Merge 461cd045c6 into ccf147d6c1

--disable-accelerator-aware-mpi

.github/ISSUE_TEMPLATE/bug-report.yml

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

.gitignore

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

Grid/Grid_Eigen_Dense.h

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

Grid/Makefile.am

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

Grid/algorithms/Algorithms.h

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

Grid/algorithms/FFT.h

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

Grid/algorithms/LinearOperator.h

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

Grid/algorithms/approx/Zolotarev.cc

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

Grid/algorithms/approx/Zolotarev.h

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

Grid/algorithms/blas/BatchedBlas.cc

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

Grid/algorithms/blas/BatchedBlas.h

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

Grid/algorithms/iterative/ConjugateGradient.h

@@ -191,7 +191,7 @@ public:
 	std::cout << GridLogMessage << "\tAxpyNorm   " << AxpyNormTimer.Elapsed() <<std::endl;
 	std::cout << GridLogMessage << "\tLinearComb " << LinearCombTimer.Elapsed() <<std::endl;
 
-	std::cout << GridLogMessage << "\tMobius flop rate " << DwfFlops/ usecs<< " Gflops " <<std::endl;
+	std::cout << GridLogDebug << "\tMobius flop rate " << DwfFlops/ usecs<< " Gflops " <<std::endl;
 
         if (ErrorOnNoConverge) assert(true_residual / Tolerance < 10000.0);
 

Grid/algorithms/iterative/ConjugateGradientMixedPrecBatched.h

@@ -0,0 +1,213 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./lib/algorithms/iterative/ConjugateGradientMixedPrecBatched.h
+
+    Copyright (C) 2015
+
+    Author: Raoul Hodgson <raoul.hodgson@ed.ac.uk>
+
+    This program is free software; you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation; either version 2 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License along
+    with this program; if not, write to the Free Software Foundation, Inc.,
+    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+    See the full license in the file "LICENSE" in the top level distribution directory
+*************************************************************************************/
+/*  END LEGAL */
+#ifndef GRID_CONJUGATE_GRADIENT_MIXED_PREC_BATCHED_H
+#define GRID_CONJUGATE_GRADIENT_MIXED_PREC_BATCHED_H
+
+NAMESPACE_BEGIN(Grid);
+
+//Mixed precision restarted defect correction CG
+template<class FieldD,class FieldF, 
+  typename std::enable_if< getPrecision<FieldD>::value == 2, int>::type = 0,
+  typename std::enable_if< getPrecision<FieldF>::value == 1, int>::type = 0> 
+class MixedPrecisionConjugateGradientBatched : public LinearFunction<FieldD> {
+public:
+  using LinearFunction<FieldD>::operator();
+  RealD   Tolerance;
+  RealD   InnerTolerance; //Initial tolerance for inner CG. Defaults to Tolerance but can be changed
+  Integer MaxInnerIterations;
+  Integer MaxOuterIterations;
+  Integer MaxPatchupIterations;
+  GridBase* SinglePrecGrid; //Grid for single-precision fields
+  RealD OuterLoopNormMult; //Stop the outer loop and move to a final double prec solve when the residual is OuterLoopNormMult * Tolerance
+  LinearOperatorBase<FieldF> &Linop_f;
+  LinearOperatorBase<FieldD> &Linop_d;
+
+  //Option to speed up *inner single precision* solves using a LinearFunction that produces a guess
+  LinearFunction<FieldF> *guesser;
+  bool updateResidual;
+  
+  MixedPrecisionConjugateGradientBatched(RealD tol, 
+          Integer maxinnerit, 
+          Integer maxouterit, 
+          Integer maxpatchit,
+          GridBase* _sp_grid, 
+          LinearOperatorBase<FieldF> &_Linop_f, 
+          LinearOperatorBase<FieldD> &_Linop_d,
+          bool _updateResidual=true) :
+    Linop_f(_Linop_f), Linop_d(_Linop_d),
+    Tolerance(tol), InnerTolerance(tol), MaxInnerIterations(maxinnerit), MaxOuterIterations(maxouterit), MaxPatchupIterations(maxpatchit), SinglePrecGrid(_sp_grid),
+    OuterLoopNormMult(100.), guesser(NULL), updateResidual(_updateResidual) { };
+
+  void useGuesser(LinearFunction<FieldF> &g){
+    guesser = &g;
+  }
+  
+  void operator() (const FieldD &src_d_in, FieldD &sol_d){
+    std::vector<FieldD> srcs_d_in{src_d_in};
+    std::vector<FieldD> sols_d{sol_d};
+
+    (*this)(srcs_d_in,sols_d);
+
+    sol_d = sols_d[0];
+  }
+
+  void operator() (const std::vector<FieldD> &src_d_in, std::vector<FieldD> &sol_d){
+    assert(src_d_in.size() == sol_d.size());
+    int NBatch = src_d_in.size();
+
+    std::cout << GridLogMessage << "NBatch = " << NBatch << std::endl;
+
+    Integer TotalOuterIterations = 0; //Number of restarts
+    std::vector<Integer> TotalInnerIterations(NBatch,0);     //Number of inner CG iterations
+    std::vector<Integer> TotalFinalStepIterations(NBatch,0); //Number of CG iterations in final patch-up step
+  
+    GridStopWatch TotalTimer;
+    TotalTimer.Start();
+
+    GridStopWatch InnerCGtimer;
+    GridStopWatch PrecChangeTimer;
+    
+    int cb = src_d_in[0].Checkerboard();
+    
+    std::vector<RealD> src_norm;
+    std::vector<RealD> norm;
+    std::vector<RealD> stop;
+    
+    GridBase* DoublePrecGrid = src_d_in[0].Grid();
+    FieldD tmp_d(DoublePrecGrid);
+    tmp_d.Checkerboard() = cb;
+    
+    FieldD tmp2_d(DoublePrecGrid);
+    tmp2_d.Checkerboard() = cb;
+
+    std::vector<FieldD> src_d;
+    std::vector<FieldF> src_f;
+    std::vector<FieldF> sol_f;
+
+    for (int i=0; i<NBatch; i++) {
+      sol_d[i].Checkerboard() = cb;
+
+      src_norm.push_back(norm2(src_d_in[i]));
+      norm.push_back(0.);
+      stop.push_back(src_norm[i] * Tolerance*Tolerance);
+
+      src_d.push_back(src_d_in[i]); //source for next inner iteration, computed from residual during operation
+
+      src_f.push_back(SinglePrecGrid);
+      src_f[i].Checkerboard() = cb;
+
+      sol_f.push_back(SinglePrecGrid);
+      sol_f[i].Checkerboard() = cb;
+    }
+    
+    RealD inner_tol = InnerTolerance;
+    
+    ConjugateGradient<FieldF> CG_f(inner_tol, MaxInnerIterations);
+    CG_f.ErrorOnNoConverge = false;
+    
+    Integer &outer_iter = TotalOuterIterations; //so it will be equal to the final iteration count
+      
+    for(outer_iter = 0; outer_iter < MaxOuterIterations; outer_iter++){
+      std::cout << GridLogMessage << std::endl;
+      std::cout << GridLogMessage << "Outer iteration " << outer_iter << std::endl;
+      
+      bool allConverged = true;
+      
+      for (int i=0; i<NBatch; i++) {
+        //Compute double precision rsd and also new RHS vector.
+        Linop_d.HermOp(sol_d[i], tmp_d);
+        norm[i] = axpy_norm(src_d[i], -1., tmp_d, src_d_in[i]); //src_d is residual vector
+        
+        std::cout<<GridLogMessage<<"MixedPrecisionConjugateGradientBatched: Outer iteration " << outer_iter <<" solve " << i << " residual "<< norm[i] << " target "<< stop[i] <<std::endl;
+
+        PrecChangeTimer.Start();
+        precisionChange(src_f[i], src_d[i]);
+        PrecChangeTimer.Stop();
+        
+        sol_f[i] = Zero();
+      
+        if(norm[i] > OuterLoopNormMult * stop[i]) {
+          allConverged = false;
+        }
+      }
+      if (allConverged) break;
+
+      if (updateResidual) {
+        RealD normMax = *std::max_element(std::begin(norm), std::end(norm));
+        RealD stopMax = *std::max_element(std::begin(stop), std::end(stop));
+        while( normMax * inner_tol * inner_tol < stopMax) inner_tol *= 2;  // inner_tol = sqrt(stop/norm) ??
+        CG_f.Tolerance = inner_tol;
+      }
+
+      //Optionally improve inner solver guess (eg using known eigenvectors)
+      if(guesser != NULL) {
+        (*guesser)(src_f, sol_f);
+      }
+
+      for (int i=0; i<NBatch; i++) {
+        //Inner CG
+        InnerCGtimer.Start();
+        CG_f(Linop_f, src_f[i], sol_f[i]);
+        InnerCGtimer.Stop();
+        TotalInnerIterations[i] += CG_f.IterationsToComplete;
+        
+        //Convert sol back to double and add to double prec solution
+        PrecChangeTimer.Start();
+        precisionChange(tmp_d, sol_f[i]);
+        PrecChangeTimer.Stop();
+        
+        axpy(sol_d[i], 1.0, tmp_d, sol_d[i]);
+      }
+
+    }
+    
+    //Final trial CG
+    std::cout << GridLogMessage << std::endl;
+    std::cout<<GridLogMessage<<"MixedPrecisionConjugateGradientBatched: Starting final patch-up double-precision solve"<<std::endl;
+    
+    for (int i=0; i<NBatch; i++) {
+      ConjugateGradient<FieldD> CG_d(Tolerance, MaxPatchupIterations);
+      CG_d(Linop_d, src_d_in[i], sol_d[i]);
+      TotalFinalStepIterations[i] += CG_d.IterationsToComplete;
+    }
+
+    TotalTimer.Stop();
+
+    std::cout << GridLogMessage << std::endl;
+    for (int i=0; i<NBatch; i++) {
+      std::cout<<GridLogMessage<<"MixedPrecisionConjugateGradientBatched: solve " << i << " Inner CG iterations " << TotalInnerIterations[i] << " Restarts " << TotalOuterIterations << " Final CG iterations " << TotalFinalStepIterations[i] << std::endl;
+    }
+    std::cout << GridLogMessage << std::endl;
+    std::cout<<GridLogMessage<<"MixedPrecisionConjugateGradientBatched: Total time " << TotalTimer.Elapsed() << " Precision change " << PrecChangeTimer.Elapsed() << " Inner CG total " << InnerCGtimer.Elapsed() << std::endl;
+    
+  }
+};
+
+NAMESPACE_END(Grid);
+
+#endif

Grid/algorithms/iterative/ConjugateGradientMultiShiftCleanup.h

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

Grid/algorithms/iterative/ConjugateGradientMultiShiftMixedPrec.h

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

Grid/algorithms/iterative/ImplicitlyRestartedLanczos.h

@@ -419,14 +419,15 @@ until convergence
 	}
       }
 
-      if ( Nconv < Nstop )
+      if ( Nconv < Nstop ) {
 	std::cout << GridLogIRL << "Nconv ("<<Nconv<<") < Nstop ("<<Nstop<<")"<<std::endl;
-
+	std::cout << GridLogIRL << "returning Nstop vectors, the last "<< Nstop-Nconv << "of which might meet convergence criterion only approximately" <<std::endl;
+      }
       eval=eval2;
       
       //Keep only converged
-      eval.resize(Nconv);// Nstop?
-      evec.resize(Nconv,grid);// Nstop?
+      eval.resize(Nstop);// was Nconv
+      evec.resize(Nstop,grid);// was Nconv
       basisSortInPlace(evec,eval,reverse);
       
     }

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/allocator/MemoryManager.cc

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

Grid/allocator/MemoryManager.h

@@ -35,6 +35,7 @@ NAMESPACE_BEGIN(Grid);
 // Move control to configure.ac and Config.h?
 
 #define GRID_ALLOC_SMALL_LIMIT (4096)
+#define GRID_ALLOC_HUGE_LIMIT  (2147483648)
 
 #define STRINGIFY(x) #x
 #define TOSTRING(x) STRINGIFY(x)
@@ -70,6 +71,21 @@ enum ViewMode {
   CpuWriteDiscard = 0x10 // same for now
 };
 
+struct MemoryStatus {
+  uint64_t     DeviceBytes;
+  uint64_t     DeviceLRUBytes;
+  uint64_t     DeviceMaxBytes;
+  uint64_t     HostToDeviceBytes;
+  uint64_t     DeviceToHostBytes;
+  uint64_t     HostToDeviceXfer;
+  uint64_t     DeviceToHostXfer;
+  uint64_t     DeviceEvictions;
+  uint64_t     DeviceDestroy;
+  uint64_t     DeviceAllocCacheBytes;
+  uint64_t     HostAllocCacheBytes;
+};
+
+
 class MemoryManager {
 private:
 
@@ -83,7 +99,7 @@ private:
   } AllocationCacheEntry;
 
   static const int NallocCacheMax=128; 
-  static const int NallocType=6;
+  static const int NallocType=9;
   static AllocationCacheEntry Entries[NallocType][NallocCacheMax];
   static int Victim[NallocType];
   static int Ncache[NallocType];
@@ -121,7 +137,26 @@ private:
   static uint64_t     DeviceToHostXfer;
   static uint64_t     DeviceEvictions;
   static uint64_t     DeviceDestroy;
- 
+  
+  static uint64_t     DeviceCacheBytes();
+  static uint64_t     HostCacheBytes();
+
+  static MemoryStatus GetFootprint(void) {
+    MemoryStatus stat;
+    stat.DeviceBytes       = DeviceBytes;
+    stat.DeviceLRUBytes    = DeviceLRUBytes;
+    stat.DeviceMaxBytes    = DeviceMaxBytes;
+    stat.HostToDeviceBytes = HostToDeviceBytes;
+    stat.DeviceToHostBytes = DeviceToHostBytes;
+    stat.HostToDeviceXfer  = HostToDeviceXfer;
+    stat.DeviceToHostXfer  = DeviceToHostXfer;
+    stat.DeviceEvictions   = DeviceEvictions;
+    stat.DeviceDestroy     = DeviceDestroy;
+    stat.DeviceAllocCacheBytes = DeviceCacheBytes();
+    stat.HostAllocCacheBytes   = HostCacheBytes();
+    return stat;
+  };
+  
  private:
 #ifndef GRID_UVM
   //////////////////////////////////////////////////////////////////////

