A pipelined vector processor and memory architecture for cyclostationary processing

Abstract

This work describes a scaleable, high performance, pipelined, vector processor architecture. Special emphasis is placed on performing fast Fourier transforms with mixed-radix butterfly operations. The initial motivation for the architecture was the computation of cyclostationary algorithms. However, the resulting architecture is capable of general purpose vector processing as well. A major factor affecting the performance of the architecture is the memory system design. The use of pipelining techniques, coupled with vector processing, places a substantial burden on the memory system performance. The memory design is based on an interleaved memory philosophy with a buffering technique referred to as split transaction memory (STM). A crucial aspect of the memory design is the memory decoding scheme. A design methodology is described for the specification of permutation matrices that yield near optimal performance for the memory system. Another important aspect of this work is the development of a software based simulator that allows a STM to be specified. The simulator, operating at the register transfer level, emulates the processing of an address stream by STM and records the events for post-processing. The STM simulator was used to evaluate three types of vector processing address patterns: constant stride, constant geometry radix-r butterfly, and digit reversed. A random address pattern was also analyzed in the context of general-purpose computing. STM simulation verified the near-optimal performance of the STM.