Compression-relaxation: a new approach to performance driven placement for regular architectures

We present a new iterative algorithm for performance driven placement applicable to regular architectures such as FPGAs. Our algorithm has two phases in each iteration: a compression phase and a relaxation phase. We employ a novel compression strategy based on the longest path tree of a cone for improving the timing performance of a given placement. Compression might cause a feasible placement to become infeasible. The concept of a slack neighborhood graph is introduced and is used in the relaxation phase to transform an infeasible placement to a feasible one using a mincost flow formulation. Our analytical results regarding the bounds on delay increase during relaxation are validated by the rapid convergence of our algorithm on benchmark circuits. We obtain placements that have 13% less critical path delay (on the average) than those generated by the Xilinx automatic place and route tool (apr) on technology mapped MCNC benchmark circuits with significantly less CPU time than apr.