Recursive least squares problem implementation on a generalized interconnection of DSP processors

The problem of implementing a recursive least squares (RLS) problem on a multiprocessor network comprising general-purpose digital signal processor is addressed. As an example of an application scenario, the adaptive beamforming problem has been chosen, where an RLS formulation has been adopted for obtaining the optimum weights. Although optimal array processors have been proposed for this problem, implementation on a generalized network affords a flexible as well as a modular and reconfigurable solution, suitable for a highly dynamic environment with changing applications and problem sizes. The parallelization issues are explored and a scheme for an efficient implementation is described. An implementation of the minimum-variance distortionless response (MVDR) problem on a network of DSP32C processors is evaluated. A 1024 size problem implementation is evaluated with the DSP32C as the computation engine, and it is shown that, when the number of processors in the system is constrained to be 10, near-optimal performance can be attained with four processors. Thus, this approach is a useful aid in determining optimal system performance parameters.<>