[petsc-users] Preconditioning with SLEPc
Jose E. Roman
jroman at dsic.upv.es
Tue Aug 7 10:13:37 CDT 2018
Currently there is no straightforward way to do this in SLEPc. For a preconditioned eigensolver such as JD you could use this:
http://slepc.upv.es/documentation/current/docs/manualpages/ST/STPrecondSetMatForPC.html
in combination with pc=PCMAT.
For a Krylov solver you will have to do something like this:
call EPSSetWhichEigenpairs(eps,EPS_TARGET_REAL,ierr);CHKERRA(ierr)
call EPSSetTarget(eps,sigma,ierr);CHKERRA(ierr)
call EPSGetST(eps,st,ierr);CHKERRA(ierr)
call STSetType(st,STSINVERT,ierr);CHKERRA(ierr)
call STGetKSP(st,ksp,ierr);CHKERRA(ierr)
call KSPGetPC(ksp,pc,ierr);CHKERRA(ierr)
call KSPSetType(ksp,KSPBCGS,ierr);CHKERRA(ierr)
call PCSetType(pc,PCNONE,ierr);CHKERRA(ierr) ! dummy preconditioner
! provide the preconditioner matrix after setup
call EPSSetUp(eps,ierr);CHKERRA(ierr)
call PCSetOperators(pc,PETSC_NULL_MAT,my_matmul_K,ierr);CHKERRA(ierr)
call PCSetType(pc,PCMAT,ierr);CHKERRA(ierr)
We may change the interface in future versions so that this can be done in a more natural way.
Jose
> El 7 ago 2018, a las 16:42, Kenneth Hall, SC.D. <kenneth.c.hall at duke.edu> escribió:
>
> Hello,
>
> I am new to SLEPc and Petsc, so I would appreciate any help I can get with my question. I have a large eigenvalue problem I am solving, based on the linearization of a nonlinear code. The eigenvalue problem takes the form
>
> A.x = lambda.B.x
>
> I can generate the matrix B explicitly, but I can only form products of A.x, and not A itself, so I use a shell matrix. I also have an iterative solver I could potentially use as a preconditioner. That is, I have an operator K^-1 where
>
> K^-1 \approx (A - sigma.B)^-1
>
> But again, I don’t have K^-1 explicitly, rather I can form products of K^-1 with a vector, so I would need to use a shell matrix.
>
> I have had success with the un-preconditioned case, but can’t figure out how to do the preconditioned case. Any suggestions would be greatly appreciated. I work with Fortran. A code snippet with the un-preconditioned set-up follows.
>
> Thanks in advance.
> Kenneth Hall
>
>
> ===============================================================
>
> !
> !.... initialize slepc
> call SlepcInitialize(PETSC_NULL_CHARACTER,ierr) ; if (ierr .ne. 0) then ; stop 'SlepcInitialize failed' ; end if
> !
> !.... get size and rank of MPI process
> call MPI_Comm_size(PETSC_COMM_WORLD,sz ,ierr) ; CHKERRA(ierr)
> call MPI_Comm_rank(PETSC_COMM_WORLD,rank,ierr) ; CHKERRA(ierr)
> !
> if (sz .ne. 1) then ; SETERRA(PETSC_COMM_SELF,1,'This is a uniprocessor example only!'); endif
> !
> !.... get user options
> call PetscOptionsGetInt (PETSC_NULL_OPTIONS,PETSC_NULL_CHARACTER,'-n' ,n ,flg,ierr) ; CHKERRA(ierr)
> call PetscOptionsHasName(PETSC_NULL_OPTIONS,PETSC_NULL_CHARACTER,'-terse',terse,ierr) ; CHKERRA(ierr)
> !
> !.... write out option list
> if (rank .eq. 0 ) then
> write(*,*) 'n = ',n
> end if
> !
> !.... Register the matrix-vector subroutine for the operator that defines the eigensystem, Ax=kBx
> call MatCreateShell(PETSC_COMM_WORLD,n,n,n,n,PETSC_NULL_INTEGER,A,ierr) ; CHKERRA(ierr)
> call MatShellSetOperation(A,MATOP_MULT,my_matmul_A,ierr) ; CHKERRA(ierr)
> !
> !.... set up B matrix
> call MatCreate(PETSC_COMM_WORLD,B,ierr) ; CHKERRA(ierr)
> call MatSetSizes(B,n,n,n,n,ierr) ; CHKERRA(ierr)
> call MatSetFromOptions(B,ierr) ; CHKERRA(ierr)
> call MatSetUp(B,ierr) ; CHKERRA(ierr)
> call My_Matrix_B(B,n)
> !
> !.... set up eigenvectors
> call VecCreate(PETSC_COMM_WORLD,xr,ierr) ; CHKERRA(ierr)
> call VecSetSizes(xr,n,n,ierr) ; CHKERRA(ierr)
> call VecSetFromOptions(xr,ierr) ; CHKERRA(ierr)
> call VecSetUp(xr,ierr) ; CHKERRA(ierr)
> !
> call VecCreate(PETSC_COMM_WORLD,xi,ierr) ; CHKERRA(ierr)
> call VecSetSizes(xi,n,n,ierr) ; CHKERRA(ierr)
> call VecSetFromOptions(xi,ierr) ; CHKERRA(ierr)
> call VecSetUp(xi,ierr) ; CHKERRA(ierr)
> !
> !.... Create eigensolver context
> call EPSCreate(PETSC_COMM_WORLD,eps,ierr) ; CHKERRA(ierr)
> !
> !.... Set operators. In this case, it is a generalized eigenvalue problem
> call EPSSetOperators(eps,A,B,ierr) ; CHKERRA(ierr)
> call EPSSetProblemType(eps,EPS_GNHEP,ierr) ; CHKERRA(ierr)
> !
> !.... set default options
> call EPSSetType(eps,EPSKRYLOVSCHUR,ierr) ; CHKERRA(ierr)
> !
> !.... set default eigenvalues to find
> call EPSSetWhichEigenpairs(eps,EPS_LARGEST_REAL,ierr) ; CHKERRA(ierr)
> !
> !.... number of requested eigenvalues
> call EPSSetDimensions(eps,nev,PETSC_DEFAULT_INTEGER,PETSC_DEFAULT_INTEGER,ierr) ; CHKERRA(ierr)
> !
> !.... set stopping criteria
> call EPSSetTolerances(eps,tol,maxit,ierr) ; CHKERRA(ierr)
> !
> !.... set up monitors
> call PetscViewerAndFormatCreate(PETSC_VIEWER_STDOUT_WORLD,PETSC_VIEWER_DEFAULT,vf,ierr) ; CHKERRA(ierr)
> call EPSMonitorSet(eps,EPSMONITORFIRST,vf,PetscViewerAndFormatDestroy,ierr) ; CHKERRA(ierr)
> call SlepcConvMonitorCreate(PETSC_VIEWER_STDOUT_WORLD,PETSC_VIEWER_DEFAULT,ctx,ierr) ; CHKERRA(ierr)
> call EPSMonitorSet(eps,EPSMONITORCONVERGED,ctx,SlepcConvMonitorDestroy,ierr) ; CHKERRA(ierr)
> !
> !.... Set solver parameters at runtime
> call EPSSetFromOptions(eps,ierr) ; CHKERRA(ierr)
> !
> !.... Solve the eigensystem
> call EPSSolve(eps,ierr) ; CHKERRA(ierr)
>
More information about the petsc-users
mailing list