28 lines
832 B
C
28 lines
832 B
C
|
#pragma once
|
||
|
|
|
||
|
|
#include <c10/util/BFloat16.h>
|
||
|
|
#include <c10/util/Half.h>
|
||
|
|
|
||
|
|
namespace c10 {
|
||
|
|
|
||
|
|
// Note: Explicit implementation of copysign for Half and BFloat16
|
||
|
|
// is needed to workaround g++-7/8 crash on aarch64, but also makes
|
||
|
|
// copysign faster for the half-precision types
|
||
|
|
template <typename T, typename U>
|
||
|
|
inline auto copysign(const T& a, const U& b) {
|
||
|
|
return std::copysign(a, b);
|
||
|
|
}
|
||
|
|
|
||
|
|
// Implement copysign for half precision floats using bit ops
|
||
|
|
// Sign is the most significant bit for both half and bfloat16 types
|
||
|
|
inline c10::Half copysign(c10::Half a, c10::Half b) {
|
||
|
|
return c10::Half((a.x & 0x7fff) | (b.x & 0x8000), c10::Half::from_bits());
|
||
|
|
}
|
||
|
|
|
||
|
|
inline c10::BFloat16 copysign(c10::BFloat16 a, c10::BFloat16 b) {
|
||
|
|
return c10::BFloat16(
|
||
|
|
(a.x & 0x7fff) | (b.x & 0x8000), c10::BFloat16::from_bits());
|
||
|
|
}
|
||
|
|
|
||
|
|
} // namespace c10
|