I'm intersted in using accelerators for getting better performance for molecular simulations software. What approaches are being used to achieve portability for applications across multiple hardware environments for the core kernels?