xref: /aosp_15_r20/external/XNNPACK/src/f16-raddstoreexpminusmax/gen/neonfp16arith-rr2-p2-x80-acc5.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_x80_acc5(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_x80_acc5(
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   float16x8_t vacc2 = vmovq_n_f16(0.0f);
42   float16x8_t vacc3 = vmovq_n_f16(0.0f);
43   float16x8_t vacc4 = vmovq_n_f16(0.0f);
44   for (; batch >= 80 * sizeof(__fp16); batch -= 80 * sizeof(__fp16)) {
45     const float16x8_t vi0 = vld1q_f16(i); i += 8;
46     const float16x8_t vi1 = vld1q_f16(i); i += 8;
47     const float16x8_t vi2 = vld1q_f16(i); i += 8;
48     const float16x8_t vi3 = vld1q_f16(i); i += 8;
49     const float16x8_t vi4 = vld1q_f16(i); i += 8;
50     const float16x8_t vi5 = vld1q_f16(i); i += 8;
51     const float16x8_t vi6 = vld1q_f16(i); i += 8;
52     const float16x8_t vi7 = vld1q_f16(i); i += 8;
53     const float16x8_t vi8 = vld1q_f16(i); i += 8;
54     const float16x8_t vi9 = vld1q_f16(i); i += 8;
55 
56     const float16x8_t vx0 = vsubq_f16(vi0, vi_max);
57     const float16x8_t vx1 = vsubq_f16(vi1, vi_max);
58     const float16x8_t vx2 = vsubq_f16(vi2, vi_max);
59     const float16x8_t vx3 = vsubq_f16(vi3, vi_max);
60     const float16x8_t vx4 = vsubq_f16(vi4, vi_max);
61     const float16x8_t vx5 = vsubq_f16(vi5, vi_max);
62     const float16x8_t vx6 = vsubq_f16(vi6, vi_max);
63     const float16x8_t vx7 = vsubq_f16(vi7, vi_max);
64     const float16x8_t vx8 = vsubq_f16(vi8, vi_max);
65     const float16x8_t vx9 = vsubq_f16(vi9, vi_max);
66 
67     float16x8_t vn0 = vfmaq_f16(vmagic_bias, vx0, vlog2e);
68     float16x8_t vn1 = vfmaq_f16(vmagic_bias, vx1, vlog2e);
69     float16x8_t vn2 = vfmaq_f16(vmagic_bias, vx2, vlog2e);
70     float16x8_t vn3 = vfmaq_f16(vmagic_bias, vx3, vlog2e);
71     float16x8_t vn4 = vfmaq_f16(vmagic_bias, vx4, vlog2e);
72     float16x8_t vn5 = vfmaq_f16(vmagic_bias, vx5, vlog2e);
73     float16x8_t vn6 = vfmaq_f16(vmagic_bias, vx6, vlog2e);
74     float16x8_t vn7 = vfmaq_f16(vmagic_bias, vx7, vlog2e);
75     float16x8_t vn8 = vfmaq_f16(vmagic_bias, vx8, vlog2e);
76     float16x8_t vn9 = vfmaq_f16(vmagic_bias, vx9, vlog2e);
77 
78     const float16x8_t vs0 = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn0), 10));
79     const float16x8_t vs1 = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn1), 10));
80     const float16x8_t vs2 = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn2), 10));
81     const float16x8_t vs3 = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn3), 10));
82     const float16x8_t vs4 = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn4), 10));
83     const float16x8_t vs5 = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn5), 10));
84     const float16x8_t vs6 = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn6), 10));
85     const float16x8_t vs7 = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn7), 10));
86     const float16x8_t vs8 = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn8), 10));
87     const float16x8_t vs9 = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn9), 10));
88 
89     vn0 = vsubq_f16(vn0, vmagic_bias);
90     vn1 = vsubq_f16(vn1, vmagic_bias);
91     vn2 = vsubq_f16(vn2, vmagic_bias);
92     vn3 = vsubq_f16(vn3, vmagic_bias);
93     vn4 = vsubq_f16(vn4, vmagic_bias);
94     vn5 = vsubq_f16(vn5, vmagic_bias);
95     vn6 = vsubq_f16(vn6, vmagic_bias);
96     vn7 = vsubq_f16(vn7, vmagic_bias);
97     vn8 = vsubq_f16(vn8, vmagic_bias);
98     vn9 = vsubq_f16(vn9, vmagic_bias);
99 
100     float16x8_t vt0 = vfmaq_f16(vx0, vn0, vminus_ln2_hi);
101     float16x8_t vt1 = vfmaq_f16(vx1, vn1, vminus_ln2_hi);
