LEAP: A Data Driven Loop Engine on Array Processor

Yong Dou, Xicheng Lu

Abstract: This paper presents a novel architecture for array processor, called LEAP, which is a set of simple processing elements. The targeted programs are innermost loops. By using the technique called if-conversion, the control dependence can be converted to data dependence to prediction variables. Then an innermost loop can be represented by a data dependence graph, where the vertex supports the expression statements of high level programming languages. By mapping the data dependence graph to fixed PEs, each PE steps the loop iteration automatically and independently at the runtime. The execution forms multiple pipelining chains. One example program FIR and corresponding transformation are given to show the execution process of the LEAP array. The performance analysis shows the effectiveness of the LEAP architecture, compared with CISC and RISC architectures.