High-level DSP synthesis using concurrent transformations, scheduling, and allocation

This paper addresses high-level synthesis methodologies for dedicated digital signal processing (DSP) architectures used in the iterative Loop-based Minnesota Architecture Synthesis (MARS) design system. We present a novel concurrent scheduling and resource allocation algorithm which exploits inter-iteration and intra-iteration precedence constraints. The novel algorithm implicitly performs algorithmic transformations, such as pipelining and retiming, on the data-flow graphs during the scheduling process to produce solutions which are as good as those previously published and which executes in less time. MARS is capable of producing optimal and near-optimal schedules in fractions of seconds. Previous synthesis systems have focused on DSP algorithms which have single or lumped delays in the recursive loops. In contrast, MARS is capable of generating valid architectures for algorithms which have randomly distributed delays. MARS exploits these delays to produce more efficient architectures and allows our system to be more general. We are able to synthesize architectures which meet the iteration bound of any algorithm by unfolding, retiming, and pipelining the original data-flow graph