xref: /aosp_15_r20/external/XNNPACK/src/f16-raddstoreexpminusmax/gen/neonfp16arith-rr2-p2-x96-acc2.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1 // Auto-generated file. Do not edit!
2 //   Template: src/f16-raddstoreexpminusmax/neonfp16arith-rr2-p2.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/raddstoreexpminusmax.h>
16 
17 
xnn_f16_raddstoreexpminusmax_ukernel__neonfp16arith_rr2_p2_x96_acc2(size_t batch,const void * input,const void * max,void * output,void * sum,const union xnn_f16_expminus_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_f16_raddstoreexpminusmax_ukernel__neonfp16arith_rr2_p2_x96_acc2(
19     size_t batch,
20     const void* input,
21     const void* max,
22     void* output,
23     void* sum,
24     const union xnn_f16_expminus_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
25 {
26   assert(batch % sizeof(__fp16) == 0);
27 
28   const float16x8_t vi_max = vld1q_dup_f16(max);
29   const float16x8_t vlog2e = vreinterpretq_f16_u16(vld1q_dup_u16(&params->neonfp16arith_rr2_p2.log2e));
30   const float16x8_t vmagic_bias = vreinterpretq_f16_u16(vld1q_dup_u16(&params->neonfp16arith_rr2_p2.magic_bias));
31   const float16x8_t vminus_ln2_hi = vreinterpretq_f16_u16(vld1q_dup_u16(&params->neonfp16arith_rr2_p2.minus_ln2_hi));
32   const float16x8_t vminus_ln2_lo = vreinterpretq_f16_u16(vld1q_dup_u16(&params->neonfp16arith_rr2_p2.minus_ln2_lo));
33   const float16x8_t vc2 = vreinterpretq_f16_u16(vld1q_dup_u16(&params->neonfp16arith_rr2_p2.c2));
34   const float16x8_t vc1 = vreinterpretq_f16_u16(vld1q_dup_u16(&params->neonfp16arith_rr2_p2.c1));
35   const float16x8_t vdenorm_cutoff = vreinterpretq_f16_u16(vld1q_dup_u16(&params->neonfp16arith_rr2_p2.denorm_cutoff));
36 
37   const __fp16* i = (const __fp16*) input;
38   __fp16* o = (__fp16*) output;
39   float16x8_t vacc0 = vmovq_n_f16(0.0f);
40   float16x8_t vacc1 = vmovq_n_f16(0.0f);
41   for (; batch >= 96 * sizeof(__fp16); batch -= 96 * sizeof(__fp16)) {
42     const float16x8_t vi0 = vld1q_f16(i); i += 8;
43     const float16x8_t vi1 = vld1q_f16(i); i += 8;
44     const float16x8_t vi2 = vld1q_f16(i); i += 8;
45     const float16x8_t vi3 = vld1q_f16(i); i += 8;
46     const float16x8_t vi4 = vld1q_f16(i); i += 8;
47     const float16x8_t vi5 = vld1q_f16(i); i += 8;
48     const float16x8_t vi6 = vld1q_f16(i); i += 8;
49     const float16x8_t vi7 = vld1q_f16(i); i += 8;
50     const float16x8_t vi8 = vld1q_f16(i); i += 8;
51     const float16x8_t vi9 = vld1q_f16(i); i += 8;
52     const float16x8_t viA = vld1q_f16(i); i += 8;
53     const float16x8_t viB = vld1q_f16(i); i += 8;
54 
55     const float16x8_t vx0 = vsubq_f16(vi0, vi_max);
56     const float16x8_t vx1 = vsubq_f16(vi1, vi_max);
57     const float16x8_t vx2 = vsubq_f16(vi2, vi_max);
58     const float16x8_t vx3 = vsubq_f16(vi3, vi_max);
59     const float16x8_t vx4 = vsubq_f16(vi4, vi_max);
60     const float16x8_t vx5 = vsubq_f16(vi5, vi_max);
61     const float16x8_t vx6 = vsubq_f16(vi6, vi_max);
62     const float16x8_t vx7 = vsubq_f16(vi7, vi_max);
63     const float16x8_t vx8 = vsubq_f16(vi8, vi_max);
64     const float16x8_t vx9 = vsubq_f16(vi9, vi_max);
65     const float16x8_t vxA = vsubq_f16(viA, vi_max);
66     const float16x8_t vxB = vsubq_f16(viB, vi_max);
67 
68     float16x8_t vn0 = vfmaq_f16(vmagic_bias, vx0, vlog2e);
69     float16x8_t vn1 = vfmaq_f16(vmagic_bias, vx1, vlog2e);
70     float16x8_t vn2 = vfmaq_f16(vmagic_bias, vx2, vlog2e);
71     float16x8_t vn3 = vfmaq_f16(vmagic_bias, vx3, vlog2e);
