dftfe::linearAlgebraOperations::internal Namespace Reference

Contains internal functions used in linearAlgebraOperations. More...

Functions
void	setupELPAHandleParameters (const MPI_Comm &mpi_communicator, MPI_Comm &processGridCommunicatorActive, const std::shared_ptr< const dftfe::ProcessGrid > &processGrid, const unsigned int na, const unsigned int nev, const unsigned int blockSize, elpa_t &elpaHandle, const dftParameters &dftParams)
	setup ELPA parameters.
void	createProcessGridSquareMatrix (const MPI_Comm &mpi_communicator, const unsigned size, std::shared_ptr< const dftfe::ProcessGrid > &processGrid, const dftParameters &dftParams, const bool useOnlyThumbRule=false)
	Wrapper function to create a two dimensional processor grid for a square matrix in dftfe::ScaLAPACKMatrix storage format.
void	createProcessGridRectangularMatrix (const MPI_Comm &mpi_communicator, const unsigned sizeRows, const unsigned sizeColumns, std::shared_ptr< const dftfe::ProcessGrid > &processGrid, const dftParameters &dftParams)
	Wrapper function to create a two dimensional processor grid for a rectangular matrix in dftfe::ScaLAPACKMatrix storage format.
template<typename T>
void	createGlobalToLocalIdMapsScaLAPACKMat (const std::shared_ptr< const dftfe::ProcessGrid > &processGrid, const dftfe::ScaLAPACKMatrix< T > &mat, std::unordered_map< unsigned int, unsigned int > &globalToLocalRowIdMap, std::unordered_map< unsigned int, unsigned int > &globalToLocalColumnIdMap)
	Creates global row/column id to local row/column ids for dftfe::ScaLAPACKMatrix.
template<typename T>
void	sumAcrossInterCommScaLAPACKMat (const std::shared_ptr< const dftfe::ProcessGrid > &processGrid, dftfe::ScaLAPACKMatrix< T > &mat, const MPI_Comm &interComm)
	Mpi all reduce of ScaLAPACKMat across a given inter communicator. Used for band parallelization.
template<typename T>
void	scaleScaLAPACKMat (const std::shared_ptr< const dftfe::ProcessGrid > &processGrid, const std::shared_ptr< dftfe::linearAlgebra::BLASWrapper< dftfe::utils::MemorySpace::HOST > > &BLASWrapperPtr, dftfe::ScaLAPACKMatrix< T > &mat, const T scalar)
	scale a ScaLAPACKMat with a scalar
template<typename T>
void	broadcastAcrossInterCommScaLAPACKMat (const std::shared_ptr< const dftfe::ProcessGrid > &processGrid, dftfe::ScaLAPACKMatrix< T > &mat, const MPI_Comm &interComm, const unsigned int broadcastRoot)
	MPI_Bcast of ScaLAPACKMat across a given inter communicator from a given broadcast root. Used for band parallelization.
template<typename T>
void	fillParallelOverlapMatrix (const T *X, const std::shared_ptr< dftfe::linearAlgebra::BLASWrapper< dftfe::utils::MemorySpace::HOST > > &BLASWrapperPtr, const unsigned int XLocalSize, const unsigned int numberVectors, const std::shared_ptr< const dftfe::ProcessGrid > &processGrid, const MPI_Comm &interBandGroupComm, const MPI_Comm &mpiComm, dftfe::ScaLAPACKMatrix< T > &overlapMatPar, const dftParameters &dftParams)
	Computes Sc=X^{T}*Xc and stores in a parallel ScaLAPACK matrix. X^{T} is the subspaceVectorsArray stored in the column major format (N x M). Sc is the overlapMatPar.
template<typename T, typename TLowPrec>
void	fillParallelOverlapMatrixMixedPrec (const T *X, const std::shared_ptr< dftfe::linearAlgebra::BLASWrapper< dftfe::utils::MemorySpace::HOST > > &BLASWrapperPtr, const unsigned int XLocalSize, const unsigned int numberVectors, const std::shared_ptr< const dftfe::ProcessGrid > &processGrid, const MPI_Comm &interBandGroupComm, const MPI_Comm &mpiComm, dftfe::ScaLAPACKMatrix< T > &overlapMatPar, const dftParameters &dftParams)
