diezguerra/pycon-spain-swd-talk

Slides and code for the 2014 PyCon Spain Speed without Drag talk

pycon-spain-swd-talk's Issues

A possible fix for the numba version

I've tracked down the problem with the numba version, that causes the fallback to nopython mode "ruining" performance.

In the test slide you run it feeding the list of tuples. In order to perform better you have to feed the numpy array version (a list will force "object mode", while the numpy array is supported in "no python" mode -the fast backend).

A slight modification in the code is also required, as right now numba doesn't support partial indexing (indexation returning an array) in no-python mode. So changing the "numba" cell to:

import numba as nb

@nb.jit
def area_nb(points):
    acc = 0
    for i in xrange(len(points) - 1):
        acc += points[i, 0] * points[i + 1, 1] - points[i + 1, 0] * points[i, 1]
    return abs(acc) / 2

area_np_nb = nb.jit(area_np)

and using the numpy version of polygons:

print " Python Area:".ljust(21),
%timeit area(polygon)
print "Python Area Numba:".ljust(20),
%timeit area_nb(polygon_np)

print "Python NumPy:".ljust(20),
%timeit area_np(polygon_np)
print "Python NumPy Numba:".ljust(20),
%timeit area_np_nb(polygon_np)

yields a result more in line with expectations. In my dev machine:

  Python Area:        1 loops, best of 3: 281 ms per loop
 Python Area Numba:  1000 loops, best of 3: 1.31 ms per loop
 Python NumPy:       100 loops, best of 3: 4.76 ms per loop
 Python NumPy Numba: 1 loops, best of 3: 4.77 ms per loop

Note that the NumPy Numba version could be tweaked to work, but it will require external
allocation of some temporal arrays and won't be worth the effort, IMO.

Note that in order to make sure that a given function is compiled in no-python mode in the current numba version (0.15.1) you can use the njit decorator (or pass a nopython=True keyword argument to jit). In that case, if it fails to compile in the fast mode an exception will be raised pointing at the root cause of the failure (in a rather cryptic way, but this should improve in incoming versions).