compiling parallelized cython with Clang

Question

I am trying to parallelize a for loop in cython using prange. My setup file is pretty basic:

from distutils.core import setup
from Cython.Build import cythonize
setup(ext_modules=cythonize('my_cython_code.pyx'))

So I am using -fopenmp as a cython compiler flag, which I accomplish by having the first two lines of the my_cython_code module be:

# distutils: extra_compile_args = -fopenmp
# distutils: extra_link_args = -fopenmp

However, my machine runs on OSX, and the mac gcc compiler is essentially Clang, which does not support OpenMP. This means that when I try to build my parallelized cython extension module, I get a fatal compiler error:

ld: library not found for -lgomp
clang: error: linker command failed with exit code 1 (use -v to see invocation)
error: command 'gcc' failed with exit status 1

Is there a way around this problem? How can I use a different compiler that will accept OpenMP when building my extension? Are there other approaches that will let me parallelize cython for loops with standard mac architecture?

debsankha · Accepted Answer · 2022-04-03 17:44:10Z

3

After a lot of trial-and-error, I made cython code to actually parallelize on an M1 mac. The crucial step was supplying both -lomp and -lmpi flags during linking.

Install llvm using homebrew since the version bundled by apple doesn't support openmpi:

brew install llvm open-mpi libomp

The flags we need to pass to clang in order to compile openmpi code can fortunately be found out by running the command mpicc -showme:

$ mpicc -showme
clang -I/opt/homebrew/Cellar/open-mpi/4.1.2/include -L/opt/homebrew/Cellar/open-mpi/4.1.2/lib -L/opt/homebrew/opt/libevent/lib -lmpi

Then, we can set the environment variables CPATH and LIBRARY_PATH accordingly:

export LIBRARY_PATH=/opt/homebrew/Cellar/open-mpi/4.1.2/lib:/opt/homebrew/opt/libevent/lib
export CPATH=/opt/homebrew/Cellar/open-mpi/4.1.2/include

The setup.py must specify extra_compile_args and extra_link_args:

setup(
    ext_modules = cythonize([
        Extension(
            'foo',
            ['foo.pyx'],
            extra_compile_args=['-fopenmp'],
            extra_link_args=['-lmpi', '-lomp'],
        )
    ]
)

Here is a full MWE.

edited Apr 3, 2022 at 17:44

answered Apr 3, 2022 at 17:35

debsankha

313 bronze badges

Sign up to request clarification or add additional context in comments.

2 Comments

DavidW Over a year ago

I believe OpenMPI is pretty different OpenMP. It's more to do with distributed computing where the cores can't see each other. I suspect you can remove it and your example would still work.

Jeremy Caney Over a year ago

Please be sure to read Stack Overflow's self-promotion policy when referencing your own content.

Jeff Hammond · Accepted Answer · 2015-01-18 22:56:35Z

2

As you state, you are using a compiler that does not support OpenMP. You need to use one that does. Please look up how to install GCC with Macports or Homebrew if you want to use proper GCC that supports OpenMP.

1 Comment

arokem Over a year ago

For homebrew use: brew reinstall gcc --without-multilib. Be sure to set your bash path to pick this newly installed gcc. In my case I set: ln -s /usr/local/Cellar/gcc/5.2.0/bin/gcc-5 ~/usr/bin/gcc (and I have export PATH="/Users/arokem/usr/bin:/usr/local/sbin:$PATH" in my bash_profile)

Alexander Ponamarev · Accepted Answer · 2019-01-11 21:14:54Z

Apparently, Apple dropped the support of OpenMP sometime ago, therefore, you cannot compile the code that includes this dependency with a standard gcc. A good way to get around that is to install LLVM and compile with it. Here is the sequence that worked for me:

Install LLVM:

brew install llvm

Include OpenMP flags(-fopenmp -lomp) to setup.py. -lomp flag is not described anywhere for some reason, but it is important for a correct compilation:

from distutils.core import setup
from distutils.extension import Extension
from Cython.Build import cythonize, build_ext


exts = [Extension(name='name_your_module',
                  sources=['your_module.pyx'],
                  extra_compile_args=['-fopenmp'],
                  extra_link_args=['-lomp']
                  )]

setup(name = 'name_your_module',
      ext_modules=cythonize(exts,
      cmdclass={'build_ext': build_ext})

And then compile the code with LLVM:

CC=/usr/local/opt/llvm/bin/clang++ python setup.py build_ext --inplace

This should result in a parallelized .so

Collectives™ on Stack Overflow

compiling parallelized cython with Clang

3 Answers 3

2 Comments

1 Comment

Comments

Your Answer

Hot Network Questions

Collectives™ on Stack Overflow

3 Answers 3

2 Comments

1 Comment

Comments

Your Answer

Sign up or log in

Post as a guest

Related