On Fri, Jan 30, 2009 at 11:50 AM, Carl Witty <carl.wi...@gmail.com> wrote:
>
> We have a problem finding and maintaining machines old enough to not
> have SSE3, etc., so that we can reliably build Sage binaries that
> don't rely on SSE3.
>
> Evidently, vmware lets you mask CPUID results, to hide particular CPU
> features from the guest:
> http://pubs.vmware.com/vi3/bsa/wwhelp/wwhimpl/common/html/wwhelp.htm?context=bsa&file=c11changevm.html.html
>
> Maybe we could use that as an easy way to be able to build no-SSE3
> binaries.
>
> I see two problems with the idea: 1) it's poorly documented; somebody
> would have to figure out which bits to flip in the cpuid results.
>
> 2) I'm guessing this doesn't actually disable the feature.  If any
> part of Sage does SSE3 detection by trying to run SSE3 code and seeing
> if it crashes (instead of using the cpuid instruction), then the
> technique would fail.

This is certainly worth trying.

It is a very high priority, IMHO, that we figure out how to build
binaries on the
vmware "farm" (which runs on boxen.math) that don't require these cpu
instructions.
I see two ways to do that:

 (1) What you suggest above using vmware, which might work. Note that
vmware server, which I run on boxen.math, doesn't have the options you
suggest modifying.  However, I bet one can change them anyways through
a text editor by directly editing the vm configuration.

 (2) Pass options to libraries that actually try to use these cpu
flags -- e.g., ATLAS, GMP, etc. -- to tell them not to.  Then use
objdump (?) to detect whether we succeeded.

I like (2) better if it can be made to work.  Mabshoff claimed he
could make it work in a few days, so maybe it is possible.

 -- William

--~--~---------~--~----~------------~-------~--~----~
To post to this group, send email to sage-devel@googlegroups.com
To unsubscribe from this group, send email to 
sage-devel-unsubscr...@googlegroups.com
For more options, visit this group at http://groups.google.com/group/sage-devel
URLs: http://www.sagemath.org
-~----------~----~----~----~------~----~------~--~---

Reply via email to