Go to the Indico Home Page Oldest event Older event Up to category Newer event Newest event Printable version
 
Focus on:  Hide Contributions
  • Compact style
  • Indico style
  • Indico style - inline minutes
  • Indico style - numbered
  • Indico style - numbered + minutes
  • Indico Weeks View
    • Switch to the management area of this event
    Login

    Madgraph5 GPU development

    → Europe/Zurich
    Virtual (Zoom)

    Virtual (Zoom)
    Stefan Roiser
    Hide

    # Madgraph meeting Monday 07.02.2022

    Present: SR, LF, OM, CV, NN, SH, WH, AV (notes)

    ## News

    SR: got an allocation of A100 at Julich via PRACE thanks to openlab. It went very fast!

    ## Round table

    LF: discussed sycl work last week about paper

    OM: not much to report, a few discussions with Stefan and Andrea
    - One point we discussed quite a bit is the issue of alphas
    The most important point is to have a version of the bridge, even without running of alphas
    Discussd also with the other authors, they are ok to start without running of alphas
    SR: started looking at several options, discussed with AV, many options discussed on github
    - OM: I would keep calculating the scale on the Fortran side, because there are many if statements,
    it is a part of the code that was written by another author
    SR: good then if it is CPU, whether fortran or c++ it is the same
    - AV: concerned about getting same pohysics calculations.
    If we can do both Fortran and CUDA/C++ with fixed alphas to compare results, that's good too.

    CV: ntr. At the last meeting there was some discussion about documentations, we there progress?
    SR: there are some readmes here and there, maybe LF/NN?
    LF: for sycl added a readme
    AV: for cuda/c++ also there, Nathan managed to use it I think
    https://github.com/madgraph5/madgraph4gpu/wiki/Building-madgraph4gpu-and-measuring-throughput
    NN: yes confirm the wiki was ok
    SR to CV: would be good if you try them out

    NN: have been working on benchmarking together with Taylor
    NN: also tried to port to some of the latest code by Andrea, one issue is sycl has no constant memory
    AV: intersting! two points, you do not need to reproduce all memory access, as you have something
    equiv alent in kokkos/sycl; two, more interesting it is worth pointing this out in the paper
    (even if I am not sure this will mak ea big difference)
    NN: presently have 15% slower than Andrea, trying to understand what to do
    NN: also looked at the latest Bridge

    SH: one, did some work on MAc for fortran in the CI, will iterate with Andrea
    SH: two, not happy about performance of CI container,will move to virtual GPUs for that
    SR: can you move to more recent cudas on vGPUs?
    SH: most recent is 11.4, might be update later on
    AV: I think cuda 11.4 should be enough, I just need c++17 support in nvcc and it should be ok in 11.4

    WH: looked at templates that produced vectorized code, just to see what would be needed for kokkos
    Have seen Andrea has done some changes recently but have not moved to that yet
    Looked at Bridge and started thinking of ways to compare different implementations,
    maybe this could also be interesting for a CI (compare LHE file output using a python script).
    AV: by th eway with MadEvent we have the same random numbers, whether MEs are fortran or cuda,
    so we can get the exact same output in the two workflows!
    OM: yes then we can decide to move stuff later

    SR: one, some work on shared librarues, picked up by Andrea in his PRs
    SR: two, looking at alphas as just discussed
    SR: three, we have been allocated some time on a Julich A100 cluster with interactive use,
    if you are interested let me know if you want access

    AV: present slides attached to the agenda
    Discussion of several points on the slides

    ## AOB

    Next meeting in two weeks

     

    There are minutes attached to this event. Show them.