GPU2CPU for clusters and RecHIts

[VinInn]

to test it add to your configuration

process.siPixelDigis = cms.EDProducer("SiPixelRawToDigiGPU",
    CablingMapLabel = cms.string(''),
    ConvertADCtoElectrons = cms.bool(False),
    ErrorList = cms.vint32(29),
    IncludeErrors = cms.bool(True),
    InputLabel = cms.InputTag("rawDataCollector"),
    Regions = cms.PSet(

    ),
    Timing = cms.untracked.bool(False),
    UsePhase1 = cms.bool(True),
    UsePilotBlade = cms.bool(False),
    UseQualityInfo = cms.bool(False),
    UserErrorList = cms.vint32(40)
)

process.siPixelRecHitsPreSplitting = cms.EDProducer("SiPixelRecHitGPU",
    src = cms.InputTag("siPixelDigis"),  
    CPE = cms.string('PixelCPEFast'),
    VerboseLevel = cms.untracked.int32(0),
)

from FWCore.ParameterSet.MassReplace import massReplaceInputTag
massReplaceInputTag(process, 'siPixelClustersPreSplitting', 'siPixelDigis')

to ensure fair comparison, on CPU add

process.siPixelClustersPreSplitting.payloadType = 'HLT'

for data remember to change GT to

process.GlobalTag = GlobalTag(process.GlobalTag, '100X_dataRun2_asv2plusPixelGainfromHLT_v1', '')

printouts and asserts can be cleaned up later on

[VinInn]

still issues. after few events I got

Assertion `fd && k!=ij' failed.

even if the number of clusters is correct

investigating....

[VinInn]

changed the asserts in couts
2 events in 200 were affected. for the time being we go like this....

[VinInn]

problem understood: it is due to clusters larger than 255 pixels: on CPU we truncate at 255, on GPU there is no truncation (as there is no copy of digis).
Will leave with it unless it creates havoc in the comparison

[fwyzard]

@VinInn can you confirm that these blocks

• https://github.com/cms-patatrack/cmssw/blob/CMSSW_10_1_X_Patatrack/RecoLocalTracker/SiPixelRecHits/src/PixelCPEFast.cc#L293_L323
• https://github.com/cms-patatrack/cmssw/blob/CMSSW_10_1_X_Patatrack/RecoLocalTracker/SiPixelRecHits/src/PixelCPEFast.cc#L24_L34

should be dropped altogether ?

In fact, could I ask you to clean up the code from the debugging stuff before making the PRs ?

[cmsbot]

A new Pull Request was created by @VinInn (Vincenzo Innocente) for CMSSW_10_1_X_Patatrack.

It involves the following packages:

EventFilter/SiPixelRawToDigi
RecoLocalTracker/SiPixelRecHits

@cmsbot, @fwyzard can you please review it and eventually sign? Thanks.

cms-bot commands are listed here

[VinInn]

should we really remove all "debug" code or just ifdefit?
(similar code, with ifdefs, exists in production code)

[fwyzard]

At the very least you should remove cout messages from the destructor of the static objects that are used only for debugging, and now litter every single CMSSW-based command.

As for what is in production: IMHO there should be a single (at most per-package) flag used to #ifdef debug messages - and even that should be allowed only if the standard workflows actually exercise building and running with it defined; as things are now there are way too many packages that cannot be compiled with whatever debug flag they use being set, because that part of the code broke ages ago and was never fixed.

[fwyzard]

As for these messages, I've cleaned them up via #21

[cmsbot]

Pull request #18 was updated. @cmsbot, @fwyzard can you please check and sign again.

[cmsbot]

Pull request #18 was updated. @cmsbot, @fwyzard can you please check and sign again.

[cmsbot]

Pull request #18 was updated. @cmsbot, @fwyzard can you please check and sign again.

[cmsbot]

Pull request #18 was updated. @cmsbot, @fwyzard can you please check and sign again.

[VinInn]

all cout are "gone"
including those of #21

[fwyzard]

Thanks, I'll give it a try after integrating #19 and #20.

Running with the HEAD right now generates ~400MB log files for 1k events...

[makortel]

@VinInn

10824.8 crashes while my own config works..

Does it crash on cms-patatrack:CMSSW_10_1_X_Patatrack or on this branch?

(for 10824.8 I took the customization from #13 which I though was the latest merged one)

this is because process.siPixelRecHitsPreSplitting.src != cms.InputTag("siPixelDigis"),

actually it is missing the whole

from FWCore.ParameterSet.MassReplace import massReplaceInputTag
massReplaceInputTag(process, 'siPixelClustersPreSplitting', 'siPixelDigis')
adding it to the end of step3_RAW2DIGI_RECO_VALIDATION.py
it works...

Are all these additions from this PR? I can make a PR on top of yours adding them (and hope that the manual version of massReplaceInputTag does not touch too many packages...)

[VinInn]

On 1 Mar, 2018, at 1:35 PM, Matti Kortelainen ***@***.***> wrote: @VinInn 10824.8 crashes while my own config works.. Does it crash on cms-patatrack:CMSSW_10_1_X_Patatrack or on this branch?

this branch

(for 10824.8 I took the customization from #13 which I though was the latest merged one) this is because process.siPixelRecHitsPreSplitting.src != cms.InputTag("siPixelDigis"), actually it is missing the whole from FWCore.ParameterSet.MassReplace import massReplaceInputTag massReplaceInputTag(process, 'siPixelClustersPreSplitting', 'siPixelDigis') adding it to the end of step3_RAW2DIGI_RECO_VALIDATION.py it works... Are all these additions from this PR? I can make a PR on top of yours adding them (and hope that the manual version of massReplaceInputTag does not touch too many packages...)

yes, from this PR v.

[fwyzard]

@VinInn , here are the comparison plots for 1k Z->mumu events without pileup.

• reference is the pixel-only workflow from 10.1.0-pre1
• 10824.5 is the pixel-only workflow on the CPU from Patatrack + this PR + the changes mentioned here
• 10824.5 is the pixel-only workflow on the GPU from Patatrack + this PR + the changes mentioned here

Do the differences agree with what you expect ?

[VinInn]

Do the differences agree with what you expect?
yes, very very similar to what I obtained with ttbar
of course here the pt>50 GeV is more populated and interesting....

[fwyzard]

OK, I think I managed to put the necessary customisation to use these changes in the gpu modifier.
I'll merge this, and make a PR with that.

[fwyzard]

Done in #31 .