Merge branch 'feature/splitGridCGBatch' into develop

Grid/GridCore.h

@@ -56,6 +56,7 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 #include <Grid/communicator/Communicator.h> 
 #include <Grid/communicator/RingAllReduce.h>
 #include <Grid/cartesian/Cartesian.h>
+#include <Grid/perfmon/HostMemory.h>
 #include <Grid/tensors/Tensors.h>      
 #include <Grid/lattice/Lattice.h>      
 #include <Grid/cshift/Cshift.h>       

Grid/algorithms/SplitOperator.h

@@ -104,13 +104,11 @@ inline int GridSplitVectorIndex(GridBase *full,GridBase *split)
 // line (--batched-solver-split). Uses the trailing dimensions of grid's processor and shm
 // layouts, so it works for 4d and 5d grids. Returns grid's own processor layout (one
 // partition) when no split is requested. Asserts divisibility; warns when partitions
-// straddle nodes, when node boundaries are not visible, or when there are more partitions
-// than right-hand sides.
+// straddle nodes or when node boundaries are not visible.
 /////////////////////////////////////////////////////////////////////////////////////////////
 inline Coordinate BatchedSolverSplitLayout(GridBase *grid,
                                            const Coordinate &request,
                                            bool node,
-                                           int nbatch,
                                            int &partitions)
 {
   int nd  = GridDefaultMpi().size();
@@ -168,10 +166,6 @@ inline Coordinate BatchedSolverSplitLayout(GridBase *grid,
     std::cout << GridLogWarning << "BatchedSolverSplit: partitions " << split
               << " straddle node boundaries (node layout " << shm << "); inner solves will communicate off node" << std::endl;
   }
-  if ( partitions > nbatch ) {
-    std::cout << GridLogWarning << "BatchedSolverSplit: " << partitions << " partitions for "
-              << nbatch << " right-hand sides; the extra partitions only solve zero padding" << std::endl;
-  }
   return split;
 }
 

Grid/algorithms/iterative/ConjugateGradientMixedPrecBatched.h

@@ -53,9 +53,17 @@ public:
   
   // Inner solves on independent partitions of the communicator (--batched-solver-split).
   // BatchedSplit is the partition MPI layout, empty for no split; BatchedSplitNode selects
-  // one partition per node. Default from the command line; callers may override.
-  Coordinate BatchedSplit;
-  bool       BatchedSplitNode;
+  // one partition per node. Fixed at construction, default from the command line.
+  const Coordinate BatchedSplit;
+  const bool       BatchedSplitNode;
+
+  // Copy of Linop_f on the partitions, made once at construction and owned; nullptr when
+  // there is no split or the operator cannot be split. It holds the gauge field as it was at
+  // construction: a solver must not outlive a change of the operator's gauge field. The
+  // outer defect correction uses Linop_d, so a stale copy would cost convergence, not
+  // correctness.
+  SplitOperator<FieldF> *split = nullptr;
+  int                    Partitions = 1;
 
   MixedPrecisionConjugateGradientBatched(RealD tol, 
           Integer maxinnerit, 
@@ -64,11 +72,37 @@ public:
           GridBase* _sp_grid, 
           LinearOperatorBase<FieldF> &_Linop_f, 
           LinearOperatorBase<FieldD> &_Linop_d,
-          bool _updateResidual=true) :
+          bool _updateResidual=true,
+          const Coordinate &_split=GridDefaultBatchedSolverSplit(),
+          bool _split_node=GridDefaultBatchedSolverSplitNode()) :
     Linop_f(_Linop_f), Linop_d(_Linop_d),
     Tolerance(tol), InnerTolerance(tol), MaxInnerIterations(maxinnerit), MaxOuterIterations(maxouterit), MaxPatchupIterations(maxpatchit), SinglePrecGrid(_sp_grid),
     OuterLoopNormMult(100.), guesser(NULL), updateResidual(_updateResidual),
-    BatchedSplit(GridDefaultBatchedSolverSplit()), BatchedSplitNode(GridDefaultBatchedSolverSplitNode()) { };
+    BatchedSplit(_split), BatchedSplitNode(_split_node)
+  {
+    Coordinate layout = BatchedSolverSplitLayout(SinglePrecGrid,BatchedSplit,BatchedSplitNode,Partitions);
+    if ( Partitions > 1 ) {
+      double t0 = usecond();
+      split = Linop_f.SplitClone(layout);
+      double t1 = usecond();
+      if ( split == nullptr ) {
+        std::cout << GridLogMessage << "MixedPrecisionConjugateGradientBatched: operator cannot be split; serial inner solves" << std::endl;
+      } else {
+        std::cout << GridLogMessage << "MixedPrecisionConjugateGradientBatched: split clone " << (t1-t0)/1.0e6 << " s" << std::endl;
+        HostMemoryReport(SinglePrecGrid,GridLogMessage,"MixedPrecisionConjugateGradientBatched: after split clone");
+      }
+    }
+  };
+
+  // Owns split
+  MixedPrecisionConjugateGradientBatched(const MixedPrecisionConjugateGradientBatched &) = delete;
+
+  MixedPrecisionConjugateGradientBatched &operator=(const MixedPrecisionConjugateGradientBatched &) = delete;
+
+  ~MixedPrecisionConjugateGradientBatched(void)
+  {
+    delete split;
+  }
 
   void useGuesser(LinearFunction<FieldF> &g){
     guesser = &g;
@@ -98,6 +132,8 @@ public:
 
     GridStopWatch InnerCGtimer;
     GridStopWatch PrecChangeTimer;
+    GridStopWatch OuterResidualTimer;
+    GridStopWatch PatchupTimer;
     
     int cb = src_d_in[0].Checkerboard();
     
@@ -137,18 +173,10 @@ public:
     ConjugateGradient<FieldF> CG_f(inner_tol, MaxInnerIterations);
     CG_f.ErrorOnNoConverge = false;
     
-    //Optionally clone the single precision operator onto split-communicator partitions
-    GridStopWatch SplitTimer;
-    SplitOperator<FieldF> *split = nullptr;
-    int partitions;
-    Coordinate layout = BatchedSolverSplitLayout(SinglePrecGrid,BatchedSplit,BatchedSplitNode,NBatch,partitions);
-    if ( partitions > 1 ) {
-      SplitTimer.Start();
-      split = Linop_f.SplitClone(layout);
-      SplitTimer.Stop();
-      if ( split == nullptr ) {
-        std::cout << GridLogMessage << "MixedPrecisionConjugateGradientBatched: operator cannot be split; serial inner solves" << std::endl;
-      }
+    SplitTimers splitTimers;
+    if ( split != nullptr && split->Partitions > NBatch ) {
+      std::cout << GridLogWarning << "MixedPrecisionConjugateGradientBatched: " << split->Partitions << " partitions for "
+                << NBatch << " right-hand sides; the extra partitions only solve zero padding" << std::endl;
     }
     
     Integer &outer_iter = TotalOuterIterations; //so it will be equal to the final iteration count
@@ -156,13 +184,16 @@ public:
     for(outer_iter = 0; outer_iter < MaxOuterIterations; outer_iter++){
       std::cout << GridLogMessage << std::endl;
       std::cout << GridLogMessage << "Outer iteration " << outer_iter << std::endl;
+      HostMemoryReport(DoublePrecGrid,GridLogMessage,"MixedPrecisionConjugateGradientBatched: outer iteration "+std::to_string(outer_iter));
       
       bool allConverged = true;
       
       for (int i=0; i<NBatch; i++) {
         //Compute double precision rsd and also new RHS vector.
+        OuterResidualTimer.Start();
         Linop_d.HermOp(sol_d[i], tmp_d);
         norm[i] = axpy_norm(src_d[i], -1., tmp_d, src_d_in[i]); //src_d is residual vector
+        OuterResidualTimer.Stop();
         
         std::cout<<GridLogMessage<<"MixedPrecisionConjugateGradientBatched: Outer iteration " << outer_iter <<" solve " << i << " residual "<< norm[i] << " target "<< stop[i] <<std::endl;
 
