Shouldn't preallocation of arrays in numpy be faster?

Question

I am confused why test2 is not faster than test1 in the following code:

import timeit

setup = """
import numpy as np
A = np.ones((220, 220, 220))
B = np.ones((220, 220, 220))

class store:
    def __init__(self):
        self.C = np.empty((220, 220, 220))
Z = store()
"""

test1 = """
C = A + B
"""

test2 = """
Z.C = A + B
"""

print timeit.timeit(test1, setup, number=1000)
print timeit.timeit(test2, setup, number=1000)

which gave me: 40.9241290092 40.7675480843

I thought because ZC was preallocated memory, there would be less overhead every time I added A+B and needed a place to store it, ie less calls to malloc behind the scenes or something like that. What am I missing?

Answer 1

Allocation is a fast operation, addition is more expensive:

In [7]: %timeit  np.empty((220, 220, 220))
1000 loops, best of 3: 472 µs per loop

In [8]: u= np.ones((220, 220, 220))

In [9]: %timeit  u+u
10 loops, best of 3: 73.5 ms per loop 

So, even if you correctly update your array ( ZC[:]= A+B ) you will not win a lot in your case (~0.5%).