Loop Transformation In Task Compiler Of Reconfigurable Computing Processor

In recent years, reconfigurable computing has abstract more and more researchers'interesting. An efficient task compiler will help the developers to explore the performance of a reconfigurable architecture when mapping a task to it. Usually, the loops take most of the execution time of a program. It is very important to map the loops to the reconfigurable architecture.This paper focuses on the mapping of loops to reconfigurable array, especially the loop optimization and loop transformation in task compiler. A loop-centric profiling method is used to collect the execution information of loops in a program, and this will help the optimization and transformation of loops. Loop unrolling is studied in detail including the condition of loop unrolling and the implementation of loop unrolling, also scalar replacement is also used to support that REmus don't have memory access unit in the reconfigurable array. Redundant elimination technology is also used in this paper based on data flow analysis to eliminate the redundant computing in loops and DFG.The compiling technology in this paper is integrated to RCAcompiler, a task compiler for REmus, and we demonstrate the effectiveness of the technology by experience on several typical media processing and digital signal processing algorithm. The results show that our methods efficiently improve the parallel of loops in the task compiler, and improve the performance of task compiler significantly.