Compute Distance Matrices with GPU

DOI: 10.5176/2251-1652_ADPC12.07

Authors: Seongho Kim and Ming Ouyang

Abstract:

Given a data matrix where the rows are objects and the columns are variables, researchers often want to compute all the pairwise distances among the objects. Dueto the design of Nvidia GPU architecture, CUDA code can work with ease data matrices where the numbers of rows and columns are multiples of sixteen. The present work proposes apadding strategy that add additional rows and columns of zeros to the matrix so that a matrix of any size may be processed by as imple and fast CUDA kernel function. For Pearson correlation coefficient, the GPU computation 15.9 to 33.5 times faster than the CPU.

Keywords: Pairwise distance, GPU, CUDA, divergent execution,Pearson correlation coefficient