Test_evec_compression changes:

Added ability to choose one of a variety of preselected basis sizes from the command line
	Fine lanczos now checks enough evecs are generated and resizes the output to Nstop and not the actual amount that converged (which can be larger)

.github/ISSUE_TEMPLATE/bug-report.yml

@@ -1,54 +0,0 @@
-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

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

Grid/DisableWarnings.h

@@ -34,6 +34,9 @@ directory
 
 #if defined __GNUC__ && __GNUC__>=6
 #pragma GCC diagnostic ignored "-Wignored-attributes"
+#endif
+#if defined __GNUC__ 
+#pragma GCC diagnostic ignored "-Wpsabi"
 #endif
 
  //disables and intel compiler specific warning (in json.hpp)
@@ -44,22 +47,14 @@ directory
 #ifdef __NVCC__
  //disables nvcc specific warning in json.hpp
 #pragma clang diagnostic ignored "-Wdeprecated-register"
-
-#ifdef __NVCC_DIAG_PRAGMA_SUPPORT__
- //disables nvcc specific warning in json.hpp
-#pragma nv_diag_suppress unsigned_compare_with_zero
-#pragma nv_diag_suppress cast_to_qualified_type
- //disables nvcc specific warning in many files
-#pragma nv_diag_suppress esa_on_defaulted_function_ignored
-#pragma nv_diag_suppress extra_semicolon
-#else
- //disables nvcc specific warning in json.hpp
 #pragma diag_suppress unsigned_compare_with_zero
 #pragma diag_suppress cast_to_qualified_type
+
  //disables nvcc specific warning in many files
 #pragma diag_suppress esa_on_defaulted_function_ignored
 #pragma diag_suppress extra_semicolon
-#endif
+
+//Eigen only
 #endif
 
 // Disable vectorisation in Eigen on the Power8/9 and PowerPC

Grid/GridCore.h

@@ -44,10 +44,9 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 #include <Grid/GridStd.h>
 #include <Grid/threads/Pragmas.h>
 #include <Grid/perfmon/Timer.h>
-//#include <Grid/perfmon/PerfCount.h>
+#include <Grid/perfmon/PerfCount.h>
 #include <Grid/util/Util.h>
 #include <Grid/log/Log.h>
-#include <Grid/perfmon/Tracing.h>
 #include <Grid/allocator/Allocator.h>
 #include <Grid/simd/Simd.h>
 #include <Grid/threads/ThreadReduction.h>

Grid/GridStd.h

@@ -16,7 +16,6 @@
 #include <functional>
 #include <stdio.h>
 #include <stdlib.h>
-#include <strings.h>
 #include <stdio.h>
 #include <signal.h>
 #include <ctime>

Grid/Grid_Eigen_Dense.h

@@ -14,11 +14,7 @@
 /* NVCC save and restore compile environment*/
 #ifdef __NVCC__
 #pragma push
-#ifdef __NVCC_DIAG_PRAGMA_SUPPORT__
-#pragma nv_diag_suppress code_is_unreachable
-#else
 #pragma diag_suppress code_is_unreachable
-#endif
 #pragma push_macro("__CUDA_ARCH__")
 #pragma push_macro("__NVCC__")
 #pragma push_macro("__CUDACC__")
@@ -34,7 +30,7 @@
 #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

@@ -66,10 +66,6 @@ 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

@@ -55,7 +55,6 @@ NAMESPACE_CHECK(BiCGSTAB);
 #include <Grid/algorithms/iterative/ConjugateGradientMultiShift.h>
 #include <Grid/algorithms/iterative/ConjugateGradientMixedPrec.h>
 #include <Grid/algorithms/iterative/ConjugateGradientMultiShiftMixedPrec.h>
-#include <Grid/algorithms/iterative/ConjugateGradientMixedPrecBatched.h>
 #include <Grid/algorithms/iterative/BiCGSTABMixedPrec.h>
 #include <Grid/algorithms/iterative/BlockConjugateGradient.h>
 #include <Grid/algorithms/iterative/ConjugateGradientReliableUpdate.h>

Grid/algorithms/CoarsenedMatrix.h

@@ -262,7 +262,7 @@ public:
 	autoView( Tnp_v , (*Tnp), AcceleratorWrite);
 	autoView( Tnm_v , (*Tnm), AcceleratorWrite);
 	const int Nsimd = CComplex::Nsimd();
-	accelerator_for(ss, FineGrid->oSites(), Nsimd, {
+	accelerator_forNB(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));
         });
@@ -324,9 +324,9 @@ public:
   GridBase*        _cbgrid;
   int hermitian;
 
-  CartesianStencil<siteVector,siteVector,DefaultImplParams> Stencil; 
+  CartesianStencil<siteVector,siteVector,int> Stencil; 
-  CartesianStencil<siteVector,siteVector,DefaultImplParams> StencilEven;
+  CartesianStencil<siteVector,siteVector,int> StencilEven;
-  CartesianStencil<siteVector,siteVector,DefaultImplParams> StencilOdd;
+  CartesianStencil<siteVector,siteVector,int> StencilOdd;
 
   std::vector<CoarseMatrix> A;
   std::vector<CoarseMatrix> Aeven;
@@ -631,7 +631,7 @@ public:
     assert(Aself != nullptr);
   }
 
-  void DselfInternal(CartesianStencil<siteVector,siteVector,DefaultImplParams> &st, CoarseMatrix &a,
+  void DselfInternal(CartesianStencil<siteVector,siteVector,int> &st, CoarseMatrix &a,
                        const CoarseVector &in, CoarseVector &out, int dag) {
     int point = geom.npoint-1;
     autoView( out_v, out, AcceleratorWrite);
@@ -694,7 +694,7 @@ public:
     }
   }
 
-  void DhopInternal(CartesianStencil<siteVector,siteVector,DefaultImplParams> &st, std::vector<CoarseMatrix> &a,
+  void DhopInternal(CartesianStencil<siteVector,siteVector,int> &st, std::vector<CoarseMatrix> &a,
                     const CoarseVector &in, CoarseVector &out, int dag) {
     SimpleCompressor<siteVector> compressor;
 
@@ -784,9 +784,9 @@ public:
     _cbgrid(new GridRedBlackCartesian(&CoarseGrid)),
     geom(CoarseGrid._ndimension),
     hermitian(hermitian_),
-    Stencil(&CoarseGrid,geom.npoint,Even,geom.directions,geom.displacements),
+    Stencil(&CoarseGrid,geom.npoint,Even,geom.directions,geom.displacements,0),
-    StencilEven(_cbgrid,geom.npoint,Even,geom.directions,geom.displacements),
+    StencilEven(_cbgrid,geom.npoint,Even,geom.directions,geom.displacements,0),
-    StencilOdd(_cbgrid,geom.npoint,Odd,geom.directions,geom.displacements),
+    StencilOdd(_cbgrid,geom.npoint,Odd,geom.directions,geom.displacements,0),
     A(geom.npoint,&CoarseGrid),
     Aeven(geom.npoint,_cbgrid),
     Aodd(geom.npoint,_cbgrid),
@@ -804,9 +804,9 @@ public:
     _cbgrid(&CoarseRBGrid),
     geom(CoarseGrid._ndimension),
     hermitian(hermitian_),
-    Stencil(&CoarseGrid,geom.npoint,Even,geom.directions,geom.displacements),
+    Stencil(&CoarseGrid,geom.npoint,Even,geom.directions,geom.displacements,0),
-    StencilEven(&CoarseRBGrid,geom.npoint,Even,geom.directions,geom.displacements),
+    StencilEven(&CoarseRBGrid,geom.npoint,Even,geom.directions,geom.displacements,0),
-    StencilOdd(&CoarseRBGrid,geom.npoint,Odd,geom.directions,geom.displacements),
+    StencilOdd(&CoarseRBGrid,geom.npoint,Odd,geom.directions,geom.displacements,0),
     A(geom.npoint,&CoarseGrid),
     Aeven(geom.npoint,&CoarseRBGrid),
     Aodd(geom.npoint,&CoarseRBGrid),

Grid/algorithms/FFT.h

@@ -29,7 +29,7 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #define _GRID_FFT_H_
 
 #ifdef HAVE_FFTW
-#if defined(USE_MKL) || defined(GRID_SYCL)
+#ifdef USE_MKL
 #include <fftw/fftw3.h>
 #else
 #include <fftw3.h>

Grid/algorithms/LinearOperator.h

@@ -526,7 +526,6 @@ public:
       (*this)(Linop,in[k],out[k]);
     }
   };
-  virtual ~OperatorFunction(){};
 };
 
 template<class Field> class LinearFunction {
@@ -542,7 +541,6 @@ public:
       (*this)(in[i], out[i]);
     }
   }
-  virtual ~LinearFunction(){};
 };
 
 template<class Field> class IdentityLinearFunction : public LinearFunction<Field> {

Grid/algorithms/approx/Chebyshev.h

@@ -258,12 +258,26 @@ public:
     for(int n=2;n<order;n++){
 
       Linop.HermOp(*Tn,y);
+#if 0
+      auto y_v = y.View();
+      auto Tn_v = Tn->View();
+      auto Tnp_v = Tnp->View();
+      auto Tnm_v = Tnm->View();
+      constexpr int Nsimd = vector_type::Nsimd();
+      accelerator_forNB(ss, in.Grid()->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));
+      });
+      if ( Coeffs[n] != 0.0) {
+	axpy(out,Coeffs[n],*Tnp,out);
+      }
+#else
       axpby(y,xscale,mscale,y,(*Tn));
       axpby(*Tnp,2.0,-1.0,y,(*Tnm));
       if ( Coeffs[n] != 0.0) {
 	axpy(out,Coeffs[n],*Tnp,out);
       }
-
+#endif
       // Cycle pointers to avoid copies
       Field *swizzle = Tnm;
       Tnm    =Tn;
@@ -278,6 +292,7 @@ public:
 template<class Field>
 class ChebyshevLanczos : public Chebyshev<Field> {
 private:
+
   std::vector<RealD> Coeffs;
   int order;
   RealD alpha;

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(ZOLO_PRECISION epsilon, int n, int type) {
+zolotarev_data* zolotarev(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(ZOLO_PRECISION epsilon, int n, int type) {
   construct_partfrac(d);
   construct_contfrac(d);
 
-  /* Converting everything to ZOLO_PRECISION for external use only */
+  /* Converting everything to PRECISION for external use only */
 
   zd = (zolotarev_data*) malloc(sizeof(zolotarev_data));
-  zd -> A = (ZOLO_PRECISION) d -> A;
+  zd -> A = (PRECISION) d -> A;
-  zd -> Delta = (ZOLO_PRECISION) d -> Delta;
+  zd -> Delta = (PRECISION) d -> Delta;
-  zd -> epsilon = (ZOLO_PRECISION) d -> epsilon;
+  zd -> epsilon = (PRECISION) d -> epsilon;
   zd -> n = d -> n;
   zd -> type = d -> type;
   zd -> dn = d -> dn;
@@ -390,24 +390,24 @@ zolotarev_data* zolotarev(ZOLO_PRECISION epsilon, int n, int type) {
   zd -> deg_num = d -> deg_num;
   zd -> deg_denom = d -> deg_denom;
 
-  zd -> a = (ZOLO_PRECISION*) malloc(zd -> dn * sizeof(ZOLO_PRECISION));
+  zd -> a = (PRECISION*) malloc(zd -> dn * sizeof(PRECISION));
-  for (m = 0; m < zd -> dn; m++) zd -> a[m] = (ZOLO_PRECISION) d -> a[m];
+  for (m = 0; m < zd -> dn; m++) zd -> a[m] = (PRECISION) d -> a[m];
   free(d -> a);
 
-  zd -> ap = (ZOLO_PRECISION*) malloc(zd -> dd * sizeof(ZOLO_PRECISION));
+  zd -> ap = (PRECISION*) malloc(zd -> dd * sizeof(PRECISION));
-  for (m = 0; m < zd -> dd; m++) zd -> ap[m] = (ZOLO_PRECISION) d -> ap[m];
+  for (m = 0; m < zd -> dd; m++) zd -> ap[m] = (PRECISION) d -> ap[m];
   free(d -> ap);
 
-  zd -> alpha = (ZOLO_PRECISION*) malloc(zd -> da * sizeof(ZOLO_PRECISION));
+  zd -> alpha = (PRECISION*) malloc(zd -> da * sizeof(PRECISION));
-  for (m = 0; m < zd -> da; m++) zd -> alpha[m] = (ZOLO_PRECISION) d -> alpha[m];
+  for (m = 0; m < zd -> da; m++) zd -> alpha[m] = (PRECISION) d -> alpha[m];
   free(d -> alpha);
 
-  zd -> beta = (ZOLO_PRECISION*) malloc(zd -> db * sizeof(ZOLO_PRECISION));
+  zd -> beta = (PRECISION*) malloc(zd -> db * sizeof(PRECISION));
-  for (m = 0; m < zd -> db; m++) zd -> beta[m] = (ZOLO_PRECISION) d -> beta[m];
+  for (m = 0; m < zd -> db; m++) zd -> beta[m] = (PRECISION) d -> beta[m];
   free(d -> beta);
 
-  zd -> gamma = (ZOLO_PRECISION*) malloc(zd -> n * sizeof(ZOLO_PRECISION));
+  zd -> gamma = (PRECISION*) malloc(zd -> n * sizeof(PRECISION));
-  for (m = 0; m < zd -> n; m++) zd -> gamma[m] = (ZOLO_PRECISION) d -> gamma[m];
+  for (m = 0; m < zd -> n; m++) zd -> gamma[m] = (PRECISION) d -> gamma[m];
   free(d -> gamma);
 
   free(d);
@@ -426,7 +426,7 @@ void zolotarev_free(zolotarev_data *zdata)
 }
 
 
-zolotarev_data* higham(ZOLO_PRECISION epsilon, int n) {
+zolotarev_data* higham(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(ZOLO_PRECISION epsilon, int n) {
   /* Converting everything to PRECISION for external use only */
 
   zd = (zolotarev_data*) malloc(sizeof(zolotarev_data));
-  zd -> A = (ZOLO_PRECISION) d -> A;
+  zd -> A = (PRECISION) d -> A;
-  zd -> Delta = (ZOLO_PRECISION) d -> Delta;
+  zd -> Delta = (PRECISION) d -> Delta;
-  zd -> epsilon = (ZOLO_PRECISION) d -> epsilon;
+  zd -> epsilon = (PRECISION) d -> epsilon;
   zd -> n = d -> n;
   zd -> type = d -> type;
   zd -> dn = d -> dn;
@@ -493,24 +493,24 @@ zolotarev_data* higham(ZOLO_PRECISION epsilon, int n) {
   zd -> deg_num = d -> deg_num;
   zd -> deg_denom = d -> deg_denom;
 
-  zd -> a = (ZOLO_PRECISION*) malloc(zd -> dn * sizeof(ZOLO_PRECISION));
+  zd -> a = (PRECISION*) malloc(zd -> dn * sizeof(PRECISION));
-  for (m = 0; m < zd -> dn; m++) zd -> a[m] = (ZOLO_PRECISION) d -> a[m];
+  for (m = 0; m < zd -> dn; m++) zd -> a[m] = (PRECISION) d -> a[m];
   free(d -> a);
 
-  zd -> ap = (ZOLO_PRECISION*) malloc(zd -> dd * sizeof(ZOLO_PRECISION));
+  zd -> ap = (PRECISION*) malloc(zd -> dd * sizeof(PRECISION));
-  for (m = 0; m < zd -> dd; m++) zd -> ap[m] = (ZOLO_PRECISION) d -> ap[m];
+  for (m = 0; m < zd -> dd; m++) zd -> ap[m] = (PRECISION) d -> ap[m];
   free(d -> ap);
 
-  zd -> alpha = (ZOLO_PRECISION*) malloc(zd -> da * sizeof(ZOLO_PRECISION));
+  zd -> alpha = (PRECISION*) malloc(zd -> da * sizeof(PRECISION));
-  for (m = 0; m < zd -> da; m++) zd -> alpha[m] = (ZOLO_PRECISION) d -> alpha[m];
+  for (m = 0; m < zd -> da; m++) zd -> alpha[m] = (PRECISION) d -> alpha[m];
   free(d -> alpha);
 
-  zd -> beta = (ZOLO_PRECISION*) malloc(zd -> db * sizeof(ZOLO_PRECISION));
+  zd -> beta = (PRECISION*) malloc(zd -> db * sizeof(PRECISION));
-  for (m = 0; m < zd -> db; m++) zd -> beta[m] = (ZOLO_PRECISION) d -> beta[m];
+  for (m = 0; m < zd -> db; m++) zd -> beta[m] = (PRECISION) d -> beta[m];
   free(d -> beta);
 
-  zd -> gamma = (ZOLO_PRECISION*) malloc(zd -> n * sizeof(ZOLO_PRECISION));
+  zd -> gamma = (PRECISION*) malloc(zd -> n * sizeof(PRECISION));
-  for (m = 0; m < zd -> n; m++) zd -> gamma[m] = (ZOLO_PRECISION) d -> gamma[m];
+  for (m = 0; m < zd -> n; m++) zd -> gamma[m] = (PRECISION) d -> gamma[m];
   free(d -> gamma);
 
   free(d);
@@ -523,17 +523,17 @@ NAMESPACE_END(Grid);
 #ifdef TEST
 
 #undef ZERO
-#define ZERO ((ZOLO_PRECISION) 0)
+#define ZERO ((PRECISION) 0)
 #undef ONE
-#define ONE ((ZOLO_PRECISION) 1)
+#define ONE ((PRECISION) 1)
 #undef TWO
-#define TWO ((ZOLO_PRECISION) 2)
+#define TWO ((PRECISION) 2)
 
 /* Evaluate the rational approximation R(x) using the factored form */
 
-static ZOLO_PRECISION zolotarev_eval(ZOLO_PRECISION x, zolotarev_data* rdata) {
+static PRECISION zolotarev_eval(PRECISION x, zolotarev_data* rdata) {
   int m;
-  ZOLO_PRECISION R;
+  PRECISION R;
 
   if (rdata -> type == 0) {
     R = rdata -> A * x;
@@ -551,9 +551,9 @@ static ZOLO_PRECISION zolotarev_eval(ZOLO_PRECISION x, zolotarev_data* rdata) {
 
 /* Evaluate the rational approximation R(x) using the partial fraction form */
 
-static ZOLO_PRECISION zolotarev_partfrac_eval(ZOLO_PRECISION x, zolotarev_data* rdata) {
+static PRECISION zolotarev_partfrac_eval(PRECISION x, zolotarev_data* rdata) {
   int m;
-  ZOLO_PRECISION R = rdata -> alpha[rdata -> da - 1];
+  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 ZOLO_PRECISION zolotarev_partfrac_eval(ZOLO_PRECISION x, zolotarev_data*
  * 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 ZOLO_PRECISION zolotarev_contfrac_eval(ZOLO_PRECISION x, zolotarev_data* rdata) {
+static PRECISION zolotarev_contfrac_eval(PRECISION x, zolotarev_data* rdata) {
   int m;
-  ZOLO_PRECISION R = rdata -> beta[0] * x;
+  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 ZOLO_PRECISION zolotarev_cayley_eval(ZOLO_PRECISION x, zolotarev_data* rdata) {
+static PRECISION zolotarev_cayley_eval(PRECISION x, zolotarev_data* rdata) {
   int m;
-  ZOLO_PRECISION T;
+  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;
-  ZOLO_PRECISION y;
+  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((ZOLO_PRECISION) eps, n) 
+    ? higham((PRECISION) eps, n) 
-    : zolotarev((ZOLO_PRECISION) eps, n, type);
+    : zolotarev((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)
-	 "\tZOLO_PRECISION = " STRINGIFY(ZOLO_PRECISION)
+	 "\tPRECISION = " STRINGIFY(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((ZOLO_PRECISION) x, rdata);
+      y = zolotarev_eval((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((ZOLO_PRECISION) x, rdata) - y)
+      ypferr = (float)((zolotarev_partfrac_eval((PRECISION) x, rdata) - y)
 		       / rdata -> Delta);
-      ycferr = (float)((zolotarev_contfrac_eval((ZOLO_PRECISION) x, rdata) - y)
+      ycferr = (float)((zolotarev_contfrac_eval((PRECISION) x, rdata) - y)
 		       / rdata -> Delta);
-      ycaylerr = (float)((zolotarev_cayley_eval((ZOLO_PRECISION) x, rdata) - y)
+      ycaylerr = (float)((zolotarev_cayley_eval((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 ZOLO_PRECISION
+#ifndef PRECISION
-#define ZOLO_PRECISION double
+#define PRECISION double
 #endif
-#define ZPRECISION ZOLO_PRECISION
+#define ZPRECISION 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(ZOLO_PRECISION epsilon, int n) ;
+ZOLOTAREV_DATA* higham(PRECISION epsilon, int n) ;
-ZOLOTAREV_DATA* zolotarev(ZOLO_PRECISION epsilon, int n, int type);
+ZOLOTAREV_DATA* zolotarev(PRECISION epsilon, int n, int type);
 void zolotarev_free(zolotarev_data *zdata);
 #endif
 
@@ -86,4 +86,3 @@ void zolotarev_free(zolotarev_data *zdata);
 NAMESPACE_END(Approx);
 NAMESPACE_END(Grid);
 #endif
-

Grid/algorithms/blas/BatchedBlas.cc

@@ -1,34 +0,0 @@
-/*************************************************************************************
-
-    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

@@ -1,727 +0,0 @@
-/*************************************************************************************
-
-    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/iterative/ConjugateGradient.h

@@ -58,7 +58,6 @@ public:
 
   void operator()(LinearOperatorBase<Field> &Linop, const Field &src, Field &psi) {
 
-    GRID_TRACE("ConjugateGradient");
     psi.Checkerboard() = src.Checkerboard();
 
     conformable(psi, src);
@@ -118,13 +117,9 @@ public:
     GridStopWatch MatrixTimer;
     GridStopWatch SolverTimer;
 
-    RealD usecs = -usecond();
     SolverTimer.Start();
     int k;
     for (k = 1; k <= MaxIterations; k++) {
-
-      GridStopWatch IterationTimer;
-      IterationTimer.Start();
       c = cp;
 
       MatrixTimer.Start();
@@ -157,41 +152,31 @@ public:
       LinearCombTimer.Stop();
       LinalgTimer.Stop();
 
-      IterationTimer.Stop();
+      std::cout << GridLogIterative << "ConjugateGradient: Iteration " << k
-      if ( (k % 500) == 0 ) {
-	std::cout << GridLogMessage << "ConjugateGradient: Iteration " << k
                 << " residual " << sqrt(cp/ssq) << " target " << Tolerance << std::endl;
-      } else { 
-	std::cout << GridLogIterative << "ConjugateGradient: Iteration " << k
-		  << " residual " << sqrt(cp/ssq) << " target " << Tolerance << " took " << IterationTimer.Elapsed() << std::endl;
-      }
 
       // Stopping condition
       if (cp <= rsq) {
-	usecs +=usecond();
         SolverTimer.Stop();
         Linop.HermOpAndNorm(psi, mmp, d, qq);
         p = mmp - src;
-	GridBase *grid = src.Grid();
+
-	RealD DwfFlops = (1452. )*grid->gSites()*4*k
-   	               + (8+4+8+4+4)*12*grid->gSites()*k; // CG linear algebra
         RealD srcnorm = std::sqrt(norm2(src));
         RealD resnorm = std::sqrt(norm2(p));
         RealD true_residual = resnorm / srcnorm;
+
         std::cout << GridLogMessage << "ConjugateGradient Converged on iteration " << k 
 		  << "\tComputed residual " << std::sqrt(cp / ssq)
 		  << "\tTrue residual " << true_residual
 		  << "\tTarget " << Tolerance << std::endl;
 
-        std::cout << GridLogMessage << "Time breakdown "<<std::endl;
+        std::cout << GridLogIterative << "Time breakdown "<<std::endl;
-	std::cout << GridLogMessage << "\tElapsed    " << SolverTimer.Elapsed() <<std::endl;
+	std::cout << GridLogIterative << "\tElapsed    " << SolverTimer.Elapsed() <<std::endl;
-	std::cout << GridLogMessage << "\tMatrix     " << MatrixTimer.Elapsed() <<std::endl;
+	std::cout << GridLogIterative << "\tMatrix     " << MatrixTimer.Elapsed() <<std::endl;
-	std::cout << GridLogMessage << "\tLinalg     " << LinalgTimer.Elapsed() <<std::endl;
+	std::cout << GridLogIterative << "\tLinalg     " << LinalgTimer.Elapsed() <<std::endl;
-	std::cout << GridLogMessage << "\tInner      " << InnerTimer.Elapsed() <<std::endl;
+	std::cout << GridLogIterative << "\tInner      " << InnerTimer.Elapsed() <<std::endl;
-	std::cout << GridLogMessage << "\tAxpyNorm   " << AxpyNormTimer.Elapsed() <<std::endl;
+	std::cout << GridLogIterative << "\tAxpyNorm   " << AxpyNormTimer.Elapsed() <<std::endl;
-	std::cout << GridLogMessage << "\tLinearComb " << LinearCombTimer.Elapsed() <<std::endl;
+	std::cout << GridLogIterative << "\tLinearComb " << LinearCombTimer.Elapsed() <<std::endl;
-
-	std::cout << GridLogDebug << "\tMobius flop rate " << DwfFlops/ usecs<< " Gflops " <<std::endl;
 
         if (ErrorOnNoConverge) assert(true_residual / Tolerance < 10000.0);
 

Grid/algorithms/iterative/ConjugateGradientMixedPrec.h

@@ -82,6 +82,11 @@ NAMESPACE_BEGIN(Grid);
     RealD stop = src_norm * Tolerance*Tolerance;
 
     GridBase* DoublePrecGrid = src_d_in.Grid();
+
+    //Generate precision change workspaces
+    precisionChangeWorkspace wk_dp_from_sp(DoublePrecGrid, SinglePrecGrid);
+    precisionChangeWorkspace wk_sp_from_dp(SinglePrecGrid, DoublePrecGrid);
+
     FieldD tmp_d(DoublePrecGrid);
     tmp_d.Checkerboard() = cb;
     
@@ -108,10 +113,7 @@ NAMESPACE_BEGIN(Grid);
     GridStopWatch PrecChangeTimer;
     
     Integer &outer_iter = TotalOuterIterations; //so it will be equal to the final iteration count
-
+      
-    precisionChangeWorkspace pc_wk_sp_to_dp(DoublePrecGrid, SinglePrecGrid);
-    precisionChangeWorkspace pc_wk_dp_to_sp(SinglePrecGrid, DoublePrecGrid);
-    
     for(outer_iter = 0; outer_iter < MaxOuterIterations; outer_iter++){
       //Compute double precision rsd and also new RHS vector.
       Linop_d.HermOp(sol_d, tmp_d);
@@ -126,7 +128,7 @@ NAMESPACE_BEGIN(Grid);
       while(norm * inner_tol * inner_tol < stop) inner_tol *= 2;  // inner_tol = sqrt(stop/norm) ??
 
       PrecChangeTimer.Start();
-      precisionChange(src_f, src_d, pc_wk_dp_to_sp);
+      precisionChange(src_f, src_d, wk_sp_from_dp);
       PrecChangeTimer.Stop();
       
       sol_f = Zero();
@@ -145,7 +147,7 @@ NAMESPACE_BEGIN(Grid);
       
       //Convert sol back to double and add to double prec solution
       PrecChangeTimer.Start();
-      precisionChange(tmp_d, sol_f, pc_wk_sp_to_dp);
+      precisionChange(tmp_d, sol_f, wk_dp_from_sp);
       PrecChangeTimer.Stop();
       
       axpy(sol_d, 1.0, tmp_d, sol_d);

Grid/algorithms/iterative/ConjugateGradientMixedPrecBatched.h

@@ -1,213 +0,0 @@
-/*************************************************************************************
-
-    Grid physics library, www.github.com/paboyle/Grid 
-
-    Source file: ./lib/algorithms/iterative/ConjugateGradientMixedPrecBatched.h
-
-    Copyright (C) 2015
-
-    Author: Raoul Hodgson <raoul.hodgson@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 */
-#ifndef GRID_CONJUGATE_GRADIENT_MIXED_PREC_BATCHED_H
-#define GRID_CONJUGATE_GRADIENT_MIXED_PREC_BATCHED_H
-
-NAMESPACE_BEGIN(Grid);
-
-//Mixed precision restarted defect correction CG
-template<class FieldD,class FieldF, 
-  typename std::enable_if< getPrecision<FieldD>::value == 2, int>::type = 0,
-  typename std::enable_if< getPrecision<FieldF>::value == 1, int>::type = 0> 
-class MixedPrecisionConjugateGradientBatched : public LinearFunction<FieldD> {
-public:
-  using LinearFunction<FieldD>::operator();
-  RealD   Tolerance;
-  RealD   InnerTolerance; //Initial tolerance for inner CG. Defaults to Tolerance but can be changed
-  Integer MaxInnerIterations;
-  Integer MaxOuterIterations;
-  Integer MaxPatchupIterations;
-  GridBase* SinglePrecGrid; //Grid for single-precision fields
-  RealD OuterLoopNormMult; //Stop the outer loop and move to a final double prec solve when the residual is OuterLoopNormMult * Tolerance
-  LinearOperatorBase<FieldF> &Linop_f;
-  LinearOperatorBase<FieldD> &Linop_d;
-
-  //Option to speed up *inner single precision* solves using a LinearFunction that produces a guess
-  LinearFunction<FieldF> *guesser;
-  bool updateResidual;
-  
-  MixedPrecisionConjugateGradientBatched(RealD tol, 
-          Integer maxinnerit, 
-          Integer maxouterit, 
-          Integer maxpatchit,
-          GridBase* _sp_grid, 
-          LinearOperatorBase<FieldF> &_Linop_f, 
-          LinearOperatorBase<FieldD> &_Linop_d,
-          bool _updateResidual=true) :
-    Linop_f(_Linop_f), Linop_d(_Linop_d),
-    Tolerance(tol), InnerTolerance(tol), MaxInnerIterations(maxinnerit), MaxOuterIterations(maxouterit), MaxPatchupIterations(maxpatchit), SinglePrecGrid(_sp_grid),
-    OuterLoopNormMult(100.), guesser(NULL), updateResidual(_updateResidual) { };
-
-  void useGuesser(LinearFunction<FieldF> &g){
-    guesser = &g;
-  }
-  
-  void operator() (const FieldD &src_d_in, FieldD &sol_d){
-    std::vector<FieldD> srcs_d_in{src_d_in};
-    std::vector<FieldD> sols_d{sol_d};
-
-    (*this)(srcs_d_in,sols_d);
-
-    sol_d = sols_d[0];
-  }
-
-  void operator() (const std::vector<FieldD> &src_d_in, std::vector<FieldD> &sol_d){
-    assert(src_d_in.size() == sol_d.size());
-    int NBatch = src_d_in.size();
-
-    std::cout << GridLogMessage << "NBatch = " << NBatch << std::endl;
-
-    Integer TotalOuterIterations = 0; //Number of restarts
-    std::vector<Integer> TotalInnerIterations(NBatch,0);     //Number of inner CG iterations
-    std::vector<Integer> TotalFinalStepIterations(NBatch,0); //Number of CG iterations in final patch-up step
-  
-    GridStopWatch TotalTimer;
-    TotalTimer.Start();
-
-    GridStopWatch InnerCGtimer;
-    GridStopWatch PrecChangeTimer;
-    
-    int cb = src_d_in[0].Checkerboard();
-    
-    std::vector<RealD> src_norm;
-    std::vector<RealD> norm;
-    std::vector<RealD> stop;
-    
-    GridBase* DoublePrecGrid = src_d_in[0].Grid();
-    FieldD tmp_d(DoublePrecGrid);
-    tmp_d.Checkerboard() = cb;
-    
-    FieldD tmp2_d(DoublePrecGrid);
-    tmp2_d.Checkerboard() = cb;
-
-    std::vector<FieldD> src_d;
-    std::vector<FieldF> src_f;
-    std::vector<FieldF> sol_f;
-
-    for (int i=0; i<NBatch; i++) {
-      sol_d[i].Checkerboard() = cb;
-
-      src_norm.push_back(norm2(src_d_in[i]));
-      norm.push_back(0.);
-      stop.push_back(src_norm[i] * Tolerance*Tolerance);
-
-      src_d.push_back(src_d_in[i]); //source for next inner iteration, computed from residual during operation
-
-      src_f.push_back(SinglePrecGrid);
-      src_f[i].Checkerboard() = cb;
-
-      sol_f.push_back(SinglePrecGrid);
-      sol_f[i].Checkerboard() = cb;
-    }
-    
-    RealD inner_tol = InnerTolerance;
-    
-    ConjugateGradient<FieldF> CG_f(inner_tol, MaxInnerIterations);
-    CG_f.ErrorOnNoConverge = false;
-    
-    Integer &outer_iter = TotalOuterIterations; //so it will be equal to the final iteration count
-      
-    for(outer_iter = 0; outer_iter < MaxOuterIterations; outer_iter++){
-      std::cout << GridLogMessage << std::endl;
-      std::cout << GridLogMessage << "Outer iteration " << outer_iter << std::endl;
-      
-      bool allConverged = true;
-      
-      for (int i=0; i<NBatch; i++) {
-        //Compute double precision rsd and also new RHS vector.
-        Linop_d.HermOp(sol_d[i], tmp_d);
-        norm[i] = axpy_norm(src_d[i], -1., tmp_d, src_d_in[i]); //src_d is residual vector
-        
-        std::cout<<GridLogMessage<<"MixedPrecisionConjugateGradientBatched: Outer iteration " << outer_iter <<" solve " << i << " residual "<< norm[i] << " target "<< stop[i] <<std::endl;
-
-        PrecChangeTimer.Start();
-        precisionChange(src_f[i], src_d[i]);
-        PrecChangeTimer.Stop();
-        
-        sol_f[i] = Zero();
-      
-        if(norm[i] > OuterLoopNormMult * stop[i]) {
-          allConverged = false;
-        }
-      }
-      if (allConverged) break;
-
-      if (updateResidual) {
-        RealD normMax = *std::max_element(std::begin(norm), std::end(norm));
-        RealD stopMax = *std::max_element(std::begin(stop), std::end(stop));
-        while( normMax * inner_tol * inner_tol < stopMax) inner_tol *= 2;  // inner_tol = sqrt(stop/norm) ??
-        CG_f.Tolerance = inner_tol;
-      }
-
-      //Optionally improve inner solver guess (eg using known eigenvectors)
-      if(guesser != NULL) {
-        (*guesser)(src_f, sol_f);
-      }
-
-      for (int i=0; i<NBatch; i++) {
-        //Inner CG
-        InnerCGtimer.Start();
-        CG_f(Linop_f, src_f[i], sol_f[i]);
-        InnerCGtimer.Stop();
-        TotalInnerIterations[i] += CG_f.IterationsToComplete;
-        
-        //Convert sol back to double and add to double prec solution
-        PrecChangeTimer.Start();
-        precisionChange(tmp_d, sol_f[i]);
-        PrecChangeTimer.Stop();
-        
-        axpy(sol_d[i], 1.0, tmp_d, sol_d[i]);
-      }
-
-    }
-    
-    //Final trial CG
-    std::cout << GridLogMessage << std::endl;
-    std::cout<<GridLogMessage<<"MixedPrecisionConjugateGradientBatched: Starting final patch-up double-precision solve"<<std::endl;
-    
-    for (int i=0; i<NBatch; i++) {
-      ConjugateGradient<FieldD> CG_d(Tolerance, MaxPatchupIterations);
-      CG_d(Linop_d, src_d_in[i], sol_d[i]);
-      TotalFinalStepIterations[i] += CG_d.IterationsToComplete;
-    }
-
-    TotalTimer.Stop();
-
-    std::cout << GridLogMessage << std::endl;
-    for (int i=0; i<NBatch; i++) {
-      std::cout<<GridLogMessage<<"MixedPrecisionConjugateGradientBatched: solve " << i << " Inner CG iterations " << TotalInnerIterations[i] << " Restarts " << TotalOuterIterations << " Final CG iterations " << TotalFinalStepIterations[i] << std::endl;
-    }
-    std::cout << GridLogMessage << std::endl;
-    std::cout<<GridLogMessage<<"MixedPrecisionConjugateGradientBatched: Total time " << TotalTimer.Elapsed() << " Precision change " << PrecChangeTimer.Elapsed() << " Inner CG total " << InnerCGtimer.Elapsed() << std::endl;
-    
-  }
-};
-
-NAMESPACE_END(Grid);
-
-#endif

Grid/algorithms/iterative/ConjugateGradientMultiShift.h

@@ -44,7 +44,7 @@ public:
 
   using OperatorFunction<Field>::operator();
 
-  //  RealD   Tolerance;
+  RealD   Tolerance;
   Integer MaxIterations;
   Integer IterationsToComplete; //Number of iterations the CG took to finish. Filled in upon completion
   std::vector<int> IterationsToCompleteShift;  // Iterations for this shift
@@ -84,7 +84,6 @@ public:
 
   void operator() (LinearOperatorBase<Field> &Linop, const Field &src, std::vector<Field> &psi)
   {
-    GRID_TRACE("ConjugateGradientMultiShift");
   
     GridBase *grid = src.Grid();
   
@@ -325,8 +324,8 @@ public:
 
       std::cout << GridLogMessage << "Time Breakdown "<<std::endl;
       std::cout << GridLogMessage << "\tElapsed    " << SolverTimer.Elapsed()     <<std::endl;
-      std::cout << GridLogMessage << "\tAXPY     " << AXPYTimer.Elapsed()     <<std::endl;
+      std::cout << GridLogMessage << "\tAXPY    " << AXPYTimer.Elapsed()     <<std::endl;
-      std::cout << GridLogMessage << "\tMatrix   " << MatrixTimer.Elapsed()     <<std::endl;
+      std::cout << GridLogMessage << "\tMarix    " << MatrixTimer.Elapsed()     <<std::endl;
       std::cout << GridLogMessage << "\tShift    " << ShiftTimer.Elapsed()     <<std::endl;
 
       IterationsToComplete = k;	

Grid/algorithms/iterative/ConjugateGradientMultiShiftCleanup.h

@@ -1,373 +0,0 @@
-/*************************************************************************************
-
-    Grid physics library, www.github.com/paboyle/Grid 
-
-    Source file: ./lib/algorithms/iterative/ConjugateGradientMultiShift.h
-
-    Copyright (C) 2015
-
-Author: Azusa Yamaguchi <ayamaguc@staffmail.ed.ac.uk>
-Author: Peter Boyle <paboyle@ph.ed.ac.uk>
-Author: Christopher Kelly <ckelly@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);
-
-//CK 2020: A variant of the multi-shift conjugate gradient with the matrix multiplication in single precision. 
-//The residual is stored in single precision, but the search directions and solution are stored in double precision. 
-//Every update_freq iterations the residual is corrected in double precision. 
-//For safety the a final regular CG is applied to clean up if necessary
-
-//PB Pure single, then double fixup
-
-template<class FieldD, class FieldF,
-	 typename std::enable_if< getPrecision<FieldD>::value == 2, int>::type = 0,
-	 typename std::enable_if< getPrecision<FieldF>::value == 1, int>::type = 0> 
-class ConjugateGradientMultiShiftMixedPrecCleanup : public OperatorMultiFunction<FieldD>,
-					     public OperatorFunction<FieldD>
-{
-public:                                                
-
-  using OperatorFunction<FieldD>::operator();
-
-  RealD   Tolerance;
-  Integer MaxIterationsMshift;
-  Integer MaxIterations;
-  Integer IterationsToComplete; //Number of iterations the CG took to finish. Filled in upon completion
-  std::vector<int> IterationsToCompleteShift;  // Iterations for this shift
-  int verbose;
-  MultiShiftFunction shifts;
-  std::vector<RealD> TrueResidualShift;
-
-  int ReliableUpdateFreq; //number of iterations between reliable updates
-
-  GridBase* SinglePrecGrid; //Grid for single-precision fields
-  LinearOperatorBase<FieldF> &Linop_f; //single precision
-
-  ConjugateGradientMultiShiftMixedPrecCleanup(Integer maxit, const MultiShiftFunction &_shifts,
-				       GridBase* _SinglePrecGrid, LinearOperatorBase<FieldF> &_Linop_f,
-				       int _ReliableUpdateFreq) : 
-    MaxIterationsMshift(maxit),  shifts(_shifts), SinglePrecGrid(_SinglePrecGrid), Linop_f(_Linop_f), ReliableUpdateFreq(_ReliableUpdateFreq),
-    MaxIterations(20000)
-  { 
-    verbose=1;
-    IterationsToCompleteShift.resize(_shifts.order);
-    TrueResidualShift.resize(_shifts.order);
-  }
-
-  void operator() (LinearOperatorBase<FieldD> &Linop, const FieldD &src, FieldD &psi)
-  {
-    GridBase *grid = src.Grid();
-    int nshift = shifts.order;
-    std::vector<FieldD> results(nshift,grid);
-    (*this)(Linop,src,results,psi);
-  }
-  void operator() (LinearOperatorBase<FieldD> &Linop, const FieldD &src, std::vector<FieldD> &results, FieldD &psi)
-  {
-    int nshift = shifts.order;
-
-    (*this)(Linop,src,results);
-  
-    psi = shifts.norm*src;
-    for(int i=0;i<nshift;i++){
-      psi = psi + shifts.residues[i]*results[i];
-    }
-
-    return;
-  }
-
-  void operator() (LinearOperatorBase<FieldD> &Linop_d, const FieldD &src_d, std::vector<FieldD> &psi_d)
-  { 
-    GRID_TRACE("ConjugateGradientMultiShiftMixedPrecCleanup");
-    GridBase *DoublePrecGrid = src_d.Grid();
-
-    ////////////////////////////////////////////////////////////////////////
-    // Convenience references to the info stored in "MultiShiftFunction"
-    ////////////////////////////////////////////////////////////////////////
-    int nshift = shifts.order;
-
-    std::vector<RealD> &mass(shifts.poles); // Make references to array in "shifts"
-    std::vector<RealD> &mresidual(shifts.tolerances);
-    std::vector<RealD> alpha(nshift,1.0);
-
-    //Double precision search directions
-    FieldD p_d(DoublePrecGrid);
-    std::vector<FieldF> ps_f (nshift, SinglePrecGrid);// Search directions (single precision)
-    std::vector<FieldF> psi_f(nshift, SinglePrecGrid);// solutions (single precision)
-
-    FieldD tmp_d(DoublePrecGrid);
-    FieldD r_d(DoublePrecGrid);
-    FieldF r_f(SinglePrecGrid);
-    FieldD mmp_d(DoublePrecGrid);
-
-    assert(psi_d.size()==nshift);
-    assert(mass.size()==nshift);
-    assert(mresidual.size()==nshift);
-  
-    // dynamic sized arrays on stack; 2d is a pain with vector
-    RealD  bs[nshift];
-    RealD  rsq[nshift];
-    RealD  rsqf[nshift];
-    RealD  z[nshift][2];
-    int     converged[nshift];
-  
-    const int       primary =0;
-  
-    //Primary shift fields CG iteration
-    RealD a,b,c,d;
-    RealD cp,bp,qq; //prev
-  
-    // Matrix mult fields
-    FieldF p_f(SinglePrecGrid);
-    FieldF mmp_f(SinglePrecGrid);
-
-    // Check lightest mass
-    for(int s=0;s<nshift;s++){
-      assert( mass[s]>= mass[primary] );
-      converged[s]=0;
-    }
-  
-    // Wire guess to zero
-    // Residuals "r" are src
-    // First search direction "p" is also src
-    cp = norm2(src_d);
-
-    // Handle trivial case of zero src.
-    if( cp == 0. ){
-      for(int s=0;s<nshift;s++){
-	psi_d[s] = Zero();
-	psi_f[s] = Zero();
-	IterationsToCompleteShift[s] = 1;
-	TrueResidualShift[s] = 0.;
-      }
-      return;
-    }
-
-    for(int s=0;s<nshift;s++){
-      rsq[s] = cp * mresidual[s] * mresidual[s];
-      rsqf[s] =rsq[s];
-      std::cout<<GridLogMessage<<"ConjugateGradientMultiShiftMixedPrecCleanup: shift "<< s <<" target resid "<<rsq[s]<<std::endl;
-      //      ps_d[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]
-    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)
-    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;
-    //    assert(norm2(tmp_d)< 1.0e-4);
-
-    axpy(mmp_d,mass[0],p_d,mmp_d);
-    RealD rn = norm2(p_d);
-    d += rn*mass[0];
-
-    b = -cp /d;
-  
-    // Set up the various shift variables
-    int       iz=0;
-    z[0][1-iz] = 1.0;
-    z[0][iz]   = 1.0;
-    bs[0]      = b;
-    for(int s=1;s<nshift;s++){
-      z[s][1-iz] = 1.0;
-      z[s][iz]   = 1.0/( 1.0 - b*(mass[s]-mass[0]));
-      bs[s]      = b*z[s][iz]; 
-    }
-  
-    // r += b[0] A.p[0]
-    // c= norm(r)
-    c=axpy_norm(r_d,b,mmp_d,r_d);
-  
-    for(int s=0;s<nshift;s++) {
-      axpby(psi_d[s],0.,-bs[s]*alpha[s],src_d,src_d);
-      precisionChange(psi_f[s],psi_d[s]);
-    }
-  
-    ///////////////////////////////////////
-    // Timers
-    ///////////////////////////////////////
-    GridStopWatch AXPYTimer, ShiftTimer, QRTimer, MatrixTimer, SolverTimer, PrecChangeTimer, CleanupTimer;
-
-    SolverTimer.Start();
-  
-    // Iteration loop
-    int k;
-  
-    for (k=1;k<=MaxIterationsMshift;k++){    
-
-      a = c /cp;
-      AXPYTimer.Start();
-      axpy(p_d,a,p_d,r_d); 
-      AXPYTimer.Stop();
-
-      PrecChangeTimer.Start();
-      precisionChange(r_f, r_d);
-      PrecChangeTimer.Stop();
-
-      AXPYTimer.Start();
-      for(int s=0;s<nshift;s++){
-	if ( ! converged[s] ) { 
-	  if (s==0){
-	    axpy(ps_f[s],a,ps_f[s],r_f);
-	  } else{
-	    RealD as =a *z[s][iz]*bs[s] /(z[s][1-iz]*b);
-	    axpby(ps_f[s],z[s][iz],as,r_f,ps_f[s]);
-	  }
-	}
-      }
-      AXPYTimer.Stop();
-
-      cp=c;
-      PrecChangeTimer.Start();
-      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();
-      precisionChange(mmp_d, mmp_f); // From Float to Double
-      PrecChangeTimer.Stop();
-
-      d=real(innerProduct(p_d,mmp_d));    
-      axpy(mmp_d,mass[0],p_d,mmp_d);
-      RealD rn = norm2(p_d);
-      d += rn*mass[0];
-    
-      bp=b;
-      b=-cp/d;
-
-      // Toggle the recurrence history
-      bs[0] = b;
-      iz = 1-iz;
-      ShiftTimer.Start();
-      for(int s=1;s<nshift;s++){
-	if((!converged[s])){
-	  RealD z0 = z[s][1-iz];
-	  RealD z1 = z[s][iz];
-	  z[s][iz] = z0*z1*bp
-	    / (b*a*(z1-z0) + z1*bp*(1- (mass[s]-mass[0])*b)); 
-	  bs[s] = b*z[s][iz]/z0; // NB sign  rel to Mike
-	}
-      }
-      ShiftTimer.Stop();
-
-      //Update single precision solutions
-      AXPYTimer.Start();
-      for(int s=0;s<nshift;s++){
-	int ss = s;
-	if( (!converged[s]) ) { 
-	  axpy(psi_f[ss],-bs[s]*alpha[s],ps_f[s],psi_f[ss]);
-	}
-      }
-      c = axpy_norm(r_d,b,mmp_d,r_d);
-      AXPYTimer.Stop();
-    
-      // Convergence checks
-      int all_converged = 1;
-      for(int s=0;s<nshift;s++){
-      
-	if ( (!converged[s]) ){
-	  IterationsToCompleteShift[s] = k;
-	
-	  RealD css  = c * z[s][iz]* z[s][iz];
-	
-	  if(css<rsqf[s]){
-	    if ( ! converged[s] )
-	      std::cout<<GridLogMessage<<"ConjugateGradientMultiShiftMixedPrecCleanup k="<<k<<" Shift "<<s<<" has converged"<<std::endl;
-	    converged[s]=1;
-	  } else {
-	    all_converged=0;
-	  }
-
-	}
-      }
-
-      if ( all_converged || k == MaxIterationsMshift-1){
-
-	SolverTimer.Stop();
-
-	for(int s=0;s<nshift;s++){
-	  precisionChange(psi_d[s],psi_f[s]);
-	}
-
-	
-	if ( all_converged ){
-	  std::cout<<GridLogMessage<< "ConjugateGradientMultiShiftMixedPrecCleanup: All shifts have converged iteration "<<k<<std::endl;
-	  std::cout<<GridLogMessage<< "ConjugateGradientMultiShiftMixedPrecCleanup: Checking solutions"<<std::endl;
-	} else {
-	  std::cout<<GridLogMessage<< "ConjugateGradientMultiShiftMixedPrecCleanup: Not all shifts have converged iteration "<<k<<std::endl;
-	}
-	
-	// Check answers 
-	for(int s=0; s < nshift; s++) { 
-	  Linop_d.HermOpAndNorm(psi_d[s],mmp_d,d,qq);
-	  axpy(tmp_d,mass[s],psi_d[s],mmp_d);
-	  axpy(r_d,-alpha[s],src_d,tmp_d);
-	  RealD rn = norm2(r_d);
-	  RealD cn = norm2(src_d);
-	  TrueResidualShift[s] = std::sqrt(rn/cn);
-	  std::cout<<GridLogMessage<<"ConjugateGradientMultiShiftMixedPrecCleanup: shift["<<s<<"] true residual "<< TrueResidualShift[s] << " target " << mresidual[s] << std::endl;
-
-	  //If we have not reached the desired tolerance, do a (mixed precision) CG cleanup
-	  if(rn >= rsq[s]){
-	    CleanupTimer.Start();
-	    std::cout<<GridLogMessage<<"ConjugateGradientMultiShiftMixedPrecCleanup: performing cleanup step for shift " << s << std::endl;
-
-	    //Setup linear operators for final cleanup
-	    ConjugateGradientMultiShiftMixedPrecSupport::ShiftedLinop<FieldD> Linop_shift_d(Linop_d, mass[s]);
-	    ConjugateGradientMultiShiftMixedPrecSupport::ShiftedLinop<FieldF> Linop_shift_f(Linop_f, mass[s]);
-					       
-	    MixedPrecisionConjugateGradient<FieldD,FieldF> cg(mresidual[s], MaxIterations, MaxIterations, SinglePrecGrid, Linop_shift_f, Linop_shift_d); 
-	    cg(src_d, psi_d[s]);
-	    
-	    TrueResidualShift[s] = cg.TrueResidual;
-	    CleanupTimer.Stop();
-	  }
-	}
-
-	std::cout << GridLogMessage << "ConjugateGradientMultiShiftMixedPrecCleanup: Time Breakdown for body"<<std::endl;
-	std::cout << GridLogMessage << "\tSolver    " << SolverTimer.Elapsed()     <<std::endl;
-	std::cout << GridLogMessage << "\t\tAXPY    " << AXPYTimer.Elapsed()     <<std::endl;
-	std::cout << GridLogMessage << "\t\tMatrix    " << MatrixTimer.Elapsed()     <<std::endl;
-	std::cout << GridLogMessage << "\t\tShift    " << ShiftTimer.Elapsed()     <<std::endl;
-	std::cout << GridLogMessage << "\t\tPrecision Change " << PrecChangeTimer.Elapsed()     <<std::endl;
-	std::cout << GridLogMessage << "\tFinal Cleanup " << CleanupTimer.Elapsed()     <<std::endl;
-	std::cout << GridLogMessage << "\tSolver+Cleanup " << SolverTimer.Elapsed() + CleanupTimer.Elapsed() << std::endl;
-
-	IterationsToComplete = k;	
-
-	return;
-      }
-   
-    }
-    std::cout<<GridLogMessage<<"CG multi shift did not converge"<<std::endl;
-    assert(0);
-  }
-
-};
-NAMESPACE_END(Grid);
-

