[pooma-dev] fields on unstructured meshes

[James Crotinger]

>-----Original Message-----
>From: Toon Knapen [mailto:toon.knapen at si-lab.com]
>Sent: Wednesday, April 24, 2002 11:25 AM
>To: pooma-dev at pooma.codesourcery.com
>Subject: [pooma-dev] fields on unstructured meshes
>
>
>The documentation of pooma 2.3 indicates that one of the next releases will
>contain unstructured meshes. I guess it's not yet for 2.4 since I can't
>find
>them in the CVS.

That was the plan when the Pooma team was still at Los Alamos. Obviously the
docs were not given a thorough review when 2.3 was released. When 2.4 was
started, I think there was still some hope that unstructured might be a
follow-on project - probably 2.5. As the lab cut funding for 2.4 in the
middle of the project, that is very unlikely. 

>But I don't see how the Field concept will be extended to unstructured
>meshes.  A field currently supports Array-like indexing (e.g. 2 indices for
>2-dimensional meshes). But on unstructured meshes, the elements need be
>numbered using some numbering scheme that is not purely based on the
>position
>of the element. But what happens with the Array-like indexing in this case
>?

There would be no tensor product indexing. This was definitely a
research-area that we had not explored in much depth. But we did believe
that there was value to data-parallel abstractions with unstructured. 

>I'm also curious why you decided to use the dimension (of the problem) as a
>template parameter . As a consequence, you need different executables to
>simulate input-files describing a 2D and a 3D problem. But I guess the
>decision is based on the performance that can be gained by fixing the
>dimension ?

This is required for the compile-time optimizations to work. Indexing has to
know the dimensionality in order for expressions like

   A(I,J) = B(I,J+1) - B(I-1,J);

to effectively turn into a set of nested loops each containing 

   a(i,j) = b(i,j+1) - b(i-1,j); // Fortran notation
 
This has to happen at compile time in order for the expression templates to
yield the kind of efficiency that one would get using hand-coded loops. 

This does lead to code bloat in certain applications, but it is basically
unavoidable if you're going to use expression templates.

	Jim

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://sourcerytools.com/pipermail/pooma-dev/attachments/20020426/e91b3b98/attachment.html>