high performance computing on graphics processing units: hgpu.org

In this paper, we present an acceleration method for texture-based ray casting on the compute unified device architecture (CUDA) compatible graphics processing unit (GPU). Since ray casting is a memory-intensive application, our method increases the hit rate of the texture cache during rendering. To achieve this, our method dynamically selects the width and height of thread blocks (TBs) such that each warp, which is a series of 32 threads simultaneously processed on the GPU, can achieve high data locality for specific viewpoints. The objective of this selection is to allow every warp rather than every thread to access data with a small stride, because the GPU executes multiple threads at the same time. In experiments using a GeForce GTX 480 card (i.e., the latest Fermi architecture), we find that the speedup of our method ranges from a factor of 1.0 to that of 4.0, depending on viewpoints. We think that optimizing the shape of TBs is important to achieve more cache hits in the highly-threaded CUDA hardware.