high performance computing on graphics processing units: hgpu.org

The open computing language (OpenCL) is a standard open source specification for parallel computing on heterogeneous architectures. OpenCL offers a set of abstract models for substantial acceleration in parallel computing and is supported by most of the leading hardware vendors. In this paper, we present a systematic approach for employing OpenCL as a hardware abstraction layer that enables the user to utilize the supported computing resources by using different scheduling and mapping schemes. We illustrate the approach by a case study of a distributed automotive embedded system. In particular, we developed an extendable set of related advanced driving assistance system (ADAS) applications under a common application setup which is mapped and executed on different OpenCL supported devices of an embedded platform. The detailed evaluations performed using different scheduling schemes in conjunction with various OpenCL mapping configurations are presented.