[OMPD] The LLVM OMPD ompd_get_task_frame() function does not seem to support NVIDIA GPUs

[jdelsign]

The LLVM OMPD ompd_get_task_frame() function does not seem to support the NVIDIA GPUs. It treats a task handle for a GPU task as if it were a CPU task. See: https://github.com/OpenMPToolsInterface/LLVM-openmp/blob/ompd-devices/libompd/src/omp-debug.cpp#L1216

As a result, it attempts to lookup symbols that are not defined in the CUDA ELF images. For example, it attempts to lookup "ompd_access__kmp_taskdata_t__ompt_task_info_" and "ompd_sizeof__kmp_taskdata_t__ompt_task_info_", which are not defined in the NVIDIA GPU CUDA ELF symbol tables.

[jprotze]

This is fixed in the ompd-device-50 branch to the point, that the library does not try to access host data structures.
@mcfadden8 do you have an idea how the device runtime could provide this information?

[mcfadden8]

@jprotze, I am sorry, but I don't know the answer to this.