	Computes Sc=X^{T}*Xc and stores in a parallel ScaLAPACK matrix. X^{T} is the subspaceVectorsArray stored in the column major format (N x M). Sc is the overlapMatPar.
template<typename T>
void	subspaceRotation (T *subspaceVectorsArray, const std::shared_ptr< dftfe::linearAlgebra::BLASWrapper< dftfe::utils::MemorySpace::HOST > > &BLASWrapperPtr, const unsigned int subspaceVectorsArrayLocalSize, const unsigned int N, const std::shared_ptr< const dftfe::ProcessGrid > &processGrid, const MPI_Comm &interBandGroupComm, const MPI_Comm &mpiComm, const dftfe::ScaLAPACKMatrix< T > &rotationMatPar, const dftParameters &dftParams, const bool rotationMatTranspose=false, const bool isRotationMatLowerTria=false, const bool doCommAfterBandParal=true)
	Computes X^{T}=Q*X^{T} inplace. X^{T} is the subspaceVectorsArray stored in the column major format (N x M). Q is rotationMatPar (N x N).
template<typename T, typename TLowPrec>
void	subspaceRotationMixedPrec (T *subspaceVectorsArray, const std::shared_ptr< dftfe::linearAlgebra::BLASWrapper< dftfe::utils::MemorySpace::HOST > > &BLASWrapperPtr, const unsigned int subspaceVectorsArrayLocalSize, const unsigned int N, const std::shared_ptr< const dftfe::ProcessGrid > &processGrid, const MPI_Comm &interBandGroupComm, const MPI_Comm &mpiComm, const dftfe::ScaLAPACKMatrix< T > &rotationMatPar, const dftParameters &dftParams, const bool rotationMatTranspose=false, const bool doCommAfterBandParal=true)
	Computes X^{T}=Q*X^{T} inplace. X^{T} is the subspaceVectorsArray stored in the column major format (N x M). Q is rotationMatPar (N x N).
template<typename T>
void	subspaceRotationSpectrumSplit (const T *X, const std::shared_ptr< dftfe::linearAlgebra::BLASWrapper< dftfe::utils::MemorySpace::HOST > > &BLASWrapperPtr, T *Y, const unsigned int subspaceVectorsArrayLocalSize, const unsigned int N, const std::shared_ptr< const dftfe::ProcessGrid > &processGrid, const unsigned int numberTopVectors, const MPI_Comm &interBandGroupComm, const MPI_Comm &mpiComm, const dftfe::ScaLAPACKMatrix< T > &QMat, const dftParameters &dftParams, const bool QMatTranspose=false)
	Computes Y^{T}=Q*X^{T}.
template<typename T, typename TLowPrec>
void	subspaceRotationSpectrumSplitMixedPrec (const T *X, const std::shared_ptr< dftfe::linearAlgebra::BLASWrapper< dftfe::utils::MemorySpace::HOST > > &BLASWrapperPtr, T *Y, const unsigned int subspaceVectorsArrayLocalSize, const unsigned int N, const std::shared_ptr< const dftfe::ProcessGrid > &processGrid, const unsigned int numberTopVectors, const MPI_Comm &interBandGroupComm, const MPI_Comm &mpiComm, const dftfe::ScaLAPACKMatrix< T > &QMat, const dftParameters &dftParams, const bool QMatTranspose=false)
	Computes Y^{T}=Q*X^{T}.
template<typename T, typename TLowPrec>
void	subspaceRotationCGSMixedPrec (T *subspaceVectorsArray, const std::shared_ptr< dftfe::linearAlgebra::BLASWrapper< dftfe::utils::MemorySpace::HOST > > &BLASWrapperPtr, const unsigned int subspaceVectorsArrayLocalSize, const unsigned int N, const std::shared_ptr< const dftfe::ProcessGrid > &processGrid, const MPI_Comm &interBandGroupComm, const MPI_Comm &mpiComm, const dftfe::ScaLAPACKMatrix< T > &rotationMatPar, const dftParameters &dftParams, const bool rotationMatTranspose=false, const bool doCommAfterBandParal=true)
	Computes X^{T}=Q*X^{T} inplace. X^{T} is the subspaceVectorsArray stored in the column major format (N x M). Q is rotationMatPar (N x N).

