Software code generation for dynamic dataflow programs

Abstract

In this paper we address the problem of generating efficient software implementations for a large class of dataflow programs that is characterized by highly data-dependent behavior and which is therefore in general not amenable to compile-time scheduling. Previous work on implementing dataflow programs has emphasized classes of stream processing algorithms that exhibit sufficiently regular behavior to permit extensive compile-time analysis and scheduling, however many real-world stream programs, do not fall into these classes and exhibit behavior that can, for example, depend on the values and even the timing of their input data. Based on an abstract machine model, we partition the problem of implementing such programs in software into three parts, viz. reduction, composition, and code emission, and present solutions for each of them. Using the reference code of an MPEG decoder, we evaluate the resulting code quality and compare it to the state of the art compilers for the same class of stream programs, with favorable results.