An Optimized Generalized Multi-Color Point Implicit Solver for Intel GPUs Using Intel OneAPIs ESIMD

The growing popularity of Computational Fluid Dynamics (CFD) simulations among engineers necessitates the use of GPU acceleration for increased efficiency. NASA FUN3D offers GPUaccelerated CFD simulations using unstructured grids across the speed regime from incompressible to hypersonic flows involving reentry. This work focuses on the generalized multi-color point implicit solver used in FUN3D, accounting for roughly half of the run time. Specifically, this work focuses on developing three optimized multi-color linear-solver kernels for the Intel Data Center Max 1550 GPU that is available on the Argonne Leadership Computing Facility’s (ALCF) exascale machine, Aurora. These optimized kernels work for a range of block sizes. Intel oneAPI’s ESIMD is used to explicitly vectorize memory loads for block-sparse vector computation, which is the core computation of the multi-color linear solver. The performance of the optimized implementations are evaluated on Intel GPUs, testing both on Aurora and internal Intel systems. The resulting performance is within 10% of the bandwidth benchmark peak. Additionally, results on the Intel GPU are compared with the CUDA-optimized code on NVIDIA hardware. The CUDA-optimized implementations on NVIDIA A100 demonstrate superior absolute performance; however, the Intel GPU achieves higher performance as a percentage of bandwidth benchmark peak, indicating more efficient utilization of available memory bandwidth in the proposed implementation.