How are we supposed to cross compile sycl applications?

[mccakit]

For native building we have to download CUDA and HIP.
What happens when we cross compile sycl from Linux to Windows?
Since it is not a runtime(yet) I assume we have to cross compile the whole toolchain.
But how do we cross compile the toolchain when CUDA Toolkit and HIP are Native installs

Do we just download binaries for them, point them in llvm install dir with some env vars then just cross compile as normal? Or something else?

A lot of development tools don't work on non linux systems, so having one host is very important

[bratpiorka]

I think there are two different “cross compile” scenarios here, and the answer depends on which one you mean. Could you share a concrete workflow you want (e.g., build SYCL apps on Linux that produce Windows binaries, or actually build the whole DPC++/SYCL toolchain for Windows while running on Linux)?

For CUDA/HIP backends you generally need target-platform SDK artifacts (headers + import/libs + tooling) that match the OS you’re producing binaries for. For Linux to Windows cross compilation you need a Windows sysroot (Windows SDK + CRT/MSVC runtime/import libs, etc.), and Windows CUDA Toolkit / Windows HIP SDK pieces (headers + .lib stubs, not Linux .so).

If you want to build, on Linux, a SYCL-based application that runs on Windows, this is currently not a supported scenario. It might be technically possible with a carefully prepared environment and correctly configured paths, but I have never tried it. In my opinion, performing a separate native build on Windows is the easier and more reliable approach.