Kokkos Kernels - GEMMA

[crtrott]

Author: @vqd8a

[vqd8a]

This session has two talks:

• Kokkos Kernels: Stream Interfaces of SpILUK and SPTRSV:
  • Abstract: Sparse incomplete LU factorization (ILU) and sparse triangular solvers have been widely used as preconditioner in iterative solution algorithms. Kokkos Kernels has provided two kernels SPILUK and SPTRSV for incomplete LU factorization and triangular solve, respectively. These kernels were integrated into Trilinos Ifpack2's preconditioner RILUK and Ifpack2's local triangular solver. Although SPILUK and SPTRSV are robust but it is hard to achieve good parallelism due to the fact that these are based upon algorithms that are inherently sequential. In this talk, we will introduce a new feature which has been released recently in Kokkos Kernels. We focus on implementing less expensive SPILUK and SPTRSV kernels on GPUs using stream executions.
• A Distributed-Memory Schur-complement PCA Preconditioner for Gemma Ill-conditioned Problems
  • Abstract: This presentation describes our first effort toward an implementation of Schur-complement principal component analysis (PCA) preconditioner for distributed-memory computing platforms. A distributed binary tree structure for binary hierarchical matrix partitioning is proposed. This implementation targets performance portability by leveraging the abstractions provided by Kokkos and Kokkos Kernels. We evaluate the implementation via an ill-conditioned problem with an iterative solver in terms of effectiveness and efficiency.

KokkosKernels_StreamInterface_SpILUK_SpTRSV_KUG2023_v2.pptx

Gemma_DistributedSchurPCAPreconditioner_KUG2023.pptx