Loading src/TNL/DistributedContainers/DistributedMatrix.h 0 → 100644 +166 −0 Original line number Diff line number Diff line /*************************************************************************** DistributedMatrix.h - description ------------------- begin : Sep 10, 2018 copyright : (C) 2018 by Tomas Oberhuber et al. email : tomas.oberhuber@fjfi.cvut.cz ***************************************************************************/ /* See Copyright Notice in tnl/Copyright */ // Implemented by: Jakub Klinkovský #pragma once #include <type_traits> // std::add_const #include <TNL/Containers/Vector.h> #include <TNL/Matrices/SparseRow.h> #include <TNL/Communicators/MpiCommunicator.h> #include <TNL/DistributedContainers/IndexMap.h> #include <TNL/DistributedContainers/DistributedVector.h> namespace TNL { namespace DistributedContainers { // TODO: 2D distribution for dense matrices (maybe it should be in different template, // because e.g. setRowFast doesn't make sense for dense matrices) template< typename Matrix, typename Communicator = Communicators::MpiCommunicator, typename IndexMap = Subrange< typename Matrix::IndexType > > class DistributedMatrix : public Object { using CommunicationGroup = typename Communicator::CommunicationGroup; template< typename Real > using DistVector = DistributedVector< Real, typename Matrix::DeviceType, Communicator, typename Matrix::IndexType, IndexMap >; public: using MatrixType = Matrix; using RealType = typename Matrix::RealType; using DeviceType = typename Matrix::DeviceType; using IndexType = typename Matrix::IndexType; using CommunicatorType = Communicator; using IndexMapType = IndexMap; using HostType = DistributedMatrix< typename Matrix::HostType, Communicator, IndexMap >; using CudaType = DistributedMatrix< typename Matrix::CudaType, Communicator, IndexMap >; using CompressedRowLengthsVector = DistributedVector< IndexType, DeviceType, CommunicatorType, IndexType, IndexMapType >; using MatrixRow = Matrices::SparseRow< RealType, IndexType >; using ConstMatrixRow = Matrices::SparseRow< typename std::add_const< RealType >::type, typename std::add_const< IndexType >::type >; DistributedMatrix() = default; DistributedMatrix( DistributedMatrix& ) = default; DistributedMatrix( IndexMap rowIndexMap, IndexType columns, CommunicationGroup group = Communicator::AllGroup ); void setDistribution( IndexMap rowIndexMap, IndexType columns, CommunicationGroup group = Communicator::AllGroup ); const IndexMap& getRowIndexMap() const; CommunicationGroup getCommunicationGroup() const; const Matrix& getLocalMatrix() const; static String getType(); virtual String getTypeVirtual() const; // TODO: no getSerializationType method until there is support for serialization /* * Some common Matrix methods follow below. */ DistributedMatrix& operator=( const DistributedMatrix& matrix ); template< typename MatrixT > DistributedMatrix& operator=( const MatrixT& matrix ); template< typename MatrixT > void setLike( const MatrixT& matrix ); void reset(); __cuda_callable__ IndexType getRows() const; __cuda_callable__ IndexType getColumns() const; void setCompressedRowLengths( const CompressedRowLengthsVector& rowLengths ); void getCompressedRowLengths( CompressedRowLengthsVector& rowLengths ) const; IndexType getRowLength( IndexType row ) const; bool setElement( IndexType row, IndexType column, RealType value ); __cuda_callable__ bool setElementFast( IndexType row, IndexType column, RealType value ); RealType getElement( IndexType row, IndexType column ) const; __cuda_callable__ RealType getElementFast( IndexType row, IndexType column ) const; __cuda_callable__ bool setRowFast( IndexType row, const IndexType* columnIndexes, const RealType* values, IndexType elements ); __cuda_callable__ void getRowFast( IndexType row, IndexType* columns, RealType* values ) const; __cuda_callable__ MatrixRow