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> |
Diego Lopez Recas | febc20f | 2021-09-16 17:25:04 +0100 | [diff] [blame] | 10 | #include <cstring> |
Narumol Prangnawarat | c3bf6ef | 2020-02-28 12:45:21 +0000 | [diff] [blame] | 11 | #include <stdint.h> |
| 12 | |
| 13 | namespace armnn |
| 14 | { |
Narumol Prangnawarat | e664484 | 2020-03-05 17:27:45 +0000 | [diff] [blame] | 15 | class BFloat16 |
| 16 | { |
| 17 | public: |
| 18 | BFloat16() |
Narumol Prangnawarat | e9e6818 | 2020-03-11 11:34:55 +0000 | [diff] [blame] | 19 | : m_Value(0) |
Narumol Prangnawarat | e664484 | 2020-03-05 17:27:45 +0000 | [diff] [blame] | 20 | {} |
| 21 | |
Francis Murtagh | c0f0af3 | 2020-03-26 12:14:21 +0000 | [diff] [blame] | 22 | BFloat16(const BFloat16& v) = default; |
| 23 | |
Narumol Prangnawarat | e664484 | 2020-03-05 17:27:45 +0000 | [diff] [blame] | 24 | explicit BFloat16(uint16_t v) |
Narumol Prangnawarat | e9e6818 | 2020-03-11 11:34:55 +0000 | [diff] [blame] | 25 | : m_Value(v) |
Narumol Prangnawarat | e664484 | 2020-03-05 17:27:45 +0000 | [diff] [blame] | 26 | {} |
| 27 | |
| 28 | explicit BFloat16(float v) |
| 29 | { |
Narumol Prangnawarat | e9e6818 | 2020-03-11 11:34:55 +0000 | [diff] [blame] | 30 | m_Value = Float32ToBFloat16(v).Val(); |
Narumol Prangnawarat | e664484 | 2020-03-05 17:27:45 +0000 | [diff] [blame] | 31 | } |
| 32 | |
Narumol Prangnawarat | 44179c3 | 2020-03-11 14:51:27 +0000 | [diff] [blame] | 33 | operator float() const |
| 34 | { |
| 35 | return ToFloat32(); |
| 36 | } |
| 37 | |
Francis Murtagh | c0f0af3 | 2020-03-26 12:14:21 +0000 | [diff] [blame] | 38 | BFloat16& operator=(const BFloat16& other) = default; |
Narumol Prangnawarat | 44179c3 | 2020-03-11 14:51:27 +0000 | [diff] [blame] | 39 | |
Narumol Prangnawarat | e664484 | 2020-03-05 17:27:45 +0000 | [diff] [blame] | 40 | BFloat16& operator=(float v) |
| 41 | { |
Narumol Prangnawarat | e9e6818 | 2020-03-11 11:34:55 +0000 | [diff] [blame] | 42 | m_Value = Float32ToBFloat16(v).Val(); |
Narumol Prangnawarat | e664484 | 2020-03-05 17:27:45 +0000 | [diff] [blame] | 43 | return *this; |
| 44 | } |
| 45 | |
| 46 | bool operator==(const BFloat16& r) const |
| 47 | { |
Narumol Prangnawarat | e9e6818 | 2020-03-11 11:34:55 +0000 | [diff] [blame] | 48 | return m_Value == r.Val(); |
Narumol Prangnawarat | e664484 | 2020-03-05 17:27:45 +0000 | [diff] [blame] | 49 | } |
| 50 | |
Narumol Prangnawarat | e9e6818 | 2020-03-11 11:34:55 +0000 | [diff] [blame] | 51 | static BFloat16 Float32ToBFloat16(const float v) |
Narumol Prangnawarat | e664484 | 2020-03-05 17:27:45 +0000 | [diff] [blame] | 52 | { |
| 53 | if (std::isnan(v)) |
| 54 | { |
Narumol Prangnawarat | e9e6818 | 2020-03-11 11:34:55 +0000 | [diff] [blame] | 55 | return Nan(); |
Narumol Prangnawarat | e664484 | 2020-03-05 17:27:45 +0000 | [diff] [blame] | 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) |
