[petsc-users] MatZeroEntries for retaining allocated space

[Junchao Zhang]

On Tue, Apr 12, 2022 at 3:51 PM Samuel Estes <samuelestes91 at gmail.com>
wrote:

> Ah, ok thank you. So let me just make sure I completely understand. When
> using the MatPreallocatorPreallocate() routine, since the exact sparsity
> pattern is specified (as opposed to just specifying the number of nnz per
> row), there is no need to put in extra zeros to maintain the allocated
> space because the routine does that for me?
>
Yes, as Mark said. The zeroing code is at
https://gitlab.com/petsc/petsc/-/blob/main/src/mat/impls/preallocator/matpreallocator.c#L186

 But please note that you need to call  MatPreallocatorPreallocate(Mat mat,
PetscBool fill, Mat A) with fill=PETSC_TRUE to zero A. See manual at
https://petsc.org/main/docs/manualpages/Mat/MatPreallocatorPreallocate.html



> On Tue, Apr 12, 2022 at 3:46 PM Junchao Zhang <junchao.zhang at gmail.com>
> wrote:
>
>>
>>
>> On Tue, Apr 12, 2022 at 12:23 PM Samuel Estes <samuelestes91 at gmail.com>
>> wrote:
>>
>>> Maybe I should create a separate question for this but I did some
>>> testing and now I'm a bit confused on how PETSc handles preallocated space
>>> which is not assigned a value. I created a test case where I allocated a
>>> certain amount of space for a matrix. I then only assigned a value to the
>>> (0,0) entry and assembled the matrix. When I look at the MatInfo it says
>>> that there were 0 unneeded entries. Furthermore, when I use MatView, I see
>>> that there are 0 values in the matrix so there actually were values which
>>> were put there.
>>>
>>> I have a guess at what's happening and if I'm correct then it would
>>> solve my original problem but I would like to confirm here. I am using the
>>> MatPreallocator class along with the MatPreallocatorPreallocate() routine
>>> to allocate space for the matrix. My guess is that this routine
>>> automatically assigns a value of zero to the entries which I specified in
>>> the preallocator. So if I had done the original MatXXXAIJSetPreallocation()
>>> routines which just allocate a certain number of non-zeros to the rows,
>>> then I would have to explicitly assign them zero values or else I would
>>> lose unassigned space in the assembly routines. Since the MatPreallocator
>>> specifies exactly which entries will be non-zero, PETSc also actually
>>> assigns a zero value under the hood so that I don't have to explicitly
>>> assign values myself. Is this correct?
>>>
>> I think your understanding is correct.
>> I think the rule is: if you have not specified the exact sparsity pattern
>> in preallocation (for example, you only said this row had this number of
>> nonzeros, but did not give the column indices), then in MatAssebmlyEnd(),
>> unused slots will be compressed out.
>> You can call MatZeroEntries() on assembled or unassembled (but
>> preallocated) matrices. The routine just zeros the internal value array.
>> It does not help you to preserve nonzero entries.
>> With MatPreallocatorPreallocate(), the sparsity is specified, so that
>> petsc will keep the entries.
>>
>>
>>>
>>> Sorry for the convoluted question. I hope it's clear.
>>>
>>> On Tue, Apr 12, 2022 at 11:54 AM Samuel Estes <samuelestes91 at gmail.com>
>>> wrote:
>>>
>>>> Hi,
>>>>
>>>> I have a problem where I want to allocate some nonzero structure to a
>>>> sparse matrix which will be used repeatedly to solve some equations. The
>>>> assembly routines will compress out allocated space which has not been
>>>> assigned a value so I know that I need to assign zero values to any unused
>>>> entries in order to keep that space. Will MatZeroEntries do this? If not, I
>>>> could always just loop over and explicitly assign zero values but
>>>> MatZeroEntries would be more convenient. Thanks!
>>>>
>>>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-users/attachments/20220412/811751d5/attachment-0001.html>