摘要
当前主流片上总线协议——AHB存在访存带宽利用率较低的问题.本文基于SoC内DMA传输较多的特点,提出一种新的优化设计:在内存控制器内部增加MCS-DMA模块,并通过驱动程序将MCS-DMA模块与目标DMA传输绑定.一方面实现数据预取,提升了单个DMA传输时的总线带宽利用率;另一方面使访存请求在内存控制器内部流水化完成,提升多个DMA并发时的总线带宽利用率.将该设计应用到北大众志SKSoC后,单个DMA传输时的总线带宽利用率提升至100%,多个DMA并发时的总线带宽利用率从33.3%提升至85.5%,而芯片的设计面积仅增加2.9%.
Current mainstream on-chip bus protocol — AHB has a problem that the bandwidth utilization of memory accesses is quite low.This paper proposes a new optimization design based on the feature that there are massive DMA transfers in SoC.Proposed method adds MCS-DMA modules inside the memory controller and bind MCS-DMA modules to target DMA transfers via software drivers.On the one hand,it prefetches data to increase the bandwidth utilization of single DMA transfer;on the other hand,it makes memory requests pipelined inside the memory controller,which increases the bandwidth utilization of multiple parallel DMA transfers.After applying the design to PKUnity-SK SoC,the bus bandwidth utilization when transferring single DMA increases to 100%.When transferring multiple DMAs in parallel,the bus bandwidth utilization increases from 33.3% to 85.5%.However,the chip area only increases by 2.9%.
出处
《电子学报》
EI
CAS
CSCD
北大核心
2010年第3期598-604,共7页
Acta Electronica Sinica
基金
国客863高技术研究发展计划(No.2006AA010202)
关键词
系统芯片
内存控制器
直接内存访问
System-on-Chip
memory controller
direct memory access（DMA）
