Commit ea079456 authored by Jakub Klinkovský's avatar Jakub Klinkovský
Browse files

Fixed cross-device overload of copySparseMatrix_impl

parent 15f4dfee

src/TNL/Matrices/SparseOperations_impl.h

+15 −2
Original line number Diff line number Diff line
@@ -161,10 +161,11 @@ copySparseMatrix_impl( Matrix1& A, const Matrix2& B ) 
   }

}



// cross-device copy

// cross-device copy (host -> gpu)

template< typename Matrix1,

          typename Matrix2 >

typename std::enable_if< ! std::is_same< typename Matrix1::DeviceType, typename Matrix2::DeviceType >::value >::type

typename std::enable_if< ! std::is_same< typename Matrix1::DeviceType, typename Matrix2::DeviceType >::value &&

                           std::is_same< typename Matrix2::DeviceType, Devices::Host >::value >::type

copySparseMatrix_impl( Matrix1& A, const Matrix2& B )

{

   typename Matrix2::CudaType B_tmp;
@@ -172,6 +173,18 @@ copySparseMatrix_impl( Matrix1& A, const Matrix2& B ) 
   copySparseMatrix_impl( A, B_tmp );

}



// cross-device copy (gpu -> host)

template< typename Matrix1,

          typename Matrix2 >

typename std::enable_if< ! std::is_same< typename Matrix1::DeviceType, typename Matrix2::DeviceType >::value &&

                           std::is_same< typename Matrix2::DeviceType, Devices::Cuda >::value >::type

copySparseMatrix_impl( Matrix1& A, const Matrix2& B )

{

   typename Matrix1::CudaType A_tmp;

   copySparseMatrix_impl( A_tmp, B );

   A = A_tmp;

}



template< typename Matrix1, typename Matrix2 >

void

copySparseMatrix( Matrix1& A, const Matrix2& B )