This document provides an overview of IREE's systems, including entry points to get started exploring IREE's capabilities.
For information on how to set up a development environment, see Getting Started on Windows and Getting Started on Linux.
- Core IREE project
- Integrations between IREE and other frameworks, such as TensorFlow
- Language and platform bindings, such as Python
- Colab notebooks for interactively using IREE's Python bindings
- Common types and utilities used throughout IREE
- IREE's MLIR dialects, LLVM compiler passes, module translation code, etc.
- Hardware Abstraction Layer for IREE's runtime, with implementations for hardware and software backends
- Shared data storage format definitions, primarily using FlatBuffers
- Assorted tools used to optimize, translate, and evaluate IREE
- Bytecode Virtual Machine used to work with IREE modules and invoke IREE functions
IREE ingests MLIR in a high-level dialect like XLA/HLO, after which it can perform its own compiler passes to eventually translate the IR into an 'IREE module', which can be executed via IREE's runtime. IREE contains programs for running each step in that pipeline under various configurations (e.g. for tests, with a debugger attached, etc.).
The iree-opt
program invokes
MlirOptMain
to run some set of IREE's optimization passes on a provided .mlir input file.
Test .mlir files that are checked in typically include a RUN
block at the top
of the file that specifies which passes should be performed and if FileCheck
should be used to test the generated output.
For example, to run some passes on the reshape.mlir test file with Bazel on Linux, use this command:
$ bazel run //iree/tools:iree-opt -- \
-split-input-file \
-iree-index-computation \
-simplify-spirv-affine-exprs=false \
-convert-iree-to-spirv \
-verify-diagnostics \
$PWD/iree/compiler/Translation/SPIRV/XLAToSPIRV/test/reshape.mlir
The iree-translate
program translates from a .mlir input file into an IREE
module.
For example, to translate simple.mlir
to an IREE module with bazel on Linux,
use this command:
$ bazel run //iree/tools:iree-translate -- \
-iree-mlir-to-vm-bytecode-module \
--iree-hal-target-backends=interpreter-bytecode \
$PWD/iree/tools/test/simple.mlir \
-o /tmp/module.fb
Custom translations may also be layered on top of iree-translate
- see
iree/samples/custom_modules/dialect
for a sample.
The iree-run-module
program takes an already translated IREE module as input
and executes an exported main function using the provided inputs.
This program can be used in sequence with iree-translate
to translate a .mlir
file to an IREE module and then execute it. Here is an example command that
executes the simple module.fb
compiled from simple.mlir
above on the IREE
interpreter
$ bazel run //iree/tools:iree-run-module -- \
--input_file=/tmp/module.fb \
--driver=interpreter \
--entry_function=abs \
--inputs="i32=-2"
The iree-run-mlir
program takes a .mlir file as input, translates it to an
IREE bytecode module, and executes the module.
It is designed for testing and debugging, not production use cases, and therefore does some additional work that usually must be explicit, like marking every function as exported by default and running all of them.
For example, to execute the contents of a test .mlir file, use this command:
$ bazel run //iree/tools:iree-run-mlir -- \
$PWD/iree/tools/test/simple.mlir \
--input-value="i32=-2" \
--iree-hal-target-backends=interpreter-bytecode
The iree-dump-module
program prints the contents of an IREE module FlatBuffer
file.
For example:
$ bazel run //iree/tools:iree-dump-module -- /tmp/module.fb