1 // Auto-generated file. Do not edit!
2 // Template: src/f16-vsigmoid/neonfp16arith.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/vunary.h>
16
17
xnn_f16_vsigmoid_ukernel__neonfp16arith_rr2_p2_div_x40(size_t batch,const void * input,void * output,const union xnn_f16_sigmoid_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_f16_vsigmoid_ukernel__neonfp16arith_rr2_p2_div_x40(
19 size_t batch,
20 const void* input,
21 void* output,
22 const union xnn_f16_sigmoid_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
23 {
24 assert(batch % sizeof(__fp16) == 0);
25
26 const float16x8_t vmagic_bias = vreinterpretq_f16_u16(vld1q_dup_u16(¶ms->neonfp16arith_rr2_p2.magic_bias));
27 const float16x8_t vminus_log2e = vreinterpretq_f16_u16(vld1q_dup_u16(¶ms->neonfp16arith_rr2_p2.minus_log2e));
28 const float16x8_t vln2_hi = vreinterpretq_f16_u16(vld1q_dup_u16(¶ms->neonfp16arith_rr2_p2.ln2_hi));
29 const float16x8_t vln2_lo = vreinterpretq_f16_u16(vld1q_dup_u16(¶ms->neonfp16arith_rr2_p2.ln2_lo));
30 const float16x8_t vc2 = vreinterpretq_f16_u16(vld1q_dup_u16(¶ms->neonfp16arith_rr2_p2.c2));
31 const float16x8_t vc1 = vreinterpretq_f16_u16(vld1q_dup_u16(¶ms->neonfp16arith_rr2_p2.c1));
32 const float16x8_t vone = vmovq_n_f16(1.0f);
33 const float16x8_t vdenorm_cutoff = vreinterpretq_f16_u16(vld1q_dup_u16(¶ms->neonfp16arith_rr2_p2.denorm_cutoff));
34
35 const __fp16* i = (const __fp16*) input;
36 __fp16* o = (__fp16*) output;
37 for (; batch >= 40 * sizeof(__fp16); batch -= 40 * sizeof(__fp16)) {
38 const float16x8_t vx0 = vld1q_f16(i); i += 8;
39 const float16x8_t vx1 = vld1q_f16(i); i += 8;
40 const float16x8_t vx2 = vld1q_f16(i); i += 8;
41 const float16x8_t vx3 = vld1q_f16(i); i += 8;
42 const float16x8_t vx4 = vld1q_f16(i); i += 8;
43
44 const float16x8_t vz0 = vabsq_f16(vx0);
45 const float16x8_t vz1 = vabsq_f16(vx1);
46 const float16x8_t vz2 = vabsq_f16(vx2);
47 const float16x8_t vz3 = vabsq_f16(vx3);
48 const float16x8_t vz4 = vabsq_f16(vx4);
49
50 float16x8_t vn0 = vfmaq_f16(vmagic_bias, vz0, vminus_log2e);
51 float16x8_t vn1 = vfmaq_f16(vmagic_bias, vz1, vminus_log2e);
52 float16x8_t vn2 = vfmaq_f16(vmagic_bias, vz2, vminus_log2e);
53 float16x8_t vn3 = vfmaq_f16(vmagic_bias, vz3, vminus_log2e);
54 float16x8_t vn4 = vfmaq_f16(vmagic_bias, vz4, vminus_log2e);
55
56 const float16x8_t vs0 = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn0), 10));
57 const float16x8_t vs1 = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn1), 10));
58 const float16x8_t vs2 = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn2), 10));
59 const float16x8_t vs3 = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn3), 10));
