Avoid GPU data race also with OpenCL