File size: 6,827 Bytes
8b7c501
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
// Auto-generated file. Do not edit!
//   Template: src/f16-vsigmoid/neonfp16arith.c.in
//   Generator: tools/xngen
//
// Copyright 2022 Google LLC
//
// This source code is licensed under the BSD-style license found in the
// LICENSE file in the root directory of this source tree.

#include <assert.h>

#include <arm_neon.h>

#include <xnnpack/common.h>
#include <xnnpack/vunary.h>


void xnn_f16_vsigmoid_ukernel__aarch64_neonfp16arith_rr2_p2_div_x24(
    size_t batch,
    const void* input,
    void* output,
    const union xnn_f16_sigmoid_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
{
  assert(batch != 0);
  assert(batch % sizeof(uint16_t) == 0);
  assert(input != NULL);
  assert(output != NULL);

  const float16x8_t vmagic_bias = vreinterpretq_f16_u16(vld1q_dup_u16(&params->fp16arith_rr2_p2.magic_bias));
  const float16x8_t vminus_log2e = vreinterpretq_f16_u16(vld1q_dup_u16(&params->fp16arith_rr2_p2.minus_log2e));
  const float16x8_t vln2_hi = vreinterpretq_f16_u16(vld1q_dup_u16(&params->fp16arith_rr2_p2.ln2_hi));
  const float16x8_t vln2_lo = vreinterpretq_f16_u16(vld1q_dup_u16(&params->fp16arith_rr2_p2.ln2_lo));
  const float16x8_t vc2 = vreinterpretq_f16_u16(vld1q_dup_u16(&params->fp16arith_rr2_p2.c2));
  const float16x8_t vc1 = vreinterpretq_f16_u16(vld1q_dup_u16(&params->fp16arith_rr2_p2.c1));
  const float16x8_t vone = vreinterpretq_f16_u16(vmovq_n_u16(UINT16_C(0x3C00)));  // 1.0h
  const float16x8_t vdenorm_cutoff = vreinterpretq_f16_u16(vld1q_dup_u16(&params->fp16arith_rr2_p2.denorm_cutoff));

  const uint16_t* i = (const uint16_t*) input;
  uint16_t* o = (uint16_t*) output;
  for (; batch >= 24 * sizeof(uint16_t); batch -= 24 * sizeof(uint16_t)) {
    const float16x8_t vx0 = vreinterpretq_f16_u16(vld1q_u16(i)); i += 8;
    const float16x8_t vx1 = vreinterpretq_f16_u16(vld1q_u16(i)); i += 8;
    const float16x8_t vx2 = vreinterpretq_f16_u16(vld1q_u16(i)); i += 8;

    const float16x8_t vz0 = vabsq_f16(vx0);
    const float16x8_t vz1 = vabsq_f16(vx1);
    const float16x8_t vz2 = vabsq_f16(vx2);

    float16x8_t vn0 = vfmaq_f16(vmagic_bias, vz0, vminus_log2e);
    float16x8_t vn1 = vfmaq_f16(vmagic_bias, vz1, vminus_log2e);
    float16x8_t vn2 = vfmaq_f16(vmagic_bias, vz2, vminus_log2e);

    const float16x8_t vs0 = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn0), 10));
    const float16x8_t vs1 = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn1), 10));
    const float16x8_t vs2 = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn2), 10));

    vn0 = vsubq_f16(vn0, vmagic_bias);
    vn1 = vsubq_f16(vn1, vmagic_bias);
    vn2 = vsubq_f16(vn2, vmagic_bias);

    float16x8_t vt0 = vfmaq_f16(vz0, vn0, vln2_hi);
    float16x8_t vt1 = vfmaq_f16(vz1, vn1, vln2_hi);
    float16x8_t vt2 = vfmaq_f16(vz2, vn2, vln2_hi);

    vt0 = vfmaq_f16(vt0, vn0, vln2_lo);
    vt1 = vfmaq_f16(vt1, vn1, vln2_lo);
    vt2 = vfmaq_f16(vt2, vn2, vln2_lo);

    const float16x8_t vp0 = vfmaq_f16(vc1, vc2, vt0);
    const float16x8_t vp1 = vfmaq_f16(vc1, vc2, vt1);
    const float16x8_t vp2 = vfmaq_f16(vc1, vc2, vt2);

    vt0 = vmulq_f16(vt0, vs0);
    vt1 = vmulq_f16(vt1, vs1);
    vt2 = vmulq_f16(vt2, vs2);

    const float16x8_t ve0 = vfmaq_f16(vs0, vp0, vt0);
    const float16x8_t ve1 = vfmaq_f16(vs1, vp1, vt1);
    const float16x8_t ve2 = vfmaq_f16(vs2, vp2, vt2);

