[petsc-users] MatNest error with null matrix blocks

Randall Mackie rlmackie862 at gmail.com
Sun Jul 6 12:10:10 CDT 2025


Thanks Barry - this is a very helpful clarification.

Randy


> On Jul 6, 2025, at 9:59 AM, Barry Smith <bsmith at petsc.dev> wrote:
> 
>  
>    Yes, they are automatically initialized to something ((void*)-2 actually). They are NOT automatically initialized to PETSC_NULL_XXX
> 
>    #define PetscObjectIsNull(obj) (obj%v == 0 .or. obj%v == -2 .or. obj%v == -3)
> 
>    PETSc Fortran has multiple representations of NULL objects. When you want to pass in the meaning "I am a null" you pass  PETSC_NULL_XXX. Not the other representations.
> 
> 
>    Barry
> 
> Why? Well, if variables were initialized to 0 and then got passed into a routine, the routine would see the 0 and think: Oh, the person does not want to set this variable (because passing x would be the same as passing PETSC_NULL_XXX).
> 
> 
> 
> 
> 
> 
>> On Jul 4, 2025, at 9:59 PM, Randall Mackie <rlmackie862 at gmail.com> wrote:
>> 
>> Thanks Barry
>> 
>> We were confused because the online manual says:
>> 
>> PETSc objects are always automatically initialized when declared so you do not need to (and should not) do
>> 
>> type(tXXX) x = PETSC_NULL_XXX
>> XXX x = PETSC_NULL_XXX
>> 
>> Randy
>> 
>> 
>>> On Jul 4, 2025, at 6:56 PM, Barry Smith <bsmith at petsc.dev> wrote:
>>> 
>>> 
>>>    You need to pass PETSC_NULL_MAT in those locations, you cannot just pass a Mat that has never been created (that is a different kind of NULL matrix :-). I have added clearer error checking in https://urldefense.us/v3/__https://gitlab.com/petsc/petsc/-/merge_requests/8526__;!!G_uCfscf7eWS!cuhCwQPErEmmFpD3ik1xLziqtbh4SbXqQViT4PDY_DgSHIm0dkes0boQKPJnZkhSqaayUh78IUoqo7i2dMIPbibEsw$ 
>>> 
>>>   Barry
>>> 
>>> 
>>>> On Jul 4, 2025, at 10:32 AM, Randall Mackie <rlmackie862 at gmail.com> wrote:
>>>> 
>>>> In the process of upgrading our code to version 3.23, we have run into another issue, this time with nest matrices.
>>>> 
>>>> In previous versions (well to at least 3.21) it was perfectly fine to pass in null matrix blocks to create a nest matrix. There are situations where we have, for example, only diagonal blocks. That should be allowed and was and worked fine.
>>>> 
>>>> In petsc 3.23, that no longer is true, or perhaps we do not know what is the proper way to pass in a null matrix.
>>>> 
>>>> The attached example reproduces this with the following error:
>>>> 
>>>> [0]PETSC ERROR: Configure options: --force --with-clean=1 --with-scalar-type=complex --with-debugging=1 --with-fortran=1 --download-mpich=1 --with-cxx=0
>>>> [0]PETSC ERROR: #1 PetscObjectReference() at /home/rmackie/PETSc/petsc/src/sys/objects/inherit.c:620
>>>> [0]PETSC ERROR: #2 MatNestSetSubMats_Nest() at /home/rmackie/PETSc/petsc/src/mat/impls/nest/matnest.c:1407
>>>> [0]PETSC ERROR: #3 MatNestSetSubMats() at /home/rmackie/PETSc/petsc/src/mat/impls/nest/matnest.c:1518
>>>> [0]PETSC ERROR: #4 MatCreateNest() at /home/rmackie/PETSc/petsc/src/mat/impls/nest/matnest.c:1800
>>>> [0]PETSC ERROR: #5 matnest_bug_reproducer.F90:48
>>>> 
>>>> 
>>>> Thanks for helping with this, and thanks for all the work on the Fortran interfaces - I’ve been playing around with your new branch and am excited by all the work!
>>>> 
>>>> Randy
>>>> <matnest_bug_reproducer.F90>
>>> 
>> 
> 

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-users/attachments/20250706/e11f013c/attachment.html>


More information about the petsc-users mailing list