| commit | 6b8954be6c4776e775f8bfbeab18e6fd3c31d708 | [log] [tgz] |
|---|---|---|
| author | Quinn Dawkins <quinn.dawkins@gmail.com> | Fri Oct 27 22:09:40 2023 -0400 |
| committer | GitHub <noreply@github.com> | Sat Oct 28 02:09:40 2023 +0000 |
| tree | d015711ccb8ff775e76e3d11b358da7ad94d82ef | |
| parent | 960e8cbe9651246aaaf44a5664f8ffe9e44eb396 [diff] |
[NFC-ish] Split out kernel config into <BACKEND>SelectLoweringStrategy (#15321) This slightly restructures codegen in all of the backends such that the selection of the lowering strategy is separate from the construction and application of the relevant pipeline. This has two potential benefits: 1) By decoupling strategy selection from the lowering, we can move the translation info onto function entry points (and/or individual private functions) rather than the export op. This allows codegen to nest on the function rather than the variant, whereas before this change, strategy selection would still need the higher level of nesting. 2) We can now insert passes between strategy selection and application. This allows expansion of variants based on the set of required capabilities for each strategy, and contraction across variants that can now be statically resolved as identical (think an elementwise operation being independent of SubgroupNonUniform, now we can collapse two variants where the presence of SubgroupNonUniform was the only thing present). Unfortunately for the time being, we still need to register essentially the same set of dialects during the strategy selection pass because the Transform Dialect builders in TransformStrategies can more or less produce whatever they want. Once those strategies are deprecated we can drop the extra registrations.
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.