LLVM Lands Build System Changes To Make It Easier For Out-Of-Tree Projects To Use MLIR

Since being released by Google engineers last year and subsequently integrated into the LLVM ecosystem, the MLIR intermediate representation has quickly been gaining interest both among LLVM projects and other external users.

MLIR is designed with heterogeneous hardware and machine learning in mind. TensorFlow and others have been re-tooled to support MLIR among other out-of-tree 'users' and more certainly being on the way with this IR designed by Chris Lattner and others.

Thanks to LLVM there has also been efforts like running MLIR over Vulkan/SPIR-V and taking the MLIR path in supporting machine learning over Vulkan, among other interesting use-cases.

In making it easier for out-of-tree projects to employ MLIR, the LLVM code-base has added the CMake bits for the handling of configuration information to an out-of-tree project. The LLVM infrastructure has already been in place for other software projects looking to employ LLVM so that they can be imported as a CMake target. Now similarly the CMake bits are there for MLIR as of this commit.

The adoption and evolution of MLIR will certainly be interesting to watch in 2020. Those not too familiar with MLIR yet can see this Tensorflow blog post with more details back from when it was opened up in April 2019.