Algorithmic skeletons (`skeletons') abstract commonly-used patterns of parallel computation, communication, and interaction. They provide top-down design composition and control inheritance throughout the whole structure. The efficient execution of skeletal applications on a heterogeneous environment has long been of interest to the research community. Arguably, executing a coarse-grained resource-intensive skeletal workloads ought to achieve higher resource utilisation and, ultimately, better job makespan on heterogeneous systems due to the structured parallelism model. This paper presents a heterogeneous OpenCL-based GPU back-end for FastFlow, a widely-used skeletal framework. Our back-end allows the user to easily write any arbitrary OpenCL code inside an heterogeneous algorithmic skeleton and seamlessly control the allocation of OpenCL kernel over the hybrid (CPU/GPU) architecture. Our performance evaluation indicate that a skeletal program which employs our back-end is around one order of magnitude faster than a skeletal parallel program using the traditional homogeneous FastFlow skeletons with the serial version of OpenCL code.