ufl team mailing list archive

[Martin Sandve Alnæs <martinal@xxxxxxxxx>]

On Thu, May 14, 2009 at 9:43 AM, Anders Logg <logg@xxxxxxxxx> wrote:
> On Thu, May 14, 2009 at 08:40:41AM +0100, Garth N. Wells wrote:
>>
>>
>> Martin Sandve Alnæs wrote:
>> > Is there a mathematical difference?
>> >
>>
>> Splitting hairs, I would say yes and that grad(u('+')) is correct and
>> grad(u)('+') is not because grad(u) is not defined in a classical sense.
>> I prefer grad(u('+')), but FFC used to use grad(u)('+'). My preference
>> is however not strong.
>>
>> My immediate question is will the generated code differ for grad(u('+'))
>> and grad(u)('+')?
>>
>> Garth
>
> No, since the restriction will be propagated to the leaves... :-)

The leaves are terminal expressions, that is literal constants,
geometric quantities, functions and basis functions.
"Propagating to the leaves" means modifying for example

  (grad(f) + grad(g))('+')
to
  (grad(f('+')) + grad(g('+')))

such that the restrictions apply directly to the terminal expressions.
That will simplify the work of the form compilers, and allow
for example taking the jump of anything.

Basically, I think the canonical representation after some algorithms should be
  u('+').dx(i)
and with indexing either
  u('+').dx(i)[j]
or
  u('+')[j].dx(i)

If restrictions, (indexing) and derivatives are applied to a
(non-scalar) function (terminal).
Currently, both
  u.dx(i)[j]
and
  u[j].dx(i)
can occur, and restrictions are not handled.

Martin