How to build on Solaris 10 (x86) using sunperf?
Hi, I have been trying to build numpy on an opteron based Solaris 10 machine, with poor results. After skimming over the discussion archives I noticed a branch for numpy.sunperf in the svn repository. Unfortunately this is no longer available, so I was curious as to the best approach to building numpy on my problem platform. Thanks, Pete
Peter Ward wrote:
Hi, I have been trying to build numpy on an opteron based Solaris 10 machine, with poor results. After skimming over the discussion archives I noticed a branch for numpy.sunperf in the svn repository. Unfortunately this is no longer available, so I was curious as to the best approach to building numpy on my problem platform.
In general, the key to building on a platform is the config.h file that numpy generates. It knows how to do it for a few platforms, but for others it may not work. In particular, we need to know how to find: 1) float and long double versions of functions 2) whether or not and how IEEE 754 floating point signalling is handled on your platform. There are a few other things that must be done. You can generate the config.h file by hand and then figure out how to get the build scripts to do it automatically. Start with a config.h file from another platform and start tracking through the error messages. It is a tedious process. If you post error messages (or file them on the Trac page for NumPy), we may be able to help. -Travis O.
Travis E. Oliphant wrote:
Peter Ward wrote:
Hi, I have been trying to build numpy on an opteron based Solaris 10 machine, with poor results. After skimming over the discussion archives I noticed a branch for numpy.sunperf in the svn repository. Unfortunately this is no longer available, so I was curious as to the best approach to building numpy on my problem platform.
In general, the key to building on a platform is the config.h file that numpy generates. It knows how to do it for a few platforms, but for others it may not work.
In particular, we need to know how to find:
1) float and long double versions of functions 2) whether or not and how IEEE 754 floating point signalling is handled on your platform.
There are a few other things that must be done. You can generate the config.h file by hand and then figure out how to get the build scripts to do it automatically. Start with a config.h file from another platform and start tracking through the error messages.
The current config.h works fine for solaris with Sun compilers, in my experience, so the problem must be somewhere else. Peter, could you post the errors you got ? As an alternative, I am working on an alternative build system for numpy: it should work on solaris (tested on Indiana, and other people reported success on solaris 9 and 10). Unfortunately, I have been redesigning the internals quite a lot lately, and I have not tested the changes on solaris yet. If you are willing to test, it should be easy for me to make it work again on solaris in a few minutes, though. cheers, David
On Jan 23, 2008 12:32 PM, David Cournapeau <david@ar.media.kyoto-u.ac.jp> wrote:
The current config.h works fine for solaris with Sun compilers, in my experience, so the problem must be somewhere else.
I've just tried numpy SVN with sun compiler (Sun studio 12) on Indiana, and it works fine, so there is no recent regressions on this. The problem may be because of opteron (64 bits). cheers, David
participants (4)
-
David Cournapeau -
David Cournapeau -
Peter Ward -
Travis E. Oliphant