more HOST_NAME_MAX fix

Can be made to work, but could put in some more serious infrastructure
for repro testing and blame attribution (Britney test) if necessary

Grid/Grid_Eigen_Dense.h

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

Grid/algorithms/approx/Zolotarev.cc

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

Grid/algorithms/approx/Zolotarev.h

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

Grid/algorithms/blas/BatchedBlas.cc

@@ -0,0 +1,34 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: BatchedBlas.h
+
+    Copyright (C) 2023
+
+Author: Peter Boyle <pboyle@bnl.gov>
+
+    This program is free software; you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation; either version 2 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License along
+    with this program; if not, write to the Free Software Foundation, Inc.,
+    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+    See the full license in the file "LICENSE" in the top level distribution directory
+*************************************************************************************/
+/*  END LEGAL */
+#include <Grid/GridCore.h>
+#include <Grid/algorithms/blas/BatchedBlas.h>
+NAMESPACE_BEGIN(Grid);
+gridblasHandle_t GridBLAS::gridblasHandle;
+int              GridBLAS::gridblasInit;
+NAMESPACE_END(Grid);
+

Grid/algorithms/blas/BatchedBlas.h

@@ -0,0 +1,727 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: BatchedBlas.h
+
+    Copyright (C) 2023
+
+Author: Peter Boyle <pboyle@bnl.gov>
+
+    This program is free software; you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation; either version 2 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License along
+    with this program; if not, write to the Free Software Foundation, Inc.,
+    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+    See the full license in the file "LICENSE" in the top level distribution directory
+*************************************************************************************/
+/*  END LEGAL */
+#pragma once
+
+#ifdef GRID_HIP
+#include <hipblas/hipblas.h>
+#endif
+#ifdef GRID_CUDA
+#include <cublas_v2.h>
+#endif
+#ifdef GRID_SYCL
+#include <oneapi/mkl.hpp>
+#endif
+#if 0
+#define GRID_ONE_MKL
+#endif
+#ifdef GRID_ONE_MKL
+#include <oneapi/mkl.hpp>
+#endif
+///////////////////////////////////////////////////////////////////////	  
+// Need to rearrange lattice data to be in the right format for a
+// batched multiply. Might as well make these static, dense packed
+///////////////////////////////////////////////////////////////////////
+NAMESPACE_BEGIN(Grid);
+#ifdef GRID_HIP
+  typedef hipblasHandle_t gridblasHandle_t;
+#endif
+#ifdef GRID_CUDA
+  typedef cublasHandle_t gridblasHandle_t;
+#endif
+#ifdef GRID_SYCL
+  typedef cl::sycl::queue *gridblasHandle_t;
+#endif
+#ifdef GRID_ONE_MKL
+  typedef cl::sycl::queue *gridblasHandle_t;
+#endif
+#if !defined(GRID_SYCL) && !defined(GRID_CUDA) && !defined(GRID_HIP) && !defined(GRID_ONE_MKL)
+  typedef int32_t gridblasHandle_t;
+#endif
+
+enum GridBLASOperation_t { GridBLAS_OP_N, GridBLAS_OP_T, GridBLAS_OP_C } ;
+
+class GridBLAS {
+public:
+
+  
+  static gridblasHandle_t gridblasHandle;
+  static int            gridblasInit;
+  
+  static void Init(void)
+  {
+    if ( ! gridblasInit ) {
+#ifdef GRID_CUDA
+      std::cout << "cublasCreate"<<std::endl;
+      cublasCreate(&gridblasHandle);
+      cublasSetPointerMode(gridblasHandle, CUBLAS_POINTER_MODE_DEVICE);
+#endif
+#ifdef GRID_HIP
+      std::cout << "hipblasCreate"<<std::endl;
+      hipblasCreate(&gridblasHandle);
+#endif
+#ifdef GRID_SYCL
+      gridblasHandle = theGridAccelerator;
+#endif
+#ifdef GRID_ONE_MKL
+      cl::sycl::cpu_selector selector;
+      cl::sycl::device selectedDevice { selector };
+      gridblasHandle =new sycl::queue (selectedDevice);
+#endif
+      gridblasInit=1;
+    }
+  }
+  
+  // Force construct once
+  GridBLAS() { Init(); };
+  ~GridBLAS() { };
+  
+  /////////////////////////////////////////////////////////////////////////////////////
+  // BLAS GEMM conventions:
+  /////////////////////////////////////////////////////////////////////////////////////
+  // - C = alpha A * B + beta C
+  // Dimensions:
+  // - C_m.n
+  // - A_m.k
+  // - B_k.n
+  // - Flops = 8 M N K
+  // - Bytes = 2*sizeof(word) * (MN+MK+KN)
+  // M=60, N=12
+  // Flop/Byte = 8 . 60.60.12 / (60.12+60.60+60.12)/16 = 4 so expect about 4 TF/s on a GCD
+  /////////////////////////////////////////////////////////////////////////////////////
+  void synchronise(void)
+  {
+#ifdef GRID_HIP
+    auto err = hipDeviceSynchronize();
+    assert(err==hipSuccess);
+#endif
+#ifdef GRID_CUDA
+    auto err = cudaDeviceSynchronize();
+    assert(err==cudaSuccess);
+#endif
+#ifdef GRID_SYCL
+    accelerator_barrier();
+#endif
+#ifdef GRID_ONE_MKL
+    gridblasHandle->wait();
+#endif
+  }
+  
+  void gemmBatched(int m,int n, int k,
+		   ComplexD alpha,
+		   deviceVector<ComplexD*> &Amk,  // pointer list to matrices
+		   deviceVector<ComplexD*> &Bkn,
+		   ComplexD beta,
+		   deviceVector<ComplexD*> &Cmn)
+  {
+    gemmBatched(GridBLAS_OP_N,GridBLAS_OP_N,
+		m,n,k,
+		alpha,
+		Amk,
+		Bkn,
+		beta,
+		Cmn);
+  }
+  void gemmBatched(int m,int n, int k,
+		   ComplexF alpha,
+		   deviceVector<ComplexF*> &Amk,  // pointer list to matrices
+		   deviceVector<ComplexF*> &Bkn,
+		   ComplexF beta,
+		   deviceVector<ComplexF*> &Cmn)
+  {
+    gemmBatched(GridBLAS_OP_N,GridBLAS_OP_N,
+		m,n,k,
+		alpha,
+		Amk,
+		Bkn,
+		beta,
+		Cmn);
+  }
+  void gemmBatched(int m,int n, int k,
+		   RealD alpha,
+		   deviceVector<RealD*> &Amk,  // pointer list to matrices
+		   deviceVector<RealD*> &Bkn,
+		   RealD beta,
+		   deviceVector<RealD*> &Cmn)
+  {
+    gemmBatched(GridBLAS_OP_N,GridBLAS_OP_N,
+		m,n,k,
+		alpha,
+		Amk,
+		Bkn,
+		beta,
+		Cmn);
+  }
+  void gemmBatched(int m,int n, int k,
+		   RealF alpha,
+		   deviceVector<RealF*> &Amk,  // pointer list to matrices
+		   deviceVector<RealF*> &Bkn,
+		   RealF beta,
+		   deviceVector<RealF*> &Cmn)
+  {
+    gemmBatched(GridBLAS_OP_N,GridBLAS_OP_N,
+		m,n,k,
+		alpha,
+		Amk,
+		Bkn,
+		beta,
+		Cmn);
+  }
+
+  void gemmBatched(GridBLASOperation_t OpA,
+		   GridBLASOperation_t OpB,
+		   int m,int n, int k,
+		   ComplexD alpha,
+		   deviceVector<ComplexD*> &Amk,  // pointer list to matrices
+		   deviceVector<ComplexD*> &Bkn,
+		   ComplexD beta,
+		   deviceVector<ComplexD*> &Cmn)
+  {
+    RealD t2=usecond();
+    int32_t batchCount = Amk.size();
+    assert(Bkn.size()==batchCount);
+    assert(Cmn.size()==batchCount);
+
+    int lda = m; // m x k column major
+    int ldb = k; // k x n column major
+    int ldc = m; // m x b column major
+    if(OpA!=GridBLAS_OP_N)
+      lda = k;
+    if(OpB!=GridBLAS_OP_N)
+      ldb = n;
+    
+    static deviceVector<ComplexD> alpha_p(1);
+    static deviceVector<ComplexD> beta_p(1);
+    // can prestore the 1 and the zero on device
+    acceleratorCopyToDevice((void *)&alpha,(void *)&alpha_p[0],sizeof(ComplexD));
+    acceleratorCopyToDevice((void *)&beta ,(void *)&beta_p[0],sizeof(ComplexD));
+    RealD t0=usecond();
+    //    std::cout << "ZgemmBatched mnk  "<<m<<","<<n<<","<<k<<" count "<<batchCount<<std::endl;
+#ifdef GRID_HIP
+    hipblasOperation_t hOpA;
+    hipblasOperation_t hOpB;
+    if ( OpA == GridBLAS_OP_N ) hOpA = HIPBLAS_OP_N;
+    if ( OpA == GridBLAS_OP_T ) hOpA = HIPBLAS_OP_T;
+    if ( OpA == GridBLAS_OP_C ) hOpA = HIPBLAS_OP_C;
+    if ( OpB == GridBLAS_OP_N ) hOpB = HIPBLAS_OP_N;
+    if ( OpB == GridBLAS_OP_T ) hOpB = HIPBLAS_OP_T;
+    if ( OpB == GridBLAS_OP_C ) hOpB = HIPBLAS_OP_C;
+    auto err = hipblasZgemmBatched(gridblasHandle,
+				   hOpA,
+				   hOpB,
+				   m,n,k,
+				   (hipblasDoubleComplex *) &alpha_p[0],
+				   (hipblasDoubleComplex **)&Amk[0], lda,
+				   (hipblasDoubleComplex **)&Bkn[0], ldb,
+				   (hipblasDoubleComplex *) &beta_p[0],
+				   (hipblasDoubleComplex **)&Cmn[0], ldc,
+				   batchCount);
+    //	 std::cout << " hipblas return code " <<(int)err<<std::endl;
+    assert(err==HIPBLAS_STATUS_SUCCESS);
+#endif
+#ifdef GRID_CUDA
+    cublasOperation_t hOpA;
+    cublasOperation_t hOpB;
+    if ( OpA == GridBLAS_OP_N ) hOpA = CUBLAS_OP_N;
+    if ( OpA == GridBLAS_OP_T ) hOpA = CUBLAS_OP_T;
+    if ( OpA == GridBLAS_OP_C ) hOpA = CUBLAS_OP_C;
+    if ( OpB == GridBLAS_OP_N ) hOpB = CUBLAS_OP_N;
+    if ( OpB == GridBLAS_OP_T ) hOpB = CUBLAS_OP_T;
+    if ( OpB == GridBLAS_OP_C ) hOpB = CUBLAS_OP_C;
+    auto err = cublasZgemmBatched(gridblasHandle,
+				  hOpA,
+				  hOpB,
+				  m,n,k,
+				  (cuDoubleComplex *) &alpha_p[0],
+				  (cuDoubleComplex **)&Amk[0], lda,
+				  (cuDoubleComplex **)&Bkn[0], ldb,
+				  (cuDoubleComplex *) &beta_p[0],
+				  (cuDoubleComplex **)&Cmn[0], ldc,
+				  batchCount);
+    assert(err==CUBLAS_STATUS_SUCCESS);
+#endif
+#ifdef GRID_SYCL
+    //MKL’s cblas_<T>gemm_batch & OneAPI
+#warning "oneMKL implementation not built "
+#endif
+#if !defined(GRID_SYCL) && !defined(GRID_CUDA) && !defined(GRID_HIP)
+    // Need a default/reference implementation
+    int sda = lda*k;
+    int sdb = ldb*k;
+    int sdc = ldc*n;
+    for (int p = 0; p < batchCount; ++p) {
+      for (int mm = 0; mm < m; ++mm) {
+	for (int nn = 0; nn < n; ++nn) {
+	  ComplexD c_mn(0.0);
+	  for (int kk = 0; kk < k; ++kk)
+	    c_mn += Amk[p][mm + kk*lda ] * Bkn[p][kk + nn*ldb];
+	  Cmn[p][mm + nn*ldc] =  (alpha)*c_mn + (beta)*Cmn[p][mm + nn*ldc ];
+	}
+      }
+    }
+#endif
+    //    synchronise();
+     RealD t1=usecond();
+     RealD flops = 8.0*m*n*k*batchCount;
+     RealD bytes = 1.0*sizeof(ComplexD)*(m*k+k*n+m*n)*batchCount;
+     //     std::cout <<GridLogMessage<< " batched Blas copy "<<(t0-t2)/1.e3 <<" ms "<<std::endl;
+     //     std::cout <<GridLogMessage<< " batched Blas zGemm call "<<m<<","<<n<<","<<k<<" "<< flops/(t1-t0)/1.e3 <<" GF/s "<<(t1-t0)/1.e3<<" ms "<<std::endl;
+     //     std::cout <<GridLogMessage<< " batched Blas zGemm call "<<m<<","<<n<<","<<k<<" "<< bytes/(t1-t0)/1.e3 <<" GB/s "<<(t1-t0)/1.e3<<" ms "<<std::endl;
+  }
+
+  void gemmBatched(GridBLASOperation_t OpA,
+		   GridBLASOperation_t OpB,
+		   int m,int n, int k,
+		   ComplexF alpha,
+		   deviceVector<ComplexF*> &Amk,  // pointer list to matrices
+		   deviceVector<ComplexF*> &Bkn,
+		   ComplexF beta,
+		   deviceVector<ComplexF*> &Cmn)
+  {
+    RealD t2=usecond();
+    int32_t batchCount = Amk.size();
+
+    int lda = m; // m x k column major
+    int ldb = k; // k x n column major
+    int ldc = m; // m x b column major
+    if(OpA!=GridBLAS_OP_N)
+      lda = k;
+    if(OpB!=GridBLAS_OP_N)
+      ldb = n;
+    static deviceVector<ComplexF> alpha_p(1);
+    static deviceVector<ComplexF> beta_p(1);
+    // can prestore the 1 and the zero on device
+    acceleratorCopyToDevice((void *)&alpha,(void *)&alpha_p[0],sizeof(ComplexF));
+    acceleratorCopyToDevice((void *)&beta ,(void *)&beta_p[0],sizeof(ComplexF));
+    RealD t0=usecond();
+
+    assert(Bkn.size()==batchCount);
+    assert(Cmn.size()==batchCount);
+#ifdef GRID_HIP
+    hipblasOperation_t hOpA;
+    hipblasOperation_t hOpB;
+    if ( OpA == GridBLAS_OP_N ) hOpA = HIPBLAS_OP_N;
+    if ( OpA == GridBLAS_OP_T ) hOpA = HIPBLAS_OP_T;
+    if ( OpA == GridBLAS_OP_C ) hOpA = HIPBLAS_OP_C;
+    if ( OpB == GridBLAS_OP_N ) hOpB = HIPBLAS_OP_N;
+    if ( OpB == GridBLAS_OP_T ) hOpB = HIPBLAS_OP_T;
+    if ( OpB == GridBLAS_OP_C ) hOpB = HIPBLAS_OP_C;
+    auto err = hipblasCgemmBatched(gridblasHandle,
+				   hOpA,
+				   hOpB,
+				   m,n,k,
+				   (hipblasComplex *) &alpha_p[0],
+				   (hipblasComplex **)&Amk[0], lda,
+				   (hipblasComplex **)&Bkn[0], ldb,
+				   (hipblasComplex *) &beta_p[0],
+				   (hipblasComplex **)&Cmn[0], ldc,
+				   batchCount);
+
+    assert(err==HIPBLAS_STATUS_SUCCESS);
+#endif
+#ifdef GRID_CUDA
+    cublasOperation_t hOpA;
+    cublasOperation_t hOpB;
+    if ( OpA == GridBLAS_OP_N ) hOpA = CUBLAS_OP_N;
+    if ( OpA == GridBLAS_OP_T ) hOpA = CUBLAS_OP_T;
+    if ( OpA == GridBLAS_OP_C ) hOpA = CUBLAS_OP_C;
+    if ( OpB == GridBLAS_OP_N ) hOpB = CUBLAS_OP_N;
+    if ( OpB == GridBLAS_OP_T ) hOpB = CUBLAS_OP_T;
+    if ( OpB == GridBLAS_OP_C ) hOpB = CUBLAS_OP_C;
+    auto err = cublasCgemmBatched(gridblasHandle,
+				  hOpA,
+				  hOpB,
+				  m,n,k,
+				  (cuComplex *) &alpha_p[0],
+				  (cuComplex **)&Amk[0], lda,
+				  (cuComplex **)&Bkn[0], ldb,
+				  (cuComplex *) &beta_p[0],
+				  (cuComplex **)&Cmn[0], ldc,
+				  batchCount);
+    assert(err==CUBLAS_STATUS_SUCCESS);
+#endif
+#ifdef GRID_SYCL
+    //MKL’s cblas_<T>gemm_batch & OneAPI
+#warning "oneMKL implementation not built "
+#endif
+#if !defined(GRID_SYCL) && !defined(GRID_CUDA) && !defined(GRID_HIP)
+    int sda = lda*k;
+    int sdb = ldb*k;
+    int sdc = ldc*n;
+    ComplexF alphaf(real(alpha),imag(alpha));
+    ComplexF betaf(real(beta),imag(beta));
+    // Need a default/reference implementation
+    for (int p = 0; p < batchCount; ++p) {
+      for (int mm = 0; mm < m; ++mm) {
+	for (int nn = 0; nn < n; ++nn) {
+	  ComplexF c_mn(0.0);
+	  for (int kk = 0; kk < k; ++kk)
+	    c_mn += Amk[p][mm + kk*lda ] * Bkn[p][kk + nn*ldb];
+	  Cmn[p][mm + nn*ldc] =  (alphaf)*c_mn + (betaf)*Cmn[p][mm + nn*ldc ];
+	}
+      }
+    }
+#endif
+     RealD t1=usecond();
+     RealD flops = 8.0*m*n*k*batchCount;
+     RealD bytes = 1.0*sizeof(ComplexF)*(m*k+k*n+m*n)*batchCount;
+  }
+  
+  ///////////////////////////////////////////////////////////////////////////
+  // Single precision real GEMM
+  ///////////////////////////////////////////////////////////////////////////
+
+  void gemmBatched(GridBLASOperation_t OpA,
+		   GridBLASOperation_t OpB,
+		   int m,int n, int k,
+		   RealF alpha,
+		   deviceVector<RealF*> &Amk,  // pointer list to matrices
+		   deviceVector<RealF*> &Bkn,
+		   RealF beta,
+		   deviceVector<RealF*> &Cmn)
+  {
+    RealD t2=usecond();
+    int32_t batchCount = Amk.size();
+
+    int lda = m; // m x k column major
+    int ldb = k; // k x n column major
+    int ldc = m; // m x b column major
+    if(OpA!=GridBLAS_OP_N)
+      lda = k;
+    if(OpB!=GridBLAS_OP_N)
+      ldb = n;
+    static deviceVector<RealF> alpha_p(1);
+    static deviceVector<RealF> beta_p(1);
+    // can prestore the 1 and the zero on device
+    acceleratorCopyToDevice((void *)&alpha,(void *)&alpha_p[0],sizeof(RealF));
+    acceleratorCopyToDevice((void *)&beta ,(void *)&beta_p[0],sizeof(RealF));
+    RealD t0=usecond();
+
+    assert(Bkn.size()==batchCount);
+    assert(Cmn.size()==batchCount);
+#ifdef GRID_HIP
+    hipblasOperation_t hOpA;
+    hipblasOperation_t hOpB;
+    if ( OpA == GridBLAS_OP_N ) hOpA = HIPBLAS_OP_N;
+    if ( OpA == GridBLAS_OP_T ) hOpA = HIPBLAS_OP_T;
+    if ( OpA == GridBLAS_OP_C ) hOpA = HIPBLAS_OP_C;
+    if ( OpB == GridBLAS_OP_N ) hOpB = HIPBLAS_OP_N;
+    if ( OpB == GridBLAS_OP_T ) hOpB = HIPBLAS_OP_T;
+    if ( OpB == GridBLAS_OP_C ) hOpB = HIPBLAS_OP_C;
+    auto err = hipblasSgemmBatched(gridblasHandle,
+				   hOpA,
+				   hOpB,
+				   m,n,k,
+				   (float *) &alpha_p[0],
+				   (float **)&Amk[0], lda,
+				   (float **)&Bkn[0], ldb,
+				   (float *) &beta_p[0],
+				   (float **)&Cmn[0], ldc,
+				   batchCount);
+    assert(err==HIPBLAS_STATUS_SUCCESS);
+#endif
+#ifdef GRID_CUDA
+    cublasOperation_t hOpA;
+    cublasOperation_t hOpB;
+    if ( OpA == GridBLAS_OP_N ) hOpA = CUBLAS_OP_N;
+    if ( OpA == GridBLAS_OP_T ) hOpA = CUBLAS_OP_T;
+    if ( OpA == GridBLAS_OP_C ) hOpA = CUBLAS_OP_C;
+    if ( OpB == GridBLAS_OP_N ) hOpB = CUBLAS_OP_N;
+    if ( OpB == GridBLAS_OP_T ) hOpB = CUBLAS_OP_T;
+    if ( OpB == GridBLAS_OP_C ) hOpB = CUBLAS_OP_C;
+    auto err = cublasSgemmBatched(gridblasHandle,
+				  hOpA,
+				  hOpB,
+				  m,n,k,
+				  (float *) &alpha_p[0],
+				  (float **)&Amk[0], lda,
+				  (float **)&Bkn[0], ldb,
+				  (float *) &beta_p[0],
+				  (float **)&Cmn[0], ldc,
+				  batchCount);
+    assert(err==CUBLAS_STATUS_SUCCESS);
+#endif
+#ifdef GRID_SYCL
+    //MKL’s cblas_<T>gemm_batch & OneAPI
+#warning "oneMKL implementation not built "
+#endif
+#if !defined(GRID_SYCL) && !defined(GRID_CUDA) && !defined(GRID_HIP)
+    int sda = lda*k;
+    int sdb = ldb*k;
+    int sdc = ldc*n;
+    // Need a default/reference implementation
+    for (int p = 0; p < batchCount; ++p) {
+      for (int mm = 0; mm < m; ++mm) {
+	for (int nn = 0; nn < n; ++nn) {
+	  RealD c_mn(0.0);
+	  for (int kk = 0; kk < k; ++kk)
+	    c_mn += Amk[p][mm + kk*lda ] * Bkn[p][kk + nn*ldb];
+	  Cmn[p][mm + nn*ldc] =  (alpha)*c_mn + (beta)*Cmn[p][mm + nn*ldc ];
+	}
+      }
+    }
+#endif
+     RealD t1=usecond();
+     RealD flops = 2.0*m*n*k*batchCount;
+     RealD bytes = 1.0*sizeof(RealF)*(m*k+k*n+m*n)*batchCount;
+  }
+  
+  
+  ///////////////////////////////////////////////////////////////////////////
+  // Double precision real GEMM
+  ///////////////////////////////////////////////////////////////////////////
+
+  void gemmBatched(GridBLASOperation_t OpA,
+		   GridBLASOperation_t OpB,
+		   int m,int n, int k,
+		   RealD alpha,
+		   deviceVector<RealD*> &Amk,  // pointer list to matrices
+		   deviceVector<RealD*> &Bkn,
+		   RealD beta,
+		   deviceVector<RealD*> &Cmn)
+  {
+    RealD t2=usecond();
+    int32_t batchCount = Amk.size();
+
+    int lda = m; // m x k column major
+    int ldb = k; // k x n column major
+    int ldc = m; // m x b column major
+    if(OpA!=GridBLAS_OP_N)
+      lda = k;
+    if(OpB!=GridBLAS_OP_N)
+      ldb = n;
+    
+    static deviceVector<RealD> alpha_p(1);
+    static deviceVector<RealD> beta_p(1);
+    // can prestore the 1 and the zero on device
+    acceleratorCopyToDevice((void *)&alpha,(void *)&alpha_p[0],sizeof(RealD));
+    acceleratorCopyToDevice((void *)&beta ,(void *)&beta_p[0],sizeof(RealD));
+    RealD t0=usecond();
+
+    assert(Bkn.size()==batchCount);
+    assert(Cmn.size()==batchCount);
+#ifdef GRID_HIP
+    hipblasOperation_t hOpA;
+    hipblasOperation_t hOpB;
+    if ( OpA == GridBLAS_OP_N ) hOpA = HIPBLAS_OP_N;
+    if ( OpA == GridBLAS_OP_T ) hOpA = HIPBLAS_OP_T;
+    if ( OpA == GridBLAS_OP_C ) hOpA = HIPBLAS_OP_C;
+    if ( OpB == GridBLAS_OP_N ) hOpB = HIPBLAS_OP_N;
+    if ( OpB == GridBLAS_OP_T ) hOpB = HIPBLAS_OP_T;
+    if ( OpB == GridBLAS_OP_C ) hOpB = HIPBLAS_OP_C;
+    auto err = hipblasDgemmBatched(gridblasHandle,
+				   HIPBLAS_OP_N,
+				   HIPBLAS_OP_N,
+				   m,n,k,
+				   (double *) &alpha_p[0],
+				   (double **)&Amk[0], lda,
+				   (double **)&Bkn[0], ldb,
+				   (double *) &beta_p[0],
+				   (double **)&Cmn[0], ldc,
+				   batchCount);
+    assert(err==HIPBLAS_STATUS_SUCCESS);
+#endif
+#ifdef GRID_CUDA
+    cublasOperation_t hOpA;
+    cublasOperation_t hOpB;
+    if ( OpA == GridBLAS_OP_N ) hOpA = CUBLAS_OP_N;
+    if ( OpA == GridBLAS_OP_T ) hOpA = CUBLAS_OP_T;
+    if ( OpA == GridBLAS_OP_C ) hOpA = CUBLAS_OP_C;
+    if ( OpB == GridBLAS_OP_N ) hOpB = CUBLAS_OP_N;
+    if ( OpB == GridBLAS_OP_T ) hOpB = CUBLAS_OP_T;
+    if ( OpB == GridBLAS_OP_C ) hOpB = CUBLAS_OP_C;
+    auto err = cublasDgemmBatched(gridblasHandle,
+				  hOpA,
+				  hOpB,
+				  m,n,k,
+				  (double *) &alpha_p[0],
+				  (double **)&Amk[0], lda,
+				  (double **)&Bkn[0], ldb,
+				  (double *) &beta_p[0],
+				  (double **)&Cmn[0], ldc,
+				  batchCount);
+    assert(err==CUBLAS_STATUS_SUCCESS);
+#endif
+#ifdef GRID_SYCL
+    /*
+      int64_t m64=m;
+      int64_t n64=n;
+      int64_t k64=k;
+      int64_t batchCount64=batchCount;
+      oneapi::mkl::blas::column_major::gemm_batch(*theGridAccelerator,
+      onemkl::transpose::N,
+      onemkl::transpose::N,
+      &m64,&n64,&k64,
+      (double *) &alpha_p[0],
+      (double **)&Amk[0], lda,
+      (double **)&Bkn[0], ldb,
+      (double *) &beta_p[0],
+      (double **)&Cmn[0], ldc,
+      1,&batchCount64);
+     */
+    //MKL’s cblas_<T>gemm_batch & OneAPI
+#warning "oneMKL implementation not built "
+#endif
+#if !defined(GRID_SYCL) && !defined(GRID_CUDA) && !defined(GRID_HIP)
+    int sda = lda*k;
+    int sdb = ldb*k;
+    int sdc = ldc*n;
+    // Need a default/reference implementation
+    for (int p = 0; p < batchCount; ++p) {
+      for (int mm = 0; mm < m; ++mm) {
+	for (int nn = 0; nn < n; ++nn) {
+	  RealD c_mn(0.0);
+	  for (int kk = 0; kk < k; ++kk)
+	    c_mn += Amk[p][mm + kk*lda ] * Bkn[p][kk + nn*ldb];
+	  Cmn[p][mm + nn*ldc] =  (alpha)*c_mn + (beta)*Cmn[p][mm + nn*ldc ];
+	}
+      }
+    }
+#endif
+     RealD t1=usecond();
+     RealD flops = 2.0*m*n*k*batchCount;
+     RealD bytes = 1.0*sizeof(RealD)*(m*k+k*n+m*n)*batchCount;
+  }
+  
+
+  
+  ////////////////////////////////////////////////////////////////////////////////////////////////
+  // Strided case used by benchmark, but generally unused in Grid
+  // Keep a code example in double complex, but don't generate the single and real variants for now
+  ////////////////////////////////////////////////////////////////////////////////////////////////
+  
+  void gemmStridedBatched(int m,int n, int k,
+			  ComplexD alpha,
+			  ComplexD* Amk,  // pointer list to matrices
+			  ComplexD* Bkn,
+			  ComplexD beta,
+			  ComplexD* Cmn,
+			  int batchCount)
+  {
+    // Use C-row major storage, so transpose calls
+    int lda = m; // m x k column major
+    int ldb = k; // k x n column major
+    int ldc = m; // m x b column major
+    int sda = m*k;
+    int sdb = k*n;
+    int sdc = m*n;
+    deviceVector<ComplexD> alpha_p(1);
+    deviceVector<ComplexD> beta_p(1);
+    acceleratorCopyToDevice((void *)&alpha,(void *)&alpha_p[0],sizeof(ComplexD));
+    acceleratorCopyToDevice((void *)&beta ,(void *)&beta_p[0],sizeof(ComplexD));
+
+    //    std::cout << "blasZgemmStridedBatched mnk  "<<m<<","<<n<<","<<k<<" count "<<batchCount<<std::endl;
+    //    std::cout << "blasZgemmStridedBatched ld   "<<lda<<","<<ldb<<","<<ldc<<std::endl;
+    //    std::cout << "blasZgemmStridedBatched sd   "<<sda<<","<<sdb<<","<<sdc<<std::endl;
+#ifdef GRID_HIP
+    auto err = hipblasZgemmStridedBatched(gridblasHandle,
+					  HIPBLAS_OP_N,
+					  HIPBLAS_OP_N,
+					  m,n,k,
+					  (hipblasDoubleComplex *) &alpha_p[0],
+					  (hipblasDoubleComplex *) Amk, lda, sda,
+					  (hipblasDoubleComplex *) Bkn, ldb, sdb,
+					  (hipblasDoubleComplex *) &beta_p[0],
+					  (hipblasDoubleComplex *) Cmn, ldc, sdc,
+					  batchCount);
+    assert(err==HIPBLAS_STATUS_SUCCESS);
+#endif
+#ifdef GRID_CUDA
+    cublasZgemmStridedBatched(gridblasHandle,
+			      CUBLAS_OP_N,
+			      CUBLAS_OP_N,
+			      m,n,k,
+			      (cuDoubleComplex *) &alpha_p[0],
+			      (cuDoubleComplex *) Amk, lda, sda,
+			      (cuDoubleComplex *) Bkn, ldb, sdb,
+			      (cuDoubleComplex *) &beta_p[0],
+			      (cuDoubleComplex *) Cmn, ldc, sdc,
+			      batchCount);
+#endif
+#if defined(GRID_SYCL) || defined(GRID_ONE_MKL)
+    oneapi::mkl::blas::column_major::gemm_batch(*gridblasHandle,
+						oneapi::mkl::transpose::N,
+						oneapi::mkl::transpose::N,
+						m,n,k,
+						alpha,
+						(const ComplexD *)Amk,lda,sda,
+						(const ComplexD *)Bkn,ldb,sdb,
+						beta,
+						(ComplexD *)Cmn,ldc,sdc,
+						batchCount);
+#endif
+#if !defined(GRID_SYCL) && !defined(GRID_CUDA) && !defined(GRID_HIP) && !defined(GRID_ONE_MKL)
+     // Need a default/reference implementation
+     for (int p = 0; p < batchCount; ++p) {
+       for (int mm = 0; mm < m; ++mm) {
+	 for (int nn = 0; nn < n; ++nn) {
+	   ComplexD c_mn(0.0);
+	   for (int kk = 0; kk < k; ++kk)
+	     c_mn += Amk[mm + kk*lda + p*sda] * Bkn[kk + nn*ldb + p*sdb];
+	   Cmn[mm + nn*ldc + p*sdc] =  (alpha)*c_mn + (beta)*Cmn[mm + nn*ldc + p*sdc];
+	 }
+       }
+     }
+#endif
+  }
+
+  double benchmark(int M, int N, int K, int BATCH)
+  {
+    int32_t N_A = M*K*BATCH;
+    int32_t N_B = K*N*BATCH;
+    int32_t N_C = M*N*BATCH;
+    deviceVector<ComplexD> A(N_A); acceleratorMemSet(&A[0],0,N_A*sizeof(ComplexD));
+    deviceVector<ComplexD> B(N_B); acceleratorMemSet(&B[0],0,N_B*sizeof(ComplexD));
+    deviceVector<ComplexD> C(N_C); acceleratorMemSet(&C[0],0,N_C*sizeof(ComplexD));
+    ComplexD alpha(1.0);
+    ComplexD beta (1.0);
+    RealD flops = 8.0*M*N*K*BATCH;
+    int ncall=10;
+    RealD t0 = usecond();
+    for(int i=0;i<ncall;i++){
+      gemmStridedBatched(M,N,K,
+			 alpha,
+			 &A[0], // m x k 
+			 &B[0], // k x n
+			 beta, 
+			 &C[0], // m x n
+			 BATCH);
+    }
+    synchronise();
+    RealD t1 = usecond();
+    RealD bytes = 1.0*sizeof(ComplexD)*(M*N*2+N*K+M*K)*BATCH;
+    flops = 8.0*M*N*K*BATCH*ncall;
+    flops = flops/(t1-t0)/1.e3;
+    return flops; // Returns gigaflops
+  }
+
+
+
+
+};
+
+NAMESPACE_END(Grid);

