quda::MinResExt Class Reference

This computes the optimum guess for the system Ax=b in the L2 residual norm. For use in the HMD force calculations using a minimal residual chronological method. This computes the guess solution as a linear combination of a given number of previous solutions. Following Brower et al, only the orthogonalised vector basis is stored to conserve memory. More...

#include <invert_quda.h>

Collaboration diagram for quda::MinResExt:

Public Member Functions
	MinResExt (DiracMatrix &mat, bool orthogonal, bool apply_mat, bool hermitian, TimeProfile &profile)
virtual	~MinResExt ()
void	operator() (ColorSpinorField &x, ColorSpinorField &b, std::vector< std::pair< ColorSpinorField *, ColorSpinorField *> > basis)
void	operator() (ColorSpinorField &x, ColorSpinorField &b, std::vector< ColorSpinorField *> p, std::vector< ColorSpinorField *> q)

Protected Member Functions
void	solve (Complex *psi_, std::vector< ColorSpinorField *> &p, std::vector< ColorSpinorField *> &q, ColorSpinorField &b, bool hermitian)
	Solve the equation A p_k psi_k = q_k psi_k = b by minimizing the residual and using Eigen's SVD algorithm for numerical stability. More...

Protected Attributes
const DiracMatrix &	mat
bool	orthogonal
bool	apply_mat
	Whether to construct an orthogonal basis or not. More...
bool	hermitian
	Whether to compute q = Ap or assume it is provided. More...
TimeProfile &	profile
	whether A is hermitian ot not More...

Detailed Description

This computes the optimum guess for the system Ax=b in the L2 residual norm. For use in the HMD force calculations using a minimal residual chronological method. This computes the guess solution as a linear combination of a given number of previous solutions. Following Brower et al, only the orthogonalised vector basis is stored to conserve memory.

If Eigen support is enabled then Eigen's SVD algorithm is used for solving the linear system, else Gaussian elimination with partial pivots is used.

Definition at line 1171 of file invert_quda.h.

Constructor & Destructor Documentation

MinResExt()

quda::MinResExt::MinResExt	(	DiracMatrix &	mat,
		bool	orthogonal,
		bool	apply_mat,
		bool	hermitian,
		TimeProfile &	profile
	)

Parameters

mat	The operator for the linear system we wish to solve
orthogonal	Whether to construct an orthogonal basis prior to constructing the linear system
apply_mat	Whether to apply the operator in place or assume q already contains this Timing profile to use

Definition at line 7 of file inv_mre.cpp.

~MinResExt()

quda::MinResExt::~MinResExt ( )

virtual

Definition at line 12 of file inv_mre.cpp.

Member Function Documentation

operator()() [1/2]

void quda::MinResExt::operator()	(	ColorSpinorField &	x,
		ColorSpinorField &	b,
		std::vector< std::pair< ColorSpinorField *, ColorSpinorField *> >	basis
	)

Parameters

x	The optimum for the solution vector.
b	The source vector in the equation to be solved. This is not preserved and is overwritten by the new residual.
basis	Vector of pairs storing the basis (p,Ap)

Definition at line 165 of file inv_mre.cpp.

operator()() [2/2]

void quda::MinResExt::operator()	(	ColorSpinorField &	x,
		ColorSpinorField &	b,
		std::vector< ColorSpinorField *>	p,
		std::vector< ColorSpinorField *>	q
	)

Parameters

x	The optimum for the solution vector.
b	The source vector in the equation to be solved. This is not preserved.
p	The basis vectors in which we are building the guess
q	The basis vectors multiplied by A

Definition at line 82 of file inv_mre.cpp.

References apply_mat, quda::blas::ax(), quda::blas::caxpy(), quda::blas::cDotProduct(), quda::blas::copy(), getVerbosity(), hermitian, quda::TimeProfile::isRunning(), mat, quda::blas::norm2(), orthogonal, printfQuda, profile, quda::QUDA_PROFILE_CHRONO, QUDA_SUMMARIZE, solve(), quda::sqrt(), X, and quda::blas::zero().

Here is the call graph for this function:

solve()

void quda::MinResExt::solve ( Complex *  psi_, std::vector< ColorSpinorField *> &  p, std::vector< ColorSpinorField *> &  q, ColorSpinorField &  b, bool  hermitian  )

protected

Solve the equation A p_k psi_k = q_k psi_k = b by minimizing the residual and using Eigen's SVD algorithm for numerical stability.

Parameters

[out]	psi	Array of coefficients
[in]	p	Search direction vectors
[in]	q	Search direction vectors with the operator applied
[in]	hermitian	Whether the linear system is Hermitian or not

Definition at line 18 of file inv_mre.cpp.

References quda::blas::cDotProduct(), profile, quda::QUDA_PROFILE_CHRONO, and quda::QUDA_PROFILE_EIGEN.

Referenced by operator()().

Here is the call graph for this function:

Here is the caller graph for this function:

Member Data Documentation

apply_mat

bool quda::MinResExt::apply_mat

protected

Whether to construct an orthogonal basis or not.

Definition at line 1176 of file invert_quda.h.

Referenced by operator()().

hermitian

bool quda::MinResExt::hermitian

protected

Whether to compute q = Ap or assume it is provided.

Definition at line 1177 of file invert_quda.h.

Referenced by operator()().

mat

const DiracMatrix& quda::MinResExt::mat

protected

Definition at line 1174 of file invert_quda.h.

Referenced by operator()().

orthogonal

bool quda::MinResExt::orthogonal

protected

Definition at line 1175 of file invert_quda.h.

Referenced by operator()().

profile

TimeProfile& quda::MinResExt::profile

protected

whether A is hermitian ot not

Definition at line 1178 of file invert_quda.h.

Referenced by operator()(), and solve().

---

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

• include/invert_quda.h
• lib/inv_mre.cpp