Index handling and assign optimization for Algorithmic Differentiation reuse index managers
For operator overloading Algorithmic Differentiation tools, the identification of primal variables and adjoint variables is usually done via indices. Two common schemes exist for their management and distribution. The linear approach is easy to implement and supports memory optimization with respect to copy statements. On the other hand, the reuse approach requires more implementation effort but results in much smaller adjoint vectors, which are more suitable for the vector mode of Algorithmic Differentiation. In this paper, we present both approaches, how to implement them, and discuss their advantages, disadvantages and properties of the resulting Algorithmic Differentiation type. In addition, a new management scheme is presented which supports copy optimizations and the reuse of indices, thus combining the advantages of the other two. The implementations of all three schemes are compared on a simple synthetic example and on a real world example using the computational fluid dynamics solver in SU2.
READ FULL TEXT