Estimation of lower bounds in scheduling algorithms for high-level synthesis

Abstract

To produce efficient design, a high-level synthesis system should be able to analyze a variety of cost-performance tradeoffs. The system can use lower-bound performance estimated methods to identify and puune inferior designs without producint complete designs. We present a lower-bound performance estimate method that is not only faster than existing methods, but also produces better lower bounds. In most cases, the lower bound produced by our algorithm is tight. Scheduling algorithms such as branch-and-bound need fast and effective lower-bound estimate methods, often for a large number of partially scheduled dataflow graphs, to reduce the search space. We extend our method to efficiently estimate completion time of partial schedules. This problem is not addressed by existing methods in the literature. Our lower-bound estimate is shown to by very effective in reducing the size of the search space when used in a branch-and-bound scheduling algorithm. Our methods can handle multicycle operations, pipelined functional units, and chaining of operations. We also present an extension to handle conditional branches. A salient feature of the extended method is its applicability to speculative execution as well as C-select implementation of conditional branches.