Reactive programming in MLIR
While compilation frameworks such as MLIR concentrate the existing know-how in HPC compilation for virtually every execution platform, they lack a key ingredient needed in the high-performance embedded systems of the future: the ability to represent reactive control and real-time aspects of a system. They do not provide first-class representation and reasoning for systems with a cyclic execution model, synchronization with external time references (logical or physical), synchronization with other systems, tasks and I/O with multiple periods and execution modes. In practice, this poses problems when representing TensorFlow control in the tf_executor dialect, when providing streaming implementations for ML applications, when integrating ML components with signal processing pipelines...
We propose two extensions of MLIR dedicated to reactive programming. Following MLIR convention, these extensions come under the form of so-called dialects:
- lus is a high-level reactive programming dialect based on Lustre. Specification of application control at lus dialect level comes with formal correctness checking ensuring the absence of infinite or undefined behaviors. In turn, this ensures that embedded implementation is possible in bounded memory space and in bounded execution time. The lus dialect follows a dataflow programming paradigm allowing the natural specification of ML specifications with complex control and integration with signal processing pipelines.
- sync is a low-level reactive programming dialect. It directly extends the Static Single Assignment (SSA) form which stands at the core of MLIR with reactive primitives allowing synchronization with other functions and with external time references, I/O with multiple periods and execution modes.
Both dialects freely combine with the data processing operations already present in MLIR (in dialects such as tensorflow or linalg), thus allowing joint specification of all aspects of an embedded system - high-performance data processing and interaction with the environment.
Our current effort is focused in two directions:
- Promoting the lus dialect for the specification of high-performance embedded applications featuring ML and signal processing under complex reactive control.
- Improving code generation.
In developing our dialects and their compilation process, we follow the maximal reuse approach encouraged by MLIR (even when this approach comes with a steeper learning curve).
Further reading: H. Pompougnac, U. Beaugnon, A. Cohen, D. Potop - From SSA to Synchronous Concurrency and Back
The repository comprises 3 main folders:
- mlir-lus contains the implementation of the lus and sync dialects, and of the command-line mlirlus* tool allowing compilation of specifications based on these dialects.
- mlir-prime is a tool that exposes existing MLIR code transformations that are not exposed by the command-line transformation tools of MLIR.
- usecases showcases the use of our new dialects on a few signal processing and ML applications.
This software is released under the GNU General Public License, version 2.0 or later, as detailed in the LICENSE file.
The authors of mlirlus are Hugo Pompougnac and Dumitru Potop.