“…Relaxed semantics models are possible, e.g., to support commutativity to enable vectorization [4], however all input code iterations are to be executed in the optimized code. On the other hand, more aggressive techniques to automatically compute approximations have been designed, e.g., by ignoring some dependencies to enable parallelization [5], by providing alternative implementations of some code parts [6], or by skipping computations [7], [8]. In our work, we investigate a new way, called Adaptive Code Refinement (ACR), inspired by Adaptive Mesh Refinement [9], a classical numerical analysis technique providing the ability to dynamically tune a computational grid to achieve precise computation only where it matters.…”