EVENTS

Dr. Peter Pirkelbauer from Lawrence Livermore National Laboratory’s Center for Applied Scientific Computing

Compiler optimization reports contain valuable insights, but their technical complexity often makes them difficult for programmers to interpret and act upon. This talk presents CompilerGPT, a framework that connects compilers, large language models, and a user-definable evaluation harness in an iterative compile, optimize, and evaluate workflow. CompilerGPT uses optimization reports, source code, compiler diagnostics, and performance measurements to guide LLM-generated code transformations while checking correctness and tracking results.

We will describe the framework, its configurable compiler and model interfaces, and experimental results using Clang and GCC, multiple LLMs, and representative sequential and OpenMP benchmarks. We will review demonstrated speedups of up to 6.5x, along with the variability, cost, and user involvement observed across workloads. The talk will also present updated results and examine CompilerGPT in the broader context of coding agents, discussing how compiler feedback, testing, profiling, and performance evaluation can serve as tools within agentic software-development systems.

Institutions

• Helmholtz Platform for Research Software Engineering - Preparatory Study