[petsc-users] Suggestion for KSPRichardson

[Jed Brown]

Thanks, I've applied this to petsc-dev.


On Wed, Jan 2, 2013 at 3:57 PM, Gaetan Kenway <kenway at utias.utoronto.ca>wrote:

> Hi Everyone
>
> I have a small suggestion for the KSPRichardson implementation.  I often
> use a KSPRichardson object to perform a small number (2-4) applications of
> a preconditioner (Additive Schwartz for example). Since I'm using a fixed
> number of iterations, the object uses KSP_NORM_NONE to save the vecNorm.
>
> Near the end of the function KSPSolve_Richardson, (line 138 according to
> the current docs, link below), the new residual is evaluated. However, for
> the last iteration, when using KSP_NORM_NONE, the last matMult and VecAYPX
> is not necessary. This results in additional unnecessary matMults(). In one
> system I'm solving, this adds approximately 20% to the computation cost.
>  (30 secs -> 36 secs)
>
> I think the following code should work:
>  if (ksp->its < maxit  || ksp->normtype != KSP_NORM_NONE){
>         ierr = KSP_MatMult(ksp,Amat,x,r);CHKERRQ(ierr);      /*   r  <- b
> - Ax      */
>         ierr = VecAYPX(r,-1.0,b);CHKERRQ(ierr);
>       }
>
> Also, it makes the degenerate case of 1 iteration have same cost as if the
> KSPRichardson wasn't there.
>
> Hopefully this is useful.
>
> Gaetan Kenway
>
> Link to source code:
>
> http://www.mcs.anl.gov/petsc/petsc-current/src/ksp/ksp/impls/rich/rich.c.html#KSPRICHARDSON
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-users/attachments/20130102/6a6a590d/attachment.html>