dftefe::linearAlgebra::MultiVectorOps Class Reference

Stateless class of static functions for spin-aware projection and rotation of product-space multivectors. More...

#include <MultiVectorOps.h>

Static Public Member Functions
template<typename ValueTypeOperator , typename ValueTypeOperand , utils::MemorySpace memorySpace>
static void	project (const OperatorContext< ValueTypeOperator, ValueTypeOperand, memorySpace > &Op, MultiVectorProductSpaceBlocked< ValueTypeOperand, memorySpace > &X, std::vector< ScaLAPACKMatrix< blasLapack::scalar_type< ValueTypeOperator, ValueTypeOperand > > > &Ps, const ElpaScalapackManager &elpa, std::shared_ptr< MultiVector< blasLapack::scalar_type< ValueTypeOperator, ValueTypeOperand >, memorySpace > > &scratchXin, std::shared_ptr< MultiVector< blasLapack::scalar_type< ValueTypeOperator, ValueTypeOperand >, memorySpace > > &scratchXout, std::shared_ptr< MultiVector< blasLapack::scalar_type< ValueTypeOperator, ValueTypeOperand >, memorySpace > > &scratchXinSmall, std::shared_ptr< MultiVector< blasLapack::scalar_type< ValueTypeOperator, ValueTypeOperand >, memorySpace > > &scratchXoutSmall)
	Blocked overload (collinear, MultiVectorProductSpaceBlocked). More...
template<typename ValueTypeOperator , typename ValueTypeOperand , utils::MemorySpace memorySpace>
static void	project (const OperatorContext< ValueTypeOperator, ValueTypeOperand, memorySpace > &Op, MultiVectorProductSpace< ValueTypeOperand, memorySpace > &X, ScaLAPACKMatrix< blasLapack::scalar_type< ValueTypeOperator, ValueTypeOperand > > &P, const ElpaScalapackManager &elpa, std::shared_ptr< MultiVector< blasLapack::scalar_type< ValueTypeOperator, ValueTypeOperand >, memorySpace > > &scratchXin, std::shared_ptr< MultiVector< blasLapack::scalar_type< ValueTypeOperator, ValueTypeOperand >, memorySpace > > &scratchXout, std::shared_ptr< MultiVector< blasLapack::scalar_type< ValueTypeOperator, ValueTypeOperand >, memorySpace > > &scratchXinSmall, std::shared_ptr< MultiVector< blasLapack::scalar_type< ValueTypeOperator, ValueTypeOperand >, memorySpace > > &scratchXoutSmall)
	Coupled overload (unpolarized S=1 / non-collinear S=2, MultiVectorProductSpace). More...
template<typename ValueType , utils::MemorySpace memorySpace>
static void	rotate (MultiVectorProductSpaceBlocked< ValueType, memorySpace > &X, const std::vector< ScaLAPACKMatrix< ValueType > > &Qs, const ElpaScalapackManager &elpa)
	Blocked rotate overload (collinear, MultiVectorProductSpaceBlocked). More...
template<typename ValueType , utils::MemorySpace memorySpace>
static void	rotate (MultiVectorProductSpace< ValueType, memorySpace > &X, const ScaLAPACKMatrix< ValueType > &Q, const ElpaScalapackManager &elpa)
	Coupled rotate overload (unpolarized / non-collinear, MultiVectorProductSpace). More...
template<typename ValueType1 , typename ValueType2 , utils::MemorySpace memorySpace>
static void	copyToBatch (const MultiVectorProductSpace< ValueType1, memorySpace > &X, size_type srcStart, size_type numVecBatch, MultiVector< ValueType2, memorySpace > &Xbatch, LinAlgOpContext< memorySpace > &context)
template<typename ValueType1 , typename ValueType2 , utils::MemorySpace memorySpace>
static void	copyFromBatch (const MultiVector< ValueType1, memorySpace > &Ybatch, size_type dstStart, size_type numVecBatch, MultiVectorProductSpace< ValueType2, memorySpace > &Y, LinAlgOpContext< memorySpace > &context)

Detailed Description

Stateless class of static functions for spin-aware projection and rotation of product-space multivectors.

project() computes P = X^H Op X. rotate() performs X <- X * Q in-place.