Grid/algorithms/iterative/ConjugateGradientMultiShiftMixedPrec.h

@@ -81,7 +81,6 @@ public:
   using OperatorFunction<FieldD>::operator();
 
   RealD   Tolerance;
-  Integer MaxIterationsMshift;
   Integer MaxIterations;
   Integer IterationsToComplete; //Number of iterations the CG took to finish. Filled in upon completion
   std::vector<int> IterationsToCompleteShift;  // Iterations for this shift
@@ -96,9 +95,9 @@ public:
 
   ConjugateGradientMultiShiftMixedPrec(Integer maxit, const MultiShiftFunction &_shifts,
 				       GridBase* _SinglePrecGrid, LinearOperatorBase<FieldF> &_Linop_f,
-				       int _ReliableUpdateFreq) : 
+				       int _ReliableUpdateFreq
-    MaxIterationsMshift(maxit),  shifts(_shifts), SinglePrecGrid(_SinglePrecGrid), Linop_f(_Linop_f), ReliableUpdateFreq(_ReliableUpdateFreq),
+				       ) : 
-    MaxIterations(20000)
+    MaxIterations(maxit),  shifts(_shifts), SinglePrecGrid(_SinglePrecGrid), Linop_f(_Linop_f), ReliableUpdateFreq(_ReliableUpdateFreq)
   { 
     verbose=1;
     IterationsToCompleteShift.resize(_shifts.order);
@@ -128,12 +127,10 @@ public:
 
   void operator() (LinearOperatorBase<FieldD> &Linop_d, const FieldD &src_d, std::vector<FieldD> &psi_d)
   { 
-    GRID_TRACE("ConjugateGradientMultiShiftMixedPrec");
     GridBase *DoublePrecGrid = src_d.Grid();
+    precisionChangeWorkspace wk_f_from_d(SinglePrecGrid, DoublePrecGrid);
+    precisionChangeWorkspace wk_d_from_f(DoublePrecGrid, SinglePrecGrid);
 
-    precisionChangeWorkspace pc_wk_s_to_d(DoublePrecGrid,SinglePrecGrid);
-    precisionChangeWorkspace pc_wk_d_to_s(SinglePrecGrid,DoublePrecGrid);
-    
     ////////////////////////////////////////////////////////////////////////
     // Convenience references to the info stored in "MultiShiftFunction"
     ////////////////////////////////////////////////////////////////////////
@@ -158,7 +155,6 @@ public:
     // dynamic sized arrays on stack; 2d is a pain with vector
     RealD  bs[nshift];
     RealD  rsq[nshift];
-    RealD  rsqf[nshift];
     RealD  z[nshift][2];
     int     converged[nshift];
   
@@ -169,8 +165,12 @@ public:
     RealD cp,bp,qq; //prev
   
     // Matrix mult fields
+    FieldF r_f(SinglePrecGrid);
     FieldF p_f(SinglePrecGrid);
+    FieldF tmp_f(SinglePrecGrid);
     FieldF mmp_f(SinglePrecGrid);
+    FieldF src_f(SinglePrecGrid);
+    precisionChange(src_f, src_d, wk_f_from_d);
 
     // Check lightest mass
     for(int s=0;s<nshift;s++){
@@ -195,26 +195,18 @@ public:
 
     for(int s=0;s<nshift;s++){
       rsq[s] = cp * mresidual[s] * mresidual[s];
-      rsqf[s] =rsq[s];
       std::cout<<GridLogMessage<<"ConjugateGradientMultiShiftMixedPrec: shift "<< s <<" target resid "<<rsq[s]<<std::endl;
       ps_d[s] = src_d;
     }
     // r and p for primary
+    r_f=src_f; //residual maintained in single
+    p_f=src_f;
     p_d = src_d; //primary copy --- make this a reference to ps_d to save axpys
-    r_d = p_d;
+  
-    
     //MdagM+m[0]
-    precisionChange(p_f, p_d, pc_wk_d_to_s);
-
     Linop_f.HermOpAndNorm(p_f,mmp_f,d,qq); // mmp = MdagM p        d=real(dot(p, mmp)),  qq=norm2(mmp)
-    precisionChange(tmp_d, mmp_f, pc_wk_s_to_d);
+    axpy(mmp_f,mass[0],p_f,mmp_f);
-    Linop_d.HermOpAndNorm(p_d,mmp_d,d,qq); // mmp = MdagM p        d=real(dot(p, mmp)),  qq=norm2(mmp)
+    RealD rn = norm2(p_f);
-    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)< 1.0);
-
-    axpy(mmp_d,mass[0],p_d,mmp_d);
-    RealD rn = norm2(p_d);
     d += rn*mass[0];
 
     b = -cp /d;
@@ -232,7 +224,7 @@ public:
   
     // r += b[0] A.p[0]
     // c= norm(r)
-    c=axpy_norm(r_d,b,mmp_d,r_d);
+    c=axpy_norm(r_f,b,mmp_f,r_f);
   
     for(int s=0;s<nshift;s++) {
       axpby(psi_d[s],0.,-bs[s]*alpha[s],src_d,src_d);
@@ -248,9 +240,14 @@ public:
     // Iteration loop
     int k;
   
-    for (k=1;k<=MaxIterationsMshift;k++){    
+    for (k=1;k<=MaxIterations;k++){    
-
       a = c /cp;
+
+      //Update double precision search direction by residual
+      PrecChangeTimer.Start();
+      precisionChange(r_d, r_f, wk_d_from_f);
+      PrecChangeTimer.Stop();
+
       AXPYTimer.Start();
       axpy(p_d,a,p_d,r_d); 
 
@@ -267,28 +264,24 @@ public:
       AXPYTimer.Stop();
 
       PrecChangeTimer.Start();
-      precisionChange(p_f, p_d, pc_wk_d_to_s); //get back single prec search direction for linop
+      precisionChange(p_f, p_d, wk_f_from_d); //get back single prec search direction for linop
       PrecChangeTimer.Stop();
 
       cp=c;
       MatrixTimer.Start();  
-      Linop_f.HermOp(p_f,mmp_f);
+      Linop_f.HermOp(p_f,mmp_f); 
+      d=real(innerProduct(p_f,mmp_f));    
       MatrixTimer.Stop();  
 
-      PrecChangeTimer.Start();
-      precisionChange(mmp_d, mmp_f, pc_wk_s_to_d); // From Float to Double
-      PrecChangeTimer.Stop();
-
       AXPYTimer.Start();
-      d=real(innerProduct(p_d,mmp_d));    
+      axpy(mmp_f,mass[0],p_f,mmp_f);
-      axpy(mmp_d,mass[0],p_d,mmp_d);
       AXPYTimer.Stop();
-      RealD rn = norm2(p_d);
+      RealD rn = norm2(p_f);
       d += rn*mass[0];
     
       bp=b;
       b=-cp/d;
-
+    
       // Toggle the recurrence history
       bs[0] = b;
       iz = 1-iz;
@@ -314,12 +307,12 @@ public:
       }
 
       //Perform reliable update if necessary; otherwise update residual from single-prec mmp
-      c = axpy_norm(r_d,b,mmp_d,r_d);
+      RealD c_f = axpy_norm(r_f,b,mmp_f,r_f);
-
       AXPYTimer.Stop();
 
+      c = c_f;
+
       if(k % ReliableUpdateFreq == 0){
-	RealD c_old = c;
 	//Replace r with true residual
 	MatrixTimer.Start();  
 	Linop_d.HermOp(psi_d[0],mmp_d); 
@@ -328,10 +321,15 @@ public:
 	AXPYTimer.Start();
 	axpy(mmp_d,mass[0],psi_d[0],mmp_d);
 
-	c = axpy_norm(r_d, -1.0, mmp_d, src_d);
+	RealD c_d = axpy_norm(r_d, -1.0, mmp_d, src_d);
 	AXPYTimer.Stop();
 
-	std::cout<<GridLogMessage<<"ConjugateGradientMultiShiftMixedPrec k="<<k<< ", replaced |r|^2 = "<<c_old <<" with |r|^2 = "<<c<<std::endl;
+	std::cout<<GridLogMessage<<"ConjugateGradientMultiShiftMixedPrec k="<<k<< ", replaced |r|^2 = "<<c_f <<" with |r|^2 = "<<c_d<<std::endl;
+	
+	PrecChangeTimer.Start();
+	precisionChange(r_f, r_d, wk_f_from_d);
+	PrecChangeTimer.Stop();
+	c = c_d;
       }
     
       // Convergence checks
@@ -343,7 +341,7 @@ public:
 	
 	  RealD css  = c * z[s][iz]* z[s][iz];
 	
-	  if(css<rsqf[s]){
+	  if(css<rsq[s]){
 	    if ( ! converged[s] )
 	      std::cout<<GridLogMessage<<"ConjugateGradientMultiShiftMixedPrec k="<<k<<" Shift "<<s<<" has converged"<<std::endl;
 	    converged[s]=1;
@@ -354,17 +352,12 @@ public:
 	}
       }
 
-      if ( all_converged || k == MaxIterationsMshift-1){
+      if ( all_converged ){
 
 	SolverTimer.Stop();
-
+	std::cout<<GridLogMessage<< "ConjugateGradientMultiShiftMixedPrec: All shifts have converged iteration "<<k<<std::endl;
-	if ( all_converged ){
+	std::cout<<GridLogMessage<< "ConjugateGradientMultiShiftMixedPrec: Checking solutions"<<std::endl;
-	  std::cout<<GridLogMessage<< "ConjugateGradientMultiShiftMixedPrec: All shifts have converged iteration "<<k<<std::endl;
+      
-	  std::cout<<GridLogMessage<< "ConjugateGradientMultiShiftMixedPrec: Checking solutions"<<std::endl;
-	} else {
-	  std::cout<<GridLogMessage<< "ConjugateGradientMultiShiftMixedPrec: Not all shifts have converged iteration "<<k<<std::endl;
-	}
-	
 	// Check answers 
 	for(int s=0; s < nshift; s++) { 
 	  Linop_d.HermOpAndNorm(psi_d[s],mmp_d,d,qq);
@@ -405,10 +398,12 @@ public:
 
 	return;
       }
+
    
     }
+    // ugly hack
     std::cout<<GridLogMessage<<"CG multi shift did not converge"<<std::endl;
-    assert(0);
+    //  assert(0);
   }
 
 };

Grid/algorithms/iterative/ConjugateGradientReliableUpdate.h

@@ -48,7 +48,7 @@ public:
   LinearOperatorBase<FieldF> &Linop_f;
   LinearOperatorBase<FieldD> &Linop_d;
   GridBase* SinglePrecGrid;
-  RealD Delta; //reliable update parameter. A reliable update is performed when the residual drops by a factor of Delta relative to its value at the last update
+  RealD Delta; //reliable update parameter
 
   //Optional ability to switch to a different linear operator once the tolerance reaches a certain point. Useful for single/half -> single/single
   LinearOperatorBase<FieldF> *Linop_fallback;
@@ -65,9 +65,7 @@ public:
       ErrorOnNoConverge(err_on_no_conv),
       DoFinalCleanup(true),
       Linop_fallback(NULL)
-  {
+  {};
-    assert(Delta > 0. && Delta < 1. && "Expect  0 < Delta < 1");
-  };
 
   void setFallbackLinop(LinearOperatorBase<FieldF> &_Linop_fallback, const RealD _fallback_transition_tol){
     Linop_fallback = &_Linop_fallback;
@@ -75,7 +73,6 @@ public:
   }
     
   void operator()(const FieldD &src, FieldD &psi) {
-    GRID_TRACE("ConjugateGradientReliableUpdate");
     LinearOperatorBase<FieldF> *Linop_f_use = &Linop_f;
     bool using_fallback = false;
       
@@ -118,12 +115,9 @@ public:
     }
 
     //Single prec initialization
-    precisionChangeWorkspace pc_wk_sp_to_dp(src.Grid(), SinglePrecGrid);
-    precisionChangeWorkspace pc_wk_dp_to_sp(SinglePrecGrid, src.Grid());
-    
     FieldF r_f(SinglePrecGrid);
     r_f.Checkerboard() = r.Checkerboard();
-    precisionChange(r_f, r, pc_wk_dp_to_sp);
+    precisionChange(r_f, r);
 
     FieldF psi_f(r_f);
     psi_f = Zero();
@@ -139,8 +133,7 @@ public:
     GridStopWatch LinalgTimer;
     GridStopWatch MatrixTimer;
     GridStopWatch SolverTimer;
-    GridStopWatch PrecChangeTimer;
+
-    
     SolverTimer.Start();
     int k = 0;
     int l = 0;
@@ -179,9 +172,7 @@ public:
       // Stopping condition
       if (cp <= rsq) {
 	//Although not written in the paper, I assume that I have to add on the final solution
-	PrecChangeTimer.Start();
+	precisionChange(mmp, psi_f);
-	precisionChange(mmp, psi_f, pc_wk_sp_to_dp);
-	PrecChangeTimer.Stop();
 	psi = psi + mmp;
 	
 	
@@ -202,10 +193,7 @@ public:
 	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 << "\tPrecChange " << PrecChangeTimer.Elapsed() <<std::endl;
-	std::cout << GridLogMessage << "\tPrecChange avg time " << PrecChangeTimer.Elapsed()/(2*l+1) <<std::endl;
 
-	
 	IterationsToComplete = k;	
 	ReliableUpdatesPerformed = l;
 	  
@@ -225,21 +213,14 @@ public:
       else if(cp < Delta * MaxResidSinceLastRelUp) { //reliable update
 	std::cout << GridLogMessage << "ConjugateGradientReliableUpdate "
 		  << cp << "(residual) < " << Delta << "(Delta) * " << MaxResidSinceLastRelUp << "(MaxResidSinceLastRelUp) on iteration " << k << " : performing reliable update\n";
-	PrecChangeTimer.Start();
+	precisionChange(mmp, psi_f);
-	precisionChange(mmp, psi_f, pc_wk_sp_to_dp);
-	PrecChangeTimer.Stop();
 	psi = psi + mmp;
 
-	MatrixTimer.Start();
 	Linop_d.HermOpAndNorm(psi, mmp, d, qq);
-	MatrixTimer.Stop();
-	
 	r = src - mmp;
 
 	psi_f = Zero();
-	PrecChangeTimer.Start();
+	precisionChange(r_f, r);
-	precisionChange(r_f, r, pc_wk_dp_to_sp);
-	PrecChangeTimer.Stop();
 	cp = norm2(r);
 	MaxResidSinceLastRelUp = cp;
 

Grid/algorithms/iterative/Deflation.h

@@ -113,43 +113,7 @@ public:
     blockPromote(guess_coarse,guess,subspace);
     guess.Checkerboard() = src.Checkerboard();
   };
-
+};
-  void operator()(const std::vector<FineField> &src,std::vector<FineField> &guess) {
-    int Nevec = (int)evec_coarse.size();
-    int Nsrc = (int)src.size();
-    // make temp variables
-    std::vector<CoarseField> src_coarse(Nsrc,evec_coarse[0].Grid());
-    std::vector<CoarseField> guess_coarse(Nsrc,evec_coarse[0].Grid());    
-    //Preporcessing
-    std::cout << GridLogMessage << "Start BlockProject for loop" << std::endl;
-    for (int j=0;j<Nsrc;j++)
-    {
-    guess_coarse[j] = Zero();
-    std::cout << GridLogMessage << "BlockProject iter: " << j << std::endl;
-    blockProject(src_coarse[j],src[j],subspace);
-    }
-    //deflation set up for eigen vector batchsize 1 and source batch size equal number of sources
-    std::cout << GridLogMessage << "Start ProjectAccum for loop" << std::endl;
-    for (int i=0;i<Nevec;i++)
-    {
-      std::cout << GridLogMessage << "ProjectAccum Nvec: " << i << std::endl;
-      const CoarseField & tmp = evec_coarse[i];
-      for (int j=0;j<Nsrc;j++)
-      {
-        axpy(guess_coarse[j],TensorRemove(innerProduct(tmp,src_coarse[j])) / eval_coarse[i],tmp,guess_coarse[j]);
-      }
-    }
-    //postprocessing
-    std::cout << GridLogMessage << "Start BlockPromote for loop" << std::endl;
-    for (int j=0;j<Nsrc;j++)
-    {
-    std::cout << GridLogMessage << "BlockProject iter: " << j << std::endl;
-    blockPromote(guess_coarse[j],guess[j],subspace);
-    guess[j].Checkerboard() = src[j].Checkerboard();
-    }
-  };
-
-  };
 
 
 

Grid/algorithms/iterative/ImplicitlyRestartedLanczos.h

@@ -419,15 +419,14 @@ 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(Nstop);// was Nconv
+      eval.resize(Nconv);// Nstop?
-      evec.resize(Nstop,grid);// was Nconv
+      evec.resize(Nconv,grid);// Nstop?
       basisSortInPlace(evec,eval,reverse);
       
     }

Grid/algorithms/iterative/LocalCoherenceLanczos.h

@@ -146,21 +146,14 @@ public:
   LinearOperatorBase<FineField> &_Linop;
   RealD                             _coarse_relax_tol;
   std::vector<FineField>        &_subspace;
-
-  int _largestEvalIdxForReport; //The convergence of the LCL is based on the evals of the coarse grid operator, not those of the underlying fine grid operator
-                                //As a result we do not know what the eval range of the fine operator is until the very end, making tuning the Cheby bounds very difficult
-                                //To work around this issue, every restart we separately reconstruct the fine operator eval for the lowest and highest evec and print these
-                                //out alongside the evals of the coarse operator. To do so we need to know the index of the largest eval (i.e. Nstop-1)
-                                //NOTE: If largestEvalIdxForReport=-1 (default) then this is not performed
   
   ImplicitlyRestartedLanczosSmoothedTester(LinearFunction<CoarseField>   &Poly,
 					   OperatorFunction<FineField>   &smoother,
 					   LinearOperatorBase<FineField> &Linop,
 					   std::vector<FineField>        &subspace,
-					   RealD coarse_relax_tol=5.0e3,
+					   RealD coarse_relax_tol=5.0e3) 
-					   int largestEvalIdxForReport=-1) 
     : _smoother(smoother), _Linop(Linop), _Poly(Poly), _subspace(subspace),
-      _coarse_relax_tol(coarse_relax_tol), _largestEvalIdxForReport(largestEvalIdxForReport)
+      _coarse_relax_tol(coarse_relax_tol)  
   {    };
 
   //evalMaxApprox: approximation of largest eval of the fine Chebyshev operator (suitably wrapped by block projection)
@@ -186,12 +179,6 @@ public:
 	     <<" |H B[i] - eval[i]B[i]|^2 / evalMaxApprox^2 " << std::setw(25) << vv
 	     <<std::endl;
 
-    if(_largestEvalIdxForReport != -1 && (j==0 || j==_largestEvalIdxForReport)){
-      std::cout<<GridLogIRL << "Estimating true eval of fine grid operator for eval idx " << j << std::endl;
-      RealD tmp_eval;
-      ReconstructEval(j,eresid,B,tmp_eval,1.0); //don't use evalMaxApprox of coarse operator! (cf below)
-    }
-    
     int conv=0;
     if( (vv<eresid*eresid) ) conv = 1;
     return conv;
@@ -422,7 +409,7 @@ public:
     //////////////////////////////////////////////////////////////////////////////////////////////////
 
     Chebyshev<FineField>                                           ChebySmooth(cheby_smooth); //lower order Chebyshev of fine operator on fine grid used to smooth regenerated eigenvectors
-    ImplicitlyRestartedLanczosSmoothedTester<Fobj,CComplex,nbasis> ChebySmoothTester(ChebyOp,ChebySmooth,_FineOp,subspace,relax,Nstop-1); 
+    ImplicitlyRestartedLanczosSmoothedTester<Fobj,CComplex,nbasis> ChebySmoothTester(ChebyOp,ChebySmooth,_FineOp,subspace,relax); 
 
     evals_coarse.resize(Nm);
     evec_coarse.resize(Nm,_CoarseGrid);

Grid/allocator/AlignedAllocator.h

@@ -176,7 +176,6 @@ 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 deviceVector  = std::vector<T,devAllocator<T> >;
 template<class T> using cshiftVector = std::vector<T,cshiftAllocator<T> >;
 
 NAMESPACE_END(Grid);

Grid/allocator/MemoryManager.cc

@@ -4,14 +4,11 @@ NAMESPACE_BEGIN(Grid);
 
 /*Allocation types, saying which pointer cache should be used*/
 #define Cpu      (0)
-#define CpuHuge  (1)
+#define CpuSmall (1)
-#define CpuSmall (2)
+#define Acc      (2)
-#define Acc      (3)
+#define AccSmall (3)
-#define AccHuge  (4)
+#define Shared   (4)
-#define AccSmall (5)
+#define SharedSmall (5)
-#define Shared   (6)
-#define SharedHuge  (7)
-#define SharedSmall (8)
 #undef GRID_MM_VERBOSE 
 uint64_t total_shared;
 uint64_t total_device;
@@ -38,15 +35,12 @@ void MemoryManager::PrintBytes(void)
   
 }
 
-uint64_t MemoryManager::DeviceCacheBytes() { return CacheBytes[Acc] + CacheBytes[AccHuge] + CacheBytes[AccSmall]; }
-uint64_t MemoryManager::HostCacheBytes()   { return CacheBytes[Cpu] + CacheBytes[CpuHuge] + CacheBytes[CpuSmall]; }
-
 //////////////////////////////////////////////////////////////////////
 // Data tables for recently freed pooiniter caches
 //////////////////////////////////////////////////////////////////////
 MemoryManager::AllocationCacheEntry MemoryManager::Entries[MemoryManager::NallocType][MemoryManager::NallocCacheMax];
 int MemoryManager::Victim[MemoryManager::NallocType];
-int MemoryManager::Ncache[MemoryManager::NallocType] = { 2, 0, 8, 8, 0, 16, 8, 0, 16 };
+int MemoryManager::Ncache[MemoryManager::NallocType] = { 2, 8, 2, 8, 2, 8 };
 uint64_t MemoryManager::CacheBytes[MemoryManager::NallocType];
 //////////////////////////////////////////////////////////////////////
 // Actual allocation and deallocation utils
@@ -176,16 +170,6 @@ void MemoryManager::Init(void)
     }
   }
 
-  str= getenv("GRID_ALLOC_NCACHE_HUGE");
-  if ( str ) {
-    Nc = atoi(str);
-    if ( (Nc>=0) && (Nc < NallocCacheMax)) {
-      Ncache[CpuHuge]=Nc;
-      Ncache[AccHuge]=Nc;
-      Ncache[SharedHuge]=Nc;
-    }
-  }
-
   str= getenv("GRID_ALLOC_NCACHE_SMALL");
   if ( str ) {
     Nc = atoi(str);
@@ -206,9 +190,7 @@ void MemoryManager::InitMessage(void) {
   
   std::cout << GridLogMessage<< "MemoryManager::Init() setting up"<<std::endl;
 #ifdef ALLOCATION_CACHE
-  std::cout << GridLogMessage<< "MemoryManager::Init() cache pool for recent host   allocations: SMALL "<<Ncache[CpuSmall]<<" LARGE "<<Ncache[Cpu]<<" HUGE "<<Ncache[CpuHuge]<<std::endl;
+  std::cout << GridLogMessage<< "MemoryManager::Init() cache pool for recent allocations: SMALL "<<Ncache[CpuSmall]<<" LARGE "<<Ncache[Cpu]<<std::endl;
-  std::cout << GridLogMessage<< "MemoryManager::Init() cache pool for recent device allocations: SMALL "<<Ncache[AccSmall]<<" LARGE "<<Ncache[Acc]<<" Huge "<<Ncache[AccHuge]<<std::endl;
-  std::cout << GridLogMessage<< "MemoryManager::Init() cache pool for recent shared allocations: SMALL "<<Ncache[SharedSmall]<<" LARGE "<<Ncache[Shared]<<" Huge "<<Ncache[SharedHuge]<<std::endl;
 #endif
   
 #ifdef GRID_UVM
@@ -240,11 +222,8 @@ void MemoryManager::InitMessage(void) {
 void *MemoryManager::Insert(void *ptr,size_t bytes,int type) 
 {
 #ifdef ALLOCATION_CACHE
-  int cache;
+  bool small = (bytes < GRID_ALLOC_SMALL_LIMIT);
-  if      (bytes < GRID_ALLOC_SMALL_LIMIT) cache = type + 2;
+  int cache = type + small;
-  else if (bytes >= GRID_ALLOC_HUGE_LIMIT) cache = type + 1;
-  else                                     cache = type;
-
   return Insert(ptr,bytes,Entries[cache],Ncache[cache],Victim[cache],CacheBytes[cache]);  
 #else
   return ptr;
@@ -253,12 +232,11 @@ void *MemoryManager::Insert(void *ptr,size_t bytes,int type)
 
 void *MemoryManager::Insert(void *ptr,size_t bytes,AllocationCacheEntry *entries,int ncache,int &victim, uint64_t &cacheBytes) 
 {
+  assert(ncache>0);
 #ifdef GRID_OMP
   assert(omp_in_parallel()==0);
 #endif 
 
-  if (ncache == 0) return ptr;
-
   void * ret = NULL;
   int v = -1;
 
@@ -293,11 +271,8 @@ void *MemoryManager::Insert(void *ptr,size_t bytes,AllocationCacheEntry *entries
 void *MemoryManager::Lookup(size_t bytes,int type)
 {
 #ifdef ALLOCATION_CACHE
-  int cache;
+  bool small = (bytes < GRID_ALLOC_SMALL_LIMIT);
-  if      (bytes < GRID_ALLOC_SMALL_LIMIT) cache = type + 2;
+  int cache = type+small;
-  else if (bytes >= GRID_ALLOC_HUGE_LIMIT) cache = type + 1;
-  else                                     cache = type;
-
   return Lookup(bytes,Entries[cache],Ncache[cache],CacheBytes[cache]);
 #else
   return NULL;
@@ -306,6 +281,7 @@ void *MemoryManager::Lookup(size_t bytes,int type)
 
 void *MemoryManager::Lookup(size_t bytes,AllocationCacheEntry *entries,int ncache,uint64_t & cacheBytes) 
 {
+  assert(ncache>0);
 #ifdef GRID_OMP
   assert(omp_in_parallel()==0);
 #endif 

Grid/allocator/MemoryManager.h

@@ -35,12 +35,6 @@ NAMESPACE_BEGIN(Grid);
 // Move control to configure.ac and Config.h?
 
 #define GRID_ALLOC_SMALL_LIMIT (4096)
-#define GRID_ALLOC_HUGE_LIMIT  (2147483648)
-
-#define STRINGIFY(x) #x
-#define TOSTRING(x) STRINGIFY(x)
-#define FILE_LINE __FILE__ ":" TOSTRING(__LINE__)
-#define AUDIT(a) MemoryManager::Audit(FILE_LINE)
 
 /*Pinning pages is costly*/
 ////////////////////////////////////////////////////////////////////////////
@@ -71,21 +65,6 @@ enum ViewMode {
   CpuWriteDiscard = 0x10 // same for now
 };
 
-struct MemoryStatus {
-  uint64_t     DeviceBytes;
-  uint64_t     DeviceLRUBytes;
-  uint64_t     DeviceMaxBytes;
-  uint64_t     HostToDeviceBytes;
-  uint64_t     DeviceToHostBytes;
-  uint64_t     HostToDeviceXfer;
-  uint64_t     DeviceToHostXfer;
-  uint64_t     DeviceEvictions;
-  uint64_t     DeviceDestroy;
-  uint64_t     DeviceAllocCacheBytes;
-  uint64_t     HostAllocCacheBytes;
-};
-
-
 class MemoryManager {
 private:
 
@@ -99,7 +78,7 @@ private:
   } AllocationCacheEntry;
 
   static const int NallocCacheMax=128; 
-  static const int NallocType=9;
+  static const int NallocType=6;
   static AllocationCacheEntry Entries[NallocType][NallocCacheMax];
   static int Victim[NallocType];
   static int Ncache[NallocType];
@@ -113,9 +92,8 @@ private:
   static void *Insert(void *ptr,size_t bytes,AllocationCacheEntry *entries,int ncache,int &victim,uint64_t &cbytes) ;
   static void *Lookup(size_t bytes,AllocationCacheEntry *entries,int ncache,uint64_t &cbytes) ;
 
- public:
   static void PrintBytes(void);
-  static void Audit(std::string s);
+ public:
   static void Init(void);
   static void InitMessage(void);
   static void *AcceleratorAllocate(size_t bytes);
@@ -135,28 +113,7 @@ private:
   static uint64_t     DeviceToHostBytes;
   static uint64_t     HostToDeviceXfer;
   static uint64_t     DeviceToHostXfer;
-  static uint64_t     DeviceEvictions;
+ 
-  static uint64_t     DeviceDestroy;
-  
-  static uint64_t     DeviceCacheBytes();
-  static uint64_t     HostCacheBytes();
-
-  static MemoryStatus GetFootprint(void) {
-    MemoryStatus stat;
-    stat.DeviceBytes       = DeviceBytes;
-    stat.DeviceLRUBytes    = DeviceLRUBytes;
-    stat.DeviceMaxBytes    = DeviceMaxBytes;
-    stat.HostToDeviceBytes = HostToDeviceBytes;
-    stat.DeviceToHostBytes = DeviceToHostBytes;
-    stat.HostToDeviceXfer  = HostToDeviceXfer;
-    stat.DeviceToHostXfer  = DeviceToHostXfer;
-    stat.DeviceEvictions   = DeviceEvictions;
-    stat.DeviceDestroy     = DeviceDestroy;
-    stat.DeviceAllocCacheBytes = DeviceCacheBytes();
-    stat.HostAllocCacheBytes   = HostCacheBytes();
-    return stat;
-  };
-  
  private:
 #ifndef GRID_UVM
   //////////////////////////////////////////////////////////////////////
@@ -213,7 +170,6 @@ private:
 
  public:
   static void Print(void);
-  static void PrintAll(void);
   static void PrintState( void* CpuPtr);
   static int   isOpen   (void* CpuPtr);
   static void  ViewClose(void* CpuPtr,ViewMode mode);

Grid/allocator/MemoryManagerCache.cc

@@ -3,13 +3,8 @@
 
 #warning "Using explicit device memory copies"
 NAMESPACE_BEGIN(Grid);
-
+//#define dprintf(...) printf ( __VA_ARGS__ ); fflush(stdout);
-#define MAXLINE 512
+#define dprintf(...)
-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(...) 
 
 
 ////////////////////////////////////////////////////////////
@@ -28,8 +23,6 @@ uint64_t  MemoryManager::HostToDeviceBytes;
 uint64_t  MemoryManager::DeviceToHostBytes;
 uint64_t  MemoryManager::HostToDeviceXfer;
 uint64_t  MemoryManager::DeviceToHostXfer;
-uint64_t  MemoryManager::DeviceEvictions;
-uint64_t  MemoryManager::DeviceDestroy;
 
 ////////////////////////////////////
 // Priority ordering for unlocked entries
@@ -111,17 +104,15 @@ 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(%llx) %llx\n",(uint64_t)AccCache.CpuPtr,(uint64_t)AccCache.AccPtr); 
   assert(AccCache.accLock==0);
   assert(AccCache.cpuLock==0);
   assert(AccCache.CpuPtr!=(uint64_t)NULL);
   if(AccCache.AccPtr) {
     AcceleratorFree((void *)AccCache.AccPtr,AccCache.bytes);
-    DeviceDestroy++;
     DeviceBytes   -=AccCache.bytes;
     LRUremove(AccCache);
-    AccCache.AccPtr=(uint64_t) NULL;
+    dprintf("MemoryManager: Free(%llx) LRU %lld Total %lld\n",(uint64_t)AccCache.AccPtr,DeviceLRUBytes,DeviceBytes);  
-    dprintf("MemoryManager: Free(%lx) LRU %ld Total %ld\n",(uint64_t)AccCache.AccPtr,DeviceLRUBytes,DeviceBytes);  
   }
   uint64_t CpuPtr = AccCache.CpuPtr;
   EntryErase(CpuPtr);
@@ -130,36 +121,26 @@ void MemoryManager::AccDiscard(AcceleratorViewEntry &AccCache)
 void MemoryManager::Evict(AcceleratorViewEntry &AccCache)
 {
   ///////////////////////////////////////////////////////////////////////////
-  // Make CPU consistent, remove from Accelerator, remove from LRU, LEAVE CPU only entry
+  // Make CPU consistent, remove from Accelerator, remove entry
-  // Cannot be acclocked. If allocated must be in LRU pool.
+  // Cannot be locked. If allocated must be in LRU pool.
-  //
-  // Nov 2022... Felix issue: Allocating two CpuPtrs, can have an entry in LRU-q with CPUlock.
-  //                          and require to evict the AccPtr copy. Eviction was a mistake in CpuViewOpen
-  //                          but there is a weakness where CpuLock entries are attempted for erase
-  //                          Take these OUT LRU queue when CPU locked?
-  //                          Cannot take out the table as cpuLock data is important.
   ///////////////////////////////////////////////////////////////////////////
   assert(AccCache.state!=Empty);
   
-  mprintf("MemoryManager: Evict cpu %lx acc %lx cpuLock %ld accLock %ld\n",
+  dprintf("MemoryManager: Evict(%llx) %llx\n",(uint64_t)AccCache.CpuPtr,(uint64_t)AccCache.AccPtr); 
-	  (uint64_t)AccCache.CpuPtr,(uint64_t)AccCache.AccPtr,
+  assert(AccCache.accLock==0);
-	  (uint64_t)AccCache.cpuLock,(uint64_t)AccCache.accLock); 
+  assert(AccCache.cpuLock==0);
-  if (AccCache.accLock!=0) return;
-  if (AccCache.cpuLock!=0) return;
   if(AccCache.state==AccDirty) {
     Flush(AccCache);
   }
+  assert(AccCache.CpuPtr!=(uint64_t)NULL);
   if(AccCache.AccPtr) {
     AcceleratorFree((void *)AccCache.AccPtr,AccCache.bytes);
-    LRUremove(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);  
+    LRUremove(AccCache);
+    dprintf("MemoryManager: Free(%llx) footprint now %lld \n",(uint64_t)AccCache.AccPtr,DeviceBytes);  
   }
-  //  uint64_t CpuPtr = AccCache.CpuPtr;
+  uint64_t CpuPtr = AccCache.CpuPtr;
-  DeviceEvictions++;
+  EntryErase(CpuPtr);
-  //  EntryErase(CpuPtr);
 }
 void MemoryManager::Flush(AcceleratorViewEntry &AccCache)
 {
@@ -169,7 +150,7 @@ 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);
+  dprintf("MemoryManager: Flush  %llx -> %llx\n",(uint64_t)AccCache.AccPtr,(uint64_t)AccCache.CpuPtr); fflush(stdout);
   DeviceToHostBytes+=AccCache.bytes;
   DeviceToHostXfer++;
   AccCache.state=Consistent;
@@ -184,7 +165,7 @@ 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);
+  dprintf("MemoryManager: Clone %llx <- %llx\n",(uint64_t)AccCache.AccPtr,(uint64_t)AccCache.CpuPtr); fflush(stdout);
   acceleratorCopyToDevice((void *)AccCache.CpuPtr,(void *)AccCache.AccPtr,AccCache.bytes);
   HostToDeviceBytes+=AccCache.bytes;
   HostToDeviceXfer++;
@@ -210,7 +191,6 @@ void MemoryManager::CpuDiscard(AcceleratorViewEntry &AccCache)
 void MemoryManager::ViewClose(void* Ptr,ViewMode mode)
 {
   if( (mode==AcceleratorRead)||(mode==AcceleratorWrite)||(mode==AcceleratorWriteDiscard) ){
-    dprintf("AcceleratorViewClose %lx\n",(uint64_t)Ptr);
     AcceleratorViewClose((uint64_t)Ptr);
   } else if( (mode==CpuRead)||(mode==CpuWrite)){
     CpuViewClose((uint64_t)Ptr);
@@ -222,7 +202,6 @@ void *MemoryManager::ViewOpen(void* _CpuPtr,size_t bytes,ViewMode mode,ViewAdvis
 {
   uint64_t CpuPtr = (uint64_t)_CpuPtr;
   if( (mode==AcceleratorRead)||(mode==AcceleratorWrite)||(mode==AcceleratorWriteDiscard) ){
-    dprintf("AcceleratorViewOpen %lx\n",(uint64_t)CpuPtr);
     return (void *) AcceleratorViewOpen(CpuPtr,bytes,mode,hint);
   } else if( (mode==CpuRead)||(mode==CpuWrite)){
     return (void *)CpuViewOpen(CpuPtr,bytes,mode,hint);
@@ -233,16 +212,13 @@ void *MemoryManager::ViewOpen(void* _CpuPtr,size_t bytes,ViewMode mode,ViewAdvis
 }
 void  MemoryManager::EvictVictims(uint64_t bytes)
 {
-  assert(bytes<DeviceMaxBytes);
   while(bytes+DeviceLRUBytes > DeviceMaxBytes){
     if ( DeviceLRUBytes > 0){
       assert(LRU.size()>0);
-      uint64_t victim = LRU.back(); // From the LRU
+      uint64_t victim = LRU.back();
       auto AccCacheIterator = EntryLookup(victim);
       auto & AccCache = AccCacheIterator->second;
       Evict(AccCache);
-    } else {
-      return;
     }
   }
 }
@@ -265,12 +241,11 @@ uint64_t MemoryManager::AcceleratorViewOpen(uint64_t CpuPtr,size_t bytes,ViewMod
   assert(AccCache.cpuLock==0);  // Programming error
 
   if(AccCache.state!=Empty) {
-    dprintf("ViewOpen found entry %lx %lx : %ld %ld accLock %ld\n",
+    dprintf("ViewOpen found entry %llx %llx : %lld %lld\n",
 		    (uint64_t)AccCache.CpuPtr,
 		    (uint64_t)CpuPtr,
 		    (uint64_t)AccCache.bytes,
-	            (uint64_t)bytes,
+		    (uint64_t)bytes);
-		    (uint64_t)AccCache.accLock);
     assert(AccCache.CpuPtr == CpuPtr);
     assert(AccCache.bytes  ==bytes);
   }
@@ -305,7 +280,6 @@ uint64_t MemoryManager::AcceleratorViewOpen(uint64_t CpuPtr,size_t bytes,ViewMod
       AccCache.state  = Consistent; // Empty + AccRead => Consistent
     }
     AccCache.accLock= 1;
-    dprintf("Copied Empty entry into device accLock= %d\n",AccCache.accLock);
   } else if(AccCache.state==CpuDirty ){
     if(mode==AcceleratorWriteDiscard) {
       CpuDiscard(AccCache);
@@ -318,30 +292,28 @@ uint64_t MemoryManager::AcceleratorViewOpen(uint64_t CpuPtr,size_t bytes,ViewMod
       AccCache.state  = Consistent; // CpuDirty + AccRead => Consistent
     }
     AccCache.accLock++;
-    dprintf("CpuDirty entry into device ++accLock= %d\n",AccCache.accLock);
+    dprintf("Copied CpuDirty entry into device accLock %d\n",AccCache.accLock);
   } else if(AccCache.state==Consistent) {
     if((mode==AcceleratorWrite)||(mode==AcceleratorWriteDiscard))
       AccCache.state  = AccDirty;   // Consistent + AcceleratorWrite=> AccDirty
     else
       AccCache.state  = Consistent; // Consistent + AccRead => Consistent
     AccCache.accLock++;
-    dprintf("Consistent entry into device ++accLock= %d\n",AccCache.accLock);
+    dprintf("Consistent entry into device accLock %d\n",AccCache.accLock);
   } else if(AccCache.state==AccDirty) {
     if((mode==AcceleratorWrite)||(mode==AcceleratorWriteDiscard))
       AccCache.state  = AccDirty; // AccDirty + AcceleratorWrite=> AccDirty
     else
       AccCache.state  = AccDirty; // AccDirty + AccRead => AccDirty
     AccCache.accLock++;
-    dprintf("AccDirty entry ++accLock= %d\n",AccCache.accLock);
+    dprintf("AccDirty entry into device accLock %d\n",AccCache.accLock);
   } else {
     assert(0);
   }
 
-  assert(AccCache.accLock>0);
+  // If view is opened on device remove from LRU
-  // If view is opened on device must remove from LRU
   if(AccCache.LRU_valid==1){
     // must possibly remove from LRU as now locked on GPU
-    dprintf("AccCache entry removed from LRU \n");
     LRUremove(AccCache);
   }
 
@@ -362,12 +334,10 @@ void MemoryManager::AcceleratorViewClose(uint64_t CpuPtr)
   assert(AccCache.accLock>0);
 
   AccCache.accLock--;
+
   // Move to LRU queue if not locked and close on device
   if(AccCache.accLock==0) {
-    dprintf("AccleratorViewClose %lx AccLock decremented to %ld move to LRU queue\n",(uint64_t)CpuPtr,(uint64_t)AccCache.accLock);
     LRUinsert(AccCache);
-  } else {
-    dprintf("AccleratorViewClose %lx AccLock decremented to %ld\n",(uint64_t)CpuPtr,(uint64_t)AccCache.accLock);
   }
 }
 void MemoryManager::CpuViewClose(uint64_t CpuPtr)
@@ -404,10 +374,9 @@ uint64_t MemoryManager::CpuViewOpen(uint64_t CpuPtr,size_t bytes,ViewMode mode,V
   auto AccCacheIterator = EntryLookup(CpuPtr);
   auto & AccCache = AccCacheIterator->second;
 
-  // CPU doesn't need to free space
+  if (!AccCache.AccPtr) {
-  //  if (!AccCache.AccPtr) {
+     EvictVictims(bytes);
-  //    EvictVictims(bytes);
+  }
-  //  }
 
   assert((mode==CpuRead)||(mode==CpuWrite));
   assert(AccCache.accLock==0);  // Programming error
@@ -461,28 +430,20 @@ void  MemoryManager::NotifyDeletion(void *_ptr)
 void  MemoryManager::Print(void)
 {
   PrintBytes();
-  std::cout << GridLogMessage << "--------------------------------------------" << std::endl;
+  std::cout << GridLogDebug << "--------------------------------------------" << std::endl;
-  std::cout << GridLogMessage << "Memory Manager                             " << std::endl;
+  std::cout << GridLogDebug << "Memory Manager                             " << std::endl;
-  std::cout << GridLogMessage << "--------------------------------------------" << std::endl;
+  std::cout << GridLogDebug << "--------------------------------------------" << std::endl;
-  std::cout << GridLogMessage << DeviceBytes   << " bytes allocated on device " << std::endl;
+  std::cout << GridLogDebug << DeviceBytes   << " bytes allocated on device " << std::endl;
-  std::cout << GridLogMessage << DeviceLRUBytes<< " bytes evictable on device " << std::endl;
+  std::cout << GridLogDebug << DeviceLRUBytes<< " bytes evictable on device " << std::endl;
-  std::cout << GridLogMessage << DeviceMaxBytes<< " bytes max on device       " << std::endl;
+  std::cout << GridLogDebug << DeviceMaxBytes<< " bytes max on device       " << std::endl;
-  std::cout << GridLogMessage << HostToDeviceXfer << " transfers        to   device " << std::endl;
+  std::cout << GridLogDebug << HostToDeviceXfer << " transfers        to   device " << std::endl;
-  std::cout << GridLogMessage << DeviceToHostXfer << " transfers        from device " << std::endl;
+  std::cout << GridLogDebug << DeviceToHostXfer << " transfers        from device " << std::endl;
-  std::cout << GridLogMessage << HostToDeviceBytes<< " bytes transfered to   device " << std::endl;
+  std::cout << GridLogDebug << HostToDeviceBytes<< " bytes transfered to   device " << std::endl;
-  std::cout << GridLogMessage << DeviceToHostBytes<< " bytes transfered from device " << std::endl;
+  std::cout << GridLogDebug << DeviceToHostBytes<< " bytes transfered from device " << std::endl;
-  std::cout << GridLogMessage << DeviceEvictions  << " Evictions from device " << std::endl;
+  std::cout << GridLogDebug << AccViewTable.size()<< " vectors " << LRU.size()<<" evictable"<< std::endl;
-  std::cout << GridLogMessage << DeviceDestroy    << " Destroyed vectors on device " << std::endl;
+  std::cout << GridLogDebug << "--------------------------------------------" << std::endl;
-  std::cout << GridLogMessage << AccViewTable.size()<< " vectors " << LRU.size()<<" evictable"<< std::endl;
+  std::cout << GridLogDebug << "CpuAddr\t\tAccAddr\t\tState\t\tcpuLock\taccLock\tLRU_valid "<<std::endl;
-  std::cout << GridLogMessage << "--------------------------------------------" << std::endl;
+  std::cout << GridLogDebug << "--------------------------------------------" << std::endl;
-}
-void  MemoryManager::PrintAll(void)
-{
-  Print();
-  std::cout << GridLogMessage << std::endl;
-  std::cout << GridLogMessage << "--------------------------------------------" << std::endl;
-  std::cout << GridLogMessage << "CpuAddr\t\tAccAddr\t\tState\t\tcpuLock\taccLock\tLRU_valid "<<std::endl;
-  std::cout << GridLogMessage << "--------------------------------------------" << std::endl;
   for(auto it=AccViewTable.begin();it!=AccViewTable.end();it++){
     auto &AccCache = it->second;
     
@@ -492,13 +453,13 @@ void  MemoryManager::PrintAll(void)
     if ( AccCache.state==AccDirty ) str = std::string("AccDirty");
     if ( AccCache.state==Consistent)str = std::string("Consistent");
 
-    std::cout << GridLogMessage << "0x"<<std::hex<<AccCache.CpuPtr<<std::dec
+    std::cout << GridLogDebug << "0x"<<std::hex<<AccCache.CpuPtr<<std::dec
 	      << "\t0x"<<std::hex<<AccCache.AccPtr<<std::dec<<"\t" <<str
 	      << "\t" << AccCache.cpuLock
 	      << "\t" << AccCache.accLock
 	      << "\t" << AccCache.LRU_valid<<std::endl;
   }
-  std::cout << GridLogMessage << "--------------------------------------------" << std::endl;
+  std::cout << GridLogDebug << "--------------------------------------------" << std::endl;
 
