Distributed heterogeneous computing integrates paradigms like distributed, SIMD, MIMD, vector, and scalar architectures to efficiently handle diverse workloads, but coordinating execution, memory, synchronization, and task distribution remains complex without a unified model. We introduce a generic MLIR dialect that offers a unified intermediate representation for diverse hardware and programming models. This dialect enables interoperability, supports both fine- and coarse-grained parallelism, and automates task scheduling and optimization. It can be lowered to LLVM IR or other existing standard MLIR dialects through a set of transformations, enabling execution across heterogeneous systems.

Robert K Samuel is a Master’s student at the Indian Institute of Technology Madras, where he conducts research in the PACE Lab under the guidance of Prof. Rupesh Nasre. His work focuses on distributed heterogeneous computing, developing abstractions that simplify the complexity of programming across diverse and distributed hardware systems.