CI update

[j-stephan]

Summary:

• added gcc-12 to CI - handled in Add gcc-12 to CI #1765
• added clang-14 to CI - handled in Add clang-14 to CI #1766
• added Xcode 13.3.1 to CI -> we now support macOS 12 "Monterey" - handled in CI Update for macOS #1759
• added CUDA 11.7 to CI - handled in CUDA CI update #1890
• CUDA 11.6 tests now use CUDA 11.6.2 - handled in CUDA CI update #1890
• collapsed compiler support matrix - handled in Collapse compiler matrix #1860
• fixed a few consistency issues for the clang-cuda device compiler (the matrix showed support for CUDA versions that clang itself doesn't officially support yet) - handled in CUDA CI update #1890
• added CUDA 11.6 + gcc-11 job to GitLab CI - handled in CUDA CI update #1890
• added ROCm 5.1 jobs to GitLab CI - handled in CI Update for ROCm #1762.
• added Boost 1.79.0 to CI - handled in Update CMake and Boost point releases #1903
• added CMake 3.23.x to CI - handled in Update CMake and Boost point releases #1903
• updated older CMake versions to latest point releases - handled in Update CMake and Boost point releases #1903
• removed all occurrences of alpaka_CXX_STANDARD: 17 from the GitHub CI script since that is now our default - handled in Enable more C++20 jobs #1902
• added a few occurrences of alpaka_CXX_STANDARD: 20 to the GitHub CI script - handled in Enable more C++20 jobs #1902
• gcc-12 jobs will use GitHub's ubuntu-22.04 runners - handled in Add gcc-12 to CI #1765
• If sanitizers are enabled, clang jobs will install the llvm base package. This includes the llvm-symbolizer binary required for meaningful output. - handled in Simplify clang installation #1763
• If the Boost.fiber back-end is enabled in C++20 mode CMake will print an error and abort (see [RFC] Future of the Boost.fiber back-end #1714). - superseded by Drop Boost.fiber back-end #1718
• Disabled the address sanitizer for memBufTest if clang-14 is used (see clang-16: ASan fails for memBufTest. #1753). - needs more investigation
• Removed Xcode 12.4.0 from the CI -> we will no longer support macOS 10.15 "Catalina" - handled in CI Update for macOS #1759.

Fixes #1670. Fixes #1708. Fixes #1724.

[j-stephan]

Wow, case in point: One of our C++20 jobs is failing with Boost 1.79.0. Apparently Boost.fiber is internally using std::result_of which was removed in C++20.

From Boost.fiber's documentation:

Boost.fiber requires C++11!

So I guess we need to start thinking about the future of this back-end.

[bernhardmgruber]

Wow, case in point: One of our C++20 jobs is failing with Boost 1.79.0. Apparently Boost.fiber is internally using std::result_of which was removed in C++20.

That was fixed 15 days ago: boostorg/fiber@a6b64c9

However, that still leaves Boost.Fiber 1.79 and below problematic with C++20.

[bernhardmgruber]

It seems g++-12 is not installed:

2022-05-10T17:06:01.9941211Z +++ sudo update-alternatives --install /usr/bin/g++ g++ /usr/bin/g++-12 50
2022-05-10T17:06:01.9987407Z update-alternatives: error: alternative path /usr/bin/g++-12 doesn't exist

[j-stephan]

You're right. apt install g++-12 leads to an installation of clang-12 for some reason. Maybe it cannot find g++-12 and does some guesswork. 🤡

[j-stephan]

Rebased and added the following changes:

• added CUDA 11.7 to CI
• CUDA 11.6 tests now use CUDA 11.6.2
• gcc-12 jobs will use GitHub's ubuntu-22.04 runners

[j-stephan]

• gcc-12 jobs are failing because there is a Docker bug in the ubuntu-22.04 runners. This has been fixed but is apparently still in the rollout phase.
• clang-14-asan is failing because of a bug in one of our test cases. We need to investigate this.

[jkelling]

While you are at it: Could you please change the build type of check with enabled sanitizer which are currently Release to RelWithDebInfo to make the printed backtraces more useful? (e.g. linux_clang-6_release_asan, I do not know if there are others)

[j-stephan]

While you are at it: Could you please change the build type of check with enabled sanitizer which are currently Release to RelWithDebInfo to make the printed backtraces more useful?

Done.

[j-stephan]

Note: The macOS 10.15 runners have been deprecated by GitHub and will be removed at the end of August. I therefore removed our macOS 10.15 tests from the CI which will be visible upon my next push (after merging #1754 and #1755).

[psychocoderHPC]

I am a little bit unhappy that this PR is mixing features, refactorings, and adding new test combinations.

Could you split e.g CUDA keyring refactorings, job renamings and new jobs into separate PRs? This would also simplify the review and give us a better overview of why a CI job is failing.

[j-stephan]

And so it shall be done. I will prepare a number of successor PRs and close this one afterwards.

[j-stephan]

All points from the original description have been addressed.