Systematic consolidation of input and output buffers in synchronous dataflow specifications

Abstract

Synchronous Dataflow, a subset of dataflow, is a commonly used model of computation in block diagram DSP programming environments. Because of the limited amount of memory in embedded DSPs, a key problem during software synthesis from SDF specifications is the minimization of the memory used by the target code. We develop a powerful formal technique called buffer merging that attempts to overlay buffers in the SDF graph systematically in order to drastically reduce data buffering requirements. This technique is complementary to lifetime-analysis based approaches, and we show that it can be fruitfully combined to yield a hybrid algorithm that results in less memory usage than either technique used alone. We give polynomial-time algorithms based on this formalism, and show that code synthesized using this technique results in a 45% reduction, on average, of the buffering memory consumption compared to existing techniques.