Fixing cuda crashes

[vkhristenko]

PR description:

Couple of things:

• EventFilter/EcalRawToDigi depended on RecoLocalCalo/EcalRecAlgos and Producers... directly and indirectly -> removed
• EventFilter/EcalRawToDigi: All cuda kernels/device functions moved to be part of plugins
• RecoLocalCalo/EcalRecAlgos and Producers: moved all cuda kernels/device functions to be part of plugins
• RecoLocalCalo/HcalRecAlgos and Producers: moved all cuda kernels/device functions to be part of plugins

the problem was due to device-side linking across packages when it was not needed. And symbols were present in .so that were not supposed to be there. e.g. running w/o current pr just ecal digi to raw + full hcal, you get the same issue. doing nm on the pluginEventFilterEcalRawToDigi.so will show symbols from RecoLocalCalo/EcalRecAlgos.

Furthermore, I think the following situation will fail:

• we have 2 shared objs: A and B. Both have host + device funcs
• B depends on A both on host and device. Therefore B will be linked against A. dynamic for host and static for device
• ldopen both, and running any kernel from B should fail. This is what we basically have now. Am I right here? I do not understand if this is supposed to work actually or not...

PR validation:

Using the exes provided as part of the release.
@mariadalfonso do not know if you remember...
1 thing to note for hcal: i do not know what was done for the cpu, cause whenever i used to validate cpu vs gpu, i used a different branch (w/ a couple of fixes applied). Note, it was documented as part of previous prs

[mariadalfonso]

@mariadalfonso do not know if you remember...
1 thing to note for hcal: i do not know what was done for the cpu, cause whenever i used to validate cpu vs gpu, i used a different branch (w/ a couple of fixes applied). Note, it was documented as part of previous prs

Both CPU fixes are in the main release.
Few things need to be customized at HLT (as 8 pulse fit all the times + applyLegacyHBMCorrection=False) to get the equivalent results. :)

[vkhristenko]

@mariadalfonso perfect! :)

[fwyzard]

@mariadalfonso

Few things need to be customized at HLT (as 8 pulse fit all the times + applyLegacyHBMCorrection=False) to get the equivalent results. :)

For the HLT menu to use

• should I start from the "frozen" OnLine_HLT_2018.py menu, or from the "current" OnLine_HLT_GRun.py menu ?
• from the HCAL point of view, do I need to apply any customisation to either of them ?
• from the HCAL point of view, do they run correctly on 2018 data ?

[mariadalfonso]

@mariadalfonso

Few things need to be customized at HLT (as 8 pulse fit all the times + applyLegacyHBMCorrection=False) to get the equivalent results. :)

For the HLT menu to use

• should I start from the "frozen" OnLine_HLT_2018.py menu, or from the "current" OnLine_HLT_GRun.py menu ?
• from the HCAL point of view, do I need to apply any customisation to either of them ?
• from the HCAL point of view, do they run correctly on 2018 data ?

1. better start from "frozen" OnLine_HLT_2018.py
2. apply the customization synchronizeHCALHLTofflineRun2(process) of https://github.com/cms-sw/cmssw/blob/CMSSW_11_1_0_pre8/HLTrigger/Configuration/python/customizeHLTforCMSSW.py#L87
3. additional customization as below (this is what the GPU has implemented, since we have this in the Run3)

    for producer in producers_by_type(process, "HBHEPhase1Reconstructor"):
        producer.algorithm.applyLegacyHBMCorrection = cms.bool( False )
        producer.algorithm.chiSqSwitch = cms.double(-1)

With the 3 points above you a correct results in 2018 data and the two floats energy-cpu and energy-gpu that should be identical.

[fwyzard]

@vkhristenko I have a few fixes:

• fix few #include statements were still pointing to the header files in the old location
• fix include guards and #include order
• apply clang-format changes

Could you apply them with

curl -L https://github.com/cms-patatrack/cmssw/files/4774221/diff.txt | patch -p1

?

The diff is attached: diff .

[vkhristenko]

@fwyzard applied

[fwyzard]

With these changes, I finally have the full HLT running 👍