Improving tasks throughput on accelerators using OpenCL command concurrency