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_x64_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_x64_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(¶ms->neonfp16arith_rr2_p2.log2e));
30 const float16x8_t vmagic_bias = vreinterpretq_f16_u16(vld1q_dup_u16(¶ms->neonfp16arith_rr2_p2.magic_bias));
31 const float16x8_t vminus_ln2_hi = vreinterpretq_f16_u16(vld1q_dup_u16(¶ms->neonfp16arith_rr2_p2.minus_ln2_hi));
32 const float16x8_t vminus_ln2_lo = vreinterpretq_f16_u16(vld1q_dup_u16(¶ms->neonfp16arith_rr2_p2.minus_ln2_lo));
33 const float16x8_t vc2 = vreinterpretq_f16_u16(vld1q_dup_u16(¶ms->neonfp16arith_rr2_p2.c2));
34 const float16x8_t vc1 = vreinterpretq_f16_u16(vld1q_dup_u16(¶ms->neonfp16arith_rr2_p2.c1));
35 const float16x8_t vdenorm_cutoff = vreinterpretq_f16_u16(vld1q_dup_u16(¶ms->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 >= 64 * sizeof(__fp16); batch -= 64 * 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
51 const float16x8_t vx0 = vsubq_f16(vi0, vi_max);
52 const float16x8_t vx1 = vsubq_f16(vi1, vi_max);
53 const float16x8_t vx2 = vsubq_f16(vi2, vi_max);
54 const float16x8_t vx3 = vsubq_f16(vi3, vi_max);
55 const float16x8_t vx4 = vsubq_f16(vi4, vi_max);
56 const float16x8_t vx5 = vsubq_f16(vi5, vi_max);
57 const float16x8_t vx6 = vsubq_f16(vi6, vi_max);
58 const float16x8_t vx7 = vsubq_f16(vi7, vi_max);
59
60 float16x8_t vn0 = vfmaq_f16(vmagic_bias, vx0, vlog2e);
61 float16x8_t vn1 = vfmaq_f16(vmagic_bias, vx1, vlog2e);
62 float16x8_t vn2 = vfmaq_f16(vmagic_bias, vx2, vlog2e);
63 float16x8_t vn3 = vfmaq_f16(vmagic_bias, vx3, vlog2e);
64 float16x8_t vn4 = vfmaq_f16(vmagic_bias, vx4, vlog2e);
65 float16x8_t vn5 = vfmaq_f16(vmagic_bias, vx5, vlog2e);
66 float16x8_t vn6 = vfmaq_f16(vmagic_bias, vx6, vlog2e);
67 float16x8_t vn7 = vfmaq_f16(vmagic_bias, vx7, vlog2e);
68
69 const float16x8_t vs0 = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn0), 10));
70 const float16x8_t vs1 = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn1), 10));
71 const float16x8_t vs2 = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn2), 10));
72 const float16x8_t vs3 = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn3), 10));
73 const float16x8_t vs4 = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn4), 10));
74 const float16x8_t vs5 = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn5), 10));
75 const float16x8_t vs6 = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn6), 10));
76 const float16x8_t vs7 = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn7), 10));
77
78 vn0 = vsubq_f16(vn0, vmagic_bias);
79 vn1 = vsubq_f16(vn1, vmagic_bias);
80 vn2 = vsubq_f16(vn2, vmagic_bias);
81 vn3 = vsubq_f16(vn3, vmagic_bias);
82 vn4 = vsubq_f16(vn4, vmagic_bias);
83 vn5 = vsubq_f16(vn5, vmagic_bias);
84 vn6 = vsubq_f16(vn6, vmagic_bias);
85 vn7 = vsubq_f16(vn7, vmagic_bias);
86
87 float16x8_t vt0 = vfmaq_f16(vx0, vn0, vminus_ln2_hi);
