1# MLIR 2 3## Overview 4 5MLIR, or Multi-Level Intermediate Representation, is a representation format 6and library of compiler utilities that sits between the model representation 7and low-level compilers/executors that generate hardware-specific code. 8 9MLIR is, at its heart, a flexible infrastructure for modern optimizing 10compilers. This means it consists of a specification for intermediate 11representations (IR) and a code toolkit to perform transformations on that 12representation. (In compiler parlance, as you move from higher-level 13representations to lower-level representations, these transformations can be 14called “lowerings”) 15 16MLIR is highly influenced by [LLVM](https://llvm.org/) and unabashedly reuses 17many great ideas from it. It has a flexible type system, and allows 18representing, analyzing and transforming graphs combining multiple levels of 19abstraction in the same compilation unit. These abstractions include TensorFlow 20operations, nested polyhedral loop regions, and even LLVM instructions and fixed 21hardware operations and types. 22 23We expect MLIR to be of interest to many groups, including: 24 25* Compiler researchers and implementers looking to optimize performance and 26 memory consumption of machine learning models 27* Hardware makers looking for a way to connect their hardware to TensorFlow, 28 such as TPUs, portable neural hardware in phones, and other custom ASICs 29* People writing language bindings that want to take advantage of optimizing 30 compilers and hardware acceleration. 31 32The TensorFlow ecosystem contains a number of compilers and optimizers that 33operate at multiple levels of the software and hardware stack. We expect the 34gradual adoption of MLIR to simplify every aspect of this stack. 35 36<img alt="MLIR overview diagram" src="./images/mlir-infra.svg"/> 37
