high performance computing on graphics processing units: hgpu.org

In this paper, we introduce IPMACC a framework for executing OpenACC for C applications over OpenCL runtime. We use over framework to compare performance of OpenACC and OpenCL. OpenACC API abstractions remove the low-level control from programmers’ hand. To understand the low-level OpenCL optimizations that are not applicable in OpenACC, we compare highly-optimized OpenCL and OpenACC versions of a wide set of benchmarks. We show that under the investigated benchmarks, exploiting scratchpad memory as a fast-communication link is the most important optimization that is not applicable in OpenACC. We also introduce a micro-benchmarking suit to investigate the overhead of various OpenACC operations. We compare our framework to a previous open source OpenACC compiler in various aspects.