From 4b41a1ef803bf3a9c2f7b076ac4c0e1b8c8723a3 Mon Sep 17 00:00:00 2001
From: Tomas Oberhuber <tomas.oberhuber@fjfi.cvut.cz>
Date: Tue, 28 May 2013 14:25:19 +0200
Subject: [PATCH] Fixing copying arrays with different types.
---
src/core/tnlCuda.h | 21 ++++++-
src/core/tnlHost.h | 9 ++-
src/implementation/core/memory-operations.h | 41 ++++++++++++-
src/implementation/core/tnlArray_impl.h | 7 ++-
src/implementation/core/tnlHost_impl.cpp | 48 +++++++--------
src/implementation/core/tnlHost_impl.h | 67 ++++++++++++---------
6 files changed, 131 insertions(+), 62 deletions(-)
diff --git a/src/core/tnlCuda.h b/src/core/tnlCuda.h
index 6d3dd31b0f..f0de588872 100644
--- a/src/core/tnlCuda.h
+++ b/src/core/tnlCuda.h
@@ -79,9 +79,12 @@ class tnlCuda
abort();
}
- template< typename Element, typename Index, typename Device >
- static bool memcpy( Element* destination,
- const Element* source,
+ template< typename DestinationElement,
+ typename SourceElement,
+ typename Index,
+ typename Device >
+ static bool memcpy( DestinationElement* destination,
+ const SourceElement* source,
const Index size )
{
switch( Device :: getDevice() )
@@ -94,6 +97,18 @@ class tnlCuda
return true;
}
+
+ template< typename Element, typename Index, typename Device >
+ static bool memcpy( Element* destination,
+ const Element* source,
+ const Index size )
+ {
+ return tnlCuda :: memcpy< Element, Element, Index, Device >
+ ( destination,
+ source,
+ size );
+ }
+
template< typename Element, typename Index, typename Device >
static bool memcmp( const Element* data1,
const Element* data2,
diff --git a/src/core/tnlHost.h b/src/core/tnlHost.h
index d2edb0ca7f..93f7ef84c7 100644
--- a/src/core/tnlHost.h
+++ b/src/core/tnlHost.h
@@ -49,6 +49,11 @@ class tnlHost
template< typename Element, typename Index >
static const Element& getArrayElementReference(const Element* data, const Index i );
+ template< typename DestinationElement, typename SourceElement, typename Index, typename Device >
+ static bool memcpy( DestinationElement* destination,
+ const SourceElement* source,
+ const Index size );
+
template< typename Element, typename Index, typename Device >
static bool memcpy( Element* destination,
const Element* source,
@@ -56,8 +61,8 @@ class tnlHost
template< typename Element, typename Index, typename Device >
static bool memcmp( const Element* data1,
- const Element* data2,
- const Index size );
+ const Element* data2,
+ const Index size );
template< typename Element, typename Index >
static bool memset( Element* destination,
diff --git a/src/implementation/core/memory-operations.h b/src/implementation/core/memory-operations.h
index 20f72e1435..cfc3fe2a41 100644
--- a/src/implementation/core/memory-operations.h
+++ b/src/implementation/core/memory-operations.h
@@ -23,6 +23,7 @@
#include <core/cuda/reduction-operations.h>
#include <core/mfuncs.h>
#include <tnlConfig.h>
+#include <string.h>
const int tnlGPUvsCPUTransferBufferSize( 1 << 20 );
@@ -115,16 +116,26 @@ bool setMemoryCuda( Element* data,
}
+template< typename DestinationElement, typename SourceElement, typename Index >
+bool copyMemoryHostToHost( DestinationElement* destination,
+ const SourceElement* source,
+ const Index size )
+{
+ for( Index i = 0; i < size; i ++ )
+ destination[ i ] = ( DestinationElement) source[ i ];
+ return true;
+}
+
template< typename Element, typename Index >
bool copyMemoryHostToHost( Element* destination,
const Element* source,
const Index size )
{
- for( Index i = 0; i < size; i ++ )
- destination[ i ] = source[ i ];
+ memcpy( destination, source, size * sizeof( Element ) );
return true;
}
+
template< typename Element, typename Index >
bool copyMemoryHostToCuda( Element* destination,
const Element* source,
@@ -170,6 +181,32 @@ bool copyMemoryCudaToHost( Element* destination,
#endif
}
+template< typename DestinationElement,
+ typename SourceElement,
+ typename Index >
+bool copyMemoryCudaToHost( DestinationElement* destination,
+ const SourceElement* source,
+ const Index size )
+{
+#ifdef HAVE_CUDA
+ abort(); // TODO: fix this
+ cudaMemcpy( destination,
+ source,
+ size * sizeof( Element ),
