Poster 21: Optimization of GPU Kernels for Sparse Matrix Computations in Hypre

Student: Chaoyu Zhang (Arkansas State University)
Supervisor: Ruipeng Li (Lawrence Livermore National Laboratory)

Abstract: The acceleration of sparse matrix computations on GPUs can significantly enhance the performance of iterative methods for solving linear systems. In this work, we consider the kernels of Sparse Matrix Vector Multiplications (SpMV), Sparse Triangular Matrix Solves (SpTrSv) and Sparse Matrix Matrix Multiplications (SpMM), which are often demanded by Algebraic Multigrid (AMG) solvers. With the CUDA and the hardware support of the Volta GPUs on Sierra, the existing kernels should be further optimized to fully take the advantage of the new hardware, and the optimizations have shown significant performance improvement. The presented kernels have been put in HYPRE for solving large scale linear systems on HPC equipped with GPUs. These shared-memory kernels for single GPU are the building blocks of distributed matrix operations required by the solver across multiple GPUs and compute nodes. The implementations of these kernels in Hypre and the code optimizations will be discussed.

ACM-SRC Semi-Finalist: no

Poster: PDF
Poster Summary: PDF


Back to Poster Archive Listing