Re: [f2py] OpenMP - again

Pearu Peterson <[email protected]> Mon, 6 Sep 2010 20:46:28 +0300
Newsgroups gmane.comp.python.f2py.user
Message-ID <[email protected]>
On Sun, Sep 5, 2010 at 1:55 PM, Paul Anton Letnes
<[email protected]> wrote:
> Hi again!
>
> On 4. sep. 2010, at 15.48, Pearu Peterson wrote:
>
>> Here follows a solution when using setup.py file.
>> So, create a setup.py file containing
>>
>> #!/usr/bin/env python
>> def configuration(parent_package='',top_path=None):
>>    from numpy.distutils.misc_util import Configuration
>>    config = Configuration(None,parent_package,top_path)
>>    config.add_extension('wrapper',
>>                         ['fortranmodule.f95','wrapper.f95'],
>>                         libraries = ['gomp','blas'],
>>                         library_dirs = ['/opt/local/lib/gcc44'],
>>                         f2py_options = [],
>>                         define_macros = [('F2PY_REPORT_ON_ARRAY_COPY','1')],
>>                         extra_link_args = ['-fopenmp']
>>                         )
>>
>>    return config
>>
>> if __name__ == "__main__":
>>    from numpy.distutils.core import setup
>>    setup(configuration=configuration)
>> #eof
>>
>> and then run
>>
>> python setup.py build_ext --inplace --fcompiler=gnu95
>>
>> HTH,
>> Pearu
>
> Thanks for your help there - I am closer to a workable solution now. There are still a few quirks, however. I attach my source files and the output (in output.txt) I get when running "make clean; make test" in the terminal.
>
> Questions:
>
> Why do I get this warning? fortranmodule.f95 is a pure fortran module, and can be compiled as such; hence, real(wp) should work (fortranmodule.f95 line 8 and 9).
> Warning:
> getctype: "real(kind=wp)" not supported (use .f2py_f2cmap).

See

  http://cens.ioc.ee/projects/f2py2e/FAQ.html#q-what-if-fortran-90-code-uses-type-spec-kind-kind

for an explanation.

> Your setup.py lets the program compile and run nicely. However, there is no speedup and only 1 OMP thread running when using your setup.py. When calling setup.py with the env variable F90FLAGS=-fopenmp set, however, I do get several threads and a close to optimal speedup - see this line in Makefile:
> env F90FLAGS=-fopenmp python setup.py build_ext --inplace --fcompiler=gnu95

You can use the following command to enable -fopenmp for fortran compilation:

python setup.py config_fc --opt="-fopenmp -O3 -funroll-loops"
build_ext --inplace

(I wish I had time to make specifying extra compiler/linker options
uniform within a setup.py file as well as in f2py command line..)

> Finally, something which might be related to gfortran performance rather than f2py: I notice that on the Linux desktop which I've also been running this program, the program runs essentially only half as fast as on my laptop. CPUs are core 2 duo 2.4 GHz on my laptop and a more recent core 2 quad 2.83 GHz on the desktop. This seems rather odd to me - the  desktop should be roughly as fast "per core".
>
> If you find this example interesting enough to use in some f2py documentation, please feel free to. I guess other people might be trying to use OpenMP with fortran as well, as it's a rather simple way of parallelizing fortran code.

If you have time then could you create a wiki page, say f2py_OpenMP in
the scipy cookbook, something along the lines of the following pages:
  http://www.scipy.org/Cookbook/F2Py
  http://www.scipy.org/Cookbook/f2py_and_NumPy

Best regards,
Pearu