 };
 int   MemoryManager::isOpen   (void* _CpuPtr) 
@@ -512,63 +473,6 @@ int   MemoryManager::isOpen   (void* _CpuPtr)
     return 0;
   }
 }
-void MemoryManager::Audit(std::string s)
-{
-  uint64_t CpuBytes=0;
-  uint64_t AccBytes=0;
-  uint64_t LruBytes1=0;
-  uint64_t LruBytes2=0;
-  uint64_t LruCnt=0;
-  
-  std::cout << " Memory Manager::Audit() from "<<s<<std::endl;
-  for(auto it=LRU.begin();it!=LRU.end();it++){
-    uint64_t cpuPtr = *it;
-    assert(EntryPresent(cpuPtr));
-    auto AccCacheIterator = EntryLookup(cpuPtr);
-    auto & AccCache = AccCacheIterator->second;
-    LruBytes2+=AccCache.bytes;
-    assert(AccCache.LRU_valid==1);
-    assert(AccCache.LRU_entry==it);
-  }
-  std::cout << " Memory Manager::Audit() LRU queue matches table entries "<<std::endl;
-
-  for(auto it=AccViewTable.begin();it!=AccViewTable.end();it++){
-    auto &AccCache = it->second;
-    
-    std::string str;
-    if ( AccCache.state==Empty    ) str = std::string("Empty");
-    if ( AccCache.state==CpuDirty ) str = std::string("CpuDirty");
-    if ( AccCache.state==AccDirty ) str = std::string("AccDirty");
-    if ( AccCache.state==Consistent)str = std::string("Consistent");
-
-    CpuBytes+=AccCache.bytes;
-    if( AccCache.AccPtr )    AccBytes+=AccCache.bytes;
-    if( AccCache.LRU_valid ) LruBytes1+=AccCache.bytes;
-    if( AccCache.LRU_valid ) LruCnt++;
-    
-    if ( AccCache.cpuLock || AccCache.accLock ) {
-      assert(AccCache.LRU_valid==0);
-
-      std::cout << GridLogError << s<< "\n\t 0x"<<std::hex<<AccCache.CpuPtr<<std::dec
-		<< "\t0x"<<std::hex<<AccCache.AccPtr<<std::dec<<"\t" <<str
-		<< "\t cpuLock  " << AccCache.cpuLock
-		<< "\t accLock  " << AccCache.accLock
-		<< "\t LRUvalid " << AccCache.LRU_valid<<std::endl;
-    }
-
-    assert( AccCache.cpuLock== 0 ) ;
-    assert( AccCache.accLock== 0 ) ;
-  }
-  std::cout << " Memory Manager::Audit() no locked table entries "<<std::endl;
-  assert(LruBytes1==LruBytes2);
-  assert(LruBytes1==DeviceLRUBytes);
-  std::cout << " Memory Manager::Audit() evictable bytes matches sum over table "<<std::endl;
-  assert(AccBytes==DeviceBytes);
-  std::cout << " Memory Manager::Audit() device bytes matches sum over table "<<std::endl;
-  assert(LruCnt == LRU.size());
-  std::cout << " Memory Manager::Audit() LRU entry count matches "<<std::endl;
-
-}
 
 void MemoryManager::PrintState(void* _CpuPtr)
 {
@@ -585,8 +489,8 @@ void MemoryManager::PrintState(void* _CpuPtr)
     if ( AccCache.state==EvictNext) str = std::string("EvictNext");
 
     std::cout << GridLogMessage << "CpuAddr\t\tAccAddr\t\tState\t\tcpuLock\taccLock\tLRU_valid "<<std::endl;
-    std::cout << GridLogMessage << "\tx"<<std::hex<<AccCache.CpuPtr<<std::dec
+    std::cout << GridLogMessage << "0x"<<std::hex<<AccCache.CpuPtr<<std::dec
-    << "\tx"<<std::hex<<AccCache.AccPtr<<std::dec<<"\t" <<str
+    << "\t0x"<<std::hex<<AccCache.AccPtr<<std::dec<<"\t" <<str
     << "\t" << AccCache.cpuLock
     << "\t" << AccCache.accLock
     << "\t" << AccCache.LRU_valid<<std::endl;

Grid/allocator/MemoryManagerShared.cc

@@ -12,10 +12,7 @@ uint64_t  MemoryManager::HostToDeviceBytes;
 uint64_t  MemoryManager::DeviceToHostBytes;
 uint64_t  MemoryManager::HostToDeviceXfer;
 uint64_t  MemoryManager::DeviceToHostXfer;
-uint64_t  MemoryManager::DeviceEvictions;
-uint64_t  MemoryManager::DeviceDestroy;
 
-void  MemoryManager::Audit(std::string s){};
 void  MemoryManager::ViewClose(void* AccPtr,ViewMode mode){};
 void *MemoryManager::ViewOpen(void* CpuPtr,size_t bytes,ViewMode mode,ViewAdvise hint){ return CpuPtr; };
 int   MemoryManager::isOpen   (void* CpuPtr) { return 0;}
@@ -24,7 +21,6 @@ void  MemoryManager::PrintState(void* CpuPtr)
 std::cout << GridLogMessage << "Host<->Device memory movement not currently managed by Grid." << std::endl;
 };
 void  MemoryManager::Print(void){};
-void  MemoryManager::PrintAll(void){};
 void  MemoryManager::NotifyDeletion(void *ptr){};
 
 NAMESPACE_END(Grid);

Grid/communicator/Communicator_base.h

@@ -53,11 +53,10 @@ public:
   // Communicator should know nothing of the physics grid, only processor grid.
   ////////////////////////////////////////////
   int              _Nprocessors;     // How many in all
-  int              _processor;       // linear processor rank
-  unsigned long    _ndimension;
-  Coordinate _shm_processors;  // Which dimensions get relayed out over processors lanes.
   Coordinate _processors;      // Which dimensions get relayed out over processors lanes.
+  int              _processor;       // linear processor rank
   Coordinate _processor_coor;  // linear processor coordinate
+  unsigned long    _ndimension;
   static Grid_MPI_Comm      communicator_world;
   Grid_MPI_Comm             communicator;
   std::vector<Grid_MPI_Comm> communicator_halo;
@@ -98,16 +97,14 @@ public:
   int                      BossRank(void)          ;
   int                      ThisRank(void)          ;
   const Coordinate & ThisProcessorCoor(void) ;
-  const Coordinate & ShmGrid(void)  { return _shm_processors; }  ;
   const Coordinate & ProcessorGrid(void)     ;
-  int                ProcessorCount(void)    ;
+  int                      ProcessorCount(void)    ;
 
   ////////////////////////////////////////////////////////////////////////////////
   // very VERY rarely (Log, serial RNG) we need world without a grid
   ////////////////////////////////////////////////////////////////////////////////
   static int  RankWorld(void) ;
   static void BroadcastWorld(int root,void* data, int bytes);
-  static void BarrierWorld(void);
   
   ////////////////////////////////////////////////////////////
   // Reduction
@@ -131,7 +128,7 @@ public:
   template<class obj> void GlobalSum(obj &o){
     typedef typename obj::scalar_type scalar_type;
     int words = sizeof(obj)/sizeof(scalar_type);
-    scalar_type * ptr = (scalar_type *)& o; // Safe alias 
+    scalar_type * ptr = (scalar_type *)& o;
     GlobalSumVector(ptr,words);
   }
   
@@ -145,17 +142,17 @@ public:
 		      int bytes);
   
   double StencilSendToRecvFrom(void *xmit,
-			       int xmit_to_rank,int do_xmit,
+			       int xmit_to_rank,
 			       void *recv,
-			       int recv_from_rank,int do_recv,
+			       int recv_from_rank,
 			       int bytes,int dir);
 
   double StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list,
 				    void *xmit,
-				    int xmit_to_rank,int do_xmit,
+				    int xmit_to_rank,
 				    void *recv,
-				    int recv_from_rank,int do_recv,
+				    int recv_from_rank,
-				    int xbytes,int rbytes,int dir);
+				    int bytes,int dir);
   
   
   void StencilSendToRecvFromComplete(std::vector<CommsRequest_t> &waitall,int i);

Grid/communicator/Communicator_mpi3.cc

@@ -106,7 +106,7 @@ CartesianCommunicator::CartesianCommunicator(const Coordinate &processors)
   // Remap using the shared memory optimising routine
   // The remap creates a comm which must be freed
   ////////////////////////////////////////////////////
-  GlobalSharedMemory::OptimalCommunicator    (processors,optimal_comm,_shm_processors);
+  GlobalSharedMemory::OptimalCommunicator    (processors,optimal_comm);
   InitFromMPICommunicator(processors,optimal_comm);
   SetCommunicator(optimal_comm);
   ///////////////////////////////////////////////////
@@ -124,13 +124,12 @@ CartesianCommunicator::CartesianCommunicator(const Coordinate &processors,const
   int parent_ndimension = parent._ndimension; assert(_ndimension >= parent._ndimension);
   Coordinate parent_processor_coor(_ndimension,0);
   Coordinate parent_processors    (_ndimension,1);
-  Coordinate shm_processors       (_ndimension,1);
+
   // Can make 5d grid from 4d etc...
   int pad = _ndimension-parent_ndimension;
   for(int d=0;d<parent_ndimension;d++){
     parent_processor_coor[pad+d]=parent._processor_coor[d];
     parent_processors    [pad+d]=parent._processors[d];
-    shm_processors       [pad+d]=parent._shm_processors[d];
   }
 
   //////////////////////////////////////////////////////////////////////////////////////////////////////
@@ -155,7 +154,6 @@ CartesianCommunicator::CartesianCommunicator(const Coordinate &processors,const
     ccoor[d] = parent_processor_coor[d] % processors[d];
     scoor[d] = parent_processor_coor[d] / processors[d];
     ssize[d] = parent_processors[d]     / processors[d];
-    if ( processors[d] < shm_processors[d] ) shm_processors[d] = processors[d]; // subnode splitting.
   }
 
   // rank within subcomm ; srank is rank of subcomm within blocks of subcomms
@@ -337,23 +335,23 @@ void CartesianCommunicator::SendToRecvFrom(void *xmit,
 }
 // Basic Halo comms primitive
 double CartesianCommunicator::StencilSendToRecvFrom( void *xmit,
-						     int dest, int dox,
+						     int dest,
 						     void *recv,
-						     int from, int dor,
+						     int from,
 						     int bytes,int dir)
 {
   std::vector<CommsRequest_t> list;
-  double offbytes = StencilSendToRecvFromBegin(list,xmit,dest,dox,recv,from,dor,bytes,bytes,dir);
+  double offbytes = StencilSendToRecvFromBegin(list,xmit,dest,recv,from,bytes,dir);
   StencilSendToRecvFromComplete(list,dir);
   return offbytes;
 }
 
 double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list,
 							 void *xmit,
-							 int dest,int dox,
+							 int dest,
 							 void *recv,
-							 int from,int dor,
+							 int from,
-							 int xbytes,int rbytes,int dir)
+							 int bytes,int dir)
 {
   int ncomm  =communicator_halo.size();
   int commdir=dir%ncomm;
@@ -372,34 +370,39 @@ double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsReques
   double off_node_bytes=0.0;
   int tag;
 
-  if ( dor ) {
+  if ( (gfrom ==MPI_UNDEFINED) || Stencil_force_mpi ) {
-    if ( (gfrom ==MPI_UNDEFINED) || Stencil_force_mpi ) {
+    tag= dir+from*32;
-      tag= dir+from*32;
+    ierr=MPI_Irecv(recv, bytes, MPI_CHAR,from,tag,communicator_halo[commdir],&rrq);
-      ierr=MPI_Irecv(recv, rbytes, MPI_CHAR,from,tag,communicator_halo[commdir],&rrq);
+    assert(ierr==0);
-      assert(ierr==0);
+    list.push_back(rrq);
-      list.push_back(rrq);
+    off_node_bytes+=bytes;
-      off_node_bytes+=rbytes;
-    }
   }
-  
+
-  if (dox) {
+  if ( (gdest == MPI_UNDEFINED) || Stencil_force_mpi ) {
-    if ( (gdest == MPI_UNDEFINED) || Stencil_force_mpi ) {
+    tag= dir+_processor*32;
-      tag= dir+_processor*32;
+    ierr =MPI_Isend(xmit, bytes, MPI_CHAR,dest,tag,communicator_halo[commdir],&xrq);
-      ierr =MPI_Isend(xmit, xbytes, MPI_CHAR,dest,tag,communicator_halo[commdir],&xrq);
+    assert(ierr==0);
-      assert(ierr==0);
+    list.push_back(xrq);
-      list.push_back(xrq);
+    off_node_bytes+=bytes;
-      off_node_bytes+=xbytes;
+  } else {
-    } else {
+    // TODO : make a OMP loop on CPU, call threaded bcopy
-      void *shm = (void *) this->ShmBufferTranslate(dest,recv);
+    void *shm = (void *) this->ShmBufferTranslate(dest,recv);
-      assert(shm!=NULL);
+    assert(shm!=NULL);
-      acceleratorCopyDeviceToDeviceAsynch(xmit,shm,xbytes);
+    //    std::cout <<"acceleratorCopyDeviceToDeviceAsynch"<< std::endl;
-    }
+    acceleratorCopyDeviceToDeviceAsynch(xmit,shm,bytes);
+  }
+
+  if ( CommunicatorPolicy == CommunicatorPolicySequential ) {
+    this->StencilSendToRecvFromComplete(list,dir);
   }
 
   return off_node_bytes;
 }
 void CartesianCommunicator::StencilSendToRecvFromComplete(std::vector<CommsRequest_t> &list,int dir)
 {
+  //   std::cout << "Copy Synchronised\n"<<std::endl;
+  acceleratorCopySynchronise();
+
   int nreq=list.size();
 
   if (nreq==0) return;
@@ -435,10 +438,6 @@ int CartesianCommunicator::RankWorld(void){
   MPI_Comm_rank(communicator_world,&r);
   return r;
 }
-void CartesianCommunicator::BarrierWorld(void){
-  int ierr = MPI_Barrier(communicator_world);
-  assert(ierr==0);
-}
 void CartesianCommunicator::BroadcastWorld(int root,void* data, int bytes)
 {
   int ierr= MPI_Bcast(data,

Grid/communicator/Communicator_none.cc

@@ -45,14 +45,12 @@ void CartesianCommunicator::Init(int *argc, char *** arv)
 CartesianCommunicator::CartesianCommunicator(const Coordinate &processors,const CartesianCommunicator &parent,int &srank) 
   : CartesianCommunicator(processors) 
 {
-  _shm_processors = Coordinate(processors.size(),1);
   srank=0;
   SetCommunicator(communicator_world);
 }
 
 CartesianCommunicator::CartesianCommunicator(const Coordinate &processors)
 {
-  _shm_processors = Coordinate(processors.size(),1);
   _processors = processors;
   _ndimension = processors.size();  assert(_ndimension>=1);
   _processor_coor.resize(_ndimension);
@@ -104,7 +102,6 @@ int  CartesianCommunicator::RankWorld(void){return 0;}
 void CartesianCommunicator::Barrier(void){}
 void CartesianCommunicator::Broadcast(int root,void* data, int bytes) {}
 void CartesianCommunicator::BroadcastWorld(int root,void* data, int bytes) { }
-void CartesianCommunicator::BarrierWorld(void) { }
 int  CartesianCommunicator::RankFromProcessorCoor(Coordinate &coor) {  return 0;}
 void CartesianCommunicator::ProcessorCoorFromRank(int rank, Coordinate &coor){  coor = _processor_coor; }
 void CartesianCommunicator::ShiftedRanks(int dim,int shift,int &source,int &dest)
@@ -114,21 +111,21 @@ void CartesianCommunicator::ShiftedRanks(int dim,int shift,int &source,int &dest
 }
 
 double CartesianCommunicator::StencilSendToRecvFrom( void *xmit,
-						     int xmit_to_rank,int dox,
+						     int xmit_to_rank,
 						     void *recv,
-						     int recv_from_rank,int dor,
+						     int recv_from_rank,
 						     int bytes, int dir)
 {
   return 2.0*bytes;
 }
 double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list,
 							 void *xmit,
-							 int xmit_to_rank,int dox,
+							 int xmit_to_rank,
 							 void *recv,
-							 int recv_from_rank,int dor,
+							 int recv_from_rank,
-							 int xbytes,int rbytes, int dir)
+							 int bytes, int dir)
 {
-  return xbytes+rbytes;
+  return 2.0*bytes;
 }
 void CartesianCommunicator::StencilSendToRecvFromComplete(std::vector<CommsRequest_t> &waitall,int dir)
 {

Grid/communicator/SharedMemory.cc

@@ -91,59 +91,6 @@ 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

@@ -93,10 +93,9 @@ public:
   // Create an optimal reordered communicator that makes MPI_Cart_create get it right
   //////////////////////////////////////////////////////////////////////////////////////
   static void Init(Grid_MPI_Comm comm); // Typically MPI_COMM_WORLD
-  // Turns MPI_COMM_WORLD into right layout for Cartesian
+  static void OptimalCommunicator            (const Coordinate &processors,Grid_MPI_Comm & optimal_comm);  // Turns MPI_COMM_WORLD into right layout for Cartesian
-  static void OptimalCommunicator            (const Coordinate &processors,Grid_MPI_Comm & optimal_comm,Coordinate &ShmDims); 
+  static void OptimalCommunicatorHypercube   (const Coordinate &processors,Grid_MPI_Comm & optimal_comm);  // Turns MPI_COMM_WORLD into right layout for Cartesian
-  static void OptimalCommunicatorHypercube   (const Coordinate &processors,Grid_MPI_Comm & optimal_comm,Coordinate &ShmDims); 
+  static void OptimalCommunicatorSharedMemory(const Coordinate &processors,Grid_MPI_Comm & optimal_comm);  // Turns MPI_COMM_WORLD into right layout for Cartesian
-  static void OptimalCommunicatorSharedMemory(const Coordinate &processors,Grid_MPI_Comm & optimal_comm,Coordinate &ShmDims); 
   static void GetShmDims(const Coordinate &WorldDims,Coordinate &ShmDims);
   ///////////////////////////////////////////////////
   // Provide shared memory facilities off comm world

Grid/communicator/SharedMemoryMPI.cc

@@ -27,8 +27,6 @@ Author: Christoph Lehner <christoph@lhnr.de>
 *************************************************************************************/
 /*  END LEGAL */
 
-#define Mheader "SharedMemoryMpi: "
-
 #include <Grid/GridCore.h>
 #include <pwd.h>
 
@@ -38,120 +36,12 @@ Author: Christoph Lehner <christoph@lhnr.de>
 #ifdef GRID_HIP
 #include <hip/hip_runtime_api.h>
 #endif
-#ifdef GRID_SYCL
+#ifdef GRID_SYCl
-#define GRID_SYCL_LEVEL_ZERO_IPC
-#include <syscall.h>
-#define SHM_SOCKETS 
-#endif
 
-#include <sys/socket.h>
+#endif
-#include <sys/un.h>
 
 NAMESPACE_BEGIN(Grid); 
-
+#define header "SharedMemoryMpi: "
-#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
-
-
 /*Construct from an MPI communicator*/
 void GlobalSharedMemory::Init(Grid_MPI_Comm comm)
 {
@@ -174,8 +64,8 @@ void GlobalSharedMemory::Init(Grid_MPI_Comm comm)
   MPI_Comm_size(WorldShmComm     ,&WorldShmSize);
 
   if ( WorldRank == 0) {
-    std::cout << Mheader " World communicator of size " <<WorldSize << std::endl;  
+    std::cout << header " World communicator of size " <<WorldSize << std::endl;  
-    std::cout << Mheader " Node  communicator of size " <<WorldShmSize << std::endl;
+    std::cout << header " Node  communicator of size " <<WorldShmSize << std::endl;
   }
   // WorldShmComm, WorldShmSize, WorldShmRank
 
@@ -262,7 +152,7 @@ int Log2Size(int TwoToPower,int MAXLOG2)
   }
   return log2size;
 }
-void GlobalSharedMemory::OptimalCommunicator(const Coordinate &processors,Grid_MPI_Comm & optimal_comm,Coordinate &SHM)
+void GlobalSharedMemory::OptimalCommunicator(const Coordinate &processors,Grid_MPI_Comm & optimal_comm)
 {
   //////////////////////////////////////////////////////////////////////////////
   // Look and see if it looks like an HPE 8600 based on hostname conventions
@@ -275,11 +165,63 @@ void GlobalSharedMemory::OptimalCommunicator(const Coordinate &processors,Grid_M
   gethostname(name,namelen);
   int nscan = sscanf(name,"r%di%dn%d",&R,&I,&N) ;
 
-  if(nscan==3 && HPEhypercube ) OptimalCommunicatorHypercube(processors,optimal_comm,SHM);
+  if(nscan==3 && HPEhypercube ) OptimalCommunicatorHypercube(processors,optimal_comm);
-  else                          OptimalCommunicatorSharedMemory(processors,optimal_comm,SHM);
+  else                          OptimalCommunicatorSharedMemory(processors,optimal_comm);
 }
+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);
 
-void GlobalSharedMemory::OptimalCommunicatorHypercube(const Coordinate &processors,Grid_MPI_Comm & optimal_comm,Coordinate &SHM)
+  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)
 {
   ////////////////////////////////////////////////////////////////
   // Assert power of two shm_size.
@@ -352,8 +294,7 @@ void GlobalSharedMemory::OptimalCommunicatorHypercube(const Coordinate &processo
   Coordinate HyperCoor(ndimension);
 
   GetShmDims(WorldDims,ShmDims);
-  SHM = ShmDims;
+
-  
   ////////////////////////////////////////////////////////////////
   // Establish torus of processes and nodes with sub-blockings
   ////////////////////////////////////////////////////////////////
@@ -400,7 +341,7 @@ void GlobalSharedMemory::OptimalCommunicatorHypercube(const Coordinate &processo
   int ierr= MPI_Comm_split(WorldComm,0,rank,&optimal_comm);
   assert(ierr==0);
 }
-void GlobalSharedMemory::OptimalCommunicatorSharedMemory(const Coordinate &processors,Grid_MPI_Comm & optimal_comm,Coordinate &SHM)
+void GlobalSharedMemory::OptimalCommunicatorSharedMemory(const Coordinate &processors,Grid_MPI_Comm & optimal_comm)
 {
   ////////////////////////////////////////////////////////////////
   // Identify subblock of ranks on node spreading across dims
@@ -412,8 +353,6 @@ void GlobalSharedMemory::OptimalCommunicatorSharedMemory(const Coordinate &proce
   Coordinate ShmCoor(ndimension);    Coordinate NodeCoor(ndimension);   Coordinate WorldCoor(ndimension);
 
   GetShmDims(WorldDims,ShmDims);
-  SHM=ShmDims;
-
   ////////////////////////////////////////////////////////////////
   // Establish torus of processes and nodes with sub-blockings
   ////////////////////////////////////////////////////////////////
@@ -452,7 +391,7 @@ void GlobalSharedMemory::OptimalCommunicatorSharedMemory(const Coordinate &proce
 #ifdef GRID_MPI3_SHMGET
 void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 {
-  std::cout << Mheader "SharedMemoryAllocate "<< bytes<< " shmget implementation "<<std::endl;
+  std::cout << header "SharedMemoryAllocate "<< bytes<< " shmget implementation "<<std::endl;
   assert(_ShmSetup==1);
   assert(_ShmAlloc==0);
 
@@ -537,7 +476,7 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
     exit(EXIT_FAILURE);  
   }
 
-  std::cout << WorldRank << Mheader " SharedMemoryMPI.cc acceleratorAllocDevice "<< bytes 
+  std::cout << WorldRank << header " SharedMemoryMPI.cc acceleratorAllocDevice "<< bytes 
 	    << "bytes at "<< std::hex<< ShmCommBuf <<std::dec<<" for comms buffers " <<std::endl;
 
   SharedMemoryZero(ShmCommBuf,bytes);
@@ -580,21 +519,16 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
     exit(EXIT_FAILURE);  
   }
   if ( WorldRank == 0 ){
-    std::cout << WorldRank << Mheader " SharedMemoryMPI.cc acceleratorAllocDevice "<< bytes 
+    std::cout << WorldRank << header " SharedMemoryMPI.cc acceleratorAllocDevice "<< bytes 
-	      << "bytes at "<< std::hex<< ShmCommBuf << " - "<<(bytes-1+(uint64_t)ShmCommBuf) <<std::dec<<" for comms buffers " <<std::endl;
+	      << "bytes at "<< std::hex<< ShmCommBuf <<std::dec<<" for comms buffers " <<std::endl;
   }
   SharedMemoryZero(ShmCommBuf,bytes);
   std::cout<< "Setting up IPC"<<std::endl;
   ///////////////////////////////////////////////////////////////////////////////////////////////////////////
   // 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
@@ -602,32 +536,24 @@ 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 ; ze_ipc_mem_handle_t ze; } clone_mem_t;
+    typedef struct { int fd; pid_t pid ; } clone_mem_t;
 
-    auto zeDevice    = cl::sycl::get_native<cl::sycl::backend::ext_oneapi_level_zero>(theGridAccelerator->get_device());
+    auto zeDevice    = cl::sycl::get_native<cl::sycl::backend::level_zero>(theGridAccelerator->get_device());
-    auto zeContext   = cl::sycl::get_native<cl::sycl::backend::ext_oneapi_level_zero>(theGridAccelerator->get_context());
+    auto zeContext   = cl::sycl::get_native<cl::sycl::backend::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::cerr << "SharedMemoryMPI.cc zeMemGetIpcHandle failed for rank "<<r<<" "<<std::hex<<err<<std::dec<<std::endl;
+	std::cout << "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
@@ -655,7 +581,6 @@ 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,
@@ -671,10 +596,6 @@ 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;
@@ -682,22 +603,16 @@ 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);
-      myfd  = syscall(438,pidfd,handle.fd,0);
+      int myfd  = syscall(438,pidfd,handle.fd,0);
-      int err_t = errno;
+
-      if (myfd < 0) {
+      std::cout<<"Using IpcHandle myfd "<<myfd<<"\n";
-        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::cerr << "SharedMemoryMPI.cc "<<zeContext<<" "<<zeDevice<<std::endl;
+	std::cout << "SharedMemoryMPI.cc "<<zeContext<<" "<<zeDevice<<std::endl;
-	std::cerr << "SharedMemoryMPI.cc zeMemOpenIpcHandle failed for rank "<<r<<" "<<std::hex<<err<<std::dec<<std::endl; 
+	std::cout << "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;
@@ -732,7 +647,6 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 #else
     WorldShmCommBufs[r] = ShmCommBuf;
 #endif
-    MPI_Barrier(WorldShmComm);
   }
 
   _ShmAllocBytes=bytes;
@@ -744,7 +658,7 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 #ifdef GRID_MPI3_SHMMMAP
 void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 {
-  std::cout << Mheader "SharedMemoryAllocate "<< bytes<< " MMAP implementation "<< GRID_SHM_PATH <<std::endl;
+  std::cout << header "SharedMemoryAllocate "<< bytes<< " MMAP implementation "<< GRID_SHM_PATH <<std::endl;
   assert(_ShmSetup==1);
   assert(_ShmAlloc==0);
   //////////////////////////////////////////////////////////////////////////////////////////////////////////
@@ -781,7 +695,7 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
     assert(((uint64_t)ptr&0x3F)==0);
     close(fd);
     WorldShmCommBufs[r] =ptr;
-    //    std::cout << Mheader "Set WorldShmCommBufs["<<r<<"]="<<ptr<< "("<< bytes<< "bytes)"<<std::endl;
+    //    std::cout << header "Set WorldShmCommBufs["<<r<<"]="<<ptr<< "("<< bytes<< "bytes)"<<std::endl;
   }
   _ShmAlloc=1;
   _ShmAllocBytes  = bytes;
@@ -791,7 +705,7 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 #ifdef GRID_MPI3_SHM_NONE
 void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 {
-  std::cout << Mheader "SharedMemoryAllocate "<< bytes<< " MMAP anonymous implementation "<<std::endl;
+  std::cout << header "SharedMemoryAllocate "<< bytes<< " MMAP anonymous implementation "<<std::endl;
   assert(_ShmSetup==1);
   assert(_ShmAlloc==0);
   //////////////////////////////////////////////////////////////////////////////////////////////////////////
@@ -838,7 +752,7 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 ////////////////////////////////////////////////////////////////////////////////////////////
 void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 { 
-  std::cout << Mheader "SharedMemoryAllocate "<< bytes<< " SHMOPEN implementation "<<std::endl;
+  std::cout << header "SharedMemoryAllocate "<< bytes<< " SHMOPEN implementation "<<std::endl;
   assert(_ShmSetup==1);
   assert(_ShmAlloc==0); 
   MPI_Barrier(WorldShmComm);

Grid/communicator/SharedMemoryNone.cc

@@ -48,10 +48,9 @@ void GlobalSharedMemory::Init(Grid_MPI_Comm comm)
   _ShmSetup=1;
 }
 
-void GlobalSharedMemory::OptimalCommunicator(const Coordinate &processors,Grid_MPI_Comm & optimal_comm,Coordinate &SHM)
+void GlobalSharedMemory::OptimalCommunicator(const Coordinate &processors,Grid_MPI_Comm & optimal_comm)
 {
   optimal_comm = WorldComm;
-  SHM = Coordinate(processors.size(),1);
 }
 
 ////////////////////////////////////////////////////////////////////////////////////////////

Grid/cshift/Cshift_common.h

@@ -29,27 +29,8 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 
 NAMESPACE_BEGIN(Grid);
 
-extern std::vector<std::pair<int,int> > Cshift_table; 
+extern 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 
 ///////////////////////////////////////////////////////////////////
@@ -93,8 +74,8 @@ Gather_plane_simple (const Lattice<vobj> &rhs,cshiftVector<vobj> &buffer,int dim
   }
   {
     auto buffer_p = & buffer[0];
-    auto table = MapCshiftTable();
+    auto table = &Cshift_table[0];
-#ifdef ACCELERATOR_CSHIFT
+#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]));
@@ -244,7 +225,7 @@ template<class vobj> void Scatter_plane_simple (Lattice<vobj> &rhs,cshiftVector<
   
   {
     auto buffer_p = & buffer[0];
-    auto table = MapCshiftTable();
+    auto table = &Cshift_table[0];
 #ifdef ACCELERATOR_CSHIFT    
     autoView( rhs_v, rhs, AcceleratorWrite);
     accelerator_for(i,ent,vobj::Nsimd(),{
@@ -359,7 +340,7 @@ template<class vobj> void Copy_plane(Lattice<vobj>& lhs,const Lattice<vobj> &rhs
   }
 
   {
-    auto table = MapCshiftTable();
+    auto table = &Cshift_table[0];
 #ifdef ACCELERATOR_CSHIFT    
     autoView(rhs_v , rhs, AcceleratorRead);
     autoView(lhs_v , lhs, AcceleratorWrite);
@@ -411,7 +392,7 @@ template<class vobj> void Copy_plane_permute(Lattice<vobj>& lhs,const Lattice<vo
   }
 
   {
-    auto table = MapCshiftTable();
+    auto table = &Cshift_table[0];
 #ifdef ACCELERATOR_CSHIFT    
     autoView( rhs_v, rhs, AcceleratorRead);
     autoView( lhs_v, lhs, AcceleratorWrite);

Grid/cshift/Cshift_mpi.h

@@ -52,8 +52,7 @@ 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
@@ -64,8 +63,6 @@ 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;
 }
 
@@ -130,20 +127,16 @@ 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;
@@ -151,39 +144,26 @@ template<class vobj> void Cshift_comms(Lattice<vobj> &ret,const Lattice<vobj> &r
 
       int bytes = words * sizeof(vobj);
 
-      tgather-=usecond();
       Gather_plane_simple (rhs,send_buf,dimension,sx,cbmask);
-      tgather+=usecond();
 
       //      int rank           = grid->_processor;
       int recv_from_rank;
       int xmit_to_rank;
       grid->ShiftedRanks(dimension,comm_proc,xmit_to_rank,recv_from_rank);
-      
+
-      tcomms-=usecond();
+      grid->Barrier();
-      //      grid->Barrier();
 
       grid->SendToRecvFrom((void *)&send_buf[0],
 			   xmit_to_rank,
 			   (void *)&recv_buf[0],
 			   recv_from_rank,
 			   bytes);
-      xbytes+=bytes;
-      //      grid->Barrier();
-      tcomms+=usecond();
 
-      tscatter-=usecond();
+      grid->Barrier();
+
       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)
@@ -210,12 +190,6 @@ 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);
 
   ///////////////////////////////////////////////
@@ -253,9 +227,7 @@ 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++){
       
@@ -280,8 +252,7 @@ template<class vobj> void  Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vo
       if(nbr_proc){
 	grid->ShiftedRanks(dimension,nbr_proc,xmit_to_rank,recv_from_rank); 
 
-	tcomms-=usecond();
+	grid->Barrier();
-	//	grid->Barrier();
 
 	send_buf_extract_mpi = &send_buf_extract[nbr_lane][0];
 	recv_buf_extract_mpi = &recv_buf_extract[i][0];
@@ -291,9 +262,7 @@ template<class vobj> void  Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vo
 			     recv_from_rank,
 			     bytes);
 
-	xbytes+=bytes;
+	grid->Barrier();
-	//	grid->Barrier();
-	tcomms+=usecond();
 
 	rpointers[i] = &recv_buf_extract[i][0];
       } else { 
@@ -301,17 +270,9 @@ 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)
@@ -331,11 +292,6 @@ 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);
@@ -359,9 +315,7 @@ 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 {
 
@@ -370,9 +324,7 @@ 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;
@@ -380,8 +332,7 @@ template<class vobj> void Cshift_comms(Lattice<vobj> &ret,const Lattice<vobj> &r
       grid->ShiftedRanks(dimension,comm_proc,xmit_to_rank,recv_from_rank);
 
 
-      tcomms-=usecond();
+      grid->Barrier();
-      //      grid->Barrier();
 
       acceleratorCopyDeviceToDevice((void *)&send_buf_v[0],(void *)&send_buf[0],bytes);
       grid->SendToRecvFrom((void *)&send_buf[0],
@@ -389,24 +340,13 @@ 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)
@@ -432,11 +372,6 @@ 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);
 
@@ -479,10 +414,8 @@ 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++){
       
@@ -507,8 +440,7 @@ template<class vobj> void  Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vo
       if(nbr_proc){
 	grid->ShiftedRanks(dimension,nbr_proc,xmit_to_rank,recv_from_rank); 
 
-	tcomms-=usecond();
+	grid->Barrier();
-	//	grid->Barrier();
 
 	acceleratorCopyDeviceToDevice((void *)&send_buf_extract[nbr_lane][0],(void *)send_buf_extract_mpi,bytes);
 	grid->SendToRecvFrom((void *)send_buf_extract_mpi,
@@ -517,28 +449,17 @@ 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

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

Grid/lattice/Lattice.h

@@ -35,7 +35,6 @@ 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>
@@ -47,4 +46,4 @@ 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/PaddedCell.h>
+#include <Grid/lattice/Lattice_crc.h>

Grid/lattice/Lattice_ET.h

@@ -63,7 +63,7 @@ accelerator_inline vobj predicatedWhere(const iobj &predicate,
   typename std::remove_const<vobj>::type ret;
 
   typedef typename vobj::scalar_object scalar_object;
-  //  typedef typename vobj::scalar_type scalar_type;
+  typedef typename vobj::scalar_type scalar_type;
   typedef typename vobj::vector_type vector_type;
 
   const int Nsimd = vobj::vector_type::Nsimd();
@@ -345,9 +345,7 @@ 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));
@@ -458,9 +456,7 @@ 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

@@ -36,7 +36,6 @@ NAMESPACE_BEGIN(Grid);
 //////////////////////////////////////////////////////////////////////////////////////////////////////
 template<class obj1,class obj2,class obj3> inline
 void mult(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const Lattice<obj3> &rhs){
-  GRID_TRACE("mult");
   ret.Checkerboard() = lhs.Checkerboard();
   autoView( ret_v , ret, AcceleratorWrite);
   autoView( lhs_v , lhs, AcceleratorRead);
@@ -54,7 +53,6 @@ void mult(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const Lattice<obj3> &rhs){
   
 template<class obj1,class obj2,class obj3> inline
 void mac(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const Lattice<obj3> &rhs){
-  GRID_TRACE("mac");
   ret.Checkerboard() = lhs.Checkerboard();
   conformable(ret,rhs);
   conformable(lhs,rhs);
@@ -72,7 +70,6 @@ void mac(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const Lattice<obj3> &rhs){
   
 template<class obj1,class obj2,class obj3> inline
 void sub(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const Lattice<obj3> &rhs){
-  GRID_TRACE("sub");
   ret.Checkerboard() = lhs.Checkerboard();
   conformable(ret,rhs);
   conformable(lhs,rhs);
@@ -89,7 +86,6 @@ void sub(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const Lattice<obj3> &rhs){
 }
 template<class obj1,class obj2,class obj3> inline
 void add(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const Lattice<obj3> &rhs){
-  GRID_TRACE("add");
   ret.Checkerboard() = lhs.Checkerboard();
   conformable(ret,rhs);
   conformable(lhs,rhs);
@@ -110,7 +106,6 @@ void add(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const Lattice<obj3> &rhs){
 //////////////////////////////////////////////////////////////////////////////////////////////////////
 template<class obj1,class obj2,class obj3> inline
 void mult(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const obj3 &rhs){
-  GRID_TRACE("mult");
   ret.Checkerboard() = lhs.Checkerboard();
   conformable(lhs,ret);
   autoView( ret_v , ret, AcceleratorWrite);
@@ -124,7 +119,6 @@ void mult(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const obj3 &rhs){
   
 template<class obj1,class obj2,class obj3> inline
 void mac(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const obj3 &rhs){
-  GRID_TRACE("mac");
   ret.Checkerboard() = lhs.Checkerboard();
   conformable(ret,lhs);
   autoView( ret_v , ret, AcceleratorWrite);
@@ -139,7 +133,6 @@ void mac(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const obj3 &rhs){
   
 template<class obj1,class obj2,class obj3> inline
 void sub(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const obj3 &rhs){
-  GRID_TRACE("sub");
   ret.Checkerboard() = lhs.Checkerboard();
   conformable(ret,lhs);
   autoView( ret_v , ret, AcceleratorWrite);
@@ -153,7 +146,6 @@ void sub(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const obj3 &rhs){
 }
 template<class obj1,class obj2,class obj3> inline
 void add(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const obj3 &rhs){
-  GRID_TRACE("add");
   ret.Checkerboard() = lhs.Checkerboard();
   conformable(lhs,ret);
   autoView( ret_v , ret, AcceleratorWrite);
@@ -171,7 +163,6 @@ void add(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const obj3 &rhs){
 //////////////////////////////////////////////////////////////////////////////////////////////////////
 template<class obj1,class obj2,class obj3> inline
 void mult(Lattice<obj1> &ret,const obj2 &lhs,const Lattice<obj3> &rhs){
-  GRID_TRACE("mult");
   ret.Checkerboard() = rhs.Checkerboard();
   conformable(ret,rhs);
   autoView( ret_v , ret, AcceleratorWrite);
@@ -186,7 +177,6 @@ void mult(Lattice<obj1> &ret,const obj2 &lhs,const Lattice<obj3> &rhs){
   
 template<class obj1,class obj2,class obj3> inline
 void mac(Lattice<obj1> &ret,const obj2 &lhs,const Lattice<obj3> &rhs){
-  GRID_TRACE("mac");
   ret.Checkerboard() = rhs.Checkerboard();
   conformable(ret,rhs);
   autoView( ret_v , ret, AcceleratorWrite);
@@ -201,7 +191,6 @@ void mac(Lattice<obj1> &ret,const obj2 &lhs,const Lattice<obj3> &rhs){
   
 template<class obj1,class obj2,class obj3> inline
 void sub(Lattice<obj1> &ret,const obj2 &lhs,const Lattice<obj3> &rhs){
-  GRID_TRACE("sub");
   ret.Checkerboard() = rhs.Checkerboard();
   conformable(ret,rhs);
   autoView( ret_v , ret, AcceleratorWrite);
@@ -215,7 +204,6 @@ void sub(Lattice<obj1> &ret,const obj2 &lhs,const Lattice<obj3> &rhs){
 }
 template<class obj1,class obj2,class obj3> inline
 void add(Lattice<obj1> &ret,const obj2 &lhs,const Lattice<obj3> &rhs){
-  GRID_TRACE("add");
   ret.Checkerboard() = rhs.Checkerboard();
   conformable(ret,rhs);
   autoView( ret_v , ret, AcceleratorWrite);
@@ -230,7 +218,6 @@ void add(Lattice<obj1> &ret,const obj2 &lhs,const Lattice<obj3> &rhs){
   
 template<class sobj,class vobj> inline
 void axpy(Lattice<vobj> &ret,sobj a,const Lattice<vobj> &x,const Lattice<vobj> &y){
-  GRID_TRACE("axpy");
   ret.Checkerboard() = x.Checkerboard();
   conformable(ret,x);
   conformable(x,y);
@@ -244,7 +231,6 @@ void axpy(Lattice<vobj> &ret,sobj a,const Lattice<vobj> &x,const Lattice<vobj> &
 }
 template<class sobj,class vobj> inline
 void axpby(Lattice<vobj> &ret,sobj a,sobj b,const Lattice<vobj> &x,const Lattice<vobj> &y){
-  GRID_TRACE("axpby");
   ret.Checkerboard() = x.Checkerboard();
   conformable(ret,x);
   conformable(x,y);
@@ -260,52 +246,13 @@ void axpby(Lattice<vobj> &ret,sobj a,sobj b,const Lattice<vobj> &x,const Lattice
 template<class sobj,class vobj> inline
 RealD axpy_norm(Lattice<vobj> &ret,sobj a,const Lattice<vobj> &x,const Lattice<vobj> &y)
 {
-  GRID_TRACE("axpy_norm");
     return axpy_norm_fast(ret,a,x,y);
 }
 template<class sobj,class vobj> inline
 RealD axpby_norm(Lattice<vobj> &ret,sobj a,sobj b,const Lattice<vobj> &x,const Lattice<vobj> &y)
 {
-  GRID_TRACE("axpby_norm");
     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

@@ -117,7 +117,6 @@ public:
   ////////////////////////////////////////////////////////////////////////////////
   template <typename Op, typename T1> inline Lattice<vobj> & operator=(const LatticeUnaryExpression<Op,T1> &expr)
   {
-    GRID_TRACE("ExpressionTemplateEval");
     GridBase *egrid(nullptr);
     GridFromExpression(egrid,expr);
     assert(egrid!=nullptr);
@@ -141,7 +140,6 @@ public:
   }
   template <typename Op, typename T1,typename T2> inline Lattice<vobj> & operator=(const LatticeBinaryExpression<Op,T1,T2> &expr)
   {
-    GRID_TRACE("ExpressionTemplateEval");
     GridBase *egrid(nullptr);
     GridFromExpression(egrid,expr);
     assert(egrid!=nullptr);
@@ -165,7 +163,6 @@ public:
   }
   template <typename Op, typename T1,typename T2,typename T3> inline Lattice<vobj> & operator=(const LatticeTrinaryExpression<Op,T1,T2,T3> &expr)
   {
-    GRID_TRACE("ExpressionTemplateEval");
     GridBase *egrid(nullptr);
     GridFromExpression(egrid,expr);
     assert(egrid!=nullptr);
@@ -291,8 +288,8 @@ public:
     typename std::enable_if<!std::is_same<robj,vobj>::value,int>::type i=0;
     conformable(*this,r);
     this->checkerboard = r.Checkerboard();
-    auto him= r.View(AcceleratorRead);
     auto me =   View(AcceleratorWriteDiscard);
+    auto him= r.View(AcceleratorRead);
     accelerator_for(ss,me.size(),vobj::Nsimd(),{
       coalescedWrite(me[ss],him(ss));
     });
@@ -306,8 +303,8 @@ public:
   inline Lattice<vobj> & operator = (const Lattice<vobj> & r){
     this->checkerboard = r.Checkerboard();
     conformable(*this,r);
-    auto him= r.View(AcceleratorRead);
     auto me =   View(AcceleratorWriteDiscard);
+    auto him= r.View(AcceleratorRead);
     accelerator_for(ss,me.size(),vobj::Nsimd(),{
       coalescedWrite(me[ss],him(ss));
     });

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) || defined(GRID_HIP) || defined(GRID_SYCL)) )
+#if ( (!defined(GRID_CUDA)) )
   int max_threads = thread_max();
   Vector < vobj > Bt(Nm * max_threads);
   thread_region

Grid/lattice/Lattice_crc.h

@@ -29,26 +29,13 @@ 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> uint32_t crc(Lattice<vobj> & buf)
-{
-  auto ff = localNorm2(f);
-  if ( mu==-1 ) mu = f.Grid()->Nd()-1;
-  typedef typename vobj::tensor_reduced normtype;
-  typedef typename normtype::scalar_object scalar;
-  std::vector<scalar> sff;
-  sliceSum(ff,sff,mu);
-  for(int t=0;t<sff.size();t++){
-    std::cout << s<<" "<<t<<" "<<sff[t]<<std::endl;
-  }
-}
-
-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::cerr << "FingerPrint "<<__FILE__ <<" "<< __LINE__ <<" "<< #U <<" "<<crc(U)<<std::endl;
+#define CRC(U) std::cout << "FingerPrint "<<__FILE__ <<" "<< __LINE__ <<" "<< #U <<" "<<crc(U)<<std::endl;
 
 NAMESPACE_END(Grid);
 

Grid/lattice/Lattice_matrix_reduction.h

@@ -32,6 +32,7 @@ template<class vobj>
 static void sliceMaddMatrix (Lattice<vobj> &R,Eigen::MatrixXcd &aa,const Lattice<vobj> &X,const Lattice<vobj> &Y,int Orthog,RealD scale=1.0) 
 {    
   typedef typename vobj::scalar_object sobj;
+  typedef typename vobj::scalar_type scalar_type;
   typedef typename vobj::vector_type vector_type;
 
   int Nblock = X.Grid()->GlobalDimensions()[Orthog];
@@ -81,6 +82,7 @@ template<class vobj>
 static void sliceMulMatrix (Lattice<vobj> &R,Eigen::MatrixXcd &aa,const Lattice<vobj> &X,int Orthog,RealD scale=1.0) 
 {    
   typedef typename vobj::scalar_object sobj;
+  typedef typename vobj::scalar_type scalar_type;
   typedef typename vobj::vector_type vector_type;
 
   int Nblock = X.Grid()->GlobalDimensions()[Orthog];
@@ -128,6 +130,7 @@ template<class vobj>
 static void sliceInnerProductMatrix(  Eigen::MatrixXcd &mat, const Lattice<vobj> &lhs,const Lattice<vobj> &rhs,int Orthog) 
 {
   typedef typename vobj::scalar_object sobj;
+  typedef typename vobj::scalar_type scalar_type;
   typedef typename vobj::vector_type vector_type;
   
   GridBase *FullGrid  = lhs.Grid();

Grid/lattice/Lattice_peekpoke.h

@@ -96,6 +96,9 @@ void pokeSite(const sobj &s,Lattice<vobj> &l,const Coordinate &site){
 
   GridBase *grid=l.Grid();
 
+  typedef typename vobj::scalar_type scalar_type;
+  typedef typename vobj::vector_type vector_type;
+
   int Nsimd = grid->Nsimd();
 
   assert( l.Checkerboard()== l.Grid()->CheckerBoard(site));
@@ -122,17 +125,14 @@ void pokeSite(const sobj &s,Lattice<vobj> &l,const Coordinate &site){
 //////////////////////////////////////////////////////////
 // Peek a scalar object from the SIMD array
 //////////////////////////////////////////////////////////
-template<class vobj>
-typename vobj::scalar_object peekSite(const Lattice<vobj> &l,const Coordinate &site){
-  typename vobj::scalar_object s;
-  peekSite(s,l,site);
-  return s;
-}        
 template<class vobj,class sobj>
 void peekSite(sobj &s,const Lattice<vobj> &l,const Coordinate &site){
         
   GridBase *grid=l.Grid();
 
+  typedef typename vobj::scalar_type scalar_type;
+  typedef typename vobj::vector_type vector_type;
+
   int Nsimd = grid->Nsimd();
 
   assert( l.Checkerboard() == l.Grid()->CheckerBoard(site));
@@ -173,11 +173,11 @@ inline void peekLocalSite(sobj &s,const LatticeView<vobj> &l,Coordinate &site)
   idx= grid->iIndex(site);
   odx= grid->oIndex(site);
   
-  const vector_type *vp = (const vector_type *) &l[odx];
+  scalar_type * vp = (scalar_type *)&l[odx];
   scalar_type * pt = (scalar_type *)&s;
       
   for(int w=0;w<words;w++){
-    pt[w] = getlane(vp[w],idx);
+    pt[w] = vp[idx+w*Nsimd];
   }
       
   return;
@@ -210,10 +210,10 @@ inline void pokeLocalSite(const sobj &s,LatticeView<vobj> &l,Coordinate &site)
   idx= grid->iIndex(site);
   odx= grid->oIndex(site);
 
-  vector_type * vp = (vector_type *)&l[odx];
+  scalar_type * vp = (scalar_type *)&l[odx];
   scalar_type * pt = (scalar_type *)&s;
   for(int w=0;w<words;w++){
-    putlane(vp[w],pt[w],idx);
+    vp[idx+w*Nsimd] = pt[w];
   }
   return;
 };

Grid/lattice/Lattice_reduction.h

@@ -28,10 +28,6 @@ Author: Christoph Lehner <christoph@lhnr.de>
 #if defined(GRID_CUDA)||defined(GRID_HIP)
 #include <Grid/lattice/Lattice_reduction_gpu.h>
 #endif
-#if defined(GRID_SYCL)
-#include <Grid/lattice/Lattice_reduction_sycl.h>
-#endif
-#include <Grid/lattice/Lattice_slicesum_core.h>
 
 NAMESPACE_BEGIN(Grid);
 