Grid/allocator/AlignedAllocator.h

@@ -176,6 +176,7 @@ 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/lattice/Lattice.h

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

Grid/lattice/Lattice_basis.h

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

Grid/lattice/Lattice_crc.h

@@ -42,13 +42,13 @@ template<class vobj> void DumpSliceNorm(std::string s,Lattice<vobj> &f,int mu=-1
   }
 }
 
-template<class vobj> uint32_t crc(Lattice<vobj> & buf)
+template<class vobj> uint32_t crc(const Lattice<vobj> & buf)
 {
   autoView( buf_v , buf, CpuRead);
   return ::crc32(0L,(unsigned char *)&buf_v[0],(size_t)sizeof(vobj)*buf.oSites());
 }
 
-#define CRC(U) std::cout << "FingerPrint "<<__FILE__ <<" "<< __LINE__ <<" "<< #U <<" "<<crc(U)<<std::endl;
+#define CRC(U) std::cerr << "FingerPrint "<<__FILE__ <<" "<< __LINE__ <<" "<< #U <<" "<<crc(U)<<std::endl;
 
 NAMESPACE_END(Grid);
 

Grid/lattice/Lattice_reduction.h

@@ -31,6 +31,7 @@ Author: Christoph Lehner <christoph@lhnr.de>
 #if defined(GRID_SYCL)
 #include <Grid/lattice/Lattice_reduction_sycl.h>
 #endif
+#include <Grid/lattice/Lattice_slicesum_core.h>
 
 NAMESPACE_BEGIN(Grid);
 
@@ -284,6 +285,7 @@ template<class vobj>
 inline ComplexD innerProduct(const Lattice<vobj> &left,const Lattice<vobj> &right) {
   GridBase *grid = left.Grid();
   ComplexD nrm = rankInnerProduct(left,right);
+  //  std::cerr<<"flight log " << std::hexfloat << nrm <<" "<<crc(left)<<std::endl;
   grid->GlobalSum(nrm);
   return nrm;
 }
@@ -448,19 +450,10 @@ template<class vobj> inline void sliceSum(const Lattice<vobj> &Data,std::vector<
   int e1=    grid->_slice_nblock[orthogdim];
   int e2=    grid->_slice_block [orthogdim];
   int stride=grid->_slice_stride[orthogdim];
-
-  // sum over reduced dimension planes, breaking out orthog dir
-  // Parallel over orthog direction
-  autoView( Data_v, Data, CpuRead);
-  thread_for( r,rd, {
-    int so=r*grid->_ostride[orthogdim]; // base offset for start of plane 
-    for(int n=0;n<e1;n++){
-      for(int b=0;b<e2;b++){
-	int ss= so+n*stride+b;
-	lvSum[r]=lvSum[r]+Data_v[ss];
-      }
-    }
-  });
+  int ostride=grid->_ostride[orthogdim];
+  
+  //Reduce Data down to lvSum
+  sliceSumReduction(Data,lvSum,rd, e1,e2,stride,ostride,Nsimd);
 
