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
|
// Copyright 2020 NLP Odyssey Authors. All rights reserved.
// Use of this source code is governed by a BSD-style
// license that can be found in the LICENSE file.
package pytorch
// FloatBits16to32 converts the bits representation of a Half Float (16 bits)
// number to an IEEE 754 float representation (32 bits)
// From http://www.fox-toolkit.org/ftp/fasthalffloatconversion.pdf
func FloatBits16to32(u16 uint16) uint32 {
return mantissaTable[offsetTable[u16>>10]+(uint32(u16)&0x3ff)] + exponentTable[u16>>10]
}
var mantissaTable [2048]uint32
var exponentTable [64]uint32
var offsetTable [64]uint32
func init() {
initMantissaTable()
initExponentTable()
initOffsetTable()
}
func initMantissaTable() {
mantissaTable[0] = 0
for i := uint32(1); i < 1024; i++ {
mantissaTable[i] = convertMantissa(i)
}
for i := uint32(1024); i < 2048; i++ {
mantissaTable[i] = 0x38000000 + ((i - 1024) << 13)
}
}
func initExponentTable() {
exponentTable[0] = 0
exponentTable[31] = 0x47800000
exponentTable[32] = 0x80000000
exponentTable[63] = 0xC7800000
for i := uint32(1); i < 31; i++ {
exponentTable[i] = i << 23
}
for i := uint32(33); i < 63; i++ {
exponentTable[i] = 0x80000000 + (i-32)<<23
}
}
func initOffsetTable() {
offsetTable[0] = 0
offsetTable[32] = 0
for i := uint32(1); i < 31; i++ {
offsetTable[i] = 1024
}
for i := uint32(32); i < 64; i++ {
offsetTable[i] = 1024
}
}
func convertMantissa(i uint32) uint32 {
var m uint32 = i << 13 // zero pad mantissa bits
var e uint32 = 0 // zero exponent
for m&0x00800000 != 0 { // while not normalized
e -= 0x00800000 // decrement exponent (1 << 23)
m <<= 1 // shift mantissa
}
m &= ^uint32(0x00800000) // clear leading 1 bit
e += 0x38800000 // adjust bias ((127-14)<<23)
return m | e // return combined number
}
|
