[petsc-users] Reusing the preconditioner when using KSPSetComputeOperators

[Åsmund Ervik]

On 29. aug. 2014 19:03, Barry Smith wrote:
> 
>   
> On Aug 29, 2014, at 9:35 AM, Åsmund Ervik <asmund.ervik at ntnu.no> wrote:
> 
>>> On 28. aug. 2014 20:52, Barry Smith wrote:
>>>>
>>>> On Aug 28, 2014, at 4:34 AM, Åsmund Ervik <asmund.ervik at ntnu.no> wrote:
>>>>
>>>>> Hello,
>>>>>
>>>>> I am solving a pressure Poisson equation with KSP, where the initial
>>>>> guess, RHS and matrix are computed by functions that I've hooked into
>>>>> KSPSetComputeXXX. (I'm also using DMDA for my domain decomposition.)
>>>>>
>>>>> For (single-phase|two-phase) I would like to (reuse|not reuse) the
>>>>> preconditioner. How do I specify that when using this way of setting the
>>>>> operator? Is it toggled by whether or not I call KSPSetOperators before
>>>>> each KSPSolve? (The manual does not mention KSPSetComputeXXX.)
>>>>
>>>>    You should call KSPSetOperators() before each KSPSolve() (otherwise the function you  provide to compute the matrix won’t be triggered).
>>>>
>>>>    With PETSc 3.5 after the call to KSPSetOperators() call KSPSetReusePreconditioner() to tell KSP wether to reuse the preconditioner or build a new one.
>>>>    With PETSc 3.4 and earlier, the final argument to KSPSetOperators() would be MAT_SAME_PRECONDITIONER to reuse the preconditioner or MAT_SAME_NONZERO_PATTERN to construct a new preconditioner
>>>>
>>
>> Thanks Barry for the clarification. Is there an example somewhere that
>> does this? All the ones I can find which use KSPSetComputeOperators()
>> have no calls to KSPSetOperators(). I guess this is because they are
>> only doing one linear solve?
>>
>> Furthermore, what should I pass in for Amat and Pmat to the KSPSetOperators() call? PetscNullObject, or do I get the Amat from the KSP somehow?
> 
> You better call KSPGetOperators() to get them. Yes this is kind of silly.

Okay, thanks. I found I also have to call MatSetFromOptions() on Amat
and Pmat after I get them from KSPGetOperators(), otherwise
KSPSetOperators() complains that
"Object is in the wrong state! Mat object's type is not set!"

This is indeed a strange sequence of calls. If this is the "standard"
way of doing like ksp/ex34.c but for several successive solves, then it
would be nice to have a comment/instruction in ex34.c for that case.

Åsmund

-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 473 bytes
Desc: OpenPGP digital signature
URL: <http://lists.mcs.anl.gov/pipermail/petsc-users/attachments/20140901/48006838/attachment.pgp>