1 //
2 // Copyright © 2020 Arm Ltd. All rights reserved.
3 // SPDX-License-Identifier: MIT
4 //
5
6 #pragma once
7
8 #include <ostream>
9 #include <cmath>
10 #include <cstring>
11 #include <stdint.h>
12
13 namespace armnn
14 {
15 class BFloat16
16 {
17 public:
BFloat16()18 BFloat16()
19 : m_Value(0)
20 {}
21
22 BFloat16(const BFloat16& v) = default;
23
BFloat16(uint16_t v)24 explicit BFloat16(uint16_t v)
25 : m_Value(v)
26 {}
27
BFloat16(float v)28 explicit BFloat16(float v)
29 {
30 m_Value = Float32ToBFloat16(v).Val();
31 }
32
operator float() const33 operator float() const
34 {
35 return ToFloat32();
36 }
37
38 BFloat16& operator=(const BFloat16& other) = default;
39
operator =(float v)40 BFloat16& operator=(float v)
41 {
42 m_Value = Float32ToBFloat16(v).Val();
43 return *this;
44 }
45
operator ==(const BFloat16 & r) const46 bool operator==(const BFloat16& r) const
47 {
48 return m_Value == r.Val();
49 }
50
Float32ToBFloat16(const float v)51 static BFloat16 Float32ToBFloat16(const float v)
52 {
53 if (std::isnan(v))
54 {
55 return Nan();
56 }
57 else
58 {
59 // Round value to the nearest even
60 // Float32
61 // S EEEEEEEE MMMMMMLRMMMMMMMMMMMMMMM
62 // BFloat16
63 // S EEEEEEEE MMMMMML
64 // LSB (L): Least significat bit of BFloat16 (last bit of the Mantissa of BFloat16)
65 // R: Rounding bit
66 // LSB = 0, R = 0 -> round down
67 // LSB = 1, R = 0 -> round down
68 // LSB = 0, R = 1, all the rest = 0 -> round down
69 // LSB = 1, R = 1 -> round up
70 // LSB = 0, R = 1 -> round up
71 const uint32_t* u32 = reinterpret_cast<const uint32_t*>(&v);
72 uint16_t u16 = static_cast<uint16_t>(*u32 >> 16u);
73 // Mark the LSB
74 const uint16_t lsb = u16 & 0x0001;
75 // Mark the error to be truncate (the rest of 16 bits of FP32)
76 const uint16_t error = static_cast<uint16_t>((*u32 & 0x0000FFFF));
77 if ((error > 0x8000 || (error == 0x8000 && lsb == 1)))
78 {
79 u16++;
80 }
81 BFloat16 b(u16);
82 return b;
83 }
84 }
85
ToFloat32() const86 float ToFloat32() const
87 {
88 const uint32_t u32 = static_cast<uint32_t>(m_Value << 16u);
89 float f32;
90 static_assert(sizeof u32 == sizeof f32, "");
91 std::memcpy(&f32, &u32, sizeof u32);
92 return f32;
93 }
94
Val() const95 uint16_t Val() const
96 {
97 return m_Value;
98 }
99
Max()100 static BFloat16 Max()
101 {
102 uint16_t max = 0x7F7F;
103 return BFloat16(max);
104 }
105
Nan()106 static BFloat16 Nan()
107 {
108 uint16_t nan = 0x7FC0;
109 return BFloat16(nan);
110 }
111
Inf()112 static BFloat16 Inf()
113 {
114 uint16_t infVal = 0x7F80;
115 return BFloat16(infVal);
116 }
117
118 private:
119 uint16_t m_Value;
120 };
121
operator <<(std::ostream & os,const BFloat16 & b)122 inline std::ostream& operator<<(std::ostream& os, const BFloat16& b)
123 {
124 os << b.ToFloat32() << "(0x" << std::hex << b.Val() << ")";
125 return os;
126 }
127
128 } //namespace armnn
129