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_nr1recps_x8(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_nr1recps_x8(
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 >= 8 * sizeof(__fp16); batch -= 8 * sizeof(__fp16)) {
38 const float16x8_t vx = vld1q_f16(i); i += 8;
39
40 const float16x8_t vz = vabsq_f16(vx);
41
42 float16x8_t vn = vfmaq_f16(vmagic_bias, vz, vminus_log2e);
43 const float16x8_t vs = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn), 10));
44 vn = vsubq_f16(vn, vmagic_bias);
45
46 float16x8_t vt = vfmaq_f16(vz, vn, vln2_hi);
47 vt = vfmaq_f16(vt, vn, vln2_lo);
48
49 const float16x8_t vp = vfmaq_f16(vc1, vc2, vt);
50 vt = vmulq_f16(vt, vs);
51 const float16x8_t ve = vfmaq_f16(vs, vp, vt);
52 const float16x8_t vd = vaddq_f16(ve, vone);
53
54 float16x8_t vr = vrecpeq_f16(vd);
55 const float16x8_t vadj = vrecpsq_f16(vr, vd);
56 vr = vmulq_f16(vr, vadj);
57
58 float16x8_t vf = vmulq_f16(ve, vr);
59 vf = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf), vcagtq_f16(vx, vdenorm_cutoff)));
60 const uint16x8_t vm = vcltq_f16(vx, vmovq_n_f16(0.0f));
61 vf = vbslq_f16(vm, vf, vsubq_f16(vone, vf));
62
63 vst1q_f16(o, vf); o += 8;
64 }
65 if XNN_UNLIKELY(batch != 0) {
66 const float16x8_t vx = vld1q_f16(i);
67
68 const float16x8_t vz = vabsq_f16(vx);
69
70 float16x8_t vn = vfmaq_f16(vmagic_bias, vz, vminus_log2e);
71 const float16x8_t vs = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn), 10));
72 vn = vsubq_f16(vn, vmagic_bias);
73
74 float16x8_t vt = vfmaq_f16(vz, vn, vln2_hi);
75 vt = vfmaq_f16(vt, vn, vln2_lo);
76
77 const float16x8_t vp = vfmaq_f16(vc1, vc2, vt);
78 vt = vmulq_f16(vt, vs);
79 const float16x8_t ve = vfmaq_f16(vs, vp, vt);
80 const float16x8_t vd = vaddq_f16(ve, vone);
81
82 float16x8_t vr = vrecpeq_f16(vd);
83 const float16x8_t vadj = vrecpsq_f16(vr, vd);
84 vr = vmulq_f16(vr, vadj);
85
86 float16x8_t vf = vmulq_f16(ve, vr);
87 vf = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf), vcagtq_f16(vx, vdenorm_cutoff)));
88 const uint16x8_t vm = vcltq_f16(vx, vmovq_n_f16(0.0f));
89 vf = vbslq_f16(vm, vf, vsubq_f16(vone, vf));
90
91 float16x4_t vf_lo = vget_low_f16(vf);
92 if (batch & (4 * sizeof(__fp16))) {
93 vst1_f16(o, vf_lo); o += 4;
94 vf_lo = vget_high_f16(vf);
95 }
96 if (batch & (2 * sizeof(__fp16))) {
97 vst1_f16(o, vf_lo); o += 2;
98 vf_lo = vext_f16(vf_lo, vf_lo, 2);
99 }
100 if (batch & (1 * sizeof(__fp16))) {
101 vst1_lane_f16(o, vf_lo, 0);
102 }
103 }
104 }
105