Portable Linear Solvers for High-Order Spectral Element Methods on GPUs

The diversification in hardware architectures has become a challenge for computational science: software stacks implemented for a specific hardware architecture often fail to port to other systems. To counter this problem, simulation software stacks increasingly rely on portability layers or software stacks that feature backends for different hardware architectures. We present Ginkgo, a math library that takes platform portability as a central design principle and can be used for numerical calculations in the nekRS CFD code. A runtime and scalability analysis for CFD applications demonstrates that Ginkgo enables platform portability at high performance and scalability.