Limit GPU shell RHS redirection to Step and SHStep only via #define/#undef.
Compute_Constraint, Interp_Constraint, and Constraint_Out continue using
the CPU Fortran path to avoid GPU alloc-per-call overhead during
initialization and analysis phases.
Also: wrap compare_result_gpu in #ifdef RESULT_CHECK to avoid link error.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Download BSSN StateList from GPU to CPU before AHFinderDirect_find_horizons
so that AH_Interp_Points reads valid field data instead of stale CPU arrays.
The resident-sync path keeps canonical state on GPU; without this download the
Newton iteration diverges and probes outside the computational domain.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
The 8 nested while(Ppc){while(Pp){OutBdLow2Hi(single,single,...)}}
loops across RestrictProlong (3 overloads) and ProlongRestrict each
produced N_c × N_f separate transfer() → MPI_Waitall barriers.
Replace with the existing batch OutBdLow2Hi(MyList<Patch>*,...) which
merges all patch pairs into a single transfer() call with 1 MPI_Waitall.
Also add Restrict_cached, OutBdLow2Hi_cached, OutBdLow2Himix_cached
to Parallel (unused for now — kept as infrastructure for future use).
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Cache Sync in RestrictProlong: replace 11 basic Parallel::Sync() calls
with Parallel::Sync_cached() across RestrictProlong, RestrictProlong_aux,
and ProlongRestrict to avoid rebuilding grid segment lists every call.
Merge paired MPI_Allreduce in surface_integral: combine 9 pairs of
consecutive RP/IP Allreduce calls into single calls with count=2*NN.
Merge scalar MPI_Allreduce in surf_MassPAng: combine 3 groups of 7
scalar Allreduce calls (mass + angular/linear momentum) into single
calls with count=7.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Phase 1: Merge N+1 transfer() calls into a single transfer() per
Sync(PatchList), reducing N+1 MPI_Waitall barriers to 1 via new
Sync_merged() that collects all intra-patch and inter-patch grid
segment lists into combined per-rank arrays.
Phase 2: Cache grid segment lists and reuse grow-only communication
buffers across RK4 substeps via SyncCache struct. Caches are per-level
and per-variable-list (predictor/corrector), invalidated on regrid.
Eliminates redundant build_ghost_gsl/build_owned_gsl0/build_gstl
rebuilds and malloc/free cycles between regrids.
Phase 3: Split Sync into async Sync_start/Sync_finish to overlap
Cartesian ghost zone exchange (MPI_Isend/Irecv) with Shell patch
synchronization. Uses MPI tag 2 to avoid conflicts with SH->Synch()
which uses transfer() with tag 1.
Also updates makefile.inc paths and flags for local build environment.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Replace all 8 blocking MPI_Allreduce error-check calls with MPI_Iallreduce,
overlapping the reduction with subsequent Parallel::Sync/SH->Synch operations.
MPI_Wait is called after Sync completes to retrieve the error result.
This hides the Allreduce latency (46.5% of CPU time) behind the ghost zone
exchange communication that must happen anyway. Safe because Sync only copies
existing data to ghost zones and the error check + abort happens before any
further computation uses the synced data.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
In the two Step() functions that handle both Patch and Shell Patch,
defer the Patch error check until after Shell Patch computation completes,
then perform a single combined MPI_Allreduce instead of two separate ones.
This eliminates 4 MPI_Allreduce calls per timestep (2 per Step function,
Predictor + Corrector phases each). The optimization is mathematically
equivalent: in normal execution (no NaN) behavior is identical; on error,
both Patch and Shell data are dumped before MPI_Abort.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
