Changed batchedInnerProduct for portability

Spatial Polyakov Loop implementation

BLAS_benchmark/BatchBlasBench.cc

@@ -12,15 +12,13 @@
 #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>
@@ -45,6 +43,90 @@ 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) GRID_ASSERT(err==cudaSuccess);		\
+    }									\
+  }
+
+#endif
+
+
 template<class T> void acceleratorPut(T& dev,T&host)
 {
   acceleratorCopyToDevice(&host,&dev,sizeof(T));
@@ -55,9 +137,6 @@ template<class T> T acceleratorGet(T& dev)
   acceleratorCopyFromDevice(&dev,&host,sizeof(T));
   return host;
 }
-#define accelerator_barrier(dummy) { theAccelerator->wait(); }
-#endif
-
 
 /**************************************************************
  * Allocator
@@ -89,7 +168,7 @@ public:
     if ( (_Tp*)ptr == (_Tp *) NULL ) {
       printf("Grid Device Allocator got NULL for %lu bytes\n",(unsigned long) bytes );
     }
-    assert( ( (_Tp*)ptr != (_Tp *)NULL ) );
+    GRID_ASSERT( ( (_Tp*)ptr != (_Tp *)NULL ) );
     return ptr;
   }
 
@@ -197,11 +276,11 @@ public:
   {
 #ifdef GRID_HIP
     auto err = hipDeviceSynchronize();
-    assert(err==hipSuccess);
+    GRID_ASSERT(err==hipSuccess);
 #endif
 #ifdef GRID_CUDA
     auto err = cudaDeviceSynchronize();
-    assert(err==cudaSuccess);
+    GRID_ASSERT(err==cudaSuccess);
 #endif
 #ifdef GRID_SYCL
     accelerator_barrier();
@@ -211,6 +290,269 @@ 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();
+
+    GRID_ASSERT(OpA!=GridBLAS_OP_T); // Complex case expect no transpose
+    GRID_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);
+    GRID_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);
+    GRID_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();
+
+    GRID_ASSERT(OpA!=GridBLAS_OP_T); // Complex case expect no transpose
+    GRID_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);
+    GRID_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);
+    GRID_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
@@ -241,36 +583,6 @@ 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,
@@ -283,11 +595,11 @@ public:
   {
     RealD t2=usecond();
     int32_t batchCount = Amk.size();
-    assert(Bkn.size()==batchCount);
+    GRID_ASSERT(Bkn.size()==batchCount);
-    assert(Cmn.size()==batchCount);
+    GRID_ASSERT(Cmn.size()==batchCount);
 
-    assert(OpA!=GridBLAS_OP_T); // Complex case expect no transpose
+    GRID_ASSERT(OpA!=GridBLAS_OP_T); // Complex case expect no transpose
-    assert(OpB!=GridBLAS_OP_T);
+    GRID_ASSERT(OpB!=GridBLAS_OP_T);
 
     int lda = m; // m x k column major
     int ldb = k; // k x n column major
@@ -324,7 +636,7 @@ public:
 				   (hipblasDoubleComplex **)&Cmn[0], ldc,
 				   batchCount);
     //	 std::cout << " hipblas return code " <<(int)err<<std::endl;
-    assert(err==HIPBLAS_STATUS_SUCCESS);
+    GRID_ASSERT(err==HIPBLAS_STATUS_SUCCESS);
 #endif
 #ifdef GRID_CUDA
     cublasOperation_t hOpA;
@@ -345,7 +657,7 @@ public:
 				  (cuDoubleComplex *) &beta_p[0],
 				  (cuDoubleComplex **)&Cmn[0], ldc,
 				  batchCount);
-    assert(err==CUBLAS_STATUS_SUCCESS);
+    GRID_ASSERT(err==CUBLAS_STATUS_SUCCESS);
 #endif
 #ifdef GRID_SYCL
       int64_t m64=m;
@@ -492,8 +804,8 @@ public:
     RealD t2=usecond();
     int32_t batchCount = Amk.size();
 
-    assert(OpA!=GridBLAS_OP_T); // Complex case expect no transpose
+    GRID_ASSERT(OpA!=GridBLAS_OP_T); // Complex case expect no transpose
-    assert(OpB!=GridBLAS_OP_T);
+    GRID_ASSERT(OpB!=GridBLAS_OP_T);
 
     int lda = m; // m x k column major
     int ldb = k; // k x n column major
@@ -509,8 +821,8 @@ public:
     acceleratorCopyToDevice((void *)&beta ,(void *)&beta_p[0],sizeof(ComplexF));
     RealD t0=usecond();
 
-    assert(Bkn.size()==batchCount);
+    GRID_ASSERT(Bkn.size()==batchCount);
-    assert(Cmn.size()==batchCount);
+    GRID_ASSERT(Cmn.size()==batchCount);
 #ifdef GRID_HIP
     hipblasOperation_t hOpA;
     hipblasOperation_t hOpB;
@@ -531,7 +843,7 @@ public:
 				   (hipblasComplex **)&Cmn[0], ldc,
 				   batchCount);
 
-    assert(err==HIPBLAS_STATUS_SUCCESS);
+    GRID_ASSERT(err==HIPBLAS_STATUS_SUCCESS);
 #endif
 #ifdef GRID_CUDA
     cublasOperation_t hOpA;
@@ -552,7 +864,7 @@ public:
 				  (cuComplex *) &beta_p[0],
 				  (cuComplex **)&Cmn[0], ldc,
 				  batchCount);
-    assert(err==CUBLAS_STATUS_SUCCESS);
+    GRID_ASSERT(err==CUBLAS_STATUS_SUCCESS);
 #endif
 #ifdef GRID_SYCL
       int64_t m64=m;
@@ -624,301 +936,6 @@ 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)
   {
@@ -967,6 +984,47 @@ 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
+  }
+
 };
 
 
@@ -1035,6 +1093,21 @@ 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
+mpicxx -qmkl=parallel -fsycl BatchBlasBench.cc -o BatchBlasBench -DGRID_SYCL

BLAS_benchmark/compile-command-frontier

@@ -0,0 +1,5 @@
+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

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

Grid/DisableWarnings.h

@@ -51,11 +51,13 @@ directory
 #pragma nv_diag_suppress cast_to_qualified_type
  //disables nvcc specific warning in many files
 #pragma nv_diag_suppress esa_on_defaulted_function_ignored
+#pragma nv_diag_suppress declared_but_not_referenced
 #pragma nv_diag_suppress extra_semicolon
 #else
  //disables nvcc specific warning in json.hpp
 #pragma diag_suppress unsigned_compare_with_zero
 #pragma diag_suppress cast_to_qualified_type
+#pragma diag_suppress declared_but_not_referenced
  //disables nvcc specific warning in many files
 #pragma diag_suppress esa_on_defaulted_function_ignored
 #pragma diag_suppress extra_semicolon

Grid/GridStd.h

@@ -1,9 +1,17 @@
 #ifndef GRID_STD_H
 #define GRID_STD_H
 
+///////////////////
+// Grid config
+///////////////////
+#include "Config.h"
+
 ///////////////////
 // Std C++ dependencies
 ///////////////////
+#define _NBACKTRACE (256)
+extern void * Grid_backtrace_buffer[_NBACKTRACE];
+
 #include <cassert>
 #include <complex>
 #include <memory>
@@ -15,7 +23,9 @@
 #include <random>
 #include <functional>
 #include <stdio.h>
+#include <string.h>
 #include <stdlib.h>
+#include <unistd.h>
 #include <strings.h>
 #include <stdio.h>
 #include <signal.h>
@@ -23,11 +33,36 @@
 #include <sys/time.h>
 #include <chrono>
 #include <zlib.h>
+#ifdef HAVE_EXECINFO_H
+#include <execinfo.h>
+#endif
+
+void GridAbort(void);
+
+#define ASSLOG(A) ::write(STDERR_FILENO,A,::strlen(A));
+#ifdef HAVE_EXECINFO_H
+#define GRID_ASSERT(b) if(!(b)) {					\
+    fflush(stdout); \
+    ASSLOG(" GRID_ASSERT failure: ");					\
+    ASSLOG(__FILE__);							\
+    ASSLOG(" : ");							\
+    ASSLOG(#b);								\
+    ASSLOG(" : ");							\
+    int symbols = backtrace(Grid_backtrace_buffer,_NBACKTRACE);		\
+    backtrace_symbols_fd(Grid_backtrace_buffer,symbols,STDERR_FILENO);	\
+    GridAbort();							\
+  };
+#else
+#define GRID_ASSERT(b) if(!(b)) {					\
+    ASSLOG(" GRID_ASSERT failure: ");					\
+    ASSLOG(__FILE__);							\
+    ASSLOG(" : ");							\
+    ASSLOG(#b);								\
+    ASSLOG(" : ");							\
+    GridAbort();							\
+  };
+#endif
 
-///////////////////
-// Grid config
-///////////////////
-#include "Config.h"
 
 #ifdef TOFU
 #undef GRID_COMMS_THREADS

Grid/Makefile.am

@@ -54,6 +54,7 @@ Version.h: version-cache
 include Make.inc
 include Eigen.inc
 
+if BUILD_FERMION_INSTANTIATIONS
   extra_sources+=$(WILS_FERMION_FILES)
   extra_sources+=$(STAG_FERMION_FILES)
 if BUILD_ZMOBIUS
@@ -68,8 +69,11 @@ if BUILD_FERMION_REPS
 endif
 if BUILD_SP
     extra_sources+=$(SP_FERMION_FILES)
+if BUILD_FERMION_REPS
       extra_sources+=$(SP_TWOIND_FERMION_FILES)
 endif
+endif
+endif
 
 lib_LIBRARIES = libGrid.a
 

Grid/Namespace.h

@@ -29,8 +29,8 @@ directory
 #pragma once
 
 #include <type_traits>
-#include <cassert>
 #include <exception>
+#include <cassert>
 
 #define NAMESPACE_BEGIN(A) namespace A {
 #define NAMESPACE_END(A)   }

Grid/algorithms/Algorithms.h

@@ -50,6 +50,9 @@ 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>
+// Not really deflation, but useful
+#include <Grid/algorithms/blas/MomentumProject.h>
 NAMESPACE_CHECK(deflation);
 #include <Grid/algorithms/iterative/ConjugateGradient.h>
 NAMESPACE_CHECK(ConjGrad);
@@ -72,6 +75,7 @@ NAMESPACE_CHECK(BiCGSTAB);
 #include <Grid/algorithms/iterative/FlexibleCommunicationAvoidingGeneralisedMinimalResidual.h>
 #include <Grid/algorithms/iterative/MixedPrecisionFlexibleGeneralisedMinimalResidual.h>
 #include <Grid/algorithms/iterative/ImplicitlyRestartedLanczos.h>
+#include <Grid/algorithms/iterative/SimpleLanczos.h>
 #include <Grid/algorithms/iterative/PowerMethod.h>
 #include <Grid/algorithms/iterative/AdefGeneric.h>
 #include <Grid/algorithms/iterative/AdefMrhs.h>
@@ -80,4 +84,9 @@ NAMESPACE_CHECK(PowerMethod);
 NAMESPACE_CHECK(multigrid);
 #include <Grid/algorithms/FFT.h>
 
+#include <Grid/algorithms/iterative/KrylovSchur.h>
+#include <Grid/algorithms/iterative/Arnoldi.h>
+#include <Grid/algorithms/iterative/LanczosBidiagonalization.h>
+#include <Grid/algorithms/iterative/RestartedLanczosBidiagonalization.h>
+
 #endif

Grid/algorithms/FFT.h

@@ -28,6 +28,15 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #ifndef _GRID_FFT_H_
 #define _GRID_FFT_H_
 
+#ifdef GRID_CUDA
+#include <cufft.h>
+#endif
+
+#ifdef GRID_HIP
+#include <hipfft/hipfft.h>
+#endif
+
+#if !defined(GRID_CUDA) && !defined(GRID_HIP)
 #ifdef HAVE_FFTW
 #if defined(USE_MKL) || defined(GRID_SYCL)
 #include <fftw/fftw3.h>
@@ -35,88 +44,190 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #include <fftw3.h>
 #endif
 #endif
+#endif
 
 NAMESPACE_BEGIN(Grid);
 
-template<class scalar> struct FFTW { };
+#ifndef FFTW_FORWARD
+#define FFTW_FORWARD (-1)
+#define FFTW_BACKWARD (+1)
+#define FFTW_ESTIMATE (0)
+#endif
 
+template<class scalar> struct FFTW {
+};
+
+#ifdef GRID_HIP
+template<> struct FFTW<ComplexD> {
+public:
+  static const int forward=FFTW_FORWARD;
+  static const int backward=FFTW_BACKWARD;
+  typedef hipfftDoubleComplex FFTW_scalar;
+  typedef hipfftHandle        FFTW_plan;
+  static FFTW_plan fftw_plan_many_dft(int rank, int *n,int howmany,
+				      FFTW_scalar *in, int *inembed,		
+				      int istride, int idist,		
+				      FFTW_scalar *out, int *onembed,		
+				      int ostride, int odist,		
+				      int sign, unsigned flags) {
+    FFTW_plan p;
+    auto rv = hipfftPlanMany(&p,rank,n,n,istride,idist,n,ostride,odist,HIPFFT_Z2Z,howmany);
+    GRID_ASSERT(rv==HIPFFT_SUCCESS);
+    return p;
+  }	  
+    
+  inline static void fftw_execute_dft(const FFTW_plan p,FFTW_scalar *in,FFTW_scalar *out, int sign) {
+    hipfftResult rv;
+    if ( sign == forward ) rv =hipfftExecZ2Z(p,in,out,HIPFFT_FORWARD);
+    else                   rv =hipfftExecZ2Z(p,in,out,HIPFFT_BACKWARD);
+    accelerator_barrier();
+    GRID_ASSERT(rv==HIPFFT_SUCCESS);
+  }
+  inline static void fftw_destroy_plan(const FFTW_plan p) {
+    hipfftDestroy(p);
+  }
+};
+template<> struct FFTW<ComplexF> {
+public:
+  static const int forward=FFTW_FORWARD;
+  static const int backward=FFTW_BACKWARD;
+  typedef hipfftComplex      FFTW_scalar;
+  typedef hipfftHandle        FFTW_plan;
+
+  static FFTW_plan fftw_plan_many_dft(int rank, int *n,int howmany,
+				      FFTW_scalar *in, int *inembed,		
+				      int istride, int idist,		
+				      FFTW_scalar *out, int *onembed,		
+				      int ostride, int odist,		
+				      int sign, unsigned flags) {
+    FFTW_plan p;
+    auto rv = hipfftPlanMany(&p,rank,n,n,istride,idist,n,ostride,odist,HIPFFT_C2C,howmany);
+    GRID_ASSERT(rv==HIPFFT_SUCCESS);
+    return p;
+  }	  
+    
+  inline static void fftw_execute_dft(const FFTW_plan p,FFTW_scalar *in,FFTW_scalar *out, int sign) {
+    hipfftResult rv;
+    if ( sign == forward ) rv =hipfftExecC2C(p,in,out,HIPFFT_FORWARD);
+    else                   rv =hipfftExecC2C(p,in,out,HIPFFT_BACKWARD);
+    accelerator_barrier();
+    GRID_ASSERT(rv==HIPFFT_SUCCESS);
+  }
+  inline static void fftw_destroy_plan(const FFTW_plan p) {
+    hipfftDestroy(p);
+  }
+};
+#endif
+
+#ifdef GRID_CUDA
+template<> struct FFTW<ComplexD> {
+public:
+  static const int forward=FFTW_FORWARD;
+  static const int backward=FFTW_BACKWARD;
+  typedef cufftDoubleComplex FFTW_scalar;
+  typedef cufftHandle        FFTW_plan;
+
+  static FFTW_plan fftw_plan_many_dft(int rank, int *n,int howmany,
+				      FFTW_scalar *in, int *inembed,		
+				      int istride, int idist,		
+				      FFTW_scalar *out, int *onembed,		
+				      int ostride, int odist,		
+				      int sign, unsigned flags) {
+    FFTW_plan p;
+    cufftPlanMany(&p,rank,n,n,istride,idist,n,ostride,odist,CUFFT_Z2Z,howmany);
+    return p;
+  }	  
+    
+  inline static void fftw_execute_dft(const FFTW_plan p,FFTW_scalar *in,FFTW_scalar *out, int sign) {
+    if ( sign == forward ) cufftExecZ2Z(p,in,out,CUFFT_FORWARD);
+    else                   cufftExecZ2Z(p,in,out,CUFFT_INVERSE);
+    accelerator_barrier();
+  }
+  inline static void fftw_destroy_plan(const FFTW_plan p) {
+    cufftDestroy(p);
+  }
+};
+template<> struct FFTW<ComplexF> {
+public:
+  static const int forward=FFTW_FORWARD;
+  static const int backward=FFTW_BACKWARD;
+  typedef cufftComplex FFTW_scalar;
+  typedef cufftHandle        FFTW_plan;
+
+  static FFTW_plan fftw_plan_many_dft(int rank, int *n,int howmany,
+				      FFTW_scalar *in, int *inembed,		
+				      int istride, int idist,		
+				      FFTW_scalar *out, int *onembed,		
+				      int ostride, int odist,		
+				      int sign, unsigned flags) {
+    FFTW_plan p;
+    cufftPlanMany(&p,rank,n,n,istride,idist,n,ostride,odist,CUFFT_C2C,howmany);
+    return p;
+  }	  
+    
+  inline static void fftw_execute_dft(const FFTW_plan p,FFTW_scalar *in,FFTW_scalar *out, int sign) {
+    if ( sign == forward ) cufftExecC2C(p,in,out,CUFFT_FORWARD);
+    else                   cufftExecC2C(p,in,out,CUFFT_INVERSE);
+    accelerator_barrier();
+  }
+  inline static void fftw_destroy_plan(const FFTW_plan p) {
+    cufftDestroy(p);
+  }
+};
+#endif
+
+#if !defined(GRID_CUDA) && !defined(GRID_HIP)
 #ifdef HAVE_FFTW
 template<> struct FFTW<ComplexD> {
 public:
-
   typedef fftw_complex FFTW_scalar;
   typedef fftw_plan    FFTW_plan;
-
+  static FFTW_plan fftw_plan_many_dft(int rank, int *n,int howmany,
-  static FFTW_plan fftw_plan_many_dft(int rank, const int *n,int howmany,
+				      FFTW_scalar *in, int *inembed,		
-				      FFTW_scalar *in, const int *inembed,		
 				      int istride, int idist,		
-				      FFTW_scalar *out, const int *onembed,		
+				      FFTW_scalar *out, int *onembed,		
 				      int ostride, int odist,		
 				      int sign, unsigned flags) {
     return ::fftw_plan_many_dft(rank,n,howmany,in,inembed,istride,idist,out,onembed,ostride,odist,sign,flags);
   }	  
     
-  static void fftw_flops(const FFTW_plan p,double *add, double *mul, double *fmas){
+  inline static void fftw_execute_dft(const FFTW_plan p,FFTW_scalar *in,FFTW_scalar *out, int sign) {
-    ::fftw_flops(p,add,mul,fmas);
-  }
-
-  inline static void fftw_execute_dft(const FFTW_plan p,FFTW_scalar *in,FFTW_scalar *out) {
     ::fftw_execute_dft(p,in,out);
   }
   inline static void fftw_destroy_plan(const FFTW_plan p) {
     ::fftw_destroy_plan(p);
   }
 };
-
 template<> struct FFTW<ComplexF> {
 public:
-
   typedef fftwf_complex FFTW_scalar;
   typedef fftwf_plan    FFTW_plan;
-
+  static FFTW_plan fftw_plan_many_dft(int rank, int *n,int howmany,
-  static FFTW_plan fftw_plan_many_dft(int rank, const int *n,int howmany,
+				      FFTW_scalar *in, int *inembed,		
-				      FFTW_scalar *in, const int *inembed,		
 				      int istride, int idist,		
-				      FFTW_scalar *out, const int *onembed,		
+				      FFTW_scalar *out, int *onembed,		
 				      int ostride, int odist,		
 				      int sign, unsigned flags) {
     return ::fftwf_plan_many_dft(rank,n,howmany,in,inembed,istride,idist,out,onembed,ostride,odist,sign,flags);
   }	  
     
-  static void fftw_flops(const FFTW_plan p,double *add, double *mul, double *fmas){
+  inline static void fftw_execute_dft(const FFTW_plan p,FFTW_scalar *in,FFTW_scalar *out, int sign) {
-    ::fftwf_flops(p,add,mul,fmas);
-  }
-
-  inline static void fftw_execute_dft(const FFTW_plan p,FFTW_scalar *in,FFTW_scalar *out) {
     ::fftwf_execute_dft(p,in,out);
   }
   inline static void fftw_destroy_plan(const FFTW_plan p) {
     ::fftwf_destroy_plan(p);
   }
 };
-
 #endif
-
-#ifndef FFTW_FORWARD
-#define FFTW_FORWARD (-1)
-#define FFTW_BACKWARD (+1)
 #endif
 
 class FFT {
 private:
     
-  GridCartesian *vgrid;
-  GridCartesian *sgrid;
-    
-  int Nd;
   double flops;
   double flops_call;
   uint64_t usec;
     
-  Coordinate dimensions;
-  Coordinate processors;
-  Coordinate processor_coor;
-    
 public:
     
   static const int forward=FFTW_FORWARD;
@@ -126,31 +237,25 @@ public:
   double MFlops(void) {return flops/usec;}
   double USec(void)   {return (double)usec;}    
 
-  FFT ( GridCartesian * grid ) :
+  FFT ( GridCartesian * grid ) 
-    vgrid(grid),
-    Nd(grid->_ndimension),
-    dimensions(grid->_fdimensions),
-    processors(grid->_processors),
-    processor_coor(grid->_processor_coor)
   {
     flops=0;
     usec =0;
-    Coordinate layout(Nd,1);
-    sgrid = new GridCartesian(dimensions,layout,processors,*grid);
   };
     
   ~FFT ( void)  {
-    delete sgrid;
+    //    delete sgrid;
   }
     
   template<class vobj>
   void FFT_dim_mask(Lattice<vobj> &result,const Lattice<vobj> &source,Coordinate mask,int sign){
 
-    conformable(result.Grid(),vgrid);
+    //    vgrid=result.Grid();
-    conformable(source.Grid(),vgrid);
+    //    conformable(result.Grid(),vgrid);
-    Lattice<vobj> tmp(vgrid);
+    //    conformable(source.Grid(),vgrid);
-    tmp = source;
+    const int Ndim = source.Grid()->Nd();
-    for(int d=0;d<Nd;d++){
+    Lattice<vobj> tmp = source;
+    for(int d=0;d<Ndim;d++){
       if( mask[d] ) {
 	FFT_dim(result,tmp,d,sign);
 	tmp=result;
@@ -160,59 +265,70 @@ public:
 
   template<class vobj>
   void FFT_all_dim(Lattice<vobj> &result,const Lattice<vobj> &source,int sign){
-    Coordinate mask(Nd,1);
+    const int Ndim = source.Grid()->Nd();
+    Coordinate mask(Ndim,1);
     FFT_dim_mask(result,source,mask,sign);
   }
 
 
   template<class vobj>
   void FFT_dim(Lattice<vobj> &result,const Lattice<vobj> &source,int dim, int sign){
-#ifndef HAVE_FFTW
+    const int Ndim = source.Grid()->Nd();
-    assert(0);
+    GridBase *grid = source.Grid();
-#else
+    conformable(result.Grid(),source.Grid());
-    conformable(result.Grid(),vgrid);
-    conformable(source.Grid(),vgrid);
 
-    int L = vgrid->_ldimensions[dim];
+    int L = grid->_ldimensions[dim];
-    int G = vgrid->_fdimensions[dim];
+    int G = grid->_fdimensions[dim];
       
-    Coordinate layout(Nd,1);
+    Coordinate layout(Ndim,1);
-    Coordinate pencil_gd(vgrid->_fdimensions);
-      
-    pencil_gd[dim] = G*processors[dim];
-      
-    // Pencil global vol LxLxGxLxL per node
-    GridCartesian pencil_g(pencil_gd,layout,processors,*vgrid);
     
     // Construct pencils
     typedef typename vobj::scalar_object sobj;
-    typedef typename sobj::scalar_type   scalar;
+    typedef typename vobj::scalar_type   scalar;
+    typedef typename vobj::scalar_type   scalar_type;
+    typedef typename vobj::vector_type   vector_type;
       
-    Lattice<sobj> pgbuf(&pencil_g);
+    //std::cout << "CPU view" << std::endl;
-    autoView(pgbuf_v , pgbuf, CpuWrite);
     
     typedef typename FFTW<scalar>::FFTW_scalar FFTW_scalar;
     typedef typename FFTW<scalar>::FFTW_plan   FFTW_plan;
       
     int Ncomp = sizeof(sobj)/sizeof(scalar);
-    int Nlow  = 1;
+    int64_t Nlow  = 1;
+    int64_t Nhigh = 1;
+
     for(int d=0;d<dim;d++){
-      Nlow*=vgrid->_ldimensions[d];
+      Nlow*=grid->_ldimensions[d];
     }
+    for(int d=dim+1;d<Ndim;d++){
+      Nhigh*=grid->_ldimensions[d];
+    }
+    int64_t Nperp=Nlow*Nhigh;
+    
+    deviceVector<scalar> pgbuf; // Layout is [perp][component][dim]
+    pgbuf.resize(Nperp*Ncomp*G);
+    scalar *pgbuf_v = &pgbuf[0];
       
     int rank = 1;  /* 1d transforms */
     int n[] = {G}; /* 1d transforms of length G */
-    int howmany = Ncomp;
+    int howmany = Ncomp * Nperp;
     int odist,idist,istride,ostride;
-    idist   = odist   = 1;          /* Distance between consecutive FT's */
+    idist   = odist   = G;            /* Distance between consecutive FT's */
-    istride = ostride = Ncomp*Nlow; /* distance between two elements in the same FT */
+    istride = ostride = 1;            /* Distance between two elements in the same FT */
     int *inembed = n, *onembed = n;
       
     scalar div;
     if ( sign == backward ) div = 1.0/G;
     else if ( sign == forward ) div = 1.0;
-    else assert(0);
+    else GRID_ASSERT(0);
 
+    double t_pencil=0;
+    double t_fft   =0;
+    double t_total =-usecond();
+    //    std::cout << GridLogPerformance<<"Making FFTW plan" << std::endl;
+    /*
+     *
+     */
     FFTW_plan p;
     {
       FFTW_scalar *in = (FFTW_scalar *)&pgbuf_v[0];
@@ -226,68 +342,154 @@ public:
     }
       
     // Barrel shift and collect global pencil
-    Coordinate lcoor(Nd), gcoor(Nd);
+    //    std::cout << GridLogPerformance<<"Making pencil" << std::endl;
+    Coordinate lcoor(Ndim), gcoor(Ndim);
+    double t_copy=0;
+    double t_shift=0;
+    t_pencil = -usecond();
     result = source;
-    int pc = processor_coor[dim];
+    int pc = grid->_processor_coor[dim];
-    for(int p=0;p<processors[dim];p++) {
+
-      {
+    const Coordinate ldims = grid->_ldimensions;
-	autoView(r_v,result,CpuRead);
+    const Coordinate rdims = grid->_rdimensions;
-	autoView(p_v,pgbuf,CpuWrite);
+    const Coordinate sdims = grid->_simd_layout;
-	thread_for(idx, sgrid->lSites(),{
+
-          Coordinate cbuf(Nd);
+    Coordinate processors = grid->_processors;
-          sobj s;
+    Coordinate pgdims(Ndim);
-	  sgrid->LocalIndexToLocalCoor(idx,cbuf);
+    pgdims[0] = G;
-	  peekLocalSite(s,r_v,cbuf);
+    for(int d=0, dd=1;d<Ndim;d++){
-	  cbuf[dim]+=((pc+p) % processors[dim])*L;
+      if ( d!=dim ) pgdims[dd++] = ldims[d];
-	  pokeLocalSite(s,p_v,cbuf);
-        });
     }
-      if (p != processors[dim] - 1) {
+    int64_t pgvol=1;
-	result = Cshift(result,dim,L);
+    for(int d=0;d<Ndim;d++) pgvol*=pgdims[d];
+    
+    const int Nsimd = vobj::Nsimd();
+    for(int p=0;p<processors[dim];p++) {
+      t_copy-=usecond();
+      autoView(r_v,result,AcceleratorRead);
+      accelerator_for(idx, grid->oSites(), vobj::Nsimd(), {
+#ifdef GRID_SIMT
+      {
+	int lane=acceleratorSIMTlane(Nsimd); // buffer lane
+#else
+      for(int lane=0;lane<Nsimd;lane++) {
+#endif
+	Coordinate icoor;
+	Coordinate ocoor;
+	Coordinate pgcoor;
+
+	Lexicographic::CoorFromIndex(icoor,lane,sdims);
+	Lexicographic::CoorFromIndex(ocoor,idx,rdims);
+
+	pgcoor[0] = ocoor[dim] + icoor[dim]*rdims[dim] + ((pc+p)%processors[dim])*L;
+	for(int d=0,dd=1;d<Ndim;d++){
+	  if ( d!=dim ) {
+	    pgcoor[dd] = ocoor[d] + icoor[d]*rdims[d];
+	    dd++;
 	  }
 	}
 
-    // Loop over orthog coords
+	// Map coordinates in lattice layout to FFTW index
-    int NN=pencil_g.lSites();
+	int64_t pgidx;
-    GridStopWatch timer;
+	Lexicographic::IndexFromCoor(pgcoor,pgidx,pgdims);
-    timer.Start();
+
-    thread_for( idx,NN,{
+	vector_type *from = (vector_type *)&r_v[idx];
-        Coordinate cbuf(Nd);
+	scalar_type stmp;
-	pencil_g.LocalIndexToLocalCoor(idx, cbuf);
+	for(int w=0;w<Ncomp;w++){
-	if ( cbuf[dim] == 0 ) {  // restricts loop to plane at lcoor[dim]==0
+	  int64_t pg_idx = pgidx + w*pgvol;
-	  FFTW_scalar *in = (FFTW_scalar *)&pgbuf_v[idx];
+	  stmp = getlane(from[w], lane);
-	  FFTW_scalar *out= (FFTW_scalar *)&pgbuf_v[idx];
+	  pgbuf_v[pg_idx] = stmp;
-	  FFTW<scalar>::fftw_execute_dft(p,in,out);
 	}
+#ifdef GRID_SIMT
+      }
+#else
+      }
+#endif
       });
-    timer.Stop();
+
+      t_copy+=usecond();
+      if (p != processors[dim] - 1) {
+	Lattice<vobj> temp(grid);
+	t_shift-=usecond();
+	temp = Cshift(result,dim,L); result = temp;
+	t_shift+=usecond();
+      }
+    }
+    t_pencil += usecond();
+      
+    FFTW_scalar *in = (FFTW_scalar *)pgbuf_v;
+    FFTW_scalar *out= (FFTW_scalar *)pgbuf_v;
+    t_fft = -usecond();
+    FFTW<scalar>::fftw_execute_dft(p,in,out,sign);
+    t_fft += usecond();
     
     // performance counting
-    double add,mul,fma;
+    flops_call = 5.0*howmany*G*log2(G);
-    FFTW<scalar>::fftw_flops(p,&add,&mul,&fma);
+    usec = t_fft;
-    flops_call = add+mul+2.0*fma;
+    flops= flops_call;
-    usec += timer.useconds();
-    flops+= flops_call*NN;
 
-    // writing out result
+    result = Zero();
+    
+    double t_insert = -usecond();
     {
-      autoView(pgbuf_v,pgbuf,CpuRead);
+      autoView(r_v,result,AcceleratorWrite);
-      autoView(result_v,result,CpuWrite);
+      accelerator_for(idx,grid->oSites(),Nsimd,{
-      thread_for(idx,sgrid->lSites(),{
+#ifdef GRID_SIMT
-	Coordinate clbuf(Nd), cgbuf(Nd);
+      {
-	sobj s;
+	int lane=acceleratorSIMTlane(Nsimd); // buffer lane
-	sgrid->LocalIndexToLocalCoor(idx,clbuf);
+#else
-	cgbuf = clbuf;
+      for(int lane=0;lane<Nsimd;lane++) {
-	cgbuf[dim] = clbuf[dim]+L*pc;
+#endif
-	peekLocalSite(s,pgbuf_v,cgbuf);
+	Coordinate icoor(Ndim);
-	pokeLocalSite(s,result_v,clbuf);
+	Coordinate ocoor(Ndim);
+	Coordinate pgcoor(Ndim);
+
+	Lexicographic::CoorFromIndex(icoor,lane,sdims);
+	Lexicographic::CoorFromIndex(ocoor,idx,rdims);
+
+	pgcoor[0] = ocoor[dim] + icoor[dim]*rdims[dim] + pc*L;
+	for(int d=0,dd=1;d<Ndim;d++){
+	  if ( d!=dim ) {
+	    pgcoor[dd] = ocoor[d] + icoor[d]*rdims[d];
+	    dd++;
+	  }
+	}
+	// Map coordinates in lattice layout to FFTW index
+	int64_t pgidx;
+	Lexicographic::IndexFromCoor(pgcoor,pgidx,pgdims);
+
+	vector_type *to = (vector_type *)&r_v[idx];
+	scalar_type stmp;
+	for(int w=0;w<Ncomp;w++){
+	  int64_t pg_idx = pgidx + w*pgvol;
+	  stmp = pgbuf_v[pg_idx];
+	  putlane(to[w], stmp, lane);
+	}
+	
+#ifdef GRID_SIMT
+      }
+#else
+      }
+#endif
       });
     }
+
     result = result*div;
 
+    t_insert +=usecond();
+    
     // destroying plan
     FFTW<scalar>::fftw_destroy_plan(p);
-#endif
+
+    t_total +=usecond();
+
+    std::cout <<GridLogPerformance<< " FFT took   "<<t_total/1.0e6 <<" s" << std::endl;
+    std::cout <<GridLogPerformance<< " FFT pencil "<<t_pencil/1.0e6 <<" s" << std::endl;
+    std::cout <<GridLogPerformance<< "  of which copy "<<t_copy/1.0e6 <<" s" << std::endl;
+    std::cout <<GridLogPerformance<< "  of which shift"<<t_shift/1.0e6 <<" s" << std::endl;
+    std::cout <<GridLogPerformance<< " FFT kernels "<<t_fft/1.0e6 <<" s" << std::endl;
+    std::cout <<GridLogPerformance<< " FFT insert  "<<t_insert/1.0e6 <<" s" << std::endl;
+    
   }
 };
 

Grid/algorithms/LinearOperator.h

@@ -64,7 +64,7 @@ public:
 //
 // I'm not entirely happy with implementation; to share the Schur code between herm and non-herm
 // while still having a "OpAndNorm" in the abstract base I had to implement it in both cases
-// with an assert trap in the non-herm. This isn't right; there must be a better C++ way to
+// with an GRID_ASSERT trap in the non-herm. This isn't right; there must be a better C++ way to
 // do it, but I fear it required multiple inheritance and mixed in abstract base classes
 /////////////////////////////////////////////////////////////////////////////////////////////
 
@@ -103,6 +103,38 @@ 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
@@ -116,22 +148,22 @@ public:
   // Support for coarsening to a multigrid
   void OpDiag (const Field &in, Field &out) {
     _Mat.Mdiag(in,out);
-    assert(0);
+    GRID_ASSERT(0);
   }
   void OpDir  (const Field &in, Field &out,int dir,int disp) {
     _Mat.Mdir(in,out,dir,disp);
-    assert(0);
+    GRID_ASSERT(0);
   }
   void OpDirAll  (const Field &in, std::vector<Field> &out){
-    assert(0);
+    GRID_ASSERT(0);
   };
   void Op     (const Field &in, Field &out){
     _Mat.M(in,out);
-    assert(0);
+    GRID_ASSERT(0);
   }
   void AdjOp     (const Field &in, Field &out){
     _Mat.Mdag(in,out);
-    assert(0);
+    GRID_ASSERT(0);
   }
   void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){
     HermOp(in,out);
@@ -156,13 +188,13 @@ public:
   ShiftedHermOpLinearOperator(LinearOperatorBase<Field> &Mat,RealD shift): _Mat(Mat), _shift(shift){};
   // Support for coarsening to a multigrid
   void OpDiag (const Field &in, Field &out) {
-    assert(0);
+    GRID_ASSERT(0);
   }
   void OpDir  (const Field &in, Field &out,int dir,int disp) {
-    assert(0);
+    GRID_ASSERT(0);
   }
   void OpDirAll  (const Field &in, std::vector<Field> &out){
-    assert(0);
+    GRID_ASSERT(0);
   };
   void Op     (const Field &in, Field &out){
     HermOp(in,out);
@@ -239,10 +271,42 @@ public:
     _Mat.Mdag(in,out);
   }
   void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){
-    assert(0);
+    GRID_ASSERT(0);
   }
   void HermOp(const Field &in, Field &out){
-    assert(0);
+    GRID_ASSERT(0);
+  }
+};
+template<class Matrix,class Field>
+class ShiftedNonHermitianLinearOperator : public LinearOperatorBase<Field> {
+  Matrix &_Mat;
+  RealD shift;
+public:
+  ShiftedNonHermitianLinearOperator(Matrix &Mat,RealD shft): _Mat(Mat),shift(shft){};
+  // Support for coarsening to a multigrid
+  void OpDiag (const Field &in, Field &out) {
+    _Mat.Mdiag(in,out);
+    out = out + shift*in;
+  }
+  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);
+    out = out + shift * in;
+  }
+  void AdjOp     (const Field &in, Field &out){
+    _Mat.Mdag(in,out);
+    out = out + shift * in;
+  }
+  void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){
+    GRID_ASSERT(0);
+  }
+  void HermOp(const Field &in, Field &out){
+    GRID_ASSERT(0);
   }
 };
 
@@ -281,13 +345,13 @@ class SchurOperatorBase :  public LinearOperatorBase<Field> {
   }
   // Support for coarsening to a multigrid
   void OpDiag (const Field &in, Field &out) {
-    assert(0); // must coarsen the unpreconditioned system
+    GRID_ASSERT(0); // must coarsen the unpreconditioned system
   }
   void OpDir  (const Field &in, Field &out,int dir,int disp) {
-    assert(0);
+    GRID_ASSERT(0);
   }
   void OpDirAll  (const Field &in, std::vector<Field> &out){
-    assert(0);
+    GRID_ASSERT(0);
   };
 };
 template<class Matrix,class Field>
@@ -383,10 +447,10 @@ class NonHermitianSchurOperatorBase :  public LinearOperatorBase<Field>
     MpcDag(tmp,out);
   }
   virtual void HermOpAndNorm(const Field& in, Field& out, RealD& n1, RealD& n2) {
-    assert(0);
+    GRID_ASSERT(0);
   }
   virtual void HermOp(const Field& in, Field& out) {
-    assert(0);
+    GRID_ASSERT(0);
   }
   void Op(const Field& in, Field& out) {
     Mpc(in, out);
@@ -396,13 +460,13 @@ class NonHermitianSchurOperatorBase :  public LinearOperatorBase<Field>
   }
   // Support for coarsening to a multigrid
   void OpDiag(const Field& in, Field& out) {
-    assert(0); // must coarsen the unpreconditioned system
+    GRID_ASSERT(0); // must coarsen the unpreconditioned system
   }
   void OpDir(const Field& in, Field& out, int dir, int disp) {
-    assert(0);
+    GRID_ASSERT(0);
   }
   void OpDirAll(const Field& in, std::vector<Field>& out){
-    assert(0);
+    GRID_ASSERT(0);
   };
 };
 
@@ -516,7 +580,7 @@ class SchurStaggeredOperator :  public SchurOperatorBase<Field> {
  public:
   SchurStaggeredOperator (Matrix &Mat): _Mat(Mat), tmp(_Mat.RedBlackGrid()) 
   { 
-    assert( _Mat.isTrivialEE() );
+    GRID_ASSERT( _Mat.isTrivialEE() );
     mass = _Mat.Mass();
   }
   virtual void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){
@@ -547,7 +611,7 @@ class SchurStaggeredOperator :  public SchurOperatorBase<Field> {
     Mpc(in,out);
   }
   virtual void MpcDagMpc(const Field &in, Field &out) {
-    assert(0);// Never need with staggered
+    GRID_ASSERT(0);// Never need with staggered
   }
 };
 template<class Matrix,class Field> using SchurStagOperator = SchurStaggeredOperator<Matrix,Field>;
@@ -559,7 +623,7 @@ template<class Field> class OperatorFunction {
 public:
   virtual void operator() (LinearOperatorBase<Field> &Linop, const Field &in, Field &out) = 0;
   virtual void operator() (LinearOperatorBase<Field> &Linop, const std::vector<Field> &in,std::vector<Field> &out) {
-    assert(in.size()==out.size());
+    GRID_ASSERT(in.size()==out.size());
     for(int k=0;k<in.size();k++){
       (*this)(Linop,in[k],out[k]);
     }
@@ -573,7 +637,7 @@ public:
 
   virtual void operator() (const std::vector<Field> &in, std::vector<Field> &out)
   {
-    assert(in.size() == out.size());
+    GRID_ASSERT(in.size() == out.size());
 
     for (unsigned int i = 0; i < in.size(); ++i)
     {

Grid/algorithms/SparseMatrix.h

@@ -45,6 +45,11 @@ 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*=1.1;
+      delta*=1.02;
       RealD f = approx(x);
       out<< x<<" "<<f<<std::endl;
     }
@@ -131,6 +131,26 @@ 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){
@@ -249,7 +269,9 @@ public:
     RealD xscale = 2.0/(hi-lo);
     RealD mscale = -(hi+lo)/(hi-lo);
     Linop.HermOp(T0,y);
+    grid->Barrier();
     axpby(T1,xscale,mscale,y,in);
+    grid->Barrier();
 
     // sum = .5 c[0] T0 + c[1] T1
     //    out = ()*T0 + Coeffs[1]*T1;

Grid/algorithms/approx/Remez.cc

@@ -121,7 +121,7 @@ double AlgRemez::generateApprox(int num_degree, int den_degree,
   // Reallocate arrays, since degree has changed
   if (num_degree != n || den_degree != d) allocate(num_degree,den_degree);
 
-  assert(a_len<=SUM_MAX);
+  GRID_ASSERT(a_len<=SUM_MAX);
 
   step = new bigfloat[num_degree+den_degree+2];
 
@@ -151,9 +151,9 @@ double AlgRemez::generateApprox(int num_degree, int den_degree,
     equations();
     if (delta < tolerance) {
       std::cout<<"Delta too small, try increasing precision\n";
-      assert(0);
+      GRID_ASSERT(0);
     };    
-    assert( delta>= tolerance);
+    GRID_ASSERT( delta>= tolerance);
 
     search(step);
   }

Grid/algorithms/approx/Remez.h

@@ -134,7 +134,7 @@ class AlgRemez
   virtual ~AlgRemez();
 
   int getDegree(void){ 
-    assert(n==d);
+    GRID_ASSERT(n==d);
     return n;
   }
   // Reset the bounds of the approximation

Grid/algorithms/approx/RemezGeneral.cc

@@ -28,11 +28,11 @@ void AlgRemezGeneral::setupPolyProperties(int num_degree, int den_degree, PolyTy
   pow_n = num_degree;
   pow_d = den_degree;
 
-  if(pow_n % 2 == 0 && num_type_in == PolyType::Odd) assert(0);
+  if(pow_n % 2 == 0 && num_type_in == PolyType::Odd) GRID_ASSERT(0);
-  if(pow_n % 2 == 1 && num_type_in == PolyType::Even) assert(0);
+  if(pow_n % 2 == 1 && num_type_in == PolyType::Even) GRID_ASSERT(0);
 
-  if(pow_d % 2 == 0 && den_type_in == PolyType::Odd) assert(0);
+  if(pow_d % 2 == 0 && den_type_in == PolyType::Odd) GRID_ASSERT(0);
-  if(pow_d % 2 == 1 && den_type_in == PolyType::Even) assert(0);
+  if(pow_d % 2 == 1 && den_type_in == PolyType::Even) GRID_ASSERT(0);
 
   num_type = num_type_in;
   den_type = den_type_in;
@@ -112,9 +112,9 @@ double AlgRemezGeneral::generateApprox(const int num_degree, const int den_degre
     equations();
     if (delta < tolerance) {
       std::cout<<"Iteration " << iter-1 << " delta too small (" << delta << "<" << tolerance << "), try increasing precision\n";
-      assert(0);
+      GRID_ASSERT(0);
     };    
-    assert( delta>= tolerance );
+    GRID_ASSERT( delta>= tolerance );
 
     search();
   }
@@ -278,7 +278,7 @@ void AlgRemezGeneral::equations(){
       if(num_pows[j] != -1){ *aa++ = z; t++; }
       z *= x;
     }
-    assert(t == n+1);
+    GRID_ASSERT(t == n+1);
 
     z = (bigfloat)1l;
     t = 0;
@@ -286,7 +286,7 @@ void AlgRemezGeneral::equations(){
       if(den_pows[j] != -1){ *aa++ = -y * z; t++; }
       z *= x;
     }
-    assert(t == d);
+    GRID_ASSERT(t == d);
 
     B[i] = y * z;		// Right hand side vector
   }

Grid/algorithms/approx/RemezGeneral.h

@@ -106,7 +106,7 @@ class AlgRemezGeneral{
 		  bigfloat (*f)(bigfloat x, void *data), void *data);
 
   inline int getDegree(void) const{ 
-    assert(n==d);
+    GRID_ASSERT(n==d);
     return n;
   }
   // Reset the bounds of the approximation

Grid/algorithms/approx/ZMobius.cc

@@ -74,7 +74,7 @@ bigfloat epsilonMobius(bigfloat x, void* data){
 void computeZmobiusOmega(std::vector<ComplexD> &omega_out, const int Ls_out,
 			 const std::vector<RealD> &omega_in, const int Ls_in,
 			 const RealD lambda_bound){
-  assert(omega_in.size() == Ls_in);
+  GRID_ASSERT(omega_in.size() == Ls_in);
   omega_out.resize(Ls_out);
 
   //Use the Remez algorithm to generate the appropriate rational polynomial

Grid/algorithms/blas/BatchedBlas.h

@@ -28,6 +28,7 @@ Author: Peter Boyle <pboyle@bnl.gov>
 #pragma once
 
 #ifdef GRID_HIP
+#include <hip/hip_version.h>
 #include <hipblas/hipblas.h>
 #endif
 #ifdef GRID_CUDA
@@ -55,16 +56,17 @@ NAMESPACE_BEGIN(Grid);
   typedef cublasHandle_t gridblasHandle_t;
 #endif
 #ifdef GRID_SYCL
-  typedef cl::sycl::queue *gridblasHandle_t;
+  typedef sycl::queue *gridblasHandle_t;
 #endif
 #ifdef GRID_ONE_MKL
-  typedef cl::sycl::queue *gridblasHandle_t;
+  typedef sycl::queue *gridblasHandle_t;
 #endif
 #if !defined(GRID_SYCL) && !defined(GRID_CUDA) && !defined(GRID_HIP) && !defined(GRID_ONE_MKL)
   typedef int32_t gridblasHandle_t;
 #endif
 
 enum GridBLASOperation_t { GridBLAS_OP_N, GridBLAS_OP_T, GridBLAS_OP_C } ;
+enum GridBLASPrecision_t { GridBLAS_PRECISION_DEFAULT, GridBLAS_PRECISION_16F, GridBLAS_PRECISION_16BF, GridBLAS_PRECISION_TF32 };
 
 class GridBLAS {
 public:
@@ -89,15 +91,30 @@ public:
       gridblasHandle = theGridAccelerator;
 #endif
 #ifdef GRID_ONE_MKL
-      cl::sycl::gpu_selector selector;
+      sycl::gpu_selector selector;
-      cl::sycl::device selectedDevice { selector };
+      sycl::device selectedDevice { selector };
-      cl::sycl::property_list q_prop{cl::sycl::property::queue::in_order()};
+      sycl::property_list q_prop{sycl::property::queue::in_order()};
       gridblasHandle =new sycl::queue (selectedDevice,q_prop);
 #endif
       gridblasInit=1;
     }
   }
 
+#ifdef GRID_CUDA
+  cublasComputeType_t toDataType(GridBLASPrecision_t p) {
+    switch (p) {
+    case GridBLAS_PRECISION_16F:
+      return CUBLAS_COMPUTE_32F_FAST_16F;
+    case GridBLAS_PRECISION_16BF:
+      return CUBLAS_COMPUTE_32F_FAST_16BF;
+    case GridBLAS_PRECISION_TF32:
+      return CUBLAS_COMPUTE_32F_FAST_TF32;
+    default:
+      GRID_ASSERT(0);
+    }
+    return CUBLAS_COMPUTE_32F_FAST_16F;
+  }
+#endif
   // Force construct once
   GridBLAS() { Init(); };
   ~GridBLAS() { };
@@ -119,11 +136,11 @@ public:
   {
 #ifdef GRID_HIP
     auto err = hipDeviceSynchronize();
-    assert(err==hipSuccess);
+    GRID_ASSERT(err==hipSuccess);
 #endif
 #ifdef GRID_CUDA
     auto err = cudaDeviceSynchronize();
-    assert(err==cudaSuccess);
+    GRID_ASSERT(err==cudaSuccess);
 #endif
 #ifdef GRID_SYCL
     accelerator_barrier();
@@ -138,8 +155,10 @@ public:
 		   deviceVector<ComplexD*> &Amk,  // pointer list to matrices
 		   deviceVector<ComplexD*> &Bkn,
 		   ComplexD beta,
-		   deviceVector<ComplexD*> &Cmn)
+		   deviceVector<ComplexD*> &Cmn,
+		   GridBLASPrecision_t precision = GridBLAS_PRECISION_DEFAULT)
   {
+    GRID_ASSERT(precision == GridBLAS_PRECISION_DEFAULT);
     gemmBatched(GridBLAS_OP_N,GridBLAS_OP_N,
 		m,n,k,
 		alpha,
@@ -201,15 +220,17 @@ public:
 		   deviceVector<ComplexD*> &Amk,  // pointer list to matrices
 		   deviceVector<ComplexD*> &Bkn,
 		   ComplexD beta,
-		   deviceVector<ComplexD*> &Cmn)
+		   deviceVector<ComplexD*> &Cmn,
+		   GridBLASPrecision_t precision = GridBLAS_PRECISION_DEFAULT)
   {
+    GRID_ASSERT(precision == GridBLAS_PRECISION_DEFAULT);
     RealD t2=usecond();
     int32_t batchCount = Amk.size();
-    assert(Bkn.size()==batchCount);
+    GRID_ASSERT(Bkn.size()==batchCount);
-    assert(Cmn.size()==batchCount);
+    GRID_ASSERT(Cmn.size()==batchCount);
 
