SUPPORTING DYNAMIC DATA STRUCTURES IN A SHARED-MEMORY BASED GPGPU PROGRAMMING FRAMEWORK

摘要

Although Graphics Processing Unit (GPU) is expected tornbe a practical high performance computing platform, currentrnprogramming frameworks such as CUDA and OpenCLrnrequire large programming cost. Therefore, we are developingrna new framework MESI-CUDA providing sharedrnvariables to hide low-level data management in CUDA.rnHowever, handling dynamic data structures is difficult inrncurrent MESI-CUDA because shared variables cannot berndynamically created and pointer fields are not allowed inrnthem. Thus, we extended MESI-CUDA to remove suchrnrestrictions. Introducing dynamic management of sharedrnvariables and automatic pointer conversion on data transfer,rnany pointer-based dynamic data structure can be shared betweenrnthe CPU and GPU with only small changes from thernC code. As the results of the evaluation, pointer conversionrnincreased the transfer time of data structures approximatelyrn3.3 times larger in the worst case, and 1.3–2 times larger inrnthe practical cases. Considering that non-conversion alternativesrncause overhead in pointer dereferences, we regardrnthis overhead is practical in most cases.