88 float16x8_t vt1 = vfmaq_f16(vx1, vn1, vminus_ln2_hi);
89 float16x8_t vt2 = vfmaq_f16(vx2, vn2, vminus_ln2_hi);
90 float16x8_t vt3 = vfmaq_f16(vx3, vn3, vminus_ln2_hi);
91 float16x8_t vt4 = vfmaq_f16(vx4, vn4, vminus_ln2_hi);
92 float16x8_t vt5 = vfmaq_f16(vx5, vn5, vminus_ln2_hi);
93 float16x8_t vt6 = vfmaq_f16(vx6, vn6, vminus_ln2_hi);
94 float16x8_t vt7 = vfmaq_f16(vx7, vn7, vminus_ln2_hi);
95
96 vt0 = vfmaq_f16(vt0, vn0, vminus_ln2_lo);
97 vt1 = vfmaq_f16(vt1, vn1, vminus_ln2_lo);
98 vt2 = vfmaq_f16(vt2, vn2, vminus_ln2_lo);
99 vt3 = vfmaq_f16(vt3, vn3, vminus_ln2_lo);
100 vt4 = vfmaq_f16(vt4, vn4, vminus_ln2_lo);
101 vt5 = vfmaq_f16(vt5, vn5, vminus_ln2_lo);
102 vt6 = vfmaq_f16(vt6, vn6, vminus_ln2_lo);
103 vt7 = vfmaq_f16(vt7, vn7, vminus_ln2_lo);
104
105 const float16x8_t vp0 = vfmaq_f16(vc1, vc2, vt0);
106 const float16x8_t vp1 = vfmaq_f16(vc1, vc2, vt1);
107 const float16x8_t vp2 = vfmaq_f16(vc1, vc2, vt2);
108 const float16x8_t vp3 = vfmaq_f16(vc1, vc2, vt3);
109 const float16x8_t vp4 = vfmaq_f16(vc1, vc2, vt4);
110 const float16x8_t vp5 = vfmaq_f16(vc1, vc2, vt5);
111 const float16x8_t vp6 = vfmaq_f16(vc1, vc2, vt6);
112 const float16x8_t vp7 = vfmaq_f16(vc1, vc2, vt7);
113
114 vt0 = vmulq_f16(vt0, vs0);
115 vt1 = vmulq_f16(vt1, vs1);
116 vt2 = vmulq_f16(vt2, vs2);
117 vt3 = vmulq_f16(vt3, vs3);
118 vt4 = vmulq_f16(vt4, vs4);
119 vt5 = vmulq_f16(vt5, vs5);
120 vt6 = vmulq_f16(vt6, vs6);
121 vt7 = vmulq_f16(vt7, vs7);
122
123 float16x8_t vf0 = vfmaq_f16(vs0, vp0, vt0);
124 const uint16x8_t vm0 = vcltq_f16(vx0, vdenorm_cutoff);
125 float16x8_t vf1 = vfmaq_f16(vs1, vp1, vt1);
126 const uint16x8_t vm1 = vcltq_f16(vx1, vdenorm_cutoff);
127 float16x8_t vf2 = vfmaq_f16(vs2, vp2, vt2);
128 const uint16x8_t vm2 = vcltq_f16(vx2, vdenorm_cutoff);
129 float16x8_t vf3 = vfmaq_f16(vs3, vp3, vt3);
130 const uint16x8_t vm3 = vcltq_f16(vx3, vdenorm_cutoff);
131 float16x8_t vf4 = vfmaq_f16(vs4, vp4, vt4);
132 const uint16x8_t vm4 = vcltq_f16(vx4, vdenorm_cutoff);
133 float16x8_t vf5 = vfmaq_f16(vs5, vp5, vt5);
134 const uint16x8_t vm5 = vcltq_f16(vx5, vdenorm_cutoff);
135 float16x8_t vf6 = vfmaq_f16(vs6, vp6, vt6);
136 const uint16x8_t vm6 = vcltq_f16(vx6, vdenorm_cutoff);
137 float16x8_t vf7 = vfmaq_f16(vs7, vp7, vt7);
138 const uint16x8_t vm7 = vcltq_f16(vx7, vdenorm_cutoff);
139
140 vf0 = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf0), vm0));
141 vf1 = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf1), vm1));
142 vf2 = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf2), vm2));
143 vf3 = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf3), vm3));
144 vf4 = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf4), vm4));
145 vf5 = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf5), vm5));
146 vf6 = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf6), vm6));
147 vf7 = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf7), vm7));
148
149 vst1q_f16(o, vf0); o += 8;
150 vst1q_f16(o, vf1); o += 8;
