Skip to content

Commit

Permalink
Fix <cmath> intrinsics for CUDA (#1886)
Browse files Browse the repository at this point in the history 
Co-authored-by: Casey Carter <[email protected]>
  • Loading branch information
@StephanTLavavej @CaseyCarter
StephanTLavavej and CaseyCarter authored Apr 30, 2021
1 parent f675d68 commit 8eca13a
Showing 1 changed file with 69 additions and 62 deletions.
131 changes: 69 additions & 62 deletions stl/inc/cmath
View file
Original file line number Diff line number Diff line change
Expand Up @@ -12,9 +12,15 @@
#include <cstdlib>
#include <xtr1common>

#if !defined(_M_CEE) && !defined(__clang__)
#if !defined(_M_CEE) && !defined(__clang__) && !defined(__CUDACC__) && !defined(__INTEL_COMPILER)
#define _HAS_CMATH_INTRINSICS 1
#else // ^^^ intrinsics available ^^^ / vvv intrinsics unavailable vvv
#define _HAS_CMATH_INTRINSICS 0
#endif // ^^^ intrinsics unavailable ^^^

#if _HAS_CMATH_INTRINSICS
#include <intrin0.h>
#endif // !defined(_M_CEE) && !defined(__clang__)
#endif // _HAS_CMATH_INTRINSICS

#if _HAS_CXX20
#include <xutility>
Expand Down Expand Up @@ -60,23 +66,23 @@ _NODISCARD _Check_return_ inline float cbrt(_In_ float _Xx) noexcept /* strength
}

_NODISCARD _Check_return_ inline float ceil(_In_ float _Xx) noexcept /* strengthened */ {
#ifdef _M_CEE
return _CSTD ceilf(_Xx);
#if _HAS_CMATH_INTRINSICS
return __ceilf(_Xx);
#elif defined(__clang__)
return __builtin_ceilf(_Xx);
#else // ^^^ __clang__ / !__clang__ vvv
return __ceilf(_Xx);
#endif // __clang__
#else // ^^^ defined(__clang__) ^^^ / vvv intrinsics unavailable vvv
return _CSTD ceilf(_Xx);
#endif // ^^^ intrinsics unavailable ^^^
}

_NODISCARD _Check_return_ inline float copysign(_In_ float _Number, _In_ float _Sign) noexcept /* strengthened */ {
#ifdef _M_CEE
return _CSTD copysignf(_Number, _Sign);
#if _HAS_CMATH_INTRINSICS
return __copysignf(_Number, _Sign);
#elif defined(__clang__)
return __builtin_copysignf(_Number, _Sign);
#else // ^^^ __clang__ / !__clang__ vvv
return __copysignf(_Number, _Sign);
#endif // __clang__
#else // ^^^ defined(__clang__) ^^^ / vvv intrinsics unavailable vvv
return _CSTD copysignf(_Number, _Sign);
#endif // ^^^ intrinsics unavailable ^^^
}

_NODISCARD _Check_return_ inline float cos(_In_ float _Xx) noexcept /* strengthened */ {
Expand Down Expand Up @@ -116,13 +122,13 @@ _NODISCARD _Check_return_ inline float fdim(_In_ float _Xx, _In_ float _Yx) noex
}

_NODISCARD _Check_return_ inline float floor(_In_ float _Xx) noexcept /* strengthened */ {
#ifdef _M_CEE
return _CSTD floorf(_Xx);
#if _HAS_CMATH_INTRINSICS
return __floorf(_Xx);
#elif defined(__clang__)
return __builtin_floorf(_Xx);
#else // ^^^ __clang__ / !__clang__ vvv
return __floorf(_Xx);
#endif // __clang__
#else // ^^^ defined(__clang__) ^^^ / vvv intrinsics unavailable vvv
return _CSTD floorf(_Xx);
#endif // ^^^ intrinsics unavailable ^^^
}

_NODISCARD _Check_return_ inline float fma(_In_ float _Xx, _In_ float _Yx, _In_ float _Zx) noexcept /* strengthened */ {
Expand Down Expand Up @@ -230,13 +236,13 @@ _NODISCARD _Check_return_ inline float rint(_In_ float _Xx) noexcept /* strength
}

_NODISCARD _Check_return_ inline float round(_In_ float _Xx) noexcept /* strengthened */ {
#ifdef _M_CEE
return _CSTD roundf(_Xx);
#if _HAS_CMATH_INTRINSICS
return __roundf(_Xx);
#elif defined(__clang__)
return __builtin_roundf(_Xx);
#else // ^^^ __clang__ / !__clang__ vvv
return __roundf(_Xx);
#endif // __clang__
#else // ^^^ defined(__clang__) ^^^ / vvv intrinsics unavailable vvv
return _CSTD roundf(_Xx);
#endif // ^^^ intrinsics unavailable ^^^
}

