Google Git
Sign in
opensecura / 3p / openxla / iree / 41115bba05960e563791ce6ed1af26093f4fab1e
commit41115bba05960e563791ce6ed1af26093f4fab1e[log] [tgz]
authorStanley Winata <68087699+raikonenfnu@users.noreply.github.com>Tue Nov 26 15:25:35 2024 -0800
committerGitHub <noreply@github.com>Tue Nov 26 15:25:35 2024 -0800
treeecc0bb689b69c8571b540fe1f21cfa47dbdb5ff2
parent5708d428b029504498cb00e19710d7ecc4a89670 [diff]
[Codegen] Bubble up Transpose attention V and try fuse with others before attention (#19250)

Flash Attention transpose_V variant is significantly faster than the
non-transpose_V variant. This is due to many matmul intrinsics being
mmtb by default. Hence, doing FA transpose_V will allow for better/more
contiguous reads from shared memory to register, improving the attention
performance quite a bit.

This PR exposes the attention_transposeV form by generating a
linalg.transpose on the V during bubbling up of transpose S.T we can
give the graph some opportunities to fuse the transpose-V to it's
producer. I have also confirmed that if we do not find any producer, the
transpose will indeed fuse back with the attenionOp. Hence worse case,
we will get same perf as before this PR.

Additionally, we modify elementwise op fusion to try fuse transpose with
other ops before letting it get fused back into attention.

---------

Signed-off-by: Stanley Winata <stanley.winata@amd.com>
8 files changed
tree: ecc0bb689b69c8571b540fe1f21cfa47dbdb5ff2
  1. .github/
  2. build_tools/
  3. compiler/
  4. docs/
  5. experimental/
  6. integrations/
  7. lib/
  8. llvm-external-projects/
  9. runtime/
  10. samples/
  11. tests/
  12. third_party/
  13. tools/
  14. .bazel_to_cmake.cfg.py
  15. .bazelignore
  16. .bazelrc
  17. .bazelversion
  18. .clang-format
  19. .git-blame-ignore-revs
  20. .gitattributes
  21. .gitignore
  22. .gitmodules
  23. .pre-commit-config.yaml
  24. .yamllint.yml
  25. AUTHORS
  26. BUILD.bazel
  27. CITATION.cff
  28. CMakeLists.txt
  29. configure_bazel.py
  30. CONTRIBUTING.md
  31. LICENSE
  32. MAINTAINERS.md
  33. README.md
  34. RELEASING.md
  35. WORKSPACE
README.md

IREE: Intermediate Representation Execution Environment

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 Discord Status pre-commit OpenSSF Best Practices

Project Status

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

Release status

PackageRelease status
GitHub release (stable)GitHub Release
GitHub release (nightly)GitHub Release
Python iree-base-compilerPyPI version
Python iree-base-runtimePyPI version

Build status

CI PkgCI

Host platformBuild status
LinuxCI - Linux x64 clang
CI - Linux arm64 clang
macOSCI - macOS x64 clang
WindowsCI - Windows x64 MSVC

For the full list of workflows see https://iree.dev/developers/general/github-actions/.

Communication Channels

Related Project Channels

  • MLIR topic within LLVM Discourse: IREE is enabled by and heavily relies on MLIR. IREE sometimes is referred to in certain MLIR discussions. Useful if you are also interested in MLIR evolution.

Architecture Overview

IREE Architecture IREE Architecture

See our website for more information.

Presentations and Talks

Community meeting recordings: IREE YouTube channel

  • 2021-06-09: IREE Runtime Design Tech Talk (recording and slides)
  • 2020-08-20: IREE CodeGen: MLIR Open Design Meeting Presentation (recording and slides)
  • 2020-03-18: Interactive HAL IR Walkthrough (recording)
  • 2020-01-31: End-to-end MLIR Workflow in IREE: MLIR Open Design Meeting Presentation (recording and slides)

License

IREE is licensed under the terms of the Apache 2.0 License with LLVM Exceptions. See LICENSE for more information.