Matthew Brett email@example.com wrote:
Meanwhile Sturla kindly worked up a patch to numpy to work round the Accelerate segfault . I haven't tested that, but given I'd already built the wheels, I prefer the ATLAS builds because they work with multiprocessing.
It is an ugly hack. If it is used it would be better to keep it in a separate branch rather than in NumPy master.
I am not able to post a PR because my fork of NumPy is unavailable on GitHub.
I propose uploading these wheels as the default for numpy and scipy. Does anyone have any objection or comments before I go ahead and do that?
ATLAS should be fine. Performance will not be anything like the alternatives (Accelerate and OpenBLAS), but correctness is more important. I am not sure why ATLAS is preferred over OpenBLAS, though, but I don't really care.
Perhaps we could ask Intel permission to use MKL in the binary wheels? The Accelerate vs. multiptocessing issue will never be fixed for Python 2.7 and poissibly Python 3.3 (it is fixed in Python 3.4). The error is in multiprocessing, so Apple will not patch Accelerate. New features will not be added to Python 2.7 standard lib, so there is no chance of getting the required fix backported there either. Stalemate.