dolfin team mailing list archive

[Anders Logg <logg@xxxxxxxxx>]

On Mon, May 08, 2006 at 04:18:02PM +0200, Garth N. Wells wrote:
> On Mon, 2006-05-08 at 09:12 -0500, Robert C. Kirby wrote:
> > The situation in which templating versus nontemplating will matter is  
> > when the compiler is capable of inlining away a lot of function calls  
> > in templated code.  For example, indexing into a matrix is sometimes  
> > done virtually so that you can access a matrix, sum of matrices, etc  
> > with the same syntax.  If it's templated instead (and correct  
> > optimizations happen), you'll get a big win.  This is the basic  
> > technical contribution of expression templating.  Virtual functions  
> > can never be inlined (this is a "theorem") in C++.
> > 
> 
> This is exactly what I saw when testing wrappers for uBlas. The key to
> accessing matrix and vector elements effectively was inlining, and as
> you note, virtual functions can't be inlined. The discounted the
> envelope-letter design for matrices and vectors as all member functions
> are accessed via a virtual function.
> 
> Garth 

We could choose to re-add the envelope-letter design at some point
(but not now). The virtual functions are in GenericMatrix and we just
decided to use it. All the functions in the envelope (which we could
add) could be inlined.

If you just use functions that do more than an insignificant amount
of work, the overhead of the virtual functions can be neglected, as
the benchmark shows. The good thing about the new design is that if
you really want to index, then use DenseMatrix and it's still possible
to call the assembly functions on it.

/Anders


> > 
> > On May 8, 2006, at 6:24 AM, Anders Logg wrote:
> > 
> > > I have added a benchmark to src/test/main.cpp. It compares the speed
> > > of assemblhy for 2D Poisson on a 128 x 128 mesh repeated 100
> > > times. Here are some typical timings:
> > >
> > > Time with template:    4.23
> > > Time without template: 4.24
> > >
> > > Time with template:    4.24
> > > Time without template: 4.21
> > >
> > > Time with template:    4.2
> > > Time without template: 4.29
> > >
> > > Time with template:    4.22
> > > Time without template: 4.24
> > >
> > > Time with template:    4.22
> > > Time without template: 4.15
> > >
> > > Sometimes the templated version wins and sometimes the non-templated
> > > wins, so I'd say there is no overhead of removing the templating. The
> > > advantage would be that FEM and other implementations can be
> > > simplified and more code moved to .cpp from .h.
> > >
> > > Garth, maybe you can check the benchmark and see if you agree with the
> > > results. I've compiled with --enable-optimization.
> > >
> > > /Anders
> > >
> > > _______________________________________________
> > > DOLFIN-dev mailing list
> > > DOLFIN-dev@xxxxxxxxxx
> > > http://www.fenics.org/cgi-bin/mailman/listinfo/dolfin-dev
> > 
> > 
> > _______________________________________________
> > DOLFIN-dev mailing list
> > DOLFIN-dev@xxxxxxxxxx
> > http://www.fenics.org/cgi-bin/mailman/listinfo/dolfin-dev
> 
> 
> _______________________________________________
> DOLFIN-dev mailing list
> DOLFIN-dev@xxxxxxxxxx
> http://www.fenics.org/cgi-bin/mailman/listinfo/dolfin-dev