摘要： 有限差分本来是个十分简单的操作, 而3D的有限差分操作的应用领域十分的广阔, 尤其是在工程模拟领域中用到的特别多, 在CUDA上进行有限差分操作, 由于要进行大量的数据访问, 如果不进行任何访问优化的话, 内存将会达到6x甚至25x的内存访问带宽, 这里提出了一种优化方法, 把内存访问的带宽控制在了2x到3x数据量的范围内. 3D Finite Difference operation is used in variety Engineering fields, however such an operation in the cpu can be sometimes time consuming, while a naive implementation in GPU requires a lot access with the global memory, which can up to 6x or even 25x the data size, here several methods are discussed to solve this problem. 阅读全文