    const float16x8_t vd0 = vaddq_f16(ve0, vone);
    const float16x8_t vd1 = vaddq_f16(ve1, vone);
    const float16x8_t vd2 = vaddq_f16(ve2, vone);

    float16x8_t vf0 = vdivq_f16(ve0, vd0);
    float16x8_t vf1 = vdivq_f16(ve1, vd1);
    float16x8_t vf2 = vdivq_f16(ve2, vd2);

    vf0 = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf0), vcagtq_f16(vx0, vdenorm_cutoff)));
    vf1 = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf1), vcagtq_f16(vx1, vdenorm_cutoff)));
    vf2 = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf2), vcagtq_f16(vx2, vdenorm_cutoff)));

    const uint16x8_t vm0 = vcltq_f16(vx0, vreinterpretq_f16_u16(vmovq_n_u16(0)));
    const uint16x8_t vm1 = vcltq_f16(vx1, vreinterpretq_f16_u16(vmovq_n_u16(0)));
    const uint16x8_t vm2 = vcltq_f16(vx2, vreinterpretq_f16_u16(vmovq_n_u16(0)));

    vf0 = vbslq_f16(vm0, vf0, vsubq_f16(vone, vf0));
    vf1 = vbslq_f16(vm1, vf1, vsubq_f16(vone, vf1));
    vf2 = vbslq_f16(vm2, vf2, vsubq_f16(vone, vf2));

    vst1q_u16(o, vreinterpretq_u16_f16(vf0)); o += 8;
    vst1q_u16(o, vreinterpretq_u16_f16(vf1)); o += 8;
    vst1q_u16(o, vreinterpretq_u16_f16(vf2)); o += 8;
  }
  for (; batch >= 8 * sizeof(uint16_t); batch -= 8 * sizeof(uint16_t)) {
    const float16x8_t vx = vreinterpretq_f16_u16(vld1q_u16(i)); i += 8;

    const float16x8_t vz = vabsq_f16(vx);

    float16x8_t vn = vfmaq_f16(vmagic_bias, vz, vminus_log2e);
    const float16x8_t vs = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn), 10));
    vn = vsubq_f16(vn, vmagic_bias);

    float16x8_t vt = vfmaq_f16(vz, vn, vln2_hi);
    vt = vfmaq_f16(vt, vn, vln2_lo);

    const float16x8_t vp = vfmaq_f16(vc1, vc2, vt);
    vt = vmulq_f16(vt, vs);
    const float16x8_t ve = vfmaq_f16(vs, vp, vt);
    const float16x8_t vd = vaddq_f16(ve, vone);

    float16x8_t vf = vdivq_f16(ve, vd);
    vf = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf), vcagtq_f16(vx, vdenorm_cutoff)));
    const uint16x8_t vm = vcltq_f16(vx, vreinterpretq_f16_u16(vmovq_n_u16(0)));
    vf = vbslq_f16(vm, vf, vsubq_f16(vone, vf));

    vst1q_u16(o, vreinterpretq_u16_f16(vf)); o += 8;
  }
  if XNN_UNLIKELY(batch != 0) {
    const float16x8_t vx = vreinterpretq_f16_u16(vld1q_u16(i));

    const float16x8_t vz = vabsq_f16(vx);

    float16x8_t vn = vfmaq_f16(vmagic_bias, vz, vminus_log2e);
    const float16x8_t vs = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn), 10));
    vn = vsubq_f16(vn, vmagic_bias);

    float16x8_t vt = vfmaq_f16(vz, vn, vln2_hi);
    vt = vfmaq_f16(vt, vn, vln2_lo);

    const float16x8_t vp = vfmaq_f16(vc1, vc2, vt);
    vt = vmulq_f16(vt, vs);
    const float16x8_t ve = vfmaq_f16(vs, vp, vt);
    const float16x8_t vd = vaddq_f16(ve, vone);

    float16x8_t vf = vdivq_f16(ve, vd);
    vf = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf), vcagtq_f16(vx, vdenorm_cutoff)));
    const uint16x8_t vm = vcltq_f16(vx, vreinterpretq_f16_u16(vmovq_n_u16(0)));
    vf = vbslq_f16(vm, vf, vsubq_f16(vone, vf));

    float16x4_t vf_lo = vget_low_f16(vf);
    if (batch & (4 * sizeof(uint16_t))) {
      vst1_u16(o, vreinterpret_u16_f16(vf_lo)); o += 4;
      vf_lo = vget_high_f16(vf);
    }
    if (batch & (2 * sizeof(uint16_t))) {
      vst1_lane_u32((void*) o, vreinterpret_u32_f16(vf_lo), 0); o += 2;
      vf_lo = vext_f16(vf_lo, vf_lo, 2);
    }
    if (batch & (1 * sizeof(uint16_t))) {
      vst1_lane_u16(o, vreinterpret_u16_f16(vf_lo), 0);
    }
  }
}