[Numpy-discussion] (apparent) infinite loop in LAPACK/ATLAS
dwf at cs.toronto.edu
Wed Nov 4 05:09:22 EST 2009
On 4-Nov-09, at 4:23 AM, David Cournapeau wrote:
> Did you compile them without any optimized as suggested in the
> makefiles ? NOOPT should not contain -O option
Yup, it contained -O0 -fPIC (-O0 I think is in fact more strict than
having no -O option?). Have you seen this problem with broken NOOPT?
I'm not certain yet, but it looks like some combination of gcc 4.4 and
the Atom microarchitecture reintroduces this old bug, as "-ffloat-
store" seems to have fixed it.
numpy.test() runs through just fine once I recompiled LAPACK and then
ATLAS as well (just rebuilding the lapack_LINUX.a and then running
'make' in the ATLAS obj dir didn't seem to do it, so I had to wait
through several hours of tuning again).
For anyone who might stumble upon this problem again, just to be safe
I added -ffloat-store to both the flag lists in LAPACK's make.inc, and
also used -Fa fi '-ffloat-store' when running configure with ATLAS.
If this is indeed either a gcc 4.4 regression or an Atom-specific
quirk it might start popping up more and more around here.
More information about the NumPy-Discussion