[petsc-users] [petsc-3.2] LSQR new convergence test

[Alexander Grayver]

Barry,

I think the error is related to this

PetscFortranAddr cctx

definition. Speaking in terms of C++ notation, it seems that PETSc takes 
&cctx as KSPDefaultConvergedCtx, however real KSPDefaultConvergedCtx is 
*cctx.
Not sure if I'm clear now, but it's easy to see if you trace code like that:

   PetscFortranAddr      cctx
   KSP  ksp

   call KSPCreate(MPI_COMM_WORLD,ksp,ierr)
   call KSPSetType(ksp,KSPLSQR,ierr)
   call KSPDefaultConvergedCreate(cctx,ierr)
   call 
KSPSetConvergenceTest(ksp,KSPDefaultConverged,cctx,PETSC_NULL_FUNCTION,ierr)
   call KSPDestroy(ksp,ierr)

I guess the error should be reproducible from the code above.

Regards,
Alexander

On 21.09.2011 17:17, Alexander Grayver wrote:
> Thanks Barry, I added your code and solver works well now, but when I 
> call KSPDestroy afterward it crashes:
>
> [8]PETSC ERROR: --------------------- Error Message 
> ------------------------------------
> [8]PETSC ERROR: Invalid argument!
> [8]PETSC ERROR: Wrong type of object: Parameter # 1!
> [8]PETSC ERROR: 
> ------------------------------------------------------------------------
> [8]PETSC ERROR: Petsc Release Version 3.2.0, Patch 2, Fri Sep 16 
> 10:10:45 CDT 2011
> [8]PETSC ERROR: See docs/changes/index.html for recent updates.
> [8]PETSC ERROR: See docs/faq.html for hints about trouble shooting.
> [8]PETSC ERROR: See docs/index.html for manual pages.
> [8]PETSC ERROR: 
> ------------------------------------------------------------------------
> [8]PETSC ERROR: /home/test on a openmpi-i named node226 by user Wed 
> Sep 21 17:01:23 2011
> [8]PETSC ERROR: Libraries linked from 
> /home/lib/petsc-3.2-p2/openmpi-intel-complex-debug-f/lib
> [8]PETSC ERROR: Configure run at Tue Sep 20 09:28:29 2011
> [8]PETSC ERROR: Configure options 
> --with-petsc-arch=openmpi-intel-complex-debug-f 
> --with-fortran-interfaces=1 
> --with-mpi-dir=/opt/mpi/intel/openmpi-1.4.2 --with-scalar-type=complex 
> --with-blas-lapack-dir=/opt/intel/Compiler/11.1/072/mkl/lib/em64t 
> --with-precision=double --with-x=0
> [8]PETSC ERROR: 
> ------------------------------------------------------------------------
> [8]PETSC ERROR: VecDestroy() line 566 in 
> /home/lib/petsc-3.2-p2/src/vec/vec/interface/vector.c
> [8]PETSC ERROR: KSPDefaultConvergedDestroy() line 669 in 
> /home/lib/petsc-3.2-p2/src/ksp/ksp/interface/iterativ.c
> [8]PETSC ERROR: KSPDestroy() line 758 in 
> /home/lib/petsc-3.2-p2/src/ksp/ksp/interface/itfunc.c
>
> Is it my fault?
>
> Regards,
> Alexander
>
> On 21.09.2011 15:36, Barry Smith wrote:
>> On Sep 21, 2011, at 2:39 AM, Alexander Grayver wrote:
>>
>>> Hong, Barry,
>>>
>>> Thanks! Sorry for not to be clear, but Barry right, my question was 
>>> how to get back to DefaultConvergenceTest in case of LSQR solver.
>>>
>>> Barry, yes, both versions give the same two norms.
>>>
>>> Can you clarify please how to implement that in Fortran:
>>> void *cctx;
