Hi
I'm a software developer with the Einstein@Home project (on astrophysics, relativity etc). Together with apps for Android, we provide a Raspberry Pi app and are currently testing (on our test project, Albert@Home:
http://albert.phys.uwm.edu ) an ARMv7 Linux app with support for NEON. Unfortunately I don't have my Parallella yet (backer # 5000ish, might take a while....), so anybody with a prototype board is welcome to try it.
Of course the ARMv7-NEON version we have now does not use the Epiphany. The source code (including an OpenCL version we use for NVIDIA, ATI/AMD and (integrated) Intel GPUs) is open source, tho.
As much as I would like to see the Einstein@Home app optimized for the Epiphany chip, I'm really skeptic this would speed up the app: it spends most of the time doing really big FFTs (on the order of 2^22, so about length= 4 million, real to complex, single precision FFT). With the combination of a not-so-fast (wrt throughput) interface between the RAM and the Epiphany, small internal RAM of Epiphany and low (O(n log n) ) complexity of FFT, my gut feeling is that it will be very hard to implement FFT faster in Epiphany than in the ARM CPU. Right??
Any comments welcome.
Cheers
HBE