Kalman filtering is a statistical estimation algorithm that uses measurements observed over time and a physical model to minimize errors due to uncertainties in the measurements (measurement noise) and the physical process (process noise). Kalman Filters are the de facto standard for charged particle reconstruction in medium and high-energy nuclear physics experiments such as those at RHIC and envisioned at the EIC. The proposal is to develop a code generator to generate custom, tuned, optimized Kalman Filter functions that can be included in the software workflow of any application requiring Kalman filtering. The generated code will automatically generate the Jacobian, implement architecture-aware optimizations, and be tuned for the specific problem size and hardware architecture. The generated code will include C++ source code (e.g., CUDA, HIP, OpenMP 5.0, OpenCL, Kokkos, RAJA) with Fortran and Python bindings. Phase I work will focus on automatic code generation and optimization for the Linear and Extended Kalman Filter (EKF). The goal of the Phase I project will be to demonstrate the technical feasibility of a code generation tool for Kalman filters and its performance benefits. The performance portable Kalman Filter code generation tool will be developed for and tested with particle track reconstruction and particle track simulation codes such as ACTS and Geant4. This will allow for performance portable Kalman Filter codes in applications of importance to the DOE. This will allow deployment on the most cost-effective GPU platform near the colliders, as well as during post-processing on leadership class exascale resources with a range of GPU acceleration options. Kalman filtering is also used in a wide range of applications, including guidance, navigation and control of vehicles, signal processing, robot motion planning, trajectory optimization, modeling of the central nervous system among many others.
