numpy unexpectedly raises zero division error if array is long enough

Steps to reproduce:

with np.errstate(invalid='ignore', divide='raise'):
    for i in range(10000):
        z = np.zeros(i)
        try:
            z / z
        except(FloatingPointError) as exc:
            print(exc, 'at size', i)
            break

output: divide by zero encountered in true_divide at size 8001

numpy version: 1.14.2

Seems like this bug was introduced just recently, as there is no such issue with 1.13.3 and 1.14.1.

[oleksandr-pavlyk]

Nowadays implementations of universal functions that use MKL VML has all been moved to mkl_umath package.

The ufunc loops for certain types are registered with NumPy’s universal functions via PyUFunc_RegisterLoopForType.

One can deregister these loops using mkl_umath.restore().

[oleksandr-pavlyk]

Yes, it is still reproducible:

In [2]: z = np.zeros(8001)

In [3]: z/z
/tmp/miniconda3/bin/ipython:1: RuntimeWarning: divide by zero encountered in true_divide
  #!/tmp/miniconda3/bin/python
/tmp/miniconda3/bin/ipython:1: RuntimeWarning: invalid value encountered in true_divide
  #!/tmp/miniconda3/bin/python
Out[3]: array([nan, nan, nan, ..., nan, nan, nan])

It is coming from use of MKL VML function vdDiv used to perform division.

One can turn errors off by using mkl.set_vml_mode:

import mkl     # import mkl-service
saved = mkl.vml_set_mode(
    "ha", # accuracy control
     "off", # denormalized numbers handling
     "ignore" # error mode control
)

See https://software.intel.com/en-us/mkl-developer-reference-c-vmlsetmode for more details

You can restore the VML behavior with

mkl.vml_set_mode(*saved)