[Draft] Windows compatibility

[Louis-Pujol]

In this PR, we implement a working windows installation procedure for KeOps. It is still work in progress.

• Use CMAKE for compilations of nvrtc_jit, pykeops_nvrtc and cpp modules for KeOps on cpu
• Proper detection of cuda dlls and variables
• Modification of cpp code to make it consistent accross platform (e.g. long is not the same on Linux and Windows 64bit systems: https://www.ibm.com/docs/fr/ibm-mq/9.3.x?topic=platforms-standard-data-types-aix-linux-windows)
• Tests with various Windows configurations (MSVC, clang?...)

[Louis-Pujol]

Hi @bcharlier @jeanfeydy @joanglaunes ,

It seems that this version works. I've tested it on github action: https://github.com/Louis-Pujol/keops_multiplatform_test/actions/runs/16226947535 with python3.8 to 3.13 and without cuda: test_torch_bindings() and test_numpy_bindings() pass

This is a first step toward a clean integration, all the cpp files are in two versions: for windows with the extension _win.cpp or _win.h (from #281, so nvrtc_jit and pykeops_nvrtc are from KeOps2.1) .

I'll continue to work on it but it is at least easily testable. If you have a Windows machine, you can install it with:

pip install keopscore@git+https://github.com/Louis-Pujol/keops.git@windows-compatibility#subdirectory=keopscore
pip install pykeops@git+https://github.com/Louis-Pujol/keops.git@windows-compatibility#subdirectory=pykeops

If you test and have bugs, let me know ! Even if it works, you will have warnings about missing g++ and openmp, you can ignore them for now.

What I have done so far:

• add a subpackage windows_compilation: standalone cmake files generator with a compile() function and detection utilities for cuda, pybind11 and Python
• Overwrite the cuda detection in keopscore.config to make it work on windows
• For CPU code generation: replace long by int64_t
• other small changes that can be found by ctrl+f if os.name == "nt"

Still todo:

• Rewrite functioning C++ code from KeOps 2.3 source files (taking inspiration from Compile v2.1 on windows with vs2019 and cuda11.3 #281)
• Detect and use openMP if available (should be possible with cmake)
• Overwrite keopscore.config detectors of c++ compiler and openMP