Automatic generation of BLAS kernels with MLIR

• Grant number: 25/14803-5
• Support Opportunities: Scholarships in Brazil - Master
• Start date: February 01, 2026
• End date: July 31, 2027
• Field of knowledge: Physical Sciences and Mathematics - Computer Science - Computer Systems
• Principal Investigator: Guido Costa Souza de Araújo
• Grantee: Luc Joffily Ribas
• Host Institution: Instituto de Computação (IC). Universidade Estadual de Campinas (UNICAMP). Campinas , SP, Brazil

Abstract

Scientific computing and machine learning applications are highly dependent on basic linear algebra operations. The optimization of these functions is a widely studied topic, requiring programs capable of making the most of the underlying hardware capabilities. In this context, researchers have developed specialized, architecture-dependent libraries such as OpenBLAS and BLIS.The proposal of this project is to generate these high-performance codes in a compiled manner, using hardware information at compile-time to adapt the program to the processor. This is expected to reduce the high cost and development time of these libraries. In this project, we intend to develop algorithms for the automatic code generation of these libraries' kernels, based on architectural parameters such as the memory hierarchy model, cache sizes, ISA, among others. For this, we will use Multi-Level Intermediate Representation (MLIR), a compilation framework capable of maintaining high-level abstractions of the program, which allows for more precise optimizations of kernel operations like matrix multiplication. (AU)