Grouptuner: Efficient Group-Aware Compiler Auto-tuning
summary
Grouptuner is a group-aware compiler auto-tuner that groups 200 GCC options into 15 functionally related groups and performs localized group-level mutations starting from historically top-ranked sequences. Evaluated on GCC9 (also tested on GCC11 and GCC13) across Cbench, Polybench and SPEC with 500 iterations, it finds option combinations that average 12.39% improvement over -O3, reaches that quality using only 77.21% of the runtime of random search, and yields >30% max gains on some programs.
AI-written summary · about
chapters (13)
- 0:00Introduction to GroupTuner and compiler autotuning
- 1:53Existing autotuning limits
- 4:21GroupTuner's core idea
- 6:09GroupTuner structure
- 8:17Group-aware mutation search
- 9:01History-based search
- 10:28Experimental setup
- 10:49Performance results
- 11:50Why GroupTuner wins
- 13:16Search efficiency
- 13:58GCC version results
- 14:19Conclusion and future work
- 14:42Q&A
related talks
19mCausalTuner: Feature-Aware Causal Guidance for Compiler Auto-tuning
15mNot So Fast: Understanding and Mitigating Negative Impacts of Compiler Optimizations on Code Reuse G
33mImproving the Effectiveness and Generality of GCC Auto-Vectorization
52mUnderstanding Compiler Optimization
01h 00mBeating Hand Tuned Assembly in Compiled Languages
29mEmpirical Observations about Profile-Guided Optimizations for Mainstream C/C++ Compilers
14mOn-Stack Replacement for Program Generators and Source-to-Source Compilers
46mPerformance-Tuning Twitter Services with Graal and Machine Learning