| commit | b476624ca08a3ed8c0c9d24c2987911c3288930d | [log] [tgz] |
|---|---|---|
| author | MaheshRavishankar <1663364+MaheshRavishankar@users.noreply.github.com> | Mon Apr 24 21:50:28 2023 -0700 |
| committer | GitHub <noreply@github.com> | Tue Apr 25 04:50:28 2023 +0000 |
| tree | ea83509447b45fa971a4382bf4cc18efc0f96d25 | |
| parent | 34e07706941838b442fbf21d1b615b54590d7f05 [diff] |
Add support for computing dispatch workgroup count using program slices. (#13038) This PR changes the way the workgroup count calculation for dispatches are handled by default. Previous approaches required to have context of what is put into a dispatch and what values are captured as workload. Managing this implicit link gets tricky. The approach in this PR assumes that all the information required for computing the number of workgroups is captured within the dispatch after the `flow.dispatch.workgroups` op is formed. In the backend, during tile and distribute a program slice is used to capture the computation that determines the number of workgroups, and is cloned into the workgroup count region. One complication is that the signature of the `flow.dispatch.workgroups` operation is not preserved through the compilation pipeline. So in the backends it is not possible to connect the leaves of the slice back to the operands of the original `flow.dispatch.workgroups` operation that were captured as workload. To get around this, these operands are annotated within the dispatch with their position in the workload list. Another side-effect of this change is that there is no need to instantiate `tensor.dim` operations to compute the workload. In dynamic shape cases, this leads to repeated instantiations of the same `tensor.dim` value leading to compile-time explosion even though these get CSE-ed away later. To address the transform dialect path a new op transform.iree.populate_workgroup_count_region_using_num_threads_slice is s added that handles the workgroup count materialization for the transform dialect path. As a consequence - `transform.iree.tile_to_forall_and_workgroup_count_region` is deprecated in favor of using `transform.structured.tile_to_forall` and the newly added transform dialect operation above. - `transform.iree.convert_conv2d_to_img2col_and_adjust_workgroup_count_region` is also deprecated. Fixes #11608
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.