[Refactor] CI job generator: define in the version.py which backend combinations should be tested

[SimeonEhrig]

Before, the backends was enabled depending on the host and device compiler.
For each host device compiler combination only one backend combination was possible.
Now, the backend combinations are defined in the versions.py and each combination can be combined with a host device compiler combination, if no filter rule forbid it.

Is preparation for enabling GCC and Clang CPU tests in the GitLab CI.
Should solve also issue #639 -> no, see here

[SimeonEhrig]

The failed TSAN job is independent of this PR, see #2101.

[SimeonEhrig]

There is no rule, which requires an enabled OpenMP2 backend for nvcc as device compiler. Therefore both backend combinations are tested.

alpaka/script/job_generator/versions.py

Lines 37 to 53 in 9a30ae8

	# nvcc
	[
	ALPAKA_ACC_CPU_B_SEQ_T_SEQ_ENABLE,
	ALPAKA_ACC_CPU_B_SEQ_T_THREADS_ENABLE,
	ALPAKA_ACC_CPU_B_OMP2_T_SEQ_ENABLE,
	ALPAKA_ACC_CPU_B_SEQ_T_OMP2_ENABLE,
	ALPAKA_ACC_GPU_CUDA_ENABLE,
	],
	# clang-cuda
	# OpenMP is not supported for clang as cuda compiler
	# https://github.com/alpaka-group/alpaka/issues/639
	# therefore a extra combination is required
	[
	ALPAKA_ACC_CPU_B_SEQ_T_SEQ_ENABLE,
	ALPAKA_ACC_CPU_B_SEQ_T_THREADS_ENABLE,
	ALPAKA_ACC_GPU_CUDA_ENABLE,
	],

@psychocoderHPC Any opinion about the behavior?

[psychocoderHPC]

There is no rule, which requires an enabled OpenMP2 backend for nvcc as device compiler. Therefore both backend combinations are tested.

alpaka/script/job_generator/versions.py

Lines 37 to 53 in 9a30ae8

	# nvcc
	[
	ALPAKA_ACC_CPU_B_SEQ_T_SEQ_ENABLE,
	ALPAKA_ACC_CPU_B_SEQ_T_THREADS_ENABLE,
	ALPAKA_ACC_CPU_B_OMP2_T_SEQ_ENABLE,
	ALPAKA_ACC_CPU_B_SEQ_T_OMP2_ENABLE,
	ALPAKA_ACC_GPU_CUDA_ENABLE,
	],
	# clang-cuda
	# OpenMP is not supported for clang as cuda compiler
	# https://github.com/alpaka-group/alpaka/issues/639
	# therefore a extra combination is required
	[
	ALPAKA_ACC_CPU_B_SEQ_T_SEQ_ENABLE,
	ALPAKA_ACC_CPU_B_SEQ_T_THREADS_ENABLE,
	ALPAKA_ACC_GPU_CUDA_ENABLE,
	],

@psychocoderHPC Any opinion about the behavior?

@SimeonEhrig looks good for me.

[psychocoderHPC]

How can I see that this is for clang-cuda except by the comment? So I am wondering how the generator know that for clang-cuda this configuration should be used and not the first entry in BACKENDS?

[SimeonEhrig]

You cannot see it directly in code, because it is decided by filter rules. So clang-cuda requires ALPAKA_ACC_GPU_CUDA_ENABLE and does not allow the OMP2 backends:

alpaka/script/job_generator/alpaka_filter.py

Lines 48 to 56 in 9a30ae8

	# OpenMP is not supported for clang as cuda compiler
	# https://github.com/alpaka-group/alpaka/issues/639
	if row_check_name(row, DEVICE_COMPILER, "==", CLANG_CUDA) and (
	row_check_backend_version(row, ALPAKA_ACC_CPU_B_OMP2_T_SEQ_ENABLE, "==", ON_VER)
	or row_check_backend_version(
	row, ALPAKA_ACC_CPU_B_SEQ_T_OMP2_ENABLE, "==", ON_VER
	)
	):
	return False

You can check if a compiler backend passes the filter with ajc-validate tool (except the alpaka filter, I need to implement support for this 🤔 ).

[j-stephan]

Just a few suggestions:

[SimeonEhrig]

I stopped the merge process, because I found a bug in #2102 which already exist in this PR. The bug is simply not triggered, because we don't use a specific combination of enabled backends.

[SimeonEhrig]

I stopped the merge process, because I found a bug in #2102 which already exist in this PR. The bug is simply not triggered, because we don't use a specific combination of enabled backends.

@j-stephan The bug is fixed in the alpaka-job-coverage library 1.5.3. and tests passed: #2102

Therefore the PR is ready to merge. Since your last review I only increased alpaka-job-coverage library version.