commit | 1ba69d96fa77b13a9953f93ecd755361d174af53 | [log] [tgz] |
---|---|---|
author | Scott Todd <scotttodd@google.com> | Mon Nov 09 10:37:42 2020 -0800 |
committer | GitHub <noreply@github.com> | Mon Nov 09 10:37:42 2020 -0800 |
tree | 600b387d852954c826750ed70fc6c81b96d44eba | |
parent | 848b893885da381ce5b4493b37cb169ace6475ee [diff] |
Deduplicate flow.executable ops during Flow dialect transformation. (#3702) This works towards https://github.com/google/iree/issues/1144, but is not generalized beyond `flow.executable`. Deep equivalence checks are performed between pairs of `flow.executable` ops in a new pass run after outlining dispatch regions. These executables will, by construction, have different symbol names so some special comparison logic is needed. When an executable is identified as a duplicate, it is deleted and references to it are updated to point to the canonical executable. This cleans up cases where duplicate executables are generated from the same library calls throughout a program, prior to running further compiler passes or dropping down to codegen. I iterated on the comparison algorithm quite a bit, first printing ops to strings and checking string equality, then caching hashes of those strings using `getChildAnalysis<>` (see [this documentation](https://mlir.llvm.org/docs/PassManagement/#analysis-management)). That had the correctness and performance traits I wanted and was easy to read, but was thread-hostile. The comparison code I settled on is still fast even without multithreading (milliseconds on mobilebert in RelWithDebInfo), thanks to efficient structural checks and plenty of places to short circuit and return early. If we find it needs speeding up, we can 1) Write a `llvm::hash_code` analysis for executables, traversing regions/blocks and using `OperationEquivalence::computeHash()` 2) Query and compare hashes use `getChildAnalysis<>` as a first check (like a bounding box test prior to collision tests) 3) Run the existing deep equivalence check to verify that a hash collision didn't happen The new pass emits [statistics](https://mlir.llvm.org/docs/PassManagement/#pass-statistics), like these for mobilebert (debug build + `-pass-statistics` flag): ``` mlir::iree_compiler::IREE::Flow::DeduplicateExecutablesPass (S) 930 total executable(s) - Number of flow.executable ops before deduplication (S) 891 duplicate executable(s) - Number of flow.executable ops removed as duplicates (S) 39 unique executable(s) - Number of flow.executable ops remaining after deduplication ```
IREE (Intermediate Representation Execution Environment, pronounced as “eerie”) is an MLIR-based end-to-end compiler that lowers ML models to a unified IR optimized for real-time mobile/edge inference against heterogeneous hardware accelerators. IREE also provides flexible deployment solutions for the compiled ML models.
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!
For development, IREE supports both Bazel and CMake on Windows and Linux. We are working on enabling macOS support. For deployment, IREE aims to additionally cover Android and iOS.
Please see the Getting Started pages on IREE's documentation hub to configure, compile, and run IREE in your favorite development environment!
IREE hosts all its documentation and project status dashboards on GitHub Pages. We are still building up the website; please feel free to create issues for the documentation you'd like to see!
We also have some public talks that explain IREE's concepts and architecture:
IREE adopts a holistic approach towards ML model compilation: the IR produced contains both the scheduling logic, required to communicate data dependencies to low-level parallel pipelined hardware/API like Vulkan, and the execution logic, encoding dense computation on the hardware in the form of hardware/API-specific binaries like SPIR-V.
The architecture of IREE is best illustrated by the following picture:
Being compilation-based means IREE does not have a traditional runtime that dispatches “ops” to their fat kernel implementations. What IREE provides is a toolbox for different deployment scenarios. It scales from running generated code on a particular API (such as emitting C code calling external DSP kernels), to a HAL (Hardware Abstraction Layer) that allows the same generated code to target multiple APIs (like Vulkan and Direct3D 12), to a full VM allowing runtime model loading for flexible deployment options and heterogeneous execution.
IREE aims to
IREE is in the early stages of development and not yet ready for broad adoption. Check out the long-term design roadmap to get a sense of where we're headed.
We plan on a quarterly basis using OKRs. Review our latest objectives to get a sense of what we're up to in the near term.
We use GitHub Projects to track progress on IREE components and specific efforts. We use GitHub Milestones to track the work associated with plans for each quarter.
IREE is licensed under the terms of the Apache license. See LICENSE for more information.