ffc team mailing list archive

[Kristian Ølgaard <k.b.oelgaard@xxxxxxxxx>]

On 20 January 2011 14:59, Marie E. Rognes <meg@xxxxxxxxx> wrote:
> On 01/20/2011 02:55 PM, Kristian Ølgaard wrote:
>
> On 20 January 2011 14:32, Marie E. Rognes <meg@xxxxxxxxx> wrote:
>
> Repeated runs of the dolfin generate-form-files scripts seems to generate
> different code. For instance for the forms defined in
>
>     /demo/undocumented/dg-advection-diffusion/cpp/AdvectionDiffusion.ufl
>
> I get differences of the type
>
> -    G[1] = 0.500000000000000*det*n1;
> -    G[2] = 0.500000000000000*det;
> +    G[1] = 0.500000000000000*det;
> +    G[2] = 0.500000000000000*det*n1;
>
> in tabulate_tensor and corresponding differences for the tensor terms. The
> code seems equivalent, but this behaviour makes it tedious to check for
> changes in the generated code. Has anyone else experienced this?
>
> Not lately, but I haven't been looking for it either. The quadrature
> generator uses dicts heavily, so somewhere we need to throw in a
> sorted(.keys()) to get rid of this behaviour but I don't know where
> exactly.
>
> Do you have any guesses (other than in quadrature/*)? I can fix if you point
> me in the right direction.

Yes, in quadraturegenerator.py the dictionary geo_consts from the ir
is already created, and since the randomness only appears when
optimisations are on, the 'random' geo_consts dictionary must be
generated in quadratureoptimization.py. Expressions are added to the
dictionary by the function optimise_code from quadrature.symbolics but
the problem might have something to do with the order by which this
function is called (the loops in the _*() functions in
quadratureoptimization.py).

Kristian

> --
> Marie
>
>