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_nr1recps_x64(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_nr1recps_x64(
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 >= 64 * sizeof(__fp16); batch -= 64 * 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 const float16x8_t vx7 = vld1q_f16(i); i += 8;
46
47 const float16x8_t vz0 = vabsq_f16(vx0);
48 const float16x8_t vz1 = vabsq_f16(vx1);
49 const float16x8_t vz2 = vabsq_f16(vx2);
50 const float16x8_t vz3 = vabsq_f16(vx3);
51 const float16x8_t vz4 = vabsq_f16(vx4);
52 const float16x8_t vz5 = vabsq_f16(vx5);
53 const float16x8_t vz6 = vabsq_f16(vx6);
54 const float16x8_t vz7 = vabsq_f16(vx7);
55
56 float16x8_t vn0 = vfmaq_f16(vmagic_bias, vz0, vminus_log2e);
57 float16x8_t vn1 = vfmaq_f16(vmagic_bias, vz1, vminus_log2e);
58 float16x8_t vn2 = vfmaq_f16(vmagic_bias, vz2, vminus_log2e);
59 float16x8_t vn3 = vfmaq_f16(vmagic_bias, vz3, vminus_log2e);
60 float16x8_t vn4 = vfmaq_f16(vmagic_bias, vz4, vminus_log2e);
61 float16x8_t vn5 = vfmaq_f16(vmagic_bias, vz5, vminus_log2e);
62 float16x8_t vn6 = vfmaq_f16(vmagic_bias, vz6, vminus_log2e);
63 float16x8_t vn7 = vfmaq_f16(vmagic_bias, vz7, vminus_log2e);
64
65 const float16x8_t vs0 = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn0), 10));
66 const float16x8_t vs1 = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn1), 10));
67 const float16x8_t vs2 = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn2), 10));
68 const float16x8_t vs3 = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn3), 10));
69 const float16x8_t vs4 = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn4), 10));
70 const float16x8_t vs5 = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn5), 10));
71 const float16x8_t vs6 = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn6), 10));
72 const float16x8_t vs7 = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn7), 10));
73
74 vn0 = vsubq_f16(vn0, vmagic_bias);
75 vn1 = vsubq_f16(vn1, vmagic_bias);
76 vn2 = vsubq_f16(vn2, vmagic_bias);
77 vn3 = vsubq_f16(vn3, vmagic_bias);
78 vn4 = vsubq_f16(vn4, vmagic_bias);
79 vn5 = vsubq_f16(vn5, vmagic_bias);
80 vn6 = vsubq_f16(vn6, vmagic_bias);
81 vn7 = vsubq_f16(vn7, vmagic_bias);
82
83 float16x8_t vt0 = vfmaq_f16(vz0, vn0, vln2_hi);
84 float16x8_t vt1 = vfmaq_f16(vz1, vn1, vln2_hi);
85 float16x8_t vt2 = vfmaq_f16(vz2, vn2, vln2_hi);
86 float16x8_t vt3 = vfmaq_f16(vz3, vn3, vln2_hi);
87 float16x8_t vt4 = vfmaq_f16(vz4, vn4, vln2_hi);
88 float16x8_t vt5 = vfmaq_f16(vz5, vn5, vln2_hi);
89 float16x8_t vt6 = vfmaq_f16(vz6, vn6, vln2_hi);
90 float16x8_t vt7 = vfmaq_f16(vz7, vn7, vln2_hi);
91
92 vt0 = vfmaq_f16(vt0, vn0, vln2_lo);
93 vt1 = vfmaq_f16(vt1, vn1, vln2_lo);
94 vt2 = vfmaq_f16(vt2, vn2, vln2_lo);
95 vt3 = vfmaq_f16(vt3, vn3, vln2_lo);
96 vt4 = vfmaq_f16(vt4, vn4, vln2_lo);
97 vt5 = vfmaq_f16(vt5, vn5, vln2_lo);
98 vt6 = vfmaq_f16(vt6, vn6, vln2_lo);
99 vt7 = vfmaq_f16(vt7, vn7, vln2_lo);
100
101 const float16x8_t vp0 = vfmaq_f16(vc1, vc2, vt0);
102 const float16x8_t vp1 = vfmaq_f16(vc1, vc2, vt1);
103 const float16x8_t vp2 = vfmaq_f16(vc1, vc2, vt2);
104 const float16x8_t vp3 = vfmaq_f16(vc1, vc2, vt3);
105 const float16x8_t vp4 = vfmaq_f16(vc1, vc2, vt4);
106 const float16x8_t vp5 = vfmaq_f16(vc1, vc2, vt5);
107 const float16x8_t vp6 = vfmaq_f16(vc1, vc2, vt6);
108 const float16x8_t vp7 = vfmaq_f16(vc1, vc2, vt7);
109
110 vt0 = vmulq_f16(vt0, vs0);
