Faster Goal-Oriented Shortest Path Search for Bulk and Incremental Detailed Routing

TL;DR

This paper introduces advanced algorithmic techniques for VLSI detailed routing, significantly improving shortest path computations in complex grid graphs and combining classical routing approaches for enhanced efficiency.

Contribution

It presents a novel goal-oriented Dijkstra's algorithm with better estimates for large incomplete graphs and a hybrid routing approach treating input wires as reservations for efficiency.

Findings

Achieved faster shortest path computations in large grid graphs.

Demonstrated improved trade-offs between runtime and routing quality.

Implemented an efficient hybrid routing method combining global search and local corrections.

Abstract

We develop new algorithmic techniques for VLSI detailed routing. First, we improve the goal-oriented version of Dijkstra's algorithm to find shortest paths in huge incomplete grid graphs with edge costs depending on the direction and the layer, and possibly on rectangular regions. We devise estimates of the distance to the targets that offer better trade-offs between running time and quality than previously known methods, leading to an overall speed-up. Second, we combine the advantages of the two classical detailed routing approaches - global shortest path search and track assignment with local corrections - by treating input wires (such as the output of track assignment) as reservations that can be used at a discount by the respective net. We show how to implement this new approach efficiently.