File size: 8,705 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
// Auto-generated file. Do not edit!
//   Template: src/f16-f32-vcvt/neon-int32.c.in
//   Generator: tools/xngen
//
// Copyright 2021 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/vcvt.h>


void xnn_f16_f32_vcvt_ukernel__neon_int32_x24(
    size_t batch,
    const void* input,
    float* output,
    const union xnn_f16_f32_cvt_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 uint32x4_t vsign_mask = vmovq_n_u32(0x80000000);
  const uint32x4_t vexp_offset = vmovq_n_u32(0x70000000);
  const float32x4_t vexp_scale = vld1q_dup_f32(&params->neon.exp_scale);
  const uint32x4_t vmagic_bias = vmovq_n_u32(0x3F000000);
  const uint32x4_t vdenorm_cutoff = vmovq_n_u32(0x04000000);

  const uint16_t* i = (const uint16_t*) input;
  for (; batch >= 24 * sizeof(uint16_t); batch -= 24 * sizeof(uint16_t)) {
    const uint16x8_t vh0 = vld1q_u16(i); i += 8;
    const uint16x8_t vh1 = vld1q_u16(i); i += 8;
    const uint16x8_t vh2 = vld1q_u16(i); i += 8;

    const uint32x4_t vw0 = vshll_n_u16(vget_low_u16(vh0), 16);
    const uint32x4_t vw1 = vshll_n_u16(vget_high_u16(vh0), 16);
    const uint32x4_t vw2 = vshll_n_u16(vget_low_u16(vh1), 16);
    const uint32x4_t vw3 = vshll_n_u16(vget_high_u16(vh1), 16);
    const uint32x4_t vw4 = vshll_n_u16(vget_low_u16(vh2), 16);
    const uint32x4_t vw5 = vshll_n_u16(vget_high_u16(vh2), 16);

    const uint32x4_t vsign0 = vandq_u32(vw0, vsign_mask);
    const uint32x4_t vsign1 = vandq_u32(vw1, vsign_mask);
    const uint32x4_t vsign2 = vandq_u32(vw2, vsign_mask);
    const uint32x4_t vsign3 = vandq_u32(vw3, vsign_mask);
    const uint32x4_t vsign4 = vandq_u32(vw4, vsign_mask);
    const uint32x4_t vsign5 = vandq_u32(vw5, vsign_mask);

    const uint32x4_t vnonsign0 = veorq_u32(vw0, vsign0);
    const uint32x4_t vnonsign1 = veorq_u32(vw1, vsign1);
    const uint32x4_t vnonsign2 = veorq_u32(vw2, vsign2);
    const uint32x4_t vnonsign3 = veorq_u32(vw3, vsign3);
    const uint32x4_t vnonsign4 = veorq_u32(vw4, vsign4);
    const uint32x4_t vnonsign5 = veorq_u32(vw5, vsign5);

    const float32x4_t vnorm0 = vmulq_f32(vreinterpretq_f32_u32(vsraq_n_u32(vexp_offset, vnonsign0, 3)), vexp_scale);
    const float32x4_t vnorm1 = vmulq_f32(vreinterpretq_f32_u32(vsraq_n_u32(vexp_offset, vnonsign1, 3)), vexp_scale);
    const float32x4_t vnorm2 = vmulq_f32(vreinterpretq_f32_u32(vsraq_n_u32(vexp_offset, vnonsign2, 3)), vexp_scale);
    const float32x4_t vnorm3 = vmulq_f32(vreinterpretq_f32_u32(vsraq_n_u32(vexp_offset, vnonsign3, 3)), vexp_scale);
    const float32x4_t vnorm4 = vmulq_f32(vreinterpretq_f32_u32(vsraq_n_u32(vexp_offset, vnonsign4, 3)), vexp_scale);
    const float32x4_t vnorm5 = vmulq_f32(vreinterpretq_f32_u32(vsraq_n_u32(vexp_offset, vnonsign5, 3)), vexp_scale);

