Fast automatic generation of efficient custom instructions for application-aware computing
Hu Chen, Shuming Chen · 2018
Adding Custom Instructions (CIs) statically or dynamically for specific applications to base processors has proved to be an effective way to improve the performance of processors and meet the time-to-market requirements. However, it is intractable to manually analyze the applications and select the suitable instructions for complex applications. In this paper, a fast automatic Instruction-Set Extension (ISE) method called FAGECI was proposed. FAGECI can fast explore the design space by enumerating candidate instructions around typical operations of applications. Further, it can also improve resource utilization through resource sharing between the instructions based on manipulating the data flow of instructions according to their characteristics. Experimental results demonstrate that FAGECI features exactly linear complexity and can generate high-performance CIs that introduce less area overhead.