基于GPU的并行优化技术

Parallel optimize technology based on GPU

针对标准并行算法难以在图形处理器(GPU)上高效运行的问题,以累加和算法为例,基于Nvidia公司统一计算设备架构(CUDA)GPU介绍了指令优化、共享缓存冲突避免、解循环优化和线程过载优化四种优化方法。实验结果表明,并行优化能有效提高算法在GPU上的执行效率,优化后累加和算法的运算速度相比标准并行算法提高了约34倍,相比CPU串行实现提高了约70倍。

Standard parallel algorithm cannot work efficiently on GPU. This paper took reduction algorithm for example, introduced four parallel optima methods for NVIDIA＇ s graphics processor unit （GPU） which supported CUDA architecture. These methods included instruction optimize and shared memory conflict avoid and loop unroll and threads overload optimize. The experiment result shows that： parallel optimize can significantly speed up the GPU compute speed. The optimized reduction algorithm is 34 times faster than standard parallel algorithm and 70 times than CPU-based implementation.