Merge 570b72a47b into b50fb34e71

Perf on Aurora
Fastest run config on Aurora to date
2025-06-13 04:37:05 +01:00 · 2025-02-07 15:36:21 -05:00 · 2025-02-01 18:39:34 +00:00 · 2025-02-01 18:08:40 +00:00 · 2025-02-01 11:04:13 -05:00 · 2025-01-31 16:32:36 +00:00
16 changed files with 2135 additions and 127 deletions
--- a/Grid/algorithms/deflation/MultiRHSBlockCGLinalg.h
+++ b/Grid/algorithms/deflation/MultiRHSBlockCGLinalg.h
@ -144,11 +144,11 @@ public:
      acceleratorCopyDeviceToDevice(&BLAS_Y[offset],&y_v[0],sizeof(scalar_object)*vol);
    }    
    RealD t4 = usecond();
-    std::cout << "MulMatrix alloc    took "<< t1-t0<<" us"<<std::endl;
-    std::cout << "MulMatrix preamble took "<< t2-t1<<" us"<<std::endl;
-    std::cout << "MulMatrix blas     took "<< t3-t2<<" us"<<std::endl;
-    std::cout << "MulMatrix copy     took "<< t4-t3<<" us"<<std::endl;
-    std::cout << "MulMatrix total "<< t4-t0<<" us"<<std::endl;
+    std::cout <<GridLogPerformance << "MulMatrix alloc    took "<< t1-t0<<" us"<<std::endl;
+    std::cout <<GridLogPerformance<< "MulMatrix preamble took "<< t2-t1<<" us"<<std::endl;
+    std::cout <<GridLogPerformance<< "MulMatrix blas     took "<< t3-t2<<" us"<<std::endl;
+    std::cout <<GridLogPerformance<< "MulMatrix copy     took "<< t4-t3<<" us"<<std::endl;
+    std::cout <<GridLogPerformance<< "MulMatrix total "<< t4-t0<<" us"<<std::endl;
  }
  
  void InnerProductMatrix(Eigen::MatrixXcd &m , const std::vector<Field> &X, const std::vector<Field> &Y)
@ -242,16 +242,16 @@ public:
    RealD flops = 8.0*M*N*K;
    flops = flops/(t4-t3)/1.e3;
    bytes = bytes/(t4-t3)/1.e3;
-    std::cout << "InnerProductMatrix m,n,k "<< M<<","<<N<<","<<K<<std::endl;
-    std::cout << "InnerProductMatrix alloc t1 "<< t1-t0<<" us"<<std::endl;
-    std::cout << "InnerProductMatrix cp    t2 "<< t2-t1<<" us"<<std::endl;
-    std::cout << "InnerProductMatrix setup t3 "<< t3-t2<<" us"<<std::endl;
-    std::cout << "InnerProductMatrix blas t4 "<< t4-t3<<" us"<<std::endl;
-    std::cout << "InnerProductMatrix blas    "<< flops<<" GF/s"<<std::endl;
-    std::cout << "InnerProductMatrix blas    "<< bytes<<" GB/s"<<std::endl;
-    std::cout << "InnerProductMatrix gsum t5 "<< t5-t4<<" us"<<std::endl;
-    std::cout << "InnerProductMatrix cp   t6 "<< t6-t5<<" us"<<std::endl;
-    std::cout << "InnerProductMatrix took "<< t6-t0<<" us"<<std::endl;
+    std::cout <<GridLogPerformance<< "InnerProductMatrix m,n,k "<< M<<","<<N<<","<<K<<std::endl;
+    std::cout <<GridLogPerformance<< "InnerProductMatrix alloc t1 "<< t1-t0<<" us"<<std::endl;
+    std::cout <<GridLogPerformance<< "InnerProductMatrix cp    t2 "<< t2-t1<<" us"<<std::endl;
+    std::cout <<GridLogPerformance<< "InnerProductMatrix setup t3 "<< t3-t2<<" us"<<std::endl;
+    std::cout <<GridLogPerformance<< "InnerProductMatrix blas t4 "<< t4-t3<<" us"<<std::endl;
+    std::cout <<GridLogPerformance<< "InnerProductMatrix blas    "<< flops<<" GF/s"<<std::endl;
+    std::cout <<GridLogPerformance<< "InnerProductMatrix blas    "<< bytes<<" GB/s"<<std::endl;
+    std::cout <<GridLogPerformance<< "InnerProductMatrix gsum t5 "<< t5-t4<<" us"<<std::endl;
+    std::cout <<GridLogPerformance<< "InnerProductMatrix cp   t6 "<< t6-t5<<" us"<<std::endl;
+    std::cout <<GridLogPerformance<< "InnerProductMatrix took "<< t6-t0<<" us"<<std::endl;
 #else
    int nrhs;
    GridBase *grid;
@ -358,17 +358,17 @@ public:
    flops = flops/(t4-t3)/1.e3;
    bytes = bytes/(t4-t3)/1.e3;
    xybytes = 4*xybytes/(t2-t1)/1.e3;
-    std::cout << "InnerProductMatrix m,n,k "<< M<<","<<N<<","<<K<<std::endl;
-    std::cout << "InnerProductMatrix alloc t1 "<< t1-t0<<" us"<<std::endl;
-    std::cout << "InnerProductMatrix cp    t2 "<< t2-t1<<" us "<<xybytes<<" GB/s"<<std::endl;
-    std::cout << "InnerProductMatrix setup t3 "<< t3-t2<<" us"<<std::endl;
-    std::cout << "InnerProductMatrix blas t4 "<< t4-t3<<" us"<<std::endl;
-    std::cout << "InnerProductMatrix blas    "<< flops<<" GF/s"<<std::endl;
-    std::cout << "InnerProductMatrix blas    "<< bytes<<" GB/s"<<std::endl;
-    std::cout << "InnerProductMatrix cp     t5 "<< t5-t4<<" us"<<std::endl;
-    std::cout << "InnerProductMatrix lsum   t6l "<< t6l-t5<<" us"<<std::endl;
-    std::cout << "InnerProductMatrix gsum   t6 "<< t6-t6l<<" us"<<std::endl;
-    std::cout << "InnerProductMatrix took "<< t6-t0<<" us"<<std::endl;
+    std::cout <<GridLogPerformance<< "InnerProductMatrix m,n,k "<< M<<","<<N<<","<<K<<std::endl;
+    std::cout <<GridLogPerformance<< "InnerProductMatrix alloc t1 "<< t1-t0<<" us"<<std::endl;
+    std::cout <<GridLogPerformance<< "InnerProductMatrix cp    t2 "<< t2-t1<<" us "<<xybytes<<" GB/s"<<std::endl;
+    std::cout <<GridLogPerformance<< "InnerProductMatrix setup t3 "<< t3-t2<<" us"<<std::endl;
+    std::cout <<GridLogPerformance<< "InnerProductMatrix blas t4 "<< t4-t3<<" us"<<std::endl;
+    std::cout <<GridLogPerformance<< "InnerProductMatrix blas    "<< flops<<" GF/s"<<std::endl;
+    std::cout <<GridLogPerformance<< "InnerProductMatrix blas    "<< bytes<<" GB/s"<<std::endl;
+    std::cout <<GridLogPerformance<< "InnerProductMatrix cp     t5 "<< t5-t4<<" us"<<std::endl;
+    std::cout <<GridLogPerformance<< "InnerProductMatrix lsum   t6l "<< t6l-t5<<" us"<<std::endl;
+    std::cout <<GridLogPerformance<< "InnerProductMatrix gsum   t6 "<< t6-t6l<<" us"<<std::endl;
+    std::cout <<GridLogPerformance<< "InnerProductMatrix took "<< t6-t0<<" us"<<std::endl;
 #endif
  }
 };
--- a/Grid/algorithms/iterative/AdefMrhs.h
+++ b/Grid/algorithms/iterative/AdefMrhs.h
@ -63,7 +63,12 @@ class TwoLevelCGmrhs
  GridStopWatch SmoothTimer;
  GridStopWatch InsertTimer;

-  
+  /*
+    Field rrr;
+  Field sss;
+  Field qqq;
+  Field zzz;
+  */  
  // more most opertor functions
  TwoLevelCGmrhs(RealD tol,
 		 Integer maxit,
@ -74,6 +79,12 @@ class TwoLevelCGmrhs
    MaxIterations(maxit),
    _FineLinop(FineLinop),
    _Smoother(Smoother)
+    /*
+    rrr(fine),
+    sss(fine),
+    qqq(fine),
+    zzz(fine)
+*/
  {
    grid       = fine;
  };
@ -81,8 +92,8 @@ class TwoLevelCGmrhs
  // Vector case
  virtual void operator() (std::vector<Field> &src, std::vector<Field> &x)
  {
-    SolveSingleSystem(src,x);
-    //    SolvePrecBlockCG(src,x);
+    //    SolveSingleSystem(src,x);
+    SolvePrecBlockCG(src,x);
  }

 ////////////////////////////////////////////////////////////////////////////////////////////////////
@ -657,6 +668,8 @@ public:
    CoarseField PleftProjMrhs(this->coarsegridmrhs);
    CoarseField PleftMss_projMrhs(this->coarsegridmrhs);

+    //    this->rrr=in[0];
+
 #undef SMOOTHER_BLOCK_SOLVE
 #if SMOOTHER_BLOCK_SOLVE
    this->SmoothTimer.Start();
@ -669,6 +682,7 @@ public:
      this->SmoothTimer.Stop();
    }
 #endif
+    //    this->sss=Min[0];
    
    for(int rhs=0;rhs<nrhs;rhs++) {
      
@ -705,9 +719,11 @@ public:
    this->_Projector.blockPromote(tmp,PleftMss_proj);// tmp= Q[in - A Min]  
    this->PromoteTimer.Stop();
    this->FineTimer.Start();
+    //    this->qqq=tmp[0];
    for(int rhs=0;rhs<nrhs;rhs++) {
      axpy(out[rhs],1.0,Min[rhs],tmp[rhs]); // Min+tmp
    }
+    //    this->zzz=out[0];
    this->FineTimer.Stop();
  }
 };
--- a/Grid/algorithms/iterative/ImplicitlyRestartedLanczos.h
+++ b/Grid/algorithms/iterative/ImplicitlyRestartedLanczos.h
@ -245,9 +245,10 @@ until convergence
 	_HermOp(src_n,tmp);
 	//	std::cout << GridLogMessage<< tmp<<std::endl; exit(0);
 	//	std::cout << GridLogIRL << " _HermOp " << norm2(tmp) << std::endl;
-	RealD vnum = real(innerProduct(src_n,tmp)); // HermOp.
+//	RealD vnum = real(innerProduct(src_n,tmp)); // HermOp.
+	RealD vnum = real(innerProduct(tmp,tmp)); // HermOp^2.
 	RealD vden = norm2(src_n);
-	RealD na = vnum/vden;
+	RealD na = std::sqrt(vnum/vden);
 	if (fabs(evalMaxApprox/na - 1.0) < 0.0001)
 	  i=_MAX_ITER_IRL_MEVAPP_;
 	evalMaxApprox = na;
@ -255,6 +256,7 @@ until convergence
 	src_n = tmp;
      }
    }
+    std::cout << GridLogIRL << " Final evalMaxApprox  " << evalMaxApprox << std::endl;
 	
    std::vector<RealD> lme(Nm);  
    std::vector<RealD> lme2(Nm);
--- a/Grid/algorithms/iterative/PrecGeneralisedConjugateResidualNonHermitian.h
+++ b/Grid/algorithms/iterative/PrecGeneralisedConjugateResidualNonHermitian.h
@ -74,7 +74,7 @@ public:

  void operator() (const Field &src, Field &psi){

-    psi=Zero();
+    //    psi=Zero();
    RealD cp, ssq,rsq;
    ssq=norm2(src);
    rsq=Tolerance*Tolerance*ssq;
--- a/Grid/algorithms/multigrid/Aggregates.h
+++ b/Grid/algorithms/multigrid/Aggregates.h
@ -30,6 +30,8 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 /*  END LEGAL */
 #pragma once

+#include <Grid/algorithms/iterative/PrecGeneralisedConjugateResidualNonHermitian.h>
+
 NAMESPACE_BEGIN(Grid);

 inline RealD AggregatePowerLaw(RealD x)
@ -124,6 +126,53 @@ public:
    }
  }

+  virtual void CreateSubspaceGCR(GridParallelRNG  &RNG,LinearOperatorBase<FineField> &DiracOp,int nn=nbasis)
+  {
+    RealD scale;
+
+    TrivialPrecon<FineField> simple_fine;
+    PrecGeneralisedConjugateResidualNonHermitian<FineField> GCR(0.001,30,DiracOp,simple_fine,12,12);
+    FineField noise(FineGrid);
+    FineField src(FineGrid);
+    FineField guess(FineGrid);
+    FineField Mn(FineGrid);
+
+    for(int b=0;b<nn;b++){
+      
+      subspace[b] = Zero();
+      gaussian(RNG,noise);
+      scale = std::pow(norm2(noise),-0.5); 
+      noise=noise*scale;
+      
+      DiracOp.Op(noise,Mn); std::cout<<GridLogMessage << "noise   ["<<b<<"] <n|Op|n> "<<innerProduct(noise,Mn)<<std::endl;
+
+      for(int i=0;i<3;i++){
+	//  void operator() (const Field &src, Field &psi){
+#if 1
+	std::cout << GridLogMessage << " inverting on noise "<<std::endl;
+	src = noise;
+	guess=Zero();
+	GCR(src,guess);
+	subspace[b] = guess;
+#else
+	std::cout << GridLogMessage << " inverting on zero "<<std::endl;
+	src=Zero();
+	guess = noise;
+	GCR(src,guess);
+	subspace[b] = guess;
+#endif
+	noise = subspace[b];
+	scale = std::pow(norm2(noise),-0.5); 
+	noise=noise*scale;
+
+      }
+
+      DiracOp.Op(noise,Mn); std::cout<<GridLogMessage << "filtered["<<b<<"] <f|Op|f> "<<innerProduct(noise,Mn)<<std::endl;
+      subspace[b]   = noise;
+
+    }
+  }
+
  ////////////////////////////////////////////////////////////////////////////////////////////////
  // World of possibilities here. But have tried quite a lot of experiments (250+ jobs run on Summit)
  // and this is the best I found
@ -160,14 +209,21 @@ public:

    int b =0;
    {
+      ComplexD ip;
      // Filter
      Chebyshev<FineField> Cheb(lo,hi,orderfilter);
      Cheb(hermop,noise,Mn);
      // normalise
      scale = std::pow(norm2(Mn),-0.5); 	Mn=Mn*scale;
      subspace[b]   = Mn;
-      hermop.Op(Mn,tmp); 
-      std::cout<<GridLogMessage << "filt ["<<b<<"] <n|MdagM|n> "<<norm2(tmp)<<std::endl;
+
+      hermop.Op(Mn,tmp);
+      ip= innerProduct(Mn,tmp); 
+      std::cout<<GridLogMessage << "filt ["<<b<<"] <n|Op|n> "<<norm2(tmp)<<" "<<ip<<std::endl;
+
+      hermop.AdjOp(Mn,tmp); 
+      ip = innerProduct(Mn,tmp); 
+      std::cout<<GridLogMessage << "filt ["<<b<<"] <n|AdjOp|n> "<<norm2(tmp)<<" "<<ip<<std::endl;
      b++;
    }

