Merge branch 'develop' into feature/scidac-wp1

Need to consolidate files.
Plan: Make this version able to go virtual base, then absorb chulwoos
version when it is proven

.github/ISSUE_TEMPLATE/bug-report.yml

@@ -0,0 +1,54 @@
+name: Bug report
+description: Report a bug.
+title: "<insert title>"
+labels: [bug]
+
+body:
+  - type: markdown
+    attributes:
+      value: >
+        Thank you for taking the time to file a bug report.
+        Please check that the code is pointing to the HEAD of develop
+        or any commit in master which is tagged with a version number.
+
+  - type: textarea
+    attributes:
+      label: "Describe the issue:"
+      description: >
+        Describe the issue and any previous attempt to solve it.
+    validations:
+      required: true
+
+  - type: textarea
+    attributes:
+      label: "Code example:"
+      description: >
+        If relevant, show how to reproduce the issue using a minimal working
+        example.
+      placeholder: |
+        << your code here >>
+      render: shell
+    validations:
+      required: false
+
+  - type: textarea
+    attributes:
+      label: "Target platform:"
+      description: >
+        Give a description of the target platform (CPU, network, compiler).
+        Please give the full CPU part description, using for example
+        `cat /proc/cpuinfo | grep 'model name' | uniq` (Linux)
+        or `sysctl machdep.cpu.brand_string` (macOS) and the full output
+        the `--version` option of your compiler.
+    validations:
+      required: true
+
+  - type: textarea
+    attributes:
+      label: "Configure options:"
+      description: >
+        Please give the exact configure command used and attach
+        `config.log`, `grid.config.summary` and the output of `make V=1`.
+      render: shell
+    validations:
+      required: true

.gitignore

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

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/Makefile.am

@@ -66,6 +66,10 @@ if BUILD_FERMION_REPS
   extra_sources+=$(ADJ_FERMION_FILES)
   extra_sources+=$(TWOIND_FERMION_FILES)
 endif
+if BUILD_SP
+    extra_sources+=$(SP_FERMION_FILES)
+    extra_sources+=$(SP_TWOIND_FERMION_FILES)
+endif
 
 lib_LIBRARIES = libGrid.a
 

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 +47,10 @@ 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>
+NAMESPACE_CHECK(deflation);
 #include <Grid/algorithms/iterative/ConjugateGradient.h>
 NAMESPACE_CHECK(ConjGrad);
 #include <Grid/algorithms/iterative/BiCGSTAB.h>
@@ -67,10 +73,11 @@ 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/CoarsenedMatrix.h>
-NAMESPACE_CHECK(CoarsendMatrix);
+#include <Grid/algorithms/multigrid/MultiGrid.h>
+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>

Grid/algorithms/LinearOperator.h

@@ -145,6 +145,44 @@ 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
 ////////////////////////////////////////////////////////////////////
@@ -542,6 +580,7 @@ public:
       (*this)(in[i], out[i]);
     }
   }
+  virtual ~LinearFunction(){};
 };
 
 template<class Field> class IdentityLinearFunction : public LinearFunction<Field> {

Grid/algorithms/approx/Chebyshev.h

@@ -90,9 +90,8 @@ public:
     order=_order;
       
     if(order < 2) exit(-1);
-    Coeffs.resize(order);
-    Coeffs.assign(0.,order);
-    Coeffs[order-1] = 1.;
+    Coeffs.resize(order,0.0);
+    Coeffs[order-1] = 1.0;
   };
   
   // PB - more efficient low pass drops high modes above the low as 1/x uses all Chebyshev's.

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 +1,727 @@
+/*************************************************************************************
+
+    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 cl::sycl::queue *gridblasHandle_t;
+#endif
+#ifdef GRID_ONE_MKL
+  typedef cl::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
+      cl::sycl::cpu_selector selector;
+      cl::sycl::device selectedDevice { selector };
+      gridblasHandle =new sycl::queue (selectedDevice);
+#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);
+
+    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
+    //MKL’s cblas_<T>gemm_batch & OneAPI
+#warning "oneMKL implementation not built "
+#endif
+#if !defined(GRID_SYCL) && !defined(GRID_CUDA) && !defined(GRID_HIP)
+    // Need a default/reference implementation
+    int sda = lda*k;
+    int sdb = ldb*k;
+    int sdc = ldc*n;
+    for (int p = 0; p < batchCount; ++p) {
+      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)
+	    c_mn += Amk[p][mm + kk*lda ] * Bkn[p][kk + nn*ldb];
+	  Cmn[p][mm + nn*ldc] =  (alpha)*c_mn + (beta)*Cmn[p][mm + nn*ldc ];
+	}
+      }
+    }
+#endif
+    //    synchronise();
+     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();
+
+    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
+    //MKL’s cblas_<T>gemm_batch & OneAPI
+#warning "oneMKL implementation not built "
+#endif
+#if !defined(GRID_SYCL) && !defined(GRID_CUDA) && !defined(GRID_HIP)
+    int sda = lda*k;
+    int sdb = ldb*k;
+    int sdc = ldc*n;
+    ComplexF alphaf(real(alpha),imag(alpha));
+    ComplexF betaf(real(beta),imag(beta));
+    // Need a default/reference implementation
+    for (int p = 0; p < batchCount; ++p) {
+      for (int mm = 0; mm < m; ++mm) {
+	for (int nn = 0; nn < n; ++nn) {
+	  ComplexF c_mn(0.0);
+	  for (int kk = 0; kk < k; ++kk)
+	    c_mn += Amk[p][mm + kk*lda ] * Bkn[p][kk + nn*ldb];
+	  Cmn[p][mm + nn*ldc] =  (alphaf)*c_mn + (betaf)*Cmn[p][mm + nn*ldc ];
+	}
+      }
+    }
+#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();
+
+    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
+    //MKL’s cblas_<T>gemm_batch & OneAPI
+#warning "oneMKL implementation not built "
+#endif
+#if !defined(GRID_SYCL) && !defined(GRID_CUDA) && !defined(GRID_HIP)
+    int sda = lda*k;
+    int sdb = ldb*k;
+    int sdc = ldc*n;
+    // Need a default/reference implementation
+    for (int p = 0; p < batchCount; ++p) {
+      for (int mm = 0; mm < m; ++mm) {
+	for (int nn = 0; nn < n; ++nn) {
+	  RealD c_mn(0.0);
+	  for (int kk = 0; kk < k; ++kk)
+	    c_mn += Amk[p][mm + kk*lda ] * Bkn[p][kk + nn*ldb];
+	  Cmn[p][mm + nn*ldc] =  (alpha)*c_mn + (beta)*Cmn[p][mm + nn*ldc ];
+	}
+      }
+    }
+#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();
+
+    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 batchCount64=batchCount;
+      oneapi::mkl::blas::column_major::gemm_batch(*theGridAccelerator,
+      onemkl::transpose::N,
+      onemkl::transpose::N,
+      &m64,&n64,&k64,
+      (double *) &alpha_p[0],
+      (double **)&Amk[0], lda,
+      (double **)&Bkn[0], ldb,
+      (double *) &beta_p[0],
+      (double **)&Cmn[0], ldc,
+      1,&batchCount64);
+     */
+    //MKL’s cblas_<T>gemm_batch & OneAPI
+#warning "oneMKL implementation not built "
+#endif
+#if !defined(GRID_SYCL) && !defined(GRID_CUDA) && !defined(GRID_HIP)
+    int sda = lda*k;
+    int sdb = ldb*k;
+    int sdc = ldc*n;
+    // Need a default/reference implementation
+    for (int p = 0; p < batchCount; ++p) {
+      for (int mm = 0; mm < m; ++mm) {
+	for (int nn = 0; nn < n; ++nn) {
+	  RealD c_mn(0.0);
+	  for (int kk = 0; kk < k; ++kk)
+	    c_mn += Amk[p][mm + kk*lda ] * Bkn[p][kk + nn*ldb];
+	  Cmn[p][mm + nn*ldc] =  (alpha)*c_mn + (beta)*Cmn[p][mm + nn*ldc ];
+	}
+      }
+    }
+#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;
+  }
+  
+
+  
+  ////////////////////////////////////////////////////////////////////////////////////////////////
+  // Strided case used by benchmark, but generally unused in Grid
+  // Keep a code example in double complex, but don't generate the single and real variants for now
+  ////////////////////////////////////////////////////////////////////////////////////////////////
+  
+  void gemmStridedBatched(int m,int n, int k,
+			  ComplexD alpha,
+			  ComplexD* Amk,  // pointer list to matrices
+			  ComplexD* Bkn,
+			  ComplexD beta,
+			  ComplexD* Cmn,
+			  int batchCount)
+  {
+    // Use C-row major storage, so transpose calls
+    int lda = m; // m x k column major
+    int ldb = k; // k x n column major
+    int ldc = m; // m x b column major
+    int sda = m*k;
+    int sdb = k*n;
+    int sdc = m*n;
+    deviceVector<ComplexD> alpha_p(1);
+    deviceVector<ComplexD> beta_p(1);
+    acceleratorCopyToDevice((void *)&alpha,(void *)&alpha_p[0],sizeof(ComplexD));
+    acceleratorCopyToDevice((void *)&beta ,(void *)&beta_p[0],sizeof(ComplexD));
+    //    std::cout << "blasZgemmStridedBatched mnk  "<<m<<","<<n<<","<<k<<" count "<<batchCount<<std::endl;
+    //    std::cout << "blasZgemmStridedBatched ld   "<<lda<<","<<ldb<<","<<ldc<<std::endl;
+    //    std::cout << "blasZgemmStridedBatched sd   "<<sda<<","<<sdb<<","<<sdc<<std::endl;
+#ifdef GRID_HIP
+    auto err = hipblasZgemmStridedBatched(gridblasHandle,
+					  HIPBLAS_OP_N,
+					  HIPBLAS_OP_N,
+					  m,n,k,
+					  (hipblasDoubleComplex *) &alpha_p[0],
+					  (hipblasDoubleComplex *) Amk, lda, sda,
+					  (hipblasDoubleComplex *) Bkn, ldb, sdb,
+					  (hipblasDoubleComplex *) &beta_p[0],
+					  (hipblasDoubleComplex *) Cmn, ldc, sdc,
+					  batchCount);
+    assert(err==HIPBLAS_STATUS_SUCCESS);
+#endif
+#ifdef GRID_CUDA
+    cublasZgemmStridedBatched(gridblasHandle,
+			      CUBLAS_OP_N,
+			      CUBLAS_OP_N,
+			      m,n,k,
+			      (cuDoubleComplex *) &alpha_p[0],
+			      (cuDoubleComplex *) Amk, lda, sda,
+			      (cuDoubleComplex *) Bkn, ldb, sdb,
+			      (cuDoubleComplex *) &beta_p[0],
+			      (cuDoubleComplex *) Cmn, ldc, sdc,
+			      batchCount);
+#endif
+#if defined(GRID_SYCL) || defined(GRID_ONE_MKL)
+    oneapi::mkl::blas::column_major::gemm_batch(*gridblasHandle,
+						oneapi::mkl::transpose::N,
+						oneapi::mkl::transpose::N,
+						m,n,k,
+						alpha,
+						(const ComplexD *)Amk,lda,sda,
+						(const ComplexD *)Bkn,ldb,sdb,
+						beta,
+						(ComplexD *)Cmn,ldc,sdc,
+						batchCount);
+#endif
+#if !defined(GRID_SYCL) && !defined(GRID_CUDA) && !defined(GRID_HIP) && !defined(GRID_ONE_MKL)
+     // Need a default/reference implementation
+     for (int p = 0; p < batchCount; ++p) {
+       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)
+	     c_mn += Amk[mm + kk*lda + p*sda] * Bkn[kk + nn*ldb + p*sdb];
+	   Cmn[mm + nn*ldc + p*sdc] =  (alpha)*c_mn + (beta)*Cmn[mm + nn*ldc + p*sdc];
+	 }
+       }
+     }
+#endif
+  }
+
+  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<ComplexD> A(N_A); acceleratorMemSet(&A[0],0,N_A*sizeof(ComplexD));
+    deviceVector<ComplexD> B(N_B); acceleratorMemSet(&B[0],0,N_B*sizeof(ComplexD));
+    deviceVector<ComplexD> C(N_C); acceleratorMemSet(&C[0],0,N_C*sizeof(ComplexD));
+    ComplexD alpha(1.0);
+    ComplexD beta (1.0);
+    RealD flops = 8.0*M*N*K*BATCH;
+    int ncall=10;
+    RealD t0 = usecond();
+    for(int i=0;i<ncall;i++){
+      gemmStridedBatched(M,N,K,
+			 alpha,
+			 &A[0], // m x k 
+			 &B[0], // k x n
+			 beta, 
+			 &C[0], // m x n
+			 BATCH);
+    }
+    synchronise();
+    RealD t1 = usecond();
+    RealD bytes = 1.0*sizeof(ComplexD)*(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/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,
+		     ComplexD(1.0),
+		     Vd,
+		     Fd,
+		     ComplexD(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,
+		     ComplexD(1.0),
+		     Vd,
+		     Cd,
+		     ComplexD(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,
+		     ComplexD(1.0),
+		     Ed,
+		     Rd,
+		     ComplexD(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,
+		     ComplexD(1.0),
+		     Ed, // x . nev
+		     Cd, // nev . nrhs
+		     ComplexD(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

@@ -33,109 +33,111 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
    * Script A = SolverMatrix 
    * Script P = Preconditioner
    *
-   * Deflation methods considered
-   *      -- Solve P A x = P b        [ like Luscher ]
-   * DEF-1        M P A x = M P b     [i.e. left precon]
-   * DEF-2        P^T M A x = P^T M b
-   * ADEF-1       Preconditioner = M P + Q      [ Q + M + M A Q]
-   * ADEF-2       Preconditioner = P^T M + Q
-   * BNN          Preconditioner = P^T M P + Q
-   * BNN2         Preconditioner = M P + P^TM +Q - M P A M 
-   * 
    * Implement ADEF-2
    *
    * Vstart = P^Tx + Qb
    * M1 = P^TM + Q
    * M2=M3=1
-   * Vout = x
    */
+NAMESPACE_BEGIN(Grid);
 
-// abstract base
-template<class Field, class CoarseField>
-class TwoLevelFlexiblePcg : public LinearFunction<Field>
+
+template<class Field>
+class TwoLevelCG : public LinearFunction<Field>
 {
  public:
-  int verbose;
   RealD   Tolerance;
   Integer MaxIterations;
-  const int mmax = 5;
   GridBase *grid;
-  GridBase *coarsegrid;
 
-  LinearOperatorBase<Field>   *_Linop
-  OperatorFunction<Field>     *_Smoother,
-  LinearFunction<CoarseField> *_CoarseSolver;
-
-  // Need somthing that knows how to get from Coarse to fine and back again
+  // Fine operator, Smoother, CoarseSolver
+  LinearOperatorBase<Field>   &_FineLinop;
+  LinearFunction<Field>   &_Smoother;
   
   // more most opertor functions
-  TwoLevelFlexiblePcg(RealD tol,
-		     Integer maxit,
-		     LinearOperatorBase<Field> *Linop,
-		     LinearOperatorBase<Field> *SmootherLinop,
-		     OperatorFunction<Field>   *Smoother,
-		     OperatorFunction<CoarseField>  CoarseLinop
-		     ) : 
+  TwoLevelCG(RealD tol,
+	     Integer maxit,
+	     LinearOperatorBase<Field>   &FineLinop,
+	     LinearFunction<Field>       &Smoother,
+	     GridBase *fine) : 
       Tolerance(tol), 
       MaxIterations(maxit),
-      _Linop(Linop),
-      _PreconditionerLinop(PrecLinop),
-      _Preconditioner(Preconditioner)
-  { 
-    verbose=0;
+      _FineLinop(FineLinop),
+      _Smoother(Smoother)
+  {
+    grid       = fine;
   };
-
-  // The Pcg routine is common to all, but the various matrices differ from derived 
-  // implementation to derived implmentation
-  void operator() (const Field &src, Field &psi){
-  void operator() (const Field &src, Field &psi){
-
-    psi.Checkerboard() = src.Checkerboard();
-    grid             = src.Grid();
-
+  
+  virtual void operator() (const Field &src, Field &x)
+  {
+    std::cout << GridLogMessage<<"HDCG: fPcg starting single RHS"<<std::endl;
     RealD f;
     RealD rtzp,rtz,a,d,b;
     RealD rptzp;
-    RealD tn;
-    RealD guess = norm2(psi);
-    RealD ssq   = norm2(src);
-    RealD rsq   = ssq*Tolerance*Tolerance;
-    
+
     /////////////////////////////
     // Set up history vectors
     /////////////////////////////
-    std::vector<Field> p  (mmax,grid);
+    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 x  (grid); x = psi;
-    Field z  (grid);
+    Field z(grid);
     Field tmp(grid);
-    Field r  (grid);
-    Field mu (grid);
-  
+    Field  mp (grid);
+    Field  r  (grid);
+    Field  mu (grid);
+    
+    std::cout << GridLogMessage<<"HDCG: fPcg allocated"<<std::endl;
+    //Initial residual computation & set up
+    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=src;
     Vstart(x,src);
-
+    
     // r0 = b -A x0
-    HermOp(x,mmp); // Shouldn't this be something else?
+    _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
     //////////////////////////////////
-    M1(r,z,tmp,mp,SmootherMirs);
+    PcgM1(r,z);
     rtzp =real(innerProduct(r,z));
-
+    
     ///////////////////////////////////////
     // Solve for Mss mu = P A z and set p = z-mu
-    // Def2: p = 1 - Q Az = Pright z 
+    // Def2 p = 1 - Q Az = Pright z
     // Other algos M2 is trivial
     ///////////////////////////////////////
-    M2(z,p[0]);
+    PcgM2(z,p[0]);
+
+    RealD ssq =  norm2(src);
+    RealD rsq =  ssq*Tolerance*Tolerance;
+
+    std::cout << GridLogMessage<<"HDCG: k=0 residual "<<rtzp<<" rsq "<<rsq<<"\n";
+
+    Field pp(grid);
 
     for (int k=0;k<=MaxIterations;k++){
     
@@ -143,31 +145,46 @@ class TwoLevelFlexiblePcg : public LinearFunction<Field>
       int peri_kp = (k+1) % mmax;
 
       rtz=rtzp;
-      d= M3(p[peri_k],mp,mmp[peri_k],tmp);
+      d= PcgM3(p[peri_k],mmp[peri_k]);
       a = rtz/d;
     
       // 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
-      M1(r,z,tmp,mp);
-
+      PcgM1(r,z);
+      
+      {
+	RealD n1,n2;
+	n1=norm2(r);
+	n2=norm2(z);
+	std::cout << GridLogMessage<<"HDCG::fPcg iteration "<<k<<" : vector r,z "<<n1<<" "<<n2<<"\n";
+      }
       rtzp =real(innerProduct(r,z));
+      std::cout << GridLogMessage<<"HDCG::fPcg iteration "<<k<<" : inner rtzp "<<rtzp<<"\n";
 
-      M2(z,mu); // ADEF-2 this is identity. Axpy possible to eliminate
+      //    PcgM2(z,p[0]);
+      PcgM2(z,mu); // ADEF-2 this is identity. Axpy possible to eliminate
+      
+      p[peri_kp]=mu;
 
-      p[peri_kp]=p[peri_k];
-
-      // Standard search direction  p -> z + b p    ; b = 
+      // 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]));
@@ -176,75 +193,324 @@ class TwoLevelFlexiblePcg : public LinearFunction<Field>
       }
 
       RealD rrn=sqrt(rn/ssq);
-      std::cout<<GridLogMessage<<"TwoLevelfPcg: k= "<<k<<" residual = "<<rrn<<std::endl;
+      RealD rtn=sqrt(rtz/ssq);
+      RealD rtnp=sqrt(rtzp/ssq);
+
+      std::cout<<GridLogMessage<<"HDCG: fPcg k= "<<k<<" residual = "<<rrn<<"\n";
 
