Incorrect usage of float16 in algos.pyx

[FelipeLema]

Hi, everyone

According to documentation, np.float16 should not be used in C code as it is right now in pandas/algos.pyx:38 (assigning np.NINF to a np.float16 container in "cython-eventually-converted-to-C" code

As a matter of fact, when this value/variable is exposed, it is exposed as 0.0, and not -inf, either in C/C++ while using gdb or even with python.

Since MINfloat6 it is not being used in all pandas code, I recommend eliminating this line and/or commentating the reason. I tried by using np.float16(np.NINF) but the problem is that a correct C -inf value is still being C-assigned to a please-don't-use-C-container.

Why is it that this is important to me/I got to find this error? I'm dealing with 3rd party code so I cannot get a configuration that allows me to do this assignment without raising SIGFPE.

No pull request? The two possible solutions to me is either using cython-kept-as-python code with just this variable (which would impact performance) or simply not using it at all. I think I don't have the authority for either.

[jreback]

xref #9220, #2511, #10046

np.float16 is really not supported, so ok to simply remove it (I think it was just copy-pasted long time ago). But if you'd like to review the xref issues and provide a recommendation. thanks

[FelipeLema]

This is definately a duplicate of #10046. I just added another insight and a link to numpy's documentation.

[jreback]

@FelipeLema ok thanks!