Performance Tradeoffs in Shared-memory Platform Portable Implementations of a Stencil Kernel

Abstract

Building on a significant amount of current research that examines the idea of platform-portable parallel code across different types of processor families, this work focuses on two sets of related questions. First, using a performance analysis methodology that leverages multiple metrics including hardware performance counters and elapsed time on both CPU and GPU platforms, we examine the performance differences that arise when using two common platform portable parallel programming approaches, namely OpenMP and VTK-m, for a stencil-based computation, which serves as a proxy for many different types of computations in visualization and analytics. Second, we explore the performance differences that result when using coarserand finer-grained parallelism approaches that are afforded by both OpenMP and VTK-m.