Grid/allocator/MemoryManagerCache.cc

@@ -519,7 +519,6 @@ void MemoryManager::Audit(std::string s)
   uint64_t LruBytes1=0;
   uint64_t LruBytes2=0;
   uint64_t LruCnt=0;
-  uint64_t LockedBytes=0;
   
   std::cout << " Memory Manager::Audit() from "<<s<<std::endl;
   for(auto it=LRU.begin();it!=LRU.end();it++){

Grid/communicator/Communicator_mpi3.cc

@@ -348,6 +348,7 @@ double CartesianCommunicator::StencilSendToRecvFrom( void *xmit,
   return offbytes;
 }
 
+#undef NVLINK_GET // Define to use get instead of put DMA
 double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list,
 							 void *xmit,
 							 int dest,int dox,
@@ -380,9 +381,15 @@ double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsReques
       list.push_back(rrq);
       off_node_bytes+=rbytes;
     }
+#ifdef NVLINK_GET
+      void *shm = (void *) this->ShmBufferTranslate(from,xmit);
+      assert(shm!=NULL);
+      acceleratorCopyDeviceToDeviceAsynch(shm,recv,rbytes);
+#endif
   }
   
   if (dox) {
+    //  rcrc = crc32(rcrc,(unsigned char *)recv,bytes);
     if ( (gdest == MPI_UNDEFINED) || Stencil_force_mpi ) {
       tag= dir+_processor*32;
       ierr =MPI_Isend(xmit, xbytes, MPI_CHAR,dest,tag,communicator_halo[commdir],&xrq);
@@ -390,9 +397,12 @@ double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsReques
       list.push_back(xrq);
       off_node_bytes+=xbytes;
     } else {
+#ifndef NVLINK_GET
       void *shm = (void *) this->ShmBufferTranslate(dest,recv);
       assert(shm!=NULL);
       acceleratorCopyDeviceToDeviceAsynch(xmit,shm,xbytes);
+#endif
+      
     }
   }
 
@@ -400,11 +410,10 @@ double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsReques
 }
 void CartesianCommunicator::StencilSendToRecvFromComplete(std::vector<CommsRequest_t> &list,int dir)
 {
-  acceleratorCopySynchronise();
-  StencilBarrier();// Synch shared memory on a single nodes
-
   int nreq=list.size();
 
+  acceleratorCopySynchronise();
+
   if (nreq==0) return;
 
   std::vector<MPI_Status> status(nreq);

Grid/communicator/Communicator_none.cc

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

Grid/communicator/SharedMemory.cc

@@ -40,6 +40,9 @@ int                 GlobalSharedMemory::_ShmAlloc;
 uint64_t            GlobalSharedMemory::_ShmAllocBytes;
 
 std::vector<void *> GlobalSharedMemory::WorldShmCommBufs;
+#ifndef ACCELERATOR_AWARE_MPI
+void * GlobalSharedMemory::HostCommBuf;
+#endif
 
 Grid_MPI_Comm       GlobalSharedMemory::WorldShmComm;
 int                 GlobalSharedMemory::WorldShmRank;
@@ -66,6 +69,26 @@ void GlobalSharedMemory::SharedMemoryFree(void)
 /////////////////////////////////
 // Alloc, free shmem region
 /////////////////////////////////
+#ifndef ACCELERATOR_AWARE_MPI
+void *SharedMemory::HostBufferMalloc(size_t bytes){
+  void *ptr = (void *)host_heap_top;
+  host_heap_top  += bytes;
+  host_heap_bytes+= bytes;
+  if (host_heap_bytes >= host_heap_size) {
+    std::cout<< " HostBufferMalloc exceeded heap size -- try increasing with --shm <MB> flag" <<std::endl;
+    std::cout<< " Parameter specified in units of MB (megabytes) " <<std::endl;
+    std::cout<< " Current alloc is " << (bytes/(1024*1024)) <<"MB"<<std::endl;
+    std::cout<< " Current bytes is " << (host_heap_bytes/(1024*1024)) <<"MB"<<std::endl;
+    std::cout<< " Current heap  is " << (host_heap_size/(1024*1024)) <<"MB"<<std::endl;
+    assert(host_heap_bytes<host_heap_size);
+  }
+  return ptr;
+}
+void SharedMemory::HostBufferFreeAll(void) { 
+  host_heap_top  =(size_t)HostCommBuf;
+  host_heap_bytes=0;
+}
+#endif
 void *SharedMemory::ShmBufferMalloc(size_t bytes){
   //  bytes = (bytes+sizeof(vRealD))&(~(sizeof(vRealD)-1));// align up bytes
   void *ptr = (void *)heap_top;
@@ -91,6 +114,59 @@ void *SharedMemory::ShmBufferSelf(void)
   //std::cerr << "ShmBufferSelf "<<ShmRank<<" "<<std::hex<< ShmCommBufs[ShmRank] <<std::dec<<std::endl;
   return ShmCommBufs[ShmRank];
 }
+static inline int divides(int a,int b)
+{
+  return ( b == ( (b/a)*a ) );
+}
+void GlobalSharedMemory::GetShmDims(const Coordinate &WorldDims,Coordinate &ShmDims)
+{
+  ////////////////////////////////////////////////////////////////
+  // Allow user to configure through environment variable
+  ////////////////////////////////////////////////////////////////
+  char* str = getenv(("GRID_SHM_DIMS_" + std::to_string(ShmDims.size())).c_str());
+  if ( str ) {
+    std::vector<int> IntShmDims;
+    GridCmdOptionIntVector(std::string(str),IntShmDims);
+    assert(IntShmDims.size() == WorldDims.size());
+    long ShmSize = 1;
+    for (int dim=0;dim<WorldDims.size();dim++) {
+      ShmSize *= (ShmDims[dim] = IntShmDims[dim]);
+      assert(divides(ShmDims[dim],WorldDims[dim]));
+    }
+    assert(ShmSize == WorldShmSize);
+    return;
+  }
+  
+  ////////////////////////////////////////////////////////////////
+  // Powers of 2,3,5 only in prime decomposition for now
+  ////////////////////////////////////////////////////////////////
+  int ndimension = WorldDims.size();
+  ShmDims=Coordinate(ndimension,1);
+
+  std::vector<int> primes({2,3,5});
+
+  int dim = 0;
+  int last_dim = ndimension - 1;
+  int AutoShmSize = 1;
+  while(AutoShmSize != WorldShmSize) {
+    int p;
+    for(p=0;p<primes.size();p++) {
+      int prime=primes[p];
+      if ( divides(prime,WorldDims[dim]/ShmDims[dim])
+        && divides(prime,WorldShmSize/AutoShmSize)  ) {
+  AutoShmSize*=prime;
+  ShmDims[dim]*=prime;
+  last_dim = dim;
+  break;
+      }
+    }
+    if (p == primes.size() && last_dim == dim) {
+      std::cerr << "GlobalSharedMemory::GetShmDims failed" << std::endl;
+      exit(EXIT_FAILURE);
+    }
+    dim=(dim+1) %ndimension;
+  }
+}
 
 NAMESPACE_END(Grid); 
 

Grid/communicator/SharedMemory.h

@@ -75,7 +75,9 @@ public:
   static int           Hugepages;
 
   static std::vector<void *> WorldShmCommBufs;
-
+#ifndef ACCELERATOR_AWARE_MPI
+  static void *HostCommBuf;
+#endif
   static Grid_MPI_Comm WorldComm;
   static int           WorldRank;
   static int           WorldSize;
@@ -120,6 +122,13 @@ private:
   size_t heap_bytes;
   size_t heap_size;
 
+#ifndef ACCELERATOR_AWARE_MPI
+  size_t host_heap_top;  // set in free all
+  size_t host_heap_bytes;// set in free all
+  void *HostCommBuf;     // set in SetCommunicator
+  size_t host_heap_size; // set in SetCommunicator
+#endif
+  
 protected:
 
   Grid_MPI_Comm    ShmComm; // for barriers
@@ -151,7 +160,10 @@ public:
   void *ShmBufferTranslate(int rank,void * local_p);
   void *ShmBufferMalloc(size_t bytes);
   void  ShmBufferFreeAll(void) ;
-  
+#ifndef ACCELERATOR_AWARE_MPI
+  void *HostBufferMalloc(size_t bytes);
+  void HostBufferFreeAll(void);
+#endif  
   //////////////////////////////////////////////////////////////////////////
   // Make info on Nodes & ranks and Shared memory available
   //////////////////////////////////////////////////////////////////////////

Grid/communicator/SharedMemoryMPI.cc

@@ -27,9 +27,10 @@ Author: Christoph Lehner <christoph@lhnr.de>
 *************************************************************************************/
 /*  END LEGAL */
 
+#define Mheader "SharedMemoryMpi: "
+
 #include <Grid/GridCore.h>
 #include <pwd.h>
-#include <syscall.h>
 
 #ifdef GRID_CUDA
 #include <cuda_runtime_api.h>
@@ -37,12 +38,122 @@ Author: Christoph Lehner <christoph@lhnr.de>
 #ifdef GRID_HIP
 #include <hip/hip_runtime_api.h>
 #endif
-#ifdef GRID_SYCl
-
+#ifdef GRID_SYCL
+#ifdef ACCELERATOR_AWARE_MPI
+#define GRID_SYCL_LEVEL_ZERO_IPC
+#define SHM_SOCKETS
+#endif 
+#include <syscall.h>
 #endif
 
+#include <sys/socket.h>
+#include <sys/un.h>
+
 NAMESPACE_BEGIN(Grid); 
-#define header "SharedMemoryMpi: "
+
+#ifdef SHM_SOCKETS
+
+/*
+ * Barbaric extra intranode communication route in case we need sockets to pass FDs
+ * Forced by level_zero not being nicely designed
+ */
+static int sock;
+static const char *sock_path_fmt = "/tmp/GridUnixSocket.%d";
+static char sock_path[256];
+class UnixSockets {
+public:
+  static void Open(int rank)
+  {
+    int errnum;
+
+    sock = socket(AF_UNIX, SOCK_DGRAM, 0);  assert(sock>0);
+
+    struct sockaddr_un sa_un = { 0 };
+    sa_un.sun_family = AF_UNIX;
+    snprintf(sa_un.sun_path, sizeof(sa_un.sun_path),sock_path_fmt,rank);
+    unlink(sa_un.sun_path);
+    if (bind(sock, (struct sockaddr *)&sa_un, sizeof(sa_un))) {
+      perror("bind failure");
+      exit(EXIT_FAILURE);
+    }
+  }
+
+  static int RecvFileDescriptor(void)
+  {
+    int n;
+    int fd;
+    char buf[1];
+    struct iovec iov;
+    struct msghdr msg;
+    struct cmsghdr *cmsg;
+    char cms[CMSG_SPACE(sizeof(int))];
+
+    iov.iov_base = buf;
+    iov.iov_len = 1;
+
+    memset(&msg, 0, sizeof msg);
+    msg.msg_name = 0;
+    msg.msg_namelen = 0;
+    msg.msg_iov = &iov;
+    msg.msg_iovlen = 1;
+
+    msg.msg_control = (caddr_t)cms;
+    msg.msg_controllen = sizeof cms;
+
+    if((n=recvmsg(sock, &msg, 0)) < 0) {
+      perror("recvmsg failed");
+      return -1;
+    }
+    if(n == 0){
+      perror("recvmsg returned 0");
+      return -1;
+    }
+    cmsg = CMSG_FIRSTHDR(&msg);
+
+    memmove(&fd, CMSG_DATA(cmsg), sizeof(int));
+
+    return fd;
+  }
+
+  static void SendFileDescriptor(int fildes,int xmit_to_rank)
+  {
+    struct msghdr msg;
+    struct iovec iov;
+    struct cmsghdr *cmsg = NULL;
+    char ctrl[CMSG_SPACE(sizeof(int))];
+    char data = ' ';
+
+    memset(&msg, 0, sizeof(struct msghdr));
+    memset(ctrl, 0, CMSG_SPACE(sizeof(int)));
+    iov.iov_base = &data;
+    iov.iov_len = sizeof(data);
+    
+    sprintf(sock_path,sock_path_fmt,xmit_to_rank);
+    
+    struct sockaddr_un sa_un = { 0 };
+    sa_un.sun_family = AF_UNIX;
+    snprintf(sa_un.sun_path, sizeof(sa_un.sun_path),sock_path_fmt,xmit_to_rank);
+
+    msg.msg_name = (void *)&sa_un;
+    msg.msg_namelen = sizeof(sa_un);
+    msg.msg_iov = &iov;
+    msg.msg_iovlen = 1;
+    msg.msg_controllen =  CMSG_SPACE(sizeof(int));
+    msg.msg_control = ctrl;
+
+    cmsg = CMSG_FIRSTHDR(&msg);
+    cmsg->cmsg_level = SOL_SOCKET;
+    cmsg->cmsg_type = SCM_RIGHTS;
+    cmsg->cmsg_len = CMSG_LEN(sizeof(int));
+
+    *((int *) CMSG_DATA(cmsg)) = fildes;
+
+    sendmsg(sock, &msg, 0);
+  };
+};
+#endif
+
+
 /*Construct from an MPI communicator*/
 void GlobalSharedMemory::Init(Grid_MPI_Comm comm)
 {
@@ -65,8 +176,8 @@ void GlobalSharedMemory::Init(Grid_MPI_Comm comm)
   MPI_Comm_size(WorldShmComm     ,&WorldShmSize);
 
   if ( WorldRank == 0) {
-    std::cout << header " World communicator of size " <<WorldSize << std::endl;  
-    std::cout << header " Node  communicator of size " <<WorldShmSize << std::endl;
+    std::cout << Mheader " World communicator of size " <<WorldSize << std::endl;  
+    std::cout << Mheader " Node  communicator of size " <<WorldShmSize << std::endl;
   }
   // WorldShmComm, WorldShmSize, WorldShmRank
 