Two overloads per function, selected at compile time by the multivector type (most-derived-first):

MultiVectorProductSpaceBlocked (collinear): Blocked loop over S independent N×N sub-problems. Spin-s columns accessed via ptr = X.data() + s*N, lda = S*N — no intermediate copy.

MultiVectorProductSpace (unpolarized S=1 or non-collinear S=2): Single call on the full S*N columns as a coupled unit. S=1 degenerates exactly to the existing scalar path.

scratchXin / scratchXout / scratchXinSmall / scratchXoutSmall for project() are owned by the caller and persist across SCF calls to avoid repeated allocation. All other scratch (SBlock, rotation buffers) is allocated internally per call, consistent with ElpaScalapackOperations.

dynamic_cast is NOT performed inside these functions. The caller (RayleighRitzEigenSolver::solve, OrthonormalizationFunctions methods) performs one dynamic_cast per call, then passes the correctly-typed reference here so the compiler selects the right overload.

Member Function Documentation

copyFromBatch()

template<typename ValueType1 , typename ValueType2 , utils::MemorySpace memorySpace>

void dftefe::linearAlgebra::MultiVectorOps::copyFromBatch ( const MultiVector< ValueType1, memorySpace > &  Ybatch, size_type  dstStart, size_type  numVecBatch, MultiVectorProductSpace< ValueType2, memorySpace > &  Y, LinAlgOpContext< memorySpace > &  context  )

static

Here is the call graph for this function:

Here is the caller graph for this function:

copyToBatch()

template<typename ValueType1 , typename ValueType2 , utils::MemorySpace memorySpace>

void dftefe::linearAlgebra::MultiVectorOps::copyToBatch ( const MultiVectorProductSpace< ValueType1, memorySpace > &  X, size_type  srcStart, size_type  numVecBatch, MultiVector< ValueType2, memorySpace > &  Xbatch, LinAlgOpContext< memorySpace > &  context  )

static

Here is the call graph for this function:

Here is the caller graph for this function:

project() [1/2]

template<typename ValueTypeOperator , typename ValueTypeOperand , utils::MemorySpace memorySpace>

void dftefe::linearAlgebra::MultiVectorOps::project ( const OperatorContext< ValueTypeOperator, ValueTypeOperand, memorySpace > &  Op, MultiVectorProductSpace< ValueTypeOperand, memorySpace > &  X, ScaLAPACKMatrix< blasLapack::scalar_type< ValueTypeOperator, ValueTypeOperand > > &  P, const ElpaScalapackManager &  elpa, std::shared_ptr< MultiVector< blasLapack::scalar_type< ValueTypeOperator, ValueTypeOperand >, memorySpace > > &  scratchXin, std::shared_ptr< MultiVector< blasLapack::scalar_type< ValueTypeOperator, ValueTypeOperand >, memorySpace > > &  scratchXout, std::shared_ptr< MultiVector< blasLapack::scalar_type< ValueTypeOperator, ValueTypeOperand >, memorySpace > > &  scratchXinSmall, std::shared_ptr< MultiVector< blasLapack::scalar_type< ValueTypeOperator, ValueTypeOperand >, memorySpace > > &  scratchXoutSmall  )

static

Coupled overload (unpolarized S=1 / non-collinear S=2, MultiVectorProductSpace).

Computes a single overlap matrix: P = X^H Op X (S*N × S*N) The full S*N columns are treated as a coupled unit. S=1 degenerates to the existing N×N scalar path.

Parameters

Op	Operator context.
X	Coupled product-space multivector (M × S*N).
P	Output: single ScaLAPACK S*N×S*N matrix (pre-allocated).
elpa	ELPA/ScaLAPACK manager.
scratchXin	Scratch multivector (M × batchSize), caller-owned.
scratchXout	Scratch multivector (M × batchSize), caller-owned.
scratchXinSmall	Tail-batch scratch, nullptr on first call; cached.
scratchXoutSmall	Tail-batch scratch, nullptr on first call; cached.

