[petsc-users] User defined solver for PCMG

[Barry Smith]

> On May 17, 2016, at 1:49 PM, Pratapa, Phanisri P <ppratapa at gatech.edu> wrote:
> 
> @Barry: Thank you for the information.
> 
> @Jed: The notes on PCMG (http://www.mcs.anl.gov/petsc/petsc-current/docs/manualpages/PC/PCMG.html#PCMG) says:
> "By default this uses GMRES on the fine grid smoother so this should be used with KSPFGMRES or the smoother changed to not use GMRES". I understood this statement to mean that GMRES is the default smoother.

   Thanks for letting us know, this is no longer correct. We now use Chebyshev as the default smoother KSP. I have fixed the documentation.


> I want to change the Krylov method (and not the smoother) that is being preconditioned. So I guess if I just implement my solver as KSP and use PCMG as preconditioner that should solve my problem. Please let me know if this sounds fine.
> 

   Yes

> Thank you,
> 
> Regards,
> 
> Pradeep
> 
> ________________________________________
> From: Barry Smith <bsmith at mcs.anl.gov>
> Sent: Tuesday, May 17, 2016 2:26:59 PM
> To: Pratapa, Phanisri P
> Cc: petsc-users at mcs.anl.gov
> Subject: Re: [petsc-users] User defined solver for PCMG
> 
>> On May 17, 2016, at 1:14 PM, Pratapa, Phanisri P <ppratapa at gatech.edu> wrote:
>> 
>> Barry,
>> 
>> What I mean is that: I want to implement multigrid preconditioning on a new Krylov method that is not currently in PETSc. For this, I was hoping that I could replace the KSPFGMRES smoother (default) with the "new solver" I have. But I do not have the new solver as a PETSc KSP yet.
> 
>   Understood. We don't provide a KSPSHELL because we consider it so easy to implement a new KSP in PETSc that having a KSPSHELL is unnecessary. If your new Krylov method is "stand-alone" and not, for example, a modification of GMRES here is how to proceed. Say your KSP is called mykrylov
> 
>    Make a directory src/ksp/ksp/impls/mykrylov Copy src/ksp/ksp/impls/cg/cg.c and src/ksp/ksp/impls/cg/cgimpl.c and  src/ksp/ksp/impls/cg/makefile to that directory changing the .c and .h file names. Then edit the three files to reflect the new names. Edit
> src/ksp/ksp/impls/makefile and add the mykrylov directory to the list of directories (variable DIRS). Then code your new Krylov method inside the two new files you copied over following the directions given in the file.
> 
>   If your new Krylov method is an extension/modification of GMRES then it is possible to reuse most of the GMRES implementation in PETSc but a bit more involved. Under the src/ksp/ksp/impls/gmres directory are several variants fgmres, pgmres, lgmres, pipefgmres, agmres, and dgmres. I would recommend picking the one that most closely resembles your new method and copying it as above and modifying it to match your algorithm.
> 
>   Barry
> 
> 
> 
>> 
>> Thank you,
>> 
>> Regards,
>> 
>> Pradeep
>> ________________________________________
>> From: Barry Smith <bsmith at mcs.anl.gov>
>> Sent: Tuesday, May 17, 2016 2:08:18 PM
>> To: Pratapa, Phanisri P
>> Cc: petsc-users at mcs.anl.gov
>> Subject: Re: [petsc-users] User defined solver for PCMG
>> 
>>> On May 17, 2016, at 12:21 PM, Pratapa, Phanisri P <ppratapa at gatech.edu> wrote:
>>> 
>>> Hi,
>>> 
>>> I am trying to find out if one can use a user defined linear solver function in PCMG (instead of the default GMRES). According to the petsc manual, I can change the solver/smoother through the KSP context and the available solvers, but I am interested in using my own function (solver).
>> 
>>  What do you mean here by "solver"? Do you want to provide a new Krylov method that is not currently in PETSc or a new preconditioner that is specific to your problem and cannot be written as a composition of preconditioners and Krylov methods already in PETSc?
>> 
>>  An example of your own custom preconditioner could be an SOR iteration that you hand code based on the stencil and doesn't use a stored matrix. In that case you would access the PC on the level and use PCSetType(subpc,PCSHELL) and PCShellSetApply(subpc, your custom function).
>> 
>> Barry
>> 
>>> 
>>> Thank you,
>>> 
>>> Regards,
>>> 
>>> Pradeep
>> 
>