 = vfmaq_f16(vt4, vn4, vln2_lo);
91 vt5 = vfmaq_f16(vt5, vn5, vln2_lo);
92 vt6 = vfmaq_f16(vt6, vn6, vln2_lo);
93
94 const float16x8_t vp0 = vfmaq_f16(vc1, vc2, vt0);
95 const float16x8_t vp1 = vfmaq_f16(vc1, vc2, vt1);
96 const float16x8_t vp2 = vfmaq_f16(vc1, vc2, vt2);
97 const float16x8_t vp3 = vfmaq_f16(vc1, vc2, vt3);
98 const float16x8_t vp4 = vfmaq_f16(vc1, vc2, vt4);
99 const float16x8_t vp5 = vfmaq_f16(vc1, vc2, vt5);
100 const float16x8_t vp6 = vfmaq_f16(vc1, vc2, vt6);
101
102 vt0 = vmulq_f16(vt0, vs0);
103 vt1 = vmulq_f16(vt1, vs1);
104 vt2 = vmulq_f16(vt2, vs2);
105 vt3 = vmulq_f16(vt3, vs3);
106 vt4 = vmulq_f16(vt4, vs4);
107 vt5 = vmulq_f16(vt5, vs5);
108 vt6 = vmulq_f16(vt6, vs6);
109
110 const float16x8_t ve0 = vfmaq_f16(vs0, vp0, vt0);
111 const float16x8_t ve1 = vfmaq_f16(vs1, vp1, vt1);
112 const float16x8_t ve2 = vfmaq_f16(vs2, vp2, vt2);
113 const float16x8_t ve3 = vfmaq_f16(vs3, vp3, vt3);
114 const float16x8_t ve4 = vfmaq_f16(vs4, vp4, vt4);
115 const float16x8_t ve5 = vfmaq_f16(vs5, vp5, vt5);
116 const float16x8_t ve6 = vfmaq_f16(vs6, vp6, vt6);
117
118 const float16x8_t vd0 = vaddq_f16(ve0, vone);
119 const float16x8_t vd1 = vaddq_f16(ve1, vone);
120 const float16x8_t vd2 = vaddq_f16(ve2, vone);
121 const float16x8_t vd3 = vaddq_f16(ve3, vone);
122 const float16x8_t vd4 = vaddq_f16(ve4, vone);
123 const float16x8_t vd5 = vaddq_f16(ve5, vone);
124 const float16x8_t vd6 = vaddq_f16(ve6, vone);
125
126 float16x8_t vr0 = vrecpeq_f16(vd0);
127 float16x8_t vr1 = vrecpeq_f16(vd1);
128 float16x8_t vr2 = vrecpeq_f16(vd2);
129 float16x8_t vr3 = vrecpeq_f16(vd3);
130 float16x8_t vr4 = vrecpeq_f16(vd4);
131 float16x8_t vr5 = vrecpeq_f16(vd5);
132 float16x8_t vr6 = vrecpeq_f16(vd6);
133
134 const float16x8_t vadj0 = vfmsq_f16(vone, vr0, vd0);
135 const float16x8_t vadj1 = vfmsq_f16(vone, vr1, vd1);
136 const float16x8_t vadj2 = vfmsq_f16(vone, vr2, vd2);
137 const float16x8_t vadj3 = vfmsq_f16(vone, vr3, vd3);
138 const float16x8_t vadj4 = vfmsq_f16(vone, vr4, vd4);
139 const float16x8_t vadj5 = vfmsq_f16(vone, vr5, vd5);
140 const float16x8_t vadj6 = vfmsq_f16(vone, vr6, vd6);
141
142 vr0 = vfmaq_f16(vr0, vr0, vadj0);
143 vr1 = vfmaq_f16(vr1, vr1, vadj1);
144 vr2 = vfmaq_f16(vr2, vr2, vadj2);
145 vr3 = vfmaq_f16(vr3, vr3, vadj3);
146 vr4 = vfmaq_f16(vr4, vr4, vadj4);
147 vr5 = vfmaq_f16(vr5, vr5, vadj5);
148 vr6 = vfmaq_f16(vr6, vr6, vadj6);
149
150 float16x8_t vf0 = vmulq_f16(ve0, vr0);
151 float16x8_t vf1 = vmulq_f16(ve1, vr1);
152 float16x8_t vf2 = vmulq_f16(ve2, vr2);
153 float16x8_t vf3 = vmulq_f16(ve3, vr3);
154 float16x8_t vf4 = vmulq_f16(ve4, vr4);
155 float16x8_t vf5 = vmulq_f16(ve5, vr5);
156 float16x8_t vf6 = vmulq_f16(ve6, vr6);
157
158 vf0 = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf0), vcagtq_f16(vx0, vdenorm_cutoff)));
159 vf1 = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf1), vcagtq_f16(vx1, vdenorm_cutoff)));
160 vf2 = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf2), vcagtq_f16(vx2, vdenorm_cutoff)));
161 vf3 = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf3), vcagtq_f16(vx3, vdenorm_cutoff)));
162 vf4 = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf4), vcagtq_f16(vx4, vdenorm_cutoff)));
163 vf5 = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf5), vcagtq_f16(vx5, vdenorm_cutoff)));
164 vf6 = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf6), vcagtq_f16(vx6, vdenorm_cutoff)));
