> On Jan 13, 2015, at 12:38, Jason Swails <jason.swails.gmail.com> wrote:
>
> On Tue, Jan 13, 2015 at 12:20 PM, Novosielski, Ryan <novosirj.ca.rutgers.edu
>> wrote:
>
>> Wouldn't there be performance tradeoffs there? I'd think that at least
>> supporting 4.1, in my case, would result in faster execution.
>
> I thought I remembered somebody (Brent?) working on this very problem --
> i.e., that SSE_TYPES is basically ignored by configure. I'm not sure what
> came of it, but I was under the impression that he had gotten pretty far
> with it. If he sees this and easily recalls what he's done, perhaps
> (hopefully? :) he might chime in here.
>
> I think the main problem here is that Amber uses the "-xHost" flag, which I
> think activates all optimizations for the instruction set on the host
> processor. This obviously causes problems on machines where the head node
> supports more advanced vector instructions than the compute nodes, which
> seems to be what's happening here.
>
> Since I am under the impression that setting the SSE_TYPES variable does
> *not* currently work for configure, I think your best bet might be to
> backup configure2 inside $AMBERHOME/AmberTools/src and change "-xHost" to
> the Intel compiler flags corresponding to the SSE instructions you
> specifically want to enable (I'm not sure what those happen to be off the
> top of my head).
>
> Hopefully that helps you get somewhere...
I was pretty suspicious of -xHost. I'm not a developer, and my experience is limited with the Intel compiler as I've just started using it this year, but I know -xSSE3 is what I've seen recommended. So I figured there might be some connection here.
The recommendations I've seen from others talking about the Intel compiler auto CPU dispatching is to start with that as a baseline, and then -ax to add specific processor sets that you'd like to include.
Thank you for your response. I had a vague hunch that this was something that maybe just never got finished properly. Does anyone happen to know if SSE_TYPES is just something used by Amber, or if it's something that the Intel compiler respects?
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Tue Jan 13 2015 - 10:00:03 PST
