Madgraph5 GPU development

Monday 21 Feb 2022, 15:00 → 16:00 Europe/Zurich

Virtual (Zoom)

# Madgraph meeting Mon 21.02.2022

Present: AndreaV, CarlV, DavidS, StephanH, StefanR, OlivierM, NathanN, WalterH

## Round table

### StefanR

Started looking at cmake configuration to use xcode config files on Mac.

### AndreaV

Slow progress in the last couple of weeks due to many other issues.

Thanks Stefan for the Juwels access!
One good news: tests on Juwels Cluster login node got ~x8 and ~x16 speedups from AVX512/zmm!
This is on some Intel Gold processors.

For the A100 on Juwels Booster added the correct arch sm_80.
AV/SR: still not yet 100% clear what we should do for arch options

Otherwise work in progress:
- Shared memory and bridge PR: almost done but did not get there yet
  (This includes also some cleanup of Makefiles)
- clang format

### CarlV

CV Tried ggttgg, got make issue
"Command: make AVX=512y FPTYPE=d HELINL=0
Error:
src/MadgraphTest.cc:181:65: error: expected constructor, destructor, or type conversion before ';' token
  181 | GTEST_ALLOW_UNINSTANTIATED_PARAMETERIZED_TEST(MadgraphTestFloat);"
AV try 'make distclean', maybe an old installation of googletest

### DavidS

Thinking of test conditions to compare abstractions (in particular Alpaka)
DS sent an email, SR sent a few replies
SR suggests that DS paves the way and tries it out

SR: forgot to say that we should fix a cuda version and gcc version
SR: maybe DS can create a twiki with some suggestions? AV: or an issue?

### StephanH

Back from vacation, NTR

### OlivierM

NTR, waiting for Andrea (but no pressure)

### WalterH

Working on automation of sycl
Got kokkos to compile and fortran (with no bridge) to compile too
Will switch to sycl and integrate with Nathan's work

### NathanN

Just had a look at arch
https://arnon.dk/matching-sm-architectures-arch-and-gencode-for-various-nvidia-cards/

Moved sycl to the latest code by Andrea.
With common rand get exactly the same physics results.
Performance wise get 1:1 on V100 and slightly slower on A100.
AV: very nice to get the same performance 1:1 on V100!

Also seem to get vectorization results from sycl, will try to understand that a bit better

## AOB

Next meeting Monday March 7

 

15:00 → 15:10 News

15:10 → 15:30 Topical discussion

15:30 → 15:50 Round table

15:50 → 16:00 AoB