@@ -191,7 +222,7 @@ public:
       }
 
       if ( split != nullptr ) {
-        InnerSplitSolves(*split, CG_f, src_f, sol_f, TotalInnerIterations, InnerCGtimer, SplitTimer);
+        InnerSplitSolves(*split, CG_f, src_f, sol_f, TotalInnerIterations, InnerCGtimer, splitTimers);
       }
 
       for (int i=0; i<NBatch; i++) {
@@ -213,17 +244,17 @@ public:
 
     }
 
-    delete split;
-    
     //Final trial CG
     std::cout << GridLogMessage << std::endl;
     std::cout<<GridLogMessage<<"MixedPrecisionConjugateGradientBatched: Starting final patch-up double-precision solve"<<std::endl;
     
+    PatchupTimer.Start();
     for (int i=0; i<NBatch; i++) {
       ConjugateGradient<FieldD> CG_d(Tolerance, MaxPatchupIterations);
       CG_d(Linop_d, src_d_in[i], sol_d[i]);
       TotalFinalStepIterations[i] += CG_d.IterationsToComplete;
     }
+    PatchupTimer.Stop();
 
     TotalTimer.Stop();
 
@@ -232,12 +263,36 @@ public:
       std::cout<<GridLogMessage<<"MixedPrecisionConjugateGradientBatched: solve " << i << " Inner CG iterations " << TotalInnerIterations[i] << " Restarts " << TotalOuterIterations << " Final CG iterations " << TotalFinalStepIterations[i] << std::endl;
     }
     std::cout << GridLogMessage << std::endl;
-    std::cout<<GridLogMessage<<"MixedPrecisionConjugateGradientBatched: Total time " << TotalTimer.Elapsed() << " Precision change " << PrecChangeTimer.Elapsed() << " Inner CG total " << InnerCGtimer.Elapsed() << " Split setup and transfer " << SplitTimer.Elapsed() << std::endl;
+    std::cout<<GridLogMessage<<"MixedPrecisionConjugateGradientBatched: Total time " << TotalTimer.Elapsed() << " Precision change " << PrecChangeTimer.Elapsed() << " Inner CG total " << InnerCGtimer.Elapsed() << std::endl;
+    std::cout<<GridLogMessage<<"MixedPrecisionConjugateGradientBatched: Outer residual " << OuterResidualTimer.Elapsed() << " Patch-up " << PatchupTimer.Elapsed() << std::endl;
+    if ( split != nullptr ) {
+      std::cout<<GridLogMessage<<"MixedPrecisionConjugateGradientBatched: Grid_split " << splitTimers.Split.Elapsed() << " (" << splitTimers.SplitCalls << " calls)"
+               << " Grid_unsplit " << splitTimers.Unsplit.Elapsed() << " (" << splitTimers.UnsplitCalls << " calls)"
+               << " Staging copies " << splitTimers.Staging.Elapsed()
+               << " Host to device " << splitTimers.H2D.Elapsed()
+               << " Iteration count sum " << splitTimers.Reduce.Elapsed() << std::endl;
+    }
+    double accounted = PrecChangeTimer.useconds() + InnerCGtimer.useconds() + OuterResidualTimer.useconds()
+                     + PatchupTimer.useconds() + splitTimers.Split.useconds()
+                     + splitTimers.Unsplit.useconds() + splitTimers.Staging.useconds() + splitTimers.H2D.useconds()
+                     + splitTimers.Reduce.useconds();
+    std::cout<<GridLogMessage<<"MixedPrecisionConjugateGradientBatched: Unaccounted " << (TotalTimer.useconds()-accounted)/1.0e6 << " s" << std::endl;
     
   }
 
 private:
 
+  // Coarse breakdown of the split path outside the inner CG
+  struct SplitTimers {
+    GridStopWatch Split;
+    GridStopWatch Unsplit;
+    GridStopWatch Staging;
+    GridStopWatch H2D;
+    GridStopWatch Reduce;
+    int SplitCalls   = 0;
+    int UnsplitCalls = 0;
+  };
+
   ////////////////////////////////////////////////////////////////////////////////////////
   // One inner solve per partition, in groups of Partitions right-hand sides. The last
   // group is zero-padded; a zero source returns at once from CG. Collective.
@@ -248,7 +303,7 @@ private:
                         std::vector<FieldF> &sol_f,
                         std::vector<Integer> &TotalInnerIterations,
                         GridStopWatch &InnerCGtimer,
-                        GridStopWatch &SplitTimer)
+                        SplitTimers &timers)
   {
     int NBatch = src_f.size();
     int P      = split.Partitions;
@@ -263,6 +318,7 @@ private:
     for(int g=0;g<NBatch;g+=P){
 
       // Gather the group; zero-pad past the end of the batch
+      timers.Staging.Start();
       for(int p=0;p<P;p++){
         group_src[p].Checkerboard() = cb;
         group_sol[p].Checkerboard() = cb;
@@ -274,22 +330,34 @@ private:
           group_sol[p] = Zero();
         }
       }
+      timers.Staging.Stop();
 
       // The initial guess (e.g. from the guesser) travels with the source
-      SplitTimer.Start();
+      timers.Split.Start();
       Grid_split(group_src,s_src);
       Grid_split(group_sol,s_sol);
-      SplitTimer.Stop();
+      timers.Split.Stop();
+      timers.SplitCalls += 2;
+
+      // Grid_split leaves its output on the host; move it now so the copy is not in the CG time
+      timers.H2D.Start();
+      {
+        autoView(s_src_v, s_src, AcceleratorRead);
+        autoView(s_sol_v, s_sol, AcceleratorRead);
+      }
+      timers.H2D.Stop();
 
       InnerCGtimer.Start();
       CG_f(*split.Linop,s_src,s_sol);
       InnerCGtimer.Stop();
 
-      SplitTimer.Start();
+      timers.Unsplit.Start();
       Grid_unsplit(group_sol,s_sol);
-      SplitTimer.Stop();
+      timers.Unsplit.Stop();
+      timers.UnsplitCalls += 1;
 
       // One iteration count per partition, contributed by the partition's rank 0 only
+      timers.Reduce.Start();
       for(int p=0;p<P;p++){
         iters[p] = 0;
       }
@@ -297,11 +365,15 @@ private:
         iters[split.Partition] = CG_f.IterationsToComplete;
       }
       SinglePrecGrid->GlobalSumVector(&iters[0],P);
+      timers.Reduce.Stop();
 
+      // Includes the host to device copy of group_sol left by Grid_unsplit
+      timers.Staging.Start();
       for(int p=0;p<P && g+p<NBatch;p++){
         sol_f[g+p] = group_sol[p];
         TotalInnerIterations[g+p] += iters[p];
       }
+      timers.Staging.Stop();
     }
   }
 };

