Custom kernels Intel XPU backend in LibTorch C++ API

I guess that it is only about the accessor type that I need to use. Or do I need to use the raw pointer?

Since the custom kernels still depend on SYCL compiler to build source code, we support both accessor type and raw pointer. Here is an example to use raw pointer.
torch-xpu-ops/src/ATen/native/xpu/sycl/NonzeroKernel.cpp at main · intel/torch-xpu-ops