Flow-based partitioning and position constraints in VLSI placement

This paper presents a new quadratic, partitioning-based placement algorithm which is able to handle non-convex and overlapping position constraints to subsets of cells, the movebounds. Our new flow-based partitioning (FBP) combines a global MinCostFlow model for computing directions with extremely fast and highly parallelizable local realization steps. Despite its global view, the size of the MinCostFlow instance is only linear in the number of partitioning regions and does not depend on the number of cells. We prove that our partitioning scheme finds a (fractional) solution for any given placement or decides in polynomial time that none exists. In practice, BonnPlace with FBP can place huge designs with almost 10 million cells and dozens of movebounds in 90 minutes of global placement. On instances with movebounds, the netlengths of our placements are more than 32% shorter than RQL´s and our tool is 9-20 times faster. Even without movebounds, the FBP improves the quality and runtime of BonnPlace significantly and our tool shows the currently best results on the latest placement benchmarks.