Acceleration of Recursive Cross-Correlation PIV Using Multiple GPUs

A large number of PIV algorithms and systems have been proposed, many of which are highly sophisticated in terms of accuracy and spatial and temporal resolution. However, a general problem with PIV is the time cost to compute vector fields from images, which often imposes specific constraints on the measurement methods. In this paper, focusing on recursive direct cross-correlation PIV with window deformation, which is one of the most popular algorithms for PIV, we propose a technique to accelerate PIV processing using a single Graphics Processing Unit (single-GPU) and multiple GPUs (multi-GPU). In the case of using single-GPU, we show that PIV data can be processed over 100 times faster than using a CPU alone and about 30 PIV image pairs per second can be processed for certain image sizes. The scalability of the algorithm used is also discussed. In the case of using multi-GPU, the image split method and the parallel method of single-GPU-based PIV processing are measured. We show that the effect of multi-GPU can be observed over a certain amount of image data whether either of these two methods is used. Data transfer between CPU and GPUs is shown to be a bottleneck if the number of GPUs used increase.