Best results on Aurora so far

Fixed boosted free field test

.gitignore

@@ -1,3 +1,7 @@
+# Doxygen stuff
+html/*
+latex/*
+
 # Compiled Object files #
 #########################
 *.slo

BLAS_benchmark/compile-command

@@ -0,0 +1,2 @@
+
+mpicxx -qmkl=parallel -fsycl BatchBlasBench.cc -o BatchBlasBench -DGRID_SYCL

BLAS_benchmark/compile-command-frontier

@@ -0,0 +1,5 @@
+CXX=hipcc
+MPICXX=mpicxx 
+CXXFLAGS="-fPIC -I{$ROCM_PATH}/include/ -I${MPICH_DIR}/include -L/lib64 -I/opt/cray/pe/mpich/8.1.28/ofi/gnu/12.3/include -DGRID_HIP"
+LDFLAGS="-L/lib64 -L${MPICH_DIR}/lib -lmpi -L${CRAY_MPICH_ROOTDIR}/gtl/lib -lmpi_gtl_hsa -lamdhip64 -lhipblas -lrocblas -lmpi_gnu_123"
+hipcc $CXXFLAGS $LDFLAGS BatchBlasBench.cc -o BatchBlasBench

BLAS_benchmark/compile-command-sunspot

@@ -0,0 +1,2 @@
+
+mpicxx -qmkl=parallel -fsycl BatchBlasBench.cc -o BatchBlasBench -DGRID_SYCL

Grid/GridCore.h

@@ -59,6 +59,7 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 #include <Grid/lattice/Lattice.h>      
 #include <Grid/cshift/Cshift.h>       
 #include <Grid/stencil/Stencil.h>      
+#include <Grid/stencil/GeneralLocalStencil.h>      
 #include <Grid/parallelIO/BinaryIO.h>
 #include <Grid/algorithms/Algorithms.h>   
 NAMESPACE_CHECK(GridCore)

Grid/Grid_Eigen_Dense.h

@@ -34,7 +34,7 @@
 #pragma push_macro("__SYCL_DEVICE_ONLY__")
 #undef __SYCL_DEVICE_ONLY__
 #define EIGEN_DONT_VECTORIZE
-//#undef EIGEN_USE_SYCL
+#undef EIGEN_USE_SYCL
 #define __SYCL__REDEFINE__
 #endif
 

Grid/Namespace.h

@@ -30,9 +30,14 @@ directory
 
 #include <type_traits>
 #include <cassert>
+#include <exception>
 
 #define NAMESPACE_BEGIN(A) namespace A {
 #define NAMESPACE_END(A)   }
 #define GRID_NAMESPACE_BEGIN NAMESPACE_BEGIN(Grid)
 #define GRID_NAMESPACE_END   NAMESPACE_END(Grid)
 #define NAMESPACE_CHECK(x) struct namespaceTEST##x {};  static_assert(std::is_same<namespaceTEST##x, ::namespaceTEST##x>::value,"Not in :: at"  ); 
+
+#define EXCEPTION_CHECK_BEGIN(A) try {
+#define EXCEPTION_CHECK_END(A)   } catch ( std::exception e ) { BACKTRACEFP(stderr); std::cerr << __PRETTY_FUNCTION__ << " : " <<__LINE__<< " Caught exception "<<e.what()<<std::endl; throw; }
+

Grid/algorithms/Algorithms.h

@@ -29,6 +29,9 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #ifndef GRID_ALGORITHMS_H
 #define GRID_ALGORITHMS_H
 
+NAMESPACE_CHECK(blas);
+#include <Grid/algorithms/blas/BatchedBlas.h>
+
 NAMESPACE_CHECK(algorithms);
 #include <Grid/algorithms/SparseMatrix.h>
 #include <Grid/algorithms/LinearOperator.h>
@@ -44,7 +47,11 @@ NAMESPACE_CHECK(SparseMatrix);
 #include <Grid/algorithms/approx/RemezGeneral.h>
 #include <Grid/algorithms/approx/ZMobius.h>
 NAMESPACE_CHECK(approx);
-#include <Grid/algorithms/iterative/Deflation.h>
+#include <Grid/algorithms/deflation/Deflation.h>
+#include <Grid/algorithms/deflation/MultiRHSBlockProject.h>
+#include <Grid/algorithms/deflation/MultiRHSDeflation.h>
+#include <Grid/algorithms/deflation/MultiRHSBlockCGLinalg.h>
+NAMESPACE_CHECK(deflation);
 #include <Grid/algorithms/iterative/ConjugateGradient.h>
 NAMESPACE_CHECK(ConjGrad);
 #include <Grid/algorithms/iterative/BiCGSTAB.h>
@@ -67,11 +74,11 @@ NAMESPACE_CHECK(BiCGSTAB);
 #include <Grid/algorithms/iterative/MixedPrecisionFlexibleGeneralisedMinimalResidual.h>
 #include <Grid/algorithms/iterative/ImplicitlyRestartedLanczos.h>
 #include <Grid/algorithms/iterative/PowerMethod.h>
-
+#include <Grid/algorithms/iterative/AdefGeneric.h>
+#include <Grid/algorithms/iterative/AdefMrhs.h>
 NAMESPACE_CHECK(PowerMethod);
 #include <Grid/algorithms/multigrid/MultiGrid.h>
-
-NAMESPACE_CHECK(CoarsendMatrix);
+NAMESPACE_CHECK(multigrid);
 #include <Grid/algorithms/FFT.h>
 
 #endif

Grid/algorithms/FFT.h

@@ -29,7 +29,7 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #define _GRID_FFT_H_
 
 #ifdef HAVE_FFTW
-#ifdef USE_MKL
+#if defined(USE_MKL) || defined(GRID_SYCL)
 #include <fftw/fftw3.h>
 #else
 #include <fftw3.h>
@@ -168,6 +168,7 @@ public:
   template<class vobj>
   void FFT_dim(Lattice<vobj> &result,const Lattice<vobj> &source,int dim, int sign){
 #ifndef HAVE_FFTW
+    std::cerr << "FFTW is not compiled but is called"<<std::endl;
     assert(0);
 #else
     conformable(result.Grid(),vgrid);
@@ -190,6 +191,7 @@ public:
       
     Lattice<sobj> pgbuf(&pencil_g);
     autoView(pgbuf_v , pgbuf, CpuWrite);
+    std::cout << "CPU view" << std::endl;
     
     typedef typename FFTW<scalar>::FFTW_scalar FFTW_scalar;
     typedef typename FFTW<scalar>::FFTW_plan   FFTW_plan;
@@ -213,6 +215,7 @@ public:
     else if ( sign == forward ) div = 1.0;
     else assert(0);
       
+    std::cout << GridLogPerformance<<"Making FFTW plan" << std::endl;
     FFTW_plan p;
     {
       FFTW_scalar *in = (FFTW_scalar *)&pgbuf_v[0];
@@ -226,6 +229,7 @@ public:
     }
       
     // Barrel shift and collect global pencil
+    std::cout << GridLogPerformance<<"Making pencil" << std::endl;
     Coordinate lcoor(Nd), gcoor(Nd);
     result = source;
     int pc = processor_coor[dim];
@@ -247,6 +251,7 @@ public:
       }
     }
       
+    std::cout <<GridLogPerformance<< "Looping orthog" << std::endl;
     // Loop over orthog coords
     int NN=pencil_g.lSites();
     GridStopWatch timer;
@@ -269,6 +274,7 @@ public:
     usec += timer.useconds();
     flops+= flops_call*NN;
       
+    std::cout <<GridLogPerformance<< "Writing back results " << std::endl;
     // writing out result
     {
       autoView(pgbuf_v,pgbuf,CpuRead);
@@ -285,6 +291,7 @@ public:
     }
     result = result*div;
       
+    std::cout <<GridLogPerformance<< "Destroying plan " << std::endl;
     // destroying plan
     FFTW<scalar>::fftw_destroy_plan(p);
 #endif

Grid/algorithms/LinearOperator.h

@@ -103,6 +103,38 @@ public:
     _Mat.MdagM(in,out);
   }
 };
+template<class Matrix,class Field>
+class MMdagLinearOperator : public LinearOperatorBase<Field> {
+  Matrix &_Mat;
+public:
+  MMdagLinearOperator(Matrix &Mat): _Mat(Mat){};
+
+  // Support for coarsening to a multigrid
+  void OpDiag (const Field &in, Field &out) {
+    _Mat.Mdiag(in,out);
+  }
+  void OpDir  (const Field &in, Field &out,int dir,int disp) {
+    _Mat.Mdir(in,out,dir,disp);
+  }
+  void OpDirAll  (const Field &in, std::vector<Field> &out){
+    _Mat.MdirAll(in,out);
+  };
+  void Op     (const Field &in, Field &out){
+    _Mat.M(in,out);
+  }
+  void AdjOp     (const Field &in, Field &out){
+    _Mat.Mdag(in,out);
+  }
+  void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){
+    _Mat.MMdag(in,out);
+    ComplexD dot = innerProduct(in,out);
+    n1=real(dot);
+    n2=norm2(out);
+  }
+  void HermOp(const Field &in, Field &out){
+    _Mat.MMdag(in,out);
+  }
+};
 
 ////////////////////////////////////////////////////////////////////
 // Construct herm op and shift it for mgrid smoother
@@ -145,6 +177,44 @@ public:
   }
 };
 
+////////////////////////////////////////////////////////////////////
+// Create a shifted HermOp
+////////////////////////////////////////////////////////////////////
+template<class Field>
+class ShiftedHermOpLinearOperator : public LinearOperatorBase<Field> {
+  LinearOperatorBase<Field> &_Mat;
+  RealD _shift;
+public:
+  ShiftedHermOpLinearOperator(LinearOperatorBase<Field> &Mat,RealD shift): _Mat(Mat), _shift(shift){};
+  // Support for coarsening to a multigrid
+  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){
+    HermOp(in,out);
+  }
+  void AdjOp     (const Field &in, Field &out){
+    HermOp(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){
+    _Mat.HermOp(in,out);
+    out = out + _shift*in;
+  }
+};
+
+
 ////////////////////////////////////////////////////////////////////
 // Wrap an already herm matrix
 ////////////////////////////////////////////////////////////////////

Grid/algorithms/SparseMatrix.h

@@ -45,6 +45,11 @@ public:
     M(in,tmp);
     Mdag(tmp,out);
   }
+  virtual void  MMdag(const Field &in, Field &out) {
+    Field tmp (in.Grid());
+    Mdag(in,tmp);
+    M(tmp,out);
+  }
   virtual  void Mdiag    (const Field &in, Field &out)=0;
   virtual  void Mdir     (const Field &in, Field &out,int dir, int disp)=0;
   virtual  void MdirAll  (const Field &in, std::vector<Field> &out)=0;

Grid/algorithms/approx/Chebyshev.h

@@ -59,7 +59,7 @@ public:
     RealD diff = hi-lo;
     RealD delta = diff*1.0e-9;
     for (RealD x=lo; x<hi; x+=delta) {
-      delta*=1.1;
+      delta*=1.02;
       RealD f = approx(x);
       out<< x<<" "<<f<<std::endl;
     }
@@ -131,6 +131,26 @@ public:
       Coeffs[j] = s * 2.0/order;
     }
   };
+  template<class functor>
+  void Init(RealD _lo,RealD _hi,int _order, functor & func)
+  {
+    lo=_lo;
+    hi=_hi;
+    order=_order;
+      
+    if(order < 2) exit(-1);
+    Coeffs.resize(order);
+    for(int j=0;j<order;j++){
+      RealD s=0;
+      for(int k=0;k<order;k++){
+	RealD y=std::cos(M_PI*(k+0.5)/order);
+	RealD x=0.5*(y*(hi-lo)+(hi+lo));
+	RealD f=func(x);
+	s=s+f*std::cos( j*M_PI*(k+0.5)/order );
+      }
+      Coeffs[j] = s * 2.0/order;
+    }
+  };
 
     
   void JacksonSmooth(void){

Grid/algorithms/approx/MultiShiftFunction.h

@@ -40,7 +40,7 @@ public:
   RealD norm;
   RealD lo,hi;
 
-  MultiShiftFunction(int n,RealD _lo,RealD _hi): poles(n), residues(n), lo(_lo), hi(_hi) {;};
+  MultiShiftFunction(int n,RealD _lo,RealD _hi): poles(n), residues(n), tolerances(n), lo(_lo), hi(_hi) {;};
   RealD approx(RealD x);
   void csv(std::ostream &out);
   void gnuplot(std::ostream &out);

Grid/algorithms/approx/Zolotarev.cc

@@ -293,7 +293,7 @@ static void sncndnFK(INTERNAL_PRECISION u, INTERNAL_PRECISION k,
  * Set type = 0 for the Zolotarev approximation, which is zero at x = 0, and
  * type = 1 for the approximation which is infinite at x = 0. */
 
-zolotarev_data* zolotarev(PRECISION epsilon, int n, int type) {
+zolotarev_data* zolotarev(ZOLO_PRECISION epsilon, int n, int type) {
   INTERNAL_PRECISION A, c, cp, kp, ksq, sn, cn, dn, Kp, Kj, z, z0, t, M, F,
     l, invlambda, xi, xisq, *tv, s, opl;
   int m, czero, ts;
@@ -375,12 +375,12 @@ zolotarev_data* zolotarev(PRECISION epsilon, int n, int type) {
   construct_partfrac(d);
   construct_contfrac(d);
 
-  /* Converting everything to PRECISION for external use only */
+  /* Converting everything to ZOLO_PRECISION for external use only */
 
   zd = (zolotarev_data*) malloc(sizeof(zolotarev_data));
-  zd -> A = (PRECISION) d -> A;
-  zd -> Delta = (PRECISION) d -> Delta;
-  zd -> epsilon = (PRECISION) d -> epsilon;
+  zd -> A = (ZOLO_PRECISION) d -> A;
+  zd -> Delta = (ZOLO_PRECISION) d -> Delta;
+  zd -> epsilon = (ZOLO_PRECISION) d -> epsilon;
   zd -> n = d -> n;
   zd -> type = d -> type;
   zd -> dn = d -> dn;
@@ -390,24 +390,24 @@ zolotarev_data* zolotarev(PRECISION epsilon, int n, int type) {
   zd -> deg_num = d -> deg_num;
   zd -> deg_denom = d -> deg_denom;
 
-  zd -> a = (PRECISION*) malloc(zd -> dn * sizeof(PRECISION));
-  for (m = 0; m < zd -> dn; m++) zd -> a[m] = (PRECISION) d -> a[m];
+  zd -> a = (ZOLO_PRECISION*) malloc(zd -> dn * sizeof(ZOLO_PRECISION));
+  for (m = 0; m < zd -> dn; m++) zd -> a[m] = (ZOLO_PRECISION) d -> a[m];
   free(d -> a);
 
-  zd -> ap = (PRECISION*) malloc(zd -> dd * sizeof(PRECISION));
-  for (m = 0; m < zd -> dd; m++) zd -> ap[m] = (PRECISION) d -> ap[m];
+  zd -> ap = (ZOLO_PRECISION*) malloc(zd -> dd * sizeof(ZOLO_PRECISION));
+  for (m = 0; m < zd -> dd; m++) zd -> ap[m] = (ZOLO_PRECISION) d -> ap[m];
   free(d -> ap);
 
-  zd -> alpha = (PRECISION*) malloc(zd -> da * sizeof(PRECISION));
-  for (m = 0; m < zd -> da; m++) zd -> alpha[m] = (PRECISION) d -> alpha[m];
+  zd -> alpha = (ZOLO_PRECISION*) malloc(zd -> da * sizeof(ZOLO_PRECISION));
+  for (m = 0; m < zd -> da; m++) zd -> alpha[m] = (ZOLO_PRECISION) d -> alpha[m];
   free(d -> alpha);
 
-  zd -> beta = (PRECISION*) malloc(zd -> db * sizeof(PRECISION));
-  for (m = 0; m < zd -> db; m++) zd -> beta[m] = (PRECISION) d -> beta[m];
+  zd -> beta = (ZOLO_PRECISION*) malloc(zd -> db * sizeof(ZOLO_PRECISION));
+  for (m = 0; m < zd -> db; m++) zd -> beta[m] = (ZOLO_PRECISION) d -> beta[m];
   free(d -> beta);
 
-  zd -> gamma = (PRECISION*) malloc(zd -> n * sizeof(PRECISION));
-  for (m = 0; m < zd -> n; m++) zd -> gamma[m] = (PRECISION) d -> gamma[m];
+  zd -> gamma = (ZOLO_PRECISION*) malloc(zd -> n * sizeof(ZOLO_PRECISION));
+  for (m = 0; m < zd -> n; m++) zd -> gamma[m] = (ZOLO_PRECISION) d -> gamma[m];
   free(d -> gamma);
 
   free(d);
@@ -426,7 +426,7 @@ void zolotarev_free(zolotarev_data *zdata)
 }
 
 
-zolotarev_data* higham(PRECISION epsilon, int n) {
+zolotarev_data* higham(ZOLO_PRECISION epsilon, int n) {
   INTERNAL_PRECISION A, M, c, cp, z, z0, t, epssq;
   int m, czero;
   zolotarev_data *zd;
@@ -481,9 +481,9 @@ zolotarev_data* higham(PRECISION epsilon, int n) {
   /* Converting everything to PRECISION for external use only */
 
   zd = (zolotarev_data*) malloc(sizeof(zolotarev_data));
-  zd -> A = (PRECISION) d -> A;
-  zd -> Delta = (PRECISION) d -> Delta;
-  zd -> epsilon = (PRECISION) d -> epsilon;
+  zd -> A = (ZOLO_PRECISION) d -> A;
+  zd -> Delta = (ZOLO_PRECISION) d -> Delta;
+  zd -> epsilon = (ZOLO_PRECISION) d -> epsilon;
   zd -> n = d -> n;
   zd -> type = d -> type;
   zd -> dn = d -> dn;
@@ -493,24 +493,24 @@ zolotarev_data* higham(PRECISION epsilon, int n) {
   zd -> deg_num = d -> deg_num;
   zd -> deg_denom = d -> deg_denom;
 
-  zd -> a = (PRECISION*) malloc(zd -> dn * sizeof(PRECISION));
-  for (m = 0; m < zd -> dn; m++) zd -> a[m] = (PRECISION) d -> a[m];
+  zd -> a = (ZOLO_PRECISION*) malloc(zd -> dn * sizeof(ZOLO_PRECISION));
+  for (m = 0; m < zd -> dn; m++) zd -> a[m] = (ZOLO_PRECISION) d -> a[m];
   free(d -> a);
 
-  zd -> ap = (PRECISION*) malloc(zd -> dd * sizeof(PRECISION));
-  for (m = 0; m < zd -> dd; m++) zd -> ap[m] = (PRECISION) d -> ap[m];
+  zd -> ap = (ZOLO_PRECISION*) malloc(zd -> dd * sizeof(ZOLO_PRECISION));
+  for (m = 0; m < zd -> dd; m++) zd -> ap[m] = (ZOLO_PRECISION) d -> ap[m];
   free(d -> ap);
 
-  zd -> alpha = (PRECISION*) malloc(zd -> da * sizeof(PRECISION));
-  for (m = 0; m < zd -> da; m++) zd -> alpha[m] = (PRECISION) d -> alpha[m];
+  zd -> alpha = (ZOLO_PRECISION*) malloc(zd -> da * sizeof(ZOLO_PRECISION));
+  for (m = 0; m < zd -> da; m++) zd -> alpha[m] = (ZOLO_PRECISION) d -> alpha[m];
   free(d -> alpha);
 
-  zd -> beta = (PRECISION*) malloc(zd -> db * sizeof(PRECISION));
-  for (m = 0; m < zd -> db; m++) zd -> beta[m] = (PRECISION) d -> beta[m];
+  zd -> beta = (ZOLO_PRECISION*) malloc(zd -> db * sizeof(ZOLO_PRECISION));
+  for (m = 0; m < zd -> db; m++) zd -> beta[m] = (ZOLO_PRECISION) d -> beta[m];
   free(d -> beta);
 
-  zd -> gamma = (PRECISION*) malloc(zd -> n * sizeof(PRECISION));
-  for (m = 0; m < zd -> n; m++) zd -> gamma[m] = (PRECISION) d -> gamma[m];
+  zd -> gamma = (ZOLO_PRECISION*) malloc(zd -> n * sizeof(ZOLO_PRECISION));
+  for (m = 0; m < zd -> n; m++) zd -> gamma[m] = (ZOLO_PRECISION) d -> gamma[m];
   free(d -> gamma);
 
   free(d);
@@ -523,17 +523,17 @@ NAMESPACE_END(Grid);
 #ifdef TEST
 
 #undef ZERO
-#define ZERO ((PRECISION) 0)
+#define ZERO ((ZOLO_PRECISION) 0)
 #undef ONE
-#define ONE ((PRECISION) 1)
+#define ONE ((ZOLO_PRECISION) 1)
 #undef TWO
-#define TWO ((PRECISION) 2)
+#define TWO ((ZOLO_PRECISION) 2)
 
 /* Evaluate the rational approximation R(x) using the factored form */
 
-static PRECISION zolotarev_eval(PRECISION x, zolotarev_data* rdata) {
+static ZOLO_PRECISION zolotarev_eval(ZOLO_PRECISION x, zolotarev_data* rdata) {
   int m;
-  PRECISION R;
+  ZOLO_PRECISION R;
 
   if (rdata -> type == 0) {
     R = rdata -> A * x;
@@ -551,9 +551,9 @@ static PRECISION zolotarev_eval(PRECISION x, zolotarev_data* rdata) {
 
 /* Evaluate the rational approximation R(x) using the partial fraction form */
 
-static PRECISION zolotarev_partfrac_eval(PRECISION x, zolotarev_data* rdata) {
+static ZOLO_PRECISION zolotarev_partfrac_eval(ZOLO_PRECISION x, zolotarev_data* rdata) {
   int m;
-  PRECISION R = rdata -> alpha[rdata -> da - 1];
+  ZOLO_PRECISION R = rdata -> alpha[rdata -> da - 1];
   for (m = 0; m < rdata -> dd; m++)
     R += rdata -> alpha[m] / (x * x - rdata -> ap[m]);
   if (rdata -> type == 1) R += rdata -> alpha[rdata -> dd] / (x * x);
@@ -568,18 +568,18 @@ static PRECISION zolotarev_partfrac_eval(PRECISION x, zolotarev_data* rdata) {
  * non-signalling overflow this will work correctly since 1/(1/0) = 1/INF = 0,
  * but with signalling overflow you will get an error message. */
 
-static PRECISION zolotarev_contfrac_eval(PRECISION x, zolotarev_data* rdata) {
+static ZOLO_PRECISION zolotarev_contfrac_eval(ZOLO_PRECISION x, zolotarev_data* rdata) {
   int m;
-  PRECISION R = rdata -> beta[0] * x;
+  ZOLO_PRECISION R = rdata -> beta[0] * x;
   for (m = 1; m < rdata -> db; m++) R = rdata -> beta[m] * x + ONE / R;
   return R;
 }    
 
 /* Evaluate the rational approximation R(x) using Cayley form */
 
-static PRECISION zolotarev_cayley_eval(PRECISION x, zolotarev_data* rdata) {
+static ZOLO_PRECISION zolotarev_cayley_eval(ZOLO_PRECISION x, zolotarev_data* rdata) {
   int m;
-  PRECISION T;
+  ZOLO_PRECISION T;
 
   T = rdata -> type == 0 ? ONE : -ONE;
   for (m = 0; m < rdata -> n; m++)
@@ -607,7 +607,7 @@ int main(int argc, char** argv) {
   int m, n, plotpts = 5000, type = 0;
   float eps, x, ypferr, ycferr, ycaylerr, maxypferr, maxycferr, maxycaylerr;
   zolotarev_data *rdata;
-  PRECISION y;
+  ZOLO_PRECISION y;
   FILE *plot_function, *plot_error, 
     *plot_partfrac, *plot_contfrac, *plot_cayley;
 
@@ -626,13 +626,13 @@ int main(int argc, char** argv) {
   }
 
   rdata = type == 2 
-    ? higham((PRECISION) eps, n) 
-    : zolotarev((PRECISION) eps, n, type);
+    ? higham((ZOLO_PRECISION) eps, n) 
+    : zolotarev((ZOLO_PRECISION) eps, n, type);
 
   printf("Zolotarev Test: R(epsilon = %g, n = %d, type = %d)\n\t" 
 	 STRINGIFY(VERSION) "\n\t" STRINGIFY(HVERSION)
 	 "\n\tINTERNAL_PRECISION = " STRINGIFY(INTERNAL_PRECISION)
-	 "\tPRECISION = " STRINGIFY(PRECISION)
+	 "\tZOLO_PRECISION = " STRINGIFY(ZOLO_PRECISION)
 	 "\n\n\tRational approximation of degree (%d,%d), %s at x = 0\n"
 	 "\tDelta = %g (maximum error)\n\n"
 	 "\tA = %g (overall factor)\n",
@@ -681,15 +681,15 @@ int main(int argc, char** argv) {
     x = 2.4 * (float) m / plotpts - 1.2;
     if (rdata -> type == 0 || fabs(x) * (float) plotpts > 1.0) {
       /* skip x = 0 for type 1, as R(0) is singular */
-      y = zolotarev_eval((PRECISION) x, rdata);
+      y = zolotarev_eval((ZOLO_PRECISION) x, rdata);
       fprintf(plot_function, "%g %g\n", x, (float) y);
       fprintf(plot_error, "%g %g\n",
 	      x, (float)((y - ((x > 0.0 ? ONE : -ONE))) / rdata -> Delta));
-      ypferr = (float)((zolotarev_partfrac_eval((PRECISION) x, rdata) - y)
+      ypferr = (float)((zolotarev_partfrac_eval((ZOLO_PRECISION) x, rdata) - y)
 		       / rdata -> Delta);
-      ycferr = (float)((zolotarev_contfrac_eval((PRECISION) x, rdata) - y)
+      ycferr = (float)((zolotarev_contfrac_eval((ZOLO_PRECISION) x, rdata) - y)
 		       / rdata -> Delta);
-      ycaylerr = (float)((zolotarev_cayley_eval((PRECISION) x, rdata) - y)
+      ycaylerr = (float)((zolotarev_cayley_eval((ZOLO_PRECISION) x, rdata) - y)
 		       / rdata -> Delta);
       if (fabs(x) < 1.0 && fabs(x) > rdata -> epsilon) {
 	maxypferr = MAX(maxypferr, fabs(ypferr));

Grid/algorithms/approx/Zolotarev.h

@@ -9,10 +9,10 @@ NAMESPACE_BEGIN(Approx);
 #define HVERSION Header Time-stamp: <14-OCT-2004 09:26:51.00 adk@MISSCONTRARY>
 
 #ifndef ZOLOTAREV_INTERNAL
-#ifndef PRECISION
-#define PRECISION double
+#ifndef ZOLO_PRECISION
+#define ZOLO_PRECISION double
 #endif
-#define ZPRECISION PRECISION
+#define ZPRECISION ZOLO_PRECISION
 #define ZOLOTAREV_DATA zolotarev_data
 #endif
 
@@ -77,8 +77,8 @@ typedef struct {
  * zolotarev_data structure. The arguments must satisfy the constraints that
  * epsilon > 0, n > 0, and type = 0 or 1. */
 
-ZOLOTAREV_DATA* higham(PRECISION epsilon, int n) ;
-ZOLOTAREV_DATA* zolotarev(PRECISION epsilon, int n, int type);
+ZOLOTAREV_DATA* higham(ZOLO_PRECISION epsilon, int n) ;
+ZOLOTAREV_DATA* zolotarev(ZOLO_PRECISION epsilon, int n, int type);
 void zolotarev_free(zolotarev_data *zdata);
 #endif
 
@@ -86,3 +86,4 @@ void zolotarev_free(zolotarev_data *zdata);
 NAMESPACE_END(Approx);
 NAMESPACE_END(Grid);
 #endif
+

Grid/algorithms/blas/BatchedBlas.cc

@@ -0,0 +1,34 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: BatchedBlas.h
+
+    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/GridCore.h>
+#include <Grid/algorithms/blas/BatchedBlas.h>
+NAMESPACE_BEGIN(Grid);
+gridblasHandle_t GridBLAS::gridblasHandle;
+int              GridBLAS::gridblasInit;
+NAMESPACE_END(Grid);
+

Grid/algorithms/blas/BatchedBlas.h

@@ -0,0 +1,895 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: BatchedBlas.h
+
+    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 */
+#pragma once
+
+#ifdef GRID_HIP
+#include <hipblas/hipblas.h>
+#endif
+#ifdef GRID_CUDA
+#include <cublas_v2.h>
+#endif
+#ifdef GRID_SYCL
+#include <oneapi/mkl.hpp>
+#endif
+#if 0
+#define GRID_ONE_MKL
+#endif
+#ifdef GRID_ONE_MKL
+#include <oneapi/mkl.hpp>
+#endif
+
+///////////////////////////////////////////////////////////////////////	  
+// Need to rearrange lattice data to be in the right format for a
+// batched multiply. Might as well make these static, dense packed
+///////////////////////////////////////////////////////////////////////
+NAMESPACE_BEGIN(Grid);
+#ifdef GRID_HIP
+  typedef hipblasHandle_t gridblasHandle_t;
+#endif
+#ifdef GRID_CUDA
+  typedef cublasHandle_t gridblasHandle_t;
+#endif
+#ifdef GRID_SYCL
+  typedef sycl::queue *gridblasHandle_t;
+#endif
+#ifdef GRID_ONE_MKL
+  typedef sycl::queue *gridblasHandle_t;
+#endif
+#if !defined(GRID_SYCL) && !defined(GRID_CUDA) && !defined(GRID_HIP) && !defined(GRID_ONE_MKL)
+  typedef int32_t gridblasHandle_t;
+#endif
+
+enum GridBLASOperation_t { GridBLAS_OP_N, GridBLAS_OP_T, GridBLAS_OP_C } ;
+
+class GridBLAS {
+public:
+
+  
+  static gridblasHandle_t gridblasHandle;
+  static int            gridblasInit;
+  
+  static void Init(void)
+  {
+    if ( ! gridblasInit ) {
+#ifdef GRID_CUDA
+      std::cout << "cublasCreate"<<std::endl;
+      cublasCreate(&gridblasHandle);
+      cublasSetPointerMode(gridblasHandle, CUBLAS_POINTER_MODE_DEVICE);
+#endif
+#ifdef GRID_HIP
+      std::cout << "hipblasCreate"<<std::endl;
+      hipblasCreate(&gridblasHandle);
+#endif
+#ifdef GRID_SYCL
+      gridblasHandle = theGridAccelerator;
+#endif
+#ifdef GRID_ONE_MKL
+      sycl::gpu_selector selector;
+      sycl::device selectedDevice { selector };
+      sycl::property_list q_prop{sycl::property::queue::in_order()};
+      gridblasHandle =new sycl::queue (selectedDevice,q_prop);
+#endif
+      gridblasInit=1;
+    }
+  }
+  
+  // Force construct once
+  GridBLAS() { Init(); };
+  ~GridBLAS() { };
+  
+  /////////////////////////////////////////////////////////////////////////////////////
+  // BLAS GEMM conventions:
+  /////////////////////////////////////////////////////////////////////////////////////
+  // - C = alpha A * B + beta C
+  // Dimensions:
+  // - C_m.n
+  // - A_m.k
+  // - B_k.n
+  // - Flops = 8 M N K
+  // - Bytes = 2*sizeof(word) * (MN+MK+KN)
+  // M=60, N=12
+  // Flop/Byte = 8 . 60.60.12 / (60.12+60.60+60.12)/16 = 4 so expect about 4 TF/s on a GCD
+  /////////////////////////////////////////////////////////////////////////////////////
+  void synchronise(void)
+  {
+#ifdef GRID_HIP
+    auto err = hipDeviceSynchronize();
+    assert(err==hipSuccess);
+#endif
+#ifdef GRID_CUDA
+    auto err = cudaDeviceSynchronize();
+    assert(err==cudaSuccess);
+#endif
+#ifdef GRID_SYCL
+    accelerator_barrier();
+#endif
+#ifdef GRID_ONE_MKL
+    gridblasHandle->wait();
+#endif
+  }
+  
+  void gemmBatched(int m,int n, int k,
+		   ComplexD alpha,
+		   deviceVector<ComplexD*> &Amk,  // pointer list to matrices
+		   deviceVector<ComplexD*> &Bkn,
+		   ComplexD beta,
+		   deviceVector<ComplexD*> &Cmn)
+  {
+    gemmBatched(GridBLAS_OP_N,GridBLAS_OP_N,
+		m,n,k,
+		alpha,
+		Amk,
+		Bkn,
+		beta,
+		Cmn);
+  }
+  void gemmBatched(int m,int n, int k,
+		   ComplexF alpha,
+		   deviceVector<ComplexF*> &Amk,  // pointer list to matrices
+		   deviceVector<ComplexF*> &Bkn,
+		   ComplexF beta,
+		   deviceVector<ComplexF*> &Cmn)
+  {
+    gemmBatched(GridBLAS_OP_N,GridBLAS_OP_N,
+		m,n,k,
+		alpha,
+		Amk,
+		Bkn,
+		beta,
+		Cmn);
+  }
+  void gemmBatched(int m,int n, int k,
+		   RealD alpha,
+		   deviceVector<RealD*> &Amk,  // pointer list to matrices
+		   deviceVector<RealD*> &Bkn,
+		   RealD beta,
+		   deviceVector<RealD*> &Cmn)
+  {
+    gemmBatched(GridBLAS_OP_N,GridBLAS_OP_N,
+		m,n,k,
+		alpha,
+		Amk,
+		Bkn,
+		beta,
+		Cmn);
+  }
+  void gemmBatched(int m,int n, int k,
+		   RealF alpha,
+		   deviceVector<RealF*> &Amk,  // pointer list to matrices
+		   deviceVector<RealF*> &Bkn,
+		   RealF beta,
+		   deviceVector<RealF*> &Cmn)
+  {
+    gemmBatched(GridBLAS_OP_N,GridBLAS_OP_N,
+		m,n,k,
+		alpha,
+		Amk,
+		Bkn,
+		beta,
+		Cmn);
+  }
+
+  void gemmBatched(GridBLASOperation_t OpA,
+		   GridBLASOperation_t OpB,
+		   int m,int n, int k,
+		   ComplexD alpha,
+		   deviceVector<ComplexD*> &Amk,  // pointer list to matrices
+		   deviceVector<ComplexD*> &Bkn,
+		   ComplexD beta,
+		   deviceVector<ComplexD*> &Cmn)
+  {
+    RealD t2=usecond();
+    int32_t batchCount = Amk.size();
+    assert(Bkn.size()==batchCount);
+    assert(Cmn.size()==batchCount);
+
+    assert(OpA!=GridBLAS_OP_T); // Complex case expect no transpose
+    assert(OpB!=GridBLAS_OP_T);
+
+    int lda = m; // m x k column major
+    int ldb = k; // k x n column major
+    int ldc = m; // m x b column major
+    if(OpA!=GridBLAS_OP_N)
+      lda = k;
+    if(OpB!=GridBLAS_OP_N)
+      ldb = n;
+    
+    static deviceVector<ComplexD> alpha_p(1);
+    static deviceVector<ComplexD> beta_p(1);
+    // can prestore the 1 and the zero on device
+    acceleratorCopyToDevice((void *)&alpha,(void *)&alpha_p[0],sizeof(ComplexD));
+    acceleratorCopyToDevice((void *)&beta ,(void *)&beta_p[0],sizeof(ComplexD));
+    RealD t0=usecond();
+    //    std::cout << "ZgemmBatched mnk  "<<m<<","<<n<<","<<k<<" count "<<batchCount<<std::endl;
+#ifdef GRID_HIP
+    hipblasOperation_t hOpA;
+    hipblasOperation_t hOpB;
+    if ( OpA == GridBLAS_OP_N ) hOpA = HIPBLAS_OP_N;
+    if ( OpA == GridBLAS_OP_T ) hOpA = HIPBLAS_OP_T;
+    if ( OpA == GridBLAS_OP_C ) hOpA = HIPBLAS_OP_C;
+    if ( OpB == GridBLAS_OP_N ) hOpB = HIPBLAS_OP_N;
+    if ( OpB == GridBLAS_OP_T ) hOpB = HIPBLAS_OP_T;
+    if ( OpB == GridBLAS_OP_C ) hOpB = HIPBLAS_OP_C;
+    auto err = hipblasZgemmBatched(gridblasHandle,
+				   hOpA,
+				   hOpB,
+				   m,n,k,
+				   (hipblasDoubleComplex *) &alpha_p[0],
+				   (hipblasDoubleComplex **)&Amk[0], lda,
+				   (hipblasDoubleComplex **)&Bkn[0], ldb,
+				   (hipblasDoubleComplex *) &beta_p[0],
+				   (hipblasDoubleComplex **)&Cmn[0], ldc,
+				   batchCount);
+    //	 std::cout << " hipblas return code " <<(int)err<<std::endl;
+    assert(err==HIPBLAS_STATUS_SUCCESS);
+#endif
+#ifdef GRID_CUDA
+    cublasOperation_t hOpA;
+    cublasOperation_t hOpB;
+    if ( OpA == GridBLAS_OP_N ) hOpA = CUBLAS_OP_N;
+    if ( OpA == GridBLAS_OP_T ) hOpA = CUBLAS_OP_T;
+    if ( OpA == GridBLAS_OP_C ) hOpA = CUBLAS_OP_C;
+    if ( OpB == GridBLAS_OP_N ) hOpB = CUBLAS_OP_N;
+    if ( OpB == GridBLAS_OP_T ) hOpB = CUBLAS_OP_T;
+    if ( OpB == GridBLAS_OP_C ) hOpB = CUBLAS_OP_C;
+    auto err = cublasZgemmBatched(gridblasHandle,
+				  hOpA,
+				  hOpB,
+				  m,n,k,
+				  (cuDoubleComplex *) &alpha_p[0],
+				  (cuDoubleComplex **)&Amk[0], lda,
+				  (cuDoubleComplex **)&Bkn[0], ldb,
+				  (cuDoubleComplex *) &beta_p[0],
+				  (cuDoubleComplex **)&Cmn[0], ldc,
+				  batchCount);
+    assert(err==CUBLAS_STATUS_SUCCESS);
+#endif
+#ifdef GRID_SYCL
+      int64_t m64=m;
+      int64_t n64=n;
+      int64_t k64=k;
+      int64_t lda64=lda;
+      int64_t ldb64=ldb;
+      int64_t ldc64=ldc;
+      int64_t batchCount64=batchCount;
+
+      oneapi::mkl::transpose iOpA;
+      oneapi::mkl::transpose iOpB;
+      
+      if ( OpA == GridBLAS_OP_N ) iOpA = oneapi::mkl::transpose::N;
+      if ( OpA == GridBLAS_OP_T ) iOpA = oneapi::mkl::transpose::T;
+      if ( OpA == GridBLAS_OP_C ) iOpA = oneapi::mkl::transpose::C;
+      if ( OpB == GridBLAS_OP_N ) iOpB = oneapi::mkl::transpose::N;
+      if ( OpB == GridBLAS_OP_T ) iOpB = oneapi::mkl::transpose::T;
+      if ( OpB == GridBLAS_OP_C ) iOpB = oneapi::mkl::transpose::C;
+
+      oneapi::mkl::blas::column_major::gemm_batch(*gridblasHandle,
+						  &iOpA,
+						  &iOpB,
+						  &m64,&n64,&k64,
+						  (ComplexD *) &alpha_p[0],
+						  (const ComplexD **)&Amk[0], (const int64_t *)&lda64,
+						  (const ComplexD **)&Bkn[0], (const int64_t *)&ldb64,
+						  (ComplexD *) &beta_p[0],
+						  (ComplexD **)&Cmn[0], (const int64_t *)&ldc64,
+						  (int64_t)1,&batchCount64,std::vector<sycl::event>());
+      synchronise();
+#if 0
+      // This code was used to check the mat mul on Sunspot/OneMKL
+      std::cerr << " Called SYCL batched ZGEMM OpA "<< OpA << " OpB "<<OpB <<std::endl;
+      std::vector<ComplexD> A(m*k);  // pointer list to matrices
+      std::vector<ComplexD> B(k*n);
+      std::vector<ComplexD> C(m*n);
+      //      int sda = lda*k;
+      //      int sdb = ldb*k;
+      //      int sdc = ldc*n;
+      std::cerr << " Checking the GEMM results "<<std::endl;
+      for (int p = 0; p < 1; ++p) {
+	ComplexD * Amk_p;  // pointer list to matrices
+	ComplexD * Bkn_p;  // pointer list to matrices
+	ComplexD * Cmn_p;  // pointer list to matrices
+	acceleratorCopyFromDevice((void *)&Amk[p],(void *)&Amk_p,sizeof(ComplexD*));
+	acceleratorCopyFromDevice((void *)&Bkn[p],(void *)&Bkn_p,sizeof(ComplexD*));
+	acceleratorCopyFromDevice((void *)&Cmn[p],(void *)&Cmn_p,sizeof(ComplexD*));
+	std::cerr << " p " << p << " copied pointers "<<std::endl;
+	acceleratorCopyFromDevice((void *)Amk_p,(void *)&A[0],m*k*sizeof(ComplexD));
+	acceleratorCopyFromDevice((void *)Bkn_p,(void *)&B[0],k*n*sizeof(ComplexD));
+	acceleratorCopyFromDevice((void *)Cmn_p,(void *)&C[0],m*n*sizeof(ComplexD));
+	std::cerr << " p " << p << " copied matrices "<<std::endl;
+	std::cerr << " C[0] "<<C[0]<<std::endl;
+	std::cerr << " A[0] "<<A[0]<<std::endl;
+	std::cerr << " B[0] "<<B[0]<<std::endl;
+	std::cerr << " m "<<m<<std::endl;
+	std::cerr << " n "<<n<<std::endl;
+	std::cerr << " k "<<k<<std::endl;
+	for (int mm = 0; mm < m; ++mm) {
+	  for (int nn = 0; nn < n; ++nn) {
+	    ComplexD c_mn(0.0);
+	    for (int kk = 0; kk < k; ++kk) {
+	      int idx_a, idx_b;
+	      //    int lda = m; // m x k column major
+	      //    int ldb = k; // k x n column major
+	      //    int ldc = m; // m x b column major
+	      if(OpA!=GridBLAS_OP_N) {
+		idx_a =kk + mm*lda;
+	      } else {
+		idx_a =mm + kk*lda;
+	      }
+	      if(OpB!=GridBLAS_OP_N) {
+		idx_b =nn + kk*ldb;
+	      } else {
+		idx_b =kk + nn*ldb;
+	      }
+	      //	      std::cerr << " idx_a "<<idx_a<<" idx_b "<<idx_b<<std::endl;
+
+	      ComplexD Ac = A[idx_a];
+	      ComplexD Bc = B[idx_b];
+	      if(OpA==GridBLAS_OP_C) Ac = conjugate(Ac);
+	      if(OpB==GridBLAS_OP_C) Bc = conjugate(Bc);
+	      
+	      c_mn += Ac*Bc;
+	    }
+	    std::cerr << " beta "<<beta<<" alpha "<<alpha<<" C_"<<mm<<","<<nn<<" "<<c_mn<<" "<<C[mm + nn*ldc]<<std::endl;
+	  }
+	}
+      }
+#endif
+#endif
+#if !defined(GRID_SYCL) && !defined(GRID_CUDA) && !defined(GRID_HIP)
+    // Need a default/reference implementation; use Eigen
+      if ( (OpA == GridBLAS_OP_N ) && (OpB == GridBLAS_OP_N) ) {
+	thread_for (p, batchCount, {
+	  Eigen::Map<Eigen::MatrixXcd> eAmk(Amk[p],m,k);
+	  Eigen::Map<Eigen::MatrixXcd> eBkn(Bkn[p],k,n);
+	  Eigen::Map<Eigen::MatrixXcd> eCmn(Cmn[p],m,n);
+	  eCmn = beta * eCmn + alpha * eAmk * eBkn ;
+        });
+      } else if ( (OpA == GridBLAS_OP_C ) && (OpB == GridBLAS_OP_N) ) {
+	thread_for (p, batchCount, {
+	  Eigen::Map<Eigen::MatrixXcd> eAmk(Amk[p],k,m);
+	  Eigen::Map<Eigen::MatrixXcd> eBkn(Bkn[p],k,n);
+	  Eigen::Map<Eigen::MatrixXcd> eCmn(Cmn[p],m,n);
+	  eCmn = beta * eCmn + alpha * eAmk.adjoint() * eBkn ;
+	  });
+      } else if ( (OpA == GridBLAS_OP_N ) && (OpB == GridBLAS_OP_C) ) {
+	thread_for (p, batchCount, {
+	  Eigen::Map<Eigen::MatrixXcd> eAmk(Amk[p],m,k);
+	  Eigen::Map<Eigen::MatrixXcd> eBkn(Bkn[p],n,k);
+	  Eigen::Map<Eigen::MatrixXcd> eCmn(Cmn[p],m,n);
+	  eCmn = beta * eCmn + alpha * eAmk * eBkn.adjoint() ;
+	  });
+      } else if ( (OpA == GridBLAS_OP_C ) && (OpB == GridBLAS_OP_C) ) {
+	thread_for (p, batchCount, {
+	  Eigen::Map<Eigen::MatrixXcd> eAmk(Amk[p],k,m);
+	  Eigen::Map<Eigen::MatrixXcd> eBkn(Bkn[p],n,k);
+	  Eigen::Map<Eigen::MatrixXcd> eCmn(Cmn[p],m,n);
+	  eCmn = beta * eCmn + alpha * eAmk.adjoint() * eBkn.adjoint() ;
+	  } );
+      } else { 
+	assert(0);
+      }
+#endif
+     RealD t1=usecond();
+     RealD flops = 8.0*m*n*k*batchCount;
+     RealD bytes = 1.0*sizeof(ComplexD)*(m*k+k*n+m*n)*batchCount;
+     //     std::cout <<GridLogMessage<< " batched Blas copy "<<(t0-t2)/1.e3 <<" ms "<<std::endl;
+     //     std::cout <<GridLogMessage<< " batched Blas zGemm call "<<m<<","<<n<<","<<k<<" "<< flops/(t1-t0)/1.e3 <<" GF/s "<<(t1-t0)/1.e3<<" ms "<<std::endl;
+     //     std::cout <<GridLogMessage<< " batched Blas zGemm call "<<m<<","<<n<<","<<k<<" "<< bytes/(t1-t0)/1.e3 <<" GB/s "<<(t1-t0)/1.e3<<" ms "<<std::endl;
+  }
+
+  void gemmBatched(GridBLASOperation_t OpA,
+		   GridBLASOperation_t OpB,
+		   int m,int n, int k,
+		   ComplexF alpha,
+		   deviceVector<ComplexF*> &Amk,  // pointer list to matrices
+		   deviceVector<ComplexF*> &Bkn,
+		   ComplexF beta,
+		   deviceVector<ComplexF*> &Cmn)
+  {
+    RealD t2=usecond();
+    int32_t batchCount = Amk.size();
+
+    assert(OpA!=GridBLAS_OP_T); // Complex case expect no transpose
+    assert(OpB!=GridBLAS_OP_T);
+
+    int lda = m; // m x k column major
+    int ldb = k; // k x n column major
+    int ldc = m; // m x b column major
+    if(OpA!=GridBLAS_OP_N)
+      lda = k;
+    if(OpB!=GridBLAS_OP_N)
+      ldb = n;
+    static deviceVector<ComplexF> alpha_p(1);
+    static deviceVector<ComplexF> beta_p(1);
+    // can prestore the 1 and the zero on device
+    acceleratorCopyToDevice((void *)&alpha,(void *)&alpha_p[0],sizeof(ComplexF));
+    acceleratorCopyToDevice((void *)&beta ,(void *)&beta_p[0],sizeof(ComplexF));
+    RealD t0=usecond();
+
+    assert(Bkn.size()==batchCount);
+    assert(Cmn.size()==batchCount);
+#ifdef GRID_HIP
+    hipblasOperation_t hOpA;
+    hipblasOperation_t hOpB;
+    if ( OpA == GridBLAS_OP_N ) hOpA = HIPBLAS_OP_N;
+    if ( OpA == GridBLAS_OP_T ) hOpA = HIPBLAS_OP_T;
+    if ( OpA == GridBLAS_OP_C ) hOpA = HIPBLAS_OP_C;
+    if ( OpB == GridBLAS_OP_N ) hOpB = HIPBLAS_OP_N;
+    if ( OpB == GridBLAS_OP_T ) hOpB = HIPBLAS_OP_T;
+    if ( OpB == GridBLAS_OP_C ) hOpB = HIPBLAS_OP_C;
+    auto err = hipblasCgemmBatched(gridblasHandle,
+				   hOpA,
+				   hOpB,
+				   m,n,k,
+				   (hipblasComplex *) &alpha_p[0],
+				   (hipblasComplex **)&Amk[0], lda,
+				   (hipblasComplex **)&Bkn[0], ldb,
+				   (hipblasComplex *) &beta_p[0],
+				   (hipblasComplex **)&Cmn[0], ldc,
+				   batchCount);
+
+    assert(err==HIPBLAS_STATUS_SUCCESS);
+#endif
+#ifdef GRID_CUDA
+    cublasOperation_t hOpA;
+    cublasOperation_t hOpB;
+    if ( OpA == GridBLAS_OP_N ) hOpA = CUBLAS_OP_N;
+    if ( OpA == GridBLAS_OP_T ) hOpA = CUBLAS_OP_T;
+    if ( OpA == GridBLAS_OP_C ) hOpA = CUBLAS_OP_C;
+    if ( OpB == GridBLAS_OP_N ) hOpB = CUBLAS_OP_N;
+    if ( OpB == GridBLAS_OP_T ) hOpB = CUBLAS_OP_T;
+    if ( OpB == GridBLAS_OP_C ) hOpB = CUBLAS_OP_C;
+    auto err = cublasCgemmBatched(gridblasHandle,
+				  hOpA,
+				  hOpB,
+				  m,n,k,
+				  (cuComplex *) &alpha_p[0],
+				  (cuComplex **)&Amk[0], lda,
+				  (cuComplex **)&Bkn[0], ldb,
+				  (cuComplex *) &beta_p[0],
+				  (cuComplex **)&Cmn[0], ldc,
+				  batchCount);
+    assert(err==CUBLAS_STATUS_SUCCESS);
+#endif
+#ifdef GRID_SYCL
+      int64_t m64=m;
+      int64_t n64=n;
+      int64_t k64=k;
+      int64_t lda64=lda;
+      int64_t ldb64=ldb;
+      int64_t ldc64=ldc;
+      int64_t batchCount64=batchCount;
+
+      oneapi::mkl::transpose iOpA;
+      oneapi::mkl::transpose iOpB;
+      
+      if ( OpA == GridBLAS_OP_N ) iOpA = oneapi::mkl::transpose::N;
+      if ( OpA == GridBLAS_OP_T ) iOpA = oneapi::mkl::transpose::T;
+      if ( OpA == GridBLAS_OP_C ) iOpA = oneapi::mkl::transpose::C;
+      if ( OpB == GridBLAS_OP_N ) iOpB = oneapi::mkl::transpose::N;
+      if ( OpB == GridBLAS_OP_T ) iOpB = oneapi::mkl::transpose::T;
+      if ( OpB == GridBLAS_OP_C ) iOpB = oneapi::mkl::transpose::C;
+
+      oneapi::mkl::blas::column_major::gemm_batch(*gridblasHandle,
+						  &iOpA,
+						  &iOpB,
+						  &m64,&n64,&k64,
+						  (ComplexF *) &alpha_p[0],
+						  (const ComplexF **)&Amk[0], (const int64_t *)&lda64,
+						  (const ComplexF **)&Bkn[0], (const int64_t *)&ldb64,
+						  (ComplexF *) &beta_p[0],
+						  (ComplexF **)&Cmn[0], (const int64_t *)&ldc64,
+						  (int64_t)1,&batchCount64,std::vector<sycl::event>());
+    synchronise();
+#endif
+#if !defined(GRID_SYCL) && !defined(GRID_CUDA) && !defined(GRID_HIP)
+    // Need a default/reference implementation; use Eigen
+      if ( (OpA == GridBLAS_OP_N ) && (OpB == GridBLAS_OP_N) ) {
+	thread_for (p, batchCount, {
+	  Eigen::Map<Eigen::MatrixXcf> eAmk(Amk[p],m,k);
+	  Eigen::Map<Eigen::MatrixXcf> eBkn(Bkn[p],k,n);
+	  Eigen::Map<Eigen::MatrixXcf> eCmn(Cmn[p],m,n);
+	  eCmn = beta * eCmn + alpha * eAmk * eBkn ;
+	  });
+      } else if ( (OpA == GridBLAS_OP_C ) && (OpB == GridBLAS_OP_N) ) {
+	thread_for (p, batchCount, {
+	  Eigen::Map<Eigen::MatrixXcf> eAmk(Amk[p],k,m);
+	  Eigen::Map<Eigen::MatrixXcf> eBkn(Bkn[p],k,n);
+	  Eigen::Map<Eigen::MatrixXcf> eCmn(Cmn[p],m,n);
+	  eCmn = beta * eCmn + alpha * eAmk.adjoint() * eBkn ;
+	  });
+      } else if ( (OpA == GridBLAS_OP_N ) && (OpB == GridBLAS_OP_C) ) {
+	thread_for (p, batchCount, {
+	  Eigen::Map<Eigen::MatrixXcf> eAmk(Amk[p],m,k);
+	  Eigen::Map<Eigen::MatrixXcf> eBkn(Bkn[p],n,k);
+	  Eigen::Map<Eigen::MatrixXcf> eCmn(Cmn[p],m,n);
+	  eCmn = beta * eCmn + alpha * eAmk * eBkn.adjoint() ;
+	  });
+      } else if ( (OpA == GridBLAS_OP_C ) && (OpB == GridBLAS_OP_C) ) {
+	thread_for (p, batchCount, {
+	  Eigen::Map<Eigen::MatrixXcf> eAmk(Amk[p],k,m);
+	  Eigen::Map<Eigen::MatrixXcf> eBkn(Bkn[p],n,k);
+	  Eigen::Map<Eigen::MatrixXcf> eCmn(Cmn[p],m,n);
+	  eCmn = beta * eCmn + alpha * eAmk.adjoint() * eBkn.adjoint() ;
+	  } );
+      } else { 
+	assert(0);
+      }
+#endif
+     RealD t1=usecond();
+     RealD flops = 8.0*m*n*k*batchCount;
+     RealD bytes = 1.0*sizeof(ComplexF)*(m*k+k*n+m*n)*batchCount;
+  }
+  
+  ///////////////////////////////////////////////////////////////////////////
+  // Single precision real GEMM
+  ///////////////////////////////////////////////////////////////////////////
+
+  void gemmBatched(GridBLASOperation_t OpA,
+		   GridBLASOperation_t OpB,
+		   int m,int n, int k,
+		   RealF alpha,
+		   deviceVector<RealF*> &Amk,  // pointer list to matrices
+		   deviceVector<RealF*> &Bkn,
+		   RealF beta,
+		   deviceVector<RealF*> &Cmn)
+  {
+    RealD t2=usecond();
+    int32_t batchCount = Amk.size();
+
+    assert(OpA!=GridBLAS_OP_C); // Real case no conjugate
+    assert(OpB!=GridBLAS_OP_C);
+
+    int lda = m; // m x k column major
+    int ldb = k; // k x n column major
+    int ldc = m; // m x b column major
+    if(OpA!=GridBLAS_OP_N)
+      lda = k;
+    if(OpB!=GridBLAS_OP_N)
+      ldb = n;
+    static deviceVector<RealF> alpha_p(1);
+    static deviceVector<RealF> beta_p(1);
+    // can prestore the 1 and the zero on device
+    acceleratorCopyToDevice((void *)&alpha,(void *)&alpha_p[0],sizeof(RealF));
+    acceleratorCopyToDevice((void *)&beta ,(void *)&beta_p[0],sizeof(RealF));
+    RealD t0=usecond();
+
+    assert(Bkn.size()==batchCount);
+    assert(Cmn.size()==batchCount);
+#ifdef GRID_HIP
+    hipblasOperation_t hOpA;
+    hipblasOperation_t hOpB;
+    if ( OpA == GridBLAS_OP_N ) hOpA = HIPBLAS_OP_N;
+    if ( OpA == GridBLAS_OP_T ) hOpA = HIPBLAS_OP_T;
+    if ( OpA == GridBLAS_OP_C ) hOpA = HIPBLAS_OP_C;
+    if ( OpB == GridBLAS_OP_N ) hOpB = HIPBLAS_OP_N;
+    if ( OpB == GridBLAS_OP_T ) hOpB = HIPBLAS_OP_T;
+    if ( OpB == GridBLAS_OP_C ) hOpB = HIPBLAS_OP_C;
+    auto err = hipblasSgemmBatched(gridblasHandle,
+				   hOpA,
+				   hOpB,
+				   m,n,k,
+				   (float *) &alpha_p[0],
+				   (float **)&Amk[0], lda,
+				   (float **)&Bkn[0], ldb,
+				   (float *) &beta_p[0],
+				   (float **)&Cmn[0], ldc,
+				   batchCount);
+    assert(err==HIPBLAS_STATUS_SUCCESS);
+#endif
+#ifdef GRID_CUDA
+    cublasOperation_t hOpA;
+    cublasOperation_t hOpB;
+    if ( OpA == GridBLAS_OP_N ) hOpA = CUBLAS_OP_N;
+    if ( OpA == GridBLAS_OP_T ) hOpA = CUBLAS_OP_T;
+    if ( OpA == GridBLAS_OP_C ) hOpA = CUBLAS_OP_C;
+    if ( OpB == GridBLAS_OP_N ) hOpB = CUBLAS_OP_N;
+    if ( OpB == GridBLAS_OP_T ) hOpB = CUBLAS_OP_T;
+    if ( OpB == GridBLAS_OP_C ) hOpB = CUBLAS_OP_C;
+    auto err = cublasSgemmBatched(gridblasHandle,
+				  hOpA,
+				  hOpB,
+				  m,n,k,
+				  (float *) &alpha_p[0],
+				  (float **)&Amk[0], lda,
+				  (float **)&Bkn[0], ldb,
+				  (float *) &beta_p[0],
+				  (float **)&Cmn[0], ldc,
+				  batchCount);
+    assert(err==CUBLAS_STATUS_SUCCESS);
+#endif
+#ifdef GRID_SYCL
+      int64_t m64=m;
+      int64_t n64=n;
+      int64_t k64=k;
+      int64_t lda64=lda;
+      int64_t ldb64=ldb;
+      int64_t ldc64=ldc;
+      int64_t batchCount64=batchCount;
+
+      oneapi::mkl::transpose iOpA;
+      oneapi::mkl::transpose iOpB;
+      
+      if ( OpA == GridBLAS_OP_N ) iOpA = oneapi::mkl::transpose::N;
+      if ( OpA == GridBLAS_OP_T ) iOpA = oneapi::mkl::transpose::T;
+      if ( OpA == GridBLAS_OP_C ) iOpA = oneapi::mkl::transpose::C;
+      if ( OpB == GridBLAS_OP_N ) iOpB = oneapi::mkl::transpose::N;
+      if ( OpB == GridBLAS_OP_T ) iOpB = oneapi::mkl::transpose::T;
+      if ( OpB == GridBLAS_OP_C ) iOpB = oneapi::mkl::transpose::C;
+
+      oneapi::mkl::blas::column_major::gemm_batch(*gridblasHandle,
+						  &iOpA,
+						  &iOpB,
+						  &m64,&n64,&k64,
+						  (float *) &alpha_p[0],
+						  (const float **)&Amk[0], (const int64_t *)&lda64,
+						  (const float **)&Bkn[0], (const int64_t *)&ldb64,
+						  (float *) &beta_p[0],
+						  (float **)&Cmn[0], (const int64_t *)&ldc64,
+						  (int64_t)1,&batchCount64,std::vector<sycl::event>());
+      synchronise();
+#endif
+#if !defined(GRID_SYCL) && !defined(GRID_CUDA) && !defined(GRID_HIP)
+    // Need a default/reference implementation; use Eigen
+      if ( (OpA == GridBLAS_OP_N ) && (OpB == GridBLAS_OP_N) ) {
+	thread_for (p, batchCount, {
+	  Eigen::Map<Eigen::MatrixXf> eAmk(Amk[p],m,k);
+	  Eigen::Map<Eigen::MatrixXf> eBkn(Bkn[p],k,n);
+	  Eigen::Map<Eigen::MatrixXf> eCmn(Cmn[p],m,n);
+	  eCmn = beta * eCmn + alpha * eAmk * eBkn ;
+	  });
+      } else if ( (OpA == GridBLAS_OP_T ) && (OpB == GridBLAS_OP_N) ) {
+	thread_for (p, batchCount, {
+	  Eigen::Map<Eigen::MatrixXf> eAmk(Amk[p],k,m);
+	  Eigen::Map<Eigen::MatrixXf> eBkn(Bkn[p],k,n);
+	  Eigen::Map<Eigen::MatrixXf> eCmn(Cmn[p],m,n);
+	  eCmn = beta * eCmn + alpha * eAmk.transpose() * eBkn ;
+	  });
+      } else if ( (OpA == GridBLAS_OP_N ) && (OpB == GridBLAS_OP_T) ) {
+	thread_for (p, batchCount, {
+	  Eigen::Map<Eigen::MatrixXf> eAmk(Amk[p],m,k);
+	  Eigen::Map<Eigen::MatrixXf> eBkn(Bkn[p],n,k);
+	  Eigen::Map<Eigen::MatrixXf> eCmn(Cmn[p],m,n);
+	  eCmn = beta * eCmn + alpha * eAmk * eBkn.transpose() ;
+	  });
+      } else if ( (OpA == GridBLAS_OP_T ) && (OpB == GridBLAS_OP_T) ) {
+	thread_for (p, batchCount, {
+	  Eigen::Map<Eigen::MatrixXf> eAmk(Amk[p],k,m);
+	  Eigen::Map<Eigen::MatrixXf> eBkn(Bkn[p],n,k);
+	  Eigen::Map<Eigen::MatrixXf> eCmn(Cmn[p],m,n);
+	  eCmn = beta * eCmn + alpha * eAmk.transpose() * eBkn.transpose() ;
+	  } );
+      } else { 
+	assert(0);
+      }
+#endif
+     RealD t1=usecond();
+     RealD flops = 2.0*m*n*k*batchCount;
+     RealD bytes = 1.0*sizeof(RealF)*(m*k+k*n+m*n)*batchCount;
+  }
+  
+  
+  ///////////////////////////////////////////////////////////////////////////
+  // Double precision real GEMM
+  ///////////////////////////////////////////////////////////////////////////
+  void gemmBatched(GridBLASOperation_t OpA,
+		   GridBLASOperation_t OpB,
+		   int m,int n, int k,
+		   RealD alpha,
+		   deviceVector<RealD*> &Amk,  // pointer list to matrices
+		   deviceVector<RealD*> &Bkn,
+		   RealD beta,
+		   deviceVector<RealD*> &Cmn)
+  {
+    RealD t2=usecond();
+    int32_t batchCount = Amk.size();
+
+    assert(OpA!=GridBLAS_OP_C); // Real case no conjugate
+    assert(OpB!=GridBLAS_OP_C);
+
+    int lda = m; // m x k column major
+    int ldb = k; // k x n column major
+    int ldc = m; // m x b column major
+    if(OpA!=GridBLAS_OP_N)
+      lda = k;
+    if(OpB!=GridBLAS_OP_N)
+      ldb = n;
+    
+    static deviceVector<RealD> alpha_p(1);
+    static deviceVector<RealD> beta_p(1);
+    // can prestore the 1 and the zero on device
+    acceleratorCopyToDevice((void *)&alpha,(void *)&alpha_p[0],sizeof(RealD));
+    acceleratorCopyToDevice((void *)&beta ,(void *)&beta_p[0],sizeof(RealD));
+    RealD t0=usecond();
+
+    assert(Bkn.size()==batchCount);
+    assert(Cmn.size()==batchCount);
+#ifdef GRID_HIP
+    hipblasOperation_t hOpA;
+    hipblasOperation_t hOpB;
+    if ( OpA == GridBLAS_OP_N ) hOpA = HIPBLAS_OP_N;
+    if ( OpA == GridBLAS_OP_T ) hOpA = HIPBLAS_OP_T;
+    if ( OpA == GridBLAS_OP_C ) hOpA = HIPBLAS_OP_C;
+    if ( OpB == GridBLAS_OP_N ) hOpB = HIPBLAS_OP_N;
+    if ( OpB == GridBLAS_OP_T ) hOpB = HIPBLAS_OP_T;
+    if ( OpB == GridBLAS_OP_C ) hOpB = HIPBLAS_OP_C;
+    auto err = hipblasDgemmBatched(gridblasHandle,
+				   HIPBLAS_OP_N,
+				   HIPBLAS_OP_N,
+				   m,n,k,
+				   (double *) &alpha_p[0],
+				   (double **)&Amk[0], lda,
+				   (double **)&Bkn[0], ldb,
+				   (double *) &beta_p[0],
+				   (double **)&Cmn[0], ldc,
+				   batchCount);
+    assert(err==HIPBLAS_STATUS_SUCCESS);
+#endif
+#ifdef GRID_CUDA
+    cublasOperation_t hOpA;
+    cublasOperation_t hOpB;
+    if ( OpA == GridBLAS_OP_N ) hOpA = CUBLAS_OP_N;
+    if ( OpA == GridBLAS_OP_T ) hOpA = CUBLAS_OP_T;
+    if ( OpA == GridBLAS_OP_C ) hOpA = CUBLAS_OP_C;
+    if ( OpB == GridBLAS_OP_N ) hOpB = CUBLAS_OP_N;
+    if ( OpB == GridBLAS_OP_T ) hOpB = CUBLAS_OP_T;
+    if ( OpB == GridBLAS_OP_C ) hOpB = CUBLAS_OP_C;
+    auto err = cublasDgemmBatched(gridblasHandle,
+				  hOpA,
+				  hOpB,
+				  m,n,k,
+				  (double *) &alpha_p[0],
+				  (double **)&Amk[0], lda,
+				  (double **)&Bkn[0], ldb,
+				  (double *) &beta_p[0],
+				  (double **)&Cmn[0], ldc,
+				  batchCount);
+    assert(err==CUBLAS_STATUS_SUCCESS);
+#endif
+#ifdef GRID_SYCL
+      int64_t m64=m;
+      int64_t n64=n;
+      int64_t k64=k;
+      int64_t lda64=lda;
+      int64_t ldb64=ldb;
+      int64_t ldc64=ldc;
+      int64_t batchCount64=batchCount;
+
+      oneapi::mkl::transpose iOpA;
+      oneapi::mkl::transpose iOpB;
+      
+      if ( OpA == GridBLAS_OP_N ) iOpA = oneapi::mkl::transpose::N;
+      if ( OpA == GridBLAS_OP_T ) iOpA = oneapi::mkl::transpose::T;
+      if ( OpA == GridBLAS_OP_C ) iOpA = oneapi::mkl::transpose::C;
+      if ( OpB == GridBLAS_OP_N ) iOpB = oneapi::mkl::transpose::N;
+      if ( OpB == GridBLAS_OP_T ) iOpB = oneapi::mkl::transpose::T;
+      if ( OpB == GridBLAS_OP_C ) iOpB = oneapi::mkl::transpose::C;
+
+      oneapi::mkl::blas::column_major::gemm_batch(*gridblasHandle,
+						  &iOpA,
+						  &iOpB,
+						  &m64,&n64,&k64,
+						  (double *) &alpha_p[0],
+						  (const double **)&Amk[0], (const int64_t *)&lda64,
+						  (const double **)&Bkn[0], (const int64_t *)&ldb64,
+						  (double *) &beta_p[0],
+						  (double **)&Cmn[0], (const int64_t *)&ldc64,
+						  (int64_t)1,&batchCount64,std::vector<sycl::event>());
+      synchronise();
+#endif
+#if !defined(GRID_SYCL) && !defined(GRID_CUDA) && !defined(GRID_HIP)
+    // Need a default/reference implementation; use Eigen
+      if ( (OpA == GridBLAS_OP_N ) && (OpB == GridBLAS_OP_N) ) {
+	thread_for (p, batchCount, {
+	  Eigen::Map<Eigen::MatrixXd> eAmk(Amk[p],m,k);
+	  Eigen::Map<Eigen::MatrixXd> eBkn(Bkn[p],k,n);
+	  Eigen::Map<Eigen::MatrixXd> eCmn(Cmn[p],m,n);
+	  eCmn = beta * eCmn + alpha * eAmk * eBkn ;
+	  });
+      } else if ( (OpA == GridBLAS_OP_T ) && (OpB == GridBLAS_OP_N) ) {
+	thread_for (p, batchCount, {
+	  Eigen::Map<Eigen::MatrixXd> eAmk(Amk[p],k,m);
+	  Eigen::Map<Eigen::MatrixXd> eBkn(Bkn[p],k,n);
+	  Eigen::Map<Eigen::MatrixXd> eCmn(Cmn[p],m,n);
+	  eCmn = beta * eCmn + alpha * eAmk.transpose() * eBkn ;
+	  });
+      } else if ( (OpA == GridBLAS_OP_N ) && (OpB == GridBLAS_OP_T) ) {
+	thread_for (p, batchCount, {
+	  Eigen::Map<Eigen::MatrixXd> eAmk(Amk[p],m,k);
+	  Eigen::Map<Eigen::MatrixXd> eBkn(Bkn[p],n,k);
+	  Eigen::Map<Eigen::MatrixXd> eCmn(Cmn[p],m,n);
+	  eCmn = beta * eCmn + alpha * eAmk * eBkn.transpose() ;
+	  });
+      } else if ( (OpA == GridBLAS_OP_T ) && (OpB == GridBLAS_OP_T) ) {
+	thread_for (p, batchCount, {
+	  Eigen::Map<Eigen::MatrixXd> eAmk(Amk[p],k,m);
+	  Eigen::Map<Eigen::MatrixXd> eBkn(Bkn[p],n,k);
+	  Eigen::Map<Eigen::MatrixXd> eCmn(Cmn[p],m,n);
+	  eCmn = beta * eCmn + alpha * eAmk.transpose() * eBkn.transpose() ;
+	  });
+      } else { 
+	assert(0);
+      }
+#endif
+     RealD t1=usecond();
+     RealD flops = 2.0*m*n*k*batchCount;
+     RealD bytes = 1.0*sizeof(RealD)*(m*k+k*n+m*n)*batchCount;
+  }
+
+  template<class CComplex>
+  double benchmark(int M, int N, int K, int BATCH)
+  {
+    int32_t N_A = M*K*BATCH;
+    int32_t N_B = K*N*BATCH;
+    int32_t N_C = M*N*BATCH;
+    deviceVector<CComplex> A(N_A); acceleratorMemSet(&A[0],0,N_A*sizeof(CComplex));
+    deviceVector<CComplex> B(N_B); acceleratorMemSet(&B[0],0,N_B*sizeof(CComplex));
+    deviceVector<CComplex> C(N_C); acceleratorMemSet(&C[0],0,N_C*sizeof(CComplex));
+    CComplex alpha(1.0);
+    CComplex beta (1.0);
+    RealD flops = 8.0*M*N*K*BATCH;
+    int ncall=1000;
+    deviceVector<CComplex *> As(BATCH);
+    deviceVector<CComplex *> Bs(BATCH);
+    deviceVector<CComplex *> Cs(BATCH);
+    for(int b = 0 ; b < BATCH;b++) {
+      CComplex *ptr;
+      ptr = &A[b*M*K];      acceleratorPut(As[b],ptr);
+      ptr = &B[b*K*N];      acceleratorPut(Bs[b],ptr);
+      ptr = &C[b*M*N];      acceleratorPut(Cs[b],ptr);
+    }
+
+    // Warm up call
+    gemmBatched(M,N,K,
+		alpha,
+		As, // m x k 
+		Bs, // k x n
+		beta, 
+		Cs);
+    synchronise();
+
+    RealD t0 = usecond();
+    for(int i=0;i<ncall;i++){
+      gemmBatched(M,N,K,
+		  alpha,
+		  As, // m x k 
+		  Bs, // k x n
+		  beta, 
+		  Cs);
+      synchronise();
+    }
+    RealD t1 = usecond();
+    RealD bytes = 1.0*sizeof(CComplex)*(M*N*2+N*K+M*K)*BATCH;
+    flops = 8.0*M*N*K*BATCH*ncall;
+    flops = flops/(t1-t0)/1.e3;
+    return flops; // Returns gigaflops
+  }
+
+};
+
+NAMESPACE_END(Grid);

