HieraSynth: A Parallel Framework for Complete Super-Optimization with Hierarchical(…)
summary
Framework implements component-based synthesis with SMT backends, adds hierarchical instruction-choice embedding, adaptive partitioning and unrealizability pruning, and runs subspaces in parallel. Claims: handles ISAs up to 700 instructions, finds synthesizes 7–8-instruction programs that previous tools could only reach for 1–3 instructions, reports up to 10.66× larger ISA or 4.75× larger program-size scaling, finds a 5.7× faster vector kernel example, and achieves near-linear speedup to 72 cores while proving optimality under its cost model.
AI-written summary · about
chapters (10)
- 0:00HieraSynth intro
- 1:07Manual vector optimization challenges
- 2:52Super optimizer limitations
- 3:55Hierarchical decomposition
- 5:21Parallel subspace partitioning
- 6:26Adaptive granularity
- 7:51Search efficiency optimizations
- 8:33Evaluation goals
- 9:39Vector program results
- 11:02Conclusion
related talks
14mDataflow-Based Pruning for Speeding up Superoptimization
19mMinotaur: A SIMD-Oriented Synthesizing Superoptimizer
21mProgram Synthesis From Partial Traces
23mSynthesizing Efficient Super-Instruction Sets for Ethereum Virtual Machine
20mOptimal Program Synthesis via Abstract Interpretation
30mAutomated High-Level Synthesis Design Modularization via E-Graph Anti-Unification
19mHydra: Generalizing Peephole Optimizations with Program Synthesis
20mMISAAL: Synthesis-Based Automatic Generation of Efficient and Retargetable(…)