1*bb722a7dSDimitry Andric //===-- Definition of bfloat16 data type. -----------------------*- C++ -*-===// 2*bb722a7dSDimitry Andric // 3*bb722a7dSDimitry Andric // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4*bb722a7dSDimitry Andric // See https://llvm.org/LICENSE.txt for license information. 5*bb722a7dSDimitry Andric // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6*bb722a7dSDimitry Andric // 7*bb722a7dSDimitry Andric //===----------------------------------------------------------------------===// 8*bb722a7dSDimitry Andric 9*bb722a7dSDimitry Andric #ifndef LLVM_LIBC_SRC___SUPPORT_FPUTIL_BFLOAT16_H 10*bb722a7dSDimitry Andric #define LLVM_LIBC_SRC___SUPPORT_FPUTIL_BFLOAT16_H 11*bb722a7dSDimitry Andric 12*bb722a7dSDimitry Andric #include "src/__support/CPP/bit.h" 13*bb722a7dSDimitry Andric #include "src/__support/CPP/type_traits.h" 14*bb722a7dSDimitry Andric #include "src/__support/FPUtil/cast.h" 15*bb722a7dSDimitry Andric #include "src/__support/FPUtil/dyadic_float.h" 16*bb722a7dSDimitry Andric #include "src/__support/macros/config.h" 17*bb722a7dSDimitry Andric #include "src/__support/macros/properties/types.h" 18*bb722a7dSDimitry Andric 19*bb722a7dSDimitry Andric #include <stdint.h> 20*bb722a7dSDimitry Andric 21*bb722a7dSDimitry Andric namespace LIBC_NAMESPACE_DECL { 22*bb722a7dSDimitry Andric namespace fputil { 23*bb722a7dSDimitry Andric 24*bb722a7dSDimitry Andric struct BFloat16 { 25*bb722a7dSDimitry Andric uint16_t bits; 26*bb722a7dSDimitry Andric 27*bb722a7dSDimitry Andric LIBC_INLINE BFloat16() = default; 28*bb722a7dSDimitry Andric BFloat16BFloat1629*bb722a7dSDimitry Andric LIBC_INLINE constexpr explicit BFloat16(uint16_t bits) : bits(bits) {} 30*bb722a7dSDimitry Andric BFloat16BFloat1631*bb722a7dSDimitry Andric template <typename T> LIBC_INLINE constexpr explicit BFloat16(T value) { 32*bb722a7dSDimitry Andric if constexpr (cpp::is_floating_point_v<T>) { 33*bb722a7dSDimitry Andric bits = fputil::cast<bfloat16>(value).bits; 34*bb722a7dSDimitry Andric } else if constexpr (cpp::is_integral_v<T>) { 35*bb722a7dSDimitry Andric Sign sign = Sign::POS; 36*bb722a7dSDimitry Andric 37*bb722a7dSDimitry Andric if constexpr (cpp::is_signed_v<T>) { 38*bb722a7dSDimitry Andric if (value < 0) { 39*bb722a7dSDimitry Andric sign = Sign::NEG; 40*bb722a7dSDimitry Andric value = -value; 41*bb722a7dSDimitry Andric } 42*bb722a7dSDimitry Andric } 43*bb722a7dSDimitry Andric 44*bb722a7dSDimitry Andric fputil::DyadicFloat<cpp::numeric_limits<cpp::make_unsigned_t<T>>::digits> 45*bb722a7dSDimitry Andric xd(sign, 0, value); 46*bb722a7dSDimitry Andric bits = xd.template as<bfloat16, /*ShouldSignalExceptions=*/true>().bits; 47*bb722a7dSDimitry Andric 48*bb722a7dSDimitry Andric } else { 49*bb722a7dSDimitry Andric bits = fputil::cast<bfloat16>(static_cast<float>(value)).bits; 50*bb722a7dSDimitry Andric } 51*bb722a7dSDimitry Andric } 52*bb722a7dSDimitry Andric 53*bb722a7dSDimitry Andric template <cpp::enable_if_t<fputil::get_fp_type<float>() == 54*bb722a7dSDimitry Andric fputil::FPType::IEEE754_Binary32, 55*bb722a7dSDimitry Andric int> = 0> 56*bb722a7dSDimitry Andric LIBC_INLINE constexpr operator float() const { 57*bb722a7dSDimitry Andric uint32_t x_bits = static_cast<uint32_t>(bits) << 16U; 58*bb722a7dSDimitry Andric return cpp::bit_cast<float>(x_bits); 59*bb722a7dSDimitry Andric } 60*bb722a7dSDimitry Andric }; // struct BFloat16 61*bb722a7dSDimitry Andric 62*bb722a7dSDimitry Andric } // namespace fputil 63*bb722a7dSDimitry Andric } // namespace LIBC_NAMESPACE_DECL 64*bb722a7dSDimitry Andric 65*bb722a7dSDimitry Andric #endif // LLVM_LIBC_SRC___SUPPORT_FPUTIL_BFLOAT16_H 66