devtools/eigen
1
0
Fork 0
mirror of https://gitlab.com/libeigen/eigen.git synced 2026-04-10 11:34:33 +08:00
Code Issues Packages Projects Releases Wiki Activity

Merged in rmlarsen/eigen2 (pull request PR-292)

Browse Source 
Adds a fast memcpy function to Eigen.
This commit is contained in:
Benoit Steiner
2017-01-25 00:14:04 +00:00
parent 7b6aaa3440 3be5ee2352
commit e96c77668d
5 changed files with 54 additions and 19 deletions
Download Patch File Download Diff File Expand all files Collapse all files

63
Eigen/src/Core/util/Memory.h
View File

@@ -63,7 +63,7 @@ namespace Eigen {
namespace internal {
EIGEN_DEVICE_FUNC
EIGEN_DEVICE_FUNC
inline void throw_std_bad_alloc()
{
#ifdef EIGEN_EXCEPTIONS
@@ -74,6 +74,41 @@ 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 ***
*****************************************************************************/
@@ -114,7 +149,7 @@ inline void* handmade_aligned_realloc(void* ptr, std::size_t size, std::size_t =
void *previous_aligned = static_cast<char *>(original)+previous_offset;
if(aligned!=previous_aligned)
std::memmove(aligned, previous_aligned, size);
*(reinterpret_cast<void**>(aligned) - 1) = original;
return aligned;
}
@@ -142,7 +177,7 @@ EIGEN_DEVICE_FUNC inline void check_that_malloc_is_allowed()
{
eigen_assert(is_malloc_allowed() && "heap allocation is forbidden (EIGEN_RUNTIME_NO_MALLOC is defined and g_is_malloc_allowed is false)");
}
#else
#else
EIGEN_DEVICE_FUNC inline void check_that_malloc_is_allowed()
{}
#endif
@@ -471,8 +506,8 @@ EIGEN_DEVICE_FUNC inline Index first_default_aligned(const Scalar* array, Index
}
/** \internal Returns the smallest integer multiple of \a base and greater or equal to \a size
*/
template<typename Index>
*/
template<typename Index>
inline Index first_multiple(Index size, Index base)
{
return ((size+base-1)/base)*base;
@@ -493,7 +528,7 @@ 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);
memcpy(target, start, size);
fast_memcpy(target, start, size);
}
};
@@ -502,7 +537,7 @@ template<typename T> struct smart_copy_helper<T,false> {
{ std::copy(start, end, target); }
};
// intelligent memmove. falls back to std::memmove for POD types, uses std::copy otherwise.
// intelligent memmove. falls back to std::memmove for POD types, uses std::copy otherwise.
template<typename T, bool UseMemmove> struct smart_memmove_helper;
template<typename T> void smart_memmove(const T* start, const T* end, T* target)
@@ -522,15 +557,15 @@ template<typename T> struct smart_memmove_helper<T,true> {
template<typename T> struct smart_memmove_helper<T,false> {
static inline void run(const T* start, const T* end, T* target)
{
{
if (UIntPtr(target) < UIntPtr(start))
{
std::copy(start, end, target);
}
else
else
{
std::ptrdiff_t count = (std::ptrdiff_t(end)-std::ptrdiff_t(start)) / sizeof(T);
std::copy_backward(start, end, target + count);
std::copy_backward(start, end, target + count);
}
}
};
@@ -603,7 +638,7 @@ template<typename T> void swap(scoped_array<T> &a,scoped_array<T> &b)
{
std::swap(a.ptr(),b.ptr());
}
} // end namespace internal
/** \internal
@@ -622,7 +657,7 @@ template<typename T> void swap(scoped_array<T> &a,scoped_array<T> &b)
* The underlying stack allocation function can controlled with the EIGEN_ALLOCA preprocessor token.
*/
#ifdef EIGEN_ALLOCA
#if EIGEN_DEFAULT_ALIGN_BYTES>0
// We always manually re-align the result of EIGEN_ALLOCA.
// If alloca is already aligned, the compiler should be smart enough to optimize away the re-alignment.
@@ -645,7 +680,7 @@ template<typename T> void swap(scoped_array<T> &a,scoped_array<T> &b)
Eigen::internal::check_size_for_overflow<TYPE>(SIZE); \
TYPE* NAME = (BUFFER)!=0 ? BUFFER : reinterpret_cast<TYPE*>(Eigen::internal::aligned_malloc(sizeof(TYPE)*SIZE)); \
Eigen::internal::aligned_stack_memory_handler<TYPE> EIGEN_CAT(NAME,_stack_memory_destructor)((BUFFER)==0 ? NAME : 0,SIZE,true)
#endif
@@ -701,7 +736,7 @@ template<typename T> void swap(scoped_array<T> &a,scoped_array<T> &b)
* Example:
* \code
* // Matrix4f requires 16 bytes alignment:
* std::map< int, Matrix4f, std::less<int>,
* std::map< int, Matrix4f, std::less<int>,
* aligned_allocator<std::pair<const int, Matrix4f> > > my_map_mat4;
* // Vector3f does not require 16 bytes alignment, no need to use Eigen's allocator:
* std::map< int, Vector3f > my_map_vec3;

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) {
memcpy(dst + col*lddst, src + col*ldsrc, rows*sizeof(Scalar));
internal::fast_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 {
::memcpy(dst, src, n);
internal::fast_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 {
::memcpy(dst, src, n);
internal::fast_memcpy(dst, src, n);
}
EIGEN_STRONG_INLINE void memcpyHostToDevice(void* dst, const void* src, size_t n) const {
memcpy(dst, src, n);

4
unsupported/Eigen/CXX11/src/Tensor/TensorFFT.h
View File

@@ -253,7 +253,7 @@ struct TensorEvaluator<const TensorFFTOp<FFT, ArgType, FFTResultType, FFTDir>, D
// get data into line_buf
const Index stride = m_strides[dim];
if (stride == 1) {
memcpy(line_buf, &buf[base_offset], line_len*sizeof(ComplexScalar));
m_device.memcpy(line_buf, &buf[base_offset], line_len*sizeof(ComplexScalar));
} else {
Index offset = base_offset;
for (int j = 0; j < line_len; ++j, offset += stride) {
@@ -271,7 +271,7 @@ struct TensorEvaluator<const TensorFFTOp<FFT, ArgType, FFTResultType, FFTDir>, D
// write back
if (FFTDir == FFT_FORWARD && stride == 1) {
memcpy(&buf[base_offset], line_buf, line_len*sizeof(ComplexScalar));
m_device.memcpy(&buf[base_offset], line_buf, line_len*sizeof(ComplexScalar));
} else {
Index offset = base_offset;
const ComplexScalar div_factor = ComplexScalar(1.0 / line_len, 0);