Challenges Faced When Porting Ginkgo to the SYCL* Ecosystem

Yu-Hsiang (Mike) Tsai presents his experience of a platform-portable sparse linear algebra library: Ginkgo. Previously, he ported essential parts of his library (like SpMV or Krylov solvers) to oneAPI based on SYCL* v1.2.1. He presents:

  • Challenges and solutions when porting functionality and for staying up-to-date due to the changes in SYCL 2020
  • Issues with per-kernel code splitting and supporting various hardware

Yu-Hsiang generates kernels suitable for different devices and gives an approach for using precompiled kernels. He shows some experimental results of using preconditioners or algebraic multigrid methods.



Yu-Hsiang (Mike) Tsai is a PhD student at the Karlsruhe Institute of Technology. He joined the Ginkgo team in 2019. Currently, he focuses on designing mixed-precision algorithms, algebraic multigrid methods, and accelerating GPU kernels for Advanced Micro Devices (AMD), Intel, and NVIDIA*.