I have just merged PR #345. This contains a couple of useful things for the paper, following Olivier's suggestions

I have added code generation an dlogfiles for ggttg and ggttggg
I have yet another script to parse a selection of logfiles from various revisions

The summary of all results for the five processes I look at (eemumu, ggtt+0,1,2,3g) is here:
https://github.com/madgraph5/madgraph4gpu/blob/master/epochX/cudacpp/tput/summaryTable.txt

I tried several combinations

both double and float
my baseline cuda116/gcc102, but also cuda116/icx2021 (the latter is related to cuda116/clang13)
my baseline "no aggressive inlining" version with a "more aggressive inlining" version

There are quite a few differences, still to be understood/tweaked, between the two compilers and the two inlining options, but I would consider the baseline for our comparison.

Note that I give one CUDA number, and several SIMDs for C++. The nice thing is that a factor 4 between no SIMD and 512y SIMD for double (and a factor 8 for float) seems always there, also for ggttggg. The baseline of the baseline is thw single CUDA result, and the single 512y/C++ result.

As the complexity increases and the tests take longer, I reduce the number of events (or gpublocks/threads) for ggttggg. For C++, even a few events would be enough to reach plateau performance, but I always try to run a reasonable number for CUDA too. I always run in CUDA the same number of events as in C++, to compare the ME value. But in CUDA, for ggttgg and ggttggg I do a second run with more gpublocks/threads, to reach the plateau. Typically for V100 this is 64 blocks and 256 threads as bare minimum (below the performance always drops by factors). The detailed configs are here

madgraph4gpu/epochX/cudacpp/tput/throughputX.sh

Line 393 in 956f0e6

if [ "${exe%%/gg_ttggg*}" != "${exe}" ]; then

(look at exeArgs2 if it exists, else at exeArgs, for the CUDA blocks/threads).

Voila that's my full performance numbers as of today. They will still evolve (especially with split kernels etc).

I will also look at the two processes that Olivier suggested as a prrof of concept of generation.

(One final word of caution, I think I have some small functional bugs in the calculations, I will look at them. Related, or maybe independent, the different compilers start giving quite different results on ggttggg... maybe it's just the order of adding the 1000 diagrams...)

processes for the paper #344

Description

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions