GitHub-Proxy/eigen
1
0
Fork 0
mirror of https://gitlab.com/libeigen/eigen.git synced 2025-07-04 04:05:19 +08:00
Code Issues Packages Projects Releases Wiki Activity

Added proper support for fast 64bit integer division on CUDA

Browse Source
This commit is contained in:
Benoit Steiner 2015-11-20 11:09:46 -08:00
parent f37a5f1c53
commit 383d1cc2ed
1 changed files with 13 additions and 20 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

33
unsupported/Eigen/CXX11/src/Tensor/TensorIntDiv.h
View File

@ -61,13 +61,8 @@ namespace {
template <typename T> template <typename T>
struct DividerTraits { struct DividerTraits {
#if defined(__SIZEOF_INT128__) && !defined(__CUDACC__)
typedef typename UnsignedTraits<T>::type type; typedef typename UnsignedTraits<T>::type type;
static const int N = sizeof(T) * 8; static const int N = sizeof(T) * 8;
#else
typedef uint32_t type;
static const int N = 32;
#endif
}; };
template <typename T> template <typename T>
@ -79,40 +74,38 @@ namespace {
#endif #endif
} }
#if defined(__CUDA_ARCH__)
template <typename T>
EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE uint64_t muluh(const uint64_t a, const T b) {
return __umul64hi(a, b);
}
#else
template <typename T> template <typename T>
EIGEN_ALWAYS_INLINE uint64_t muluh(const uint64_t a, const T b) { EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE uint64_t muluh(const uint64_t a, const T b) {
#if defined(__SIZEOF_INT128__) && !defined(__CUDACC__) #if defined(__CUDA_ARCH__)
return __umul64hi(a, b);
#elif defined(__SIZEOF_INT128__)
__uint128_t v = static_cast<__uint128_t>(a) * static_cast<__uint128_t>(b); __uint128_t v = static_cast<__uint128_t>(a) * static_cast<__uint128_t>(b);
return static_cast<uint64_t>(v >> 64); return static_cast<uint64_t>(v >> 64);
#else #else
EIGEN_STATIC_ASSERT(sizeof(T) == 4, YOU_MADE_A_PROGRAMMING_MISTAKE); return (TensorUInt128<static_val<0>, uint64_t>(a) * TensorUInt128<static_val<0>, uint64_t>(b)).upper();
return (a * b) >> 32;
#endif #endif
} }
#endif
template <int N, typename T> template <int N, typename T>
struct DividerHelper { struct DividerHelper {
static EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE uint32_t computeMultiplier (const int log_div, const T divider) { static EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE uint32_t computeMultiplier(const int log_div, const T divider) {
EIGEN_STATIC_ASSERT(N == 32, YOU_MADE_A_PROGRAMMING_MISTAKE); EIGEN_STATIC_ASSERT(N == 32, YOU_MADE_A_PROGRAMMING_MISTAKE);
return static_cast<uint32_t>((static_cast<uint64_t>(1) << (N+log_div)) / divider - (static_cast<uint64_t>(1) << N) + 1); return static_cast<uint32_t>((static_cast<uint64_t>(1) << (N+log_div)) / divider - (static_cast<uint64_t>(1) << N) + 1);
} }
}; };
#if defined(__SIZEOF_INT128__) && !defined(__CUDACC__)
template <typename T> template <typename T>
struct DividerHelper<64, T> { struct DividerHelper<64, T> {
static EIGEN_ALWAYS_INLINE uint64_t computeMultiplier(const int log_div, const T divider) { static EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE uint64_t computeMultiplier(const int log_div, const T divider) {
#if defined(__SIZEOF_INT128__) && !defined(__CUDA_ARCH__)
return static_cast<uint64_t>((static_cast<__uint128_t>(1) << (64+log_div)) / static_cast<__uint128_t>(divider) - (static_cast<__uint128_t>(1) << 64) + 1); return static_cast<uint64_t>((static_cast<__uint128_t>(1) << (64+log_div)) / static_cast<__uint128_t>(divider) - (static_cast<__uint128_t>(1) << 64) + 1);
#else
const uint64_t shift = 1ULL << log_div;
TensorUInt128<uint64_t, uint64_t> result = (TensorUInt128<uint64_t, static_val<0> >(shift, 0) / TensorUInt128<static_val<0>, uint64_t>(divider) - TensorUInt128<static_val<1>, static_val<0> >(1, 0) + TensorUInt128<static_val<0>, static_val<1> >(1));
return static_cast<uint64_t>(result);
#endif
} }
}; };
#endif
} }