Optimise lie algebra project

BLAS_benchmark/BatchBlasBench.cc

@@ -12,13 +12,15 @@
 #include <iostream>
 #include <sys/time.h>
 
+#define GRID_SYCL
+#undef  GRID_HIP
+#undef  GRID_CUDA
 
 #ifdef GRID_HIP
 #include <hipblas/hipblas.h>
 #endif
 #ifdef GRID_CUDA
 #include <cublas_v2.h>
-
 #endif
 #ifdef GRID_SYCL
 #include <oneapi/mkl.hpp>
@@ -43,90 +45,6 @@ inline void acceleratorFreeDevice(void *ptr,size_t bytes){free(ptr,*theAccelerat
 inline void acceleratorMemSet(void *base,int value,size_t bytes) { theAccelerator->memset(base,value,bytes); theAccelerator->wait();}
 inline void acceleratorCopyToDevice(void *from,void *to,size_t bytes)  { theAccelerator->memcpy(to,from,bytes); theAccelerator->wait();}
 inline void acceleratorCopyFromDevice(void *from,void *to,size_t bytes){ theAccelerator->memcpy(to,from,bytes); theAccelerator->wait();}
-#define accelerator_barrier(dummy) { theAccelerator->wait(); }
-#endif
-
-#ifdef GRID_HIP
-hipStream_t copyStream;
-hipStream_t computeStream;
-void acceleratorInit(void)
-{
-  int device = 0;
-  auto discard = hipSetDevice(device);
-  discard = hipStreamCreate(&copyStream);
-  discard = hipStreamCreate(&computeStream);
-  printf("AcceleratorHIPInit\n");
-}
-inline void *acceleratorAllocDevice(size_t bytes)
-{
-  void *ptr=NULL;
-  auto err = hipMalloc((void **)&ptr,bytes);
-  if( err != hipSuccess ) {
-    ptr = (void *) NULL;
-    fprintf(stderr," hipMalloc failed for %ld %s \n",bytes,hipGetErrorString(err)); fflush(stderr);
-  }
-  return ptr;
-};
-inline void acceleratorFreeDevice(void *ptr,size_t bytes){ auto discard=hipFree(ptr);};
-inline void acceleratorFreeDevice(void *ptr){ auto discard=hipFree(ptr);};
-inline void acceleratorMemSet(void *base,int value,size_t bytes) { auto discard=hipMemset(base,value,bytes);}
-inline void acceleratorCopyToDevice(void *from,void *to,size_t bytes)  { auto discard=hipMemcpy(to,from,bytes, hipMemcpyHostToDevice);}
-inline void acceleratorCopyFromDevice(void *from,void *to,size_t bytes){ auto discard=hipMemcpy(to,from,bytes, hipMemcpyDeviceToHost);}
-#define accelerator_barrier(dummy)				\
-  {								\
-    auto tmp=hipStreamSynchronize(computeStream);		\
-    auto err = hipGetLastError();				\
-    if ( err != hipSuccess ) {					\
-      printf("After hipDeviceSynchronize() : HIP error %s \n", hipGetErrorString( err )); \
-      puts(__FILE__);							\
-      printf("Line %d\n",__LINE__);				\
-      exit(0);							\
-    }								\
-  }
-
-#endif
-
-#ifdef GRID_CUDA
-cudaStream_t copyStream;
-cudaStream_t computeStream;
-void acceleratorInit(void)
-{
-  int device = 0;
-  cudaSetDevice(device);
-  cudaStreamCreate(&copyStream);
-  cudaStreamCreate(&computeStream);
-}
-inline void *acceleratorAllocDevice(size_t bytes)
-{
-  void *ptr=NULL;
-  auto err = cudaMalloc((void **)&ptr,bytes);
-  if( err != cudaSuccess ) {
-    ptr = (void *) NULL;
-    printf(" cudaMalloc failed for %d %s \n",bytes,cudaGetErrorString(err));
-  }
-  return ptr;
-};
-inline void acceleratorFreeShared(void *ptr){ cudaFree(ptr);};
-inline void acceleratorFreeDevice(void *ptr){ cudaFree(ptr);};
-inline void acceleratorCopyToDevice(void *from,void *to,size_t bytes)  { cudaMemcpy(to,from,bytes, cudaMemcpyHostToDevice);}
-inline void acceleratorCopyFromDevice(void *from,void *to,size_t bytes){ cudaMemcpy(to,from,bytes, cudaMemcpyDeviceToHost);}
-inline void acceleratorMemSet(void *base,int value,size_t bytes) { cudaMemset(base,value,bytes);}
-#define accelerator_barrier(dummy)					\
-  {									\
-    cudaStreamSynchronize(computeStream);				\
-    cudaError err = cudaGetLastError();					\
-    if ( cudaSuccess != err ) {						\
-      printf("accelerator_barrier(): Cuda error %s \n",			\
-	     cudaGetErrorString( err ));				\
-      printf("File %s Line %d\n",__FILE__,__LINE__);			\
-      fflush(stdout);							\
-      if (acceleratorAbortOnGpuError) assert(err==cudaSuccess);		\
-    }									\
-  }
-
-#endif
-
-
 template<class T> void acceleratorPut(T& dev,T&host)
 {
   acceleratorCopyToDevice(&host,&dev,sizeof(T));
@@ -137,6 +55,9 @@ template<class T> T acceleratorGet(T& dev)
   acceleratorCopyFromDevice(&dev,&host,sizeof(T));
   return host;
 }
+#define accelerator_barrier(dummy) { theAccelerator->wait(); }
+#endif
+
 
 /**************************************************************
  * Allocator
@@ -290,269 +211,6 @@ public:
 #endif
   }
   
-
-  /////////////////////////////////////////////////////////////
-  // Single matrix GEMM -- fp64 and fp32
-  /////////////////////////////////////////////////////////////
-  void gemm(GridBLASOperation_t OpA,
-	    GridBLASOperation_t OpB,
-	    int m,int n, int k,
-	    ComplexD alpha,
-	    ComplexD* Amk,  // Device pointer
-	    ComplexD* Bkn,
-	    ComplexD beta,
-	    ComplexD* Cmn)
-  {
-    RealD t2=usecond();
-
-    assert(OpA!=GridBLAS_OP_T); // Complex case expect no transpose
-    assert(OpB!=GridBLAS_OP_T);
-
-    int lda = m; // m x k column major
-    int ldb = k; // k x n column major
-    int ldc = m; // m x b column major
-    if(OpA!=GridBLAS_OP_N)
-      lda = k;
-    if(OpB!=GridBLAS_OP_N)
-      ldb = n;
-    
-    static deviceVector<ComplexD> alpha_p(1);
-    static deviceVector<ComplexD> beta_p(1);
-    // can prestore the 1 and the zero on device
-    acceleratorCopyToDevice((void *)&alpha,(void *)&alpha_p[0],sizeof(ComplexD));
-    acceleratorCopyToDevice((void *)&beta ,(void *)&beta_p[0],sizeof(ComplexD));
-    RealD t0=usecond();
-
-#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 = hipblasZgemm(gridblasHandle,
-			    hOpA,
-			    hOpB,
-			    m,n,k,
-			    (hipblasDoubleComplex *) &alpha_p[0],
-			    (hipblasDoubleComplex *) Amk, lda,
-			    (hipblasDoubleComplex *) Bkn, ldb,
-			    (hipblasDoubleComplex *) &beta_p[0],
-			    (hipblasDoubleComplex *) Cmn, ldc);
-    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 = cublasZgemm(gridblasHandle,
-			   hOpA,
-			   hOpB,
-			   m,n,k,
-			   (cuDoubleComplex *) &alpha_p[0],
-			   (cuDoubleComplex *) Amk, lda,
-			   (cuDoubleComplex *) Bkn, ldb,
-			   (cuDoubleComplex *) &beta_p[0],
-			   (cuDoubleComplex *) Cmn, ldc);
-    assert(err==CUBLAS_STATUS_SUCCESS);
-#endif
-#ifdef GRID_SYCL
-      int64_t m64=m;
-      int64_t n64=n;
-      int64_t k64=k;
-      int64_t lda64=lda;
-      int64_t ldb64=ldb;
-      int64_t ldc64=ldc;
-
-      oneapi::mkl::transpose iOpA;
-      oneapi::mkl::transpose iOpB;
-      
-      if ( OpA == GridBLAS_OP_N ) iOpA = oneapi::mkl::transpose::N;
-      if ( OpA == GridBLAS_OP_T ) iOpA = oneapi::mkl::transpose::T;
-      if ( OpA == GridBLAS_OP_C ) iOpA = oneapi::mkl::transpose::C;
-      if ( OpB == GridBLAS_OP_N ) iOpB = oneapi::mkl::transpose::N;
-      if ( OpB == GridBLAS_OP_T ) iOpB = oneapi::mkl::transpose::T;
-      if ( OpB == GridBLAS_OP_C ) iOpB = oneapi::mkl::transpose::C;
-
-      oneapi::mkl::blas::column_major::gemm(*gridblasHandle,
-					    iOpA,
-					    iOpB,
-					    m64,n64,k64,
-					    (ComplexD *) &alpha_p[0],
-					    (const ComplexD *)Amk, (int64_t )lda64,
-					    (const ComplexD *)Bkn, (int64_t )ldb64,
-					    (ComplexD *) &beta_p[0],
-					    (ComplexD *)Cmn, (int64_t)ldc64);
-      synchronise();
-#endif
-#if !defined(GRID_SYCL) && !defined(GRID_CUDA) && !defined(GRID_HIP)
-    // Need a default/reference implementation; use Eigen
-      if ( (OpA == GridBLAS_OP_N ) && (OpB == GridBLAS_OP_N) ) {
-	Eigen::Map<Eigen::MatrixXcd> eAmk(Amk,m,k);
-	Eigen::Map<Eigen::MatrixXcd> eBkn(Bkn,k,n);
-	Eigen::Map<Eigen::MatrixXcd> eCmn(Cmn,m,n);
-	eCmn = beta * eCmn + alpha * eAmk * eBkn ;
-      } else if ( (OpA == GridBLAS_OP_C ) && (OpB == GridBLAS_OP_N) ) {
-	Eigen::Map<Eigen::MatrixXcd> eAmk(Amk,k,m);
-	Eigen::Map<Eigen::MatrixXcd> eBkn(Bkn,k,n);
-	Eigen::Map<Eigen::MatrixXcd> eCmn(Cmn,m,n);
-	eCmn = beta * eCmn + alpha * eAmk.adjoint() * eBkn ;
-      } else if ( (OpA == GridBLAS_OP_N ) && (OpB == GridBLAS_OP_C) ) {
-	Eigen::Map<Eigen::MatrixXcd> eAmk(Amk,m,k);
-	Eigen::Map<Eigen::MatrixXcd> eBkn(Bkn,n,k);
-	Eigen::Map<Eigen::MatrixXcd> eCmn(Cmn,m,n);
-	eCmn = beta * eCmn + alpha * eAmk * eBkn.adjoint() ;
-      } else if ( (OpA == GridBLAS_OP_C ) && (OpB == GridBLAS_OP_C) ) {
-	Eigen::Map<Eigen::MatrixXcd> eAmk(Amk,k,m);
-	Eigen::Map<Eigen::MatrixXcd> eBkn(Bkn,n,k);
-	Eigen::Map<Eigen::MatrixXcd> eCmn(Cmn,m,n);
-	eCmn = beta * eCmn + alpha * eAmk.adjoint() * eBkn.adjoint() ;
-      } else { 
-	assert(0);
-      }
-#endif
-     RealD t1=usecond();
-     RealD flops = 8.0*m*n*k;
-     RealD bytes = 1.0*sizeof(ComplexD)*(m*k+k*n+m*n);
-  }
-  void gemm(GridBLASOperation_t OpA,
-	    GridBLASOperation_t OpB,
-	    int m,int n, int k,
-	    ComplexF alpha,
-	    ComplexF* Amk,  // Device pointer
-	    ComplexF* Bkn,
-	    ComplexF beta,
-	    ComplexF* Cmn)
-  {
-    RealD t2=usecond();
-
-    assert(OpA!=GridBLAS_OP_T); // Complex case expect no transpose
-    assert(OpB!=GridBLAS_OP_T);
-
-    int lda = m; // m x k column major
-    int ldb = k; // k x n column major
-    int ldc = m; // m x b column major
-    if(OpA!=GridBLAS_OP_N)
-      lda = k;
-    if(OpB!=GridBLAS_OP_N)
-      ldb = n;
-    
-    static deviceVector<ComplexF> alpha_p(1);
-    static deviceVector<ComplexF> beta_p(1);
-    // can prestore the 1 and the zero on device
-    acceleratorCopyToDevice((void *)&alpha,(void *)&alpha_p[0],sizeof(ComplexF));
-    acceleratorCopyToDevice((void *)&beta ,(void *)&beta_p[0],sizeof(ComplexF));
-    RealD t0=usecond();
-
-#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 = hipblasCgemm(gridblasHandle,
-			    hOpA,
-			    hOpB,
-			    m,n,k,
-			    (hipblasComplex *) &alpha_p[0],
-			    (hipblasComplex *) Amk, lda,
-			    (hipblasComplex *) Bkn, ldb,
-			    (hipblasComplex *) &beta_p[0],
-			    (hipblasComplex *) Cmn, ldc);
-    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 = cublasCgemm(gridblasHandle,
-			   hOpA,
-			   hOpB,
-			   m,n,k,
-			   (cuComplex *) &alpha_p[0],
-			   (cuComplex *) Amk, lda,
-			   (cuComplex *) Bkn, ldb,
-			   (cuComplex *) &beta_p[0],
-			   (cuComplex *) Cmn, ldc);
-    assert(err==CUBLAS_STATUS_SUCCESS);
-#endif
-#ifdef GRID_SYCL
-      int64_t m64=m;
-      int64_t n64=n;
-      int64_t k64=k;
-      int64_t lda64=lda;
-      int64_t ldb64=ldb;
-      int64_t ldc64=ldc;
-
-      oneapi::mkl::transpose iOpA;
-      oneapi::mkl::transpose iOpB;
-      
-      if ( OpA == GridBLAS_OP_N ) iOpA = oneapi::mkl::transpose::N;
-      if ( OpA == GridBLAS_OP_T ) iOpA = oneapi::mkl::transpose::T;
-      if ( OpA == GridBLAS_OP_C ) iOpA = oneapi::mkl::transpose::C;
-      if ( OpB == GridBLAS_OP_N ) iOpB = oneapi::mkl::transpose::N;
-      if ( OpB == GridBLAS_OP_T ) iOpB = oneapi::mkl::transpose::T;
-      if ( OpB == GridBLAS_OP_C ) iOpB = oneapi::mkl::transpose::C;
-
-      oneapi::mkl::blas::column_major::gemm(*gridblasHandle,
-					    iOpA,
-					    iOpB,
-					    m64,n64,k64,
-					    (ComplexF *) &alpha_p[0],
-					    (const ComplexF *)Amk, (int64_t )lda64,
-					    (const ComplexF *)Bkn, (int64_t )ldb64,
-					    (ComplexF *) &beta_p[0],
-					    (ComplexF *)Cmn, (int64_t )ldc64);
-      synchronise();
-#endif
-#if !defined(GRID_SYCL) && !defined(GRID_CUDA) && !defined(GRID_HIP)
-    // Need a default/reference implementation; use Eigen
-      if ( (OpA == GridBLAS_OP_N ) && (OpB == GridBLAS_OP_N) ) {
-	Eigen::Map<Eigen::MatrixXcf> eAmk(Amk,m,k);
-	Eigen::Map<Eigen::MatrixXcf> eBkn(Bkn,k,n);
-	Eigen::Map<Eigen::MatrixXcf> eCmn(Cmn,m,n);
-	eCmn = beta * eCmn + alpha * eAmk * eBkn ;
-      } else if ( (OpA == GridBLAS_OP_C ) && (OpB == GridBLAS_OP_N) ) {
-	Eigen::Map<Eigen::MatrixXcf> eAmk(Amk,k,m);
-	Eigen::Map<Eigen::MatrixXcf> eBkn(Bkn,k,n);
-	Eigen::Map<Eigen::MatrixXcf> eCmn(Cmn,m,n);
-	eCmn = beta * eCmn + alpha * eAmk.adjoint() * eBkn ;
-      } else if ( (OpA == GridBLAS_OP_N ) && (OpB == GridBLAS_OP_C) ) {
-	Eigen::Map<Eigen::MatrixXcf> eAmk(Amk,m,k);
-	Eigen::Map<Eigen::MatrixXcf> eBkn(Bkn,n,k);
-	Eigen::Map<Eigen::MatrixXcf> eCmn(Cmn,m,n);
-	eCmn = beta * eCmn + alpha * eAmk * eBkn.adjoint() ;
-      } else if ( (OpA == GridBLAS_OP_C ) && (OpB == GridBLAS_OP_C) ) {
-	Eigen::Map<Eigen::MatrixXcf> eAmk(Amk,k,m);
-	Eigen::Map<Eigen::MatrixXcf> eBkn(Bkn,n,k);
-	Eigen::Map<Eigen::MatrixXcf> eCmn(Cmn,m,n);
-	eCmn = beta * eCmn + alpha * eAmk.adjoint() * eBkn.adjoint() ;
-      } else { 
-	assert(0);
-      }
-#endif
-     RealD t1=usecond();
-     RealD flops = 8.0*m*n*k;
-     RealD bytes = 1.0*sizeof(ComplexF)*(m*k+k*n+m*n);
-  }
-
-  
-  /////////////////////////////////////////////////////////////
   void gemmBatched(int m,int n, int k,
 		   ComplexD alpha,
 		   deviceVector<ComplexD*> &Amk,  // pointer list to matrices
@@ -583,6 +241,36 @@ public:
 		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,
@@ -936,6 +624,301 @@ public:
      RealD bytes = 1.0*sizeof(ComplexF)*(m*k+k*n+m*n)*batchCount;
   }
   
+  ///////////////////////////////////////////////////////////////////////////
+  // Single precision real GEMM
+  ///////////////////////////////////////////////////////////////////////////
+
+  void gemmBatched(GridBLASOperation_t OpA,
+		   GridBLASOperation_t OpB,
+		   int m,int n, int k,
+		   RealF alpha,
+		   deviceVector<RealF*> &Amk,  // pointer list to matrices
+		   deviceVector<RealF*> &Bkn,
+		   RealF beta,
+		   deviceVector<RealF*> &Cmn)
+  {
+    RealD t2=usecond();
+    int32_t batchCount = Amk.size();
+
+    assert(OpA!=GridBLAS_OP_C); // Real case no conjugate
+    assert(OpB!=GridBLAS_OP_C);
+
+    int lda = m; // m x k column major
+    int ldb = k; // k x n column major
+    int ldc = m; // m x b column major
+    if(OpA!=GridBLAS_OP_N)
+      lda = k;
+    if(OpB!=GridBLAS_OP_N)
+      ldb = n;
+    static deviceVector<RealF> alpha_p(1);
+    static deviceVector<RealF> beta_p(1);
+    // can prestore the 1 and the zero on device
+    acceleratorCopyToDevice((void *)&alpha,(void *)&alpha_p[0],sizeof(RealF));
+    acceleratorCopyToDevice((void *)&beta ,(void *)&beta_p[0],sizeof(RealF));
+    RealD t0=usecond();
+
+    assert(Bkn.size()==batchCount);
+    assert(Cmn.size()==batchCount);
+#ifdef GRID_HIP
+    hipblasOperation_t hOpA;
+    hipblasOperation_t hOpB;
+    if ( OpA == GridBLAS_OP_N ) hOpA = HIPBLAS_OP_N;
+    if ( OpA == GridBLAS_OP_T ) hOpA = HIPBLAS_OP_T;
+    if ( OpA == GridBLAS_OP_C ) hOpA = HIPBLAS_OP_C;
+    if ( OpB == GridBLAS_OP_N ) hOpB = HIPBLAS_OP_N;
+    if ( OpB == GridBLAS_OP_T ) hOpB = HIPBLAS_OP_T;
+    if ( OpB == GridBLAS_OP_C ) hOpB = HIPBLAS_OP_C;
+    auto err = hipblasSgemmBatched(gridblasHandle,
+				   hOpA,
+				   hOpB,
+				   m,n,k,
+				   (float *) &alpha_p[0],
+				   (float **)&Amk[0], lda,
+				   (float **)&Bkn[0], ldb,
+				   (float *) &beta_p[0],
+				   (float **)&Cmn[0], ldc,
+				   batchCount);
+    assert(err==HIPBLAS_STATUS_SUCCESS);
+#endif
+#ifdef GRID_CUDA
+    cublasOperation_t hOpA;
+    cublasOperation_t hOpB;
+    if ( OpA == GridBLAS_OP_N ) hOpA = CUBLAS_OP_N;
+    if ( OpA == GridBLAS_OP_T ) hOpA = CUBLAS_OP_T;
+    if ( OpA == GridBLAS_OP_C ) hOpA = CUBLAS_OP_C;
+    if ( OpB == GridBLAS_OP_N ) hOpB = CUBLAS_OP_N;
+    if ( OpB == GridBLAS_OP_T ) hOpB = CUBLAS_OP_T;
+    if ( OpB == GridBLAS_OP_C ) hOpB = CUBLAS_OP_C;
+    auto err = cublasSgemmBatched(gridblasHandle,
+				  hOpA,
+				  hOpB,
+				  m,n,k,
+				  (float *) &alpha_p[0],
+				  (float **)&Amk[0], lda,
+				  (float **)&Bkn[0], ldb,
+				  (float *) &beta_p[0],
+				  (float **)&Cmn[0], ldc,
+				  batchCount);
+    assert(err==CUBLAS_STATUS_SUCCESS);
+#endif
+#ifdef GRID_SYCL
+      int64_t m64=m;
+      int64_t n64=n;
+      int64_t k64=k;
+      int64_t lda64=lda;
+      int64_t ldb64=ldb;
+      int64_t ldc64=ldc;
+      int64_t batchCount64=batchCount;
+
+      oneapi::mkl::transpose iOpA;
+      oneapi::mkl::transpose iOpB;
+      
+      if ( OpA == GridBLAS_OP_N ) iOpA = oneapi::mkl::transpose::N;
+      if ( OpA == GridBLAS_OP_T ) iOpA = oneapi::mkl::transpose::T;
+      if ( OpA == GridBLAS_OP_C ) iOpA = oneapi::mkl::transpose::C;
+      if ( OpB == GridBLAS_OP_N ) iOpB = oneapi::mkl::transpose::N;
+      if ( OpB == GridBLAS_OP_T ) iOpB = oneapi::mkl::transpose::T;
+      if ( OpB == GridBLAS_OP_C ) iOpB = oneapi::mkl::transpose::C;
+
+      oneapi::mkl::blas::column_major::gemm_batch(*gridblasHandle,
+						  &iOpA,
+						  &iOpB,
+						  &m64,&n64,&k64,
+						  (float *) &alpha_p[0],
+						  (const float **)&Amk[0], (const int64_t *)&lda64,
+						  (const float **)&Bkn[0], (const int64_t *)&ldb64,
+						  (float *) &beta_p[0],
+						  (float **)&Cmn[0], (const int64_t *)&ldc64,
+						  (int64_t)1,&batchCount64,std::vector<sycl::event>());
+      synchronise();
+#endif
+#if !defined(GRID_SYCL) && !defined(GRID_CUDA) && !defined(GRID_HIP)
+    // Need a default/reference implementation; use Eigen
+      if ( (OpA == GridBLAS_OP_N ) && (OpB == GridBLAS_OP_N) ) {
+	thread_for (p, batchCount, {
+	  Eigen::Map<Eigen::MatrixXf> eAmk(Amk[p],m,k);
+	  Eigen::Map<Eigen::MatrixXf> eBkn(Bkn[p],k,n);
+	  Eigen::Map<Eigen::MatrixXf> eCmn(Cmn[p],m,n);
+	  eCmn = beta * eCmn + alpha * eAmk * eBkn ;
+	  });
+      } else if ( (OpA == GridBLAS_OP_T ) && (OpB == GridBLAS_OP_N) ) {
+	thread_for (p, batchCount, {
+	  Eigen::Map<Eigen::MatrixXf> eAmk(Amk[p],k,m);
+	  Eigen::Map<Eigen::MatrixXf> eBkn(Bkn[p],k,n);
+	  Eigen::Map<Eigen::MatrixXf> eCmn(Cmn[p],m,n);
+	  eCmn = beta * eCmn + alpha * eAmk.transpose() * eBkn ;
+	  });
+      } else if ( (OpA == GridBLAS_OP_N ) && (OpB == GridBLAS_OP_T) ) {
+	thread_for (p, batchCount, {
+	  Eigen::Map<Eigen::MatrixXf> eAmk(Amk[p],m,k);
+	  Eigen::Map<Eigen::MatrixXf> eBkn(Bkn[p],n,k);
+	  Eigen::Map<Eigen::MatrixXf> eCmn(Cmn[p],m,n);
+	  eCmn = beta * eCmn + alpha * eAmk * eBkn.transpose() ;
+	  });
+      } else if ( (OpA == GridBLAS_OP_T ) && (OpB == GridBLAS_OP_T) ) {
+	thread_for (p, batchCount, {
+	  Eigen::Map<Eigen::MatrixXf> eAmk(Amk[p],k,m);
+	  Eigen::Map<Eigen::MatrixXf> eBkn(Bkn[p],n,k);
+	  Eigen::Map<Eigen::MatrixXf> eCmn(Cmn[p],m,n);
+	  eCmn = beta * eCmn + alpha * eAmk.transpose() * eBkn.transpose() ;
+	  } );
+      } else { 
+	assert(0);
+      }
+#endif
+     RealD t1=usecond();
+     RealD flops = 2.0*m*n*k*batchCount;
+     RealD bytes = 1.0*sizeof(RealF)*(m*k+k*n+m*n)*batchCount;
+  }
+  
+  
+  ///////////////////////////////////////////////////////////////////////////
+  // Double precision real GEMM
+  ///////////////////////////////////////////////////////////////////////////
+  void gemmBatched(GridBLASOperation_t OpA,
+		   GridBLASOperation_t OpB,
+		   int m,int n, int k,
+		   RealD alpha,
+		   deviceVector<RealD*> &Amk,  // pointer list to matrices
+		   deviceVector<RealD*> &Bkn,
+		   RealD beta,
+		   deviceVector<RealD*> &Cmn)
+  {
+    RealD t2=usecond();
+    int32_t batchCount = Amk.size();
+
+    assert(OpA!=GridBLAS_OP_C); // Real case no conjugate
+    assert(OpB!=GridBLAS_OP_C);
+
+    int lda = m; // m x k column major
+    int ldb = k; // k x n column major
+    int ldc = m; // m x b column major
+    if(OpA!=GridBLAS_OP_N)
+      lda = k;
+    if(OpB!=GridBLAS_OP_N)
+      ldb = n;
+    
+    static deviceVector<RealD> alpha_p(1);
+    static deviceVector<RealD> beta_p(1);
+    // can prestore the 1 and the zero on device
+    acceleratorCopyToDevice((void *)&alpha,(void *)&alpha_p[0],sizeof(RealD));
+    acceleratorCopyToDevice((void *)&beta ,(void *)&beta_p[0],sizeof(RealD));
+    RealD t0=usecond();
+
+    assert(Bkn.size()==batchCount);
+    assert(Cmn.size()==batchCount);
+#ifdef GRID_HIP
+    hipblasOperation_t hOpA;
+    hipblasOperation_t hOpB;
+    if ( OpA == GridBLAS_OP_N ) hOpA = HIPBLAS_OP_N;
+    if ( OpA == GridBLAS_OP_T ) hOpA = HIPBLAS_OP_T;
+    if ( OpA == GridBLAS_OP_C ) hOpA = HIPBLAS_OP_C;
+    if ( OpB == GridBLAS_OP_N ) hOpB = HIPBLAS_OP_N;
+    if ( OpB == GridBLAS_OP_T ) hOpB = HIPBLAS_OP_T;
+    if ( OpB == GridBLAS_OP_C ) hOpB = HIPBLAS_OP_C;
+    auto err = hipblasDgemmBatched(gridblasHandle,
+				   HIPBLAS_OP_N,
+				   HIPBLAS_OP_N,
+				   m,n,k,
+				   (double *) &alpha_p[0],
+				   (double **)&Amk[0], lda,
+				   (double **)&Bkn[0], ldb,
+				   (double *) &beta_p[0],
+				   (double **)&Cmn[0], ldc,
+				   batchCount);
+    assert(err==HIPBLAS_STATUS_SUCCESS);
+#endif
+#ifdef GRID_CUDA
+    cublasOperation_t hOpA;
+    cublasOperation_t hOpB;
+    if ( OpA == GridBLAS_OP_N ) hOpA = CUBLAS_OP_N;
+    if ( OpA == GridBLAS_OP_T ) hOpA = CUBLAS_OP_T;
+    if ( OpA == GridBLAS_OP_C ) hOpA = CUBLAS_OP_C;
+    if ( OpB == GridBLAS_OP_N ) hOpB = CUBLAS_OP_N;
+    if ( OpB == GridBLAS_OP_T ) hOpB = CUBLAS_OP_T;
+    if ( OpB == GridBLAS_OP_C ) hOpB = CUBLAS_OP_C;
+    auto err = cublasDgemmBatched(gridblasHandle,
+				  hOpA,
+				  hOpB,
+				  m,n,k,
+				  (double *) &alpha_p[0],
+				  (double **)&Amk[0], lda,
+				  (double **)&Bkn[0], ldb,
+				  (double *) &beta_p[0],
+				  (double **)&Cmn[0], ldc,
+				  batchCount);
+    assert(err==CUBLAS_STATUS_SUCCESS);
+#endif
+#ifdef GRID_SYCL
+      int64_t m64=m;
+      int64_t n64=n;
+      int64_t k64=k;
+      int64_t lda64=lda;
+      int64_t ldb64=ldb;
+      int64_t ldc64=ldc;
+      int64_t batchCount64=batchCount;
+
+      oneapi::mkl::transpose iOpA;
+      oneapi::mkl::transpose iOpB;
+      
+      if ( OpA == GridBLAS_OP_N ) iOpA = oneapi::mkl::transpose::N;
+      if ( OpA == GridBLAS_OP_T ) iOpA = oneapi::mkl::transpose::T;
+      if ( OpA == GridBLAS_OP_C ) iOpA = oneapi::mkl::transpose::C;
+      if ( OpB == GridBLAS_OP_N ) iOpB = oneapi::mkl::transpose::N;
+      if ( OpB == GridBLAS_OP_T ) iOpB = oneapi::mkl::transpose::T;
+      if ( OpB == GridBLAS_OP_C ) iOpB = oneapi::mkl::transpose::C;
+
+      oneapi::mkl::blas::column_major::gemm_batch(*gridblasHandle,
+						  &iOpA,
+						  &iOpB,
+						  &m64,&n64,&k64,
+						  (double *) &alpha_p[0],
+						  (const double **)&Amk[0], (const int64_t *)&lda64,
+						  (const double **)&Bkn[0], (const int64_t *)&ldb64,
+						  (double *) &beta_p[0],
+						  (double **)&Cmn[0], (const int64_t *)&ldc64,
+						  (int64_t)1,&batchCount64,std::vector<sycl::event>());
+      synchronise();
+#endif
+#if !defined(GRID_SYCL) && !defined(GRID_CUDA) && !defined(GRID_HIP)
+    // Need a default/reference implementation; use Eigen
+      if ( (OpA == GridBLAS_OP_N ) && (OpB == GridBLAS_OP_N) ) {
+	thread_for (p, batchCount, {
+	  Eigen::Map<Eigen::MatrixXd> eAmk(Amk[p],m,k);
+	  Eigen::Map<Eigen::MatrixXd> eBkn(Bkn[p],k,n);
+	  Eigen::Map<Eigen::MatrixXd> eCmn(Cmn[p],m,n);
+	  eCmn = beta * eCmn + alpha * eAmk * eBkn ;
+	  });
+      } else if ( (OpA == GridBLAS_OP_T ) && (OpB == GridBLAS_OP_N) ) {
+	thread_for (p, batchCount, {
+	  Eigen::Map<Eigen::MatrixXd> eAmk(Amk[p],k,m);
+	  Eigen::Map<Eigen::MatrixXd> eBkn(Bkn[p],k,n);
+	  Eigen::Map<Eigen::MatrixXd> eCmn(Cmn[p],m,n);
+	  eCmn = beta * eCmn + alpha * eAmk.transpose() * eBkn ;
+	  });
+      } else if ( (OpA == GridBLAS_OP_N ) && (OpB == GridBLAS_OP_T) ) {
+	thread_for (p, batchCount, {
+	  Eigen::Map<Eigen::MatrixXd> eAmk(Amk[p],m,k);
+	  Eigen::Map<Eigen::MatrixXd> eBkn(Bkn[p],n,k);
+	  Eigen::Map<Eigen::MatrixXd> eCmn(Cmn[p],m,n);
+	  eCmn = beta * eCmn + alpha * eAmk * eBkn.transpose() ;
+	  });
+      } else if ( (OpA == GridBLAS_OP_T ) && (OpB == GridBLAS_OP_T) ) {
+	thread_for (p, batchCount, {
+	  Eigen::Map<Eigen::MatrixXd> eAmk(Amk[p],k,m);
+	  Eigen::Map<Eigen::MatrixXd> eBkn(Bkn[p],n,k);
+	  Eigen::Map<Eigen::MatrixXd> eCmn(Cmn[p],m,n);
+	  eCmn = beta * eCmn + alpha * eAmk.transpose() * eBkn.transpose() ;
+	  });
+      } else { 
+	assert(0);
+      }
+#endif
+     RealD t1=usecond();
+     RealD flops = 2.0*m*n*k*batchCount;
+     RealD bytes = 1.0*sizeof(RealD)*(m*k+k*n+m*n)*batchCount;
+  }
+
   template<class CComplex>
   double benchmark(int M, int N, int K, int BATCH)
   {
@@ -984,47 +967,6 @@ public:
     return flops; // Returns gigaflops
   }
 
-  template<class CComplex>
-  double benchmark(int M, int N, int K)
-  {
-    int32_t N_A = M*K;
-    int32_t N_B = K*N;
-    int32_t N_C = M*N;
-    deviceVector<CComplex> A(N_A); acceleratorMemSet(&A[0],0,N_A*sizeof(CComplex));
-    deviceVector<CComplex> B(N_B); acceleratorMemSet(&B[0],0,N_B*sizeof(CComplex));
-    deviceVector<CComplex> C(N_C); acceleratorMemSet(&C[0],0,N_C*sizeof(CComplex));
-    CComplex alpha(1.0);
-    CComplex beta (1.0);
-    RealD flops = 8.0*M*N*K;
-    int ncall=10;
-
-    gemm(GridBLAS_OP_C,GridBLAS_OP_N,
-	 M,N,K,
-	 alpha,
-	 &A[0], // m x k 
-	 &B[0], // k x n
-	 beta, 
-	 &C[0]);
-    synchronise();
-
-    RealD t0 = usecond();
-    for(int i=0;i<ncall;i++){
-      gemm(GridBLAS_OP_N,GridBLAS_OP_N,
-	   M,N,K,
-	   alpha,
-	   &A[0], // m x k 
-	   &B[0], // k x n
-	   beta, 
-	   &C[0]);
-      synchronise();
-    }
-    RealD t1 = usecond();
-    RealD bytes = 1.0*sizeof(CComplex)*(M*N*2+N*K+M*K);
-    flops = 8.0*M*N*K*ncall;
-    flops = flops/(t1-t0)/1.e3;
-    return flops; // Returns gigaflops
-  }
-
 };
 
