[petsc-users] Error when using TSSetRHSJacobian+DMSetMatrixPreallocateOnly

[Vijay S. Mahadevan]

>     Why are you setting this flag?  If you don't set it then the code will go through.

Yes it passes through with DMDA. But with my custom DM implementation
(MOAB based), this fails unless I explicitly set the NNZ pattern in
DMCreateMatrix. I currently don't do that and the only way to
replicate this behavior with DMDA was setting the flag explicitly.

>    There is no reason to use a user created matrix since the DM can provide the correct one. One of the "main jobs" of DM is to create that matrix so we really don't intend people to use a DM but then create the matrix themselves.

Agreed, and that's what I expected when I passed in NULL args instead
of user matrix. But the DM created matrix errors out during Duplicate
while user provided matrix passes through, even though both of them
are unassembled. And this is what is confusing.

Vijay

On Wed, Sep 18, 2013 at 10:46 AM, Barry Smith <bsmith at mcs.anl.gov> wrote:
>
>    The code did not "crash", it ended with an error message
>
>
> On Sep 18, 2013, at 10:32 AM, "Vijay S. Mahadevan" <vijay.m at gmail.com> wrote:
>
>> All,
>>
>> When I call DMSetMatrixPreallocateOnly
>
>     Why are you setting this flag?  If you don't set it then the code will go through.
>
>> and then TSSetRHSJacobian
>> without a matrix input argument (NULL instead), then the code crashes
>> during TSSolve and more specifically during MatDuplicate. This is
>> replicable with src/ts/examples/tutorials/ex2.c. Here's the error with
>
>    Jed is reworking the logic here so the code will change.
>
>
>> stack trace.
>>
>> [0]PETSC ERROR: --------------------- Error Message
>>
>>
>> I hope this is not supposed to happen. The fix seems to be to call
>> TSSetRHSJacobian with a user created matrix instead of letting DM
>> create one,
>
>    There is no reason to use a user created matrix since the DM can provide the correct one. One of the "main jobs" of DM is to create that matrix so we really don't intend people to use a DM but then create the matrix themselves.
>
>    Barry
>
>
>> which rectifies this issue. This behavior feels
>> inconsistent and is there a way to fix this.
>>
>> Vijay
>>
>> PS: Just replace the following lines in ex2.c to replicate the error.
>>
>> /*  ierr = TSSetRHSJacobian(ts,A,A,RHSJacobian,&appctx);CHKERRQ(ierr); */
>>  ierr = TSSetRHSJacobian(ts,NULL,NULL,RHSJacobian,&appctx);CHKERRQ(ierr);
>>  ierr = DMSetMatrixPreallocateOnly(appctx.da,PETSC_TRUE);CHKERRQ(ierr);
>