Efficient OpenCL-based concurrent tasks offloading on accelerators