Opened 5 years ago

Closed 5 years ago

#3697 closed task (fixed)

Replace non-standard Windows numpy and scipy packages with PyPi ones

Reported by: Tom Goddard Owned by: Tom Goddard
Priority: moderate Milestone: 1.2
Component: Third Party Version:
Keywords: Cc: chimera-programmers, Tristan Croll
Blocked By: Blocking:
Notify when closed: Platform: all
Project: ChimeraX

Description

For unclear reasons we are using custom compiled numpy and scipy wheels on Windows that use Intel Math Kernel Library. I think Conrad did this and said those packages were faster. They complicate our ChimeraX build and package updating and the numpy mkl wheel is very large, 200 Mbytes vs 13 Mbytes for the PyPi numpy wheel. The PyPi numpy includes OpenBLAS while the MKL numpy includes MKL libraries that are a few hundred Mbytes. I could find no quantitative evidence that numpy MKL was much faster than OpenBLAS for common numpy uses and some sites claimed OpenBLAS was faster on AMD processors.

I think we should use PyPi numpy and scipy on packages on Windows as we currently do for Mac and Linux.

Change History (1)

comment:1 by Tom Goddard, 5 years ago

Resolution: → fixed
Status: assigned → closed

Done.

With PyPi numpy and scipy ChimeraX took 16.6 second to compute 3q3g helix tubes, and took 17.1 seconds with MKL numpy and scipy.

Note: See TracTickets for help on using tickets.