[CUDA][HIP] get_native() API doesn't work for device

[densamoilov]

get_native() doesn't work for device when CUDA or HIP backend is used.

Reproducer:

#include <sycl/sycl.hpp>

int main() {
    auto dev = sycl::device(sycl::gpu_selector{});
    auto ctx = sycl::context(dev);

    auto cuda_ctx = sycl::get_native<sycl::backend::ext_oneapi_cuda>(ctx);
    auto cuda_dev = sycl::get_native<sycl::backend::ext_oneapi_cuda>(dev);
    return 0;
}

Error message:

In file included from /buildspace/dpcpp/llvm/build/install/bin/../include/sycl/sycl.hpp:16:
/buildspace/dpcpp/llvm/build/install/bin/../include/sycl/backend.hpp:128:10: error: reinterpret_cast from 'pi_native_handle' (aka 'unsigned long') to 'backend_return_t<(backend)'\x03', device>' (aka 'int') is not allowed
  return reinterpret_cast<backend_return_t<BackendName, SyclObjectT>>(
         ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1.cpp:8:31: note: in instantiation of function template specialization 'sycl::get_native<sycl::backend::ext_oneapi_cuda, sycl::device>' requested here
    auto cuda_dev = sycl::get_native<sycl::backend::ext_oneapi_cuda>(dev);
                              ^
1 error generated.

I tried DPC++ daily 2022-08-22 and oneAPI DPC++ Compiler 2022-06, the result is same.

[densamoilov]

Type of a native device for CUDA and HIP backends is just int and not a pointer therefore using reinterpret cast here causes the issue.

[fxzjshm]

Sorry to disturb, #6649 fixes only for CUDA backend but not HIP, so this issue shouldn't be closed.

[steffenlarsen]

@npmiller for awareness.