Toward Efficient Permutation for Hierarchical N:M Sparsity on GPUs

TL;DR

This paper introduces a hierarchical N:M sparsity technique for neural networks, along with a specialized channel permutation method and GPU kernel, to improve accuracy and efficiency in sparse model compression on GPUs.

Contribution

It proposes gyro-permutation, a novel channel permutation strategy tailored for hierarchical N:M sparsity, and develops a GPU kernel to support layer permutation during sparse network execution.

Findings

Gyro-permutation significantly improves HiNM sparsity accuracy.

The GPU kernel enables efficient layer permutation during inference.

Hierarchical N:M sparsity achieves accuracy comparable to unstructured sparsity.

Abstract

N:M sparsity pruning is a powerful technique for compressing deep neural networks, utilizing NVIDIA's Sparse Tensor Core technology. This method benefits from hardware support for sparse indexing, enabling the adoption of fine-grained sparsity to maintain model accuracy while minimizing the overhead typically associated with irregular data access. Although restricted to a fixed level of sparsity due to its reliance on hardware, N:M sparsity can be combined with coarser sparsity techniques to achieve diverse compression ratios. Initially, column-wise vector sparsity is applied to a dense model, followed by row-wise N:M sparsity on the preserved column vectors. We call this multi-level approach as hierarchical N:M (HiNM) sparsity. Similar to earlier single-level sparsity techniques, HiNM sparsity necessitates an effective channel permutation strategy to maximize the accuracy of the compressed networks. However, it introduces further complexities by requiring the rearrangement of both input and output channels, addressing challenges such as permutation sequence, HiNM-sparsity-aware permutation, and maintaining consistency in channel ordering across layers. In this paper, we introduce a channel permutation method designed specifically for HiNM sparsity, named gyro-permutation. This method is crafted to exploit the unique characteristics of HiNM pruning, incorporating a strategic policy in each permutation phase, including channel sampling, clustering, and assignment, to circumvent local minima. Additionally, we have developed a GPU kernel that facilitates independent layer permutation during the execution of HiNM sparse networks. Our extensive experimental evaluations on various DNN models demonstrate that our gyro-permutation significantly enhances the accuracy of HiNM sparse networks, allowing them to reach performance levels comparable to those of unstructured sparse networks.