Wrong results when calculating power with an array

Since Numba 0.38.0, the following function leads to wrong results:

import numpy as np
import numba as nb

def test(a, b):
    return (1. / a) ** b

test_nb = nb.njit(test)

Example:

>>> test_nb(0.1 * np.ones(50), np.ones(50))
array([ 1.,  1.,  1.,  1.,  1.,  1.,  1.,  1.,  1.,  1.,  1.,  1.,  1.,
        1.,  1.,  1.,  1.,  1.,  1.,  1.,  1.,  1.,  1.,  1.,  1.,  1.,
        1.,  1.,  1.,  1.,  1.,  1.,  1.,  1.,  1.,  1.,  1.,  1.,  1.,
        1.,  1.,  1.,  1.,  1.,  1.,  1.,  1.,  1., 10., 10.])

Expected output (with NumPy):

>>> test(0.1 * np.ones(50), np.ones(50))
array([10., 10., 10., 10., 10., 10., 10., 10., 10., 10., 10., 10., 10.,
       10., 10., 10., 10., 10., 10., 10., 10., 10., 10., 10., 10., 10.,
       10., 10., 10., 10., 10., 10., 10., 10., 10., 10., 10., 10., 10.,
       10., 10., 10., 10., 10., 10., 10., 10., 10., 10., 10.])

I have only encountered this bug on Windows, and it does not occur on Numba 0.37.0. Relevant output from conda list:

llvmlite                  0.23.0                   py36_1    conda-forge
numba                     0.38.0           py36h830ac7b_0
numpy                     1.14.2           py36h5c71026_0
python                    3.6.0                         0

Let me know if you need more information.

[stuartarchibald]

@dionhaefner thanks for the report. Given what you present, please could you try setting the environment variable NUMBA_DISABLE_INTEL_SVML to non-zero and re-running. I’m suggesting this as Windows 64 bit is the only platform with SVML by default and it appeared in 0.38. Thanks.

[seibert]

Are we sure that the problem is only Windows? If the conda-forge llvmlite is installed on a Linux box with SVML present, what happens?

I’m worried that instead we need to be able to detect when llvmlite has been linked against an unpatched LLVM (regardless of platform), and then disable SVML support in that case, regardless of whether the library is present.