Systolic array synthesis by static analysis of program dependencies

We present a technique for mapping recurrence equations to systolic arrays. While this problem has been studied in fairly great detail, the recurrence equations that are analysed here are a generalization of those studied previously. In a n earlier paper (14] we have showed how systolic arrays can be synthesized from such generalized recurrence equations by a combination of affine transformations and explicit pipelining. This paper extends the results in two directions. Firstly, a multistage pipelining technique is proposed, which permits the synthesis of systolic arrays with irregular data flow. Secondly we develop analysis techniques for the synthesis of systolic arrays whose computation is governed by control signals in a systematic manner which is amenable to mechanization. The full paper also discusses how these techniques can be applied to the mapping problem for more general architectures.