Demonstration of Square Root Kalman Filter on Warp Parallel Computer

Abstract

A parallel algorithm for solving an n-state, m-measurement square root Kalman filter on an (n+m+1)-cell linear array is described. The approach is to combine the time and measurement updates of the covariance matrix and use fast Givens rotations to triangularize an appropriate matrix. The rotations of the rows in this matrix are done in parallel to distribute the computations. The cell-to-cell communication is minimized by arranging the matrix in a near triangular form. We demonstrate the parallel algorithm on Warp for an extended-Kalman filter commonly used in target tracking applications. The Warp implementation is written in a high-level language and achieves two orders of magnitude speedup over the same filter running on a Sun workstation. Efficient algorithm mapping for parallel computations is the key to achieving the high speed filter performance. With the availability of iWarp chips in 1990, the results open new solutions to on-board target tracking and other Kalman filtering applications.