microsoft / triton-shared

Shared Middle-Layer for Triton Compilation
MIT License
134 stars 26 forks source link

Add `TritonTilingExt` dialect to support ops not representable in linalg #45

Closed nhat-nguyen closed 8 months ago

nhat-nguyen commented 8 months ago

The TritonTilingExt dialect leverages linalg's TilingInterface to add tiling & fusion support for operators that cannot be represented in linalg Much of the tiling implementation is borrowed from linalg's tiling code. As an example, I have added a barebone version of cumsum that represents the cumulative sum of the inner-most dimension of a tensor. Other backends can then lower the op to lower-level implementation as needed after applying tiling & fusion.