getRow( IndexType row ); __cuda_callable__ ConstMatrixRow getRow( IndexType row ) const; // multiplication with a global vector template< typename Vector, typename RealOut > void vectorProduct( const Vector& inVector, DistVector< RealOut >& outVector ) const; // optimization for matrix-vector multiplication void updateVectorProductPrefetchPattern(); // multiplication with a distributed vector template< typename RealIn, typename RealOut > void vectorProduct( const DistVector< RealIn >& inVector, DistVector< RealOut >& outVector ) const; protected: IndexMap rowIndexMap; CommunicationGroup group = Communicator::NullGroup; Matrix localMatrix; private: // TODO: disabled until they are implemented using Object::save; using Object::load; using Object::boundLoad; }; } // namespace DistributedContainers } // namespace TNL #include "DistributedMatrix_impl.h" src/TNL/DistributedContainers/DistributedMatrix_impl.h 0 → 100644 +338 −0 Original line number Diff line number Diff line /*************************************************************************** DistributedMatrix.h - description ------------------- begin : Sep 10, 2018 copyright : (C) 2018 by Tomas Oberhuber et al. email : tomas.oberhuber@fjfi.cvut.cz ***************************************************************************/ /* See Copyright Notice in tnl/Copyright */ // Implemented by: Jakub Klinkovský #pragma once #include "DistributedMatrix.h" namespace TNL { namespace DistributedContainers { template< typename Matrix, typename Communicator, typename IndexMap > DistributedMatrix< Matrix, Communicator, IndexMap >:: DistributedMatrix( IndexMap rowIndexMap, IndexType columns, CommunicationGroup group ) { setDistribution( rowIndexMap, columns, group ); } template< typename Matrix, typename Communicator, typename IndexMap > void DistributedMatrix< Matrix, Communicator, IndexMap >:: setDistribution( IndexMap rowIndexMap, IndexType columns, CommunicationGroup group ) { this->rowIndexMap = rowIndexMap; this->group = group; if( group != Communicator::NullGroup ) localMatrix.setDimensions( rowIndexMap.getLocalSize(), columns ); } template< typename Matrix, typename Communicator, typename IndexMap > const IndexMap& DistributedMatrix< Matrix, Communicator, IndexMap >:: getRowIndexMap() const { return rowIndexMap; } template< typename Matrix, typename Communicator, typename IndexMap > typename Communicator::CommunicationGroup DistributedMatrix< Matrix, Communicator, IndexMap >:: getCommunicationGroup() const { return group; } template< typename Matrix, typename Communicator, typename IndexMap > const Matrix& DistributedMatrix< Matrix, Communicator, IndexMap >:: getLocalMatrix() const { return localMatrix; } template< typename Matrix, typename Communicator, typename IndexMap > String DistributedMatrix< Matrix, Communicator, IndexMap >:: getType() { return String( "DistributedContainers::DistributedMatrix< " ) + Matrix::getType() + ", " + // TODO: communicators don't have a getType method "<Communicator>, " + IndexMap::getType() + " >"; } template< typename Matrix, typename Communicator, typename IndexMap > String DistributedMatrix< Matrix, Communicator, IndexMap >:: getTypeVirtual() const { return getType(); } /* * Some common Matrix methods follow below. */ template< typename Matrix, typename Communicator, typename IndexMap > DistributedMatrix< Matrix, Communicator, IndexMap >& DistributedMatrix< Matrix, Communicator, IndexMap >:: operator=( const DistributedMatrix& matrix ) { setLike( matrix ); localMatrix = matrix.getLocalMatrix(); return *this; } template< typename Matrix, typename Communicator, typename IndexMap > template< typename MatrixT > DistributedMatrix< Matrix, Communicator, IndexMap >& DistributedMatrix< Matrix, Communicator, IndexMap >:: operator=( const MatrixT& matrix ) { setLike( matrix ); localMatrix = matrix.getLocalMatrix(); return *this; } template< typename Matrix, typename Communicator, typename IndexMap > template< typename MatrixT > void DistributedMatrix< Matrix, Communicator, IndexMap >:: setLike( const MatrixT& matrix ) { rowIndexMap = matrix.getRowIndexMap(); group = matrix.getCommunicationGroup(); localMatrix.setLike( matrix.getLocalMatrix() ); } template< typename Matrix, typename Communicator, typename IndexMap > void DistributedMatrix< Matrix, Communicator, IndexMap >:: reset() { rowIndexMap.reset(); group = Communicator::NullGroup; localMatrix.reset(); } template< typename Matrix, typename Communicator, typename IndexMap > __cuda_callable__ typename Matrix::IndexType DistributedMatrix< Matrix, Communicator, IndexMap >:: getRows() const { return rowIndexMap.getGlobalSize(); } template< typename Matrix, typename Communicator, typename IndexMap > __cuda_callable__ typename Matrix::IndexType DistributedMatrix< Matrix, Communicator, IndexMap >:: getColumns() const { return localMatrix.getColumns(); } template< typename Matrix, typename Communicator, typename IndexMap > void DistributedMatrix< Matrix, Communicator, IndexMap >:: setCompressedRowLengths( const CompressedRowLengthsVector& rowLengths ) { TNL_ASSERT_EQ( rowLengths.getSize(), getRows(), "row lengths vector has wrong size" ); TNL_ASSERT_EQ( rowLengths.getIndexMap(), getRowIndexMap(), "row lengths vector has wrong distribution" ); TNL_ASSERT_EQ( rowLengths.getCommunicationGroup(), getCommunicationGroup(), "row lengths vector has wrong communication group" ); if( getCommunicationGroup() != CommunicatorType::NullGroup ) localMatrix.setCompressedRowLengths( rowLengths.getLocalVectorView() ); } template< typename Matrix, typename Communicator, typename IndexMap > void DistributedMatrix< Matrix, Communicator, IndexMap >:: getCompressedRowLengths( CompressedRowLengthsVector& rowLengths ) const { if( getCommunicationGroup() != CommunicatorType::NullGroup ) { rowLengths.setDistribution( getRowIndexMap(), getCommunicationGroup() ); localMatrix.getCompressedRowLengths( rowLengths.getLocalVectorView() ); } } template< typename Matrix, typename Communicator, typename IndexMap > typename Matrix::IndexType DistributedMatrix< Matrix, Communicator, IndexMap >:: getRowLength( IndexType row ) const { const IndexType localRow = rowIndexMap.getLocalIndex( row ); return localMatrix.getRowLength( localRow ); } template< typename Matrix, typename Communicator, typename IndexMap > bool DistributedMatrix< Matrix, Communicator, IndexMap >:: setElement( IndexType row, IndexType column, RealType value ) { const IndexType localRow = rowIndexMap.getLocalIndex( row ); return localMatrix.setElement( localRow, column, value ); } template< typename Matrix, typename Communicator, typename IndexMap > __cuda_callable__ bool DistributedMatrix< Matrix, Communicator, IndexMap >:: setElementFast( IndexType row, IndexType column, RealType value ) { const IndexType localRow = rowIndexMap.getLocalIndex( row ); return localMatrix.setElementFast( localRow, column, value ); } template< typename Matrix, typename Communicator, typename IndexMap > typename Matrix::RealType DistributedMatrix< Matrix, Communicator, IndexMap >:: getElement( IndexType row, IndexType column ) const { const IndexType localRow = rowIndexMap.getLocalIndex( row ); return localMatrix.getElement( localRow, column ); } template< typename Matrix, typename Communicator, typename IndexMap > __cuda_callable__ typename Matrix::RealType DistributedMatrix< Matrix, Communicator, IndexMap >:: getElementFast( IndexType row, IndexType column ) const { const IndexType localRow = rowIndexMap.getLocalIndex( row ); return localMatrix.getElementFast( localRow, column ); } template< typename Matrix, typename Communicator, typename IndexMap > __cuda_callable__ bool DistributedMatrix< Matrix, Communicator, IndexMap >:: setRowFast( IndexType row, const