@@ -95,7 +91,10 @@ inline typename vobj::scalar_objectD sumD_cpu(const vobj *arg, Integer osites)
   for(int i=0;i<nthread;i++){
     ssum = ssum+sumarray[i];
   } 
-  return ssum;
+  
+  typedef typename vobj::scalar_object ssobj;
+  ssobj ret = ssum;
+  return ret;
 }
 /*
 Threaded max, don't use for now
@@ -128,7 +127,7 @@ inline Double max(const Double *arg, Integer osites)
 template<class vobj>
 inline typename vobj::scalar_object sum(const vobj *arg, Integer osites)
 {
-#if defined(GRID_CUDA)||defined(GRID_HIP)||defined(GRID_SYCL)
+#if defined(GRID_CUDA)||defined(GRID_HIP)
   return sum_gpu(arg,osites);
 #else
   return sum_cpu(arg,osites);
@@ -137,61 +136,25 @@ inline typename vobj::scalar_object sum(const vobj *arg, Integer osites)
 template<class vobj>
 inline typename vobj::scalar_objectD sumD(const vobj *arg, Integer osites)
 {
-#if defined(GRID_CUDA)||defined(GRID_HIP)||defined(GRID_SYCL)
+#if defined(GRID_CUDA)||defined(GRID_HIP)
   return sumD_gpu(arg,osites);
 #else
   return sumD_cpu(arg,osites);
 #endif  
 }
-template<class vobj>
-inline typename vobj::scalar_objectD sumD_large(const vobj *arg, Integer osites)
-{
-#if defined(GRID_CUDA)||defined(GRID_HIP)||defined(GRID_SYCL)
-  return sumD_gpu_large(arg,osites);
-#else
-  return sumD_cpu(arg,osites);
-#endif  
-}
-
-template<class vobj>
-inline typename vobj::scalar_object rankSum(const Lattice<vobj> &arg)
-{
-  Integer osites = arg.Grid()->oSites();
-#if defined(GRID_CUDA)||defined(GRID_HIP)||defined(GRID_SYCL)
-  autoView( arg_v, arg, AcceleratorRead);
-  return sum_gpu(&arg_v[0],osites);
-#else
-  autoView(arg_v, arg, CpuRead);
-  return sum_cpu(&arg_v[0],osites);
-#endif  
-}
 
 template<class vobj>
 inline typename vobj::scalar_object sum(const Lattice<vobj> &arg)
 {
-  auto ssum = rankSum(arg);
+#if defined(GRID_CUDA)||defined(GRID_HIP)
-  arg.Grid()->GlobalSum(ssum);
-  return ssum;
-}
-
-template<class vobj>
-inline typename vobj::scalar_object rankSumLarge(const Lattice<vobj> &arg)
-{
-#if defined(GRID_CUDA)||defined(GRID_HIP)||defined(GRID_SYCL)
   autoView( arg_v, arg, AcceleratorRead);
   Integer osites = arg.Grid()->oSites();
-  return sum_gpu_large(&arg_v[0],osites);
+  auto ssum= sum_gpu(&arg_v[0],osites);
 #else
   autoView(arg_v, arg, CpuRead);
   Integer osites = arg.Grid()->oSites();
-  return sum_cpu(&arg_v[0],osites);
+  auto ssum= sum_cpu(&arg_v[0],osites);
-#endif
+#endif  
-}
-
-template<class vobj>
-inline typename vobj::scalar_object sum_large(const Lattice<vobj> &arg)
-{
-  auto ssum = rankSumLarge(arg);
   arg.Grid()->GlobalSum(ssum);
   return ssum;
 }
@@ -234,6 +197,7 @@ template<class vobj> inline RealD maxLocalNorm2(const Lattice<vobj> &arg)
 template<class vobj>
 inline ComplexD rankInnerProduct(const Lattice<vobj> &left,const Lattice<vobj> &right)
 {
+  typedef typename vobj::scalar_type scalar_type;
   typedef typename vobj::vector_typeD vector_type;
   ComplexD  nrm;
   
@@ -243,55 +207,33 @@ inline ComplexD rankInnerProduct(const Lattice<vobj> &left,const Lattice<vobj> &
   const uint64_t sites = grid->oSites();
   
   // Might make all code paths go this way.
-#if 0
   typedef decltype(innerProductD(vobj(),vobj())) inner_t;
   Vector<inner_t> inner_tmp(sites);
   auto inner_tmp_v = &inner_tmp[0];
-  {
-    autoView( left_v , left, AcceleratorRead);
-    autoView( right_v,right, AcceleratorRead);
-    // This code could read coalesce
-    // GPU - SIMT lane compliance...
-    accelerator_for( ss, sites, nsimd,{
-	auto x_l = left_v(ss);
-	auto y_l = right_v(ss);
-	coalescedWrite(inner_tmp_v[ss],innerProductD(x_l,y_l));
-    });
-  }
-#else
-  typedef decltype(innerProduct(vobj(),vobj())) inner_t;
-  Vector<inner_t> inner_tmp(sites);
-  auto inner_tmp_v = &inner_tmp[0];
     
   {
     autoView( left_v , left, AcceleratorRead);
     autoView( right_v,right, AcceleratorRead);
 
     // GPU - SIMT lane compliance...
-    accelerator_for( ss, sites, nsimd,{
+    accelerator_for( ss, sites, 1,{
-	auto x_l = left_v(ss);
+	auto x_l = left_v[ss];
-	auto y_l = right_v(ss);
+	auto y_l = right_v[ss];
-	coalescedWrite(inner_tmp_v[ss],innerProduct(x_l,y_l));
+	inner_tmp_v[ss]=innerProductD(x_l,y_l);
     });
   }
-#endif
+
   // This is in single precision and fails some tests
-  auto anrm = sumD(inner_tmp_v,sites);  
+  auto anrm = sum(inner_tmp_v,sites);  
   nrm = anrm;
   return nrm;
 }
 
-
 template<class vobj>
 inline ComplexD innerProduct(const Lattice<vobj> &left,const Lattice<vobj> &right) {
   GridBase *grid = left.Grid();
-  uint32_t csum=0;
-  //  Uint32Checksum(left,csum);
   ComplexD nrm = rankInnerProduct(left,right);
-  RealD local = real(nrm);
-  GridNormLog(real(nrm),csum); // Could log before and after global sum to distinguish local and MPI
   grid->GlobalSum(nrm);
-  GridMPINormLog(local,real(nrm)); 
   return nrm;
 }
 
@@ -315,7 +257,8 @@ axpby_norm_fast(Lattice<vobj> &z,sobj a,sobj b,const Lattice<vobj> &x,const Latt
   conformable(z,x);
   conformable(x,y);
 
-  //  typedef typename vobj::vector_typeD vector_type;
+  typedef typename vobj::scalar_type scalar_type;
+  typedef typename vobj::vector_typeD vector_type;
   RealD  nrm;
   
   GridBase *grid = x.Grid();
@@ -327,29 +270,17 @@ axpby_norm_fast(Lattice<vobj> &z,sobj a,sobj b,const Lattice<vobj> &x,const Latt
   autoView( x_v, x, AcceleratorRead);
   autoView( y_v, y, AcceleratorRead);
   autoView( z_v, z, AcceleratorWrite);
-#if 0
+
   typedef decltype(innerProductD(x_v[0],y_v[0])) inner_t;
   Vector<inner_t> inner_tmp(sites);
   auto inner_tmp_v = &inner_tmp[0];
 
-  accelerator_for( ss, sites, nsimd,{
+  accelerator_for( ss, sites, 1,{
-      auto tmp = a*x_v(ss)+b*y_v(ss);
+      auto tmp = a*x_v[ss]+b*y_v[ss];
-      coalescedWrite(inner_tmp_v[ss],innerProductD(tmp,tmp));
+      inner_tmp_v[ss]=innerProductD(tmp,tmp);
-      coalescedWrite(z_v[ss],tmp);
+      z_v[ss]=tmp;
   });
   nrm = real(TensorRemove(sum(inner_tmp_v,sites)));
-#else
-  typedef decltype(innerProduct(x_v[0],y_v[0])) inner_t;
-  Vector<inner_t> inner_tmp(sites);
-  auto inner_tmp_v = &inner_tmp[0];
-
-  accelerator_for( ss, sites, nsimd,{
-      auto tmp = a*x_v(ss)+b*y_v(ss);
-      coalescedWrite(inner_tmp_v[ss],innerProduct(tmp,tmp));
-      coalescedWrite(z_v[ss],tmp);
-  });
-  nrm = real(TensorRemove(sumD(inner_tmp_v,sites)));
-#endif
   grid->GlobalSum(nrm);
   return nrm; 
 }
@@ -359,6 +290,7 @@ innerProductNorm(ComplexD& ip, RealD &nrm, const Lattice<vobj> &left,const Latti
 {
   conformable(left,right);
 
+  typedef typename vobj::scalar_type scalar_type;
   typedef typename vobj::vector_typeD vector_type;
   Vector<ComplexD> tmp(2);
 
@@ -455,10 +387,19 @@ template<class vobj> inline void sliceSum(const Lattice<vobj> &Data,std::vector<
   int e1=    grid->_slice_nblock[orthogdim];
   int e2=    grid->_slice_block [orthogdim];
   int stride=grid->_slice_stride[orthogdim];
-  int ostride=grid->_ostride[orthogdim];
+
-  
+  // sum over reduced dimension planes, breaking out orthog dir
-  //Reduce Data down to lvSum
+  // Parallel over orthog direction
-  sliceSumReduction(Data,lvSum,rd, e1,e2,stride,ostride,Nsimd);
+  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];
+      }
+    }
+  });
 
   // Sum across simd lanes in the plane, breaking out orthog dir.
   Coordinate icoor(Nd);
@@ -493,15 +434,6 @@ template<class vobj> inline void sliceSum(const Lattice<vobj> &Data,std::vector<
   int words = fd*sizeof(sobj)/sizeof(scalar_type);
   grid->GlobalSumVector(ptr, words);
 }
-template<class vobj> inline
-std::vector<typename vobj::scalar_object> 
-sliceSum(const Lattice<vobj> &Data,int orthogdim)
-{
-  std::vector<typename vobj::scalar_object> result;
-  sliceSum(Data,result,orthogdim);
-  return result;
-}
-
 
 template<class vobj>
 static void sliceInnerProductVector( std::vector<ComplexD> & result, const Lattice<vobj> &lhs,const Lattice<vobj> &rhs,int orthogdim) 
@@ -606,8 +538,7 @@ static void sliceNorm (std::vector<RealD> &sn,const Lattice<vobj> &rhs,int Ortho
 template<class vobj>
 static void sliceMaddVector(Lattice<vobj> &R,std::vector<RealD> &a,const Lattice<vobj> &X,const Lattice<vobj> &Y,
 			    int orthogdim,RealD scale=1.0) 
-{
+{    
-  // perhaps easier to just promote A to a field and use regular madd
   typedef typename vobj::scalar_object sobj;
   typedef typename vobj::scalar_type scalar_type;
   typedef typename vobj::vector_type vector_type;
@@ -638,7 +569,8 @@ static void sliceMaddVector(Lattice<vobj> &R,std::vector<RealD> &a,const Lattice
     for(int l=0;l<Nsimd;l++){
       grid->iCoorFromIindex(icoor,l);
       int ldx =r+icoor[orthogdim]*rd;
-      av.putlane(scalar_type(a[ldx])*zscale,l);
+      scalar_type *as =(scalar_type *)&av;
+      as[l] = scalar_type(a[ldx])*zscale;
     }
 
     tensor_reduced at; at=av;
@@ -678,6 +610,7 @@ template<class vobj>
 static void sliceMaddMatrix (Lattice<vobj> &R,Eigen::MatrixXcd &aa,const Lattice<vobj> &X,const Lattice<vobj> &Y,int Orthog,RealD scale=1.0) 
 {    
   typedef typename vobj::scalar_object sobj;
+  typedef typename vobj::scalar_type scalar_type;
   typedef typename vobj::vector_type vector_type;
 
   int Nblock = X.Grid()->GlobalDimensions()[Orthog];
@@ -731,6 +664,7 @@ template<class vobj>
 static void sliceMulMatrix (Lattice<vobj> &R,Eigen::MatrixXcd &aa,const Lattice<vobj> &X,int Orthog,RealD scale=1.0) 
 {    
   typedef typename vobj::scalar_object sobj;
+  typedef typename vobj::scalar_type scalar_type;
   typedef typename vobj::vector_type vector_type;
 
   int Nblock = X.Grid()->GlobalDimensions()[Orthog];
@@ -784,6 +718,7 @@ template<class vobj>
 static void sliceInnerProductMatrix(  Eigen::MatrixXcd &mat, const Lattice<vobj> &lhs,const Lattice<vobj> &rhs,int Orthog) 
 {
   typedef typename vobj::scalar_object sobj;
+  typedef typename vobj::scalar_type scalar_type;
   typedef typename vobj::vector_type vector_type;
   
   GridBase *FullGrid  = lhs.Grid();

Grid/lattice/Lattice_reduction_gpu.h

@@ -23,27 +23,27 @@ unsigned int nextPow2(Iterator x) {
 }
 
 template <class Iterator>
-int getNumBlocksAndThreads(const Iterator n, const size_t sizeofsobj, Iterator &threads, Iterator &blocks) {
+void getNumBlocksAndThreads(const Iterator n, const size_t sizeofsobj, Iterator &threads, Iterator &blocks) {
   
   int device;
 #ifdef GRID_CUDA
   cudaGetDevice(&device);
 #endif
 #ifdef GRID_HIP
-  auto r=hipGetDevice(&device);
+  hipGetDevice(&device);
 #endif
   
   Iterator warpSize            = gpu_props[device].warpSize;
   Iterator sharedMemPerBlock   = gpu_props[device].sharedMemPerBlock;
   Iterator maxThreadsPerBlock  = gpu_props[device].maxThreadsPerBlock;
   Iterator multiProcessorCount = gpu_props[device].multiProcessorCount;
-  /*  
+  
   std::cout << GridLogDebug << "GPU has:" << std::endl;
   std::cout << GridLogDebug << "\twarpSize            = " << warpSize << std::endl;
   std::cout << GridLogDebug << "\tsharedMemPerBlock   = " << sharedMemPerBlock << std::endl;
   std::cout << GridLogDebug << "\tmaxThreadsPerBlock  = " << maxThreadsPerBlock << std::endl;
   std::cout << GridLogDebug << "\tmultiProcessorCount = " << multiProcessorCount << std::endl;
-  */  
+  
   if (warpSize != WARP_SIZE) {
     std::cout << GridLogError << "The warp size of the GPU in use does not match the warp size set when compiling Grid." << std::endl;
     exit(EXIT_FAILURE);
@@ -53,12 +53,12 @@ int getNumBlocksAndThreads(const Iterator n, const size_t sizeofsobj, Iterator &
   threads = warpSize;
   if ( threads*sizeofsobj > sharedMemPerBlock ) {
     std::cout << GridLogError << "The object is too large for the shared memory." << std::endl;
-    return 0;
+    exit(EXIT_FAILURE);
   }
   while( 2*threads*sizeofsobj < sharedMemPerBlock && 2*threads <= maxThreadsPerBlock ) threads *= 2;
   // keep all the streaming multiprocessors busy
   blocks = nextPow2(multiProcessorCount);
-  return 1;
+  
 }
 
 template <class sobj, class Iterator>
@@ -198,7 +198,7 @@ __global__ void reduceKernel(const vobj *lat, sobj *buffer, Iterator n) {
 // Possibly promote to double and sum
 /////////////////////////////////////////////////////////////////////////////////////////////////////////
 template <class vobj>
-inline typename vobj::scalar_objectD sumD_gpu_small(const vobj *lat, Integer osites) 
+inline typename vobj::scalar_objectD sumD_gpu(const vobj *lat, Integer osites) 
 {
   typedef typename vobj::scalar_objectD sobj;
   typedef decltype(lat) Iterator;
@@ -207,77 +207,17 @@ inline typename vobj::scalar_objectD sumD_gpu_small(const vobj *lat, Integer osi
   Integer size = osites*nsimd;
 
   Integer numThreads, numBlocks;
-  int ok = getNumBlocksAndThreads(size, sizeof(sobj), numThreads, numBlocks);
+  getNumBlocksAndThreads(size, sizeof(sobj), numThreads, numBlocks);
-  assert(ok);
-
   Integer smemSize = numThreads * sizeof(sobj);
-  // Move out of UVM
+
-  // Turns out I had messed up the synchronise after move to compute stream
-  // as running this on the default stream fools the synchronise
-#undef UVM_BLOCK_BUFFER  
-#ifndef UVM_BLOCK_BUFFER  
-  commVector<sobj> buffer(numBlocks);
-  sobj *buffer_v = &buffer[0];
-  sobj result;
-  reduceKernel<<< numBlocks, numThreads, smemSize, computeStream >>>(lat, buffer_v, size);
-  accelerator_barrier();
-  acceleratorCopyFromDevice(buffer_v,&result,sizeof(result));
-#else
   Vector<sobj> buffer(numBlocks);
   sobj *buffer_v = &buffer[0];
-  sobj result;
+  
-  reduceKernel<<< numBlocks, numThreads, smemSize, computeStream >>>(lat, buffer_v, size);
+  reduceKernel<<< numBlocks, numThreads, smemSize >>>(lat, buffer_v, size);
   accelerator_barrier();
-  result = *buffer_v;
+  auto result = buffer_v[0];
-#endif
   return result;
 }
-
-template <class vobj>
-inline typename vobj::scalar_objectD sumD_gpu_large(const vobj *lat, Integer osites)
-{
-  typedef typename vobj::vector_type  vector;
-  typedef typename vobj::scalar_typeD scalarD;
-  typedef typename vobj::scalar_objectD sobj;
-  sobj ret;
-  scalarD *ret_p = (scalarD *)&ret;
-  
-  const int words = sizeof(vobj)/sizeof(vector);
-
-  Vector<vector> buffer(osites);
-  vector *dat = (vector *)lat;
-  vector *buf = &buffer[0];
-  iScalar<vector> *tbuf =(iScalar<vector> *)  &buffer[0];
-  for(int w=0;w<words;w++) {
-
-    accelerator_for(ss,osites,1,{
-	buf[ss] = dat[ss*words+w];
-      });
-      
-    ret_p[w] = sumD_gpu_small(tbuf,osites);
-  }
-  return ret;
-}
-
-template <class vobj>
-inline typename vobj::scalar_objectD sumD_gpu(const vobj *lat, Integer osites)
-{
-  typedef typename vobj::scalar_objectD sobj;
-  sobj ret;
-  
-  Integer nsimd= vobj::Nsimd();
-  Integer size = osites*nsimd;
-  Integer numThreads, numBlocks;
-  int ok = getNumBlocksAndThreads(size, sizeof(sobj), numThreads, numBlocks);
-  
-  if ( ok ) {
-    ret = sumD_gpu_small(lat,osites);
-  } else {
-    ret = sumD_gpu_large(lat,osites);
-  }
-  return ret;
-}
-
 /////////////////////////////////////////////////////////////////////////////////////////////////////////
 // Return as same precision as input performing reduction in double precision though
 /////////////////////////////////////////////////////////////////////////////////////////////////////////
@@ -290,13 +230,6 @@ inline typename vobj::scalar_object sum_gpu(const vobj *lat, Integer osites)
   return result;
 }
 
-template <class vobj>
+
-inline typename vobj::scalar_object sum_gpu_large(const vobj *lat, Integer osites)
-{
-  typedef typename vobj::scalar_object sobj;
-  sobj result;
-  result = sumD_gpu_large(lat,osites);
-  return result;
-}
 
 NAMESPACE_END(Grid);

Grid/lattice/Lattice_reduction_sycl.h

@@ -1,125 +0,0 @@
-NAMESPACE_BEGIN(Grid);
-
-/////////////////////////////////////////////////////////////////////////////////////////////////////////
-// Possibly promote to double and sum
-/////////////////////////////////////////////////////////////////////////////////////////////////////////
-
-template <class vobj>
-inline typename vobj::scalar_objectD sumD_gpu_tensor(const vobj *lat, Integer osites) 
-{
-  typedef typename vobj::scalar_object sobj;
-  typedef typename vobj::scalar_objectD sobjD;
-  sobj *mysum =(sobj *) malloc_shared(sizeof(sobj),*theGridAccelerator);
-  sobj identity; zeroit(identity);
-  sobj ret ; 
-
-  Integer nsimd= vobj::Nsimd();
-  
-  theGridAccelerator->submit([&](cl::sycl::handler &cgh) {
-     auto Reduction = cl::sycl::reduction(mysum,identity,std::plus<>());
-     cgh.parallel_for(cl::sycl::range<1>{osites},
-		      Reduction,
-		      [=] (cl::sycl::id<1> item, auto &sum) {
-      auto osite   = item[0];
-      sum +=Reduce(lat[osite]);
-     });
-   });
-  theGridAccelerator->wait();
-  ret = mysum[0];
-  free(mysum,*theGridAccelerator);
-  sobjD dret; convertType(dret,ret);
-  return dret;
-}
-
-template <class vobj>
-inline typename vobj::scalar_objectD sumD_gpu_large(const vobj *lat, Integer osites)
-{
-  return sumD_gpu_tensor(lat,osites);
-}
-template <class vobj>
-inline typename vobj::scalar_objectD sumD_gpu_small(const vobj *lat, Integer osites)
-{
-  return sumD_gpu_large(lat,osites);
-}
-
-template <class vobj>
-inline typename vobj::scalar_objectD sumD_gpu(const vobj *lat, Integer osites)
-{
-  return sumD_gpu_large(lat,osites);
-}
-
-/////////////////////////////////////////////////////////////////////////////////////////////////////////
-// Return as same precision as input performing reduction in double precision though
-/////////////////////////////////////////////////////////////////////////////////////////////////////////
-template <class vobj>
-inline typename vobj::scalar_object sum_gpu(const vobj *lat, Integer osites) 
-{
-  typedef typename vobj::scalar_object sobj;
-  sobj result;
-  result = sumD_gpu(lat,osites);
-  return result;
-}
-
-template <class vobj>
-inline typename vobj::scalar_object sum_gpu_large(const vobj *lat, Integer osites)
-{
-  typedef typename vobj::scalar_object sobj;
-  sobj result;
-  result = sumD_gpu_large(lat,osites);
-  return result;
-}
-
-NAMESPACE_END(Grid);
-
-/*
-template<class Double> Double svm_reduce(Double *vec,uint64_t L)
-{
-  Double sumResult; zeroit(sumResult);
-  Double *d_sum =(Double *)cl::sycl::malloc_shared(sizeof(Double),*theGridAccelerator);
-  Double identity;  zeroit(identity);
-  theGridAccelerator->submit([&](cl::sycl::handler &cgh) {
-     auto Reduction = cl::sycl::reduction(d_sum,identity,std::plus<>());
-     cgh.parallel_for(cl::sycl::range<1>{L},
-		      Reduction,
-		      [=] (cl::sycl::id<1> index, auto &sum) {
-	 sum +=vec[index];
-     });
-   });
-  theGridAccelerator->wait();
-  Double ret = d_sum[0];
-  free(d_sum,*theGridAccelerator);
-  std::cout << " svm_reduce finished "<<L<<" sites sum = " << ret <<std::endl;
-  return ret;
-}
-
-template <class vobj>
-inline typename vobj::scalar_objectD sumD_gpu_repack(const vobj *lat, Integer osites)
-{
-  typedef typename vobj::vector_type  vector;
-  typedef typename vobj::scalar_type  scalar;
-
-  typedef typename vobj::scalar_typeD scalarD;
-  typedef typename vobj::scalar_objectD sobjD;
-
-  sobjD ret;
-  scalarD *ret_p = (scalarD *)&ret;
-  
-  const int nsimd = vobj::Nsimd();
-  const int words = sizeof(vobj)/sizeof(vector);
-
-  Vector<scalar> buffer(osites*nsimd);
-  scalar *buf = &buffer[0];
-  vector *dat = (vector *)lat;
-
-  for(int w=0;w<words;w++) {
-
-    accelerator_for(ss,osites,nsimd,{
-	int lane = acceleratorSIMTlane(nsimd);
-	buf[ss*nsimd+lane] = dat[ss*words+w].getlane(lane);
-    });
-    //Precision change at this point is to late to gain precision
-    ret_p[w] = svm_reduce(buf,nsimd*osites);
-  }
-  return ret;
-}
-*/

Grid/lattice/Lattice_rng.h

@@ -32,8 +32,9 @@
 #include <random>
 
 #ifdef RNG_SITMO
-#include <Grid/sitmo_rng/sitmo_prng_engine.hpp>
+#include <Grid/random/sitmo_prng_engine.hpp>
 #endif 
+#include <Grid/random/gaussian.h>
 
 #if defined(RNG_SITMO)
 #define RNG_FAST_DISCARD
@@ -142,8 +143,8 @@ public:
 
   std::vector<RngEngine>                             _generators;
   std::vector<std::uniform_real_distribution<RealD> > _uniform;
-  std::vector<std::normal_distribution<RealD> >       _gaussian;
+  std::vector<Grid::gaussian_distribution<RealD> >       _gaussian;
-  std::vector<std::discrete_distribution<int32_t> >   _bernoulli;
+  //  std::vector<std::discrete_distribution<int32_t> >   _bernoulli;
   std::vector<std::uniform_int_distribution<uint32_t> > _uid;
 
   ///////////////////////
@@ -152,7 +153,6 @@ 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.
@@ -163,9 +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;
@@ -180,9 +180,6 @@ 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
@@ -247,8 +244,8 @@ public:
   GridSerialRNG() : GridRNGbase() {
     _generators.resize(1);
     _uniform.resize(1,std::uniform_real_distribution<RealD>{0,1});
-    _gaussian.resize(1,std::normal_distribution<RealD>(0.0,1.0) );
+    _gaussian.resize(1,gaussian_distribution<RealD>(0.0,1.0) );
-    _bernoulli.resize(1,std::discrete_distribution<int32_t>{1,1});
+    //    _bernoulli.resize(1,std::discrete_distribution<int32_t>{1,1});
     _uid.resize(1,std::uniform_int_distribution<uint32_t>() );
   }
 
@@ -361,8 +358,8 @@ public:
 
     _generators.resize(_vol);
     _uniform.resize(_vol,std::uniform_real_distribution<RealD>{0,1});
-    _gaussian.resize(_vol,std::normal_distribution<RealD>(0.0,1.0) );
+    _gaussian.resize(_vol,gaussian_distribution<RealD>(0.0,1.0) );
-    _bernoulli.resize(_vol,std::discrete_distribution<int32_t>{1,1});
+    //    _bernoulli.resize(_vol,std::discrete_distribution<int32_t>{1,1});
     _uid.resize(_vol,std::uniform_int_distribution<uint32_t>() );
   }
 
@@ -411,7 +408,7 @@ public:
       std::cout << GridLogMessage << "Seed SHA256: " << GridChecksum::sha256_string(seeds) << std::endl;
       SeedFixedIntegers(seeds);
     }
-  void SeedFixedIntegers(const std::vector<int> &seeds, int britney=0){
+  void SeedFixedIntegers(const std::vector<int> &seeds){
 
     // Everyone generates the same seed_seq based on input seeds
     CartesianCommunicator::BroadcastWorld(0,(void *)&seeds[0],sizeof(int)*seeds.size());
@@ -428,29 +425,22 @@ public:
     // MT implementation does not implement fast discard even though
     // in principle this is possible
     ////////////////////////////////////////////////
-    thread_for( lidx, _grid->lSites(), {
 
-	int gidx;
+    // Everybody loops over global volume.
+    thread_for( gidx, _grid->_gsites, {
+	// Where is it?
+	int rank;
 	int o_idx;
 	int i_idx;
-	int rank;
+
-	Coordinate pcoor;
-	Coordinate lcoor;
 	Coordinate gcoor;
-	_grid->LocalIndexToLocalCoor(lidx,lcoor);
+	_grid->GlobalIndexToGlobalCoor(gidx,gcoor);
-	pcoor=_grid->ThisProcessorCoor();
-	_grid->ProcessorCoorLocalCoorToGlobalCoor(pcoor,lcoor,gcoor);
-	_grid->GlobalCoorToGlobalIndex(gcoor,gidx);
-
 	_grid->GlobalCoorToRankIndex(rank,o_idx,i_idx,gcoor);
-
-	assert(rank == _grid->ThisRank() );
 	
-	int l_idx=generator_idx(o_idx,i_idx);
+	// If this is one of mine we take it
-	_generators[l_idx] = master_engine;
+	if( rank == _grid->ThisRank() ){
-	if ( britney ) { 
+	  int l_idx=generator_idx(o_idx,i_idx);
-	  Skip(_generators[l_idx],l_idx); // Skip to next RNG sequence
+	  _generators[l_idx] = master_engine;
-	} else { 	
 	  Skip(_generators[l_idx],gidx); // Skip to next RNG sequence
 	}
     });
@@ -526,11 +516,11 @@ public:
 
 template <class vobj> inline void random(GridParallelRNG &rng,Lattice<vobj> &l)   { rng.fill(l,rng._uniform);  }
 template <class vobj> inline void gaussian(GridParallelRNG &rng,Lattice<vobj> &l) { rng.fill(l,rng._gaussian); }
-template <class vobj> inline void bernoulli(GridParallelRNG &rng,Lattice<vobj> &l){ rng.fill(l,rng._bernoulli);}
+//template <class vobj> inline void bernoulli(GridParallelRNG &rng,Lattice<vobj> &l){ rng.fill(l,rng._bernoulli);}
 
 template <class sobj> inline void random(GridSerialRNG &rng,sobj &l)   { rng.fill(l,rng._uniform  ); }
 template <class sobj> inline void gaussian(GridSerialRNG &rng,sobj &l) { rng.fill(l,rng._gaussian ); }
-template <class sobj> inline void bernoulli(GridSerialRNG &rng,sobj &l){ rng.fill(l,rng._bernoulli); }
+//template <class sobj> inline void bernoulli(GridSerialRNG &rng,sobj &l){ rng.fill(l,rng._bernoulli); }
 
 NAMESPACE_END(Grid);
 #endif

Grid/lattice/Lattice_slicesum_core.h

@@ -1,224 +0,0 @@
-#pragma once
-
-#if defined(GRID_CUDA)
-
-#include <cub/cub.cuh>
-#define gpucub cub
-#define gpuError_t cudaError_t
-#define gpuSuccess cudaSuccess
-
-#elif defined(GRID_HIP)
-
-#include <hipcub/hipcub.hpp>
-#define gpucub hipcub
-#define gpuError_t hipError_t
-#define gpuSuccess hipSuccess
-
-#endif
-
-
-NAMESPACE_BEGIN(Grid);
-
-
-#if defined(GRID_CUDA) || defined(GRID_HIP)
-template<class vobj> inline void sliceSumReduction_cub_small(const vobj *Data, 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);
-  
-
-}
-#endif 
-
-
-#if defined(GRID_SYCL)
-template<class vobj> inline void sliceSumReduction_sycl_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;
-
-  vobj *mysum = (vobj *) malloc_shared(rd*sizeof(vobj),*theGridAccelerator);
-  vobj vobj_zero;
-  zeroit(vobj_zero);
-  for (int r = 0; r<rd; r++) { 
-    mysum[r] = vobj_zero; 
-  }
-
-  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[ss]));
-
-  });
-
-  for (int r = 0; r < rd; r++) {
-      theGridAccelerator->submit([&](cl::sycl::handler &cgh) {
-          auto Reduction = cl::sycl::reduction(&mysum[r],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();
-  for (int r = 0; r < rd; r++) {
-    lvSum[r] = mysum[r];
-  }
-  free(mysum,*theGridAccelerator);
-}
-#endif
-
-template<class vobj> inline void sliceSumReduction_large(const vobj *Data, 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];
-    });
-
-    #if defined(GRID_CUDA) || defined(GRID_HIP)
-      sliceSumReduction_cub_small(buf,lvSum_small,rd,e1,e2,stride, ostride,Nsimd);
-    #elif defined(GRID_SYCL)
-      sliceSumReduction_sycl_small(buf,lvSum_small,rd,e1,e2,stride, ostride,Nsimd);
-    #endif
-
-    for (int r = 0; r < rd; r++) {
-      lvSum_ptr[w+words*r]=lvSum_small[r];
-    }
-
-  }
-
-  
-}
-
-template<class vobj> inline void sliceSumReduction_gpu(const Lattice<vobj> &Data, Vector<vobj> &lvSum, const int rd, const int e1, const int e2, const int stride, const int ostride, const int Nsimd)
-{
-  autoView(Data_v, Data, AcceleratorRead); //reduction libraries cannot deal with large vobjs so we split into small/large case.
-    if constexpr (sizeof(vobj) <= 256) { 
-
-      #if defined(GRID_CUDA) || defined(GRID_HIP)
-        sliceSumReduction_cub_small(&Data_v[0], lvSum, rd, e1, e2, stride, ostride, Nsimd);
-      #elif defined (GRID_SYCL)
-        sliceSumReduction_sycl_small(&Data_v[0], lvSum, rd, e1, e2, stride, ostride, Nsimd);
-      #endif
-
-    }
-    else {
-      sliceSumReduction_large(&Data_v[0], lvSum, rd, e1, e2, stride, ostride, Nsimd);
-    }
-}
-
-
-template<class vobj> inline void sliceSumReduction_cpu(const Lattice<vobj> &Data, 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) || defined(GRID_SYCL)
-  
-  sliceSumReduction_gpu(Data, lvSum, rd, e1, e2, stride, ostride, Nsimd);
-  
-  #else
-  sliceSumReduction_cpu(Data, lvSum, rd, e1, e2, stride, ostride, Nsimd);
-
-  #endif
-}
-
-
-NAMESPACE_END(Grid);

Grid/lattice/Lattice_trace.h

@@ -66,65 +66,6 @@ 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

@@ -194,11 +194,11 @@ accelerator_inline void convertType(vComplexD2 & out, const ComplexD & in) {
 #endif
 
 accelerator_inline void convertType(vComplexF & out, const vComplexD2 & in) {
-  precisionChange(out,in);
+  out.v = Optimization::PrecisionChange::DtoS(in._internal[0].v,in._internal[1].v);
 }
 
 accelerator_inline void convertType(vComplexD2 & out, const vComplexF & in) {
-  precisionChange(out,in);
+  Optimization::PrecisionChange::StoD(in.v,out._internal[0].v,out._internal[1].v);
 }
 
 template<typename T1,typename T2>
@@ -288,36 +288,7 @@ inline void blockProject(Lattice<iVector<CComplex,nbasis > > &coarseData,
     blockZAXPY(fineDataRed,ip,Basis[v],fineDataRed); 
   }
 }
-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,
-                               const VLattice &Basis)
-{
-  int NBatch = fineData.size();
-  assert(coarseData.size() == NBatch);
 
-  GridBase * fine  = fineData[0].Grid();
-  GridBase * coarse= coarseData[0].Grid();
-
-  Lattice<iScalar<CComplex>> ip(coarse);
-  std::vector<Lattice<vobj>> fineDataCopy = fineData;
-
-  autoView(ip_, ip, AcceleratorWrite);
-  for(int v=0;v<nbasis;v++) {
-    for (int k=0; k<NBatch; k++) {
-      autoView( coarseData_ , coarseData[k], AcceleratorWrite);
-      blockInnerProductD(ip,Basis[v],fineDataCopy[k]); // ip = <basis|fine>
-      accelerator_for( sc, coarse->oSites(), vobj::Nsimd(), {
-        convertType(coarseData_[sc](v),ip_[sc]);
-      });
-
-      // improve numerical stability of projection
-      // |fine> = |fine> - <basis|fine> |basis>
-      ip=-ip;
-      blockZAXPY(fineDataCopy[k],ip,Basis[v],fineDataCopy[k]); 
-    }
-  }
-}
 
 template<class vobj,class vobj2,class CComplex>
   inline void blockZAXPY(Lattice<vobj> &fineZ,
@@ -469,13 +440,15 @@ inline void blockSum(Lattice<vobj> &coarseData,const Lattice<vobj> &fineData)
   Coordinate fine_rdimensions = fine->_rdimensions;
   Coordinate coarse_rdimensions = coarse->_rdimensions;
 
+  vobj zz = Zero();
+  
   accelerator_for(sc,coarse->oSites(),1,{
 
       // One thread per sub block
       Coordinate coor_c(_ndimension);
       Lexicographic::CoorFromIndex(coor_c,sc,coarse_rdimensions);  // Block coordinate
 
-      vobj cd = Zero();
+      vobj cd = zz;
       
       for(int sb=0;sb<blockVol;sb++){
 
@@ -617,26 +590,6 @@ inline void blockPromote(const Lattice<iVector<CComplex,nbasis > > &coarseData,
 }
 #endif
 
-template<class vobj,class CComplex,int nbasis,class VLattice>
-inline void batchBlockPromote(const std::vector<Lattice<iVector<CComplex,nbasis>>> &coarseData,
-                               std::vector<Lattice<vobj>> &fineData,
-                               const VLattice &Basis)
-{
-  int NBatch = coarseData.size();
-  assert(fineData.size() == NBatch);
-
-  GridBase * fine   = fineData[0].Grid();
-  GridBase * coarse = coarseData[0].Grid();
-  for (int k=0; k<NBatch; k++)
-    fineData[k]=Zero();
-  for (int i=0;i<nbasis;i++) {
-    for (int k=0; k<NBatch; k++) {
-      Lattice<iScalar<CComplex>> ip = PeekIndex<0>(coarseData[k],i);
-      blockZAXPY(fineData[k],ip,Basis[i],fineData[k]);
-    }
-  }
-}
-
 // Useful for precision conversion, or indeed anything where an operator= does a conversion on scalars.
 // Simd layouts need not match since we use peek/poke Local
 template<class vobj,class vvobj>
@@ -695,68 +648,8 @@ void localCopyRegion(const Lattice<vobj> &From,Lattice<vobj> & To,Coordinate Fro
   for(int d=0;d<nd;d++){
     assert(Fg->_processors[d]  == Tg->_processors[d]);
   }
+
   // the above should guarantee that the operations are local
-  
-#if 1
-
-  size_t nsite = 1;
-  for(int i=0;i<nd;i++) nsite *= RegionSize[i];
-  
-  size_t tbytes = 4*nsite*sizeof(int);
-  int *table = (int*)malloc(tbytes);
- 
-  thread_for(idx, nsite, {
-      Coordinate from_coor, to_coor;
-      size_t rem = idx;
-      for(int i=0;i<nd;i++){
-	size_t base_i  = rem % RegionSize[i]; rem /= RegionSize[i];
-	from_coor[i] = base_i + FromLowerLeft[i];
-	to_coor[i] = base_i + ToLowerLeft[i];
-      }
-      
-      int foidx = Fg->oIndex(from_coor);
-      int fiidx = Fg->iIndex(from_coor);
-      int toidx = Tg->oIndex(to_coor);
-      int tiidx = Tg->iIndex(to_coor);
-      int* tt = table + 4*idx;
-      tt[0] = foidx;
-      tt[1] = fiidx;
-      tt[2] = toidx;
-      tt[3] = tiidx;
-    });
-  
-  int* table_d = (int*)acceleratorAllocDevice(tbytes);
-  acceleratorCopyToDevice(table,table_d,tbytes);
-
-  typedef typename vobj::vector_type vector_type;
-  typedef typename vobj::scalar_type scalar_type;
-
-  autoView(from_v,From,AcceleratorRead);
-  autoView(to_v,To,AcceleratorWrite);
-  
-  accelerator_for(idx,nsite,1,{
-      static const int words=sizeof(vobj)/sizeof(vector_type);
-      int* tt = table_d + 4*idx;
-      int from_oidx = *tt++;
-      int from_lane = *tt++;
-      int to_oidx = *tt++;
-      int to_lane = *tt;
-
-      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);
-      }
-    });
-  
-  acceleratorFreeDevice(table_d);    
-  free(table);
-  
-
-#else  
   Coordinate ldf = Fg->_ldimensions;
   Coordinate rdf = Fg->_rdimensions;
   Coordinate isf = Fg->_istride;
@@ -765,9 +658,9 @@ void localCopyRegion(const Lattice<vobj> &From,Lattice<vobj> & To,Coordinate Fro
   Coordinate ist = Tg->_istride;
   Coordinate ost = Tg->_ostride;
 
-  autoView( t_v , To, CpuWrite);
+  autoView( t_v , To, AcceleratorWrite);
-  autoView( f_v , From, CpuRead);
+  autoView( f_v , From, AcceleratorRead);
-  thread_for(idx,Fg->lSites(),{
+  accelerator_for(idx,Fg->lSites(),1,{
     sobj s;
     Coordinate Fcoor(nd);
     Coordinate Tcoor(nd);
@@ -780,24 +673,17 @@ void localCopyRegion(const Lattice<vobj> &From,Lattice<vobj> & To,Coordinate Fro
       Tcoor[d] = ToLowerLeft[d]+ Fcoor[d]-FromLowerLeft[d];
     }
     if (in_region) {
-#if 0      
+      Integer idx_f = 0; for(int d=0;d<nd;d++) idx_f+=isf[d]*(Fcoor[d]/rdf[d]);
-      Integer idx_f = 0; for(int d=0;d<nd;d++) idx_f+=isf[d]*(Fcoor[d]/rdf[d]); // inner index from
+      Integer idx_t = 0; for(int d=0;d<nd;d++) idx_t+=ist[d]*(Tcoor[d]/rdt[d]);
-      Integer idx_t = 0; for(int d=0;d<nd;d++) idx_t+=ist[d]*(Tcoor[d]/rdt[d]); // inner index to
+      Integer odx_f = 0; for(int d=0;d<nd;d++) odx_f+=osf[d]*(Fcoor[d]%rdf[d]);
-      Integer odx_f = 0; for(int d=0;d<nd;d++) odx_f+=osf[d]*(Fcoor[d]%rdf[d]); // outer index from
+      Integer odx_t = 0; for(int d=0;d<nd;d++) odx_t+=ost[d]*(Tcoor[d]%rdt[d]);
-      Integer odx_t = 0; for(int d=0;d<nd;d++) odx_t+=ost[d]*(Tcoor[d]%rdt[d]); // outer index to
       scalar_type * fp = (scalar_type *)&f_v[odx_f];
       scalar_type * tp = (scalar_type *)&t_v[odx_t];
       for(int w=0;w<words;w++){
-	tp[w].putlane(fp[w].getlane(idx_f),idx_t);
+	tp[idx_t+w*Nsimd] = fp[idx_f+w*Nsimd];  // FIXME IF RRII layout, type pun no worke
       }
-#else
-    peekLocalSite(s,f_v,Fcoor);
-    pokeLocalSite(s,t_v,Tcoor);
-#endif
     }
   });
-
-#endif
 }
 
 
@@ -890,8 +776,6 @@ void ExtractSlice(Lattice<vobj> &lowDim,const Lattice<vobj> & higherDim,int slic
 }
 
 
-//Insert subvolume orthogonal to direction 'orthog' with slice index 'slice_lo' from 'lowDim' onto slice index 'slice_hi' of higherDim
-//The local dimensions of both 'lowDim' and 'higherDim' orthogonal to 'orthog' should be the same
 template<class vobj>
 void InsertSliceLocal(const Lattice<vobj> &lowDim, Lattice<vobj> & higherDim,int slice_lo,int slice_hi, int orthog)
 {
@@ -908,70 +792,11 @@ 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->_processors[d]  == hg->_processors[d]);
-      assert(lg->_ldimensions[d] == hg->_ldimensions[d]);
+    assert(lg->_ldimensions[d] == hg->_ldimensions[d]);
-    }
+  }
   }
 
-#if 1
-  size_t nsite = lg->lSites()/lg->LocalDimensions()[orthog];
-  size_t tbytes = 4*nsite*sizeof(int);
-  int *table = (int*)malloc(tbytes);
-  
-  thread_for(idx,nsite,{
-    Coordinate lcoor(nl);
-    Coordinate hcoor(nh);
-    lcoor[orthog] = slice_lo;
-    hcoor[orthog] = slice_hi;
-    size_t rem = idx;
-    for(int mu=0;mu<nl;mu++){
-      if(mu != orthog){
-	int xmu = rem % lg->LocalDimensions()[mu];  rem /= lg->LocalDimensions()[mu];
-	lcoor[mu] = hcoor[mu] = xmu;
-      }
-    }
-    int loidx = lg->oIndex(lcoor);
-    int liidx = lg->iIndex(lcoor);
-    int hoidx = hg->oIndex(hcoor);
-    int hiidx = hg->iIndex(hcoor);
-    int* tt = table + 4*idx;
-    tt[0] = loidx;
-    tt[1] = liidx;
-    tt[2] = hoidx;
-    tt[3] = hiidx;
-    });
-   
-  int* table_d = (int*)acceleratorAllocDevice(tbytes);
-  acceleratorCopyToDevice(table,table_d,tbytes);
-
-  typedef typename vobj::vector_type vector_type;
-  typedef typename vobj::scalar_type scalar_type;
-
-  autoView(lowDim_v,lowDim,AcceleratorRead);
-  autoView(higherDim_v,higherDim,AcceleratorWrite);
-  
-  accelerator_for(idx,nsite,1,{
-      static const int words=sizeof(vobj)/sizeof(vector_type);
-      int* tt = table_d + 4*idx;
-      int from_oidx = *tt++;
-      int from_lane = *tt++;
-      int to_oidx = *tt++;
-      int to_lane = *tt;
-
-      const vector_type* from = (const vector_type *)&lowDim_v[from_oidx];
-      vector_type* to = (vector_type *)&higherDim_v[to_oidx];
-      
-      scalar_type stmp;
-      for(int w=0;w<words;w++){
-	stmp = getlane(from[w], from_lane);
-	putlane(to[w], stmp, to_lane);
-      }
-    });
-  
-  acceleratorFreeDevice(table_d);    
-  free(table);
-  
-#else
   // the above should guarantee that the operations are local
   autoView(lowDimv,lowDim,CpuRead);
   autoView(higherDimv,higherDim,CpuWrite);
@@ -987,7 +812,6 @@ void InsertSliceLocal(const Lattice<vobj> &lowDim, Lattice<vobj> & higherDim,int
       pokeLocalSite(s,higherDimv,hcoor);
     }
   });
-#endif
 }
 
 
@@ -1256,80 +1080,11 @@ vectorizeFromRevLexOrdArray( std::vector<sobj> &in, Lattice<vobj> &out)
   });
 }
 
-//Very fast precision change. Requires in/out objects to reside on same Grid (e.g. by using double2 for the double-precision field)
-template<class VobjOut, class VobjIn>
-void precisionChangeFast(Lattice<VobjOut> &out, const Lattice<VobjIn> &in)
-{
-  typedef typename VobjOut::vector_type Vout;
-  typedef typename VobjIn::vector_type Vin;
-  const int N = sizeof(VobjOut)/sizeof(Vout);
-  conformable(out.Grid(),in.Grid());
-  out.Checkerboard() = in.Checkerboard();
-  int nsimd = out.Grid()->Nsimd();
-  autoView( out_v  , out, AcceleratorWrite);
-  autoView(  in_v ,   in, AcceleratorRead);
-  accelerator_for(idx,out.Grid()->oSites(),1,{
-      Vout *vout = (Vout *)&out_v[idx];
-      Vin  *vin  = (Vin  *)&in_v[idx];
-      precisionChange(vout,vin,N);
-  });
-}
-//Convert a Lattice from one precision to another (original, slow implementation)
-template<class VobjOut, class VobjIn>
-void precisionChangeOrig(Lattice<VobjOut> &out, const Lattice<VobjIn> &in)
-{
-  assert(out.Grid()->Nd() == in.Grid()->Nd());
-  for(int d=0;d<out.Grid()->Nd();d++){
-    assert(out.Grid()->FullDimensions()[d] == in.Grid()->FullDimensions()[d]);
-  }
-  out.Checkerboard() = in.Checkerboard();
-  GridBase *in_grid=in.Grid();
-  GridBase *out_grid = out.Grid();
-
-  typedef typename VobjOut::scalar_object SobjOut;
-  typedef typename VobjIn::scalar_object SobjIn;
-
-  int ndim = out.Grid()->Nd();
-  int out_nsimd = out_grid->Nsimd();
-  int in_nsimd = in_grid->Nsimd();
-  std::vector<Coordinate > out_icoor(out_nsimd);
-      
-  for(int lane=0; lane < out_nsimd; lane++){
-    out_icoor[lane].resize(ndim);
-    out_grid->iCoorFromIindex(out_icoor[lane], lane);
-  }
-        
-  std::vector<SobjOut> in_slex_conv(in_grid->lSites());
-  unvectorizeToLexOrdArray(in_slex_conv, in);
-    
-  autoView( out_v , out, CpuWrite);
-  thread_for(out_oidx,out_grid->oSites(),{
-    Coordinate out_ocoor(ndim);
-    out_grid->oCoorFromOindex(out_ocoor, out_oidx);
-
-    ExtractPointerArray<SobjOut> ptrs(out_nsimd);      
-
-    Coordinate lcoor(out_grid->Nd());
-      
-    for(int lane=0; lane < out_nsimd; lane++){
-      for(int mu=0;mu<ndim;mu++)
-	lcoor[mu] = out_ocoor[mu] + out_grid->_rdimensions[mu]*out_icoor[lane][mu];
-	
-      int llex; Lexicographic::IndexFromCoor(lcoor, llex, out_grid->_ldimensions);
-      ptrs[lane] = &in_slex_conv[llex];
-    }
-    merge(out_v[out_oidx], ptrs, 0);
-  });
-}
-
 //The workspace for a precision change operation allowing for the reuse of the mapping to save time on subsequent calls
 class precisionChangeWorkspace{
   std::pair<Integer,Integer>* fmap_device; //device pointer
-  //maintain grids for checking
-  GridBase* _out_grid;
-  GridBase* _in_grid;
 public:
-  precisionChangeWorkspace(GridBase *out_grid, GridBase *in_grid): _out_grid(out_grid), _in_grid(in_grid){
+  precisionChangeWorkspace(GridBase *out_grid, GridBase *in_grid){
     //Build a map between the sites and lanes of the output field and the input field as we cannot use the Grids on the device
     assert(out_grid->Nd() == in_grid->Nd());
     for(int d=0;d<out_grid->Nd();d++){
@@ -1376,46 +1131,20 @@ public:
   
   std::pair<Integer,Integer> const* getMap() const{ return fmap_device; }
 
-  void checkGrids(GridBase* out, GridBase* in) const{
-    conformable(out, _out_grid);
-    conformable(in, _in_grid);
-  }
-  
   ~precisionChangeWorkspace(){
     acceleratorFreeDevice(fmap_device);
   }
 };
 
