dolfin team mailing list archive

[Anders Logg <logg@xxxxxxxxx>]

On Sat, Dec 06, 2008 at 04:03:25PM +0100, Johan Hake wrote:
> On Saturday 06 December 2008 15:43:13 Anders Logg wrote:
> > On Sat, Dec 06, 2008 at 03:34:40PM +0100, Johan Hake wrote:
> > > On Saturday 06 December 2008 14:56:59 Anders Logg wrote:
> > > > On Sat, Dec 06, 2008 at 12:05:41PM +0100, DOLFIN wrote:
> > > > > One or more new changesets pushed to the primary dolfin repository.
> > > > > A short summary of the last three changesets is included below.
> > > > >
> > > > > changeset:   5262:0e349fbe09ce4179252652fe5c1d58725951bc3f
> > > > > tag:         tip
> > > > > user:        "Johan Hake <hake@xxxxxxxxx>"
> > > > > date:        Sat Dec 06 12:05:42 2008 +0100
> > > > > files:       demo/pde/stokes/taylor-hood/python/demo.py
> > > > > dolfin/function/SpecialFunctions.h dolfin/swig/dolfin_function_pre.i
> > > > > site-packages/dolfin/function.py description:
> > > > > Added support for f.split()
> > > > >  - Renamed operator[] to f._sub instead of f.sub
> > > > >  - f.sub(i) now returns an instantiated sub function
> > > > >  - f.split() uses f.sub() to return a tuple of all sub functions
> > > > >  - stoke/taylor-hood demo now runs.
> > > >
> > > > Excellent!
> > > >
> > > > It runs now but the solution looks completely crazy. The problem is
> > > > that the boundary conditions are not set correctly since we use V and
> > > > Q to set the boundary conditions for the sub systems and they don't
> > > > know the offsets (DofMap::offset()).
> > >
> > > The problem here is that these spaces need to be SubSpaces?
> >
> > Yes.
> >
> > > If that is the case, we could extract the subspaces after a
> > > MixedFunctionSpace is created and then store these in the spaces
> > > attribute, either as pure cpp.SubSpaces or add another python class,
> > > SubSpace, which is a cpp.SubSpace and stores the original ffc.element
> > > too, or something?
> >
> > I was thinking something like this:
> >
> > W = VectorFunctionSpace(mesh, "triangle", 2) + FunctionSpace(mesh,
> > "triangle", 1) V, Q = W.split()
> 
> Doesn't look too nice, and it will create a non intuitive work flow. Combining 
> two interfaces aren't easy!
> 
> > The split function needs to both create SubSpaces (which will lead to
> > DofMaps with correct offsets) and set the element correctly, which can
> > be done by looking at the spaces attribute in MixedFunctionSpace.
> >
> > It's a bit weird since what we really do is
> >
> > V = VectorFunctionSpace(mesh, "triangle", 2)
> > Q = FunctionSpace(mesh, "triangle", 1)
> > W = V + Q
> > V, Q = W.split()
> >
> > First, we put V and Q together to create a mixed function space. Then
> > we split W again into V and Q. After the split V and Q know that they
> > are part of the bigger space (at least they know the offset into the
> > bigger space).
> 
> Yes I see this. Instead, or in addition, of using split, we could implement 
> W.sub(i) that returns a subspace, which is then used when setting the bc, and 
> other relevant places.
> 
> > The other option would be to let + have a side effect on V and Q but
> > that does not seem to be a good solution.
> 
> I thought of this too. But I cannot se how we could do it with the present 
> implementation of FunctionSpace/Subspace. 

If we can get make_subspace working as you suggest below, then it
would be easy to modify __add__ for FunctionSpaceBase:

  W = MixedFunctionSpace([self, other])
  self.make_subspace(W, 0)
  other.make_subspace(W, 1)
  return W

This would make the interface look nice but someone (you know who you
are) might think we are insane... :-)

> If we could just call a member function in a FunctionSpace, let say 
> V.make_subspace(W,i), and this would then create the needed stuff in V, to 
> make it SubSpaceable. But now SubSpace is a subclass which effectively 
> prevents this approach.
> 
> This approach is abit intrucive too. But what difference would it make for the 
> user, i.e., if V in addition to be a FunctionSpace now also is a subspace of 
> W?

The only reason we need this is to be able to set boundary conditions
for sub systems. For that we need two things: the offset into the
global system vector and the DofMap for the subsystem in question.
There might be other ways to define the interface for DirichletBC,
for example:

  bc = DirichletBC(W, 0, ...)

-- 
Anders

Attachment: signature.asc
Description: Digital signature