kokkos_launch_compiler + CUDA auto-detect arch

[jrmadsen]

• enables CUDA as a language to auto-detect architecture when kokkos_launch_compiler is being used
• removes -DKokkos_ARCH_VOLTA70=ON from Jenkins test which tests the compiler-launcher
• Closes Make GPU Auto-Detection Work With Kokkos Launch Compiler #3758

[masterleinad]

Is there any problem with always doing this and removing the first run?

[jrmadsen]

Is there any problem with always doing this and removing the first run?

I honestly cannot think of any other than it might cause users unfamiliar with the Kokkos build system to see CUDA getting enabled as a language and lead them to believe we are actually compiling things with CMAKE_CUDA_COMPILER

[jrmadsen]

As far as I know there is no way to suppress this from getting emitted when the enable_language(CUDA) is done:

-- The CUDA compiler identification is NVIDIA 11.0.221
-- Detecting CUDA compiler ABI info
-- Detecting CUDA compiler ABI info - done
-- Check for working CUDA compiler: /usr/local/cuda/bin/nvcc - skipped
-- Detecting CUDA compile features
-- Detecting CUDA compile features - done

[jrmadsen]

Even though I can't think of why it may cause the auto-detect to emit something different, at least with the current implementation, if the users sets CMAKE_CXX_COMPILER to nvcc_wrapper or mpicxx + MPICH_CXX=nvcc_wrapper (or whatever the env variable is), we know the compiler being used to detect the auto-arch will be the compiler used when compiling the code. Supercomputers have a lot of toolchains and I don't feel like I can say with 100% certainty that if cmake picks up nvcc from cuda 9.2, it will correctly detect the Ampere arch. To me, keeping the way it currently is will produce the least amount of surprises and removing it will only save us ~8 lines of cmake code.

[masterleinad]

Only some minor comments. Otherwise, I'm fine with the changes here.

[masterleinad]

Looks OK to me.

[masterleinad]

Retest this please.

[dalg24]

Looks fine to me. Pasting below snippet from configuration log in CI build for reference

-- CUDA auto-detection of architecture failed with /usr/bin/g++-8. Enabling CUDA language ONLY to auto-detect architecture...
-- Looking for a CUDA compiler
-- Looking for a CUDA compiler - /usr/local/cuda/bin/nvcc
-- The CUDA compiler identification is NVIDIA 11.0.221
-- Check for working CUDA compiler: /usr/local/cuda/bin/nvcc
-- Check for working CUDA compiler: /usr/local/cuda/bin/nvcc - works
-- Detecting CUDA compiler ABI info
-- Detecting CUDA compiler ABI info - done
-- Detecting CUDA compile features
-- Detecting CUDA compile features - done
-- Detected CUDA Compute Capability 70
-- Setting Kokkos_ARCH_VOLTA70=ON

You must fix the indentation though. The rest of the cmake/kokkos_arch.cmake uses two spaces and your patch uses a mix of two and four spaces.

[jrmadsen]

You must fix the indentation though. The rest of the cmake/kokkos_arch.cmake uses two spaces and your patch uses a mix of two and four spaces.

Nah, my IDE aligns to the original indentation of two spaces and then uses 4 spaces from then on out because 4 spaces is better than 2 😉

[masterleinad]

Retest this please.