Exploring the number of register windows in ASIP synthesis

ASIPs (Application Specific Instruction Set Processors) are one of the key components of many embedded systems which are typically application specific. An ASIP can be defined by a set of architectural features, number of register windows being one of them. The work described here focuses on generating the transfer time penalties for some of the mediabench benchmark applications namely JPEG and MPEG encoder and decoder, for different number of register windows. The problem has been solved in two steps. First the "spills" for different number of windows were counted where a spill refers to the situation where a context switch cannot be accommodated in the register windows thereby adding the overhead of transferring some data to the memory. This part of the problem was solved by mapping it to the "regular language recognition problem". In the next step, actual time penalties for different system configurations, were computed. Here, a system configuration consists of the memory configuration, bus width. and speed and processor cycle time. Thus, this work may also drive the design space exploration process. Results and expected performance gains by selecting different number of register windows is presented