       // Stopping condition
       if ( rn <= rsq ) { 
 
-	HermOp(x,mmp); // Shouldn't this be something else?
+	HDCGTimer.Stop();
+	std::cout<<GridLogMessage<<"HDCG: fPcg converged in "<<k<<" iterations and "<<HDCGTimer.Elapsed()<<std::endl;;
+	
+	_FineLinop.HermOp(x,mmp[0]);			  
 	axpy(tmp,-1.0,src,mmp[0]);
 	
-	RealD psinorm = sqrt(norm2(x));
-	RealD srcnorm = sqrt(norm2(src));
-	RealD tmpnorm = sqrt(norm2(tmp));
-	RealD true_residual = tmpnorm/srcnorm;
-	std::cout<<GridLogMessage<<"TwoLevelfPcg:   true residual is "<<true_residual<<std::endl;
-	std::cout<<GridLogMessage<<"TwoLevelfPcg: target residual was"<<Tolerance<<std::endl;
-	return k;
+	RealD  mmpnorm = sqrt(norm2(mmp[0]));
+	RealD  xnorm   = sqrt(norm2(x));
+	RealD  srcnorm = sqrt(norm2(src));
+	RealD  tmpnorm = sqrt(norm2(tmp));
+	RealD  true_residual = tmpnorm/srcnorm;
+	std::cout<<GridLogMessage
+	       <<"HDCG: true residual is "<<true_residual
+	       <<" solution "<<xnorm
+	       <<" source "<<srcnorm
+	       <<" mmp "<<mmpnorm	  
+	       <<std::endl;
+      
+	return;
       }
+
     }
-    // Non-convergence
-    assert(0);
+    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 M(Field & in,Field & out,Field & tmp) {
+  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) {
+    out=in;
   }
 
-  virtual void M1(Field & in, Field & out) {// the smoother
+  virtual RealD PcgM3(const Field & p, Field & mmp){
+    RealD dd;
+    _FineLinop.HermOp(p,mmp);
+    ComplexD dot = innerProduct(p,mmp);
+    dd=real(dot);
+    return dd;
+  }
 
+  /////////////////////////////////////////////////////////////////////
+  // Only Def1 has non-trivial Vout.
+  /////////////////////////////////////////////////////////////////////
+
+};
+  
+template<class Field, class CoarseField, class Aggregation>
+class TwoLevelADEF2 : public TwoLevelCG<Field>
+{
+ public:
+  ///////////////////////////////////////////////////////////////////////////////////
+  // Need something that knows how to get from Coarse to fine and back again
+  //  void ProjectToSubspace(CoarseVector &CoarseVec,const FineField &FineVec){
+  //  void PromoteFromSubspace(const CoarseVector &CoarseVec,FineField &FineVec){
+  ///////////////////////////////////////////////////////////////////////////////////
+  GridBase *coarsegrid;
+  Aggregation &_Aggregates;                    
+  LinearFunction<CoarseField> &_CoarseSolver;
+  LinearFunction<CoarseField> &_CoarseSolverPrecise;
+  ///////////////////////////////////////////////////////////////////////////////////
+  
+  // more most opertor functions
+  TwoLevelADEF2(RealD tol,
+		Integer maxit,
+		LinearOperatorBase<Field>    &FineLinop,
+		LinearFunction<Field>        &Smoother,
+		LinearFunction<CoarseField>  &CoarseSolver,
+		LinearFunction<CoarseField>  &CoarseSolverPrecise,
+		Aggregation &Aggregates
+		) :
+      TwoLevelCG<Field>(tol,maxit,FineLinop,Smoother,Aggregates.FineGrid),
+      _CoarseSolver(CoarseSolver),
+      _CoarseSolverPrecise(CoarseSolverPrecise),
+      _Aggregates(Aggregates)
+  {
+    coarsegrid = Aggregates.CoarseGrid;
+  };
+
+  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(grid);
-    Field Min(grid);
 
-    PcgM(in,Min); // Smoother call
+    Field tmp(this->grid);
+    Field Min(this->grid);
+    CoarseField PleftProj(this->coarsegrid);
+    CoarseField PleftMss_proj(this->coarsegrid);
 
-    HermOp(Min,out);
+    GridStopWatch SmootherTimer;
+    GridStopWatch MatrixTimer;
+    SmootherTimer.Start();
+    this->_Smoother(in,Min);
+    SmootherTimer.Stop();
+
+    MatrixTimer.Start();
+    this->_FineLinop.HermOp(Min,out);
+    MatrixTimer.Stop();
     axpy(tmp,-1.0,out,in);          // tmp  = in - A Min
 
-    ProjectToSubspace(tmp,PleftProj);     
-    ApplyInverse(PleftProj,PleftMss_proj); // Ass^{-1} [in - A Min]_s
-    PromoteFromSubspace(PleftMss_proj,tmp);// tmp = Q[in - A Min]  
+    GridStopWatch ProjTimer;
+    GridStopWatch CoarseTimer;
+    GridStopWatch PromTimer;
+    ProjTimer.Start();
+    this->_Aggregates.ProjectToSubspace(PleftProj,tmp);     
+    ProjTimer.Stop();
+    CoarseTimer.Start();
+    this->_CoarseSolver(PleftProj,PleftMss_proj); // Ass^{-1} [in - A Min]_s
+    CoarseTimer.Stop();
+    PromTimer.Start();
+    this->_Aggregates.PromoteFromSubspace(PleftMss_proj,tmp);// tmp = Q[in - A Min]  
+    PromTimer.Stop();
+    std::cout << GridLogPerformance << "PcgM1 breakdown "<<std::endl;
+    std::cout << GridLogPerformance << "\tSmoother   " << SmootherTimer.Elapsed() <<std::endl;
+    std::cout << GridLogPerformance << "\tMatrix     " << MatrixTimer.Elapsed() <<std::endl;
+    std::cout << GridLogPerformance << "\tProj       " << ProjTimer.Elapsed() <<std::endl;
+    std::cout << GridLogPerformance << "\tCoarse     " << CoarseTimer.Elapsed() <<std::endl;
+    std::cout << GridLogPerformance << "\tProm       " << PromTimer.Elapsed() <<std::endl;
+
     axpy(out,1.0,Min,tmp); // Min+tmp
   }
 
-  virtual void M2(const Field & in, Field & out) {
-    out=in;
-    // Must override for Def2 only
-    //  case PcgDef2:
-    //    Pright(in,out);
-    //    break;
-  }
-
-  virtual RealD M3(const Field & p, Field & mmp){
-    double d,dd;
-    HermOpAndNorm(p,mmp,d,dd);
-    return dd;
-    // Must override for Def1 only
-    //  case PcgDef1:
-    //    d=linop_d->Mprec(p,mmp,tmp,0,1);// Dag no
-    //      linop_d->Mprec(mmp,mp,tmp,1);// Dag yes
-    //    Pleft(mp,mmp);
-    //    d=real(linop_d->inner(p,mmp));
-  }
-
-  virtual void VstartDef2(Field & xconst Field & src){
-    //case PcgDef2:
-    //case PcgAdef2: 
-    //case PcgAdef2f:
-    //case PcgV11f:
+  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]
@@ -256,142 +522,78 @@ class TwoLevelFlexiblePcg : public LinearFunction<Field>
     //                   = src_s - (A guess)_s - src_s  + (A guess)_s 
     //                   = 0 
     ///////////////////////////////////
-    Field r(grid);
-    Field mmp(grid);
-    
-    HermOp(x,mmp);
-    axpy (r, -1.0, mmp, src);        // r_{-1} = src - A x
-    ProjectToSubspace(r,PleftProj);     
-    ApplyInverseCG(PleftProj,PleftMss_proj); // Ass^{-1} r_s
-    PromoteFromSubspace(PleftMss_proj,mmp);  
-    x=x+mmp;
+    Field r(this->grid);
+    Field mmp(this->grid);
+    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>
+{
+public:
+  const std::vector<Field> &evec;
+  const std::vector<RealD> &eval;
+  
+  TwoLevelADEF1defl(RealD tol,
+		   Integer maxit,
+		   LinearOperatorBase<Field>   &FineLinop,
+		   LinearFunction<Field>   &Smoother,
+		   std::vector<Field> &_evec,
+		   std::vector<RealD> &_eval) : 
+    TwoLevelCG<Field>(tol,maxit,FineLinop,Smoother,_evec[0].Grid()),
+    evec(_evec),
+    eval(_eval)
+  {};
+
+  // Can just inherit existing M2
+  // Can just inherit existing M3
+
+  // Simple vstart - do nothing
   virtual void Vstart(Field & x,const Field & src){
-    return;
+    x=src; // Could apply Q
+  };
+
+  // Override PcgM1
+  virtual void PcgM1(Field & in, Field & out)
+  {
+    GRID_TRACE("EvecPreconditioner ");
+    int N=evec.size();
+    Field Pin(this->grid);
+    Field Qin(this->grid);
+
+    //MP  + Q = M(1-AQ) + Q = M
+    // // If we are eigenvector deflating in coarse space
+    // // Q   = Sum_i |phi_i> 1/lambda_i <phi_i|
+    // // A Q = Sum_i |phi_i> <phi_i|
+    // // M(1-AQ) = M(1-proj) + Q
+    Qin.Checkerboard()=in.Checkerboard();
+    Qin = Zero();
+    Pin = in;
+    for (int i=0;i<N;i++) {
+      const Field& tmp = evec[i];
+      auto ip = TensorRemove(innerProduct(tmp,in));
+      axpy(Qin, ip / eval[i],tmp,Qin);
+      axpy(Pin, -ip ,tmp,Pin);
+    }
+
+    this->_Smoother(Pin,out);
+
+    out = out + Qin;
   }
+};
 
-  /////////////////////////////////////////////////////////////////////
-  // Only Def1 has non-trivial Vout. Override in Def1
-  /////////////////////////////////////////////////////////////////////
-  virtual void   Vout  (Field & in, Field & out,Field & src){
-    out = in;
-    //case PcgDef1:
-    //    //Qb + PT x
-    //    ProjectToSubspace(src,PleftProj);     
-    //    ApplyInverse(PleftProj,PleftMss_proj); // Ass^{-1} r_s
-    //    PromoteFromSubspace(PleftMss_proj,tmp);  
-    //    
-    //    Pright(in,out);
-    //    
-    //    linop_d->axpy(out,tmp,out,1.0);
-    //    break;
-  }
+NAMESPACE_END(Grid);
 
-  ////////////////////////////////////////////////////////////////////////////////////////////////
-  // Pright and Pleft are common to all implementations
-  ////////////////////////////////////////////////////////////////////////////////////////////////
-  virtual void Pright(Field & in,Field & out){
-    // P_R  = [ 1              0 ] 
-    //        [ -Mss^-1 Msb    0 ] 
-    Field in_sbar(grid);
-
-    ProjectToSubspace(in,PleftProj);     
-    PromoteFromSubspace(PleftProj,out);  
-    axpy(in_sbar,-1.0,out,in);       // in_sbar = in - in_s 
-
-    HermOp(in_sbar,out);
-    ProjectToSubspace(out,PleftProj);           // Mssbar in_sbar  (project)
-
-    ApplyInverse     (PleftProj,PleftMss_proj); // Mss^{-1} Mssbar 
-    PromoteFromSubspace(PleftMss_proj,out);     // 
-
-    axpy(out,-1.0,out,in_sbar);     // in_sbar - Mss^{-1} Mssbar in_sbar
-  }
-  virtual void Pleft (Field & in,Field & out){
-    // P_L  = [ 1  -Mbs Mss^-1] 
-    //        [ 0   0         ] 
-    Field in_sbar(grid);
-    Field    tmp2(grid);
-    Field    Mtmp(grid);
-
-    ProjectToSubspace(in,PleftProj);     
-    PromoteFromSubspace(PleftProj,out);  
-    axpy(in_sbar,-1.0,out,in);      // in_sbar = in - in_s
-
-    ApplyInverse(PleftProj,PleftMss_proj); // Mss^{-1} in_s
-    PromoteFromSubspace(PleftMss_proj,out);
-
-    HermOp(out,Mtmp);
-
-    ProjectToSubspace(Mtmp,PleftProj);      // Msbar s Mss^{-1}
-    PromoteFromSubspace(PleftProj,tmp2);
-
-    axpy(out,-1.0,tmp2,Mtmp);
-    axpy(out,-1.0,out,in_sbar);     // in_sbar - Msbars Mss^{-1} in_s
-  }
-}
-
-template<class Field>
-class TwoLevelFlexiblePcgADef2 : public TwoLevelFlexiblePcg<Field> {
- public:
-  virtual void M(Field & in,Field & out,Field & tmp){
-
-  } 
-  virtual void M1(Field & in, Field & out,Field & tmp,Field & mp){
-
-  }
-  virtual void M2(Field & in, Field & out){
-
-  }
-  virtual RealD M3(Field & p, Field & mp,Field & mmp, Field & tmp){
-
-  }
-  virtual void Vstart(Field & in, Field & src, Field & r, Field & mp, Field & mmp, Field & tmp){
-
-  }
-}
-/*
-template<class Field>
-class TwoLevelFlexiblePcgAD : public TwoLevelFlexiblePcg<Field> {
- public:
-  virtual void M(Field & in,Field & out,Field & tmp); 
-  virtual void M1(Field & in, Field & out,Field & tmp,Field & mp);
-  virtual void M2(Field & in, Field & out);
-  virtual RealD M3(Field & p, Field & mp,Field & mmp, Field & tmp);
-  virtual void Vstart(Field & in, Field & src, Field & r, Field & mp, Field & mmp, Field & tmp);
-}
-
-template<class Field>
-class TwoLevelFlexiblePcgDef1 : public TwoLevelFlexiblePcg<Field> {
- public:
-  virtual void M(Field & in,Field & out,Field & tmp); 
-  virtual void M1(Field & in, Field & out,Field & tmp,Field & mp);
-  virtual void M2(Field & in, Field & out);
-  virtual RealD M3(Field & p, Field & mp,Field & mmp, Field & tmp);
-  virtual void Vstart(Field & in, Field & src, Field & r, Field & mp, Field & mmp, Field & tmp);
-  virtual void   Vout  (Field & in, Field & out,Field & src,Field & tmp);
-}
-
-template<class Field>
-class TwoLevelFlexiblePcgDef2 : public TwoLevelFlexiblePcg<Field> {
- public:
-  virtual void M(Field & in,Field & out,Field & tmp); 
-  virtual void M1(Field & in, Field & out,Field & tmp,Field & mp);
-  virtual void M2(Field & in, Field & out);
-  virtual RealD M3(Field & p, Field & mp,Field & mmp, Field & tmp);
-  virtual void Vstart(Field & in, Field & src, Field & r, Field & mp, Field & mmp, Field & tmp);
-}
-
-template<class Field>
-class TwoLevelFlexiblePcgV11: public TwoLevelFlexiblePcg<Field> {
- public:
-  virtual void M(Field & in,Field & out,Field & tmp); 
-  virtual void M1(Field & in, Field & out,Field & tmp,Field & mp);
-  virtual void M2(Field & in, Field & out);
-  virtual RealD M3(Field & p, Field & mp,Field & mmp, Field & tmp);
-  virtual void Vstart(Field & in, Field & src, Field & r, Field & mp, Field & mmp, Field & tmp);
-}
-*/
 #endif

Grid/algorithms/iterative/AdefMrhs.h

@@ -0,0 +1,414 @@
+    /*************************************************************************************
+
+    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;
+
+  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)
+  {
+    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);
+
+    for(int rhs=0;rhs<nrhs;rhs++) {
+
+      this->SmoothTimer.Start();
+      this->_Smoother(in[rhs],Min[rhs]);
+      this->SmoothTimer.Stop();
+
+      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/ConjugateGradient.h

@@ -103,7 +103,7 @@ public:
     // Check if guess is really REALLY good :)
     if (cp <= rsq) {
       TrueResidual = std::sqrt(a/ssq);
-      std::cout << GridLogMessage << "ConjugateGradient guess is converged already " << std::endl;
+      std::cout << GridLogMessage << "ConjugateGradient guess is converged already : cp " << cp <<" rsq "<<rsq <<" ssq "<<ssq<< std::endl;
       IterationsToComplete = 0;	
       return;
     }
@@ -183,13 +183,13 @@ public:
 		  << "\tTrue residual " << true_residual
 		  << "\tTarget " << Tolerance << std::endl;
 
-        std::cout << GridLogMessage << "Time breakdown "<<std::endl;
 	std::cout << GridLogMessage << "\tElapsed    " << SolverTimer.Elapsed() <<std::endl;
-	std::cout << GridLogMessage << "\tMatrix     " << MatrixTimer.Elapsed() <<std::endl;
-	std::cout << GridLogMessage << "\tLinalg     " << LinalgTimer.Elapsed() <<std::endl;
-	std::cout << GridLogMessage << "\tInner      " << InnerTimer.Elapsed() <<std::endl;
-	std::cout << GridLogMessage << "\tAxpyNorm   " << AxpyNormTimer.Elapsed() <<std::endl;
-	std::cout << GridLogMessage << "\tLinearComb " << LinearCombTimer.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 << "\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;
 
@@ -207,7 +207,8 @@ public:
 
     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 "<< TrueResidual<< std::endl;
 
     if (ErrorOnNoConverge) assert(0);
     IterationsToComplete = k;

Grid/algorithms/iterative/ConjugateGradientMultiShift.h

@@ -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

@@ -166,16 +166,16 @@ public:
       rsqf[s] =rsq[s];
       std::cout<<GridLogMessage<<"ConjugateGradientMultiShiftMixedPrecCleanup: shift "<< s <<" target resid "<<rsq[s]<<std::endl;
       //      ps_d[s] = src_d;
-      precisionChangeFast(ps_f[s],src_d);
+      precisionChange(ps_f[s],src_d);
     }
     // r and p for primary
     p_d = src_d; //primary copy --- make this a reference to ps_d to save axpys
     r_d = p_d;
     
     //MdagM+m[0]
-    precisionChangeFast(p_f,p_d);
+    precisionChange(p_f,p_d);
     Linop_f.HermOpAndNorm(p_f,mmp_f,d,qq); // mmp = MdagM p        d=real(dot(p, mmp)),  qq=norm2(mmp)
-    precisionChangeFast(tmp_d,mmp_f);
+    precisionChange(tmp_d,mmp_f);
     Linop_d.HermOpAndNorm(p_d,mmp_d,d,qq); // mmp = MdagM p        d=real(dot(p, mmp)),  qq=norm2(mmp)
     tmp_d = tmp_d - mmp_d;
     std::cout << " Testing operators match "<<norm2(mmp_d)<<" f "<<norm2(mmp_f)<<" diff "<< norm2(tmp_d)<<std::endl;
@@ -204,7 +204,7 @@ public:
   
     for(int s=0;s<nshift;s++) {
       axpby(psi_d[s],0.,-bs[s]*alpha[s],src_d,src_d);
-      precisionChangeFast(psi_f[s],psi_d[s]);
+      precisionChange(psi_f[s],psi_d[s]);
     }
   
     ///////////////////////////////////////
@@ -225,7 +225,7 @@ public:
       AXPYTimer.Stop();
 
       PrecChangeTimer.Start();
-      precisionChangeFast(r_f, r_d);
+      precisionChange(r_f, r_d);
       PrecChangeTimer.Stop();
 
       AXPYTimer.Start();
@@ -243,13 +243,13 @@ public:
 
       cp=c;
       PrecChangeTimer.Start();
-      precisionChangeFast(p_f, p_d); //get back single prec search direction for linop
+      precisionChange(p_f, p_d); //get back single prec search direction for linop
       PrecChangeTimer.Stop();
       MatrixTimer.Start();  
       Linop_f.HermOp(p_f,mmp_f);
       MatrixTimer.Stop();  
       PrecChangeTimer.Start();
-      precisionChangeFast(mmp_d, mmp_f); // From Float to Double
+      precisionChange(mmp_d, mmp_f); // From Float to Double
       PrecChangeTimer.Stop();
 
       d=real(innerProduct(p_d,mmp_d));    
@@ -311,7 +311,7 @@ public:
 	SolverTimer.Stop();
 
 	for(int s=0;s<nshift;s++){
-	  precisionChangeFast(psi_d[s],psi_f[s]);
+	  precisionChange(psi_d[s],psi_f[s]);
 	}
 
 	

