A generic infrastructure for OpenCL performance analysis

OpenCL is an open standard for programming of parallel heterogeneous systems. It is designed for portability, therefore being utilized in the area of embedded system programming as well as high performance computing (HPC). Due to the applicability on different platforms, OpenCL library vendors have a certain freedom in implementing parts of the OpenCL execution model. Multiple versions of the standard increase the diversity of OpenCL implementations. Sophisticated performance analysis tools are required to optimize the performance of an OpenCL program for a specific OpenCL implementation and therewith efficiently utilize available hardware resources. This paper presents a generic tools interface for performance measurement of OpenCL programs. We depict the functionality and implementation of a respective measurement library and its integration in the performance infrastructure Score-P. Tests with a synthetic and several real-world OpenCL applications on AMD, ARM, NVIDIA, and Intel platforms validate the applicability of our framework.