Architecture and applications of the HiPAR video signal processor

Abstract

We propose the architecture of a highly parallel DSP (HiPAR-DSP) as a flexible and programmable processor for image and video processing. The design is based on an analysis of image processing algorithms in terms of available parallelization resources, demands on program control, and required data access mechanisms. This led to a very long instruction word (VLIW)-controlled ASIMD RISC-architecture with four or sixteen data paths, employing data-level parallelism, parallel instructions, micro-instruction pipelining, and data transfer concurrently to data processing. Common data access patterns for image processing algorithms are supported by use of a shared on-chip memory with parallel matrix type access patterns and a separate data-cache per data path. By properly balancing processing and controlling capabilities as internal and external memory bandwidth, this approach is optimized to make the best use of currently available silicon resources. A high clock frequency is achieved by implementation of classic RISC features. The architecture fully supports high level language programming. With the 16 data path version and a 100 MHz clock, a sustained performance of more than 2 billion arithmetic operations per second (GOPS) is achieved for a wide range of algorithms. The examples show the parallel implementation of image processing algorithms like histogramming, Hough transform, or search in a sorted list with efficient use of the processor resources. A prototype of the architecture with four parallel data paths is available, using a 0.6 /spl mu/m CMOS technology.