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_x48(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_x48(
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 for (; batch >= 48 * sizeof(__fp16); batch -= 48 * sizeof(__fp16)) {
41 const float16x8_t vi0 = vld1q_f16(i); i += 8;
42 const float16x8_t vi1 = vld1q_f16(i); i += 8;
43 const float16x8_t vi2 = vld1q_f16(i); i += 8;
44 const float16x8_t vi3 = vld1q_f16(i); i += 8;
45 const float16x8_t vi4 = vld1q_f16(i); i += 8;
46 const float16x8_t vi5 = vld1q_f16(i); i += 8;
47
48 const float16x8_t vx0 = vsubq_f16(vi0, vi_max);
49 const float16x8_t vx1 = vsubq_f16(vi1, vi_max);
50 const float16x8_t vx2 = vsubq_f16(vi2, vi_max);
51 const float16x8_t vx3 = vsubq_f16(vi3, vi_max);
52 const float16x8_t vx4 = vsubq_f16(vi4, vi_max);
53 const float16x8_t vx5 = vsubq_f16(vi5, vi_max);
54
55 float16x8_t vn0 = vfmaq_f16(vmagic_bias, vx0, vlog2e);
56 float16x8_t vn1 = vfmaq_f16(vmagic_bias, vx1, vlog2e);
57 float16x8_t vn2 = vfmaq_f16(vmagic_bias, vx2, vlog2e);
58 float16x8_t vn3 = vfmaq_f16(vmagic_bias, vx3, vlog2e);
59 float16x8_t vn4 = vfmaq_f16(vmagic_bias, vx4, vlog2e);
60 float16x8_t vn5 = vfmaq_f16(vmagic_bias, vx5, vlog2e);
61
62 const float16x8_t vs0 = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn0), 10));
63 const float16x8_t vs1 = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn1), 10));
64 const float16x8_t vs2 = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn2), 10));
65 const float16x8_t vs3 = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn3), 10));
66 const float16x8_t vs4 = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn4), 10));
67 const float16x8_t vs5 = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn5), 10));
68
69 vn0 = vsubq_f16(vn0, vmagic_bias);
70 vn1 = vsubq_f16(vn1, vmagic_bias);
71 vn2 = vsubq_f16(vn2, vmagic_bias);
72 vn3 = vsubq_f16(vn3, vmagic_bias);
73 vn4 = vsubq_f16(vn4, vmagic_bias);
74 vn5 = vsubq_f16(vn5, vmagic_bias);
75
76 float16x8_t vt0 = vfmaq_f16(vx0, vn0, vminus_ln2_hi);
77 float16x8_t vt1 = vfmaq_f16(vx1, vn1, vminus_ln2_hi);
78 float16x8_t vt2 = vfmaq_f16(vx2, vn2, vminus_ln2_hi);
79 float16x8_t vt3 = vfmaq_f16(vx3, vn3, vminus_ln2_hi);
80 float16x8_t vt4 = vfmaq_f16(vx4, vn4, vminus_ln2_hi);
81 float16x8_t vt5 = vfmaq_f16(vx5, vn5, vminus_ln2_hi);
82
83 vt0 = vfmaq_f16(vt0, vn0, vminus_ln2_lo);
84 vt1 = vfmaq_f16(vt1, vn1, vminus_ln2_lo);
85 vt2 = vfmaq_f16(vt2, vn2, vminus_ln2_lo);
86 vt3 = vfmaq_f16(vt3, vn3, vminus_ln2_lo);
87 vt4 = vfmaq_f16(vt4, vn4, vminus_ln2_lo);
88 vt5 = vfmaq_f16(vt5, vn5, vminus_ln2_lo);
89
90 const float16x8_t vp0 = vfmaq_f16(vc1, vc2, vt0);
91 const float16x8_t vp1 = vfmaq_f16(vc1, vc2, vt1);
92 const float16x8_t vp2 = vfmaq_f16(vc1, vc2, vt2);
93 const float16x8_t vp3 = vfmaq_f16(vc1, vc2, vt3);
94 const float16x8_t vp4 = vfmaq_f16(vc1, vc2, vt4);
95 const float16x8_t vp5 = vfmaq_f16(vc1, vc2, vt5);
96
97 vt0 = vmulq_f16(vt0, vs0);
98 vt1 = vmulq_f16(vt1, vs1);
99 vt2 = vmulq_f16(vt2, vs2);
100 vt3 = vmulq_f16(vt3, vs3);
101 vt4 = vmulq_f16(vt4, vs4);
102 vt5 = vmulq_f16(vt5, vs5);
103
104 float16x8_t vf0 = vfmaq_f16(vs0, vp0, vt0);
105 const uint16x8_t vm0 = vcltq_f16(vx0, vdenorm_cutoff);
106 float16x8_t vf1 = vfmaq_f16(vs1, vp1, vt1);
107 const uint16x8_t vm1 = vcltq_f16(vx1, vdenorm_cutoff);
108 float16x8_t vf2 = vfmaq_f16(vs2, vp2, vt2);
109 const uint16x8_t vm2 = vcltq_f16(vx2, vdenorm_cutoff);
110 float16x8_t vf3 = vfmaq_f16(vs3, vp3, vt3);
111 const uint16x8_t vm3 = vcltq_f16(vx3, vdenorm_cutoff);
