采用GPU的提升纹理缓存命中光线投射方法

Improving Texture Cache-Hit Rate of GPU-Based Ray Casting

提出一种改善纹理缓存命中率的方法.首先,分析图形处理器(GPU)中三维纹理组织的布局特性;进而提出根据视点的变化动态选择线程配置的策略,目的在于最小化warp级的投射光线纹理访存跨距;最后,算法用CUDA(compute unified device architecture)实现并验证.实验结果表明:当视点分别围绕x,y,z坐标轴旋转时,改进后算法的帧速率分别为改进前的1.08,1.14,0.98倍.

This paper presents a method of improving the texture cache hitrate for GPU-based volume rendering.Firstly,we analyze the data layout of 3Dtexture in GPU.Based on it,a dynamic strategy of selecting the thread block shape according to the viewpoint is proposed.The strategy can minimize the access stride for the warp-level threads.Finally,we realize the method in CUDA（compute unified device architecture）and testify the effectiveness.The experimental results show that when the viewpoint rotates around the x-,y-,z-axis,the frame rates are 1.08,1.14 and 0.98 time faster than that of static thread block shape configuration,respectively.