Grid/lattice/Lattice_transfer.h

@@ -1550,174 +1550,277 @@ void precisionChange(Lattice<VobjOut> &out, const Lattice<VobjIn> &in){
 ////////////////////////////////////////////////////////////////////////////////
 // Communicate between grids
 ////////////////////////////////////////////////////////////////////////////////
-///////////////////////////////////////////////////////////////////////////////////////////////////////////
-// SIMPLE CASE:
-///////////////////////////////////////////////////////////////////////////////////////////////////////////
+/////////////////////////////////////////////////////////////////////////////////////////////
+// Grid_split / Grid_unsplit
 //
-// Mesh of nodes (2x2) ; subdivide to  1x1 subdivisions
+// The split grid divides the full grid's processor layout Pf into partitions of layout
+// Ps; ratio R = Pf/Ps per dimension, nvector = prod R partitions. A rank at full processor
+// coordinate c belongs to partition q = c/Ps at split coordinate s = c%Ps, and partition q
+// holds vector index lex(q,R) (GridSplitVectorIndex). Local volumes satisfy
+// Ls = Lf * R per dimension.
 //
-// Lex ord:   
-//          N0 va0 vb0 vc0 vd0       N1 va1 vb1 vc1 vd1  
-//          N2 va2 vb2 vc2 vd2       N3 va3 vb3 vc3 vd3 
+// Every full-grid rank's local block of field v lies wholly inside one split-grid rank's
+// local volume, in the sub-box at offset o = c%R (units of Lf), on the rank of partition
+// q(v) at split coordinate c/R. The redistribution is therefore nvector whole-block
+// point-to-point exchanges, nothing finer, and no data moves twice. Step k (a coordinate in
+// the R box) pairs every rank with one sender and one receiver:
 //
-// Ratio = full[dim] / split[dim]
+//   send  field q = (c%R + k)%R      to   full coordinate q*Ps + c/R
+//   recv  sub-box o = (c/Ps - k)%R   from full coordinate (c%Ps)*R + o
 //
-// For each dimension do an all to all; get Nvec -> Nvec / ratio
-//                                          Ldim -> Ldim * ratio
-//                                          LocalVol -> LocalVol * ratio
-// full AllToAll(0)
-//          N0 va0 vb0 va1 vb1       N1 vc0 vd0 vc1 vd1   
-//          N2 va2 vb2 va3 vb3       N3 vc2 vd2 vc3 vd3 
+// Within a step the send map is a permutation of the ranks and the receive map is its
+// inverse, so each step is one symmetric SendToRecvFrom per rank; a rank paired with
+// itself copies locally. Grid_unsplit runs the same steps with the roles exchanged.
 //
-// REARRANGE
-//          N0 va01 vb01      N1 vc01 vd01
-//          N2 va23 vb23      N3 vc23 vd23
-//
-// full AllToAll(1)           // Not what is wanted. FIXME
-//          N0 va01 va23      N1 vc01 vc23 
-//          N2 vb01 vb23      N3 vd01 vd23
-// 
-// REARRANGE
-//          N0 va0123      N1 vc0123
-//          N2 vb0123      N3 vd0123
-//
-// Must also rearrange data to get into the NEW lex order of grid at each stage. Some kind of "insert/extract".
-// NB: Easiest to programme if keep in lex order.
-/*
- *  Let chunk = (fvol*nvec)/sP be size of a chunk.         ( Divide lexico vol * nvec into fP/sP = M chunks )
- *  
- *  2nd A2A (over sP nodes; subdivide the fP into sP chunks of M)
- * 
- *     node 0     1st chunk of node 0M..(1M-1); 2nd chunk of node 0M..(1M-1)..   data chunk x M x sP = fL / sP * M * sP = fL * M growth
- *     node 1     1st chunk of node 1M..(2M-1); 2nd chunk of node 1M..(2M-1)..
- *     node 2     1st chunk of node 2M..(3M-1); 2nd chunk of node 2M..(3M-1)..
- *     node 3     1st chunk of node 3M..(3M-1); 2nd chunk of node 2M..(3M-1)..
- *  etc...
- */
+// Data stays on the device. A block on the full grid is the local volume in lexicographic
+// order; the split grid's local volume is staged box-major (sub-box index, then
+// lexicographic within the sub-box), so every exchanged block is contiguous. Packing and
+// unpacking are accelerator kernels. Buffers are deviceVectors: the staged split volume
+// plus one block. Without ACCELERATOR_AWARE_MPI each block is staged through host memory.
+/////////////////////////////////////////////////////////////////////////////////////////////
+
+// Scalar sites of in, written to buf in box-major order: sub-boxes of extent bdims taken
+// lexicographically, sites lexicographically within each. bdims = local dimensions gives
+// plain lexicographic order.
+template<class vobj>
+void GridSplitUnvectorise(const Lattice<vobj> &in,typename vobj::scalar_object *buf,const Coordinate &bdims)
+{
+  GridBase *grid = in.Grid();
+  int nd = grid->_ndimension;
+  Coordinate rdims = grid->_rdimensions;
+  Coordinate simd  = grid->_simd_layout;
+  Coordinate nbox(nd);
+  int64_t bvol = 1;
+  for(int d=0;d<nd;d++){
+    GRID_ASSERT( (grid->_ldimensions[d] % bdims[d]) == 0 );
+    nbox[d] = grid->_ldimensions[d] / bdims[d];
+    bvol   *= bdims[d];
+  }
+  const int Nsimd = grid->Nsimd();
+  autoView( in_v , in, AcceleratorRead);
+  accelerator_for(ss, grid->oSites(), 1, {
+    Coordinate ocoor(nd);
+    Coordinate icoor(nd);
+    Coordinate bcoor(nd);
+    Coordinate wcoor(nd);
+    Lexicographic::CoorFromIndex(ocoor,ss,rdims);
+    for(int lane=0;lane<Nsimd;lane++){
+      Lexicographic::CoorFromIndex(icoor,lane,simd);
+      for(int d=0;d<nd;d++){
+        int l = ocoor[d] + rdims[d]*icoor[d];
+        bcoor[d] = l / bdims[d];
+        wcoor[d] = l % bdims[d];
+      }
+      int64_t b;
+      int64_t w;
+      Lexicographic::IndexFromCoor(bcoor,b,nbox);
+      Lexicographic::IndexFromCoor(wcoor,w,bdims);
+      buf[b*bvol+w] = extractLane(lane,in_v[ss]);
+    }
+  });
+}
+
+// Inverse of GridSplitUnvectorise: every site of out is written
+template<class vobj>
+void GridSplitVectorise(const typename vobj::scalar_object *buf,Lattice<vobj> &out,const Coordinate &bdims)
+{
+  GridBase *grid = out.Grid();
+  int nd = grid->_ndimension;
+  Coordinate rdims = grid->_rdimensions;
+  Coordinate simd  = grid->_simd_layout;
+  Coordinate nbox(nd);
+  int64_t bvol = 1;
+  for(int d=0;d<nd;d++){
+    GRID_ASSERT( (grid->_ldimensions[d] % bdims[d]) == 0 );
+    nbox[d] = grid->_ldimensions[d] / bdims[d];
+    bvol   *= bdims[d];
+  }
+  const int Nsimd = grid->Nsimd();
+  autoView( out_v , out, AcceleratorWriteDiscard);
+  accelerator_for(ss, grid->oSites(), 1, {
+    Coordinate ocoor(nd);
+    Coordinate icoor(nd);
+    Coordinate bcoor(nd);
+    Coordinate wcoor(nd);
+    Lexicographic::CoorFromIndex(ocoor,ss,rdims);
+    for(int lane=0;lane<Nsimd;lane++){
+      Lexicographic::CoorFromIndex(icoor,lane,simd);
+      for(int d=0;d<nd;d++){
+        int l = ocoor[d] + rdims[d]*icoor[d];
+        bcoor[d] = l / bdims[d];
+        wcoor[d] = l % bdims[d];
+      }
+      int64_t b;
+      int64_t w;
+      Lexicographic::IndexFromCoor(bcoor,b,nbox);
+      Lexicographic::IndexFromCoor(wcoor,w,bdims);
+      insertLane(lane,out_v[ss],buf[b*bvol+w]);
+    }
+  });
+}
+
+// Step schedule shared by Grid_split and Grid_unsplit; see the description above
+class GridSplitSchedule
+{
+public:
+  int        nd;
+  int        nvector;
+  Coordinate ratio;      // R
+  Coordinate fldims;     // full grid local dimensions = sub-box extent on the split grid
+  GridBase  *full_grid;
+
+  // Step k: this rank sends field Field(k) to rank FieldRank(k) and receives sub-box
+  // Box(k) from rank BoxRank(k) (Grid_split); Grid_unsplit reverses both
+  std::vector<int> Field;
+  std::vector<int> FieldRank;
+  std::vector<int> Box;
+  std::vector<int> BoxRank;
+
+  GridSplitSchedule(GridBase *_full_grid,GridBase *split_grid) : full_grid(_full_grid)
+  {
+    nd = full_grid->_ndimension;
+    GRID_ASSERT(split_grid->_ndimension == nd);
+    ratio.resize(nd);
+    fldims = full_grid->_ldimensions;
+    Coordinate sprocs = split_grid->_processors;
+    Coordinate fcoor  = full_grid->_processor_coor;
+    nvector = 1;
+    for(int d=0;d<nd;d++){
+      GRID_ASSERT(full_grid->_gdimensions[d] == split_grid->_gdimensions[d]);
+      GRID_ASSERT(full_grid->_fdimensions[d] == split_grid->_fdimensions[d]);
+      GRID_ASSERT( (full_grid->_processors[d] % sprocs[d]) == 0 );
+      ratio[d] = full_grid->_processors[d] / sprocs[d];
+      GRID_ASSERT(split_grid->_ldimensions[d] == fldims[d]*ratio[d]);
+      nvector *= ratio[d];
+    }
+    GRID_ASSERT(nvector*split_grid->_Nprocessors == full_grid->_Nprocessors);
+
+    Field.resize(nvector);
+    FieldRank.resize(nvector);
+    Box.resize(nvector);
+    BoxRank.resize(nvector);
+    for(int k=0;k<nvector;k++){
+      Coordinate kc(nd);
+      Coordinate q(nd);
+      Coordinate dest(nd);
+      Coordinate o(nd);
+      Coordinate from(nd);
+      Lexicographic::CoorFromIndex(kc,k,ratio);
+      for(int d=0;d<nd;d++){
+        q[d]    = (fcoor[d] % ratio[d] + kc[d]) % ratio[d];
+        dest[d] = q[d]*sprocs[d] + fcoor[d] / ratio[d];
+        o[d]    = (fcoor[d] / sprocs[d] - kc[d] + ratio[d]) % ratio[d];
+        from[d] = (fcoor[d] % sprocs[d])*ratio[d] + o[d];
+      }
+      Lexicographic::IndexFromCoor(q,Field[k],ratio);
+      Lexicographic::IndexFromCoor(o,Box[k],ratio);
+      FieldRank[k] = full_grid->RankFromProcessorCoor(dest);
+      BoxRank[k]   = full_grid->RankFromProcessorCoor(from);
+    }
+  }
+
+  // One step's exchange; returns the bytes that left this rank
+  uint64_t Exchange(void *xmit,int dest,void *recv,int from,uint64_t bytes)
+  {
+    int me = full_grid->ThisRank();
+    // The send map is a permutation and the receive map its inverse: self pairs with self
+    GRID_ASSERT( (dest == me) == (from == me) );
+    if ( dest == me ) {
+      acceleratorCopyDeviceToDevice(xmit,recv,bytes);
+      return 0;
+    }
+    GRID_ASSERT( (bytes % sizeof(int32_t)) == 0 );
+    GRID_ASSERT( bytes/sizeof(int32_t) < (1ULL<<31) );   // SendToRecvFrom counts int32 words
+#ifdef ACCELERATOR_AWARE_MPI
+    full_grid->SendToRecvFrom(xmit,dest,recv,from,bytes);
+#else
+    std::vector<char> hxmit(bytes);
+    std::vector<char> hrecv(bytes);
+    acceleratorCopyFromDevice(xmit,&hxmit[0],bytes);
+    full_grid->SendToRecvFrom(&hxmit[0],dest,&hrecv[0],from,bytes);
+    acceleratorCopyToDevice(&hrecv[0],recv,bytes);
+#endif
+    return bytes;
+  }
+};
+
+// Common body: full[v] may alias one another (the single field overload)
 template<class Vobj>
-void Grid_split(std::vector<Lattice<Vobj> > & full,Lattice<Vobj>   & split)
+void Grid_split(const std::vector<const Lattice<Vobj> *> &full,Lattice<Vobj> &split)
 {
   typedef typename Vobj::scalar_object Sobj;
 