@@ -169,59 +280,7 @@ void GlobalSharedMemory::OptimalCommunicator(const Coordinate &processors,Grid_M
   if(nscan==3 && HPEhypercube ) OptimalCommunicatorHypercube(processors,optimal_comm,SHM);
   else                          OptimalCommunicatorSharedMemory(processors,optimal_comm,SHM);
 }
-static inline int divides(int a,int b)
-{
-  return ( b == ( (b/a)*a ) );
-}
-void GlobalSharedMemory::GetShmDims(const Coordinate &WorldDims,Coordinate &ShmDims)
-{
-  ////////////////////////////////////////////////////////////////
-  // Allow user to configure through environment variable
-  ////////////////////////////////////////////////////////////////
-  char* str = getenv(("GRID_SHM_DIMS_" + std::to_string(ShmDims.size())).c_str());
-  if ( str ) {
-    std::vector<int> IntShmDims;
-    GridCmdOptionIntVector(std::string(str),IntShmDims);
-    assert(IntShmDims.size() == WorldDims.size());
-    long ShmSize = 1;
-    for (int dim=0;dim<WorldDims.size();dim++) {
-      ShmSize *= (ShmDims[dim] = IntShmDims[dim]);
-      assert(divides(ShmDims[dim],WorldDims[dim]));
-    }
-    assert(ShmSize == WorldShmSize);
-    return;
-  }
-  
-  ////////////////////////////////////////////////////////////////
-  // Powers of 2,3,5 only in prime decomposition for now
-  ////////////////////////////////////////////////////////////////
-  int ndimension = WorldDims.size();
-  ShmDims=Coordinate(ndimension,1);
 
-  std::vector<int> primes({2,3,5});
-
-  int dim = 0;
-  int last_dim = ndimension - 1;
-  int AutoShmSize = 1;
-  while(AutoShmSize != WorldShmSize) {
-    int p;
-    for(p=0;p<primes.size();p++) {
-      int prime=primes[p];
-      if ( divides(prime,WorldDims[dim]/ShmDims[dim])
-        && divides(prime,WorldShmSize/AutoShmSize)  ) {
-	AutoShmSize*=prime;
-	ShmDims[dim]*=prime;
-	last_dim = dim;
-	break;
-      }
-    }
-    if (p == primes.size() && last_dim == dim) {
-      std::cerr << "GlobalSharedMemory::GetShmDims failed" << std::endl;
-      exit(EXIT_FAILURE);
-    }
-    dim=(dim+1) %ndimension;
-  }
-}
 void GlobalSharedMemory::OptimalCommunicatorHypercube(const Coordinate &processors,Grid_MPI_Comm & optimal_comm,Coordinate &SHM)
 {
   ////////////////////////////////////////////////////////////////
@@ -395,7 +454,7 @@ void GlobalSharedMemory::OptimalCommunicatorSharedMemory(const Coordinate &proce
 #ifdef GRID_MPI3_SHMGET
 void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 {
-  std::cout << header "SharedMemoryAllocate "<< bytes<< " shmget implementation "<<std::endl;
+  std::cout << Mheader "SharedMemoryAllocate "<< bytes<< " shmget implementation "<<std::endl;
   assert(_ShmSetup==1);
   assert(_ShmAlloc==0);
 
@@ -455,46 +514,6 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 // Hugetlbfs mapping intended
 ////////////////////////////////////////////////////////////////////////////////////////////
 #if defined(GRID_CUDA) ||defined(GRID_HIP)  || defined(GRID_SYCL)
-
-//if defined(GRID_SYCL)
-#if 0
-void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
-{
-  void * ShmCommBuf ; 
-  assert(_ShmSetup==1);
-  assert(_ShmAlloc==0);
-
-  //////////////////////////////////////////////////////////////////////////////////////////////////////////
-  // allocate the pointer array for shared windows for our group
-  //////////////////////////////////////////////////////////////////////////////////////////////////////////
-  MPI_Barrier(WorldShmComm);
-  WorldShmCommBufs.resize(WorldShmSize);
-
-  ///////////////////////////////////////////////////////////////////////////////////////////////////////////
-  // Each MPI rank should allocate our own buffer
-  ///////////////////////////////////////////////////////////////////////////////////////////////////////////
-  ShmCommBuf = acceleratorAllocDevice(bytes);
-
-  if (ShmCommBuf == (void *)NULL ) {
-    std::cerr << " SharedMemoryMPI.cc acceleratorAllocDevice failed NULL pointer for " << bytes<<" bytes " << std::endl;
-    exit(EXIT_FAILURE);  
-  }
-
-  std::cout << WorldRank << header " SharedMemoryMPI.cc acceleratorAllocDevice "<< bytes 
-	    << "bytes at "<< std::hex<< ShmCommBuf <<std::dec<<" for comms buffers " <<std::endl;
-
-  SharedMemoryZero(ShmCommBuf,bytes);
-
-  assert(WorldShmSize == 1);
-  for(int r=0;r<WorldShmSize;r++){
-    WorldShmCommBufs[r] = ShmCommBuf;
-  }
-  _ShmAllocBytes=bytes;
-  _ShmAlloc=1;
-}
-#endif
-
-#if defined(GRID_CUDA) ||defined(GRID_HIP) ||defined(GRID_SYCL)  
 void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 {
   void * ShmCommBuf ; 
@@ -517,13 +536,16 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
   ///////////////////////////////////////////////////////////////////////////////////////////////////////////
   // Each MPI rank should allocate our own buffer
   ///////////////////////////////////////////////////////////////////////////////////////////////////////////
+#ifndef ACCELERATOR_AWARE_MPI
+  HostCommBuf= malloc(bytes);
+#endif  
   ShmCommBuf = acceleratorAllocDevice(bytes);
   if (ShmCommBuf == (void *)NULL ) {
     std::cerr << " SharedMemoryMPI.cc acceleratorAllocDevice failed NULL pointer for " << bytes<<" bytes " << std::endl;
     exit(EXIT_FAILURE);  
   }
   if ( WorldRank == 0 ){
-    std::cout << WorldRank << header " SharedMemoryMPI.cc acceleratorAllocDevice "<< bytes 
+    std::cout << WorldRank << Mheader " SharedMemoryMPI.cc acceleratorAllocDevice "<< bytes 
 	      << "bytes at "<< std::hex<< ShmCommBuf << " - "<<(bytes-1+(uint64_t)ShmCommBuf) <<std::dec<<" for comms buffers " <<std::endl;
   }
   SharedMemoryZero(ShmCommBuf,bytes);
@@ -531,8 +553,13 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
   ///////////////////////////////////////////////////////////////////////////////////////////////////////////
   // Loop over ranks/gpu's on our node
   ///////////////////////////////////////////////////////////////////////////////////////////////////////////
+#ifdef SHM_SOCKETS
+  UnixSockets::Open(WorldShmRank);
+#endif
   for(int r=0;r<WorldShmSize;r++){
 
+    MPI_Barrier(WorldShmComm);
+
 #ifndef GRID_MPI3_SHM_NONE
     //////////////////////////////////////////////////
     // If it is me, pass around the IPC access key
@@ -540,24 +567,32 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
     void * thisBuf = ShmCommBuf;
     if(!Stencil_force_mpi) {
 #ifdef GRID_SYCL_LEVEL_ZERO_IPC
-    typedef struct { int fd; pid_t pid ; } clone_mem_t;
+    typedef struct { int fd; pid_t pid ; ze_ipc_mem_handle_t ze; } clone_mem_t;
 
-    auto zeDevice    = cl::sycl::get_native<cl::sycl::backend::level_zero>(theGridAccelerator->get_device());
-    auto zeContext   = cl::sycl::get_native<cl::sycl::backend::level_zero>(theGridAccelerator->get_context());
+    auto zeDevice    = cl::sycl::get_native<cl::sycl::backend::ext_oneapi_level_zero>(theGridAccelerator->get_device());
+    auto zeContext   = cl::sycl::get_native<cl::sycl::backend::ext_oneapi_level_zero>(theGridAccelerator->get_context());
       
     ze_ipc_mem_handle_t ihandle;
     clone_mem_t handle;
-
+    
     if ( r==WorldShmRank ) { 
       auto err = zeMemGetIpcHandle(zeContext,ShmCommBuf,&ihandle);
       if ( err != ZE_RESULT_SUCCESS ) {
-	std::cout << "SharedMemoryMPI.cc zeMemGetIpcHandle failed for rank "<<r<<" "<<std::hex<<err<<std::dec<<std::endl;
+	std::cerr << "SharedMemoryMPI.cc zeMemGetIpcHandle failed for rank "<<r<<" "<<std::hex<<err<<std::dec<<std::endl;
 	exit(EXIT_FAILURE);
       } else {
 	std::cout << "SharedMemoryMPI.cc zeMemGetIpcHandle succeeded for rank "<<r<<" "<<std::hex<<err<<std::dec<<std::endl;
       }
       memcpy((void *)&handle.fd,(void *)&ihandle,sizeof(int));
       handle.pid = getpid();
+      memcpy((void *)&handle.ze,(void *)&ihandle,sizeof(ihandle));
+#ifdef SHM_SOCKETS
+      for(int rr=0;rr<WorldShmSize;rr++){
+	if(rr!=r){
+	  UnixSockets::SendFileDescriptor(handle.fd,rr);
+	}
+      }
+#endif
     }
 #endif
 #ifdef GRID_CUDA
@@ -585,6 +620,7 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
     // Share this IPC handle across the Shm Comm
     //////////////////////////////////////////////////
     { 
+      MPI_Barrier(WorldShmComm);
       int ierr=MPI_Bcast(&handle,
 			 sizeof(handle),
 			 MPI_BYTE,
@@ -600,6 +636,10 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 #ifdef GRID_SYCL_LEVEL_ZERO_IPC
     if ( r!=WorldShmRank ) {
       thisBuf = nullptr;
+      int myfd;
+#ifdef SHM_SOCKETS
+      myfd=UnixSockets::RecvFileDescriptor();
+#else
       std::cout<<"mapping seeking remote pid/fd "
 	       <<handle.pid<<"/"
 	       <<handle.fd<<std::endl;
@@ -607,16 +647,22 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
       int pidfd = syscall(SYS_pidfd_open,handle.pid,0);
       std::cout<<"Using IpcHandle pidfd "<<pidfd<<"\n";
       //      int myfd  = syscall(SYS_pidfd_getfd,pidfd,handle.fd,0);
-      int myfd  = syscall(438,pidfd,handle.fd,0);
-
-      std::cout<<"Using IpcHandle myfd "<<myfd<<"\n";
-      
+      myfd  = syscall(438,pidfd,handle.fd,0);
+      int err_t = errno;
+      if (myfd < 0) {
+        fprintf(stderr,"pidfd_getfd returned %d errno was %d\n", myfd,err_t); fflush(stderr);
+	perror("pidfd_getfd failed ");
+	assert(0);
+      }
+#endif
+      std::cout<<"Using IpcHandle mapped remote pid "<<handle.pid <<" FD "<<handle.fd <<" to myfd "<<myfd<<"\n";
+      memcpy((void *)&ihandle,(void *)&handle.ze,sizeof(ihandle));
       memcpy((void *)&ihandle,(void *)&myfd,sizeof(int));
 
       auto err = zeMemOpenIpcHandle(zeContext,zeDevice,ihandle,0,&thisBuf);
       if ( err != ZE_RESULT_SUCCESS ) {
-	std::cout << "SharedMemoryMPI.cc "<<zeContext<<" "<<zeDevice<<std::endl;
-	std::cout << "SharedMemoryMPI.cc zeMemOpenIpcHandle failed for rank "<<r<<" "<<std::hex<<err<<std::dec<<std::endl; 
+	std::cerr << "SharedMemoryMPI.cc "<<zeContext<<" "<<zeDevice<<std::endl;
+	std::cerr << "SharedMemoryMPI.cc zeMemOpenIpcHandle failed for rank "<<r<<" "<<std::hex<<err<<std::dec<<std::endl; 
 	exit(EXIT_FAILURE);
       } else {
 	std::cout << "SharedMemoryMPI.cc zeMemOpenIpcHandle succeeded for rank "<<r<<std::endl;
@@ -651,18 +697,18 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 #else
     WorldShmCommBufs[r] = ShmCommBuf;
 #endif
+    MPI_Barrier(WorldShmComm);
   }
 
   _ShmAllocBytes=bytes;
   _ShmAlloc=1;
 }
-#endif
 
 #else 
 #ifdef GRID_MPI3_SHMMMAP
 void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 {
-  std::cout << header "SharedMemoryAllocate "<< bytes<< " MMAP implementation "<< GRID_SHM_PATH <<std::endl;
+  std::cout << Mheader "SharedMemoryAllocate "<< bytes<< " MMAP implementation "<< GRID_SHM_PATH <<std::endl;
   assert(_ShmSetup==1);
   assert(_ShmAlloc==0);
   //////////////////////////////////////////////////////////////////////////////////////////////////////////
@@ -699,7 +745,7 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
     assert(((uint64_t)ptr&0x3F)==0);
     close(fd);
     WorldShmCommBufs[r] =ptr;
-    //    std::cout << header "Set WorldShmCommBufs["<<r<<"]="<<ptr<< "("<< bytes<< "bytes)"<<std::endl;
+    //    std::cout << Mheader "Set WorldShmCommBufs["<<r<<"]="<<ptr<< "("<< bytes<< "bytes)"<<std::endl;
   }
   _ShmAlloc=1;
   _ShmAllocBytes  = bytes;
@@ -709,7 +755,7 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 #ifdef GRID_MPI3_SHM_NONE
 void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 {
-  std::cout << header "SharedMemoryAllocate "<< bytes<< " MMAP anonymous implementation "<<std::endl;
+  std::cout << Mheader "SharedMemoryAllocate "<< bytes<< " MMAP anonymous implementation "<<std::endl;
   assert(_ShmSetup==1);
   assert(_ShmAlloc==0);
   //////////////////////////////////////////////////////////////////////////////////////////////////////////
@@ -756,7 +802,7 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 ////////////////////////////////////////////////////////////////////////////////////////////
 void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 { 
-  std::cout << header "SharedMemoryAllocate "<< bytes<< " SHMOPEN implementation "<<std::endl;
+  std::cout << Mheader "SharedMemoryAllocate "<< bytes<< " SHMOPEN implementation "<<std::endl;
   assert(_ShmSetup==1);
   assert(_ShmAlloc==0); 
   MPI_Barrier(WorldShmComm);
@@ -880,6 +926,12 @@ void SharedMemory::SetCommunicator(Grid_MPI_Comm comm)
   }
   ShmBufferFreeAll();
 
+#ifndef ACCELERATOR_AWARE_MPI
+  host_heap_size = heap_size;
+  HostCommBuf= GlobalSharedMemory::HostCommBuf;
+  HostBufferFreeAll();
+#endif  
+
   /////////////////////////////////////////////////////////////////////
   // find comm ranks in our SHM group (i.e. which ranks are on our node)
   /////////////////////////////////////////////////////////////////////

