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_nr1fma_x16(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_nr1fma_x16(
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 >= 16 * sizeof(__fp16); batch -= 16 * sizeof(__fp16)) {
38 const float16x8_t vx0 = vld1q_f16(i); i += 8;
39 const float16x8_t vx1 = vld1q_f16(i); i += 8;
40
41 const float16x8_t vz0 = vabsq_f16(vx0);
42 const float16x8_t vz1 = vabsq_f16(vx1);
43
44 float16x8_t vn0 = vfmaq_f16(vmagic_bias, vz0, vminus_log2e);
45 float16x8_t vn1 = vfmaq_f16(vmagic_bias, vz1, vminus_log2e);
46
47 const float16x8_t vs0 = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn0), 10));
48 const float16x8_t vs1 = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn1), 10));
49
50 vn0 = vsubq_f16(vn0, vmagic_bias);
51 vn1 = vsubq_f16(vn1, vmagic_bias);
52
53 float16x8_t vt0 = vfmaq_f16(vz0, vn0, vln2_hi);
54 float16x8_t vt1 = vfmaq_f16(vz1, vn1, vln2_hi);
55
56 vt0 = vfmaq_f16(vt0, vn0, vln2_lo);
57 vt1 = vfmaq_f16(vt1, vn1, vln2_lo);
58
59 const float16x8_t vp0 = vfmaq_f16(vc1, vc2, vt0);
60 const float16x8_t vp1 = vfmaq_f16(vc1, vc2, vt1);
61
62 vt0 = vmulq_f16(vt0, vs0);
63 vt1 = vmulq_f16(vt1, vs1);
64
65 const float16x8_t ve0 = vfmaq_f16(vs0, vp0, vt0);
66 const float16x8_t ve1 = vfmaq_f16(vs1, vp1, vt1);
67
68 const float16x8_t vd0 = vaddq_f16(ve0, vone);
69 const float16x8_t vd1 = vaddq_f16(ve1, vone);
70
71 float16x8_t vr0 = vrecpeq_f16(vd0);
72 float16x8_t vr1 = vrecpeq_f16(vd1);
73
74 const float16x8_t vadj0 = vfmsq_f16(vone, vr0, vd0);
75 const float16x8_t vadj1 = vfmsq_f16(vone, vr1, vd1);
76
77 vr0 = vfmaq_f16(vr0, vr0, vadj0);
78 vr1 = vfmaq_f16(vr1, vr1, vadj1);
79
80 float16x8_t vf0 = vmulq_f16(ve0, vr0);
81 float16x8_t vf1 = vmulq_f16(ve1, vr1);
82
83 vf0 = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf0), vcagtq_f16(vx0, vdenorm_cutoff)));
84 vf1 = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf1), vcagtq_f16(vx1, vdenorm_cutoff)));
85
86 const uint16x8_t vm0 = vcltq_f16(vx0, vmovq_n_f16(0.0f));
87 const uint16x8_t vm1 = vcltq_f16(vx1, vmovq_n_f16(0.0f));
88
89 vf0 = vbslq_f16(vm0, vf0, vsubq_f16(vone, vf0));
90 vf1 = vbslq_f16(vm1, vf1, vsubq_f16(vone, vf1));
91
92 vst1q_f16(o, vf0); o += 8;
93 vst1q_f16(o, vf1); o += 8;
94 }
95 for (; batch >= 8 * sizeof(__fp16); batch -= 8 * sizeof(__fp16)) {
96 const float16x8_t vx = vld1q_f16(i); i += 8;
97
98 const float16x8_t vz = vabsq_f16(vx);
99
100 float16x8_t vn = vfmaq_f16(vmagic_bias, vz, vminus_log2e);
101 const float16x8_t vs = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn), 10));
102 vn = vsubq_f16(vn, vmagic_bias);
103
104 float16x8_t vt = vfmaq_f16(vz, vn, vln2_hi);
105 vt = vfmaq_f16(vt, vn, vln2_lo);
106
107 const float16x8_t vp = vfmaq_f16(vc1, vc2, vt);
108 vt = vmulq_f16(vt, vs);
109 const float16x8_t ve = vfmaq_f16(vs, vp, vt);
110 const float16x8_t vd = vaddq_f16(ve, vone);
111
112 float16x8_t vr = vrecpeq_f16(vd);
113 const float16x8_t vadj = vfmsq_f16(vone, vr, vd);
114 vr = vfmaq_f16(vr, vr, vadj);
115
116 float16x8_t vf = vmulq_f16(ve, vr);
117 vf = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf), vcagtq_f16(vx, vdenorm_cutoff)));
118 const uint16x8_t vm = vcltq_f16(vx, vmovq_n_f16(0.0f));
119 vf = vbslq_f16(vm, vf, vsubq_f16(vone, vf));
120
121 vst1q_f16(o, vf); o += 8;
122 }
123 if XNN_UNLIKELY(batch != 0) {
124 const float16x8_t vx = vld1q_f16(i);
125
126 const float16x8_t vz = vabsq_f16(vx);
127
128 float16x8_t vn = vfmaq_f16(vmagic_bias, vz, vminus_log2e);
129 const float16x8_t vs = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn), 10));
130 vn = vsubq_f16(vn, vmagic_bias);
131
132 float16x8_t vt = vfmaq_f16(vz, vn, vln2_hi);
133 vt = vfmaq_f16(vt, vn, vln2_lo);
134
135 const float16x8_t vp = vfmaq_f16(vc1, vc2, vt);
136 vt = vmulq_f16(vt, vs);
137 const float16x8_t ve = vfmaq_f16(vs, vp, vt);
138 const float16x8_t vd = vaddq_f16(ve, vone);
139
140 float16x8_t vr = vrecpeq_f16(vd);
141 const float16x8_t vadj = vfmsq_f16(vone, vr, vd);
142 vr = vfmaq_f16(vr, vr, vadj);
143
144 float16x8_t vf = vmulq_f16(ve, vr);
145 vf = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf), vcagtq_f16(vx, vdenorm_cutoff)));
146 const uint16x8_t vm = vcltq_f16(vx, vmovq_n_f16(0.0f));
147 vf = vbslq_f16(vm, vf, vsubq_f16(vone, vf));
148
149 float16x4_t vf_lo = vget_low_f16(vf);
150 if (batch & (4 * sizeof(__fp16))) {
151 vst1_f16(o, vf_lo); o += 4;
152 vf_lo = vget_high_f16(vf);
153 }
154 if (batch & (2 * sizeof(__fp16))) {
155 vst1_f16(o, vf_lo); o += 2;
156 vf_lo = vext_f16(vf_lo, vf_lo, 2);
157 }
158 if (batch & (1 * sizeof(__fp16))) {
159 vst1_lane_f16(o, vf_lo, 0);
160 }
161 }
162 }
163