On Sun, Mar 29, 2020 at 6:20 PM Jed Brown <[email protected]> wrote:

We are interested in assembling point-block diagonals for use in pbjacobi smoothing. This is much simpler than AIJ,

Would the coloring/DD algo on the outside be the same, just with a different assembly kernel? Or are you looking a fusing the element construction with assembly? That seems reasonable now that I think about it. It's the same coloring, right? And, do we have an appropriate coloring method now? Thanks, Mark

but lots of people will need AIJ so it's certainly worthwhile.

Mark Adams <[email protected]> writes:

...

Is there any plans/interest in assembling AIJ matrices on GPUs?

I have a somewhat strangely balanced FE operator and I have each thread block create one element matrix, one thread per integration point with an expensive computation at each integration point. Then I assemble the matrices on the CPU.

I want to parallelize assembly on GPU machines, with coloring or DD I assume. Does anyone else care about this?

Thanks, Mark