Grid/cshift/Cshift_common.h

@@ -29,8 +29,27 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 
 NAMESPACE_BEGIN(Grid);
 
-extern Vector<std::pair<int,int> > Cshift_table; 
+extern std::vector<std::pair<int,int> > Cshift_table; 
+extern commVector<std::pair<int,int> > Cshift_table_device; 
 
+inline std::pair<int,int> *MapCshiftTable(void)
+{
+  // GPU version
+#ifdef ACCELERATOR_CSHIFT    
+  uint64_t sz=Cshift_table.size();
+  if (Cshift_table_device.size()!=sz )    {
+    Cshift_table_device.resize(sz);
+  }
+  acceleratorCopyToDevice((void *)&Cshift_table[0],
+			  (void *)&Cshift_table_device[0],
+			  sizeof(Cshift_table[0])*sz);
+
+  return &Cshift_table_device[0];
+#else 
+  return &Cshift_table[0];
+#endif
+  // CPU version use identify map
+}
 ///////////////////////////////////////////////////////////////////
 // Gather for when there is no need to SIMD split 
 ///////////////////////////////////////////////////////////////////
@@ -74,8 +93,8 @@ Gather_plane_simple (const Lattice<vobj> &rhs,cshiftVector<vobj> &buffer,int dim
   }
   {
     auto buffer_p = & buffer[0];
-    auto table = &Cshift_table[0];
-#ifdef ACCELERATOR_CSHIFT    
+    auto table = MapCshiftTable();
+#ifdef ACCELERATOR_CSHIFT
     autoView(rhs_v , rhs, AcceleratorRead);
     accelerator_for(i,ent,vobj::Nsimd(),{
 	coalescedWrite(buffer_p[table[i].first],coalescedRead(rhs_v[table[i].second]));
@@ -225,7 +244,7 @@ template<class vobj> void Scatter_plane_simple (Lattice<vobj> &rhs,cshiftVector<
   
   {
     auto buffer_p = & buffer[0];
-    auto table = &Cshift_table[0];
+    auto table = MapCshiftTable();
 #ifdef ACCELERATOR_CSHIFT    
     autoView( rhs_v, rhs, AcceleratorWrite);
     accelerator_for(i,ent,vobj::Nsimd(),{
@@ -340,7 +359,7 @@ template<class vobj> void Copy_plane(Lattice<vobj>& lhs,const Lattice<vobj> &rhs
   }
 
   {
-    auto table = &Cshift_table[0];
+    auto table = MapCshiftTable();
 #ifdef ACCELERATOR_CSHIFT    
     autoView(rhs_v , rhs, AcceleratorRead);
     autoView(lhs_v , lhs, AcceleratorWrite);
@@ -392,7 +411,7 @@ template<class vobj> void Copy_plane_permute(Lattice<vobj>& lhs,const Lattice<vo
   }
 
   {
-    auto table = &Cshift_table[0];
+    auto table = MapCshiftTable();
 #ifdef ACCELERATOR_CSHIFT    
     autoView( rhs_v, rhs, AcceleratorRead);
     autoView( lhs_v, lhs, AcceleratorWrite);

Grid/cshift/Cshift_mpi.h

@@ -52,7 +52,8 @@ template<class vobj> Lattice<vobj> Cshift(const Lattice<vobj> &rhs,int dimension
   int comm_dim        = rhs.Grid()->_processors[dimension] >1 ;
   int splice_dim      = rhs.Grid()->_simd_layout[dimension]>1 && (comm_dim);
 
-
+  RealD t1,t0;
+  t0=usecond();
   if ( !comm_dim ) {
     //std::cout << "CSHIFT: Cshift_local" <<std::endl;
     Cshift_local(ret,rhs,dimension,shift); // Handles checkerboarding
@@ -63,6 +64,8 @@ template<class vobj> Lattice<vobj> Cshift(const Lattice<vobj> &rhs,int dimension
     //std::cout << "CSHIFT: Cshift_comms" <<std::endl;
     Cshift_comms(ret,rhs,dimension,shift);
   }
+  t1=usecond();
+  //  std::cout << GridLogPerformance << "Cshift took "<< (t1-t0)/1e3 << " ms"<<std::endl;
   return ret;
 }
 
@@ -127,16 +130,20 @@ template<class vobj> void Cshift_comms(Lattice<vobj> &ret,const Lattice<vobj> &r
     
   int cb= (cbmask==0x2)? Odd : Even;
   int sshift= rhs.Grid()->CheckerBoardShiftForCB(rhs.Checkerboard(),dimension,shift,cb);
-
+  RealD tcopy=0.0;
+  RealD tgather=0.0;
+  RealD tscatter=0.0;
+  RealD tcomms=0.0;
+  uint64_t xbytes=0;
   for(int x=0;x<rd;x++){       
 
     int sx        =  (x+sshift)%rd;
     int comm_proc = ((x+sshift)/rd)%pd;
     
     if (comm_proc==0) {
-
+      tcopy-=usecond();
       Copy_plane(ret,rhs,dimension,x,sx,cbmask); 
-
+      tcopy+=usecond();
     } else {
 
       int words = buffer_size;
@@ -144,26 +151,39 @@ template<class vobj> void Cshift_comms(Lattice<vobj> &ret,const Lattice<vobj> &r
 
       int bytes = words * sizeof(vobj);
 
+      tgather-=usecond();
       Gather_plane_simple (rhs,send_buf,dimension,sx,cbmask);
+      tgather+=usecond();
 
       //      int rank           = grid->_processor;
       int recv_from_rank;
       int xmit_to_rank;
       grid->ShiftedRanks(dimension,comm_proc,xmit_to_rank,recv_from_rank);
-
-      grid->Barrier();
+      
+      tcomms-=usecond();
+      //      grid->Barrier();
 
       grid->SendToRecvFrom((void *)&send_buf[0],
 			   xmit_to_rank,
 			   (void *)&recv_buf[0],
 			   recv_from_rank,
 			   bytes);
+      xbytes+=bytes;
+      //      grid->Barrier();
+      tcomms+=usecond();
 
-      grid->Barrier();
-
+      tscatter-=usecond();
       Scatter_plane_simple (ret,recv_buf,dimension,x,cbmask);
+      tscatter+=usecond();
     }
   }
+  /*
+  std::cout << GridLogPerformance << " Cshift copy    "<<tcopy/1e3<<" ms"<<std::endl;
+  std::cout << GridLogPerformance << " Cshift gather  "<<tgather/1e3<<" ms"<<std::endl;
+  std::cout << GridLogPerformance << " Cshift scatter "<<tscatter/1e3<<" ms"<<std::endl;
+  std::cout << GridLogPerformance << " Cshift comm    "<<tcomms/1e3<<" ms"<<std::endl;
+  std::cout << GridLogPerformance << " Cshift BW      "<<(2.0*xbytes)/tcomms<<" MB/s "<<2*xbytes<< " Bytes "<<std::endl;
+  */
 }
 
 template<class vobj> void  Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vobj> &rhs,int dimension,int shift,int cbmask)
@@ -190,6 +210,12 @@ template<class vobj> void  Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vo
   assert(shift>=0);
   assert(shift<fd);
 
+  RealD tcopy=0.0;
+  RealD tgather=0.0;
+  RealD tscatter=0.0;
+  RealD tcomms=0.0;
+  uint64_t xbytes=0;
+  
   int permute_type=grid->PermuteType(dimension);
 
   ///////////////////////////////////////////////
@@ -227,7 +253,9 @@ template<class vobj> void  Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vo
       pointers[i] = &send_buf_extract[i][0];
     }
     int sx   = (x+sshift)%rd;
+    tgather-=usecond();
     Gather_plane_extract(rhs,pointers,dimension,sx,cbmask);
+    tgather+=usecond();
 
     for(int i=0;i<Nsimd;i++){
       
@@ -252,7 +280,8 @@ template<class vobj> void  Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vo
       if(nbr_proc){
 	grid->ShiftedRanks(dimension,nbr_proc,xmit_to_rank,recv_from_rank); 
 
-	grid->Barrier();
+	tcomms-=usecond();
+	//	grid->Barrier();
 
 	send_buf_extract_mpi = &send_buf_extract[nbr_lane][0];
 	recv_buf_extract_mpi = &recv_buf_extract[i][0];
@@ -262,7 +291,9 @@ template<class vobj> void  Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vo
 			     recv_from_rank,
 			     bytes);
 
-	grid->Barrier();
+	xbytes+=bytes;
+	//	grid->Barrier();
+	tcomms+=usecond();
 
 	rpointers[i] = &recv_buf_extract[i][0];
       } else { 
@@ -270,9 +301,17 @@ template<class vobj> void  Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vo
       }
 
     }
+    tscatter-=usecond();
     Scatter_plane_merge(ret,rpointers,dimension,x,cbmask);
+    tscatter+=usecond();
   }
-
+  /*
+  std::cout << GridLogPerformance << " Cshift (s) copy    "<<tcopy/1e3<<" ms"<<std::endl;
+  std::cout << GridLogPerformance << " Cshift (s) gather  "<<tgather/1e3<<" ms"<<std::endl;
+  std::cout << GridLogPerformance << " Cshift (s) scatter "<<tscatter/1e3<<" ms"<<std::endl;
+  std::cout << GridLogPerformance << " Cshift (s) comm    "<<tcomms/1e3<<" ms"<<std::endl;
+  std::cout << GridLogPerformance << " Cshift BW      "<<(2.0*xbytes)/tcomms<<" MB/s "<<2*xbytes<< " Bytes "<<std::endl;
+  */
 }
 #else 
 template<class vobj> void Cshift_comms(Lattice<vobj> &ret,const Lattice<vobj> &rhs,int dimension,int shift,int cbmask)
@@ -292,6 +331,11 @@ template<class vobj> void Cshift_comms(Lattice<vobj> &ret,const Lattice<vobj> &r
   assert(comm_dim==1);
   assert(shift>=0);
   assert(shift<fd);
+  RealD tcopy=0.0;
+  RealD tgather=0.0;
+  RealD tscatter=0.0;
+  RealD tcomms=0.0;
+  uint64_t xbytes=0;
   
   int buffer_size = rhs.Grid()->_slice_nblock[dimension]*rhs.Grid()->_slice_block[dimension];
   static cshiftVector<vobj> send_buf_v; send_buf_v.resize(buffer_size);
@@ -315,7 +359,9 @@ template<class vobj> void Cshift_comms(Lattice<vobj> &ret,const Lattice<vobj> &r
     
     if (comm_proc==0) {
 
+      tcopy-=usecond();
       Copy_plane(ret,rhs,dimension,x,sx,cbmask); 
+      tcopy+=usecond();
 
     } else {
 
@@ -324,7 +370,9 @@ template<class vobj> void Cshift_comms(Lattice<vobj> &ret,const Lattice<vobj> &r
 
       int bytes = words * sizeof(vobj);
 
+      tgather-=usecond();
       Gather_plane_simple (rhs,send_buf_v,dimension,sx,cbmask);
+      tgather+=usecond();
 
       //      int rank           = grid->_processor;
       int recv_from_rank;
@@ -332,7 +380,8 @@ template<class vobj> void Cshift_comms(Lattice<vobj> &ret,const Lattice<vobj> &r
       grid->ShiftedRanks(dimension,comm_proc,xmit_to_rank,recv_from_rank);
 
 
-      grid->Barrier();
+      tcomms-=usecond();
+      //      grid->Barrier();
 
       acceleratorCopyDeviceToDevice((void *)&send_buf_v[0],(void *)&send_buf[0],bytes);
       grid->SendToRecvFrom((void *)&send_buf[0],
@@ -340,13 +389,24 @@ template<class vobj> void Cshift_comms(Lattice<vobj> &ret,const Lattice<vobj> &r
 			   (void *)&recv_buf[0],
 			   recv_from_rank,
 			   bytes);
+      xbytes+=bytes;
       acceleratorCopyDeviceToDevice((void *)&recv_buf[0],(void *)&recv_buf_v[0],bytes);
 
-      grid->Barrier();
+      //      grid->Barrier();
+      tcomms+=usecond();
 
+      tscatter-=usecond();
       Scatter_plane_simple (ret,recv_buf_v,dimension,x,cbmask);
+      tscatter+=usecond();
     }
   }
+  /*
+  std::cout << GridLogPerformance << " Cshift copy    "<<tcopy/1e3<<" ms"<<std::endl;
+  std::cout << GridLogPerformance << " Cshift gather  "<<tgather/1e3<<" ms"<<std::endl;
+  std::cout << GridLogPerformance << " Cshift scatter "<<tscatter/1e3<<" ms"<<std::endl;
+  std::cout << GridLogPerformance << " Cshift comm    "<<tcomms/1e3<<" ms"<<std::endl;
+  std::cout << GridLogPerformance << " Cshift BW      "<<(2.0*xbytes)/tcomms<<" MB/s "<<2*xbytes<< " Bytes "<<std::endl;
+  */
 }
 
 template<class vobj> void  Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vobj> &rhs,int dimension,int shift,int cbmask)
@@ -372,6 +432,11 @@ template<class vobj> void  Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vo
   assert(simd_layout==2);
   assert(shift>=0);
   assert(shift<fd);
+  RealD tcopy=0.0;
+  RealD tgather=0.0;
+  RealD tscatter=0.0;
+  RealD tcomms=0.0;
+  uint64_t xbytes=0;
 
   int permute_type=grid->PermuteType(dimension);
 
@@ -414,8 +479,10 @@ template<class vobj> void  Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vo
     for(int i=0;i<Nsimd;i++){       
       pointers[i] = &send_buf_extract[i][0];
     }
+    tgather-=usecond();
     int sx   = (x+sshift)%rd;
     Gather_plane_extract(rhs,pointers,dimension,sx,cbmask);
+    tgather+=usecond();
 
     for(int i=0;i<Nsimd;i++){
       
@@ -440,7 +507,8 @@ template<class vobj> void  Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vo
       if(nbr_proc){
 	grid->ShiftedRanks(dimension,nbr_proc,xmit_to_rank,recv_from_rank); 
 
-	grid->Barrier();
+	tcomms-=usecond();
+	//	grid->Barrier();
 
 	acceleratorCopyDeviceToDevice((void *)&send_buf_extract[nbr_lane][0],(void *)send_buf_extract_mpi,bytes);
 	grid->SendToRecvFrom((void *)send_buf_extract_mpi,
@@ -449,17 +517,28 @@ template<class vobj> void  Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vo
 			     recv_from_rank,
 			     bytes);
 	acceleratorCopyDeviceToDevice((void *)recv_buf_extract_mpi,(void *)&recv_buf_extract[i][0],bytes);
+	xbytes+=bytes;
 
