CausalTuner: Feature-Aware Causal Guidance for Compiler Auto-tuning
summary
Framework applies LLVM-based tuning on AMD hardware, extracts program features from optimized bitcode, clusters similar programs with synergy clustering, runs Fast Causal Inference to mine feature→pass causal rules, then performs a Two-Phase Rule-Injected Search (subspace sampling + retro-evolutionary fine-tuning with 1–5% conservative vs ~15% aggressive mutation). Evaluation on cBench, Polybench, SPEC CPU2017 and llama.cpp reports a 17.77% average speedup, 23.21% token-throughput gain on llama, and dramatically faster cold-start convergence (claimed ~1430% acceleration in first 50 iterations).
AI-written summary · about
chapters (15)
- 0:00CausalTuner introduction
- 0:43Compiler optimization challenges
- 2:49Search-based tuning problems
- 5:38CausalTuner framework
- 6:43Synergy clustering stage one
- 7:46Causal rule mining stage two
- 8:48Online search with guidance
- 9:53Experimental setup benchmarks
- 10:36Performance vs autotuners
- 12:26Ablation study components
- 13:08Generalization to workloads
- 14:13Cross-platform validation
- 14:57Interpretability of guidance
- 16:01Conclusion and future
- 16:45Q&A
related talks
20mGrouptuner: Efficient Group-Aware Compiler Auto-tuning
22mMulti-level Machine Learning-Guided Autotuning for Efficient Code Generation on a Deep(…)
11mAccelerated Software for a Post-Moore World
21mTowards Compiler-Guided Static Analysis
22mBeacons: An End-to-End Compiler Framework for Predicting and Utilizing Dynamic ...
56mPerformance tuning Twitter services with Graal and Machine Learning
01h 00mBeating Hand Tuned Assembly in Compiled Languages
10mAuto-tuning vllm