>>> KSPDefaultConvergedCreate(&cctx);
>>> KSPSetConvergenceTest(ksp,KSPDefaultConverged,cctx);
>>     It should be almost the same
>>
>>      PetscFortranAddr cctx
>>
>>      call KSPDefaultConvergedCreate(cctx,ierr)
>>      call 
>> KSPSetConvergenceTest(ksp,KSPDefaultConverged,cctx,PETSC_NULL_FUNCTION,ierr)
>>
>>     Barry
>>
>>
>>> I'm a little bit confused about cctx parameter.
>>>
>>> Regards,
>>> Alexander
>>>
>>>
>>> On 21.09.2011 01:10, Barry Smith wrote:
>>>> It has its own monitor that provides additional information  
>>>> -ksp_monitor_lsqr
>>>>
>>>> You can also remove the new convergence test and get back the old 
>>>> one with code like
>>>>
>>>> void *cctx;
>>>> KSPDefaultConvergedCreate(&cctx);
>>>> KSPSetConvergenceTest(ksp,KSPDefaultConverged,cctx);
>>>>
>>>> after the KSPType is set to LSQR. So if you are happy with the old 
>>>> test.
>>>>
>>>>
>>>>    Do both versions give the same first two norms?
>>>>
>>>>> 0 KSP Residual norm 9.386670021557e-17
>>>>>   1 KSP Residual norm 8.258308650175e-17
>>>>    Barry
>>>>
>>>>
>>>> On Sep 20, 2011, at 4:40 AM, Alexander Grayver wrote:
>>>>
>>>>> Hello,
>>>>>
>>>>> In comparison with petsc-3.1-p7 in the latest petsc LSQR solver 
>>>>> behaves differently.
>>>>>
>>>>> In petsc31 I had:
>>>>>
>>>>>   0 KSP Residual norm 9.386670021557e-17
>>>>> ...
>>>>> 95 KSP Residual norm 9.341367317075e-18
>>>>> Linear solve converged due to CONVERGED_RTOL iterations 95
>>>>> KSP Object:
>>>>>   type: lsqr
>>>>>   maximum iterations=200, initial guess is zero
>>>>>   tolerances:  relative=0.1, absolute=1e-50, divergence=10000
>>>>>   left preconditioning
>>>>>   using PRECONDITIONED norm type for convergence test
>>>>> PC Object:
>>>>>   type: none
>>>>>   linear system matrix = precond matrix:
>>>>>   Matrix Object:
>>>>>     type=shell, rows=19584, cols=19584
>>>>>
>>>>> relative residual norm =    0.9951737192872134E-01
>>>>>
>>>>> Now, with petsc32 I have:
>>>>>
>>>>>   0 KSP Residual norm 9.386670021557e-17
>>>>>   1 KSP Residual norm 8.258308650175e-17
>>>>> Linear solve converged due to CONVERGED_RTOL_NORMAL iterations 1
>>>>> KSP Object: 12 MPI processes
>>>>>   type: lsqr
>>>>>   maximum iterations=200, initial guess is zero
>>>>>   tolerances:  relative=0.1, absolute=1e-50, divergence=10000
>>>>>   left preconditioning
>>>>>   using UNPRECONDITIONED norm type for convergence test
>>>>> PC Object: 12 MPI processes
>>>>>   type: none
>>>>>   linear system matrix = precond matrix:
>>>>>   Matrix Object:   12 MPI processes
>>>>>     type: shell
>>>>>     rows=19584, cols=19584
>>>>> relative residual norm =    0.8797910900468064E+00
>>>>>
>>>>> So I found new convergence test here which sets 
>>>>> CONVERGED_RTOL_NORMAL:
>>>>> http://www.mcs.anl.gov/petsc/petsc-as/snapshots/petsc-current/docs/manualpages/KSP/KSPLSQRDefaultConverged.html 
>>>>>
>>>>>
>>>>> The question is, how to interpret this new test and make it works 
>>>>> properly for me?
>>>>> Thanks in advance.
>>>>>
>>>>> Regards,
>>>>> Alexander
>