Order-of-magnitude slowdown in FIAT

[Andrew McRae]

Hello all,

We at Firedrake recently incorporated the changes between 5500635 and ec68699 into our fork of FIAT (i.e. all of Nico/Aslak's recent changes)

After running our test-suite through cProfile, the time spent in FIAT's polynomial_set init or 'below' jumped from 1.28% to 18.21%, accompanied by a 15-20% increase in the total test time. Nearly all the time is spent in calls to numpy_lambdify and calls to form_derivative, both in expansions.py.

Any idea what's happened?

[Marie Elisabeth Rognes]

@amcrae Is this still a problem?

[Andrew McRae]

Hi Marie,

I have no reason to think this has been fixed. I can try 'benchmarking' again tomorrow. Our test suite has changed over time so it won't be like-for-like, but it should still be obvious whether it's O(1%) or O(10%).

[Andrew McRae]

FWIW, convenience link: https://bitbucket.org/fenics-project/fiat/branches/compare/ec68699..5500635

[Andrew McRae]

Yep, still present. There's something crazily expensive about the numpy_lambdify, it seems.

[Andrew McRae]

Actually that's slightly unfair, basically all that time is spent inside sympy's lambdify, it seems. And there must be a lot of "stringifying" of arrays going on, because 85% of that time is spent in "printer:231:doprint".

[Miklós Homolya]

If it has to do something with numpy_lambdify and sympy, then Lawrence's patch from #8 might help?