Implementation Of Two-dimensional DFT Parallel Algorithm On CUDA

Fourier transform is a kind of important algorithm in the field of digital signal processing, It play an important role in the several image processing such as image enhancement, Image denoising, image feature detection and image compression etc. Its features of linearity and symmetry properties earn it someplace in the field of digital image processing. One of the qualities of digital image processing is its large amount of digital processing and more time cost while the processing is on, which asks a high requirement of the storage space and processing spee especially in the situation which requests high on the timeliness. The integrity of the real-time image processing system depends on the complexity of the arithmetic and the technology of the software and hardware. The theme of this thesis is how to implement the high-efficiency without decreasing arithmetic.In recent years, GPU have been developing fastly and the processing power of single precision floating point and the external memory bandwidth of the mainstream GPU are much higher than the same period of the CPU. In such situation, it is going to a kind of a trend to carry out the Large-scale parallel computing by CPU. These charactristics make it a new computing resource in the digital image, which is useful for bringing out a two-dimensional Fourier transform algorithm that is suitable for the parallel-computing model of CUDA to implement its high- efficiency which is studied.This innovation from the digital image to start a large-scale parallel computing, parallel design of two-dimensional Fourier transform algorithm, the class with the CUDA C language and run. The image data is divided into different blocks (Block), between each Block is independent of two-dimensional Fourier transform, when all of the data processing is completed, copy the image data from GPU into memory.The main works are as followings:Firstly, DFT parallel algorithm is designed; Secondly, a good operating environment is established for CUDA, in selected experimental data, the reference to the practical application of the digital image is taken to be matrix; Finally, the experimental data and comparing the CPU to run the data prove that CUDA parallel computing on large-scale takes advantage, and a two-dimensional discrete Fourier transform proposed in this thesis can achieve high efficiency.