Choose timezone
Your profile timezone:
Madgraph5 GPU development
→
Europe/Zurich
513/1-024 (CERN)
Videoconference
Madgraph5 GPU development
- Zoom Meeting ID
- 63816708295
- Host
- Stefan Roiser
- Useful links
- Join via phone
- Zoom URL
-
# Madgraph on GPU dev meeting (Tue 3 Sep 2024)
https://indico.cern.ch/event/1355160/
Present: OM, SR, AV (notes), DM, ZW, CV
Excused: Amaya
## Andrea
AV shows some slides. Discussion.
SR: about build times, was thinking of doing some tests splitting the different objects.
AV: please go ahead, start from my WIP PR if you find it useful
AV: note that now it's a single .o file with all templates precompiled (in HELINL=L).
AV: in any case I suspect that the build time is slow in compiling CppProcess.o (or in the link?), not in HelAmps.o?
AV: will note down that I should check the build time of CPPProcess.o with respect to HelAmps.o. <===
AV: note a technicality, the HelAmps.o is recompiled in each P* because memory access changes.
Instead FFV.o in fortran are compiled only once in src/lib.
OM: note that also Fortran has many files that are rebuilt in every P* even if not strictly needed.
For instance the genps.o is rebuilt every time (instead FFV is built only once).
## Daniele
DM shows some slides. Discussion.
AV: worrying that Jin's fortran number is different from all four numbers in DM's table.
Maybe this is simply because Jin was using an old version of the mg5amc+cudacpp code?
DM: yes this could be relevant.
OM/DM: discussion on details in the matrix1_optim.f code.
Maybe som eissue in the MULTICHANNEL and CONFSUB handling?
DM: will investigate.
## Olivier
Worked with Andrea on the merges he described.
Gives a few more details on the gpucpp_for360 from AV's slide.
One "fun" fact is a bug report about changing the reference frame for calculating MEs.
In some cases eg related to polarization the Lorentz invariance is broken.
There is some Lorenz boost back and forth, but numerical precision means the event gets an E-16 non-zero pT.
This changes the cross section by 10%, which should not happen...
## Zenny
Wrapped up the functionality for reweighting. The C++ works.
Have a pipeline for getting a reweighted
AV: can you write many weights?
ZW: yes
AV: do you actually recompute the refence ME and take a ratio?
ZW: yes absolutely, need to recompute the ME for the reference value.
SR: have you tried on Robert's use case?
ZW: not yet but will try
## Carl
CV: semester is starting for AT so he needs to teach and will have a bit less time
SR: is this slot inconvenient?
CV: will cross check, but maybe the schedule is not known yet
OM: this week cannot have a meeting with AT, but next meeting I could
SR/AV: let us know when you meet just in case
## Stefan
SR: played with the new CI of Olivier, it looks very nice
Only one issue is that the order of execution of pipelines changes.
A bit inconvenient as it appears differently
OM: I can order them alphabetically
SR: or I can do that
AV: do you have everything you need in master?
SR: am using PR #981 by OM
AV: this is now closed as it is a duplicate of 882 which was merged today (with the june24 stuff), please check if this is enough
SR: ok will check
## AOB
DM: will be at the CSC the next two weeks
AV: on holiday the next 7 days, will connect to CMS on Tue but moved it to 4pm (thanks for the understanding)
SR: we have opened a new position and we hired Enrico Bothmann who will join in Jan 2025
SR: next meeting will be Tue 17 Sep
CV: will be at CERN but unfortunately at another meeting at the same time
There are minutes attached to this event.
Show them.
