An idea that occurred to me after reading Fernando's email. A function could be called at numpy import time that specifically checks for the instruction set on the CPU running and makes sure that is completely covers the instruction set available through all the various calls, including to BLAS. If this kind of thing were added, numpy could fail with a loud warning rather than dying with mysterious errors later on. The trouble would seem that you can switch your BLAS shared library without re-compiling numpy, so numpy would have to do a run-time query of ATLAS, etc. for compilation issues. Which is likely library-dependent, and furthermore, not having looked into BLAS implementations, I'm not sure that (m)any of them provide such information. Do they? Is this idea technically possible?
-Andrew Fernando Perez wrote: > On Dec 10, 2007 4:41 PM, Robert Kern <[EMAIL PROTECTED]> wrote: > >> The current situation is untenable. I will gladly accept a slow BLAS for an >> official binary that won't segfault anywhere. We can look for a faster BLAS >> later. > > Just to add a note to this: John Hunter and I just finished teaching a > python workshop here in Boulder, and one attendee had a recurring > all-out crash on WinXP. Eventually John was able to track it to a bad > BLAS call, but the death was an 'illegal instruction'. We then noticed > that this was on an older Pentium III laptop, and I'd be willing to > bet that the problem is an ATLAS compiled with SSE2 support. The PIII > chip only has plain SSE, not SSE2, and that's the kind of crash I've > seen when accidentally running code compiled in my office machine (a > P4) on my laptop (a similarly old PIII). > > It may very well be that it's OK to ship binaries with ATLAS, but just > to build them without any fancy instruction support (no SSE, SSE2 or > anything else of that kind, just plain x87 code). > > > Cheers, > > f > _______________________________________________ > Numpy-discussion mailing list > Numpy-discussion@scipy.org > http://projects.scipy.org/mailman/listinfo/numpy-discussion _______________________________________________ Numpy-discussion mailing list Numpy-discussion@scipy.org http://projects.scipy.org/mailman/listinfo/numpy-discussion