111 vt1 = vmulq_f16(vt1, vs1);
112 vt2 = vmulq_f16(vt2, vs2);
113 vt3 = vmulq_f16(vt3, vs3);
114 vt4 = vmulq_f16(vt4, vs4);
115 vt5 = vmulq_f16(vt5, vs5);
116 vt6 = vmulq_f16(vt6, vs6);
117 vt7 = vmulq_f16(vt7, vs7);
118
119 const float16x8_t ve0 = vfmaq_f16(vs0, vp0, vt0);
120 const float16x8_t ve1 = vfmaq_f16(vs1, vp1, vt1);
121 const float16x8_t ve2 = vfmaq_f16(vs2, vp2, vt2);
122 const float16x8_t ve3 = vfmaq_f16(vs3, vp3, vt3);
123 const float16x8_t ve4 = vfmaq_f16(vs4, vp4, vt4);
124 const float16x8_t ve5 = vfmaq_f16(vs5, vp5, vt5);
125 const float16x8_t ve6 = vfmaq_f16(vs6, vp6, vt6);
126 const float16x8_t ve7 = vfmaq_f16(vs7, vp7, vt7);
127
128 const float16x8_t vd0 = vaddq_f16(ve0, vone);
129 const float16x8_t vd1 = vaddq_f16(ve1, vone);
130 const float16x8_t vd2 = vaddq_f16(ve2, vone);
131 const float16x8_t vd3 = vaddq_f16(ve3, vone);
132 const float16x8_t vd4 = vaddq_f16(ve4, vone);
133 const float16x8_t vd5 = vaddq_f16(ve5, vone);
134 const float16x8_t vd6 = vaddq_f16(ve6, vone);
135 const float16x8_t vd7 = vaddq_f16(ve7, vone);
136
137 float16x8_t vr0 = vrecpeq_f16(vd0);
138 float16x8_t vr1 = vrecpeq_f16(vd1);
139 float16x8_t vr2 = vrecpeq_f16(vd2);
140 float16x8_t vr3 = vrecpeq_f16(vd3);
141 float16x8_t vr4 = vrecpeq_f16(vd4);
142 float16x8_t vr5 = vrecpeq_f16(vd5);
143 float16x8_t vr6 = vrecpeq_f16(vd6);
144 float16x8_t vr7 = vrecpeq_f16(vd7);
145
146 const float16x8_t vadj0 = vrecpsq_f16(vr0, vd0);
147 const float16x8_t vadj1 = vrecpsq_f16(vr1, vd1);
148 const float16x8_t vadj2 = vrecpsq_f16(vr2, vd2);
149 const float16x8_t vadj3 = vrecpsq_f16(vr3, vd3);
150 const float16x8_t vadj4 = vrecpsq_f16(vr4, vd4);
151 const float16x8_t vadj5 = vrecpsq_f16(vr5, vd5);
152 const float16x8_t vadj6 = vrecpsq_f16(vr6, vd6);
153 const float16x8_t vadj7 = vrecpsq_f16(vr7, vd7);
154
155 vr0 = vmulq_f16(vr0, vadj0);
156 vr1 = vmulq_f16(vr1, vadj1);
157 vr2 = vmulq_f16(vr2, vadj2);
158 vr3 = vmulq_f16(vr3, vadj3);
159 vr4 = vmulq_f16(vr4, vadj4);
160 vr5 = vmulq_f16(vr5, vadj5);
161 vr6 = vmulq_f16(vr6, vadj6);
162 vr7 = vmulq_f16(vr7, vadj7);
163
164 float16x8_t vf0 = vmulq_f16(ve0, vr0);
165 float16x8_t vf1 = vmulq_f16(ve1, vr1);
166 float16x8_t vf2 = vmulq_f16(ve2, vr2);
167 float16x8_t vf3 = vmulq_f16(ve3, vr3);
168 float16x8_t vf4 = vmulq_f16(ve4, vr4);
169 float16x8_t vf5 = vmulq_f16(ve5, vr5);
170 float16x8_t vf6 = vmulq_f16(ve6, vr6);
171 float16x8_t vf7 = vmulq_f16(ve7, vr7);
172
173 vf0 = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf0), vcagtq_f16(vx0, vdenorm_cutoff)));
174 vf1 = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf1), vcagtq_f16(vx1, vdenorm_cutoff)));
175 vf2 = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf2), vcagtq_f16(vx2, vdenorm_cutoff)));
176 vf3 = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf3), vcagtq_f16(vx3, vdenorm_cutoff)));
177 vf4 = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf4), vcagtq_f16(vx4, vdenorm_cutoff)));
178 vf5 = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf5), vcagtq_f16(vx5, vdenorm_cutoff)));
179 vf6 = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf6), vcagtq_f16(vx6, vdenorm_cutoff)));
180 vf7 = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf7), vcagtq_f16(vx7, vdenorm_cutoff)));
181
182 const uint16x8_t vm0 = vcltq_f16(vx0, vmovq_n_f16(0.0f));
183 const uint16x8_t vm1 = vcltq_f16(vx1, vmovq_n_f16(0.0f));