_NODISCARD _Check_return_ inline float scalbln(_In_ float _Xx, _In_ long _Yx) noexcept /* strengthened */ {
Expand Down Expand Up @@ -272,13 +278,13 @@ _NODISCARD _Check_return_ inline float tgamma(_In_ float _Xx) noexcept /* streng
}

_NODISCARD _Check_return_ inline float trunc(_In_ float _Xx) noexcept /* strengthened */ {
#ifdef _M_CEE
return _CSTD truncf(_Xx);
#if _HAS_CMATH_INTRINSICS
return __truncf(_Xx);
#elif defined(__clang__)
return __builtin_truncf(_Xx);
#else // ^^^ __clang__ / !__clang__ vvv
return __truncf(_Xx);
#endif // __clang__
#else // ^^^ defined(__clang__) ^^^ / vvv intrinsics unavailable vvv
return _CSTD truncf(_Xx);
#endif // ^^^ intrinsics unavailable ^^^
}

_NODISCARD _Check_return_ inline long double acos(_In_ long double _Xx) noexcept /* strengthened */ {
Expand Down Expand Up @@ -315,24 +321,24 @@ _NODISCARD _Check_return_ inline long double cbrt(_In_ long double _Xx) noexcept
}

_NODISCARD _Check_return_ inline long double ceil(_In_ long double _Xx) noexcept /* strengthened */ {
#ifdef _M_CEE
return _CSTD ceill(_Xx);
#if _HAS_CMATH_INTRINSICS
return __ceil(static_cast<double>(_Xx));
#elif defined(__clang__)
return __builtin_ceill(_Xx);
#else // ^^^ __clang__ / !__clang__ vvv
return __ceil(static_cast<double>(_Xx));
#endif // __clang__
#else // ^^^ defined(__clang__) ^^^ / vvv intrinsics unavailable vvv
return _CSTD ceill(_Xx);
#endif // ^^^ intrinsics unavailable ^^^
}

_NODISCARD _Check_return_ inline long double copysign(_In_ long double _Number, _In_ long double _Sign) noexcept
/* strengthened */ {
#ifdef _M_CEE
return _CSTD copysignl(_Number, _Sign);
#if _HAS_CMATH_INTRINSICS
return __copysign(static_cast<double>(_Number), static_cast<double>(_Sign));
#elif defined(__clang__)
return __builtin_copysignl(_Number, _Sign);
#else // ^^^ __clang__ / !__clang__ vvv
return __copysign(static_cast<double>(_Number), static_cast<double>(_Sign));
#endif // __clang__
#else // ^^^ defined(__clang__) ^^^ / vvv intrinsics unavailable vvv
return _CSTD copysignl(_Number, _Sign);
#endif // ^^^ intrinsics unavailable ^^^
}

_NODISCARD _Check_return_ inline long double cos(_In_ long double _Xx) noexcept /* strengthened */ {
Expand Down Expand Up @@ -373,13 +379,13 @@ _NODISCARD _Check_return_ inline long double fdim(_In_ long double _Xx, _In_ lon
}

_NODISCARD _Check_return_ inline long double floor(_In_ long double _Xx) noexcept /* strengthened */ {
#ifdef _M_CEE
return _CSTD floorl(_Xx);
#if _HAS_CMATH_INTRINSICS
return __floor(static_cast<double>(_Xx));
#elif defined(__clang__)
return __builtin_floorl(_Xx);
#else // ^^^ __clang__ / !__clang__ vvv
return __floor(static_cast<double>(_Xx));
#endif // __clang__
#else // ^^^ defined(__clang__) ^^^ / vvv intrinsics unavailable vvv
return _CSTD floorl(_Xx);
#endif // ^^^ intrinsics unavailable ^^^
}

_NODISCARD _Check_return_ inline long double fma(
Expand Down Expand Up @@ -496,13 +502,13 @@ _NODISCARD _Check_return_ inline long double rint(_In_ long double _Xx) noexcept
}

_NODISCARD _Check_return_ inline long double round(_In_ long double _Xx) noexcept /* strengthened */ {
#ifdef _M_CEE
return _CSTD roundl(_Xx);
#if _HAS_CMATH_INTRINSICS
return __round(static_cast<double>(_Xx));
#elif defined(__clang__)
return __builtin_roundl(_Xx);
#else // ^^^ __clang__ / !__clang__ vvv
return __round(static_cast<double>(_Xx));
#endif // __clang__
#else // ^^^ defined(__clang__) ^^^ / vvv intrinsics unavailable vvv
return _CSTD roundl(_Xx);
#endif // ^^^ intrinsics unavailable ^^^
}

