matshell

[Romain Veltz]

Dear users,

I am using petsc4py to solve odes. I am trying sundial or arkimex but the solution explodes whereas it should be bounded (everything is fine with ROSW). I think there is maybe an issue with the jacobian ( snes.setUseMF(True) , precision... ).

I am considering the use of matshell because I know how to compute J.x (J is dense and large, millions^2 of entries ). I can't find simple examples for this in petsc4py. Would it be possible to do something like the old code:

https://code.google.com/p/petsc4py/issues/detail?id=11

Thank you for your help,

Bests,

Romain

[Lisandro Dalcin]

See this code, that's the way to define a "shell" matrix in petsc4py:

https://bitbucket.org/petsc/petsc4py/src/master/demo/poisson2d/poisson2d.py

[Romain Veltz]

Thank you for the link, this is what I needed.

Let me ask another question please (I googled it but it not easy to find).

I want to make the computation in def mult(…): y = J(u).x where u is the current solution.

I want to use this with SNES but I don't know how pass u to the function def mult

Can you provide me with this last information please?

Thank you

[Lisandro Dalcin]

I would suggest to add a method to your Mat class, let's call it "def computeJacobian(self, snes, u, *args)", in such a way that you can pass it to SNES.setJacobian(). That method should store the values of the "u" vector (you should make a copy) to be used later in "mult()".

PS: This bug tracker is not a forum to ask questions, next time please email petsc-users or my personal mail account.

[Lisandro Dalcin]

Not an issue