Narumol Prangnawarat | c3bf6ef | 2020-02-28 12:45:21 +0000 | [diff] [blame] | 1 | // |
| 2 | // Copyright © 2020 Arm Ltd. All rights reserved. |
| 3 | // SPDX-License-Identifier: MIT |
| 4 | // |
| 5 | |
| 6 | #pragma once |
| 7 | |
Narumol Prangnawarat | e664484 | 2020-03-05 17:27:45 +0000 | [diff] [blame] | 8 | #include <ostream> |
Narumol Prangnawarat | 8832522 | 2020-03-06 14:45:57 +0000 | [diff] [blame] | 9 | #include <cmath> |
Narumol Prangnawarat | c3bf6ef | 2020-02-28 12:45:21 +0000 | [diff] [blame] | 10 | #include <stdint.h> |
| 11 | |
| 12 | namespace armnn |
| 13 | { |
Narumol Prangnawarat | e664484 | 2020-03-05 17:27:45 +0000 | [diff] [blame] | 14 | class BFloat16 |
| 15 | { |
| 16 | public: |
| 17 | BFloat16() |
Narumol Prangnawarat | e9e6818 | 2020-03-11 11:34:55 +0000 | [diff] [blame] | 18 | : m_Value(0) |
Narumol Prangnawarat | e664484 | 2020-03-05 17:27:45 +0000 | [diff] [blame] | 19 | {} |
| 20 | |
Francis Murtagh | c0f0af3 | 2020-03-26 12:14:21 +0000 | [diff] [blame] | 21 | BFloat16(const BFloat16& v) = default; |
| 22 | |
Narumol Prangnawarat | e664484 | 2020-03-05 17:27:45 +0000 | [diff] [blame] | 23 | explicit BFloat16(uint16_t v) |
Narumol Prangnawarat | e9e6818 | 2020-03-11 11:34:55 +0000 | [diff] [blame] | 24 | : m_Value(v) |
Narumol Prangnawarat | e664484 | 2020-03-05 17:27:45 +0000 | [diff] [blame] | 25 | {} |
| 26 | |
| 27 | explicit BFloat16(float v) |
| 28 | { |
Narumol Prangnawarat | e9e6818 | 2020-03-11 11:34:55 +0000 | [diff] [blame] | 29 | m_Value = Float32ToBFloat16(v).Val(); |
Narumol Prangnawarat | e664484 | 2020-03-05 17:27:45 +0000 | [diff] [blame] | 30 | } |
| 31 | |
Narumol Prangnawarat | 44179c3 | 2020-03-11 14:51:27 +0000 | [diff] [blame] | 32 | operator float() const |
| 33 | { |
| 34 | return ToFloat32(); |
| 35 | } |
| 36 | |
Francis Murtagh | c0f0af3 | 2020-03-26 12:14:21 +0000 | [diff] [blame] | 37 | BFloat16& operator=(const BFloat16& other) = default; |
Narumol Prangnawarat | 44179c3 | 2020-03-11 14:51:27 +0000 | [diff] [blame] | 38 | |
Narumol Prangnawarat | e664484 | 2020-03-05 17:27:45 +0000 | [diff] [blame] | 39 | BFloat16& operator=(float v) |
| 40 | { |
Narumol Prangnawarat | e9e6818 | 2020-03-11 11:34:55 +0000 | [diff] [blame] | 41 | m_Value = Float32ToBFloat16(v).Val(); |
Narumol Prangnawarat | e664484 | 2020-03-05 17:27:45 +0000 | [diff] [blame] | 42 | return *this; |
| 43 | } |
| 44 | |
| 45 | bool operator==(const BFloat16& r) const |
| 46 | { |
Narumol Prangnawarat | e9e6818 | 2020-03-11 11:34:55 +0000 | [diff] [blame] | 47 | return m_Value == r.Val(); |
Narumol Prangnawarat | e664484 | 2020-03-05 17:27:45 +0000 | [diff] [blame] | 48 | } |
| 49 | |
Narumol Prangnawarat | e9e6818 | 2020-03-11 11:34:55 +0000 | [diff] [blame] | 50 | static BFloat16 Float32ToBFloat16(const float v) |
Narumol Prangnawarat | e664484 | 2020-03-05 17:27:45 +0000 | [diff] [blame] | 51 | { |
| 52 | if (std::isnan(v)) |
| 53 | { |
Narumol Prangnawarat | e9e6818 | 2020-03-11 11:34:55 +0000 | [diff] [blame] | 54 | return Nan(); |
Narumol Prangnawarat | e664484 | 2020-03-05 17:27:45 +0000 | [diff] [blame] | 55 | } |
| 56 | else |
| 57 | { |
| 58 | // Round value to the nearest even |
| 59 | // Float32 |
| 60 | // S EEEEEEEE MMMMMMLRMMMMMMMMMMMMMMM |
| 61 | // BFloat16 |
| 62 | // S EEEEEEEE MMMMMML |
| 63 | // LSB (L): Least significat bit of BFloat16 (last bit of the Mantissa of BFloat16) |
