GPU Performance Optimization Targeting OpenCL Model

Graphic processing units attract more and more general-purpose computing due to high performance/cost ratio.In order to fully exploit the capability of GPU for general-purpose computing under heterogeneous processing platforms,this paper proposes performance optimization methods targeting OpenCL model.Polyhedral representation of a source program is built to optimize and allocate GPU memory system.By checking memory access patterns of the source program,access instances those can be grouped together are discovered by means of graph coloring.Subsequently,data space transformation is utilized to alter irregular memory access patterns for the sake of improving the off-chip memory bandwidth by taking advantage of vector data types.Meanwhile,data reuse information is detected to allocate data into distinct fast memory regions according to both the properties of data accesses and the characteristics of the OpenCL memory model,with the purpose of making best usage of the fast on-chip memory.Experimental results on benchmarks showed that the optimized programs achieved a speedup of 1.6X~8.4X in comparison with the un-optimized versions,demonstrated the effectiveness of the proposed methods.