An automatic mapping technique for OpenACC kernel code based on deeply fused and heterogeneous many-core architecture

Abstract

Now the OpenACC has become a popular programming interface for many-core application programming. Internationally, a lot of research have been done on OpenACC for CPU + GPU heterogeneous many-core architecture. Among them, the PGI OpenACC compiler developed by NVIDIA is the most advanced one. But there are few research on OpenACC related to the Home Grown Heterogeneous Many-Core (HGHM) Architecture that is different from GPU. This paper proposes an automatic mapping technique for OpenACC kernel code based on the OpenACC compiler to a heterogeneous and deeply fused many-core architecture. Our approach uses the static analysis and feedback dynamic analysis of the compiler to perform the automatic mapping of the program parallel kernel code to many-core devices, and it greatly improves the transformation quality of the compiler. Experimental results show that this technique can greatly improve the efficiency of using OpenACC to port applications to heterogeneous and fused many-core system without impacting program acceleration performance.