vgvassilev/clad

What it solves

Clad provides automatic differentiation (AD) for C++, allowing developers to automatically generate C++ code for computing derivatives of mathematical functions. This eliminates the need to manually derive and code the gradient, Hessian, or Jacobian of complex functions, which is often error-prone and tedious.

How it works

Clad is implemented as a plugin for the Clang compiler and is based on the LLVM compiler infrastructure. It uses source-transformation to analyze C++ source code and automatically generate the corresponding C++ code for the derivative. It supports both forward-mode and reverse-mode AD, as well as combinations of both for higher-order derivatives.

Who it’s for

It is designed for C++ developers and researchers who need to compute derivatives of mathematical functions for optimization, machine learning, or scientific computing, and those who need to estimate floating-point errors in their programs.

Highlights

• Multiple AD Modes: Supports forward-mode, reverse-mode, and combined modes for Hessian and Jacobian matrices.
• Comprehensive API: Provides dedicated functions like clad::differentiate, clad::gradient, clad::hessian, and clad::jacobian.
• Floating-Point Error Estimation: Includes a specialized tool (clad::estimate_error) to compute the floating-point error of a program.
• Compiler Integration: Works as a Clang plugin or within Jupyter Notebooks via the xeus-cpp kernel.
• C++ Feature Support: Extensive coverage of modern C++ features with a robust fallback and recovery system.