Strange behavior in the performance of operator (** p)

Numba
1

I noticed that that the operation value ** p shows huge difference in the computing time between the two cases p=2.0 and p=2. (Also: between p=3.0 and p=3, and so on.)

Minimum Working Example:

@njit(fastmath=True)
def my_func(a, p):
    n = len(a)
    x = 0
    for i in range(n):
        for j in range(n):
            x = x + (a[i] ** p) + (a[j] ** p)
    
    return x

And,
input:

n = 20000

seed = 0
np.random.seed(seed)
a = np.random.rand(n)

And, try the following code for p=2.0 and p=2

my_func(a, p)

tic = time.time()
my_func(a, p)
toc = time.time()
print(toc - tic)

And, p=2 is a lot faster (like by 80% in my pc)

And, if we remove the @njit decorator, the p=2.0 becomes (slightly) faster than p=2 !!

Can anyone help me understand what is going on here?

2

That integer exponents are faster doesn’t seem particularly surprising see a discussion like this

For plain-python, I’m not sure but I’d guess it’s because of the implementation of integer objects in python. I vaguely remember some effort to improve that for C integers in 3.11-ish

1 Like
3

Thank your response! Yeah… I was told that the integer exponents can result in unrolling a for-loop and thus it has higher speed! Thanks for link.

I was wondering if numba can detect that (i.e. check if int(p) == p, where p is exponent), and when the condition becomes true, just replace it with integer to speed up the calculation.