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_x40(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_x40(
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 >= 40 * sizeof(__fp16); batch -= 40 * 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
47 const float16x8_t vx0 = vsubq_f16(vi0, vi_max);
48 const float16x8_t vx1 = vsubq_f16(vi1, vi_max);
49 const float16x8_t vx2 = vsubq_f16(vi2, vi_max);
50 const float16x8_t vx3 = vsubq_f16(vi3, vi_max);
51 const float16x8_t vx4 = vsubq_f16(vi4, vi_max);
52
53 float16x8_t vn0 = vfmaq_f16(vmagic_bias, vx0, vlog2e);
54 float16x8_t vn1 = vfmaq_f16(vmagic_bias, vx1, vlog2e);
55 float16x8_t vn2 = vfmaq_f16(vmagic_bias, vx2, vlog2e);
56 float16x8_t vn3 = vfmaq_f16(vmagic_bias, vx3, vlog2e);
57 float16x8_t vn4 = vfmaq_f16(vmagic_bias, vx4, vlog2e);
58
59 const float16x8_t vs0 = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn0), 10));
60 const float16x8_t vs1 = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn1), 10));
61 const float16x8_t vs2 = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn2), 10));
62 const float16x8_t vs3 = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn3), 10));
63 const float16x8_t vs4 = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn4), 10));
64
65 vn0 = vsubq_f16(vn0, vmagic_bias);
66 vn1 = vsubq_f16(vn1, vmagic_bias);
67 vn2 = vsubq_f16(vn2, vmagic_bias);
68 vn3 = vsubq_f16(vn3, vmagic_bias);
69 vn4 = vsubq_f16(vn4, vmagic_bias);
70
71 float16x8_t vt0 = vfmaq_f16(vx0, vn0, vminus_ln2_hi);
72 float16x8_t vt1 = vfmaq_f16(vx1, vn1, vminus_ln2_hi);
73 float16x8_t vt2 = vfmaq_f16(vx2, vn2, vminus_ln2_hi);
74 float16x8_t vt3 = vfmaq_f16(vx3, vn3, vminus_ln2_hi);
75 float16x8_t vt4 = vfmaq_f16(vx4, vn4, vminus_ln2_hi);
76
77 vt0 = vfmaq_f16(vt0, vn0, vminus_ln2_lo);
78 vt1 = vfmaq_f16(vt1, vn1, vminus_ln2_lo);
79 vt2 = vfmaq_f16(vt2, vn2, vminus_ln2_lo);
80 vt3 = vfmaq_f16(vt3, vn3, vminus_ln2_lo);
81 vt4 = vfmaq_f16(vt4, vn4, vminus_ln2_lo);
82
83 const float16x8_t vp0 = vfmaq_f16(vc1, vc2, vt0);
84 const float16x8_t vp1 = vfmaq_f16(vc1, vc2, vt1);
85 const float16x8_t vp2 = vfmaq_f16(vc1, vc2, vt2);
86 const float16x8_t vp3 = vfmaq_f16(vc1, vc2, vt3);
87 const float16x8_t vp4 = vfmaq_f16(vc1, vc2, vt4);
88
89 vt0 = vmulq_f16(vt0, vs0);
90 vt1 = vmulq_f16(vt1, vs1);
91 vt2 = vmulq_f16(vt2, vs2);
92 vt3 = vmulq_f16(vt3, vs3);
93 vt4 = vmulq_f16(vt4, vs4);
94
95 float16x8_t vf0 = vfmaq_f16(vs0, vp0, vt0);
96 const uint16x8_t vm0 = vcltq_f16(vx0, vdenorm_cutoff);
97 float16x8_t vf1 = vfmaq_f16(vs1, vp1, vt1);
98 const uint16x8_t vm1 = vcltq_f16(vx1, vdenorm_cutoff);
99 float16x8_t vf2 = vfmaq_f16(vs2, vp2, vt2);
100 const uint16x8_t vm2 = vcltq_f16(vx2, vdenorm_cutoff);
101 float16x8_t vf3 = vfmaq_f16(vs3, vp3, vt3);
102 const uint16x8_t vm3 = vcltq_f16(vx3, vdenorm_cutoff);
103 float16x8_t vf4 = vfmaq_f16(vs4, vp4, vt4);
104 const uint16x8_t vm4 = vcltq_f16(vx4, vdenorm_cutoff);
105
106 vf0 = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf0), vm0));
