Scheduling algorithms for PIPE (pipelined image-processing engine)

Abstract

In this paper we present heuristic scheduling algorithms for the National Bureau of Standards/Aspen Inc. Pipelined Image-Processing Engine (PIPE). PIPE is a special-purpose machine for low-level image processing consisting of a linearly connected array of processing stages. A program is specified as a directed acyclic graph (DAG). Our first algorithm schedules planar DAGs. It works top-down through the graph and uses the greedy approach to schedule operations on a stage. It uses several heuristics to control the movement of images between stages. The worst case time for the schedule generated by the algorithm is O( N) times the optimal schedule, where N is the maximum width of the graph. We generalize this algorithm to work on nonplanar graphs, using heuristics for repositioning images on the stages of PIPE. The worst case time for the more general algorithm is also O( N) times the optimal schedule. Finally, we analyze the problem of optimizing throughput and latency for a sequence of DAGs on PIPE.