ab-Stream: A Framework for programming Many-core
State Key Laboratory of High Performance Computing, National University of Defense Technology, Changsha 410073, China
Przeglad Elektrotechniczny (Electrical Review), R. 88 NR 7b, 2012
@article{gan2012ab,
title={ab-Stream: A Framework for programming Many-core},
author={GAN, X. and WANG, Z. and SHEN, L. and ZHU, Q.},
year={2012}
}
The common approach to program many-core processor is to write processor-specific code with low level APIs for different processors, which could achieve good performance but would result in serious portability issues: programmers are required to write a specific version code for target architecture. Therefore, we present ab-Stream, an extensible framework for programming many-threaded processor based on SUIF Intermediate Representation. ab-Stream abstracts many-core many-threaded processor into a unified architecture and ab-Stream program is an OpenMP-like program with different directives for many-core processor. Furthermore, a prototype of ab-Stream was implemented to map ab-Stream programs into many-core GPU. Experiments show that our implementation can execute transformed code correctly and efficiently on CUDA-enabled GPUs. Furthermore, performance of ab-Stream version code produced by our prototype can outperform original GPU code and is close to handoptimized GPU code.
July 15, 2012 by hgpu