_NODISCARD _Check_return_ inline long double scalbln(_In_ long double _Xx, _In_ long _Yx) noexcept /* strengthened */ {
Expand Down Expand Up @@ -538,13 +544,13 @@ _NODISCARD _Check_return_ inline long double tgamma(_In_ long double _Xx) noexce
}

_NODISCARD _Check_return_ inline long double trunc(_In_ long double _Xx) noexcept /* strengthened */ {
#ifdef _M_CEE
return _CSTD truncl(_Xx);
#if _HAS_CMATH_INTRINSICS
return __trunc(static_cast<double>(_Xx));
#elif defined(__clang__)
return __builtin_truncl(_Xx);
#else // ^^^ __clang__ / !__clang__ vvv
return __trunc(static_cast<double>(_Xx));
#endif // __clang__
#else // ^^^ defined(__clang__) ^^^ / vvv intrinsics unavailable vvv
return _CSTD truncl(_Xx);
#endif // ^^^ intrinsics unavailable ^^^
}


Expand Down Expand Up @@ -598,13 +604,13 @@ _STD _Common_float_type_t<_Ty1, _Ty2> remquo(_Ty1 _Left, _Ty2 _Right, int* _Pquo
#define _GENERIC_MATH1R(FUN, RET) _GENERIC_MATH1_BASE(FUN, RET, _CSTD FUN)
#define _GENERIC_MATH1(FUN) _GENERIC_MATH1R(FUN, double)

#ifdef _M_CEE
#define _GENERIC_MATH1I(FUN, CLANG_INTRIN, MSVC_INTRIN) _GENERIC_MATH1_BASE(FUN, double, _CSTD FUN)
#if _HAS_CMATH_INTRINSICS
#define _GENERIC_MATH1I(FUN, CLANG_INTRIN, MSVC_INTRIN) _GENERIC_MATH1_BASE(FUN, double, MSVC_INTRIN)
#elif defined(__clang__)
#define _GENERIC_MATH1I(FUN, CLANG_INTRIN, MSVC_INTRIN) _GENERIC_MATH1_BASE(FUN, double, CLANG_INTRIN)
#else // ^^^ __clang__ / !__clang__ vvv
#define _GENERIC_MATH1I(FUN, CLANG_INTRIN, MSVC_INTRIN) _GENERIC_MATH1_BASE(FUN, double, MSVC_INTRIN)
#endif // __clang__
#else // ^^^ defined(__clang__) ^^^ / vvv intrinsics unavailable vvv
#define _GENERIC_MATH1I(FUN, CLANG_INTRIN, MSVC_INTRIN) _GENERIC_MATH1_BASE(FUN, double, _CSTD FUN)
#endif // ^^^ intrinsics unavailable ^^^

#define _GENERIC_MATH1X(FUN, ARG2) \
template <class _Ty, _STD enable_if_t<_STD is_integral_v<_Ty>, int> = 0> \
Expand All @@ -622,13 +628,13 @@ _STD _Common_float_type_t<_Ty1, _Ty2> remquo(_Ty1 _Left, _Ty2 _Right, int* _Pquo

#define _GENERIC_MATH2(FUN) _GENERIC_MATH2_BASE(FUN, _CSTD FUN)

#ifdef _M_CEE
#define _GENERIC_MATH2I(FUN, CLANG_INTRIN, MSVC_INTRIN) _GENERIC_MATH2_BASE(FUN, _CSTD FUN)
#if _HAS_CMATH_INTRINSICS
#define _GENERIC_MATH2I(FUN, CLANG_INTRIN, MSVC_INTRIN) _GENERIC_MATH2_BASE(FUN, MSVC_INTRIN)
#elif defined(__clang__)
#define _GENERIC_MATH2I(FUN, CLANG_INTRIN, MSVC_INTRIN) _GENERIC_MATH2_BASE(FUN, CLANG_INTRIN)
#else // ^^^ __clang__ / !__clang__ vvv
#define _GENERIC_MATH2I(FUN, CLANG_INTRIN, MSVC_INTRIN) _GENERIC_MATH2_BASE(FUN, MSVC_INTRIN)
#endif // __clang__
#else // ^^^ defined(__clang__) ^^^ / vvv intrinsics unavailable vvv
#define _GENERIC_MATH2I(FUN, CLANG_INTRIN, MSVC_INTRIN) _GENERIC_MATH2_BASE(FUN, _CSTD FUN)
#endif // ^^^ intrinsics unavailable ^^^

// The following order matches N4820 26.8.1 [cmath.syn].
_GENERIC_MATH1(acos)
Expand Down Expand Up @@ -702,6 +708,7 @@ _GENERIC_MATH2(fmin)
#undef _GENERIC_MATH2_BASE
#undef _GENERIC_MATH2
#undef _GENERIC_MATH2I
#undef _HAS_CMATH_INTRINSICS


_STD_BEGIN
Expand Down

0 comments on commit 8eca13a

Please sign in to comment.