   // Sum across simd lanes in the plane, breaking out orthog dir.
   Coordinate icoor(Nd);
@@ -504,6 +497,7 @@ sliceSum(const Lattice<vobj> &Data,int orthogdim)
   return result;
 }
 
+
 template<class vobj>
 static void sliceInnerProductVector( std::vector<ComplexD> & result, const Lattice<vobj> &lhs,const Lattice<vobj> &rhs,int orthogdim) 
 {

Grid/lattice/Lattice_slicesum_core.h

@@ -0,0 +1,213 @@
+#pragma once
+#include <type_traits>
+#if defined(GRID_CUDA)
+
+#include <cub/cub.cuh>
+#define gpucub cub
+#define gpuError_t cudaError_t
+#define gpuSuccess cudaSuccess
+
+#elif defined(GRID_HIP)
+
+#include <hipcub/hipcub.hpp>
+#define gpucub hipcub
+#define gpuError_t hipError_t
+#define gpuSuccess hipSuccess
+
+#endif
+
+
+NAMESPACE_BEGIN(Grid);
+
+
+#if defined(GRID_CUDA) || defined(GRID_HIP)
+template<class vobj> inline void sliceSumReduction_cub_small(const vobj *Data, Vector<vobj> &lvSum, const int rd, const int e1, const int e2, const int stride, const int ostride, const int Nsimd) {
+  size_t subvol_size = e1*e2;
+  commVector<vobj> reduction_buffer(rd*subvol_size);
+  auto rb_p = &reduction_buffer[0];
+  vobj zero_init;
+  zeroit(zero_init);
+
+  
+  void *temp_storage_array = NULL;
+  size_t temp_storage_bytes = 0;
+  vobj *d_out;
+  int* d_offsets;
+
+  std::vector<int> offsets(rd+1,0);
+
+  for (int i = 0; i < offsets.size(); i++) {
+    offsets[i] = i*subvol_size;
+  }
+  
+  //Allocate memory for output and offset arrays on device
+  d_out = static_cast<vobj*>(acceleratorAllocDevice(rd*sizeof(vobj)));
+  
+  d_offsets = static_cast<int*>(acceleratorAllocDevice((rd+1)*sizeof(int)));
+  
+  //copy offsets to device
+  acceleratorCopyToDeviceAsync(&offsets[0],d_offsets,sizeof(int)*(rd+1),computeStream);
+  
+  
+  gpuError_t gpuErr = gpucub::DeviceSegmentedReduce::Reduce(temp_storage_array, temp_storage_bytes, rb_p,d_out, rd, d_offsets, d_offsets+1, ::gpucub::Sum(), zero_init, computeStream);
+  if (gpuErr!=gpuSuccess) {
+    std::cout << GridLogError << "Lattice_slicesum_gpu.h: Encountered error during gpucub::DeviceSegmentedReduce::Reduce (setup)! Error: " << gpuErr <<std::endl;
+    exit(EXIT_FAILURE);
+  }
+
+  //allocate memory for temp_storage_array  
+  temp_storage_array = acceleratorAllocDevice(temp_storage_bytes);
+  
+  //prepare buffer for reduction
+  //use non-blocking accelerator_for to avoid syncs (ok because we submit to same computeStream)
+  //use 2d accelerator_for to avoid launch latencies found when serially looping over rd 
+  accelerator_for2dNB( s,subvol_size, r,rd, Nsimd,{ 
+  
+    int n = s / e2;
+    int b = s % e2;
+    int so=r*ostride; // base offset for start of plane 
+    int ss= so+n*stride+b;
+
+    coalescedWrite(rb_p[r*subvol_size+s], coalescedRead(Data[ss]));
+
+  });
+  
+  //issue segmented reductions in computeStream
+  gpuErr = gpucub::DeviceSegmentedReduce::Reduce(temp_storage_array, temp_storage_bytes, rb_p, d_out, rd, d_offsets, d_offsets+1,::gpucub::Sum(), zero_init, computeStream);
+  if (gpuErr!=gpuSuccess) {
+    std::cout << GridLogError << "Lattice_slicesum_gpu.h: Encountered error during gpucub::DeviceSegmentedReduce::Reduce! Error: " << gpuErr <<std::endl;
+    exit(EXIT_FAILURE);
+  }
+  
+  acceleratorCopyFromDeviceAsync(d_out,&lvSum[0],rd*sizeof(vobj),computeStream);
+  
+  //sync after copy
+  accelerator_barrier();
+ 
+  acceleratorFreeDevice(temp_storage_array);
+  acceleratorFreeDevice(d_out);
+  acceleratorFreeDevice(d_offsets);
+  
+
+}
+
+template<class vobj> inline void sliceSumReduction_cub_large(const vobj *Data, Vector<vobj> &lvSum, const int rd, const int e1, const int e2, const int stride, const int ostride, const int Nsimd) {
+  typedef typename vobj::vector_type vector;
+  const int words = sizeof(vobj)/sizeof(vector);
+  const int osites = rd*e1*e2;
+  commVector<vector>buffer(osites);
+  vector *dat = (vector *)Data;
+  vector *buf = &buffer[0];
+  Vector<vector> lvSum_small(rd);
+  vector *lvSum_ptr = (vector *)&lvSum[0];
+
+  for (int w = 0; w < words; w++) {
+    accelerator_for(ss,osites,1,{
+	    buf[ss] = dat[ss*words+w];
+    });
+
+    sliceSumReduction_cub_small(buf,lvSum_small,rd,e1,e2,stride, ostride,Nsimd);
+      
+    for (int r = 0; r < rd; r++) {
+      lvSum_ptr[w+words*r]=lvSum_small[r];
+    }
+
+  }
+
+  
+}
+
+template<class vobj> inline void sliceSumReduction_cub(const Lattice<vobj> &Data, Vector<vobj> &lvSum, const int rd, const int e1, const int e2, const int stride, const int ostride, const int Nsimd)
+{
+  autoView(Data_v, Data, AcceleratorRead); //hipcub/cub cannot deal with large vobjs so we split into small/large case.
+    if constexpr (sizeof(vobj) <= 256) { 
+      sliceSumReduction_cub_small(&Data_v[0], lvSum, rd, e1, e2, stride, ostride, Nsimd);
+    }
+    else {
+      sliceSumReduction_cub_large(&Data_v[0], lvSum, rd, e1, e2, stride, ostride, Nsimd);
+    }
+}
+#endif
+
+
+#if defined(GRID_SYCL)
+template<class vobj> inline void sliceSumReduction_sycl(const Lattice<vobj> &Data, Vector <vobj> &lvSum, const int  &rd, const int &e1, const int &e2, const int &stride, const int &ostride, const int &Nsimd)
+{
+  typedef typename vobj::scalar_object sobj;
+  size_t subvol_size = e1*e2;
+
+  vobj *mysum = (vobj *) malloc_shared(sizeof(vobj),*theGridAccelerator);
+  vobj vobj_zero;
+  zeroit(vobj_zero);
+    
+  commVector<vobj> reduction_buffer(rd*subvol_size);    
+
+  auto rb_p = &reduction_buffer[0];
+
+  autoView(Data_v, Data, AcceleratorRead);
+
+  //prepare reduction buffer 
+  accelerator_for2d( s,subvol_size, r,rd, (size_t)Nsimd,{ 
+  
+      int n = s / e2;
+      int b = s % e2;
+      int so=r*ostride; // base offset for start of plane 
+      int ss= so+n*stride+b;
+
+      coalescedWrite(rb_p[r*subvol_size+s], coalescedRead(Data_v[ss]));
+
+  });
+
+  for (int r = 0; r < rd; r++) {
+      mysum[0] = vobj_zero; //dirty hack: cannot pass vobj_zero as identity to sycl::reduction as its not device_copyable
+      theGridAccelerator->submit([&](cl::sycl::handler &cgh) {
+          auto Reduction = cl::sycl::reduction(mysum,std::plus<>());
+          cgh.parallel_for(cl::sycl::range<1>{subvol_size},
+          Reduction,
+          [=](cl::sycl::id<1> item, auto &sum) {
+              auto s = item[0];
+              sum += rb_p[r*subvol_size+s];
+          });
+      });
+      theGridAccelerator->wait();
+      lvSum[r] = mysum[0];
+  }
+  
+  free(mysum,*theGridAccelerator);
+}
+#endif
+
+template<class vobj> inline void sliceSumReduction_cpu(const Lattice<vobj> &Data, Vector<vobj> &lvSum, const int &rd, const int &e1, const int &e2, const int &stride, const int &ostride, const int &Nsimd)
+{
+  // sum over reduced dimension planes, breaking out orthog dir
+  // Parallel over orthog direction
+  autoView( Data_v, Data, CpuRead);
+  thread_for( r,rd, {
+    int so=r*ostride; // base offset for start of plane 
+    for(int n=0;n<e1;n++){
+      for(int b=0;b<e2;b++){
+        int ss= so+n*stride+b;
+        lvSum[r]=lvSum[r]+Data_v[ss];
+      }
+    }
+  });
+}
+
+template<class vobj> inline void sliceSumReduction(const Lattice<vobj> &Data, Vector<vobj> &lvSum, const int &rd, const int &e1, const int &e2, const int &stride, const int &ostride, const int &Nsimd) 
+{
+  #if defined(GRID_CUDA) || defined(GRID_HIP)
+  
+  sliceSumReduction_cub(Data, lvSum, rd, e1, e2, stride, ostride, Nsimd);
+  
+  #elif defined(GRID_SYCL)
+  
+  sliceSumReduction_sycl(Data, lvSum, rd, e1, e2, stride, ostride, Nsimd);
+  
+  #else
+  sliceSumReduction_cpu(Data, lvSum, rd, e1, e2, stride, ostride, Nsimd);
+
+  #endif
+}
+
+
+NAMESPACE_END(Grid);

Grid/lattice/Lattice_transfer.h

@@ -469,15 +469,13 @@ 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 = zz;
+      vobj cd = Zero();
       
       for(int sb=0;sb<blockVol;sb++){
 

Grid/lattice/Lattice_view.h

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

Grid/perfmon/Tracing.h

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

Grid/qcd/action/fermion/WilsonTMFermion.h

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

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

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

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

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

Grid/qcd/hmc/integrators/Integrator.h

@@ -237,7 +237,7 @@ public:
 
     for (int level = 0; level < as.size(); ++level) {
       int multiplier = as.at(level).multiplier;
-      ActionLevel<Field> * Level = new ActionLevel<Field>(multiplier);
+      ActionLevel<Field, RepresentationPolicy> * Level = new ActionLevel<Field, RepresentationPolicy>(multiplier);
       Level->push_back(new EmptyAction<Field>); 
       LevelForces.push_back(*Level);
       // does it copy by value or reference??

Grid/qcd/smearing/HISQSmearing.h

@@ -138,7 +138,7 @@ public:
         for(int nu=0;nu<Nd;nu++) {
             appendShift(shifts,mu);
             appendShift(shifts,nu);
-            appendShift(shifts,NO_SHIFT);
+            appendShift(shifts,shiftSignal::NO_SHIFT);
             appendShift(shifts,mu,Back(nu));
             appendShift(shifts,Back(nu));
             appendShift(shifts,Back(mu));
@@ -173,7 +173,6 @@ public:
             auto gStencil_v = gStencil.View(); 
 
             accelerator_for(site,Nsites,Simd::Nsimd(),{ // ----------- 3-link constructs
-//            for(int site=0;site<Nsites;site++){ // ----------- 3-link constructs
                 stencilElement SE0, SE1, SE2, SE3, SE4, SE5;
                 U3matrix U0, U1, U2, U3, U4, U5, W;
                 for(int nu=0;nu<Nd;nu++) {
@@ -216,7 +215,6 @@ public:
             })
 
             accelerator_for(site,Nsites,Simd::Nsimd(),{ // ----------- 5-link 
-//            for(int site=0;site<Nsites;site++){ // ----------- 5-link
                 stencilElement SE0, SE1, SE2, SE3, SE4, SE5;
                 U3matrix U0, U1, U2, U3, U4, U5, W;
                 int sigmaIndex = 0;
@@ -254,7 +252,6 @@ public:
             })
 
             accelerator_for(site,Nsites,Simd::Nsimd(),{ // ----------- 7-link
-//            for(int site=0;site<Nsites;site++){ // ----------- 7-link
                 stencilElement SE0, SE1, SE2, SE3, SE4, SE5;
                 U3matrix U0, U1, U2, U3, U4, U5, W;
                 int sigmaIndex = 0;

Grid/simd/Grid_vector_types.h

@@ -1133,4 +1133,13 @@ static_assert(sizeof(SIMD_Ftype) == sizeof(SIMD_Itype), "SIMD vector lengths inc
 
 NAMESPACE_END(Grid);
 
+#ifdef GRID_SYCL
+template<> struct sycl::is_device_copyable<Grid::vComplexF> : public std::true_type {};
+template<> struct sycl::is_device_copyable<Grid::vComplexD> : public std::true_type {};
+template<> struct sycl::is_device_copyable<Grid::vRealF   > : public std::true_type {};
+template<> struct sycl::is_device_copyable<Grid::vRealD   > : public std::true_type {};
+template<> struct sycl::is_device_copyable<Grid::vInteger > : public std::true_type {};
+#endif
+
+
 #endif

Grid/stencil/GeneralLocalStencil.h

@@ -141,8 +141,14 @@ public:
 ////////////////////////////////////////////////
 // Some machinery to streamline making a stencil 
 ////////////////////////////////////////////////
-#define BACKWARD_CONST 16
-#define NO_SHIFT -1
+
+class shiftSignal {
+public:
+    enum {
+        BACKWARD_CONST = 16,
+        NO_SHIFT       = -1
+    };
+};
 
 // TODO: put a check somewhere that BACKWARD_CONST > Nd!
 
@@ -150,16 +156,16 @@ public:
 inline int Back(const int dir) {
     // generalShift will use BACKWARD_CONST to determine whether we step forward or 
     // backward. Trick inspired by SIMULATeQCD. 
-    return dir + BACKWARD_CONST;
+    return dir + shiftSignal::BACKWARD_CONST;
 }
 
 /*!  @brief shift one unit in direction dir */
 template<typename... Args>
 void generalShift(Coordinate& shift, int dir) {
-    if (dir >= BACKWARD_CONST) {
-        dir -= BACKWARD_CONST;
+    if (dir >= shiftSignal::BACKWARD_CONST) {
+        dir -= shiftSignal::BACKWARD_CONST;
         shift[dir]+=-1;
-    } else if (dir == NO_SHIFT) {
+    } else if (dir == shiftSignal::NO_SHIFT) {
         ; // do nothing
     } else {
         shift[dir]+=1;
@@ -169,10 +175,10 @@ void generalShift(Coordinate& shift, int dir) {
 /*!  @brief follow a path of directions, shifting one unit in each direction */
 template<typename... Args>
 void generalShift(Coordinate& shift, int dir, Args... args) {
-    if (dir >= BACKWARD_CONST) {
-        dir -= BACKWARD_CONST;
+    if (dir >= shiftSignal::BACKWARD_CONST) {
+        dir -= shiftSignal::BACKWARD_CONST;
         shift[dir]+=-1;
-    } else if (dir == NO_SHIFT) {
+    } else if (dir == shiftSignal::NO_SHIFT) {
         ; // do nothing
     } else {
         shift[dir]+=1;

Grid/stencil/Stencil.h

@@ -706,7 +706,7 @@ public:
 	}
       }
     }
-    std::cout << GridLogDebug << "BuildSurfaceList size is "<<surface_list.size()<<std::endl;
+    //std::cout << "BuildSurfaceList size is "<<surface_list.size()<<std::endl;
   }
   /// Introduce a block structure and switch off comms on boundaries
   void DirichletBlock(const Coordinate &dirichlet_block)
@@ -761,7 +761,8 @@ public:
 		   int checkerboard,
 		   const std::vector<int> &directions,
 		   const std::vector<int> &distances,
-		   Parameters p=Parameters())
+		   Parameters p=Parameters(),
+		   bool preserve_shm=false)
   {
     face_table_computed=0;
     _grid    = grid;
@@ -855,7 +856,9 @@ public:
     /////////////////////////////////////////////////////////////////////////////////
     const int Nsimd = grid->Nsimd();
 
-    _grid->ShmBufferFreeAll();
+    // Allow for multiple stencils to exist simultaneously
+    if (!preserve_shm)
+      _grid->ShmBufferFreeAll();
 
     int maxl=2;
     u_simd_send_buf.resize(maxl);

Grid/tensors/Tensor_traits.h

@@ -404,3 +404,12 @@ NAMESPACE_BEGIN(Grid);
   };
 NAMESPACE_END(Grid);
 
+
+#ifdef GRID_SYCL
+template<typename T> struct
+sycl::is_device_copyable<T, typename std::enable_if<
+			      Grid::isGridTensor<T>::value  && (!std::is_trivially_copyable<T>::value),
+			      void>::type>
+  : public std::true_type {};
+#endif
+

Grid/threads/Accelerator.h

@@ -225,6 +225,8 @@ inline void acceleratorFreeShared(void *ptr){ cudaFree(ptr);};
 inline void acceleratorFreeDevice(void *ptr){ cudaFree(ptr);};
 inline void acceleratorCopyToDevice(void *from,void *to,size_t bytes)  { cudaMemcpy(to,from,bytes, cudaMemcpyHostToDevice);}
 inline void acceleratorCopyFromDevice(void *from,void *to,size_t bytes){ cudaMemcpy(to,from,bytes, cudaMemcpyDeviceToHost);}
+inline void acceleratorCopyToDeviceAsync(void *from, void *to, size_t bytes, cudaStream_t stream = copyStream) { cudaMemcpyAsync(to,from,bytes, cudaMemcpyHostToDevice, stream);}
+inline void acceleratorCopyFromDeviceAsync(void *from, void *to, size_t bytes, cudaStream_t stream = copyStream) { cudaMemcpyAsync(to,from,bytes, cudaMemcpyDeviceToHost, stream);}
 inline void acceleratorMemSet(void *base,int value,size_t bytes) { cudaMemset(base,value,bytes);}
 inline void acceleratorCopyDeviceToDeviceAsynch(void *from,void *to,size_t bytes) // Asynch
 {
@@ -253,17 +255,13 @@ inline int  acceleratorIsCommunicable(void *ptr)
 #define GRID_SYCL_LEVEL_ZERO_IPC
 
