[Numpy-discussion] [SciPy-user] Managing Python with NumPy and many external libraries on multiple Windows machines

[David Cournapeau]

Stéfan van der Walt wrote:
>
> I see a number of docstrings already have
>
> .. versionadded:: 1.3.0
>   

Yes, but I don't think it is enough. In particular, I don't like so much
the fact that the C API documentation for every function is separate
from the code. There are several mismatches between the refguide and the
actual code ATM. Maybe we could use sphinx to get the doc from the .c
code to document the C api functions, so everything is at one place ?


> I assume you are referring to ufunc_api_order.txt and
> numpy_api_order.txt.

yes.

>   Fixing them by hashing should be fine; we must
> just be very careful not to change those files by accident.

I think as a rule, we should not change those by accident :) For pure C
libraries, the public headers should always be treated very carefully -
unfortunately, we don't have anything like that for python C extensions,
but those .txt files are the closest thing we have.

I put some code in the detect_capi_mismatch svn branch:
    - the checksum is computed on a string representation of the ordered
list of both numpy and ufunc C API
    - the ordered list is the one returned by the parsing function used
to generate those headers
    - I keep a checksum of each new version in a separate file. A
warning is raised if there is a mismatch

cheers,

David