MC: A Software Solution to Reduce Errors Caused by Incoherent Caches on GPUs in Unstructured Graphic Algorithm

摘要

Recently, researchers have focused on addressing incoherent caches on GPUs as current GPUs lack hardware to support that. Moreover, the support for inter-block communication also lacks which limits the scalability of parallel programming especially in the unstructured algorithm in which program would share data between different threads. Barrier synchronization can be a solution but it becomes invalid because of incoherent caches. In this paper, we propose a set of rules for programming on current GPUs to avoid the errors caused by incoherent caches when applying barrier synchronization. We also leverage these rules into an unstructured graphic algorithm - constrained Delaunay triangulation. In comparison with traditional ways such as (1) disabling L1 cache and (2) using keyword volatile, we find that when vertices is over 300K the error rate is lesser than (1) by 91.19% and than (2) by 84.2% on average.