dolfin team mailing list archive

["Garth N. Wells" <g.n.wells@xxxxxxxxxx>]

On Mon, 2006-07-03 at 17:31 +0200, Anders Logg wrote:
> I've extended the uBlas Krylov solver so that it now works with
> matrices of type uBlasKrylovMatrix, only defined by having two
> functions size() and mult(). I had to make some minor changes to avoid
> using axpy_prod() but the tests I made showed now reduction in
> speed. (There is one extra copy of values into a vector compared to
> the original code.)
> 

Make sure that you use noalias() whenever possible when working with
uBlas data types. In the code, r += b should be noalias(r) += b or
r.plus_assign(b). Otherwise, a copy of r is make internally which gets
expensive for large vectors. 
  

> I have temporarily disabled the preconditioning. See all comments
> 
>     // FIXME: Preconditioner temporarily disabled
> 
> This was necessary since uBlasPreconditioner assumes uBlasSparseMatrix.
> We need to make it so that it only assumes uBlasKrylovMatrix. Then we
> can have a special subclass, like
> 
>     class uBlasILUPreconditioner : public uBlasPreconditioner
> 
> or similarly. This class should then take in the constructor as
> argument a uBlasSparseMatrix. Then we can overload solve()
> uBlasKrylovSolver so we have two versions: one with and one without an
> extra preconditioner argument.
> 
> I don't think it's possible to let the Krylov solver create a
> preconditioner on its own since it only knows how to do mult(), that
> is, it only knows that the matrix is uBlasKrylovMatrix. So it can't
> create a preconditioner that assumes it can do M.assign(A).
> 
> Does that make sense?
> 

Sounds ok. The preconditioner interface does need to be sorted out. My
first implementation was very simple just to get something working.

Garth  

> /Anders
> _______________________________________________
> DOLFIN-dev mailing list
> DOLFIN-dev@xxxxxxxxxx
> http://www.fenics.org/mailman/listinfo/dolfin-dev