Loops over list 20x faster than over ndarray

Question

I can't figure out why among the following loop the f?c are 20 times slower than the f? .

I understand type definition allows Cython to leverage C speed.

What am I missing here?

Thanks

%%cython
import numpy as np
cimport numpy as np


cpdef f1(l):
    for k in l:
        k

cpdef f1c(np.ndarray npl):
    cdef int i = 0
    for i in range(npl.shape[0]):
        npl[i]

cpdef f2(n):
    for k in n:
        k

cpdef f2c(np.ndarray npn):
    cdef int i = 0
    for i in range(npn.shape[0]):
        npn[i]

and the timings:

l = ["lol"]*100000
npl = np.array(l, dtype=np.str)

n = [1]*100000
npn = np.array(n, dtype=np.int)



%timeit f1(l)

%timeit f1c(npl)

%timeit f2(n)

%timeit f2c(npn)

1000 loops, best of 3: 484 µs per loop
100 loops, best of 3: 13.1 ms per loop
1000 loops, best of 3: 483 µs per loop
100 loops, best of 3: 12.4 ms per loop

Answer 1

The loop over the numpy is at least one order of magnitude faster when you specify the data type and the number of dimensions of the array:

def f2c(np.ndarray[np.int_t, ndim=1] npn):
    cdef int i = 0
    for i in range(npn.shape[0]):
        npn[i]

similarly, I got a two times faster loop for the string case:

def f1c(np.ndarray[object, ndim=1] npl):
    cdef int i = 0
    for i in range(npl.shape[0]):
        npl[i]

where in this case you must use:

npl = np.array(l, dtype=object)