Auto-Tuning Compiler Optimizations beyond Numerical Parameters

Number 108058

Program optimization is a challenging task due to growing demands in modern computing and a rapidly evolving hardware environment. This often leads to highly optimized programs that lack performance portability across different classes of devices, manufacturers, and hardware generations.

Auto-tuning has provided a pathway to automate this complex process by optimizing configuration options automatically based on runtime feedback. However, current auto-tuning techniques are limited in their expressiveness, focusing mainly on numerical parameters and failing to address more fundamental optimization aspects, such as code transformations or optimization schedules.

In this project-seminar, we aim to investigate and develop methods that extend the capabilities of auto-tuning beyond its current limitations. To achieve this, we will design a versatile interface for specifying diverse optimization spaces, develop efficient methods for managing these spaces, and create robust strategies for their effective exploration. We will evaluate the efficacy of our approach through the application to state-of-the-art compiler frameworks, such as Halide, TVM, or MLIR.

---

 

Organizer

People:  Johannes Lenfers  Bastian Köpcke