Victor Eijkhout <[email protected]> writes:

On Dec 31, 2013, at 3:10 PM, Jed Brown <[email protected]> wrote:

...

Hmm, I don't really see the API for specifying the distributions. rightshift->operate(">>1") is a pretty special-purpose interface. Presumably this part is non-collective,

No, it's collective. As I argue in section 1.1, I want sequential semantics.

Okay, but surely a more general interface for describing the distribution must involve setting non-collectively (at least the data is not collective, so you'll have to build the two-sided representation).

It's actually not all that special purpose. The shift notation is enough if you have stencils with offsets that are constant during program execution.

Uh, only for structured grids, and for multiple dimensions, you either need this communication abstraction to know about multi-dimensional layout (essentially putting DMDA inside VecScatter) or you can't use the shift notation.

And you did read section 2.3, right?

Yeah, it's cute but limited and IMO, not very useful in practice.