Efficient scheduling of behavioural descriptions in high-level synthesis

A new heuristic scheduling algorithm for time constrained datapath synthesis is described. The algorithm is based on the distribution graph concept where a least mean square error function is used to schedule operations in sequence, resulting in a computationally efficient solution with the capability of including other high-level synthesis features such as register cost without significant increase in execution time. This new proposed method contrasts with previously published algorithms where the influence of all operations on the schedule is first evaluated before the most appropriate operation is selected and scheduled. An important feature of the presented algorithm is its ability to solve different scheduling problems, including conditional statements, multicycled functional units and structural pipelining. To illustrate the efficiency of the algorithm a set of benchmark examples has been synthesised and compared. It has been shown that the new algorithm produces high quality solutions when compared to other heuristic algorithms. Furthermore, it is simple to implement and computationally efficient, with execution times increasing approximately linearly with increasing time constraints allowing complex designs to be synthesised in an acceptable timescale. As an example, it takes < 30 s to obtain an optimal schedule for the discrete cosine transform when the time constraint of a maximum 36 control steps is imposed.