Accelerating Evolutionary GPU Kernel Optimization with Record-Replay

Abstract: Optimizing graphics processing unit (GPU) kernels requires navigating interacting choices across source code, compiler transformations and launch parameters. In this talk, I will present Record-Remix-Replay (R³), a hierarchical optimization framework that uses language learning model (LLM)-guided evolutionary search to explore CUDA/HIP kernel rewrites and record-replay-based auto-tuning to efficiently evaluate each candidate and tune lower-level compiler and runtime decisions. I will describe how R³ makes broad search over the GPU optimization stack practical and present results on scientific GPU workloads, including a QUDA lattice quantum chromodynamics case study on AMD MI300A GPUs.

Bio: Daniel Nichols is the Sidney Fernbach postdoctoral fellow at Lawrence Livermore National Laboratory. He received his Ph.D. in computer science from the University of Maryland, College Park, where he worked with the Parallel Software and Systems Group.

Series: See all upcoming talks at the LANS Seminars page.