-	grid->Barrier();
+	//	grid->Barrier();
+	tcomms+=usecond();
 	rpointers[i] = &recv_buf_extract[i][0];
       } else { 
 	rpointers[i] = &send_buf_extract[nbr_lane][0];
       }
 
     }
+    tscatter-=usecond();
     Scatter_plane_merge(ret,rpointers,dimension,x,cbmask);
-  }
+    tscatter+=usecond();
 
+  }
+  /*
+  std::cout << GridLogPerformance << " Cshift (s) copy    "<<tcopy/1e3<<" ms"<<std::endl;
+  std::cout << GridLogPerformance << " Cshift (s) gather  "<<tgather/1e3<<" ms"<<std::endl;
+  std::cout << GridLogPerformance << " Cshift (s) scatter "<<tscatter/1e3<<" ms"<<std::endl;
+  std::cout << GridLogPerformance << " Cshift (s) comm    "<<tcomms/1e3<<" ms"<<std::endl;
+  std::cout << GridLogPerformance << " Cshift BW      "<<(2.0*xbytes)/tcomms<<" MB/s"<<std::endl;
+  */
 }
 #endif
 NAMESPACE_END(Grid); 

Grid/cshift/Cshift_table.cc

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

Grid/lattice/Lattice.h

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

Grid/lattice/Lattice_ET.h

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

Grid/lattice/Lattice_arith.h

@@ -270,5 +270,42 @@ RealD axpby_norm(Lattice<vobj> &ret,sobj a,sobj b,const Lattice<vobj> &x,const L
     return axpby_norm_fast(ret,a,b,x,y);
 }
 
+/// Trace product
+template<class obj> auto traceProduct(const Lattice<obj> &rhs_1,const Lattice<obj> &rhs_2)
+  -> Lattice<decltype(trace(obj()))>
+{
+  typedef decltype(trace(obj())) robj;
+  Lattice<robj> ret_i(rhs_1.Grid());
+  autoView( rhs1 , rhs_1, AcceleratorRead);
+  autoView( rhs2 , rhs_2, AcceleratorRead);
+  autoView( ret , ret_i, AcceleratorWrite);
+  ret.Checkerboard() = rhs_1.Checkerboard();
+  accelerator_for(ss,rhs1.size(),obj::Nsimd(),{
+      coalescedWrite(ret[ss],traceProduct(rhs1(ss),rhs2(ss)));
+  });
+  return ret_i;
+}
+
+template<class obj1,class obj2> auto traceProduct(const Lattice<obj1> &rhs_1,const obj2 &rhs2)
+  -> Lattice<decltype(trace(obj1()))>
+{
+  typedef decltype(trace(obj1())) robj;
+  Lattice<robj> ret_i(rhs_1.Grid());
+  autoView( rhs1 , rhs_1, AcceleratorRead);
+  autoView( ret , ret_i, AcceleratorWrite);
+  ret.Checkerboard() = rhs_1.Checkerboard();
+  accelerator_for(ss,rhs1.size(),obj1::Nsimd(),{
+      coalescedWrite(ret[ss],traceProduct(rhs1(ss),rhs2));
+  });
+  return ret_i;
+}
+template<class obj1,class obj2> auto traceProduct(const obj2 &rhs_2,const Lattice<obj1> &rhs_1)
+  -> Lattice<decltype(trace(obj1()))>
+{
+  return traceProduct(rhs_1,rhs_2);
+}
+
+
+
 NAMESPACE_END(Grid);
 #endif

Grid/lattice/Lattice_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);
 
@@ -153,33 +154,44 @@ inline typename vobj::scalar_objectD sumD_large(const vobj *arg, Integer osites)
 }
 
 template<class vobj>
-inline typename vobj::scalar_object sum(const Lattice<vobj> &arg)
+inline typename vobj::scalar_object rankSum(const Lattice<vobj> &arg)
 {
   Integer osites = arg.Grid()->oSites();
 #if defined(GRID_CUDA)||defined(GRID_HIP)||defined(GRID_SYCL)
-  typename vobj::scalar_object ssum;
   autoView( arg_v, arg, AcceleratorRead);
-  ssum= sum_gpu(&arg_v[0],osites);
+  return sum_gpu(&arg_v[0],osites);
 #else
   autoView(arg_v, arg, CpuRead);
-  auto ssum= sum_cpu(&arg_v[0],osites);
+  return sum_cpu(&arg_v[0],osites);
 #endif  
+}
+
+template<class vobj>
+inline typename vobj::scalar_object sum(const Lattice<vobj> &arg)
+{
+  auto ssum = rankSum(arg);
   arg.Grid()->GlobalSum(ssum);
   return ssum;
 }
 
 template<class vobj>
-inline typename vobj::scalar_object sum_large(const Lattice<vobj> &arg)
+inline typename vobj::scalar_object rankSumLarge(const Lattice<vobj> &arg)
 {
 #if defined(GRID_CUDA)||defined(GRID_HIP)||defined(GRID_SYCL)
   autoView( arg_v, arg, AcceleratorRead);
   Integer osites = arg.Grid()->oSites();
-  auto ssum= sum_gpu_large(&arg_v[0],osites);
+  return sum_gpu_large(&arg_v[0],osites);
 #else
   autoView(arg_v, arg, CpuRead);
   Integer osites = arg.Grid()->oSites();
-  auto ssum= sum_cpu(&arg_v[0],osites);
+  return sum_cpu(&arg_v[0],osites);
 #endif
+}
+
+template<class vobj>
+inline typename vobj::scalar_object sum_large(const Lattice<vobj> &arg)
+{
+  auto ssum = rankSumLarge(arg);
   arg.Grid()->GlobalSum(ssum);
   return ssum;
 }
@@ -269,11 +281,29 @@ inline ComplexD rankInnerProduct(const Lattice<vobj> &left,const Lattice<vobj> &
   return nrm;
 }
 
+
 template<class vobj>
 inline ComplexD innerProduct(const Lattice<vobj> &left,const Lattice<vobj> &right) {
   GridBase *grid = left.Grid();
+
+#ifdef GRID_SYCL
+  uint64_t csum=0;
+  if ( FlightRecorder::LoggingMode != FlightRecorder::LoggingModeNone)
+  {
+    // Hack
+    // Fast integer xor checksum. Can also be used in comms now.
+    autoView(l_v,left,AcceleratorRead);
+    Integer words = left.Grid()->oSites()*sizeof(vobj)/sizeof(uint64_t);
+    uint64_t *base= (uint64_t *)&l_v[0];
+    csum=svm_xor(base,words);
+  }
+  FlightRecorder::CsumLog(csum);
+#endif
   ComplexD nrm = rankInnerProduct(left,right);
+  RealD local = real(nrm);
+  FlightRecorder::NormLog(real(nrm)); 
   grid->GlobalSum(nrm);
+  FlightRecorder::ReductionLog(local,real(nrm)); 
   return nrm;
 }
 
@@ -437,19 +467,10 @@ 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);
@@ -493,6 +514,7 @@ sliceSum(const Lattice<vobj> &Data,int orthogdim)
   return result;
 }
 
+
 template<class vobj>
 static void sliceInnerProductVector( std::vector<ComplexD> & result, const Lattice<vobj> &lhs,const Lattice<vobj> &rhs,int orthogdim) 
 {

Grid/lattice/Lattice_reduction_gpu.h

@@ -30,7 +30,7 @@ int getNumBlocksAndThreads(const Iterator n, const size_t sizeofsobj, Iterator &
   cudaGetDevice(&device);
 #endif
 #ifdef GRID_HIP
-  hipGetDevice(&device);
+  auto r=hipGetDevice(&device);
 #endif
   
   Iterator warpSize            = gpu_props[device].warpSize;
@@ -211,25 +211,22 @@ inline typename vobj::scalar_objectD sumD_gpu_small(const vobj *lat, Integer osi
   assert(ok);
 
   Integer smemSize = numThreads * sizeof(sobj);
-  // UVM seems to be buggy under later CUDA drivers
-  // This fails on A100 and driver 5.30.02 / CUDA 12.1
-  // Fails with multiple NVCC versions back to 11.4,
-  // which worked with earlier drivers.
-  // Not sure which driver had first fail and this bears checking
-  // Is awkward as must install multiple driver versions
+  // Move out of UVM
+  // Turns out I had messed up the synchronise after move to compute stream
+  // as running this on the default stream fools the synchronise
 #undef UVM_BLOCK_BUFFER  
 #ifndef UVM_BLOCK_BUFFER  
   commVector<sobj> buffer(numBlocks);
   sobj *buffer_v = &buffer[0];
   sobj result;
-  reduceKernel<<< numBlocks, numThreads, smemSize >>>(lat, buffer_v, size);
+  reduceKernel<<< numBlocks, numThreads, smemSize, computeStream >>>(lat, buffer_v, size);
   accelerator_barrier();
   acceleratorCopyFromDevice(buffer_v,&result,sizeof(result));
 #else
   Vector<sobj> buffer(numBlocks);
   sobj *buffer_v = &buffer[0];
   sobj result;
-  reduceKernel<<< numBlocks, numThreads, smemSize >>>(lat, buffer_v, size);
+  reduceKernel<<< numBlocks, numThreads, smemSize, computeStream >>>(lat, buffer_v, size);
   accelerator_barrier();
   result = *buffer_v;
 #endif

Grid/lattice/Lattice_reduction_sycl.h

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

Grid/lattice/Lattice_rng.h

@@ -152,6 +152,7 @@ public:
 #ifdef RNG_FAST_DISCARD
   static void Skip(RngEngine &eng,uint64_t site)
   {
+#if 0
     /////////////////////////////////////////////////////////////////////////////////////
     // Skip by 2^40 elements between successive lattice sites
     // This goes by 10^12.
@@ -162,9 +163,9 @@ public:
     // tens of seconds per trajectory so this is clean in all reasonable cases,
     // and margin of safety is orders of magnitude.
     // We could hack Sitmo to skip in the higher order words of state if necessary
-      //
-      // Replace with 2^30 ; avoid problem on large volumes
-      //
+    //
+    // Replace with 2^30 ; avoid problem on large volumes
+    //
     /////////////////////////////////////////////////////////////////////////////////////
     //      uint64_t skip = site+1;  //   Old init Skipped then drew.  Checked compat with faster init
     const int shift = 30;
@@ -179,6 +180,9 @@ public:
     assert((skip >> shift)==site); // check for overflow
 
     eng.discard(skip);
+#else
+    eng.discardhi(site);
+#endif
     //      std::cout << " Engine  " <<site << " state " <<eng<<std::endl;
   } 
 #endif
@@ -407,7 +411,7 @@ public:
       std::cout << GridLogMessage << "Seed SHA256: " << GridChecksum::sha256_string(seeds) << std::endl;
       SeedFixedIntegers(seeds);
     }
-  void SeedFixedIntegers(const std::vector<int> &seeds){
+  void SeedFixedIntegers(const std::vector<int> &seeds, int britney=0){
 
     // Everyone generates the same seed_seq based on input seeds
     CartesianCommunicator::BroadcastWorld(0,(void *)&seeds[0],sizeof(int)*seeds.size());
@@ -424,7 +428,6 @@ public:
     // MT implementation does not implement fast discard even though
     // in principle this is possible
     ////////////////////////////////////////////////
-#if 1
     thread_for( lidx, _grid->lSites(), {
 
 	int gidx;
@@ -440,43 +443,17 @@ public:
 	_grid->GlobalCoorToGlobalIndex(gcoor,gidx);
 
 	_grid->GlobalCoorToRankIndex(rank,o_idx,i_idx,gcoor);
-#if 1
-	assert(rank == _grid->ThisRank() );
-#else
-// 
-	if (rank != _grid->ThisRank() ){
-	std::cout <<"rank "<<rank<<" _grid->ThisRank() "<<_grid->ThisRank()<< std::endl;
-//	exit(-42);
-//	assert(0);
-	}
-#endif
 
+	assert(rank == _grid->ThisRank() );
 	
 	int l_idx=generator_idx(o_idx,i_idx);
 	_generators[l_idx] = master_engine;
-	Skip(_generators[l_idx],gidx); // Skip to next RNG sequence
-    });
-#else
-    // Everybody loops over global volume.
-    thread_for( gidx, _grid->_gsites, {
-
-	// Where is it?
-	int rank;
-	int o_idx;
-	int i_idx;
-
-	Coordinate gcoor;
-	_grid->GlobalIndexToGlobalCoor(gidx,gcoor);
-	_grid->GlobalCoorToRankIndex(rank,o_idx,i_idx,gcoor);
-	
-	// If this is one of mine we take it
-	if( rank == _grid->ThisRank() ){
-	  int l_idx=generator_idx(o_idx,i_idx);
-	  _generators[l_idx] = master_engine;
+	if ( britney ) { 
+	  Skip(_generators[l_idx],l_idx); // Skip to next RNG sequence
+	} else { 	
 	  Skip(_generators[l_idx],gidx); // Skip to next RNG sequence
 	}
     });
-#endif
 #else 
     ////////////////////////////////////////////////////////////////
     // Machine and thread decomposition dependent seeding is efficient

