ufl team mailing list archive

[kent-and@xxxxxxxxx]

> UFL doesn't do any simplifications and rewriting at all, by design.
> Rewriting rules are not reversible in general, but can always be
> applied later. The exception is that obvious zeros are removed in
> some key places, e.g. 0*f -> 0, 0+f -> f, and during AD.
>
> The same applies for the flattening of expressions like (f*f)*f -> (f*f*f)
> which is implemented as a function. We may collapse (f*f*f) to 3*f
> in that routine as well, but in general the problem of deciding which
> simplifications it is a good idea to do does not have a unique solution.
>

ok, but I guess the simple solution in this example is to implement
a pow function in a similar fashion as cos, etc.
If you think it is a good idea, I can implement it. (It would be good for
me to
try to implement something. )

But, this is an example in which AD produces "an explosion", if one does
not have
simplification algoriithms (or a pow function with a derivative). Is this
the only example then ?

Kent

> The form compiler may or may not be able to use information
> from the original expression form in heuristics for optimization.
>
> The latex compiler will produce bad output (not close to user input)
> if we do too much automatic simplifications.
>
> --
> Martin
>
>
> 2008/9/24 Kent Andre <kent-and@xxxxxxxxx>:
>>
>> I tried:
>> a = f*f*f*f*f*f*v*dx
>> b = derivative(a, f)
>>
>> which gets a bit larger than it should:
>>
>> { (((((((((((v_0 * w_0) + (w_0 * v_0)) * w_0) + ((w_0 * w_0) * v_0)) *
>> w_0) + (((w_0 * w_0) * w_0) * v_0)) * w_0) + ((((w_0 * w_0) * w_0) *
>> w_0) * v_0)) * w_0) + (((((w_0 * w_0) * w_0) * w
>> _0) * w_0) * v_0)) * v_-2) } * dx0
>>
>> Or is this an efficient way to represent it ?
>>
>> Kent
>>
>>
>> On on., 2008-09-24 at 15:37 +0200, UFL wrote:
>>> One or more new changesets pushed to the primary ufl repository.
>>> A short summary of the last three changesets is included below.
>>>
>>> changeset:   261:47f24c5a799264152018ebb14d5c623206cb8c12
>>> tag:         tip
>>> user:        "Kent-Andre Mardal <kent-and@xxxxxxxxx>"
>>> date:        Wed Sep 24 15:37:24 2008 +0200
>>> files:       demo/ad_ex1.ufl
>>> description:
>>> ad test
>>>
>>>
>>> changeset:   260:4e87970569562901a101b2bc90b916c90e552cfd
>>> user:        "Martin Sandve Alnæs <martinal@xxxxxxxxx>"
>>> date:        Wed Sep 24 15:28:22 2008 +0200
>>> files:       test/classcoverage.py test/mass.py ufl/algorithms/ad.py
>>> ufl/algorithms/transformations.py ufl/algorithms/ufl2latex.py
>>> ufl/formoperators.py
>>> description:
>>> Some bugfixes and test fixes.
>>> Some tests fail, but that's ok for now.
>>>
>>>
>>> changeset:   259:5693657007c3c3fbb954aaf20aecabb4761f5608
>>> user:        "Martin Sandve Alnæs <martinal@xxxxxxxxx>"
>>> date:        Wed Sep 24 14:06:55 2008 +0200
>>> files:       ufl/algorithms/ad.py ufl/base.py ufl/baseoperators.py
>>> ufl/common.py
>>> description:
>>> More AD rules, most important ones are in place now.
>>>
>>> ----------------------------------------------------------------------
>>> For more details, visit http://www.fenics.org/hg/ufl
>>> _______________________________________________
>>> UFL-dev mailing list
>>> UFL-dev@xxxxxxxxxx
>>> http://fenics.org/mailman/listinfo/ufl-dev
>>
>> _______________________________________________
>> UFL-dev mailing list
>> UFL-dev@xxxxxxxxxx
>> http://fenics.org/mailman/listinfo/ufl-dev
>>
>