Max Aehle(AG Scientific Computing, RPTU Kaiserslautern-Landau)
hosted by Seminar Series on Scientific Computing
"Reverse-Mode Automatic Differentiation of Compiled Programs"
Algorithmic differentiation (AD) is a set of techniques to obtain accurate derivatives of computer-implemented functions. For gradient-based numerical optimization purposes, the reverse mode of AD is especially suited - the run-time it needs to compute a gradient of the objective function is proportional to the run-time of the objective function, and independent of the number of design parameters.
In practice, classical AD tools require that the source code of the computer-implemented function is available, in a limited set of programming languages. As a step towards making AD applicable to cross-language or partially closed-source client programs, we developed the new AD tool Derivgrind . Derivgrind leverages the dynamic binary instrumentation framework Valgrind to add forward-mode AD logic to the machine code of compiled computer code.
In this talk, we present the new index-handling and tape-recording capabilities that we added to Derivgrind during the last months . In combination with a simple tape evaluator program, they enable operator-overloading-style reverse-mode AD for compiled programs.
 Max Aehle, Johannes Blühdorn, Max Sagebaum, Nicolas R. Gauger. Forward-Mode Automatic Differentiation of Compiled Programs. arXiv:2209.01895
Thursday, 02.02.2023, 11:45
Hybrid (Room 32-349 and via Zoom)