+ cudaMemcpyDeviceToHost );
+ if( ! checkCudaDevice )
+ {
+ cerr << "Transfer of data from CUDA device to host failed." << endl;
+ return false;
+ }
+ return true;
+#else
+ cerr << "CUDA support is missing on this system " << __FILE__ << " line " << __LINE__ << "." << endl;
+ return false;
+#endif
+}
+
+
template< typename Element, typename Index >
bool copyMemoryCudaToCuda( Element* destination,
const Element* source,
diff --git a/src/implementation/core/tnlArray_impl.h b/src/implementation/core/tnlArray_impl.h
index 466ec12e33..c96ad724d1 100644
--- a/src/implementation/core/tnlArray_impl.h
+++ b/src/implementation/core/tnlArray_impl.h
@@ -213,9 +213,10 @@ tnlArray< Element, Device, Index >&
<< "Source size: " << array. getSize() << endl
<< "Target name: " << this -> getName() << endl
<< "Target size: " << this -> getSize() << endl );
- Device :: template memcpy< typename Array :: ElementType,
- typename Array :: IndexType,
- typename Array :: DeviceType >
+ Device :: template memcpy< Element,
+ typename Array :: ElementType,
+ typename Array :: IndexType,
+ typename Array :: DeviceType >
( this -> getData(),
array. getData(),
array. getSize() );
diff --git a/src/implementation/core/tnlHost_impl.cpp b/src/implementation/core/tnlHost_impl.cpp
index 7408bdcad5..77ab56b2ad 100644
--- a/src/implementation/core/tnlHost_impl.cpp
+++ b/src/implementation/core/tnlHost_impl.cpp
@@ -80,30 +80,30 @@ template const float& tnlHost :: getArrayElementReference< float, lo
template const double& tnlHost :: getArrayElementReference< double, long int >( const double* data, const long int i );
template const long double& tnlHost :: getArrayElementReference< long double, long int >( const long double* data, const long int i );
-template bool tnlHost :: memcpy< char, int, tnlHost >( char* destination, const char* source, const int size );
-template bool tnlHost :: memcpy< int, int, tnlHost >( int* destination, const int* source, const int size );
-template bool tnlHost :: memcpy< long int, int, tnlHost >( long int* destination, const long int* source, const int size );
-template bool tnlHost :: memcpy< float, int, tnlHost >( float* destination, const float* source, const int size );
-template bool tnlHost :: memcpy< double, int, tnlHost >( double* destination, const double* source, const int size );
-template bool tnlHost :: memcpy< long double, int, tnlHost >( long double* destination, const long double* source, const int size );
-template bool tnlHost :: memcpy< char, long int, tnlHost >( char* destination, const char* source, const long int size );
-template bool tnlHost :: memcpy< int, long int, tnlHost >( int* destination, const int* source, const long int size );
-template bool tnlHost :: memcpy< long int, long int, tnlHost >( long int* destination, const long int* source, const long int size );
-template bool tnlHost :: memcpy< float, long int, tnlHost >( float* destination, const float* source, const long int size );
-template bool tnlHost :: memcpy< double, long int, tnlHost >( double* destination, const double* source, const long int size );
-template bool tnlHost :: memcpy< long double, long int, tnlHost >( long double* destination, const long double* source, const long int size );
-template bool tnlHost :: memcpy< char, int, tnlCuda >( char* destination, const char* source, const int size );
-template bool tnlHost :: memcpy< int, int, tnlCuda >( int* destination, const int* source, const int size );
-template bool tnlHost :: memcpy< long int, int, tnlCuda >( long int* destination, const long int* source, const int size );
-template bool tnlHost :: memcpy< float, int, tnlCuda >( float* destination, const float* source, const int size );
-template bool tnlHost :: memcpy< double, int, tnlCuda >( double* destination, const double* source, const int size );
-template bool tnlHost :: memcpy< long double, int, tnlCuda >( long double* destination, const long double* source, const int size );
-template bool tnlHost :: memcpy< char, long int, tnlCuda >( char* destination, const char* source, const long int size );