Grid/algorithms/iterative/ConjugateGradientMultiShiftMixedPrec.h

@@ -211,7 +211,7 @@ public:
     Linop_d.HermOpAndNorm(p_d,mmp_d,d,qq); // mmp = MdagM p        d=real(dot(p, mmp)),  qq=norm2(mmp)
     tmp_d = tmp_d - mmp_d;
     std::cout << " Testing operators match "<<norm2(mmp_d)<<" f "<<norm2(mmp_f)<<" diff "<< norm2(tmp_d)<<std::endl;
-    //    assert(norm2(tmp_d)< 1.0e-4);
+    assert(norm2(tmp_d)< 1.0);
 
     axpy(mmp_d,mass[0],p_d,mmp_d);
     RealD rn = norm2(p_d);

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;
   }
 };
@@ -419,14 +421,15 @@ until convergence
 	}
       }
 
-      if ( Nconv < Nstop )
+      if ( Nconv < Nstop ) {
 	std::cout << GridLogIRL << "Nconv ("<<Nconv<<") < Nstop ("<<Nstop<<")"<<std::endl;
-
+	std::cout << GridLogIRL << "returning Nstop vectors, the last "<< Nstop-Nconv << "of which might meet convergence criterion only approximately" <<std::endl;
+      }
       eval=eval2;
       
       //Keep only converged
-      eval.resize(Nconv);// Nstop?
-      evec.resize(Nconv,grid);// Nstop?
+      eval.resize(Nstop);// was Nconv
+      evec.resize(Nstop,grid);// was Nconv
       basisSortInPlace(evec,eval,reverse);
       
     }
@@ -456,7 +459,7 @@ until convergence
 	    std::vector<Field>& evec,
 	    Field& w,int Nm,int k)
   {
-    std::cout<<GridLogIRL << "Lanczos step " <<k<<std::endl;
+    std::cout<<GridLogDebug << "Lanczos step " <<k<<std::endl;
     const RealD tiny = 1.0e-20;
     assert( k< Nm );
 
@@ -464,7 +467,7 @@ until convergence
 
     Field& evec_k = evec[k];
 
-    _PolyOp(evec_k,w);    std::cout<<GridLogIRL << "PolyOp" <<std::endl;
+    _PolyOp(evec_k,w);    std::cout<<GridLogDebug << "PolyOp" <<std::endl;
 
     if(k>0) w -= lme[k-1] * evec[k-1];
 
@@ -479,18 +482,18 @@ until convergence
     lme[k] = beta;
 
     if ( (k>0) && ( (k % orth_period) == 0 )) {
-      std::cout<<GridLogIRL << "Orthogonalising " <<k<<std::endl;
+      std::cout<<GridLogDebug << "Orthogonalising " <<k<<std::endl;
       orthogonalize(w,evec,k); // orthonormalise
-      std::cout<<GridLogIRL << "Orthogonalised " <<k<<std::endl;
+      std::cout<<GridLogDebug << "Orthogonalised " <<k<<std::endl;
     }
 
     if(k < Nm-1) evec[k+1] = w;
 
-    std::cout<<GridLogIRL << "alpha[" << k << "] = " << zalph << " beta[" << k << "] = "<<beta<<std::endl;
+    std::cout<<GridLogIRL << "Lanczos step alpha[" << k << "] = " << zalph << " beta[" << k << "] = "<<beta<<std::endl;
     if ( beta < tiny ) 
       std::cout<<GridLogIRL << " beta is tiny "<<beta<<std::endl;
 
-    std::cout<<GridLogIRL << "Lanczos step complete " <<k<<std::endl;
+    std::cout<<GridLogDebug << "Lanczos step complete " <<k<<std::endl;
   }
 
   void diagonalize_Eigen(std::vector<RealD>& lmd, std::vector<RealD>& lme, 

Grid/algorithms/iterative/NormalEquations.h

@@ -33,7 +33,7 @@ NAMESPACE_BEGIN(Grid);
 ///////////////////////////////////////////////////////////////////////////////////////////////////////
 // Take a matrix and form an NE solver calling a Herm solver
 ///////////////////////////////////////////////////////////////////////////////////////////////////////
-template<class Field> class NormalEquations {
+template<class Field> class NormalEquations : public LinearFunction<Field>{
 private:
   SparseMatrixBase<Field> & _Matrix;
   OperatorFunction<Field> & _HermitianSolver;
@@ -60,7 +60,7 @@ public:
   }     
 };
 
-template<class Field> class HPDSolver {
+template<class Field> class HPDSolver : public LinearFunction<Field> {
 private:
   LinearOperatorBase<Field> & _Matrix;
   OperatorFunction<Field> & _HermitianSolver;
@@ -78,13 +78,13 @@ public:
   void operator() (const Field &in, Field &out){
  
     _Guess(in,out);
-    _HermitianSolver(_Matrix,in,out);  // Mdag M out = Mdag in
+    _HermitianSolver(_Matrix,in,out);  //M out = in
 
   }     
 };
 
 
-template<class Field> class MdagMSolver {
+template<class Field> class MdagMSolver : public LinearFunction<Field> {
 private:
   SparseMatrixBase<Field> & _Matrix;
   OperatorFunction<Field> & _HermitianSolver;

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_ = 50; 
+    const int _MAX_ITER_EST_ = 100; 
 
     for (int i=0;i<_MAX_ITER_EST_;i++) { 
       

Grid/algorithms/multigrid/Aggregates.h

@@ -0,0 +1,478 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./lib/algorithms/Aggregates.h
+
+    Copyright (C) 2015
+
+Author: Azusa Yamaguchi <ayamaguc@staffmail.ed.ac.uk>
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
+Author: paboyle <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
+
+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;
+
+  typedef Lattice< CComplex >   CoarseScalar; // used for inner products on fine field
+  typedef Lattice<Fobj >        FineField;
+
+  GridBase *CoarseGrid;
+  GridBase *FineGrid;
+  std::vector<Lattice<Fobj> > subspace;
+  int checkerboard;
+  int Checkerboard(void){return checkerboard;}
+  Aggregation(GridBase *_CoarseGrid,GridBase *_FineGrid,int _checkerboard) : 
+    CoarseGrid(_CoarseGrid),
+    FineGrid(_FineGrid),
+    subspace(nbasis,_FineGrid),
+    checkerboard(_checkerboard)
+  {
+  };
+  
+  
+  void Orthogonalise(void){
+    CoarseScalar InnerProd(CoarseGrid); 
+    //    std::cout << GridLogMessage <<" Block Gramm-Schmidt pass 1"<<std::endl;
+    blockOrthogonalise(InnerProd,subspace);
+  } 
+  void ProjectToSubspace(CoarseVector &CoarseVec,const FineField &FineVec){
+    blockProject(CoarseVec,FineVec,subspace);
+  }
+  void PromoteFromSubspace(const CoarseVector &CoarseVec,FineField &FineVec){
+    FineVec.Checkerboard() = subspace[0].Checkerboard();
+    blockPromote(CoarseVec,FineVec,subspace);
+  }
+
+  virtual void CreateSubspaceRandom(GridParallelRNG  &RNG) {
+    int nn=nbasis;
+    RealD scale;
+    FineField noise(FineGrid);
+    for(int b=0;b<nn;b++){
+      subspace[b] = Zero();
+      gaussian(RNG,noise);
+      scale = std::pow(norm2(noise),-0.5); 
+      noise=noise*scale;
+      subspace[b] = noise;
+    }
+  }
+  virtual void CreateSubspace(GridParallelRNG  &RNG,LinearOperatorBase<FineField> &hermop,int nn=nbasis)
+  {
+
+    RealD scale;
+
+    ConjugateGradient<FineField> CG(1.0e-2,100,false);
+    FineField noise(FineGrid);
+    FineField Mn(FineGrid);
+
+    for(int b=0;b<nn;b++){
+      
+      subspace[b] = Zero();
+      gaussian(RNG,noise);
+      scale = std::pow(norm2(noise),-0.5); 
+      noise=noise*scale;
+      
+      hermop.Op(noise,Mn); std::cout<<GridLogMessage << "noise   ["<<b<<"] <n|MdagM|n> "<<norm2(Mn)<<std::endl;
+
+      for(int i=0;i<1;i++){
+
+	CG(hermop,noise,subspace[b]);
+
+	noise = subspace[b];
+	scale = std::pow(norm2(noise),-0.5); 
+	noise=noise*scale;
+
+      }
+
+      hermop.Op(noise,Mn); std::cout<<GridLogMessage << "filtered["<<b<<"] <f|MdagM|f> "<<norm2(Mn)<<std::endl;
+      subspace[b]   = noise;
+
+    }
+  }
+
+  ////////////////////////////////////////////////////////////////////////////////////////////////
+  // World of possibilities here. But have tried quite a lot of experiments (250+ jobs run on Summit)
+  // and this is the best I found
+  ////////////////////////////////////////////////////////////////////////////////////////////////
+
+  virtual void CreateSubspaceChebyshev(GridParallelRNG  &RNG,LinearOperatorBase<FineField> &hermop,
+				       int nn,
+				       double hi,
+				       double lo,
+				       int orderfilter,
+				       int ordermin,
+				       int orderstep,
+				       double filterlo
+				       ) {
+
+    RealD scale;
+
+    FineField noise(FineGrid);
+    FineField Mn(FineGrid);
+    FineField tmp(FineGrid);
+
+    // New normalised noise
+    gaussian(RNG,noise);
+    scale = std::pow(norm2(noise),-0.5); 
+    noise=noise*scale;
+
+    std::cout << GridLogMessage<<" Chebyshev subspace pass-1 : ord "<<orderfilter<<" ["<<lo<<","<<hi<<"]"<<std::endl;
+    std::cout << GridLogMessage<<" Chebyshev subspace pass-2 : nbasis"<<nn<<" min "
+	      <<ordermin<<" step "<<orderstep
+	      <<" lo"<<filterlo<<std::endl;
+
+    // Initial matrix element
+    hermop.Op(noise,Mn); std::cout<<GridLogMessage << "noise <n|MdagM|n> "<<norm2(Mn)<<std::endl;
+
+    int b =0;
+    {
+      // Filter
+      Chebyshev<FineField> Cheb(lo,hi,orderfilter);
+      Cheb(hermop,noise,Mn);
+      // normalise
+      scale = std::pow(norm2(Mn),-0.5); 	Mn=Mn*scale;
+      subspace[b]   = Mn;
+      hermop.Op(Mn,tmp); 
+      std::cout<<GridLogMessage << "filt ["<<b<<"] <n|MdagM|n> "<<norm2(tmp)<<std::endl;
+      b++;
+    }
+
+    // Generate a full sequence of Chebyshevs
+    {
+      lo=filterlo;
+      noise=Mn;
+
+      FineField T0(FineGrid); T0 = noise;  
+      FineField T1(FineGrid); 
+      FineField T2(FineGrid);
+      FineField y(FineGrid);
+      
+      FineField *Tnm = &T0;
+      FineField *Tn  = &T1;
+      FineField *Tnp = &T2;
+
+      // Tn=T1 = (xscale M + mscale)in
+      RealD xscale = 2.0/(hi-lo);
+      RealD mscale = -(hi+lo)/(hi-lo);
+      hermop.HermOp(T0,y);
+      T1=y*xscale+noise*mscale;
+
+      for(int n=2;n<=ordermin+orderstep*(nn-2);n++){
+	
+	hermop.HermOp(*Tn,y);
+
+	autoView( y_v , y, AcceleratorWrite);
+	autoView( Tn_v , (*Tn), AcceleratorWrite);
+	autoView( Tnp_v , (*Tnp), AcceleratorWrite);
+	autoView( Tnm_v , (*Tnm), AcceleratorWrite);
+	const int Nsimd = CComplex::Nsimd();
+	accelerator_for(ss, FineGrid->oSites(), Nsimd, {
+	  coalescedWrite(y_v[ss],xscale*y_v(ss)+mscale*Tn_v(ss));
+	  coalescedWrite(Tnp_v[ss],2.0*y_v(ss)-Tnm_v(ss));
+        });
+
+	// Possible more fine grained control is needed than a linear sweep,
+	// but huge productivity gain if this is simple algorithm and not a tunable
+	int m =1;
+	if ( n>=ordermin ) m=n-ordermin;
+	if ( (m%orderstep)==0 ) { 
+	  Mn=*Tnp;
+	  scale = std::pow(norm2(Mn),-0.5);         Mn=Mn*scale;
+	  subspace[b] = Mn;
+	  hermop.Op(Mn,tmp); 
+	  std::cout<<GridLogMessage << n<<" filt ["<<b<<"] <n|MdagM|n> "<<norm2(tmp)<<std::endl;
+	  b++;
+	}
+
+	// Cycle pointers to avoid copies
+	FineField *swizzle = Tnm;
+	Tnm    =Tn;
+	Tn     =Tnp;
+	Tnp    =swizzle;
+	  
+      }
+    }
+    assert(b==nn);
+  }
+  virtual void CreateSubspaceChebyshev(GridParallelRNG  &RNG,LinearOperatorBase<FineField> &hermop,
+				       int nn,
+				       double hi,
+				       double lo,
+				       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<<" ["<<lo<<","<<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(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;
+      hermop.Op(Mn,tmp); 
+      std::cout<<GridLogMessage << "filt ["<<b<<"] <n|MdagM|n> "<<norm2(tmp)<<std::endl;
+    }
+
+  }
+  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

@@ -56,243 +56,6 @@ inline void blockMaskedInnerProduct(Lattice<CComplex> &CoarseInner,
   blockSum(CoarseInner,fine_inner_msk);
 }
 
-
-class Geometry {
-public:
-  int npoint;
-  int base;
-  std::vector<int> directions   ;
-  std::vector<int> displacements;
-  std::vector<int> points_dagger;
-
-  Geometry(int _d)  {
-    
-    base = (_d==5) ? 1:0;
-
-    // make coarse grid stencil for 4d , not 5d
-    if ( _d==5 ) _d=4;
-
-    npoint = 2*_d+1;
-    directions.resize(npoint);
-    displacements.resize(npoint);
-    points_dagger.resize(npoint);
-    for(int d=0;d<_d;d++){
-      directions[d   ] = d+base;
-      directions[d+_d] = d+base;
-      displacements[d  ] = +1;
-      displacements[d+_d]= -1;
-      points_dagger[d   ] = d+_d;
-      points_dagger[d+_d] = d;
-    }
-    directions   [2*_d]=0;
-    displacements[2*_d]=0;
-    points_dagger[2*_d]=2*_d;
-  }
-
-  int point(int dir, int disp) {
-    assert(disp == -1 || disp == 0 || disp == 1);
-    assert(base+0 <= dir && dir < base+4);
-
-    // directions faster index = new indexing
-    // 4d (base = 0):
-    // point 0  1  2  3  4  5  6  7  8
-    // dir   0  1  2  3  0  1  2  3  0
-    // disp +1 +1 +1 +1 -1 -1 -1 -1  0
-    // 5d (base = 1):
-    // point 0  1  2  3  4  5  6  7  8
-    // dir   1  2  3  4  1  2  3  4  0
-    // disp +1 +1 +1 +1 -1 -1 -1 -1  0
-
-    // displacements faster index = old indexing
-    // 4d (base = 0):
-    // point 0  1  2  3  4  5  6  7  8
-    // dir   0  0  1  1  2  2  3  3  0
-    // disp +1 -1 +1 -1 +1 -1 +1 -1  0
-    // 5d (base = 1):
-    // point 0  1  2  3  4  5  6  7  8
-    // dir   1  1  2  2  3  3  4  4  0
-    // disp +1 -1 +1 -1 +1 -1 +1 -1  0
-
-    if(dir == 0 and disp == 0)
-      return 8;
-    else // New indexing
-      return (1 - disp) / 2 * 4 + dir - base;
-    // else // Old indexing
-    //   return (4 * (dir - base) + 1 - disp) / 2;
-  }
-};
-  
-template<class Fobj,class CComplex,int nbasis>
-class Aggregation   {
-public:
-  typedef iVector<CComplex,nbasis >             siteVector;
-  typedef Lattice<siteVector>                 CoarseVector;
-  typedef Lattice<iMatrix<CComplex,nbasis > > CoarseMatrix;
-
-  typedef Lattice< CComplex >   CoarseScalar; // used for inner products on fine field
-  typedef Lattice<Fobj >        FineField;
-
-  GridBase *CoarseGrid;
-  GridBase *FineGrid;
-  std::vector<Lattice<Fobj> > subspace;
-  int checkerboard;
-  int Checkerboard(void){return checkerboard;}
-  Aggregation(GridBase *_CoarseGrid,GridBase *_FineGrid,int _checkerboard) : 
-    CoarseGrid(_CoarseGrid),
-    FineGrid(_FineGrid),
-    subspace(nbasis,_FineGrid),
-    checkerboard(_checkerboard)
-  {
-  };
-  
-  void Orthogonalise(void){
-    CoarseScalar InnerProd(CoarseGrid); 
-    std::cout << GridLogMessage <<" Block Gramm-Schmidt pass 1"<<std::endl;
-    blockOrthogonalise(InnerProd,subspace);
-  } 
-  void ProjectToSubspace(CoarseVector &CoarseVec,const FineField &FineVec){
-    blockProject(CoarseVec,FineVec,subspace);
-  }
-  void PromoteFromSubspace(const CoarseVector &CoarseVec,FineField &FineVec){
-    FineVec.Checkerboard() = subspace[0].Checkerboard();
-    blockPromote(CoarseVec,FineVec,subspace);
-  }
-
-  virtual void CreateSubspace(GridParallelRNG  &RNG,LinearOperatorBase<FineField> &hermop,int nn=nbasis) {
-
-    RealD scale;
-
-    ConjugateGradient<FineField> CG(1.0e-2,100,false);
-    FineField noise(FineGrid);
-    FineField Mn(FineGrid);
-
-    for(int b=0;b<nn;b++){
-      
-      subspace[b] = Zero();
-      gaussian(RNG,noise);
-      scale = std::pow(norm2(noise),-0.5); 
-      noise=noise*scale;
-      
-      hermop.Op(noise,Mn); std::cout<<GridLogMessage << "noise   ["<<b<<"] <n|MdagM|n> "<<norm2(Mn)<<std::endl;
-
-      for(int i=0;i<1;i++){
-
-	CG(hermop,noise,subspace[b]);
-
-	noise = subspace[b];
-	scale = std::pow(norm2(noise),-0.5); 
-	noise=noise*scale;
-
-      }
-
-      hermop.Op(noise,Mn); std::cout<<GridLogMessage << "filtered["<<b<<"] <f|MdagM|f> "<<norm2(Mn)<<std::endl;
-      subspace[b]   = noise;
-
-    }
-  }
-
-  ////////////////////////////////////////////////////////////////////////////////////////////////
-  // World of possibilities here. But have tried quite a lot of experiments (250+ jobs run on Summit)
-  // and this is the best I found
-  ////////////////////////////////////////////////////////////////////////////////////////////////
-
-  virtual void CreateSubspaceChebyshev(GridParallelRNG  &RNG,LinearOperatorBase<FineField> &hermop,
-				       int nn,
-				       double hi,
-				       double lo,
-				       int orderfilter,
-				       int ordermin,
-				       int orderstep,
-				       double filterlo
-				       ) {
-
-    RealD scale;
-
-    FineField noise(FineGrid);
-    FineField Mn(FineGrid);
-    FineField tmp(FineGrid);
-
-    // New normalised noise
-    gaussian(RNG,noise);
-    scale = std::pow(norm2(noise),-0.5); 
-    noise=noise*scale;
-
-    // Initial matrix element
-    hermop.Op(noise,Mn); std::cout<<GridLogMessage << "noise <n|MdagM|n> "<<norm2(Mn)<<std::endl;
-
-    int b =0;
-    {
-      // Filter
-      Chebyshev<FineField> Cheb(lo,hi,orderfilter);
-      Cheb(hermop,noise,Mn);
-      // normalise
-      scale = std::pow(norm2(Mn),-0.5); 	Mn=Mn*scale;
-      subspace[b]   = Mn;
-      hermop.Op(Mn,tmp); 
-      std::cout<<GridLogMessage << "filt ["<<b<<"] <n|MdagM|n> "<<norm2(tmp)<<std::endl;
-      b++;
-    }
-
-    // Generate a full sequence of Chebyshevs
-    {
-      lo=filterlo;
-      noise=Mn;
-
-      FineField T0(FineGrid); T0 = noise;  
-      FineField T1(FineGrid); 
-      FineField T2(FineGrid);
-      FineField y(FineGrid);
-      
-      FineField *Tnm = &T0;
-      FineField *Tn  = &T1;
-      FineField *Tnp = &T2;
-
-      // Tn=T1 = (xscale M + mscale)in
-      RealD xscale = 2.0/(hi-lo);
-      RealD mscale = -(hi+lo)/(hi-lo);
-      hermop.HermOp(T0,y);
-      T1=y*xscale+noise*mscale;
-
-      for(int n=2;n<=ordermin+orderstep*(nn-2);n++){
-	
-	hermop.HermOp(*Tn,y);
-
-	autoView( y_v , y, AcceleratorWrite);
-	autoView( Tn_v , (*Tn), AcceleratorWrite);
-	autoView( Tnp_v , (*Tnp), AcceleratorWrite);
-	autoView( Tnm_v , (*Tnm), AcceleratorWrite);
-	const int Nsimd = CComplex::Nsimd();
-	accelerator_for(ss, FineGrid->oSites(), Nsimd, {
-	  coalescedWrite(y_v[ss],xscale*y_v(ss)+mscale*Tn_v(ss));
-	  coalescedWrite(Tnp_v[ss],2.0*y_v(ss)-Tnm_v(ss));
-        });
-
-	// Possible more fine grained control is needed than a linear sweep,
-	// but huge productivity gain if this is simple algorithm and not a tunable
-	int m =1;
-	if ( n>=ordermin ) m=n-ordermin;
-	if ( (m%orderstep)==0 ) { 
-	  Mn=*Tnp;
-	  scale = std::pow(norm2(Mn),-0.5);         Mn=Mn*scale;
-	  subspace[b] = Mn;
-	  hermop.Op(Mn,tmp); 
-	  std::cout<<GridLogMessage << n<<" filt ["<<b<<"] <n|MdagM|n> "<<norm2(tmp)<<std::endl;
-	  b++;
-	}
-
-	// Cycle pointers to avoid copies
-	FineField *swizzle = Tnm;
-	Tnm    =Tn;
-	Tn     =Tnp;
-	Tnp    =swizzle;
-	  
-      }
-    }
-    assert(b==nn);
-  }
-
-};
-
 // Fine Object == (per site) type of fine field
 // nbasis      == number of deflation vectors
 template<class Fobj,class CComplex,int nbasis>

