Abstract
Computational electromagnetics simulations use direct solvers to find the solution to linear systems of equations. The performance of state-of-the-art libraries, used to enhance the execution time of direct solvers, are tied to the underlying architecture and problem instance. Libraries often trade portability and generality for performance and do not use domain and hardware information for optimization. We propose to develop a domain-specific compiler called Axb, which uses application and architecture information, to automatically generate high-performance code for direct sparse solvers on modern parallel architecture. Our preliminary results show that Axb can improve the performance of direct solvers up to 10X compared to state-of-art-libraries for tested problems.