    const float32x4_t vdenorm0 = vsubq_f32(vreinterpretq_f32_u32(vsriq_n_u32(vmagic_bias, vnonsign0, 16)), vreinterpretq_f32_u32(vmagic_bias));
    const float32x4_t vdenorm1 = vsubq_f32(vreinterpretq_f32_u32(vsriq_n_u32(vmagic_bias, vnonsign1, 16)), vreinterpretq_f32_u32(vmagic_bias));
    const float32x4_t vdenorm2 = vsubq_f32(vreinterpretq_f32_u32(vsriq_n_u32(vmagic_bias, vnonsign2, 16)), vreinterpretq_f32_u32(vmagic_bias));
    const float32x4_t vdenorm3 = vsubq_f32(vreinterpretq_f32_u32(vsriq_n_u32(vmagic_bias, vnonsign3, 16)), vreinterpretq_f32_u32(vmagic_bias));
    const float32x4_t vdenorm4 = vsubq_f32(vreinterpretq_f32_u32(vsriq_n_u32(vmagic_bias, vnonsign4, 16)), vreinterpretq_f32_u32(vmagic_bias));
    const float32x4_t vdenorm5 = vsubq_f32(vreinterpretq_f32_u32(vsriq_n_u32(vmagic_bias, vnonsign5, 16)), vreinterpretq_f32_u32(vmagic_bias));

    const uint32x4_t vxmask0 = vcgtq_u32(vnonsign0, vdenorm_cutoff);
    const uint32x4_t vxmask1 = vcgtq_u32(vnonsign1, vdenorm_cutoff);
    const uint32x4_t vxmask2 = vcgtq_u32(vnonsign2, vdenorm_cutoff);
    const uint32x4_t vxmask3 = vcgtq_u32(vnonsign3, vdenorm_cutoff);
    const uint32x4_t vxmask4 = vcgtq_u32(vnonsign4, vdenorm_cutoff);
    const uint32x4_t vxmask5 = vcgtq_u32(vnonsign5, vdenorm_cutoff);

    const uint32x4_t vf0 = vorrq_u32(vsign0, vreinterpretq_u32_f32(vbslq_f32(vxmask0, vnorm0, vdenorm0)));
    const uint32x4_t vf1 = vorrq_u32(vsign1, vreinterpretq_u32_f32(vbslq_f32(vxmask1, vnorm1, vdenorm1)));
    const uint32x4_t vf2 = vorrq_u32(vsign2, vreinterpretq_u32_f32(vbslq_f32(vxmask2, vnorm2, vdenorm2)));
    const uint32x4_t vf3 = vorrq_u32(vsign3, vreinterpretq_u32_f32(vbslq_f32(vxmask3, vnorm3, vdenorm3)));
    const uint32x4_t vf4 = vorrq_u32(vsign4, vreinterpretq_u32_f32(vbslq_f32(vxmask4, vnorm4, vdenorm4)));
    const uint32x4_t vf5 = vorrq_u32(vsign5, vreinterpretq_u32_f32(vbslq_f32(vxmask5, vnorm5, vdenorm5)));

    vst1q_f32(output, vreinterpretq_f32_u32(vf0)); output += 4;
    vst1q_f32(output, vreinterpretq_f32_u32(vf1)); output += 4;
    vst1q_f32(output, vreinterpretq_f32_u32(vf2)); output += 4;
    vst1q_f32(output, vreinterpretq_f32_u32(vf3)); output += 4;
    vst1q_f32(output, vreinterpretq_f32_u32(vf4)); output += 4;
    vst1q_f32(output, vreinterpretq_f32_u32(vf5)); output += 4;
  }
  for (; batch >= 8 * sizeof(uint16_t); batch -= 8 * sizeof(uint16_t)) {
    const uint16x8_t vh = vld1q_u16(i); i += 8;

    const uint32x4_t vw_lo = vshll_n_u16(vget_low_u16(vh), 16);
    const uint32x4_t vw_hi = vshll_n_u16(vget_high_u16(vh), 16);

    const uint32x4_t vsign_lo = vandq_u32(vw_lo, vsign_mask);
    const uint32x4_t vsign_hi = vandq_u32(vw_hi, vsign_mask);

    const uint32x4_t vnonsign_lo = veorq_u32(vw_lo, vsign_lo);
    const uint32x4_t vnonsign_hi = veorq_u32(vw_hi, vsign_hi);

