File size: 3,734 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
// Auto-generated file. Do not edit!
//   Template: src/f16-f32-vcvt/scalar.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 <xnnpack/common.h>
#include <xnnpack/math.h>
#include <xnnpack/vcvt.h>


void xnn_f16_f32_vcvt_ukernel__scalar_x4(
    size_t batch,
    const void* input,
    float* output,
    const union xnn_f16_f32_cvt_params params[restrict XNN_MIN_ELEMENTS(1)])
{
  assert(batch != 0);
  assert(batch % sizeof(uint16_t) == 0);
  assert(input != NULL);
  assert(output != NULL);

  const uint32_t vsign_mask = params->scalar.sign_mask;
  const uint32_t vexp_offset = params->scalar.exp_offset;
  const float vexp_scale = params->scalar.exp_scale;
  const uint32_t vmagic_mask = params->scalar.magic_mask;
  const float vmagic_bias = params->scalar.magic_bias;
  const uint32_t vdenorm_cutoff = params->scalar.denorm_cutoff;

  const uint16_t* i = (const uint16_t*) input;
  uint32_t* o = (uint32_t*) output;
  for (; batch >= 4 * sizeof(uint16_t); batch -= 4 * sizeof(uint16_t)) {
    const uint16_t vh0 = i[0];
    const uint16_t vh1 = i[1];
    const uint16_t vh2 = i[2];
    const uint16_t vh3 = i[3];
    i += 4;

    const uint32_t vw0 = (uint32_t) vh0 << 16;
    const uint32_t vw1 = (uint32_t) vh1 << 16;
    const uint32_t vw2 = (uint32_t) vh2 << 16;
    const uint32_t vw3 = (uint32_t) vh3 << 16;

    const uint32_t vsign0 = vw0 & vsign_mask;
    const uint32_t vsign1 = vw1 & vsign_mask;
    const uint32_t vsign2 = vw2 & vsign_mask;
    const uint32_t vsign3 = vw3 & vsign_mask;

    const uint32_t v2w0 = vw0 + vw0;
    const uint32_t v2w1 = vw1 + vw1;
    const uint32_t v2w2 = vw2 + vw2;
    const uint32_t v2w3 = vw3 + vw3;

    const uint32_t vnorm0 = float_as_uint32(uint32_as_float((v2w0 >> 4) + vexp_offset) * vexp_scale);
    const uint32_t vnorm1 = float_as_uint32(uint32_as_float((v2w1 >> 4) + vexp_offset) * vexp_scale);
    const uint32_t vnorm2 = float_as_uint32(uint32_as_float((v2w2 >> 4) + vexp_offset) * vexp_scale);
    const uint32_t vnorm3 = float_as_uint32(uint32_as_float((v2w3 >> 4) + vexp_offset) * vexp_scale);

    const uint32_t vdenorm0 = float_as_uint32(uint32_as_float((v2w0 >> 17) | vmagic_mask) - vmagic_bias);
    const uint32_t vdenorm1 = float_as_uint32(uint32_as_float((v2w1 >> 17) | vmagic_mask) - vmagic_bias);
    const uint32_t vdenorm2 = float_as_uint32(uint32_as_float((v2w2 >> 17) | vmagic_mask) - vmagic_bias);
    const uint32_t vdenorm3 = float_as_uint32(uint32_as_float((v2w3 >> 17) | vmagic_mask) - vmagic_bias);

    const uint32_t vf0 = vsign0 | (XNN_UNPREDICTABLE(v2w0 < vdenorm_cutoff) ? vdenorm0 : vnorm0);
    const uint32_t vf1 = vsign1 | (XNN_UNPREDICTABLE(v2w1 < vdenorm_cutoff) ? vdenorm1 : vnorm1);
    const uint32_t vf2 = vsign2 | (XNN_UNPREDICTABLE(v2w2 < vdenorm_cutoff) ? vdenorm2 : vnorm2);
    const uint32_t vf3 = vsign3 | (XNN_UNPREDICTABLE(v2w3 < vdenorm_cutoff) ? vdenorm3 : vnorm3);

    o[0] = vf0;
    o[1] = vf1;
    o[2] = vf2;
    o[3] = vf3;
    o += 4;
  }
  if XNN_UNLIKELY(batch != 0) {
    do {
      const uint16_t vh = *i++;

      const uint32_t vw = (uint32_t) vh << 16;
      const uint32_t vsign = vw & vsign_mask;
      const uint32_t v2w = vw + vw;
      const uint32_t vnorm = float_as_uint32(uint32_as_float((v2w >> 4) + vexp_offset) * vexp_scale);
      const uint32_t vdenorm = float_as_uint32(uint32_as_float((v2w >> 17) | vmagic_mask) - vmagic_bias);
      const uint32_t vf = vsign | (XNN_UNPREDICTABLE(v2w < vdenorm_cutoff) ? vdenorm : vnorm);

      *o++ = vf;

      batch -= sizeof(uint16_t);
    } while (batch != 0);
  }
}