Cmake: Add 3.28:

[johnwparent]

No description provided.

[johnwparent]

cc @alalazo

[tldahlgren]

FYI. Confirmed the 5 new version sha256s.

[cedricchevalier19]

@johnwparent can you describe a little bit more the conflict with Kokkos?

We have observed an issue using CMake 3.28.4 and Cuda (kokkos/kokkos#6898), but it should be solved in the last version 4.3.00 of Kokkos.

Thank you very much in advance.

[johnwparent]

FYI. Confirmed the 5 new version sha256s.

Thanks as always!!

[johnwparent]

@johnwparent can you describe a little bit more the conflict with Kokkos?

We have observed an issue using CMake 3.28.4 and Cuda (kokkos/kokkos#6898), but it should be solved in the last version 4.3.00 of Kokkos.

Thank you very much in advance.

@cedricchevalier19

With CMake 3.28 (all patch versions so far) we're seeing a CI failure specifically with kokkos-kernels and trilinos right now, but this would be true with any package consuming the kokkos CMake config file. The error we observe is an eventual make (or CMake for 3.28.4) failure caused by this (from the kokkos config):

  CMake Warning (dev) in CMakeLists.txt:
  Policy CMP0111 is not set: An imported target missing its location property
  fails during generation.  Run "cmake --help-policy CMP0111" for policy
  details.  Use the cmake_policy command to set the policy and suppress this
  warning.

  IMPORTED_LOCATION not set for imported target "CUDA::cuda_driver"
  configuration "RELEASE".
This warning is for project developers.  Use -Wno-dev to suppress it.

CMake Error at kokkos/lib64/cmake/Kokkos/KokkosConfig.cmake:41 (SET_TARGET_PROPERTIES):
  The link interface of target "CUDA::cudart" contains:

    CUDA::cudart_static_deps

  but the target was not found.  Possible reasons include:

    * There is a typo in the target name.
    * A find_package call is missing for an IMPORTED target.
    * An ALIAS target is missing.

  Call Stack (most recent call first):
  CMakeLists.txt:129 (FIND_PACKAGE)

when tools eventually can't resolve the path to that target.
From a grep of the kokkos code base we can see some custom logic around this and similar targets

grep -rin "cudart_static" .
./cmake/Modules/CudaToolkit.cmake:130:- ``CUDA::cudart_static``
./cmake/Modules/CudaToolkit.cmake:806:  _CUDAToolkit_find_and_add_import_lib(cudart_static)
./cmake/Modules/CudaToolkit.cmake:808:  # setup dependencies that are required for cudart_static when building
./cmake/Modules/CudaToolkit.cmake:811:  if(NOT TARGET CUDA::cudart_static_deps
./cmake/Modules/CudaToolkit.cmake:812:     AND TARGET CUDA::cudart_static)
./cmake/Modules/CudaToolkit.cmake:814:    add_library(CUDA::cudart_static_deps IMPORTED INTERFACE)
./cmake/Modules/CudaToolkit.cmake:815:    target_link_libraries(CUDA::cudart_static INTERFACE CUDA::cudart_static_deps)
./cmake/Modules/CudaToolkit.cmake:819:      target_link_libraries(CUDA::cudart_static_deps INTERFACE Threads::Threads ${CMAKE_DL_LIBS})
./cmake/Modules/CudaToolkit.cmake:827:        message(WARNING "Could not find librt library, needed by CUDA::cudart_static")
./cmake/Modules/CudaToolkit.cmake:829:        target_link_libraries(CUDA::cudart_static_deps INTERFACE ${CUDAToolkit_rt_LIBRARY})

So it seems like when we updated the CUDA detection behavior in CMake, the Kokkos logic around these targets broke and is now producing bad KokkosConfig.cmake files

@cedricchevalier19 is the Fix in Kokkos in a release that's part of Spack? If so I'll scope the conflict appropriately.

[cedricchevalier19]

Fix is present in 4.3.00 which is available in Spack.

For other versions, conflict should be declared only when +cuda is set.

[johnwparent]

Fix is present in 4.3.00 which is available in Spack.

For other versions, conflict should be declared only when +cuda is set.

@cedricchevalier19
Thank you, I've updated the conflict to be more specific.

[cedricchevalier19]

It is fine on the Kokkos side. I'll let @alalazo and the spack team decide if they want to separate the Kokkos conflicts modification.

[johnwparent]

Note: requires #43780