Madgraph5 GPU development

Europe/Zurich
513/R-070 - Openlab Space (CERN)

513/R-070 - Openlab Space

CERN

15
Show room on map
Stefan Roiser

# Madgraph4gpu dev meeting Tue 25.04.2023

https://indico.cern.ch/event/1275610/

 

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

Excused: SH

 

## Discussion on licensing

 

AV presents a proposal for copyright and licensing.

General agreement, go ahead and merge it.

We can rediscuss this later on if needed.

Let's focus on technical stuff and development.

TC: especially let's focus on completing functionality for the release, rather than on further speed improvements

 

OM: technically best is to create two repos under mg5amcnlo,

and have the plugins there for sycl and cudacpp

 

## Short round table

 

ZW: poster at CHEP was upgraded to a talk!

 

OM: new phase space integrator based on machine learning

Very good results for the moment, factor 5-10 better in unweighting efficiency 

(here 5-10 for W->du, but also same range for other processes).

Here GPUs will be very useful for training.

 

TC: ntr

 

CV: ntr

 

NN: ntr

 

JT: containers for both cudacpp and sycl are ready!

https://gitlab.cern.ch/AIGROUP-it-sc-accelerators/madgraph4gpu-ci/container_registry

(code is here https://gitlab.cern.ch/AIGROUP-it-sc-accelerators/madgraph4gpu-ci) 

Can download and build everything both on cudacpp and sycl.

Using the llvm of sycl, not oneapi for licensing reasons.

AV: great! maybe NN can cross check that you get the same results,

since this may be a different sycl version

 

AV: nta - will merge the PR after the meeting!

 

SR: in contact with nvidia to include our code as a benchmark in their suite

 

## AOB

 

SR: next meeting? agreement on May 2nd.

 

There are minutes attached to this event. Show them.