Hyperbolic Diffusion in Flux Reconstruction: Optimisation through Kernel Fusion within Tensor-Product Elements

Abstract: Novel methods are presented in this initial study for the fusion of GPU kernels in the artificial compressibility method (ACM), using tensor product elements with constant Jacobians and flux reconstruction. This is made possible through the hyperbolisation of the diffusion terms, which eliminates the expensive algorithmic steps needed to form the viscous stresses. Two fusion approaches are presented, which offer differing levels of parallelism. This is found to be necessary for the change in workload as the order of accuracy of the elements is increased. Several further optimisations of these approaches are demonstrated, including a generation time memory manager which maximises resource usage. The fused kernels are able to achieve 3-4 times speedup, which compares favourably with a theoretical maximum speedup of 4. In three dimensional test cases, the generated fused kernels are found to reduce total runtime by ${\sim}25\%$, and, when compared to the standard ACM formulation, simulations demonstrate that a speedup of $2.3$ times can be achieved.