Grid/algorithms/deflation/MultiRHSBlockCGLinalg.h

@@ -0,0 +1,376 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: MultiRHSBlockCGLinalg.h
+
+    Copyright (C) 2024
+
+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 */
+#pragma once
+
+NAMESPACE_BEGIN(Grid);
+
+
+/* Need helper object for BLAS accelerated mrhs blockCG */
+template<class Field>
+class MultiRHSBlockCGLinalg
+{
+public:
+
+  typedef typename Field::scalar_type   scalar;
+  typedef typename Field::scalar_object scalar_object;
+  typedef typename Field::vector_object vector_object;
+
+  deviceVector<scalar> BLAS_X;      // nrhs x vol -- the sources
+  deviceVector<scalar> BLAS_Y;      // nrhs x vol -- the result
+  deviceVector<scalar> BLAS_C;      // nrhs x nrhs -- the coefficients 
+  deviceVector<scalar> BLAS_Cred;   // nrhs x nrhs x oSites -- reduction buffer
+  deviceVector<scalar *> Xdip;
+  deviceVector<scalar *> Ydip;
+  deviceVector<scalar *> Cdip;
+  
+  MultiRHSBlockCGLinalg() {};
+  ~MultiRHSBlockCGLinalg(){ Deallocate(); };
+  
+  void Deallocate(void)
+  {
+    Xdip.resize(0);
+    Ydip.resize(0);
+    Cdip.resize(0);
+    BLAS_Cred.resize(0);
+    BLAS_C.resize(0);
+    BLAS_X.resize(0);
+    BLAS_Y.resize(0);
+  }
+  void MaddMatrix(std::vector<Field> &AP, Eigen::MatrixXcd &m , const std::vector<Field> &X,const std::vector<Field> &Y,RealD scale=1.0)
+  {
+    std::vector<Field> Y_copy(AP.size(),AP[0].Grid());
+    for(int r=0;r<AP.size();r++){
+      Y_copy[r] = Y[r];
+    }
+    MulMatrix(AP,m,X);
+    for(int r=0;r<AP.size();r++){
+      AP[r] = scale*AP[r]+Y_copy[r];
+    }
+  }
+  void MulMatrix(std::vector<Field> &Y, Eigen::MatrixXcd &m , const std::vector<Field> &X)
+  {
+    typedef typename Field::scalar_type scomplex;
+    GridBase *grid;
+    uint64_t vol;
+    uint64_t words;
+
+    int nrhs = Y.size();
+    grid  = X[0].Grid();
+    vol   = grid->lSites();
+    words = sizeof(scalar_object)/sizeof(scalar);
+    int64_t vw = vol * words;
+
+    RealD t0 = usecond();
+    BLAS_X.resize(nrhs * vw); // cost free if size doesn't change
+    BLAS_Y.resize(nrhs * vw); // cost free if size doesn't change
+    BLAS_C.resize(nrhs * nrhs);// cost free if size doesn't change
+    RealD t1 = usecond();
+
+    /////////////////////////////////////////////
+    // Copy in the multi-rhs sources
+    /////////////////////////////////////////////
+    for(int r=0;r<nrhs;r++){
+      int64_t offset = r*vw;
+      autoView(x_v,X[r],AcceleratorRead);
+      acceleratorCopyDeviceToDevice(&x_v[0],&BLAS_X[offset],sizeof(scalar_object)*vol);
+    }
+
+    // Assumes Eigen storage contiguous
+    acceleratorCopyToDevice(&m(0,0),&BLAS_C[0],BLAS_C.size()*sizeof(scalar));
+    
+  /*
+   * in Fortran column major notation (cuBlas order)
+   *
+   * Xxr = [X1(x)][..][Xn(x)]
+   * Yxr = [Y1(x)][..][Ym(x)]
+   * Y = X . C
+   */
+    deviceVector<scalar *> Xd(1);
+    deviceVector<scalar *> Yd(1);
+    deviceVector<scalar *> Cd(1);
+
+    scalar * Xh = & BLAS_X[0];
+    scalar * Yh = & BLAS_Y[0];
+    scalar * Ch = & BLAS_C[0];
+
+    acceleratorPut(Xd[0],Xh);
+    acceleratorPut(Yd[0],Yh);
+    acceleratorPut(Cd[0],Ch);
+
+    RealD t2 = usecond();
+    GridBLAS BLAS;
+    /////////////////////////////////////////
+    // Y = X*C (transpose?)
+    /////////////////////////////////////////
+    BLAS.gemmBatched(GridBLAS_OP_N,GridBLAS_OP_N, 
+    		     vw,nrhs,nrhs,
+		     scalar(1.0),
+		     Xd,
+		     Cd,
+		     scalar(0.0),  // wipe out Y
+		     Yd);
+    BLAS.synchronise();
+    RealD t3 = usecond();
+
+    // Copy back Y = m X 
+    for(int r=0;r<nrhs;r++){
+      int64_t offset = r*vw;
+      autoView(y_v,Y[r],AcceleratorWrite);
+      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;
+  }
+  
+  void InnerProductMatrix(Eigen::MatrixXcd &m , const std::vector<Field> &X, const std::vector<Field> &Y)
+  {
+#if 0    
+    int nrhs;
+    GridBase *grid;
+    uint64_t vol;
+    uint64_t words;
+
+    nrhs = X.size();
+    assert(X.size()==Y.size());
+    conformable(X[0],Y[0]);
+
+    grid  = X[0].Grid();
+    vol   = grid->lSites();
+    words = sizeof(scalar_object)/sizeof(scalar);
+    int64_t vw = vol * words;
+
+    RealD t0 = usecond();
+    BLAS_X.resize(nrhs * vw); // cost free if size doesn't change
+    BLAS_Y.resize(nrhs * vw); // cost free if size doesn't change
+    BLAS_C.resize(nrhs * nrhs);// cost free if size doesn't change
+    RealD t1 = usecond();
+
+    /////////////////////////////////////////////
+    // Copy in the multi-rhs sources
+    /////////////////////////////////////////////
+    for(int r=0;r<nrhs;r++){
+      int64_t offset = r*vw;
+      autoView(x_v,X[r],AcceleratorRead);
+      acceleratorCopyDeviceToDevice(&x_v[0],&BLAS_X[offset],sizeof(scalar_object)*vol);
+      autoView(y_v,Y[r],AcceleratorRead);
+      acceleratorCopyDeviceToDevice(&y_v[0],&BLAS_Y[offset],sizeof(scalar_object)*vol);
+    }
+    RealD t2 = usecond();
+
+  /*
+   * in Fortran column major notation (cuBlas order)
+   *
+   * Xxr = [X1(x)][..][Xn(x)]
+   *
+   * Yxr = [Y1(x)][..][Ym(x)]
+   *
+   * C_rs = X^dag Y
+   */
+    deviceVector<scalar *> Xd(1);
+    deviceVector<scalar *> Yd(1);
+    deviceVector<scalar *> Cd(1);
+
+    scalar * Xh = & BLAS_X[0];
+    scalar * Yh = & BLAS_Y[0];
+    scalar * Ch = & BLAS_C[0];
+
+    acceleratorPut(Xd[0],Xh);
+    acceleratorPut(Yd[0],Yh);
+    acceleratorPut(Cd[0],Ch);
+
+    GridBLAS BLAS;
+
+    RealD t3 = usecond();
+    /////////////////////////////////////////
+    // C_rs = X^dag Y
+    /////////////////////////////////////////
+    BLAS.gemmBatched(GridBLAS_OP_C,GridBLAS_OP_N, 
+    		     nrhs,nrhs,vw,
+		     ComplexD(1.0),
+		     Xd,
+		     Yd,
+		     ComplexD(0.0),  // wipe out C
+		     Cd);
+    BLAS.synchronise();
+    RealD t4 = usecond();
+
+    std::vector<scalar> HOST_C(BLAS_C.size());      // nrhs . nrhs -- the coefficients 
+    acceleratorCopyFromDevice(&BLAS_C[0],&HOST_C[0],BLAS_C.size()*sizeof(scalar));
+    grid->GlobalSumVector(&HOST_C[0],nrhs*nrhs);
+
+    RealD t5 = usecond();
+    for(int rr=0;rr<nrhs;rr++){
+      for(int r=0;r<nrhs;r++){
+	int off = r+nrhs*rr;
+	m(r,rr)=HOST_C[off];
+      }
+    }
+    RealD t6 = usecond();
+    uint64_t M=nrhs;
+    uint64_t N=nrhs;
+    uint64_t K=vw;
+    RealD bytes = 1.0*sizeof(ComplexD)*(M*N*2+N*K+M*K);
+    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;
+#else
+    int nrhs;
+    GridBase *grid;
+    uint64_t vol;
+    uint64_t words;
+
+    nrhs = X.size();
+    assert(X.size()==Y.size());
+    conformable(X[0],Y[0]);
+
+    grid  = X[0].Grid();
+    int rd0 =  grid->_rdimensions[0] * grid->_rdimensions[1];
+    vol   = grid->oSites()/rd0;
+    words = rd0*sizeof(vector_object)/sizeof(scalar);
+    int64_t vw = vol * words;
+    assert(vw == grid->lSites()*sizeof(scalar_object)/sizeof(scalar));
+
+    RealD t0 = usecond();
+    BLAS_X.resize(nrhs * vw); // cost free if size doesn't change
+    BLAS_Y.resize(nrhs * vw); // cost free if size doesn't change
+    BLAS_Cred.resize(nrhs * nrhs * vol);// cost free if size doesn't change
+    RealD t1 = usecond();
+
+    /////////////////////////////////////////////
+    // Copy in the multi-rhs sources -- layout batched BLAS ready
+    /////////////////////////////////////////////
+    for(int r=0;r<nrhs;r++){
+      autoView(x_v,X[r],AcceleratorRead);
+      autoView(y_v,Y[r],AcceleratorRead);
+      scalar *from_x=(scalar *)&x_v[0];
+      scalar *from_y=(scalar *)&y_v[0];
+      scalar *BX = &BLAS_X[0];
+      scalar *BY = &BLAS_Y[0];
+      accelerator_for(ssw,vw,1,{
+	  uint64_t ss=ssw/words;
+	  uint64_t  w=ssw%words;
+	  uint64_t offset = w+r*words+ss*nrhs*words; // [ss][rhs][words]
+	  BX[offset] = from_x[ssw];
+	  BY[offset] = from_y[ssw];
+	});
+    }
+    RealD t2 = usecond();
+
+  /*
+   * in Fortran column major notation (cuBlas order)
+   *
+   * Xxr = [X1(x)][..][Xn(x)]
+   *
+   * Yxr = [Y1(x)][..][Ym(x)]
+   *
+   * C_rs = X^dag Y
+   */
+    Xdip.resize(vol);
+    Ydip.resize(vol);
+    Cdip.resize(vol);
+    std::vector<scalar *> Xh(vol);
+    std::vector<scalar *> Yh(vol);
+    std::vector<scalar *> Ch(vol);
+    for(uint64_t ss=0;ss<vol;ss++){
+
+      Xh[ss] = & BLAS_X[ss*nrhs*words];
+      Yh[ss] = & BLAS_Y[ss*nrhs*words];
+      Ch[ss] = & BLAS_Cred[ss*nrhs*nrhs];
+
+    }
+    acceleratorCopyToDevice(&Xh[0],&Xdip[0],vol*sizeof(scalar *));
+    acceleratorCopyToDevice(&Yh[0],&Ydip[0],vol*sizeof(scalar *));
+    acceleratorCopyToDevice(&Ch[0],&Cdip[0],vol*sizeof(scalar *));
+    
+    GridBLAS BLAS;
+
+    RealD t3 = usecond();
+    /////////////////////////////////////////
+    // C_rs = X^dag Y
+    /////////////////////////////////////////
+    BLAS.gemmBatched(GridBLAS_OP_C,GridBLAS_OP_N, 
+    		     nrhs,nrhs,words,
+		     ComplexD(1.0),
+		     Xdip,
+		     Ydip,
+		     ComplexD(0.0),  // wipe out C
+		     Cdip);
+    BLAS.synchronise();
+    RealD t4 = usecond();
+
+    std::vector<scalar> HOST_C(BLAS_Cred.size());      // nrhs . nrhs -- the coefficients 
+    acceleratorCopyFromDevice(&BLAS_Cred[0],&HOST_C[0],BLAS_Cred.size()*sizeof(scalar));
+
+    RealD t5 = usecond();
+    m = Eigen::MatrixXcd::Zero(nrhs,nrhs);
+    for(int ss=0;ss<vol;ss++){
+      Eigen::Map<Eigen::MatrixXcd> eC((std::complex<double> *)&HOST_C[ss*nrhs*nrhs],nrhs,nrhs);
+      m = m + eC;
+    }
+    RealD t6l = usecond();
+    grid->GlobalSumVector((scalar *) &m(0,0),nrhs*nrhs);
+    RealD t6 = usecond();
+    uint64_t M=nrhs;
+    uint64_t N=nrhs;
+    uint64_t K=vw;
+    RealD xybytes = grid->lSites()*sizeof(scalar_object);
+    RealD bytes = 1.0*sizeof(ComplexD)*(M*N*2+N*K+M*K);
+    RealD flops = 8.0*M*N*K;
+    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;
+#endif
+  }
+};
+
+NAMESPACE_END(Grid);

Grid/algorithms/deflation/MultiRHSBlockProject.h

@@ -0,0 +1,513 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: MultiRHSDeflation.h
+
+    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 */
+#pragma once
+
+NAMESPACE_BEGIN(Grid);
+
+
+/* 
+   MultiRHS block projection
+
+   Import basis -> nblock x nbasis x  (block x internal) 
+   Import vector of fine lattice objects -> nblock x nrhs x (block x internal) 
+
+   => coarse_(nrhs x nbasis )^block = via batched GEMM
+
+//template<class vobj,class CComplex,int nbasis,class VLattice>
+//inline void blockProject(Lattice<iVector<CComplex,nbasis > > &coarseData,
+//			   const VLattice &fineData,
+//			   const VLattice &Basis)
+*/
+
+template<class Field>
+class MultiRHSBlockProject
+{
+public:
+
+  typedef typename Field::scalar_type   scalar;
+  typedef typename Field::scalar_object scalar_object;
+  typedef Field Fermion;
+
+  int nbasis;
+  GridBase *coarse_grid;
+  GridBase *fine_grid;
+  uint64_t block_vol;
+  uint64_t fine_vol;
+  uint64_t coarse_vol;
+  uint64_t words;
+
+  // Row major layout "C" order:
+  // BLAS_V[coarse_vol][nbasis][block_vol][words]
+  // BLAS_F[coarse_vol][nrhs][block_vol][words]
+  // BLAS_C[coarse_vol][nrhs][nbasis]
+  /*
+   * in Fortran column major notation (cuBlas order)
+   *
+   * Vxb = [v1(x)][..][vn(x)] ... x coarse vol
+   *
+   * Fxr = [r1(x)][..][rm(x)] ... x coarse vol
+   *
+   * Block project:
+   * C_br = V^dag F x coarse vol
+   *
+   * Block promote:
+   * F_xr = Vxb Cbr x coarse_vol
+   */  
+  deviceVector<scalar> BLAS_V;      // words * block_vol * nbasis x coarse_vol 
+  deviceVector<scalar> BLAS_F;      // nrhs x fine_vol * words   -- the sources
+  deviceVector<scalar> BLAS_C;      // nrhs x coarse_vol * nbasis -- the coarse coeffs
+
+  RealD blasNorm2(deviceVector<scalar> &blas)
+  {
+    scalar ss(0.0);
+    std::vector<scalar> tmp(blas.size());
+    acceleratorCopyFromDevice(&blas[0],&tmp[0],blas.size()*sizeof(scalar));
+    for(int64_t s=0;s<blas.size();s++){
+      ss=ss+tmp[s]*adj(tmp[s]);
+    }
+    coarse_grid->GlobalSum(ss);
+    return real(ss);
+  }
+  
+  MultiRHSBlockProject(){};
+ ~MultiRHSBlockProject(){ Deallocate(); };
+  
+  void Deallocate(void)
+  {
+    nbasis=0;
+    coarse_grid=nullptr;
+    fine_grid=nullptr;
+    fine_vol=0;
+    block_vol=0;
+    coarse_vol=0;
+    words=0;
+    BLAS_V.resize(0);
+    BLAS_F.resize(0);
+    BLAS_C.resize(0);
+  }
+  void Allocate(int _nbasis,GridBase *_fgrid,GridBase *_cgrid)
+  {
+    nbasis=_nbasis;
+
+    fine_grid=_fgrid;
+    coarse_grid=_cgrid;
+
+    fine_vol   = fine_grid->lSites();
+    coarse_vol = coarse_grid->lSites();
+    block_vol = fine_vol/coarse_vol;
+    
+    words = sizeof(scalar_object)/sizeof(scalar);
+
+    BLAS_V.resize (fine_vol * words * nbasis );
+  }
+  void ImportFineGridVectors(std::vector <Field > &vecs, deviceVector<scalar> &blas)
+  {
+    int nvec = vecs.size();
+    typedef typename Field::vector_object vobj;
+    //    std::cout << GridLogMessage <<" BlockProjector importing "<<nvec<< " fine grid vectors" <<std::endl;
+
+    assert(vecs[0].Grid()==fine_grid);
+
+    subdivides(coarse_grid,fine_grid); // require they map
+
+    int _ndimension = coarse_grid->_ndimension;
+    assert(block_vol == fine_grid->oSites() / coarse_grid->oSites());
+    
+    Coordinate  block_r      (_ndimension);
+    for(int d=0 ; d<_ndimension;d++){
+      block_r[d] = fine_grid->_rdimensions[d] / coarse_grid->_rdimensions[d];
+    }
+
+    uint64_t sz = blas.size();
+
+    acceleratorMemSet(&blas[0],0,blas.size()*sizeof(scalar));
+
+    Coordinate fine_rdimensions = fine_grid->_rdimensions;
+    Coordinate coarse_rdimensions = coarse_grid->_rdimensions;
+    int64_t bv= block_vol;
+    for(int v=0;v<vecs.size();v++){
+
+      //      std::cout << " BlockProjector importing vector"<<v<<" "<<norm2(vecs[v])<<std::endl;
+      autoView( fineData   , vecs[v], AcceleratorRead);
+
+      auto blasData_p  = &blas[0];
+      auto fineData_p  = &fineData[0];
+
+      int64_t osites = fine_grid->oSites();
+
+      // loop over fine sites
+      const int Nsimd = vobj::Nsimd();
+      //      std::cout << "sz "<<sz<<std::endl;
+      //      std::cout << "prod "<<Nsimd * coarse_grid->oSites() * block_vol * nvec * words<<std::endl;
+      assert(sz == Nsimd * coarse_grid->oSites() * block_vol * nvec * words);
+      uint64_t lwords= words; // local variable for copy in to GPU
+      accelerator_for(sf,osites,Nsimd,{
+#ifdef GRID_SIMT
+        {
+	  int lane=acceleratorSIMTlane(Nsimd); // buffer lane
+#else
+	  for(int lane=0;lane<Nsimd;lane++) {
+#endif
+	  // One thread per fine site
+	  Coordinate coor_f(_ndimension);
+	  Coordinate coor_b(_ndimension);
+	  Coordinate coor_c(_ndimension);
+
+	  // Fine site to fine coor
+	  Lexicographic::CoorFromIndex(coor_f,sf,fine_rdimensions);
+
+	  for(int d=0;d<_ndimension;d++) coor_b[d] = coor_f[d]%block_r[d];
+	  for(int d=0;d<_ndimension;d++) coor_c[d] = coor_f[d]/block_r[d];
+	  
+	  int sc;// coarse site
+	  int sb;// block site
+	  Lexicographic::IndexFromCoor(coor_c,sc,coarse_rdimensions);
+	  Lexicographic::IndexFromCoor(coor_b,sb,block_r);
+
+          scalar_object data = extractLane(lane,fineData[sf]);
+
+	  // BLAS layout address calculation
+	  // words * block_vol * nbasis x coarse_vol
+	  // coarse oSite x block vole x lanes
+	  int64_t site = (lane*osites + sc*bv)*nvec
+   	               + v*bv
+	               + sb;
+
+	  //	  assert(site*lwords<sz);
+
+	  scalar_object * ptr = (scalar_object *)&blasData_p[site*lwords];
+
+	  *ptr = data;
+#ifdef GRID_SIMT
+	}
+#else
+	}
+#endif
+      });
+      //      std::cout << " import fine Blas norm "<<blasNorm2(blas)<<std::endl;
+      //      std::cout << " BlockProjector imported vector"<<v<<std::endl;
+    }
+  }
+  void ExportFineGridVectors(std::vector <Field> &vecs, deviceVector<scalar> &blas)
+  {
+    typedef typename Field::vector_object vobj;
+
+    int nvec = vecs.size();
+
+    assert(vecs[0].Grid()==fine_grid);
+
+    subdivides(coarse_grid,fine_grid); // require they map
+
+    int _ndimension = coarse_grid->_ndimension;
+    assert(block_vol == fine_grid->oSites() / coarse_grid->oSites());
+    
+    Coordinate  block_r      (_ndimension);
+    for(int d=0 ; d<_ndimension;d++){
+      block_r[d] = fine_grid->_rdimensions[d] / coarse_grid->_rdimensions[d];
+    }
+    Coordinate fine_rdimensions = fine_grid->_rdimensions;
+    Coordinate coarse_rdimensions = coarse_grid->_rdimensions;
+
+    //    std::cout << " export fine Blas norm "<<blasNorm2(blas)<<std::endl;
+
+    int64_t bv= block_vol;
+    for(int v=0;v<vecs.size();v++){
+
+      autoView( fineData   , vecs[v], AcceleratorWrite);
+
+      auto blasData_p  = &blas[0];
+      auto fineData_p    = &fineData[0];
+
+      int64_t osites = fine_grid->oSites();
+      uint64_t lwords = words;
+      //      std::cout << " Nsimd is "<<vobj::Nsimd() << std::endl;
+      //      std::cout << " lwords is "<<lwords << std::endl;
+      //      std::cout << " sizeof(scalar_object) is "<<sizeof(scalar_object) << std::endl;
+      // loop over fine sites
+      accelerator_for(sf,osites,vobj::Nsimd(),{
+      
+#ifdef GRID_SIMT
+        {
+	  int lane=acceleratorSIMTlane(vobj::Nsimd()); // buffer lane
+#else
+	  for(int lane=0;lane<vobj::Nsimd();lane++) {
+#endif
+	  // One thread per fine site
+	  Coordinate coor_f(_ndimension);
+	  Coordinate coor_b(_ndimension);
+	  Coordinate coor_c(_ndimension);
+
+	  Lexicographic::CoorFromIndex(coor_f,sf,fine_rdimensions);
+
+	  for(int d=0;d<_ndimension;d++) coor_b[d] = coor_f[d]%block_r[d];
+	  for(int d=0;d<_ndimension;d++) coor_c[d] = coor_f[d]/block_r[d];
+	  
+	  int sc;
+	  int sb;
+	  Lexicographic::IndexFromCoor(coor_c,sc,coarse_rdimensions);
+	  Lexicographic::IndexFromCoor(coor_b,sb,block_r);
+
+	  // BLAS layout address calculation
+	  // words * block_vol * nbasis x coarse_vol 	  
+	  int64_t site = (lane*osites + sc*bv)*nvec
+   	               + v*bv
+	               + sb;
+
+	  scalar_object * ptr = (scalar_object *)&blasData_p[site*lwords];
+
+	  scalar_object data = *ptr;
+
+	  insertLane(lane,fineData[sf],data);
+#ifdef GRID_SIMT
+	}
+#else
+	}
+#endif
+      });
+    }
+  }
+  template<class vobj>
+  void ImportCoarseGridVectors(std::vector <Lattice<vobj> > &vecs, deviceVector<scalar> &blas)
+  {
+    int nvec = vecs.size();
+    typedef typename vobj::scalar_object coarse_scalar_object;
+
+    //    std::cout << " BlockProjector importing "<<nvec<< " coarse grid vectors" <<std::endl;
+
+    assert(vecs[0].Grid()==coarse_grid);
+
+    int _ndimension = coarse_grid->_ndimension;
+
+    uint64_t sz = blas.size();
+
+    Coordinate coarse_rdimensions = coarse_grid->_rdimensions;
+    
+    for(int v=0;v<vecs.size();v++){
+
+      //      std::cout << " BlockProjector importing coarse vector"<<v<<" "<<norm2(vecs[v])<<std::endl;
+      autoView( coarseData   , vecs[v], AcceleratorRead);
+
+      auto blasData_p  = &blas[0];
+      auto coarseData_p  = &coarseData[0];
+
+      int64_t osites = coarse_grid->oSites();
+
+      // loop over fine sites
+      const int Nsimd = vobj::Nsimd();
+      uint64_t cwords=sizeof(typename vobj::scalar_object)/sizeof(scalar);
+      assert(cwords==nbasis);
+      
+      accelerator_for(sc,osites,Nsimd,{
+#ifdef GRID_SIMT
+        {
+	  int lane=acceleratorSIMTlane(Nsimd); // buffer lane
+#else
+	  for(int lane=0;lane<Nsimd;lane++) {
+#endif
+           // C_br per site
+	    int64_t blas_site = (lane*osites + sc)*nvec*cwords + v*cwords;
+	    
+	    coarse_scalar_object data = extractLane(lane,coarseData[sc]);
+
+	    coarse_scalar_object * ptr = (coarse_scalar_object *)&blasData_p[blas_site];
+
+	    *ptr = data;
+#ifdef GRID_SIMT
+	}
+#else
+	}
+#endif
+      });
+      //      std::cout << " import coarsee Blas norm "<<blasNorm2(blas)<<std::endl;
+    }
+  }
+  template<class vobj>
+  void ExportCoarseGridVectors(std::vector <Lattice<vobj> > &vecs, deviceVector<scalar> &blas)
+  {
+    int nvec = vecs.size();
+    typedef typename vobj::scalar_object coarse_scalar_object;
+    //    std::cout << GridLogMessage<<" BlockProjector exporting "<<nvec<< " coarse grid vectors" <<std::endl;
+
+    assert(vecs[0].Grid()==coarse_grid);
+
+    int _ndimension = coarse_grid->_ndimension;
+    
+    uint64_t sz = blas.size();
+
+    Coordinate coarse_rdimensions = coarse_grid->_rdimensions;
+    
+    //    std::cout << " export coarsee Blas norm "<<blasNorm2(blas)<<std::endl;
+    for(int v=0;v<vecs.size();v++){
+
+      //  std::cout << " BlockProjector exporting coarse vector"<<v<<std::endl;
+      autoView( coarseData   , vecs[v], AcceleratorWrite);
+
+      auto blasData_p  = &blas[0];
+      auto coarseData_p  = &coarseData[0];
+
+      int64_t osites = coarse_grid->oSites();
+
+      // loop over fine sites
+      const int Nsimd = vobj::Nsimd();
+      uint64_t cwords=sizeof(typename vobj::scalar_object)/sizeof(scalar);
+      assert(cwords==nbasis);
+      
+      accelerator_for(sc,osites,Nsimd,{
+	  // Wrap in a macro "FOR_ALL_LANES(lane,{ ... });
+#ifdef GRID_SIMT
+        {
+	  int lane=acceleratorSIMTlane(Nsimd); // buffer lane
+#else
+	  for(int lane=0;lane<Nsimd;lane++) {
+#endif
+	    int64_t blas_site = (lane*osites + sc)*nvec*cwords + v*cwords;
+	    coarse_scalar_object * ptr = (coarse_scalar_object *)&blasData_p[blas_site];
+	    coarse_scalar_object data = *ptr;
+	    insertLane(lane,coarseData[sc],data);
+#ifdef GRID_SIMT
+	}
+#else
+	}
+#endif
+      });
+    }
+  }
+  void ImportBasis(std::vector < Field > &vecs)
+  {
+    //    std::cout << " BlockProjector Import basis size "<<vecs.size()<<std::endl;
+    ImportFineGridVectors(vecs,BLAS_V);
+  }
+
+  template<class cobj>
+  void blockProject(std::vector<Field> &fine,std::vector< Lattice<cobj> > & coarse)
+  {
+    int nrhs=fine.size();
+    int _nbasis = sizeof(typename cobj::scalar_object)/sizeof(scalar);
+    //    std::cout << "blockProject nbasis " <<nbasis<<" " << _nbasis<<std::endl;
+    assert(nbasis==_nbasis);
+    
+    BLAS_F.resize (fine_vol * words * nrhs );
+    BLAS_C.resize (coarse_vol * nbasis * nrhs );
+
+    /////////////////////////////////////////////
+    // Copy in the multi-rhs sources to same data layout
+    /////////////////////////////////////////////
+    //    std::cout << "BlockProject import fine"<<std::endl;
+    ImportFineGridVectors(fine,BLAS_F);
+    
+    deviceVector<scalar *> Vd(coarse_vol);
+    deviceVector<scalar *> Fd(coarse_vol);
+    deviceVector<scalar *> Cd(coarse_vol);
+
+    //    std::cout << "BlockProject pointers"<<std::endl;
+    for(int c=0;c<coarse_vol;c++){
+      // BLAS_V[coarse_vol][nbasis][block_vol][words]
+      // BLAS_F[coarse_vol][nrhs][block_vol][words]
+      // BLAS_C[coarse_vol][nrhs][nbasis]
+      scalar * Vh = & BLAS_V[c*nbasis*block_vol*words];
+      scalar * Fh = & BLAS_F[c*nrhs*block_vol*words];
+      scalar * Ch = & BLAS_C[c*nrhs*nbasis];
+
+      acceleratorPut(Vd[c],Vh);
+      acceleratorPut(Fd[c],Fh);
+      acceleratorPut(Cd[c],Ch);
+    }
+
+    GridBLAS BLAS;
+
+    //    std::cout << "BlockProject BLAS"<<std::endl;
+    int64_t vw = block_vol * words;
+    /////////////////////////////////////////
+    // C_br = V^dag R
+    /////////////////////////////////////////
+    BLAS.gemmBatched(GridBLAS_OP_C,GridBLAS_OP_N, 
+    		     nbasis,nrhs,vw,
+		     scalar(1.0),
+		     Vd,
+		     Fd,
+		     scalar(0.0),  // wipe out C
+		     Cd);
+    BLAS.synchronise();
+    //    std::cout << "BlockProject done"<<std::endl;
+    ExportCoarseGridVectors(coarse, BLAS_C);
+    //    std::cout << "BlockProject done"<<std::endl;
+
+  }
+
+  template<class cobj>
+  void blockPromote(std::vector<Field> &fine,std::vector<Lattice<cobj> > & coarse)
+  {
+    int nrhs=fine.size();
+    int _nbasis = sizeof(typename cobj::scalar_object)/sizeof(scalar);
+    assert(nbasis==_nbasis);
+    
+    BLAS_F.resize (fine_vol * words * nrhs );
+    BLAS_C.resize (coarse_vol * nbasis * nrhs );
+
+    ImportCoarseGridVectors(coarse, BLAS_C);
+
+    GridBLAS BLAS;
+
+    deviceVector<scalar *> Vd(coarse_vol);
+    deviceVector<scalar *> Fd(coarse_vol);
+    deviceVector<scalar *> Cd(coarse_vol);
+
+    for(int c=0;c<coarse_vol;c++){
+      // BLAS_V[coarse_vol][nbasis][block_vol][words]
+      // BLAS_F[coarse_vol][nrhs][block_vol][words]
+      // BLAS_C[coarse_vol][nrhs][nbasis]
+      scalar * Vh = & BLAS_V[c*nbasis*block_vol*words];
+      scalar * Fh = & BLAS_F[c*nrhs*block_vol*words];
+      scalar * Ch = & BLAS_C[c*nrhs*nbasis];
+      acceleratorPut(Vd[c],Vh);
+      acceleratorPut(Fd[c],Fh);
+      acceleratorPut(Cd[c],Ch);
+    }
+
+    /////////////////////////////////////////
+    // Block promote:
+    // F_xr = Vxb Cbr (x coarse_vol)
+    /////////////////////////////////////////
+
+    int64_t vw = block_vol * words;
+    BLAS.gemmBatched(GridBLAS_OP_N,GridBLAS_OP_N, 
+    		     vw,nrhs,nbasis,
+		     scalar(1.0),
+		     Vd,
+		     Cd,
+		     scalar(0.0),  // wipe out C
+		     Fd);
+    BLAS.synchronise();
+    //    std::cout << " blas call done"<<std::endl;
+    
+    ExportFineGridVectors(fine, BLAS_F);
+    //    std::cout << " exported "<<std::endl;
+  }
+};
+
+NAMESPACE_END(Grid);

Grid/algorithms/deflation/MultiRHSDeflation.h

@@ -0,0 +1,233 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: MultiRHSDeflation.h
+
+    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 */
+#pragma once
+
+NAMESPACE_BEGIN(Grid);
+
+
+/* Need helper object for BLAS accelerated mrhs projection
+
+   i) MultiRHS Deflation
+
+   Import Evecs -> nev x vol x internal 
+   Import vector of Lattice objects -> nrhs x vol x internal
+   => Cij (nrhs x Nev) via GEMM.
+   => Guess  (nrhs x vol x internal)  = C x evecs (via GEMM)
+   Export
+
+   
+   ii) MultiRHS block projection
+
+   Import basis -> nblock x nbasis x  (block x internal) 
+   Import vector of fine lattice objects -> nblock x nrhs x (block x internal) 
+
+   => coarse_(nrhs x nbasis )^block = via batched GEMM
+
+   iii)   Alternate interface: 
+   Import higher dim Lattice object-> vol x nrhs layout
+   
+*/
+template<class Field>
+class MultiRHSDeflation
+{
+public:
+
+  typedef typename Field::scalar_type   scalar;
+  typedef typename Field::scalar_object scalar_object;
+
+  int nev;
+  std::vector<RealD> eval;
+  GridBase *grid;
+  uint64_t vol;
+  uint64_t words;
+  
+  deviceVector<scalar> BLAS_E;      //  nev x vol -- the eigenbasis   (up to a 1/sqrt(lambda))
+  deviceVector<scalar> BLAS_R;      // nrhs x vol -- the sources
+  deviceVector<scalar> BLAS_G;      // nrhs x vol -- the guess
+  deviceVector<scalar> BLAS_C;      // nrhs x nev -- the coefficients 
+  
+  MultiRHSDeflation(){};
+  ~MultiRHSDeflation(){ Deallocate(); };
+  
+  void Deallocate(void)
+  {
+    nev=0;
+    grid=nullptr;
+    vol=0;
+    words=0;
+    BLAS_E.resize(0);
+    BLAS_R.resize(0);
+    BLAS_C.resize(0);
+    BLAS_G.resize(0);
+  }
+  void Allocate(int _nev,GridBase *_grid)
+  {
+    nev=_nev;
+    grid=_grid;
+    vol   = grid->lSites();
+    words = sizeof(scalar_object)/sizeof(scalar);
+    eval.resize(nev);
+    BLAS_E.resize (vol * words * nev );
+    std::cout << GridLogMessage << " Allocate for "<<nev<<" eigenvectors and volume "<<vol<<std::endl;
+  }
+  void ImportEigenVector(Field &evec,RealD &_eval, int ev)
+  {
+    //    std::cout << " ev " <<ev<<" eval "<<_eval<< std::endl;
+    assert(ev<eval.size());
+    eval[ev] = _eval;
+
+    int64_t offset = ev*vol*words;
+    autoView(v,evec,AcceleratorRead);
+    acceleratorCopyDeviceToDevice(&v[0],&BLAS_E[offset],sizeof(scalar_object)*vol);
+
+  }
+  void ImportEigenBasis(std::vector<Field> &evec,std::vector<RealD> &_eval)
+  {
+    ImportEigenBasis(evec,_eval,0,evec.size());
+  }
+  // Could use to import a batch of eigenvectors
+  void ImportEigenBasis(std::vector<Field> &evec,std::vector<RealD> &_eval, int _ev0, int _nev)
+  {
+    assert(_ev0+_nev<=evec.size());
+
+    Allocate(_nev,evec[0].Grid());
+    
+    // Imports a sub-batch of eigenvectors, _ev0, ..., _ev0+_nev-1
+    for(int e=0;e<nev;e++){
+      std::cout << "Importing eigenvector "<<e<<" evalue "<<_eval[_ev0+e]<<std::endl;
+      ImportEigenVector(evec[_ev0+e],_eval[_ev0+e],e);
+    }
+  }
+  void DeflateSources(std::vector<Field> &source,std::vector<Field> & guess)
+  {
+    int nrhs = source.size();
+    assert(source.size()==guess.size());
+    assert(grid == guess[0].Grid());
+    conformable(guess[0],source[0]);
+
+    int64_t vw = vol * words;
+
+    RealD t0 = usecond();
+    BLAS_R.resize(nrhs * vw); // cost free if size doesn't change
+    BLAS_G.resize(nrhs * vw); // cost free if size doesn't change
+    BLAS_C.resize(nev * nrhs);// cost free if size doesn't change
+
+    /////////////////////////////////////////////
+    // Copy in the multi-rhs sources
+    /////////////////////////////////////////////
+    //    for(int r=0;r<nrhs;r++){
+    //      std::cout << " source["<<r<<"] = "<<norm2(source[r])<<std::endl;
+    //    }
+    for(int r=0;r<nrhs;r++){
+      int64_t offset = r*vw;
+      autoView(v,source[r],AcceleratorRead);
+      acceleratorCopyDeviceToDevice(&v[0],&BLAS_R[offset],sizeof(scalar_object)*vol);
+    }
+
+  /*
+   * in Fortran column major notation (cuBlas order)
+   *
+   * Exe = [e1(x)][..][en(x)]
+   *
+   * Rxr = [r1(x)][..][rm(x)]
+   *
+   * C_er = E^dag R
+   * C_er = C_er / lambda_e 
+   * G_xr = Exe Cer
+   */
+    deviceVector<scalar *> Ed(1);
+    deviceVector<scalar *> Rd(1);
+    deviceVector<scalar *> Cd(1);
+    deviceVector<scalar *> Gd(1);
+
+    scalar * Eh = & BLAS_E[0];
+    scalar * Rh = & BLAS_R[0];
+    scalar * Ch = & BLAS_C[0];
+    scalar * Gh = & BLAS_G[0];
+
+    acceleratorPut(Ed[0],Eh);
+    acceleratorPut(Rd[0],Rh);
+    acceleratorPut(Cd[0],Ch);
+    acceleratorPut(Gd[0],Gh);
+
+    GridBLAS BLAS;
+
+    /////////////////////////////////////////
+    // C_er = E^dag R
+    /////////////////////////////////////////
+    BLAS.gemmBatched(GridBLAS_OP_C,GridBLAS_OP_N, 
+    		     nev,nrhs,vw,
+		     scalar(1.0),
+		     Ed,
+		     Rd,
+		     scalar(0.0),  // wipe out C
+		     Cd);
+    BLAS.synchronise();
+
+    assert(BLAS_C.size()==nev*nrhs);
+
+    std::vector<scalar> HOST_C(BLAS_C.size());      // nrhs . nev -- the coefficients 
+    acceleratorCopyFromDevice(&BLAS_C[0],&HOST_C[0],BLAS_C.size()*sizeof(scalar));
+    grid->GlobalSumVector(&HOST_C[0],nev*nrhs);
+    for(int e=0;e<nev;e++){
+      RealD lam(1.0/eval[e]);
+      for(int r=0;r<nrhs;r++){
+	int off = e+nev*r;
+	HOST_C[off]=HOST_C[off] * lam;
+	//	std::cout << "C["<<e<<"]["<<r<<"] ="<<HOST_C[off]<< " eval[e] "<<eval[e] <<std::endl;
+      }
+    }
+    acceleratorCopyToDevice(&HOST_C[0],&BLAS_C[0],BLAS_C.size()*sizeof(scalar));
+
+    
+    /////////////////////////////////////////
+    // Guess G_xr = Exe Cer
+    /////////////////////////////////////////
+    BLAS.gemmBatched(GridBLAS_OP_N,GridBLAS_OP_N, 
+		     vw,nrhs,nev,
+		     scalar(1.0),
+		     Ed, // x . nev
+		     Cd, // nev . nrhs
+		     scalar(0.0),
+		     Gd);
+    BLAS.synchronise();
+
+    ///////////////////////////////////////
+    // Copy out the multirhs
+    ///////////////////////////////////////
+    for(int r=0;r<nrhs;r++){
+      int64_t offset = r*vw;
+      autoView(v,guess[r],AcceleratorWrite);
+      acceleratorCopyDeviceToDevice(&BLAS_G[offset],&v[0],sizeof(scalar_object)*vol);
+    }
+    RealD t1 = usecond();
+    std::cout << GridLogMessage << "MultiRHSDeflation for "<<nrhs<<" sources with "<<nev<<" eigenvectors took " << (t1-t0)/1e3 <<" ms"<<std::endl;
+  }
+};
+
+NAMESPACE_END(Grid);

Grid/algorithms/iterative/AdefGeneric.h

@@ -41,6 +41,7 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
    */
 NAMESPACE_BEGIN(Grid);
 
+
 template<class Field>
 class TwoLevelCG : public LinearFunction<Field>
 {
@@ -69,35 +70,54 @@ class TwoLevelCG : public LinearFunction<Field>
   
   virtual void operator() (const Field &src, Field &x)
   {
-    Field resid(grid);
+    std::cout << GridLogMessage<<"HDCG: fPcg starting single RHS"<<std::endl;
     RealD f;
     RealD rtzp,rtz,a,d,b;
     RealD rptzp;
 
-    Field p(grid);
+    /////////////////////////////
+    // Set up history vectors
+    /////////////////////////////
+    int mmax = 5;
+    std::cout << GridLogMessage<<"HDCG: fPcg allocating"<<std::endl;
+    std::vector<Field> p(mmax,grid);
+    std::vector<Field> mmp(mmax,grid);
+    std::vector<RealD> pAp(mmax);
     Field z(grid);
     Field tmp(grid);
-    Field mmp(grid);
+    Field  mp (grid);
     Field  r  (grid);
     Field  mu (grid);
-    Field rp (grid);
     
+    std::cout << GridLogMessage<<"HDCG: fPcg allocated"<<std::endl;
     //Initial residual computation & set up
-    double tn;
+    RealD guess   = norm2(x);
+    std::cout << GridLogMessage<<"HDCG: fPcg guess nrm "<<guess<<std::endl;
+    RealD src_nrm = norm2(src);
+    std::cout << GridLogMessage<<"HDCG: fPcg src nrm "<<src_nrm<<std::endl;
+    
+    if ( src_nrm == 0.0 ) {
+      std::cout << GridLogMessage<<"HDCG: fPcg given trivial source norm "<<src_nrm<<std::endl;
+      x=Zero();
+    }
+    RealD tn;
     
     GridStopWatch HDCGTimer;
     HDCGTimer.Start();
     //////////////////////////
     // x0 = Vstart -- possibly modify guess
     //////////////////////////
-    x=Zero();
     Vstart(x,src);
     
     // r0 = b -A x0
-    _FineLinop.HermOp(x,mmp);
-
-    axpy(r, -1.0, mmp, src);    // Recomputes r=src-x0
-    rp=r;
+    _FineLinop.HermOp(x,mmp[0]);
+    axpy (r, -1.0,mmp[0], src);    // Recomputes r=src-Ax0
+    {
+      double n1 = norm2(x);
+      double n2 = norm2(mmp[0]);
+      double n3 = norm2(r);
+      std::cout<<GridLogMessage<<"x,vstart,r = "<<n1<<" "<<n2<<" "<<n3<<std::endl;
+    }
 
     //////////////////////////////////
     // Compute z = M1 x
@@ -106,58 +126,88 @@ class TwoLevelCG : public LinearFunction<Field>
     rtzp =real(innerProduct(r,z));
     
     ///////////////////////////////////////
-    // Except Def2, M2 is trivial
+    // Solve for Mss mu = P A z and set p = z-mu
+    // Def2 p = 1 - Q Az = Pright z
+    // Other algos M2 is trivial
     ///////////////////////////////////////
-    p=z;
+    PcgM2(z,p[0]);
 
     RealD ssq =  norm2(src);
     RealD rsq =  ssq*Tolerance*Tolerance;
 
-    std::cout<<GridLogMessage<<"HDCG: k=0 residual "<<rtzp<<" target rsq "<<rsq<<" ssq "<<ssq<<std::endl;
+    std::cout << GridLogMessage<<"HDCG: k=0 residual "<<rtzp<<" rsq "<<rsq<<"\n";
 
-    for (int k=1;k<=MaxIterations;k++){
+    Field pp(grid);
+
+    for (int k=0;k<=MaxIterations;k++){
+    
+      int peri_k  = k % mmax;
+      int peri_kp = (k+1) % mmax;
 
       rtz=rtzp;
-      d= PcgM3(p,mmp);
+      d= PcgM3(p[peri_k],mmp[peri_k]);
       a = rtz/d;
     
-      axpy(x,a,p,x);
-      RealD rn = axpy_norm(r,-a,mmp,r);
+      // Memorise this
+      pAp[peri_k] = d;
       
+      axpy(x,a,p[peri_k],x);
+      RealD rn = axpy_norm(r,-a,mmp[peri_k],r);
+
+      // Compute z = M x
       PcgM1(r,z);
       
-      rtzp =real(innerProduct(r,z));
-
-      int ipcg=1; // almost free inexact preconditioned CG
-      if (ipcg) {
-	rptzp =real(innerProduct(rp,z));
-      } else {
-	rptzp =0;
+      {
+	RealD n1,n2;
+	n1=norm2(r);
+	n2=norm2(z);
+	std::cout << GridLogMessage<<"HDCG::fPcg iteration "<<k<<" : vector r,z "<<n1<<" "<<n2<<"\n";
       }
-      b = (rtzp-rptzp)/rtz;
+      rtzp =real(innerProduct(r,z));
+      std::cout << GridLogMessage<<"HDCG::fPcg iteration "<<k<<" : inner rtzp "<<rtzp<<"\n";
 
+      //    PcgM2(z,p[0]);
       PcgM2(z,mu); // ADEF-2 this is identity. Axpy possible to eliminate
       
-      axpy(p,b,p,mu);  // mu = A r
+      p[peri_kp]=mu;
+
+      // Standard search direction  p -> z + b p    
+      b = (rtzp)/rtz;
+      
+      int northog;
+      // k=zero  <=> peri_kp=1;        northog = 1
+      // k=1     <=> peri_kp=2;        northog = 2
+      // ...               ...                  ...
+      // k=mmax-2<=> peri_kp=mmax-1;   northog = mmax-1
+      // k=mmax-1<=> peri_kp=0;        northog = 1
+
+      //    northog     = (peri_kp==0)?1:peri_kp; // This is the fCG(mmax) algorithm
+      northog     = (k>mmax-1)?(mmax-1):k;        // This is the fCG-Tr(mmax-1) algorithm
+    
+      std::cout<<GridLogMessage<<"HDCG::fPcg iteration "<<k<<" : orthogonalising to last "<<northog<<" vectors\n";
+      for(int back=0; back < northog; back++){
+	int peri_back = (k-back)%mmax;
+	RealD pbApk= real(innerProduct(mmp[peri_back],p[peri_kp]));
+	RealD beta = -pbApk/pAp[peri_back];
+	axpy(p[peri_kp],beta,p[peri_back],p[peri_kp]);
+      }
 
       RealD rrn=sqrt(rn/ssq);
       RealD rtn=sqrt(rtz/ssq);
-      std::cout<<GridLogMessage<<"HDCG: Pcg k= "<<k<<" residual = "<<rrn<<std::endl;
+      RealD rtnp=sqrt(rtzp/ssq);
 
-      if ( ipcg ) {
-	axpy(rp,0.0,r,r);
-      }
+      std::cout<<GridLogMessage<<"HDCG: fPcg k= "<<k<<" residual = "<<rrn<<"\n";
 
       // Stopping condition
       if ( rn <= rsq ) { 
 
 	HDCGTimer.Stop();
-	std::cout<<GridLogMessage<<"HDCG: Pcg converged in "<<k<<" iterations and "<<HDCGTimer.Elapsed()<<std::endl;;
+	std::cout<<GridLogMessage<<"HDCG: fPcg converged in "<<k<<" iterations and "<<HDCGTimer.Elapsed()<<std::endl;;
 	
-	_FineLinop.HermOp(x,mmp);			  
-	axpy(tmp,-1.0,src,mmp);
+	_FineLinop.HermOp(x,mmp[0]);			  
+	axpy(tmp,-1.0,src,mmp[0]);
 	
-	RealD  mmpnorm = sqrt(norm2(mmp));
+	RealD  mmpnorm = sqrt(norm2(mmp[0]));
 	RealD  xnorm   = sqrt(norm2(x));
 	RealD  srcnorm = sqrt(norm2(src));
 	RealD  tmpnorm = sqrt(norm2(tmp));
@@ -173,18 +223,196 @@ class TwoLevelCG : public LinearFunction<Field>
       }
 
     }
-    std::cout<<GridLogMessage<<"HDCG: not converged"<<std::endl;
+    HDCGTimer.Stop();
+    std::cout<<GridLogMessage<<"HDCG: not converged "<<HDCGTimer.Elapsed()<<std::endl;
     RealD  xnorm   = sqrt(norm2(x));
     RealD  srcnorm = sqrt(norm2(src));
     std::cout<<GridLogMessage<<"HDCG: non-converged solution "<<xnorm<<" source "<<srcnorm<<std::endl;
+  }
 
+
+
+  virtual void operator() (std::vector<Field> &src, std::vector<Field> &x)
+  {
+    std::cout << GridLogMessage<<"HDCG: mrhs fPcg starting"<<std::endl;
+    src[0].Grid()->Barrier();
+    int nrhs = src.size();
+    std::vector<RealD> f(nrhs);
+    std::vector<RealD> rtzp(nrhs);
+    std::vector<RealD> rtz(nrhs);
+    std::vector<RealD> a(nrhs);
+    std::vector<RealD> d(nrhs);
+    std::vector<RealD> b(nrhs);
+    std::vector<RealD> rptzp(nrhs);
+    /////////////////////////////
+    // Set up history vectors
+    /////////////////////////////
+    int mmax = 3;
+    std::cout << GridLogMessage<<"HDCG: fPcg allocating"<<std::endl;
+    src[0].Grid()->Barrier();
+    std::vector<std::vector<Field> > p(nrhs);   for(int r=0;r<nrhs;r++)  p[r].resize(mmax,grid);
+    std::cout << GridLogMessage<<"HDCG: fPcg allocated p"<<std::endl;
+    src[0].Grid()->Barrier();
+    std::vector<std::vector<Field> > mmp(nrhs); for(int r=0;r<nrhs;r++) mmp[r].resize(mmax,grid);
+    std::cout << GridLogMessage<<"HDCG: fPcg allocated mmp"<<std::endl;
+    src[0].Grid()->Barrier();
+    std::vector<std::vector<RealD> > pAp(nrhs); for(int r=0;r<nrhs;r++) pAp[r].resize(mmax);
+    std::cout << GridLogMessage<<"HDCG: fPcg allocated pAp"<<std::endl;
+    src[0].Grid()->Barrier();
+    std::vector<Field> z(nrhs,grid);
+    std::vector<Field>  mp (nrhs,grid);
+    std::vector<Field>  r  (nrhs,grid);
+    std::vector<Field>  mu (nrhs,grid);
+    std::cout << GridLogMessage<<"HDCG: fPcg allocated z,mp,r,mu"<<std::endl;
+    src[0].Grid()->Barrier();
+
+    //Initial residual computation & set up
+    std::vector<RealD> src_nrm(nrhs);
+    for(int rhs=0;rhs<nrhs;rhs++) {
+      src_nrm[rhs]=norm2(src[rhs]);
+      assert(src_nrm[rhs]!=0.0);
+    }
+    std::vector<RealD> tn(nrhs);
+
+    GridStopWatch HDCGTimer;
+    HDCGTimer.Start();
+    //////////////////////////
+    // x0 = Vstart -- possibly modify guess
+    //////////////////////////
+    Vstart(x,src);
+
+    for(int rhs=0;rhs<nrhs;rhs++){
+      // r0 = b -A x0
+      _FineLinop.HermOp(x[rhs],mmp[rhs][0]);
+      axpy (r[rhs], -1.0,mmp[rhs][0], src[rhs]);    // Recomputes r=src-Ax0
+    }
+
+    //////////////////////////////////
+    // Compute z = M1 x
+    //////////////////////////////////
+    // This needs a multiRHS version for acceleration
+    PcgM1(r,z);
+
+    std::vector<RealD> ssq(nrhs);
+    std::vector<RealD> rsq(nrhs);
+    std::vector<Field> pp(nrhs,grid);
+
+    for(int rhs=0;rhs<nrhs;rhs++){
+      rtzp[rhs] =real(innerProduct(r[rhs],z[rhs]));
+      p[rhs][0]=z[rhs];
+      ssq[rhs]=norm2(src[rhs]);
+      rsq[rhs]=  ssq[rhs]*Tolerance*Tolerance;
+      std::cout << GridLogMessage<<"mrhs HDCG: "<<rhs<<" k=0 residual "<<rtzp[rhs]<<" rsq "<<rsq[rhs]<<"\n";
+    }
+
+    std::vector<RealD> rn(nrhs);
+    for (int k=0;k<=MaxIterations;k++){
+    
+      int peri_k  = k % mmax;
+      int peri_kp = (k+1) % mmax;
+
+      for(int rhs=0;rhs<nrhs;rhs++){
+	rtz[rhs]=rtzp[rhs];
+	d[rhs]= PcgM3(p[rhs][peri_k],mmp[rhs][peri_k]);
+	a[rhs] = rtz[rhs]/d[rhs];
+    
+	// Memorise this
+	pAp[rhs][peri_k] = d[rhs];
+
+	axpy(x[rhs],a[rhs],p[rhs][peri_k],x[rhs]);
+	rn[rhs] = axpy_norm(r[rhs],-a[rhs],mmp[rhs][peri_k],r[rhs]);
+      }
+
+      // Compute z = M x (for *all* RHS)
+      PcgM1(r,z);
+      std::cout << GridLogMessage<<"HDCG::fPcg M1 complete"<<std::endl;
+      grid->Barrier();
+      
+      RealD max_rn=0.0;
+      for(int rhs=0;rhs<nrhs;rhs++){
+
+	rtzp[rhs] =real(innerProduct(r[rhs],z[rhs]));
+
+	std::cout << GridLogMessage<<"HDCG::fPcg rhs"<<rhs<<" iteration "<<k<<" : inner rtzp "<<rtzp[rhs]<<"\n";
+	
+	mu[rhs]=z[rhs];
+
+	p[rhs][peri_kp]=mu[rhs];
+
+	// Standard search direction p == z + b p 
+	b[rhs] = (rtzp[rhs])/rtz[rhs];
+
+	int northog = (k>mmax-1)?(mmax-1):k;        // This is the fCG-Tr(mmax-1) algorithm
+	std::cout<<GridLogMessage<<"HDCG::fPcg iteration "<<k<<" : orthogonalising to last "<<northog<<" vectors\n";
+	for(int back=0; back < northog; back++){
+	  int peri_back = (k-back)%mmax;
+	  RealD pbApk= real(innerProduct(mmp[rhs][peri_back],p[rhs][peri_kp]));
+	  RealD beta = -pbApk/pAp[rhs][peri_back];
+	  axpy(p[rhs][peri_kp],beta,p[rhs][peri_back],p[rhs][peri_kp]);
+	}
+
+	RealD rrn=sqrt(rn[rhs]/ssq[rhs]);
+	RealD rtn=sqrt(rtz[rhs]/ssq[rhs]);
+	RealD rtnp=sqrt(rtzp[rhs]/ssq[rhs]);
+	
+	std::cout<<GridLogMessage<<"HDCG: rhs "<<rhs<<"fPcg k= "<<k<<" residual = "<<rrn<<"\n";
+	if ( rrn > max_rn ) max_rn = rrn;
+      }
+
+      // Stopping condition based on worst case
+      if ( max_rn <= Tolerance ) { 
+
+	HDCGTimer.Stop();
+	std::cout<<GridLogMessage<<"HDCG: mrhs fPcg converged in "<<k<<" iterations and "<<HDCGTimer.Elapsed()<<std::endl;;
+
+	for(int rhs=0;rhs<nrhs;rhs++){
+	  _FineLinop.HermOp(x[rhs],mmp[rhs][0]);			  
+	  Field tmp(grid);
+	  axpy(tmp,-1.0,src[rhs],mmp[rhs][0]);
+      
+	  RealD  mmpnorm = sqrt(norm2(mmp[rhs][0]));
+	  RealD  xnorm   = sqrt(norm2(x[rhs]));
+	  RealD  srcnorm = sqrt(norm2(src[rhs]));
+	  RealD  tmpnorm = sqrt(norm2(tmp));
+	  RealD  true_residual = tmpnorm/srcnorm;
+	  std::cout<<GridLogMessage
+		   <<"HDCG: true residual ["<<rhs<<"] is "<<true_residual
+		   <<" solution "<<xnorm
+		   <<" source "<<srcnorm
+		   <<" mmp "<<mmpnorm	  
+		   <<std::endl;
+	}
 	return;
       }
       
