MG3MConv: Multi-Grained Matrix-Multiplication-Mapping Convolution Algorithm toward the SW26010 Processor

TL;DR

This paper introduces MG3MConv, a high-performance, adaptable convolution algorithm optimized for the SW26010 processor, achieving significant efficiency improvements over existing GPU-based solutions across various CNN models.

Contribution

Proposes MG3MConv, a novel multi-grained matrix-multiplication-mapping convolution algorithm tailored for SW26010, with architecture-specific optimizations and diversified task mapping schemes.

Findings

Achieves up to 84.78% hardware efficiency on SW26010.

Outperforms cuDNN in most convolution scenarios.

Reaches 67.04% efficiency on VGG, surpassing cuDNN and swDNN.

Abstract

As the core of artificial intelligence applications, the research of convolution has become a hot topic in high performance computing. With the rapid development of the emerging SW26010 processor in artificial intelligence, there is an urgent need for high-performance convolution algorithms on the processor. However, the current support of convolution on SW26010 is still rudimentary. The only studies provide sufficient runtime peak performance but lack the adaptability to various convolution scenes. To perfect convolution algorithms on SW26010, we propose a multi-grained matrix-multiplication-mapping convolution algorithm called MG3MConv, which targets the architectural features of SW26010. MG3MConv supports diversified mapping schemes of convolution tasks based on the concept of the thread block proposed in this paper. All the architecture-oriented optimization methods are elaborately designed from four levels to fully exploit the hardware efficiency of SW26010. The experiments show that the hardware efficiency of MG3MConv can reach 84.78% in max, which is 1.75 times compared with that of cuDNN based on NVIDIA K80m GPU. Moreover, MG3MConv can overperform cuDNN in most convolution scenes. We also use six representative CNNs as real-world cases, and the hardware efficiency of MG3MConv reaches up to 67.04% on the VGG network model, which is 1.37 times and 1.96 times that of cuDNN and swDNN, respectively.