Pipelining and Vector Processing

Abstract

Objectives • To introduce the principles of pipelining; • To illustrate how resource conflicts are handled in pipelined systems; • To give instruction pipeline details of some example processors; • To describe vector processing basics; • To present details on the Cray X-MP vector machine; • To discuss performance of pipelined and vector processors. We briefly introduced the concept of pipelining in Chapter 1. Pipelining essentially allows overlapped execution of multiple operations. We begin this chapter with an introduction to the basic concepts of pipelining. Since pipelining is a highly coordinated activity, resource conflicts in the pipeline can seriously degrade its performance by causing pipeline stalls. We give some example scenarios that can potentially cause such stalls. Several types of hazards cause pipeline stalls. We discuss three types: resource hazards, data hazards, and control hazards. When two or more instructions in the pipeline require the same resource, a resource hazard occurs. In Section 8.2, we describe some solutions for resource hazards. Data dependencies among the instructions lead to the data hazards. Section 8.3 gives some examples and possible solutions to minimize the impact of data dependencies on the pipeline performance. Pipelining performs best when the execution is sequential. Pipeline performance deteriorates if the control flow is altered. Thus, it becomes important to devise ways to handle branches. Section 8.4 discusses several ways to handle such control hazards. The next section presents 273