Madgraph5 GPU development
Virtual (Zoom)
# MG dev meeting - Mon 08.11.2021
Present: SR, AV (notes), TC, DS, NathanN, LF, CarlV
Apologies: OM
## Round table
Taylor
- finalised codegen for kokkos
- debugging issues, some nans in MEs
AV: check if in all events or only some events, you must exclud ebad events from final averages
NN: also got some nans in sycl
- SR: trying nvidia? TC: yes for the moment will focus on that, then mov eto openmp etc
Andrea
- investigated factor ~2x performace slowdown in nvidia ggttgg wrt older implementations
> around 30% from number of threads/blocks, now trying best grid
> around 40% from removing static for constexpr array, now put back
> around 40% from regression from cuda 11.2 to 11.2, reported as bug to nvidia and now using 11.1
- looking at fortran throughput again
- sent some feedback to StephanH about CI test code
- busy with the LHCC review last week
- ECFA workshop talk tomorrow, will circulate slides later this afternoon
Nathan
- investigating nans observed in sycl, mainly a simple mistake (random numbers not copied from host to device)
- trying other hardwares including intel and nvidia GPUs, had some issues with sycl/intel on AMD GPUs
Laurence
- in the past had stopped hipsycl for a couple of reasons, eg not feature complete (eg missing abs), which is why I moved to oneapi
(hipsycl is developed by the university of Heidelberg)
- fixed a few issues on sycl and opened github tickets about issues I found
David
- more progress on alpaka code generation (but little time due to another project)
- code generation is essentially done, need to fix makefiles and then iterate over compilation and runtime tests
Carl
- ntr, was on vacation and busy with cms
- SR: we will probably have quite some analysis work to do when we finalise the comparison of abstraction layers
Stefan
- integrated epochX4 tag into my fortran bridging installation, need to adapt check.cc executable
- got this nice bulletin from some colleagues, https://llvm-gpu-news.github.io/
## AOB
Next meeting: Mon 22.11.2021