High Performance OpenMP for GPUs
TimeWednesday, November 14th11:30am - 12pm
DescriptionOpenMP has a 20 year history in HPC. It was the parallel programming model of choice for large SMP servers, and later proved effective for hybrid MPI+OpenMP programming on scalable systems. As core counts rose, best practices for OpenMP implementations and programming have evolved. The early focus was on maximizing efficiency of loop-level parallelism. More recently a global, outer-parallel approach has been used to improve scalability on nodes with large numbers of CPU cores. GPUs are optimized to deliver high performance and throughput on massively parallel, regular, loop-structured algorithms. In this talk, we'll survey the features of the latest OpenMP standards, describe how OpenMP programs can exploit the performance potential of today's GPUs, and summarize the implications for a high-performance OpenMP implementation that supports both multicore CPUs and GPU accelerators.