Choose timezone
Your profile timezone:
Madgraph5 GPU development - !!! 4pm CET !!!
→
Europe/Zurich
513/R-070 - Openlab Space (CERN)
Stefan Roiser
# Madgraph dev meeting (Mon 14.11.2022)
Present: SR, ZW, JT, AV, OM, NN, TC
## Presentation by OM on portability frameworks
OM presents the slides attached to the agenda.
He comments on the paper presented by Taylor last week, https://arxiv.org/abs/2203.09945
TC: sherpa (blockgen) has both cuda and kokkos, initially was cuda
SR: one more external group is LHCb HLT, they just have one header file to abstract CUDA and HIP,
they are similar to use in that they use very little functionality
## Presentation by TC on portability frameworks
AV: would not agree that cuda is difficult to learn... code looks much easier than kokkos, especially
Also gives a better idea of what goes on behind the scenes, less of a black box
AV by chat: "thanks Olivier and Taylor for the presentations, sorry I will need to leave soon. Two very quick comments. One, I do not agree that CUDA is difficult to learn: if I look at Taylor's (very nice!) slide, especially kokkos looks very very complex, cuda seems much easier. Two, I would separate the portability of different GPUs and the portability of doing both CPUs and GPUs; in particular I would try to understand which functionality we are taking about. Namely, on CPU we need vectorization and I still do not see this here. But also, thread pools in CPUs and GPUs are quite different, my take is that in the c++ native implementation we have not even really looksed at multithreading, instead on GPUs the threads are something we need all the time."
[AV leaves, end of notes - discussion continues]
There are minutes attached to this event.
Show them.
