Towards a Multi-device Version of the HYFMGPU Algorithm for Hyperspectral Scenes Registration

Hyperspectral image registration is a relevant task for real-time applications like environmental disasters management or search and rescue scenarios. Traditional algo-rithms were not devoted to real-time performance, the HYFMGPU algorithm having arisen as a solution to such a lack. Sensors are expected to evolve and thus generate images with finer resolutions and wider wavelength ranges, so a multi-GPU implementa-tion seems to be necessary in a near future. This work presents a first approach to such a multi-device version, identifying some stages of the pipeline as the most suitable to run in parallel in several GPUs. An MPI+CUDA variation of the original HYFMGPU algorithm is implemented, achieving speedups of 1.83× in 2 GPUs and 3.08× in 4 GPUs for the stages of the pipeline distributed among several devices. Different issues related to communications-derived time overloads and to some CUDA-based libraries particularities, as long as some optimization possibilities out of the currently distributed stages, were also detected. We plan to tackle them in further development stages of this multi-GPU implementation.

keywords: Hyperspectral imaging, image registration, Fourier transforms, multi-GPU, CUDA, OpenMP, remote sensing.