[SciPy-user] 64 bit support

George Nurser agn at noc.soton.ac.uk
Tue Mar 7 09:37:46 CST 2006


On 7 Mar 2006, at 00:13, Paul Ray wrote:

> Hi,
>
> Does anyone have experience building SciPy on 64 bit linux boxes?
>
> Our machine is a dual Opteron, running something like RedHat
> Enterprise 4.2:
> xxi1 : 74>uname -a
> Linux xxi1.nrl.navy.mil 2.6.9-11.ELsmp #1 SMP Thu Jun 16 11:18:13 CDT
> 2005 x86_64 x86_64 x86_64 GNU/Linux
>
> We've been running into problems with numpy/scipy looking for
> libraries in /usr/lib when it should be using /usr/lib64, so that a
> simple install with setup.py install doesn't work.  It crashes when
> you run and it loads those 32 bit libraries.  I think all I need to
> do is ensure that the library search paths include /usr/lib64 and /
> usr/X11R6/lib64 and NOT /usr/lib or /usr/X11R6/lib.  What is the easy
> or "right" way to do this?  Is there an option to setup.py?  Should I
> make a site.cfg (and where should it reside?)?  Should I edit
> system_info.py?
>
> I assume that this configuration is becoming more common as 64-bit
> processors show up in more machines, so I'm surprised that setup.py
> doesn't already figure it out and "do the right thing".  Are we doing
> something boneheaded?


I managed to get numpy/scipy/matplotlib set up on a single core (SUN)  
Opteron running a similar RedHat
uname -a
Linux nohow 2.6.9-22.0.2.EL #1 Thu Jan 5 17:03:08 EST 2006 x86_64  
x86_64 x86_64 GNU/Linux
(BTW - was the uname -a issued last summer, or is your clock wrong:)


I'm sure that I did not do it the easy way -- i used a site.cfg --  
but it worked in the end.

Since I do not have root privileges  on this computer, and in any  
case I used the acml libraries, I used a site.cfg file for numpy and  
scipy.

I believe that the code has been changed to allow this site.cfg file  
to reside either in (in decreasing order of priority)
1. the directory where you are doing setup.py from
2. your home directory $HOME
3. the numpy/distutils subdirectory of the root numpy source  
directory when installing numpy OR (when installing scipy)
  the distutils subdirectory of the *installed* numpy directory.
This copying of the site.cfg is now automatically done when numpy is  
installed.

numpy  and scipy only seemed to need the site.cfg file to find the  
acml libraries (or in your case, presumably  the Atlas libraries).  
The standard setup.py worked fine otherwise.
FFTW of course has to be set, so that $FFTW/include and $FFTW/lib  
hold the include files and fftw libraries.

For matplotlib, I did have to modfy the setupext.py file, replacing
     libdirs = [os.path.join(p, 'lib')     for p in basedir 
[sys.platform]
                if os.path.exists(p)]
  by
     libdirs = [os.path.join(p, 'lib64')     for p in basedir 
[sys.platform]
                if os.path.exists(p)] + \
	        [os.path.join(p, 'lib')     for p in basedir[sys.platform]
                if os.path.exists(p)]
in function add_base_flags (~line 104)


HTH. George.






More information about the SciPy-user mailing list