BLD: should we use tempita for cython templating

[jreback]

see here

might replace some of the manual code paths in generate_code.py

[jreback]

@gfyoung can you comment on your experience

[gfyoung]

+1 from me - it's currently used in scipy and numpy and I in fact put it to good use in one of my PR's against numpy. Also, since tempita comes with Cython (Cython.Tempita --> no extra build dependency), if it's possible to use in order to cut down on manual code (which it seems based on a quick skim of the code in pandas/src/generate_code.py), certainly!

[sinhrks]

Quite interesting. I'm just adding new cython templates to support more sparse dtypes, and it can be much simpler using it.

[jreback]

@gfyoung can u post a link to the PR where u r using this?

[gfyoung]

Link to PR here. Key file to examine is randint_helpers.pxi.in and the templating code here. The key part is getting the template right, as once you do, filling it in is really easy.

Useful discussion points regarding Tempita in my PR start here, as there are links to other libraries that use Tempita as well.

[jreback]

thanks @gfyoung

note that I have found fused types pretty useful as well. here is an example of a numeric type which is used in window.pyx & some of the algos.pyx. These are pretty easy to define and work with; and somewhat remove the need for templating.

[gfyoung]

@jreback : yes, I remember you mentioning that in that same PR <a href=https://github.com/numpy/numpy/pull/6938#discussion_r64143615">here. I guess not all maintainers are a fan of it though, and it did take some getting used to as well when I first attempted it.

[jreback]

right they are in newer cython (though our min 0.19.1 certainly has it)

[gfyoung]

Well different repository, different rules 😄 - fused_types may not be as applicable in this case though given that there are some things in the template to generate_code.py that aren't fuse-type-able FLOABW.

[jreback]

@gfyoung yeah, though if I could use a fused type I think we should. Just easier to read/maintain. They seem to act very much like 'real' types, with a tiny exception. You need to this to avoid warnings when compiling.

cdef numeric val

if numeric in cython.floating and val == val:
     # do something if we are not-NaN

[shoyer]

I agree that fused types are nice, but in my experience they end up falling short. Tempita looks pretty nice, better than our ad-hoc current system.