[F&E]: Improving Compilation Time

[luigigisolfi]

Description:

This issue will be taken care of during the Tudat Summer Camp 2026.

Suggestions for implementation:

No response

Additional context:

@HannoSpreeuw , @jhidding , @DominicDirkx could you add some context?

[larshinueber]

In addition to the ideas @alfonsoSR brought up during the meeting, I once more stumbled upon nanobind. They claim a ~3 times improvement in binary size and compilation time compared to pybind11 (see benchmark), while having an "extremely similar" API to pybind.
@jhidding @HannoSpreeuw do you have any experience with nanobind by chance?

[alfonsoSR]

The other day I was thinking about it and I came up with a sort of creative approach. Better to discuss it in a developer meeting, but if tudat (C++) is no longer a standalone library, we could just move the high-level interfaces to Python.

Sort of a radical thing to do, but stuff like create_system_of_bodies is not performance critical, and nothing recompiles faster than Python :)

It would also remove complexity from the library and make the source code a bit more accesible to more people, also much easier to debug. I can explain the concept better in the next meeting, but the more I think about it, the more I like it

[HannoSpreeuw]

In addition to the ideas @alfonsoSR brought up during the meeting, I once more stumbled upon nanobind. They claim a ~3 times improvement in binary size and compilation time compared to pybind11 (see benchmark), while having an "extremely similar" API to pybind. @jhidding @HannoSpreeuw do you have any experience with nanobind by chance?

Wow. Interesting. No, I don't know nanobind, but I would like to check next week if I can confirm the factor 3 shorter compilation time, for tudatpy.

[HannoSpreeuw]

we could just move the high-level interfaces to Python.

I like this approach. But, if you would enforce this concept across the entire codebase, what do you end up with? Only workhorse routines in C++, the rest in Python?

This is how I have often developed, top layer in Python, workhorses in ANSI C, or Numba compiled Python. I mean, if you are left with only workhorses to be compiled, you can choose from a number of languages, including Rust.

[larshinueber]

Another tool I came across is include-what-you-use (https://github.com/include-what-you-use/include-what-you-use) to detect which includes can be removed or simplified. I ran it on tudatpy, see the output attached.
The output is quite verbose and I believe it struggles a bit with Eigen types, as it includes headers from the user src directly (e.g. #include "src/Core/DenseBase.h"), but I think it could be a starting point to review the includes.

iwyu_output.txt