Legacy Fortran Code to GPU-Enablement, Agentic AI Conversion to C++
Anthony Marinov A poster for the 2026 MICDE Predictive Science Symposium describing our progress on our GPU acceleration work as part of the SHAPE project.
We (at the time of presenting this poster) managed to successfully offload the main parallel loop to the GPU of our MI300A system utilizing unified memory, as a proof of concept and viability. We also successfully utilized LLM’s and agentic AI to quickly translate the codebase from Fortran into C++ in a matter of weeks.
Our future work will focus on optimizing the code for GPU acceleration, identifying and parallelizing other bottlenecks, and integrating additional features from the simulation research groups.
