ffc team mailing list archive

[Anders Logg <logg@xxxxxxxxx>]

On Tue, Apr 13, 2010 at 08:19:33PM +0800, Garth N. Wells wrote:
>
>
> On 13/04/10 17:59, Anders Logg wrote:
> >On Tue, Apr 13, 2010 at 11:31:38AM +0200, Mehdi Nikbakht wrote:
> >>On Tue, 2010-04-13 at 10:43 +0200, Anders Logg wrote:
> >>>On Tue, Apr 13, 2010 at 10:31:51AM +0200, Mehdi Nikbakht wrote:
> >>>>On Tue, 2010-04-13 at 09:45 +0200, Anders Logg wrote:
> >>>>>On Tue, Apr 13, 2010 at 07:17:28AM +0800, Garth N. Wells wrote:
> >>>>>>
> >>>>>>
> >>>>>>On 12/04/10 23:35, Anders Logg wrote:
> >>>>>>>On Mon, Apr 12, 2010 at 10:20:13PM +0800, Garth N. Wells wrote:
> >>>>>>>>
> >>>>>>>>
> >>>>>>>>On 12/04/10 21:49, Anders Logg wrote:
> >>>>>>>>>On Mon, Apr 12, 2010 at 09:34:38PM +0800, Garth N. Wells wrote:
> >>>>>>>>>>
> >>>>>>>>>>
> >>>>>>>>>>On 12/04/10 21:29, Anders Logg wrote:
> >>>>>>>>>>>On Mon, Apr 12, 2010 at 09:21:32PM +0800, Garth N. Wells wrote:
> >>>>>>>>>>>>
> >>>>>>>>>>>>
> >>>>>>>>>>>>On 12/04/10 21:19, Garth N. Wells wrote:
> >>>>>>>>>>>>>
> >>>>>>>>>>>>>
> >>>>>>>>>>>>>On 12/04/10 20:47, Anders Logg wrote:
> >>>>>>>>>>>>>>We are doing some work where we need to do run-time quadrature over
> >>>>>>>>>>>>>>arbitrary polyhedra.
> >>>>>>>>>>>>>
> >>>>>>>>>>>>>We (Mehdi and I) do this already (using UFC), so I don't see why a new
> >>>>>>>>>>>>>function is required. Can you explain why evaluate_tensor is not enough?
> >>>>>>>>>>>>>
> >>>>>>>>>>>>
> >>>>>>>>>>>>I meant 'tabulate_tensor'.
> >>>>>>>>>>>
> >>>>>>>>>>>Which function do you call for evaluating the integrand?
> >>>>>>>>>>>
> >>>>>>>>>>
> >>>>>>>>>>We evaluate it inside ufc::tabulate_tensor. We construct our forms
> >>>>>>>>>>with an extra argument, say an object "CutCellIntegrator", which can
> >>>>>>>>>>provide quadrature schemes which depend on the considered cell.
> >>>>>>>>>
> >>>>>>>>>That would require a special purpose code generator.
> >>>>>>>>
> >>>>>>>>What's wrong with that? FFC won't (and shouldn't) be able to do
> >>>>>>>>everything. Just adding a function to UFC won't make FFC do what we
> >>>>>>>>do now. We reuse FFC (import modules) and add special purpose
> >>>>>>>>extensions.
> >>>>>>>
> >>>>>>>Exactly, it won't make FFC do what we need, but we could *use* FFC to
> >>>>>>>do what we need (without adding a special-purpose code generator).
> >>>>>>>
> >>>>>>>>>Having
> >>>>>>>>>evaluate_integrand would allow more flexibility for users to implement
> >>>>>>>>>their own special quadrature scheme.
> >>>>>>>>>
> >>>>>>>>
> >>>>>>>>We make "CutCellIntegrator" an abstract base class, so the user has
> >>>>>>>>*complete* freedom to define the quadrature scheme and the generated
> >>>>>>>>code does not depend on the scheme, since the scheme may depend on
> >>>>>>>>things like how the cell 'cut' is represented.
> >>>>>>>
> >>>>>>>Then it sounds to me like that generated code is not at all special,
> >>>>>>>but instead general purpose and should be added to UFC/FFC.
> >>>>>>>
> >>>>>>>And the most general interface would (I think) be an interface for
> >>>>>>>evaluating the integrand at a given point. We already have the same
> >>>>>>>for basis functions (evaluate_basis_function) so it is a natural
> >>>>>>>extension.
> >>>>>>>
> >>>>>>
> >>>>>>I still don't see the need for 'evaluate_integrand' unless you plan
> >>>>>>to call it directly from the assembler side (i.e. DOLFIN). Is that
> >>>>>>the case? Perhaps you can give me a concrete example of how you plan
> >>>>>>to use it.
> >>>>>
> >>>>>Yes, that's the plan. In pseudo-code, this is what we want to do:
> >>>>>
> >>>>>   for polyhedron in intersection.cut_cells:
> >>>>>
> >>>>>     quadrature_rule = QuadratureRule(polyhedron)
> >>>>>     AK = 0
> >>>>>
> >>>>>     for (x, w) in quadrature_rule:
> >>>>>
> >>>>>       AK += w * evaluate_integrand(x)
> >>>>>
> >>>>>     A += AK
> >>>>>
> >>>>>All data representing the geometry, the polyhedra, mappings from those
> >>>>>polyhedra to the original cells etc is in the Intersection class (in
> >>>>>the sandbox):
> >>>>>
> >>>>>   intersection = Intersection(mesh0, mesh1)
> >>>>>
> >>>>>Eventually, we might want to move part of the functionality into
> >>>>>either DOLFIN or FFC, but having access to and evaluate_integrand
> >>>>>function makes it possible to experiment (from the C++ side) without
> >>>>>the need for building complex abstractions at this point.
> >>>>
> >>>>As far as I understood you want to compute the integration points for
> >>>>polyhedrons inside Dolfin and evaluate_integrands will just compute that
> >>>>integrand in that specific integration points. If it is the case how
> >>>>would you determine what is the order of quadrature rule that you want
> >>>>to use?
> >>>
> >>>The quadrature rule would be a simple option for the user to
> >>>set. Currently we only have one general rule implemented which is
> >>>barycentric quadrature.
> >>
> >>Then if you use higher order elements, you need to update your
> >>quadrature rule manually. I think it would be nice to compute your own
> >>quadrature rule inside tabualte_tensor by using the standard quadrature
> >>rule.
> >
> >Yes, but the problem is that the computation of the quadrature rule is
> >nontrivial and it might be better to do it from C++ than as part of
> >the generated code, especially when the code depends on external
> >libraries like CGAL. See BarycenterQuadrature.cpp.
> >
>
> What we plan is that the generated code would provide a 1D scheme
> (or possible just a polynomial order) and the C++ quadrature object
> could apply it on a sub-triangulations.

