FFTs map very nicely to systolic arrays. There are quite a few papers on this, particularly related to FPGAs. Folks should be using MKL for FFTs, since it should be faster than any other general-purpose implementation out there. If that’s not the case, please let us know. MKL supports the FFTW interface, although this abstraction serves no purpose because MKL already knows the right implementation for each processor, so if you are porting something anyways, you may want to use the MKL FFT interface. Documenting the problem sizes that are important to your applications is the most important thing to do right now. If you are calling FFTs in batches, or the code could be changed to do that, particularly with the same dimensions, that is useful to understand. For example, if you have a plane-wave DFT code that does hundreds of rank-N FFTs (i.e. N is constant), then that means the graph doesn’t need to be reloaded between calls, which will improve the efficiency. For what it’s worth, I created an 3D FFT proxy for MiniDFT 1.1 (does exact-exchange, unlike MiniDFT 1.06) that was small enough to run in Sniper. It’s currently in Fortran but I can convert it for CSA activities. Best, Jeff From: anl-intel-csa <[email protected]> on behalf of "Vazquez Mayagoitia, Alvaro" <[email protected]> Date: Wednesday, June 28, 2017 at 3:08 PM To: "Pope, Adrian C." <[email protected]> Cc: "[email protected]" <[email protected]> Subject: Re: [Anl-intel-csa] 2D FFTs on CSA Hi, Adrian I added fftpack to the git repo which uses r2r transformations , I haven't gone further since the code is pure fortran 77. It would be good to know if a good replacement could be tested in CSA. We use a lot FFTs in periodic quantum chemistry codes. Thanks - Alvaro Vazquez-Mayagoitia, PhD Theoretical and computational chemistry Argonne National Laboratory https://web.alcf.anl.gov/~vama/ On Jun 28, 2017 1:33 PM, "Pope, Adrian C." <[email protected]> wrote: Update: Christopher Hughes put me in contact with Jongsoo Park for FFTs on CSA - we’ve been chatting over email, and I’m looping in Venkat Vishwanath since he may have some APS use cases. If anyone else is particularly interested this topic just let me know and I’ll include you in discussions. Cheers, Adrian

On Jun 23, 2017, at 12:16 PM, Pope, Adrian C. <[email protected]> wrote:

One of the potential LSST app kernels does many 2D R2R FFTs with sizes ~32x32. They pulled out a simplified version of the code for us and it currently uses FFTW3 interface signatures, so that would be easiest, but we could also adapt to a different interface. I think someone on the call said there was some Intel work on 2D FFT on CSA in the graph(500?) project. If so, could someone send a pointer to that code?

There may also be another LSST app kernel variant that focuses on 2D FFTs with sizes more like ~4000x4000. I’m still tracking down the details on that, but it would be good to know the size ranges of any 2D FFT work on CSA so far.

Thanks! Adrian

_______________________________________________ anl-intel-csa mailing list [email protected] https://lists.jlse.anl.gov/mailman/listinfo/anl-intel-csa

_______________________________________________ anl-intel-csa mailing list [email protected] https://lists.jlse.anl.gov/mailman/listinfo/anl-intel-csa