Engineers engaged on Google’s TensorFlow machine studying framework have revealed a subproject, MLIR, that’s supposed to be a standard intermediate language for machine studying frameworks.
MLIR, quick for Multi-Level Intermediate Representation, will enable initiatives utilizing TensorFlow and different machine studying libraries to be compiled to extra environment friendly code that takes most benefit of underlying {hardware}. What’s extra, MLIR might in time be utilized by compilers usually, extending its optimization advantages past machine studying initiatives.
MLIR isn’t a language like C++ or Python. It represents an intermediate compilation step between these higher-level languages and machine code. The compiler framework LLVM makes use of an intermediate illustration, or IR, of its personal. One of LLVM’s originators, Chris Lattner, is a co-creator of MLIR. Making MLIR an LLVM co-project may very well be a strategy to unfold its adoption.
In a slide presentation on the EuroLLVM convention earlier this month, Lattner and fellow Googler Tatiana Shpeisman defined how TensorFlow already generates a number of IRs internally, however that these disparate IRs don’t profit from each other. MLIR gives a single, normal IR for all of these TensorFlow subsystems. TensorFlow is presently migrating to make use of MLIR internally.
Another profit MLIR might present is parallelized compilation. MLIR is designed to permit a compiler to work on completely different segments of code in parallel, permitting machine studying fashions—and different kinds of purposes—to be pushed to manufacturing extra shortly.
MLIR might present different advantages to languages and frameworks exterior machine studying. For instance, LLVM-based languages like Swift and Rust have needed to develop their very own inside IRs, as a result of many optimizations utilized in these languages can’t be expressed in LLVM. MLIR might present a regular strategy to specific these optimizations, which might in flip be re-used for different languages.
The MLIR challenge is open supply. An official specification is obtainable for individuals who need to generate MLIR.
