Cluda: Gallium3D API Implemented Atop Nvidia CUDA Driver API (gitlab.freedesktop.org)

Cluda is a new project that implements the Gallium3D driver API on top of NVIDIA’s CUDA driver API, effectively letting Mesa’s Gallium-based stack talk to NVIDIA GPUs via cuDriver (libcuda) rather than a native Gallium kernel driver. The implementation maps Gallium abstractions—contexts, buffer objects, command submission and synchronization—onto CUDA primitives (contexts, cuMem allocations, kernels and streams), and routes shader and compute workloads into CUDA-executable code (e.g., by compiling Mesa’s intermediate representation to PTX or CUDA kernels). That approach provides a non‑standard but pragmatic path for running Mesa/Gallium frontends on NVIDIA hardware where a lightweight, CUDA-backed path is useful. For the AI/ML community this matters because it can broaden how GPU-accelerated compute stacks are prototyped and deployed: researchers and projects that target Gallium/Mesa can experiment on NVIDIA hardware without relying on vendor OpenGL/Vulkan drivers or heavy kernel-driver work. It could simplify integration with CUDA-first ML tooling and speed iteration for compute-shader driven workloads, but comes with caveats—performance depends on translation overhead and CUDA’s exposed features, some GPU functionality may be missing or emulated, and compatibility/security/licensing constraints remain. Overall Cluda is a notable interoperability experiment that could lower friction between Mesa’s ecosystem and NVIDIA’s CUDA platform.