Webinar : A hybrid approach to Adjoint Algorithmic Differentiation
Wed 29th June 2022
Online

Time: 16:00 AEST 15:00 JST | 14:00 HKT | 14:00 SGT

The AD world basically splits in two. Source transformation can give incredibly efficient adjoints, but is restricted to "simple languages" like a subset of C. On the other hand, operator overloading has successfully handled large production codes, but in general, is less efficient than source transformation. But there's nothing stopping us combining these two ideas. It turns out that this is somewhat tricky to do, but by no means impossible.

In this webinar, NAG Senior Software Engineer, Johannes Lotz, talks about this hybrid approach to Adjoint AD. This approach provides source transformation-like performance whilst being as easy to apply as operator overloading.

This webinar will provide:

  • An overview of AD
  • Results from the front lines: what our customers have achieved
  • The general idea behind the hybrid approach to AD
  • Results on in-house code and QuantLib

Your Presenter:

Dr Johannes Lotz has been the technical lead in the NAG Automatic Differentiation (AD) product team since January 2021. After his PhD at the RWTH Aachen University on "Hybrid Approaches to Adjoint Code Generation using dco/c++", he continued as a postdoctoral researcher in the broad field of AD. As the main developer of the software package dco/c++, Johannes has been working closely with clients from different areas (e.g., computational finance), both in industry and academia. For over 10 years, Johannes has successfully integrated AD solutions into large-scale numerical codes of various shapes and complexity. His main focus is on AD of C++, CUDA, and Fortran codes for CPU and GPU architectures. Johannes is constantly applying his expertise to the innovation of AD techniques and products focusing on adding customer value, while he strengthens know-how and thought leadership in AD.

To register, click here.