dolfin team mailing list archive

[Martin Sandve Alnæs <martinal@xxxxxxxxx>]

On Mon, Apr 27, 2009 at 3:32 PM, Garth N. Wells <gnw20@xxxxxxxxx> wrote:
>
>
> Anders Logg wrote:
>> On Mon, Apr 27, 2009 at 09:21:58AM -0400, Shawn Walker wrote:
>>> On Mon, 27 Apr 2009, Anders Logg wrote:
>>>
>>>> On Mon, Apr 27, 2009 at 11:26:13AM +0200, Kent Andre wrote:
>>>>> On lø., 2009-04-25 at 00:14 +0200, Anders Logg wrote:
>>>>>> On Wed, Apr 22, 2009 at 05:28:30PM -0400, Shawn Walker wrote:
>>>>>>> Here is the changeset that adds a `higher_order_coordinates' variable for
>>>>>>> storing higher order mesh data.  This is a very minor change so please
>>>>>>> push this.
>>>>>>>
>>>>>>> A changeset for DOLFIN is coming immediately after this.
>>>>>>>
>>>>>>> - Shawn
>>>>>> I'm not sure what to do about this. It's problematic to add
>>>>>> experimental work to UFC since it must be stable. In particular, any
>>>>>> small change to ufc.h means that all forms must be recompiled
>>>>>> everywhere for everyone.
>>>>>>
>>>>>> So before we make a change to UFC, we need to know exactly what we
>>>>>> need. Which also means I can't import your DOLFIN patch since it
>>>>>> depends on the UFC patch.
>>>>>>
>>>>>> I see you've added
>>>>>>
>>>>>>     double** higher_order_coordinates;
>>>>>>
>>>>>> to ufc::cell. This is analogous to what is now implemented in
>>>>>> MeshGeometry and the mesh XML format so I think it's good.
>>>>>>
>>>>>> The question is what other information we need. As it works now (for
>>>>>> the standard ufc::cell), UFC code generated by a form compiler knows
>>>>>> what to expect from for a ufc::cell argument. If higher order mappings
>>>>>> should work the same way, then the generated code and thus the form
>>>>>> compilers need to know which mapping should be used and also the
>>>>>> length of higher_order_coordinates. Is this what you were thinking?
>>>>>>
>>>>>> Before we do much more about it, more people need to weigh in on it as
>>>>>> it affects DOLFIN, UFC, SyFi and FFC.
>>>>>>
>>>>>
>>>>> But is there any other way around this. It would be nice with higher
>>>>> order meshes and UFC should not stop this.
>>>>>
>>>>> An alternative to changing the cell class would be to make a subclass
>>>>> of cell. Would this work ?
>>>> How about just using the current ufc::cell data structure as it is but
>>>> let coordinates hold all the coordinates?
>>>>
>>>> This could also be the final solution. Then everything that's needed
>>>> is an extra argument to tabulate_tensor that tells the generated code
>>>> whether the cell is affinely mapped or not. The flag could simply be
>>>> an integer: 1 means affine, 2 means quadratic etc.
>>> But you still need to modify the ufc::cell code, I think.  There is also
>>> an implicit assumption that the higher order coordinates should contain
>>> the standard mesh vertex coordinates.  Of course, this is true for most
>>> practical cases.  But for more fancy mappings, maybe this is not the
>>> case.
>>
>> It seems to me that a reasonable assumption would be to limit the
>> cases to P1, P2, P3, etc, that is, mappings that can be written down
>> using standard Lagrange bases so then the vertices will always be
>> included. They would also be first in the list meaning that the code
>> would actually work (but might not give accurate results) even if it
>> were generated for affine mappings.
>>
>>> Also, in the ufc::cell code, you currently read in the cell coordinates
>>> using info in MeshTopology.  However, the higher order coordinate info
>>> resides in MeshGeometry (which is where it belongs).  So you would still
>>> need to modify ufc.h.   Remember, there is higher order cell data that is
>>> contained in MeshGeometry.
>>
>> Where is MeshTopology used for this? I looked in UFCCell.h which is
>> where the coordinates are copied to ufc::cell and there MeshGeometry
>> is used.
>>
>>> Is it really that hard to change ufc.h?  Other things have to be
>>> recompiled, but isn't that automatic?
>>
>> Yes, it's easy to change, but a main point with UFC is that we
>> shouldn't change it.
>>
>
> UFC will need to be extended as time goes on, but it is hard to know
> from the outset how it should be done. What about using some IFDEF's or
> non-pure virtual functions in the development version to allow
> experimentation? These can then either be removed or added to UFC at
> release time.
>
> Garth

Or subclasses with non-pure virtual functions:

class experimental_cell_integral: public ufc::cell_integral
{
  void foo() const { throw ...("Experimental feature not implemented."); }
};

or

namespace eufc {
  class cell_integral: public ufc::cell_integral
  {
    void foo() const { throw ...("Experimental feature not implemented."); }
  };
}

We can define these in "experimental_ufc.h" or "eufc.h" to keep the
official header file constant.

Then the DOLFIN code that uses experimental features must be clearly marked:

  ufc::cell_integral *itg = form.create_cell_integral(0)
  eufc::cell_integral *eitg = dynamic_cast<eufc::cell_integral>(itg);

and can then use "if(eitg)" to select between experimental and
non-experimental code.

Martin