I am still worried about it when Real and ParticelReal are different.
What I am worried about are two things.  (1) The bisection is doing (Real - Real)*Real, whereas the particle codes do (ParticleReal-Real)*Real.  (2) Roundoff domain the hi is stored in Real.  It's cast ParticleReal when we use it in enforcePeriodic.  It's not clear to me that's 100% robust.
So I think instead of RealBox roundoff_domain, we should have GpuArray<double,AMREX_SPACEDIM> roundoff_hi_d and GpuArray<float,AMREX_SPACEDIM> roundoff_hi_f.
The bisection should mimic exactly (double_or_float - Real) * Real.
In particle codes, we can have a function that always return the roundoff hi in ParticleReal.   That function can contains ifdef to decide which one to use, roundoff_hi_f or roundoff_hi_d.
Anymore I think the PR is good enough for WarpX right now.  So feel free to merge it as is.
But in long term, it might bite us again.