Towards making the most of NLP-based device mapping optimization for OpenCL kernels