Characterizing Optimizations to Memory Access Patterns using Architecture-Independent Program Features

Abstract

High-performance computing developers are faced with the challenge of optimizing the performance of OpenCL workloads on diverse architectures. e Architecture-Independent Workload Characterization (AIWC) tool is a plugin for the Oclgrind OpenCL simulator that gathers metrics of OpenCL programs that can be used to understand and predict program performance on an arbitrary given hardware architecture. However, AIWC metrics are not always easily interpreted and do not re ect some important memory access pa erns a ecting e ciency across architectures. We propose a new metric of parallel spatial locality -the closeness of memory accesses simultaneously issued by OpenCL workitems (threads). We implement the parallel spatial locality metric in the AIWC framework, and analyse gathered results on matrix multiply and the Extended OpenDwarfs OpenCL benchmarks. e di erences in the observed parallel spatial locality metric across implementations of matrix multiply re ect the optimizations performed. e new metric can be used to distinguish between the OpenDwarfs benchmarks based on the memory access pa erns a ecting their performance on various architectures. e improvements suggested to AIWC will help HPC developers be er understand memory access pa erns of complex codes and guide optimization of codes for arbitrary hardware targets.