 
@@ -1093,21 +1035,6 @@ static void BLAS(void)
       std::cout<< M<<"\t\t"<<N<<"\t\t"<<K<<"\t\t"<<BATCH<<"\t\t"<<p<<std::endl;
     }}
   fprintf(FP,"\n\n\n");
-
-  std::cout << "----------------------------------------------------------"<<std::endl;
-  std::cout << "  M  "<<"\t\t"<<"N"<<"\t\t\t"<<"K"<<"\t\t"<<"Gflop/s / rank (inner product matrix)"<<std::endl;
-  std::cout << "----------------------------------------------------------"<<std::endl;
-  {
-    int M=12;
-    int N=12;
-    std::vector<int> ks({4*1024*1024, 2*1024*1024, 1024*1024, 256*1024, 1024 });
-    for( int kk=0;kk<ks.size();kk++ ) {
-      int K = ks[kk];
-      double p=blas.benchmark<CComplex>(M,N,K);
-      fprintf(FP,"%d, %d, %d, %d, %f\n", M, N, K, 1, p);
-      std::cout<< M<<"\t\t"<<N<<"\t\t"<<K<<"\t\t"<<1<<"\t\t"<<p<<std::endl;
-    }
-  }
   std::cout << "=================================================================================="<<std::endl;
 };
 

BLAS_benchmark/compile-command

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

BLAS_benchmark/compile-command-frontier

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

BLAS_benchmark/compile-command-sunspot

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

Grid/algorithms/Algorithms.h

@@ -50,7 +50,6 @@ NAMESPACE_CHECK(approx);
 #include <Grid/algorithms/deflation/Deflation.h>
 #include <Grid/algorithms/deflation/MultiRHSBlockProject.h>
 #include <Grid/algorithms/deflation/MultiRHSDeflation.h>
-#include <Grid/algorithms/deflation/MultiRHSBlockCGLinalg.h>
 NAMESPACE_CHECK(deflation);
 #include <Grid/algorithms/iterative/ConjugateGradient.h>
 NAMESPACE_CHECK(ConjGrad);

Grid/algorithms/FFT.h

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

Grid/algorithms/LinearOperator.h

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

Grid/algorithms/SparseMatrix.h

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

Grid/algorithms/approx/Chebyshev.h

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

Grid/algorithms/deflation/MultiRHSBlockCGLinalg.h

@@ -1,376 +0,0 @@
-/*************************************************************************************
-
-    Grid physics library, www.github.com/paboyle/Grid 
-
-    Source file: MultiRHSBlockCGLinalg.h
-
-    Copyright (C) 2024
-
-Author: Peter Boyle <pboyle@bnl.gov>
-
-    This program is free software; you can redistribute it and/or modify
-    it under the terms of the GNU General Public License as published by
-    the Free Software Foundation; either version 2 of the License, or
-    (at your option) any later version.
-
-    This program is distributed in the hope that it will be useful,
-    but WITHOUT ANY WARRANTY; without even the implied warranty of
-    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
-    GNU General Public License for more details.
-
-    You should have received a copy of the GNU General Public License along
-    with this program; if not, write to the Free Software Foundation, Inc.,
-    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
-
-    See the full license in the file "LICENSE" in the top level distribution directory
-*************************************************************************************/
-/*  END LEGAL */
-#pragma once
-
-NAMESPACE_BEGIN(Grid);
-
-
-/* Need helper object for BLAS accelerated mrhs blockCG */
-template<class Field>
-class MultiRHSBlockCGLinalg
-{
-public:
-
-  typedef typename Field::scalar_type   scalar;
-  typedef typename Field::scalar_object scalar_object;
-  typedef typename Field::vector_object vector_object;
-
-  deviceVector<scalar> BLAS_X;      // nrhs x vol -- the sources
-  deviceVector<scalar> BLAS_Y;      // nrhs x vol -- the result
-  deviceVector<scalar> BLAS_C;      // nrhs x nrhs -- the coefficients 
-  deviceVector<scalar> BLAS_Cred;   // nrhs x nrhs x oSites -- reduction buffer
-  deviceVector<scalar *> Xdip;
-  deviceVector<scalar *> Ydip;
-  deviceVector<scalar *> Cdip;
-  
-  MultiRHSBlockCGLinalg() {};
-  ~MultiRHSBlockCGLinalg(){ Deallocate(); };
-  
-  void Deallocate(void)
-  {
-    Xdip.resize(0);
-    Ydip.resize(0);
-    Cdip.resize(0);
-    BLAS_Cred.resize(0);
-    BLAS_C.resize(0);
-    BLAS_X.resize(0);
-    BLAS_Y.resize(0);
-  }
-  void MaddMatrix(std::vector<Field> &AP, Eigen::MatrixXcd &m , const std::vector<Field> &X,const std::vector<Field> &Y,RealD scale=1.0)
-  {
-    std::vector<Field> Y_copy(AP.size(),AP[0].Grid());
-    for(int r=0;r<AP.size();r++){
-      Y_copy[r] = Y[r];
-    }
-    MulMatrix(AP,m,X);
-    for(int r=0;r<AP.size();r++){
-      AP[r] = scale*AP[r]+Y_copy[r];
-    }
-  }
-  void MulMatrix(std::vector<Field> &Y, Eigen::MatrixXcd &m , const std::vector<Field> &X)
-  {
-    typedef typename Field::scalar_type scomplex;
-    GridBase *grid;
-    uint64_t vol;
-    uint64_t words;
-
-    int nrhs = Y.size();
-    grid  = X[0].Grid();
-    vol   = grid->lSites();
-    words = sizeof(scalar_object)/sizeof(scalar);
-    int64_t vw = vol * words;
-
-    RealD t0 = usecond();
-    BLAS_X.resize(nrhs * vw); // cost free if size doesn't change
-    BLAS_Y.resize(nrhs * vw); // cost free if size doesn't change
-    BLAS_C.resize(nrhs * nrhs);// cost free if size doesn't change
-    RealD t1 = usecond();
-
-    /////////////////////////////////////////////
-    // Copy in the multi-rhs sources
-    /////////////////////////////////////////////
-    for(int r=0;r<nrhs;r++){
-      int64_t offset = r*vw;
-      autoView(x_v,X[r],AcceleratorRead);
-      acceleratorCopyDeviceToDevice(&x_v[0],&BLAS_X[offset],sizeof(scalar_object)*vol);
-    }
-
-    // Assumes Eigen storage contiguous
-    acceleratorCopyToDevice(&m(0,0),&BLAS_C[0],BLAS_C.size()*sizeof(scalar));
-    
-  /*
-   * in Fortran column major notation (cuBlas order)
-   *
-   * Xxr = [X1(x)][..][Xn(x)]
-   * Yxr = [Y1(x)][..][Ym(x)]
-   * Y = X . C
-   */
-    deviceVector<scalar *> Xd(1);
-    deviceVector<scalar *> Yd(1);
-    deviceVector<scalar *> Cd(1);
-
-    scalar * Xh = & BLAS_X[0];
-    scalar * Yh = & BLAS_Y[0];
-    scalar * Ch = & BLAS_C[0];
-
-    acceleratorPut(Xd[0],Xh);
-    acceleratorPut(Yd[0],Yh);
-    acceleratorPut(Cd[0],Ch);
-
-    RealD t2 = usecond();
-    GridBLAS BLAS;
-    /////////////////////////////////////////
-    // Y = X*C (transpose?)
-    /////////////////////////////////////////
-    BLAS.gemmBatched(GridBLAS_OP_N,GridBLAS_OP_N, 
-    		     vw,nrhs,nrhs,
-		     scalar(1.0),
-		     Xd,
-		     Cd,
-		     scalar(0.0),  // wipe out Y
-		     Yd);
-    BLAS.synchronise();
-    RealD t3 = usecond();
-
-    // Copy back Y = m X 
-    for(int r=0;r<nrhs;r++){
-      int64_t offset = r*vw;
-      autoView(y_v,Y[r],AcceleratorWrite);
-      acceleratorCopyDeviceToDevice(&BLAS_Y[offset],&y_v[0],sizeof(scalar_object)*vol);
-    }    
-    RealD t4 = usecond();
-    std::cout << "MulMatrix alloc    took "<< t1-t0<<" us"<<std::endl;
-    std::cout << "MulMatrix preamble took "<< t2-t1<<" us"<<std::endl;
-    std::cout << "MulMatrix blas     took "<< t3-t2<<" us"<<std::endl;
-    std::cout << "MulMatrix copy     took "<< t4-t3<<" us"<<std::endl;
-    std::cout << "MulMatrix total "<< t4-t0<<" us"<<std::endl;
-  }
-  
-  void InnerProductMatrix(Eigen::MatrixXcd &m , const std::vector<Field> &X, const std::vector<Field> &Y)
-  {
-#if 0    
-    int nrhs;
-    GridBase *grid;
-    uint64_t vol;
-    uint64_t words;
-
-    nrhs = X.size();
-    assert(X.size()==Y.size());
-    conformable(X[0],Y[0]);
-
-    grid  = X[0].Grid();
-    vol   = grid->lSites();
-    words = sizeof(scalar_object)/sizeof(scalar);
-    int64_t vw = vol * words;
-
-    RealD t0 = usecond();
-    BLAS_X.resize(nrhs * vw); // cost free if size doesn't change
-    BLAS_Y.resize(nrhs * vw); // cost free if size doesn't change
-    BLAS_C.resize(nrhs * nrhs);// cost free if size doesn't change
-    RealD t1 = usecond();
-
-    /////////////////////////////////////////////
-    // Copy in the multi-rhs sources
-    /////////////////////////////////////////////
-    for(int r=0;r<nrhs;r++){
-      int64_t offset = r*vw;
-      autoView(x_v,X[r],AcceleratorRead);
-      acceleratorCopyDeviceToDevice(&x_v[0],&BLAS_X[offset],sizeof(scalar_object)*vol);
-      autoView(y_v,Y[r],AcceleratorRead);
-      acceleratorCopyDeviceToDevice(&y_v[0],&BLAS_Y[offset],sizeof(scalar_object)*vol);
-    }
-    RealD t2 = usecond();
-
-  /*
-   * in Fortran column major notation (cuBlas order)
-   *
-   * Xxr = [X1(x)][..][Xn(x)]
-   *
-   * Yxr = [Y1(x)][..][Ym(x)]
-   *
-   * C_rs = X^dag Y
-   */
-    deviceVector<scalar *> Xd(1);
-    deviceVector<scalar *> Yd(1);
-    deviceVector<scalar *> Cd(1);
-
-    scalar * Xh = & BLAS_X[0];
-    scalar * Yh = & BLAS_Y[0];
-    scalar * Ch = & BLAS_C[0];
-
-    acceleratorPut(Xd[0],Xh);
-    acceleratorPut(Yd[0],Yh);
-    acceleratorPut(Cd[0],Ch);
-
-    GridBLAS BLAS;
-
-    RealD t3 = usecond();
-    /////////////////////////////////////////
-    // C_rs = X^dag Y
-    /////////////////////////////////////////
-    BLAS.gemmBatched(GridBLAS_OP_C,GridBLAS_OP_N, 
-    		     nrhs,nrhs,vw,
-		     ComplexD(1.0),
-		     Xd,
-		     Yd,
-		     ComplexD(0.0),  // wipe out C
-		     Cd);
-    BLAS.synchronise();
-    RealD t4 = usecond();
-
-    std::vector<scalar> HOST_C(BLAS_C.size());      // nrhs . nrhs -- the coefficients 
-    acceleratorCopyFromDevice(&BLAS_C[0],&HOST_C[0],BLAS_C.size()*sizeof(scalar));
-    grid->GlobalSumVector(&HOST_C[0],nrhs*nrhs);
-
-    RealD t5 = usecond();
-    for(int rr=0;rr<nrhs;rr++){
-      for(int r=0;r<nrhs;r++){
-	int off = r+nrhs*rr;
-	m(r,rr)=HOST_C[off];
-      }
-    }
-    RealD t6 = usecond();
-    uint64_t M=nrhs;
-    uint64_t N=nrhs;
-    uint64_t K=vw;
-    RealD bytes = 1.0*sizeof(ComplexD)*(M*N*2+N*K+M*K);
-    RealD flops = 8.0*M*N*K;
-    flops = flops/(t4-t3)/1.e3;
-    bytes = bytes/(t4-t3)/1.e3;
-    std::cout << "InnerProductMatrix m,n,k "<< M<<","<<N<<","<<K<<std::endl;
-    std::cout << "InnerProductMatrix alloc t1 "<< t1-t0<<" us"<<std::endl;
-    std::cout << "InnerProductMatrix cp    t2 "<< t2-t1<<" us"<<std::endl;
-    std::cout << "InnerProductMatrix setup t3 "<< t3-t2<<" us"<<std::endl;
-    std::cout << "InnerProductMatrix blas t4 "<< t4-t3<<" us"<<std::endl;
-    std::cout << "InnerProductMatrix blas    "<< flops<<" GF/s"<<std::endl;
-    std::cout << "InnerProductMatrix blas    "<< bytes<<" GB/s"<<std::endl;
-    std::cout << "InnerProductMatrix gsum t5 "<< t5-t4<<" us"<<std::endl;
-    std::cout << "InnerProductMatrix cp   t6 "<< t6-t5<<" us"<<std::endl;
-    std::cout << "InnerProductMatrix took "<< t6-t0<<" us"<<std::endl;
-#else
-    int nrhs;
-    GridBase *grid;
-    uint64_t vol;
-    uint64_t words;
-
-    nrhs = X.size();
-    assert(X.size()==Y.size());
-    conformable(X[0],Y[0]);
-
-    grid  = X[0].Grid();
-    int rd0 =  grid->_rdimensions[0] * grid->_rdimensions[1];
-    vol   = grid->oSites()/rd0;
-    words = rd0*sizeof(vector_object)/sizeof(scalar);
-    int64_t vw = vol * words;
-    assert(vw == grid->lSites()*sizeof(scalar_object)/sizeof(scalar));
-
-    RealD t0 = usecond();
-    BLAS_X.resize(nrhs * vw); // cost free if size doesn't change
-    BLAS_Y.resize(nrhs * vw); // cost free if size doesn't change
-    BLAS_Cred.resize(nrhs * nrhs * vol);// cost free if size doesn't change
-    RealD t1 = usecond();
-
-    /////////////////////////////////////////////
-    // Copy in the multi-rhs sources -- layout batched BLAS ready
-    /////////////////////////////////////////////
-    for(int r=0;r<nrhs;r++){
-      autoView(x_v,X[r],AcceleratorRead);
-      autoView(y_v,Y[r],AcceleratorRead);
-      scalar *from_x=(scalar *)&x_v[0];
-      scalar *from_y=(scalar *)&y_v[0];
-      scalar *BX = &BLAS_X[0];
-      scalar *BY = &BLAS_Y[0];
-      accelerator_for(ssw,vw,1,{
-	  uint64_t ss=ssw/words;
-	  uint64_t  w=ssw%words;
-	  uint64_t offset = w+r*words+ss*nrhs*words; // [ss][rhs][words]
-	  BX[offset] = from_x[ssw];
-	  BY[offset] = from_y[ssw];
-	});
-    }
-    RealD t2 = usecond();
-
-  /*
-   * in Fortran column major notation (cuBlas order)
-   *
-   * Xxr = [X1(x)][..][Xn(x)]
-   *
-   * Yxr = [Y1(x)][..][Ym(x)]
-   *
-   * C_rs = X^dag Y
-   */
-    Xdip.resize(vol);
-    Ydip.resize(vol);
-    Cdip.resize(vol);
-    std::vector<scalar *> Xh(vol);
-    std::vector<scalar *> Yh(vol);
-    std::vector<scalar *> Ch(vol);
-    for(uint64_t ss=0;ss<vol;ss++){
-
-      Xh[ss] = & BLAS_X[ss*nrhs*words];
-      Yh[ss] = & BLAS_Y[ss*nrhs*words];
-      Ch[ss] = & BLAS_Cred[ss*nrhs*nrhs];
-
-    }
-    acceleratorCopyToDevice(&Xh[0],&Xdip[0],vol*sizeof(scalar *));
-    acceleratorCopyToDevice(&Yh[0],&Ydip[0],vol*sizeof(scalar *));
-    acceleratorCopyToDevice(&Ch[0],&Cdip[0],vol*sizeof(scalar *));
-    
-    GridBLAS BLAS;
-
-    RealD t3 = usecond();
-    /////////////////////////////////////////
-    // C_rs = X^dag Y
-    /////////////////////////////////////////
-    BLAS.gemmBatched(GridBLAS_OP_C,GridBLAS_OP_N, 
-    		     nrhs,nrhs,words,
-		     ComplexD(1.0),
-		     Xdip,
-		     Ydip,
-		     ComplexD(0.0),  // wipe out C
-		     Cdip);
-    BLAS.synchronise();
-    RealD t4 = usecond();
-
-    std::vector<scalar> HOST_C(BLAS_Cred.size());      // nrhs . nrhs -- the coefficients 
-    acceleratorCopyFromDevice(&BLAS_Cred[0],&HOST_C[0],BLAS_Cred.size()*sizeof(scalar));
-
-    RealD t5 = usecond();
-    m = Eigen::MatrixXcd::Zero(nrhs,nrhs);
-    for(int ss=0;ss<vol;ss++){
-      Eigen::Map<Eigen::MatrixXcd> eC((std::complex<double> *)&HOST_C[ss*nrhs*nrhs],nrhs,nrhs);
-      m = m + eC;
-    }
-    RealD t6l = usecond();
-    grid->GlobalSumVector((scalar *) &m(0,0),nrhs*nrhs);
-    RealD t6 = usecond();
-    uint64_t M=nrhs;
-    uint64_t N=nrhs;
-    uint64_t K=vw;
-    RealD xybytes = grid->lSites()*sizeof(scalar_object);
-    RealD bytes = 1.0*sizeof(ComplexD)*(M*N*2+N*K+M*K);
-    RealD flops = 8.0*M*N*K;
-    flops = flops/(t4-t3)/1.e3;
-    bytes = bytes/(t4-t3)/1.e3;
-    xybytes = 4*xybytes/(t2-t1)/1.e3;
-    std::cout << "InnerProductMatrix m,n,k "<< M<<","<<N<<","<<K<<std::endl;
-    std::cout << "InnerProductMatrix alloc t1 "<< t1-t0<<" us"<<std::endl;
-    std::cout << "InnerProductMatrix cp    t2 "<< t2-t1<<" us "<<xybytes<<" GB/s"<<std::endl;
-    std::cout << "InnerProductMatrix setup t3 "<< t3-t2<<" us"<<std::endl;
-    std::cout << "InnerProductMatrix blas t4 "<< t4-t3<<" us"<<std::endl;
-    std::cout << "InnerProductMatrix blas    "<< flops<<" GF/s"<<std::endl;
-    std::cout << "InnerProductMatrix blas    "<< bytes<<" GB/s"<<std::endl;
-    std::cout << "InnerProductMatrix cp     t5 "<< t5-t4<<" us"<<std::endl;
-    std::cout << "InnerProductMatrix lsum   t6l "<< t6l-t5<<" us"<<std::endl;
-    std::cout << "InnerProductMatrix gsum   t6 "<< t6-t6l<<" us"<<std::endl;
-    std::cout << "InnerProductMatrix took "<< t6-t0<<" us"<<std::endl;
-#endif
-  }
-};
-
-NAMESPACE_END(Grid);

Grid/algorithms/deflation/MultiRHSBlockProject.h

@@ -447,10 +447,10 @@ public:
     /////////////////////////////////////////
     BLAS.gemmBatched(GridBLAS_OP_C,GridBLAS_OP_N, 
     		     nbasis,nrhs,vw,
-		     scalar(1.0),
+		     ComplexD(1.0),
 		     Vd,
 		     Fd,
-		     scalar(0.0),  // wipe out C
+		     ComplexD(0.0),  // wipe out C
 		     Cd);
     BLAS.synchronise();
     //    std::cout << "BlockProject done"<<std::endl;
@@ -497,10 +497,10 @@ public:
     int64_t vw = block_vol * words;
     BLAS.gemmBatched(GridBLAS_OP_N,GridBLAS_OP_N, 
     		     vw,nrhs,nbasis,
-		     scalar(1.0),
+		     ComplexD(1.0),
 		     Vd,
 		     Cd,
-		     scalar(0.0),  // wipe out C
+		     ComplexD(0.0),  // wipe out C
 		     Fd);
     BLAS.synchronise();
     //    std::cout << " blas call done"<<std::endl;

Grid/algorithms/deflation/MultiRHSDeflation.h

@@ -182,10 +182,10 @@ public:
     /////////////////////////////////////////
     BLAS.gemmBatched(GridBLAS_OP_C,GridBLAS_OP_N, 
     		     nev,nrhs,vw,
-		     scalar(1.0),
+		     ComplexD(1.0),
 		     Ed,
 		     Rd,
-		     scalar(0.0),  // wipe out C
+		     ComplexD(0.0),  // wipe out C
 		     Cd);
     BLAS.synchronise();
 
@@ -210,10 +210,10 @@ public:
     /////////////////////////////////////////
     BLAS.gemmBatched(GridBLAS_OP_N,GridBLAS_OP_N, 
 		     vw,nrhs,nev,
-		     scalar(1.0),
+		     ComplexD(1.0),
 		     Ed, // x . nev
 		     Cd, // nev . nrhs
-		     scalar(0.0),
+		     ComplexD(0.0),
 		     Gd);
     BLAS.synchronise();
 

Grid/algorithms/iterative/AdefMrhs.h

@@ -53,7 +53,6 @@ class TwoLevelCGmrhs
   // Fine operator, Smoother, CoarseSolver
   LinearOperatorBase<Field>   &_FineLinop;
   LinearFunction<Field>   &_Smoother;
-  MultiRHSBlockCGLinalg<Field> _BlockCGLinalg;
 
   GridStopWatch ProjectTimer;
   GridStopWatch PromoteTimer;
@@ -80,301 +79,6 @@ class TwoLevelCGmrhs
   
   // Vector case
   virtual void operator() (std::vector<Field> &src, std::vector<Field> &x)
-  {
-    SolveSingleSystem(src,x);
-    //    SolvePrecBlockCG(src,x);
-  }
-
-////////////////////////////////////////////////////////////////////////////////////////////////////
-// Thin QR factorisation (google it)
-////////////////////////////////////////////////////////////////////////////////////////////////////
-  ////////////////////////////////////////////////////////////////////////////////////////////////////
-  //Dimensions
-  // R_{ferm x Nblock} =  Q_{ferm x Nblock} x  C_{Nblock x Nblock} -> ferm x Nblock
-  //
-  // Rdag R = m_rr = Herm = L L^dag        <-- Cholesky decomposition (LLT routine in Eigen)
-  //
-  //   Q  C = R => Q = R C^{-1}
-  //
-  // Want  Ident = Q^dag Q = C^{-dag} R^dag R C^{-1} = C^{-dag} L L^dag C^{-1} = 1_{Nblock x Nblock} 
-  //
-  // Set C = L^{dag}, and then Q^dag Q = ident 
-  //
-  // Checks:
-  // Cdag C = Rdag R ; passes.
-  // QdagQ  = 1      ; passes
-  ////////////////////////////////////////////////////////////////////////////////////////////////////
-  void ThinQRfact (Eigen::MatrixXcd &m_zz,
-		   Eigen::MatrixXcd &C,
-		   Eigen::MatrixXcd &Cinv,
-		   std::vector<Field> &  Q,
-		   std::vector<Field> & MQ,
-		   const std::vector<Field> & Z,
-		   const std::vector<Field> & MZ)
-  {
-    RealD t0=usecond();
-    _BlockCGLinalg.InnerProductMatrix(m_zz,MZ,Z);
-    RealD t1=usecond();
-
-    m_zz = 0.5*(m_zz+m_zz.adjoint());
-    
-    Eigen::MatrixXcd L    = m_zz.llt().matrixL(); 
-    
-    C    = L.adjoint();
-    Cinv = C.inverse();
-    
-    RealD t3=usecond();
-    _BlockCGLinalg.MulMatrix( Q,Cinv,Z);
-    _BlockCGLinalg.MulMatrix(MQ,Cinv,MZ);
-    RealD t4=usecond();
-    std::cout << " ThinQRfact IP    :"<< t1-t0<<" us"<<std::endl;
-    std::cout << " ThinQRfact Eigen :"<< t3-t1<<" us"<<std::endl;
-    std::cout << " ThinQRfact MulMat:"<< t4-t3<<" us"<<std::endl;
-  }
-
-  virtual void SolvePrecBlockCG (std::vector<Field> &src, std::vector<Field> &X)
-  {
-    std::cout << GridLogMessage<<"HDCG: mrhs fPrecBlockcg starting"<<std::endl;
-    src[0].Grid()->Barrier();
-    int nrhs = src.size();
-    //    std::vector<RealD> f(nrhs);
-    //    std::vector<RealD> rtzp(nrhs);
-    //    std::vector<RealD> rtz(nrhs);
-    //    std::vector<RealD> a(nrhs);
-    //    std::vector<RealD> d(nrhs);
-    //    std::vector<RealD> b(nrhs);
-    //    std::vector<RealD> rptzp(nrhs);
-
-    ////////////////////////////////////////////
-    //Initial residual computation & set up
-    ////////////////////////////////////////////
-    std::vector<RealD> ssq(nrhs);
-    for(int rhs=0;rhs<nrhs;rhs++){
-      ssq[rhs]=norm2(src[rhs]); assert(ssq[rhs]!=0.0);
-    }      
-
-    ///////////////////////////
-    // Fields -- eliminate duplicates between fPcg and block cg
-    ///////////////////////////
-    std::vector<Field> Mtmp(nrhs,grid);
-    std::vector<Field> tmp(nrhs,grid);
-    std::vector<Field>   Z(nrhs,grid); // Rename Z to R
-    std::vector<Field>  MZ(nrhs,grid); // Rename MZ to Z
-    std::vector<Field>   Q(nrhs,grid); // 
-    std::vector<Field>  MQ(nrhs,grid); // Rename to P
-    std::vector<Field>   D(nrhs,grid);
-    std::vector<Field>  AD(nrhs,grid);
-    
-    /************************************************************************
-     * Preconditioned Block conjugate gradient rQ
-     * Generalise Sebastien Birk Thesis, after Dubrulle 2001.
-     * Introduce preconditioning following Saad Ch9
-     ************************************************************************
-     * Dimensions:
-     *
-     *   X,B etc... ==(Nferm x nrhs)
-     *  Matrix A==(Nferm x Nferm)
-     *  
-     * Nferm = Nspin x Ncolour x Ncomplex x Nlattice_site
-     * QC => Thin QR factorisation (google it)
-     *
-     * R = B-AX
-     * Z = Mi R
-     * QC = Z
-     * D = Q 
-     * for k: 
-     *   R  = AD
-     *   Z  = Mi R
-     *   M  = [D^dag R]^{-1}
-     *   X  = X + D M C
-     *   QS = Q - Z.M
-     *   D  = Q + D S^dag
-     *   C  = S C
-     */
-    Eigen::MatrixXcd m_DZ     = Eigen::MatrixXcd::Identity(nrhs,nrhs);
-    Eigen::MatrixXcd m_M      = Eigen::MatrixXcd::Identity(nrhs,nrhs);
-    Eigen::MatrixXcd m_zz     = Eigen::MatrixXcd::Zero(nrhs,nrhs);
-    Eigen::MatrixXcd m_rr     = Eigen::MatrixXcd::Zero(nrhs,nrhs);
-    
-    Eigen::MatrixXcd m_C      = Eigen::MatrixXcd::Zero(nrhs,nrhs);
-    Eigen::MatrixXcd m_Cinv   = Eigen::MatrixXcd::Zero(nrhs,nrhs);
-    Eigen::MatrixXcd m_S      = Eigen::MatrixXcd::Zero(nrhs,nrhs);
-    Eigen::MatrixXcd m_Sinv   = Eigen::MatrixXcd::Zero(nrhs,nrhs);
-    
-    Eigen::MatrixXcd m_tmp    = Eigen::MatrixXcd::Identity(nrhs,nrhs);
-    Eigen::MatrixXcd m_tmp1   = Eigen::MatrixXcd::Identity(nrhs,nrhs);
-
-    GridStopWatch HDCGTimer;
-
-    //////////////////////////
-    // x0 = Vstart -- possibly modify guess
-    //////////////////////////
-    Vstart(X,src);
-
-    //////////////////////////
-    // R = B-AX
-    //////////////////////////
-    for(int rhs=0;rhs<nrhs;rhs++){
-      // r0 = b -A x0
-      _FineLinop.HermOp(X[rhs],tmp[rhs]);
-      axpy (Z[rhs], -1.0,tmp[rhs], src[rhs]);    // Computes R=Z=src - A X0
-    }
-
-    //////////////////////////////////
-    // Compute MZ = M1 Z = M1 B - M1 A x0
-    //////////////////////////////////
-    PcgM1(Z,MZ);  
-
-    //////////////////////////////////
-    // QC = Z
-    //////////////////////////////////
-    ThinQRfact (m_zz, m_C, m_Cinv, Q, MQ, Z, MZ);
-
-    //////////////////////////////////
-    // D=MQ
-    //////////////////////////////////
-    for(int b=0;b<nrhs;b++) D[b]=MQ[b]; // LLT rotation of the MZ basis of search dirs
-
-    std::cout << GridLogMessage<<"PrecBlockCGrQ vec computed initial residual and QR fact " <<std::endl;
-
-    ProjectTimer.Reset();
-    PromoteTimer.Reset();
-    DeflateTimer.Reset();
-    CoarseTimer.Reset();
-    SmoothTimer.Reset();
-    FineTimer.Reset();
-    InsertTimer.Reset();
-
-    GridStopWatch M1Timer;
-    GridStopWatch M2Timer;
-    GridStopWatch M3Timer;
-    GridStopWatch LinalgTimer;
-    GridStopWatch InnerProdTimer;
-
-    HDCGTimer.Start();
-
-    std::vector<RealD> rn(nrhs);
-    for (int k=0;k<=MaxIterations;k++){
-
-      ////////////////////
-      // Z  = AD
-      ////////////////////
-      M3Timer.Start();
-      for(int b=0;b<nrhs;b++) _FineLinop.HermOp(D[b], Z[b]);      
-      M3Timer.Stop();
-
-      ////////////////////
-      // MZ  = M1 Z <==== the Multigrid preconditioner
-      ////////////////////
-      M1Timer.Start();
-      PcgM1(Z,MZ);
-      M1Timer.Stop();
-
-      FineTimer.Start();
-      ////////////////////
-      // M  = [D^dag Z]^{-1} = (<Ddag MZ>_M)^{-1} inner prod, generalising Saad derivation of Precon CG
-      ////////////////////
-      InnerProdTimer.Start();
-      _BlockCGLinalg.InnerProductMatrix(m_DZ,D,Z);
-      InnerProdTimer.Stop();
-      m_M       = m_DZ.inverse();
-
-      ///////////////////////////
-      // X  = X + D MC
-      ///////////////////////////
-      m_tmp     = m_M * m_C;
-      LinalgTimer.Start();
-      _BlockCGLinalg.MaddMatrix(X,m_tmp, D,X);     // D are the search directions and X takes the updates 
-      LinalgTimer.Stop();
-
-      ///////////////////////////
-      // QS = Q - M Z
-      // (MQ) S = MQ - M (M1Z)
-      ///////////////////////////
-      LinalgTimer.Start();
-      _BlockCGLinalg.MaddMatrix(tmp ,m_M, Z, Q,-1.0);
-      _BlockCGLinalg.MaddMatrix(Mtmp,m_M,MZ,MQ,-1.0);
-      ThinQRfact (m_zz, m_S, m_Sinv, Q, MQ, tmp, Mtmp);
-      LinalgTimer.Stop();
-
-      ////////////////////////////
-      // D  = MQ + D S^dag
-      ////////////////////////////
-      m_tmp = m_S.adjoint();
-      LinalgTimer.Start();
-      _BlockCGLinalg.MaddMatrix(D,m_tmp,D,MQ);
-      LinalgTimer.Stop();
-
-      ////////////////////////////
-      // C  = S C
-      ////////////////////////////
-      m_C = m_S*m_C;
-      
-      ////////////////////////////
-      // convergence monitor
-      ////////////////////////////
-      m_rr = m_C.adjoint() * m_C;
-      
-      FineTimer.Stop();
-
-      RealD max_resid=0;
-      RealD rrsum=0;
-      RealD sssum=0;
-      RealD rr;
-
-      for(int b=0;b<nrhs;b++) {
-	rrsum+=real(m_rr(b,b));
-	sssum+=ssq[b];
-	rr = real(m_rr(b,b))/ssq[b];
-	if ( rr > max_resid ) max_resid = rr;
-      }
-      std::cout << GridLogMessage <<
-	  "\t Prec BlockCGrQ Iteration "<<k<<" ave resid "<< std::sqrt(rrsum/sssum) << " max "<< std::sqrt(max_resid) <<std::endl;
-
-
-      if ( max_resid < Tolerance*Tolerance ) { 
-
-	HDCGTimer.Stop();
-	std::cout<<GridLogMessage<<"HDCG: mrhs PrecBlockCGrQ converged in "<<k<<" iterations and "<<HDCGTimer.Elapsed()<<std::endl;;
-	std::cout<<GridLogMessage<<"HDCG: mrhs PrecBlockCGrQ : Linalg  "<<LinalgTimer.Elapsed()<<std::endl;;
-	std::cout<<GridLogMessage<<"HDCG: mrhs PrecBlockCGrQ : fine H  "<<M3Timer.Elapsed()<<std::endl;;
-	std::cout<<GridLogMessage<<"HDCG: mrhs PrecBlockCGrQ : prec M1 "<<M1Timer.Elapsed()<<std::endl;;
-	std::cout<<GridLogMessage<<"**** M1 breakdown:"<<std::endl;
-	std::cout<<GridLogMessage<<"HDCG: mrhs PrecBlockCGrQ : Project "<<ProjectTimer.Elapsed()<<std::endl;;
-	std::cout<<GridLogMessage<<"HDCG: mrhs PrecBlockCGrQ : Promote "<<PromoteTimer.Elapsed()<<std::endl;;
-	std::cout<<GridLogMessage<<"HDCG: mrhs PrecBlockCGrQ : Deflate "<<DeflateTimer.Elapsed()<<std::endl;;
-	std::cout<<GridLogMessage<<"HDCG: mrhs PrecBlockCGrQ : Coarse  "<<CoarseTimer.Elapsed()<<std::endl;;
-	std::cout<<GridLogMessage<<"HDCG: mrhs PrecBlockCGrQ : Fine    "<<FineTimer.Elapsed()<<std::endl;;
-	std::cout<<GridLogMessage<<"HDCG: mrhs PrecBlockCGrQ : Smooth  "<<SmoothTimer.Elapsed()<<std::endl;;
-	std::cout<<GridLogMessage<<"HDCG: mrhs PrecBlockCGrQ : Insert  "<<InsertTimer.Elapsed()<<std::endl;;
-
-	for(int rhs=0;rhs<nrhs;rhs++){
-
-	  _FineLinop.HermOp(X[rhs],tmp[rhs]);			  
-
-	  Field mytmp(grid);
-	  axpy(mytmp,-1.0,src[rhs],tmp[rhs]);
-      
-	  RealD  xnorm   = sqrt(norm2(X[rhs]));
-	  RealD  srcnorm = sqrt(norm2(src[rhs]));
-	  RealD  tmpnorm = sqrt(norm2(mytmp));
-	  RealD  true_residual = tmpnorm/srcnorm;
-	  std::cout<<GridLogMessage
-		   <<"HDCG: true residual ["<<rhs<<"] is "<<true_residual
-		   <<" solution "<<xnorm
-		   <<" source "<<srcnorm
-		   <<std::endl;
-	}
-	return;
-      }
-      
-    }
-    HDCGTimer.Stop();
-    std::cout<<GridLogMessage<<"HDCG: PrecBlockCGrQ not converged "<<HDCGTimer.Elapsed()<<std::endl;
-    assert(0);
-  }
-
-  virtual void SolveSingleSystem (std::vector<Field> &src, std::vector<Field> &x)
   {
     std::cout << GridLogMessage<<"HDCG: mrhs fPcg starting"<<std::endl;
     src[0].Grid()->Barrier();
@@ -657,23 +361,15 @@ public:
     CoarseField PleftProjMrhs(this->coarsegridmrhs);
     CoarseField PleftMss_projMrhs(this->coarsegridmrhs);
 