+    }
+    HDCGTimer.Stop();
+    std::cout<<GridLogMessage<<"HDCG: not converged "<<HDCGTimer.Elapsed()<<std::endl;
+    for(int rhs=0;rhs<nrhs;rhs++){
+      RealD  xnorm   = sqrt(norm2(x[rhs]));
+      RealD  srcnorm = sqrt(norm2(src[rhs]));
+      std::cout<<GridLogMessage<<"HDCG: non-converged solution "<<xnorm<<" source "<<srcnorm<<std::endl;
+    }
+  }
+  
 
  public:
 
+  virtual void PcgM1(std::vector<Field> & in,std::vector<Field> & out)
+  {
+    std::cout << "PcgM1 default (cheat) mrhs version"<<std::endl;
+    for(int rhs=0;rhs<in.size();rhs++){
+      this->PcgM1(in[rhs],out[rhs]);
+    }
+  }
   virtual void PcgM1(Field & in, Field & out)     =0;
+  virtual void Vstart(std::vector<Field> & x,std::vector<Field> & src)
+  {
+    std::cout << "Vstart default (cheat) mrhs version"<<std::endl;
+    for(int rhs=0;rhs<x.size();rhs++){
+      this->Vstart(x[rhs],src[rhs]);
+    }
+  }
   virtual void Vstart(Field & x,const Field & src)=0;
 
   virtual void PcgM2(const Field & in, Field & out) {
@@ -239,6 +467,7 @@ class TwoLevelADEF2 : public TwoLevelCG<Field>
 
   virtual void PcgM1(Field & in, Field & out)
   {
+    GRID_TRACE("MultiGridPreconditioner ");
     // [PTM+Q] in = [1 - Q A] M in + Q in = Min + Q [ in -A Min]
 
     Field tmp(this->grid);
@@ -281,6 +510,7 @@ class TwoLevelADEF2 : public TwoLevelCG<Field>
 
   virtual void Vstart(Field & x,const Field & src)
   {
+    std::cout << GridLogMessage<<"HDCG: fPcg Vstart "<<std::endl;
     ///////////////////////////////////
     // Choose x_0 such that 
     // x_0 = guess +  (A_ss^inv) r_s = guess + Ass_inv [src -Aguess]
@@ -297,14 +527,18 @@ class TwoLevelADEF2 : public TwoLevelCG<Field>
     CoarseField PleftProj(this->coarsegrid);
     CoarseField PleftMss_proj(this->coarsegrid);
 
+    std::cout << GridLogMessage<<"HDCG: fPcg Vstart projecting "<<std::endl;
     this->_Aggregates.ProjectToSubspace(PleftProj,src);     
+    std::cout << GridLogMessage<<"HDCG: fPcg Vstart coarse solve "<<std::endl;
     this->_CoarseSolverPrecise(PleftProj,PleftMss_proj); // Ass^{-1} r_s
+    std::cout << GridLogMessage<<"HDCG: fPcg Vstart promote "<<std::endl;
     this->_Aggregates.PromoteFromSubspace(PleftMss_proj,x);  
 
   }
 
 };
 
+  
 template<class Field>
 class TwoLevelADEF1defl : public TwoLevelCG<Field>
 {
@@ -334,6 +568,7 @@ public:
   // Override PcgM1
   virtual void PcgM1(Field & in, Field & out)
   {
+    GRID_TRACE("EvecPreconditioner ");
     int N=evec.size();
     Field Pin(this->grid);
     Field Qin(this->grid);

Grid/algorithms/iterative/AdefMrhs.h

@@ -0,0 +1,718 @@
+    /*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./lib/algorithms/iterative/AdefGeneric.h
+
+    Copyright (C) 2015
+
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+
+    This program is free software; you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation; either version 2 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License along
+    with this program; if not, write to the Free Software Foundation, Inc.,
+    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+    See the full license in the file "LICENSE" in the top level distribution directory
+    *************************************************************************************/
+    /*  END LEGAL */
+#pragma once
+
+
+  /*
+   * Compared to Tang-2009:  P=Pleft. P^T = PRight Q=MssInv. 
+   * Script A = SolverMatrix 
+   * Script P = Preconditioner
+   *
+   * Implement ADEF-2
+   *
+   * Vstart = P^Tx + Qb
+   * M1 = P^TM + Q
+   * M2=M3=1
+   */
+NAMESPACE_BEGIN(Grid);
+
+
+template<class Field>
+class TwoLevelCGmrhs
+{
+ public:
+  RealD   Tolerance;
+  Integer MaxIterations;
+  GridBase *grid;
+
+  // Fine operator, Smoother, CoarseSolver
+  LinearOperatorBase<Field>   &_FineLinop;
+  LinearFunction<Field>   &_Smoother;
+  MultiRHSBlockCGLinalg<Field> _BlockCGLinalg;
+
+  GridStopWatch ProjectTimer;
+  GridStopWatch PromoteTimer;
+  GridStopWatch DeflateTimer;
+  GridStopWatch CoarseTimer;
+  GridStopWatch FineTimer;
+  GridStopWatch SmoothTimer;
+  GridStopWatch InsertTimer;
+
+  
+  // more most opertor functions
+  TwoLevelCGmrhs(RealD tol,
+		 Integer maxit,
+		 LinearOperatorBase<Field>   &FineLinop,
+		 LinearFunction<Field>       &Smoother,
+		 GridBase *fine) : 
+    Tolerance(tol), 
+    MaxIterations(maxit),
+    _FineLinop(FineLinop),
+    _Smoother(Smoother)
+  {
+    grid       = fine;
+  };
+  
+  // Vector case
+  virtual void operator() (std::vector<Field> &src, std::vector<Field> &x)
+  {
+    SolveSingleSystem(src,x);
+    //    SolvePrecBlockCG(src,x);
+  }
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+// Thin QR factorisation (google it)
+////////////////////////////////////////////////////////////////////////////////////////////////////
+  ////////////////////////////////////////////////////////////////////////////////////////////////////
+  //Dimensions
+  // R_{ferm x Nblock} =  Q_{ferm x Nblock} x  C_{Nblock x Nblock} -> ferm x Nblock
+  //
+  // Rdag R = m_rr = Herm = L L^dag        <-- Cholesky decomposition (LLT routine in Eigen)
+  //
+  //   Q  C = R => Q = R C^{-1}
+  //
+  // Want  Ident = Q^dag Q = C^{-dag} R^dag R C^{-1} = C^{-dag} L L^dag C^{-1} = 1_{Nblock x Nblock} 
+  //
+  // Set C = L^{dag}, and then Q^dag Q = ident 
+  //
+  // Checks:
+  // Cdag C = Rdag R ; passes.
+  // QdagQ  = 1      ; passes
+  ////////////////////////////////////////////////////////////////////////////////////////////////////
+  void ThinQRfact (Eigen::MatrixXcd &m_zz,
+		   Eigen::MatrixXcd &C,
+		   Eigen::MatrixXcd &Cinv,
+		   std::vector<Field> &  Q,
+		   std::vector<Field> & MQ,
+		   const std::vector<Field> & Z,
+		   const std::vector<Field> & MZ)
+  {
+    RealD t0=usecond();
+    _BlockCGLinalg.InnerProductMatrix(m_zz,MZ,Z);
+    RealD t1=usecond();
+
+    m_zz = 0.5*(m_zz+m_zz.adjoint());
+    
+    Eigen::MatrixXcd L    = m_zz.llt().matrixL(); 
+    
+    C    = L.adjoint();
+    Cinv = C.inverse();
+    
+    RealD t3=usecond();
+    _BlockCGLinalg.MulMatrix( Q,Cinv,Z);
+    _BlockCGLinalg.MulMatrix(MQ,Cinv,MZ);
+    RealD t4=usecond();
+    std::cout << " ThinQRfact IP    :"<< t1-t0<<" us"<<std::endl;
+    std::cout << " ThinQRfact Eigen :"<< t3-t1<<" us"<<std::endl;
+    std::cout << " ThinQRfact MulMat:"<< t4-t3<<" us"<<std::endl;
+  }
+
+  virtual void SolvePrecBlockCG (std::vector<Field> &src, std::vector<Field> &X)
+  {
+    std::cout << GridLogMessage<<"HDCG: mrhs fPrecBlockcg starting"<<std::endl;
+    src[0].Grid()->Barrier();
+    int nrhs = src.size();
+    //    std::vector<RealD> f(nrhs);
+    //    std::vector<RealD> rtzp(nrhs);
+    //    std::vector<RealD> rtz(nrhs);
+    //    std::vector<RealD> a(nrhs);
+    //    std::vector<RealD> d(nrhs);
+    //    std::vector<RealD> b(nrhs);
+    //    std::vector<RealD> rptzp(nrhs);
+
+    ////////////////////////////////////////////
+    //Initial residual computation & set up
+    ////////////////////////////////////////////
+    std::vector<RealD> ssq(nrhs);
+    for(int rhs=0;rhs<nrhs;rhs++){
+      ssq[rhs]=norm2(src[rhs]); assert(ssq[rhs]!=0.0);
+    }      
+
+    ///////////////////////////
+    // Fields -- eliminate duplicates between fPcg and block cg
+    ///////////////////////////
+    std::vector<Field> Mtmp(nrhs,grid);
+    std::vector<Field> tmp(nrhs,grid);
+    std::vector<Field>   Z(nrhs,grid); // Rename Z to R
+    std::vector<Field>  MZ(nrhs,grid); // Rename MZ to Z
+    std::vector<Field>   Q(nrhs,grid); // 
+    std::vector<Field>  MQ(nrhs,grid); // Rename to P
+    std::vector<Field>   D(nrhs,grid);
+    std::vector<Field>  AD(nrhs,grid);
+    
+    /************************************************************************
+     * Preconditioned Block conjugate gradient rQ
+     * Generalise Sebastien Birk Thesis, after Dubrulle 2001.
+     * Introduce preconditioning following Saad Ch9
+     ************************************************************************
+     * Dimensions:
+     *
+     *   X,B etc... ==(Nferm x nrhs)
+     *  Matrix A==(Nferm x Nferm)
+     *  
+     * Nferm = Nspin x Ncolour x Ncomplex x Nlattice_site
+     * QC => Thin QR factorisation (google it)
+     *
+     * R = B-AX
+     * Z = Mi R
+     * QC = Z
+     * D = Q 
+     * for k: 
+     *   R  = AD
+     *   Z  = Mi R
+     *   M  = [D^dag R]^{-1}
+     *   X  = X + D M C
+     *   QS = Q - Z.M
+     *   D  = Q + D S^dag
+     *   C  = S C
+     */
+    Eigen::MatrixXcd m_DZ     = Eigen::MatrixXcd::Identity(nrhs,nrhs);
+    Eigen::MatrixXcd m_M      = Eigen::MatrixXcd::Identity(nrhs,nrhs);
+    Eigen::MatrixXcd m_zz     = Eigen::MatrixXcd::Zero(nrhs,nrhs);
+    Eigen::MatrixXcd m_rr     = Eigen::MatrixXcd::Zero(nrhs,nrhs);
+    
+    Eigen::MatrixXcd m_C      = Eigen::MatrixXcd::Zero(nrhs,nrhs);
+    Eigen::MatrixXcd m_Cinv   = Eigen::MatrixXcd::Zero(nrhs,nrhs);
+    Eigen::MatrixXcd m_S      = Eigen::MatrixXcd::Zero(nrhs,nrhs);
+    Eigen::MatrixXcd m_Sinv   = Eigen::MatrixXcd::Zero(nrhs,nrhs);
+    
+    Eigen::MatrixXcd m_tmp    = Eigen::MatrixXcd::Identity(nrhs,nrhs);
+    Eigen::MatrixXcd m_tmp1   = Eigen::MatrixXcd::Identity(nrhs,nrhs);
+
+    GridStopWatch HDCGTimer;
+
+    //////////////////////////
+    // x0 = Vstart -- possibly modify guess
+    //////////////////////////
+    Vstart(X,src);
+
+    //////////////////////////
+    // R = B-AX
+    //////////////////////////
+    for(int rhs=0;rhs<nrhs;rhs++){
+      // r0 = b -A x0
+      _FineLinop.HermOp(X[rhs],tmp[rhs]);
+      axpy (Z[rhs], -1.0,tmp[rhs], src[rhs]);    // Computes R=Z=src - A X0
+    }
+
+    //////////////////////////////////
+    // Compute MZ = M1 Z = M1 B - M1 A x0
+    //////////////////////////////////
+    PcgM1(Z,MZ);  
+
+    //////////////////////////////////
+    // QC = Z
+    //////////////////////////////////
+    ThinQRfact (m_zz, m_C, m_Cinv, Q, MQ, Z, MZ);
+
+    //////////////////////////////////
+    // D=MQ
+    //////////////////////////////////
+    for(int b=0;b<nrhs;b++) D[b]=MQ[b]; // LLT rotation of the MZ basis of search dirs
+
+    std::cout << GridLogMessage<<"PrecBlockCGrQ vec computed initial residual and QR fact " <<std::endl;
+
+    ProjectTimer.Reset();
+    PromoteTimer.Reset();
+    DeflateTimer.Reset();
+    CoarseTimer.Reset();
+    SmoothTimer.Reset();
+    FineTimer.Reset();
+    InsertTimer.Reset();
+
+    GridStopWatch M1Timer;
+    GridStopWatch M2Timer;
+    GridStopWatch M3Timer;
+    GridStopWatch LinalgTimer;
+    GridStopWatch InnerProdTimer;
+
+    HDCGTimer.Start();
+
+    std::vector<RealD> rn(nrhs);
+    for (int k=0;k<=MaxIterations;k++){
+
+      ////////////////////
+      // Z  = AD
+      ////////////////////
+      M3Timer.Start();
+      for(int b=0;b<nrhs;b++) _FineLinop.HermOp(D[b], Z[b]);      
+      M3Timer.Stop();
+
+      ////////////////////
+      // MZ  = M1 Z <==== the Multigrid preconditioner
+      ////////////////////
+      M1Timer.Start();
+      PcgM1(Z,MZ);
+      M1Timer.Stop();
+
+      FineTimer.Start();
+      ////////////////////
+      // M  = [D^dag Z]^{-1} = (<Ddag MZ>_M)^{-1} inner prod, generalising Saad derivation of Precon CG
+      ////////////////////
+      InnerProdTimer.Start();
+      _BlockCGLinalg.InnerProductMatrix(m_DZ,D,Z);
+      InnerProdTimer.Stop();
+      m_M       = m_DZ.inverse();
+
+      ///////////////////////////
+      // X  = X + D MC
+      ///////////////////////////
+      m_tmp     = m_M * m_C;
+      LinalgTimer.Start();
+      _BlockCGLinalg.MaddMatrix(X,m_tmp, D,X);     // D are the search directions and X takes the updates 
+      LinalgTimer.Stop();
+
+      ///////////////////////////
+      // QS = Q - M Z
+      // (MQ) S = MQ - M (M1Z)
+      ///////////////////////////
+      LinalgTimer.Start();
+      _BlockCGLinalg.MaddMatrix(tmp ,m_M, Z, Q,-1.0);
+      _BlockCGLinalg.MaddMatrix(Mtmp,m_M,MZ,MQ,-1.0);
+      ThinQRfact (m_zz, m_S, m_Sinv, Q, MQ, tmp, Mtmp);
+      LinalgTimer.Stop();
+
+      ////////////////////////////
+      // D  = MQ + D S^dag
+      ////////////////////////////
+      m_tmp = m_S.adjoint();
+      LinalgTimer.Start();
+      _BlockCGLinalg.MaddMatrix(D,m_tmp,D,MQ);
+      LinalgTimer.Stop();
+
+      ////////////////////////////
+      // C  = S C
+      ////////////////////////////
+      m_C = m_S*m_C;
+      
+      ////////////////////////////
+      // convergence monitor
+      ////////////////////////////
+      m_rr = m_C.adjoint() * m_C;
+      
+      FineTimer.Stop();
+
+      RealD max_resid=0;
+      RealD rrsum=0;
+      RealD sssum=0;
+      RealD rr;
+
+      for(int b=0;b<nrhs;b++) {
+	rrsum+=real(m_rr(b,b));
+	sssum+=ssq[b];
+	rr = real(m_rr(b,b))/ssq[b];
+	if ( rr > max_resid ) max_resid = rr;
+      }
+      std::cout << GridLogMessage <<
+	  "\t Prec BlockCGrQ Iteration "<<k<<" ave resid "<< std::sqrt(rrsum/sssum) << " max "<< std::sqrt(max_resid) <<std::endl;
+
+
+      if ( max_resid < Tolerance*Tolerance ) { 
+
+	HDCGTimer.Stop();
+	std::cout<<GridLogMessage<<"HDCG: mrhs PrecBlockCGrQ converged in "<<k<<" iterations and "<<HDCGTimer.Elapsed()<<std::endl;;
+	std::cout<<GridLogMessage<<"HDCG: mrhs PrecBlockCGrQ : Linalg  "<<LinalgTimer.Elapsed()<<std::endl;;
+	std::cout<<GridLogMessage<<"HDCG: mrhs PrecBlockCGrQ : fine H  "<<M3Timer.Elapsed()<<std::endl;;
+	std::cout<<GridLogMessage<<"HDCG: mrhs PrecBlockCGrQ : prec M1 "<<M1Timer.Elapsed()<<std::endl;;
+	std::cout<<GridLogMessage<<"**** M1 breakdown:"<<std::endl;
+	std::cout<<GridLogMessage<<"HDCG: mrhs PrecBlockCGrQ : Project "<<ProjectTimer.Elapsed()<<std::endl;;
+	std::cout<<GridLogMessage<<"HDCG: mrhs PrecBlockCGrQ : Promote "<<PromoteTimer.Elapsed()<<std::endl;;
+	std::cout<<GridLogMessage<<"HDCG: mrhs PrecBlockCGrQ : Deflate "<<DeflateTimer.Elapsed()<<std::endl;;
+	std::cout<<GridLogMessage<<"HDCG: mrhs PrecBlockCGrQ : Coarse  "<<CoarseTimer.Elapsed()<<std::endl;;
+	std::cout<<GridLogMessage<<"HDCG: mrhs PrecBlockCGrQ : Fine    "<<FineTimer.Elapsed()<<std::endl;;
+	std::cout<<GridLogMessage<<"HDCG: mrhs PrecBlockCGrQ : Smooth  "<<SmoothTimer.Elapsed()<<std::endl;;
+	std::cout<<GridLogMessage<<"HDCG: mrhs PrecBlockCGrQ : Insert  "<<InsertTimer.Elapsed()<<std::endl;;
+
+	for(int rhs=0;rhs<nrhs;rhs++){
+
+	  _FineLinop.HermOp(X[rhs],tmp[rhs]);			  
+
+	  Field mytmp(grid);
+	  axpy(mytmp,-1.0,src[rhs],tmp[rhs]);
+      
+	  RealD  xnorm   = sqrt(norm2(X[rhs]));
+	  RealD  srcnorm = sqrt(norm2(src[rhs]));
+	  RealD  tmpnorm = sqrt(norm2(mytmp));
+	  RealD  true_residual = tmpnorm/srcnorm;
+	  std::cout<<GridLogMessage
+		   <<"HDCG: true residual ["<<rhs<<"] is "<<true_residual
+		   <<" solution "<<xnorm
+		   <<" source "<<srcnorm
+		   <<std::endl;
+	}
+	return;
+      }
+      
+    }
+    HDCGTimer.Stop();
+    std::cout<<GridLogMessage<<"HDCG: PrecBlockCGrQ not converged "<<HDCGTimer.Elapsed()<<std::endl;
+    assert(0);
+  }
+
+  virtual void SolveSingleSystem (std::vector<Field> &src, std::vector<Field> &x)
+  {
+    std::cout << GridLogMessage<<"HDCG: mrhs fPcg starting"<<std::endl;
+    src[0].Grid()->Barrier();
+    int nrhs = src.size();
+    std::vector<RealD> f(nrhs);
+    std::vector<RealD> rtzp(nrhs);
+    std::vector<RealD> rtz(nrhs);
+    std::vector<RealD> a(nrhs);
+    std::vector<RealD> d(nrhs);
+    std::vector<RealD> b(nrhs);
+    std::vector<RealD> rptzp(nrhs);
+    /////////////////////////////
+    // Set up history vectors
+    /////////////////////////////
+    int mmax = 3;
+
+    std::vector<std::vector<Field> > p(nrhs);   for(int r=0;r<nrhs;r++)  p[r].resize(mmax,grid);
+    std::vector<std::vector<Field> > mmp(nrhs); for(int r=0;r<nrhs;r++) mmp[r].resize(mmax,grid);
+    std::vector<std::vector<RealD> > pAp(nrhs); for(int r=0;r<nrhs;r++) pAp[r].resize(mmax);
+
+    std::vector<Field> z(nrhs,grid);
+    std::vector<Field>  mp (nrhs,grid);
+    std::vector<Field>  r  (nrhs,grid);
+    std::vector<Field>  mu (nrhs,grid);
+
+    //Initial residual computation & set up
+    std::vector<RealD> src_nrm(nrhs);
+    for(int rhs=0;rhs<nrhs;rhs++) {
+      src_nrm[rhs]=norm2(src[rhs]);
+      assert(src_nrm[rhs]!=0.0);
+    }
+    std::vector<RealD> tn(nrhs);
+
+    GridStopWatch HDCGTimer;
+    //////////////////////////
+    // x0 = Vstart -- possibly modify guess
+    //////////////////////////
+    Vstart(x,src);
+
+    for(int rhs=0;rhs<nrhs;rhs++){
+      // r0 = b -A x0
+      _FineLinop.HermOp(x[rhs],mmp[rhs][0]);
+      axpy (r[rhs], -1.0,mmp[rhs][0], src[rhs]);    // Recomputes r=src-Ax0
+    }
+
+    //////////////////////////////////
+    // Compute z = M1 x
+    //////////////////////////////////
+    // This needs a multiRHS version for acceleration
+    PcgM1(r,z);
+
+    std::vector<RealD> ssq(nrhs);
+    std::vector<RealD> rsq(nrhs);
+    std::vector<Field> pp(nrhs,grid);
+
+    for(int rhs=0;rhs<nrhs;rhs++){
+      rtzp[rhs] =real(innerProduct(r[rhs],z[rhs]));
+      p[rhs][0]=z[rhs];
+      ssq[rhs]=norm2(src[rhs]);
+      rsq[rhs]=  ssq[rhs]*Tolerance*Tolerance;
+      //      std::cout << GridLogMessage<<"mrhs HDCG: "<<rhs<<" k=0 residual "<<rtzp[rhs]<<" rsq "<<rsq[rhs]<<"\n";
+    }
+
+    ProjectTimer.Reset();
+    PromoteTimer.Reset();
+    DeflateTimer.Reset();
+    CoarseTimer.Reset();
+    SmoothTimer.Reset();
+    FineTimer.Reset();
+    InsertTimer.Reset();
+
+    GridStopWatch M1Timer;
+    GridStopWatch M2Timer;
+    GridStopWatch M3Timer;
+    GridStopWatch LinalgTimer;
+
+    HDCGTimer.Start();
+
+    std::vector<RealD> rn(nrhs);
+    for (int k=0;k<=MaxIterations;k++){
+    
+      int peri_k  = k % mmax;
+      int peri_kp = (k+1) % mmax;
+
+      for(int rhs=0;rhs<nrhs;rhs++){
+	rtz[rhs]=rtzp[rhs];
+	M3Timer.Start();
+	d[rhs]= PcgM3(p[rhs][peri_k],mmp[rhs][peri_k]);
+	M3Timer.Stop();
+	a[rhs] = rtz[rhs]/d[rhs];
+
+	LinalgTimer.Start();
+	// Memorise this
+	pAp[rhs][peri_k] = d[rhs];
+
+	axpy(x[rhs],a[rhs],p[rhs][peri_k],x[rhs]);
+	rn[rhs] = axpy_norm(r[rhs],-a[rhs],mmp[rhs][peri_k],r[rhs]);
+	LinalgTimer.Stop();
+      }
+
+      // Compute z = M x (for *all* RHS)
+      M1Timer.Start();
+      PcgM1(r,z);
+      M1Timer.Stop();
+      
+      RealD max_rn=0.0;
+      LinalgTimer.Start();
+      for(int rhs=0;rhs<nrhs;rhs++){
+
+	rtzp[rhs] =real(innerProduct(r[rhs],z[rhs]));
+
+	//	std::cout << GridLogMessage<<"HDCG::fPcg rhs"<<rhs<<" iteration "<<k<<" : inner rtzp "<<rtzp[rhs]<<"\n";
+	mu[rhs]=z[rhs];
+
+	p[rhs][peri_kp]=mu[rhs];
+
+	// Standard search direction p == z + b p 
+	b[rhs] = (rtzp[rhs])/rtz[rhs];
+
+	int northog = (k>mmax-1)?(mmax-1):k;        // This is the fCG-Tr(mmax-1) algorithm
+	for(int back=0; back < northog; back++){
+	  int peri_back = (k-back)%mmax;
+	  RealD pbApk= real(innerProduct(mmp[rhs][peri_back],p[rhs][peri_kp]));
+	  RealD beta = -pbApk/pAp[rhs][peri_back];
+	  axpy(p[rhs][peri_kp],beta,p[rhs][peri_back],p[rhs][peri_kp]);
+	}
+
+	RealD rrn=sqrt(rn[rhs]/ssq[rhs]);
+	RealD rtn=sqrt(rtz[rhs]/ssq[rhs]);
+	RealD rtnp=sqrt(rtzp[rhs]/ssq[rhs]);
+	
+	std::cout<<GridLogMessage<<"HDCG:fPcg rhs "<<rhs<<" k= "<<k<<" residual = "<<rrn<<"\n";
+	if ( rrn > max_rn ) max_rn = rrn;
+      }
+      LinalgTimer.Stop();
+
+      // Stopping condition based on worst case
+      if ( max_rn <= Tolerance ) { 
+
+	HDCGTimer.Stop();
+	std::cout<<GridLogMessage<<"HDCG: mrhs fPcg converged in "<<k<<" iterations and "<<HDCGTimer.Elapsed()<<std::endl;;
+	std::cout<<GridLogMessage<<"HDCG: mrhs fPcg : Linalg  "<<LinalgTimer.Elapsed()<<std::endl;;
+	std::cout<<GridLogMessage<<"HDCG: mrhs fPcg : fine M3 "<<M3Timer.Elapsed()<<std::endl;;
+	std::cout<<GridLogMessage<<"HDCG: mrhs fPcg : prec M1 "<<M1Timer.Elapsed()<<std::endl;;
+	std::cout<<GridLogMessage<<"**** M1 breakdown:"<<std::endl;
+	std::cout<<GridLogMessage<<"HDCG: mrhs fPcg : Project "<<ProjectTimer.Elapsed()<<std::endl;;
+	std::cout<<GridLogMessage<<"HDCG: mrhs fPcg : Promote "<<PromoteTimer.Elapsed()<<std::endl;;
+	std::cout<<GridLogMessage<<"HDCG: mrhs fPcg : Deflate "<<DeflateTimer.Elapsed()<<std::endl;;
+	std::cout<<GridLogMessage<<"HDCG: mrhs fPcg : Coarse  "<<CoarseTimer.Elapsed()<<std::endl;;
+	std::cout<<GridLogMessage<<"HDCG: mrhs fPcg : Fine    "<<FineTimer.Elapsed()<<std::endl;;
+	std::cout<<GridLogMessage<<"HDCG: mrhs fPcg : Smooth  "<<SmoothTimer.Elapsed()<<std::endl;;
+	std::cout<<GridLogMessage<<"HDCG: mrhs fPcg : Insert  "<<InsertTimer.Elapsed()<<std::endl;;
+
+	for(int rhs=0;rhs<nrhs;rhs++){
+	  _FineLinop.HermOp(x[rhs],mmp[rhs][0]);			  
+	  Field tmp(grid);
+	  axpy(tmp,-1.0,src[rhs],mmp[rhs][0]);
+      
+	  RealD  mmpnorm = sqrt(norm2(mmp[rhs][0]));
+	  RealD  xnorm   = sqrt(norm2(x[rhs]));
+	  RealD  srcnorm = sqrt(norm2(src[rhs]));
+	  RealD  tmpnorm = sqrt(norm2(tmp));
+	  RealD  true_residual = tmpnorm/srcnorm;
+	  std::cout<<GridLogMessage
+		   <<"HDCG: true residual ["<<rhs<<"] is "<<true_residual
+		   <<" solution "<<xnorm
+		   <<" source "<<srcnorm
+		   <<" mmp "<<mmpnorm	  
+		   <<std::endl;
+	}
+	return;
+      }
+      
+    }
+    HDCGTimer.Stop();
+    std::cout<<GridLogMessage<<"HDCG: not converged "<<HDCGTimer.Elapsed()<<std::endl;
+    for(int rhs=0;rhs<nrhs;rhs++){
+      RealD  xnorm   = sqrt(norm2(x[rhs]));
+      RealD  srcnorm = sqrt(norm2(src[rhs]));
+      std::cout<<GridLogMessage<<"HDCG: non-converged solution "<<xnorm<<" source "<<srcnorm<<std::endl;
+    }
+  }
+  
+
+ public:
+
+  virtual void PcgM1(std::vector<Field> & in,std::vector<Field> & out) = 0;
+  virtual void Vstart(std::vector<Field> & x,std::vector<Field> & src) = 0;
+  virtual void PcgM2(const Field & in, Field & out) {
+    out=in;
+  }
+
+  virtual RealD PcgM3(const Field & p, Field & mmp){
+    RealD dd;
+    _FineLinop.HermOp(p,mmp);
+    ComplexD dot = innerProduct(p,mmp);
+    dd=real(dot);
+    return dd;
+  }
+
+};
+
+template<class Field, class CoarseField>
+class TwoLevelADEF2mrhs : public TwoLevelCGmrhs<Field>
+{
+public:
+  GridBase *coarsegrid;
+  GridBase *coarsegridmrhs;
+  LinearFunction<CoarseField> &_CoarseSolverMrhs;
+  LinearFunction<CoarseField> &_CoarseSolverPreciseMrhs;
+  MultiRHSBlockProject<Field>    &_Projector;
+  MultiRHSDeflation<CoarseField> &_Deflator;
+
+  
+  TwoLevelADEF2mrhs(RealD tol,
+		    Integer maxit,
+		    LinearOperatorBase<Field>    &FineLinop,
+		    LinearFunction<Field>        &Smoother,
+		    LinearFunction<CoarseField>  &CoarseSolverMrhs,
+		    LinearFunction<CoarseField>  &CoarseSolverPreciseMrhs,
+		    MultiRHSBlockProject<Field>    &Projector,
+		    MultiRHSDeflation<CoarseField> &Deflator,
+		    GridBase *_coarsemrhsgrid) :
+    TwoLevelCGmrhs<Field>(tol, maxit,FineLinop,Smoother,Projector.fine_grid),
+    _CoarseSolverMrhs(CoarseSolverMrhs),
+    _CoarseSolverPreciseMrhs(CoarseSolverPreciseMrhs),
+    _Projector(Projector),
+    _Deflator(Deflator)
+  {
+    coarsegrid = Projector.coarse_grid;
+    coarsegridmrhs = _coarsemrhsgrid;// Thi could be in projector
+  };
+
+  // Override Vstart
+  virtual void Vstart(std::vector<Field> & x,std::vector<Field> & src)
+  {
+    int nrhs=x.size();
+    ///////////////////////////////////
+    // Choose x_0 such that 
+    // x_0 = guess +  (A_ss^inv) r_s = guess + Ass_inv [src -Aguess]
+    //                               = [1 - Ass_inv A] Guess + Assinv src
+    //                               = P^T guess + Assinv src 
+    //                               = Vstart  [Tang notation]
+    // This gives:
+    // W^T (src - A x_0) = src_s - A guess_s - r_s
+    //                   = src_s - (A guess)_s - src_s  + (A guess)_s 
+    //                   = 0 
+    ///////////////////////////////////
+    std::vector<CoarseField> PleftProj(nrhs,this->coarsegrid);
+    std::vector<CoarseField> PleftMss_proj(nrhs,this->coarsegrid);
+    CoarseField PleftProjMrhs(this->coarsegridmrhs);
+    CoarseField PleftMss_projMrhs(this->coarsegridmrhs);
+
+    this->_Projector.blockProject(src,PleftProj);
+    this->_Deflator.DeflateSources(PleftProj,PleftMss_proj);
+    for(int rhs=0;rhs<nrhs;rhs++) {
+      InsertSliceFast(PleftProj[rhs],PleftProjMrhs,rhs,0);
+      InsertSliceFast(PleftMss_proj[rhs],PleftMss_projMrhs,rhs,0); // the guess
+    }
+    
+    this->_CoarseSolverPreciseMrhs(PleftProjMrhs,PleftMss_projMrhs); // Ass^{-1} r_s
+
+    for(int rhs=0;rhs<nrhs;rhs++) {
+      ExtractSliceFast(PleftMss_proj[rhs],PleftMss_projMrhs,rhs,0);
+    }
+    this->_Projector.blockPromote(x,PleftMss_proj);
+  }
+
+  virtual void PcgM1(std::vector<Field> & in,std::vector<Field> & out){
+
+    int nrhs=in.size();
+
+    // [PTM+Q] in = [1 - Q A] M in + Q in = Min + Q [ in -A Min]
+    std::vector<Field> tmp(nrhs,this->grid);
+    std::vector<Field> Min(nrhs,this->grid);
+
+    std::vector<CoarseField> PleftProj(nrhs,this->coarsegrid);
+    std::vector<CoarseField> PleftMss_proj(nrhs,this->coarsegrid);
+
+    CoarseField PleftProjMrhs(this->coarsegridmrhs);
+    CoarseField PleftMss_projMrhs(this->coarsegridmrhs);
+
+#undef SMOOTHER_BLOCK_SOLVE
+#if SMOOTHER_BLOCK_SOLVE
+    this->SmoothTimer.Start();
+    this->_Smoother(in,Min);
+    this->SmoothTimer.Stop();
+#else
+    for(int rhs=0;rhs<nrhs;rhs++) {
+      this->SmoothTimer.Start();
+      this->_Smoother(in[rhs],Min[rhs]);
+      this->SmoothTimer.Stop();
+    }
+#endif
+    
+    for(int rhs=0;rhs<nrhs;rhs++) {
+      
+      this->FineTimer.Start();
+      this->_FineLinop.HermOp(Min[rhs],out[rhs]);
+      axpy(tmp[rhs],-1.0,out[rhs],in[rhs]);          // resid  = in - A Min
+      this->FineTimer.Stop();
+
+    }
+
+    this->ProjectTimer.Start();
+    this->_Projector.blockProject(tmp,PleftProj);
+    this->ProjectTimer.Stop();
+    this->DeflateTimer.Start();
+    this->_Deflator.DeflateSources(PleftProj,PleftMss_proj);
+    this->DeflateTimer.Stop();
+    this->InsertTimer.Start();
+    for(int rhs=0;rhs<nrhs;rhs++) {
+      InsertSliceFast(PleftProj[rhs],PleftProjMrhs,rhs,0);
+      InsertSliceFast(PleftMss_proj[rhs],PleftMss_projMrhs,rhs,0); // the guess
+    }
+    this->InsertTimer.Stop();
+
+    this->CoarseTimer.Start();
+    this->_CoarseSolverMrhs(PleftProjMrhs,PleftMss_projMrhs); // Ass^{-1} [in - A Min]_s
+    this->CoarseTimer.Stop();
+
+    this->InsertTimer.Start();
+    for(int rhs=0;rhs<nrhs;rhs++) {
+      ExtractSliceFast(PleftMss_proj[rhs],PleftMss_projMrhs,rhs,0);
+    }
+    this->InsertTimer.Stop();
+    this->PromoteTimer.Start();
+    this->_Projector.blockPromote(tmp,PleftMss_proj);// tmp= Q[in - A Min]  
+    this->PromoteTimer.Stop();
+    this->FineTimer.Start();
+    for(int rhs=0;rhs<nrhs;rhs++) {
+      axpy(out[rhs],1.0,Min[rhs],tmp[rhs]); // Min+tmp
+    }
+    this->FineTimer.Stop();
+  }
+};
+
+
+NAMESPACE_END(Grid);
+
+

Grid/algorithms/iterative/BlockConjugateGradient.h

@@ -31,6 +31,58 @@ directory
 
 NAMESPACE_BEGIN(Grid);
 
+template<class Field>
+void InnerProductMatrix(Eigen::MatrixXcd &m , const std::vector<Field> &X, const std::vector<Field> &Y){
+  typedef typename Field::scalar_type scomplex;
+  int Nblock = X.size();
+  for(int b=0;b<Nblock;b++){
+  for(int bp=0;bp<Nblock;bp++) {
+    m(b,bp) = innerProduct(X[b],Y[bp]);  
+  }}
+}
+template<class Field>
+void MaddMatrix(std::vector<Field> &AP, Eigen::MatrixXcd &m , const std::vector<Field> &X,const std::vector<Field> &Y,RealD scale=1.0){
+  // Should make this cache friendly with site outermost, parallel_for
+  // Deal with case AP aliases with either Y or X
+  //
+  //Could pack "X" and "AP" into a Nblock x Volume dense array.
+  // AP(Nrhs x vol) = Y(Nrhs x vol) + scale * m(nrhs x nrhs) * X(nrhs*vol)
+  typedef typename Field::scalar_type scomplex;
+  int Nblock = AP.size();
+  std::vector<Field> tmp(Nblock,X[0]);
+  for(int b=0;b<Nblock;b++){
+    tmp[b]   = Y[b];
+    for(int bp=0;bp<Nblock;bp++) {
+      tmp[b] = tmp[b] +scomplex(scale*m(bp,b))*X[bp]; 
+    }
+  }
+  for(int b=0;b<Nblock;b++){
+    AP[b] = tmp[b];
+  }
+}
+template<class Field>
+void MulMatrix(std::vector<Field> &AP, Eigen::MatrixXcd &m , const std::vector<Field> &X){
+  // Should make this cache friendly with site outermost, parallel_for
+  typedef typename Field::scalar_type scomplex;
+  int Nblock = AP.size();
+  for(int b=0;b<Nblock;b++){
+    AP[b] = Zero();
+    for(int bp=0;bp<Nblock;bp++) {
+      AP[b] += scomplex(m(bp,b))*X[bp]; 
+    }
+  }
+}
+template<class Field>
+double normv(const std::vector<Field> &P){
+  int Nblock = P.size();
+  double nn = 0.0;
+  for(int b=0;b<Nblock;b++) {
+    nn+=norm2(P[b]);
+  }
+  return nn;
+}
+
+
 enum BlockCGtype { BlockCG, BlockCGrQ, CGmultiRHS, BlockCGVec, BlockCGrQVec };
 
 //////////////////////////////////////////////////////////////////////////
@@ -87,10 +139,19 @@ void ThinQRfact (Eigen::MatrixXcd &m_rr,
   sliceInnerProductMatrix(m_rr,R,R,Orthog);
 
   // Force manifest hermitian to avoid rounding related
+  /*
+  int rank=m_rr.rows();
+  for(int r=0;r<rank;r++){
+  for(int s=0;s<rank;s++){
+    std::cout << "QR m_rr["<<r<<","<<s<<"] "<<m_rr(r,s)<<std::endl;
+  }}
+  */
   m_rr = 0.5*(m_rr+m_rr.adjoint());
 
   Eigen::MatrixXcd L    = m_rr.llt().matrixL(); 
 
+//  ComplexD det = L.determinant();
+//  std::cout << " Det m_rr "<<det<<std::endl;
   C    = L.adjoint();
   Cinv = C.inverse();
   ////////////////////////////////////////////////////////////////////////////////////////////////////
@@ -110,11 +171,20 @@ void ThinQRfact (Eigen::MatrixXcd &m_rr,
 		 const std::vector<Field> & R)
 {
   InnerProductMatrix(m_rr,R,R);
-
+  /*
+  int rank=m_rr.rows();
+  for(int r=0;r<rank;r++){
+  for(int s=0;s<rank;s++){
+    std::cout << "QRvec m_rr["<<r<<","<<s<<"] "<<m_rr(r,s)<<std::endl;
+  }}
+  */
   m_rr = 0.5*(m_rr+m_rr.adjoint());
 
   Eigen::MatrixXcd L    = m_rr.llt().matrixL(); 
 
+  //  ComplexD det = L.determinant();
+  //  std::cout << " Det m_rr "<<det<<std::endl;
+
   C    = L.adjoint();
   Cinv = C.inverse();
 
@@ -186,6 +256,7 @@ void BlockCGrQsolve(LinearOperatorBase<Field> &Linop, const Field &B, Field &X)
   sliceNorm(ssq,B,Orthog);
   RealD sssum=0;
   for(int b=0;b<Nblock;b++) sssum+=ssq[b];
+  for(int b=0;b<Nblock;b++) std::cout << "src["<<b<<"]" << ssq[b] <<std::endl;
 
   sliceNorm(residuals,B,Orthog);
   for(int b=0;b<Nblock;b++){ assert(std::isnan(residuals[b])==0); }
@@ -221,6 +292,9 @@ void BlockCGrQsolve(LinearOperatorBase<Field> &Linop, const Field &B, Field &X)
   Linop.HermOp(X, AD);
   tmp = B - AD;  
 
+  sliceNorm(residuals,tmp,Orthog);
+  for(int b=0;b<Nblock;b++) std::cout << "res["<<b<<"]" << residuals[b] <<std::endl;
+  
   ThinQRfact (m_rr, m_C, m_Cinv, Q, tmp);
   D=Q;
 
@@ -236,6 +310,8 @@ void BlockCGrQsolve(LinearOperatorBase<Field> &Linop, const Field &B, Field &X)
   GridStopWatch SolverTimer;
   SolverTimer.Start();
 
+  RealD max_resid=0;
+
   int k;
   for (k = 1; k <= MaxIterations; k++) {
 
@@ -280,7 +356,7 @@ void BlockCGrQsolve(LinearOperatorBase<Field> &Linop, const Field &B, Field &X)
      */
     m_rr = m_C.adjoint() * m_C;
 
-    RealD max_resid=0;
+    max_resid=0;
     RealD rrsum=0;
     RealD rr;
 
@@ -322,7 +398,9 @@ void BlockCGrQsolve(LinearOperatorBase<Field> &Linop, const Field &B, Field &X)
     }
 
   }
-  std::cout << GridLogMessage << "BlockConjugateGradient(rQ) did NOT converge" << std::endl;
+
+  std::cout << GridLogMessage << "BlockConjugateGradient(rQ) did NOT converge "<<k<<" / "<<MaxIterations
+	    <<" residual "<< std::sqrt(max_resid)<< std::endl;
 
   if (ErrorOnNoConverge) assert(0);
   IterationsToComplete = k;
@@ -466,43 +544,6 @@ void CGmultiRHSsolve(LinearOperatorBase<Field> &Linop, const Field &Src, Field &
   IterationsToComplete = k;
 }
 
-void InnerProductMatrix(Eigen::MatrixXcd &m , const std::vector<Field> &X, const std::vector<Field> &Y){
-  for(int b=0;b<Nblock;b++){
-  for(int bp=0;bp<Nblock;bp++) {
-    m(b,bp) = innerProduct(X[b],Y[bp]);  
-  }}
-}
-void MaddMatrix(std::vector<Field> &AP, Eigen::MatrixXcd &m , const std::vector<Field> &X,const std::vector<Field> &Y,RealD scale=1.0){
-  // Should make this cache friendly with site outermost, parallel_for
-  // Deal with case AP aliases with either Y or X
-  std::vector<Field> tmp(Nblock,X[0]);
-  for(int b=0;b<Nblock;b++){
-    tmp[b]   = Y[b];
-    for(int bp=0;bp<Nblock;bp++) {
-      tmp[b] = tmp[b] + scomplex(scale*m(bp,b))*X[bp]; 
-    }
-  }
-  for(int b=0;b<Nblock;b++){
-    AP[b] = tmp[b];
-  }
-}
-void MulMatrix(std::vector<Field> &AP, Eigen::MatrixXcd &m , const std::vector<Field> &X){
-  // Should make this cache friendly with site outermost, parallel_for
-  for(int b=0;b<Nblock;b++){
-    AP[b] = Zero();
-    for(int bp=0;bp<Nblock;bp++) {
-      AP[b] += scomplex(m(bp,b))*X[bp]; 
-    }
-  }
-}
-double normv(const std::vector<Field> &P){
-  double nn = 0.0;
-  for(int b=0;b<Nblock;b++) {
-    nn+=norm2(P[b]);
-  }
-  return nn;
-}
-
 ////////////////////////////////////////////////////////////////////////////
 // BlockCGrQvec implementation:
 //--------------------------
@@ -549,6 +590,7 @@ void BlockCGrQsolveVec(LinearOperatorBase<Field> &Linop, const std::vector<Field
 
   RealD sssum=0;
   for(int b=0;b<Nblock;b++){ ssq[b] = norm2(B[b]);}
+  for(int b=0;b<Nblock;b++){ std::cout << "ssq["<<b<<"] "<<ssq[b]<<std::endl;}
   for(int b=0;b<Nblock;b++) sssum+=ssq[b];
 
   for(int b=0;b<Nblock;b++){ residuals[b] = norm2(B[b]);}
@@ -585,6 +627,7 @@ void BlockCGrQsolveVec(LinearOperatorBase<Field> &Linop, const std::vector<Field
   for(int b=0;b<Nblock;b++) {
     Linop.HermOp(X[b], AD[b]);
     tmp[b] = B[b] - AD[b];  
+    std::cout << "r0["<<b<<"] "<<norm2(tmp[b])<<std::endl;
   }
 
   ThinQRfact (m_rr, m_C, m_Cinv, Q, tmp);

Grid/algorithms/iterative/ConjugateGradient.h

@@ -38,6 +38,7 @@ NAMESPACE_BEGIN(Grid);
 // single input vec, single output vec.
 /////////////////////////////////////////////////////////////
 
+
 template <class Field>
 class ConjugateGradient : public OperatorFunction<Field> {
 public:
@@ -54,11 +55,26 @@ public:
   ConjugateGradient(RealD tol, Integer maxit, bool err_on_no_conv = true)
     : Tolerance(tol),
       MaxIterations(maxit),
-      ErrorOnNoConverge(err_on_no_conv){};
+      ErrorOnNoConverge(err_on_no_conv)
+  {};
+
+  virtual void LogIteration(int k,RealD a,RealD b){
+    //    std::cout << "ConjugageGradient::LogIteration() "<<std::endl;
+  };
+  virtual void LogBegin(void){
+    std::cout << "ConjugageGradient::LogBegin() "<<std::endl;
+  };
 
     void operator()(LinearOperatorBase<Field> &Linop, const Field &src, Field &psi) {
 
+      this->LogBegin();
+
       GRID_TRACE("ConjugateGradient");
+    GridStopWatch PreambleTimer;
+    GridStopWatch ConstructTimer;
+    GridStopWatch NormTimer;
+    GridStopWatch AssignTimer;
+    PreambleTimer.Start();
     psi.Checkerboard() = src.Checkerboard();
 
     conformable(psi, src);
@@ -66,22 +82,32 @@ public:
     RealD cp, c, a, d, b, ssq, qq;
     //RealD b_pred;
 
-    Field p(src);
-    Field mmp(src);
-    Field r(src);
+    // Was doing copies
+    ConstructTimer.Start();
+    Field p  (src.Grid());
+    Field mmp(src.Grid());
+    Field r  (src.Grid());
+    ConstructTimer.Stop();
 
     // Initial residual computation & set up
+    NormTimer.Start();
+    ssq = norm2(src);
     RealD guess = norm2(psi);
+    NormTimer.Stop();
     assert(std::isnan(guess) == 0);
-    
+    AssignTimer.Start();
+    if ( guess == 0.0 ) {
+      r = src;
+      p = r;
+      a = ssq;
+    } else { 
       Linop.HermOpAndNorm(psi, mmp, d, b);
-    
       r = src - mmp;
       p = r;
-
       a = norm2(p);
+    }
     cp = a;
-    ssq = norm2(src);
+    AssignTimer.Stop();
 
     // Handle trivial case of zero src
     if (ssq == 0.){
@@ -111,6 +137,7 @@ public:
     std::cout << GridLogIterative << std::setprecision(8)
               << "ConjugateGradient: k=0 residual " << cp << " target " << rsq << std::endl;
 
+    PreambleTimer.Stop();
     GridStopWatch LinalgTimer;
     GridStopWatch InnerTimer;
     GridStopWatch AxpyNormTimer;
@@ -156,6 +183,7 @@ public:
       }
       LinearCombTimer.Stop();
       LinalgTimer.Stop();
+      LogIteration(k,a,b);
 
       IterationTimer.Stop();
       if ( (k % 500) == 0 ) {
@@ -183,13 +211,14 @@ public:
 		  << "\tTrue residual " << true_residual
 		  << "\tTarget " << Tolerance << std::endl;
 
-	std::cout << GridLogMessage << "\tElapsed    " << SolverTimer.Elapsed() <<std::endl;
+	//	std::cout << GridLogMessage << "\tPreamble   " << PreambleTimer.Elapsed() <<std::endl;
+	std::cout << GridLogMessage << "\tSolver Elapsed    " << SolverTimer.Elapsed() <<std::endl;
         std::cout << GridLogPerformance << "Time breakdown "<<std::endl;
 	std::cout << GridLogPerformance << "\tMatrix     " << MatrixTimer.Elapsed() <<std::endl;
 	std::cout << GridLogPerformance << "\tLinalg     " << LinalgTimer.Elapsed() <<std::endl;
-	std::cout << GridLogPerformance << "\tInner      " << InnerTimer.Elapsed() <<std::endl;
-	std::cout << GridLogPerformance << "\tAxpyNorm   " << AxpyNormTimer.Elapsed() <<std::endl;
-	std::cout << GridLogPerformance << "\tLinearComb " << LinearCombTimer.Elapsed() <<std::endl;
+	std::cout << GridLogPerformance << "\t\tInner      " << InnerTimer.Elapsed() <<std::endl;
+	std::cout << GridLogPerformance << "\t\tAxpyNorm   " << AxpyNormTimer.Elapsed() <<std::endl;
+	std::cout << GridLogPerformance << "\t\tLinearComb " << LinearCombTimer.Elapsed() <<std::endl;
 
 	std::cout << GridLogDebug << "\tMobius flop rate " << DwfFlops/ usecs<< " Gflops " <<std::endl;
 
@@ -202,17 +231,143 @@ public:
       }
     }
     // Failed. Calculate true residual before giving up                                                         
-    Linop.HermOpAndNorm(psi, mmp, d, qq);
-    p = mmp - src;
+    // Linop.HermOpAndNorm(psi, mmp, d, qq);
+    //    p = mmp - src;
+    //TrueResidual = sqrt(norm2(p)/ssq);
+    //    TrueResidual = 1;
 
-    TrueResidual = sqrt(norm2(p)/ssq);
-
-    std::cout << GridLogMessage << "ConjugateGradient did NOT converge "<<k<<" / "<< MaxIterations<< std::endl;
+    std::cout << GridLogMessage << "ConjugateGradient did NOT converge "<<k<<" / "<< MaxIterations
+    	      <<" residual "<< std::sqrt(cp / ssq)<< std::endl;
+    SolverTimer.Stop();
+    std::cout << GridLogMessage << "\tPreamble   " << PreambleTimer.Elapsed() <<std::endl;
+    std::cout << GridLogMessage << "\tConstruct  " << ConstructTimer.Elapsed() <<std::endl;
+    std::cout << GridLogMessage << "\tNorm       " << NormTimer.Elapsed() <<std::endl;
+    std::cout << GridLogMessage << "\tAssign     " << AssignTimer.Elapsed() <<std::endl;
+    std::cout << GridLogMessage << "\tSolver     " << SolverTimer.Elapsed() <<std::endl;
+    std::cout << GridLogMessage << "Solver breakdown "<<std::endl;
+    std::cout << GridLogMessage << "\tMatrix     " << MatrixTimer.Elapsed() <<std::endl;
+    std::cout << GridLogMessage<< "\tLinalg     " << LinalgTimer.Elapsed() <<std::endl;
+    std::cout << GridLogPerformance << "\t\tInner      " << InnerTimer.Elapsed() <<std::endl;
+    std::cout << GridLogPerformance << "\t\tAxpyNorm   " << AxpyNormTimer.Elapsed() <<std::endl;
+    std::cout << GridLogPerformance << "\t\tLinearComb " << LinearCombTimer.Elapsed() <<std::endl;
 
     if (ErrorOnNoConverge) assert(0);
     IterationsToComplete = k;
 
   }
 };
+
+
+template <class Field>
+class ConjugateGradientPolynomial : public ConjugateGradient<Field> {
+public:
+  // Optionally record the CG polynomial
+  std::vector<double> ak;
+  std::vector<double> bk;
+  std::vector<double> poly_p;
+  std::vector<double> poly_r;
+  std::vector<double> poly_Ap;
+  std::vector<double> polynomial;
+
+public:
+  ConjugateGradientPolynomial(RealD tol, Integer maxit, bool err_on_no_conv = true)
+    : ConjugateGradient<Field>(tol,maxit,err_on_no_conv)
+  { };
+  void PolyHermOp(LinearOperatorBase<Field> &Linop, const Field &src, Field &psi)
+  {
+    Field tmp(src.Grid());
+    Field AtoN(src.Grid());
+    AtoN = src;
+    psi=AtoN*polynomial[0];
+    for(int n=1;n<polynomial.size();n++){
+      tmp = AtoN;
+      Linop.HermOp(tmp,AtoN);
+      psi = psi + polynomial[n]*AtoN;
+    }
+  }
+  void CGsequenceHermOp(LinearOperatorBase<Field> &Linop, const Field &src, Field &x)
+  {
+    Field Ap(src.Grid());
+    Field r(src.Grid());
+    Field p(src.Grid());
+    p=src;
+    r=src;
+    x=Zero();
+    x.Checkerboard()=src.Checkerboard();
+    for(int k=0;k<ak.size();k++){
+      x = x + ak[k]*p;
+      Linop.HermOp(p,Ap);
+      r = r - ak[k] * Ap;
+      p = r + bk[k] * p;
+    }
+  }
+  void Solve(LinearOperatorBase<Field> &Linop, const Field &src, Field &psi)
+  {
+    psi=Zero();
+    this->operator ()(Linop,src,psi);
+  }
+  virtual void LogBegin(void)
+  {
+    std::cout << "ConjugageGradientPolynomial::LogBegin() "<<std::endl;
+    ak.resize(0);
+    bk.resize(0);
+    polynomial.resize(0);
+    poly_Ap.resize(0);
+    poly_Ap.resize(0);
+    poly_p.resize(1);
+    poly_r.resize(1);
+    poly_p[0]=1.0;
+    poly_r[0]=1.0;
+  };
+  virtual void LogIteration(int k,RealD a,RealD b)
+  {
+    // With zero guess,
+    // p = r = src
+    //
+    // iterate:
+    //   x =  x + a p
+    //   r =  r - a A p
+    //   p =  r + b p
+    //
+    // [0]
+    // r = x
+    // p = x
+    // Ap=0
+    //
+    // [1]
+    // Ap = A x + 0  ==> shift poly P right by 1 and add 0.
+    // x  = x + a p  ==> add polynomials term by term 
+    // r  = r - a A p  ==> add polynomials term by term
+    // p  = r + b p  ==> add polynomials term by term
+    //
+    std::cout << "ConjugageGradientPolynomial::LogIteration() "<<k<<std::endl;
+    ak.push_back(a);
+    bk.push_back(b);
+    //  Ap= right_shift(p)
+    poly_Ap.resize(k+1);
+    poly_Ap[0]=0.0;
+    for(int i=0;i<k;i++){
+      poly_Ap[i+1]=poly_p[i];
+    }
+
+    //  x = x + a p
+    polynomial.resize(k);
+    polynomial[k-1]=0.0;
+    for(int i=0;i<k;i++){
+      polynomial[i] = polynomial[i] + a * poly_p[i];
+    }
+    
+    //  r = r - a Ap
+    //  p = r + b p
+    poly_r.resize(k+1);
+    poly_p.resize(k+1);
+    poly_r[k] = poly_p[k] = 0.0;
+    for(int i=0;i<k+1;i++){
+      poly_r[i] = poly_r[i] - a * poly_Ap[i];
+      poly_p[i] = poly_r[i] + b * poly_p[i];
+    }
+  }
+};
+
 NAMESPACE_END(Grid);
 #endif

Grid/algorithms/iterative/ConjugateGradientMixedPrec.h

@@ -116,14 +116,14 @@ NAMESPACE_BEGIN(Grid);
       //Compute double precision rsd and also new RHS vector.
       Linop_d.HermOp(sol_d, tmp_d);
       RealD norm = axpy_norm(src_d, -1., tmp_d, src_d_in); //src_d is residual vector
-      
+      std::cout<<GridLogMessage<<" rsd norm "<<norm<<std::endl;
       std::cout<<GridLogMessage<<"MixedPrecisionConjugateGradient: Outer iteration " <<outer_iter<<" residual "<< norm<< " target "<< stop<<std::endl;
 
       if(norm < OuterLoopNormMult * stop){
 	std::cout<<GridLogMessage<<"MixedPrecisionConjugateGradient: Outer iteration converged on iteration " <<outer_iter <<std::endl;
 	break;
       }
-      while(norm * inner_tol * inner_tol < stop) inner_tol *= 2;  // inner_tol = sqrt(stop/norm) ??
+      while(norm * inner_tol * inner_tol < stop*1.01) inner_tol *= 2;  // inner_tol = sqrt(stop/norm) ??
 
