Successfully deploying the ASIP

Abstract

Apart from the design flow for building the architecture of a programmable platform, the software development design flow impacts the overall time-tomarket considerably. A programmable platform becomes useless if it cannot be programmed effectively. That means, on the one hand, software development tools must offer to the programmer a reasonable abstraction of the underlying hardware. On the other hand, this abstraction must still be capable of exporting critical features of the architecture so that reasonable performance can be achieved. Finally, manually reimplementing these tools for every single design alternative during design space exploration is not feasible. A tool flow has to support the automatic generation of software development tools. This chapter describes how programming models are defined in the Mescal project to enable the successful deployment of ASIPs. In addition, the automatic generation of fast cycle-accurate, bit-true simulators eases the deployment of the programmable platform since software can be developed before actual silicon is available. We would like to note that a comprehensive discussion of all issues involved with the deployment of a programmable platform is beyond the scope of this book. In this chapter, we briefly discuss all currently available functionality in Tipi that eases the deployment of ASIPs. We recognize areas of current and future research in this field. This includes the question of how to compile from a more general application description, rather than a sequential language, down to a programmable platform. For one particular application domain - network processing - we introduce a specific programming abstraction. This abstraction, a so-called programming model, is natural for the application domain and allows the programmer to exploit the resources of the underlying network processor considerably faster than by using completely hand-written code. We also show how the cumbersome task of finding a feasible mapping of threads of execution onto processing resources can be automated. We are currently generalizing these experiences to less restricted application and architecture specifications so that compilation from heterogeneous models of computation to Tipi multiprocessor systems will be possible. © 2005 Springer Science+Business Media, Inc.