Analytical Characterization and Design Space Exploration for Optimization of CNNs

TL;DR

This paper introduces an analytical modeling approach to optimize loop-level configurations for CNNs on multi-core CPUs, significantly reducing data movement and improving performance over existing auto-tuning methods.

Contribution

It presents a novel analytical method for efficiently exploring the CNN optimization space, outperforming or matching state-of-the-art auto-tuning techniques.

Findings

Achieves comparable or better performance than existing libraries.

Reduces data movement through optimized loop transformations.

Provides a scalable approach for CNN optimization on multi-core CPUs.

Abstract

Moving data through the memory hierarchy is a fundamental bottleneck that can limit the performance of core algorithms of machine learning, such as convolutional neural networks (CNNs). Loop-level optimization, including loop tiling and loop permutation, are fundamental transformations to reduce data movement. However, the search space for finding the best loop-level optimization configuration is explosively large. This paper develops an analytical modeling approach for finding the best loop-level optimization configuration for CNNs on multi-core CPUs. Experimental evaluation shows that this approach achieves comparable or better performance than state-of-the-art libraries and auto-tuning based optimizers for CNNs.