De/Re-Composition of Data-Parallel Computations via Multi-Dimensional Homomorphism - PLDI 2024

Explore a groundbreaking approach to optimizing data-parallel computations in this 18-minute conference talk from PLDI 2024. Dive into the world of Multi-Dimensional Homomorphisms (MDHs) and learn how they can be used to systematically decompose and recompose computations for efficient execution on modern architectures. Discover how this method applies to a wide range of data-parallel computations, including linear algebra routines, stencil computations, and deep learning algorithms. Understand the power of this approach in expressing various state-of-the-art strategies and its potential for automatic optimization through auto-tuning. Gain insights into how this technique achieves higher performance than vendor-provided solutions on real-world datasets across diverse computational domains.

[PLDI24] [TOPLAS] (De/Re)-Composition of Data-Parallel Computations via Multi-Dimensional(…)