Program compression in long instruction word application-specific instruction-set processors

Abstract

Modern day embedded systems set high requirements for the processing hardware to minimize the area, and more importantly, the power consumption. Moreover, the ever increasing complexity of embedded applications requires more and more processing power. Application-specific architectures, where the hardware resources can be tailored for a given application, have been introduced to meet these requirements.

Parallel processor architectures have also become favorable as they provide more processing power by utilizing the instruction level parallelism. However, parallel processor architectures result in large program codes, which require large memories that increase not only the area, but also the power consumption of the system due to increased memory I/O bandwidth. Program compression methods have been proposed to tackle this problem and reduce the size of the program code and, therefore, also the area and power consumption.

The focus of this Thesis is on program compression in parallel processor architectures. State-of-the-art program compression methods are surveyed and compared against the derived comparison metrics. Based on the survey, three compression methods are chosen to be evaluated on transport triggered architecture, a parallel processor architecture template used to design application-specific instruction-set processors. The methods are adapted to exploit the characteristics of the architecture. In addition to code density evaluations in terms of compression ratio, an evaluation methodology based on hardware implementations is proposed. It allows to evaluate the effects of compression on the actual area and power consumption of the system.

Program compression may also result in poor instruction-set orthogonality, which makes the programming after compression more difficult and worsens the performance. The orthogonality may turn out to be so poor that the program code cannot be modified anymore. A novel methodology with a small overhead in area and power consumption is proposed to allow to modify the program code also after compression.