From d31061cd497d78c03ab6de9d418915de5e70b762 Mon Sep 17 00:00:00 2001
From: Peter Boyle <paboyle@ph.ed.ac.uk>
Date: Fri, 2 Oct 2026 11:57:27 -0400
Subject: [PATCH] Batched split grid

---
 Grid/util/Init.cc | 15 +++++++++++++++
 Grid/util/Init.h  |  5 +++++
 2 files changed, 20 insertions(+)

diff --git a/Grid/util/Init.cc b/Grid/util/Init.cc
index 0c5787879..e6f621cde 100644
--- a/Grid/util/Init.cc
+++ b/Grid/util/Init.cc
@@ -95,6 +95,8 @@ NAMESPACE_BEGIN(Grid);
 //////////////////////////////////////////////////////
 static Coordinate Grid_default_latt;
 static Coordinate Grid_default_mpi;
+static Coordinate Grid_default_batched_solver_split;
+static bool       Grid_default_batched_solver_split_node = false;
 
 int GridThread::_threads =1;
 int GridThread::_hyperthreads=1;
@@ -108,6 +110,8 @@ char *GridHostname(void)
 }
 const Coordinate &GridDefaultLatt(void)     {return Grid_default_latt;};
 const Coordinate &GridDefaultMpi(void)      {return Grid_default_mpi;};
+const Coordinate &GridDefaultBatchedSolverSplit(void)     {return Grid_default_batched_solver_split;};
+bool              GridDefaultBatchedSolverSplitNode(void) {return Grid_default_batched_solver_split_node;};
 const Coordinate GridDefaultSimd(int dims,int nsimd)
 {
   Coordinate layout(dims);
@@ -562,6 +566,7 @@ void Grid_init(int *argc,char ***argv)
     std::cout<<GridLogMessage<<"Performance:"<<std::endl;
     std::cout<<GridLogMessage<<std::endl;
     std::cout<<GridLogMessage<<"  --comms-overlap    : Overlap comms with compute "<<std::endl;    
+    std::cout<<GridLogMessage<<"  --batched-solver-split a.b.c.d|node : batched solvers run independent solves on partitions of this MPI layout (node = one per node)"<<std::endl;
     std::cout<<GridLogMessage<<std::endl;
     std::cout<<GridLogMessage<<"  --dslash-generic: Wilson kernel for generic Nc"<<std::endl;    
     std::cout<<GridLogMessage<<"  --dslash-unroll : Wilson kernel for Nc=3"<<std::endl;    
@@ -613,6 +618,16 @@ void Grid_init(int *argc,char ***argv)
 		  Grid_default_latt,
 		  Grid_default_mpi);
 
+  if( GridCmdOptionExists(*argv,*argv+*argc,"--batched-solver-split") ){
+    std::string arg = GridCmdOptionPayload(*argv,*argv+*argc,"--batched-solver-split");
+    if ( arg == "node" ) {
+      Grid_default_batched_solver_split_node = true;
+    } else {
+      GridCmdOptionIntVector(arg,Grid_default_batched_solver_split);
+      GRID_ASSERT(Grid_default_batched_solver_split.size() == Grid_default_mpi.size());
+    }
+  }
+
   if( GridCmdOptionExists(*argv,*argv+*argc,"--decomposition") ){
     std::cout<<GridLogMessage<<"Grid Default Decomposition patterns\n";
     std::cout<<GridLogMessage<<"\tOpenMP threads : "<<GridThread::GetThreads()<<std::endl;
diff --git a/Grid/util/Init.h b/Grid/util/Init.h
index b370077a8..aa1f7e7ee 100644
--- a/Grid/util/Init.h
+++ b/Grid/util/Init.h
@@ -49,6 +49,11 @@ void Grid_unquiesce_nodes(void);
 const Coordinate  GridDefaultSimd(int dims,int nsimd);
 const Coordinate &GridDefaultLatt(void);
 const Coordinate &GridDefaultMpi(void);
+// --batched-solver-split a.b.c.d : MPI layout of each partition for batched solves.
+// --batched-solver-split node    : one partition per node.
+// Not given: empty Coordinate and false, meaning no split.
+const Coordinate &GridDefaultBatchedSolverSplit(void);
+bool              GridDefaultBatchedSolverSplitNode(void);
 const int        &GridThreads(void)  ;
 void              GridSetThreads(int t) ;
 void GridLogTimestamp(int);