-    assert(OpA!=GridBLAS_OP_T); // Complex case expect no transpose
+    //assert(OpA!=GridBLAS_OP_T); // Complex case expect no transpose
-    assert(OpB!=GridBLAS_OP_T);
+    //assert(OpB!=GridBLAS_OP_T);
 
     int lda = m; // m x k column major
     int ldb = k; // k x n column major
@@ -235,6 +256,18 @@ public:
     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;
+#if defined(HIP_VERSION_MAJOR) && (HIP_VERSION_MAJOR >=7)
+    auto err = hipblasZgemmBatched(gridblasHandle,
+				   hOpA,
+				   hOpB,
+				   m,n,k,
+				   (hipDoubleComplex *) &alpha_p[0],
+				   (hipDoubleComplex **)&Amk[0], lda,
+				   (hipDoubleComplex **)&Bkn[0], ldb,
+				   (hipDoubleComplex *) &beta_p[0],
+				   (hipDoubleComplex **)&Cmn[0], ldc,
+				   batchCount);
+#else
     auto err = hipblasZgemmBatched(gridblasHandle,
                                    hOpA,
                                    hOpB,
@@ -245,8 +278,9 @@ public:
                                    (hipblasDoubleComplex *) &beta_p[0],
                                    (hipblasDoubleComplex **)&Cmn[0], ldc,
                                    batchCount);
+#endif
     //	 std::cout << " hipblas return code " <<(int)err<<std::endl;
-    assert(err==HIPBLAS_STATUS_SUCCESS);
+    GRID_ASSERT(err==HIPBLAS_STATUS_SUCCESS);
 #endif
 #ifdef GRID_CUDA
     cublasOperation_t hOpA;
@@ -267,7 +301,7 @@ public:
 				  (cuDoubleComplex *) &beta_p[0],
 				  (cuDoubleComplex **)&Cmn[0], ldc,
 				  batchCount);
-    assert(err==CUBLAS_STATUS_SUCCESS);
+    GRID_ASSERT(err==CUBLAS_STATUS_SUCCESS);
 #endif
 #ifdef GRID_SYCL
       int64_t m64=m;
@@ -367,28 +401,67 @@ public:
 	  Eigen::Map<Eigen::MatrixXcd> eAmk(Amk[p],m,k);
 	  Eigen::Map<Eigen::MatrixXcd> eBkn(Bkn[p],k,n);
 	  Eigen::Map<Eigen::MatrixXcd> eCmn(Cmn[p],m,n);
+	  if (std::abs(beta) != 0.0)
 	    eCmn = beta * eCmn + alpha * eAmk * eBkn ;
+	  else
+	    eCmn = alpha * eAmk * eBkn ;
         });
       } else if ( (OpA == GridBLAS_OP_C ) && (OpB == GridBLAS_OP_N) ) {
 	thread_for (p, batchCount, {
 	  Eigen::Map<Eigen::MatrixXcd> eAmk(Amk[p],k,m);
 	  Eigen::Map<Eigen::MatrixXcd> eBkn(Bkn[p],k,n);
 	  Eigen::Map<Eigen::MatrixXcd> eCmn(Cmn[p],m,n);
+	  if (std::abs(beta) != 0.0)
 	    eCmn = beta * eCmn + alpha * eAmk.adjoint() * eBkn ;
+	  else
+	    eCmn = alpha * eAmk.adjoint() * eBkn ;
+	  });
+      } else if ( (OpA == GridBLAS_OP_T ) && (OpB == GridBLAS_OP_N) ) {
+	thread_for (p, batchCount, {
+	  Eigen::Map<Eigen::MatrixXcd> eAmk(Amk[p],k,m);
+	  Eigen::Map<Eigen::MatrixXcd> eBkn(Bkn[p],k,n);
+	  Eigen::Map<Eigen::MatrixXcd> eCmn(Cmn[p],m,n);
+	  if (std::abs(beta) != 0.0)
+	    eCmn = beta * eCmn + alpha * eAmk.transpose() * eBkn ;
+	  else
+	    eCmn = alpha * eAmk.transpose() * eBkn ;
 	  });
       } else if ( (OpA == GridBLAS_OP_N ) && (OpB == GridBLAS_OP_C) ) {
 	thread_for (p, batchCount, {
 	  Eigen::Map<Eigen::MatrixXcd> eAmk(Amk[p],m,k);
 	  Eigen::Map<Eigen::MatrixXcd> eBkn(Bkn[p],n,k);
 	  Eigen::Map<Eigen::MatrixXcd> eCmn(Cmn[p],m,n);
+	  if (std::abs(beta) != 0.0)
 	    eCmn = beta * eCmn + alpha * eAmk * eBkn.adjoint() ;
+	  else
+	    eCmn = alpha * eAmk * eBkn.adjoint() ;
+	  });
+      } else if ( (OpA == GridBLAS_OP_N ) && (OpB == GridBLAS_OP_T) ) {
+	thread_for (p, batchCount, {
+	  Eigen::Map<Eigen::MatrixXcd> eAmk(Amk[p],m,k);
+	  Eigen::Map<Eigen::MatrixXcd> eBkn(Bkn[p],n,k);
+	  Eigen::Map<Eigen::MatrixXcd> eCmn(Cmn[p],m,n);
+	  eCmn = beta * eCmn + alpha * eAmk * eBkn.transpose() ;
 	  });
       } else if ( (OpA == GridBLAS_OP_C ) && (OpB == GridBLAS_OP_C) ) {
 	thread_for (p, batchCount, {
 	  Eigen::Map<Eigen::MatrixXcd> eAmk(Amk[p],k,m);
 	  Eigen::Map<Eigen::MatrixXcd> eBkn(Bkn[p],n,k);
 	  Eigen::Map<Eigen::MatrixXcd> eCmn(Cmn[p],m,n);
+	  if (std::abs(beta) != 0.0)
 	    eCmn = beta * eCmn + alpha * eAmk.adjoint() * eBkn.adjoint() ;
+	  else
+	    eCmn = alpha * eAmk.adjoint() * eBkn.adjoint() ;
+	  } );
+      } else if ( (OpA == GridBLAS_OP_T ) && (OpB == GridBLAS_OP_T) ) {
+	thread_for (p, batchCount, {
+	  Eigen::Map<Eigen::MatrixXcd> eAmk(Amk[p],k,m);
+	  Eigen::Map<Eigen::MatrixXcd> eBkn(Bkn[p],n,k);
+	  Eigen::Map<Eigen::MatrixXcd> eCmn(Cmn[p],m,n);
+	  if (std::abs(beta) != 0.0)
+	    eCmn = beta * eCmn + alpha * eAmk.transpose() * eBkn.transpose() ;
+	  else
+	    eCmn = alpha * eAmk.transpose() * eBkn.transpose() ;
 	  } );
       } else { 
 	assert(0);
@@ -409,13 +482,14 @@ public:
 		   deviceVector<ComplexF*> &Amk,  // pointer list to matrices
 		   deviceVector<ComplexF*> &Bkn,
 		   ComplexF beta,
-		   deviceVector<ComplexF*> &Cmn)
+		   deviceVector<ComplexF*> &Cmn,
+		   GridBLASPrecision_t precision = GridBLAS_PRECISION_DEFAULT)
   {
     RealD t2=usecond();
     int32_t batchCount = Amk.size();
 
-    assert(OpA!=GridBLAS_OP_T); // Complex case expect no transpose
+    //assert(OpA!=GridBLAS_OP_T); // Complex case expect no transpose
-    assert(OpB!=GridBLAS_OP_T);
+    //assert(OpB!=GridBLAS_OP_T);
 
     int lda = m; // m x k column major
     int ldb = k; // k x n column major
@@ -431,9 +505,10 @@ public:
     acceleratorCopyToDevice((void *)&beta ,(void *)&beta_p[0],sizeof(ComplexF));
     RealD t0=usecond();
 
-    assert(Bkn.size()==batchCount);
+    GRID_ASSERT(Bkn.size()==batchCount);
-    assert(Cmn.size()==batchCount);
+    GRID_ASSERT(Cmn.size()==batchCount);
 #ifdef GRID_HIP
+    GRID_ASSERT(precision == GridBLAS_PRECISION_DEFAULT);
     hipblasOperation_t hOpA;
     hipblasOperation_t hOpB;
     if ( OpA == GridBLAS_OP_N ) hOpA = HIPBLAS_OP_N;
@@ -442,6 +517,18 @@ public:
     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;
+#if defined(HIP_VERSION_MAJOR) && (HIP_VERSION_MAJOR >=7)
+    auto err = hipblasCgemmBatched(gridblasHandle,
+				   hOpA,
+				   hOpB,
+				   m,n,k,
+				   (hipComplex *) &alpha_p[0],
+				   (hipComplex **)&Amk[0], lda,
+				   (hipComplex **)&Bkn[0], ldb,
+				   (hipComplex *) &beta_p[0],
+				   (hipComplex **)&Cmn[0], ldc,
+				   batchCount);
+#else
     auto err = hipblasCgemmBatched(gridblasHandle,
                                    hOpA,
                                    hOpB,
@@ -453,7 +540,8 @@ public:
                                    (hipblasComplex **)&Cmn[0], ldc,
                                    batchCount);
 
-    assert(err==HIPBLAS_STATUS_SUCCESS);
+#endif
+    GRID_ASSERT(err==HIPBLAS_STATUS_SUCCESS);
 #endif
 #ifdef GRID_CUDA
     cublasOperation_t hOpA;
@@ -464,7 +552,9 @@ public:
     if ( OpB == GridBLAS_OP_N ) hOpB = CUBLAS_OP_N;
     if ( OpB == GridBLAS_OP_T ) hOpB = CUBLAS_OP_T;
     if ( OpB == GridBLAS_OP_C ) hOpB = CUBLAS_OP_C;
-    auto err = cublasCgemmBatched(gridblasHandle,
+    cublasStatus_t err;
+    if (precision == GridBLAS_PRECISION_DEFAULT) {
+      err = cublasCgemmBatched(gridblasHandle,
 			       hOpA,
 			       hOpB,
 			       m,n,k,
@@ -474,9 +564,23 @@ public:
 			       (cuComplex *) &beta_p[0],
 			       (cuComplex **)&Cmn[0], ldc,
 			       batchCount);
-    assert(err==CUBLAS_STATUS_SUCCESS);
+    } else {
+      cublasComputeType_t compute_precision = toDataType(precision);
+      err = cublasGemmBatchedEx(gridblasHandle,
+				hOpA,
+				hOpB,
+				m,n,k,
+				(void *) &alpha_p[0],
+				(void **)&Amk[0], CUDA_C_32F, lda,
+				(void **)&Bkn[0], CUDA_C_32F, ldb,
+				(void *) &beta_p[0],
+				(void **)&Cmn[0], CUDA_C_32F, ldc,
+				batchCount, compute_precision, CUBLAS_GEMM_DEFAULT);
+    }
+    GRID_ASSERT(err==CUBLAS_STATUS_SUCCESS);
 #endif
 #ifdef GRID_SYCL
+    GRID_ASSERT(precision == GridBLAS_PRECISION_DEFAULT);
     int64_t m64=m;
     int64_t n64=n;
     int64_t k64=k;
@@ -508,34 +612,77 @@ public:
     synchronise();
 #endif
 #if !defined(GRID_SYCL) && !defined(GRID_CUDA) && !defined(GRID_HIP)
+    GRID_ASSERT(precision == GridBLAS_PRECISION_DEFAULT);
     // Need a default/reference implementation; use Eigen
       if ( (OpA == GridBLAS_OP_N ) && (OpB == GridBLAS_OP_N) ) {
 	thread_for (p, batchCount, {
 	  Eigen::Map<Eigen::MatrixXcf> eAmk(Amk[p],m,k);
 	  Eigen::Map<Eigen::MatrixXcf> eBkn(Bkn[p],k,n);
 	  Eigen::Map<Eigen::MatrixXcf> eCmn(Cmn[p],m,n);
+	  if (std::abs(beta) != 0.0)
 	    eCmn = beta * eCmn + alpha * eAmk * eBkn ;
+	  else
+	    eCmn = alpha * eAmk * eBkn ;
 	  });
       } else if ( (OpA == GridBLAS_OP_C ) && (OpB == GridBLAS_OP_N) ) {
 	thread_for (p, batchCount, {
 	  Eigen::Map<Eigen::MatrixXcf> eAmk(Amk[p],k,m);
 	  Eigen::Map<Eigen::MatrixXcf> eBkn(Bkn[p],k,n);
 	  Eigen::Map<Eigen::MatrixXcf> eCmn(Cmn[p],m,n);
+	  if (std::abs(beta) != 0.0)
 	    eCmn = beta * eCmn + alpha * eAmk.adjoint() * eBkn ;
+	  else
+	    eCmn = alpha * eAmk.adjoint() * eBkn ;
+	  });
+      } else if ( (OpA == GridBLAS_OP_T ) && (OpB == GridBLAS_OP_N) ) {
+	thread_for (p, batchCount, {
+	  Eigen::Map<Eigen::MatrixXcf> eAmk(Amk[p],k,m);
+	  Eigen::Map<Eigen::MatrixXcf> eBkn(Bkn[p],k,n);
+	  Eigen::Map<Eigen::MatrixXcf> eCmn(Cmn[p],m,n);
+	  if (std::abs(beta) != 0.0)
+	    eCmn = beta * eCmn + alpha * eAmk.transpose() * eBkn ;
+	  else
+	    eCmn = alpha * eAmk.transpose() * eBkn ;
 	  });
       } else if ( (OpA == GridBLAS_OP_N ) && (OpB == GridBLAS_OP_C) ) {
 	thread_for (p, batchCount, {
 	  Eigen::Map<Eigen::MatrixXcf> eAmk(Amk[p],m,k);
 	  Eigen::Map<Eigen::MatrixXcf> eBkn(Bkn[p],n,k);
 	  Eigen::Map<Eigen::MatrixXcf> eCmn(Cmn[p],m,n);
+	  if (std::abs(beta) != 0.0)
 	    eCmn = beta * eCmn + alpha * eAmk * eBkn.adjoint() ;
+	  else
+	    eCmn = alpha * eAmk * eBkn.adjoint() ;
+	  });
+      } else if ( (OpA == GridBLAS_OP_N ) && (OpB == GridBLAS_OP_T) ) {
+	thread_for (p, batchCount, {
+	  Eigen::Map<Eigen::MatrixXcf> eAmk(Amk[p],m,k);
+	  Eigen::Map<Eigen::MatrixXcf> eBkn(Bkn[p],n,k);
+	  Eigen::Map<Eigen::MatrixXcf> eCmn(Cmn[p],m,n);
+	  if (std::abs(beta) != 0.0)
+	    eCmn = beta * eCmn + alpha * eAmk * eBkn.transpose() ;
+	  else
+	    eCmn = alpha * eAmk * eBkn.transpose() ;
 	  });
       } else if ( (OpA == GridBLAS_OP_C ) && (OpB == GridBLAS_OP_C) ) {
 	thread_for (p, batchCount, {
 	  Eigen::Map<Eigen::MatrixXcf> eAmk(Amk[p],k,m);
 	  Eigen::Map<Eigen::MatrixXcf> eBkn(Bkn[p],n,k);
 	  Eigen::Map<Eigen::MatrixXcf> eCmn(Cmn[p],m,n);
+	  if (std::abs(beta) != 0.0)
 	    eCmn = beta * eCmn + alpha * eAmk.adjoint() * eBkn.adjoint() ;
+	  else
+	    eCmn = alpha * eAmk.adjoint() * eBkn.adjoint() ;
+	  } );
+      } else if ( (OpA == GridBLAS_OP_T ) && (OpB == GridBLAS_OP_T) ) {
+	thread_for (p, batchCount, {
+	  Eigen::Map<Eigen::MatrixXcf> eAmk(Amk[p],k,m);
+	  Eigen::Map<Eigen::MatrixXcf> eBkn(Bkn[p],n,k);
+	  Eigen::Map<Eigen::MatrixXcf> eCmn(Cmn[p],m,n);
+	  if (std::abs(beta) != 0.0)
+	    eCmn = beta * eCmn + alpha * eAmk.transpose() * eBkn.transpose() ;
+	  else
+	    eCmn = alpha * eAmk.transpose() * eBkn.transpose() ;
 	  } );
       } else { 
 	assert(0);
@@ -562,8 +709,8 @@ public:
     RealD t2=usecond();
     int32_t batchCount = Amk.size();
 
-    assert(OpA!=GridBLAS_OP_C); // Real case no conjugate
+    GRID_ASSERT(OpA!=GridBLAS_OP_C); // Real case no conjugate
-    assert(OpB!=GridBLAS_OP_C);
+    GRID_ASSERT(OpB!=GridBLAS_OP_C);
 
     int lda = m; // m x k column major
     int ldb = k; // k x n column major
@@ -579,8 +726,8 @@ public:
     acceleratorCopyToDevice((void *)&beta ,(void *)&beta_p[0],sizeof(RealF));
     RealD t0=usecond();
 
-    assert(Bkn.size()==batchCount);
+    GRID_ASSERT(Bkn.size()==batchCount);
-    assert(Cmn.size()==batchCount);
+    GRID_ASSERT(Cmn.size()==batchCount);
 #ifdef GRID_HIP
     hipblasOperation_t hOpA;
     hipblasOperation_t hOpB;
@@ -600,7 +747,7 @@ public:
 				   (float *) &beta_p[0],
 				   (float **)&Cmn[0], ldc,
 				   batchCount);
-    assert(err==HIPBLAS_STATUS_SUCCESS);
+    GRID_ASSERT(err==HIPBLAS_STATUS_SUCCESS);
 #endif
 #ifdef GRID_CUDA
     cublasOperation_t hOpA;
@@ -621,7 +768,7 @@ public:
 				  (float *) &beta_p[0],
 				  (float **)&Cmn[0], ldc,
 				  batchCount);
-    assert(err==CUBLAS_STATUS_SUCCESS);
+    GRID_ASSERT(err==CUBLAS_STATUS_SUCCESS);
 #endif
 #ifdef GRID_SYCL
       int64_t m64=m;
@@ -661,28 +808,40 @@ public:
 	  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);
+	  if (std::abs(beta) != 0.0)
 	    eCmn = beta * eCmn + alpha * eAmk * eBkn ;
+	  else
+	    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);
+	  if (std::abs(beta) != 0.0)
 	    eCmn = beta * eCmn + alpha * eAmk.transpose() * eBkn ;
+	  else
+	    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);
+	  if (std::abs(beta) != 0.0)
 	    eCmn = beta * eCmn + alpha * eAmk * eBkn.transpose() ;
+	  else
+	    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);
+	  if (std::abs(beta) != 0.0)
 	    eCmn = beta * eCmn + alpha * eAmk.transpose() * eBkn.transpose() ;
+	  else
+	    eCmn = alpha * eAmk.transpose() * eBkn.transpose() ;
 	  });
       } else { 
 	assert(0);
@@ -709,8 +868,8 @@ public:
     RealD t2=usecond();
     int32_t batchCount = Amk.size();
 
-    assert(OpA!=GridBLAS_OP_C); // Real case no conjugate
+    GRID_ASSERT(OpA!=GridBLAS_OP_C); // Real case no conjugate
-    assert(OpB!=GridBLAS_OP_C);
+    GRID_ASSERT(OpB!=GridBLAS_OP_C);
 
     int lda = m; // m x k column major
     int ldb = k; // k x n column major
@@ -727,8 +886,8 @@ public:
     acceleratorCopyToDevice((void *)&beta ,(void *)&beta_p[0],sizeof(RealD));
     RealD t0=usecond();
 
-    assert(Bkn.size()==batchCount);
+    GRID_ASSERT(Bkn.size()==batchCount);
-    assert(Cmn.size()==batchCount);
+    GRID_ASSERT(Cmn.size()==batchCount);
 #ifdef GRID_HIP
     hipblasOperation_t hOpA;
     hipblasOperation_t hOpB;
@@ -748,7 +907,7 @@ public:
 				   (double *) &beta_p[0],
 				   (double **)&Cmn[0], ldc,
 				   batchCount);
-    assert(err==HIPBLAS_STATUS_SUCCESS);
+    GRID_ASSERT(err==HIPBLAS_STATUS_SUCCESS);
 #endif
 #ifdef GRID_CUDA
     cublasOperation_t hOpA;
@@ -769,7 +928,7 @@ public:
 				  (double *) &beta_p[0],
 				  (double **)&Cmn[0], ldc,
 				  batchCount);
-    assert(err==CUBLAS_STATUS_SUCCESS);
+    GRID_ASSERT(err==CUBLAS_STATUS_SUCCESS);
 #endif
 #ifdef GRID_SYCL
       int64_t m64=m;
@@ -809,28 +968,40 @@ public:
 	  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);
+	  if (std::abs(beta) != 0.0)
 	    eCmn = beta * eCmn + alpha * eAmk * eBkn ;
+	  else
+	    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);
+	  if (std::abs(beta) != 0.0)
 	    eCmn = beta * eCmn + alpha * eAmk.transpose() * eBkn ;
+	  else
+	    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);
+	  if (std::abs(beta) != 0.0)
 	    eCmn = beta * eCmn + alpha * eAmk * eBkn.transpose() ;
+	  else
+	    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);
+	  if (std::abs(beta) != 0.0)
 	    eCmn = beta * eCmn + alpha * eAmk.transpose() * eBkn.transpose() ;
+	  else
+	    eCmn = alpha * eAmk.transpose() * eBkn.transpose() ;
 	  });
       } else { 
 	assert(0);
@@ -841,6 +1012,372 @@ public:
      RealD bytes = 1.0*sizeof(RealD)*(m*k+k*n+m*n)*batchCount;
   }
 
+  /*
+    Inverse and Determinant
+
+    - CPU version uses Eigen
+    - GPU version uses LAPACK-compatible getrf / getri
+
+    Design comment: Eigen does not expose getrf / getri in a LAPACK compatible manner.
+                    Overhead to go through getrf / getri for CPU version too large.
+		    Current interface therefore only guarantees the inverse and determinant
+		    functions on all platforms but not the getrf / getri ones.
+  */
+#if !defined(GRID_SYCL) && !defined(GRID_CUDA) && !defined(GRID_HIP)
+
+  void inverseBatched(int64_t n,
+		      deviceVector<ComplexD*> &Ann,
+		      deviceVector<ComplexD*> &Cnn) {
+
+    int64_t batchCount = Ann.size();
+    GRID_ASSERT(batchCount == Cnn.size());
+    thread_for(p,batchCount, {
+	Eigen::Map<Eigen::MatrixXcd> eAnn(Ann[p],n,n);
+	Eigen::Map<Eigen::MatrixXcd> eCnn(Cnn[p],n,n);
+	eCnn = eAnn.inverse();
+      });
+  }
+
+  void inverseBatched(int64_t n,
+		      deviceVector<ComplexF*> &Ann,
+		      deviceVector<ComplexF*> &Cnn) {
+
+    int64_t batchCount = Ann.size();
+    GRID_ASSERT(batchCount == Cnn.size());
+    thread_for(p,batchCount, {
+	Eigen::Map<Eigen::MatrixXcf> eAnn(Ann[p],n,n);
+	Eigen::Map<Eigen::MatrixXcf> eCnn(Cnn[p],n,n);
+	eCnn = eAnn.inverse();
+      });
+  }
+
+  void determinantBatched(int64_t n,
+			  deviceVector<ComplexD*> &Ann,
+			  deviceVector<ComplexD*> &C) {
+
+    int64_t batchCount = Ann.size();
+    GRID_ASSERT(batchCount == C.size());
+    thread_for(p,batchCount, {
+	Eigen::Map<Eigen::MatrixXcd> eAnn(Ann[p],n,n);
+	*C[p] = eAnn.determinant();
+      });
+  }
+
+  void determinantBatched(int64_t n,
+			  deviceVector<ComplexF*> &Ann,
+			  deviceVector<ComplexF*> &C) {
+
+    int64_t batchCount = Ann.size();
+    GRID_ASSERT(batchCount == C.size());
+    thread_for(p,batchCount, {
+	Eigen::Map<Eigen::MatrixXcf> eAnn(Ann[p],n,n);
+	*C[p] = eAnn.determinant();
+      });
+  }
+  
+#else
+
+#ifdef GRID_SYCL
+  template<typename T>
+  void getrfBatchedSYCL(int64_t n,
+			deviceVector<T*> &Ann,
+			deviceVector<int64_t> &ipiv,
+			deviceVector<int64_t> &info) {
+    
+    int64_t batchCount = Ann.size();
+
+    static deviceVector<T> scratchpad;
+    int64_t sp_size = oneapi::mkl::lapack::getrf_batch_scratchpad_size<T>(*gridblasHandle, &n, &n, &n, (int64_t)1, &batchCount);
+    if (sp_size > scratchpad.size())
+      scratchpad.resize(sp_size);
+
+    static deviceVector<int64_t*> _ipiv;
+    if (batchCount > _ipiv.size())
+      _ipiv.resize(batchCount);
+    int64_t** p_ipiv = &_ipiv[0];
+    int64_t* pipiv = &ipiv[0];
+
+    accelerator_for(i, batchCount, 1, { p_ipiv[i] = &pipiv[i*n]; });
+
+    oneapi::mkl::lapack::getrf_batch(*gridblasHandle,
+				    &n, &n,
+				    (T **)&Ann[0],
+				    &n,
+				    (int64_t**)&_ipiv[0],
+				    (int64_t)1, &batchCount,
+				    (T*)&scratchpad[0], (int64_t)scratchpad.size(),
+				    std::vector<sycl::event>());
+    synchronise();
+  }
+#endif
+
+  void getrfBatched(int64_t n,
+		    deviceVector<ComplexD*> &Ann,
+		    deviceVector<int64_t> &ipiv,
+		    deviceVector<int64_t> &info)
+  {
+    int64_t batchCount = Ann.size();
+    GRID_ASSERT(ipiv.size()==batchCount*n);
+    GRID_ASSERT(info.size()==batchCount);
+
+#ifdef GRID_HIP
+#if defined(HIP_VERSION_MAJOR) && (HIP_VERSION_MAJOR >=7)
+    auto err = hipblasZgetrfBatched(gridblasHandle,(int)n,
+				    (hipDoubleComplex **)&Ann[0], (int)n,
+				    (int*) &ipiv[0],
+				    (int*) &info[0],
+				    (int)batchCount);
+#else
+    auto err = hipblasZgetrfBatched(gridblasHandle,(int)n,
+                                    (hipblasDoubleComplex **)&Ann[0], (int)n,
+                                    (int*) &ipiv[0],
+                                    (int*) &info[0],
+                                    (int)batchCount);
+#endif
+    GRID_ASSERT(err==HIPBLAS_STATUS_SUCCESS);
+#endif
+#ifdef GRID_CUDA
+    auto err = cublasZgetrfBatched(gridblasHandle, (int)n,
+				   (cuDoubleComplex **)&Ann[0], (int)n,
+				   (int*) &ipiv[0],
+				   (int*) &info[0],
+				   (int)batchCount);
+    GRID_ASSERT(err==CUBLAS_STATUS_SUCCESS);
+#endif
+#ifdef GRID_SYCL
+    getrfBatchedSYCL(n, Ann, ipiv, info);
+#endif
+  }
+
+  void getrfBatched(int64_t n,
+		    deviceVector<ComplexF*> &Ann,
+		    deviceVector<int64_t> &ipiv,
+		    deviceVector<int64_t> &info)
+  {
+    int64_t batchCount = Ann.size();
+    GRID_ASSERT(ipiv.size()==batchCount*n);
+    GRID_ASSERT(info.size()==batchCount);
+
+#ifdef GRID_HIP
+#if defined(HIP_VERSION_MAJOR) && (HIP_VERSION_MAJOR >=7)
+    auto err = hipblasCgetrfBatched(gridblasHandle,(int)n,
+				    (hipComplex **)&Ann[0], (int)n,
+				    (int*) &ipiv[0],
+				    (int*) &info[0],
+				    (int)batchCount);
+#else
+    auto err = hipblasCgetrfBatched(gridblasHandle,(int)n,
+                                    (hipblasComplex **)&Ann[0], (int)n,
+                                    (int*) &ipiv[0],
+                                    (int*) &info[0],
+                                    (int)batchCount);
+#endif
+
+    GRID_ASSERT(err==HIPBLAS_STATUS_SUCCESS);
+#endif
+#ifdef GRID_CUDA
+    auto err = cublasCgetrfBatched(gridblasHandle, (int)n,
+				   (cuComplex **)&Ann[0], (int)n,
+				   (int*) &ipiv[0],
+				   (int*) &info[0],
+				   (int)batchCount);
+    GRID_ASSERT(err==CUBLAS_STATUS_SUCCESS);
+#endif
+#ifdef GRID_SYCL
+    getrfBatchedSYCL(n, Ann, ipiv, info);
+#endif
+  }
+
+#ifdef GRID_SYCL
+  template<typename T>
+  void getriBatchedSYCL(int64_t n,
+			deviceVector<T*> &Ann,
+			deviceVector<int64_t> &ipiv,
+			deviceVector<int64_t> &info,
+			deviceVector<T*> &Cnn) {
+
+    int64_t batchCount = Ann.size();
+
+    static deviceVector<T> scratchpad;
+    int64_t sp_size = oneapi::mkl::lapack::getri_batch_scratchpad_size<T>(*gridblasHandle, &n, &n, (int64_t)1, &batchCount);
+    if (sp_size > scratchpad.size())
+      scratchpad.resize(sp_size);
+
+    static deviceVector<int64_t*> _ipiv;
+    if (batchCount > _ipiv.size())
+      _ipiv.resize(batchCount);
+    int64_t** p_ipiv = &_ipiv[0];
+    int64_t* pipiv = &ipiv[0];
+
+    accelerator_for(i, batchCount, 1, { p_ipiv[i] = &pipiv[i*n]; });
+
+    oneapi::mkl::lapack::getri_batch(*gridblasHandle,
+				     &n,
+				     (T **)&Ann[0],
+				     &n,
+				     (int64_t**)p_ipiv,
+				     (int64_t)1, &batchCount,
+				     (T *)&scratchpad[0], (int64_t)scratchpad.size(),
+				     std::vector<sycl::event>());
+
+    synchronise();
+
+    T** pA = &Ann[0];
+    T** pC = &Cnn[0];
+    accelerator_for(i, batchCount*n*n, 1, {
+	auto j = i / batchCount;
+	auto k = i % batchCount;
+	pC[k][j] = pA[k][j];
+      });
+  }
+
+#endif
+
+  void getriBatched(int64_t n,
+		    deviceVector<ComplexD*> &Ann,
+		    deviceVector<int64_t> &ipiv,
+		    deviceVector<int64_t> &info,
+		    deviceVector<ComplexD*> &Cnn)
+  {
+    int64_t batchCount = Ann.size();
+    GRID_ASSERT(ipiv.size()==batchCount*n);
+    GRID_ASSERT(info.size()==batchCount);
+    GRID_ASSERT(Cnn.size()==batchCount);
+
+#ifdef GRID_HIP
+#if defined(HIP_VERSION_MAJOR) && (HIP_VERSION_MAJOR >=7)
+    auto err = hipblasZgetriBatched(gridblasHandle,(int)n,
+				    (hipDoubleComplex **)&Ann[0], (int)n,
+				    (int*) &ipiv[0],
+				    (hipDoubleComplex **)&Cnn[0], (int)n,
+				    (int*) &info[0],
+				    (int)batchCount);
+#else
+    auto err = hipblasZgetriBatched(gridblasHandle,(int)n,
+                                    (hipblasDoubleComplex **)&Ann[0], (int)n,
+                                    (int*) &ipiv[0],
+                                    (hipblasDoubleComplex **)&Cnn[0], (int)n,
+                                    (int*) &info[0],
+                                    (int)batchCount);
+
+#endif
+    GRID_ASSERT(err==HIPBLAS_STATUS_SUCCESS);
+#endif
+#ifdef GRID_CUDA
+    auto err = cublasZgetriBatched(gridblasHandle, (int)n,
+				   (cuDoubleComplex **)&Ann[0], (int)n,
+				   (int*) &ipiv[0],
+				   (cuDoubleComplex **)&Cnn[0], (int)n,
+				   (int*) &info[0],
+				   (int)batchCount);
+    GRID_ASSERT(err==CUBLAS_STATUS_SUCCESS);
+#endif
+#ifdef GRID_SYCL
+    getriBatchedSYCL(n, Ann, ipiv, info, Cnn);
+#endif
+  }
+
+  void getriBatched(int64_t n,
+		    deviceVector<ComplexF*> &Ann,
+		    deviceVector<int64_t> &ipiv,
+		    deviceVector<int64_t> &info,
+		    deviceVector<ComplexF*> &Cnn)
+  {
+    int64_t batchCount = Ann.size();
+    GRID_ASSERT(ipiv.size()==batchCount*n);
+    GRID_ASSERT(info.size()==batchCount);
+    GRID_ASSERT(Cnn.size()==batchCount);
+
+#ifdef GRID_HIP
+#if defined(HIP_VERSION_MAJOR) && (HIP_VERSION_MAJOR >=7)
+    auto err = hipblasCgetriBatched(gridblasHandle,(int)n,
+				    (hipComplex **)&Ann[0], (int)n,
+				    (int*) &ipiv[0],
+				    (hipComplex **)&Cnn[0], (int)n,
+				    (int*) &info[0],
+				    (int)batchCount);
+#else
+    auto err = hipblasCgetriBatched(gridblasHandle,(int)n,
+                                    (hipblasComplex **)&Ann[0], (int)n,
+                                    (int*) &ipiv[0],
+                                    (hipblasComplex **)&Cnn[0], (int)n,
+                                    (int*) &info[0],
+                                    (int)batchCount);
+#endif
+    GRID_ASSERT(err==HIPBLAS_STATUS_SUCCESS);
+#endif
+#ifdef GRID_CUDA
+    auto err = cublasCgetriBatched(gridblasHandle, (int)n,
+				   (cuComplex **)&Ann[0], (int)n,
+				   (int*) &ipiv[0],
+				   (cuComplex **)&Cnn[0], (int)n,
+				   (int*) &info[0],
+				   (int)batchCount);
+    GRID_ASSERT(err==CUBLAS_STATUS_SUCCESS);
+#endif
+#ifdef GRID_SYCL
+    getriBatchedSYCL(n, Ann, ipiv, info, Cnn);
+#endif
+  }
+
+  template<typename dtype>
+  void inverseBatched(int64_t n,
+		      deviceVector<dtype*> &Ann, // this will be overwritten with LU decomposition
+		      deviceVector<dtype*> &Cnn  // this will be overwritten with the inverse
+		      ) {
+
+    int64_t batchCount = Ann.size();
+    RealD t0 = usecond();
+    deviceVector<int64_t> ipiv(batchCount*n);
+    deviceVector<int64_t> info(batchCount);
+
+    //RealD t1 = usecond();
+    getrfBatched(n, Ann, ipiv, info);
+    // test info for non-invertibility?  set to nan if yes?
+    getriBatched(n, Ann, ipiv, info, Cnn);
+    //synchronise();
+    //RealD t2 = usecond();
+    //std::cout << GridLogMessage << "Temp " << t1-t0 << " rf/ri " << t2-t1  << std::endl;
+  }
+
+  template<typename dtype>
+  void determinantBatched(int64_t n,
+			  deviceVector<dtype*> &Ann, // this will be overwritten with LU decomposition
+			  deviceVector<dtype*> &C    // this will be overwritten with determinant
+			  ) {
+
+    int64_t batchCount = Ann.size();
+    //RealD t0 = usecond();
+    deviceVector<int64_t> ipiv(batchCount*n);
+    deviceVector<int64_t> info(batchCount);
+    
+    dtype** pAnn = (dtype**)&Ann[0];
+    dtype** pC = (dtype**)&C[0];
+#if defined(GRID_CUDA) || defined(GRID_HIP)
+    int* pipiv = (int*)&ipiv[0];
+#else
+    int64_t* pipiv = (int64_t*)&ipiv[0];
+#endif
+
+    //RealD t1 = usecond();
+    getrfBatched(n, Ann, ipiv, info);
+    //RealD t2 = usecond();
+    accelerator_for(i,batchCount,1,{
+	dtype det = 1.0;
+	for (int64_t j=0;j<n;j++) {
+	  det *= pAnn[i][n*j + j];
+	  // branchless signs
+	  det *= (pipiv[i*n + j] == j+1) ? (1.0) : (-1.0);
+	}
+	*pC[i] = det;
+      });
+    
+    //RealD t3 = usecond();
+    //std::cout << GridLogMessage << "Temp " << t1 - t0 << " rf/ri " << t2-t1  << "final" << t3 - t2 << std::endl;
+  }
+#endif
+  
+  
   template<class CComplex>
   double benchmark(int M, int N, int K, int BATCH)
   {

Grid/algorithms/blas/MomentumProject.h

@@ -0,0 +1,300 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: MomentumProject.h
+
+    Copyright (C) 2025
+
+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);
+/* 
+   MultiMomProject
+
+   Import vectors -> nxyz x (ncomponent x nt)
+   Import complex phases -> nmom x nxy
+
+   apply = via (possibly batched) GEMM
+*/
+template<class Field, class ComplexField>
+class MomentumProject
+{
+public:
+
+  typedef typename Field::scalar_type   scalar;
+  typedef typename Field::scalar_object scalar_object;
+
+  GridBase *grid;
+  uint64_t nmom;
+  uint64_t nxyz;
+  uint64_t nt;
+  uint64_t nbtw;
+  uint64_t words;
+
+  deviceVector<scalar> BLAS_V;      // 
+  deviceVector<scalar> BLAS_M;      // 
+  deviceVector<scalar> BLAS_P;      // 
+  
+  MomentumProject(){};
+ ~MomentumProject(){ Deallocate(); };
+  
+  void Deallocate(void)
+  {
+    grid=nullptr;
+    nmom=0;
+    nxyz=0;
+    nt=0;
+    nbtw=0;
+    words=0;
+    BLAS_V.resize(0);
+    BLAS_M.resize(0);
+    BLAS_P.resize(0);
+  }
+  void Allocate(int _nmom,GridBase *_grid)
+  {
+    grid=_grid;
+    Coordinate ldims = grid->LocalDimensions();
+
+    nmom=_nmom;
+    nt   = ldims[grid->Nd()-1];
+    nxyz = grid->lSites()/nt;
+    words = sizeof(scalar_object)/sizeof(scalar);
+    nbtw = nt * words;
+
+    BLAS_V.resize (nxyz * nt * words );
+    BLAS_M.resize (nmom * nxyz       );
+    BLAS_P.resize (nmom * nt * words );
+  }
+  void ImportMomenta(const std::vector <ComplexField> &momenta)
+  {
+    GRID_ASSERT(momenta.size()==nmom);
+    //    might as well just make the momenta here
+    typedef typename Field::vector_object vobj;
+
+    int nd = grid->_ndimension;
+
+    uint64_t sz = BLAS_M.size();
+
+    GRID_ASSERT(momenta.size()==nmom)
+    GRID_ASSERT(momenta[0].Grid()==grid);
+    GRID_ASSERT(sz = nxyz * nmom);
+    
+    Coordinate rdimensions = grid->_rdimensions;
+    Coordinate ldims       = grid->LocalDimensions();
+    int64_t osites         = grid->oSites();
+    Coordinate simd        = grid->_simd_layout;
+    const int Nsimd        = vobj::Nsimd();
+    uint64_t lwords        = words; // local variable for copy in to GPU
+    int64_t Nxyz = nxyz;
+    auto blasData_p  = &BLAS_M[0];
+    for(int m=0;m<momenta.size();m++){
+
+      autoView( Data   , momenta[m], AcceleratorRead);
+      auto Data_p  = &Data[0];
+
+      accelerator_for(xyz,nxyz,1,{
+	  //////////////////////////////////////////
+	  // isite -- map lane within buffer to lane within lattice
+	  ////////////////////////////////////////////
+	    Coordinate lcoor(nd,0);
+	    Lexicographic::CoorFromIndex(lcoor,xyz,ldims);
+	    
+	    Coordinate icoor(nd);
+	    Coordinate ocoor(nd);
+	    for (int d = 0; d < nd; d++) {
+	      icoor[d] = lcoor[d]/rdimensions[d];
+	      ocoor[d] = lcoor[d]%rdimensions[d];
+	    }
+	    int64_t osite;
+	    int64_t isite;
+	    Lexicographic::IndexFromCoor(ocoor,osite,rdimensions);
+	    Lexicographic::IndexFromCoor(icoor,isite,simd);
+	    
+	    // BLAS_M[nmom][slice_vol]
+	    // Fortran Column major BLAS layout is M_xyz,mom
+	    scalar data = extractLane(isite,Data[osite]);
+	    uint64_t idx = xyz+m*Nxyz;
+	    blasData_p[idx] = data;
+	});
+    }
+  }
+  void ImportVector(Field &vec)
+  {
+    typedef typename Field::vector_object vobj;
+
+    int nd = grid->_ndimension;
+
+    uint64_t sz = BLAS_V.size();
+
+    GRID_ASSERT(sz = nxyz * words * nt);
+    
+    Coordinate rdimensions = grid->_rdimensions;
+    Coordinate ldims= grid->LocalDimensions();
+    int64_t osites = grid->oSites();
+    Coordinate simd = grid->_simd_layout;
+    const int Nsimd = vobj::Nsimd();
+    uint64_t lwords= words; // local variable for copy in to GPU
+
+    auto blasData_p  = &BLAS_V[0];
+    autoView( Data   , vec, AcceleratorRead);
+    auto Data_p  = &Data[0];
+
+    int64_t nwords = words;// for capture
+    int64_t Nt     = nt;// for capture
+    
+    accelerator_for(sf,osites,Nsimd,{
+#ifdef GRID_SIMT
+        {
+	  int lane=acceleratorSIMTlane(Nsimd); // buffer lane
+#else
+	  for(int lane=0;lane<Nsimd;lane++) {
+#endif
+	  //////////////////////////////////////////
+	  // isite -- map lane within buffer to lane within lattice
+	  ////////////////////////////////////////////
+	    Coordinate lcoor(nd,0);
+	    Coordinate icoor(nd);
+	    Coordinate ocoor(nd);
+	    
+	    Lexicographic::CoorFromIndex(icoor,lane,simd);
+	    Lexicographic::CoorFromIndex(ocoor,sf,rdimensions);
+
+	  
+	    int64_t l_xyz = 0;
+	    for (int d = 0; d < nd; d++) {
+	      lcoor[d] = rdimensions[d]*icoor[d] + ocoor[d];
+	    }
+	    uint64_t l_t   = lcoor[nd-1];
+
+	    Coordinate xyz_coor = lcoor;
+	    xyz_coor[nd-1] =0;
+	    Lexicographic::IndexFromCoor(xyz_coor,l_xyz,ldims);
+
+	    
+	    scalar_object data = extractLane(lane,Data[sf]);
+	    scalar *data_words = (scalar *) &data;
+	    for(int w = 0 ; w < nwords; w++) {
+	      // BLAS_V[slice_vol][nt][words]
+	      // Fortran Column major BLAS layout is V_(t,w)_xyz
+	      uint64_t idx = w+l_t*nwords + l_xyz * nwords * Nt;
+	      blasData_p[idx] = data_words[w];
+	    }
+#ifdef GRID_SIMT
+	}
+#else
+	}
+#endif
+	});
+  }
+  void ExportMomentumProjection(std::vector<typename Field::scalar_object> &projection)
+  {
+    projection.resize(nmom*nt);
+    acceleratorCopyFromDevice(&BLAS_P[0],(scalar *)&projection[0],BLAS_P.size()*sizeof(scalar));
+    // Could decide on a layout late?
+  }
+
+  // Row major layout "C" order:
+  // BLAS_V[slice_vol][nt][words]
+  // BLAS_M[nmom][slice_vol]
+  // BLAS_P[nmom][nt][words]
+  //
+  // Fortran Column major BLAS layout is V_(w,t)_xyz
+  // Fortran Column major BLAS layout is M_xyz,mom
+  // Fortran Column major BLAS layout is P_(w,t),mom
+  //
+  // Projected
+  //
+  // P = (V * M)_(w,t),mom
+  //
+  void Project(Field &data,std::vector< typename Field::scalar_object > & projected_gdata)
+  {
+    double t_import=0;
+    double t_export=0;
+    double t_gemm  =0;
+    double t_allreduce=0;
+    t_import-=usecond();
+    this->ImportVector(data);
+
+    std::vector< typename Field::scalar_object > projected_planes;
+
+    deviceVector<scalar *> Vd(1);
+    deviceVector<scalar *> Md(1);
+    deviceVector<scalar *> Pd(1);
+
+    scalar * Vh = & BLAS_V[0];
+    scalar * Mh = & BLAS_M[0];
+    scalar * Ph = & BLAS_P[0];
+
+    acceleratorPut(Vd[0],Vh);
+    acceleratorPut(Md[0],Mh);
+    acceleratorPut(Pd[0],Ph);
+    t_import+=usecond();
+
+    GridBLAS BLAS;
+
+    /////////////////////////////////////////
+    // P_im = VMmx . Vxi
+    /////////////////////////////////////////
+    t_gemm-=usecond();
+    BLAS.gemmBatched(GridBLAS_OP_N,GridBLAS_OP_N, 
+    		     words*nt,nmom,nxyz,
+		     scalar(1.0),
+		     Vd,
+		     Md,
+		     scalar(0.0),  // wipe out result
+		     Pd);
+    BLAS.synchronise();
+    t_gemm+=usecond();
+
+    t_export-=usecond();
+    ExportMomentumProjection(projected_planes); // resizes
+    t_export+=usecond();
+
+    /////////////////////////////////
+    // Reduce across MPI ranks
+    /////////////////////////////////
+    int nd = grid->Nd();
+    int gt = grid->GlobalDimensions()[nd-1];
+    int lt = grid->LocalDimensions()[nd-1];
+    projected_gdata.resize(gt*nmom);
+    for(int t=0;t<gt*nmom;t++){ // global Nt array with zeroes for stuff not on this node
+      projected_gdata[t]=Zero();
+    }
+    for(int t=0;t<lt;t++){
+    for(int m=0;m<nmom;m++){
+      int st = grid->LocalStarts()[nd-1];
+      projected_gdata[t+st + gt*m] = projected_planes[t+lt*m];
+    }}
+    t_allreduce-=usecond();
+    grid->GlobalSumVector((scalar *)&projected_gdata[0],gt*nmom*words);
+    t_allreduce+=usecond();
+
+    std::cout << GridLogPerformance<<" MomentumProject t_import  "<<t_import<<"us"<<std::endl;
+    std::cout << GridLogPerformance<<" MomentumProject t_export  "<<t_export<<"us"<<std::endl;
+    std::cout << GridLogPerformance<<" MomentumProject t_gemm    "<<t_gemm<<"us"<<std::endl;
+    std::cout << GridLogPerformance<<" MomentumProject t_reduce  "<<t_allreduce<<"us"<<std::endl;
+
+  }
+};
+
+NAMESPACE_END(Grid);

Grid/algorithms/deflation/Deflation.h

@@ -69,8 +69,8 @@ public:
   DeflatedGuesser(const std::vector<Field> & _evec, const std::vector<RealD> & _eval, const unsigned int _N)
   : evec(_evec), eval(_eval), N(_N)
   {
-    assert(evec.size()==eval.size());
+    GRID_ASSERT(evec.size()==eval.size());
-    assert(N <= evec.size());
+    GRID_ASSERT(N <= evec.size());
   } 
 
   virtual void operator()(const Field &src,Field &guess) {
@@ -141,8 +141,7 @@ public:
     }
     //postprocessing
     std::cout << GridLogMessage << "Start BlockPromote for loop" << std::endl;
-    for (int j=0;j<Nsrc;j++)
+    for (int j=0;j<Nsrc;j++) {
-    {
       std::cout << GridLogMessage << "BlockProject iter: " << j << std::endl;
       blockPromote(guess_coarse[j],guess[j],subspace);
       guess[j].Checkerboard() = src[j].Checkerboard();