151 vst1q_f16(o, vf2); o += 8;
152 vst1q_f16(o, vf3); o += 8;
153 vst1q_f16(o, vf4); o += 8;
154 vst1q_f16(o, vf5); o += 8;
155 vst1q_f16(o, vf6); o += 8;
156 vst1q_f16(o, vf7); o += 8;
157
158 vacc0 = vaddq_f16(vacc0, vf0);
159 vacc1 = vaddq_f16(vacc1, vf1);
160 vacc0 = vaddq_f16(vacc0, vf2);
161 vacc1 = vaddq_f16(vacc1, vf3);
162 vacc0 = vaddq_f16(vacc0, vf4);
163 vacc1 = vaddq_f16(vacc1, vf5);
164 vacc0 = vaddq_f16(vacc0, vf6);
165 vacc1 = vaddq_f16(vacc1, vf7);
166 }
167 vacc0 = vaddq_f16(vacc0, vacc1);
168
169 float16x8_t vacc = vacc0;
170 for (; batch >= 8 * sizeof(__fp16); batch -= 8 * sizeof(__fp16)) {
171 const float16x8_t vi = vld1q_f16(i); i += 8;
172
173 const float16x8_t vx = vsubq_f16(vi, vi_max);
174
175 float16x8_t vn = vfmaq_f16(vmagic_bias, vx, vlog2e);
176 const float16x8_t vs = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn), 10));
177 vn = vsubq_f16(vn, vmagic_bias);
178
179 float16x8_t vt = vfmaq_f16(vx, vn, vminus_ln2_hi);
180 vt = vfmaq_f16(vt, vn, vminus_ln2_lo);
181
182 const float16x8_t vp = vfmaq_f16(vc1, vc2, vt);
183 vt = vmulq_f16(vt, vs);
184
185 float16x8_t vf = vfmaq_f16(vs, vp, vt);
186 const uint16x8_t vm = vcltq_f16(vx, vdenorm_cutoff);
187 vf = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf), vm));
188
189 vst1q_f16(o, vf); o += 8;
190
191 vacc = vaddq_f16(vacc, vf);
192 }
193 float16x4_t vacc_lo = vadd_f16(vget_low_f16(vacc), vget_high_f16(vacc));
194 if (batch != 0) {
195 assert(batch >= 1 * sizeof(__fp16));
196 assert(batch <= 7 * sizeof(__fp16));
197 const float16x8_t vi = vld1q_f16(i);
198
199 const float16x8_t vx = vsubq_f16(vi, vi_max);
200
201 float16x8_t vn = vfmaq_f16(vmagic_bias, vx, vlog2e);
202 const float16x8_t vs = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn), 10));
203 vn = vsubq_f16(vn, vmagic_bias);
204
205 float16x8_t vt = vfmaq_f16(vx, vn, vminus_ln2_hi);
206 vt = vfmaq_f16(vt, vn, vminus_ln2_lo);
207
208 const float16x8_t vp = vfmaq_f16(vc1, vc2, vt);
209 vt = vmulq_f16(vt, vs);
210
211 float16x8_t vf = vfmaq_f16(vs, vp, vt);
212 const uint16x8_t vm = vcltq_f16(vx, vdenorm_cutoff);
213 vf = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf), vm));
214
215 float16x4_t vf_lo = vget_low_f16(vf);
216 if (batch & (4 * sizeof(__fp16))) {
217 vst1_f16(o, vf_lo); o += 4;
218 vacc_lo = vadd_f16(vacc_lo, vf_lo);
219 vf_lo = vget_high_f16(vf);
220 }
221 if (batch & (2 * sizeof(__fp16))) {
222 vst1_lane_u32((void*) o, vreinterpret_u32_f16(vf_lo), 0); o += 2;
223 vacc_lo = vadd_f16(vacc_lo, vreinterpret_f16_u64(vshl_n_u64(vreinterpret_u64_f16(vf_lo), 32)));
224 vf_lo = vext_f16(vf_lo, vf_lo, 2);
225 }
226 if (batch & (1 * sizeof(__fp16))) {
227 vst1_lane_f16(o, vf_lo, 0);
228 vacc_lo = vadd_f16(vacc_lo, vreinterpret_f16_u64(vshl_n_u64(vreinterpret_u64_f16(vf_lo), 48)));
229 }
230 }
231 vacc_lo = vpadd_f16(vacc_lo, vacc_lo);
232 *((__fp16*) sum) = vget_lane_f16(vacc_lo, 0) + vget_lane_f16(vacc_lo, 1);
233 }
234