search

joezuntz / pycamb

Python bindings for the Code for Anisotropies in the Microwave Background (CAMB)
8 stars 9 forks source link

Support for multi-threading #3

Closed steven-murray closed 10 years ago

steven-murray commented 11 years ago

Hey,

This is a nice application. I was wondering how to get it to support multi-threading using open mp? CAMB itself does this, but when I build with this setup it doesn't seem to multithread at all. I have tried using libraries = ['gomp'] in the Extension part of the setup script, to no avail.

I also added a '@ThreadNum' variable to the generatePycamb.py file to change the number of threads used. I'm probably missing something really simple!

joezuntz commented 11 years ago

Hi Steven,

You can do this by compiling with F90FLAGS=-fopenmp python setup.py install though you may need the ['gomp'] thing that you already did as well. I should add a note in the README on this!

Cheers, Joe

steven-murray commented 11 years ago

Hi thanks for your prompt reply! So I tried doing what you suggest, I also remembered just after I contacted you that the right way to link to openmp with gfortran was the -fopenmp option (rather than the -lgomp option).

But it still doesn't give me multithreading... it's very odd. Perhaps a better way is to premake the CAMB library and then link to it directly?

joezuntz commented 11 years ago

That's odd. There should be a line in the setup output somewhere that looks like: Fortran f90 compiler: /usr/local/bin/gfortran -fopenmp -m64 -fPIC -O3 -funroll-loops

Does yours include the openmp flag? Good to clean by removing the build dir first, of course. Also, you mayt need to set the environment variable OMP_NUM_THREADS=2 (e.g)

The older version of the code was built as you describe - if you go back in the git history to before rainwoodman's commits you should find the version built like that.

Joe

steven-murray commented 11 years ago

Hi, My line reads: Fortran f90 compiler: /usr/local/bin/gfortran -Wall -fno-second-underscore -m64 -fPIC -O3 -funroll-loops

ie. there's evidently no -fopenmp there. Not sure why this is. I cleaned my build directory.

I'll have a little google around see what I can dig up...

steven-murray commented 11 years ago

As usual, documentation on f2py is scarce and out of date.

Anyway, it works when I use the 'old' way. Maybe one day I'll figure out this way as well :)

steven-murray commented 11 years ago

It may be that I am using numpy v1.7.0

When I use the machine at work with numpy 1.6.1 it works fine with the flag specified as you say.

I have noticed that in the f2py of the newer numpy, the option extra_f90_compile_args is no longer available. Perhaps this has something to do with it...