Accelerating Finite Impulse Response Filtering Using Tensor Cores

This paper studies how to accelerate a single channel 2D image convolution using NVIDIA's Tensor Core. Tensor Core is a dedicated arithmetic unit for speeding up matrix products and was proposed to speed up the convolution in machine learning. From the Volta architecture of NVIDIA's GPU, Tensor Core is utilized for various applications. The Tensor Core is not limited to machine learning but various applications. This paper focus acceleration of image convolution of a single-channel filter of finite impulse response (FIR) filtering. Usually, Tensor Core is used for multi-channel convolution; thus, direct usage of Tensor Core cannot maximize the performance of the usual FIR filtering. Therefore, we propose three programming patterns for parallel processing of FIR filtering: kernel-loop unrolling, kernel-image-loop unrolling, and diagonal kernel-image-loop unrolling. Our experiments compare the proposed loop unrolling methods for linear time-invariant (LTI) and variant filters (LTV): Gaussian filtering and bilateral filtering. The results show that Tensor Core can accelerate the convolution for LTI filters but cannot accelerate linear LTV filters. Also, we compare the two GPU architectures of Turing and Ampere, and we can accelerate the LTI filter using TCs in both architectures.