Static Profile-Driven Compilation for FPGAs.
FPL '01: Proceedings of the 11th International Conference on Field-Programmable Logic and Applications(2001)
摘要
We describe a static profiling methodology to extract hotspots from netlists. Hot-spots are small regular sub-circuits the optimization of which has a big impact on the final result. We have built a tool that can extract and characterize hot-spots from large netlists very quickly. The tool can be used to direct human attention on portions of circuits that need hand-optimization, as well as to automatically direct efforts of FPGA tools. We show impressive throughput improvements when compiling to the PipeRench reconfigurable architecture and use hot-spots to enable fast architectural design space exploration for FPGAs by predicting the FPGA CLB structure that produces the best final area-delay. Our prediction is fairly accurate and only takes a few hours as compared to weeks for an exhaustive analysis. We also demonstrate better results when targeting FPGAs with heterogeneous CLBs.
更多查看译文
关键词
fpgas,compilation,profile-driven
AI 理解论文
溯源树
样例
生成溯源树,研究论文发展脉络