Madgraph5 GPU development

513/R-070 - Openlab Space (CERN)

513/R-070 - Openlab Space


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

# Madgraph dev meeting 06 June 2023

Present: SR, AV, SH, JT
Remote: WH, NN, CV
Excused: ZW

## Round table

WH: ntr, wanted to follow up on susy samples but did not have time

NN: updated the sycl branch with the latest changes from cudacpp

JT: been working on HIP macros to add to the cudacpp implementation
Just got access to an AMD GPU and will soon test it

SH: presents some slides about batch size for the GPU
(caveat: this includes the GPU unweighting he developed)
- AV on ggttggg: are you sure the slowdown at 32k is not a fluctuation? SH: unlikely, this is mean of 5 runs
- SR/AV: maybe would be nice to test not just powers of two but also number of SMs times a power of two
- Discussion on stack size. SR: try massif. SH: can move arrays to heap by using allocatable. SV: please try that out! There is a todo on Olivier to try malloc... if you have idea, that would be interesting! This would solve issue of needing MEMAX and USED, it would allow using C++ with 32-size arrays and not 16k-size arrays!

CV: colleague is doing some tests and finding some issues, will report.
As mentioned before, we have some A100s which we can use to do some tests, please suggest a version
SR: the HEAD in github master on epochX is ok

AV: presents some slides
General message is trying to get a full workflow running for some processes like ggtt(g),
keeping performance and other process-specific issues later

SR: discussions with various people

## AOB

AV/SR: we will try to have a workshop with also other generators

Next meeting: 20 June
AV will be in LesHouches, may connect or not

There are minutes attached to this event. Show them.