@ -213,8 +269,18 @@ public:
 	  Mn=*Tnp;
 	  scale = std::pow(norm2(Mn),-0.5);         Mn=Mn*scale;
 	  subspace[b] = Mn;
-	  hermop.Op(Mn,tmp); 
-	  std::cout<<GridLogMessage << n<<" filt ["<<b<<"] <n|MdagM|n> "<<norm2(tmp)<<std::endl;
+
+
+	  ComplexD ip;
+
+	  hermop.Op(Mn,tmp);
+	  ip= innerProduct(Mn,tmp); 
+	  std::cout<<GridLogMessage << "filt ["<<b<<"] <n|Op|n> "<<norm2(tmp)<<" "<<ip<<std::endl;
+
+	  hermop.AdjOp(Mn,tmp); 
+	  ip = innerProduct(Mn,tmp); 
+	  std::cout<<GridLogMessage << "filt ["<<b<<"] <n|AdjOp|n> "<<norm2(tmp)<<" "<<ip<<std::endl;
+	  
 	  b++;
 	}

@ -228,6 +294,70 @@ public:
    }
    assert(b==nn);
  }
+
+
+  virtual void CreateSubspacePolyCheby(GridParallelRNG  &RNG,LinearOperatorBase<FineField> &hermop,
+				       int nn,
+				       double hi,
+				       double lo1,
+				       int orderfilter,
+				       double lo2,
+				       int orderstep)
+  {
+    RealD scale;
+
+    FineField noise(FineGrid);
+    FineField Mn(FineGrid);
+    FineField tmp(FineGrid);
+
+    // New normalised noise
+    gaussian(RNG,noise);
+    scale = std::pow(norm2(noise),-0.5); 
+    noise=noise*scale;
+
+    std::cout << GridLogMessage<<" CreateSubspacePolyCheby "<<std::endl;
+    // Initial matrix element
+    hermop.Op(noise,Mn);
+    std::cout<<GridLogMessage << "noise <n|MdagM|n> "<<norm2(Mn)<<std::endl;
+
+    int b =0;
+    {
+      // Filter
+      std::cout << GridLogMessage << "Cheby "<<lo1<<","<<hi<<" "<<orderstep<<std::endl;
+      Chebyshev<FineField> Cheb(lo1,hi,orderfilter);
+      Cheb(hermop,noise,Mn);
+      // normalise
+      scale = std::pow(norm2(Mn),-0.5); 	Mn=Mn*scale;
+      subspace[b]   = Mn;
+      hermop.Op(Mn,tmp); 
+      std::cout<<GridLogMessage << "filt ["<<b<<"] <n|MdagM|n> "<<norm2(tmp)<<std::endl;
+      std::cout<<GridLogMessage << "filt ["<<b<<"] <n|n> "<<norm2(Mn)<<std::endl;
+    }
+
+    // Generate a full sequence of Chebyshevs
+    for(int n=1;n<nn;n++){
+      std::cout << GridLogMessage << "Cheby "<<lo2<<","<<hi<<" "<<orderstep<<std::endl;
+      Chebyshev<FineField> Cheb(lo2,hi,orderstep);
+      Cheb(hermop,subspace[n-1],Mn);
+
+      for(int m=0;m<n;m++){
+	ComplexD c = innerProduct(subspace[m],Mn);
+	Mn = Mn - c*subspace[m];
+      }
+      
+      // normalise
+      scale = std::pow(norm2(Mn),-0.5);
+      Mn=Mn*scale;
+      
+      subspace[n]=Mn;
+      
+      hermop.Op(Mn,tmp); 
+      std::cout<<GridLogMessage << "filt ["<<n<<"] <n|MdagM|n> "<<norm2(tmp)<<std::endl;
+      std::cout<<GridLogMessage << "filt ["<<n<<"] <n|n> "<<norm2(Mn)<<std::endl;
+
+    }
+  }
+
  virtual void CreateSubspaceChebyshev(GridParallelRNG  &RNG,LinearOperatorBase<FineField> &hermop,
 				       int nn,
 				       double hi,
--- a/Grid/communicator/Communicator_base.h
+++ b/Grid/communicator/Communicator_base.h
@ -127,7 +127,7 @@ public:
  void GlobalSumVector(ComplexD *c,int N);
  void GlobalXOR(uint32_t &);
  void GlobalXOR(uint64_t &);
-  
+
  template<class obj> void GlobalSumP2P(obj &o)
  {
    std::vector<obj> column;
--- a/benchmarks/Benchmark_usqcd.cc
+++ b/benchmarks/Benchmark_usqcd.cc
@ -175,8 +175,8 @@ public:
 	    timestat.statistics(t_time);
 	  
 	    dbytes=dbytes*ppn;
-	    double xbytes    = dbytes*0.5;
-	    double bidibytes = dbytes;
+	    double xbytes    = dbytes;
+	    double bidibytes = dbytes*2.0;
 	  
 	    std::cout<<GridLogMessage << lat<<"\t"<<Ls<<"\t "
 		     << bytes << " \t "
--- a/systems/Aurora/benchmarks/bench16.pbs
+++ b/systems/Aurora/benchmarks/bench16.pbs
@ -0,0 +1,80 @@
+#!/bin/bash
+
+##PBS -q LatticeQCD_aesp_CNDA
+#PBS -q debug-scaling
+##PBS -q prod
+#PBS -l select=16
+#PBS -l walltime=00:20:00
+#PBS -A LatticeQCD_aesp_CNDA
+
+cd $PBS_O_WORKDIR
+
+source ../sourceme.sh
+
+cp $PBS_NODEFILE nodefile
+
+export OMP_NUM_THREADS=4
+export MPICH_OFI_NIC_POLICY=GPU
+
+#export MPIR_CVAR_CH4_OFI_ENABLE_GPU_PIPELINE=1
+#unset MPIR_CVAR_CH4_OFI_GPU_PIPELINE_D2H_ENGINE_TYPE
+#unset MPIR_CVAR_CH4_OFI_GPU_PIPELINE_H2D_ENGINE_TYPE
+#unset MPIR_CVAR_GPU_USE_IMMEDIATE_COMMAND_LIST
+#export MPIR_CVAR_CH4_OFI_GPU_PIPELINE_D2H_ENGINE_TYPE=0
+#export MPIR_CVAR_CH4_OFI_GPU_PIPELINE_H2D_ENGINE_TYPE=0
+#export MPIR_CVAR_GPU_USE_IMMEDIATE_COMMAND_LIST=1
+#export MPIR_CVAR_CH4_OFI_GPU_PIPELINE_BUFFER_SZ=1048576
+#export MPIR_CVAR_CH4_OFI_GPU_PIPELINE_THRESHOLD=131072
+#export MPIR_CVAR_CH4_OFI_GPU_PIPELINE_NUM_BUFFERS_PER_CHUNK=16
+#export MPIR_CVAR_CH4_OFI_GPU_PIPELINE_MAX_NUM_BUFFERS=16
+
+#
+# Local vol 16.16.16.32
+#
+
+VOL 128.64.128.96
+MPI 4.4.4.3
+NPROC 192
+mpiexec -np 192 -ppn 12 -envall ./gpu_tile.sh ./Benchmark_dwf_fp32 --mpi 4.4.4.3 --grid 128.64.128.96 --shm-mpi 0 --shm 2048 --device-mem 32000 --accelerator-threads 32 --comms-overlap
+
+
+LX=32
+LY=16
+LZ=32
+LT=32
+
+NX=2
+NY=2
+NZ=4
+NT=1
+
+GX=2
+GY=2
+GZ=1
+GT=3
+
+PX=$((NX * GX ))
+PY=$((NY * GY ))
+PZ=$((NZ * GZ ))
+PT=$((NT * GT ))
+
+VX=$((PX * LX ))
+VY=$((PY * LY ))
+VZ=$((PZ * LZ ))
+VT=$((PT * LT ))
+
+NP=$((PX*PY*PZ*PT))
+VOL=${VX}.${VY}.${VZ}.${VT}
+AT=8
+MPI=${PX}.${PY}.${PZ}.${PT}
+
+CMD="mpiexec -np $NP -ppn 12  -envall \
+	     ./gpu_tile.sh ./Benchmark_dwf_fp32 --mpi $MPI --grid $VOL \
+		--shm-mpi 0 --shm 2048 --device-mem 32000 --accelerator-threads $AT --comms-overlap "
+
+echo VOL $VOL
+echo MPI $MPI
+echo NPROC $NP
+echo $CMD
+$CMD
+
--- a/systems/Aurora/benchmarks/gpu_tile.sh
+++ b/systems/Aurora/benchmarks/gpu_tile.sh
@ -30,8 +30,8 @@ echo "rank $PALS_RANKID ; local rank $PALS_LOCAL_RANKID ; ZE_AFFINITY_MASK=$ZE_A