Grid/lattice/Lattice_slicesum_core.h

@@ -0,0 +1,224 @@
+#pragma once
+
+#if defined(GRID_CUDA)
+
+#include <cub/cub.cuh>
+#define gpucub cub
+#define gpuError_t cudaError_t
+#define gpuSuccess cudaSuccess
+
+#elif defined(GRID_HIP)
+
+#include <hipcub/hipcub.hpp>
+#define gpucub hipcub
+#define gpuError_t hipError_t
+#define gpuSuccess hipSuccess
+
+#endif
+
+
+NAMESPACE_BEGIN(Grid);
+
+
+#if defined(GRID_CUDA) || defined(GRID_HIP)
+template<class vobj> inline void sliceSumReduction_cub_small(const vobj *Data, Vector<vobj> &lvSum, const int rd, const int e1, const int e2, const int stride, const int ostride, const int Nsimd) {
+  size_t subvol_size = e1*e2;
+  commVector<vobj> reduction_buffer(rd*subvol_size);
+  auto rb_p = &reduction_buffer[0];
+  vobj zero_init;
+  zeroit(zero_init);
+
+  
+  void *temp_storage_array = NULL;
+  size_t temp_storage_bytes = 0;
+  vobj *d_out;
+  int* d_offsets;
+
+  std::vector<int> offsets(rd+1,0);
+
+  for (int i = 0; i < offsets.size(); i++) {
+    offsets[i] = i*subvol_size;
+  }
+  
+  //Allocate memory for output and offset arrays on device
+  d_out = static_cast<vobj*>(acceleratorAllocDevice(rd*sizeof(vobj)));
+  
+  d_offsets = static_cast<int*>(acceleratorAllocDevice((rd+1)*sizeof(int)));
+  
+  //copy offsets to device
+  acceleratorCopyToDeviceAsync(&offsets[0],d_offsets,sizeof(int)*(rd+1),computeStream);
+  
+  
+  gpuError_t gpuErr = gpucub::DeviceSegmentedReduce::Reduce(temp_storage_array, temp_storage_bytes, rb_p,d_out, rd, d_offsets, d_offsets+1, ::gpucub::Sum(), zero_init, computeStream);
+  if (gpuErr!=gpuSuccess) {
+    std::cout << GridLogError << "Lattice_slicesum_gpu.h: Encountered error during gpucub::DeviceSegmentedReduce::Reduce (setup)! Error: " << gpuErr <<std::endl;
+    exit(EXIT_FAILURE);
+  }
+
+  //allocate memory for temp_storage_array  
+  temp_storage_array = acceleratorAllocDevice(temp_storage_bytes);
+  
+  //prepare buffer for reduction
+  //use non-blocking accelerator_for to avoid syncs (ok because we submit to same computeStream)
+  //use 2d accelerator_for to avoid launch latencies found when serially looping over rd 
+  accelerator_for2dNB( s,subvol_size, r,rd, Nsimd,{ 
+  
+    int n = s / e2;
+    int b = s % e2;
+    int so=r*ostride; // base offset for start of plane 
+    int ss= so+n*stride+b;
+
+    coalescedWrite(rb_p[r*subvol_size+s], coalescedRead(Data[ss]));
+
+  });
+  
+  //issue segmented reductions in computeStream
+  gpuErr = gpucub::DeviceSegmentedReduce::Reduce(temp_storage_array, temp_storage_bytes, rb_p, d_out, rd, d_offsets, d_offsets+1,::gpucub::Sum(), zero_init, computeStream);
+  if (gpuErr!=gpuSuccess) {
+    std::cout << GridLogError << "Lattice_slicesum_gpu.h: Encountered error during gpucub::DeviceSegmentedReduce::Reduce! Error: " << gpuErr <<std::endl;
+    exit(EXIT_FAILURE);
+  }
+  
+  acceleratorCopyFromDeviceAsync(d_out,&lvSum[0],rd*sizeof(vobj),computeStream);
+  
+  //sync after copy
+  accelerator_barrier();
+ 
+  acceleratorFreeDevice(temp_storage_array);
+  acceleratorFreeDevice(d_out);
+  acceleratorFreeDevice(d_offsets);
+  
+
+}
+#endif 
+
+
+#if defined(GRID_SYCL)
+template<class vobj> inline void sliceSumReduction_sycl_small(const vobj *Data, Vector <vobj> &lvSum, const int  &rd, const int &e1, const int &e2, const int &stride, const int &ostride, const int &Nsimd)
+{
+  size_t subvol_size = e1*e2;
+
+  vobj *mysum = (vobj *) malloc_shared(rd*sizeof(vobj),*theGridAccelerator);
+  vobj vobj_zero;
+  zeroit(vobj_zero);
+  for (int r = 0; r<rd; r++) { 
+    mysum[r] = vobj_zero; 
+  }
+
+  commVector<vobj> reduction_buffer(rd*subvol_size);    
+
+  auto rb_p = &reduction_buffer[0];
+
+  // autoView(Data_v, Data, AcceleratorRead);
+
+  //prepare reduction buffer 
+  accelerator_for2d( s,subvol_size, r,rd, (size_t)Nsimd,{ 
+  
+      int n = s / e2;
+      int b = s % e2;
+      int so=r*ostride; // base offset for start of plane 
+      int ss= so+n*stride+b;
+
+      coalescedWrite(rb_p[r*subvol_size+s], coalescedRead(Data[ss]));
+
+  });
+
+  for (int r = 0; r < rd; r++) {
+      theGridAccelerator->submit([&](cl::sycl::handler &cgh) {
+          auto Reduction = cl::sycl::reduction(&mysum[r],std::plus<>());
+          cgh.parallel_for(cl::sycl::range<1>{subvol_size},
+          Reduction,
+          [=](cl::sycl::id<1> item, auto &sum) {
+              auto s = item[0];
+              sum += rb_p[r*subvol_size+s];
+          });
+      });
+      
+     
+  }
+  theGridAccelerator->wait();
+  for (int r = 0; r < rd; r++) {
+    lvSum[r] = mysum[r];
+  }
+  free(mysum,*theGridAccelerator);
+}
+#endif
+
+template<class vobj> inline void sliceSumReduction_large(const vobj *Data, Vector<vobj> &lvSum, const int rd, const int e1, const int e2, const int stride, const int ostride, const int Nsimd) {
+  typedef typename vobj::vector_type vector;
+  const int words = sizeof(vobj)/sizeof(vector);
+  const int osites = rd*e1*e2;
+  commVector<vector>buffer(osites);
+  vector *dat = (vector *)Data;
+  vector *buf = &buffer[0];
+  Vector<vector> lvSum_small(rd);
+  vector *lvSum_ptr = (vector *)&lvSum[0];
+
+  for (int w = 0; w < words; w++) {
+    accelerator_for(ss,osites,1,{
+	    buf[ss] = dat[ss*words+w];
+    });
+
+    #if defined(GRID_CUDA) || defined(GRID_HIP)
+      sliceSumReduction_cub_small(buf,lvSum_small,rd,e1,e2,stride, ostride,Nsimd);
+    #elif defined(GRID_SYCL)
+      sliceSumReduction_sycl_small(buf,lvSum_small,rd,e1,e2,stride, ostride,Nsimd);
+    #endif
+
+    for (int r = 0; r < rd; r++) {
+      lvSum_ptr[w+words*r]=lvSum_small[r];
+    }
+
+  }
+
+  
+}
+
+template<class vobj> inline void sliceSumReduction_gpu(const Lattice<vobj> &Data, Vector<vobj> &lvSum, const int rd, const int e1, const int e2, const int stride, const int ostride, const int Nsimd)
+{
+  autoView(Data_v, Data, AcceleratorRead); //reduction libraries cannot deal with large vobjs so we split into small/large case.
+    if constexpr (sizeof(vobj) <= 256) { 
+
+      #if defined(GRID_CUDA) || defined(GRID_HIP)
+        sliceSumReduction_cub_small(&Data_v[0], lvSum, rd, e1, e2, stride, ostride, Nsimd);
+      #elif defined (GRID_SYCL)
+        sliceSumReduction_sycl_small(&Data_v[0], lvSum, rd, e1, e2, stride, ostride, Nsimd);
+      #endif
+
+    }
+    else {
+      sliceSumReduction_large(&Data_v[0], lvSum, rd, e1, e2, stride, ostride, Nsimd);
+    }
+}
+
+
+template<class vobj> inline void sliceSumReduction_cpu(const Lattice<vobj> &Data, Vector<vobj> &lvSum, const int &rd, const int &e1, const int &e2, const int &stride, const int &ostride, const int &Nsimd)
+{
+  // sum over reduced dimension planes, breaking out orthog dir
+  // Parallel over orthog direction
+  autoView( Data_v, Data, CpuRead);
+  thread_for( r,rd, {
+    int so=r*ostride; // base offset for start of plane 
+    for(int n=0;n<e1;n++){
+      for(int b=0;b<e2;b++){
+        int ss= so+n*stride+b;
+        lvSum[r]=lvSum[r]+Data_v[ss];
+      }
+    }
+  });
+}
+
+template<class vobj> inline void sliceSumReduction(const Lattice<vobj> &Data, Vector<vobj> &lvSum, const int &rd, const int &e1, const int &e2, const int &stride, const int &ostride, const int &Nsimd) 
+{
+  #if defined(GRID_CUDA) || defined(GRID_HIP) || defined(GRID_SYCL)
+  
+  sliceSumReduction_gpu(Data, lvSum, rd, e1, e2, stride, ostride, Nsimd);
+  
+  #else
+  sliceSumReduction_cpu(Data, lvSum, rd, e1, e2, stride, ostride, Nsimd);
+
+  #endif
+}
+
+
+NAMESPACE_END(Grid);

Grid/lattice/Lattice_trace.h

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

Grid/lattice/Lattice_transfer.h

@@ -288,7 +288,36 @@ inline void blockProject(Lattice<iVector<CComplex,nbasis > > &coarseData,
     blockZAXPY(fineDataRed,ip,Basis[v],fineDataRed); 
   }
 }
+template<class vobj,class CComplex,int nbasis,class VLattice>
+inline void batchBlockProject(std::vector<Lattice<iVector<CComplex,nbasis>>> &coarseData,
+                               const std::vector<Lattice<vobj>> &fineData,
+                               const VLattice &Basis)
+{
+  int NBatch = fineData.size();
+  assert(coarseData.size() == NBatch);
 
+  GridBase * fine  = fineData[0].Grid();
+  GridBase * coarse= coarseData[0].Grid();
+
+  Lattice<iScalar<CComplex>> ip(coarse);
+  std::vector<Lattice<vobj>> fineDataCopy = fineData;
+
+  autoView(ip_, ip, AcceleratorWrite);
+  for(int v=0;v<nbasis;v++) {
+    for (int k=0; k<NBatch; k++) {
+      autoView( coarseData_ , coarseData[k], AcceleratorWrite);
+      blockInnerProductD(ip,Basis[v],fineDataCopy[k]); // ip = <basis|fine>
+      accelerator_for( sc, coarse->oSites(), vobj::Nsimd(), {
+        convertType(coarseData_[sc](v),ip_[sc]);
+      });
+
+      // improve numerical stability of projection
+      // |fine> = |fine> - <basis|fine> |basis>
+      ip=-ip;
+      blockZAXPY(fineDataCopy[k],ip,Basis[v],fineDataCopy[k]); 
+    }
+  }
+}
 
 template<class vobj,class vobj2,class CComplex>
   inline void blockZAXPY(Lattice<vobj> &fineZ,
@@ -440,15 +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++){
 
@@ -590,6 +617,26 @@ inline void blockPromote(const Lattice<iVector<CComplex,nbasis > > &coarseData,
 }
 #endif
 
+template<class vobj,class CComplex,int nbasis,class VLattice>
+inline void batchBlockPromote(const std::vector<Lattice<iVector<CComplex,nbasis>>> &coarseData,
+                               std::vector<Lattice<vobj>> &fineData,
+                               const VLattice &Basis)
+{
+  int NBatch = coarseData.size();
+  assert(fineData.size() == NBatch);
+
+  GridBase * fine   = fineData[0].Grid();
+  GridBase * coarse = coarseData[0].Grid();
+  for (int k=0; k<NBatch; k++)
+    fineData[k]=Zero();
+  for (int i=0;i<nbasis;i++) {
+    for (int k=0; k<NBatch; k++) {
+      Lattice<iScalar<CComplex>> ip = PeekIndex<0>(coarseData[k],i);
+      blockZAXPY(fineData[k],ip,Basis[i],fineData[k]);
+    }
+  }
+}
+
 // Useful for precision conversion, or indeed anything where an operator= does a conversion on scalars.
 // Simd layouts need not match since we use peek/poke Local
 template<class vobj,class vvobj>
@@ -648,8 +695,68 @@ void localCopyRegion(const Lattice<vobj> &From,Lattice<vobj> & To,Coordinate Fro
   for(int d=0;d<nd;d++){
     assert(Fg->_processors[d]  == Tg->_processors[d]);
   }
-
   // the above should guarantee that the operations are local
+  
+#if 1
+
+  size_t nsite = 1;
+  for(int i=0;i<nd;i++) nsite *= RegionSize[i];
+  
+  size_t tbytes = 4*nsite*sizeof(int);
+  int *table = (int*)malloc(tbytes);
+ 
+  thread_for(idx, nsite, {
+      Coordinate from_coor, to_coor;
+      size_t rem = idx;
+      for(int i=0;i<nd;i++){
+	size_t base_i  = rem % RegionSize[i]; rem /= RegionSize[i];
+	from_coor[i] = base_i + FromLowerLeft[i];
+	to_coor[i] = base_i + ToLowerLeft[i];
+      }
+      
+      int foidx = Fg->oIndex(from_coor);
+      int fiidx = Fg->iIndex(from_coor);
+      int toidx = Tg->oIndex(to_coor);
+      int tiidx = Tg->iIndex(to_coor);
+      int* tt = table + 4*idx;
+      tt[0] = foidx;
+      tt[1] = fiidx;
+      tt[2] = toidx;
+      tt[3] = tiidx;
+    });
+  
+  int* table_d = (int*)acceleratorAllocDevice(tbytes);
+  acceleratorCopyToDevice(table,table_d,tbytes);
+
+  typedef typename vobj::vector_type vector_type;
+  typedef typename vobj::scalar_type scalar_type;
+
+  autoView(from_v,From,AcceleratorRead);
+  autoView(to_v,To,AcceleratorWrite);
+  
+  accelerator_for(idx,nsite,1,{
+      static const int words=sizeof(vobj)/sizeof(vector_type);
+      int* tt = table_d + 4*idx;
+      int from_oidx = *tt++;
+      int from_lane = *tt++;
+      int to_oidx = *tt++;
+      int to_lane = *tt;
+
+      const vector_type* from = (const vector_type *)&from_v[from_oidx];
+      vector_type* to = (vector_type *)&to_v[to_oidx];
+      
+      scalar_type stmp;
+      for(int w=0;w<words;w++){
+	stmp = getlane(from[w], from_lane);
+	putlane(to[w], stmp, to_lane);
+      }
+    });
+  
+  acceleratorFreeDevice(table_d);    
+  free(table);
+  
+
+#else  
   Coordinate ldf = Fg->_ldimensions;
   Coordinate rdf = Fg->_rdimensions;
   Coordinate isf = Fg->_istride;
@@ -658,9 +765,9 @@ void localCopyRegion(const Lattice<vobj> &From,Lattice<vobj> & To,Coordinate Fro
   Coordinate ist = Tg->_istride;
   Coordinate ost = Tg->_ostride;
 