Grid/algorithms/deflation/MultiRHSBlockCGLinalg.h

@@ -0,0 +1,376 @@
+/*************************************************************************************
+
+    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 <<GridLogPerformance << "MulMatrix alloc    took "<< t1-t0<<" us"<<std::endl;
+    std::cout <<GridLogPerformance<< "MulMatrix preamble took "<< t2-t1<<" us"<<std::endl;
+    std::cout <<GridLogPerformance<< "MulMatrix blas     took "<< t3-t2<<" us"<<std::endl;
+    std::cout <<GridLogPerformance<< "MulMatrix copy     took "<< t4-t3<<" us"<<std::endl;
+    std::cout <<GridLogPerformance<< "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();
+    GRID_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 <<GridLogPerformance<< "InnerProductMatrix m,n,k "<< M<<","<<N<<","<<K<<std::endl;
+    std::cout <<GridLogPerformance<< "InnerProductMatrix alloc t1 "<< t1-t0<<" us"<<std::endl;
+    std::cout <<GridLogPerformance<< "InnerProductMatrix cp    t2 "<< t2-t1<<" us"<<std::endl;
+    std::cout <<GridLogPerformance<< "InnerProductMatrix setup t3 "<< t3-t2<<" us"<<std::endl;
+    std::cout <<GridLogPerformance<< "InnerProductMatrix blas t4 "<< t4-t3<<" us"<<std::endl;
+    std::cout <<GridLogPerformance<< "InnerProductMatrix blas    "<< flops<<" GF/s"<<std::endl;
+    std::cout <<GridLogPerformance<< "InnerProductMatrix blas    "<< bytes<<" GB/s"<<std::endl;
+    std::cout <<GridLogPerformance<< "InnerProductMatrix gsum t5 "<< t5-t4<<" us"<<std::endl;
+    std::cout <<GridLogPerformance<< "InnerProductMatrix cp   t6 "<< t6-t5<<" us"<<std::endl;
+    std::cout <<GridLogPerformance<< "InnerProductMatrix took "<< t6-t0<<" us"<<std::endl;
+#else
+    int nrhs;
+    GridBase *grid;
+    uint64_t vol;
+    uint64_t words;
+
+    nrhs = X.size();
+    GRID_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;
+    GRID_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 <<GridLogPerformance<< "InnerProductMatrix m,n,k "<< M<<","<<N<<","<<K<<std::endl;
+    std::cout <<GridLogPerformance<< "InnerProductMatrix alloc t1 "<< t1-t0<<" us"<<std::endl;
+    std::cout <<GridLogPerformance<< "InnerProductMatrix cp    t2 "<< t2-t1<<" us "<<xybytes<<" GB/s"<<std::endl;
+    std::cout <<GridLogPerformance<< "InnerProductMatrix setup t3 "<< t3-t2<<" us"<<std::endl;
+    std::cout <<GridLogPerformance<< "InnerProductMatrix blas t4 "<< t4-t3<<" us"<<std::endl;
+    std::cout <<GridLogPerformance<< "InnerProductMatrix blas    "<< flops<<" GF/s"<<std::endl;
+    std::cout <<GridLogPerformance<< "InnerProductMatrix blas    "<< bytes<<" GB/s"<<std::endl;
+    std::cout <<GridLogPerformance<< "InnerProductMatrix cp     t5 "<< t5-t4<<" us"<<std::endl;
+    std::cout <<GridLogPerformance<< "InnerProductMatrix lsum   t6l "<< t6l-t5<<" us"<<std::endl;
+    std::cout <<GridLogPerformance<< "InnerProductMatrix gsum   t6 "<< t6-t6l<<" us"<<std::endl;
+    std::cout <<GridLogPerformance<< "InnerProductMatrix took "<< t6-t0<<" us"<<std::endl;
+#endif
+  }
+};
+
+NAMESPACE_END(Grid);

Grid/algorithms/deflation/MultiRHSBlockProject.h

@@ -131,12 +131,12 @@ public:
     typedef typename Field::vector_object vobj;
     //    std::cout << GridLogMessage <<" BlockProjector importing "<<nvec<< " fine grid vectors" <<std::endl;
 
-    assert(vecs[0].Grid()==fine_grid);
+    GRID_ASSERT(vecs[0].Grid()==fine_grid);
 
     subdivides(coarse_grid,fine_grid); // require they map
 
     int _ndimension = coarse_grid->_ndimension;
-    assert(block_vol == fine_grid->oSites() / coarse_grid->oSites());
+    GRID_ASSERT(block_vol == fine_grid->oSites() / coarse_grid->oSites());
     
     Coordinate  block_r      (_ndimension);
     for(int d=0 ; d<_ndimension;d++){
@@ -164,7 +164,7 @@ public:
       const int Nsimd = vobj::Nsimd();
       //      std::cout << "sz "<<sz<<std::endl;
       //      std::cout << "prod "<<Nsimd * coarse_grid->oSites() * block_vol * nvec * words<<std::endl;
-      assert(sz == Nsimd * coarse_grid->oSites() * block_vol * nvec * words);
+      GRID_ASSERT(sz == Nsimd * coarse_grid->oSites() * block_vol * nvec * words);
       uint64_t lwords= words; // local variable for copy in to GPU
       accelerator_for(sf,osites,Nsimd,{
 #ifdef GRID_SIMT
@@ -198,7 +198,7 @@ public:
    	               + v*bv
 	               + sb;
 
-	  //	  assert(site*lwords<sz);
+	  //	  GRID_ASSERT(site*lwords<sz);
 
 	  scalar_object * ptr = (scalar_object *)&blasData_p[site*lwords];
 
@@ -219,12 +219,12 @@ public:
 
     int nvec = vecs.size();
 
-    assert(vecs[0].Grid()==fine_grid);
+    GRID_ASSERT(vecs[0].Grid()==fine_grid);
 
     subdivides(coarse_grid,fine_grid); // require they map
 
     int _ndimension = coarse_grid->_ndimension;
-    assert(block_vol == fine_grid->oSites() / coarse_grid->oSites());
+    GRID_ASSERT(block_vol == fine_grid->oSites() / coarse_grid->oSites());
     
     Coordinate  block_r      (_ndimension);
     for(int d=0 ; d<_ndimension;d++){
@@ -299,7 +299,7 @@ public:
 
     //    std::cout << " BlockProjector importing "<<nvec<< " coarse grid vectors" <<std::endl;
 
-    assert(vecs[0].Grid()==coarse_grid);
+    GRID_ASSERT(vecs[0].Grid()==coarse_grid);
 
     int _ndimension = coarse_grid->_ndimension;
 
@@ -320,7 +320,7 @@ public:
       // loop over fine sites
       const int Nsimd = vobj::Nsimd();
       uint64_t cwords=sizeof(typename vobj::scalar_object)/sizeof(scalar);
-      assert(cwords==nbasis);
+      GRID_ASSERT(cwords==nbasis);
       
       accelerator_for(sc,osites,Nsimd,{
 #ifdef GRID_SIMT
@@ -353,7 +353,7 @@ public:
     typedef typename vobj::scalar_object coarse_scalar_object;
     //    std::cout << GridLogMessage<<" BlockProjector exporting "<<nvec<< " coarse grid vectors" <<std::endl;
 
-    assert(vecs[0].Grid()==coarse_grid);
+    GRID_ASSERT(vecs[0].Grid()==coarse_grid);
 
     int _ndimension = coarse_grid->_ndimension;
     
@@ -375,7 +375,7 @@ public:
       // loop over fine sites
       const int Nsimd = vobj::Nsimd();
       uint64_t cwords=sizeof(typename vobj::scalar_object)/sizeof(scalar);
-      assert(cwords==nbasis);
+      GRID_ASSERT(cwords==nbasis);
       
       accelerator_for(sc,osites,Nsimd,{
 	  // Wrap in a macro "FOR_ALL_LANES(lane,{ ... });
@@ -409,7 +409,7 @@ public:
     int nrhs=fine.size();
     int _nbasis = sizeof(typename cobj::scalar_object)/sizeof(scalar);
     //    std::cout << "blockProject nbasis " <<nbasis<<" " << _nbasis<<std::endl;
-    assert(nbasis==_nbasis);
+    GRID_ASSERT(nbasis==_nbasis);
     
     BLAS_F.resize (fine_vol * words * nrhs );
     BLAS_C.resize (coarse_vol * nbasis * nrhs );
@@ -447,10 +447,10 @@ public:
     /////////////////////////////////////////
     BLAS.gemmBatched(GridBLAS_OP_C,GridBLAS_OP_N, 
     		     nbasis,nrhs,vw,
-		     ComplexD(1.0),
+		     scalar(1.0),
 		     Vd,
 		     Fd,
-		     ComplexD(0.0),  // wipe out C
+		     scalar(0.0),  // wipe out C
 		     Cd);
     BLAS.synchronise();
     //    std::cout << "BlockProject done"<<std::endl;
@@ -464,7 +464,7 @@ public:
   {
     int nrhs=fine.size();
     int _nbasis = sizeof(typename cobj::scalar_object)/sizeof(scalar);
-    assert(nbasis==_nbasis);
+    GRID_ASSERT(nbasis==_nbasis);
     
     BLAS_F.resize (fine_vol * words * nrhs );
     BLAS_C.resize (coarse_vol * nbasis * nrhs );
@@ -497,10 +497,10 @@ public:
     int64_t vw = block_vol * words;
     BLAS.gemmBatched(GridBLAS_OP_N,GridBLAS_OP_N, 
     		     vw,nrhs,nbasis,
-		     ComplexD(1.0),
+		     scalar(1.0),
 		     Vd,
 		     Cd,
-		     ComplexD(0.0),  // wipe out C
+		     scalar(0.0),  // wipe out C
 		     Fd);
     BLAS.synchronise();
     //    std::cout << " blas call done"<<std::endl;

Grid/algorithms/deflation/MultiRHSDeflation.h

@@ -98,7 +98,7 @@ public:
   void ImportEigenVector(Field &evec,RealD &_eval, int ev)
   {
     //    std::cout << " ev " <<ev<<" eval "<<_eval<< std::endl;
-    assert(ev<eval.size());
+    GRID_ASSERT(ev<eval.size());
     eval[ev] = _eval;
 
     int64_t offset = ev*vol*words;
@@ -113,7 +113,7 @@ public:
   // Could use to import a batch of eigenvectors
   void ImportEigenBasis(std::vector<Field> &evec,std::vector<RealD> &_eval, int _ev0, int _nev)
   {
-    assert(_ev0+_nev<=evec.size());
+    GRID_ASSERT(_ev0+_nev<=evec.size());
 
     Allocate(_nev,evec[0].Grid());
     
@@ -126,8 +126,8 @@ public:
   void DeflateSources(std::vector<Field> &source,std::vector<Field> & guess)
   {
     int nrhs = source.size();
-    assert(source.size()==guess.size());
+    GRID_ASSERT(source.size()==guess.size());
-    assert(grid == guess[0].Grid());
+    GRID_ASSERT(grid == guess[0].Grid());
     conformable(guess[0],source[0]);
 
     int64_t vw = vol * words;
@@ -182,14 +182,14 @@ public:
     /////////////////////////////////////////
     BLAS.gemmBatched(GridBLAS_OP_C,GridBLAS_OP_N, 
     		     nev,nrhs,vw,
-		     ComplexD(1.0),
+		     scalar(1.0),
 		     Ed,
 		     Rd,
-		     ComplexD(0.0),  // wipe out C
+		     scalar(0.0),  // wipe out C
 		     Cd);
     BLAS.synchronise();
 
-    assert(BLAS_C.size()==nev*nrhs);
+    GRID_ASSERT(BLAS_C.size()==nev*nrhs);
 
     std::vector<scalar> HOST_C(BLAS_C.size());      // nrhs . nev -- the coefficients 
     acceleratorCopyFromDevice(&BLAS_C[0],&HOST_C[0],BLAS_C.size()*sizeof(scalar));
@@ -210,10 +210,10 @@ public:
     /////////////////////////////////////////
     BLAS.gemmBatched(GridBLAS_OP_N,GridBLAS_OP_N, 
 		     vw,nrhs,nev,
-		     ComplexD(1.0),
+		     scalar(1.0),
 		     Ed, // x . nev
 		     Cd, // nev . nrhs
-		     ComplexD(0.0),
+		     scalar(0.0),
 		     Gd);
     BLAS.synchronise();
 

Grid/algorithms/iterative/AdefGeneric.h

@@ -270,7 +270,7 @@ class TwoLevelCG : public LinearFunction<Field>
     std::vector<RealD> src_nrm(nrhs);
     for(int rhs=0;rhs<nrhs;rhs++) {
       src_nrm[rhs]=norm2(src[rhs]);
-      assert(src_nrm[rhs]!=0.0);
+      GRID_ASSERT(src_nrm[rhs]!=0.0);
     }
     std::vector<RealD> tn(nrhs);
 

Grid/algorithms/iterative/AdefMrhs.h

@@ -53,6 +53,7 @@ class TwoLevelCGmrhs
   // Fine operator, Smoother, CoarseSolver
   LinearOperatorBase<Field>   &_FineLinop;
   LinearFunction<Field>   &_Smoother;
+  MultiRHSBlockCGLinalg<Field> _BlockCGLinalg;
 
   GridStopWatch ProjectTimer;
   GridStopWatch PromoteTimer;
@@ -62,7 +63,12 @@ class TwoLevelCGmrhs
   GridStopWatch SmoothTimer;
   GridStopWatch InsertTimer;
 
-  
+  /*
+    Field rrr;
+  Field sss;
+  Field qqq;
+  Field zzz;
+  */  
   // more most opertor functions
   TwoLevelCGmrhs(RealD tol,
 		 Integer maxit,
@@ -73,12 +79,313 @@ class TwoLevelCGmrhs
     MaxIterations(maxit),
     _FineLinop(FineLinop),
     _Smoother(Smoother)
+    /*
+    rrr(fine),
+    sss(fine),
+    qqq(fine),
+    zzz(fine)
+*/
   {
     grid       = fine;
   };
   
   // 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]); GRID_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;
+    GRID_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();
@@ -108,7 +415,7 @@ class TwoLevelCGmrhs
     std::vector<RealD> src_nrm(nrhs);
     for(int rhs=0;rhs<nrhs;rhs++) {
       src_nrm[rhs]=norm2(src[rhs]);
-      assert(src_nrm[rhs]!=0.0);
+      GRID_ASSERT(src_nrm[rhs]!=0.0);
     }
     std::vector<RealD> tn(nrhs);
 
@@ -361,15 +668,26 @@ public:
     CoarseField PleftProjMrhs(this->coarsegridmrhs);
     CoarseField PleftMss_projMrhs(this->coarsegridmrhs);
 
-    for(int rhs=0;rhs<nrhs;rhs++) {
+    //    this->rrr=in[0];
 
+#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
+    //    this->sss=Min[0];
+    
+    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();
 
@@ -401,9 +719,11 @@ public:
     this->_Projector.blockPromote(tmp,PleftMss_proj);// tmp= Q[in - A Min]  
     this->PromoteTimer.Stop();
     this->FineTimer.Start();
+    //    this->qqq=tmp[0];
     for(int rhs=0;rhs<nrhs;rhs++) {
       axpy(out[rhs],1.0,Min[rhs],tmp[rhs]); // Min+tmp
     }
+    //    this->zzz=out[0];
     this->FineTimer.Stop();
   }
 };

Grid/algorithms/iterative/Arnoldi.h

@@ -0,0 +1,433 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid 
+
+Source file: ./lib/algorithms/iterative/Arnoldi.h
+
+Copyright (C) 2015
+
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+Author: paboyle <paboyle@ph.ed.ac.uk>
+Author: Patrick Oare <poare@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 */
+#ifndef GRID_ARNOLDI_H
+#define GRID_ARNOLDI_H
+
+NAMESPACE_BEGIN(Grid); 
+
+//Moved to KrylovSchur
+#if 0
+/**
+<<<<<<< HEAD
+ * Options for which Ritz values to keep in implicit restart.
+ */
+enum RitzFilter {
+  EvalNormSmall,           // Keep evals with smallest norm
+  EvalNormLarge,           // Keep evals with largest norm
+  EvalReSmall,             // Keep evals with smallest real part
+  EvalReLarge              // Keep evals with largest real part
+};
+
+// Select comparison function from RitzFilter
+struct ComplexComparator
+{
+  RitzFilter f;
+  ComplexComparator (RitzFilter _f) : f(_f) {}
+  bool operator()(std::complex<double> z1, std::complex<double> z2) { 
+    switch (f) {
+      RealD tmp1, tmp2;
+      tmp1=std::abs(std::imag(z1));
+      tmp2=std::abs(std::imag(z2));
+      case EvalNormSmall:
+        return std::abs(z1) < std::abs(z2);
+      case EvalNormLarge:
+        return std::abs(z1) > std::abs(z2);
+// Terrible hack
+//        return std::abs(std::real(z1)) < std::abs(std::real(z2));
+//	if ( std::abs(std::real(z1))  >4.) tmp1 +=1.;
+//	if ( std::abs(std::real(z2))  >4.) tmp2 +=1.;
+      case EvalReSmall:
+	  return tmp1 < tmp2;
+//        return std::abs(std::imag(z1)) < std::abs(std::imag(z2));
+      case EvalReLarge:
+	  return tmp1 > tmp2;
+//        return std::abs(std::real(z1)) > std::abs(std::real(z2));
+      default:
+        assert(0);
+    }
+  }
+};
+
+=======
+>>>>>>> 68af1bba67dd62881ead5ab1e54962a5486a0791
+#endif
+
+/**
+ * Implementation of the Arnoldi algorithm.
+ */
+template<class Field> 
+class Arnoldi {
+
+  private:
+  
+    std::string cname = std::string("Arnoldi");
+    int MaxIter;   // Max iterations
+    RealD Tolerance;
+    RealD ssq;
+    RealD rtol;
+    int Nm;           // Number of basis vectors to track (equals MaxIter if no restart)
+    int Nk;           // Number of basis vectors to keep every restart (equals -1 if no restart)
+    int Nstop;       // Stop after converging Nstop eigenvectors.
+
+    LinearOperatorBase<Field> &Linop;
+    GridBase *Grid;
+
+    RealD approxLambdaMax;
+    RealD beta_k;
+    Field f;
+    std::vector<Field> basis;               // orthonormal Arnoldi basis
+    Eigen::MatrixXcd Hess;                  // Hessenberg matrix of size Nbasis (after construction)
+    Eigen::MatrixXcd Qt;                    // Transpose of basis rotation which projects out high modes.
+
+    Eigen::VectorXcd evals;                 // evals of Hess
+    Eigen::MatrixXcd littleEvecs;           // Nm x Nm evecs matrix
+    std::vector<Field> evecs;               // Vector of evec fields
+
+    RitzFilter ritzFilter;                        // how to sort evals
+
+  public:       
+
+    Arnoldi(LinearOperatorBase<Field> &_Linop, GridBase *_Grid, RealD _Tolerance, RitzFilter filter = EvalReSmall)
+      : Linop(_Linop), Grid(_Grid), Tolerance(_Tolerance), ritzFilter(filter), f(_Grid), MaxIter(-1), Nm(-1), Nk(-1), 
+          Nstop (-1), evals (0), evecs (), ssq (0.0), rtol (0.0), beta_k (0.0), approxLambdaMax (0.0)
+    {
+      f = Zero();
+    };
+
+    /**
+     * Runs the Arnoldi loop with(out) implicit restarting. For each iteration:
+     *   - Runs an Arnoldi step.
+     *   - Computes the eigensystem of the Hessenberg matrix.
+     *   - Performs implicit restarting.
+     */
+    void operator()(const Field& v0, int _maxIter, int _Nm, int _Nk, int _Nstop, bool doubleOrthog = false) {
+      MaxIter = _maxIter;
+      Nm = _Nm; Nk = _Nk;
+      Nstop = _Nstop;
+      
+      ssq = norm2(v0);
+      RealD approxLambdaMax = approxMaxEval(v0);
+      rtol = Tolerance * approxLambdaMax;
+
+      ComplexComparator compareComplex (ritzFilter);
+      std::cout << GridLogMessage << "Comparing Ritz values with: " << ritzFilter << std::endl;
+
+      int start = 1;
+      Field startVec = v0;
+      littleEvecs = Eigen::MatrixXcd::Zero(Nm, Nm);
+      for (int i = 0; i < MaxIter; i++) {
+        std::cout << GridLogMessage << "Restart Iteration " << i << std::endl;
+
+        // Perform Arnoldi steps to compute Krylov basis and Rayleigh quotient (Hess)
+        arnoldiIteration(startVec, Nm, start, doubleOrthog);
+        startVec = f;
+
+        // compute eigensystem and sort evals
+        // compute_eigensystem();
+        compute_eigensystem(Hess);
+        std::cout << GridLogMessage << "Eigenvalues after Arnoldi step: " << std::endl << evals << std::endl;
+
+        std::sort(evals.begin(), evals.end(), compareComplex);
+        std::cout << GridLogMessage << "Ritz values after sorting (first Nk preserved): " << std::endl << evals << std::endl;
+        // SU(N)::tepidConfiguration
+
+        // Implicit restart to de-weight unwanted eigenvalues
+        implicitRestart(_Nm, _Nk);      // probably can delete _Nm and _Nk from function args
+        start = Nk;
+
+        // check convergence and return if needed.
+        int Nconv = converged();
+        std::cout << GridLogMessage << "Number of evecs converged: " << Nconv << std::endl;
+        if (Nconv >= Nstop || i == MaxIter - 1) {
+          std::cout << GridLogMessage << "Converged with " << Nconv << " / " << Nstop << " eigenvectors on iteration " 
+                        << i << "." << std::endl;
+          basisRotate(evecs, Qt, 0, Nk, 0, Nk, Nm);
+          std::cout << GridLogMessage << "Eigenvalues [first " << Nconv << " converged]: " << std::endl << evals << std::endl;
+          return;
+        }
+      }      
+    }
+
+    /**
+     * Approximates the maximum eigenvalue of Linop.Op to normalize the residual and test for convergence. 
+     * 
+     * Parameters
+     * ----------
+     * Field& v0
+     *  Source field to start with. Must have non-zero norm.
+     * int MAX_ITER (default = 50)
+     *  Maximum number of iterations for power approximation. 
+     * 
+     * Returns
+     * -------
+     * RealD lamApprox
+     *  Approximation of largest eigenvalue. 
+     */
+    RealD approxMaxEval(const Field& v0, int MAX_ITER = 50) {
+      assert (norm2(v0) > 1e-8);                        // must have relatively large source norm to start
+      RealD lamApprox = 0.0;
+      RealD denom = 1.0; RealD num = 1.0;
+      Field v0cp (Grid); Field tmp (Grid);
+      v0cp = v0;
+      denom = std::sqrt(norm2(v0cp));
+      for (int i = 0; i < MAX_ITER; i++) {
+        Linop.Op(v0cp, tmp);                               // CAREFUL: do not do Op(tmp, tmp)
+        v0cp = tmp;
+        num = std::sqrt(norm2(v0cp));                      // num = |A^{n+1} v0|
+        lamApprox = num / denom;                           // lam = |A^{n+1} v0| / |A^n v0|
+        std::cout << GridLogDebug << "Approx for max eval: " << lamApprox << std::endl;
+        denom = num;                                       // denom = |A^{n} v0|
+      }
+      return lamApprox;
+    }
+
+    /**
+     * Constructs the Arnoldi basis for the Krylov space K_n(D, src). (TODO make private)
+     * 
+     * Parameters
+     * ----------
+     * v0 : Field&
+     *  Source to generate Krylov basis. 
+     * Nm : int
+     *  Final size of the basis desired. If the basis becomes complete before a basis of size Nm is constructed 
+     *  (determined by relative tolerance Tolerance), stops iteration there. 
+     * doubleOrthog : bool (default = false)
+     *  Whether to double orthogonalize the basis (for numerical cancellations) or not. 
+     * start        : int (default = 0)
+     *  If non-zero, assumes part of the Arnoldi basis has already been constructed. 
+     */
+    void arnoldiIteration(const Field& v0, int Nm, int start = 1, bool doubleOrthog = false)
+    {
+
+      ComplexD coeff;
+      Field w (Grid);           // A acting on last Krylov vector. 
+
+      if (start == 1) {       // initialize everything that we need.
+        RealD v0Norm = 1 / std::sqrt(ssq);
+        basis.push_back(v0Norm * v0);                // normalized source
+
+        Hess = Eigen::MatrixXcd::Zero(Nm, Nm);
+        f = Zero();
+      } else {
+        assert( start == basis.size() );      // should be starting at the end of basis (start = Nk)
+        Eigen::MatrixXcd HessCp = Hess;
+        Hess = Eigen::MatrixXcd::Zero(Nm, Nm);
+        Hess(Eigen::seqN(0, Nk), Eigen::seqN(0, Nk)) = HessCp;
+      }
+
+      // Construct next Arnoldi vector by normalizing w_i = Dv_i - \sum_j v_j h_{ji}
+      for (int i = start - 1; i < Nm; i++) {
+
+        Linop.Op(basis.back(), w);
+        for (int j = 0; j < basis.size(); j++) {
+          coeff = innerProduct(basis[j], w);       // coeff = h_{ij}. Note that since {vi} is ONB it's OK to subtract it off after. 
+          Hess(j, i) = coeff;
+          w -= coeff * basis[j];
+        }
+
+        if (doubleOrthog) {
+          // TODO implement
+        }
+
+        // add w_i to the pile
+        if (i < Nm - 1) {
+          coeff = std::sqrt(norm2(w));
+          Hess(i+1, i) = coeff;
+          basis.push_back(
+            (1.0/coeff) * w
+          );
+        }
+
+        // after iterations, update f and beta_k = ||f||
+        f = w;                                // make sure f is not normalized
+        beta_k = std::sqrt(norm2(f));         // beta_k = ||f_k|| determines convergence.
+      }
+
+      std::cout << GridLogMessage << "|f|^2 after Arnoldi step = " << norm2(f) << std::endl;
+      std::cout << GridLogDebug << "Computed Hessenberg matrix = " << std::endl << Hess << std::endl;
+
+      return;
+    }
+
+    /**
+     * Approximates the eigensystem of the linear operator by computing the eigensystem of 
+     * the Hessenberg matrix. Assumes that the Hessenberg matrix has already been constructed (by 
+     * calling the operator() function).
+     * 
+     * TODO implement in parent class eventually.
+     * 
+     * Parameters
+     * ----------
+     * Eigen::MatrixXcd& S
+     *  Schur matrix (upper triangular) similar to original Rayleigh quotient.
+     */
+    void compute_eigensystem(Eigen::MatrixXcd& S)
+    {
+
+      std::cout << GridLogMessage << "Computing eigenvalues." << std::endl;
+
+      evecs.clear();
+
+      Eigen::ComplexEigenSolver<Eigen::MatrixXcd> es;
+      es.compute(S);
+      evals = es.eigenvalues();
+      littleEvecs = es.eigenvectors();
+
+      // Convert evecs to lattice fields
+      for (int k = 0; k < evals.size(); k++) {
+        Eigen::VectorXcd vec = littleEvecs.col(k);
+        Field tmp (basis[0].Grid());
+        tmp = Zero();
+        for (int j = 0; j < basis.size(); j++) {
+          tmp = tmp + vec[j] * basis[j];
+        }
+        evecs.push_back(tmp);
+      }
+
+      std::cout << GridLogMessage << "Eigenvalues: " << std::endl << evals << std::endl;
+
+    }
+
+    /**
+     * Verifies the factorization DV = V^\dag H + f e^\dag with the last-computed 
+     * V, H, f. 
+     */
+    // RealD verifyFactorization() {
+    //   int k = basis.size();         // number of basis vectors, also the size of H.
+    //   std::vector<Field> factorized (k, Zero());
+    //   Field tmp (FGrid); tmp = Zero();
+    //   for (int i = 0; i < basis.size(); i++) {
+    //     Linop.Op(basis[i], tmp);
+    //   }
+    //   // basisRotate(basis, Q, 0, Nk, 0, Nk, Nm);
+    //   // Linop.Op(, )
+    // }
+
+    /* Getters */
+    Eigen::MatrixXcd    getHessenbergMat()  { return Hess; }
+    Field               getF()              { return f; }
+    std::vector<Field>  getBasis()          { return basis; }
+    Eigen::VectorXcd    getEvals()          { return evals; }
+    std::vector<Field>  getEvecs()          { return evecs; }
+
+    /**
+     * Implements implicit restarting for Arnoldi. Assumes eigenvalues are sorted. 
+     * 
+     * Parameters
+     * ----------
+     * int _Nm
+     *  Size of basis to keep (Hessenberg is MxM).
+     * int Nk
+     *  Number of basis vectors to keep at each restart.
+     */
+    void implicitRestart(int _Nm, int _Nk) {
+      assert ( _Nk <= _Nm );
+      Nm = _Nm; Nk = _Nk;
+      int Np = Nm - Nk;       // keep Nk smallest (or largest, depends on sort function) evecs
+      
+      std::cout << GridLogMessage << "Computing QR Factorizations." << std::endl;
+
+      Eigen::MatrixXcd Q = Eigen::MatrixXcd::Identity(Nm, Nm);
+      Eigen::MatrixXcd Qi (Nm, Nm);
+      Eigen::MatrixXcd R (Nm, Nm);
+
+      for (int i = Nk; i < Nm; i++) {        // keep the first Nk eigenvalues and iterate through the last Np. Should loop Np times
+
+        // Useful debugging output
+        std::cout << GridLogDebug << "Computing QR factorization for i = " << i << std::endl;
+        std::cout << GridLogDebug << "Eval shift = " << evals[i] << std::endl;
+        std::cout << GridLogDebug << "Hess before rotation: " << Hess << std::endl;
+
+        // QR factorize 
+        Eigen::HouseholderQR<Eigen::MatrixXcd> QR (Hess - evals[i] * Eigen::MatrixXcd::Identity(Nm, Nm));
+        Qi = QR.householderQ();
+        Q = Q * Qi;
+        Hess = Qi.adjoint() * Hess * Qi;
+
+        std::cout << GridLogDebug << "Qt up to i = " << Q.transpose() << std::endl;
+
+      }
+
+      std::cout << GridLogDebug << "Hess after all rotations: " << std::endl << Hess << std::endl; 
+
+      // form Arnoldi vector f: f is normal to the basis vectors and its norm \beta is used to determine the Ritz estimate. 
+      std::complex<double> beta = Hess(Nk, Nk-1);
+      std::complex<double> sigma = Q(Nm-1, Nk-1);
+      f = basis[Nk] * beta + f * sigma;
+      RealD betak = std::sqrt(norm2(f));
+      std::cout << GridLogMessage << "|f|^2 after implicit restart = " << norm2(f) << std::endl;
+
+      // Rotate basis by Qt
+      Qt = Q.transpose();
+      basisRotate(basis, Qt, 0, Nk + 1, 0, Nm, Nm);
+
+      // rotate
+      basisRotate(evecs, Qt, 0, Nk + 1, 0, Nm, Nm);
+
+      // Truncate the basis and restart
+      basis = std::vector<Field> (basis.begin(), basis.begin() + Nk);
+      // evecs = std::vector<Field> (evecs.begin(), evecs.begin() + Nk);
+      Hess = Hess(Eigen::seqN(0, Nk), Eigen::seqN(0, Nk));
+
+      std::cout << "evecs size: " << evecs.size() << std::endl;
+
+    }
+  
+    /**
+     * Computes the number of Arnoldi eigenvectors that have converged. An eigenvector s is considered converged 
+     * for a tolerance epsilon if 
+     *    r(s) := |\beta e_m^T s| < epsilon
+     * where beta is the norm of f_{m+1}.
+     * 
+     * Parameters
+     * ----------
+     * 
+     * Returns
+     * -------
+     * int : Number of converged eigenvectors.
+     */
+    int converged() {
+      int Nconv = 0;
+      for (int k = 0; k < evecs.size(); k++) {
+        RealD emTs = std::abs(littleEvecs(Nm - 1, k));           // e_m^T s
+        RealD ritzEstimate = beta_k * emTs;
+        // TODO should be ritzEstimate < Tolerance * lambda_max
+        std::cout << GridLogMessage << "Ritz estimate for evec " << k << " = " << ritzEstimate << std::endl;
+        if (ritzEstimate < rtol) {
+          Nconv++;
+        }
+      }
+      return Nconv;
+    }
+
+};
+    
+NAMESPACE_END(Grid);
+#endif

Grid/algorithms/iterative/BiCGSTAB.h

@@ -47,7 +47,7 @@ class BiCGSTAB : public OperatorFunction<Field>
   public:
     using OperatorFunction<Field>::operator();
     
-    bool ErrorOnNoConverge;  // throw an assert when the CG fails to converge.
+    bool ErrorOnNoConverge;  // throw an GRID_ASSERT when the CG fails to converge.
                              // Defaults true.
     RealD Tolerance;
     Integer MaxIterations;
@@ -77,7 +77,7 @@ class BiCGSTAB : public OperatorFunction<Field>
 
       // Initial residual computation & set up
       RealD guess = norm2(psi);
-      assert(std::isnan(guess) == 0);
+      GRID_ASSERT(std::isnan(guess) == 0);
     
       Linop.Op(psi, v);
       b = norm2(v);
@@ -214,7 +214,7 @@ class BiCGSTAB : public OperatorFunction<Field>
           std::cout << GridLogMessage << "\tAxpyNorm   " << AxpyNormTimer.Elapsed() << std::endl;
           std::cout << GridLogMessage << "\tLinearComb " << LinearCombTimer.Elapsed() << std::endl;
 
-          if(ErrorOnNoConverge){ assert(true_residual / Tolerance < 10000.0); }
+          if(ErrorOnNoConverge){ GRID_ASSERT(true_residual / Tolerance < 10000.0); }
 
           IterationsToComplete = k;	
 
@@ -224,7 +224,7 @@ class BiCGSTAB : public OperatorFunction<Field>
       
       std::cout << GridLogMessage << "BiCGSTAB did NOT converge" << std::endl;
 
-      if(ErrorOnNoConverge){ assert(0); }
+      if(ErrorOnNoConverge){ GRID_ASSERT(0); }
       IterationsToComplete = k;
     }
 };

Grid/algorithms/iterative/BlockConjugateGradient.h

@@ -31,6 +31,58 @@ 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 };
 
 //////////////////////////////////////////////////////////////////////////
@@ -46,7 +98,7 @@ class BlockConjugateGradient : public OperatorFunction<Field> {
   int Nblock;
 
   BlockCGtype CGtype;
-  bool ErrorOnNoConverge;  // throw an assert when the CG fails to converge.
+  bool ErrorOnNoConverge;  // throw an GRID_ASSERT when the CG fails to converge.
                            // Defaults true.
   RealD Tolerance;
   Integer MaxIterations;
@@ -87,10 +139,19 @@ 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();
   ////////////////////////////////////////////////////////////////////////////////////////////////////
@@ -110,11 +171,20 @@ 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();
 
@@ -131,7 +201,7 @@ void operator()(LinearOperatorBase<Field> &Linop, const Field &Src, Field &Psi)
   } else if (CGtype == CGmultiRHS ) {
     CGmultiRHSsolve(Linop,Src,Psi);
   } else {
-    assert(0);
+    GRID_ASSERT(0);
   }
 }
 virtual void operator()(LinearOperatorBase<Field> &Linop, const std::vector<Field> &Src, std::vector<Field> &Psi) 
@@ -139,7 +209,7 @@ virtual void operator()(LinearOperatorBase<Field> &Linop, const std::vector<Fiel
   if ( CGtype == BlockCGrQVec ) {
     BlockCGrQsolveVec(Linop,Src,Psi);
   } else {
-    assert(0);
+    GRID_ASSERT(0);
   }
 }
 
@@ -186,12 +256,13 @@ 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); }
+  for(int b=0;b<Nblock;b++){ GRID_ASSERT(std::isnan(residuals[b])==0); }
 
   sliceNorm(residuals,X,Orthog);
-  for(int b=0;b<Nblock;b++){ assert(std::isnan(residuals[b])==0); }
+  for(int b=0;b<Nblock;b++){ GRID_ASSERT(std::isnan(residuals[b])==0); }
 
   /************************************************************************
    * Block conjugate gradient rQ (Sebastien Birk Thesis, after Dubrulle 2001)
@@ -221,6 +292,9 @@ 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;
 
@@ -236,6 +310,8 @@ 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++) {
 
@@ -280,7 +356,7 @@ void BlockCGrQsolve(LinearOperatorBase<Field> &Linop, const Field &B, Field &X)
      */
     m_rr = m_C.adjoint() * m_C;
 
-    RealD max_resid=0;
+    max_resid=0;
     RealD rrsum=0;
     RealD rr;
 
@@ -322,9 +398,11 @@ void BlockCGrQsolve(LinearOperatorBase<Field> &Linop, const Field &B, Field &X)
     }
 
   }
-  std::cout << GridLogMessage << "BlockConjugateGradient(rQ) did NOT converge" << std::endl;
 
-  if (ErrorOnNoConverge) assert(0);
+  std::cout << GridLogMessage << "BlockConjugateGradient(rQ) did NOT converge "<<k<<" / "<<MaxIterations
+	    <<" residual "<< std::sqrt(max_resid)<< std::endl;
+
+  if (ErrorOnNoConverge) GRID_ASSERT(0);
   IterationsToComplete = k;
 }
 //////////////////////////////////////////////////////////////////////////
@@ -360,10 +438,10 @@ void CGmultiRHSsolve(LinearOperatorBase<Field> &Linop, const Field &Src, Field &
   for(int b=0;b<Nblock;b++) sssum+=ssq[b];
 
   sliceNorm(residuals,Src,Orthog);
-  for(int b=0;b<Nblock;b++){ assert(std::isnan(residuals[b])==0); }
+  for(int b=0;b<Nblock;b++){ GRID_ASSERT(std::isnan(residuals[b])==0); }
 
   sliceNorm(residuals,Psi,Orthog);
-  for(int b=0;b<Nblock;b++){ assert(std::isnan(residuals[b])==0); }
+  for(int b=0;b<Nblock;b++){ GRID_ASSERT(std::isnan(residuals[b])==0); }
 
   // Initial search dir is guess
   Linop.HermOp(Psi, AP);
@@ -462,47 +540,10 @@ void CGmultiRHSsolve(LinearOperatorBase<Field> &Linop, const Field &Src, Field &
   }
   std::cout << GridLogMessage << "MultiRHSConjugateGradient did NOT converge" << std::endl;
 
-  if (ErrorOnNoConverge) assert(0);
+  if (ErrorOnNoConverge) GRID_ASSERT(0);
   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:
 //--------------------------
@@ -513,7 +554,7 @@ double normv(const std::vector<Field> &P){
 void BlockCGrQsolveVec(LinearOperatorBase<Field> &Linop, const std::vector<Field> &B, std::vector<Field> &X) 
 {
   Nblock = B.size();
-  assert(Nblock == X.size());
+  GRID_ASSERT(Nblock == X.size());
 
   std::cout<<GridLogMessage<<" Block Conjugate Gradient Vec rQ : Nblock "<<Nblock<<std::endl;
 
@@ -549,13 +590,14 @@ 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]);}
-  for(int b=0;b<Nblock;b++){ assert(std::isnan(residuals[b])==0); }
+  for(int b=0;b<Nblock;b++){ GRID_ASSERT(std::isnan(residuals[b])==0); }
 
   for(int b=0;b<Nblock;b++){ residuals[b] = norm2(X[b]);}
-  for(int b=0;b<Nblock;b++){ assert(std::isnan(residuals[b])==0); }
+  for(int b=0;b<Nblock;b++){ GRID_ASSERT(std::isnan(residuals[b])==0); }
 
   /************************************************************************
    * Block conjugate gradient rQ (Sebastien Birk Thesis, after Dubrulle 2001)
@@ -585,6 +627,7 @@ 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);
@@ -688,7 +731,7 @@ void BlockCGrQsolveVec(LinearOperatorBase<Field> &Linop, const std::vector<Field
   }
   std::cout << GridLogMessage << "BlockConjugateGradient(rQ) did NOT converge" << std::endl;
 
-  if (ErrorOnNoConverge) assert(0);
+  if (ErrorOnNoConverge) GRID_ASSERT(0);
   IterationsToComplete = k;
 }
 

Grid/algorithms/iterative/CommunicationAvoidingGeneralisedMinimalResidual.h

@@ -36,7 +36,7 @@ class CommunicationAvoidingGeneralisedMinimalResidual : public OperatorFunction<
  public:
   using OperatorFunction<Field>::operator();
 
-  bool ErrorOnNoConverge; // Throw an assert when CAGMRES fails to converge,
+  bool ErrorOnNoConverge; // Throw an GRID_ASSERT when CAGMRES fails to converge,
                           // defaults to true
 
   RealD   Tolerance;
@@ -82,7 +82,7 @@ class CommunicationAvoidingGeneralisedMinimalResidual : public OperatorFunction<
     conformable(psi, src);
 
     RealD guess = norm2(psi);
-    assert(std::isnan(guess) == 0);
+    GRID_ASSERT(std::isnan(guess) == 0);
 
     RealD cp;
     RealD ssq = norm2(src);
@@ -137,7 +137,7 @@ class CommunicationAvoidingGeneralisedMinimalResidual : public OperatorFunction<
     std::cout << GridLogMessage << "CommunicationAvoidingGeneralisedMinimalResidual did NOT converge" << std::endl;
 
     if (ErrorOnNoConverge)
-      assert(0);
+      GRID_ASSERT(0);
   }
 
   RealD outerLoopBody(LinearOperatorBase<Field> &LinOp, const Field &src, Field &psi, RealD rsq) {
@@ -185,7 +185,7 @@ class CommunicationAvoidingGeneralisedMinimalResidual : public OperatorFunction<
       }
     }
 
-    assert(0); // Never reached
+    GRID_ASSERT(0); // Never reached
     return cp;
   }
 

Grid/algorithms/iterative/ConjugateGradient.h

@@ -38,13 +38,14 @@ NAMESPACE_BEGIN(Grid);
 // single input vec, single output vec.
 /////////////////////////////////////////////////////////////
 
+
 template <class Field>
 class ConjugateGradient : public OperatorFunction<Field> {
 public:
 
   using OperatorFunction<Field>::operator();
   
-  bool ErrorOnNoConverge;  // throw an assert when the CG fails to converge.
+  bool ErrorOnNoConverge;  // throw an GRID_ASSERT when the CG fails to converge.
                            // Defaults true.
   RealD Tolerance;
   Integer MaxIterations;
@@ -57,10 +58,22 @@ 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();
 
@@ -70,14 +83,19 @@ 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);
-    assert(std::isnan(guess) == 0);
+    NormTimer.Stop();
+    GRID_ASSERT(std::isnan(guess) == 0);
+    AssignTimer.Start();
     if ( guess == 0.0 ) {
       r = src;
       p = r;
@@ -89,6 +107,7 @@ public:
       a = norm2(p);
     }
     cp = a;
+    AssignTimer.Stop();
 
     // Handle trivial case of zero src
     if (ssq == 0.){
@@ -164,6 +183,7 @@ public:
       }
       LinearCombTimer.Stop();
       LinalgTimer.Stop();
+      LogIteration(k,a,b);
 
       IterationTimer.Stop();
       if ( (k % 500) == 0 ) {
@@ -202,7 +222,7 @@ public:
 
 	std::cout << GridLogDebug << "\tMobius flop rate " << DwfFlops/ usecs<< " Gflops " <<std::endl;
 
-        if (ErrorOnNoConverge) assert(true_residual / Tolerance < 10000.0);
+        if (ErrorOnNoConverge) GRID_ASSERT(true_residual / Tolerance < 10000.0);
 
 	IterationsToComplete = k;	
 	TrueResidual = true_residual;
@@ -220,6 +240,9 @@ 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;
@@ -228,10 +251,123 @@ public:
     std::cout << GridLogPerformance << "\t\tAxpyNorm   " << AxpyNormTimer.Elapsed() <<std::endl;
     std::cout << GridLogPerformance << "\t\tLinearComb " << LinearCombTimer.Elapsed() <<std::endl;
 
-    if (ErrorOnNoConverge) assert(0);
+    if (ErrorOnNoConverge) GRID_ASSERT(0);
     IterationsToComplete = k;
 
   }
 };
+
+
+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/ConjugateGradientMixedPrec.h

@@ -116,14 +116,14 @@ NAMESPACE_BEGIN(Grid);
       //Compute double precision rsd and also new RHS vector.
       Linop_d.HermOp(sol_d, tmp_d);
       RealD norm = axpy_norm(src_d, -1., tmp_d, src_d_in); //src_d is residual vector
-      
+      std::cout<<GridLogMessage<<" rsd norm "<<norm<<std::endl;
       std::cout<<GridLogMessage<<"MixedPrecisionConjugateGradient: Outer iteration " <<outer_iter<<" residual "<< norm<< " target "<< stop<<std::endl;
 
       if(norm < OuterLoopNormMult * stop){
 	std::cout<<GridLogMessage<<"MixedPrecisionConjugateGradient: Outer iteration converged on iteration " <<outer_iter <<std::endl;
 	break;
       }
-      while(norm * inner_tol * inner_tol < stop) inner_tol *= 2;  // inner_tol = sqrt(stop/norm) ??
+      while(norm * inner_tol * inner_tol < stop*1.01) inner_tol *= 2;  // inner_tol = sqrt(stop/norm) ??
 