 if [ $PALS_RANKID = "0" ]
 then
-    numactl -p $NUMAP -N $NUMAP unitrace --chrome-kernel-logging --chrome-mpi-logging --chrome-sycl-logging --demangle "$@"
-#    numactl -p $NUMAP -N $NUMAP  "$@"
+#    numactl -p $NUMAP -N $NUMAP unitrace --chrome-kernel-logging --chrome-mpi-logging --chrome-sycl-logging --demangle "$@"
+    numactl -p $NUMAP -N $NUMAP  "$@"
 else 
    numactl -p $NUMAP -N $NUMAP  "$@"
 fi
--- a/systems/Frontier/sourceme.sh
+++ b/systems/Frontier/sourceme.sh
@ -2,7 +2,7 @@
 spack load c-lime
 module load emacs 
 module load PrgEnv-gnu
-module load rocm
+module load rocm/6.0.0
 module load cray-mpich
 module load gmp
 module load cray-fftw
--- a/tests/debug/Test_general_coarse_hdcg_phys48_blockcg.cc
+++ b/tests/debug/Test_general_coarse_hdcg_phys48_blockcg.cc
@ -0,0 +1,781 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./tests/Test_general_coarse_hdcg.cc
+
+    Copyright (C) 2023
+
+Author: Peter Boyle <pboyle@bnl.gov>
+
+    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 */
+#include <Grid/Grid.h>
+#include <Grid/algorithms/iterative/ImplicitlyRestartedBlockLanczos.h>
+#include <Grid/algorithms/iterative/ImplicitlyRestartedBlockLanczosCoarse.h>
+#include <Grid/algorithms/iterative/AdefMrhs.h>
+#include <Grid/algorithms/iterative/PowerSpectrum.h>
+#include <Grid/algorithms/iterative/BlockConjugateGradient.h>
+
+using namespace std;
+using namespace Grid;
+
+template<class aggregation>
+void SaveFineEvecs(aggregation &Agg,std::string file)
+{
+#ifdef HAVE_LIME
+  emptyUserRecord record;
+  ScidacWriter WR(Agg[0].Grid()->IsBoss());
+  WR.open(file);
+  for(int b=0;b<Agg.size();b++){
+    WR.writeScidacFieldRecord(Agg[b],record,0,Grid::BinaryIO::BINARYIO_LEXICOGRAPHIC);
+  }
+  WR.close();
+#endif
+}
+template<class aggregation>
+void SaveBasis(aggregation &Agg,std::string file)
+{
+#ifdef HAVE_LIME
+  emptyUserRecord record;
+  ScidacWriter WR(Agg.FineGrid->IsBoss());
+  WR.open(file);
+  for(int b=0;b<Agg.subspace.size();b++){
+    WR.writeScidacFieldRecord(Agg.subspace[b],record,0,Grid::BinaryIO::BINARYIO_LEXICOGRAPHIC);
+    //    WR.writeScidacFieldRecord(Agg.subspace[b],record);
+  }
+  WR.close();
+#endif
+}
+template<class aggregation>
+void LoadBasis(aggregation &Agg, std::string file)
+{
+#ifdef HAVE_LIME
+  emptyUserRecord record;
+  ScidacReader RD ;
+  RD.open(file);
+  for(int b=0;b<Agg.subspace.size();b++){
+    RD.readScidacFieldRecord(Agg.subspace[b],record,Grid::BinaryIO::BINARYIO_LEXICOGRAPHIC);
+    //    RD.readScidacFieldRecord(Agg.subspace[b],record,0);
+  }    
+  RD.close();
+#endif
+}
+
+template<class aggregation>
+void LoadBasisSkip(aggregation &Agg, std::string file,int N,LatticeFermionF & tmp)
+{
+#ifdef HAVE_LIME
+  emptyUserRecord record;
+  ScidacReader RD ;
+  
+  RD.open(file);
+  for(int b=0;b<Agg.subspace.size();b++){
+    for(int n=0;n<N;n++){
+      RD.readScidacFieldRecord(tmp,record,Grid::BinaryIO::BINARYIO_LEXICOGRAPHIC);
+      if(n==0) precisionChange(Agg.subspace[b],tmp);
+    }
+    //    RD.readScidacFieldRecord(Agg.subspace[b],record,0);
+  }    
+  RD.close();
+#endif
+}
+template<class aggregation>
+void LoadBasisSum(aggregation &Agg, std::string file,int N,LatticeFermionF & tmp)
+{
+#ifdef HAVE_LIME
+  emptyUserRecord record;
+  ScidacReader RD ;
+  
+  LatticeFermionF sum(tmp.Grid());
+  RD.open(file);
+  for(int b=0;b<Agg.subspace.size();b++){
+    sum=Zero();
+    for(int n=0;n<N;n++){
+      RD.readScidacFieldRecord(tmp,record,Grid::BinaryIO::BINARYIO_LEXICOGRAPHIC);
+      sum=sum+tmp;
+    }
+    precisionChange(Agg.subspace[b],sum);
+    //    RD.readScidacFieldRecord(Agg.subspace[b],record,0);
+  }    
+  RD.close();
+#endif
+}
+
+template<class CoarseVector>
+void SaveEigenvectors(std::vector<RealD>            &eval,
+		      std::vector<CoarseVector>     &evec,
+		      std::string evec_file,
+		      std::string eval_file)
+{
+#ifdef HAVE_LIME
+  emptyUserRecord record;
+  ScidacWriter WR(evec[0].Grid()->IsBoss());
+  WR.open(evec_file);
+  for(int b=0;b<evec.size();b++){
+    WR.writeScidacFieldRecord(evec[b],record,0,0);
+  }
+  WR.close();
+  XmlWriter WRx(eval_file);
+  write(WRx,"evals",eval);
+#endif
+}
+template<class CoarseVector>
+void LoadEigenvectors(std::vector<RealD>            &eval,
+		      std::vector<CoarseVector>     &evec,
+		      std::string evec_file,
+		      std::string eval_file)
+{
+#ifdef HAVE_LIME
+    XmlReader RDx(eval_file);
+    read(RDx,"evals",eval);
+    emptyUserRecord record;
+
+    Grid::ScidacReader RD ;
+    RD.open(evec_file);
+    assert(evec.size()==eval.size());
+    for(int k=0;k<eval.size();k++) {
+      RD.readScidacFieldRecord(evec[k],record);
+    }
+    RD.close();
+#endif
+}
+
+// Want Op in CoarsenOp to call MatPcDagMatPc
+template<class Field>
+class HermOpAdaptor : public LinearOperatorBase<Field>
+{
+  LinearOperatorBase<Field> & wrapped;
+public:
+  HermOpAdaptor(LinearOperatorBase<Field> &wrapme) : wrapped(wrapme)  {};
+  void Op     (const Field &in, Field &out)   { wrapped.HermOp(in,out);  }
+  void HermOp(const Field &in, Field &out)    { wrapped.HermOp(in,out); }
+  void AdjOp     (const Field &in, Field &out){ wrapped.HermOp(in,out);  }
+  void OpDiag (const Field &in, Field &out)                  {    assert(0);  }
+  void OpDir  (const Field &in, Field &out,int dir,int disp) {    assert(0);  }
+  void OpDirAll  (const Field &in, std::vector<Field> &out)  {    assert(0);  };
+  void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){    assert(0);  }
+};
+
+template<class Field> class FixedCGPolynomial : public LinearFunction<Field>
+{
+public:
+  using LinearFunction<Field>::operator();
+  typedef LinearOperatorBase<Field> FineOperator;
+  FineOperator   & _SmootherOperator;
+  ConjugateGradientPolynomial<Field>  CG;
+  int iters;
+  bool record;
+  int replay_count;
+  FixedCGPolynomial(int _iters, FineOperator &SmootherOperator) :
+    _SmootherOperator(SmootherOperator),
+    iters(_iters),
+    record(true),
+    CG(0.0,_iters,false)
+  {
+    std::cout << GridLogMessage<<" FixedCGPolynomial order "<<iters<<std::endl;
+    replay_count = 0;
+  };
+  void operator() (const Field &in, Field &out) 
+  {
+#if 1
+    GridBase *grid = in.Grid();
+    Field Mx0(grid);
+    Field r0(grid);
+    Field Minvr0(grid);
+
+    _SmootherOperator.HermOp(out,Mx0);
+
+    r0 = in - Mx0;
+
+    Minvr0 = Zero();
+    Minvr0.Checkerboard()=in.Checkerboard();
+    
+    if ( record ) {
+      std::cout << " FixedCGPolynomial recording polynomial "<<std::endl;
+      CG.Solve(_SmootherOperator,r0,Minvr0);
+      record = false;
+      /*
+      std::cout << "P(x) = 0 "<<std::endl;
+      for(int i=0;i<CG.polynomial.size();i++){
+	std::cout<<" + "<< CG.polynomial[i]<<" * (x**"<<i<<")"<<std::endl;
+	}
+      */
+      Field tmp(Minvr0.Grid());
+      CG.CGsequenceHermOp(_SmootherOperator,r0,tmp);
+      tmp = tmp - Minvr0;
+      std::cout << " CGsequence error "<<norm2(tmp)<<" / "<<norm2(out)<<std::endl;
+    } else {
+      std::cout << " FixedCGPolynomial replaying polynomial "<<std::endl;
+      CG.CGsequenceHermOp(_SmootherOperator,r0,Minvr0);
+      if ( replay_count %5== 0 ) record=true;
+      replay_count++;
+    }
+    out = out + Minvr0;
+    _SmootherOperator.HermOp(out,r0);
+    r0 = r0 - in;
+    RealD rr=norm2(r0);
+    RealD ss=norm2(in);
+    std::cout << " FixedCGPolynomial replayed polynomial resid "<<::sqrt(rr/ss)<<std::endl;
+#else
+    out = Zero();
+    out.Checkerboard()=in.Checkerboard();
+    if ( record ) {
+      std::cout << " FixedCGPolynomial recording polynomial "<<std::endl;
+      CG.Solve(_SmootherOperator,in,out);
+      record = false;
+      std::cout << "P(x) = 0 "<<std::endl;
+      for(int i=0;i<CG.polynomial.size();i++){
+	std::cout<<" + "<< CG.polynomial[i]<<" * (x**"<<i<<")"<<std::endl;
+      }
+      Field tmp(in.Grid());
+      CG.CGsequenceHermOp(_SmootherOperator,in,tmp);
+      tmp = tmp - out;
+      std::cout << " CGsequence error "<<norm2(tmp)<<" / "<<norm2(out)<<std::endl;
+    } else {
+      std::cout << " FixedCGPolynomial replaying polynomial "<<std::endl;
+      CG.CGsequenceHermOp(_SmootherOperator,in,out);
+      if ( replay_count %5== 5 ) record=true;
+      replay_count++;
+    }
+#endif
+    
+  }
+  void operator() (const std::vector<Field> &in, std::vector<Field> &out)
+  {
+    for(int i=0;i<out.size();i++){
+      out[i]=Zero();
+    }
+    int blockDim = 0;//not used for BlockCGVec
+    BlockConjugateGradient<Field>    BCGV  (BlockCGrQVec,blockDim,0.0,iters,false);
+    BCGV(_SmootherOperator,in,out);
+  }
+  
+};
+template<class Field> class CGSmoother : public LinearFunction<Field>
+{
+public:
+  using LinearFunction<Field>::operator();
+  typedef LinearOperatorBase<Field> FineOperator;
+  FineOperator   & _SmootherOperator;
+  int iters;
+  CGSmoother(int _iters, FineOperator &SmootherOperator) :
+    _SmootherOperator(SmootherOperator),
+    iters(_iters)
+  {
+    std::cout << GridLogMessage<<" Mirs smoother order "<<iters<<std::endl;
+  };
+  void operator() (const Field &in, Field &out) 
+  {
+    ConjugateGradient<Field>  CG(0.0,iters,false); // non-converge is just fine in a smoother
+
+    out=Zero();
+
+    CG(_SmootherOperator,in,out);
+  }
+};
+
+
+RealD InverseApproximation(RealD x){
+  return 1.0/x;
+}
+template<class Field> class ChebyshevSmoother : public LinearFunction<Field>
+{
+public:
+  using LinearFunction<Field>::operator();
+  typedef LinearOperatorBase<Field> FineOperator;
+  FineOperator   & _SmootherOperator;
+  Chebyshev<Field> Cheby;
+  ChebyshevSmoother(RealD _lo,RealD _hi,int _ord, FineOperator &SmootherOperator) :
+    _SmootherOperator(SmootherOperator),
+    Cheby(_lo,_hi,_ord,InverseApproximation)
+  {
+    std::cout << GridLogMessage<<" Chebyshev smoother order "<<_ord<<" ["<<_lo<<","<<_hi<<"]"<<std::endl;
+  };
+  void operator() (const Field &in, Field &out) 
+  {
+    //    Field r(out.Grid());
+    Cheby(_SmootherOperator,in,out);
+    //    _SmootherOperator.HermOp(out,r);
+    //    r=r-in;
+    //    RealD rr=norm2(r);
+    //    RealD ss=norm2(in);
+    //    std::cout << GridLogMessage<<" Chebyshev smoother resid "<<::sqrt(rr/ss)<<std::endl;
+  }
+};
+
+
+template<class Field> class ChebyshevInverter : public LinearFunction<Field>
+{
+public:
+  using LinearFunction<Field>::operator();
+  typedef LinearOperatorBase<Field> FineOperator;
+  FineOperator   & _Operator;
+  Chebyshev<Field> Cheby;
+  ChebyshevInverter(RealD _lo,RealD _hi,int _ord, FineOperator &Operator) :
+    _Operator(Operator),
+    Cheby(_lo,_hi,_ord,InverseApproximation)
+  {
+    std::cout << GridLogMessage<<" Chebyshev Inverter order "<<_ord<<" ["<<_lo<<","<<_hi<<"]"<<std::endl;
+  };
+  void operator() (const Field &in, Field &out) 
+  {
+    Field r(in.Grid());
+    Field AinvR(in.Grid());
+    _Operator.HermOp(out,r);
+    r = in - r; // b - A x
+    Cheby(_Operator,r,AinvR); // A^{-1} ( b - A x ) ~ A^{-1} b - x
+    out = out + AinvR;
+    _Operator.HermOp(out,r);
+    r = in - r; // b - A x
+    RealD rr = norm2(r);
+    RealD ss = norm2(in);
+    std::cout << "ChebshevInverse resid " <<::sqrt(rr/ss)<<std::endl;
+  }
+};
+
+
+
+int main (int argc, char ** argv)
+{
+  Grid_init(&argc,&argv);
+
+  int sample=1;
+  if( GridCmdOptionExists(argv,argv+argc,"--sample") ){
+    std::string arg;
+    arg = GridCmdOptionPayload(argv,argv+argc,"--sample");
+    GridCmdOptionInt(arg,sample);
+  }
+  
+  const int Ls=24;
+  const int nbasis = 62;
+  const int cb = 0 ;
+  RealD mass=0.00078;
+
+  if( GridCmdOptionExists(argv,argv+argc,"--mass") ){
+    std::string arg;
+    arg = GridCmdOptionPayload(argv,argv+argc,"--mass");
+    GridCmdOptionFloat(arg,mass);
+  }
+
+  RealD M5=1.8;
+  RealD b=1.5;
+  RealD c=0.5;
+
+  std::cout << GridLogMessage << " *************************** " <<std::endl;
+  std::cout << GridLogMessage << " Mass " <<mass<<std::endl;
+  std::cout << GridLogMessage << " M5   " <<M5<<std::endl;
+  std::cout << GridLogMessage << " Ls   " <<Ls<<std::endl;
+  std::cout << GridLogMessage << " b    " <<b<<std::endl;
+  std::cout << GridLogMessage << " c    " <<c<<std::endl;
+  std::cout << GridLogMessage << " *************************** " <<std::endl;
+  
+  GridCartesian         * UGrid   = SpaceTimeGrid::makeFourDimGrid(GridDefaultLatt(),
+								   GridDefaultSimd(Nd,vComplex::Nsimd()),
+								   GridDefaultMpi());
+  GridRedBlackCartesian * UrbGrid = SpaceTimeGrid::makeFourDimRedBlackGrid(UGrid);
+  GridCartesian         * FGrid   = SpaceTimeGrid::makeFiveDimGrid(Ls,UGrid);
+  GridRedBlackCartesian * FrbGrid = SpaceTimeGrid::makeFiveDimRedBlackGrid(Ls,UGrid);
+
+  //////////////////////////////////////////
+  // Single precision grids -- lanczos + smoother
+  //////////////////////////////////////////
+  GridCartesian         * UGridF   = SpaceTimeGrid::makeFourDimGrid(GridDefaultLatt(),
+								   GridDefaultSimd(Nd,vComplexF::Nsimd()),
+								   GridDefaultMpi());
+  GridRedBlackCartesian * UrbGridF = SpaceTimeGrid::makeFourDimRedBlackGrid(UGridF);
+  GridCartesian         * FGridF   = SpaceTimeGrid::makeFiveDimGrid(Ls,UGridF);
+  GridRedBlackCartesian * FrbGridF = SpaceTimeGrid::makeFiveDimRedBlackGrid(Ls,UGridF);
+
+  ///////////////////////// Configuration /////////////////////////////////
+  LatticeGaugeField Umu(UGrid);
+
+  FieldMetaData header;
+  std::string file("ckpoint_lat.1000");
+  NerscIO::readConfiguration(Umu,header,file);
+
+  //////////////////////// Fermion action //////////////////////////////////
+  MobiusFermionD Ddwf(Umu,*FGrid,*FrbGrid,*UGrid,*UrbGrid,mass,M5,b,c);
+  SchurDiagMooeeOperator<MobiusFermionD, LatticeFermion> HermOpEO(Ddwf);
+
+  std::cout << "**************************************"<<std::endl;
+  std::cout << "         Fine Power method            "<<std::endl;
+  std::cout << "**************************************"<<std::endl;
+
+  {
+    LatticeFermionD pm_src(FrbGrid);
+    pm_src = ComplexD(1.0);
+    PowerMethod<LatticeFermionD>       fPM;
+    fPM(HermOpEO,pm_src);
+  }
+
+  if(0)
+  {
+
+    std::cout << "**************************************"<<std::endl;
+    std::cout << "         Fine Lanczos           "<<std::endl;
+    std::cout << "**************************************"<<std::endl;
+
+    typedef LatticeFermionF FermionField;
+    LatticeGaugeFieldF UmuF(UGridF);
+    precisionChange(UmuF,Umu);
+    MobiusFermionF DdwfF(UmuF,*FGridF,*FrbGridF,*UGridF,*UrbGridF,mass,M5,b,c);
+    SchurDiagMooeeOperator<MobiusFermionF, LatticeFermionF> HermOpEOF(DdwfF);
+
+    const int Fine_Nstop = 200;
+    const int Fine_Nk = 200;
+    const int Fine_Np = 200;
+    const int Fine_Nm = Fine_Nk+Fine_Np;
+    const int Fine_MaxIt= 10;
+
+    RealD Fine_resid = 1.0e-4;
+    std::cout << GridLogMessage << "Fine Lanczos "<<std::endl;
+    std::cout << GridLogMessage << "Nstop "<<Fine_Nstop<<std::endl;
+    std::cout << GridLogMessage << "Nk "<<Fine_Nk<<std::endl;
+    std::cout << GridLogMessage << "Np "<<Fine_Np<<std::endl;
+    std::cout << GridLogMessage << "resid "<<Fine_resid<<std::endl;
+
+    Chebyshev<FermionField> Cheby(0.002,92.0,401);
+    //    Chebyshev<FermionField> Cheby(0.1,92.0,401);
+    FunctionHermOp<FermionField> OpCheby(Cheby,HermOpEOF);
+    PlainHermOp<FermionField> Op     (HermOpEOF);
+    ImplicitlyRestartedLanczos<FermionField> IRL(OpCheby,Op,Fine_Nstop,Fine_Nk,Fine_Nm,Fine_resid,Fine_MaxIt);
+    std::vector<RealD>          Fine_eval(Fine_Nm);
+    FermionField                Fine_src(FrbGridF); 
+    Fine_src = ComplexF(1.0);
+    std::vector<FermionField> Fine_evec(Fine_Nm,FrbGridF);
+
+    int Fine_Nconv;
+    std::cout << GridLogMessage <<" Calling IRL.calc single prec"<<std::endl;
+    IRL.calc(Fine_eval,Fine_evec,Fine_src,Fine_Nconv);
+
+    std::string evec_file("/lustre/orion/phy157/proj-shared/phy157_dwf/paboyle/Subspace.phys48.evecF");
+    SaveFineEvecs(Fine_evec,evec_file);
+  }
+
+
+  //////////////////////////////////////////
+  // Construct a coarsened grid with 4^4 cell
+  //////////////////////////////////////////
+  Coordinate Block({4,4,6,4});
+  Coordinate clatt = GridDefaultLatt();
+  for(int d=0;d<clatt.size();d++){
+    clatt[d] = clatt[d]/Block[d];
+  }
+
+  GridCartesian *Coarse4d =  SpaceTimeGrid::makeFourDimGrid(clatt,
+							    GridDefaultSimd(Nd,vComplex::Nsimd()),
+							    GridDefaultMpi());;
+  GridCartesian *Coarse5d =  SpaceTimeGrid::makeFiveDimGrid(1,Coarse4d);
+
+  ///////////////////////// RNGs /////////////////////////////////
+  std::vector<int> seeds4({1,2,3,4});
+  std::vector<int> seeds5({5,6,7,8});
+  std::vector<int> cseeds({5,6,7,8});
+
+  GridParallelRNG          RNG5(FGrid);   RNG5.SeedFixedIntegers(seeds5);
+  GridParallelRNG          RNG4(UGrid);   RNG4.SeedFixedIntegers(seeds4);
+  GridParallelRNG          CRNG(Coarse5d);CRNG.SeedFixedIntegers(cseeds);
+
+  
+  typedef HermOpAdaptor<LatticeFermionD> HermFineMatrix;
+  HermFineMatrix FineHermOp(HermOpEO);
+
+  ////////////////////////////////////////////////////////////
+  ///////////// Coarse basis and Little Dirac Operator ///////
+  ////////////////////////////////////////////////////////////
+  typedef GeneralCoarsenedMatrix<vSpinColourVector,vTComplex,nbasis> LittleDiracOperator;
+  typedef LittleDiracOperator::CoarseVector CoarseVector;
+
+  NextToNextToNextToNearestStencilGeometry5D geom(Coarse5d);
+
+  typedef Aggregation<vSpinColourVector,vTComplex,nbasis> Subspace;
+  Subspace Aggregates(Coarse5d,FrbGrid,cb);
+
+  ////////////////////////////////////////////////////////////
+  // Need to check about red-black grid coarsening
+  ////////////////////////////////////////////////////////////
+  std::string subspace_file("/lustre/orion/phy157/proj-shared/phy157_dwf/paboyle/Subspace.phys48.mixed.2500.60");
+  //  //  std::string subspace_file("/lustre/orion/phy157/proj-shared/phy157_dwf/paboyle/Subspace.phys48.new.62");
+  //  std::string refine_file("/lustre/orion/phy157/proj-shared/phy157_dwf/paboyle/Subspace.phys48.evecF");
+  std::string refine_file("/lustre/orion/phy157/proj-shared/phy157_dwf/paboyle/Refine.phys48.mixed.2500.60");
+  std::string ldop_file("/lustre/orion/phy157/proj-shared/phy157_dwf/paboyle/LittleDiracOp.phys48.mixed.60");
+  std::string evec_file("/lustre/orion/phy157/proj-shared/phy157_dwf/paboyle/evecs.scidac");
+  std::string eval_file("/lustre/orion/phy157/proj-shared/phy157_dwf/paboyle/eval.xml");
+  bool load_agg=true;
+  bool load_refine=true;
+  bool load_mat=false;
+  bool load_evec=false;
+
+  int refine=1;
+  if ( load_agg ) {
+    if ( !(refine) || (!load_refine) ) { 
+      LoadBasis(Aggregates,subspace_file);
+    }
+  } else {
+    //    Aggregates.CreateSubspaceMultishift(RNG5,HermOpEO,
+    //    					0.0003,1.0e-5,2000); // Lo, tol, maxit
+    //    Aggregates.CreateSubspaceChebyshev(RNG5,HermOpEO,nbasis,95.,0.01,1500);// <== last run
+    Aggregates.CreateSubspaceChebyshevNew(RNG5,HermOpEO,95.); 
+    SaveBasis(Aggregates,subspace_file);
+  }
+
+  std::cout << "**************************************"<<std::endl;
+  std::cout << "Building MultiRHS Coarse operator"<<std::endl;
+  std::cout << "**************************************"<<std::endl;
+  ConjugateGradient<CoarseVector>  coarseCG(4.0e-2,20000,true);
+    
+  const int nrhs=24;
+    
+  Coordinate mpi=GridDefaultMpi();
+  Coordinate rhMpi ({1,1,mpi[0],mpi[1],mpi[2],mpi[3]});
+  Coordinate rhLatt({nrhs,1,clatt[0],clatt[1],clatt[2],clatt[3]});
+  Coordinate rhSimd({vComplex::Nsimd(),1, 1,1,1,1});
+    
+  GridCartesian *CoarseMrhs = new GridCartesian(rhLatt,rhSimd,rhMpi); 
+  typedef MultiGeneralCoarsenedMatrix<vSpinColourVector,vTComplex,nbasis> MultiGeneralCoarsenedMatrix_t;
+  MultiGeneralCoarsenedMatrix_t mrhs(geom,CoarseMrhs);
+
+  std::cout << "**************************************"<<std::endl;
+  std::cout << "         Coarse Lanczos               "<<std::endl;
+  std::cout << "**************************************"<<std::endl;
+
+  typedef HermitianLinearOperator<MultiGeneralCoarsenedMatrix_t,CoarseVector> MrhsHermMatrix;
+  Chebyshev<CoarseVector>      IRLCheby(0.005,42.0,301);  // 1 iter
+  MrhsHermMatrix MrhsCoarseOp     (mrhs);
+
+  //  CoarseVector pm_src(CoarseMrhs);
+  //  pm_src = ComplexD(1.0);
+  //  PowerMethod<CoarseVector>       cPM;   cPM(MrhsCoarseOp,pm_src);
+
+  int Nk=192;
+  int Nm=384;
+  int Nstop=Nk;
+  int Nconv_test_interval=1;
+  
+  ImplicitlyRestartedBlockLanczosCoarse<CoarseVector> IRL(MrhsCoarseOp,
+							  Coarse5d,
+							  CoarseMrhs,
+							  nrhs,
+							  IRLCheby,
+							  Nstop,
+							  Nconv_test_interval,
+							  nrhs,
+							  Nk,
+							  Nm,
+							  1e-5,10);
+
+  int Nconv;
+  std::vector<RealD>            eval(Nm);
+  std::vector<CoarseVector>     evec(Nm,Coarse5d);
+  std::vector<CoarseVector>     c_src(nrhs,Coarse5d);
+
+  ///////////////////////
+  // Deflation guesser object
+  ///////////////////////
+  MultiRHSDeflation<CoarseVector> MrhsGuesser;
