A cycle-accurate compilation algorithm for custom pipelined datapaths

Traditional high level synthesis (HLS) techniques generate a datapath and controller for a given behavioral description. The growing wiring cost and delay of today technologies require aggressive optimizations, such as interconnect pipelining, that cannot be done after generating the datapath and without invalidating the schedule. On the other hand, the increasing manufacturing complexities demand approaches that favor design for manufacturability (DFM).To address these problems we propose an approach in which the datapath of the architecture is fully allocated before scheduling and binding. We compile a C program directly to the datapath and generate the controller. We can support the entire ANSI C syntax because the datapath can be as complex as the datapath of a processor. Since there is no instruction abstraction in this architecture we call it No-Instruction-Set-Computer (NISC). As the first step towards realization of a NISC-based design flow, we present an algorithm that maps an application on a given datapath by performing scheduling and binding simultaneously. With this algorithm, we achieved up to 70% speedup on a NISC with a datapath similar to that of MIPS, compared to a MIPS gcc compiler. It also efficiently handles different datapath features such as pipelining, forwarding and multi-cycle units.