Workshop: LLVM-HPC2026: The Twelfth Workshop on the LLVM Compiler Infrastructure in HPC
Authors: Addison Alvey-Blanco (Sandia National Laboratories, University of Illinois at Urbana-Champaign) and Brian Kelley and Sivasankaran Rajamanickam (Sandia National Laboratories)
Abstract: We present a platform-agnostic transformation pipeline that blends kernel fusion and algebraic optimization. Our approach begins with tensor-form linear algebra operations in a Multi-Level Intermediate Representation (MLIR) module and automatically discovers compatible producer-consumer regions. The operations in each region are represented and composed in Einstein notation. We use exact operand-subset dynamic programming to select a binary contraction order, lexicographically minimizing scalar multiplication work followed by logical intermediate volume. The selected contractions are reconstructed in MLIR, and an automatic partitioner selects connected fusion groups according to removable logical storage, parallel-domain compatibility, and reduction topology. After bufferization, the selected groups are outlined, their concrete memory dependences are checked, compatible parallel loops are fused, and eligible intermediates are forwarded as scalars. We implement the complete transformation in Linear Algebra Performance through Intermediate Subprograms (LAPIS), an MLIR-to-Kokkos compiler framework, so one optimized module can target backends supported by Kokkos execution spaces.
Back to LLVM-HPC2026: The Twelfth Workshop on the LLVM Compiler Infrastructure in HPC Archive Listing
Back to Full Workshop Archive Listing