Superlinear Speedup on GPGPU Using Laplacian Algorithm with Convolution Filtering as A Case Study

Abstract

In this paper, the main idea is to investigate the hypothesis that superlinear speedup occurs when the concurrent threads on General Purposes Graphic Processing Units (GPGPU) carry heavy workloads. In order to evaluate this hypothesis, Laplacian image edge detection algorithm with convolution filtering is chosen as a case study. In this work, local memories of GPGPU are utilized in order to achieve the superlinear speedup. The convolution filtering kernels of the Laplacian edge detection algorithm are invoked in these local memories. By this, the low latency of the GPGPGU local memory are deployed efficiently and this subsequently leads to a higher speedup. The results obtained presented that the superlinear speedup is achieved when the size of the convolution kernel is large. In this study, when the convolution kernel size is 7×7, superlinear speedup is observed for image dataset of sizes between 1KB-2500KB.