Fast convolution kernels on pascal GPU with high memory efficiency

TL;DR

This paper introduces two high-performance, memory-efficient convolution kernels for Pascal GPUs that significantly outperform Nvidia's cuDNN library, especially in single-channel scenarios.

Contribution

The paper presents novel convolution kernels optimized for Pascal GPUs that improve performance and memory efficiency over existing libraries.

Findings

2.6X performance improvement for single-channel convolution

1.4X performance improvement for multi-channel convolution

Effective hiding of global memory access delay

Abstract

The convolution computation is widely used in many fields, especially in CNNs. Because of the rapid growth of the training data in CNNs, GPUs have been used for the acceleration, and memory-efficient algorithms are focused because of thier high performance. In this paper, we propose two convolution kernels for single-channel convolution and multi-channel convolution respectively. Our two methods achieve high performance by hiding the access delay of the global memory efficiently, and achieving high ratio of floating point Fused Multiply-Add operations per fetched data from the global memory. In comparison to the latest Cudnn library developed by Nvidia aimed to accelerate the deep-learning computation, the average performance improvement by our research is 2.6X for the single-channel, and 1.4X for the multi-channel.