72     float16x8_t vn4 = vfmaq_f16(vmagic_bias, vx4, vlog2e);
73     float16x8_t vn5 = vfmaq_f16(vmagic_bias, vx5, vlog2e);
74     float16x8_t vn6 = vfmaq_f16(vmagic_bias, vx6, vlog2e);
75     float16x8_t vn7 = vfmaq_f16(vmagic_bias, vx7, vlog2e);
76     float16x8_t vn8 = vfmaq_f16(vmagic_bias, vx8, vlog2e);
77     float16x8_t vn9 = vfmaq_f16(vmagic_bias, vx9, vlog2e);
78     float16x8_t vnA = vfmaq_f16(vmagic_bias, vxA, vlog2e);
79     float16x8_t vnB = vfmaq_f16(vmagic_bias, vxB, vlog2e);
80 
81     const float16x8_t vs0 = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn0), 10));
82     const float16x8_t vs1 = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn1), 10));
83     const float16x8_t vs2 = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn2), 10));
84     const float16x8_t vs3 = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn3), 10));
85     const float16x8_t vs4 = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn4), 10));
86     const float16x8_t vs5 = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn5), 10));
87     const float16x8_t vs6 = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn6), 10));
88     const float16x8_t vs7 = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn7), 10));
89     const float16x8_t vs8 = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn8), 10));
90     const float16x8_t vs9 = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn9), 10));
91     const float16x8_t vsA = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vnA), 10));
92     const float16x8_t vsB = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vnB), 10));
93 
94     vn0 = vsubq_f16(vn0, vmagic_bias);
95     vn1 = vsubq_f16(vn1, vmagic_bias);
96     vn2 = vsubq_f16(vn2, vmagic_bias);
97     vn3 = vsubq_f16(vn3, vmagic_bias);
98     vn4 = vsubq_f16(vn4, vmagic_bias);
99     vn5 = vsubq_f16(vn5, vmagic_bias);
100     vn6 = vsubq_f16(vn6, vmagic_bias);
101     vn7 = vsubq_f16(vn7, vmagic_bias);
102     vn8 = vsubq_f16(vn8, vmagic_bias);
103     vn9 = vsubq_f16(vn9, vmagic_bias);
104     vnA = vsubq_f16(vnA, vmagic_bias);
105     vnB = vsubq_f16(vnB, vmagic_bias);
106 
107     float16x8_t vt0 = vfmaq_f16(vx0, vn0, vminus_ln2_hi);
108     float16x8_t vt1 = vfmaq_f16(vx1, vn1, vminus_ln2_hi);
109     float16x8_t vt2 = vfmaq_f16(vx2, vn2, vminus_ln2_hi);
110     float16x8_t vt3 = vfmaq_f16(vx3, vn3, vminus_ln2_hi);
111     float16x8_t vt4 = vfmaq_f16(vx4, vn4, vminus_ln2_hi);
112     float16x8_t vt5 = vfmaq_f16(vx5, vn5, vminus_ln2_hi);
113     float16x8_t vt6 = vfmaq_f16(vx6, vn6, vminus_ln2_hi);
114     float16x8_t vt7 = vfmaq_f16(vx7, vn7, vminus_ln2_hi);
115     float16x8_t vt8 = vfmaq_f16(vx8, vn8, vminus_ln2_hi);
116     float16x8_t vt9 = vfmaq_f16(vx9, vn9, vminus_ln2_hi);
117     float16x8_t vtA = vfmaq_f16(vxA, vnA, vminus_ln2_hi);
118     float16x8_t vtB = vfmaq_f16(vxB, vnB, vminus_ln2_hi);
119 
120     vt0 = vfmaq_f16(vt0, vn0, vminus_ln2_lo);
121     vt1 = vfmaq_f16(vt1, vn1, vminus_ln2_lo);
122     vt2 = vfmaq_f16(vt2, vn2, vminus_ln2_lo);
123     vt3 = vfmaq_f16(vt3, vn3, vminus_ln2_lo);
124     vt4 = vfmaq_f16(vt4, vn4, vminus_ln2_lo);
125     vt5 = vfmaq_f16(vt5, vn5, vminus_ln2_lo);
126     vt6 = vfmaq_f16(vt6, vn6, vminus_ln2_lo);
127     vt7 = vfmaq_f16(vt7, vn7, vminus_ln2_lo);
128     vt8 = vfmaq_f16(vt8, vn8, vminus_ln2_lo);
129     vt9 = vfmaq_f16(vt9, vn9, vminus_ln2_lo);
130     vtA = vfmaq_f16(vtA, vnA, vminus_ln2_lo);
131     vtB = vfmaq_f16(vtB, vnB, vminus_ln2_lo);
132 
133     const float16x8_t vp0 = vfmaq_f16(vc1, vc2, vt0);
134     const float16x8_t vp1 = vfmaq_f16(vc1, vc2, vt1);
135     const float16x8_t vp2 = vfmaq_f16(vc1, vc2, vt2);
136     const float16x8_t vp3 = vfmaq_f16(vc1, vc2, vt3);
137     const float16x8_t vp4 = vfmaq_f16(vc1, vc2, vt4);
138     const float16x8_t vp5 = vfmaq_f16(vc1, vc2, vt5);
139     const float16x8_t vp6 = vfmaq_f16(vc1, vc2, vt6);
140     const float16x8_t vp7 = vfmaq_f16(vc1, vc2, vt7);
141     const float16x8_t vp8 = vfmaq_f16(vc1, vc2, vt8);
142     const float16x8_t vp9 = vfmaq_f16(vc1, vc2, vt9);
143     const float16x8_t vpA = vfmaq_f16(vc1, vc2, vtA);
144     const float16x8_t vpB = vfmaq_f16(vc1, vc2, vtB);
145 
146     vt0 = vmulq_f16(vt0, vs0);
147     vt1 = vmulq_f16(vt1, vs1);
148     vt2 = vmulq_f16(vt2, vs2);
149     vt3 = vmulq_f16(vt3, vs3);
150     vt4 = vmulq_f16(vt4, vs4);
151     vt5 = vmulq_f16(vt5, vs5);
152     vt6 = vmulq_f16(vt6, vs6);
153     vt7 = vmulq_f16(vt7, vs7);
154     vt8 = vmulq_f16(vt8, vs8);
155     vt9 = vmulq_f16(vt9, vs9);
156     vtA = vmulq_f16(vtA, vsA);
157     vtB = vmulq_f16(vtB, vsB);
158 
159     float16x8_t vf0 = vfmaq_f16(vs0, vp0, vt0);
160     const uint16x8_t vm0 = vcltq_f16(vx0, vdenorm_cutoff);
161     float16x8_t vf1 = vfmaq_f16(vs1, vp1, vt1);
162     const uint16x8_t vm1 = vcltq_f16(vx1, vdenorm_cutoff);
163     float16x8_t vf2 = vfmaq_f16(vs2, vp2, vt2);
164     const uint16x8_t vm2 = vcltq_f16(vx2, vdenorm_cutoff);
165     float16x8_t vf3 = vfmaq_f16(vs3, vp3, vt3);
166     const uint16x8_t vm3 = vcltq_f16(vx3, vdenorm_cutoff);
167     float16x8_t vf4 = vfmaq_f16(vs4, vp4, vt4);
168     const uint16x8_t vm4 = vcltq_f16(vx4, vdenorm_cutoff);
169     float16x8_t vf5 = vfmaq_f16(vs5, vp5, vt5);
170     const uint16x8_t vm5 = vcltq_f16(vx5, vdenorm_cutoff);
171     float16x8_t vf6 = vfmaq_f16(vs6, vp6, vt6);
172     const uint16x8_t vm6 = vcltq_f16(vx6, vdenorm_cutoff);
173     float16x8_t vf7 = vfmaq_f16(vs7, vp7, vt7);
174     const uint16x8_t vm7 = vcltq_f16(vx7, vdenorm_cutoff);
175     float16x8_t vf8 = vfmaq_f16(vs8, vp8, vt8);
176     const uint16x8_t vm8 = vcltq_f16(vx8, vdenorm_cutoff);
177     float16x8_t vf9 = vfmaq_f16(vs9, vp9, vt9);
178     const uint16x8_t vm9 = vcltq_f16(vx9, vdenorm_cutoff);
179     float16x8_t vfA = vfmaq_f16(vsA, vpA, vtA);
180     const uint16x8_t vmA = vcltq_f16(vxA, vdenorm_cutoff);
181     float16x8_t vfB = vfmaq_f16(vsB, vpB, vtB);
182     const uint16x8_t vmB = vcltq_f16(vxB, vdenorm_cutoff);
183 
184     vf0 = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf0), vm0));
185     vf1 = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf1), vm1));
186     vf2 = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf2), vm2));
187     vf3 = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf3), vm3));
188     vf4 = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf4), vm4));
189     vf5 = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf5), vm5));
190     vf6 = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf6), vm6));
191     vf7 = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf7), vm7));
192     vf8 = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf8), vm8));
193     vf9 = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf9), vm9));
194     vfA = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vfA), vmA));
195     vfB = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vfB), vmB));
196 
197     vst1q_f16(o, vf0); o += 8;
198     vst1q_f16(o, vf1); o += 8;