       PrecChangeTimer.Start();
       precisionChange(src_f, src_d, pc_wk_dp_to_sp);

Grid/algorithms/iterative/ConjugateGradientMixedPrecBatched.h

@@ -77,7 +77,7 @@ public:
   }
 
   void operator() (const std::vector<FieldD> &src_d_in, std::vector<FieldD> &sol_d){
-    assert(src_d_in.size() == sol_d.size());
+    GRID_ASSERT(src_d_in.size() == sol_d.size());
     int NBatch = src_d_in.size();
 
     std::cout << GridLogMessage << "NBatch = " << NBatch << std::endl;

Grid/algorithms/iterative/ConjugateGradientMultiShift.h

@@ -98,15 +98,15 @@ public:
     std::vector<RealD> alpha(nshift,1.0);
     std::vector<Field>   ps(nshift,grid);// Search directions
 
-    assert(psi.size()==nshift);
+    GRID_ASSERT(psi.size()==nshift);
-    assert(mass.size()==nshift);
+    GRID_ASSERT(mass.size()==nshift);
-    assert(mresidual.size()==nshift);
+    GRID_ASSERT(mresidual.size()==nshift);
   
-    // dynamic sized arrays on stack; 2d is a pain with vector
+    // remove dynamic sized arrays on stack; 2d is a pain with vector
-    RealD  bs[nshift];
+    std::vector<RealD>  bs(nshift);
-    RealD  rsq[nshift];
+    std::vector<RealD>  rsq(nshift);
-    RealD  z[nshift][2];
+    std::vector<std::array<RealD,2> >  z(nshift);
-    int     converged[nshift];
+    std::vector<int>     converged(nshift);
   
     const int       primary =0;
   
@@ -122,7 +122,7 @@ public:
   
     // Check lightest mass
     for(int s=0;s<nshift;s++){
-      assert( mass[s]>= mass[primary] );
+      GRID_ASSERT( mass[s]>= mass[primary] );
       converged[s]=0;
     }
   
@@ -338,7 +338,7 @@ public:
     }
     // ugly hack
     std::cout<<GridLogMessage<<"CG multi shift did not converge"<<std::endl;
-    //  assert(0);
+    //  GRID_ASSERT(0);
   }
 
 };

Grid/algorithms/iterative/ConjugateGradientMultiShiftCleanup.h

@@ -118,16 +118,16 @@ public:
     FieldF r_f(SinglePrecGrid);
     FieldD mmp_d(DoublePrecGrid);
 
-    assert(psi_d.size()==nshift);
+    GRID_ASSERT(psi_d.size()==nshift);
-    assert(mass.size()==nshift);
+    GRID_ASSERT(mass.size()==nshift);
-    assert(mresidual.size()==nshift);
+    GRID_ASSERT(mresidual.size()==nshift);
   
     // dynamic sized arrays on stack; 2d is a pain with vector
-    RealD  bs[nshift];
+    std::vector<RealD>  bs(nshift);
-    RealD  rsq[nshift];
+    std::vector<RealD>  rsq(nshift);
-    RealD  rsqf[nshift];
+    std::vector<RealD>  rsqf(nshift);
-    RealD  z[nshift][2];
+    std::vector<std::array<RealD,2> >  z(nshift);
-    int     converged[nshift];
+    std::vector<int>     converged(nshift);
   
     const int       primary =0;
   
@@ -141,7 +141,7 @@ public:
 
     // Check lightest mass
     for(int s=0;s<nshift;s++){
-      assert( mass[s]>= mass[primary] );
+      GRID_ASSERT( mass[s]>= mass[primary] );
       converged[s]=0;
     }
   
@@ -179,7 +179,7 @@ public:
     Linop_d.HermOpAndNorm(p_d,mmp_d,d,qq); // mmp = MdagM p        d=real(dot(p, mmp)),  qq=norm2(mmp)
     tmp_d = tmp_d - mmp_d;
     std::cout << " Testing operators match "<<norm2(mmp_d)<<" f "<<norm2(mmp_f)<<" diff "<< norm2(tmp_d)<<std::endl;
-    //    assert(norm2(tmp_d)< 1.0e-4);
+    //    GRID_ASSERT(norm2(tmp_d)< 1.0e-4);
 
     axpy(mmp_d,mass[0],p_d,mmp_d);
     RealD rn = norm2(p_d);
@@ -365,7 +365,7 @@ public:
    
     }
     std::cout<<GridLogMessage<<"CG multi shift did not converge"<<std::endl;
-    assert(0);
+    GRID_ASSERT(0);
   }
 
 };

Grid/algorithms/iterative/ConjugateGradientMultiShiftMixedPrec.h

@@ -48,12 +48,12 @@ public:
 
   ShiftedLinop(LinearOperatorBase<Field> &_linop_base, RealD _shift): linop_base(_linop_base), shift(_shift){}
 
-  void OpDiag (const Field &in, Field &out){ assert(0); }
+  void OpDiag (const Field &in, Field &out){ GRID_ASSERT(0); }
-  void OpDir  (const Field &in, Field &out,int dir,int disp){ assert(0); }
+  void OpDir  (const Field &in, Field &out,int dir,int disp){ GRID_ASSERT(0); }
-  void OpDirAll  (const Field &in, std::vector<Field> &out){ assert(0); }
+  void OpDirAll  (const Field &in, std::vector<Field> &out){ GRID_ASSERT(0); }
   
-  void Op     (const Field &in, Field &out){ assert(0); }
+  void Op     (const Field &in, Field &out){ GRID_ASSERT(0); }
-  void AdjOp  (const Field &in, Field &out){ assert(0); }
+  void AdjOp  (const Field &in, Field &out){ GRID_ASSERT(0); }
 
   void HermOp(const Field &in, Field &out){
     linop_base.HermOp(in, out);
@@ -151,16 +151,16 @@ public:
     FieldD r_d(DoublePrecGrid);
     FieldD mmp_d(DoublePrecGrid);
 
-    assert(psi_d.size()==nshift);
+    GRID_ASSERT(psi_d.size()==nshift);
-    assert(mass.size()==nshift);
+    GRID_ASSERT(mass.size()==nshift);
-    assert(mresidual.size()==nshift);
+    GRID_ASSERT(mresidual.size()==nshift);
   
     // dynamic sized arrays on stack; 2d is a pain with vector
-    RealD  bs[nshift];
+    std::vector<RealD>  bs(nshift);
-    RealD  rsq[nshift];
+    std::vector<RealD>  rsq(nshift);
-    RealD  rsqf[nshift];
+    std::vector<RealD>  rsqf(nshift);
-    RealD  z[nshift][2];
+    std::vector<std::array<RealD,2> >  z(nshift);
-    int     converged[nshift];
+    std::vector<int>     converged(nshift);
   
     const int       primary =0;
   
@@ -174,7 +174,7 @@ public:
 
     // Check lightest mass
     for(int s=0;s<nshift;s++){
-      assert( mass[s]>= mass[primary] );
+      GRID_ASSERT( mass[s]>= mass[primary] );
       converged[s]=0;
     }
   
@@ -211,7 +211,7 @@ public:
     Linop_d.HermOpAndNorm(p_d,mmp_d,d,qq); // mmp = MdagM p        d=real(dot(p, mmp)),  qq=norm2(mmp)
     tmp_d = tmp_d - mmp_d;
     std::cout << " Testing operators match "<<norm2(mmp_d)<<" f "<<norm2(mmp_f)<<" diff "<< norm2(tmp_d)<<std::endl;
-    assert(norm2(tmp_d)< 1.0);
+    GRID_ASSERT(norm2(tmp_d)< 1.0);
 
     axpy(mmp_d,mass[0],p_d,mmp_d);
     RealD rn = norm2(p_d);
@@ -408,7 +408,7 @@ public:
    
     }
     std::cout<<GridLogMessage<<"CG multi shift did not converge"<<std::endl;
-    assert(0);
+    GRID_ASSERT(0);
   }
 
 };

Grid/algorithms/iterative/ConjugateGradientReliableUpdate.h

@@ -35,7 +35,7 @@ template<class FieldD,class FieldF,
 	 typename std::enable_if< getPrecision<FieldF>::value == 1, int>::type = 0> 
 class ConjugateGradientReliableUpdate : public LinearFunction<FieldD> {
 public:
-  bool ErrorOnNoConverge;  // throw an assert when the CG fails to converge.
+  bool ErrorOnNoConverge;  // throw an GRID_ASSERT when the CG fails to converge.
   // Defaults true.
   RealD Tolerance;
   Integer MaxIterations;
@@ -66,7 +66,7 @@ public:
       DoFinalCleanup(true),
       Linop_fallback(NULL)
   {
-    assert(Delta > 0. && Delta < 1. && "Expect  0 < Delta < 1");
+    GRID_ASSERT(Delta > 0. && Delta < 1. && "Expect  0 < Delta < 1");
   };
 
   void setFallbackLinop(LinearOperatorBase<FieldF> &_Linop_fallback, const RealD _fallback_transition_tol){
@@ -90,7 +90,7 @@ public:
 
     // Initial residual computation & set up
     RealD guess = norm2(psi);
-    assert(std::isnan(guess) == 0);
+    GRID_ASSERT(std::isnan(guess) == 0);
     
     Linop_d.HermOpAndNorm(psi, mmp, d, b);
     
@@ -217,7 +217,7 @@ public:
 	  CG(Linop_d,src,psi);
 	  IterationsToCleanup = CG.IterationsToComplete;
 	}
-	else if (ErrorOnNoConverge) assert(true_residual / Tolerance < 10000.0);
+	else if (ErrorOnNoConverge) GRID_ASSERT(true_residual / Tolerance < 10000.0);
 
 	std::cout << GridLogMessage << "ConjugateGradientReliableUpdate complete.\n";
 	return;
@@ -263,7 +263,7 @@ public:
     std::cout << GridLogMessage << "ConjugateGradientReliableUpdate did NOT converge"
 	      << std::endl;
       
-    if (ErrorOnNoConverge) assert(0);
+    if (ErrorOnNoConverge) GRID_ASSERT(0);
     IterationsToComplete = k;
     ReliableUpdatesPerformed = l;      
   }    

Grid/algorithms/iterative/ConjugateGradientTimeslice.h

@@ -0,0 +1,277 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./lib/algorithms/iterative/ConjugateGradientTimeslice.h
+
+Copyright (C) 2015
+
+Author: Azusa Yamaguchi <ayamaguc@staffmail.ed.ac.uk>
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+Author: paboyle <paboyle@ph.ed.ac.uk>
+
+This program is free software; you can redistribute it and/or modify
+it under the terms of the GNU General Public License as published by
+the Free Software Foundation; either version 2 of the License, or
+(at your option) any later version.
+
+This program is distributed in the hope that it will be useful,
+but WITHOUT ANY WARRANTY; without even the implied warranty of
+MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+GNU General Public License for more details.
+
+You should have received a copy of the GNU General Public License along
+with this program; if not, write to the Free Software Foundation, Inc.,
+51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+See the full license in the file "LICENSE" in the top level distribution
+directory
+*************************************************************************************/
+			   /*  END LEGAL */
+#ifndef GRID_CONJUGATE_GRADIENT_TIMESLICE_H
+#define GRID_CONJUGATE_GRADIENT_TIMESLICE_H
+
+NAMESPACE_BEGIN(Grid);
+
+/////////////////////////////////////////////////////////////
+// Base classes for iterative processes based on operators
+// single input vec, single output vec.
+/////////////////////////////////////////////////////////////
+
+/**
+ * Simple modification of conjugate gradient that outputs the residual as a function 
+ * of time, in order to study the large wavelength behavior of the solver. 
+ */
+
+
+template <class Field>
+class ConjugateGradientTimeslice : public OperatorFunction<Field> {
+public:
+
+  using OperatorFunction<Field>::operator();
+  
+  bool ErrorOnNoConverge;  // throw an assert when the CG fails to converge.
+                           // Defaults true.
+  RealD Tolerance;
+  Integer MaxIterations;
+  Integer IterationsToComplete; //Number of iterations the CG took to finish. Filled in upon completion
+  RealD TrueResidual;
+  
+  ConjugateGradientTimeslice(RealD tol, Integer maxit, bool err_on_no_conv = true)
+    : Tolerance(tol),
+      MaxIterations(maxit),
+      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("ConjugateGradientTimeslice");
+    GridStopWatch PreambleTimer;
+    GridStopWatch ConstructTimer;
+    GridStopWatch NormTimer;
+    GridStopWatch AssignTimer;
+    PreambleTimer.Start();
+    psi.Checkerboard() = src.Checkerboard();
+
+    conformable(psi, src);
+
+    RealD cp, c, a, d, b, ssq, qq;
+    //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);                 // Norm of source vector ||b||^2
+
+    ssqtx = localNorm2(src);          // Norm |b(x, t)|^2 as a field
+    std::vector<RealD> ssqt;          // Norm of source not summed over time slices, ssq(t) = \sum_x |b(x, t)|^2
+    sliceSum(ssqtx, ssqt, Tdir);      // TODO make sure Tdir is globally defined
+
+    RealD guess = norm2(psi);         // Norm of initial guess ||psi||^2
+    NormTimer.Stop();
+    assert(std::isnan(guess) == 0);
+    AssignTimer.Start();
+    if ( guess == 0.0 ) {
+      r = src;
+      p = r;
+      a = ssq;
+    } else { 
+      Linop.HermOpAndNorm(psi, mmp, d, b);        // 
+      r = src - mmp;      // Initial residual r0 = b - A guess
+      p = r;              // initial conj vector p0 = r0
+      a = norm2(p);
+    }
+    cp = a;
+    AssignTimer.Stop();
+
+    // Handle trivial case of zero src
+    if (ssq == 0.){
+      psi = Zero();
+      IterationsToComplete = 1;
+      TrueResidual = 0.;
+      return;
+    }
+
+    std::cout << GridLogIterative << std::setprecision(8) << "ConjugateGradient: guess " << guess << std::endl;
+    std::cout << GridLogIterative << std::setprecision(8) << "ConjugateGradient:   src " << ssq << std::endl;
+    std::cout << GridLogIterative << std::setprecision(8) << "ConjugateGradient:    mp " << d << std::endl;
+    std::cout << GridLogIterative << std::setprecision(8) << "ConjugateGradient:   mmp " << b << std::endl;
+    std::cout << GridLogIterative << std::setprecision(8) << "ConjugateGradient:  cp,r " << cp << std::endl;
+    std::cout << GridLogIterative << std::setprecision(8) << "ConjugateGradient:     p " << a << std::endl;
+
+    RealD rsq = Tolerance * Tolerance * ssq;
+
+    // Check if guess is really REALLY good :)
+    if (cp <= rsq) {
+      TrueResidual = std::sqrt(a/ssq);
+      std::cout << GridLogMessage << "ConjugateGradient guess is converged already " << std::endl;
+      IterationsToComplete = 0;	
+      return;
+    }
+
+    std::cout << GridLogIterative << std::setprecision(8)
+              << "ConjugateGradient: k=0 residual " << cp << " target " << rsq << std::endl;
+
+    PreambleTimer.Stop();
+    GridStopWatch LinalgTimer;
+    GridStopWatch InnerTimer;
+    GridStopWatch AxpyNormTimer;
+    GridStopWatch LinearCombTimer;
+    GridStopWatch MatrixTimer;
+    GridStopWatch SolverTimer;
+
+    RealD usecs = -usecond();
+    SolverTimer.Start();
+    int k;
+    for (k = 1; k <= MaxIterations; k++) {
+
+      GridStopWatch IterationTimer;
+      IterationTimer.Start();
+      c = cp;
+
+      MatrixTimer.Start();
+      Linop.HermOp(p, mmp);         // Computes mmp = Ap
+      MatrixTimer.Stop();
+
+      LinalgTimer.Start();
+
+      InnerTimer.Start();
+      ComplexD dc  = innerProduct(p,mmp);         // p^\dagger A p
+      InnerTimer.Stop();
+      d = dc.real();
+      a = c / d;
+
+      // What is axpy? Some accelerator or something? Check Lattice_arith.h
+      AxpyNormTimer.Start();
+
+      // axpy_norm computes ax+by for vectors x and y compatible with a GPU. Here b is set to 1 (see the function in Lattice_reduction.h). 
+      // The first argument passes r by reference, so it stores r --> -a * Ap + 1 * r, i.e. it performs an update on 
+      // r_k --> r_{k+1} = r_k - \alpha_k A p_k. The function returns the norm squared of the first variable, i.e. ||r_{k+1}||^2.
+      cp = axpy_norm(r, -a, mmp, r);
+      AxpyNormTimer.Stop();
+      b = cp / c;
+
+      LinearCombTimer.Start();
+      {
+        autoView( psi_v , psi, AcceleratorWrite);
+        autoView( p_v   , p,   AcceleratorWrite);
+        autoView( r_v   , r,   AcceleratorWrite);
+        accelerator_for(ss,p_v.size(), Field::vector_object::Nsimd(),{
+            coalescedWrite(psi_v[ss], a      *  p_v(ss) + psi_v(ss));
+            coalescedWrite(p_v[ss]  , b      *  p_v(ss) + r_v  (ss));
+        });
+      }
+      LinearCombTimer.Stop();
+      LinalgTimer.Stop();
+      LogIteration(k,a,b);
+
+      IterationTimer.Stop();
+      if ( (k % 500) == 0 ) {
+        std::cout << GridLogMessage << "ConjugateGradient: Iteration " << k
+                << " residual " << sqrt(cp/ssq) << " target " << Tolerance << std::endl;
+      } else { 
+        std::cout << GridLogIterative << "ConjugateGradient: Iteration " << k
+                << " residual " << sqrt(cp/ssq) << " target " << Tolerance << " took " << IterationTimer.Elapsed() << std::endl;
+      }
+
+      // Stopping condition
+      if (cp <= rsq) {
+        usecs +=usecond();
+        SolverTimer.Stop();
+        Linop.HermOpAndNorm(psi, mmp, d, qq);
+        p = mmp - src;
+        GridBase *grid = src.Grid();
+        RealD DwfFlops = (1452. )*grid->gSites()*4*k
+   	               + (8+4+8+4+4)*12*grid->gSites()*k; // CG linear algebra
+        RealD srcnorm = std::sqrt(norm2(src));
+        RealD resnorm = std::sqrt(norm2(p));
+        RealD true_residual = resnorm / srcnorm;
+        std::cout << GridLogMessage << "ConjugateGradient Converged on iteration " << k 
+          << "\tComputed residual " << std::sqrt(cp / ssq)
+          << "\tTrue residual " << true_residual
+          << "\tTarget " << Tolerance << std::endl;
+
+        // GridLogMessage logs the message to the terminal output; GridLogPerformance probably writes to a log file?
+        //	std::cout << GridLogMessage << "\tPreamble   " << PreambleTimer.Elapsed() <<std::endl;
+        std::cout << GridLogMessage << "\tSolver Elapsed    " << SolverTimer.Elapsed() <<std::endl;
+        std::cout << GridLogPerformance << "Time breakdown "<<std::endl;
+        std::cout << GridLogPerformance << "\tMatrix     " << MatrixTimer.Elapsed() <<std::endl;
+        std::cout << GridLogPerformance << "\tLinalg     " << LinalgTimer.Elapsed() <<std::endl;
+        std::cout << GridLogPerformance << "\t\tInner      " << InnerTimer.Elapsed() <<std::endl;
+        std::cout << GridLogPerformance << "\t\tAxpyNorm   " << AxpyNormTimer.Elapsed() <<std::endl;
+        std::cout << GridLogPerformance << "\t\tLinearComb " << LinearCombTimer.Elapsed() <<std::endl;
+
+        std::cout << GridLogDebug << "\tMobius flop rate " << DwfFlops/ usecs<< " Gflops " <<std::endl;
+
+        if (ErrorOnNoConverge) assert(true_residual / Tolerance < 10000.0);
+
+        IterationsToComplete = k;	
+        TrueResidual = true_residual;
+
+        return;
+      }
+    }
+    // Failed. Calculate true residual before giving up                                                         
+    // Linop.HermOpAndNorm(psi, mmp, d, qq);
+    //    p = mmp - src;
+    //TrueResidual = sqrt(norm2(p)/ssq);
+    //    TrueResidual = 1;
+
+    std::cout << GridLogMessage << "ConjugateGradient did NOT converge "<<k<<" / "<< MaxIterations
+    	      <<" 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;
+    std::cout << GridLogMessage<< "\tLinalg     " << LinalgTimer.Elapsed() <<std::endl;
+    std::cout << GridLogPerformance << "\t\tInner      " << InnerTimer.Elapsed() <<std::endl;
+    std::cout << GridLogPerformance << "\t\tAxpyNorm   " << AxpyNormTimer.Elapsed() <<std::endl;
+    std::cout << GridLogPerformance << "\t\tLinearComb " << LinearCombTimer.Elapsed() <<std::endl;
+
+    if (ErrorOnNoConverge) assert(0);
+    IterationsToComplete = k;
+
+  }
+};
+
+NAMESPACE_END(Grid);
+#endif

Grid/algorithms/iterative/ConjugateResidual.h

@@ -106,7 +106,7 @@ public:
     }
 
     std::cout<<GridLogMessage<<"ConjugateResidual did NOT converge"<<std::endl;
-    assert(0);
+    GRID_ASSERT(0);
   }
 };
 NAMESPACE_END(Grid);

Grid/algorithms/iterative/FlexibleCommunicationAvoidingGeneralisedMinimalResidual.h

@@ -36,7 +36,7 @@ class FlexibleCommunicationAvoidingGeneralisedMinimalResidual : public OperatorF
  public:
   using OperatorFunction<Field>::operator();
 
-  bool ErrorOnNoConverge; // Throw an assert when FCAGMRES fails to converge,
+  bool ErrorOnNoConverge; // Throw an GRID_ASSERT when FCAGMRES fails to converge,
                           // defaults to true
 
   RealD   Tolerance;
@@ -87,7 +87,7 @@ class FlexibleCommunicationAvoidingGeneralisedMinimalResidual : public OperatorF
     conformable(psi, src);
 
     RealD guess = norm2(psi);
-    assert(std::isnan(guess) == 0);
+    GRID_ASSERT(std::isnan(guess) == 0);
 
     RealD cp;
     RealD ssq = norm2(src);
@@ -144,7 +144,7 @@ class FlexibleCommunicationAvoidingGeneralisedMinimalResidual : public OperatorF
     std::cout << GridLogMessage << "FlexibleCommunicationAvoidingGeneralisedMinimalResidual did NOT converge" << std::endl;
 
     if (ErrorOnNoConverge)
-      assert(0);
+      GRID_ASSERT(0);
   }
 
   RealD outerLoopBody(LinearOperatorBase<Field> &LinOp, const Field &src, Field &psi, RealD rsq) {
@@ -191,7 +191,7 @@ class FlexibleCommunicationAvoidingGeneralisedMinimalResidual : public OperatorF
       }
     }
 
-    assert(0); // Never reached
+    GRID_ASSERT(0); // Never reached
     return cp;
   }
 

Grid/algorithms/iterative/FlexibleGeneralisedMinimalResidual.h

@@ -36,7 +36,7 @@ class FlexibleGeneralisedMinimalResidual : public OperatorFunction<Field> {
  public:
   using OperatorFunction<Field>::operator();
 
-  bool ErrorOnNoConverge; // Throw an assert when FGMRES fails to converge,
+  bool ErrorOnNoConverge; // Throw an GRID_ASSERT when FGMRES fails to converge,
                           // defaults to true
 
   RealD   Tolerance;
@@ -85,7 +85,7 @@ class FlexibleGeneralisedMinimalResidual : public OperatorFunction<Field> {
     conformable(psi, src);
 
     RealD guess = norm2(psi);
-    assert(std::isnan(guess) == 0);
+    GRID_ASSERT(std::isnan(guess) == 0);
 
     RealD cp;
     RealD ssq = norm2(src);
@@ -142,7 +142,7 @@ class FlexibleGeneralisedMinimalResidual : public OperatorFunction<Field> {
     std::cout << GridLogMessage << "FlexibleGeneralisedMinimalResidual did NOT converge" << std::endl;
 
     if (ErrorOnNoConverge)
-      assert(0);
+      GRID_ASSERT(0);
   }
 
   RealD outerLoopBody(LinearOperatorBase<Field> &LinOp, const Field &src, Field &psi, RealD rsq) {
@@ -189,7 +189,7 @@ class FlexibleGeneralisedMinimalResidual : public OperatorFunction<Field> {
       }
     }
 
-    assert(0); // Never reached
+    GRID_ASSERT(0); // Never reached
     return cp;
   }
 

Grid/algorithms/iterative/GeneralisedMinimalResidual.h

@@ -36,7 +36,7 @@ class GeneralisedMinimalResidual : public OperatorFunction<Field> {
  public:
   using OperatorFunction<Field>::operator();
 
-  bool ErrorOnNoConverge; // Throw an assert when GMRES fails to converge,
+  bool ErrorOnNoConverge; // Throw an GRID_ASSERT when GMRES fails to converge,
                           // defaults to true
 
   RealD   Tolerance;
@@ -80,7 +80,7 @@ class GeneralisedMinimalResidual : public OperatorFunction<Field> {
     conformable(psi, src);
 
     RealD guess = norm2(psi);
-    assert(std::isnan(guess) == 0);
+    GRID_ASSERT(std::isnan(guess) == 0);
 
     RealD cp;
     RealD ssq = norm2(src);
@@ -135,7 +135,7 @@ class GeneralisedMinimalResidual : public OperatorFunction<Field> {
     std::cout << GridLogMessage << "GeneralisedMinimalResidual did NOT converge" << std::endl;
 
     if (ErrorOnNoConverge)
-      assert(0);
+      GRID_ASSERT(0);
   }
 
   RealD outerLoopBody(LinearOperatorBase<Field> &LinOp, const Field &src, Field &psi, RealD rsq) {
@@ -181,7 +181,7 @@ class GeneralisedMinimalResidual : public OperatorFunction<Field> {
       }
     }
 
-    assert(0); // Never reached
+    GRID_ASSERT(0); // Never reached
     return cp;
   }
 

Grid/algorithms/iterative/ImplicitlyRestartedBlockLanczos.h

@@ -175,7 +175,7 @@ public:
       eresid(_eresid),  MaxIter(_MaxIter),
       diagonalisation(_diagonalisation),split_test(0),
       Nevec_acc(_Nu)
-  { assert( (Nk%Nu==0) && (Nm%Nu==0) ); };
+  { GRID_ASSERT( (Nk%Nu==0) && (Nm%Nu==0) ); };
 
   ////////////////////////////////
   // Helpers
@@ -206,7 +206,7 @@ public:
           Glog<<"orthogonalize after: "<<j<<" of "<<k<<" "<< ip <<std::endl;
       }
     }
-    assert(normalize(w,if_print) != 0);
+    GRID_ASSERT(normalize(w,if_print) != 0);
   }
   void reorthogonalize(Field& w, std::vector<Field>& evec, int k)
   {
@@ -225,7 +225,7 @@ public:
       w[i] = w[i] - ip * evec[j];
     }}
     for(int i=0; i<_Nu; ++i)
-    assert(normalize(w[i],if_print) !=0);
+    GRID_ASSERT(normalize(w[i],if_print) !=0);
   }
 
 
@@ -244,7 +244,7 @@ public:
     const uint64_t sites = grid->lSites();
 
     int Nbatch = R/Nevec_acc;
-    assert( R%Nevec_acc == 0 );
+    GRID_ASSERT( R%Nevec_acc == 0 );
 //    Glog << "nBatch, Nevec_acc, R, Nu = " 
 //         << Nbatch << "," << Nevec_acc << "," << R << "," << Nu << std::endl;
     
@@ -302,7 +302,7 @@ public:
       }
     }
     for (int i=0; i<Nu; ++i) {
-      assert(normalize(w[i],do_print)!=0);
+      GRID_ASSERT(normalize(w[i],do_print)!=0);
     }
     
     Glog << "cuBLAS Zgemm done"<< std::endl;
@@ -374,8 +374,8 @@ cudaStat = cudaMallocManaged((void **)&evec_acc, Nevec_acc*sites*12*sizeof(CUDA_
   {
     std::string fname = std::string(cname+"::calc_irbl()"); 
     GridBase *grid = evec[0].Grid();
-    assert(grid == src[0].Grid());
+    GRID_ASSERT(grid == src[0].Grid());
-    assert( Nu = src.size() );
+    GRID_ASSERT( Nu = src.size() );
     
     Glog << std::string(74,'*') << std::endl;
     Glog << fname + " starting iteration 0 /  "<< MaxIter<< std::endl;
@@ -396,7 +396,7 @@ cudaStat = cudaMallocManaged((void **)&evec_acc, Nevec_acc*sites*12*sizeof(CUDA_
     }
     Glog << std::string(74,'*') << std::endl;
     
-    assert(Nm == evec.size() && Nm == eval.size());
+    GRID_ASSERT(Nm == evec.size() && Nm == eval.size());
 
     std::vector<std::vector<ComplexD>> lmd(Nu,std::vector<ComplexD>(Nm,0.0));  
     std::vector<std::vector<ComplexD>> lme(Nu,std::vector<ComplexD>(Nm,0.0));  
@@ -579,8 +579,8 @@ cudaStat = cudaMallocManaged((void **)&evec_acc, Nevec_acc*sites*12*sizeof(CUDA_
   {
     std::string fname = std::string(cname+"::calc_rbl()"); 
     GridBase *grid = evec[0].Grid();
-    assert(grid == src[0].Grid());
+    GRID_ASSERT(grid == src[0].Grid());
-    assert( Nu = src.size() );
+    GRID_ASSERT( Nu = src.size() );
 
     int Np = (Nm-Nk);
     if (Np > 0 && MaxIter > 1) Np /= MaxIter;
@@ -607,7 +607,7 @@ cudaStat = cudaMallocManaged((void **)&evec_acc, Nevec_acc*sites*12*sizeof(CUDA_
     }
     Glog << std::string(74,'*') << std::endl;
     
-    assert(Nm == evec.size() && Nm == eval.size());
+    GRID_ASSERT(Nm == evec.size() && Nm == eval.size());
 	
     std::vector<std::vector<ComplexD>> lmd(Nu,std::vector<ComplexD>(Nm,0.0));  
     std::vector<std::vector<ComplexD>> lme(Nu,std::vector<ComplexD>(Nm,0.0));  
@@ -785,7 +785,7 @@ private:
     
     int Nu = w.size();
     int Nm = evec.size();
-    assert( b < Nm/Nu );
+    GRID_ASSERT( b < Nm/Nu );
 //    GridCartesian *grid = evec[0]._grid;
     
     // converts block index to full indicies for an interval [L,R)
@@ -796,7 +796,7 @@ private:
 
     Glog << "Using split grid"<< std::endl;
 //   LatticeGaugeField s_Umu(SGrid);
-   assert((Nu%mrhs)==0);
+   GRID_ASSERT((Nu%mrhs)==0);
    std::vector<Field>   in(mrhs,f_grid);
      
     Field s_in(sf_grid);
@@ -906,7 +906,7 @@ if(split_test){
     
     for (int u=0; u<Nu; ++u) {
 //      Glog << "norm2(w[" << u << "])= "<< norm2(w[u]) << std::endl;
-      assert (!isnan(norm2(w[u])));
+      GRID_ASSERT (!isnan(norm2(w[u])));
       for (int k=L+u; k<R; ++k) {
         Glog <<" In block "<< b << "," <<" beta[" << u << "," << k-L << "] = " << lme[u][k] << std::endl;
       }
@@ -929,8 +929,8 @@ if(split_test){
 			 Eigen::MatrixXcd & Qt, // Nm x Nm
 			 GridBase *grid)
   {
-    assert( Nk%Nu == 0 && Nm%Nu == 0 );
+    GRID_ASSERT( Nk%Nu == 0 && Nm%Nu == 0 );
-    assert( Nk <= Nm );
+    GRID_ASSERT( Nk <= Nm );
     Eigen::MatrixXcd BlockTriDiag = Eigen::MatrixXcd::Zero(Nk,Nk);
     
     for ( int u=0; u<Nu; ++u ) {
@@ -970,8 +970,8 @@ if(split_test){
 			 GridBase *grid)
   {
     Glog << "diagonalize_lapack: Nu= "<<Nu<<" Nk= "<<Nk<<" Nm= "<<std::endl;
-    assert( Nk%Nu == 0 && Nm%Nu == 0 );
+    GRID_ASSERT( Nk%Nu == 0 && Nm%Nu == 0 );
-    assert( Nk <= Nm );
+    GRID_ASSERT( Nk <= Nm );
     Eigen::MatrixXcd BlockTriDiag = Eigen::MatrixXcd::Zero(Nk,Nk);
     
     for ( int u=0; u<Nu; ++u ) {
@@ -1119,7 +1119,7 @@ if (1){
       diagonalize_lapack(eval,lmd,lme,Nu,Nk,Nm,Qt,grid);
 #endif
     } else { 
-      assert(0);
+      GRID_ASSERT(0);
     }
   }
   
@@ -1131,8 +1131,8 @@ if (1){
          Eigen::MatrixXcd& M)
   {
     //Glog << "unpackHermitBlockTriDiagMatToEigen() begin" << '\n'; 
-    assert( Nk%Nu == 0 && Nm%Nu == 0 );
+    GRID_ASSERT( Nk%Nu == 0 && Nm%Nu == 0 );
-    assert( Nk <= Nm );
+    GRID_ASSERT( Nk <= Nm );
     M = Eigen::MatrixXcd::Zero(Nk,Nk);
     
     // rearrange 
@@ -1159,8 +1159,8 @@ if (1){
          Eigen::MatrixXcd& M)
   {
     //Glog << "packHermitBlockTriDiagMatfromEigen() begin" << '\n'; 
-    assert( Nk%Nu == 0 && Nm%Nu == 0 );
+    GRID_ASSERT( Nk%Nu == 0 && Nm%Nu == 0 );
-    assert( Nk <= Nm );
+    GRID_ASSERT( Nk <= Nm );
     
     // rearrange 
     for ( int u=0; u<Nu; ++u ) {

Grid/algorithms/iterative/ImplicitlyRestartedBlockLanczosCoarse.h

@@ -121,7 +121,7 @@ public:
       eresid(_eresid),  MaxIter(_MaxIter),
       diagonalisation(_diagonalisation),
       Nevec_acc(_Nu)
-  { assert( (Nk%Nu==0) && (Nm%Nu==0) ); };
+  { GRID_ASSERT( (Nk%Nu==0) && (Nm%Nu==0) ); };
 
   ////////////////////////////////
   // Helpers
@@ -151,7 +151,7 @@ public:
           Glog<<"orthogonalize after: "<<j<<" of "<<k<<" "<< ip <<std::endl;
       }
     }
-    assert(normalize(w,if_print) != 0);
+    GRID_ASSERT(normalize(w,if_print) != 0);
   }
   void reorthogonalize(Field& w, std::vector<Field>& evec, int k)
   {
@@ -169,7 +169,7 @@ public:
       w[i] = w[i] - ip * evec[j];
     }}
     for(int i=0; i<_Nu; ++i)
-    assert(normalize(w[i],if_print) !=0);
+    GRID_ASSERT(normalize(w[i],if_print) !=0);
   }
   
   void orthogonalize_blockhead(Field& w, std::vector<Field>& evec, int k, int Nu)
@@ -205,8 +205,8 @@ public:
   {
     std::string fname = std::string(cname+"::calc_irbl()"); 
     GridBase *grid = evec[0].Grid();
-    assert(grid == src[0].Grid());
+    GRID_ASSERT(grid == src[0].Grid());
-    assert( Nu = src.size() );
+    GRID_ASSERT( Nu = src.size() );
     
     Glog << std::string(74,'*') << std::endl;
     Glog << fname + " starting iteration 0 /  "<< MaxIter<< std::endl;
@@ -227,7 +227,7 @@ public:
     }
     Glog << std::string(74,'*') << std::endl;
     
-    assert(Nm == evec.size() && Nm == eval.size());
+    GRID_ASSERT(Nm == evec.size() && Nm == eval.size());
 
     std::vector<std::vector<ComplexD>> lmd(Nu,std::vector<ComplexD>(Nm,0.0));  
     std::vector<std::vector<ComplexD>> lme(Nu,std::vector<ComplexD>(Nm,0.0));  
@@ -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<<"] ";
@@ -413,8 +413,8 @@ public:
   {
     std::string fname = std::string(cname+"::calc_rbl()"); 
     GridBase *grid = evec[0].Grid();
-    assert(grid == src[0].Grid());
+    GRID_ASSERT(grid == src[0].Grid());
-    assert( Nu = src.size() );
+    GRID_ASSERT( Nu = src.size() );
 
     int Np = (Nm-Nk);
     if (Np > 0 && MaxIter > 1) Np /= MaxIter;
@@ -441,7 +441,7 @@ public:
     }
     Glog << std::string(74,'*') << std::endl;
     
-    assert(Nm == evec.size() && Nm == eval.size());
+    GRID_ASSERT(Nm == evec.size() && Nm == eval.size());
 	
     std::vector<std::vector<ComplexD>> lmd(Nu,std::vector<ComplexD>(Nm,0.0));  
     std::vector<std::vector<ComplexD>> lme(Nu,std::vector<ComplexD>(Nm,0.0));  
@@ -622,7 +622,7 @@ private:
     
     int Nu = w.size();
     int Nm = evec.size();
-    assert( b < Nm/Nu );
+    GRID_ASSERT( b < Nm/Nu );
     
     // converts block index to full indicies for an interval [L,R)
     int L = Nu*b;
@@ -630,7 +630,7 @@ private:
 
     Real beta;
 
-    assert((Nu%mrhs)==0);
+    GRID_ASSERT((Nu%mrhs)==0);
     std::vector<Field>   in(mrhs,f_grid);
     std::vector<Field>   out(mrhs,f_grid);
 
@@ -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) {
@@ -711,12 +711,12 @@ private:
     
     for (int u=0; u<Nu; ++u) {
       //      Glog << "norm2(w[" << u << "])= "<< norm2(w[u]) << std::endl;
-      assert (!isnan(norm2(w[u])));
+      GRID_ASSERT (!isnan(norm2(w[u])));
       for (int k=L+u; k<R; ++k) {
 	//        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) {
@@ -734,8 +734,8 @@ private:
 			 Eigen::MatrixXcd & Qt, // Nm x Nm
 			 GridBase *grid)
   {
-    assert( Nk%Nu == 0 && Nm%Nu == 0 );
+    GRID_ASSERT( Nk%Nu == 0 && Nm%Nu == 0 );
-    assert( Nk <= Nm );
+    GRID_ASSERT( Nk <= Nm );
     Eigen::MatrixXcd BlockTriDiag = Eigen::MatrixXcd::Zero(Nk,Nk);
     
     for ( int u=0; u<Nu; ++u ) {
@@ -775,8 +775,8 @@ private:
 			 GridBase *grid)
   {
     Glog << "diagonalize_lapack: Nu= "<<Nu<<" Nk= "<<Nk<<" Nm= "<<std::endl;
-    assert( Nk%Nu == 0 && Nm%Nu == 0 );
+    GRID_ASSERT( Nk%Nu == 0 && Nm%Nu == 0 );
-    assert( Nk <= Nm );
+    GRID_ASSERT( Nk <= Nm );
     Eigen::MatrixXcd BlockTriDiag = Eigen::MatrixXcd::Zero(Nk,Nk);
     
     for ( int u=0; u<Nu; ++u ) {
@@ -924,7 +924,7 @@ if (1){
       diagonalize_lapack(eval,lmd,lme,Nu,Nk,Nm,Qt,grid);
 #endif
     } else { 
-      assert(0);
+      GRID_ASSERT(0);
     }
   }
   
@@ -935,9 +935,9 @@ 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 );
+    GRID_ASSERT( Nk%Nu == 0 && Nm%Nu == 0 );
-    assert( Nk <= Nm );
+    GRID_ASSERT( Nk <= Nm );
     M = Eigen::MatrixXcd::Zero(Nk,Nk);
     
     // rearrange 
@@ -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; 
   }
  
 
@@ -963,9 +963,9 @@ 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 );
+    GRID_ASSERT( Nk%Nu == 0 && Nm%Nu == 0 );
-    assert( Nk <= Nm );
+    GRID_ASSERT( Nk <= Nm );
     
     // rearrange 
     for ( int u=0; u<Nu; ++u ) {
@@ -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/ImplicitlyRestartedLanczos.h

@@ -53,6 +53,18 @@ enum IRLdiagonalisation {
   IRLdiagonaliseWithEigen
 };
 
+enum IRLeigsort { 
+  IRLeigsortMax,
+  IRLeigsortSqMin
+};
+
+#if 0
+bool square_comp(RealD a, RealD b){
+	if (a*a<b*b) return true;
+	return false;
+}
+#endif
+
 template<class Field> class ImplicitlyRestartedLanczosHermOpTester  : public ImplicitlyRestartedLanczosTester<Field>
 {
  public:
@@ -119,8 +131,9 @@ class ImplicitlyRestartedLanczos {
   /////////////////////////
   // Constructor
   /////////////////////////
-  
  public:
+  IRLeigsort EigSort;
+
   
   //////////////////////////////////////////////////////////////////
   // PAB:
@@ -154,6 +167,7 @@ public:
     Nstop(_Nstop)  ,      Nk(_Nk),      Nm(_Nm),
     eresid(_eresid),      betastp(_betastp),
     MaxIter(_MaxIter)  ,      MinRestart(_MinRestart),
+    EigSort(IRLeigsortMax), 
     orth_period(_orth_period), diagonalisation(_diagonalisation)  { };
 
     ImplicitlyRestartedLanczos(LinearFunction<Field> & PolyOp,
@@ -170,6 +184,7 @@ public:
     Nstop(_Nstop)  ,      Nk(_Nk),      Nm(_Nm),
     eresid(_eresid),      betastp(_betastp),
     MaxIter(_MaxIter)  ,      MinRestart(_MinRestart),
+    EigSort(IRLeigsortMax), 
     orth_period(_orth_period), diagonalisation(_diagonalisation)  { };
 
   ////////////////////////////////
@@ -211,7 +226,7 @@ until convergence
   void calc(std::vector<RealD>& eval, std::vector<Field>& evec,  const Field& src, int& Nconv, bool reverse=false)
   {
     GridBase *grid = src.Grid();
-    assert(grid == evec[0].Grid());
+    GRID_ASSERT(grid == evec[0].Grid());
     
     //    GridLogIRL.TimingMode(1);
     std::cout << GridLogIRL <<"**************************************************************************"<< std::endl;
@@ -231,7 +246,7 @@ until convergence
     }
     std::cout << GridLogIRL <<"**************************************************************************"<< std::endl;
 	
-    assert(Nm <= evec.size() && Nm <= eval.size());
+    GRID_ASSERT(Nm <= evec.size() && Nm <= eval.size());
     
     // quickly get an idea of the largest eigenvalue to more properly normalize the residuum
     RealD evalMaxApprox = 0.0;
@@ -245,9 +260,10 @@ until convergence
 	_HermOp(src_n,tmp);
 	//	std::cout << GridLogMessage<< tmp<<std::endl; exit(0);
 	//	std::cout << GridLogIRL << " _HermOp " << norm2(tmp) << std::endl;
-	RealD vnum = real(innerProduct(src_n,tmp)); // HermOp.
+//	RealD vnum = real(innerProduct(src_n,tmp)); // HermOp.
+	RealD vnum = real(innerProduct(tmp,tmp)); // HermOp^2.
 	RealD vden = norm2(src_n);
-	RealD na = vnum/vden;
+	RealD na = std::sqrt(vnum/vden);
 	if (fabs(evalMaxApprox/na - 1.0) < 0.0001)
 	  i=_MAX_ITER_IRL_MEVAPP_;
 	evalMaxApprox = na;
@@ -255,6 +271,7 @@ until convergence
 	src_n = tmp;
       }
     }
+    std::cout << GridLogIRL << " Final evalMaxApprox  " << evalMaxApprox << std::endl;
 	
     std::vector<RealD> lme(Nm);  
     std::vector<RealD> lme2(Nm);
@@ -314,8 +331,12 @@ until convergence
       // sorting
       //////////////////////////////////
       eval2_copy = eval2;
+//      if (EigSort==IRLeigsortMax)
+//      std::partial_sort(eval2.begin(),eval2.begin()+Nm,eval2.end(),square_comp);
+//      else
       std::partial_sort(eval2.begin(),eval2.begin()+Nm,eval2.end(),std::greater<RealD>());
       std::cout<<GridLogIRL <<" evals sorted "<<std::endl;
+//      eval2_copy = eval2;
       const int chunk=8;
       for(int io=0; io<k2;io+=chunk){
 	std::cout<<GridLogIRL << "eval "<< std::setw(3) << io ;
@@ -331,11 +352,12 @@ until convergence
       //////////////////////////////////
       Qt = Eigen::MatrixXd::Identity(Nm,Nm);
       for(int ip=k2; ip<Nm; ++ip){ 
+//        std::cout<<GridLogIRL <<"QR decompose "<<eval2[ip]<<std::endl;
 	QR_decomp(eval,lme,Nm,Nm,Qt,eval2[ip],k1,Nm);
       }
       std::cout<<GridLogIRL <<"QR decomposed "<<std::endl;
 
-      assert(k2<Nm);      assert(k2<Nm);      assert(k1>0);
+      GRID_ASSERT(k2<Nm);      GRID_ASSERT(k2<Nm);      GRID_ASSERT(k1>0);
 
       basisRotate(evec,Qt,k1-1,k2+1,0,Nm,Nm); /// big constraint on the basis
       std::cout<<GridLogIRL <<"basisRotated  by Qt *"<<k1-1<<","<<k2+1<<")"<<std::endl;
@@ -373,7 +395,8 @@ until convergence
 
 	//  power of two search pattern;  not every evalue in eval2 is assessed.
 	int allconv =1;
-	for(int jj = 1; jj<=Nstop; jj*=2){
+//	for(int jj = 1; jj<=Nstop; jj*=2){
+	for(int jj = 1; jj<=Nstop; jj++){
 	  int j = Nstop-jj;
 	  RealD e = eval2_copy[j]; // Discard the evalue
 	  basisRotateJ(B,evec,Qt,j,0,Nk,Nm);	    
@@ -461,7 +484,7 @@ until convergence
   {
     std::cout<<GridLogDebug << "Lanczos step " <<k<<std::endl;
     const RealD tiny = 1.0e-20;
-    assert( k< Nm );
+    GRID_ASSERT( k< Nm );
 
     GridStopWatch gsw_op,gsw_o;
 
@@ -595,7 +618,7 @@ until convergence
     }  else if ( diagonalisation == IRLdiagonaliseWithEigen ) { 
       diagonalize_Eigen(lmd,lme,Nk,Nm,Qt,grid);
     } else { 
-      assert(0);
+      GRID_ASSERT(0);
     }
   }
 
@@ -685,7 +708,7 @@ void diagonalize_lapack(std::vector<RealD>& lmd,
     }
   }
 #else 
-  assert(0);
+  GRID_ASSERT(0);
 #endif
 }
 

Grid/algorithms/iterative/LanczosBidiagonalization.h

@@ -0,0 +1,276 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./Grid/algorithms/iterative/LanczosBidiagonalization.h
+
+Copyright (C) 2015
+
+Author: Chulwoo Jung <chulwoo@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 */
+#ifndef GRID_LANCZOS_BIDIAGONALIZATION_H
+#define GRID_LANCZOS_BIDIAGONALIZATION_H
+
+NAMESPACE_BEGIN(Grid);
+
+/**
+ * Lanczos Bidiagonalization (Golub-Kahan)
+ *
+ * For a linear operator A with adjoint A^dag, constructs the bidiagonal
+ * decomposition:
+ *
+ *   A  V_m = U_m B_m
+ *   A^dag U_m = V_m B_m^T + beta_{m+1} v_{m+1} e_m^T
+ *
+ * where:
+ *   V_m = [v_1, ..., v_m]  right Lanczos vectors (orthonormal)
+ *   U_m = [u_1, ..., u_m]  left  Lanczos vectors (orthonormal)
+ *   B_m is upper bidiagonal with diag(alpha_1,...,alpha_m) and
+ *       superdiag(beta_2,...,beta_m)
+ *
+ * The singular values of A are approximated by those of B_m.
+ * The singular values of B_m are the square roots of the eigenvalues of
+ * the symmetric tridiagonal matrix B_m^T B_m.
+ *
+ * Usage:
+ *   LanczosBidiagonalization<Field> lb(Linop, grid);
+ *   lb.run(src, Nm, tol);
+ *   // Access results via getters.
+ */
+template <class Field>
+class LanczosBidiagonalization {
+
+  public: 
+  LinearOperatorBase<Field> &Linop;
+  GridBase *Grid;
+
+  int Nm;           // number of Lanczos steps taken
+  RealD Tolerance;  // convergence threshold on beta_{k+1} / alpha_k
+
+  std::vector<Field>  V;       // right Lanczos vectors v_1 ... v_m
+  std::vector<Field>  U;       // left  Lanczos vectors u_1 ... u_m
+  std::vector<RealD>  alpha;   // diagonal of bidiagonal matrix
+  std::vector<RealD>  beta;    // super-diagonal (beta[k] couples u_k and v_{k+1})
+
+  // SVD of the bidiagonal matrix (filled after computeSVD())
+  Eigen::VectorXd  singularValues;
+  Eigen::MatrixXd  leftSVecs;   // columns are left  singular vectors of B
+  Eigen::MatrixXd  rightSVecs;  // columns are right singular vectors of B
+
+public:
+
+  LanczosBidiagonalization(LinearOperatorBase<Field> &_Linop, GridBase *_Grid,
+                           RealD _tol = 1.0e-8)
+    : Linop(_Linop), Grid(_Grid), Tolerance(_tol), Nm(0)
+  {}
+
+  /**
+   * Run the Golub-Kahan Lanczos bidiagonalization.
+   *
+   * Parameters
+   * ----------
+   * src  : starting vector (need not be normalised)
+   * Nmax : maximum number of Lanczos steps
+   * reorth : if true, full reorthogonalisation of both V and U bases
+   */
+  void run(const Field &src, int Nmax, bool reorth = true)
+  {
+    assert(norm2(src) > 0.0);
+
+    V.clear(); U.clear();
+    alpha.clear(); beta.clear();
+    Nm = 0;
+
+    Field p(Grid), r(Grid);
+
+    // --- initialise: v_1 = src / ||src|| ---
+    Field v(Grid);
+    v = src;
+    RealD nrm = std::sqrt(norm2(v));
+    v = (1.0 / nrm) * v;
+    V.push_back(v);
+
+    for (int k = 0; k < Nmax; ++k) {
+
+      // p = A v_k
+      Linop.Op(V[k], p);
+
+      // p = p - beta_k * u_{k-1}   (remove previous left vector)
+      if (k > 0) {
+        p = p - beta[k-1] * U[k-1];
+      }
+
+      // alpha_k = ||p||
+      RealD ak = std::sqrt(norm2(p));
+      if (ak < 1.0e-14) {
+        std::cout << GridLogMessage
+                  << "LanczosBidiagonalization: lucky breakdown at step "
+                  << k << " (alpha = " << ak << ")" << std::endl;
+        break;
+      }
+      alpha.push_back(ak);
+
+      // u_k = p / alpha_k
+      Field u(Grid);
+      u = (1.0 / ak) * p;
+
+      // full reorthogonalisation of u against previous U
+      if (reorth) {
+        for (int j = 0; j < (int)U.size(); ++j) {
+          ComplexD ip = innerProduct(U[j], u);
+          u = u - ip * U[j];
+        }
+        RealD unrm = std::sqrt(norm2(u));
+        if (unrm > 1.0e-14) u = (1.0 / unrm) * u;
+      }
+      U.push_back(u);
+
+      // r = A^dag u_k - alpha_k * v_k
+      Linop.AdjOp(U[k], r);
+      r = r - ak * V[k];
+
+      // full reorthogonalisation of r against previous V
+      if (reorth) {
+        for (int j = 0; j < (int)V.size(); ++j) {
+          ComplexD ip = innerProduct(V[j], r);
+          r = r - ip * V[j];
+        }
+      }
+
+      // beta_{k+1} = ||r||
+      RealD bk = std::sqrt(norm2(r));
+      beta.push_back(bk);
+
+      Nm = k + 1;
+
+      std::cout << GridLogMessage
+                << "LanczosBidiagonalization step " << k
+                << "  alpha = " << ak
+                << "  beta  = " << bk << std::endl;
+
+      // convergence: residual beta / alpha small enough
+      if (bk / ak < Tolerance) {
+        std::cout << GridLogMessage
+                  << "LanczosBidiagonalization converged at step " << k
+                  << "  (beta/alpha = " << bk / ak << ")" << std::endl;
+        break;
+      }
+
+      if (k == Nmax - 1) break;   // no v_{k+2} needed after last step
+
+      // v_{k+1} = r / beta_{k+1}
+      Field vnext(Grid);
+      vnext = (1.0 / bk) * r;
+      V.push_back(vnext);
+    }
+  }
+
+  /**
+   * Compute the SVD of the bidiagonal matrix B using Eigen.
+   * Singular values are stored in descending order.
+   */
+  void computeSVD()
+  {
+    int m = Nm;
+    Eigen::MatrixXd B = Eigen::MatrixXd::Zero(m, m);
+
+    for (int k = 0; k < m; ++k) {
+      B(k, k) = alpha[k];
+      if (k + 1 < m && k < (int)beta.size())
+        B(k, k+1) = beta[k];
+    }
+
+    Eigen::JacobiSVD<Eigen::MatrixXd> svd(B,
+        Eigen::ComputeThinU | Eigen::ComputeThinV);
+
+    singularValues = svd.singularValues();   // already sorted descending
+    leftSVecs      = svd.matrixU();
+    rightSVecs     = svd.matrixV();
+
+    std::cout << GridLogMessage
+              << "LanczosBidiagonalization: singular values of B_" << m
+              << std::endl;
+    for (int k = 0; k < m; ++k)
+      std::cout << GridLogMessage << "  sigma[" << k << "] = "
+                << singularValues(k) << std::endl;
+  }
+
+  /**
+   * Return the k-th approximate left singular vector of A in the full
+   * lattice space.  computeSVD() must have been called first.
+   */
+  Field leftSingularVector(int k)
+  {
+    assert(k < (int)leftSVecs.cols());
+    Field svec(Grid);
+    svec = Zero();
+    for (int j = 0; j < Nm; ++j)
+      svec = svec + leftSVecs(j, k) * U[j];
+    return svec;
+  }
+
+  /**
+   * Return the k-th approximate right singular vector of A in the full
+   * lattice space.  computeSVD() must have been called first.
+   */
+  Field rightSingularVector(int k)
+  {
+    assert(k < (int)rightSVecs.cols());
+    Field svec(Grid);
+    svec = Zero();
+    for (int j = 0; j < Nm; ++j)
+      svec = svec + rightSVecs(j, k) * V[j];
+    return svec;
+  }
+
+  /**
+   * Verify the bidiagonalization: returns max residual
+   *   max_k || A v_k - alpha_k u_k - beta_k u_{k-1} ||
+   */
+  RealD verify()
+  {
+    Field tmp(Grid);
+    RealD maxres = 0.0;
+    for (int k = 0; k < Nm; ++k) {
+      Linop.Op(V[k], tmp);
+      tmp = tmp - alpha[k] * U[k];
+      if (k > 0 && k-1 < (int)beta.size())
+        tmp = tmp - beta[k-1] * U[k-1];
+      RealD res = std::sqrt(norm2(tmp));
+      if (res > maxres) maxres = res;
+      std::cout << GridLogMessage
+                << "LanczosBidiagonalization verify step " << k
+                << "  ||A v_k - alpha_k u_k - beta_{k-1} u_{k-1}|| = "
+                << res << std::endl;
+    }
+    return maxres;
+  }
+
+  /* Getters */
+  int                       getNm()           const { return Nm; }
+  const std::vector<Field>& getV()            const { return V; }
+  const std::vector<Field>& getU()            const { return U; }
+  const std::vector<RealD>& getAlpha()        const { return alpha; }
+  const std::vector<RealD>& getBeta()         const { return beta; }
+  Eigen::VectorXd           getSingularValues() const { return singularValues; }
+};
+
+NAMESPACE_END(Grid);
+#endif

