摘要
相对于传统的串行程序移植,并行系统间的代码移植因体系结构间的巨大差异而变得极为复杂。为此,针对统一计算设备架构(CUDA)程序向其他异构多核平台的移植,提出CUDA架构到Cell的映射方案。通过模型映射、并行粒度提升、共享变量清除和运行时优化,使CUDA程序的大规模并行线程可以在Cell平台上正确执行。实验结果证明,翻译后的程序在Cell的执行效率可达到Cell平台上手动编写程序的72%。
Compared with traditional serial program migration,parallel program migration becomes sophisticated for the huge diversity of different architectures.To migrate Compute Unified Device Architecture（CUDA） programs to other heterogeneous multi-cores,a method of mapping CUDA architecture to Cell is proposed.Through executing model mapping,enhancing parallel granularity,memory mapping and optimization,the mass threads in CUDA can execute correctly in Cell architecture by source code migration.Experimental result shows the executing speed of translated programs can achieve 72% of native compiled programs.
出处
《计算机工程》
CAS
CSCD
2012年第24期279-282,共4页
Computer Engineering
基金
国家"863"计划基金资助项目(2009AA012201)
"核高基"重大专项(2009ZX01036-001-001)
河南省重大科技攻关计划基金资助项目(092101210501)
关键词
源代码移植
异构多核
模型映射
共享变量清除
运行时优化
source code migration
heterogeneous multi-core
model mapping
shared variable removing
runtime optimization