Detailed Description

Contains internal functions used in linearAlgebraOperations.

Author

Sambit Das

Function Documentation

broadcastAcrossInterCommScaLAPACKMat()

template<typename T>

void dftfe::linearAlgebraOperations::internal::broadcastAcrossInterCommScaLAPACKMat	(	const std::shared_ptr< const dftfe::ProcessGrid > &	processGrid,
		dftfe::ScaLAPACKMatrix< T > &	mat,
		const MPI_Comm &	interComm,
		const unsigned int	broadcastRoot )

MPI_Bcast of ScaLAPACKMat across a given inter communicator from a given broadcast root. Used for band parallelization.

createGlobalToLocalIdMapsScaLAPACKMat()

template<typename T>

void dftfe::linearAlgebraOperations::internal::createGlobalToLocalIdMapsScaLAPACKMat	(	const std::shared_ptr< const dftfe::ProcessGrid > &	processGrid,
		const dftfe::ScaLAPACKMatrix< T > &	mat,
		std::unordered_map< unsigned int, unsigned int > &	globalToLocalRowIdMap,
		std::unordered_map< unsigned int, unsigned int > &	globalToLocalColumnIdMap )

Creates global row/column id to local row/column ids for dftfe::ScaLAPACKMatrix.

createProcessGridRectangularMatrix()

void dftfe::linearAlgebraOperations::internal::createProcessGridRectangularMatrix	(	const MPI_Comm &	mpi_communicator,
		const unsigned	sizeRows,
		const unsigned	sizeColumns,
		std::shared_ptr< const dftfe::ProcessGrid > &	processGrid,
		const dftParameters &	dftParams )

Wrapper function to create a two dimensional processor grid for a rectangular matrix in dftfe::ScaLAPACKMatrix storage format.

createProcessGridSquareMatrix()

void dftfe::linearAlgebraOperations::internal::createProcessGridSquareMatrix	(	const MPI_Comm &	mpi_communicator,
		const unsigned	size,
		std::shared_ptr< const dftfe::ProcessGrid > &	processGrid,
		const dftParameters &	dftParams,
		const bool	useOnlyThumbRule = false )

Wrapper function to create a two dimensional processor grid for a square matrix in dftfe::ScaLAPACKMatrix storage format.

fillParallelOverlapMatrix()

template<typename T>

void dftfe::linearAlgebraOperations::internal::fillParallelOverlapMatrix	(	const T *	X,
		const std::shared_ptr< dftfe::linearAlgebra::BLASWrapper< dftfe::utils::MemorySpace::HOST > > &	BLASWrapperPtr,
		const unsigned int	XLocalSize,
		const unsigned int	numberVectors,
		const std::shared_ptr< const dftfe::ProcessGrid > &	processGrid,
		const MPI_Comm &	interBandGroupComm,
		const MPI_Comm &	mpiComm,
		dftfe::ScaLAPACKMatrix< T > &	overlapMatPar,
		const dftParameters &	dftParams )

Computes Sc=X^{T}*Xc and stores in a parallel ScaLAPACK matrix. X^{T} is the subspaceVectorsArray stored in the column major format (N x M). Sc is the overlapMatPar.

The overlap matrix computation and filling is done in a blocked approach which avoids creation of full serial overlap matrix memory, and also avoids creation of another full X memory.

fillParallelOverlapMatrixMixedPrec()

template<typename T, typename TLowPrec>

void dftfe::linearAlgebraOperations::internal::fillParallelOverlapMatrixMixedPrec	(	const T *	X,
		const std::shared_ptr< dftfe::linearAlgebra::BLASWrapper< dftfe::utils::MemorySpace::HOST > > &	BLASWrapperPtr,
		const unsigned int	XLocalSize,
		const unsigned int	numberVectors,
		const std::shared_ptr< const dftfe::ProcessGrid > &	processGrid,
		const MPI_Comm &	interBandGroupComm,
		const MPI_Comm &	mpiComm,
		dftfe::ScaLAPACKMatrix< T > &	overlapMatPar,
		const dftParameters &	dftParams )

Computes Sc=X^{T}*Xc and stores in a parallel ScaLAPACK matrix. X^{T} is the subspaceVectorsArray stored in the column major format (N x M). Sc is the overlapMatPar.