Grid/algorithms/multigrid/GeneralCoarsenedMatrix.h

@@ -0,0 +1,619 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./lib/algorithms/GeneralCoarsenedMatrix.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
+
+#include <Grid/qcd/QCD.h> // needed for Dagger(Yes|No), Inverse(Yes|No)
+
+#include <Grid/lattice/PaddedCell.h>
+#include <Grid/stencil/GeneralLocalStencil.h>
+
+NAMESPACE_BEGIN(Grid);
+
+// Fine Object == (per site) type of fine field
+// nbasis      == number of deflation vectors
+template<class Fobj,class CComplex,int nbasis>
+class GeneralCoarsenedMatrix : public SparseMatrixBase<Lattice<iVector<CComplex,nbasis > > >  {
+public:
+
+  typedef GeneralCoarsenedMatrix<Fobj,CComplex,nbasis> GeneralCoarseOp;
+  typedef iVector<CComplex,nbasis >           siteVector;
+  typedef iMatrix<CComplex,nbasis >           siteMatrix;
+  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
+  ////////////////////
+  int hermitian;
+  GridBase      *       _FineGrid; 
+  GridCartesian *       _CoarseGrid; 
+  NonLocalStencilGeometry &geom;
+  PaddedCell Cell;
+  GeneralLocalStencil Stencil;
+  
+  std::vector<CoarseMatrix> _A;
+  std::vector<CoarseMatrix> _Adag;
+  std::vector<CoarseVector> MultTemporaries;
+
+  ///////////////////////
+  // Interface
+  ///////////////////////
+  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;
+    std::cout << GridLogMessage <<"GeneralCoarsenedMatrix::ProjectNearestNeighbour "<< CopyMe._A[0].Grid()<<std::endl;
+    for(int p=0;p<geom.npoint;p++){
+      for(int pp=0;pp<CopyMe.geom.npoint;pp++){
+ 	// Search for the same relative shift
+	// 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]);
+	  nfound++;
+	}
+      }
+    }
+    assert(nfound==geom.npoint);
+    ExchangeCoarseLinks();
+  }
+  */
+  
+  GeneralCoarsenedMatrix(NonLocalStencilGeometry &_geom,GridBase *FineGrid, GridCartesian * CoarseGrid)
+    : geom(_geom),
+      _FineGrid(FineGrid),
+      _CoarseGrid(CoarseGrid),
+      hermitian(1),
+      Cell(_geom.Depth(),_CoarseGrid),
+      Stencil(Cell.grids.back(),geom.shifts)
+  {
+    {
+      int npoint = _geom.npoint;
+    }
+    _A.resize(geom.npoint,CoarseGrid);
+    //    _Adag.resize(geom.npoint,CoarseGrid);
+  }
+  void M (const CoarseVector &in, CoarseVector &out)
+  {
+    Mult(_A,in,out);
+  }
+  void Mdag (const CoarseVector &in, CoarseVector &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 ttemps=0; RealD tcopy=0;
+    RealD tmult2=0;
+
+    ttot=-usecond();
+    conformable(CoarseGrid(),in.Grid());
+    conformable(in.Grid(),out.Grid());
+    out.Checkerboard() = in.Checkerboard();
+    CoarseVector tin=in;
+
+    texch-=usecond();
+    CoarseVector pin = Cell.ExchangePeriodic(tin);
+    texch+=usecond();
+
+    CoarseVector pout(pin.Grid());
+
+    int npoint = geom.npoint;
+    typedef LatticeView<Cobj> Aview;
+    typedef LatticeView<Cvec> Vview;
+      
+    const int Nsimd = CComplex::Nsimd();
+    
+    int64_t osites=pin.Grid()->oSites();
+
+    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, AcceleratorWriteDiscard);
+      autoView( Stencil_v  , Stencil, AcceleratorRead);
+      tviews+=usecond();
+
+      // 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();
+      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(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 = 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 << 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;
+
+  };
+  
+  void PopulateAdag(void)
+  {
+    for(int64_t bidx=0;bidx<CoarseGrid()->gSites() ;bidx++){
+      Coordinate bcoor;
+      CoarseGrid()->GlobalIndexToGlobalCoor(bidx,bcoor);
+      
+      for(int p=0;p<geom.npoint;p++){
+	Coordinate scoor = bcoor;
+	for(int mu=0;mu<bcoor.size();mu++){
+	  int L = CoarseGrid()->GlobalDimensions()[mu];
+	  scoor[mu] = (bcoor[mu] - geom.shifts[p][mu] + L) % L; // Modulo arithmetic
+	}
+	// Flip to poke/peekLocalSite and not too bad
+	auto link = peekSite(_A[p],scoor);
+	int pp = geom.Reverse(p);
+	pokeSite(adj(link),_Adag[pp],bcoor);
+      }
+    }
+  }
+  /////////////////////////////////////////////////////////////
+  // 
+  // A) Only reduced flops option is to use a padded cell of depth 4
+  // and apply MpcDagMpc in the padded cell.
+  //
+  // Makes for ONE application of MpcDagMpc per vector instead of 30 or 80.
+  // With the effective cell size around (B+8)^4 perhaps 12^4/4^4 ratio
+  // Cost is 81x more, same as stencil size.
+  //
+  // But: can eliminate comms and do as local dirichlet.
+  //
+  // Local exchange gauge field once.
+  // Apply to all vectors, local only computation.
+  // Must exchange ghost subcells in reverse process of PaddedCell to take inner products
+  //
+  // B) Can reduce cost: pad by 1, apply Deo      (4^4+6^4+8^4+8^4 )/ (4x 4^4)
+  //                     pad by 2, apply Doe
+  //                     pad by 3, apply Deo
+  //                     then break out 8x directions; cost is ~10x MpcDagMpc per vector
+  //
+  // => almost factor of 10 in setup cost, excluding data rearrangement
+  //
+  // Intermediates -- ignore the corner terms, leave approximate and force Hermitian
+  // Intermediates -- pad by 2 and apply 1+8+24 = 33 times.
+  /////////////////////////////////////////////////////////////
+
+    //////////////////////////////////////////////////////////
+    // BFM HDCG style approach: Solve a system of equations to get Aij
+    //////////////////////////////////////////////////////////
+    /*
+     *     Here, k,l index which possible shift within the 3^Nd "ball" connected by MdagM.
+     *
+     *     conj(phases[block]) proj[k][ block*Nvec+j ] =  \sum_ball  e^{i q_k . delta} < phi_{block,j} | MdagM | phi_{(block+delta),i} > 
+     *                                                 =  \sum_ball e^{iqk.delta} A_ji
+     *
+     *     Must invert matrix M_k,l = e^[i q_k . delta_l]
+     *
+     *     Where q_k = delta_k . (2*M_PI/global_nb[mu])
+     */
+#if 0
+  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 tinv=0.0;
+
+    /////////////////////////////////////////////////////////////
+    // Orthogonalise the subblocks over the basis
+    /////////////////////////////////////////////////////////////
+    CoarseScalar InnerProd(CoarseGrid()); 
+    blockOrthogonalise(InnerProd,Subspace.subspace);
+
+    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
+    CoarseVector coarseInner(CoarseGrid());
+
+    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
+	/////////////////////////////////////////////////////
+	// Stick a phase on every block
+	/////////////////////////////////////////////////////
+	tphase-=usecond();
+	CoarseComplexField coor(CoarseGrid());
+	CoarseComplexField pha(CoarseGrid());	pha=Zero();
+	for(int mu=0;mu<Nd;mu++){
+	  LatticeCoordinate(coor,mu);
+	  RealD TwoPiL =  M_PI * 2.0/ clatt[mu];
+	  pha = pha + (TwoPiL * geom.shifts[p][mu]) * coor;
+	}
+	pha  =exp(pha*ci);
+	phaV=Zero();
+	blockZAXPY(phaV,pha,Subspace.subspace[i],phaV);
+	tphase+=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();
+
+	tproj-=usecond();
+	blockProject(coarseInner,MphaV,Subspace.subspace);
+	coarseInner = conjugate(pha) * coarseInner;
+
+	ComputeProj[p] = coarseInner;
+	tproj+=usecond();
+
+      }
+
+      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();
+    }
+
+    // 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 mat    "<<tmat <<" us"<<std::endl;
+    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.ExchangePeriodic(_A[p]);
+      //      _Adag[p]= Cell.ExchangePeriodic(_Adag[p]);
+    }
+  }
+  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/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

@@ -0,0 +1,238 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./lib/algorithms/GeneralCoarsenedMatrix.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);
+
+
+/////////////////////////////////////////////////////////////////
+// Geometry class in cartesian case
+/////////////////////////////////////////////////////////////////
+
+class Geometry {
+public:
+  int npoint;
+  int base;
+  std::vector<int> directions   ;
+  std::vector<int> displacements;
+  std::vector<int> points_dagger;
+
+  Geometry(int _d)  {
+    
+    base = (_d==5) ? 1:0;
+
+    // make coarse grid stencil for 4d , not 5d
+    if ( _d==5 ) _d=4;
+
+    npoint = 2*_d+1;
+    directions.resize(npoint);
+    displacements.resize(npoint);
+    points_dagger.resize(npoint);
+    for(int d=0;d<_d;d++){
+      directions[d   ] = d+base;
+      directions[d+_d] = d+base;
+      displacements[d  ] = +1;
+      displacements[d+_d]= -1;
+      points_dagger[d   ] = d+_d;
+      points_dagger[d+_d] = d;
+    }
+    directions   [2*_d]=0;
+    displacements[2*_d]=0;
+    points_dagger[2*_d]=2*_d;
+  }
+
+  int point(int dir, int disp) {
+    assert(disp == -1 || disp == 0 || disp == 1);
+    assert(base+0 <= dir && dir < base+4);
+
+    // directions faster index = new indexing
+    // 4d (base = 0):
+    // point 0  1  2  3  4  5  6  7  8
+    // dir   0  1  2  3  0  1  2  3  0
+    // disp +1 +1 +1 +1 -1 -1 -1 -1  0
+    // 5d (base = 1):
+    // point 0  1  2  3  4  5  6  7  8
+    // dir   1  2  3  4  1  2  3  4  0
+    // disp +1 +1 +1 +1 -1 -1 -1 -1  0
+
+    // displacements faster index = old indexing
+    // 4d (base = 0):
+    // point 0  1  2  3  4  5  6  7  8
+    // dir   0  0  1  1  2  2  3  3  0
+    // disp +1 -1 +1 -1 +1 -1 +1 -1  0
+    // 5d (base = 1):
+    // point 0  1  2  3  4  5  6  7  8
+    // dir   1  1  2  2  3  3  4  4  0
+    // disp +1 -1 +1 -1 +1 -1 +1 -1  0
+
+    if(dir == 0 and disp == 0)
+      return 8;
+    else // New indexing
+      return (1 - disp) / 2 * 4 + dir - base;
+    // else // Old indexing
+    //   return (4 * (dir - base) + 1 - disp) / 2;
+  }
+};
+
+/////////////////////////////////////////////////////////////////
+// Less local equivalent of Geometry class in cartesian case
+/////////////////////////////////////////////////////////////////
+class NonLocalStencilGeometry {
+public:
+  //  int depth;
+  int skip;
+  int hops;
+  int npoint;
+  std::vector<Coordinate> shifts;
+  Coordinate stencil_size;
+  Coordinate stencil_lo;
+  Coordinate stencil_hi;
+  GridCartesian *grid;
+  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
+  int DimSkip(void){return skip;};
+
+  virtual ~NonLocalStencilGeometry() {};
+
+  int  Reverse(int point)
+  {
+    int Nd = Grid()->Nd();
+    Coordinate shft = shifts[point];
+    Coordinate rev(Nd);
+    for(int mu=0;mu<Nd;mu++) rev[mu]= -shft[mu];
+    for(int p=0;p<npoint;p++){
+      if(rev==shifts[p]){
+	return p;
+      }
+    }
+    assert(0);
+    return -1;
+  }
+  void BuildShifts(void)
+  {
+    this->shifts.resize(0);
+    int Nd = this->grid->Nd();
+
+    int dd = this->DimSkip();
+    for(int s0=this->stencil_lo[dd+0];s0<=this->stencil_hi[dd+0];s0++){
+    for(int s1=this->stencil_lo[dd+1];s1<=this->stencil_hi[dd+1];s1++){
+    for(int s2=this->stencil_lo[dd+2];s2<=this->stencil_hi[dd+2];s2++){
+    for(int s3=this->stencil_lo[dd+3];s3<=this->stencil_hi[dd+3];s3++){
+      Coordinate sft(Nd,0);
+      sft[dd+0] = s0;
+      sft[dd+1] = s1;
+      sft[dd+2] = s2;
+      sft[dd+3] = s3;
+      int nhops = abs(s0)+abs(s1)+abs(s2)+abs(s3);
+      if(nhops<=this->hops) this->shifts.push_back(sft);
+    }}}}
+    this->npoint = this->shifts.size();
+    std::cout << GridLogMessage << "NonLocalStencilGeometry has "<< this->npoint << " terms in stencil "<<std::endl;
+  }
+  
+  NonLocalStencilGeometry(GridCartesian *_coarse_grid,int _hops,int _skip) : grid(_coarse_grid), hops(_hops), skip(_skip)
+  {
+    Coordinate latt = grid->GlobalDimensions();
+    stencil_size.resize(grid->Nd());
+    stencil_lo.resize(grid->Nd());
+    stencil_hi.resize(grid->Nd());
+    for(int d=0;d<grid->Nd();d++){
+     if ( latt[d] == 1 ) {
+      stencil_lo[d] = 0;
+      stencil_hi[d] = 0;
+      stencil_size[d]= 1;
+     } else if ( latt[d] == 2 ) {
+      stencil_lo[d] = -1;
+      stencil_hi[d] = 0;
+      stencil_size[d]= 2;
+     } else if ( latt[d] > 2 ) {
+       stencil_lo[d] = -1;
+       stencil_hi[d] =  1;
+       stencil_size[d]= 3;
+     }
+    }
+    this->BuildShifts();
+  };
+
+};
+
+// Need to worry about red-black now
+class NonLocalStencilGeometry4D : public NonLocalStencilGeometry {
+public:
+  virtual int DerivedDimSkip(void) { return 0;};
+  NonLocalStencilGeometry4D(GridCartesian *Coarse,int _hops) : NonLocalStencilGeometry(Coarse,_hops,0) { };
+  virtual ~NonLocalStencilGeometry4D() {};
+};
+class NonLocalStencilGeometry5D : public NonLocalStencilGeometry {
+public:
+  virtual int DerivedDimSkip(void) { return 1; }; 
+  NonLocalStencilGeometry5D(GridCartesian *Coarse,int _hops) : NonLocalStencilGeometry(Coarse,_hops,1)  { };
+  virtual ~NonLocalStencilGeometry5D() {};
+};
+/*
+ * Bunch of different options classes
+ */
+class NextToNextToNextToNearestStencilGeometry4D : public NonLocalStencilGeometry4D {
+public:
+  NextToNextToNextToNearestStencilGeometry4D(GridCartesian *Coarse) :  NonLocalStencilGeometry4D(Coarse,4)
+  {
+  };
+};
+class NextToNextToNextToNearestStencilGeometry5D : public  NonLocalStencilGeometry5D {
+public:
+  NextToNextToNextToNearestStencilGeometry5D(GridCartesian *Coarse) :  NonLocalStencilGeometry5D(Coarse,4)
+  {
+  };
+};
+class NextToNearestStencilGeometry4D : public  NonLocalStencilGeometry4D {
+public:
+  NextToNearestStencilGeometry4D(GridCartesian *Coarse) :  NonLocalStencilGeometry4D(Coarse,2)
+  {
+  };
+};
+class NextToNearestStencilGeometry5D : public  NonLocalStencilGeometry5D {
+public:
+  NextToNearestStencilGeometry5D(GridCartesian *Coarse) :  NonLocalStencilGeometry5D(Coarse,2)
+  {
+  };
+};
+class NearestStencilGeometry4D : public  NonLocalStencilGeometry4D {
+public:
+  NearestStencilGeometry4D(GridCartesian *Coarse) :  NonLocalStencilGeometry4D(Coarse,1)
+  {
+  };
+};
+class NearestStencilGeometry5D : public  NonLocalStencilGeometry5D {
+public:
+  NearestStencilGeometry5D(GridCartesian *Coarse) :  NonLocalStencilGeometry5D(Coarse,1)
+  {
+  };
+};
+
+NAMESPACE_END(Grid);

Grid/algorithms/multigrid/MultiGrid.h

@@ -0,0 +1,34 @@
+    /*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid
+
+    Source file: Grid/algorithms/multigrid/MultiGrid.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
+
+#include <Grid/algorithms/multigrid/Aggregates.h>
+#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

@@ -175,8 +175,56 @@ template<class T> using cshiftAllocator = std::allocator<T>;
 
 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> using commVector    = std::vector<T,devAllocator<T> >;
+template<class T> using deviceVector  = 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:
+  accelerator_inline T & operator[](size_t i) const { return this->data[i]; };
+  vecView(std::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(std::vector<T> &vec,ViewMode _mode)
+{
+  vecView<T> ret(vec,_mode); // does the open
+  return ret;                // must be closed
+}
+
+// Little autoscope assister
+template<class View> 
+class VectorViewCloser
+{
+  View v;  // Take a copy of view and call view close when I go out of scope automatically
+ public:
+  VectorViewCloser(View &_v) : v(_v) {};
+  ~VectorViewCloser() { auto ptr = v.cpu_ptr; v.ViewClose();  MemoryManager::NotifyDeletion(ptr);}
+};
+
+#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.h

@@ -209,9 +209,9 @@ 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 NotifyDeletion(void * CpuPtr);
   static void Print(void);
   static void PrintAll(void);
   static void PrintState( void* CpuPtr);

Grid/allocator/MemoryManagerCache.cc

@@ -8,7 +8,7 @@ NAMESPACE_BEGIN(Grid);
 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 dprintf(...) snprintf (print_buffer,MAXLINE, __VA_ARGS__ ); std::cout << GridLogDebug << print_buffer;
 //#define dprintf(...) 
 