-  autoView( t_v , To, AcceleratorWrite);
-  autoView( f_v , From, AcceleratorRead);
-  accelerator_for(idx,Fg->lSites(),1,{
+  autoView( t_v , To, CpuWrite);
+  autoView( f_v , From, CpuRead);
+  thread_for(idx,Fg->lSites(),{
     sobj s;
     Coordinate Fcoor(nd);
     Coordinate Tcoor(nd);
@@ -673,17 +780,24 @@ void localCopyRegion(const Lattice<vobj> &From,Lattice<vobj> & To,Coordinate Fro
       Tcoor[d] = ToLowerLeft[d]+ Fcoor[d]-FromLowerLeft[d];
     }
     if (in_region) {
-      Integer idx_f = 0; for(int d=0;d<nd;d++) idx_f+=isf[d]*(Fcoor[d]/rdf[d]);
-      Integer idx_t = 0; for(int d=0;d<nd;d++) idx_t+=ist[d]*(Tcoor[d]/rdt[d]);
-      Integer odx_f = 0; for(int d=0;d<nd;d++) odx_f+=osf[d]*(Fcoor[d]%rdf[d]);
-      Integer odx_t = 0; for(int d=0;d<nd;d++) odx_t+=ost[d]*(Tcoor[d]%rdt[d]);
-      vector_type * fp = (vector_type *)&f_v[odx_f];
-      vector_type * tp = (vector_type *)&t_v[odx_t];
+#if 0      
+      Integer idx_f = 0; for(int d=0;d<nd;d++) idx_f+=isf[d]*(Fcoor[d]/rdf[d]); // inner index from
+      Integer idx_t = 0; for(int d=0;d<nd;d++) idx_t+=ist[d]*(Tcoor[d]/rdt[d]); // inner index to
+      Integer odx_f = 0; for(int d=0;d<nd;d++) odx_f+=osf[d]*(Fcoor[d]%rdf[d]); // outer index from
+      Integer odx_t = 0; for(int d=0;d<nd;d++) odx_t+=ost[d]*(Tcoor[d]%rdt[d]); // outer index to
+      scalar_type * fp = (scalar_type *)&f_v[odx_f];
+      scalar_type * tp = (scalar_type *)&t_v[odx_t];
       for(int w=0;w<words;w++){
 	tp[w].putlane(fp[w].getlane(idx_f),idx_t);
       }
+#else
+    peekLocalSite(s,f_v,Fcoor);
+    pokeLocalSite(s,t_v,Tcoor);
+#endif
     }
   });
+
+#endif
 }
 
 
@@ -776,6 +890,8 @@ void ExtractSlice(Lattice<vobj> &lowDim,const Lattice<vobj> & higherDim,int slic
 }
 
 
+//Insert subvolume orthogonal to direction 'orthog' with slice index 'slice_lo' from 'lowDim' onto slice index 'slice_hi' of higherDim
+//The local dimensions of both 'lowDim' and 'higherDim' orthogonal to 'orthog' should be the same
 template<class vobj>
 void InsertSliceLocal(const Lattice<vobj> &lowDim, Lattice<vobj> & higherDim,int slice_lo,int slice_hi, int orthog)
 {
@@ -792,11 +908,70 @@ void InsertSliceLocal(const Lattice<vobj> &lowDim, Lattice<vobj> & higherDim,int
 
   for(int d=0;d<nh;d++){
     if ( d!=orthog ) {
-    assert(lg->_processors[d]  == hg->_processors[d]);
-    assert(lg->_ldimensions[d] == hg->_ldimensions[d]);
-  }
+      assert(lg->_processors[d]  == hg->_processors[d]);
+      assert(lg->_ldimensions[d] == hg->_ldimensions[d]);
+    }
   }
 
+#if 1
+  size_t nsite = lg->lSites()/lg->LocalDimensions()[orthog];
+  size_t tbytes = 4*nsite*sizeof(int);
+  int *table = (int*)malloc(tbytes);
+  
+  thread_for(idx,nsite,{
+    Coordinate lcoor(nl);
+    Coordinate hcoor(nh);
+    lcoor[orthog] = slice_lo;
+    hcoor[orthog] = slice_hi;
+    size_t rem = idx;
+    for(int mu=0;mu<nl;mu++){
+      if(mu != orthog){
+	int xmu = rem % lg->LocalDimensions()[mu];  rem /= lg->LocalDimensions()[mu];
+	lcoor[mu] = hcoor[mu] = xmu;
+      }
+    }
+    int loidx = lg->oIndex(lcoor);
+    int liidx = lg->iIndex(lcoor);
+    int hoidx = hg->oIndex(hcoor);
+    int hiidx = hg->iIndex(hcoor);
+    int* tt = table + 4*idx;
+    tt[0] = loidx;
+    tt[1] = liidx;
+    tt[2] = hoidx;
+    tt[3] = hiidx;
+    });
+   
+  int* table_d = (int*)acceleratorAllocDevice(tbytes);
+  acceleratorCopyToDevice(table,table_d,tbytes);
+
+  typedef typename vobj::vector_type vector_type;
+  typedef typename vobj::scalar_type scalar_type;
+
+  autoView(lowDim_v,lowDim,AcceleratorRead);
+  autoView(higherDim_v,higherDim,AcceleratorWrite);
+  
+  accelerator_for(idx,nsite,1,{
+      static const int words=sizeof(vobj)/sizeof(vector_type);
+      int* tt = table_d + 4*idx;
+      int from_oidx = *tt++;
+      int from_lane = *tt++;
+      int to_oidx = *tt++;
+      int to_lane = *tt;
+
+      const vector_type* from = (const vector_type *)&lowDim_v[from_oidx];
+      vector_type* to = (vector_type *)&higherDim_v[to_oidx];
+      
+      scalar_type stmp;
+      for(int w=0;w<words;w++){
+	stmp = getlane(from[w], from_lane);
+	putlane(to[w], stmp, to_lane);
+      }
+    });
+  
+  acceleratorFreeDevice(table_d);    
+  free(table);
+  
+#else
   // the above should guarantee that the operations are local
   autoView(lowDimv,lowDim,CpuRead);
   autoView(higherDimv,higherDim,CpuWrite);
@@ -812,6 +987,7 @@ void InsertSliceLocal(const Lattice<vobj> &lowDim, Lattice<vobj> & higherDim,int
       pokeLocalSite(s,higherDimv,hcoor);
     }
   });
+#endif
 }
 
 

Grid/lattice/Lattice_view.h

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

Grid/lattice/PaddedCell.h

@@ -0,0 +1,174 @@
+/*************************************************************************************
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./lib/lattice/PaddedCell.h
+
+    Copyright (C) 2019
+
+Author: Peter Boyle pboyle@bnl.gov
+
+    This program is free software; you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation; either version 2 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License along
+    with this program; if not, write to the Free Software Foundation, Inc.,
+    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+    See the full license in the file "LICENSE" in the top level distribution directory
+*************************************************************************************/
+/*  END LEGAL */
+#pragma once
+
+#include<Grid/cshift/Cshift.h>
+
+NAMESPACE_BEGIN(Grid);
+
+//Allow the user to specify how the C-shift is performed, e.g. to respect the appropriate boundary conditions
+template<typename vobj>
+struct CshiftImplBase{
+  virtual Lattice<vobj> Cshift(const Lattice<vobj> &in, int dir, int shift) const = 0;
+  virtual ~CshiftImplBase(){}
+};
+template<typename vobj>
+struct CshiftImplDefault: public CshiftImplBase<vobj>{
+  Lattice<vobj> Cshift(const Lattice<vobj> &in, int dir, int shift) const override{ return Grid::Cshift(in,dir,shift); }
+};
+template<typename Gimpl>
+struct CshiftImplGauge: public CshiftImplBase<typename Gimpl::GaugeLinkField::vector_object>{
+  typename Gimpl::GaugeLinkField Cshift(const typename Gimpl::GaugeLinkField &in, int dir, int shift) const override{ return Gimpl::CshiftLink(in,dir,shift); }
+};  
+
+class PaddedCell {
+public:
+  GridCartesian * unpadded_grid;
+  int dims;
+  int depth;
+  std::vector<GridCartesian *> grids;
+
+  ~PaddedCell()
+  {
+    DeleteGrids();
+  }
+  PaddedCell(int _depth,GridCartesian *_grid)
+  {
+    unpadded_grid = _grid;
+    depth=_depth;
+    dims=_grid->Nd();
+    AllocateGrids();
+    Coordinate local     =unpadded_grid->LocalDimensions();
+    for(int d=0;d<dims;d++){
+      assert(local[d]>=depth);
+    }
+  }
+  void DeleteGrids(void)
+  {
+    for(int d=0;d<grids.size();d++){
+      delete grids[d];
+    }
+    grids.resize(0);
+  };
+  void AllocateGrids(void)
+  {
+    Coordinate local     =unpadded_grid->LocalDimensions();
+    Coordinate simd      =unpadded_grid->_simd_layout;
+    Coordinate processors=unpadded_grid->_processors;
+    Coordinate plocal    =unpadded_grid->LocalDimensions();
+    Coordinate global(dims);
+
+    // expand up one dim at a time
+    for(int d=0;d<dims;d++){
+
+      plocal[d] += 2*depth; 
+
+      for(int d=0;d<dims;d++){
+	global[d] = plocal[d]*processors[d];
+      }
+
+      grids.push_back(new GridCartesian(global,simd,processors));
+    }
+  };
+  template<class vobj>
+  inline Lattice<vobj> Extract(const Lattice<vobj> &in) const
+  {
+    Lattice<vobj> out(unpadded_grid);
+
+    Coordinate local     =unpadded_grid->LocalDimensions();
+    Coordinate fll(dims,depth); // depends on the MPI spread
+    Coordinate tll(dims,0); // depends on the MPI spread
+    localCopyRegion(in,out,fll,tll,local);
+    return out;
+  }
+  template<class vobj>
+  inline Lattice<vobj> Exchange(const Lattice<vobj> &in, const CshiftImplBase<vobj> &cshift = CshiftImplDefault<vobj>()) const
+  {
+    GridBase *old_grid = in.Grid();
+    int dims = old_grid->Nd();
+    Lattice<vobj> tmp = in;
+    for(int d=0;d<dims;d++){
+      tmp = Expand(d,tmp,cshift); // rvalue && assignment
+    }
+    return tmp;
+  }
+  // expand up one dim at a time
+  template<class vobj>
+  inline Lattice<vobj> Expand(int dim, const Lattice<vobj> &in, const CshiftImplBase<vobj> &cshift = CshiftImplDefault<vobj>()) const
+  {
+    GridBase *old_grid = in.Grid();
+    GridCartesian *new_grid = grids[dim];//These are new grids
+    Lattice<vobj>  padded(new_grid);
+    Lattice<vobj> shifted(old_grid);    
+    Coordinate local     =old_grid->LocalDimensions();
+    Coordinate plocal    =new_grid->LocalDimensions();
+    if(dim==0) conformable(old_grid,unpadded_grid);
+    else       conformable(old_grid,grids[dim-1]);
+
+    std::cout << " dim "<<dim<<" local "<<local << " padding to "<<plocal<<std::endl;
+
+    double tins=0, tshift=0;
+    
+    // Middle bit
+    double t = usecond();
+    for(int x=0;x<local[dim];x++){
+      InsertSliceLocal(in,padded,x,depth+x,dim);
+    }
+    tins += usecond() - t;
+    
+    // High bit
+    t = usecond();
+    shifted = cshift.Cshift(in,dim,depth);
+    tshift += usecond() - t;
+
+    t=usecond();
+    for(int x=0;x<depth;x++){
+      InsertSliceLocal(shifted,padded,local[dim]-depth+x,depth+local[dim]+x,dim);
+    }
+    tins += usecond() - t;
+    
+    // Low bit
+    t = usecond();
+    shifted = cshift.Cshift(in,dim,-depth);
+    tshift += usecond() - t;
+    
+    t = usecond();
+    for(int x=0;x<depth;x++){
+      InsertSliceLocal(shifted,padded,x,x,dim);
+    }
+    tins += usecond() - t;
+
+    std::cout << GridLogPerformance << "PaddedCell::Expand timings: cshift:" << tshift/1000 << "ms, insert-slice:" << tins/1000 << "ms" << std::endl;
+    
+    return padded;
+  }
+
+};
+ 
+
+NAMESPACE_END(Grid);
+

Grid/log/Log.h

@@ -179,11 +179,11 @@ extern GridLogger GridLogSolver;
 extern GridLogger GridLogError;
 extern GridLogger GridLogWarning;
 extern GridLogger GridLogMessage;
-extern GridLogger GridLogDebug  ;
+extern GridLogger GridLogDebug;
 extern GridLogger GridLogPerformance;
 extern GridLogger GridLogDslash;
-extern GridLogger GridLogIterative  ;
-extern GridLogger GridLogIntegrator  ;
+extern GridLogger GridLogIterative;
+extern GridLogger GridLogIntegrator;
 extern GridLogger GridLogHMC;
 extern GridLogger GridLogMemory;
 extern GridLogger GridLogTracing;
@@ -191,6 +191,41 @@ extern Colours    GridLogColours;
 
 std::string demangle(const char* name) ;
 
+template<typename... Args>
+inline std::string sjoin(Args&&... args) noexcept {
+    std::ostringstream msg;
+    (msg << ... << args);
+    return msg.str();
+}
+
+/*!  @brief make log messages work like python print */
+template <typename... Args>
+inline void Grid_log(Args&&... args) {
+    std::string msg = sjoin(std::forward<Args>(args)...);
+    std::cout << GridLogMessage << msg << std::endl;
+}
+
+/*!  @brief make warning messages work like python print */
+template <typename... Args>
+inline void Grid_warn(Args&&... args) {
+    std::string msg = sjoin(std::forward<Args>(args)...);
+    std::cout << "\033[33m" << GridLogWarning << msg << "\033[0m" << std::endl;
+}
+
+/*!  @brief make error messages work like python print */
+template <typename... Args>
+inline void Grid_error(Args&&... args) {
+    std::string msg = sjoin(std::forward<Args>(args)...);
+    std::cout << "\033[31m" << GridLogError << msg << "\033[0m" << std::endl;
+}
+
+/*!  @brief make pass messages work like python print */
+template <typename... Args>
+inline void Grid_pass(Args&&... args) {
+    std::string msg = sjoin(std::forward<Args>(args)...);
+    std::cout << "\033[32m" << GridLogMessage << msg << "\033[0m" << std::endl;
+}
+
 #define _NBACKTRACE (256)
 extern void * Grid_backtrace_buffer[_NBACKTRACE];
 