 
-//We would like to use precisionChangeFast when possible. However usage of this requires the Grids to be the same (runtime check)
+//Convert a lattice of one precision to another. The input workspace contains the mapping data.
-//*and* the precisionChange(VobjOut::vector_type, VobjIn, int) function to be defined for the types; this requires an extra compile-time check which we do using some SFINAE trickery
-template<class VobjOut, class VobjIn>
-auto _precisionChangeFastWrap(Lattice<VobjOut> &out, const Lattice<VobjIn> &in, int dummy)->decltype( precisionChange( ((typename VobjOut::vector_type*)0), ((typename VobjIn::vector_type*)0), 1), int()){
-  if(out.Grid() == in.Grid()){
-    precisionChangeFast(out,in);
-    return 1;
-  }else{
-    return 0;
-  }
-}
-template<class VobjOut, class VobjIn>
-int _precisionChangeFastWrap(Lattice<VobjOut> &out, const Lattice<VobjIn> &in, long dummy){ //note long here is intentional; it means the above is preferred if available
-  return 0;
-}
-
-
-//Convert a lattice of one precision to another. Much faster than original implementation but requires a pregenerated workspace
-//which contains the mapping data.
 template<class VobjOut, class VobjIn>
 void precisionChange(Lattice<VobjOut> &out, const Lattice<VobjIn> &in, const precisionChangeWorkspace &workspace){
-  if(_precisionChangeFastWrap(out,in,0)) return;
+  static_assert( std::is_same<typename VobjOut::DoublePrecision, typename VobjIn::DoublePrecision>::value == 1, "copyLane: tensor types must be the same" ); //if tensor types are same the DoublePrecision type must be the same
-  
-  static_assert( std::is_same<typename VobjOut::scalar_typeD, typename VobjIn::scalar_typeD>::value == 1, "precisionChange: tensor types must be the same" ); //if tensor types are same the DoublePrecision type must be the same
 
   out.Checkerboard() = in.Checkerboard();
   constexpr int Nsimd_out = VobjOut::Nsimd();
 
-  workspace.checkGrids(out.Grid(),in.Grid());
   std::pair<Integer,Integer> const* fmap_device = workspace.getMap();
 
   //Do the copy/precision change
@@ -1432,18 +1161,15 @@ void precisionChange(Lattice<VobjOut> &out, const Lattice<VobjIn> &in, const pre
     });
 }
 
-//Convert a Lattice from one precision to another. Much faster than original implementation but slower than precisionChangeFast
+//Convert a Lattice from one precision to another
-//or precisionChange called with pregenerated workspace, as it needs to internally generate the workspace on the host and copy to device
+//Generate the workspace in place; if multiple calls with the same mapping are performed, consider pregenerating the workspace and reusing
 template<class VobjOut, class VobjIn>
 void precisionChange(Lattice<VobjOut> &out, const Lattice<VobjIn> &in){
-  if(_precisionChangeFastWrap(out,in,0)) return;   
   precisionChangeWorkspace workspace(out.Grid(), in.Grid());
   precisionChange(out, in, workspace);
 }
 
 
-
-
 ////////////////////////////////////////////////////////////////////////////////
 // Communicate between grids
 ////////////////////////////////////////////////////////////////////////////////

Grid/lattice/Lattice_view.h

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

Grid/lattice/PaddedCell.h

@@ -1,174 +0,0 @@
-/*************************************************************************************
-    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); }
-};  
-
-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();
-    for(int d=0;d<dims;d++){
-      assert(local[d]>=depth);
-    }
-  }
-  void DeleteGrids(void)
-  {
-    for(int d=0;d<grids.size();d++){
-      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);
-
-    // expand up one dim at a time
-    for(int d=0;d<dims;d++){
-
-      plocal[d] += 2*depth; 
-
-      for(int d=0;d<dims;d++){
-	global[d] = plocal[d]*processors[d];
-      }
-
-      grids.push_back(new GridCartesian(global,simd,processors));
-    }
-  };
-  template<class vobj>
-  inline Lattice<vobj> Extract(const Lattice<vobj> &in) const
-  {
-    Lattice<vobj> out(unpadded_grid);
-
-    Coordinate local     =unpadded_grid->LocalDimensions();
-    Coordinate fll(dims,depth); // depends on the MPI spread
-    Coordinate tll(dims,0); // depends on the MPI spread
-    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;
-  }
-  // 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
-  {
-    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;
-    
-    // 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;
-  }
-
-};
- 
-
-NAMESPACE_END(Grid);
-

Grid/log/Log.cc

@@ -65,40 +65,32 @@ GridLogger GridLogSolver (1, "Solver", GridLogColours, "NORMAL");
 GridLogger GridLogError  (1, "Error" , GridLogColours, "RED");
 GridLogger GridLogWarning(1, "Warning", GridLogColours, "YELLOW");
 GridLogger GridLogMessage(1, "Message", GridLogColours, "NORMAL");
-GridLogger GridLogMemory (1, "Memory", GridLogColours, "NORMAL");
-GridLogger GridLogTracing(1, "Tracing", GridLogColours, "NORMAL");
 GridLogger GridLogDebug  (1, "Debug", GridLogColours, "PURPLE");
 GridLogger GridLogPerformance(1, "Performance", GridLogColours, "GREEN");
-GridLogger GridLogDslash     (1, "Dslash", GridLogColours, "BLUE");
 GridLogger GridLogIterative  (1, "Iterative", GridLogColours, "BLUE");
 GridLogger GridLogIntegrator (1, "Integrator", GridLogColours, "BLUE");
 GridLogger GridLogHMC (1, "HMC", GridLogColours, "BLUE");
 
 void GridLogConfigure(std::vector<std::string> &logstreams) {
-  GridLogError.Active(1);
+  GridLogError.Active(0);
   GridLogWarning.Active(0);
   GridLogMessage.Active(1); // at least the messages should be always on
-  GridLogMemory.Active(0); 
-  GridLogTracing.Active(0); 
   GridLogIterative.Active(0);
   GridLogDebug.Active(0);
   GridLogPerformance.Active(0);
-  GridLogDslash.Active(0);
   GridLogIntegrator.Active(1);
   GridLogColours.Active(0);
   GridLogHMC.Active(1);
 
   for (int i = 0; i < logstreams.size(); i++) {
-    if (logstreams[i] == std::string("Tracing"))     GridLogTracing.Active(1);
+    if (logstreams[i] == std::string("Error"))       GridLogError.Active(1);
-    if (logstreams[i] == std::string("Memory"))      GridLogMemory.Active(1);
     if (logstreams[i] == std::string("Warning"))     GridLogWarning.Active(1);
     if (logstreams[i] == std::string("NoMessage"))   GridLogMessage.Active(0);
     if (logstreams[i] == std::string("Iterative"))   GridLogIterative.Active(1);
     if (logstreams[i] == std::string("Debug"))       GridLogDebug.Active(1);
     if (logstreams[i] == std::string("Performance")) GridLogPerformance.Active(1);
-    if (logstreams[i] == std::string("Dslash"))      GridLogDslash.Active(1);
+    if (logstreams[i] == std::string("NoIntegrator"))  GridLogIntegrator.Active(0);
-    if (logstreams[i] == std::string("NoIntegrator"))GridLogIntegrator.Active(0);
+    if (logstreams[i] == std::string("NoHMC"))         GridLogHMC.Active(0);
-    if (logstreams[i] == std::string("NoHMC"))       GridLogHMC.Active(0);
     if (logstreams[i] == std::string("Colours"))     GridLogColours.Active(1);
   }
 }

Grid/log/Log.h

@@ -138,8 +138,7 @@ public:
         stream << std::setw(log.topWidth);
       }
       stream << log.topName << log.background()<< " : ";
-      //      stream << log.colour() <<  std::left;
+      stream << log.colour() <<  std::left;
-      stream <<  std::left;
       if (log.chanWidth > 0)
       {
         stream << std::setw(log.chanWidth);
@@ -154,9 +153,9 @@ public:
 	stream << log.evidence()
 	       << now	       << log.background() << " : " ;
       }
-      //      stream << log.colour();
+      stream << log.colour();
-      stream <<  std::right;
       stream.flags(f);
+
       return stream;
     } else { 
       return devnull;
@@ -179,53 +178,15 @@ 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 GridLogIterative;
+extern GridLogger GridLogIntegrator  ;
-extern GridLogger GridLogIntegrator;
 extern GridLogger GridLogHMC;
-extern GridLogger GridLogMemory;
-extern GridLogger GridLogTracing;
 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/IldgIO.h

@@ -31,7 +31,6 @@ directory
 #include <fstream>
 #include <iomanip>
 #include <iostream>
-#include <string>
 #include <map>
 
 #include <pwd.h>
@@ -655,8 +654,7 @@ class IldgWriter : public ScidacWriter {
     // Fill ILDG header data struct
     //////////////////////////////////////////////////////
     ildgFormat ildgfmt ;
-    const std::string stNC = std::to_string( Nc ) ;
+    ildgfmt.field     = std::string("su3gauge");
-    ildgfmt.field          = std::string("su"+stNC+"gauge");
 
     if ( format == std::string("IEEE32BIG") ) { 
       ildgfmt.precision = 32;
@@ -873,8 +871,7 @@ class IldgReader : public GridLimeReader {
     } else { 
 
       assert(found_ildgFormat);
-      const std::string stNC = std::to_string( Nc ) ;
+      assert ( ildgFormat_.field == std::string("su3gauge") );
-      assert ( ildgFormat_.field == std::string("su"+stNC+"gauge") );
 
       ///////////////////////////////////////////////////////////////////////////////////////
       // Populate our Grid metadata as best we can
@@ -882,7 +879,7 @@ class IldgReader : public GridLimeReader {
 
       std::ostringstream vers; vers << ildgFormat_.version;
       FieldMetaData_.hdr_version = vers.str();
-      FieldMetaData_.data_type = std::string("4D_SU"+stNC+"_GAUGE_"+stNC+"x"+stNC);
+      FieldMetaData_.data_type = std::string("4D_SU3_GAUGE_3X3");
 
       FieldMetaData_.nd=4;
       FieldMetaData_.dimension.resize(4);

Grid/parallelIO/MetaData.h

@@ -6,8 +6,8 @@
 
     Copyright (C) 2015
 
+
     Author: Peter Boyle <paboyle@ph.ed.ac.uk>
-    Author: Jamie Hudspith <renwick.james.hudspth@gmail.com>
 
     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
@@ -182,8 +182,8 @@ class GaugeStatistics
 public:
   void operator()(Lattice<vLorentzColourMatrixD> & data,FieldMetaData &header)
   {
-    header.link_trace = WilsonLoops<Impl>::linkTrace(data);
+    header.link_trace=WilsonLoops<Impl>::linkTrace(data);
-    header.plaquette  = WilsonLoops<Impl>::avgPlaquette(data);
+    header.plaquette =WilsonLoops<Impl>::avgPlaquette(data);
   }
 };
 typedef GaugeStatistics<PeriodicGimplD> PeriodicGaugeStatistics;
@@ -203,24 +203,20 @@ template<> inline void PrepareMetaData<vLorentzColourMatrixD>(Lattice<vLorentzCo
 //////////////////////////////////////////////////////////////////////
 inline void reconstruct3(LorentzColourMatrix & cm)
 {
-  assert( Nc < 4 && Nc > 1 ) ;
+  const int x=0;
+  const int y=1;
+  const int z=2;
   for(int mu=0;mu<Nd;mu++){
-    #if Nc == 2
+    cm(mu)()(2,x) = adj(cm(mu)()(0,y)*cm(mu)()(1,z)-cm(mu)()(0,z)*cm(mu)()(1,y)); //x= yz-zy
-      cm(mu)()(1,0) = -adj(cm(mu)()(0,y)) ;
+    cm(mu)()(2,y) = adj(cm(mu)()(0,z)*cm(mu)()(1,x)-cm(mu)()(0,x)*cm(mu)()(1,z)); //y= zx-xz
-      cm(mu)()(1,1) =  adj(cm(mu)()(0,x)) ;
+    cm(mu)()(2,z) = adj(cm(mu)()(0,x)*cm(mu)()(1,y)-cm(mu)()(0,y)*cm(mu)()(1,x)); //z= xy-yx
-    #else
-      const int x=0 , y=1 , z=2 ; // a little disinenuous labelling
-      cm(mu)()(2,x) = adj(cm(mu)()(0,y)*cm(mu)()(1,z)-cm(mu)()(0,z)*cm(mu)()(1,y)); //x= yz-zy
-      cm(mu)()(2,y) = adj(cm(mu)()(0,z)*cm(mu)()(1,x)-cm(mu)()(0,x)*cm(mu)()(1,z)); //y= zx-xz
-      cm(mu)()(2,z) = adj(cm(mu)()(0,x)*cm(mu)()(1,y)-cm(mu)()(0,y)*cm(mu)()(1,x)); //z= xy-yx
-    #endif
   }
 }
 
 ////////////////////////////////////////////////////////////////////////////////
 // Some data types for intermediate storage
 ////////////////////////////////////////////////////////////////////////////////
-template<typename vtype> using iLorentzColour2x3 = iVector<iVector<iVector<vtype, Nc>, Nc-1>, Nd >;
+template<typename vtype> using iLorentzColour2x3 = iVector<iVector<iVector<vtype, Nc>, 2>, Nd >;
 
 typedef iLorentzColour2x3<Complex>  LorentzColour2x3;
 typedef iLorentzColour2x3<ComplexF> LorentzColour2x3F;
@@ -282,6 +278,7 @@ struct GaugeSimpleMunger{
 
 template <class fobj, class sobj>
 struct GaugeSimpleUnmunger {
+
   void operator()(sobj &in, fobj &out) {
     for (int mu = 0; mu < Nd; mu++) {
       for (int i = 0; i < Nc; i++) {
@@ -320,8 +317,8 @@ template<class fobj,class sobj>
 struct Gauge3x2munger{
   void operator() (fobj &in,sobj &out){
     for(int mu=0;mu<Nd;mu++){
-      for(int i=0;i<Nc-1;i++){
+      for(int i=0;i<2;i++){
-	for(int j=0;j<Nc;j++){
+	for(int j=0;j<3;j++){
 	  out(mu)()(i,j) = in(mu)(i)(j);
 	}}
     }
@@ -333,8 +330,8 @@ template<class fobj,class sobj>
 struct Gauge3x2unmunger{
   void operator() (sobj &in,fobj &out){
     for(int mu=0;mu<Nd;mu++){
-      for(int i=0;i<Nc-1;i++){
+      for(int i=0;i<2;i++){
-	for(int j=0;j<Nc;j++){
+	for(int j=0;j<3;j++){
 	  out(mu)(i)(j) = in(mu)()(i,j);
 	}}
     }

Grid/parallelIO/NerscIO.h

@@ -9,7 +9,6 @@
     Author: Matt Spraggs <matthew.spraggs@gmail.com>
     Author: Peter Boyle <paboyle@ph.ed.ac.uk>
     Author: paboyle <paboyle@ph.ed.ac.uk>
-    Author: Jamie Hudspith <renwick.james.hudspth@gmail.com>
 
     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
@@ -31,8 +30,6 @@
 #ifndef GRID_NERSC_IO_H
 #define GRID_NERSC_IO_H
 
-#include <string>
-
 NAMESPACE_BEGIN(Grid);
 
 using namespace Grid;
@@ -150,17 +147,15 @@ public:
 
     std::string format(header.floating_point);
 
-    const int ieee32big = (format == std::string("IEEE32BIG"));
+    int ieee32big = (format == std::string("IEEE32BIG"));
-    const int ieee32    = (format == std::string("IEEE32"));
+    int ieee32    = (format == std::string("IEEE32"));
-    const int ieee64big = (format == std::string("IEEE64BIG"));
+    int ieee64big = (format == std::string("IEEE64BIG"));
-    const int ieee64    = (format == std::string("IEEE64") || \
+    int ieee64    = (format == std::string("IEEE64") || format == std::string("IEEE64LITTLE"));
-			   format == std::string("IEEE64LITTLE"));
 
     uint32_t nersc_csum,scidac_csuma,scidac_csumb;
     // depending on datatype, set up munger;
     // munger is a function of <floating point, Real, data_type>
-    const std::string stNC = std::to_string( Nc ) ;
+    if ( header.data_type == std::string("4D_SU3_GAUGE") ) {
-    if ( header.data_type == std::string("4D_SU"+stNC+"_GAUGE") ) {
       if ( ieee32 || ieee32big ) {
 	BinaryIO::readLatticeObject<vLorentzColourMatrixD, LorentzColour2x3F> 
 	  (Umu,file,Gauge3x2munger<LorentzColour2x3F,LorentzColourMatrix>(), offset,format,
@@ -171,7 +166,7 @@ public:
 	  (Umu,file,Gauge3x2munger<LorentzColour2x3D,LorentzColourMatrix>(),offset,format,
 	   nersc_csum,scidac_csuma,scidac_csumb);
       }
-    } else if ( header.data_type == std::string("4D_SU"+stNC+"_GAUGE_"+stNC+"x"+stNC) ) {
+    } else if ( header.data_type == std::string("4D_SU3_GAUGE_3x3") ) {
       if ( ieee32 || ieee32big ) {
 	BinaryIO::readLatticeObject<vLorentzColourMatrixD,LorentzColourMatrixF>
 	  (Umu,file,GaugeSimpleMunger<LorentzColourMatrixF,LorentzColourMatrix>(),offset,format,
@@ -216,29 +211,27 @@ public:
   template<class GaugeStats=PeriodicGaugeStatistics>
   static inline void writeConfiguration(Lattice<vLorentzColourMatrixD > &Umu,
 					std::string file, 
-					std::string ens_label = std::string("DWF"),
+					std::string ens_label = std::string("DWF"))
-					std::string ens_id = std::string("UKQCD"),
-					unsigned int sequence_number = 1)
   {
-    writeConfiguration(Umu,file,0,1,ens_label,ens_id,sequence_number);
+    writeConfiguration(Umu,file,0,1,ens_label);
   }
   template<class GaugeStats=PeriodicGaugeStatistics>
   static inline void writeConfiguration(Lattice<vLorentzColourMatrixD > &Umu,
 					std::string file, 
 					int two_row,
 					int bits32,
-					std::string ens_label = std::string("DWF"),
+					std::string ens_label = std::string("DWF"))
-					std::string ens_id = std::string("UKQCD"),
-					unsigned int sequence_number = 1)
   {
     typedef vLorentzColourMatrixD vobj;
     typedef typename vobj::scalar_object sobj;
 
     FieldMetaData header;
-    header.sequence_number = sequence_number;
+    ///////////////////////////////////////////
-    header.ensemble_id     = ens_id;
+    // Following should become arguments
+    ///////////////////////////////////////////
+    header.sequence_number = 1;
+    header.ensemble_id     = std::string("UKQCD");
     header.ensemble_label  = ens_label;
-    header.hdr_version     = "1.0" ;
 
     typedef LorentzColourMatrixD fobj3D;
     typedef LorentzColour2x3D    fobj2D;
@@ -252,14 +245,10 @@ public:
 
     uint64_t offset;
 
-    // Sod it -- always write NcxNc double
+    // Sod it -- always write 3x3 double
-    header.floating_point  = std::string("IEEE64BIG");
+    header.floating_point = std::string("IEEE64BIG");
-    const std::string stNC = std::to_string( Nc ) ;
+    header.data_type      = std::string("4D_SU3_GAUGE_3x3");
-    if( two_row ) {
+    GaugeSimpleUnmunger<fobj3D,sobj> munge;
-      header.data_type = std::string("4D_SU" + stNC + "_GAUGE" );
-    } else {
-      header.data_type = std::string("4D_SU" + stNC + "_GAUGE_" + stNC + "x" + stNC );
-    }
     if ( grid->IsBoss() ) { 
       truncate(file);
       offset = writeHeader(header,file);
@@ -267,15 +256,8 @@ public:
     grid->Broadcast(0,(void *)&offset,sizeof(offset));
 
     uint32_t nersc_csum,scidac_csuma,scidac_csumb;
-    if( two_row ) {
+    BinaryIO::writeLatticeObject<vobj,fobj3D>(Umu,file,munge,offset,header.floating_point,
-      Gauge3x2unmunger<fobj2D,sobj> munge;
+					      nersc_csum,scidac_csuma,scidac_csumb);
-      BinaryIO::writeLatticeObject<vobj,fobj2D>(Umu,file,munge,offset,header.floating_point,
-						nersc_csum,scidac_csuma,scidac_csumb);
-    } else {
-      GaugeSimpleUnmunger<fobj3D,sobj> munge;
-      BinaryIO::writeLatticeObject<vobj,fobj3D>(Umu,file,munge,offset,header.floating_point,
-						nersc_csum,scidac_csuma,scidac_csumb);
-    }
     header.checksum = nersc_csum;
     if ( grid->IsBoss() ) { 
       writeHeader(header,file);
@@ -307,7 +289,8 @@ public:
     header.plaquette=0.0;
     MachineCharacteristics(header);
 
-    uint64_t offset;
+	uint64_t offset;
+  
 #ifdef RNG_RANLUX
     header.floating_point = std::string("UINT64");
     header.data_type      = std::string("RANLUX48");
@@ -347,7 +330,7 @@ public:
 
     GridBase *grid = parallel.Grid();
 
-    uint64_t offset = readHeader(file,grid,header);
+	uint64_t offset = readHeader(file,grid,header);
 
     FieldMetaData clone(header);
 

Grid/perfmon/PerfCount.cc

@@ -27,12 +27,9 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 /*  END LEGAL */
 
 #include <Grid/GridCore.h>
-
-#include <Grid/perfmon/Timer.h>
 #include <Grid/perfmon/PerfCount.h>
-NAMESPACE_BEGIN(Grid);
 
-GridTimePoint theProgramStart = GridClock::now();
+NAMESPACE_BEGIN(Grid);
 
 #define CacheControl(L,O,R) ((PERF_COUNT_HW_CACHE_##L)|(PERF_COUNT_HW_CACHE_OP_##O<<8)| (PERF_COUNT_HW_CACHE_RESULT_##R<<16))
 #define RawConfig(A,B) (A<<8|B)

Grid/perfmon/PerfCount.h

@@ -30,12 +30,6 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 #ifndef GRID_PERFCOUNT_H
 #define GRID_PERFCOUNT_H
 
-
-#ifndef __SSC_START
-#define __SSC_START
-#define __SSC_STOP
-#endif
-
 #include <sys/time.h>
 #include <ctime>
 #include <chrono>
@@ -78,9 +72,17 @@ static long perf_event_open(struct perf_event_attr *hw_event, pid_t pid,
 inline uint64_t cyclecount(void){ 
   return 0;
 }
+#define __SSC_MARK(mark) __asm__ __volatile__ ("movl %0, %%ebx; .byte 0x64, 0x67, 0x90 " ::"i"(mark):"%ebx")
+#define __SSC_STOP  __SSC_MARK(0x110)
+#define __SSC_START __SSC_MARK(0x111)
+
 
 #else
 
+#define __SSC_MARK(mark) 
+#define __SSC_STOP  
+#define __SSC_START 
+
 /*
  * cycle counters arch dependent
  */

Grid/perfmon/Timer.h

@@ -35,8 +35,17 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 
 NAMESPACE_BEGIN(Grid)
 
-//typedef  std::chrono::system_clock          GridClock;
+// Dress the output; use std::chrono
-typedef  std::chrono::high_resolution_clock   GridClock;
+// C++11 time facilities better?
+inline double usecond(void) {
+  struct timeval tv;
+#ifdef TIMERS_ON
+  gettimeofday(&tv,NULL);
+#endif
+  return 1.0*tv.tv_usec + 1.0e6*tv.tv_sec;
+}
+
+typedef  std::chrono::system_clock          GridClock;
 typedef  std::chrono::time_point<GridClock> GridTimePoint;
 
 typedef  std::chrono::seconds               GridSecs;
@@ -44,15 +53,6 @@ typedef  std::chrono::milliseconds          GridMillisecs;
 typedef  std::chrono::microseconds          GridUsecs;
 typedef  std::chrono::microseconds          GridTime;
 
-extern GridTimePoint theProgramStart;
-// Dress the output; use std::chrono
-// C++11 time facilities better?
-inline double usecond(void) {
-  auto usecs = std::chrono::duration_cast<GridUsecs>(GridClock::now()-theProgramStart); 
-  return 1.0*usecs.count();
-}
-
-
 inline std::ostream& operator<< (std::ostream & stream, const GridSecs & time)
 {
   stream << time.count()<<" s";

Grid/perfmon/Tracing.h

@@ -1,70 +0,0 @@
-#pragma once
-
-NAMESPACE_BEGIN(Grid);
-
-#ifdef GRID_TRACING_NVTX
-#include <nvToolsExt.h>
-class GridTracer {
-public:
-  GridTracer(const char* name) {
-    nvtxRangePushA(name);
-  }
-  ~GridTracer() {
-    nvtxRangePop();
-  }
-};
-inline void tracePush(const char *name) { nvtxRangePushA(name); }
-inline void tracePop(const char *name) { nvtxRangePop(); }
-inline int  traceStart(const char *name) {  }
-inline void traceStop(int ID) {  }
-#endif
-
-#ifdef GRID_TRACING_ROCTX
-#include <roctracer/roctx.h>
-class GridTracer {
- public:
-  GridTracer(const char* name) {
-    roctxRangePushA(name);
-    std::cout << "roctxRangePush "<<name<<std::endl;
-  }
-  ~GridTracer() {
-    roctxRangePop();
-    std::cout << "roctxRangePop "<<std::endl;
-  }
-};
-inline void tracePush(const char *name) { roctxRangePushA(name); }
-inline void tracePop(const char *name) { roctxRangePop(); }
-inline int  traceStart(const char *name) { return roctxRangeStart(name); }
-inline void traceStop(int ID) { roctxRangeStop(ID); }
-#endif
-
-#ifdef GRID_TRACING_TIMER
-class GridTracer {
- public:
-  const char *name;
-  double elapsed;
-  GridTracer(const char* _name) {
-    name = _name;
-    elapsed=-usecond();
-  }
-  ~GridTracer() {
-    elapsed+=usecond();
-    std::cout << GridLogTracing << name << " took " <<elapsed<< " us" <<std::endl;
-  }
-};
-inline void tracePush(const char *name) {  }
-inline void tracePop(const char *name) {  }
-inline int  traceStart(const char *name) { return 0; }
-inline void traceStop(int ID) {  }
-#endif
-
-#ifdef GRID_TRACING_NONE
-#define GRID_TRACE(name) 
-inline void tracePush(const char *name) {  }
-inline void tracePop(const char *name) {  }
-inline int  traceStart(const char *name) { return 0;  }
-inline void traceStop(int ID) {  }
-#else
-#define GRID_TRACE(name) GridTracer uniq_name_using_macros##__COUNTER__(name);
-#endif
-NAMESPACE_END(Grid);

Grid/pugixml/pugixml.cc

@@ -16,12 +16,8 @@
 
 #ifdef __NVCC__
 #pragma push
-#ifdef __NVCC_DIAG_PRAGMA_SUPPORT__
-#pragma nv_diag_suppress declared_but_not_referenced // suppress "function was declared but never referenced warning"
-#else
 #pragma diag_suppress declared_but_not_referenced // suppress "function was declared but never referenced warning"
 #endif
-#endif
 
 #include "pugixml.h"
 

Grid/qcd/QCD.h

@@ -104,7 +104,6 @@ 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> > >;
@@ -127,7 +126,6 @@ typedef iSpinMatrix<ComplexD >          SpinMatrixD;
 typedef iSpinMatrix<vComplex >          vSpinMatrix;
 typedef iSpinMatrix<vComplexF>          vSpinMatrixF;
 typedef iSpinMatrix<vComplexD>          vSpinMatrixD;
-typedef iSpinMatrix<vComplexD2>         vSpinMatrixD2;
 
 // Colour Matrix
 typedef iColourMatrix<Complex  >        ColourMatrix;
@@ -137,7 +135,6 @@ typedef iColourMatrix<ComplexD >        ColourMatrixD;
 typedef iColourMatrix<vComplex >        vColourMatrix;
 typedef iColourMatrix<vComplexF>        vColourMatrixF;
 typedef iColourMatrix<vComplexD>        vColourMatrixD;
-typedef iColourMatrix<vComplexD2>       vColourMatrixD2;
 
 // SpinColour matrix
 typedef iSpinColourMatrix<Complex  >    SpinColourMatrix;
@@ -147,7 +144,6 @@ typedef iSpinColourMatrix<ComplexD >    SpinColourMatrixD;
 typedef iSpinColourMatrix<vComplex >    vSpinColourMatrix;
 typedef iSpinColourMatrix<vComplexF>    vSpinColourMatrixF;
 typedef iSpinColourMatrix<vComplexD>    vSpinColourMatrixD;
-typedef iSpinColourMatrix<vComplexD2>   vSpinColourMatrixD2;
 
 // SpinColourSpinColour matrix
 typedef iSpinColourSpinColourMatrix<Complex  >    SpinColourSpinColourMatrix;
@@ -157,7 +153,6 @@ typedef iSpinColourSpinColourMatrix<ComplexD >    SpinColourSpinColourMatrixD;
 typedef iSpinColourSpinColourMatrix<vComplex >    vSpinColourSpinColourMatrix;
 typedef iSpinColourSpinColourMatrix<vComplexF>    vSpinColourSpinColourMatrixF;
 typedef iSpinColourSpinColourMatrix<vComplexD>    vSpinColourSpinColourMatrixD;
-typedef iSpinColourSpinColourMatrix<vComplexD2>   vSpinColourSpinColourMatrixD2;
 
 // SpinColourSpinColour matrix
 typedef iSpinColourSpinColourMatrix<Complex  >    SpinColourSpinColourMatrix;
@@ -167,46 +162,33 @@ typedef iSpinColourSpinColourMatrix<ComplexD >    SpinColourSpinColourMatrixD;
 typedef iSpinColourSpinColourMatrix<vComplex >    vSpinColourSpinColourMatrix;
 typedef iSpinColourSpinColourMatrix<vComplexF>    vSpinColourSpinColourMatrixF;
 typedef iSpinColourSpinColourMatrix<vComplexD>    vSpinColourSpinColourMatrixD;
-typedef iSpinColourSpinColourMatrix<vComplexD2>   vSpinColourSpinColourMatrixD2;
 
 // LorentzColour
 typedef iLorentzColourMatrix<Complex  > LorentzColourMatrix;
 typedef iLorentzColourMatrix<ComplexF > LorentzColourMatrixF;
 typedef iLorentzColourMatrix<ComplexD > LorentzColourMatrixD;
 
-typedef iLorentzColourMatrix<vComplex >  vLorentzColourMatrix;
+typedef iLorentzColourMatrix<vComplex > vLorentzColourMatrix;
-typedef iLorentzColourMatrix<vComplexF>  vLorentzColourMatrixF;
+typedef iLorentzColourMatrix<vComplexF> vLorentzColourMatrixF;
-typedef iLorentzColourMatrix<vComplexD>  vLorentzColourMatrixD;
+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;
 typedef iDoubleStoredColourMatrix<ComplexD > DoubleStoredColourMatrixD;
 
-typedef iDoubleStoredColourMatrix<vComplex >  vDoubleStoredColourMatrix;
+typedef iDoubleStoredColourMatrix<vComplex > vDoubleStoredColourMatrix;
-typedef iDoubleStoredColourMatrix<vComplexF>  vDoubleStoredColourMatrixF;
+typedef iDoubleStoredColourMatrix<vComplexF> vDoubleStoredColourMatrixF;
-typedef iDoubleStoredColourMatrix<vComplexD>  vDoubleStoredColourMatrixD;
+typedef iDoubleStoredColourMatrix<vComplexD> vDoubleStoredColourMatrixD;
-typedef iDoubleStoredColourMatrix<vComplexD2> vDoubleStoredColourMatrixD2;
 
 //G-parity flavour matrix
 typedef iGparityFlavourMatrix<Complex> GparityFlavourMatrix;
 typedef iGparityFlavourMatrix<ComplexF> GparityFlavourMatrixF;
 typedef iGparityFlavourMatrix<ComplexD> GparityFlavourMatrixD;
 
-typedef iGparityFlavourMatrix<vComplex>   vGparityFlavourMatrix;
+typedef iGparityFlavourMatrix<vComplex> vGparityFlavourMatrix;
-typedef iGparityFlavourMatrix<vComplexF>  vGparityFlavourMatrixF;
+typedef iGparityFlavourMatrix<vComplexF> vGparityFlavourMatrixF;
-typedef iGparityFlavourMatrix<vComplexD>  vGparityFlavourMatrixD;
+typedef iGparityFlavourMatrix<vComplexD> vGparityFlavourMatrixD;
-typedef iGparityFlavourMatrix<vComplexD2> vGparityFlavourMatrixD2;
 
 
 // Spin vector
@@ -217,7 +199,6 @@ typedef iSpinVector<ComplexD>           SpinVectorD;
 typedef iSpinVector<vComplex >           vSpinVector;
 typedef iSpinVector<vComplexF>           vSpinVectorF;
 typedef iSpinVector<vComplexD>           vSpinVectorD;
-typedef iSpinVector<vComplexD2>          vSpinVectorD2;
 
 // Colour vector
 typedef iColourVector<Complex >         ColourVector;
@@ -227,7 +208,6 @@ typedef iColourVector<ComplexD>         ColourVectorD;
 typedef iColourVector<vComplex >         vColourVector;
 typedef iColourVector<vComplexF>         vColourVectorF;
 typedef iColourVector<vComplexD>         vColourVectorD;
-typedef iColourVector<vComplexD2>        vColourVectorD2;
 
 // SpinColourVector
 typedef iSpinColourVector<Complex >     SpinColourVector;
@@ -237,7 +217,6 @@ typedef iSpinColourVector<ComplexD>     SpinColourVectorD;
 typedef iSpinColourVector<vComplex >     vSpinColourVector;
 typedef iSpinColourVector<vComplexF>     vSpinColourVectorF;
 typedef iSpinColourVector<vComplexD>     vSpinColourVectorD;
-typedef iSpinColourVector<vComplexD2>    vSpinColourVectorD2;
 
 // HalfSpin vector
 typedef iHalfSpinVector<Complex >       HalfSpinVector;
@@ -247,17 +226,15 @@ typedef iHalfSpinVector<ComplexD>       HalfSpinVectorD;
 typedef iHalfSpinVector<vComplex >       vHalfSpinVector;
 typedef iHalfSpinVector<vComplexF>       vHalfSpinVectorF;
 typedef iHalfSpinVector<vComplexD>       vHalfSpinVectorD;
-typedef iHalfSpinVector<vComplexD2>      vHalfSpinVectorD2;
 
 // HalfSpinColour vector
 typedef iHalfSpinColourVector<Complex > HalfSpinColourVector;
 typedef iHalfSpinColourVector<ComplexF> HalfSpinColourVectorF;
 typedef iHalfSpinColourVector<ComplexD> HalfSpinColourVectorD;
     
-typedef iHalfSpinColourVector<vComplex >  vHalfSpinColourVector;
+typedef iHalfSpinColourVector<vComplex > vHalfSpinColourVector;
-typedef iHalfSpinColourVector<vComplexF>  vHalfSpinColourVectorF;
+typedef iHalfSpinColourVector<vComplexF> vHalfSpinColourVectorF;
-typedef iHalfSpinColourVector<vComplexD>  vHalfSpinColourVectorD;
+typedef iHalfSpinColourVector<vComplexD> vHalfSpinColourVectorD;
-typedef iHalfSpinColourVector<vComplexD2> vHalfSpinColourVectorD2;
 
 //G-parity flavour vector
 typedef iGparityFlavourVector<Complex >         GparityFlavourVector;
@@ -267,7 +244,7 @@ typedef iGparityFlavourVector<ComplexD>         GparityFlavourVectorD;
 typedef iGparityFlavourVector<vComplex >         vGparityFlavourVector;
 typedef iGparityFlavourVector<vComplexF>         vGparityFlavourVectorF;
 typedef iGparityFlavourVector<vComplexD>         vGparityFlavourVectorD;
-typedef iGparityFlavourVector<vComplexD2>        vGparityFlavourVectorD2;
+
     
 // singlets
 typedef iSinglet<Complex >         TComplex;     // FIXME This is painful. Tensor singlet complex type.
@@ -277,7 +254,6 @@ typedef iSinglet<ComplexD>         TComplexD;    // FIXME This is painful. Tenso
 typedef iSinglet<vComplex >        vTComplex ;   // what if we don't know the tensor structure
 typedef iSinglet<vComplexF>        vTComplexF;   // what if we don't know the tensor structure
 typedef iSinglet<vComplexD>        vTComplexD;   // what if we don't know the tensor structure
-typedef iSinglet<vComplexD2>       vTComplexD2;   // what if we don't know the tensor structure
 
 typedef iSinglet<Real >            TReal;        // Shouldn't need these; can I make it work without?
 typedef iSinglet<RealF>            TRealF;       // Shouldn't need these; can I make it work without?
@@ -295,62 +271,47 @@ typedef iSinglet<Integer >         TInteger;
 typedef Lattice<vColourMatrix>          LatticeColourMatrix;
 typedef Lattice<vColourMatrixF>         LatticeColourMatrixF;
 typedef Lattice<vColourMatrixD>         LatticeColourMatrixD;
-typedef Lattice<vColourMatrixD2>        LatticeColourMatrixD2;
 
 typedef Lattice<vSpinMatrix>            LatticeSpinMatrix;
 typedef Lattice<vSpinMatrixF>           LatticeSpinMatrixF;
 typedef Lattice<vSpinMatrixD>           LatticeSpinMatrixD;
-typedef Lattice<vSpinMatrixD2>          LatticeSpinMatrixD2;
 
 typedef Lattice<vSpinColourMatrix>      LatticeSpinColourMatrix;
 typedef Lattice<vSpinColourMatrixF>     LatticeSpinColourMatrixF;
 typedef Lattice<vSpinColourMatrixD>     LatticeSpinColourMatrixD;
-typedef Lattice<vSpinColourMatrixD2>    LatticeSpinColourMatrixD2;
 
 typedef Lattice<vSpinColourSpinColourMatrix>      LatticeSpinColourSpinColourMatrix;
 typedef Lattice<vSpinColourSpinColourMatrixF>     LatticeSpinColourSpinColourMatrixF;
 typedef Lattice<vSpinColourSpinColourMatrixD>     LatticeSpinColourSpinColourMatrixD;
-typedef Lattice<vSpinColourSpinColourMatrixD2>    LatticeSpinColourSpinColourMatrixD2;
 
-typedef Lattice<vLorentzColourMatrix>   LatticeLorentzColourMatrix;
+typedef Lattice<vLorentzColourMatrix>  LatticeLorentzColourMatrix;
-typedef Lattice<vLorentzColourMatrixF>  LatticeLorentzColourMatrixF;
+typedef Lattice<vLorentzColourMatrixF> LatticeLorentzColourMatrixF;
-typedef Lattice<vLorentzColourMatrixD>  LatticeLorentzColourMatrixD;
+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<vDoubleStoredColourMatrix>  LatticeDoubleStoredColourMatrix;
-typedef Lattice<vDoubleStoredColourMatrixF>  LatticeDoubleStoredColourMatrixF;
+typedef Lattice<vDoubleStoredColourMatrixF> LatticeDoubleStoredColourMatrixF;
-typedef Lattice<vDoubleStoredColourMatrixD>  LatticeDoubleStoredColourMatrixD;
+typedef Lattice<vDoubleStoredColourMatrixD> LatticeDoubleStoredColourMatrixD;
-typedef Lattice<vDoubleStoredColourMatrixD2> LatticeDoubleStoredColourMatrixD2;
 
 typedef Lattice<vSpinVector>            LatticeSpinVector;
 typedef Lattice<vSpinVectorF>           LatticeSpinVectorF;
 typedef Lattice<vSpinVectorD>           LatticeSpinVectorD;
-typedef Lattice<vSpinVectorD2>          LatticeSpinVectorD2;
 
 typedef Lattice<vColourVector>          LatticeColourVector;
 typedef Lattice<vColourVectorF>         LatticeColourVectorF;
 typedef Lattice<vColourVectorD>         LatticeColourVectorD;
-typedef Lattice<vColourVectorD2>        LatticeColourVectorD2;
 
 typedef Lattice<vSpinColourVector>      LatticeSpinColourVector;
 typedef Lattice<vSpinColourVectorF>     LatticeSpinColourVectorF;
 typedef Lattice<vSpinColourVectorD>     LatticeSpinColourVectorD;
-typedef Lattice<vSpinColourVectorD2>    LatticeSpinColourVectorD2;
 
 typedef Lattice<vHalfSpinVector>        LatticeHalfSpinVector;
 typedef Lattice<vHalfSpinVectorF>       LatticeHalfSpinVectorF;
 typedef Lattice<vHalfSpinVectorD>       LatticeHalfSpinVectorD;
-typedef Lattice<vHalfSpinVectorD2>      LatticeHalfSpinVectorD2;
 
-typedef Lattice<vHalfSpinColourVector>   LatticeHalfSpinColourVector;
+typedef Lattice<vHalfSpinColourVector>  LatticeHalfSpinColourVector;
-typedef Lattice<vHalfSpinColourVectorF>  LatticeHalfSpinColourVectorF;
+typedef Lattice<vHalfSpinColourVectorF> LatticeHalfSpinColourVectorF;
-typedef Lattice<vHalfSpinColourVectorD>  LatticeHalfSpinColourVectorD;
+typedef Lattice<vHalfSpinColourVectorD> LatticeHalfSpinColourVectorD;
-typedef Lattice<vHalfSpinColourVectorD2> LatticeHalfSpinColourVectorD2;
 
 typedef Lattice<vTReal>            LatticeReal;
 typedef Lattice<vTRealF>           LatticeRealF;
@@ -359,7 +320,6 @@ typedef Lattice<vTRealD>           LatticeRealD;
 typedef Lattice<vTComplex>         LatticeComplex;
 typedef Lattice<vTComplexF>        LatticeComplexF;
 typedef Lattice<vTComplexD>        LatticeComplexD;
-typedef Lattice<vTComplexD2>       LatticeComplexD2;
 
 typedef Lattice<vTInteger>         LatticeInteger; // Predicates for "where"
 
@@ -367,42 +327,37 @@ typedef Lattice<vTInteger>         LatticeInteger; // Predicates for "where"
 ///////////////////////////////////////////
 // Physical names for things
 ///////////////////////////////////////////
-typedef LatticeHalfSpinColourVector   LatticeHalfFermion;
+typedef LatticeHalfSpinColourVector  LatticeHalfFermion;
-typedef LatticeHalfSpinColourVectorF  LatticeHalfFermionF;
+typedef LatticeHalfSpinColourVectorF LatticeHalfFermionF;
-typedef LatticeHalfSpinColourVectorD  LatticeHalfFermionD;
+typedef LatticeHalfSpinColourVectorF LatticeHalfFermionD;
-typedef LatticeHalfSpinColourVectorD2 LatticeHalfFermionD2;
 
 typedef LatticeSpinColourVector      LatticeFermion;
 typedef LatticeSpinColourVectorF     LatticeFermionF;
 typedef LatticeSpinColourVectorD     LatticeFermionD;
-typedef LatticeSpinColourVectorD2    LatticeFermionD2;
 
 typedef LatticeSpinColourMatrix                LatticePropagator;
 typedef LatticeSpinColourMatrixF               LatticePropagatorF;
 typedef LatticeSpinColourMatrixD               LatticePropagatorD;
-typedef LatticeSpinColourMatrixD2              LatticePropagatorD2;
 
 typedef LatticeLorentzColourMatrix             LatticeGaugeField;
 typedef LatticeLorentzColourMatrixF            LatticeGaugeFieldF;
 typedef LatticeLorentzColourMatrixD            LatticeGaugeFieldD;
-typedef LatticeLorentzColourMatrixD2           LatticeGaugeFieldD2;
 
 typedef LatticeDoubleStoredColourMatrix        LatticeDoubledGaugeField;
 typedef LatticeDoubleStoredColourMatrixF       LatticeDoubledGaugeFieldF;
 typedef LatticeDoubleStoredColourMatrixD       LatticeDoubledGaugeFieldD;
-typedef LatticeDoubleStoredColourMatrixD2      LatticeDoubledGaugeFieldD2;
 
 template<class GF> using LorentzScalar = Lattice<iScalar<typename GF::vector_object::element> >;
 
+// Uhgg... typing this hurt  ;)
+// (my keyboard got burning hot when I typed this, must be the anti-Fermion)
 typedef Lattice<vColourVector>          LatticeStaggeredFermion;    
 typedef Lattice<vColourVectorF>         LatticeStaggeredFermionF;    
 typedef Lattice<vColourVectorD>         LatticeStaggeredFermionD;    
-typedef Lattice<vColourVectorD2>        LatticeStaggeredFermionD2;    
 
 typedef Lattice<vColourMatrix>          LatticeStaggeredPropagator; 
 typedef Lattice<vColourMatrixF>         LatticeStaggeredPropagatorF; 
 typedef Lattice<vColourMatrixD>         LatticeStaggeredPropagatorD; 
-typedef Lattice<vColourMatrixD2>        LatticeStaggeredPropagatorD2; 
 
 //////////////////////////////////////////////////////////////////////////////
 // Peek and Poke named after physics attributes
@@ -521,20 +476,9 @@ template<class vobj> void pokeLorentz(vobj &lhs,const decltype(peekIndex<Lorentz
 // Fermion <-> propagator assignements
 //////////////////////////////////////////////
 //template <class Prop, class Ferm>
-#define FAST_FERM_TO_PROP
 template <class Fimpl>
 void FermToProp(typename Fimpl::PropagatorField &p, const typename Fimpl::FermionField &f, const int s, const int c)
 {
-#ifdef FAST_FERM_TO_PROP
-  autoView(p_v,p,CpuWrite);
-  autoView(f_v,f,CpuRead);
-  thread_for(idx,p_v.oSites(),{
-      for(int ss = 0; ss < Ns; ++ss) {
-      for(int cc = 0; cc < Fimpl::Dimension; ++cc) {
-	p_v[idx]()(ss,s)(cc,c) = f_v[idx]()(ss)(cc); // Propagator sink index is LEFT, suitable for left mult by gauge link (e.g.)
-      }}
-    });
-#else
   for(int j = 0; j < Ns; ++j)
     {
       auto pjs = peekSpin(p, j, s);
@@ -546,23 +490,12 @@ void FermToProp(typename Fimpl::PropagatorField &p, const typename Fimpl::Fermio
 	}
       pokeSpin(p, pjs, j, s);
     }
-#endif
 }
     
 //template <class Prop, class Ferm>
 template <class Fimpl>
 void PropToFerm(typename Fimpl::FermionField &f, const typename Fimpl::PropagatorField &p, const int s, const int c)
 {
-#ifdef FAST_FERM_TO_PROP
-  autoView(p_v,p,CpuRead);
-  autoView(f_v,f,CpuWrite);
-  thread_for(idx,p_v.oSites(),{
-      for(int ss = 0; ss < Ns; ++ss) {
-      for(int cc = 0; cc < Fimpl::Dimension; ++cc) {
-	f_v[idx]()(ss)(cc) = p_v[idx]()(ss,s)(cc,c); // LEFT index is copied across for s,c right index
-      }}
-    });
-#else
   for(int j = 0; j < Ns; ++j)
     {
       auto pjs = peekSpin(p, j, s);
@@ -574,7 +507,6 @@ void PropToFerm(typename Fimpl::FermionField &f, const typename Fimpl::Propagato
 	}
       pokeSpin(f, fj, j);
     }
-#endif
 }
     
 //////////////////////////////////////////////

Grid/qcd/action/ActionBase.h

@@ -34,117 +34,21 @@ 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;
-  RealD deriv_max_sum;
-  RealD Fdt_norm_sum;
-  RealD Fdt_max_sum;
-  int   deriv_num;
-  RealD deriv_us;
-  RealD S_us;
-  RealD refresh_us;
-  void  reset_timer(void)        {
-    deriv_us = S_us = refresh_us = 0.0;
-    deriv_norm_sum = deriv_max_sum=0.0;
-    Fdt_max_sum =  Fdt_norm_sum = 0.0;
-    deriv_num=0;
-  }
-  void  deriv_log(RealD nrm, RealD max,RealD Fdt_nrm,RealD Fdt_max) {
-    if ( max > deriv_max_sum ) {
-      deriv_max_sum=max;
-    }
-    deriv_norm_sum+=nrm;
-    if ( Fdt_max > Fdt_max_sum ) {
-      Fdt_max_sum=Fdt_max;
-    }
-    Fdt_norm_sum+=Fdt_nrm; deriv_num++;
-  }
-  RealD deriv_max_average(void)       { return deriv_max_sum; };
-  RealD deriv_norm_average(void)      { return deriv_norm_sum/deriv_num; };
-  RealD Fdt_max_average(void)         { return Fdt_max_sum; };
-  RealD Fdt_norm_average(void)        { return Fdt_norm_sum/deriv_num; };
-  RealD deriv_timer(void)        { return deriv_us; };
-  RealD S_timer(void)            { return S_us; };
-  RealD refresh_timer(void)      { return refresh_us; };
-  void deriv_timer_start(void)   { deriv_us-=usecond(); }
-  void deriv_timer_stop(void)    { deriv_us+=usecond(); }
-  void refresh_timer_start(void) { refresh_us-=usecond(); }
-  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

