Madgraph5 GPU development

Name: Madgraph5 GPU development
Start: 2022-10-31T15:00:00+01:00
End: 2022-10-31T16:30:00+01:00
Location: CERN

Monday 31 Oct 2022, 15:00 → 16:30 Europe/Zurich

513/R-070 - Openlab Space (CERN)

Videoconference

Stefan Roiser

stefan.roiser@cern.ch

+41 75 4115334

Hide

Round table

ACAT conference ran last week, so not much work on code. Talks on accelerated event generators by Andrea and Taylor
Nathan: Progress on the bridge code for Kokkos
Carl: Performance test have been set up. Tesla T4 ready for testing.
Request: README files for cuda should be moved into cuda directory as for other abstraction layers
Rui: Joined recently. Goal is to integrate Madgraph & GPUs into Athena (ATLAS)
Olivier: Possibility to consolidate the portability frameworks? (e.g. choose to continue only one of SYCL and Kokkos?)

Taylor Something that other Madgraph developers should be asked about?

Walter Hard to choose the solution that will be adopted in the future.

Taylor Converting between portability frameworks is a bit of work but generally possible.

Action Taylor / Olivier: Comparison (pros/cons) of different abstraction frameworks as base for discussion in future meeting (–> HEPCCE report)

Jorgen: Reviewing CI on github.
- Not all processes are being tested.
- Plan to run automated nightly performance tests in the future.
- Stephan: Reason for partial coverage is turnaround time. Maybe add nightly tests with full coverage? Compromise between faster turnaround for PRs and full coverage.

(See slides)

Numbers presented are for a standalone example
It was ensured that cublas vs. madgraph-style are executing the same number of matrix multiplications
Interesting result: single-precision is 2x faster for madgraph style, but same speed using cublas.
Possibility to manually call tensor cores? Theoretically yes, but very low level.
Note that the madgraph-style example does a factor two more work (not taking into account the symmetric matrix)

There are minutes attached to this event. Show them.