Fixed issue 14

[oleksandr-pavlyk]

Used pthread_atfork to restore environment variable MKL_THREADING_LAYER for the forked process.

@peterbell10 @rgommers

Closes #14

Using the environment as specified in #14, and

# file iss14.py
def do_test(*args):
    import os
    print("in do_test   ", os.getenv("MKL_THREADING_LAYER", None))
    import torch; import numpy as np
    x = np.random.rand(100)
    x.dot(x)

if __name__ == '__main__':
    import multiprocessing as mp
    import os
    print("in __main__ before mkl   ", os.getenv("MKL_THREADING_LAYER", None))
    import mkl
    print("in __main__ after mkl   ", os.getenv("MKL_THREADING_LAYER", None))

    mp = mp.get_context('spawn')
    with mp.Pool(2) as p:
        p.map(do_test, range(2))

With this fix I am getting:

(debug_pytorch) [22:44:14 ansatnuc04 wd]$ python iss14.py
in __main__ before mkl    None
in __main__ after mkl    None
in do_test    None
in do_test    None

(debug_pytorch) [22:44:18 ansatnuc04 wd]$ MKL_THREADING_LAYER=GNU python iss14.py
in __main__ before mkl    GNU
in __main__ after mkl    GNU
in do_test    GNU
in do_test    GNU

[anton-malakhov]

LGTM

[rgommers]

Thanks @oleksandr-pavlyk!

[rgommers]

It would be good to know when this makes its way into conda defaults and conda-forge. For conda-forge I know where to look (https://github.com/conda-forge/mkl-service-feedstock). Not sure about defaults, except try conda search mkl-service sometime in the future. I assume this will be in the mkl-service 2.4.0 release?

[oleksandr-pavlyk]

Yes, I was thinking of releasing 2.4.0 today. I am looking through new functionality in MKL to add to service.

mkl_set_num_stripes, mkl_get_num_stripes

Also adding support for more code in mkl_enable_instructions function.

[oleksandr-pavlyk]

@rgommers Fixed build break in conda-forge/mkl-service-feedstock#17

[rgommers]

thanks!