by dobkeratops » Fri Aug 07, 2015 11:48 pm
>> What we really have to do is to make AI-like compiler, which can translate (C++) programs into machine code for specified series of numbers of Epiphany cores.
Does anyone know any details of REX computings' approach ?
Their website seems to indicate they have some sort of elaborate plan for profile-guided optimisations ..
I can imagine having multiple versions of similar high-order-functions for parallelism (things like par_foreach managing DMA for you) and/or some template parameters, allowing you to tweak your implementation but it sounds like they plan to absorb some of this complexity into a compiler, which could be cleaner, if it works.
How far could this go. Imagine something to automatically break large loop bodies up into pipeline stages (a pipelined par_foreach would be an easy way to leverage a lot of cores..)