Implementation Of High Level Synthesis Design Spcae Exploration Method For Computationally Intensive Applications

Different directive settings of High Level Synthesis can lead to designs with widely different performance of design circuit without changing the source code.However,it requires a huge cost on time for only a single High Level Synthesis.The desdign space exploration of High Level Synthesis can find appropriate directives to meet the design requirements.In addition,computationally intensive applications consume amount of computational resources and memory operations.As a result,an appropriate design can greatly improve the performance of the system.According to the above problems,the High Level Synthesis Design Spcae Exploration methods are studied and a system framework of High Level Synthesis Design Spcae Exploration method for computing intensive applications is designed.Aiming at the theoretical model of High Level Synthesis,this paper take High Level Synthesis tool as a black box function.The code of the directives is used as the input of the black box function and the result of the High Level Synthesis tool is taken as the output.Bayesian optimization method is used for Design Spcae Exploration experiments.Aiming at the discreteness of the High Level Synthesis design space,the choice of the new exploration point is improved,which select the point with the largest collection function from the unexplored design space as the new exploration point.The simulation experiment is carried out in the end.According to the characteristics of massive calculation and frequent access operation to the memory,the High Level Synthesis design code general matrix multiplication is chosen as the benchmark code for this experiment.After introducing and analyzing the related High Level Synthesis directives,the design space and the exploration goals are designed.Based on Python and its related library functions,the experimental data collection script is implemented by analyzing the design process of the High Level Synthesis and its relative Tcl scripts.Based on the experimental data obtaining from the collection script,the Design Spcae Exploration experiment is executed.At first,the framework of Design Spcae Exploration method is demonstrated and implemented.Based on this,it is found that the one-of-k code has the same calculation result for the distance between different instructions,and the normalized code is used to improve it;the influence of the number of initial points and the convergence ability of different models on the Design Spcae Exploration is studied.Experimental results show that the Bayesian Optimization Design Spcae Exploration based on the improved normalized coding has good convergence ability.