可将科学计算中大量算法的计算形式视为由流计算和相当比例的通用计算混合而成.针对低并行度计算以及不易流化(Streamlization)的数据结构对流计算整体性能具有较大影响,提出了一种软、硬件可控的适应性片上存储结构DAMS Cache.该结构能够同时适应混杂流计算中流数据以及标量数据的存储需求；采用了适应性动态存储资源分配策略和适应性动态地址映射策略解决地址映射冲突问题；通过全硬件支持非规则流、条件流的存储与访问,混合数据替换策略能够充分挖掘数据的生产者-消费者局部性及时间、空间局部性.验证评估实验表明,相对Cache以及SPM(Scratchpad Memory),DAMS Cache算法的适应性较好,面向混杂流计算的性能较优.%In scientific applications, the computing process is composed of stream computing and general computing. The performance of computing is limited by low parallelism and hard streamlization. We propose what we call DAMS-Cache (DAMS stands for Dynamical Address Mapping Stream) , which can be controlled by both the software and hardware that manage an on-chip memory structure that can, we believe, suppress effectively the above-mentioned limitation. Sections 1 through 4 of the full paper explain DAMS-Cache, whose core consists of; (1) the memory system can support both stream computing and scalar data processing in high performance, which supports irregular stream and conditionally-loaded stream computing by hardware; (2) the DAMS Cache explores both coarse-grained producer-consumer locality and fine-grained temporal/spatial locality by using replacement strategy of mixed data structure; (3 ) in order to avoid address mapping conflict, the dynamically adaptive storage resource assignment strategy and dynamically adaptive address mapping strategy are also presented. Section 5 presents performance analysis of DAMS Cache; it shows preliminarily that DAMS Cache has better adaptability and high performance compared with Scratchpad Memory and traditional Cache.
展开▼
