[petsc-users] Bug when multipling a SBAIJ matrix with a vector

[Barry Smith]

   I've asked Hong and Shri to fix the bug I introduced that sometimes makes it require a diagonal. Once the bug is fixed 3.1 will allow empty diagonals again.

   Barry

On Oct 21, 2010, at 8:08 AM, Jed Brown wrote:

> On Thu, Oct 21, 2010 at 14:41, Andreas Hauffe
> <andreas.hauffe at tu-dresden.de> wrote:
>> I get no error and this example delivers the right result for PETSC 3.0. What
>> did change from 3.0 to 3.1?
> 
> Many of the matrix kernels were optimized, this one was changed in
> relatively simple ways, but the data structures for factorization were
> changed completely.  In this case, Barry's name is on the commit
> message
> 
> http://petsc.cs.iit.edu/petsc/petsc-dev/rev/0308bd570415#l2.33
> 
> He should have put that behavioral change in the 3.1 changelog.  Or
> maybe the cost of handling empty rows is actually not significant so
> it should still be supported.  I can't tell from the commit message,
> but empty rows are still supported by the code in MatMult_SeqSBAIJ_2,
> so it probably should work for block size 1 as well.
> 
> Do you really not get an error message when you run your example code
> with 3.1 built with debug support?  I just ran your code with 3.1 and
> I get a (bad) error message.
> 
> Jed