ffc team mailing list archive

[Anders Logg <logg@xxxxxxxxx>]

On Wed, Mar 19, 2008 at 12:08:16PM +0100, Kristian Oelgaard wrote:
> Quoting Anders Logg <logg@xxxxxxxxx>:
> 
> > On Tue, Mar 18, 2008 at 02:37:49PM -0400, Marie Rognes wrote:
> > > Kristian Oelgaard wrote:
> > > > Quoting Marie Rognes <meg@xxxxxxxxxxx>:
> > > >
> > > >   
> > > >>
> > > >> However, this time it is actually not simplify that is the main
> > culprit:
> > > >>
> > > >>     After index reassignment:
> > > >>      
> > > >>           0.5 | vi0[0]*vi1[0]*dX(0) + 0.5 | vi0[0]*vi1[0]*dX(0) + 0.5 |
> > > >>     vi0[1]*vi1[1]*dX(0) + 0.5 | vi0[1]*vi1[2]*dX(0) + 0.5 |
> > > >>     vi0[2]*vi1[2]*dX(0) + 0.5 | vi0[2]*vi1[1]*dX(0) + 0.5 |
> > > >>     vi0[3]*vi1[3]*dX(0) + 0.5 | vi0[3]*vi1[3]*dX(0)
> > > >>      
> > > >>       After simplification:
> > > >>      
> > > >>           0.5 | vi0[b0]*vi1[b0]*dX(0) + 0.5 | vi0[b0]*vi1[b0]*dX(0) +
> > > >>     0.5 | vi0[1]*vi1[2]*dX(0) + 0.5 | vi0[2]*vi1[1]*dX(0)
> > > >>
> > > >>
> > > >> where the index b0 has range [0, 1, 2, 3] in the first term and [0, 3] 
> > > >> in the second term. This makes sense.
> > > >>     
> > > >
> > > > Maybe, on a side note: the fact that 'b0' can have a different range in
> > each
> > > > monomial makes it pretty difficult to debug/understand what is going on
> > by just
> > > > looking at the pretty print version of the form. The index should be
> > different
> > > > in my opinion.
> > > >
> > > >   
> > > 
> > > Agree. Maybe we could have two versions of printing indices, one with 
> > > the ranges (if set) ala
> > > 
> > >     b0 (0, 1, 2, 3)
> > > 
> > > for debug purposes and a denser one for non-debug purposes?
> > > 
> > > Actually, that would be a starting point for an accurate factorization 
> > > check too.
> > 
> > See if it works now. It was a simple fix (assuming it works) to
> > include the index range in the computation of signatures (see
> > signature.py).
> 
> Yes, it works, but the output could be optimised:
> 
>     // Compute geometry tensors
>     const double G0_ = det;
>     const double G1_ = det;
>     const double G2_ = det;
>     const double G3_ = det;
>     
>     // Compute element tensor
>     A[0] = 0.25*G0_ + 0.25*G1_;
>     A[1] = 0;
>     A[2] = 0;
>     A[3] = 0;
>     A[4] = 0;
>     A[5] = 0.25*G0_;
>     A[6] = 0.25*G2_;
>     A[7] = 0;
>     A[8] = 0;
>     A[9] = 0.25*G3_;
>     A[10] = 0.25*G0_;
>     A[11] = 0;
>     A[12] = 0;
>     A[13] = 0;
>     A[14] = 0;
>     A[15] = 0.25*G0_ + 0.25*G1_;
> 
> So something in factorization should change.

Yes, but I don't want to mess with it unless I really have to.
A lot of work needs to be done when we move to UFL anyway, so
I'd rather wait.

> And it would be nice to include some information about the range of the indices
> in the form signature:
> 
>   After simplification:
> 
>       0.5 | vi0[b0]*vi1[b0]*dX(0) + 0.5 | vi0[b0]*vi1[b0]*dX(0) + 0.5 |
> vi0[1]*vi1[2]*dX(0) + 0.5 | vi0[2]*vi1[1]*dX(0)
> 
> b0 in first product 0-3, in second 0-1.

ok, see if you like it now.

-- 
Anders