1 pointby tribe-iro4 hours ago1 comment
  • tribe-iro4 hours ago
    iro-cuda-ffi is a small, stable bridge that lets Rust programs call CUDA C++ kernels that are built ahead of time with NVIDIA’s own toolchain (nvcc). It’s designed for teams who already have CUDA code, or want to keep writing it in standard .cu files, and still drive it safely from Rust. That means you can keep using familiar NVIDIA tooling like Nsight and cuda-gdb, and you don’t have to rewrite or wrap kernels as strings or scripts at runtime. What makes it different is the approach: it doesn’t do runtime compilation or a Rust‑specific GPU language; it uses the same nvcc‑compiled CUDA C++ you already have. That keeps the existing CUDA code and tests intact, while giving Rust a clean, minimal ABI to launch kernels with. The result is a straightforward way to reuse NVIDIA‑optimized code paths inside Rust without changing your CUDA workflow.