Faster GPU-based convolutional gridding via thread coarsening

TL;DR

This paper introduces a GPU optimization technique called thread coarsening to significantly accelerate convolutional gridding in interferometric imaging, achieving up to 3.2x speedup on certain hardware.

Contribution

It applies thread coarsening to existing GPU algorithms, substantially improving their efficiency for convolutional gridding tasks.

Findings

Up to 3.2x performance gain on GTX 980

Up to 1.9x performance gain on GTX 980 for quad-polarization

Significant gains on Radeon R9 290X

Abstract

Convolutional gridding is a processor-intensive step in interferometric imaging. While it is possible to use graphics processing units (GPUs) to accelerate this operation, existing methods use only a fraction of the available flops. We apply thread coarsening to improve the efficiency of an existing algorithm, and observe performance gains of up to $3.2\times$ for single-polarization gridding and $1.9\times$ for quad-polarization gridding on a GeForce GTX 980, and smaller but still significant gains on a Radeon R9 290X.