PDL: Dialects for Representing and Transforming Pattern Rewrites

Hi all,

Just wanted to give an update on the PDL (Pattern Descriptor Language) work done by an intern(Jeff Niu) that we hosted at Google last year. This work is intended to provide a way of representing and transforming pattern rewrites in MLIR, with MLIR. For those unfamiliar with this work, by using dialects to represent patterns we get quite a few benefits: pattern merging(i.e. faster pattern matches), new frontends besides tablgen/DRR(think python, swift, whatever you want), runtime pattern generation, etc. This work was originally proposed and discussed here (https://groups.google.com/a/tensorflow.org/g/mlir/c/j_bn74ByxlQ), with an accompanying presentation to the MLIR ODM group(https://drive.google.com/corp/drive/folders/1hb_sXbdMbIz95X-aaa6Vf5wSYRwsJuve).

Iā€™m happy to give an update that this work has started to find its way upstream(sorry for the delay), with a first set of revisions already out for review(rooted at D84578). I welcome all interested to review those revisions and otherwise get involved with the work.

Thanks,
River

1 Like