dolfin team mailing list archive

[Anders Logg <logg@xxxxxxxxx>]

On Sun, Dec 16, 2007 at 12:52:38PM +0100, Ola Skavhaug wrote:
> Anders Logg skrev den 16/12-2007 følgende:
> > On Sun, Dec 16, 2007 at 09:46:56AM +0000, Garth N. Wells wrote:
> > > 
> > > 
> > > Gideon Simpson wrote:
> > > > I'm interested in implementing the Elman et al. preconditioner for 
> > > > Stokes in the fenics suite.  This involves building an approximate schur 
> > > > complement for the block strucutre matrix
> > > > A B^T
> > > > B 0
> > > > representing the weak form of the stokes operator.  has anyone 
> > > > experimented, if not with this sort, some sort of block structure using 
> > > > the fenics front end?
> > > > 
> > > >
> > > 
> > > No, when using the default uBLAS backend, there aren't any sophisticated 
> > > preconditioners. You can create a preconditioner which derives from 
> > > uBLASPreconditioner.
> > > 
> > > You can also look at using PETSc which provides a variety of 
> > > preconditioners.
> > > 
> > > Garth
> > 
> > Since this is quite common, maybe we should consider adding a some
> > abstractions for block systems/preconditioners: BlockMatrix,
> > BlockPreconditioner?
> > 
> > PyCC has something like this (but only in Python) that we can look at.
> > Ola knows more about it.
> 
> It would be quite simple to port the block matrices in PyCC to C++. It will
> be placed a bit down on my todo list, however.
> 
> We have two types of block-matrices:
> 
> BlockMatrix
> DiagBlockMatrix
> 
> Additionally, we have a zero-matrix, since block systems often contain zeros,
> and we want to make sure the matrix dimensions in the block system are
> correct.
> 
> Ola

ok, let's try to get this going at some point (but my todo-list is
also expanding).

-- 
Anders