-#undef SMOOTHER_BLOCK_SOLVE
-#if SMOOTHER_BLOCK_SOLVE
-    this->SmoothTimer.Start();
-    this->_Smoother(in,Min);
-    this->SmoothTimer.Stop();
-#else
     for(int rhs=0;rhs<nrhs;rhs++) {
+
       this->SmoothTimer.Start();
       this->_Smoother(in[rhs],Min[rhs]);
       this->SmoothTimer.Stop();
-    }
-#endif
-    
-    for(int rhs=0;rhs<nrhs;rhs++) {
 
       this->FineTimer.Start();
       this->_FineLinop.HermOp(Min[rhs],out[rhs]);
+
       axpy(tmp[rhs],-1.0,out[rhs],in[rhs]);          // resid  = in - A Min
       this->FineTimer.Stop();
 

Grid/algorithms/iterative/BlockConjugateGradient.h

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

Grid/algorithms/iterative/ConjugateGradient.h

@@ -38,7 +38,6 @@ NAMESPACE_BEGIN(Grid);
 // single input vec, single output vec.
 /////////////////////////////////////////////////////////////
 
-
 template <class Field>
 class ConjugateGradient : public OperatorFunction<Field> {
 public:
@@ -58,22 +57,10 @@ public:
       ErrorOnNoConverge(err_on_no_conv)
   {};
 
-  virtual void LogIteration(int k,RealD a,RealD b){
-    //    std::cout << "ConjugageGradient::LogIteration() "<<std::endl;
-  };
-  virtual void LogBegin(void){
-    std::cout << "ConjugageGradient::LogBegin() "<<std::endl;
-  };
-
   void operator()(LinearOperatorBase<Field> &Linop, const Field &src, Field &psi) {
 
-      this->LogBegin();
-
     GRID_TRACE("ConjugateGradient");
     GridStopWatch PreambleTimer;
-    GridStopWatch ConstructTimer;
-    GridStopWatch NormTimer;
-    GridStopWatch AssignTimer;
     PreambleTimer.Start();
     psi.Checkerboard() = src.Checkerboard();
 
@@ -83,19 +70,14 @@ public:
     //RealD b_pred;
 
     // Was doing copies
-    ConstructTimer.Start();
     Field p(src.Grid());
     Field mmp(src.Grid());
     Field r(src.Grid());
-    ConstructTimer.Stop();
 
     // Initial residual computation & set up
-    NormTimer.Start();
     ssq = norm2(src);
     RealD guess = norm2(psi);
-    NormTimer.Stop();
     assert(std::isnan(guess) == 0);
-    AssignTimer.Start();
     if ( guess == 0.0 ) {
       r = src;
       p = r;
@@ -107,7 +89,6 @@ public:
       a = norm2(p);
     }
     cp = a;
-    AssignTimer.Stop();
 
     // Handle trivial case of zero src
     if (ssq == 0.){
@@ -183,7 +164,6 @@ public:
       }
       LinearCombTimer.Stop();
       LinalgTimer.Stop();
-      LogIteration(k,a,b);
 
       IterationTimer.Stop();
       if ( (k % 500) == 0 ) {
@@ -240,9 +220,6 @@ public:
     	      <<" residual "<< std::sqrt(cp / ssq)<< std::endl;
     SolverTimer.Stop();
     std::cout << GridLogMessage << "\tPreamble   " << PreambleTimer.Elapsed() <<std::endl;
-    std::cout << GridLogMessage << "\tConstruct  " << ConstructTimer.Elapsed() <<std::endl;
-    std::cout << GridLogMessage << "\tNorm       " << NormTimer.Elapsed() <<std::endl;
-    std::cout << GridLogMessage << "\tAssign     " << AssignTimer.Elapsed() <<std::endl;
     std::cout << GridLogMessage << "\tSolver     " << SolverTimer.Elapsed() <<std::endl;
     std::cout << GridLogMessage << "Solver breakdown "<<std::endl;
     std::cout << GridLogMessage << "\tMatrix     " << MatrixTimer.Elapsed() <<std::endl;
@@ -256,118 +233,5 @@ public:
 
   }
 };
-
-
-template <class Field>
-class ConjugateGradientPolynomial : public ConjugateGradient<Field> {
-public:
-  // Optionally record the CG polynomial
-  std::vector<double> ak;
-  std::vector<double> bk;
-  std::vector<double> poly_p;
-  std::vector<double> poly_r;
-  std::vector<double> poly_Ap;
-  std::vector<double> polynomial;
-
-public:
-  ConjugateGradientPolynomial(RealD tol, Integer maxit, bool err_on_no_conv = true)
-    : ConjugateGradient<Field>(tol,maxit,err_on_no_conv)
-  { };
-  void PolyHermOp(LinearOperatorBase<Field> &Linop, const Field &src, Field &psi)
-  {
-    Field tmp(src.Grid());
-    Field AtoN(src.Grid());
-    AtoN = src;
-    psi=AtoN*polynomial[0];
-    for(int n=1;n<polynomial.size();n++){
-      tmp = AtoN;
-      Linop.HermOp(tmp,AtoN);
-      psi = psi + polynomial[n]*AtoN;
-    }
-  }
-  void CGsequenceHermOp(LinearOperatorBase<Field> &Linop, const Field &src, Field &x)
-  {
-    Field Ap(src.Grid());
-    Field r(src.Grid());
-    Field p(src.Grid());
-    p=src;
-    r=src;
-    x=Zero();
-    x.Checkerboard()=src.Checkerboard();
-    for(int k=0;k<ak.size();k++){
-      x = x + ak[k]*p;
-      Linop.HermOp(p,Ap);
-      r = r - ak[k] * Ap;
-      p = r + bk[k] * p;
-    }
-  }
-  void Solve(LinearOperatorBase<Field> &Linop, const Field &src, Field &psi)
-  {
-    psi=Zero();
-    this->operator ()(Linop,src,psi);
-  }
-  virtual void LogBegin(void)
-  {
-    std::cout << "ConjugageGradientPolynomial::LogBegin() "<<std::endl;
-    ak.resize(0);
-    bk.resize(0);
-    polynomial.resize(0);
-    poly_Ap.resize(0);
-    poly_Ap.resize(0);
-    poly_p.resize(1);
-    poly_r.resize(1);
-    poly_p[0]=1.0;
-    poly_r[0]=1.0;
-  };
-  virtual void LogIteration(int k,RealD a,RealD b)
-  {
-    // With zero guess,
-    // p = r = src
-    //
-    // iterate:
-    //   x =  x + a p
-    //   r =  r - a A p
-    //   p =  r + b p
-    //
-    // [0]
-    // r = x
-    // p = x
-    // Ap=0
-    //
-    // [1]
-    // Ap = A x + 0  ==> shift poly P right by 1 and add 0.
-    // x  = x + a p  ==> add polynomials term by term 
-    // r  = r - a A p  ==> add polynomials term by term
-    // p  = r + b p  ==> add polynomials term by term
-    //
-    std::cout << "ConjugageGradientPolynomial::LogIteration() "<<k<<std::endl;
-    ak.push_back(a);
-    bk.push_back(b);
-    //  Ap= right_shift(p)
-    poly_Ap.resize(k+1);
-    poly_Ap[0]=0.0;
-    for(int i=0;i<k;i++){
-      poly_Ap[i+1]=poly_p[i];
-    }
-
-    //  x = x + a p
-    polynomial.resize(k);
-    polynomial[k-1]=0.0;
-    for(int i=0;i<k;i++){
-      polynomial[i] = polynomial[i] + a * poly_p[i];
-    }
-    
-    //  r = r - a Ap
-    //  p = r + b p
-    poly_r.resize(k+1);
-    poly_p.resize(k+1);
-    poly_r[k] = poly_p[k] = 0.0;
-    for(int i=0;i<k+1;i++){
-      poly_r[i] = poly_r[i] - a * poly_Ap[i];
-      poly_p[i] = poly_r[i] + b * poly_p[i];
-    }
-  }
-};
-
 NAMESPACE_END(Grid);
 #endif

Grid/algorithms/iterative/ConjugateGradientMultiShift.h

@@ -102,11 +102,11 @@ public:
     assert(mass.size()==nshift);
     assert(mresidual.size()==nshift);
   
-    // remove dynamic sized arrays on stack; 2d is a pain with vector
+    // dynamic sized arrays on stack; 2d is a pain with vector
-    std::vector<RealD>  bs(nshift);
+    RealD  bs[nshift];
-    std::vector<RealD>  rsq(nshift);
+    RealD  rsq[nshift];
-    std::vector<std::array<RealD,2> >  z(nshift);
+    RealD  z[nshift][2];
-    std::vector<int>     converged(nshift);
+    int     converged[nshift];
   
     const int       primary =0;
   

Grid/algorithms/iterative/ConjugateGradientMultiShiftCleanup.h

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

Grid/algorithms/iterative/ConjugateGradientMultiShiftMixedPrec.h

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

Grid/algorithms/iterative/ImplicitlyRestartedBlockLanczosCoarse.h

@@ -279,16 +279,16 @@ public:
       Qt = Eigen::MatrixXcd::Identity(Nm,Nm);
       diagonalize(eval2,lmd2,lme2,Nu,Nm,Nm,Qt,grid);
       _sort.push(eval2,Nm);
-      //      Glog << "#Ritz value before shift: "<< std::endl;
+      Glog << "#Ritz value before shift: "<< std::endl;
       for(int i=0; i<Nm; ++i){
-	//	std::cout.precision(13);
+	std::cout.precision(13);
-	//	std::cout << "[" << std::setw(4)<< std::setiosflags(std::ios_base::right) <<i<<"] ";
+	std::cout << "[" << std::setw(4)<< std::setiosflags(std::ios_base::right) <<i<<"] ";
-	//	std::cout << "Rval = "<<std::setw(20)<< std::setiosflags(std::ios_base::left)<< eval2[i] << std::endl;
+	std::cout << "Rval = "<<std::setw(20)<< std::setiosflags(std::ios_base::left)<< eval2[i] << std::endl;
       }
       
       //----------------------------------------------------------------------
       if ( Nm>Nk ) {
-	//        Glog <<" #Apply shifted QR transformations "<<std::endl;
+        Glog <<" #Apply shifted QR transformations "<<std::endl;
         //int k2 = Nk+Nu;
         int k2 = Nk;
       
@@ -326,7 +326,7 @@ public:
         Qt = Eigen::MatrixXcd::Identity(Nm,Nm);
         diagonalize(eval2,lmd2,lme2,Nu,Nk,Nm,Qt,grid);
         _sort.push(eval2,Nk);
-	//	Glog << "#Ritz value after shift: "<< std::endl;
+	Glog << "#Ritz value after shift: "<< std::endl;
         for(int i=0; i<Nk; ++i){
 	  //          std::cout.precision(13);
 	  //          std::cout << "[" << std::setw(4)<< std::setiosflags(std::ios_base::right) <<i<<"] ";
@@ -644,7 +644,7 @@ private:
       //      for (int u=0; u<mrhs; ++u) Glog << " out["<<u<<"] = "<<norm2(out[u])<<std::endl;
       k_start +=mrhs;
     }
-    //    Glog << "LinAlg "<< std::endl;
+    Glog << "LinAlg "<< std::endl;
     
     if (b>0) {
       for (int u=0; u<Nu; ++u) {
@@ -678,7 +678,7 @@ private:
       }
       w_copy[u] = w[u];
     }
-    //    Glog << "LinAlg done"<< std::endl;
+    Glog << "LinAlg done"<< std::endl;
     
     // In block version, the steps 6 and 7 in Lanczos construction is
     // replaced by the QR decomposition of new basis block.
@@ -691,15 +691,15 @@ private:
     }
 
     // re-orthogonalization for numerical stability
-    //    Glog << "Gram Schmidt"<< std::endl;
+    Glog << "Gram Schmidt"<< std::endl;
     orthogonalize(w,Nu,evec,R);
     // QR part
     for (int u=1; u<Nu; ++u) {
       orthogonalize(w[u],w,u);
     }
-    //    Glog << "Gram Schmidt done "<< std::endl;
+    Glog << "Gram Schmidt done "<< std::endl;
     
-    //    Glog << "LinAlg "<< std::endl;
+    Glog << "LinAlg "<< std::endl;
     for (int u=0; u<Nu; ++u) {
       //for (int v=0; v<Nu; ++v) {
       for (int v=u; v<Nu; ++v) {
@@ -716,7 +716,7 @@ private:
 	//        Glog <<" In block "<< b << "," <<" beta[" << u << "," << k-L << "] = " << lme[u][k] << std::endl;
       }
     }
-    //    Glog << "LinAlg done "<< std::endl;
+    Glog << "LinAlg done "<< std::endl;
 
     if (b < Nm/Nu-1) {
       for (int u=0; u<Nu; ++u) {
@@ -935,7 +935,7 @@ if (1){
          int Nu, int Nb, int Nk, int Nm,
          Eigen::MatrixXcd& M)
   {
-    //    Glog << "unpackHermitBlockTriDiagMatToEigen() begin" << '\n'; 
+    Glog << "unpackHermitBlockTriDiagMatToEigen() begin" << '\n'; 
     assert( Nk%Nu == 0 && Nm%Nu == 0 );
     assert( Nk <= Nm );
     M = Eigen::MatrixXcd::Zero(Nk,Nk);
@@ -953,7 +953,7 @@ if (1){
         M(u+(k/Nu)*Nu,k-Nu) = lme[u][k-Nu];
       }
     }
-    //    Glog << "unpackHermitBlockTriDiagMatToEigen() end" << std::endl; 
+    Glog << "unpackHermitBlockTriDiagMatToEigen() end" << std::endl; 
   }
  
 
@@ -963,7 +963,7 @@ if (1){
          int Nu, int Nb, int Nk, int Nm,
          Eigen::MatrixXcd& M)
   {
-    //    Glog << "packHermitBlockTriDiagMatfromEigen() begin" << '\n'; 
+    Glog << "packHermitBlockTriDiagMatfromEigen() begin" << '\n'; 
     assert( Nk%Nu == 0 && Nm%Nu == 0 );
     assert( Nk <= Nm );
     
@@ -979,7 +979,7 @@ if (1){
         lme[u][k-Nu] = M(u+(k/Nu)*Nu,k-Nu);
       }
     }
-    //    Glog << "packHermitBlockTriDiagMatfromEigen() end" <<std::endl; 
+    Glog << "packHermitBlockTriDiagMatfromEigen() end" <<std::endl; 
   }
 
 
@@ -988,7 +988,7 @@ if (1){
 		            RealD Dsh,
 		            Eigen::MatrixXcd& Qprod)
   {
-    //    Glog << "shiftedQRDecompEigen() begin" << '\n'; 
+    Glog << "shiftedQRDecompEigen() begin" << '\n'; 
     Eigen::MatrixXcd Q = Eigen::MatrixXcd::Zero(Nm,Nm);
     Eigen::MatrixXcd R = Eigen::MatrixXcd::Zero(Nm,Nm);
     Eigen::MatrixXcd Mtmp = Eigen::MatrixXcd::Zero(Nm,Nm);
@@ -1004,7 +1004,7 @@ if (1){
                         // lower triangular part used to represent series
                         // of Q sequence.
 
-    //    Glog << "shiftedQRDecompEigen() Housholder & QR" << '\n'; 
+    Glog << "shiftedQRDecompEigen() Housholder & QR" << '\n'; 
     // equivalent operation of Qprod *= Q
     //M = Eigen::MatrixXcd::Zero(Nm,Nm);
     
@@ -1025,7 +1025,7 @@ if (1){
     
     Mtmp = Eigen::MatrixXcd::Zero(Nm,Nm);
 
-    //    Glog << "shiftedQRDecompEigen() Mtmp create" << '\n'; 
+    Glog << "shiftedQRDecompEigen() Mtmp create" << '\n'; 
     for (int i=0; i<Nm; ++i) {
       for (int j=0; j<Nm-(Nu+1); ++j) {
         for (int k=0; k<Nu+1+j; ++k) {
@@ -1033,7 +1033,7 @@ if (1){
         }
       }
     }
-    //    Glog << "shiftedQRDecompEigen() Mtmp loop1" << '\n'; 
+    Glog << "shiftedQRDecompEigen() Mtmp loop1" << '\n'; 
     for (int i=0; i<Nm; ++i) {
       for (int j=Nm-(Nu+1); j<Nm; ++j) {
         for (int k=0; k<Nm; ++k) {
@@ -1041,7 +1041,7 @@ if (1){
         }
       }
     }
-    //    Glog << "shiftedQRDecompEigen() Mtmp loop2" << '\n'; 
+    Glog << "shiftedQRDecompEigen() Mtmp loop2" << '\n'; 
     
     //static int ntimes = 2;
     //for (int j=0; j<Nm-(ntimes*Nu); ++j) {
@@ -1067,13 +1067,13 @@ if (1){
         Mtmp(j,i) = conj(Mtmp(i,j));
       }
     }
-    //    Glog << "shiftedQRDecompEigen() Mtmp loop3" << '\n'; 
+    Glog << "shiftedQRDecompEigen() Mtmp loop3" << '\n'; 
 
     for (int i=0; i<Nm; ++i) {
       Mtmp(i,i) = real(Mtmp(i,i)) + Dsh;
     }
     
-    //    Glog << "shiftedQRDecompEigen() Mtmp loop4" << '\n'; 
+    Glog << "shiftedQRDecompEigen() Mtmp loop4" << '\n'; 
     M = Mtmp;
 
     //M = Q.adjoint()*(M*Q);
@@ -1085,7 +1085,7 @@ if (1){
     //  }
     //}
     
-    //    Glog << "shiftedQRDecompEigen() end" <<std::endl; 
+    Glog << "shiftedQRDecompEigen() end" <<std::endl; 
   }
 
   void exampleQRDecompEigen(void)

Grid/algorithms/iterative/NormalEquations.h

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

Grid/algorithms/iterative/PowerMethod.h

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

Grid/algorithms/iterative/PowerSpectrum.h

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

Grid/algorithms/multigrid/CoarsenedMatrix.h

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

Grid/allocator/AlignedAllocator.h

@@ -174,11 +174,21 @@ template<typename _Tp>  inline bool operator!=(const devAllocator<_Tp>&, const d
 ////////////////////////////////////////////////////////////////////////////////
 // Template typedefs
 ////////////////////////////////////////////////////////////////////////////////
-template<class T> using hostVector          = std::vector<T,alignedAllocator<T> >;           // Needs autoview
+#ifdef ACCELERATOR_CSHIFT
-template<class T> using Vector              = std::vector<T,uvmAllocator<T> >;               // 
+// Cshift on device
-template<class T> using uvmVector           = std::vector<T,uvmAllocator<T> >;               // auto migrating page
+template<class T> using cshiftAllocator = devAllocator<T>;
-template<class T> using deviceVector        = std::vector<T,devAllocator<T> >;               // device vector
+#else
+// Cshift on host
+template<class T> using cshiftAllocator = std::allocator<T>;
+#endif
 
+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> >;
+
+/*
 template<class T> class vecView
 {
  protected:
@@ -187,9 +197,8 @@ template<class T> class vecView
   ViewMode mode;
   void * cpu_ptr;
  public:
-  // Rvalue accessor
   accelerator_inline T & operator[](size_t i) const { return this->data[i]; };
-  vecView(Vector<T> &refer_to_me,ViewMode _mode)
+  vecView(std::vector<T> &refer_to_me,ViewMode _mode)
   {
     cpu_ptr = &refer_to_me[0];
     size = refer_to_me.size();
@@ -205,15 +214,26 @@ template<class T> class vecView
   }
 };
 
-template<class T> vecView<T> VectorView(Vector<T> &vec,ViewMode _mode)
+template<class T> vecView<T> VectorView(std::vector<T> &vec,ViewMode _mode)
 {
   vecView<T> ret(vec,_mode); // does the open
   return ret;                // must be closed
 }
 
+// Little autoscope assister
+template<class View> 
+class VectorViewCloser
+{
+  View v;  // Take a copy of view and call view close when I go out of scope automatically
+ public:
+  VectorViewCloser(View &_v) : v(_v) {};
+  ~VectorViewCloser() { auto ptr = v.cpu_ptr; v.ViewClose();  MemoryManager::NotifyDeletion(ptr);}
+};
+
 #define autoVecView(v_v,v,mode)					\
   auto v_v = VectorView(v,mode);				\
   ViewCloser<decltype(v_v)> _autoView##v_v(v_v);
+*/
 
 NAMESPACE_END(Grid);
 

Grid/allocator/MemoryStats.cc

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

Grid/cartesian/Cartesian_base.h

@@ -82,7 +82,6 @@ public:
   bool _isCheckerBoarded; 
   int        LocallyPeriodic;
   Coordinate _checker_dim_mask;
-  int              _checker_dim;
 
 public:
 
@@ -92,6 +91,7 @@ public:
   ////////////////////////////////////////////////////////////////
   virtual int CheckerBoarded(int dim)=0;
   virtual int CheckerBoard(const Coordinate &site)=0;
+  virtual int CheckerDim(void){ return 0; };
   virtual int CheckerBoardDestination(int source_cb,int shift,int dim)=0;
   virtual int CheckerBoardShift(int source_cb,int dim,int shift,int osite)=0;
   virtual int CheckerBoardShiftForCB(int source_cb,int dim,int shift,int cb)=0;

Grid/cartesian/Cartesian_full.h

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

Grid/cartesian/Cartesian_red_black.h

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

Grid/cshift/Cshift.h

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

Grid/cshift/Cshift_common.h

@@ -30,11 +30,12 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 NAMESPACE_BEGIN(Grid);
 
 extern std::vector<std::pair<int,int> > Cshift_table; 
-extern deviceVector<std::pair<int,int> > Cshift_table_device; 
+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);
@@ -44,13 +45,16 @@ inline std::pair<int,int> *MapCshiftTable(void)
 			  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 
 ///////////////////////////////////////////////////////////////////
 template<class vobj> void 
-Gather_plane_simple (const Lattice<vobj> &rhs,deviceVector<vobj> &buffer,int dimension,int plane,int cbmask, int off=0)
+Gather_plane_simple (const Lattice<vobj> &rhs,cshiftVector<vobj> &buffer,int dimension,int plane,int cbmask, int off=0)
 {
   int rd = rhs.Grid()->_rdimensions[dimension];
 
@@ -90,10 +94,17 @@ Gather_plane_simple (const Lattice<vobj> &rhs,deviceVector<vobj> &buffer,int dim
   {
     auto buffer_p = & buffer[0];
     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]));
     });
+#else
+    autoView(rhs_v , rhs, CpuRead);
+    thread_for(i,ent,{
+      buffer_p[table[i].first]=rhs_v[table[i].second];
+    });
+#endif
   }
 }
 
@@ -118,6 +129,7 @@ Gather_plane_extract(const Lattice<vobj> &rhs,
   int n1=rhs.Grid()->_slice_stride[dimension];
 
   if ( cbmask ==0x3){
+#ifdef ACCELERATOR_CSHIFT
     autoView(rhs_v , rhs, AcceleratorRead);
     accelerator_for(nn,e1*e2,1,{
 	int n = nn%e1;
@@ -128,10 +140,21 @@ Gather_plane_extract(const Lattice<vobj> &rhs,
 	vobj temp =rhs_v[so+o+b];
 	extract<vobj>(temp,pointers,offset);
       });
+#else
+    autoView(rhs_v , rhs, CpuRead);
+    thread_for2d(n,e1,b,e2,{
+	int o      =   n*n1;
+	int offset = b+n*e2;
+	
+	vobj temp =rhs_v[so+o+b];
+	extract<vobj>(temp,pointers,offset);
+      });
+#endif
   } else { 
     Coordinate rdim=rhs.Grid()->_rdimensions;
     Coordinate cdm =rhs.Grid()->_checker_dim_mask;
     std::cout << " Dense packed buffer WARNING " <<std::endl; // Does this get called twice once for each cb?
+#ifdef ACCELERATOR_CSHIFT    
     autoView(rhs_v , rhs, AcceleratorRead);
     accelerator_for(nn,e1*e2,1,{
 	int n = nn%e1;
@@ -152,13 +175,33 @@ Gather_plane_extract(const Lattice<vobj> &rhs,
 	  extract<vobj>(temp,pointers,offset);
 	}
       });
+#else
+    autoView(rhs_v , rhs, CpuRead);
+    thread_for2d(n,e1,b,e2,{
+
+	Coordinate coor;
+
+	int o=n*n1;
+	int oindex = o+b;
+
+       	int cb = RedBlackCheckerBoardFromOindex(oindex, rdim, cdm);
+
+	int ocb=1<<cb;
+	int offset = b+n*e2;
+
+	if ( ocb & cbmask ) {
+	  vobj temp =rhs_v[so+o+b];
+	  extract<vobj>(temp,pointers,offset);
+	}
+      });
+#endif
   }
 }
 
 //////////////////////////////////////////////////////
 // Scatter for when there is no need to SIMD split
 //////////////////////////////////////////////////////
-template<class vobj> void Scatter_plane_simple (Lattice<vobj> &rhs,deviceVector<vobj> &buffer, int dimension,int plane,int cbmask)
+template<class vobj> void Scatter_plane_simple (Lattice<vobj> &rhs,cshiftVector<vobj> &buffer, int dimension,int plane,int cbmask)
 {
   int rd = rhs.Grid()->_rdimensions[dimension];
 
@@ -202,10 +245,17 @@ template<class vobj> void Scatter_plane_simple (Lattice<vobj> &rhs,deviceVector<
   {
     auto buffer_p = & buffer[0];
     auto table = MapCshiftTable();
+#ifdef ACCELERATOR_CSHIFT    
     autoView( rhs_v, rhs, AcceleratorWrite);
     accelerator_for(i,ent,vobj::Nsimd(),{
 	coalescedWrite(rhs_v[table[i].first],coalescedRead(buffer_p[table[i].second]));
     });
+#else
+    autoView( rhs_v, rhs, CpuWrite);
+    thread_for(i,ent,{
+      rhs_v[table[i].first]=buffer_p[table[i].second];
+    });
+#endif
   }
 }
 
@@ -228,6 +278,7 @@ template<class vobj> void Scatter_plane_merge(Lattice<vobj> &rhs,ExtractPointerA
   if(cbmask ==0x3 ) {
     int _slice_stride = rhs.Grid()->_slice_stride[dimension];
     int _slice_block = rhs.Grid()->_slice_block[dimension];
+#ifdef ACCELERATOR_CSHIFT    
     autoView( rhs_v , rhs, AcceleratorWrite);
     accelerator_for(nn,e1*e2,1,{
 	int n = nn%e1;
@@ -236,6 +287,14 @@ template<class vobj> void Scatter_plane_merge(Lattice<vobj> &rhs,ExtractPointerA
 	int offset = b+n*_slice_block;
 	merge(rhs_v[so+o+b],pointers,offset);
       });
+#else
+    autoView( rhs_v , rhs, CpuWrite);
+    thread_for2d(n,e1,b,e2,{
+	int o      = n*_slice_stride;
+	int offset = b+n*_slice_block;
+	merge(rhs_v[so+o+b],pointers,offset);
+    });
+#endif
   } else { 
 
     // Case of SIMD split AND checker dim cannot currently be hit, except in 
@@ -301,11 +360,19 @@ template<class vobj> void Copy_plane(Lattice<vobj>& lhs,const Lattice<vobj> &rhs
 
   {
     auto table = MapCshiftTable();
+#ifdef ACCELERATOR_CSHIFT    
     autoView(rhs_v , rhs, AcceleratorRead);
     autoView(lhs_v , lhs, AcceleratorWrite);
     accelerator_for(i,ent,vobj::Nsimd(),{
       coalescedWrite(lhs_v[table[i].first],coalescedRead(rhs_v[table[i].second]));
     });
+#else
+    autoView(rhs_v , rhs, CpuRead);
+    autoView(lhs_v , lhs, CpuWrite);
+    thread_for(i,ent,{
+      lhs_v[table[i].first]=rhs_v[table[i].second];
+    });
+#endif
   }
 }
 
@@ -345,11 +412,19 @@ template<class vobj> void Copy_plane_permute(Lattice<vobj>& lhs,const Lattice<vo
 
   {
     auto table = MapCshiftTable();
+#ifdef ACCELERATOR_CSHIFT    
     autoView( rhs_v, rhs, AcceleratorRead);
     autoView( lhs_v, lhs, AcceleratorWrite);
     accelerator_for(i,ent,1,{
       permute(lhs_v[table[i].first],rhs_v[table[i].second],permute_type);
     });
+#else
+    autoView( rhs_v, rhs, CpuRead);
+    autoView( lhs_v, lhs, CpuWrite);
+    thread_for(i,ent,{
+      permute(lhs_v[table[i].first],rhs_v[table[i].second],permute_type);
+    });
+#endif
   }
 }
 

Grid/cshift/Cshift_mpi.h

@@ -94,7 +94,7 @@ template<class vobj> void Cshift_comms_simd(Lattice<vobj>& ret,const Lattice<vob
   sshift[0] = rhs.Grid()->CheckerBoardShiftForCB(rhs.Checkerboard(),dimension,shift,Even);
   sshift[1] = rhs.Grid()->CheckerBoardShiftForCB(rhs.Checkerboard(),dimension,shift,Odd);
 
-  //  std::cout << "Cshift_comms_simd dim "<<dimension<<"cb "<<rhs.Checkerboard()<<"shift "<<shift<<" sshift " << sshift[0]<<" "<<sshift[1]<<std::endl;
+  //std::cout << "Cshift_comms_simd dim "<<dimension<<"cb "<<rhs.checkerboard<<"shift "<<shift<<" sshift " << sshift[0]<<" "<<sshift[1]<<std::endl;
   if ( sshift[0] == sshift[1] ) {
     //std::cout << "Single pass Cshift_comms" <<std::endl;
     Cshift_comms_simd(ret,rhs,dimension,shift,0x3);
@@ -104,6 +104,8 @@ template<class vobj> void Cshift_comms_simd(Lattice<vobj>& ret,const Lattice<vob
     Cshift_comms_simd(ret,rhs,dimension,shift,0x2);// both with block stride loop iteration
   }
 }
+#define ACCELERATOR_CSHIFT_NO_COPY
+#ifdef ACCELERATOR_CSHIFT_NO_COPY
 template<class vobj> void Cshift_comms(Lattice<vobj> &ret,const Lattice<vobj> &rhs,int dimension,int shift,int cbmask)
 {
   typedef typename vobj::vector_type vector_type;
@@ -123,8 +125,8 @@ template<class vobj> void Cshift_comms(Lattice<vobj> &ret,const Lattice<vobj> &r
   assert(shift<fd);
   
   int buffer_size = rhs.Grid()->_slice_nblock[dimension]*rhs.Grid()->_slice_block[dimension];
-  static deviceVector<vobj> send_buf; send_buf.resize(buffer_size);
+  static cshiftVector<vobj> send_buf; send_buf.resize(buffer_size);
-  static deviceVector<vobj> recv_buf; recv_buf.resize(buffer_size);
+  static cshiftVector<vobj> recv_buf; recv_buf.resize(buffer_size);
     
   int cb= (cbmask==0x2)? Odd : Even;
   int sshift= rhs.Grid()->CheckerBoardShiftForCB(rhs.Checkerboard(),dimension,shift,cb);
@@ -159,7 +161,7 @@ template<class vobj> void Cshift_comms(Lattice<vobj> &ret,const Lattice<vobj> &r
       grid->ShiftedRanks(dimension,comm_proc,xmit_to_rank,recv_from_rank);
       
       tcomms-=usecond();
-      grid->Barrier();
+      //      grid->Barrier();
 
       grid->SendToRecvFrom((void *)&send_buf[0],
 			   xmit_to_rank,
@@ -167,7 +169,7 @@ template<class vobj> void Cshift_comms(Lattice<vobj> &ret,const Lattice<vobj> &r
 			   recv_from_rank,
 			   bytes);
       xbytes+=bytes;
-      grid->Barrier();
+      //      grid->Barrier();
       tcomms+=usecond();
 
       tscatter-=usecond();
@@ -175,11 +177,13 @@ template<class vobj> void Cshift_comms(Lattice<vobj> &ret,const Lattice<vobj> &r
       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)
@@ -220,8 +224,8 @@ template<class vobj> void  Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vo
   int buffer_size = grid->_slice_nblock[dimension]*grid->_slice_block[dimension];
   //  int words = sizeof(vobj)/sizeof(vector_type);
 
-  static std::vector<deviceVector<scalar_object> >  send_buf_extract; send_buf_extract.resize(Nsimd);
+  static std::vector<cshiftVector<scalar_object> >  send_buf_extract; send_buf_extract.resize(Nsimd);
-  static std::vector<deviceVector<scalar_object> >  recv_buf_extract; recv_buf_extract.resize(Nsimd);
+  static std::vector<cshiftVector<scalar_object> >  recv_buf_extract; recv_buf_extract.resize(Nsimd);
   scalar_object *  recv_buf_extract_mpi;
   scalar_object *  send_buf_extract_mpi;
  
@@ -277,7 +281,7 @@ template<class vobj> void  Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vo
 	grid->ShiftedRanks(dimension,nbr_proc,xmit_to_rank,recv_from_rank); 
 
 	tcomms-=usecond();
-	grid->Barrier();
+	//	grid->Barrier();
 
 	send_buf_extract_mpi = &send_buf_extract[nbr_lane][0];
 	recv_buf_extract_mpi = &recv_buf_extract[i][0];
@@ -288,7 +292,7 @@ template<class vobj> void  Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vo
 			     bytes);
 
 	xbytes+=bytes;
-	grid->Barrier();
+	//	grid->Barrier();
 	tcomms+=usecond();
 
 	rpointers[i] = &recv_buf_extract[i][0];
@@ -301,12 +305,242 @@ template<class vobj> void  Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vo
     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)
+{
+  typedef typename vobj::vector_type vector_type;
+  typedef typename vobj::scalar_type scalar_type;
+
+  GridBase *grid=rhs.Grid();
+  Lattice<vobj> temp(rhs.Grid());
+
+  int fd              = rhs.Grid()->_fdimensions[dimension];
+  int rd              = rhs.Grid()->_rdimensions[dimension];
+  int pd              = rhs.Grid()->_processors[dimension];
+  int simd_layout     = rhs.Grid()->_simd_layout[dimension];
+  int comm_dim        = rhs.Grid()->_processors[dimension] >1 ;
+  assert(simd_layout==1);
+  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);
+  static cshiftVector<vobj> recv_buf_v; recv_buf_v.resize(buffer_size);
+  vobj *send_buf;
+  vobj *recv_buf;
+  {
+    grid->ShmBufferFreeAll();
+    size_t bytes = buffer_size*sizeof(vobj);
+    send_buf=(vobj *)grid->ShmBufferMalloc(bytes);
+    recv_buf=(vobj *)grid->ShmBufferMalloc(bytes);
+  }
+    
+  int cb= (cbmask==0x2)? Odd : Even;
+  int sshift= rhs.Grid()->CheckerBoardShiftForCB(rhs.Checkerboard(),dimension,shift,cb);
+
+  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;
+      if (cbmask != 0x3) words=words>>1;
+
+      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;
+      int xmit_to_rank;
+      grid->ShiftedRanks(dimension,comm_proc,xmit_to_rank,recv_from_rank);
+
+
+      tcomms-=usecond();
+      //      grid->Barrier();
+
+      acceleratorCopyDeviceToDevice((void *)&send_buf_v[0],(void *)&send_buf[0],bytes);
+      grid->SendToRecvFrom((void *)&send_buf[0],
+			   xmit_to_rank,
+			   (void *)&recv_buf[0],
+			   recv_from_rank,
+			   bytes);
+      xbytes+=bytes;
+      acceleratorCopyDeviceToDevice((void *)&recv_buf[0],(void *)&recv_buf_v[0],bytes);
+
+      //      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)
+{
+  GridBase *grid=rhs.Grid();
+  const int Nsimd = grid->Nsimd();
+  typedef typename vobj::vector_type vector_type;
+  typedef typename vobj::scalar_object scalar_object;
+  typedef typename vobj::scalar_type scalar_type;
+   
+  int fd = grid->_fdimensions[dimension];
+  int rd = grid->_rdimensions[dimension];
+  int ld = grid->_ldimensions[dimension];
+  int pd = grid->_processors[dimension];
+  int simd_layout     = grid->_simd_layout[dimension];
+  int comm_dim        = grid->_processors[dimension] >1 ;
+
+  //std::cout << "Cshift_comms_simd dim "<< dimension << " fd "<<fd<<" rd "<<rd
+  //    << " ld "<<ld<<" pd " << pd<<" simd_layout "<<simd_layout 
+  //    << " comm_dim " << comm_dim << " cbmask " << cbmask <<std::endl;
+
+  assert(comm_dim==1);
+  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);
+
+  ///////////////////////////////////////////////
+  // Simd direction uses an extract/merge pair
+  ///////////////////////////////////////////////
+  int buffer_size = grid->_slice_nblock[dimension]*grid->_slice_block[dimension];
+  //  int words = sizeof(vobj)/sizeof(vector_type);
+
+  static std::vector<cshiftVector<scalar_object> >  send_buf_extract; send_buf_extract.resize(Nsimd);
+  static std::vector<cshiftVector<scalar_object> >  recv_buf_extract; recv_buf_extract.resize(Nsimd);
+  scalar_object *  recv_buf_extract_mpi;
+  scalar_object *  send_buf_extract_mpi;
+  {
+    size_t bytes = sizeof(scalar_object)*buffer_size;
+    grid->ShmBufferFreeAll();
+    send_buf_extract_mpi = (scalar_object *)grid->ShmBufferMalloc(bytes);
+    recv_buf_extract_mpi = (scalar_object *)grid->ShmBufferMalloc(bytes);
+  }
+  for(int s=0;s<Nsimd;s++){
+    send_buf_extract[s].resize(buffer_size);
+    recv_buf_extract[s].resize(buffer_size);
+  }
+
+  int bytes = buffer_size*sizeof(scalar_object);
+
+  ExtractPointerArray<scalar_object>  pointers(Nsimd); // 
+  ExtractPointerArray<scalar_object> rpointers(Nsimd); // received pointers
+
+  ///////////////////////////////////////////
+  // Work out what to send where
+  ///////////////////////////////////////////
+  int cb    = (cbmask==0x2)? Odd : Even;
+  int sshift= grid->CheckerBoardShiftForCB(rhs.Checkerboard(),dimension,shift,cb);
+
+  // loop over outer coord planes orthog to dim
+  for(int x=0;x<rd;x++){       
+
+    // FIXME call local permute copy if none are offnode.
+    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++){
+      
+      int inner_bit = (Nsimd>>(permute_type+1));
+      int ic= (i&inner_bit)? 1:0;
+
+      int my_coor          = rd*ic + x;
+      int nbr_coor         = my_coor+sshift;
+      int nbr_proc = ((nbr_coor)/ld) % pd;// relative shift in processors
+
+      int nbr_ic   = (nbr_coor%ld)/rd;    // inner coord of peer
+      int nbr_ox   = (nbr_coor%rd);       // outer coord of peer
+      int nbr_lane = (i&(~inner_bit));
+
+      int recv_from_rank;
+      int xmit_to_rank;
+
+      if (nbr_ic) nbr_lane|=inner_bit;
+
+      assert (sx == nbr_ox);
+
+      if(nbr_proc){
+	grid->ShiftedRanks(dimension,nbr_proc,xmit_to_rank,recv_from_rank); 
+
+	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,
+			     xmit_to_rank,
+			     (void *)recv_buf_extract_mpi,
+			     recv_from_rank,
+			     bytes);
+	acceleratorCopyDeviceToDevice((void *)recv_buf_extract_mpi,(void *)&recv_buf_extract[i][0],bytes);
+	xbytes+=bytes;
+
+	//	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); 
 
