OpenCL FFT for GPUs FFT issues in OpenCL


I was looking for OpenCL FFT implementation. I found two 1.) the clFFT from Apple and 2.) the AMD OpenCL FFT. Since I want to use it on nVidia GPUs
and the later one requrires AMD SDK to be present at compile time I’ve choosen the first one. The problem is that it is optimized to G80
architecture and for long FFT base (N > 2**17 ) I experience buggy behaviour for Inverse FFT when using interleaved complex representation.
Performing an FFT on simle sine wave and immediately Inverse FFT-ing it back, gives a very noisy sine with some random breaks in it.
However, for plannar representation it seems working fine. Has anybody else experienced similar behaviour ? (I need interleaved rep., to be
compliant with fftw and for other reasons.)


You can try out ArrayFire OpenCL library (free). Here is the documentation link: