Hi Nelson!
I know that a better algorithm exists that updates the u's on a "checkers board", so that all new (say "black positions") values are updated using the old (say "white positions") values instead of the completely random update of the code below. I intend to verify if this option passes through llvm without problems.
I'm not sure whether this will be helpful to you, but a little while ago I posted an example of some Chapel code that uses the Checkerboard algorithm you're talking about on another thread. See this message: Code optimization - #24 by jcorrado
That code is part of a 3D Poisson solver that uses SOR, so it might be similar enough to be a useful reference for you.
Please let me know if you have any questions.
Best,
Jeremiah