Locality-Aware Automatic Differentiation on the GPU for Mesh-Based Computations

This paper proposes a GPU AD system built specifically for mesh-based computation, where locality and sparsity are the whole game. Instead of tracing a big global graph like general-purpose autodiff stacks, it evaluates derivatives per mesh element, keeps the work in registers/shared memory, and assembles sparse outputs directly on the GPU. It also supports matrix-free operations like Hessian-vector products, plus dynamic sparsity updates when interaction changes the neighborhood structure.

For game-dev-adjacent graphics work, the practical angle is speed and scalability in solvers that already live on the GPU: cloth, elastic simulation, ARAP deformation, parameterization, smoothing, frame fields, and spherical optimization. The authors say it outperforms PyTorch, JAX, Warp, DrJIT, EnzymeAD, and Thallo across Newton, Gauss-Newton, L-BFGS, and gradient descent workflows. The paper was submitted Aug. 30, 2025, revised June 4, 2026, and later appeared in ACM TOG 45(4).