Abstract: Tensor computations have become a mainstay in numerical fields and in data science. They constitute a critical class of applications that include domains such as computational chemistry, material sciences, quantum mechanics and machine learning. Their use in distributed-memory, high-performance computing clusters with multi-node GPUs impose fundamental scientific and engineering challenges that require rethinking, redesigning and jointly optimizing tensor operators. In this keynote I will introduce and discuss the DiMage family of mappers, their design, implementation, semantics and various results. DiMage mappers represent tensor networks as constraint problems solved with the Z3 SMT solver. Unlike prior works, DiMage mappers leverage the producer-consumer relations among operators in the network to decide the optimal placement of both data tensor and compute tensor operators. In addition, this family of mappers automatically decide the grid shape of processing elements (PEs), compute- and data- mappings, at the granularity of a dimension, while generating the MPI/NCCL communication. Future directions and remaining challenges will be discussed.

Biography: Dr. Martin Kong is an Assistant Professor in the Department of Computer Science and Engineering (CSE) at The Ohio State University. His research interests include high-performance computing, correctness and validation of scientific programs in parallel and distributed programming models, quantum computing, and polyhedral compilation. He has served as General Chair of the ACM SIGPLAN Compiler Construction (CC) 2026 Conference, and is an Associated Editor in ACM Transactions on Architecture and Code Optimization (TACO). Dr. Kong has served in the Program Committee of several premier conferences such as SC, CGO, IPDPS, PPoPP, among others.