The overlap matrix computation and filling is done in a blocked approach which avoids creation of full serial overlap matrix memory, and also avoids creation of another full X memory.

scaleScaLAPACKMat()

template<typename T>

void dftfe::linearAlgebraOperations::internal::scaleScaLAPACKMat	(	const std::shared_ptr< const dftfe::ProcessGrid > &	processGrid,
		const std::shared_ptr< dftfe::linearAlgebra::BLASWrapper< dftfe::utils::MemorySpace::HOST > > &	BLASWrapperPtr,
		dftfe::ScaLAPACKMatrix< T > &	mat,
		const T	scalar )

scale a ScaLAPACKMat with a scalar

setupELPAHandleParameters()

void dftfe::linearAlgebraOperations::internal::setupELPAHandleParameters	(	const MPI_Comm &	mpi_communicator,
		MPI_Comm &	processGridCommunicatorActive,
		const std::shared_ptr< const dftfe::ProcessGrid > &	processGrid,
		const unsigned int	na,
		const unsigned int	nev,
		const unsigned int	blockSize,
		elpa_t &	elpaHandle,
		const dftParameters &	dftParams )

setup ELPA parameters.

subspaceRotation()

template<typename T>

void dftfe::linearAlgebraOperations::internal::subspaceRotation	(	T *	subspaceVectorsArray,
		const std::shared_ptr< dftfe::linearAlgebra::BLASWrapper< dftfe::utils::MemorySpace::HOST > > &	BLASWrapperPtr,
		const unsigned int	subspaceVectorsArrayLocalSize,
		const unsigned int	N,
		const std::shared_ptr< const dftfe::ProcessGrid > &	processGrid,
		const MPI_Comm &	interBandGroupComm,
		const MPI_Comm &	mpiComm,
		const dftfe::ScaLAPACKMatrix< T > &	rotationMatPar,
		const dftParameters &	dftParams,
		const bool	rotationMatTranspose = false,
		const bool	isRotationMatLowerTria = false,
		const bool	doCommAfterBandParal = true )

Computes X^{T}=Q*X^{T} inplace. X^{T} is the subspaceVectorsArray stored in the column major format (N x M). Q is rotationMatPar (N x N).

The subspace rotation inside this function is done in a blocked approach which avoids creation of full serial rotation matrix memory, and also avoids creation of another full subspaceVectorsArray memory. subspaceVectorsArrayLocalSize=N*M

subspaceRotationCGSMixedPrec()

template<typename T, typename TLowPrec>

void dftfe::linearAlgebraOperations::internal::subspaceRotationCGSMixedPrec	(	T *	subspaceVectorsArray,
		const std::shared_ptr< dftfe::linearAlgebra::BLASWrapper< dftfe::utils::MemorySpace::HOST > > &	BLASWrapperPtr,
		const unsigned int	subspaceVectorsArrayLocalSize,
		const unsigned int	N,
		const std::shared_ptr< const dftfe::ProcessGrid > &	processGrid,
		const MPI_Comm &	interBandGroupComm,
		const MPI_Comm &	mpiComm,
		const dftfe::ScaLAPACKMatrix< T > &	rotationMatPar,
		const dftParameters &	dftParams,
		const bool	rotationMatTranspose = false,
		const bool	doCommAfterBandParal = true )

Computes X^{T}=Q*X^{T} inplace. X^{T} is the subspaceVectorsArray stored in the column major format (N x M). Q is rotationMatPar (N x N).

The subspace rotation inside this function is done in a blocked approach which avoids creation of full serial rotation matrix memory, and also avoids creation of another full subspaceVectorsArray memory. subspaceVectorsArrayLocalSize=N*M

subspaceRotationMixedPrec()

template<typename T, typename TLowPrec>

void dftfe::linearAlgebraOperations::internal::subspaceRotationMixedPrec	(	T *	subspaceVectorsArray,
		const std::shared_ptr< dftfe::linearAlgebra::BLASWrapper< dftfe::utils::MemorySpace::HOST > > &	BLASWrapperPtr,
		const unsigned int	subspaceVectorsArrayLocalSize,
		const unsigned int	N,
		const std::shared_ptr< const dftfe::ProcessGrid > &	processGrid,
		const MPI_Comm &	interBandGroupComm,
		const MPI_Comm &	mpiComm,
		const dftfe::ScaLAPACKMatrix< T > &	rotationMatPar,
		const dftParameters &	dftParams,
		const bool	rotationMatTranspose = false,
		const bool	doCommAfterBandParal = true )

Computes X^{T}=Q*X^{T} inplace. X^{T} is the subspaceVectorsArray stored in the column major format (N x M). Q is rotationMatPar (N x N).