 NAMESPACE_END(Grid);
-#if 0
-#include <CL/sycl.hpp>
-#include <CL/sycl/usm.hpp>
-#include <level_zero/ze_api.h>
-#include <CL/sycl/backend/level_zero.hpp>
-#else
+
+// Force deterministic reductions
+#define SYCL_REDUCTION_DETERMINISTIC
 #include <sycl/CL/sycl.hpp>
 #include <sycl/usm.hpp>
 #include <level_zero/ze_api.h>
 #include <sycl/ext/oneapi/backend/level_zero.hpp>
-#endif
 
 NAMESPACE_BEGIN(Grid);
 
@@ -287,23 +285,24 @@ accelerator_inline int acceleratorSIMTlane(int Nsimd) {
 
 #define accelerator_for2dNB( iter1, num1, iter2, num2, nsimd, ... )	\
   theGridAccelerator->submit([&](cl::sycl::handler &cgh) {		\
-      unsigned long nt=acceleratorThreads();				\
-      unsigned long unum1 = num1;					\
-      unsigned long unum2 = num2;					\
-      if(nt < 8)nt=8;							\
-      cl::sycl::range<3> local {nt,1,nsimd};				\
-      cl::sycl::range<3> global{unum1,unum2,nsimd};			\
-      cgh.parallel_for(					\
-      cl::sycl::nd_range<3>(global,local), \
-      [=] (cl::sycl::nd_item<3> item) /*mutable*/     \
-      [[intel::reqd_sub_group_size(16)]]	      \
-      {						      \
-      auto iter1    = item.get_global_id(0);	      \
-      auto iter2    = item.get_global_id(1);	      \
-      auto lane     = item.get_global_id(2);	      \
-      { __VA_ARGS__ };				      \
-     });	   			              \
-    });
+    unsigned long nt=acceleratorThreads();				\
+    if(nt < 8)nt=8;							\
+    unsigned long unum1 = num1;						\
+    unsigned long unum2 = num2;						\
+    unsigned long unum1_divisible_by_nt = ((unum1 + nt - 1) / nt) * nt;	\
+    cl::sycl::range<3> local {nt,1,nsimd};				\
+    cl::sycl::range<3> global{unum1_divisible_by_nt,unum2,nsimd};	\
+    cgh.parallel_for(							\
+		     cl::sycl::nd_range<3>(global,local),		\
+		     [=] (cl::sycl::nd_item<3> item) /*mutable*/	\
+		     [[intel::reqd_sub_group_size(16)]]			\
+		     {							\
+		       auto iter1    = item.get_global_id(0);		\
+		       auto iter2    = item.get_global_id(1);		\
+		       auto lane     = item.get_global_id(2);		\
+		       { if (iter1 < unum1){ __VA_ARGS__ } };		\
+		     });						\
+  });
 
 #define accelerator_barrier(dummy) { theGridAccelerator->wait(); }
 
@@ -442,6 +441,8 @@ inline void acceleratorFreeShared(void *ptr){ auto r=hipFree(ptr);};
 inline void acceleratorFreeDevice(void *ptr){ auto r=hipFree(ptr);};
 inline void acceleratorCopyToDevice(void *from,void *to,size_t bytes)  { auto r=hipMemcpy(to,from,bytes, hipMemcpyHostToDevice);}
 inline void acceleratorCopyFromDevice(void *from,void *to,size_t bytes){ auto r=hipMemcpy(to,from,bytes, hipMemcpyDeviceToHost);}
+inline void acceleratorCopyToDeviceAsync(void *from, void *to, size_t bytes, hipStream_t stream = copyStream) { auto r = hipMemcpyAsync(to,from,bytes, hipMemcpyHostToDevice, stream);}
+inline void acceleratorCopyFromDeviceAsync(void *from, void *to, size_t bytes, hipStream_t stream = copyStream) { auto r = hipMemcpyAsync(to,from,bytes, hipMemcpyDeviceToHost, stream);}
 //inline void acceleratorCopyDeviceToDeviceAsynch(void *from,void *to,size_t bytes)  { hipMemcpy(to,from,bytes, hipMemcpyDeviceToDevice);}
 //inline void acceleratorCopySynchronise(void) {  }
 inline void acceleratorMemSet(void *base,int value,size_t bytes) { auto r=hipMemset(base,value,bytes);}

Grid/util/Init.cc

@@ -77,6 +77,10 @@ feenableexcept (unsigned int excepts)
 }
 #endif
 
+#ifndef HOST_NAME_MAX
+#define HOST_NAME_MAX _POSIX_HOST_NAME_MAX
+#endif
+
 NAMESPACE_BEGIN(Grid);
 
 //////////////////////////////////////////////////////
@@ -393,6 +397,9 @@ void Grid_init(int *argc,char ***argv)
   std::cout << GridLogMessage << "MPI is initialised and logging filters activated "<<std::endl;
   std::cout << GridLogMessage << "================================================ "<<std::endl;
 
+  char hostname[HOST_NAME_MAX+1];
+  gethostname(hostname, HOST_NAME_MAX+1);
+  std::cout << GridLogMessage << "This rank is running on host "<< hostname<<std::endl;
 
