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_x48(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_x48(
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 >= 48 * sizeof(__fp16); batch -= 48 * 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
45 const float16x8_t vz0 = vabsq_f16(vx0);
46 const float16x8_t vz1 = vabsq_f16(vx1);
47 const float16x8_t vz2 = vabsq_f16(vx2);
48 const float16x8_t vz3 = vabsq_f16(vx3);
49 const float16x8_t vz4 = vabsq_f16(vx4);
50 const float16x8_t vz5 = vabsq_f16(vx5);
51
52 float16x8_t vn0 = vfmaq_f16(vmagic_bias, vz0, vminus_log2e);
53 float16x8_t vn1 = vfmaq_f16(vmagic_bias, vz1, vminus_log2e);
54 float16x8_t vn2 = vfmaq_f16(vmagic_bias, vz2, vminus_log2e);
55 float16x8_t vn3 = vfmaq_f16(vmagic_bias, vz3, vminus_log2e);
56 float16x8_t vn4 = vfmaq_f16(vmagic_bias, vz4, vminus_log2e);
57 float16x8_t vn5 = vfmaq_f16(vmagic_bias, vz5, vminus_log2e);
58
59 const float16x8_t vs0 = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn0), 10));
60 const float16x8_t vs1 = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn1), 10));
61 const float16x8_t vs2 = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn2), 10));
62 const float16x8_t vs3 = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn3), 10));
63 const float16x8_t vs4 = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn4), 10));
64 const float16x8_t vs5 = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn5), 10));
65
66 vn0 = vsubq_f16(vn0, vmagic_bias);
67 vn1 = vsubq_f16(vn1, vmagic_bias);
68 vn2 = vsubq_f16(vn2, vmagic_bias);
69 vn3 = vsubq_f16(vn3, vmagic_bias);
70 vn4 = vsubq_f16(vn4, vmagic_bias);
71 vn5 = vsubq_f16(vn5, vmagic_bias);
72
73 float16x8_t vt0 = vfmaq_f16(vz0, vn0, vln2_hi);
74 float16x8_t vt1 = vfmaq_f16(vz1, vn1, vln2_hi);
75 float16x8_t vt2 = vfmaq_f16(vz2, vn2, vln2_hi);
76 float16x8_t vt3 = vfmaq_f16(vz3, vn3, vln2_hi);
77 float16x8_t vt4 = vfmaq_f16(vz4, vn4, vln2_hi);
78 float16x8_t vt5 = vfmaq_f16(vz5, vn5, vln2_hi);
79
80 vt0 = vfmaq_f16(vt0, vn0, vln2_lo);
81 vt1 = vfmaq_f16(vt1, vn1, vln2_lo);
82 vt2 = vfmaq_f16(vt2, vn2, vln2_lo);
83 vt3 = vfmaq_f16(vt3, vn3, vln2_lo);
84 vt4 = vfmaq_f16(vt4, vn4, vln2_lo);
85 vt5 = vfmaq_f16(vt5, vn5, vln2_lo);
86
87 const float16x8_t vp0 = vfmaq_f16(vc1, vc2, vt0);
88 const float16x8_t vp1 = vfmaq_f16(vc1, vc2, vt1);
89 const float16x8_t vp2 = vfmaq_f16(vc1, vc2, vt2);
90 const float16x8_t vp3 = vfmaq_f16(vc1, vc2, vt3);
91 const float16x8_t vp4 = vfmaq_f16(vc1, vc2, vt4);
92 const float16x8_t vp5 = vfmaq_f16(vc1, vc2, vt5);
93
94 vt0 = vmulq_f16(vt0, vs0);
95 vt1 = vmulq_f16(vt1, vs1);
96 vt2 = vmulq_f16(vt2, vs2);
97 vt3 = vmulq_f16(vt3, vs3);
98 vt4 = vmulq_f16(vt4, vs4);
99 vt5 = vmulq_f16(vt5, vs5);
100
