Skip to content

Latest commit

 

History

History
68 lines (50 loc) · 4.49 KB

README.md

File metadata and controls

68 lines (50 loc) · 4.49 KB
Raw

mkFit-external

External header-only packages required for standalone build of mkFit from CMSSW packages RecoTracker/MkFitCore and RecoTracker/MkFitCMS.

  1. nlohman/json

  2. Extract of SMatrix package from ROOT

Setting up standalone build

Almost all sources will be used from an existing RecoTracker/MkFitCore and RecoTracker/MkFitCMS checkout. Others will be automatically taken from this repo (i.e., trackreco/mkFit-external). Some data files (geometry binary dumps) are automatically retrived from the web (from phi1.t2.ucsd.edu).

If you have an existing CMSSW environment setup, this should do the trick:

mkdir standalone
cd standalone
$CMSSW_BASE/src/RecoTracker/MkFitCore/standalone/configure $CMSSW_BASE/src
export -n INTEL_LICENSE_FILE
make -j 32
# or, for build with ROOT enabled
make WITH_ROOT=1 -j 32

Without CMSSW environment you still need to run the above configure script for initial setup. tbb can be taken from the system ... but will likely require some sym-link magick as CMSSW includes pick it up from oneapi/tbb directory which is not the standard tbb deployment on GNU/Linux.

Notes on running standalone:

export LD_LIBRARY_PATH=.
# or
export LD_LIBRARY_PATH=.:${LD_LIBRARY_PATH}

# phase1 multi-iteration test
./mkFit --cmssw-n2seeds --input-file /data2/slava77/samples/2021/11834.0_TTbar_14TeV+2021/AVE_50_BX01_25ns/memoryFile.fv6.default.211008-c6b7c67.bin --read-cmssw-tracks --build-mimi --num-iters-cmssw 9 --backward-fit --num-events 100 --quality-val --num-thr 32 --num-thr-ev 8

# phase2 initialStep test
./mkFit --input-file /data2/slava77/analysis/CMSSW_12_4_0-mkFit/work/18f4388/39634.0_TTbar_14TeV+2026D88PU/memoryFile.fv7.clean.writeAll.recT.allSeeds.220712-703c46d.bin  --geom CMS-phase2 --seed-input cmssw --seed-cleaning n2 --build-mimi --num-iters-cmssw 1 --num-events 100 --quality-val

# binary file maker
 ./writeMemoryFile  --input ~/mnt/phi3-slash/ceph/cms/store/user/slava77/CMSSW_11_2_0_mkFit/work/24feee2/2021/10muPt0p2to1HS/trackingNtuple.root --output 10mu-lowpT-1000.bin  --maxevt 1000 --write-rec-tracks

# JSON default dumper
./mkFit --json-save-iterations mkfit-phase1-%s.json --geom CMS-phase1 --num-events 0
./mkFit --json-save-iterations mkfit-phase2-%s.json --geom CMS-phase2 --num-events 0

# Validation in standalone/
SA_PATH=<path-to>/RecoTracker/MkFitCore/standalone
ln -s ${SA_PATH}/plotting .
ln -s ${SA_PATH}/val_scripts .
ln -s ${SA_PATH}/web .
ln -s ${SA_PATH}/xeon_scripts .
# Note the last argument, this is for phase2
val_scripts/validation-cmssw-benchmarks-multiiter.sh forPR --mtv-like-val TTbar_phase2
web/collectBenchmarks-multi.sh <dir-name> forPR

RntDumpers -- ROOT NTuple Dumpers

This is the latest attempt to have a somewhat standard way of dumping tunbing / debugging information directly into ROOT RNTuples or TTrees. Before we had dumpers via ROOT text tree files (slurped into ROOT via TTree::ReadFile()) -- this was slow and produced HUGE intermediate files. This resides in MkFitCore/standalone/RntDumpers.

So far there is a single implementation of such dumper, implemented in MkFinder_selectHitIndices.icc with call-out scattered through Mkfinder.cc as if-defed pieces of code, #ifdef RNT_DUMP_MkF_SelHitIdcs. To enable it, build WITH_ROOT and uncomment CPPFLAGS += -DRNT_DUMP_MkF_SelHitIdcs in Makefile.config, around line 98.

This will dump one tree entry per layer loop in MkBuilder::find_tracks_in_layers(). It contains a std::vector<CandInfo> ci structs -- detailed inforamtion about each candidate processed in this loop, including the search windows and propagation dcetails as well as detailes about considered hits. All the structrues used in dumping are defined in RntStruct.h and "standard" conversions from mkFit types to them are in RntConversions.h. Further, each CandInfo object contains a std::vector<HitMatchInfo> hmi -- this contains information about how each hit was matched in selectHitIndicesV2() and also about how it was scored (if it passed all the intermediate checks) in findCandidatesCloneEngine().

Notes:

  1. A lot of score information is missing because the hit chi2 is too large or because it is not passing the compatibility tests. One might need to rework score dumping logick to proceed anyway (or loosen the cuts).
  2. Invalid hits that are added to canidadets are clearly not dumped.
  3. One might need enlarge bin-search window scale-up factors in selectHitIndicesV2() for some very specific cases ... they are already plenty large for most use-cases (i.e., they work OK for pixelLess iteration).