Masked array performance

[bjlittle]

Given #3301 and SciTools/iris-agg-regrid#37 we should consider whether masked arrays with no masked points should be automatically converted to ordinary ndarrays for obvious performance benefits.

Note that python-netcdf4 now automatically always returns masked arrays.

This could be considered as a breaking change in iris behaviour, and so puts it plum in 3.0.0 territory.

---

As a user of iris, I want comparable regridding performance when using masked array data where no points are masked (i.e. mask=False) and non-masked arrays. I would also like to have the ability to apply this functionality to other areas of my code if I deem it necessary.

I would like the capacility to be avilable to users to apply else where in their code.

Acceptance Criteria

• Create a function/decorator that can be applied to my iris function to give equivalent performance if my cube data is masked, but the mask is False.
• The output cube data should remain masked if it was originally masked. Remain unmasked if originally unmasked.
• The function/decorator is applied to appropriate iris regridding methods.
• Have an ASV benchmark to prove the benefits of the solution.
• The new functionality should be in the what's new.
• The new functionality is documented in the user guide.

[pp-mo]

masked array data where no points are masked (i.e. mask=False)

That "i.e." isn't quite an equation : You can still have a mask array, but with no True values.
For operations where masks are costly, like regridding, it will still be worth the cost of a test + convert.
But it probably needs to be on a case-by-case basis.
We should have some common code for it, though.

[pp-mo]

... meanwhile, could we not routinely convert the mask=False cases on loading, if that is generally useful ??

[rcomer]

This week I learned that a whole new implementation of masked arrays is afoot:
https://github.com/ahaldane/ndarray_ducktypes/blob/master/doc/MaskedArray.md

Note that the owner of that repo is a numpy "member" and another member indicated a consensus for replacing numpy masked arrays "at some point" with something developed outside numpy: numpy/numpy#16022 (comment)