-  int full_vecs   = full.size();
-
-  GRID_ASSERT(full_vecs>=1);
-
-  GridBase * full_grid = full[0].Grid();
+  GRID_ASSERT(full.size()>=1);
+  GridBase *full_grid  = full[0]->Grid();
   GridBase *split_grid = split.Grid();
 
-  int       ndim  = full_grid->_ndimension;
-  int  full_nproc = full_grid->_Nprocessors;
-  int split_nproc =split_grid->_Nprocessors;
-
-  ////////////////////////////////
-  // Checkerboard management
-  ////////////////////////////////
-  int cb = full[0].Checkerboard();
+  int cb = full[0]->Checkerboard();
   split.Checkerboard() = cb;
-
-  //////////////////////////////
-  // Checks
-  //////////////////////////////
-  GRID_ASSERT(full_grid->_ndimension==split_grid->_ndimension);
-  for(int n=0;n<full_vecs;n++){
-    GRID_ASSERT(full[n].Checkerboard() == cb);
-    for(int d=0;d<ndim;d++){
-      GRID_ASSERT(full[n].Grid()->_gdimensions[d]==split.Grid()->_gdimensions[d]);
-      GRID_ASSERT(full[n].Grid()->_fdimensions[d]==split.Grid()->_fdimensions[d]);
-    }
+  for(int n=0;n<full.size();n++){
+    GRID_ASSERT(full[n]->Grid() == full_grid);
+    GRID_ASSERT(full[n]->Checkerboard() == cb);
   }
 
-  int   nvector   =full_nproc/split_nproc; 
-  GRID_ASSERT(nvector*split_nproc==full_nproc);
-  GRID_ASSERT(nvector == full_vecs);
+  double t_start = usecond();
+  GridSplitSchedule sched(full_grid,split_grid);
+  GRID_ASSERT(sched.nvector == full.size());
 
-  Coordinate ratio(ndim);
-  for(int d=0;d<ndim;d++){
-    ratio[d] = full_grid->_processors[d]/ split_grid->_processors[d];
+  uint64_t fvol  = full_grid->lSites();
+  uint64_t bytes = fvol*sizeof(Sobj);
+  deviceVector<Sobj> staged(split_grid->lSites());
+  deviceVector<Sobj> block(fvol);
+  Sobj *staged_p = &staged[0];
+  Sobj *block_p  = &block[0];
+  double t_alloc = usecond() - t_start;
+
+  double t_pack = 0;
+  double t_comm = 0;
+  uint64_t offrank = 0;
+  for(int k=0;k<sched.nvector;k++){
+    double t0 = usecond();
+    GridSplitUnvectorise(*full[sched.Field[k]],block_p,full_grid->_ldimensions);
+    double t1 = usecond();
+    offrank += sched.Exchange((void *)block_p,sched.FieldRank[k],
+                              (void *)&staged_p[sched.Box[k]*fvol],sched.BoxRank[k],bytes);
+    double t2 = usecond();
+    t_pack += t1-t0;
+    t_comm += t2-t1;
   }
 