+
+  //////////////////////////////////////////
+  // Block projector for coarse/fine
+  //////////////////////////////////////////
+  MultiRHSBlockProject<LatticeFermionD> MrhsProjector;
+
+  //////////////////////////
+  // Extra HDCG parameters
+  //////////////////////////
+  int maxit=300;
+  ConjugateGradient<CoarseVector>  CG(5.0e-2,maxit,false);
+  ConjugateGradient<CoarseVector>  CGstart(5.0e-2,maxit,false);
+  RealD lo=2.0;
+  int ord = 7;
+  //  int ord = 11;
+
+  int blockDim = 0;//not used for BlockCG
+  BlockConjugateGradient<CoarseVector>    BCG  (BlockCGrQ,blockDim,5.0e-5,maxit,true);
+
+  DoNothingGuesser<CoarseVector> DoNothing;
+  //  HPDSolver<CoarseVector> HPDSolveMrhs(MrhsCoarseOp,CG,DoNothing);
+  //  HPDSolver<CoarseVector> HPDSolveMrhsStart(MrhsCoarseOp,CGstart,DoNothing);
+  //  HPDSolver<CoarseVector> HPDSolveMrhs(MrhsCoarseOp,BCG,DoNothing);
+  //  HPDSolver<CoarseVector> HPDSolveMrhsRefine(MrhsCoarseOp,BCG,DoNothing);
+  //  FixedCGPolynomial<CoarseVector>  HPDSolveMrhs(maxit,MrhsCoarseOp);
+
+  ChebyshevInverter<CoarseVector> HPDSolveMrhs(1.0e-2,40.0,120,MrhsCoarseOp);  //
+  //  ChebyshevInverter<CoarseVector> HPDSolveMrhs(1.0e-2,40.0,110,MrhsCoarseOp);  // 114 iter with Chebysmooth and BlockCG
+  //  ChebyshevInverter<CoarseVector> HPDSolveMrhs(1.0e-2,40.0,120,MrhsCoarseOp); // 138 iter with Chebysmooth
+  //  ChebyshevInverter<CoarseVector> HPDSolveMrhs(1.0e-2,40.0,200,MrhsCoarseOp); // 139 iter
+  //  ChebyshevInverter<CoarseVector> HPDSolveMrhs(3.0e-3,40.0,200,MrhsCoarseOp); // 137 iter, CG smooth, flex
+  //  ChebyshevInverter<CoarseVector> HPDSolveMrhs(1.0e-3,40.0,200,MrhsCoarseOp); // 146 iter, CG smooth, flex
+  //  ChebyshevInverter<CoarseVector> HPDSolveMrhs(3.0e-4,40.0,200,MrhsCoarseOp); // 156 iter, CG smooth, flex
+
+  /////////////////////////////////////////////////
+  // Mirs smoother
+  /////////////////////////////////////////////////
+  ShiftedHermOpLinearOperator<LatticeFermionD> ShiftedFineHermOp(HermOpEO,lo);
+  //  FixedCGPolynomial<LatticeFermionD> CGsmooth(ord,ShiftedFineHermOp) ;
+  //  CGSmoother<LatticeFermionD> CGsmooth(ord,ShiftedFineHermOp) ;
+  ChebyshevSmoother<LatticeFermionD> CGsmooth(2.0,92.0,8,HermOpEO) ;
+  
+  if ( load_refine ) {
+    LoadBasis(Aggregates,refine_file);
+    //    LatticeFermionF conv_tmp(FrbGridF);
+    //    LoadBasisSum(Aggregates,refine_file,sample,conv_tmp);
+  } else {
+    Aggregates.RefineSubspace(HermOpEO,0.001,1.0e-3,3000); // 172 iters
+    SaveBasis(Aggregates,refine_file);
+  }
+  Aggregates.Orthogonalise();
+
+  std::cout << "**************************************"<<std::endl;
+  std::cout << "Coarsen after refine"<<std::endl;
+  std::cout << "**************************************"<<std::endl;
+  mrhs.CoarsenOperator(FineHermOp,Aggregates,Coarse5d);
+
+  std::cout << "**************************************"<<std::endl;
+  std::cout << " Recompute coarse evecs  "<<std::endl;
+  std::cout << "**************************************"<<std::endl;
+  evec.resize(Nm,Coarse5d);
+  eval.resize(Nm);
+  for(int r=0;r<nrhs;r++){
+    random(CRNG,c_src[r]);
+  }
+ IRL.calc(eval,evec,c_src,Nconv,LanczosType::irbl);
+
+  std::cout << "**************************************"<<std::endl;
+  std::cout << " Reimport coarse evecs  "<<std::endl;
+  std::cout << "**************************************"<<std::endl;
+  MrhsGuesser.ImportEigenBasis(evec,eval);
+
+  std::cout << "**************************************"<<std::endl;
+  std::cout << " Setting up mRHS HDCG"<<std::endl;
+  std::cout << "**************************************"<<std::endl;
+  MrhsProjector.Allocate(nbasis,FrbGrid,Coarse5d);
+  MrhsProjector.ImportBasis(Aggregates.subspace);
+      
+  std::cout << "**************************************"<<std::endl;
+  std::cout << "Calling mRHS HDCG"<<std::endl;
+  std::cout << "**************************************"<<std::endl;
+  TwoLevelADEF2mrhs<LatticeFermion,CoarseVector>
+    HDCGmrhs(1.0e-8, 300,
+	     FineHermOp,
+	     CGsmooth,
+	     HPDSolveMrhs,    // Used in M1
+	     HPDSolveMrhs,          // Used in Vstart
+	     MrhsProjector,
+	     MrhsGuesser,
+	     CoarseMrhs);
+    
+  std::vector<LatticeFermionD> src_mrhs(nrhs,FrbGrid);
+  std::vector<LatticeFermionD> res_mrhs(nrhs,FrbGrid);
+  LatticeFermionD result_accurate(FrbGrid);
+  LatticeFermionD result_sloppy(FrbGrid);
+  LatticeFermionD error(FrbGrid);
+  LatticeFermionD residual(FrbGrid);
+
+  for(int r=0;r<nrhs;r++){
+    random(RNG5,src_mrhs[r]);
+    res_mrhs[r]=Zero();
+  }
+  HDCGmrhs(src_mrhs,res_mrhs);
+  result_accurate = res_mrhs[0];
+
+#if 0
+
+  std::vector<RealD>   bins({1.0e-3,1.0e-2,1.0e-1,1.0,10.0,100.0});
+  std::vector<int>   orders({6000  ,4000  ,1000  ,500,500 ,500});
+  PowerSpectrum GraphicEqualizer(bins,orders);
+  
+  std::cout << "**************************************"<<std::endl;
+  std::cout << GridLogMessage << " PowerSpectrum of rrr "<<std::endl;
+  std::cout << "**************************************"<<std::endl;
+  GraphicEqualizer(FineHermOp,HDCGmrhs.rrr);
+  std::cout << "**************************************"<<std::endl;
+  std::cout << GridLogMessage << " PowerSpectrum of sss "<<std::endl;
+  std::cout << "**************************************"<<std::endl;
+  GraphicEqualizer(FineHermOp,HDCGmrhs.sss);
+  std::cout << "**************************************"<<std::endl;
+  std::cout << GridLogMessage << " PowerSpectrum of qqq "<<std::endl;
+  std::cout << "**************************************"<<std::endl;
+  GraphicEqualizer(FineHermOp,HDCGmrhs.qqq);
+  std::cout << "**************************************"<<std::endl;
+  std::cout << GridLogMessage << " PowerSpectrum of zzz "<<std::endl;
+  std::cout << "**************************************"<<std::endl;
+  GraphicEqualizer(FineHermOp,HDCGmrhs.zzz);
+
+  std::vector<RealD> tols({1.0e-3,1.0e-4,1.0e-5});
+
+
+  for(auto tol : tols) {
+    
+    TwoLevelADEF2mrhs<LatticeFermion,CoarseVector>
+      HDCGmrhsSloppy(tol, 500,
+		     FineHermOp,
+		     CGsmooth,
+		     HPDSolveMrhs,    // Used in M1
+		     HPDSolveMrhs,    // Used in Vstart
+		     MrhsProjector,
+		     MrhsGuesser,
+		     CoarseMrhs);
+  
+    //  Solve again to 10^-5
+    for(int r=0;r<nrhs;r++){
+      res_mrhs[r]=Zero();
+    }
+    HDCGmrhsSloppy(src_mrhs,res_mrhs);
+    
+    result_sloppy = res_mrhs[0];
+    error = result_sloppy - result_accurate;
+    FineHermOp.HermOp(result_sloppy,residual);
+    residual = residual - src_mrhs[0];
+    
+    std::cout << "**************************************"<<std::endl;
+    std::cout << GridLogMessage << " Converged to tolerance "<< tol<<std::endl;
+    std::cout << GridLogMessage << " Absolute error "<<norm2(error)<<std::endl;
+    std::cout << GridLogMessage << " Residual       "<<norm2(residual)<<std::endl;
+    std::cout << "**************************************"<<std::endl;
+
+    std::cout << "**************************************"<<std::endl;
+    std::cout << GridLogMessage << " PowerSpectrum of error   "<<std::endl;
+    std::cout << "**************************************"<<std::endl;
+    GraphicEqualizer(FineHermOp,error);
+    std::cout << "**************************************"<<std::endl;
+    std::cout << GridLogMessage << " PowerSpectrum of residual   "<<std::endl;
+    std::cout << "**************************************"<<std::endl;
+    GraphicEqualizer(FineHermOp,residual);
+
+  };
+#endif
+  
+  // Standard CG
+#if 0
+  {
+  std::cout << "**************************************"<<std::endl;
+  std::cout << "Calling red black CG"<<std::endl;
+  std::cout << "**************************************"<<std::endl;
+      
+    LatticeFermion result(FrbGrid); result=Zero();
+    LatticeFermion    src(FrbGrid); random(RNG5,src);
+    result=Zero();
+
+    ConjugateGradient<LatticeFermionD>  CGfine(1.0e-8,30000,false);
+    CGfine(HermOpEO, src, result);
+  }
+#endif  
+  Grid_finalize();
+  return 0;
+}
--- a/tests/debug/Test_general_coarse_hdcg_phys48_lanczos_subspace.cc
+++ b/tests/debug/Test_general_coarse_hdcg_phys48_lanczos_subspace.cc
@ -0,0 +1,355 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./tests/Test_general_coarse_hdcg.cc
+
+    Copyright (C) 2023
+
+Author: Peter Boyle <pboyle@bnl.gov>
+
+    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 */
+#include <Grid/Grid.h>
+#include <Grid/algorithms/iterative/ImplicitlyRestartedBlockLanczos.h>
+#include <Grid/algorithms/iterative/ImplicitlyRestartedBlockLanczosCoarse.h>
+#include <Grid/algorithms/iterative/AdefMrhs.h>
+
+using namespace std;
+using namespace Grid;
+
+template<class aggregation>
+void SaveFineEvecs(aggregation &Agg,std::string file)
+{
+#ifdef HAVE_LIME
+  emptyUserRecord record;
+  ScidacWriter WR(Agg[0].Grid()->IsBoss());
+  WR.open(file);
+  for(int b=0;b<Agg.size();b++){
+    WR.writeScidacFieldRecord(Agg[b],record,0,Grid::BinaryIO::BINARYIO_LEXICOGRAPHIC);
+  }
+  WR.close();
+#endif
+}
+template<class aggregation>
+void SaveBasis(aggregation &Agg,std::string file)
+{
+#ifdef HAVE_LIME
+  emptyUserRecord record;
+  ScidacWriter WR(Agg.FineGrid->IsBoss());
+  WR.open(file);
+  for(int b=0;b<Agg.subspace.size();b++){
+    WR.writeScidacFieldRecord(Agg.subspace[b],record,0,Grid::BinaryIO::BINARYIO_LEXICOGRAPHIC);
+    //    WR.writeScidacFieldRecord(Agg.subspace[b],record);
+  }
+  WR.close();
+#endif
+}
+template<class aggregation>
+void LoadBasis(aggregation &Agg, std::string file)
+{
+#ifdef HAVE_LIME
+  emptyUserRecord record;
+  ScidacReader RD ;
+  RD.open(file);
+  for(int b=0;b<Agg.subspace.size();b++){
+    RD.readScidacFieldRecord(Agg.subspace[b],record,Grid::BinaryIO::BINARYIO_LEXICOGRAPHIC);
+    //    RD.readScidacFieldRecord(Agg.subspace[b],record,0);
+  }    
+  RD.close();
+#endif
+}
+template<class aggregation>
+void LoadFineEvecs(aggregation &Agg, std::string file,LatticeFermionF & conv_tmp)
+{
+#ifdef HAVE_LIME
+  emptyUserRecord record;
+  ScidacReader RD ;
+  RD.open(file);
+  for(int b=0;b<Agg.size();b++){
+    RD.readScidacFieldRecord(conv_tmp,record,Grid::BinaryIO::BINARYIO_LEXICOGRAPHIC);
+    precisionChange(Agg[b],conv_tmp);
+  }    
+  RD.close();
+#endif
+}
+template<class CoarseVector>
+void SaveEigenvectors(std::vector<RealD>            &eval,
+		      std::vector<CoarseVector>     &evec,
+		      std::string evec_file,
+		      std::string eval_file)
+{
+#ifdef HAVE_LIME
+  emptyUserRecord record;
+  ScidacWriter WR(evec[0].Grid()->IsBoss());
+  WR.open(evec_file);
+  for(int b=0;b<evec.size();b++){
+    WR.writeScidacFieldRecord(evec[b],record,0,0);
+  }
+  WR.close();
+  XmlWriter WRx(eval_file);
+  write(WRx,"evals",eval);
+#endif
+}
+template<class CoarseVector>
+void LoadEigenvectors(std::vector<RealD>            &eval,
+		      std::vector<CoarseVector>     &evec,
+		      std::string evec_file,
+		      std::string eval_file)
+{
+#ifdef HAVE_LIME
+    XmlReader RDx(eval_file);
+    read(RDx,"evals",eval);
+    emptyUserRecord record;
+
+    Grid::ScidacReader RD ;
+    RD.open(evec_file);
+    assert(evec.size()==eval.size());
+    for(int k=0;k<eval.size();k++) {
+      RD.readScidacFieldRecord(evec[k],record);
+    }
+    RD.close();
+#endif
+}
+
+// Want Op in CoarsenOp to call MatPcDagMatPc
+template<class Field>
+class HermOpAdaptor : public LinearOperatorBase<Field>
+{
+  LinearOperatorBase<Field> & wrapped;
+public:
+  HermOpAdaptor(LinearOperatorBase<Field> &wrapme) : wrapped(wrapme)  {};
+  void Op     (const Field &in, Field &out)   { wrapped.HermOp(in,out);  }
+  void HermOp(const Field &in, Field &out)    { wrapped.HermOp(in,out); }
+  void AdjOp     (const Field &in, Field &out){ wrapped.HermOp(in,out);  }
+  void OpDiag (const Field &in, Field &out)                  {    assert(0);  }
+  void OpDir  (const Field &in, Field &out,int dir,int disp) {    assert(0);  }
+  void OpDirAll  (const Field &in, std::vector<Field> &out)  {    assert(0);  };
+  void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){    assert(0);  }
+};
+
+template<class Field> class CGSmoother : public LinearFunction<Field>
+{
+public:
+  using LinearFunction<Field>::operator();
+  typedef LinearOperatorBase<Field> FineOperator;
+  FineOperator   & _SmootherOperator;
+  int iters;
+  CGSmoother(int _iters, FineOperator &SmootherOperator) :
+    _SmootherOperator(SmootherOperator),
+    iters(_iters)
+  {
+    std::cout << GridLogMessage<<" Mirs smoother order "<<iters<<std::endl;
+  };
+  void operator() (const Field &in, Field &out) 
+  {
+    ConjugateGradient<Field>  CG(0.0,iters,false); // non-converge is just fine in a smoother
+
+    out=Zero();
+
+    CG(_SmootherOperator,in,out);
+  }
+};
+
+
+int main (int argc, char ** argv)
+{
+  Grid_init(&argc,&argv);
+
+  const int Ls=24;
+  const int nbasis = 62;
+  const int cb = 0 ;
+  RealD mass=0.00078;
+  RealD M5=1.8;
+  RealD b=1.5;
+  RealD c=0.5;
+
+  GridCartesian         * UGrid   = SpaceTimeGrid::makeFourDimGrid(GridDefaultLatt(),
+								   GridDefaultSimd(Nd,vComplex::Nsimd()),
+								   GridDefaultMpi());
+  GridRedBlackCartesian * UrbGrid = SpaceTimeGrid::makeFourDimRedBlackGrid(UGrid);
+  GridCartesian         * FGrid   = SpaceTimeGrid::makeFiveDimGrid(Ls,UGrid);
+  GridRedBlackCartesian * FrbGrid = SpaceTimeGrid::makeFiveDimRedBlackGrid(Ls,UGrid);
+
+  // Construct a coarsened grid with 4^4 cell
+  Coordinate Block({4,4,6,4});
+  Coordinate clatt = GridDefaultLatt();
+  for(int d=0;d<clatt.size();d++){
+    clatt[d] = clatt[d]/Block[d];
+  }
+
+  //////////////////////////////////////////
+  // Double precision grids 
+  //////////////////////////////////////////
+  GridCartesian *Coarse4d =  SpaceTimeGrid::makeFourDimGrid(clatt,
+							    GridDefaultSimd(Nd,vComplex::Nsimd()),
+							    GridDefaultMpi());;
+  GridCartesian *Coarse5d =  SpaceTimeGrid::makeFiveDimGrid(1,Coarse4d);
+
+  //////////////////////////////////////////
+  // Single precision grids -- lanczos + smoother
+  //////////////////////////////////////////
+  GridCartesian         * UGridF   = SpaceTimeGrid::makeFourDimGrid(GridDefaultLatt(),
+								   GridDefaultSimd(Nd,vComplexF::Nsimd()),
+								   GridDefaultMpi());
+  GridRedBlackCartesian * UrbGridF = SpaceTimeGrid::makeFourDimRedBlackGrid(UGridF);
+  GridCartesian         * FGridF   = SpaceTimeGrid::makeFiveDimGrid(Ls,UGridF);
+  GridRedBlackCartesian * FrbGridF = SpaceTimeGrid::makeFiveDimRedBlackGrid(Ls,UGridF);
+  ///////////////////////// RNGs /////////////////////////////////
+  std::vector<int> seeds4({1,2,3,4});
+  std::vector<int> seeds5({5,6,7,8});
+  std::vector<int> cseeds({5,6,7,8});
+
+  GridParallelRNG          RNG5(FGrid);   RNG5.SeedFixedIntegers(seeds5);
+  GridParallelRNG          RNG4(UGrid);   RNG4.SeedFixedIntegers(seeds4);
+  GridParallelRNG          CRNG(Coarse5d);CRNG.SeedFixedIntegers(cseeds);
+
+  ///////////////////////// Configuration /////////////////////////////////
+  LatticeGaugeField Umu(UGrid);
+
+  FieldMetaData header;
+  std::string file("ckpoint_lat.1000");
+  NerscIO::readConfiguration(Umu,header,file);
+
+  //////////////////////// Fermion action //////////////////////////////////
+  MobiusFermionD Ddwf(Umu,*FGrid,*FrbGrid,*UGrid,*UrbGrid,mass,M5,b,c);
+
+  SchurDiagMooeeOperator<MobiusFermionD, LatticeFermion> HermOpEO(Ddwf);
+  
+  const int Fine_Nstop = 200;
+  const int Fine_Nk = 100;
+  const int Fine_Np = 100;
+  const int Fine_Nm = Fine_Nk+Fine_Np;
+
+  typedef LatticeFermion FermionField;
+  std::vector<RealD>        Fine_eval;
+  std::vector<FermionField> Fine_evec;
+
+  LatticeFermionF conv_tmp(FrbGridF);
+  Fine_eval.resize(Fine_Nstop);
+  Fine_evec.resize(Fine_Nstop,FrbGrid);
+  std::string evec_file("/lustre/orion/phy157/proj-shared/phy157_dwf/paboyle/Subspace.phys48.evecF");
+  LoadFineEvecs(Fine_evec,evec_file,conv_tmp);
+  
+  typedef HermOpAdaptor<LatticeFermionD> HermFineMatrix;
+  HermFineMatrix FineHermOp(HermOpEO);
+
+  ////////////////////////////////////////////////////////////
+  ///////////// Coarse basis and Little Dirac Operator ///////
+  ////////////////////////////////////////////////////////////
+  typedef GeneralCoarsenedMatrix<vSpinColourVector,vTComplex,nbasis> LittleDiracOperator;
+  typedef LittleDiracOperator::CoarseVector CoarseVector;
+
+  NextToNextToNextToNearestStencilGeometry5D geom(Coarse5d);
+
+  typedef Aggregation<vSpinColourVector,vTComplex,nbasis> Subspace;
+  Subspace Aggregates(Coarse5d,FrbGrid,cb);
+
+  ////////////////////////////////////////////////////////////
+  // Need to check about red-black grid coarsening
+  ////////////////////////////////////////////////////////////
+  //  std::string subspace_file("/lustre/orion/phy157/proj-shared/phy157_dwf/paboyle/Subspace.phys48.mixed.2500.60");
+  //  //  std::string subspace_file("/lustre/orion/phy157/proj-shared/phy157_dwf/paboyle/Subspace.phys48.new.62");
+  //  std::string refine_file("/lustre/orion/phy157/proj-shared/phy157_dwf/paboyle/Subspace.phys48.evec");
+  std::string refine_file("/lustre/orion/phy157/proj-shared/phy157_dwf/paboyle/Refine.phys48.mixed.2500.60");
+  //  std::string ldop_file("/lustre/orion/phy157/proj-shared/phy157_dwf/paboyle/LittleDiracOp.phys48.mixed.60");
+  //  std::string evec_file("/lustre/orion/phy157/proj-shared/phy157_dwf/paboyle/evecs.scidac");
+  //  std::string eval_file("/lustre/orion/phy157/proj-shared/phy157_dwf/paboyle/eval.xml");
+  bool load_agg=true;
+  bool load_refine=true;
+
+  //////////////////////////////////////////
+  // Block projector for coarse/fine
+  //////////////////////////////////////////
+  MultiRHSBlockProject<LatticeFermionD> MrhsProjector;
+
+
+  /////////////////////////////////////////////////
+  // Mirs smoother
+  /////////////////////////////////////////////////
+  int ord=8;
+  RealD lo=2.0;
+  RealD MirsShift = lo;
+  ShiftedHermOpLinearOperator<LatticeFermionD> ShiftedFineHermOp(HermOpEO,MirsShift);
+  CGSmoother<LatticeFermionD> CGsmooth(ord,ShiftedFineHermOp) ;
+  
+  LoadBasis(Aggregates,refine_file);
+  Aggregates.Orthogonalise();
+
+  std::cout << "**************************************"<<std::endl;
+  std::cout << " Using filtered subspace"<<std::endl;
+  std::cout << "**************************************"<<std::endl;
+  MrhsProjector.Allocate(nbasis,FrbGrid,Coarse5d);
+  MrhsProjector.ImportBasis(Aggregates.subspace);
+
+  FermionField Ftmp(FrbGrid);
+  std::vector<FermionField> Fine_ev(1,FrbGrid);
+  std::vector<FermionField> Fine_ev_compressed(1,FrbGrid);
+  std::vector<CoarseVector>  c_evec(1,Coarse5d);
+  for(int ev=0;ev<Fine_evec.size();ev++){
+    Fine_ev[0] = Fine_evec[ev];
+    MrhsProjector.blockProject(Fine_ev,c_evec);
+    MrhsProjector.blockPromote(Fine_ev_compressed,c_evec);
+    Ftmp = Fine_ev_compressed[0];
+    RealD div = 1.0/ sqrt(norm2(Ftmp));
+    Ftmp = Ftmp * div;
+    std::cout << GridLogMessage<<" "<<ev<<" uncomp "<< norm2(Fine_ev[0])  <<std::endl;
+    std::cout << GridLogMessage<<" "<<ev<<" comp   "<< norm2(Ftmp)  <<std::endl;
+    Ftmp = Fine_ev[0] - Ftmp;
+    std::cout << GridLogMessage<<" "<<ev<<" diff "<< norm2(Ftmp)  <<std::endl;
+    CGsmooth(Fine_ev_compressed[0],Ftmp);
+    Ftmp = Ftmp *lo;
+    std::cout << GridLogMessage<<" "<<ev<<" smoothed "<< norm2(Ftmp)  <<std::endl;
+    div = 1.0/ sqrt(norm2(Ftmp));
+    Ftmp=Ftmp*div;
+    Ftmp = Fine_ev[0]-Ftmp;
+    std::cout << GridLogMessage<<" "<<ev<<" diff "<< norm2(Ftmp)  <<std::endl;
+  }
+
+  std::cout << "**************************************"<<std::endl;
+  std::cout << " Using eigenvector subspace "<<std::endl;
+  std::cout << "**************************************"<<std::endl;
+  for(int i=0;i<Aggregates.subspace.size();i++){
+    Aggregates.subspace[i] = Fine_evec[i];
+  }
+  Aggregates.Orthogonalise();
+  MrhsProjector.ImportBasis(Aggregates.subspace);
+  for(int ev=0;ev<Fine_evec.size();ev++){
+    Fine_ev[0] = Fine_evec[ev];
+    MrhsProjector.blockProject(Fine_ev,c_evec);
+    MrhsProjector.blockPromote(Fine_ev_compressed,c_evec);
+    Ftmp = Fine_ev_compressed[0];
+    RealD div = 1.0/ sqrt(norm2(Ftmp));
+    Ftmp = Ftmp * div;
+    std::cout << GridLogMessage<<" "<<ev<<" uncomp "<< norm2(Fine_ev[0])  <<std::endl;
+    std::cout << GridLogMessage<<" "<<ev<<" comp   "<< norm2(Ftmp)  <<std::endl;
+    Ftmp = Fine_ev[0] - Ftmp;
+    std::cout << GridLogMessage<<" "<<ev<<" diff "<< norm2(Ftmp)  <<std::endl;
+    CGsmooth(Fine_ev_compressed[0],Ftmp);
+    Ftmp = Ftmp *lo;
+    std::cout << GridLogMessage<<" "<<ev<<" smoothed "<< norm2(Ftmp)  <<std::endl;
+    div = 1.0/ sqrt(norm2(Ftmp));
+    Ftmp=Ftmp*div;
+    Ftmp = Fine_ev[0]-Ftmp;
+    std::cout << GridLogMessage<<" "<<ev<<" diff "<< norm2(Ftmp)  <<std::endl;
+  }
+
+  // Standard CG
+  Grid_finalize();
+  return 0;
+}
--- a/tests/debug/Test_general_coarse_pvdagm.cc
+++ b/tests/debug/Test_general_coarse_pvdagm.cc
@ -36,28 +36,6 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 using namespace std;
 using namespace Grid;

