OpenCL Tracking

Todo List

Last updated 11.2.2013

Task	Responsible	Status
Extract RECO hits and MC truth matching for efficiency and fake rate calculation	Daniel	in progress
Recheck ( Phi / Phi' ) plot from Tracking POG talk	Daniel	open
Recheck the gabs in the Phi plot from Tracking POG talk - probably tracks with missing hits -> fix selection	Daniel	open
Integrate Rieman fit in Triplet filter	Daniel	done
Compare runtime of OpenCL algorithm to some form of CMSSW tracking	Daniel, Thomas	to be discussed
Retune cut values - different cut values for different layer	Daniel	todo
Check whether the hit charge can be useful	Thomas	todo
Efficiency and Fake rate of the Triplet joining process	Thomas	todo
Check for conversion tracks, where will they be stored ?	Thomas	todo

Algorithms & Data Flow

Insert algoithm name, possible parameters and the necessary input and output data here.

Data Organzation / Pre-partioning

Hits

The most obvious and straight-forward organization of hit data is in buffers which hold only the hits of one detection layer ( 1,2,3 pixel layers ) Furthermore, individual buffers exist for the barrel and endcap region. A more fine-grained partioning is possible.

Contraints to reduce the number of initial Tracklets

Name	Note	Constraint Formula	Detector Region
Compatiple Cluster Shape	-	-
MaxCurvature	constraint on \phi and dz depending on position in barrel
Maxd\phi	constraint of change in \phi at each layer		Barrel
Maxd\theta (\eta)	constraint of change in \theta \eta at each layer		Endcap
InteractionPoint	extrapolate candidate to beam line to check whether physically possible --> relaxed in a later iteration?
Overall Track Segmnet Length

Compatible Cluster Shape

The Cluster shape of the local hit reconstruction can give an indication whether hits might belong to the same track. This information is already used in some CMSSW Seeding steps, but this information must be extracted after the RECO step. It is no available on SIM level.

Necessary Input Data:

• Cluster Shape information of all three hits: find out how this is represented

MaxCurvature

Necessary Input Data:

• Global position of all three hits

Maxd\phi

Necessary Input Data:

• Global position of all three hits

Maxd\theta (\eta)

Amount of energy deposit in the hit

on the cluster level

Necessary Input Data:

• Global position of all three hits

InteractionPoint

Depending on the sophistication of this check either the global origin (0,0,0) is assumed to be the interaction point or a user-defined point must be provided

Necessary Input Data:

• Interaction Point in Global Coordinates
• Already paremtrized curve to be able to extrapolate to the interaction point

Constraint Implementation

To keep the kernels small and branching-free, all constraints should be implemented in seperated kernels, as the constraints will run in sequence and not concurrently. Each constraint kernel will operate on an input list of possible tracklets and an output list of tracklets which have passed the constraint. For the first constraint kernel, not a tracklet list is used as input, but simply all hits in the region of interest. This kernel will create the first tracklet candidate list. The tracklet output list of constraint kernel can be the input for the next kernel. The tracklet list can also contain more detailed information ( like fitted parameters ) to be used by later constraint or combination kernels.

During development, dedicated buffers to hold the result of each kernel run can be employed to ease debugging. In release code, a double-buffer scheme can be employed to reuse only two buffers. Once all constraint kernels were run, the final tracklet list can be handed to the kernel which combines the available tracklets.

Track Segments merging

Once the initial set of track segments has been found, they must be merged to form the complete tracks. Multiple criterias are possible to check whether neighboring track segments belong to one track.

Name	Note	Constraint Formula	Detector Region
Compatiple Direction in the x-z plane	-	-
Compatible Curvature in the r-phi plane	constraint on \phi and dz depending on position in barrel
Shared Hits in the same region

Random Implementation Notes

• the fastest Constraints should be applied first to shrink down the number of possible tracklet candidates before going to the more compute intensive Constraints
• always keep in mind, that a kernel should be able to run on the data of many events at the same time !

Open Points

• How to handle StripHits which are not confined in one dimension ( store also the error on the hit location and make the erro along the strip very large ? )
• How to handle missing hits on tracks ( plot how many missing hits regular tracks have on average ) > tracks with missing hits can be recovered if the rest of the track has been found correctly. The missing tracklet can be recreated extrapolating to the missing hit

Related Work

Alice HLT:

http://ieeexplore.ieee.org/ielx5/23/5985592/05934702.pdf?tp=&arnumber=5934702&isnumber=5985592

CBM:

http://iopscience.iop.org/1742-6596/219/3/032048/pdf/1742-6596_219_3_032048.pdf

http://www-alt.gsi.de/documents/DOC-2006-Nov-45-1.pdf

http://144.206.159.178/ft/787/72682/13400987.pdf

Belle 2:

http://indico.mppmu.mpg.de/indico/getFile.py/access?contribId=1&sessionId=0&resId=0&materialId=slides&confId=1954

-- ThomasHauth - 23-Oct-2012