ALT: Joint Tuning for Data Layouts and Loops in Deep Learning Compilation

What is it about?

This is a work to improve deep learning inference performance by compilation techniques. The main contribution is to enable flexible data layout optimization without re-implementing operators. With this, we can jointly tune data layouts, and existing loops easily.

Featured Image

Photo by Markus Spiske on Unsplash

Photo by Markus Spiske on Unsplash

Why is it important?

It shows that the bidirectional tuning flow is more beneficial than simply putting different optimization techniques into different system layers and no interaction is involved. In order to do this, a new infra for versatile data layout transformation is developed where some primitive functions help to transform data layouts without re-implementing operators like that in existing works. And, a bidirectional auto-tuning framework is built atop the transformation, which enables interaction between two important optimization techniques.