 
@@ -111,7 +111,7 @@ 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 +141,7 @@ 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 +155,7 @@ 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 +169,7 @@ 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 AccPtr %lx -> CpuPtr %lx\n",(uint64_t)AccCache.AccPtr,(uint64_t)AccCache.CpuPtr); fflush(stdout);
   DeviceToHostBytes+=AccCache.bytes;
   DeviceToHostXfer++;
   AccCache.state=Consistent;
@@ -184,7 +184,7 @@ 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 AccPtr %lx <- CpuPtr %lx\n",(uint64_t)AccCache.AccPtr,(uint64_t)AccCache.CpuPtr); fflush(stdout);
   acceleratorCopyToDevice((void *)AccCache.CpuPtr,(void *)AccCache.AccPtr,AccCache.bytes);
   HostToDeviceBytes+=AccCache.bytes;
   HostToDeviceXfer++;
@@ -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)
@@ -519,7 +520,6 @@ void MemoryManager::Audit(std::string s)
   uint64_t LruBytes1=0;
   uint64_t LruBytes2=0;
   uint64_t LruCnt=0;
-  uint64_t LockedBytes=0;
   
   std::cout << " Memory Manager::Audit() from "<<s<<std::endl;
   for(auto it=LRU.begin();it!=LRU.end();it++){

Grid/cartesian/Cartesian_base.h

@@ -70,8 +70,8 @@ public:
   Coordinate _istride;    // Inner stride i.e. within simd lane
   int _osites;                  // _isites*_osites = product(dimensions).
   int _isites;
-  int _fsites;                  // _isites*_osites = product(dimensions).
-  int _gsites;
+  int64_t _fsites;                  // _isites*_osites = product(dimensions).
+  int64_t _gsites;
   Coordinate _slice_block;// subslice information
   Coordinate _slice_stride;
   Coordinate _slice_nblock;
@@ -183,7 +183,7 @@ public:
   inline int Nsimd(void)  const { return _isites; };// Synonymous with iSites
   inline int oSites(void) const { return _osites; };
   inline int lSites(void) const { return _isites*_osites; }; 
-  inline int gSites(void) const { return _isites*_osites*_Nprocessors; }; 
+  inline int64_t gSites(void) const { return (int64_t)_isites*(int64_t)_osites*(int64_t)_Nprocessors; }; 
   inline int Nd    (void) const { return _ndimension;};
 
   inline const Coordinate LocalStarts(void)             { return _lstart;    };
@@ -214,7 +214,7 @@ public:
   ////////////////////////////////////////////////////////////////
   // Global addressing
   ////////////////////////////////////////////////////////////////
-  void GlobalIndexToGlobalCoor(int gidx,Coordinate &gcoor){
+  void GlobalIndexToGlobalCoor(int64_t gidx,Coordinate &gcoor){
     assert(gidx< gSites());
     Lexicographic::CoorFromIndex(gcoor,gidx,_gdimensions);
   }
@@ -222,7 +222,7 @@ public:
     assert(lidx<lSites());
     Lexicographic::CoorFromIndex(lcoor,lidx,_ldimensions);
   }
-  void GlobalCoorToGlobalIndex(const Coordinate & gcoor,int & gidx){
+  void GlobalCoorToGlobalIndex(const Coordinate & gcoor,int64_t & gidx){
     gidx=0;
     int mult=1;
     for(int mu=0;mu<_ndimension;mu++) {

Grid/communicator/Communicator_base.h

@@ -138,6 +138,14 @@ public:
   ////////////////////////////////////////////////////////////
   // Face exchange, buffer swap in translational invariant way
   ////////////////////////////////////////////////////////////
+  void CommsComplete(std::vector<CommsRequest_t> &list);
+  void SendToRecvFromBegin(std::vector<CommsRequest_t> &list,
+			   void *xmit,
+			   int dest,
+			   void *recv,
+			   int from,
+			   int bytes,int dir);
+  
   void SendToRecvFrom(void *xmit,
 		      int xmit_to_rank,
 		      void *recv,

Grid/communicator/Communicator_mpi3.cc

@@ -306,6 +306,44 @@ 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<CommsRequest_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<CommsRequest_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,
@@ -348,6 +386,7 @@ double CartesianCommunicator::StencilSendToRecvFrom( void *xmit,
   return offbytes;
 }
 
+#undef NVLINK_GET // Define to use get instead of put DMA
 double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list,
 							 void *xmit,
 							 int dest,int dox,
@@ -380,9 +419,15 @@ double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsReques
       list.push_back(rrq);
       off_node_bytes+=rbytes;
     }
+#ifdef NVLINK_GET
+      void *shm = (void *) this->ShmBufferTranslate(from,xmit);
+      assert(shm!=NULL);
+      acceleratorCopyDeviceToDeviceAsynch(shm,recv,rbytes);
+#endif
   }
   
   if (dox) {
+    //  rcrc = crc32(rcrc,(unsigned char *)recv,bytes);
     if ( (gdest == MPI_UNDEFINED) || Stencil_force_mpi ) {
       tag= dir+_processor*32;
       ierr =MPI_Isend(xmit, xbytes, MPI_CHAR,dest,tag,communicator_halo[commdir],&xrq);
@@ -390,9 +435,12 @@ double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsReques
       list.push_back(xrq);
       off_node_bytes+=xbytes;
     } else {
+#ifndef NVLINK_GET
       void *shm = (void *) this->ShmBufferTranslate(dest,recv);
       assert(shm!=NULL);
       acceleratorCopyDeviceToDeviceAsynch(xmit,shm,xbytes);
+#endif
+      
     }
   }
 
@@ -402,6 +450,8 @@ void CartesianCommunicator::StencilSendToRecvFromComplete(std::vector<CommsReque
 {
   int nreq=list.size();
 
+  acceleratorCopySynchronise();
+
   if (nreq==0) return;
 
   std::vector<MPI_Status> status(nreq);

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);
@@ -128,7 +139,7 @@ double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsReques
 							 int recv_from_rank,int dor,
 							 int xbytes,int rbytes, int dir)
 {
-  return 2.0*bytes;
+  return xbytes+rbytes;
 }
 void CartesianCommunicator::StencilSendToRecvFromComplete(std::vector<CommsRequest_t> &waitall,int dir)
 {

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;
@@ -91,6 +114,59 @@ void *SharedMemory::ShmBufferSelf(void)
   //std::cerr << "ShmBufferSelf "<<ShmRank<<" "<<std::hex<< ShmCommBufs[ShmRank] <<std::dec<<std::endl;
   return ShmCommBufs[ShmRank];
 }
+static inline int divides(int a,int b)
+{
+  return ( b == ( (b/a)*a ) );
+}
+void GlobalSharedMemory::GetShmDims(const Coordinate &WorldDims,Coordinate &ShmDims)
+{
+  ////////////////////////////////////////////////////////////////
+  // Allow user to configure through environment variable
+  ////////////////////////////////////////////////////////////////
+  char* str = getenv(("GRID_SHM_DIMS_" + std::to_string(ShmDims.size())).c_str());
+  if ( str ) {
+    std::vector<int> IntShmDims;
+    GridCmdOptionIntVector(std::string(str),IntShmDims);
+    assert(IntShmDims.size() == WorldDims.size());
+    long ShmSize = 1;
+    for (int dim=0;dim<WorldDims.size();dim++) {
+      ShmSize *= (ShmDims[dim] = IntShmDims[dim]);
+      assert(divides(ShmDims[dim],WorldDims[dim]));
+    }
+    assert(ShmSize == WorldShmSize);
+    return;
+  }
+  
+  ////////////////////////////////////////////////////////////////
+  // Powers of 2,3,5 only in prime decomposition for now
+  ////////////////////////////////////////////////////////////////
+  int ndimension = WorldDims.size();
+  ShmDims=Coordinate(ndimension,1);
+
+  std::vector<int> primes({2,3,5});
+
+  int dim = 0;
+  int last_dim = ndimension - 1;
+  int AutoShmSize = 1;
+  while(AutoShmSize != WorldShmSize) {
+    int p;
+    for(p=0;p<primes.size();p++) {
+      int prime=primes[p];
+      if ( divides(prime,WorldDims[dim]/ShmDims[dim])
+        && divides(prime,WorldShmSize/AutoShmSize)  ) {
+  AutoShmSize*=prime;
+  ShmDims[dim]*=prime;
+  last_dim = dim;
+  break;
+      }
+    }
+    if (p == primes.size() && last_dim == dim) {
+      std::cerr << "GlobalSharedMemory::GetShmDims failed" << std::endl;
+      exit(EXIT_FAILURE);
+    }
+    dim=(dim+1) %ndimension;
+  }
+}
 
 NAMESPACE_END(Grid); 
 

Grid/communicator/SharedMemory.h

@@ -75,7 +75,9 @@ 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 +122,13 @@ 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 +160,10 @@ 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

@@ -27,9 +27,10 @@ Author: Christoph Lehner <christoph@lhnr.de>
 *************************************************************************************/
 /*  END LEGAL */
 
+#define Mheader "SharedMemoryMpi: "
+
 #include <Grid/GridCore.h>
 #include <pwd.h>
-#include <syscall.h>
 
 #ifdef GRID_CUDA
 #include <cuda_runtime_api.h>
@@ -38,12 +39,121 @@ 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
+#define SHM_SOCKETS
+#endif 
+#include <syscall.h>
+#endif
+
+#include <sys/socket.h>
+#include <sys/un.h>
+
+NAMESPACE_BEGIN(Grid); 
+
+#ifdef SHM_SOCKETS
+
+/*
+ * Barbaric extra intranode communication route in case we need sockets to pass FDs
+ * Forced by level_zero not being nicely designed
+ */
+static int sock;
+static const char *sock_path_fmt = "/tmp/GridUnixSocket.%d";
+static char sock_path[256];
+class UnixSockets {
+public:
+  static void Open(int rank)
+  {
+    int errnum;
+
+    sock = socket(AF_UNIX, SOCK_DGRAM, 0);  assert(sock>0);
+
+    struct sockaddr_un sa_un = { 0 };
+    sa_un.sun_family = AF_UNIX;
+    snprintf(sa_un.sun_path, sizeof(sa_un.sun_path),sock_path_fmt,rank);
+    unlink(sa_un.sun_path);
+    if (bind(sock, (struct sockaddr *)&sa_un, sizeof(sa_un))) {
+      perror("bind failure");
+      exit(EXIT_FAILURE);
+    }
+  }
+
+  static int RecvFileDescriptor(void)
+  {
+    int n;
+    int fd;
+    char buf[1];
+    struct iovec iov;
+    struct msghdr msg;
+    struct cmsghdr *cmsg;
+    char cms[CMSG_SPACE(sizeof(int))];
+
+    iov.iov_base = buf;
+    iov.iov_len = 1;
+
+    memset(&msg, 0, sizeof msg);
+    msg.msg_name = 0;
+    msg.msg_namelen = 0;
+    msg.msg_iov = &iov;
+    msg.msg_iovlen = 1;
+
+    msg.msg_control = (caddr_t)cms;
+    msg.msg_controllen = sizeof cms;
+
+    if((n=recvmsg(sock, &msg, 0)) < 0) {
+      perror("recvmsg failed");
+      return -1;
+    }
+    if(n == 0){
+      perror("recvmsg returned 0");
+      return -1;
+    }
+    cmsg = CMSG_FIRSTHDR(&msg);
+
+    memmove(&fd, CMSG_DATA(cmsg), sizeof(int));
+
+    return fd;
+  }
+
+  static void SendFileDescriptor(int fildes,int xmit_to_rank)
+  {
+    struct msghdr msg;
+    struct iovec iov;
+    struct cmsghdr *cmsg = NULL;
+    char ctrl[CMSG_SPACE(sizeof(int))];
+    char data = ' ';
+
+    memset(&msg, 0, sizeof(struct msghdr));
+    memset(ctrl, 0, CMSG_SPACE(sizeof(int)));
+    iov.iov_base = &data;
+    iov.iov_len = sizeof(data);
+    
+    sprintf(sock_path,sock_path_fmt,xmit_to_rank);
+    
+    struct sockaddr_un sa_un = { 0 };
+    sa_un.sun_family = AF_UNIX;
+    snprintf(sa_un.sun_path, sizeof(sa_un.sun_path),sock_path_fmt,xmit_to_rank);
+
+    msg.msg_name = (void *)&sa_un;
+    msg.msg_namelen = sizeof(sa_un);
+    msg.msg_iov = &iov;
+    msg.msg_iovlen = 1;
+    msg.msg_controllen =  CMSG_SPACE(sizeof(int));
+    msg.msg_control = ctrl;
+
+    cmsg = CMSG_FIRSTHDR(&msg);
+    cmsg->cmsg_level = SOL_SOCKET;
+    cmsg->cmsg_type = SCM_RIGHTS;
+    cmsg->cmsg_len = CMSG_LEN(sizeof(int));
+
+    *((int *) CMSG_DATA(cmsg)) = fildes;
+
+    sendmsg(sock, &msg, 0);
+  };
+};
 #endif
 
 
-NAMESPACE_BEGIN(Grid); 
-#define header "SharedMemoryMpi: "
 /*Construct from an MPI communicator*/
 void GlobalSharedMemory::Init(Grid_MPI_Comm comm)
 {
@@ -66,8 +176,8 @@ void GlobalSharedMemory::Init(Grid_MPI_Comm comm)
   MPI_Comm_size(WorldShmComm     ,&WorldShmSize);
 
   if ( WorldRank == 0) {
-    std::cout << header " World communicator of size " <<WorldSize << std::endl;  
-    std::cout << header " Node  communicator of size " <<WorldShmSize << std::endl;
+    std::cout << Mheader " World communicator of size " <<WorldSize << std::endl;  
+    std::cout << Mheader " Node  communicator of size " <<WorldShmSize << std::endl;
   }
   // WorldShmComm, WorldShmSize, WorldShmRank
 
@@ -170,59 +280,7 @@ void GlobalSharedMemory::OptimalCommunicator(const Coordinate &processors,Grid_M
   if(nscan==3 && HPEhypercube ) OptimalCommunicatorHypercube(processors,optimal_comm,SHM);
   else                          OptimalCommunicatorSharedMemory(processors,optimal_comm,SHM);
 }
-static inline int divides(int a,int b)
-{
-  return ( b == ( (b/a)*a ) );
-}
-void GlobalSharedMemory::GetShmDims(const Coordinate &WorldDims,Coordinate &ShmDims)
-{
-  ////////////////////////////////////////////////////////////////
-  // Allow user to configure through environment variable
-  ////////////////////////////////////////////////////////////////
-  char* str = getenv(("GRID_SHM_DIMS_" + std::to_string(ShmDims.size())).c_str());
-  if ( str ) {
-    std::vector<int> IntShmDims;
-    GridCmdOptionIntVector(std::string(str),IntShmDims);
-    assert(IntShmDims.size() == WorldDims.size());
-    long ShmSize = 1;
-    for (int dim=0;dim<WorldDims.size();dim++) {
-      ShmSize *= (ShmDims[dim] = IntShmDims[dim]);
-      assert(divides(ShmDims[dim],WorldDims[dim]));
-    }
-    assert(ShmSize == WorldShmSize);
-    return;
-  }
-  
-  ////////////////////////////////////////////////////////////////
-  // Powers of 2,3,5 only in prime decomposition for now
-  ////////////////////////////////////////////////////////////////
-  int ndimension = WorldDims.size();
-  ShmDims=Coordinate(ndimension,1);
 
-  std::vector<int> primes({2,3,5});
-
-  int dim = 0;
-  int last_dim = ndimension - 1;
-  int AutoShmSize = 1;
-  while(AutoShmSize != WorldShmSize) {
-    int p;
-    for(p=0;p<primes.size();p++) {
-      int prime=primes[p];
-      if ( divides(prime,WorldDims[dim]/ShmDims[dim])
-        && divides(prime,WorldShmSize/AutoShmSize)  ) {
-	AutoShmSize*=prime;
-	ShmDims[dim]*=prime;
-	last_dim = dim;
-	break;
-      }
-    }
-    if (p == primes.size() && last_dim == dim) {
-      std::cerr << "GlobalSharedMemory::GetShmDims failed" << std::endl;
-      exit(EXIT_FAILURE);
-    }
-    dim=(dim+1) %ndimension;
-  }
-}
 void GlobalSharedMemory::OptimalCommunicatorHypercube(const Coordinate &processors,Grid_MPI_Comm & optimal_comm,Coordinate &SHM)
 {
   ////////////////////////////////////////////////////////////////
@@ -396,7 +454,7 @@ void GlobalSharedMemory::OptimalCommunicatorSharedMemory(const Coordinate &proce
 #ifdef GRID_MPI3_SHMGET
 void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 {
-  std::cout << header "SharedMemoryAllocate "<< bytes<< " shmget implementation "<<std::endl;
+  std::cout << Mheader "SharedMemoryAllocate "<< bytes<< " shmget implementation "<<std::endl;
   assert(_ShmSetup==1);
   assert(_ShmAlloc==0);
 
@@ -456,46 +514,6 @@ 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 << header " 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)
 {
   void * ShmCommBuf ; 
@@ -518,13 +536,16 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
   ///////////////////////////////////////////////////////////////////////////////////////////////////////////
   // Each MPI rank should allocate our own buffer
   ///////////////////////////////////////////////////////////////////////////////////////////////////////////
+#ifndef ACCELERATOR_AWARE_MPI
+  HostCommBuf= malloc(bytes);
+#endif  
   ShmCommBuf = acceleratorAllocDevice(bytes);
   if (ShmCommBuf == (void *)NULL ) {
     std::cerr << " SharedMemoryMPI.cc acceleratorAllocDevice failed NULL pointer for " << bytes<<" bytes " << std::endl;
     exit(EXIT_FAILURE);  
   }
   if ( WorldRank == 0 ){
-    std::cout << WorldRank << header " SharedMemoryMPI.cc acceleratorAllocDevice "<< bytes 
+    std::cout << WorldRank << Mheader " SharedMemoryMPI.cc acceleratorAllocDevice "<< bytes 
 	      << "bytes at "<< std::hex<< ShmCommBuf << " - "<<(bytes-1+(uint64_t)ShmCommBuf) <<std::dec<<" for comms buffers " <<std::endl;
   }
   SharedMemoryZero(ShmCommBuf,bytes);
@@ -532,8 +553,13 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
   ///////////////////////////////////////////////////////////////////////////////////////////////////////////
   // Loop over ranks/gpu's on our node
   ///////////////////////////////////////////////////////////////////////////////////////////////////////////
+#ifdef SHM_SOCKETS
+  UnixSockets::Open(WorldShmRank);
+#endif
   for(int r=0;r<WorldShmSize;r++){
 