 #endif

Grid/cshift/Cshift_table.cc

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

Grid/lattice/Lattice_base.h

@@ -236,20 +236,17 @@ public:
   template<class sobj> inline Lattice<vobj> & operator = (const sobj & r){
     vobj vtmp;
     vtmp = r;
-#if 0
+#if 1
-    deviceVector<vobj> vvtmp(1);
-    acceleratorPut(vvtmp[0],vtmp);
-    vobj *vvtmp_p = & vvtmp[0];
-    auto me  = View(AcceleratorWrite);
-    accelerator_for(ss,me.size(),vobj::Nsimd(),{
-	auto stmp=coalescedRead(*vvtmp_p);
-	coalescedWrite(me[ss],stmp);
-    });
-#else    
     auto me  = View(CpuWrite);
     thread_for(ss,me.size(),{
        me[ss]= r;
       });
+#else    
+    auto me  = View(AcceleratorWrite);
+    accelerator_for(ss,me.size(),vobj::Nsimd(),{
+	auto stmp=coalescedRead(vtmp);
+	coalescedWrite(me[ss],stmp);
+    });
 #endif    
     me.ViewClose();
     return *this;

Grid/lattice/Lattice_basis.h

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

Grid/lattice/Lattice_peekpoke.h

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

Grid/lattice/Lattice_reduction.h

@@ -46,7 +46,7 @@ inline typename vobj::scalar_object sum_cpu(const vobj *arg, Integer osites)
   //  const int Nsimd = vobj::Nsimd();
   const int nthread = GridThread::GetThreads();
 
-  std::vector<sobj> sumarray(nthread);
+  Vector<sobj> sumarray(nthread);
   for(int i=0;i<nthread;i++){
     sumarray[i]=Zero();
   }
@@ -75,7 +75,7 @@ inline typename vobj::scalar_objectD sumD_cpu(const vobj *arg, Integer osites)
 
   const int nthread = GridThread::GetThreads();
 
-  std::vector<sobj> sumarray(nthread);
+  Vector<sobj> sumarray(nthread);
   for(int i=0;i<nthread;i++){
     sumarray[i]=Zero();
   }
@@ -343,6 +343,18 @@ axpby_norm_fast(Lattice<vobj> &z,sobj a,sobj b,const Lattice<vobj> &x,const Latt
   autoView( x_v, x, AcceleratorRead);
   autoView( y_v, y, AcceleratorRead);
   autoView( z_v, z, AcceleratorWrite);
+#if 0
+  typedef decltype(innerProductD(x_v[0],y_v[0])) inner_t;
+  Vector<inner_t> inner_tmp(sites);
+  auto inner_tmp_v = &inner_tmp[0];
+
+  accelerator_for( ss, sites, nsimd,{
+      auto tmp = a*x_v(ss)+b*y_v(ss);
+      coalescedWrite(inner_tmp_v[ss],innerProductD(tmp,tmp));
+      coalescedWrite(z_v[ss],tmp);
+  });
+  nrm = real(TensorRemove(sum(inner_tmp_v,sites)));
+#else
   typedef decltype(innerProduct(x_v[0],y_v[0])) inner_t;
   deviceVector<inner_t> inner_tmp;
   inner_tmp.resize(sites);
@@ -354,6 +366,7 @@ axpby_norm_fast(Lattice<vobj> &z,sobj a,sobj b,const Lattice<vobj> &x,const Latt
       coalescedWrite(z_v[ss],tmp);
   });
   nrm = real(TensorRemove(sumD(inner_tmp_v,sites)));
+#endif
   grid->GlobalSum(nrm);
   return nrm; 
 }
@@ -364,7 +377,7 @@ innerProductNorm(ComplexD& ip, RealD &nrm, const Lattice<vobj> &left,const Latti
   conformable(left,right);
 
   typedef typename vobj::vector_typeD vector_type;
-  std::vector<ComplexD> tmp(2);
+  Vector<ComplexD> tmp(2);
 
   GridBase *grid = left.Grid();
 
@@ -374,8 +387,8 @@ innerProductNorm(ComplexD& ip, RealD &nrm, const Lattice<vobj> &left,const Latti
   // GPU
   typedef decltype(innerProductD(vobj(),vobj())) inner_t;
   typedef decltype(innerProductD(vobj(),vobj())) norm_t;
-  deviceVector<inner_t> inner_tmp(sites);
+  Vector<inner_t> inner_tmp(sites);
-  deviceVector<norm_t>  norm_tmp(sites);
+  Vector<norm_t>  norm_tmp(sites);
   auto inner_tmp_v = &inner_tmp[0];
   auto norm_tmp_v = &norm_tmp[0];
   {
@@ -425,9 +438,7 @@ inline auto sum(const LatticeTrinaryExpression<Op,T1,T2,T3> & expr)
 // sliceSum, sliceInnerProduct, sliceAxpy, sliceNorm etc...
 //////////////////////////////////////////////////////////////////////////////////////////////////////////////
 
-template<class vobj> inline void sliceSum(const Lattice<vobj> &Data,
+template<class vobj> inline void sliceSum(const Lattice<vobj> &Data,std::vector<typename vobj::scalar_object> &result,int orthogdim)
-					  std::vector<typename vobj::scalar_object> &result,
-					  int orthogdim)
 {
   ///////////////////////////////////////////////////////
   // FIXME precision promoted summation
@@ -449,8 +460,8 @@ template<class vobj> inline void sliceSum(const Lattice<vobj> &Data,
   int ld=grid->_ldimensions[orthogdim];
   int rd=grid->_rdimensions[orthogdim];
 
-  std::vector<vobj> lvSum(rd); // will locally sum vectors first
+  Vector<vobj> lvSum(rd); // will locally sum vectors first
-  std::vector<sobj> lsSum(ld,Zero());                    // sum across these down to scalars
+  Vector<sobj> lsSum(ld,Zero());                    // sum across these down to scalars
   ExtractBuffer<sobj> extracted(Nsimd);                  // splitting the SIMD
 
   result.resize(fd); // And then global sum to return the same vector to every node 
@@ -508,20 +519,7 @@ sliceSum(const Lattice<vobj> &Data,int orthogdim)
   return result;
 }
 
-/*
-Reimplement
 
-1)
-template<class vobj>
-static void sliceMaddMatrix (Lattice<vobj> &R,Eigen::MatrixXcd &aa,const Lattice<vobj> &X,const Lattice<vobj> &Y,int Orthog,RealD scale=1.0) 
-
-2)
-template<class vobj>
-static void sliceInnerProductMatrix(  Eigen::MatrixXcd &mat, const Lattice<vobj> &lhs,const Lattice<vobj> &rhs,int Orthog) 
-
-3)
--- Make Slice Mul Matrix call sliceMaddMatrix
- */
 template<class vobj>
 static void sliceInnerProductVector( std::vector<ComplexD> & result, const Lattice<vobj> &lhs,const Lattice<vobj> &rhs,int orthogdim) 
 {
@@ -541,8 +539,8 @@ static void sliceInnerProductVector( std::vector<ComplexD> & result, const Latti
   int ld=grid->_ldimensions[orthogdim];
   int rd=grid->_rdimensions[orthogdim];
 
-  std::vector<vector_type> lvSum(rd); // will locally sum vectors first
+  Vector<vector_type> lvSum(rd); // will locally sum vectors first
-  std::vector<scalar_type > lsSum(ld,scalar_type(0.0));                    // sum across these down to scalars
+  Vector<scalar_type > lsSum(ld,scalar_type(0.0));                    // sum across these down to scalars
   ExtractBuffer<iScalar<scalar_type> > extracted(Nsimd);   // splitting the SIMD  
 
   result.resize(fd); // And then global sum to return the same vector to every node for IO to file
@@ -672,96 +670,203 @@ static void sliceMaddVector(Lattice<vobj> &R,std::vector<RealD> &a,const Lattice
   }
 };
 
+/*
 inline GridBase         *makeSubSliceGrid(const GridBase *BlockSolverGrid,int Orthog)
 {
   int NN    = BlockSolverGrid->_ndimension;
   int nsimd = BlockSolverGrid->Nsimd();
   
-  std::vector<int> latt_phys(NN-1);
+  std::vector<int> latt_phys(0);
-  Coordinate simd_phys;
+  std::vector<int> simd_phys(0);
-  std::vector<int>  mpi_phys(NN-1);
+  std::vector<int>  mpi_phys(0);
-  Coordinate checker_dim_mask(NN-1);
-  int checker_dim=-1;
   
-  int dd;
   for(int d=0;d<NN;d++){
     if( d!=Orthog ) { 
-      latt_phys[dd]=BlockSolverGrid->_fdimensions[d];
+      latt_phys.push_back(BlockSolverGrid->_fdimensions[d]);
-      mpi_phys[dd] =BlockSolverGrid->_processors[d];
+      simd_phys.push_back(BlockSolverGrid->_simd_layout[d]);
-      checker_dim_mask[dd] = BlockSolverGrid->_checker_dim_mask[d];
+      mpi_phys.push_back(BlockSolverGrid->_processors[d]);
-      if ( d == BlockSolverGrid->_checker_dim ) checker_dim = dd;
-      dd++;
     }
   }
-  simd_phys=GridDefaultSimd(latt_phys.size(),nsimd);
+  return (GridBase *)new GridCartesian(latt_phys,simd_phys,mpi_phys); 
-  GridCartesian *tmp         = new GridCartesian(latt_phys,simd_phys,mpi_phys);
-  if(BlockSolverGrid->_isCheckerBoarded) {
-    GridRedBlackCartesian *ret = new GridRedBlackCartesian(tmp,checker_dim_mask,checker_dim);
-    delete tmp;
-    return (GridBase *) ret;
-  } else { 
-    return (GridBase *) tmp;
-  }
 }
+*/
 
 template<class vobj>
 static void sliceMaddMatrix (Lattice<vobj> &R,Eigen::MatrixXcd &aa,const Lattice<vobj> &X,const Lattice<vobj> &Y,int Orthog,RealD scale=1.0) 
 {    
-  GridBase *FullGrid = X.Grid();
-  GridBase *SliceGrid = makeSubSliceGrid(FullGrid,Orthog);
-
-  Lattice<vobj> Ys(SliceGrid);
-  Lattice<vobj> Rs(SliceGrid);
-  Lattice<vobj> Xs(SliceGrid);
-  Lattice<vobj> RR(FullGrid);
-
-  RR = R; // Copies checkerboard for insert
-  
   typedef typename vobj::scalar_object sobj;
   typedef typename vobj::vector_type vector_type;
-  int Nslice = X.Grid()->GlobalDimensions()[Orthog];
+
-  for(int i=0;i<Nslice;i++){
+  int Nblock = X.Grid()->GlobalDimensions()[Orthog];
-    ExtractSlice(Ys,Y,i,Orthog);
+
-    ExtractSlice(Rs,R,i,Orthog);
+  GridBase *FullGrid  = X.Grid();
-    Rs=Ys;
+  //  GridBase *SliceGrid = makeSubSliceGrid(FullGrid,Orthog);
-    for(int j=0;j<Nslice;j++){
+
-      ExtractSlice(Xs,X,j,Orthog);
+  //  Lattice<vobj> Xslice(SliceGrid);
-      Rs = Rs + Xs*(scale*aa(j,i));
+  //  Lattice<vobj> Rslice(SliceGrid);
+
+  assert( FullGrid->_simd_layout[Orthog]==1);
+  //  int nh =  FullGrid->_ndimension;
+  //  int nl = SliceGrid->_ndimension;
+  //  int nl = nh-1;
+
+  //FIXME package in a convenient iterator
+  //Should loop over a plane orthogonal to direction "Orthog"
+  int stride=FullGrid->_slice_stride[Orthog];
+  int block =FullGrid->_slice_block [Orthog];
+  int nblock=FullGrid->_slice_nblock[Orthog];
+  int ostride=FullGrid->_ostride[Orthog];
+
+  autoView( X_v, X, CpuRead);
+  autoView( Y_v, Y, CpuRead);
+  autoView( R_v, R, CpuWrite);
+  thread_region
+  {
+    Vector<vobj> s_x(Nblock);
+
+    thread_for_collapse_in_region(2, n,nblock, {
+     for(int b=0;b<block;b++){
+      int o  = n*stride + b;
+
+      for(int i=0;i<Nblock;i++){
+	s_x[i] = X_v[o+i*ostride];
       }
-    InsertSlice(Rs,RR,i,Orthog);
+
+      vobj dot;
+      for(int i=0;i<Nblock;i++){
+	dot = Y_v[o+i*ostride];
+	for(int j=0;j<Nblock;j++){
+	  dot = dot + s_x[j]*(scale*aa(j,i));
+	}
+	R_v[o+i*ostride]=dot;
+      }
+    }});
   }
-  R=RR; // Copy back handles arguments aliasing case
-  delete SliceGrid;
 };
 
 template<class vobj>
 static void sliceMulMatrix (Lattice<vobj> &R,Eigen::MatrixXcd &aa,const Lattice<vobj> &X,int Orthog,RealD scale=1.0) 
 {    
-  R=Zero();
+  typedef typename vobj::scalar_object sobj;
-  sliceMaddMatrix(R,aa,X,R,Orthog,scale);
+  typedef typename vobj::vector_type vector_type;
+
+  int Nblock = X.Grid()->GlobalDimensions()[Orthog];
+
+  GridBase *FullGrid  = X.Grid();
+  //  GridBase *SliceGrid = makeSubSliceGrid(FullGrid,Orthog);
+  //  Lattice<vobj> Xslice(SliceGrid);
+  //  Lattice<vobj> Rslice(SliceGrid);
+
+  assert( FullGrid->_simd_layout[Orthog]==1);
+  //  int nh =  FullGrid->_ndimension;
+  //  int nl = SliceGrid->_ndimension;
+  //  int nl=1;
+
+  //FIXME package in a convenient iterator
+  // thread_for2d_in_region
+  //Should loop over a plane orthogonal to direction "Orthog"
+  int stride=FullGrid->_slice_stride[Orthog];
+  int block =FullGrid->_slice_block [Orthog];
+  int nblock=FullGrid->_slice_nblock[Orthog];
+  int ostride=FullGrid->_ostride[Orthog];
+  autoView( R_v, R, CpuWrite);
+  autoView( X_v, X, CpuRead);
+  thread_region
+  {
+    std::vector<vobj> s_x(Nblock);
+
+
+    thread_for_collapse_in_region( 2 ,n,nblock,{
+    for(int b=0;b<block;b++){
+      int o  = n*stride + b;
+
+      for(int i=0;i<Nblock;i++){
+	s_x[i] = X_v[o+i*ostride];
+      }
+
+      vobj dot;
+      for(int i=0;i<Nblock;i++){
+	dot = s_x[0]*(scale*aa(0,i));
+	for(int j=1;j<Nblock;j++){
+	  dot = dot + s_x[j]*(scale*aa(j,i));
+	}
+	R_v[o+i*ostride]=dot;
+      }
+    }});
+  }
 };
 
 
 template<class vobj>
 static void sliceInnerProductMatrix(  Eigen::MatrixXcd &mat, const Lattice<vobj> &lhs,const Lattice<vobj> &rhs,int Orthog) 
 {
-  GridBase *SliceGrid = makeSubSliceGrid(lhs.Grid(),Orthog);
-
-  Lattice<vobj> ls(SliceGrid);
-  Lattice<vobj> rs(SliceGrid);
-  
   typedef typename vobj::scalar_object sobj;
   typedef typename vobj::vector_type vector_type;
-  int Nslice = lhs.Grid()->GlobalDimensions()[Orthog];
+  
-  mat = Eigen::MatrixXcd::Zero(Nslice,Nslice);
+  GridBase *FullGrid  = lhs.Grid();
-  for(int s=0;s<Nslice;s++){
+  //  GridBase *SliceGrid = makeSubSliceGrid(FullGrid,Orthog);
-    ExtractSlice(ls,lhs,s,Orthog);
+  
-    for(int ss=0;ss<Nslice;ss++){
+  int Nblock = FullGrid->GlobalDimensions()[Orthog];
-      ExtractSlice(rs,rhs,ss,Orthog);
+  
-      mat(s,ss) = innerProduct(ls,rs);
+  //  Lattice<vobj> Lslice(SliceGrid);
+  //  Lattice<vobj> Rslice(SliceGrid);
+  
+  mat = Eigen::MatrixXcd::Zero(Nblock,Nblock);
+
+  assert( FullGrid->_simd_layout[Orthog]==1);
+  //  int nh =  FullGrid->_ndimension;
+  //  int nl = SliceGrid->_ndimension;
+  //  int nl = nh-1;
+
+  //FIXME package in a convenient iterator
+  //Should loop over a plane orthogonal to direction "Orthog"
+  int stride=FullGrid->_slice_stride[Orthog];
+  int block =FullGrid->_slice_block [Orthog];
+  int nblock=FullGrid->_slice_nblock[Orthog];
+  int ostride=FullGrid->_ostride[Orthog];
+
+  typedef typename vobj::vector_typeD vector_typeD;
+
+  autoView( lhs_v, lhs, CpuRead);
+  autoView( rhs_v, rhs, CpuRead);
+  thread_region
+  {
+    std::vector<vobj> Left(Nblock);
+    std::vector<vobj> Right(Nblock);
+    Eigen::MatrixXcd  mat_thread = Eigen::MatrixXcd::Zero(Nblock,Nblock);
+
+    thread_for_collapse_in_region( 2, n,nblock,{
+    for(int b=0;b<block;b++){
+
+      int o  = n*stride + b;
+
+      for(int i=0;i<Nblock;i++){
+	Left [i] = lhs_v[o+i*ostride];
+	Right[i] = rhs_v[o+i*ostride];
+      }
+
+      for(int i=0;i<Nblock;i++){
+      for(int j=0;j<Nblock;j++){
+	auto tmp = innerProduct(Left[i],Right[j]);
+	auto rtmp = TensorRemove(tmp);
+	auto red  =  Reduce(rtmp);
+	mat_thread(i,j) += std::complex<double>(real(red),imag(red));
+      }}
+    }});
+    thread_critical
+    {
+      mat += mat_thread;
     }  
   }
-  delete SliceGrid;
+
+  for(int i=0;i<Nblock;i++){
+  for(int j=0;j<Nblock;j++){
+    ComplexD sum = mat(i,j);
+    FullGrid->GlobalSum(sum);
+    mat(i,j)=sum;
+  }}
+
+  return;
 }
 
 NAMESPACE_END(Grid);

Grid/lattice/Lattice_reduction_gpu.h

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

Grid/lattice/Lattice_reduction_sycl.h

@@ -4,20 +4,23 @@ NAMESPACE_BEGIN(Grid);
 // Possibly promote to double and sum
 /////////////////////////////////////////////////////////////////////////////////////////////////////////
 
-
 template <class vobj>
 inline typename vobj::scalar_objectD sumD_gpu_tensor(const vobj *lat, Integer osites) 
 {
   typedef typename vobj::scalar_object sobj;
   typedef typename vobj::scalar_objectD sobjD;
-
+  static Vector<sobj> mysum;
+  mysum.resize(1);
+  sobj *mysum_p = & mysum[0];
   sobj identity; zeroit(identity);
-  sobj ret; zeroit(ret);
+  mysum[0] = identity;
+  sobj ret ; 
+
   Integer nsimd= vobj::Nsimd();
-  { 
+
-    sycl::buffer<sobj, 1> abuff(&ret, {1});
+  const cl::sycl::property_list PropList ({ cl::sycl::property::reduction::initialize_to_identity() });
   theGridAccelerator->submit([&](cl::sycl::handler &cgh) {
-      auto Reduction = cl::sycl::reduction(abuff,cgh,identity,std::plus<>());
+    auto Reduction = cl::sycl::reduction(mysum_p,identity,std::plus<>(),PropList);
      cgh.parallel_for(cl::sycl::range<1>{osites},
 		      Reduction,
 		      [=] (cl::sycl::id<1> item, auto &sum) {
@@ -25,7 +28,9 @@ inline typename vobj::scalar_objectD sumD_gpu_tensor(const vobj *lat, Integer os
       sum +=Reduce(lat[osite]);
      });
    });
-  }
+  theGridAccelerator->wait();
+  ret = mysum[0];
+  //  free(mysum,*theGridAccelerator);
   sobjD dret; convertType(dret,ret);
   return dret;
 }
@@ -71,22 +76,59 @@ inline typename vobj::scalar_object sum_gpu_large(const vobj *lat, Integer osite
 
 template<class Word> Word svm_xor(Word *vec,uint64_t L)
 {
+  Word xorResult; xorResult = 0;
+  static Vector<Word> d_sum;
+  d_sum.resize(1);
+  Word *d_sum_p=&d_sum[0];
   Word identity;  identity=0;
-  Word ret = 0;
+  d_sum[0] = identity;
-  { 
+  const cl::sycl::property_list PropList ({ cl::sycl::property::reduction::initialize_to_identity() });
-    sycl::buffer<Word, 1> abuff(&ret, {1});
   theGridAccelerator->submit([&](cl::sycl::handler &cgh) {
-      auto Reduction = cl::sycl::reduction(abuff,cgh,identity,std::bit_xor<>());
+    auto Reduction = cl::sycl::reduction(d_sum_p,identity,std::bit_xor<>(),PropList);
      cgh.parallel_for(cl::sycl::range<1>{L},
 		      Reduction,
 		      [=] (cl::sycl::id<1> index, auto &sum) {
 	 sum^=vec[index];
      });
    });
-  }
   theGridAccelerator->wait();
+  Word ret = d_sum[0];
+  //  free(d_sum,*theGridAccelerator);
   return ret;
 }
 
 NAMESPACE_END(Grid);
 
+/*
+
+template <class vobj>
+inline typename vobj::scalar_objectD sumD_gpu_repack(const vobj *lat, Integer osites)
+{
+  typedef typename vobj::vector_type  vector;
+  typedef typename vobj::scalar_type  scalar;
+
+  typedef typename vobj::scalar_typeD scalarD;
+  typedef typename vobj::scalar_objectD sobjD;
+
+  sobjD ret;
+  scalarD *ret_p = (scalarD *)&ret;
+  
+  const int nsimd = vobj::Nsimd();
+  const int words = sizeof(vobj)/sizeof(vector);
+
+  Vector<scalar> buffer(osites*nsimd);
+  scalar *buf = &buffer[0];
+  vector *dat = (vector *)lat;
+
+  for(int w=0;w<words;w++) {
+
+    accelerator_for(ss,osites,nsimd,{
+	int lane = acceleratorSIMTlane(nsimd);
+	buf[ss*nsimd+lane] = dat[ss*words+w].getlane(lane);
+    });
+    //Precision change at this point is to late to gain precision
+    ret_p[w] = svm_reduce(buf,nsimd*osites);
+  }
+  return ret;
+}
+*/

Grid/lattice/Lattice_slicesum_core.h

@@ -21,18 +21,9 @@ NAMESPACE_BEGIN(Grid);
 
 
 #if defined(GRID_CUDA) || defined(GRID_HIP)
-template<class vobj>
+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) {
-inline void sliceSumReduction_cub_small(const vobj *Data,
-					std::vector<vobj> &lvSum,
-					const int rd,
-					const int e1,
-					const int e2,
-					const int stride,
-					const int ostride,
-					const int Nsimd)
-{
   size_t subvol_size = e1*e2;
-  deviceVector<vobj> reduction_buffer(rd*subvol_size);
+  commVector<vobj> reduction_buffer(rd*subvol_size);
   auto rb_p = &reduction_buffer[0];
   vobj zero_init;
   zeroit(zero_init);
@@ -103,15 +94,7 @@ inline void sliceSumReduction_cub_small(const vobj *Data,
 
 
 #if defined(GRID_SYCL)
-template<class vobj>
+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)
-inline void sliceSumReduction_sycl_small(const vobj *Data,
-					 std::vector <vobj> &lvSum,
-					 const int  &rd,
-					 const int &e1,
-					 const int &e2,
-					 const int &stride,
-					 const int &ostride,
-					 const int &Nsimd)
 {
   size_t subvol_size = e1*e2;
 
@@ -122,7 +105,7 @@ inline void sliceSumReduction_sycl_small(const vobj *Data,
     mysum[r] = vobj_zero; 
   }
 
-  deviceVector<vobj> reduction_buffer(rd*subvol_size);    
+  commVector<vobj> reduction_buffer(rd*subvol_size);    
 
   auto rb_p = &reduction_buffer[0];
 
@@ -161,23 +144,14 @@ inline void sliceSumReduction_sycl_small(const vobj *Data,
 }
 #endif
 
-template<class vobj>
+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) {
-inline void sliceSumReduction_large(const vobj *Data,
-				    std::vector<vobj> &lvSum,
-				    const int rd,
-				    const int e1,
-				    const int e2,
-				    const int stride,
-				    const int ostride,
-				    const int Nsimd)
-{
   typedef typename vobj::vector_type vector;
   const int words = sizeof(vobj)/sizeof(vector);
   const int osites = rd*e1*e2;
-  deviceVector<vector>buffer(osites);
+  commVector<vector>buffer(osites);
   vector *dat = (vector *)Data;
   vector *buf = &buffer[0];
-  std::vector<vector> lvSum_small(rd);
+  Vector<vector> lvSum_small(rd);
   vector *lvSum_ptr = (vector *)&lvSum[0];
 
   for (int w = 0; w < words; w++) {
@@ -194,18 +168,13 @@ inline void sliceSumReduction_large(const vobj *Data,
     for (int r = 0; r < rd; r++) {
       lvSum_ptr[w+words*r]=lvSum_small[r];
     }
-  }
+
   }
 
-template<class vobj>
+  
-inline void sliceSumReduction_gpu(const Lattice<vobj> &Data,
+}
-				  std::vector<vobj> &lvSum,
+
-				  const int rd,
+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)
-				  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) { 
@@ -223,15 +192,7 @@ inline void sliceSumReduction_gpu(const Lattice<vobj> &Data,
 }
 
 
-template<class vobj>
+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)
-inline void sliceSumReduction_cpu(const Lattice<vobj> &Data,
-				  std::vector<vobj> &lvSum,
-				  const int &rd,
-				  const int &e1,
-				  const int &e2,
-				  const int &stride,
-				  const int &ostride,
-				  const int &Nsimd)
 {
   // sum over reduced dimension planes, breaking out orthog dir
   // Parallel over orthog direction
@@ -247,19 +208,15 @@ inline void sliceSumReduction_cpu(const Lattice<vobj> &Data,
   });
 }
 
-template<class vobj> inline void sliceSumReduction(const Lattice<vobj> &Data,
+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) 
-						   std::vector<vobj> &lvSum,
-						   const int &rd,
-						   const int &e1,
-						   const int &e2,
-						   const int &stride,
-						   const int &ostride,
-						   const int &Nsimd) 
 {
   #if defined(GRID_CUDA) || defined(GRID_HIP) || defined(GRID_SYCL)
+  
   sliceSumReduction_gpu(Data, lvSum, rd, e1, e2, stride, ostride, Nsimd);
+  
   #else
   sliceSumReduction_cpu(Data, lvSum, rd, e1, e2, stride, ostride, Nsimd);
+
   #endif
 }
 

Grid/lattice/Lattice_transfer.h

@@ -43,49 +43,20 @@ inline void subdivides(GridBase *coarse,GridBase *fine)
   }
 }
  
- 
 ////////////////////////////////////////////////////////////////////////////////////////////
 // remove and insert a half checkerboard
 ////////////////////////////////////////////////////////////////////////////////////////////
+
 template<class vobj> inline void pickCheckerboard(int cb,Lattice<vobj> &half,const Lattice<vobj> &full)
 {
-  half.Checkerboard() = cb;
+  acceleratorPickCheckerboard(cb,half,full);
-
-  autoView( half_v, half, CpuWrite);
-  autoView( full_v, full, CpuRead);
-  thread_for(ss, full.Grid()->oSites(),{
-    int cbos;
-    Coordinate coor;
-    full.Grid()->oCoorFromOindex(coor,ss);
-    cbos=half.Grid()->CheckerBoard(coor);
-
-    if (cbos==cb) {
-      int ssh=half.Grid()->oIndex(coor);
-      half_v[ssh] = full_v[ss];
-    }
-  });
 }
 template<class vobj> inline void setCheckerboard(Lattice<vobj> &full,const Lattice<vobj> &half)
 {
-  int cb = half.Checkerboard();
+  acceleratorSetCheckerboard(full,half);
-  autoView( half_v , half, CpuRead);
-  autoView( full_v , full, CpuWrite);
-  thread_for(ss,full.Grid()->oSites(),{
-
-    Coordinate coor;
-    int cbos;
-
-    full.Grid()->oCoorFromOindex(coor,ss);
-    cbos=half.Grid()->CheckerBoard(coor);
-      
-    if (cbos==cb) {
-      int ssh=half.Grid()->oIndex(coor);
-      full_v[ss]=half_v[ssh];
-    }
-  });
 }
 
-template<class vobj> inline void acceleratorPickCheckerboard(int cb,Lattice<vobj> &half,const Lattice<vobj> &full, int checker_dim_half=0)
+template<class vobj> inline void acceleratorPickCheckerboard(int cb,Lattice<vobj> &half,const Lattice<vobj> &full, int dummy=0)
 {
   half.Checkerboard() = cb;
   autoView(half_v, half, AcceleratorWrite);
@@ -95,6 +66,7 @@ template<class vobj> inline void acceleratorPickCheckerboard(int cb,Lattice<vobj
   unsigned long ndim_half          = half.Grid()->_ndimension;
   Coordinate checker_dim_mask_half = half.Grid()->_checker_dim_mask;
   Coordinate ostride_half          = half.Grid()->_ostride;
+  int checker_dim_half             = half.Grid()->CheckerDim();
   accelerator_for(ss, full.Grid()->oSites(),full.Grid()->Nsimd(),{
     
     Coordinate coor;
@@ -119,7 +91,7 @@ template<class vobj> inline void acceleratorPickCheckerboard(int cb,Lattice<vobj
     }
   });
 }
-template<class vobj> inline void acceleratorSetCheckerboard(Lattice<vobj> &full,const Lattice<vobj> &half, int checker_dim_half=0)
+template<class vobj> inline void acceleratorSetCheckerboard(Lattice<vobj> &full,const Lattice<vobj> &half, int dummy=0)
 {
   int cb = half.Checkerboard();
   autoView(half_v , half, AcceleratorRead);
@@ -129,6 +101,7 @@ template<class vobj> inline void acceleratorSetCheckerboard(Lattice<vobj> &full,
   unsigned long ndim_half          = half.Grid()->_ndimension;
   Coordinate checker_dim_mask_half = half.Grid()->_checker_dim_mask;
   Coordinate ostride_half          = half.Grid()->_ostride;
+  int checker_dim_half             = half.Grid()->CheckerDim();
   accelerator_for(ss,full.Grid()->oSites(),full.Grid()->Nsimd(),{
 
     Coordinate coor;
@@ -981,14 +954,8 @@ void InsertSlice(const Lattice<vobj> &lowDim,Lattice<vobj> & higherDim,int slice
     hcoor[orthog] = slice;
     for(int d=0;d<nh;d++){
       if ( d!=orthog ) { 
-	hcoor[d]=lcoor[ddl];
+	hcoor[d]=lcoor[ddl++];
-	if ( hg->_checker_dim == d ) {
-	  hcoor[d]=hcoor[d]*2; // factor in the full coor for peekLocalSite
-	  lcoor[ddl]=lcoor[ddl]*2; // factor in the full coor for peekLocalSite
       }
-	ddl++;
-      }
-      
     }
     peekLocalSite(s,lowDimv,lcoor);
     pokeLocalSite(s,higherDimv,hcoor);
@@ -1009,7 +976,6 @@ void ExtractSlice(Lattice<vobj> &lowDim,const Lattice<vobj> & higherDim,int slic
   assert(orthog<nh);
   assert(orthog>=0);
   assert(hg->_processors[orthog]==1);
-  lowDim.Checkerboard() = higherDim.Checkerboard();
 
   int dl; dl = 0;
   for(int d=0;d<nh;d++){
@@ -1027,16 +993,11 @@ void ExtractSlice(Lattice<vobj> &lowDim,const Lattice<vobj> & higherDim,int slic
     Coordinate lcoor(nl);
     Coordinate hcoor(nh);
     lg->LocalIndexToLocalCoor(idx,lcoor);
-    hcoor[orthog] = slice;
     int ddl=0;
+    hcoor[orthog] = slice;
     for(int d=0;d<nh;d++){
       if ( d!=orthog ) { 
-	hcoor[d]=lcoor[ddl];
+	hcoor[d]=lcoor[ddl++];
-	if ( hg->_checker_dim == d ) {
-	  hcoor[d]=hcoor[d]*2;     // factor in the full gridd coor for peekLocalSite
-	  lcoor[ddl]=lcoor[ddl]*2; // factor in the full coor for peekLocalSite
-	}
-	ddl++;
       }
     }
     peekLocalSite(s,higherDimv,hcoor);

