On Enhancing SIMD-controlled DSPs for Performing Recursive Filtering

Abstract

Many digital signal processors (DSPs) and also microprocessors are employing the single-instruction multiple-data (SIMD) paradigm for controlling their data paths. Although the SIMD paradigm can provide high computational power and efficiency, not all applications can profit from this feature.One important application, particularly in audio processing, of DSPs are recursive (IIR) filters. Due to their datadependencies they can not exploit the capabilities of SIMD-controlled DSPs as non-recursive (FIR) filters can. This paper shows, how the SIMD control scheme can be enhanced to accommodate recursive filters without introducing much hardware overhead.Three methods for calculating recursive filters on SIMD-controlled DSPs and their requirements for control and data transfer are presented. They can be applied to direct-form as well as cascade-form realizations. The performance and hardware requirements of these methods are evaluated to determine the most efficient solution in terms of the AT-product.

Short-Vector SIMD Parallelization in Signal Processing

“…They have therefore limited power if there are too many data dependencies, as there are in IIR filtering. Examples can be found in [20,21]. We will use a rather straight forward approach that will be improved by algebraic transforms in the next section.…”

Parallelization of IIR filters using SIMD extensions

“…The parallelization of IIR filters is more difficult due to data dependencies. There are several approaches including space-time transformations of loop iterations [9], [10] and algebraic transformations [11]. The approach presented in this paper also uses algebraic transformations, although different ones.…”