On Fri, Nov 30, 2018 at 1:39 AM Ralf Gommers <ralf.gommers@gmail.com> wrote:On Sun, Nov 25, 2018 at 5:22 AM Sourav Singh <ssouravsingh12@gmail.com> wrote:Hello,On Sat, Nov 24, 2018 at 11:26 AM Ralf Gommers <ralf.gommers@gmail.com> wrote:Hi Sourav,On Mon, Nov 19, 2018 at 5:47 AM Sourav Singh <ssouravsingh12@gmail.com> wrote:Hello,I am writing the mail regarding project proposal for GSoC 2019.I am interested in working with SciPy for the upcoming Google Summer of Code and I am interested in either working on enhancing the randomized numerical linear algebra functionality or fixing the scipy.fftpack module.I understand that it is quite early to talk, since the organization proposals haven't started yet, but I would like to know if a project proposal can be started right now so that the proposals can be better fleshed out by the time the application period starts?That's the earliest start I've seen (which is a good thing)!I suspect that we'll have trouble finding a mentor for the randomized linear algebra, however fftpack could be a good topic. Have you seen https://github.com/numpy/numpy/pull/11888? I suspect that by the time GSoC starts, we'll have merged that into NumPy, and would like to take it over for SciPy. A related topic would be to add a backend system for fft implementations (pyfftw, mkl-fft).And a third topic would be to make the APIs of numpy.fft and scipy.fftpack agree. That would be the lowest on my list because it's hard (requiring lots of discussion on impact on users if we deprecate things etc.), but it's an important one as well.I am interested in adding pocketfft to scipy, which is linked in the numpy PR. Would it be possible to combine this with the backend system project for scipy?Yes, this would be possible - that certainly should fit in a single GSoC. It's also possible that Pocketfft will already be merged by then, given that the author of Pocketfft is quite proactive. The NumPy PR is only waiting for the branching of the 1.16.x release, after which it can in principle be merged (C99 is the blocker now). In that case, implementing a backend system + working on matching signatures for numpy and scipy would still be a nice project.Cheers,RalfI would be happy to help mentor an FFT-related project and think having a backend system would be very nice.
pyFFTW's scipy interfaces are mostly complete, aside from some real-valued transforms for which work was started in:There were also GPU-based FFTs matching the scipy API proposed recently in https://github.com/cupy/cupy/pull/1745