Memory Saving Discrete Fourier Transform on GPUs

This paper will show an alternative method to compute the two-dimensional Discrete Fourier Transform. While current GPU Fourier transform libraries need a large buffer for storing intermediate results, our method can compute the same output with far less memory. This will function by exploiting the separability of the Fourier transform. Using this scheme, it is possible to transform rows and columns independently. As multiple lines can be transformed at once, the available memory on the device can be used to reduce the number of necessary kernel calls drastically. We will also prove that our approach can compete with the timings of the two-dimensional transform provided by NVIDIAs CUFFT library but consumes at the same time far less memory by enabling the transformation of much bigger data sets on the device.