Kokkos Kernels - Batched Algorithms

[crtrott]

Authors: @srajama1 @kliegeois

[kliegeois]

KUG2023.pdf

[kliegeois]

Applications such as multilevel finite element methods, modeling collision in plasma and reentry simulations are reformulating algorithms to match the hierarchical parallel manycore CPUs/GPUs. This reformulation results in a need for linear solvers at different levels of hierarchical parallel algorithms. Several of these linear systems are small or medium size, independent, linear systems that can be solved at the same time. Depending on the problems, some of these linear systems are sparse and others dense. The community has been focused on solving dense linear systems as a batch, or solving multiple small dense linear systems. However, the work has focused primarily on one level of hierarchical parallelism - using an entire accelerator to solve the linear systems. We have been focused on solving batched dense/sparse linear systems, both at the device level and at the team level so applications can use these options effectively. We will present recent results on CPUs/GPUs using our hierarchical parallel batched linear solver implementations.