A GPGPU Programming Framework based on a Shared-Memory Model

Kazuhiko Ohno, Dai Michiura, Masaki Matsumoto, Takahiro Sasaki, and Toshio Kondo


parallel programming language, compiler, GPU, CUDA, virtual shared memory


Although general purpose computation on GPU (GPGPU) seems promising method for high performance computing, current programming frameworks such as CUDA and OpenCL are much difficult to achieve high performance. Therefore, we propose a new framework for easier GPGPU programming. Our framework provides shared variables which can be accessed from both CPU and GPU. Our compiler translates user’s shared-memory based program into CUDA program automatically generating memory allocation and data transfer code. The compiler also overlaps kernel executions and data transfers by optimizing scheduling. The evaluation results show that programs in our framework can match for hand-optimized CUDA programs, automatically generating optimized data transfer code.

Important Links:

Go Back