IndexType* columnIndexes, const RealType* values, IndexType elements ) { const IndexType localRow = rowIndexMap.getLocalIndex( row ); return localMatrix.setRowFast( localRow, columnIndexes, values, elements ); } template< typename Matrix, typename Communicator, typename IndexMap > __cuda_callable__ void DistributedMatrix< Matrix, Communicator, IndexMap >:: getRowFast( IndexType row, IndexType* columns, RealType* values ) const { const IndexType localRow = rowIndexMap.getLocalIndex( row ); return localMatrix.getRowFast( localRow, columns, values ); } template< typename Matrix, typename Communicator, typename IndexMap > __cuda_callable__ typename DistributedMatrix< Matrix, Communicator, IndexMap >::MatrixRow DistributedMatrix< Matrix, Communicator, IndexMap >:: getRow( IndexType row ) { const IndexType localRow = rowIndexMap.getLocalIndex( row ); return localMatrix.getRow( localRow ); } template< typename Matrix, typename Communicator, typename IndexMap > __cuda_callable__ typename DistributedMatrix< Matrix, Communicator, IndexMap >::ConstMatrixRow DistributedMatrix< Matrix, Communicator, IndexMap >:: getRow( IndexType row ) const { const IndexType localRow = rowIndexMap.getLocalIndex( row ); return localMatrix.getRow( localRow ); } template< typename Matrix, typename Communicator, typename IndexMap > template< typename Vector, typename RealOut > void DistributedMatrix< Matrix, Communicator, IndexMap >:: vectorProduct( const Vector& inVector, DistVector< RealOut >& outVector ) const { TNL_ASSERT_EQ( inVector.getSize(), getColumns(), "input vector has wrong size" ); TNL_ASSERT_EQ( outVector.getSize(), getRows(), "output vector has wrong size" ); TNL_ASSERT_EQ( outVector.getIndexMap(), getRowIndexMap(), "output vector has wrong distribution" ); TNL_ASSERT_EQ( outVector.getCommunicationGroup(), getCommunicationGroup(), "output vector has wrong communication group" ); auto outView = outVector.getLocalVectorView(); localMatrix.vectorProduct( inVector, outView ); } } // namespace DistributedContainers } // namespace TNL src/UnitTests/DistributedContainers/CMakeLists.txt +15 −0 Original line number Diff line number Diff line Loading @@ -12,6 +12,12 @@ if( BUILD_CUDA ) TARGET_LINK_LIBRARIES( DistributedVectorTest ${GTEST_BOTH_LIBRARIES} tnl ) CUDA_ADD_EXECUTABLE( DistributedMatrixTest DistributedMatrixTest.cu OPTIONS ${CXX_TESTS_FLAGS} ) TARGET_LINK_LIBRARIES( DistributedMatrixTest ${GTEST_BOTH_LIBRARIES} tnl ) else() ADD_EXECUTABLE( DistributedArrayTest DistributedArrayTest.cpp ) TARGET_COMPILE_OPTIONS( DistributedArrayTest PRIVATE ${CXX_TESTS_FLAGS} ) Loading @@ -24,6 +30,12 @@ else() TARGET_LINK_LIBRARIES( DistributedVectorTest ${GTEST_BOTH_LIBRARIES} tnl ) ADD_EXECUTABLE( DistributedMatrixTest DistributedMatrixTest.cpp ) TARGET_COMPILE_OPTIONS( DistributedMatrixTest PRIVATE ${CXX_TESTS_FLAGS} ) TARGET_LINK_LIBRARIES( DistributedMatrixTest ${GTEST_BOTH_LIBRARIES} tnl ) endif() SET( mpi_test_parameters -np 4 -H localhost:4 "${EXECUTABLE_OUTPUT_PATH}/DistributedArrayTest${CMAKE_EXECUTABLE_SUFFIX}" ) Loading @@ -32,4 +44,7 @@ ADD_TEST( NAME DistributedArrayTest COMMAND "mpirun" ${mpi_test_parameters}) SET( mpi_test_parameters -np 4 -H localhost:4 "${EXECUTABLE_OUTPUT_PATH}/DistributedVectorTest${CMAKE_EXECUTABLE_SUFFIX}" ) ADD_TEST( NAME DistributedVectorTest COMMAND "mpirun" ${mpi_test_parameters}) SET( mpi_test_parameters -np 4 -H localhost:4 "${EXECUTABLE_OUTPUT_PATH}/DistributedMatrixTest${CMAKE_EXECUTABLE_SUFFIX}" ) ADD_TEST( NAME DistributedMatrixTest COMMAND "mpirun" ${mpi_test_parameters}) endif() src/UnitTests/DistributedContainers/DistributedMatrixTest.cpp 0 → 100644 +1 −0 Original line number Diff line number Diff line #include "DistributedMatrixTest.h" src/UnitTests/DistributedContainers/DistributedMatrixTest.cu 0 → 100644 +1 −0 Original line number Diff line number Diff line #include "DistributedMatrixTest.h" Loading
