| commit | 1eb5713823a105c51719954f3e3e2fff2952c09a | [log] [tgz] |
|---|---|---|
| author | Quinn Dawkins <quinn.dawkins@gmail.com> | Wed Nov 08 22:38:35 2023 -0800 |
| committer | GitHub <noreply@github.com> | Thu Nov 09 01:38:35 2023 -0500 |
| tree | 705f88bb786dd5087b4a173a26c268c674e15fc0 | |
| parent | 9b4606bed2ffdadb7b9c6776af0ec47440d3bd7f [diff] |
[HAL] Add configuration stage in executable lowerings (#15362) The core idea behind this split is to allow for passes to be added in between strategy selection and translation, thereby expanding and contracting the set of variants per executable based on the known feature requirements for those strategies. The reason a split is needed is because strategy selection itself should not care what other variants of the executable are being compiled, however to deduplicate them we need to nest at the executable level. Codegen was already partially structured to enable this (by way of encoding the lowering pipelines as an IR attribute). This PR just recovers the separation that was already there. In terms of immediate benefits, this adds an option to dump executable configurations (essentially post-kernel config dispatches) and substitute them in, making it significantly easier to get the default lowering configuration for a dispatch from the compiler, tweak it, and plug it back in. Additionally benchmark dumps now happen after this stage for the same reason.
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.