    const float32x4_t vnorm_lo = vmulq_f32(vreinterpretq_f32_u32(vsraq_n_u32(vexp_offset, vnonsign_lo, 3)), vexp_scale);
    const float32x4_t vnorm_hi = vmulq_f32(vreinterpretq_f32_u32(vsraq_n_u32(vexp_offset, vnonsign_hi, 3)), vexp_scale);

    const float32x4_t vdenorm_lo = vsubq_f32(vreinterpretq_f32_u32(vsriq_n_u32(vmagic_bias, vnonsign_lo, 16)), vreinterpretq_f32_u32(vmagic_bias));
    const float32x4_t vdenorm_hi = vsubq_f32(vreinterpretq_f32_u32(vsriq_n_u32(vmagic_bias, vnonsign_hi, 16)), vreinterpretq_f32_u32(vmagic_bias));

    const uint32x4_t vxmask_lo = vcgtq_u32(vnonsign_lo, vdenorm_cutoff);
    const uint32x4_t vf_lo = vorrq_u32(vsign_lo, vreinterpretq_u32_f32(vbslq_f32(vxmask_lo, vnorm_lo, vdenorm_lo)));

    const uint32x4_t vxmask_hi = vcgtq_u32(vnonsign_hi, vdenorm_cutoff);
    const uint32x4_t vf_hi = vorrq_u32(vsign_hi, vreinterpretq_u32_f32(vbslq_f32(vxmask_hi, vnorm_hi, vdenorm_hi)));

    vst1q_f32(output, vreinterpretq_f32_u32(vf_lo)); output += 4;
    vst1q_f32(output, vreinterpretq_f32_u32(vf_hi)); output += 4;
  }
  if XNN_UNPREDICTABLE(batch != 0) {
    const uint16x8_t vh = vld1q_u16(i); i += 8;

    const uint32x4_t vw_lo = vshll_n_u16(vget_low_u16(vh), 16);
    const uint32x4_t vw_hi = vshll_n_u16(vget_high_u16(vh), 16);

    const uint32x4_t vsign_lo = vandq_u32(vw_lo, vsign_mask);
    const uint32x4_t vsign_hi = vandq_u32(vw_hi, vsign_mask);

    const uint32x4_t vnonsign_lo = veorq_u32(vw_lo, vsign_lo);
    const uint32x4_t vnonsign_hi = veorq_u32(vw_hi, vsign_hi);

    const float32x4_t vnorm_lo = vmulq_f32(vreinterpretq_f32_u32(vsraq_n_u32(vexp_offset, vnonsign_lo, 3)), vexp_scale);
    const float32x4_t vnorm_hi = vmulq_f32(vreinterpretq_f32_u32(vsraq_n_u32(vexp_offset, vnonsign_hi, 3)), vexp_scale);

    const float32x4_t vdenorm_lo = vsubq_f32(vreinterpretq_f32_u32(vsriq_n_u32(vmagic_bias, vnonsign_lo, 16)), vreinterpretq_f32_u32(vmagic_bias));
    const float32x4_t vdenorm_hi = vsubq_f32(vreinterpretq_f32_u32(vsriq_n_u32(vmagic_bias, vnonsign_hi, 16)), vreinterpretq_f32_u32(vmagic_bias));

    const uint32x4_t vxmask_lo = vcgtq_u32(vnonsign_lo, vdenorm_cutoff);
    uint32x4_t vf = vorrq_u32(vsign_lo, vreinterpretq_u32_f32(vbslq_f32(vxmask_lo, vnorm_lo, vdenorm_lo)));

    if (batch & (4 * sizeof(uint16_t))) {
      vst1q_f32(output, vreinterpretq_f32_u32(vf)); output += 4;

      const uint32x4_t vxmask_hi = vcgtq_u32(vnonsign_hi, vdenorm_cutoff);
      vf = vorrq_u32(vsign_hi, vreinterpretq_u32_f32(vbslq_f32(vxmask_hi, vnorm_hi, vdenorm_hi)));
    }
    uint32x2_t vf_lo = vget_low_u32(vf);
    if (batch & (2 * sizeof(uint16_t))) {
      vst1_f32(output, vreinterpret_f32_u32(vf_lo)); output += 2;
      vf_lo = vget_high_u32(vf);
    }
    if (batch & (1 * sizeof(uint16_t))) {
      vst1_lane_f32(output, vreinterpret_f32_u32(vf_lo), 0);
    }
  }
}