dolfin team mailing list archive

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

Anders Logg wrote:
> I have made a class diagram for the linear algebra classes, which can
> be found either in doc/manual/{dia,eps}/class-diagram-la.{dia,eps} or 
> at
>
>     http://home.simula.no/~logg/tmp/class-diagram.eps

Something wrong with this link.

> > From this diagram, it is clear there are some inconsistencies. Here
> are some suggestions:
>
> 1. Rename PETScSparseMatrix --> PETScMatrix since there is only one
> PETSc matrix.
>
> 2. Rename DenseVector --> uBlasVector. DenseVector can be a typedef.
>
> 3. Break up LinearSolver in two classes: PETScLinearSolver and
> uBlasLinearSolver and make both pure virtual interfaces.
>
> 4. Rename PETScLU --> PETScLUSolver to make it consistent.
>
> 5. Rename VirtualMatrix --> PETScKrylovMatrix to make it consistent
> with uBlasKrylovMatrix.
>
> 6. Rename Preconditioner --> PETScPreconditioner to make it consistent
> with uBlasPreconditioner.
>
> 7. Typedefs for common default types should include the following:
>
>    Matrix
>    Vector
>
>    DenseMatrix
>    DenseVector    (something that works with DenseMatrix)
>
>    SparseMatrix
>    SparseVector   (something that works with SparseVector)
>
>    LinearSolver
>    KrylovSolver
>    GMRES
>    LU
>
>    VirtualMatrix  (something that works with KrylovSolver)
>    Preconditioner (something that works with KrylovSolver)

All looks good. Can add

  Rename EigenvalueSolver --> PETScEigenvalueSolver.

> Here's a little list of things I think we should fix before releasing 0.6.2:
>
>  - Fix the above cleanups
>
>  - Make sure the linear algebra works nicely in PyDOLFIN
>
>  - Update the class diagram
>
>  - Complete the linear algebra chapter in the manual
>
>  - Add a linear algebra benchmark in src/bench
>
>  - Add a linear algebra test in src/test
>
>  - Add a linear algebra demo in src/demo

We should also get everything compiling working with the --disable-petsc 
flag (some demos don't compile at the moment) and remove PETSc-specific 
operations if they're not necessary.

Garth

> It would be great if we could all help out so everyone agrees on and
> knows the design of the linear algebra library. We would then have a
> well functioning and stable linear algebra library that would require
> little maintenance so we can focus on other things (like porting to
> the new mesh library).
>
> /Anders
> _______________________________________________
> DOLFIN-dev mailing list
> DOLFIN-dev@xxxxxxxxxx
> http://www.fenics.org/mailman/listinfo/dolfin-dev