| 64 | // R: Rounding bit |
| 65 | // LSB = 0, R = 0 -> round down |
| 66 | // LSB = 1, R = 0 -> round down |
| 67 | // LSB = 0, R = 1, all the rest = 0 -> round down |
| 68 | // LSB = 1, R = 1 -> round up |
| 69 | // LSB = 0, R = 1 -> round up |
| 70 | const uint32_t* u32 = reinterpret_cast<const uint32_t*>(&v); |
| 71 | uint16_t u16 = static_cast<uint16_t>(*u32 >> 16u); |
| 72 | // Mark the LSB |
| 73 | const uint16_t lsb = u16 & 0x0001; |
| 74 | // Mark the error to be truncate (the rest of 16 bits of FP32) |
Francis Murtagh | c0f0af3 | 2020-03-26 12:14:21 +0000 | [diff] [blame] | 75 | const uint16_t error = static_cast<uint16_t>((*u32 & 0x0000FFFF)); |
Narumol Prangnawarat | e664484 | 2020-03-05 17:27:45 +0000 | [diff] [blame] | 76 | if ((error > 0x8000 || (error == 0x8000 && lsb == 1))) |
| 77 | { |
| 78 | u16++; |
| 79 | } |
| 80 | BFloat16 b(u16); |
| 81 | return b; |
| 82 | } |
| 83 | } |
| 84 | |
Narumol Prangnawarat | e9e6818 | 2020-03-11 11:34:55 +0000 | [diff] [blame] | 85 | float ToFloat32() const |
Narumol Prangnawarat | e664484 | 2020-03-05 17:27:45 +0000 | [diff] [blame] | 86 | { |
Francis Murtagh | c0f0af3 | 2020-03-26 12:14:21 +0000 | [diff] [blame] | 87 | const uint32_t u32 = static_cast<uint32_t>(m_Value << 16u); |
Narumol Prangnawarat | e664484 | 2020-03-05 17:27:45 +0000 | [diff] [blame] | 88 | const float* f32 = reinterpret_cast<const float*>(&u32); |
| 89 | return *f32; |
| 90 | } |
| 91 | |
Narumol Prangnawarat | e9e6818 | 2020-03-11 11:34:55 +0000 | [diff] [blame] | 92 | uint16_t Val() const |
Narumol Prangnawarat | e664484 | 2020-03-05 17:27:45 +0000 | [diff] [blame] | 93 | { |
Narumol Prangnawarat | e9e6818 | 2020-03-11 11:34:55 +0000 | [diff] [blame] | 94 | return m_Value; |
Narumol Prangnawarat | e664484 | 2020-03-05 17:27:45 +0000 | [diff] [blame] | 95 | } |
| 96 | |
Narumol Prangnawarat | e9e6818 | 2020-03-11 11:34:55 +0000 | [diff] [blame] | 97 | static BFloat16 Max() |
Narumol Prangnawarat | e664484 | 2020-03-05 17:27:45 +0000 | [diff] [blame] | 98 | { |
| 99 | uint16_t max = 0x7F7F; |
| 100 | return BFloat16(max); |
| 101 | } |
| 102 | |
Narumol Prangnawarat | e9e6818 | 2020-03-11 11:34:55 +0000 | [diff] [blame] | 103 | static BFloat16 Nan() |
Narumol Prangnawarat | e664484 | 2020-03-05 17:27:45 +0000 | [diff] [blame] | 104 | { |
| 105 | uint16_t nan = 0x7FC0; |
| 106 | return BFloat16(nan); |
| 107 | } |
| 108 | |
Narumol Prangnawarat | e9e6818 | 2020-03-11 11:34:55 +0000 | [diff] [blame] | 109 | static BFloat16 Inf() |
Narumol Prangnawarat | e664484 | 2020-03-05 17:27:45 +0000 | [diff] [blame] | 110 | { |
| 111 | uint16_t infVal = 0x7F80; |
| 112 | return BFloat16(infVal); |
| 113 | } |
| 114 | |
| 115 | private: |
Narumol Prangnawarat | e9e6818 | 2020-03-11 11:34:55 +0000 | [diff] [blame] | 116 | uint16_t m_Value; |
Narumol Prangnawarat | e664484 | 2020-03-05 17:27:45 +0000 | [diff] [blame] | 117 | }; |
| 118 | |
| 119 | inline std::ostream& operator<<(std::ostream& os, const BFloat16& b) |
| 120 | { |
Narumol Prangnawarat | e9e6818 | 2020-03-11 11:34:55 +0000 | [diff] [blame] | 121 | os << b.ToFloat32() << "(0x" << std::hex << b.Val() << ")"; |
Narumol Prangnawarat | e664484 | 2020-03-05 17:27:45 +0000 | [diff] [blame] | 122 | return os; |
| 123 | } |
| 124 | |
Narumol Prangnawarat | c3bf6ef | 2020-02-28 12:45:21 +0000 | [diff] [blame] | 125 | } //namespace armnn |