@@ -30,8 +30,6 @@ 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>
@@ -39,10 +37,6 @@ NAMESPACE_CHECK(ActionSet);
 #include <Grid/qcd/action/ActionParams.h>
 NAMESPACE_CHECK(ActionParams);
 
-#include <Grid/qcd/action/filters/MomentumFilter.h>
-#include <Grid/qcd/action/filters/DirichletFilter.h>
-#include <Grid/qcd/action/filters/DDHMCFilter.h>
-
 ////////////////////////////////////////////
 // Gauge Actions
 ////////////////////////////////////////////

Grid/qcd/action/ActionParams.h

@@ -34,45 +34,28 @@ directory
 
 NAMESPACE_BEGIN(Grid);
 
-
+// These can move into a params header and be given MacroMagic serialisation
 struct GparityWilsonImplParams {
-  Coordinate twists;
+  Coordinate twists; //Here the first Nd-1 directions are treated as "spatial", and a twist value of 1 indicates G-parity BCs in that direction. 
                      //mu=Nd-1 is assumed to be the time direction and a twist value of 1 indicates antiperiodic BCs
-  Coordinate dirichlet; // Blocksize of dirichlet BCs
+  GparityWilsonImplParams() : twists(Nd, 0) {};
-  int  partialDirichlet;
-  GparityWilsonImplParams() : twists(Nd, 0) {
-    dirichlet.resize(0);
-    partialDirichlet=0;
-  };
 };
   
 struct WilsonImplParams {
   bool overlapCommsCompute;
-  Coordinate dirichlet; // Blocksize of dirichlet BCs
-  int  partialDirichlet;
   AcceleratorVector<Real,Nd> twist_n_2pi_L;
   AcceleratorVector<Complex,Nd> boundary_phases;
   WilsonImplParams()  {
-    dirichlet.resize(0);
-    partialDirichlet=0;
     boundary_phases.resize(Nd, 1.0);
       twist_n_2pi_L.resize(Nd, 0.0);
   };
   WilsonImplParams(const AcceleratorVector<Complex,Nd> phi) : boundary_phases(phi), overlapCommsCompute(false) {
     twist_n_2pi_L.resize(Nd, 0.0);
-    partialDirichlet=0;
-    dirichlet.resize(0);
   }
 };
 
 struct StaggeredImplParams {
-  Coordinate dirichlet; // Blocksize of dirichlet BCs
+  StaggeredImplParams()  {};
-  int  partialDirichlet;
-  StaggeredImplParams()
-  {
-    partialDirichlet=0;
-    dirichlet.resize(0);
-  };
 };
   
   struct OneFlavourRationalParams : Serializable {
@@ -81,7 +64,6 @@ struct StaggeredImplParams {
 				    RealD, hi, 
 				    int,   MaxIter, 
 				    RealD, tolerance, 
-				    RealD, mdtolerance, 
 				    int,   degree, 
 				    int,   precision,
 				    int,   BoundsCheckFreq,
@@ -97,19 +79,18 @@ struct StaggeredImplParams {
                            	int _degree    = 10,
 				int _precision = 64,
 				int _BoundsCheckFreq=20,
-				RealD mdtol    = 1.0e-6,
 				double _BoundsCheckTol=1e-6)
       : lo(_lo),
 	hi(_hi),
 	MaxIter(_maxit),
 	tolerance(tol),
-        mdtolerance(mdtol),
 	degree(_degree),
         precision(_precision),
         BoundsCheckFreq(_BoundsCheckFreq),
         BoundsCheckTol(_BoundsCheckTol){};
   };
-  
+
+
   /*Action parameters for the generalized rational action
     The approximation is for (M^dag M)^{1/inv_pow}
     where inv_pow is the denominator of the fractional power.
@@ -152,6 +133,7 @@ struct StaggeredImplParams {
   };
 
 
+  
 NAMESPACE_END(Grid);
 
 #endif

Grid/qcd/action/fermion/CayleyFermion5D.h

@@ -68,17 +68,9 @@ public:
   ///////////////////////////////////////////////////////////////
   // Support for MADWF tricks
   ///////////////////////////////////////////////////////////////
-  RealD Mass(void) { return (mass_plus + mass_minus) / 2.0; };
+  RealD Mass(void) { return mass; };
-  RealD MassPlus(void) { return mass_plus; };
-  RealD MassMinus(void) { return mass_minus; };
-
   void  SetMass(RealD _mass) { 
-    mass_plus=mass_minus=_mass; 
+    mass=_mass; 
-    SetCoefficientsInternal(_zolo_hi,_gamma,_b,_c);  // Reset coeffs
-  } ;
-  void  SetMass(RealD _mass_plus, RealD _mass_minus) { 
-    mass_plus=_mass_plus;
-    mass_minus=_mass_minus;
     SetCoefficientsInternal(_zolo_hi,_gamma,_b,_c);  // Reset coeffs
   } ;
   void  P(const FermionField &psi, FermionField &chi);
@@ -116,7 +108,7 @@ public:
   void   MeooeDag5D    (const FermionField &in, FermionField &out);
 
   //    protected:
-  RealD mass_plus, mass_minus;
+  RealD mass;
 
   // Save arguments to SetCoefficientsInternal
   Vector<Coeff_t> _gamma;
@@ -183,6 +175,16 @@ public:
 		  GridRedBlackCartesian &FourDimRedBlackGrid,
 		  RealD _mass,RealD _M5,const ImplParams &p= ImplParams());
 
+  void CayleyReport(void);
+  void CayleyZeroCounters(void);
+
+  double M5Dflops;
+  double M5Dcalls;
+  double M5Dtime;
+
+  double MooeeInvFlops;
+  double MooeeInvCalls;
+  double MooeeInvTime;
 
 protected:
   virtual void SetCoefficientsZolotarev(RealD zolohi,Approx::zolotarev_data *zdata,RealD b,RealD c);

Grid/qcd/action/fermion/CloverHelpers.h

@@ -1,334 +0,0 @@
-/*************************************************************************************
-
-    Grid physics library, www.github.com/paboyle/Grid
-
-    Source file: ./lib/qcd/action/fermion/WilsonCloverFermionImplementation.h
-
-    Copyright (C) 2017 - 2022
-
-    Author: paboyle <paboyle@ph.ed.ac.uk>
-    Author: Daniel Richtmann <daniel.richtmann@gmail.com>
-    Author: Mattia Bruno <mattia.bruno@cern.ch>
-
-    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/Grid.h>
-#include <Grid/qcd/spin/Dirac.h>
-#include <Grid/qcd/action/fermion/WilsonCloverHelpers.h>
-
-////////////////////////////////////////////
-// Standard Clover
-//   (4+m0) + csw * clover_term
-// Exp Clover
-//   (4+m0) * exp(csw/(4+m0) clover_term)
-//   = (4+m0) + csw * clover_term + ...
-////////////////////////////////////////////
-
-NAMESPACE_BEGIN(Grid);
-
-
-//////////////////////////////////
-// Generic Standard Clover
-//////////////////////////////////
-
-template<class Impl>
-class CloverHelpers: public WilsonCloverHelpers<Impl> {
-public:
-
-  INHERIT_IMPL_TYPES(Impl);
-  INHERIT_CLOVER_TYPES(Impl);
-
-  typedef WilsonCloverHelpers<Impl> Helpers;
-
-  static void Instantiate(CloverField& CloverTerm, CloverField& CloverTermInv, RealD csw_t, RealD diag_mass) {
-    GridBase *grid = CloverTerm.Grid();
-    CloverTerm += diag_mass;
-
-    int lvol = grid->lSites();
-    int DimRep = Impl::Dimension;
-    {
-      autoView(CTv,CloverTerm,CpuRead);
-      autoView(CTIv,CloverTermInv,CpuWrite);
-      thread_for(site, lvol, {
-        Coordinate lcoor;
-        grid->LocalIndexToLocalCoor(site, lcoor);
-        Eigen::MatrixXcd EigenCloverOp = Eigen::MatrixXcd::Zero(Ns * DimRep, Ns * DimRep);
-        Eigen::MatrixXcd EigenInvCloverOp = Eigen::MatrixXcd::Zero(Ns * DimRep, Ns * DimRep);
-        typename SiteClover::scalar_object Qx = Zero(), Qxinv = Zero();
-        peekLocalSite(Qx, CTv, lcoor);
-
-        for (int j = 0; j < Ns; j++)
-          for (int k = 0; k < Ns; k++)
-            for (int a = 0; a < DimRep; a++)
-              for (int b = 0; b < DimRep; b++){
-                auto zz =  Qx()(j, k)(a, b);
-                EigenCloverOp(a + j * DimRep, b + k * DimRep) = std::complex<double>(zz);
-              }
-
-        EigenInvCloverOp = EigenCloverOp.inverse();
-        for (int j = 0; j < Ns; j++)
-          for (int k = 0; k < Ns; k++)
-            for (int a = 0; a < DimRep; a++)
-              for (int b = 0; b < DimRep; b++)
-                Qxinv()(j, k)(a, b) = EigenInvCloverOp(a + j * DimRep, b + k * DimRep);
-               pokeLocalSite(Qxinv, CTIv, lcoor);
-      });
-    }
-  }
-
-  static GaugeLinkField Cmunu(std::vector<GaugeLinkField> &U, GaugeLinkField &lambda, int mu, int nu) {
-    return Helpers::Cmunu(U, lambda, mu, nu);
-  }
-
-};
-
-
-//////////////////////////////////
-// Generic Exp Clover
-//////////////////////////////////
-
-template<class Impl>
-class ExpCloverHelpers: public WilsonCloverHelpers<Impl> {
-public:
-
-  INHERIT_IMPL_TYPES(Impl);
-  INHERIT_CLOVER_TYPES(Impl);
-
-  template <typename vtype> using iImplClover = iScalar<iMatrix<iMatrix<vtype, Impl::Dimension>, Ns>>;
-  typedef WilsonCloverHelpers<Impl> Helpers;
-
-  // Can this be avoided?
-  static void IdentityTimesC(const CloverField& in, RealD c) {
-    int DimRep = Impl::Dimension;
-
-    autoView(in_v, in, AcceleratorWrite);
-
-    accelerator_for(ss, in.Grid()->oSites(), 1, {
-      for (int sa=0; sa<Ns; sa++)
-        for (int ca=0; ca<DimRep; ca++)
-          in_v[ss]()(sa,sa)(ca,ca) = c;
-    });
-  }
-
-  static int getNMAX(RealD prec, RealD R) {
-    /* compute stop condition for exponential */
-    int NMAX=1;
-    RealD cond=R*R/2.;
-
-    while (cond*std::exp(R)>prec) {
-      NMAX++;
-      cond*=R/(double)(NMAX+1);
-    }
-    return NMAX;
-  }
-
-  static int getNMAX(Lattice<iImplClover<vComplexD2>> &t, RealD R) {return getNMAX(1e-12,R);}
-  static int getNMAX(Lattice<iImplClover<vComplexD>> &t, RealD R) {return getNMAX(1e-12,R);}
-  static int getNMAX(Lattice<iImplClover<vComplexF>> &t, RealD R) {return getNMAX(1e-6,R);}
-
-  static void Instantiate(CloverField& Clover, CloverField& CloverInv, RealD csw_t, RealD diag_mass) {
-    GridBase* grid = Clover.Grid();
-    CloverField ExpClover(grid);
-
-    int NMAX = getNMAX(Clover, 3.*csw_t/diag_mass);
-
-    Clover *= (1.0/diag_mass);
-
-    // Taylor expansion, slow but generic
-    // Horner scheme: a0 + a1 x + a2 x^2 + .. = a0 + x (a1 + x(...))
-    // qN = cN
-    // qn = cn + qn+1 X
-    std::vector<RealD> cn(NMAX+1);
-    cn[0] = 1.0;
-    for (int i=1; i<=NMAX; i++)
-      cn[i] = cn[i-1] / RealD(i);
-
-    ExpClover = Zero();
-    IdentityTimesC(ExpClover, cn[NMAX]);
-    for (int i=NMAX-1; i>=0; i--)
-      ExpClover = ExpClover * Clover + cn[i];
-
-    // prepare inverse
-    CloverInv = (-1.0)*Clover;
-
-    Clover = ExpClover * diag_mass;
-
-    ExpClover = Zero();
-    IdentityTimesC(ExpClover, cn[NMAX]);
-    for (int i=NMAX-1; i>=0; i--)
-      ExpClover = ExpClover * CloverInv + cn[i];
-
-    CloverInv = ExpClover * (1.0/diag_mass);
-
-  }
-
-  static GaugeLinkField Cmunu(std::vector<GaugeLinkField> &U, GaugeLinkField &lambda, int mu, int nu) {
-    assert(0);
-    return lambda;
-  }
-
-};
-
-
-//////////////////////////////////
-// Compact Standard Clover
-//////////////////////////////////
-
-
-template<class Impl>
-class CompactCloverHelpers: public CompactWilsonCloverHelpers<Impl>,
-                            public WilsonCloverHelpers<Impl> {
-public:
-
-  INHERIT_IMPL_TYPES(Impl);
-  INHERIT_CLOVER_TYPES(Impl);
-  INHERIT_COMPACT_CLOVER_TYPES(Impl);
-
-  typedef WilsonCloverHelpers<Impl> Helpers;
-  typedef CompactWilsonCloverHelpers<Impl> CompactHelpers;
-
-  static void InstantiateClover(CloverField& Clover, CloverField& CloverInv, RealD csw_t, RealD diag_mass) {
-    Clover += diag_mass;
-  }
-
-  static void InvertClover(CloverField& InvClover,
-                            const CloverDiagonalField& diagonal,
-                            const CloverTriangleField& triangle,
-                            CloverDiagonalField&       diagonalInv,
-                            CloverTriangleField&       triangleInv,
-                            bool fixedBoundaries) {
-
-    CompactHelpers::Invert(diagonal, triangle, diagonalInv, triangleInv);
-  }
-
-  // TODO: implement Cmunu for better performances with compact layout, but don't do it
-  // here, but rather in WilsonCloverHelpers.h -> CompactWilsonCloverHelpers
-  static GaugeLinkField Cmunu(std::vector<GaugeLinkField> &U, GaugeLinkField &lambda, int mu, int nu) {
-    return Helpers::Cmunu(U, lambda, mu, nu);
-  }
-};
-
-//////////////////////////////////
-// Compact Exp Clover
-//////////////////////////////////
-
-template<class Impl>
-class CompactExpCloverHelpers: public CompactWilsonCloverHelpers<Impl> {
-public:
-
-  INHERIT_IMPL_TYPES(Impl);
-  INHERIT_CLOVER_TYPES(Impl);
-  INHERIT_COMPACT_CLOVER_TYPES(Impl);
-
-  template <typename vtype> using iImplClover = iScalar<iMatrix<iMatrix<vtype, Impl::Dimension>, Ns>>;
-  typedef CompactWilsonCloverHelpers<Impl> CompactHelpers;
-
-  // Can this be avoided?
-  static void IdentityTimesC(const CloverField& in, RealD c) {
-    int DimRep = Impl::Dimension;
-
-    autoView(in_v, in, AcceleratorWrite);
-
-    accelerator_for(ss, in.Grid()->oSites(), 1, {
-      for (int sa=0; sa<Ns; sa++)
-        for (int ca=0; ca<DimRep; ca++)
-          in_v[ss]()(sa,sa)(ca,ca) = c;
-    });
-  }
-
-  static int getNMAX(RealD prec, RealD R) {
-    /* compute stop condition for exponential */
-    int NMAX=1;
-    RealD cond=R*R/2.;
-
-    while (cond*std::exp(R)>prec) {
-      NMAX++;
-      cond*=R/(double)(NMAX+1);
-    }
-    return NMAX;
-  }
-
-  static int getNMAX(Lattice<iImplClover<vComplexD>> &t, RealD R) {return getNMAX(1e-12,R);}
-  static int getNMAX(Lattice<iImplClover<vComplexF>> &t, RealD R) {return getNMAX(1e-6,R);}
-
-  static void InstantiateClover(CloverField& Clover, CloverField& CloverInv, RealD csw_t, RealD diag_mass) {
-
-    GridBase* grid = Clover.Grid();
-    CloverField ExpClover(grid);
-
-    int NMAX = getNMAX(Clover, 3.*csw_t/diag_mass);
-
-    Clover *= (1.0/diag_mass);
-
-    // Taylor expansion, slow but generic
-    // Horner scheme: a0 + a1 x + a2 x^2 + .. = a0 + x (a1 + x(...))
-    // qN = cN
-    // qn = cn + qn+1 X
-    std::vector<RealD> cn(NMAX+1);
-    cn[0] = 1.0;
-    for (int i=1; i<=NMAX; i++)
-      cn[i] = cn[i-1] / RealD(i);
-
-    ExpClover = Zero();
-    IdentityTimesC(ExpClover, cn[NMAX]);
-    for (int i=NMAX-1; i>=0; i--)
-      ExpClover = ExpClover * Clover + cn[i];
-
-    // prepare inverse
-    CloverInv = (-1.0)*Clover;
-
-    Clover = ExpClover * diag_mass;
-
-    ExpClover = Zero();
-    IdentityTimesC(ExpClover, cn[NMAX]);
-    for (int i=NMAX-1; i>=0; i--)
-      ExpClover = ExpClover * CloverInv + cn[i];
-
-    CloverInv = ExpClover * (1.0/diag_mass);
-
-  }
-
-  static void InvertClover(CloverField& InvClover,
-                            const CloverDiagonalField& diagonal,
-                            const CloverTriangleField& triangle,
-                            CloverDiagonalField&       diagonalInv,
-                            CloverTriangleField&       triangleInv,
-                            bool fixedBoundaries) {
-
-    if (fixedBoundaries)
-    {
-      CompactHelpers::Invert(diagonal, triangle, diagonalInv, triangleInv);
-    }
-    else
-    {
-      CompactHelpers::ConvertLayout(InvClover, diagonalInv, triangleInv);
-    }
-  }
-
-  static GaugeLinkField Cmunu(std::vector<GaugeLinkField> &U, GaugeLinkField &lambda, int mu, int nu) {
-    assert(0);
-    return lambda;
-  }
-
-};
-
-
-NAMESPACE_END(Grid);

Grid/qcd/action/fermion/CompactWilsonCloverFermion.h

@@ -31,7 +31,6 @@
 
 #include <Grid/qcd/action/fermion/WilsonCloverTypes.h>
 #include <Grid/qcd/action/fermion/WilsonCloverHelpers.h>
-#include <Grid/qcd/action/fermion/CloverHelpers.h>
 
 NAMESPACE_BEGIN(Grid);
 
@@ -86,7 +85,7 @@ NAMESPACE_BEGIN(Grid);
 //           + (2 * 1 + 4 * 1/2) triangle parts = 4 triangle parts =  60 complex words per site
 //                                                                 =  84 complex words per site
 
-template<class Impl, class CloverHelpers>
+template<class Impl>
 class CompactWilsonCloverFermion : public WilsonFermion<Impl>,
                                    public WilsonCloverHelpers<Impl>,
                                    public CompactWilsonCloverHelpers<Impl> {
@@ -225,7 +224,7 @@ public:
   RealD csw_t;
   RealD cF;
 
-  bool fixedBoundaries;
+  bool open_boundaries;
 
   CloverDiagonalField Diagonal,    DiagonalEven,    DiagonalOdd;
   CloverDiagonalField DiagonalInv, DiagonalInvEven, DiagonalInvOdd;

Grid/qcd/action/fermion/DWFSlow.h

@@ -1,291 +0,0 @@
-/*************************************************************************************
-
-Grid physics library, www.github.com/paboyle/Grid
-
-Source file: ./lib/qcd/action/fermion/DWFSlow.h
-
-Copyright (C) 2022
-
-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);
-
-template <class Impl>
-class DWFSlowFermion : public FermionOperator<Impl>
-{
-public:
-  INHERIT_IMPL_TYPES(Impl);
-
-  ///////////////////////////////////////////////////////////////
-  // Implement the abstract base
-  ///////////////////////////////////////////////////////////////
-  GridBase *GaugeGrid(void) { return _grid4; }
-  GridBase *GaugeRedBlackGrid(void) { return _cbgrid4; }
-  GridBase *FermionGrid(void) { return _grid; }
-  GridBase *FermionRedBlackGrid(void) { return _cbgrid; }
-
-  FermionField _tmp;
-  FermionField &tmp(void) { return _tmp; }
-
-  //////////////////////////////////////////////////////////////////
-  // override multiply; cut number routines if pass dagger argument
-  // and also make interface more uniformly consistent
-  //////////////////////////////////////////////////////////////////
-  virtual void  M(const FermionField &in, FermionField &out)
-  {
-    FermionField tmp(_grid);
-    out = (5.0 - M5) * in;
-    Dhop(in,tmp,DaggerNo);
-    out = out + tmp;
-  }
-  virtual void  Mdag(const FermionField &in, FermionField &out)
-  {
-    FermionField tmp(_grid);
-    out = (5.0 - M5) * in;
-    Dhop(in,tmp,DaggerYes);
-    out = out + tmp;
-  };
-
-  /////////////////////////////////////////////////////////
-  // half checkerboard operations 5D redblack so just site identiy
-  /////////////////////////////////////////////////////////
-  void Meooe(const FermionField &in, FermionField &out)
-  {
-    if ( in.Checkerboard() == Odd ) {
-      this->DhopEO(in,out,DaggerNo);
-    } else {
-      this->DhopOE(in,out,DaggerNo);
-    }
-  }
-  void MeooeDag(const FermionField &in, FermionField &out)
-  {
-    if ( in.Checkerboard() == Odd ) {
-      this->DhopEO(in,out,DaggerYes);
-    } else {
-      this->DhopOE(in,out,DaggerYes);
-    }
-  };
-
-  // allow override for twisted mass and clover
-  virtual void Mooee(const FermionField &in, FermionField &out)
-  {
-    out = (5.0 - M5) * in;
-  }
-  virtual void MooeeDag(const FermionField &in, FermionField &out)
-  {
-    out = (5.0 - M5) * in;
-  }
-  virtual void MooeeInv(const FermionField &in, FermionField &out)
-  {
-    out = (1.0/(5.0 - M5)) * in;
-  };
-  virtual void MooeeInvDag(const FermionField &in, FermionField &out)
-  {
-    out = (1.0/(5.0 - M5)) * in;
-  };
-
-  virtual void  MomentumSpacePropagator(FermionField &out,const FermionField &in,RealD _mass,std::vector<double> twist) {} ;
-
-  ////////////////////////
-  // Derivative interface
-  ////////////////////////
-  // Interface calls an internal routine
-  void DhopDeriv(GaugeField &mat,const FermionField &U,const FermionField &V,int dag)  { assert(0);};
-  void DhopDerivOE(GaugeField &mat,const FermionField &U,const FermionField &V,int dag){ assert(0);};
-  void DhopDerivEO(GaugeField &mat,const FermionField &U,const FermionField &V,int dag){ assert(0);};
-
-  ///////////////////////////////////////////////////////////////
-  // non-hermitian hopping term; half cb or both
-  ///////////////////////////////////////////////////////////////
-  void Dhop(const FermionField &in, FermionField &out, int dag)
-  {
-    FermionField tmp(in.Grid());
-    Dhop5(in,out,MassField,MassField,dag );
-    for(int mu=0;mu<4;mu++){
-      DhopDirU(in,Umu[mu],Umu[mu],tmp,mu,dag );    out = out + tmp;
-    }
-  };
-  void DhopOE(const FermionField &in, FermionField &out, int dag)
-  {
-    FermionField tmp(in.Grid());
-    assert(in.Checkerboard()==Even);
-    Dhop5(in,out,MassFieldOdd,MassFieldEven,dag);
-    for(int mu=0;mu<4;mu++){
-      DhopDirU(in,UmuOdd[mu],UmuEven[mu],tmp,mu,dag );    out = out + tmp;
-    }
-  };
-  void DhopEO(const FermionField &in, FermionField &out, int dag)
-  {
-    FermionField tmp(in.Grid());
-    assert(in.Checkerboard()==Odd);
-    Dhop5(in,out, MassFieldEven,MassFieldOdd ,dag );  
-    for(int mu=0;mu<4;mu++){
-      DhopDirU(in,UmuEven[mu],UmuOdd[mu],tmp,mu,dag );    out = out + tmp;
-    }
-  };
-
-  ///////////////////////////////////////////////////////////////
-  // Multigrid assistance; force term uses too
-  ///////////////////////////////////////////////////////////////
-  void Mdir(const FermionField &in, FermionField &out, int dir, int disp){ assert(0);};
-  void MdirAll(const FermionField &in, std::vector<FermionField> &out)   { assert(0);};
-  void DhopDir(const FermionField &in, FermionField &out, int dir, int disp) { assert(0);};
-  void DhopDirAll(const FermionField &in, std::vector<FermionField> &out)    { assert(0);};
-  void DhopDirCalc(const FermionField &in, FermionField &out, int dirdisp,int gamma, int dag) { assert(0);};
-
-  void DhopDirU(const FermionField &in, const GaugeLinkField &U5e, const GaugeLinkField &U5o, FermionField &out, int mu, int dag)
-  {
-    RealD     sgn= 1.0;
-    if (dag ) sgn=-1.0;
-
-    Gamma::Algebra Gmu [] = {
-			 Gamma::Algebra::GammaX,
-			 Gamma::Algebra::GammaY,
-			 Gamma::Algebra::GammaZ,
-			 Gamma::Algebra::GammaT
-    };
-
-    //    mass is  1,1,1,1,-m has to multiply the round the world term
-    FermionField tmp (in.Grid());
-    tmp = U5e * Cshift(in,mu+1,1);
-    out = tmp - Gamma(Gmu[mu])*tmp*sgn;
-    
-    tmp = Cshift(adj(U5o)*in,mu+1,-1);
-    out = out + tmp + Gamma(Gmu[mu])*tmp*sgn;
-
-    out = -0.5*out;
-  };
-
-  void Dhop5(const FermionField &in, FermionField &out, ComplexField &massE, ComplexField &massO, int dag)
-  {
-    // Mass term.... must multiple the round world with mass = 1,1,1,1, -m
-    RealD     sgn= 1.0;
-    if (dag ) sgn=-1.0;
-
-    Gamma G5(Gamma::Algebra::Gamma5);
-
-    FermionField tmp (in.Grid());
-    tmp = massE*Cshift(in,0,1);
-    out = tmp - G5*tmp*sgn;
-    
-    tmp = Cshift(massO*in,0,-1);
-    out = out + tmp + G5*tmp*sgn;
-    out = -0.5*out;
-  };
-
-  // Constructor
-  DWFSlowFermion(GaugeField &_Umu, GridCartesian &Fgrid,
-		 GridRedBlackCartesian &Hgrid, RealD _mass, RealD _M5)
-    :
-    _grid(&Fgrid),
-    _cbgrid(&Hgrid),
-    _grid4(_Umu.Grid()),
-    Umu(Nd,&Fgrid),
-    UmuEven(Nd,&Hgrid),
-    UmuOdd(Nd,&Hgrid),
-    MassField(&Fgrid),
-    MassFieldEven(&Hgrid),
-    MassFieldOdd(&Hgrid),
-    M5(_M5),
-    mass(_mass),
-    _tmp(&Hgrid)
-    {
-      Ls=Fgrid._fdimensions[0];
-      ImportGauge(_Umu);
-
-      typedef typename FermionField::scalar_type scalar;
-
-      Lattice<iScalar<vInteger> > coor(&Fgrid);
-      LatticeCoordinate(coor, 0); // Scoor
-      ComplexField one(&Fgrid);
-      MassField =scalar(-mass);
-      one       =scalar(1.0);
-      MassField =where(coor==Integer(Ls-1),MassField,one);
-      for(int mu=0;mu<Nd;mu++){
-	pickCheckerboard(Even,UmuEven[mu],Umu[mu]);
-	pickCheckerboard(Odd ,UmuOdd[mu],Umu[mu]);
-      }
-      pickCheckerboard(Even,MassFieldEven,MassField);
-      pickCheckerboard(Odd ,MassFieldOdd,MassField);
-      
-    }
-  
-  // DoubleStore impl dependent
-  void ImportGauge(const GaugeField &_Umu4)
-  {
-    GaugeLinkField U4(_grid4);
-    for(int mu=0;mu<Nd;mu++){
-      U4 = PeekIndex<LorentzIndex>(_Umu4, mu);
-      for(int s=0;s<this->Ls;s++){
-	InsertSlice(U4,Umu[mu],s,0);
-      }
-    }
-  }
-
-  ///////////////////////////////////////////////////////////////
-  // Data members require to support the functionality
-  ///////////////////////////////////////////////////////////////
-
-public:
-  virtual RealD Mass(void) { return mass; }
-  virtual int   isTrivialEE(void) { return 1; };
-  RealD mass;
-  RealD M5;
-  int Ls;
-
-  GridBase *_grid4;
-  GridBase *_grid;
-  GridBase *_cbgrid4;
-  GridBase *_cbgrid;
-
-  // Copy of the gauge field , with even and odd subsets
-  std::vector<GaugeLinkField> Umu;
-  std::vector<GaugeLinkField> UmuEven;
-  std::vector<GaugeLinkField> UmuOdd;
-  ComplexField MassField;
-  ComplexField MassFieldEven;
-  ComplexField MassFieldOdd;
-
-  ///////////////////////////////////////////////////////////////
-  // Conserved current utilities
-  ///////////////////////////////////////////////////////////////
-  void ContractConservedCurrent(PropagatorField &q_in_1,
-                                PropagatorField &q_in_2,
-                                PropagatorField &q_out,
-                                PropagatorField &phys_src,
-                                Current curr_type,
-                                unsigned int mu){}
-  void SeqConservedCurrent(PropagatorField &q_in,
-                           PropagatorField &q_out,
-                           PropagatorField &phys_src,
-                           Current curr_type,
-                           unsigned int mu,
-                           unsigned int tmin,
-			   unsigned int tmax,
-			   ComplexField &lattice_cmplx){}
-};
-
-typedef DWFSlowFermion<WilsonImplF> DWFSlowFermionF;
-typedef DWFSlowFermion<WilsonImplD> DWFSlowFermionD;
-
-NAMESPACE_END(Grid);

Grid/qcd/action/fermion/Fermion.h

@@ -47,7 +47,6 @@ Author: Peter Boyle <pabobyle@ph.ed.ac.uk>
 ////////////////////////////////////////////
 // Fermion operators / actions
 ////////////////////////////////////////////
-#include <Grid/qcd/action/fermion/DWFSlow.h>       // Slow DWF
 
 #include <Grid/qcd/action/fermion/WilsonFermion.h>       // 4d wilson like
 NAMESPACE_CHECK(Wilson);
@@ -113,161 +112,200 @@ NAMESPACE_CHECK(DWFutils);
 // Cayley 5d
 NAMESPACE_BEGIN(Grid);
 
-typedef WilsonFermion<WilsonImplD2> WilsonFermionD2;
+typedef WilsonFermion<WilsonImplR> WilsonFermionR;
 typedef WilsonFermion<WilsonImplF> WilsonFermionF;
 typedef WilsonFermion<WilsonImplD> WilsonFermionD;
 
+//typedef WilsonFermion<WilsonImplRL> WilsonFermionRL;
+//typedef WilsonFermion<WilsonImplFH> WilsonFermionFH;
+//typedef WilsonFermion<WilsonImplDF> WilsonFermionDF;
+
+typedef WilsonFermion<WilsonAdjImplR> WilsonAdjFermionR;
 typedef WilsonFermion<WilsonAdjImplF> WilsonAdjFermionF;
 typedef WilsonFermion<WilsonAdjImplD> WilsonAdjFermionD;
 
+typedef WilsonFermion<WilsonTwoIndexSymmetricImplR> WilsonTwoIndexSymmetricFermionR;
 typedef WilsonFermion<WilsonTwoIndexSymmetricImplF> WilsonTwoIndexSymmetricFermionF;
 typedef WilsonFermion<WilsonTwoIndexSymmetricImplD> WilsonTwoIndexSymmetricFermionD;
 
+typedef WilsonFermion<WilsonTwoIndexAntiSymmetricImplR> WilsonTwoIndexAntiSymmetricFermionR;
 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<WilsonImplR> WilsonTMFermionR;
 typedef WilsonTMFermion<WilsonImplF> WilsonTMFermionF;
 typedef WilsonTMFermion<WilsonImplD> WilsonTMFermionD;
 
 // Clover fermions
-template <typename WImpl> using WilsonClover = WilsonCloverFermion<WImpl, CloverHelpers<WImpl>>;
+typedef WilsonCloverFermion<WilsonImplR> WilsonCloverFermionR;
-template <typename WImpl> using WilsonExpClover = WilsonCloverFermion<WImpl, ExpCloverHelpers<WImpl>>;
+typedef WilsonCloverFermion<WilsonImplF> WilsonCloverFermionF;
+typedef WilsonCloverFermion<WilsonImplD> WilsonCloverFermionD;
 
-typedef WilsonClover<WilsonImplD2> WilsonCloverFermionD2;
+typedef WilsonCloverFermion<WilsonAdjImplR> WilsonCloverAdjFermionR;
-typedef WilsonClover<WilsonImplF> WilsonCloverFermionF;
+typedef WilsonCloverFermion<WilsonAdjImplF> WilsonCloverAdjFermionF;
-typedef WilsonClover<WilsonImplD> WilsonCloverFermionD;
+typedef WilsonCloverFermion<WilsonAdjImplD> WilsonCloverAdjFermionD;
 
-typedef WilsonExpClover<WilsonImplD2> WilsonExpCloverFermionD2;
+typedef WilsonCloverFermion<WilsonTwoIndexSymmetricImplR> WilsonCloverTwoIndexSymmetricFermionR;
-typedef WilsonExpClover<WilsonImplF> WilsonExpCloverFermionF;
+typedef WilsonCloverFermion<WilsonTwoIndexSymmetricImplF> WilsonCloverTwoIndexSymmetricFermionF;
-typedef WilsonExpClover<WilsonImplD> WilsonExpCloverFermionD;
+typedef WilsonCloverFermion<WilsonTwoIndexSymmetricImplD> WilsonCloverTwoIndexSymmetricFermionD;
 
-typedef WilsonClover<WilsonAdjImplF> WilsonCloverAdjFermionF;
+typedef WilsonCloverFermion<WilsonTwoIndexAntiSymmetricImplR> WilsonCloverTwoIndexAntiSymmetricFermionR;
-typedef WilsonClover<WilsonAdjImplD> WilsonCloverAdjFermionD;
+typedef WilsonCloverFermion<WilsonTwoIndexAntiSymmetricImplF> WilsonCloverTwoIndexAntiSymmetricFermionF;
-
+typedef WilsonCloverFermion<WilsonTwoIndexAntiSymmetricImplD> WilsonCloverTwoIndexAntiSymmetricFermionD;
-typedef WilsonClover<WilsonTwoIndexSymmetricImplF> WilsonCloverTwoIndexSymmetricFermionF;
-typedef WilsonClover<WilsonTwoIndexSymmetricImplD> WilsonCloverTwoIndexSymmetricFermionD;
-
-typedef WilsonClover<WilsonTwoIndexAntiSymmetricImplF> WilsonCloverTwoIndexAntiSymmetricFermionF;
-typedef WilsonClover<WilsonTwoIndexAntiSymmetricImplD> WilsonCloverTwoIndexAntiSymmetricFermionD;
 
 // Compact Clover fermions
-template <typename WImpl> using CompactWilsonClover = CompactWilsonCloverFermion<WImpl, CompactCloverHelpers<WImpl>>;
+typedef CompactWilsonCloverFermion<WilsonImplR> CompactWilsonCloverFermionR;
-template <typename WImpl> using CompactWilsonExpClover = CompactWilsonCloverFermion<WImpl, CompactExpCloverHelpers<WImpl>>;
+typedef CompactWilsonCloverFermion<WilsonImplF> CompactWilsonCloverFermionF;
+typedef CompactWilsonCloverFermion<WilsonImplD> CompactWilsonCloverFermionD;
 
-typedef CompactWilsonClover<WilsonImplD2> CompactWilsonCloverFermionD2;
+typedef CompactWilsonCloverFermion<WilsonAdjImplR> CompactWilsonCloverAdjFermionR;
-typedef CompactWilsonClover<WilsonImplF> CompactWilsonCloverFermionF;
+typedef CompactWilsonCloverFermion<WilsonAdjImplF> CompactWilsonCloverAdjFermionF;
-typedef CompactWilsonClover<WilsonImplD> CompactWilsonCloverFermionD;
+typedef CompactWilsonCloverFermion<WilsonAdjImplD> CompactWilsonCloverAdjFermionD;
 
-typedef CompactWilsonExpClover<WilsonImplD2> CompactWilsonExpCloverFermionD2;
+typedef CompactWilsonCloverFermion<WilsonTwoIndexSymmetricImplR> CompactWilsonCloverTwoIndexSymmetricFermionR;
-typedef CompactWilsonExpClover<WilsonImplF> CompactWilsonExpCloverFermionF;
+typedef CompactWilsonCloverFermion<WilsonTwoIndexSymmetricImplF> CompactWilsonCloverTwoIndexSymmetricFermionF;
-typedef CompactWilsonExpClover<WilsonImplD> CompactWilsonExpCloverFermionD;
+typedef CompactWilsonCloverFermion<WilsonTwoIndexSymmetricImplD> CompactWilsonCloverTwoIndexSymmetricFermionD;
 
-typedef CompactWilsonClover<WilsonAdjImplF> CompactWilsonCloverAdjFermionF;
+typedef CompactWilsonCloverFermion<WilsonTwoIndexAntiSymmetricImplR> CompactWilsonCloverTwoIndexAntiSymmetricFermionR;
-typedef CompactWilsonClover<WilsonAdjImplD> CompactWilsonCloverAdjFermionD;
+typedef CompactWilsonCloverFermion<WilsonTwoIndexAntiSymmetricImplF> CompactWilsonCloverTwoIndexAntiSymmetricFermionF;
-
+typedef CompactWilsonCloverFermion<WilsonTwoIndexAntiSymmetricImplD> CompactWilsonCloverTwoIndexAntiSymmetricFermionD;
-typedef CompactWilsonClover<WilsonTwoIndexSymmetricImplF> CompactWilsonCloverTwoIndexSymmetricFermionF;
-typedef CompactWilsonClover<WilsonTwoIndexSymmetricImplD> CompactWilsonCloverTwoIndexSymmetricFermionD;
-
-typedef CompactWilsonClover<WilsonTwoIndexAntiSymmetricImplF> CompactWilsonCloverTwoIndexAntiSymmetricFermionF;
-typedef CompactWilsonClover<WilsonTwoIndexAntiSymmetricImplD> CompactWilsonCloverTwoIndexAntiSymmetricFermionD;
 
 // Domain Wall fermions
+typedef DomainWallFermion<WilsonImplR> DomainWallFermionR;
 typedef DomainWallFermion<WilsonImplF> DomainWallFermionF;
 typedef DomainWallFermion<WilsonImplD> DomainWallFermionD;
-typedef DomainWallFermion<WilsonImplD2> DomainWallFermionD2;
 
-typedef DomainWallEOFAFermion<WilsonImplD2> DomainWallEOFAFermionD2;
+//typedef DomainWallFermion<WilsonImplRL> DomainWallFermionRL;
+//typedef DomainWallFermion<WilsonImplFH> DomainWallFermionFH;
+//typedef DomainWallFermion<WilsonImplDF> DomainWallFermionDF;
+
+typedef DomainWallEOFAFermion<WilsonImplR> DomainWallEOFAFermionR;
 typedef DomainWallEOFAFermion<WilsonImplF> DomainWallEOFAFermionF;
 typedef DomainWallEOFAFermion<WilsonImplD> DomainWallEOFAFermionD;
 
-typedef MobiusFermion<WilsonImplD2> MobiusFermionD2;
+//typedef DomainWallEOFAFermion<WilsonImplRL> DomainWallEOFAFermionRL;
+//typedef DomainWallEOFAFermion<WilsonImplFH> DomainWallEOFAFermionFH;
+//typedef DomainWallEOFAFermion<WilsonImplDF> DomainWallEOFAFermionDF;
+
+typedef MobiusFermion<WilsonImplR> MobiusFermionR;
 typedef MobiusFermion<WilsonImplF> MobiusFermionF;
 typedef MobiusFermion<WilsonImplD> MobiusFermionD;
 
-typedef MobiusEOFAFermion<WilsonImplD2> MobiusEOFAFermionD2;
+//typedef MobiusFermion<WilsonImplRL> MobiusFermionRL;
+//typedef MobiusFermion<WilsonImplFH> MobiusFermionFH;
+//typedef MobiusFermion<WilsonImplDF> MobiusFermionDF;
+
+typedef MobiusEOFAFermion<WilsonImplR> MobiusEOFAFermionR;
 typedef MobiusEOFAFermion<WilsonImplF> MobiusEOFAFermionF;
 typedef MobiusEOFAFermion<WilsonImplD> MobiusEOFAFermionD;
 
-typedef ZMobiusFermion<ZWilsonImplD2> ZMobiusFermionD2;
+//typedef MobiusEOFAFermion<WilsonImplRL> MobiusEOFAFermionRL;
+//typedef MobiusEOFAFermion<WilsonImplFH> MobiusEOFAFermionFH;
+//typedef MobiusEOFAFermion<WilsonImplDF> MobiusEOFAFermionDF;
+
+typedef ZMobiusFermion<ZWilsonImplR> ZMobiusFermionR;
 typedef ZMobiusFermion<ZWilsonImplF> ZMobiusFermionF;
 typedef ZMobiusFermion<ZWilsonImplD> ZMobiusFermionD;
 
-typedef ScaledShamirFermion<WilsonImplD2> ScaledShamirFermionD2;
+//typedef ZMobiusFermion<ZWilsonImplRL> ZMobiusFermionRL;
+//typedef ZMobiusFermion<ZWilsonImplFH> ZMobiusFermionFH;
+//typedef ZMobiusFermion<ZWilsonImplDF> ZMobiusFermionDF;
+
+// Ls vectorised
+typedef ScaledShamirFermion<WilsonImplR> ScaledShamirFermionR;
 typedef ScaledShamirFermion<WilsonImplF> ScaledShamirFermionF;
 typedef ScaledShamirFermion<WilsonImplD> ScaledShamirFermionD;
 
-typedef MobiusZolotarevFermion<WilsonImplD2> MobiusZolotarevFermionD2;
+typedef MobiusZolotarevFermion<WilsonImplR> MobiusZolotarevFermionR;
 typedef MobiusZolotarevFermion<WilsonImplF> MobiusZolotarevFermionF;
 typedef MobiusZolotarevFermion<WilsonImplD> MobiusZolotarevFermionD;
-typedef ShamirZolotarevFermion<WilsonImplD2> ShamirZolotarevFermionD2;
+typedef ShamirZolotarevFermion<WilsonImplR> ShamirZolotarevFermionR;
 typedef ShamirZolotarevFermion<WilsonImplF> ShamirZolotarevFermionF;
 typedef ShamirZolotarevFermion<WilsonImplD> ShamirZolotarevFermionD;
 
-typedef OverlapWilsonCayleyTanhFermion<WilsonImplD2> OverlapWilsonCayleyTanhFermionD2;
+typedef OverlapWilsonCayleyTanhFermion<WilsonImplR> OverlapWilsonCayleyTanhFermionR;
 typedef OverlapWilsonCayleyTanhFermion<WilsonImplF> OverlapWilsonCayleyTanhFermionF;
 typedef OverlapWilsonCayleyTanhFermion<WilsonImplD> OverlapWilsonCayleyTanhFermionD;
-typedef OverlapWilsonCayleyZolotarevFermion<WilsonImplD2> OverlapWilsonCayleyZolotarevFermionD2;
+typedef OverlapWilsonCayleyZolotarevFermion<WilsonImplR> OverlapWilsonCayleyZolotarevFermionR;
 typedef OverlapWilsonCayleyZolotarevFermion<WilsonImplF> OverlapWilsonCayleyZolotarevFermionF;
 typedef OverlapWilsonCayleyZolotarevFermion<WilsonImplD> OverlapWilsonCayleyZolotarevFermionD;
 
 // Continued fraction
-typedef OverlapWilsonContFracTanhFermion<WilsonImplD2> OverlapWilsonContFracTanhFermionD2;
+typedef OverlapWilsonContFracTanhFermion<WilsonImplR> OverlapWilsonContFracTanhFermionR;
 typedef OverlapWilsonContFracTanhFermion<WilsonImplF> OverlapWilsonContFracTanhFermionF;
 typedef OverlapWilsonContFracTanhFermion<WilsonImplD> OverlapWilsonContFracTanhFermionD;
-typedef OverlapWilsonContFracZolotarevFermion<WilsonImplD2> OverlapWilsonContFracZolotarevFermionD2;
+typedef OverlapWilsonContFracZolotarevFermion<WilsonImplR> OverlapWilsonContFracZolotarevFermionR;
 typedef OverlapWilsonContFracZolotarevFermion<WilsonImplF> OverlapWilsonContFracZolotarevFermionF;
 typedef OverlapWilsonContFracZolotarevFermion<WilsonImplD> OverlapWilsonContFracZolotarevFermionD;
 
 // Partial fraction
-typedef OverlapWilsonPartialFractionTanhFermion<WilsonImplD2> OverlapWilsonPartialFractionTanhFermionD2;
+typedef OverlapWilsonPartialFractionTanhFermion<WilsonImplR> OverlapWilsonPartialFractionTanhFermionR;
 typedef OverlapWilsonPartialFractionTanhFermion<WilsonImplF> OverlapWilsonPartialFractionTanhFermionF;
 typedef OverlapWilsonPartialFractionTanhFermion<WilsonImplD> OverlapWilsonPartialFractionTanhFermionD;
 
-typedef OverlapWilsonPartialFractionZolotarevFermion<WilsonImplD2> OverlapWilsonPartialFractionZolotarevFermionD2;
+typedef OverlapWilsonPartialFractionZolotarevFermion<WilsonImplR> OverlapWilsonPartialFractionZolotarevFermionR;
 typedef OverlapWilsonPartialFractionZolotarevFermion<WilsonImplF> OverlapWilsonPartialFractionZolotarevFermionF;
 typedef OverlapWilsonPartialFractionZolotarevFermion<WilsonImplD> OverlapWilsonPartialFractionZolotarevFermionD;
 
 // Gparity cases; partial list until tested
+typedef WilsonFermion<GparityWilsonImplR>     GparityWilsonFermionR;
 typedef WilsonFermion<GparityWilsonImplF>     GparityWilsonFermionF;
 typedef WilsonFermion<GparityWilsonImplD>     GparityWilsonFermionD;
 
+//typedef WilsonFermion<GparityWilsonImplRL>     GparityWilsonFermionRL;
+//typedef WilsonFermion<GparityWilsonImplFH>     GparityWilsonFermionFH;
+//typedef WilsonFermion<GparityWilsonImplDF>     GparityWilsonFermionDF;
+
+typedef DomainWallFermion<GparityWilsonImplR> GparityDomainWallFermionR;
 typedef DomainWallFermion<GparityWilsonImplF> GparityDomainWallFermionF;
 typedef DomainWallFermion<GparityWilsonImplD> GparityDomainWallFermionD;
 
-typedef DomainWallEOFAFermion<GparityWilsonImplR> GparityDomainWallEOFAFermionD2;
+//typedef DomainWallFermion<GparityWilsonImplRL> GparityDomainWallFermionRL;
+//typedef DomainWallFermion<GparityWilsonImplFH> GparityDomainWallFermionFH;
+//typedef DomainWallFermion<GparityWilsonImplDF> GparityDomainWallFermionDF;
+
+typedef DomainWallEOFAFermion<GparityWilsonImplR> GparityDomainWallEOFAFermionR;
 typedef DomainWallEOFAFermion<GparityWilsonImplF> GparityDomainWallEOFAFermionF;
 typedef DomainWallEOFAFermion<GparityWilsonImplD> GparityDomainWallEOFAFermionD;
 
-typedef WilsonTMFermion<GparityWilsonImplR> GparityWilsonTMFermionD2;
+//typedef DomainWallEOFAFermion<GparityWilsonImplRL> GparityDomainWallEOFAFermionRL;
+//typedef DomainWallEOFAFermion<GparityWilsonImplFH> GparityDomainWallEOFAFermionFH;
+//typedef DomainWallEOFAFermion<GparityWilsonImplDF> GparityDomainWallEOFAFermionDF;
+
+typedef WilsonTMFermion<GparityWilsonImplR> GparityWilsonTMFermionR;
 typedef WilsonTMFermion<GparityWilsonImplF> GparityWilsonTMFermionF;
 typedef WilsonTMFermion<GparityWilsonImplD> GparityWilsonTMFermionD;
 
-typedef MobiusFermion<GparityWilsonImplR> GparityMobiusFermionD2;
+//typedef WilsonTMFermion<GparityWilsonImplRL> GparityWilsonTMFermionRL;
+//typedef WilsonTMFermion<GparityWilsonImplFH> GparityWilsonTMFermionFH;
+//typedef WilsonTMFermion<GparityWilsonImplDF> GparityWilsonTMFermionDF;
+
+typedef MobiusFermion<GparityWilsonImplR> GparityMobiusFermionR;
 typedef MobiusFermion<GparityWilsonImplF> GparityMobiusFermionF;
 typedef MobiusFermion<GparityWilsonImplD> GparityMobiusFermionD;
 
