The x-lib: a really simple library for message passing

[mhonman]

During "Preparing for Parallella" I was stuck by how much more difficult it seemed to structure parallel programs on the Epiphany when compared to Occam and the Transputer (I'll blame Iain Barron's keynote for reminding me of all their good points). On the other hand the low-latency communication mesh provided by Epiphany is a brilliant feature that frees the programmer from the limitations of Occam's purely-local memory model and the point-to-point links of the Transputer hardware.*

Prodded by the comment that "someone should look into implementing a channel abstraction for Epiphany" and tempted by Ben's offer of access to the Sundance Zedboard, well, here's something that might be a start.

It is provisionally called the x-lib, being eXperimental stuff that might become useful eXtensions to the Adapteva SDK.

The basic idea for pairwise synchronisation between cores comes from the e_barrier code in the SDK (i.e. write to peer's internal memory, and busy-wait on own internal memory). Thanks to Yaniv for that and the patient explanations along the way.

A test/demonstration application can be found here with the x-lib code: . It works on a Zedboard with Epiphany-III and esdk.5.13.07.10, other configurations are untested.

There is a rather work-in progress manual there too. Since we don't really know if the x-lib API will be useful, or how much it might change before it becomes useful, the manual does not contain a comprehensive listing of the function calls and the behaviour. However I've endeavoured to provide information on the purpose and behaviour of the x-lib functions within comments in the header and source files.

I think the basic synchronisation concept should be extensible to support one-message-in-transit scenarios, and the simultaneous data exchanges that usually feature in applications that have a grid-structured computational domain.

Feedback would be very much appreciated, even if it is "too slow", "too unwieldy", "can't I do it XYZ way?" etc.

*IMO the key conceptual difference between then and now is that cores are so cheap that there is no need for processes to share a core, thus techniques like busy-waiting become socially acceptable.

[mhonman]