[Numpy-discussion] Linking against MKL but still slow?

[Christoph Dann]

Dear list,

so far I used Enthoughts Python Distribution which contains a compiled
version of numpy linked against MKL. Now, I want to implement my own
extensions to numpy, so I need my build numpy on my own. So, I
installed Intel Parallel studio including MKL and the C / Fortran
compilers.

I linked against the same libraries as Enthought:

In [2]: np.show_config()
lapack_opt_info:
    libraries = ['mkl_lapack95_lp64', 'iomp5', 'mkl_def',
'mkl_intel_lp64', 'mkl_intel_thread', 'mkl_core', 'pthread']
    library_dirs = ['/opt/intel/lib/intel64', '/opt/intel/mkl/lib/intel64']
    define_macros = [('SCIPY_MKL_H', None)]
    include_dirs = ['/opt/intel/include/']
blas_opt_info:
    libraries = ['iomp5', 'mkl_def', 'mkl_intel_lp64',
'mkl_intel_thread', 'mkl_core', 'pthread']
    library_dirs = ['/opt/intel/lib/intel64', '/opt/intel/mkl/lib/intel64']
    define_macros = [('SCIPY_MKL_H', None)]
    include_dirs = ['/opt/intel/include/']
lapack_mkl_info:
    libraries = ['mkl_lapack95_lp64', 'iomp5', 'mkl_def',
'mkl_intel_lp64', 'mkl_intel_thread', 'mkl_core', 'pthread']
    library_dirs = ['/opt/intel/lib/intel64', '/opt/intel/mkl/lib/intel64']
    define_macros = [('SCIPY_MKL_H', None)]
    include_dirs = ['/opt/intel/include/']
blas_mkl_info:
    libraries = ['iomp5', 'mkl_def', 'mkl_intel_lp64',
'mkl_intel_thread', 'mkl_core', 'pthread']
    library_dirs = ['/opt/intel/lib/intel64', '/opt/intel/mkl/lib/intel64']
    define_macros = [('SCIPY_MKL_H', None)]
    include_dirs = ['/opt/intel/include/']
mkl_info:
    libraries = ['iomp5', 'mkl_def', 'mkl_intel_lp64',
'mkl_intel_thread', 'mkl_core', 'pthread']
    library_dirs = ['/opt/intel/lib/intel64', '/opt/intel/mkl/lib/intel64']
    define_macros = [('SCIPY_MKL_H', None)]
    include_dirs = ['/opt/intel/include/']


and used the intel compilers to build numpy. I have an Intel i7
processor and compile including AVX instructions. Yet, EPD is double
as fast as my own build executing the simple benchmark from:
http://dpinte.wordpress.com/2010/01/15/numpy-performance-improvement-with-the-mkl/
I expected at least comparable performance. How is such a decrease
possible? Did I miss a significant part to make numpy really fast?

Thanks
Christoph