Question

I want to speed up the following code using cython:

class A(object):
    cdef fun(self):
        return 3


class B(object):
    cdef fun(self):
        return 2

def test():
    cdef int x, y, i, s = 0
    a = [ [A(), B()], [B(), A()]]
    for i in xrange(1000):
        for x in xrange(2):
            for y in xrange(2):
                s += a[x][y].fun()
    return s

The only thing that comes to mind is something like this:

def test():
    cdef int x, y, i, s = 0
    types = [ [0, 1], [1, 0]]
    data = [[...], [...]]
    for i in xrange(1000):
        for x in xrange(2):
            for y in xrange(2):
                if types[x,y] == 0:
                   s+= A(data[x,y]).fun()
                else:
                   s+= B(data[x,y]).fun() 
    return s

Basically, the solution in C++ will be to have array of pointers to some base class with virtual method fun(), then you could iterate through it pretty quickly. Is there a way to do it using python/cython?

BTW: would it be faster to use numpy’s 2D array with dtype=object_, instead of python lists?

Answer 1

Looks like code like this gives about 20x speedup:

import numpy as np
cimport numpy as np
cdef class Base(object):
    cdef int fun(self):
        return -1

cdef class A(Base):
    cdef int fun(self):
        return 3


cdef class B(Base):
    cdef int fun(self):
        return 2

def test():
    bbb = np.array([[A(), B()], [B(), A()]], dtype=np.object_)
    cdef np.ndarray[dtype=object, ndim=2] a = bbb

    cdef int i, x, y
    cdef int s = 0
    cdef Base u

    for i in xrange(1000):
        for x in xrange(2):
            for y in xrange(2):
                u = a[x,y]                
                s += u.fun()
    return s

It even checks, that A and B are inherited from Base, probably there is way to disable it in release builds and get additional speedup

EDIT: Check could be removed using

u = <Base>a[x,y]