利用访存模式构建GPU高效率数据访问

Employing Access Model to Structure GPU Efficient Data Access

针对访存相对密集的应用,提出了一种基于访存模式高效率数据访问技术。该技术结合应用程序的访存特性和GPU的片上高速共享存储器特性减少应用程序对高延迟片外存储访问的次数从而提高系统数据访问的效率,通过在不同架构的GPU上进行了验证,分别取得了N卡最高9倍和A卡最高8倍的加速效果,并对各个优化策略在不同架构GPU上取得效果的原因进行了分析。

Aiming at the relatively intensive applications of memory access,an efficient data access technology based on access model is proposed.This techonology combines the memory access characteristics of applications and the on-chip high-speed shared memory characteristics of GPU to reduce the number of applications accessing off-chip menory with high latency so as to improve the efficiency of system data access.By testifying on GPU of different architectures,the acceleration effects of N card up to nine times and A card up to eight times are achieved respectively,and the reasons for the effectiveness of each optimization strategy on different architecture GPU are analyzed.