Grid/perfmon/Tracing.h

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

Grid/qcd/QCD.h

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

Grid/qcd/action/ActionBase.h

@@ -34,10 +34,24 @@ directory
 
 NAMESPACE_BEGIN(Grid);
 
+///////////////////////////////////
+// Smart configuration base class
+///////////////////////////////////
+template< class Field >
+class ConfigurationBase
+{
+public:
+  ConfigurationBase() {}
+  virtual ~ConfigurationBase() {}
+  virtual void set_Field(Field& U) =0;
+  virtual void smeared_force(Field&) = 0;
+  virtual Field& get_SmearedU() =0;
+  virtual Field &get_U(bool smeared = false) = 0;
+};
+
 template <class GaugeField >
 class Action 
 {
-
 public:
   bool is_smeared = false;
   RealD deriv_norm_sum;
@@ -77,16 +91,60 @@ public:
   void refresh_timer_stop(void)  { refresh_us+=usecond(); }
   void S_timer_start(void)       { S_us-=usecond(); }
   void S_timer_stop(void)        { S_us+=usecond(); }
+  /////////////////////////////
   // Heatbath?
+  /////////////////////////////
   virtual void refresh(const GaugeField& U, GridSerialRNG &sRNG, GridParallelRNG& pRNG) = 0; // refresh pseudofermions
   virtual RealD S(const GaugeField& U) = 0;                             // evaluate the action
   virtual RealD Sinitial(const GaugeField& U) { return this->S(U); } ;  // if the refresh computes the action, can cache it. Alternately refreshAndAction() ?
   virtual void deriv(const GaugeField& U, GaugeField& dSdU) = 0;        // evaluate the action derivative
+
+  /////////////////////////////////////////////////////////////
+  // virtual smeared interface through configuration container
+  /////////////////////////////////////////////////////////////
+  virtual void refresh(ConfigurationBase<GaugeField> & U, GridSerialRNG &sRNG, GridParallelRNG& pRNG)
+  {
+    refresh(U.get_U(is_smeared),sRNG,pRNG);
+  }
+  virtual RealD S(ConfigurationBase<GaugeField>& U)
+  {
+    return S(U.get_U(is_smeared));
+  }
+  virtual RealD Sinitial(ConfigurationBase<GaugeField>& U) 
+  {
+    return Sinitial(U.get_U(is_smeared));
+  }
+  virtual void deriv(ConfigurationBase<GaugeField>& U, GaugeField& dSdU)
+  {
+    deriv(U.get_U(is_smeared),dSdU); 
+    if ( is_smeared ) {
+      U.smeared_force(dSdU);
+    }
+  }
+  ///////////////////////////////
+  // Logging
+  ///////////////////////////////
   virtual std::string action_name()    = 0;                             // return the action name
   virtual std::string LogParameters()  = 0;                             // prints action parameters
   virtual ~Action(){}
 };
 
+template <class GaugeField >
+class EmptyAction : public Action <GaugeField>
+{
+  virtual void refresh(const GaugeField& U, GridSerialRNG &sRNG, GridParallelRNG& pRNG) { assert(0);}; // refresh pseudofermions
+  virtual RealD S(const GaugeField& U) { return 0.0;};                             // evaluate the action
+  virtual void deriv(const GaugeField& U, GaugeField& dSdU) { assert(0); };        // evaluate the action derivative
+
+  ///////////////////////////////
+  // Logging
+  ///////////////////////////////
+  virtual std::string action_name()    { return std::string("Level Force Log"); };
+  virtual std::string LogParameters()  { return std::string("No parameters");};
+};
+
+
+
 NAMESPACE_END(Grid);
 
 #endif // ACTION_BASE_H

Grid/qcd/action/ActionCore.h

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

Grid/qcd/action/fermion/Fermion.h

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

Grid/qcd/action/fermion/WilsonCompressor.h

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

Grid/qcd/action/fermion/WilsonImpl.h

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

Grid/qcd/action/fermion/WilsonTMFermion.h

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

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

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

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

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

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

@@ -423,14 +423,14 @@ void WilsonKernels<Impl>::DhopDirKernel( StencilImpl &st, DoubledGaugeField &U,S
 #define KERNEL_CALL(A) KERNEL_CALLNB(A); accelerator_barrier();
 
 #define KERNEL_CALL_EXT(A)						\
-  const uint64_t    NN = Nsite*Ls;					\
   const uint64_t    sz = st.surface_list.size();			\
   auto ptr = &st.surface_list[0];					\
   accelerator_forNB( ss, sz, Simd::Nsimd(), {				\
       int sF = ptr[ss];							\
-      int sU = ss/Ls;							\
+      int sU = sF/Ls;							\
       WilsonKernels<Impl>::A(st_v,U_v,buf,sF,sU,in_v,out_v);		\
-    });									
+    });									\
+  accelerator_barrier();
 
 #define ASM_CALL(A)							\
   thread_for( sss, Nsite, {						\
@@ -462,12 +462,9 @@ void WilsonKernels<Impl>::DhopKernel(int Opt,StencilImpl &st,  DoubledGaugeField
     autoView(st_v , st,AcceleratorRead);
 
    if( interior && exterior ) {
+     acceleratorFenceComputeStream();
      if (Opt == WilsonKernelsStatic::OptGeneric    ) { KERNEL_CALL(GenericDhopSite); return;}
-#ifdef SYCL_HACK     
-     if (Opt == WilsonKernelsStatic::OptHandUnroll ) { KERNEL_CALL(HandDhopSiteSycl);    return; }
-#else
      if (Opt == WilsonKernelsStatic::OptHandUnroll ) { KERNEL_CALL(HandDhopSite);    return;}
-#endif     
 #ifndef GRID_CUDA
      if (Opt == WilsonKernelsStatic::OptInlineAsm  ) {  ASM_CALL(AsmDhopSite);    return;}
 #endif
@@ -478,8 +475,10 @@ void WilsonKernels<Impl>::DhopKernel(int Opt,StencilImpl &st,  DoubledGaugeField
      if (Opt == WilsonKernelsStatic::OptInlineAsm  ) {  ASM_CALL(AsmDhopSiteInt);    return;}
 #endif
    } else if( exterior ) {
-     if (Opt == WilsonKernelsStatic::OptGeneric    ) { KERNEL_CALL(GenericDhopSiteExt); return;}
-     if (Opt == WilsonKernelsStatic::OptHandUnroll ) { KERNEL_CALL(HandDhopSiteExt);    return;}
+     // dependent on result of merge
+     acceleratorFenceComputeStream();
+     if (Opt == WilsonKernelsStatic::OptGeneric    ) { KERNEL_CALL_EXT(GenericDhopSiteExt); return;}
+     if (Opt == WilsonKernelsStatic::OptHandUnroll ) { KERNEL_CALL_EXT(HandDhopSiteExt);    return;}
 #ifndef GRID_CUDA
      if (Opt == WilsonKernelsStatic::OptInlineAsm  ) {  ASM_CALL(AsmDhopSiteExt);    return;}
 #endif
@@ -497,26 +496,26 @@ void WilsonKernels<Impl>::DhopKernel(int Opt,StencilImpl &st,  DoubledGaugeField
     autoView(st_v ,st,AcceleratorRead);
 
    if( interior && exterior ) {
+     acceleratorFenceComputeStream();
      if (Opt == WilsonKernelsStatic::OptGeneric    ) { KERNEL_CALL(GenericDhopSiteDag); return;}
      if (Opt == WilsonKernelsStatic::OptHandUnroll ) { KERNEL_CALL(HandDhopSiteDag);    return;}
 #ifndef GRID_CUDA
      if (Opt == WilsonKernelsStatic::OptInlineAsm  ) {  ASM_CALL(AsmDhopSiteDag);     return;}
 #endif
-     acceleratorFenceComputeStream();
    } else if( interior ) {
-     if (Opt == WilsonKernelsStatic::OptGeneric    ) { KERNEL_CALL(GenericDhopSiteDagInt); return;}
-     if (Opt == WilsonKernelsStatic::OptHandUnroll ) { KERNEL_CALL(HandDhopSiteDagInt);    return;}
+     if (Opt == WilsonKernelsStatic::OptGeneric    ) { KERNEL_CALLNB(GenericDhopSiteDagInt); return;}
+     if (Opt == WilsonKernelsStatic::OptHandUnroll ) { KERNEL_CALLNB(HandDhopSiteDagInt);    return;}
 #ifndef GRID_CUDA
      if (Opt == WilsonKernelsStatic::OptInlineAsm  ) {  ASM_CALL(AsmDhopSiteDagInt);     return;}
 #endif
    } else if( exterior ) {
+     // Dependent on result of merge
      acceleratorFenceComputeStream();
-     if (Opt == WilsonKernelsStatic::OptGeneric    ) { KERNEL_CALL(GenericDhopSiteDagExt); return;}
-     if (Opt == WilsonKernelsStatic::OptHandUnroll ) { KERNEL_CALL(HandDhopSiteDagExt);    return;}
+     if (Opt == WilsonKernelsStatic::OptGeneric    ) { KERNEL_CALL_EXT(GenericDhopSiteDagExt); return;}
+     if (Opt == WilsonKernelsStatic::OptHandUnroll ) { KERNEL_CALL_EXT(HandDhopSiteDagExt);    return;}
 #ifndef GRID_CUDA
      if (Opt == WilsonKernelsStatic::OptInlineAsm  ) {  ASM_CALL(AsmDhopSiteDagExt);     return;}
 #endif
-     acceleratorFenceComputeStream();
    }
    assert(0 && " Kernel optimisation case not covered ");
   }

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

Grid/qcd/action/fermion/instantiation/SpWilsonTwoIndexSymmetricImplF/WilsonCloverFermionInstantiationSpWilsonTwoIndexSymmetricImplF.cc

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

Grid/qcd/action/fermion/instantiation/SpWilsonTwoIndexSymmetricImplF/WilsonFermionInstantiationSpWilsonTwoIndexSymmetricImplF.cc

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

Grid/qcd/action/fermion/instantiation/SpWilsonTwoIndexSymmetricImplF/WilsonKernelsInstantiationSpWilsonTwoIndexSymmetricImplF.cc

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

Grid/qcd/action/fermion/instantiation/SpWilsonTwoIndexSymmetricImplF/WilsonTMFermionInstantiationSpWilsonTwoIndexSymmetricImplF.cc

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

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

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

Grid/qcd/action/fermion/instantiation/WilsonImplD2/CayleyFermion5DInstantiationWilsonImplD2.cc

@@ -1 +0,0 @@
-../CayleyFermion5DInstantiation.cc.master

Grid/qcd/action/fermion/instantiation/WilsonImplD2/ContinuedFractionFermion5DInstantiationWilsonImplD2.cc

@@ -1 +0,0 @@
-../ContinuedFractionFermion5DInstantiation.cc.master

Grid/qcd/action/fermion/instantiation/WilsonImplD2/DomainWallEOFAFermionInstantiationWilsonImplD2.cc

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

Grid/qcd/action/fermion/instantiation/WilsonImplD2/MobiusEOFAFermionInstantiationWilsonImplD2.cc

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

Grid/qcd/action/fermion/instantiation/WilsonImplD2/PartialFractionFermion5DInstantiationWilsonImplD2.cc

@@ -1 +0,0 @@
-../PartialFractionFermion5DInstantiation.cc.master

Grid/qcd/action/fermion/instantiation/WilsonImplD2/WilsonFermion5DInstantiationWilsonImplD2.cc

@@ -1 +0,0 @@
-../WilsonFermion5DInstantiation.cc.master

Grid/qcd/action/fermion/instantiation/WilsonImplD2/impl.h

@@ -1 +0,0 @@
-#define IMPLEMENTATION WilsonImplD2

Grid/qcd/action/fermion/instantiation/ZWilsonImplD2/CayleyFermion5DInstantiationZWilsonImplD2.cc

@@ -1 +0,0 @@
-../CayleyFermion5DInstantiation.cc.master

Grid/qcd/action/fermion/instantiation/ZWilsonImplD2/ContinuedFractionFermion5DInstantiationZWilsonImplD2.cc

@@ -1 +0,0 @@
-../ContinuedFractionFermion5DInstantiation.cc.master

Grid/qcd/action/fermion/instantiation/ZWilsonImplD2/DomainWallEOFAFermionInstantiationZWilsonImplD2.cc

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

Grid/qcd/action/fermion/instantiation/ZWilsonImplD2/MobiusEOFAFermionInstantiationZWilsonImplD2.cc

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

Grid/qcd/action/fermion/instantiation/ZWilsonImplD2/PartialFractionFermion5DInstantiationZWilsonImplD2.cc

@@ -1 +0,0 @@
-../PartialFractionFermion5DInstantiation.cc.master

Grid/qcd/action/fermion/instantiation/ZWilsonImplD2/WilsonFermion5DInstantiationZWilsonImplD2.cc

@@ -1 +0,0 @@
-../WilsonFermion5DInstantiation.cc.master

Grid/qcd/action/fermion/instantiation/ZWilsonImplD2/impl.h

@@ -1 +0,0 @@
-#define IMPLEMENTATION ZWilsonImplD2

Grid/qcd/action/fermion/instantiation/generate_instantiations.sh

@@ -10,12 +10,18 @@ WILSON_IMPL_LIST=" \
 	   WilsonImplF \
 	   WilsonImplD \
 	   WilsonImplD2 \
+	   SpWilsonImplF \
+	   SpWilsonImplD \
 	   WilsonAdjImplF \
 	   WilsonAdjImplD \
 	   WilsonTwoIndexSymmetricImplF \
 	   WilsonTwoIndexSymmetricImplD \
 	   WilsonTwoIndexAntiSymmetricImplF \
 	   WilsonTwoIndexAntiSymmetricImplD \
+	   SpWilsonTwoIndexAntiSymmetricImplF \
+	   SpWilsonTwoIndexAntiSymmetricImplD \
+	   SpWilsonTwoIndexSymmetricImplF \
+	   SpWilsonTwoIndexSymmetricImplD \
 	   GparityWilsonImplF \
 	   GparityWilsonImplD "