Phil Miller - NOAA
Phil Miller - NOAA
Per the latter, I'm trying this with `AMReX_CUDA_FASTMATH=OFF` to rule that out.
I'll also try an OpenMP build and a debug build. In these runs, I'm using a slightly modified version of AMReX to change the assertion in `enforcePeriodic` to `AMREX_ALWAYS_ASSERT`, to...
~~The build without fastmath seemed to be producing nonsensical results.~~ OpenMP next
Ugh, I screwed up the `no_fast_math` build, and accidentally had single precision fields. So, I'll rebuild and rerun that
In the OpenMP build, that assertion in `enforcePeriodic` triggers! @peterscherpelz @KZhu-ME
Following https://github.com/AMReX-Codes/amrex/pull/2839/ might we need the round-off domain fix for `plo` as well?
Digging through the code, it does seem that particles for which rounding leaves them outside the domain after a periodic shift can in fact get discarded, since they will not...
I do wonder if instead of the bisection approach to computing `rhi/rlo` could be replaced with a suitably careful use of routines from the `nextafter` family Edit: Or probably, more...
Weekly AMReX update was merged, and 22.10 release tagged. This is done.
The reason for the change was essentially so that a DRV would return a layout matching its dynamic state, not the static 7 ranks of the underlying View that it...