107 vf1 = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf1), vm1));
108 vf2 = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf2), vm2));
109 vf3 = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf3), vm3));
110 vf4 = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf4), vm4));
111
112 vst1q_f16(o, vf0); o += 8;
113 vst1q_f16(o, vf1); o += 8;
114 vst1q_f16(o, vf2); o += 8;
115 vst1q_f16(o, vf3); o += 8;
116 vst1q_f16(o, vf4); o += 8;
117
118 vacc0 = vaddq_f16(vacc0, vf0);
119 vacc0 = vaddq_f16(vacc0, vf1);
120 vacc0 = vaddq_f16(vacc0, vf2);
121 vacc0 = vaddq_f16(vacc0, vf3);
122 vacc0 = vaddq_f16(vacc0, vf4);
123 }
124
125 float16x8_t vacc = vacc0;
126 for (; batch >= 8 * sizeof(__fp16); batch -= 8 * sizeof(__fp16)) {
127 const float16x8_t vi = vld1q_f16(i); i += 8;
128
129 const float16x8_t vx = vsubq_f16(vi, vi_max);
130
131 float16x8_t vn = vfmaq_f16(vmagic_bias, vx, vlog2e);
132 const float16x8_t vs = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn), 10));
133 vn = vsubq_f16(vn, vmagic_bias);
134
135 float16x8_t vt = vfmaq_f16(vx, vn, vminus_ln2_hi);
136 vt = vfmaq_f16(vt, vn, vminus_ln2_lo);
137
138 const float16x8_t vp = vfmaq_f16(vc1, vc2, vt);
139 vt = vmulq_f16(vt, vs);
140
141 float16x8_t vf = vfmaq_f16(vs, vp, vt);
142 const uint16x8_t vm = vcltq_f16(vx, vdenorm_cutoff);
143 vf = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf), vm));
144
145 vst1q_f16(o, vf); o += 8;
146
147 vacc = vaddq_f16(vacc, vf);
148 }
149 float16x4_t vacc_lo = vadd_f16(vget_low_f16(vacc), vget_high_f16(vacc));
150 if (batch != 0) {
151 assert(batch >= 1 * sizeof(__fp16));
152 assert(batch <= 7 * sizeof(__fp16));
153 const float16x8_t vi = vld1q_f16(i);
154
155 const float16x8_t vx = vsubq_f16(vi, vi_max);
156
157 float16x8_t vn = vfmaq_f16(vmagic_bias, vx, vlog2e);
158 const float16x8_t vs = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn), 10));
159 vn = vsubq_f16(vn, vmagic_bias);
160
161 float16x8_t vt = vfmaq_f16(vx, vn, vminus_ln2_hi);
162 vt = vfmaq_f16(vt, vn, vminus_ln2_lo);
163
164 const float16x8_t vp = vfmaq_f16(vc1, vc2, vt);
165 vt = vmulq_f16(vt, vs);
166
167 float16x8_t vf = vfmaq_f16(vs, vp, vt);
168 const uint16x8_t vm = vcltq_f16(vx, vdenorm_cutoff);
169 vf = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf), vm));
170
171 float16x4_t vf_lo = vget_low_f16(vf);
172 if (batch & (4 * sizeof(__fp16))) {
173 vst1_f16(o, vf_lo); o += 4;
174 vacc_lo = vadd_f16(vacc_lo, vf_lo);
175 vf_lo = vget_high_f16(vf);
176 }
177 if (batch & (2 * sizeof(__fp16))) {
178 vst1_lane_u32((void*) o, vreinterpret_u32_f16(vf_lo), 0); o += 2;
179 vacc_lo = vadd_f16(vacc_lo, vreinterpret_f16_u64(vshl_n_u64(vreinterpret_u64_f16(vf_lo), 32)));
180 vf_lo = vext_f16(vf_lo, vf_lo, 2);
181 }
182 if (batch & (1 * sizeof(__fp16))) {
183 vst1_lane_f16(o, vf_lo, 0);
184 vacc_lo = vadd_f16(vacc_lo, vreinterpret_f16_u64(vshl_n_u64(vreinterpret_u64_f16(vf_lo), 48)));
185 }
186 }
187 vacc_lo = vpadd_f16(vacc_lo, vacc_lo);
188 *((__fp16*) sum) = vget_lane_f16(vacc_lo, 0) + vget_lane_f16(vacc_lo, 1);
189 }
190