Grid/lattice/PaddedCell.h

@@ -54,7 +54,7 @@ struct CshiftImplGauge: public CshiftImplBase<typename Gimpl::GaugeLinkField::ve
  *
  */
 
-template<class vobj> inline void ScatterSlice(const deviceVector<vobj> &buf,
+template<class vobj> inline void ScatterSlice(const cshiftVector<vobj> &buf,
 					      Lattice<vobj> &lat,
 					      int x,
 					      int dim,
@@ -140,7 +140,7 @@ template<class vobj> inline void ScatterSlice(const deviceVector<vobj> &buf,
   });
 }
 
-template<class vobj> inline void GatherSlice(deviceVector<vobj> &buf,
+template<class vobj> inline void GatherSlice(cshiftVector<vobj> &buf,
 					     const Lattice<vobj> &lat,
 					     int x,
 					     int dim,
@@ -462,8 +462,8 @@ public:
     int rNsimd = Nsimd / simd[dimension];
     assert( buffer_size == from.Grid()->_slice_nblock[dimension]*from.Grid()->_slice_block[dimension] / simd[dimension]);
 
-    static deviceVector<vobj> send_buf; 
+    static cshiftVector<vobj> send_buf; 
-    static deviceVector<vobj> recv_buf;
+    static cshiftVector<vobj> recv_buf;
     send_buf.resize(buffer_size*2*depth);    
     recv_buf.resize(buffer_size*2*depth);
 

Grid/qcd/action/fermion/CayleyFermion5D.h

@@ -90,16 +90,16 @@ public:
   void M5D(const FermionField &psi,
 	   const FermionField &phi,
 	   FermionField &chi,
-	   std::vector<Coeff_t> &lower,
+	   Vector<Coeff_t> &lower,
-	   std::vector<Coeff_t> &diag,
+	   Vector<Coeff_t> &diag,
-	   std::vector<Coeff_t> &upper);
+	   Vector<Coeff_t> &upper);
 
   void M5Ddag(const FermionField &psi,
 	      const FermionField &phi,
 	      FermionField &chi,
-	      std::vector<Coeff_t> &lower,
+	      Vector<Coeff_t> &lower,
-	      std::vector<Coeff_t> &diag,
+	      Vector<Coeff_t> &diag,
-	      std::vector<Coeff_t> &upper);
+	      Vector<Coeff_t> &upper);
 
   virtual void   Instantiatable(void)=0;
 
@@ -119,35 +119,35 @@ public:
   RealD mass_plus, mass_minus;
 
   // Save arguments to SetCoefficientsInternal
-  std::vector<Coeff_t> _gamma;
+  Vector<Coeff_t> _gamma;
   RealD                _zolo_hi;
   RealD                _b;
   RealD                _c;
 
   // Cayley form Moebius (tanh and zolotarev)
-  std::vector<Coeff_t> omega;
+  Vector<Coeff_t> omega;
-  std::vector<Coeff_t> bs;    // S dependent coeffs
+  Vector<Coeff_t> bs;    // S dependent coeffs
-  std::vector<Coeff_t> cs;
+  Vector<Coeff_t> cs;
-  std::vector<Coeff_t> as;
+  Vector<Coeff_t> as;
   // For preconditioning Cayley form
-  std::vector<Coeff_t> bee;
+  Vector<Coeff_t> bee;
-  std::vector<Coeff_t> cee;
+  Vector<Coeff_t> cee;
-  std::vector<Coeff_t> aee;
+  Vector<Coeff_t> aee;
-  std::vector<Coeff_t> beo;
+  Vector<Coeff_t> beo;
-  std::vector<Coeff_t> ceo;
+  Vector<Coeff_t> ceo;
-  std::vector<Coeff_t> aeo;
+  Vector<Coeff_t> aeo;
   // LDU factorisation of the eeoo matrix
-  std::vector<Coeff_t> lee;
+  Vector<Coeff_t> lee;
-  std::vector<Coeff_t> leem;
+  Vector<Coeff_t> leem;
-  std::vector<Coeff_t> uee;
+  Vector<Coeff_t> uee;
-  std::vector<Coeff_t> ueem;
+  Vector<Coeff_t> ueem;
-  std::vector<Coeff_t> dee;
+  Vector<Coeff_t> dee;
 
   // Matrices of 5d ee inverse params
-  //  std::vector<iSinglet<Simd> >  MatpInv;
+  Vector<iSinglet<Simd> >  MatpInv;
-  //  std::vector<iSinglet<Simd> >  MatmInv;
+  Vector<iSinglet<Simd> >  MatmInv;
-  //  std::vector<iSinglet<Simd> >  MatpInvDag;
+  Vector<iSinglet<Simd> >  MatpInvDag;
-  //  std::vector<iSinglet<Simd> >  MatmInvDag;
+  Vector<iSinglet<Simd> >  MatmInvDag;
 
   ///////////////////////////////////////////////////////////////
   // Conserved current utilities
@@ -187,7 +187,7 @@ public:
 protected:
   virtual void SetCoefficientsZolotarev(RealD zolohi,Approx::zolotarev_data *zdata,RealD b,RealD c);
   virtual void SetCoefficientsTanh(Approx::zolotarev_data *zdata,RealD b,RealD c);
-  virtual void SetCoefficientsInternal(RealD zolo_hi,std::vector<Coeff_t> & gamma,RealD b,RealD c);
+  virtual void SetCoefficientsInternal(RealD zolo_hi,Vector<Coeff_t> & gamma,RealD b,RealD c);
 };
 
 NAMESPACE_END(Grid);

Grid/qcd/action/fermion/ContinuedFractionFermion5D.h

@@ -90,12 +90,12 @@ protected:
   RealD mass;
   RealD R;
   RealD ZoloHiInv;
-  std::vector<double> Beta;
+  Vector<double> Beta;
-  std::vector<double> cc;;
+  Vector<double> cc;;
-  std::vector<double> cc_d;;
+  Vector<double> cc_d;;
-  std::vector<double> sqrt_cc;
+  Vector<double> sqrt_cc;
-  std::vector<double> See;
+  Vector<double> See;
-  std::vector<double> Aee;
+  Vector<double> Aee;
 
 };
 

Grid/qcd/action/fermion/DomainWallEOFAFermion.h

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

Grid/qcd/action/fermion/ImprovedStaggeredFermion.h

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

Grid/qcd/action/fermion/ImprovedStaggeredFermion5D.h

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

Grid/qcd/action/fermion/MobiusEOFAFermion.h

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

Grid/qcd/action/fermion/NaiveStaggeredFermion.h

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

Grid/qcd/action/fermion/PartialFractionFermion5D.h

@@ -94,8 +94,8 @@ protected:
   RealD R;
   RealD amax;
   RealD scale;
-  std::vector<double> p; 
+  Vector<double> p; 
-  std::vector<double> q;
+  Vector<double> q;
 
 };
 

Grid/qcd/action/fermion/SchurDiagTwoKappa.h

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

Grid/qcd/action/fermion/StaggeredKernels.h

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

Grid/qcd/action/fermion/WilsonCompressor.h

@@ -47,7 +47,7 @@ public:
   static int PartialCompressionFactor(GridBase *grid) { return 1;}
 #endif
   template<class vobj,class cobj,class compressor>
-  static void Gather_plane_simple (deviceVector<std::pair<int,int> >& table,
+  static void Gather_plane_simple (commVector<std::pair<int,int> >& table,
 				   const Lattice<vobj> &rhs,
 				   cobj *buffer,
 				   compressor &compress,
@@ -109,7 +109,7 @@ public:
   // Reorder the fifth dim to be s=Ls-1 , s=0, s=1,...,Ls-2.
   ////////////////////////////////////////////////////////////////////////////////////////////
   template<class vobj,class cobj,class compressor>
-  static void Gather_plane_exchange(deviceVector<std::pair<int,int> >& table,const Lattice<vobj> &rhs,
+  static void Gather_plane_exchange(commVector<std::pair<int,int> >& table,const Lattice<vobj> &rhs,
 				    std::vector<cobj *> pointers,int dimension,int plane,int cbmask,
 				    compressor &compress,int type,int partial)
   {
@@ -197,7 +197,7 @@ public:
 #endif
   
   template<class vobj,class cobj,class compressor>
-  static void Gather_plane_simple (deviceVector<std::pair<int,int> >& table,
+  static void Gather_plane_simple (commVector<std::pair<int,int> >& table,
 					 const Lattice<vobj> &rhs,
 					 cobj *buffer,
 					 compressor &compress,
@@ -208,7 +208,7 @@ public:
     else        FaceGatherSimple::Gather_plane_simple(table,rhs,buffer,compress,off,so,partial);
   }
   template<class vobj,class cobj,class compressor>
-  static void Gather_plane_exchange(deviceVector<std::pair<int,int> >& table,const Lattice<vobj> &rhs,
+  static void Gather_plane_exchange(commVector<std::pair<int,int> >& table,const Lattice<vobj> &rhs,
 				    std::vector<cobj *> pointers,int dimension,int plane,int cbmask,
 				    compressor &compress,int type,int partial)
   {
@@ -402,6 +402,7 @@ public:
 
   typedef CartesianStencil<vobj,cobj,Parameters> Base;
   typedef typename Base::View_type View_type;
+  typedef typename Base::StencilVector StencilVector;
 
   //  Vector<int> surface_list;
   WilsonStencil(GridBase *grid,

Grid/qcd/action/fermion/WilsonFermion.h

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

Grid/qcd/action/fermion/WilsonFermion5D.h

@@ -135,18 +135,21 @@ public:
 		     int dag);
     
   void DhopInternal(StencilImpl & st,
+		    LebesgueOrder &lo,
 		    DoubledGaugeField &U,
 		    const FermionField &in, 
 		    FermionField &out,
 		    int dag);
 
   void DhopInternalOverlappedComms(StencilImpl & st,
+				   LebesgueOrder &lo,
 				   DoubledGaugeField &U,
 				   const FermionField &in, 
 				   FermionField &out,
 				   int dag);
 
   void DhopInternalSerialComms(StencilImpl & st,
+			       LebesgueOrder &lo,
 			       DoubledGaugeField &U,
 			       const FermionField &in, 
 			       FermionField &out,
@@ -200,6 +203,9 @@ public:
   DoubledGaugeField UmuEven;
   DoubledGaugeField UmuOdd;
     
+  LebesgueOrder Lebesgue;
+  LebesgueOrder LebesgueEvenOdd;
+    
   // Comms buffer
   //  std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> >  comm_buf;
 

Grid/qcd/action/fermion/ZMobiusFermion.h

@@ -58,7 +58,7 @@ public:
   {
     //    RealD eps = 1.0;
     std::cout<<GridLogMessage << "ZMobiusFermion (b="<<b<<",c="<<c<<") with Ls= "<<this->Ls<<" gamma passed in"<<std::endl;
-    std::vector<Coeff_t> zgamma(this->Ls);
+    Vector<Coeff_t> zgamma(this->Ls);
     for(int s=0;s<this->Ls;s++){
       zgamma[s] = gamma[s];
     }

Grid/qcd/action/fermion/implementation/CayleyFermion5DImplementation.h

@@ -156,18 +156,18 @@ template<class Impl>
 void CayleyFermion5D<Impl>::M5D   (const FermionField &psi, FermionField &chi)
 {
   int Ls=this->Ls;
-  std::vector<Coeff_t> diag (Ls,1.0);
+  Vector<Coeff_t> diag (Ls,1.0);
-  std::vector<Coeff_t> upper(Ls,-1.0); upper[Ls-1]=mass_minus;
+  Vector<Coeff_t> upper(Ls,-1.0); upper[Ls-1]=mass_minus;
-  std::vector<Coeff_t> lower(Ls,-1.0); lower[0]   =mass_plus;
+  Vector<Coeff_t> lower(Ls,-1.0); lower[0]   =mass_plus;
   M5D(psi,chi,chi,lower,diag,upper);
 }
 template<class Impl>
 void CayleyFermion5D<Impl>::Meooe5D    (const FermionField &psi, FermionField &Din)
 {
   int Ls=this->Ls;
-  std::vector<Coeff_t> diag = bs;
+  Vector<Coeff_t> diag = bs;
-  std::vector<Coeff_t> upper= cs;
+  Vector<Coeff_t> upper= cs;
-  std::vector<Coeff_t> lower= cs; 
+  Vector<Coeff_t> lower= cs; 
   upper[Ls-1]=-mass_minus*upper[Ls-1];
   lower[0]   =-mass_plus*lower[0];
   M5D(psi,psi,Din,lower,diag,upper);
@@ -176,9 +176,9 @@ void CayleyFermion5D<Impl>::Meooe5D    (const FermionField &psi, FermionField &D
 template<class Impl> void CayleyFermion5D<Impl>::Meo5D     (const FermionField &psi, FermionField &chi)
 {
   int Ls=this->Ls;
-  std::vector<Coeff_t> diag = beo;
+  Vector<Coeff_t> diag = beo;
-  std::vector<Coeff_t> upper(Ls);
+  Vector<Coeff_t> upper(Ls);
-  std::vector<Coeff_t> lower(Ls);
+  Vector<Coeff_t> lower(Ls);
   for(int i=0;i<Ls;i++) {
     upper[i]=-ceo[i];
     lower[i]=-ceo[i];
@@ -191,9 +191,9 @@ template<class Impl>
 void CayleyFermion5D<Impl>::Mooee       (const FermionField &psi, FermionField &chi)
 {
   int Ls=this->Ls;
-  std::vector<Coeff_t> diag = bee;
+  Vector<Coeff_t> diag = bee;
-  std::vector<Coeff_t> upper(Ls);
+  Vector<Coeff_t> upper(Ls);
-  std::vector<Coeff_t> lower(Ls);
+  Vector<Coeff_t> lower(Ls);
   for(int i=0;i<Ls;i++) {
     upper[i]=-cee[i];
     lower[i]=-cee[i];
@@ -206,9 +206,9 @@ template<class Impl>
 void CayleyFermion5D<Impl>::MooeeDag    (const FermionField &psi, FermionField &chi)
 {
   int Ls=this->Ls;
-  std::vector<Coeff_t> diag = bee;
+  Vector<Coeff_t> diag = bee;
-  std::vector<Coeff_t> upper(Ls);
+  Vector<Coeff_t> upper(Ls);
-  std::vector<Coeff_t> lower(Ls);
+  Vector<Coeff_t> lower(Ls);
 
   for (int s=0;s<Ls;s++){
     // Assemble the 5d matrix
@@ -236,9 +236,9 @@ template<class Impl>
 void CayleyFermion5D<Impl>::M5Ddag (const FermionField &psi, FermionField &chi)
 {
   int Ls=this->Ls;
-  std::vector<Coeff_t> diag(Ls,1.0);
+  Vector<Coeff_t> diag(Ls,1.0);
-  std::vector<Coeff_t> upper(Ls,-1.0);
+  Vector<Coeff_t> upper(Ls,-1.0);
-  std::vector<Coeff_t> lower(Ls,-1.0);
+  Vector<Coeff_t> lower(Ls,-1.0);
   upper[Ls-1]=-mass_plus*upper[Ls-1];
   lower[0]   =-mass_minus*lower[0];
   M5Ddag(psi,chi,chi,lower,diag,upper);
@@ -248,9 +248,9 @@ template<class Impl>
 void CayleyFermion5D<Impl>::MeooeDag5D    (const FermionField &psi, FermionField &Din)
 {
   int Ls=this->Ls;
-  std::vector<Coeff_t> diag =bs;
+  Vector<Coeff_t> diag =bs;
-  std::vector<Coeff_t> upper=cs;
+  Vector<Coeff_t> upper=cs;
-  std::vector<Coeff_t> lower=cs; 
+  Vector<Coeff_t> lower=cs; 
 
   for (int s=0;s<Ls;s++){
     if ( s== 0 ) {
@@ -394,7 +394,7 @@ void CayleyFermion5D<Impl>::MeoDeriv(GaugeField &mat,const FermionField &U,const
 template<class Impl>
 void CayleyFermion5D<Impl>::SetCoefficientsTanh(Approx::zolotarev_data *zdata,RealD b,RealD c)
 {
-  std::vector<Coeff_t> gamma(this->Ls);
+  Vector<Coeff_t> gamma(this->Ls);
   for(int s=0;s<this->Ls;s++) gamma[s] = zdata->gamma[s];
   SetCoefficientsInternal(1.0,gamma,b,c);
 }
@@ -402,13 +402,13 @@ void CayleyFermion5D<Impl>::SetCoefficientsTanh(Approx::zolotarev_data *zdata,Re
 template<class Impl>
 void CayleyFermion5D<Impl>::SetCoefficientsZolotarev(RealD zolo_hi,Approx::zolotarev_data *zdata,RealD b,RealD c)
 {
-  std::vector<Coeff_t> gamma(this->Ls);
+  Vector<Coeff_t> gamma(this->Ls);
   for(int s=0;s<this->Ls;s++) gamma[s] = zdata->gamma[s];
   SetCoefficientsInternal(zolo_hi,gamma,b,c);
 }
 //Zolo
 template<class Impl>
-void CayleyFermion5D<Impl>::SetCoefficientsInternal(RealD zolo_hi,std::vector<Coeff_t> & gamma,RealD b,RealD c)
+void CayleyFermion5D<Impl>::SetCoefficientsInternal(RealD zolo_hi,Vector<Coeff_t> & gamma,RealD b,RealD c)
 {
   int Ls=this->Ls;
 

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

@@ -43,9 +43,9 @@ void
 CayleyFermion5D<Impl>::M5D(const FermionField &psi_i,
 			       const FermionField &phi_i, 
 			       FermionField &chi_i,
-			       std::vector<Coeff_t> &lower,
+			       Vector<Coeff_t> &lower,
-			       std::vector<Coeff_t> &diag,
+			       Vector<Coeff_t> &diag,
-			       std::vector<Coeff_t> &upper)
+			       Vector<Coeff_t> &upper)
 {
   
   chi_i.Checkerboard()=psi_i.Checkerboard();
@@ -55,16 +55,12 @@ CayleyFermion5D<Impl>::M5D(const FermionField &psi_i,
   autoView(chi , chi_i,AcceleratorWrite);
   assert(phi.Checkerboard() == psi.Checkerboard());
 
+  auto pdiag = &diag[0];
+  auto pupper = &upper[0];
+  auto plower = &lower[0];
+
   int Ls =this->Ls;
 
-  static deviceVector<Coeff_t> d_diag(Ls) ; acceleratorCopyToDevice(&diag[0] ,&d_diag[0],Ls*sizeof(Coeff_t));
-  static deviceVector<Coeff_t> d_upper(Ls); acceleratorCopyToDevice(&upper[0],&d_upper[0],Ls*sizeof(Coeff_t));
-  static deviceVector<Coeff_t> d_lower(Ls); acceleratorCopyToDevice(&lower[0],&d_lower[0],Ls*sizeof(Coeff_t));
-  
-  auto pdiag = &d_diag[0];
-  auto pupper = &d_upper[0];
-  auto plower = &d_lower[0];
-
   // 10 = 3 complex mult + 2 complex add
   // Flops = 10.0*(Nc*Ns) *Ls*vol (/2 for red black counting)
   uint64_t nloop = grid->oSites();
@@ -86,9 +82,9 @@ void
 CayleyFermion5D<Impl>::M5Ddag(const FermionField &psi_i,
 			      const FermionField &phi_i, 
 			      FermionField &chi_i,
-			      std::vector<Coeff_t> &lower,
+			      Vector<Coeff_t> &lower,
-			      std::vector<Coeff_t> &diag,
+			      Vector<Coeff_t> &diag,
-			      std::vector<Coeff_t> &upper)
+			      Vector<Coeff_t> &upper)
 {
   chi_i.Checkerboard()=psi_i.Checkerboard();
   GridBase *grid=psi_i.Grid();
@@ -97,16 +93,12 @@ CayleyFermion5D<Impl>::M5Ddag(const FermionField &psi_i,
   autoView(chi , chi_i,AcceleratorWrite);
   assert(phi.Checkerboard() == psi.Checkerboard());
 
+  auto pdiag = &diag[0];
+  auto pupper = &upper[0];
+  auto plower = &lower[0];
+
   int Ls=this->Ls;
 
-  static deviceVector<Coeff_t> d_diag(Ls) ; acceleratorCopyToDevice(&diag[0] ,&d_diag[0],Ls*sizeof(Coeff_t));
-  static deviceVector<Coeff_t> d_upper(Ls); acceleratorCopyToDevice(&upper[0],&d_upper[0],Ls*sizeof(Coeff_t));
-  static deviceVector<Coeff_t> d_lower(Ls); acceleratorCopyToDevice(&lower[0],&d_lower[0],Ls*sizeof(Coeff_t));
-  
-  auto pdiag = &d_diag[0];
-  auto pupper = &d_upper[0];
-  auto plower = &d_lower[0];
-
   // Flops = 6.0*(Nc*Ns) *Ls*vol
   uint64_t nloop = grid->oSites();
   accelerator_for(sss,nloop,Simd::Nsimd(),{
@@ -134,17 +126,11 @@ CayleyFermion5D<Impl>::MooeeInv    (const FermionField &psi_i, FermionField &chi
 
   int Ls=this->Ls;
 
-  static deviceVector<Coeff_t> d_lee(Ls); acceleratorCopyToDevice(&lee[0],&d_lee[0],Ls*sizeof(Coeff_t));
+  auto plee  = & lee [0];
-  static deviceVector<Coeff_t> d_dee(Ls); acceleratorCopyToDevice(&dee[0],&d_dee[0],Ls*sizeof(Coeff_t));
+  auto pdee  = & dee [0];
-  static deviceVector<Coeff_t> d_uee(Ls); acceleratorCopyToDevice(&uee[0],&d_uee[0],Ls*sizeof(Coeff_t));
+  auto puee  = & uee [0];
-  static deviceVector<Coeff_t> d_leem(Ls); acceleratorCopyToDevice(&leem[0],&d_leem[0],Ls*sizeof(Coeff_t));
+  auto pleem = & leem[0];
-  static deviceVector<Coeff_t> d_ueem(Ls); acceleratorCopyToDevice(&ueem[0],&d_ueem[0],Ls*sizeof(Coeff_t));
+  auto pueem = & ueem[0];
-
-  auto plee  = & d_lee [0];
-  auto pdee  = & d_dee [0];
-  auto puee  = & d_uee [0];
-  auto pleem = & d_leem[0];
-  auto pueem = & d_ueem[0];
 
   uint64_t nloop = grid->oSites()/Ls;
   accelerator_for(sss,nloop,Simd::Nsimd(),{
@@ -196,17 +182,11 @@ CayleyFermion5D<Impl>::MooeeInvDag (const FermionField &psi_i, FermionField &chi
   autoView(psi , psi_i,AcceleratorRead);
   autoView(chi , chi_i,AcceleratorWrite);
 
-  static deviceVector<Coeff_t> d_lee(Ls); acceleratorCopyToDevice(&lee[0],&d_lee[0],Ls*sizeof(Coeff_t));
+  auto plee  = & lee [0];
-  static deviceVector<Coeff_t> d_dee(Ls); acceleratorCopyToDevice(&dee[0],&d_dee[0],Ls*sizeof(Coeff_t));
+  auto pdee  = & dee [0];
-  static deviceVector<Coeff_t> d_uee(Ls); acceleratorCopyToDevice(&uee[0],&d_uee[0],Ls*sizeof(Coeff_t));
+  auto puee  = & uee [0];
-  static deviceVector<Coeff_t> d_leem(Ls); acceleratorCopyToDevice(&leem[0],&d_leem[0],Ls*sizeof(Coeff_t));
+  auto pleem = & leem[0];
-  static deviceVector<Coeff_t> d_ueem(Ls); acceleratorCopyToDevice(&ueem[0],&d_ueem[0],Ls*sizeof(Coeff_t));
+  auto pueem = & ueem[0];
-
-  auto plee  = & d_lee [0];
-  auto pdee  = & d_dee [0];
-  auto puee  = & d_uee [0];
-  auto pleem = & d_leem[0];
-  auto pueem = & d_ueem[0];
 
   assert(psi.Checkerboard() == psi.Checkerboard());
 

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

@@ -1,5 +1,3 @@
-#if 0
-
 /*************************************************************************************
 
     Grid physics library, www.github.com/paboyle/Grid 
@@ -820,5 +818,3 @@ CayleyFermion5D<Impl>::MooeeInternal(const FermionField &psi, FermionField &chi,
 }
 
 NAMESPACE_END(Grid);
-
-#endif

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

@@ -41,7 +41,7 @@ NAMESPACE_BEGIN(Grid);
 // Pplus  backwards..
 template<class Impl>
 void DomainWallEOFAFermion<Impl>::M5D(const FermionField& psi_i, const FermionField& phi_i,FermionField& chi_i, 
-				      std::vector<Coeff_t>& lower, std::vector<Coeff_t>& diag, std::vector<Coeff_t>& upper)
+				      Vector<Coeff_t>& lower, Vector<Coeff_t>& diag, Vector<Coeff_t>& upper)
 {
   chi_i.Checkerboard() = psi_i.Checkerboard();
   int Ls = this->Ls;
@@ -50,15 +50,9 @@ void DomainWallEOFAFermion<Impl>::M5D(const FermionField& psi_i, const FermionFi
   autoView( psi , psi_i, AcceleratorRead);
   autoView( chi , chi_i, AcceleratorWrite);
   assert(phi.Checkerboard() == psi.Checkerboard());
-
+  auto pdiag = &diag[0];
-  static deviceVector<Coeff_t> d_diag(Ls); acceleratorCopyToDevice(&diag[0],&d_diag[0],Ls*sizeof(Coeff_t));
+  auto pupper = &upper[0];
-  static deviceVector<Coeff_t> d_upper(Ls);acceleratorCopyToDevice(&upper[0],&d_upper[0],Ls*sizeof(Coeff_t));
+  auto plower = &lower[0];
-  static deviceVector<Coeff_t> d_lower(Ls);acceleratorCopyToDevice(&lower[0],&d_lower[0],Ls*sizeof(Coeff_t));
-  
-  auto pdiag = &d_diag[0];
-  auto pupper = &d_upper[0];
-  auto plower = &d_lower[0];
-
   // Flops = 6.0*(Nc*Ns) *Ls*vol
   
   auto nloop=grid->oSites()/Ls;
@@ -79,7 +73,7 @@ void DomainWallEOFAFermion<Impl>::M5D(const FermionField& psi_i, const FermionFi
 
 template<class Impl>
 void DomainWallEOFAFermion<Impl>::M5Ddag(const FermionField& psi_i, const FermionField& phi_i, FermionField& chi_i, 
-					 std::vector<Coeff_t>& lower, std::vector<Coeff_t>& diag, std::vector<Coeff_t>& upper)
+					 Vector<Coeff_t>& lower, Vector<Coeff_t>& diag, Vector<Coeff_t>& upper)
 {
   chi_i.Checkerboard() = psi_i.Checkerboard();
   GridBase* grid = psi_i.Grid();
@@ -89,14 +83,9 @@ void DomainWallEOFAFermion<Impl>::M5Ddag(const FermionField& psi_i, const Fermio
   autoView( phi , phi_i, AcceleratorRead);
   autoView( chi , chi_i, AcceleratorWrite);
   assert(phi.Checkerboard() == psi.Checkerboard());
-
+  auto pdiag = &diag[0];
-  static deviceVector<Coeff_t> d_diag(Ls); acceleratorCopyToDevice(&diag[0],&d_diag[0],Ls*sizeof(Coeff_t));
+  auto pupper = &upper[0];
-  static deviceVector<Coeff_t> d_upper(Ls);acceleratorCopyToDevice(&upper[0],&d_upper[0],Ls*sizeof(Coeff_t));
+  auto plower = &lower[0];
-  static deviceVector<Coeff_t> d_lower(Ls);acceleratorCopyToDevice(&lower[0],&d_lower[0],Ls*sizeof(Coeff_t));
-  
-  auto pdiag = &d_diag[0];
-  auto pupper = &d_upper[0];
-  auto plower = &d_lower[0];
 
   // Flops = 6.0*(Nc*Ns) *Ls*vol
 
@@ -125,17 +114,12 @@ void DomainWallEOFAFermion<Impl>::MooeeInv(const FermionField& psi_i, FermionFie
   autoView( chi, chi_i, AcceleratorWrite);
   int Ls = this->Ls;
 
-  static deviceVector<Coeff_t> d_lee(Ls); acceleratorCopyToDevice(&this->lee[0],&d_lee[0],Ls*sizeof(Coeff_t));
+  auto plee  = & this->lee[0];
-  static deviceVector<Coeff_t> d_dee(Ls); acceleratorCopyToDevice(&this->dee[0],&d_dee[0],Ls*sizeof(Coeff_t));
+  auto pdee  = & this->dee[0];
-  static deviceVector<Coeff_t> d_uee(Ls); acceleratorCopyToDevice(&this->uee[0],&d_uee[0],Ls*sizeof(Coeff_t));
+  auto puee  = & this->uee[0];
-  static deviceVector<Coeff_t> d_leem(Ls); acceleratorCopyToDevice(&this->leem[0],&d_leem[0],Ls*sizeof(Coeff_t));
-  static deviceVector<Coeff_t> d_ueem(Ls); acceleratorCopyToDevice(&this->ueem[0],&d_ueem[0],Ls*sizeof(Coeff_t));
 
-  auto plee  = & d_lee [0];
+  auto pleem = & this->leem[0];
-  auto pdee  = & d_dee [0];
+  auto pueem = & this->ueem[0];
-  auto puee  = & d_uee [0];
-  auto pleem = & d_leem[0];
-  auto pueem = & d_ueem[0];
 
   uint64_t nloop=grid->oSites()/Ls;
   accelerator_for(sss,nloop,Simd::Nsimd(),{

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

@@ -131,9 +131,9 @@ void DomainWallEOFAFermion<Impl>::M5D(const FermionField& psi, FermionField& chi
     else{ shiftm = -shift*(mq3-mq2); }
   }
 
-  std::vector<Coeff_t> diag(Ls,1.0);
+  Vector<Coeff_t> diag(Ls,1.0);
-  std::vector<Coeff_t> upper(Ls,-1.0); upper[Ls-1] = mq1 + shiftm;
+  Vector<Coeff_t> upper(Ls,-1.0); upper[Ls-1] = mq1 + shiftm;
-  std::vector<Coeff_t> lower(Ls,-1.0); lower[0]    = mq1 + shiftp;
+  Vector<Coeff_t> lower(Ls,-1.0); lower[0]    = mq1 + shiftp;
 
 #if(0)
   std::cout << GridLogMessage << "DomainWallEOFAFermion::M5D(FF&,FF&):" << std::endl;
@@ -168,9 +168,9 @@ void DomainWallEOFAFermion<Impl>::M5Ddag(const FermionField& psi, FermionField&
     else{ shiftm = -shift*(mq3-mq2); }
   }
 
-  std::vector<Coeff_t> diag(Ls,1.0);
+  Vector<Coeff_t> diag(Ls,1.0);
-  std::vector<Coeff_t> upper(Ls,-1.0); upper[Ls-1] = mq1 + shiftp;
+  Vector<Coeff_t> upper(Ls,-1.0); upper[Ls-1] = mq1 + shiftp;
-  std::vector<Coeff_t> lower(Ls,-1.0); lower[0]    = mq1 + shiftm;
+  Vector<Coeff_t> lower(Ls,-1.0); lower[0]    = mq1 + shiftm;
 
   this->M5Ddag(psi, chi, chi, lower, diag, upper);
 }
@@ -181,9 +181,9 @@ void DomainWallEOFAFermion<Impl>::Mooee(const FermionField& psi, FermionField& c
 {
   int Ls = this->Ls;
 
-  std::vector<Coeff_t> diag = this->bee;
+  Vector<Coeff_t> diag = this->bee;
-  std::vector<Coeff_t> upper(Ls);
+  Vector<Coeff_t> upper(Ls);
-  std::vector<Coeff_t> lower(Ls);
+  Vector<Coeff_t> lower(Ls);
 
   for(int s=0; s<Ls; s++){
     upper[s] = -this->cee[s];
@@ -200,9 +200,9 @@ void DomainWallEOFAFermion<Impl>::MooeeDag(const FermionField& psi, FermionField
 {
   int Ls = this->Ls;
 
-  std::vector<Coeff_t> diag = this->bee;
+  Vector<Coeff_t> diag = this->bee;
-  std::vector<Coeff_t> upper(Ls);
+  Vector<Coeff_t> upper(Ls);
-  std::vector<Coeff_t> lower(Ls);
+  Vector<Coeff_t> lower(Ls);
 
   for(int s=0; s<Ls; s++){
     upper[s] = -this->cee[s];
@@ -218,7 +218,7 @@ void DomainWallEOFAFermion<Impl>::MooeeDag(const FermionField& psi, FermionField
 
 //Zolo
 template<class Impl>
-void DomainWallEOFAFermion<Impl>::SetCoefficientsInternal(RealD zolo_hi, std::vector<Coeff_t>& gamma, RealD b, RealD c)
+void DomainWallEOFAFermion<Impl>::SetCoefficientsInternal(RealD zolo_hi, Vector<Coeff_t>& gamma, RealD b, RealD c)
 {
   int   Ls    = this->Ls;
   int   pm    = this->pm;

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

@@ -61,6 +61,8 @@ ImprovedStaggeredFermion5D<Impl>::ImprovedStaggeredFermion5D(GridCartesian
   UUUmu(&FourDimGrid),
   UUUmuEven(&FourDimRedBlackGrid),
   UUUmuOdd(&FourDimRedBlackGrid),
+  Lebesgue(&FourDimGrid),
+  LebesgueEvenOdd(&FourDimRedBlackGrid),
   _tmp(&FiveDimRedBlackGrid)
 {
 
@@ -275,18 +277,18 @@ void ImprovedStaggeredFermion5D<Impl>::DhopDerivOE(GaugeField &mat,
 
 /*CHANGE */
 template<class Impl>
-void ImprovedStaggeredFermion5D<Impl>::DhopInternal(StencilImpl & st, 
+void ImprovedStaggeredFermion5D<Impl>::DhopInternal(StencilImpl & st, LebesgueOrder &lo,
 						    DoubledGaugeField & U,DoubledGaugeField & UUU,
 						    const FermionField &in, FermionField &out,int dag)
 {
   if ( StaggeredKernelsStatic::Comms == StaggeredKernelsStatic::CommsAndCompute )
-    DhopInternalOverlappedComms(st,U,UUU,in,out,dag);
+    DhopInternalOverlappedComms(st,lo,U,UUU,in,out,dag);
   else
-    DhopInternalSerialComms(st,U,UUU,in,out,dag);
+    DhopInternalSerialComms(st,lo,U,UUU,in,out,dag);
 }
 
 template<class Impl>
-void ImprovedStaggeredFermion5D<Impl>::DhopInternalOverlappedComms(StencilImpl & st, 
+void ImprovedStaggeredFermion5D<Impl>::DhopInternalOverlappedComms(StencilImpl & st, LebesgueOrder &lo,
 								   DoubledGaugeField & U,DoubledGaugeField & UUU,
 								   const FermionField &in, FermionField &out,int dag)
 {
@@ -311,7 +313,7 @@ void ImprovedStaggeredFermion5D<Impl>::DhopInternalOverlappedComms(StencilImpl &
   {
     int interior=1;
     int exterior=0;
-    Kernels::DhopImproved(st,U,UUU,in,out,dag,interior,exterior);
+    Kernels::DhopImproved(st,lo,U,UUU,in,out,dag,interior,exterior);
   }
 
   st.CommsMerge(compressor);
@@ -321,12 +323,12 @@ void ImprovedStaggeredFermion5D<Impl>::DhopInternalOverlappedComms(StencilImpl &
   {
     int interior=0;
     int exterior=1;
-    Kernels::DhopImproved(st,U,UUU,in,out,dag,interior,exterior);
+    Kernels::DhopImproved(st,lo,U,UUU,in,out,dag,interior,exterior);
   }
 }
 
 template<class Impl>
-void ImprovedStaggeredFermion5D<Impl>::DhopInternalSerialComms(StencilImpl & st, 
+void ImprovedStaggeredFermion5D<Impl>::DhopInternalSerialComms(StencilImpl & st, LebesgueOrder &lo,
 						    DoubledGaugeField & U,DoubledGaugeField & UUU,
 						    const FermionField &in, FermionField &out,int dag)
 {
@@ -339,7 +341,7 @@ void ImprovedStaggeredFermion5D<Impl>::DhopInternalSerialComms(StencilImpl & st,
   {
     int interior=1;
     int exterior=1;
-    Kernels::DhopImproved(st,U,UUU,in,out,dag,interior,exterior);
+    Kernels::DhopImproved(st,lo,U,UUU,in,out,dag,interior,exterior);
   }
 }
 /*CHANGE END*/
@@ -355,7 +357,7 @@ void ImprovedStaggeredFermion5D<Impl>::DhopOE(const FermionField &in, FermionFie
   assert(in.Checkerboard()==Even);
   out.Checkerboard() = Odd;
 
