[petsc-dev] Problem mixing sieve and TS

[Jed Brown]

On Thu, Oct 27, 2011 at 00:16, Paul Sicsic <paul.sicsic at polytechnique.edu>wrote:

> The X you passed in is not a valid Mat (we could see that from the earlier
>> call). More specifically, according to valgrind, it is uninitialized memory.
>> The shift is still infinite coming into this function, which I don't
>> understand.
>>
>> Why do you say the X matrix is not valid ?
> If I run the program with -mat_view, our Mass (AppCtx%M) and rigidity
> matrix (AppCTx%K) are assembled.
>

The Mat that makes it into MatAXPY is not valid. What does your code look
like that makes that call?


>
>
>> From the singularity at 0, it looks like perhaps the initial time step
>> size is being set to zero. Next time you get a stack trace, can you on
>> ts->time_step?
>>
>
Yes, it should.


>
> tssolve_ (ts=0x7fff42d866e0, x=0x7fff42d86388, ftime=0x7fff42d866f0,
> __ierr=0x7fff42d85adc) at
> /opt/HPC/petsc-dev/src/ts/interface/ftn-auto/tsf.c:397
> 397             (Vec)PetscToPointer((x) ),ftime);
> (gdb) p ts->time_step
> $1 = 6.9531990170672068e-310
>
> Should ts->time_step be inizialized entering TSSolve ?
>
> We had called (As in example ex22f) :
>      Call TSSetInitialTimeStep(AppCtx%TS, 0.0, .01,  ierr); CHKERRQ(iErr)
>      Call TSSetFromOptions(AppCtx%TS,  iErr); CHKERRQ(iErr)
>

This is one reason why I HATE, HATE, HATE Fortran. When you pass the
literals 0.0 and .01, they are passed as single instead of double precision.
You need to define variables to hold these values:

PetscReal zero,dt

zero = 0.0
dt = .01
call TSSetInitialTimeStep(AppCtx%TS, zero, dt, ierr)
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-dev/attachments/20111027/7dad600d/attachment.html>