Coroutines in numba

问题

I'm working on something that requires fast coroutines and I believe numba could speed up my code.

Here's a silly example: a function that squares its input, and adds to it the number of times its been called.

def make_square_plus_count():
    i = 0
    def square_plus_count(x):
        nonlocal i
        i += 1
        return x**2 + i
    return square_plus_count

You can't even nopython=False JIT this, presumably due to the nonlocal keyword.

But you don't need nonlocal if you use a class instead:

def make_square_plus_count():
    @numba.jitclass({'i': numba.uint64})
    class State:
        def __init__(self):
            self.i = 0

    state = State()

    @numba.jit()
    def square_plus_count(x):
        state.i += 1
        return x**2 + state.i
    return square_plus_count

This at least works, but it breaks if you do nopython=True.

Is there a solution for this that will compile with nopython=True?

回答1:

If you're going to use a state-class anyway you could also use methods instead of a closure (should be no-python compiled):

import numba

@numba.jitclass({'i': numba.uint64})
class State(object):
    def __init__(self):
        self.i = 0

    def square_plus_count(self, x):
        self.i += 1
        return x**2 + self.i

square_with_call_count = State().square_plus_count  # using the method
print([square_with_call_count(i) for i in range(10)])
# [1, 3, 7, 13, 21, 31, 43, 57, 73, 91]

However timings show that this is actually slower than a pure python closure implementation. I expect that as long as you don't use nonlocal numpy-arrays or do operations on arrays in your method (or closure) this will be less efficient!

来源：https://stackoverflow.com/questions/41842656/coroutines-in-numba