[CI] Add libclc tests to pre- and post-commit validation

[jchlanda]

Run check-libclc as part of continues integration system. Time-wise it shouldn't be noticeable, on an mi100 node it takes ~3.5sec.

Also adjust XFAILs.

[bader]

I'm not sure if it works. libclc is not enabled by default (only with CUDA/HIP back-ends), so I assume there might be problems if this library is not built. Right?

I suggest adding a separate command to this GitHub Actions script: https://github.com/intel/llvm/blob/sycl/.github/workflows/sycl_linux_build_and_test.yml#L136.
@alexbatashev can help with that.

[jchlanda]

You are right: https://github.com/jchlanda/llvm/blob/jakub/check_sycl_check_libclc/buildbot/configure.py#L62
Do you think it's worth adding it to the list of llvm_enabled_projects by default?

[bader]

No, it's a dead code by default, but it must be enabled in pre-/post-commit CI validation.
It might be worth adding when CUDA or HIP backends included to the build, but I'd like to avoid the drawback mentioned in the description:

The only slight drawback is that if check-libclc fails it will not proceed to running check-sycl.

To avoid this, we can run check-libclc separately from check-sycl as we do for the compiler (i.e. check-clang).
If we need simplify the workflow for DPC++ compiler developers we can make check-sycl-toolchain CMake target with dependencies on check-* testing the compiler, runtime and other DPC++ components, which can be executed concurrently and do not block each other. Python script instead of cmake target works as well.

[alexbatashev]

@jchlanda Alexey is right, you don't need to add dependencies to the check-sycl. Please, adjust configure.py call here and just add a new step for the build job as @bader described earlier. All the logs will be public, both pre-commit and post-commit testing will be enabled, and you'll be able to modify the run command as you see fit.

[jchlanda]

I've folded it into: b8f99bc
It'll build libclc in ci_default mode and run it through linux build and test. I've decided against adding another target.

[alexbatashev]

LGTM

[alexbatashev]

@jchlanda do these logs look right to you?
https://github.com/intel/llvm/runs/4392597907?check_suite_focus=true

UPD. I guess since no targets were requested, no tests were run (-DLIBCLC_TARGETS_TO_BUILD= in the logs).

[jchlanda]

[jchlanda]

@jchlanda do these logs look right to you? https://github.com/intel/llvm/runs/4392597907?check_suite_focus=true

UPD. I guess since no targets were requested, no tests were run (-DLIBCLC_TARGETS_TO_BUILD= in the logs).

Yeah, that's exactly it, it was enabled but, like you say, no libclc targets were selected: https://github.com/intel/llvm/runs/4392597907?check_suite_focus=true#step:5:605
I'm not sure what the best way to handle it would be. Is there a safe libclc target that can be enabled by default on the CI machines?

[bader]

@jchlanda, @alexbatashev, sorry, I forgot that this comment is not addressed yet:

I guess since no targets were requested, no tests were run (-DLIBCLC_TARGETS_TO_BUILD= in the logs).

So, check-libclc is executed by CI, but it does nothing today. We need another change to make this step useful. :)

[jchlanda]

yeap, but I don't know if there is a default libclc target that CI machines should be building?

[bader]

yeap, but I don't know if there is a default libclc target that CI machines should be building?

I assume you wanted to test libclc for nvptx and amdgcn targets. So, I suppose corresponding triples can be set in the python script. Right? Enabling these in CI doesn't seem to require specific HW and can executed. @alexbatashev, what do you think about this? AFIAK, other targets are not used in our project.
https://github.com/intel/llvm/blob/sycl/buildbot/configure.py#L70
https://github.com/intel/llvm/blob/sycl/buildbot/configure.py#L77

[jchlanda]

yeap, but I don't know if there is a default libclc target that CI machines should be building?

I assume you wanted to test libclc for nvptx and amdgcn targets. So, I suppose corresponding triples can be set in the python script. Right? Enabling these in CI doesn't seem to require specific HW and can executed. @alexbatashev, what do you think about this? AFIAK, other targets are not used in our project. https://github.com/intel/llvm/blob/sycl/buildbot/configure.py#L70 https://github.com/intel/llvm/blob/sycl/buildbot/configure.py#L77

Out of the box check-libclc will complain about the device libraries, I've just tried amdgcn--;amdgcn--amdhsa on my dev machine, that has no ROCm installed and it fails with:

clang-14: error: cannot find ROCm device library; provide its path via '--rocm-path' or '--rocm-device-lib-path', or pass '-nogpulib' to build without ROCm device library

[bader]

Do we need ROCm device library to run libclc tests? I see that test just emit LLVM.

[jchlanda]

Do we need ROCm device library to run libclc tests? I see that test just emit LLVM.

I've addressed it in: #5091