AMReX: Conduit with SoA Particles Broken about warpx HOT 4 CLOSED

Oh no, thanks for reporting @nicolemarsaglia!

Earlier this year, we transitioned the old data structure layout of AMReX in WarpX from a mix of AoS + SoA to be purely SoA now. I am pretty sure that must have broken it.

Here are the PRs:

• AMReX-Codes/amrex#3350
• AMReX-Codes/amrex#3639
• #4653

Do those look correct to you?
I noticed at the time that CI was missing for Ascent in AMReX and maybe a bug slipped through...

from warpx.

Do you think the specific line that fails on

        n_coords["values/x"].set_external(const_cast<ParticleReal*>(&soa.GetRealData(0)[0]),
                                          num_particles);

needs a guard for zero particles on an MPI rank? That can definitely happen and accessing [0] in that case would be a memory violation...

from warpx.

I think there is another issue there further down the code:
https://github.com/AMReX-Codes/amrex/blob/c09da9977004a32ab4f209fba6114c7889a0a5bd/Src/Extern/Conduit/AMReX_Conduit_Blueprint_ParticlesI.H#L185-L196

We now have a 64bit cpuid attribute to uniquely identify particles. In the linked code, it uses an intermediate version where we split both into 32bit numbers, which was not final. Sorry this slipped through. Do you or @cyrush know how to update this for Conduit?

You need to use soa.GetIdCPUData() for this branch instead of the current logic - and again, there can be zero particles on an MPI rank.

from warpx.

Fixed via AMReX-Codes/amrex#4065.

This will come in after the WarpX 24.08 release (which depends on AMReX 24.08, which is already tagged), via the next weekly AMReX update.

from warpx.