112 float16x8_t vf4 = vfmaq_f16(vs4, vp4, vt4);
113 const uint16x8_t vm4 = vcltq_f16(vx4, vdenorm_cutoff);
114 float16x8_t vf5 = vfmaq_f16(vs5, vp5, vt5);
115 const uint16x8_t vm5 = vcltq_f16(vx5, vdenorm_cutoff);
116
117 vf0 = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf0), vm0));
118 vf1 = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf1), vm1));
119 vf2 = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf2), vm2));
120 vf3 = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf3), vm3));
121 vf4 = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf4), vm4));
122 vf5 = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf5), vm5));
123
124 vst1q_f16(o, vf0); o += 8;
125 vst1q_f16(o, vf1); o += 8;
126 vst1q_f16(o, vf2); o += 8;
127 vst1q_f16(o, vf3); o += 8;
128 vst1q_f16(o, vf4); o += 8;
129 vst1q_f16(o, vf5); o += 8;
130
131 vacc0 = vaddq_f16(vacc0, vf0);
132 vacc0 = vaddq_f16(vacc0, vf1);
133 vacc0 = vaddq_f16(vacc0, vf2);
134 vacc0 = vaddq_f16(vacc0, vf3);
135 vacc0 = vaddq_f16(vacc0, vf4);
136 vacc0 = vaddq_f16(vacc0, vf5);
137 }
138
139 float16x8_t vacc = vacc0;
140 for (; batch >= 8 * sizeof(__fp16); batch -= 8 * sizeof(__fp16)) {
141 const float16x8_t vi = vld1q_f16(i); i += 8;
142
143 const float16x8_t vx = vsubq_f16(vi, vi_max);
144
145 float16x8_t vn = vfmaq_f16(vmagic_bias, vx, vlog2e);
146 const float16x8_t vs = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn), 10));
147 vn = vsubq_f16(vn, vmagic_bias);
148
149 float16x8_t vt = vfmaq_f16(vx, vn, vminus_ln2_hi);
150 vt = vfmaq_f16(vt, vn, vminus_ln2_lo);
151
152 const float16x8_t vp = vfmaq_f16(vc1, vc2, vt);
153 vt = vmulq_f16(vt, vs);
154
155 float16x8_t vf = vfmaq_f16(vs, vp, vt);
156 const uint16x8_t vm = vcltq_f16(vx, vdenorm_cutoff);
157 vf = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf), vm));
158
159 vst1q_f16(o, vf); o += 8;
160
161 vacc = vaddq_f16(vacc, vf);
162 }
163 float16x4_t vacc_lo = vadd_f16(vget_low_f16(vacc), vget_high_f16(vacc));
164 if (batch != 0) {
165 assert(batch >= 1 * sizeof(__fp16));
166 assert(batch <= 7 * sizeof(__fp16));
167 const float16x8_t vi = vld1q_f16(i);
168
169 const float16x8_t vx = vsubq_f16(vi, vi_max);
170
171 float16x8_t vn = vfmaq_f16(vmagic_bias, vx, vlog2e);
172 const float16x8_t vs = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn), 10));
173 vn = vsubq_f16(vn, vmagic_bias);
174
175 float16x8_t vt = vfmaq_f16(vx, vn, vminus_ln2_hi);
176 vt = vfmaq_f16(vt, vn, vminus_ln2_lo);
177
178 const float16x8_t vp = vfmaq_f16(vc1, vc2, vt);
179 vt = vmulq_f16(vt, vs);
180
181 float16x8_t vf = vfmaq_f16(vs, vp, vt);
182 const uint16x8_t vm = vcltq_f16(vx, vdenorm_cutoff);
183 vf = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf), vm));
184
185 float16x4_t vf_lo = vget_low_f16(vf);
186 if (batch & (4 * sizeof(__fp16))) {
187 vst1_f16(o, vf_lo); o += 4;
188 vacc_lo = vadd_f16(vacc_lo, vf_lo);
189 vf_lo = vget_high_f16(vf);
190 }
191 if (batch & (2 * sizeof(__fp16))) {
192 vst1_lane_u32((void*) o, vreinterpret_u32_f16(vf_lo), 0); o += 2;
193 vacc_lo = vadd_f16(vacc_lo, vreinterpret_f16_u64(vshl_n_u64(vreinterpret_u64_f16(vf_lo), 32)));
194 vf_lo = vext_f16(vf_lo, vf_lo, 2);
195 }
196 if (batch & (1 * sizeof(__fp16))) {
197 vst1_lane_f16(o, vf_lo, 0);
198 vacc_lo = vadd_f16(vacc_lo, vreinterpret_f16_u64(vshl_n_u64(vreinterpret_u64_f16(vf_lo), 48)));
199 }
200 }
201 vacc_lo = vpadd_f16(vacc_lo, vacc_lo);
202 *((__fp16*) sum) = vget_lane_f16(vacc_lo, 0) + vget_lane_f16(vacc_lo, 1);
203 }
204