The increasing diversity of heterogeneous architectures presents a major compiler challenge, with current toolchains often being rigid, vendor-specific, and difficult to extend. To address this, we are developing the OFA-Compiler (One For All) ecosystem, a suite of modular MLIR-based tools for heterogeneous targets. This presentation introduces OFA-Flex, the component of this ecosystem designed as a design-space exploration (DSE) framework.

 

The core of OFA-Flex is a powerful pattern that decomposes the compilation process into three distinct and modular stages: scheduling, resource allocation, and placement. These stages are implemented ashigh-level algorithms operating on MLIR Linalg-on-tensors dialect. To ensure modularity, a custom dialect is used to communicate decisions between stages, allowing them to operate independently yet cohesively. The OFA-Flex key innovation is its "pluggable" nature: users can select from a library of provided algorithms for each stage or seamlessly integrate new, custom ones.

 

The execution plan generated by OFA-Flex is then handed off to OFA-Backend. This second component of our ecosystem generates high-performance code by lowering the abstract operations and linking them to a library of optimized hardware kernels

Dr. Omar Ragheb is a Research Manager at the advanced computing R&D centre in Fujitsu Canada. His work focuses on creating compiler technologies for heterogeneous computing systems. He is a lead contributor to the One for All(OFA)-Compiler project, which exploresflexible, MLIR-based compilation strategies. Prior to his role at Fujitsu, he gained industry experience at LegUp Computing (acquired by Microchip) and through internships at AMD and the RIKEN Center for Computational Science. Dr. Ragheb earned his PhD from the University of Toronto under the supervision of Professor Jason Anderson."

 

Stephen Wicklund is a Researcher at the advanced computing R&D centre in Fujitsu Canada. He earned his Master's of Applied Science from the University of Toronto under the supervision of Professor Jason Anderson. His master's research focused on MLIR-based compiler optimizations for coarse-grained reconfigurable architectures (CGRAs). His current research focuses on MLIR-based compilers for heterogeneous architectures, with an emphasis on compiler infrastructure and optimization techniques for emerging systems. He contributes to the OFA-Compiler project, where he works on compiler optimizations and intermediate representations for heterogeneous platforms.