| commit | ba3e6a7223bcc9c5e29e00b363bb437b219c1de1 | [log] [tgz] |
|---|---|---|
| author | Benjamin Maxwell <benjamin.maxwell@arm.com> | Tue Oct 10 18:35:40 2023 +0100 |
| committer | GitHub <noreply@github.com> | Tue Oct 10 10:35:40 2023 -0700 |
| tree | 892dbed0e383a991a6d4a83b109c96aabb0d48ef | |
| parent | 729bb75cfeeac27c8559a54f3934cf4bfebe91eb [diff] |
[CPU][SVE] Enable scalable vectorization and tiling for non-padded matmuls (#15108) This patch implements a "vertical slice" that allows for the scalable vectorization of matmuls on AArch64 targets with SVE. This only updates lowerings along that path, so more is needed to enable scalable vectorization everywhere. This required a few changes: - The default vector sizes of matmuls on AArch64+SVE are now (8, [32], 16) * That is a middle scalable dimension (i.e. 32 x vscale) - `iree-llvmcpu-tile-and-fuse` now generates vscale bounded loops for scalable tiles - `TilingConfig` now returns a pair of tile sizes and scalable flags (`SizesAndScalableFlags`) for vector sizes * This allows connecting the scalable sizes to the generic vectorizer (which passes them down to the linalg vectorizer) A few unit tests have been added for this, but a complete e2e test is not possible without SVE testing infrastructure.
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.