CODEWITHSUNDEEP
pythonnumba
Dave31415
Dave31415

Reputation: 2946

numba doesn't appear to work

Here is some code that I wrote to test numba

import time
from numba import autojit

@autojit
def triple_count(num):
    start = time.time()
    count=0.0
    r=range(num)
    for i in r:
        for j in r:
            for k in r:
                count += i*j*k

    runtime=time.time()-start
    rate=num/runtime 
    print "%0.4f seconds" % runtime
    print "rate : %0.3f per second" % rate
    print "count: %s" % count

The result when @autojit is commented out 

2.3796 seconds
rate : 126.073 per second
count: 9.0216784125e+13

The result when using @autojit

0.0000 seconds

rate : 8065969.231 per second

count: 0.0

What is going wrong here?

Upvotes: 1

Views: 300

Answers (1)

JoshAdel
JoshAdel

Reputation: 68682

I would first start out by timing things differently, not inserting timing calls within the method and instead return the result and then time the method call:

from numba import autojit

def triple_count(num):
    count=0.0
    r=range(num)
    for i in r:
        for j in r:
            for k in r:
                count += i*j*k

    return count

triple_count_numba = autojit(triple_count)

For num=200:

In [7]: triple_count(200)
Out[7]:
7880599000000.0

In [8]:triple_count_numba(200)
Out[8]:
7880599000000.0

In [9]: %timeit triple_count(200)
1 loops, best of 3: 747 ms per loop

In [10]: %timeit triple_count_numba(200)
100 loops, best of 3: 6.59 ms per loop

So the numba produces the same answer as the vanilla python and gets it quite a bit faster. With num=500 it's 11.5 s vs 106 ms. I'm using numba 0.14.0. You could use timeit programatically in a standard python script, but I'm using it in an IPython notebook using the %magics for convenience.

Upvotes: 1

Related Questions