commit | 0d9e58740c78364ec7b4084d5dbeebd88185c4c2 | [log] [tgz] |
---|---|---|
author | Kunwar Grover <groverkss@gmail.com> | Tue Jun 25 22:09:19 2024 +0100 |
committer | GitHub <noreply@github.com> | Tue Jun 25 22:09:19 2024 +0100 |
tree | 222da9ed55a9214120a3626a8b98bb118165d203 | |
parent | 5fcdddb7c76001de17f475e75c0f1e3ba4ea957b [diff] |
[NestedLayout] Make layout use strides instead of basis (#17284) This patch makes the layout store how threads and subgroups are distributed as "strides" instead of basis. Strides are just a mapping from `virtual tid --> tid` (where tid is thread id). The mapping is a dot product between `virtual tid` and `strides`: ``` tid = vtid * stride ``` If we put some restrictions on `strides` and make the mapping can be made invertible with the inverse mapping: ``` vtid_i = tid floordiv stride_i mod size_i ``` The advantage with strides is that strides are only dependent on the dimension they are working on, while basis actually depend on the order of dimension because to get the distribution for dimension `i`, you need to know the distribution for dimensions `i-1`. (If you think about it, this is just skewing to remove dependency of`i` on `i-1`). With strides, dimensions are essentially independent, and we do not need any orders in the layout, which hugely simplifies the layout. This is much closer to what CuTe layouts do (somewhat). Along with this change, distribution now actually uses 2 new ops to query the `tid -> virtual tid` map, which makes distribution independent of these changes. The reason for implementing this was that with strides, the layouts actually have a canonical form, unlike with basis, because different active ids, order, basis permutations can represent the same layout. Having a canonical form helps a lot when try to match two different matmul anchors.
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.
Community meeting recordings: IREE YouTube channel
IREE is licensed under the terms of the Apache 2.0 License with LLVM Exceptions. See LICENSE for more information.