-  uint64_t lsites = full_grid->lSites();
-  uint64_t     sz = lsites * nvector;
-  std::vector<Sobj> tmpdata(sz);
-  std::vector<Sobj> alldata(sz);
-  std::vector<Sobj> scalardata(lsites); 
+  double t0 = usecond();
+  GridSplitVectorise(staged_p,split,sched.fldims);
+  double t_unpack = usecond() - t0;
+  double t_total  = usecond() - t_start;
 
-  for(int v=0;v<nvector;v++){
-    unvectorizeToLexOrdArray(scalardata,full[v]);    
-    thread_for(site,lsites,{
-      alldata[v*lsites+site] = scalardata[site];
-    });
+  std::cout << GridLogPerformance << "Grid_split: " << sched.nvector << " x " << bytes/1.0e6
+            << " MB/rank total " << t_total/1.0e6 << " s : alloc " << t_alloc/1.0e6
+            << " pack " << t_pack/1.0e6
+            << " exchange " << t_comm/1.0e6 << " (" << offrank/1.0e6 << " MB off rank, "
+            << offrank/(t_comm+1.0e-9)/1.0e3 << " GB/s)"
+            << " unpack " << t_unpack/1.0e6 << std::endl;
+  // Staging buffers are still live here, so this is the high-water point of the call
+  if ( GridLogPerformance.isActive() ) {
+    HostMemoryReport(full_grid,GridLogPerformance,"Grid_split");
   }
-
-  int nvec = nvector; // Counts down to 1 as we collapse dims
-  Coordinate ldims = full_grid->_ldimensions;
-
-  for(int d=ndim-1;d>=0;d--){
-
-    if ( ratio[d] != 1 ) {
-
-      full_grid ->AllToAll(d,alldata,tmpdata);
-      if ( split_grid->_processors[d] > 1 ) {
-	alldata=tmpdata;
-	split_grid->AllToAll(d,alldata,tmpdata);
-      }
-
-      auto rdims = ldims; 
-      auto     M = ratio[d];
-      auto rsites= lsites*M;// increases rsites by M
-      nvec      /= M;       // Reduce nvec by subdivision factor
-      rdims[d]  *= M;       // increase local dim by same factor
-
-      int sP =   split_grid->_processors[d];
-      int fP =    full_grid->_processors[d];
-
-      int fvol   = lsites;
-      
-      int chunk  = (nvec*fvol)/sP;          GRID_ASSERT(chunk*sP == nvec*fvol);
-
-      // Loop over reordered data post A2A
-      thread_for(c, chunk, {
-	Coordinate coor(ndim);
-	for(int m=0;m<M;m++){
-	  for(int s=0;s<sP;s++){
-	    
-	    // addressing; use lexico
-	    int lex_r;
-	    uint64_t lex_c        = c+chunk*m+chunk*M*s;
-	    uint64_t lex_fvol_vec = c+chunk*s;
-	    uint64_t lex_fvol     = lex_fvol_vec%fvol;
-	    uint64_t lex_vec      = lex_fvol_vec/fvol;
-
-	    // which node sets an adder to the coordinate
-	    Lexicographic::CoorFromIndex(coor, lex_fvol, ldims);	  
-	    coor[d] += m*ldims[d];
-	    Lexicographic::IndexFromCoor(coor, lex_r, rdims);	  
-	    lex_r += lex_vec * rsites;
-
-	    // LexicoFind coordinate & vector number within split lattice
-	    alldata[lex_r] = tmpdata[lex_c];
-
-	  }
-	}
-      });
-      ldims[d]*= ratio[d];
-      lsites  *= ratio[d];
-
-    }
-  }
-  vectorizeFromLexOrdArray(alldata,split);    
 }
 
+template<class Vobj>
+void Grid_split(std::vector<Lattice<Vobj> > & full,Lattice<Vobj>   & split)
+{
+  std::vector<const Lattice<Vobj> *> full_p(full.size());
+  for(int n=0;n<full.size();n++){
+    full_p[n] = &full[n];
+  }
+  Grid_split(full_p,split);
+}
+
+// The same field to every partition
 template<class Vobj>
 void Grid_split(Lattice<Vobj> &full,Lattice<Vobj>   & split)
 {
   int nvector = full.Grid()->_Nprocessors / split.Grid()->_Nprocessors;
-  std::vector<Lattice<Vobj> > full_v(nvector,full.Grid());
-  for(int n=0;n<nvector;n++){
-    full_v[n] = full;
-  }
-  Grid_split(full_v,split);
+  std::vector<const Lattice<Vobj> *> full_p(nvector,&full);
+  Grid_split(full_p,split);
 }
 
 template<class Vobj>
@@ -1725,119 +1828,57 @@ void Grid_unsplit(std::vector<Lattice<Vobj> > & full,Lattice<Vobj>   & split)
 {
   typedef typename Vobj::scalar_object Sobj;
 
-  int full_vecs   = full.size();
-
-  GRID_ASSERT(full_vecs>=1);
-
-  GridBase * full_grid = full[0].Grid();
+  GRID_ASSERT(full.size()>=1);
+  GridBase *full_grid  = full[0].Grid();
   GridBase *split_grid = split.Grid();
 
-  int       ndim  = full_grid->_ndimension;
-  int  full_nproc = full_grid->_Nprocessors;
-  int split_nproc =split_grid->_Nprocessors;
-
-  ////////////////////////////////
-  // Checkerboard management
-  ////////////////////////////////
   int cb = full[0].Checkerboard();
   split.Checkerboard() = cb;
-
-  //////////////////////////////
-  // Checks
-  //////////////////////////////
-  GRID_ASSERT(full_grid->_ndimension==split_grid->_ndimension);
-  for(int n=0;n<full_vecs;n++){
+  for(int n=0;n<full.size();n++){
+    GRID_ASSERT(full[n].Grid() == full_grid);
     GRID_ASSERT(full[n].Checkerboard() == cb);
-    for(int d=0;d<ndim;d++){
-      GRID_ASSERT(full[n].Grid()->_gdimensions[d]==split.Grid()->_gdimensions[d]);
-      GRID_ASSERT(full[n].Grid()->_fdimensions[d]==split.Grid()->_fdimensions[d]);
-    }
   }
 
-  int   nvector   =full_nproc/split_nproc; 
-  GRID_ASSERT(nvector*split_nproc==full_nproc);
-  GRID_ASSERT(nvector == full_vecs);
+  double t_start = usecond();
+  GridSplitSchedule sched(full_grid,split_grid);
+  GRID_ASSERT(sched.nvector == full.size());
 
