Oracle-Free Repair Synthesis for Floating-Point Programs
summary
Introduces microstructure — local neighborhoods where repeating floating-point error patterns let the tool infer accurate values without high-precision oracles — and builds Assessor, an oracle-free dynamic synthesis that samples neighborhoods, estimates expected value and error variance, and fits polynomial patches only on erroneous intervals. Evaluated on 36 numerical functions (17 using GSL/library code), repairs cut errors by orders of magnitude, synthesizes patches in under one second, and produces patches that run ~9 ns versus 36 ns for originals.
AI-written summary · about
chapters (7)
- 0:00Introduction
- 1:05Floating-point errors background
- 3:13Oracles are hard to obtain
- 5:20Microstructure insight for repair
- 8:12How Assessor repairs
- 10:06Evaluation results
- 12:57Key takeaways
related talks
50mRecent Advances in Floating-point (Static) Analyses
22mProgram Synthesis for Artifacts beyond Programs
16mMemory-Efficient Fixpoint Computation
59mUnder-approximation for Scalable Bug Detection
22mModular Optimization-Based Roundoff Error Analysis of Floating-Point Programs
21mExact Loop Bound Analysis
20mAutomatic and Incremental Repair for Speculative Information Leaks
28mLLOR: Automated Repair of OpenMP Programs