165
166 const uint16x8_t vm0 = vcltq_f16(vx0, vmovq_n_f16(0.0f));
167 const uint16x8_t vm1 = vcltq_f16(vx1, vmovq_n_f16(0.0f));
168 const uint16x8_t vm2 = vcltq_f16(vx2, vmovq_n_f16(0.0f));
169 const uint16x8_t vm3 = vcltq_f16(vx3, vmovq_n_f16(0.0f));
170 const uint16x8_t vm4 = vcltq_f16(vx4, vmovq_n_f16(0.0f));
171 const uint16x8_t vm5 = vcltq_f16(vx5, vmovq_n_f16(0.0f));
172 const uint16x8_t vm6 = vcltq_f16(vx6, vmovq_n_f16(0.0f));
173
174 vf0 = vbslq_f16(vm0, vf0, vsubq_f16(vone, vf0));
175 vf1 = vbslq_f16(vm1, vf1, vsubq_f16(vone, vf1));
176 vf2 = vbslq_f16(vm2, vf2, vsubq_f16(vone, vf2));
177 vf3 = vbslq_f16(vm3, vf3, vsubq_f16(vone, vf3));
178 vf4 = vbslq_f16(vm4, vf4, vsubq_f16(vone, vf4));
179 vf5 = vbslq_f16(vm5, vf5, vsubq_f16(vone, vf5));
180 vf6 = vbslq_f16(vm6, vf6, vsubq_f16(vone, vf6));
181
182 vst1q_f16(o, vf0); o += 8;
183 vst1q_f16(o, vf1); o += 8;
184 vst1q_f16(o, vf2); o += 8;
185 vst1q_f16(o, vf3); o += 8;
186 vst1q_f16(o, vf4); o += 8;
187 vst1q_f16(o, vf5); o += 8;
188 vst1q_f16(o, vf6); o += 8;
189 }
190 for (; batch >= 8 * sizeof(__fp16); batch -= 8 * sizeof(__fp16)) {
191 const float16x8_t vx = vld1q_f16(i); i += 8;
192
193 const float16x8_t vz = vabsq_f16(vx);
194
195 float16x8_t vn = vfmaq_f16(vmagic_bias, vz, vminus_log2e);
196 const float16x8_t vs = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn), 10));
197 vn = vsubq_f16(vn, vmagic_bias);
198
199 float16x8_t vt = vfmaq_f16(vz, vn, vln2_hi);
200 vt = vfmaq_f16(vt, vn, vln2_lo);
201
202 const float16x8_t vp = vfmaq_f16(vc1, vc2, vt);
203 vt = vmulq_f16(vt, vs);
204 const float16x8_t ve = vfmaq_f16(vs, vp, vt);
205 const float16x8_t vd = vaddq_f16(ve, vone);
206
207 float16x8_t vr = vrecpeq_f16(vd);
208 const float16x8_t vadj = vfmsq_f16(vone, vr, vd);
209 vr = vfmaq_f16(vr, vr, vadj);
210
211 float16x8_t vf = vmulq_f16(ve, vr);
212 vf = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf), vcagtq_f16(vx, vdenorm_cutoff)));
213 const uint16x8_t vm = vcltq_f16(vx, vmovq_n_f16(0.0f));
214 vf = vbslq_f16(vm, vf, vsubq_f16(vone, vf));
215
216 vst1q_f16(o, vf); o += 8;
217 }
218 if XNN_UNLIKELY(batch != 0) {
219 const float16x8_t vx = vld1q_f16(i);
220
221 const float16x8_t vz = vabsq_f16(vx);
222
223 float16x8_t vn = vfmaq_f16(vmagic_bias, vz, vminus_log2e);
224 const float16x8_t vs = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn), 10));
225 vn = vsubq_f16(vn, vmagic_bias);
226
227 float16x8_t vt = vfmaq_f16(vz, vn, vln2_hi);
228 vt = vfmaq_f16(vt, vn, vln2_lo);
229
230 const float16x8_t vp = vfmaq_f16(vc1, vc2, vt);
231 vt = vmulq_f16(vt, vs);
232 const float16x8_t ve = vfmaq_f16(vs, vp, vt);
233 const float16x8_t vd = vaddq_f16(ve, vone);
234
235 float16x8_t vr = vrecpeq_f16(vd);
236 const float16x8_t vadj = vfmsq_f16(vone, vr, vd);
237 vr = vfmaq_f16(vr, vr, vadj);
238
239 float16x8_t vf = vmulq_f16(ve, vr);
240 vf = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf), vcagtq_f16(vx, vdenorm_cutoff)));
241 const uint16x8_t vm = vcltq_f16(vx, vmovq_n_f16(0.0f));
242 vf = vbslq_f16(vm, vf, vsubq_f16(vone, vf));
243
244 float16x4_t vf_lo = vget_low_f16(vf);
245 if (batch & (4 * sizeof(__fp16))) {
246 vst1_f16(o, vf_lo); o += 4;
247 vf_lo = vget_high_f16(vf);
248 }
249 if (batch & (2 * sizeof(__fp16))) {
250 vst1_f16(o, vf_lo); o += 2;
251 vf_lo = vext_f16(vf_lo, vf_lo, 2);
252 }
253 if (batch & (1 * sizeof(__fp16))) {
254 vst1_lane_f16(o, vf_lo, 0);
255 }
256 }
257 }
258