Accelerating Computational Fluid Dynamics on RISC-V: Vectorization of OpenFOAM’s Multigrid Solver

Computational Fluid Dynamics (CFD) applications rely heavily on memorybound sparse linear algebra kernels like Sparse Matrix-Vector multiplication (SpMV). Legacy frameworks such as OpenFOAM exploit MPI parallelism but lack effective support for modern vector architectures, largely due to internal matrix formats that prevent contiguous memory accesses. With the emergence of the RISC-V Vector Extension (RVV), vector architectures offer a compelling alternative to traditional SIMD and GPUs. This thesis investigates this potential by optimizing a state-of-the-art OpenFOAM multigrid solver across two contrasting RISC-V architectures: the long-vector EPAC accelerator prototype and the commercial short-vector Sophon SG2044 processor. To overcome native memory bottlenecks, the main contributions of this work include: i) vectorizing the SpMV kernel using RVV intrinsics and vector-friendly sparse formats; ii) developing a custom smoother plugin for OpenFOAM that performs runtime data conversion; and iii) a comprehensive speedup and scalability evaluation scaling the problem size for the isolated SpMV kernel, the custom smoother, and the end-to-end CFD simulation. Experimental results demonstrate a 6x speedup for the custom smoother on the EPAC test chip and a 1.5x speedup on the SG2044. Furthermore, endto-end CFD simulation evaluations reveal overall performance improvements of 1.62x and 1.16x on the EPAC and SG2044 architectures, respectively. Ultimately, this work proves that legacy CFD codes can be successfully modernized and accelerated using emerging RISC-V hardware.