-  Coordinate ratio(ndim);
-  for(int d=0;d<ndim;d++){
-    ratio[d] = full_grid->_processors[d]/ split_grid->_processors[d];
+  uint64_t fvol  = full_grid->lSites();
+  uint64_t bytes = fvol*sizeof(Sobj);
+  deviceVector<Sobj> staged(split_grid->lSites());
+  deviceVector<Sobj> block(fvol);
+  Sobj *staged_p = &staged[0];
+  Sobj *block_p  = &block[0];
+  double t_alloc = usecond() - t_start;
+
+  double t0 = usecond();
+  GridSplitUnvectorise(split,staged_p,sched.fldims);
+  double t_pack = usecond() - t0;
+
+  double t_comm   = 0;
+  double t_unpack = 0;
+  uint64_t offrank = 0;
+  for(int k=0;k<sched.nvector;k++){
+    double t1 = usecond();
+    offrank += sched.Exchange((void *)&staged_p[sched.Box[k]*fvol],sched.BoxRank[k],
+                              (void *)block_p,sched.FieldRank[k],bytes);
+    double t2 = usecond();
+    GridSplitVectorise(block_p,full[sched.Field[k]],full_grid->_ldimensions);
+    double t3 = usecond();
+    t_comm   += t2-t1;
+    t_unpack += t3-t2;
   }
+  double t_total = usecond() - t_start;
 
-  uint64_t lsites = full_grid->lSites();
-  uint64_t     sz = lsites * nvector;
-  std::vector<Sobj> tmpdata(sz);
-  std::vector<Sobj> alldata(sz);
-  std::vector<Sobj> scalardata(lsites); 
-
-  unvectorizeToLexOrdArray(alldata,split);    
-
-  /////////////////////////////////////////////////////////////////
-  // Start from split grid and work towards full grid
-  /////////////////////////////////////////////////////////////////
-
-  int nvec = 1;
-  uint64_t rsites        = split_grid->lSites();
-  Coordinate rdims = split_grid->_ldimensions;
-
-  for(int d=0;d<ndim;d++){
-
-    if ( ratio[d] != 1 ) {
-
-      auto     M = ratio[d];
-
-      int sP =   split_grid->_processors[d];
-      int fP =    full_grid->_processors[d];
-      
-      auto ldims = rdims;  ldims[d]  /= M;  // Decrease local dims by same factor
-      auto lsites= rsites/M;                // Decreases rsites by M
-      
-      int fvol   = lsites;
-      int chunk  = (nvec*fvol)/sP;          GRID_ASSERT(chunk*sP == nvec*fvol);
-	
-      {
-	// Loop over reordered data post A2A
-	thread_for(c, chunk,{
-	  Coordinate coor(ndim);
-	  for(int m=0;m<M;m++){
-	    for(int s=0;s<sP;s++){
-
-	      // addressing; use lexico
-	      int lex_r;
-	      uint64_t lex_c = c+chunk*m+chunk*M*s;
-	      uint64_t lex_fvol_vec = c+chunk*s;
-	      uint64_t lex_fvol     = lex_fvol_vec%fvol;
-	      uint64_t lex_vec      = lex_fvol_vec/fvol;
-	      
-	      // which node sets an adder to the coordinate
-	      Lexicographic::CoorFromIndex(coor, lex_fvol, ldims);	  
-	      coor[d] += m*ldims[d];
-	      Lexicographic::IndexFromCoor(coor, lex_r, rdims);	  
-	      lex_r += lex_vec * rsites;
-	      
-	      // LexicoFind coordinate & vector number within split lattice
-	      tmpdata[lex_c] = alldata[lex_r];
-	    }
-	  }
-        });
-      }
-
-      if ( split_grid->_processors[d] > 1 ) {
-	split_grid->AllToAll(d,tmpdata,alldata);
-	tmpdata=alldata;
-      }
-      full_grid ->AllToAll(d,tmpdata,alldata);
-      rdims[d]/= M;
-      rsites  /= M;
-      nvec    *= M;       // Increase nvec by subdivision factor
-    }
-  }
-
-  lsites = full_grid->lSites();
-  for(int v=0;v<nvector;v++){
-    thread_for(site, lsites,{
-      scalardata[site] = alldata[v*lsites+site];
-    });
-    vectorizeFromLexOrdArray(scalardata,full[v]);    
+  std::cout << GridLogPerformance << "Grid_unsplit: " << sched.nvector << " x " << bytes/1.0e6
+            << " MB/rank total " << t_total/1.0e6 << " s : alloc " << t_alloc/1.0e6
+            << " pack " << t_pack/1.0e6
+            << " exchange " << t_comm/1.0e6 << " (" << offrank/1.0e6 << " MB off rank, "
+            << offrank/(t_comm+1.0e-9)/1.0e3 << " GB/s)"
+            << " unpack " << t_unpack/1.0e6 << std::endl;
+  // Staging buffers are still live here, so this is the high-water point of the call
+  if ( GridLogPerformance.isActive() ) {
+    HostMemoryReport(full_grid,GridLogPerformance,"Grid_unsplit");
   }
 }
 

