| Line | Branch | Exec | Source |
|---|---|---|---|
| 1 | // | ||
| 2 | // SPDX-FileCopyrightText: Copyright 2024-2025 Arm Limited and/or its affiliates <open-source-office@arm.com> | ||
| 3 | // | ||
| 4 | // SPDX-License-Identifier: Apache-2.0 | ||
| 5 | // | ||
| 6 | |||
| 7 | #pragma once | ||
| 8 | |||
| 9 | #include <cstdint> | ||
| 10 | #include <cstring> | ||
| 11 | #include <iosfwd> | ||
| 12 | #include <type_traits> | ||
| 13 | |||
| 14 | #include "test/common/cpu_info.hpp" | ||
| 15 | #include "test/common/type_traits.hpp" | ||
| 16 | |||
| 17 | extern "C" { | ||
| 18 | |||
| 19 | /// Converts single-precision floating-point to half-precision brain floating-point. | ||
| 20 | /// | ||
| 21 | /// @params[in] value The single-precision floating-point value. | ||
| 22 | /// | ||
| 23 | /// @return The half-precision brain floating-point value reinterpreted as 16-bit unsigned integer. | ||
| 24 | uint16_t kai_test_float_to_bfloat16_bfcvt(float value); | ||
| 25 | |||
| 26 | } // extern "C" | ||
| 27 | namespace kai::test { | ||
| 28 | |||
| 29 | /// Half-precision brain floating-point. | ||
| 30 | template <bool hardware_support> | ||
| 31 | class BFloat16 { | ||
| 32 | public: | ||
| 33 | /// Constructor. | ||
| 34 | BFloat16() = default; | ||
| 35 | |||
| 36 | using p_F32BF16convert = uint16_t (*)(float); | ||
| 37 | |||
| 38 | /// Creates a new object from the specified numeric value. | ||
| 39 | 3392113970 | explicit BFloat16(float value) : m_data(f32_bf16_convertfn(value)) { | |
| 40 | 3392113970 | } | |
| 41 | |||
| 42 | /// Creates a new half-precision brain floating-point value from the raw data. | ||
| 43 | /// | ||
| 44 | /// @param[in] data The binary representation of the floating-point value. | ||
| 45 | /// | ||
| 46 | /// @return The half-precision brain floating-point value. | ||
| 47 | static constexpr BFloat16 from_binary(uint16_t data) { | ||
| 48 | BFloat16 value{}; | ||
| 49 | value.m_data = data; | ||
| 50 | return value; | ||
| 51 | } | ||
| 52 | |||
| 53 | /// Assigns to the specified numeric value which will be converted to `bfloat16_t`. | ||
| 54 | template <typename T, std::enable_if_t<is_arithmetic<T>, bool> = true> | ||
| 55 | BFloat16& operator=(T value) { | ||
| 56 | const auto value_f32 = static_cast<float>(value); | ||
| 57 | m_data = f32_bf16_convertfn(value_f32); | ||
| 58 | return *this; | ||
| 59 | } | ||
| 60 | |||
| 61 | /// Converts to single-precision floating-point. | ||
| 62 | 1686405429 | explicit operator float() const { | |
| 63 | 1686405429 | float value_f32 = 0.0F; | |
| 64 | 1686405429 | uint32_t value_u32 = static_cast<uint32_t>(m_data) << 16; | |
| 65 | |||
| 66 | 1686405429 | memcpy(&value_f32, &value_u32, sizeof(float)); | |
| 67 | |||
| 68 | 3372810858 | return value_f32; | |
| 69 | 1686405429 | } | |
| 70 | |||
| 71 | private: | ||
| 72 | /// Equality operator. | ||
| 73 | 3 | [[nodiscard]] friend bool operator==(BFloat16 lhs, BFloat16 rhs) { | |
| 74 | 3 | return lhs.m_data == rhs.m_data; | |
| 75 | } | ||
| 76 | |||
| 77 | /// Inequality operator. | ||
| 78 | 3 | [[nodiscard]] friend bool operator!=(BFloat16 lhs, BFloat16 rhs) { | |
| 79 | 3 | return lhs.m_data != rhs.m_data; | |
| 80 | } | ||
| 81 | |||
| 82 | /// Writes the value to the output stream. | ||
| 83 | /// | ||
| 84 | /// @param[in] os Output stream to be written to. | ||
| 85 | /// @param[in] value Value to be written. | ||
| 86 | /// | ||
| 87 | /// @return The output stream. | ||
| 88 | friend std::ostream& operator<<(std::ostream& os, BFloat16<> value); | ||
| 89 | |||
| 90 | 1399264618 | static uint16_t float_to_bfloat16_round_towards_zero(float value) { | |
| 91 | 1399264618 | uint32_t value_u32; | |
| 92 | |||
| 93 | 1399264618 | memcpy(&value_u32, &value, sizeof(value)); | |
| 94 | |||
| 95 | 2798529236 | return value_u32 >> 16; | |
| 96 | 1399264618 | } | |
| 97 | |||
| 98 | inline static p_F32BF16convert f32_bf16_convertfn = (hardware_support && cpu_has_bf16()) | ||
| 99 | ? &kai_test_float_to_bfloat16_bfcvt | ||
| 100 | : &float_to_bfloat16_round_towards_zero; | ||
| 101 | |||
| 102 | uint16_t m_data; | ||
| 103 | }; | ||
| 104 | |||
| 105 | } // namespace kai::test | ||
| 106 |