src/TNL/DistributedContainers/DistributedMatrix.h 0 → 100644 +166 −0 Original line number Diff line number Diff line /*************************************************************************** DistributedMatrix.h - description ------------------- begin : Sep 10, 2018 copyright : (C) 2018 by Tomas Oberhuber et al. email : tomas.oberhuber@fjfi.cvut.cz ***************************************************************************/ /* See Copyright Notice in tnl/Copyright */ // Implemented by: Jakub Klinkovský #pragma once #include <type_traits> // std::add_const #include <TNL/Containers/Vector.h> #include <TNL/Matrices/SparseRow.h> #include <TNL/Communicators/MpiCommunicator.h> #include <TNL/DistributedContainers/IndexMap.h> #include <TNL/DistributedContainers/DistributedVector.h> namespace TNL { namespace DistributedContainers { // TODO: 2D distribution for dense matrices (maybe it should be in different template, // because e.g. setRowFast doesn't make sense for dense matrices) template< typename Matrix, typename Communicator = Communicators::MpiCommunicator, typename IndexMap = Subrange< typename Matrix::IndexType > > class DistributedMatrix : public Object { using CommunicationGroup = typename Communicator::CommunicationGroup; template< typename Real > using DistVector = DistributedVector< Real, typename Matrix::DeviceType, Communicator, typename Matrix::IndexType, IndexMap >; public: using MatrixType = Matrix; using RealType = typename Matrix::RealType; using DeviceType = typename Matrix::DeviceType; using IndexType = typename Matrix::IndexType; using CommunicatorType = Communicator; using IndexMapType = IndexMap; using HostType = DistributedMatrix< typename Matrix::HostType, Communicator, IndexMap >; using CudaType = DistributedMatrix< typename Matrix::CudaType, Communicator, IndexMap >; using CompressedRowLengthsVector = DistributedVector< IndexType, DeviceType, CommunicatorType, IndexType, IndexMapType >; using MatrixRow = Matrices::SparseRow< RealType, IndexType >; using ConstMatrixRow = Matrices::SparseRow< typename std::add_const< RealType >::type, typename std::add_const< IndexType >::type >; DistributedMatrix() = default; DistributedMatrix( DistributedMatrix& ) = default; DistributedMatrix( IndexMap rowIndexMap, IndexType columns, CommunicationGroup group = Communicator::AllGroup ); void setDistribution( IndexMap rowIndexMap, IndexType columns, CommunicationGroup group = Communicator::AllGroup ); const IndexMap& getRowIndexMap() const; CommunicationGroup getCommunicationGroup() const; const Matrix& getLocalMatrix() const; static String getType(); virtual String getTypeVirtual() const; // TODO: no getSerializationType method until there is support for serialization /* * Some common Matrix methods follow below. */ DistributedMatrix& operator=( const DistributedMatrix& matrix ); template< typename MatrixT > DistributedMatrix& operator=( const MatrixT& matrix ); template< typename MatrixT > void setLike( const MatrixT& matrix ); void reset(); __cuda_callable__ IndexType getRows() const; __cuda_callable__ IndexType getColumns() const; void setCompressedRowLengths( const CompressedRowLengthsVector& rowLengths ); void getCompressedRowLengths( CompressedRowLengthsVector& rowLengths ) const; IndexType getRowLength( IndexType row ) const; bool setElement( IndexType row, IndexType column, RealType value ); __cuda_callable__ bool setElementFast( IndexType row, IndexType column, RealType value ); RealType getElement( IndexType row, IndexType column ) const; __cuda_callable__ RealType getElementFast( IndexType row, IndexType column ) const; __cuda_callable__ bool setRowFast( IndexType row, const IndexType* columnIndexes, const RealType* values, IndexType elements ); __cuda_callable__ void getRowFast( IndexType row, IndexType* columns, RealType* values ) const; __cuda_callable__ MatrixRow