The cuFFTAdvisor is a tool for predicting the optimal settings of the cuFFT library. It is usable for programs that can, within a defined tolerance, change the size of a signal transformed to or from Fourier space to improve FT or inverse FT performance. cuFFTAdvisor implements a heuristic to find data sizes with which cuFFT can perform well. Optionally, it seeks the best data sizes empirically, using autotuning.

The cuFFTAdvisor was developed for CUDA 8.0 and Pascal GPU architecture. Since then, several new GPU architectures and CUDA releases have appeared. The thesis aims to check whether the heuristic is still working well. The student checks whether the latest cuFFT using modern hardware has a similar dependence between signal size and performance as is with CUDA 8.0. If needed, they update the heuristic to work well with the latest cuFFT. The second goal of the thesis is the adaptation of cuFFTAdvisor for 2D signals: the student observes how the performance is related to the size of the signal and adapts the heuristics to work with 2D signals.