The subspace rotation inside this function is done in a blocked approach which avoids creation of full serial rotation matrix memory, and also avoids creation of another full subspaceVectorsArray memory. subspaceVectorsArrayLocalSize=N*M

subspaceRotationSpectrumSplit()

template<typename T>

void dftfe::linearAlgebraOperations::internal::subspaceRotationSpectrumSplit	(	const T *	X,
		const std::shared_ptr< dftfe::linearAlgebra::BLASWrapper< dftfe::utils::MemorySpace::HOST > > &	BLASWrapperPtr,
		T *	Y,
		const unsigned int	subspaceVectorsArrayLocalSize,
		const unsigned int	N,
		const std::shared_ptr< const dftfe::ProcessGrid > &	processGrid,
		const unsigned int	numberTopVectors,
		const MPI_Comm &	interBandGroupComm,
		const MPI_Comm &	mpiComm,
		const dftfe::ScaLAPACKMatrix< T > &	QMat,
		const dftParameters &	dftParams,
		const bool	QMatTranspose = false )

Computes Y^{T}=Q*X^{T}.

X^{T} is stored in the column major format (N x M). Q is extracted from the supplied QMat as Q=QMat{1:numberTopVectors}. If QMat is in column major format set QMatTranspose=false, otherwise set to true if in row major format. The dimensions (in row major format) of QMat could be either a) (N x numberTopVectors) or b) (N x N) where numberTopVectors!=N. In this case it is assumed that Q is stored in the first numberTopVectors columns of QMat. The subspace rotation inside this function is done in a blocked approach which avoids creation of full serial rotation matrix memory, and also avoids creation of another full X memory. subspaceVectorsArrayLocalSize=N*M

subspaceRotationSpectrumSplitMixedPrec()

template<typename T, typename TLowPrec>

void dftfe::linearAlgebraOperations::internal::subspaceRotationSpectrumSplitMixedPrec	(	const T *	X,
		const std::shared_ptr< dftfe::linearAlgebra::BLASWrapper< dftfe::utils::MemorySpace::HOST > > &	BLASWrapperPtr,
		T *	Y,
		const unsigned int	subspaceVectorsArrayLocalSize,
		const unsigned int	N,
		const std::shared_ptr< const dftfe::ProcessGrid > &	processGrid,
		const unsigned int	numberTopVectors,
		const MPI_Comm &	interBandGroupComm,
		const MPI_Comm &	mpiComm,
		const dftfe::ScaLAPACKMatrix< T > &	QMat,
		const dftParameters &	dftParams,
		const bool	QMatTranspose = false )

Computes Y^{T}=Q*X^{T}.

X^{T} is stored in the column major format (N x M). Q is extracted from the supplied QMat as Q=QMat{1:numberTopVectors}. If QMat is in column major format set QMatTranspose=false, otherwise set to true if in row major format. The dimensions (in row major format) of QMat could be either a) (N x numberTopVectors) or b) (N x N) where numberTopVectors!=N. In this case it is assumed that Q is stored in the first numberTopVectors columns of QMat. The subspace rotation inside this function is done in a blocked approach which avoids creation of full serial rotation matrix memory, and also avoids creation of another full X memory. subspaceVectorsArrayLocalSize=N*M

sumAcrossInterCommScaLAPACKMat()

template<typename T>

void dftfe::linearAlgebraOperations::internal::sumAcrossInterCommScaLAPACKMat	(	const std::shared_ptr< const dftfe::ProcessGrid > &	processGrid,
		dftfe::ScaLAPACKMatrix< T > &	mat,
		const MPI_Comm &	interComm )

Mpi all reduce of ScaLAPACKMat across a given inter communicator. Used for band parallelization.