Towards Faster and Better Analytical Placements

Abstract

In this thesis, we explore ways to improve the current state of the art in analytical VLSI standard cell placement techniques. We present a convex formulation of the placement problem which can be made arbitrarily close to a linear wire length model. An interior point method (IPM) for solving this convex problem is presented which utilizes a preconditioned version of the conjugate gradient squared (CGS) iterative method. Preconditioners are computed using a drop tolerance incomplete $LDL^T$ factorization. Both the IPM and the iterative method have been implemented in parallel in a shared memory environment. Moderate speedups are achieved on two and four processor symmetric multiprocessor workstations. The parallel implementation often requires fewer CGS and IPM iterations due to very slight numerical differences in the vector computations. The limiting factor is memory-to-CPU bandwidth and the inability to bind slave threads to a particular processor.

Fetch the entire document in postscript format.


 frickin' computers Mark Frazer -- mjfrazer@gmail.com  frickin' computers