numpy unexpectedly raises zero division error if array is long enough

[evfro]

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.

[charris]

I cannot reproduce this. What platform?

[evfro]

Sorry for not mentioning,
Windows 10, version 1709

[mattip]

What do you get when you run import sys; print(sys.version), if not anaconda how did you obtain NumPy?

[evfro]

3.6.5 |Anaconda, Inc.| (default, Apr 29 2018, 16:14:56) [GCC 7.2.0]

[mattip]

How do you get Anaconda on windows 10 compiled with GCC 7.2.0? Is that standard? I thought they use MSVC?

[evfro]

My bad, I've mixed terminals. I have double checked and I'm also not able to reproduce it on Windows.
So the problem appears only on Ubuntu (at least 14.04.4 LTS and 16.04.3 LTS) both with
Python 3.6.5 |Anaconda, Inc.| (default, Apr 29 2018, 16:14:56) [GCC 7.2.0] on linux and numpy version 1.14.2

[evfro]

Hi! Any update on that issue? I'm getting this error also with newer versions of Python: Python 3.7.0 (default, Jun 28 2018, 13:15:42) [GCC 7.2.0] :: Anaconda, Inc. on linux, and Numpy: 1.15.2 on Ubuntu system.

[charris]

I still cannot reproduce this, so it is hard to debug :)

[charris]

Anyone out there running ubuntu who can check this?

[charris]

I don't see anything suspicious between NumPy 1.14.1 and 1.14.2

A total of 5 pull requests were merged for this release.

* `#10674 <https://github.com/numpy/numpy/pull/10674>`__: BUG: Further back-compat fix for subclassed array repr
* `#10725 <https://github.com/numpy/numpy/pull/10725>`__: BUG: dragon4 fractional output mode adds too many trailing zeros
* `#10726 <https://github.com/numpy/numpy/pull/10726>`__: BUG: Fix f2py generated code to work on PyPy
* `#10727 <https://github.com/numpy/numpy/pull/10727>`__: BUG: Fix missing NPY_VISIBILITY_HIDDEN on npy_longdouble_to_PyLong
* `#10729 <https://github.com/numpy/numpy/pull/10729>`__: DOC: Create 1.14.2 notes and changelog.

[mattip]

Maybe Ananconda switched compilers / mkl libraries. Please check with them.

[mattip]

@evfro, @oleksandr-pavlyk can we close this?

[evfro]

I've found a workaround and not using the lines of code that generate that issue.
That said, I've checked whether the issue still persists with the recent compatible conda/numpy/mkl environment and, unfortunately, it does.

import sys; print(sys.version)
3.6.9 |Anaconda, Inc.| (default, Jul 30 2019, 19:07:31) [GCC 7.3.0]

numpy.__version__
1.16.4

conda list mkl
mkl 2019.4 243
mkl-service 2.0.2 py36h7b6447c_0

[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)

[mattip]

NumPy specifically resets the error state when it is obvious (by emitting NAN) that divide-by-zero occurred. I think the MKL port should do the same inside the ufuncs without requiring global error mode control.

[mattip]

Closing, as it seems this is a "won't fix" from the mkl port of NumPy. Please reopen if there is something actionable NumPy can do.

[phil-blain]

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

2021 link: https://software.intel.com/content/www/us/en/develop/documentation/onemkl-developer-reference-c/top/vector-mathematical-functions/vm-service-functions/vmlsetmode.html#vmlsetmode

[astroboylrx]

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

Thanks a lot for explaining that. It's so hard to find relevant information since too many webpages are about real zero-division cases.

I ran into this issue as well and it only occurs with numpy with MKL. I tried to turn errors off with your code snippet and it
does removed half of such errors.
@oleksandr-pavlyk @phil-blain I wonder what might be responsible for the other half "divide by zero encountered in true_divide" errors? Any thoughts or suggestions would be greatly appreciated!

[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().