102     float16x8_t vt2 = vfmaq_f16(vx2, vn2, vminus_ln2_hi);
103     float16x8_t vt3 = vfmaq_f16(vx3, vn3, vminus_ln2_hi);
104     float16x8_t vt4 = vfmaq_f16(vx4, vn4, vminus_ln2_hi);
105     float16x8_t vt5 = vfmaq_f16(vx5, vn5, vminus_ln2_hi);
106     float16x8_t vt6 = vfmaq_f16(vx6, vn6, vminus_ln2_hi);
107     float16x8_t vt7 = vfmaq_f16(vx7, vn7, vminus_ln2_hi);
108     float16x8_t vt8 = vfmaq_f16(vx8, vn8, vminus_ln2_hi);
109     float16x8_t vt9 = vfmaq_f16(vx9, vn9, vminus_ln2_hi);
110 
111     vt0 = vfmaq_f16(vt0, vn0, vminus_ln2_lo);
112     vt1 = vfmaq_f16(vt1, vn1, vminus_ln2_lo);
113     vt2 = vfmaq_f16(vt2, vn2, vminus_ln2_lo);
114     vt3 = vfmaq_f16(vt3, vn3, vminus_ln2_lo);
115     vt4 = vfmaq_f16(vt4, vn4, vminus_ln2_lo);
116     vt5 = vfmaq_f16(vt5, vn5, vminus_ln2_lo);
117     vt6 = vfmaq_f16(vt6, vn6, vminus_ln2_lo);
118     vt7 = vfmaq_f16(vt7, vn7, vminus_ln2_lo);
119     vt8 = vfmaq_f16(vt8, vn8, vminus_ln2_lo);
120     vt9 = vfmaq_f16(vt9, vn9, vminus_ln2_lo);
121 
122     const float16x8_t vp0 = vfmaq_f16(vc1, vc2, vt0);
123     const float16x8_t vp1 = vfmaq_f16(vc1, vc2, vt1);
124     const float16x8_t vp2 = vfmaq_f16(vc1, vc2, vt2);
125     const float16x8_t vp3 = vfmaq_f16(vc1, vc2, vt3);
126     const float16x8_t vp4 = vfmaq_f16(vc1, vc2, vt4);
127     const float16x8_t vp5 = vfmaq_f16(vc1, vc2, vt5);
128     const float16x8_t vp6 = vfmaq_f16(vc1, vc2, vt6);
129     const float16x8_t vp7 = vfmaq_f16(vc1, vc2, vt7);
130     const float16x8_t vp8 = vfmaq_f16(vc1, vc2, vt8);
131     const float16x8_t vp9 = vfmaq_f16(vc1, vc2, vt9);
132 
133     vt0 = vmulq_f16(vt0, vs0);
134     vt1 = vmulq_f16(vt1, vs1);
135     vt2 = vmulq_f16(vt2, vs2);
136     vt3 = vmulq_f16(vt3, vs3);
137     vt4 = vmulq_f16(vt4, vs4);
138     vt5 = vmulq_f16(vt5, vs5);
139     vt6 = vmulq_f16(vt6, vs6);
140     vt7 = vmulq_f16(vt7, vs7);
141     vt8 = vmulq_f16(vt8, vs8);
142     vt9 = vmulq_f16(vt9, vs9);
143 
144     float16x8_t vf0 = vfmaq_f16(vs0, vp0, vt0);
145     const uint16x8_t vm0 = vcltq_f16(vx0, vdenorm_cutoff);
146     float16x8_t vf1 = vfmaq_f16(vs1, vp1, vt1);
147     const uint16x8_t vm1 = vcltq_f16(vx1, vdenorm_cutoff);
148     float16x8_t vf2 = vfmaq_f16(vs2, vp2, vt2);
149     const uint16x8_t vm2 = vcltq_f16(vx2, vdenorm_cutoff);
150     float16x8_t vf3 = vfmaq_f16(vs3, vp3, vt3);
151     const uint16x8_t vm3 = vcltq_f16(vx3, vdenorm_cutoff);
152     float16x8_t vf4 = vfmaq_f16(vs4, vp4, vt4);
153     const uint16x8_t vm4 = vcltq_f16(vx4, vdenorm_cutoff);
154     float16x8_t vf5 = vfmaq_f16(vs5, vp5, vt5);
155     const uint16x8_t vm5 = vcltq_f16(vx5, vdenorm_cutoff);
156     float16x8_t vf6 = vfmaq_f16(vs6, vp6, vt6);
157     const uint16x8_t vm6 = vcltq_f16(vx6, vdenorm_cutoff);
158     float16x8_t vf7 = vfmaq_f16(vs7, vp7, vt7);
159     const uint16x8_t vm7 = vcltq_f16(vx7, vdenorm_cutoff);
160     float16x8_t vf8 = vfmaq_f16(vs8, vp8, vt8);
161     const uint16x8_t vm8 = vcltq_f16(vx8, vdenorm_cutoff);
162     float16x8_t vf9 = vfmaq_f16(vs9, vp9, vt9);
163     const uint16x8_t vm9 = vcltq_f16(vx9, vdenorm_cutoff);
164 
165     vf0 = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf0), vm0));
166     vf1 = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf1), vm1));
167     vf2 = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf2), vm2));
168     vf3 = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf3), vm3));
169     vf4 = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf4), vm4));
170     vf5 = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf5), vm5));
171     vf6 = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf6), vm6));
172     vf7 = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf7), vm7));