-template bool tnlHost :: memcpy< int, long int, tnlCuda >( int* destination, const int* source, const long int size );
-template bool tnlHost :: memcpy< long int, long int, tnlCuda >( long int* destination, const long int* source, const long int size );
-template bool tnlHost :: memcpy< float, long int, tnlCuda >( float* destination, const float* source, const long int size );
-template bool tnlHost :: memcpy< double, long int, tnlCuda >( double* destination, const double* source, const long int size );
-template bool tnlHost :: memcpy< long double, long int, tnlCuda >( long double* destination, const long double* source, const long int size );
+template bool tnlHost :: memcpy< char, char, int, tnlHost >( char* destination, const char* source, const int size );
+template bool tnlHost :: memcpy< int, int, int, tnlHost >( int* destination, const int* source, const int size );
+template bool tnlHost :: memcpy< long int, long int, int, tnlHost >( long int* destination, const long int* source, const int size );
+template bool tnlHost :: memcpy< float, float, int, tnlHost >( float* destination, const float* source, const int size );
+template bool tnlHost :: memcpy< double, double, int, tnlHost >( double* destination, const double* source, const int size );
+template bool tnlHost :: memcpy< long double, long double, int, tnlHost >( long double* destination, const long double* source, const int size );
+template bool tnlHost :: memcpy< char, char, long int, tnlHost >( char* destination, const char* source, const long int size );
+template bool tnlHost :: memcpy< int, int, long int, tnlHost >( int* destination, const int* source, const long int size );
+template bool tnlHost :: memcpy< long int, long int, long int, tnlHost >( long int* destination, const long int* source, const long int size );
+template bool tnlHost :: memcpy< float, float, long int, tnlHost >( float* destination, const float* source, const long int size );
+template bool tnlHost :: memcpy< double, double, long int, tnlHost >( double* destination, const double* source, const long int size );
+template bool tnlHost :: memcpy< long double, long double, long int, tnlHost >( long double* destination, const long double* source, const long int size );
+template bool tnlHost :: memcpy< char, char, int, tnlCuda >( char* destination, const char* source, const int size );
+template bool tnlHost :: memcpy< int, int, int, tnlCuda >( int* destination, const int* source, const int size );
+template bool tnlHost :: memcpy< long int, long int, int, tnlCuda >( long int* destination, const long int* source, const int size );
+template bool tnlHost :: memcpy< float, float, int, tnlCuda >( float* destination, const float* source, const int size );
+template bool tnlHost :: memcpy< double, double, int, tnlCuda >( double* destination, const double* source, const int size );
+template bool tnlHost :: memcpy< long double, long double, int, tnlCuda >( long double* destination, const long double* source, const int size );
+template bool tnlHost :: memcpy< char, char, long int, tnlCuda >( char* destination, const char* source, const long int size );
+template bool tnlHost :: memcpy< int, int, long int, tnlCuda >( int* destination, const int* source, const long int size );
+template bool tnlHost :: memcpy< long int, long int, long int, tnlCuda >( long int* destination, const long int* source, const long int size );
+template bool tnlHost :: memcpy< float, float, long int, tnlCuda >( float* destination, const float* source, const long int size );
+template bool tnlHost :: memcpy< double, double, long int, tnlCuda >( double* destination, const double* source, const long int size );
+template bool tnlHost :: memcpy< long double, long double, long int, tnlCuda >( long double* destination, const long double* source, const long int size );
template bool tnlHost :: memcmp< char, int, tnlHost >( const char* data1, const char* data2, const int size );
template bool tnlHost :: memcmp< int, int, tnlHost >( const int* data1, const int* data2, const int size );
diff --git a/src/implementation/core/tnlHost_impl.h b/src/implementation/core/tnlHost_impl.h
