Python API implicitly requires Kokkos

[Sichao25]

Currently, the Python API requires Kokkos, because some Kokkos views are used without a guard (e.g. #ifdef OMEGA_H_USE_KOKKOS).

I think this could be resolved in the source code by guarding the specific functions though I haven't taken a deeper look yet to confirm the scope of the change.

However, depending on the desired behavior for the Python API, we could instead make Kokkos an explicit dependency of the Python variant. This could be included in spack/spack-packages#5792

What do you think? @cwsmith @jacobmerson

[Fuad-HH]

I was just checking this. I guess only changing code src/PyOmega_h_numpy_transform.hpp will be enough to remove kokkos dependency. Are there other places you know of that includes kokkos directly?

I can even add it in my pull request (#229).

[Sichao25]

That is the only place I know for now.

[Fuad-HH]

My pull request removed complete kokkos dependence expect one. I tried building without kokkos and with python bindings. Got the following error:

src/PyOmega_h_mesh.cpp:436:45: error: 'get_vtx_patches' is not a member of 'Omega_h::Mesh'
  436 |     .def("get_vtx_patches", &Omega_h::Mesh::get_vtx_patches,
      |                                             ^~~~~~~~~~~~~~~

This is because Omega_h::Mesh::get_vtx_patches definition is under the OMEGA_H_USE_KOKKOS guard (here).

I tried removing the binding for it and Omega_h compiled without any problem and the python tests passed. However, this is used in pcms: here. I am not sure what is the best way to fix this.

[Fuad-HH]

This is fixed in #231 and #229 .

[Sichao25]

Thanks!