-template<class Field>
-class HermOpAdaptor : public LinearOperatorBase<Field>
-{
-  LinearOperatorBase<Field> & wrapped;
-public:
-  HermOpAdaptor(LinearOperatorBase<Field> &wrapme) : wrapped(wrapme)  {};
-  void OpDiag (const Field &in, Field &out) {    assert(0);  }
-  void OpDir  (const Field &in, Field &out,int dir,int disp) {    assert(0);  }
-  void OpDirAll  (const Field &in, std::vector<Field> &out){    assert(0);  };
-  void Op     (const Field &in, Field &out){
-    wrapped.HermOp(in,out);
-  }
-  void AdjOp     (const Field &in, Field &out){
-    wrapped.HermOp(in,out);
-  }
-  void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){    assert(0);  }
-  void HermOp(const Field &in, Field &out){
-    wrapped.HermOp(in,out);
-  }
-  
-};
-
 template<class Matrix,class Field>
 class PVdagMLinearOperator : public LinearOperatorBase<Field> {
  Matrix &_Mat;
@ -69,78 +47,169 @@ public:
  void OpDir  (const Field &in, Field &out,int dir,int disp) {    assert(0);  }
  void OpDirAll  (const Field &in, std::vector<Field> &out){    assert(0);  };
  void Op     (const Field &in, Field &out){
+    std::cout << "Op: PVdag M "<<std::endl;
    Field tmp(in.Grid());
    _Mat.M(in,tmp);
    _PV.Mdag(tmp,out);
  }
  void AdjOp     (const Field &in, Field &out){
+    std::cout << "AdjOp: Mdag PV "<<std::endl;
    Field tmp(in.Grid());
-    _PV.M(tmp,out);
-    _Mat.Mdag(in,tmp);
+    _PV.M(in,tmp);
+    _Mat.Mdag(tmp,out);
  }
  void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){    assert(0);  }
  void HermOp(const Field &in, Field &out){
-    std::cout << "HermOp"<<std::endl;
+    std::cout << "HermOp: Mdag PV PVdag M"<<std::endl;
+    Field tmp(in.Grid());
+    //    _Mat.M(in,tmp);
+    //    _PV.Mdag(tmp,out);
+    //    _PV.M(out,tmp);
+    //    _Mat.Mdag(tmp,out);
+    Op(in,tmp);
+    AdjOp(tmp,out);
+    //    std::cout << "HermOp done "<<norm2(out)<<std::endl;
+  }
+};
+template<class Matrix,class Field>
+class ShiftedPVdagMLinearOperator : public LinearOperatorBase<Field> {
+  Matrix &_Mat;
+  Matrix &_PV;
+  RealD shift;
+public:
+  ShiftedPVdagMLinearOperator(RealD _shift,Matrix &Mat,Matrix &PV): shift(_shift),_Mat(Mat),_PV(PV){};
+
+  void OpDiag (const Field &in, Field &out) {    assert(0);  }
+  void OpDir  (const Field &in, Field &out,int dir,int disp) {    assert(0);  }
+  void OpDirAll  (const Field &in, std::vector<Field> &out){    assert(0);  };
+  void Op     (const Field &in, Field &out){
+    std::cout << "Op: PVdag M "<<std::endl;
    Field tmp(in.Grid());
    _Mat.M(in,tmp);
    _PV.Mdag(tmp,out);
-    _PV.M(out,tmp);
-    _Mat.Mdag(tmp,out);
-    std::cout << "HermOp done "<<norm2(out)<<std::endl;
-    
+    out = out + shift * in;
  }
-};
-
-template<class Field> class DumbOperator  : public LinearOperatorBase<Field> {
-public:
-  LatticeComplex scale;
-  DumbOperator(GridBase *grid) : scale(grid)
-  {
-    scale = 0.0;
-    LatticeComplex scalesft(grid);
-    LatticeComplex scaletmp(grid);
-    for(int d=0;d<4;d++){
-      Lattice<iScalar<vInteger> > x(grid); LatticeCoordinate(x,d+1);
-      LatticeCoordinate(scaletmp,d+1);
-      scalesft = Cshift(scaletmp,d+1,1);
-      scale = 100.0*scale + where( mod(x    ,2)==(Integer)0, scalesft,scaletmp);
-    }
-    std::cout << " scale\n" << scale << std::endl;
-  }
-  // Support for coarsening to a multigrid
-  void OpDiag (const Field &in, Field &out) {};
-  void OpDir  (const Field &in, Field &out,int dir,int disp){};
-  void OpDirAll  (const Field &in, std::vector<Field> &out) {};
-
-  void Op     (const Field &in, Field &out){
-    out = scale * in;
-  }
-  void AdjOp  (const Field &in, Field &out){
-    out = scale * in;
+  void AdjOp     (const Field &in, Field &out){
+    std::cout << "AdjOp: Mdag PV "<<std::endl;
+    Field tmp(in.Grid());
+    _PV.M(tmp,out);
+    _Mat.Mdag(in,tmp);
+    out = out + shift * in;
  }
+  void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){    assert(0);  }
  void HermOp(const Field &in, Field &out){
-    double n1, n2;
-    HermOpAndNorm(in,out,n1,n2);
-  }
-  void HermOpAndNorm(const Field &in, Field &out,double &n1,double &n2){
-    ComplexD dot;
-
-    out = scale * in;
-
-    dot= innerProduct(in,out);
-    n1=real(dot);
-
-    dot = innerProduct(out,out);
-    n2=real(dot);
+    std::cout << "HermOp: Mdag PV PVdag M"<<std::endl;
+    Field tmp(in.Grid());
+    Op(in,tmp);
+    AdjOp(tmp,out);
  }
 };
+template<class Fobj,class CComplex,int nbasis>
+class MGPreconditioner : public LinearFunction< Lattice<Fobj> > {
+public:
+  using LinearFunction<Lattice<Fobj> >::operator();

