sycl::devices created via interop API are not uniqued

[Ralender]

Describe the bug
If a device is created via the interop API it will always create a new device_impl even when there already y exist a device_impl with the same backend handle. this issue is not limited to opencl or any device in particular.

To Reproduce
create a file named test.cpp containing

#include <sycl/sycl.hpp>

int main() {
  sycl::device sdev(sycl::default_selector{});
  auto ocl_dev = sycl::get_native<sycl::backend::opencl>(sdev);
  sycl::device dev(sycl::make_device<sycl::backend::opencl>(ocl_dev));
  assert(sdev == dev);
}

clang++ -fsycl -o a.out test_interop_opencl.cpp

clang++ being the sycl capable compiler

./a.out

this will hit the assert even thought (in my case the opencl implementation returned the same device handle)

Environment (please complete the following information):

• OS: Any
• Target device and vendor: Any
• DPC++ version: e25f199

[JackAKirk]

Fix proposed here: #6204

[keryell]

Thanks!

[JackAKirk]

Reopened this since the fix was only for the cuda backend and opencl behavior is unchanged from the originally reported issue.

[AlexeySachkov]

I think that this is fixed by #13483. Tagging @aelovikov-intel to comment here

[aelovikov-intel]

I think that this is fixed by #13483. Tagging @aelovikov-intel to comment here

I believe I couldn't fix it for OpenCL because there is CPU backend where the number of all potential (sub-)sub-devices is exponential.

[AlexeySachkov]

I think that this is fixed by #13483. Tagging @aelovikov-intel to comment here

I believe I couldn't fix it for OpenCL because there is CPU backend where the number of all potential (sub-)sub-devices is exponential.

Ok, but it is at least fixed for top-level devices as illustrated by a test you added, right?

[aelovikov-intel]

Ok, but it is at least fixed for top-level devices as illustrated by a test you added, right?

Can't say that I remember, but it definitely looks so based on the changes in the PR.