An Hybrid Data Transfer Optimization Technique for GPGPU

Eric Petit (1) , François Bodin (1) , Romain Dolbeau (1)

1 CAPS - Compilation, parallel architectures and system

Graphical Processing Units (GPU) can provide tremendous computing power. Current NVidia and ATI hardware display a peak performance of hundreds of gigaflops. However, because of the data transfer speed between CPU and GPU is limited, those devices are difficult to use to accelerate numerical applications. In this paper we propose a software hybrid technique for automatically optimizing data transfer based on static and dynamic information on data accesses.

Mots clés: Astex; speculative; thread; automatic parallelization; codelet; hmpp; multicore; GPU; SoC; GPGPU; coprocessor; communication optimisation

Domaines: Calcul parallèle, distribué et partagé [cs.DC]

Eric Petit, François Bodin, Romain Dolbeau. An Hybrid Data Transfer Optimization Technique for GPGPU. Compilers for Parallel Computers workshop (CPC2007), 2007, Lisbon, Portugal. ⟨hal-00418717⟩