commit | 1a635643583e8787030b9fe57fae5d0543644f65 | [log] [tgz] |
---|---|---|
author | Benjamin Maxwell <benjamin.maxwell@arm.com> | Thu Sep 28 19:00:55 2023 +0100 |
committer | GitHub <noreply@github.com> | Thu Sep 28 11:00:55 2023 -0700 |
tree | a8eddb83c1d3f8c3520e84cccf74d01bdacadeb9 | |
parent | 15ace7e2f687cbf5c7821cbb7f28e95a604452a2 [diff] |
Refactor IREECodegenAttrs to use typed array parameters (#15032) This reworks the attributes in IREECodegenAttrs so that typed [Optional]ArrayRefParameters are used instead of untyped ArrayAttrs. Doing this removes the manual type-checking and manual conversions to/from SmallVector<T>. This did require a small restructuring of the lowering_config attribute, so instead of having multiple nested arrays IntegerAttrs for tiling levels, it now has one array of TilingLevelAttrs. The TilingLevelAttr then contains the tile sizes and tile interchange values for a particular level. The TilingLevelAttr has a custom assembly format that keeps its syntax as close as possible to the previous syntax, to avoid as much test churn as possible. We plan to follow up this patch with another change that extends the TilingLevelAttr to support scalable sizes (using the canonical MLIR syntax), which will need a custom parser anyway, so we don't see this as a dealbreaker. Example configs: ``` // No tile interchange (no change in syntax) #iree_codegen.lowering_config<tile_sizes = [[2, 64, 64, 64], [1, 1, 1, 4], [0, 0, 0, 0]]> // With tile interchange // Before: #iree_codegen.lowering_config< tile_sizes = [[32, 64, 0], [8, 32, 0], [0, 0, 16]], tile_interchange = [[1, 0, 2], [], []]> // After: #iree_codegen.lowering_config<tile_sizes = [ {sizes=[32, 64, 0], interchange=[1, 0, 2]}, [8, 32, 0], [0, 0, 16]]> ``` --- These changes makes the `lowering_config` easier to extend for a follow on patch that starts adding scalable tile sizes.
IREE (Intermediate Representation Execution Environment, pronounced as “eerie”) is an MLIR-based end-to-end compiler and runtime that lowers Machine Learning (ML) models to a unified IR that scales up to meet the needs of the datacenter and down to satisfy the constraints and special considerations of mobile and edge deployments.
See our website for project details, user guides, and instructions on building from source.
IREE is still in its early phase. We have settled down on the overarching infrastructure and are actively improving various software components as well as project logistics. It is still quite far from ready for everyday use and is made available without any support at the moment. With that said, we welcome any kind of feedback on any communication channels!
See our website for more information.
IREE is licensed under the terms of the Apache 2.0 License with LLVM Exceptions. See LICENSE for more information.