Stable, Fast and Accurate: Kernelized Attention with Relative Positional Encoding Shengjie Luo 1, Di He