   /////////////////////////////////////////////////////////
   // Reporting

benchmarks/Benchmark_usqcd.cc

@@ -0,0 +1,968 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./benchmarks/Benchmark_usqcd.cc
+
+    Copyright (C) 2015
+
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+Author: paboyle <paboyle@ph.ed.ac.uk>
+
+    This program is free software; you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation; either version 2 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License along
+    with this program; if not, write to the Free Software Foundation, Inc.,
+    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+    See the full license in the file "LICENSE" in the top level distribution directory
+    *************************************************************************************/
+    /*  END LEGAL */
+#include <Grid/Grid.h>
+#include <Grid/algorithms/blas/BatchedBlas.h>
+
+using namespace Grid;
+
+std::vector<int> L_list;
+std::vector<int> Ls_list;
+std::vector<double> mflop_list;
+
+double mflop_ref;
+double mflop_ref_err;
+
+int NN_global;
+
+FILE * FP;
+
+struct time_statistics{
+  double mean;
+  double err;
+  double min;
+  double max;
+
+  void statistics(std::vector<double> v){
+      double sum = std::accumulate(v.begin(), v.end(), 0.0);
+      mean = sum / v.size();
+
+      std::vector<double> diff(v.size());
+      std::transform(v.begin(), v.end(), diff.begin(), [=](double x) { return x - mean; });
+      double sq_sum = std::inner_product(diff.begin(), diff.end(), diff.begin(), 0.0);
+      err = std::sqrt(sq_sum / (v.size()*(v.size() - 1)));
+
+      auto result = std::minmax_element(v.begin(), v.end());
+      min = *result.first;
+      max = *result.second;
+}
+};
+
+void comms_header(){
+  std::cout <<GridLogMessage << " L  "<<"\t"<<" Ls  "<<"\t"
+            <<"bytes\t MB/s uni  \t\t MB/s bidi "<<std::endl;
+};
+
+struct controls {
+  int Opt;
+  int CommsOverlap;
+  Grid::CartesianCommunicator::CommunicatorPolicy_t CommsAsynch;
+};
+
+class Benchmark {
+public:
+  static void Decomposition (void ) {
+
+    int threads = GridThread::GetThreads();
+    std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
+    std::cout<<GridLogMessage << "= Grid is setup to use "<<threads<<" threads"<<std::endl;
+    std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
+    std::cout<<GridLogMessage<<"Grid Default Decomposition patterns\n";
+    std::cout<<GridLogMessage<<"\tOpenMP threads : "<<GridThread::GetThreads()<<std::endl;
+    std::cout<<GridLogMessage<<"\tMPI tasks      : "<<GridCmdVectorIntToString(GridDefaultMpi())<<std::endl;
+    std::cout<<GridLogMessage<<"\tvReal          : "<<sizeof(vReal )*8    <<"bits ; " <<GridCmdVectorIntToString(GridDefaultSimd(4,vReal::Nsimd()))<<std::endl;
+    std::cout<<GridLogMessage<<"\tvRealF         : "<<sizeof(vRealF)*8    <<"bits ; " <<GridCmdVectorIntToString(GridDefaultSimd(4,vRealF::Nsimd()))<<std::endl;
+    std::cout<<GridLogMessage<<"\tvRealD         : "<<sizeof(vRealD)*8    <<"bits ; " <<GridCmdVectorIntToString(GridDefaultSimd(4,vRealD::Nsimd()))<<std::endl;
+    std::cout<<GridLogMessage<<"\tvComplex       : "<<sizeof(vComplex )*8 <<"bits ; " <<GridCmdVectorIntToString(GridDefaultSimd(4,vComplex::Nsimd()))<<std::endl;
+    std::cout<<GridLogMessage<<"\tvComplexF      : "<<sizeof(vComplexF)*8 <<"bits ; " <<GridCmdVectorIntToString(GridDefaultSimd(4,vComplexF::Nsimd()))<<std::endl;
+    std::cout<<GridLogMessage<<"\tvComplexD      : "<<sizeof(vComplexD)*8 <<"bits ; " <<GridCmdVectorIntToString(GridDefaultSimd(4,vComplexD::Nsimd()))<<std::endl;
+    std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
+
+  }
+
+  static void Comms(void)
+  {
+    int Nloop=200;
+    int nmu=0;
+    int maxlat=32;
+
+    Coordinate simd_layout = GridDefaultSimd(Nd,vComplexD::Nsimd());
+    Coordinate mpi_layout  = GridDefaultMpi();
+
+    for(int mu=0;mu<Nd;mu++) if (mpi_layout[mu]>1) nmu++;
+
+    std::vector<double> t_time(Nloop);
+    time_statistics timestat;
+
+    std::cout<<GridLogMessage << "===================================================================================================="<<std::endl;
+    std::cout<<GridLogMessage << "= Benchmarking threaded STENCIL halo exchange in "<<nmu<<" dimensions"<<std::endl;
+    std::cout<<GridLogMessage << "===================================================================================================="<<std::endl;
+    comms_header();
+
+    fprintf(FP,"Communications\n\n");
+    fprintf(FP,"Packet bytes, direction, GB/s per node\n");
+    for(int lat=16;lat<=maxlat;lat+=8){
+      //      for(int Ls=8;Ls<=8;Ls*=2){
+      { int Ls=12;
+
+	Coordinate latt_size  ({lat*mpi_layout[0],
+	      lat*mpi_layout[1],
+	      lat*mpi_layout[2],
+	      lat*mpi_layout[3]});
+
+	GridCartesian     Grid(latt_size,simd_layout,mpi_layout);
+	RealD Nrank = Grid._Nprocessors;
+	RealD Nnode = Grid.NodeCount();
+	RealD ppn = Nrank/Nnode;
+
+	std::vector<HalfSpinColourVectorD *> xbuf(8);
+	std::vector<HalfSpinColourVectorD *> rbuf(8);
+	//Grid.ShmBufferFreeAll();
+	uint64_t bytes=lat*lat*lat*Ls*sizeof(HalfSpinColourVectorD);
+	for(int d=0;d<8;d++){
+	  xbuf[d] = (HalfSpinColourVectorD *)acceleratorAllocDevice(bytes);
+	  rbuf[d] = (HalfSpinColourVectorD *)acceleratorAllocDevice(bytes);
+	  //	  bzero((void *)xbuf[d],lat*lat*lat*Ls*sizeof(HalfSpinColourVectorD));
+	  //	  bzero((void *)rbuf[d],lat*lat*lat*Ls*sizeof(HalfSpinColourVectorD));
+	}
+
+	//	int ncomm;
+	double dbytes;
+
+        for(int dir=0;dir<8;dir++) {
+	  int mu =dir % 4;
+	  if (mpi_layout[mu]>1 ) {
+
+	    std::vector<double> times(Nloop);
+	    for(int i=0;i<Nloop;i++){
+
+	      dbytes=0;	        
+	      double start=usecond();
+	      int xmit_to_rank;
+	      int recv_from_rank;
+
+	      if ( dir == mu ) { 
+		int comm_proc=1;
+		Grid.ShiftedRanks(mu,comm_proc,xmit_to_rank,recv_from_rank);
+	      } else { 
+		int comm_proc = mpi_layout[mu]-1;
+		Grid.ShiftedRanks(mu,comm_proc,xmit_to_rank,recv_from_rank);
+	      }
+	      Grid.SendToRecvFrom((void *)&xbuf[dir][0], xmit_to_rank,
+				  (void *)&rbuf[dir][0], recv_from_rank,
+				  bytes);
+	      dbytes+=bytes;
+	     
+	      double stop=usecond();
+	      t_time[i] = stop-start; // microseconds
+
+	    }
+	    timestat.statistics(t_time);
+	  
+	    dbytes=dbytes*ppn;
+	    double xbytes    = dbytes*0.5;
+	    double bidibytes = dbytes;
+	  
+	    std::cout<<GridLogMessage << lat<<"\t"<<Ls<<"\t "
+		     << bytes << " \t "
+		     <<xbytes/timestat.mean
+		     << "\t\t"
+		     << bidibytes/timestat.mean<< std::endl;
+	    fprintf(FP,"%ld, %d, %f\n",(long)bytes,dir,bidibytes/timestat.mean/1000.);
+	  }
+	}
+	for(int d=0;d<8;d++){
+	  acceleratorFreeDevice(xbuf[d]);
+	  acceleratorFreeDevice(rbuf[d]);
+	}
+      }
+    }
+    fprintf(FP,"\n\n");
+    
+    return;
+  }
+
+  
+  static void Memory(void)
+  {
+    const int Nvec=8;
+    typedef Lattice< iVector< vReal,Nvec> > LatticeVec;
+    typedef iVector<vReal,Nvec> Vec;
+
+    Coordinate simd_layout = GridDefaultSimd(Nd,vReal::Nsimd());
+    Coordinate mpi_layout  = GridDefaultMpi();
+
+    fprintf(FP,"Memory Bandwidth\n\n");
+    fprintf(FP,"Bytes, GB/s per node\n");
+    std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
+    std::cout<<GridLogMessage << "= Benchmarking a*x + y bandwidth"<<std::endl;
+    std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
+    std::cout<<GridLogMessage << "  L  "<<"\t\t"<<"bytes"<<"\t\t\t"<<"GB/s"<<"\t\t"<<"Gflop/s"<<"\t\t seconds"<< "\t\tGB/s / node"<<std::endl;
+    std::cout<<GridLogMessage << "----------------------------------------------------------"<<std::endl;
+  
+    //    uint64_t NP;
+    uint64_t NN;
+
+
+  uint64_t lmax=40;
+#define NLOOP (1000*lmax*lmax*lmax*lmax/lat/lat/lat/lat)
+
+    GridSerialRNG          sRNG;      sRNG.SeedFixedIntegers(std::vector<int>({45,12,81,9}));
+    for(int lat=8;lat<=lmax;lat+=8){
+
+      Coordinate latt_size  ({lat*mpi_layout[0],lat*mpi_layout[1],lat*mpi_layout[2],lat*mpi_layout[3]});
+      int64_t vol= latt_size[0]*latt_size[1]*latt_size[2]*latt_size[3];
+
+      GridCartesian     Grid(latt_size,simd_layout,mpi_layout);
+
+      //      NP= Grid.RankCount();
+      NN =Grid.NodeCount();
+
+      Vec rn ; random(sRNG,rn);
+
+      LatticeVec z(&Grid); z=Zero();
+      LatticeVec x(&Grid); x=Zero();
+      LatticeVec y(&Grid); y=Zero();
+      double a=2.0;
+
+      uint64_t Nloop=NLOOP;
+
+      double start=usecond();
+      for(int i=0;i<Nloop;i++){
+	z=a*x-y;
+      }
+      double stop=usecond();
+      double time = (stop-start)/Nloop*1000;
+     
+      double flops=vol*Nvec*2;// mul,add
+      double bytes=3.0*vol*Nvec*sizeof(Real);
+      std::cout<<GridLogMessage<<std::setprecision(3) 
+	       << lat<<"\t\t"<<bytes<<"   \t\t"<<bytes/time<<"\t\t"<<flops/time<<"\t\t"<<(stop-start)/1000./1000.
+	       << "\t\t"<< bytes/time/NN <<std::endl;
+
+      fprintf(FP,"%ld, %f\n",(long)bytes,bytes/time/NN);
+
+    }
+    fprintf(FP,"\n\n");
+  };
+
+
+  static void BLAS(void)
+  {
+    //int nbasis, int nrhs, int coarseVol
+    int  basis[] = { 16,32,64 };
+    int  rhs[]   = { 8,16,32 };
+    int  vol  = 4*4*4*4;
+
+    GridBLAS blas;
+    
+    std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
+    std::cout<<GridLogMessage << "= batched GEMM (double precision) "<<std::endl;
+    std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
+    std::cout<<GridLogMessage << "  M  "<<"\t\t"<<"N"<<"\t\t\t"<<"K"<<"\t\t"<<"Gflop/s / rank (coarse mrhs)"<<std::endl;
+    std::cout<<GridLogMessage << "----------------------------------------------------------"<<std::endl;
+  
+    fprintf(FP,"GEMM\n\n M, N, K, BATCH, GF/s per rank\n");
+
+    for(int b=0;b<3;b++){
+    for(int r=0;r<3;r++){
+      int M=basis[b];
+      int N=rhs[r];
+      int K=basis[b];
+      int BATCH=vol;
+      double p=blas.benchmark(M,N,K,BATCH);
+
+      fprintf(FP,"%d, %d, %d, %d, %f\n", M, N, K, BATCH, p);
+      
+      std::cout<<GridLogMessage<<std::setprecision(3) 
+	       << M<<"\t\t"<<N<<"\t\t"<<K<<"\t\t"<<BATCH<<"\t\t"<<p<<std::endl;
+    }}
+    std::cout<<GridLogMessage << "----------------------------------------------------------"<<std::endl;
+    std::cout<<GridLogMessage << "  M  "<<"\t\t"<<"N"<<"\t\t\t"<<"K"<<"\t\t"<<"Gflop/s / rank (block project)"<<std::endl;
+    std::cout<<GridLogMessage << "----------------------------------------------------------"<<std::endl;
+    for(int b=0;b<3;b++){
+    for(int r=0;r<3;r++){
+      int M=basis[b];
+      int N=rhs[r];
+      int K=vol;
+      int BATCH=vol;
+      double p=blas.benchmark(M,N,K,BATCH);
+
+      fprintf(FP,"%d, %d, %d, %d, %f\n", M, N, K, BATCH, p);
+      std::cout<<GridLogMessage<<std::setprecision(3) 
+	       << M<<"\t\t"<<N<<"\t\t"<<K<<"\t\t"<<BATCH<<"\t\t"<<p<<std::endl;
+    }}
+    std::cout<<GridLogMessage << "----------------------------------------------------------"<<std::endl;
+    std::cout<<GridLogMessage << "  M  "<<"\t\t"<<"N"<<"\t\t\t"<<"K"<<"\t\t"<<"Gflop/s / rank (block promote)"<<std::endl;
+    std::cout<<GridLogMessage << "----------------------------------------------------------"<<std::endl;
+    for(int b=0;b<3;b++){
+    for(int r=0;r<3;r++){
+      int M=rhs[r];
+      int N=vol;
+      int K=basis[b];
+      int BATCH=vol;
+      double p=blas.benchmark(M,N,K,BATCH);
+
+      fprintf(FP,"%d, %d, %d, %d, %f\n", M, N, K, BATCH, p);
+      std::cout<<GridLogMessage<<std::setprecision(3) 
+	       << M<<"\t\t"<<N<<"\t\t"<<K<<"\t\t"<<BATCH<<"\t\t"<<p<<std::endl;
+    }}
+    fprintf(FP,"\n\n\n");
+    std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
+  };
+  
+
+  static void SU4(void)
+  {
+    const int Nc4=4;
+    typedef Lattice< iMatrix< vComplexF,Nc4> > LatticeSU4;
+
+    Coordinate simd_layout = GridDefaultSimd(Nd,vComplexF::Nsimd());
+    Coordinate mpi_layout  = GridDefaultMpi();
+    
+    std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
+    std::cout<<GridLogMessage << "= Benchmarking z = y*x SU(4) bandwidth"<<std::endl;
+    std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
+    std::cout<<GridLogMessage << "  L  "<<"\t\t"<<"bytes"<<"\t\t\t"<<"GB/s"<<"\t\t"<<"Gflop/s"<<"\t\t seconds"<< "\t\tGB/s / node"<<std::endl;
+    std::cout<<GridLogMessage << "----------------------------------------------------------"<<std::endl;
+  
+    uint64_t NN;
+
+
+    uint64_t lmax=32;
+
+    GridSerialRNG          sRNG;      sRNG.SeedFixedIntegers(std::vector<int>({45,12,81,9}));
+    for(int lat=8;lat<=lmax;lat+=8){
+
+      Coordinate latt_size  ({lat*mpi_layout[0],lat*mpi_layout[1],lat*mpi_layout[2],lat*mpi_layout[3]});
+      int64_t vol= latt_size[0]*latt_size[1]*latt_size[2]*latt_size[3];
+
+      GridCartesian     Grid(latt_size,simd_layout,mpi_layout);
+
+      NN =Grid.NodeCount();
+
+
+      LatticeSU4 z(&Grid); z=Zero();
+      LatticeSU4 x(&Grid); x=Zero();
+      LatticeSU4 y(&Grid); y=Zero();
+      //      double a=2.0;
+
+      uint64_t Nloop=NLOOP;
+
+      double start=usecond();
+      for(int i=0;i<Nloop;i++){
+	z=x*y;
+      }
+      double stop=usecond();
+      double time = (stop-start)/Nloop*1000;
+     
+      double flops=vol*Nc4*Nc4*(6+(Nc4-1)*8);// mul,add
+      double bytes=3.0*vol*Nc4*Nc4*2*sizeof(RealF);
+      std::cout<<GridLogMessage<<std::setprecision(3) 
+	       << lat<<"\t\t"<<bytes<<"   \t\t"<<bytes/time<<"\t\t"<<flops/time<<"\t\t"<<(stop-start)/1000./1000.
+	       << "\t\t"<< bytes/time/NN <<std::endl;
+
+    }
+  };
+
+
+  static double DWF(int Ls,int L)
+  {
+    RealD mass=0.1;
+    RealD M5  =1.8;
+
+    double mflops;
+    double mflops_best = 0;
+    double mflops_worst= 0;
+    std::vector<double> mflops_all;
+
+    ///////////////////////////////////////////////////////
+    // Set/Get the layout & grid size
+    ///////////////////////////////////////////////////////
+    int threads = GridThread::GetThreads();
+    Coordinate mpi = GridDefaultMpi(); assert(mpi.size()==4);
+    Coordinate local({L,L,L,L});
+    Coordinate latt4({local[0]*mpi[0],local[1]*mpi[1],local[2]*mpi[2],local[3]*mpi[3]});
+
+    GridCartesian         * TmpGrid   = SpaceTimeGrid::makeFourDimGrid(latt4, 
+								       GridDefaultSimd(Nd,vComplex::Nsimd()),
+								       GridDefaultMpi());
+    uint64_t NP = TmpGrid->RankCount();
+    uint64_t NN = TmpGrid->NodeCount();
+    NN_global=NN;
+    uint64_t SHM=NP/NN;
+
+
+    ///////// Welcome message ////////////
+    std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
+    std::cout<<GridLogMessage << "Benchmark DWF on "<<L<<"^4 local volume "<<std::endl;
+    std::cout<<GridLogMessage << "* Nc             : "<<Nc<<std::endl;
+    std::cout<<GridLogMessage << "* Global volume  : "<<GridCmdVectorIntToString(latt4)<<std::endl;
+    std::cout<<GridLogMessage << "* Ls             : "<<Ls<<std::endl;
+    std::cout<<GridLogMessage << "* ranks          : "<<NP  <<std::endl;
+    std::cout<<GridLogMessage << "* nodes          : "<<NN  <<std::endl;
+    std::cout<<GridLogMessage << "* ranks/node     : "<<SHM <<std::endl;
+    std::cout<<GridLogMessage << "* ranks geom     : "<<GridCmdVectorIntToString(mpi)<<std::endl;
+    std::cout<<GridLogMessage << "* Using "<<threads<<" threads"<<std::endl;
+    std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
+
+    ///////// Lattice Init ////////////
+    GridCartesian         * UGrid   = SpaceTimeGrid::makeFourDimGrid(latt4, GridDefaultSimd(Nd,vComplexF::Nsimd()),GridDefaultMpi());
+    GridRedBlackCartesian * UrbGrid = SpaceTimeGrid::makeFourDimRedBlackGrid(UGrid);
+    GridCartesian         * FGrid   = SpaceTimeGrid::makeFiveDimGrid(Ls,UGrid);
+    GridRedBlackCartesian * FrbGrid = SpaceTimeGrid::makeFiveDimRedBlackGrid(Ls,UGrid);
+
+    
+    ///////// RNG Init ////////////
+    std::vector<int> seeds4({1,2,3,4});
+    std::vector<int> seeds5({5,6,7,8});
+    GridParallelRNG          RNG4(UGrid);  RNG4.SeedFixedIntegers(seeds4);
+    GridParallelRNG          RNG5(FGrid);  RNG5.SeedFixedIntegers(seeds5);
+    std::cout << GridLogMessage << "Initialised RNGs" << std::endl;
+
+    typedef DomainWallFermionF Action;
+    typedef typename Action::FermionField Fermion;
+    typedef LatticeGaugeFieldF Gauge;
+    
+    ///////// Source preparation ////////////
+    Gauge Umu(UGrid);  SU<Nc>::HotConfiguration(RNG4,Umu); 
+    Fermion src   (FGrid); random(RNG5,src);
+    Fermion src_e (FrbGrid);
+    Fermion src_o (FrbGrid);
+    Fermion r_e   (FrbGrid);
+    Fermion r_o   (FrbGrid);
+    Fermion r_eo  (FGrid);
+    Action Dw(Umu,*FGrid,*FrbGrid,*UGrid,*UrbGrid,mass,M5);
+
+    {
+
+      pickCheckerboard(Even,src_e,src);
+      pickCheckerboard(Odd,src_o,src);
+
+#ifdef AVX512
+      const int num_cases = 3;
+#else 
+      const int num_cases = 2;
+#endif      
+      std::string fmt("G/S/C ; G/O/C ; G/S/S ; G/O/S ");
+
+      controls Cases [] = {
+	{  WilsonKernelsStatic::OptGeneric   ,  WilsonKernelsStatic::CommsAndCompute  ,CartesianCommunicator::CommunicatorPolicyConcurrent },
+	{  WilsonKernelsStatic::OptHandUnroll,  WilsonKernelsStatic::CommsAndCompute  ,CartesianCommunicator::CommunicatorPolicyConcurrent },
+	{  WilsonKernelsStatic::OptInlineAsm ,  WilsonKernelsStatic::CommsAndCompute  ,CartesianCommunicator::CommunicatorPolicyConcurrent }
+      }; 
+
+      for(int c=0;c<num_cases;c++) {
+	
+	WilsonKernelsStatic::Comms = Cases[c].CommsOverlap;
+	WilsonKernelsStatic::Opt   = Cases[c].Opt;
+	CartesianCommunicator::SetCommunicatorPolicy(Cases[c].CommsAsynch);
+
+	std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
+	if ( WilsonKernelsStatic::Opt == WilsonKernelsStatic::OptGeneric   ) std::cout << GridLogMessage<< "* Using GENERIC Nc WilsonKernels" <<std::endl;
+	if ( WilsonKernelsStatic::Opt == WilsonKernelsStatic::OptInlineAsm ) std::cout << GridLogMessage<< "* Using ASM      WilsonKernels" <<std::endl;
+	if ( WilsonKernelsStatic::Opt == WilsonKernelsStatic::OptHandUnroll) std::cout << GridLogMessage<< "* Using UNROLLED WilsonKernels" <<std::endl;
+	if ( WilsonKernelsStatic::Comms == WilsonKernelsStatic::CommsAndCompute ) std::cout << GridLogMessage<< "* Using Overlapped Comms/Compute" <<std::endl;
+	if ( WilsonKernelsStatic::Comms == WilsonKernelsStatic::CommsThenCompute) std::cout << GridLogMessage<< "* Using sequential Comms/Compute" <<std::endl;
+	std::cout << GridLogMessage<< "* SINGLE precision "<<std::endl;
+	std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
+
+	int nwarm = 10;
+	double t0=usecond();
+	FGrid->Barrier();
+	for(int i=0;i<nwarm;i++){
+	  Dw.DhopEO(src_o,r_e,DaggerNo);
+	}
+	FGrid->Barrier();
+	double t1=usecond();
+	uint64_t ncall = 500;
+
+	FGrid->Broadcast(0,&ncall,sizeof(ncall));
+
+	//	std::cout << GridLogMessage << " Estimate " << ncall << " calls per second"<<std::endl;
+
+	time_statistics timestat;
+	std::vector<double> t_time(ncall);
+	for(uint64_t i=0;i<ncall;i++){
+	  t0=usecond();
+	  Dw.DhopEO(src_o,r_e,DaggerNo);
+	  t1=usecond();
+	  t_time[i] = t1-t0;
+	}
+	FGrid->Barrier();
+	
+	double volume=Ls;  for(int mu=0;mu<Nd;mu++) volume=volume*latt4[mu];
+
+	// Nc=3 gives
+	// 1344= 3*(2*8+6)*2*8 + 8*3*2*2 + 3*4*2*8
+	// 1344 = Nc* (6+(Nc-1)*8)*2*Nd + Nd*Nc*2*2  + Nd*Nc*Ns*2
+	//	double flops=(1344.0*volume)/2;
+	double fps = Nc* (6+(Nc-1)*8)*Ns*Nd + 2*Nd*Nc*Ns  + 2*Nd*Nc*Ns*2;
+
+	double flops=(fps*volume)/2;
+	double mf_hi, mf_lo, mf_err;
+
+	timestat.statistics(t_time);
+	mf_hi = flops/timestat.min;
+	mf_lo = flops/timestat.max;
+	mf_err= flops/timestat.min * timestat.err/timestat.mean;
+
+	mflops = flops/timestat.mean;
+	mflops_all.push_back(mflops);
+	if ( mflops_best == 0   ) mflops_best = mflops;
+	if ( mflops_worst== 0   ) mflops_worst= mflops;
+	if ( mflops>mflops_best ) mflops_best = mflops;
+	if ( mflops<mflops_worst) mflops_worst= mflops;
+
+	std::cout<<GridLogMessage<< "Deo FlopsPerSite is "<<fps<<std::endl;
+	std::cout<<GridLogMessage << std::fixed << std::setprecision(1)<<"Deo mflop/s =   "<< mflops << " ("<<mf_err<<") " << mf_lo<<"-"<<mf_hi <<std::endl;
+	std::cout<<GridLogMessage << std::fixed << std::setprecision(1)<<"Deo mflop/s per rank   "<< mflops/NP<<std::endl;
+	std::cout<<GridLogMessage << std::fixed << std::setprecision(1)<<"Deo mflop/s per node   "<< mflops/NN<<std::endl;
+
+      }
+
+      std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
+      std::cout<<GridLogMessage << L<<"^4 x "<<Ls<< " Deo Best  mflop/s        =   "<< mflops_best << " ; " << mflops_best/NN<<" per node " <<std::endl;
+      std::cout<<GridLogMessage << L<<"^4 x "<<Ls<< " Deo Worst mflop/s        =   "<< mflops_worst<< " ; " << mflops_worst/NN<<" per node " <<std::endl;
+      std::cout<<GridLogMessage <<fmt << std::endl;
+      std::cout<<GridLogMessage ;
+
+      for(int i=0;i<mflops_all.size();i++){
+	std::cout<<mflops_all[i]/NN<<" ; " ;
+      }
+      std::cout<<std::endl;
+      std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
+
+    }
+    return mflops_best;
+  }
+
+
+  static double Staggered(int L)
+  {
+    double mflops;
+    double mflops_best = 0;
+    double mflops_worst= 0;
+    std::vector<double> mflops_all;
+
+    ///////////////////////////////////////////////////////
+    // Set/Get the layout & grid size
+    ///////////////////////////////////////////////////////
+    int threads = GridThread::GetThreads();
+    Coordinate mpi = GridDefaultMpi(); assert(mpi.size()==4);
+    Coordinate local({L,L,L,L});
+    Coordinate latt4({local[0]*mpi[0],local[1]*mpi[1],local[2]*mpi[2],local[3]*mpi[3]});
+    
+    GridCartesian         * TmpGrid   = SpaceTimeGrid::makeFourDimGrid(latt4,
+								       GridDefaultSimd(Nd,vComplex::Nsimd()),
+								       GridDefaultMpi());
+    uint64_t NP = TmpGrid->RankCount();
+    uint64_t NN = TmpGrid->NodeCount();
+    NN_global=NN;
+    uint64_t SHM=NP/NN;
+
+
+    ///////// Welcome message ////////////
+    std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
+    std::cout<<GridLogMessage << "Benchmark ImprovedStaggered on "<<L<<"^4 local volume "<<std::endl;
+    std::cout<<GridLogMessage << "* Global volume  : "<<GridCmdVectorIntToString(latt4)<<std::endl;
+    std::cout<<GridLogMessage << "* ranks          : "<<NP  <<std::endl;
+    std::cout<<GridLogMessage << "* nodes          : "<<NN  <<std::endl;
+    std::cout<<GridLogMessage << "* ranks/node     : "<<SHM <<std::endl;
+    std::cout<<GridLogMessage << "* ranks geom     : "<<GridCmdVectorIntToString(mpi)<<std::endl;
+    std::cout<<GridLogMessage << "* Using "<<threads<<" threads"<<std::endl;
+    std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
+
+    ///////// Lattice Init ////////////
+    GridCartesian         * FGrid   = SpaceTimeGrid::makeFourDimGrid(latt4, GridDefaultSimd(Nd,vComplexF::Nsimd()),GridDefaultMpi());
+    GridRedBlackCartesian * FrbGrid = SpaceTimeGrid::makeFourDimRedBlackGrid(FGrid);
+    
+    ///////// RNG Init ////////////
+    std::vector<int> seeds4({1,2,3,4});
+    GridParallelRNG          RNG4(FGrid);  RNG4.SeedFixedIntegers(seeds4);
+    std::cout << GridLogMessage << "Initialised RNGs" << std::endl;
+
+    RealD mass=0.1;
+    RealD c1=9.0/8.0;
+    RealD c2=-1.0/24.0;
+    RealD u0=1.0;
+
+    typedef ImprovedStaggeredFermionF Action;
+    typedef typename Action::FermionField Fermion; 
+    typedef LatticeGaugeFieldF Gauge;
+    
+    Gauge Umu(FGrid);  SU<Nc>::HotConfiguration(RNG4,Umu); 
+
+    typename Action::ImplParams params;
+    Action Ds(Umu,Umu,*FGrid,*FrbGrid,mass,c1,c2,u0,params);
+
+    ///////// Source preparation ////////////
+    Fermion src   (FGrid); random(RNG4,src);
+    Fermion src_e (FrbGrid);
+    Fermion src_o (FrbGrid);
+    Fermion r_e   (FrbGrid);
+    Fermion r_o   (FrbGrid);
+    Fermion r_eo  (FGrid);
+  
+    {
+
+      pickCheckerboard(Even,src_e,src);
+      pickCheckerboard(Odd,src_o,src);
+    
+      const int num_cases = 2;
+      std::string fmt("G/S/C ; G/O/C ; G/S/S ; G/O/S ");
+      
+      controls Cases [] = {
+	{  StaggeredKernelsStatic::OptGeneric   ,  StaggeredKernelsStatic::CommsAndCompute  ,CartesianCommunicator::CommunicatorPolicyConcurrent  },
+	{  StaggeredKernelsStatic::OptHandUnroll,  StaggeredKernelsStatic::CommsAndCompute  ,CartesianCommunicator::CommunicatorPolicyConcurrent  },
+	{  StaggeredKernelsStatic::OptInlineAsm ,  StaggeredKernelsStatic::CommsAndCompute  ,CartesianCommunicator::CommunicatorPolicyConcurrent  }
+      }; 
+
+      for(int c=0;c<num_cases;c++) {
+	
+	StaggeredKernelsStatic::Comms = Cases[c].CommsOverlap;
+	StaggeredKernelsStatic::Opt   = Cases[c].Opt;
+	CartesianCommunicator::SetCommunicatorPolicy(Cases[c].CommsAsynch);
+      
+	std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
+	if ( StaggeredKernelsStatic::Opt == StaggeredKernelsStatic::OptGeneric   ) std::cout << GridLogMessage<< "* Using GENERIC Nc StaggeredKernels" <<std::endl;
+	std::cout << GridLogMessage<< "* SINGLE precision "<<std::endl;
+	std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
+	
+	int nwarm = 10;
+	double t0=usecond();
+	FGrid->Barrier();
+	for(int i=0;i<nwarm;i++){
+	  Ds.DhopEO(src_o,r_e,DaggerNo);
+	}
+	FGrid->Barrier();
+	double t1=usecond();
+	uint64_t ncall = 500;
+
+	FGrid->Broadcast(0,&ncall,sizeof(ncall));
+
+	//	std::cout << GridLogMessage << " Estimate " << ncall << " calls per second"<<std::endl;
+
+	time_statistics timestat;
+	std::vector<double> t_time(ncall);
+	for(uint64_t i=0;i<ncall;i++){
+	  t0=usecond();
+	  Ds.DhopEO(src_o,r_e,DaggerNo);
+	  t1=usecond();
+	  t_time[i] = t1-t0;
+	}
+	FGrid->Barrier();
+	
+	double volume=1;  for(int mu=0;mu<Nd;mu++) volume=volume*latt4[mu];
+	double flops=(1146.0*volume)/2;
+	double mf_hi, mf_lo, mf_err;
+	
+	timestat.statistics(t_time);
+	mf_hi = flops/timestat.min;
+	mf_lo = flops/timestat.max;
+	mf_err= flops/timestat.min * timestat.err/timestat.mean;
+
+	mflops = flops/timestat.mean;
+	mflops_all.push_back(mflops);
+	if ( mflops_best == 0   ) mflops_best = mflops;
+	if ( mflops_worst== 0   ) mflops_worst= mflops;
+	if ( mflops>mflops_best ) mflops_best = mflops;
+	if ( mflops<mflops_worst) mflops_worst= mflops;
+	
+	std::cout<<GridLogMessage << std::fixed << std::setprecision(1)<<"Deo mflop/s =   "<< mflops << " ("<<mf_err<<") " << mf_lo<<"-"<<mf_hi <<std::endl;
+	std::cout<<GridLogMessage << std::fixed << std::setprecision(1)<<"Deo mflop/s per rank   "<< mflops/NP<<std::endl;
+	std::cout<<GridLogMessage << std::fixed << std::setprecision(1)<<"Deo mflop/s per node   "<< mflops/NN<<std::endl;
+      
+      }
+
+      std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
+      std::cout<<GridLogMessage << L<<"^4  Deo Best  mflop/s        =   "<< mflops_best << " ; " << mflops_best/NN<<" per node " <<std::endl;
+      std::cout<<GridLogMessage << L<<"^4  Deo Worst mflop/s        =   "<< mflops_worst<< " ; " << mflops_worst/NN<<" per node " <<std::endl;
+      std::cout<<GridLogMessage <<fmt << std::endl;
+      std::cout<<GridLogMessage ;
+
+      for(int i=0;i<mflops_all.size();i++){
+	std::cout<<mflops_all[i]/NN<<" ; " ;
+      }
+      std::cout<<std::endl;
+    }
+    std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
+    return mflops_best;
+  }
+
+  static double Clover(int L)
+  {
+    double mflops;
+    double mflops_best = 0;
+    double mflops_worst= 0;
+    std::vector<double> mflops_all;
+
+    ///////////////////////////////////////////////////////
+    // Set/Get the layout & grid size
+    ///////////////////////////////////////////////////////
+    int threads = GridThread::GetThreads();
+    Coordinate mpi = GridDefaultMpi(); assert(mpi.size()==4);
+    Coordinate local({L,L,L,L});
+    Coordinate latt4({local[0]*mpi[0],local[1]*mpi[1],local[2]*mpi[2],local[3]*mpi[3]});
+    
+    GridCartesian         * TmpGrid   = SpaceTimeGrid::makeFourDimGrid(latt4,
+								       GridDefaultSimd(Nd,vComplex::Nsimd()),
+								       GridDefaultMpi());
+    uint64_t NP = TmpGrid->RankCount();
+    uint64_t NN = TmpGrid->NodeCount();
+    NN_global=NN;
+    uint64_t SHM=NP/NN;
+
+
+    ///////// Welcome message ////////////
+    std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
+    std::cout<<GridLogMessage << "Benchmark Clover on "<<L<<"^4 local volume "<<std::endl;
+    std::cout<<GridLogMessage << "* Global volume  : "<<GridCmdVectorIntToString(latt4)<<std::endl;
+    std::cout<<GridLogMessage << "* ranks          : "<<NP  <<std::endl;
+    std::cout<<GridLogMessage << "* nodes          : "<<NN  <<std::endl;
+    std::cout<<GridLogMessage << "* ranks/node     : "<<SHM <<std::endl;
+    std::cout<<GridLogMessage << "* ranks geom     : "<<GridCmdVectorIntToString(mpi)<<std::endl;
+    std::cout<<GridLogMessage << "* Using "<<threads<<" threads"<<std::endl;
+    std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
+
+    ///////// Lattice Init ////////////
+    GridCartesian         * FGrid   = SpaceTimeGrid::makeFourDimGrid(latt4, GridDefaultSimd(Nd,vComplexF::Nsimd()),GridDefaultMpi());
+    GridRedBlackCartesian * FrbGrid = SpaceTimeGrid::makeFourDimRedBlackGrid(FGrid);
+    
+    ///////// RNG Init ////////////
+    std::vector<int> seeds4({1,2,3,4});
+    GridParallelRNG          RNG4(FGrid);  RNG4.SeedFixedIntegers(seeds4);
+    std::cout << GridLogMessage << "Initialised RNGs" << std::endl;
+
+    RealD mass=0.1;
+    RealD csw=1.0;
+
+    typedef WilsonCloverFermionF Action;
+    typedef typename Action::FermionField Fermion; 
+    typedef LatticeGaugeFieldF Gauge;
+    
+    Gauge Umu(FGrid);  SU<Nc>::HotConfiguration(RNG4,Umu); 
+
+    Action Dc(Umu,*FGrid,*FrbGrid,mass,csw,csw);
+
+    ///////// Source preparation ////////////
+    Fermion src   (FGrid); random(RNG4,src);
+    Fermion r     (FGrid);
+  
+    {
+
+      const int num_cases = 1;
+      std::string fmt("G/S/C ; G/O/C ; G/S/S ; G/O/S ");
+      
+      controls Cases [] = {
+	{  WilsonKernelsStatic::OptGeneric   ,  WilsonKernelsStatic::CommsAndCompute  ,CartesianCommunicator::CommunicatorPolicyConcurrent  },
+      }; 
+
+      for(int c=0;c<num_cases;c++) {
+	
+	WilsonKernelsStatic::Comms = Cases[c].CommsOverlap;
+	WilsonKernelsStatic::Opt   = Cases[c].Opt;
+	CartesianCommunicator::SetCommunicatorPolicy(Cases[c].CommsAsynch);
+      
+	std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
+	std::cout << GridLogMessage<< "* SINGLE precision "<<std::endl;
+	std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
+	
+	int nwarm = 10;
+	double t0=usecond();
+	FGrid->Barrier();
+	for(int i=0;i<nwarm;i++){
+	  Dc.M(src,r);
+	}
+	FGrid->Barrier();
+	double t1=usecond();
+	uint64_t ncall = 500;
+
+	FGrid->Broadcast(0,&ncall,sizeof(ncall));
+
+	//	std::cout << GridLogMessage << " Estimate " << ncall << " calls per second"<<std::endl;
+
+	time_statistics timestat;
+	std::vector<double> t_time(ncall);
+	for(uint64_t i=0;i<ncall;i++){
+	  t0=usecond();
+	  Dc.M(src,r);
+	  t1=usecond();
+	  t_time[i] = t1-t0;
+	}
+	FGrid->Barrier();
+	
+	double volume=1;  for(int mu=0;mu<Nd;mu++) volume=volume*latt4[mu];
+	double flops=(1344+ 24+6*6*8*2)*volume;
+	double mf_hi, mf_lo, mf_err;
+	
+	timestat.statistics(t_time);
+	mf_hi = flops/timestat.min;
+	mf_lo = flops/timestat.max;
+	mf_err= flops/timestat.min * timestat.err/timestat.mean;
+
+	mflops = flops/timestat.mean;
+	mflops_all.push_back(mflops);
+	if ( mflops_best == 0   ) mflops_best = mflops;
+	if ( mflops_worst== 0   ) mflops_worst= mflops;
+	if ( mflops>mflops_best ) mflops_best = mflops;
+	if ( mflops<mflops_worst) mflops_worst= mflops;
+	
+	std::cout<<GridLogMessage << std::fixed << std::setprecision(1)<<"Dclov mflop/s =   "<< mflops << " ("<<mf_err<<") " << mf_lo<<"-"<<mf_hi <<std::endl;
+	std::cout<<GridLogMessage << std::fixed << std::setprecision(1)<<"Dclov mflop/s per rank   "<< mflops/NP<<std::endl;
+	std::cout<<GridLogMessage << std::fixed << std::setprecision(1)<<"Dclov mflop/s per node   "<< mflops/NN<<std::endl;
+      
+      }
+
+      std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
+      std::cout<<GridLogMessage << L<<"^4  Deo Best  mflop/s        =   "<< mflops_best << " ; " << mflops_best/NN<<" per node " <<std::endl;
+      std::cout<<GridLogMessage << L<<"^4  Deo Worst mflop/s        =   "<< mflops_worst<< " ; " << mflops_worst/NN<<" per node " <<std::endl;
+      std::cout<<GridLogMessage <<fmt << std::endl;
+      std::cout<<GridLogMessage ;
+
+      for(int i=0;i<mflops_all.size();i++){
+	std::cout<<mflops_all[i]/NN<<" ; " ;
+      }
+      std::cout<<std::endl;
+    }
+    std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
+    return mflops_best;
+  }
+};
+
+
+
+
+int main (int argc, char ** argv)
+{
+  Grid_init(&argc,&argv);
+
+  if (GlobalSharedMemory::WorldRank==0) { 
+    FP = fopen("Benchmark_usqcd.csv","w");
+  } else {
+    FP = fopen("/dev/null","w");
+  }
+
+  CartesianCommunicator::SetCommunicatorPolicy(CartesianCommunicator::CommunicatorPolicySequential);
+  LebesgueOrder::Block = std::vector<int>({2,2,2,2});
+
+  Benchmark::Decomposition();
+
+  int do_su4=0;
+  int do_memory=1;
+  int do_comms =1;
+  int do_blas  =1;
+
+  int sel=4;
+  std::vector<int> L_list({8,12,16,24,32});
+  int selm1=sel-1;
+
+  std::vector<double> clover;
+  std::vector<double> dwf4;
+  std::vector<double> staggered;
+
+  int Ls=1;
+  std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
+  std::cout<<GridLogMessage << " Clover dslash 4D vectorised (temporarily Wilson)" <<std::endl;
+  std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
+  for(int l=0;l<L_list.size();l++){
+    clover.push_back(Benchmark::DWF(1,L_list[l]));
+  }
+
+  Ls=12;
+  std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
+  std::cout<<GridLogMessage << " Domain wall dslash 4D vectorised" <<std::endl;
+  std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
+  for(int l=0;l<L_list.size();l++){
+    double result = Benchmark::DWF(Ls,L_list[l]) ;
+    dwf4.push_back(result);
+  }
+
+  std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
+  std::cout<<GridLogMessage << " Improved Staggered dslash 4D vectorised" <<std::endl;
+  std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
+  for(int l=0;l<L_list.size();l++){
+    double result = Benchmark::Staggered(L_list[l]) ;
+    staggered.push_back(result);
+  }
+
+  std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
+  std::cout<<GridLogMessage << " Summary table Ls="<<Ls <<std::endl;
+  std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
+  std::cout<<GridLogMessage << "L \t\t Clover \t\t DWF4 \t\t Staggered" <<std::endl;
+  for(int l=0;l<L_list.size();l++){
+    std::cout<<GridLogMessage << L_list[l] <<" \t\t "<< clover[l]<<" \t\t "<<dwf4[l] << " \t\t "<< staggered[l]<<std::endl;
+  }
+  std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
+
+  int NN=NN_global;
+  if ( do_memory ) {
+    std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
+    std::cout<<GridLogMessage << " Memory benchmark " <<std::endl;
+    std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
+    Benchmark::Memory();
+  }
+
+  if ( do_blas ) {
+#if defined(GRID_CUDA) || defined(GRID_HIP)     || defined(GRID_SYCL)   
+    std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
+    std::cout<<GridLogMessage << " Batched BLAS benchmark " <<std::endl;
+    std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
+    Benchmark::BLAS();
+#endif
+  }
+
+  if ( do_su4 ) {
+    std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
+    std::cout<<GridLogMessage << " SU(4) benchmark " <<std::endl;
+    std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
+    Benchmark::SU4();
+  }
+  
+  if ( do_comms ) {
+    std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
+    std::cout<<GridLogMessage << " Communications benchmark " <<std::endl;
+    std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
+    Benchmark::Comms();
+  }
+
+    std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
+    std::cout<<GridLogMessage << " Per Node Summary table Ls="<<Ls <<std::endl;
+    std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
+    std::cout<<GridLogMessage << " L \t\t Clover\t\t DWF4\t\t Staggered (GF/s per node)" <<std::endl;
+    fprintf(FP,"Per node summary table\n");
+    fprintf(FP,"\n");
+    fprintf(FP,"L , Wilson, DWF4, Staggered, GF/s per node\n");
+    fprintf(FP,"\n");
+    for(int l=0;l<L_list.size();l++){
+      std::cout<<GridLogMessage << L_list[l] <<" \t\t "<< clover[l]/NN<<" \t "<<dwf4[l]/NN<< " \t "<<staggered[l]/NN<<std::endl;
+      fprintf(FP,"%d , %.0f, %.0f, %.0f\n",L_list[l],clover[l]/NN/1000.,dwf4[l]/NN/1000.,staggered[l]/NN/1000.);
+    }
+    fprintf(FP,"\n");
+
+    std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
+
+    std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
+    std::cout<<GridLogMessage << " Comparison point     result: "  << 0.5*(dwf4[sel]+dwf4[selm1])/NN << " Mflop/s per node"<<std::endl;
+    std::cout<<GridLogMessage << " Comparison point is 0.5*("<<dwf4[sel]/NN<<"+"<<dwf4[selm1]/NN << ") "<<std::endl;
+    std::cout<<std::setprecision(3);
+    std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
+
+  Grid_finalize();
+  fclose(FP);
+}

