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 dev meeting

    Present: SR, OM, AV, NN, WH

    ## Round table

    ### Nathan

    Got frameworks up and running (Cuda, kokkos, sycl, alpaka) on some of the devices.
    Working on getting all those up and running.
    Could show some preview of results next week.

    SR: will Taylor go to Bologna?
    NN: send him a mail, but he will probably not go.
    Not clear if we US can go there...

    ### Walter

    NTR

    ### Olivier

    Gave a talk at a meeting with CMS on April 11.
    Main goal was to warn them in advance of what we do and tell them to be ready.
    (In the past there were often long delays up to two years to pick up new features).
    Their feedback is definitely that they are interested in vectorization.
    They would also like helicity recycling in the vectorization.
    Not so intersted for now in GPUs.
    They seem definitely not intesred in multithreading.
    SR: a bit more optimistic about GPUs maybe...
    OM: yes but they seem intersted for high-RAM CPU processes, but then GPU is even worse.
    AV: also a bit more optimisic on GPUs, maybe other people will push for this.
    SR: maybe we should take the process Sapta is intersted in (V+4jets?) and show them numbers for GPUs?

    With respect to processes interesting to physicists, finally fixed issues with MLM in 'vector' mode.
    All is ok at least in cross section calculation, still need to check unweighted generation.
    Anyway nothing crashes which is already a good point!
    AV: maybe add an MLM process to our list? eg madevent + me export cudacpp for MLM
    OM: yes (eg Sapta V+up to 4 jets is MLM), but do not expect you need anything to change in ME

    ### Stefan

    Contributed unisgned int. Andrea pushed this.

    Alphas - Andrea will discuss this.
    Note that there is no code generation. Happy to do after that.
    AV: yes, note however that would prefer to have codegen before mergin
    SR agrre on that

    ### Andrea

    Show a few slides. Discussion.

    NN: is gpu threads/blocks hardcoded for now? could we make it configurable, #defines, env var or similar?
    AV: was thinking of reading a card file from C++, also to configure GPUs, CPUs, heterogeneous etc.
    NN/all: sounds good.

    ## AOB

    9 May? OM in a workshop the whole day.
    Agreed: next meeting on 16 May!

     

    There are minutes attached to this event. Show them.