Alexey Bader

> is there a CUDA installation on the CI machine that does the Jenkins build? Building the libdevice objects requires normal nvptx backend `-c` compilation. However the log is saying...

@hdelan, do I understand it correctly, that this change adds dependency on CUDA SDK to build libclc for NVPTX? I guess today we use clang to emit NVVM, which doesn't...

Okay. Thanks. Let's test.

/verify with https://github.com/intel/llvm-test-suite/pull/1112

Tests pass on Linux, but failed on Windows. I think failures are not related to your patch and probably will be fixed by https://github.com/intel/llvm/pull/6517. @aelovikov-intel, am I right?

> The other failing tests I do not believe are related to this PR. Agree. @npmiller took care about the failures on HIP - see https://github.com/intel/llvm/pull/6519#issuecomment-1203755950. Could you do the...

@intel/llvm-reviewers-cuda, @intel/llvm-reviewers-runtime, ping.

This looks like a limitation of level zero API/plug-in. See https://github.com/intel/llvm/blob/sycl/sycl/plugins/level_zero/pi_level_zero.cpp#L5686-L5688. I suppose if you switch to OpenCL back-end you should see expected results.

> **what is the time 0 for opencl backend?** Sorry, I don't get what do you mean here. Could you add more context, please?

You can find the description of returned values in OpenCL spec: https://registry.khronos.org/OpenCL/specs/3.0-unified/html/OpenCL_API.html#profiling-operations. DPC++ runtime returns to user values reported by OpenCL API as is.