Grid/perfmon/HostMemory.h

@@ -0,0 +1,227 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid
+
+    Source file: ./lib/perfmon/HostMemory.h
+
+    Copyright (C) 2026
+
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+
+    This program is free software; you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation; either version 2 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License along
+    with this program; if not, write to the Free Software Foundation, Inc.,
+    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+    See the full license in the file "LICENSE" in the top level distribution directory
+*************************************************************************************/
+/*  END LEGAL */
+#pragma once
+
+#include <sys/resource.h>
+#include <unistd.h>
+#include <cctype>
+#include <fstream>
+#ifdef __APPLE__
+#include <mach/mach.h>
+#endif
+
+NAMESPACE_BEGIN(Grid);
+
+/////////////////////////////////////////////////////////////////////////////////////////////
+// Host memory as the OOM killer sees it, plus the device allocator footprint.
+// Targets are Apple or Linux; on Apple only the RSS and device fields are filled.
+//
+//  RSS      : this process, now and its high-water mark (getrusage).
+//  Node     : MemAvailable and MemTotal from /proc/meminfo (Linux).
+//  Cgroup   : the nearest enclosing memory cgroup with a finite limit (the one the batch
+//             system enforces), its usage now, its high-water mark and the limit (Linux,
+//             v2 or v1 hierarchy). Cgroup usage includes page cache, tmpfs and kernel memory
+//             charged to the job, none of which appear in any process RSS.
+//  Device   : MemoryManager Lattice footprint on the device and its free-block cache. Device
+//             allocations outside the MemoryManager (comms buffers) are not included.
+//
+// All values in GB; -1 where the source does not exist on this platform or kernel.
+/////////////////////////////////////////////////////////////////////////////////////////////
+struct HostMemoryStatus
+{
+  RealD RSS           = -1;
+  RealD RSSPeak       = -1;
+  RealD NodeAvailable = -1;
+  RealD NodeTotal     = -1;
+  RealD CgroupCurrent = -1;
+  RealD CgroupPeak    = -1;
+  RealD CgroupLimit   = -1;
+  RealD DeviceLattice = -1;
+  RealD DeviceCache   = -1;
+};
+
+#ifndef __APPLE__ // Linux
+
+// One number from a file; -1 if the file is absent or holds no number (e.g. "max")
+inline RealD HostMemoryReadBytes(const std::string &file)
+{
+  std::ifstream f(file);
+  if ( !f.good() ) {
+    return -1;
+  }
+  std::string word;
+  f >> word;
+  if ( word.empty() || !isdigit(word[0]) ) {
+    return -1;
+  }
+  return std::stod(word);
+}
+
+// Value of a "Key: N kB" line in /proc/meminfo, in bytes; -1 if absent
+inline RealD HostMemoryMeminfo(const std::string &key)
+{
+  std::ifstream f("/proc/meminfo");
+  std::string name;
+  RealD value;
+  std::string unit;
+  while ( f >> name >> value >> unit ) {
+    if ( name == key + ":" ) {
+      return value*1024.0;
+    }
+  }
+  return -1;
+}
+
+// Fill the cgroup fields from the nearest enclosing memory cgroup with a finite limit, or
+// from this process's own cgroup (limit left at -1) if no ancestor has one
+inline void HostMemoryCgroup(HostMemoryStatus &s)
+{
+  std::ifstream f("/proc/self/cgroup");
+  std::string line;
+  std::string root;
+  std::string path;
+  std::string current;
+  std::string peak;
+  std::string limit;
+  while ( std::getline(f,line) ) {
+    size_t c1 = line.find(':');
+    size_t c2 = line.find(':',c1+1);
+    if ( c1 == std::string::npos || c2 == std::string::npos ) {
+      continue;
+    }
+    std::string controllers = line.substr(c1+1,c2-c1-1);
+    // v2 unified hierarchy: "0::/path"
+    if ( controllers.empty() && root.empty() ) {
+      root    = "/sys/fs/cgroup";
+      path    = line.substr(c2+1);
+      current = "memory.current";
+      peak    = "memory.peak";
+      limit   = "memory.max";
+    }
+    // v1 memory controller: "N:memory:/path", takes precedence
+    if ( (","+controllers+",").find(",memory,") != std::string::npos ) {
+      root    = "/sys/fs/cgroup/memory";
+      path    = line.substr(c2+1);
+      current = "memory.usage_in_bytes";
+      peak    = "memory.max_usage_in_bytes";
+      limit   = "memory.limit_in_bytes";
+      break;
+    }
+  }
+  if ( root.empty() ) {
+    return;
+  }
+
+  // v1 reports "no limit" as a number near 2^63
+  const RealD unlimited = 1.0e18;
+  std::string dir = path;
+  while ( true ) {
+    std::string base = root + dir + "/";
+    RealD lim = HostMemoryReadBytes(base+limit);
+    if ( lim > 0 && lim < unlimited ) {
+      s.CgroupCurrent = HostMemoryReadBytes(base+current)/1.0e9;
+      s.CgroupPeak    = HostMemoryReadBytes(base+peak)/1.0e9;
+      s.CgroupLimit   = lim/1.0e9;
+      return;
+    }
+    if ( dir.empty() || dir == "/" ) {
+      break;
+    }
+    dir = dir.substr(0,dir.rfind('/'));
+  }
+  std::string base = root + path + "/";
+  s.CgroupCurrent = HostMemoryReadBytes(base+current)/1.0e9;
+  s.CgroupPeak    = HostMemoryReadBytes(base+peak)/1.0e9;
+}
+
+#endif
+
+// This rank only; no communication
+inline HostMemoryStatus HostMemoryQuery(void)
+{
+  HostMemoryStatus s;
+
+  struct rusage ru;
+  getrusage(RUSAGE_SELF,&ru);
+#ifdef __APPLE__
+  s.RSSPeak = ru.ru_maxrss/1.0e9;                // bytes on macOS
+  mach_task_basic_info_data_t info;
+  mach_msg_type_number_t count = MACH_TASK_BASIC_INFO_COUNT;
+  if ( task_info(mach_task_self(),MACH_TASK_BASIC_INFO,(task_info_t)&info,&count) == KERN_SUCCESS ) {
+    s.RSS = info.resident_size/1.0e9;
+  }
+#else // Linux
+  s.RSSPeak = ru.ru_maxrss*1024.0/1.0e9;         // kilobytes on Linux
+  std::ifstream statm("/proc/self/statm");
+  long pages    = 0;
+  long resident = 0;
+  if ( statm >> pages >> resident ) {
+    s.RSS = resident*(RealD)sysconf(_SC_PAGESIZE)/1.0e9;
+  }
+  RealD avail = HostMemoryMeminfo("MemAvailable");
+  RealD total = HostMemoryMeminfo("MemTotal");
+  if ( avail >= 0 ) {
+    s.NodeAvailable = avail/1.0e9;
+  }
+  if ( total >= 0 ) {
+    s.NodeTotal = total/1.0e9;
+  }
+  HostMemoryCgroup(s);
+#endif
+
+  s.DeviceLattice = MemoryManager::DeviceBytes/1.0e9;
+  s.DeviceCache   = MemoryManager::DeviceCacheBytes()/1.0e9;
+  return s;
+}
+
+/////////////////////////////////////////////////////////////////////////////////////////////
+// Collective over grid. One line on log: maxima over ranks, except node MemAvailable which
+// is the minimum over ranks (the tightest node).
+/////////////////////////////////////////////////////////////////////////////////////////////
+inline void HostMemoryReport(GridBase *grid,GridLogger &log,const std::string &label)
+{
+  HostMemoryStatus s = HostMemoryQuery();
+  RealD minus_available = -s.NodeAvailable;
+  grid->GlobalMax(s.RSS);
+  grid->GlobalMax(s.RSSPeak);
+  grid->GlobalMax(minus_available);
+  grid->GlobalMax(s.NodeTotal);
+  grid->GlobalMax(s.CgroupCurrent);
+  grid->GlobalMax(s.CgroupPeak);
+  grid->GlobalMax(s.CgroupLimit);
+  grid->GlobalMax(s.DeviceLattice);
+  grid->GlobalMax(s.DeviceCache);
+  std::cout << log << "HOSTMEM " << label
+            << " : rank RSS " << s.RSS << " peak " << s.RSSPeak
+            << " ; node available " << -minus_available << " of " << s.NodeTotal
+            << " ; cgroup " << s.CgroupCurrent << " peak " << s.CgroupPeak << " limit " << s.CgroupLimit
+            << " ; device lattice " << s.DeviceLattice << " cache " << s.DeviceCache
+            << " GB" << std::endl;
+}
+
+NAMESPACE_END(Grid);

Grid/qcd/action/fermion/FermionOperator.h

@@ -70,6 +70,8 @@ public:
     GRID_ASSERT(U != nullptr);
     GRID_ASSERT(mpi_split.size() == Nd);
 
+    double t0 = usecond();
+
     SplitOperator<FermionField> *split = new SplitOperator<FermionField>();
 
     split->GaugeGrid   = new GridCartesian(U->FullDimensions(),U->_simd_layout,mpi_split,*U);
@@ -85,6 +87,7 @@ public:
       split->FermionGrid   = split->GaugeGrid;
       split->FermionRBGrid = split->GaugeRBGrid;
     }
+    std::cout << GridLogPerformance << "MakeSplitGrids: " << (usecond()-t0)/1.0e6 << " s" << std::endl;
     return split;
   }
 

Grid/qcd/action/fermion/WilsonFermion5D.h

@@ -192,9 +192,14 @@ public:
   // Redistribute this operator's doubled gauge field onto clone's split grid. Collective.
   void SplitDoubledGaugeInto(WilsonFermion5D<Impl> &clone)
   {
+    double t0 = usecond();
     DoubledGaugeField sUmu(clone.GaugeGrid());
     Grid_split(Umu,sUmu);
+    double t1 = usecond();
     clone.ImportDoubledGauge(sUmu);
+    double t2 = usecond();
+    std::cout << GridLogPerformance << "SplitDoubledGaugeInto: Grid_split " << (t1-t0)/1.0e6
+              << " s import " << (t2-t1)/1.0e6 << " s" << std::endl;
   }
 
