Jan André Reuter

The issue here is, that OneAPI implements the `_emi` callbacks as their normal ones. This affects at least `target`, `target_data_op` and `target_submit`. You can still register callbacks with the callback...

> is it possible to execute llvm-clang plugin/pass using syclcc? if so how we can test llvm-clang plugin/pass using syclcc? Looking at the help of `acpp`, unknown flags are passed...

Here's the output with the environment variable. It seems like it depends on the compiler if the code above works or not: ```console $ amdclang -O0 -fopenmp --offload-arch=gfx90a reproducer.c -I/opt/rocm/include/hip...

Seems like the difference is the flag `--hip-link` which is added by `hipcc` during compilation ```console $ amdclang -fopenmp --offload-arch=gfx90a -I/opt/rocm/include/hip -D__HIP_PLATFORM_AMD__ reproducer.c -lamdhip64 --hip-link $ ./a.out Number of devices...

After testing a bit more, I found out why `--hip-link` solves the issue. The flag also removes the library for OpenMP target offloading `-lomptarget` (`libomptarget.so`). Adding the flag causes the...

Using a very small test program just calling `hipGetDeviceCount` from main works totally fine with `amdclang` and OpenMP target offloading (tested on a system with no AMD GPU): ```console $...

I don't know if that's a thing which can be fixed from the HIP side or if work at the LLVM toolchain is required. However, I would at least expect...

The issue probably is not resolved yet with ROCm, as it was just recently resolved in LLVM 19git. I'll recheck tomorrow. In general, I would consider this as an LLVM...

I can confirm that the issue is fixed with LLVM 19git and will therefore eventually also land in ROCm. As the limitation seems to come from tying to call `hip`...

> When exactly does this need to be used? I.e. what is the effect of using "oneapi" over "intel"? If I understood you correctly the configure used by those projects...