SmoothE Paper Accepted to ASPLOS 2025!
As a joint work between Cornell and University of Maryland, SmoothE tackles a major bottleneck in compiler optimization and program synthesis by introducing the first differentiable e-graph extraction algorithm. Unlike traditional methods that struggle with scalability or oversimplified cost models, SmoothE leverages a probabilistic, gradient-based approach to extract optimal expressions from large equivalence classes with support of complex, non-linear objectives and GPU-acceleration.
Built in PyTorch and evaluated on diverse, real-world e-graphs, SmoothE achieves an impressive balance of efficiency and solution quality. Congratulations to the team behind this innovative work
Yixiao Du
publication