Major optimization: Pre-build grid segment lists (GSLs) once per Step() call
via SyncPreparePlan(), then reuse them across all 4 RK4 substep SyncBegin calls
via SyncBeginWithPlan(). This eliminates the O(cpusize * blocks^2) GSL rebuild
cost that was incurred on every ghost zone exchange.
Applied async SyncBegin/SyncEnd overlap pattern to Z4c_class.C (ABEtype==2,
the default configuration), which was still using blocking Parallel::Sync.
Both the regular and CPBC variants of Z4c Step() are now optimized.
Co-authored-by: copilot-swe-agent[bot] <198982749+copilot@users.noreply.github.com>
Split the blocking Parallel::Sync into async SyncBegin (initiates local copy +
MPI_Isend/Irecv) and SyncEnd (MPI_Waitall + unpack). This allows overlapping MPI
ghost zone exchange with error checking and Shell patch computation.
Modified Step() in bssn_class.C for both PSTR==0 and PSTR==1/2/3 versions to
start Sync before error checks, overlapping the MPI_Allreduce with the ongoing
ghost zone transfers.
Co-authored-by: copilot-swe-agent[bot] <198982749+copilot@users.noreply.github.com>