getRow( IndexType row ); __cuda_callable__ ConstMatrixRow getRow( IndexType row ) const; // multiplication with a global vector template< typename Vector, typename RealOut > void vectorProduct( const Vector& inVector, DistVector< RealOut >& outVector ) const; // optimization for matrix-vector multiplication void updateVectorProductPrefetchPattern(); // multiplication with a distributed vector template< typename RealIn, typename RealOut > void vectorProduct( const DistVector< RealIn >& inVector, DistVector< RealOut >& outVector ) const; protected: IndexMap rowIndexMap; CommunicationGroup group = Communicator::NullGroup; Matrix localMatrix; private: // TODO: disabled until they are implemented using Object::save; using Object::load; using Object::boundLoad; }; } // namespace DistributedContainers } // namespace TNL #include "DistributedMatrix_impl.h"
src/TNL/DistributedContainers/DistributedMatrix_impl.h 0 → 100644 +338 −0 Original line number Diff line number Diff line /*************************************************************************** DistributedMatrix.h - description ------------------- begin : Sep 10, 2018 copyright : (C) 2018 by Tomas Oberhuber et al. email : tomas.oberhuber@fjfi.cvut.cz ***************************************************************************/ /* See Copyright Notice in tnl/Copyright */ // Implemented by: Jakub Klinkovský #pragma once #include "DistributedMatrix.h" namespace TNL { namespace DistributedContainers { template< typename Matrix, typename Communicator, typename IndexMap > DistributedMatrix< Matrix, Communicator, IndexMap >:: DistributedMatrix( IndexMap rowIndexMap, IndexType columns, CommunicationGroup group ) { setDistribution( rowIndexMap, columns, group ); } template< typename Matrix, typename Communicator, typename IndexMap > void DistributedMatrix< Matrix, Communicator, IndexMap >:: setDistribution( IndexMap rowIndexMap, IndexType columns, CommunicationGroup group ) { this->rowIndexMap = rowIndexMap; this->group = group; if( group != Communicator::NullGroup ) localMatrix.setDimensions( rowIndexMap.getLocalSize(), columns ); } template< typename Matrix, typename Communicator, typename IndexMap > const IndexMap& DistributedMatrix< Matrix, Communicator, IndexMap >:: getRowIndexMap() const { return rowIndexMap; } template< typename Matrix, typename Communicator, typename IndexMap > typename Communicator::CommunicationGroup DistributedMatrix< Matrix, Communicator, IndexMap >:: getCommunicationGroup() const { return group; } template< typename Matrix, typename Communicator, typename IndexMap > const Matrix& DistributedMatrix< Matrix, Communicator, IndexMap >:: getLocalMatrix() const { return localMatrix; } template< typename Matrix, typename Communicator, typename IndexMap > String DistributedMatrix< Matrix, Communicator, IndexMap >:: getType() { return String( "DistributedContainers::DistributedMatrix< " ) + Matrix::getType() + ", " + // TODO: communicators don't have a getType method "<Communicator>, " + IndexMap::getType() + " >"; } template< typename Matrix, typename Communicator, typename IndexMap > String DistributedMatrix< Matrix, Communicator, IndexMap >:: getTypeVirtual() const { return getType(); } /* * Some common Matrix methods follow below. */ template< typename Matrix, typename Communicator, typename IndexMap > DistributedMatrix< Matrix, Communicator, IndexMap >& DistributedMatrix< Matrix, Communicator, IndexMap >:: operator=( const DistributedMatrix& matrix ) { setLike( matrix ); localMatrix = matrix.getLocalMatrix(); return *this; } template< typename Matrix, typename Communicator, typename IndexMap > template< typename MatrixT > DistributedMatrix< Matrix, Communicator, IndexMap >& DistributedMatrix< Matrix, Communicator, IndexMap >:: operator=( const MatrixT& matrix ) { setLike( matrix ); localMatrix = matrix.getLocalMatrix(); return *this; } template< typename Matrix, typename Communicator, typename IndexMap > template< typename MatrixT > void DistributedMatrix< Matrix, Communicator, IndexMap >:: setLike( const MatrixT& matrix ) { rowIndexMap = matrix.getRowIndexMap(); group = matrix.getCommunicationGroup(); localMatrix.setLike( matrix.getLocalMatrix() ); } template< typename Matrix, typename Communicator, typename IndexMap > void DistributedMatrix< Matrix, Communicator, IndexMap >:: reset() { rowIndexMap.reset(); group = Communicator::NullGroup; localMatrix.reset(); } template< typename Matrix, typename Communicator, typename IndexMap > __cuda_callable__ typename Matrix::IndexType DistributedMatrix< Matrix, Communicator, IndexMap >:: getRows() const { return rowIndexMap.getGlobalSize(); } template< typename Matrix, typename Communicator, typename IndexMap > __cuda_callable__ typename Matrix::IndexType DistributedMatrix< Matrix, Communicator, IndexMap >:: getColumns() const { return localMatrix.getColumns(); } template< typename Matrix, typename Communicator, typename IndexMap > void DistributedMatrix< Matrix, Communicator, IndexMap >:: setCompressedRowLengths( const CompressedRowLengthsVector& rowLengths ) { TNL_ASSERT_EQ( rowLengths.getSize(), getRows(), "row lengths vector has wrong size" ); TNL_ASSERT_EQ( rowLengths.getIndexMap(), getRowIndexMap(), "row lengths vector has wrong distribution" ); TNL_ASSERT_EQ( rowLengths.getCommunicationGroup(), getCommunicationGroup(), "row lengths vector has wrong communication group" ); if( getCommunicationGroup() != CommunicatorType::NullGroup ) localMatrix.setCompressedRowLengths( rowLengths.getLocalVectorView() ); } template< typename Matrix, typename Communicator, typename IndexMap > void DistributedMatrix< Matrix, Communicator, IndexMap >:: getCompressedRowLengths( CompressedRowLengthsVector& rowLengths ) const { if( getCommunicationGroup() != CommunicatorType::NullGroup ) { rowLengths.setDistribution( getRowIndexMap(), getCommunicationGroup() ); localMatrix.getCompressedRowLengths( rowLengths.getLocalVectorView() ); } } template< typename Matrix, typename Communicator, typename IndexMap > typename Matrix::IndexType DistributedMatrix< Matrix, Communicator, IndexMap >:: getRowLength( IndexType row ) const { const IndexType localRow = rowIndexMap.getLocalIndex( row ); return localMatrix.getRowLength( localRow ); } template< typename Matrix, typename Communicator, typename IndexMap > bool DistributedMatrix< Matrix, Communicator, IndexMap >:: setElement( IndexType row, IndexType column, RealType value ) { const IndexType localRow = rowIndexMap.getLocalIndex( row ); return localMatrix.setElement( localRow, column, value ); } template< typename Matrix, typename Communicator, typename IndexMap > __cuda_callable__ bool DistributedMatrix< Matrix, Communicator, IndexMap >:: setElementFast( IndexType row, IndexType column, RealType value ) { const IndexType localRow = rowIndexMap.getLocalIndex( row ); return localMatrix.setElementFast( localRow, column, value ); } template< typename Matrix, typename Communicator, typename IndexMap > typename Matrix::RealType DistributedMatrix< Matrix, Communicator, IndexMap >:: getElement( IndexType row, IndexType column ) const { const IndexType localRow = rowIndexMap.getLocalIndex( row ); return localMatrix.getElement( localRow, column ); } template< typename Matrix, typename Communicator, typename IndexMap > __cuda_callable__ typename Matrix::RealType DistributedMatrix< Matrix, Communicator, IndexMap >:: getElementFast( IndexType row, IndexType column ) const { const IndexType localRow = rowIndexMap.getLocalIndex( row ); return localMatrix.getElementFast( localRow, column ); } template< typename Matrix, typename Communicator, typename IndexMap > __cuda_callable__ bool DistributedMatrix< Matrix, Communicator, IndexMap >:: setRowFast( IndexType row, const