ENH: expose datetime64 conversion functions #16364
Conversation
| @@ -287,8 +286,7 @@ convert_timedelta_to_pyobject(npy_timedelta td, PyArray_DatetimeMetaData *meta); | |||
| /* | |||
| * Converts a datetime based on the given metadata into a datetimestruct | |||
| */ | |||
| NPY_NO_EXPORT int | |||
| convert_datetime_to_datetimestruct(PyArray_DatetimeMetaData *meta, | |||
| int convert_datetime_to_datetimestruct(PyArray_DatetimeMetaData *meta, | |||
There was a problem hiding this comment.
I don't think this is sufficient, extension modules aren't allowed to link against each other. This needs to go through the function table like the other NUMPY_API functions.
There was a problem hiding this comment.
Thanks, ill see if i can track that down. Before I do, is this likely to be accepted once working?
There was a problem hiding this comment.
I think we'd need to think about the function names, but in principle it seems sensible to me to expose the npy_datetimestruct type and some APIs that consume it.
|
For completeness we might want to upstream the timedelta changes as well: Probably in a separate PR, maybe a pre-cursor. The above is used in JSON so I think we wouldn't be able to completely simplify our dependency structure without tacking that as well |
|
@eric-wieser Not sure why this is tagged random. |
@bashtage best guess is that on a previous PR implementing test_cython we adapted the setup from random.tests.test_extending into a fixture and there was an offhand mention of using that pattern in the random tests. |
|
Closing in favor of #21199. |
These would allow pandas to remove our copy/pasted implementations.
cc @WillAyd @mroeschke anything else we might need?