Fine granularity clustering for large scale placement problems

Abstract

In this paper we present a linear-time Fine Granularity Clustering (FGC) algorithm to reduce the size of large scale placement problems. FGC absorbs as many nets as possible into Fine Clusters. The absorbed nets are expected to be short in any good placement; therefore the clustering process does not affect the quality of results. We compare FGC with a connectivity-based clustering algorithm proposed in [1] and simulated-annealing-based algorithm in TimberWolf [2], both of which also reduce the number of external nets between clusters. The experimental results show that our algorithm achieves better net absorption than the previous approaches while using much less CPU time for large scale problems. With our FGC algorithm, we propose a Fast Placer Implementation (FPI) framework, which combines our FGC-based size reduction with traditional placement techniques to handle largescale placement problems. We compared FPI placement results with a public-domain fast standard cell placer Capo[4] on large scale benchmarks. The results show that FPI can reduce CPU time for large scale placement by a factor of 3~5x while obtaining placement results of comparable or better quality.

FastPlace: efficient analytical placement using cell shifting, iterative local refinement,and a hybrid net model

“…The runtime of FastPlace can be further reduced by: (a) Employing a hierarchical framework (e.g. [19]) to reduce the problem size. The reduced problem can then be solved by FastPlace.…”

“…Note that, when the placement problem is so large that a flat analytical approach cannot handle it effectively, a hierarchical analytical approach is beneficial. One way to convert to a hierarchical approach is to incorporate the fine granularity clustering technique proposed by Hu et al [19]. This technique essentially introduces a two-level hierarchy to reduce the size of large-scale placement problems.…”

FastPlace 3.0: A Fast Multilevel Quadratic Placement Algorithm with Placement Congestion Control

“…Since this clustering is performed before any placement, we restrict it to finegrain clustering to minimize any loss in placement quality due to incorrect clustering. In fact, it was demonstrated in [12] that building fine-grain clusters can improve placement efficiency with negligible loss in placement quality.…”

Clusters first? a study of circuit structure and placement

“…The initial success of the single-level clustering method of Cong and Smith [6] has parallels to a recent placement effort [12]. In our preliminary experiments, a simple top-down combinatorial approach was able to find optimal placements for meshlike graphs, when provided with an initial optimal clustering tree [10].…”

“…Recently, Hu and Marek-Sadowska [12] presented a placement approach that first applies a limited amount of clustering, reducing the size of the placement problem; the authors then used Capo to perform placement. The primary impact of the approach was improved run times; there was also a modest improvement in placement wire length.…”