Abstract
Abstract DATuner is an autotuning program, but its running time is very long and uneven. In our paper, a novel methodology for autotuning and exploring parameter space is proposed. Combined with Structural Equation Model and statistical techniques, the prior pruning was proposed. Firstly, using benchmarks in DATuner, VTR (Verilog to Routing) was run to generate training data set; Secondly, with training data set as input, modeling software —— AMOS was run to construct structural equation model, and then model was validated and evaluated. If the fitness of the model does not reach requirement, then the connections between variables are modified to get final model whose fitness is sufficient. Finally, the parameters’ correlation was analyzed which resulted in routing failure. Applying the prior pruning to DATuner, its autotuning time is reduced and running efficiency has been improved. Compared with unmodified DATuner, DATuner with prior pruning decreases by an average of 44% on the autotuning time. The proposed method greatly improves autotuning efficiency. In addition, DATuner’s framework was optimized.
Published Version
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call