Making Scientific Computing Libraries Forward Compatible

<p>NSF's Software Infrastructure for Sustained Innovation funds the development of community software in support of scientific computing innovation. A requirement is that the developed software be sustainable. Design-by-Transformation (DxT) is an approach to software development that views libraries not as instantiated in code, but as expert knowledge that is combined with knowledge about a target architecture by a tool (DxTer) that synthesizes the library implementation. We argue that this approach makes libraries to some degree forward compatible in that a (disruptive) new architectural advance can be accommodated by encoding knowledge about that architecture. This is particularly important when bugs are not correctness bugs, but instead performance bugs that affect how fast an answer is obtained and/or how much energy is consumed to compute the answer. DxT allows a human expert to focus on developing the primitives from which libraries are constructed and new insights as opposed to the rote application of known ideas to entire libraries. We summarize our success in the domain of dense linear algebra as evidence of DxT's potential.