bootstrap.sh

@@ -1,12 +1,12 @@
 #!/usr/bin/env bash
 set -e
 
-EIGEN_URL='https://gitlab.com/libeigen/eigen/-/archive/3.3.7/eigen-3.3.7.tar.bz2'
-EIGEN_SHA256SUM='685adf14bd8e9c015b78097c1dc22f2f01343756f196acdc76a678e1ae352e11'
+EIGEN_URL='https://gitlab.com/libeigen/eigen/-/archive/3.4.0/eigen-3.4.0.tar.bz2'
+EIGEN_SHA256SUM='b4c198460eba6f28d34894e3a5710998818515104d6e74e5cc331ce31e46e626'
 
 
 echo "-- deploying Eigen source..."
-ARC=`basename ${EIGEN_URL}`
+ARC=$(basename ${EIGEN_URL})
 wget ${EIGEN_URL} --no-check-certificate
 if command -v sha256sum; then
    echo "$EIGEN_SHA256SUM  $(basename "$EIGEN_URL")" \
@@ -14,13 +14,8 @@ if command -v sha256sum; then
 else
    echo "WARNING: could not verify checksum, please install sha256sum" >&2
 fi
-./scripts/update_eigen.sh ${ARC}
-rm ${ARC}
-# patch for non-portable includes in Eigen 3.3.5
-# apparently already fixed in Eigen HEAD so it should not be 
-# a problem in the future (A.P.)
-patch Eigen/unsupported/Eigen/CXX11/Tensor scripts/eigen-3.3.5.Tensor.patch
-
+./scripts/update_eigen.sh "${ARC}"
+rm "${ARC}"
 echo '-- generating Make.inc files...'
 ./scripts/filelist
 echo '-- generating configure script...'

scripts/eigen-3.3.5.Tensor.patch

@@ -1,19 +0,0 @@
---- ./Eigen/unsupported/Eigen/CXX11/Tensor	2018-07-23 10:33:42.000000000 +0100
-+++ Tensor	2018-08-28 16:15:56.000000000 +0100
-@@ -25,7 +25,7 @@
- #include <utility>
- #endif
- 
--#include <Eigen/src/Core/util/DisableStupidWarnings.h>
-+#include "../../../Eigen/src/Core/util/DisableStupidWarnings.h"
- 
- #include "../SpecialFunctions"
- #include "src/util/CXX11Meta.h"
-@@ -147,6 +147,6 @@
- 
- #include "src/Tensor/TensorIO.h"
- 
--#include <Eigen/src/Core/util/ReenableStupidWarnings.h>
-+#include "../../../Eigen/src/Core/util/ReenableStupidWarnings.h"
- 
- //#endif // EIGEN_CXX11_TENSOR_MODULE

systems/Aurora/benchmarks/bench1024.pbs

@@ -25,12 +25,16 @@ export MPIR_CVAR_CH4_OFI_ENABLE_GPU_PIPELINE=1
 
 export MPIR_CVAR_CH4_OFI_GPU_PIPELINE_D2H_ENGINE_TYPE=0
 export MPIR_CVAR_CH4_OFI_GPU_PIPELINE_H2D_ENGINE_TYPE=0
-export MPIR_CVAR_GPU_USE_IMMEDIATE_COMMAND_LIST=1
+#export MPIR_CVAR_GPU_USE_IMMEDIATE_COMMAND_LIST=1
 export MPIR_CVAR_CH4_OFI_GPU_PIPELINE_BUFFER_SZ=1048576
 export MPIR_CVAR_CH4_OFI_GPU_PIPELINE_THRESHOLD=131072
 export MPIR_CVAR_CH4_OFI_GPU_PIPELINE_NUM_BUFFERS_PER_CHUNK=16
 export MPIR_CVAR_CH4_OFI_GPU_PIPELINE_MAX_NUM_BUFFERS=16
 export MPICH_OFI_NIC_POLICY=GPU
+export FI_CXI_CQ_FILL_PERCENT=10
+export FI_CXI_DEFAULT_CQ_SIZE=262144
+#export FI_CXI_DEFAULT_CQ_SIZE=131072
+#export FI_CXI_CQ_FILL_PERCENT=20
 