+  typedef Aggregation<Fobj,CComplex,nbasis> Aggregates;
+  typedef typename Aggregation<Fobj,CComplex,nbasis>::FineField    FineField;
+  typedef typename Aggregation<Fobj,CComplex,nbasis>::CoarseVector CoarseVector;
+  typedef typename Aggregation<Fobj,CComplex,nbasis>::CoarseMatrix CoarseMatrix;
+  typedef LinearOperatorBase<FineField>                            FineOperator;
+  typedef LinearFunction    <FineField>                            FineSmoother;
+  typedef LinearOperatorBase<CoarseVector>                         CoarseOperator;
+  typedef LinearFunction    <CoarseVector>                         CoarseSolver;
+  Aggregates     & _Aggregates;
+  FineOperator   & _FineOperator;
+  FineSmoother   & _PreSmoother;
+  FineSmoother   & _PostSmoother;
+  CoarseOperator & _CoarseOperator;
+  CoarseSolver   & _CoarseSolve;
+
+  int    level;  void Level(int lv) {level = lv; };
+
+  MGPreconditioner(Aggregates &Agg,
+		   FineOperator &Fine,
+		   FineSmoother &PreSmoother,
+		   FineSmoother &PostSmoother,
+		   CoarseOperator &CoarseOperator_,
+		   CoarseSolver &CoarseSolve_)
+    : _Aggregates(Agg),
+      _FineOperator(Fine),
+      _PreSmoother(PreSmoother),
+      _PostSmoother(PostSmoother),
+      _CoarseOperator(CoarseOperator_),
+      _CoarseSolve(CoarseSolve_),
+      level(1)  {  }
+
+  virtual void operator()(const FineField &in, FineField & out) 
+  {
+    GridBase *CoarseGrid = _Aggregates.CoarseGrid;
+    //    auto CoarseGrid = _CoarseOperator.Grid();
+    CoarseVector Csrc(CoarseGrid);
+    CoarseVector Csol(CoarseGrid);
+    FineField vec1(in.Grid());
+    FineField vec2(in.Grid());
+
+    std::cout<<GridLogMessage << "Calling PreSmoother " <<std::endl;
+
+    //    std::cout<<GridLogMessage << "Calling PreSmoother input residual "<<norm2(in) <<std::endl;
+    double t;
+    // Fine Smoother
+    //    out = in;
+    out = Zero();
+    t=-usecond();
+    _PreSmoother(in,out);
+    t+=usecond();
+
+    std::cout<<GridLogMessage << "PreSmoother took "<< t/1000.0<< "ms" <<std::endl;
+
+    // Update the residual
+    _FineOperator.Op(out,vec1);  sub(vec1, in ,vec1);   
+    //    std::cout<<GridLogMessage <<"Residual-1 now " <<norm2(vec1)<<std::endl;
+
+    // Fine to Coarse 
+    t=-usecond();
+    _Aggregates.ProjectToSubspace  (Csrc,vec1);
+    t+=usecond();
+    std::cout<<GridLogMessage << "Project to coarse took "<< t/1000.0<< "ms" <<std::endl;
+
+    // Coarse correction
+    t=-usecond();
+    Csol = Zero();
+    _CoarseSolve(Csrc,Csol);
+    //Csol=Zero();
+    t+=usecond();
+    std::cout<<GridLogMessage << "Coarse solve took "<< t/1000.0<< "ms" <<std::endl;
+
+    // Coarse to Fine
+    t=-usecond();  
+    //    _CoarseOperator.PromoteFromSubspace(_Aggregates,Csol,vec1);
+    _Aggregates.PromoteFromSubspace(Csol,vec1); 
+    add(out,out,vec1);
+    t+=usecond();
+    std::cout<<GridLogMessage << "Promote to this level took "<< t/1000.0<< "ms" <<std::endl;
+
+    // Residual
+    _FineOperator.Op(out,vec1);  sub(vec1 ,in , vec1);  
+    //    std::cout<<GridLogMessage <<"Residual-2 now " <<norm2(vec1)<<std::endl;
+
+    // Fine Smoother
+    t=-usecond();
+    //    vec2=vec1;
+    vec2=Zero();
+    _PostSmoother(vec1,vec2);
+    t+=usecond();
+    std::cout<<GridLogMessage << "PostSmoother took "<< t/1000.0<< "ms" <<std::endl;
+
+    add( out,out,vec2);
+    std::cout<<GridLogMessage << "Done " <<std::endl;
+  }
+};

