Validation-Centric AI-Assisted GPU Porting for Legacy Weather Simulation Code Exceeding 250,000 Lines

Recently, LLM-based CLI agents have emerged as practical tools for accelerating GPU Porting of large-scale legacy scientific applications. This study proposes a Validation-centric AI-assisted workflow that adapts CReSS, a Fortran-based weather simulation code exceeding 250,000 lines, to GPU environments while maintaining scientific validity.

This workflow utilizes AI agents to perform the following processes:

• OpenMP region extraction and application of OpenACC transformations
• Generation of dump-based kernel benchmarks from physically meaningful simulation states
• Element-wise comparison with reference data and application-level validation

Experimental results showed that the workflow generated numerically verified GPU implementations for 162 target kernels, achieving a 5.1x application-level performance improvement compared to actual development costs. Notably, it detected numerical discrepancies in 5 kernels caused by floating-point and intrinsic function differences, providing feedback to developers.

The study suggests that AI-assisted porting of large-scale legacy code requires, beyond simple code generation, context management across sessions, runtime state reconstruction, and validation-centric workflow design.