Packer for PrimaryVertices

For the SOACollection part, you can simply call .pack() / .unpack(). For the rest, you either need to write the packer manually or convert your structure to SOACollection to be able to leverage the automatic packer.

That seems very easy. Thanks a lot!

Here is an example: https://gitlab.cern.ch/lhcb/LHCb/-/blob/master/Event/EventBase/tests/src/test_SOAPacker.cpp#L60

The template parameter is true to enable some debug-mode prints.

Let me know if you have issues with it.

@wouter I had a look at the PrimaryVertices class and it doesn't seem very far from being all SOACollections.

I think the PrimaryVerticesContainer isn't needed, and PrimaryVertices can be made into an SOACollection owning the PVTracks collection (similar to what is done for CaloHypos https://gitlab.cern.ch/lhcb/LHCb/-/blob/master/Event/RecEvent/include/Event/CaloHypos_v2.h#L38).

namespace VerticesTag {
  struct Tracks : Event::vector_field<Event::index_field<PVTracks>> {};
  struct Pos : Event::Vec_field<3> {};
  struct Cov : Event::MatSym_field<3> {};
  struct Chi2 : Event::float_field {};
  struct nDoF : Event::int_field {};
}

Alternatively, if PVTracks are small objects and not shared between vertices, they can be inlined directly using a struct_field (similar to entries in CaloClusters https://gitlab.cern.ch/lhcb/LHCb/-/blob/master/Event/RecEvent/include/Event/CaloClusters_v2.h#L48).

I can do the conversion to SOACollection if you want.

Hi Arthur, I cannot refuse that offer :-) The PV tracks are not very small. However, they are currently already ordered by vertex. (Every vertex has an indices point to the first and last+1 track in the track list.) There are a few things to worry about:

The format of the PVTracks is currently tight to the PV fnding algorithm. So, we need to be careful to change the data layout.
There is a reason that the existing Vertex object derives from LHCb::VertexBase: there is a lot of old code with particles pointing to vertices which still works exactly because of the inheritance. So, we need to be careful that that remains working (or adapt all the downstream code, which is probably a lot of work).
Every VertexBase object also has a 'key' which comes from KeyedObject. We need the key as well (because it is used elsewhere), but this can just the index in the container (e.g. 'offset()').
If the tracks are inside the PVs, then indeed we do not need the PrimaryVerticesContainer object. However, if we get rid of it, then we still need a place for the list that has 'pointers' from the original Velo tracks to the PV tracks (https://gitlab.cern.ch/lhcb/LHCb/-/blob/master/Event/RecEvent/include/Event/PrimaryVertices.h#L182).
Note that we store both the cov matrix and its inverse. It's convenient to have both, but we need to persist only the inverse. The 'load(Buffer)' function will then need to fill the cov matrix. (In PVTrack there are some fields that are only cached for speed, I think: These objects can also be created from the original Velo container.)

If I were to implement the persistence, I would actually not change to SoA, but rather add the save(Buffer) and load(Buffer, version) functions to PrimaryVertexContainer.

@ahennequ if you do this can you please take care of making the Allen converters put out the same format, anticipating eventual discussions about which algorithms to use in HLT2 next year?

cc @adudziak

added to epic &12

mentioned in merge request !4299 (merged)

closed with merge request !4299 (merged)

mentioned in commit 194b5323

Packer for PrimaryVertices

Child items ...

Activity