 int main (int argc, char ** argv)
 {
  Grid_init(&argc,&argv);

-  const int Ls=2;
+  const int Ls=16;

  GridCartesian         * UGrid   = SpaceTimeGrid::makeFourDimGrid(GridDefaultLatt(), GridDefaultSimd(Nd,vComplex::Nsimd()),GridDefaultMpi());
  GridRedBlackCartesian * UrbGrid = SpaceTimeGrid::makeFourDimRedBlackGrid(UGrid);
@ -151,7 +220,8 @@ int main (int argc, char ** argv)
  // Construct a coarsened grid
  Coordinate clatt = GridDefaultLatt();
  for(int d=0;d<clatt.size();d++){
-    clatt[d] = clatt[d]/4;
+    clatt[d] = clatt[d]/2;
+    //    clatt[d] = clatt[d]/4;
  }
  GridCartesian *Coarse4d =  SpaceTimeGrid::makeFourDimGrid(clatt, GridDefaultSimd(Nd,vComplex::Nsimd()),GridDefaultMpi());;
  GridCartesian *Coarse5d =  SpaceTimeGrid::makeFiveDimGrid(1,Coarse4d);
@ -173,15 +243,14 @@ int main (int argc, char ** argv)
  FieldMetaData header;
  std::string file("ckpoint_lat.4000");
  NerscIO::readConfiguration(Umu,header,file);
-  //Umu = 1.0;
  
-  RealD mass=0.5;
+  RealD mass=0.01;
  RealD M5=1.8;

  DomainWallFermionD Ddwf(Umu,*FGrid,*FrbGrid,*UGrid,*UrbGrid,mass,M5);
  DomainWallFermionD Dpv(Umu,*FGrid,*FrbGrid,*UGrid,*UrbGrid,1.0,M5);

-  const int nbasis = 1;
+  const int nbasis = 20;
  const int cb = 0 ;
  LatticeFermion prom(FGrid);

@ -193,25 +262,51 @@ int main (int argc, char ** argv)
  std::cout<<GridLogMessage<<std::endl;
  std::cout<<GridLogMessage<<"*******************************************"<<std::endl;
  std::cout<<GridLogMessage<<std::endl;
-  
-  PVdagMLinearOperator<DomainWallFermionD,LatticeFermionD> PVdagM(Ddwf,Dpv);
-  HermOpAdaptor<LatticeFermionD> HOA(PVdagM);
+
+  typedef PVdagMLinearOperator<DomainWallFermionD,LatticeFermionD> PVdagM_t;
+  typedef ShiftedPVdagMLinearOperator<DomainWallFermionD,LatticeFermionD> ShiftedPVdagM_t;
+  PVdagM_t PVdagM(Ddwf,Dpv);
+  //  ShiftedPVdagM_t ShiftedPVdagM(2.0,Ddwf,Dpv); // 355
+  //  ShiftedPVdagM_t ShiftedPVdagM(1.0,Ddwf,Dpv); // 246
+  //  ShiftedPVdagM_t ShiftedPVdagM(0.5,Ddwf,Dpv); // 183
+  //  ShiftedPVdagM_t ShiftedPVdagM(0.25,Ddwf,Dpv); // 145
+  //  ShiftedPVdagM_t ShiftedPVdagM(0.1,Ddwf,Dpv); // 134
+  //  ShiftedPVdagM_t ShiftedPVdagM(0.1,Ddwf,Dpv); // 127 -- NULL space via inverse iteration
+  //  ShiftedPVdagM_t ShiftedPVdagM(0.1,Ddwf,Dpv); // 57 -- NULL space via inverse iteration; 3 iterations
+  //  ShiftedPVdagM_t ShiftedPVdagM(0.25,Ddwf,Dpv); // 57 , tighter inversion
+  //  ShiftedPVdagM_t ShiftedPVdagM(0.25,Ddwf,Dpv); // nbasis 20 -- 49 iters
+  //  ShiftedPVdagM_t ShiftedPVdagM(0.25,Ddwf,Dpv); // nbasis 20 -- 70 iters; asymmetric 
+  //  ShiftedPVdagM_t ShiftedPVdagM(0.25,Ddwf,Dpv); // 58; Loosen coarse, tighten fine
+  //  ShiftedPVdagM_t ShiftedPVdagM(0.1,Ddwf,Dpv); // 56 ... 
+  //  ShiftedPVdagM_t ShiftedPVdagM(0.1,Ddwf,Dpv); // 51 ...  with 24 vecs
+  //  ShiftedPVdagM_t ShiftedPVdagM(0.1,Ddwf,Dpv); // 31 ...  with 24 vecs and 2^4 blocking
+  //  ShiftedPVdagM_t ShiftedPVdagM(0.1,Ddwf,Dpv); // 43 ...  with 16 vecs and 2^4 blocking, sloppier
+  //  ShiftedPVdagM_t ShiftedPVdagM(0.1,Ddwf,Dpv); // 35  ...  with 20 vecs and 2^4 blocking
+  //  ShiftedPVdagM_t ShiftedPVdagM(0.1,Ddwf,Dpv); // 35  ...  with 20 vecs and 2^4 blocking, looser coarse
+  //  ShiftedPVdagM_t ShiftedPVdagM(0.1,Ddwf,Dpv); // 64  ...  with 20 vecs, Christoph setup, and 2^4 blocking, looser coarse
+  ShiftedPVdagM_t ShiftedPVdagM(0.01,Ddwf,Dpv); // 
+

  // Run power method on HOA??
-  PowerMethod<LatticeFermion>       PM;   PM(HOA,src);
+  //  PowerMethod<LatticeFermion>       PM;   PM(PVdagM,src);
 
  // Warning: This routine calls PVdagM.Op, not PVdagM.HermOp
  typedef Aggregation<vSpinColourVector,vTComplex,nbasis> Subspace;
  Subspace AggregatesPD(Coarse5d,FGrid,cb);
+  /*
  AggregatesPD.CreateSubspaceChebyshev(RNG5,
-				       HOA,
+				       PVdagM,
 				       nbasis,
-				       5000.0,
-				       0.02,
-				       100,
-				       50,
-				       50,
+				       4000.0,
+				       2.0,
+				       200,
+				       200,
+				       200,
 				       0.0);
+  */
+  AggregatesPD.CreateSubspaceGCR(RNG5,
+				 PVdagM,
+				 nbasis);
  
  LittleDiracOperator LittleDiracOpPV(geom,FGrid,Coarse5d);
  LittleDiracOpPV.CoarsenOperator(PVdagM,AggregatesPD);
@ -257,6 +352,60 @@ int main (int argc, char ** argv)
  std::cout<<GridLogMessage<<" ldop error: "<<norm2(c_proj)<<std::endl;
  //  std::cout<<GridLogMessage<<" error "<< c_proj<<std::endl;

