What would be the “GPU version” of the numpy.fft package in python?

Question

I want to compute the fft of a big signal (big sample size) at a shorter span of time (thus, GPU). I have tried the following modules:

• PyFFT - does not support 2D transforms and non powers of 2
• gpyfft - transform size is also not arbitrary (powers of 2, 3, 5)
• scikits.cuda - only returns N/2+1 coefficients. would be problematic when taking transforms of complex inputs
• numpy.fft - arbitrary inputs, complete transform, supports 2d transforms

What would be the python wrapped "GPU version" of the numpy.fft package? (Optional: Can it be CUDA based?)

Answer 1

If you want CUDA based library for computing FFT, where the transform size is arbitrary, and support 1D, 2D, and 3D FFTs; then you may need to have a look at cuFFT. http://docs.nvidia.com/cuda/cufft/#axzz410MycIFQ
Note: this is cuda based library, not python.