173     vf8 = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf8), vm8));
174     vf9 = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf9), vm9));
175 
176     vst1q_f16(o, vf0); o += 8;
177     vst1q_f16(o, vf1); o += 8;
178     vst1q_f16(o, vf2); o += 8;
179     vst1q_f16(o, vf3); o += 8;
180     vst1q_f16(o, vf4); o += 8;
181     vst1q_f16(o, vf5); o += 8;
182     vst1q_f16(o, vf6); o += 8;
183     vst1q_f16(o, vf7); o += 8;
184     vst1q_f16(o, vf8); o += 8;
185     vst1q_f16(o, vf9); o += 8;
186 
187     vacc0 = vaddq_f16(vacc0, vf0);
188     vacc1 = vaddq_f16(vacc1, vf1);
189     vacc2 = vaddq_f16(vacc2, vf2);
190     vacc3 = vaddq_f16(vacc3, vf3);
191     vacc4 = vaddq_f16(vacc4, vf4);
192     vacc0 = vaddq_f16(vacc0, vf5);
193     vacc1 = vaddq_f16(vacc1, vf6);
194     vacc2 = vaddq_f16(vacc2, vf7);
195     vacc3 = vaddq_f16(vacc3, vf8);
196     vacc4 = vaddq_f16(vacc4, vf9);
197   }
198   vacc0 = vaddq_f16(vacc0, vacc1);
199   vacc2 = vaddq_f16(vacc2, vacc3);
200   vacc0 = vaddq_f16(vacc0, vacc2);
201   vacc0 = vaddq_f16(vacc0, vacc4);
202 
203   float16x8_t vacc = vacc0;
204   for (; batch >= 8 * sizeof(__fp16); batch -= 8 * sizeof(__fp16)) {
205     const float16x8_t vi = vld1q_f16(i); i += 8;
206 
207     const float16x8_t vx = vsubq_f16(vi, vi_max);
208 
209     float16x8_t vn = vfmaq_f16(vmagic_bias, vx, vlog2e);
210     const float16x8_t vs = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn), 10));
211     vn = vsubq_f16(vn, vmagic_bias);
212 
213     float16x8_t vt = vfmaq_f16(vx, vn, vminus_ln2_hi);
214     vt = vfmaq_f16(vt, vn, vminus_ln2_lo);
215 
216     const float16x8_t vp = vfmaq_f16(vc1, vc2, vt);
217     vt = vmulq_f16(vt, vs);
218 
219     float16x8_t vf = vfmaq_f16(vs, vp, vt);
220     const uint16x8_t vm = vcltq_f16(vx, vdenorm_cutoff);
221     vf = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf), vm));
222 
223     vst1q_f16(o, vf); o += 8;
224 
225     vacc = vaddq_f16(vacc, vf);
226   }
227   float16x4_t vacc_lo = vadd_f16(vget_low_f16(vacc), vget_high_f16(vacc));
228   if (batch != 0) {
229     assert(batch >= 1 * sizeof(__fp16));
230     assert(batch <= 7 * sizeof(__fp16));
231     const float16x8_t vi = vld1q_f16(i);
232 
233     const float16x8_t vx = vsubq_f16(vi, vi_max);
234 
235     float16x8_t vn = vfmaq_f16(vmagic_bias, vx, vlog2e);
236     const float16x8_t vs = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn), 10));
237     vn = vsubq_f16(vn, vmagic_bias);
238 
239     float16x8_t vt = vfmaq_f16(vx, vn, vminus_ln2_hi);
240     vt = vfmaq_f16(vt, vn, vminus_ln2_lo);
241 
242     const float16x8_t vp = vfmaq_f16(vc1, vc2, vt);
243     vt = vmulq_f16(vt, vs);
244 
245     float16x8_t vf = vfmaq_f16(vs, vp, vt);
246     const uint16x8_t vm = vcltq_f16(vx, vdenorm_cutoff);
247     vf = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf), vm));
248 
249     float16x4_t vf_lo = vget_low_f16(vf);
250     if (batch & (4 * sizeof(__fp16))) {
251       vst1_f16(o, vf_lo); o += 4;
252       vacc_lo = vadd_f16(vacc_lo, vf_lo);
253       vf_lo = vget_high_f16(vf);
254     }
255     if (batch & (2 * sizeof(__fp16))) {
256       vst1_lane_u32((void*) o, vreinterpret_u32_f16(vf_lo), 0); o += 2;
257       vacc_lo = vadd_f16(vacc_lo, vreinterpret_f16_u64(vshl_n_u64(vreinterpret_u64_f16(vf_lo), 32)));
258       vf_lo = vext_f16(vf_lo, vf_lo, 2);
259     }
260     if (batch & (1 * sizeof(__fp16))) {
261       vst1_lane_f16(o, vf_lo, 0);
262       vacc_lo = vadd_f16(vacc_lo, vreinterpret_f16_u64(vshl_n_u64(vreinterpret_u64_f16(vf_lo), 48)));
263     }
264   }
265   vacc_lo = vpadd_f16(vacc_lo, vacc_lo);
266   *((__fp16*) sum) = vget_lane_f16(vacc_lo, 0) + vget_lane_f16(vacc_lo, 1);
267 }
268