High Performance Code Generation in MLIR: An Early Case Study with GEMM

TL;DR

This paper presents an early case study demonstrating how MLIR, a new compiler infrastructure, can generate high-performance code for GEMM, comparable to manual libraries, with improved modularity and automation.

Contribution

It introduces MLIR's capabilities for high-performance code generation, focusing on memrefs, affine dialect, and polyhedral utilities, highlighting its potential for modular and automated code production.

Findings

MLIR can generate competitive GEMM code.

MLIR's infrastructure supports modular and reusable code.

Potential for automating high-performance code generation.

Abstract

This article is primarily meant to present an early case study on using MLIR, a new compiler intermediate representation infrastructure, for high-performance code generation. Aspects of MLIR covered in particular include memrefs, the affine dialect, and polyhedral utilities and pass infrastructure surrounding those. This article is also aimed at showing the role compiler infrastructure could play in generating code that is competitive with highly tuned manually developed libraries, albeit in a more modular, reusable, and automatable way.