GPU accelerated scalable parallel coordinates plots

Abstract

Parallel coordinates are a powerful technique to visually analyze multi-parameter data, i.e., sets of datapoints with potentially many associated parameter values per datapoint. When these sets are large, line rendering becomes a severe performance bottleneck, and since many lines fall into the same pixel the numerical precision of the color buffer is quickly reached. We propose a scalable GPU realization of parallel coordinates building upon 2D pairwise attribute bins, to significantly reduce the number of lines to be rendered. Our approach comprises a GPU compute pipeline that combines shader-based scattering with atomic increment operations to efficiently count how often a line is drawn. These counts are then used to draw all pairwise sub-plots in the parallel coordinates plot, by analytically calculating the opacity for each count and rendering a line with end points determined by the 2D coordinates of the bin. In this way, framebuffer precision issues that are paramount in classical approaches can be overcome. We demonstrate the efficiency of the proposed realization for visualizing a weather forecast ensemble comprising 2.7 billion datapoints, each carrying 7 prognostic floating-point variables like temperature, precipitation and pressure, plus spatial and simulation input variables. We compare our pipeline to a rasterization-based approach regarding performance, and demonstrate interactive brushing at 4 s per frame at full HD viewport resolution.