Bonelli, Federico Edoardo
A framework to implement a multifrontal scheme on GPU architectures with OpenCL
Thursday 31st March 2011
In this work we analyze an open-source multifrontal solver implementation (UMF-PACK) and modify it to transfer the computation load on an OpenCL device, typically a GPU. To achieve this result the dbOpenCL library has been created, which allows a neat integration of OpenCL code into existent C or C++ code. An analysis and proling of both the original and the modied UMFPACK implementation is given while solving an example problem built with the LifeV finite element library.