GPU-friendly data structures for keeping track of particle events?

[BenWibking]

Is there an appropriate (sparse) data structure that could be used to keep track of discrete events associated with specific particles?

The use case is: a particle flows through a dynamically interesting region of the simulation and we want to record some properties about the particle (say, it's location) and the hydro variables at that position. This can happen to individual particles multiple times without limit. So most particles will have no associated "events" recorded, but some particles may have dozens or hundreds of "events".

This naturally suggests a sparse or jagged data structure for each particle, but that does not easily work performantly on GPU. Is there an existing AMReX data structure that would work for this purpose?

Should we create a separate particle container (say, EventParticles) to store the events, and then associate the real particles with the "event" particles?

[AlexanderSinn]

I would create a separate Struct-of-Array data structure to store the events. So like a EventParticles particle container but not really thinking of the elements as particles and also not necessarily using an amrex particle container. Each iteration step, a reduction over particles is performed to determine how many more events should be stored, then the events data structure is resized, and finally a second loop over particles stores the events with atomic add or a prefix sum used to determine the index into the events array. With this approach the previously recorded events of a particular particle are not reachable from that particle, however the use case doesn't seem to need this.

[BenWibking]

Is there any downside to using a separate ParticleContainer for this?

The events will always have associated positions, and while the individual events don't need to have unique IDs, they can be labeled with the cpu_id and particle_id of their associated real particle. Is there any issue with duplicating particle IDs in this way?