Abstract

LEGO is a compiler for generating dynamic tiling algorithms for sparse tensor accelerators. Users describe policies that adjust tile boundaries in response to sparse structural properties, separating those policies from tensor expressions, storage formats, and accelerator-specific dispatch code. This separation makes policies reusable across workloads and backends. In an ExTensor simulation, tiles from a generated Dynamic Reflexive Tiling-style algorithm achieve runtimes within 1% of the handwritten baseline.

Article

pdf

BibTeX

@article{ravipatigautham2026,
  title={Portable Dynamic Tiling for Sparse Tensor Accelerators},
  author={Sai Gautham Ravipati and Fredrik Kjolstad and Priyanka Raina and Olivia Hsu},
  journal={Workshop on Languages, Tools, and Techniques for Accelerator Design (LATTE) co-located with ASPLOS},
  year={2026},
  month={March}
}