ffc team mailing list archive

[Anders Logg <logg@xxxxxxxxx>]

On Fri, Mar 27, 2009 at 09:49:47AM +0100, Kristian Oelgaard wrote:
> Quoting "Garth N. Wells" <gnw20@xxxxxxxxx>:
> 
> >
> >
> > Anders Logg wrote:
> > > FFC is now fully functional w.r.t. compilation of UFL forms, at least
> > > judging by the demos in FFC which now all produce correct results.
> > >
> > > The Poisson demo in DOLFIN has also been moved to UFL.
> > >
> > > I therefore suggest we make new releases of FFC, UFL and DOLFIN:
> > >
> > >   DOLFIN 0.9.2
> > >   FFC 0.6.2
> > >   UFL 0.1.0 ?
> > >   SyFi/SFC?
> > >
> > > This batch of releases will support both the old .form format and the
> > > new .ufl format. After the release, we can get started on cleaning up
> > > and removing the .form format.
> > >
> > > In it's simplest form, the UFL form language looks very much like the
> > > FFC form language. Here's a comparison for the Poisson demo:
> > >
> > > FFC
> > > ---
> > >
> > >   element = FiniteElement("Lagrange", "triangle", 1)
> > >
> > >   v = TestFunction(element)
> > >   u = TrialFunction(element)
> > >   f = Function(element)
> > >
> > >   a = dot(grad(v), grad(u))*dx
> > >   L = v*f*dx
> > >
> > > UFL
> > > ---
> > >
> > >   element = FiniteElement("Lagrange", triangle, 1)
> > >
> > >   v = TestFunction(element)
> > >   u = TrialFunction(element)
> > >   f = Function(element)
> > >
> > >   a = inner(grad(v), grad(u))*dx # dot also works
> > >   L = v*f*dx
> > >
> > > However, UFL does much more, like treatment of tensor expressions,
> > > nonlinear operators like sin, cos, exp, sqrt, differentiation of forms
> > > w.r.t. variables etc. For more information, look at Martin's summary:
> > >
> > >   http://www.fenics.org/wiki/UFL
> > >
> > > UFL includes a script form2ufl that can be used to convert from the
> > > old FFC form language to the new UFL form language. The script is not
> > > foolproof but it may help in converting to the new format.
> > >
> > > So, let's make a new release now (after getting the buildbot green
> > > again for DOLFIN) and then make the full transition after the release.
> > >
> > > It's also time to promote UFL from a development project (/dev) to a
> > > proper project (/hg) since now FFC, SFC and DOLFIN will depend on it.
> > > Any objections?
> > >
> >
> > It would be worth checking with Kristian how far away the quadrature
> > optimisations are for FFC + UFL. If it's only a few days, it would be
> > good to wait on the FFC release since the new quadrature code is quite a
> > bit slower than the old code (around a factor of 5 in some of my tests).
> 
> It's already there, just uncomment the optimise_options lines at the top of the
> uflquadraturegenerator.py file.
> However, there might still be bugs in the generated code and it compiles slower
> (I've seen a factor of around 6-7 in some cases)
> 
> Some benchmark tests:
> 
> PressureEquation       FFC comp.    num. ops  assemble 2.5e6 times
> Old FFC (*.form),+opt: 45.0s         26679          28.2s
> New FFC (*.ufl), -opt:  0.9s        166086         110.0s
> New FFC (*.ufl), +opt:  3.4s         15771          21.8s
> 
> DGElastoDynamics  FFC comp. | ops. a, dx | ops. a, dS | ops. L, dx | ops. L, dS
> (*.form), +opt:     16m20s        791        2098         1571         4305
> (*.ufl),  -opt:         7s      25776       93600        10008        23280
> (*.ufl),  -opt:        45s       1217        2529         1375         2923
> 
> For the elasto-dynmics case, running 10 time steps on a 9000 element mesh takes
> roughly 14.5s for both of the two optimised versions. The new FFC without
> optimisation runs 1 second slower (15.6s).
> 
> Kristian

I don't think we need to wait for further testing / optimization. The
old .form format is still functioning and it's the default choice.

If we make a release, it will be easier to add the optimizations since
we can remove all the old stuff and reset the references for the
regression tests to the new format (simplifying and speeding up
testing).

-- 
Anders

Attachment: signature.asc
Description: Digital signature