       PrecChangeTimer.Start();
       precisionChange(src_f, src_d, pc_wk_dp_to_sp);

Grid/algorithms/iterative/ConjugateGradientMultiShift.h

@@ -102,11 +102,11 @@ public:
     assert(mass.size()==nshift);
     assert(mresidual.size()==nshift);
   
-    // dynamic sized arrays on stack; 2d is a pain with vector
-    RealD  bs[nshift];
-    RealD  rsq[nshift];
-    RealD  z[nshift][2];
-    int     converged[nshift];
+    // remove dynamic sized arrays on stack; 2d is a pain with vector
+    std::vector<RealD>  bs(nshift);
+    std::vector<RealD>  rsq(nshift);
+    std::vector<std::array<RealD,2> >  z(nshift);
+    std::vector<int>     converged(nshift);
   
     const int       primary =0;
   
@@ -144,7 +144,7 @@ public:
     for(int s=0;s<nshift;s++){
       rsq[s] = cp * mresidual[s] * mresidual[s];
       std::cout<<GridLogMessage<<"ConjugateGradientMultiShift: shift "<<s
-	       <<" target resid "<<rsq[s]<<std::endl;
+	       <<" target resid^2 "<<rsq[s]<<std::endl;
       ps[s] = src;
     }
     // r and p for primary

Grid/algorithms/iterative/ConjugateGradientMultiShiftCleanup.h

@@ -123,11 +123,11 @@ public:
     assert(mresidual.size()==nshift);
   
     // dynamic sized arrays on stack; 2d is a pain with vector
-    RealD  bs[nshift];
-    RealD  rsq[nshift];
-    RealD  rsqf[nshift];
-    RealD  z[nshift][2];
-    int     converged[nshift];
+    std::vector<RealD>  bs(nshift);
+    std::vector<RealD>  rsq(nshift);
+    std::vector<RealD>  rsqf(nshift);
+    std::vector<std::array<RealD,2> >  z(nshift);
+    std::vector<int>     converged(nshift);
   
     const int       primary =0;
   

Grid/algorithms/iterative/ConjugateGradientMultiShiftMixedPrec.h

@@ -156,11 +156,11 @@ public:
     assert(mresidual.size()==nshift);
   
     // dynamic sized arrays on stack; 2d is a pain with vector
-    RealD  bs[nshift];
-    RealD  rsq[nshift];
-    RealD  rsqf[nshift];
-    RealD  z[nshift][2];
-    int     converged[nshift];
+    std::vector<RealD>  bs(nshift);
+    std::vector<RealD>  rsq(nshift);
+    std::vector<RealD>  rsqf(nshift);
+    std::vector<std::array<RealD,2> >  z(nshift);
+    std::vector<int>     converged(nshift);
   
     const int       primary =0;
   

Grid/algorithms/iterative/ImplicitlyRestartedLanczos.h

@@ -79,14 +79,16 @@ template<class Field> class ImplicitlyRestartedLanczosHermOpTester  : public Imp
     RealD vv = norm2(v) / ::pow(evalMaxApprox,2.0);
 
     std::cout.precision(13);
-    std::cout<<GridLogIRL  << "[" << std::setw(3)<<j<<"] "
-	     <<"eval = "<<std::setw(25)<< eval << " (" << eval_poly << ")"
-	     <<" |H B[i] - eval[i]B[i]|^2 / evalMaxApprox^2 " << std::setw(25) << vv
-	     <<std::endl;
 
     int conv=0;
     if( (vv<eresid*eresid) ) conv = 1;
 
+    std::cout<<GridLogIRL  << "[" << std::setw(3)<<j<<"] "
+	     <<"eval = "<<std::setw(25)<< eval << " (" << eval_poly << ")"
+	     <<" |H B[i] - eval[i]B[i]|^2 / evalMaxApprox^2 " << std::setw(25) << vv
+	     <<" target " << eresid*eresid << " conv " <<conv
+	     <<std::endl;
+
     return conv;
   }
 };

Grid/algorithms/iterative/NormalEquations.h

@@ -60,6 +60,32 @@ public:
   }     
 };
 
+template<class Field> class NormalResidual : public LinearFunction<Field>{
+private:
+  SparseMatrixBase<Field> & _Matrix;
+  OperatorFunction<Field> & _HermitianSolver;
+  LinearFunction<Field>   & _Guess;
+public:
+
+  /////////////////////////////////////////////////////
+  // Wrap the usual normal equations trick
+  /////////////////////////////////////////////////////
+ NormalResidual(SparseMatrixBase<Field> &Matrix, OperatorFunction<Field> &HermitianSolver,
+		 LinearFunction<Field> &Guess) 
+   :  _Matrix(Matrix), _HermitianSolver(HermitianSolver), _Guess(Guess) {}; 
+
+  void operator() (const Field &in, Field &out){
+ 
+    Field res(in.Grid());
+    Field tmp(in.Grid());
+
+    MMdagLinearOperator<SparseMatrixBase<Field>,Field> MMdagOp(_Matrix);
+    _Guess(in,res);
+    _HermitianSolver(MMdagOp,in,res);  // M Mdag res = in ;
+    _Matrix.Mdag(res,out);             // out = Mdag res
+  }     
+};
+
 template<class Field> class HPDSolver : public LinearFunction<Field> {
 private:
   LinearOperatorBase<Field> & _Matrix;

Grid/algorithms/iterative/PowerMethod.h

@@ -20,7 +20,7 @@ template<class Field> class PowerMethod
     RealD evalMaxApprox = 0.0; 
     auto src_n = src; 
     auto tmp = src; 
-    const int _MAX_ITER_EST_ = 100; 
+    const int _MAX_ITER_EST_ = 200; 
 
     for (int i=0;i<_MAX_ITER_EST_;i++) { 
       
@@ -30,18 +30,17 @@ template<class Field> class PowerMethod
       RealD vden = norm2(src_n); 
       RealD na = vnum/vden; 
 
-      std::cout << GridLogIterative << "PowerMethod: Current approximation of largest eigenvalue " << na << std::endl;
+      std::cout << GridLogMessage << "PowerMethod: Current approximation of largest eigenvalue " << na << std::endl;
       
-      if ( (fabs(evalMaxApprox/na - 1.0) < 0.001) || (i==_MAX_ITER_EST_-1) ) { 
- 	evalMaxApprox = na; 
-	std::cout << GridLogMessage << " Approximation of largest eigenvalue: " << evalMaxApprox << std::endl;
- 	return evalMaxApprox; 
-      } 
+      //      if ( (fabs(evalMaxApprox/na - 1.0) < 0.0001) || (i==_MAX_ITER_EST_-1) ) { 
+	// 	evalMaxApprox = na; 
+	// 	return evalMaxApprox; 
+      //      } 
       evalMaxApprox = na; 
       src_n = tmp;
     }
-    assert(0);
-    return 0;
+    std::cout << GridLogMessage << " Approximation of largest eigenvalue: " << evalMaxApprox << std::endl;
+    return evalMaxApprox;
   }
 };
 }

Grid/algorithms/iterative/PowerSpectrum.h

@@ -0,0 +1,76 @@
+#pragma once
+namespace Grid {
+
+class Band
+{
+  RealD lo, hi;
+public:
+  Band(RealD _lo,RealD _hi)
+  {
+    lo=_lo;
+    hi=_hi;
+  }
+  RealD operator() (RealD x){
+    if ( x>lo && x<hi ){
+      return 1.0;
+    } else {
+      return 0.0;
+    }
+  }
+};
+
+class PowerSpectrum
+{ 
+ public: 
+
+  template<typename T>  static RealD normalise(T& v) 
+  {
+    RealD nn = norm2(v);
+    nn = sqrt(nn);
+    v = v * (1.0/nn);
+    return nn;
+  }
+
+  std::vector<RealD> ranges;
+  std::vector<int> order;
+  
+  PowerSpectrum(  std::vector<RealD> &bins, std::vector<int> &_order ) : ranges(bins), order(_order)  { };
+
+  template<class Field>
+  RealD operator()(LinearOperatorBase<Field> &HermOp, const Field &src) 
+  { 
+    GridBase *grid = src.Grid(); 
+    int N=ranges.size();
+    RealD hi = ranges[N-1];
+
+    RealD lo_band = 0.0;
+    RealD hi_band;
+    RealD nn=norm2(src);
+    RealD ss=0.0;
+
+    Field tmp = src;
+
+    for(int b=0;b<N;b++){
+      hi_band = ranges[b];
+      Band Notch(lo_band,hi_band);
+      
+      Chebyshev<Field> polynomial;
+      polynomial.Init(0.0,hi,order[b],Notch);
+      polynomial.JacksonSmooth();
+
+      polynomial(HermOp,src,tmp) ;
+
+      RealD p=norm2(tmp);
+      ss=ss+p;
+      std::cout << GridLogMessage << " PowerSpectrum Band["<<lo_band<<","<<hi_band<<"] power "<<norm2(tmp)/nn<<std::endl;
+      
+      lo_band=hi_band;
+    }
+    std::cout << GridLogMessage << " PowerSpectrum total power "<<ss/nn<<std::endl;
+    std::cout << GridLogMessage << " PowerSpectrum total power (unnormalised) "<<nn<<std::endl;
+
+    return 0;
+  };
+};
+  
+}

Grid/algorithms/iterative/SchurRedBlack.h

@@ -499,6 +499,87 @@ namespace Grid {
       }
   };
 
+  ///////////////////////////////////////////////////////////////////////////////////////////////////////
+  // Site diagonal is identity, left preconditioned by Mee^inv
+  // ( 1 - Mee^inv Meo Moo^inv Moe ) phi = Mee_inv ( Mee - Meo Moo^inv Moe Mee^inv  ) phi =  Mee_inv eta
+  //
+  // Solve:
+  // ( 1 - Mee^inv Meo Moo^inv Moe )^dag ( 1 - Mee^inv Meo Moo^inv Moe ) phi = ( 1 - Mee^inv Meo Moo^inv Moe )^dag  Mee_inv eta
+  //
+  // Old notation e<->o
+  //
+  // Left precon by Moo^-1
+  //  b) (Doo^{dag} M_oo^-dag) (Moo^-1 Doo) psi_o =  [ (D_oo)^dag M_oo^-dag ] Moo^-1 L^{-1}  eta_o
+  //                                   eta_o'     = (D_oo)^dag  M_oo^-dag Moo^-1 (eta_o - Moe Mee^{-1} eta_e)
+  ///////////////////////////////////////////////////////////////////////////////////////////////////////
+  template<class Field> class SchurRedBlackDiagOneSolve : public SchurRedBlackBase<Field> {
+  public:
+    typedef CheckerBoardedSparseMatrixBase<Field> Matrix;
+
+    /////////////////////////////////////////////////////
+    // Wrap the usual normal equations Schur trick
+    /////////////////////////////////////////////////////
+  SchurRedBlackDiagOneSolve(OperatorFunction<Field> &HermitianRBSolver, const bool initSubGuess = false,
+      const bool _solnAsInitGuess = false)  
+    : SchurRedBlackBase<Field>(HermitianRBSolver,initSubGuess,_solnAsInitGuess) {};
+
+    virtual void RedBlackSource(Matrix & _Matrix,const Field &src, Field &src_e,Field &src_o)
+    {
+      GridBase *grid = _Matrix.RedBlackGrid();
+      GridBase *fgrid= _Matrix.Grid();
+
+      SchurDiagOneOperator<Matrix,Field> _HermOpEO(_Matrix);
+      
+      Field   tmp(grid);
+      Field  Mtmp(grid);
+
+      pickCheckerboard(Even,src_e,src);
+      pickCheckerboard(Odd ,src_o,src);
+    
+      /////////////////////////////////////////////////////
+      // src_o = Mpcdag *MooeeInv * (source_o - Moe MeeInv source_e)
+      /////////////////////////////////////////////////////
+      _Matrix.MooeeInv(src_e,tmp);     assert(  tmp.Checkerboard() ==Even);
+      _Matrix.Meooe   (tmp,Mtmp);      assert( Mtmp.Checkerboard() ==Odd);     
+      Mtmp=src_o-Mtmp;                 
+      _Matrix.MooeeInv(Mtmp,tmp);      assert( tmp.Checkerboard() ==Odd);     
+      
+      // get the right MpcDag
+      _HermOpEO.MpcDag(tmp,src_o);     assert(src_o.Checkerboard() ==Odd);       
+    }
+
+    virtual void RedBlackSolution(Matrix & _Matrix,const Field &sol_o, const Field &src_e,Field &sol)
+    {
+      GridBase *grid = _Matrix.RedBlackGrid();
+      GridBase *fgrid= _Matrix.Grid();
+
+      Field   tmp(grid);
+      Field   sol_e(grid);
+
+
+      ///////////////////////////////////////////////////
+      // sol_e = M_ee^-1 * ( src_e - Meo sol_o )...
+      ///////////////////////////////////////////////////
+      _Matrix.Meooe(sol_o,tmp);    assert(  tmp.Checkerboard()   ==Even);
+      tmp = src_e-tmp;             assert(  src_e.Checkerboard() ==Even);
+      _Matrix.MooeeInv(tmp,sol_e); assert(  sol_e.Checkerboard() ==Even);
+     
+      setCheckerboard(sol,sol_e);  assert(  sol_e.Checkerboard() ==Even);
+      setCheckerboard(sol,sol_o);  assert(  sol_o.Checkerboard() ==Odd );
+    };
+
+    virtual void RedBlackSolve   (Matrix & _Matrix,const Field &src_o, Field &sol_o)
+    {
+      SchurDiagOneOperator<Matrix,Field> _HermOpEO(_Matrix);
+      this->_HermitianRBSolver(_HermOpEO,src_o,sol_o);
+    };
+    virtual void RedBlackSolve   (Matrix & _Matrix,const std::vector<Field> &src_o,  std::vector<Field> &sol_o)
+    {
+      SchurDiagOneOperator<Matrix,Field> _HermOpEO(_Matrix);
+      this->_HermitianRBSolver(_HermOpEO,src_o,sol_o); 
+    }
+  };
+
   ///////////////////////////////////////////////////////////////////////////////////////////////////////
   // Site diagonal is identity, right preconditioned by Mee^inv
   // ( 1 - Meo Moo^inv Moe Mee^inv  ) phi =( 1 - Meo Moo^inv Moe Mee^inv  ) Mee psi =  = eta  = eta

Grid/algorithms/multigrid/Aggregates.h

@@ -32,9 +32,18 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 
 NAMESPACE_BEGIN(Grid);
 
+inline RealD AggregatePowerLaw(RealD x)
+{
+  //  return std::pow(x,-4);
+  //  return std::pow(x,-3);
+  return std::pow(x,-5);
+}
+
 template<class Fobj,class CComplex,int nbasis>
 class Aggregation {
 public:
+  constexpr int Nbasis(void) { return nbasis; };
+  
   typedef iVector<CComplex,nbasis >             siteVector;
   typedef Lattice<siteVector>                 CoarseVector;
   typedef Lattice<iMatrix<CComplex,nbasis > > CoarseMatrix;
@@ -246,9 +255,54 @@ public:
       // Initial matrix element
       hermop.Op(noise,Mn);
       if(b==0) std::cout<<GridLogMessage << "noise <n|MdagM|n> "<<norm2(Mn)<<std::endl;
+
       // Filter
       Chebyshev<FineField> Cheb(lo,hi,orderfilter);
       Cheb(hermop,noise,Mn);
+      scale = std::pow(norm2(Mn),-0.5); 	Mn=Mn*scale;
+
+      // Refine
+      Chebyshev<FineField> PowerLaw(lo,hi,1000,AggregatePowerLaw);
+      noise = Mn;
+      PowerLaw(hermop,noise,Mn);
+      scale = std::pow(norm2(Mn),-0.5); 	Mn=Mn*scale;
+
+      // normalise
+      subspace[b]   = Mn;
+      hermop.Op(Mn,tmp); 
+      std::cout<<GridLogMessage << "filt ["<<b<<"] <n|MdagM|n> "<<norm2(tmp)<<std::endl;
+    }
+
+  }
+
+  virtual void CreateSubspaceChebyshevPowerLaw(GridParallelRNG  &RNG,LinearOperatorBase<FineField> &hermop,
+					       int nn,
+					       double hi,
+					       int orderfilter
+					       ) {
+
+    RealD scale;
+
+    FineField noise(FineGrid);
+    FineField Mn(FineGrid);
+    FineField tmp(FineGrid);
+
+    // New normalised noise
+    std::cout << GridLogMessage<<" Chebyshev subspace pure noise : ord "<<orderfilter<<" [0,"<<hi<<"]"<<std::endl;
+    std::cout << GridLogMessage<<" Chebyshev subspace pure noise  : nbasis "<<nn<<std::endl;
+
+    for(int b =0;b<nbasis;b++)
+    {
+      gaussian(RNG,noise);
+      scale = std::pow(norm2(noise),-0.5); 
+      noise=noise*scale;
+
+      // Initial matrix element
+      hermop.Op(noise,Mn);
+      if(b==0) std::cout<<GridLogMessage << "noise <n|MdagM|n> "<<norm2(Mn)<<std::endl;
+      // Filter
+      Chebyshev<FineField> Cheb(0.0,hi,orderfilter,AggregatePowerLaw);
+      Cheb(hermop,noise,Mn);
       // normalise
       scale = std::pow(norm2(Mn),-0.5); 	Mn=Mn*scale;
       subspace[b]   = Mn;
@@ -257,6 +311,168 @@ public:
     }
 
   }
+  virtual void CreateSubspaceChebyshevNew(GridParallelRNG  &RNG,LinearOperatorBase<FineField> &hermop,
+					  double hi
+					  ) {
+
+    RealD scale;
+
+    FineField noise(FineGrid);
+    FineField Mn(FineGrid);
+    FineField tmp(FineGrid);
+
+    // New normalised noise
+    for(int b =0;b<nbasis;b++)
+    {
+      gaussian(RNG,noise);
+      scale = std::pow(norm2(noise),-0.5); 
+      noise=noise*scale;
+
+      // Initial matrix element
+      hermop.Op(noise,Mn);
+      if(b==0) std::cout<<GridLogMessage << "noise <n|MdagM|n> "<<norm2(Mn)<<std::endl;
+      // Filter
+      //#opt2(x) =  acheb(x,3,90,300)* acheb(x,1,90,50) * acheb(x,0.5,90,200) * acheb(x,0.05,90,400) * acheb(x,0.01,90,1500)
+      /*266
+      Chebyshev<FineField> Cheb1(3.0,hi,300);
+      Chebyshev<FineField> Cheb2(1.0,hi,50);
+      Chebyshev<FineField> Cheb3(0.5,hi,300);
+      Chebyshev<FineField> Cheb4(0.05,hi,500);
+      Chebyshev<FineField> Cheb5(0.01,hi,2000);
+      */
+      /* 242 */
+      /*
+      Chebyshev<FineField> Cheb3(0.1,hi,300);
+      Chebyshev<FineField> Cheb2(0.02,hi,1000);
+      Chebyshev<FineField> Cheb1(0.003,hi,2000);
+      8?
+      */
+      /* How many??
+      */
+      Chebyshev<FineField> Cheb2(0.001,hi,2500); // 169 iters on HDCG after refine
+      Chebyshev<FineField> Cheb1(0.02,hi,600);
+
+      //      Chebyshev<FineField> Cheb2(0.001,hi,1500);
+      //      Chebyshev<FineField> Cheb1(0.02,hi,600);
+      Cheb1(hermop,noise,Mn); scale = std::pow(norm2(Mn),-0.5); 	noise=Mn*scale;
+      hermop.Op(noise,tmp); std::cout<<GridLogMessage << "Cheb1 <n|MdagM|n> "<<norm2(tmp)<<std::endl;
+      Cheb2(hermop,noise,Mn); scale = std::pow(norm2(Mn),-0.5); 	noise=Mn*scale;
+      hermop.Op(noise,tmp); std::cout<<GridLogMessage << "Cheb2 <n|MdagM|n> "<<norm2(tmp)<<std::endl;
+      //      Cheb3(hermop,noise,Mn); scale = std::pow(norm2(Mn),-0.5); 	noise=Mn*scale;
+      //      hermop.Op(noise,tmp); std::cout<<GridLogMessage << "Cheb3 <n|MdagM|n> "<<norm2(tmp)<<std::endl;
+      //      Cheb4(hermop,noise,Mn); scale = std::pow(norm2(Mn),-0.5); 	noise=Mn*scale;
+      //      hermop.Op(noise,tmp); std::cout<<GridLogMessage << "Cheb4 <n|MdagM|n> "<<norm2(tmp)<<std::endl;
+      //      Cheb5(hermop,noise,Mn); scale = std::pow(norm2(Mn),-0.5); 	noise=Mn*scale;
+      //      hermop.Op(noise,tmp); std::cout<<GridLogMessage << "Cheb5 <n|MdagM|n> "<<norm2(tmp)<<std::endl;
+      subspace[b]   = noise;
+      hermop.Op(subspace[b],tmp); 
+      std::cout<<GridLogMessage << "filt ["<<b<<"] <n|MdagM|n> "<<norm2(tmp)<< " norm " << norm2(noise)<<std::endl;
+    }
+
+  }
+
+  virtual void CreateSubspaceMultishift(GridParallelRNG  &RNG,LinearOperatorBase<FineField> &hermop,
+					double Lo,double tol,int maxit)
+  {
+
+    RealD scale;
+
+    FineField noise(FineGrid);
+    FineField Mn(FineGrid);
+    FineField tmp(FineGrid);
+
+    // New normalised noise
+    std::cout << GridLogMessage<<" Multishift subspace : Lo "<<Lo<<std::endl;
+
+    // Filter
+    // [ 1/6(x+Lo)  - 1/2(x+2Lo) + 1/2(x+3Lo)  -1/6(x+4Lo) = Lo^3 /[ (x+1Lo)(x+2Lo)(x+3Lo)(x+4Lo) ]
+    //
+    // 1/(x+Lo)  - 1/(x+2 Lo)
+    double epsilon      = Lo/3;
+    std::vector<RealD> alpha({1.0/6.0,-1.0/2.0,1.0/2.0,-1.0/6.0});
+    std::vector<RealD> shifts({Lo,Lo+epsilon,Lo+2*epsilon,Lo+3*epsilon});
+    std::vector<RealD> tols({tol,tol,tol,tol});
+    std::cout << "sizes "<<alpha.size()<<" "<<shifts.size()<<" "<<tols.size()<<std::endl;
+
+    MultiShiftFunction msf(4,0.0,95.0);
+    std::cout << "msf constructed "<<std::endl;
+    msf.poles=shifts;
+    msf.residues=alpha;
+    msf.tolerances=tols;
+    msf.norm=0.0;
+    msf.order=alpha.size();
+    ConjugateGradientMultiShift<FineField> MSCG(maxit,msf);
+    
+    for(int b =0;b<nbasis;b++)
+    {
+      gaussian(RNG,noise);
+      scale = std::pow(norm2(noise),-0.5); 
+      noise=noise*scale;
+
+      // Initial matrix element
+      hermop.Op(noise,Mn);
+      if(b==0) std::cout<<GridLogMessage << "noise <n|MdagM|n> "<<norm2(Mn)<<std::endl;
+
+      MSCG(hermop,noise,Mn);
+      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;
+
+    }
+
+  }
+  virtual void RefineSubspace(LinearOperatorBase<FineField> &hermop,
+			      double Lo,double tol,int maxit)
+  {
+    FineField tmp(FineGrid);
+    for(int b =0;b<nbasis;b++)
+    {
+      ConjugateGradient<FineField>  CGsloppy(tol,maxit,false);
+      ShiftedHermOpLinearOperator<FineField> ShiftedFineHermOp(hermop,Lo);
+      tmp=Zero();
+      CGsloppy(hermop,subspace[b],tmp);
+      RealD scale = std::pow(norm2(tmp),-0.5); 	tmp=tmp*scale;
+      subspace[b]=tmp;
+      hermop.Op(subspace[b],tmp);
+      std::cout<<GridLogMessage << "filt ["<<b<<"] <n|MdagM|n> "<<norm2(tmp)<<std::endl;
+    }
+  }
+  virtual void RefineSubspaceHDCG(LinearOperatorBase<FineField> &hermop,
+				  TwoLevelADEF2mrhs<FineField,CoarseVector> & theHDCG,
+				  int nrhs)
+  {
+    std::vector<FineField> src_mrhs(nrhs,FineGrid);
+    std::vector<FineField> res_mrhs(nrhs,FineGrid);
+    FineField tmp(FineGrid);
+    for(int b =0;b<nbasis;b+=nrhs)
+    {
+      tmp = subspace[b];
+      RealD scale = std::pow(norm2(tmp),-0.5); 	tmp=tmp*scale;
+      subspace[b] =tmp;
+      hermop.Op(subspace[b],tmp);
+      std::cout<<GridLogMessage << "before filt ["<<b<<"] <n|MdagM|n> "<<norm2(tmp)<<std::endl;
+
+      for(int r=0;r<MIN(nbasis-b,nrhs);r++){
+	src_mrhs[r] = subspace[b+r];
+      }
+      for(int r=0;r<nrhs;r++){
+	res_mrhs[r] = Zero();
+      }
+      theHDCG(src_mrhs,res_mrhs);
+
+      for(int r=0;r<MIN(nbasis-b,nrhs);r++){
+	tmp = res_mrhs[r];
+	RealD scale = std::pow(norm2(tmp),-0.5); tmp=tmp*scale;
+	subspace[b+r]=tmp;
+      }
+      hermop.Op(subspace[b],tmp);
+      std::cout<<GridLogMessage << "after filt ["<<b<<"] <n|MdagM|n> "<<norm2(tmp)<<std::endl;
+    }
+  }
+
+  
   
 };
 NAMESPACE_END(Grid);
+

Grid/algorithms/multigrid/CoarsenedMatrix.h

@@ -99,7 +99,7 @@ public:
   CoarseMatrix AselfInvEven;
   CoarseMatrix AselfInvOdd;
 
-  Vector<RealD> dag_factor;
+  deviceVector<RealD> dag_factor;
 
   ///////////////////////
   // Interface
@@ -124,9 +124,13 @@ public:
     int npoint = geom.npoint;
     typedef LatticeView<Cobj> Aview;
       
-    Vector<Aview> AcceleratorViewContainer;
+    deviceVector<Aview> AcceleratorViewContainer(geom.npoint);
+    hostVector<Aview>   hAcceleratorViewContainer(geom.npoint);
   
-    for(int p=0;p<geom.npoint;p++) AcceleratorViewContainer.push_back(A[p].View(AcceleratorRead));
+    for(int p=0;p<geom.npoint;p++) {
+      hAcceleratorViewContainer[p] = A[p].View(AcceleratorRead);
+      acceleratorPut(AcceleratorViewContainer[p],hAcceleratorViewContainer[p]);
+    }
     Aview *Aview_p = & AcceleratorViewContainer[0];
 
     const int Nsimd = CComplex::Nsimd();
@@ -161,7 +165,7 @@ public:
       coalescedWrite(out_v[ss](b),res);
       });
 
-    for(int p=0;p<geom.npoint;p++) AcceleratorViewContainer[p].ViewClose();
+    for(int p=0;p<geom.npoint;p++) hAcceleratorViewContainer[p].ViewClose();
   };
 
   void Mdag (const CoarseVector &in, CoarseVector &out)
@@ -190,9 +194,14 @@ public:
     int npoint = geom.npoint;
     typedef LatticeView<Cobj> Aview;
 
-    Vector<Aview> AcceleratorViewContainer;
 
-    for(int p=0;p<geom.npoint;p++) AcceleratorViewContainer.push_back(A[p].View(AcceleratorRead));
+    deviceVector<Aview> AcceleratorViewContainer(geom.npoint);
+    hostVector<Aview>   hAcceleratorViewContainer(geom.npoint);
+  
+    for(int p=0;p<geom.npoint;p++) {
+      hAcceleratorViewContainer[p] = A[p].View(AcceleratorRead);
+      acceleratorPut(AcceleratorViewContainer[p],hAcceleratorViewContainer[p]);
+    }
     Aview *Aview_p = & AcceleratorViewContainer[0];
 
     const int Nsimd = CComplex::Nsimd();
@@ -201,10 +210,10 @@ public:
 
     int osites=Grid()->oSites();
 
-    Vector<int> points(geom.npoint, 0);
-    for(int p=0; p<geom.npoint; p++)
-      points[p] = geom.points_dagger[p];
-
+    deviceVector<int> points(geom.npoint);
+    for(int p=0; p<geom.npoint; p++) { 
+      acceleratorPut(points[p],geom.points_dagger[p]);
+    }
     auto points_p = &points[0];
 
     RealD* dag_factor_p = &dag_factor[0];
@@ -236,7 +245,7 @@ public:
       coalescedWrite(out_v[ss](b),res);
       });
 
-    for(int p=0;p<geom.npoint;p++) AcceleratorViewContainer[p].ViewClose();
+    for(int p=0;p<geom.npoint;p++) hAcceleratorViewContainer[p].ViewClose();
   }
 
   void MdirComms(const CoarseVector &in)
@@ -251,8 +260,14 @@ public:
     out.Checkerboard() = in.Checkerboard();
 
     typedef LatticeView<Cobj> Aview;
-    Vector<Aview> AcceleratorViewContainer;
-    for(int p=0;p<geom.npoint;p++) AcceleratorViewContainer.push_back(A[p].View(AcceleratorRead));
+
+    deviceVector<Aview> AcceleratorViewContainer(geom.npoint);
+    hostVector<Aview>   hAcceleratorViewContainer(geom.npoint);
+  
+    for(int p=0;p<geom.npoint;p++) {
+      hAcceleratorViewContainer[p] = A[p].View(AcceleratorRead);
+      acceleratorPut(AcceleratorViewContainer[p],hAcceleratorViewContainer[p]);
+    }
     Aview *Aview_p = & AcceleratorViewContainer[0];
 
     autoView( out_v , out, AcceleratorWrite);
@@ -285,7 +300,7 @@ public:
       }
       coalescedWrite(out_v[ss](b),res);
     });
-    for(int p=0;p<geom.npoint;p++) AcceleratorViewContainer[p].ViewClose();
+    for(int p=0;p<geom.npoint;p++) hAcceleratorViewContainer[p].ViewClose();
   }
   void MdirAll(const CoarseVector &in,std::vector<CoarseVector> &out)
   {
@@ -469,14 +484,20 @@ public:
 
     // determine in what order we need the points
     int npoint = geom.npoint-1;
-    Vector<int> points(npoint, 0);
-    for(int p=0; p<npoint; p++)
-      points[p] = (dag && !hermitian) ? geom.points_dagger[p] : p;
-
+    deviceVector<int> points(npoint);
+    for(int p=0; p<npoint; p++) {
+      int val = (dag && !hermitian) ? geom.points_dagger[p] : p;
+      acceleratorPut(points[p], val);
+    }
     auto points_p = &points[0];
 
-    Vector<Aview> AcceleratorViewContainer;
-    for(int p=0;p<npoint;p++) AcceleratorViewContainer.push_back(a[p].View(AcceleratorRead));
+    deviceVector<Aview> AcceleratorViewContainer(geom.npoint);
+    hostVector<Aview>   hAcceleratorViewContainer(geom.npoint);
+  
+    for(int p=0;p<geom.npoint;p++) {
+      hAcceleratorViewContainer[p] = a[p].View(AcceleratorRead);
+      acceleratorPut(AcceleratorViewContainer[p],hAcceleratorViewContainer[p]);
+    }
     Aview *Aview_p = & AcceleratorViewContainer[0];
 
     const int Nsimd = CComplex::Nsimd();
@@ -539,7 +560,7 @@ public:
       });
     }
 
-    for(int p=0;p<npoint;p++) AcceleratorViewContainer[p].ViewClose();
+    for(int p=0;p<npoint;p++) hAcceleratorViewContainer[p].ViewClose();
   }
   
   CoarsenedMatrix(GridCartesian &CoarseGrid, int hermitian_=0) 	:
@@ -590,11 +611,13 @@ public:
     }
 
     // GPU readable prefactor
+    std::vector<RealD> h_dag_factor(nbasis*nbasis);
     thread_for(i, nbasis*nbasis, {
       int j = i/nbasis;
       int k = i%nbasis;
-      dag_factor[i] = dag_factor_eigen(j, k);
+      h_dag_factor[i] = dag_factor_eigen(j, k);
     });
+    acceleratorCopyToDevice(&h_dag_factor[0],&dag_factor[0],dag_factor.size()*sizeof(RealD));
   }
 
   void CoarsenOperator(GridBase *FineGrid,LinearOperatorBase<Lattice<Fobj> > &linop,

Grid/algorithms/multigrid/GeneralCoarsenedMatrix.h

@@ -47,8 +47,10 @@ public:
   typedef Lattice<siteVector>                 CoarseVector;
   typedef Lattice<iMatrix<CComplex,nbasis > > CoarseMatrix;
   typedef iMatrix<CComplex,nbasis >  Cobj;
+  typedef iVector<CComplex,nbasis >  Cvec;
   typedef Lattice< CComplex >   CoarseScalar; // used for inner products on fine field
   typedef Lattice<Fobj >        FineField;
+  typedef Lattice<CComplex >    FineComplexField;
   typedef CoarseVector Field;
   ////////////////////
   // Data members
@@ -62,14 +64,26 @@ public:
   
   std::vector<CoarseMatrix> _A;
   std::vector<CoarseMatrix> _Adag;
+  std::vector<CoarseVector> MultTemporaries;
 
   ///////////////////////
   // Interface
   ///////////////////////
-  GridBase      * Grid(void)           { return _FineGrid; };   // this is all the linalg routines need to know
+  GridBase      * Grid(void)           { return _CoarseGrid; };   // this is all the linalg routines need to know
   GridBase      * FineGrid(void)       { return _FineGrid; };   // this is all the linalg routines need to know
   GridCartesian * CoarseGrid(void)     { return _CoarseGrid; };   // this is all the linalg routines need to know
 
+  /*  void ShiftMatrix(RealD shift)
+  {
+    int Nd=_FineGrid->Nd(); 
+    Coordinate zero_shift(Nd,0);
+    for(int p=0;p<geom.npoint;p++){
+      if ( zero_shift==geom.shifts[p] ) {
+	_A[p] = _A[p]+shift;
+	//	_Adag[p] = _Adag[p]+shift;
+      }
+    }    
+  }
   void ProjectNearestNeighbour(RealD shift, GeneralCoarseOp &CopyMe)
   {
     int nfound=0;
@@ -80,7 +94,7 @@ public:
 	// Avoids brutal handling of Grid pointers
 	if ( CopyMe.geom.shifts[pp]==geom.shifts[p] ) {
 	  _A[p] = CopyMe.Cell.Extract(CopyMe._A[pp]);
-	  _Adag[p] = CopyMe.Cell.Extract(CopyMe._Adag[pp]);
+	  //	  _Adag[p] = CopyMe.Cell.Extract(CopyMe._Adag[pp]);
 	  nfound++;
 	}
       }
@@ -88,6 +102,7 @@ public:
     assert(nfound==geom.npoint);
     ExchangeCoarseLinks();
   }
+  */
   
   GeneralCoarsenedMatrix(NonLocalStencilGeometry &_geom,GridBase *FineGrid, GridCartesian * CoarseGrid)
     : geom(_geom),
@@ -99,19 +114,9 @@ public:
   {
     {
       int npoint = _geom.npoint;
-      autoView( Stencil_v  , Stencil, AcceleratorRead);
-      int osites=Stencil.Grid()->oSites();
-      for(int ss=0;ss<osites;ss++){
-	for(int point=0;point<npoint;point++){
-	  auto SE = Stencil_v.GetEntry(point,ss);
-	  int o = SE->_offset;
-	  assert( o< osites);
     }
-      }    
-    }
-
     _A.resize(geom.npoint,CoarseGrid);
-    _Adag.resize(geom.npoint,CoarseGrid);
+    //    _Adag.resize(geom.npoint,CoarseGrid);
   }
   void M (const CoarseVector &in, CoarseVector &out)
   {
@@ -119,16 +124,16 @@ public:
   }
   void Mdag (const CoarseVector &in, CoarseVector &out)
   {
-    if ( hermitian ) M(in,out);
-    else Mult(_Adag,in,out);
+    assert(hermitian);
+    Mult(_A,in,out);
+    //    if ( hermitian ) M(in,out);
+    //    else Mult(_Adag,in,out);
   }
   void Mult (std::vector<CoarseMatrix> &A,const CoarseVector &in, CoarseVector &out)
   {
-    RealD tviews=0;
-    RealD ttot=0;
-    RealD tmult=0;
-    RealD texch=0;
-    RealD text=0;
+    RealD tviews=0;    RealD ttot=0;    RealD tmult=0;   RealD texch=0;    RealD text=0; RealD ttemps=0; RealD tcopy=0;
+    RealD tmult2=0;
+
     ttot=-usecond();
     conformable(CoarseGrid(),in.Grid());
     conformable(in.Grid(),out.Grid());
@@ -136,71 +141,106 @@ public:
     CoarseVector tin=in;
 
     texch-=usecond();
-    CoarseVector pin  = Cell.Exchange(tin);
+    CoarseVector pin = Cell.ExchangePeriodic(tin);
     texch+=usecond();
 
-    CoarseVector pout(pin.Grid()); pout=Zero();
+    CoarseVector pout(pin.Grid());
 
     int npoint = geom.npoint;
     typedef LatticeView<Cobj> Aview;
+    typedef LatticeView<Cvec> Vview;
       
     const int Nsimd = CComplex::Nsimd();
     
-    int osites=pin.Grid()->oSites();
-    //    int gsites=pin.Grid()->gSites();
+    int64_t osites=pin.Grid()->oSites();
 
-    RealD flops = 1.0* npoint * nbasis * nbasis * 8 * osites;
-    RealD bytes = (1.0*osites*sizeof(siteMatrix)*npoint+2.0*osites*sizeof(siteVector))*npoint;
-      
-    //    for(int point=0;point<npoint;point++){
-    //      conformable(A[point],pin);
-    //    }
+    RealD flops = 1.0* npoint * nbasis * nbasis * 8.0 * osites * CComplex::Nsimd();
+    RealD bytes = 1.0*osites*sizeof(siteMatrix)*npoint
+                + 2.0*osites*sizeof(siteVector)*npoint;
       
     {
       tviews-=usecond();
       autoView( in_v , pin, AcceleratorRead);
-      autoView( out_v , pout, AcceleratorWrite);
+      autoView( out_v , pout, AcceleratorWriteDiscard);
       autoView( Stencil_v  , Stencil, AcceleratorRead);
       tviews+=usecond();
 
-      for(int point=0;point<npoint;point++){
+      // Static and prereserve to keep UVM region live and not resized across multiple calls
+      ttemps-=usecond();
+      MultTemporaries.resize(npoint,pin.Grid());       
+      ttemps+=usecond();
+      std::vector<Aview> AcceleratorViewContainer_h;
+      std::vector<Vview> AcceleratorVecViewContainer_h; 
+
       tviews-=usecond();
-	autoView( A_v, A[point],AcceleratorRead);
+      for(int p=0;p<npoint;p++) {
+	AcceleratorViewContainer_h.push_back(      A[p].View(AcceleratorRead));
+	AcceleratorVecViewContainer_h.push_back(MultTemporaries[p].View(AcceleratorWrite));
+      }
       tviews+=usecond();
+
+      static deviceVector<Aview> AcceleratorViewContainer; AcceleratorViewContainer.resize(npoint);
+      static deviceVector<Vview> AcceleratorVecViewContainer; AcceleratorVecViewContainer.resize(npoint); 
+      
+      auto Aview_p = &AcceleratorViewContainer[0];
+      auto Vview_p = &AcceleratorVecViewContainer[0];
+      tcopy-=usecond();
+      acceleratorCopyToDevice(&AcceleratorViewContainer_h[0],&AcceleratorViewContainer[0],npoint *sizeof(Aview));
+      acceleratorCopyToDevice(&AcceleratorVecViewContainer_h[0],&AcceleratorVecViewContainer[0],npoint *sizeof(Vview));
+      tcopy+=usecond();
+
       tmult-=usecond();
-	accelerator_for(sss, osites*nbasis, Nsimd, {
-
-	    typedef decltype(coalescedRead(in_v[0]))    calcVector;
-
-	    int ss = sss/nbasis;
-	    int b  = sss%nbasis;
-
+      accelerator_for(spb, osites*nbasis*npoint, Nsimd, {
+	  typedef decltype(coalescedRead(in_v[0](0))) calcComplex;
+	  int32_t ss   = spb/(nbasis*npoint);
+	  int32_t bp   = spb%(nbasis*npoint);
+	  int32_t point= bp/nbasis;
+	  int32_t b    = bp%nbasis;
 	  auto SE  = Stencil_v.GetEntry(point,ss);
 	  auto nbr = coalescedReadGeneralPermute(in_v[SE->_offset],SE->_permute,Nd);
-	    auto res = out_v(ss)(b);
-	    for(int bb=0;bb<nbasis;bb++) {
-	      res = res + coalescedRead(A_v[ss](b,bb))*nbr(bb);
+	  auto res = coalescedRead(Aview_p[point][ss](0,b))*nbr(0);
+	  for(int bb=1;bb<nbasis;bb++) {
+	    res = res + coalescedRead(Aview_p[point][ss](bb,b))*nbr(bb);
+	  }
+	  coalescedWrite(Vview_p[point][ss](b),res);
+      });
+      tmult2-=usecond();
+      accelerator_for(sb, osites*nbasis, Nsimd, {
+	  int ss = sb/nbasis;
+	  int b  = sb%nbasis;
+	  auto res = coalescedRead(Vview_p[0][ss](b));
+	  for(int point=1;point<npoint;point++){
+	    res = res + coalescedRead(Vview_p[point][ss](b));
 	  }
 	  coalescedWrite(out_v[ss](b),res);
       });
-
+      tmult2+=usecond();
       tmult+=usecond();
+      for(int p=0;p<npoint;p++) {
+	AcceleratorViewContainer_h[p].ViewClose();
+	AcceleratorVecViewContainer_h[p].ViewClose();
       }
     }
+
     text-=usecond();
     out = Cell.Extract(pout);
     text+=usecond();
     ttot+=usecond();
     
-    std::cout << GridLogDebug<<"Coarse Mult Aviews "<<tviews<<" us"<<std::endl;
-    std::cout << GridLogDebug<<"Coarse Mult exch "<<texch<<" us"<<std::endl;
-    std::cout << GridLogDebug<<"Coarse Mult mult "<<tmult<<" us"<<std::endl;
-    std::cout << GridLogDebug<<"Coarse Mult ext  "<<text<<" us"<<std::endl;
-    std::cout << GridLogDebug<<"Coarse Mult tot  "<<ttot<<" us"<<std::endl;
-    std::cout << GridLogDebug<<"Coarse Kernel flop/s "<< flops/tmult<<" mflop/s"<<std::endl;
-    std::cout << GridLogDebug<<"Coarse Kernel bytes/s"<< bytes/tmult<<" MB/s"<<std::endl;
-    std::cout << GridLogDebug<<"Coarse overall flops/s "<< flops/ttot<<" mflop/s"<<std::endl;
-    std::cout << GridLogDebug<<"Coarse total bytes   "<< bytes/1e6<<" MB"<<std::endl;
+    std::cout << GridLogPerformance<<"Coarse 1rhs Mult Aviews "<<tviews<<" us"<<std::endl;
+    std::cout << GridLogPerformance<<"Coarse Mult exch "<<texch<<" us"<<std::endl;
+    std::cout << GridLogPerformance<<"Coarse Mult mult "<<tmult<<" us"<<std::endl;
+    std::cout << GridLogPerformance<<" of which mult2  "<<tmult2<<" us"<<std::endl;
+    std::cout << GridLogPerformance<<"Coarse Mult ext  "<<text<<" us"<<std::endl;
+    std::cout << GridLogPerformance<<"Coarse Mult temps "<<ttemps<<" us"<<std::endl;
+    std::cout << GridLogPerformance<<"Coarse Mult copy  "<<tcopy<<" us"<<std::endl;
+    std::cout << GridLogPerformance<<"Coarse Mult tot  "<<ttot<<" us"<<std::endl;
+    //    std::cout << GridLogPerformance<<std::endl;
+    std::cout << GridLogPerformance<<"Coarse Kernel flops "<< flops<<std::endl;
+    std::cout << GridLogPerformance<<"Coarse Kernel flop/s "<< flops/tmult<<" mflop/s"<<std::endl;
+    std::cout << GridLogPerformance<<"Coarse Kernel bytes/s "<< bytes/tmult<<" MB/s"<<std::endl;
+    std::cout << GridLogPerformance<<"Coarse overall flops/s "<< flops/ttot<<" mflop/s"<<std::endl;
+    std::cout << GridLogPerformance<<"Coarse total bytes   "<< bytes/1e6<<" MB"<<std::endl;
 
   };
   
@@ -262,6 +302,7 @@ public:
      *
      *     Where q_k = delta_k . (2*M_PI/global_nb[mu])
      */
+#if 0
   void CoarsenOperator(LinearOperatorBase<Lattice<Fobj> > &linop,
 		       Aggregation<Fobj,CComplex,nbasis> & Subspace)
   {
@@ -378,22 +419,17 @@ public:
 	autoView( FT_v  , FT[k], AcceleratorRead);
 	accelerator_for(sss, osites, 1, {
 	    for(int j=0;j<nbasis;j++){
-	      A_v[sss](j,i) = FT_v[sss](j);
+	      A_v[sss](i,j) = FT_v[sss](j);
 	    }
         });
       }
       tinv+=usecond();
     }
 
-    for(int p=0;p<geom.npoint;p++){
-      Coordinate coor({0,0,0,0,0});
-      auto sval = peekSite(_A[p],coor);
-    }
-
     // Only needed if nonhermitian
     if ( ! hermitian ) {
-      std::cout << GridLogMessage<<"PopulateAdag  "<<std::endl;
-      PopulateAdag();
+      //      std::cout << GridLogMessage<<"PopulateAdag  "<<std::endl;
+      //      PopulateAdag();
     }
 
     // Need to write something to populate Adag from A
@@ -404,10 +440,173 @@ public:
     std::cout << GridLogMessage<<"CoarsenOperator proj   "<<tproj<<" us"<<std::endl;
     std::cout << GridLogMessage<<"CoarsenOperator inv    "<<tinv<<" us"<<std::endl;
   }
+#else
+  void CoarsenOperator(LinearOperatorBase<Lattice<Fobj> > &linop,
+		       Aggregation<Fobj,CComplex,nbasis> & Subspace)
+  {
+    std::cout << GridLogMessage<< "GeneralCoarsenMatrix "<< std::endl;
+    GridBase *grid = FineGrid();
+
+    RealD tproj=0.0;
+    RealD teigen=0.0;
+    RealD tmat=0.0;
+    RealD tphase=0.0;
+    RealD tphaseBZ=0.0;
+    RealD tinv=0.0;
+
+    /////////////////////////////////////////////////////////////
+    // Orthogonalise the subblocks over the basis
+    /////////////////////////////////////////////////////////////
+    CoarseScalar InnerProd(CoarseGrid()); 
+    blockOrthogonalise(InnerProd,Subspace.subspace);
+
+    //    for(int s=0;s<Subspace.subspace.size();s++){
+      //      std::cout << " subspace norm "<<norm2(Subspace.subspace[s])<<std::endl;
+    //    }
+    const int npoint = geom.npoint;
+      
+    Coordinate clatt = CoarseGrid()->GlobalDimensions();
+    int Nd = CoarseGrid()->Nd();
+
+      /*
+       *     Here, k,l index which possible momentum/shift within the N-points connected by MdagM.
+       *     Matrix index i is mapped to this shift via 
+       *               geom.shifts[i]
+       *
+       *     conj(pha[block]) proj[k (which mom)][j (basis vec cpt)][block] 
+       *       =  \sum_{l in ball}  e^{i q_k . delta_l} < phi_{block,j} | MdagM | phi_{(block+delta_l),i} > 
+       *       =  \sum_{l in ball} e^{iqk.delta_l} A_ji^{b.b+l}
+       *       = M_{kl} A_ji^{b.b+l}
+       *
+       *     Must assemble and invert matrix M_k,l = e^[i q_k . delta_l]
+       *  
+       *     Where q_k = delta_k . (2*M_PI/global_nb[mu])
+       *
+       *     Then A{ji}^{b,b+l} = M^{-1}_{lm} ComputeProj_{m,b,i,j}
+       */
+    teigen-=usecond();
+    Eigen::MatrixXcd Mkl    = Eigen::MatrixXcd::Zero(npoint,npoint);
+    Eigen::MatrixXcd invMkl = Eigen::MatrixXcd::Zero(npoint,npoint);
+    ComplexD ci(0.0,1.0);
+    for(int k=0;k<npoint;k++){ // Loop over momenta
+
+      for(int l=0;l<npoint;l++){ // Loop over nbr relative
+	ComplexD phase(0.0,0.0);
+	for(int mu=0;mu<Nd;mu++){
+	  RealD TwoPiL =  M_PI * 2.0/ clatt[mu];
+	  phase=phase+TwoPiL*geom.shifts[k][mu]*geom.shifts[l][mu];
+	}
+	phase=exp(phase*ci);
+	Mkl(k,l) = phase;
+      }
+    }
+    invMkl = Mkl.inverse();
+    teigen+=usecond();
+
+    ///////////////////////////////////////////////////////////////////////
+    // Now compute the matrix elements of linop between the orthonormal
+    // set of vectors.
+    ///////////////////////////////////////////////////////////////////////
+    FineField phaV(grid); // Phased block basis vector
+    FineField MphaV(grid);// Matrix applied
+    std::vector<FineComplexField> phaF(npoint,grid);
+    std::vector<CoarseComplexField> pha(npoint,CoarseGrid());
+    
+    CoarseVector coarseInner(CoarseGrid());
+    
+    typedef typename CComplex::scalar_type SComplex;
+    FineComplexField one(grid); one=SComplex(1.0);
+    FineComplexField zz(grid); zz = Zero();
+    tphase=-usecond();
+    for(int p=0;p<npoint;p++){ // Loop over momenta in npoint
+      /////////////////////////////////////////////////////
+      // Stick a phase on every block
+      /////////////////////////////////////////////////////
+      CoarseComplexField coor(CoarseGrid());
+      pha[p]=Zero();
+      for(int mu=0;mu<Nd;mu++){
+	LatticeCoordinate(coor,mu);
+	RealD TwoPiL =  M_PI * 2.0/ clatt[mu];
+	pha[p] = pha[p] + (TwoPiL * geom.shifts[p][mu]) * coor;
+      }
+      pha[p]  =exp(pha[p]*ci);
+
+      blockZAXPY(phaF[p],pha[p],one,zz);
+      
+    }
+    tphase+=usecond();
+    
+    std::vector<CoarseVector> ComputeProj(npoint,CoarseGrid());
+    std::vector<CoarseVector>          FT(npoint,CoarseGrid());
+    for(int i=0;i<nbasis;i++){// Loop over basis vectors
+      std::cout << GridLogMessage<< "CoarsenMatrixColoured vec "<<i<<"/"<<nbasis<< std::endl;
+      for(int p=0;p<npoint;p++){ // Loop over momenta in npoint
+	tphaseBZ-=usecond();
+	phaV = phaF[p]*Subspace.subspace[i];
+	tphaseBZ+=usecond();
+
+	/////////////////////////////////////////////////////////////////////
+	// Multiple phased subspace vector by matrix and project to subspace
+	// Remove local bulk phase to leave relative phases
+	/////////////////////////////////////////////////////////////////////
+	tmat-=usecond();
+	linop.Op(phaV,MphaV);
+	tmat+=usecond();
+	//	std::cout << i << " " <<p << " MphaV "<<norm2(MphaV)<<" "<<norm2(phaV)<<std::endl;
+
+	tproj-=usecond();
+	blockProject(coarseInner,MphaV,Subspace.subspace);
+	coarseInner = conjugate(pha[p]) * coarseInner;
+
+	ComputeProj[p] = coarseInner;
+	tproj+=usecond();
+	//	std::cout << i << " " <<p << " ComputeProj "<<norm2(ComputeProj[p])<<std::endl;
+
+      }
+
+      tinv-=usecond();
+      for(int k=0;k<npoint;k++){
+	FT[k] = Zero();
+	for(int l=0;l<npoint;l++){
+	  FT[k]= FT[k]+ invMkl(l,k)*ComputeProj[l];
+	}
+      
+	int osites=CoarseGrid()->oSites();
+	autoView( A_v  , _A[k], AcceleratorWrite);
+	autoView( FT_v  , FT[k], AcceleratorRead);
+	accelerator_for(sss, osites, 1, {
+	    for(int j=0;j<nbasis;j++){
+	      A_v[sss](i,j) = FT_v[sss](j);
+	    }
+        });
+      }
+      tinv+=usecond();
+    }
+
+    // Only needed if nonhermitian
+    if ( ! hermitian ) {
+      //      std::cout << GridLogMessage<<"PopulateAdag  "<<std::endl;
+      //      PopulateAdag();
+    }
+
+    for(int p=0;p<geom.npoint;p++){
+      std::cout << " _A["<<p<<"] "<<norm2(_A[p])<<std::endl;
+    }
+
+    // Need to write something to populate Adag from A
+    ExchangeCoarseLinks();
+    std::cout << GridLogMessage<<"CoarsenOperator eigen  "<<teigen<<" us"<<std::endl;
+    std::cout << GridLogMessage<<"CoarsenOperator phase  "<<tphase<<" us"<<std::endl;
+    std::cout << GridLogMessage<<"CoarsenOperator phaseBZ "<<tphaseBZ<<" us"<<std::endl;
+    std::cout << GridLogMessage<<"CoarsenOperator mat    "<<tmat <<" us"<<std::endl;
+    std::cout << GridLogMessage<<"CoarsenOperator proj   "<<tproj<<" us"<<std::endl;
+    std::cout << GridLogMessage<<"CoarsenOperator inv    "<<tinv<<" us"<<std::endl;
+  }
+#endif  
   void ExchangeCoarseLinks(void){
     for(int p=0;p<geom.npoint;p++){
-      _A[p] = Cell.Exchange(_A[p]);
-      _Adag[p]= Cell.Exchange(_Adag[p]);
+      _A[p] = Cell.ExchangePeriodic(_A[p]);
+      //      _Adag[p]= Cell.ExchangePeriodic(_Adag[p]);
     }
   }
   virtual  void Mdiag    (const Field &in, Field &out){ assert(0);};
@@ -416,4 +615,5 @@ public:
 };
 
 
+  
 NAMESPACE_END(Grid);

