Abstract: Floating-point arithmetic is foundational to scientific computing in HPC, and choices about floating-point precision can have a significant effect on the accuracy and speed of HPC codes. Unfortunately, current precision optimization tools require significant user interaction, and few work on the scale of HPC codes due to significant analysis overhead. We propose an automatic search and replacement system that finds the maximum speedup using mixed precision given a required level of accuracy. To achieve this, we integrated three existing analysis tools into a system that requires minimal input from the user. If a speedup is found, our system can provide a ready-to-compile mixed-precision version of the original program.
Best Poster Finalist (BP): no
Poster summary: PDF
Reproducibility Description Appendix: PDF
Back to Poster Archive Listing