On Thu, May 24, 2012 at 6:26 PM, Barry Smith <[email protected]> wrote:

...

So do we manage blocks using internal C++ templates, "templates in C", C generated using some other system (m4 anyone?), or something else entirely?

Yes! Finally, we acknowledge that this a problem.

1) C++ templates are not a solution to anything. ANYTHING.

2) I am assuming "templates in C" would work somewhat like a templating engine. I tried this for the last TOMS paper with Andy. Its was just not a big payoff for the work put in, and definitely did not justify incorporating another package.

3) I prefer C generated from another system, like the one I use for FEM (which I am not attached to). We will definitely need this for GPU kernels, and I am guessing thread kernels if they are going to be worth something.

For this particular example (and perhaps many others for sparse matrices) it is a matter of writing "the same algorithm" for a different data structure (the split storage). Perhaps what is needed is a "sparse matrix/graph/mesh" language, for which one can write kernels/code fragments "independent of the data structure" from which C/whatever is generated? But I don't have a clue what the language would look like. I think a general purpose tool (for example templates) is unlikely to be useful for us.

Haven't we already identified most of these primitives? Doesn't it generally come down to selecting a (block) row (upper, lower, or the entire thing) and applying a blocked sparse-dense-{minus,plus}-dot?