Grid/algorithms/multigrid/GeneralCoarsenedMatrixMultiRHS.h

@@ -0,0 +1,729 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./lib/algorithms/GeneralCoarsenedMatrixMultiRHS.h
+
+    Copyright (C) 2015
+
+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 */
+#pragma once
+
+
+NAMESPACE_BEGIN(Grid);
+
+
+// Fine Object == (per site) type of fine field
+// nbasis      == number of deflation vectors
+template<class Fobj,class CComplex,int nbasis>
+class MultiGeneralCoarsenedMatrix : public SparseMatrixBase<Lattice<iVector<CComplex,nbasis > > >  {
+public:
+  typedef typename CComplex::scalar_object SComplex;
+  typedef GeneralCoarsenedMatrix<Fobj,CComplex,nbasis> GeneralCoarseOp;
+  typedef MultiGeneralCoarsenedMatrix<Fobj,CComplex,nbasis> MultiGeneralCoarseOp;
+
+  typedef iVector<CComplex,nbasis >           siteVector;
+  typedef iMatrix<CComplex,nbasis >           siteMatrix;
+  typedef iVector<SComplex,nbasis >           calcVector;
+  typedef iMatrix<SComplex,nbasis >           calcMatrix;
+  typedef Lattice<iScalar<CComplex> >         CoarseComplexField;
+  typedef Lattice<siteVector>                 CoarseVector;
+  typedef Lattice<iMatrix<CComplex,nbasis > > CoarseMatrix;
+  typedef iMatrix<CComplex,nbasis >  Cobj;
+  typedef iVector<CComplex,nbasis >  Cvec;
+  typedef Lattice< CComplex >   CoarseScalar; // used for inner products on fine field
+  typedef Lattice<Fobj >        FineField;
+  typedef Lattice<CComplex >    FineComplexField;
+  typedef CoarseVector Field;
+
+  ////////////////////
+  // Data members
+  ////////////////////
+  GridCartesian *       _CoarseGridMulti; 
+  NonLocalStencilGeometry geom;
+  NonLocalStencilGeometry geom_srhs;
+  PaddedCell Cell;
+  GeneralLocalStencil Stencil;
+
+  deviceVector<calcVector> BLAS_B;
+  deviceVector<calcVector> BLAS_C;
+  std::vector<deviceVector<calcMatrix> > BLAS_A;
+
+  std::vector<deviceVector<ComplexD *> > BLAS_AP;
+  std::vector<deviceVector<ComplexD *> > BLAS_BP;
+  deviceVector<ComplexD *>               BLAS_CP;
+
+  ///////////////////////
+  // Interface
+  ///////////////////////
+  GridBase      * Grid(void)           { return _CoarseGridMulti; };   // this is all the linalg routines need to know
+  GridCartesian * CoarseGrid(void)     { return _CoarseGridMulti; };   // this is all the linalg routines need to know
+
+  // Can be used to do I/O on the operator matrices externally
+  void SetMatrix (int p,CoarseMatrix & A)
+  {
+    assert(A.size()==geom_srhs.npoint);
+    GridtoBLAS(A[p],BLAS_A[p]);
+  }
+  void GetMatrix (int p,CoarseMatrix & A)
+  {
+    assert(A.size()==geom_srhs.npoint);
+    BLAStoGrid(A[p],BLAS_A[p]);
+  }
+  void CopyMatrix (GeneralCoarseOp &_Op)
+  {
+    for(int p=0;p<geom.npoint;p++){
+      auto Aup = _Op.Cell.Extract(_Op._A[p]);
+      //Unpadded
+      GridtoBLAS(Aup,BLAS_A[p]);
+    }
+  }
+  /*
+  void CheckMatrix (GeneralCoarseOp &_Op)
+  {
+    std::cout <<"************* Checking the little direc operator mRHS"<<std::endl;
+    for(int p=0;p<geom.npoint;p++){
+      //Unpadded
+      auto Aup = _Op.Cell.Extract(_Op._A[p]);
+      auto Ack = Aup;
+      BLAStoGrid(Ack,BLAS_A[p]);
+      std::cout << p<<" Ack "<<norm2(Ack)<<std::endl;
+      std::cout << p<<" Aup "<<norm2(Aup)<<std::endl;
+    }
+    std::cout <<"************* "<<std::endl;
+  }
+  */
+  
+  MultiGeneralCoarsenedMatrix(NonLocalStencilGeometry &_geom,GridCartesian *CoarseGridMulti) :
+    _CoarseGridMulti(CoarseGridMulti),
+    geom_srhs(_geom),
+    geom(_CoarseGridMulti,_geom.hops,_geom.skip+1),
+    Cell(geom.Depth(),_CoarseGridMulti),
+    Stencil(Cell.grids.back(),geom.shifts) // padded cell stencil
+  {
+    int32_t padded_sites   = Cell.grids.back()->lSites();
+    int32_t unpadded_sites = CoarseGridMulti->lSites();
+    
+    int32_t nrhs  = CoarseGridMulti->FullDimensions()[0];  // # RHS
+    int32_t orhs  = nrhs/CComplex::Nsimd();
+
+    padded_sites   = padded_sites/nrhs;
+    unpadded_sites = unpadded_sites/nrhs;
+    
+    /////////////////////////////////////////////////
+    // Device data vector storage
+    /////////////////////////////////////////////////
+    BLAS_A.resize(geom.npoint);
+    for(int p=0;p<geom.npoint;p++){
+      BLAS_A[p].resize (unpadded_sites); // no ghost zone, npoint elements
+    }
+    
+    BLAS_B.resize(nrhs *padded_sites);   // includes ghost zone
+    BLAS_C.resize(nrhs *unpadded_sites); // no ghost zone
+    BLAS_AP.resize(geom.npoint);
+    BLAS_BP.resize(geom.npoint);
+    for(int p=0;p<geom.npoint;p++){
+      BLAS_AP[p].resize(unpadded_sites);
+      BLAS_BP[p].resize(unpadded_sites);
+    }
+    BLAS_CP.resize(unpadded_sites);
+
+    /////////////////////////////////////////////////
+    // Pointers to data
+    /////////////////////////////////////////////////
+
+    // Site identity mapping for A
+    for(int p=0;p<geom.npoint;p++){
+      for(int ss=0;ss<unpadded_sites;ss++){
+	ComplexD *ptr = (ComplexD *)&BLAS_A[p][ss];
+	acceleratorPut(BLAS_AP[p][ss],ptr);
+      }
+    }
+    // Site identity mapping for C
+    for(int ss=0;ss<unpadded_sites;ss++){
+      ComplexD *ptr = (ComplexD *)&BLAS_C[ss*nrhs];
+      acceleratorPut(BLAS_CP[ss],ptr);
+    }
+
+    // Neighbour table is more complicated
+    int32_t j=0; // Interior point counter (unpadded)
+    for(int32_t s=0;s<padded_sites;s++){ // 4 volume, padded
+      int ghost_zone=0;
+      for(int32_t point = 0 ; point < geom.npoint; point++){
+	int i=s*orhs*geom.npoint+point;
+	if( Stencil._entries[i]._wrap ) { // stencil is indexed by the oSite of the CoarseGridMulti, hence orhs factor
+	  ghost_zone=1; // If general stencil wrapped in any direction, wrap=1
+	}
+      }
+
+      if( ghost_zone==0) {
+	for(int32_t point = 0 ; point < geom.npoint; point++){
+	  int i=s*orhs*geom.npoint+point;
+ 	  int32_t nbr = Stencil._entries[i]._offset*CComplex::Nsimd(); // oSite -> lSite
+	  assert(nbr<BLAS_B.size());
+	  ComplexD * ptr = (ComplexD *)&BLAS_B[nbr];
+	  acceleratorPut(BLAS_BP[point][j],ptr); // neighbour indexing in ghost zone volume
+	}
+	j++;
+      }
+    }
+    assert(j==unpadded_sites);
+  }
+  template<class vobj> void GridtoBLAS(const Lattice<vobj> &from,deviceVector<typename vobj::scalar_object> &to)
+  {
+  typedef typename vobj::scalar_object sobj;
+  typedef typename vobj::scalar_type scalar_type;
+  typedef typename vobj::vector_type vector_type;
+
+  GridBase *Fg = from.Grid();
+  assert(!Fg->_isCheckerBoarded);
+  int nd = Fg->_ndimension;
+
+  to.resize(Fg->lSites());
+
+  Coordinate LocalLatt = Fg->LocalDimensions();
+  size_t nsite = 1;
+  for(int i=0;i<nd;i++) nsite *= LocalLatt[i];
+
+  ////////////////////////////////////////////////////////////////////////////////////////////////
+  // do the index calc on the GPU
+  ////////////////////////////////////////////////////////////////////////////////////////////////
+  Coordinate f_ostride = Fg->_ostride;
+  Coordinate f_istride = Fg->_istride;
+  Coordinate f_rdimensions = Fg->_rdimensions;
+
+  autoView(from_v,from,AcceleratorRead);
+  auto to_v = &to[0];
+
+  const int words=sizeof(vobj)/sizeof(vector_type);
+  accelerator_for(idx,nsite,1,{
+      
+      Coordinate from_coor, base;
+      Lexicographic::CoorFromIndex(base,idx,LocalLatt);
+      for(int i=0;i<nd;i++){
+	from_coor[i] = base[i];
+      }
+      int from_oidx = 0; for(int d=0;d<nd;d++) from_oidx+=f_ostride[d]*(from_coor[d]%f_rdimensions[d]);
+      int from_lane = 0; for(int d=0;d<nd;d++) from_lane+=f_istride[d]*(from_coor[d]/f_rdimensions[d]);
+
+      const vector_type* from = (const vector_type *)&from_v[from_oidx];
+      scalar_type* to = (scalar_type *)&to_v[idx];
+      
+      scalar_type stmp;
+      for(int w=0;w<words;w++){
+	stmp = getlane(from[w], from_lane);
+	to[w] = stmp;
+      }
+    });
+  }    
+  template<class vobj> void BLAStoGrid(Lattice<vobj> &grid,deviceVector<typename vobj::scalar_object> &in)
+  {
+  typedef typename vobj::scalar_object sobj;
+  typedef typename vobj::scalar_type scalar_type;
+  typedef typename vobj::vector_type vector_type;
+
+  GridBase *Tg = grid.Grid();
+  assert(!Tg->_isCheckerBoarded);
+  int nd = Tg->_ndimension;
+  
+  assert(in.size()==Tg->lSites());
+
+  Coordinate LocalLatt = Tg->LocalDimensions();
+  size_t nsite = 1;
+  for(int i=0;i<nd;i++) nsite *= LocalLatt[i];
+
+  ////////////////////////////////////////////////////////////////////////////////////////////////
+  // do the index calc on the GPU
+  ////////////////////////////////////////////////////////////////////////////////////////////////
+  Coordinate t_ostride = Tg->_ostride;
+  Coordinate t_istride = Tg->_istride;
+  Coordinate t_rdimensions = Tg->_rdimensions;
+
+  autoView(to_v,grid,AcceleratorWrite);
+  auto from_v = &in[0];
+
+  const int words=sizeof(vobj)/sizeof(vector_type);
+  accelerator_for(idx,nsite,1,{
+      
+      Coordinate to_coor, base;
+      Lexicographic::CoorFromIndex(base,idx,LocalLatt);
+      for(int i=0;i<nd;i++){
+	to_coor[i] = base[i];
+      }
+      int to_oidx = 0; for(int d=0;d<nd;d++) to_oidx+=t_ostride[d]*(to_coor[d]%t_rdimensions[d]);
+      int to_lane = 0; for(int d=0;d<nd;d++) to_lane+=t_istride[d]*(to_coor[d]/t_rdimensions[d]);
+
+      vector_type* to = (vector_type *)&to_v[to_oidx];
+      scalar_type* from = (scalar_type *)&from_v[idx];
+      
+      scalar_type stmp;
+      for(int w=0;w<words;w++){
+	stmp=from[w];
+	putlane(to[w], stmp, to_lane);
+      }
+    });
+  }
+  void CoarsenOperator(LinearOperatorBase<Lattice<Fobj> > &linop,
+		       Aggregation<Fobj,CComplex,nbasis> & Subspace,
+		       GridBase *CoarseGrid)
+  {
+#if 0
+    std::cout << GridLogMessage<< "GeneralCoarsenMatrixMrhs "<< std::endl;
+
+    GridBase *grid = Subspace.FineGrid;
+
+    /////////////////////////////////////////////////////////////
+    // Orthogonalise the subblocks over the basis
+    /////////////////////////////////////////////////////////////
+    CoarseScalar InnerProd(CoarseGrid); 
+    blockOrthogonalise(InnerProd,Subspace.subspace);
+
+    const int npoint = geom_srhs.npoint;
+
+    Coordinate clatt = CoarseGrid->GlobalDimensions();
+    int Nd = CoarseGrid->Nd();
+      /*
+       *     Here, k,l index which possible momentum/shift within the N-points connected by MdagM.
+       *     Matrix index i is mapped to this shift via 
+       *               geom.shifts[i]
+       *
+       *     conj(pha[block]) proj[k (which mom)][j (basis vec cpt)][block] 
+       *       =  \sum_{l in ball}  e^{i q_k . delta_l} < phi_{block,j} | MdagM | phi_{(block+delta_l),i} > 
+       *       =  \sum_{l in ball} e^{iqk.delta_l} A_ji^{b.b+l}
+       *       = M_{kl} A_ji^{b.b+l}
+       *
+       *     Must assemble and invert matrix M_k,l = e^[i q_k . delta_l]
+       *  
+       *     Where q_k = delta_k . (2*M_PI/global_nb[mu])
+       *
+       *     Then A{ji}^{b,b+l} = M^{-1}_{lm} ComputeProj_{m,b,i,j}
+       */
+    Eigen::MatrixXcd Mkl    = Eigen::MatrixXcd::Zero(npoint,npoint);
+    Eigen::MatrixXcd invMkl = Eigen::MatrixXcd::Zero(npoint,npoint);
+    ComplexD ci(0.0,1.0);
+    for(int k=0;k<npoint;k++){ // Loop over momenta
+
+      for(int l=0;l<npoint;l++){ // Loop over nbr relative
+	ComplexD phase(0.0,0.0);
+	for(int mu=0;mu<Nd;mu++){
+	  RealD TwoPiL =  M_PI * 2.0/ clatt[mu];
+	  phase=phase+TwoPiL*geom_srhs.shifts[k][mu]*geom_srhs.shifts[l][mu];
+	}
+	phase=exp(phase*ci);
+	Mkl(k,l) = phase;
+      }
+    }
+    invMkl = Mkl.inverse();
+
+    ///////////////////////////////////////////////////////////////////////
+    // Now compute the matrix elements of linop between the orthonormal
+    // set of vectors.
+    ///////////////////////////////////////////////////////////////////////
+    FineField phaV(grid); // Phased block basis vector
+    FineField MphaV(grid);// Matrix applied
+    std::vector<FineComplexField> phaF(npoint,grid);
+    std::vector<CoarseComplexField> pha(npoint,CoarseGrid);
+    
+    CoarseVector coarseInner(CoarseGrid);
+    
+    typedef typename CComplex::scalar_type SComplex;
+    FineComplexField one(grid); one=SComplex(1.0);
+    FineComplexField zz(grid); zz = Zero();
+    for(int p=0;p<npoint;p++){ // Loop over momenta in npoint
+      /////////////////////////////////////////////////////
+      // Stick a phase on every block
+      /////////////////////////////////////////////////////
+      CoarseComplexField coor(CoarseGrid);
+      pha[p]=Zero();
+      for(int mu=0;mu<Nd;mu++){
+	LatticeCoordinate(coor,mu);
+	RealD TwoPiL =  M_PI * 2.0/ clatt[mu];
+	pha[p] = pha[p] + (TwoPiL * geom_srhs.shifts[p][mu]) * coor;
+      }
+      pha[p]  =exp(pha[p]*ci);	
+
+      blockZAXPY(phaF[p],pha[p],one,zz);
+    }
+
+    // Could save on temporary storage here
+    std::vector<CoarseMatrix> _A;
+    _A.resize(geom_srhs.npoint,CoarseGrid);
+
+    std::vector<CoarseVector> ComputeProj(npoint,CoarseGrid);
+    CoarseVector          FT(CoarseGrid);
+    for(int i=0;i<nbasis;i++){// Loop over basis vectors
+      std::cout << GridLogMessage<< "CoarsenMatrixColoured vec "<<i<<"/"<<nbasis<< std::endl;
+      for(int p=0;p<npoint;p++){ // Loop over momenta in npoint
+
+	phaV = phaF[p]*Subspace.subspace[i];
+
+	/////////////////////////////////////////////////////////////////////
+	// Multiple phased subspace vector by matrix and project to subspace
+	// Remove local bulk phase to leave relative phases
+	/////////////////////////////////////////////////////////////////////
+	linop.Op(phaV,MphaV);
+
+	// Fixme, could use batched block projector here
+	blockProject(coarseInner,MphaV,Subspace.subspace);
+
+	coarseInner = conjugate(pha[p]) * coarseInner;
+
+	ComputeProj[p] = coarseInner;
+      }
+
+      // Could do this with a block promote or similar BLAS call via the MultiRHSBlockProjector with a const matrix.
+      for(int k=0;k<npoint;k++){
+
+	FT = Zero();
+	for(int l=0;l<npoint;l++){
+	  FT= FT+ invMkl(l,k)*ComputeProj[l];
+	}
+      
+	int osites=CoarseGrid->oSites();
+	autoView( A_v  , _A[k], AcceleratorWrite);
+	autoView( FT_v  , FT, AcceleratorRead);
+	accelerator_for(sss, osites, 1, {
+	    for(int j=0;j<nbasis;j++){
+	      A_v[sss](i,j) = FT_v[sss](j);
+	    }
+        });
+      }
+    }
+
+    // Only needed if nonhermitian
+    //    if ( ! hermitian ) {
+    //      std::cout << GridLogMessage<<"PopulateAdag  "<<std::endl;
+    //      PopulateAdag();
+    //    }
+    // Need to write something to populate Adag from A
+
+    for(int p=0;p<geom_srhs.npoint;p++){
+      GridtoBLAS(_A[p],BLAS_A[p]);
+    }
+    /*
+Grid : Message : 11698.730546 s : CoarsenOperator eigen  1334 us
+Grid : Message : 11698.730563 s : CoarsenOperator phase  34729 us
+Grid : Message : 11698.730565 s : CoarsenOperator phaseBZ 2423814 us
+Grid : Message : 11698.730566 s : CoarsenOperator mat    127890998 us
+Grid : Message : 11698.730567 s : CoarsenOperator proj   515840840 us
+Grid : Message : 11698.730568 s : CoarsenOperator inv    103948313 us
+Takes 600s to compute matrix elements, DOMINATED by the block project.
+Easy to speed up with the batched block project.
+Store npoint vectors, get npoint x Nbasis block projection, and 81 fold faster.
+
+// Block project below taks to 240s
+Grid : Message : 328.193418 s : CoarsenOperator phase      38338 us
+Grid : Message : 328.193434 s : CoarsenOperator phaseBZ  1711226 us
+Grid : Message : 328.193436 s : CoarsenOperator mat    122213270 us
+//Grid : Message : 328.193438 s : CoarsenOperator proj   1181154 us <-- this is mistimed
+//Grid : Message : 11698.730568 s : CoarsenOperator inv  103948313 us <-- Cut this ~10x if lucky by loop fusion
+     */
+#else
+    RealD tproj=0.0;
+    RealD tmat=0.0;
+    RealD tphase=0.0;
+    RealD tphaseBZ=0.0;
+    RealD tinv=0.0;
+
+    std::cout << GridLogMessage<< "GeneralCoarsenMatrixMrhs "<< std::endl;
+
+    GridBase *grid = Subspace.FineGrid;
+
+    /////////////////////////////////////////////////////////////
+    // Orthogonalise the subblocks over the basis
+    /////////////////////////////////////////////////////////////
+    CoarseScalar InnerProd(CoarseGrid); 
+    blockOrthogonalise(InnerProd,Subspace.subspace);
+
+
+    MultiRHSBlockProject<Lattice<Fobj> >    Projector;
+    Projector.Allocate(nbasis,grid,CoarseGrid);
+    Projector.ImportBasis(Subspace.subspace);
+    
+    const int npoint = geom_srhs.npoint;
+
+    Coordinate clatt = CoarseGrid->GlobalDimensions();
+    int Nd = CoarseGrid->Nd();
+      /*
+       *     Here, k,l index which possible momentum/shift within the N-points connected by MdagM.
+       *     Matrix index i is mapped to this shift via 
+       *               geom.shifts[i]
+       *
+       *     conj(pha[block]) proj[k (which mom)][j (basis vec cpt)][block] 
+       *       =  \sum_{l in ball}  e^{i q_k . delta_l} < phi_{block,j} | MdagM | phi_{(block+delta_l),i} > 
+       *       =  \sum_{l in ball} e^{iqk.delta_l} A_ji^{b.b+l}
+       *       = M_{kl} A_ji^{b.b+l}
+       *
+       *     Must assemble and invert matrix M_k,l = e^[i q_k . delta_l]
+       *  
+       *     Where q_k = delta_k . (2*M_PI/global_nb[mu])
+       *
+       *     Then A{ji}^{b,b+l} = M^{-1}_{lm} ComputeProj_{m,b,i,j}
+       */
+    Eigen::MatrixXcd Mkl    = Eigen::MatrixXcd::Zero(npoint,npoint);
+    Eigen::MatrixXcd invMkl = Eigen::MatrixXcd::Zero(npoint,npoint);
+    ComplexD ci(0.0,1.0);
+    for(int k=0;k<npoint;k++){ // Loop over momenta
+
+      for(int l=0;l<npoint;l++){ // Loop over nbr relative
+	ComplexD phase(0.0,0.0);
+	for(int mu=0;mu<Nd;mu++){
+	  RealD TwoPiL =  M_PI * 2.0/ clatt[mu];
+	  phase=phase+TwoPiL*geom_srhs.shifts[k][mu]*geom_srhs.shifts[l][mu];
+	}
+	phase=exp(phase*ci);
+	Mkl(k,l) = phase;
+      }
+    }
+    invMkl = Mkl.inverse();
+
+    ///////////////////////////////////////////////////////////////////////
+    // Now compute the matrix elements of linop between the orthonormal
+    // set of vectors.
+    ///////////////////////////////////////////////////////////////////////
+    FineField phaV(grid); // Phased block basis vector
+    FineField MphaV(grid);// Matrix applied
+    std::vector<FineComplexField> phaF(npoint,grid);
+    std::vector<CoarseComplexField> pha(npoint,CoarseGrid);
+    
+    CoarseVector coarseInner(CoarseGrid);
+    
+    tphase=-usecond();
+    typedef typename CComplex::scalar_type SComplex;
+    FineComplexField one(grid); one=SComplex(1.0);
+    FineComplexField zz(grid); zz = Zero();
+    for(int p=0;p<npoint;p++){ // Loop over momenta in npoint
+      /////////////////////////////////////////////////////
+      // Stick a phase on every block
+      /////////////////////////////////////////////////////
+      CoarseComplexField coor(CoarseGrid);
+      pha[p]=Zero();
+      for(int mu=0;mu<Nd;mu++){
+	LatticeCoordinate(coor,mu);
+	RealD TwoPiL =  M_PI * 2.0/ clatt[mu];
+	pha[p] = pha[p] + (TwoPiL * geom_srhs.shifts[p][mu]) * coor;
+      }
+      pha[p]  =exp(pha[p]*ci);	
+
+      blockZAXPY(phaF[p],pha[p],one,zz);
+    }
+    tphase+=usecond();
+
+    // Could save on temporary storage here
+    std::vector<CoarseMatrix> _A;
+    _A.resize(geom_srhs.npoint,CoarseGrid);
+
+    // Count use small chunks than npoint == 81 and save memory
+    int batch = 9;
+    std::vector<FineField>    _MphaV(batch,grid);
+    std::vector<CoarseVector> TmpProj(batch,CoarseGrid);
+
+    std::vector<CoarseVector> ComputeProj(npoint,CoarseGrid);
+    CoarseVector          FT(CoarseGrid);
+    for(int i=0;i<nbasis;i++){// Loop over basis vectors
+      std::cout << GridLogMessage<< "CoarsenMatrixColoured vec "<<i<<"/"<<nbasis<< std::endl;
+
+      //      std::cout << GridLogMessage << " phasing the fine vector "<<std::endl;
+      // Fixme : do this in batches
+      for(int p=0;p<npoint;p+=batch){ // Loop over momenta in npoint
+
+	for(int b=0;b<MIN(batch,npoint-p);b++){
+	  tphaseBZ-=usecond();
+	  phaV = phaF[p+b]*Subspace.subspace[i];
+	  tphaseBZ+=usecond();
+
+	  /////////////////////////////////////////////////////////////////////
+	  // Multiple phased subspace vector by matrix and project to subspace
+	  // Remove local bulk phase to leave relative phases
+	  /////////////////////////////////////////////////////////////////////
+	  // Memory footprint was an issue
+	  tmat-=usecond();
+	  linop.Op(phaV,MphaV);
+	  _MphaV[b] = MphaV;
+	  tmat+=usecond();
+	}      
+
+	//	std::cout << GridLogMessage << " Calling block project "<<std::endl;
+	tproj-=usecond();
+	Projector.blockProject(_MphaV,TmpProj);
+	tproj+=usecond();
+	
+	//	std::cout << GridLogMessage << " conj phasing the coarse vectors "<<std::endl;
+	for(int b=0;b<MIN(batch,npoint-p);b++){
+	  ComputeProj[p+b] = conjugate(pha[p+b])*TmpProj[b];
+	}
+      }
+
+      // Could do this with a block promote or similar BLAS call via the MultiRHSBlockProjector with a const matrix.
+      
+      // std::cout << GridLogMessage << " Starting FT inv "<<std::endl;
+      tinv-=usecond();
+      for(int k=0;k<npoint;k++){
+	FT = Zero();
+	// 81 kernel calls as many ComputeProj vectors
+	// Could fuse with a vector of views, but ugly
+	// Could unroll the expression and run fewer kernels -- much more attractive
+	// Could also do non blocking.
+#if 0	
+	for(int l=0;l<npoint;l++){
+	  FT= FT+ invMkl(l,k)*ComputeProj[l];
+	}
+#else
+	const int radix = 9;
+	int ll;
+	for(ll=0;ll+radix-1<npoint;ll+=radix){
+	  // When ll = npoint-radix, ll+radix-1 = npoint-1, and we do it all.
+	  FT = FT 
+	    + invMkl(ll+0,k)*ComputeProj[ll+0]
+	    + invMkl(ll+1,k)*ComputeProj[ll+1]
+	    + invMkl(ll+2,k)*ComputeProj[ll+2]
+	    + invMkl(ll+3,k)*ComputeProj[ll+3]
+	    + invMkl(ll+4,k)*ComputeProj[ll+4]
+	    + invMkl(ll+5,k)*ComputeProj[ll+5]
+	    + invMkl(ll+6,k)*ComputeProj[ll+6]
+	    + invMkl(ll+7,k)*ComputeProj[ll+7]
+	    + invMkl(ll+8,k)*ComputeProj[ll+8];
+	}
+	for(int l=ll;l<npoint;l++){
+	  FT= FT+ invMkl(l,k)*ComputeProj[l];
+	}
+#endif
+      
+	// 1 kernel call -- must be cheaper
+	int osites=CoarseGrid->oSites();
+	autoView( A_v  , _A[k], AcceleratorWrite);
+	autoView( FT_v  , FT, AcceleratorRead);
+	accelerator_for(sss, osites, 1, {
+	    for(int j=0;j<nbasis;j++){
+	      A_v[sss](i,j) = FT_v[sss](j);
+	    }
+        });
+      }
+      tinv+=usecond();
+    }
+
+    // Only needed if nonhermitian
+    //    if ( ! hermitian ) {
+    //      std::cout << GridLogMessage<<"PopulateAdag  "<<std::endl;
+    //      PopulateAdag();
+    //    }
+    // Need to write something to populate Adag from A
+    //    std::cout << GridLogMessage << " Calling GridtoBLAS "<<std::endl;
+    for(int p=0;p<geom_srhs.npoint;p++){
+      GridtoBLAS(_A[p],BLAS_A[p]);
+    }
+    std::cout << GridLogMessage<<"CoarsenOperator phase  "<<tphase<<" us"<<std::endl;
+    std::cout << GridLogMessage<<"CoarsenOperator phaseBZ "<<tphaseBZ<<" us"<<std::endl;
+    std::cout << GridLogMessage<<"CoarsenOperator mat    "<<tmat <<" us"<<std::endl;
+    std::cout << GridLogMessage<<"CoarsenOperator proj   "<<tproj<<" us"<<std::endl;
+    std::cout << GridLogMessage<<"CoarsenOperator inv    "<<tinv<<" us"<<std::endl;
+#endif
+  }
+  void Mdag(const CoarseVector &in, CoarseVector &out)
+  {
+    this->M(in,out);
+  }
+  void M (const CoarseVector &in, CoarseVector &out)
+  {
+    //    std::cout << GridLogMessage << "New Mrhs coarse"<<std::endl;
+    conformable(CoarseGrid(),in.Grid());
+    conformable(in.Grid(),out.Grid());
+    out.Checkerboard() = in.Checkerboard();
+
+    RealD t_tot;
+    RealD t_exch;
+    RealD t_GtoB;
+    RealD t_BtoG;
+    RealD t_mult;
+
+    t_tot=-usecond();
+    CoarseVector tin=in;
+    t_exch=-usecond();
+    CoarseVector pin = Cell.ExchangePeriodic(tin); //padded input
+    t_exch+=usecond();
+
+    CoarseVector pout(pin.Grid());
+
+    int npoint = geom.npoint;
+    typedef calcMatrix* Aview;
+    typedef LatticeView<Cvec> Vview;
+      
+    const int Nsimd = CComplex::Nsimd();
+
+    int64_t nrhs  =pin.Grid()->GlobalDimensions()[0];
+    assert(nrhs>=1);
+
+    RealD flops,bytes;
+    int64_t osites=in.Grid()->oSites(); // unpadded
+    int64_t unpadded_vol = CoarseGrid()->lSites()/nrhs;
+    
+    flops = 1.0* npoint * nbasis * nbasis * 8.0 * osites * CComplex::Nsimd();
+    bytes = 1.0*osites*sizeof(siteMatrix)*npoint/pin.Grid()->GlobalDimensions()[0]
+          + 2.0*osites*sizeof(siteVector)*npoint;
+    
+
+    t_GtoB=-usecond();
+    GridtoBLAS(pin,BLAS_B);
+    t_GtoB+=usecond();
+
+    GridBLAS BLAS;
+
+    t_mult=-usecond();
+    for(int p=0;p<geom.npoint;p++){
+      RealD c = 1.0;
+      if (p==0) c = 0.0;
+      ComplexD beta(c);
+
+      BLAS.gemmBatched(nbasis,nrhs,nbasis,
+		       ComplexD(1.0),
+		       BLAS_AP[p], 
+		       BLAS_BP[p], 
+		       ComplexD(c), 
+		       BLAS_CP);
+    }
+    BLAS.synchronise();
+    t_mult+=usecond();
+
+    t_BtoG=-usecond();
+    BLAStoGrid(out,BLAS_C);
+    t_BtoG+=usecond();
+    t_tot+=usecond();
+    /*
+    std::cout << GridLogMessage << "New Mrhs coarse DONE "<<std::endl;
+    std::cout << GridLogMessage<<"Coarse Mult exch "<<t_exch<<" us"<<std::endl;
+    std::cout << GridLogMessage<<"Coarse Mult mult "<<t_mult<<" us"<<std::endl;
+    std::cout << GridLogMessage<<"Coarse Mult GtoB  "<<t_GtoB<<" us"<<std::endl;
+    std::cout << GridLogMessage<<"Coarse Mult BtoG  "<<t_BtoG<<" us"<<std::endl;
+    std::cout << GridLogMessage<<"Coarse Mult tot  "<<t_tot<<" us"<<std::endl;
+    */
+    //    std::cout << GridLogMessage<<std::endl;
+    //    std::cout << GridLogMessage<<"Coarse Kernel flops "<< flops<<std::endl;
+    //    std::cout << GridLogMessage<<"Coarse Kernel flop/s "<< flops/t_mult<<" mflop/s"<<std::endl;
+    //    std::cout << GridLogMessage<<"Coarse Kernel bytes/s "<< bytes/t_mult/1000<<" GB/s"<<std::endl;
+    //    std::cout << GridLogMessage<<"Coarse overall flops/s "<< flops/t_tot<<" mflop/s"<<std::endl;
+    //    std::cout << GridLogMessage<<"Coarse total bytes   "<< bytes/1e6<<" MB"<<std::endl;
+  };
+  virtual  void Mdiag    (const Field &in, Field &out){ assert(0);};
+  virtual  void Mdir     (const Field &in, Field &out,int dir, int disp){assert(0);};
+  virtual  void MdirAll  (const Field &in, std::vector<Field> &out){assert(0);};
+};
+  
+NAMESPACE_END(Grid);

Grid/algorithms/multigrid/Geometry.h

@@ -104,7 +104,8 @@ public:
 /////////////////////////////////////////////////////////////////
 class NonLocalStencilGeometry {
 public:
-  int depth;
+  //  int depth;
+  int skip;
   int hops;
   int npoint;
   std::vector<Coordinate> shifts;
@@ -115,8 +116,7 @@ public:
   GridCartesian *Grid() {return grid;};
   int Depth(void){return 1;};   // Ghost zone depth
   int Hops(void){return hops;}; // # of hops=> level of corner fill in in stencil
-
-  virtual int DimSkip(void) =0;
+  int DimSkip(void){return skip;};
 
   virtual ~NonLocalStencilGeometry() {};
 
@@ -156,7 +156,7 @@ public:
     std::cout << GridLogMessage << "NonLocalStencilGeometry has "<< this->npoint << " terms in stencil "<<std::endl;
   }
   
-  NonLocalStencilGeometry(GridCartesian *_coarse_grid,int _hops) : grid(_coarse_grid), hops(_hops)
+  NonLocalStencilGeometry(GridCartesian *_coarse_grid,int _hops,int _skip) : grid(_coarse_grid), hops(_hops), skip(_skip)
   {
     Coordinate latt = grid->GlobalDimensions();
     stencil_size.resize(grid->Nd());
@@ -177,6 +177,7 @@ public:
        stencil_size[d]= 3;
      }
     }
+    this->BuildShifts();
   };
 
 };
@@ -184,14 +185,14 @@ public:
 // Need to worry about red-black now
 class NonLocalStencilGeometry4D : public NonLocalStencilGeometry {
 public:
-  virtual int DimSkip(void) { return 0;};
-  NonLocalStencilGeometry4D(GridCartesian *Coarse,int _hops) : NonLocalStencilGeometry(Coarse,_hops) { };
+  virtual int DerivedDimSkip(void) { return 0;};
+  NonLocalStencilGeometry4D(GridCartesian *Coarse,int _hops) : NonLocalStencilGeometry(Coarse,_hops,0) { };
   virtual ~NonLocalStencilGeometry4D() {};
 };
 class NonLocalStencilGeometry5D : public NonLocalStencilGeometry {
 public:
-  virtual int DimSkip(void) { return 1; }; 
-  NonLocalStencilGeometry5D(GridCartesian *Coarse,int _hops) : NonLocalStencilGeometry(Coarse,_hops)  { };
+  virtual int DerivedDimSkip(void) { return 1; }; 
+  NonLocalStencilGeometry5D(GridCartesian *Coarse,int _hops) : NonLocalStencilGeometry(Coarse,_hops,1)  { };
   virtual ~NonLocalStencilGeometry5D() {};
 };
 /*
@@ -201,42 +202,36 @@ class NextToNextToNextToNearestStencilGeometry4D : public NonLocalStencilGeometr
 public:
   NextToNextToNextToNearestStencilGeometry4D(GridCartesian *Coarse) :  NonLocalStencilGeometry4D(Coarse,4)
   {
-    this->BuildShifts();
   };
 };
 class NextToNextToNextToNearestStencilGeometry5D : public  NonLocalStencilGeometry5D {
 public:
   NextToNextToNextToNearestStencilGeometry5D(GridCartesian *Coarse) :  NonLocalStencilGeometry5D(Coarse,4)
   {
-    this->BuildShifts();
   };
 };
 class NextToNearestStencilGeometry4D : public  NonLocalStencilGeometry4D {
 public:
   NextToNearestStencilGeometry4D(GridCartesian *Coarse) :  NonLocalStencilGeometry4D(Coarse,2)
   {
-    this->BuildShifts();
   };
 };
 class NextToNearestStencilGeometry5D : public  NonLocalStencilGeometry5D {
 public:
   NextToNearestStencilGeometry5D(GridCartesian *Coarse) :  NonLocalStencilGeometry5D(Coarse,2)
   {
-    this->BuildShifts();
   };
 };
 class NearestStencilGeometry4D : public  NonLocalStencilGeometry4D {
 public:
   NearestStencilGeometry4D(GridCartesian *Coarse) :  NonLocalStencilGeometry4D(Coarse,1)
   {
-    this->BuildShifts();
   };
 };
 class NearestStencilGeometry5D : public  NonLocalStencilGeometry5D {
 public:
   NearestStencilGeometry5D(GridCartesian *Coarse) :  NonLocalStencilGeometry5D(Coarse,1)
   {
-    this->BuildShifts();
   };
 };
 

Grid/algorithms/multigrid/MultiGrid.h

@@ -31,3 +31,4 @@ Author: Peter Boyle <pboyle@bnl.gov>
 #include <Grid/algorithms/multigrid/Geometry.h>
 #include <Grid/algorithms/multigrid/CoarsenedMatrix.h>
 #include <Grid/algorithms/multigrid/GeneralCoarsenedMatrix.h>
+#include <Grid/algorithms/multigrid/GeneralCoarsenedMatrixMultiRHS.h>

Grid/allocator/AlignedAllocator.h

@@ -54,6 +54,9 @@ public:
     size_type bytes = __n*sizeof(_Tp);
     profilerAllocate(bytes);
     _Tp *ptr = (_Tp*) MemoryManager::CpuAllocate(bytes);
+    if ( (_Tp*)ptr == (_Tp *) NULL ) {
+      printf("Grid CPU Allocator got NULL for %lu bytes\n",(unsigned long) bytes );
+    }
     assert( ( (_Tp*)ptr != (_Tp *)NULL ) );
     return ptr;
   }
@@ -100,6 +103,9 @@ public:
     size_type bytes = __n*sizeof(_Tp);
     profilerAllocate(bytes);
     _Tp *ptr = (_Tp*) MemoryManager::SharedAllocate(bytes);
+    if ( (_Tp*)ptr == (_Tp *) NULL ) {
+      printf("Grid Shared Allocator got NULL for %lu bytes\n",(unsigned long) bytes );
+    }
     assert( ( (_Tp*)ptr != (_Tp *)NULL ) );
     return ptr;
   }
@@ -145,6 +151,9 @@ public:
     size_type bytes = __n*sizeof(_Tp);
     profilerAllocate(bytes);
     _Tp *ptr = (_Tp*) MemoryManager::AcceleratorAllocate(bytes);
+    if ( (_Tp*)ptr == (_Tp *) NULL ) {
+      printf("Grid Device Allocator got NULL for %lu bytes\n",(unsigned long) bytes );
+    }
     assert( ( (_Tp*)ptr != (_Tp *)NULL ) );
     return ptr;
   }
@@ -165,18 +174,46 @@ template<typename _Tp>  inline bool operator!=(const devAllocator<_Tp>&, const d
 ////////////////////////////////////////////////////////////////////////////////
 // Template typedefs
 ////////////////////////////////////////////////////////////////////////////////
-#ifdef ACCELERATOR_CSHIFT
-// Cshift on device
-template<class T> using cshiftAllocator = devAllocator<T>;
-#else
-// Cshift on host
-template<class T> using cshiftAllocator = std::allocator<T>;
-#endif
+template<class T> using hostVector          = std::vector<T,alignedAllocator<T> >;           // Needs autoview
+template<class T> using Vector              = std::vector<T,uvmAllocator<T> >;               // 
+template<class T> using uvmVector           = std::vector<T,uvmAllocator<T> >;               // auto migrating page
+template<class T> using deviceVector        = std::vector<T,devAllocator<T> >;               // device vector
 
-template<class T> using Vector        = std::vector<T,uvmAllocator<T> >;           
-template<class T> using stencilVector = std::vector<T,alignedAllocator<T> >;           
-template<class T> using commVector = std::vector<T,devAllocator<T> >;
-template<class T> using cshiftVector = std::vector<T,cshiftAllocator<T> >;
+template<class T> class vecView
+{
+ protected:
+  T * data;
+  uint64_t size;
+  ViewMode mode;
+  void * cpu_ptr;
+ public:
+  // Rvalue accessor
+  accelerator_inline T & operator[](size_t i) const { return this->data[i]; };
+  vecView(Vector<T> &refer_to_me,ViewMode _mode)
+  {
+    cpu_ptr = &refer_to_me[0];
+    size = refer_to_me.size();
+    mode = _mode;
+    data =(T *) MemoryManager::ViewOpen(cpu_ptr,
+					size*sizeof(T),
+					mode,
+					AdviseDefault);
+  }
+  void ViewClose(void)
+  { // Inform the manager
+    MemoryManager::ViewClose(this->cpu_ptr,this->mode);    
+  }
+};
+
+template<class T> vecView<T> VectorView(Vector<T> &vec,ViewMode _mode)
+{
+  vecView<T> ret(vec,_mode); // does the open
+  return ret;                // must be closed
+}
+
+#define autoVecView(v_v,v,mode)					\
+  auto v_v = VectorView(v,mode);				\
+  ViewCloser<decltype(v_v)> _autoView##v_v(v_v);
 
 NAMESPACE_END(Grid);
 

Grid/allocator/MemoryManager.cc

@@ -16,6 +16,44 @@ NAMESPACE_BEGIN(Grid);
 uint64_t total_shared;
 uint64_t total_device;
 uint64_t total_host;;
+
+#if defined(__has_feature)
+#if __has_feature(leak_sanitizer)
+#define ASAN_LEAK_CHECK
+#endif
+#endif
+
+#ifdef ASAN_LEAK_CHECK
+#include <sanitizer/asan_interface.h>
+#include <sanitizer/common_interface_defs.h>
+#include <sanitizer/lsan_interface.h>
+#define LEAK_CHECK(A) { __lsan_do_recoverable_leak_check(); }
+#else
+#define LEAK_CHECK(A) { }
+#endif
+
+void MemoryManager::DisplayMallinfo(void)
+{
+#ifdef __linux__
+  struct mallinfo mi; // really want mallinfo2, but glibc version isn't uniform
+  
+  mi = mallinfo();
+
+  std::cout << "MemoryManager: Total non-mmapped bytes (arena):       "<< (size_t)mi.arena<<std::endl;
+  std::cout << "MemoryManager: # of free chunks (ordblks):            "<< (size_t)mi.ordblks<<std::endl;
+  std::cout << "MemoryManager: # of free fastbin blocks (smblks):     "<< (size_t)mi.smblks<<std::endl;
+  std::cout << "MemoryManager: # of mapped regions (hblks):           "<< (size_t)mi.hblks<<std::endl;
+  std::cout << "MemoryManager: Bytes in mapped regions (hblkhd):      "<< (size_t)mi.hblkhd<<std::endl;
+  std::cout << "MemoryManager: Max. total allocated space (usmblks):  "<< (size_t)mi.usmblks<<std::endl;
+  std::cout << "MemoryManager: Free bytes held in fastbins (fsmblks): "<< (size_t)mi.fsmblks<<std::endl;
+  std::cout << "MemoryManager: Total allocated space (uordblks):      "<< (size_t)mi.uordblks<<std::endl;
+  std::cout << "MemoryManager: Total free space (fordblks):           "<< (size_t)mi.fordblks<<std::endl;
+  std::cout << "MemoryManager: Topmost releasable block (keepcost):   "<< (size_t)mi.keepcost<<std::endl;
+#endif
+  LEAK_CHECK();
+ 
+}
+
 void MemoryManager::PrintBytes(void)
 {
   std::cout << " MemoryManager : ------------------------------------ "<<std::endl;
@@ -35,7 +73,7 @@ void MemoryManager::PrintBytes(void)
 #ifdef GRID_CUDA
   cuda_mem();
 #endif
-  
+  DisplayMallinfo();
 }
 
 uint64_t MemoryManager::DeviceCacheBytes() { return CacheBytes[Acc] + CacheBytes[AccHuge] + CacheBytes[AccSmall]; }

Grid/allocator/MemoryManager.h

@@ -209,9 +209,10 @@ private:
   static void     CpuViewClose(uint64_t Ptr);
   static uint64_t CpuViewOpen(uint64_t  CpuPtr,size_t bytes,ViewMode mode,ViewAdvise hint);
 #endif
-  static void NotifyDeletion(void * CpuPtr);
 
  public:
+  static void DisplayMallinfo(void);
+  static void NotifyDeletion(void * CpuPtr);
   static void Print(void);
   static void PrintAll(void);
   static void PrintState( void* CpuPtr);

Grid/allocator/MemoryManagerCache.cc

@@ -1,17 +1,15 @@
 #include <Grid/GridCore.h>
 #ifndef GRID_UVM
 
-#warning "Using explicit device memory copies"
 NAMESPACE_BEGIN(Grid);
 
 #define MAXLINE 512
 static char print_buffer [ MAXLINE ];
 
-#define mprintf(...) snprintf (print_buffer,MAXLINE, __VA_ARGS__ ); std::cout << GridLogMemory << print_buffer;
-#define dprintf(...) snprintf (print_buffer,MAXLINE, __VA_ARGS__ ); std::cout << GridLogMemory << print_buffer;
+#define mprintf(...) snprintf (print_buffer,MAXLINE, __VA_ARGS__ ); std::cout << GridLogMemory << print_buffer << std::endl;
+#define dprintf(...) snprintf (print_buffer,MAXLINE, __VA_ARGS__ ); std::cout << GridLogDebug  << print_buffer << std::endl;
 //#define dprintf(...) 
 
-
 ////////////////////////////////////////////////////////////
 // For caching copies of data on device
 ////////////////////////////////////////////////////////////
@@ -111,7 +109,7 @@ void MemoryManager::AccDiscard(AcceleratorViewEntry &AccCache)
   ///////////////////////////////////////////////////////////
   assert(AccCache.state!=Empty);
   
-  mprintf("MemoryManager: Discard(%lx) %lx\n",(uint64_t)AccCache.CpuPtr,(uint64_t)AccCache.AccPtr); 
+  dprintf("MemoryManager: Discard(%lx) %lx\n",(uint64_t)AccCache.CpuPtr,(uint64_t)AccCache.AccPtr); 
   assert(AccCache.accLock==0);
   assert(AccCache.cpuLock==0);
   assert(AccCache.CpuPtr!=(uint64_t)NULL);
@@ -141,7 +139,7 @@ void MemoryManager::Evict(AcceleratorViewEntry &AccCache)
   ///////////////////////////////////////////////////////////////////////////
   assert(AccCache.state!=Empty);
   
-  mprintf("MemoryManager: Evict cpu %lx acc %lx cpuLock %ld accLock %ld\n",
+  mprintf("MemoryManager: Evict CpuPtr %lx AccPtr %lx cpuLock %ld accLock %ld\n",
 	  (uint64_t)AccCache.CpuPtr,(uint64_t)AccCache.AccPtr,
 	  (uint64_t)AccCache.cpuLock,(uint64_t)AccCache.accLock); 
   if (AccCache.accLock!=0) return;
@@ -155,7 +153,7 @@ void MemoryManager::Evict(AcceleratorViewEntry &AccCache)
     AccCache.AccPtr=(uint64_t)NULL;
     AccCache.state=CpuDirty; // CPU primary now
     DeviceBytes   -=AccCache.bytes;
-    dprintf("MemoryManager: Free(%lx) footprint now %ld \n",(uint64_t)AccCache.AccPtr,DeviceBytes);  
+    dprintf("MemoryManager: Free(AccPtr %lx) footprint now %ld \n",(uint64_t)AccCache.AccPtr,DeviceBytes);  
   }
   //  uint64_t CpuPtr = AccCache.CpuPtr;
   DeviceEvictions++;
@@ -169,7 +167,7 @@ void MemoryManager::Flush(AcceleratorViewEntry &AccCache)
   assert(AccCache.AccPtr!=(uint64_t)NULL);
   assert(AccCache.CpuPtr!=(uint64_t)NULL);
   acceleratorCopyFromDevice((void *)AccCache.AccPtr,(void *)AccCache.CpuPtr,AccCache.bytes);
-  mprintf("MemoryManager: Flush  %lx -> %lx\n",(uint64_t)AccCache.AccPtr,(uint64_t)AccCache.CpuPtr); fflush(stdout);
+  mprintf("MemoryManager: acceleratorCopyFromDevice Flush size %ld AccPtr %lx -> CpuPtr %lx\n",(uint64_t)AccCache.bytes,(uint64_t)AccCache.AccPtr,(uint64_t)AccCache.CpuPtr); fflush(stdout);
   DeviceToHostBytes+=AccCache.bytes;
   DeviceToHostXfer++;
   AccCache.state=Consistent;
@@ -184,7 +182,9 @@ void MemoryManager::Clone(AcceleratorViewEntry &AccCache)
     AccCache.AccPtr=(uint64_t)AcceleratorAllocate(AccCache.bytes);
     DeviceBytes+=AccCache.bytes;
   }
-  mprintf("MemoryManager: Clone %lx <- %lx\n",(uint64_t)AccCache.AccPtr,(uint64_t)AccCache.CpuPtr); fflush(stdout);
+  mprintf("MemoryManager: acceleratorCopyToDevice   Clone size %ld AccPtr %lx <- CpuPtr %lx\n",
+	  (uint64_t)AccCache.bytes,
+	  (uint64_t)AccCache.AccPtr,(uint64_t)AccCache.CpuPtr); fflush(stdout);
   acceleratorCopyToDevice((void *)AccCache.CpuPtr,(void *)AccCache.AccPtr,AccCache.bytes);
   HostToDeviceBytes+=AccCache.bytes;
   HostToDeviceXfer++;
@@ -265,7 +265,7 @@ uint64_t MemoryManager::AcceleratorViewOpen(uint64_t CpuPtr,size_t bytes,ViewMod
   assert(AccCache.cpuLock==0);  // Programming error
 
   if(AccCache.state!=Empty) {
-    dprintf("ViewOpen found entry %lx %lx : %ld %ld accLock %ld\n",
+    dprintf("ViewOpen found entry %lx %lx : sizes %ld %ld accLock %ld\n",
 		    (uint64_t)AccCache.CpuPtr,
 		    (uint64_t)CpuPtr,
 		    (uint64_t)AccCache.bytes,
@@ -474,6 +474,7 @@ void  MemoryManager::Print(void)
   std::cout << GridLogMessage << DeviceEvictions  << " Evictions from device " << std::endl;
   std::cout << GridLogMessage << DeviceDestroy    << " Destroyed vectors on device " << std::endl;
   std::cout << GridLogMessage << AccViewTable.size()<< " vectors " << LRU.size()<<" evictable"<< std::endl;
+  acceleratorMem();
   std::cout << GridLogMessage << "--------------------------------------------" << std::endl;
 }
 void  MemoryManager::PrintAll(void)

Grid/allocator/MemoryStats.cc

@@ -15,10 +15,10 @@ void check_huge_pages(void *Buf,uint64_t BYTES)
   uint64_t virt_pfn = (uint64_t)Buf / page_size;
   off_t offset = sizeof(uint64_t) * virt_pfn;
   uint64_t npages = (BYTES + page_size-1) / page_size;
-  uint64_t pagedata[npages];
+  std::vector<uint64_t> pagedata(npages);
   uint64_t ret = lseek(fd, offset, SEEK_SET);
   assert(ret == offset);
-  ret = ::read(fd, pagedata, sizeof(uint64_t)*npages);
+  ret = ::read(fd, &pagedata[0], sizeof(uint64_t)*npages);
   assert(ret == sizeof(uint64_t) * npages);
   int nhugepages = npages / 512;
   int n4ktotal, nnothuge;

Grid/cartesian/Cartesian_base.h

@@ -82,6 +82,7 @@ public:
   bool _isCheckerBoarded; 
   int        LocallyPeriodic;
   Coordinate _checker_dim_mask;
+  int              _checker_dim;
 
 public:
 

Grid/cartesian/Cartesian_full.h

@@ -38,7 +38,7 @@ class GridCartesian: public GridBase {
 
 public:
   int dummy;
-  Coordinate _checker_dim_mask;
+  //  Coordinate _checker_dim_mask;
   virtual int  CheckerBoardFromOindexTable (int Oindex) {
     return 0;
   }
@@ -106,6 +106,7 @@ public:
     _rdimensions.resize(_ndimension);
     _simd_layout.resize(_ndimension);
     _checker_dim_mask.resize(_ndimension);;
+    _checker_dim = -1;
     _lstart.resize(_ndimension);
     _lend.resize(_ndimension);
 

Grid/cartesian/Cartesian_red_black.h

@@ -57,9 +57,10 @@ class GridRedBlackCartesian : public GridBase
 {
 public:
   //  Coordinate _checker_dim_mask;
-  int              _checker_dim;
+  //  int              _checker_dim;
   std::vector<int> _checker_board;
 
+  virtual int isCheckerBoarded(void) const { return 1; };
   virtual int CheckerBoarded(int dim){
     if( dim==_checker_dim) return 1;
     else return 0;

Grid/communicator/Communicator_base.cc

@@ -57,18 +57,29 @@ int                      CartesianCommunicator::ProcessorCount(void)    { return
 // very VERY rarely (Log, serial RNG) we need world without a grid
 ////////////////////////////////////////////////////////////////////////////////
 
+#ifdef USE_GRID_REDUCTION
+void CartesianCommunicator::GlobalSum(ComplexF &c)
+{
+  GlobalSumP2P(c);
+}
+void CartesianCommunicator::GlobalSum(ComplexD &c)
+{
+  GlobalSumP2P(c);
+}
+#else
 void CartesianCommunicator::GlobalSum(ComplexF &c)
 {
   GlobalSumVector((float *)&c,2);
 }
-void CartesianCommunicator::GlobalSumVector(ComplexF *c,int N)
-{
-  GlobalSumVector((float *)c,2*N);
-}
 void CartesianCommunicator::GlobalSum(ComplexD &c)
 {
   GlobalSumVector((double *)&c,2);
 }
+#endif
+void CartesianCommunicator::GlobalSumVector(ComplexF *c,int N)
+{
+  GlobalSumVector((float *)c,2*N);
+}
 void CartesianCommunicator::GlobalSumVector(ComplexD *c,int N)
 {
   GlobalSumVector((double *)c,2*N);

Grid/communicator/Communicator_base.h

@@ -128,6 +128,34 @@ public:
   void GlobalXOR(uint32_t &);
   void GlobalXOR(uint64_t &);
   
+  template<class obj> void GlobalSumP2P(obj &o)
+  {
+    std::vector<obj> column;
+    obj accum = o;
+    int source,dest;
+    for(int d=0;d<_ndimension;d++){
+      column.resize(_processors[d]);
+      column[0] = accum;
+      std::vector<MpiCommsRequest_t> list;
+      for(int p=1;p<_processors[d];p++){
+	ShiftedRanks(d,p,source,dest);
+	SendToRecvFromBegin(list,
+			    &column[0],
+			    dest,
+			    &column[p],
+			    source,
+			    sizeof(obj),d*100+p);
+
+      }
+      CommsComplete(list);
+      for(int p=1;p<_processors[d];p++){
+	accum = accum + column[p];
+      }
+    }
+    Broadcast(0,accum);
+    o=accum;
+  }
+
   template<class obj> void GlobalSum(obj &o){
     typedef typename obj::scalar_type scalar_type;
     int words = sizeof(obj)/sizeof(scalar_type);
@@ -138,6 +166,14 @@ public:
   ////////////////////////////////////////////////////////////
   // Face exchange, buffer swap in translational invariant way
   ////////////////////////////////////////////////////////////
+  void CommsComplete(std::vector<MpiCommsRequest_t> &list);
+  void SendToRecvFromBegin(std::vector<MpiCommsRequest_t> &list,
+			   void *xmit,
+			   int dest,
+			   void *recv,
+			   int from,
+			   int bytes,int dir);
+  
   void SendToRecvFrom(void *xmit,
 		      int xmit_to_rank,
 		      void *recv,
@@ -150,6 +186,12 @@ public:
 			       int recv_from_rank,int do_recv,
 			       int bytes,int dir);
 
+  double StencilSendToRecvFromPrepare(std::vector<CommsRequest_t> &list,
+				      void *xmit,
+				      int xmit_to_rank,int do_xmit,
+				      void *recv,
+				      int recv_from_rank,int do_recv,
+				      int xbytes,int rbytes,int dir);
   double StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list,
 				    void *xmit,
 				    int xmit_to_rank,int do_xmit,

Grid/communicator/Communicator_mpi3.cc

@@ -257,6 +257,25 @@ CartesianCommunicator::~CartesianCommunicator()
     }
   }
 }
+#ifdef USE_GRID_REDUCTION
+void CartesianCommunicator::GlobalSum(float &f){
+  CartesianCommunicator::GlobalSumP2P(f);
+}
+void CartesianCommunicator::GlobalSum(double &d)
+{
+  CartesianCommunicator::GlobalSumP2P(d);
+}
+#else
+void CartesianCommunicator::GlobalSum(float &f){
+  int ierr=MPI_Allreduce(MPI_IN_PLACE,&f,1,MPI_FLOAT,MPI_SUM,communicator);
+  assert(ierr==0);
+}
+void CartesianCommunicator::GlobalSum(double &d)
+{
+  int ierr = MPI_Allreduce(MPI_IN_PLACE,&d,1,MPI_DOUBLE,MPI_SUM,communicator);
+  assert(ierr==0);
+}
+#endif
 void CartesianCommunicator::GlobalSum(uint32_t &u){
   int ierr=MPI_Allreduce(MPI_IN_PLACE,&u,1,MPI_UINT32_T,MPI_SUM,communicator);
   assert(ierr==0);
@@ -287,25 +306,54 @@ void CartesianCommunicator::GlobalMax(double &d)
   int ierr = MPI_Allreduce(MPI_IN_PLACE,&d,1,MPI_DOUBLE,MPI_MAX,communicator);
   assert(ierr==0);
 }
-void CartesianCommunicator::GlobalSum(float &f){
-  int ierr=MPI_Allreduce(MPI_IN_PLACE,&f,1,MPI_FLOAT,MPI_SUM,communicator);
-  assert(ierr==0);
-}
 void CartesianCommunicator::GlobalSumVector(float *f,int N)
 {
   int ierr=MPI_Allreduce(MPI_IN_PLACE,f,N,MPI_FLOAT,MPI_SUM,communicator);
   assert(ierr==0);
 }
-void CartesianCommunicator::GlobalSum(double &d)
-{
-  int ierr = MPI_Allreduce(MPI_IN_PLACE,&d,1,MPI_DOUBLE,MPI_SUM,communicator);
-  assert(ierr==0);
-}
 void CartesianCommunicator::GlobalSumVector(double *d,int N)
 {
   int ierr = MPI_Allreduce(MPI_IN_PLACE,d,N,MPI_DOUBLE,MPI_SUM,communicator);
   assert(ierr==0);
 }
+
+void CartesianCommunicator::SendToRecvFromBegin(std::vector<MpiCommsRequest_t> &list,
+						void *xmit,
+						int dest,
+						void *recv,
+						int from,
+						int bytes,int dir)
+{
+  MPI_Request xrq;
+  MPI_Request rrq;
+
+  assert(dest != _processor);
+  assert(from != _processor);
+
+  int tag;
+
+  tag= dir+from*32;
+  int ierr=MPI_Irecv(recv, bytes, MPI_CHAR,from,tag,communicator,&rrq);
+  assert(ierr==0);
+  list.push_back(rrq);
+  
+  tag= dir+_processor*32;
+  ierr =MPI_Isend(xmit, bytes, MPI_CHAR,dest,tag,communicator,&xrq);
+  assert(ierr==0);
+  list.push_back(xrq);
+}
+void CartesianCommunicator::CommsComplete(std::vector<MpiCommsRequest_t> &list)
+{
+  int nreq=list.size();
+
+  if (nreq==0) return;
+
+  std::vector<MPI_Status> status(nreq);
+  int ierr = MPI_Waitall(nreq,&list[0],&status[0]);
+  assert(ierr==0);
+  list.resize(0);
+}
+
 // Basic Halo comms primitive
 void CartesianCommunicator::SendToRecvFrom(void *xmit,
 					   int dest,
@@ -313,7 +361,7 @@ void CartesianCommunicator::SendToRecvFrom(void *xmit,
 					   int from,
 					   int bytes)
 {
-  std::vector<CommsRequest_t> reqs(0);
+  std::vector<MpiCommsRequest_t> reqs(0);
   unsigned long  xcrc = crc32(0L, Z_NULL, 0);
   unsigned long  rcrc = crc32(0L, Z_NULL, 0);
 
@@ -343,11 +391,23 @@ double CartesianCommunicator::StencilSendToRecvFrom( void *xmit,
 						     int bytes,int dir)
 {
   std::vector<CommsRequest_t> list;
-  double offbytes = StencilSendToRecvFromBegin(list,xmit,dest,dox,recv,from,dor,bytes,bytes,dir);
+  double offbytes = StencilSendToRecvFromPrepare(list,xmit,dest,dox,recv,from,dor,bytes,bytes,dir);
+  offbytes       += StencilSendToRecvFromBegin(list,xmit,dest,dox,recv,from,dor,bytes,bytes,dir);
   StencilSendToRecvFromComplete(list,dir);
   return offbytes;
 }
 
+
+#ifdef ACCELERATOR_AWARE_MPI
+double CartesianCommunicator::StencilSendToRecvFromPrepare(std::vector<CommsRequest_t> &list,
+							   void *xmit,
+							   int dest,int dox,
+							   void *recv,
+							   int from,int dor,
+							   int xbytes,int rbytes,int dir)
+{
+  return 0.0; // Do nothing -- no preparation required
+}
 double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list,
 							 void *xmit,
 							 int dest,int dox,
@@ -395,7 +455,231 @@ double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsReques
       acceleratorCopyDeviceToDeviceAsynch(xmit,shm,xbytes);
     }
   }
+  return off_node_bytes;
+}
 
+void CartesianCommunicator::StencilSendToRecvFromComplete(std::vector<CommsRequest_t> &list,int dir)
+{
+  int nreq=list.size();
+
+  acceleratorCopySynchronise();
+
+  if (nreq==0) return;
+  std::vector<MPI_Status> status(nreq);
+  int ierr = MPI_Waitall(nreq,&list[0],&status[0]);
+  assert(ierr==0);
+  list.resize(0);
+  this->StencilBarrier(); 
+}
+
+#else /* NOT     ... ACCELERATOR_AWARE_MPI */
+///////////////////////////////////////////
+// Pipeline mode through host memory
+///////////////////////////////////////////
+  /*
+   * In prepare (phase 1):
+   * PHASE 1: (prepare)
+   * - post MPI receive buffers asynch
+   * - post device - host send buffer transfer asynch
+   * PHASE 2: (Begin)
+   * - complete all copies
+   * - post MPI send asynch
+   * - post device - device transfers
+   * PHASE 3: (Complete)
+   * - MPI_waitall
+   * - host-device transfers
+   *
+   *********************************
+   * NB could split this further:
+   *--------------------------------
+   * PHASE 1: (Prepare)
+   * - post MPI receive buffers asynch
+   * - post device - host send buffer transfer asynch
+   * PHASE 2: (BeginInterNode)
+   * - complete all copies 
+   * - post MPI send asynch
+   * PHASE 3: (BeginIntraNode)
+   * - post device - device transfers
+   * PHASE 4: (Complete)
+   * - MPI_waitall
+   * - host-device transfers asynch
+   * - (complete all copies) 
+   */
+double CartesianCommunicator::StencilSendToRecvFromPrepare(std::vector<CommsRequest_t> &list,
+							   void *xmit,
+							   int dest,int dox,
+							   void *recv,
+							   int from,int dor,
+							   int xbytes,int rbytes,int dir)
+{
+/*
+ * Bring sequence from Stencil.h down to lower level.
+ * Assume using XeLink is ok
+ */  
+  int ncomm  =communicator_halo.size();
+  int commdir=dir%ncomm;
+
+  MPI_Request xrq;
+  MPI_Request rrq;
+
+  int ierr;
+  int gdest = ShmRanks[dest];
+  int gfrom = ShmRanks[from];
+  int gme   = ShmRanks[_processor];
+
+  assert(dest != _processor);
+  assert(from != _processor);
+  assert(gme  == ShmRank);
+  double off_node_bytes=0.0;
+  int tag;
+
+  void * host_recv = NULL;
+  void * host_xmit = NULL;
+
+  /*
+   * PHASE 1: (Prepare)
+   * - post MPI receive buffers asynch
+   * - post device - host send buffer transfer asynch
+   */
+  
+  if ( dor ) {
+    if ( (gfrom ==MPI_UNDEFINED) || Stencil_force_mpi ) {
+      tag= dir+from*32;
+      host_recv = this->HostBufferMalloc(rbytes);
+      ierr=MPI_Irecv(host_recv, rbytes, MPI_CHAR,from,tag,communicator_halo[commdir],&rrq);
+      assert(ierr==0);
+      CommsRequest_t srq;
+      srq.PacketType = InterNodeRecv;
+      srq.bytes      = rbytes;
+      srq.req        = rrq;
+      srq.host_buf   = host_recv;
+      srq.device_buf = recv;
+      list.push_back(srq);
+      off_node_bytes+=rbytes;
+    }
+  }
+  
+  if (dox) {
+    if ( (gdest == MPI_UNDEFINED) || Stencil_force_mpi ) {
+#undef DEVICE_TO_HOST_CONCURRENT // pipeline
+#ifdef DEVICE_TO_HOST_CONCURRENT
+      tag= dir+_processor*32;
+
+      host_xmit = this->HostBufferMalloc(xbytes);
+      acceleratorCopyFromDeviceAsynch(xmit, host_xmit,xbytes); // Make this Asynch
+      
+      //      ierr =MPI_Isend(host_xmit, xbytes, MPI_CHAR,dest,tag,communicator_halo[commdir],&xrq);
+      //      assert(ierr==0);
+      //      off_node_bytes+=xbytes;
+
+      CommsRequest_t srq;
+      srq.PacketType = InterNodeXmit;
+      srq.bytes      = xbytes;
+      //      srq.req        = xrq;
+      srq.host_buf   = host_xmit;
+      srq.device_buf = xmit;
+      list.push_back(srq);
+#else
+      tag= dir+_processor*32;
+
+      host_xmit = this->HostBufferMalloc(xbytes);
+      const int chunks=1;
+      for(int n=0;n<chunks;n++){
+	void * host_xmitc = (void *)( (uint64_t) host_xmit + n*xbytes/chunks);
+	void * xmitc      = (void *)( (uint64_t) xmit      + n*xbytes/chunks);
+	acceleratorCopyFromDeviceAsynch(xmitc, host_xmitc,xbytes/chunks); // Make this Asynch
+      }
+      acceleratorCopySynchronise(); // Complete all pending copy transfers
+      
+      ierr =MPI_Isend(host_xmit, xbytes, MPI_CHAR,dest,tag,communicator_halo[commdir],&xrq);
+      assert(ierr==0);
+      off_node_bytes+=xbytes;
+
+      CommsRequest_t srq;
+      srq.PacketType = InterNodeXmit;
+      srq.bytes      = xbytes;
+      srq.req        = xrq;
+      srq.host_buf   = host_xmit;
+      srq.device_buf = xmit;
+      list.push_back(srq);
+#endif
+    }
+  }
+
+  return off_node_bytes;
+}
+
+double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list,
+							 void *xmit,
+							 int dest,int dox,
+							 void *recv,
+							 int from,int dor,
+							 int xbytes,int rbytes,int dir)
+{
+  int ncomm  =communicator_halo.size();
+  int commdir=dir%ncomm;
+
+  MPI_Request xrq;
+  MPI_Request rrq;
+
+  int ierr;
+  int gdest = ShmRanks[dest];
+  int gfrom = ShmRanks[from];
+  int gme   = ShmRanks[_processor];
+
+  assert(dest != _processor);
+  assert(from != _processor);
+  assert(gme  == ShmRank);
+  double off_node_bytes=0.0;
+  int tag;
+
+  void * host_xmit = NULL;
+
+  ////////////////////////////////
+  // Receives already posted
+  // Copies already started
+  ////////////////////////////////
+  /*  
+   * PHASE 2: (Begin)
+   * - complete all copies
+   * - post MPI send asynch
+   */
+
+  //  static int printed;
+  //  if((printed<8) && this->IsBoss() ) {
+  //    printf("dir %d doX %d doR %d Face size %ld %ld\n",dir,dox,dor,xbytes,rbytes);
+  //    printed++;
+  //  }
+  
+  if (dox) {
+
+    if ( (gdest == MPI_UNDEFINED) || Stencil_force_mpi ) {
+#ifdef DEVICE_TO_HOST_CONCURRENT
+      tag= dir+_processor*32;
+      // Find the send in the prepared list
+      int list_idx=-1;
+      for(int idx = 0; idx<list.size();idx++){
+
+	if ( (list[idx].device_buf==xmit)
+	   &&(list[idx].PacketType==InterNodeXmit)
+	   &&(list[idx].bytes==xbytes) ) {
+
+	  list_idx = idx;
+	  host_xmit = list[idx].host_buf;
+	}
+      }
+      assert(list_idx != -1); // found it
+      ierr =MPI_Isend(host_xmit, xbytes, MPI_CHAR,dest,tag,communicator_halo[commdir],&xrq);
+      assert(ierr==0);
+      list[list_idx].req        = xrq; // Update the MPI request in the list
+      off_node_bytes+=xbytes;
+#endif      
+    } else {
+      void *shm = (void *) this->ShmBufferTranslate(dest,recv);
+      assert(shm!=NULL);
+      acceleratorCopyDeviceToDeviceAsynch(xmit,shm,xbytes);
+    }
+  }
   return off_node_bytes;
 }
 void CartesianCommunicator::StencilSendToRecvFromComplete(std::vector<CommsRequest_t> &list,int dir)
@@ -403,12 +687,32 @@ void CartesianCommunicator::StencilSendToRecvFromComplete(std::vector<CommsReque
   int nreq=list.size();
 
   if (nreq==0) return;
-
   std::vector<MPI_Status> status(nreq);
-  int ierr = MPI_Waitall(nreq,&list[0],&status[0]);
-  assert(ierr==0);
-  list.resize(0);
+  std::vector<MPI_Request> MpiRequests(nreq);
+
+  for(int r=0;r<nreq;r++){
+    MpiRequests[r] = list[r].req;
   }
+  
+  int ierr = MPI_Waitall(nreq,&MpiRequests[0],&status[0]);
+  assert(ierr==0);
+
+  for(int r=0;r<nreq;r++){
+    if ( list[r].PacketType==InterNodeRecv ) {
+      acceleratorCopyToDeviceAsynch(list[r].host_buf,list[r].device_buf,list[r].bytes);
+    }
+  }
+  
+  acceleratorCopySynchronise(); // Complete all pending copy transfers
+  list.resize(0);               // Delete the list
+  this->HostBufferFreeAll();    // Clean up the buffer allocs
+  this->StencilBarrier(); 
+}
+#endif
+////////////////////////////////////////////
+// END PIPELINE MODE / NO CUDA AWARE MPI
+////////////////////////////////////////////
+
 void CartesianCommunicator::StencilBarrier(void)
 {
   MPI_Barrier  (ShmComm);

Grid/communicator/Communicator_none.cc

@@ -91,6 +91,17 @@ void CartesianCommunicator::SendToRecvFrom(void *xmit,
 {
   assert(0);
 }
+void CartesianCommunicator::CommsComplete(std::vector<CommsRequest_t> &list){ assert(0);}
+void CartesianCommunicator::SendToRecvFromBegin(std::vector<CommsRequest_t> &list,
+						void *xmit,
+						int dest,
+						void *recv,
+						int from,
+						int bytes,int dir)
+{
+  assert(0);
+}
+
 void CartesianCommunicator::AllToAll(int dim,void  *in,void *out,uint64_t words,uint64_t bytes)
 {
   bcopy(in,out,bytes*words);
@@ -121,6 +132,15 @@ double CartesianCommunicator::StencilSendToRecvFrom( void *xmit,
 {
   return 2.0*bytes;
 }
+double CartesianCommunicator::StencilSendToRecvFromPrepare(std::vector<CommsRequest_t> &list,
+							   void *xmit,
+							   int xmit_to_rank,int dox,
+							   void *recv,
+							   int recv_from_rank,int dor,
+							   int xbytes,int rbytes, int dir)
+{
+  return xbytes+rbytes;
+}
 double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list,
 							 void *xmit,
 							 int xmit_to_rank,int dox,

Grid/communicator/SharedMemory.cc

@@ -40,6 +40,9 @@ int                 GlobalSharedMemory::_ShmAlloc;
 uint64_t            GlobalSharedMemory::_ShmAllocBytes;
 
 std::vector<void *> GlobalSharedMemory::WorldShmCommBufs;
+#ifndef ACCELERATOR_AWARE_MPI
+void * GlobalSharedMemory::HostCommBuf;
+#endif
 
 Grid_MPI_Comm       GlobalSharedMemory::WorldShmComm;
 int                 GlobalSharedMemory::WorldShmRank;
@@ -66,6 +69,26 @@ void GlobalSharedMemory::SharedMemoryFree(void)
 /////////////////////////////////
 // Alloc, free shmem region
 /////////////////////////////////
+#ifndef ACCELERATOR_AWARE_MPI
+void *SharedMemory::HostBufferMalloc(size_t bytes){
+  void *ptr = (void *)host_heap_top;
+  host_heap_top  += bytes;
+  host_heap_bytes+= bytes;
+  if (host_heap_bytes >= host_heap_size) {
+    std::cout<< " HostBufferMalloc exceeded heap size -- try increasing with --shm <MB> flag" <<std::endl;
+    std::cout<< " Parameter specified in units of MB (megabytes) " <<std::endl;
+    std::cout<< " Current alloc is " << (bytes/(1024*1024)) <<"MB"<<std::endl;
+    std::cout<< " Current bytes is " << (host_heap_bytes/(1024*1024)) <<"MB"<<std::endl;
+    std::cout<< " Current heap  is " << (host_heap_size/(1024*1024)) <<"MB"<<std::endl;
+    assert(host_heap_bytes<host_heap_size);
+  }
+  return ptr;
+}
+void SharedMemory::HostBufferFreeAll(void) { 
+  host_heap_top  =(size_t)HostCommBuf;
+  host_heap_bytes=0;
+}
+#endif
 void *SharedMemory::ShmBufferMalloc(size_t bytes){
   //  bytes = (bytes+sizeof(vRealD))&(~(sizeof(vRealD)-1));// align up bytes
   void *ptr = (void *)heap_top;

