CMS Mg5amc@NLO integration

Tuesday 10 Sept 2024, 16:00 → 17:00 Europe/Zurich

513/1-024 (CERN)

# Madgraph GPU meeting with CMS (Tue 10.09.2024)
https://indico.cern.ch/event/1373476/

Present: OM, ZW, SR, AV (notes), Jin, Sapta

## Intro

Sapta will move to ATLAS beginning of January 2025.

SR: main Sherpa/Pepper developer Enrico Bothmann will join CERN IT in Jan 2025.

## Jin

Jin shows some slides.

Gridpack production.

Slide 3. Now see clear improvement in DY+3j.
But fortran/cpp gridpack production still running for DY+4j.
==> AV: should try to produce gridpack with cuda and then run evt gen in cpp or fortran (but still WIP)

Slide 4. Clear improvements in tt+2j and tt+3j.

Event generation.

Slide 6. No improvement from AVX2 in DY event generation
AV: DY+3j should be better? JC: no also in DY+3j event generation faster in Fortran?
==> AV: very strange, should follow up. Gridpack production faster for DY+3j and not evt gen?

Slide 9. Inclusive vs exclusive.
ZW: note that in CUDA the ME is almost negligible, while it is dominant in fortran/cpp for many jets.

Slide 10. Discussion.

About EvtGen for 17 Sep
==> AV: for evt gen, would test again DY+3j, maybe even for 17 Sep

About DY+4j
==> AV: for DY+4j, would push for multi-backend option
OM: yes this is an option... note that a compile script in gridpacks exist,
CMS cannot use it because they drop the source code
AV: here you do not need the source code because you just change a symlink

Sapta
Various talks upcoming
One talk on Friday Sep 13, past of software days, specifically on MadgraphGPU and CDR,
this is a talk by the generator conveners
Next week Jin will give a talk on Friday 20 Sep

Presentation
Sapta: focus on important plots?
==> AV: yes for instance have the bottom right plot of slide 8 to become the main one.
Then it would be nice to have a similar plot right bottom of slide 6, but need DY up to 3jets.

Sapta: focus on event generation?
SR yes this is where we spend money
AV yes agree, but also mention grdpack creation if logistics is easier

Jin: how to treat process grouping?
AV: maybe compare fortran vecsize1 with process grouping and withpout
OM: yes

Jin: should test DY+3j inclusive?
SR/AV/OM: yes

Jin: when multibackend?
AV: will try Thu/Fri, tomorrow I am off
[After the meeting AV gave some instructions in https://github.com/madgraph5/madgraph4gpu/pull/948#issuecomment-2342332159]

16:00 → 16:50 Discussion

Speakers: Jin Choi (Seoul National University (KR)), Saptaparna Bhattacharya (Wayne State University (US))

mg4gpu.status.240910.pdf