OCLS: A Simplified High-Level Abstraction Based Framework for Heterogeneous Systems

In contrast with the increasing popularity of heterogeneous systems, programming on these systems remains complex and time-consuming. Developers have to access heterogeneous processors through explicitly and error-prone operations provided by low-level approaches like OpenCL. We present OCLS (OpenCL Simplified), a high-level abstraction based framework and its implementation as a minimal library on the top of OpenCL. OCLS shields hardware details, simplifies the development process and handles the environment configuration and data movement implicitly. Its APIs act like ordinary functions and require little prior training. OCLS thus reduces heterogeneous programming effort and relieves the programmers of low-level programming. We evaluated OCLS across a set of different benchmarks. The size of benchmarks rewritten in OCLS reduced by an average ratio of 35.4 %. In the experiment on both GPU and Intel MIC platforms with data sets in different size, OCLS yielded better performance than original OpenCL programs and showed a good stability and portability.