Tim, On Apr 2, 2012, at 11:30 AM, Timothy Bates wrote:
> Hi All, > > HPC seem to be maintaining the gcc toolchain up to date (they have GCC 4.7 > compiled with autovectoring using OpenMP…) > > http://hpc.sourceforge.net > > BUT the page http://r.research.att.com/tools/ says "do not use compilers > from HPC, they won't work correctly!” Is that the case? > Two reasons: a) they do not use Apple's drivers, so those are incompatible with most "regular" flags on Mac OS X (including most basic ones like -arch). b) last time I checked they were broken, i.e. the distribution did not even include libraries that the compiler linked against and it had OS version issues (i.e. it worked only on a very specific version which was not even what they were advertized for). I would hope that the latter point may have been rectified in the meantime, but I don't know. Gaurav never responded to my comments so I stopped worrying about that build. (There was a point c) where his compilers don't support ppc cross-compilation but that is less relevant now). It is stil possible to build FSF gcc and Apple drivers - that's what we used a while ago when Apple's branch was broken. But note that even the most recent compilers are not much better, OMP performance is unusable for R's purpose so last time I checked there were no noticeable gains after all the work, but more recent reports are welcome. > Also, I wondered if http://www.macports.org might be the way to go to get a > version of gcc with a non-crashing OpenMP library? > MacPorts are quite notorious for the quality of the binaries and conflicts they cause, so I would be wary about that. If you compile everything from scratch (R and libraries), then the HPC compilers may work - you just have to stick to FSF flags. I am still weighting the options - the most reasonable way at the moment is clang because it is supported by Apple and under active development (personally, I have switched to clang because it's much better for development), but there is no OpenMP yet for clang, although it is (allegedly) brewing. But as I said, at least for R itself, the threading performance problem is deeper, so just updating the compiler or OMP doesn't seem to help (I didn't try MPC, though). > PS: The att.com page talks about install disks for OS X, but I think it’s all > via the app store now, including X Code. Yes, it varies by Xcode version and your OS X version. App store is the last resort, I prefer ADC which has always worked and still works. I think the FAQ is up to date. Cheers, Simon _______________________________________________ R-SIG-Mac mailing list [email protected] https://stat.ethz.ch/mailman/listinfo/r-sig-mac
