Automatic Derivation And Optimization Of Data Information Directive In AceMesh

With the continuous development of computer architecture,heterogeneous multi-core has become the mainstream architecture of high-performance computer.With the emergence of heterogeneous multi-core platforms,the former mainstream parallel programming model on many-core platforms will no longer be applicable.Today,the research and development of parallel programming model for heterogeneous multi-core platform is the key to efficient use of high-performance computers.China's domestic the Sunway TaihuLight supercomputer occupies the top three in the world's TOP500 list all year round.The characteristic programming models present on this platform are mainly OpenACC*and Athread.These two programming models divide coarse-grained tasks with control flow drive,and the internal utilization of the computing core is not perfect.In order to make better use of the computing performance of domestic heterogeneous platforms,the parallel compilation group of the Academy of Sciences of the Chinese Academy of Sciences proposed the parallel programming model AceMesh for domestic heterogeneous platforms.The model is Data-driven and uses a more fine-grained partition method to build and execute task graphs,and to mine the memory cache reuse of LDM by various means.The AceMesh parallel programming model initially provided developers with interfaces in the form of low-level library APIs.To reduce the difficulty for developers and provide high-availability portability,the research team extended the parallel programming interface with directive-based programming languages.The AceMesh compiler will generate a parallel program for the specified platform for the written Compiler Directives application and platform features.Developers only need to pay attention to the directive-based programming languages and the application itself.Since AceMesh is in its infancy,early design flaws have caused AceMesh Compiler and the directive-based programming languages to be tightly coupled.Developers write Compiler Directives of AceMesh related to data information is more complicated,and at the same time can not meet the diversification of the program.This paper improves AceMesh directive for domestic heterogeneous platform by implementing the automatic analysis of the parallel region by AceMesh Compiler.Using Affine analysis to collect data information of parallel area to solve the problem of complicated copy clause and insufficient expression in acemesh guidance,and putting for-ward a new core transformation algorithm;expanding the scope of array partition and array declaration guidance,reducing the redundancy of guidance;using the formula based on full binary tree to analyze the data results of AceMesh Compiler automatic analysis,reducing the logarithm of task graph registration guidance According to the description of dependency.Finally,the applicability and generality of the simplified guidance are evaluated by the benchmark program of the AGCM(atmospheric general circulation model)and the benchmark program of the NPB(NAS parallel benchmarks).