Grid/communicator/SharedMemory.h

@@ -46,8 +46,22 @@ NAMESPACE_BEGIN(Grid);
 
 #if defined (GRID_COMMS_MPI3) 
 typedef MPI_Comm    Grid_MPI_Comm;
+typedef MPI_Request MpiCommsRequest_t;
+#ifdef ACCELERATOR_AWARE_MPI
 typedef MPI_Request CommsRequest_t;
 #else
+enum PacketType_t { InterNodeXmit, InterNodeRecv, IntraNodeXmit, IntraNodeRecv };
+typedef struct {
+  PacketType_t PacketType;
+  void *host_buf;
+  void *device_buf;
+  unsigned long bytes;
+  MpiCommsRequest_t req;
+} CommsRequest_t;
+#endif
+
+#else 
+typedef int MpiCommsRequest_t;
 typedef int CommsRequest_t;
 typedef int Grid_MPI_Comm;
 #endif
@@ -75,7 +89,9 @@ public:
   static int           Hugepages;
 
   static std::vector<void *> WorldShmCommBufs;
-
+#ifndef ACCELERATOR_AWARE_MPI
+  static void *HostCommBuf;
+#endif
   static Grid_MPI_Comm WorldComm;
   static int           WorldRank;
   static int           WorldSize;
@@ -120,6 +136,13 @@ private:
   size_t heap_bytes;
   size_t heap_size;
 
+#ifndef ACCELERATOR_AWARE_MPI
+  size_t host_heap_top;  // set in free all
+  size_t host_heap_bytes;// set in free all
+  void *HostCommBuf;     // set in SetCommunicator
+  size_t host_heap_size; // set in SetCommunicator
+#endif
+  
 protected:
 
   Grid_MPI_Comm    ShmComm; // for barriers
@@ -151,7 +174,10 @@ public:
   void *ShmBufferTranslate(int rank,void * local_p);
   void *ShmBufferMalloc(size_t bytes);
   void  ShmBufferFreeAll(void) ;
-  
+#ifndef ACCELERATOR_AWARE_MPI
+  void *HostBufferMalloc(size_t bytes);
+  void HostBufferFreeAll(void);
+#endif  
   //////////////////////////////////////////////////////////////////////////
   // Make info on Nodes & ranks and Shared memory available
   //////////////////////////////////////////////////////////////////////////

Grid/communicator/SharedMemoryMPI.cc

@@ -39,9 +39,16 @@ Author: Christoph Lehner <christoph@lhnr.de>
 #include <hip/hip_runtime_api.h>
 #endif
 #ifdef GRID_SYCL
+#ifdef ACCELERATOR_AWARE_MPI
 #define GRID_SYCL_LEVEL_ZERO_IPC
-#include <syscall.h>
 #define SHM_SOCKETS
+#else
+#ifdef HAVE_NUMAIF_H
+  #warning " Using NUMAIF "
+#include <numaif.h>
+#endif 
+#endif 
+#include <syscall.h>
 #endif
 
 #include <sys/socket.h>
@@ -511,46 +518,6 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 ////////////////////////////////////////////////////////////////////////////////////////////
 // Hugetlbfs mapping intended
 ////////////////////////////////////////////////////////////////////////////////////////////
-#if defined(GRID_CUDA) ||defined(GRID_HIP)  || defined(GRID_SYCL)
-
-//if defined(GRID_SYCL)
-#if 0
-void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
-{
-  void * ShmCommBuf ; 
-  assert(_ShmSetup==1);
-  assert(_ShmAlloc==0);
-
-  //////////////////////////////////////////////////////////////////////////////////////////////////////////
-  // allocate the pointer array for shared windows for our group
-  //////////////////////////////////////////////////////////////////////////////////////////////////////////
-  MPI_Barrier(WorldShmComm);
-  WorldShmCommBufs.resize(WorldShmSize);
-
-  ///////////////////////////////////////////////////////////////////////////////////////////////////////////
-  // Each MPI rank should allocate our own buffer
-  ///////////////////////////////////////////////////////////////////////////////////////////////////////////
-  ShmCommBuf = acceleratorAllocDevice(bytes);
-
-  if (ShmCommBuf == (void *)NULL ) {
-    std::cerr << " SharedMemoryMPI.cc acceleratorAllocDevice failed NULL pointer for " << bytes<<" bytes " << std::endl;
-    exit(EXIT_FAILURE);  
-  }
-
-  std::cout << WorldRank << Mheader " SharedMemoryMPI.cc acceleratorAllocDevice "<< bytes 
-	    << "bytes at "<< std::hex<< ShmCommBuf <<std::dec<<" for comms buffers " <<std::endl;
-
-  SharedMemoryZero(ShmCommBuf,bytes);
-
-  assert(WorldShmSize == 1);
-  for(int r=0;r<WorldShmSize;r++){
-    WorldShmCommBufs[r] = ShmCommBuf;
-  }
-  _ShmAllocBytes=bytes;
-  _ShmAlloc=1;
-}
-#endif
-
 #if defined(GRID_CUDA) ||defined(GRID_HIP)  || defined(GRID_SYCL)
 void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 {
@@ -574,6 +541,40 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
   ///////////////////////////////////////////////////////////////////////////////////////////////////////////
   // Each MPI rank should allocate our own buffer
   ///////////////////////////////////////////////////////////////////////////////////////////////////////////
+#ifndef ACCELERATOR_AWARE_MPI
+  printf("Host buffer allocate for GPU non-aware MPI\n");
+#if 0
+  HostCommBuf= acceleratorAllocHost(bytes);
+#else 
+  HostCommBuf= malloc(bytes); /// CHANGE THIS TO malloc_host
+#ifdef HAVE_NUMAIF_H
+  #warning "Moving host buffers to specific NUMA domain"
+  int numa;
+  char *numa_name=(char *)getenv("MPI_BUF_NUMA");
+  if(numa_name) {
+    unsigned long page_size = sysconf(_SC_PAGESIZE);
+    numa = atoi(numa_name);
+    unsigned long page_count = bytes/page_size;
+    std::vector<void *> pages(page_count);
+    std::vector<int>    nodes(page_count,numa);
+    std::vector<int>    status(page_count,-1);
+    for(unsigned long p=0;p<page_count;p++){
+      pages[p] =(void *) ((uint64_t) HostCommBuf + p*page_size);
+    }
+    int ret = move_pages(0,
+			 page_count,
+			 &pages[0],
+			 &nodes[0],
+			 &status[0],
+			 MPOL_MF_MOVE);
+    printf("Host buffer move to numa domain %d : move_pages returned %d\n",numa,ret);
+    if (ret) perror(" move_pages failed for reason:");
+  }
+#endif  
+  acceleratorPin(HostCommBuf,bytes);
+#endif  
+
+#endif  
   ShmCommBuf = acceleratorAllocDevice(bytes);
   if (ShmCommBuf == (void *)NULL ) {
     std::cerr << " SharedMemoryMPI.cc acceleratorAllocDevice failed NULL pointer for " << bytes<<" bytes " << std::endl;
@@ -604,8 +605,8 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 #ifdef GRID_SYCL_LEVEL_ZERO_IPC
     typedef struct { int fd; pid_t pid ; ze_ipc_mem_handle_t ze; } clone_mem_t;
 
-    auto zeDevice    = cl::sycl::get_native<cl::sycl::backend::ext_oneapi_level_zero>(theGridAccelerator->get_device());
-    auto zeContext   = cl::sycl::get_native<cl::sycl::backend::ext_oneapi_level_zero>(theGridAccelerator->get_context());
+    auto zeDevice    = sycl::get_native<sycl::backend::ext_oneapi_level_zero>(theGridAccelerator->get_device());
+    auto zeContext   = sycl::get_native<sycl::backend::ext_oneapi_level_zero>(theGridAccelerator->get_context());
       
     ze_ipc_mem_handle_t ihandle;
     clone_mem_t handle;
@@ -738,7 +739,6 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
   _ShmAllocBytes=bytes;
   _ShmAlloc=1;
 }
-#endif
 
 #else 
 #ifdef GRID_MPI3_SHMMMAP
@@ -962,6 +962,12 @@ void SharedMemory::SetCommunicator(Grid_MPI_Comm comm)
   }
   ShmBufferFreeAll();
 
+#ifndef ACCELERATOR_AWARE_MPI
+  host_heap_size = heap_size;
+  HostCommBuf= GlobalSharedMemory::HostCommBuf;
+  HostBufferFreeAll();
+#endif  
+
   /////////////////////////////////////////////////////////////////////
   // find comm ranks in our SHM group (i.e. which ranks are on our node)
   /////////////////////////////////////////////////////////////////////

Grid/cshift/Cshift.h

@@ -51,7 +51,6 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #endif 
 
 NAMESPACE_BEGIN(Grid);
-
 template<class Expression,typename std::enable_if<is_lattice_expr<Expression>::value,void>::type * = nullptr> 
 auto Cshift(const Expression &expr,int dim,int shift)  -> decltype(closure(expr)) 
 {

Grid/cshift/Cshift_common.h

@@ -29,13 +29,28 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 
 NAMESPACE_BEGIN(Grid);
 
-extern Vector<std::pair<int,int> > Cshift_table; 
+extern std::vector<std::pair<int,int> > Cshift_table; 
+extern deviceVector<std::pair<int,int> > Cshift_table_device; 
 
+inline std::pair<int,int> *MapCshiftTable(void)
+{
+  // GPU version
+  uint64_t sz=Cshift_table.size();
+  if (Cshift_table_device.size()!=sz )    {
+    Cshift_table_device.resize(sz);
+  }
+  acceleratorCopyToDevice((void *)&Cshift_table[0],
+			  (void *)&Cshift_table_device[0],
+			  sizeof(Cshift_table[0])*sz);
+
+  return &Cshift_table_device[0];
+  // CPU version use identify map
+}
 ///////////////////////////////////////////////////////////////////
 // Gather for when there is no need to SIMD split 
 ///////////////////////////////////////////////////////////////////
 template<class vobj> void 
-Gather_plane_simple (const Lattice<vobj> &rhs,cshiftVector<vobj> &buffer,int dimension,int plane,int cbmask, int off=0)
+Gather_plane_simple (const Lattice<vobj> &rhs,deviceVector<vobj> &buffer,int dimension,int plane,int cbmask, int off=0)
 {
   int rd = rhs.Grid()->_rdimensions[dimension];
 
@@ -74,18 +89,11 @@ Gather_plane_simple (const Lattice<vobj> &rhs,cshiftVector<vobj> &buffer,int dim
   }
   {
     auto buffer_p = & buffer[0];
-    auto table = &Cshift_table[0];
-#ifdef ACCELERATOR_CSHIFT    
+    auto table = MapCshiftTable();
     autoView(rhs_v , rhs, AcceleratorRead);
     accelerator_for(i,ent,vobj::Nsimd(),{
 	coalescedWrite(buffer_p[table[i].first],coalescedRead(rhs_v[table[i].second]));
     });
-#else
-    autoView(rhs_v , rhs, CpuRead);
-    thread_for(i,ent,{
-      buffer_p[table[i].first]=rhs_v[table[i].second];
-    });
-#endif
   }
 }
 
@@ -110,7 +118,6 @@ Gather_plane_extract(const Lattice<vobj> &rhs,
   int n1=rhs.Grid()->_slice_stride[dimension];
 
   if ( cbmask ==0x3){
-#ifdef ACCELERATOR_CSHIFT
     autoView(rhs_v , rhs, AcceleratorRead);
     accelerator_for(nn,e1*e2,1,{
 	int n = nn%e1;
@@ -121,21 +128,10 @@ Gather_plane_extract(const Lattice<vobj> &rhs,
 	vobj temp =rhs_v[so+o+b];
 	extract<vobj>(temp,pointers,offset);
       });
-#else
-    autoView(rhs_v , rhs, CpuRead);
-    thread_for2d(n,e1,b,e2,{
-	int o      =   n*n1;
-	int offset = b+n*e2;
-	
-	vobj temp =rhs_v[so+o+b];
-	extract<vobj>(temp,pointers,offset);
-      });
-#endif
   } else { 
     Coordinate rdim=rhs.Grid()->_rdimensions;
     Coordinate cdm =rhs.Grid()->_checker_dim_mask;
     std::cout << " Dense packed buffer WARNING " <<std::endl; // Does this get called twice once for each cb?
-#ifdef ACCELERATOR_CSHIFT    
     autoView(rhs_v , rhs, AcceleratorRead);
     accelerator_for(nn,e1*e2,1,{
 	int n = nn%e1;
@@ -156,33 +152,13 @@ Gather_plane_extract(const Lattice<vobj> &rhs,
 	  extract<vobj>(temp,pointers,offset);
 	}
       });
-#else
-    autoView(rhs_v , rhs, CpuRead);
-    thread_for2d(n,e1,b,e2,{
-
-	Coordinate coor;
-
-	int o=n*n1;
-	int oindex = o+b;
-
-       	int cb = RedBlackCheckerBoardFromOindex(oindex, rdim, cdm);
-
-	int ocb=1<<cb;
-	int offset = b+n*e2;
-
-	if ( ocb & cbmask ) {
-	  vobj temp =rhs_v[so+o+b];
-	  extract<vobj>(temp,pointers,offset);
-	}
-      });
-#endif
   }
 }
 
 //////////////////////////////////////////////////////
 // Scatter for when there is no need to SIMD split
 //////////////////////////////////////////////////////
-template<class vobj> void Scatter_plane_simple (Lattice<vobj> &rhs,cshiftVector<vobj> &buffer, int dimension,int plane,int cbmask)
+template<class vobj> void Scatter_plane_simple (Lattice<vobj> &rhs,deviceVector<vobj> &buffer, int dimension,int plane,int cbmask)
 {
   int rd = rhs.Grid()->_rdimensions[dimension];
 
@@ -225,18 +201,11 @@ template<class vobj> void Scatter_plane_simple (Lattice<vobj> &rhs,cshiftVector<
   
   {
     auto buffer_p = & buffer[0];
-    auto table = &Cshift_table[0];
-#ifdef ACCELERATOR_CSHIFT    
+    auto table = MapCshiftTable();
     autoView( rhs_v, rhs, AcceleratorWrite);
     accelerator_for(i,ent,vobj::Nsimd(),{
 	coalescedWrite(rhs_v[table[i].first],coalescedRead(buffer_p[table[i].second]));
     });
-#else
-    autoView( rhs_v, rhs, CpuWrite);
-    thread_for(i,ent,{
-      rhs_v[table[i].first]=buffer_p[table[i].second];
-    });
-#endif
   }
 }
 
@@ -259,7 +228,6 @@ template<class vobj> void Scatter_plane_merge(Lattice<vobj> &rhs,ExtractPointerA
   if(cbmask ==0x3 ) {
     int _slice_stride = rhs.Grid()->_slice_stride[dimension];
     int _slice_block = rhs.Grid()->_slice_block[dimension];
-#ifdef ACCELERATOR_CSHIFT    
     autoView( rhs_v , rhs, AcceleratorWrite);
     accelerator_for(nn,e1*e2,1,{
 	int n = nn%e1;
@@ -268,14 +236,6 @@ template<class vobj> void Scatter_plane_merge(Lattice<vobj> &rhs,ExtractPointerA
 	int offset = b+n*_slice_block;
 	merge(rhs_v[so+o+b],pointers,offset);
       });
-#else
-    autoView( rhs_v , rhs, CpuWrite);
-    thread_for2d(n,e1,b,e2,{
-	int o      = n*_slice_stride;
-	int offset = b+n*_slice_block;
-	merge(rhs_v[so+o+b],pointers,offset);
-    });
-#endif
   } else { 
 
     // Case of SIMD split AND checker dim cannot currently be hit, except in 
@@ -297,30 +257,6 @@ template<class vobj> void Scatter_plane_merge(Lattice<vobj> &rhs,ExtractPointerA
   }
 }
 
-#if (defined(GRID_CUDA) || defined(GRID_HIP)) && defined(ACCELERATOR_CSHIFT)
-
-template <typename T>
-T iDivUp(T a, T b) // Round a / b to nearest higher integer value
-{ return (a % b != 0) ? (a / b + 1) : (a / b); }
-
-template <typename T>
-__global__ void populate_Cshift_table(T* vector, T lo, T ro, T e1, T e2, T stride)
-{
-    int idx = blockIdx.x*blockDim.x + threadIdx.x;
-    if (idx >= e1*e2) return;
-
-    int n, b, o;
-
-    n = idx / e2;
-    b = idx % e2;
-    o = n*stride + b;
-
-    vector[2*idx + 0] = lo + o;
-    vector[2*idx + 1] = ro + o;
-}
-
-#endif
-
 //////////////////////////////////////////////////////
 // local to node block strided copies
 //////////////////////////////////////////////////////
@@ -345,20 +281,12 @@ template<class vobj> void Copy_plane(Lattice<vobj>& lhs,const Lattice<vobj> &rhs
   int ent=0;
 
   if(cbmask == 0x3 ){
-#if (defined(GRID_CUDA) || defined(GRID_HIP)) && defined(ACCELERATOR_CSHIFT)
-    ent = e1*e2;
-    dim3 blockSize(acceleratorThreads());
-    dim3 gridSize(iDivUp((unsigned int)ent, blockSize.x));
-    populate_Cshift_table<<<gridSize, blockSize>>>(&Cshift_table[0].first, lo, ro, e1, e2, stride);
-    accelerator_barrier();
-#else
     for(int n=0;n<e1;n++){
       for(int b=0;b<e2;b++){
         int o =n*stride+b;
 	Cshift_table[ent++] = std::pair<int,int>(lo+o,ro+o);
       }
     }
-#endif
   } else { 
     for(int n=0;n<e1;n++){
       for(int b=0;b<e2;b++){
@@ -372,20 +300,12 @@ template<class vobj> void Copy_plane(Lattice<vobj>& lhs,const Lattice<vobj> &rhs
   }
 
   {
-    auto table = &Cshift_table[0];
-#ifdef ACCELERATOR_CSHIFT    
+    auto table = MapCshiftTable();
     autoView(rhs_v , rhs, AcceleratorRead);
     autoView(lhs_v , lhs, AcceleratorWrite);
     accelerator_for(i,ent,vobj::Nsimd(),{
       coalescedWrite(lhs_v[table[i].first],coalescedRead(rhs_v[table[i].second]));
     });
-#else
-    autoView(rhs_v , rhs, CpuRead);
-    autoView(lhs_v , lhs, CpuWrite);
-    thread_for(i,ent,{
-      lhs_v[table[i].first]=rhs_v[table[i].second];
-    });
-#endif
   }
 }
 
@@ -409,19 +329,11 @@ template<class vobj> void Copy_plane_permute(Lattice<vobj>& lhs,const Lattice<vo
   int ent=0;
 
   if ( cbmask == 0x3 ) {
-#if (defined(GRID_CUDA) || defined(GRID_HIP)) && defined(ACCELERATOR_CSHIFT)
-    ent = e1*e2;
-    dim3 blockSize(acceleratorThreads());
-    dim3 gridSize(iDivUp((unsigned int)ent, blockSize.x));
-    populate_Cshift_table<<<gridSize, blockSize>>>(&Cshift_table[0].first, lo, ro, e1, e2, stride);
-    accelerator_barrier();
-#else
     for(int n=0;n<e1;n++){
     for(int b=0;b<e2;b++){
       int o  =n*stride;
       Cshift_table[ent++] = std::pair<int,int>(lo+o+b,ro+o+b);
     }}
-#endif
   } else {
     for(int n=0;n<e1;n++){
     for(int b=0;b<e2;b++){
@@ -432,20 +344,12 @@ template<class vobj> void Copy_plane_permute(Lattice<vobj>& lhs,const Lattice<vo
   }
 
   {
-    auto table = &Cshift_table[0];
-#ifdef ACCELERATOR_CSHIFT    
+    auto table = MapCshiftTable();
     autoView( rhs_v, rhs, AcceleratorRead);
     autoView( lhs_v, lhs, AcceleratorWrite);
     accelerator_for(i,ent,1,{
       permute(lhs_v[table[i].first],rhs_v[table[i].second],permute_type);
     });
-#else
-    autoView( rhs_v, rhs, CpuRead);
-    autoView( lhs_v, lhs, CpuWrite);
-    thread_for(i,ent,{
-      permute(lhs_v[table[i].first],rhs_v[table[i].second],permute_type);
-    });
-#endif
   }
 }
 

Grid/cshift/Cshift_mpi.h

@@ -31,7 +31,7 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 
 
 NAMESPACE_BEGIN(Grid); 
-
+const int Cshift_verbose=0;
 template<class vobj> Lattice<vobj> Cshift(const Lattice<vobj> &rhs,int dimension,int shift)
 {
   typedef typename vobj::vector_type vector_type;
@@ -52,7 +52,8 @@ template<class vobj> Lattice<vobj> Cshift(const Lattice<vobj> &rhs,int dimension
   int comm_dim        = rhs.Grid()->_processors[dimension] >1 ;
   int splice_dim      = rhs.Grid()->_simd_layout[dimension]>1 && (comm_dim);
 
-
+  RealD t1,t0;
+  t0=usecond();
   if ( !comm_dim ) {
     //    std::cout << "CSHIFT: Cshift_local" <<std::endl;
     Cshift_local(ret,rhs,dimension,shift); // Handles checkerboarding
@@ -63,9 +64,11 @@ template<class vobj> Lattice<vobj> Cshift(const Lattice<vobj> &rhs,int dimension
     //    std::cout << "CSHIFT: Cshift_comms" <<std::endl;
     Cshift_comms(ret,rhs,dimension,shift);
   }
+  t1=usecond();
+  if(Cshift_verbose) std::cout << GridLogPerformance << "Cshift took "<< (t1-t0)/1e3 << " ms"<<std::endl;
   return ret;
 }
-
+#if 1
 template<class vobj> void Cshift_comms(Lattice<vobj>& ret,const Lattice<vobj> &rhs,int dimension,int shift)
 {
   int sshift[2];
@@ -91,7 +94,7 @@ template<class vobj> void Cshift_comms_simd(Lattice<vobj>& ret,const Lattice<vob
   sshift[0] = rhs.Grid()->CheckerBoardShiftForCB(rhs.Checkerboard(),dimension,shift,Even);
   sshift[1] = rhs.Grid()->CheckerBoardShiftForCB(rhs.Checkerboard(),dimension,shift,Odd);
 
-  //std::cout << "Cshift_comms_simd dim "<<dimension<<"cb "<<rhs.checkerboard<<"shift "<<shift<<" sshift " << sshift[0]<<" "<<sshift[1]<<std::endl;
+  //  std::cout << "Cshift_comms_simd dim "<<dimension<<"cb "<<rhs.Checkerboard()<<"shift "<<shift<<" sshift " << sshift[0]<<" "<<sshift[1]<<std::endl;
   if ( sshift[0] == sshift[1] ) {
     //    std::cout << "Single pass Cshift_comms" <<std::endl;
     Cshift_comms_simd(ret,rhs,dimension,shift,0x3);
@@ -101,8 +104,6 @@ template<class vobj> void Cshift_comms_simd(Lattice<vobj>& ret,const Lattice<vob
     Cshift_comms_simd(ret,rhs,dimension,shift,0x2);// both with block stride loop iteration
   }
 }
-#define ACCELERATOR_CSHIFT_NO_COPY
-#ifdef ACCELERATOR_CSHIFT_NO_COPY
 template<class vobj> void Cshift_comms(Lattice<vobj> &ret,const Lattice<vobj> &rhs,int dimension,int shift,int cbmask)
 {
   typedef typename vobj::vector_type vector_type;
@@ -122,21 +123,25 @@ template<class vobj> void Cshift_comms(Lattice<vobj> &ret,const Lattice<vobj> &r
   assert(shift<fd);
   
   int buffer_size = rhs.Grid()->_slice_nblock[dimension]*rhs.Grid()->_slice_block[dimension];
-  static cshiftVector<vobj> send_buf; send_buf.resize(buffer_size);
-  static cshiftVector<vobj> recv_buf; recv_buf.resize(buffer_size);
+  static deviceVector<vobj> send_buf; send_buf.resize(buffer_size);
+  static deviceVector<vobj> recv_buf; recv_buf.resize(buffer_size);
     
   int cb= (cbmask==0x2)? Odd : Even;
   int sshift= rhs.Grid()->CheckerBoardShiftForCB(rhs.Checkerboard(),dimension,shift,cb);
-
+  RealD tcopy=0.0;
+  RealD tgather=0.0;
+  RealD tscatter=0.0;
+  RealD tcomms=0.0;
+  uint64_t xbytes=0;
   for(int x=0;x<rd;x++){       
 
     int sx        =  (x+sshift)%rd;
     int comm_proc = ((x+sshift)/rd)%pd;
     
     if (comm_proc==0) {
-
+      tcopy-=usecond();
       Copy_plane(ret,rhs,dimension,x,sx,cbmask); 
-
+      tcopy+=usecond();
     } else {
 
       int words = buffer_size;
@@ -144,13 +149,16 @@ template<class vobj> void Cshift_comms(Lattice<vobj> &ret,const Lattice<vobj> &r
 
       int bytes = words * sizeof(vobj);
 
+      tgather-=usecond();
       Gather_plane_simple (rhs,send_buf,dimension,sx,cbmask);
+      tgather+=usecond();
 
       //      int rank           = grid->_processor;
       int recv_from_rank;
       int xmit_to_rank;
       grid->ShiftedRanks(dimension,comm_proc,xmit_to_rank,recv_from_rank);
       
+      tcomms-=usecond();
       grid->Barrier();
 
       grid->SendToRecvFrom((void *)&send_buf[0],
@@ -158,12 +166,22 @@ template<class vobj> void Cshift_comms(Lattice<vobj> &ret,const Lattice<vobj> &r
 			   (void *)&recv_buf[0],
 			   recv_from_rank,
 			   bytes);
-
+      xbytes+=bytes;
       grid->Barrier();
+      tcomms+=usecond();
 
+      tscatter-=usecond();
       Scatter_plane_simple (ret,recv_buf,dimension,x,cbmask);
+      tscatter+=usecond();
     }
   }
+  if (Cshift_verbose){
+    std::cout << GridLogPerformance << " Cshift copy    "<<tcopy/1e3<<" ms"<<std::endl;
+    std::cout << GridLogPerformance << " Cshift gather  "<<tgather/1e3<<" ms"<<std::endl;
+    std::cout << GridLogPerformance << " Cshift scatter "<<tscatter/1e3<<" ms"<<std::endl;
+    std::cout << GridLogPerformance << " Cshift comm    "<<tcomms/1e3<<" ms"<<std::endl;
+    std::cout << GridLogPerformance << " Cshift BW      "<<(2.0*xbytes)/tcomms<<" MB/s "<<2*xbytes<< " Bytes "<<std::endl;
+  }
 }
 
 template<class vobj> void  Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vobj> &rhs,int dimension,int shift,int cbmask)
@@ -190,6 +208,12 @@ template<class vobj> void  Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vo
   assert(shift>=0);
   assert(shift<fd);
 
+  RealD tcopy=0.0;
+  RealD tgather=0.0;
+  RealD tscatter=0.0;
+  RealD tcomms=0.0;
+  uint64_t xbytes=0;
+  
   int permute_type=grid->PermuteType(dimension);
 
   ///////////////////////////////////////////////
@@ -198,8 +222,8 @@ template<class vobj> void  Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vo
   int buffer_size = grid->_slice_nblock[dimension]*grid->_slice_block[dimension];
   //  int words = sizeof(vobj)/sizeof(vector_type);
 
-  static std::vector<cshiftVector<scalar_object> >  send_buf_extract; send_buf_extract.resize(Nsimd);
-  static std::vector<cshiftVector<scalar_object> >  recv_buf_extract; recv_buf_extract.resize(Nsimd);
+  static std::vector<deviceVector<scalar_object> >  send_buf_extract; send_buf_extract.resize(Nsimd);
+  static std::vector<deviceVector<scalar_object> >  recv_buf_extract; recv_buf_extract.resize(Nsimd);
   scalar_object *  recv_buf_extract_mpi;
   scalar_object *  send_buf_extract_mpi;
  
@@ -227,7 +251,9 @@ template<class vobj> void  Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vo
       pointers[i] = &send_buf_extract[i][0];
     }
     int sx   = (x+sshift)%rd;
+    tgather-=usecond();
     Gather_plane_extract(rhs,pointers,dimension,sx,cbmask);
+    tgather+=usecond();
 
     for(int i=0;i<Nsimd;i++){
       
@@ -252,6 +278,7 @@ template<class vobj> void  Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vo
       if(nbr_proc){
 	grid->ShiftedRanks(dimension,nbr_proc,xmit_to_rank,recv_from_rank); 
 
+	tcomms-=usecond();
 	grid->Barrier();
 
 	send_buf_extract_mpi = &send_buf_extract[nbr_lane][0];
@@ -262,7 +289,9 @@ template<class vobj> void  Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vo
 			     recv_from_rank,
 			     bytes);
 
+	xbytes+=bytes;
 	grid->Barrier();
+	tcomms+=usecond();
 
 	rpointers[i] = &recv_buf_extract[i][0];
       } else { 
@@ -270,9 +299,17 @@ template<class vobj> void  Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vo
       }
 
     }
+    tscatter-=usecond();
     Scatter_plane_merge(ret,rpointers,dimension,x,cbmask);
+    tscatter+=usecond();
+  }
+  if(Cshift_verbose){
+    std::cout << GridLogPerformance << " Cshift (s) copy    "<<tcopy/1e3<<" ms"<<std::endl;
+    std::cout << GridLogPerformance << " Cshift (s) gather  "<<tgather/1e3<<" ms"<<std::endl;
+    std::cout << GridLogPerformance << " Cshift (s) scatter "<<tscatter/1e3<<" ms"<<std::endl;
+    std::cout << GridLogPerformance << " Cshift (s) comm    "<<tcomms/1e3<<" ms"<<std::endl;
+    std::cout << GridLogPerformance << " Cshift BW      "<<(2.0*xbytes)/tcomms<<" MB/s "<<2*xbytes<< " Bytes "<<std::endl;
   }
-
 }
 #else 
 template<class vobj> void Cshift_comms(Lattice<vobj> &ret,const Lattice<vobj> &rhs,int dimension,int shift,int cbmask)
@@ -292,6 +329,11 @@ template<class vobj> void Cshift_comms(Lattice<vobj> &ret,const Lattice<vobj> &r
   assert(comm_dim==1);
   assert(shift>=0);
   assert(shift<fd);
+  RealD tcopy=0.0;
+  RealD tgather=0.0;
+  RealD tscatter=0.0;
+  RealD tcomms=0.0;
+  uint64_t xbytes=0;
   
   int buffer_size = rhs.Grid()->_slice_nblock[dimension]*rhs.Grid()->_slice_block[dimension];
   static cshiftVector<vobj> send_buf_v; send_buf_v.resize(buffer_size);
@@ -315,7 +357,9 @@ template<class vobj> void Cshift_comms(Lattice<vobj> &ret,const Lattice<vobj> &r
     
     if (comm_proc==0) {
 
+      tcopy-=usecond();
       Copy_plane(ret,rhs,dimension,x,sx,cbmask); 
+      tcopy+=usecond();
 
     } else {
 
@@ -324,7 +368,9 @@ template<class vobj> void Cshift_comms(Lattice<vobj> &ret,const Lattice<vobj> &r
 
       int bytes = words * sizeof(vobj);
 
+      tgather-=usecond();
       Gather_plane_simple (rhs,send_buf_v,dimension,sx,cbmask);
+      tgather+=usecond();
 
       //      int rank           = grid->_processor;
       int recv_from_rank;
@@ -332,7 +378,8 @@ template<class vobj> void Cshift_comms(Lattice<vobj> &ret,const Lattice<vobj> &r
       grid->ShiftedRanks(dimension,comm_proc,xmit_to_rank,recv_from_rank);
 
 
-      grid->Barrier();
+      tcomms-=usecond();
+      //      grid->Barrier();
 
       acceleratorCopyDeviceToDevice((void *)&send_buf_v[0],(void *)&send_buf[0],bytes);
       grid->SendToRecvFrom((void *)&send_buf[0],
@@ -340,13 +387,24 @@ template<class vobj> void Cshift_comms(Lattice<vobj> &ret,const Lattice<vobj> &r
 			   (void *)&recv_buf[0],
 			   recv_from_rank,
 			   bytes);
+      xbytes+=bytes;
       acceleratorCopyDeviceToDevice((void *)&recv_buf[0],(void *)&recv_buf_v[0],bytes);
 
-      grid->Barrier();
+      //      grid->Barrier();
+      tcomms+=usecond();
 
+      tscatter-=usecond();
       Scatter_plane_simple (ret,recv_buf_v,dimension,x,cbmask);
+      tscatter+=usecond();
     }
   }
+  if(Cshift_verbose){
+    std::cout << GridLogPerformance << " Cshift copy    "<<tcopy/1e3<<" ms"<<std::endl;
+    std::cout << GridLogPerformance << " Cshift gather  "<<tgather/1e3<<" ms"<<std::endl;
+    std::cout << GridLogPerformance << " Cshift scatter "<<tscatter/1e3<<" ms"<<std::endl;
+    std::cout << GridLogPerformance << " Cshift comm    "<<tcomms/1e3<<" ms"<<std::endl;
+    std::cout << GridLogPerformance << " Cshift BW      "<<(2.0*xbytes)/tcomms<<" MB/s "<<2*xbytes<< " Bytes "<<std::endl;
+  }
 }
 
 template<class vobj> void  Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vobj> &rhs,int dimension,int shift,int cbmask)
@@ -372,6 +430,11 @@ template<class vobj> void  Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vo
   assert(simd_layout==2);
   assert(shift>=0);
   assert(shift<fd);
+  RealD tcopy=0.0;
+  RealD tgather=0.0;
+  RealD tscatter=0.0;
+  RealD tcomms=0.0;
+  uint64_t xbytes=0;
 
   int permute_type=grid->PermuteType(dimension);
 
@@ -414,8 +477,10 @@ template<class vobj> void  Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vo
     for(int i=0;i<Nsimd;i++){       
       pointers[i] = &send_buf_extract[i][0];
     }
+    tgather-=usecond();
     int sx   = (x+sshift)%rd;
     Gather_plane_extract(rhs,pointers,dimension,sx,cbmask);
+    tgather+=usecond();
 
     for(int i=0;i<Nsimd;i++){
       
@@ -440,7 +505,8 @@ template<class vobj> void  Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vo
       if(nbr_proc){
 	grid->ShiftedRanks(dimension,nbr_proc,xmit_to_rank,recv_from_rank); 
 
-	grid->Barrier();
+	tcomms-=usecond();
+	//	grid->Barrier();
 
 	acceleratorCopyDeviceToDevice((void *)&send_buf_extract[nbr_lane][0],(void *)send_buf_extract_mpi,bytes);
 	grid->SendToRecvFrom((void *)send_buf_extract_mpi,
@@ -449,19 +515,31 @@ template<class vobj> void  Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vo
 			     recv_from_rank,
 			     bytes);
 	acceleratorCopyDeviceToDevice((void *)recv_buf_extract_mpi,(void *)&recv_buf_extract[i][0],bytes);
+	xbytes+=bytes;
 
-	grid->Barrier();
+	//	grid->Barrier();
+	tcomms+=usecond();
 	rpointers[i] = &recv_buf_extract[i][0];
       } else { 
 	rpointers[i] = &send_buf_extract[nbr_lane][0];
       }
 
     }
+    tscatter-=usecond();
     Scatter_plane_merge(ret,rpointers,dimension,x,cbmask);
-  }
+    tscatter+=usecond();
 
   }
+  if(Cshift_verbose){
+    std::cout << GridLogPerformance << " Cshift (s) copy    "<<tcopy/1e3<<" ms"<<std::endl;
+    std::cout << GridLogPerformance << " Cshift (s) gather  "<<tgather/1e3<<" ms"<<std::endl;
+    std::cout << GridLogPerformance << " Cshift (s) scatter "<<tscatter/1e3<<" ms"<<std::endl;
+    std::cout << GridLogPerformance << " Cshift (s) comm    "<<tcomms/1e3<<" ms"<<std::endl;
+    std::cout << GridLogPerformance << " Cshift BW      "<<(2.0*xbytes)/tcomms<<" MB/s"<<std::endl;
+  }
+}
 #endif
+
 NAMESPACE_END(Grid); 
 
 #endif

Grid/cshift/Cshift_table.cc

@@ -1,4 +1,5 @@
 #include <Grid/GridCore.h>       
 NAMESPACE_BEGIN(Grid);
-Vector<std::pair<int,int> > Cshift_table; 
+std::vector<std::pair<int,int> > Cshift_table; 
+deviceVector<std::pair<int,int> > Cshift_table_device; 
 NAMESPACE_END(Grid);

Grid/lattice/Lattice.h

@@ -35,6 +35,7 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #include <Grid/lattice/Lattice_transpose.h>
 #include <Grid/lattice/Lattice_local.h>
 #include <Grid/lattice/Lattice_reduction.h>
+#include <Grid/lattice/Lattice_crc.h>
 #include <Grid/lattice/Lattice_peekpoke.h>
 #include <Grid/lattice/Lattice_reality.h>
 #include <Grid/lattice/Lattice_real_imag.h>
@@ -46,5 +47,4 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #include <Grid/lattice/Lattice_unary.h>
 #include <Grid/lattice/Lattice_transfer.h>
 #include <Grid/lattice/Lattice_basis.h>
-#include <Grid/lattice/Lattice_crc.h>
 #include <Grid/lattice/PaddedCell.h>

Grid/lattice/Lattice_arith.h

@@ -257,18 +257,68 @@ void axpby(Lattice<vobj> &ret,sobj a,sobj b,const Lattice<vobj> &x,const Lattice
   });
 }
 
+#define FAST_AXPY_NORM
 template<class sobj,class vobj> inline
 RealD axpy_norm(Lattice<vobj> &ret,sobj a,const Lattice<vobj> &x,const Lattice<vobj> &y)
 {
   GRID_TRACE("axpy_norm");
+#ifdef FAST_AXPY_NORM
   return axpy_norm_fast(ret,a,x,y);
+#else
+  ret = a*x+y;
+  RealD nn=norm2(ret);
+  return nn;
+#endif
 }
 template<class sobj,class vobj> inline
 RealD axpby_norm(Lattice<vobj> &ret,sobj a,sobj b,const Lattice<vobj> &x,const Lattice<vobj> &y)
 {
   GRID_TRACE("axpby_norm");
+#ifdef FAST_AXPY_NORM
   return axpby_norm_fast(ret,a,b,x,y);
+#else
+  ret = a*x+b*y;
+  RealD nn=norm2(ret);
+  return nn;
+#endif
 }
 
+/// Trace product
+template<class obj> auto traceProduct(const Lattice<obj> &rhs_1,const Lattice<obj> &rhs_2)
+  -> Lattice<decltype(trace(obj()))>
+{
+  typedef decltype(trace(obj())) robj;
+  Lattice<robj> ret_i(rhs_1.Grid());
+  autoView( rhs1 , rhs_1, AcceleratorRead);
+  autoView( rhs2 , rhs_2, AcceleratorRead);
+  autoView( ret , ret_i, AcceleratorWrite);
+  ret.Checkerboard() = rhs_1.Checkerboard();
+  accelerator_for(ss,rhs1.size(),obj::Nsimd(),{
+      coalescedWrite(ret[ss],traceProduct(rhs1(ss),rhs2(ss)));
+  });
+  return ret_i;
+}
+
+template<class obj1,class obj2> auto traceProduct(const Lattice<obj1> &rhs_1,const obj2 &rhs2)
+  -> Lattice<decltype(trace(obj1()))>
+{
+  typedef decltype(trace(obj1())) robj;
+  Lattice<robj> ret_i(rhs_1.Grid());
+  autoView( rhs1 , rhs_1, AcceleratorRead);
+  autoView( ret , ret_i, AcceleratorWrite);
+  ret.Checkerboard() = rhs_1.Checkerboard();
+  accelerator_for(ss,rhs1.size(),obj1::Nsimd(),{
+      coalescedWrite(ret[ss],traceProduct(rhs1(ss),rhs2));
+  });
+  return ret_i;
+}
+template<class obj1,class obj2> auto traceProduct(const obj2 &rhs_2,const Lattice<obj1> &rhs_1)
+  -> Lattice<decltype(trace(obj1()))>
+{
+  return traceProduct(rhs_1,rhs_2);
+}
+
+
+
 NAMESPACE_END(Grid);
 #endif

Grid/lattice/Lattice_base.h

@@ -234,10 +234,23 @@ public:
   }
 
   template<class sobj> inline Lattice<vobj> & operator = (const sobj & r){
+    vobj vtmp;
+    vtmp = r;
+#if 0
+    deviceVector<vobj> vvtmp(1);
+    acceleratorPut(vvtmp[0],vtmp);
+    vobj *vvtmp_p = & vvtmp[0];
+    auto me  = View(AcceleratorWrite);
+    accelerator_for(ss,me.size(),vobj::Nsimd(),{
+	auto stmp=coalescedRead(*vvtmp_p);
+	coalescedWrite(me[ss],stmp);
+    });
+#else    
     auto me  = View(CpuWrite);
     thread_for(ss,me.size(),{
        me[ss]= r;
       });
+#endif    
     me.ViewClose();
     return *this;
   }

Grid/lattice/Lattice_basis.h

@@ -53,36 +53,19 @@ void basisRotate(VField &basis,Matrix& Qt,int j0, int j1, int k0,int k1,int Nm)
   typedef decltype(basis[0]) Field;
   typedef decltype(basis[0].View(AcceleratorRead)) View;
 
-  Vector<View> basis_v; basis_v.reserve(basis.size());
-  typedef typename std::remove_reference<decltype(basis_v[0][0])>::type vobj;
+  hostVector<View>  h_basis_v(basis.size());
+  deviceVector<View> d_basis_v(basis.size());
+  typedef typename std::remove_reference<decltype(h_basis_v[0][0])>::type vobj;
   typedef typename std::remove_reference<decltype(Qt(0,0))>::type Coeff_t;
+
   GridBase* grid = basis[0].Grid();
       
   for(int k=0;k<basis.size();k++){
-    basis_v.push_back(basis[k].View(AcceleratorWrite));
+    h_basis_v[k] = basis[k].View(AcceleratorWrite);
+    acceleratorPut(d_basis_v[k],h_basis_v[k]);
   }
 
-#if ( (!defined(GRID_CUDA)) )
-  int max_threads = thread_max();
-  Vector < vobj > Bt(Nm * max_threads);
-  thread_region
-    {
-      vobj* B = &Bt[Nm * thread_num()];
-      thread_for_in_region(ss, grid->oSites(),{
-	  for(int j=j0; j<j1; ++j) B[j]=0.;
-      
-	  for(int j=j0; j<j1; ++j){
-	    for(int k=k0; k<k1; ++k){
-	      B[j] +=Qt(j,k) * basis_v[k][ss];
-	    }
-	  }
-	  for(int j=j0; j<j1; ++j){
-	    basis_v[j][ss] = B[j];
-	  }
-	});
-    }
-#else
-  View *basis_vp = &basis_v[0];
+  View *basis_vp = &d_basis_v[0];
 
   int nrot = j1-j0;
   if (!nrot) // edge case not handled gracefully by Cuda
@@ -91,17 +74,19 @@ void basisRotate(VField &basis,Matrix& Qt,int j0, int j1, int k0,int k1,int Nm)
   uint64_t oSites   =grid->oSites();
   uint64_t siteBlock=(grid->oSites()+nrot-1)/nrot; // Maximum 1 additional vector overhead
 
-  Vector <vobj> Bt(siteBlock * nrot); 
+  deviceVector <vobj> Bt(siteBlock * nrot); 
   auto Bp=&Bt[0];
 
   // GPU readable copy of matrix
-  Vector<Coeff_t> Qt_jv(Nm*Nm);
+  hostVector<Coeff_t> h_Qt_jv(Nm*Nm);
+  deviceVector<Coeff_t> Qt_jv(Nm*Nm);
   Coeff_t *Qt_p = & Qt_jv[0];
   thread_for(i,Nm*Nm,{
       int j = i/Nm;
       int k = i%Nm;
-      Qt_p[i]=Qt(j,k);
+      h_Qt_jv[i]=Qt(j,k);
   });
+  acceleratorCopyToDevice(&h_Qt_jv[0],Qt_p,Nm*Nm*sizeof(Coeff_t));
 
   // Block the loop to keep storage footprint down
   for(uint64_t s=0;s<oSites;s+=siteBlock){
@@ -137,9 +122,8 @@ void basisRotate(VField &basis,Matrix& Qt,int j0, int j1, int k0,int k1,int Nm)
 	coalescedWrite(basis_vp[jj][sss],coalescedRead(Bp[ss*nrot+j]));
       });
   }
-#endif
 
-  for(int k=0;k<basis.size();k++) basis_v[k].ViewClose();
+  for(int k=0;k<basis.size();k++) h_basis_v[k].ViewClose();
 }
 
 // Extract a single rotated vector
@@ -152,16 +136,19 @@ void basisRotateJ(Field &result,std::vector<Field> &basis,Eigen::MatrixXd& Qt,in
 
   result.Checkerboard() = basis[0].Checkerboard();
 
-  Vector<View> basis_v; basis_v.reserve(basis.size());
+  hostVector<View>  h_basis_v(basis.size());
+  deviceVector<View> d_basis_v(basis.size());
   for(int k=0;k<basis.size();k++){
-    basis_v.push_back(basis[k].View(AcceleratorRead));
+    h_basis_v[k]=basis[k].View(AcceleratorRead);
+    acceleratorPut(d_basis_v[k],h_basis_v[k]);
   }
-  vobj zz=Zero();
-  Vector<double> Qt_jv(Nm);
-  double * Qt_j = & Qt_jv[0];
-  for(int k=0;k<Nm;++k) Qt_j[k]=Qt(j,k);
 
-  auto basis_vp=& basis_v[0];
+  vobj zz=Zero();
+  deviceVector<double> Qt_jv(Nm);
+  double * Qt_j = & Qt_jv[0];
+  for(int k=0;k<Nm;++k) acceleratorPut(Qt_j[k],Qt(j,k));
+
+  auto basis_vp=& d_basis_v[0];
   autoView(result_v,result,AcceleratorWrite);
   accelerator_for(ss, grid->oSites(),vobj::Nsimd(),{
     vobj zzz=Zero();
@@ -171,7 +158,7 @@ void basisRotateJ(Field &result,std::vector<Field> &basis,Eigen::MatrixXd& Qt,in
     }
     coalescedWrite(result_v[ss], B);
   });
-  for(int k=0;k<basis.size();k++) basis_v[k].ViewClose();
+  for(int k=0;k<basis.size();k++) h_basis_v[k].ViewClose();
 }
 
 template<class Field>

Grid/lattice/Lattice_crc.h

@@ -29,7 +29,7 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 
 NAMESPACE_BEGIN(Grid);
 
-template<class vobj> void DumpSliceNorm(std::string s,Lattice<vobj> &f,int mu=-1)
+template<class vobj> void DumpSliceNorm(std::string s,const Lattice<vobj> &f,int mu=-1)
 {
   auto ff = localNorm2(f);
   if ( mu==-1 ) mu = f.Grid()->Nd()-1;
@@ -42,13 +42,13 @@ template<class vobj> void DumpSliceNorm(std::string s,Lattice<vobj> &f,int mu=-1
   }
 }
 
-template<class vobj> uint32_t crc(Lattice<vobj> & buf)
+template<class vobj> uint32_t crc(const Lattice<vobj> & buf)
 {
   autoView( buf_v , buf, CpuRead);
   return ::crc32(0L,(unsigned char *)&buf_v[0],(size_t)sizeof(vobj)*buf.oSites());
 }
 
-#define CRC(U) std::cout << "FingerPrint "<<__FILE__ <<" "<< __LINE__ <<" "<< #U <<" "<<crc(U)<<std::endl;
+#define CRC(U) std::cerr << "FingerPrint "<<__FILE__ <<" "<< __LINE__ <<" "<< #U <<" "<<crc(U)<<std::endl;
 
 NAMESPACE_END(Grid);
 

Grid/lattice/Lattice_peekpoke.h

@@ -165,7 +165,7 @@ inline void peekLocalSite(sobj &s,const LatticeView<vobj> &l,Coordinate &site)
 
   int Nsimd = grid->Nsimd();
 
-  assert( l.Checkerboard()== grid->CheckerBoard(site));
+  //  assert( l.Checkerboard()== grid->CheckerBoard(site));
   assert( sizeof(sobj)*Nsimd == sizeof(vobj));
 
   static const int words=sizeof(vobj)/sizeof(vector_type);
@@ -179,7 +179,7 @@ inline void peekLocalSite(sobj &s,const LatticeView<vobj> &l,Coordinate &site)
   for(int w=0;w<words;w++){
     pt[w] = getlane(vp[w],idx);
   }
-      
+  //  std::cout << "peekLocalSite "<<site<<" "<<odx<<","<<idx<<" "<<s<<std::endl;
   return;
 };
 template<class vobj,class sobj>
@@ -202,7 +202,7 @@ inline void pokeLocalSite(const sobj &s,LatticeView<vobj> &l,Coordinate &site)
 
   int Nsimd = grid->Nsimd();
 
-  assert( l.Checkerboard()== grid->CheckerBoard(site));
+  //  assert( l.Checkerboard()== grid->CheckerBoard(site));
   assert( sizeof(sobj)*Nsimd == sizeof(vobj));
 
   static const int words=sizeof(vobj)/sizeof(vector_type);

Grid/lattice/Lattice_reduction.h

@@ -31,6 +31,7 @@ Author: Christoph Lehner <christoph@lhnr.de>
 #if defined(GRID_SYCL)
 #include <Grid/lattice/Lattice_reduction_sycl.h>
 #endif
+#include <Grid/lattice/Lattice_slicesum_core.h>
 
 NAMESPACE_BEGIN(Grid);
 
@@ -45,7 +46,7 @@ inline typename vobj::scalar_object sum_cpu(const vobj *arg, Integer osites)
   //  const int Nsimd = vobj::Nsimd();
   const int nthread = GridThread::GetThreads();
 
-  Vector<sobj> sumarray(nthread);
+  std::vector<sobj> sumarray(nthread);
   for(int i=0;i<nthread;i++){
     sumarray[i]=Zero();
   }
@@ -74,7 +75,7 @@ inline typename vobj::scalar_objectD sumD_cpu(const vobj *arg, Integer osites)
 
   const int nthread = GridThread::GetThreads();
 
-  Vector<sobj> sumarray(nthread);
+  std::vector<sobj> sumarray(nthread);
   for(int i=0;i<nthread;i++){
     sumarray[i]=Zero();
   }
@@ -203,6 +204,27 @@ template<class vobj> inline RealD norm2(const Lattice<vobj> &arg){
   return real(nrm); 
 }
 
+
+template<class Op,class T1>
+inline auto norm2(const LatticeUnaryExpression<Op,T1> & expr)  ->RealD
+{
+  return norm2(closure(expr));
+}
+
+template<class Op,class T1,class T2>
+inline auto norm2(const LatticeBinaryExpression<Op,T1,T2> & expr)      ->RealD
+{
+  return norm2(closure(expr));
+}
+
+
+template<class Op,class T1,class T2,class T3>
+inline auto norm2(const LatticeTrinaryExpression<Op,T1,T2,T3> & expr)      ->RealD
+{
+  return norm2(closure(expr));
+}
+
+
 //The global maximum of the site norm2
 template<class vobj> inline RealD maxLocalNorm2(const Lattice<vobj> &arg)
 {
@@ -242,24 +264,8 @@ inline ComplexD rankInnerProduct(const Lattice<vobj> &left,const Lattice<vobj> &
   const uint64_t sites = grid->oSites();
   
   // Might make all code paths go this way.
-#if 0
-  typedef decltype(innerProductD(vobj(),vobj())) inner_t;
-  Vector<inner_t> inner_tmp(sites);
-  auto inner_tmp_v = &inner_tmp[0];
-  {
-    autoView( left_v , left, AcceleratorRead);
-    autoView( right_v,right, AcceleratorRead);
-    // This code could read coalesce
-    // GPU - SIMT lane compliance...
-    accelerator_for( ss, sites, nsimd,{
-	auto x_l = left_v(ss);
-	auto y_l = right_v(ss);
-	coalescedWrite(inner_tmp_v[ss],innerProductD(x_l,y_l));
-    });
-  }
-#else
   typedef decltype(innerProduct(vobj(),vobj())) inner_t;
-  Vector<inner_t> inner_tmp(sites);
+  deviceVector<inner_t> inner_tmp(sites);
   auto inner_tmp_v = &inner_tmp[0];
     
   {
@@ -273,18 +279,57 @@ inline ComplexD rankInnerProduct(const Lattice<vobj> &left,const Lattice<vobj> &
 	coalescedWrite(inner_tmp_v[ss],innerProduct(x_l,y_l));
     });
   }
-#endif
   // This is in single precision and fails some tests
   auto anrm = sumD(inner_tmp_v,sites);  
   nrm = anrm;
   return nrm;
 }
 
+
 template<class vobj>
 inline ComplexD innerProduct(const Lattice<vobj> &left,const Lattice<vobj> &right) {
   GridBase *grid = left.Grid();
+
+  bool ok;
+#ifdef GRID_SYCL
+  uint64_t csum=0;
+  uint64_t csum2=0;
+  if ( FlightRecorder::LoggingMode != FlightRecorder::LoggingModeNone)
+  {
+    // Hack
+    // Fast integer xor checksum. Can also be used in comms now.
+    autoView(l_v,left,AcceleratorRead);
+    Integer words = left.Grid()->oSites()*sizeof(vobj)/sizeof(uint64_t);
+    uint64_t *base= (uint64_t *)&l_v[0];
+    csum=svm_xor(base,words);
+    ok = FlightRecorder::CsumLog(csum);
+    if ( !ok ) {
+      csum2=svm_xor(base,words);
+      std::cerr<< " Bad CSUM " << std::hex<< csum << " recomputed as "<<csum2<<std::dec<<std::endl;
+    } else {
+      //      csum2=svm_xor(base,words);
+      //      std::cerr<< " ok CSUM " << std::hex<< csum << " recomputed as "<<csum2<<std::dec<<std::endl;
+    }
+    assert(ok);
+  }
+#endif
+  FlightRecorder::StepLog("rank inner product");
   ComplexD nrm = rankInnerProduct(left,right);
+  //  ComplexD nrmck=nrm;
+  RealD local = real(nrm);
+  ok = FlightRecorder::NormLog(real(nrm));
+  if ( !ok ) {
+    ComplexD nrm2 = rankInnerProduct(left,right);
+    RealD local2 = real(nrm2);
+    std::cerr<< " Bad NORM " << local << " recomputed as "<<local2<<std::endl;
+    assert(ok);
+  }
+  FlightRecorder::StepLog("Start global sum");
+  //  grid->GlobalSumP2P(nrm);
   grid->GlobalSum(nrm);
+  FlightRecorder::StepLog("Finished global sum");
+  //  std::cout << " norm "<< nrm << " p2p norm "<<nrmck<<std::endl;
+  FlightRecorder::ReductionLog(local,real(nrm)); 
   return nrm;
 }
 
@@ -320,20 +365,9 @@ axpby_norm_fast(Lattice<vobj> &z,sobj a,sobj b,const Lattice<vobj> &x,const Latt
   autoView( x_v, x, AcceleratorRead);
   autoView( y_v, y, AcceleratorRead);
   autoView( z_v, z, AcceleratorWrite);
-#if 0
-  typedef decltype(innerProductD(x_v[0],y_v[0])) inner_t;
-  Vector<inner_t> inner_tmp(sites);
-  auto inner_tmp_v = &inner_tmp[0];
-
-  accelerator_for( ss, sites, nsimd,{
-      auto tmp = a*x_v(ss)+b*y_v(ss);
-      coalescedWrite(inner_tmp_v[ss],innerProductD(tmp,tmp));
-      coalescedWrite(z_v[ss],tmp);
-  });
-  nrm = real(TensorRemove(sum(inner_tmp_v,sites)));
-#else
   typedef decltype(innerProduct(x_v[0],y_v[0])) inner_t;
-  Vector<inner_t> inner_tmp(sites);
+  deviceVector<inner_t> inner_tmp;
+  inner_tmp.resize(sites);
   auto inner_tmp_v = &inner_tmp[0];
 
   accelerator_for( ss, sites, nsimd,{
@@ -341,9 +375,44 @@ axpby_norm_fast(Lattice<vobj> &z,sobj a,sobj b,const Lattice<vobj> &x,const Latt
       coalescedWrite(inner_tmp_v[ss],innerProduct(tmp,tmp));
       coalescedWrite(z_v[ss],tmp);
   });
-  nrm = real(TensorRemove(sumD(inner_tmp_v,sites)));
+  bool ok;
+#ifdef GRID_SYCL
+  uint64_t csum=0;
+  uint64_t csum2=0;
+  if ( FlightRecorder::LoggingMode != FlightRecorder::LoggingModeNone)
+  {
+    // z_v
+    {
+      Integer words = sites*sizeof(vobj)/sizeof(uint64_t);
+      uint64_t *base= (uint64_t *)&z_v[0];
+      csum=svm_xor(base,words);
+      ok = FlightRecorder::CsumLog(csum);
+      if ( !ok ) {
+	csum2=svm_xor(base,words);
+	std::cerr<< " Bad z_v CSUM " << std::hex<< csum << " recomputed as "<<csum2<<std::dec<<std::endl;
+      }
+      assert(ok);
+    }
+    // inner_v
+    {
+      Integer words = sites*sizeof(inner_t)/sizeof(uint64_t);
+      uint64_t *base= (uint64_t *)&inner_tmp_v[0];
+      csum=svm_xor(base,words);
+      ok = FlightRecorder::CsumLog(csum);
+      if ( !ok ) {
+	csum2=svm_xor(base,words);
+	std::cerr<< " Bad inner_tmp_v CSUM " << std::hex<< csum << " recomputed as "<<csum2<<std::dec<<std::endl;
+      }
+      assert(ok);
+    }
+  }
 #endif
+  nrm = real(TensorRemove(sumD(inner_tmp_v,sites)));
+  ok = FlightRecorder::NormLog(real(nrm));
+  assert(ok);
+  RealD local = real(nrm);
   grid->GlobalSum(nrm);
+  FlightRecorder::ReductionLog(local,real(nrm));
   return nrm; 
 }
  
@@ -353,7 +422,7 @@ innerProductNorm(ComplexD& ip, RealD &nrm, const Lattice<vobj> &left,const Latti
   conformable(left,right);
 
   typedef typename vobj::vector_typeD vector_type;
-  Vector<ComplexD> tmp(2);
+  std::vector<ComplexD> tmp(2);
 
   GridBase *grid = left.Grid();
 
@@ -363,8 +432,8 @@ innerProductNorm(ComplexD& ip, RealD &nrm, const Lattice<vobj> &left,const Latti
   // GPU
   typedef decltype(innerProductD(vobj(),vobj())) inner_t;
   typedef decltype(innerProductD(vobj(),vobj())) norm_t;
-  Vector<inner_t> inner_tmp(sites);
-  Vector<norm_t>  norm_tmp(sites);
+  deviceVector<inner_t> inner_tmp(sites);
+  deviceVector<norm_t>  norm_tmp(sites);
   auto inner_tmp_v = &inner_tmp[0];
   auto norm_tmp_v = &norm_tmp[0];
   {
@@ -414,7 +483,9 @@ inline auto sum(const LatticeTrinaryExpression<Op,T1,T2,T3> & expr)
 // sliceSum, sliceInnerProduct, sliceAxpy, sliceNorm etc...
 //////////////////////////////////////////////////////////////////////////////////////////////////////////////
 
-template<class vobj> inline void sliceSum(const Lattice<vobj> &Data,std::vector<typename vobj::scalar_object> &result,int orthogdim)
+template<class vobj> inline void sliceSum(const Lattice<vobj> &Data,
+					  std::vector<typename vobj::scalar_object> &result,
+					  int orthogdim)
 {
   ///////////////////////////////////////////////////////
   // FIXME precision promoted summation
@@ -436,8 +507,8 @@ template<class vobj> inline void sliceSum(const Lattice<vobj> &Data,std::vector<
   int ld=grid->_ldimensions[orthogdim];
   int rd=grid->_rdimensions[orthogdim];
 
-  Vector<vobj> lvSum(rd); // will locally sum vectors first
-  Vector<sobj> lsSum(ld,Zero());                    // sum across these down to scalars
+  std::vector<vobj> lvSum(rd); // will locally sum vectors first
+  std::vector<sobj> lsSum(ld,Zero());                    // sum across these down to scalars
   ExtractBuffer<sobj> extracted(Nsimd);                  // splitting the SIMD
 
   result.resize(fd); // And then global sum to return the same vector to every node 
@@ -448,19 +519,10 @@ template<class vobj> inline void sliceSum(const Lattice<vobj> &Data,std::vector<
   int e1=    grid->_slice_nblock[orthogdim];
   int e2=    grid->_slice_block [orthogdim];
   int stride=grid->_slice_stride[orthogdim];
+  int ostride=grid->_ostride[orthogdim];
   
-  // sum over reduced dimension planes, breaking out orthog dir
-  // Parallel over orthog direction
-  autoView( Data_v, Data, CpuRead);
-  thread_for( r,rd, {
-    int so=r*grid->_ostride[orthogdim]; // base offset for start of plane 
-    for(int n=0;n<e1;n++){
-      for(int b=0;b<e2;b++){
-	int ss= so+n*stride+b;
-	lvSum[r]=lvSum[r]+Data_v[ss];
-      }
-    }
-  });
+  //Reduce Data down to lvSum
+  sliceSumReduction(Data,lvSum,rd, e1,e2,stride,ostride,Nsimd);
 
   // Sum across simd lanes in the plane, breaking out orthog dir.
   Coordinate icoor(Nd);
@@ -494,6 +556,8 @@ template<class vobj> inline void sliceSum(const Lattice<vobj> &Data,std::vector<
   scalar_type * ptr = (scalar_type *) &result[0];
   int words = fd*sizeof(sobj)/sizeof(scalar_type);
   grid->GlobalSumVector(ptr, words);
+  //  std::cout << GridLogMessage << " sliceSum local"<<t_sum<<" us, host+mpi "<<t_rest<<std::endl;
+  
 }
 template<class vobj> inline
 std::vector<typename vobj::scalar_object> 
@@ -504,6 +568,20 @@ sliceSum(const Lattice<vobj> &Data,int orthogdim)
   return result;
 }
 
