Update CMakeLists to support newer CMake versions

[Qiangong2]

FIND_PACKAGE(CUDA) was deprecated starting with cmake 3.10 and flat out removed in cmake 3.27. For machines in environments with strict update requirements, this breaks compilation on newer systems.

FIND_PACKAGE(CUDAToolkit) works the exact same way, and is supported by cmake.

Thanks to the NERSC team for identifying the issue.

[biochem-fan]

Thank you very much but can you make the pull request compatible with older versions of CMake?

[Qiangong2]

CUDAToolkit was added in 3.10, and distributions as far back as RHEL/CentOS 7 all have at least 3.10 available in their repos. A user would need to be running RHEL 6 or older (which don't even support cmake 3.0) to not have CUDAToolkit support in their cmake install.

This pr essentially makes the requirement cmake 3.10 instead of cmake 3.0.

[biochem-fan]

Isn't it just adding if-else with CMake version checks?

[FilipeMaia]

@Qiangong2, as @biochem-fan said, I would suggest suggest:

IF(CMAKE_VERSION VERSION_LESS 3.10)
FIND_PACKAGE(CUDA)
ELSE(CMAKE_VERSION)
# newer version of CMake use a different module name
FIND_PACKAGE(CUDAToolkit)
ENDIF(CMAKE_VERSION)

[Qiangong2]

Added in the if/else statement as requested. I can verify it builds and runs without issue.

[biochem-fan]

@Qiangong2 @FilipeMaia

Thank you very much.

I can verify it builds and runs without issue

Did you test on versions both before and after 3.10?
If so, I will merge this.

[Qiangong2]

I don't have a machine old enough to test if it works before 3.10, but since the if statement would use FIND_PACKAGE(CUDA) if it returned true, I believe it is functional.

[biochem-fan]

In FindCUDA, we can control arguments to nvcc via the CUDA_NVCC_FLAGS variable.
What is the new way to do this? We need to set the C++ standard and enable OpenMP in nvcc.
Failure to do so resulted in silent but serious problems before.

I looked at FindCUDA documentation:

It is no longer necessary to use this module or call find_package(CUDA) for compiling CUDA code. Instead, list CUDA among the languages named in the top-level call to the project() command, or call the enable_language() command with CUDA. Then one can add CUDA (.cu) sources directly to targets similar to other languages.

None of the pages linked from here mention CUDA_NVCC_FLAGS. Is the variable no longer supported after FindCUDA removal? What is the alternative way?