-typedef MobiusEOFAFermion<GparityWilsonImplR> GparityMobiusEOFAFermionD2;
+//typedef MobiusFermion<GparityWilsonImplRL> GparityMobiusFermionRL;
+//typedef MobiusFermion<GparityWilsonImplFH> GparityMobiusFermionFH;
+//typedef MobiusFermion<GparityWilsonImplDF> GparityMobiusFermionDF;
+
+typedef MobiusEOFAFermion<GparityWilsonImplR> GparityMobiusEOFAFermionR;
 typedef MobiusEOFAFermion<GparityWilsonImplF> GparityMobiusEOFAFermionF;
 typedef MobiusEOFAFermion<GparityWilsonImplD> GparityMobiusEOFAFermionD;
 
+//typedef MobiusEOFAFermion<GparityWilsonImplRL> GparityMobiusEOFAFermionRL;
+//typedef MobiusEOFAFermion<GparityWilsonImplFH> GparityMobiusEOFAFermionFH;
+//typedef MobiusEOFAFermion<GparityWilsonImplDF> GparityMobiusEOFAFermionDF;
+
+typedef ImprovedStaggeredFermion<StaggeredImplR> ImprovedStaggeredFermionR;
 typedef ImprovedStaggeredFermion<StaggeredImplF> ImprovedStaggeredFermionF;
 typedef ImprovedStaggeredFermion<StaggeredImplD> ImprovedStaggeredFermionD;
 
+typedef NaiveStaggeredFermion<StaggeredImplR> NaiveStaggeredFermionR;
 typedef NaiveStaggeredFermion<StaggeredImplF> NaiveStaggeredFermionF;
 typedef NaiveStaggeredFermion<StaggeredImplD> NaiveStaggeredFermionD;
 
+typedef ImprovedStaggeredFermion5D<StaggeredImplR> ImprovedStaggeredFermion5DR;
 typedef ImprovedStaggeredFermion5D<StaggeredImplF> ImprovedStaggeredFermion5DF;
 typedef ImprovedStaggeredFermion5D<StaggeredImplD> ImprovedStaggeredFermion5DD;
 

Grid/qcd/action/fermion/FermionOperator.h

@@ -49,8 +49,6 @@ public:
 
   virtual FermionField &tmp(void) = 0;
 
-  virtual void DirichletBlock(const Coordinate & _Block) { assert(0); };
-  
   GridBase * Grid(void)   { return FermionGrid(); };   // this is all the linalg routines need to know
   GridBase * RedBlackGrid(void) { return FermionRedBlackGrid(); };
 

Grid/qcd/action/fermion/ImprovedStaggeredFermion.h

@@ -47,6 +47,18 @@ public:
   FermionField _tmp;
   FermionField &tmp(void) { return _tmp; }
 
+  ////////////////////////////////////////
+  // Performance monitoring
+  ////////////////////////////////////////
+  void Report(void);
+  void ZeroCounters(void);
+  double DhopTotalTime;
+  double DhopCalls;
+  double DhopCommTime;
+  double DhopComputeTime;
+  double DhopComputeTime2;
+  double DhopFaceTime;
+
   ///////////////////////////////////////////////////////////////
   // Implement the abstract base
   ///////////////////////////////////////////////////////////////

Grid/qcd/action/fermion/ImprovedStaggeredFermion5D.h

@@ -52,6 +52,18 @@ public:
   FermionField _tmp;
   FermionField &tmp(void) { return _tmp; }
 
+  ////////////////////////////////////////
+  // Performance monitoring
+  ////////////////////////////////////////
+  void Report(void);
+  void ZeroCounters(void);
+  double DhopTotalTime;
+  double DhopCalls;
+  double DhopCommTime;
+  double DhopComputeTime;
+  double DhopComputeTime2;
+  double DhopFaceTime;
+
   ///////////////////////////////////////////////////////////////
   // Implement the abstract base
   ///////////////////////////////////////////////////////////////

Grid/qcd/action/fermion/NaiveStaggeredFermion.h

@@ -47,6 +47,18 @@ public:
   FermionField _tmp;
   FermionField &tmp(void) { return _tmp; }
 
+  ////////////////////////////////////////
+  // Performance monitoring
+  ////////////////////////////////////////
+  void Report(void);
+  void ZeroCounters(void);
+  double DhopTotalTime;
+  double DhopCalls;
+  double DhopCommTime;
+  double DhopComputeTime;
+  double DhopComputeTime2;
+  double DhopFaceTime;
+
   ///////////////////////////////////////////////////////////////
   // Implement the abstract base
   ///////////////////////////////////////////////////////////////

Grid/qcd/action/fermion/WilsonCloverFermion.h

@@ -32,7 +32,6 @@
 
 #include <Grid/qcd/action/fermion/WilsonCloverTypes.h>
 #include <Grid/qcd/action/fermion/WilsonCloverHelpers.h>
-#include <Grid/qcd/action/fermion/CloverHelpers.h>
 
 NAMESPACE_BEGIN(Grid);
 
@@ -52,7 +51,7 @@ NAMESPACE_BEGIN(Grid);
 // csw_r = csw_t to recover the isotropic version
 //////////////////////////////////////////////////////////////////
 
-template<class Impl, class CloverHelpers>
+template <class Impl>
 class WilsonCloverFermion : public WilsonFermion<Impl>,
                             public WilsonCloverHelpers<Impl>
 {

Grid/qcd/action/fermion/WilsonCloverHelpers.h

@@ -209,8 +209,6 @@ public:
 };
 
 
-////////////////////////////////////////////////////////
-
 template<class Impl> class CompactWilsonCloverHelpers {
 public:
 

Grid/qcd/action/fermion/WilsonCloverTypes.h

@@ -47,6 +47,8 @@ class CompactWilsonCloverTypes {
 public:
   INHERIT_IMPL_TYPES(Impl);
 
+  static_assert(Nd == 4 && Nc == 3 && Ns == 4 && Impl::Dimension == 3, "Wrong dimensions");
+
   static constexpr int Nred      = Nc * Nhs;        // 6
   static constexpr int Nblock    = Nhs;             // 2
   static constexpr int Ndiagonal = Nred;            // 6

Grid/qcd/action/fermion/WilsonCompressor.h

@@ -32,218 +32,17 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 
 NAMESPACE_BEGIN(Grid);
 
-///////////////////////////////////////////////////////////////
-// Wilson compressor will need FaceGather policies for:
-// Periodic, Dirichlet, and partial Dirichlet for DWF
-///////////////////////////////////////////////////////////////
-const int dwf_compressor_depth=2;
-#define DWF_COMPRESS
-class FaceGatherPartialDWF
-{
-public:
-#ifdef DWF_COMPRESS
-  static int PartialCompressionFactor(GridBase *grid) {return grid->_fdimensions[0]/(2*dwf_compressor_depth);};
-#else
-  static int PartialCompressionFactor(GridBase *grid) { return 1;}
-#endif
-  template<class vobj,class cobj,class compressor>
-  static void Gather_plane_simple (commVector<std::pair<int,int> >& table,
-				   const Lattice<vobj> &rhs,
-				   cobj *buffer,
-				   compressor &compress,
-				   int off,int so,int partial)
-  {
-    //DWF only hack: If a direction that is OFF node we use Partial Dirichlet
-    //  Shrinks local and remote comms buffers
-    GridBase *Grid = rhs.Grid();
-    int Ls = Grid->_rdimensions[0];
-#ifdef DWF_COMPRESS
-    int depth=dwf_compressor_depth;
-#else 
-    int depth=Ls/2;
-#endif
-    std::pair<int,int> *table_v = & table[0];
-    auto rhs_v = rhs.View(AcceleratorRead);
-    int vol=table.size()/Ls;
-    accelerator_forNB( idx,table.size(), vobj::Nsimd(), {
-	Integer i=idx/Ls;
-	Integer s=idx%Ls;
-	Integer sc=depth+s-(Ls-depth);
-	if(s<depth)     compress.Compress(buffer[off+i+s*vol],rhs_v[so+table_v[idx].second]);
-	if(s>=Ls-depth) compress.Compress(buffer[off+i+sc*vol],rhs_v[so+table_v[idx].second]);
-    });
-    rhs_v.ViewClose();
-  }
-  template<class decompressor,class Decompression>
-  static void DecompressFace(decompressor decompress,Decompression &dd)
-  {
-    auto Ls = dd.dims[0];
-#ifdef DWF_COMPRESS
-    int depth=dwf_compressor_depth;
-#else
-    int depth=Ls/2;
-#endif    
-    // Just pass in the Grid
-    auto kp = dd.kernel_p;
-    auto mp = dd.mpi_p;
-    int size= dd.buffer_size;
-    int vol= size/Ls;
-    accelerator_forNB(o,size,1,{
-	int idx=o/Ls;
-	int   s=o%Ls;
-	if ( s < depth ) {
-	  int oo=s*vol+idx;
-	  kp[o]=mp[oo];
-	} else if ( s >= Ls-depth ) {
-	  int sc = depth + s - (Ls-depth);
-	  int oo=sc*vol+idx; 
-	  kp[o]=mp[oo];
-	} else {
-	  kp[o] = Zero();//fill rest with zero if partial dirichlet
-	}
-    });
-  }
-  ////////////////////////////////////////////////////////////////////////////////////////////
-  // Need to gather *interior portions* for ALL s-slices in simd directions
-  // Do the gather as need to treat SIMD lanes differently, and insert zeroes on receive side
-  // Reorder the fifth dim to be s=Ls-1 , s=0, s=1,...,Ls-2.
-  ////////////////////////////////////////////////////////////////////////////////////////////
-  template<class vobj,class cobj,class compressor>
-  static void Gather_plane_exchange(commVector<std::pair<int,int> >& table,const Lattice<vobj> &rhs,
-				    std::vector<cobj *> pointers,int dimension,int plane,int cbmask,
-				    compressor &compress,int type,int partial)
-  {
-    GridBase *Grid = rhs.Grid();
-    int Ls = Grid->_rdimensions[0];
-#ifdef DWF_COMPRESS
-    int depth=dwf_compressor_depth;
-#else
-    int depth = Ls/2;
-#endif
-    
-    // insertion of zeroes...
-    assert( (table.size()&0x1)==0);
-    int num=table.size()/2;
-    int so  = plane*rhs.Grid()->_ostride[dimension]; // base offset for start of plane
-    
-    auto rhs_v = rhs.View(AcceleratorRead);
-    auto p0=&pointers[0][0];
-    auto p1=&pointers[1][0];
-    auto tp=&table[0];
-    int nnum=num/Ls;
-    accelerator_forNB(j, num, vobj::Nsimd(), {
-	//  Reorders both local and remote comms buffers
-	//  
-	int s  = j % Ls;
-	int sp1 = (s+depth)%Ls;  // peri incremented s slice
-	
-	int hxyz= j/Ls;
-
-	int xyz0= hxyz*2; // xyzt part of coor
-	int xyz1= hxyz*2+1;
-	
-	int jj= hxyz + sp1*nnum ; // 0,1,2,3 -> Ls-1 slice , 0-slice, 1-slice ....
-	
-	int kk0= xyz0*Ls + s ; // s=0 goes to s=1
-	int kk1= xyz1*Ls + s ; // s=Ls-1 -> s=0
-	compress.CompressExchange(p0[jj],p1[jj],
-				  rhs_v[so+tp[kk0 ].second], // Same s, consecutive xyz sites
-				  rhs_v[so+tp[kk1 ].second], 
-				  type);
-    });
-    rhs_v.ViewClose();
-  }
-  // Merge routine is for SIMD faces
-  template<class decompressor,class Merger>
-  static void MergeFace(decompressor decompress,Merger &mm)
-  {
-    auto Ls = mm.dims[0];
-#ifdef DWF_COMPRESS
-    int depth=dwf_compressor_depth;
-#else
-    int depth = Ls/2;
-#endif
-    int  num= mm.buffer_size/2; // relate vol and Ls to buffer size
-    auto mp = &mm.mpointer[0];
-    auto vp0= &mm.vpointers[0][0]; // First arg is exchange first
-    auto vp1= &mm.vpointers[1][0];
-    auto type= mm.type;
-    int nnum = num/Ls;
-    accelerator_forNB(o,num,Merger::Nsimd,{
-
-	int  s=o%Ls;
-	int hxyz=o/Ls; // xyzt related component
-	int xyz0=hxyz*2;
-	int xyz1=hxyz*2+1;
-
-	int sp = (s+depth)%Ls; 
-	int jj= hxyz + sp*nnum ; // 0,1,2,3 -> Ls-1 slice , 0-slice, 1-slice ....
-
-	int oo0= s+xyz0*Ls;
-	int oo1= s+xyz1*Ls;
-
-	// same ss0, ss1 pair goes to new layout
-	decompress.Exchange(mp[oo0],mp[oo1],vp0[jj],vp1[jj],type);
-      });
-  }
-};
-class FaceGatherDWFMixedBCs
-{
-public:
-#ifdef DWF_COMPRESS
-  static int PartialCompressionFactor(GridBase *grid) {return grid->_fdimensions[0]/(2*dwf_compressor_depth);};
-#else 
-  static int PartialCompressionFactor(GridBase *grid) {return 1;}
-#endif
-  
-  template<class vobj,class cobj,class compressor>
-  static void Gather_plane_simple (commVector<std::pair<int,int> >& table,
-					 const Lattice<vobj> &rhs,
-					 cobj *buffer,
-					 compressor &compress,
-					 int off,int so,int partial)
-  {
-    //    std::cout << " face gather simple DWF partial "<<partial <<std::endl;
-    if(partial) FaceGatherPartialDWF::Gather_plane_simple(table,rhs,buffer,compress,off,so,partial);
-    else        FaceGatherSimple::Gather_plane_simple(table,rhs,buffer,compress,off,so,partial);
-  }
-  template<class vobj,class cobj,class compressor>
-  static void Gather_plane_exchange(commVector<std::pair<int,int> >& table,const Lattice<vobj> &rhs,
-				    std::vector<cobj *> pointers,int dimension,int plane,int cbmask,
-				    compressor &compress,int type,int partial)
-  {
-    //    std::cout << " face gather exch DWF partial "<<partial <<std::endl;
-    if(partial) FaceGatherPartialDWF::Gather_plane_exchange(table,rhs,pointers,dimension, plane,cbmask,compress,type,partial);
-    else        FaceGatherSimple::Gather_plane_exchange    (table,rhs,pointers,dimension, plane,cbmask,compress,type,partial);
-  }
-  template<class decompressor,class Merger>
-  static void MergeFace(decompressor decompress,Merger &mm)
-  {
-    int partial = mm.partial;
-    //    std::cout << " merge DWF partial "<<partial <<std::endl;
-    if ( partial ) FaceGatherPartialDWF::MergeFace(decompress,mm);
-    else           FaceGatherSimple::MergeFace(decompress,mm);
-  }
-
-  template<class decompressor,class Decompression>
-  static void DecompressFace(decompressor decompress,Decompression &dd)
-  {
-    int partial = dd.partial;
-    //    std::cout << " decompress DWF partial "<<partial <<std::endl;
-    if ( partial ) FaceGatherPartialDWF::DecompressFace(decompress,dd);
-    else           FaceGatherSimple::DecompressFace(decompress,dd);
-  }
-};
-
 /////////////////////////////////////////////////////////////////////////////////////////////
-// optimised versions supporting half precision too??? Deprecate
+// optimised versions supporting half precision too
 /////////////////////////////////////////////////////////////////////////////////////////////
 
+template<class _HCspinor,class _Hspinor,class _Spinor, class projector,typename SFINAE = void >
+class WilsonCompressorTemplate;
+
 
-//Could make FaceGather a template param, but then behaviour is runtime not compile time
 template<class _HCspinor,class _Hspinor,class _Spinor, class projector>
-class WilsonCompressorTemplate  : public FaceGatherDWFMixedBCs
+class WilsonCompressorTemplate< _HCspinor, _Hspinor, _Spinor, projector,
-//  : public FaceGatherSimple
+				typename std::enable_if<std::is_same<_HCspinor,_Hspinor>::value>::type >
 {
 public:
   
@@ -280,81 +79,172 @@ public:
   /*****************************************************/
   /* Exchange includes precision change if mpi data is not same */
   /*****************************************************/
-  accelerator_inline void Exchange(SiteHalfSpinor &mp0,
+  accelerator_inline void Exchange(SiteHalfSpinor *mp,
-				   SiteHalfSpinor &mp1,
+				   const SiteHalfSpinor * __restrict__ vp0,
-				   const SiteHalfSpinor & vp0,
+				   const SiteHalfSpinor * __restrict__ vp1,
-				   const SiteHalfSpinor & vp1,
+				   Integer type,Integer o) const {
-				   Integer type) const {
 #ifdef GRID_SIMT
-    exchangeSIMT(mp0,mp1,vp0,vp1,type);
+    exchangeSIMT(mp[2*o],mp[2*o+1],vp0[o],vp1[o],type);
 #else
     SiteHalfSpinor tmp1;
     SiteHalfSpinor tmp2;
-    exchange(tmp1,tmp2,vp0,vp1,type);
+    exchange(tmp1,tmp2,vp0[o],vp1[o],type);
-    vstream(mp0,tmp1);
+    vstream(mp[2*o  ],tmp1);
-    vstream(mp1,tmp2);
+    vstream(mp[2*o+1],tmp2);
 #endif
   }
-  
+
 
   /*****************************************************/
   /* Have a decompression step if mpi data is not same */
   /*****************************************************/
-  accelerator_inline void Decompress(SiteHalfSpinor &out,
+  accelerator_inline void Decompress(SiteHalfSpinor * __restrict__ out,
-				     SiteHalfSpinor &in) const {    
+				     SiteHalfSpinor * __restrict__ in, Integer o) const {    
-    out = in;
+    assert(0);
   }
 
   /*****************************************************/
   /* Compress Exchange                                 */
   /*****************************************************/
-  accelerator_inline void CompressExchange(SiteHalfSpinor &out0,
+  accelerator_inline void CompressExchange(SiteHalfSpinor * __restrict__ out0,
-					   SiteHalfSpinor &out1,
+					   SiteHalfSpinor * __restrict__ out1,
-					   const SiteSpinor &in0,
+					   const SiteSpinor * __restrict__ in,
-					   const SiteSpinor &in1,
+					   Integer j,Integer k, Integer m,Integer type) const
-					   Integer type) const
   {
 #ifdef GRID_SIMT
     typedef SiteSpinor vobj;
     typedef SiteHalfSpinor hvobj;
-    typedef decltype(coalescedRead(in0))    sobj;
+    typedef decltype(coalescedRead(*in))    sobj;
-    typedef decltype(coalescedRead(out0)) hsobj;
+    typedef decltype(coalescedRead(*out0)) hsobj;
 
-    constexpr unsigned int Nsimd = vobj::Nsimd();
+    unsigned int Nsimd = vobj::Nsimd();
     unsigned int mask = Nsimd >> (type + 1);
     int lane = acceleratorSIMTlane(Nsimd);
     int j0 = lane &(~mask); // inner coor zero
     int j1 = lane |(mask) ; // inner coor one
-    const vobj *vp0 = &in0;
+    const vobj *vp0 = &in[k];
-    const vobj *vp1 = &in1;
+    const vobj *vp1 = &in[m];
     const vobj *vp = (lane&mask) ? vp1:vp0;
     auto sa = coalescedRead(*vp,j0);
     auto sb = coalescedRead(*vp,j1);
     hsobj psa, psb;
     projector::Proj(psa,sa,mu,dag);
     projector::Proj(psb,sb,mu,dag);
-    coalescedWrite(out0,psa);
+    coalescedWrite(out0[j],psa);
-    coalescedWrite(out1,psb);
+    coalescedWrite(out1[j],psb);
 #else
     SiteHalfSpinor temp1, temp2;
     SiteHalfSpinor temp3, temp4;
-    projector::Proj(temp1,in0,mu,dag);
+    projector::Proj(temp1,in[k],mu,dag);
-    projector::Proj(temp2,in1,mu,dag);
+    projector::Proj(temp2,in[m],mu,dag);
     exchange(temp3,temp4,temp1,temp2,type);
-    vstream(out0,temp3);
+    vstream(out0[j],temp3);
-    vstream(out1,temp4);
+    vstream(out1[j],temp4);
 #endif
   }
 
   /*****************************************************/
   /* Pass the info to the stencil */
   /*****************************************************/
-  accelerator_inline bool DecompressionStep(void) const {
+  accelerator_inline bool DecompressionStep(void) const { return false; }
-    return false;
-  }
 
 };
 
+#if 0
+template<class _HCspinor,class _Hspinor,class _Spinor, class projector>
+class WilsonCompressorTemplate< _HCspinor, _Hspinor, _Spinor, projector,
+				typename std::enable_if<!std::is_same<_HCspinor,_Hspinor>::value>::type >
+{
+public:
+  
+  int mu,dag;  
+
+  void Point(int p) { mu=p; };
+
+  WilsonCompressorTemplate(int _dag=0){
+    dag = _dag;
+  }
+
+  typedef _Spinor         SiteSpinor;
+  typedef _Hspinor     SiteHalfSpinor;
+  typedef _HCspinor SiteHalfCommSpinor;
+  typedef typename SiteHalfCommSpinor::vector_type vComplexLow;
+  typedef typename SiteHalfSpinor::vector_type     vComplexHigh;
+  constexpr static int Nw=sizeof(SiteHalfSpinor)/sizeof(vComplexHigh);
+
+  accelerator_inline int CommDatumSize(void) const {
+    return sizeof(SiteHalfCommSpinor);
+  }
+
+  /*****************************************************/
+  /* Compress includes precision change if mpi data is not same */
+  /*****************************************************/
+  accelerator_inline void Compress(SiteHalfSpinor &buf,const SiteSpinor &in) const {
+    SiteHalfSpinor hsp;
+    SiteHalfCommSpinor *hbuf = (SiteHalfCommSpinor *)buf;
+    projector::Proj(hsp,in,mu,dag);
+    precisionChange((vComplexLow *)&hbuf[o],(vComplexHigh *)&hsp,Nw);
+  }
+  accelerator_inline void Compress(SiteHalfSpinor &buf,const SiteSpinor &in) const {
+#ifdef GRID_SIMT
+    typedef decltype(coalescedRead(buf)) sobj;
+    sobj sp;
+    auto sin = coalescedRead(in);
+    projector::Proj(sp,sin,mu,dag);
+    coalescedWrite(buf,sp);
+#else
+    projector::Proj(buf,in,mu,dag);
+#endif
+  }
+
+  /*****************************************************/
+  /* Exchange includes precision change if mpi data is not same */
+  /*****************************************************/
+  accelerator_inline void Exchange(SiteHalfSpinor *mp,
+                       SiteHalfSpinor *vp0,
+                       SiteHalfSpinor *vp1,
+		       Integer type,Integer o) const {
+    SiteHalfSpinor vt0,vt1;
+    SiteHalfCommSpinor *vpp0 = (SiteHalfCommSpinor *)vp0;
+    SiteHalfCommSpinor *vpp1 = (SiteHalfCommSpinor *)vp1;
+    precisionChange((vComplexHigh *)&vt0,(vComplexLow *)&vpp0[o],Nw);
+    precisionChange((vComplexHigh *)&vt1,(vComplexLow *)&vpp1[o],Nw);
+    exchange(mp[2*o],mp[2*o+1],vt0,vt1,type);
+  }
+
+  /*****************************************************/
+  /* Have a decompression step if mpi data is not same */
+  /*****************************************************/
+  accelerator_inline void Decompress(SiteHalfSpinor *out, SiteHalfSpinor *in, Integer o) const {
+    SiteHalfCommSpinor *hin=(SiteHalfCommSpinor *)in;
+    precisionChange((vComplexHigh *)&out[o],(vComplexLow *)&hin[o],Nw);
+  }
+
+  /*****************************************************/
+  /* Compress Exchange                                 */
+  /*****************************************************/
+  accelerator_inline void CompressExchange(SiteHalfSpinor *out0,
+			       SiteHalfSpinor *out1,
+			       const SiteSpinor *in,
+			       Integer j,Integer k, Integer m,Integer type) const {
+    SiteHalfSpinor temp1, temp2,temp3,temp4;
+    SiteHalfCommSpinor *hout0 = (SiteHalfCommSpinor *)out0;
+    SiteHalfCommSpinor *hout1 = (SiteHalfCommSpinor *)out1;
+    projector::Proj(temp1,in[k],mu,dag);
+    projector::Proj(temp2,in[m],mu,dag);
+    exchange(temp3,temp4,temp1,temp2,type);
+    precisionChange((vComplexLow *)&hout0[j],(vComplexHigh *)&temp3,Nw);
+    precisionChange((vComplexLow *)&hout1[j],(vComplexHigh *)&temp4,Nw);
+  }
+
+  /*****************************************************/
+  /* Pass the info to the stencil */
+  /*****************************************************/
+  accelerator_inline bool DecompressionStep(void) const { return true; }
+
+};
+#endif
+
 #define DECLARE_PROJ(Projector,Compressor,spProj)			\
   class Projector {							\
   public:								\
@@ -404,7 +294,11 @@ public:
   typedef typename Base::View_type View_type;
   typedef typename Base::StencilVector StencilVector;
 
-  //  Vector<int> surface_list;
+  void ZeroCountersi(void)  {  }
+  void Reporti(int calls)  {  }
+
+  std::vector<int> surface_list;
+
   WilsonStencil(GridBase *grid,
 		int npoints,
 		int checkerboard,
@@ -412,11 +306,11 @@ public:
 		const std::vector<int> &distances,Parameters p)  
     : CartesianStencil<vobj,cobj,Parameters> (grid,npoints,checkerboard,directions,distances,p) 
   { 
-    //    surface_list.resize(0);
+    ZeroCountersi();
+    surface_list.resize(0);
     this->same_node.resize(npoints);
   };
 
-  /*
   void BuildSurfaceList(int Ls,int vol4){
 
     // find same node for SHM
@@ -437,8 +331,7 @@ public:
       }
     }
   }
-  */
+
-  
   template < class compressor>
   void HaloExchangeOpt(const Lattice<vobj> &source,compressor &compress) 
   {
@@ -484,26 +377,24 @@ public:
 
     int dag = compress.dag;
     int face_idx=0;
-#define vet_same_node(a,b) \
-      { auto tmp = b;  }
     if ( dag ) { 
-      vet_same_node(this->same_node[Xp],this->HaloGatherDir(source,XpCompress,Xp,face_idx));
+      assert(this->same_node[Xp]==this->HaloGatherDir(source,XpCompress,Xp,face_idx));
-      vet_same_node(this->same_node[Yp],this->HaloGatherDir(source,YpCompress,Yp,face_idx));
+      assert(this->same_node[Yp]==this->HaloGatherDir(source,YpCompress,Yp,face_idx));
-      vet_same_node(this->same_node[Zp],this->HaloGatherDir(source,ZpCompress,Zp,face_idx));
+      assert(this->same_node[Zp]==this->HaloGatherDir(source,ZpCompress,Zp,face_idx));
-      vet_same_node(this->same_node[Tp],this->HaloGatherDir(source,TpCompress,Tp,face_idx));
+      assert(this->same_node[Tp]==this->HaloGatherDir(source,TpCompress,Tp,face_idx));
-      vet_same_node(this->same_node[Xm],this->HaloGatherDir(source,XmCompress,Xm,face_idx));
+      assert(this->same_node[Xm]==this->HaloGatherDir(source,XmCompress,Xm,face_idx));
-      vet_same_node(this->same_node[Ym],this->HaloGatherDir(source,YmCompress,Ym,face_idx));
+      assert(this->same_node[Ym]==this->HaloGatherDir(source,YmCompress,Ym,face_idx));
-      vet_same_node(this->same_node[Zm],this->HaloGatherDir(source,ZmCompress,Zm,face_idx));
+      assert(this->same_node[Zm]==this->HaloGatherDir(source,ZmCompress,Zm,face_idx));
-      vet_same_node(this->same_node[Tm],this->HaloGatherDir(source,TmCompress,Tm,face_idx));
+      assert(this->same_node[Tm]==this->HaloGatherDir(source,TmCompress,Tm,face_idx));
     } else {
-      vet_same_node(this->same_node[Xp],this->HaloGatherDir(source,XmCompress,Xp,face_idx));
+      assert(this->same_node[Xp]==this->HaloGatherDir(source,XmCompress,Xp,face_idx));
-      vet_same_node(this->same_node[Yp],this->HaloGatherDir(source,YmCompress,Yp,face_idx));
+      assert(this->same_node[Yp]==this->HaloGatherDir(source,YmCompress,Yp,face_idx));
-      vet_same_node(this->same_node[Zp],this->HaloGatherDir(source,ZmCompress,Zp,face_idx));
+      assert(this->same_node[Zp]==this->HaloGatherDir(source,ZmCompress,Zp,face_idx));
-      vet_same_node(this->same_node[Tp],this->HaloGatherDir(source,TmCompress,Tp,face_idx));
+      assert(this->same_node[Tp]==this->HaloGatherDir(source,TmCompress,Tp,face_idx));
-      vet_same_node(this->same_node[Xm],this->HaloGatherDir(source,XpCompress,Xm,face_idx));
+      assert(this->same_node[Xm]==this->HaloGatherDir(source,XpCompress,Xm,face_idx));
-      vet_same_node(this->same_node[Ym],this->HaloGatherDir(source,YpCompress,Ym,face_idx));
+      assert(this->same_node[Ym]==this->HaloGatherDir(source,YpCompress,Ym,face_idx));
-      vet_same_node(this->same_node[Zm],this->HaloGatherDir(source,ZpCompress,Zm,face_idx));
+      assert(this->same_node[Zm]==this->HaloGatherDir(source,ZpCompress,Zm,face_idx));
-      vet_same_node(this->same_node[Tm],this->HaloGatherDir(source,TpCompress,Tm,face_idx));
+      assert(this->same_node[Tm]==this->HaloGatherDir(source,TpCompress,Tm,face_idx));
     }
     this->face_table_computed=1;
     assert(this->u_comm_offset==this->_unified_buffer_size);

Grid/qcd/action/fermion/WilsonFermion.h

@@ -74,6 +74,20 @@ public:
   FermionField _tmp;
   FermionField &tmp(void) { return _tmp; }
 
+  void Report(void);
+  void ZeroCounters(void);
+  double DhopCalls;
+  double DhopCommTime;
+  double DhopComputeTime;
+  double DhopComputeTime2;
+  double DhopFaceTime;
+  double DhopTotalTime;
+
+  double DerivCalls;
+  double DerivCommTime;
+  double DerivComputeTime;
+  double DerivDhopComputeTime;
+
   //////////////////////////////////////////////////////////////////
   // override multiply; cut number routines if pass dagger argument
   // and also make interface more uniformly consistent

Grid/qcd/action/fermion/WilsonFermion5D.h

@@ -75,8 +75,19 @@ public:
   FermionField _tmp;
   FermionField &tmp(void) { return _tmp; }
 
-  int Dirichlet;
+  void Report(void);
-  Coordinate Block; 
+  void ZeroCounters(void);
+  double DhopCalls;
+  double DhopCommTime;
+  double DhopComputeTime;
+  double DhopComputeTime2;
+  double DhopFaceTime;
+  double DhopTotalTime;
+
+  double DerivCalls;
+  double DerivCommTime;
+  double DerivComputeTime;
+  double DerivDhopComputeTime;
 
   ///////////////////////////////////////////////////////////////
   // Implement the abstract base
@@ -162,10 +173,7 @@ public:
 		  GridCartesian         &FourDimGrid,
 		  GridRedBlackCartesian &FourDimRedBlackGrid,
 		  double _M5,const ImplParams &p= ImplParams());
-
+    
-  virtual void DirichletBlock(const Coordinate & block)
-  {
-  }
   // Constructors
   /*
     WilsonFermion5D(int simd, 

Grid/qcd/action/fermion/WilsonImpl.h

@@ -37,7 +37,7 @@ NAMESPACE_BEGIN(Grid);
 template <class S, class Representation = FundamentalRepresentation,class Options = CoeffReal >
 class WilsonImpl : public PeriodicGaugeImpl<GaugeImplTypes<S, Representation::Dimension > > {
 public:
-  
+
   static const int Dimension = Representation::Dimension;
   static const bool isFundamental = Representation::isFundamental;
   static const bool LsVectorised=false;
@@ -242,13 +242,19 @@ public:
 typedef WilsonImpl<vComplex,  FundamentalRepresentation, CoeffReal > WilsonImplR;  // Real.. whichever prec
 typedef WilsonImpl<vComplexF, FundamentalRepresentation, CoeffReal > WilsonImplF;  // Float
 typedef WilsonImpl<vComplexD, FundamentalRepresentation, CoeffReal > WilsonImplD;  // Double
-typedef WilsonImpl<vComplexD2, FundamentalRepresentation, CoeffReal > WilsonImplD2;  // Double
+
+//typedef WilsonImpl<vComplex,  FundamentalRepresentation, CoeffRealHalfComms > WilsonImplRL;  // Real.. whichever prec
+//typedef WilsonImpl<vComplexF, FundamentalRepresentation, CoeffRealHalfComms > WilsonImplFH;  // Float
+//typedef WilsonImpl<vComplexD, FundamentalRepresentation, CoeffRealHalfComms > WilsonImplDF;  // Double
 
 typedef WilsonImpl<vComplex,  FundamentalRepresentation, CoeffComplex > ZWilsonImplR; // Real.. whichever prec
 typedef WilsonImpl<vComplexF, FundamentalRepresentation, CoeffComplex > ZWilsonImplF; // Float
 typedef WilsonImpl<vComplexD, FundamentalRepresentation, CoeffComplex > ZWilsonImplD; // Double
-typedef WilsonImpl<vComplexD2, FundamentalRepresentation, CoeffComplex > ZWilsonImplD2; // Double
 
+//typedef WilsonImpl<vComplex,  FundamentalRepresentation, CoeffComplexHalfComms > ZWilsonImplRL; // Real.. whichever prec
+//typedef WilsonImpl<vComplexF, FundamentalRepresentation, CoeffComplexHalfComms > ZWilsonImplFH; // Float
+//typedef WilsonImpl<vComplexD, FundamentalRepresentation, CoeffComplexHalfComms > ZWilsonImplDF; // Double
+ 
 typedef WilsonImpl<vComplex,  AdjointRepresentation, CoeffReal > WilsonAdjImplR;   // Real.. whichever prec
 typedef WilsonImpl<vComplexF, AdjointRepresentation, CoeffReal > WilsonAdjImplF;  // Float
 typedef WilsonImpl<vComplexD, AdjointRepresentation, CoeffReal > WilsonAdjImplD;  // Double
@@ -261,22 +267,6 @@ 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/WilsonKernels.h

@@ -52,6 +52,13 @@ public:
   typedef AcceleratorVector<int,STENCIL_MAX> StencilVector;   
 public:
 
+#ifdef GRID_SYCL
+#define SYCL_HACK
+#endif  
+#ifdef SYCL_HACK
+  static void HandDhopSiteSycl(StencilVector st_perm,StencilEntry *st_p, SiteDoubledGaugeField *U,SiteHalfSpinor  *buf,
+			       int ss,int sU,const SiteSpinor *in, SiteSpinor *out);
+#endif
   
   static void DhopKernel(int Opt,StencilImpl &st,  DoubledGaugeField &U, SiteHalfSpinor * buf,
 			 int Ls, int Nsite, const FermionField &in, FermionField &out,

Grid/qcd/action/fermion/WilsonTMFermion.h

@@ -63,9 +63,7 @@ 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/CayleyFermion5DImplementation.h

@@ -47,7 +47,7 @@ CayleyFermion5D<Impl>::CayleyFermion5D(GaugeField &_Umu,
 			FiveDimRedBlackGrid,
 			FourDimGrid,
 			FourDimRedBlackGrid,_M5,p),
-  mass_plus(_mass), mass_minus(_mass)
+  mass(_mass)
 { 
 }
 
@@ -152,13 +152,65 @@ void CayleyFermion5D<Impl>::DminusDag(const FermionField &psi, FermionField &chi
   }
 }
 
+template<class Impl> void CayleyFermion5D<Impl>::CayleyReport(void)
+{
+  this->Report();
+  Coordinate latt = GridDefaultLatt();          
+  RealD volume = this->Ls;  for(int mu=0;mu<Nd;mu++) volume=volume*latt[mu];
+  RealD NP     = this->_FourDimGrid->_Nprocessors;
+  if ( M5Dcalls > 0 ) {
+    std::cout << GridLogMessage << "#### M5D calls report " << std::endl;
+    std::cout << GridLogMessage << "CayleyFermion5D Number of M5D Calls     : " << M5Dcalls   << std::endl;
+    std::cout << GridLogMessage << "CayleyFermion5D ComputeTime/Calls       : " << M5Dtime / M5Dcalls << " us" << std::endl;
+
+    // Flops = 10.0*(Nc*Ns) *Ls*vol
+    RealD mflops = 10.0*(Nc*Ns)*volume*M5Dcalls/M5Dtime/2; // 2 for red black counting
+    std::cout << GridLogMessage << "Average mflops/s per call                : " << mflops << std::endl;
+    std::cout << GridLogMessage << "Average mflops/s per call per rank       : " << mflops/NP << std::endl;
+
+    // Bytes = sizeof(Real) * (Nc*Ns*Nreim) * Ls * vol * (read+write) (/2 for red black counting)
+    // read = 2 ( psi[ss+s+1] and psi[ss+s-1] count as 1 )
+    // write = 1
+    RealD Gbytes = sizeof(Real) * (Nc*Ns*2) * volume * 3 /2. * 1.e-9;
+    std::cout << GridLogMessage << "Average bandwidth (GB/s)                 : " << Gbytes/M5Dtime*M5Dcalls*1.e6 << std::endl;
+  }
+
+  if ( MooeeInvCalls > 0 ) {
+
+    std::cout << GridLogMessage << "#### MooeeInv calls report " << std::endl;
+    std::cout << GridLogMessage << "CayleyFermion5D Number of MooeeInv Calls     : " << MooeeInvCalls   << std::endl;
+    std::cout << GridLogMessage << "CayleyFermion5D ComputeTime/Calls            : " << MooeeInvTime / MooeeInvCalls << " us" << std::endl;
+#ifdef GRID_CUDA
+    RealD mflops = ( -16.*Nc*Ns+this->Ls*(1.+18.*Nc*Ns) )*volume*MooeeInvCalls/MooeeInvTime/2; // 2 for red black counting
+    std::cout << GridLogMessage << "Average mflops/s per call                : " << mflops << std::endl;
+    std::cout << GridLogMessage << "Average mflops/s per call per rank       : " << mflops/NP << std::endl;
+#else
+    // Flops = MADD * Ls *Ls *4dvol * spin/colour/complex
+    RealD mflops = 2.0*24*this->Ls*volume*MooeeInvCalls/MooeeInvTime/2; // 2 for red black counting
+    std::cout << GridLogMessage << "Average mflops/s per call                : " << mflops << std::endl;
+    std::cout << GridLogMessage << "Average mflops/s per call per rank       : " << mflops/NP << std::endl;
+#endif
+  }
+
+}
+template<class Impl> void CayleyFermion5D<Impl>::CayleyZeroCounters(void)
+{
+  this->ZeroCounters();
+  M5Dflops=0;
+  M5Dcalls=0;
+  M5Dtime=0;
+  MooeeInvFlops=0;
+  MooeeInvCalls=0;
+  MooeeInvTime=0;
+}
+
 template<class Impl>  
 void CayleyFermion5D<Impl>::M5D   (const FermionField &psi, FermionField &chi)
 {
   int Ls=this->Ls;
   Vector<Coeff_t> diag (Ls,1.0);
-  Vector<Coeff_t> upper(Ls,-1.0); upper[Ls-1]=mass_minus;
+  Vector<Coeff_t> upper(Ls,-1.0); upper[Ls-1]=mass;
-  Vector<Coeff_t> lower(Ls,-1.0); lower[0]   =mass_plus;
+  Vector<Coeff_t> lower(Ls,-1.0); lower[0]   =mass;
   M5D(psi,chi,chi,lower,diag,upper);
 }
 template<class Impl>
@@ -168,8 +220,8 @@ void CayleyFermion5D<Impl>::Meooe5D    (const FermionField &psi, FermionField &D
   Vector<Coeff_t> diag = bs;
   Vector<Coeff_t> upper= cs;
   Vector<Coeff_t> lower= cs; 
-  upper[Ls-1]=-mass_minus*upper[Ls-1];
+  upper[Ls-1]=-mass*upper[Ls-1];
-  lower[0]   =-mass_plus*lower[0];
+  lower[0]   =-mass*lower[0];
   M5D(psi,psi,Din,lower,diag,upper);
 }
 // FIXME Redunant with the above routine; check this and eliminate
@@ -183,8 +235,8 @@ template<class Impl> void CayleyFermion5D<Impl>::Meo5D     (const FermionField &
     upper[i]=-ceo[i];
     lower[i]=-ceo[i];
   }
-  upper[Ls-1]=-mass_minus*upper[Ls-1];
+  upper[Ls-1]=-mass*upper[Ls-1];
-  lower[0]   =-mass_plus*lower[0];
+  lower[0]   =-mass*lower[0];
   M5D(psi,psi,chi,lower,diag,upper);
 }
 template<class Impl>
@@ -198,8 +250,8 @@ void CayleyFermion5D<Impl>::Mooee       (const FermionField &psi, FermionField &
     upper[i]=-cee[i];
     lower[i]=-cee[i];
   }
-  upper[Ls-1]=-mass_minus*upper[Ls-1];
+  upper[Ls-1]=-mass*upper[Ls-1];
-  lower[0]   =-mass_plus*lower[0];
+  lower[0]   =-mass*lower[0];
   M5D(psi,psi,chi,lower,diag,upper);
 }
 template<class Impl>
@@ -214,9 +266,9 @@ void CayleyFermion5D<Impl>::MooeeDag    (const FermionField &psi, FermionField &
     // Assemble the 5d matrix
     if ( s==0 ) {
       upper[s] = -cee[s+1] ;
-      lower[s] = mass_minus*cee[Ls-1];
+      lower[s] = mass*cee[Ls-1];
     } else if ( s==(Ls-1)) { 
-      upper[s] = mass_plus*cee[0];
+      upper[s] = mass*cee[0];
       lower[s] = -cee[s-1];
     } else {
       upper[s]=-cee[s+1];
@@ -239,8 +291,8 @@ void CayleyFermion5D<Impl>::M5Ddag (const FermionField &psi, FermionField &chi)
   Vector<Coeff_t> diag(Ls,1.0);
   Vector<Coeff_t> upper(Ls,-1.0);
   Vector<Coeff_t> lower(Ls,-1.0);
-  upper[Ls-1]=-mass_plus*upper[Ls-1];
+  upper[Ls-1]=-mass*upper[Ls-1];
-  lower[0]   =-mass_minus*lower[0];
+  lower[0]   =-mass*lower[0];
   M5Ddag(psi,chi,chi,lower,diag,upper);
 }
 
@@ -255,9 +307,9 @@ void CayleyFermion5D<Impl>::MeooeDag5D    (const FermionField &psi, FermionField
   for (int s=0;s<Ls;s++){
     if ( s== 0 ) {
       upper[s] = cs[s+1];
-      lower[s] =-mass_minus*cs[Ls-1];
+      lower[s] =-mass*cs[Ls-1];
     } else if ( s==(Ls-1) ) { 
-      upper[s] =-mass_plus*cs[0];
+      upper[s] =-mass*cs[0];
       lower[s] = cs[s-1];
     } else { 
       upper[s] = cs[s+1];
@@ -500,7 +552,7 @@ void CayleyFermion5D<Impl>::SetCoefficientsInternal(RealD zolo_hi,Vector<Coeff_t
       
       lee[i] =-cee[i+1]/bee[i]; // sub-diag entry on the ith column
       
-      leem[i]=mass_minus*cee[Ls-1]/bee[0];
+      leem[i]=mass*cee[Ls-1]/bee[0];
       for(int j=0;j<i;j++) {
 	assert(bee[j+1]!=Coeff_t(0.0));
 	leem[i]*= aee[j]/bee[j+1];
@@ -508,7 +560,7 @@ void CayleyFermion5D<Impl>::SetCoefficientsInternal(RealD zolo_hi,Vector<Coeff_t
       
       uee[i] =-aee[i]/bee[i];   // up-diag entry on the ith row
       
-      ueem[i]=mass_plus;
+      ueem[i]=mass;
       for(int j=1;j<=i;j++) ueem[i]*= cee[j]/bee[j];
       ueem[i]*= aee[0]/bee[0];
       
@@ -521,7 +573,7 @@ void CayleyFermion5D<Impl>::SetCoefficientsInternal(RealD zolo_hi,Vector<Coeff_t
   }
 	
   { 
-    Coeff_t delta_d=mass_minus*cee[Ls-1];
+    Coeff_t delta_d=mass*cee[Ls-1];
     for(int j=0;j<Ls-1;j++) {
       assert(bee[j] != Coeff_t(0.0));
       delta_d *= cee[j]/bee[j];
@@ -590,10 +642,7 @@ void CayleyFermion5D<Impl>::ContractConservedCurrent( PropagatorField &q_in_1,
 						      Current curr_type,
 						      unsigned int mu)
 {
-
+#if (!defined(GRID_HIP))
-  assert(mass_plus == mass_minus);
-  RealD mass = mass_plus;
-  
   Gamma::Algebra Gmu [] = {
     Gamma::Algebra::GammaX,
     Gamma::Algebra::GammaY,
@@ -712,7 +761,7 @@ void CayleyFermion5D<Impl>::ContractConservedCurrent( PropagatorField &q_in_1,
     else          q_out +=     C;
     
   }
-
+#endif
 }
 
 template <class Impl>
@@ -728,8 +777,6 @@ void CayleyFermion5D<Impl>::SeqConservedCurrent(PropagatorField &q_in,
   assert(mu>=0);
   assert(mu<Nd);
 
-  assert(mass_plus == mass_minus);
-  RealD mass = mass_plus;
 
 #if 0
   int tshift = (mu == Nd-1) ? 1 : 0;
@@ -779,6 +826,7 @@ void CayleyFermion5D<Impl>::SeqConservedCurrent(PropagatorField &q_in,
   }
 #endif
 
+#if (!defined(GRID_HIP))
   int tshift = (mu == Nd-1) ? 1 : 0;
   unsigned int LLt    = GridDefaultLatt()[Tp];
   ////////////////////////////////////////////////
@@ -898,6 +946,7 @@ void CayleyFermion5D<Impl>::SeqConservedCurrent(PropagatorField &q_in,
 
     InsertSlice(L_Q, q_out, s , 0);
   }
+#endif
 }
 #undef Pp
 #undef Pm
@@ -905,6 +954,88 @@ void CayleyFermion5D<Impl>::SeqConservedCurrent(PropagatorField &q_in,
 #undef TopRowWithSource
 
 
+
+#if 0
+template<class Impl>
+void CayleyFermion5D<Impl>::MooeeInternalCompute(int dag, int inv,
+						 Vector<iSinglet<Simd> > & Matp,
+						 Vector<iSinglet<Simd> > & Matm)
+{
+  int Ls=this->Ls;
+
+  GridBase *grid = this->FermionRedBlackGrid();
+  int LLs = grid->_rdimensions[0];
+
+  if ( LLs == Ls ) {
+    return; // Not vectorised in 5th direction
+  }
+
+  Eigen::MatrixXcd Pplus  = Eigen::MatrixXcd::Zero(Ls,Ls);
+  Eigen::MatrixXcd Pminus = Eigen::MatrixXcd::Zero(Ls,Ls);
+  
+  for(int s=0;s<Ls;s++){
+    Pplus(s,s) = bee[s];
+    Pminus(s,s)= bee[s];
+  }
+  
+  for(int s=0;s<Ls-1;s++){
+    Pminus(s,s+1) = -cee[s];
+  }
+  
+  for(int s=0;s<Ls-1;s++){
+    Pplus(s+1,s) = -cee[s+1];
+  }
+  Pplus (0,Ls-1) = mass*cee[0];
+  Pminus(Ls-1,0) = mass*cee[Ls-1];
+  
+  Eigen::MatrixXcd PplusMat ;
+  Eigen::MatrixXcd PminusMat;
+  
+  if ( inv ) {
+    PplusMat =Pplus.inverse();
+    PminusMat=Pminus.inverse();
+  } else { 
+    PplusMat =Pplus;
+    PminusMat=Pminus;
+  }
+  
+  if(dag){
+    PplusMat.adjointInPlace();
+    PminusMat.adjointInPlace();
+  }
+  
+  typedef typename SiteHalfSpinor::scalar_type scalar_type;
+  const int Nsimd=Simd::Nsimd();
+  Matp.resize(Ls*LLs);
+  Matm.resize(Ls*LLs);
+
+  for(int s2=0;s2<Ls;s2++){
+    for(int s1=0;s1<LLs;s1++){
+      int istride = LLs;
+      int ostride = 1;
+      Simd Vp;
+      Simd Vm;
+      scalar_type *sp = (scalar_type *)&Vp;
+      scalar_type *sm = (scalar_type *)&Vm;
+      for(int l=0;l<Nsimd;l++){
+	if ( switcheroo<Coeff_t>::iscomplex() ) {
+	  sp[l] = PplusMat (l*istride+s1*ostride,s2);
+	  sm[l] = PminusMat(l*istride+s1*ostride,s2);
+	} else { 
+	  // if real
+	  scalar_type tmp;
+	  tmp = PplusMat (l*istride+s1*ostride,s2);
+	  sp[l] = scalar_type(tmp.real(),tmp.real());
+	  tmp = PminusMat(l*istride+s1*ostride,s2);
+	  sm[l] = scalar_type(tmp.real(),tmp.real());
+	}
+      }
+      Matp[LLs*s2+s1] = Vp;
+      Matm[LLs*s2+s1] = Vm;
+    }}
+}
+#endif
+
 NAMESPACE_END(Grid);
 