 # 12 ppn, 32 nodes, 384 ranks
 #
@@ -45,12 +49,12 @@ CMD="mpiexec -np 12288 -ppn 12  -envall \
 	     ./gpu_tile_compact.sh \
 	     ./Benchmark_dwf_fp32 --mpi 8.8.8.24 --grid 128.128.128.384 \
 		--shm-mpi 1 --shm 2048 --device-mem 32000 --accelerator-threads 32 --comms-overlap"
-$CMD | tee 1024node.dwf.small
+$CMD | tee 1024node.dwf.small.cq
 
 CMD="mpiexec -np 12288 -ppn 12  -envall \
 	     ./gpu_tile_compact.sh \
 	     ./Benchmark_dwf_fp32 --mpi 16.8.8.12 --grid 256.256.256.384 \
 		--shm-mpi 1 --shm 2048 --device-mem 32000 --accelerator-threads 32 --comms-overlap"
-$CMD | tee 1024node.dwf
+$CMD | tee 1024node.dwf.cq
 
 

systems/Aurora/benchmarks/bench12.pbs

@@ -17,6 +17,7 @@ source ../sourceme.sh
 export OMP_NUM_THREADS=3
 export MPIR_CVAR_CH4_OFI_ENABLE_GPU_PIPELINE=1
 
+
 #unset MPIR_CVAR_CH4_OFI_GPU_PIPELINE_D2H_ENGINE_TYPE
 #unset MPIR_CVAR_CH4_OFI_GPU_PIPELINE_H2D_ENGINE_TYPE
 #unset MPIR_CVAR_GPU_USE_IMMEDIATE_COMMAND_LIST
@@ -35,11 +36,25 @@ CMD="mpiexec -np 24 -ppn 12  -envall \
 	     ./Benchmark_comms_host_device --mpi 2.3.2.2 --grid 32.24.32.192 \
 		--shm-mpi 1 --shm 2048 --device-mem 32000 --accelerator-threads 32"
 
-$CMD 
+#$CMD 
 
 CMD="mpiexec -np 24 -ppn 12  -envall \
 	     ./gpu_tile_compact.sh \
 	     ./Benchmark_dwf_fp32 --mpi 2.3.2.2 --grid 64.96.64.64 --comms-overlap \
 		--shm-mpi 1 --shm 2048 --device-mem 32000 --accelerator-threads 32"
 
+#$CMD 
+
+CMD="mpiexec -np 1 -ppn 1  -envall \
+	     ./gpu_tile_compact.sh \
+	     ./Benchmark_dwf --mpi 1.1.1.1 --grid 16.32.32.32 --comms-sequential \
+		--shm-mpi 1 --shm 2048 --device-mem 32000 --accelerator-threads 32"
+
+$CMD 
+
+CMD="mpiexec -np 1 -ppn 1  -envall \
+	     ./gpu_tile_compact.sh \
+	     ./Benchmark_dwf_fp32 --mpi 1.1.1.1 --grid 16.32.32.32 --comms-sequential \
+		--shm-mpi 1 --shm 2048 --device-mem 32000 --accelerator-threads 32"
+
 $CMD 

systems/Aurora/config-command

@@ -11,6 +11,6 @@ TOOLS=$HOME/tools
 	--enable-unified=no \
 	MPICXX=mpicxx \
 	CXX=icpx \
-	LDFLAGS="-fiopenmp -fsycl -fsycl-device-code-split=per_kernel -fsycl-device-lib=all -lze_loader -L$TOOLS/lib64/" \
-	CXXFLAGS="-fiopenmp -fsycl-unnamed-lambda -fsycl -I$INSTALL/include -Wno-tautological-compare -I$HOME/ -I$TOOLS/include"
+	LDFLAGS="-fiopenmp -fsycl -fsycl-device-code-split=per_kernel -fsycl-device-lib=all -lze_loader -L$TOOLS/lib64/ -L${MKLROOT}/lib -qmkl=parallel " \
+	CXXFLAGS="-fiopenmp -fsycl-unnamed-lambda -fsycl -I$INSTALL/include -Wno-tautological-compare -I$HOME/ -I$TOOLS/include -qmkl=parallel"
 

systems/Aurora/sourceme.sh

@@ -3,10 +3,24 @@
 module use /soft/modulefiles
 module load intel_compute_runtime/release/agama-devel-682.22
 
+export FI_CXI_DEFAULT_CQ_SIZE=131072
+export FI_CXI_CQ_FILL_PERCENT=20
+
+export SYCL_PROGRAM_COMPILE_OPTIONS="-ze-opt-large-register-file"
+#export SYCL_PROGRAM_COMPILE_OPTIONS="-ze-intel-enable-auto-large-GRF-mode"
+
+#
+# -ftarget-register-alloc-mode=pvc:default 
+# -ftarget-register-alloc-mode=pvc:small
+# -ftarget-register-alloc-mode=pvc:large
+# -ftarget-register-alloc-mode=pvc:auto
+#
+
 export HTTP_PROXY=http://proxy.alcf.anl.gov:3128
 export HTTPS_PROXY=http://proxy.alcf.anl.gov:3128
 export http_proxy=http://proxy.alcf.anl.gov:3128
 export https_proxy=http://proxy.alcf.anl.gov:3128
 #export MPIR_CVAR_CH4_OFI_ENABLE_HMEM=1
 git config --global http.proxy http://proxy.alcf.anl.gov:3128
- 
+
+export SYCL_PROGRAM_COMPILE_OPTIONS="-ze-opt-large-register-file"

systems/Aurora/tests/repro16.pbs

@@ -0,0 +1,40 @@
+#!/bin/bash
+
+## qsub -q EarlyAppAccess -A Aurora_Deployment -I -l select=1 -l walltime=60:00
+
+#PBS -q EarlyAppAccess
+#PBS -l select=16
+#PBS -l walltime=01:00:00
+#PBS -A LatticeQCD_aesp_CNDA
+
+#export OMP_PROC_BIND=spread
+#unset OMP_PLACES
+
+cd $PBS_O_WORKDIR
+
+source ../sourceme.sh
+
+cat $PBS_NODEFILE
+
+export OMP_NUM_THREADS=3
+export MPIR_CVAR_CH4_OFI_ENABLE_GPU_PIPELINE=1
+
+#unset MPIR_CVAR_CH4_OFI_GPU_PIPELINE_D2H_ENGINE_TYPE
+#unset MPIR_CVAR_CH4_OFI_GPU_PIPELINE_H2D_ENGINE_TYPE
+#unset MPIR_CVAR_GPU_USE_IMMEDIATE_COMMAND_LIST
+
+export MPIR_CVAR_CH4_OFI_GPU_PIPELINE_D2H_ENGINE_TYPE=0
+export MPIR_CVAR_CH4_OFI_GPU_PIPELINE_H2D_ENGINE_TYPE=0
+export MPIR_CVAR_GPU_USE_IMMEDIATE_COMMAND_LIST=1
+export MPIR_CVAR_CH4_OFI_GPU_PIPELINE_BUFFER_SZ=1048576
+export MPIR_CVAR_CH4_OFI_GPU_PIPELINE_THRESHOLD=131072
+export MPIR_CVAR_CH4_OFI_GPU_PIPELINE_NUM_BUFFERS_PER_CHUNK=16
+export MPIR_CVAR_CH4_OFI_GPU_PIPELINE_MAX_NUM_BUFFERS=16
+export MPICH_OFI_NIC_POLICY=GPU
+
+# 12 ppn, 16 nodes, 192 ranks
+CMD="mpiexec -np 192 -ppn 12  -envall \
+	     ./gpu_tile_compact.sh \
+	     ./Test_dwf_mixedcg_prec --mpi 2.4.4.6 --grid 64.128.128.192 \
+		--shm-mpi 1 --shm 4096 --device-mem 32000 --accelerator-threads 32 --seconds 3000"
+$CMD 

systems/Aurora/tests/solver/stag16.pbs

@@ -0,0 +1,40 @@
+#!/bin/bash
+
+## qsub -q EarlyAppAccess -A Aurora_Deployment -I -l select=1 -l walltime=60:00
+
+#PBS -q EarlyAppAccess
+#PBS -l select=16
+#PBS -l walltime=01:00:00
+#PBS -A LatticeQCD_aesp_CNDA
+
+#export OMP_PROC_BIND=spread
+#unset OMP_PLACES
+
+cd $PBS_O_WORKDIR
+
+source ../../sourceme.sh
+
+cat $PBS_NODEFILE
+
+export OMP_NUM_THREADS=3
+export MPIR_CVAR_CH4_OFI_ENABLE_GPU_PIPELINE=1
+
+#unset MPIR_CVAR_CH4_OFI_GPU_PIPELINE_D2H_ENGINE_TYPE
+#unset MPIR_CVAR_CH4_OFI_GPU_PIPELINE_H2D_ENGINE_TYPE
+#unset MPIR_CVAR_GPU_USE_IMMEDIATE_COMMAND_LIST
+
+export MPIR_CVAR_CH4_OFI_GPU_PIPELINE_D2H_ENGINE_TYPE=0
+export MPIR_CVAR_CH4_OFI_GPU_PIPELINE_H2D_ENGINE_TYPE=0
+export MPIR_CVAR_GPU_USE_IMMEDIATE_COMMAND_LIST=1
+export MPIR_CVAR_CH4_OFI_GPU_PIPELINE_BUFFER_SZ=1048576
+export MPIR_CVAR_CH4_OFI_GPU_PIPELINE_THRESHOLD=131072
+export MPIR_CVAR_CH4_OFI_GPU_PIPELINE_NUM_BUFFERS_PER_CHUNK=16
+export MPIR_CVAR_CH4_OFI_GPU_PIPELINE_MAX_NUM_BUFFERS=16
+export MPICH_OFI_NIC_POLICY=GPU
+
+# 12 ppn, 16 nodes, 192 ranks
+CMD="mpiexec -np 192 -ppn 12  -envall \
+	     ./gpu_tile_compact.sh \
+	     ./Test_staggered_cg_prec --mpi 2.4.4.6 --grid 128.128.128.192 \
+	     --shm-mpi 1 --shm 4096 --device-mem 32000 --accelerator-threads 32 --seconds 3000"
+$CMD 

systems/Booster/benchmarks/Benchmark_usqcd.csv

@@ -0,0 +1,70 @@
+Memory Bandwidth
+
+Bytes, GB/s per node
+3145728, 225.900365
+50331648, 2858.859504
+254803968, 4145.556367
+805306368, 4905.772480
+1966080000, 4978.312557
+
+
+GEMM
+
+ M, N, K, BATCH, GF/s per rank
+16, 8, 16, 256, 1.713639
+16, 16, 16, 256, 288.268316
+16, 32, 16, 256, 597.053950
+32, 8, 32, 256, 557.382591
+32, 16, 32, 256, 1100.145311
+32, 32, 32, 256, 1885.080449
+64, 8, 64, 256, 1725.163599
+64, 16, 64, 256, 3389.336566
+64, 32, 64, 256, 4168.252422
+16, 8, 256, 256, 1326.262134
+16, 16, 256, 256, 2318.095475
+16, 32, 256, 256, 3555.436503
+32, 8, 256, 256, 1920.139170
+32, 16, 256, 256, 3486.174753
+32, 32, 256, 256, 5320.821724
+64, 8, 256, 256, 2539.597502
+64, 16, 256, 256, 5003.456775
+64, 32, 256, 256, 7837.531562
+8, 256, 16, 256, 1427.848170
+16, 256, 16, 256, 2222.147815
+32, 256, 16, 256, 2877.121715
+8, 256, 32, 256, 1922.890086
+16, 256, 32, 256, 3199.469082
+32, 256, 32, 256, 4845.405343
+8, 256, 64, 256, 2639.483343
+16, 256, 64, 256, 5012.800299
+32, 256, 64, 256, 7216.006882
+
+
+
+Communications
+
+Packet bytes, direction, GB/s per node
+4718592, 2, 206.570734
+4718592, 3, 207.501847
+4718592, 6, 189.730277
+4718592, 7, 204.301218
+15925248, 2, 307.882997
+15925248, 3, 287.901076
+15925248, 6, 295.603109
+15925248, 7, 300.682033
+37748736, 2, 331.740364
+37748736, 3, 338.610627
+37748736, 6, 332.580657
+37748736, 7, 336.336579
+
+
+Per node summary table
+
+L , Wilson, DWF4, Staggered, GF/s per node
+
+8 , 16, 1165, 10
+12 , 473, 4901, 163
+16 , 1436, 8464, 442
+24 , 4133, 10139, 1530
+32 , 5726, 11487, 2518
+

systems/Booster/config-command

@@ -5,10 +5,12 @@ LIME=/p/home/jusers/boyle2/juwels/gm2dwf/boyle/
     --enable-gen-simd-width=64 \
     --enable-shm=nvlink \
     --enable-accelerator=cuda \
+    --disable-gparity \
+    --disable-fermion-reps \
     --with-lime=$LIME \
-    --disable-accelerator-cshift \
+    --enable-accelerator-cshift \
     --disable-unified \
     CXX=nvcc \
     LDFLAGS="-cudart shared " \
-    CXXFLAGS="-ccbin mpicxx -gencode arch=compute_80,code=sm_80 -std=c++14 -cudart shared"
+    CXXFLAGS="-ccbin mpicxx -gencode arch=compute_80,code=sm_80 -std=c++17 -cudart shared -lcublas"
 

systems/Booster/sourceme.sh

@@ -1,5 +1,5 @@
-module load GCC/9.3.0       
-module load  GMP/6.2.0   
-module load MPFR/4.1.0     
-module load OpenMPI/4.1.0rc1  
-module load CUDA/11.3
+module load GCC
+module load GMP
+module load MPFR
+module load OpenMPI
+module load CUDA

systems/Frontier/config-command

@@ -16,7 +16,7 @@ CLIME=`spack find --paths c-lime@2-3-9 | grep c-lime| cut -c 15-`
 --disable-fermion-reps \
 CXX=hipcc MPICXX=mpicxx \
 CXXFLAGS="-fPIC -I{$ROCM_PATH}/include/ -I${MPICH_DIR}/include -L/lib64 -fgpu-sanitize" \
- LDFLAGS="-L/lib64 -L${MPICH_DIR}/lib -lmpi -L${CRAY_MPICH_ROOTDIR}/gtl/lib -lmpi_gtl_hsa -lamdhip64 "
+ LDFLAGS="-L/lib64 -L${MPICH_DIR}/lib -lmpi -L${CRAY_MPICH_ROOTDIR}/gtl/lib -lmpi_gtl_hsa -lamdhip64  -lhipblas -lrocblas"
 
 
 

systems/PVC-OEM/setup.sh

@@ -1,3 +1,5 @@
 export https_proxy=http://proxy-chain.intel.com:911
 module load intel-release
 module load intel/mpich
+export MPIR_CVAR_CH4_OFI_ENABLE_GPU_PIPELINE=1
+export SYCL_PROGRAM_COMPILE_OPTIONS="-ze-opt-large-register-file"

systems/mac-arm/config-command-mpi

@@ -1,4 +1,3 @@
-BREW=/opt/local/
-MPICXX=mpicxx ../../configure --enable-simd=GEN --enable-comms=mpi-auto --enable-unified=yes --prefix $HOME/QCD/GridInstall --with-lime=/Users/peterboyle/QCD/SciDAC/install/ --with-openssl=$BREW --disable-fermion-reps --disable-gparity --disable-debug
+CXXFLAGS=-I/opt/local/include LDFLAGS=-L/opt/local/lib/ CXX=c++-13 MPICXX=mpicxx ../../configure --enable-simd=GEN --enable-comms=mpi-auto --enable-unified=yes --prefix $HOME/QCD/GridInstall --with-lime=/Users/peterboyle/QCD/SciDAC/install/ --with-openssl=$BREW --disable-fermion-reps --disable-gparity --disable-debug 
 
 

tests/Test_dwf_mixedcg_prec.cc

@@ -30,27 +30,20 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 using namespace std;
 using namespace Grid;
 
-template<class d>
-struct scal {
-  d internal;
-};
-
-  Gamma::Algebra Gmu [] = {
-    Gamma::Algebra::GammaX,
-    Gamma::Algebra::GammaY,
-    Gamma::Algebra::GammaZ,
-    Gamma::Algebra::GammaT
-  };
+#ifndef HOST_NAME_MAX
+#define HOST_NAME_MAX _POSIX_HOST_NAME_MAX
+#endif
 
 int main (int argc, char ** argv)
 {
+  char hostname[HOST_NAME_MAX+1];
+  gethostname(hostname, HOST_NAME_MAX+1);
+  std::string host(hostname);
+  
   Grid_init(&argc,&argv);
 
   const int Ls=12;
 
-  std::cout << GridLogMessage << "::::: NB: to enable a quick bit reproducibility check use the --checksums flag. " << std::endl;
-
-  { 
   GridCartesian         * UGrid   = SpaceTimeGrid::makeFourDimGrid(GridDefaultLatt(), GridDefaultSimd(Nd,vComplexD::Nsimd()),GridDefaultMpi());
   GridRedBlackCartesian * UrbGrid = SpaceTimeGrid::makeFourDimRedBlackGrid(UGrid);
   GridCartesian         * FGrid   = SpaceTimeGrid::makeFiveDimGrid(Ls,UGrid);
@@ -92,7 +85,14 @@ int main (int argc, char ** argv)
   SchurDiagMooeeOperator<DomainWallFermionD,LatticeFermionD> HermOpEO(Ddwf);
   SchurDiagMooeeOperator<DomainWallFermionF,LatticeFermionF> HermOpEO_f(Ddwf_f);
 
