Cray...Code dependent on IEEE floating point?

[mcguire]

Hi folks. I'm new to the project, just started yesterday.

Can someone tell me if the distributedfolding code depends on IEEE floating point math? I ask because I have a Cray vector supercomputer here (all but the latest Cray PVP systems lack IEEE support) and if I can get my hands on the source code I might be able vectorize the code.

Thanks,
-Dave

[bwkaz]

Well I don't have access to source, so I don't know how much it depends on IEEE 754, but I do remember hearing (I think it was in the thread "altivec enhancements", if you want to search for that one, but it is quite long) that vectorization won't really help the DF algorithm all that much. The vast majority of the time is spent chasing pointers, not doing math, IIRC. ;)

However, you would probably want to wait until the developer (goes by the handle Brian The Fist) chimes in. He may have more info than I can give.

In any case, welcome aboard!

[mcguire]

Hmm, well the Cray PVP machines' main strength is of course in vectorizable code. Oh well.

However, the machine does have truly amazing memory bandwidth (good for chasing those pointers), as well as eight (soon to be twelve) processors, so it might do reasonably well anyway.

-Dave

[Brian the Fist]

Sounds cool. So do I get telnet access? :D
Seriously though, yes, we use some floating point to compute atom co-ordinates but most time is spent following memory pointers through complex data structures.

(Some of) our code is not open source but vecotrizing is not expected to speed it up. The code is entirely in ANSI C which I assume would compile on your CRAY though.

[runestar]

If you get telnet access, does he get like an honorary staff member position or something? ;)


RuneStar