An Analysis of GPU-accelerated Algorithms for Visualizing Geological Data with Various Spatial Scales: Case Studies from Field Measurements in Taiwan
  Kuang-Yi Chen     Yu-Chi Lai     Ping-Yu Chang     Seng-Rong Song  
Kuang-Yi Chen, Yu-Chi Lai, Ping-Yu Chang,Seng-Rong Song, An Analysis of GPU-accelerated Algorithms for Visualizing Geological Data with Various Spatial Scales: Case Studies from Field Measurements in Taiwan, Springer2013 SUBMITTED
Abstract

Spatial interpolations and extrapolations are crucial tasks for presenting data from various geophysical surveys. However, spatial modeling algorithms such as Kriging are time-consuming. Careful observation can reveal that such spatial modeling computation concerning each grid point not only is independent from the computation in other grid points but also can be processed in parallel without a ecting the result. Therefore, several works are proposed to use a graphics processing unit (GPU) to accelerate the computation of inverse distance (ID) and Kriging (KR) but their methods does not use the full computation power of GPU
and there is still room for improvement. This study focuses on the usage of the shared memory in GPU to fully accelerate the parallel computation of spatial
modeling methods. Proper usage of shared memory is critical for GPU parallelization. A GPU-based kD-tree is also implemented for parallelizing the computation of nearest neighbor (NN) and limited inverse distance (LID). All four algorithms including KR, ID, NN, and LID are applied to three real geospatial data sets in Taiwan to compare the performance between the GPU and CPU implementations. These data sets are in two formats: electrical impedance and temperature with di erent spatial scales. The spatial extent of surveying areas ranges from 0.02 km2 to 36,000 km2. Our results show that GPU acceleration can be achieved under these different spatial scales. Furthermore, the performance and interpolated results of four algorithms are compared to analyze the applied conditions of these algorithms. We demonstrate that GPU can mortise the time of data transfer and reach its full parallelization power when the grid size grows. At the end, the comparison shows that the GPU can greatly improve the performance from 50 times faster for nearest neighbor to 400 times faster for Kriging. The computation of Kriging can take advantage of all existing cores to reach the fully parallelization power of a GPU. Our future studies will focus on the use of GPUs in geophysical inversion tasks.

Bibtex

xxx

PDF