Using Accelerator Libraries for Science
Abstract:
Accelerator systems are superior in performance and energy efficiency when executing dedicated algorithms and, therefore, to play an important role in many scientific fields. However, designing these special algorithms often requires expert hardware and software knowledge.
This course presents accelerator libraries which hide the algorithmic complexities behind interfaces, so only C, C++ or FORTRAN knowledge is required to run them on accelerator systems. In this way many scientific investigations can be accelerated without specific programming knowledge. Moreover, the library developers ensure compatibility and improved performance on new accelerator generations, so that the application code that uses the libraries does not have to be modified again. The focus lies on GPU libraries, as they cover the widest range of scientific tools for accelerators