Grid/algorithms/iterative/LocalCoherenceLanczos.h

@@ -80,7 +80,7 @@ public:
   ProjectedHermOp(LinearOperatorBase<FineField>& linop, std::vector<FineField> & _subspace) : 
     _Linop(linop), subspace(_subspace)
   {  
-    assert(subspace.size() >0);
+    GRID_ASSERT(subspace.size() >0);
   };
 
   void operator()(const CoarseField& in, CoarseField& out) {
@@ -346,12 +346,12 @@ public:
 
   void testFine(RealD resid) 
   {
-    assert(evals_fine.size() == nbasis);
+    GRID_ASSERT(evals_fine.size() == nbasis);
-    assert(subspace.size() == nbasis);
+    GRID_ASSERT(subspace.size() == nbasis);
     PlainHermOp<FineField>    Op(_FineOp);
     ImplicitlyRestartedLanczosHermOpTester<FineField> SimpleTester(Op);
     for(int k=0;k<nbasis;k++){
-      assert(SimpleTester.ReconstructEval(k,resid,subspace[k],evals_fine[k],1.0)==1);
+      GRID_ASSERT(SimpleTester.ReconstructEval(k,resid,subspace[k],evals_fine[k],1.0)==1);
     }
   }
 
@@ -359,8 +359,8 @@ public:
   //hence the smoother can be tuned after running the coarse Lanczos by using a different smoother here
   void testCoarse(RealD resid,ChebyParams cheby_smooth,RealD relax) 
   {
-    assert(evals_fine.size() == nbasis);
+    GRID_ASSERT(evals_fine.size() == nbasis);
-    assert(subspace.size() == nbasis);
+    GRID_ASSERT(subspace.size() == nbasis);
     //////////////////////////////////////////////////////////////////////////////////////////////////
     // create a smoother and see if we can get a cheap convergence test and smooth inside the IRL
     //////////////////////////////////////////////////////////////////////////////////////////////////
@@ -380,7 +380,7 @@ public:
   void calcFine(ChebyParams cheby_parms,int Nstop,int Nk,int Nm,RealD resid, 
 		RealD MaxIt, RealD betastp, int MinRes)
   {
-    assert(nbasis<=Nm);
+    GRID_ASSERT(nbasis<=Nm);
     Chebyshev<FineField>      Cheby(cheby_parms);
     FunctionHermOp<FineField> ChebyOp(Cheby,_FineOp);
     PlainHermOp<FineField>    Op(_FineOp);
@@ -400,8 +400,8 @@ public:
     IRL.calc(evals_fine,subspace,src,Nconv,false);
     
     // Shrink down to number saved
-    assert(Nstop>=nbasis);
+    GRID_ASSERT(Nstop>=nbasis);
-    assert(Nconv>=nbasis);
+    GRID_ASSERT(Nconv>=nbasis);
     evals_fine.resize(nbasis);
     subspace.resize(nbasis,_FineGrid);
   }
@@ -433,7 +433,7 @@ public:
     ImplicitlyRestartedLanczos<CoarseField> IRL(ChebyOp,ChebyOp,ChebySmoothTester,Nstop,Nk,Nm,resid,MaxIt,betastp,MinRes);
     int Nconv=0;
     IRL.calc(evals_coarse,evec_coarse,src,Nconv,false);
-    assert(Nconv>=Nstop);
+    GRID_ASSERT(Nconv>=Nstop);
     evals_coarse.resize(Nstop);
     evec_coarse.resize (Nstop,_CoarseGrid);
     for (int i=0;i<Nstop;i++){

Grid/algorithms/iterative/MinimalResidual.h

@@ -35,7 +35,7 @@ template<class Field> class MinimalResidual : public OperatorFunction<Field> {
  public:
   using OperatorFunction<Field>::operator();
 
-  bool ErrorOnNoConverge; // throw an assert when the MR fails to converge.
+  bool ErrorOnNoConverge; // throw an GRID_ASSERT when the MR fails to converge.
                           // Defaults true.
   RealD   Tolerance;
   Integer MaxIterations;
@@ -59,7 +59,7 @@ template<class Field> class MinimalResidual : public OperatorFunction<Field> {
 
     // Initial residual computation & set up
     RealD guess = norm2(psi);
-    assert(std::isnan(guess) == 0);
+    GRID_ASSERT(std::isnan(guess) == 0);
 
     RealD ssq = norm2(src);
     RealD rsq = Tolerance * Tolerance * ssq;
@@ -136,7 +136,7 @@ template<class Field> class MinimalResidual : public OperatorFunction<Field> {
         std::cout << GridLogMessage << "MR Time elapsed: Linalg  " << LinalgTimer.Elapsed() << std::endl;
 
         if (ErrorOnNoConverge)
-          assert(true_residual / Tolerance < 10000.0);
+          GRID_ASSERT(true_residual / Tolerance < 10000.0);
 
         IterationsToComplete = k;
 
@@ -148,7 +148,7 @@ template<class Field> class MinimalResidual : public OperatorFunction<Field> {
               << std::endl;
 
     if (ErrorOnNoConverge)
-      assert(0);
+      GRID_ASSERT(0);
 
     IterationsToComplete = k;
   }

Grid/algorithms/iterative/MixedPrecisionFlexibleGeneralisedMinimalResidual.h

@@ -37,7 +37,7 @@ class MixedPrecisionFlexibleGeneralisedMinimalResidual : public OperatorFunction
 
   using OperatorFunction<FieldD>::operator();
 
-  bool ErrorOnNoConverge; // Throw an assert when MPFGMRES fails to converge,
+  bool ErrorOnNoConverge; // Throw an GRID_ASSERT when MPFGMRES fails to converge,
                           // defaults to true
 
   RealD   Tolerance;
@@ -91,7 +91,7 @@ class MixedPrecisionFlexibleGeneralisedMinimalResidual : public OperatorFunction
     conformable(psi, src);
 
     RealD guess = norm2(psi);
-    assert(std::isnan(guess) == 0);
+    GRID_ASSERT(std::isnan(guess) == 0);
 
     RealD cp;
     RealD ssq = norm2(src);
@@ -150,7 +150,7 @@ class MixedPrecisionFlexibleGeneralisedMinimalResidual : public OperatorFunction
     std::cout << GridLogMessage << "MPFGMRES did NOT converge" << std::endl;
 
     if (ErrorOnNoConverge)
-      assert(0);
+      GRID_ASSERT(0);
   }
 
   RealD outerLoopBody(LinearOperatorBase<FieldD> &LinOp, const FieldD &src, FieldD &psi, RealD rsq) {
@@ -197,7 +197,7 @@ class MixedPrecisionFlexibleGeneralisedMinimalResidual : public OperatorFunction
       }
     }
 
-    assert(0); // Never reached
+    GRID_ASSERT(0); // Never reached
     return cp;
   }
 

Grid/algorithms/iterative/NormalEquations.h

@@ -60,6 +60,32 @@ 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_ = 100; 
+    const int _MAX_ITER_EST_ = 200; 
 
     for (int i=0;i<_MAX_ITER_EST_;i++) { 
       
@@ -30,18 +30,17 @@ template<class Field> class PowerMethod
       RealD vden = norm2(src_n); 
       RealD na = vnum/vden; 
 
-      std::cout << GridLogIterative << "PowerMethod: Current approximation of largest eigenvalue " << na << std::endl;
+      std::cout << GridLogMessage << "PowerMethod: Current approximation of largest eigenvalue " << na << std::endl;
       
-      if ( (fabs(evalMaxApprox/na - 1.0) < 0.001) || (i==_MAX_ITER_EST_-1) ) { 
+      //      if ( (fabs(evalMaxApprox/na - 1.0) < 0.0001) || (i==_MAX_ITER_EST_-1) ) { 
- 	evalMaxApprox = na; 
+	// 	evalMaxApprox = na; 
-	std::cout << GridLogMessage << " Approximation of largest eigenvalue: " << evalMaxApprox << std::endl;
+	// 	return evalMaxApprox; 
- 	return evalMaxApprox; 
+      //      } 
-      } 
       evalMaxApprox = na; 
       src_n = tmp;
     }
-    assert(0);
+    std::cout << GridLogMessage << " Approximation of largest eigenvalue: " << evalMaxApprox << std::endl;
-    return 0;
+    return evalMaxApprox;
   }
 };
 }

Grid/algorithms/iterative/PowerSpectrum.h

@@ -0,0 +1,76 @@
+#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/iterative/PrecConjugateResidual.h

@@ -112,7 +112,7 @@ public:
     }
 
     std::cout<<GridLogMessage<<"PrecConjugateResidual did NOT converge"<<std::endl;
-    assert(0);
+    GRID_ASSERT(0);
   }
 };
 NAMESPACE_END(Grid);

Grid/algorithms/iterative/PrecGeneralisedConjugateResidual.h

@@ -118,7 +118,7 @@ public:
 
     }
     GCRLogLevel<<"Variable Preconditioned GCR did not converge"<<std::endl;
-    //    assert(0);
+    //    GRID_ASSERT(0);
   }
 
   RealD GCRnStep(const Field &src, Field &psi,RealD rsq){
@@ -221,7 +221,7 @@ public:
       int northog = ((kp)>(mmax-1))?(mmax-1):(kp);  // if more than mmax done, we orthog all mmax history.
       for(int back=0;back<northog;back++){
 
-	int peri_back=(k-back)%mmax;   	  assert((k-back)>=0);
+	int peri_back=(k-back)%mmax;   	  GRID_ASSERT((k-back)>=0);
 
 	b=-real(innerProduct(q[peri_back],Az))/qq[peri_back];
 	p[peri_kp]=p[peri_kp]+b*p[peri_back];
@@ -231,7 +231,7 @@ public:
       qq[peri_kp]=norm2(q[peri_kp]); // could use axpy_norm
       LinalgTimer.Stop();
     }
-    assert(0); // never reached
+    GRID_ASSERT(0); // never reached
     return cp;
   }
 };

Grid/algorithms/iterative/PrecGeneralisedConjugateResidualNonHermitian.h

@@ -66,15 +66,26 @@ public:
     Linop(_Linop),
     Preconditioner(Prec),
     mmax(_mmax),
-    nstep(_nstep)
+    nstep(_nstep)         // what is nstep vs mmax? one is the number of inner iterations
   { 
     level=1;
     verbose=1;
   };
 
+  // virtual method stubs for updating GCR polynomial
+  virtual void LogBegin(void){
+    std::cout << "GCR::LogBegin() "<<std::endl;
+  };
+  virtual void LogIteration(int k, ComplexD a, std::vector<ComplexD> betas){
+    std::cout << "GCR::LogIteration() "<<std::endl;
+  };
+  virtual void LogComplete(std::vector<ComplexD>& alphas, std::vector<std::vector<ComplexD>>& betas) {
+    std::cout << "GCR::LogComplete() "<<std::endl;
+  };
+
   void operator() (const Field &src, Field &psi){
 
-    psi=Zero();
+    //    psi=Zero();
     RealD cp, ssq,rsq;
     ssq=norm2(src);
     rsq=Tolerance*Tolerance*ssq;
@@ -96,7 +107,6 @@ public:
       GCRLogLevel <<"PGCR("<<mmax<<","<<nstep<<") "<< steps <<" steps cp = "<<cp<<" target "<<rsq <<std::endl;
 
       if(cp<rsq) {
-
         SolverTimer.Stop();
 
         Linop.Op(psi,r);
@@ -113,7 +123,7 @@ public:
 
     }
     GCRLogLevel<<"Variable Preconditioned GCR did not converge"<<std::endl;
-    //    assert(0);
+    //    GRID_ASSERT(0);
   }
 
   RealD GCRnStep(const Field &src, Field &psi,RealD rsq){
@@ -135,9 +145,9 @@ public:
     ////////////////////////////////
     // history for flexible orthog
     ////////////////////////////////
-    std::vector<Field> q(mmax,grid);
+    std::vector<Field> q(mmax,grid);        // q = Ap
-    std::vector<Field> p(mmax,grid);
+    std::vector<Field> p(mmax,grid);        // store mmax conjugate momenta
-    std::vector<RealD> qq(mmax);
+    std::vector<RealD> qq(mmax);            // qq = (Ap)^2 = <p|A^\dagger A |p> (denom of \alpha)
       
     GCRLogLevel<< "PGCR nStep("<<nstep<<")"<<std::endl;
 
@@ -157,6 +167,8 @@ public:
     LinalgTimer.Stop();
     GCRLogLevel<< "PGCR true residual r = src - A psi   "<< norm2(r) <<std::endl;
 
+    this->LogBegin();       // initialize polynomial GCR if needed (TODO think about placement of this)
+    
     /////////////////////
     // p = Prec(r)
     /////////////////////
@@ -179,31 +191,44 @@ public:
     q[0]= Az;
     qq[0]= zAAz;
 
+    std::cout << "||init p - src||: " << norm2(p[0] - src) << std::endl;   // for debugging
+    
     cp =norm2(r);
     LinalgTimer.Stop();
 
+    std::vector<ComplexD> all_alphas;
+    std::vector<std::vector<ComplexD>> all_betas;
+
     for(int k=0;k<nstep;k++){
 
       steps++;
 
       int kp     = k+1;
-      int peri_k = k %mmax;
+      int peri_k = k %mmax;     // only store mmax vectors; just roll around if needed
       int peri_kp= kp%mmax;
 
+      // std::cout << "peri_kp = " << peri_kp << std::endl;
+
       LinalgTimer.Start();
       rq= innerProduct(q[peri_k],r); // what if rAr not real?
-      a = rq/qq[peri_k];
+      a = rq/qq[peri_k];              // compute alpha_j
 
-      axpy(psi,a,p[peri_k],psi);         
+      all_alphas.push_back(a);
 
-      cp = axpy_norm(r,-a,q[peri_k],r);
+      axpy(psi,a,p[peri_k],psi);      // update psi --> psi + \alpha p
+
+      cp = axpy_norm(r,-a,q[peri_k],r);       // update r --> r - \alpha D p. Note q = Dp
       LinalgTimer.Stop();
 
-      GCRLogLevel<< "PGCR step["<<steps<<"]  resid " << cp << " target " <<rsq<<std::endl; 
+      // LogIterationA(k + 1, a);
 
-      if((k==nstep-1)||(cp<rsq)){
+      GCRLogLevel<< "GCR step["<<steps<<"]  resid " << cp << " target " <<rsq<<std::endl; 
-	return cp;
+
-      }
+      // moving this to end of loop so that it doesn't exit beforehand
+      // TODO if I want to uncomment this, I have to split the LogIteration again and put LogIterationA() beforehand
+      // if((k==nstep-1)||(cp<rsq)){
+      //   return cp;
+      // }
 
 
       PrecTimer.Start();
@@ -221,22 +246,205 @@ public:
       q[peri_kp]=Az;
       p[peri_kp]=z;
 
+      // Field Dsrc (grid);
+      // Linop.Op(src, Dsrc);
+      // std::cout << "||q[peri_kp] - D(src)||: " << norm2(q[peri_kp] - Dsrc) << std::endl;   // for debugging
+
+          // // delete after testing
+          // std::cout << "Testing Dsq on one for GCR: " << std::endl;
+          // Field myField (grid);
+          // myField = 1.0;
+          // Field out1 (grid); Field out2 (grid);
+          // Linop.HermOp(myField, out1);
+          // Linop.Op(myField, out2);
+          // std::cout << "Dsq.Hermop(ones) has norm " << norm2(out1) << std::endl;
+          // std::cout << "Dsq.Op(ones) has norm " << norm2(out2) << std::endl;
+
+      // basically northog = k+1 if mmax is large
       int northog = ((kp)>(mmax-1))?(mmax-1):(kp);  // if more than mmax done, we orthog all mmax history.
+      // std::cout << "northog: " << northog << std::endl;
+      std::vector<ComplexD> betas (northog);
+      // std::cout << "peri_kp: " << peri_kp << std::endl;
+      // we iterate backwards counting down from the current k+1 index (peri_kp) because we 
       for(int back=0;back<northog;back++){
 
-	int peri_back=(k-back)%mmax;   	  assert((k-back)>=0);
+	int peri_back=(k-back)%mmax;   	  GRID_ASSERT((k-back)>=0);
 
-	b=-real(innerProduct(q[peri_back],Az))/qq[peri_back];
+        // b=-real(innerProduct(q[peri_back],Az))/qq[peri_back];
+        b=-(innerProduct(q[peri_back],Az))/qq[peri_back];     // TODO try complex beta
         p[peri_kp]=p[peri_kp]+b*p[peri_back];
         q[peri_kp]=q[peri_kp]+b*q[peri_back];
 
+        // LogIterationB(peri_back, b);
+        // betas[back] = b;    // may need to change the indexing if I ever do it with restarts
+        // std::cout << "[DEBUG] pushing beta for back = " << back << ", peri_back = " << peri_back << std::endl;
+
+        betas[peri_back] = b;    // may need to change the indexing if I ever do it with restarts
+
       }
       qq[peri_kp]=norm2(q[peri_kp]); // could use axpy_norm
       LinalgTimer.Stop();
+
+      // log iteration and update GCR polynomial if necessary.
+      all_betas.push_back(betas);
+      LogIteration(k + 1, a, betas);
+
+      // finish if necessary
+      if((k==nstep-1)||(cp<rsq)){
+        std::cout << "All alphas: " << std::endl << all_alphas << std::endl;
+        std::cout << "All betas: " << std::endl << all_betas << std::endl;
+        LogComplete(all_alphas, all_betas);
+        std::cout << "Exiting GCR." << std::endl;
+        return cp;
       }
-    assert(0); // never reached
+
+    }
+    GRID_ASSERT(0); // never reached
     return cp;
   }
 };
+
+class PolynomialFile: Serializable {
+  public:
+    GRID_SERIALIZABLE_CLASS_MEMBERS(PolynomialFile, 
+      std::vector<std::vector<std::complex<double>>>, data,
+      std::vector<std::vector<std::complex<double>>>, betas,
+      std::vector<std::complex<double>>,              alphas
+    );
+};
+
+// Optionally record the GCR polynomial. [PO]: TODO
+template <class Field>
+class PGCRPolynomial : public PrecGeneralisedConjugateResidualNonHermitian<Field> {
+public:
+  std::vector<ComplexD> ak;
+  std::vector<std::vector<ComplexD>> bk;
+  // std::vector<ComplexD> poly_p;
+  std::vector<std::vector<ComplexD>> poly_p;
+  std::vector<ComplexD> poly_Ap;        // polynomial in Ap_j (only store it for last p)
+  std::vector<ComplexD> poly_r;
+  std::vector<ComplexD> polynomial;
+
+  PolynomialFile& PF;
+
+public:
+  PGCRPolynomial(RealD tol, Integer maxit,LinearOperatorBase<Field> &_Linop, LinearFunction<Field> &Prec, int _mmax, int _nstep, PolynomialFile& _PF)
+    : PrecGeneralisedConjugateResidualNonHermitian<Field>(tol, maxit, _Linop, Prec, _mmax, _nstep), PF(_PF)
+  {};
+
+  // think this applies the polynomial in A = Linop to a field src. The coeffs are 
+  // stored in the vector `polynomial`.
+  void PolyOp(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;
+      this->Linop.Op(tmp,AtoN);               // iterate A^n
+      psi = psi + polynomial[n]*AtoN;       // psi += poly_n A^n src
+    }
+  }
+
+  // [PO TODO] debug this
+  void PGCRsequence(const Field &src, Field &x)
+  {
+    Field Ap(src.Grid());
+    Field r(src.Grid());
+    // Field p(src.Grid());
+    // p=src;
+    std::vector<Field> p;
+    p.push_back(src);
+    r=src;
+    x=Zero();
+    x.Checkerboard()=src.Checkerboard();
+    for(int k=0;k<ak.size();k++){
+      x = x + ak[k]*p[k];
+      this->Linop.Op(p[k], Ap);
+      r = r - ak[k] * Ap;
+      // p[k] = r;
+      p.push_back(r);
+      for (int i = 0; i < k; i++) {     // [PO TODO] check indices
+        p[k+1] += bk[i, k+1] * p[i];
+      }
+      // p = r + bk[k] * p;
+    }
+  }
+
+  void Solve(const Field &src, Field &psi)
+  {
+    psi=Zero();
+    this->operator()(src, psi);
+  }
+
+  virtual void LogBegin(void)
+  {
+    std::cout << "PGCR::LogBegin() "<<std::endl;
+    ak.resize(0);
+    bk.resize(0);
+    polynomial.resize(0);
+    poly_Ap.push_back(0.0);     // start with (0.0); during first iteration should change to (0.0, 1.0)
+    std::vector<ComplexD> p0_tmp;
+    p0_tmp.push_back(1.0);
+    poly_p.push_back(p0_tmp);
+    poly_r.push_back(1.0);
+  };
+
+  // Updates vector psi and r and initializes vector p[k+1]
+  virtual void LogIteration(int k, ComplexD a, std::vector<ComplexD> betas){
+    std::cout << "PGCR::LogIteration(k = " << k << ")" << std::endl;
+    ak.push_back(a);
+    bk.push_back(betas);
+
+    // update Ap by pushing p[k] to the right
+    poly_Ap.push_back(0.0);   // need to pad the end with an element
+    poly_Ap[0] = 0.0;         // technically this should be unnecessary, as the first component is never set
+    for(int i = 0; i < k; i++){
+      poly_Ap[i+1]=poly_p[k-1][i];        // A\vec{p} = (0, \vec{p}) bc A shifts components of p to the right
+    }
+
+    // update psi_{k+1} --> psi_k + a_k p_k
+    polynomial.push_back(0.0);
+    for(int i = 0; i < k; i++) {
+      polynomial[i] += a * poly_p[k-1][i];
+    }
+    {
+      std::vector<std::complex<double>> poly_stdcmplx(polynomial.begin(), polynomial.end());
+      PF.data.push_back(poly_stdcmplx);
+    }
+
+    //  r_{k+1} --> r_k - a_k A p_k
+    //  p_{k+1} --> r_k + \sum_{i=0}^k \beta_{ik} p_i, input betas = (\beta_{ik})_i
+    poly_r.push_back(0.0);        // should be of size k+1 if we start with k = 1
+    std::vector<ComplexD> p_next (k + 1, ComplexD(0.0));     // p_{k+1} = same size as r_{k+1}
+    for(int i = 0; i < k + 1; i++){
+      poly_r[i] = poly_r[i] - a * poly_Ap[i];     // update r_{k+1} --> r_k - \alpha_k A p_k
+      p_next[i] = poly_r[i];                 // init new vector as r_{k+1}
+    }
+
+    // p_{k+1} --> p_{k+1} + \sum_i \beta_{ij} p_i
+    int nbeta = betas.size();
+    std::cout << "Betas: " << betas << std::endl;
+    for (int j = 0; j < nbeta; j++) {
+      for (int i = 0; i < j+1; i++) {
+        p_next[i] += betas[j] * poly_p[j][i];
+      }
+    }
+    poly_p.push_back(p_next);                 // add p_{k+1} to the list of p's
+  }
+
+  virtual void LogComplete(std::vector<ComplexD>& alphas, std::vector<std::vector<ComplexD>>& betas) {
+    /** Logs all alphas and betas to complete the iterations. */
+    std::cout << "PGCR::LogComplete() "<<std::endl;
+    for (int i = 0; i < alphas.size(); i++) {
+      PF.alphas.push_back(std::complex<double>(alphas[i].real(), alphas[i].imag()));
+      std::vector<std::complex<double>> beta_stdcmplx(betas[i].begin(), betas[i].end());
+      PF.betas.push_back(beta_stdcmplx);
+    }
+  };
+
+};
+
 NAMESPACE_END(Grid);
 #endif

Grid/algorithms/iterative/QuasiMinimalResidual.h

@@ -79,7 +79,7 @@ class QuasiMinimalResidual : public OperatorFunction<Field> {
 
     LinOp.Op(x,r); r = b - r;
 
-    assert(normb> 0.0);
+    GRID_ASSERT(normb> 0.0);
 
     resid = norm2(r)/normb;
     if (resid <= Tolerance) {
@@ -105,8 +105,8 @@ class QuasiMinimalResidual : public OperatorFunction<Field> {
     for (int i = 1; i <= MaxIterations; i++) {
 
       // Breakdown tests
-      assert( rho != 0.0);
+      GRID_ASSERT( rho != 0.0);
-      assert( xi  != 0.0);
+      GRID_ASSERT( xi  != 0.0);
 
       v = (1. / rho) * v_tld;
       y = (1. / rho) * y;
@@ -134,10 +134,10 @@ class QuasiMinimalResidual : public OperatorFunction<Field> {
       ep=Zep.real();
       std::cout << "Zep "<<Zep <<std::endl;
       // Complex Audit
-      assert(abs(ep)>0);
+      GRID_ASSERT(abs(ep)>0);
 
       beta = ep / delta;
-      assert(abs(beta)>0);
+      GRID_ASSERT(abs(beta)>0);
 
       v_tld = p_tld - beta * v;
       y = v_tld;
@@ -158,7 +158,7 @@ class QuasiMinimalResidual : public OperatorFunction<Field> {
       std::cout << "theta "<<theta<<std::endl;
       std::cout << "gamma "<<gamma<<std::endl;
 
-      assert(abs(gamma)> 0.0);
+      GRID_ASSERT(abs(gamma)> 0.0);
 
       eta = -eta * rho_1 * gamma* gamma / (beta * gamma_1 * gamma_1);
 
@@ -178,7 +178,7 @@ class QuasiMinimalResidual : public OperatorFunction<Field> {
       }
       std::cout << "Iteration "<<i<<" resid " << resid<<std::endl;
     }
-    assert(0);
+    GRID_ASSERT(0);
     return;                            // no convergence
   }
 #else

Grid/algorithms/iterative/RestartedLanczosBidiagonalization.h

@@ -0,0 +1,753 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./Grid/algorithms/iterative/RestartedLanczosBidiagonalization.h
+
+Copyright (C) 2015
+
+Author: Chulwoo Jung <chulwoo@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 */
+#ifndef GRID_RESTARTED_LANCZOS_BIDIAGONALIZATION_H
+#define GRID_RESTARTED_LANCZOS_BIDIAGONALIZATION_H
+
+NAMESPACE_BEGIN(Grid);
+
+/**
+ * Implicitly Restarted Lanczos Bidiagonalization (IRLBA)
+ *
+ * Computes the p largest (or p smallest) singular triplets of a linear
+ * operator A using the Golub-Kahan-Lanczos bidiagonalization with implicit
+ * restart via thick-restart / QR shifts.
+ *
+ * Algorithm (Baglama & Reichel, SIAM J. Sci. Comput. 27(1):19-42, 2005):
+ *
+ *   Outer loop:
+ *     1. Extend the p-step (or seed) bidiagonalization to k steps:
+ *           A  V_k = U_k B_k
+ *           A^dag U_k = V_k B_k^T + beta_{k+1} v_{k+1} e_k^T
+ *     2. Compute SVD:  B_k = X Sigma Y^T
+ *     3. Check convergence of the p desired singular values via
+ *           |beta_{k+1} * y_{k,i}|  <  tol * sigma_i
+ *        where y_{k,i} is the last component of the i-th right singular vector.
+ *     4. Apply k-p implicit QR shifts to implicitly compress the basis
+ *        to p steps (Sorensen-Lehoucq thick restart):
+ *           B_p^+ = X_p^T B_k Y_p   (upper bidiagonal, p x p)
+ *        and update the lattice vectors:
+ *           V_p^+ = V_k Y_p
+ *           U_p^+ = U_k X_p
+ *        The new residual coupling is
+ *           beta_p^+ v_{p+1}^+ = beta_{k+1} v_{k+1} * (e_k^T Y_p)_p
+ *                               + B_k(p,p+1) * (orthogonal tail from QR)
+ *     5. Go to step 1.
+ *
+ * Template parameter
+ * ------------------
+ *   Field : lattice field type (must support Grid algebra operations)
+ *
+ * Usage
+ * -----
+ *   RestartedLanczosBidiagonalization<Field> irlba(Linop, grid, p, k, tol, maxIter);
+ *   irlba.run(src);
+ *   // Results available via getters.
+ */
+template <class Field>
+class RestartedLanczosBidiagonalization {
+
+public:
+  LinearOperatorBase<Field> &Linop;
+  GridBase *Grid;
+
+  int    Nk;       // number of desired singular triplets
+  int    Nm;       // Lanczos basis size (Nm > Nk)
+  RealD  Tolerance;
+  int    MaxIter;
+  bool   largest; // if true, target largest singular values; otherwise smallest
+
+  // Converged singular triplets (filled after run())
+  std::vector<RealD>  singularValues;   // sigma_0 >= sigma_1 >= ...
+  std::vector<Field>  leftVectors;      // approximate left singular vectors
+  std::vector<Field>  rightVectors;     // approximate right singular vectors
+
+private:
+  // Working bases (size up to Nm+1)
+  std::vector<Field>  V;    // right Lanczos vectors
+  std::vector<Field>  U;    // left  Lanczos vectors
+  std::vector<RealD>  alpha;
+  std::vector<RealD>  beta;
+
+  // After a thick restart, the column at index restart_col of U^dag A V
+  // has extra non-zero entries (rows 0..restart_col-2) beyond what the
+  // upper bidiagonal captures.  fvec[j] = <U[j] | A V[restart_col]> for
+  // j = 0..restart_col-1.  (fvec[restart_col-1] == beta[restart_col-1].)
+  // reset_col == -1 means no restart has occurred yet (pure bidiagonal).
+  std::vector<RealD>  fvec;
+  int                 restart_col;
+
+public:
+
+  RestartedLanczosBidiagonalization(LinearOperatorBase<Field> &_Linop,
+                                    GridBase *_Grid,
+                                    int _Nk, int _Nm,
+                                    RealD _tol   = 1.0e-8,
+                                    int   _maxIt = 300,
+                                    bool  _largest = true)
+    : Linop(_Linop), Grid(_Grid),
+      Nk(_Nk), Nm(_Nm),
+      Tolerance(_tol), MaxIter(_maxIt),
+      largest(_largest)
+  {
+    assert(Nm > Nk);
+  }
+
+  /**
+   * Run IRLBA starting from src.
+   * On exit, singularValues, leftVectors, rightVectors are filled with
+   * the Nk converged singular triplets.
+   */
+  void run(const Field &src)
+  {
+    assert(norm2(src) > 0.0);
+
+    singularValues.clear();
+    leftVectors.clear();
+    rightVectors.clear();
+
+    // Allocate working bases
+    V.clear(); U.clear();
+    alpha.clear(); beta.clear();
+    fvec.clear(); restart_col = -1;
+    V.reserve(Nm + 1);
+    U.reserve(Nm);
+
+    // Seed: v_0 = src / ||src||
+    Field vtmp(Grid);
+    vtmp = src;
+    RealD nrm = std::sqrt(norm2(vtmp));
+    vtmp = (1.0 / nrm) * vtmp;
+    V.push_back(vtmp);
+
+    int pStart = 0;  // current basis size at start of extension
+    RealD betaRestart = 0.0; // coupling from previous restart
+
+    for (int iter = 0; iter < MaxIter; ++iter) {
+
+      // ----------------------------------------------------------------
+      // Step 1: extend from pStart steps to Nm steps
+      // ----------------------------------------------------------------
+      extendBasis(pStart, Nm, betaRestart);
+//      verify();
+
+      // ----------------------------------------------------------------
+      // Step 2: SVD of the Nm x Nm B matrix.
+      // iter=0 (pStart==0): B is exactly bidiagonal — use buildBidiagonal.
+      // iter>0 (pStart==Nk): after a thick restart, column restart_col of
+      // U^dag A V has extra off-diagonal entries captured by fvec; use
+      // buildFullB so the Ritz values and restart vectors are computed from
+      // the exact projected matrix A V = U B_full.
+      // ----------------------------------------------------------------
+      Eigen::MatrixXd B = (pStart == 0) ? buildBidiagonal(Nm) : buildFullB(Nm);
+      Eigen::JacobiSVD<Eigen::MatrixXd> svd(B,
+          Eigen::ComputeThinU | Eigen::ComputeThinV);
+
+      Eigen::VectorXd sigma = svd.singularValues();  // descending
+      Eigen::MatrixXd X     = svd.matrixU();          // Nm x Nm left SVecs of B
+      Eigen::MatrixXd Y     = svd.matrixV();          // Nm x Nm right SVecs of B
+
+      // If targeting smallest, reorder so desired ones come first
+      Eigen::VectorXi order = sortOrder(sigma);
+
+      // ----------------------------------------------------------------
+      // Step 3: check convergence of the Nk desired singular values
+      // ----------------------------------------------------------------
+      RealD betaK = beta.back();  // beta_{k+1}
+      // In our convention A V = U B (exact), the residual is in the A^dag
+      // direction: A^dag u_j - sigma_j v_j = betaK * X[Nm-1,j] * V[Nm].
+      // Convergence criterion: |betaK * X[Nm-1, idx]| < tol * sigma_idx.
+      int nconv = 0;
+      for (int i = 0; i < Nk; ++i) {
+        int idx = order(i);
+        RealD res = std::abs(betaK * X(Nm - 1, idx));
+        RealD thr = Tolerance * std::max(sigma(idx), 1.0e-14);
+        std::cout << GridLogMessage
+                  << "IRLBA iter " << iter
+                  << "  sigma[" << i << "] = " << sigma(idx)
+                  << "  res = " << res
+                  << "  thr = " << thr << std::endl;
+        if (res < thr) ++nconv;
+        else break;  // residuals not strictly ordered but break is conservative
+      }
+
+      if (nconv >= Nk) {
+        std::cout << GridLogMessage
+                  << "IRLBA converged: " << nconv << " singular values after "
+                  << iter + 1 << " iterations." << std::endl;
+        // Collect converged triplets
+        extractTriplets(Nm, sigma, X, Y, order, Nk);
+        return;
+      }
+
+      // ----------------------------------------------------------------
+      // Step 4: implicit restart — compress to Nk steps
+      // ----------------------------------------------------------------
+      implicitRestart(Nm, Nk, sigma, X, Y, order, betaK, betaRestart);
+//      verify();
+
+      // Lucky breakdown: exact invariant subspace found; convergence is exact.
+      // B_p^+ = diag(alpha[0..Nk-1]); extract directly from restart basis.
+      if (betaRestart < 1.0e-14) {
+        std::cout << GridLogMessage
+                  << "IRLBA: lucky breakdown after restart (betaRestart = 0)."
+                  << " Extracting " << Nk << " exact Ritz triplets." << std::endl;
+        // Re-run SVD on the p-step diagonal B^+ to get sorted Ritz triplets.
+        Eigen::MatrixXd Bp = buildBidiagonal(Nk);
+        Eigen::JacobiSVD<Eigen::MatrixXd> svdp(Bp,
+            Eigen::ComputeThinU | Eigen::ComputeThinV);
+        Eigen::VectorXi ordp = sortOrder(svdp.singularValues());
+        extractTriplets(Nk, svdp.singularValues(), svdp.matrixU(),
+                        svdp.matrixV(), ordp, Nk);
+        return;
+      }
+
+      pStart = Nk;
+    }
+
+    std::cout << GridLogMessage
+              << "IRLBA: did not converge in " << MaxIter
+              << " iterations. Returning best approximations." << std::endl;
+
+    // Return best available approximations
+    Eigen::MatrixXd B = buildFullB((int)alpha.size());
+    Eigen::JacobiSVD<Eigen::MatrixXd> svd(B,
+        Eigen::ComputeThinU | Eigen::ComputeThinV);
+    Eigen::VectorXd sigma = svd.singularValues();
+    Eigen::MatrixXd X     = svd.matrixU();
+    Eigen::MatrixXd Y     = svd.matrixV();
+    Eigen::VectorXi order = sortOrder(sigma);
+    int nout = std::min(Nk, (int)alpha.size());
+    extractTriplets((int)alpha.size(), sigma, X, Y, order, nout);
+  }
+
+  /* Getters */
+  int getNk() const { return (int)singularValues.size(); }
+  const std::vector<RealD>&  getSingularValues() const { return singularValues; }
+  const std::vector<Field>&  getLeftVectors()    const { return leftVectors; }
+  const std::vector<Field>&  getRightVectors()   const { return rightVectors; }
+
+  /**
+   * Print B_k and U^dag A V to verify the bidiagonalization relation
+   *   A V_m = U_m B_m   (exact in our GK convention)
+   * On the first call (pStart=0), max|B - U^dag A V| should be ~machine epsilon.
+   * After a restart and extension, the column p of U^dag A V deviates from B
+   * by O(betaK): this is expected because the thick restart breaks the Krylov
+   * structure at column p, introducing off-diagonal terms proportional to betaK.
+   * These terms vanish as betaK -> 0 (convergence), so the algorithm is correct.
+   */
+  void verify()
+  {
+    int m  = (int)alpha.size();
+    int nU = (int)U.size();
+    int nV = (int)V.size();
+    if (m == 0) { std::cout << GridLogMessage << "IRLBA verify: empty basis" << std::endl; return; }
+
+    // Build reference matrix Bref (nU x nV):
+    //   Columns 0..m-1 : buildFullB(m)  (bidiagonal + fvec column at restart_col)
+    //   Column  m      : residual column, two cases:
+    //     (a) restart_col == m (right after implicitRestart, before extendBasis):
+    //         V[m] = sgn*V_old[Nm], so <U[i]|A|V[m]> = fvec[i] for all i
+    //     (b) otherwise (pure GK or after extendBasis):
+    //         only entry (m-1, m) = beta[m-1]  (GK recurrence residual)
+    Eigen::MatrixXd Bref = Eigen::MatrixXd::Zero(nU, nV);
+    {
+      Eigen::MatrixXd Bfull = buildFullB(m);
+      int cols = std::min(m, nV);
+      Bref.block(0, 0, m, cols) = Bfull.block(0, 0, m, cols);
+    }
+    if (nV > m && m > 0) {
+      if (restart_col == m && (int)fvec.size() == m) {
+        // Case (a): right after implicitRestart
+        for (int i = 0; i < m; ++i) Bref(i, m) = fvec[i];
+      } else if ((int)beta.size() >= m) {
+        // Case (b): standard GK residual column
+        Bref(m - 1, m) = beta[m - 1];
+      }
+    }
+
+    // Compute M[i,j] = <U[i] | A | V[j]>
+    Eigen::MatrixXd M = Eigen::MatrixXd::Zero(nU, nV);
+    Field Avj(Grid);
+    for (int j = 0; j < nV; ++j) {
+      Linop.Op(V[j], Avj);
+      for (int i = 0; i < nU; ++i) {
+        ComplexD ip = innerProduct(U[i], Avj);
+        M(i, j) = ip.real();
+      }
+    }
+
+    // Print Bref
+    std::cout << GridLogMessage
+              << "IRLBA verify: Bref (" << nU << "x" << nV << "):" << std::endl;
+    for (int i = 0; i < nU; ++i) {
+      std::cout << GridLogMessage << "  row " << i << ": ";
+      for (int j = 0; j < nV; ++j) std::cout << Bref(i,j) << " ";
+      std::cout << std::endl;
+    }
+
+    // Print U^dag A V
+    std::cout << GridLogMessage
+              << "IRLBA verify: U^dag A V (" << nU << "x" << nV << "):" << std::endl;
+    for (int i = 0; i < nU; ++i) {
+      std::cout << GridLogMessage << "  row " << i << ": ";
+      for (int j = 0; j < nV; ++j) std::cout << M(i,j) << " ";
+      std::cout << std::endl;
+    }
+
+    // Max deviation over the full nU x nV matrix
+    RealD maxdev = (Bref - M).cwiseAbs().maxCoeff();
+    std::cout << GridLogMessage
+              << "IRLBA verify: max|Bref - U^dag A V| = " << maxdev << std::endl;
+
+    // Beta
+    std::cout << GridLogMessage << "IRLBA verify: beta[0.." << (int)beta.size()-1 << "] = ";
+    for (auto b : beta) std::cout << b << " ";
+    std::cout << std::endl;
+  }
+
+private:
+
+  // ------------------------------------------------------------------
+  // Build the m x m upper-bidiagonal matrix from alpha[0..m-1], beta[0..m-2]
+  // ------------------------------------------------------------------
+  Eigen::MatrixXd buildBidiagonal(int m) const
+  {
+    Eigen::MatrixXd B = Eigen::MatrixXd::Zero(m, m);
+    for (int k = 0; k < m; ++k) {
+      B(k, k) = alpha[k];
+      if (k + 1 < m && k < (int)beta.size())
+        B(k, k + 1) = beta[k];
+    }
+    return B;
+  }
+
+  // ------------------------------------------------------------------
+  // Build the full m x m B matrix, including the non-bidiagonal column
+  // at restart_col that arises after a thick restart.
+  //
+  // After restart, A V[restart_col] has projections onto all U[0..restart_col-1]
+  // (not just U[restart_col-1]).  These are stored in fvec[0..restart_col-1]
+  // and make column restart_col of U^dag A V non-bidiagonal.
+  // ------------------------------------------------------------------
+  Eigen::MatrixXd buildFullB(int m) const
+  {
+    Eigen::MatrixXd B = buildBidiagonal(m);
+    if (restart_col >= 0 && restart_col < m && (int)fvec.size() > 0) {
+      for (int j = 0; j < restart_col && j < (int)fvec.size(); ++j){
+        B(j, restart_col) = fvec[j];
+        std::cout << GridLogDebug << "buildFullB: B  " <<j<<" "<<restart_col<<B(j, restart_col)<<std::endl;
+      }
+    }
+    return B;
+  }
+
+  // ------------------------------------------------------------------
+  // Return a permutation vector that puts the desired Nk singular values
+  // first (largest first if largest==true, smallest first otherwise).
+  // Eigen's JacobiSVD already returns sigma in descending order, so for
+  // largest we just return 0,1,...,m-1; for smallest we reverse.
+  // ------------------------------------------------------------------
+  Eigen::VectorXi sortOrder(const Eigen::VectorXd &sigma) const
+  {
+    int m = (int)sigma.size();
+    Eigen::VectorXi ord(m);
+    if (largest) {
+      for (int i = 0; i < m; ++i) ord(i) = i;
+    } else {
+      for (int i = 0; i < m; ++i) ord(i) = m - 1 - i;
+    }
+    return ord;
+  }
+
+  // ------------------------------------------------------------------
+  // Extend the Lanczos bidiagonalization from pStart to kEnd steps.
+  // On first call pStart==0 (V[0] already set).
+  // On restart calls V[0..pStart], U[0..pStart-1], alpha[0..pStart-1],
+  // beta[0..pStart-1] are already set; betaRestart is the coupling
+  // beta_{pStart} that drives the first new U step.
+  // ------------------------------------------------------------------
+  void extendBasis(int pStart, int kEnd, RealD betaRestart)
+  {
+    // Truncate containers to pStart (Lattice has no default constructor)
+    if ((int)V.size() > pStart + 1) V.erase(V.begin() + pStart + 1, V.end());
+    if ((int)U.size() > pStart)     U.erase(U.begin() + pStart,     U.end());
+    alpha.resize(pStart);
+    beta.resize(pStart);
+
+    Field p(Grid), r(Grid);
+
+    for (int k = pStart; k < kEnd; ++k) {
+
+      // p = A v_k
+      Linop.Op(V[k], p);
+
+      // Remove previous left vector coupling
+      if (k > 0) {
+        p = p - beta[k - 1] * U[k - 1];
+      }
+      // On the first step after a restart, beta[pStart-1] was already set;
+      // but V[pStart] was already constructed including the beta correction,
+      // so no extra subtraction needed here beyond the standard recurrence.
+
+      // Reorthogonalize p against U, then alpha_k = ||p||, u_k = p/alpha_k
+      reorthogonalize(p, U);
+      RealD ak = std::sqrt(norm2(p));
+      if (ak < 1.0e-14) {
+        std::cout << GridLogMessage
+                  << "IRLBA extendBasis: lucky breakdown at step " << k
+                  << " (alpha = " << ak << ")" << std::endl;
+        alpha.push_back(ak);
+        Field zero(Grid); zero = Zero();
+        U.push_back(zero);
+        beta.push_back(0.0);
+        V.push_back(zero);
+        break;
+      }
+      alpha.push_back(ak);
+
+      Field u(Grid);
+      u = (1.0 / ak) * p;
+      U.push_back(u);
+
+      // r = A^dag u_k - alpha_k v_k, reorthogonalize, then beta_{k+1} = ||r||
+      Linop.AdjOp(U[k], r);
+      r = r - ak * V[k];
+      reorthogonalize(r, V);
+
+      RealD bk = std::sqrt(norm2(r));
+      beta.push_back(bk);
+
+      std::cout << GridLogMessage
+                << "IRLBA extend step " << k
+                << "  alpha = " << ak
+                << "  beta  = " << bk << std::endl;
+
+      // Always push v_{k+1} (needed as residual direction for restart)
+      if (bk < 1.0e-14) {
+        std::cout << GridLogMessage
+                  << "IRLBA extendBasis: lucky breakdown (beta = 0) at step "
+                  << k << std::endl;
+        Field zero(Grid); zero = Zero();
+        V.push_back(zero);
+        break;
+      }
+      Field vnext(Grid);
+      vnext = (1.0 / bk) * r;
+      V.push_back(vnext);
+
+      if (k == kEnd - 1) break;  // v_{k+1} pushed above; stop here
+    }
+  }
+
+public:
+  // ------------------------------------------------------------------
+  // Block reorthogonalization helpers.
+  // Declared public because CUDA extended lambdas cannot live inside
+  // private/protected member functions.
+  //
+  // batchInnerProducts: computes c[j] = <basis[j], vec> for all j
+  //   in a single GPU pass (one accelerator_barrier instead of n).
+  //   Queues n pairs of (per-site kernel, reduceKernel) to computeStream
+  //   without intermediate CPU syncs, then syncs once at the end.
+  //
+  // batchUpdate: computes vec -= sum_j c[j]*basis[j] in one GPU kernel.
+  //
+  // reorthogonalize: two-pass Classical Gram-Schmidt (CGS2) using the
+  //   two helpers above.  Each pass costs 2 GPU syncs (1 IP + 1 update)
+  //   instead of 2n syncs per pass in the old sequential MGS.
+  // ------------------------------------------------------------------
+
+  void batchInnerProducts(const Field &vec,
+                          const std::vector<Field> &basis,
+                          std::vector<ComplexD> &c)
+  {
+    int n = (int)basis.size();
+    c.resize(n);
+    if (n == 0) return;
+
+    typedef typename Field::vector_object         vobj;
+    typedef decltype(innerProduct(vobj(), vobj())) inner_t;
+    typedef decltype(basis[0].View(AcceleratorRead)) View;
+
+    GridBase *grid = vec.Grid();
+    uint64_t oSites = grid->oSites();
+    uint64_t nsimd  = grid->Nsimd();
+
+    // all_ip[j * oSites + ss] = per-site inner product of basis[j] and vec at site ss.
+    // Layout: n contiguous blocks of oSites each.
+    deviceVector<inner_t> all_ip((uint64_t)n * oSites);
+    inner_t *all_ip_p = &all_ip[0];
+
+    hostVector<View>   h_basis_v(n);
+    deviceVector<View> d_basis_v(n);
+    for (int j = 0; j < n; ++j) {
+      h_basis_v[j] = basis[j].View(AcceleratorRead);
+      acceleratorPut(d_basis_v[j], h_basis_v[j]);
+    }
+    View *basis_vp = &d_basis_v[0];
+
+    // Queue n per-site kernels to the accelerator stream — no intermediate barriers.
+    autoView(vec_v, vec, AcceleratorRead);
+    for (int j = 0; j < n; ++j) {
+      int      jj      = j;
+      uint64_t oSites_ = oSites;
+      accelerator_for(ss, oSites, nsimd, {
+        auto x = coalescedRead(basis_vp[jj][ss]);
+        auto y = coalescedRead(vec_v[ss]);
+        coalescedWrite(all_ip_p[jj * oSites_ + ss], innerProduct(x, y));
+      });
+    }
+
+    // ONE sync after all n kernels
+    accelerator_barrier();
+
+    // Copy all per-site results to host
+    hostVector<inner_t> all_ip_h((uint64_t)n * oSites);
+    acceleratorCopyFromDevice(all_ip_p, &all_ip_h[0], (uint64_t)n * oSites * sizeof(inner_t));
+
+    // Reduce on host: sum over oSites, then collapse SIMD lanes via Reduce(TensorRemove(...))
+    // TensorRemove strips the iSinglet tensor wrapper to expose the SIMD scalar type.
+    // Reduce sums all nsimd lanes and returns a plain scalar (RealD or ComplexD).
+    std::vector<ComplexD> raw(n);
+    for (int j = 0; j < n; ++j) {
+      inner_t sum = Zero();
+      for (uint64_t ss = 0; ss < oSites; ++ss)
+        sum += all_ip_h[(uint64_t)j * oSites + ss];
+      raw[j] = ComplexD(Reduce(TensorRemove(sum)));
+    }
+    grid->GlobalSumVector(&raw[0], n);
+    for (int j = 0; j < n; ++j) c[j] = raw[j];
+
+    for (int j = 0; j < n; ++j) h_basis_v[j].ViewClose();
+  }
+
+  void batchUpdate(Field &vec,
+                   const std::vector<Field> &basis,
+                   const std::vector<ComplexD> &c)
+  {
+    int n = (int)basis.size();
+    if (n == 0) return;
+
+    typedef typename Field::vector_object vobj;
+    typedef decltype(basis[0].View(AcceleratorRead)) View;
+
+    GridBase *grid = vec.Grid();
+    uint64_t oSites = grid->oSites();
+    uint64_t nsimd  = grid->Nsimd();
+
+    // Split complex coefficients into real/imag double arrays on device.
+    // Using doubles avoids potential ComplexD-device-code compatibility issues.
+    hostVector<double>   h_re(n), h_im(n);
+    deviceVector<double> d_re(n), d_im(n);
+    for (int k = 0; k < n; ++k) {
+      h_re[k] = c[k].real();
+      h_im[k] = c[k].imag();
+    }
+    acceleratorCopyToDevice(&h_re[0], &d_re[0], n * sizeof(double));
+    acceleratorCopyToDevice(&h_im[0], &d_im[0], n * sizeof(double));
+    double *re_p = &d_re[0];
+    double *im_p = &d_im[0];
+
+    // Basis views
+    hostVector<View>   h_basis_v(n);
+    deviceVector<View> d_basis_v(n);
+    for (int k = 0; k < n; ++k) {
+      h_basis_v[k] = basis[k].View(AcceleratorRead);
+      acceleratorPut(d_basis_v[k], h_basis_v[k]);
+    }
+    View *basis_vp = &d_basis_v[0];
+
+    // Single kernel: vec[ss] -= sum_k (re[k] + i*im[k]) * basis[k][ss]
+    autoView(vec_v, vec, AcceleratorWrite);
+    accelerator_for(ss, oSites, nsimd, {
+      auto v = coalescedRead(vec_v[ss]);
+      for (int k = 0; k < n; ++k) {
+        auto b = coalescedRead(basis_vp[k][ss]);
+        v = v - re_p[k] * b - timesI(im_p[k] * b);
+      }
+      coalescedWrite(vec_v[ss], v);
+    });
+
+    for (int k = 0; k < n; ++k) h_basis_v[k].ViewClose();
+  }
+
+  // ------------------------------------------------------------------
+  // Full reorthogonalization using two-pass Classical Gram-Schmidt (CGS2).
+  // Each pass calls batchInnerProducts (1 GPU sync) + batchUpdate (1 sync),
+  // replacing the old 2n GPU syncs per pass from sequential MGS.
+  // ------------------------------------------------------------------
+  void reorthogonalize(Field &vec, const std::vector<Field> &basis)
+  {
+    if (basis.empty()) return;
+    std::vector<ComplexD> c;
+    for (int pass = 0; pass < 2; ++pass) {
+      batchInnerProducts(vec, basis, c);
+      batchUpdate(vec, basis, c);
+    }
+  }
+
+  // ------------------------------------------------------------------
+  // Implicit restart: given the Nm-step bidiagonalization and its SVD,
+  // compress to Nk steps via implicit QR shifts applied to B_k.
+  //
+  // The "shifts" are the Nm - Nk singular values we want to deflate
+  // (those NOT in the desired set).  We apply them as implicit QR steps
+  // to the bidiagonal matrix, then update the lattice bases accordingly.
+  //
+  // After this call:
+  //   V[0..Nk],  U[0..Nk-1],  alpha[0..Nk-1],  beta[0..Nk-1]  are updated.
+  //   betaRestart  ← new beta_Nk coupling for the next extension.
+  // ------------------------------------------------------------------
+  void implicitRestart(int k, int p,
+                       const Eigen::VectorXd &sigma,
+                       const Eigen::MatrixXd &X,
+                       const Eigen::MatrixXd &Y,
+                       const Eigen::VectorXi &order,
+                       RealD betaK,
+                       RealD &betaRestart)
+  {
+    // Thick restart (Baglama & Reichel, Sec. 2.2):
+    //
+    // Given B_k = X Sigma Y^T, define the new p-step basis by:
+    //   V^+_i = V_k * y_{order(i)}      (right sing. vec. of B_k)
+    //   U^+_i = U_k * x_{order(i)}      (left  sing. vec. of B_k)
+    //
+    // Then A V^+_i = A V_k y_{order(i)} = U_k B_k y_{order(i)}
+    //             = sigma_{order(i)} U_k x_{order(i)} = sigma_{order(i)} U^+_i
+    //
+    // So B_p^+ = diag(sigma_{order(0)}, ..., sigma_{order(p-1)}) — DIAGONAL,
+    // all internal betas are zero.
+    //
+    // The residual coupling comes from A^dag U_k = V_k B_k^T + betaK V[k] e_{k-1}^T:
+    //   A^dag U^+_{p-1} - sigma_{order(p-1)} V^+_{p-1}
+    //     = V_k (B_k^T x_{order(p-1)} - sigma_{order(p-1)} y_{order(p-1)})
+    //       + betaK * X(k-1, order(p-1)) * V[k]
+    //     = betaK * X(k-1, order(p-1)) * V[k]   (since B_k^T x_j = sigma_j y_j)
+    //
+    // Therefore: betaRestart = |betaK * X(k-1, order(p-1))|
+    //            V[p] = sign(X(k-1, order(p-1))) * V[k]
+
+    // ---- Build new lattice vectors ----
+    std::vector<Field> Vnew, Unew;
+    Vnew.reserve(p + 1);
+    Unew.reserve(p);
+
+    for (int i = 0; i < p; ++i) {
+      int idx = order(i);
+      Field vi(Grid); vi = Zero();
+      for (int j = 0; j < k; ++j)
+        vi = vi + Y(j, idx) * V[j];
+      Vnew.push_back(vi);
+    }
+
+    for (int i = 0; i < p; ++i) {
+      int idx = order(i);
+      Field ui(Grid); ui = Zero();
+      for (int j = 0; j < k; ++j)
+        ui = ui + X(j, idx) * U[j];
+      Unew.push_back(ui);
+    }
+
+    // New v_{p} (0-indexed: V[p]) = sign * V[k]
+    // From A^dag U_k = V_k B_k^T + betaK V[k] e_{k-1}^T:
+    //   A^dag U^+_j - sigma_j V^+_j = betaK * X(k-1, order(j)) * V[k]
+    // The last Ritz pair (j=p-1) defines betaRestart and the sign of V[p].
+    // All p couplings (j=0..p-1) are stored in fvec so that buildFullB can
+    // reconstruct the exact column p of U^dag A V after the next extension.
+    RealD coeff = betaK * X(k - 1, order(p - 1));
+    betaRestart  = std::abs(coeff);
+    RealD sgn = (coeff >= 0.0) ? 1.0 : -1.0;
+
+    fvec.resize(p);
+    for (int j = 0; j < p; ++j)
+      fvec[j] = betaK * X(k - 1, order(j)) * sgn;
+    // fvec[p-1] == betaRestart by construction
+    restart_col = p;
+
+    Field vp(Grid);
+    if (betaRestart > 1.0e-14) {
+      vp = sgn * V[k];
+    } else {
+      betaRestart = 0.0;
+      vp = Zero();
+    }
+    Vnew.push_back(vp);  // V[p]
+
+    // ---- New alpha, beta ----
+    // B_p^+ is diagonal: alpha^+_i = sigma_{order(i)}, all internal beta = 0
+    std::vector<RealD> alpha_new(p), beta_new(p);
+    for (int i = 0; i < p; ++i) alpha_new[i] = sigma(order(i));
+    for (int i = 0; i < p - 1; ++i) beta_new[i] = 0.0;
+    beta_new[p - 1] = betaRestart;
+
+    // ---- Commit new state ----
+    V = Vnew;
+    U = Unew;
+    alpha = alpha_new;
+    beta  = beta_new;
+
+    std::cout << GridLogMessage
+              << "IRLBA restart: compressed to " << p << " steps,"
+              << "  new beta_p = " << betaRestart << std::endl;
+  }
+
+  // ------------------------------------------------------------------
+  // Extract the desired singular triplets into the public output vectors.
+  // ------------------------------------------------------------------
+  void extractTriplets(int m,
+                       const Eigen::VectorXd &sigma,
+                       const Eigen::MatrixXd &X,
+                       const Eigen::MatrixXd &Y,
+                       const Eigen::VectorXi &order,
+                       int nout)
+  {
+    singularValues.resize(nout);
+    leftVectors.clear();   leftVectors.reserve(nout);
+    rightVectors.clear();  rightVectors.reserve(nout);
+
+    for (int i = 0; i < nout; ++i) {
+      int idx = order(i);
+      singularValues[i] = sigma(idx);
+
+      // Left singular vector of A:  svec_L = U_m * x_i
+      Field svL(Grid); svL = Zero();
+      for (int j = 0; j < m && j < (int)U.size(); ++j)
+        svL = svL + X(j, idx) * U[j];
+      leftVectors.push_back(svL);
+
+      // Right singular vector of A:  svec_R = V_m * y_i
+      Field svR(Grid); svR = Zero();
+      for (int j = 0; j < m && j < (int)V.size(); ++j)
+        svR = svR + Y(j, idx) * V[j];
+      rightVectors.push_back(svR);
+    }
+  }
+};
+
+NAMESPACE_END(Grid);
+#endif

