CUDA下地质图像边缘检测算法并行优化

摘要

为提高地质图像边缘检测Prewitt算法计算速度, 结合算法计算密集和数据密集的特点, 从核函数计算层面, 提出基于调整线程块坐标优化线程发散方法和采用本地变量访存优化指令延迟设计思想;从CPU-GPU数据传输开销层面, 提出基于CUDA流优化数据传输开销方法.经测试, 当设置线程块规模为32*32、采用独立的局部变量替代索引访存和使用CUDA工作流分块计算时, 对大于6168*6168尺寸的地质图像加速比可提高120倍以上.该并行优化方案易于实现, 可应用于大规模地质图像边缘检测领域.%To improve the computational speed of the Prewitt algorithm for edge detection of geological images, the computationally intensive and data-intensive features of the algorithm were combined.From the aspect of kernel function calculation, a method of optimizing thread divergence based on adjusting thread block coordinates and adopting local variables was proposed to optimize instruction latency.From the aspect of CPU-GPU data transmission overhead, a method based on CUDA stream to optimize data transmission overhead was proposed.Testing results show that, when setting the thread block size to 32*32, using independent local variables instead of indexed memory access, and using CUDA workflow block calculations, the acceleration ratio for geological images larger than 6168*6168 can be increased by more than 120 times.This parallel optimization scheme is easy to implement and can be applied in the field of large-scale geological image edge detection.