Re: [sage-support] CPU flags PNI and "SSSE3" for Ubuntu binaries
William Stein
>
> Hi folks,
>
> In IRC, there was a report about the "SSSE3" and PNI flags for the
> following Ubuntu binary:
>
> sage-4.0.2-linux-Ubuntu_9.04-i686-Linux.tar.gz
>
> Here's the IRC log:
>
> 15:39 < nifgraup> Hi,I would like to report a packaging/build bug in sage
> 4.0.2, can someone guide me?
> 15:45 < nifgraup> I'm running ubuntu 9.04 (kubuntu) and the
> sage-4.0.2-linux-Ubuntu_9.04-i686-Linux package
> 15:46 < nifgraup> and sage complains:
> 15:46 < nifgraup> The following processor flags were on the build machine but
> are not on this computer:
> 15:46 < nifgraup> pni ssse3
> 15:47 < nifgraup> the sage-4.0.2-linux-Ubuntu_9.04-i686-Linux package should
> not be built with these flags
>
> We've had problems with the SSE3 flag before.
Unfortunately, there has never ever been any Sage binary that was
successfully built without pni + ssse3 flags and that was built on
sage.math (the compile farm that builds all the sage binaries). At no
point were such binaries ever released. I thought that the binaries
that were explicitly labeled "sse2" that we released for sage-3.4.2
were so built but that wasn't the case as it turned out with testing.
Fixing this is on the todo list, and I hope it happens for sage-4.1.1.
Mostly atlas and mpir are the only two packages that "cause trouble",
though I'm not convinced of that -- I tried building a binary on
sage.math (on a 32-bit vmware machine), then putting it on an *old*
Pentium withouts ssse3 pni, and it of course wouldn't start. Then I
rebuilt atlas and mpir, and it seemed to work fine. I then ran the
test suite, which did *not* come close to passing -- so other things
evidently fail. If anybody wants to help with this, the plan is the
following.
1. Fix the atlas and mpir spkg's so they definitely support the
SAGE_FAT_BINARY environment variable (this might be the case already).
2. Build sage binary using the SAGE_FAT_BINARY variable set on a
modern 32-bit linux box.
3. Move that binary to an old 32-bit linux box.
4. Try it out -- if it fails to start then the SAGE_FAT_BINARY flag
doesn't work. Go to 1.
5. If Sage starts, try "random_matrix(ZZ,300).det()", which will test
ATLAS. If this fails, go to 1.
6. Run the full Sage test suite. Inspect and see what is still built
using ssse3/pni, etc. Go to 1.
The relevant trac ticket is here: http://sagetrac.org/sage_trac/ticket/6246
-- William