High Level Synthesis of Multi-precision Data Flow Graphs

A number of DSP algorithms involve linear transforms employing weighted sum computations, where the weights are fixed at design time. Add-shift implementation of such a computation results in a data flow graph that has multiple precision variables and functional units. We explore the potential of precision sensitive approach for the high level synthesis of such multi-precision DFGs. We focus on fixed latency implementation of these DFGs. We present register allocation, functional unit binding and scheduling algorithms to exploit the multi-precision nature of such DFGs for area efficient implementation. The proposed approach is fairly generic and could be applied to multi-precision DFGs involving any type of functional units. Significant improvements of upto $27\%$ have been obtained over the conventional high-level synthesis approach