0ee3ebcb5b
Since cling was ported to LLVM 13, it is required that the `__cuda_register_globals` function and the `__cuda_fatbin_wrapper` and `__cuda_gpubin_handle` variables are unique when defining a CUDA kernel. Otherwise, the JIT is lazy and reuses the compiled version of `__cuda_register_globals`, `__cuda_fatbin_wrapper` and `__cuda_gpubin_handle` from the first CUDA kernel definition for all subsequent CUDA kernel definitions, which in practice means that the PTX code from the first kernel is re-registered each time. Increase the default CUDA SM level to 35 because SM 20 is deprecated or removed in the current CUDA SDK versions.