FBTuner: A Feedback-Directed Approach for Safe Mixed-Precision Tuning
Xinyi Li, Ganesh Lalitha Gopalakrishnan · 2024
Porting high-performance computing (HPC) applications to lower or mixed-precision formats offers potential benefits, such as reduced computation and power consumption. However, this process presents challenges, including higher rounding errors, poor convergence, floating-point exceptions, and incurs considerable effort, particularly when leveraging low-precision hardware. Current precision-tuning approaches fail to comprehensively address these challenges and do not effectively utilize low-precision hardware.While not as accurate as converting the code to lower precision, this is a very effective tradeoff in terms of design-space search, as many precision settings may need to be explored during precision tuning.FBTuner empowers designers to confidently implement mixed precision in their projects while addressing the key challenges of porting HPC programs to lower or mixed-precision formats. We plan to demonstrate FBTuner on HPC proxy applications and demonstrate that the porting does not result in significant loss of accuracy or increase the number of iterations to attain convergence.