FFT and interpolation with MatLab and CUDA question about expected performance of CUDA vs Matlab


I am currently working on speeding up the post processing of some images and have been looking at CUDA for a while.
The code I have is extensive and is written in Matlab so it will take lots of effort to change it into C for CUDA.
I am getting the following time in seconds in Matlab for execution of critical bottlenecks of the algorithm

FFT: 0.00012299
Spline interpolation: 0.0021951

(for an array of 1024 elements)

Does anybody knows if I should get a considerable improvement if I use CUDA for this operations ? Is the 0.00012299 for FFT a time that can be considerably improved by CUDA?
I would like to know this since I would not like to implement everything in CUDA and not getting noticeable results



I am doing Hilbert transform and the spline interpolation. Are these functions included in the math libraries of CUDA?
Thanks again!

Hi Pablo,

Not sure if you’ve seen Jacket yet, but just in case, I wanted to point you to our website (http://www.accelereyes.com). Jacket handles FFTs and interpolation, so it should be able to accomplish what you need.



Hi John!

Thank for your response. Unfortunately this option was discussed in the group but it was beyond our current budget.

Another question to everyone:

Are there interpolation libraries in CUDA or CUBLAS and CUFFT are the only ones? Is there an efficent way to implement cubic interpolation in CUDA?