Cost of Soundness in Mixed-Precision Tuning
summary
Compares five open-source mixed-precision tuners on 100 FPBench straight-line programs and measures speedup, error budget usage, and runtime. Key findings: sound static with regime inference matches or outperforms dynamic tuners on these benchmarks, while standalone static tools are overly conservative and dynamic tools sometimes exceed the allowed error by up to 9 orders of magnitude. Experiments use uniform 128/64-bit baselines (quad cost 5s vs double <1s) and show a mixed assignment that is 10% faster than quad but still meets tight error bounds; regime inference raises optimization time (slowest approach, vs some tools finishing <0.5s).
AI-written summary · about
chapters (8)
- 0:00Mixed precision tuning intro
- 0:21What mixed precision tuning does
- 3:14Sound vs dynamic tuners
- 5:00Evaluating tuners with benchmarks
- 6:42Cost of soundness metric
- 9:10Comparison study results
- 12:00Key findings and limits
- 12:43Q&A on ML kernel techniques
related talks
23mBeating the Benchmarks: Co-evolving the ISA and Development Tools
01h 09mThere's Treasure Everywhere
50mRecent Advances in Floating-point (Static) Analyses
01h 00mFastware
22mModular Optimization-Based Roundoff Error Analysis of Floating-Point Programs
56mWrite Fast Code Like a Native
01h 02mFastware
42mAdventures in Efficiency and Performance