A Comparison of GPGPU Computing Frameworks on Embedded Systems

Abstract Graphics processing units have found their way onto the die of embedded CPUs. Embedded devices have thus gained access to on-die parallel co-processors that can be put to good use with the help of the low-level OpenCL standard API, but also using programming frameworks that help making use of this additional compute power. This paper compares several programming frameworks with different coding styles. We report on the coding effort needed and the performance achieved on a Congatec conga-QG embedded computer on a module using two representative codes taken from the SHOC benchmark suite as comparison.