[Numpy-discussion] 2-d in-place operation performance vs 1-d non in-place

[George Sakkis]

I was surprised to see that an in-place modification of a 2-d array
turns out to be slower from the respective non-mutating operation on 1-
d arrays, although the latter creates new array objects. Here is the
benchmarking code:

import timeit

for n in 10,100,1000,10000:
   setup = 'from numpy.random import random;' \
           'm=random((%d,2));' \
           'u1=random(%d);' \
           'u2=u1.reshape((u1.size,1))' % (n,n)
   timers = [timeit.Timer(stmt,setup) for stmt in
       # 1-d operations; create new arrays
       'a0 = m[:,0]-u1; a1 = m[:,1]-u1',
       # 2-d in place operation
       'm -= u2'
   ]
   print n, [min(timer.repeat(3,1000)) for timer in timers]


And some results (Python 2.5, WinXP):

10 [0.010832382327921563, 0.0045706926438974782]
100 [0.010882668048592767, 0.021704993232380093]
1000 [0.018272154701226007, 0.19477587235249172]
10000 [0.073787590322233698, 1.9234369172618306]

So the 2-d in-place modification time grows linearly with the array
size but the 1-d operations are much more efficient, despite
allocating new arrays while doing so. What gives ?

George