high performance computing on graphics processing units: hgpu.org

Graphic processing Units (GPUs) are gaining ground in high-performance computing. CUDA (an extension to C) is most widely used parallel programming framework for general purpose GPU computations. However, the task of writing optimized CUDA program is complex even for experts. We present a method for restructuring loops into an optimized CUDA kernels based on a 3-step algorithm which are loop tiling, coalesced memory access, and maximizing machine utilization. For this we identify the GPU constraints for maximum performance such that the memory usage (global memory and shared memory), number of blocks, and number of threads per block. In addition we identify the condition for maximizing utilization of the GPU resources. We also establish the relationships between the influencing parameters and propose a method for finding possible tiling solutions with coalesced memory access that best meets the identified constraints. We also present a simplified algorithm for restructuring loops and rewrite them as an efficient CUDA Kernel. The execution model of synthesized kernel consists of uniformly distributing the kernel threads to keep all cores busy while transferring a tailored data locality which is accessed using coalesced pattern to amortize the long latency of the secondary memory. In the evaluation, we implement some simple applications using the proposed restructuring strategy and evaluate the performance in terms of execution time and GPU throughput.