IndexType* columnIndexes, const RealType* values, IndexType elements ) { const IndexType localRow = rowIndexMap.getLocalIndex( row ); return localMatrix.setRowFast( localRow, columnIndexes, values, elements ); } template< typename Matrix, typename Communicator, typename IndexMap > __cuda_callable__ void DistributedMatrix< Matrix, Communicator, IndexMap >:: getRowFast( IndexType row, IndexType* columns, RealType* values ) const { const IndexType localRow = rowIndexMap.getLocalIndex( row ); return localMatrix.getRowFast( localRow, columns, values ); } template< typename Matrix, typename Communicator, typename IndexMap > __cuda_callable__ typename DistributedMatrix< Matrix, Communicator, IndexMap >::MatrixRow DistributedMatrix< Matrix, Communicator, IndexMap >:: getRow( IndexType row ) { const IndexType localRow = rowIndexMap.getLocalIndex( row ); return localMatrix.getRow( localRow ); } template< typename Matrix, typename Communicator, typename IndexMap > __cuda_callable__ typename DistributedMatrix< Matrix, Communicator, IndexMap >::ConstMatrixRow DistributedMatrix< Matrix, Communicator, IndexMap >:: getRow( IndexType row ) const { const IndexType localRow = rowIndexMap.getLocalIndex( row ); return localMatrix.getRow( localRow ); } template< typename Matrix, typename Communicator, typename IndexMap > template< typename Vector, typename RealOut > void DistributedMatrix< Matrix, Communicator, IndexMap >:: vectorProduct( const Vector& inVector, DistVector< RealOut >& outVector ) const { TNL_ASSERT_EQ( inVector.getSize(), getColumns(), "input vector has wrong size" ); TNL_ASSERT_EQ( outVector.getSize(), getRows(), "output vector has wrong size" ); TNL_ASSERT_EQ( outVector.getIndexMap(), getRowIndexMap(), "output vector has wrong distribution" ); TNL_ASSERT_EQ( outVector.getCommunicationGroup(), getCommunicationGroup(), "output vector has wrong communication group" ); auto outView = outVector.getLocalVectorView(); localMatrix.vectorProduct( inVector, outView ); } } // namespace DistributedContainers } // namespace TNL
src/UnitTests/DistributedContainers/CMakeLists.txt +15 −0 Original line number Diff line number Diff line Loading @@ -12,6 +12,12 @@ if( BUILD_CUDA ) TARGET_LINK_LIBRARIES( DistributedVectorTest ${GTEST_BOTH_LIBRARIES} tnl ) CUDA_ADD_EXECUTABLE( DistributedMatrixTest DistributedMatrixTest.cu OPTIONS ${CXX_TESTS_FLAGS} ) TARGET_LINK_LIBRARIES( DistributedMatrixTest ${GTEST_BOTH_LIBRARIES} tnl ) else() ADD_EXECUTABLE( DistributedArrayTest DistributedArrayTest.cpp ) TARGET_COMPILE_OPTIONS( DistributedArrayTest PRIVATE ${CXX_TESTS_FLAGS} ) Loading @@ -24,6 +30,12 @@ else() TARGET_LINK_LIBRARIES( DistributedVectorTest ${GTEST_BOTH_LIBRARIES} tnl ) ADD_EXECUTABLE( DistributedMatrixTest DistributedMatrixTest.cpp ) TARGET_COMPILE_OPTIONS( DistributedMatrixTest PRIVATE ${CXX_TESTS_FLAGS} ) TARGET_LINK_LIBRARIES( DistributedMatrixTest ${GTEST_BOTH_LIBRARIES} tnl ) endif() SET( mpi_test_parameters -np 4 -H localhost:4 "${EXECUTABLE_OUTPUT_PATH}/DistributedArrayTest${CMAKE_EXECUTABLE_SUFFIX}" ) Loading @@ -32,4 +44,7 @@ ADD_TEST( NAME DistributedArrayTest COMMAND "mpirun" ${mpi_test_parameters}) SET( mpi_test_parameters -np 4 -H localhost:4 "${EXECUTABLE_OUTPUT_PATH}/DistributedVectorTest${CMAKE_EXECUTABLE_SUFFIX}" ) ADD_TEST( NAME DistributedVectorTest COMMAND "mpirun" ${mpi_test_parameters}) SET( mpi_test_parameters -np 4 -H localhost:4 "${EXECUTABLE_OUTPUT_PATH}/DistributedMatrixTest${CMAKE_EXECUTABLE_SUFFIX}" ) ADD_TEST( NAME DistributedMatrixTest COMMAND "mpirun" ${mpi_test_parameters}) endif()
src/UnitTests/DistributedContainers/DistributedMatrixTest.cpp 0 → 100644 +1 −0 Original line number Diff line number Diff line #include "DistributedMatrixTest.h"
src/UnitTests/DistributedContainers/DistributedMatrixTest.cu 0 → 100644 +1 −0 Original line number Diff line number Diff line #include "DistributedMatrixTest.h"