199     vst1q_f16(o, vf2); o += 8;
200     vst1q_f16(o, vf3); o += 8;
201     vst1q_f16(o, vf4); o += 8;
202     vst1q_f16(o, vf5); o += 8;
203     vst1q_f16(o, vf6); o += 8;
204     vst1q_f16(o, vf7); o += 8;
205     vst1q_f16(o, vf8); o += 8;
206     vst1q_f16(o, vf9); o += 8;
207     vst1q_f16(o, vfA); o += 8;
208     vst1q_f16(o, vfB); o += 8;
209 
210     vacc0 = vaddq_f16(vacc0, vf0);
211     vacc1 = vaddq_f16(vacc1, vf1);
212     vacc0 = vaddq_f16(vacc0, vf2);
213     vacc1 = vaddq_f16(vacc1, vf3);
214     vacc0 = vaddq_f16(vacc0, vf4);
215     vacc1 = vaddq_f16(vacc1, vf5);
216     vacc0 = vaddq_f16(vacc0, vf6);
217     vacc1 = vaddq_f16(vacc1, vf7);
218     vacc0 = vaddq_f16(vacc0, vf8);
219     vacc1 = vaddq_f16(vacc1, vf9);
220     vacc0 = vaddq_f16(vacc0, vfA);
221     vacc1 = vaddq_f16(vacc1, vfB);
222   }
223   vacc0 = vaddq_f16(vacc0, vacc1);
224 
225   float16x8_t vacc = vacc0;
226   for (; batch >= 8 * sizeof(__fp16); batch -= 8 * sizeof(__fp16)) {
227     const float16x8_t vi = vld1q_f16(i); i += 8;
228 
229     const float16x8_t vx = vsubq_f16(vi, vi_max);
230 
231     float16x8_t vn = vfmaq_f16(vmagic_bias, vx, vlog2e);
232     const float16x8_t vs = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn), 10));
233     vn = vsubq_f16(vn, vmagic_bias);
234 
235     float16x8_t vt = vfmaq_f16(vx, vn, vminus_ln2_hi);
236     vt = vfmaq_f16(vt, vn, vminus_ln2_lo);
237 
238     const float16x8_t vp = vfmaq_f16(vc1, vc2, vt);
239     vt = vmulq_f16(vt, vs);
240 
241     float16x8_t vf = vfmaq_f16(vs, vp, vt);
242     const uint16x8_t vm = vcltq_f16(vx, vdenorm_cutoff);
243     vf = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf), vm));
244 
245     vst1q_f16(o, vf); o += 8;
246 
247     vacc = vaddq_f16(vacc, vf);
248   }
249   float16x4_t vacc_lo = vadd_f16(vget_low_f16(vacc), vget_high_f16(vacc));
250   if (batch != 0) {
251     assert(batch >= 1 * sizeof(__fp16));
252     assert(batch <= 7 * sizeof(__fp16));
253     const float16x8_t vi = vld1q_f16(i);
254 
255     const float16x8_t vx = vsubq_f16(vi, vi_max);
256 
257     float16x8_t vn = vfmaq_f16(vmagic_bias, vx, vlog2e);
258     const float16x8_t vs = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn), 10));
259     vn = vsubq_f16(vn, vmagic_bias);
260 
261     float16x8_t vt = vfmaq_f16(vx, vn, vminus_ln2_hi);
262     vt = vfmaq_f16(vt, vn, vminus_ln2_lo);
263 
264     const float16x8_t vp = vfmaq_f16(vc1, vc2, vt);
265     vt = vmulq_f16(vt, vs);
266 
267     float16x8_t vf = vfmaq_f16(vs, vp, vt);
268     const uint16x8_t vm = vcltq_f16(vx, vdenorm_cutoff);
269     vf = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf), vm));
270 
271     float16x4_t vf_lo = vget_low_f16(vf);
272     if (batch & (4 * sizeof(__fp16))) {
273       vst1_f16(o, vf_lo); o += 4;
274       vacc_lo = vadd_f16(vacc_lo, vf_lo);
275       vf_lo = vget_high_f16(vf);
276     }
277     if (batch & (2 * sizeof(__fp16))) {
278       vst1_lane_u32((void*) o, vreinterpret_u32_f16(vf_lo), 0); o += 2;
279       vacc_lo = vadd_f16(vacc_lo, vreinterpret_f16_u64(vshl_n_u64(vreinterpret_u64_f16(vf_lo), 32)));
280       vf_lo = vext_f16(vf_lo, vf_lo, 2);
281     }
282     if (batch & (1 * sizeof(__fp16))) {
283       vst1_lane_f16(o, vf_lo, 0);
284       vacc_lo = vadd_f16(vacc_lo, vreinterpret_f16_u64(vshl_n_u64(vreinterpret_u64_f16(vf_lo), 48)));
285     }
286   }
287   vacc_lo = vpadd_f16(vacc_lo, vacc_lo);
288   *((__fp16*) sum) = vget_lane_f16(vacc_lo, 0) + vget_lane_f16(vacc_lo, 1);
289 }
290