101 const float16x8_t ve0 = vfmaq_f16(vs0, vp0, vt0);
102 const float16x8_t ve1 = vfmaq_f16(vs1, vp1, vt1);
103 const float16x8_t ve2 = vfmaq_f16(vs2, vp2, vt2);
104 const float16x8_t ve3 = vfmaq_f16(vs3, vp3, vt3);
105 const float16x8_t ve4 = vfmaq_f16(vs4, vp4, vt4);
106 const float16x8_t ve5 = vfmaq_f16(vs5, vp5, vt5);
107
108 const float16x8_t vd0 = vaddq_f16(ve0, vone);
109 const float16x8_t vd1 = vaddq_f16(ve1, vone);
110 const float16x8_t vd2 = vaddq_f16(ve2, vone);
111 const float16x8_t vd3 = vaddq_f16(ve3, vone);
112 const float16x8_t vd4 = vaddq_f16(ve4, vone);
113 const float16x8_t vd5 = vaddq_f16(ve5, vone);
114
115 float16x8_t vr0 = vrecpeq_f16(vd0);
116 float16x8_t vr1 = vrecpeq_f16(vd1);
117 float16x8_t vr2 = vrecpeq_f16(vd2);
118 float16x8_t vr3 = vrecpeq_f16(vd3);
119 float16x8_t vr4 = vrecpeq_f16(vd4);
120 float16x8_t vr5 = vrecpeq_f16(vd5);
121
122 const float16x8_t vadj0 = vrecpsq_f16(vr0, vd0);
123 const float16x8_t vadj1 = vrecpsq_f16(vr1, vd1);
124 const float16x8_t vadj2 = vrecpsq_f16(vr2, vd2);
125 const float16x8_t vadj3 = vrecpsq_f16(vr3, vd3);
126 const float16x8_t vadj4 = vrecpsq_f16(vr4, vd4);
127 const float16x8_t vadj5 = vrecpsq_f16(vr5, vd5);
128
129 vr0 = vmulq_f16(vr0, vadj0);
130 vr1 = vmulq_f16(vr1, vadj1);
131 vr2 = vmulq_f16(vr2, vadj2);
132 vr3 = vmulq_f16(vr3, vadj3);
133 vr4 = vmulq_f16(vr4, vadj4);
134 vr5 = vmulq_f16(vr5, vadj5);
135
136 float16x8_t vf0 = vmulq_f16(ve0, vr0);
137 float16x8_t vf1 = vmulq_f16(ve1, vr1);
138 float16x8_t vf2 = vmulq_f16(ve2, vr2);
139 float16x8_t vf3 = vmulq_f16(ve3, vr3);
140 float16x8_t vf4 = vmulq_f16(ve4, vr4);
141 float16x8_t vf5 = vmulq_f16(ve5, vr5);
142
143 vf0 = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf0), vcagtq_f16(vx0, vdenorm_cutoff)));
144 vf1 = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf1), vcagtq_f16(vx1, vdenorm_cutoff)));
145 vf2 = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf2), vcagtq_f16(vx2, vdenorm_cutoff)));
146 vf3 = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf3), vcagtq_f16(vx3, vdenorm_cutoff)));
147 vf4 = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf4), vcagtq_f16(vx4, vdenorm_cutoff)));
148 vf5 = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf5), vcagtq_f16(vx5, vdenorm_cutoff)));
149
150 const uint16x8_t vm0 = vcltq_f16(vx0, vmovq_n_f16(0.0f));
151 const uint16x8_t vm1 = vcltq_f16(vx1, vmovq_n_f16(0.0f));
152 const uint16x8_t vm2 = vcltq_f16(vx2, vmovq_n_f16(0.0f));
153 const uint16x8_t vm3 = vcltq_f16(vx3, vmovq_n_f16(0.0f));
154 const uint16x8_t vm4 = vcltq_f16(vx4, vmovq_n_f16(0.0f));
155 const uint16x8_t vm5 = vcltq_f16(vx5, vmovq_n_f16(0.0f));
156
157 vf0 = vbslq_f16(vm0, vf0, vsubq_f16(vone, vf0));
158 vf1 = vbslq_f16(vm1, vf1, vsubq_f16(vone, vf1));