184 const uint16x8_t vm2 = vcltq_f16(vx2, vmovq_n_f16(0.0f));
185 const uint16x8_t vm3 = vcltq_f16(vx3, vmovq_n_f16(0.0f));
186 const uint16x8_t vm4 = vcltq_f16(vx4, vmovq_n_f16(0.0f));
187 const uint16x8_t vm5 = vcltq_f16(vx5, vmovq_n_f16(0.0f));
188 const uint16x8_t vm6 = vcltq_f16(vx6, vmovq_n_f16(0.0f));
189 const uint16x8_t vm7 = vcltq_f16(vx7, vmovq_n_f16(0.0f));
190
191 vf0 = vbslq_f16(vm0, vf0, vsubq_f16(vone, vf0));
192 vf1 = vbslq_f16(vm1, vf1, vsubq_f16(vone, vf1));
193 vf2 = vbslq_f16(vm2, vf2, vsubq_f16(vone, vf2));
194 vf3 = vbslq_f16(vm3, vf3, vsubq_f16(vone, vf3));
195 vf4 = vbslq_f16(vm4, vf4, vsubq_f16(vone, vf4));
196 vf5 = vbslq_f16(vm5, vf5, vsubq_f16(vone, vf5));
197 vf6 = vbslq_f16(vm6, vf6, vsubq_f16(vone, vf6));
198 vf7 = vbslq_f16(vm7, vf7, vsubq_f16(vone, vf7));
199
200 vst1q_f16(o, vf0); o += 8;
201 vst1q_f16(o, vf1); o += 8;
202 vst1q_f16(o, vf2); o += 8;
203 vst1q_f16(o, vf3); o += 8;
204 vst1q_f16(o, vf4); o += 8;
205 vst1q_f16(o, vf5); o += 8;
206 vst1q_f16(o, vf6); o += 8;
207 vst1q_f16(o, vf7); o += 8;
208 }
209 for (; batch >= 8 * sizeof(__fp16); batch -= 8 * sizeof(__fp16)) {
210 const float16x8_t vx = vld1q_f16(i); i += 8;
211
212 const float16x8_t vz = vabsq_f16(vx);
213
214 float16x8_t vn = vfmaq_f16(vmagic_bias, vz, vminus_log2e);
215 const float16x8_t vs = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn), 10));
216 vn = vsubq_f16(vn, vmagic_bias);
217
218 float16x8_t vt = vfmaq_f16(vz, vn, vln2_hi);
219 vt = vfmaq_f16(vt, vn, vln2_lo);
220
221 const float16x8_t vp = vfmaq_f16(vc1, vc2, vt);
222 vt = vmulq_f16(vt, vs);
223 const float16x8_t ve = vfmaq_f16(vs, vp, vt);
224 const float16x8_t vd = vaddq_f16(ve, vone);
225
226 float16x8_t vr = vrecpeq_f16(vd);
227 const float16x8_t vadj = vrecpsq_f16(vr, vd);
228 vr = vmulq_f16(vr, vadj);
229
230 float16x8_t vf = vmulq_f16(ve, vr);
231 vf = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf), vcagtq_f16(vx, vdenorm_cutoff)));
232 const uint16x8_t vm = vcltq_f16(vx, vmovq_n_f16(0.0f));
233 vf = vbslq_f16(vm, vf, vsubq_f16(vone, vf));
234
235 vst1q_f16(o, vf); o += 8;
236 }
237 if XNN_UNLIKELY(batch != 0) {
238 const float16x8_t vx = vld1q_f16(i);
239
240 const float16x8_t vz = vabsq_f16(vx);
241
242 float16x8_t vn = vfmaq_f16(vmagic_bias, vz, vminus_log2e);
243 const float16x8_t vs = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn), 10));
244 vn = vsubq_f16(vn, vmagic_bias);
245
246 float16x8_t vt = vfmaq_f16(vz, vn, vln2_hi);
247 vt = vfmaq_f16(vt, vn, vln2_lo);
248
249 const float16x8_t vp = vfmaq_f16(vc1, vc2, vt);
250 vt = vmulq_f16(vt, vs);
251 const float16x8_t ve = vfmaq_f16(vs, vp, vt);
252 const float16x8_t vd = vaddq_f16(ve, vone);
253
254 float16x8_t vr = vrecpeq_f16(vd);
255 const float16x8_t vadj = vrecpsq_f16(vr, vd);
256 vr = vmulq_f16(vr, vadj);
257
258 float16x8_t vf = vmulq_f16(ve, vr);
259 vf = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf), vcagtq_f16(vx, vdenorm_cutoff)));
260 const uint16x8_t vm = vcltq_f16(vx, vmovq_n_f16(0.0f));
261 vf = vbslq_f16(vm, vf, vsubq_f16(vone, vf));
262
263 float16x4_t vf_lo = vget_low_f16(vf);
264 if (batch & (4 * sizeof(__fp16))) {
265 vst1_f16(o, vf_lo); o += 4;
266 vf_lo = vget_high_f16(vf);
267 }
268 if (batch & (2 * sizeof(__fp16))) {
269 vst1_f16(o, vf_lo); o += 2;
270 vf_lo = vext_f16(vf_lo, vf_lo, 2);
271 }
272 if (batch & (1 * sizeof(__fp16))) {
273 vst1_lane_f16(o, vf_lo, 0);
274 }
275 }
276 }
277