60 const float16x8_t vs4 = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn4), 10));
61
62 vn0 = vsubq_f16(vn0, vmagic_bias);
63 vn1 = vsubq_f16(vn1, vmagic_bias);
64 vn2 = vsubq_f16(vn2, vmagic_bias);
65 vn3 = vsubq_f16(vn3, vmagic_bias);
66 vn4 = vsubq_f16(vn4, vmagic_bias);
67
68 float16x8_t vt0 = vfmaq_f16(vz0, vn0, vln2_hi);
69 float16x8_t vt1 = vfmaq_f16(vz1, vn1, vln2_hi);
70 float16x8_t vt2 = vfmaq_f16(vz2, vn2, vln2_hi);
71 float16x8_t vt3 = vfmaq_f16(vz3, vn3, vln2_hi);
72 float16x8_t vt4 = vfmaq_f16(vz4, vn4, vln2_hi);
73
74 vt0 = vfmaq_f16(vt0, vn0, vln2_lo);
75 vt1 = vfmaq_f16(vt1, vn1, vln2_lo);
76 vt2 = vfmaq_f16(vt2, vn2, vln2_lo);
77 vt3 = vfmaq_f16(vt3, vn3, vln2_lo);
78 vt4 = vfmaq_f16(vt4, vn4, vln2_lo);
79
80 const float16x8_t vp0 = vfmaq_f16(vc1, vc2, vt0);
81 const float16x8_t vp1 = vfmaq_f16(vc1, vc2, vt1);
82 const float16x8_t vp2 = vfmaq_f16(vc1, vc2, vt2);
83 const float16x8_t vp3 = vfmaq_f16(vc1, vc2, vt3);
84 const float16x8_t vp4 = vfmaq_f16(vc1, vc2, vt4);
85
86 vt0 = vmulq_f16(vt0, vs0);
87 vt1 = vmulq_f16(vt1, vs1);
88 vt2 = vmulq_f16(vt2, vs2);
89 vt3 = vmulq_f16(vt3, vs3);
90 vt4 = vmulq_f16(vt4, vs4);
91
92 const float16x8_t ve0 = vfmaq_f16(vs0, vp0, vt0);
93 const float16x8_t ve1 = vfmaq_f16(vs1, vp1, vt1);
94 const float16x8_t ve2 = vfmaq_f16(vs2, vp2, vt2);
95 const float16x8_t ve3 = vfmaq_f16(vs3, vp3, vt3);
96 const float16x8_t ve4 = vfmaq_f16(vs4, vp4, vt4);
97
98 const float16x8_t vd0 = vaddq_f16(ve0, vone);
99 const float16x8_t vd1 = vaddq_f16(ve1, vone);
100 const float16x8_t vd2 = vaddq_f16(ve2, vone);
101 const float16x8_t vd3 = vaddq_f16(ve3, vone);
102 const float16x8_t vd4 = vaddq_f16(ve4, vone);
103
104 float16x8_t vf0 = vdivq_f16(ve0, vd0);
105 float16x8_t vf1 = vdivq_f16(ve1, vd1);
106 float16x8_t vf2 = vdivq_f16(ve2, vd2);
107 float16x8_t vf3 = vdivq_f16(ve3, vd3);
108 float16x8_t vf4 = vdivq_f16(ve4, vd4);
109
110 vf0 = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf0), vcagtq_f16(vx0, vdenorm_cutoff)));
111 vf1 = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf1), vcagtq_f16(vx1, vdenorm_cutoff)));
112 vf2 = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf2), vcagtq_f16(vx2, vdenorm_cutoff)));
113 vf3 = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf3), vcagtq_f16(vx3, vdenorm_cutoff)));
114 vf4 = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf4), vcagtq_f16(vx4, vdenorm_cutoff)));
115
116 const uint16x8_t vm0 = vcltq_f16(vx0, vmovq_n_f16(0.0f));
117 const uint16x8_t vm1 = vcltq_f16(vx1, vmovq_n_f16(0.0f));
118 const uint16x8_t vm2 = vcltq_f16(vx2, vmovq_n_f16(0.0f));
119 const uint16x8_t vm3 = vcltq_f16(vx3, vmovq_n_f16(0.0f));
