[petsc-users] MatSetValues on same column multiple times

[John Mousel]

To be more clear, I do the following:

<preallocate>
<insert 0's where I want columns>
<flush>
<insert real entries>
<flush>
<clean up special rows by adding corrections to coefficients>

What I want to clarify is if I touch the same column more than once, say
I'm looping over neighbors in an octree mesh, and it's just easier to
insert them multiple times as I find them for fluxes in different
directions instead of finding one exact coefficient for each neighbor, how
much preallocation would I need to specify? The number of times it gets
touched, or just one?

John


On Tue, Sep 25, 2012 at 4:50 PM, Matthew Knepley <knepley at gmail.com> wrote:

> On Tue, Sep 25, 2012 at 5:44 PM, John Mousel <john.mousel at gmail.com>wrote:
>
>>
>> If I'm adding to a particular column in multiple passes, do I need to
>> preallocate enough memory for each pass, or just 1 for each column in the
>> row?
>> Right now I'm drastically over-allocating, and I'm trying to reduce this.
>> If I give the true number of columns in the row, I keep getting new entry
>> errors.
>>
>> New nonzero at (8045,60) caused a malloc!
>>
>
> You have to be more specific. It sounds like what you are doing is
>
> <preallocate>
> <put in some nonzeros>
> <assemble>
> <try to put in more nonzeros>
>
> This will not work since assembly squeezes out extra space. Why are you
> assembling in between?
> If you need to clear buffers, you can just to MAT_ASSEMBLY_FLUSH.
>
>    Matt
>
>
>> I have it spit into phases where I set zeros in the non-zero pattern,
>> then I come back and fill in entries in a bulk pass, and then do a second
>> pass to clean up entries in special rows.
>> Could someone clarify how this works.
>>
>> John
>>
>
>
>
> --
> What most experimenters take for granted before they begin their
> experiments is infinitely more interesting than any results to which their
> experiments lead.
> -- Norbert Wiener
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-users/attachments/20120925/a7726fc4/attachment.html>