RCCL

Wrapper of the RCCL (Radeon Collective Communication Library) for Julia. The bindings to the C functions were generated with Clang.jl. The API is very similar to that of NCCL.jl and passes the same set of tests as that.

General notes

A couple of implementation details differ between RCCL.jl and NCCL.jl due to discrepancies between the AMDGPU.jl and CUDA.jl; specifically:

• CUDA.jl exports a CUstream type which is simply an alias of cuStream_t which then gets wrapped in a CuStream struct. CUstream can be passed to all NCCL C functions since it is basically a C type. AMDGPU.jl exposes a Julia struct called HIPStream which contains a handle to a hipStream_t C type (which itself is not exported). Thus, when @ccalling RCCL functions, a simple Ptr{Cvoid} is passed. This should change if somehow a hipStream_t is exposed by AMDGPU.jl;
• CUDA.jl has a