120 const uint16x8_t vm4 = vcltq_f16(vx4, vmovq_n_f16(0.0f));
121
122 vf0 = vbslq_f16(vm0, vf0, vsubq_f16(vone, vf0));
123 vf1 = vbslq_f16(vm1, vf1, vsubq_f16(vone, vf1));
124 vf2 = vbslq_f16(vm2, vf2, vsubq_f16(vone, vf2));
125 vf3 = vbslq_f16(vm3, vf3, vsubq_f16(vone, vf3));
126 vf4 = vbslq_f16(vm4, vf4, vsubq_f16(vone, vf4));
127
128 vst1q_f16(o, vf0); o += 8;
129 vst1q_f16(o, vf1); o += 8;
130 vst1q_f16(o, vf2); o += 8;
131 vst1q_f16(o, vf3); o += 8;
132 vst1q_f16(o, vf4); o += 8;
133 }
134 for (; batch >= 8 * sizeof(__fp16); batch -= 8 * sizeof(__fp16)) {
135 const float16x8_t vx = vld1q_f16(i); i += 8;
136
137 const float16x8_t vz = vabsq_f16(vx);
138
139 float16x8_t vn = vfmaq_f16(vmagic_bias, vz, vminus_log2e);
140 const float16x8_t vs = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn), 10));
141 vn = vsubq_f16(vn, vmagic_bias);
142
143 float16x8_t vt = vfmaq_f16(vz, vn, vln2_hi);
144 vt = vfmaq_f16(vt, vn, vln2_lo);
145
146 const float16x8_t vp = vfmaq_f16(vc1, vc2, vt);
147 vt = vmulq_f16(vt, vs);
148 const float16x8_t ve = vfmaq_f16(vs, vp, vt);
149 const float16x8_t vd = vaddq_f16(ve, vone);
150
151 float16x8_t vf = vdivq_f16(ve, vd);
152 vf = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf), vcagtq_f16(vx, vdenorm_cutoff)));
153 const uint16x8_t vm = vcltq_f16(vx, vmovq_n_f16(0.0f));
154 vf = vbslq_f16(vm, vf, vsubq_f16(vone, vf));
155
156 vst1q_f16(o, vf); o += 8;
157 }
158 if XNN_UNLIKELY(batch != 0) {
159 const float16x8_t vx = vld1q_f16(i);
160
161 const float16x8_t vz = vabsq_f16(vx);
162
163 float16x8_t vn = vfmaq_f16(vmagic_bias, vz, vminus_log2e);
164 const float16x8_t vs = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn), 10));
165 vn = vsubq_f16(vn, vmagic_bias);
166
167 float16x8_t vt = vfmaq_f16(vz, vn, vln2_hi);
168 vt = vfmaq_f16(vt, vn, vln2_lo);
169
170 const float16x8_t vp = vfmaq_f16(vc1, vc2, vt);
171 vt = vmulq_f16(vt, vs);
172 const float16x8_t ve = vfmaq_f16(vs, vp, vt);
173 const float16x8_t vd = vaddq_f16(ve, vone);
174
175 float16x8_t vf = vdivq_f16(ve, vd);
176 vf = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf), vcagtq_f16(vx, vdenorm_cutoff)));
177 const uint16x8_t vm = vcltq_f16(vx, vmovq_n_f16(0.0f));
178 vf = vbslq_f16(vm, vf, vsubq_f16(vone, vf));
179
180 float16x4_t vf_lo = vget_low_f16(vf);
181 if (batch & (4 * sizeof(__fp16))) {
182 vst1_f16(o, vf_lo); o += 4;
183 vf_lo = vget_high_f16(vf);
184 }
185 if (batch & (2 * sizeof(__fp16))) {
186 vst1_f16(o, vf_lo); o += 2;
187 vf_lo = vext_f16(vf_lo, vf_lo, 2);
188 }
189 if (batch & (1 * sizeof(__fp16))) {
190 vst1_lane_f16(o, vf_lo, 0);
191 }
192 }
193 }
194