index c309906620..7be6841860 100644
--- a/src/implementation/core/tnlHost_impl.h
+++ b/src/implementation/core/tnlHost_impl.h
@@ -67,10 +67,10 @@ const Element& tnlHost :: getArrayElementReference(const Element* data, const In
return data[ i ];
};
-template< typename Element, typename Index, typename Device >
-bool tnlHost :: memcpy( Element* destination,
- const Element* source,
- const Index size )
+template< typename DestinationElement, typename SourceElement, typename Index, typename Device >
+bool tnlHost :: memcpy( DestinationElement* destination,
+ const SourceElement* source,
+ const Index size )
{
switch( Device :: getDevice() )
{
@@ -82,6 +82,17 @@ bool tnlHost :: memcpy( Element* destination,
return true;
};
+template< typename Element, typename Index, typename Device >
+bool tnlHost :: memcpy( Element* destination,
+ const Element* source,
+ const Index size )
+{
+ return tnlHost :: memcpy< Element, Element, Index, Device >
+ ( destination,
+ source,
+ size );
+};
+
template< typename Element, typename Index, typename Device >
bool tnlHost :: memcmp( const Element* data1,
const Element* data2,
@@ -168,30 +179,30 @@ extern template const float& tnlHost :: getArrayElementReference< float,
extern template const double& tnlHost :: getArrayElementReference< double, long int >( const double* data, const long int i );
extern template const long double& tnlHost :: getArrayElementReference< long double, long int >( const long double* data, const long int i );
-extern template bool tnlHost :: memcpy< char, int, tnlHost >( char* destination, const char* source, const int size );
-extern template bool tnlHost :: memcpy< int, int, tnlHost >( int* destination, const int* source, const int size );
-extern template bool tnlHost :: memcpy< long int, int, tnlHost >( long int* destination, const long int* source, const int size );
-extern template bool tnlHost :: memcpy< float, int, tnlHost >( float* destination, const float* source, const int size );
-extern template bool tnlHost :: memcpy< double, int, tnlHost >( double* destination, const double* source, const int size );
-extern template bool tnlHost :: memcpy< long double, int, tnlHost >( long double* destination, const long double* source, const int size );
-extern template bool tnlHost :: memcpy< char, long int, tnlHost >( char* destination, const char* source, const long int size );
-extern template bool tnlHost :: memcpy< int, long int, tnlHost >( int* destination, const int* source, const long int size );
-extern template bool tnlHost :: memcpy< long int, long int, tnlHost >( long int* destination, const long int* source, const long int size );
-extern template bool tnlHost :: memcpy< float, long int, tnlHost >( float* destination, const float* source, const long int size );
-extern template bool tnlHost :: memcpy< double, long int, tnlHost >( double* destination, const double* source, const long int size );
-extern template bool tnlHost :: memcpy< long double, long int, tnlHost >( long double* destination, const long double* source, const long int size );
-extern template bool tnlHost :: memcpy< char, int, tnlCuda >( char* destination, const char* source, const int size );
-extern template bool tnlHost :: memcpy< int, int, tnlCuda >( int* destination, const int* source, const int size );
-extern template bool tnlHost :: memcpy< long int, int, tnlCuda >( long int* destination, const long int* source, const int size );
-extern template bool tnlHost :: memcpy< float, int, tnlCuda >( float* destination, const float* source, const int size );
-extern template bool tnlHost :: memcpy< double, int, tnlCuda >( double* destination, const double* source, const int size );
-extern template bool tnlHost :: memcpy< long double, int, tnlCuda >( long double* destination, const long double* source, const int size );
-extern template bool tnlHost :: memcpy< char, long int, tnlCuda >( char* destination, const char* source, const long int size );
-extern template bool tnlHost :: memcpy< int, long int, tnlCuda >( int* destination, const int* source, const long int size );
