A Directive-Based Data Layout Abstraction for Performance Portability of OpenACC Applications

摘要

Directive-based programming interfaces such as OpenACC and OpenMP are becoming more prevalent in application development targeting accelerators, in particular when porting existing CPU-only code. Unlike vendor-specific alternatives such as CUDA, they are designed to be portable across different accelerators, and therefore once necessary directives are added to an existing CPU-only code, it can be executed on different accelerator architectures depending on the availability of supporting compilers. However, it does not automatically mean that such code runs efficiently on different architectures, and in fact, architecture-specific coding such as choosing optimal data layouts is almost mandatory for optimal performance, imposing a significant burden if implemented manually. Towards realizing performance portability in accelerator programming, we propose a set of extended directives that allow the programmer to optimize data layouts for a given accelerator without modifying original program code. Unlike the manual approach, the code change is confined in the directives with the original code kept as it is. This paper evaluates the effectiveness of our proposed extensions in the OpenACC standard by extending UPACS and CCS-QCD OpenACC applications. A prototype source-to-source translator for the extensions achieves 123% and 120% of the baseline performance, respectively, which are comparable to manually tuned versions.