Whole-Program FP Error Tracking
Scaling floating-point error tracking with shadow execution
Floating-point rounding error accumulates silently in scientific codes. Shadow execution tracks it by running every floating-point operation a second time in higher precision and comparing the two results.
This project scales shadow execution to whole programs: optimized builds, applications split across multiple modules, and MPI workloads. It instruments programs with LLVM so that error can be followed across module and process boundaries.
Status: in preparation for submission to CGO ‘27.
Tech: C, C++, Python, LLVM