From Static to Quasi-Dynamic: Reconsidering Scheduling and Memory in SDF Compilers

Synchronous Dataflow (SDF) is a popular Model of Computation (MoC) for signal processing applications, thanks to its static analizability and provably deadlock-free execution. SDF compilers have traditionally prioritized static schedulers for their low runtime latency, particularly for signal processing and embedded systems. However, the conventional static technique of converting of multi-rate graphs to single-rate equivalents introduces scalability challenges through exponential graph growth and lost coarse-grained symmetry information. We revisit the idea of runtime scheduling for static dataflow and explore a tradeoff between latency, binary size and compilation time.Our approach introduces a specialized “quasi-dynamic” runtime scheduler and memory manager that leverages static topology information to precompute buffer sizes and allows out-of-order execution of actor firings, while thread assignment and memory allocation are handled at runtime through the OpenMP runtime and standard allocation primitives (malloc). By preserving multi-rate actor relationships, we avoid the node explosion inherent to single-rate conversion and improve the feasibility of SDF for a previously unreachable class of workloads. Comparative analysis of the current prototype against the Preesm static scheduler on the degridder radio astronomy application show, on average, 12.2x compilation time speedup and 4.4x smaller binary sizes, while only suffering 1.4x slowdown.