dolfin team mailing list archive

["Garth N. Wells" <gnw20@xxxxxxxxx>]

Anders Logg wrote:
> On Thu, Sep 25, 2008 at 05:20:53PM +0200, Anders Logg wrote:
> > On Thu, Sep 25, 2008 at 04:18:23PM +0100, Garth N. Wells wrote:
> > > Anders Logg wrote:
> > > > On Thu, Sep 25, 2008 at 03:53:01PM +0100, Garth N. Wells wrote:
> > > > > Anders Logg wrote:
> > > > > > On Thu, Sep 25, 2008 at 03:11:27PM +0100, Garth N. Wells wrote:
> > > > > > > Anders Logg wrote:
> > > > > > > > On Thu, Sep 25, 2008 at 12:41:33PM +0100, Garth N. Wells wrote:
> > > > > > > > > Nuno David Lopes wrote:
> > > > > > > > > > Is there a simple way of setting an initial guess for an Iterative 
> > > > > > > > > > LinearSolver? 
> > > > > > > > > > In Umfpack and PETSc the default initial guess is the zero vector right?
> > > > > > > > > At the moment, yes (note the UMFPACK is an LU solver, so an initial 
> > > > > > > > > guess doesn't do anything).
> > > > > > > > >
> > > > > > > > > It's very simple, and I've been meaning to add an option for using an 
> > > > > > > > > initial guess. It's also useful for Newton solvers. I'll add something 
> > > > > > > > > in the next few days.
> > > > > > > > >
> > > > > > > > > Garth
> > > > > > > > It would be natural to let the x argument always be the initial
> > > > > > > > guess. I thought we already did this.
> > > > > > > >
> > > > > > > > Would it be enough to make sure that Vector::init() does not reset the
> > > > > > > > values to zero?
> > > > > > > Yes.
> > > > > > >
> > > > > > > The danger is if someone sends an uninitialised vector to the solver.
> > > > > > >
> > > > > > > Garth
> > > > > > We can just put something like this in the init() functions:
> > > > > >
> > > > > >   if (x && size() == N)
> > > > > >     return;
> > > > > >
> > > > > > I think we had this a while back but at some point VecZeroEntries was
> > > > > > inserted.
> > > > > I think that we should have an option whether or not to use an initial 
> > > > > guess. The default can be to use the guess (as in your above code extract).
> > > > >
> > > > > Garth
> > > > I think it would work nicely if the solver calls x.init() and that
> > > > will initialize x to a zero vector only if it has not already been
> > > > initialized. 
> > > Sounds a bit dangerous to me that Vector::init() sometimes zeroes a the 
> > > vector, and sometimes it doesn't. Should we get rid of Vector::init() 
> > > and use Vector::resize() and Vector::zero() instead? No ambiguity then.
> > >
> > > Garth
> > That sounds good.
>
> On second thought, we shouldn't mess with the init() functions unless
> we want to change the meaning of init() for all the classes deriving
> from GenericTensor, including also matrices for all backends. Also,
> resize() sometimes means that existing values will be preserved when
> the size is increased.

I've already started on it. There are a lot of places in the code where 
we only need to resize and not zero, so I like using resize() and 
zero(). We even had places with

   x.init(N);
   x.zero();

There is no ambiguity now with resize and zero.


> So the best solution is to do something like
>
>   if (x.size() != N)
>     x.init();

This is what I've done.

Garth

> in the Krylov solvers instead of just init().
>
>
>
> ------------------------------------------------------------------------
>
> _______________________________________________
> DOLFIN-dev mailing list
> DOLFIN-dev@xxxxxxxxxx
> http://www.fenics.org/mailman/listinfo/dolfin-dev