+    MPI_Barrier(WorldShmComm);
+
 #ifndef GRID_MPI3_SHM_NONE
     //////////////////////////////////////////////////
     // If it is me, pass around the IPC access key
@@ -541,24 +567,32 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
     void * thisBuf = ShmCommBuf;
     if(!Stencil_force_mpi) {
 #ifdef GRID_SYCL_LEVEL_ZERO_IPC
-    typedef struct { int fd; pid_t pid ; } clone_mem_t;
+    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::level_zero>(theGridAccelerator->get_device());
-    auto zeContext   = cl::sycl::get_native<cl::sycl::backend::level_zero>(theGridAccelerator->get_context());
+    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());
       
     ze_ipc_mem_handle_t ihandle;
     clone_mem_t handle;
-
+    
     if ( r==WorldShmRank ) { 
       auto err = zeMemGetIpcHandle(zeContext,ShmCommBuf,&ihandle);
       if ( err != ZE_RESULT_SUCCESS ) {
-	std::cout << "SharedMemoryMPI.cc zeMemGetIpcHandle failed for rank "<<r<<" "<<std::hex<<err<<std::dec<<std::endl;
+	std::cerr << "SharedMemoryMPI.cc zeMemGetIpcHandle failed for rank "<<r<<" "<<std::hex<<err<<std::dec<<std::endl;
 	exit(EXIT_FAILURE);
       } else {
 	std::cout << "SharedMemoryMPI.cc zeMemGetIpcHandle succeeded for rank "<<r<<" "<<std::hex<<err<<std::dec<<std::endl;
       }
       memcpy((void *)&handle.fd,(void *)&ihandle,sizeof(int));
       handle.pid = getpid();
+      memcpy((void *)&handle.ze,(void *)&ihandle,sizeof(ihandle));
+#ifdef SHM_SOCKETS
+      for(int rr=0;rr<WorldShmSize;rr++){
+	if(rr!=r){
+	  UnixSockets::SendFileDescriptor(handle.fd,rr);
+	}
+      }
+#endif
     }
 #endif
 #ifdef GRID_CUDA
@@ -586,6 +620,7 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
     // Share this IPC handle across the Shm Comm
     //////////////////////////////////////////////////
     { 
+      MPI_Barrier(WorldShmComm);
       int ierr=MPI_Bcast(&handle,
 			 sizeof(handle),
 			 MPI_BYTE,
@@ -601,6 +636,10 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 #ifdef GRID_SYCL_LEVEL_ZERO_IPC
     if ( r!=WorldShmRank ) {
       thisBuf = nullptr;
+      int myfd;
+#ifdef SHM_SOCKETS
+      myfd=UnixSockets::RecvFileDescriptor();
+#else
       std::cout<<"mapping seeking remote pid/fd "
 	       <<handle.pid<<"/"
 	       <<handle.fd<<std::endl;
@@ -608,16 +647,22 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
       int pidfd = syscall(SYS_pidfd_open,handle.pid,0);
       std::cout<<"Using IpcHandle pidfd "<<pidfd<<"\n";
       //      int myfd  = syscall(SYS_pidfd_getfd,pidfd,handle.fd,0);
-      int myfd  = syscall(438,pidfd,handle.fd,0);
-
-      std::cout<<"Using IpcHandle myfd "<<myfd<<"\n";
-      
+      myfd  = syscall(438,pidfd,handle.fd,0);
+      int err_t = errno;
+      if (myfd < 0) {
+        fprintf(stderr,"pidfd_getfd returned %d errno was %d\n", myfd,err_t); fflush(stderr);
+	perror("pidfd_getfd failed ");
+	assert(0);
+      }
+#endif
+      std::cout<<"Using IpcHandle mapped remote pid "<<handle.pid <<" FD "<<handle.fd <<" to myfd "<<myfd<<"\n";
+      memcpy((void *)&ihandle,(void *)&handle.ze,sizeof(ihandle));
       memcpy((void *)&ihandle,(void *)&myfd,sizeof(int));
 
       auto err = zeMemOpenIpcHandle(zeContext,zeDevice,ihandle,0,&thisBuf);
       if ( err != ZE_RESULT_SUCCESS ) {
-	std::cout << "SharedMemoryMPI.cc "<<zeContext<<" "<<zeDevice<<std::endl;
-	std::cout << "SharedMemoryMPI.cc zeMemOpenIpcHandle failed for rank "<<r<<" "<<std::hex<<err<<std::dec<<std::endl; 
+	std::cerr << "SharedMemoryMPI.cc "<<zeContext<<" "<<zeDevice<<std::endl;
+	std::cerr << "SharedMemoryMPI.cc zeMemOpenIpcHandle failed for rank "<<r<<" "<<std::hex<<err<<std::dec<<std::endl; 
 	exit(EXIT_FAILURE);
       } else {
 	std::cout << "SharedMemoryMPI.cc zeMemOpenIpcHandle succeeded for rank "<<r<<std::endl;
@@ -652,18 +697,18 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 #else
     WorldShmCommBufs[r] = ShmCommBuf;
 #endif
+    MPI_Barrier(WorldShmComm);
   }
 
   _ShmAllocBytes=bytes;
   _ShmAlloc=1;
 }
-#endif
 
 #else 
 #ifdef GRID_MPI3_SHMMMAP
 void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 {
-  std::cout << header "SharedMemoryAllocate "<< bytes<< " MMAP implementation "<< GRID_SHM_PATH <<std::endl;
+  std::cout << Mheader "SharedMemoryAllocate "<< bytes<< " MMAP implementation "<< GRID_SHM_PATH <<std::endl;
   assert(_ShmSetup==1);
   assert(_ShmAlloc==0);
   //////////////////////////////////////////////////////////////////////////////////////////////////////////
@@ -700,7 +745,7 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
     assert(((uint64_t)ptr&0x3F)==0);
     close(fd);
     WorldShmCommBufs[r] =ptr;
-    //    std::cout << header "Set WorldShmCommBufs["<<r<<"]="<<ptr<< "("<< bytes<< "bytes)"<<std::endl;
+    //    std::cout << Mheader "Set WorldShmCommBufs["<<r<<"]="<<ptr<< "("<< bytes<< "bytes)"<<std::endl;
   }
   _ShmAlloc=1;
   _ShmAllocBytes  = bytes;
@@ -710,7 +755,7 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 #ifdef GRID_MPI3_SHM_NONE
 void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 {
-  std::cout << header "SharedMemoryAllocate "<< bytes<< " MMAP anonymous implementation "<<std::endl;
+  std::cout << Mheader "SharedMemoryAllocate "<< bytes<< " MMAP anonymous implementation "<<std::endl;
   assert(_ShmSetup==1);
   assert(_ShmAlloc==0);
   //////////////////////////////////////////////////////////////////////////////////////////////////////////
@@ -757,7 +802,7 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 ////////////////////////////////////////////////////////////////////////////////////////////
 void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 { 
-  std::cout << header "SharedMemoryAllocate "<< bytes<< " SHMOPEN implementation "<<std::endl;
+  std::cout << Mheader "SharedMemoryAllocate "<< bytes<< " SHMOPEN implementation "<<std::endl;
   assert(_ShmSetup==1);
   assert(_ShmAlloc==0); 
   MPI_Barrier(WorldShmComm);
@@ -881,6 +926,12 @@ 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_common.h

@@ -29,8 +29,27 @@ 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 commVector<std::pair<int,int> > Cshift_table_device; 
 
+inline std::pair<int,int> *MapCshiftTable(void)
+{
+  // GPU version
+#ifdef ACCELERATOR_CSHIFT    
+  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];
+#else 
+  return &Cshift_table[0];
+#endif
+  // CPU version use identify map
+}
 ///////////////////////////////////////////////////////////////////
 // Gather for when there is no need to SIMD split 
 ///////////////////////////////////////////////////////////////////
@@ -74,8 +93,8 @@ 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();
+#ifdef ACCELERATOR_CSHIFT
     autoView(rhs_v , rhs, AcceleratorRead);
     accelerator_for(i,ent,vobj::Nsimd(),{
 	coalescedWrite(buffer_p[table[i].first],coalescedRead(rhs_v[table[i].second]));
@@ -225,7 +244,7 @@ template<class vobj> void Scatter_plane_simple (Lattice<vobj> &rhs,cshiftVector<
   
   {
     auto buffer_p = & buffer[0];
-    auto table = &Cshift_table[0];
+    auto table = MapCshiftTable();
 #ifdef ACCELERATOR_CSHIFT    
     autoView( rhs_v, rhs, AcceleratorWrite);
     accelerator_for(i,ent,vobj::Nsimd(),{
@@ -297,30 +316,6 @@ 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 +340,12 @@ 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++){
@@ -372,7 +359,7 @@ template<class vobj> void Copy_plane(Lattice<vobj>& lhs,const Lattice<vobj> &rhs
   }
 
   {
-    auto table = &Cshift_table[0];
+    auto table = MapCshiftTable();
 #ifdef ACCELERATOR_CSHIFT    
     autoView(rhs_v , rhs, AcceleratorRead);
     autoView(lhs_v , lhs, AcceleratorWrite);
@@ -409,19 +396,11 @@ 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++){
@@ -432,7 +411,7 @@ template<class vobj> void Copy_plane_permute(Lattice<vobj>& lhs,const Lattice<vo
   }
 
   {
-    auto table = &Cshift_table[0];
+    auto table = MapCshiftTable();
 #ifdef ACCELERATOR_CSHIFT    
     autoView( rhs_v, rhs, AcceleratorRead);
     autoView( lhs_v, lhs, AcceleratorWrite);

Grid/cshift/Cshift_mpi.h

@@ -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
@@ -63,6 +64,8 @@ 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();
+  //  std::cout << GridLogPerformance << "Cshift took "<< (t1-t0)/1e3 << " ms"<<std::endl;
   return ret;
 }
 
@@ -127,16 +130,20 @@ template<class vobj> void Cshift_comms(Lattice<vobj> &ret,const Lattice<vobj> &r
     
   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;
@@ -144,26 +151,39 @@ 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);
-
-      grid->Barrier();
+      
+      tcomms-=usecond();
+      //      grid->Barrier();
 
       grid->SendToRecvFrom((void *)&send_buf[0],
 			   xmit_to_rank,
 			   (void *)&recv_buf[0],
 			   recv_from_rank,
 			   bytes);
+      xbytes+=bytes;
+      //      grid->Barrier();
+      tcomms+=usecond();
 
-      grid->Barrier();
-
+      tscatter-=usecond();
       Scatter_plane_simple (ret,recv_buf,dimension,x,cbmask);
+      tscatter+=usecond();
     }
   }
+  /*
+  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 +210,12 @@ 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);
 
   ///////////////////////////////////////////////
@@ -227,7 +253,9 @@ 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++){
       
@@ -252,7 +280,8 @@ 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();
 
 	send_buf_extract_mpi = &send_buf_extract[nbr_lane][0];
 	recv_buf_extract_mpi = &recv_buf_extract[i][0];
@@ -262,7 +291,9 @@ template<class vobj> void  Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vo
 			     recv_from_rank,
 			     bytes);
 
-	grid->Barrier();
+	xbytes+=bytes;
+	//	grid->Barrier();
+	tcomms+=usecond();
 
 	rpointers[i] = &recv_buf_extract[i][0];
       } else { 
@@ -270,9 +301,17 @@ 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();
   }
-
+  /*
+  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 +331,11 @@ 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 +359,9 @@ 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 +370,9 @@ 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 +380,8 @@ 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 +389,24 @@ 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();
     }
   }
+  /*
+  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 +432,11 @@ 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 +479,10 @@ 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 +507,8 @@ 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,
@@ -449,17 +517,28 @@ 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();
 
+  }
+  /*
+  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); 

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; 
+commVector<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>
@@ -46,4 +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_ET.h

@@ -345,7 +345,9 @@ GridUnopClass(UnaryNot, Not(a));
 GridUnopClass(UnaryTrace, trace(a));
 GridUnopClass(UnaryTranspose, transpose(a));
 GridUnopClass(UnaryTa, Ta(a));
+GridUnopClass(UnarySpTa, SpTa(a));
 GridUnopClass(UnaryProjectOnGroup, ProjectOnGroup(a));
+GridUnopClass(UnaryProjectOnSpGroup, ProjectOnSpGroup(a));
 GridUnopClass(UnaryTimesI, timesI(a));
 GridUnopClass(UnaryTimesMinusI, timesMinusI(a));
 GridUnopClass(UnaryAbs, abs(a));
@@ -456,7 +458,9 @@ GRID_DEF_UNOP(operator!, UnaryNot);
 GRID_DEF_UNOP(trace, UnaryTrace);
 GRID_DEF_UNOP(transpose, UnaryTranspose);
 GRID_DEF_UNOP(Ta, UnaryTa);
+GRID_DEF_UNOP(SpTa, UnarySpTa);
 GRID_DEF_UNOP(ProjectOnGroup, UnaryProjectOnGroup);
+GRID_DEF_UNOP(ProjectOnSpGroup, UnaryProjectOnSpGroup);
 GRID_DEF_UNOP(timesI, UnaryTimesI);
 GRID_DEF_UNOP(timesMinusI, UnaryTimesMinusI);
 GRID_DEF_UNOP(abs, UnaryAbs);  // abs overloaded in cmath C++98; DON'T do the

Grid/lattice/Lattice_arith.h

@@ -270,5 +270,42 @@ RealD axpby_norm(Lattice<vobj> &ret,sobj a,sobj b,const Lattice<vobj> &x,const L
     return axpby_norm_fast(ret,a,b,x,y);
 }
 
+/// 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

@@ -234,9 +234,12 @@ public:
   }
 
   template<class sobj> inline Lattice<vobj> & operator = (const sobj & r){
-    auto me  = View(CpuWrite);
-    thread_for(ss,me.size(),{
-	me[ss]= r;
+    vobj vtmp;
+    vtmp = r;
+    auto me  = View(AcceleratorWrite);
+    accelerator_for(ss,me.size(),vobj::Nsimd(),{
+	auto stmp=coalescedRead(vtmp);
+	coalescedWrite(me[ss],stmp);
     });
     me.ViewClose();
     return *this;
@@ -360,7 +363,7 @@ public:
 
 template<class vobj> std::ostream& operator<< (std::ostream& stream, const Lattice<vobj> &o){
   typedef typename vobj::scalar_object sobj;
-  for(int g=0;g<o.Grid()->_gsites;g++){
+  for(int64_t g=0;g<o.Grid()->_gsites;g++){
 
     Coordinate gcoor;
     o.Grid()->GlobalIndexToGlobalCoor(g,gcoor);

Grid/lattice/Lattice_basis.h

@@ -62,7 +62,7 @@ void basisRotate(VField &basis,Matrix& Qt,int j0, int j1, int k0,int k1,int Nm)
     basis_v.push_back(basis[k].View(AcceleratorWrite));
   }
 
-#if ( (!defined(GRID_CUDA)) )
+#if ( !(defined(GRID_CUDA) || defined(GRID_HIP) || defined(GRID_SYCL)) )
   int max_threads = thread_max();
   Vector < vobj > Bt(Nm * max_threads);
   thread_region

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_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);
 
@@ -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)
 {
@@ -280,11 +302,29 @@ inline ComplexD rankInnerProduct(const Lattice<vobj> &left,const Lattice<vobj> &
   return nrm;
 }
 
+
 template<class vobj>
 inline ComplexD innerProduct(const Lattice<vobj> &left,const Lattice<vobj> &right) {
   GridBase *grid = left.Grid();
+
+#ifdef GRID_SYCL
+  uint64_t csum=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);
+  }
+  FlightRecorder::CsumLog(csum);
+#endif
   ComplexD nrm = rankInnerProduct(left,right);
+  RealD local = real(nrm);
+  FlightRecorder::NormLog(real(nrm)); 
   grid->GlobalSum(nrm);
+  FlightRecorder::ReductionLog(local,real(nrm)); 
   return nrm;
 }
 
@@ -448,19 +488,10 @@ 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];
-
-  // 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];
-      }
-    }
-  });
+  int ostride=grid->_ostride[orthogdim];
+  
+  //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);
@@ -504,6 +535,7 @@ sliceSum(const Lattice<vobj> &Data,int orthogdim)
   return result;
 }
 
+
 template<class vobj>
 static void sliceInnerProductVector( std::vector<ComplexD> & result, const Lattice<vobj> &lhs,const Lattice<vobj> &rhs,int orthogdim) 
 {

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;

Grid/lattice/Lattice_reduction_sycl.h

@@ -69,29 +69,30 @@ inline typename vobj::scalar_object sum_gpu_large(const vobj *lat, Integer osite
   return result;
 }
 
-NAMESPACE_END(Grid);
 
-/*
-template<class Double> Double svm_reduce(Double *vec,uint64_t L)
+template<class Word> Word svm_xor(Word *vec,uint64_t L)
 {
-  Double sumResult; zeroit(sumResult);
-  Double *d_sum =(Double *)cl::sycl::malloc_shared(sizeof(Double),*theGridAccelerator);
-  Double identity;  zeroit(identity);
+  Word xorResult; xorResult = 0;
+  Word *d_sum =(Word *)cl::sycl::malloc_shared(sizeof(Word),*theGridAccelerator);
+  Word identity;  identity=0;
   theGridAccelerator->submit([&](cl::sycl::handler &cgh) {
-     auto Reduction = cl::sycl::reduction(d_sum,identity,std::plus<>());
+     auto Reduction = cl::sycl::reduction(d_sum,identity,std::bit_xor<>());
      cgh.parallel_for(cl::sycl::range<1>{L},
 		      Reduction,
 		      [=] (cl::sycl::id<1> index, auto &sum) {
-	 sum +=vec[index];
+	 sum ^=vec[index];
      });
    });
   theGridAccelerator->wait();
-  Double ret = d_sum[0];
+  Word ret = d_sum[0];
   free(d_sum,*theGridAccelerator);
-  std::cout << " svm_reduce finished "<<L<<" sites sum = " << ret <<std::endl;
   return ret;
 }
 
+NAMESPACE_END(Grid);
+
+/*
+
 template <class vobj>
 inline typename vobj::scalar_objectD sumD_gpu_repack(const vobj *lat, Integer osites)
 {

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.
@@ -162,9 +163,9 @@ public:
     // tens of seconds per trajectory so this is clean in all reasonable cases,
     // and margin of safety is orders of magnitude.
     // We could hack Sitmo to skip in the higher order words of state if necessary
-      //
-      // Replace with 2^30 ; avoid problem on large volumes
-      //
+    //
+    // Replace with 2^30 ; avoid problem on large volumes
+    //
     /////////////////////////////////////////////////////////////////////////////////////
     //      uint64_t skip = site+1;  //   Old init Skipped then drew.  Checked compat with faster init
     const int shift = 30;
@@ -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
@@ -361,9 +365,14 @@ public:
     _bernoulli.resize(_vol,std::discrete_distribution<int32_t>{1,1});
     _uid.resize(_vol,std::uniform_int_distribution<uint32_t>() );
   }
-
-  template <class vobj,class distribution> inline void fill(Lattice<vobj> &l,std::vector<distribution> &dist){
-
+  template <class vobj,class distribution> inline void fill(Lattice<vobj> &l,std::vector<distribution> &dist)
+  {
+    if ( l.Grid()->_isCheckerBoarded ) {
+      Lattice<vobj> tmp(_grid);
+      fill(tmp,dist);
+      pickCheckerboard(l.Checkerboard(),l,tmp);
+      return;
+    }
     typedef typename vobj::scalar_object scalar_object;
     typedef typename vobj::scalar_type scalar_type;
     typedef typename vobj::vector_type vector_type;
@@ -407,7 +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());
@@ -424,10 +433,9 @@ public:
     // MT implementation does not implement fast discard even though
     // in principle this is possible
     ////////////////////////////////////////////////
-#if 1
     thread_for( lidx, _grid->lSites(), {
 