-  std::cout << GridLogMessage << "::::::::::::: Starting mixed CG" << std::endl;
+  int nsecs=600;
+  if( GridCmdOptionExists(argv,argv+argc,"--seconds") ){
+    std::string arg = GridCmdOptionPayload(argv,argv+argc,"--seconds");
+    GridCmdOptionInt(arg,nsecs);
+  }
+  
+  std::cout << GridLogMessage << "::::::::::::: Starting mixed CG for "<<nsecs <<" seconds" << std::endl;
+
   MixedPrecisionConjugateGradient<LatticeFermionD,LatticeFermionF> mCG(1.0e-8, 10000, 50, FrbGrid_f, HermOpEO_f, HermOpEO);
   double t1,t2,flops;
   double MdagMsiteflops = 1452; // Mobius (real coeffs)
@@ -101,7 +101,14 @@ int main (int argc, char ** argv)
   std:: cout << " MdagM site flops = "<< 4*MdagMsiteflops<<std::endl;
   std:: cout << " CG    site flops = "<< CGsiteflops <<std::endl;
   int iters;
-  for(int i=0;i<10;i++){
+
+  time_t start = time(NULL);
+
+  uint32_t csum, csumref;
+  csumref=0;
+  int iter=0;
+  do {
+    std::cerr << "******************* SINGLE PRECISION SOLVE "<<iter<<std::endl;
     result_o = Zero();
     t1=usecond();
     mCG(src_o,result_o);
@@ -111,10 +118,28 @@ int main (int argc, char ** argv)
     flops+= CGsiteflops*FrbGrid->gSites()*iters;
     std::cout << " SinglePrecision iterations/sec "<< iters/(t2-t1)*1000.*1000.<<std::endl;
     std::cout << " SinglePrecision GF/s "<< flops/(t2-t1)/1000.<<std::endl;
-  }
-  std::cout << GridLogMessage << "::::::::::::: Starting regular CG" << std::endl;
+
+    csum = crc(result_o);
+
+    if ( csumref == 0 ) {
+      csumref = csum;
+    } else {
+      if ( csum != csumref ) { 
+	std::cerr << host<<" FAILURE " <<iter <<" csum "<<std::hex<<csum<< " != "<<csumref <<std::dec<<std::endl;
+	assert(0);
+      } else {
+	std::cout << host <<" OK " <<iter <<" csum "<<std::hex<<csum<<std::dec<<" -- OK! "<<std::endl;
+      }
+    }
+    iter ++;
+  } while (time(NULL) < (start + nsecs/2) );
+    
+  std::cout << GridLogMessage << "::::::::::::: Starting double precision CG" << std::endl;
   ConjugateGradient<LatticeFermionD> CG(1.0e-8,10000);
-  for(int i=0;i<1;i++){
+  csumref=0;
+  int i=0;
+  do { 
+    std::cerr << "******************* DOUBLE PRECISION SOLVE "<<i<<std::endl;
     result_o_2 = Zero();
     t1=usecond();
     CG(HermOpEO,src_o,result_o_2);
@@ -122,46 +147,30 @@ int main (int argc, char ** argv)
     iters = CG.IterationsToComplete;
     flops = MdagMsiteflops*4*FrbGrid->gSites()*iters; 
     flops+= CGsiteflops*FrbGrid->gSites()*iters;
-    
+
     std::cout << " DoublePrecision iterations/sec "<< iters/(t2-t1)*1000.*1000.<<std::endl;
     std::cout << " DoublePrecision GF/s "<< flops/(t2-t1)/1000.<<std::endl;
-  }
-  
-  //  MemoryManager::Print();
+
+    csum = crc(result_o);
+
+    if ( csumref == 0 ) {
+      csumref = csum;
+    } else {
+      if ( csum != csumref ) { 
+	std::cerr << i <<" csum "<<std::hex<<csum<< " != "<<csumref <<std::dec<<std::endl;
+	assert(0);
+      } else {
+	std::cout << i <<" csum "<<std::hex<<csum<<std::dec<<" -- OK! "<<std::endl;
+      }
+    }
+    i++;
+  } while (time(NULL) < (start + nsecs) );
 
   LatticeFermionD diff_o(FrbGrid);
   RealD diff = axpy_norm(diff_o, -1.0, result_o, result_o_2);
 
   std::cout << GridLogMessage << "::::::::::::: Diff between mixed and regular CG: " << diff << std::endl;
-
-  #ifdef HAVE_LIME
-  if( GridCmdOptionExists(argv,argv+argc,"--checksums") ){
+  assert(diff < 1e-4);
   
-  std::string file1("./Propagator1");
-  emptyUserRecord record;
-  uint32_t nersc_csum;
-  uint32_t scidac_csuma;
-  uint32_t scidac_csumb;
-  typedef SpinColourVectorD   FermionD;
-  typedef vSpinColourVectorD vFermionD;
-
-  BinarySimpleMunger<FermionD,FermionD> munge;
-  std::string format = getFormatString<vFermionD>();
-  
-  BinaryIO::writeLatticeObject<vFermionD,FermionD>(result_o,file1,munge, 0, format,
-						   nersc_csum,scidac_csuma,scidac_csumb);
-
-  std::cout << GridLogMessage << " Mixed checksums "<<std::hex << scidac_csuma << " "<<scidac_csumb<<std::endl;
-
-  BinaryIO::writeLatticeObject<vFermionD,FermionD>(result_o_2,file1,munge, 0, format,
-						   nersc_csum,scidac_csuma,scidac_csumb);
-
-  std::cout << GridLogMessage << " CG checksums "<<std::hex << scidac_csuma << " "<<scidac_csumb<<std::endl;
-  }
-  #endif
-  }
-  
-  MemoryManager::Print();
-
   Grid_finalize();
 }

tests/core/Test_sliceSum.cc

@@ -0,0 +1,321 @@
+#include <Grid/Grid.h>
+
+template<class vobj> inline void sliceSumCPU(const Grid::Lattice<vobj> &Data,std::vector<typename vobj::scalar_object> &result,int orthogdim)
+{
+  using namespace Grid;
+  ///////////////////////////////////////////////////////
+  // FIXME precision promoted summation
+  // may be important for correlation functions
+  // But easily avoided by using double precision fields
+  ///////////////////////////////////////////////////////
+  typedef typename vobj::scalar_object sobj;
+  typedef typename vobj::scalar_object::scalar_type scalar_type;
+  GridBase  *grid = Data.Grid();
+  assert(grid!=NULL);
+
+  const int    Nd = grid->_ndimension;
+  const int Nsimd = grid->Nsimd();
+
+  assert(orthogdim >= 0);
+  assert(orthogdim < Nd);
+
+  int fd=grid->_fdimensions[orthogdim];
+  int ld=grid->_ldimensions[orthogdim];
+  int rd=grid->_rdimensions[orthogdim];
+
+  Vector<vobj> lvSum(rd); // will locally sum vectors first
+  Vector<sobj> lsSum(ld,Zero());                    // sum across these down to scalars
+  ExtractBuffer<sobj> extracted(Nsimd);                  // splitting the SIMD
+
+  result.resize(fd); // And then global sum to return the same vector to every node 
+  for(int r=0;r<rd;r++){
+    lvSum[r]=Zero();
+  }
+
+  int e1=    grid->_slice_nblock[orthogdim];
+  int e2=    grid->_slice_block [orthogdim];
+  int stride=grid->_slice_stride[orthogdim];
+  int ostride=grid->_ostride[orthogdim];
+  
+  //Reduce Data down to lvSum
+  sliceSumReduction_cpu(Data,lvSum,rd, e1,e2,stride,ostride,Nsimd);
+
+  // Sum across simd lanes in the plane, breaking out orthog dir.
+  Coordinate icoor(Nd);
+
+  for(int rt=0;rt<rd;rt++){
+
+    extract(lvSum[rt],extracted);
+
+    for(int idx=0;idx<Nsimd;idx++){
+
+      grid->iCoorFromIindex(icoor,idx);
+
+      int ldx =rt+icoor[orthogdim]*rd;
+
+      lsSum[ldx]=lsSum[ldx]+extracted[idx];
+
+    }
+  }
+  
+  // sum over nodes.
+  for(int t=0;t<fd;t++){
+    int pt = t/ld; // processor plane
+    int lt = t%ld;
+    if ( pt == grid->_processor_coor[orthogdim] ) {
+      result[t]=lsSum[lt];
+    } else {
+      result[t]=Zero();
+    }
+
+  }
+  scalar_type * ptr = (scalar_type *) &result[0];
+  int words = fd*sizeof(sobj)/sizeof(scalar_type);
+  grid->GlobalSumVector(ptr, words);
+}
+
+
+int main (int argc, char ** argv) {
+    
+    using namespace Grid;
+
+    Grid_init(&argc,&argv);
+
+
+    Coordinate latt_size({64,64,64,16});
+    auto simd_layout = GridDefaultSimd(Nd, vComplexD::Nsimd());
+    auto mpi_layout = GridDefaultMpi();
+    GridCartesian Grid(latt_size, simd_layout, mpi_layout);
+
+    std::vector<int> seeds({1, 2, 3, 4});
+
+    GridParallelRNG pRNG(&Grid);
+    pRNG.SeedFixedIntegers(seeds);
+
+    LatticeComplexD test_data(&Grid);
+    gaussian(pRNG,test_data);
+
+    std::vector<TComplexD> reduction_reference;
+    std::vector<TComplexD> reduction_result;
+
+    //warmup
+    for (int sweeps = 0; sweeps < 5; sweeps++) {
+      reduction_result = sliceSum(test_data,0);
+    }
+
+    int trace_id = traceStart("sliceSum benchmark - ComplexD");
+    std::cout << GridLogMessage << "Testing ComplexD" << std::endl;
+    std::cout << GridLogMessage << "sizeof(ComplexD) = " << sizeof(ComplexD) << std::endl;
+    std::cout << GridLogMessage << "sizeof(vComplexD) = " << sizeof(vComplexD) << std::endl;
+    for (int i = 0; i < Nd; i++) {
+
+      RealD t=-usecond();
+
+      tracePush("sliceSum");
+      sliceSumCPU(test_data,reduction_reference,i);
+      tracePop("sliceSum");
+
+      t+=usecond();
+      std::cout << GridLogMessage << "Orthog. dir. = " << i << std::endl;
+      std::cout << GridLogMessage << "CPU sliceSum took "<<t<<" usecs"<<std::endl;
+      
+      
+      RealD tgpu=-usecond();
+
+      tracePush("sliceSumGpu");
+      reduction_result = sliceSum(test_data,i);
+      tracePop("sliceSumGpu");
+
+      tgpu+=usecond();
+
+      std::cout << GridLogMessage <<"GPU sliceSum took "<<tgpu<<" usecs"<<std::endl<<std::endl;;
+
+
+      for(int t=0;t<reduction_reference.size();t++) {
+
+        auto diff = reduction_reference[t]-reduction_result[t];
+        assert(abs(TensorRemove(diff)) < 1e-8 );
+
+      }
+
+    
+    }
+    traceStop(trace_id);
+
+    LatticeSpinVectorD test_data_cv(&Grid);
+    gaussian(pRNG,test_data_cv);
+
+    std::vector<SpinVectorD> reduction_reference_cv;
+    std::vector<SpinVectorD> reduction_result_cv;
+
+    //warmup
+    for (int sweeps = 0; sweeps < 5; sweeps++) {
+      reduction_result_cv = sliceSum(test_data_cv,0);
+    }
+    trace_id = traceStart("sliceSum benchmark - SpinVectorD");
+
+    std::cout << GridLogMessage << "Testing SpinVectorD" << std::endl;
+    std::cout << GridLogMessage << "sizeof(SpinVectorD) = " << sizeof(SpinVectorD) << std::endl;
+    std::cout << GridLogMessage << "sizeof(vSpinVectorD) = " << sizeof(vSpinVectorD) << std::endl;
+    for (int i = 0; i < Nd; i++) {
+
+      RealD t=-usecond();
+
+      tracePush("sliceSum");
+      sliceSumCPU(test_data_cv,reduction_reference_cv,i);
+      tracePop("sliceSum");
+
+      t+=usecond();
+      std::cout << GridLogMessage << "Orthog. dir. = " << i << std::endl;
+      std::cout << GridLogMessage << "CPU sliceSum took "<<t<<" usecs"<<std::endl;
+      
+      
+      RealD tgpu=-usecond();
+
+      tracePush("sliceSumGpu");
+      reduction_result_cv = sliceSum(test_data_cv,i);
+      tracePop("sliceSumGpu");
+
+      tgpu+=usecond();
+
+      std::cout << GridLogMessage <<"GPU sliceSum took "<<tgpu<<" usecs"<<std::endl<<std::endl;;
+
+
+      for(int t=0;t<reduction_reference_cv.size();t++) {
+
+        auto diff = reduction_reference_cv[t]-reduction_result_cv[t];
+        assert(abs(diff()(0)()) < 1e-8 );
+        assert(abs(diff()(1)()) < 1e-8 );
+        assert(abs(diff()(2)()) < 1e-8 );
+        assert(abs(diff()(3)()) < 1e-8 );
+
+      }
+
+    
+    }
+    traceStop(trace_id);
+
+    LatticeSpinColourVectorD test_data_scv(&Grid);
+    gaussian(pRNG,test_data_scv);
+
+    std::vector<SpinColourVectorD> reduction_reference_scv;
+    std::vector<SpinColourVectorD> reduction_result_scv;
+
+    //warmup
+    for (int sweeps = 0; sweeps < 5; sweeps++) {
+      reduction_result_scv = sliceSum(test_data_scv,0);
+    }
+    trace_id = traceStart("sliceSum benchmark - SpinColourVectorD");
+
+    std::cout << GridLogMessage << "Testing SpinColourVectorD" << std::endl;
+    std::cout << GridLogMessage << "sizeof(SpinColourVectorD) = " << sizeof(SpinColourVectorD) << std::endl;
+    std::cout << GridLogMessage << "sizeof(vSpinColourVectorD) = " << sizeof(vSpinColourVectorD) << std::endl;
+    for (int i = 0; i < Nd; i++) {
+
+      RealD t=-usecond();
+
+      tracePush("sliceSum");
+      sliceSumCPU(test_data_scv,reduction_reference_scv,i);
+      tracePop("sliceSum");
+
+      t+=usecond();
+      std::cout << GridLogMessage << "Orthog. dir. = " << i << std::endl;
+      std::cout << GridLogMessage << "CPU sliceSum took "<<t<<" usecs"<<std::endl;
+      
+      
+      RealD tgpu=-usecond();
+
+      tracePush("sliceSumGpu");
+      reduction_result_scv = sliceSum(test_data_scv,i);
+      tracePop("sliceSumGpu");
+
+      tgpu+=usecond();
+
+      std::cout << GridLogMessage <<"GPU sliceSum took "<<tgpu<<" usecs"<<std::endl<<std::endl;;
+
+
+      for(int t=0;t<reduction_reference_scv.size();t++) {
+
+        auto diff = reduction_reference_scv[t]-reduction_result_scv[t];
+        // std::cout << diff <<std::endl;
+        assert(abs(diff()(0)(0)) < 1e-8 );
+        assert(abs(diff()(0)(1)) < 1e-8 );
+        assert(abs(diff()(0)(2)) < 1e-8 );
+        assert(abs(diff()(1)(0)) < 1e-8 );
+        assert(abs(diff()(1)(1)) < 1e-8 );
+        assert(abs(diff()(1)(2)) < 1e-8 );    
+        assert(abs(diff()(2)(0)) < 1e-8 );
+        assert(abs(diff()(2)(1)) < 1e-8 );
+        assert(abs(diff()(2)(2)) < 1e-8 );    
+        assert(abs(diff()(3)(0)) < 1e-8 );
+        assert(abs(diff()(3)(1)) < 1e-8 );
+        assert(abs(diff()(3)(2)) < 1e-8 );
+
+      }
+
+    
+    }
+    traceStop(trace_id);
+
+    LatticeSpinColourMatrixD test_data_scm(&Grid);
+    gaussian(pRNG,test_data_scm);
+
+    std::vector<SpinColourMatrixD> reduction_reference_scm;
+    std::vector<SpinColourMatrixD> reduction_result_scm;
+
+    //warmup
+    for (int sweeps = 0; sweeps < 5; sweeps++) {
+      reduction_result_scm = sliceSum(test_data_scm,0);
+    }
+    trace_id = traceStart("sliceSum benchmark - SpinColourMatrixD");
+
+    std::cout << GridLogMessage << "Testing SpinColourMatrixD" << std::endl;
+    std::cout << GridLogMessage << "sizeof(SpinColourMatrixD) = " << sizeof(SpinColourMatrixD) << std::endl;
+    std::cout << GridLogMessage << "sizeof(vSpinColourMatrixD) = " << sizeof(vSpinColourMatrixD) << std::endl;
+    for (int i = 0; i < Nd; i++) {
+
+      RealD t=-usecond();
+
+      tracePush("sliceSum");
+      sliceSumCPU(test_data_scm,reduction_reference_scm,i);
+      tracePop("sliceSum");
+
+      t+=usecond();
+      std::cout << GridLogMessage << "Orthog. dir. = " << i << std::endl;
+      std::cout << GridLogMessage << "CPU sliceSum took "<<t<<" usecs"<<std::endl;
+      
+      
+      RealD tgpu=-usecond();
+
+      tracePush("sliceSumGpu");
+      reduction_result_scm = sliceSum(test_data_scm,i);
+      tracePop("sliceSumGpu");
+
+      tgpu+=usecond();
+
+      std::cout << GridLogMessage <<"GPU sliceSum took "<<tgpu<<" usecs"<<std::endl<<std::endl;;
+
+
+      for(int t=0;t<reduction_reference_scm.size();t++) {
+
+        auto diff = reduction_reference_scm[t]-reduction_result_scm[t];
+        // std::cout << diff <<std::endl;
+        for (int is = 0; is < Ns; is++) {
+          for (int js = 0; js < Ns; js++) {
+            for (int ic = 0; ic < Nc; ic++) {
+              for (int jc = 0; jc < Nc; jc++) {
+                assert(abs(diff()(is,js)(ic,jc)) < 1e-8);
+              }
+            }
+          }
+        }
+
+      }
+
+    
+    }
+    traceStop(trace_id);
+
+    Grid_finalize();
+    return 0;
+}