Port OpenVDB Python bindings from pybind11 to nanobind

[matthewdcong]

No description provided.

[Idclip]

I have no reason to think we would want to move away from pybind11. Can you explain your thinking here

[matthewdcong]

We discussed this in last week's TSC meeting. To summarize, I'm concurrently developing Python bindings for NanoVDB which expose GPU interop via nanobind, a feature which does not exist in pybind11. Since we need to pass objects between OpenVDB and NanoVDB at the C++ level even under a Python abstraction (e.g. in openToNanoVDB and nanoToOpenVDB), we need both to use the same binding library.

From the OpenVDB perspective, the port to nanobind provides no change in the Python interface (i.e. Python code that works will continue to work from the user's standpoint), and furthermore nanobind provides faster runtime, compile time, and smaller library files than pybind11.

[danrbailey]

We discussed this in last week's TSC meeting. To summarize, I'm concurrently developing Python bindings for NanoVDB which expose GPU interop via nanobind, a feature which does not exist in pybind11. Since we need to pass objects between OpenVDB and NanoVDB at the C++ level even under a Python abstraction (e.g. in openToNanoVDB and nanoToOpenVDB), we need both to use the same binding library.

From the OpenVDB perspective, the port to nanobind provides no change in the Python interface (i.e. Python code that works will continue to work from the user's standpoint), and furthermore nanobind provides faster runtime, compile time, and smaller library files than pybind11.

Mainly for @Idclip's benefit - we also discussed what we see as the main issue with this proposed change, which is the dependencies. Nanobind requires robinmap which is a new dependency compared to pybind11 and is not header-only when compared to pybind11. While many of the ASWF projects are actively switching from boost python to pybind11, there's not much adoption of nanobind currently so this proposed migration does introduce some additional barriers. One of the points of discussion were whether it might be feasible to embed a version of nanobind similar to what we do with half to reduce this barrier. Of course, Git submodules and similar mechanisms are not as convenient when behind a firewall. Embedding doesn't seem as straight-forward a decision as for half, but Matt was going to think a bit more about what that means in terms of symbols, namespaces, CMake integration, etc. Hopefully we can find a good time to discuss with you as well to get your thoughts. The benefits are clear so it's worth considering how exactly we might go about doing this. We haven't completely ruled out having pybind11 for OpenVDB and nanobind for NanoVDB as well.

[apradhana]

Git subtree can be an alternative to managing dependency. During our TSC meeting last week, we also talked about NanoBind's dependency to robin_map. It may be possible to apply a patch to a git subtree copy of NanoBind so that it doesn't depend on robin_map.

Moving to NanoBind also seems to future-proofing Python binding given the reasons that @matthewdcong explained.

[matthewdcong]

While nanobind for NanoVDB and pybind11 for OpenVDB might be a good interim solution, I'm not a huge fan of it long-term as it prevents a user from using Open and Nano synergistically in Python. For example, I think we'd want to support a workflow where host-side modifications are made in OpenVDB, followed by a conversion to NanoVDB for device-side computation, and a conversion back to OpenVDB for more host-side co-processing if possible.

[Idclip]

We discussed this in last week's TSC meeting. To summarize, I'm concurrently developing Python bindings for NanoVDB which expose GPU interop via nanobind, a feature which does not exist in pybind11. Since we need to pass objects between OpenVDB and NanoVDB at the C++ level even under a Python abstraction (e.g. in openToNanoVDB and nanoToOpenVDB), we need both to use the same binding library.
From the OpenVDB perspective, the port to nanobind provides no change in the Python interface (i.e. Python code that works will continue to work from the user's standpoint), and furthermore nanobind provides faster runtime, compile time, and smaller library files than pybind11.

Mainly for @Idclip's benefit - we also discussed what we see as the main issue with this proposed change, which is the dependencies. Nanobind requires robinmap which is a new dependency compared to pybind11 and is not header-only when compared to pybind11. While many of the ASWF projects are actively switching from boost python to pybind11, there's not much adoption of nanobind currently so this proposed migration does introduce some additional barriers. One of the points of discussion were whether it might be feasible to embed a version of nanobind similar to what we do with half to reduce this barrier. Of course, Git submodules and similar mechanisms are not as convenient when behind a firewall. Embedding doesn't seem as straight-forward a decision as for half, but Matt was going to think a bit more about what that means in terms of symbols, namespaces, CMake integration, etc. Hopefully we can find a good time to discuss with you as well to get your thoughts. The benefits are clear so it's worth considering how exactly we might go about doing this. We haven't completely ruled out having pybind11 for OpenVDB and nanobind for NanoVDB as well.

Thank you Dan, this has exactly described my concern. Faster compiler times/smaller binaries and "faster" runtime (I'd be surprised to see this having a significant impact the the types of bindings we expose - is the binding lookup really that slow compared to the VDB operations we're invoking? happy to be proven wrong but would like to see results) are great, but IMO are really not much of a win when compared to the added potential build complexity and ASWF homogenization. We just moved away from boost python for exactly this reason.

Moving to NanoBind also seems to future-proofing Python binding given the reasons that @matthewdcong explained.

Nanobind is newer, this does not necessarily mean it's more future proof and I don't see enough evidence to reflect this.

I do however appreciate the mentioned limitation with GPU interop - @matthewdcong does this pybind11 issue reflect the problem or is it more involved? pybind/pybind11#3858 I would like to understand this more and whether there are any other solutions available to us with pybind11

[matthewdcong]

Yeah, the pybind11 issue outlines part of the problem. As they mention in the discussion, backporting a basic typecaster for a DLPack struct to pybind11 is doable, but it becomes much more involved to transparently handle all the frameworks (NumPy, PyTorch, Tensorflow, etc.) in a unified manner with datatype casting, shape checking, and host and device memory management.

[matthewdcong]

While an additional dependency may introduce a small barrier to adoption, Python-level host and device interoperability with widely-used frameworks such as PyTorch and Tensorflow has the potential to significantly increase the audience for the VDB data structure within the large and growing ML community.