+/*
+Reimplement
+
+1)
+template<class vobj>
+static void sliceMaddMatrix (Lattice<vobj> &R,Eigen::MatrixXcd &aa,const Lattice<vobj> &X,const Lattice<vobj> &Y,int Orthog,RealD scale=1.0) 
+
+2)
+template<class vobj>
+static void sliceInnerProductMatrix(  Eigen::MatrixXcd &mat, const Lattice<vobj> &lhs,const Lattice<vobj> &rhs,int Orthog) 
+
+3)
+-- Make Slice Mul Matrix call sliceMaddMatrix
+ */
 template<class vobj>
 static void sliceInnerProductVector( std::vector<ComplexD> & result, const Lattice<vobj> &lhs,const Lattice<vobj> &rhs,int orthogdim) 
 {
@@ -523,8 +601,8 @@ static void sliceInnerProductVector( std::vector<ComplexD> & result, const Latti
   int ld=grid->_ldimensions[orthogdim];
   int rd=grid->_rdimensions[orthogdim];
 
-  Vector<vector_type> lvSum(rd); // will locally sum vectors first
-  Vector<scalar_type > lsSum(ld,scalar_type(0.0));                    // sum across these down to scalars
+  std::vector<vector_type> lvSum(rd); // will locally sum vectors first
+  std::vector<scalar_type > lsSum(ld,scalar_type(0.0));                    // sum across these down to scalars
   ExtractBuffer<iScalar<scalar_type> > extracted(Nsimd);   // splitting the SIMD  
 
   result.resize(fd); // And then global sum to return the same vector to every node for IO to file
@@ -654,203 +732,96 @@ static void sliceMaddVector(Lattice<vobj> &R,std::vector<RealD> &a,const Lattice
   }
 };
 
-/*
 inline GridBase         *makeSubSliceGrid(const GridBase *BlockSolverGrid,int Orthog)
 {
   int NN    = BlockSolverGrid->_ndimension;
   int nsimd = BlockSolverGrid->Nsimd();
   
-  std::vector<int> latt_phys(0);
-  std::vector<int> simd_phys(0);
-  std::vector<int>  mpi_phys(0);
+  std::vector<int> latt_phys(NN-1);
+  Coordinate simd_phys;
+  std::vector<int>  mpi_phys(NN-1);
+  Coordinate checker_dim_mask(NN-1);
+  int checker_dim=-1;
 
+  int dd;
   for(int d=0;d<NN;d++){
     if( d!=Orthog ) { 
-      latt_phys.push_back(BlockSolverGrid->_fdimensions[d]);
-      simd_phys.push_back(BlockSolverGrid->_simd_layout[d]);
-      mpi_phys.push_back(BlockSolverGrid->_processors[d]);
+      latt_phys[dd]=BlockSolverGrid->_fdimensions[d];
+      mpi_phys[dd] =BlockSolverGrid->_processors[d];
+      checker_dim_mask[dd] = BlockSolverGrid->_checker_dim_mask[d];
+      if ( d == BlockSolverGrid->_checker_dim ) checker_dim = dd;
+      dd++;
     }
   }
-  return (GridBase *)new GridCartesian(latt_phys,simd_phys,mpi_phys); 
+  simd_phys=GridDefaultSimd(latt_phys.size(),nsimd);
+  GridCartesian *tmp         = new GridCartesian(latt_phys,simd_phys,mpi_phys);
+  if(BlockSolverGrid->_isCheckerBoarded) {
+    GridRedBlackCartesian *ret = new GridRedBlackCartesian(tmp,checker_dim_mask,checker_dim);
+    delete tmp;
+    return (GridBase *) ret;
+  } else { 
+    return (GridBase *) tmp;
+  }
 }
-*/
 
 template<class vobj>
 static void sliceMaddMatrix (Lattice<vobj> &R,Eigen::MatrixXcd &aa,const Lattice<vobj> &X,const Lattice<vobj> &Y,int Orthog,RealD scale=1.0) 
 {    
+  GridBase *FullGrid = X.Grid();
+  GridBase *SliceGrid = makeSubSliceGrid(FullGrid,Orthog);
+
+  Lattice<vobj> Ys(SliceGrid);
+  Lattice<vobj> Rs(SliceGrid);
+  Lattice<vobj> Xs(SliceGrid);
+  Lattice<vobj> RR(FullGrid);
+
+  RR = R; // Copies checkerboard for insert
+  
   typedef typename vobj::scalar_object sobj;
   typedef typename vobj::vector_type vector_type;
-
-  int Nblock = X.Grid()->GlobalDimensions()[Orthog];
-
-  GridBase *FullGrid  = X.Grid();
-  //  GridBase *SliceGrid = makeSubSliceGrid(FullGrid,Orthog);
-
-  //  Lattice<vobj> Xslice(SliceGrid);
-  //  Lattice<vobj> Rslice(SliceGrid);
-
-  assert( FullGrid->_simd_layout[Orthog]==1);
-  //  int nh =  FullGrid->_ndimension;
-  //  int nl = SliceGrid->_ndimension;
-  //  int nl = nh-1;
-
-  //FIXME package in a convenient iterator
-  //Should loop over a plane orthogonal to direction "Orthog"
-  int stride=FullGrid->_slice_stride[Orthog];
-  int block =FullGrid->_slice_block [Orthog];
-  int nblock=FullGrid->_slice_nblock[Orthog];
-  int ostride=FullGrid->_ostride[Orthog];
-
-  autoView( X_v, X, CpuRead);
-  autoView( Y_v, Y, CpuRead);
-  autoView( R_v, R, CpuWrite);
-  thread_region
-  {
-    Vector<vobj> s_x(Nblock);
-
-    thread_for_collapse_in_region(2, n,nblock, {
-     for(int b=0;b<block;b++){
-      int o  = n*stride + b;
-
-      for(int i=0;i<Nblock;i++){
-	s_x[i] = X_v[o+i*ostride];
+  int Nslice = X.Grid()->GlobalDimensions()[Orthog];
+  for(int i=0;i<Nslice;i++){
+    ExtractSlice(Ys,Y,i,Orthog);
+    ExtractSlice(Rs,R,i,Orthog);
+    Rs=Ys;
+    for(int j=0;j<Nslice;j++){
+      ExtractSlice(Xs,X,j,Orthog);
+      Rs = Rs + Xs*(scale*aa(j,i));
     }
-
-      vobj dot;
-      for(int i=0;i<Nblock;i++){
-	dot = Y_v[o+i*ostride];
-	for(int j=0;j<Nblock;j++){
-	  dot = dot + s_x[j]*(scale*aa(j,i));
-	}
-	R_v[o+i*ostride]=dot;
-      }
-    }});
+    InsertSlice(Rs,RR,i,Orthog);
   }
+  R=RR; // Copy back handles arguments aliasing case
+  delete SliceGrid;
 };
 
 template<class vobj>
 static void sliceMulMatrix (Lattice<vobj> &R,Eigen::MatrixXcd &aa,const Lattice<vobj> &X,int Orthog,RealD scale=1.0)
 {
-  typedef typename vobj::scalar_object sobj;
-  typedef typename vobj::vector_type vector_type;
-
-  int Nblock = X.Grid()->GlobalDimensions()[Orthog];
-
-  GridBase *FullGrid  = X.Grid();
-  //  GridBase *SliceGrid = makeSubSliceGrid(FullGrid,Orthog);
-  //  Lattice<vobj> Xslice(SliceGrid);
-  //  Lattice<vobj> Rslice(SliceGrid);
-
-  assert( FullGrid->_simd_layout[Orthog]==1);
-  //  int nh =  FullGrid->_ndimension;
-  //  int nl = SliceGrid->_ndimension;
-  //  int nl=1;
-
-  //FIXME package in a convenient iterator
-  // thread_for2d_in_region
-  //Should loop over a plane orthogonal to direction "Orthog"
-  int stride=FullGrid->_slice_stride[Orthog];
-  int block =FullGrid->_slice_block [Orthog];
-  int nblock=FullGrid->_slice_nblock[Orthog];
-  int ostride=FullGrid->_ostride[Orthog];
-  autoView( R_v, R, CpuWrite);
-  autoView( X_v, X, CpuRead);
-  thread_region
-  {
-    std::vector<vobj> s_x(Nblock);
-
-
-    thread_for_collapse_in_region( 2 ,n,nblock,{
-    for(int b=0;b<block;b++){
-      int o  = n*stride + b;
-
-      for(int i=0;i<Nblock;i++){
-	s_x[i] = X_v[o+i*ostride];
-      }
-
-      vobj dot;
-      for(int i=0;i<Nblock;i++){
-	dot = s_x[0]*(scale*aa(0,i));
-	for(int j=1;j<Nblock;j++){
-	  dot = dot + s_x[j]*(scale*aa(j,i));
-	}
-	R_v[o+i*ostride]=dot;
-      }
-    }});
-  }
+  R=Zero();
+  sliceMaddMatrix(R,aa,X,R,Orthog,scale);
 };
 
 
 template<class vobj>
 static void sliceInnerProductMatrix(  Eigen::MatrixXcd &mat, const Lattice<vobj> &lhs,const Lattice<vobj> &rhs,int Orthog) 
 {
+  GridBase *SliceGrid = makeSubSliceGrid(lhs.Grid(),Orthog);
+
+  Lattice<vobj> ls(SliceGrid);
+  Lattice<vobj> rs(SliceGrid);
+  
   typedef typename vobj::scalar_object sobj;
   typedef typename vobj::vector_type vector_type;
-  
-  GridBase *FullGrid  = lhs.Grid();
-  //  GridBase *SliceGrid = makeSubSliceGrid(FullGrid,Orthog);
-  
-  int Nblock = FullGrid->GlobalDimensions()[Orthog];
-  
-  //  Lattice<vobj> Lslice(SliceGrid);
-  //  Lattice<vobj> Rslice(SliceGrid);
-  
-  mat = Eigen::MatrixXcd::Zero(Nblock,Nblock);
-
-  assert( FullGrid->_simd_layout[Orthog]==1);
-  //  int nh =  FullGrid->_ndimension;
-  //  int nl = SliceGrid->_ndimension;
-  //  int nl = nh-1;
-
-  //FIXME package in a convenient iterator
-  //Should loop over a plane orthogonal to direction "Orthog"
-  int stride=FullGrid->_slice_stride[Orthog];
-  int block =FullGrid->_slice_block [Orthog];
-  int nblock=FullGrid->_slice_nblock[Orthog];
-  int ostride=FullGrid->_ostride[Orthog];
-
-  typedef typename vobj::vector_typeD vector_typeD;
-
-  autoView( lhs_v, lhs, CpuRead);
-  autoView( rhs_v, rhs, CpuRead);
-  thread_region
-  {
-    std::vector<vobj> Left(Nblock);
-    std::vector<vobj> Right(Nblock);
-    Eigen::MatrixXcd  mat_thread = Eigen::MatrixXcd::Zero(Nblock,Nblock);
-
-    thread_for_collapse_in_region( 2, n,nblock,{
-    for(int b=0;b<block;b++){
-
-      int o  = n*stride + b;
-
-      for(int i=0;i<Nblock;i++){
-	Left [i] = lhs_v[o+i*ostride];
-	Right[i] = rhs_v[o+i*ostride];
-      }
-
-      for(int i=0;i<Nblock;i++){
-      for(int j=0;j<Nblock;j++){
-	auto tmp = innerProduct(Left[i],Right[j]);
-	auto rtmp = TensorRemove(tmp);
-	auto red  =  Reduce(rtmp);
-	mat_thread(i,j) += std::complex<double>(real(red),imag(red));
-      }}
-    }});
-    thread_critical
-    {
-      mat += mat_thread;
+  int Nslice = lhs.Grid()->GlobalDimensions()[Orthog];
+  mat = Eigen::MatrixXcd::Zero(Nslice,Nslice);
+  for(int s=0;s<Nslice;s++){
+    ExtractSlice(ls,lhs,s,Orthog);
+    for(int ss=0;ss<Nslice;ss++){
+      ExtractSlice(rs,rhs,ss,Orthog);
+      mat(s,ss) = innerProduct(ls,rs);
     }
   }
-
-  for(int i=0;i<Nblock;i++){
-  for(int j=0;j<Nblock;j++){
-    ComplexD sum = mat(i,j);
-    FullGrid->GlobalSum(sum);
-    mat(i,j)=sum;
-  }}
-
-  return;
+  delete SliceGrid;
 }
 
 NAMESPACE_END(Grid);

Grid/lattice/Lattice_reduction_gpu.h

@@ -30,7 +30,7 @@ int getNumBlocksAndThreads(const Iterator n, const size_t sizeofsobj, Iterator &
   cudaGetDevice(&device);
 #endif
 #ifdef GRID_HIP
-  hipGetDevice(&device);
+  auto r=hipGetDevice(&device);
 #endif
   
   Iterator warpSize            = gpu_props[device].warpSize;
@@ -214,22 +214,12 @@ inline typename vobj::scalar_objectD sumD_gpu_small(const vobj *lat, Integer osi
   // Move out of UVM
   // Turns out I had messed up the synchronise after move to compute stream
   // as running this on the default stream fools the synchronise
-#undef UVM_BLOCK_BUFFER  
-#ifndef UVM_BLOCK_BUFFER  
-  commVector<sobj> buffer(numBlocks);
+  deviceVector<sobj> buffer(numBlocks);
   sobj *buffer_v = &buffer[0];
   sobj result;
   reduceKernel<<< numBlocks, numThreads, smemSize, computeStream >>>(lat, buffer_v, size);
   accelerator_barrier();
   acceleratorCopyFromDevice(buffer_v,&result,sizeof(result));
-#else
-  Vector<sobj> buffer(numBlocks);
-  sobj *buffer_v = &buffer[0];
-  sobj result;
-  reduceKernel<<< numBlocks, numThreads, smemSize, computeStream >>>(lat, buffer_v, size);
-  accelerator_barrier();
-  result = *buffer_v;
-#endif
   return result;
 }
 
@@ -244,7 +234,7 @@ inline typename vobj::scalar_objectD sumD_gpu_large(const vobj *lat, Integer osi
   
   const int words = sizeof(vobj)/sizeof(vector);
 
-  Vector<vector> buffer(osites);
+  deviceVector<vector> buffer(osites);
   vector *dat = (vector *)lat;
   vector *buf = &buffer[0];
   iScalar<vector> *tbuf =(iScalar<vector> *)  &buffer[0];

Grid/lattice/Lattice_reduction_sycl.h

@@ -4,29 +4,28 @@ NAMESPACE_BEGIN(Grid);
 // Possibly promote to double and sum
 /////////////////////////////////////////////////////////////////////////////////////////////////////////
 
+
 template <class vobj>
 inline typename vobj::scalar_objectD sumD_gpu_tensor(const vobj *lat, Integer osites) 
 {
   typedef typename vobj::scalar_object sobj;
   typedef typename vobj::scalar_objectD sobjD;
-  sobj *mysum =(sobj *) malloc_shared(sizeof(sobj),*theGridAccelerator);
+
   sobj identity; zeroit(identity);
-  sobj ret ; 
-
+  sobj ret; zeroit(ret);
   Integer nsimd= vobj::Nsimd();
-  
-  theGridAccelerator->submit([&](cl::sycl::handler &cgh) {
-     auto Reduction = cl::sycl::reduction(mysum,identity,std::plus<>());
-     cgh.parallel_for(cl::sycl::range<1>{osites},
+  { 
+    sycl::buffer<sobj, 1> abuff(&ret, {1});
+    theGridAccelerator->submit([&](sycl::handler &cgh) {
+      auto Reduction = sycl::reduction(abuff,cgh,identity,std::plus<>());
+      cgh.parallel_for(sycl::range<1>{osites},
                       Reduction,
-		      [=] (cl::sycl::id<1> item, auto &sum) {
+                      [=] (sycl::id<1> item, auto &sum) {
                         auto osite   = item[0];
                         sum +=Reduce(lat[osite]);
                       });
     });
-  theGridAccelerator->wait();
-  ret = mysum[0];
-  free(mysum,*theGridAccelerator);
+  }
   sobjD dret; convertType(dret,ret);
   return dret;
 }
@@ -69,57 +68,25 @@ inline typename vobj::scalar_object sum_gpu_large(const vobj *lat, Integer osite
   return result;
 }
 
+
+template<class Word> Word svm_xor(Word *vec,uint64_t L)
+{
+  Word identity;  identity=0;
+  Word ret = 0;
+  { 
+    sycl::buffer<Word, 1> abuff(&ret, {1});
+    theGridAccelerator->submit([&](sycl::handler &cgh) {
+      auto Reduction = sycl::reduction(abuff,cgh,identity,std::bit_xor<>());
+      cgh.parallel_for(sycl::range<1>{L},
+                      Reduction,
+                      [=] (sycl::id<1> index, auto &sum) {
+                        sum ^=vec[index];
+                      });
+    });
+  }
+  theGridAccelerator->wait();
+  return ret;
+}
+
 NAMESPACE_END(Grid);
 
-/*
-template<class Double> Double svm_reduce(Double *vec,uint64_t L)
-{
-  Double sumResult; zeroit(sumResult);
-  Double *d_sum =(Double *)cl::sycl::malloc_shared(sizeof(Double),*theGridAccelerator);
-  Double identity;  zeroit(identity);
-  theGridAccelerator->submit([&](cl::sycl::handler &cgh) {
-     auto Reduction = cl::sycl::reduction(d_sum,identity,std::plus<>());
-     cgh.parallel_for(cl::sycl::range<1>{L},
-		      Reduction,
-		      [=] (cl::sycl::id<1> index, auto &sum) {
-	 sum +=vec[index];
-     });
-   });
-  theGridAccelerator->wait();
-  Double ret = d_sum[0];
-  free(d_sum,*theGridAccelerator);
-  std::cout << " svm_reduce finished "<<L<<" sites sum = " << ret <<std::endl;
-  return ret;
-}
-
-template <class vobj>
-inline typename vobj::scalar_objectD sumD_gpu_repack(const vobj *lat, Integer osites)
-{
-  typedef typename vobj::vector_type  vector;
-  typedef typename vobj::scalar_type  scalar;
-
-  typedef typename vobj::scalar_typeD scalarD;
-  typedef typename vobj::scalar_objectD sobjD;
-
-  sobjD ret;
-  scalarD *ret_p = (scalarD *)&ret;
-  
-  const int nsimd = vobj::Nsimd();
-  const int words = sizeof(vobj)/sizeof(vector);
-
-  Vector<scalar> buffer(osites*nsimd);
-  scalar *buf = &buffer[0];
-  vector *dat = (vector *)lat;
-
-  for(int w=0;w<words;w++) {
-
-    accelerator_for(ss,osites,nsimd,{
-	int lane = acceleratorSIMTlane(nsimd);
-	buf[ss*nsimd+lane] = dat[ss*words+w].getlane(lane);
-    });
-    //Precision change at this point is to late to gain precision
-    ret_p[w] = svm_reduce(buf,nsimd*osites);
-  }
-  return ret;
-}
-*/

Grid/lattice/Lattice_rng.h

@@ -152,6 +152,7 @@ public:
 #ifdef RNG_FAST_DISCARD
   static void Skip(RngEngine &eng,uint64_t site)
   {
+#if 0
     /////////////////////////////////////////////////////////////////////////////////////
     // Skip by 2^40 elements between successive lattice sites
     // This goes by 10^12.
@@ -179,6 +180,9 @@ public:
     assert((skip >> shift)==site); // check for overflow
 
     eng.discard(skip);
+#else
+    eng.discardhi(site);
+#endif
     //      std::cout << " Engine  " <<site << " state " <<eng<<std::endl;
   } 
 #endif
@@ -412,7 +416,7 @@ public:
       std::cout << GridLogMessage << "Seed SHA256: " << GridChecksum::sha256_string(seeds) << std::endl;
       SeedFixedIntegers(seeds);
     }
-  void SeedFixedIntegers(const std::vector<int> &seeds){
+  void SeedFixedIntegers(const std::vector<int> &seeds, int britney=0){
 
     // Everyone generates the same seed_seq based on input seeds
     CartesianCommunicator::BroadcastWorld(0,(void *)&seeds[0],sizeof(int)*seeds.size());
@@ -429,7 +433,6 @@ public:
     // MT implementation does not implement fast discard even though
     // in principle this is possible
     ////////////////////////////////////////////////
-#if 1
     thread_for( lidx, _grid->lSites(), {
 
 	int64_t gidx;
@@ -450,29 +453,12 @@ public:
 	
 	int l_idx=generator_idx(o_idx,i_idx);
 	_generators[l_idx] = master_engine;
-	Skip(_generators[l_idx],gidx); // Skip to next RNG sequence
-    });
-#else
-    // Everybody loops over global volume.
-    thread_for( gidx, _grid->_gsites, {
-
-	// Where is it?
-	int rank;
-	int o_idx;
-	int i_idx;
-
-	Coordinate gcoor;
-	_grid->GlobalIndexToGlobalCoor(gidx,gcoor);
-	_grid->GlobalCoorToRankIndex(rank,o_idx,i_idx,gcoor);
-	
-	// If this is one of mine we take it
-	if( rank == _grid->ThisRank() ){
-	  int l_idx=generator_idx(o_idx,i_idx);
-	  _generators[l_idx] = master_engine;
+	if ( britney ) { 
+	  Skip(_generators[l_idx],l_idx); // Skip to next RNG sequence
+	} else { 	
 	  Skip(_generators[l_idx],gidx); // Skip to next RNG sequence
 	}
     });
-#endif
 #else 
     ////////////////////////////////////////////////////////////////
     // Machine and thread decomposition dependent seeding is efficient

Grid/lattice/Lattice_slicesum_core.h

@@ -0,0 +1,267 @@
+#pragma once
+
+#if defined(GRID_CUDA)
+
+#include <cub/cub.cuh>
+#define gpucub cub
+#define gpuError_t cudaError_t
+#define gpuSuccess cudaSuccess
+
+#elif defined(GRID_HIP)
+
+#include <hipcub/hipcub.hpp>
+#define gpucub hipcub
+#define gpuError_t hipError_t
+#define gpuSuccess hipSuccess
+
+#endif
+
+
+NAMESPACE_BEGIN(Grid);
+
+
+#if defined(GRID_CUDA) || defined(GRID_HIP)
+template<class vobj>
+inline void sliceSumReduction_cub_small(const vobj *Data,
+					std::vector<vobj> &lvSum,
+					const int rd,
+					const int e1,
+					const int e2,
+					const int stride,
+					const int ostride,
+					const int Nsimd)
+{
+  size_t subvol_size = e1*e2;
+  deviceVector<vobj> reduction_buffer(rd*subvol_size);
+  auto rb_p = &reduction_buffer[0];
+  vobj zero_init;
+  zeroit(zero_init);
+
+  
+  void *temp_storage_array = NULL;
+  size_t temp_storage_bytes = 0;
+  vobj *d_out;
+  int* d_offsets;
+
+  std::vector<int> offsets(rd+1,0);
+
+  for (int i = 0; i < offsets.size(); i++) {
+    offsets[i] = i*subvol_size;
+  }
+  
+  //Allocate memory for output and offset arrays on device
+  d_out = static_cast<vobj*>(acceleratorAllocDevice(rd*sizeof(vobj)));
+  
+  d_offsets = static_cast<int*>(acceleratorAllocDevice((rd+1)*sizeof(int)));
+  
+  //copy offsets to device
+  acceleratorCopyToDeviceAsync(&offsets[0],d_offsets,sizeof(int)*(rd+1),computeStream);
+  
+  
+  gpuError_t gpuErr = gpucub::DeviceSegmentedReduce::Reduce(temp_storage_array, temp_storage_bytes, rb_p,d_out, rd, d_offsets, d_offsets+1, ::gpucub::Sum(), zero_init, computeStream);
+  if (gpuErr!=gpuSuccess) {
+    std::cout << GridLogError << "Lattice_slicesum_gpu.h: Encountered error during gpucub::DeviceSegmentedReduce::Reduce (setup)! Error: " << gpuErr <<std::endl;
+    exit(EXIT_FAILURE);
+  }
+
+  //allocate memory for temp_storage_array  
+  temp_storage_array = acceleratorAllocDevice(temp_storage_bytes);
+  
+  //prepare buffer for reduction
+  //use non-blocking accelerator_for to avoid syncs (ok because we submit to same computeStream)
+  //use 2d accelerator_for to avoid launch latencies found when serially looping over rd 
+  accelerator_for2dNB( s,subvol_size, r,rd, Nsimd,{ 
+  
+    int n = s / e2;
+    int b = s % e2;
+    int so=r*ostride; // base offset for start of plane 
+    int ss= so+n*stride+b;
+
+    coalescedWrite(rb_p[r*subvol_size+s], coalescedRead(Data[ss]));
+
+  });
+  
+  //issue segmented reductions in computeStream
+  gpuErr = gpucub::DeviceSegmentedReduce::Reduce(temp_storage_array, temp_storage_bytes, rb_p, d_out, rd, d_offsets, d_offsets+1,::gpucub::Sum(), zero_init, computeStream);
+  if (gpuErr!=gpuSuccess) {
+    std::cout << GridLogError << "Lattice_slicesum_gpu.h: Encountered error during gpucub::DeviceSegmentedReduce::Reduce! Error: " << gpuErr <<std::endl;
+    exit(EXIT_FAILURE);
+  }
+  
+  acceleratorCopyFromDeviceAsync(d_out,&lvSum[0],rd*sizeof(vobj),computeStream);
+  
+  //sync after copy
+  accelerator_barrier();
+ 
+  acceleratorFreeDevice(temp_storage_array);
+  acceleratorFreeDevice(d_out);
+  acceleratorFreeDevice(d_offsets);
+  
+
+}
+#endif 
+
+
+#if defined(GRID_SYCL)
+template<class vobj>
+inline void sliceSumReduction_sycl_small(const vobj *Data,
+					 std::vector <vobj> &lvSum,
+					 const int  &rd,
+					 const int &e1,
+					 const int &e2,
+					 const int &stride,
+					 const int &ostride,
+					 const int &Nsimd)
+{
+  size_t subvol_size = e1*e2;
+
+  vobj *mysum = (vobj *) malloc_shared(rd*sizeof(vobj),*theGridAccelerator);
+  vobj vobj_zero;
+  zeroit(vobj_zero);
+  for (int r = 0; r<rd; r++) { 
+    mysum[r] = vobj_zero; 
+  }
+
+  deviceVector<vobj> reduction_buffer(rd*subvol_size);    
+
+  auto rb_p = &reduction_buffer[0];
+
+  // autoView(Data_v, Data, AcceleratorRead);
+
+  //prepare reduction buffer 
+  accelerator_for2d( s,subvol_size, r,rd, (size_t)Nsimd,{ 
+  
+      int n = s / e2;
+      int b = s % e2;
+      int so=r*ostride; // base offset for start of plane 
+      int ss= so+n*stride+b;
+
+      coalescedWrite(rb_p[r*subvol_size+s], coalescedRead(Data[ss]));
+
+  });
+
+  for (int r = 0; r < rd; r++) {
+      theGridAccelerator->submit([&](sycl::handler &cgh) {
+          auto Reduction = sycl::reduction(&mysum[r],std::plus<>());
+          cgh.parallel_for(sycl::range<1>{subvol_size},
+          Reduction,
+          [=](sycl::id<1> item, auto &sum) {
+              auto s = item[0];
+              sum += rb_p[r*subvol_size+s];
+          });
+      });
+      
+     
+  }
+  theGridAccelerator->wait();
+  for (int r = 0; r < rd; r++) {
+    lvSum[r] = mysum[r];
+  }
+  free(mysum,*theGridAccelerator);
+}
+#endif
+
+template<class vobj>
+inline void sliceSumReduction_large(const vobj *Data,
+				    std::vector<vobj> &lvSum,
+				    const int rd,
+				    const int e1,
+				    const int e2,
+				    const int stride,
+				    const int ostride,
+				    const int Nsimd)
+{
+  typedef typename vobj::vector_type vector;
+  const int words = sizeof(vobj)/sizeof(vector);
+  const int osites = rd*e1*e2;
+  deviceVector<vector>buffer(osites);
+  vector *dat = (vector *)Data;
+  vector *buf = &buffer[0];
+  std::vector<vector> lvSum_small(rd);
+  vector *lvSum_ptr = (vector *)&lvSum[0];
+
+  for (int w = 0; w < words; w++) {
+    accelerator_for(ss,osites,1,{
+	    buf[ss] = dat[ss*words+w];
+    });
+
+    #if defined(GRID_CUDA) || defined(GRID_HIP)
+      sliceSumReduction_cub_small(buf,lvSum_small,rd,e1,e2,stride, ostride,Nsimd);
+    #elif defined(GRID_SYCL)
+      sliceSumReduction_sycl_small(buf,lvSum_small,rd,e1,e2,stride, ostride,Nsimd);
+    #endif
+
+    for (int r = 0; r < rd; r++) {
+      lvSum_ptr[w+words*r]=lvSum_small[r];
+    }
+  }
+}
+
+template<class vobj>
+inline void sliceSumReduction_gpu(const Lattice<vobj> &Data,
+				  std::vector<vobj> &lvSum,
+				  const int rd,
+				  const int e1,
+				  const int e2,
+				  const int stride,
+				  const int ostride,
+				  const int Nsimd)
+{
+  autoView(Data_v, Data, AcceleratorRead); //reduction libraries cannot deal with large vobjs so we split into small/large case.
+    if constexpr (sizeof(vobj) <= 256) { 
+
+      #if defined(GRID_CUDA) || defined(GRID_HIP)
+        sliceSumReduction_cub_small(&Data_v[0], lvSum, rd, e1, e2, stride, ostride, Nsimd);
+      #elif defined (GRID_SYCL)
+        sliceSumReduction_sycl_small(&Data_v[0], lvSum, rd, e1, e2, stride, ostride, Nsimd);
+      #endif
+
+    }
+    else {
+      sliceSumReduction_large(&Data_v[0], lvSum, rd, e1, e2, stride, ostride, Nsimd);
+    }
+}
+
+
+template<class vobj>
+inline void sliceSumReduction_cpu(const Lattice<vobj> &Data,
+				  std::vector<vobj> &lvSum,
+				  const int &rd,
+				  const int &e1,
+				  const int &e2,
+				  const int &stride,
+				  const int &ostride,
+				  const int &Nsimd)
+{
+  // sum over reduced dimension planes, breaking out orthog dir
+  // Parallel over orthog direction
+  autoView( Data_v, Data, CpuRead);
+  thread_for( r,rd, {
+    int so=r*ostride; // base offset for start of plane 
+    for(int n=0;n<e1;n++){
+      for(int b=0;b<e2;b++){
+        int ss= so+n*stride+b;
+        lvSum[r]=lvSum[r]+Data_v[ss];
+      }
+    }
+  });
+}
+
+template<class vobj> inline void sliceSumReduction(const Lattice<vobj> &Data,
+						   std::vector<vobj> &lvSum,
+						   const int &rd,
+						   const int &e1,
+						   const int &e2,
+						   const int &stride,
+						   const int &ostride,
+						   const int &Nsimd) 
+{
+#if defined(GRID_CUDA) || defined(GRID_HIP) || defined(GRID_SYCL)
+  sliceSumReduction_gpu(Data, lvSum, rd, e1, e2, stride, ostride, Nsimd);
+#else
+  sliceSumReduction_cpu(Data, lvSum, rd, e1, e2, stride, ostride, Nsimd);
+#endif
+}
+
+
+NAMESPACE_END(Grid);

Grid/lattice/Lattice_transfer.h

@@ -276,18 +276,33 @@ inline void blockProject(Lattice<iVector<CComplex,nbasis > > &coarseData,
 
   autoView( coarseData_ , coarseData, AcceleratorWrite);
   autoView( ip_         , ip,         AcceleratorWrite);
+  RealD t_IP=0;
+  RealD t_co=0;
+  RealD t_za=0;
   for(int v=0;v<nbasis;v++) {
+    t_IP-=usecond();
     blockInnerProductD(ip,Basis[v],fineDataRed); // ip = <basis|fine>
+    t_IP+=usecond();
+    t_co-=usecond();
     accelerator_for( sc, coarse->oSites(), vobj::Nsimd(), {
 	convertType(coarseData_[sc](v),ip_[sc]);
     });
+    t_co+=usecond();
 
     // improve numerical stability of projection
     // |fine> = |fine> - <basis|fine> |basis>
     ip=-ip;
+    t_za-=usecond();
     blockZAXPY(fineDataRed,ip,Basis[v],fineDataRed); 
+    t_za+=usecond();
   }
+  //  std::cout << GridLogPerformance << " blockProject : blockInnerProduct :  "<<t_IP<<" us"<<std::endl;
+  //  std::cout << GridLogPerformance << " blockProject : conv              :  "<<t_co<<" us"<<std::endl;
+  //  std::cout << GridLogPerformance << " blockProject : blockZaxpy        :  "<<t_za<<" us"<<std::endl;
 }
+// This only minimises data motion from CPU to GPU
+// there is chance of better implementation that does a vxk loop of inner products to data share
+// at the GPU thread level
 template<class vobj,class CComplex,int nbasis,class VLattice>
 inline void batchBlockProject(std::vector<Lattice<iVector<CComplex,nbasis>>> &coarseData,
                                const std::vector<Lattice<vobj>> &fineData,
@@ -393,8 +408,15 @@ template<class vobj,class CComplex>
   Lattice<dotp> coarse_inner(coarse);
 
   // Precision promotion
+  RealD t;
+  t=-usecond();
   fine_inner = localInnerProductD<vobj>(fineX,fineY);
+  //  t+=usecond(); std::cout << GridLogPerformance << " blockInnerProduct : localInnerProductD "<<t<<" us"<<std::endl;
+  
+  t=-usecond();
   blockSum(coarse_inner,fine_inner);
+  //  t+=usecond(); std::cout << GridLogPerformance << " blockInnerProduct : blockSum "<<t<<" us"<<std::endl;
+  t=-usecond();
   {
     autoView( CoarseInner_  , CoarseInner,AcceleratorWrite);
     autoView( coarse_inner_ , coarse_inner,AcceleratorRead);
@@ -402,6 +424,7 @@ template<class vobj,class CComplex>
       convertType(CoarseInner_[ss], TensorRemove(coarse_inner_[ss]));
     });
   }
+  //  t+=usecond(); std::cout << GridLogPerformance << " blockInnerProduct : convertType "<<t<<" us"<<std::endl;
  
 }
 
@@ -444,6 +467,9 @@ inline void blockNormalise(Lattice<CComplex> &ip,Lattice<vobj> &fineX)
 template<class vobj>
 inline void blockSum(Lattice<vobj> &coarseData,const Lattice<vobj> &fineData) 
 {
+  const int maxsubsec=256;
+  typedef iVector<vobj,maxsubsec> vSubsec;
+
   GridBase * fine  = fineData.Grid();
   GridBase * coarse= coarseData.Grid();
 
@@ -471,16 +497,32 @@ inline void blockSum(Lattice<vobj> &coarseData,const Lattice<vobj> &fineData)
 
   vobj zz = Zero();
 
-  accelerator_for(sc,coarse->oSites(),vobj::Nsimd(),{
+  // Somewhat lazy calculation
+  // Find the biggest power of two subsection divisor less than or equal to maxsubsec
+  int subsec=maxsubsec;
+  int subvol;
+  subvol=blockVol/subsec;
+  while(subvol*subsec!=blockVol){
+    subsec = subsec/2;
+    subvol=blockVol/subsec;
+  };
+
+  Lattice<vSubsec> coarseTmp(coarse);
+  autoView( coarseTmp_, coarseTmp, AcceleratorWriteDiscard);
+  auto coarseTmp_p= &coarseTmp_[0];
+  
+  // Sum within subsecs in a first kernel
+  accelerator_for(sce,subsec*coarse->oSites(),vobj::Nsimd(),{
+
+      int sc=sce/subsec;
+      int e=sce%subsec;
       
       // One thread per sub block
       Coordinate coor_c(_ndimension);
       Lexicographic::CoorFromIndex(coor_c,sc,coarse_rdimensions);  // Block coordinate
 
       auto cd = coalescedRead(zz);
-      
-      for(int sb=0;sb<blockVol;sb++){
-
+      for(int sb=e*subvol;sb<MIN((e+1)*subvol,blockVol);sb++){
 	int sf;
 	Coordinate coor_b(_ndimension);
 	Coordinate coor_f(_ndimension);
@@ -491,9 +533,18 @@ inline void blockSum(Lattice<vobj> &coarseData,const Lattice<vobj> &fineData)
 	cd=cd+coalescedRead(fineData_p[sf]);
       }
 
-      coalescedWrite(coarseData_p[sc],cd);
+      coalescedWrite(coarseTmp_[sc](e),cd);
 
     });
+   // Sum across subsecs in a second kernel
+   accelerator_for(sc,coarse->oSites(),vobj::Nsimd(),{
+      auto cd = coalescedRead(coarseTmp_p[sc](0));
+      for(int e=1;e<subsec;e++){
+	cd=cd+coalescedRead(coarseTmp_p[sc](e));
+      }
+      coalescedWrite(coarseData_p[sc],cd);
+   });
+
   return;
 }
 
@@ -550,7 +601,7 @@ inline void blockOrthogonalise(Lattice<CComplex> &ip,std::vector<Lattice<vobj> >
   blockOrthonormalize(ip,Basis);
 }
 
-#if 0
+#ifdef GRID_ACCELERATED
 // TODO: CPU optimized version here
 template<class vobj,class CComplex,int nbasis>
 inline void blockPromote(const Lattice<iVector<CComplex,nbasis > > &coarseData,
@@ -576,26 +627,37 @@ inline void blockPromote(const Lattice<iVector<CComplex,nbasis > > &coarseData,
   autoView( fineData_   , fineData, AcceleratorWrite);
   autoView( coarseData_ , coarseData, AcceleratorRead);
 
+  typedef LatticeView<vobj> Vview;
+  std::vector<Vview> AcceleratorVecViewContainer_h; 
+  for(int v=0;v<nbasis;v++) {
+    AcceleratorVecViewContainer_h.push_back(Basis[v].View(AcceleratorRead));
+  }
+  static deviceVector<Vview> AcceleratorVecViewContainer; AcceleratorVecViewContainer.resize(nbasis); 
+  acceleratorCopyToDevice(&AcceleratorVecViewContainer_h[0],&AcceleratorVecViewContainer[0],nbasis *sizeof(Vview));
+  auto Basis_p = &AcceleratorVecViewContainer[0];
   // Loop with a cache friendly loop ordering
-  accelerator_for(sf,fine->oSites(),1,{
+  Coordinate frdimensions=fine->_rdimensions;
+  Coordinate crdimensions=coarse->_rdimensions;
+  accelerator_for(sf,fine->oSites(),vobj::Nsimd(),{
     int sc;
     Coordinate coor_c(_ndimension);
     Coordinate coor_f(_ndimension);
 
-    Lexicographic::CoorFromIndex(coor_f,sf,fine->_rdimensions);
+    Lexicographic::CoorFromIndex(coor_f,sf,frdimensions);
     for(int d=0;d<_ndimension;d++) coor_c[d]=coor_f[d]/block_r[d];
-    Lexicographic::IndexFromCoor(coor_c,sc,coarse->_rdimensions);
+    Lexicographic::IndexFromCoor(coor_c,sc,crdimensions);
 
-    for(int i=0;i<nbasis;i++) {
-      /*      auto basis_ = Basis[i],  );*/
-      if(i==0) fineData_[sf]=coarseData_[sc](i) *basis_[sf]);
-      else     fineData_[sf]=fineData_[sf]+coarseData_[sc](i)*basis_[sf]);
-    }
+    auto sum= coarseData_(sc)(0) *Basis_p[0](sf);
+    for(int i=1;i<nbasis;i++) sum = sum + coarseData_(sc)(i)*Basis_p[i](sf);
+    coalescedWrite(fineData_[sf],sum);
   });
+  for(int v=0;v<nbasis;v++) {
+    AcceleratorVecViewContainer_h[v].ViewClose();
+  }
   return;
-  
 }
 #else
+// CPU version
 template<class vobj,class CComplex,int nbasis,class VLattice>
 inline void blockPromote(const Lattice<iVector<CComplex,nbasis > > &coarseData,
 			 Lattice<vobj>   &fineData,
@@ -682,7 +744,11 @@ void localCopyRegion(const Lattice<vobj> &From,Lattice<vobj> & To,Coordinate Fro
   typedef typename vobj::scalar_type scalar_type;
   typedef typename vobj::vector_type vector_type;
 
-  static const int words=sizeof(vobj)/sizeof(vector_type);
+  const int words=sizeof(vobj)/sizeof(vector_type);
+
+  //////////////////////////////////////////////////////////////////////////////////////////
+  // checks should guarantee that the operations are local
+  //////////////////////////////////////////////////////////////////////////////////////////
 
   GridBase *Fg = From.Grid();
   GridBase *Tg = To.Grid();
@@ -697,39 +763,20 @@ void localCopyRegion(const Lattice<vobj> &From,Lattice<vobj> & To,Coordinate Fro
   for(int d=0;d<nd;d++){
     assert(Fg->_processors[d]  == Tg->_processors[d]);
   }
-  // the above should guarantee that the operations are local
 
-#if 1
+  ///////////////////////////////////////////////////////////
+  // do the index calc on the GPU
+  ///////////////////////////////////////////////////////////
+  Coordinate f_ostride = Fg->_ostride;
+  Coordinate f_istride = Fg->_istride;
+  Coordinate f_rdimensions = Fg->_rdimensions;
+  Coordinate t_ostride = Tg->_ostride;
+  Coordinate t_istride = Tg->_istride;
+  Coordinate t_rdimensions = Tg->_rdimensions;
 
   size_t nsite = 1;
   for(int i=0;i<nd;i++) nsite *= RegionSize[i];
 
-  size_t tbytes = 4*nsite*sizeof(int);
-  int *table = (int*)malloc(tbytes);
- 
-  thread_for(idx, nsite, {
-      Coordinate from_coor, to_coor;
-      size_t rem = idx;
-      for(int i=0;i<nd;i++){
-	size_t base_i  = rem % RegionSize[i]; rem /= RegionSize[i];
-	from_coor[i] = base_i + FromLowerLeft[i];
-	to_coor[i] = base_i + ToLowerLeft[i];
-      }
-      
-      int foidx = Fg->oIndex(from_coor);
-      int fiidx = Fg->iIndex(from_coor);
-      int toidx = Tg->oIndex(to_coor);
-      int tiidx = Tg->iIndex(to_coor);
-      int* tt = table + 4*idx;
-      tt[0] = foidx;
-      tt[1] = fiidx;
-      tt[2] = toidx;
-      tt[3] = tiidx;
-    });
-  
-  int* table_d = (int*)acceleratorAllocDevice(tbytes);
-  acceleratorCopyToDevice(table,table_d,tbytes);
-
   typedef typename vobj::vector_type vector_type;
   typedef typename vobj::scalar_type scalar_type;
 
@@ -737,12 +784,17 @@ void localCopyRegion(const Lattice<vobj> &From,Lattice<vobj> & To,Coordinate Fro
   autoView(to_v,To,AcceleratorWrite);
 
   accelerator_for(idx,nsite,1,{
-      static const int words=sizeof(vobj)/sizeof(vector_type);
-      int* tt = table_d + 4*idx;
-      int from_oidx = *tt++;
-      int from_lane = *tt++;
-      int to_oidx = *tt++;
-      int to_lane = *tt;
+
+      Coordinate from_coor, to_coor, base;
+      Lexicographic::CoorFromIndex(base,idx,RegionSize);
+      for(int i=0;i<nd;i++){
+	from_coor[i] = base[i] + FromLowerLeft[i];
+	to_coor[i] = base[i] + ToLowerLeft[i];
+      }
+      int from_oidx = 0; for(int d=0;d<nd;d++) from_oidx+=f_ostride[d]*(from_coor[d]%f_rdimensions[d]);
+      int from_lane = 0; for(int d=0;d<nd;d++) from_lane+=f_istride[d]*(from_coor[d]/f_rdimensions[d]);
+      int to_oidx   = 0; for(int d=0;d<nd;d++) to_oidx+=t_ostride[d]*(to_coor[d]%t_rdimensions[d]);
+      int to_lane   = 0; for(int d=0;d<nd;d++) to_lane+=t_istride[d]*(to_coor[d]/t_rdimensions[d]);
 
       const vector_type* from = (const vector_type *)&from_v[from_oidx];
       vector_type* to = (vector_type *)&to_v[to_oidx];
@@ -753,55 +805,145 @@ void localCopyRegion(const Lattice<vobj> &From,Lattice<vobj> & To,Coordinate Fro
 	putlane(to[w], stmp, to_lane);
       }
   });
-  
-  acceleratorFreeDevice(table_d);    
-  free(table);
-  
-
-#else  
-  Coordinate ldf = Fg->_ldimensions;
-  Coordinate rdf = Fg->_rdimensions;
-  Coordinate isf = Fg->_istride;
-  Coordinate osf = Fg->_ostride;
-  Coordinate rdt = Tg->_rdimensions;
-  Coordinate ist = Tg->_istride;
-  Coordinate ost = Tg->_ostride;
-
-  autoView( t_v , To, CpuWrite);
-  autoView( f_v , From, CpuRead);
-  thread_for(idx,Fg->lSites(),{
-    sobj s;
-    Coordinate Fcoor(nd);
-    Coordinate Tcoor(nd);
-    Lexicographic::CoorFromIndex(Fcoor,idx,ldf);
-    int in_region=1;
-    for(int d=0;d<nd;d++){
-      if ( (Fcoor[d] < FromLowerLeft[d]) || (Fcoor[d]>=FromLowerLeft[d]+RegionSize[d]) ){ 
-	in_region=0;
 }
-      Tcoor[d] = ToLowerLeft[d]+ Fcoor[d]-FromLowerLeft[d];
+
+template<class vobj>
+void InsertSliceFast(const Lattice<vobj> &From,Lattice<vobj> & To,int slice, int orthog)
+{
+  typedef typename vobj::scalar_object sobj;
+  typedef typename vobj::scalar_type scalar_type;
+  typedef typename vobj::vector_type vector_type;
+
+  const int words=sizeof(vobj)/sizeof(vector_type);
+
+  //////////////////////////////////////////////////////////////////////////////////////////
+  // checks should guarantee that the operations are local
+  //////////////////////////////////////////////////////////////////////////////////////////
+  GridBase *Fg = From.Grid();
+  GridBase *Tg = To.Grid();
+  assert(!Fg->_isCheckerBoarded);
+  assert(!Tg->_isCheckerBoarded);
+  int Nsimd = Fg->Nsimd();
+  int nF = Fg->_ndimension;
+  int nT = Tg->_ndimension;
+  assert(nF+1 == nT);
+
+  ///////////////////////////////////////////////////////////
+  // do the index calc on the GPU
+  ///////////////////////////////////////////////////////////
+  Coordinate f_ostride = Fg->_ostride;
+  Coordinate f_istride = Fg->_istride;
+  Coordinate f_rdimensions = Fg->_rdimensions;
+  Coordinate t_ostride = Tg->_ostride;
+  Coordinate t_istride = Tg->_istride;
+  Coordinate t_rdimensions = Tg->_rdimensions;
+  Coordinate RegionSize = Fg->_ldimensions;
+  size_t nsite = 1;
+  for(int i=0;i<nF;i++) nsite *= RegionSize[i]; // whole volume of lower dim grid
+
+  typedef typename vobj::vector_type vector_type;
+  typedef typename vobj::scalar_type scalar_type;
+
+  autoView(from_v,From,AcceleratorRead);
+  autoView(to_v,To,AcceleratorWrite);
+
+  accelerator_for(idx,nsite,1,{
+
+      Coordinate from_coor(nF), to_coor(nT);
+      Lexicographic::CoorFromIndex(from_coor,idx,RegionSize);
+      int j=0;
+      for(int i=0;i<nT;i++){
+	if ( i!=orthog ) { 
+	  to_coor[i] = from_coor[j];
+	  j++;
+	} else {
+	  to_coor[i] = slice;
 	}
-    if (in_region) {
-#if 0      
-      Integer idx_f = 0; for(int d=0;d<nd;d++) idx_f+=isf[d]*(Fcoor[d]/rdf[d]); // inner index from
-      Integer idx_t = 0; for(int d=0;d<nd;d++) idx_t+=ist[d]*(Tcoor[d]/rdt[d]); // inner index to
-      Integer odx_f = 0; for(int d=0;d<nd;d++) odx_f+=osf[d]*(Fcoor[d]%rdf[d]); // outer index from
-      Integer odx_t = 0; for(int d=0;d<nd;d++) odx_t+=ost[d]*(Tcoor[d]%rdt[d]); // outer index to
-      scalar_type * fp = (scalar_type *)&f_v[odx_f];
-      scalar_type * tp = (scalar_type *)&t_v[odx_t];
+      }
+      int from_oidx = 0; for(int d=0;d<nF;d++) from_oidx+=f_ostride[d]*(from_coor[d]%f_rdimensions[d]);
+      int from_lane = 0; for(int d=0;d<nF;d++) from_lane+=f_istride[d]*(from_coor[d]/f_rdimensions[d]);
+      int to_oidx   = 0; for(int d=0;d<nT;d++) to_oidx+=t_ostride[d]*(to_coor[d]%t_rdimensions[d]);
+      int to_lane   = 0; for(int d=0;d<nT;d++) to_lane+=t_istride[d]*(to_coor[d]/t_rdimensions[d]);
+
+      const vector_type* from = (const vector_type *)&from_v[from_oidx];
+      vector_type* to = (vector_type *)&to_v[to_oidx];
+      
+      scalar_type stmp;
       for(int w=0;w<words;w++){
-	tp[w].putlane(fp[w].getlane(idx_f),idx_t);
-      }
-#else
-    peekLocalSite(s,f_v,Fcoor);
-    pokeLocalSite(s,t_v,Tcoor);
-#endif
+	stmp = getlane(from[w], from_lane);
+	putlane(to[w], stmp, to_lane);
       }
   });
-
-#endif
 }
 
+template<class vobj>
+void ExtractSliceFast(Lattice<vobj> &To,const Lattice<vobj> & From,int slice, int orthog)
+{
+  typedef typename vobj::scalar_object sobj;
+  typedef typename vobj::scalar_type scalar_type;
+  typedef typename vobj::vector_type vector_type;
+
+  const int words=sizeof(vobj)/sizeof(vector_type);
+
+  //////////////////////////////////////////////////////////////////////////////////////////
+  // checks should guarantee that the operations are local
+  //////////////////////////////////////////////////////////////////////////////////////////
+  GridBase *Fg = From.Grid();
+  GridBase *Tg = To.Grid();
+  assert(!Fg->_isCheckerBoarded);
+  assert(!Tg->_isCheckerBoarded);
+  int Nsimd = Fg->Nsimd();
+  int nF = Fg->_ndimension;
+  int nT = Tg->_ndimension;
+  assert(nT+1 == nF);
+
+  ///////////////////////////////////////////////////////////
+  // do the index calc on the GPU
+  ///////////////////////////////////////////////////////////
+  Coordinate f_ostride = Fg->_ostride;
+  Coordinate f_istride = Fg->_istride;
+  Coordinate f_rdimensions = Fg->_rdimensions;
+  Coordinate t_ostride = Tg->_ostride;
+  Coordinate t_istride = Tg->_istride;
+  Coordinate t_rdimensions = Tg->_rdimensions;
+  Coordinate RegionSize = Tg->_ldimensions;
+  size_t nsite = 1;
+  for(int i=0;i<nT;i++) nsite *= RegionSize[i]; // whole volume of lower dim grid
+
+  typedef typename vobj::vector_type vector_type;
+  typedef typename vobj::scalar_type scalar_type;
+
+  autoView(from_v,From,AcceleratorRead);
+  autoView(to_v,To,AcceleratorWrite);
+
+  accelerator_for(idx,nsite,1,{
+
+      Coordinate from_coor(nF), to_coor(nT);
+      Lexicographic::CoorFromIndex(to_coor,idx,RegionSize);
+      int j=0;
+      for(int i=0;i<nF;i++){
+	if ( i!=orthog ) { 
+	  from_coor[i] = to_coor[j];
+	  j++;
+	} else {
+	  from_coor[i] = slice;
+	}
+      }
+      int from_oidx = 0; for(int d=0;d<nF;d++) from_oidx+=f_ostride[d]*(from_coor[d]%f_rdimensions[d]);
+      int from_lane = 0; for(int d=0;d<nF;d++) from_lane+=f_istride[d]*(from_coor[d]/f_rdimensions[d]);
+      int to_oidx   = 0; for(int d=0;d<nT;d++) to_oidx+=t_ostride[d]*(to_coor[d]%t_rdimensions[d]);
+      int to_lane   = 0; for(int d=0;d<nT;d++) to_lane+=t_istride[d]*(to_coor[d]/t_rdimensions[d]);
+
+      const vector_type* from = (const vector_type *)&from_v[from_oidx];
+      vector_type* to = (vector_type *)&to_v[to_oidx];
+      
+      scalar_type stmp;
+      for(int w=0;w<words;w++){
+	stmp = getlane(from[w], from_lane);
+	putlane(to[w], stmp, to_lane);
+      }
+  });
+}
 
 template<class vobj>
 void InsertSlice(const Lattice<vobj> &lowDim,Lattice<vobj> & higherDim,int slice, int orthog)
@@ -839,8 +981,14 @@ void InsertSlice(const Lattice<vobj> &lowDim,Lattice<vobj> & higherDim,int slice
     hcoor[orthog] = slice;
     for(int d=0;d<nh;d++){
       if ( d!=orthog ) { 
-	hcoor[d]=lcoor[ddl++];
+	hcoor[d]=lcoor[ddl];
+	if ( hg->_checker_dim == d ) {
+	  hcoor[d]=hcoor[d]*2; // factor in the full coor for peekLocalSite
+	  lcoor[ddl]=lcoor[ddl]*2; // factor in the full coor for peekLocalSite
 	}
+	ddl++;
+      }
+      
     }
     peekLocalSite(s,lowDimv,lcoor);
     pokeLocalSite(s,higherDimv,hcoor);
@@ -861,6 +1009,7 @@ void ExtractSlice(Lattice<vobj> &lowDim,const Lattice<vobj> & higherDim,int slic
   assert(orthog<nh);
   assert(orthog>=0);
   assert(hg->_processors[orthog]==1);
+  lowDim.Checkerboard() = higherDim.Checkerboard();
 
   int dl; dl = 0;
   for(int d=0;d<nh;d++){
@@ -878,11 +1027,16 @@ void ExtractSlice(Lattice<vobj> &lowDim,const Lattice<vobj> & higherDim,int slic
     Coordinate lcoor(nl);
     Coordinate hcoor(nh);
     lg->LocalIndexToLocalCoor(idx,lcoor);
-    int ddl=0;
     hcoor[orthog] = slice;
+    int ddl=0;
     for(int d=0;d<nh;d++){
       if ( d!=orthog ) { 
-	hcoor[d]=lcoor[ddl++];
+	hcoor[d]=lcoor[ddl];
+	if ( hg->_checker_dim == d ) {
+	  hcoor[d]=hcoor[d]*2;     // factor in the full gridd coor for peekLocalSite
+	  lcoor[ddl]=lcoor[ddl]*2; // factor in the full coor for peekLocalSite
+	}
+	ddl++;
       }
     }
     peekLocalSite(s,higherDimv,hcoor);
@@ -891,9 +1045,7 @@ void ExtractSlice(Lattice<vobj> &lowDim,const Lattice<vobj> & higherDim,int slic
 
 }
 
-
-//Insert subvolume orthogonal to direction 'orthog' with slice index 'slice_lo' from 'lowDim' onto slice index 'slice_hi' of higherDim
-//The local dimensions of both 'lowDim' and 'higherDim' orthogonal to 'orthog' should be the same
+//Can I implement with local copyregion??
 template<class vobj>
 void InsertSliceLocal(const Lattice<vobj> &lowDim, Lattice<vobj> & higherDim,int slice_lo,int slice_hi, int orthog)
 {
@@ -914,121 +1066,18 @@ void InsertSliceLocal(const Lattice<vobj> &lowDim, Lattice<vobj> & higherDim,int
       assert(lg->_ldimensions[d] == hg->_ldimensions[d]);
     }
   }
-
-#if 1
-  size_t nsite = lg->lSites()/lg->LocalDimensions()[orthog];
-  size_t tbytes = 4*nsite*sizeof(int);
-  int *table = (int*)malloc(tbytes);
-  
-  thread_for(idx,nsite,{
-    Coordinate lcoor(nl);
-    Coordinate hcoor(nh);
-    lcoor[orthog] = slice_lo;
-    hcoor[orthog] = slice_hi;
-    size_t rem = idx;
-    for(int mu=0;mu<nl;mu++){
-      if(mu != orthog){
-	int xmu = rem % lg->LocalDimensions()[mu];  rem /= lg->LocalDimensions()[mu];
-	lcoor[mu] = hcoor[mu] = xmu;
-      }
-    }
-    int loidx = lg->oIndex(lcoor);
-    int liidx = lg->iIndex(lcoor);
-    int hoidx = hg->oIndex(hcoor);
-    int hiidx = hg->iIndex(hcoor);
-    int* tt = table + 4*idx;
-    tt[0] = loidx;
-    tt[1] = liidx;
-    tt[2] = hoidx;
-    tt[3] = hiidx;
-    });
-   
-  int* table_d = (int*)acceleratorAllocDevice(tbytes);
-  acceleratorCopyToDevice(table,table_d,tbytes);
-
-  typedef typename vobj::vector_type vector_type;
-  typedef typename vobj::scalar_type scalar_type;
-
-  autoView(lowDim_v,lowDim,AcceleratorRead);
-  autoView(higherDim_v,higherDim,AcceleratorWrite);
-  
-  accelerator_for(idx,nsite,1,{
-      static const int words=sizeof(vobj)/sizeof(vector_type);
-      int* tt = table_d + 4*idx;
-      int from_oidx = *tt++;
-      int from_lane = *tt++;
-      int to_oidx = *tt++;
-      int to_lane = *tt;
-
-      const vector_type* from = (const vector_type *)&lowDim_v[from_oidx];
-      vector_type* to = (vector_type *)&higherDim_v[to_oidx];
-      
-      scalar_type stmp;
-      for(int w=0;w<words;w++){
-	stmp = getlane(from[w], from_lane);
-	putlane(to[w], stmp, to_lane);
-      }
-    });
-  
-  acceleratorFreeDevice(table_d);    
-  free(table);
-  
-#else
-  // the above should guarantee that the operations are local
-  autoView(lowDimv,lowDim,CpuRead);
-  autoView(higherDimv,higherDim,CpuWrite);
-  thread_for(idx,lg->lSites(),{
-    sobj s;
-    Coordinate lcoor(nl);
-    Coordinate hcoor(nh);
-    lg->LocalIndexToLocalCoor(idx,lcoor);
-    if( lcoor[orthog] == slice_lo ) { 
-      hcoor=lcoor;
-      hcoor[orthog] = slice_hi;
-      peekLocalSite(s,lowDimv,lcoor);
-      pokeLocalSite(s,higherDimv,hcoor);
-    }
-  });
-#endif
+  Coordinate sz = lg->_ldimensions;
+  sz[orthog]=1;
+  Coordinate f_ll(nl,0); f_ll[orthog]=slice_lo;
+  Coordinate t_ll(nh,0); t_ll[orthog]=slice_hi;
+  localCopyRegion(lowDim,higherDim,f_ll,t_ll,sz);
 }
 
 
 template<class vobj>
 void ExtractSliceLocal(Lattice<vobj> &lowDim,const Lattice<vobj> & higherDim,int slice_lo,int slice_hi, int orthog)
 {
-  typedef typename vobj::scalar_object sobj;
-
-  GridBase *lg = lowDim.Grid();
-  GridBase *hg = higherDim.Grid();
-  int nl = lg->_ndimension;
-  int nh = hg->_ndimension;
-
-  assert(nl == nh);
-  assert(orthog<nh);
-  assert(orthog>=0);
-
-  for(int d=0;d<nh;d++){
-    if ( d!=orthog ) {
-    assert(lg->_processors[d]  == hg->_processors[d]);
-    assert(lg->_ldimensions[d] == hg->_ldimensions[d]);
-  }
-  }
-
-  // the above should guarantee that the operations are local
-  autoView(lowDimv,lowDim,CpuWrite);
-  autoView(higherDimv,higherDim,CpuRead);
-  thread_for(idx,lg->lSites(),{
-    sobj s;
-    Coordinate lcoor(nl);
-    Coordinate hcoor(nh);
-    lg->LocalIndexToLocalCoor(idx,lcoor);
-    if( lcoor[orthog] == slice_lo ) { 
-      hcoor=lcoor;
-      hcoor[orthog] = slice_hi;
-      peekLocalSite(s,higherDimv,hcoor);
-      pokeLocalSite(s,lowDimv,lcoor);
-    }
-  });
+  InsertSliceLocal(higherDim,lowDim,slice_hi,slice_lo,orthog);
 }
 
 
@@ -1740,5 +1789,35 @@ void Grid_unsplit(std::vector<Lattice<Vobj> > & full,Lattice<Vobj>   & split)
   }
 }
 
+//////////////////////////////////////////////////////
+// Faster but less accurate blockProject
+//////////////////////////////////////////////////////
+template<class vobj,class CComplex,int nbasis,class VLattice>
+inline void blockProjectFast(Lattice<iVector<CComplex,nbasis > > &coarseData,
+			     const             Lattice<vobj>   &fineData,
+			     const VLattice &Basis)
+{
+  GridBase * fine  = fineData.Grid();
+  GridBase * coarse= coarseData.Grid();
+
+  Lattice<iScalar<CComplex> > ip(coarse);
+
+  autoView( coarseData_ , coarseData, AcceleratorWrite);
+  autoView( ip_         , ip,         AcceleratorWrite);
+  RealD t_IP=0;
+  RealD t_co=0;
+  for(int v=0;v<nbasis;v++) {
+    t_IP-=usecond();
+    blockInnerProductD(ip,Basis[v],fineData); 
+    t_IP+=usecond();
+    t_co-=usecond();
+    accelerator_for( sc, coarse->oSites(), vobj::Nsimd(), {
+	convertType(coarseData_[sc](v),ip_[sc]);
+      });
+    t_co+=usecond();
+  }
+}
+
+
 NAMESPACE_END(Grid);
 

Grid/lattice/Lattice_view.h

@@ -45,6 +45,7 @@ public:
   };
   // Host only
   GridBase * getGrid(void) const { return _grid; };
+  vobj* getHostPointer(void) const { return _odata; };
 };
 
 /////////////////////////////////////////////////////////////////////////////////////////

Grid/lattice/PaddedCell.h

@@ -45,6 +45,188 @@ struct CshiftImplGauge: public CshiftImplBase<typename Gimpl::GaugeLinkField::ve
   typename Gimpl::GaugeLinkField Cshift(const typename Gimpl::GaugeLinkField &in, int dir, int shift) const override{ return Gimpl::CshiftLink(in,dir,shift); }
 };  
 
