xref: /aosp_15_r20/external/XNNPACK/src/f16-vsigmoid/gen/vsigmoid-neonfp16arith-rr2-p2-div-x64.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
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_div_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_div_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(&params->neonfp16arith_rr2_p2.magic_bias));
27   const float16x8_t vminus_log2e = vreinterpretq_f16_u16(vld1q_dup_u16(&params->neonfp16arith_rr2_p2.minus_log2e));
28   const float16x8_t vln2_hi = vreinterpretq_f16_u16(vld1q_dup_u16(&params->neonfp16arith_rr2_p2.ln2_hi));
29   const float16x8_t vln2_lo = vreinterpretq_f16_u16(vld1q_dup_u16(&params->neonfp16arith_rr2_p2.ln2_lo));
30   const float16x8_t vc2 = vreinterpretq_f16_u16(vld1q_dup_u16(&params->neonfp16arith_rr2_p2.c2));
31   const float16x8_t vc1 = vreinterpretq_f16_u16(vld1q_dup_u16(&params->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(&params->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 vf0 = vdivq_f16(ve0, vd0);
138     float16x8_t vf1 = vdivq_f16(ve1, vd1);
139     float16x8_t vf2 = vdivq_f16(ve2, vd2);
140     float16x8_t vf3 = vdivq_f16(ve3, vd3);
141     float16x8_t vf4 = vdivq_f16(ve4, vd4);
142     float16x8_t vf5 = vdivq_f16(ve5, vd5);
143     float16x8_t vf6 = vdivq_f16(ve6, vd6);
144     float16x8_t vf7 = vdivq_f16(ve7, vd7);
145 
146     vf0 = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf0), vcagtq_f16(vx0, vdenorm_cutoff)));
147     vf1 = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf1), vcagtq_f16(vx1, vdenorm_cutoff)));
148     vf2 = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf2), vcagtq_f16(vx2, vdenorm_cutoff)));
149     vf3 = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf3), vcagtq_f16(vx3, vdenorm_cutoff)));
150     vf4 = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf4), vcagtq_f16(vx4, vdenorm_cutoff)));
151     vf5 = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf5), vcagtq_f16(vx5, vdenorm_cutoff)));
152     vf6 = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf6), vcagtq_f16(vx6, vdenorm_cutoff)));
153     vf7 = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf7), vcagtq_f16(vx7, vdenorm_cutoff)));
154 
155     const uint16x8_t vm0 = vcltq_f16(vx0, vmovq_n_f16(0.0f));
156     const uint16x8_t vm1 = vcltq_f16(vx1, vmovq_n_f16(0.0f));
157     const uint16x8_t vm2 = vcltq_f16(vx2, vmovq_n_f16(0.0f));
158     const uint16x8_t vm3 = vcltq_f16(vx3, vmovq_n_f16(0.0f));
159     const uint16x8_t vm4 = vcltq_f16(vx4, vmovq_n_f16(0.0f));
160     const uint16x8_t vm5 = vcltq_f16(vx5, vmovq_n_f16(0.0f));
161     const uint16x8_t vm6 = vcltq_f16(vx6, vmovq_n_f16(0.0f));
162     const uint16x8_t vm7 = vcltq_f16(vx7, vmovq_n_f16(0.0f));
163 
164     vf0 = vbslq_f16(vm0, vf0, vsubq_f16(vone, vf0));
165     vf1 = vbslq_f16(vm1, vf1, vsubq_f16(vone, vf1));
166     vf2 = vbslq_f16(vm2, vf2, vsubq_f16(vone, vf2));
167     vf3 = vbslq_f16(vm3, vf3, vsubq_f16(vone, vf3));
168     vf4 = vbslq_f16(vm4, vf4, vsubq_f16(vone, vf4));
169     vf5 = vbslq_f16(vm5, vf5, vsubq_f16(vone, vf5));
170     vf6 = vbslq_f16(vm6, vf6, vsubq_f16(vone, vf6));
171     vf7 = vbslq_f16(vm7, vf7, vsubq_f16(vone, vf7));
172 
173     vst1q_f16(o, vf0); o += 8;
174     vst1q_f16(o, vf1); o += 8;
175     vst1q_f16(o, vf2); o += 8;
176     vst1q_f16(o, vf3); o += 8;
177     vst1q_f16(o, vf4); o += 8;
178     vst1q_f16(o, vf5); o += 8;
179     vst1q_f16(o, vf6); o += 8;
180     vst1q_f16(o, vf7); o += 8;
181   }
182   for (; batch >= 8 * sizeof(__fp16); batch -= 8 * sizeof(__fp16)) {
183     const float16x8_t vx = vld1q_f16(i); i += 8;
184 
185     const float16x8_t vz = vabsq_f16(vx);
186 
187     float16x8_t vn = vfmaq_f16(vmagic_bias, vz, vminus_log2e);
188     const float16x8_t vs = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn), 10));
189     vn = vsubq_f16(vn, vmagic_bias);
190 
191     float16x8_t vt = vfmaq_f16(vz, vn, vln2_hi);
192     vt = vfmaq_f16(vt, vn, vln2_lo);
193 
194     const float16x8_t vp = vfmaq_f16(vc1, vc2, vt);
195     vt = vmulq_f16(vt, vs);
196     const float16x8_t ve = vfmaq_f16(vs, vp, vt);
197     const float16x8_t vd = vaddq_f16(ve, vone);
198 
199     float16x8_t vf = vdivq_f16(ve, vd);
200     vf = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf), vcagtq_f16(vx, vdenorm_cutoff)));
201     const uint16x8_t vm = vcltq_f16(vx, vmovq_n_f16(0.0f));
202     vf = vbslq_f16(vm, vf, vsubq_f16(vone, vf));
203 
204     vst1q_f16(o, vf); o += 8;
205   }
206   if XNN_UNLIKELY(batch != 0) {
207     const float16x8_t vx = vld1q_f16(i);
208 
209     const float16x8_t vz = vabsq_f16(vx);
210 
211     float16x8_t vn = vfmaq_f16(vmagic_bias, vz, vminus_log2e);
212     const float16x8_t vs = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn), 10));
213     vn = vsubq_f16(vn, vmagic_bias);
214 
215     float16x8_t vt = vfmaq_f16(vz, vn, vln2_hi);
216     vt = vfmaq_f16(vt, vn, vln2_lo);
217 
218     const float16x8_t vp = vfmaq_f16(vc1, vc2, vt);
219     vt = vmulq_f16(vt, vs);
220     const float16x8_t ve = vfmaq_f16(vs, vp, vt);
221     const float16x8_t vd = vaddq_f16(ve, vone);
222 
223     float16x8_t vf = vdivq_f16(ve, vd);
224     vf = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf), vcagtq_f16(vx, vdenorm_cutoff)));
225     const uint16x8_t vm = vcltq_f16(vx, vmovq_n_f16(0.0f));
226     vf = vbslq_f16(vm, vf, vsubq_f16(vone, vf));
227 
228     float16x4_t vf_lo = vget_low_f16(vf);
229     if (batch & (4 * sizeof(__fp16))) {
230       vst1_f16(o, vf_lo); o += 4;
231       vf_lo = vget_high_f16(vf);
232     }
233     if (batch & (2 * sizeof(__fp16))) {
234       vst1_f16(o, vf_lo); o += 2;
235       vf_lo = vext_f16(vf_lo, vf_lo, 2);
236     }
237     if (batch & (1 * sizeof(__fp16))) {
238       vst1_lane_f16(o, vf_lo, 0);
239     }
240   }
241 }
242