AI-Assisted GPU Porting of a 250k+ Line Legacy Weather Simulation Code
Summary
An arXiv CS paper describes a validation-centric AI-assisted workflow for porting a 250k+ line legacy weather simulation to GPUs. The approach uses AI agents to extract OpenMP regions, generate dump-based kernel benchmarks, apply OpenACC transformations, and validate results against reference data, achieving a 5.1x application-level speedup and uncovering numerical discrepancies for feedback to developers. The work highlights the importance of session-wide context and robust validation in AI-assisted porting of large scientific codes.