Assess switch from pybind11 to nanobind

[garth-wells]

There may be benefits to switching from pybind11 to nanobind (https://github.com/wjakob/nanobind).

This will require a careful assessment, and the blockers resolved.

[chrisrichardson]

Known issues:

• No complex number support (yet) in nanobind
• No support for const pointers, e.g. std::shared_ptr<const Mesh> needs to be cast to std::shared_ptr<Mesh> before being used in nanobind - this will require a lot of new lambda functions in the wrappers.

[garth-wells]

See wjakob/nanobind#292 on support for complex.

[chrisrichardson]

Const is still a pain

[wjakob]

I just saw this issue because of a mention in the nanobind repo. Just to clarify: pybind11 handles shared_ptr<const...> and nanobind does not?

(If so, note that neither of these two frameworks honors const-ness, so this will likely not do what you want even for the pybind11 side.)

[chrisrichardson]

@wjakob. Yes. It is just an interfacing and convenience issue. Clearly python does not generally care about const (though I guess numpy has some idea).
The problem is we are wrapping functions with signatures including e.g. vector<vector<std:shared_ptr<const Foo>>> as input args. Now we have to unwrap/copy everything and write a lambda. Not impossible, just more laborious. pybind seems to conveniently just ignore the const.

[wjakob]

The shared_ptr issue should now be fixed on the master branch: wjakob/nanobind@21ee32b

[wjakob]

std::complex<T> is supported as of the latest release.

[garth-wells]

std::complex<T> is supported as of the latest release.

Thanks @wjakob. Does nb::ndarray support complex floats?

[chrisrichardson]

I've updated the branch https://github.com/FEniCS/dolfinx/tree/nanobind to sync with main.
I think shared_ptr may be now OK. Complex numbers are in nanobind for STL, but we still need ndarray to support std::complex.

[wjakob]

Another user has made a PR about this exact feature ;) -> wjakob/nanobind#319 (not sure if it's related to FEniCS)

[chrisrichardson]

Thanks for the heads-up @wjakob. I think they are looking for alien life (SETI Institute), which is a bit outside of our domain.

[luigifcruz]

I mean, FEniCS can be useful for radio astronomy too! 😄

Just as a clarification, my PR (wjakob/nanobind#319) was written for the Python bindings of BLADE, a GPU-accelerated software being used as the DSP-backend on the Allen Telescope Array for astronomy and SETI purposes.

[wjakob]

Thanks to @luigifcruz, this is now merged 🎉 . I just pushed out nanobind 1.7.0 with this change.

[garth-wells]

Thanks @luigifcruz, thanks @wjakob!

[garth-wells]

Switch (almost?) complete in #2820. Can be merged once new nanobind release made.