The International Conference for High Performance Computing, Networking, Storage, and Analysis
The International Conference for High Performance Computing, Networking, Storage, and Analysis
Authors: Weiling Yang, Jianbin Fang, Dezun Dong, and Xing Su (National University of Defense Technology (NUDT), China) and Zheng Wang (University of Leeds)
Abstract: General matrix multiplication (GEMM) is a key subroutine in high-performance computing. While the existing BLAS libraries can obtain near peak hardware performance on large GEMM, they deliver low performance on small and irregularly-shaped GEMM. We propose and implement an algorithm for small and irregularly-shaped GEMM, which can effectively reduce the overhead of packing operation. In addition, we also propose a novel parallelization method to tap the computing potential of multi-core CPUs. Our work achieves high hardware utilization through a series of optimizations. Our experiments show that on ARMv8-based processors, LibShalom can achieve about a 1.05x - 3x performance speedup over mainstream BLAS libraries, including OpenBLAS, ARMPL, BLIS, LIBXSMM and BALSFEO.
Presentation: file
Back to Technical Papers Archive Listing