A Compiler Transformation to Overlap Communication with Dependent Computation

Hiding communication latency is essential toachieve scalable performance on current and future parallelsystems. In this extended abstract, we present a novel compilertransformation that overlaps communication with computationto hide communication latency. Unlike prior work, we are ableto achieve this overlap even in the presence of an overlap-inhibitingdata dependence between the communication andcomputation. We do so by transforming the data dependenceinto an overlap-amenable one. To achieve this overlap, theMaunam compiler transforms the code by employing arrayexpansion, partial loop peeling, loop alignment, and arraycontraction. This transformation is useful for optimization ofsystolic, communication avoiding algorithms