Paul raises good points, too. Manuel -- do you have anyone that could work on this? I'd be happy to supply the configury magic to check and see if the GCC intrinsics are available (assumedly outputing an AM_CONDITIONAL and/or AC_DEFINE to let the code base know the decision) if someone else can do the work of actually implementing the *use* of the GCC atomics. Perhaps someone might even be inspired to check out the atomics openpa, the Gasnet tools, etc....?
Our typical modus operandi for such projects is to fork off into a mercurial branch, iterate on the work there (particularly if it involves external collaborators), and then bring it back into the main OMPI SVN when done. On Jul 20, 2010, at 12:12 PM, Barrett, Brian W wrote: > Jeff - > > I think falling back to GCC built-in if available is a rational idea. We've > been using them in another project without any problems. They are > potentially a bit slower than the hand-crafted assembly because they > generally use full memory barriers when we only need read memory barriers, > but that's not a bad thing for the portability case. > > There's still an issue with the high resolution timers, but there are > actually rational fall-backs for that, so we're probably ok there. > > Brian > > On Jul 20, 2010, at 9:49 AM, Jeff Squyres wrote: > > > *** This mail mainly targeted at Brian and George *** > > > > Debian maintainer Manuel Prinz raised an idea to me this morning: > > > > The Debian community compiles and tests Debian on a huge range of hardware > > platforms. It's been a long-standing issue that Open MPI doesn't support > > all of them (e.g., MIPS, ARM, ...). Specifically, we don't have assembly > > to support all of those platforms. > > > > The Debian community asks: if building with a recent GCC on one of these > > platforms where OMPI doesn't have native assembly, can we fall back to the > > GCC intrinsic atomics? > > > > https://svn.open-mpi.org/trac/ompi/ticket/2495 > > > > Additionally, there's then OpenPA project from Argonne that supports a > > bunch of atomics on a bunch of platforms. George told me at one point that > > he didn't think it was sufficient for Open MPI's needs. Do we know if > > that's still true? > > > > -- > > Jeff Squyres > > jsquy...@cisco.com > > For corporate legal information go to: > > http://www.cisco.com/web/about/doing_business/legal/cri/ > > > > > > > -- Jeff Squyres jsquy...@cisco.com For corporate legal information go to: http://www.cisco.com/web/about/doing_business/legal/cri/
