面向GPU平台的二维FFT的加速技术研究  

Research on the Acceleration Technology of 2D FFT Targeting GPU Platform

作　　者：陈博伦 何卫锋[1] CHEN Bo-lun;HE Wei-feng(Department of Microelectronics and Nanoscience,Shanghai Jiaotong University,Shanghai 200240)

机构地区：[1]上海交通大学微纳电子学系,上海200240

出　　处：《现代计算机》2020年第12期68-72,共5页Modern Computer

摘　　要：定量分析一维FFT的分解基数选取、旋转因子计算、倒序排列等各个阶段在GPU上并行执行时的特征,并提出以蝶形算子访存跨度为依据的共享内存访问机制,解决共享内存访问效率过低的问题。在此基础上,提出批量列处理机制,解决二维FFT列变换时的全局内存访问不连续问题。实验结果表明,在图像尺寸从1024×1024到4096×4096像素的情况下,加速后的二维FFT程序的性能较CUFFT库函数提升5%-13%。Quantitatively analyze the characteristics of the 1D FFT's radix selection,rotation factor calculation,and reverse order execution on the GPU in parallel,and propose a shared memory access mechanism based on butterfly operator fetch span to solve the problem of inefficient shared memory access.On this basis,a batch column processing mechanism is proposed to solve the problem of discontinuous global memo⁃ry access during 2D FFT’s column transformation.The experimental results show that the performance of the 2D FFT program designed is improved by 5%-13%compared with the CUFFT library function when the image size is from 1024×1024 to 4096×4096 pixels.

关 键 词：FFT GPU 并行计算 性能分析 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]