+
+  /**********
+   * Some solvers
+   **********
+   */
+
+  ///////////////////////////////////////
+  // Coarse grid solver test
+  ///////////////////////////////////////
+
+  std::cout<<GridLogMessage<<"******************* "<<std::endl;
+  std::cout<<GridLogMessage<<" Coarse Grid Solve -- Level 3 "<<std::endl;
+  std::cout<<GridLogMessage<<"******************* "<<std::endl;
+  TrivialPrecon<CoarseVector> simple;
+  NonHermitianLinearOperator<LittleDiracOperator,CoarseVector> LinOpCoarse(LittleDiracOpPV);
+  //  PrecGeneralisedConjugateResidualNonHermitian<CoarseVector>  L2PGCR(1.0e-4, 100, LinOpCoarse,simple,10,10); 
+  PrecGeneralisedConjugateResidualNonHermitian<CoarseVector>  L2PGCR(3.0e-2, 100, LinOpCoarse,simple,10,10); 
+  L2PGCR.Level(3);
+  c_res=Zero();
+  L2PGCR(c_src,c_res);
+
+  ////////////////////////////////////////
+  // Fine grid smoother
+  ////////////////////////////////////////
+  std::cout<<GridLogMessage<<"******************* "<<std::endl;
+  std::cout<<GridLogMessage<<" Fine Grid Smoother -- Level 2 "<<std::endl;
+  std::cout<<GridLogMessage<<"******************* "<<std::endl;
+  TrivialPrecon<LatticeFermionD> simple_fine;
+  //  NonHermitianLinearOperator<PVdagM_t,LatticeFermionD> LinOpSmooth(PVdagM);
+  PrecGeneralisedConjugateResidualNonHermitian<LatticeFermionD> SmootherGCR(0.01,1,ShiftedPVdagM,simple_fine,16,16);
+  SmootherGCR.Level(2);
+  
+  LatticeFermionD f_src(FGrid);
+  LatticeFermionD f_res(FGrid);
+
+  f_src = one;  // 1 in every element for vector 1.
+  f_res=Zero();
+  SmootherGCR(f_src,f_res);
+
+  typedef MGPreconditioner<vSpinColourVector,  vTComplex,nbasis> TwoLevelMG;
+
+  TwoLevelMG TwoLevelPrecon(AggregatesPD,
+			    PVdagM,
+			    simple_fine,
+			    SmootherGCR,
+			    LinOpCoarse,
+			    L2PGCR);
+  
+  PrecGeneralisedConjugateResidualNonHermitian<LatticeFermion> L1PGCR(1.0e-8,1000,PVdagM,TwoLevelPrecon,16,16);
+  L1PGCR.Level(1);
+
+  f_res=Zero();
+  L1PGCR(f_src,f_res);
+
  std::cout<<GridLogMessage<<std::endl;
  std::cout<<GridLogMessage<<"*******************************************"<<std::endl;
  std::cout<<GridLogMessage<<std::endl;
--- a/tests/lanczos/LanParams.xml
+++ b/tests/lanczos/LanParams.xml
@ -0,0 +1,6 @@
+<?xml version="1.0"?>
+<grid>
+  <LanczosParameters>
+    <mass>-3.5</mass>
+  </LanczosParameters>
+</grid>
--- a/tests/lanczos/Test_wilson_DWFKernel.cc
+++ b/tests/lanczos/Test_wilson_DWFKernel.cc
@ -0,0 +1,278 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./tests/Test_dwf_lanczos.cc
+
+Copyright (C) 2015
+
+Author: Chulwoo Jung <chulwoo@bnl.gov>
+
+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 */
+#include <Grid/Grid.h>
+
+using namespace std;
+using namespace Grid;
+ ;
+
+typedef WilsonFermionD FermionOp;
+typedef typename WilsonFermionD::FermionField FermionField;
+
+
+RealD AllZero(RealD x) { return 0.; }
+
+namespace Grid {
+
+#if 0
+template<typename Field>
+class RationalHermOp : public LinearFunction<Field> {
+public:
+  using LinearFunction<Field>::operator();  
+//  OperatorFunction<Field>   & _poly;
+  LinearOperatorBase<Field> &_Linop;
+  RealD _massDen, _massNum;
+   
+  FunctionHermOp(LinearOperatorBase<Field>& linop, RealD massDen,RealD massNum)
+    :  _Linop(linop) ,_massDen(massDen),_massNum(massNum) {};
+      
+  void operator()(const Field& in, Field& out) {
+//    _poly(_Linop,in,out);
+  } 
+};
+#endif
+
+template<class Matrix,class Field>
+class InvG5LinearOperator : public LinearOperatorBase<Field> {
+  Matrix &_Mat;
+  RealD _num;
+  RealD _Tol;
+  Integer _MaxIt;
+  Gamma g5;
+
+public:
+  InvG5LinearOperator(Matrix &Mat,RealD num): _Mat(Mat),_num(num), _Tol(1e-12),_MaxIt(10000), g5(Gamma::Algebra::Gamma5) {};
+
+  // Support for coarsening to a multigrid
+  void OpDiag (const Field &in, Field &out) {
+    assert(0);
+    _Mat.Mdiag(in,out);
+  }
+  void OpDir  (const Field &in, Field &out,int dir,int disp) {
+    assert(0);
+    _Mat.Mdir(in,out,dir,disp);
+  }
+  void OpDirAll  (const Field &in, std::vector<Field> &out){
+    assert(0);
+    _Mat.MdirAll(in,out);
+  };
+  void Op     (const Field &in, Field &out){
+    assert(0);
+    _Mat.M(in,out);
+  }
+  void AdjOp     (const Field &in, Field &out){
+    assert(0);
+    _Mat.Mdag(in,out);
+  }
+  void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){
+    HermOp(in,out);
+    ComplexD dot = innerProduct(in,out);
+    n1=real(dot);
+    n2=norm2(out);
+  }
+  void HermOp(const Field &in, Field &out){
+     Field tmp(in.Grid());
+     MdagMLinearOperator<Matrix,Field> denom(_Mat);
+     ConjugateGradient<Field> CG(_Tol,_MaxIt); 
+     _Mat.M(in,tmp);
+     tmp += _num*in;
+     _Mat.Mdag(tmp,out);
+     CG(denom,out,tmp);
+     out = g5*tmp;
+  }
+};
+
+
+struct LanczosParameters: Serializable {
+  GRID_SERIALIZABLE_CLASS_MEMBERS(LanczosParameters,
+		  		RealD, mass , 
+				RealD, resid,
+	  			RealD, ChebyLow,
+	  			RealD, ChebyHigh,
+	  			Integer, ChebyOrder)
+//                                  Integer, StartTrajectory,
+//                                  Integer, Trajectories, /* @brief Number of sweeps in this run */
+//                                  bool, MetropolisTest,
+//                                  Integer, NoMetropolisUntil,
+//                                  std::string, StartingType,
+//                                  Integer, SW,
+//				  RealD, Kappa,
+//                                  IntegratorParameters, MD)
+
+  LanczosParameters() {
+    ////////////////////////////// Default values
+      mass = 0;
+//    MetropolisTest    = true;
+//    NoMetropolisUntil = 10;
+//    StartTrajectory   = 0;
+//    SW                = 2;
+//    Trajectories      = 10;
+//    StartingType      = "HotStart";
+    /////////////////////////////////
+  }
+
+  template <class ReaderClass >
+  LanczosParameters(Reader<ReaderClass> & TheReader){
+    initialize(TheReader);
+  }
+
+  template < class ReaderClass > 
+  void initialize(Reader<ReaderClass> &TheReader){
+//    std::cout << GridLogMessage << "Reading HMC\n";
+    read(TheReader, "HMC", *this);
+  }
+
+
+  void print_parameters() const {
+//    std::cout << GridLogMessage << "[HMC parameters] Trajectories            : " << Trajectories << "\n";
+//    std::cout << GridLogMessage << "[HMC parameters] Start trajectory        : " << StartTrajectory << "\n";
+//    std::cout << GridLogMessage << "[HMC parameters] Metropolis test (on/off): " << std::boolalpha << MetropolisTest << "\n";
+//    std::cout << GridLogMessage << "[HMC parameters] Thermalization trajs    : " << NoMetropolisUntil << "\n";
+//    std::cout << GridLogMessage << "[HMC parameters] Starting type           : " << StartingType << "\n";
+//    MD.print_parameters();
+  }
+  
+};
+
+}
+
+int main(int argc, char** argv) {
+  Grid_init(&argc, &argv);
+
+  GridCartesian* UGrid = SpaceTimeGrid::makeFourDimGrid(
+      GridDefaultLatt(), GridDefaultSimd(Nd, vComplex::Nsimd()),
+      GridDefaultMpi());
+  GridRedBlackCartesian* UrbGrid =
+      SpaceTimeGrid::makeFourDimRedBlackGrid(UGrid);
+  GridCartesian* FGrid = UGrid;
+  GridRedBlackCartesian* FrbGrid = UrbGrid;
+//  printf("UGrid=%p UrbGrid=%p FGrid=%p FrbGrid=%p\n", UGrid, UrbGrid, FGrid, FrbGrid);
+
+  std::vector<int> seeds4({1, 2, 3, 4});
+  std::vector<int> seeds5({5, 6, 7, 8});
+  GridParallelRNG RNG5(FGrid);
+  RNG5.SeedFixedIntegers(seeds5);
+  GridParallelRNG RNG4(UGrid);
+  RNG4.SeedFixedIntegers(seeds4);
+  GridParallelRNG RNG5rb(FrbGrid);
+  RNG5.SeedFixedIntegers(seeds5);
+
+  LatticeGaugeField Umu(UGrid);
+//  SU<Nc>::HotConfiguration(RNG4, Umu);
+
+  FieldMetaData header;
+  std::string file("./config");
+
+  int precision32 = 0;
+  int tworow      = 0;
+//  NerscIO::writeConfiguration(Umu,file,tworow,precision32);
+  NerscIO::readConfiguration(Umu,header,file);
+
+/*
+  std::vector<LatticeColourMatrix> U(4, UGrid);
+  for (int mu = 0; mu < Nd; mu++) {
+    U[mu] = PeekIndex<LorentzIndex>(Umu, mu);
+  }
+*/
+
+  int Nstop = 5;
+  int Nk = 10;
+  int Np = 90;
+  int Nm = Nk + Np;
+  int MaxIt = 10000;
+  RealD resid = 1.0e-5;
+
+  RealD mass = -1.0;
+
+  LanczosParameters LanParams;
+#if 1
+  {
+    XmlReader  HMCrd("LanParams.xml");
+    read(HMCrd,"LanczosParameters",LanParams);
+  }
+#else
+  {
+    LanParams.mass = mass;
+  }
+#endif
+  std::cout << GridLogMessage<< LanParams <<std::endl;
+  { 
+    XmlWriter HMCwr("LanParams.xml.out");
+    write(HMCwr,"LanczosParameters",LanParams);
+  }
+
+  mass=LanParams.mass;
+  resid=LanParams.resid;
+
+
+while ( mass > - 5.0){
+  FermionOp WilsonOperator(Umu,*FGrid,*FrbGrid,2.+mass);
+  InvG5LinearOperator<FermionOp,LatticeFermion> HermOp(WilsonOperator,-2.); /// <-----
+  //SchurDiagTwoOperator<FermionOp,FermionField> HermOp(WilsonOperator);
+//  Gamma5HermitianLinearOperator <FermionOp,LatticeFermion> HermOp2(WilsonOperator); /// <-----
+
+  std::vector<double> Coeffs{0, 0, 1.};
+  Polynomial<FermionField> PolyX(Coeffs);
+  Chebyshev<FermionField> Cheby(LanParams.ChebyLow,LanParams.ChebyHigh,LanParams.ChebyOrder);
+
+       FunctionHermOp<FermionField> OpCheby(Cheby,HermOp);
+//     InvHermOp<FermionField> Op(WilsonOperator,HermOp);
+     PlainHermOp<FermionField> Op     (HermOp);
+//     PlainHermOp<FermionField> Op2     (HermOp2);
+
+  ImplicitlyRestartedLanczos<FermionField> IRL(OpCheby, Op, Nstop, Nk, Nm, resid, MaxIt);
+
+  std::vector<RealD> eval(Nm);
+  FermionField src(FGrid);
+  gaussian(RNG5, src);
+  std::vector<FermionField> evec(Nm, FGrid);
+  for (int i = 0; i < 1; i++) {
+    std::cout << i << " / " << Nm << " grid pointer " << evec[i].Grid()
+              << std::endl;
+  };
+
+  int Nconv;
+  IRL.calc(eval, evec, src, Nconv);
+
+  std::cout << mass <<" : " << eval << std::endl;
+
+  Gamma g5(Gamma::Algebra::Gamma5) ;
+  ComplexD dot;
+  FermionField tmp(FGrid);
+  for (int i = 0; i < Nstop ; i++) {
+    tmp = g5*evec[i];
+    dot = innerProduct(tmp,evec[i]);
+    std::cout << mass << " : " << eval[i]  << " " << real(dot) << " " << imag(dot)  << std::endl ;
+  }
+  src  = evec[0]+evec[1]+evec[2];
+  mass += -0.1;
+}
+
+  Grid_finalize();
+}
--- a/tests/lanczos/Test_wilson_specflow.cc
+++ b/tests/lanczos/Test_wilson_specflow.cc
@ -0,0 +1,211 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./tests/Test_dwf_lanczos.cc
+
+Copyright (C) 2015
+
+Author: Chulwoo Jung <chulwoo@bnl.gov>
+
+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 */
+#include <Grid/Grid.h>
+
+using namespace std;
+using namespace Grid;
+ ;
+
+typedef WilsonFermionD FermionOp;
+typedef typename WilsonFermionD::FermionField FermionField;
+
+
+RealD AllZero(RealD x) { return 0.; }
+
+namespace Grid {
+
+struct LanczosParameters: Serializable {
+  GRID_SERIALIZABLE_CLASS_MEMBERS(LanczosParameters,
+		  		RealD, mass , 
+	  			RealD, ChebyLow,
+	  			RealD, ChebyHigh,
+	  			Integer, ChebyOrder)
+//                                  Integer, StartTrajectory,
+//                                  Integer, Trajectories, /* @brief Number of sweeps in this run */
+//                                  bool, MetropolisTest,
+//                                  Integer, NoMetropolisUntil,
+//                                  std::string, StartingType,
+//                                  Integer, SW,
+//				  RealD, Kappa,
+//                                  IntegratorParameters, MD)
+
+  LanczosParameters() {
+    ////////////////////////////// Default values
+      mass = 0;
+//    MetropolisTest    = true;
+//    NoMetropolisUntil = 10;
+//    StartTrajectory   = 0;
+//    SW                = 2;
+//    Trajectories      = 10;
+//    StartingType      = "HotStart";
+    /////////////////////////////////
+  }
+
+  template <class ReaderClass >
+  LanczosParameters(Reader<ReaderClass> & TheReader){
+    initialize(TheReader);
+  }
+
+  template < class ReaderClass > 
+  void initialize(Reader<ReaderClass> &TheReader){
+//    std::cout << GridLogMessage << "Reading HMC\n";
+    read(TheReader, "HMC", *this);
+  }
+
+
+  void print_parameters() const {
+//    std::cout << GridLogMessage << "[HMC parameters] Trajectories            : " << Trajectories << "\n";
+//    std::cout << GridLogMessage << "[HMC parameters] Start trajectory        : " << StartTrajectory << "\n";
+//    std::cout << GridLogMessage << "[HMC parameters] Metropolis test (on/off): " << std::boolalpha << MetropolisTest << "\n";
+//    std::cout << GridLogMessage << "[HMC parameters] Thermalization trajs    : " << NoMetropolisUntil << "\n";
+//    std::cout << GridLogMessage << "[HMC parameters] Starting type           : " << StartingType << "\n";
+//    MD.print_parameters();
+  }
+  
+};
+
+}
+
+int main(int argc, char** argv) {
+  Grid_init(&argc, &argv);
+
+  GridCartesian* UGrid = SpaceTimeGrid::makeFourDimGrid(
+      GridDefaultLatt(), GridDefaultSimd(Nd, vComplex::Nsimd()),
+      GridDefaultMpi());
+  GridRedBlackCartesian* UrbGrid =
+      SpaceTimeGrid::makeFourDimRedBlackGrid(UGrid);
+  GridCartesian* FGrid = UGrid;
+  GridRedBlackCartesian* FrbGrid = UrbGrid;
+//  printf("UGrid=%p UrbGrid=%p FGrid=%p FrbGrid=%p\n", UGrid, UrbGrid, FGrid, FrbGrid);
+
+  std::vector<int> seeds4({1, 2, 3, 4});
+  std::vector<int> seeds5({5, 6, 7, 8});
+  GridParallelRNG RNG5(FGrid);
+  RNG5.SeedFixedIntegers(seeds5);
+  GridParallelRNG RNG4(UGrid);
+  RNG4.SeedFixedIntegers(seeds4);
+  GridParallelRNG RNG5rb(FrbGrid);
+  RNG5.SeedFixedIntegers(seeds5);
+
+  LatticeGaugeField Umu(UGrid);
+//  SU<Nc>::HotConfiguration(RNG4, Umu);
+
+  FieldMetaData header;
+  std::string file("./config");
+
+  int precision32 = 0;
+  int tworow      = 0;
+//  NerscIO::writeConfiguration(Umu,file,tworow,precision32);
+  NerscIO::readConfiguration(Umu,header,file);
+
+/*
+  std::vector<LatticeColourMatrix> U(4, UGrid);
+  for (int mu = 0; mu < Nd; mu++) {
+    U[mu] = PeekIndex<LorentzIndex>(Umu, mu);
+  }
+*/
+
+  int Nstop = 10;
+  int Nk = 20;
+  int Np = 80;
+  int Nm = Nk + Np;
+  int MaxIt = 10000;
+  RealD resid = 1.0e-5;
+
+  RealD mass = -1.0;
+
+  LanczosParameters LanParams;
+#if 1
+  {
+    XmlReader  HMCrd("LanParams.xml");
+    read(HMCrd,"LanczosParameters",LanParams);
+  }
+#else
+  {
+    LanParams.mass = mass;
+  }
+#endif
+  std::cout << GridLogMessage<< LanParams <<std::endl;
+  { 
+    XmlWriter HMCwr("LanParams.xml.out");
+    write(HMCwr,"LanczosParameters",LanParams);
+  }
+
+  mass=LanParams.mass;
+
+
+while ( mass > - 5.0){
+  FermionOp WilsonOperator(Umu,*FGrid,*FrbGrid,mass);
+  MdagMLinearOperator<FermionOp,FermionField> HermOp(WilsonOperator); /// <-----
+  //SchurDiagTwoOperator<FermionOp,FermionField> HermOp(WilsonOperator);
+  Gamma5HermitianLinearOperator <FermionOp,LatticeFermion> HermOp2(WilsonOperator); /// <-----
+
+  std::vector<double> Coeffs{0, 1.};
+  Polynomial<FermionField> PolyX(Coeffs);
+//  Chebyshev<FermionField> Cheby(0.5, 60., 31);
+//                                  RealD, ChebyLow,
+//                                RealD, ChebyHigh,
+//                                Integer, ChebyOrder)
+
+  Chebyshev<FermionField> Cheby(LanParams.ChebyLow,LanParams.ChebyHigh,LanParams.ChebyOrder);
+
+  FunctionHermOp<FermionField> OpCheby(Cheby,HermOp);
+     PlainHermOp<FermionField> Op     (HermOp);
+     PlainHermOp<FermionField> Op2     (HermOp2);
+
+  ImplicitlyRestartedLanczos<FermionField> IRL(OpCheby, Op2, Nstop, Nk, Nm, resid, MaxIt);
+
+  std::vector<RealD> eval(Nm);
+  FermionField src(FGrid);
+  gaussian(RNG5, src);
+  std::vector<FermionField> evec(Nm, FGrid);
+  for (int i = 0; i < 1; i++) {
+    std::cout << i << " / " << Nm << " grid pointer " << evec[i].Grid()
+              << std::endl;
+  };
+
+  int Nconv;
+  IRL.calc(eval, evec, src, Nconv);
+
+  std::cout << mass <<" : " << eval << std::endl;
+
+  Gamma g5(Gamma::Algebra::Gamma5) ;
+  ComplexD dot;
+  FermionField tmp(FGrid);
+  for (int i = 0; i < Nstop ; i++) {
+    tmp = g5*evec[i];
+    dot = innerProduct(tmp,evec[i]);
+    std::cout << mass << " : " << eval[i]  << " " << real(dot) << " " << imag(dot)  << std::endl ;
+  }
+  src  = evec[0]+evec[1]+evec[2];
+  mass += -0.1;
+}
+
+  Grid_finalize();
+}
Author	SHA1	Message	Date
chulwoo1	73e27a16aa	Merge `570b72a47b` into `b50fb34e71`	2025-02-07 15:36:21 -05:00
Peter Boyle	b50fb34e71	Perf on Aurora	2025-02-01 18:39:34 +00:00
Peter Boyle	de84d730ff	Fastest run config on Aurora to date	2025-02-01 18:08:40 +00:00
Peter Boyle	c74d11e3d7	PVdagM MG	2025-02-01 11:04:13 -05:00
Peter Boyle	c4fc972fec	Merge branch 'feature/deprecate-uvm' into develop	2025-01-31 16:32:36 +00:00
Chulwoo Jung	570b72a47b	Bugfix. Sorry!	2025-01-21 15:37:39 -05:00
Chulwoo Jung	a5798a89ed	Merge branch 'develop' into specflow	2025-01-21 12:13:24 -05:00
Peter Boyle	3f3661a86f	Heading towards PVdagM multigrid	2025-01-17 14:33:35 +00:00
Chulwoo Jung	f7e2f9a401	Checking in spectral flow and DWF/Mobius kernel eigenvalue measurement	2025-01-16 20:47:33 +00:00
Chulwoo Jung	2848a9b558	DWF Kernel lanczos working(?)	2025-01-16 01:29:56 +00:00
Peter Boyle	5a4f9bf2e3	Force the ROCM version	2024-10-29 18:12:31 -04:00
Peter Boyle	f617468e04	Update Lattice_base.h	2024-10-11 10:39:16 -04:00
Peter Boyle	ee4046fe92	Added a dimension ordered column sum based reduction for scalar. Removes dependence on MPI_Allreduce and allows for work around on systems where this is bollox.	2024-09-27 09:26:03 -04:00
Peter Boyle	2a9cfeb9ea	New files	2024-09-26 14:23:29 -04:00
Peter Boyle	1147b8ea40	Cheby poly setup	2024-09-26 14:20:32 -04:00
Peter Boyle	3f9119b39d	Remove vectors used for the power spectrum table in paper	2024-09-26 14:19:41 -04:00
Peter Boyle	35e8225abd	Verbose control	2024-09-26 14:18:35 -04:00
Peter Boyle	bdbfbb7a14	Merge branch 'develop' of https://github.com/paboyle/Grid into develop	2024-09-26 14:05:45 -04:00
Peter Boyle	f7d4be8d96	Calculate bytes correctly	2024-09-26 14:04:44 -04:00