Workshops Archive

Linalg vs MDH: Why Reduction Semantics Matter


Workshop: LLVM-HPC2026: The Twelfth Workshop on the LLVM Compiler Infrastructure in HPC

Authors: Richard Schulze, Jens Hunloh, and Lars Hunloh (University of Muenster); Tobias Grosser (University of Cambridge); and Ari Rasch (University of Muenster)

Abstract: Data-parallel computations are performance-critical building blocks in domains like scientific computing and machine learning. In the MLIR ecosystem, the Linalg dialect—in particular linalg.generic—is a central abstraction for such computations, yet it does not treat reduction operators as first-class, compiler-visible constructs.

To address this gap, we propose the mdh dialect: it makes reduction (combine) operators explicit—including user-defined operators beyond standard pointwise cases—thereby enabling reduction-aware code generation and richer expressivity for reduction-based computations. Our experiments on GPU and CPU—across linear algebra, deep learning, data mining, and quantum chemistry workloads—isolate the effect of explicit reduction operators, estimating the relative performance opportunity if linalg.generic exposed such operators.


Back to LLVM-HPC2026: The Twelfth Workshop on the LLVM Compiler Infrastructure in HPC Archive Listing

Back to Full Workshop Archive Listing