Improved Python scientific calculation speed
ftigeot at wolfpond.org
Wed Feb 21 15:08:26 PST 2018
I have recently discovered the default packaging options for scientific
software in dports were far from optimal.
math/py-numpy used an old, slow implementation of BLAS by default.
The faster implementation I have replaced it with, OpenBLAS, was
built without any MP support. Only one core was being used for calculations
with Keras, Theano, etc...
These two commits improve performance tremendously with modern hardware:
The changes should trigger down to dports and packages in a few days.
Some custom py-numpy code went from ~= 2000 to ~= 600 seconds per run on a
6-core Coffeelake desktop PC.
More information about the Users