AI agents help port 250k-line weather code to GPUs with validation

The paper describes a validation-centric AI-assisted workflow that ports a legacy Fortran weather simulation with over 250,000 lines to GPU hardware. It uses an AI agent to identify OpenMP regions, generate dump-based kernel benchmarks from meaningful simulation states, and apply OpenACC transformations. The resulting GPU kernels are checked element-wise against dumped reference data and then tested with full-application runs. This approach produced numerically valid implementations for 162 of the target kernels.

Development remained practical because the AI agent handled much of the boilerplate work, reducing manual effort. Wall-clock time spent on the porting effort was kept within feasible limits for the research team. The workflow avoided lengthy manual refactoring by relying on automated extraction and transformation steps. Overall, the port achieved a 5.1x application-level speedup on a real typhoon simulation.

The workflow depends on having dump-based validation data that captures runtime states. It requires the AI agent to maintain session-spanning context and reconstruct those states accurately. Small omissions in static analysis can lead to costly recovery steps, as mismatches must be traced back to source differences. Without such dump capabilities, the validation-centric approach may not be applicable.

The case study shows that AI-assisted GPU porting must be designed around validation, not just code generation. Preserving scientific credibility demands checking numerical results against trusted references. For large legacy scientific applications, this validation focus helps catch subtle floating-point and intrinsic-function differences. Thus, similar workflows could benefit other domains where code correctness is as important as performance.