Long pipelines in single-chip digital signal processors-concepts and case study

摘要

The effectiveness of long pipelines in single-chip digital signal processors for complex algorithms was studied using a processor model with 25 pipeline stages. The processor is based on a Harvard architecture. Pipelining is used to reduce the instruction cycle time compared to current signal processors. Key features of the processor model are data-stationary pipeline control, local resolution of pipeline hazards with buffering, multiple branch prediction, a mixed relative-incremental addressing scheme, and asynchronous communication between pipeline and environment. The processor is implemented as a software model. The results show that high pipeline utilization can be achieved for a variety of algorithms leading to a significantly higher performance than achieved by conventional single-chip signal processors with Harvard architecture.