On the GPU Performance of 3D Stencil Computations Implemented in OpenCL.

Aiming at a close examination of the OpenCL performance myth, we study in this paper OpenCL implementations of several representative 3D stencil computations. It is found that typical optimization techniques such as array padding, plane sweeping and chunking give similar performance boosts to the OpenCL implementations, as those obtained in corresponding CUDA programs. The key to good performance lies in maximizing the use of on-chip resources of a GPU, same for both OpenCL and CUDA programming. In most cases, the achieved FLOPS rates on NVIDIA's Fermi and Kepler GPUs are fully comparable between the two programming alternatives. For four typical 3D stencil computations, the performance of the OpenCL implementations is on average 9% and 2% faster than that of the CUDA counterparts on GTX590 and Tesla K20, respectively. At the moment, the only clear advantage of CUDA programming for stencil computations arises from CUDA's ability of using the read-only data cache on NVIDIA's Kepler GPUs. The skepticism about OpenCL's GPU performance thus seems unjustified for 3D stencil computations.