-  DhopInternal(StencilEven,UmuOdd,UUUmuOdd,in,out,dag);
+  DhopInternal(StencilEven,LebesgueEvenOdd,UmuOdd,UUUmuOdd,in,out,dag);
 }
 template<class Impl>
 void ImprovedStaggeredFermion5D<Impl>::DhopEO(const FermionField &in, FermionField &out,int dag)
@@ -366,7 +368,7 @@ void ImprovedStaggeredFermion5D<Impl>::DhopEO(const FermionField &in, FermionFie
   assert(in.Checkerboard()==Odd);
   out.Checkerboard() = Even;
 
-  DhopInternal(StencilOdd,UmuEven,UUUmuEven,in,out,dag);
+  DhopInternal(StencilOdd,LebesgueEvenOdd,UmuEven,UUUmuEven,in,out,dag);
 }
 template<class Impl>
 void ImprovedStaggeredFermion5D<Impl>::Dhop(const FermionField &in, FermionField &out,int dag)
@@ -376,7 +378,7 @@ void ImprovedStaggeredFermion5D<Impl>::Dhop(const FermionField &in, FermionField
 
   out.Checkerboard() = in.Checkerboard();
 
-  DhopInternal(Stencil,Umu,UUUmu,in,out,dag);
+  DhopInternal(Stencil,Lebesgue,Umu,UUUmu,in,out,dag);
 }
 
 /////////////////////////////////////////////////////////////////////////

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

@@ -48,6 +48,8 @@ ImprovedStaggeredFermion<Impl>::ImprovedStaggeredFermion(GridCartesian &Fgrid, G
     StencilEven(&Hgrid, npoint, Even, directions, displacements,p),  // source is Even
     StencilOdd(&Hgrid, npoint, Odd, directions, displacements,p),  // source is Odd
     mass(_mass),
+    Lebesgue(_grid),
+    LebesgueEvenOdd(_cbgrid),
     Umu(&Fgrid),
     UmuEven(&Hgrid),
     UmuOdd(&Hgrid),
@@ -337,7 +339,7 @@ void ImprovedStaggeredFermion<Impl>::Dhop(const FermionField &in, FermionField &
 
   out.Checkerboard() = in.Checkerboard();
 
-  DhopInternal(Stencil, Umu, UUUmu, in, out, dag);
+  DhopInternal(Stencil, Lebesgue, Umu, UUUmu, in, out, dag);
 }
 
 template <class Impl>
@@ -349,7 +351,7 @@ void ImprovedStaggeredFermion<Impl>::DhopOE(const FermionField &in, FermionField
   assert(in.Checkerboard() == Even);
   out.Checkerboard() = Odd;
 
-  DhopInternal(StencilEven, UmuOdd, UUUmuOdd, in, out, dag);
+  DhopInternal(StencilEven, LebesgueEvenOdd, UmuOdd, UUUmuOdd, in, out, dag);
 }
 
 template <class Impl>
@@ -361,7 +363,7 @@ void ImprovedStaggeredFermion<Impl>::DhopEO(const FermionField &in, FermionField
   assert(in.Checkerboard() == Odd);
   out.Checkerboard() = Even;
 
-  DhopInternal(StencilOdd, UmuEven, UUUmuEven, in, out, dag);
+  DhopInternal(StencilOdd, LebesgueEvenOdd, UmuEven, UUUmuEven, in, out, dag);
 }
 
 template <class Impl>
@@ -392,19 +394,19 @@ void ImprovedStaggeredFermion<Impl>::DhopDir(const FermionField &in, FermionFiel
 
 
 template <class Impl>
-void ImprovedStaggeredFermion<Impl>::DhopInternal(StencilImpl &st, 
+void ImprovedStaggeredFermion<Impl>::DhopInternal(StencilImpl &st, LebesgueOrder &lo,
 						  DoubledGaugeField &U,
 						  DoubledGaugeField &UUU,
 						  const FermionField &in,
 						  FermionField &out, int dag) 
 {
   if ( StaggeredKernelsStatic::Comms == StaggeredKernelsStatic::CommsAndCompute )
-    DhopInternalOverlappedComms(st,U,UUU,in,out,dag);
+    DhopInternalOverlappedComms(st,lo,U,UUU,in,out,dag);
   else
-    DhopInternalSerialComms(st,U,UUU,in,out,dag);
+    DhopInternalSerialComms(st,lo,U,UUU,in,out,dag);
 }
 template <class Impl>
-void ImprovedStaggeredFermion<Impl>::DhopInternalOverlappedComms(StencilImpl &st, 
+void ImprovedStaggeredFermion<Impl>::DhopInternalOverlappedComms(StencilImpl &st, LebesgueOrder &lo,
 								 DoubledGaugeField &U,
 								 DoubledGaugeField &UUU,
 								 const FermionField &in,
@@ -427,7 +429,7 @@ void ImprovedStaggeredFermion<Impl>::DhopInternalOverlappedComms(StencilImpl &st
   {
     int interior=1;
     int exterior=0;
-    Kernels::DhopImproved(st,U,UUU,in,out,dag,interior,exterior);
+    Kernels::DhopImproved(st,lo,U,UUU,in,out,dag,interior,exterior);
   }
 
   st.CommunicateComplete(requests);
@@ -438,13 +440,13 @@ void ImprovedStaggeredFermion<Impl>::DhopInternalOverlappedComms(StencilImpl &st
   {
     int interior=0;
     int exterior=1;
-    Kernels::DhopImproved(st,U,UUU,in,out,dag,interior,exterior);
+    Kernels::DhopImproved(st,lo,U,UUU,in,out,dag,interior,exterior);
   }
 }
 
 
 template <class Impl>
-void ImprovedStaggeredFermion<Impl>::DhopInternalSerialComms(StencilImpl &st, 
+void ImprovedStaggeredFermion<Impl>::DhopInternalSerialComms(StencilImpl &st, LebesgueOrder &lo,
 							     DoubledGaugeField &U,
 							     DoubledGaugeField &UUU,
 							     const FermionField &in,
@@ -458,7 +460,7 @@ void ImprovedStaggeredFermion<Impl>::DhopInternalSerialComms(StencilImpl &st,
   {
     int interior=1;
     int exterior=1;
-    Kernels::DhopImproved(st,U,UUU,in,out,dag,interior,exterior);
+    Kernels::DhopImproved(st,lo,U,UUU,in,out,dag,interior,exterior);
   }
 };
 

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

@@ -39,7 +39,7 @@ NAMESPACE_BEGIN(Grid);
  
 template<class Impl>
 void MobiusEOFAFermion<Impl>::M5D(const FermionField &psi_i, const FermionField &phi_i, FermionField &chi_i,
-				  std::vector<Coeff_t> &lower, std::vector<Coeff_t> &diag, std::vector<Coeff_t> &upper)
+				  Vector<Coeff_t> &lower, Vector<Coeff_t> &diag, Vector<Coeff_t> &upper)
 {
   chi_i.Checkerboard() = psi_i.Checkerboard();
   GridBase *grid = psi_i.Grid();
@@ -50,13 +50,9 @@ void MobiusEOFAFermion<Impl>::M5D(const FermionField &psi_i, const FermionField
 
   assert(phi.Checkerboard() == psi.Checkerboard());
 
-  static deviceVector<Coeff_t> d_diag(Ls); acceleratorCopyToDevice(&diag[0],&d_diag[0],Ls*sizeof(Coeff_t));
+  auto pdiag = &diag[0];
-  static deviceVector<Coeff_t> d_upper(Ls);acceleratorCopyToDevice(&upper[0],&d_upper[0],Ls*sizeof(Coeff_t));
+  auto pupper = &upper[0];
-  static deviceVector<Coeff_t> d_lower(Ls);acceleratorCopyToDevice(&lower[0],&d_lower[0],Ls*sizeof(Coeff_t));
+  auto plower = &lower[0];
-  
-  auto pdiag = &d_diag[0];
-  auto pupper = &d_upper[0];
-  auto plower = &d_lower[0];
 
   // Flops = 6.0*(Nc*Ns) *Ls*vol
   int nloop = grid->oSites()/Ls;
@@ -78,8 +74,8 @@ void MobiusEOFAFermion<Impl>::M5D(const FermionField &psi_i, const FermionField
 
 template<class Impl>
 void MobiusEOFAFermion<Impl>::M5D_shift(const FermionField &psi_i, const FermionField &phi_i, FermionField &chi_i,
-					std::vector<Coeff_t> &lower, std::vector<Coeff_t> &diag, std::vector<Coeff_t> &upper,
+					Vector<Coeff_t> &lower, Vector<Coeff_t> &diag, Vector<Coeff_t> &upper,
-					std::vector<Coeff_t> &shift_coeffs)
+					Vector<Coeff_t> &shift_coeffs)
 {
   chi_i.Checkerboard() = psi_i.Checkerboard();
   GridBase *grid = psi_i.Grid();
@@ -93,15 +89,10 @@ void MobiusEOFAFermion<Impl>::M5D_shift(const FermionField &psi_i, const Fermion
 
   assert(phi.Checkerboard() == psi.Checkerboard());
 
-  static deviceVector<Coeff_t> d_diag(Ls); acceleratorCopyToDevice(&diag[0],&d_diag[0],Ls*sizeof(Coeff_t));
+  auto pdiag = &diag[0];
-  static deviceVector<Coeff_t> d_upper(Ls);acceleratorCopyToDevice(&upper[0],&d_upper[0],Ls*sizeof(Coeff_t));
+  auto pupper = &upper[0];
-  static deviceVector<Coeff_t> d_lower(Ls);acceleratorCopyToDevice(&lower[0],&d_lower[0],Ls*sizeof(Coeff_t));
+  auto plower = &lower[0];
-  static deviceVector<Coeff_t> d_shift_coeffs(Ls);acceleratorCopyToDevice(&shift_coeffs[0],&d_shift_coeffs[0],Ls*sizeof(Coeff_t));
+  auto pshift_coeffs = &shift_coeffs[0];
-  
-  auto pdiag = &d_diag[0];
-  auto pupper = &d_upper[0];
-  auto plower = &d_lower[0];
-  auto pshift_coeffs = &d_shift_coeffs[0];
 
   // Flops = 6.0*(Nc*Ns) *Ls*vol
   int nloop = grid->oSites()/Ls;
@@ -128,7 +119,7 @@ void MobiusEOFAFermion<Impl>::M5D_shift(const FermionField &psi_i, const Fermion
 
 template<class Impl>
 void MobiusEOFAFermion<Impl>::M5Ddag(const FermionField &psi_i, const FermionField &phi_i, FermionField &chi_i,
-				     std::vector<Coeff_t> &lower, std::vector<Coeff_t> &diag, std::vector<Coeff_t> &upper)
+				     Vector<Coeff_t> &lower, Vector<Coeff_t> &diag, Vector<Coeff_t> &upper)
 {
   chi_i.Checkerboard() = psi_i.Checkerboard();
   GridBase *grid = psi_i.Grid();
@@ -139,13 +130,9 @@ void MobiusEOFAFermion<Impl>::M5Ddag(const FermionField &psi_i, const FermionFie
 
   assert(phi.Checkerboard() == psi.Checkerboard());
 
-  static deviceVector<Coeff_t> d_diag(Ls); acceleratorCopyToDevice(&diag[0],&d_diag[0],Ls*sizeof(Coeff_t));
+  auto pdiag = &diag[0];
-  static deviceVector<Coeff_t> d_upper(Ls);acceleratorCopyToDevice(&upper[0],&d_upper[0],Ls*sizeof(Coeff_t));
+  auto pupper = &upper[0];
-  static deviceVector<Coeff_t> d_lower(Ls);acceleratorCopyToDevice(&lower[0],&d_lower[0],Ls*sizeof(Coeff_t));
+  auto plower = &lower[0];
-  
-  auto pdiag = &d_diag[0];
-  auto pupper = &d_upper[0];
-  auto plower = &d_lower[0];
 
   // Flops = 6.0*(Nc*Ns) *Ls*vol
   int nloop = grid->oSites()/Ls;
@@ -167,8 +154,8 @@ void MobiusEOFAFermion<Impl>::M5Ddag(const FermionField &psi_i, const FermionFie
 
 template<class Impl>
 void MobiusEOFAFermion<Impl>::M5Ddag_shift(const FermionField &psi_i, const FermionField &phi_i, FermionField &chi_i,
-					   std::vector<Coeff_t> &lower, std::vector<Coeff_t> &diag, std::vector<Coeff_t> &upper,
+					   Vector<Coeff_t> &lower, Vector<Coeff_t> &diag, Vector<Coeff_t> &upper,
-					   std::vector<Coeff_t> &shift_coeffs)
+					   Vector<Coeff_t> &shift_coeffs)
 {
   chi_i.Checkerboard() = psi_i.Checkerboard();
   GridBase *grid = psi_i.Grid();
@@ -180,15 +167,10 @@ void MobiusEOFAFermion<Impl>::M5Ddag_shift(const FermionField &psi_i, const Ferm
 
   assert(phi.Checkerboard() == psi.Checkerboard());
 
-  static deviceVector<Coeff_t> d_diag(Ls); acceleratorCopyToDevice(&diag[0],&d_diag[0],Ls*sizeof(Coeff_t));
+  auto pdiag = &diag[0];
-  static deviceVector<Coeff_t> d_upper(Ls);acceleratorCopyToDevice(&upper[0],&d_upper[0],Ls*sizeof(Coeff_t));
+  auto pupper = &upper[0];
-  static deviceVector<Coeff_t> d_lower(Ls);acceleratorCopyToDevice(&lower[0],&d_lower[0],Ls*sizeof(Coeff_t));
+  auto plower = &lower[0];
-  static deviceVector<Coeff_t> d_shift_coeffs(Ls);acceleratorCopyToDevice(&shift_coeffs[0],&d_shift_coeffs[0],Ls*sizeof(Coeff_t));
+  auto pshift_coeffs = &shift_coeffs[0];
-  
-  auto pdiag = &d_diag[0];
-  auto pupper = &d_upper[0];
-  auto plower = &d_lower[0];
-  auto pshift_coeffs = &d_shift_coeffs[0];
 
   // Flops = 6.0*(Nc*Ns) *Ls*vol
   auto pm = this->pm;
@@ -230,17 +212,11 @@ void MobiusEOFAFermion<Impl>::MooeeInv(const FermionField &psi_i, FermionField &
   autoView(psi , psi_i, AcceleratorRead);
   autoView(chi , chi_i, AcceleratorWrite);
 
-  static deviceVector<Coeff_t> d_lee(Ls); acceleratorCopyToDevice(&this->lee[0],&d_lee[0],Ls*sizeof(Coeff_t));
+  auto plee = & this->lee [0];
-  static deviceVector<Coeff_t> d_dee(Ls); acceleratorCopyToDevice(&this->dee[0],&d_dee[0],Ls*sizeof(Coeff_t));
+  auto pdee = & this->dee [0];
-  static deviceVector<Coeff_t> d_uee(Ls); acceleratorCopyToDevice(&this->uee[0],&d_uee[0],Ls*sizeof(Coeff_t));
+  auto puee = & this->uee [0];
-  static deviceVector<Coeff_t> d_leem(Ls); acceleratorCopyToDevice(&this->leem[0],&d_leem[0],Ls*sizeof(Coeff_t));
+  auto pleem= & this->leem[0];
-  static deviceVector<Coeff_t> d_ueem(Ls); acceleratorCopyToDevice(&this->ueem[0],&d_ueem[0],Ls*sizeof(Coeff_t));
+  auto pueem= & this->ueem[0];
-
-  auto plee  = & d_lee [0];
-  auto pdee  = & d_dee [0];
-  auto puee  = & d_uee [0];
-  auto pleem = & d_leem[0];
-  auto pueem = & d_ueem[0];
 
   if(this->shift != 0.0){ MooeeInv_shift(psi_i,chi_i); return; }
 
@@ -292,24 +268,14 @@ void MobiusEOFAFermion<Impl>::MooeeInv_shift(const FermionField &psi_i, FermionF
   autoView(psi , psi_i, AcceleratorRead);
   autoView(chi , chi_i, AcceleratorWrite);
 
-  // Move into object and constructor
-  static deviceVector<Coeff_t> d_lee(Ls); acceleratorCopyToDevice(&this->lee[0],&d_lee[0],Ls*sizeof(Coeff_t));
-  static deviceVector<Coeff_t> d_dee(Ls); acceleratorCopyToDevice(&this->dee[0],&d_dee[0],Ls*sizeof(Coeff_t));
-  static deviceVector<Coeff_t> d_uee(Ls); acceleratorCopyToDevice(&this->uee[0],&d_uee[0],Ls*sizeof(Coeff_t));
-  static deviceVector<Coeff_t> d_leem(Ls); acceleratorCopyToDevice(&this->leem[0],&d_leem[0],Ls*sizeof(Coeff_t));
-  static deviceVector<Coeff_t> d_ueem(Ls); acceleratorCopyToDevice(&this->ueem[0],&d_ueem[0],Ls*sizeof(Coeff_t));
-
   auto pm = this->pm;
-  auto plee  = & d_lee [0];
+  auto plee = & this->lee [0];
-  auto pdee  = & d_dee [0];
+  auto pdee = & this->dee [0];
-  auto puee  = & d_uee [0];
+  auto puee = & this->uee [0];
-  auto pleem = & d_leem[0];
+  auto pleem= & this->leem[0];
-  auto pueem = & d_ueem[0];
+  auto pueem= & this->ueem[0];
-
+  auto pMooeeInv_shift_lc   = &MooeeInv_shift_lc[0];
-  static deviceVector<Coeff_t> d_MooeeInv_shift_lc(Ls); acceleratorCopyToDevice(&MooeeInv_shift_lc[0],&d_MooeeInv_shift_lc[0],Ls*sizeof(Coeff_t));
+  auto pMooeeInv_shift_norm = &MooeeInv_shift_norm[0];
-  static deviceVector<Coeff_t> d_MooeeInv_shift_norm(Ls); acceleratorCopyToDevice(&MooeeInv_shift_norm[0],&d_MooeeInv_shift_norm[0],Ls*sizeof(Coeff_t));
-  auto pMooeeInv_shift_lc   = &d_MooeeInv_shift_lc[0];
-  auto pMooeeInv_shift_norm = &d_MooeeInv_shift_norm[0];
 
   int nloop = grid->oSites()/Ls;
   accelerator_for(sss,nloop,Simd::Nsimd(),{
@@ -367,17 +333,11 @@ void MobiusEOFAFermion<Impl>::MooeeInvDag(const FermionField &psi_i, FermionFiel
   autoView(psi , psi_i, AcceleratorRead);
   autoView(chi , chi_i, AcceleratorWrite);
 
-  static deviceVector<Coeff_t> d_lee(Ls); acceleratorCopyToDevice(&this->lee[0],&d_lee[0],Ls*sizeof(Coeff_t));
+  auto plee = & this->lee [0];
-  static deviceVector<Coeff_t> d_dee(Ls); acceleratorCopyToDevice(&this->dee[0],&d_dee[0],Ls*sizeof(Coeff_t));
+  auto pdee = & this->dee [0];
-  static deviceVector<Coeff_t> d_uee(Ls); acceleratorCopyToDevice(&this->uee[0],&d_uee[0],Ls*sizeof(Coeff_t));
+  auto puee = & this->uee [0];
-  static deviceVector<Coeff_t> d_leem(Ls); acceleratorCopyToDevice(&this->leem[0],&d_leem[0],Ls*sizeof(Coeff_t));
+  auto pleem= & this->leem[0];
-  static deviceVector<Coeff_t> d_ueem(Ls); acceleratorCopyToDevice(&this->ueem[0],&d_ueem[0],Ls*sizeof(Coeff_t));
+  auto pueem= & this->ueem[0];
-
-  auto plee  = & d_lee [0];
-  auto pdee  = & d_dee [0];
-  auto puee  = & d_uee [0];
-  auto pleem = & d_leem[0];
-  auto pueem = & d_ueem[0];
 
   int nloop = grid->oSites()/Ls;
   accelerator_for(sss,nloop,Simd::Nsimd(),{
@@ -426,28 +386,14 @@ void MobiusEOFAFermion<Impl>::MooeeInvDag_shift(const FermionField &psi_i, Fermi
   autoView(chi , chi_i, AcceleratorWrite);
   int Ls = this->Ls;
 
-  static deviceVector<Coeff_t> d_lee(Ls); acceleratorCopyToDevice(&this->lee[0],&d_lee[0],Ls*sizeof(Coeff_t));
-  static deviceVector<Coeff_t> d_dee(Ls); acceleratorCopyToDevice(&this->dee[0],&d_dee[0],Ls*sizeof(Coeff_t));
-  static deviceVector<Coeff_t> d_uee(Ls); acceleratorCopyToDevice(&this->uee[0],&d_uee[0],Ls*sizeof(Coeff_t));
-  static deviceVector<Coeff_t> d_leem(Ls); acceleratorCopyToDevice(&this->leem[0],&d_leem[0],Ls*sizeof(Coeff_t));
-  static deviceVector<Coeff_t> d_ueem(Ls); acceleratorCopyToDevice(&this->ueem[0],&d_ueem[0],Ls*sizeof(Coeff_t));
-
   auto pm = this->pm;
-  auto plee  = & d_lee [0];
+  auto plee = & this->lee [0];
-  auto pdee  = & d_dee [0];
+  auto pdee = & this->dee [0];
-  auto puee  = & d_uee [0];
+  auto puee = & this->uee [0];
-  auto pleem = & d_leem[0];
+  auto pleem= & this->leem[0];
-  auto pueem = & d_ueem[0];
+  auto pueem= & this->ueem[0];
-
+  auto pMooeeInvDag_shift_lc   = &MooeeInvDag_shift_lc[0];
-  static deviceVector<Coeff_t> d_MooeeInvDag_shift_lc(Ls);
+  auto pMooeeInvDag_shift_norm = &MooeeInvDag_shift_norm[0];
-  static deviceVector<Coeff_t> d_MooeeInvDag_shift_norm(Ls);
-  acceleratorCopyToDevice(&MooeeInvDag_shift_lc[0],&d_MooeeInvDag_shift_lc[0],Ls*sizeof(Coeff_t));
-  acceleratorCopyToDevice(&MooeeInvDag_shift_norm[0],&d_MooeeInvDag_shift_norm[0],Ls*sizeof(Coeff_t));
-  auto pMooeeInvDag_shift_lc   = &d_MooeeInvDag_shift_lc[0];
-  auto pMooeeInvDag_shift_norm = &d_MooeeInvDag_shift_norm[0];
-
-  //  auto pMooeeInvDag_shift_lc   = &MooeeInvDag_shift_lc[0];
-  //  auto pMooeeInvDag_shift_norm = &MooeeInvDag_shift_norm[0];
 
   int nloop = grid->oSites()/Ls;
   accelerator_for(sss,nloop,Simd::Nsimd(),{

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

@@ -196,9 +196,9 @@ void MobiusEOFAFermion<Impl>::M5D(const FermionField& psi, FermionField& chi)
 {
   int Ls = this->Ls;
 
-  std::vector<Coeff_t> diag(Ls,1.0);
+  Vector<Coeff_t> diag(Ls,1.0);
-  std::vector<Coeff_t> upper(Ls,-1.0);  upper[Ls-1] = this->mq1;
+  Vector<Coeff_t> upper(Ls,-1.0);  upper[Ls-1] = this->mq1;
-  std::vector<Coeff_t> lower(Ls,-1.0);  lower[0]    = this->mq1;
+  Vector<Coeff_t> lower(Ls,-1.0);  lower[0]    = this->mq1;
 
   // no shift term
   if(this->shift == 0.0){ this->M5D(psi, chi, chi, lower, diag, upper); }
@@ -212,9 +212,9 @@ void MobiusEOFAFermion<Impl>::M5Ddag(const FermionField& psi, FermionField& chi)
 {
   int Ls = this->Ls;
 
-  std::vector<Coeff_t> diag(Ls,1.0);
+  Vector<Coeff_t> diag(Ls,1.0);
-  std::vector<Coeff_t> upper(Ls,-1.0);  upper[Ls-1] = this->mq1;
+  Vector<Coeff_t> upper(Ls,-1.0);  upper[Ls-1] = this->mq1;
-  std::vector<Coeff_t> lower(Ls,-1.0);  lower[0]    = this->mq1;
+  Vector<Coeff_t> lower(Ls,-1.0);  lower[0]    = this->mq1;
 
   // no shift term
   if(this->shift == 0.0){ this->M5Ddag(psi, chi, chi, lower, diag, upper); }
@@ -230,9 +230,9 @@ void MobiusEOFAFermion<Impl>::Mooee(const FermionField& psi, FermionField& chi)
   int Ls = this->Ls;
 
   // coefficients of Mooee
-  std::vector<Coeff_t> diag = this->bee;
+  Vector<Coeff_t> diag = this->bee;
-  std::vector<Coeff_t> upper(Ls);
+  Vector<Coeff_t> upper(Ls);
-  std::vector<Coeff_t> lower(Ls);
+  Vector<Coeff_t> lower(Ls);
   for(int s=0; s<Ls; s++){
     upper[s] = -this->cee[s];
     lower[s] = -this->cee[s];
@@ -253,9 +253,9 @@ void MobiusEOFAFermion<Impl>::MooeeDag(const FermionField& psi, FermionField& ch
   int Ls = this->Ls;
 
   // coefficients of MooeeDag
-  std::vector<Coeff_t> diag = this->bee;
+  Vector<Coeff_t> diag = this->bee;
-  std::vector<Coeff_t> upper(Ls);
+  Vector<Coeff_t> upper(Ls);
-  std::vector<Coeff_t> lower(Ls);
+  Vector<Coeff_t> lower(Ls);
   for(int s=0; s<Ls; s++){
     if(s==0) {
       upper[s] = -this->cee[s+1];
@@ -314,10 +314,10 @@ void MobiusEOFAFermion<Impl>::SetCoefficientsPrecondShiftOps()
   // Tridiagonal solve for MooeeInvDag_shift_lc
   {
     Coeff_t m(0.0);
-    std::vector<Coeff_t> d = Mooee_shift;
+    Vector<Coeff_t> d = Mooee_shift;
-    std::vector<Coeff_t> u(Ls,0.0);
+    Vector<Coeff_t> u(Ls,0.0);
-    std::vector<Coeff_t> y(Ls,0.0);
+    Vector<Coeff_t> y(Ls,0.0);
-    std::vector<Coeff_t> q(Ls,0.0);
+    Vector<Coeff_t> q(Ls,0.0);
     if(pm == 1){ u[0] = 1.0; }
     else{ u[Ls-1] = 1.0; }
 

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

@@ -48,6 +48,8 @@ NaiveStaggeredFermion<Impl>::NaiveStaggeredFermion(GridCartesian &Fgrid, GridRed
     StencilEven(&Hgrid, npoint, Even, directions, displacements,p),  // source is Even
     StencilOdd(&Hgrid, npoint, Odd, directions, displacements,p),  // source is Odd
     mass(_mass),
+    Lebesgue(_grid),
+    LebesgueEvenOdd(_cbgrid),
     Umu(&Fgrid),
     UmuEven(&Hgrid),
     UmuOdd(&Hgrid),
@@ -266,7 +268,7 @@ void NaiveStaggeredFermion<Impl>::Dhop(const FermionField &in, FermionField &out
 
   out.Checkerboard() = in.Checkerboard();
 
-  DhopInternal(Stencil, Umu, in, out, dag);
+  DhopInternal(Stencil, Lebesgue, Umu, in, out, dag);
 }
 
 template <class Impl>
@@ -278,7 +280,7 @@ void NaiveStaggeredFermion<Impl>::DhopOE(const FermionField &in, FermionField &o
   assert(in.Checkerboard() == Even);
   out.Checkerboard() = Odd;
 
-  DhopInternal(StencilEven, UmuOdd, in, out, dag);
+  DhopInternal(StencilEven, LebesgueEvenOdd, UmuOdd, in, out, dag);
 }
 
 template <class Impl>
@@ -290,7 +292,7 @@ void NaiveStaggeredFermion<Impl>::DhopEO(const FermionField &in, FermionField &o
   assert(in.Checkerboard() == Odd);
   out.Checkerboard() = Even;
 
-  DhopInternal(StencilOdd, UmuEven, in, out, dag);
+  DhopInternal(StencilOdd, LebesgueEvenOdd, UmuEven, in, out, dag);
 }
 
 template <class Impl>
@@ -321,18 +323,18 @@ void NaiveStaggeredFermion<Impl>::DhopDir(const FermionField &in, FermionField &
 
 
 template <class Impl>
-void NaiveStaggeredFermion<Impl>::DhopInternal(StencilImpl &st,
+void NaiveStaggeredFermion<Impl>::DhopInternal(StencilImpl &st, LebesgueOrder &lo,
 					       DoubledGaugeField &U,
 					       const FermionField &in,
 					       FermionField &out, int dag) 
 {
   if ( StaggeredKernelsStatic::Comms == StaggeredKernelsStatic::CommsAndCompute )
-    DhopInternalOverlappedComms(st,U,in,out,dag);
+    DhopInternalOverlappedComms(st,lo,U,in,out,dag);
   else
-    DhopInternalSerialComms(st,U,in,out,dag);
+    DhopInternalSerialComms(st,lo,U,in,out,dag);
 }
 template <class Impl>
-void NaiveStaggeredFermion<Impl>::DhopInternalOverlappedComms(StencilImpl &st,
+void NaiveStaggeredFermion<Impl>::DhopInternalOverlappedComms(StencilImpl &st, LebesgueOrder &lo,
 							      DoubledGaugeField &U,
 							      const FermionField &in,
 							      FermionField &out, int dag) 
@@ -354,7 +356,7 @@ void NaiveStaggeredFermion<Impl>::DhopInternalOverlappedComms(StencilImpl &st,
   {
     int interior=1;
     int exterior=0;
-    Kernels::DhopNaive(st,U,in,out,dag,interior,exterior);
+    Kernels::DhopNaive(st,lo,U,in,out,dag,interior,exterior);
   }
 
   st.CommunicateComplete(requests);
@@ -365,12 +367,12 @@ void NaiveStaggeredFermion<Impl>::DhopInternalOverlappedComms(StencilImpl &st,
   {
     int interior=0;
     int exterior=1;
-    Kernels::DhopNaive(st,U,in,out,dag,interior,exterior);
+    Kernels::DhopNaive(st,lo,U,in,out,dag,interior,exterior);
   }
 }
 
 template <class Impl>
-void NaiveStaggeredFermion<Impl>::DhopInternalSerialComms(StencilImpl &st,
+void NaiveStaggeredFermion<Impl>::DhopInternalSerialComms(StencilImpl &st, LebesgueOrder &lo,
 							  DoubledGaugeField &U,
 							  const FermionField &in,
 							  FermionField &out, int dag) 
@@ -383,7 +385,7 @@ void NaiveStaggeredFermion<Impl>::DhopInternalSerialComms(StencilImpl &st,
   {
     int interior=1;
     int exterior=1;
-    Kernels::DhopNaive(st,U,in,out,dag,interior,exterior);
+    Kernels::DhopNaive(st,lo,U,in,out,dag,interior,exterior);
   }
 };
 

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

@@ -375,6 +375,23 @@ void StaggeredKernels<Impl>::DhopSiteHandExt(StencilView &st,
   }
 }
 
+/*
+#define DHOP_SITE_HAND_INSTANTIATE(IMPL)				\
+  template void StaggeredKernels<IMPL>::DhopSiteHand(StencilImpl &st, LebesgueOrder &lo, \
+						     DoubledGaugeFieldView &U,DoubledGaugeFieldView &UUU, \
+						     SiteSpinor *buf, int LLs, int sU, \
+						     const FermionFieldView &in, FermionFieldView &out, int dag); \
+									\
+  template void StaggeredKernels<IMPL>::DhopSiteHandInt(StencilImpl &st, LebesgueOrder &lo, \
+						     DoubledGaugeFieldView &U,DoubledGaugeFieldView &UUU, \
+						     SiteSpinor *buf, int LLs, int sU, \
+						     const FermionFieldView &in, FermionFieldView &out, int dag); \
+									\
+  template void StaggeredKernels<IMPL>::DhopSiteHandExt(StencilImpl &st, LebesgueOrder &lo, \
+						     DoubledGaugeFieldView &U,DoubledGaugeFieldView &UUU, \
+						     SiteSpinor *buf, int LLs, int sU, \
+						     const FermionFieldView &in, FermionFieldView &out, int dag); \
+*/
 #undef LOAD_CHI
 #undef HAND_DECLARATIONS
 

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

@@ -256,7 +256,7 @@ void StaggeredKernels<Impl>::DhopDirKernel(StencilImpl &st, DoubledGaugeFieldVie
   });
 
 template <class Impl> 
-void StaggeredKernels<Impl>::DhopImproved(StencilImpl &st, 
+void StaggeredKernels<Impl>::DhopImproved(StencilImpl &st, LebesgueOrder &lo, 
 					  DoubledGaugeField &U, DoubledGaugeField &UUU, 
 					  const FermionField &in, FermionField &out, int dag, int interior,int exterior)
 {
@@ -294,7 +294,7 @@ void StaggeredKernels<Impl>::DhopImproved(StencilImpl &st,
   assert(0 && " Kernel optimisation case not covered ");
 }
 template <class Impl> 
-void StaggeredKernels<Impl>::DhopNaive(StencilImpl &st, 
+void StaggeredKernels<Impl>::DhopNaive(StencilImpl &st, LebesgueOrder &lo, 
 				       DoubledGaugeField &U,
 				       const FermionField &in, FermionField &out, int dag, int interior,int exterior)
 {

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

@@ -58,9 +58,15 @@ WilsonFermion5D<Impl>::WilsonFermion5D(GaugeField &_Umu,
   Umu(_FourDimGrid),
   UmuEven(_FourDimRedBlackGrid),
   UmuOdd (_FourDimRedBlackGrid),
+  Lebesgue(_FourDimGrid),
+  LebesgueEvenOdd(_FourDimRedBlackGrid),
   _tmp(&FiveDimRedBlackGrid),
   Dirichlet(0)
 {
+  Stencil.lo     = &Lebesgue;
+  StencilEven.lo = &LebesgueEvenOdd;
+  StencilOdd.lo  = &LebesgueEvenOdd;
+  
   // some assertions
   assert(FiveDimGrid._ndimension==5);
   assert(FourDimGrid._ndimension==4);
@@ -299,19 +305,19 @@ void WilsonFermion5D<Impl>::DhopDerivOE(GaugeField &mat,
 }
 
 template<class Impl>
-void WilsonFermion5D<Impl>::DhopInternal(StencilImpl & st,
+void WilsonFermion5D<Impl>::DhopInternal(StencilImpl & st, LebesgueOrder &lo,
                                          DoubledGaugeField & U,
                                          const FermionField &in, FermionField &out,int dag)
 {
   if ( WilsonKernelsStatic::Comms == WilsonKernelsStatic::CommsAndCompute )
-    DhopInternalOverlappedComms(st,U,in,out,dag);
+    DhopInternalOverlappedComms(st,lo,U,in,out,dag);
   else 
-    DhopInternalSerialComms(st,U,in,out,dag);
+    DhopInternalSerialComms(st,lo,U,in,out,dag);
 }
 
 
 template<class Impl>
-void WilsonFermion5D<Impl>::DhopInternalOverlappedComms(StencilImpl & st,
+void WilsonFermion5D<Impl>::DhopInternalOverlappedComms(StencilImpl & st, LebesgueOrder &lo,
 							DoubledGaugeField & U,
 							const FermionField &in, FermionField &out,int dag)
 {
@@ -325,12 +331,10 @@ void WilsonFermion5D<Impl>::DhopInternalOverlappedComms(StencilImpl & st,
   // Start comms  // Gather intranode and extra node differentiated??
   /////////////////////////////
   {
-    //    std::cout << " WilsonFermion5D gather " <<std::endl;
     GRID_TRACE("Gather");
     st.HaloExchangeOptGather(in,compressor); // Put the barrier in the routine
   }
   
