Focus on:  Hide Contributions
Login

Madgraph5 GPU development

→ Europe/Zurich
513/1-024 (CERN)

513/1-024

CERN

50
Show room on map
Madgraph5 GPU development
Zoom Meeting ID
63816708295
Host
Stefan Roiser
Useful links
Join via phone
Zoom URL
Hide

Attendees: Olivier, Stefan, Daniele, Carl, Ameya

Apologies: Andrea

 

Daniele
—

  • Update on DY+4j (see attached link)
  • cross sections numbers are different than from Jin
    • for DY+3 maybe a mistake with the order 
    • for DY+4 very different, this needs to be understood
  • Olivier, maybe related to cuts? 
  • Also a difference with sde strategies for the fortran code needs to be understood. 
  • Olivier, try to use fixed scale computations (in the run card)
  • Note that in cudacpp (even fortran) we currently use sde_strategy = 1


Olivier
—

  • Jin has reported floating point division by 0, being investigated by Olivier and Andrea
  • Try to fix ttbar CI issue, think there is no issue by itself. In fortran there are sometimes two resets for the helicities. In Fortran reimplemented 1 reset. 
  • The threshold for good helicities can be set now in the run card. 
  • Release schedule by the end of the month is realistic
    • Merge master_june24 into master (and in gpcupp)
    • Understand better Jin report (divide by 0)
  • Need to fix upstream madgraph for usage with the plugin code. We can release without and fix this at a later stage if absolutely needed.
  • Still need to run a script to extract the plugin code into a separate repository. 
  • Upstream madgraph 3.6.0 is ready for release, but not the interaction witth the plugin (see above). 


Carl & Ameya
— 

Discussing a possible project for Ameya,

Ameya, have time now until end of August, when the next semester starts less time for contribution. 

  • Move phase space generation onto GPU, Olivier will send a paper about the strategy to Ameya and pointers in the Fortran code. Olivier and Ameya to followup.
  • Other possibilities discussed:

  • Olivier presents slides about differences in the directory structure of fortran and cudacpp. Even in fortran we use different files for matrix elements where the only difference are the couplings used. The ida being to combine those and reduce the number of files, functions, … 

  • Discussion about Madnis, current developers are moving around. Contribution can also be discussed. 

     

Aob

---

Next Meeting 27 Aug

There are minutes attached to this event. Show them.
Powered by Indico v3.3.6-pre