159 vf2 = vbslq_f16(vm2, vf2, vsubq_f16(vone, vf2));
160 vf3 = vbslq_f16(vm3, vf3, vsubq_f16(vone, vf3));
161 vf4 = vbslq_f16(vm4, vf4, vsubq_f16(vone, vf4));
162 vf5 = vbslq_f16(vm5, vf5, vsubq_f16(vone, vf5));
163
164 vst1q_f16(o, vf0); o += 8;
165 vst1q_f16(o, vf1); o += 8;
166 vst1q_f16(o, vf2); o += 8;
167 vst1q_f16(o, vf3); o += 8;
168 vst1q_f16(o, vf4); o += 8;
169 vst1q_f16(o, vf5); o += 8;
170 }
171 for (; batch >= 8 * sizeof(__fp16); batch -= 8 * sizeof(__fp16)) {
172 const float16x8_t vx = vld1q_f16(i); i += 8;
173
174 const float16x8_t vz = vabsq_f16(vx);
175
176 float16x8_t vn = vfmaq_f16(vmagic_bias, vz, vminus_log2e);
177 const float16x8_t vs = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn), 10));
178 vn = vsubq_f16(vn, vmagic_bias);
179
180 float16x8_t vt = vfmaq_f16(vz, vn, vln2_hi);
181 vt = vfmaq_f16(vt, vn, vln2_lo);
182
183 const float16x8_t vp = vfmaq_f16(vc1, vc2, vt);
184 vt = vmulq_f16(vt, vs);
185 const float16x8_t ve = vfmaq_f16(vs, vp, vt);
186 const float16x8_t vd = vaddq_f16(ve, vone);
187
188 float16x8_t vr = vrecpeq_f16(vd);
189 const float16x8_t vadj = vrecpsq_f16(vr, vd);
190 vr = vmulq_f16(vr, vadj);
191
192 float16x8_t vf = vmulq_f16(ve, vr);
193 vf = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf), vcagtq_f16(vx, vdenorm_cutoff)));
194 const uint16x8_t vm = vcltq_f16(vx, vmovq_n_f16(0.0f));
195 vf = vbslq_f16(vm, vf, vsubq_f16(vone, vf));
196
197 vst1q_f16(o, vf); o += 8;
198 }
199 if XNN_UNLIKELY(batch != 0) {
200 const float16x8_t vx = vld1q_f16(i);
201
202 const float16x8_t vz = vabsq_f16(vx);
203
204 float16x8_t vn = vfmaq_f16(vmagic_bias, vz, vminus_log2e);
205 const float16x8_t vs = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn), 10));
206 vn = vsubq_f16(vn, vmagic_bias);
207
208 float16x8_t vt = vfmaq_f16(vz, vn, vln2_hi);
209 vt = vfmaq_f16(vt, vn, vln2_lo);
210
211 const float16x8_t vp = vfmaq_f16(vc1, vc2, vt);
212 vt = vmulq_f16(vt, vs);
213 const float16x8_t ve = vfmaq_f16(vs, vp, vt);
214 const float16x8_t vd = vaddq_f16(ve, vone);
215
216 float16x8_t vr = vrecpeq_f16(vd);
217 const float16x8_t vadj = vrecpsq_f16(vr, vd);
218 vr = vmulq_f16(vr, vadj);
219
220 float16x8_t vf = vmulq_f16(ve, vr);
221 vf = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf), vcagtq_f16(vx, vdenorm_cutoff)));
222 const uint16x8_t vm = vcltq_f16(vx, vmovq_n_f16(0.0f));
223 vf = vbslq_f16(vm, vf, vsubq_f16(vone, vf));
224
225 float16x4_t vf_lo = vget_low_f16(vf);
226 if (batch & (4 * sizeof(__fp16))) {
227 vst1_f16(o, vf_lo); o += 4;
228 vf_lo = vget_high_f16(vf);
229 }
230 if (batch & (2 * sizeof(__fp16))) {
231 vst1_f16(o, vf_lo); o += 2;
232 vf_lo = vext_f16(vf_lo, vf_lo, 2);
233 }
234 if (batch & (1 * sizeof(__fp16))) {
235 vst1_lane_f16(o, vf_lo, 0);
236 }
237 }
238 }
239