[NumPy-Tickets] [NumPy] #1504: Numpy Can't Link Against 64 bit threaded MKL libraries

[NumPy Trac]

#1504: Numpy Can't Link Against 64 bit threaded MKL libraries
-----------------------------------+----------------------------------------
 Reporter:  SevenThunders          |       Owner:  somebody
     Type:  defect                 |      Status:  new     
 Priority:  normal                 |   Milestone:  2.0.0   
Component:  Other                  |     Version:  1.4.0   
 Keywords:  mkl undefined symbols  |  
-----------------------------------+----------------------------------------
 The latest MKL libraries (> 10.2) have somewhat complicated link
 requirements.  They require grouping flags and now apparently generate
 runtime errors if one tries to use numpy (e.g. dot()) after linking
 against the dynamic libraries in linux.
 I'm getting the errors described here:
 http://software.intel.com/en-us/forums/showthread.php?t=74037

 The claim is that the dynamic loading of the mkl libraries is not in the
 correct order and cause undefined symbols errors.  The proposed solution
 is to link using static libraries,  but the current build scheme does not
 offer that kind of control.

 I thought perhaps the scons fork might get me there,  but I currently
 can't install that  because it seems to require a development version of
 NumScons  (0.11 ?)  for which there are no downloadable binaries/packages.

 Seeing as how mkl is the currently the best way to get good numpy
 performance on Intel machines,  I would think this would be a fairly high
 priority item.  I know I'm not the only one with this problem.

-- 
Ticket URL: <http://projects.scipy.org/numpy/ticket/1504>
NumPy <http://projects.scipy.org/numpy>
My example project