-  //  std::cout << " WilsonFermion5D Communicate Begin " <<std::endl;
   std::vector<std::vector<CommsRequest_t> > requests;
   auto id=traceStart("Communicate overlapped");
   st.CommunicateBegin(requests);
@@ -339,7 +343,6 @@ void WilsonFermion5D<Impl>::DhopInternalOverlappedComms(StencilImpl & st,
   // Overlap with comms
   /////////////////////////////
   {
-    //  std::cout << " WilsonFermion5D Comms merge " <<std::endl;
     GRID_TRACE("MergeSHM");
     st.CommsMergeSHM(compressor);// Could do this inside parallel region overlapped with comms
   }
@@ -347,7 +350,6 @@ void WilsonFermion5D<Impl>::DhopInternalOverlappedComms(StencilImpl & st,
   /////////////////////////////
   // do the compute interior
   /////////////////////////////
-  //  std::cout << " WilsonFermion5D Interior " <<std::endl;
   int Opt = WilsonKernelsStatic::Opt; // Why pass this. Kernels should know
   if (dag == DaggerYes) {
     GRID_TRACE("DhopDagInterior");
@@ -360,7 +362,6 @@ void WilsonFermion5D<Impl>::DhopInternalOverlappedComms(StencilImpl & st,
   /////////////////////////////
   // Complete comms
   /////////////////////////////
-  //  std::cout << " WilsonFermion5D Comms Complete " <<std::endl;
   st.CommunicateComplete(requests);
   traceStop(id);
 
@@ -368,13 +369,11 @@ void WilsonFermion5D<Impl>::DhopInternalOverlappedComms(StencilImpl & st,
   // do the compute exterior
   /////////////////////////////
   {
-    //    std::cout << " WilsonFermion5D Comms Merge " <<std::endl;
     GRID_TRACE("Merge");
     st.CommsMerge(compressor);
   }
   
 
-  //  std::cout << " WilsonFermion5D Exterior " <<std::endl;
   if (dag == DaggerYes) {
     GRID_TRACE("DhopDagExterior");
     Kernels::DhopDagKernel(Opt,st,U,st.CommBuf(),LLs,U.oSites(),in,out,0,1);
@@ -382,12 +381,11 @@ void WilsonFermion5D<Impl>::DhopInternalOverlappedComms(StencilImpl & st,
     GRID_TRACE("DhopExterior");
     Kernels::DhopKernel   (Opt,st,U,st.CommBuf(),LLs,U.oSites(),in,out,0,1);
   }
-  //  std::cout << " WilsonFermion5D Done " <<std::endl;
 }
 
 
 template<class Impl>
-void WilsonFermion5D<Impl>::DhopInternalSerialComms(StencilImpl & st, 
+void WilsonFermion5D<Impl>::DhopInternalSerialComms(StencilImpl & st, LebesgueOrder &lo,
 						    DoubledGaugeField & U,
 						    const FermionField &in, 
 						    FermionField &out,int dag)
@@ -397,13 +395,11 @@ void WilsonFermion5D<Impl>::DhopInternalSerialComms(StencilImpl & st,
 
   int LLs = in.Grid()->_rdimensions[0];
 
-  //  std::cout << " WilsonFermion5D Halo exch " <<std::endl;
   {
     GRID_TRACE("HaloExchange");
     st.HaloExchangeOpt(in,compressor);
   }
   
-  //  std::cout << " WilsonFermion5D Dhop " <<std::endl;
   int Opt = WilsonKernelsStatic::Opt;
   if (dag == DaggerYes) {
     GRID_TRACE("DhopDag");
@@ -412,7 +408,6 @@ void WilsonFermion5D<Impl>::DhopInternalSerialComms(StencilImpl & st,
     GRID_TRACE("Dhop");
     Kernels::DhopKernel(Opt,st,U,st.CommBuf(),LLs,U.oSites(),in,out);
   }
-  //  std::cout << " WilsonFermion5D Done " <<std::endl;
 }
 
 
@@ -425,7 +420,7 @@ void WilsonFermion5D<Impl>::DhopOE(const FermionField &in, FermionField &out,int
   assert(in.Checkerboard()==Even);
   out.Checkerboard() = Odd;
 
-  DhopInternal(StencilEven,UmuOdd,in,out,dag);
+  DhopInternal(StencilEven,LebesgueEvenOdd,UmuOdd,in,out,dag);
 }
 template<class Impl>
 void WilsonFermion5D<Impl>::DhopEO(const FermionField &in, FermionField &out,int dag)
@@ -436,7 +431,7 @@ void WilsonFermion5D<Impl>::DhopEO(const FermionField &in, FermionField &out,int
   assert(in.Checkerboard()==Odd);
   out.Checkerboard() = Even;
 
-  DhopInternal(StencilOdd,UmuEven,in,out,dag);
+  DhopInternal(StencilOdd,LebesgueEvenOdd,UmuEven,in,out,dag);
 }
 template<class Impl>
 void WilsonFermion5D<Impl>::Dhop(const FermionField &in, FermionField &out,int dag)
@@ -446,7 +441,7 @@ void WilsonFermion5D<Impl>::Dhop(const FermionField &in, FermionField &out,int d
 
   out.Checkerboard() = in.Checkerboard();
 
-  DhopInternal(Stencil,Umu,in,out,dag);
+  DhopInternal(Stencil,Lebesgue,Umu,in,out,dag);
 }
 template<class Impl>
 void WilsonFermion5D<Impl>::DW(const FermionField &in, FermionField &out,int dag)

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

@@ -52,12 +52,17 @@ WilsonFermion<Impl>::WilsonFermion(GaugeField &_Umu, GridCartesian &Fgrid,
     StencilEven(&Hgrid, npoint, Even, directions,displacements,p),  // source is Even
     StencilOdd(&Hgrid, npoint, Odd, directions,displacements,p),  // source is Odd
     mass(_mass),
+    Lebesgue(_grid),
+    LebesgueEvenOdd(_cbgrid),
     Umu(&Fgrid),
     UmuEven(&Hgrid),
     UmuOdd(&Hgrid),
       _tmp(&Hgrid),
       anisotropyCoeff(anis)
 {
+  Stencil.lo     = &Lebesgue;
+  StencilEven.lo = &LebesgueEvenOdd;
+  StencilOdd.lo  = &LebesgueEvenOdd;
   // Allocate the required comms buffer
   ImportGauge(_Umu);
   if  (anisotropyCoeff.isAnisotropic){
@@ -309,7 +314,7 @@ void WilsonFermion<Impl>::Dhop(const FermionField &in, FermionField &out, int da
 
   out.Checkerboard() = in.Checkerboard();
 
-  DhopInternal(Stencil, Umu, in, out, dag);
+  DhopInternal(Stencil, Lebesgue, Umu, in, out, dag);
 }
 
 template <class Impl>
@@ -321,7 +326,7 @@ void WilsonFermion<Impl>::DhopOE(const FermionField &in, FermionField &out, int
   assert(in.Checkerboard() == Even);
   out.Checkerboard() = Odd;
 
-  DhopInternal(StencilEven, UmuOdd, in, out, dag);
+  DhopInternal(StencilEven, LebesgueEvenOdd, UmuOdd, in, out, dag);
 }
 
 template <class Impl>
@@ -333,7 +338,7 @@ void WilsonFermion<Impl>::DhopEO(const FermionField &in, FermionField &out,int d
   assert(in.Checkerboard() == Odd);
   out.Checkerboard() = Even;
 
-  DhopInternal(StencilOdd, UmuEven, in, out, dag);
+  DhopInternal(StencilOdd, LebesgueEvenOdd, UmuEven, in, out, dag);
 }
 
 template <class Impl>
@@ -386,21 +391,21 @@ void WilsonFermion<Impl>::DhopDirCalc(const FermionField &in, FermionField &out,
 };
 
 template <class Impl>
-void WilsonFermion<Impl>::DhopInternal(StencilImpl &st, 
+void WilsonFermion<Impl>::DhopInternal(StencilImpl &st, LebesgueOrder &lo,
                                        DoubledGaugeField &U,
                                        const FermionField &in,
                                        FermionField &out, int dag)
 {
 #ifdef GRID_OMP
   if ( WilsonKernelsStatic::Comms == WilsonKernelsStatic::CommsAndCompute )
-    DhopInternalOverlappedComms(st,U,in,out,dag);
+    DhopInternalOverlappedComms(st,lo,U,in,out,dag);
   else
 #endif
-    DhopInternalSerial(st,U,in,out,dag);
+    DhopInternalSerial(st,lo,U,in,out,dag);
 }
 
 template <class Impl>
-void WilsonFermion<Impl>::DhopInternalOverlappedComms(StencilImpl &st, 
+void WilsonFermion<Impl>::DhopInternalOverlappedComms(StencilImpl &st, LebesgueOrder &lo,
 						      DoubledGaugeField &U,
 						      const FermionField &in,
 						      FermionField &out, int dag)
@@ -469,7 +474,7 @@ void WilsonFermion<Impl>::DhopInternalOverlappedComms(StencilImpl &st,
 
 
 template <class Impl>
-void WilsonFermion<Impl>::DhopInternalSerial(StencilImpl &st, 
+void WilsonFermion<Impl>::DhopInternalSerial(StencilImpl &st, LebesgueOrder &lo,
                                        DoubledGaugeField &U,
                                        const FermionField &in,
                                        FermionField &out, int dag)

Grid/qcd/action/fermion/implementation/WilsonKernelsAsmAvx512.h

@@ -40,11 +40,11 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 /// Switch off the 5d vectorised code optimisations
 #undef DWFVEC5D
 
-static std::vector<vComplexF> signsF;
+static Vector<vComplexF> signsF;
 
   template<typename vtype>    
-  int setupSigns(std::vector<vtype>& signs ){
+  int setupSigns(Vector<vtype>& signs ){
-    std::vector<vtype> bother(2);
+    Vector<vtype> bother(2);
     signs = bother;
     vrsign(signs[0]);
     visign(signs[1]);
@@ -364,7 +364,7 @@ WilsonKernels<ZDomainWallVec5dImplF>::AsmDhopSiteDagExt(StencilView &st, Doubled
 
 #include <simd/Intel512double.h>
     
-static std::vector<vComplexD> signsD;
+static Vector<vComplexD> signsD;
 static int signInitD = setupSigns(signsD);
     
 #define MAYBEPERM(A,perm) if (perm) { A ; }

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

@@ -434,7 +434,7 @@ void WilsonKernels<Impl>::DhopDirKernel( StencilImpl &st, DoubledGaugeField &U,S
 
 #define ASM_CALL(A)							\
   thread_for( sss, Nsite, {						\
-    int ss = sss; /*st.lo->Reorder(sss);*/			\
+    int ss = st.lo->Reorder(sss);					\
     int sU = ss;							\
     int sF = ss*Ls;							\
     WilsonKernels<Impl>::A(st_v,U_v,buf,sF,sU,Ls,1,in_v,out_v);		\

Grid/qcd/representations/adjoint.h

@@ -40,7 +40,7 @@ public:
     U = Zero();
     LatticeColourMatrix tmp(Uin.Grid());
 
-    std::vector<typename SU<ncolour>::Matrix> ta(Dimension);
+    Vector<typename SU<ncolour>::Matrix> ta(Dimension);
 
     // Debug lines
     // LatticeMatrix uno(Uin.Grid());

Grid/qcd/representations/two_index.h

@@ -43,7 +43,7 @@ public:
     U = Zero();
     LatticeColourMatrix tmp(Uin.Grid());
 
-    std::vector<typename GaugeGroup<ncolour,group_name>::Matrix> eij(Dimension);
+    Vector<typename GaugeGroup<ncolour,group_name>::Matrix> eij(Dimension);
 
     for (int a = 0; a < Dimension; a++)
       GaugeGroupTwoIndex<ncolour, S, group_name>::base(a, eij[a]);

Grid/qcd/smearing/GaugeConfigurationMasked.h

@@ -32,7 +32,9 @@ private:
   //  Smear_Stout<Gimpl> *StoutSmearing;
   //  std::vector<GaugeField> SmearedSet;
   
+  GridRedBlackCartesian * UrbGrid; // keep a copy of the redblack grid for life of object
   std::vector<LatticeLorentzComplex> masks;
+  std::vector<int> cbs;
 
   typedef typename SU3Adjoint::AMatrix AdjMatrix;
   typedef typename SU3Adjoint::LatticeAdjMatrix  AdjMatrixField;
@@ -147,6 +149,25 @@ private:
     }
     pokeLorentz(Fdet, Fdet_pol, nu);
   }
+
+  void Compute_MpInvJx_dNxxdSy(int cb,
+			       const GaugeLinkField &PlaqL,
+			       const GaugeLinkField &PlaqR,
+			       AdjMatrixField MpInvJx,
+			       AdjVectorField &Fdet2 )
+  {
+    GaugeLinkField PlaqLeo(UrbGrid);
+    GaugeLinkField PlaqReo(UrbGrid);
+    AdjMatrixField MpInvJxeo(UrbGrid);
+    AdjVectorField Fdet2eo(UrbGrid);
+    pickCheckerboard(cb,PlaqLeo,PlaqL);
+    pickCheckerboard(cb,PlaqReo,PlaqR);
+    pickCheckerboard(cb,MpInvJxeo,MpInvJx);
+    Fdet2eo.Checkerboard()=cb;
+    Compute_MpInvJx_dNxxdSy(PlaqLeo,PlaqReo,MpInvJxeo,Fdet2eo);
+    setCheckerboard(Fdet2,Fdet2eo);
+  }
+  
   void Compute_MpInvJx_dNxxdSy(const GaugeLinkField &PlaqL,const GaugeLinkField &PlaqR, AdjMatrixField MpInvJx,AdjVectorField &Fdet2 )
   {
     GaugeLinkField UtaU(PlaqL.Grid());
@@ -278,6 +299,7 @@ public:
     ////////////////////////////////////////////////////////////////////////////////
     // Mask the gauge field
     ////////////////////////////////////////////////////////////////////////////////
+    int cb = cbs[smr];
     auto mask=PeekIndex<LorentzIndex>(masks[smr],mu); // the cb mask
     
     Umsk = U;
@@ -442,7 +464,7 @@ public:
     AdjMatrixField MpInvJx_nu(grid);
     MpInvJx = (-1.0)*MpAdInv * JxAd;// rho is on the plaq factor
 
-    Compute_MpInvJx_dNxxdSy(PlaqL,PlaqR,MpInvJx,FdetV);
+    Compute_MpInvJx_dNxxdSy(cb,PlaqL,PlaqR,MpInvJx,FdetV);
     Fdet2_mu=FdetV;
     Fdet1_mu=Zero();
     
@@ -499,7 +521,7 @@ public:
 
 	time=-usecond();
 	PlaqR=(-1.0)*PlaqR;
-	Compute_MpInvJx_dNxxdSy(PlaqL,PlaqR,MpInvJx,FdetV);
+	Compute_MpInvJx_dNxxdSy(cb,PlaqL,PlaqR,MpInvJx,FdetV);
 	Fdet2_nu = FdetV;
 	time+=usecond();
 	std::cout << GridLogMessage << "Compute_MpInvJx_dNxxSy (occurs 6x) took "<<time<< " us"<<std::endl;
@@ -520,7 +542,7 @@ public:
 	
 
 	MpInvJx_nu = Cshift(MpInvJx,mu,-1);
-	Compute_MpInvJx_dNxxdSy(PlaqL,PlaqR,MpInvJx_nu,FdetV);
+	Compute_MpInvJx_dNxxdSy(cb,PlaqL,PlaqR,MpInvJx_nu,FdetV);
 	Fdet2_nu = Fdet2_nu+FdetV;
 	
 	///////////////// -ve nu /////////////////
@@ -539,7 +561,7 @@ public:
 	Fdet1_nu = Fdet1_nu + transpose(Nxy)*dJdXe_nMpInv_y;
 
 	MpInvJx_nu = Cshift(MpInvJx,nu,1);
-	Compute_MpInvJx_dNxxdSy(PlaqL,PlaqR,MpInvJx_nu,FdetV);
+	Compute_MpInvJx_dNxxdSy(cb,PlaqL,PlaqR,MpInvJx_nu,FdetV);
 	Fdet2_nu = Fdet2_nu+FdetV;
 	
 	// x==
@@ -560,7 +582,7 @@ public:
 
 	MpInvJx_nu = Cshift(MpInvJx,mu,-1);
 	MpInvJx_nu = Cshift(MpInvJx_nu,nu,1);
-	Compute_MpInvJx_dNxxdSy(PlaqL,PlaqR,MpInvJx_nu,FdetV);
+	Compute_MpInvJx_dNxxdSy(cb,PlaqL,PlaqR,MpInvJx_nu,FdetV);
 	Fdet2_nu = Fdet2_nu+FdetV;
 
 	/////////////////////////////////////////////////////////////////////
@@ -589,7 +611,7 @@ public:
 
 	MpInvJx_nu = Cshift(MpInvJx,nu,-1);
 
-	Compute_MpInvJx_dNxxdSy(PlaqL,PlaqR,MpInvJx_nu,FdetV);
+	Compute_MpInvJx_dNxxdSy(cb,PlaqL,PlaqR,MpInvJx_nu,FdetV);
 	Fdet2_mu = Fdet2_mu+FdetV;
 
 	//  __
@@ -609,7 +631,7 @@ public:
 
 	MpInvJx_nu = Cshift(MpInvJx,nu,1);
 
-	Compute_MpInvJx_dNxxdSy(PlaqL,PlaqR,MpInvJx_nu,FdetV);
+	Compute_MpInvJx_dNxxdSy(cb,PlaqL,PlaqR,MpInvJx_nu,FdetV);
 	Fdet2_mu = Fdet2_mu+FdetV;
 	
       }
@@ -931,6 +953,10 @@ private:
 public:
 
   /* Standard constructor */
+  virtual ~SmearedConfigurationMasked()
+  {
+    delete UrbGrid;
+  }
   SmearedConfigurationMasked(GridCartesian* _UGrid, unsigned int Nsmear, Smear_Stout<Gimpl>& Stout)
     : SmearedConfiguration<Gimpl>(_UGrid, Nsmear,Stout)
   {
@@ -939,7 +965,6 @@ public:
     // was resized in base class
     assert(this->SmearedSet.size()==Nsmear);
     
-    GridRedBlackCartesian * UrbGrid;
     UrbGrid = SpaceTimeGrid::makeFourDimRedBlackGrid(_UGrid);
     LatticeComplex one(_UGrid); one = ComplexD(1.0,0.0);
     LatticeComplex tmp(_UGrid);
@@ -947,11 +972,12 @@ public:
     for (unsigned int i = 0; i < this->smearingLevels; ++i) {
 
       masks.push_back(*(new LatticeLorentzComplex(_UGrid)));
-
       int mu= (i/2) %Nd;
       int cb= (i%2);
       LatticeComplex tmpcb(UrbGrid);
 
+      cbs.push_back(cb);
+	
       masks[i]=Zero();
       ////////////////////
       // Setup the mask
@@ -962,7 +988,6 @@ public:
       PokeIndex<LorentzIndex>(masks[i],tmp, mu);
 	
     }
-    delete UrbGrid;
   }
   
   virtual void smeared_force(GaugeField &SigmaTilde) 

Grid/qcd/utils/A2Autils.h

@@ -158,12 +158,12 @@ void A2Autils<FImpl>::MesonField(TensorType &mat,
   int MFrvol = rd*Lblock*Rblock*Nmom;
   int MFlvol = ld*Lblock*Rblock*Nmom;
 
-  std::vector<SpinMatrix_v > lvSum(MFrvol);
+  Vector<SpinMatrix_v > lvSum(MFrvol);
   thread_for( r, MFrvol,{
     lvSum[r] = Zero();
   });
 
-  std::vector<SpinMatrix_s > lsSum(MFlvol);             
+  Vector<SpinMatrix_s > lsSum(MFlvol);             
   thread_for(r,MFlvol,{
     lsSum[r]=scalar_type(0.0);
   });
@@ -346,12 +346,12 @@ void A2Autils<FImpl>::PionFieldXX(Eigen::Tensor<ComplexD,3> &mat,
   int MFrvol = rd*Lblock*Rblock;
   int MFlvol = ld*Lblock*Rblock;
 
-  std::vector<vector_type > lvSum(MFrvol);
+  Vector<vector_type > lvSum(MFrvol);
   thread_for(r,MFrvol,{
     lvSum[r] = Zero();
   });
 
-  std::vector<scalar_type > lsSum(MFlvol);             
+  Vector<scalar_type > lsSum(MFlvol);             
   thread_for(r,MFlvol,{
     lsSum[r]=scalar_type(0.0);
   });
@@ -493,12 +493,12 @@ void A2Autils<FImpl>::PionFieldWVmom(Eigen::Tensor<ComplexD,4> &mat,
   int MFrvol = rd*Lblock*Rblock*Nmom;
   int MFlvol = ld*Lblock*Rblock*Nmom;
 
-  std::vector<vector_type > lvSum(MFrvol);
+  Vector<vector_type > lvSum(MFrvol);
   thread_for(r,MFrvol,{
     lvSum[r] = Zero();
   });
 
-  std::vector<scalar_type > lsSum(MFlvol);             
+  Vector<scalar_type > lsSum(MFlvol);             
   thread_for(r,MFlvol,{
     lsSum[r]=scalar_type(0.0);
   });
@@ -700,13 +700,13 @@ void A2Autils<FImpl>::AslashField(TensorType &mat,
     int MFrvol = rd*Lblock*Rblock*Nem;
     int MFlvol = ld*Lblock*Rblock*Nem;
 
-    std::vector<vector_type> lvSum(MFrvol);
+    Vector<vector_type> lvSum(MFrvol);
     thread_for(r,MFrvol,
     {
       lvSum[r] = Zero();
     });
 
-    std::vector<scalar_type> lsSum(MFlvol);             
+    Vector<scalar_type> lsSum(MFlvol);             
     thread_for(r,MFlvol,
     {
         lsSum[r] = scalar_type(0.0);

Grid/qcd/utils/BaryonUtils.h

@@ -971,9 +971,7 @@ void BaryonUtils<FImpl>::BaryonGamma3pt(
   autoView( vq_ti , q_ti     , AcceleratorRead);
   autoView( vq_tf , q_tf     , AcceleratorRead);
 
-  deviceVector<mobj> my_Dq_spec(2);
+  Vector<mobj> my_Dq_spec{Dq_spec1,Dq_spec2};
-  acceleratorPut(my_Dq_spec[0],Dq_spec1);
-  acceleratorPut(my_Dq_spec[1],Dq_spec2);
   mobj * Dq_spec_p = &my_Dq_spec[0];
 
   if (group == 1) {
@@ -1302,8 +1300,7 @@ void BaryonUtils<FImpl>::SigmaToNucleonEye(const PropagatorField &qq_loop,
   autoView( vd_tf   , qd_tf    , AcceleratorRead);
   autoView( vs_ti   , qs_ti    , AcceleratorRead);
 
-  deviceVector<mobj> my_Dq_spec(1);
+  Vector<mobj> my_Dq_spec{Du_spec};
-  acceleratorPut(my_Dq_spec[0],Du_spec);
   mobj * Dq_spec_p = &my_Dq_spec[0];
 
   if(op == "Q1"){
@@ -1356,8 +1353,7 @@ void BaryonUtils<FImpl>::SigmaToNucleonNonEye(const PropagatorField &qq_ti,
   autoView( vd_tf , qd_tf    , AcceleratorRead  );
   autoView( vs_ti , qs_ti    , AcceleratorRead  );
   
-  deviceVector<mobj> my_Dq_spec(1);
+  Vector<mobj> my_Dq_spec{Du_spec};
-  acceleratorPut(my_Dq_spec[0],Du_spec);
   mobj * Dq_spec_p = &my_Dq_spec[0];
 
   if(op == "Q1"){
@@ -1548,9 +1544,7 @@ void BaryonUtils<FImpl>::XiToSigmaEye(const PropagatorField &qq_loop,
   autoView( vd_tf   , qd_tf    , AcceleratorRead);
   autoView( vs_ti   , qs_ti    , AcceleratorRead);
 
-  deviceVector<mobj> my_Dq_spec(2);
+  Vector<mobj> my_Dq_spec{Dd_spec,Ds_spec};
-  acceleratorPut(my_Dq_spec[0],Dd_spec);
-  acceleratorPut(my_Dq_spec[0],Ds_spec);
   mobj * Dq_spec_p = &my_Dq_spec[0];
 
   if(op == "Q1"){

Grid/qcd/utils/GaugeGroup.h

@@ -418,33 +418,33 @@ static void LieAlgebraProject(LatticeAlgebraMatrix &out,const LatticeMatrix &in,
   int hNNm1= NNm1/2;
   RealD sqrt_2 = sqrt(2.0);
   Complex ci(0.0,1.0);
+
+  const int nsimd=  Matrix::Nsimd();
+  accelerator_for(ss,grid->oSites(),nsimd,{
       for(int su2Index=0;su2Index<hNNm1;su2Index++){
 	int i1, i2;
 	su2SubGroupIndex(i1, i2, su2Index);
 	int ax = su2Index*2;
 	int ay = su2Index*2+1;
-    accelerator_for(ss,grid->oSites(),1,{
 	// in is traceless ANTI-hermitian whereas Grid generators are Hermitian.
 	// trace( Ta x Ci in)
 	// Bet I need to move to real part with mult by -i
-	out_v[ss]()()(ax,b) = 0.5*(real(in_v[ss]()()(i2,i1)) - real(in_v[ss]()()(i1,i2)));
+	coalescedWrite(out_v[ss]()()(ax,b),0.5*(real(in_v(ss)()()(i2,i1)) - real(in_v(ss)()()(i1,i2))));
-	out_v[ss]()()(ay,b) = 0.5*(imag(in_v[ss]()()(i1,i2)) + imag(in_v[ss]()()(i2,i1)));
+	coalescedWrite(out_v[ss]()()(ay,b),0.5*(imag(in_v(ss)()()(i1,i2)) + imag(in_v(ss)()()(i2,i1))));
-      });
       }
       for(int diagIndex=0;diagIndex<N-1;diagIndex++){
 	int k = diagIndex + 1; // diagIndex starts from 0
 	int a = NNm1+diagIndex;
 	RealD scale = 1.0/sqrt(2.0*k*(k+1));
-    accelerator_for(ss,grid->oSites(),vComplex::Nsimd(),{
+	auto tmp = in_v(ss)()()(0,0);
-	auto tmp = in_v[ss]()()(0,0);
 	for(int i=1;i<k;i++){
-	  tmp=tmp+in_v[ss]()()(i,i);
+	  tmp=tmp+in_v(ss)()()(i,i);
+	}
+	tmp = tmp - in_v(ss)()()(k,k)*k;
+	coalescedWrite(out_v[ss]()()(a,b),imag(tmp) * scale);
       }
-	tmp = tmp - in_v[ss]()()(k,k)*k;
-	out_v[ss]()()(a,b) =imag(tmp) * scale;
     });
 }
-}
 
   
 };

Grid/qcd/utils/SUn.impl.h

@@ -118,7 +118,7 @@ static void generatorDiagonal(int diagIndex, iGroupMatrix<cplx> &ta) {
 ////////////////////////////////////////////////////////////////////////
 // Map a su2 subgroup number to the pair of rows that are non zero
 ////////////////////////////////////////////////////////////////////////
-static void su2SubGroupIndex(int &i1, int &i2, int su2_index, GroupName::SU) {
+static accelerator_inline void su2SubGroupIndex(int &i1, int &i2, int su2_index, GroupName::SU) {
   assert((su2_index >= 0) && (su2_index < (ncolour * (ncolour - 1)) / 2));
 
   int spare = su2_index;

Grid/qcd/utils/SUnAdjoint.h

@@ -62,7 +62,7 @@ public:
     // returns i(T_Adj)^index necessary for the projectors
     // see definitions above
     iAdjTa = Zero();
-    iSUnMatrix<cplx> ta[ncolour * ncolour - 1];
+    Vector<iSUnMatrix<cplx> > ta(ncolour * ncolour - 1);
     iSUnMatrix<cplx> tmp;
 
     // FIXME not very efficient to get all the generators everytime

Grid/stencil/GeneralLocalStencil.h

@@ -72,7 +72,7 @@ public:
   }
 
   // Resident in managed memory
-  deviceVector<GeneralStencilEntry>  _entries; 
+  Vector<GeneralStencilEntry>  _entries; 
 
   GeneralLocalStencil(GridBase *grid, const std::vector<Coordinate> &shifts)
   {
@@ -141,7 +141,7 @@ public:
 	  ////////////////////////////////////////////////
 	  // Store in look up table
 	  ////////////////////////////////////////////////
-	  acceleratorPut(this->_entries[lex],SE);
+	  this->_entries[lex] = SE;
 	}
       });
   }

Grid/stencil/Lebesgue.cc

@@ -1,4 +1,3 @@
-#if 0
 /*************************************************************************************
 
     Grid physics library, www.github.com/paboyle/Grid 
@@ -242,4 +241,3 @@ void LebesgueOrder::ZGraph(void)
 }
 NAMESPACE_END(Grid);
 
-#endif

Grid/stencil/Lebesgue.h

@@ -72,7 +72,7 @@ public:
   void ThreadInterleave(void);
 
 private:
-  deviceVector<IndexInteger> _LebesgueReorder;
+  Vector<IndexInteger> _LebesgueReorder;
 
 };    
 

Grid/stencil/SimpleCompressor.h

@@ -19,7 +19,7 @@ public:
   static int PartialCompressionFactor(GridBase *grid) {return 1;};
   // Decompress is after merge so ok
   template<class vobj,class cobj,class compressor> 
-  static void Gather_plane_simple (deviceVector<std::pair<int,int> >& table,
+  static void Gather_plane_simple (commVector<std::pair<int,int> >& table,
 				   const Lattice<vobj> &rhs,
 				   cobj *buffer,
 				   compressor &compress,
@@ -35,7 +35,7 @@ public:
     rhs_v.ViewClose();
   }
   template<class vobj,class cobj,class compressor>
-  static void Gather_plane_exchange(deviceVector<std::pair<int,int> >& table,const Lattice<vobj> &rhs,
+  static void Gather_plane_exchange(commVector<std::pair<int,int> >& table,const Lattice<vobj> &rhs,
 				    std::vector<cobj *> pointers,int dimension,int plane,int cbmask,
 				    compressor &compress,int type,int partial)
   {
@@ -83,6 +83,25 @@ public:
 // Wilson compressor will add alternate policies for Dirichlet
 // and possibly partial Dirichlet for DWF
 ////////////////////////////////////
+/*
+class FaceGatherDirichlet
+{
+  // If it's dirichlet we don't assemble comms buffers
+  //
+  // Rely on zeroes in gauge field to drive the correct result
+  // NAN propgagation: field will locally wrap, so fermion should NOT contain NAN and just permute
+  template<class vobj,class cobj,class compressor>
+  static void Gather_plane_simple (commVector<std::pair<int,int> >& table,const Lattice<vobj> &rhs,cobj *buffer,compressor &compress, int off,int so){};
+  template<class vobj,class cobj,class compressor>
+  static void Gather_plane_exchange(commVector<std::pair<int,int> >& table,const Lattice<vobj> &rhs,
+				   Vector<cobj *> pointers,int dimension,int plane,int cbmask,
+				   compressor &compress,int type) {}
+  template<class decompressor,class Merger>
+  static void Merge(decompressor decompress,Merge &mm)  {  }
+  template<class decompressor,class Decompression>
+  static void Decompress(decompressor decompress,Decompression &dd) {}
+};
+*/
 
 template<class vobj,class FaceGather>
 class SimpleCompressorGather : public FaceGather {

Grid/stencil/Stencil.h

@@ -31,6 +31,7 @@
 #define STENCIL_MAX (16)
 
 #include <Grid/stencil/SimpleCompressor.h>   // subdir aggregate
+#include <Grid/stencil/Lebesgue.h>   // subdir aggregate
 #include <Grid/stencil/GeneralLocalStencil.h>
 
 //////////////////////////////////////////////////////////////////////////////////////////
@@ -255,6 +256,7 @@ protected:
   GridBase *                        _grid;
 public:
   GridBase *Grid(void) const { return _grid; }
+  LebesgueOrder *lo;
 
   ////////////////////////////////////////////////////////////////////////
   // Needed to conveniently communicate gparity parameters into GPU memory
@@ -271,11 +273,11 @@ public:
   int face_table_computed;
   int partialDirichlet;
   int fullDirichlet;
-  std::vector<deviceVector<std::pair<int,int> > > face_table ;
+  std::vector<commVector<std::pair<int,int> > > face_table ;
-  deviceVector<int> surface_list;
+  Vector<int> surface_list;
 
-  std::vector<StencilEntry>  _entries; // Resident in host memory
+  stencilVector<StencilEntry>  _entries; // Resident in managed memory
-  deviceVector<StencilEntry>     _entries_device; // Resident in device memory
+  commVector<StencilEntry>     _entries_device; // Resident in device memory
   std::vector<Packet> Packets;
   std::vector<Merge> Mergers;
   std::vector<Merge> MergersSHM;
@@ -365,9 +367,10 @@ public:
   void CommunicateBegin(std::vector<std::vector<CommsRequest_t> > &reqs)
   {
     // All GPU kernel tasks must complete
-    accelerator_barrier();     // All kernels should ALREADY be complete
+    //    accelerator_barrier();     // All kernels should ALREADY be complete
-    _grid->StencilBarrier();   // Everyone is here, so noone running slow and still using receive buffer
+    //    _grid->StencilBarrier();   // Everyone is here, so noone running slow and still using receive buffer
                                // But the HaloGather had a barrier too.
+#ifdef ACCELERATOR_AWARE_MPI
     for(int i=0;i<Packets.size();i++){
       _grid->StencilSendToRecvFromBegin(MpiReqs,
 					Packets[i].send_buf,
@@ -376,6 +379,23 @@ public:
 					Packets[i].from_rank,Packets[i].do_recv,
 					Packets[i].xbytes,Packets[i].rbytes,i);
     }
+#else
+#warning "Using COPY VIA HOST BUFFERS IN STENCIL"
+    for(int i=0;i<Packets.size();i++){
+      // Introduce a host buffer with a cheap slab allocator and zero cost wipe all
+      Packets[i].host_send_buf = _grid->HostBufferMalloc(Packets[i].xbytes);
+      Packets[i].host_recv_buf = _grid->HostBufferMalloc(Packets[i].rbytes);
+      if ( Packets[i].do_send ) {
+	acceleratorCopyFromDevice(Packets[i].send_buf, Packets[i].host_send_buf,Packets[i].xbytes);
+      }
+      _grid->StencilSendToRecvFromBegin(MpiReqs,
+					Packets[i].host_send_buf,
+					Packets[i].to_rank,Packets[i].do_send,
+					Packets[i].host_recv_buf,
+					Packets[i].from_rank,Packets[i].do_recv,
+					Packets[i].xbytes,Packets[i].rbytes,i);
+    }
+#endif
     // Get comms started then run checksums
     // Having this PRIOR to the dslash seems to make Sunspot work... (!)
     for(int i=0;i<Packets.size();i++){
@@ -390,9 +410,18 @@ public:
     if   ( this->partialDirichlet ) DslashLogPartial();
     else if ( this->fullDirichlet ) DslashLogDirichlet();
     else DslashLogFull();
-    acceleratorCopySynchronise();// is in the StencilSendToRecvFromComplete
+    // acceleratorCopySynchronise() is in the StencilSendToRecvFromComplete
-    accelerator_barrier(); 
+    //    accelerator_barrier(); 
     _grid->StencilBarrier(); 
+#ifndef ACCELERATOR_AWARE_MPI
+#warning "Using COPY VIA HOST BUFFERS IN STENCIL"
+    for(int i=0;i<Packets.size();i++){
+      if ( Packets[i].do_recv ) {
+	acceleratorCopyToDevice(Packets[i].host_recv_buf, Packets[i].recv_buf,Packets[i].rbytes);
+      }
+    }
+    _grid->HostBufferFreeAll();
+#endif
     // run any checksums
     for(int i=0;i<Packets.size();i++){
       if ( Packets[i].do_recv )
@@ -473,7 +502,7 @@ public:
   template<class compressor>
   void HaloGather(const Lattice<vobj> &source,compressor &compress)
   {
-    accelerator_barrier();
+    //    accelerator_barrier();
     _grid->StencilBarrier();// Synch shared memory on a single nodes
 
     assert(source.Grid()==_grid);
@@ -487,7 +516,6 @@ public:
       HaloGatherDir(source,compress,point,face_idx);
     }
     accelerator_barrier(); // All my local gathers are complete
-    _grid->StencilBarrier();// Synch shared memory on a single nodes
     face_table_computed=1;
     assert(u_comm_offset==_unified_buffer_size);
   }
@@ -640,7 +668,7 @@ public:
     for(int point=0;point<this->_npoints;point++){
       this->same_node[point] = this->SameNode(point);
     }
-    int32_t surface_list_size=0;
+
     for(int site = 0 ;site< vol4;site++){
       int local = 1;
       for(int point=0;point<this->_npoints;point++){
@@ -650,30 +678,11 @@ public:
       }
       if(local == 0) {
 	for(int s=0;s<Ls;s++){
-	  surface_list_size++;
+	  surface_list.push_back(site*Ls+s);
 	}
       }
     }
-    std::cout << "BuildSurfaceList size is "<<surface_list.size()<<std::endl;
+    //std::cout << "BuildSurfaceList size is "<<surface_list.size()<<std::endl;
-    surface_list.resize(surface_list_size);
-    std::vector<int> surface_list_host(surface_list_size);
-    int32_t ss=0;
-    for(int site = 0 ;site< vol4;site++){
-      int local = 1;
-      for(int point=0;point<this->_npoints;point++){
-	if( (!this->GetNodeLocal(site*Ls,point)) && (!this->same_node[point]) ){
-	  local = 0;
-	}
-      }
-      if(local == 0) {
-	for(int s=0;s<Ls;s++){
-	  int idx=site*Ls+s;
-	  surface_list_host[ss]= idx;
-	  ss++;
-	}
-      }
-    }
-    acceleratorCopyToDevice(&surface_list_host[0],&surface_list[0],surface_list_size*sizeof(int));
   }
   /// Introduce a block structure and switch off comms on boundaries
   void DirichletBlock(const Coordinate &dirichlet_block)

Grid/threads/Accelerator.cc

@@ -207,10 +207,10 @@ cl::sycl::queue *theCopyAccelerator;
 void acceleratorInit(void)
 {
   int nDevices = 1;
-  //  cl::sycl::gpu_selector selector;
+  cl::sycl::gpu_selector selector;
-  //  cl::sycl::device selectedDevice { selector };
+  cl::sycl::device selectedDevice { selector };
-  theGridAccelerator = new sycl::queue (sycl::gpu_selector_v);
+  theGridAccelerator = new sycl::queue (selectedDevice);
-  theCopyAccelerator = new sycl::queue (sycl::gpu_selector_v);
+  theCopyAccelerator = new sycl::queue (selectedDevice);
   //  theCopyAccelerator = theGridAccelerator; // Should proceed concurrenlty anyway.
 
