Galvez J1
1Nci, Canberra, Australia
Biography:
Jorge is an HPC specialist who has experience in the computational chemistry and geophysical fluid dynamics fields. His main strength is GPU programming and porting codebases from CPU to GPUs.
Abstract:
Scientific software development is primarily driven by graduate students and postdoctoral researchers who often lack formal training in high-performance computing and software engineering. Consequently, the pace of discovery in fields like Bioinformatics, Computational Chemistry, and Geophysics is increasingly throttled by software quality, usability, and the difficulty of extending legacy code to new methodologies. Without rigorous oversight, these codebases rapidly accumulate technical debt and domain-specific workarounds.
A critical challenge in these environments is the reliance on "testing the physics" rather than the code. This lack of unit-level verification creates a slow feedback loop, making it nearly impossible to refactor for modern hardware. This is particularly evident when attempting to migrate from CPU-based architectures to GPUs, where structural re-architecture is often required for performance portability.
This work explores the application of agentic coding systems and modern AI-driven development workflows to bridge this gap. I present a series of modernization workflows that leverage autonomous agents to automate the migration of build systems, the generation of comprehensive unit testing suites, and the re-architecturing of legacy code. These workflows demonstrate how agentic systems can prepare scientific software for GPU porting in a performance-portable manner, reducing the burden on researchers and accelerating the transition to next-generation computing architectures.
