On Tuesday 15 September 2009 18:37:08 Dr. David Kirkby wrote: > Jason Moxham wrote: > > Thanks , but I mean the machine on which the original binarys were > > compiled on , I assume it was some box on skynet? If my fix is correct > > then gcc on that particular machine has been compiled with the default > > cpu set too high. > > I can't say for Linux, but I'm not aware there is any 'default CPU > setting' in gcc.
There used to be , at least this is how I remember it , although this is from 8+ years ago. From what I can remember there is no setting for running gcc , but it is set when you compile the gcc binarys , so gcc with no args will default to whatever was set at compile time. > Certainly on SPARC, gcc will not make use of any of the > sparcv9 instructions, even if it runs on a machine supporting them. They > have to be enabled. > > I suspect the problem is that programs like 'atlas' and 'mpir' poke > around at a low-level, find out what instructions the machine supports, > then build themselves optimally for that CPU. I somewhat doubt the > problems is anything to do with how gcc was compiled. > > I think the best fix would be to look at the source of ATLAS and MPIR > and see if we can change it in some way so that this default behavior > can be overridden. > MPIR(fat build) is built with no options just the default gcc , and this is what I'm told the live dvd uses. This is all a guess , until I can reproduce it. > It would be good if programs like 'mpir' and 'atlas' could take a > configure option something like > > ./configure --prefix=/usr/local --assume-minimal-instruction-set > MPIR does have these options , either they dont work or are used wrong or the OS setup is wrong , but without a access to a pair of machines that have this problem , I can only speculate. > If that could be done, then we would need to have a way in Sage of > setting those when building binaries. > > We could request such an option was added to mpir and atlas. There may > already be one, but we are not using it when building binaries. > > There is another issue which must be considered when building binaries. > Building them on version 'n' of an operating system does not guaranatee > they will work on 'n-1'. On Solaris, building on a late release of > Solaris 10, does not guarantee the code will run on the previous release > of Solaris 10, though it probably will in most cases. Hence platforms > used for creating binaries should be running old versions of the OS. > Yet another issue , who and what machine builds them? > Dave > > --~--~---------~--~----~------------~-------~--~----~ To post to this group, send an email to sage-devel@googlegroups.com To unsubscribe from this group, send an email to sage-devel-unsubscr...@googlegroups.com For more options, visit this group at http://groups.google.com/group/sage-devel URL: http://www.sagemath.org -~----------~----~----~----~------~----~------~--~---
