Automated Floating-Point Precision Tuning: Progress, Challenges, and Future Opportuni

This invited talk by Cindy Rubio-Gonzalez covers a decade of research on automated floating-point precision tuning, the process of finding mixed-precision versions of numerical programs that run faster while meeting a user-specified error tolerance. She frames the core challenges: an exponential search space over variables and precision formats, the cost of empirically evaluating every candidate configuration, generalization beyond the representative inputs used during tuning, and the difficulty of specifying acceptable error. She then walks through a lineage of tools developed with her students at UC Davis. An early delta-debugging-based search systematically lowers variable precision and guarantees a local minimum; a later white-box tool groups variables into usage-based communities to shrink the search space and reduce costly type casting; FPLearner trains graph neural networks on a floating-point-tailored program representation to predict performance and accuracy, cutting search time by up to 61% without losing speedup; and the most recent tool, presented at the conference, applies deep reinforcement learning by modeling the search as a Markov decision process, finding better speedups and often smaller errors while exploring a broader configuration space. A case study with NCAR applied tuning to Fortran weather and climate model hotspots with hundreds of variables, doubling hotspot speed but losing gains to casting overhead in the full model. Future directions include LLMs, GPU and multi-format support, compiler-aware tuning, and verification to guarantee reliability.