-extern template bool tnlHost :: memcpy< long int, long int, tnlCuda >( long int* destination, const long int* source, const long int size );
-extern template bool tnlHost :: memcpy< float, long int, tnlCuda >( float* destination, const float* source, const long int size );
-extern template bool tnlHost :: memcpy< double, long int, tnlCuda >( double* destination, const double* source, const long int size );
-extern template bool tnlHost :: memcpy< long double, long int, tnlCuda >( long double* destination, const long double* source, const long int size );
+extern template bool tnlHost :: memcpy< char, char, int, tnlHost >( char* destination, const char* source, const int size );
+extern template bool tnlHost :: memcpy< int, int, int, tnlHost >( int* destination, const int* source, const int size );
+extern template bool tnlHost :: memcpy< long int, long int, int, tnlHost >( long int* destination, const long int* source, const int size );
+extern template bool tnlHost :: memcpy< float, float, int, tnlHost >( float* destination, const float* source, const int size );
+extern template bool tnlHost :: memcpy< double, double, int, tnlHost >( double* destination, const double* source, const int size );
+extern template bool tnlHost :: memcpy< long double, long double, int, tnlHost >( long double* destination, const long double* source, const int size );
+extern template bool tnlHost :: memcpy< char, char, long int, tnlHost >( char* destination, const char* source, const long int size );
+extern template bool tnlHost :: memcpy< int, int, long int, tnlHost >( int* destination, const int* source, const long int size );
+extern template bool tnlHost :: memcpy< long int, long int, long int, tnlHost >( long int* destination, const long int* source, const long int size );
+extern template bool tnlHost :: memcpy< float, float, long int, tnlHost >( float* destination, const float* source, const long int size );
+extern template bool tnlHost :: memcpy< double, double, long int, tnlHost >( double* destination, const double* source, const long int size );
+extern template bool tnlHost :: memcpy< long double, long double, long int, tnlHost >( long double* destination, const long double* source, const long int size );
+extern template bool tnlHost :: memcpy< char, char, int, tnlCuda >( char* destination, const char* source, const int size );
+extern template bool tnlHost :: memcpy< int, int, int, tnlCuda >( int* destination, const int* source, const int size );
+extern template bool tnlHost :: memcpy< long int, long int, int, tnlCuda >( long int* destination, const long int* source, const int size );
+extern template bool tnlHost :: memcpy< float, float, int, tnlCuda >( float* destination, const float* source, const int size );
+extern template bool tnlHost :: memcpy< double, double, int, tnlCuda >( double* destination, const double* source, const int size );
+extern template bool tnlHost :: memcpy< long double, long double, int, tnlCuda >( long double* destination, const long double* source, const int size );
+extern template bool tnlHost :: memcpy< char, char, long int, tnlCuda >( char* destination, const char* source, const long int size );
+extern template bool tnlHost :: memcpy< int, int, long int, tnlCuda >( int* destination, const int* source, const long int size );
+extern template bool tnlHost :: memcpy< long int, long int, long int, tnlCuda >( long int* destination, const long int* source, const long int size );
+extern template bool tnlHost :: memcpy< float, float, long int, tnlCuda >( float* destination, const float* source, const long int size );
+extern template bool tnlHost :: memcpy< double, double, long int, tnlCuda >( double* destination, const double* source, const long int size );
+extern template bool tnlHost :: memcpy< long double, long double, long int, tnlCuda >( long double* destination, const long double* source, const long int size );
extern template bool tnlHost :: memcmp< char, int, tnlHost >( const char* data1, const char* data2, const int size );
extern template bool tnlHost :: memcmp< int, int, tnlHost >( const int* data1, const int* data2, const int size );
--
GitLab