cmake update for main VASP version

[the-hampel]

Hi @AdhocMan ,

I made various changes to the cmake files to work for all supported platforms. These are quite drastic changes, but I thought I would like to update you on the progress I made and how the version looks that we merge into the master branch of VASP.

One of the most important changes is here that I am decided to preprocess all fortran files first manually to solve any dependency issues. This works now very reliably.

If you have any comments on these I am happy to incorporate any suggestions you have. If you see any problems let me also now.

And thank you again for your work on this. This was a great help.

[AdhocMan]

Thanks for providing these changes!
It appears to work fine in our GH200 cluster environment, including parallel compilation with make.
I have added a few comments on some parts.

[AdhocMan]

I'd strongly recommend not to do this detection logic.
On some clusters a GPU may not be available at compilation and this will fail. Additionally, the nvcc executable may not be in PATH, so this would have to be looked for through enable_language(CUDA).
The detection of the default architecture is already handled by setting CMAKE_CUDA_ARCHITECTURES to native, since -gpu=ccnative lets nfortran do the detection.

[the-hampel]

Yes I see what you mean. I found this solution somewhere in the internet and later discovered that cmake can detect cuda by itself with check_language(CUDA) . Do you know if this command actually also checks that nvcc is working? Similar to how it is done with gcc? But anyway I can drop this. I see the point that test compiling via nvcc might be okay, even if no GPU is present, but executing is a problem. I will remove that part.

[the-hampel]

see b801e20

[AdhocMan]

Using the check_language feature relates to the broader question of whether you'd want to set options automatically. One could apply similar logic to other features like OpenMP. Personally, I'd tend to prefer fixed settings.
But generally speaking, there should always be a way to disable any automatic option selection, since it's much better for package managers like spack to fail if an selected option does not work. Using set(VASP_CUDA ON) will always override the user provided setting.
If there is a need for default settings for certain use cases, maybe have a look at https://cmake.org/cmake/help/latest/manual/cmake-presets.7.html. This should allow to define sets of options to be used together, for example when building with CUDA.

Also, adding the check if(${CMAKE_CUDA_ARCHITECTURES} STREQUAL "native") is unessary, since the flag will already be correctly set through

    set(VASP_CUDA_ARCH ${CMAKE_CUDA_ARCHITECTURES})
    list(POP_FRONT VASP_CUDA_ARCH VASP_FIRST_ARCH)
    set(VASP_GPU_FLAG "-gpu=cc${VASP_FIRST_ARCH}")

[the-hampel]

Thanks for taking such detailed look at the PR @AdhocMan. These are all very helpful and I tried to address all of them. I answered to all comments separately above and incorporated the changes.