[pypy-dev] PyPy translation on Xeon Phi (pka MIC)

[wlavrijsen at lbl.gov]

Armin,

> Wim's reply is not helpful at all for me, as it is mostly given as a
> series of three-letter acronyms I've never heard about :-)

well, what Yury said. :) Sorry, but I was replying to an Intel guy ...

To first order, think of it as a server that you ssh to, and then you
find yourself on a perfectly ordinary Linux machine, albeit one that
presents you with 240 "cpus."

For the current generation, you have to deal with cross-compilation, b/c
its vector instruction set is unique. Beyond that, it's just x86. (The
next generation uses AVX-512, so no more cross-compilation necessary.)
Since PyPy does not generate vector instructions (nor deals with any
offloading), the easiest I found was to generate C, and cross-compile
that, all on the host; not try to run the translator on the Phi. Even
so, I didn't find anything worthwhile: loosing vectorization is okay
(for high energy physics codes, as they're too branchy anyway), but you
really need to be able to scale out.

Btw., we're in big for its successor:

   https://www.nersc.gov/users/computational-systems/cori/

Best regards,
            Wim
-- 
WLavrijsen at lbl.gov    --    +1 (510) 486 6411    --    www.lavrijsen.net