A Mechanized Semantics for Dataflow Circuits

This paper proposes a mechanized formal semantics for dataflow circuits: rather than following a predetermined, static schedule, the execution of the circuit components is constrained solely by the availability of their input data. We model circuit components as abstract computing units, asynchronously connected with each other through unidirectional, unbounded FIFO. In contrast to Kahn's classic, denotational semantic framework, our semantics is operational. It intends to reflect Dennis' dataflow paradigm with firing, while still formalizing the observable behaviors of circuits as channels histories.

The components we handle are either stateless or stateful, and may be non-deterministic. We formalize sufficient conditions to achieve the determinacy of circuits executions: all possible schedules of such circuits lead to a unique observable behavior. We provide two equivalent views for circuits. The first one is a direct and natural representation as graphs of components. The second is a core, structured term calculus, which enables constructing and reasoning about circuits in a inductive way. We prove that both representations are semantically equivalent.

We conduct our formalization within the Coq proof assistant. We experimentally validate its relevance by applying our general semantic framework to dataflow circuits generated with Dynamatic, a recent HLS tool exploiting dataflow circuits to generate dynamically scheduled, elastic circuits.