CFD/eigen
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Fix CUDA constexpr issues for numeric_limits.

Browse Source 
Some CUDA/HIP constants fail on device with `constexpr` since they
internally rely on non-constexpr functions, e.g.
```
\#define CUDART_INF_F            __int_as_float(0x7f800000)
```
This fails for cuda-clang (though passes with nvcc). These constants are
currently used by `device::numeric_limits`.  For portability, we
need to remove `constexpr` from the affected functions.

For C++11 or higher, we should be able to rely on the `std::numeric_limits`
versions anyways, since the methods themselves are now `constexpr`, so
should be supported on device (clang/hipcc natively, nvcc with
`--expr-relaxed-constexpr`).
This commit is contained in:
Antonio Sanchez 2021-03-25 12:37:25 -07:00 committed by Antonio Sánchez
parent af1247fbc1
commit 78ee3d6261
3 changed files with 26 additions and 10 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
Eigen/src/Core/util/Macros.h
View File

@ -763,8 +763,6 @@
#if EIGEN_MAX_CPP_VER>=14 && (EIGEN_COMP_CXXVER >= 11 && (EIGEN_COMP_CLANG || EIGEN_COMP_NVCC >= 70500))
#define EIGEN_HAS_CONSTEXPR 1
#endif
#elif defined(EIGEN_HIPCC)
// Skip const_expr on the HIP platform
#elif EIGEN_MAX_CPP_VER>=14 && (__has_feature(cxx_relaxed_constexpr) || (EIGEN_COMP_CXXVER >= 14) || \
(EIGEN_GNUC_AT_LEAST(4,8) && (EIGEN_COMP_CXXVER >= 11)) || \
(EIGEN_COMP_CLANG >= 306 && (EIGEN_COMP_CXXVER >= 11)))

16
Eigen/src/Core/util/Meta.h
View File

@ -275,7 +275,7 @@ template<bool Condition, typename T=void> struct enable_if;
template<typename T> struct enable_if<true,T>
{ typedef T type; };
#if defined(EIGEN_GPU_COMPILE_PHASE)
#if defined(EIGEN_GPU_COMPILE_PHASE) && !EIGEN_HAS_CXX11
#if !defined(__FLT_EPSILON__)
#define __FLT_EPSILON__ FLT_EPSILON
#define __DBL_EPSILON__ DBL_EPSILON
@ -296,7 +296,7 @@ template<> struct numeric_limits<float>
{
EIGEN_DEVICE_FUNC EIGEN_CONSTEXPR
static float epsilon() { return __FLT_EPSILON__; }
EIGEN_DEVICE_FUNC EIGEN_CONSTEXPR
EIGEN_DEVICE_FUNC
static float (max)() {
#if defined(EIGEN_CUDA_ARCH)
return CUDART_MAX_NORMAL_F;
@ -306,7 +306,7 @@ template<> struct numeric_limits<float>
}
EIGEN_DEVICE_FUNC EIGEN_CONSTEXPR
static float (min)() { return FLT_MIN; }
EIGEN_DEVICE_FUNC EIGEN_CONSTEXPR
EIGEN_DEVICE_FUNC
static float infinity() {
#if defined(EIGEN_CUDA_ARCH)
return CUDART_INF_F;
@ -314,7 +314,7 @@ template<> struct numeric_limits<float>
return HIPRT_INF_F;
#endif
}
EIGEN_DEVICE_FUNC EIGEN_CONSTEXPR
EIGEN_DEVICE_FUNC
static float quiet_NaN() {
#if defined(EIGEN_CUDA_ARCH)
return CUDART_NAN_F;
@ -331,7 +331,7 @@ template<> struct numeric_limits<double>
static double (max)() { return DBL_MAX; }
EIGEN_DEVICE_FUNC EIGEN_CONSTEXPR
static double (min)() { return DBL_MIN; }
EIGEN_DEVICE_FUNC EIGEN_CONSTEXPR
EIGEN_DEVICE_FUNC
static double infinity() {
#if defined(EIGEN_CUDA_ARCH)
return CUDART_INF;
@ -339,7 +339,7 @@ template<> struct numeric_limits<double>
return HIPRT_INF;
#endif
}
EIGEN_DEVICE_FUNC EIGEN_CONSTEXPR
EIGEN_DEVICE_FUNC
static double quiet_NaN() {
#if defined(EIGEN_CUDA_ARCH)
return CUDART_NAN;
@ -414,7 +414,7 @@ template<> struct numeric_limits<bool>
}
#endif
#endif // defined(EIGEN_GPU_COMPILE_PHASE) && !EIGEN_HAS_CXX11
/** \internal
* A base class do disable default copy ctor and copy assignment operator.
@ -761,7 +761,7 @@ template<typename T> EIGEN_DEVICE_FUNC void swap(T &a, T &b) { T tmp = b; b =
template<typename T> EIGEN_STRONG_INLINE void swap(T &a, T &b) { std::swap(a,b); }
#endif
#if defined(EIGEN_GPU_COMPILE_PHASE)
#if defined(EIGEN_GPU_COMPILE_PHASE) && !EIGEN_HAS_CXX11
using internal::device::numeric_limits;
#else
using std::numeric_limits;

18
test/gpu_basic.cu
View File

@ -343,6 +343,21 @@ struct matrix_inverse {
}
};
template<typename T>
struct numeric_limits_test {
EIGEN_DEVICE_FUNC
void operator()(int i, const typename T::Scalar* in, typename T::Scalar* out) const
{
EIGEN_UNUSED_VARIABLE(in)
int out_idx = i * 5;
out[out_idx++] = numext::numeric_limits<float>::epsilon();
out[out_idx++] = (numext::numeric_limits<float>::max)();
out[out_idx++] = (numext::numeric_limits<float>::min)();
out[out_idx++] = numext::numeric_limits<float>::infinity();
out[out_idx++] = numext::numeric_limits<float>::quiet_NaN();
}
};
template<typename Type1, typename Type2>
bool verifyIsApproxWithInfsNans(const Type1& a, const Type2& b, typename Type1::Scalar* = 0) // Enabled for Eigen's type only
{
@ -434,6 +449,9 @@ EIGEN_DECLARE_TEST(gpu_basic)
CALL_SUBTEST( run_and_compare_to_gpu(complex_operators<Vector3cf>(), nthreads, cfin, cfout) );
CALL_SUBTEST( test_with_infs_nans(complex_sqrt<Vector3cf>(), nthreads, cfin, cfout) );
// numeric_limits
CALL_SUBTEST( test_with_infs_nans(numeric_limits_test<Vector3f>(), 1, in, out) );
#if defined(__NVCC__)
// FIXME
// These subtests compiles only with nvcc and fail with HIPCC and clang-cuda