计算机体系结构论文CAPRI Prediction of Compaction-Adequacy for Handling Control-Divergence in GPGPU Architectures (isca2012).pdfVIP

CAPRI: Prediction of Compaction-Adequacy for Handling Control-Divergence in GPGPU Architectures Minsoo Rhu Mattan Erez Electrical and Computer Engineering Department Electrical and Computer Engineering Department The University of Texas at Austin The University of Texas at Austin minsoo.rhu@ mattan.erez@ Abstract control. When a conditional branch requires some lanes to execute the true path and other lanes to execute the Wide SIMD-based GPUs have evolved into a promising plat- false path, performance degrades. This is because only form for running general purpose workloads. Current pro- grammable GPUs allow even code with irregular control to ex- a single instruction is issued to all SIMD lanes, imply- ecute well on their SIMD pipelines. To do this, each SIMD lane ing that only a subset of the lanes should actually exe- is considered to execute a logical thread where hardware ensures cute operations and commit results. Recent research has that control flow is accurate by automatically applying masked shown that the impact of this control divergence problem execution. The masked execution, however, often degrades perfor- can be reduced by dynamically forming SIMD-instructions mance because the issue slots of masked lanes are wasted. This from large collections of threads [8, 19]. These collections degradation can be mitigated by dynamically compacting