dolfin team mailing list archive

[Anders Logg <logg@xxxxxxxxx>]

On Mon, Feb 13, 2006 at 05:06:50PM +0100, Johan Jansson wrote:
> On Mon, Feb 13, 2006 at 09:39:14AM -0600, Anders Logg wrote:
> > On Mon, Feb 13, 2006 at 02:46:04PM +0100, Johan Jansson wrote:
> > 
> > > I think two remaining features to implement are:
> > > 
> > > 1. Loading Functions (DiscreteFunction) from file.
> > 
> > ok, which format do we want to do initially? DOLFIN XML?
> > 
> > Maybe we should also consider some binary format that is quicker to
> > write and parse? Could be useful mostly during computations (and less
> > useful for storing solutions for future reference) for example when
> > solving the dual problem and linearizing around a previously computed
> > solution (or time sequence of solutions) that doesn't fit in memory.
> 
> Isn't a natural candidate the VTK format? This is the standard output
> format for DOLFIN, isn't it natural to be able to read it back in?

The output is currently limited to values at vertices. Do we want to
support reading and writing arbitrary functions? Since we define
discrete Functions as triples (vector, mesh, element), we would have
to read and write all three. It's no problem reading and writing the
vector and the mesh since it's all data, but the element is more code
than data (dofmap(), vertexeval() etc), so it's more difficult to
store to a file.

Perhaps we could precompile a number of basic element types (we
already have linears precompiled) and the element type would only be a
string?

One way would be to support the VTK format for linears (values at
vertices) and have a more general DOLFIN XML format that supports all
types of elements.

> > > 2. An interface for projecting/interpolating between Functions. The
> > > main example which requires some manual work in solvers is defining
> > > the initial values by a UserFunction and wanting to represent that as
> > > a DiscreteFunction.
> > 
> > I'm not sure what you mean. I thought I solved this with the new
> > functionality I added to the Function class last week. You can define
> > a Function for the initial value by overloading eval(), or even by
> > 
> > Function u = 0.0;
> > 
> > and then u can change from being a user-defined Function to a discrete
> > Function. See new convection-diffusion demo.
> > 
> > You can also do
> > 
> > Function u(x, mesh, element);
> > u = f;
> > 
> > to interpolate f to u.
> > 
> 
> Ok, great! I'm sorry, I didn't realize this was what you added last
> week.
> 
> I think interpolation is enough, projection can wait.

ok, good.

> > I have limited time to work on new features before Geilo, but I can
> > probably fit something in if I get some help.
> > 
> 
> Yes, I also have limited time. But maybe we can at least start looking
> at Function input and decide what we should do, i.e. if we should
> design a DOLFIN XML data format for Functions or use the VTK format.

ok. We just need to decide how general we want to be.

/Anders