| 1 | //===-- Definition of bfloat16 data type. -----------------------*- C++ -*-===// |
| 2 | // |
| 3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
| 4 | // See https://llvm.org/LICENSE.txt for license information. |
| 5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
| 6 | // |
| 7 | //===----------------------------------------------------------------------===// |
| 8 | |
| 9 | #ifndef LLVM_LIBC_SRC___SUPPORT_FPUTIL_BFLOAT16_H |
| 10 | #define LLVM_LIBC_SRC___SUPPORT_FPUTIL_BFLOAT16_H |
| 11 | |
| 12 | #include "hdr/stdint_proxy.h" |
| 13 | #include "src/__support/CPP/bit.h" |
| 14 | #include "src/__support/CPP/type_traits.h" |
| 15 | #include "src/__support/FPUtil/cast.h" |
| 16 | #include "src/__support/FPUtil/comparison_operations.h" |
| 17 | #include "src/__support/FPUtil/dyadic_float.h" |
| 18 | #include "src/__support/FPUtil/generic/add_sub.h" |
| 19 | #include "src/__support/FPUtil/generic/div.h" |
| 20 | #include "src/__support/FPUtil/generic/mul.h" |
| 21 | #include "src/__support/macros/config.h" |
| 22 | #include "src/__support/macros/properties/types.h" |
| 23 | |
| 24 | namespace LIBC_NAMESPACE_DECL { |
| 25 | namespace fputil { |
| 26 | |
| 27 | struct BFloat16 { |
| 28 | uint16_t bits; |
| 29 | |
| 30 | LIBC_INLINE BFloat16() = default; |
| 31 | |
| 32 | template <typename T> |
| 33 | LIBC_INLINE constexpr explicit BFloat16(T value) |
| 34 | : bits(static_cast<uint16_t>(0U)) { |
| 35 | if constexpr (cpp::is_floating_point_v<T>) { |
| 36 | bits = fputil::cast<bfloat16>(value).bits; |
| 37 | } else if constexpr (cpp::is_integral_v<T>) { |
| 38 | Sign sign = Sign::POS; |
| 39 | |
| 40 | if constexpr (cpp::is_signed_v<T>) { |
| 41 | if (value < 0) { |
| 42 | sign = Sign::NEG; |
| 43 | value = -value; |
| 44 | } |
| 45 | } |
| 46 | |
| 47 | fputil::DyadicFloat<cpp::numeric_limits<cpp::make_unsigned_t<T>>::digits> |
| 48 | xd(sign, 0, value); |
| 49 | bits = xd.template as<bfloat16, /*ShouldSignalExceptions=*/true>().bits; |
| 50 | |
| 51 | } else if constexpr (cpp::is_convertible_v<T, BFloat16>) { |
| 52 | bits = value.operator BFloat16().bits; |
| 53 | } else { |
| 54 | bits = fputil::cast<bfloat16>(x: static_cast<float>(value)).bits; |
| 55 | } |
| 56 | } |
| 57 | |
| 58 | template <cpp::enable_if_t<fputil::get_fp_type<float>() == |
| 59 | fputil::FPType::IEEE754_Binary32, |
| 60 | int> = 0> |
| 61 | LIBC_INLINE constexpr operator float() const { |
| 62 | uint32_t x_bits = static_cast<uint32_t>(bits) << 16U; |
| 63 | return cpp::bit_cast<float>(from: x_bits); |
| 64 | } |
| 65 | |
| 66 | template <typename T, cpp::enable_if_t<cpp::is_integral_v<T>, int> = 0> |
| 67 | LIBC_INLINE constexpr explicit operator T() const { |
| 68 | return static_cast<T>(static_cast<float>(*this)); |
| 69 | } |
| 70 | |
| 71 | LIBC_INLINE constexpr bool operator==(BFloat16 other) const { |
| 72 | return fputil::equals(x: *this, y: other); |
| 73 | } |
| 74 | |
| 75 | LIBC_INLINE constexpr bool operator!=(BFloat16 other) const { |
| 76 | return !fputil::equals(x: *this, y: other); |
| 77 | } |
| 78 | |
| 79 | LIBC_INLINE constexpr bool operator<(BFloat16 other) const { |
| 80 | return fputil::less_than(x: *this, y: other); |
| 81 | } |
| 82 | |
| 83 | LIBC_INLINE constexpr bool operator<=(BFloat16 other) const { |
| 84 | return fputil::less_than_or_equals(x: *this, y: other); |
| 85 | } |
| 86 | |
| 87 | LIBC_INLINE constexpr bool operator>(BFloat16 other) const { |
| 88 | return fputil::greater_than(x: *this, y: other); |
| 89 | } |
| 90 | |
| 91 | LIBC_INLINE constexpr bool operator>=(BFloat16 other) const { |
| 92 | return fputil::greater_than_or_equals(x: *this, y: other); |
| 93 | } |
| 94 | |
| 95 | LIBC_INLINE LIBC_BIT_CAST_CONSTEXPR BFloat16 operator-() const { |
| 96 | fputil::FPBits<bfloat16> result(*this); |
| 97 | result.set_sign(result.is_pos() ? Sign::NEG : Sign::POS); |
| 98 | return result.get_val(); |
| 99 | } |
| 100 | |
| 101 | LIBC_INLINE constexpr BFloat16 operator+(BFloat16 other) const { |
| 102 | return fputil::generic::add<BFloat16>(x: *this, y: other); |
| 103 | } |
| 104 | |
| 105 | LIBC_INLINE constexpr BFloat16 operator-(BFloat16 other) const { |
| 106 | return fputil::generic::sub<BFloat16>(x: *this, y: other); |
| 107 | } |
| 108 | |
| 109 | LIBC_INLINE constexpr BFloat16 operator*(BFloat16 other) const { |
| 110 | return fputil::generic::mul<bfloat16>(x: *this, y: other); |
| 111 | } |
| 112 | |
| 113 | LIBC_INLINE constexpr BFloat16 operator/(BFloat16 other) const { |
| 114 | return fputil::generic::div<bfloat16>(x: *this, y: other); |
| 115 | } |
| 116 | |
| 117 | LIBC_INLINE constexpr BFloat16 &operator*=(BFloat16 other) { |
| 118 | *this = *this * other; |
| 119 | return *this; |
| 120 | } |
| 121 | LIBC_INLINE constexpr BFloat16 &operator+=(BFloat16 other) { |
| 122 | *this = *this + other; |
| 123 | return *this; |
| 124 | } |
| 125 | LIBC_INLINE constexpr BFloat16 &operator-=(BFloat16 other) { |
| 126 | *this = *this - other; |
| 127 | return *this; |
| 128 | } |
| 129 | LIBC_INLINE constexpr BFloat16 &operator/=(BFloat16 other) { |
| 130 | *this = *this / other; |
| 131 | return *this; |
| 132 | } |
| 133 | }; // struct BFloat16 |
| 134 | |
| 135 | } // namespace fputil |
| 136 | } // namespace LIBC_NAMESPACE_DECL |
| 137 | |
| 138 | #endif // LLVM_LIBC_SRC___SUPPORT_FPUTIL_BFLOAT16_H |
| 139 | |