Permille cross section discrepancy fortran vs cudacpp (gg_tt.mad) after merging june24 and goodhel together

[valassi]

VERY BAD. There is a NEW permille cross section discrepancy fortran vs cudacpp after merging june24 and goodhel together.

This is still WIP, but anyway it does not look good.

Note:

• the test was ok (fortran==cudacpp) in both june24 and goodhel branches
• the test only fails when merging the two together (WIP)
• what changes is that the fortran xsec moves, while the cudacpp xsec remains the same
• the numbers of events are the same, so it does not look like a problem generated by the numbers of events?
• I found this on the very first test I executed, on gg_tt.mad, so I guess other processes may have this too...

[valassi]

I vaguely wonder if this is related to https://github.com/mg5amcnlo/mg5amcnlo/pull/120

I had tried to generated code with that 120, and there was only a few lines changing inggtt.mad, but this included some initializations. Maybe the initializations are missing and this will fix it? (HOWEVER june24 tests by themselves were all ok, so this sounds far fetched).

[oliviermattelaer]

Can you check at commit https://github.com/madgraph5/madgraph4gpu/pull/992/commits/7d0a553afb140a20173527a2d1537ba397d2e3d3

This was part of the master_goodhel but seems lost in the merging with june24 or something like that. This part of this branch: master_for360 (which is only two commit in top of your goodhel branch) and the CI are all green in that case.

If you agree, can you close this thread?

[valassi]

Can you check at commit 7d0a553

This was part of the master_goodhel but seems lost in the merging with june24 or something like that. This part of this branch: master_for360 (which is only two commit in top of your goodhel branch) and the CI are all green in that case.

Hi Olivier thanks, I will check, no time to do it now sorry, might have to wait next week. (I thought I had seen that comments out, but maybe it was a different branch).

If you agree, can you close this thread?

Let me test until it's fixed and then I will close it.

Just to mention (for myself mainly?) a few things I noticed yesterday, before I leave and forget:

• issue 120 does not seem related, I tried to cheery pick some commits but the issue remains, even if three variables are now initialised https://github.com/valassi/madgraph4gpu/commit/b4ede27173664f47419f2600f9f3fd7e8e0aeae6 and https://github.com/valassi/madgraph4gpu/commit/a97109b18b2f62ad96211eba145beef575ecac8c
• one thing I did notice and find strange (but may well be not related here) is that executing madevent with the same input file more than once gives different numbers of events, and this is different in cpp and fortran: cudacpp eventually settles on 1621/1616 events, fortran on 1623/1618; however the first execution in each case gives the same events as the other one
• this is related to another thing I notices or I do not understand: my madX.sh script is full of handling of ftn26 files, because this used to be the 'WRITE(26,*)' in fortran, but now using strace I never see ftn26, instead I see (in fortran only? not cudacpp?) a tmp file... something else to be understood

(Rephrasing the last point, as usual clearly identifying the inputs and outputs of madevent executables is often challenging, and this is essential to these sorts of tests: equal inputs must give equal outputs, but if there are hidden inputs/outputs, it's complex)

[valassi]

Can you check at commit 7d0a553

This was part of the master_goodhel but seems lost in the merging with june24 or something like that. This part of this branch: master_for360 (which is only two commit in top of your goodhel branch) and the CI are all green in that case.

If you agree, can you close this thread?

Hi @oliviermattelaer I finally checked that this issues is fixed in your #992. I assume that it was indeed fixed by your commit https://github.com/madgraph5/madgraph4gpu/pull/992/commits/7d0a553afb140a20173527a2d1537ba397d2e3d3

I am closing this as fixed. Thanks!