I don't understand. How would the 1D scheme connect to a
triangulation?

Btw, we have considered subtriangulating the polyhedra to get a
quadrature rule, but decided to just to barycentric quadrature for now
to save time, both implementation and run time...

> >>>>Since to evaluate integrands you need to tabulate basis functions and
> >>>>their derivatives on arbitrary points. How do you want to tabulate basis
> >>>>functions and their derivatives inside evaluate_integrands?
> >>>
> >>>That's a good point. We would need to evaluate the basis functions and
> >>>their derivatives at a given arbitrary point which is not known at
> >>>compile-time.
> >>
> >>Yes, you need to use the tabulate_basis* functions implemented by
> >>Kristian. Then I am not the only one who is using them. Good for
> >>Kristian. ;)
> >
> >ok, good. Then there is no principal problem of generating code for
> >evaluate_integrand (if it is allowed to call tabaulate_basis).
> >
> >I don't see what the problem is of adding evaluate_integrand. It is a
> >natural extension (we have evaluate_basis already), it would be
> >"simple" to implement (Kristian can correct me if I'm wrong) and it
> >makes generated code useful for assembly over cut meshes (without the
> >need for writing a special-purpose code generator).
> >
>
> It's not a problem - just good to scrutinise heavily additions to

Agree.

> UFC to avoid bloat. I'm happy for it to be added. We should make
> clear that it's for testing/special cases since the performance for
> computing element tensors will be poor compared to tabulate_tensor.

ok. So we have the following proposed additions to UFC:

  evaluate_integrand
  tabulate_tensor(x, w)
  evaluate_basis (two versions)

  + quite a few other suggestions as blueprints
  https://blueprints.launchpad.net/ufc

Perhaps it's a good time to think about UFC 2.0 and take care of all
the blueprints. Or perhaps we should wait until after DOLFIN 1.0?

I have limited resources at the moment so we might want to postpone
this.

--
Anders

Attachment: signature.asc
Description: Digital signature