[SciPy-User] building numpy + scipy with MKL 10.3 and icc + ifort 12.1

[Steve Schmerler]

Hello

I built numpy 1.6.2 and scipy 0.10.1 with MKL 10.3, using ifort+icc or
ifort+gcc, following [1,2] on a Red Hat based box. Build details are at
the end of this mail.

numpy
-----
Using gcc, all tests pass. Using icc, I get some messages like::

    .../numpy/core/tests/test_umath.py:570: RuntimeWarning: invalid value encountered in fmax
    assert_equal(np.fmax(arg1, arg2), out)

all in test_umath.py. No test fails however. Is this a problem?

scipy
-----
With both gcc and icc, I get 3 errors. 

    ======================================================================
    ERROR: test_simple_fat (test_decomp.TestRQ)
    ----------------------------------------------------------------------
    Traceback (most recent call last):
      File "/home/schmerler/soft/lib/python2.7/site-packages/scipy/linalg/tests/test_decomp.py", line 1155, in test_simple_fat
        r,q = rq(a)
      File "/home/schmerler/soft/lib/python2.7/site-packages/scipy/linalg/decomp_qr.py", line 293, in rq
        rq, tau, work, info = gerqf(a1, lwork=lwork, overwrite_a=overwrite_a)
    error: (lwork>=n||lwork==-1) failed for 1st keyword lwork: dgerqf:lwork=2

    ======================================================================
    ERROR: test_simple_trap (test_decomp.TestRQ)
    ----------------------------------------------------------------------
    Traceback (most recent call last):
      File "/home/schmerler/soft/lib/python2.7/site-packages/scipy/linalg/tests/test_decomp.py", line 1143, in test_simple_trap
        r,q = rq(a)
      File "/home/schmerler/soft/lib/python2.7/site-packages/scipy/linalg/decomp_qr.py", line 293, in rq
        rq, tau, work, info = gerqf(a1, lwork=lwork, overwrite_a=overwrite_a)
    error: (lwork>=n||lwork==-1) failed for 1st keyword lwork: dgerqf:lwork=2

and one which says ".../atlas_version.so: undefined symbol:
ATL_buildinfo" but that's probably because I don't have ATLAS.

Have others seen the TestRQ error with MKL? I haven't built against
normal blas/lapack yet to check if this may be a scipy issue.

Thanks.


----------------------------------------------------------------------
Build details
----------------------------------------------------------------------
python 2.7.1 
numpy 1.6.2 
scipy 0.10.1 
MKL 10.3,
ifort/icc 12.1 
gcc 4.1.2 (same was used to build python)

numpy site.cfg:

    [DEFAULT]
    library_dirs = /usr/lib64
    include_dirs = /usr/include
    [mkl]
    library_dirs = /cm/shared/apps/intel-cs/composer_xe_2011_sp1.9.293/mkl/lib/intel64
    include_dirs = /cm/shared/apps/intel-cs/composer_xe_2011_sp1.9.293/mkl/include
    lapack_libs = mkl_intel_lp64, mkl_sequential, mkl_core
    mkl_libs = mkl_intel_lp64, mkl_sequential, mkl_core, mkl_mc3, mkl_def

modified:
    numpy/distutils/intelccompiler.py (IntelEM64TCCompiler)
        icc -O2 -xhost -fPIC -fp-model strict -fomit-frame-pointer
    numpy/distutils/fcompiler/intel.py
        ifort -O2 -xhost -fp-model strict -fPIC

install (numpy,scipy):
    gcc + ifort
        python setup.py build --fcompiler=intelem
    icc + ifort
        python setup.py build --compiler=intelem --fcompiler=intelem

[1] http://software.intel.com/en-us/articles/numpy-scipy-with-mkl/
[2] http://www.scipy.org/Installing_SciPy/Linux#head-0b5ce001569a20ddbbdb2187578000372a09acb1

best,
Steve