Automatic conversion numpy.dtype <-> nanobind::dlpack::dtype

[cosama]

Probably a bit of a edge case, but I have a class that I would like to initialize with a numpy.dtype as an argument. This seems to not be possible.

I was also thinking about having the nanobind::dlpack::dtype being returned from a function, exposed as a numpy.dtype in python and I think that would not work either.

Not a huge issue for now, but maybe somebody else will think about doing this and might see this before going down the rabbit hole.

[hpkfft]

A function can have nb::object obj as a parameter. Then you can get its attributes. For example, np.float32 has attribute itemsize with value 4. See also: https://nanobind.readthedocs.io/en/latest/api_core.html#attribute-access and https://nanobind.readthedocs.io/en/latest/api_core.html#_CPPv4N8nanobind4int_E

To return a numpy.dtype, you can make the object you wish to return by getting the NumPy module and calling its dtype() attribute. Something like:

nb::module_ numpy = nb::module_::import_("numpy");
nb::object dt = numpy.attr("dtype")("float32")
return dt;

Note: I did not test the above, so maybe I made some dumb mistake. :)