Here is the call graph for this function:

project() [2/2]

template<typename ValueTypeOperator , typename ValueTypeOperand , utils::MemorySpace memorySpace>

void dftefe::linearAlgebra::MultiVectorOps::project ( const OperatorContext< ValueTypeOperator, ValueTypeOperand, memorySpace > &  Op, MultiVectorProductSpaceBlocked< ValueTypeOperand, memorySpace > &  X, std::vector< ScaLAPACKMatrix< blasLapack::scalar_type< ValueTypeOperator, ValueTypeOperand > > > &  Ps, const ElpaScalapackManager &  elpa, std::shared_ptr< MultiVector< blasLapack::scalar_type< ValueTypeOperator, ValueTypeOperand >, memorySpace > > &  scratchXin, std::shared_ptr< MultiVector< blasLapack::scalar_type< ValueTypeOperator, ValueTypeOperand >, memorySpace > > &  scratchXout, std::shared_ptr< MultiVector< blasLapack::scalar_type< ValueTypeOperator, ValueTypeOperand >, memorySpace > > &  scratchXinSmall, std::shared_ptr< MultiVector< blasLapack::scalar_type< ValueTypeOperator, ValueTypeOperand >, memorySpace > > &  scratchXoutSmall  )

static

Blocked overload (collinear, MultiVectorProductSpaceBlocked).

Computes S independent overlap matrices: Ps[s] = Xs^H Op Xs (N×N each) where Xs is the spin-s block of X accessed as ptr = X.data() + s * X.numVectorsPerSpace(), lda = X.numVectors()

Parameters

Op	Operator context (e.g. B metric or identity).
X	Blocked product-space multivector (M × S*N).
Ps	Output: S ScaLAPACK N×N matrices (pre-allocated by caller).
elpa	ELPA/ScaLAPACK manager (provides process grid).
scratchXin	Scratch multivector (M × batchSize), caller-owned.
scratchXout	Scratch multivector (M × batchSize), caller-owned.
scratchXinSmall	Tail-batch scratch, nullptr on first call; cached.
scratchXoutSmall	Tail-batch scratch, nullptr on first call; cached.

Here is the call graph for this function:

Here is the caller graph for this function:

rotate() [1/2]

template<typename ValueType , utils::MemorySpace memorySpace>

void dftefe::linearAlgebra::MultiVectorOps::rotate ( MultiVectorProductSpace< ValueType, memorySpace > &  X, const ScaLAPACKMatrix< ValueType > &  Q, const ElpaScalapackManager &  elpa  )

static

Coupled rotate overload (unpolarized / non-collinear, MultiVectorProductSpace).

Performs a single in-place subspace rotation on all S*N columns: X <- X * Q One M×S*N scratch buffer (unavoidable for in-place operation). S=1 degenerates to the existing scalar rotation path.

Parameters

X	Coupled multivector, modified in place.
Q	Single rotation matrix (S*N × S*N).
elpa	ELPA/ScaLAPACK manager.

Here is the call graph for this function:

rotate() [2/2]

template<typename ValueType , utils::MemorySpace memorySpace>

void dftefe::linearAlgebra::MultiVectorOps::rotate ( MultiVectorProductSpaceBlocked< ValueType, memorySpace > &  X, const std::vector< ScaLAPACKMatrix< ValueType > > &  Qs, const ElpaScalapackManager &  elpa  )

static

Blocked rotate overload (collinear, MultiVectorProductSpaceBlocked).

Performs S independent in-place subspace rotations: Xs <- Xs * Qs[s] using ptr = X.data() + s*N, lda = S*N. scratchRotBlock (M×N) is reused across s iterations — S× smaller than the coupled scratch.

Parameters

X	Blocked multivector, modified in place.
Qs	S rotation matrices (N×N each).
elpa	ELPA/ScaLAPACK manager.

Here is the call graph for this function:

Here is the caller graph for this function:

---

The documentation for this class was generated from the following files:

• src/linearAlgebra/MultiVectorOps.h
• src/linearAlgebra/MultiVectorOps.t.cpp