tl;dr: Researchers built a multi-vendor GPU compilation framework directly into rustc and LLVM, using Rust's ownership and aliasing guarantees to safely manage GPU data transfers without vendor lock-in or unsafe raw pointers. The approach uses a two-pass compilation pipeline to handle ABI mismatches between host and device targets. Benchmarks on RAJAPerf show the generated GPU kernels achieve performance competitive with hand-optimized CUDA and HIP C++ baselines.