[petsc-users] about the new EPSSetStoppingTestFunction functionality in SLEPC

Jose E. Roman jroman at dsic.upv.es
Wed Jun 1 10:51:34 CDT 2016


On 1 Jun 2016, Giacomo Mulas <gmulas at oa-cagliari.inaf.it> wrote:
> 
> 
> On Wed, 1 Jun 2016, Jose E. Roman wrote:
> 
>> This functionality is currently very basic, not intended for doing
>> sophisticated computations.  Currently there is no public API for
>> accessing eigenvectors during EPSSolve.  Adding this would require a
>> non-trivial effort.
> 
> Well, the currently converged eigenvectors (I need only the vectors, not the
> eigenvalues) _must_ be somewhere, in some format in the eps structure.  Of
> course I would need to only read them without modifying anything.  And
> actually not even all of them, only some components.  Could you tell me if
> there is some part of the code I can read to understand where they are
> stored, so I can try to read them even without high-level interfaces?

Try EPSComputeRitzVector()

> 
> Alternatively, a sort of inelegant but effective way of doing this would be
> to
> 1) make a complete copy, a snapshot, of the EPS structure, store it via the
> private ctx variable of MyStoppingTest every time a new eigenpair is found, and have MyStoppingTest return EPS_CONVERGED_USER
> 2) outside EPSSolve, do the actual convergence check, using all usual fancy
> EPSGetConverged and EPSGetEigenvector calls, which are then available 3) if more eigenpairs are needed, restore eps from the snapshot, so EPSSolve
> does not have to start from scratch every time, and have it resume from
> where it was.
> To achieve this, I would need to know what I need to copy to "snapshot" an
> eps structure, and how to resume a partial EPSSolve calculation from where
> it was interrupted.

SLEPc is not prepared for this kind of stuff. Sorry.

> 
> Would either solution (read converged eigenvectors directly off memory from
> the eps structure or snapshot/resume EPSSolve) be viable without much effort
> on your part (except telling me where to look in the code and possibly
> answering some additional question if there's something I can't figure out)?



More information about the petsc-users mailing list