A research paper presenting a zero-overhead GPU offload framework built natively into rustc and LLVM that uses Rust's ownership and type system to safely manage GPU data transfers. The two-pass compilation pipeline resolves cross-vendor ABI mismatches and achieves competitive kernel performance against hand-optimized CUDA and HIP C++ baselines on RAJAPerf.
Background
GPU programming has historically been dominated by CUDA (NVIDIA) and HIP (AMD), with Rust emerging as a safer alternative but lacking native GPU offload support in its compiler. This work integrates GPU offloading directly into rustc and LLVM to bridge that gap.
- Source
- Lobsters
- Published
- Aug 18, 2026 at 08:16 PM
- Score
- 6.0 / 10