-	int gidx;
+	int64_t gidx;
 	int o_idx;
 	int i_idx;
 	int rank;
@@ -445,29 +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 +1,213 @@
+#pragma once
+#include <type_traits>
+#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, 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;
+  commVector<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);
+  
+
+}
+
+template<class vobj> inline void sliceSumReduction_cub_large(const vobj *Data, 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;
+  commVector<vector>buffer(osites);
+  vector *dat = (vector *)Data;
+  vector *buf = &buffer[0];
+  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];
+    });
+
+    sliceSumReduction_cub_small(buf,lvSum_small,rd,e1,e2,stride, ostride,Nsimd);
+      
+    for (int r = 0; r < rd; r++) {
+      lvSum_ptr[w+words*r]=lvSum_small[r];
+    }
+
+  }
+
+  
+}
+
+template<class vobj> inline void sliceSumReduction_cub(const Lattice<vobj> &Data, 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); //hipcub/cub cannot deal with large vobjs so we split into small/large case.
+    if constexpr (sizeof(vobj) <= 256) { 
+      sliceSumReduction_cub_small(&Data_v[0], lvSum, rd, e1, e2, stride, ostride, Nsimd);
+    }
+    else {
+      sliceSumReduction_cub_large(&Data_v[0], lvSum, rd, e1, e2, stride, ostride, Nsimd);
+    }
+}
+#endif
+
+
+#if defined(GRID_SYCL)
+template<class vobj> inline void sliceSumReduction_sycl(const Lattice<vobj> &Data, Vector <vobj> &lvSum, const int  &rd, const int &e1, const int &e2, const int &stride, const int &ostride, const int &Nsimd)
+{
+  typedef typename vobj::scalar_object sobj;
+  size_t subvol_size = e1*e2;
+
+  vobj *mysum = (vobj *) malloc_shared(sizeof(vobj),*theGridAccelerator);
+  vobj vobj_zero;
+  zeroit(vobj_zero);
+    
+  commVector<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_v[ss]));
+
+  });
+
+  for (int r = 0; r < rd; r++) {
+      mysum[0] = vobj_zero; //dirty hack: cannot pass vobj_zero as identity to sycl::reduction as its not device_copyable
+      theGridAccelerator->submit([&](cl::sycl::handler &cgh) {
+          auto Reduction = cl::sycl::reduction(mysum,std::plus<>());
+          cgh.parallel_for(cl::sycl::range<1>{subvol_size},
+          Reduction,
+          [=](cl::sycl::id<1> item, auto &sum) {
+              auto s = item[0];
+              sum += rb_p[r*subvol_size+s];
+          });
+      });
+      theGridAccelerator->wait();
+      lvSum[r] = mysum[0];
+  }
+  
+  free(mysum,*theGridAccelerator);
+}
+#endif
+
+template<class vobj> inline void sliceSumReduction_cpu(const Lattice<vobj> &Data, 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, 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)
+  
+  sliceSumReduction_cub(Data, lvSum, rd, e1, e2, stride, ostride, Nsimd);
+  
+  #elif defined(GRID_SYCL)
+  
+  sliceSumReduction_sycl(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_trace.h

@@ -66,6 +66,65 @@ inline auto TraceIndex(const Lattice<vobj> &lhs) -> Lattice<decltype(traceIndex<
   return ret;
 };
 
+template<int N, class Vec>
+Lattice<iScalar<iScalar<iScalar<Vec> > > > Determinant(const Lattice<iScalar<iScalar<iMatrix<Vec, N> > > > &Umu)
+{
+  GridBase *grid=Umu.Grid();
+  auto lvol = grid->lSites();
+  Lattice<iScalar<iScalar<iScalar<Vec> > > > ret(grid);
+  typedef typename Vec::scalar_type scalar;
+  autoView(Umu_v,Umu,CpuRead);
+  autoView(ret_v,ret,CpuWrite);
+  thread_for(site,lvol,{
+    Eigen::MatrixXcd EigenU = Eigen::MatrixXcd::Zero(N,N);
+    Coordinate lcoor;
+    grid->LocalIndexToLocalCoor(site, lcoor);
+    iScalar<iScalar<iMatrix<scalar, N> > > Us;
+    peekLocalSite(Us, Umu_v, lcoor);
+    for(int i=0;i<N;i++){
+      for(int j=0;j<N;j++){
+	scalar tmp= Us()()(i,j);
+	ComplexD ztmp(real(tmp),imag(tmp));
+	EigenU(i,j)=ztmp;
+      }}
+    ComplexD detD  = EigenU.determinant();
+    typename Vec::scalar_type det(detD.real(),detD.imag());
+    pokeLocalSite(det,ret_v,lcoor);
+  });
+  return ret;
+}
+
+template<int N>
+Lattice<iScalar<iScalar<iMatrix<vComplexD, N> > > > Inverse(const Lattice<iScalar<iScalar<iMatrix<vComplexD, N> > > > &Umu)
+{
+  GridBase *grid=Umu.Grid();
+  auto lvol = grid->lSites();
+  Lattice<iScalar<iScalar<iMatrix<vComplexD, N> > > > ret(grid);
+  
+  autoView(Umu_v,Umu,CpuRead);
+  autoView(ret_v,ret,CpuWrite);
+  thread_for(site,lvol,{
+    Eigen::MatrixXcd EigenU = Eigen::MatrixXcd::Zero(N,N);
+    Coordinate lcoor;
+    grid->LocalIndexToLocalCoor(site, lcoor);
+    iScalar<iScalar<iMatrix<ComplexD, N> > > Us;
+    iScalar<iScalar<iMatrix<ComplexD, N> > > Ui;
+    peekLocalSite(Us, Umu_v, lcoor);
+    for(int i=0;i<N;i++){
+      for(int j=0;j<N;j++){
+	EigenU(i,j) = Us()()(i,j);
+      }}
+    Eigen::MatrixXcd EigenUinv = EigenU.inverse();
+    for(int i=0;i<N;i++){
+      for(int j=0;j<N;j++){
+	Ui()()(i,j) = EigenUinv(i,j);
+      }}
+    pokeLocalSite(Ui,ret_v,lcoor);
+  });
+  return ret;
+}
+
+
 NAMESPACE_END(Grid);
 #endif
 

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();
 
@@ -463,37 +489,62 @@ inline void blockSum(Lattice<vobj> &coarseData,const Lattice<vobj> &fineData)
   autoView( coarseData_ , coarseData, AcceleratorWrite);
   autoView( fineData_   , fineData, AcceleratorRead);
 
-  auto coarseData_p = &coarseData_[0];
-  auto fineData_p = &fineData_[0];
+  auto coarseData_p  = &coarseData_[0];
+  auto fineData_p    = &fineData_[0];
   
   Coordinate fine_rdimensions = fine->_rdimensions;
   Coordinate coarse_rdimensions = coarse->_rdimensions;
 
   vobj zz = Zero();
-  
-  accelerator_for(sc,coarse->oSites(),1,{
 
+  // 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
 
-      vobj cd = zz;
-      
-      for(int sb=0;sb<blockVol;sb++){
-
+      auto cd = coalescedRead(zz);
+      for(int sb=e*subvol;sb<MIN((e+1)*subvol,blockVol);sb++){
 	int sf;
 	Coordinate coor_b(_ndimension);
 	Coordinate coor_f(_ndimension);
 	Lexicographic::CoorFromIndex(coor_b,sb,block_r);               // Block sub coordinate
 	for(int d=0;d<_ndimension;d++) coor_f[d]=coor_c[d]*block_r[d] + coor_b[d];
 	Lexicographic::IndexFromCoor(coor_f,sf,fine_rdimensions);
-
-	cd=cd+fineData_p[sf];
+	
+	cd=cd+coalescedRead(fineData_p[sf]);
       }
 
-      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();
@@ -698,43 +764,186 @@ void localCopyRegion(const Lattice<vobj> &From,Lattice<vobj> & To,Coordinate Fro
     assert(Fg->_processors[d]  == Tg->_processors[d]);
   }
 
-  // the above should guarantee that the operations are local
-  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;
+  ///////////////////////////////////////////////////////////
+  // 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;
 
-  autoView( t_v , To, AcceleratorWrite);
-  autoView( f_v , From, AcceleratorRead);
-  accelerator_for(idx,Fg->lSites(),1,{
-    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;
+  size_t nsite = 1;
+  for(int i=0;i<nd;i++) nsite *= RegionSize[i];
+
+  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, 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];
       }
-      Tcoor[d] = ToLowerLeft[d]+ Fcoor[d]-FromLowerLeft[d];
-    }
-    if (in_region) {
-      Integer idx_f = 0; for(int d=0;d<nd;d++) idx_f+=isf[d]*(Fcoor[d]/rdf[d]);
-      Integer idx_t = 0; for(int d=0;d<nd;d++) idx_t+=ist[d]*(Tcoor[d]/rdt[d]);
-      Integer odx_f = 0; for(int d=0;d<nd;d++) odx_f+=osf[d]*(Fcoor[d]%rdf[d]);
-      Integer odx_t = 0; for(int d=0;d<nd;d++) odx_t+=ost[d]*(Tcoor[d]%rdt[d]);
-      vector_type * fp = (vector_type *)&f_v[odx_f];
-      vector_type * tp = (vector_type *)&t_v[odx_t];
+      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];
+      
+      scalar_type stmp;
       for(int w=0;w<words;w++){
-	tp[w].putlane(fp[w].getlane(idx_f),idx_t);
+	stmp = getlane(from[w], from_lane);
+	putlane(to[w], stmp, to_lane);
       }
-    }
   });
 }
 
+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;
+	}
+      }
+      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 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)
@@ -824,7 +1033,7 @@ void ExtractSlice(Lattice<vobj> &lowDim,const Lattice<vobj> & higherDim,int slic
 
 }
 
-
+//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)
 {
@@ -841,65 +1050,22 @@ void InsertSliceLocal(const Lattice<vobj> &lowDim, Lattice<vobj> & higherDim,int
 
   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,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);
+      assert(lg->_processors[d]  == hg->_processors[d]);
+      assert(lg->_ldimensions[d] == hg->_ldimensions[d]);
     }
-  });
+  }
+  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);
 }
 
 
@@ -925,7 +1091,7 @@ void Replicate(const Lattice<vobj> &coarse,Lattice<vobj> & fine)
 
   Coordinate fcoor(nd);
   Coordinate ccoor(nd);
-  for(int g=0;g<fg->gSites();g++){
+  for(int64_t g=0;g<fg->gSites();g++){
 
     fg->GlobalIndexToGlobalCoor(g,fcoor);
     for(int d=0;d<nd;d++){
@@ -1611,5 +1777,35 @@ 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

@@ -0,0 +1,571 @@
+/*************************************************************************************
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./lib/lattice/PaddedCell.h
+
+    Copyright (C) 2019
+
+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
+
+#include<Grid/cshift/Cshift.h>
+
+NAMESPACE_BEGIN(Grid);
+
+//Allow the user to specify how the C-shift is performed, e.g. to respect the appropriate boundary conditions
+template<typename vobj>
+struct CshiftImplBase{
+  virtual Lattice<vobj> Cshift(const Lattice<vobj> &in, int dir, int shift) const = 0;
+  virtual ~CshiftImplBase(){}
+};
+template<typename vobj>
+struct CshiftImplDefault: public CshiftImplBase<vobj>{
+  Lattice<vobj> Cshift(const Lattice<vobj> &in, int dir, int shift) const override{ return Grid::Cshift(in,dir,shift); }
+};
+template<typename Gimpl>
+struct CshiftImplGauge: public CshiftImplBase<typename Gimpl::GaugeLinkField::vector_object>{
+  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 cshiftVector<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(cshiftVector<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;
+  int dims;
+  int depth;
+  std::vector<GridCartesian *> grids;
+
+  ~PaddedCell()
+  {
+    DeleteGrids();
+  }
+  PaddedCell(int _depth,GridCartesian *_grid)
+  {
+    unpadded_grid = _grid;
+    depth=_depth;
+    dims=_grid->Nd();
+    AllocateGrids();
+    Coordinate local     =unpadded_grid->LocalDimensions();
+    Coordinate procs     =unpadded_grid->ProcessorGrid();
+    for(int d=0;d<dims;d++){
+      if ( procs[d] > 1 ) assert(local[d]>=depth);
+    }
+  }
+  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)
+  {
+    Coordinate local     =unpadded_grid->LocalDimensions();
+    Coordinate simd      =unpadded_grid->_simd_layout;
+    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];
+	}
+
+	old_grid = new GridCartesian(global,simd,processors);
+      }
+      grids.push_back(old_grid);
+    }
+  };
+  template<class vobj>
+  inline Lattice<vobj> Extract(const Lattice<vobj> &in) const
+  {
+    Coordinate processors=unpadded_grid->_processors;
+
+    Lattice<vobj> out(unpadded_grid);
+
+    Coordinate local     =unpadded_grid->LocalDimensions();
+    // depends on the MPI spread      
+    Coordinate fll(dims,depth);
+    Coordinate tll(dims,0); // depends on the MPI spread
+    for(int d=0;d<dims;d++){
+      if( processors[d]==1 ) fll[d]=0;
+    }
+    localCopyRegion(in,out,fll,tll,local);
+    return out;
+  }
+  template<class vobj>
+  inline Lattice<vobj> Exchange(const Lattice<vobj> &in, const CshiftImplBase<vobj> &cshift = CshiftImplDefault<vobj>()) const
+  {
+    GridBase *old_grid = in.Grid();
+    int dims = old_grid->Nd();
+    Lattice<vobj> tmp = in;
+    for(int d=0;d<dims;d++){
+      tmp = Expand(d,tmp,cshift); // rvalue && assignment
+    }
+    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
+  {
+    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]);
+
+    double tins=0, tshift=0;
+
+    int islocal = 0 ;
+    if ( processors[dim] == 1 ) islocal = 1;
+
+    if ( islocal ) {
+
+      // replace with a copy and maybe grid swizzle
+      // return in;??
+      double t = usecond();
+      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++){
+	InsertSliceLocal(in,padded,x,depth+x,dim);
+      }
+      tins += usecond() - t;
+    
+      // High bit
+      t = usecond();
+      shifted = cshift.Cshift(in,dim,depth);
+      tshift += usecond() - t;
+
+      t=usecond();
+      for(int x=0;x<depth;x++){
+	InsertSliceLocal(shifted,padded,local[dim]-depth+x,depth+local[dim]+x,dim);
+      }
+      tins += usecond() - t;
+    
+      // Low bit
+      t = usecond();
+      shifted = cshift.Cshift(in,dim,-depth);
+      tshift += usecond() - t;
+    
+      t = usecond();
+      for(int x=0;x<depth;x++){
+	InsertSliceLocal(shifted,padded,x,x,dim);
+      }
+      tins += usecond() - t;
+
+    }
+    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 cshiftVector<vobj> send_buf; 
+    static cshiftVector<vobj> recv_buf;
+    send_buf.resize(buffer_size*2*depth);    
+    recv_buf.resize(buffer_size*2*depth);
+
+    std::vector<CommsRequest_t> fwd_req;   
+    std::vector<CommsRequest_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

@@ -179,11 +179,11 @@ extern GridLogger GridLogSolver;
 extern GridLogger GridLogError;
 extern GridLogger GridLogWarning;
 extern GridLogger GridLogMessage;
-extern GridLogger GridLogDebug  ;
+extern GridLogger GridLogDebug;
 extern GridLogger GridLogPerformance;
 extern GridLogger GridLogDslash;
-extern GridLogger GridLogIterative  ;
-extern GridLogger GridLogIntegrator  ;
+extern GridLogger GridLogIterative;
+extern GridLogger GridLogIntegrator;
 extern GridLogger GridLogHMC;
 extern GridLogger GridLogMemory;
 extern GridLogger GridLogTracing;
@@ -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/QCD.h

@@ -104,6 +104,7 @@ template<typename vtype> using iSpinMatrix                = iScalar<iMatrix<iSca
 template<typename vtype> using iColourMatrix              = iScalar<iScalar<iMatrix<vtype, Nc> > > ;
 template<typename vtype> using iSpinColourMatrix          = iScalar<iMatrix<iMatrix<vtype, Nc>, Ns> >;
 template<typename vtype> using iLorentzColourMatrix       = iVector<iScalar<iMatrix<vtype, Nc> >, Nd > ;
+template<typename vtype> using iLorentzComplex            = iVector<iScalar<iScalar<vtype> >, Nd > ;
 template<typename vtype> using iDoubleStoredColourMatrix  = iVector<iScalar<iMatrix<vtype, Nc> >, Nds > ;
 template<typename vtype> using iSpinVector                = iScalar<iVector<iScalar<vtype>, Ns> >;
 template<typename vtype> using iColourVector              = iScalar<iScalar<iVector<vtype, Nc> > >;
@@ -178,6 +179,15 @@ typedef iLorentzColourMatrix<vComplexF>  vLorentzColourMatrixF;
 typedef iLorentzColourMatrix<vComplexD>  vLorentzColourMatrixD;
 typedef iLorentzColourMatrix<vComplexD2> vLorentzColourMatrixD2;
 
+// LorentzComplex
+typedef iLorentzComplex<Complex  > LorentzComplex;
+typedef iLorentzComplex<ComplexF > LorentzComplexF;
+typedef iLorentzComplex<ComplexD > LorentzComplexD;
+
+typedef iLorentzComplex<vComplex > vLorentzComplex;
+typedef iLorentzComplex<vComplexF> vLorentzComplexF;
+typedef iLorentzComplex<vComplexD> vLorentzComplexD;
+
 // DoubleStored gauge field
 typedef iDoubleStoredColourMatrix<Complex  > DoubleStoredColourMatrix;
 typedef iDoubleStoredColourMatrix<ComplexF > DoubleStoredColourMatrixF;
@@ -307,6 +317,10 @@ typedef Lattice<vLorentzColourMatrixF>  LatticeLorentzColourMatrixF;
 typedef Lattice<vLorentzColourMatrixD>  LatticeLorentzColourMatrixD;
 typedef Lattice<vLorentzColourMatrixD2> LatticeLorentzColourMatrixD2;
 
+typedef Lattice<vLorentzComplex>  LatticeLorentzComplex;
+typedef Lattice<vLorentzComplexF> LatticeLorentzComplexF;
+typedef Lattice<vLorentzComplexD> LatticeLorentzComplexD;
+
 // DoubleStored gauge field
 typedef Lattice<vDoubleStoredColourMatrix>   LatticeDoubleStoredColourMatrix;
 typedef Lattice<vDoubleStoredColourMatrixF>  LatticeDoubleStoredColourMatrixF;

Grid/qcd/action/ActionBase.h

@@ -34,10 +34,24 @@ directory
 
 NAMESPACE_BEGIN(Grid);
 
+///////////////////////////////////
+// Smart configuration base class
+///////////////////////////////////
+template< class Field >
+class ConfigurationBase
+{
+public:
+  ConfigurationBase() {}
+  virtual ~ConfigurationBase() {}
+  virtual void set_Field(Field& U) =0;
+  virtual void smeared_force(Field&) = 0;
+  virtual Field& get_SmearedU() =0;
+  virtual Field &get_U(bool smeared = false) = 0;
+};
+
 template <class GaugeField >
 class Action 
 {
-
 public:
   bool is_smeared = false;
   RealD deriv_norm_sum;
@@ -77,16 +91,60 @@ public:
   void refresh_timer_stop(void)  { refresh_us+=usecond(); }
   void S_timer_start(void)       { S_us-=usecond(); }
   void S_timer_stop(void)        { S_us+=usecond(); }
+  /////////////////////////////
   // Heatbath?
+  /////////////////////////////
   virtual void refresh(const GaugeField& U, GridSerialRNG &sRNG, GridParallelRNG& pRNG) = 0; // refresh pseudofermions
   virtual RealD S(const GaugeField& U) = 0;                             // evaluate the action
   virtual RealD Sinitial(const GaugeField& U) { return this->S(U); } ;  // if the refresh computes the action, can cache it. Alternately refreshAndAction() ?
   virtual void deriv(const GaugeField& U, GaugeField& dSdU) = 0;        // evaluate the action derivative
+
+  /////////////////////////////////////////////////////////////
+  // virtual smeared interface through configuration container
+  /////////////////////////////////////////////////////////////
+  virtual void refresh(ConfigurationBase<GaugeField> & U, GridSerialRNG &sRNG, GridParallelRNG& pRNG)
+  {
+    refresh(U.get_U(is_smeared),sRNG,pRNG);
+  }
+  virtual RealD S(ConfigurationBase<GaugeField>& U)
+  {
+    return S(U.get_U(is_smeared));
+  }
+  virtual RealD Sinitial(ConfigurationBase<GaugeField>& U) 
+  {
+    return Sinitial(U.get_U(is_smeared));
+  }
+  virtual void deriv(ConfigurationBase<GaugeField>& U, GaugeField& dSdU)
+  {
+    deriv(U.get_U(is_smeared),dSdU); 
+    if ( is_smeared ) {
+      U.smeared_force(dSdU);
+    }
+  }
+  ///////////////////////////////
+  // Logging
+  ///////////////////////////////
   virtual std::string action_name()    = 0;                             // return the action name
   virtual std::string LogParameters()  = 0;                             // prints action parameters
   virtual ~Action(){}
 };
 
+template <class GaugeField >
+class EmptyAction : public Action <GaugeField>
+{
+  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/ActionCore.h

