Optimizing OpenCL Local Work Group Size With Machine Learning