A parallel clustering algorithm for placement

In order to improve the layout quality of a VLSI design, many placement tools employ clustering algorithms to prune the optimization space and produce a design that can be enhanced while considering multiple design constraints. An intelligent clustering algorithm can guide a placement tool to reduce...

Full description

Saved in:
Bibliographic Details
Main Authors: Momeni, Amir, Mistry, Perhaad, Kaeli, David
Format: Conference Proceeding
Language:English
Subjects:
Algorithm design and analysis
Benchmark testing
Clustering
Clustering algorithms
GPU
Graphics processing units
Kernel
Merging
OpenCL
Parallel
Placement
Very large scale integration
Online Access:Request full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:In order to improve the layout quality of a VLSI design, many placement tools employ clustering algorithms to prune the optimization space and produce a design that can be enhanced while considering multiple design constraints. An intelligent clustering algorithm can guide a placement tool to reduce wire length, reduce cycle time, consider additional metrics or optimize a design based on a combination of these objectives. Given the myriad of choices, clustering algorithms can be time consuming to run, and this can impact our ability to fully explore the clustering space. Heterogeneous systems have been growing in popularity due to their attractive processing capabilities. Heterogeneous computing systems can be leveraged to improve the performance of clustering algorithms. In this paper, we present an OpenCL-based parallel clustering algorithm used during placement that targets heterogeneous systems. Our algorithm splits the computation, and exploits both CPU and GPU concurrently to balance the memory usage and the computational load in the heterogeneous system. We compare our parallel algorithm CL-Choice to a number of previously proposed algorithms. We evaluate these implementations in terms of placement speed and overall design quality. Simulation results show that our parallel implementation run on a Graphics Processing Unit (GPU) can achieve a 27× improvement over a number of serial algorithms in terms of speed, while not significantly impacting design quality.
ISSN:1948-3287
1948-3295
DOI:10.1109/ISQED.2014.6783347