@@ -30,6 +30,8 @@ directory
 #ifndef QCD_ACTION_CORE
 #define QCD_ACTION_CORE
 
+#include <Grid/qcd/action/gauge/GaugeImplementations.h>
+
 #include <Grid/qcd/action/ActionBase.h>
 NAMESPACE_CHECK(ActionBase);
 #include <Grid/qcd/action/ActionSet.h>

Grid/qcd/action/fermion/Fermion.h

@@ -126,6 +126,16 @@ typedef WilsonFermion<WilsonTwoIndexSymmetricImplD> WilsonTwoIndexSymmetricFermi
 typedef WilsonFermion<WilsonTwoIndexAntiSymmetricImplF> WilsonTwoIndexAntiSymmetricFermionF;
 typedef WilsonFermion<WilsonTwoIndexAntiSymmetricImplD> WilsonTwoIndexAntiSymmetricFermionD;
 
+// Sp(2n)
+typedef WilsonFermion<SpWilsonImplF> SpWilsonFermionF;
+typedef WilsonFermion<SpWilsonImplD> SpWilsonFermionD;
+
+typedef WilsonFermion<SpWilsonTwoIndexAntiSymmetricImplF> SpWilsonTwoIndexAntiSymmetricFermionF;
+typedef WilsonFermion<SpWilsonTwoIndexAntiSymmetricImplD> SpWilsonTwoIndexAntiSymmetricFermionD;
+
+typedef WilsonFermion<SpWilsonTwoIndexSymmetricImplF> SpWilsonTwoIndexSymmetricFermionF;
+typedef WilsonFermion<SpWilsonTwoIndexSymmetricImplD> SpWilsonTwoIndexSymmetricFermionD;
+
 // Twisted mass fermion
 typedef WilsonTMFermion<WilsonImplD2> WilsonTMFermionD2;
 typedef WilsonTMFermion<WilsonImplF> WilsonTMFermionF;

Grid/qcd/action/fermion/WilsonCompressor.h

@@ -507,6 +507,7 @@ public:
     }
     this->face_table_computed=1;
     assert(this->u_comm_offset==this->_unified_buffer_size);
+    accelerator_barrier();
   }
 
 };

Grid/qcd/action/fermion/WilsonImpl.h

@@ -261,6 +261,22 @@ typedef WilsonImpl<vComplex,  TwoIndexAntiSymmetricRepresentation, CoeffReal > W
 typedef WilsonImpl<vComplexF, TwoIndexAntiSymmetricRepresentation, CoeffReal > WilsonTwoIndexAntiSymmetricImplF;  // Float
 typedef WilsonImpl<vComplexD, TwoIndexAntiSymmetricRepresentation, CoeffReal > WilsonTwoIndexAntiSymmetricImplD;  // Double
 
+//sp 2n
+
+typedef WilsonImpl<vComplex,  SpFundamentalRepresentation, CoeffReal > SpWilsonImplR;  // Real.. whichever prec
+typedef WilsonImpl<vComplexF, SpFundamentalRepresentation, CoeffReal > SpWilsonImplF;  // Float
+typedef WilsonImpl<vComplexD, SpFundamentalRepresentation, CoeffReal > SpWilsonImplD;  // Double
+
+typedef WilsonImpl<vComplex,  SpTwoIndexAntiSymmetricRepresentation, CoeffReal > SpWilsonTwoIndexAntiSymmetricImplR;  // Real.. whichever prec
+typedef WilsonImpl<vComplexF, SpTwoIndexAntiSymmetricRepresentation, CoeffReal > SpWilsonTwoIndexAntiSymmetricImplF;  // Float
+typedef WilsonImpl<vComplexD, SpTwoIndexAntiSymmetricRepresentation, CoeffReal > SpWilsonTwoIndexAntiSymmetricImplD;  // Double
+
+typedef WilsonImpl<vComplex,  SpTwoIndexSymmetricRepresentation, CoeffReal > SpWilsonTwoIndexSymmetricImplR;  // Real.. whichever prec
+typedef WilsonImpl<vComplexF, SpTwoIndexSymmetricRepresentation, CoeffReal > SpWilsonTwoIndexSymmetricImplF;  // Float
+typedef WilsonImpl<vComplexD, SpTwoIndexSymmetricRepresentation, CoeffReal > SpWilsonTwoIndexSymmetricImplD;  // Double
+
+typedef WilsonImpl<vComplex,  SpTwoIndexSymmetricRepresentation, CoeffReal > SpWilsonAdjImplR;  // Real.. whichever prec    // adj = 2indx symmetric for Sp(2N)
+typedef WilsonImpl<vComplexF, SpTwoIndexSymmetricRepresentation, CoeffReal > SpWilsonAdjImplF;  // Float     // adj = 2indx symmetric for Sp(2N)
+typedef WilsonImpl<vComplexD, SpTwoIndexSymmetricRepresentation, CoeffReal > SpWilsonAdjImplD;  // Double    // adj = 2indx symmetric for Sp(2N)
 
 NAMESPACE_END(Grid);
-

Grid/qcd/action/fermion/WilsonTMFermion.h

@@ -63,7 +63,9 @@ public:
   virtual void MooeeDag(const FermionField &in, FermionField &out) ;
   virtual void MooeeInv(const FermionField &in, FermionField &out) ;
   virtual void MooeeInvDag(const FermionField &in, FermionField &out) ;
-
+  virtual void M(const FermionField &in, FermionField &out) ;
+  virtual void Mdag(const FermionField &in, FermionField &out) ;
+  
 private:
   RealD mu; // TwistedMass parameter
 

Grid/qcd/action/fermion/implementation/StaggeredKernelsImplementation.h

@@ -280,20 +280,16 @@ void StaggeredKernels<Impl>::DhopImproved(StencilImpl &st, LebesgueOrder &lo,
 
   if( interior && exterior ) { 
     if (Opt == OptGeneric    ) { KERNEL_CALL(DhopSiteGeneric,1); return;}
-#ifndef GRID_CUDA
     if (Opt == OptHandUnroll ) { KERNEL_CALL(DhopSiteHand,1);    return;}
+#ifndef GRID_CUDA
     if (Opt == OptInlineAsm  ) {  ASM_CALL(DhopSiteAsm);     return;}
 #endif
   } else if( interior ) {
     if (Opt == OptGeneric    ) { KERNEL_CALL(DhopSiteGenericInt,1); return;}
-#ifndef GRID_CUDA
     if (Opt == OptHandUnroll ) { KERNEL_CALL(DhopSiteHandInt,1);    return;}
-#endif
   } else if( exterior ) { 
     if (Opt == OptGeneric    ) { KERNEL_CALL(DhopSiteGenericExt,1); return;}
-#ifndef GRID_CUDA
     if (Opt == OptHandUnroll ) { KERNEL_CALL(DhopSiteHandExt,1);    return;}
-#endif
   }
   assert(0 && " Kernel optimisation case not covered ");
 }
@@ -322,19 +318,13 @@ void StaggeredKernels<Impl>::DhopNaive(StencilImpl &st, LebesgueOrder &lo,
   
   if( interior && exterior ) { 
     if (Opt == OptGeneric    ) { KERNEL_CALL(DhopSiteGeneric,0); return;}
-#ifndef GRID_CUDA
     if (Opt == OptHandUnroll ) { KERNEL_CALL(DhopSiteHand,0);    return;}
-#endif
   } else if( interior ) {
     if (Opt == OptGeneric    ) { KERNEL_CALL(DhopSiteGenericInt,0); return;}
-#ifndef GRID_CUDA
     if (Opt == OptHandUnroll ) { KERNEL_CALL(DhopSiteHandInt,0);    return;}
-#endif
   } else if( exterior ) { 
     if (Opt == OptGeneric    ) { KERNEL_CALL(DhopSiteGenericExt,0); return;}
-#ifndef GRID_CUDA
     if (Opt == OptHandUnroll ) { KERNEL_CALL(DhopSiteHandExt,0);    return;}
-#endif
   }
 }
 

Grid/qcd/action/fermion/implementation/WilsonFermion5DImplementation.h

@@ -332,8 +332,7 @@ void WilsonFermion5D<Impl>::DhopInternalOverlappedComms(StencilImpl & st, Lebesg
   /////////////////////////////
   {
     GRID_TRACE("Gather");
-    st.HaloExchangeOptGather(in,compressor);
-    accelerator_barrier();
+    st.HaloExchangeOptGather(in,compressor); // Put the barrier in the routine
   }
   
   std::vector<std::vector<CommsRequest_t> > requests;

Grid/qcd/action/fermion/implementation/WilsonKernelsImplementation.h

@@ -423,14 +423,14 @@ void WilsonKernels<Impl>::DhopDirKernel( StencilImpl &st, DoubledGaugeField &U,S
 #define KERNEL_CALL(A) KERNEL_CALLNB(A); accelerator_barrier();
 
 #define KERNEL_CALL_EXT(A)						\
-  const uint64_t    NN = Nsite*Ls;					\
   const uint64_t    sz = st.surface_list.size();			\
   auto ptr = &st.surface_list[0];					\
   accelerator_forNB( ss, sz, Simd::Nsimd(), {				\
       int sF = ptr[ss];							\
-      int sU = ss/Ls;							\
+      int sU = sF/Ls;							\
       WilsonKernels<Impl>::A(st_v,U_v,buf,sF,sU,in_v,out_v);		\
-    });									
+    });									\
+  accelerator_barrier();
 
 #define ASM_CALL(A)							\
   thread_for( sss, Nsite, {						\
@@ -462,6 +462,7 @@ void WilsonKernels<Impl>::DhopKernel(int Opt,StencilImpl &st,  DoubledGaugeField
     autoView(st_v , st,AcceleratorRead);
 
    if( interior && exterior ) {
+     acceleratorFenceComputeStream();
      if (Opt == WilsonKernelsStatic::OptGeneric    ) { KERNEL_CALL(GenericDhopSite); return;}
      if (Opt == WilsonKernelsStatic::OptHandUnroll ) { KERNEL_CALL(HandDhopSite);    return;}
 #ifndef GRID_CUDA
@@ -474,9 +475,10 @@ void WilsonKernels<Impl>::DhopKernel(int Opt,StencilImpl &st,  DoubledGaugeField
      if (Opt == WilsonKernelsStatic::OptInlineAsm  ) {  ASM_CALL(AsmDhopSiteInt);    return;}
 #endif
    } else if( exterior ) {
+     // dependent on result of merge
      acceleratorFenceComputeStream();
-     if (Opt == WilsonKernelsStatic::OptGeneric    ) { KERNEL_CALL(GenericDhopSiteExt); return;}
-     if (Opt == WilsonKernelsStatic::OptHandUnroll ) { KERNEL_CALL(HandDhopSiteExt);    return;}
+     if (Opt == WilsonKernelsStatic::OptGeneric    ) { KERNEL_CALL_EXT(GenericDhopSiteExt); return;}
+     if (Opt == WilsonKernelsStatic::OptHandUnroll ) { KERNEL_CALL_EXT(HandDhopSiteExt);    return;}
 #ifndef GRID_CUDA
      if (Opt == WilsonKernelsStatic::OptInlineAsm  ) {  ASM_CALL(AsmDhopSiteExt);    return;}
 #endif
@@ -494,6 +496,7 @@ void WilsonKernels<Impl>::DhopKernel(int Opt,StencilImpl &st,  DoubledGaugeField
     autoView(st_v ,st,AcceleratorRead);
 
    if( interior && exterior ) {
+     acceleratorFenceComputeStream();
      if (Opt == WilsonKernelsStatic::OptGeneric    ) { KERNEL_CALL(GenericDhopSiteDag); return;}
      if (Opt == WilsonKernelsStatic::OptHandUnroll ) { KERNEL_CALL(HandDhopSiteDag);    return;}
 #ifndef GRID_CUDA
@@ -506,9 +509,10 @@ void WilsonKernels<Impl>::DhopKernel(int Opt,StencilImpl &st,  DoubledGaugeField
      if (Opt == WilsonKernelsStatic::OptInlineAsm  ) {  ASM_CALL(AsmDhopSiteDagInt);     return;}
 #endif
    } else if( exterior ) {
+     // Dependent on result of merge
      acceleratorFenceComputeStream();
-     if (Opt == WilsonKernelsStatic::OptGeneric    ) { KERNEL_CALL(GenericDhopSiteDagExt); return;}
-     if (Opt == WilsonKernelsStatic::OptHandUnroll ) { KERNEL_CALL(HandDhopSiteDagExt);    return;}
+     if (Opt == WilsonKernelsStatic::OptGeneric    ) { KERNEL_CALL_EXT(GenericDhopSiteDagExt); return;}
+     if (Opt == WilsonKernelsStatic::OptHandUnroll ) { KERNEL_CALL_EXT(HandDhopSiteDagExt);    return;}
 #ifndef GRID_CUDA
      if (Opt == WilsonKernelsStatic::OptInlineAsm  ) {  ASM_CALL(AsmDhopSiteDagExt);     return;}
 #endif

Grid/qcd/action/fermion/implementation/WilsonTMFermionImplementation.h

@@ -93,5 +93,25 @@ void WilsonTMFermion<Impl>::MooeeInvDag(const FermionField &in, FermionField &ou
   RealD b    = tm /sq;
   axpibg5x(out,in,a,b);
 }
+template<class Impl>
+void WilsonTMFermion<Impl>::M(const FermionField &in, FermionField &out) {
+  out.Checkerboard() = in.Checkerboard();
+  this->Dhop(in, out, DaggerNo);
+  FermionField tmp(out.Grid());
+  RealD a = 4.0+this->mass;
+  RealD b = this->mu;
+  axpibg5x(tmp,in,a,b);
+  axpy(out, 1.0, tmp, out);
+}
+template<class Impl>
+void WilsonTMFermion<Impl>::Mdag(const FermionField &in, FermionField &out) {
+  out.Checkerboard() = in.Checkerboard();
+  this->Dhop(in, out, DaggerYes);
+  FermionField tmp(out.Grid());
+  RealD a = 4.0+this->mass;
+  RealD b = -this->mu;
+  axpibg5x(tmp,in,a,b);
+  axpy(out, 1.0, tmp, out);
+}
 
 NAMESPACE_END(Grid);

Grid/qcd/action/fermion/instantiation/SpWilsonImplD/WilsonCloverFermionInstantiationSpWilsonImplD.cc

@@ -0,0 +1 @@
+../WilsonCloverFermionInstantiation.cc.master

Grid/qcd/action/fermion/instantiation/SpWilsonImplD/WilsonFermionInstantiationSpWilsonImplD.cc

@@ -0,0 +1 @@
+../WilsonFermionInstantiation.cc.master

Grid/qcd/action/fermion/instantiation/SpWilsonImplD/WilsonKernelsInstantiationSpWilsonImplD.cc

@@ -0,0 +1 @@
+../WilsonKernelsInstantiation.cc.master

Grid/qcd/action/fermion/instantiation/SpWilsonImplD/WilsonTMFermionInstantiationSpWilsonImplD.cc

@@ -0,0 +1 @@
+../WilsonTMFermionInstantiation.cc.master

Grid/qcd/action/fermion/instantiation/SpWilsonImplD/impl.h

@@ -0,0 +1 @@
+#define IMPLEMENTATION SpWilsonImplD

Grid/qcd/action/fermion/instantiation/SpWilsonImplF/WilsonCloverFermionInstantiationSpWilsonImplF.cc

@@ -0,0 +1 @@
+../WilsonCloverFermionInstantiation.cc.master

Grid/qcd/action/fermion/instantiation/SpWilsonImplF/WilsonFermionInstantiationSpWilsonImplF.cc

@@ -0,0 +1 @@
+../WilsonFermionInstantiation.cc.master

Grid/qcd/action/fermion/instantiation/SpWilsonImplF/WilsonKernelsInstantiationSpWilsonImplF.cc

@@ -0,0 +1 @@
+../WilsonKernelsInstantiation.cc.master

Grid/qcd/action/fermion/instantiation/SpWilsonImplF/WilsonTMFermionInstantiationSpWilsonImplF.cc

@@ -0,0 +1 @@
+../WilsonTMFermionInstantiation.cc.master

Grid/qcd/action/fermion/instantiation/SpWilsonImplF/impl.h

@@ -0,0 +1 @@
+#define IMPLEMENTATION SpWilsonImplF

Grid/qcd/action/fermion/instantiation/SpWilsonTwoIndexAntiSymmetricImplD/WilsonCloverFermionInstantiationSpWilsonTwoIndexAntiSymmetricImplD.cc

@@ -0,0 +1 @@
+../WilsonCloverFermionInstantiation.cc.master

Grid/qcd/action/fermion/instantiation/SpWilsonTwoIndexAntiSymmetricImplD/WilsonFermionInstantiationSpWilsonTwoIndexAntiSymmetricImplD.cc

@@ -0,0 +1 @@
+../WilsonFermionInstantiation.cc.master

Grid/qcd/action/fermion/instantiation/SpWilsonTwoIndexAntiSymmetricImplD/WilsonKernelsInstantiationSpWilsonTwoIndexAntiSymmetricImplD.cc

@@ -0,0 +1 @@
+../WilsonKernelsInstantiation.cc.master

Grid/qcd/action/fermion/instantiation/SpWilsonTwoIndexAntiSymmetricImplD/WilsonTMFermionInstantiationSpWilsonTwoIndexAntiSymmetricImplD.cc

@@ -0,0 +1 @@
+../WilsonTMFermionInstantiation.cc.master

Grid/qcd/action/fermion/instantiation/SpWilsonTwoIndexAntiSymmetricImplD/impl.h

@@ -0,0 +1 @@
+#define IMPLEMENTATION SpWilsonTwoIndexAntiSymmetricImplD

Grid/qcd/action/fermion/instantiation/SpWilsonTwoIndexAntiSymmetricImplF/WilsonCloverFermionInstantiationSpWilsonTwoIndexAntiSymmetricImplF.cc

@@ -0,0 +1 @@
+../WilsonCloverFermionInstantiation.cc.master

Grid/qcd/action/fermion/instantiation/SpWilsonTwoIndexAntiSymmetricImplF/WilsonFermionInstantiationSpWilsonTwoIndexAntiSymmetricImplF.cc

@@ -0,0 +1 @@
+../WilsonFermionInstantiation.cc.master

Grid/qcd/action/fermion/instantiation/SpWilsonTwoIndexAntiSymmetricImplF/WilsonKernelsInstantiationSpWilsonTwoIndexAntiSymmetricImplF.cc

@@ -0,0 +1 @@
+../WilsonKernelsInstantiation.cc.master

Grid/qcd/action/fermion/instantiation/SpWilsonTwoIndexAntiSymmetricImplF/WilsonTMFermionInstantiationSpWilsonTwoIndexAntiSymmetricImplF.cc

@@ -0,0 +1 @@
+../WilsonTMFermionInstantiation.cc.master

Grid/qcd/action/fermion/instantiation/SpWilsonTwoIndexAntiSymmetricImplF/impl.h

@@ -0,0 +1 @@
+#define IMPLEMENTATION SpWilsonTwoIndexAntiSymmetricImplF

Grid/qcd/action/fermion/instantiation/SpWilsonTwoIndexSymmetricImplD/WilsonCloverFermionInstantiationSpWilsonTwoIndexSymmetricImplD.cc

@@ -0,0 +1 @@
+../WilsonCloverFermionInstantiation.cc.master

Grid/qcd/action/fermion/instantiation/SpWilsonTwoIndexSymmetricImplD/WilsonFermionInstantiationSpWilsonTwoIndexSymmetricImplD.cc

@@ -0,0 +1 @@
+../WilsonFermionInstantiation.cc.master

Grid/qcd/action/fermion/instantiation/SpWilsonTwoIndexSymmetricImplD/WilsonKernelsInstantiationSpWilsonTwoIndexSymmetricImplD.cc

@@ -0,0 +1 @@
+../WilsonKernelsInstantiation.cc.master

Grid/qcd/action/fermion/instantiation/SpWilsonTwoIndexSymmetricImplD/WilsonTMFermionInstantiationSpWilsonTwoIndexSymmetricImplD.cc

@@ -0,0 +1 @@
+../WilsonTMFermionInstantiation.cc.master