[Glean] Operator Fusion

Published: June 28, 2020 by SingularityKChen (Last updated: June 29, 2020 )

• Categories:
• Glean 53

• Tags:
• OperatorFusion 1

• Operator Fusion

Operator Fusion¹

There are many opportunities, where fused operators—in terms of fused chains of basic operators—can significantly improve performance.

First, fusion allows eliminating the unnecessary materialization of intermediates, whose allocation and write is often a bottle-neck.

Second, fusion can eliminate unnecessary scans of inputs by exploiting temporal cell or row locality.

Third, multiple aggregates with shared inputs ever-age similar opportunities for DAGs (directed acyclic graphs) of multiple aggregates over common subexpressions (CSEs).

Fourth, “sparse drivers” allow sparsity exploitation across entire chains of operations.

1. Boehm, Matthias, et al. "On optimizing operator fusion plans for large-scale machine learning in systemml." arXiv preprint arXiv:1801.00829 (2018). ↩