一种面向众核架构的数据流编译框架

A Compilation Framework of Dataflow Programs for Many-Core Architecture

数据流编程模型将程序设计与媒体处理相结合,已大量应用到各个领域.众核处理器已经成为主流和工业标准,如何利用众核架构的特性来提高流应用执行性能已成为目前研究工作的一大难点.文中提出了一个高效的流编译框架来优化流应用的执行,该框架包含3个优化策略:设计一个最优的软件流水调度方法;提出一个高效的数据存储分配算法;并采用合理的众核间的映射策略,减小通信以及同步的开销.文中在Godson-T上实现了该编译器框架,实验结果表明,该方法比优化前有较大性能改进.

Domain specific programming like Dataflow Programming Model which combines the features of media applications and programming languages has applied to many fields. Many-core architecture has become the mainstream solution and industry standard, how to use the character- istic of many-core architecture to improve the performance of stream applications has become a difficulty in present research work. In order to solve these problems, we propose an efficient stream compilation framework for many-core architecture to optimize the execution of stream applications, which is composed of three optimization strategy. In the first phrase, rate-optimal software pipelining schedule is constructed to improve parallelism. Then, a buffer allocation algorithm is proposed to allocate the data for pipelining schedule and redundant buffer copy operation is eliminated. The last phase maps the logical cores to the physical cores to reduce the communi- cation overhead. We also implement the framework on Godson-T and the experiments show that our method obtains about an average 58% improvement.