IREE Compiler

This directory contains IREE's main MLIR-based compiler.

Paths of particular interest include:

└── compiler/
    ├── setup.py     (builds/installs the compiler Python API)
    ├── API/         (C and Python APIs)
    ├── Bindings/    (used to generate different ABI bindings)
    ├── Codegen/     (device code generation for assorted APIs)
    ├── Dialect/
    │   ├── Flow/    (tensor program modeling and compute workload partitioning)
    │   ├── HAL/     (Hardware Abstraction Layer for buffer and execution management)
    │   │   └── Target/
    │   │       ├── LLVM/
    │   │       ├── VMVX/
    │   │       ├── VulkanSPIRV/
    │   │       └── etc.
    │   ├── Stream/  (device placement and asynchronous scheduling)
    │   ├── Util/    (common types across other IREE dialects)
    │   └── VM/      (abstract Virtual Machine)
    ├── InputConversion/  (conversions from frontend/input dialects)
    └── Translation/      (translation pipeline definitions)

Noteworthy compiler components not included here include:

├── integrations/
│   └── tensorflow/
│       └── iree_tf_compiler/  (passes and tools for importing TensorFlow programs)
└── llvm-external-projects/
    ├── iree-compiler-api/     (C and Python APIs for the compiler)
    └── iree-dialects/         (public IREE dialects for other projects to target)

The general flow of data from a frontend framework down to a compiled program is:

program written in framework

    |
    |  import tool
    V

imported .mlir file

    |
    |  main IREE compiler
    V

compiled program

where import tools live outside of the compiler/ directory (and possibly outside of the IREE repo itself).

Refer to IREE's presentations and talks and this architecture diagram for details on how the pieces fit together:

Coding Style

Like the rest of the project, the compiler/ directory uses clang-format with the BasedOnStyle: Google option. However, since the code in this directory makes heavy use of LLVM and MLIR, it also adheres to LLVM style for variable naming (use int variableName instead of int variable_name and pointer alignment (use int *a instead of int* a).

Read more:

• https://google.github.io/styleguide/cppguide.html
• https://llvm.org/docs/CodingStandards.html
• https://clang.llvm.org/docs/ClangFormatStyleOptions.html

Standalone Python Builds

The included setup.py file can be used to build Python binaries or directly install the IREE compiler API. Do note that the compiler is quite heavy and unless you are developing it and on a significant machine, you will want to use released binaries.

There are two ways to build/install Python packages:

• Directly from the source tree (this is how official releases are done).
• From the build directory while developing.

It is recommended to use your favorite method for managing virtual environemnts instead of modifying the system installation.

Only relatively recent versions of pip are supported. Always use the latest via pip install --upgrade pip.

You can build either from the source or build tree (assumes that CMake has been configured and the project built). The latter is typically used by project developers who are already setup for development and want to incrementally generate Python packages without rebuilding.

To build a wheel that can be installed on the same Python version and OS:

python -m pip wheel iree/compiler

To directly install:

python -m pip install iree/compiler

In order to sanity check once the package is installed:

python iree/compiler/Api/python/test/transforms/ireec/compile_sample_module.py