   ///////////////////////////////////////////////////////////////

tests/solver/Test_dwf_mrhs_cg.cc

@@ -56,14 +56,12 @@ int main (int argc, char ** argv)
   /////////////////////////////////////////////
   // Split into 1^4 mpi communicators
   /////////////////////////////////////////////
-  for(int i=0;i<argc;i++){
-    if(std::string(argv[i]) == "--split"){
-      for(int k=0;k<mpi_layout.size();k++){
-	std::stringstream ss; 
-	ss << argv[i+1+k]; 
-	ss >> mpi_split[k];
-      }
-      break;
+  if ( GridCmdOptionExists(argv,argv+argc,"--split") ) {
+    std::vector<int> split;
+    GridCmdOptionIntVector(GridCmdOptionPayload(argv,argv+argc,"--split"),split);
+    GRID_ASSERT(split.size() == mpi_layout.size());
+    for(int k=0;k<mpi_layout.size();k++){
+      mpi_split[k] = split[k];
     }
   }
 

tests/solver/Test_dwf_mrhs_cg_mpi.cc

@@ -59,14 +59,12 @@ int main (int argc, char ** argv)
   // Split into 1^4 mpi communicators
   /////////////////////////////////////////////
 
-  for(int i=0;i<argc;i++){
-    if(std::string(argv[i]) == "--split"){
-      for(int k=0;k<mpi_layout.size();k++){
-	std::stringstream ss; 
-	ss << argv[i+1+k]; 
-	ss >> mpi_split[k];
-      }
-      break;
+  if ( GridCmdOptionExists(argv,argv+argc,"--split") ) {
+    std::vector<int> split;
+    GridCmdOptionIntVector(GridCmdOptionPayload(argv,argv+argc,"--split"),split);
+    GRID_ASSERT(split.size() == mpi_layout.size());
+    for(int k=0;k<mpi_layout.size();k++){
+      mpi_split[k] = split[k];
     }
   }
 

tests/solver/Test_mobius_bcg.cc

@@ -63,14 +63,12 @@ int main (int argc, char ** argv)
   // Split into 1^4 mpi communicators
   /////////////////////////////////////////////
 
-  for(int i=0;i<argc;i++){
-    if(std::string(argv[i]) == "--split"){
-      for(int k=0;k<mpi_layout.size();k++){
-	std::stringstream ss; 
-	ss << argv[i+1+k]; 
-	ss >> mpi_split[k];
-      }
-      break;
+  if ( GridCmdOptionExists(argv,argv+argc,"--split") ) {
+    std::vector<int> split;
+    GridCmdOptionIntVector(GridCmdOptionPayload(argv,argv+argc,"--split"),split);
+    GRID_ASSERT(split.size() == mpi_layout.size());
+    for(int k=0;k<mpi_layout.size();k++){
+      mpi_split[k] = split[k];
     }
   }
 

tests/solver/Test_split_grid.cc

@@ -56,14 +56,12 @@ int main (int argc, char ** argv)
   // Split into 1^4 mpi communicators
   /////////////////////////////////////////////
 
-  for(int i=0;i<argc;i++){
-    if(std::string(argv[i]) == "--split"){
-      for(int k=0;k<mpi_layout.size();k++){
-	std::stringstream ss; 
-	ss << argv[i+1+k]; 
-	ss >> mpi_split[k];
-      }
-      break;
+  if ( GridCmdOptionExists(argv,argv+argc,"--split") ) {
+    std::vector<int> split;
+    GridCmdOptionIntVector(GridCmdOptionPayload(argv,argv+argc,"--split"),split);
+    GRID_ASSERT(split.size() == mpi_layout.size());
+    for(int k=0;k<mpi_layout.size();k++){
+      mpi_split[k] = split[k];
     }
   }
 

tests/solver/Test_split_mixedprec_batched.cc

@@ -145,16 +145,19 @@ void CheckSplitBatched(const std::string &name,
     sol_split[i] = Zero();
   }
 
-  MixedPrecisionConjugateGradientBatched<FieldD,FieldF> mCG(TOLERANCE,10000,50,1000,grid_f,Linop_f,Linop_d);
-
-  std::cout << GridLogMessage << name << ": unsplit batched solve" << std::endl;
-  mCG.BatchedSplit     = Coordinate();
-  mCG.BatchedSplitNode = false;
-  mCG(src,sol_ref);
-
-  std::cout << GridLogMessage << name << ": split batched solve, partition layout " << layout << std::endl;
-  mCG.BatchedSplit = layout;
-  mCG(src,sol_split);
+  {
+    std::cout << GridLogMessage << name << ": unsplit batched solve" << std::endl;
+    MixedPrecisionConjugateGradientBatched<FieldD,FieldF> mCG(TOLERANCE,10000,50,1000,grid_f,Linop_f,Linop_d,
+                                                              true,Coordinate(),false);
+    mCG(src,sol_ref);
+  }
+  {
+    std::cout << GridLogMessage << name << ": split batched solve, partition layout " << layout << std::endl;
+    MixedPrecisionConjugateGradientBatched<FieldD,FieldF> mCG(TOLERANCE,10000,50,1000,grid_f,Linop_f,Linop_d,
+                                                              true,layout,false);
+    GRID_ASSERT( (mCG.Partitions == 1) || (mCG.split != nullptr) );
+    mCG(src,sol_split);
+  }
 
   FieldD Msol(src[0].Grid());
   Msol.Checkerboard() = cb;

tests/solver/Test_split_mobius_batched.cc

@@ -93,6 +93,7 @@ void ReportMemory(GridBase *grid,const std::string &phase)
             << " : host RSS " << rss << " GB, peak " << peak
             << " GB; allocator cache host " << hostcache << " GB, device " << devcache
             << " GB (max over ranks)" << std::endl;
+  HostMemoryReport(grid,GridLogMessage,phase);
 }
 
 typedef LatticeFermionD FieldD;
@@ -212,25 +213,24 @@ int main (int argc, char ** argv)
 
   ReportMemory(UGrid_d,"operators and sources ready");
 
-  MixedPrecisionConjugateGradientBatched<FieldD,FieldF> mCG(tol,10000,50,10000,FrbGrid_f,Linop_f,Linop_d);
-
-  Coordinate split     = mCG.BatchedSplit;
-  bool       splitnode = mCG.BatchedSplitNode;
-
   if ( unsplit ) {
-    mCG.BatchedSplit     = Coordinate();
-    mCG.BatchedSplitNode = false;
+    MixedPrecisionConjugateGradientBatched<FieldD,FieldF> mCG(tol,10000,50,10000,FrbGrid_f,Linop_f,Linop_d,
+                                                              true,Coordinate(),false);
     SolveAndReport("UNSPLIT",mCG,Linop_d,src,sol);
     ReportMemory(UGrid_d,"after unsplit solve");
   }
 
-  mCG.BatchedSplit     = split;
-  mCG.BatchedSplitNode = splitnode;
-  // Repeated split solves expose allocations not released between calls
-  for(int r=0;r<repeat;r++){
-    SolveAndReport("SPLIT",mCG,Linop_d,src,sol);
-    ReportMemory(UGrid_d,"after split solve "+std::to_string(r));
+  if ( repeat > 0 ) {
+    // Split layout from the command line; the split operator is cloned here, once
+    MixedPrecisionConjugateGradientBatched<FieldD,FieldF> mCG(tol,10000,50,10000,FrbGrid_f,Linop_f,Linop_d);
+    ReportMemory(UGrid_d,"after split solver construction");
+    // Repeated split solves reuse the clone and expose allocations not released between calls
+    for(int r=0;r<repeat;r++){
+      SolveAndReport("SPLIT",mCG,Linop_d,src,sol);
+      ReportMemory(UGrid_d,"after split solve "+std::to_string(r));
+    }
   }
+  ReportMemory(UGrid_d,"after split solver destruction");
 
   Grid_finalize();
 }