Grid/algorithms/iterative/SchurRedBlack.h

@@ -327,9 +327,9 @@ namespace Grid {
       /////////////////////////////////////////////////////
       // src_o = (source_o - Moe MeeInv source_e)
       /////////////////////////////////////////////////////
-      _Matrix.MooeeInv(src_e,tmp);     assert(  tmp.Checkerboard() ==Even);
+      _Matrix.MooeeInv(src_e,tmp);     GRID_ASSERT(  tmp.Checkerboard() ==Even);
-      _Matrix.Meooe   (tmp,Mtmp);      assert( Mtmp.Checkerboard() ==Odd);     
+      _Matrix.Meooe   (tmp,Mtmp);      GRID_ASSERT( Mtmp.Checkerboard() ==Odd);     
-      tmp=src_o-Mtmp;                  assert(  tmp.Checkerboard() ==Odd);     
+      tmp=src_o-Mtmp;                  GRID_ASSERT(  tmp.Checkerboard() ==Odd);     
 
       _Matrix.Mooee(tmp,src_o); // Extra factor of "m" in source from dumb choice of matrix norm.
     }
@@ -347,17 +347,17 @@ namespace Grid {
       ///////////////////////////////////////////////////
       // sol_e = M_ee^-1 * ( src_e - Meo sol_o )...
       ///////////////////////////////////////////////////
-      _Matrix.Meooe(sol_o,tmp);        assert(  tmp.Checkerboard()   ==Even);
+      _Matrix.Meooe(sol_o,tmp);        GRID_ASSERT(  tmp.Checkerboard()   ==Even);
-      src_e = src_e-tmp;               assert(  src_e.Checkerboard() ==Even);
+      src_e = src_e-tmp;               GRID_ASSERT(  src_e.Checkerboard() ==Even);
-      _Matrix.MooeeInv(src_e,sol_e);   assert(  sol_e.Checkerboard() ==Even);
+      _Matrix.MooeeInv(src_e,sol_e);   GRID_ASSERT(  sol_e.Checkerboard() ==Even);
      
-      setCheckerboard(sol,sol_e); assert(  sol_e.Checkerboard() ==Even);
+      setCheckerboard(sol,sol_e); GRID_ASSERT(  sol_e.Checkerboard() ==Even);
-      setCheckerboard(sol,sol_o); assert(  sol_o.Checkerboard() ==Odd );
+      setCheckerboard(sol,sol_o); GRID_ASSERT(  sol_o.Checkerboard() ==Odd );
     }
     virtual void RedBlackSolve   (Matrix & _Matrix,const Field &src_o, Field &sol_o)
     {
       SchurStaggeredOperator<Matrix,Field> _HermOpEO(_Matrix);
-      this->_HermitianRBSolver(_HermOpEO,src_o,sol_o);  assert(sol_o.Checkerboard()==Odd);
+      this->_HermitianRBSolver(_HermOpEO,src_o,sol_o);  GRID_ASSERT(sol_o.Checkerboard()==Odd);
     };
     virtual void RedBlackSolve   (Matrix & _Matrix,const std::vector<Field> &src_o,  std::vector<Field> &sol_o)
     {
@@ -396,13 +396,13 @@ namespace Grid {
       /////////////////////////////////////////////////////
       // src_o = Mdag * (source_o - Moe MeeInv source_e)
       /////////////////////////////////////////////////////
-      _Matrix.MooeeInv(src_e,tmp);     assert(  tmp.Checkerboard() ==Even);
+      _Matrix.MooeeInv(src_e,tmp);     GRID_ASSERT(  tmp.Checkerboard() ==Even);
-      _Matrix.Meooe   (tmp,Mtmp);      assert( Mtmp.Checkerboard() ==Odd);     
+      _Matrix.Meooe   (tmp,Mtmp);      GRID_ASSERT( Mtmp.Checkerboard() ==Odd);     
-      tmp=src_o-Mtmp;                  assert(  tmp.Checkerboard() ==Odd);     
+      tmp=src_o-Mtmp;                  GRID_ASSERT(  tmp.Checkerboard() ==Odd);     
 
       // get the right MpcDag
       SchurDiagMooeeOperator<Matrix,Field> _HermOpEO(_Matrix);
-      _HermOpEO.MpcDag(tmp,src_o);     assert(src_o.Checkerboard() ==Odd);       
+      _HermOpEO.MpcDag(tmp,src_o);     GRID_ASSERT(src_o.Checkerboard() ==Odd);       
 
     }
     virtual void RedBlackSolution(Matrix & _Matrix,const Field &sol_o, const Field &src_e,Field &sol)
@@ -416,17 +416,17 @@ namespace Grid {
       ///////////////////////////////////////////////////
       // sol_e = M_ee^-1 * ( src_e - Meo sol_o )...
       ///////////////////////////////////////////////////
-      _Matrix.Meooe(sol_o,tmp);          assert(  tmp.Checkerboard()   ==Even);
+      _Matrix.Meooe(sol_o,tmp);          GRID_ASSERT(  tmp.Checkerboard()   ==Even);
-      src_e_i = src_e-tmp;               assert(  src_e_i.Checkerboard() ==Even);
+      src_e_i = src_e-tmp;               GRID_ASSERT(  src_e_i.Checkerboard() ==Even);
-      _Matrix.MooeeInv(src_e_i,sol_e);   assert(  sol_e.Checkerboard() ==Even);
+      _Matrix.MooeeInv(src_e_i,sol_e);   GRID_ASSERT(  sol_e.Checkerboard() ==Even);
      
-      setCheckerboard(sol,sol_e); assert(  sol_e.Checkerboard() ==Even);
+      setCheckerboard(sol,sol_e); GRID_ASSERT(  sol_e.Checkerboard() ==Even);
-      setCheckerboard(sol,sol_o); assert(  sol_o.Checkerboard() ==Odd );
+      setCheckerboard(sol,sol_o); GRID_ASSERT(  sol_o.Checkerboard() ==Odd );
     }
     virtual void RedBlackSolve   (Matrix & _Matrix,const Field &src_o, Field &sol_o)
     {
       SchurDiagMooeeOperator<Matrix,Field> _HermOpEO(_Matrix);
-      this->_HermitianRBSolver(_HermOpEO,src_o,sol_o);  assert(sol_o.Checkerboard()==Odd);
+      this->_HermitianRBSolver(_HermOpEO,src_o,sol_o);  GRID_ASSERT(sol_o.Checkerboard()==Odd);
     };
     virtual void RedBlackSolve   (Matrix & _Matrix,const std::vector<Field> &src_o,  std::vector<Field> &sol_o)
     {
@@ -461,9 +461,9 @@ namespace Grid {
         /////////////////////////////////////////////////////
         // src_o = Mdag * (source_o - Moe MeeInv source_e)
         /////////////////////////////////////////////////////
-        _Matrix.MooeeInv(src_e, tmp);   assert(   tmp.Checkerboard() == Even );
+        _Matrix.MooeeInv(src_e, tmp);   GRID_ASSERT(   tmp.Checkerboard() == Even );
-        _Matrix.Meooe   (tmp, Mtmp);    assert(  Mtmp.Checkerboard() == Odd  );     
+        _Matrix.Meooe   (tmp, Mtmp);    GRID_ASSERT(  Mtmp.Checkerboard() == Odd  );     
-        src_o -= Mtmp;                  assert( src_o.Checkerboard() == Odd  );     
+        src_o -= Mtmp;                  GRID_ASSERT( src_o.Checkerboard() == Odd  );     
       }
       
       virtual void RedBlackSolution(Matrix& _Matrix, const Field& sol_o, const Field& src_e, Field& sol)
@@ -478,18 +478,18 @@ namespace Grid {
         ///////////////////////////////////////////////////
         // sol_e = M_ee^-1 * ( src_e - Meo sol_o )...
         ///////////////////////////////////////////////////
-        _Matrix.Meooe(sol_o, tmp);         assert(     tmp.Checkerboard() == Even );
+        _Matrix.Meooe(sol_o, tmp);         GRID_ASSERT(     tmp.Checkerboard() == Even );
-        src_e_i = src_e - tmp;             assert( src_e_i.Checkerboard() == Even );
+        src_e_i = src_e - tmp;             GRID_ASSERT( src_e_i.Checkerboard() == Even );
-        _Matrix.MooeeInv(src_e_i, sol_e);  assert(   sol_e.Checkerboard() == Even );
+        _Matrix.MooeeInv(src_e_i, sol_e);  GRID_ASSERT(   sol_e.Checkerboard() == Even );
        
-        setCheckerboard(sol, sol_e); assert( sol_e.Checkerboard() == Even );
+        setCheckerboard(sol, sol_e); GRID_ASSERT( sol_e.Checkerboard() == Even );
-        setCheckerboard(sol, sol_o); assert( sol_o.Checkerboard() == Odd  );
+        setCheckerboard(sol, sol_o); GRID_ASSERT( sol_o.Checkerboard() == Odd  );
       }
 
       virtual void RedBlackSolve(Matrix& _Matrix, const Field& src_o, Field& sol_o)
       {
         NonHermitianSchurDiagMooeeOperator<Matrix,Field> _OpEO(_Matrix);
-        this->_HermitianRBSolver(_OpEO, src_o, sol_o);  assert(sol_o.Checkerboard() == Odd);
+        this->_HermitianRBSolver(_OpEO, src_o, sol_o);  GRID_ASSERT(sol_o.Checkerboard() == Odd);
       }
 
       virtual void RedBlackSolve(Matrix& _Matrix, const std::vector<Field>& src_o, std::vector<Field>& sol_o)
@@ -539,13 +539,13 @@ namespace Grid {
       /////////////////////////////////////////////////////
       // src_o = Mpcdag *MooeeInv * (source_o - Moe MeeInv source_e)
       /////////////////////////////////////////////////////
-      _Matrix.MooeeInv(src_e,tmp);     assert(  tmp.Checkerboard() ==Even);
+      _Matrix.MooeeInv(src_e,tmp);     GRID_ASSERT(  tmp.Checkerboard() ==Even);
-      _Matrix.Meooe   (tmp,Mtmp);      assert( Mtmp.Checkerboard() ==Odd);     
+      _Matrix.Meooe   (tmp,Mtmp);      GRID_ASSERT( Mtmp.Checkerboard() ==Odd);     
       Mtmp=src_o-Mtmp;                 
-      _Matrix.MooeeInv(Mtmp,tmp);      assert( tmp.Checkerboard() ==Odd);     
+      _Matrix.MooeeInv(Mtmp,tmp);      GRID_ASSERT( tmp.Checkerboard() ==Odd);     
       
       // get the right MpcDag
-      _HermOpEO.MpcDag(tmp,src_o);     assert(src_o.Checkerboard() ==Odd);       
+      _HermOpEO.MpcDag(tmp,src_o);     GRID_ASSERT(src_o.Checkerboard() ==Odd);       
     }
 
     virtual void RedBlackSolution(Matrix & _Matrix,const Field &sol_o, const Field &src_e,Field &sol)
@@ -560,12 +560,12 @@ namespace Grid {
       ///////////////////////////////////////////////////
       // sol_e = M_ee^-1 * ( src_e - Meo sol_o )...
       ///////////////////////////////////////////////////
-      _Matrix.Meooe(sol_o,tmp);    assert(  tmp.Checkerboard()   ==Even);
+      _Matrix.Meooe(sol_o,tmp);    GRID_ASSERT(  tmp.Checkerboard()   ==Even);
-      tmp = src_e-tmp;             assert(  src_e.Checkerboard() ==Even);
+      tmp = src_e-tmp;             GRID_ASSERT(  src_e.Checkerboard() ==Even);
-      _Matrix.MooeeInv(tmp,sol_e); assert(  sol_e.Checkerboard() ==Even);
+      _Matrix.MooeeInv(tmp,sol_e); GRID_ASSERT(  sol_e.Checkerboard() ==Even);
      
-      setCheckerboard(sol,sol_e);  assert(  sol_e.Checkerboard() ==Even);
+      setCheckerboard(sol,sol_e);  GRID_ASSERT(  sol_e.Checkerboard() ==Even);
-      setCheckerboard(sol,sol_o);  assert(  sol_o.Checkerboard() ==Odd );
+      setCheckerboard(sol,sol_o);  GRID_ASSERT(  sol_o.Checkerboard() ==Odd );
     };
 
     virtual void RedBlackSolve   (Matrix & _Matrix,const Field &src_o, Field &sol_o)
@@ -612,12 +612,12 @@ namespace Grid {
       /////////////////////////////////////////////////////
       // src_o = Mdag * (source_o - Moe MeeInv source_e)
       /////////////////////////////////////////////////////
-      _Matrix.MooeeInv(src_e,tmp);     assert(  tmp.Checkerboard() ==Even);
+      _Matrix.MooeeInv(src_e,tmp);     GRID_ASSERT(  tmp.Checkerboard() ==Even);
-      _Matrix.Meooe   (tmp,Mtmp);      assert( Mtmp.Checkerboard() ==Odd);     
+      _Matrix.Meooe   (tmp,Mtmp);      GRID_ASSERT( Mtmp.Checkerboard() ==Odd);     
-      tmp=src_o-Mtmp;                  assert(  tmp.Checkerboard() ==Odd);     
+      tmp=src_o-Mtmp;                  GRID_ASSERT(  tmp.Checkerboard() ==Odd);     
 
       // get the right MpcDag
-      _HermOpEO.MpcDag(tmp,src_o);     assert(src_o.Checkerboard() ==Odd);       
+      _HermOpEO.MpcDag(tmp,src_o);     GRID_ASSERT(src_o.Checkerboard() ==Odd);       
     }
 
     virtual void RedBlackSolution(Matrix & _Matrix,const Field &sol_o, const Field &src_e,Field &sol)
@@ -638,12 +638,12 @@ namespace Grid {
       ///////////////////////////////////////////////////
       // sol_e = M_ee^-1 * ( src_e - Meo sol_o )...
       ///////////////////////////////////////////////////
-      _Matrix.Meooe(sol_o_i,tmp);    assert(  tmp.Checkerboard()   ==Even);
+      _Matrix.Meooe(sol_o_i,tmp);    GRID_ASSERT(  tmp.Checkerboard()   ==Even);
-      tmp = src_e-tmp;               assert(  src_e.Checkerboard() ==Even);
+      tmp = src_e-tmp;               GRID_ASSERT(  src_e.Checkerboard() ==Even);
-      _Matrix.MooeeInv(tmp,sol_e);   assert(  sol_e.Checkerboard() ==Even);
+      _Matrix.MooeeInv(tmp,sol_e);   GRID_ASSERT(  sol_e.Checkerboard() ==Even);
      
-      setCheckerboard(sol,sol_e);    assert(  sol_e.Checkerboard() ==Even);
+      setCheckerboard(sol,sol_e);    GRID_ASSERT(  sol_e.Checkerboard() ==Even);
-      setCheckerboard(sol,sol_o_i);  assert(  sol_o_i.Checkerboard() ==Odd );
+      setCheckerboard(sol,sol_o_i);  GRID_ASSERT(  sol_o_i.Checkerboard() ==Odd );
     };
 
     virtual void RedBlackSolve   (Matrix & _Matrix,const Field &src_o, Field &sol_o)
@@ -684,9 +684,9 @@ namespace Grid {
         /////////////////////////////////////////////////////
         // src_o = Mdag * (source_o - Moe MeeInv source_e)
         /////////////////////////////////////////////////////
-        _Matrix.MooeeInv(src_e, tmp);   assert(   tmp.Checkerboard() == Even );
+        _Matrix.MooeeInv(src_e, tmp);   GRID_ASSERT(   tmp.Checkerboard() == Even );
-        _Matrix.Meooe   (tmp, Mtmp);    assert(  Mtmp.Checkerboard() == Odd  );     
+        _Matrix.Meooe   (tmp, Mtmp);    GRID_ASSERT(  Mtmp.Checkerboard() == Odd  );     
-        src_o -= Mtmp;                  assert( src_o.Checkerboard() == Odd  );     
+        src_o -= Mtmp;                  GRID_ASSERT( src_o.Checkerboard() == Odd  );     
       }
 
       virtual void RedBlackSolution(Matrix& _Matrix, const Field& sol_o, const Field& src_e, Field& sol)
@@ -707,12 +707,12 @@ namespace Grid {
         ///////////////////////////////////////////////////
         // sol_e = M_ee^-1 * ( src_e - Meo sol_o )...
         ///////////////////////////////////////////////////
-        _Matrix.Meooe(sol_o_i, tmp);    assert(   tmp.Checkerboard() == Even );
+        _Matrix.Meooe(sol_o_i, tmp);    GRID_ASSERT(   tmp.Checkerboard() == Even );
-        tmp = src_e - tmp;              assert( src_e.Checkerboard() == Even );
+        tmp = src_e - tmp;              GRID_ASSERT( src_e.Checkerboard() == Even );
-        _Matrix.MooeeInv(tmp, sol_e);   assert( sol_e.Checkerboard() == Even );
+        _Matrix.MooeeInv(tmp, sol_e);   GRID_ASSERT( sol_e.Checkerboard() == Even );
        
-        setCheckerboard(sol, sol_e);    assert(   sol_e.Checkerboard() == Even );
+        setCheckerboard(sol, sol_e);    GRID_ASSERT(   sol_e.Checkerboard() == Even );
-        setCheckerboard(sol, sol_o_i);  assert( sol_o_i.Checkerboard() == Odd  );
+        setCheckerboard(sol, sol_o_i);  GRID_ASSERT( sol_o_i.Checkerboard() == Odd  );
       };
 
       virtual void RedBlackSolve(Matrix& _Matrix, const Field& src_o, Field& sol_o)

Grid/algorithms/iterative/SimpleLanczos.h

@@ -0,0 +1,931 @@
+    /*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./lib/algorithms/iterative/ImplicitlyRestartedLanczos.h
+
+    Copyright (C) 2015
+
+Author: Chulwoo Jung <chulwoo@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 */
+#ifndef GRID_LANC_H
+#define GRID_LANC_H
+
+#include <string.h>		//memset
+
+#ifdef USE_LAPACK
+#ifdef USE_MKL
+#include<mkl_lapack.h>
+#else
+void LAPACK_dstegr (char *jobz, char *range, int *n, double *d, double *e,
+		    double *vl, double *vu, int *il, int *iu, double *abstol,
+		    int *m, double *w, double *z, int *ldz, int *isuppz,
+		    double *work, int *lwork, int *iwork, int *liwork,
+		    int *info);
+//#include <lapacke/lapacke.h>
+#endif
+#endif
+
+//#include <Grid/algorithms/densematrix/DenseMatrix.h>
+
+// eliminate temorary vector in calc()
+#define MEM_SAVE
+
+namespace Grid
+{
+
+  struct Bisection
+  {
+
+#if 0
+    static void get_eig2 (int row_num, std::vector < RealD > &ALPHA,
+			  std::vector < RealD > &BETA,
+			  std::vector < RealD > &eig)
+    {
+      int i, j;
+        std::vector < RealD > evec1 (row_num + 3);
+        std::vector < RealD > evec2 (row_num + 3);
+      RealD eps2;
+        ALPHA[1] = 0.;
+        BETHA[1] = 0.;
+      for (i = 0; i < row_num - 1; i++)
+	{
+	  ALPHA[i + 1] = A[i * (row_num + 1)].real ();
+	  BETHA[i + 2] = A[i * (row_num + 1) + 1].real ();
+	}
+      ALPHA[row_num] = A[(row_num - 1) * (row_num + 1)].real ();
+        bisec (ALPHA, BETHA, row_num, 1, row_num, 1e-10, 1e-10, evec1, eps2);
+        bisec (ALPHA, BETHA, row_num, 1, row_num, 1e-16, 1e-16, evec2, eps2);
+
+      // Do we really need to sort here?
+      int begin = 1;
+      int end = row_num;
+      int swapped = 1;
+      while (swapped)
+	{
+	  swapped = 0;
+	  for (i = begin; i < end; i++)
+	    {
+	      if (mag (evec2[i]) > mag (evec2[i + 1]))
+		{
+		  swap (evec2 + i, evec2 + i + 1);
+		  swapped = 1;
+		}
+	    }
+	  end--;
+	  for (i = end - 1; i >= begin; i--)
+	    {
+	      if (mag (evec2[i]) > mag (evec2[i + 1]))
+		{
+		  swap (evec2 + i, evec2 + i + 1);
+		  swapped = 1;
+		}
+	    }
+	  begin++;
+	}
+
+      for (i = 0; i < row_num; i++)
+	{
+	  for (j = 0; j < row_num; j++)
+	    {
+	      if (i == j)
+		H[i * row_num + j] = evec2[i + 1];
+	      else
+		H[i * row_num + j] = 0.;
+	    }
+	}
+    }
+#endif
+
+    static void bisec (std::vector < RealD > &c,
+		       std::vector < RealD > &b,
+		       int n,
+		       int m1,
+		       int m2,
+		       RealD eps1,
+		       RealD relfeh, std::vector < RealD > &x, RealD & eps2)
+    {
+      std::vector < RealD > wu (n + 2);
+
+      RealD h, q, x1, xu, x0, xmin, xmax;
+      int i, a, k;
+
+      b[1] = 0.0;
+      xmin = c[n] - fabs (b[n]);
+      xmax = c[n] + fabs (b[n]);
+      for (i = 1; i < n; i++)
+	{
+	  h = fabs (b[i]) + fabs (b[i + 1]);
+	  if (c[i] + h > xmax)
+	    xmax = c[i] + h;
+	  if (c[i] - h < xmin)
+	    xmin = c[i] - h;
+	}
+      xmax *= 2.;
+
+      eps2 = relfeh * ((xmin + xmax) > 0.0 ? xmax : -xmin);
+      if (eps1 <= 0.0)
+	eps1 = eps2;
+      eps2 = 0.5 * eps1 + 7.0 * (eps2);
+      x0 = xmax;
+      for (i = m1; i <= m2; i++)
+	{
+	  x[i] = xmax;
+	  wu[i] = xmin;
+	}
+
+      for (k = m2; k >= m1; k--)
+	{
+	  xu = xmin;
+	  i = k;
+	  do
+	    {
+	      if (xu < wu[i])
+		{
+		  xu = wu[i];
+		  i = m1 - 1;
+		}
+	      i--;
+	    }
+	  while (i >= m1);
+	  if (x0 > x[k])
+	    x0 = x[k];
+	  while ((x0 - xu) > 2 * relfeh * (fabs (xu) + fabs (x0)) + eps1)
+	    {
+	      x1 = (xu + x0) / 2;
+
+	      a = 0;
+	      q = 1.0;
+	      for (i = 1; i <= n; i++)
+		{
+		  q =
+		    c[i] - x1 -
+		    ((q != 0.0) ? b[i] * b[i] / q : fabs (b[i]) / relfeh);
+		  if (q < 0)
+		    a++;
+		}
+//      printf("x1=%0.14e a=%d\n",x1,a);
+	      if (a < k)
+		{
+		  if (a < m1)
+		    {
+		      xu = x1;
+		      wu[m1] = x1;
+		    }
+		  else
+		    {
+		      xu = x1;
+		      wu[a + 1] = x1;
+		      if (x[a] > x1)
+			x[a] = x1;
+		    }
+		}
+	      else
+		x0 = x1;
+	    }
+	  printf ("x0=%0.14e xu=%0.14e k=%d\n", x0, xu, k);
+	  x[k] = (x0 + xu) / 2;
+	}
+    }
+  };
+
+/////////////////////////////////////////////////////////////
+// Implicitly restarted lanczos
+/////////////////////////////////////////////////////////////
+
+
+  template < class Field > class SimpleLanczos
+  {
+
+    const RealD small = 1.0e-16;
+  public:
+    int lock;
+    int get;
+    int Niter;
+    int converged;
+
+    int Nstop;			// Number of evecs checked for convergence
+    int Nk;			// Number of converged sought
+    int Np;			// Np -- Number of spare vecs in kryloc space
+    int Nm;			// Nm -- total number of vectors
+
+
+    RealD OrthoTime;
+
+    RealD eresid;
+
+//    SortEigen < Field > _sort;
+
+    LinearFunction < Field > &_Linop;
+
+//    OperatorFunction < Field > &_poly;
+
+    /////////////////////////
+    // Constructor
+    /////////////////////////
+    void init (void)
+    {
+    };
+//    void Abort (int ff, std::vector < RealD > &evals, DenseVector < Denstd::vector  < RealD > >&evecs);
+
+    SimpleLanczos (LinearFunction < Field > &Linop,	// op
+//		   OperatorFunction < Field > &poly,	// polynmial
+		   int _Nstop,	// sought vecs
+		   int _Nk,	// sought vecs
+		   int _Nm,	// spare vecs
+		   RealD _eresid,	// resid in lmdue deficit 
+		   int _Niter):	// Max iterations
+     
+      _Linop (Linop),
+ //     _poly (poly),
+      Nstop (_Nstop), Nk (_Nk), Nm (_Nm), eresid (_eresid), Niter (_Niter)
+    {
+      Np = Nm - Nk;
+      assert (Np > 0);
+    };
+
+    /////////////////////////
+    // Sanity checked this routine (step) against Saad.
+    /////////////////////////
+    void RitzMatrix (std::vector < Field > &evec, int k)
+    {
+
+      if (1)
+	return;
+
+      GridBase *grid = evec[0].Grid();
+      Field w (grid);
+      std::cout << GridLogMessage << "RitzMatrix " << std::endl;
+      for (int i = 0; i < k; i++)
+	{
+	  _Linop(evec[i], w);
+//      _poly(_Linop,evec[i],w);
+	  std::cout << GridLogMessage << "[" << i << "] ";
+	  for (int j = 0; j < k; j++)
+	    {
+	      ComplexD in = innerProduct (evec[j], w);
+	      if (fabs ((double) i - j) > 1)
+		{
+		  if (abs (in) > 1.0e-9)
+		    {
+		      std::cout << GridLogMessage << "oops" << std::endl;
+		      abort ();
+		    }
+		  else
+		    std::cout << GridLogMessage << " 0 ";
+		}
+	      else
+		{
+		  std::cout << GridLogMessage << " " << in << " ";
+		}
+	    }
+	  std::cout << GridLogMessage << std::endl;
+	}
+    }
+
+    void step (std::vector < RealD > &lmd,
+	       std::vector < RealD > &lme,
+	       Field & last, Field & current, Field & next, uint64_t k)
+    {
+      if (lmd.size () <= k)
+	lmd.resize (k + Nm);
+      if (lme.size () <= k)
+	lme.resize (k + Nm);
+
+
+//      _poly(_Linop,current,next );   // 3. wk:=Avk−βkv_{k−1}
+      _Linop(current, next);	// 3. wk:=Avk−βkv_{k−1}
+      if (k > 0)
+	{
+	  next -= lme[k - 1] * last;
+	}
+//      std::cout<<GridLogMessage << "<last|next>" << innerProduct(last,next) <<std::endl;
+
+      ComplexD zalph = innerProduct (current, next);	// 4. αk:=(wk,vk)
+      RealD alph = real (zalph);
+
+      next = next - alph * current;	// 5. wk:=wk−αkvk
+//      std::cout<<GridLogMessage << "<current|next>" << innerProduct(current,next) <<std::endl;
+
+      RealD beta = normalise (next);	// 6. βk+1 := ∥wk∥2. If βk+1 = 0 then Stop
+      // 7. vk+1 := wk/βk+1
+//       norm=beta;
+
+      int interval = Nm / 100 + 1;
+      if ((k % interval) == 0)
+	std::
+	  cout << GridLogMessage << k << " : alpha = " << zalph << " beta " <<
+	  beta << std::endl;
+      const RealD tiny = 1.0e-20;
+      if (beta < tiny)
+	{
+	  std::cout << GridLogMessage << " beta is tiny " << beta << std::
+	    endl;
+	}
+      lmd[k] = alph;
+      lme[k] = beta;
+
+    }
+
+    void qr_decomp (std::vector < RealD > &lmd,
+		    std::vector  < RealD > &lme,
+		    int Nk,
+		    int Nm,
+		    std::vector  < RealD > &Qt, RealD Dsh, int kmin, int kmax)
+    {
+      int k = kmin - 1;
+      RealD x;
+
+      RealD Fden = 1.0 / hypot (lmd[k] - Dsh, lme[k]);
+      RealD c = (lmd[k] - Dsh) * Fden;
+      RealD s = -lme[k] * Fden;
+
+      RealD tmpa1 = lmd[k];
+      RealD tmpa2 = lmd[k + 1];
+      RealD tmpb = lme[k];
+
+      lmd[k] = c * c * tmpa1 + s * s * tmpa2 - 2.0 * c * s * tmpb;
+      lmd[k + 1] = s * s * tmpa1 + c * c * tmpa2 + 2.0 * c * s * tmpb;
+      lme[k] = c * s * (tmpa1 - tmpa2) + (c * c - s * s) * tmpb;
+      x = -s * lme[k + 1];
+      lme[k + 1] = c * lme[k + 1];
+
+      for (int i = 0; i < Nk; ++i)
+	{
+	  RealD Qtmp1 = Qt[i + Nm * k];
+	  RealD Qtmp2 = Qt[i + Nm * (k + 1)];
+	  Qt[i + Nm * k] = c * Qtmp1 - s * Qtmp2;
+	  Qt[i + Nm * (k + 1)] = s * Qtmp1 + c * Qtmp2;
+	}
+
+      // Givens transformations
+      for (int k = kmin; k < kmax - 1; ++k)
+	{
+
+	  RealD Fden = 1.0 / hypot (x, lme[k - 1]);
+	  RealD c = lme[k - 1] * Fden;
+	  RealD s = -x * Fden;
+
+	  RealD tmpa1 = lmd[k];
+	  RealD tmpa2 = lmd[k + 1];
+	  RealD tmpb = lme[k];
+
+	  lmd[k] = c * c * tmpa1 + s * s * tmpa2 - 2.0 * c * s * tmpb;
+	  lmd[k + 1] = s * s * tmpa1 + c * c * tmpa2 + 2.0 * c * s * tmpb;
+	  lme[k] = c * s * (tmpa1 - tmpa2) + (c * c - s * s) * tmpb;
+	  lme[k - 1] = c * lme[k - 1] - s * x;
+
+	  if (k != kmax - 2)
+	    {
+	      x = -s * lme[k + 1];
+	      lme[k + 1] = c * lme[k + 1];
+	    }
+
+	  for (int i = 0; i < Nk; ++i)
+	    {
+	      RealD Qtmp1 = Qt[i + Nm * k];
+	      RealD Qtmp2 = Qt[i + Nm * (k + 1)];
+	      Qt[i + Nm * k] = c * Qtmp1 - s * Qtmp2;
+	      Qt[i + Nm * (k + 1)] = s * Qtmp1 + c * Qtmp2;
+	    }
+	}
+    }
+
+#if 0
+#ifdef USE_LAPACK
+#ifdef USE_MKL
+#define LAPACK_INT MKL_INT
+#else
+#define LAPACK_INT long long
+#endif
+    void diagonalize_lapack (std::vector  < RealD > &lmd, std::vector  < RealD > &lme, int N1,	// all
+			     int N2,	// get
+			     GridBase * grid)
+    {
+      const int size = Nm;
+      LAPACK_INT NN = N1;
+      double evals_tmp[NN];
+      double DD[NN];
+      double EE[NN];
+      for (int i = 0; i < NN; i++)
+	for (int j = i - 1; j <= i + 1; j++)
+	  if (j < NN && j >= 0)
+	    {
+	      if (i == j)
+		DD[i] = lmd[i];
+	      if (i == j)
+		evals_tmp[i] = lmd[i];
+	      if (j == (i - 1))
+		EE[j] = lme[j];
+	    }
+      LAPACK_INT evals_found;
+      LAPACK_INT lwork =
+	((18 * NN) >
+	 (1 + 4 * NN + NN * NN) ? (18 * NN) : (1 + 4 * NN + NN * NN));
+      LAPACK_INT liwork = 3 + NN * 10;
+      LAPACK_INT iwork[liwork];
+      double work[lwork];
+      LAPACK_INT isuppz[2 * NN];
+      char jobz = 'N';		// calculate evals only
+      char range = 'I';		// calculate il-th to iu-th evals
+      //    char range = 'A'; // calculate all evals
+      char uplo = 'U';		// refer to upper half of original matrix
+      char compz = 'I';		// Compute eigenvectors of tridiagonal matrix
+      int ifail[NN];
+      LAPACK_INT info;
+//  int total = QMP_get_number_of_nodes();
+//  int node = QMP_get_node_number();
+//  GridBase *grid = evec[0]._grid;
+      int total = grid->_Nprocessors;
+      int node = grid->_processor;
+      int interval = (NN / total) + 1;
+      double vl = 0.0, vu = 0.0;
+      LAPACK_INT il = interval * node + 1, iu = interval * (node + 1);
+      if (iu > NN)
+	iu = NN;
+      double tol = 0.0;
+      if (1)
+	{
+	  memset (evals_tmp, 0, sizeof (double) * NN);
+	  if (il <= NN)
+	    {
+	      printf ("total=%d node=%d il=%d iu=%d\n", total, node, il, iu);
+#ifdef USE_MKL
+	      dstegr (&jobz, &range, &NN,
+#else
+	      LAPACK_dstegr (&jobz, &range, &NN,
+#endif
+			     (double *) DD, (double *) EE, &vl, &vu, &il, &iu,	// these four are ignored if second parameteris 'A'
+			     &tol,	// tolerance
+			     &evals_found, evals_tmp, (double *) NULL, &NN,
+			     isuppz, work, &lwork, iwork, &liwork, &info);
+	      for (int i = iu - 1; i >= il - 1; i--)
+		{
+		  printf ("node=%d evals_found=%d evals_tmp[%d] = %g\n", node,
+			  evals_found, i - (il - 1), evals_tmp[i - (il - 1)]);
+		  evals_tmp[i] = evals_tmp[i - (il - 1)];
+		  if (il > 1)
+		    evals_tmp[i - (il - 1)] = 0.;
+		}
+	    }
+	  {
+	    grid->GlobalSumVector (evals_tmp, NN);
+	  }
+	}
+// cheating a bit. It is better to sort instead of just reversing it, but the document of the routine says evals are sorted in increasing order. qr gives evals in decreasing order.
+    }
+#undef LAPACK_INT
+#endif
+
+
+    void diagonalize (std::vector  < RealD > &lmd,
+		      std::vector  < RealD > &lme,
+		      int N2, int N1, GridBase * grid)
+    {
+
+#ifdef USE_LAPACK
+      const int check_lapack = 0;	// just use lapack if 0, check against lapack if 1
+
+      if (!check_lapack)
+	return diagonalize_lapack (lmd, lme, N2, N1, grid);
+
+//      diagonalize_lapack(lmd2,lme2,Nm2,Nm,Qt,grid);
+#endif
+    }
+#endif
+
+    static RealD normalise (Field & v)
+    {
+      RealD nn = norm2 (v);
+      nn = sqrt (nn);
+      v = v * (1.0 / nn);
+      return nn;
+    }
+
+    void orthogonalize (Field & w, std::vector < Field > &evec, int k)
+    {
+      double t0 = -usecond () / 1e6;
+      typedef typename Field::scalar_type MyComplex;
+      MyComplex ip;
+
+      if (0)
+	{
+	  for (int j = 0; j < k; ++j)
+	    {
+	      normalise (evec[j]);
+	      for (int i = 0; i < j; i++)
+		{
+		  ip = innerProduct (evec[i], evec[j]);	// are the evecs normalised? ; this assumes so.
+		  evec[j] = evec[j] - ip * evec[i];
+		}
+	    }
+	}
+
+      for (int j = 0; j < k; ++j)
+	{
+	  ip = innerProduct (evec[j], w);	// are the evecs normalised? ; this assumes so.
+	  w = w - ip * evec[j];
+	}
+      normalise (w);
+      t0 += usecond () / 1e6;
+      OrthoTime += t0;
+    }
+
+    void setUnit_Qt (int Nm, std::vector < RealD > &Qt)
+    {
+      for (int i = 0; i < Qt.size (); ++i)
+	Qt[i] = 0.0;
+      for (int k = 0; k < Nm; ++k)
+	Qt[k + k * Nm] = 1.0;
+    }
+
+
+    void calc (std::vector < RealD > &eval, const Field & src, int &Nconv)
+    {
+
+      GridBase *grid = src.Grid();
+//      assert(grid == src._grid);
+
+      std::
+	cout << GridLogMessage << " -- Nk = " << Nk << " Np = " << Np << std::
+	endl;
+      std::cout << GridLogMessage << " -- Nm = " << Nm << std::endl;
+      std::cout << GridLogMessage << " -- size of eval   = " << eval.
+	size () << std::endl;
+
+//      assert(c.size() && Nm == eval.size());
+
+      std::vector < RealD > lme (Nm);
+      std::vector < RealD > lmd (Nm);
+
+
+      Field current (grid);
+      Field last (grid);
+      Field next (grid);
+
+      Nconv = 0;
+
+      RealD beta_k;
+
+      // Set initial vector
+      // (uniform vector) Why not src??
+      //      evec[0] = 1.0;
+      current = src;
+      std::cout << GridLogMessage << "norm2(src)= " << norm2 (src) << std::
+	endl;
+      normalise (current);
+      std::
+	cout << GridLogMessage << "norm2(evec[0])= " << norm2 (current) <<
+	std::endl;
+
+      // Initial Nk steps
+      OrthoTime = 0.;
+      double t0 = usecond () / 1e6;
+      RealD norm;		// sqrt norm of last vector
+
+      uint64_t iter = 0;
+
+      bool initted = false;
+      std::vector < RealD > low (Nstop * 10);
+      std::vector < RealD > high (Nstop * 10);
+      RealD cont = 0.;
+      while (1) {
+	  cont = 0.;
+	  std::vector < RealD > lme2 (Nm);
+	  std::vector < RealD > lmd2 (Nm);
+	  for (uint64_t k = 0; k < Nm; ++k, iter++) {
+	      step (lmd, lme, last, current, next, iter);
+	      last = current;
+	      current = next;
+	    }
+	  double t1 = usecond () / 1e6;
+	  std::cout << GridLogMessage << "IRL::Initial steps: " << t1 -
+	    t0 << "seconds" << std::endl;
+	  t0 = t1;
+	  std::
+	    cout << GridLogMessage << "IRL::Initial steps:OrthoTime " <<
+	    OrthoTime << "seconds" << std::endl;
+
+	  // getting eigenvalues
+	  lmd2.resize (iter + 2);
+	  lme2.resize (iter + 2);
+	  for (uint64_t k = 0; k < iter; ++k) {
+	      lmd2[k + 1] = lmd[k];
+	      lme2[k + 2] = lme[k];
+	    }
+	  t1 = usecond () / 1e6;
+	  std::cout << GridLogMessage << "IRL:: copy: " << t1 -
+	    t0 << "seconds" << std::endl;
+	  t0 = t1;
+	  {
+	    int total = grid->_Nprocessors;
+	    int node = grid->_processor;
+	    int interval = (Nstop / total) + 1;
+	    int iu = (iter + 1) - (interval * node + 1);
+	    int il = (iter + 1) - (interval * (node + 1));
+	    std::vector < RealD > eval2 (iter + 3);
+	    RealD eps2;
+	    Bisection::bisec (lmd2, lme2, iter, il, iu, 1e-16, 1e-10, eval2,
+			      eps2);
+//        diagonalize(eval2,lme2,iter,Nk,grid);
+	    RealD diff = 0.;
+	    for (int i = il; i <= iu; i++) {
+		if (initted)
+		  diff =
+		    fabs (eval2[i] - high[iu-i]) / (fabs (eval2[i]) +
+						      fabs (high[iu-i]));
+		if (initted && (diff > eresid))
+		  cont = 1.;
+		if (initted)
+		  printf ("eval[%d]=%0.14e %0.14e, %0.14e\n", i, eval2[i],
+			  high[iu-i], diff);
+		high[iu-i] = eval2[i];
+	      }
+	    il = (interval * node + 1);
+	    iu = (interval * (node + 1));
+	    Bisection::bisec (lmd2, lme2, iter, il, iu, 1e-16, 1e-10, eval2,
+			      eps2);
+	    for (int i = il; i <= iu; i++) {
+		if (initted)
+		  diff =
+		    fabs (eval2[i] - low[i]) / (fabs (eval2[i]) +
+						fabs (low[i]));
+		if (initted && (diff > eresid))
+		  cont = 1.;
+		if (initted)
+		  printf ("eval[%d]=%0.14e %0.14e, %0.14e\n", i, eval2[i],
+			  low[i], diff);
+		low[i] = eval2[i];
+	      }
+	    t1 = usecond () / 1e6;
+	    std::cout << GridLogMessage << "IRL:: diagonalize: " << t1 -
+	      t0 << "seconds" << std::endl;
+	    t0 = t1;
+	  }
+
+	  for (uint64_t k = 0; k < Nk; ++k) {
+//          eval[k] = eval2[k];
+	    }
+	  if (initted)
+	    {
+	      grid->GlobalSumVector (&cont, 1);
+	      if (cont < 1.) return;
+	    }
+	  initted = true;
+	}
+
+    }
+
+
+
+
+
+#if 0
+
+/**
+   There is some matrix Q such that for any vector y
+   Q.e_1 = y and Q is unitary.
+**/
+    template < class T >
+      static T orthQ (DenseMatrix < T > &Q, std::vector < T > y)
+    {
+      int N = y.size ();	//Matrix Size
+      Fill (Q, 0.0);
+      T tau;
+      for (int i = 0; i < N; i++)
+	{
+	  Q[i][0] = y[i];
+	}
+      T sig = conj (y[0]) * y[0];
+      T tau0 = fabs (sqrt (sig));
+
+      for (int j = 1; j < N; j++)
+	{
+	  sig += conj (y[j]) * y[j];
+	  tau = abs (sqrt (sig));
+
+	  if (abs (tau0) > 0.0)
+	    {
+
+	      T gam = conj ((y[j] / tau) / tau0);
+	      for (int k = 0; k <= j - 1; k++)
+		{
+		  Q[k][j] = -gam * y[k];
+		}
+	      Q[j][j] = tau0 / tau;
+	    }
+	  else
+	    {
+	      Q[j - 1][j] = 1.0;
+	    }
+	  tau0 = tau;
+	}
+      return tau;
+    }
+
+/**
+	There is some matrix Q such that for any vector y
+	Q.e_k = y and Q is unitary.
+**/
+    template < class T >
+      static T orthU (DenseMatrix < T > &Q, std::vector < T > y)
+    {
+      T tau = orthQ (Q, y);
+      SL (Q);
+      return tau;
+    }
+
+
+/**
+	Wind up with a matrix with the first con rows untouched
+
+say con = 2
+	Q is such that Qdag H Q has {x, x, val, 0, 0, 0, 0, ...} as 1st colum
+	and the matrix is upper hessenberg
+	and with f and Q appropriately modidied with Q is the arnoldi factorization
+
+**/
+
+    template < class T > static void Lock (DenseMatrix < T > &H,	///Hess mtx     
+					   DenseMatrix < T > &Q,	///Lock Transform
+					   T val,	///value to be locked
+					   int con,	///number already locked
+					   RealD small, int dfg, bool herm)
+    {
+      //ForceTridiagonal(H);
+
+      int M = H.dim;
+      DenseVector < T > vec;
+      Resize (vec, M - con);
+
+      DenseMatrix < T > AH;
+      Resize (AH, M - con, M - con);
+      AH = GetSubMtx (H, con, M, con, M);
+
+      DenseMatrix < T > QQ;
+      Resize (QQ, M - con, M - con);
+
+      Unity (Q);
+      Unity (QQ);
+
+      DenseVector < T > evals;
+      Resize (evals, M - con);
+      DenseMatrix < T > evecs;
+      Resize (evecs, M - con, M - con);
+
+      Wilkinson < T > (AH, evals, evecs, small);
+
+      int k = 0;
+      RealD cold = abs (val - evals[k]);
+      for (int i = 1; i < M - con; i++)
+	{
+	  RealD cnew = abs (val - evals[i]);
+	  if (cnew < cold)
+	    {
+	      k = i;
+	      cold = cnew;
+	    }
+	}
+      vec = evecs[k];
+
+      ComplexD tau;
+      orthQ (QQ, vec);
+      //orthQM(QQ,AH,vec);
+
+      AH = Hermitian (QQ) * AH;
+      AH = AH * QQ;
+
+      for (int i = con; i < M; i++)
+	{
+	  for (int j = con; j < M; j++)
+	    {
+	      Q[i][j] = QQ[i - con][j - con];
+	      H[i][j] = AH[i - con][j - con];
+	    }
+	}
+
+      for (int j = M - 1; j > con + 2; j--)
+	{
+
+	  DenseMatrix < T > U;
+	  Resize (U, j - 1 - con, j - 1 - con);
+	  DenseVector < T > z;
+	  Resize (z, j - 1 - con);
+	  T nm = norm (z);
+	  for (int k = con + 0; k < j - 1; k++)
+	    {
+	      z[k - con] = conj (H (j, k + 1));
+	    }
+	  normalise (z);
+
+	  RealD tmp = 0;
+	  for (int i = 0; i < z.size () - 1; i++)
+	    {
+	      tmp = tmp + abs (z[i]);
+	    }
+
+	  if (tmp < small / ((RealD) z.size () - 1.0))
+	    {
+	      continue;
+	    }
+
+	  tau = orthU (U, z);
+
+	  DenseMatrix < T > Hb;
+	  Resize (Hb, j - 1 - con, M);
+
+	  for (int a = 0; a < M; a++)
+	    {
+	      for (int b = 0; b < j - 1 - con; b++)
+		{
+		  T sum = 0;
+		  for (int c = 0; c < j - 1 - con; c++)
+		    {
+		      sum += H[a][con + 1 + c] * U[c][b];
+		    }		//sum += H(a,con+1+c)*U(c,b);}
+		  Hb[b][a] = sum;
+		}
+	    }
+
+	  for (int k = con + 1; k < j; k++)
+	    {
+	      for (int l = 0; l < M; l++)
+		{
+		  H[l][k] = Hb[k - 1 - con][l];
+		}
+	    }			//H(Hb[k-1-con][l] , l,k);}}
+
+	  DenseMatrix < T > Qb;
+	  Resize (Qb, M, M);
+
+	  for (int a = 0; a < M; a++)
+	    {
+	      for (int b = 0; b < j - 1 - con; b++)
+		{
+		  T sum = 0;
+		  for (int c = 0; c < j - 1 - con; c++)
+		    {
+		      sum += Q[a][con + 1 + c] * U[c][b];
+		    }		//sum += Q(a,con+1+c)*U(c,b);}
+		  Qb[b][a] = sum;
+		}
+	    }
+
+	  for (int k = con + 1; k < j; k++)
+	    {
+	      for (int l = 0; l < M; l++)
+		{
+		  Q[l][k] = Qb[k - 1 - con][l];
+		}
+	    }			//Q(Qb[k-1-con][l] , l,k);}}
+
+	  DenseMatrix < T > Hc;
+	  Resize (Hc, M, M);
+
+	  for (int a = 0; a < j - 1 - con; a++)
+	    {
+	      for (int b = 0; b < M; b++)
+		{
+		  T sum = 0;
+		  for (int c = 0; c < j - 1 - con; c++)
+		    {
+		      sum += conj (U[c][a]) * H[con + 1 + c][b];
+		    }		//sum += conj( U(c,a) )*H(con+1+c,b);}
+		  Hc[b][a] = sum;
+		}
+	    }
+
+	  for (int k = 0; k < M; k++)
+	    {
+	      for (int l = con + 1; l < j; l++)
+		{
+		  H[l][k] = Hc[k][l - 1 - con];
+		}
+	    }			//H(Hc[k][l-1-con] , l,k);}}
+
+	}
+    }
+#endif
+
+
+  };
+
+}
+#endif

