AtomGroup.rotate is slow

[mimischi]

Expected behaviour

Using universe.atoms.rotate(rotation_matrix) should be as fast as np.dot(universe.atoms.postions, rotation_matrix).

Actual behaviour

The rotate method is almost 7x slower than np.dot. In the test system it's almost unnoticeable (70μs vs 490μs), but in my production system it's a difference of ~30 seconds (with numpy) and ~2m30s (with MDAnalysis).

Code to reproduce the behaviour

See this notebook.

In my production system I first translate and then rotate my system. To isolate the exact problem, I've used all variations: rotate only, translate only and rotate+translate. Also I've done the tests on both single frames and while iterating over all frames.

Currently version of MDAnalysis:

0.16.2

[richardjgowers]

One difference between the numpy and mda versions of things is MDA ensures that the positions it works on are all unique. It would be interesting to see if commenting out the uniqueness check in MDA improves performance a lot for your large system

[kain88-de]

what position uniqueness check?

[richardjgowers]

@kain88-de unique atoms more than positions I guess https://github.com/MDAnalysis/mdanalysis/blob/develop/package/MDAnalysis/core/groups.py#L827

this line looks needlessly slow too, no reason to go via atoms.unique, could just be x = self.universe....

[kain88-de]

I was wondering about the use of unique here anyway.