 

Grid/qcd/action/fermion/implementation/CayleyFermion5Dcache.h

@@ -63,18 +63,23 @@ CayleyFermion5D<Impl>::M5D(const FermionField &psi_i,
 
   // 10 = 3 complex mult + 2 complex add
   // Flops = 10.0*(Nc*Ns) *Ls*vol (/2 for red black counting)
-  uint64_t nloop = grid->oSites();
+  M5Dcalls++;
+  M5Dtime-=usecond();
+
+  uint64_t nloop = grid->oSites()/Ls;
   accelerator_for(sss,nloop,Simd::Nsimd(),{
-    uint64_t s = sss%Ls;
+    uint64_t ss= sss*Ls;
-    uint64_t ss= sss-s;
     typedef decltype(coalescedRead(psi[0])) spinor;
     spinor tmp1, tmp2;
-    uint64_t idx_u = ss+((s+1)%Ls);
+    for(int s=0;s<Ls;s++){
-    uint64_t idx_l = ss+((s+Ls-1)%Ls);
+      uint64_t idx_u = ss+((s+1)%Ls);
-    spProj5m(tmp1,psi(idx_u));
+      uint64_t idx_l = ss+((s+Ls-1)%Ls);
-    spProj5p(tmp2,psi(idx_l));
+      spProj5m(tmp1,psi(idx_u));
-    coalescedWrite(chi[ss+s],pdiag[s]*phi(ss+s)+pupper[s]*tmp1+plower[s]*tmp2);
+      spProj5p(tmp2,psi(idx_l));
+      coalescedWrite(chi[ss+s],pdiag[s]*phi(ss+s)+pupper[s]*tmp1+plower[s]*tmp2);
+    }
   });
+  M5Dtime+=usecond();
 }
 
 template<class Impl>  
@@ -100,18 +105,23 @@ CayleyFermion5D<Impl>::M5Ddag(const FermionField &psi_i,
   int Ls=this->Ls;
 
   // Flops = 6.0*(Nc*Ns) *Ls*vol
-  uint64_t nloop = grid->oSites();
+  M5Dcalls++;
+  M5Dtime-=usecond();
+
+  uint64_t nloop = grid->oSites()/Ls;
   accelerator_for(sss,nloop,Simd::Nsimd(),{
-    uint64_t s = sss%Ls;
+    uint64_t ss=sss*Ls;
-    uint64_t ss= sss-s;
     typedef decltype(coalescedRead(psi[0])) spinor;
     spinor tmp1,tmp2;
-    uint64_t idx_u = ss+((s+1)%Ls);
+    for(int s=0;s<Ls;s++){
-    uint64_t idx_l = ss+((s+Ls-1)%Ls);
+      uint64_t idx_u = ss+((s+1)%Ls);
-    spProj5p(tmp1,psi(idx_u));
+      uint64_t idx_l = ss+((s+Ls-1)%Ls);
-    spProj5m(tmp2,psi(idx_l));
+      spProj5p(tmp1,psi(idx_u));
-    coalescedWrite(chi[ss+s],pdiag[s]*phi(ss+s)+pupper[s]*tmp1+plower[s]*tmp2);
+      spProj5m(tmp2,psi(idx_l));
+      coalescedWrite(chi[ss+s],pdiag[s]*phi(ss+s)+pupper[s]*tmp1+plower[s]*tmp2);
+    }
   });
+  M5Dtime+=usecond();
 }
 
 template<class Impl>
@@ -132,6 +142,8 @@ CayleyFermion5D<Impl>::MooeeInv    (const FermionField &psi_i, FermionField &chi
   auto pleem = & leem[0];
   auto pueem = & ueem[0];
 
+  MooeeInvCalls++;
+  MooeeInvTime-=usecond();
   uint64_t nloop = grid->oSites()/Ls;
   accelerator_for(sss,nloop,Simd::Nsimd(),{
     uint64_t ss=sss*Ls;
@@ -168,6 +180,8 @@ CayleyFermion5D<Impl>::MooeeInv    (const FermionField &psi_i, FermionField &chi
       coalescedWrite(chi[ss+s],res);
     }
   });
+
+  MooeeInvTime+=usecond();
   
 }
 
@@ -190,6 +204,10 @@ CayleyFermion5D<Impl>::MooeeInvDag (const FermionField &psi_i, FermionField &chi
 
   assert(psi.Checkerboard() == psi.Checkerboard());
 
+  MooeeInvCalls++;
+  MooeeInvTime-=usecond();
+
+
   uint64_t nloop = grid->oSites()/Ls;
   accelerator_for(sss,nloop,Simd::Nsimd(),{
     uint64_t ss=sss*Ls;
@@ -226,6 +244,7 @@ CayleyFermion5D<Impl>::MooeeInvDag (const FermionField &psi_i, FermionField &chi
       coalescedWrite(chi[ss+s],res);
     }
   });
+  MooeeInvTime+=usecond();
 
 }
 

Grid/qcd/action/fermion/implementation/CayleyFermion5Dvec.h

@@ -94,6 +94,10 @@ CayleyFermion5D<Impl>::M5D(const FermionField &psi_i,
       d_p[ss] = diag[s];
     }}
 
+
+  M5Dcalls++;
+  M5Dtime-=usecond();
+
   assert(Nc==3);
 
   thread_loop( (int ss=0;ss<grid->oSites();ss+=LLs),{ // adds LLs
@@ -194,6 +198,7 @@ CayleyFermion5D<Impl>::M5D(const FermionField &psi_i,
     }
 #endif
   });
+  M5Dtime+=usecond();
 }
 
 template<class Impl>  
@@ -237,6 +242,8 @@ CayleyFermion5D<Impl>::M5Ddag(const FermionField &psi_i,
       d_p[ss] = diag[s];
     }}
 
+  M5Dcalls++;
+  M5Dtime-=usecond();
   thread_loop( (int ss=0;ss<grid->oSites();ss+=LLs),{ // adds LLs
 #if 0
     alignas(64) SiteHalfSpinor hp;
@@ -332,6 +339,7 @@ CayleyFermion5D<Impl>::M5Ddag(const FermionField &psi_i,
     }
 #endif
   });
+  M5Dtime+=usecond();
 }
 
 
@@ -805,6 +813,9 @@ CayleyFermion5D<Impl>::MooeeInternal(const FermionField &psi, FermionField &chi,
   }
   assert(_Matp->size()==Ls*LLs);
 
+  MooeeInvCalls++;
+  MooeeInvTime-=usecond();
+
   if ( switcheroo<Coeff_t>::iscomplex() ) {
     thread_loop( (auto site=0;site<vol;site++),{
       MooeeInternalZAsm(psi,chi,LLs,site,*_Matp,*_Matm);
@@ -814,7 +825,7 @@ CayleyFermion5D<Impl>::MooeeInternal(const FermionField &psi, FermionField &chi,
       MooeeInternalAsm(psi,chi,LLs,site,*_Matp,*_Matm);
     });
   }
-
+  MooeeInvTime+=usecond();
 }
 
 NAMESPACE_END(Grid);

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

@@ -32,23 +32,22 @@
 #include <Grid/qcd/spin/Dirac.h>
 #include <Grid/qcd/action/fermion/CompactWilsonCloverFermion.h>
 
-
 NAMESPACE_BEGIN(Grid);
-template<class Impl, class CloverHelpers>
+template<class Impl>
-CompactWilsonCloverFermion<Impl, CloverHelpers>::CompactWilsonCloverFermion(GaugeField& _Umu,
+CompactWilsonCloverFermion<Impl>::CompactWilsonCloverFermion(GaugeField& _Umu,
-                                                                            GridCartesian& Fgrid,
+                                                             GridCartesian& Fgrid,
-                                                                            GridRedBlackCartesian& Hgrid,
+                                                             GridRedBlackCartesian& Hgrid,
-                                                                            const RealD _mass,
+                                                             const RealD _mass,
-                                                                            const RealD _csw_r,
+                                                             const RealD _csw_r,
-                                                                            const RealD _csw_t,
+                                                             const RealD _csw_t,
-                                                                            const RealD _cF,
+                                                             const RealD _cF,
-                                                                            const WilsonAnisotropyCoefficients& clover_anisotropy,
+                                                             const WilsonAnisotropyCoefficients& clover_anisotropy,
-                                                                            const ImplParams& impl_p)
+                                                             const ImplParams& impl_p)
   : WilsonBase(_Umu, Fgrid, Hgrid, _mass, impl_p, clover_anisotropy)
   , csw_r(_csw_r)
   , csw_t(_csw_t)
   , cF(_cF)
-  , fixedBoundaries(impl_p.boundary_phases[Nd-1] == 0.0)
+  , open_boundaries(impl_p.boundary_phases[Nd-1] == 0.0)
   , Diagonal(&Fgrid),        Triangle(&Fgrid)
   , DiagonalEven(&Hgrid),    TriangleEven(&Hgrid)
   , DiagonalOdd(&Hgrid),     TriangleOdd(&Hgrid)
@@ -59,85 +58,80 @@ CompactWilsonCloverFermion<Impl, CloverHelpers>::CompactWilsonCloverFermion(Gaug
   , BoundaryMask(&Fgrid)
   , BoundaryMaskEven(&Hgrid), BoundaryMaskOdd(&Hgrid)
 {
-  assert(Nd == 4 && Nc == 3 && Ns == 4 && Impl::Dimension == 3);
-
   csw_r *= 0.5;
   csw_t *= 0.5;
   if (clover_anisotropy.isAnisotropic)
     csw_r /= clover_anisotropy.xi_0;
 
   ImportGauge(_Umu);
-  if (fixedBoundaries) {
+  if (open_boundaries)
-    this->BoundaryMaskEven.Checkerboard() = Even;
-    this->BoundaryMaskOdd.Checkerboard() = Odd;
     CompactHelpers::SetupMasks(this->BoundaryMask, this->BoundaryMaskEven, this->BoundaryMaskOdd);
-  }
 }
 
-template<class Impl, class CloverHelpers>
+template<class Impl>
-void CompactWilsonCloverFermion<Impl, CloverHelpers>::Dhop(const FermionField& in, FermionField& out, int dag) {
+void CompactWilsonCloverFermion<Impl>::Dhop(const FermionField& in, FermionField& out, int dag) {
   WilsonBase::Dhop(in, out, dag);
-  if(fixedBoundaries) ApplyBoundaryMask(out);
+  if(open_boundaries) ApplyBoundaryMask(out);
 }
 
-template<class Impl, class CloverHelpers>
+template<class Impl>
-void CompactWilsonCloverFermion<Impl, CloverHelpers>::DhopOE(const FermionField& in, FermionField& out, int dag) {
+void CompactWilsonCloverFermion<Impl>::DhopOE(const FermionField& in, FermionField& out, int dag) {
   WilsonBase::DhopOE(in, out, dag);
-  if(fixedBoundaries) ApplyBoundaryMask(out);
+  if(open_boundaries) ApplyBoundaryMask(out);
 }
 
-template<class Impl, class CloverHelpers>
+template<class Impl>
-void CompactWilsonCloverFermion<Impl, CloverHelpers>::DhopEO(const FermionField& in, FermionField& out, int dag) {
+void CompactWilsonCloverFermion<Impl>::DhopEO(const FermionField& in, FermionField& out, int dag) {
   WilsonBase::DhopEO(in, out, dag);
-  if(fixedBoundaries) ApplyBoundaryMask(out);
+  if(open_boundaries) ApplyBoundaryMask(out);
 }
 
-template<class Impl, class CloverHelpers>
+template<class Impl>
-void CompactWilsonCloverFermion<Impl, CloverHelpers>::DhopDir(const FermionField& in, FermionField& out, int dir, int disp) {
+void CompactWilsonCloverFermion<Impl>::DhopDir(const FermionField& in, FermionField& out, int dir, int disp) {
   WilsonBase::DhopDir(in, out, dir, disp);
-  if(this->fixedBoundaries) ApplyBoundaryMask(out);
+  if(this->open_boundaries) ApplyBoundaryMask(out);
 }
 
-template<class Impl, class CloverHelpers>
+template<class Impl>
-void CompactWilsonCloverFermion<Impl, CloverHelpers>::DhopDirAll(const FermionField& in, std::vector<FermionField>& out) {
+void CompactWilsonCloverFermion<Impl>::DhopDirAll(const FermionField& in, std::vector<FermionField>& out) {
   WilsonBase::DhopDirAll(in, out);
-  if(this->fixedBoundaries) {
+  if(this->open_boundaries) {
     for(auto& o : out) ApplyBoundaryMask(o);
   }
 }
 
-template<class Impl, class CloverHelpers>
+template<class Impl>
-void CompactWilsonCloverFermion<Impl, CloverHelpers>::M(const FermionField& in, FermionField& out) {
+void CompactWilsonCloverFermion<Impl>::M(const FermionField& in, FermionField& out) {
   out.Checkerboard() = in.Checkerboard();
   WilsonBase::Dhop(in, out, DaggerNo); // call base to save applying bc
   Mooee(in, Tmp);
   axpy(out, 1.0, out, Tmp);
-  if(fixedBoundaries) ApplyBoundaryMask(out);
+  if(open_boundaries) ApplyBoundaryMask(out);
 }
 
-template<class Impl, class CloverHelpers>
+template<class Impl>
-void CompactWilsonCloverFermion<Impl, CloverHelpers>::Mdag(const FermionField& in, FermionField& out) {
+void CompactWilsonCloverFermion<Impl>::Mdag(const FermionField& in, FermionField& out) {
   out.Checkerboard() = in.Checkerboard();
   WilsonBase::Dhop(in, out, DaggerYes);  // call base to save applying bc
   MooeeDag(in, Tmp);
   axpy(out, 1.0, out, Tmp);
-  if(fixedBoundaries) ApplyBoundaryMask(out);
+  if(open_boundaries) ApplyBoundaryMask(out);
 }
 
-template<class Impl, class CloverHelpers>
+template<class Impl>
-void CompactWilsonCloverFermion<Impl, CloverHelpers>::Meooe(const FermionField& in, FermionField& out) {
+void CompactWilsonCloverFermion<Impl>::Meooe(const FermionField& in, FermionField& out) {
   WilsonBase::Meooe(in, out);
-  if(fixedBoundaries) ApplyBoundaryMask(out);
+  if(open_boundaries) ApplyBoundaryMask(out);
 }
 
-template<class Impl, class CloverHelpers>
+template<class Impl>
-void CompactWilsonCloverFermion<Impl, CloverHelpers>::MeooeDag(const FermionField& in, FermionField& out) {
+void CompactWilsonCloverFermion<Impl>::MeooeDag(const FermionField& in, FermionField& out) {
   WilsonBase::MeooeDag(in, out);
-  if(fixedBoundaries) ApplyBoundaryMask(out);
+  if(open_boundaries) ApplyBoundaryMask(out);
 }
 
-template<class Impl, class CloverHelpers>
+template<class Impl>
-void CompactWilsonCloverFermion<Impl, CloverHelpers>::Mooee(const FermionField& in, FermionField& out) {
+void CompactWilsonCloverFermion<Impl>::Mooee(const FermionField& in, FermionField& out) {
   if(in.Grid()->_isCheckerBoarded) {
     if(in.Checkerboard() == Odd) {
       MooeeInternal(in, out, DiagonalOdd, TriangleOdd);
@@ -147,16 +141,16 @@ void CompactWilsonCloverFermion<Impl, CloverHelpers>::Mooee(const FermionField&
   } else {
     MooeeInternal(in, out, Diagonal, Triangle);
   }
-  if(fixedBoundaries) ApplyBoundaryMask(out);
+  if(open_boundaries) ApplyBoundaryMask(out);
 }
 
-template<class Impl, class CloverHelpers>
+template<class Impl>
-void CompactWilsonCloverFermion<Impl, CloverHelpers>::MooeeDag(const FermionField& in, FermionField& out) {
+void CompactWilsonCloverFermion<Impl>::MooeeDag(const FermionField& in, FermionField& out) {
   Mooee(in, out); // blocks are hermitian
 }
 
-template<class Impl, class CloverHelpers>
+template<class Impl>
-void CompactWilsonCloverFermion<Impl, CloverHelpers>::MooeeInv(const FermionField& in, FermionField& out) {
+void CompactWilsonCloverFermion<Impl>::MooeeInv(const FermionField& in, FermionField& out) {
   if(in.Grid()->_isCheckerBoarded) {
     if(in.Checkerboard() == Odd) {
       MooeeInternal(in, out, DiagonalInvOdd, TriangleInvOdd);
@@ -166,27 +160,27 @@ void CompactWilsonCloverFermion<Impl, CloverHelpers>::MooeeInv(const FermionFiel
   } else {
     MooeeInternal(in, out, DiagonalInv, TriangleInv);
   }
-  if(fixedBoundaries) ApplyBoundaryMask(out);
+  if(open_boundaries) ApplyBoundaryMask(out);
 }
 
-template<class Impl, class CloverHelpers>
+template<class Impl>
-void CompactWilsonCloverFermion<Impl, CloverHelpers>::MooeeInvDag(const FermionField& in, FermionField& out) {
+void CompactWilsonCloverFermion<Impl>::MooeeInvDag(const FermionField& in, FermionField& out) {
   MooeeInv(in, out); // blocks are hermitian
 }
 
-template<class Impl, class CloverHelpers>
+template<class Impl>
-void CompactWilsonCloverFermion<Impl, CloverHelpers>::Mdir(const FermionField& in, FermionField& out, int dir, int disp) {
+void CompactWilsonCloverFermion<Impl>::Mdir(const FermionField& in, FermionField& out, int dir, int disp) {
   DhopDir(in, out, dir, disp);
 }
 
-template<class Impl, class CloverHelpers>
+template<class Impl>
-void CompactWilsonCloverFermion<Impl, CloverHelpers>::MdirAll(const FermionField& in, std::vector<FermionField>& out) {
+void CompactWilsonCloverFermion<Impl>::MdirAll(const FermionField& in, std::vector<FermionField>& out) {
   DhopDirAll(in, out);
 }
 
-template<class Impl, class CloverHelpers>
+template<class Impl>
-void CompactWilsonCloverFermion<Impl, CloverHelpers>::MDeriv(GaugeField& force, const FermionField& X, const FermionField& Y, int dag) {
+void CompactWilsonCloverFermion<Impl>::MDeriv(GaugeField& force, const FermionField& X, const FermionField& Y, int dag) {
-  assert(!fixedBoundaries); // TODO check for changes required for open bc
+  assert(!open_boundaries); // TODO check for changes required for open bc
 
   // NOTE: code copied from original clover term
   conformable(X.Grid(), Y.Grid());
@@ -257,7 +251,7 @@ void CompactWilsonCloverFermion<Impl, CloverHelpers>::MDeriv(GaugeField& force,
       }
       PropagatorField Slambda = Gamma(sigma[count]) * Lambda; // sigma checked
       Impl::TraceSpinImpl(lambda, Slambda);                   // traceSpin ok
-      force_mu -= factor*CloverHelpers::Cmunu(U, lambda, mu, nu);   // checked
+      force_mu -= factor*Helpers::Cmunu(U, lambda, mu, nu);   // checked
       count++;
     }
 
@@ -267,18 +261,18 @@ void CompactWilsonCloverFermion<Impl, CloverHelpers>::MDeriv(GaugeField& force,
   force += clover_force;
 }
 
-template<class Impl, class CloverHelpers>
+template<class Impl>
-void CompactWilsonCloverFermion<Impl, CloverHelpers>::MooDeriv(GaugeField& mat, const FermionField& U, const FermionField& V, int dag) {
+void CompactWilsonCloverFermion<Impl>::MooDeriv(GaugeField& mat, const FermionField& U, const FermionField& V, int dag) {
   assert(0);
 }
 
-template<class Impl, class CloverHelpers>
+template<class Impl>
-void CompactWilsonCloverFermion<Impl, CloverHelpers>::MeeDeriv(GaugeField& mat, const FermionField& U, const FermionField& V, int dag) {
+void CompactWilsonCloverFermion<Impl>::MeeDeriv(GaugeField& mat, const FermionField& U, const FermionField& V, int dag) {
   assert(0);
 }
 
-template<class Impl, class CloverHelpers>
+template<class Impl>
-void CompactWilsonCloverFermion<Impl, CloverHelpers>::MooeeInternal(const FermionField&        in,
+void CompactWilsonCloverFermion<Impl>::MooeeInternal(const FermionField&        in,
                     FermionField&              out,
                     const CloverDiagonalField& diagonal,
                     const CloverTriangleField& triangle) {
@@ -291,8 +285,8 @@ void CompactWilsonCloverFermion<Impl, CloverHelpers>::MooeeInternal(const Fermio
   CompactHelpers::MooeeKernel(diagonal.oSites(), 1, in, out, diagonal, triangle);
 }
 
-template<class Impl, class CloverHelpers>
+template<class Impl>
-void CompactWilsonCloverFermion<Impl, CloverHelpers>::ImportGauge(const GaugeField& _Umu) {
+void CompactWilsonCloverFermion<Impl>::ImportGauge(const GaugeField& _Umu) {
   // NOTE: parts copied from original implementation
 
   // Import gauge into base class
@@ -305,7 +299,6 @@ void CompactWilsonCloverFermion<Impl, CloverHelpers>::ImportGauge(const GaugeFie
   GridBase* grid = _Umu.Grid();
   typename Impl::GaugeLinkField Bx(grid), By(grid), Bz(grid), Ex(grid), Ey(grid), Ez(grid);
   CloverField TmpOriginal(grid);
-  CloverField TmpInverse(grid);
 
   // Compute the field strength terms mu>nu
   double t2 = usecond();
@@ -325,30 +318,22 @@ void CompactWilsonCloverFermion<Impl, CloverHelpers>::ImportGauge(const GaugeFie
   TmpOriginal += Helpers::fillCloverXT(Ex) * csw_t;
   TmpOriginal += Helpers::fillCloverYT(Ey) * csw_t;
   TmpOriginal += Helpers::fillCloverZT(Ez) * csw_t;
-
+  TmpOriginal += this->diag_mass;
-  // Instantiate the clover term
-  // - In case of the standard clover the mass term is added
-  // - In case of the exponential clover the clover term is exponentiated
-  double t4 = usecond();
-  CloverHelpers::InstantiateClover(TmpOriginal, TmpInverse, csw_t, this->diag_mass);
 
   // Convert the data layout of the clover term
-  double t5 = usecond();
+  double t4 = usecond();
   CompactHelpers::ConvertLayout(TmpOriginal, Diagonal, Triangle);
 
-  // Modify the clover term at the temporal boundaries in case of open boundary conditions
+  // Possible modify the boundary values
-  double t6 = usecond();
+  double t5 = usecond();
-  if(fixedBoundaries) CompactHelpers::ModifyBoundaries(Diagonal, Triangle, csw_t, cF, this->diag_mass);
+  if(open_boundaries) CompactHelpers::ModifyBoundaries(Diagonal, Triangle, csw_t, cF, this->diag_mass);
 
-  // Invert the Clover term
+  // Invert the clover term in the improved layout
-  // In case of the exponential clover with (anti-)periodic boundary conditions exp(-Clover) saved
+  double t6 = usecond();
-  // in TmpInverse can be used. In all other cases the clover term has to be explictly inverted.
+  CompactHelpers::Invert(Diagonal, Triangle, DiagonalInv, TriangleInv);
-  // TODO: For now this inversion is explictly done on the CPU
-  double t7 = usecond();
-  CloverHelpers::InvertClover(TmpInverse, Diagonal, Triangle, DiagonalInv, TriangleInv, fixedBoundaries);
 
   // Fill the remaining clover fields
-  double t8 = usecond();
+  double t7 = usecond();
   pickCheckerboard(Even, DiagonalEven,    Diagonal);
   pickCheckerboard(Even, TriangleEven,    Triangle);
   pickCheckerboard(Odd,  DiagonalOdd,     Diagonal);
@@ -359,19 +344,20 @@ void CompactWilsonCloverFermion<Impl, CloverHelpers>::ImportGauge(const GaugeFie
   pickCheckerboard(Odd,  TriangleInvOdd,  TriangleInv);
 
   // Report timings
-  double t9 = usecond();
+  double t8 = usecond();
-
+#if 0
-  std::cout << GridLogDebug << "CompactWilsonCloverFermion::ImportGauge timings:" << std::endl;
+  std::cout << GridLogMessage << "CompactWilsonCloverFermion::ImportGauge timings:"
-  std::cout << GridLogDebug << "WilsonFermion::Importgauge = " << (t1 - t0) / 1e6 << std::endl;
+            << " WilsonFermion::Importgauge = " << (t1 - t0) / 1e6
-  std::cout << GridLogDebug << "allocations =                " << (t2 - t1) / 1e6 << std::endl;
+            << ", allocations = "               << (t2 - t1) / 1e6
-  std::cout << GridLogDebug << "field strength =             " << (t3 - t2) / 1e6 << std::endl;
+            << ", field strength = "            << (t3 - t2) / 1e6
-  std::cout << GridLogDebug << "fill clover =                " << (t4 - t3) / 1e6 << std::endl;
+            << ", fill clover = "               << (t4 - t3) / 1e6
-  std::cout << GridLogDebug << "instantiate clover =         " << (t5 - t4) / 1e6 << std::endl;
+            << ", convert = "                   << (t5 - t4) / 1e6
-  std::cout << GridLogDebug << "convert layout =             " << (t6 - t5) / 1e6 << std::endl;
+            << ", boundaries = "                << (t6 - t5) / 1e6
-  std::cout << GridLogDebug << "modify boundaries =          " << (t7 - t6) / 1e6 << std::endl;
+            << ", inversions = "                << (t7 - t6) / 1e6
-  std::cout << GridLogDebug << "invert clover =              " << (t8 - t7) / 1e6 << std::endl;
+            << ", pick cbs = "                  << (t8 - t7) / 1e6
-  std::cout << GridLogDebug << "pick cbs =                   " << (t9 - t8) / 1e6 << std::endl;
+            << ", total = "                     << (t8 - t0) / 1e6
-  std::cout << GridLogDebug << "total =                      " << (t9 - t0) / 1e6 << std::endl;
+            << std::endl;
+#endif
 }
 
 NAMESPACE_END(Grid);

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

@@ -54,6 +54,8 @@ void DomainWallEOFAFermion<Impl>::M5D(const FermionField& psi_i, const FermionFi
   auto pupper = &upper[0];
   auto plower = &lower[0];
   // Flops = 6.0*(Nc*Ns) *Ls*vol
+  this->M5Dcalls++;
+  this->M5Dtime -= usecond();
   
   auto nloop=grid->oSites()/Ls;
   accelerator_for(sss,nloop,Simd::Nsimd(),{
@@ -69,6 +71,7 @@ void DomainWallEOFAFermion<Impl>::M5D(const FermionField& psi_i, const FermionFi
     }
   });
 
+  this->M5Dtime += usecond();
 }
 
 template<class Impl>
@@ -88,6 +91,8 @@ void DomainWallEOFAFermion<Impl>::M5Ddag(const FermionField& psi_i, const Fermio
   auto plower = &lower[0];
 
   // Flops = 6.0*(Nc*Ns) *Ls*vol
+  this->M5Dcalls++;
+  this->M5Dtime -= usecond();
 
   auto nloop=grid->oSites()/Ls;
   accelerator_for(sss,nloop,Simd::Nsimd(),{
@@ -103,6 +108,7 @@ void DomainWallEOFAFermion<Impl>::M5Ddag(const FermionField& psi_i, const Fermio
     }
   });
 
+  this->M5Dtime += usecond();
 }
 
 template<class Impl>
@@ -121,6 +127,8 @@ void DomainWallEOFAFermion<Impl>::MooeeInv(const FermionField& psi_i, FermionFie
   auto pleem = & this->leem[0];
   auto pueem = & this->ueem[0];
 
+  this->MooeeInvCalls++;
+  this->MooeeInvTime -= usecond();
   uint64_t nloop=grid->oSites()/Ls;
   accelerator_for(sss,nloop,Simd::Nsimd(),{
     uint64_t ss=sss*Ls;
@@ -156,6 +164,7 @@ void DomainWallEOFAFermion<Impl>::MooeeInv(const FermionField& psi_i, FermionFie
       coalescedWrite(chi[ss+s],res);
     }
   });
+  this->MooeeInvTime += usecond();
 }
 
 template<class Impl>
@@ -176,6 +185,8 @@ void DomainWallEOFAFermion<Impl>::MooeeInvDag(const FermionField& psi_i, Fermion
 
   assert(psi.Checkerboard() == psi.Checkerboard());
 
+  this->MooeeInvCalls++;
+  this->MooeeInvTime -= usecond();
   auto nloop = grid->oSites()/Ls;
   accelerator_for(sss,nloop,Simd::Nsimd(),{
     uint64_t ss=sss*Ls;
@@ -212,6 +223,7 @@ void DomainWallEOFAFermion<Impl>::MooeeInvDag(const FermionField& psi_i, Fermion
     }
   });
 
+  this->MooeeInvTime += usecond();
 }
 
 NAMESPACE_END(Grid);

Grid/qcd/action/fermion/implementation/ImprovedStaggeredFermion5DImplementation.h

@@ -298,33 +298,45 @@ void ImprovedStaggeredFermion5D<Impl>::DhopInternalOverlappedComms(StencilImpl &
   int LLs = in.Grid()->_rdimensions[0];
   int len =  U.Grid()->oSites();
 
+  DhopFaceTime-=usecond();
   st.Prepare();
   st.HaloGather(in,compressor);
+  DhopFaceTime+=usecond();
 
+  DhopCommTime -=usecond();
   std::vector<std::vector<CommsRequest_t> > requests;
   st.CommunicateBegin(requests);
 
   //  st.HaloExchangeOptGather(in,compressor); // Wilson compressor
+  DhopFaceTime-=usecond();
   st.CommsMergeSHM(compressor);// Could do this inside parallel region overlapped with comms
+  DhopFaceTime+=usecond();
 
   //////////////////////////////////////////////////////////////////////////////////////////////////////
   // Remove explicit thread mapping introduced for OPA reasons.
   //////////////////////////////////////////////////////////////////////////////////////////////////////
+  DhopComputeTime-=usecond();
   {
     int interior=1;
     int exterior=0;
     Kernels::DhopImproved(st,lo,U,UUU,in,out,dag,interior,exterior);
   }
+  DhopComputeTime+=usecond();
 
+  DhopFaceTime-=usecond();
   st.CommsMerge(compressor);
+  DhopFaceTime+=usecond();
 
   st.CommunicateComplete(requests);
+  DhopCommTime +=usecond();
 
+  DhopComputeTime2-=usecond();
   {
     int interior=0;
     int exterior=1;
     Kernels::DhopImproved(st,lo,U,UUU,in,out,dag,interior,exterior);
   }
+  DhopComputeTime2+=usecond();
 }
 
 template<class Impl>
@@ -335,14 +347,22 @@ void ImprovedStaggeredFermion5D<Impl>::DhopInternalSerialComms(StencilImpl & st,
   Compressor compressor;
   int LLs = in.Grid()->_rdimensions[0];
 
+ //double t1=usecond();
+  DhopTotalTime -= usecond();
+  DhopCommTime -= usecond();
   st.HaloExchange(in,compressor);
+  DhopCommTime += usecond();
   
+  DhopComputeTime -= usecond();
   // Dhop takes the 4d grid from U, and makes a 5d index for fermion
   {
     int interior=1;
     int exterior=1;
     Kernels::DhopImproved(st,lo,U,UUU,in,out,dag,interior,exterior);
   }
+  DhopComputeTime += usecond();
+  DhopTotalTime   += usecond();
+
 }
 /*CHANGE END*/
 
@@ -351,6 +371,7 @@ void ImprovedStaggeredFermion5D<Impl>::DhopInternalSerialComms(StencilImpl & st,
 template<class Impl>
 void ImprovedStaggeredFermion5D<Impl>::DhopOE(const FermionField &in, FermionField &out,int dag)
 {
+  DhopCalls+=1;
   conformable(in.Grid(),FermionRedBlackGrid());    // verifies half grid
   conformable(in.Grid(),out.Grid()); // drops the cb check
 
@@ -362,6 +383,7 @@ void ImprovedStaggeredFermion5D<Impl>::DhopOE(const FermionField &in, FermionFie
 template<class Impl>
 void ImprovedStaggeredFermion5D<Impl>::DhopEO(const FermionField &in, FermionField &out,int dag)
 {
+  DhopCalls+=1;
   conformable(in.Grid(),FermionRedBlackGrid());    // verifies half grid
   conformable(in.Grid(),out.Grid()); // drops the cb check
 
@@ -373,6 +395,7 @@ void ImprovedStaggeredFermion5D<Impl>::DhopEO(const FermionField &in, FermionFie
 template<class Impl>
 void ImprovedStaggeredFermion5D<Impl>::Dhop(const FermionField &in, FermionField &out,int dag)
 {
+  DhopCalls+=2;
   conformable(in.Grid(),FermionGrid()); // verifies full grid
   conformable(in.Grid(),out.Grid());
 
@@ -381,6 +404,58 @@ void ImprovedStaggeredFermion5D<Impl>::Dhop(const FermionField &in, FermionField
   DhopInternal(Stencil,Lebesgue,Umu,UUUmu,in,out,dag);
 }
 
+template<class Impl>
+void ImprovedStaggeredFermion5D<Impl>::Report(void) 
+{
+  Coordinate latt = GridDefaultLatt();          
+  RealD volume = Ls;  for(int mu=0;mu<Nd;mu++) volume=volume*latt[mu];
+  RealD NP = _FourDimGrid->_Nprocessors;
+  RealD NN = _FourDimGrid->NodeCount();
+
+  std::cout << GridLogMessage << "#### Dhop calls report " << std::endl;
+
+  std::cout << GridLogMessage << "ImprovedStaggeredFermion5D Number of DhopEO Calls   : " 
+	    << DhopCalls   << std::endl;
+  std::cout << GridLogMessage << "ImprovedStaggeredFermion5D TotalTime   /Calls       : " 
+	    << DhopTotalTime   / DhopCalls << " us" << std::endl;
+  std::cout << GridLogMessage << "ImprovedStaggeredFermion5D CommTime    /Calls       : " 
+	    << DhopCommTime    / DhopCalls << " us" << std::endl;
+  std::cout << GridLogMessage << "ImprovedStaggeredFermion5D ComputeTime/Calls        : " 
+	    << DhopComputeTime / DhopCalls << " us" << std::endl;
+
+  // Average the compute time
+  _FourDimGrid->GlobalSum(DhopComputeTime);
+  DhopComputeTime/=NP;
+
+  RealD mflops = 1154*volume*DhopCalls/DhopComputeTime/2; // 2 for red black counting
+  std::cout << GridLogMessage << "Average mflops/s per call                : " << mflops << std::endl;
+  std::cout << GridLogMessage << "Average mflops/s per call per rank       : " << mflops/NP << std::endl;
+  std::cout << GridLogMessage << "Average mflops/s per call per node       : " << mflops/NN << std::endl;
+  
+  RealD Fullmflops = 1154*volume*DhopCalls/(DhopTotalTime)/2; // 2 for red black counting
+  std::cout << GridLogMessage << "Average mflops/s per call (full)         : " << Fullmflops << std::endl;
+  std::cout << GridLogMessage << "Average mflops/s per call per rank (full): " << Fullmflops/NP << std::endl;
+  std::cout << GridLogMessage << "Average mflops/s per call per node (full): " << Fullmflops/NN << std::endl;
+
+  std::cout << GridLogMessage << "ImprovedStaggeredFermion5D Stencil"    <<std::endl;  Stencil.Report();
+  std::cout << GridLogMessage << "ImprovedStaggeredFermion5D StencilEven"<<std::endl;  StencilEven.Report();
+  std::cout << GridLogMessage << "ImprovedStaggeredFermion5D StencilOdd" <<std::endl;  StencilOdd.Report();
+}
+template<class Impl>
+void ImprovedStaggeredFermion5D<Impl>::ZeroCounters(void) 
+{
+  DhopCalls       = 0;
+  DhopTotalTime    = 0;
+  DhopCommTime    = 0;
+  DhopComputeTime = 0;
+  DhopFaceTime    = 0;
+
+
+  Stencil.ZeroCounters();
+  StencilEven.ZeroCounters();
+  StencilOdd.ZeroCounters();
+}
+
 /////////////////////////////////////////////////////////////////////////
 // Implement the general interface. Here we use SAME mass on all slices
 /////////////////////////////////////////////////////////////////////////

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

@@ -334,6 +334,7 @@ void ImprovedStaggeredFermion<Impl>::DhopDerivEO(GaugeField &mat, const FermionF
 template <class Impl>
 void ImprovedStaggeredFermion<Impl>::Dhop(const FermionField &in, FermionField &out, int dag) 
 {
+  DhopCalls+=2;
   conformable(in.Grid(), _grid);  // verifies full grid
   conformable(in.Grid(), out.Grid());
 
@@ -345,6 +346,7 @@ void ImprovedStaggeredFermion<Impl>::Dhop(const FermionField &in, FermionField &
 template <class Impl>
 void ImprovedStaggeredFermion<Impl>::DhopOE(const FermionField &in, FermionField &out, int dag) 
 {
+  DhopCalls+=1;
   conformable(in.Grid(), _cbgrid);    // verifies half grid
   conformable(in.Grid(), out.Grid());  // drops the cb check
 
@@ -357,6 +359,7 @@ void ImprovedStaggeredFermion<Impl>::DhopOE(const FermionField &in, FermionField
 template <class Impl>
 void ImprovedStaggeredFermion<Impl>::DhopEO(const FermionField &in, FermionField &out, int dag) 
 {
+  DhopCalls+=1;
   conformable(in.Grid(), _cbgrid);    // verifies half grid
   conformable(in.Grid(), out.Grid());  // drops the cb check
 
@@ -415,33 +418,47 @@ void ImprovedStaggeredFermion<Impl>::DhopInternalOverlappedComms(StencilImpl &st
   Compressor compressor; 
   int len =  U.Grid()->oSites();
 
+  DhopTotalTime   -= usecond();
+
+  DhopFaceTime    -= usecond();
   st.Prepare();
   st.HaloGather(in,compressor);
+  DhopFaceTime    += usecond();
 
+  DhopCommTime -=usecond();
   std::vector<std::vector<CommsRequest_t> > requests;
   st.CommunicateBegin(requests);
 
+  DhopFaceTime-=usecond();
   st.CommsMergeSHM(compressor);
+  DhopFaceTime+= usecond();
 
   //////////////////////////////////////////////////////////////////////////////////////////////////////
   // Removed explicit thread comms
   //////////////////////////////////////////////////////////////////////////////////////////////////////
+  DhopComputeTime    -= usecond();
   {
     int interior=1;
     int exterior=0;
     Kernels::DhopImproved(st,lo,U,UUU,in,out,dag,interior,exterior);
   }
+  DhopComputeTime    += usecond();
 
   st.CommunicateComplete(requests);
+  DhopCommTime +=usecond();
 
   // First to enter, last to leave timing
+  DhopFaceTime    -= usecond();
   st.CommsMerge(compressor);
+  DhopFaceTime    -= usecond();
 
+  DhopComputeTime2    -= usecond();
   {
     int interior=0;
     int exterior=1;
     Kernels::DhopImproved(st,lo,U,UUU,in,out,dag,interior,exterior);
   }
+  DhopComputeTime2    += usecond();
 }
 
 
@@ -454,16 +471,78 @@ void ImprovedStaggeredFermion<Impl>::DhopInternalSerialComms(StencilImpl &st, Le
 {
   assert((dag == DaggerNo) || (dag == DaggerYes));
 
+  DhopTotalTime   -= usecond();
+
+  DhopCommTime    -= usecond();
   Compressor compressor;
   st.HaloExchange(in, compressor);
+  DhopCommTime    += usecond();
 
+  DhopComputeTime -= usecond();
   {
     int interior=1;
     int exterior=1;
     Kernels::DhopImproved(st,lo,U,UUU,in,out,dag,interior,exterior);
   }
+  DhopComputeTime += usecond();
+  DhopTotalTime   += usecond();
 };
 
+  ////////////////////////////////////////////////////////////////
+  // Reporting
+  ////////////////////////////////////////////////////////////////
+template<class Impl>
+void ImprovedStaggeredFermion<Impl>::Report(void) 
+{
+  Coordinate latt = _grid->GlobalDimensions();
+  RealD volume = 1;  for(int mu=0;mu<Nd;mu++) volume=volume*latt[mu];
+  RealD NP = _grid->_Nprocessors;
+  RealD NN = _grid->NodeCount();
+
+  std::cout << GridLogMessage << "#### Dhop calls report " << std::endl;
+
+  std::cout << GridLogMessage << "ImprovedStaggeredFermion Number of DhopEO Calls   : " 
+	    << DhopCalls   << std::endl;
+  std::cout << GridLogMessage << "ImprovedStaggeredFermion TotalTime   /Calls       : " 
+	    << DhopTotalTime   / DhopCalls << " us" << std::endl;
+  std::cout << GridLogMessage << "ImprovedStaggeredFermion CommTime    /Calls       : " 
+	    << DhopCommTime    / DhopCalls << " us" << std::endl;
+  std::cout << GridLogMessage << "ImprovedStaggeredFermion ComputeTime/Calls        : " 
+	    << DhopComputeTime / DhopCalls << " us" << std::endl;
+
+  // Average the compute time
+  _grid->GlobalSum(DhopComputeTime);
+  DhopComputeTime/=NP;
+
+  RealD mflops = 1154*volume*DhopCalls/DhopComputeTime/2; // 2 for red black counting
+  std::cout << GridLogMessage << "Average mflops/s per call                : " << mflops << std::endl;
+  std::cout << GridLogMessage << "Average mflops/s per call per rank       : " << mflops/NP << std::endl;
+  std::cout << GridLogMessage << "Average mflops/s per call per node       : " << mflops/NN << std::endl;
+  
+  RealD Fullmflops = 1154*volume*DhopCalls/(DhopTotalTime)/2; // 2 for red black counting
+  std::cout << GridLogMessage << "Average mflops/s per call (full)         : " << Fullmflops << std::endl;
+  std::cout << GridLogMessage << "Average mflops/s per call per rank (full): " << Fullmflops/NP << std::endl;
+  std::cout << GridLogMessage << "Average mflops/s per call per node (full): " << Fullmflops/NN << std::endl;
+
+  std::cout << GridLogMessage << "ImprovedStaggeredFermion Stencil"    <<std::endl;  Stencil.Report();
+  std::cout << GridLogMessage << "ImprovedStaggeredFermion StencilEven"<<std::endl;  StencilEven.Report();
+  std::cout << GridLogMessage << "ImprovedStaggeredFermion StencilOdd" <<std::endl;  StencilOdd.Report();
+}
+template<class Impl>
+void ImprovedStaggeredFermion<Impl>::ZeroCounters(void) 
+{
+  DhopCalls       = 0;
+  DhopTotalTime   = 0;
+  DhopCommTime    = 0;
+  DhopComputeTime = 0;
+  DhopFaceTime    = 0;
+
+  Stencil.ZeroCounters();
+  StencilEven.ZeroCounters();
+  StencilOdd.ZeroCounters();
+}
+
+
 //////////////////////////////////////////////////////// 
 // Conserved current - not yet implemented.
 ////////////////////////////////////////////////////////

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

@@ -55,6 +55,9 @@ void MobiusEOFAFermion<Impl>::M5D(const FermionField &psi_i, const FermionField
   auto plower = &lower[0];
 
   // Flops = 6.0*(Nc*Ns) *Ls*vol
+  this->M5Dcalls++;
+  this->M5Dtime -= usecond();
+
   int nloop = grid->oSites()/Ls;
   accelerator_for(sss,nloop,Simd::Nsimd(),{
     uint64_t ss = sss*Ls;
@@ -70,6 +73,7 @@ void MobiusEOFAFermion<Impl>::M5D(const FermionField &psi_i, const FermionField
     }
   });
 
+  this->M5Dtime += usecond();
 }
 
 template<class Impl>
@@ -95,6 +99,9 @@ void MobiusEOFAFermion<Impl>::M5D_shift(const FermionField &psi_i, const Fermion
   auto pshift_coeffs = &shift_coeffs[0];
 
   // Flops = 6.0*(Nc*Ns) *Ls*vol
+  this->M5Dcalls++;
+  this->M5Dtime -= usecond();
+
   int nloop = grid->oSites()/Ls;
   accelerator_for(sss,nloop,Simd::Nsimd(),{
     uint64_t ss = sss*Ls;
@@ -115,6 +122,7 @@ void MobiusEOFAFermion<Impl>::M5D_shift(const FermionField &psi_i, const Fermion
     }
   });
 
+  this->M5Dtime += usecond();
 }
 
 template<class Impl>
@@ -135,6 +143,9 @@ void MobiusEOFAFermion<Impl>::M5Ddag(const FermionField &psi_i, const FermionFie
   auto plower = &lower[0];
 
   // Flops = 6.0*(Nc*Ns) *Ls*vol
+  this->M5Dcalls++;
+  this->M5Dtime -= usecond();
+
   int nloop = grid->oSites()/Ls;
   accelerator_for(sss,nloop,Simd::Nsimd(), {
     uint64_t ss = sss*Ls;
@@ -150,6 +161,8 @@ void MobiusEOFAFermion<Impl>::M5Ddag(const FermionField &psi_i, const FermionFie
       coalescedWrite(chi[ss+s], pdiag[s]*phi(ss+s) + pupper[s]*tmp1 + plower[s]*tmp2);
     }
   });
+
+  this->M5Dtime += usecond();
 }
 
 template<class Impl>
@@ -173,6 +186,9 @@ void MobiusEOFAFermion<Impl>::M5Ddag_shift(const FermionField &psi_i, const Ferm
   auto pshift_coeffs = &shift_coeffs[0];
 
   // Flops = 6.0*(Nc*Ns) *Ls*vol
+  this->M5Dcalls++;
+  this->M5Dtime -= usecond();
+
   auto pm = this->pm;
 
   int nloop = grid->oSites()/Ls;
@@ -201,6 +217,7 @@ void MobiusEOFAFermion<Impl>::M5Ddag_shift(const FermionField &psi_i, const Ferm
     }
   });
 
+  this->M5Dtime += usecond();
 }
 
 template<class Impl>
@@ -220,6 +237,9 @@ void MobiusEOFAFermion<Impl>::MooeeInv(const FermionField &psi_i, FermionField &
 
   if(this->shift != 0.0){ MooeeInv_shift(psi_i,chi_i); return; }
 
+  this->MooeeInvCalls++;
+  this->MooeeInvTime -= usecond();
+
   int nloop = grid->oSites()/Ls;
   accelerator_for(sss,nloop,Simd::Nsimd(),{
     uint64_t ss=sss*Ls;
@@ -257,6 +277,7 @@ void MobiusEOFAFermion<Impl>::MooeeInv(const FermionField &psi_i, FermionField &
     }
   });
    
+  this->MooeeInvTime += usecond();
 }
 
 template<class Impl>
@@ -276,6 +297,8 @@ void MobiusEOFAFermion<Impl>::MooeeInv_shift(const FermionField &psi_i, FermionF
   auto pueem= & this->ueem[0];
   auto pMooeeInv_shift_lc   = &MooeeInv_shift_lc[0];
   auto pMooeeInv_shift_norm = &MooeeInv_shift_norm[0];
+  this->MooeeInvCalls++;
+  this->MooeeInvTime -= usecond();
 
   int nloop = grid->oSites()/Ls;
   accelerator_for(sss,nloop,Simd::Nsimd(),{
@@ -320,6 +343,7 @@ void MobiusEOFAFermion<Impl>::MooeeInv_shift(const FermionField &psi_i, FermionF
       }
   });
 
+  this->MooeeInvTime += usecond();
 }
 
 template<class Impl>
@@ -339,6 +363,9 @@ void MobiusEOFAFermion<Impl>::MooeeInvDag(const FermionField &psi_i, FermionFiel
   auto pleem= & this->leem[0];
   auto pueem= & this->ueem[0];
 
+  this->MooeeInvCalls++;
+  this->MooeeInvTime -= usecond();
+
   int nloop = grid->oSites()/Ls;
   accelerator_for(sss,nloop,Simd::Nsimd(),{
     uint64_t ss=sss*Ls;
@@ -375,6 +402,7 @@ void MobiusEOFAFermion<Impl>::MooeeInvDag(const FermionField &psi_i, FermionFiel
       coalescedWrite(chi[ss+s],res);
     }
   });
+  this->MooeeInvTime += usecond();
 }
 
 template<class Impl>
@@ -395,6 +423,9 @@ void MobiusEOFAFermion<Impl>::MooeeInvDag_shift(const FermionField &psi_i, Fermi
   auto pMooeeInvDag_shift_lc   = &MooeeInvDag_shift_lc[0];
   auto pMooeeInvDag_shift_norm = &MooeeInvDag_shift_norm[0];
 
+  this->MooeeInvCalls++;
+  this->MooeeInvTime -= usecond();
+
   int nloop = grid->oSites()/Ls;
   accelerator_for(sss,nloop,Simd::Nsimd(),{
       uint64_t ss=sss*Ls;
@@ -438,6 +469,7 @@ void MobiusEOFAFermion<Impl>::MooeeInvDag_shift(const FermionField &psi_i, Fermi
       }
   });
 
+  this->MooeeInvTime += usecond();
 }
 
 NAMESPACE_END(Grid);