Madgraph5 GPU development

Monday 23 Jan 2023, 15:00 → 16:30 Europe/Zurich

513/R-070 - Openlab Space (CERN)

# Madgraph4gpu dev meeting Mon 23.01.2023

Present: SR, SH, ZW, JT, AV, OM, TC, NN, CV

## Round table

CV: ntr, ready to do some performance tests if you have stuff

TC: ntr

NN: working on sycl branch with performance plots with sycl vectorization
Added color selection and helicity selection, but getting some strange results in tmad scripts, 
different cross sections in fortran and sycl, but debug -2 seems to give the same MEs
AV: there is some cleanup to do on the upstream mg5amcnlo, suggest to look at it again when done
NN: sounds good

OM: discussed with AV last week on upstream mg5amcnlo
Had tried to use lo_vect_diag branch, but this was a failure
because it did not meet its goal of making some files smaller,
and also introduced other new issues.
In the end decided to give up on lo_vect_diag.
Learnt some useful things on madgraph on the way however, some cleanup to do!

AV: did some cleanup in collaboration with OM.
Had initially agreed to base my MRs on his lo_vect_diag,
then last week we agreed to drop that.
This week will resum work in that scenario, without lo_vect_diag.
AV: will give a talk at the Compute Accelerator Forum in 2.5 weeks.

JT: expanded the grafan scripts.
JT shows some plots. On GPUs, CUDACPP and SYCL are comparable, 
with SYCL a bit better for simple processes and cuda better for complex processes.

ZW: set up a prototype to use the bridge to reweight existing LHE events at LO.
Rewrote an LHE parser entirely in C++, as the boost parser has issues with large files (slow and large RAM).
Will then work on a specific example from CMS.
OM: in the discussion we also discussed some new ways to do reweighting,
wrote a couple of pages and will circulate them

SH: having a look at on-GPU maximum weight computation,
will discuss offline with OM this week to have a chat
OM: very good, ping me on skype

SR: busy writing proposals unrelated to madgraph
Related to madgraph, one project CYCLOPS is related to RISC-V processors
which could be very interesting for extreme vectorization
AV: more details in a EuroHPC plenary from ACAT
TC: it would be interesting if madgraph could get on one of the three exascale machines
(it was mentioned in the ACAT plenary that there would be those three machines on an aggressive timescale)
SR: we have access to LUMI through openlab
OM: have access to LUMI via Belgium
TC: with NN we are in contact with Aurora and Frontier
AV: we should also not forget ARM and Power CPUs, not just new GPUs

## AOB

OM: one interesting process for two subprocesses is 
 Define p = u d
 Generate q q > w+ w- q q

SR: next meeting Tue Feb 7 (OM not available Mon Feb 6)

1 News

2 Topical discussion

3 Round table

4 AoB