PIK-Convolution: Step Convolution Acceleration based on Multi-GPU Architecture

SimpleConvolution is the most important and time-consuming part of convolutional neural networks (CNN) for image processing. Each slide of the window in two-dimensional convolution will cause repeated memory access. To solve above problems, we propose an algorithm integrated with the hardware level, parallel in kernel Convolution(PIK-Convolution), which reduces the granularity of a SingleConvolution and avoid the problem of storage access pressure caused by spatial locality. Unlike Image to column(im2col) algorithm, we avoid the problem of non-adjacent data in memory directly from hardware perspective. With MGPUSim, a multi-GPU architecture platform, we implement our algorithm. A large number of experimental results show that the proposed algorithm performs well on the new architecture. Compared with the unoptimized algorithm, the overall efficiency of the new algorithm is optimized by 2.2×-7.45×, and the memory access efficiency is an average of 30× of the unoptimized algorithm. Compared to the state-of-the-art convolution acceleration, our algorithm has a performance optimization of 1.5×-2.2.