Rapid Development of Efficient Codes for PDE Simulation and PDE-Constrained Optimization

http://venus.usc.edu/seminars/CAM/Long.html High-performance algorithms for PDE-constrained optimization have been adopted slowly in practice because of the need to make intrusive changes to existing simulation codes. To overcome this barrier to entry, we have developed Sundance, a set of symbolic software tools enabling rapid development of efficient, differentiable, parallel simulation codes. Despite the high level of abstraction and generality, and the convenience of the user interface, simulators developed with Sundance prove to outperform special-purpose hand-tuned codes (even for forward problems, before we see the additional gains enabled through superior optimization algorithms). In this talk we will describe the design of Sundance and its software interface for interoperability with other codes, and discuss how, far from seeing an "abstraction penalty", we exploit abstraction and generality to achieve very high performance.

Location: 203 KAP

Audiences: Everyone Is Invited

Contact: Roger Ghanem