mirror/eigen
2
0
Fork 0
mirror of https://gitlab.com/libeigen/eigen.git synced 2025-03-13 18:37:27 +08:00
Code Issues Packages Projects Releases Wiki Activity

Revert PR-292. After further investigation, the memcpy->memmove change was only good for Haswell on older versions of glibc. Adding a switch for small sizes is perhaps useful for string copies, but also has an overhead for larger sizes, making it a poor trade-off for general memcpy.

Browse Source 
This PR also removes a couple of unnecessary semi-colons in Eigen/src/Core/AssignEvaluator.h that caused compiler warning everywhere.
This commit is contained in:
Rasmus Munk Larsen 2017-01-26 12:46:06 -08:00
parent 25a1703579
commit edaa0fc5d1
5 changed files with 6 additions and 41 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
Eigen/src/Core/AssignEvaluator.h
View File

@ -708,7 +708,7 @@ void resize_if_allowed(DstXprType &dst, const SrcXprType& src, const Functor &/*
EIGEN_ONLY_USED_FOR_DEBUG(dst);
EIGEN_ONLY_USED_FOR_DEBUG(src);
eigen_assert(dst.rows() == src.rows() && dst.cols() == src.cols());
};
}
template<typename DstXprType,typename SrcXprType, typename T1, typename T2>
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
@ -719,7 +719,7 @@ void resize_if_allowed(DstXprType &dst, const SrcXprType& src, const internal::a
if(((dst.rows()!=dstRows) || (dst.cols()!=dstCols)))
dst.resize(dstRows, dstCols);
eigen_assert(dst.rows() == dstRows && dst.cols() == dstCols);
};
}
template<typename DstXprType, typename SrcXprType, typename Functor>
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void call_dense_assignment_loop(DstXprType& dst, const SrcXprType& src, const Functor &func)

37
Eigen/src/Core/util/Memory.h
View File

@ -74,41 +74,6 @@ inline void throw_std_bad_alloc()
#endif
}
EIGEN_DEVICE_FUNC
inline void fast_memcpy(void* dst, const void* src, size_t size) {
#if defined(__CUDA__) || defined(__ANDROID__)
::memcpy(dst, src, size);
#else
switch(size) {
// Most compilers will generate inline code for fixed sizes,
// which is significantly faster for small copies.
case 1: memcpy(dst, src, 1); break;
case 2: memcpy(dst, src, 2); break;
case 3: memcpy(dst, src, 3); break;
case 4: memcpy(dst, src, 4); break;
case 5: memcpy(dst, src, 5); break;
case 6: memcpy(dst, src, 6); break;
case 7: memcpy(dst, src, 7); break;
case 8: memcpy(dst, src, 8); break;
case 9: memcpy(dst, src, 9); break;
case 10: memcpy(dst, src, 10); break;
case 11: memcpy(dst, src, 11); break;
case 12: memcpy(dst, src, 12); break;
case 13: memcpy(dst, src, 13); break;
case 14: memcpy(dst, src, 14); break;
case 15: memcpy(dst, src, 15); break;
case 16: memcpy(dst, src, 16); break;
#ifdef EIGEN_OS_LINUX
// On Linux, memmove appears to be faster than memcpy for
// large sizes, strangely enough.
default: memmove(dst, src, size); break;
#else
default: memcpy(dst, src, size); break;
#endif
}
#endif
}
/*****************************************************************************
*** Implementation of handmade aligned functions ***
*****************************************************************************/
@ -528,7 +493,7 @@ template<typename T> struct smart_copy_helper<T,true> {
IntPtr size = IntPtr(end)-IntPtr(start);
if(size==0) return;
eigen_internal_assert(start!=0 && end!=0 && target!=0);
fast_memcpy(target, start, size);
memcpy(target, start, size);
}
};

2
unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h
View File

@ -56,7 +56,7 @@ void pack_simple(Scalar * dst, const Scalar * src, Index cols, Index rows, Index
} else {
// Naive memcpy calls
for (Index col = 0; col < cols; ++col) {
internal::fast_memcpy(dst + col*lddst, src + col*ldsrc, rows*sizeof(Scalar));
memcpy(dst + col*lddst, src + col*ldsrc, rows*sizeof(Scalar));
}
}
}

2
unsupported/Eigen/CXX11/src/Tensor/TensorDeviceDefault.h
View File

@ -22,7 +22,7 @@ struct DefaultDevice {
internal::aligned_free(buffer);
}
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void memcpy(void* dst, const void* src, size_t n) const {
internal::fast_memcpy(dst, src, n);
::memcpy(dst, src, n);
}
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void memcpyHostToDevice(void* dst, const void* src, size_t n) const {
memcpy(dst, src, n);

2
unsupported/Eigen/CXX11/src/Tensor/TensorDeviceThreadPool.h
View File

@ -106,7 +106,7 @@ struct ThreadPoolDevice {
}
EIGEN_STRONG_INLINE void memcpy(void* dst, const void* src, size_t n) const {
internal::fast_memcpy(dst, src, n);
::memcpy(dst, src, n);
}
EIGEN_STRONG_INLINE void memcpyHostToDevice(void* dst, const void* src, size_t n) const {
memcpy(dst, src, n);