| commit | 9fd55d2843081b0255d3c18c5f3716d98ce22570 | [log] [tgz] |
|---|---|---|
| author | Quinn Dawkins <quinn.dawkins@gmail.com> | Thu Jun 20 16:31:42 2024 -0400 |
| committer | GitHub <noreply@github.com> | Thu Jun 20 16:31:42 2024 -0400 |
| tree | a8f0c93f9f38b51bd72389038a56c7c98c9347ce | |
| parent | d01fb2341c12d81ff54455712c9a33dea70b13ff [diff] |
[Codegen][GPU] Update greedy tile + fuse pipeline to generate mfma (#17617) This adds intrinsic packing and reshape propagation patterns to LLVMGPUTileAndFuse to allow for generating mfma operations. This adds a few passes to invoke a few necessary patterns for the pipeline to generate (good) code. 1. PropagateReshapesByExpansion to propagate reshapes introduced after decomposing tensor.pack/unpack towards the edges of the kernel in the hopes that the destination can line up properly. 2. IREE::GPU::PackToIntrinsics to pack based on the lowering config specified mma kind. 3. IREE::GPU::DistributeMmaToLanes to distribute iree_gpu.multi_mma ops to lanes, similar to another tiling level. There are a few known outstanding issues. 1. We run `ConvertToDestinationPassingStyle` twice to re-link the kernel destination with the body after decomposing `tensor.unpack`. This is to work around an issue with EliminateEmptyTensors being unable to analyze `flow.dispatch.tensor.store` ops with slicing behavior properly. After workgroup distribution is refactored to generate an scf.forall, this needs to be revisited. 4. iree_gpu.shuffle_tensor lowering to `tensor.insert_slice` is still broken. This will need to be reworked to support dynamic shapes. 5. Currently, because of the way the layout works, only MFMA_16x16x16 works. To support other layouts we will need another level of expanding to the intrinsic implicit layout and then propagating those expand_shapes. This will likely need to happen after reduction tiling unless we want to teach tile + fuse to swap tensor.expand_shape ops with tensor.extract_slice.
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.