 #ifdef GRID_SYCL_LEVEL_ZERO_IPC

Grid/util/Init.cc

@@ -464,12 +464,16 @@ void Grid_init(int *argc,char ***argv)
     std::cout<<GridLogMessage<<std::endl;
     std::cout<<GridLogMessage<<"Performance:"<<std::endl;
     std::cout<<GridLogMessage<<std::endl;
+    std::cout<<GridLogMessage<<"  --comms-concurrent : Asynchronous MPI calls; several dirs at a time "<<std::endl;    
+    std::cout<<GridLogMessage<<"  --comms-sequential : Synchronous MPI calls; one dirs at a time "<<std::endl;    
     std::cout<<GridLogMessage<<"  --comms-overlap    : Overlap comms with compute "<<std::endl;    
     std::cout<<GridLogMessage<<std::endl;
     std::cout<<GridLogMessage<<"  --dslash-generic: Wilson kernel for generic Nc"<<std::endl;    
     std::cout<<GridLogMessage<<"  --dslash-unroll : Wilson kernel for Nc=3"<<std::endl;    
     std::cout<<GridLogMessage<<"  --dslash-asm    : Wilson kernel for AVX512"<<std::endl;    
     std::cout<<GridLogMessage<<std::endl;
+    std::cout<<GridLogMessage<<"  --lebesgue      : Cache oblivious Lebesgue curve/Morton order/Z-graph stencil looping"<<std::endl;    
+    std::cout<<GridLogMessage<<"  --cacheblocking n.m.o.p : Hypercuboidal cache blocking"<<std::endl;    
     std::cout<<GridLogMessage<<std::endl;
     exit(EXIT_SUCCESS);
   }
@@ -497,8 +501,28 @@ void Grid_init(int *argc,char ***argv)
     WilsonKernelsStatic::Comms = WilsonKernelsStatic::CommsThenCompute;
     StaggeredKernelsStatic::Comms = StaggeredKernelsStatic::CommsThenCompute;
   }
+  if( GridCmdOptionExists(*argv,*argv+*argc,"--comms-concurrent") ){
+    CartesianCommunicator::SetCommunicatorPolicy(CartesianCommunicator::CommunicatorPolicyConcurrent);
+  }
+  if( GridCmdOptionExists(*argv,*argv+*argc,"--comms-sequential") ){
+    CartesianCommunicator::SetCommunicatorPolicy(CartesianCommunicator::CommunicatorPolicySequential);
+  }
 
+  if( GridCmdOptionExists(*argv,*argv+*argc,"--lebesgue") ){
+    LebesgueOrder::UseLebesgueOrder=1;
+  }
   CartesianCommunicator::nCommThreads = 1;
+#ifdef GRID_COMMS_THREADS  
+  if( GridCmdOptionExists(*argv,*argv+*argc,"--comms-threads") ){
+    arg= GridCmdOptionPayload(*argv,*argv+*argc,"--comms-threads");
+    GridCmdOptionInt(arg,CartesianCommunicator::nCommThreads);
+    assert(CartesianCommunicator::nCommThreads > 0);
+  }
+#endif  
+  if( GridCmdOptionExists(*argv,*argv+*argc,"--cacheblocking") ){
+    arg= GridCmdOptionPayload(*argv,*argv+*argc,"--cacheblocking");
+    GridCmdOptionIntVector(arg,LebesgueOrder::Block);
+  }
   if( GridCmdOptionExists(*argv,*argv+*argc,"--notimestamp") ){
     GridLogTimestamp(0);
   } else {
@@ -549,31 +573,8 @@ void GridLogLayout() {
 
 void * Grid_backtrace_buffer[_NBACKTRACE];
 
-void Grid_usr_signal_handler(int sig,siginfo_t *si,void * ptr)
-{
-  fprintf(stderr,"Signal handler on host %s\n",hostname);
-  fprintf(stderr,"Caught signal %d\n",si->si_signo);
-  fprintf(stderr,"  mem address %llx\n",(unsigned long long)si->si_addr);
-  fprintf(stderr,"         code %d\n",si->si_code);
-  // x86 64bit
-#ifdef __linux__
-#ifdef __x86_64__
-  ucontext_t * uc= (ucontext_t *)ptr;
-  struct sigcontext *sc = (struct sigcontext *)&uc->uc_mcontext;
-  fprintf(stderr,"  instruction %llx\n",(unsigned long long)sc->rip);
-#endif
-#endif
-  fflush(stderr);
-  BACKTRACEFP(stderr);
-  fprintf(stderr,"Called backtrace\n");
-  fflush(stdout);
-  fflush(stderr);
-  return;
-}
-
 void Grid_sa_signal_handler(int sig,siginfo_t *si,void * ptr)
 {
-  fprintf(stderr,"Signal handler on host %s\n",hostname);
   fprintf(stderr,"Caught signal %d\n",si->si_signo);
   fprintf(stderr,"  mem address %llx\n",(unsigned long long)si->si_addr);
   fprintf(stderr,"         code %d\n",si->si_code);
@@ -584,7 +585,7 @@ void Grid_sa_signal_handler(int sig,siginfo_t *si,void * ptr)
   ucontext_t * uc= (ucontext_t *)ptr;
   struct sigcontext *sc = (struct sigcontext *)&uc->uc_mcontext;
   fprintf(stderr,"  instruction %llx\n",(unsigned long long)sc->rip);
-#define REG(A)  fprintf(stderr,"  %s %lx\n",#A,sc-> A);
+#define REG(A)  printf("  %s %lx\n",#A,sc-> A);
   REG(rdi);
   REG(rsi);
   REG(rbp);
@@ -617,8 +618,8 @@ void Grid_sa_signal_handler(int sig,siginfo_t *si,void * ptr)
 
 void Grid_exit_handler(void)
 {
-  //  BACKTRACEFP(stdout);
+  BACKTRACEFP(stdout);
-  //  fflush(stdout);
+  fflush(stdout);
 }
 void Grid_debug_handler_init(void)
 {
@@ -626,10 +627,10 @@ void Grid_debug_handler_init(void)
   sigemptyset (&sa.sa_mask);
   sa.sa_sigaction= Grid_sa_signal_handler;
   sa.sa_flags    = SA_SIGINFO;
-  //  sigaction(SIGSEGV,&sa,NULL);
+  sigaction(SIGSEGV,&sa,NULL);
   sigaction(SIGTRAP,&sa,NULL);
   sigaction(SIGBUS,&sa,NULL);
-  //  sigaction(SIGUSR2,&sa,NULL);
+  sigaction(SIGUSR2,&sa,NULL);
 
   feenableexcept( FE_INVALID|FE_OVERFLOW|FE_DIVBYZERO);
 
@@ -637,14 +638,7 @@ void Grid_debug_handler_init(void)
   sigaction(SIGKILL,&sa,NULL);
   sigaction(SIGILL,&sa,NULL);
 
-  // Non terminating SIGUSR1/2 handler
+  atexit(Grid_exit_handler);
-  struct sigaction sa_ping;
-  sigemptyset (&sa_ping.sa_mask);
-  sa_ping.sa_sigaction= Grid_usr_signal_handler;
-  sa_ping.sa_flags    = SA_SIGINFO;
-  sigaction(SIGHUP,&sa_ping,NULL);
-
-  //  atexit(Grid_exit_handler);
 }
 
 NAMESPACE_END(Grid);

benchmarks/Benchmark_ITT.cc

@@ -644,6 +644,11 @@ int main (int argc, char ** argv)
   Grid_init(&argc,&argv);
 
   CartesianCommunicator::SetCommunicatorPolicy(CartesianCommunicator::CommunicatorPolicySequential);
+#ifdef KNL
+  LebesgueOrder::Block = std::vector<int>({8,2,2,2});
+#else
+  LebesgueOrder::Block = std::vector<int>({2,2,2,2});
+#endif
   Benchmark::Decomposition();
 
   int do_su4=1;

benchmarks/Benchmark_memory_asynch.cc

@@ -70,7 +70,7 @@ int main (int argc, char ** argv)
     pRNG.SeedFixedIntegers(std::vector<int>({56,17,89,101}));
 
     std::vector<double> stop(threads);
-    std::vector<Vec> sum(threads);
+    Vector<Vec> sum(threads);
 
     std::vector<LatticeVec> x(threads,&Grid);
     for(int t=0;t<threads;t++){

benchmarks/Benchmark_mooee.cc

@@ -78,9 +78,9 @@ int main (int argc, char ** argv)
     double t0,t1;
     
     typedef typename DomainWallFermionD::Coeff_t Coeff_t;
-    std::vector<Coeff_t> diag = Dw.bs;
+    Vector<Coeff_t> diag = Dw.bs;
-    std::vector<Coeff_t> upper= Dw.cs;
+    Vector<Coeff_t> upper= Dw.cs;
-    std::vector<Coeff_t> lower= Dw.cs;
+    Vector<Coeff_t> lower= Dw.cs;
     upper[Ls-1]=-Dw.mass_minus*upper[Ls-1];
     lower[0]   =-Dw.mass_plus*lower[0];
     

benchmarks/Benchmark_usqcd.cc

@@ -861,7 +861,7 @@ int main (int argc, char ** argv)
   }
 
   CartesianCommunicator::SetCommunicatorPolicy(CartesianCommunicator::CommunicatorPolicySequential);
-  //  LebesgueOrder::Block = std::vector<int>({2,2,2,2});
+  LebesgueOrder::Block = std::vector<int>({2,2,2,2});
 
   Benchmark::Decomposition();
 

configure.ac

@@ -225,6 +225,18 @@ case ${ac_SFW_FP16} in
       AC_MSG_ERROR(["SFW FP16 option not supported ${ac_SFW_FP16}"]);;
 esac
 
+############### Default to accelerator cshift, but revert to host if UCX is buggy or other reasons
+AC_ARG_ENABLE([accelerator-aware-mpi],
+    [AS_HELP_STRING([--enable-accelerator-aware-mpi=yes|no],[run mpi transfers from device])],
+    [ac_ACCELERATOR_AWARE_MPI=${enable_accelerator_aware_mpi}], [ac_ACCELERATOR_AWARE_MPI=yes])
+
+case ${ac_ACCELERATOR_AWARE_MPI} in
+    yes)
+      AC_DEFINE([ACCELERATOR_CSHIFT],[1],[ Cshift runs on host])
+      AC_DEFINE([ACCELERATOR_AWARE_MPI],[1],[ Stencil can use device pointers]);;
+    *);;
+esac
+
 
 ############### SYCL/CUDA/HIP/none
 AC_ARG_ENABLE([accelerator],
@@ -652,6 +664,16 @@ case ${ac_SHM_FAST_PATH} in
      *) ;;
 esac
 
+############### communication type selection
+AC_ARG_ENABLE([comms-threads],[AS_HELP_STRING([--enable-comms-threads | --disable-comms-threads],[Use multiple threads in MPI calls])],[ac_COMMS_THREADS=${enable_comms_threads}],[ac_COMMS_THREADS=yes])
+
+case ${ac_COMMS_THREADS} in
+     yes)
+        AC_DEFINE([GRID_COMMS_THREADING],[1],[GRID_COMMS_NONE] )
+      ;;
+     *) ;;
+esac
+
 ############### communication type selection
 AC_ARG_ENABLE([comms],[AS_HELP_STRING([--enable-comms=none|mpi|mpi-auto],[Select communications])],[ac_COMMS=${enable_comms}],[ac_COMMS=none])
 

systems/Aurora-AOT/config-command

@@ -1,23 +0,0 @@
-#Ahead of time compile for PVC
-export LDFLAGS="-fiopenmp -fsycl -fsycl-device-code-split=per_kernel -fsycl-targets=spir64_gen -Xs -device -Xs pvc -fsycl-device-lib=all -lze_loader -L${MKLROOT}/lib -qmkl=parallel  -fsycl  -lsycl " 
-export CXXFLAGS="-O3 -fiopenmp -fsycl-unnamed-lambda -fsycl -Wno-tautological-compare -qmkl=parallel  -fsycl -fno-exceptions -fsycl-targets=spir64_gen -Xs -device -Xs pvc "
-
-#JIT compile 
-#export LDFLAGS="-fiopenmp -fsycl -fsycl-device-code-split=per_kernel  -fsycl-device-lib=all -lze_loader -L${MKLROOT}/lib -qmkl=parallel  -fsycl  -lsycl " 
-#export CXXFLAGS="-O3 -fiopenmp -fsycl-unnamed-lambda -fsycl -Wno-tautological-compare -qmkl=parallel  -fsycl -fno-exceptions "
-
-../../configure \
-	--enable-simd=GPU \
-	--enable-gen-simd-width=64 \
-	--enable-comms=mpi-auto \
-	--enable-debug \
-	--disable-gparity \
-	--disable-fermion-reps \
-	--with-lime=$CLIME \
-	--enable-shm=nvlink \
-	--enable-accelerator=sycl \
-	--enable-accelerator-aware-mpi=yes\
-	--enable-unified=no \
-	MPICXX=mpicxx \
-	CXX=icpx 
-

systems/Aurora-AOT/sourceme.sh

@@ -1,15 +0,0 @@
-#module load oneapi/release/2023.12.15.001
-#module load mpich/icc-all-debug-pmix-gpu/52.2
-#module load mpich-config/mode/deterministic
-#module load intel_compute_runtime/release/821.35
-
-source ~/spack/share/spack/setup-env.sh 
-spack load c-lime
-spack load openssl
-export CLIME=`spack find --paths c-lime | grep ^c-lime | awk '{print $2}' `
-export HTTP_PROXY=http://proxy.alcf.anl.gov:3128
-export HTTPS_PROXY=http://proxy.alcf.anl.gov:3128
-export http_proxy=http://proxy.alcf.anl.gov:3128
-export https_proxy=http://proxy.alcf.anl.gov:3128
-git config --global http.proxy http://proxy.alcf.anl.gov:3128
-export SYCL_PROGRAM_COMPILE_OPTIONS="-ze-opt-large-register-file"

systems/Aurora-AOT/tests/reproBigJob.pbs

@@ -1,74 +0,0 @@
-#!/bin/bash
-
-#PBS -l select=512
-#PBS -q EarlyAppAccess
-#PBS -A LatticeQCD_aesp_CNDA
-#PBS -l walltime=6:00:00
-#PBS -N reproBigJob
-#PBS -k doe
-
-#export OMP_PROC_BIND=spread
-#unset OMP_PLACES
-
-#module load oneapi/eng-compiler/2023.05.15.003
-#module load mpich/51.2/icc-all-deterministic-pmix-gpu
-
-# 56 cores / 6 threads ~9
-export OMP_NUM_THREADS=6
-export MPIR_CVAR_CH4_OFI_ENABLE_GPU_PIPELINE=1
-export MPIR_CVAR_CH4_OFI_GPU_PIPELINE_D2H_ENGINE_TYPE=0
-export MPIR_CVAR_CH4_OFI_GPU_PIPELINE_H2D_ENGINE_TYPE=0
-export MPIR_CVAR_CH4_OFI_GPU_PIPELINE_BUFFER_SZ=10485760
-export MPIR_CVAR_CH4_OFI_GPU_PIPELINE_THRESHOLD=131072
-export MPIR_CVAR_CH4_OFI_GPU_PIPELINE_NUM_BUFFERS_PER_CHUNK=16
-export MPIR_CVAR_CH4_OFI_GPU_PIPELINE_MAX_NUM_BUFFERS=16
-#export MPIR_CVAR_GPU_USE_IMMEDIATE_COMMAND_LIST=1
-
-#export SYCL_PI_LEVEL_ZERO_USE_IMMEDIATE_COMMANDLISTS=1
-export SYCL_PI_LEVEL_ZERO_USE_COPY_ENGINE=1
-export SYCL_PI_LEVEL_ZERO_USE_COPY_ENGINE_FOR_D2D_COPY=1
-export SYCL_PROGRAM_COMPILE_OPTIONS="-ze-opt-large-register-file"
-
-export GRID_PRINT_ENTIRE_LOG=0
-export GRID_CHECKSUM_RECV_BUF=0
-export GRID_CHECKSUM_SEND_BUF=0
-
-export MPICH_OFI_NIC_POLICY=GPU
-
-#export MPIR_CVAR_ALLREDUCE_DEVICE_COLLECTIVE=0
-#export MPIR_CVAR_REDUCE_DEVICE_COLLECTIVE=0
-#export MPIR_CVAR_ALLREDUCE_INTRA_ALGORITHM=recursive_doubling
-#unset MPIR_CVAR_CH4_COLL_SELECTION_TUNING_JSON_FILE
-#unset MPIR_CVAR_COLL_SELECTION_TUNING_JSON_FILE
-#unset MPIR_CVAR_CH4_POSIX_COLL_SELECTION_TUNING_JSON_FILE
-
-cd $PBS_O_WORKDIR
-
-cp $PBS_NODEFILE nodefile
-
-DIR=reproBigJob.$PBS_JOBID
-
-mkdir -p $DIR
-cd $DIR
-
-cp $PBS_NODEFILE nodefile
-
-BINARY=../Test_dwf_mixedcg_prec
-
-echo > pingjob <<EOF
-while read node ; 
-do
-	echo ssh $node killall -s USR1 -- ../Test_dwf_mixedcg_prec
-done < nodefile
-EOF
-
-CMD="mpiexec -np 6144 -ppn 12  -envall --hostfile nodefile \
-	     ../gpu_tile_compact.sh \
-	     $BINARY --mpi 8.8.8.12 --grid 128.128.128.288 \
-	--shm-mpi 0 --shm 4096 --device-mem 32000 --accelerator-threads 32 --seconds 18000 --debug-stdout --log Message --debug-signals --comms-overlap"
-
-echo $CMD > command-line
-env > environment
-$CMD
-grep Oops Grid.stderr.* > failures.$PBS_JOBID
-rm core.*

systems/Aurora/benchmarks/bench1.pbs

@@ -1,38 +1,67 @@
 #!/bin/bash
 
-#PBS -q EarlyAppAccess
+#PBS -q debug
 #PBS -l select=1
 #PBS -l walltime=00:20:00
 #PBS -A LatticeQCD_aesp_CNDA
 
+#export OMP_PROC_BIND=spread
+#unset OMP_PLACES
+
 cd $PBS_O_WORKDIR
 
 source ../sourceme.sh
+module load pti-gpu
 
-cp $PBS_NODEFILE nodefile
+#cat $PBS_NODEFILE
 
 export OMP_NUM_THREADS=4
 export MPIR_CVAR_CH4_OFI_ENABLE_GPU_PIPELINE=1
-unset MPIR_CVAR_CH4_OFI_GPU_PIPELINE_D2H_ENGINE_TYPE
+
-unset MPIR_CVAR_CH4_OFI_GPU_PIPELINE_H2D_ENGINE_TYPE
+#unset MPIR_CVAR_CH4_OFI_GPU_PIPELINE_D2H_ENGINE_TYPE
-unset MPIR_CVAR_GPU_USE_IMMEDIATE_COMMAND_LIST
+#unset MPIR_CVAR_CH4_OFI_GPU_PIPELINE_H2D_ENGINE_TYPE
+#unset MPIR_CVAR_GPU_USE_IMMEDIATE_COMMAND_LIST
+
 #export MPIR_CVAR_CH4_OFI_GPU_PIPELINE_D2H_ENGINE_TYPE=0
 #export MPIR_CVAR_CH4_OFI_GPU_PIPELINE_H2D_ENGINE_TYPE=0
 #export MPIR_CVAR_GPU_USE_IMMEDIATE_COMMAND_LIST=1
-export MPIR_CVAR_CH4_OFI_GPU_PIPELINE_BUFFER_SZ=1048576
+#export MPIR_CVAR_CH4_OFI_GPU_PIPELINE_BUFFER_SZ=1048576
-export MPIR_CVAR_CH4_OFI_GPU_PIPELINE_THRESHOLD=131072
+#export MPIR_CVAR_CH4_OFI_GPU_PIPELINE_THRESHOLD=131072
-export MPIR_CVAR_CH4_OFI_GPU_PIPELINE_NUM_BUFFERS_PER_CHUNK=16
+#export MPIR_CVAR_CH4_OFI_GPU_PIPELINE_NUM_BUFFERS_PER_CHUNK=16
-export MPIR_CVAR_CH4_OFI_GPU_PIPELINE_MAX_NUM_BUFFERS=16
+#export MPIR_CVAR_CH4_OFI_GPU_PIPELINE_MAX_NUM_BUFFERS=16
 export MPICH_OFI_NIC_POLICY=GPU
 
+# 12 ppn, 2 nodes, 24 ranks
+#
 CMD="mpiexec -np 12 -ppn 12  -envall \
-	     ./Benchmark_dwf_fp32 --mpi 2.1.2.3 --grid 32.32.64.48 \
+	     ./gpu_tile_compact.sh \
-		--shm-mpi 1 --shm 2048 --device-mem 32000 --accelerator-threads 32 --debug-signals"
+	     ./Benchmark_comms_host_device --mpi 2.2.1.3 --grid 24.32.32.24 \
+		--shm-mpi 0 --shm 2048 --device-mem 32000 --accelerator-threads 32" 
+#$CMD | tee 1node.comms
 
-#for f in 1 2 3 4 5 6 7 8
+
-for f in 1
+CMD="mpiexec -np 1 -ppn 1  -envall \
-do
+	     ./gpu_tile_compact.sh \
-echo $CMD
+	     ./Benchmark_dwf_fp32 --mpi 1.1.1.1 --grid 16.32.32.32 \
-$CMD | tee 1node.32.32.64.48.dwf.hbm.$f
+		--shm-mpi 0 --shm 2048 --device-mem 32000 --accelerator-threads 32 "
-done
+#$CMD | tee 1tile.dwf
+
+CMD="mpiexec -np 12 -ppn 12  -envall \
+	     ./gpu_tile_compact.sh \
+	     ./Benchmark_dwf_fp32 --mpi 2.2.1.3 --grid 32.32.32.48 \
+		--shm-mpi 0 --shm 2048 --device-mem 32000 --accelerator-threads 32 --comms-overlap"
+$CMD | tee 1node.32.32.32.48.dwf
+
+
+CMD="mpiexec -np 12 -ppn 12  -envall \
+	     ./gpu_tile_compact.sh \
+	     ./Benchmark_dwf_fp32 --mpi 2.2.1.3 --grid 64.64.32.96 \
+		--shm-mpi 0 --shm 2048 --device-mem 32000 --accelerator-threads 32 --comms-overlap"
+#$CMD | tee 1node.64.64.32.96.dwf
+
+CMD="mpiexec -np 12 -ppn 12  -envall \
+	     ./gpu_tile_compact.sh \
+	     ./Benchmark_dwf_fp32 --mpi 2.2.1.3 --grid 64.32.32.48 \
+		--shm-mpi 0 --shm 2048 --device-mem 32000 --accelerator-threads 32 --comms-overlap"
+#$CMD | tee 1node.64.32.32.48.dwf
 

systems/Aurora/benchmarks/bench2.pbs

@@ -1,6 +1,6 @@
 #!/bin/bash
 
-#PBS -q EarlyAppAccess
+#PBS -q workq
 #PBS -l select=2
 #PBS -l walltime=00:20:00
 #PBS -A LatticeQCD_aesp_CNDA
@@ -11,16 +11,17 @@
 cd $PBS_O_WORKDIR
 
 source ../sourceme.sh
-#module load pti-gpu
+module load pti-gpu
 
-
+#cat $PBS_NODEFILE
-cp $PBS_NODEFILE nodefile
 
 export OMP_NUM_THREADS=4
 export MPIR_CVAR_CH4_OFI_ENABLE_GPU_PIPELINE=1
+
 #unset MPIR_CVAR_CH4_OFI_GPU_PIPELINE_D2H_ENGINE_TYPE
 #unset MPIR_CVAR_CH4_OFI_GPU_PIPELINE_H2D_ENGINE_TYPE
 #unset MPIR_CVAR_GPU_USE_IMMEDIATE_COMMAND_LIST
+
 export MPIR_CVAR_CH4_OFI_GPU_PIPELINE_D2H_ENGINE_TYPE=0
 export MPIR_CVAR_CH4_OFI_GPU_PIPELINE_H2D_ENGINE_TYPE=0
 export MPIR_CVAR_GPU_USE_IMMEDIATE_COMMAND_LIST=1
@@ -33,26 +34,22 @@ export MPICH_OFI_NIC_POLICY=GPU
 # 12 ppn, 2 nodes, 24 ranks
 #
 CMD="mpiexec -np 24 -ppn 12  -envall \
-	     ./gpu_tile.sh \
+	     ./gpu_tile_compact.sh \
 	     ./Benchmark_comms_host_device --mpi 2.2.2.3 --grid 24.32.32.24 \
 		--shm-mpi 0 --shm 2048 --device-mem 32000 --accelerator-threads 32" 
-#$CMD | tee 2node.comms.hbm
+$CMD | tee 2node.comms
 
 
 CMD="mpiexec -np 24 -ppn 12  -envall \
+	     ./gpu_tile_compact.sh \
 	     ./Benchmark_dwf_fp32 --mpi 2.2.2.3 --grid 32.32.64.48 \
-		--shm-mpi 1 --shm 2048 --device-mem 32000 --accelerator-threads 32 --comms-overlap --debug-signals"
+		--shm-mpi 0 --shm 2048 --device-mem 32000 --accelerator-threads 32 --comms-overlap"
+$CMD | tee 2node.32.32.64.48.dwf
 
-#for f in 1 2 3 4 5 6 7 8
-for f in 1
-do
-echo $CMD
-$CMD | tee 2node.32.32.64.48.dwf.hbm.$f
-done
 
 CMD="mpiexec -np 24 -ppn 12  -envall \
-	     ./gpu_tile.sh \
+	     ./gpu_tile_compact.sh \
 	     ./Benchmark_dwf_fp32 --mpi 2.2.2.3 --grid 64.64.64.96 \
 		--shm-mpi 0 --shm 2048 --device-mem 32000 --accelerator-threads 32 --comms-overlap"
-#$CMD | tee 2node.64.64.64.96.dwf.hbm
+$CMD | tee 2node.64.64.64.96.dwf
 

systems/Aurora/config-command

@@ -1,6 +1,6 @@
 
-export LDFLAGS="-fiopenmp -fsycl -fsycl-device-code-split=per_kernel  -fsycl-device-lib=all -lze_loader -L${MKLROOT}/lib -qmkl=parallel  -fsycl  -lsycl " 
+export LDFLAGS="-fiopenmp -fsycl -fsycl-device-code-split=per_kernel -fsycl-targets=spir64_gen -Xs -device -Xs pvc -fsycl-device-lib=all -lze_loader -L${MKLROOT}/lib -qmkl=parallel  -fsycl  -lsycl " 
-export CXXFLAGS="-O3 -fiopenmp -fsycl-unnamed-lambda -fsycl -I$INSTALL/include -Wno-tautological-compare -I$HOME/ -qmkl=parallel  -fsycl -fno-exceptions "
+export CXXFLAGS="-O3 -fiopenmp -fsycl-unnamed-lambda -fsycl -I$INSTALL/include -Wno-tautological-compare -I$HOME/ -qmkl=parallel  -fsycl -fno-exceptions -fsycl-targets=spir64_gen -Xs -device -Xs pvc "
 ../../configure \
 	--enable-simd=GPU \
 	--enable-gen-simd-width=64 \

systems/Aurora/sourceme.sh

@@ -1,14 +1,40 @@
-module load oneapi/release/2023.12.15.001
-#module load mpich/icc-all-debug-pmix-gpu/52.2
-#module load mpich-config/mode/deterministic
-#module load intel_compute_runtime/release/821.35
 source ~/spack/share/spack/setup-env.sh 
 spack load c-lime
-spack load openssl
 export CLIME=`spack find --paths c-lime | grep ^c-lime | awk '{print $2}' `
+#spack load libefence
+#export EFENCE=`spack find --paths libefence | grep ^libefence | awk '{print $2}' `
+#export LD_LIBRARY_PATH=${EFENCE}/lib:$LD_LIBRARY_PATH
+#spack load gperftools
+export TCMALLOC=/home/paboyle/gperftools/install
+export LD_LIBRARY_PATH=${TCMALLOC}/lib:$LD_LIBRARY_PATH
+export INTELGT_AUTO_ATTACH_DISABLE=1
+
+#export ONEAPI_DEVICE_SELECTOR=level_zero:0.0
+#module load oneapi/release/2023.12.15.001
+#module use /soft/modulefiles
+#module load intel_compute_runtime/release/agama-devel-682.22
+
+#export FI_CXI_DEFAULT_CQ_SIZE=131072
+#export FI_CXI_CQ_FILL_PERCENT=20
+#export SYCL_PROGRAM_COMPILE_OPTIONS="-ze-opt-large-register-file"
+#export SYCL_PROGRAM_COMPILE_OPTIONS="-ze-intel-enable-auto-large-GRF-mode"
+
+#
+# -ftarget-register-alloc-mode=pvc:default 
+# -ftarget-register-alloc-mode=pvc:small
+# -ftarget-register-alloc-mode=pvc:large
+# -ftarget-register-alloc-mode=pvc:auto
+#export MPIR_CVAR_CH4_OFI_ENABLE_HMEM=1
+
 export HTTP_PROXY=http://proxy.alcf.anl.gov:3128
 export HTTPS_PROXY=http://proxy.alcf.anl.gov:3128
 export http_proxy=http://proxy.alcf.anl.gov:3128
 export https_proxy=http://proxy.alcf.anl.gov:3128
 git config --global http.proxy http://proxy.alcf.anl.gov:3128
+
+#source ~/spack/share/spack/setup-env.sh
+#spack load gperftools
+#export TCMALLOC=`spack find --paths gperftools | grep ^gperftools | awk '{print $2}' `
+#export LD_LIBRARY_PATH=${TCMALLOC}/lib:$LD_LIBRARY_PATH
+
 export SYCL_PROGRAM_COMPILE_OPTIONS="-ze-opt-large-register-file"

systems/Aurora/tests/repro16.pbs

@@ -2,8 +2,7 @@
 
 ## qsub -q EarlyAppAccess -A Aurora_Deployment -I -l select=1 -l walltime=60:00
 
-#PBS -l select=16
+#PBS -l select=16:system=sunspot,place=scatter
-#PBS -q EarlyAppAccess
 #PBS -A LatticeQCD_aesp_CNDA
 #PBS -l walltime=01:00:00
 #PBS -N dwf
@@ -14,14 +13,19 @@
 
 cd $PBS_O_WORKDIR
 
-source ../sourceme.sh
+#source ../sourceme.sh
 
 cat $PBS_NODEFILE
 
+#export MPICH_COLL_SYNC=1
+#export MPICH_ENV_DISPLAY=1
+export MPICH_
 export OMP_NUM_THREADS=3
 export MPIR_CVAR_CH4_OFI_ENABLE_GPU_PIPELINE=1
+module load oneapi/eng-compiler/2023.05.15.003
+module load mpich/51.2/icc-all-deterministic-pmix-gpu
+#export LD_LIBRARY_PATH=/soft/restricted/CNDA/updates/2023.05.15.001/oneapi/compiler/eng-20230512/compiler/linux/lib/:$LD_LIBRARY_PATH
 
-#module load mpich/51.2/icc-all-deterministic-pmix-gpu
 #unset MPIR_CVAR_CH4_OFI_GPU_PIPELINE_D2H_ENGINE_TYPE
 #unset MPIR_CVAR_CH4_OFI_GPU_PIPELINE_H2D_ENGINE_TYPE
 #unset MPIR_CVAR_GPU_USE_IMMEDIATE_COMMAND_LIST

systems/Aurora/tests/repro1gpu.pbs

@@ -1,7 +1,6 @@
 #!/bin/bash
 
-#PBS -l select=16
+#PBS -l select=16:system=sunspot,place=scatter
-#PBS -q EarlyAppAccess
 #PBS -A LatticeQCD_aesp_CNDA
 #PBS -l walltime=02:00:00
 #PBS -N repro1gpu
@@ -10,9 +9,8 @@
 #export OMP_PROC_BIND=spread
 #unset OMP_PLACES
 
-
+module load oneapi/eng-compiler/2023.05.15.003
-#module load oneapi/eng-compiler/2023.05.15.003
+module load mpich/51.2/icc-all-deterministic-pmix-gpu
-#module load mpich/51.2/icc-all-deterministic-pmix-gpu
 
 # 56 cores / 6 threads ~9
 export OMP_NUM_THREADS=6
@@ -36,8 +34,6 @@ export SYCL_PROGRAM_COMPILE_OPTIONS="-ze-opt-large-register-file"
 
 cd $PBS_O_WORKDIR
 
-source ../sourceme.sh
-
 NN=`cat $PBS_NODEFILE | wc -l`
 echo $PBS_NODEFILE
 cat $PBS_NODEFILE