Damien L-G
Damien L-G
> > I know there is `__CUDA_ARCH__` but that macro is only defined inside device functions. Is there a different way to detect the architecture? > > We define `KOKKOS_IMPL_ARCH_NVIDIA_GPU`...
We export a `kokkos_compilation()` function with CMake that you can use to do this portably. https://github.com/kokkos/kokkos/blob/f53b26d44285a93ca6f17669fdf47315cb392e9e/cmake/KokkosConfigCommon.cmake.in#L203-L218 Admittedly this is not covered in the documentation. You would do ```CMake kokkos_compilation(SOURCE exercise_4_solution.cpp)
And to be clear, `nvcc_wrapper` is not supported on Windows, as Daniel pointed out in https://github.com/kokkos/kokkos/issues/7144#issuecomment-2237331863, one must configure with `-D Kokkos_ENABLE_COMPILE_AS_CMAKE_LANGUAGE=ON`. The tutorials currently do not support MSVC+CUDA so...
We recommend to use the Threads backend on Windows.
(had forgotten to commit changes to `Kokkos_ViewAlloc.hpp`)
> > Reduce the number of instantiations in the code. > > Do you have any data to suggest this was a problem? Every single view initialize (scalar) and view...
> Anyway, I don't feel strongly about it if you feel that the project is going to be really diligent in not making a mistake with calling it. These are...
It does not **compile** as the warning is being treated as an error. We are aware of that issue, it is currently showing on all PRs, and clearly unrelated to...
Planning to ignore the timeout in the MI300A build on HPSF CI