Grid/algorithms/multigrid/Aggregates.h

@@ -30,6 +30,8 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 /*  END LEGAL */
 #pragma once
 
+#include <Grid/algorithms/iterative/PrecGeneralisedConjugateResidualNonHermitian.h>
+
 NAMESPACE_BEGIN(Grid);
 
 inline RealD AggregatePowerLaw(RealD x)
@@ -95,7 +97,7 @@ public:
 
     RealD scale;
 
-    ConjugateGradient<FineField> CG(1.0e-2,100,false);
+    ConjugateGradient<FineField> CG(1.0e-4,2000,false);
     FineField noise(FineGrid);
     FineField Mn(FineGrid);
 
@@ -108,7 +110,7 @@ public:
       
       hermop.Op(noise,Mn); std::cout<<GridLogMessage << "noise   ["<<b<<"] <n|MdagM|n> "<<norm2(Mn)<<std::endl;
 
-      for(int i=0;i<1;i++){
+      for(int i=0;i<4;i++){
 
 	CG(hermop,noise,subspace[b]);
 
@@ -124,6 +126,56 @@ public:
     }
   }
 
+  virtual void CreateSubspaceGCR(GridParallelRNG  &RNG,LinearOperatorBase<FineField> &DiracOp,int nn=nbasis)
+  {
+    RealD scale;
+
+    TrivialPrecon<FineField> simple_fine;
+    //    PrecGeneralisedConjugateResidualNonHermitian<FineField> GCR(0.001,10,DiracOp,simple_fine,30,30);
+    //    PrecGeneralisedConjugateResidualNonHermitian<FineField> GCR(0.001,10,DiracOp,simple_fine,12,12);
+    //    PrecGeneralisedConjugateResidualNonHermitian<FineField> GCR(0.001,30,DiracOp,simple_fine,12,12);
+    PrecGeneralisedConjugateResidualNonHermitian<FineField> GCR(0.001,30,DiracOp,simple_fine,10,10);
+    FineField noise(FineGrid);
+    FineField src(FineGrid);
+    FineField guess(FineGrid);
+    FineField Mn(FineGrid);
+
+    for(int b=0;b<nn;b++){
+      
+      subspace[b] = Zero();
+      gaussian(RNG,noise);
+      scale = std::pow(norm2(noise),-0.5); 
+      noise=noise*scale;
+      
+      DiracOp.Op(noise,Mn); std::cout<<GridLogMessage << "noise   ["<<b<<"] <n|Op|n> "<<innerProduct(noise,Mn)<<std::endl;
+
+      for(int i=0;i<3;i++){
+	//  void operator() (const Field &src, Field &psi){
+#if 1
+	if (i==0)std::cout << GridLogMessage << " inverting on noise "<<std::endl;
+	src = noise;
+	guess=Zero();
+	GCR(src,guess);
+	subspace[b] = guess;
+#else
+	if (i==0)std::cout << GridLogMessage << " inverting on zero "<<std::endl;
+	src=Zero();
+	guess = noise;
+	GCR(src,guess);
+	subspace[b] = guess;
+#endif
+	noise = subspace[b];
+	scale = std::pow(norm2(noise),-0.5); 
+	noise=noise*scale;
+
+      }
+
+      DiracOp.Op(noise,Mn); std::cout<<GridLogMessage << "filtered["<<b<<"] <f|Op|f> "<<innerProduct(noise,Mn)<<" <f|OpDagOp|f>"<<norm2(Mn)<<std::endl;
+      subspace[b]   = noise;
+
+    }
+  }
+
   ////////////////////////////////////////////////////////////////////////////////////////////////
   // World of possibilities here. But have tried quite a lot of experiments (250+ jobs run on Summit)
   // and this is the best I found
@@ -160,14 +212,21 @@ public:
 
     int b =0;
     {
+      ComplexD ip;
       // Filter
       Chebyshev<FineField> Cheb(lo,hi,orderfilter);
       Cheb(hermop,noise,Mn);
       // normalise
       scale = std::pow(norm2(Mn),-0.5); 	Mn=Mn*scale;
       subspace[b]   = Mn;
+
       hermop.Op(Mn,tmp);
-      std::cout<<GridLogMessage << "filt ["<<b<<"] <n|MdagM|n> "<<norm2(tmp)<<std::endl;
+      ip= innerProduct(Mn,tmp); 
+      std::cout<<GridLogMessage << "filt ["<<b<<"] <n|Op|n> "<<norm2(tmp)<<" "<<ip<<std::endl;
+
+      hermop.AdjOp(Mn,tmp); 
+      ip = innerProduct(Mn,tmp); 
+      std::cout<<GridLogMessage << "filt ["<<b<<"] <n|AdjOp|n> "<<norm2(tmp)<<" "<<ip<<std::endl;
       b++;
     }
 
@@ -213,8 +272,18 @@ public:
 	  Mn=*Tnp;
 	  scale = std::pow(norm2(Mn),-0.5);         Mn=Mn*scale;
 	  subspace[b] = Mn;
+
+
+	  ComplexD ip;
+
 	  hermop.Op(Mn,tmp);
-	  std::cout<<GridLogMessage << n<<" filt ["<<b<<"] <n|MdagM|n> "<<norm2(tmp)<<std::endl;
+	  ip= innerProduct(Mn,tmp); 
+	  std::cout<<GridLogMessage << "filt ["<<b<<"] <n|Op|n> "<<norm2(tmp)<<" "<<ip<<std::endl;
+
+	  hermop.AdjOp(Mn,tmp); 
+	  ip = innerProduct(Mn,tmp); 
+	  std::cout<<GridLogMessage << "filt ["<<b<<"] <n|AdjOp|n> "<<norm2(tmp)<<" "<<ip<<std::endl;
+	  
 	  b++;
 	}
 
@@ -226,8 +295,72 @@ public:
 	  
       }
     }
-    assert(b==nn);
+    GRID_ASSERT(b==nn);
   }
+
+
+  virtual void CreateSubspacePolyCheby(GridParallelRNG  &RNG,LinearOperatorBase<FineField> &hermop,
+				       int nn,
+				       double hi,
+				       double lo1,
+				       int orderfilter,
+				       double lo2,
+				       int orderstep)
+  {
+    RealD scale;
+
+    FineField noise(FineGrid);
+    FineField Mn(FineGrid);
+    FineField tmp(FineGrid);
+
+    // New normalised noise
+    gaussian(RNG,noise);
+    scale = std::pow(norm2(noise),-0.5); 
+    noise=noise*scale;
+
+    std::cout << GridLogMessage<<" CreateSubspacePolyCheby "<<std::endl;
+    // Initial matrix element
+    hermop.Op(noise,Mn);
+    std::cout<<GridLogMessage << "noise <n|MdagM|n> "<<norm2(Mn)<<std::endl;
+
+    int b =0;
+    {
+      // Filter
+      std::cout << GridLogMessage << "Cheby "<<lo1<<","<<hi<<" "<<orderstep<<std::endl;
+      Chebyshev<FineField> Cheb(lo1,hi,orderfilter);
+      Cheb(hermop,noise,Mn);
+      // normalise
+      scale = std::pow(norm2(Mn),-0.5); 	Mn=Mn*scale;
+      subspace[b]   = Mn;
+      hermop.Op(Mn,tmp); 
+      std::cout<<GridLogMessage << "filt ["<<b<<"] <n|MdagM|n> "<<norm2(tmp)<<std::endl;
+      std::cout<<GridLogMessage << "filt ["<<b<<"] <n|n> "<<norm2(Mn)<<std::endl;
+    }
+
+    // Generate a full sequence of Chebyshevs
+    for(int n=1;n<nn;n++){
+      std::cout << GridLogMessage << "Cheby "<<lo2<<","<<hi<<" "<<orderstep<<std::endl;
+      Chebyshev<FineField> Cheb(lo2,hi,orderstep);
+      Cheb(hermop,subspace[n-1],Mn);
+
+      for(int m=0;m<n;m++){
+	ComplexD c = innerProduct(subspace[m],Mn);
+	Mn = Mn - c*subspace[m];
+      }
+      
+      // normalise
+      scale = std::pow(norm2(Mn),-0.5);
+      Mn=Mn*scale;
+      
+      subspace[n]=Mn;
+      
+      hermop.Op(Mn,tmp); 
+      std::cout<<GridLogMessage << "filt ["<<n<<"] <n|MdagM|n> "<<norm2(tmp)<<std::endl;
+      std::cout<<GridLogMessage << "filt ["<<n<<"] <n|n> "<<norm2(Mn)<<std::endl;
+
+    }
+  }
+
   virtual void CreateSubspaceChebyshev(GridParallelRNG  &RNG,LinearOperatorBase<FineField> &hermop,
 				       int nn,
 				       double hi,

Grid/algorithms/multigrid/CoarsenedMatrix.h

@@ -99,7 +99,7 @@ public:
   CoarseMatrix AselfInvEven;
   CoarseMatrix AselfInvOdd;
 
-  Vector<RealD> dag_factor;
+  deviceVector<RealD> dag_factor;
 
   ///////////////////////
   // Interface
@@ -124,9 +124,13 @@ public:
     int npoint = geom.npoint;
     typedef LatticeView<Cobj> Aview;
       
-    Vector<Aview> AcceleratorViewContainer;
+    deviceVector<Aview> AcceleratorViewContainer(geom.npoint);
+    hostVector<Aview>   hAcceleratorViewContainer(geom.npoint);
   
-    for(int p=0;p<geom.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();
@@ -161,7 +165,7 @@ public:
       coalescedWrite(out_v[ss](b),res);
       });
 
-    for(int p=0;p<geom.npoint;p++) AcceleratorViewContainer[p].ViewClose();
+    for(int p=0;p<geom.npoint;p++) hAcceleratorViewContainer[p].ViewClose();
   };
 
   void Mdag (const CoarseVector &in, CoarseVector &out)
@@ -190,9 +194,14 @@ public:
     int npoint = geom.npoint;
     typedef LatticeView<Cobj> Aview;
 
-    Vector<Aview> AcceleratorViewContainer;
 
-    for(int p=0;p<geom.npoint;p++) AcceleratorViewContainer.push_back(A[p].View(AcceleratorRead));
+    deviceVector<Aview> AcceleratorViewContainer(geom.npoint);
+    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();
@@ -201,10 +210,10 @@ public:
 
     int osites=Grid()->oSites();
 
-    Vector<int> points(geom.npoint, 0);
+    deviceVector<int> points(geom.npoint);
-    for(int p=0; p<geom.npoint; p++)
+    for(int p=0; p<geom.npoint; p++) { 
-      points[p] = geom.points_dagger[p];
+      acceleratorPut(points[p],geom.points_dagger[p]);
-
+    }
     auto points_p = &points[0];
 
     RealD* dag_factor_p = &dag_factor[0];
@@ -236,7 +245,7 @@ public:
       coalescedWrite(out_v[ss](b),res);
       });
 
-    for(int p=0;p<geom.npoint;p++) AcceleratorViewContainer[p].ViewClose();
+    for(int p=0;p<geom.npoint;p++) hAcceleratorViewContainer[p].ViewClose();
   }
 
   void MdirComms(const CoarseVector &in)
@@ -251,8 +260,14 @@ public:
     out.Checkerboard() = in.Checkerboard();
 
     typedef LatticeView<Cobj> Aview;
-    Vector<Aview> AcceleratorViewContainer;
+
-    for(int p=0;p<geom.npoint;p++) AcceleratorViewContainer.push_back(A[p].View(AcceleratorRead));
+    deviceVector<Aview> AcceleratorViewContainer(geom.npoint);
+    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);
@@ -285,7 +300,7 @@ public:
       }
       coalescedWrite(out_v[ss](b),res);
     });
-    for(int p=0;p<geom.npoint;p++) AcceleratorViewContainer[p].ViewClose();
+    for(int p=0;p<geom.npoint;p++) hAcceleratorViewContainer[p].ViewClose();
   }
   void MdirAll(const CoarseVector &in,std::vector<CoarseVector> &out)
   {
@@ -294,7 +309,7 @@ public:
     if ((out.size()!=ndir)&&(out.size()!=ndir+1)) { 
       std::cout <<"MdirAll out size "<< out.size()<<std::endl;
       std::cout <<"MdirAll ndir "<< ndir<<std::endl;
-      assert(0);
+      GRID_ASSERT(0);
     }
     for(int p=0;p<ndir;p++){
       MdirCalc(in,out[p],p);
@@ -358,7 +373,7 @@ public:
     conformable(in.Grid(), _cbgrid);    // verifies half grid
     conformable(in.Grid(), out.Grid()); // drops the cb check
 
-    assert(in.Checkerboard() == Even);
+    GRID_ASSERT(in.Checkerboard() == Even);
     out.Checkerboard() = Odd;
 
     DhopInternal(StencilEven, Aodd, in, out, dag);
@@ -368,7 +383,7 @@ public:
     conformable(in.Grid(), _cbgrid);    // verifies half grid
     conformable(in.Grid(), out.Grid()); // drops the cb check
 
-    assert(in.Checkerboard() == Odd);
+    GRID_ASSERT(in.Checkerboard() == Odd);
     out.Checkerboard() = Even;
 
     DhopInternal(StencilOdd, Aeven, in, out, dag);
@@ -376,7 +391,7 @@ public:
 
   void MooeeInternal(const CoarseVector &in, CoarseVector &out, int dag, int inv) {
     out.Checkerboard() = in.Checkerboard();
-    assert(in.Checkerboard() == Odd || in.Checkerboard() == Even);
+    GRID_ASSERT(in.Checkerboard() == Odd || in.Checkerboard() == Even);
 
     CoarseMatrix *Aself = nullptr;
     if(in.Grid()->_isCheckerBoarded) {
@@ -391,7 +406,7 @@ public:
       Aself = (inv) ? &AselfInv : &A[geom.npoint-1];
       DselfInternal(Stencil, *Aself, in, out, dag);
     }
-    assert(Aself != nullptr);
+    GRID_ASSERT(Aself != nullptr);
   }
 
   void DselfInternal(CartesianStencil<siteVector,siteVector,DefaultImplParams> &st, CoarseMatrix &a,
@@ -469,14 +484,20 @@ public:
 
     // determine in what order we need the points
     int npoint = geom.npoint-1;
-    Vector<int> points(npoint, 0);
+    deviceVector<int> points(npoint);
-    for(int p=0; p<npoint; p++)
+    for(int p=0; p<npoint; p++) {
-      points[p] = (dag && !hermitian) ? geom.points_dagger[p] : p;
+      int val = (dag && !hermitian) ? geom.points_dagger[p] : p;
-
+      acceleratorPut(points[p], val);
+    }
     auto points_p = &points[0];
 
-    Vector<Aview> AcceleratorViewContainer;
+    deviceVector<Aview> AcceleratorViewContainer(geom.npoint);
-    for(int p=0;p<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();
@@ -539,7 +560,7 @@ public:
       });
     }
 
-    for(int p=0;p<npoint;p++) AcceleratorViewContainer[p].ViewClose();
+    for(int p=0;p<npoint;p++) hAcceleratorViewContainer[p].ViewClose();
   }
   
   CoarsenedMatrix(GridCartesian &CoarseGrid, int hermitian_=0) 	:
@@ -590,11 +611,13 @@ 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;
-      dag_factor[i] = dag_factor_eigen(j, k);
+      h_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,
@@ -674,7 +697,7 @@ public:
     evenmask = where(mod(bcb,2)==(Integer)0,one,zero);
     oddmask  = one-evenmask;
 
-    assert(self_stencil!=-1);
+    GRID_ASSERT(self_stencil!=-1);
 
     for(int i=0;i<nbasis;i++){
 

Grid/algorithms/multigrid/GeneralCoarsenedMatrix.h

@@ -99,7 +99,7 @@ public:
 	}
       }
     }
-    assert(nfound==geom.npoint);
+    GRID_ASSERT(nfound==geom.npoint);
     ExchangeCoarseLinks();
   }
   */
@@ -124,7 +124,7 @@ public:
   }
   void Mdag (const CoarseVector &in, CoarseVector &out)
   {
-    assert(hermitian);
+    GRID_ASSERT(hermitian);
     Mult(_A,in,out);
     //    if ( hermitian ) M(in,out);
     //    else Mult(_Adag,in,out);
@@ -441,8 +441,20 @@ public:
     std::cout << GridLogMessage<<"CoarsenOperator inv    "<<tinv<<" us"<<std::endl;
   }
 #else
+  //////////////////////////////////////////////////////////////////////
+  // Galerkin projection of matrix
+  //////////////////////////////////////////////////////////////////////
   void CoarsenOperator(LinearOperatorBase<Lattice<Fobj> > &linop,
 		       Aggregation<Fobj,CComplex,nbasis> & Subspace)
+  {
+    CoarsenOperator(linop,Subspace,Subspace);
+  }
+  //////////////////////////////////////////////////////////////////////
+  // Petrov - Galerkin projection of matrix
+  //////////////////////////////////////////////////////////////////////
+  void CoarsenOperator(LinearOperatorBase<Lattice<Fobj> > &linop,
+		       Aggregation<Fobj,CComplex,nbasis> & U,
+		       Aggregation<Fobj,CComplex,nbasis> & V)
   {
     std::cout << GridLogMessage<< "GeneralCoarsenMatrix "<< std::endl;
     GridBase *grid = FineGrid();
@@ -458,11 +470,9 @@ public:
     // Orthogonalise the subblocks over the basis
     /////////////////////////////////////////////////////////////
     CoarseScalar InnerProd(CoarseGrid()); 
-    blockOrthogonalise(InnerProd,Subspace.subspace);
+    blockOrthogonalise(InnerProd,V.subspace);
+    blockOrthogonalise(InnerProd,U.subspace);
 
-    //    for(int s=0;s<Subspace.subspace.size();s++){
-      //      std::cout << " subspace norm "<<norm2(Subspace.subspace[s])<<std::endl;
-    //    }
     const int npoint = geom.npoint;
       
     Coordinate clatt = CoarseGrid()->GlobalDimensions();
@@ -542,7 +552,7 @@ public:
       std::cout << GridLogMessage<< "CoarsenMatrixColoured vec "<<i<<"/"<<nbasis<< std::endl;
       for(int p=0;p<npoint;p++){ // Loop over momenta in npoint
 	tphaseBZ-=usecond();
-	phaV = phaF[p]*Subspace.subspace[i];
+	phaV = phaF[p]*V.subspace[i];
 	tphaseBZ+=usecond();
 
 	/////////////////////////////////////////////////////////////////////
@@ -555,7 +565,7 @@ public:
 	//	std::cout << i << " " <<p << " MphaV "<<norm2(MphaV)<<" "<<norm2(phaV)<<std::endl;
 
 	tproj-=usecond();
-	blockProject(coarseInner,MphaV,Subspace.subspace);
+	blockProject(coarseInner,MphaV,U.subspace);
 	coarseInner = conjugate(pha[p]) * coarseInner;
 
 	ComputeProj[p] = coarseInner;
@@ -609,7 +619,7 @@ public:
       //      _Adag[p]= Cell.ExchangePeriodic(_Adag[p]);
     }
   }
-  virtual  void Mdiag    (const Field &in, Field &out){ assert(0);};
+  virtual  void Mdiag    (const Field &in, Field &out){ GRID_ASSERT(0);};
   virtual  void Mdir     (const Field &in, Field &out,int dir, int disp){assert(0);};
   virtual  void MdirAll  (const Field &in, std::vector<Field> &out){assert(0);};
 };

Grid/algorithms/multigrid/GeneralCoarsenedMatrixMultiRHS.h

@@ -80,12 +80,12 @@ public:
   // Can be used to do I/O on the operator matrices externally
   void SetMatrix (int p,CoarseMatrix & A)
   {
-    assert(A.size()==geom_srhs.npoint);
+    GRID_ASSERT(A.size()==geom_srhs.npoint);
     GridtoBLAS(A[p],BLAS_A[p]);
   }
   void GetMatrix (int p,CoarseMatrix & A)
   {
-    assert(A.size()==geom_srhs.npoint);
+    GRID_ASSERT(A.size()==geom_srhs.npoint);
     BLAStoGrid(A[p],BLAS_A[p]);
   }
   void CopyMatrix (GeneralCoarseOp &_Op)
@@ -178,14 +178,14 @@ public:
 	for(int32_t point = 0 ; point < geom.npoint; point++){
 	  int i=s*orhs*geom.npoint+point;
  	  int32_t nbr = Stencil._entries[i]._offset*CComplex::Nsimd(); // oSite -> lSite
-	  assert(nbr<BLAS_B.size());
+	  GRID_ASSERT(nbr<BLAS_B.size());
 	  ComplexD * ptr = (ComplexD *)&BLAS_B[nbr];
 	  acceleratorPut(BLAS_BP[point][j],ptr); // neighbour indexing in ghost zone volume
 	}
 	j++;
       }
     }
-    assert(j==unpadded_sites);
+    GRID_ASSERT(j==unpadded_sites);
   }
   template<class vobj> void GridtoBLAS(const Lattice<vobj> &from,deviceVector<typename vobj::scalar_object> &to)
   {
@@ -194,7 +194,7 @@ public:
   typedef typename vobj::vector_type vector_type;
 
   GridBase *Fg = from.Grid();
-  assert(!Fg->_isCheckerBoarded);
+  GRID_ASSERT(!Fg->_isCheckerBoarded);
   int nd = Fg->_ndimension;
 
   to.resize(Fg->lSites());
@@ -241,10 +241,10 @@ public:
   typedef typename vobj::vector_type vector_type;
 
   GridBase *Tg = grid.Grid();
-  assert(!Tg->_isCheckerBoarded);
+  GRID_ASSERT(!Tg->_isCheckerBoarded);
   int nd = Tg->_ndimension;
   
-  assert(in.size()==Tg->lSites());
+  GRID_ASSERT(in.size()==Tg->lSites());
 
   Coordinate LocalLatt = Tg->LocalDimensions();
   size_t nsite = 1;
@@ -669,7 +669,7 @@ Grid : Message : 328.193436 s : CoarsenOperator mat    122213270 us
     const int Nsimd = CComplex::Nsimd();
 
     int64_t nrhs  =pin.Grid()->GlobalDimensions()[0];
-    assert(nrhs>=1);
+    GRID_ASSERT(nrhs>=1);
 
     RealD flops,bytes;
     int64_t osites=in.Grid()->oSites(); // unpadded
@@ -721,7 +721,7 @@ Grid : Message : 328.193436 s : CoarsenOperator mat    122213270 us
     //    std::cout << GridLogMessage<<"Coarse overall flops/s "<< flops/t_tot<<" mflop/s"<<std::endl;
     //    std::cout << GridLogMessage<<"Coarse total bytes   "<< bytes/1e6<<" MB"<<std::endl;
   };
-  virtual  void Mdiag    (const Field &in, Field &out){ assert(0);};
+  virtual  void Mdiag    (const Field &in, Field &out){ GRID_ASSERT(0);};
   virtual  void Mdir     (const Field &in, Field &out,int dir, int disp){assert(0);};
   virtual  void MdirAll  (const Field &in, std::vector<Field> &out){assert(0);};
 };

Grid/algorithms/multigrid/Geometry.h

@@ -67,8 +67,8 @@ public:
   }
 
   int point(int dir, int disp) {
-    assert(disp == -1 || disp == 0 || disp == 1);
+    GRID_ASSERT(disp == -1 || disp == 0 || disp == 1);
-    assert(base+0 <= dir && dir < base+4);
+    GRID_ASSERT(base+0 <= dir && dir < base+4);
 
     // directions faster index = new indexing
     // 4d (base = 0):
@@ -131,7 +131,7 @@ public:
 	return p;
       }
     }
-    assert(0);
+    GRID_ASSERT(0);
     return -1;
   }
   void BuildShifts(void)

Grid/allocator/AlignedAllocator.h

@@ -57,7 +57,7 @@ public:
     if ( (_Tp*)ptr == (_Tp *) NULL ) {
       printf("Grid CPU Allocator got NULL for %lu bytes\n",(unsigned long) bytes );
     }
-    assert( ( (_Tp*)ptr != (_Tp *)NULL ) );
+    GRID_ASSERT( ( (_Tp*)ptr != (_Tp *)NULL ) );
     return ptr;
   }
 
@@ -69,7 +69,7 @@ public:
   }
 
   // FIXME: hack for the copy constructor: it must be avoided to avoid single thread loop
-  void construct(pointer __p, const _Tp& __val) { assert(0);};
+  void construct(pointer __p, const _Tp& __val) { };
   void construct(pointer __p) { };
   void destroy(pointer __p) { };
 };
@@ -106,7 +106,7 @@ public:
     if ( (_Tp*)ptr == (_Tp *) NULL ) {
       printf("Grid Shared Allocator got NULL for %lu bytes\n",(unsigned long) bytes );
     }
-    assert( ( (_Tp*)ptr != (_Tp *)NULL ) );
+    GRID_ASSERT( ( (_Tp*)ptr != (_Tp *)NULL ) );
     return ptr;
   }
 
@@ -154,7 +154,7 @@ public:
     if ( (_Tp*)ptr == (_Tp *) NULL ) {
       printf("Grid Device Allocator got NULL for %lu bytes\n",(unsigned long) bytes );
     }
-    assert( ( (_Tp*)ptr != (_Tp *)NULL ) );
+    GRID_ASSERT( ( (_Tp*)ptr != (_Tp *)NULL ) );
     return ptr;
   }
 
@@ -174,19 +174,10 @@ template<typename _Tp>  inline bool operator!=(const devAllocator<_Tp>&, const d
 ////////////////////////////////////////////////////////////////////////////////
 // Template typedefs
 ////////////////////////////////////////////////////////////////////////////////
-#ifdef ACCELERATOR_CSHIFT
+template<class T> using hostVector          = std::vector<T,alignedAllocator<T> >;           // Needs autoview
-// Cshift on device
+template<class T> using Vector              = std::vector<T,uvmAllocator<T> >;               // Really want to deprecate
-template<class T> using cshiftAllocator = devAllocator<T>;
+template<class T> using uvmVector           = std::vector<T,uvmAllocator<T> >;               // auto migrating page
-#else
+template<class T> using deviceVector        = std::vector<T,devAllocator<T> >;               // device vector
-// 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
@@ -197,8 +188,9 @@ 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(std::vector<T> &refer_to_me,ViewMode _mode)
+  vecView(Vector<T> &refer_to_me,ViewMode _mode)
   {
     cpu_ptr = &refer_to_me[0];
     size = refer_to_me.size();
@@ -214,22 +206,12 @@ template<class T> class vecView
   }
 };
 
-template<class T> vecView<T> VectorView(std::vector<T> &vec,ViewMode _mode)
+template<class T> vecView<T> VectorView(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);

Grid/allocator/MemoryManager.cc

@@ -292,7 +292,7 @@ void *MemoryManager::Insert(void *ptr,size_t bytes,int type)
 void *MemoryManager::Insert(void *ptr,size_t bytes,AllocationCacheEntry *entries,int ncache,int &victim, uint64_t &cacheBytes) 
 {
 #ifdef GRID_OMP
-  assert(omp_in_parallel()==0);
+  GRID_ASSERT(omp_in_parallel()==0);
 #endif 
 
   if (ncache == 0) return ptr;
@@ -345,7 +345,7 @@ void *MemoryManager::Lookup(size_t bytes,int type)
 void *MemoryManager::Lookup(size_t bytes,AllocationCacheEntry *entries,int ncache,uint64_t & cacheBytes) 
 {
 #ifdef GRID_OMP
-  assert(omp_in_parallel()==0);
+  GRID_ASSERT(omp_in_parallel()==0);
 #endif 
   for(int e=0;e<ncache;e++){
     if ( entries[e].valid && ( entries[e].bytes == bytes ) ) {

Grid/allocator/MemoryManagerCache.cc

@@ -1,16 +1,15 @@
 #include <Grid/GridCore.h>
 #ifndef GRID_UVM
 
-#warning "Using explicit device memory copies"
 NAMESPACE_BEGIN(Grid);
 
 #define MAXLINE 512
 static char print_buffer [ MAXLINE ];
 
-#define mprintf(...) snprintf (print_buffer,MAXLINE, __VA_ARGS__ ); std::cout << GridLogMemory << print_buffer;
+#define mprintf(...) snprintf (print_buffer,MAXLINE, __VA_ARGS__ ); std::cout << GridLogMemory << print_buffer << std::endl;
-#define dprintf(...) snprintf (print_buffer,MAXLINE, __VA_ARGS__ ); std::cout << GridLogDebug << print_buffer;
+#define dprintf(...) snprintf (print_buffer,MAXLINE, __VA_ARGS__ ); std::cout << GridLogDebug  << print_buffer << std::endl;
 //#define dprintf(...) 
-
+//#define mprintf(...) 
 
 ////////////////////////////////////////////////////////////
 // For caching copies of data on device
@@ -51,12 +50,12 @@ int   MemoryManager::EntryPresent(uint64_t CpuPtr)
 {
   if(AccViewTable.empty()) return 0;
 
-  auto count = AccViewTable.count(CpuPtr);  assert((count==0)||(count==1));
+  auto count = AccViewTable.count(CpuPtr);  GRID_ASSERT((count==0)||(count==1));
   return count;
 }
 void  MemoryManager::EntryCreate(uint64_t CpuPtr,size_t bytes,ViewMode mode,ViewAdvise hint)
 {
-  assert(!EntryPresent(CpuPtr));
+  GRID_ASSERT(!EntryPresent(CpuPtr));
   AcceleratorViewEntry AccCache;
   AccCache.CpuPtr = CpuPtr;
   AccCache.AccPtr = (uint64_t)NULL;
@@ -70,9 +69,9 @@ void  MemoryManager::EntryCreate(uint64_t CpuPtr,size_t bytes,ViewMode mode,View
 }
 MemoryManager::AccViewTableIterator MemoryManager::EntryLookup(uint64_t CpuPtr)
 {
-  assert(EntryPresent(CpuPtr));
+  GRID_ASSERT(EntryPresent(CpuPtr));
   auto AccCacheIterator = AccViewTable.find(CpuPtr);
-  assert(AccCacheIterator!=AccViewTable.end());
+  GRID_ASSERT(AccCacheIterator!=AccViewTable.end());
   return AccCacheIterator;
 }
 void MemoryManager::EntryErase(uint64_t CpuPtr)
@@ -82,7 +81,7 @@ void MemoryManager::EntryErase(uint64_t CpuPtr)
 }
 void  MemoryManager::LRUinsert(AcceleratorViewEntry &AccCache)
 {
-  assert(AccCache.LRU_valid==0);
+  GRID_ASSERT(AccCache.LRU_valid==0);
   if (AccCache.transient) { 
     LRU.push_back(AccCache.CpuPtr);
     AccCache.LRU_entry = --LRU.end();
@@ -95,7 +94,7 @@ void  MemoryManager::LRUinsert(AcceleratorViewEntry &AccCache)
 }
 void  MemoryManager::LRUremove(AcceleratorViewEntry &AccCache)
 {
-  assert(AccCache.LRU_valid==1);
+  GRID_ASSERT(AccCache.LRU_valid==1);
   LRU.erase(AccCache.LRU_entry);
   AccCache.LRU_valid = 0;
   DeviceLRUBytes-=AccCache.bytes;
@@ -109,19 +108,19 @@ void MemoryManager::AccDiscard(AcceleratorViewEntry &AccCache)
   // Remove from Accelerator, remove entry, without flush
   // Cannot be locked. If allocated Must be in LRU pool.
   ///////////////////////////////////////////////////////////
-  assert(AccCache.state!=Empty);
+  GRID_ASSERT(AccCache.state!=Empty);
   
-  dprintf("MemoryManager: Discard(%lx) %lx\n",(uint64_t)AccCache.CpuPtr,(uint64_t)AccCache.AccPtr); 
+  dprintf("MemoryManager: Discard(%lx) %lx",(uint64_t)AccCache.CpuPtr,(uint64_t)AccCache.AccPtr); 
-  assert(AccCache.accLock==0);
+  GRID_ASSERT(AccCache.accLock==0);
-  assert(AccCache.cpuLock==0);
+  GRID_ASSERT(AccCache.cpuLock==0);
-  assert(AccCache.CpuPtr!=(uint64_t)NULL);
+  GRID_ASSERT(AccCache.CpuPtr!=(uint64_t)NULL);
   if(AccCache.AccPtr) {
     AcceleratorFree((void *)AccCache.AccPtr,AccCache.bytes);
     DeviceDestroy++;
     DeviceBytes   -=AccCache.bytes;
     LRUremove(AccCache);
     AccCache.AccPtr=(uint64_t) NULL;
-    dprintf("MemoryManager: Free(%lx) LRU %ld Total %ld\n",(uint64_t)AccCache.AccPtr,DeviceLRUBytes,DeviceBytes);  
+    dprintf("MemoryManager: Free(%lx) LRU %ld Total %ld",(uint64_t)AccCache.AccPtr,DeviceLRUBytes,DeviceBytes);  
   }
   uint64_t CpuPtr = AccCache.CpuPtr;
   EntryErase(CpuPtr);
@@ -139,9 +138,9 @@ void MemoryManager::Evict(AcceleratorViewEntry &AccCache)
   //                          Take these OUT LRU queue when CPU locked?
   //                          Cannot take out the table as cpuLock data is important.
   ///////////////////////////////////////////////////////////////////////////
-  assert(AccCache.state!=Empty);
+  GRID_ASSERT(AccCache.state!=Empty);
   
-  mprintf("MemoryManager: Evict CpuPtr %lx AccPtr %lx cpuLock %ld accLock %ld\n",
+  mprintf("MemoryManager: Evict CpuPtr %lx AccPtr %lx cpuLock %ld accLock %ld",
 	  (uint64_t)AccCache.CpuPtr,(uint64_t)AccCache.AccPtr,
 	  (uint64_t)AccCache.cpuLock,(uint64_t)AccCache.accLock); 
   if (AccCache.accLock!=0) return;
@@ -155,7 +154,7 @@ void MemoryManager::Evict(AcceleratorViewEntry &AccCache)
     AccCache.AccPtr=(uint64_t)NULL;
     AccCache.state=CpuDirty; // CPU primary now
     DeviceBytes   -=AccCache.bytes;
-    dprintf("MemoryManager: Free(AccPtr %lx) footprint now %ld \n",(uint64_t)AccCache.AccPtr,DeviceBytes);  
+    dprintf("MemoryManager: Free(AccPtr %lx) footprint now %ld ",(uint64_t)AccCache.AccPtr,DeviceBytes);  
   }
   //  uint64_t CpuPtr = AccCache.CpuPtr;
   DeviceEvictions++;
@@ -163,28 +162,30 @@ void MemoryManager::Evict(AcceleratorViewEntry &AccCache)
 }
 void MemoryManager::Flush(AcceleratorViewEntry &AccCache)
 {
-  assert(AccCache.state==AccDirty);
+  GRID_ASSERT(AccCache.state==AccDirty);
-  assert(AccCache.cpuLock==0);
+  GRID_ASSERT(AccCache.cpuLock==0);
-  assert(AccCache.accLock==0);
+  GRID_ASSERT(AccCache.accLock==0);
-  assert(AccCache.AccPtr!=(uint64_t)NULL);
+  GRID_ASSERT(AccCache.AccPtr!=(uint64_t)NULL);
-  assert(AccCache.CpuPtr!=(uint64_t)NULL);
+  GRID_ASSERT(AccCache.CpuPtr!=(uint64_t)NULL);
   acceleratorCopyFromDevice((void *)AccCache.AccPtr,(void *)AccCache.CpuPtr,AccCache.bytes);
-  mprintf("MemoryManager: acceleratorCopyFromDevice Flush AccPtr %lx -> CpuPtr %lx\n",(uint64_t)AccCache.AccPtr,(uint64_t)AccCache.CpuPtr); fflush(stdout);
+  mprintf("MemoryManager: acceleratorCopyFromDevice Flush size %ld AccPtr %lx -> CpuPtr %lx",(uint64_t)AccCache.bytes,(uint64_t)AccCache.AccPtr,(uint64_t)AccCache.CpuPtr); fflush(stdout);
   DeviceToHostBytes+=AccCache.bytes;
   DeviceToHostXfer++;
   AccCache.state=Consistent;
 }
 void MemoryManager::Clone(AcceleratorViewEntry &AccCache)
 {
-  assert(AccCache.state==CpuDirty);
+  GRID_ASSERT(AccCache.state==CpuDirty);
-  assert(AccCache.cpuLock==0);
+  GRID_ASSERT(AccCache.cpuLock==0);
-  assert(AccCache.accLock==0);
+  GRID_ASSERT(AccCache.accLock==0);
-  assert(AccCache.CpuPtr!=(uint64_t)NULL);
+  GRID_ASSERT(AccCache.CpuPtr!=(uint64_t)NULL);
   if(AccCache.AccPtr==(uint64_t)NULL){
     AccCache.AccPtr=(uint64_t)AcceleratorAllocate(AccCache.bytes);
     DeviceBytes+=AccCache.bytes;
   }
-  mprintf("MemoryManager: acceleratorCopyToDevice   Clone AccPtr %lx <- CpuPtr %lx\n",(uint64_t)AccCache.AccPtr,(uint64_t)AccCache.CpuPtr); fflush(stdout);
+  mprintf("MemoryManager: acceleratorCopyToDevice   Clone size %ld AccPtr %lx <- CpuPtr %lx",
+	  (uint64_t)AccCache.bytes,
+	  (uint64_t)AccCache.AccPtr,(uint64_t)AccCache.CpuPtr); fflush(stdout);
   acceleratorCopyToDevice((void *)AccCache.CpuPtr,(void *)AccCache.AccPtr,AccCache.bytes);
   HostToDeviceBytes+=AccCache.bytes;
   HostToDeviceXfer++;
@@ -193,10 +194,10 @@ void MemoryManager::Clone(AcceleratorViewEntry &AccCache)
 
 void MemoryManager::CpuDiscard(AcceleratorViewEntry &AccCache)
 {
-  assert(AccCache.state!=Empty);
+  GRID_ASSERT(AccCache.state!=Empty);
-  assert(AccCache.cpuLock==0);
+  GRID_ASSERT(AccCache.cpuLock==0);
-  assert(AccCache.accLock==0);
+  GRID_ASSERT(AccCache.accLock==0);
-  assert(AccCache.CpuPtr!=(uint64_t)NULL);
+  GRID_ASSERT(AccCache.CpuPtr!=(uint64_t)NULL);
   if(AccCache.AccPtr==(uint64_t)NULL){
     AccCache.AccPtr=(uint64_t)AcceleratorAllocate(AccCache.bytes);
     DeviceBytes+=AccCache.bytes;
@@ -210,33 +211,36 @@ void MemoryManager::CpuDiscard(AcceleratorViewEntry &AccCache)
 void MemoryManager::ViewClose(void* Ptr,ViewMode mode)
 {
   if( (mode==AcceleratorRead)||(mode==AcceleratorWrite)||(mode==AcceleratorWriteDiscard) ){
-    dprintf("AcceleratorViewClose %lx\n",(uint64_t)Ptr);
+    dprintf("AcceleratorViewClose %lx",(uint64_t)Ptr);
     AcceleratorViewClose((uint64_t)Ptr);
   } else if( (mode==CpuRead)||(mode==CpuWrite)){
     CpuViewClose((uint64_t)Ptr);
   } else { 
-    assert(0);
+    GRID_ASSERT(0);
   }
 }
 void *MemoryManager::ViewOpen(void* _CpuPtr,size_t bytes,ViewMode mode,ViewAdvise hint)
 {
   uint64_t CpuPtr = (uint64_t)_CpuPtr;
   if( (mode==AcceleratorRead)||(mode==AcceleratorWrite)||(mode==AcceleratorWriteDiscard) ){
-    dprintf("AcceleratorViewOpen %lx\n",(uint64_t)CpuPtr);
+    dprintf("AcceleratorViewOpen %lx",(uint64_t)CpuPtr);
     return (void *) AcceleratorViewOpen(CpuPtr,bytes,mode,hint);
   } else if( (mode==CpuRead)||(mode==CpuWrite)){
     return (void *)CpuViewOpen(CpuPtr,bytes,mode,hint);
   } else { 
-    assert(0);
+    GRID_ASSERT(0);
     return NULL;
   }
 }
 void  MemoryManager::EvictVictims(uint64_t bytes)
 {
-  assert(bytes<DeviceMaxBytes);
+  if(bytes>=DeviceMaxBytes) {
+    printf("EvictVictims bytes %ld DeviceMaxBytes %ld\n",bytes,DeviceMaxBytes);
+  }
+  GRID_ASSERT(bytes<DeviceMaxBytes);
   while(bytes+DeviceLRUBytes > DeviceMaxBytes){
     if ( DeviceLRUBytes > 0){
-      assert(LRU.size()>0);
+      GRID_ASSERT(LRU.size()>0);
       uint64_t victim = LRU.back(); // From the LRU
       auto AccCacheIterator = EntryLookup(victim);
       auto & AccCache = AccCacheIterator->second;
@@ -260,19 +264,19 @@ uint64_t MemoryManager::AcceleratorViewOpen(uint64_t CpuPtr,size_t bytes,ViewMod
   if (!AccCache.AccPtr) {
     EvictVictims(bytes); 
   } 
-  assert((mode==AcceleratorRead)||(mode==AcceleratorWrite)||(mode==AcceleratorWriteDiscard));
+  GRID_ASSERT((mode==AcceleratorRead)||(mode==AcceleratorWrite)||(mode==AcceleratorWriteDiscard));
 
-  assert(AccCache.cpuLock==0);  // Programming error
+  GRID_ASSERT(AccCache.cpuLock==0);  // Programming error
 
   if(AccCache.state!=Empty) {
-    dprintf("ViewOpen found entry %lx %lx : %ld %ld accLock %ld\n",
+    dprintf("ViewOpen found entry %lx %lx : sizes %ld %ld accLock %ld",
 		    (uint64_t)AccCache.CpuPtr,
 		    (uint64_t)CpuPtr,
 		    (uint64_t)AccCache.bytes,
 	            (uint64_t)bytes,
 		    (uint64_t)AccCache.accLock);
-    assert(AccCache.CpuPtr == CpuPtr);
+    GRID_ASSERT(AccCache.CpuPtr == CpuPtr);
-    assert(AccCache.bytes  ==bytes);
+    GRID_ASSERT(AccCache.bytes  ==bytes);
   }
 /*
  *  State transitions and actions
@@ -289,7 +293,7 @@ uint64_t MemoryManager::AcceleratorViewOpen(uint64_t CpuPtr,size_t bytes,ViewMod
  *  AccWrite AccDirty   AccDirty       -        - 
  */
   if(AccCache.state==Empty) {
-    assert(AccCache.LRU_valid==0);
+    GRID_ASSERT(AccCache.LRU_valid==0);
     AccCache.CpuPtr = CpuPtr;
     AccCache.AccPtr = (uint64_t)NULL;
     AccCache.bytes  = bytes;
@@ -305,7 +309,7 @@ uint64_t MemoryManager::AcceleratorViewOpen(uint64_t CpuPtr,size_t bytes,ViewMod
       AccCache.state  = Consistent; // Empty + AccRead => Consistent
     }
     AccCache.accLock= 1;
-    dprintf("Copied Empty entry into device accLock= %d\n",AccCache.accLock);
+    dprintf("Copied Empty entry into device accLock= %d",AccCache.accLock);
   } else if(AccCache.state==CpuDirty ){
     if(mode==AcceleratorWriteDiscard) {
       CpuDiscard(AccCache);
@@ -318,30 +322,30 @@ uint64_t MemoryManager::AcceleratorViewOpen(uint64_t CpuPtr,size_t bytes,ViewMod
       AccCache.state  = Consistent; // CpuDirty + AccRead => Consistent
     }
     AccCache.accLock++;
-    dprintf("CpuDirty entry into device ++accLock= %d\n",AccCache.accLock);
+    dprintf("CpuDirty entry into device ++accLock= %d",AccCache.accLock);
   } else if(AccCache.state==Consistent) {
     if((mode==AcceleratorWrite)||(mode==AcceleratorWriteDiscard))
       AccCache.state  = AccDirty;   // Consistent + AcceleratorWrite=> AccDirty
     else
       AccCache.state  = Consistent; // Consistent + AccRead => Consistent
     AccCache.accLock++;
-    dprintf("Consistent entry into device ++accLock= %d\n",AccCache.accLock);
+    dprintf("Consistent entry into device ++accLock= %d",AccCache.accLock);
   } else if(AccCache.state==AccDirty) {
     if((mode==AcceleratorWrite)||(mode==AcceleratorWriteDiscard))
       AccCache.state  = AccDirty; // AccDirty + AcceleratorWrite=> AccDirty
     else
       AccCache.state  = AccDirty; // AccDirty + AccRead => AccDirty
     AccCache.accLock++;
-    dprintf("AccDirty entry ++accLock= %d\n",AccCache.accLock);
+    dprintf("AccDirty entry ++accLock= %d",AccCache.accLock);
   } else {
-    assert(0);
+    GRID_ASSERT(0);
   }
 
-  assert(AccCache.accLock>0);
+  GRID_ASSERT(AccCache.accLock>0);
   // If view is opened on device must remove from LRU
   if(AccCache.LRU_valid==1){
     // must possibly remove from LRU as now locked on GPU
-    dprintf("AccCache entry removed from LRU \n");
+    dprintf("AccCache entry removed from LRU ");
     LRUremove(AccCache);
   }
 
