Research On OpenACC-based Automatic Parallelization Technology

In recent years, GPU is widely used as an accelerated computing device. Besides, the computing technology with General-Purpose GPU which abbreviated as GPGPU has become a necessary means of dealing with any large amounts of data. However, the difficulty of GPU computing has been increased by complex hardware environment and different CPU programming models. To enable more researchers to use GPU computing, the graphics card manufacturers and compiler companies develop and promote the OpenACC standards.On the basis of OpenACC, we proposed a new framework named GENerate OpenACC, or GENACC for short. GENACC can automatically accelerate the serial code. The framework analyze the static program of the source code, then recognize the hot snippet and analyze the computation property. And finally it add OpenACC directive to the source code. Compared to other automatic parallelization systems, the GENACC framework has three advantages on portability, cross-platform and capability.In addition, we constructed a new computation cost evaluation model to deal with the problems existing in GENACC framework, such as how to estimate the cost of computation time about analysis of hot snippet and how to judge hardware acceleration. The model analyze hot snippet and hierarchically model for hardware properties of computing platform. Therefore, it can construct models for computation time and the energy cost of computing. We can use the model as important method to evaluate computation cost and the rationality of computing device. We also give a new time evaluation model on the base of LogGP model and Roofline model. According to thickness size, we model by three levels to compute the energy cost.Finally, we experiment on computation cost evaluation model and GENACC framework through two test sets, EPCC and NPB. The experiment results show that GENACC can accurately produce compiler directive, and the source codes show fine performance on different amount of data.