Float division is no longer IEEE 754 after njit

[astrojuanlu]

After applying numba.njit to a function, float division behavior is no longer that of NumPy, but the one of Python:

# coding: utf-8
import numpy as np
from numba import njit

def func_ok(a, b):
    c = np.sin(a * b)
    return np.divide(a - b, c)

numba_func_ok = njit(func_ok)


def func_fail(a, b):
    c = np.sin(a * b)
    return (a - b) / c

numba_func_fail = njit(func_fail)

if __name__ == '__main__':
    print(func_ok(0., 0.))  # nan
    print(numba_func_ok(0., 0.))  # nan
    print(func_fail(0., 0.))  # nan
    print(numba_func_fail(0., 0.))  # ZeroDivisionError

My workaround is to use numpy.divide, which luckily is supported in nopython mode.

[pitrou]

Unfortunately we can't implement both semantics at the same time. We chose to implement the Python semantics over scalars and the Numpy semantics over Numpy arrays.

[seibert]

We could provide a mechanism to disable raising zero division errors. This would also have a performance benefit for those who are willing to deal with the NaN.

[seibert]

this is now available with the error_model="python" or error_model="numpy" argument to @jit