摘要
CUDA是NVIDIA公司推出的GPU编程模型,它为高效利用GPU计算能力提供了强大的支持。但CUDA线程无法直接访问I/O设备、网卡等外围设备,在CUDA线程与外围设备的交互功能方面,目前CUDA的支持十分有限,仅支持非实时的屏幕打印(printf)。因此提出了一种交互型库函数框架FILiC,它通过设备和主机之间的巧妙交互,高效实现了CUDA线程实时的较完整I/O等函数;并且该框架具有很好的可扩展性,CUDA程序员或者编译器开发者可基于该框架按需求开发新的CUDA线程交互功能。
NVIDIA developed the CUDA programming model which provides a way to accelerate more general applications by GPU. But CUDA threads cannot access peripherals directly. As far as library functions interacting with peripherals, only ' printf' is allowed in CUDA threads by now. We described a framework named FILiC for interactive library, which implements I/O functions in CUDA threads efficiently by interactions between the device and the host. And FIL- iC is a framework with good scalability - CUDA progammers and compiler developers can use it to design some new library functions which interact with peripherals for CUDA threads.
出处
《计算机科学》
CSCD
北大核心
2012年第3期124-127,138,共5页
Computer Science