Francis Murtagh | c0f0af3 | 2020-03-26 12:14:21 +0000 | [diff] [blame] | 76 | const uint16_t error = static_cast<uint16_t>((*u32 & 0x0000FFFF)); |
Narumol Prangnawarat | e664484 | 2020-03-05 17:27:45 +0000 | [diff] [blame] | 77 | if ((error > 0x8000 || (error == 0x8000 && lsb == 1))) |
| 78 | { |
| 79 | u16++; |
| 80 | } |
| 81 | BFloat16 b(u16); |
| 82 | return b; |
| 83 | } |
| 84 | } |
| 85 | |
Narumol Prangnawarat | e9e6818 | 2020-03-11 11:34:55 +0000 | [diff] [blame] | 86 | float ToFloat32() const |
Narumol Prangnawarat | e664484 | 2020-03-05 17:27:45 +0000 | [diff] [blame] | 87 | { |
Francis Murtagh | c0f0af3 | 2020-03-26 12:14:21 +0000 | [diff] [blame] | 88 | const uint32_t u32 = static_cast<uint32_t>(m_Value << 16u); |
Diego Lopez Recas | febc20f | 2021-09-16 17:25:04 +0100 | [diff] [blame] | 89 | float f32; |
| 90 | static_assert(sizeof u32 == sizeof f32, ""); |
| 91 | std::memcpy(&f32, &u32, sizeof u32); |
| 92 | return f32; |
Narumol Prangnawarat | e664484 | 2020-03-05 17:27:45 +0000 | [diff] [blame] | 93 | } |
| 94 | |
Narumol Prangnawarat | e9e6818 | 2020-03-11 11:34:55 +0000 | [diff] [blame] | 95 | uint16_t Val() const |
Narumol Prangnawarat | e664484 | 2020-03-05 17:27:45 +0000 | [diff] [blame] | 96 | { |
Narumol Prangnawarat | e9e6818 | 2020-03-11 11:34:55 +0000 | [diff] [blame] | 97 | return m_Value; |
Narumol Prangnawarat | e664484 | 2020-03-05 17:27:45 +0000 | [diff] [blame] | 98 | } |
| 99 | |
Narumol Prangnawarat | e9e6818 | 2020-03-11 11:34:55 +0000 | [diff] [blame] | 100 | static BFloat16 Max() |
Narumol Prangnawarat | e664484 | 2020-03-05 17:27:45 +0000 | [diff] [blame] | 101 | { |
| 102 | uint16_t max = 0x7F7F; |
| 103 | return BFloat16(max); |
| 104 | } |
| 105 | |
Narumol Prangnawarat | e9e6818 | 2020-03-11 11:34:55 +0000 | [diff] [blame] | 106 | static BFloat16 Nan() |
Narumol Prangnawarat | e664484 | 2020-03-05 17:27:45 +0000 | [diff] [blame] | 107 | { |
| 108 | uint16_t nan = 0x7FC0; |
| 109 | return BFloat16(nan); |
| 110 | } |
| 111 | |
Narumol Prangnawarat | e9e6818 | 2020-03-11 11:34:55 +0000 | [diff] [blame] | 112 | static BFloat16 Inf() |
Narumol Prangnawarat | e664484 | 2020-03-05 17:27:45 +0000 | [diff] [blame] | 113 | { |
| 114 | uint16_t infVal = 0x7F80; |
| 115 | return BFloat16(infVal); |
| 116 | } |
| 117 | |
| 118 | private: |
Narumol Prangnawarat | e9e6818 | 2020-03-11 11:34:55 +0000 | [diff] [blame] | 119 | uint16_t m_Value; |
Narumol Prangnawarat | e664484 | 2020-03-05 17:27:45 +0000 | [diff] [blame] | 120 | }; |
| 121 | |
| 122 | inline std::ostream& operator<<(std::ostream& os, const BFloat16& b) |
| 123 | { |
Narumol Prangnawarat | e9e6818 | 2020-03-11 11:34:55 +0000 | [diff] [blame] | 124 | os << b.ToFloat32() << "(0x" << std::hex << b.Val() << ")"; |
Narumol Prangnawarat | e664484 | 2020-03-05 17:27:45 +0000 | [diff] [blame] | 125 | return os; |
| 126 | } |
| 127 | |
Narumol Prangnawarat | c3bf6ef | 2020-02-28 12:45:21 +0000 | [diff] [blame] | 128 | } //namespace armnn |