[petsc-users] reuse an assembled matrix when a linear problem is solved with KSP and DM

[Barry Smith]

On Sep 17, 2013, at 8:35 PM, Roc Wang <pengxwang at hotmail.com> wrote:

> Hello,
> 
>     I am trying to solve a linear problem by using KSPSolve and DM. In the example  /src/ksp/ksp/examples/tutorials/ex45.c, the Matrix and the RHS vector are computed by calling  KSPSetComputeRHS and  KSPSetComputeOperators in which functions ComputeRHS() and ComputeMatrix() are called and the matrix is assembled in ComputeMatrix().   
> 
>     Since the linear system is solved in an iteration, it will save time to reuse the matrix assembled in ComputeMatrix().  That is the ComputeMatrix() is called only once, but ComputeRHS() is called in each iteration.

   What do you need mean by an "iteration" here? Do you mean one iteration of a KSP solvers, of which several are needed to produce the approximate solution or something else? 

   The ComputeRHS() function is only called ONCE for each KSPSolve (even though KSP may take several iterations to solve the linear system).

   The ComputeMatrix() function is actually only called ONCE period, unless you call KSPSetComputeOperators() again, even if you call KSPSolve() several times it uses the same matrix until you call KSPSetComputerOperators() again.

>  Can anyone tell me what should I do to keep the assembled matrix?
> 
>      I am thinking not to use  KSPSetComputeOperators for assembling the matrix, but use DMCreateMatrix() alternatively. Am I right?

   That is a completely orthogonal issue; DMCreateMatrix() only gives you an empty sparse matrix with the right nonzero structure for the mesh associated with DM. You still need to compute its entries yourself, possibly in a ComputeMatrix function you write.

   Barry

> 
>      Any hits would be appreciated very much. Thanks alot.
> 
>