+
+/*
+ *
+ * TODO: 
+ *  -- address elementsof vobj via thread block in Scatter/Gather
+ *  -- overlap comms with motion in Face_exchange
+ *
+ */
+
+template<class vobj> inline void ScatterSlice(const deviceVector<vobj> &buf,
+					      Lattice<vobj> &lat,
+					      int x,
+					      int dim,
+					      int offset=0)
+{
+  const int Nsimd=vobj::Nsimd();
+  typedef typename vobj::scalar_object sobj;
+  typedef typename vobj::scalar_type scalar_type;
+  typedef typename vobj::vector_type vector_type;
+
+  GridBase *grid = lat.Grid();
+  Coordinate simd = grid->_simd_layout;
+  int Nd          = grid->Nd();
+  int block       = grid->_slice_block[dim];
+  int stride      = grid->_slice_stride[dim];
+  int nblock      = grid->_slice_nblock[dim];
+  int rd          = grid->_rdimensions[dim];
+
+  int ox = x%rd;
+  int ix = x/rd;
+
+  int isites = 1; for(int d=0;d<Nd;d++) if( d!=dim) isites*=simd[d];
+
+  Coordinate rsimd= simd;  rsimd[dim]=1; // maybe reduce Nsimd
+
+  int rNsimd = 1; for(int d=0;d<Nd;d++) rNsimd*=rsimd[d];
+  int rNsimda= Nsimd/simd[dim]; // should be equal
+  assert(rNsimda==rNsimd);
+  int face_ovol=block*nblock;
+
+  //  assert(buf.size()==face_ovol*rNsimd);
+
+  /*This will work GPU ONLY unless rNsimd is put in the lexico index*/
+  //Let's make it work on GPU and then make a special accelerator_for that
+  //doesn't hide the SIMD direction and keeps explicit in the threadIdx
+  //for cross platform
+  // FIXME -- can put internal indices into thread loop
+  auto buf_p = & buf[0];
+  autoView(lat_v, lat, AcceleratorWrite);
+  accelerator_for(ss, face_ovol/simd[dim],Nsimd,{
+
+    // scalar layout won't coalesce
+#ifdef GRID_SIMT
+      {
+	int blane=acceleratorSIMTlane(Nsimd); // buffer lane
+#else
+      for(int blane=0;blane<Nsimd;blane++) {
+#endif
+	int olane=blane%rNsimd;               // reduced lattice lane
+	int obit =blane/rNsimd;
+
+	///////////////////////////////////////////////////////////////
+	// osite -- potentially one bit from simd in the buffer: (ss<<1)|obit
+	///////////////////////////////////////////////////////////////
+	int ssp = ss*simd[dim]+obit;
+	int b    = ssp%block;
+	int n    = ssp/block;
+	int osite= b+n*stride + ox*block;
+	
+	////////////////////////////////////////////
+	// isite -- map lane within buffer to lane within lattice
+	////////////////////////////////////////////
+	Coordinate icoor;
+	int lane;
+	Lexicographic::CoorFromIndex(icoor,olane,rsimd);
+	icoor[dim]=ix;
+	Lexicographic::IndexFromCoor(icoor,lane,simd);
+	
+	///////////////////////////////////////////
+	// Transfer into lattice - will coalesce
+	///////////////////////////////////////////
+	//	sobj obj = extractLane(blane,buf_p[ss+offset]);
+	//	insertLane(lane,lat_v[osite],obj);
+	const int words=sizeof(vobj)/sizeof(vector_type);
+	vector_type * from = (vector_type *)&buf_p[ss+offset];
+	vector_type * to   = (vector_type *)&lat_v[osite];
+	scalar_type stmp;
+	for(int w=0;w<words;w++){
+	  stmp = getlane(from[w], blane);
+	  putlane(to[w], stmp, lane);
+	}
+      }
+  });
+}
+
+template<class vobj> inline void GatherSlice(deviceVector<vobj> &buf,
+					     const Lattice<vobj> &lat,
+					     int x,
+					     int dim,
+					     int offset=0)
+{
+  const int Nsimd=vobj::Nsimd();
+  typedef typename vobj::scalar_object sobj;
+  typedef typename vobj::scalar_type scalar_type;
+  typedef typename vobj::vector_type vector_type;
+
+  autoView(lat_v, lat, AcceleratorRead);
+
+  GridBase *grid = lat.Grid();
+  Coordinate simd = grid->_simd_layout;
+  int Nd          = grid->Nd();
+  int block       = grid->_slice_block[dim];
+  int stride      = grid->_slice_stride[dim];
+  int nblock      = grid->_slice_nblock[dim];
+  int rd          = grid->_rdimensions[dim];
+
+  int ox = x%rd;
+  int ix = x/rd;
+
+  int isites = 1; for(int d=0;d<Nd;d++) if( d!=dim) isites*=simd[d];
+
+  Coordinate rsimd= simd;  rsimd[dim]=1; // maybe reduce Nsimd
+
+  int rNsimd = 1; for(int d=0;d<Nd;d++) rNsimd*=rsimd[d];
+  
+  int face_ovol=block*nblock;
+
+  //  assert(buf.size()==face_ovol*rNsimd);
+
+  /*This will work GPU ONLY unless rNsimd is put in the lexico index*/
+  //Let's make it work on GPU and then make a special accelerator_for that
+  //doesn't hide the SIMD direction and keeps explicit in the threadIdx
+  //for cross platform
+  //For CPU perhaps just run a loop over Nsimd
+  auto buf_p = & buf[0];
+  accelerator_for(ss, face_ovol/simd[dim],Nsimd,{
+
+    // scalar layout won't coalesce
+#ifdef GRID_SIMT
+      {
+	int blane=acceleratorSIMTlane(Nsimd); // buffer lane
+#else
+      for(int blane=0;blane<Nsimd;blane++) {
+#endif
+	int olane=blane%rNsimd;               // reduced lattice lane
+	int obit =blane/rNsimd;
+	
+	////////////////////////////////////////////
+	// osite
+	////////////////////////////////////////////
+	int ssp = ss*simd[dim]+obit;
+	int b    = ssp%block;
+	int n    = ssp/block;
+	int osite= b+n*stride + ox*block;
+
+	////////////////////////////////////////////
+	// isite -- map lane within buffer to lane within lattice
+	////////////////////////////////////////////
+	Coordinate icoor;
+	int lane;
+	Lexicographic::CoorFromIndex(icoor,olane,rsimd);
+	icoor[dim]=ix;
+	Lexicographic::IndexFromCoor(icoor,lane,simd);
+	
+	///////////////////////////////////////////
+	// Take out of lattice
+	///////////////////////////////////////////
+	//	sobj obj = extractLane(lane,lat_v[osite]);
+	//	insertLane(blane,buf_p[ss+offset],obj);
+	const int words=sizeof(vobj)/sizeof(vector_type);
+	vector_type * to    = (vector_type *)&buf_p[ss+offset];
+	vector_type * from  = (vector_type *)&lat_v[osite];
+	scalar_type stmp;
+	for(int w=0;w<words;w++){
+	  stmp = getlane(from[w], lane);
+	  putlane(to[w], stmp, blane);
+	}
+      }
+  });
+}
+
+
 class PaddedCell {
 public:
   GridCartesian * unpadded_grid;
@@ -70,9 +252,12 @@ public:
   }
   void DeleteGrids(void)
   {
+    Coordinate processors=unpadded_grid->_processors;
     for(int d=0;d<grids.size();d++){
+      if ( processors[d] > 1 ) { 
 	delete grids[d];
       }
+    }
     grids.resize(0);
   };
   void AllocateGrids(void)
@@ -82,19 +267,20 @@ public:
     Coordinate processors=unpadded_grid->_processors;
     Coordinate plocal    =unpadded_grid->LocalDimensions();
     Coordinate global(dims);
-
+    GridCartesian *old_grid = unpadded_grid;
     // expand up one dim at a time
     for(int d=0;d<dims;d++){
 
       if ( processors[d] > 1 ) { 
 	plocal[d] += 2*depth; 
-      }
       
 	for(int d=0;d<dims;d++){
 	  global[d] = plocal[d]*processors[d];
 	}
 
-      grids.push_back(new GridCartesian(global,simd,processors));
+	old_grid = new GridCartesian(global,simd,processors);
+      }
+      grids.push_back(old_grid);
     }
   };
   template<class vobj>
@@ -125,6 +311,17 @@ public:
     }
     return tmp;
   }
+  template<class vobj>
+  inline Lattice<vobj> ExchangePeriodic(const Lattice<vobj> &in) const
+  {
+    GridBase *old_grid = in.Grid();
+    int dims = old_grid->Nd();
+    Lattice<vobj> tmp = in;
+    for(int d=0;d<dims;d++){
+      tmp = ExpandPeriodic(d,tmp); // rvalue && assignment
+    }
+    return tmp;
+  }
   // expand up one dim at a time
   template<class vobj>
   inline Lattice<vobj> Expand(int dim, const Lattice<vobj> &in, const CshiftImplBase<vobj> &cshift = CshiftImplDefault<vobj>()) const
@@ -139,8 +336,6 @@ public:
     if(dim==0) conformable(old_grid,unpadded_grid);
     else       conformable(old_grid,grids[dim-1]);
 
-    //    std::cout << " dim "<<dim<<" local "<<local << " padding to "<<plocal<<std::endl;
-
     double tins=0, tshift=0;
 
     int islocal = 0 ;
@@ -148,13 +343,23 @@ public:
 
     if ( islocal ) {
 
+      // replace with a copy and maybe grid swizzle
+      // return in;??
       double t = usecond();
-      for(int x=0;x<local[dim];x++){
-	InsertSliceLocal(in,padded,x,x,dim);
-      }
+      padded = in;
       tins += usecond() - t;
       
     } else {
+
+      //////////////////////////////////////////////
+      // Replace sequence with
+      // ---------------------
+      // (i) Gather high face(s); start comms
+      // (ii) Gather low  face(s); start comms
+      // (iii) Copy middle bit with localCopyRegion
+      // (iv) Complete high face(s), insert slice(s)
+      // (iv) Complete low  face(s), insert slice(s)
+      //////////////////////////////////////////////
       // Middle bit
       double t = usecond();
       for(int x=0;x<local[dim];x++){
@@ -183,14 +388,184 @@ public:
 	InsertSliceLocal(shifted,padded,x,x,dim);
       }
       tins += usecond() - t;
+
     }
-    std::cout << GridLogDebug << "PaddedCell::Expand timings: cshift:" << tshift/1000 << "ms, insert-slice:" << tins/1000 << "ms" << std::endl;
+    std::cout << GridLogPerformance << "PaddedCell::Expand timings: cshift:" << tshift/1000 << "ms, insert-slice:" << tins/1000 << "ms" << std::endl;
     
     return padded;
   }
 
+  template<class vobj>
+  inline Lattice<vobj> ExpandPeriodic(int dim, const Lattice<vobj> &in) const
+  {
+    Coordinate processors=unpadded_grid->_processors;
+    GridBase *old_grid = in.Grid();
+    GridCartesian *new_grid = grids[dim];//These are new grids
+    Lattice<vobj>  padded(new_grid);
+    //    Lattice<vobj> shifted(old_grid);    
+    Coordinate local     =old_grid->LocalDimensions();
+    Coordinate plocal    =new_grid->LocalDimensions();
+    if(dim==0) conformable(old_grid,unpadded_grid);
+    else       conformable(old_grid,grids[dim-1]);
+
+    //    std::cout << " dim "<<dim<<" local "<<local << " padding to "<<plocal<<std::endl;
+    double tins=0, tshift=0;
+
+    int islocal = 0 ;
+    if ( processors[dim] == 1 ) islocal = 1;
+
+    if ( islocal ) {
+      padded=in; // slightly different interface could avoid a copy operation
+    } else {
+      Face_exchange(in,padded,dim,depth);
+      return padded;
+    }
+    return padded;
+  }
+  template<class vobj>
+  void Face_exchange(const Lattice<vobj> &from,
+		     Lattice<vobj> &to,
+		     int dimension,int depth) const
+  {
+    typedef typename vobj::vector_type vector_type;
+    typedef typename vobj::scalar_type scalar_type;
+    typedef typename vobj::scalar_object sobj;
+
+    RealD t_gather=0.0;
+    RealD t_scatter=0.0;
+    RealD t_comms=0.0;
+    RealD t_copy=0.0;
+    
+    //    std::cout << GridLogMessage << "dimension " <<dimension<<std::endl;
+    //    DumpSliceNorm(std::string("Face_exchange from"),from,dimension);
+    GridBase *grid=from.Grid();
+    GridBase *new_grid=to.Grid();
+
+    Coordinate lds = from.Grid()->_ldimensions;
+    Coordinate nlds=   to.Grid()->_ldimensions;
+    Coordinate simd= from.Grid()->_simd_layout;
+    int ld    = lds[dimension];
+    int nld   = to.Grid()->_ldimensions[dimension];
+    const int Nsimd = vobj::Nsimd();
+
+    assert(depth<=lds[dimension]); // A must be on neighbouring node
+    assert(depth>0);   // A caller bug if zero
+    assert(ld+2*depth==nld);
+    ////////////////////////////////////////////////////////////////////////////
+    // Face size and byte calculations
+    ////////////////////////////////////////////////////////////////////////////
+    int buffer_size = 1;
+    for(int d=0;d<lds.size();d++){
+      if ( d!= dimension) buffer_size=buffer_size*lds[d];
+    }
+    buffer_size = buffer_size  / Nsimd;
+    int rNsimd = Nsimd / simd[dimension];
+    assert( buffer_size == from.Grid()->_slice_nblock[dimension]*from.Grid()->_slice_block[dimension] / simd[dimension]);
+
+    static deviceVector<vobj> send_buf; 
+    static deviceVector<vobj> recv_buf;
+    send_buf.resize(buffer_size*2*depth);    
+    recv_buf.resize(buffer_size*2*depth);
+
+    std::vector<MpiCommsRequest_t> fwd_req;   
+    std::vector<MpiCommsRequest_t> bwd_req;   
+
+    int words = buffer_size;
+    int bytes = words * sizeof(vobj);
+
+    ////////////////////////////////////////////////////////////////////////////
+    // Communication coords
+    ////////////////////////////////////////////////////////////////////////////
+    int comm_proc = 1;
+    int xmit_to_rank;
+    int recv_from_rank;
+    grid->ShiftedRanks(dimension,comm_proc,xmit_to_rank,recv_from_rank);
+
+    ////////////////////////////////////////////////////////////////////////////
+    // Gather all surface terms up to depth "d"
+    ////////////////////////////////////////////////////////////////////////////
+    RealD t;
+    RealD t_tot=-usecond();
+    int plane=0;
+    for ( int d=0;d < depth ; d ++ ) {
+      int tag = d*1024 + dimension*2+0;
+
+      t=usecond();
+      GatherSlice(send_buf,from,d,dimension,plane*buffer_size); plane++;
+      t_gather+=usecond()-t;
+
+      t=usecond();
+      grid->SendToRecvFromBegin(fwd_req,
+				(void *)&send_buf[d*buffer_size], xmit_to_rank,
+				(void *)&recv_buf[d*buffer_size], recv_from_rank, bytes, tag);
+      t_comms+=usecond()-t;
+     }
+    for ( int d=0;d < depth ; d ++ ) {
+      int tag = d*1024 + dimension*2+1;
+
+      t=usecond();
+      GatherSlice(send_buf,from,ld-depth+d,dimension,plane*buffer_size); plane++;
+      t_gather+= usecond() - t;
+
+      t=usecond();
+      grid->SendToRecvFromBegin(bwd_req,
+				(void *)&send_buf[(d+depth)*buffer_size], recv_from_rank,
+				(void *)&recv_buf[(d+depth)*buffer_size], xmit_to_rank, bytes,tag);
+      t_comms+=usecond()-t;
+    }
+
+    ////////////////////////////////////////////////////////////////////////////
+    // Copy interior -- overlap this with comms
+    ////////////////////////////////////////////////////////////////////////////
+    int Nd = new_grid->Nd();
+    Coordinate LL(Nd,0);
+    Coordinate sz = grid->_ldimensions;
+    Coordinate toLL(Nd,0);
+    toLL[dimension]=depth;
+    t=usecond();
+    localCopyRegion(from,to,LL,toLL,sz);
+    t_copy= usecond() - t;
+    
+    ////////////////////////////////////////////////////////////////////////////
+    // Scatter all faces
+    ////////////////////////////////////////////////////////////////////////////
+    plane=0;
+
+    t=usecond();
+    grid->CommsComplete(fwd_req);
+    t_comms+= usecond() - t;
+
+    t=usecond();
+    for ( int d=0;d < depth ; d ++ ) {
+      ScatterSlice(recv_buf,to,nld-depth+d,dimension,plane*buffer_size); plane++;
+    }
+    t_scatter= usecond() - t;
+
+    t=usecond();
+    grid->CommsComplete(bwd_req);
+    t_comms+= usecond() - t;
+    
+    t=usecond();
+    for ( int d=0;d < depth ; d ++ ) {
+      ScatterSlice(recv_buf,to,d,dimension,plane*buffer_size); plane++;
+    }
+    t_scatter+= usecond() - t;
+    t_tot+=usecond();
+
+    std::cout << GridLogPerformance << "PaddedCell::Expand new timings: gather :" << t_gather/1000  << "ms"<<std::endl;
+    std::cout << GridLogPerformance << "PaddedCell::Expand new timings: scatter:" << t_scatter/1000   << "ms"<<std::endl;
+    std::cout << GridLogPerformance << "PaddedCell::Expand new timings: copy   :" << t_copy/1000      << "ms"<<std::endl;
+    std::cout << GridLogPerformance << "PaddedCell::Expand new timings: comms  :" << t_comms/1000     << "ms"<<std::endl;
+    std::cout << GridLogPerformance << "PaddedCell::Expand new timings: total  :" << t_tot/1000     << "ms"<<std::endl;
+    std::cout << GridLogPerformance << "PaddedCell::Expand new timings: gather :" << depth*4.0*bytes/t_gather << "MB/s"<<std::endl;
+    std::cout << GridLogPerformance << "PaddedCell::Expand new timings: scatter:" << depth*4.0*bytes/t_scatter<< "MB/s"<<std::endl;
+    std::cout << GridLogPerformance << "PaddedCell::Expand new timings: comms  :" << (RealD)4.0*bytes/t_comms   << "MB/s"<<std::endl;
+    std::cout << GridLogPerformance << "PaddedCell::Expand new timings: face bytes  :" << depth*bytes/1e6 << "MB"<<std::endl;
+  }
+  
 };
  
 
 NAMESPACE_END(Grid);
 
+

Grid/log/Log.h

@@ -191,6 +191,41 @@ extern Colours    GridLogColours;
 
 std::string demangle(const char* name) ;
 
+template<typename... Args>
+inline std::string sjoin(Args&&... args) noexcept {
+    std::ostringstream msg;
+    (msg << ... << args);
+    return msg.str();
+}
+
+/*!  @brief make log messages work like python print */
+template <typename... Args>
+inline void Grid_log(Args&&... args) {
+    std::string msg = sjoin(std::forward<Args>(args)...);
+    std::cout << GridLogMessage << msg << std::endl;
+}
+
+/*!  @brief make warning messages work like python print */
+template <typename... Args>
+inline void Grid_warn(Args&&... args) {
+    std::string msg = sjoin(std::forward<Args>(args)...);
+    std::cout << "\033[33m" << GridLogWarning << msg << "\033[0m" << std::endl;
+}
+
+/*!  @brief make error messages work like python print */
+template <typename... Args>
+inline void Grid_error(Args&&... args) {
+    std::string msg = sjoin(std::forward<Args>(args)...);
+    std::cout << "\033[31m" << GridLogError << msg << "\033[0m" << std::endl;
+}
+
+/*!  @brief make pass messages work like python print */
+template <typename... Args>
+inline void Grid_pass(Args&&... args) {
+    std::string msg = sjoin(std::forward<Args>(args)...);
+    std::cout << "\033[32m" << GridLogMessage << msg << "\033[0m" << std::endl;
+}
+
 #define _NBACKTRACE (256)
 extern void * Grid_backtrace_buffer[_NBACKTRACE];
 

Grid/parallelIO/BinaryIO.h

@@ -165,7 +165,7 @@ class BinaryIO {
 	 * FIXME -- 128^3 x 256 x 16 will overflow.
 	 */
 	
-	int global_site;
+	int64_t global_site;
 
 	Lexicographic::CoorFromIndex(coor,local_site,local_vol);
 
@@ -175,8 +175,8 @@ class BinaryIO {
 
 	Lexicographic::IndexFromCoor(coor,global_site,global_vol);
 
-	uint32_t gsite29   = global_site%29;
-	uint32_t gsite31   = global_site%31;
+	uint64_t gsite29   = global_site%29;
+	uint64_t gsite31   = global_site%31;
 	
 	site_crc = crc32(0,(unsigned char *)site_buf,sizeof(fobj));
 	//	std::cout << "Site "<<local_site << " crc "<<std::hex<<site_crc<<std::dec<<std::endl;
@@ -545,7 +545,9 @@ class BinaryIO {
 				       const std::string &format,
 				       uint32_t &nersc_csum,
 				       uint32_t &scidac_csuma,
-				       uint32_t &scidac_csumb)
+				       uint32_t &scidac_csumb,
+				       int control=BINARYIO_LEXICOGRAPHIC
+				       )
   {
     typedef typename vobj::scalar_object sobj;
     typedef typename vobj::Realified::scalar_type word;    word w=0;
@@ -556,7 +558,7 @@ class BinaryIO {
     std::vector<sobj> scalardata(lsites); 
     std::vector<fobj>     iodata(lsites); // Munge, checksum, byte order in here
     
-    IOobject(w,grid,iodata,file,offset,format,BINARYIO_READ|BINARYIO_LEXICOGRAPHIC,
+    IOobject(w,grid,iodata,file,offset,format,BINARYIO_READ|control,
 	     nersc_csum,scidac_csuma,scidac_csumb);
 
     GridStopWatch timer; 
@@ -582,7 +584,8 @@ class BinaryIO {
 					  const std::string &format,
 					  uint32_t &nersc_csum,
 					  uint32_t &scidac_csuma,
-					  uint32_t &scidac_csumb)
+					  uint32_t &scidac_csumb,
+					  int control=BINARYIO_LEXICOGRAPHIC)
   {
     typedef typename vobj::scalar_object sobj;
     typedef typename vobj::Realified::scalar_type word;    word w=0;
@@ -607,7 +610,7 @@ class BinaryIO {
     while (attemptsLeft >= 0)
     {
       grid->Barrier();
-      IOobject(w,grid,iodata,file,offset,format,BINARYIO_WRITE|BINARYIO_LEXICOGRAPHIC,
+      IOobject(w,grid,iodata,file,offset,format,BINARYIO_WRITE|control,
 	             nersc_csum,scidac_csuma,scidac_csumb);
       if (checkWrite)
       {
@@ -617,7 +620,7 @@ class BinaryIO {
 
         std::cout << GridLogMessage << "writeLatticeObject: read back object" << std::endl;
         grid->Barrier();
-        IOobject(w,grid,ckiodata,file,ckoffset,format,BINARYIO_READ|BINARYIO_LEXICOGRAPHIC,
+        IOobject(w,grid,ckiodata,file,ckoffset,format,BINARYIO_READ|control,
 	               cknersc_csum,ckscidac_csuma,ckscidac_csumb);
         if ((cknersc_csum != nersc_csum) or (ckscidac_csuma != scidac_csuma) or (ckscidac_csumb != scidac_csumb))
         {

Grid/parallelIO/IldgIO.h

@@ -162,8 +162,14 @@ template<class vobj> void ScidacMetaData(Lattice<vobj> & field,
  {
    uint32_t scidac_checksuma = stoull(scidacChecksum_.suma,0,16);
    uint32_t scidac_checksumb = stoull(scidacChecksum_.sumb,0,16);
-   if ( scidac_csuma !=scidac_checksuma) return 0;
-   if ( scidac_csumb !=scidac_checksumb) return 0;
+   std::cout << GridLogMessage << " scidacChecksumVerify computed "<<scidac_csuma<<" expected "<<scidac_checksuma <<std::endl;
+   std::cout << GridLogMessage << " scidacChecksumVerify computed "<<scidac_csumb<<" expected "<<scidac_checksumb <<std::endl;
+   if ( scidac_csuma !=scidac_checksuma) {
+     return 0;
+   };
+   if ( scidac_csumb !=scidac_checksumb) {
+     return 0;
+   };
    return 1;
  }
 
@@ -206,7 +212,7 @@ class GridLimeReader : public BinaryIO {
   // Read a generic lattice field and verify checksum
   ////////////////////////////////////////////
   template<class vobj>
-  void readLimeLatticeBinaryObject(Lattice<vobj> &field,std::string record_name)
+  void readLimeLatticeBinaryObject(Lattice<vobj> &field,std::string record_name,int control=BINARYIO_LEXICOGRAPHIC)
   {
     typedef typename vobj::scalar_object sobj;
     scidacChecksum scidacChecksum_;
@@ -238,7 +244,7 @@ class GridLimeReader : public BinaryIO {
 	uint64_t offset= ftello(File);
 	//	std::cout << " ReadLatticeObject from offset "<<offset << std::endl;
 	BinarySimpleMunger<sobj,sobj> munge;
-	BinaryIO::readLatticeObject< vobj, sobj >(field, filename, munge, offset, format,nersc_csum,scidac_csuma,scidac_csumb);
+	BinaryIO::readLatticeObject< vobj, sobj >(field, filename, munge, offset, format,nersc_csum,scidac_csuma,scidac_csumb,control);
 	std::cout << GridLogMessage << "SciDAC checksum A " << std::hex << scidac_csuma << std::dec << std::endl;
 	std::cout << GridLogMessage << "SciDAC checksum B " << std::hex << scidac_csumb << std::dec << std::endl;
 	/////////////////////////////////////////////
@@ -408,7 +414,7 @@ class GridLimeWriter : public BinaryIO
   // in communicator used by the field.Grid()
   ////////////////////////////////////////////////////
   template<class vobj>
-  void writeLimeLatticeBinaryObject(Lattice<vobj> &field,std::string record_name)
+  void writeLimeLatticeBinaryObject(Lattice<vobj> &field,std::string record_name,int control=BINARYIO_LEXICOGRAPHIC)
   {
     ////////////////////////////////////////////////////////////////////
     // NB: FILE and iostream are jointly writing disjoint sequences in the
@@ -459,7 +465,7 @@ class GridLimeWriter : public BinaryIO
     ///////////////////////////////////////////
     std::string format = getFormatString<vobj>();
     BinarySimpleMunger<sobj,sobj> munge;
-    BinaryIO::writeLatticeObject<vobj,sobj>(field, filename, munge, offset1, format,nersc_csum,scidac_csuma,scidac_csumb);
+    BinaryIO::writeLatticeObject<vobj,sobj>(field, filename, munge, offset1, format,nersc_csum,scidac_csuma,scidac_csumb,control);
 
     ///////////////////////////////////////////
     // Wind forward and close the record
@@ -512,7 +518,8 @@ class ScidacWriter : public GridLimeWriter {
   ////////////////////////////////////////////////
   template <class vobj, class userRecord>
   void writeScidacFieldRecord(Lattice<vobj> &field,userRecord _userRecord,
-                              const unsigned int recordScientificPrec = 0) 
+                              const unsigned int recordScientificPrec = 0,
+			      int control=BINARYIO_LEXICOGRAPHIC)
   {
     GridBase * grid = field.Grid();
 
@@ -534,7 +541,7 @@ class ScidacWriter : public GridLimeWriter {
       writeLimeObject(0,0,_scidacRecord,_scidacRecord.SerialisableClassName(),std::string(SCIDAC_PRIVATE_RECORD_XML));
     }
     // Collective call
-    writeLimeLatticeBinaryObject(field,std::string(ILDG_BINARY_DATA));      // Closes message with checksum
+    writeLimeLatticeBinaryObject(field,std::string(ILDG_BINARY_DATA),control);      // Closes message with checksum
   }
 };
 
@@ -553,7 +560,8 @@ class ScidacReader : public GridLimeReader {
   // Write generic lattice field in scidac format
   ////////////////////////////////////////////////
   template <class vobj, class userRecord>
-  void readScidacFieldRecord(Lattice<vobj> &field,userRecord &_userRecord) 
+  void readScidacFieldRecord(Lattice<vobj> &field,userRecord &_userRecord,
+			     int control=BINARYIO_LEXICOGRAPHIC) 
   {
     typedef typename vobj::scalar_object sobj;
     GridBase * grid = field.Grid();
@@ -571,7 +579,7 @@ class ScidacReader : public GridLimeReader {
     readLimeObject(header ,std::string("FieldMetaData"),std::string(GRID_FORMAT)); // Open message 
     readLimeObject(_userRecord,_userRecord.SerialisableClassName(),std::string(SCIDAC_RECORD_XML));
     readLimeObject(_scidacRecord,_scidacRecord.SerialisableClassName(),std::string(SCIDAC_PRIVATE_RECORD_XML));
-    readLimeLatticeBinaryObject(field,std::string(ILDG_BINARY_DATA));
+    readLimeLatticeBinaryObject(field,std::string(ILDG_BINARY_DATA),control);
   }
   void skipPastBinaryRecord(void) {
     std::string rec_name(ILDG_BINARY_DATA);

Grid/perfmon/Tracing.h

@@ -34,7 +34,7 @@ class GridTracer {
 };
 inline void tracePush(const char *name) { roctxRangePushA(name); }
 inline void tracePop(const char *name) { roctxRangePop(); }
-inline int  traceStart(const char *name) { roctxRangeStart(name); }
+inline int  traceStart(const char *name) { return roctxRangeStart(name); }
 inline void traceStop(int ID) { roctxRangeStop(ID); }
 #endif
 

Grid/qcd/action/ActionBase.h

@@ -129,6 +129,26 @@ public:
   virtual ~Action(){}
 };
 
+template <class GaugeField >
+class EmptyAction : public Action <GaugeField>
+{
+  using Action<GaugeField>::refresh;
+  using Action<GaugeField>::Sinitial;
+  using Action<GaugeField>::deriv;
+
+  virtual void refresh(const GaugeField& U, GridSerialRNG &sRNG, GridParallelRNG& pRNG) { assert(0);}; // refresh pseudofermions
+  virtual RealD S(const GaugeField& U) { return 0.0;};                             // evaluate the action
+  virtual void deriv(const GaugeField& U, GaugeField& dSdU) { assert(0); };        // evaluate the action derivative
+
+  ///////////////////////////////
+  // Logging
+  ///////////////////////////////
+  virtual std::string action_name()    { return std::string("Level Force Log"); };
+  virtual std::string LogParameters()  { return std::string("No parameters");};
+};
+
+
+
 NAMESPACE_END(Grid);
 
 #endif // ACTION_BASE_H

Grid/qcd/action/fermion/AbstractEOFAFermion.h

@@ -55,6 +55,11 @@ public:
   RealD alpha; // Mobius scale
   RealD k;     // EOFA normalization constant
 
+  // Device resident
+  deviceVector<Coeff_t> d_shift_coefficients;
+  deviceVector<Coeff_t> d_MooeeInv_shift_lc;
+  deviceVector<Coeff_t> d_MooeeInv_shift_norm;
+  
   virtual void Instantiatable(void) = 0;
 
   // EOFA-specific operations
@@ -92,6 +97,11 @@ public:
     this->k = this->alpha * (_mq3-_mq2) * std::pow(this->alpha+1.0,2*Ls) /
       ( std::pow(this->alpha+1.0,Ls) + _mq2*std::pow(this->alpha-1.0,Ls) ) /
       ( std::pow(this->alpha+1.0,Ls) + _mq3*std::pow(this->alpha-1.0,Ls) );
+    
+    d_shift_coefficients.resize(Ls);
+    d_MooeeInv_shift_lc.resize(Ls);
+    d_MooeeInv_shift_norm.resize(Ls);
+
   };
 };
 

Grid/qcd/action/fermion/CayleyFermion5D.h

@@ -90,16 +90,16 @@ public:
   void M5D(const FermionField &psi,
 	   const FermionField &phi,
 	   FermionField &chi,
-	   Vector<Coeff_t> &lower,
-	   Vector<Coeff_t> &diag,
-	   Vector<Coeff_t> &upper);
+	   std::vector<Coeff_t> &lower,
+	   std::vector<Coeff_t> &diag,
+	   std::vector<Coeff_t> &upper);
 
   void M5Ddag(const FermionField &psi,
 	      const FermionField &phi,
 	      FermionField &chi,
-	      Vector<Coeff_t> &lower,
-	      Vector<Coeff_t> &diag,
-	      Vector<Coeff_t> &upper);
+	      std::vector<Coeff_t> &lower,
+	      std::vector<Coeff_t> &diag,
+	      std::vector<Coeff_t> &upper);
 
   virtual void   Instantiatable(void)=0;
 
@@ -119,35 +119,51 @@ public:
   RealD mass_plus, mass_minus;
 
   // Save arguments to SetCoefficientsInternal
-  Vector<Coeff_t> _gamma;
+  std::vector<Coeff_t> _gamma;
   RealD                _zolo_hi;
   RealD                _b;
   RealD                _c;
 
+  // possible boost
+  std::vector<ComplexD> qmu;
+  void set_qmu(std::vector<ComplexD> _qmu) { qmu=_qmu; assert(qmu.size()==Nd);};
+  void addQmu(const FermionField &in, FermionField &out, int dag);
+  
   // Cayley form Moebius (tanh and zolotarev)
-  Vector<Coeff_t> omega;
-  Vector<Coeff_t> bs;    // S dependent coeffs
-  Vector<Coeff_t> cs;
-  Vector<Coeff_t> as;
+  std::vector<Coeff_t> omega;
+  std::vector<Coeff_t> bs;    // S dependent coeffs
+  std::vector<Coeff_t> cs;
+  std::vector<Coeff_t> as;
   // For preconditioning Cayley form
-  Vector<Coeff_t> bee;
-  Vector<Coeff_t> cee;
-  Vector<Coeff_t> aee;
-  Vector<Coeff_t> beo;
-  Vector<Coeff_t> ceo;
-  Vector<Coeff_t> aeo;
+  std::vector<Coeff_t> bee;
+  std::vector<Coeff_t> cee;
+  std::vector<Coeff_t> aee;
+  std::vector<Coeff_t> beo;
+  std::vector<Coeff_t> ceo;
+  std::vector<Coeff_t> aeo;
   // LDU factorisation of the eeoo matrix
-  Vector<Coeff_t> lee;
-  Vector<Coeff_t> leem;
-  Vector<Coeff_t> uee;
-  Vector<Coeff_t> ueem;
-  Vector<Coeff_t> dee;
+  std::vector<Coeff_t> lee;
+  std::vector<Coeff_t> leem;
+  std::vector<Coeff_t> uee;
+  std::vector<Coeff_t> ueem;
+  std::vector<Coeff_t> dee;
+
+  // Device memory
+  deviceVector<Coeff_t> d_diag;
+  deviceVector<Coeff_t> d_upper;
+  deviceVector<Coeff_t> d_lower;
+
+  deviceVector<Coeff_t> d_lee;
+  deviceVector<Coeff_t> d_dee;
+  deviceVector<Coeff_t> d_uee;
+  deviceVector<Coeff_t> d_leem;
+  deviceVector<Coeff_t> d_ueem;
 
   // Matrices of 5d ee inverse params
-  Vector<iSinglet<Simd> >  MatpInv;
-  Vector<iSinglet<Simd> >  MatmInv;
-  Vector<iSinglet<Simd> >  MatpInvDag;
-  Vector<iSinglet<Simd> >  MatmInvDag;
+  //  std::vector<iSinglet<Simd> >  MatpInv;
+  //  std::vector<iSinglet<Simd> >  MatmInv;
+  //  std::vector<iSinglet<Simd> >  MatpInvDag;
+  //  std::vector<iSinglet<Simd> >  MatmInvDag;
 
   ///////////////////////////////////////////////////////////////
   // Conserved current utilities
@@ -187,7 +203,7 @@ public:
 protected:
   virtual void SetCoefficientsZolotarev(RealD zolohi,Approx::zolotarev_data *zdata,RealD b,RealD c);
   virtual void SetCoefficientsTanh(Approx::zolotarev_data *zdata,RealD b,RealD c);
-  virtual void SetCoefficientsInternal(RealD zolo_hi,Vector<Coeff_t> & gamma,RealD b,RealD c);
+  virtual void SetCoefficientsInternal(RealD zolo_hi,std::vector<Coeff_t> & gamma,RealD b,RealD c);
 };
 
 NAMESPACE_END(Grid);

Grid/qcd/action/fermion/ContinuedFractionFermion5D.h

@@ -60,6 +60,50 @@ public:
   //      virtual void   Instantiatable(void)=0;
   virtual void   Instantiatable(void) =0;
 
+  void FreePropagator(const FermionField &in,FermionField &out,RealD mass,std::vector<Complex> boundary, std::vector<double> twist)
+  {
+    std::cout << "Free Propagator for PartialFraction"<<std::endl;
+    FermionField in_k(in.Grid());
+    FermionField prop_k(in.Grid());
+    
+    FFT theFFT((GridCartesian *) in.Grid());
+
+    //phase for boundary condition
+    ComplexField coor(in.Grid());
+    ComplexField ph(in.Grid());  ph = Zero();
+    FermionField in_buf(in.Grid()); in_buf = Zero();
+    typedef typename Simd::scalar_type Scalar;
+    Scalar ci(0.0,1.0);
+    assert(twist.size() == Nd);//check that twist is Nd
+    assert(boundary.size() == Nd);//check that boundary conditions is Nd
+    int shift = 0;
+    for(unsigned int nu = 0; nu < Nd; nu++)
+      {
+	// Shift coordinate lattice index by 1 to account for 5th dimension.
+	LatticeCoordinate(coor, nu + shift);
+	double boundary_phase = ::acos(real(boundary[nu]));
+	ph = ph + boundary_phase*coor*((1./(in.Grid()->_fdimensions[nu+shift])));
+	//momenta for propagator shifted by twist+boundary
+	twist[nu] = twist[nu] + boundary_phase/((2.0*M_PI));
+      }
+    in_buf = exp(ci*ph*(-1.0))*in;
+
+    theFFT.FFT_all_dim(in_k,in,FFT::forward);
+    this->MomentumSpacePropagatorHw(prop_k,in_k,mass,twist);
+    theFFT.FFT_all_dim(out,prop_k,FFT::backward);
+    
+    //phase for boundary condition
+    out = out * exp(ci*ph);
+  };
+
+  virtual void FreePropagator(const FermionField &in,FermionField &out,RealD mass) {
+    std::vector<double> twist(Nd,0.0); //default: periodic boundarys in all directions
+    std::vector<Complex> boundary;
+    for(int i=0;i<Nd;i++) boundary.push_back(1);//default: periodic boundary conditions
+    FreePropagator(in,out,mass,boundary,twist);
+  };
+
+  
   // Efficient support for multigrid coarsening
   virtual void  Mdir (const FermionField &in, FermionField &out,int dir,int disp);
   virtual void  MdirAll(const FermionField &in, std::vector<FermionField> &out);
@@ -90,12 +134,12 @@ protected:
   RealD mass;
   RealD R;
   RealD ZoloHiInv;
-  Vector<double> Beta;
-  Vector<double> cc;;
-  Vector<double> cc_d;;
-  Vector<double> sqrt_cc;
-  Vector<double> See;
-  Vector<double> Aee;
+  std::vector<double> Beta;
+  std::vector<double> cc;;
+  std::vector<double> cc_d;;
+  std::vector<double> sqrt_cc;
+  std::vector<double> See;
+  std::vector<double> Aee;
 
 };
 

Grid/qcd/action/fermion/DomainWallEOFAFermion.h

@@ -69,10 +69,10 @@ public:
   // Instantiate different versions depending on Impl
   /////////////////////////////////////////////////////
   void M5D(const FermionField& psi, const FermionField& phi, FermionField& chi,
-	   Vector<Coeff_t>& lower, Vector<Coeff_t>& diag, Vector<Coeff_t>& upper);
+	   std::vector<Coeff_t>& lower, std::vector<Coeff_t>& diag, std::vector<Coeff_t>& upper);
 
   void M5Ddag(const FermionField& psi, const FermionField& phi, FermionField& chi,
-	      Vector<Coeff_t>& lower, Vector<Coeff_t>& diag, Vector<Coeff_t>& upper);
+	      std::vector<Coeff_t>& lower, std::vector<Coeff_t>& diag, std::vector<Coeff_t>& upper);
 
   virtual void RefreshShiftCoefficients(RealD new_shift);
 
@@ -83,7 +83,7 @@ public:
 			RealD _M5, const ImplParams& p=ImplParams());
 
 protected:
-  void SetCoefficientsInternal(RealD zolo_hi, Vector<Coeff_t>& gamma, RealD b, RealD c);
+  void SetCoefficientsInternal(RealD zolo_hi, std::vector<Coeff_t>& gamma, RealD b, RealD c);
 };
 
 NAMESPACE_END(Grid);

Grid/qcd/action/fermion/ImprovedStaggeredFermion.h

@@ -102,11 +102,11 @@ public:
 		     GaugeField &mat, 
 		     const FermionField &A, const FermionField &B, int dag);
 
-  void DhopInternal(StencilImpl &st, LebesgueOrder &lo, DoubledGaugeField &U,DoubledGaugeField &UUU,
+  void DhopInternal(StencilImpl &st, DoubledGaugeField &U,DoubledGaugeField &UUU,
                     const FermionField &in, FermionField &out, int dag);
-  void DhopInternalSerialComms(StencilImpl &st, LebesgueOrder &lo, DoubledGaugeField &U,DoubledGaugeField &UUU,
+  void DhopInternalSerialComms(StencilImpl &st, DoubledGaugeField &U,DoubledGaugeField &UUU,
                     const FermionField &in, FermionField &out, int dag);
-  void DhopInternalOverlappedComms(StencilImpl &st, LebesgueOrder &lo, DoubledGaugeField &U,DoubledGaugeField &UUU,
+  void DhopInternalOverlappedComms(StencilImpl &st, DoubledGaugeField &U,DoubledGaugeField &UUU,
                     const FermionField &in, FermionField &out, int dag);
 
   //////////////////////////////////////////////////////////////////////////
@@ -164,8 +164,6 @@ public:
   DoubledGaugeField UUUmuEven;
   DoubledGaugeField UUUmuOdd;
 
-  LebesgueOrder Lebesgue;
-  LebesgueOrder LebesgueEvenOdd;
   
   ///////////////////////////////////////////////////////////////
   // Conserved current utilities

Grid/qcd/action/fermion/ImprovedStaggeredFermion5D.h

@@ -100,7 +100,6 @@ public:
 		     int dag);
     
   void DhopInternal(StencilImpl & st,
-		    LebesgueOrder &lo,
 		    DoubledGaugeField &U,
 		    DoubledGaugeField &UUU,
 		    const FermionField &in, 
@@ -108,7 +107,6 @@ public:
 		    int dag);
     
     void DhopInternalOverlappedComms(StencilImpl & st,
-		      LebesgueOrder &lo,
 		      DoubledGaugeField &U,
 		      DoubledGaugeField &UUU,
 		      const FermionField &in, 
@@ -116,7 +114,6 @@ public:
 		      int dag);
 
     void DhopInternalSerialComms(StencilImpl & st,
-		      LebesgueOrder &lo,
 		      DoubledGaugeField &U,
 		      DoubledGaugeField &UUU,
 		      const FermionField &in, 
@@ -192,8 +189,6 @@ public:
   DoubledGaugeField UUUmuEven;
   DoubledGaugeField UUUmuOdd;
     
-  LebesgueOrder Lebesgue;
-  LebesgueOrder LebesgueEvenOdd;
     
   // Comms buffer
   //  std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> >  comm_buf;

Grid/qcd/action/fermion/MobiusEOFAFermion.h

@@ -42,11 +42,11 @@ public:
 
 public:
   // Shift operator coefficients for red-black preconditioned Mobius EOFA
-  Vector<Coeff_t> Mooee_shift;
-  Vector<Coeff_t> MooeeInv_shift_lc;
-  Vector<Coeff_t> MooeeInv_shift_norm;
-  Vector<Coeff_t> MooeeInvDag_shift_lc;
-  Vector<Coeff_t> MooeeInvDag_shift_norm;
+  std::vector<Coeff_t> Mooee_shift;
+  std::vector<Coeff_t> MooeeInv_shift_lc;
+  std::vector<Coeff_t> MooeeInv_shift_norm;
+  std::vector<Coeff_t> MooeeInvDag_shift_lc;
+  std::vector<Coeff_t> MooeeInvDag_shift_norm;
 
   virtual void Instantiatable(void) {};
 
@@ -74,18 +74,18 @@ public:
   // Instantiate different versions depending on Impl
   /////////////////////////////////////////////////////
   void M5D(const FermionField& psi, const FermionField& phi, FermionField& chi,
-	   Vector<Coeff_t>& lower, Vector<Coeff_t>& diag, Vector<Coeff_t>& upper);
+	   std::vector<Coeff_t>& lower, std::vector<Coeff_t>& diag, std::vector<Coeff_t>& upper);
 
   void M5D_shift(const FermionField& psi, const FermionField& phi, FermionField& chi,
-		 Vector<Coeff_t>& lower, Vector<Coeff_t>& diag, Vector<Coeff_t>& upper,
-		 Vector<Coeff_t>& shift_coeffs);
+		 std::vector<Coeff_t>& lower, std::vector<Coeff_t>& diag, std::vector<Coeff_t>& upper,
+		 std::vector<Coeff_t>& shift_coeffs);
 
   void M5Ddag(const FermionField& psi, const FermionField& phi, FermionField& chi,
-	      Vector<Coeff_t>& lower, Vector<Coeff_t>& diag, Vector<Coeff_t>& upper);
+	      std::vector<Coeff_t>& lower, std::vector<Coeff_t>& diag, std::vector<Coeff_t>& upper);
 
   void M5Ddag_shift(const FermionField& psi, const FermionField& phi, FermionField& chi,
-		    Vector<Coeff_t>& lower, Vector<Coeff_t>& diag, Vector<Coeff_t>& upper,
-		    Vector<Coeff_t>& shift_coeffs);
+		    std::vector<Coeff_t>& lower, std::vector<Coeff_t>& diag, std::vector<Coeff_t>& upper,
+		    std::vector<Coeff_t>& shift_coeffs);
 
   virtual void RefreshShiftCoefficients(RealD new_shift);
 

Grid/qcd/action/fermion/NaiveStaggeredFermion.h

@@ -102,11 +102,11 @@ public:
 		     GaugeField &mat, 
 		     const FermionField &A, const FermionField &B, int dag);
 
-  void DhopInternal(StencilImpl &st, LebesgueOrder &lo, DoubledGaugeField &U,
+  void DhopInternal(StencilImpl &st, DoubledGaugeField &U,
                     const FermionField &in, FermionField &out, int dag);
-  void DhopInternalSerialComms(StencilImpl &st, LebesgueOrder &lo, DoubledGaugeField &U,
+  void DhopInternalSerialComms(StencilImpl &st, DoubledGaugeField &U,
 			       const FermionField &in, FermionField &out, int dag);
-  void DhopInternalOverlappedComms(StencilImpl &st, LebesgueOrder &lo, DoubledGaugeField &U,
+  void DhopInternalOverlappedComms(StencilImpl &st, DoubledGaugeField &U,
 				   const FermionField &in, FermionField &out, int dag);
 
   //////////////////////////////////////////////////////////////////////////
@@ -152,9 +152,6 @@ public:
   DoubledGaugeField UmuEven;
   DoubledGaugeField UmuOdd;
 
-  LebesgueOrder Lebesgue;
-  LebesgueOrder LebesgueEvenOdd;
-  
   ///////////////////////////////////////////////////////////////
   // Conserved current utilities
   ///////////////////////////////////////////////////////////////

Grid/qcd/action/fermion/OverlapWilsonCayleyZolotarevFermion.h

@@ -41,6 +41,10 @@ public:
 public:
 
   // Constructors
+  virtual void   Instantiatable(void){};
+  void  MomentumSpacePropagator(FermionField &out,const FermionField &in,RealD _m,std::vector<double> twist) {
+    this->MomentumSpacePropagatorHw(out,in,_m,twist);
+  };
 
   OverlapWilsonCayleyZolotarevFermion(GaugeField &_Umu,
 				      GridCartesian         &FiveDimGrid,

Grid/qcd/action/fermion/OverlapWilsonContfracTanhFermion.h

@@ -41,6 +41,9 @@ public:
 public:
 
   virtual void   Instantiatable(void){};
+  void  MomentumSpacePropagator(FermionField &out,const FermionField &in,RealD _m,std::vector<double> twist) {
+    this->MomentumSpacePropagatorHw(out,in,_m,twist);
+  };
   // Constructors
   OverlapWilsonContFracTanhFermion(GaugeField &_Umu,
 				   GridCartesian         &FiveDimGrid,

Grid/qcd/action/fermion/OverlapWilsonContfracZolotarevFermion.h

@@ -40,6 +40,9 @@ public:
   INHERIT_IMPL_TYPES(Impl);
 
   virtual void   Instantiatable(void){};
+  void  MomentumSpacePropagator(FermionField &out,const FermionField &in,RealD _m,std::vector<double> twist) {
+    this->MomentumSpacePropagatorHw(out,in,_m,twist);
+  };
   // Constructors
   OverlapWilsonContFracZolotarevFermion(GaugeField &_Umu,
 					GridCartesian         &FiveDimGrid,

Grid/qcd/action/fermion/OverlapWilsonPartialFractionTanhFermion.h

@@ -41,6 +41,9 @@ public:
 public:
 
   virtual void   Instantiatable(void){};
+  void  MomentumSpacePropagator(FermionField &out,const FermionField &in,RealD _m,std::vector<double> twist) {
+    this->MomentumSpacePropagatorHw(out,in,_m,twist);
+  };
   // Constructors
   OverlapWilsonPartialFractionTanhFermion(GaugeField &_Umu,
 					  GridCartesian         &FiveDimGrid,

Grid/qcd/action/fermion/OverlapWilsonPartialFractionZolotarevFermion.h

@@ -40,6 +40,11 @@ public:
   INHERIT_IMPL_TYPES(Impl);
 
   virtual void   Instantiatable(void){};
+
+  void  MomentumSpacePropagator(FermionField &out,const FermionField &in,RealD _m,std::vector<double> twist) {
+    this->MomentumSpacePropagatorHw(out,in,_m,twist);
+  };
+
   // Constructors
   OverlapWilsonPartialFractionZolotarevFermion(GaugeField &_Umu,
 					       GridCartesian         &FiveDimGrid,

Grid/qcd/action/fermion/PartialFractionFermion5D.h

@@ -39,7 +39,7 @@ class PartialFractionFermion5D : public WilsonFermion5D<Impl>
 public:
   INHERIT_IMPL_TYPES(Impl);
 
-  const int part_frac_chroma_convention=1;
+  const int part_frac_chroma_convention=0;
 
   void   Meooe_internal(const FermionField &in, FermionField &out,int dag);
   void   Mooee_internal(const FermionField &in, FermionField &out,int dag);
@@ -83,19 +83,78 @@ public:
 			   GridRedBlackCartesian &FourDimRedBlackGrid,
 			   RealD _mass,RealD M5,const ImplParams &p= ImplParams());
 
+  PartialFractionFermion5D(GaugeField &_Umu,
+			   GridCartesian         &FiveDimGrid,
+			   GridRedBlackCartesian &FiveDimRedBlackGrid,
+			   GridCartesian         &FourDimGrid,
+			   GridRedBlackCartesian &FourDimRedBlackGrid,
+			   RealD _mass,RealD M5,std::vector<RealD> &_qmu,const ImplParams &p= ImplParams());
+
+  void FreePropagator(const FermionField &in,FermionField &out,RealD mass,std::vector<Complex> boundary, std::vector<double> twist)
+  {
+    std::cout << "Free Propagator for PartialFraction"<<std::endl;
+    FermionField in_k(in.Grid());
+    FermionField prop_k(in.Grid());
+    
+    FFT theFFT((GridCartesian *) in.Grid());
+
+    //phase for boundary condition
+    ComplexField coor(in.Grid());
+    ComplexField ph(in.Grid());  ph = Zero();
+    FermionField in_buf(in.Grid()); in_buf = Zero();
+    typedef typename Simd::scalar_type Scalar;
+    Scalar ci(0.0,1.0);
+    assert(twist.size() == Nd);//check that twist is Nd
+    assert(boundary.size() == Nd);//check that boundary conditions is Nd
+    int shift = 0;
+    for(unsigned int nu = 0; nu < Nd; nu++)
+      {
+	// Shift coordinate lattice index by 1 to account for 5th dimension.
+	LatticeCoordinate(coor, nu + shift);
+	double boundary_phase = ::acos(real(boundary[nu]));
+	ph = ph + boundary_phase*coor*((1./(in.Grid()->_fdimensions[nu+shift])));
+	//momenta for propagator shifted by twist+boundary
+	twist[nu] = twist[nu] + boundary_phase/((2.0*M_PI));
+      }
+    in_buf = exp(ci*ph*(-1.0))*in;
+
+    theFFT.FFT_all_dim(in_k,in,FFT::forward);
+    if ( this->qmu.size() ){
+      this->MomentumSpacePropagatorHwQ(prop_k,in_k,mass,twist,this->qmu);
+    } else {
+      this->MomentumSpacePropagatorHw(prop_k,in_k,mass,twist);
+    }
+    theFFT.FFT_all_dim(out,prop_k,FFT::backward);
+    
+    //phase for boundary condition
+    out = out * exp(ci*ph);
+  };
+
+  virtual void FreePropagator(const FermionField &in,FermionField &out,RealD mass) {
+    std::vector<double> twist(Nd,0.0); //default: periodic boundarys in all directions
+    std::vector<Complex> boundary;
+    for(int i=0;i<Nd;i++) boundary.push_back(1);//default: periodic boundary conditions
+    FreePropagator(in,out,mass,boundary,twist);
+  };
+
+  void set_qmu(std::vector<RealD> _qmu) { qmu=_qmu; assert(qmu.size()==Nd);};
+  void addQmu(const FermionField &in, FermionField &out, int dag);
+
 protected:
 
   virtual void SetCoefficientsTanh(Approx::zolotarev_data *zdata,RealD scale);
   virtual void SetCoefficientsZolotarev(RealD zolo_hi,Approx::zolotarev_data *zdata);
 
+  std::vector<RealD> qmu;
+
   // Part frac
   RealD mass;
   RealD dw_diag;
   RealD R;
   RealD amax;
   RealD scale;
-  Vector<double> p; 
-  Vector<double> q;
+  std::vector<double> p; 
+  std::vector<double> q;
 
 };
 

Grid/qcd/action/fermion/SchurDiagTwoKappa.h

@@ -35,7 +35,7 @@ template<class Matrix, class Field>
 class KappaSimilarityTransform {
 public:
   INHERIT_IMPL_TYPES(Matrix);
-  Vector<Coeff_t> kappa, kappaDag, kappaInv, kappaInvDag;
+  std::vector<Coeff_t> kappa, kappaDag, kappaInv, kappaInvDag;
 
   KappaSimilarityTransform (Matrix &zmob) {
     for (int i=0;i<(int)zmob.bs.size();i++) {

Grid/qcd/action/fermion/StaggeredKernels.h

@@ -49,10 +49,10 @@ template<class Impl> class StaggeredKernels : public FermionOperator<Impl> , pub
    
  public:
 
-  void DhopImproved(StencilImpl &st, LebesgueOrder &lo, 
+  void DhopImproved(StencilImpl &st,
 		    DoubledGaugeField &U, DoubledGaugeField &UUU, 
 		    const FermionField &in, FermionField &out, int dag, int interior,int exterior);
-  void DhopNaive(StencilImpl &st, LebesgueOrder &lo, 
+  void DhopNaive(StencilImpl &st,
 		 DoubledGaugeField &U,
 		 const FermionField &in, FermionField &out, int dag, int interior,int exterior);
   

Grid/qcd/action/fermion/WilsonCompressor.h

@@ -47,7 +47,7 @@ public:
   static int PartialCompressionFactor(GridBase *grid) { return 1;}
 #endif
   template<class vobj,class cobj,class compressor>
-  static void Gather_plane_simple (commVector<std::pair<int,int> >& table,
+  static void Gather_plane_simple (deviceVector<std::pair<int,int> >& table,
 				   const Lattice<vobj> &rhs,
 				   cobj *buffer,
 				   compressor &compress,
@@ -109,7 +109,7 @@ public:
   // Reorder the fifth dim to be s=Ls-1 , s=0, s=1,...,Ls-2.
   ////////////////////////////////////////////////////////////////////////////////////////////
   template<class vobj,class cobj,class compressor>
-  static void Gather_plane_exchange(commVector<std::pair<int,int> >& table,const Lattice<vobj> &rhs,
+  static void Gather_plane_exchange(deviceVector<std::pair<int,int> >& table,const Lattice<vobj> &rhs,
 				    std::vector<cobj *> pointers,int dimension,int plane,int cbmask,
 				    compressor &compress,int type,int partial)
   {
@@ -197,7 +197,7 @@ public:
 #endif
   
   template<class vobj,class cobj,class compressor>
-  static void Gather_plane_simple (commVector<std::pair<int,int> >& table,
+  static void Gather_plane_simple (deviceVector<std::pair<int,int> >& table,
 					 const Lattice<vobj> &rhs,
 					 cobj *buffer,
 					 compressor &compress,
@@ -208,7 +208,7 @@ public:
     else        FaceGatherSimple::Gather_plane_simple(table,rhs,buffer,compress,off,so,partial);
   }
   template<class vobj,class cobj,class compressor>
-  static void Gather_plane_exchange(commVector<std::pair<int,int> >& table,const Lattice<vobj> &rhs,
+  static void Gather_plane_exchange(deviceVector<std::pair<int,int> >& table,const Lattice<vobj> &rhs,
 				    std::vector<cobj *> pointers,int dimension,int plane,int cbmask,
 				    compressor &compress,int type,int partial)
   {
@@ -402,7 +402,6 @@ public:
 
   typedef CartesianStencil<vobj,cobj,Parameters> Base;
   typedef typename Base::View_type View_type;
-  typedef typename Base::StencilVector StencilVector;
 
   //  Vector<int> surface_list;
   WilsonStencil(GridBase *grid,
@@ -416,29 +415,6 @@ public:
     this->same_node.resize(npoints);
   };
   
-  /*
-  void BuildSurfaceList(int Ls,int vol4){
-
-    // find same node for SHM
-    // Here we know the distance is 1 for WilsonStencil
-    for(int point=0;point<this->_npoints;point++){
-      this->same_node[point] = this->SameNode(point);
-    }
-    
-    for(int site = 0 ;site< vol4;site++){
-      int local = 1;
-      for(int point=0;point<this->_npoints;point++){
-	if( (!this->GetNodeLocal(site*Ls,point)) && (!this->same_node[point]) ){ 
-	  local = 0;
-	}
-      }
-      if(local == 0) { 
-	surface_list.push_back(site);
-      }
-    }
-  }
-  */
-  
   template < class compressor>
   void HaloExchangeOpt(const Lattice<vobj> &source,compressor &compress) 
   {

Grid/qcd/action/fermion/WilsonFermion.h

@@ -126,13 +126,16 @@ public:
   void DerivInternal(StencilImpl &st, DoubledGaugeField &U, GaugeField &mat,
                      const FermionField &A, const FermionField &B, int dag);
 
-  void DhopInternal(StencilImpl &st, LebesgueOrder &lo, DoubledGaugeField &U,
+  void DhopInternal(StencilImpl &st,
+		    DoubledGaugeField &U,
                     const FermionField &in, FermionField &out, int dag);
 
-  void DhopInternalSerial(StencilImpl &st, LebesgueOrder &lo, DoubledGaugeField &U,
+  void DhopInternalSerial(StencilImpl &st,
+			  DoubledGaugeField &U,
 			  const FermionField &in, FermionField &out, int dag);
 
-  void DhopInternalOverlappedComms(StencilImpl &st, LebesgueOrder &lo, DoubledGaugeField &U,
+  void DhopInternalOverlappedComms(StencilImpl &st,
+				   DoubledGaugeField &U,
 				   const FermionField &in, FermionField &out, int dag);
 
   // Constructor
@@ -168,9 +171,6 @@ public:
   DoubledGaugeField UmuEven;
   DoubledGaugeField UmuOdd;
 
-  LebesgueOrder Lebesgue;
-  LebesgueOrder LebesgueEvenOdd;
-
   WilsonAnisotropyCoefficients anisotropyCoeff;
 
   ///////////////////////////////////////////////////////////////

Grid/qcd/action/fermion/WilsonFermion5D.h

@@ -109,6 +109,8 @@ public:
   void MomentumSpacePropagatorHt_5d(FermionField &out,const FermionField &in,RealD mass,std::vector<double> twist) ;
   void MomentumSpacePropagatorHt(FermionField &out,const FermionField &in,RealD mass,std::vector<double> twist) ;
   void MomentumSpacePropagatorHw(FermionField &out,const FermionField &in,RealD mass,std::vector<double> twist) ;
+  void MomentumSpacePropagatorHwQ(FermionField &out,const FermionField &in,RealD mass,std::vector<double> twist,
+				  std::vector<double> qmu) ;
 
   // Implement hopping term non-hermitian hopping term; half cb or both
   // Implement s-diagonal DW
@@ -117,6 +119,9 @@ public:
   void DhopOE(const FermionField &in, FermionField &out,int dag);
   void DhopEO(const FermionField &in, FermionField &out,int dag);
 
+  void DhopComms  (const FermionField &in, FermionField &out);
+  void DhopCalc   (const FermionField &in, FermionField &out,uint64_t *ids);
+  
   // add a DhopComm
   // -- suboptimal interface will presently trigger multiple comms.
   void DhopDir(const FermionField &in, FermionField &out,int dir,int disp);
@@ -135,21 +140,18 @@ public:
 		     int dag);
     
   void DhopInternal(StencilImpl & st,
-		    LebesgueOrder &lo,
 		    DoubledGaugeField &U,
 		    const FermionField &in, 
 		    FermionField &out,
 		    int dag);
 
   void DhopInternalOverlappedComms(StencilImpl & st,
-				   LebesgueOrder &lo,
 				   DoubledGaugeField &U,
 				   const FermionField &in, 
 				   FermionField &out,
 				   int dag);
 
   void DhopInternalSerialComms(StencilImpl & st,
-			       LebesgueOrder &lo,
 			       DoubledGaugeField &U,
 			       const FermionField &in, 
 			       FermionField &out,
@@ -203,9 +205,6 @@ public:
   DoubledGaugeField UmuEven;
   DoubledGaugeField UmuOdd;
     
-  LebesgueOrder Lebesgue;
-  LebesgueOrder LebesgueEvenOdd;
-    
   // Comms buffer
   //  std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> >  comm_buf;
 

Grid/qcd/action/fermion/WilsonKernels.h

@@ -57,6 +57,10 @@ public:
 			 int Ls, int Nsite, const FermionField &in, FermionField &out,
 			 int interior=1,int exterior=1) ;
 
+  static void DhopKernel(int Opt,StencilImpl &st,  DoubledGaugeField &U, SiteHalfSpinor * buf,
+			 int Ls, int Nsite, const FermionField &in, FermionField &out,
+			 uint64_t *ids);
+  
   static void DhopDagKernel(int Opt,StencilImpl &st,  DoubledGaugeField &U, SiteHalfSpinor * buf,
 			    int Ls, int Nsite, const FermionField &in, FermionField &out,
 			    int interior=1,int exterior=1) ;