Detailed Description

Backend-neutral access to cub's block- and device-level primitives.

hipCUB is a close mirror of cub: same algorithm names, same two-call (size-then-run) shape, same block-level collectives — it differs only in header root (<hipcub/hipcub.hpp> vs <cub/cub.cuh>) and namespace (hipcub vs cub). Aliasing the namespace here lets the existing cub::DeviceScan, cub::BlockScan, cub::DeviceReduce, cub::DeviceRadixSort, and cub::BlockRadixSort call sites compile unchanged on both backends.

Include this instead of <cub/...> directly. Doing so matters beyond tidiness: on systems where a CUDA toolkit is also installed and on the compiler's default include path (CPATH — the usual setup on Cray/HPE machines with a cudatoolkit module loaded), a bare #include <cub/cub.cuh> silently resolves to NVIDIA's cub during a HIP build and drags the whole CUDA runtime in with it.

The scratch-storage boilerplate that wraps the device-level calls lives in backend/algorithms.h; this header is just the include + namespace.