Hello Sebastian, I do not really see any problems in implementing this method using current version of pyfft. You just need to pad your image according to the required behavior at the borders and pass it to the FFT plan. The only difference with CUFFT is that pyfft requires 2^n image sizes, while CUFFT has more freedom in this respect (although the paper still recommends using multiples of 512, so the difference is not that big).
Best regards, Bogdan On Sun, Jan 8, 2012 at 9:25 AM, Sebastian Nowozin <[email protected]> wrote: > Dear all, > > nvidia has a great description on how 2D FFT can be used for 2D > convolutions with different boundary handlings, at: > http://developer.download.nvidia.com/compute/cuda/2_2/sdk/website/projects/convolutionFFT2D/doc/convolutionFFT2D.pdf > > I wondered whether anyone has used pyopencl/pyfft in a similar manner, > i.e. implementing 2D convolutions with some common boundary handling > (such as clamp-to-border) or whether such functionality is planned for > pyfft in the future? > > Thanks, > Sebastian > > _______________________________________________ > PyOpenCL mailing list > [email protected] > http://lists.tiker.net/listinfo/pyopencl _______________________________________________ PyOpenCL mailing list [email protected] http://lists.tiker.net/listinfo/pyopencl
