Code-aware Neural Surrogates
Accelerating performance-critical kernels in HPC applications
Large scientific simulations spend most of their time in a small number of performance-critical kernels. This project integrates machine learning with program analysis and compiler optimizations to replace such kernels with learned neural surrogates, speeding up large-scale HPC codes while keeping the rest of the application unchanged.
The approach targets C, C++, and Fortran applications and is evaluated across ten HPC benchmarks.
Status: in preparation for submission to IPDPS ‘27. Supported by the U.S. Department of Energy.
Tech: C, C++, Fortran, Python, OpenMP, CUDA, ONNX, LLVM/Clang, LFortran, PyTorch, NumPy, TensorRT