A synthesis system for application-specific arrays implementing matrix computations.

摘要

We propose a theoretical and methodological approach for synthesis of application specific arrays. The system, called ALIAS (ALgorithms Into Arrays Synthesizer), is intended to be an interactive design environment facilitating mapping between matrix algorithm specification and hardware implementations.;The input to the system is the single-assignment form of the algorithms and the output from the system is a synthesized array or a mapping of the algorithm on a user-specified array. A priori knowledge is not required for the user since the symbolic transformations are automatically done by the tool. ALIAS provides automated transformations of input matrix algorithms and interactive mapping to processor arrays.;A key feature of the proposed approach is MGD (Mesh Graph Descriptor) notation, used to formally characterize the design transformations in both the behavioral and structural domains. This notation is designed to be transparent to the user and, more importantly, to allow denotational transformations to take place.;ALIAS consists of three main modules: MAP (Matrix Algorithm Parser), DIP (Display Interface Program), and DM (Descriptor Manipulator). MAP takes inputs which are the single-assignment equivalents of matrix algorithms and produces MGD specifications for the dependency graph corresponding to that algorithm. DIP displays the MMG (Multi-Mesh Graph) structure to the user in interactive manner. DM module then applies transformers selected by the user to the MGD descriptors in deriving the array implementations of the algorithm. Given the design parameters and constraints, DM can be selectively used to obtain optimal designs. The user can choose to display various screens showing the MMG, node scheduling and assignment, or the array design along with the number of execution steps.