@@ -358,16 +362,16 @@ void MemoryManager::AcceleratorViewClose(uint64_t CpuPtr)
   auto AccCacheIterator = EntryLookup(CpuPtr);
   auto & AccCache = AccCacheIterator->second;
 
-  assert(AccCache.cpuLock==0);
+  GRID_ASSERT(AccCache.cpuLock==0);
-  assert(AccCache.accLock>0);
+  GRID_ASSERT(AccCache.accLock>0);
 
   AccCache.accLock--;
   // Move to LRU queue if not locked and close on device
   if(AccCache.accLock==0) {
-    dprintf("AccleratorViewClose %lx AccLock decremented to %ld move to LRU queue\n",(uint64_t)CpuPtr,(uint64_t)AccCache.accLock);
+    dprintf("AccleratorViewClose %lx AccLock decremented to %ld move to LRU queue",(uint64_t)CpuPtr,(uint64_t)AccCache.accLock);
     LRUinsert(AccCache);
   } else {
-    dprintf("AccleratorViewClose %lx AccLock decremented to %ld\n",(uint64_t)CpuPtr,(uint64_t)AccCache.accLock);
+    dprintf("AccleratorViewClose %lx AccLock decremented to %ld",(uint64_t)CpuPtr,(uint64_t)AccCache.accLock);
   }
 }
 void MemoryManager::CpuViewClose(uint64_t CpuPtr)
@@ -375,8 +379,8 @@ void MemoryManager::CpuViewClose(uint64_t CpuPtr)
   auto AccCacheIterator = EntryLookup(CpuPtr);
   auto & AccCache = AccCacheIterator->second;
 
-  assert(AccCache.cpuLock>0);
+  GRID_ASSERT(AccCache.cpuLock>0);
-  assert(AccCache.accLock==0);
+  GRID_ASSERT(AccCache.accLock==0);
 
   AccCache.cpuLock--;
 }
@@ -409,12 +413,12 @@ uint64_t MemoryManager::CpuViewOpen(uint64_t CpuPtr,size_t bytes,ViewMode mode,V
   //    EvictVictims(bytes);
   //  }
 
-  assert((mode==CpuRead)||(mode==CpuWrite));
+  GRID_ASSERT((mode==CpuRead)||(mode==CpuWrite));
-  assert(AccCache.accLock==0);  // Programming error
+  GRID_ASSERT(AccCache.accLock==0);  // Programming error
 
   if(AccCache.state!=Empty) {
-    assert(AccCache.CpuPtr == CpuPtr);
+    GRID_ASSERT(AccCache.CpuPtr == CpuPtr);
-    assert(AccCache.bytes==bytes);
+    GRID_ASSERT(AccCache.bytes==bytes);
   }
 
   if(AccCache.state==Empty) {
@@ -429,20 +433,20 @@ uint64_t MemoryManager::CpuViewOpen(uint64_t CpuPtr,size_t bytes,ViewMode mode,V
     AccCache.state = CpuDirty; // CpuDirty +CpuRead/CpuWrite => CpuDirty
     AccCache.cpuLock++;
   } else if(AccCache.state==Consistent) {
-    assert(AccCache.AccPtr != (uint64_t)NULL);
+    GRID_ASSERT(AccCache.AccPtr != (uint64_t)NULL);
     if(mode==CpuWrite)
       AccCache.state = CpuDirty;   // Consistent +CpuWrite => CpuDirty
     else 
       AccCache.state = Consistent; // Consistent +CpuRead  => Consistent
     AccCache.cpuLock++;
   } else if(AccCache.state==AccDirty) {
-    assert(AccCache.AccPtr != (uint64_t)NULL);
+    GRID_ASSERT(AccCache.AccPtr != (uint64_t)NULL);
     Flush(AccCache);
     if(mode==CpuWrite) AccCache.state = CpuDirty;   // AccDirty +CpuWrite => CpuDirty, Flush
     else            AccCache.state = Consistent; // AccDirty +CpuRead  => Consistent, Flush
     AccCache.cpuLock++;
   } else {
-    assert(0); // should be unreachable
+    GRID_ASSERT(0); // should be unreachable
   }
 
   AccCache.transient= transient? EvictNext : 0;
@@ -524,12 +528,12 @@ void MemoryManager::Audit(std::string s)
   std::cout << " Memory Manager::Audit() from "<<s<<std::endl;
   for(auto it=LRU.begin();it!=LRU.end();it++){
     uint64_t cpuPtr = *it;
-    assert(EntryPresent(cpuPtr));
+    GRID_ASSERT(EntryPresent(cpuPtr));
     auto AccCacheIterator = EntryLookup(cpuPtr);
     auto & AccCache = AccCacheIterator->second;
     LruBytes2+=AccCache.bytes;
-    assert(AccCache.LRU_valid==1);
+    GRID_ASSERT(AccCache.LRU_valid==1);
-    assert(AccCache.LRU_entry==it);
+    GRID_ASSERT(AccCache.LRU_entry==it);
   }
   std::cout << " Memory Manager::Audit() LRU queue matches table entries "<<std::endl;
 
@@ -548,7 +552,7 @@ void MemoryManager::Audit(std::string s)
     if( AccCache.LRU_valid ) LruCnt++;
     
     if ( AccCache.cpuLock || AccCache.accLock ) {
-      assert(AccCache.LRU_valid==0);
+      GRID_ASSERT(AccCache.LRU_valid==0);
 
       std::cout << GridLogError << s<< "\n\t 0x"<<std::hex<<AccCache.CpuPtr<<std::dec
 		<< "\t0x"<<std::hex<<AccCache.AccPtr<<std::dec<<"\t" <<str
@@ -557,16 +561,16 @@ void MemoryManager::Audit(std::string s)
 		<< "\t LRUvalid " << AccCache.LRU_valid<<std::endl;
     }
 
-    assert( AccCache.cpuLock== 0 ) ;
+    GRID_ASSERT( AccCache.cpuLock== 0 ) ;
-    assert( AccCache.accLock== 0 ) ;
+    GRID_ASSERT( AccCache.accLock== 0 ) ;
   }
   std::cout << " Memory Manager::Audit() no locked table entries "<<std::endl;
-  assert(LruBytes1==LruBytes2);
+  GRID_ASSERT(LruBytes1==LruBytes2);
-  assert(LruBytes1==DeviceLRUBytes);
+  GRID_ASSERT(LruBytes1==DeviceLRUBytes);
   std::cout << " Memory Manager::Audit() evictable bytes matches sum over table "<<std::endl;
-  assert(AccBytes==DeviceBytes);
+  GRID_ASSERT(AccBytes==DeviceBytes);
   std::cout << " Memory Manager::Audit() device bytes matches sum over table "<<std::endl;
-  assert(LruCnt == LRU.size());
+  GRID_ASSERT(LruCnt == LRU.size());
   std::cout << " Memory Manager::Audit() LRU entry count matches "<<std::endl;
 
 }

Grid/allocator/MemoryStats.cc

@@ -10,16 +10,16 @@ void check_huge_pages(void *Buf,uint64_t BYTES)
 {
 #ifdef __linux__
   int fd = open("/proc/self/pagemap", O_RDONLY);
-  assert(fd >= 0);
+  GRID_ASSERT(fd >= 0);
   const int page_size = 4096;
   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;
-  uint64_t pagedata[npages];
+  std::vector<uint64_t> pagedata(npages);
   uint64_t ret = lseek(fd, offset, SEEK_SET);
-  assert(ret == offset);
+  GRID_ASSERT(ret == offset);
-  ret = ::read(fd, pagedata, sizeof(uint64_t)*npages);
+  ret = ::read(fd, &pagedata[0], sizeof(uint64_t)*npages);
-  assert(ret == sizeof(uint64_t) * npages);
+  GRID_ASSERT(ret == sizeof(uint64_t) * npages);
   int nhugepages = npages / 512;
   int n4ktotal, nnothuge;
   n4ktotal = 0;

Grid/cartesian/Cartesian_base.h

@@ -82,6 +82,7 @@ public:
   bool _isCheckerBoarded; 
   int        LocallyPeriodic;
   Coordinate _checker_dim_mask;
+  int              _checker_dim;
 
 public:
 
@@ -91,7 +92,6 @@ 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;
@@ -165,7 +165,7 @@ public:
     //
     if ( _simd_layout[dimension] > 2 ) { 
       for(int d=0;d<_ndimension;d++){
-	if ( d != dimension ) assert ( (_simd_layout[d]==1)  );
+	if ( d != dimension ) GRID_ASSERT ( (_simd_layout[d]==1)  );
       }
       permute_type = RotateBit; // How to specify distance; this is not just direction.
       return permute_type;
@@ -187,7 +187,7 @@ public:
   inline int64_t gSites(void) const { return (int64_t)_isites*(int64_t)_osites*(int64_t)_Nprocessors; }; 
   inline int Nd    (void) const { return _ndimension;};
 
-  inline const Coordinate LocalStarts(void)             { return _lstart;    };
+  inline const Coordinate &LocalStarts(void)            { return _lstart;    };
   inline const Coordinate &FullDimensions(void)         { return _fdimensions;};
   inline const Coordinate &GlobalDimensions(void)       { return _gdimensions;};
   inline const Coordinate &LocalDimensions(void)        { return _ldimensions;};
@@ -216,11 +216,11 @@ public:
   // Global addressing
   ////////////////////////////////////////////////////////////////
   void GlobalIndexToGlobalCoor(int64_t gidx,Coordinate &gcoor){
-    assert(gidx< gSites());
+    GRID_ASSERT(gidx< gSites());
     Lexicographic::CoorFromIndex(gcoor,gidx,_gdimensions);
   }
   void LocalIndexToLocalCoor(int lidx,Coordinate &lcoor){
-    assert(lidx<lSites());
+    GRID_ASSERT(lidx<lSites());
     Lexicographic::CoorFromIndex(lcoor,lidx,_ldimensions);
   }
   void GlobalCoorToGlobalIndex(const Coordinate & gcoor,int64_t & gidx){

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;
   }
@@ -106,6 +106,7 @@ public:
     _rdimensions.resize(_ndimension);
     _simd_layout.resize(_ndimension);
     _checker_dim_mask.resize(_ndimension);;
+    _checker_dim = -1;
     _lstart.resize(_ndimension);
     _lend.resize(_ndimension);
 
@@ -127,10 +128,10 @@ public:
         // Use a reduced simd grid
         _ldimensions[d] = _gdimensions[d] / _processors[d]; //local dimensions
         //std::cout << _ldimensions[d] << "  " << _gdimensions[d] << "  " << _processors[d] << std::endl;
-        assert(_ldimensions[d] * _processors[d] == _gdimensions[d]);
+        GRID_ASSERT(_ldimensions[d] * _processors[d] == _gdimensions[d]);
 
         _rdimensions[d] = _ldimensions[d] / _simd_layout[d]; //overdecomposition
-        assert(_rdimensions[d] * _simd_layout[d] == _ldimensions[d]);
+        GRID_ASSERT(_rdimensions[d] * _simd_layout[d] == _ldimensions[d]);
 
         _lstart[d] = _processor_coor[d] * _ldimensions[d];
         _lend[d] = _processor_coor[d] * _ldimensions[d] + _ldimensions[d] - 1;

Grid/cartesian/Cartesian_red_black.h

@@ -57,17 +57,17 @@ class GridRedBlackCartesian : public GridBase
 {
 public:
   //  Coordinate _checker_dim_mask;
-  int              _checker_dim;
+  //  int              _checker_dim;
   std::vector<int> _checker_board;
 
-  virtual int CheckerDim(void){ return _checker_dim; };
+  virtual int isCheckerBoarded(void) const { return 1; };
   virtual int CheckerBoarded(int dim){
     if( dim==_checker_dim) return 1;
     else return 0;
   }
   virtual int CheckerBoard(const Coordinate &site){
     int linear=0;
-    assert(site.size()==_ndimension);
+    GRID_ASSERT(site.size()==_ndimension);
     for(int d=0;d<_ndimension;d++){ 
       if(_checker_dim_mask[d])
 	linear=linear+site[d];
@@ -160,11 +160,11 @@ public:
 
       _isCheckerBoarded = true;
     _checker_dim = checker_dim;
-    assert(checker_dim_mask[checker_dim] == 1);
+    GRID_ASSERT(checker_dim_mask[checker_dim] == 1);
     _ndimension = dimensions.size();
-    assert(checker_dim_mask.size() == _ndimension);
+    GRID_ASSERT(checker_dim_mask.size() == _ndimension);
-    assert(processor_grid.size() == _ndimension);
+    GRID_ASSERT(processor_grid.size() == _ndimension);
-    assert(simd_layout.size() == _ndimension);
+    GRID_ASSERT(simd_layout.size() == _ndimension);
 
     _fdimensions.resize(_ndimension);
     _gdimensions.resize(_ndimension);
@@ -190,20 +190,20 @@ public:
 
         if (d == _checker_dim)
 	  {
-	    assert((_gdimensions[d] & 0x1) == 0);
+	    GRID_ASSERT((_gdimensions[d] & 0x1) == 0);
 	    _gdimensions[d] = _gdimensions[d] / 2; // Remove a checkerboard
 	    _gsites /= 2;
 	  }
         _ldimensions[d] = _gdimensions[d] / _processors[d];
-        assert(_ldimensions[d] * _processors[d] == _gdimensions[d]);
+        GRID_ASSERT(_ldimensions[d] * _processors[d] == _gdimensions[d]);
         _lstart[d] = _processor_coor[d] * _ldimensions[d];
         _lend[d] = _processor_coor[d] * _ldimensions[d] + _ldimensions[d] - 1;
 
         // Use a reduced simd grid
         _simd_layout[d] = simd_layout[d];
         _rdimensions[d] = _ldimensions[d] / _simd_layout[d]; // this is not checking if this is integer
-        assert(_rdimensions[d] * _simd_layout[d] == _ldimensions[d]);
+        GRID_ASSERT(_rdimensions[d] * _simd_layout[d] == _ldimensions[d]);
-        assert(_rdimensions[d] > 0);
+        GRID_ASSERT(_rdimensions[d] > 0);
 
         // all elements of a simd vector must have same checkerboard.
         // If Ls vectorised, this must still be the case; e.g. dwf rb5d

Grid/communicator/Communicator_base.cc

@@ -57,18 +57,29 @@ int                      CartesianCommunicator::ProcessorCount(void)    { return
 // very VERY rarely (Log, serial RNG) we need world without a grid
 ////////////////////////////////////////////////////////////////////////////////
 
+#ifdef USE_GRID_REDUCTION
+void CartesianCommunicator::GlobalSum(ComplexF &c)
+{
+  GlobalSumP2P(c);
+}
+void CartesianCommunicator::GlobalSum(ComplexD &c)
+{
+  GlobalSumP2P(c);
+}
+#else
 void CartesianCommunicator::GlobalSum(ComplexF &c)
 {
   GlobalSumVector((float *)&c,2);
 }
-void CartesianCommunicator::GlobalSumVector(ComplexF *c,int N)
-{
-  GlobalSumVector((float *)c,2*N);
-}
 void CartesianCommunicator::GlobalSum(ComplexD &c)
 {
   GlobalSumVector((double *)&c,2);
 }
+#endif
+void CartesianCommunicator::GlobalSumVector(ComplexF *c,int N)
+{
+  GlobalSumVector((float *)c,2*N);
+}
 void CartesianCommunicator::GlobalSumVector(ComplexD *c,int N)
 {
   GlobalSumVector((double *)c,2*N);

Grid/communicator/Communicator_base.h

@@ -33,6 +33,8 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 ///////////////////////////////////
 #include <Grid/communicator/SharedMemory.h>
 
+#define NVLINK_GET
+
 NAMESPACE_BEGIN(Grid);
 
 extern bool Stencil_force_mpi ;
@@ -106,7 +108,7 @@ public:
   // very VERY rarely (Log, serial RNG) we need world without a grid
   ////////////////////////////////////////////////////////////////////////////////
   static int  RankWorld(void) ;
-  static void BroadcastWorld(int root,void* data, int bytes);
+  static void BroadcastWorld(int root,void* data, uint64_t bytes);
   static void BarrierWorld(void);
   
   ////////////////////////////////////////////////////////////
@@ -128,6 +130,35 @@ public:
   void GlobalXOR(uint32_t &);
   void GlobalXOR(uint64_t &);
 
+  template<class obj> void GlobalSumP2P(obj &o)
+  {
+    std::vector<obj> column;
+    obj accum = o;
+    int source,dest;
+    for(int d=0;d<_ndimension;d++){
+      column.resize(_processors[d]);
+      column[0] = accum;
+      std::vector<MpiCommsRequest_t> list;
+      for(int p=1;p<_processors[d];p++){
+	ShiftedRanks(d,p,source,dest);
+	SendToRecvFromBegin(list,
+			    &column[0],
+			    dest,
+			    &column[p],
+			    source,
+			    sizeof(obj),d*100+p);
+
+      }
+      if (!list.empty()) // avoid triggering GRID_ASSERT in comms == none
+	CommsComplete(list);
+      for(int p=1;p<_processors[d];p++){
+	accum = accum + column[p];
+      }
+    }
+    Broadcast(0,accum);
+    o=accum;
+  }
+
   template<class obj> void GlobalSum(obj &o){
     typedef typename obj::scalar_type scalar_type;
     int words = sizeof(obj)/sizeof(scalar_type);
@@ -138,32 +169,44 @@ public:
   ////////////////////////////////////////////////////////////
   // Face exchange, buffer swap in translational invariant way
   ////////////////////////////////////////////////////////////
-  void CommsComplete(std::vector<CommsRequest_t> &list);
+  void CommsComplete(std::vector<MpiCommsRequest_t> &list);
-  void SendToRecvFromBegin(std::vector<CommsRequest_t> &list,
+  void SendToRecvFromBegin(std::vector<MpiCommsRequest_t> &list,
 			   void *xmit,
 			   int dest,
 			   void *recv,
 			   int from,
-			   int bytes,int dir);
+			   uint64_t bytes,int dir);
   
   void SendToRecvFrom(void *xmit,
 		      int xmit_to_rank,
 		      void *recv,
 		      int recv_from_rank,
-		      int bytes);
+		      uint64_t bytes);
   
+  int IsOffNode(int rank);
   double StencilSendToRecvFrom(void *xmit,
 			       int xmit_to_rank,int do_xmit,
 			       void *recv,
 			       int recv_from_rank,int do_recv,
-			       int bytes,int dir);
+			       uint64_t bytes,int dir);
 
-  double StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list,
+  double StencilSendToRecvFromPrepare(std::vector<CommsRequest_t> &list,
 				      void *xmit,
 				      int xmit_to_rank,int do_xmit,
 				      void *recv,
 				      int recv_from_rank,int do_recv,
-				    int xbytes,int rbytes,int dir);
+				      uint64_t xbytes,uint64_t rbytes,int dir);
+
+  // Could do a PollHtoD and have a CommsMerge dependence
+  void StencilSendToRecvFromPollDtoH (std::vector<CommsRequest_t> &list);
+  void StencilSendToRecvFromPollIRecv(std::vector<CommsRequest_t> &list);
+
+  double StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list,
+				    void *xmit,void *xmit_comp,
+				    int xmit_to_rank,int do_xmit,
+				    void *recv,void *recv_comp,
+				    int recv_from_rank,int do_recv,
+				    uint64_t xbytes,uint64_t rbytes,int dir);
   
   
   void StencilSendToRecvFromComplete(std::vector<CommsRequest_t> &waitall,int i);
@@ -177,20 +220,20 @@ public:
   ////////////////////////////////////////////////////////////
   // Broadcast a buffer and composite larger
   ////////////////////////////////////////////////////////////
-  void Broadcast(int root,void* data, int bytes);
+  void Broadcast(int root,void* data, uint64_t bytes);
 
   ////////////////////////////////////////////////////////////
   // All2All down one dimension
   ////////////////////////////////////////////////////////////
   template<class T> void AllToAll(int dim,std::vector<T> &in, std::vector<T> &out){
-    assert(dim>=0);
+    GRID_ASSERT(dim>=0);
-    assert(dim<_ndimension);
+    GRID_ASSERT(dim<_ndimension);
-    assert(in.size()==out.size());
+    GRID_ASSERT(in.size()==out.size());
     int numnode = _processors[dim];
     uint64_t bytes=sizeof(T);
     uint64_t words=in.size()/numnode;
-    assert(numnode * words == in.size());
+    GRID_ASSERT(numnode * words == in.size());
-    assert(words < (1ULL<<31));
+    GRID_ASSERT(words < (1ULL<<31));
     AllToAll(dim,(void *)&in[0],(void *)&out[0],words,bytes);
   }
   void AllToAll(int dim  ,void *in,void *out,uint64_t words,uint64_t bytes);

Grid/communicator/Communicator_mpi3.cc

@@ -28,9 +28,17 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #include <Grid/GridCore.h>
 #include <Grid/communicator/SharedMemory.h>
 
+void GridAbort(void) { MPI_Abort(MPI_COMM_WORLD,SIGABRT); }
+extern void * Grid_backtrace_buffer[_NBACKTRACE];
+
 NAMESPACE_BEGIN(Grid);
 
+
 Grid_MPI_Comm       CartesianCommunicator::communicator_world;
+#ifdef GRID_CHECKSUM_COMMS
+uint64_t checksum_index = 1;
+#endif
+
 
 ////////////////////////////////////////////
 // First initialise of comms system
@@ -55,11 +63,11 @@ void CartesianCommunicator::Init(int *argc, char ***argv)
 #endif
     //If only 1 comms thread we require any threading mode other than SINGLE, but for multiple comms threads we need MULTIPLE
     if( (nCommThreads == 1) && (provided == MPI_THREAD_SINGLE) ) {
-      assert(0);
+      GRID_ASSERT(0);
     }
 
     if( (nCommThreads > 1) && (provided != MPI_THREAD_MULTIPLE) ) {
-      assert(0);
+      GRID_ASSERT(0);
     }
   }
 
@@ -80,20 +88,20 @@ void CartesianCommunicator::Init(int *argc, char ***argv)
 void CartesianCommunicator::ShiftedRanks(int dim,int shift,int &source,int &dest)
 {
   int ierr=MPI_Cart_shift(communicator,dim,shift,&source,&dest);
-  assert(ierr==0);
+  GRID_ASSERT(ierr==0);
 }
 int CartesianCommunicator::RankFromProcessorCoor(Coordinate &coor)
 {
   int rank;
   int ierr=MPI_Cart_rank  (communicator, &coor[0], &rank);
-  assert(ierr==0);
+  GRID_ASSERT(ierr==0);
   return rank;
 }
 void  CartesianCommunicator::ProcessorCoorFromRank(int rank, Coordinate &coor)
 {
   coor.resize(_ndimension);
   int ierr=MPI_Cart_coords  (communicator, rank, _ndimension,&coor[0]);
-  assert(ierr==0);
+  GRID_ASSERT(ierr==0);
 }
 
 ////////////////////////////////////////////////////////////////////////////////////////////////////////
@@ -120,8 +128,8 @@ CartesianCommunicator::CartesianCommunicator(const Coordinate &processors)
 //////////////////////////////////
 CartesianCommunicator::CartesianCommunicator(const Coordinate &processors,const CartesianCommunicator &parent,int &srank)
 {
-  _ndimension = processors.size();  assert(_ndimension>=1);
+  _ndimension = processors.size();  GRID_ASSERT(_ndimension>=1);
-  int parent_ndimension = parent._ndimension; assert(_ndimension >= parent._ndimension);
+  int parent_ndimension = parent._ndimension; GRID_ASSERT(_ndimension >= parent._ndimension);
   Coordinate parent_processor_coor(_ndimension,0);
   Coordinate parent_processors    (_ndimension,1);
   Coordinate shm_processors       (_ndimension,1);
@@ -145,7 +153,7 @@ CartesianCommunicator::CartesianCommunicator(const Coordinate &processors,const
     childsize *= processors[d];
   }
   int Nchild = Nparent/childsize;
-  assert (childsize * Nchild == Nparent);
+  GRID_ASSERT (childsize * Nchild == Nparent);
 
   Coordinate ccoor(_ndimension); // coor within subcommunicator
   Coordinate scoor(_ndimension); // coor of split within parent
@@ -171,12 +179,12 @@ CartesianCommunicator::CartesianCommunicator(const Coordinate &processors,const
     // Split the communicator
     ////////////////////////////////////////////////////////////////
     int ierr= MPI_Comm_split(parent.communicator,srank,crank,&comm_split);
-    assert(ierr==0);
+    GRID_ASSERT(ierr==0);
 
   } else {
     srank = 0;
     int ierr = MPI_Comm_dup (parent.communicator,&comm_split);
-    assert(ierr==0);
+    GRID_ASSERT(ierr==0);
   }
 
   //////////////////////////////////////////////////////////////////////////////////////////////////////
@@ -201,7 +209,7 @@ CartesianCommunicator::CartesianCommunicator(const Coordinate &processors,const
     }
   }
   for(int d=0;d<processors.size();d++){
-    assert(_processor_coor[d] == ccoor[d] );
+    GRID_ASSERT(_processor_coor[d] == ccoor[d] );
   }
 }
 
@@ -243,7 +251,7 @@ void CartesianCommunicator::InitFromMPICommunicator(const Coordinate &processors
   for(int i=0;i<_ndimension*2;i++){
     MPI_Comm_dup(communicator,&communicator_halo[i]);
   }
-  assert(Size==_Nprocessors);
+  GRID_ASSERT(Size==_Nprocessors);
 }
 
 CartesianCommunicator::~CartesianCommunicator()
@@ -257,82 +265,103 @@ CartesianCommunicator::~CartesianCommunicator()
     }
   }
 }
-void CartesianCommunicator::GlobalSum(uint32_t &u){
+#ifdef USE_GRID_REDUCTION
-  int ierr=MPI_Allreduce(MPI_IN_PLACE,&u,1,MPI_UINT32_T,MPI_SUM,communicator);
-  assert(ierr==0);
-}
-void CartesianCommunicator::GlobalSum(uint64_t &u){
-  int ierr=MPI_Allreduce(MPI_IN_PLACE,&u,1,MPI_UINT64_T,MPI_SUM,communicator);
-  assert(ierr==0);
-}
-void CartesianCommunicator::GlobalSumVector(uint64_t* u,int N){
-  int ierr=MPI_Allreduce(MPI_IN_PLACE,u,N,MPI_UINT64_T,MPI_SUM,communicator);
-  assert(ierr==0);
-}
-void CartesianCommunicator::GlobalXOR(uint32_t &u){
-  int ierr=MPI_Allreduce(MPI_IN_PLACE,&u,1,MPI_UINT32_T,MPI_BXOR,communicator);
-  assert(ierr==0);
-}
-void CartesianCommunicator::GlobalXOR(uint64_t &u){
-  int ierr=MPI_Allreduce(MPI_IN_PLACE,&u,1,MPI_UINT64_T,MPI_BXOR,communicator);
-  assert(ierr==0);
-}
-void CartesianCommunicator::GlobalMax(float &f)
-{
-  int ierr=MPI_Allreduce(MPI_IN_PLACE,&f,1,MPI_FLOAT,MPI_MAX,communicator);
-  assert(ierr==0);
-}
-void CartesianCommunicator::GlobalMax(double &d)
-{
-  int ierr = MPI_Allreduce(MPI_IN_PLACE,&d,1,MPI_DOUBLE,MPI_MAX,communicator);
-  assert(ierr==0);
-}
 void CartesianCommunicator::GlobalSum(float &f){
-  int ierr=MPI_Allreduce(MPI_IN_PLACE,&f,1,MPI_FLOAT,MPI_SUM,communicator);
+  FlightRecorder::StepLog("GlobalSumP2P");
-  assert(ierr==0);
+  CartesianCommunicator::GlobalSumP2P(f);
-}
-void CartesianCommunicator::GlobalSumVector(float *f,int N)
-{
-  int ierr=MPI_Allreduce(MPI_IN_PLACE,f,N,MPI_FLOAT,MPI_SUM,communicator);
-  assert(ierr==0);
 }
 void CartesianCommunicator::GlobalSum(double &d)
 {
+  FlightRecorder::StepLog("GlobalSumP2P");
+  CartesianCommunicator::GlobalSumP2P(d);
+}
+#else
+void CartesianCommunicator::GlobalSum(float &f){
+  FlightRecorder::StepLog("AllReduce float");
+  int ierr=MPI_Allreduce(MPI_IN_PLACE,&f,1,MPI_FLOAT,MPI_SUM,communicator);
+  GRID_ASSERT(ierr==0);
+}
+void CartesianCommunicator::GlobalSum(double &d)
+{
+  FlightRecorder::StepLog("AllReduce double");
   int ierr = MPI_Allreduce(MPI_IN_PLACE,&d,1,MPI_DOUBLE,MPI_SUM,communicator);
-  assert(ierr==0);
+  GRID_ASSERT(ierr==0);
+}
+#endif
+void CartesianCommunicator::GlobalSum(uint32_t &u){
+  FlightRecorder::StepLog("AllReduce uint32_t");
+  int ierr=MPI_Allreduce(MPI_IN_PLACE,&u,1,MPI_UINT32_T,MPI_SUM,communicator);
+  GRID_ASSERT(ierr==0);
+}
+void CartesianCommunicator::GlobalSum(uint64_t &u){
+  FlightRecorder::StepLog("AllReduce uint64_t");
+  int ierr=MPI_Allreduce(MPI_IN_PLACE,&u,1,MPI_UINT64_T,MPI_SUM,communicator);
+  GRID_ASSERT(ierr==0);
+}
+void CartesianCommunicator::GlobalSumVector(uint64_t* u,int N){
+  FlightRecorder::StepLog("AllReduceVector");
+  int ierr=MPI_Allreduce(MPI_IN_PLACE,u,N,MPI_UINT64_T,MPI_SUM,communicator);
+  GRID_ASSERT(ierr==0);
+}
+void CartesianCommunicator::GlobalXOR(uint32_t &u){
+  int ierr=MPI_Allreduce(MPI_IN_PLACE,&u,1,MPI_UINT32_T,MPI_BXOR,communicator);
+  GRID_ASSERT(ierr==0);
+}
+void CartesianCommunicator::GlobalXOR(uint64_t &u){
+  FlightRecorder::StepLog("GlobalXOR");
+  int ierr=MPI_Allreduce(MPI_IN_PLACE,&u,1,MPI_UINT64_T,MPI_BXOR,communicator);
+  GRID_ASSERT(ierr==0);
+}
+void CartesianCommunicator::GlobalMax(float &f)
+{
+  FlightRecorder::StepLog("GlobalMax");
+  int ierr=MPI_Allreduce(MPI_IN_PLACE,&f,1,MPI_FLOAT,MPI_MAX,communicator);
+  GRID_ASSERT(ierr==0);
+}
+void CartesianCommunicator::GlobalMax(double &d)
+{
+  FlightRecorder::StepLog("GlobalMax");
+  int ierr = MPI_Allreduce(MPI_IN_PLACE,&d,1,MPI_DOUBLE,MPI_MAX,communicator);
+  GRID_ASSERT(ierr==0);
+}
+void CartesianCommunicator::GlobalSumVector(float *f,int N)
+{
+  FlightRecorder::StepLog("GlobalSumVector(float *)");
+  int ierr=MPI_Allreduce(MPI_IN_PLACE,f,N,MPI_FLOAT,MPI_SUM,communicator);
+  GRID_ASSERT(ierr==0);
 }
 void CartesianCommunicator::GlobalSumVector(double *d,int N)
 {
+  FlightRecorder::StepLog("GlobalSumVector(double *)");
   int ierr = MPI_Allreduce(MPI_IN_PLACE,d,N,MPI_DOUBLE,MPI_SUM,communicator);
-  assert(ierr==0);
+  GRID_ASSERT(ierr==0);
 }
 
-void CartesianCommunicator::SendToRecvFromBegin(std::vector<CommsRequest_t> &list,
+void CartesianCommunicator::SendToRecvFromBegin(std::vector<MpiCommsRequest_t> &list,
 						void *xmit,
 						int dest,
 						void *recv,
 						int from,
-						int bytes,int dir)
+						uint64_t bytes,int dir)
 {
   MPI_Request xrq;
   MPI_Request rrq;
 
-  assert(dest != _processor);
+  GRID_ASSERT(dest != _processor);
-  assert(from != _processor);
+  GRID_ASSERT(from != _processor);
-
   int tag;
 
   tag= dir+from*32;
-  int ierr=MPI_Irecv(recv, bytes, MPI_CHAR,from,tag,communicator,&rrq);
+  int ierr=MPI_Irecv(recv,(int)( bytes/sizeof(int32_t)), MPI_INT32_T,from,tag,communicator,&rrq);
-  assert(ierr==0);
+  GRID_ASSERT(ierr==0);
   list.push_back(rrq);
   
   tag= dir+_processor*32;
-  ierr =MPI_Isend(xmit, bytes, MPI_CHAR,dest,tag,communicator,&xrq);
+  ierr =MPI_Isend(xmit,(int)(bytes/sizeof(int32_t)), MPI_INT32_T,dest,tag,communicator,&xrq);
-  assert(ierr==0);
+  GRID_ASSERT(ierr==0);
   list.push_back(xrq);
 }
-void CartesianCommunicator::CommsComplete(std::vector<CommsRequest_t> &list)
+void CartesianCommunicator::CommsComplete(std::vector<MpiCommsRequest_t> &list)
 {
   int nreq=list.size();
 
@@ -340,7 +369,7 @@ void CartesianCommunicator::CommsComplete(std::vector<CommsRequest_t> &list)
 
   std::vector<MPI_Status> status(nreq);
   int ierr = MPI_Waitall(nreq,&list[0],&status[0]);
-  assert(ierr==0);
+  GRID_ASSERT(ierr==0);
   list.resize(0);
 }
 
@@ -349,50 +378,63 @@ void CartesianCommunicator::SendToRecvFrom(void *xmit,
 					   int dest,
 					   void *recv,
 					   int from,
-					   int bytes)
+					   uint64_t bytes)
 {
-  std::vector<CommsRequest_t> reqs(0);
+  std::vector<MpiCommsRequest_t> reqs(0);
-  unsigned long  xcrc = crc32(0L, Z_NULL, 0);
-  unsigned long  rcrc = crc32(0L, Z_NULL, 0);
 
   int myrank = _processor;
   int ierr;
 
   // Enforce no UVM in comms, device or host OK
-  assert(acceleratorIsCommunicable(xmit));
+  GRID_ASSERT(acceleratorIsCommunicable(xmit));
-  assert(acceleratorIsCommunicable(recv));
+  GRID_ASSERT(acceleratorIsCommunicable(recv));
 
   // Give the CPU to MPI immediately; can use threads to overlap optionally
   //  printf("proc %d SendToRecvFrom %d bytes Sendrecv \n",_processor,bytes);
-  ierr=MPI_Sendrecv(xmit,bytes,MPI_CHAR,dest,myrank,
+  ierr=MPI_Sendrecv(xmit,(int)(bytes/sizeof(int32_t)),MPI_INT32_T,dest,myrank,
-		    recv,bytes,MPI_CHAR,from, from,
+		    recv,(int)(bytes/sizeof(int32_t)),MPI_INT32_T,from, from,
 		    communicator,MPI_STATUS_IGNORE);
-  assert(ierr==0);
+  GRID_ASSERT(ierr==0);
 
-  //  xcrc = crc32(xcrc,(unsigned char *)xmit,bytes);
-  //  rcrc = crc32(rcrc,(unsigned char *)recv,bytes);
-  //  printf("proc %d SendToRecvFrom %d bytes xcrc %lx rcrc %lx\n",_processor,bytes,xcrc,rcrc); fflush
 }
 // Basic Halo comms primitive
 double CartesianCommunicator::StencilSendToRecvFrom( void *xmit,
 						     int dest, int dox,
 						     void *recv,
 						     int from, int dor,
-						     int bytes,int dir)
+						     uint64_t bytes,int dir)
 {
   std::vector<CommsRequest_t> list;
-  double offbytes = StencilSendToRecvFromBegin(list,xmit,dest,dox,recv,from,dor,bytes,bytes,dir);
+  double offbytes = StencilSendToRecvFromPrepare(list,xmit,dest,dox,recv,from,dor,bytes,bytes,dir);
+  offbytes       += StencilSendToRecvFromBegin(list,xmit,xmit,dest,dox,recv,recv,from,dor,bytes,bytes,dir);
   StencilSendToRecvFromComplete(list,dir);
   return offbytes;
 }
+int CartesianCommunicator::IsOffNode(int rank)
+{
+  int grank = ShmRanks[rank];
+  if ( grank == MPI_UNDEFINED ) return true;
+  else return false;
+}
 
-#undef NVLINK_GET // Define to use get instead of put DMA
+#ifdef ACCELERATOR_AWARE_MPI
-double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list,
+void CartesianCommunicator::StencilSendToRecvFromPollIRecv(std::vector<CommsRequest_t> &list) {};
+void CartesianCommunicator::StencilSendToRecvFromPollDtoH(std::vector<CommsRequest_t> &list) {};
+double CartesianCommunicator::StencilSendToRecvFromPrepare(std::vector<CommsRequest_t> &list,
 							   void *xmit,
 							   int dest,int dox,
 							   void *recv,
 							   int from,int dor,
-							 int xbytes,int rbytes,int dir)
+							   uint64_t xbytes,uint64_t rbytes,int dir)
+{
+  return 0.0; // Do nothing -- no preparation required
+}
+double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list,
+							 void *xmit,void *xmit_comp,
+							 int dest,int dox,
+							 void *recv,void *recv_comp,
+							 int from,int dor,
+							 uint64_t xbytes,uint64_t rbytes,int dir)
 {
   int ncomm  =communicator_halo.size();
   int commdir=dir%ncomm;
@@ -405,62 +447,431 @@ double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsReques
   int gfrom = ShmRanks[from];
   int gme   = ShmRanks[_processor];
 
-  assert(dest != _processor);
+  GRID_ASSERT(dest != _processor);
-  assert(from != _processor);
+  GRID_ASSERT(from != _processor);
-  assert(gme  == ShmRank);
+  GRID_ASSERT(gme  == ShmRank);
   double off_node_bytes=0.0;
   int tag;
   
   if ( dor ) {
     if ( (gfrom ==MPI_UNDEFINED) || Stencil_force_mpi ) {
       tag= dir+from*32;
-      ierr=MPI_Irecv(recv, rbytes, MPI_CHAR,from,tag,communicator_halo[commdir],&rrq);
+      //      std::cout << " StencilSendToRecvFrom "<<dir<<" MPI_Irecv "<<std::hex<<recv<<std::dec<<std::endl;
-      assert(ierr==0);
+      ierr=MPI_Irecv(recv_comp,(int)(rbytes/sizeof(int32_t)), MPI_INT32_T,from,tag,communicator_halo[commdir],&rrq);
+      GRID_ASSERT(ierr==0);
       list.push_back(rrq);
       off_node_bytes+=rbytes;
     }
 #ifdef NVLINK_GET
+    else { 
       void *shm = (void *) this->ShmBufferTranslate(from,xmit);
-      assert(shm!=NULL);
+      GRID_ASSERT(shm!=NULL);
+      //      std::cout << " StencilSendToRecvFrom "<<dir<<" CopyDeviceToDevice recv "<<std::hex<<recv<<" remote "<<shm <<std::dec<<std::endl;
       acceleratorCopyDeviceToDeviceAsynch(shm,recv,rbytes);
+    }
 #endif
   }
-  
+  // This is a NVLINK PUT  
   if (dox) {
-    //  rcrc = crc32(rcrc,(unsigned char *)recv,bytes);
     if ( (gdest == MPI_UNDEFINED) || Stencil_force_mpi ) {
       tag= dir+_processor*32;
-      ierr =MPI_Isend(xmit, xbytes, MPI_CHAR,dest,tag,communicator_halo[commdir],&xrq);
+      ierr =MPI_Isend(xmit_comp,(int)(xbytes/sizeof(int32_t)), MPI_INT32_T,dest,tag,communicator_halo[commdir],&xrq);
-      assert(ierr==0);
+      GRID_ASSERT(ierr==0);
       list.push_back(xrq);
       off_node_bytes+=xbytes;
     } else {
 #ifndef NVLINK_GET
       void *shm = (void *) this->ShmBufferTranslate(dest,recv);
-      assert(shm!=NULL);
+      GRID_ASSERT(shm!=NULL);
       acceleratorCopyDeviceToDeviceAsynch(xmit,shm,xbytes);
 #endif
+    }
+  }
+  return off_node_bytes;
+}
 
+void CartesianCommunicator::StencilSendToRecvFromComplete(std::vector<CommsRequest_t> &list,int dir)
+{
+  int nreq=list.size();
+  /*finishes Get/Put*/
+  acceleratorCopySynchronise();
+
+  if (nreq==0) return;
+  std::vector<MPI_Status> status(nreq);
+  int ierr = MPI_Waitall(nreq,&list[0],&status[0]);
+  GRID_ASSERT(ierr==0);
+  list.resize(0);
+  this->StencilBarrier(); 
+}
+
+#else /* NOT     ... ACCELERATOR_AWARE_MPI */
+///////////////////////////////////////////
+// Pipeline mode through host memory
+///////////////////////////////////////////
+  /*
+   * In prepare (phase 1):
+   * PHASE 1: (prepare)
+   * - post MPI receive buffers asynch
+   * - post device - host send buffer transfer asynch
+   * PHASE 2: (Begin)
+   * - complete all copies
+   * - post MPI send asynch
+   * - post device - device transfers
+   * PHASE 3: (Complete)
+   * - MPI_waitall
+   * - host-device transfers
+   *
+   *********************************
+   * NB could split this further:
+   *--------------------------------
+   * PHASE 1: (Prepare)
+   * - post MPI receive buffers asynch
+   * - post device - host send buffer transfer asynch
+   * PHASE 2: (BeginInterNode)
+   * - complete all copies 
+   * - post MPI send asynch
+   * PHASE 3: (BeginIntraNode)
+   * - post device - device transfers
+   * PHASE 4: (Complete)
+   * - MPI_waitall
+   * - host-device transfers asynch
+   * - (complete all copies) 
+   */
+double CartesianCommunicator::StencilSendToRecvFromPrepare(std::vector<CommsRequest_t> &list,
+							   void *xmit,
+							   int dest,int dox,
+							   void *recv,
+							   int from,int dor,
+							   uint64_t xbytes,uint64_t rbytes,int dir)
+{
+/*
+ * Bring sequence from Stencil.h down to lower level.
+ * Assume using XeLink is ok
+ */  
+  int ncomm  =communicator_halo.size();
+  int commdir=dir%ncomm;
+
+  MPI_Request xrq;
+  MPI_Request rrq;
+
+  int ierr;
+  int gdest = ShmRanks[dest];
+  int gfrom = ShmRanks[from];
+  int gme   = ShmRanks[_processor];
+
+  GRID_ASSERT(dest != _processor);
+  GRID_ASSERT(from != _processor);
+  GRID_ASSERT(gme  == ShmRank);
+  double off_node_bytes=0.0;
+  int tag;
+
+  void * host_recv = NULL;
+  void * host_xmit = NULL;
+
+  /*
+   * PHASE 1: (Prepare)
+   * - post MPI receive buffers asynch
+   * - post device - host send buffer transfer asynch
+   */
+  
+#ifdef GRID_CHECKSUM_COMMS
+  rbytes += 8;
+  xbytes += 8;
+#endif
+
+  if ( dor ) {
+    if ( (gfrom ==MPI_UNDEFINED) || Stencil_force_mpi ) {
+      tag= dir+from*32;
+      host_recv = this->HostBufferMalloc(rbytes);
+      ierr=MPI_Irecv(host_recv,(int)(rbytes/sizeof(int32_t)), MPI_INT32_T,from,tag,communicator_halo[commdir],&rrq);
+      GRID_ASSERT(ierr==0);
+      CommsRequest_t srq;
+      srq.PacketType = InterNodeRecv;
+      srq.bytes      = rbytes;
+      srq.req        = rrq;
+      srq.host_buf   = host_recv;
+      srq.device_buf = recv;
+      srq.tag        = tag;
+      list.push_back(srq);
+      off_node_bytes+=rbytes;
+    }
+  }
+  
+  if (dox) {
+    if ( (gdest == MPI_UNDEFINED) || Stencil_force_mpi ) {
+
+      tag= dir+_processor*32;
+
+      host_xmit = this->HostBufferMalloc(xbytes);
+      CommsRequest_t srq;
+
+#ifdef GRID_CHECKSUM_COMMS
+      uint64_t xbytes_data = xbytes - 8;
+      srq.ev = acceleratorCopyFromDeviceAsynch(xmit, host_xmit,xbytes_data); // Make this Asynch
+      GRID_ASSERT(xbytes % 8 == 0);
+      // flip one bit so that a zero buffer is not consistent
+      uint64_t xsum = checksum_gpu((uint64_t*)xmit, xbytes_data / 8) ^ (checksum_index + 1 + 1000 * tag); 
+      *(uint64_t*)(((char*)host_xmit) + xbytes_data) = xsum;
+#else
+      srq.ev = acceleratorCopyFromDeviceAsynch(xmit, host_xmit,xbytes); // Make this Asynch
+#endif
+      
+      //      ierr =MPI_Isend(host_xmit, xbytes, MPI_CHAR,dest,tag,communicator_halo[commdir],&xrq);
+      //      GRID_ASSERT(ierr==0);
+      //      off_node_bytes+=xbytes;
+
+      srq.PacketType = InterNodeXmit;
+      srq.bytes      = xbytes;
+      //      srq.req        = xrq;
+      srq.host_buf   = host_xmit;
+      srq.device_buf = xmit;
+      srq.tag        = tag;
+      srq.dest       = dest;
+      srq.commdir    = commdir;
+      list.push_back(srq);
     }
   }
 
   return off_node_bytes;
 }
+/*
+ * In the interest of better pipelining, poll for completion on each DtoH and 
+ * start MPI_ISend in the meantime
+ */
+void CartesianCommunicator::StencilSendToRecvFromPollIRecv(std::vector<CommsRequest_t> &list)
+{
+  int pending = 0;
+  do {
+
+    pending = 0;
+
+    for(int idx = 0; idx<list.size();idx++){
+
+      if ( list[idx].PacketType==InterNodeRecv ) {
+
+	int flag = 0;
+	MPI_Status status;
+	int ierr = MPI_Test(&list[idx].req,&flag,&status);
+	assert(ierr==0);
+
+	if ( flag ) {
+	  //	  std::cout << " PollIrecv "<<idx<<" flag "<<flag<<std::endl;
+#ifdef GRID_CHECKSUM_COMMS
+ 	  acceleratorCopyToDeviceAsynch(list[idx].host_buf,list[idx].device_buf,list[idx].bytes - 8);
+#else
+	  acceleratorCopyToDeviceAsynch(list[idx].host_buf,list[idx].device_buf,list[idx].bytes);
+#endif
+	  list[idx].PacketType=InterNodeReceiveHtoD;
+	} else {
+	  pending ++;
+	}
+      }
+    }
+    //    std::cout << " PollIrecv "<<pending<<" pending requests"<<std::endl;
+  } while ( pending );
+  
+}
+void CartesianCommunicator::StencilSendToRecvFromPollDtoH(std::vector<CommsRequest_t> &list)
+{
+  int pending = 0;
+  do {
+
+    pending = 0;
+
+    for(int idx = 0; idx<list.size();idx++){
+
+      if ( list[idx].PacketType==InterNodeXmit ) {
+
+	if ( acceleratorEventIsComplete(list[idx].ev) ) {
+
+	  void *host_xmit = list[idx].host_buf;
+	  uint64_t xbytes = list[idx].bytes;
+	  int dest        = list[idx].dest;
+	  int tag         = list[idx].tag;
+	  int commdir     = list[idx].commdir;
+	  ///////////////////
+	  // Send packet
+	  ///////////////////
+
+	  //	  std::cout << " DtoH is complete for index "<<idx<<" calling MPI_Isend "<<std::endl;
+	  
+	  MPI_Request xrq;
+	  int ierr =MPI_Isend(host_xmit, (int)(xbytes/sizeof(int32_t)), MPI_INT32_T,dest,tag,communicator_halo[commdir],&xrq);
+	  GRID_ASSERT(ierr==0);
+
+	  list[idx].req        = xrq; // Update the MPI request in the list
+
+	  list[idx].PacketType=InterNodeXmitISend;
+
+	} else {
+	  // not done, so return to polling loop
+	  pending++;
+	}
+      }
+    }
+  } while (pending);
+}  
+
+double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list,
+							 void *xmit,void *xmit_comp,
+							 int dest,int dox,
+							 void *recv,void *recv_comp,
+							 int from,int dor,
+							 uint64_t xbytes,uint64_t rbytes,int dir)
+{
+  int ncomm  =communicator_halo.size();
+  int commdir=dir%ncomm;
+
+  MPI_Request xrq;
+  MPI_Request rrq;
+
+  int ierr;
+  int gdest = ShmRanks[dest];
+  int gfrom = ShmRanks[from];
+  int gme   = ShmRanks[_processor];
+
+  GRID_ASSERT(dest != _processor);
+  GRID_ASSERT(from != _processor);
+  GRID_ASSERT(gme  == ShmRank);
+  double off_node_bytes=0.0;
+  int tag;
+
+  void * host_xmit = NULL;
+
+  ////////////////////////////////
+  // Receives already posted
+  // Copies already started
+  ////////////////////////////////
+  /*  
+   * PHASE 2: (Begin)
+   * - complete all copies
+   * - post MPI send asynch
+   */
+#ifdef NVLINK_GET
+  if ( dor ) {
+
+    if ( ! ( (gfrom ==MPI_UNDEFINED) || Stencil_force_mpi ) ) {
+      // Intranode
+      void *shm = (void *) this->ShmBufferTranslate(from,xmit);
+      GRID_ASSERT(shm!=NULL);
+
+      CommsRequest_t srq;
+
+      srq.ev = acceleratorCopyDeviceToDeviceAsynch(shm,recv,rbytes);
+
+      srq.PacketType = IntraNodeRecv;
+      srq.bytes      = xbytes;
+      //      srq.req        = xrq;
+      srq.host_buf   = NULL;
+      srq.device_buf = xmit;
+      srq.tag        = -1;
+      srq.dest       = dest;
+      srq.commdir    = dir;
+      list.push_back(srq);
+    }
+  }  
+#else
+  if (dox) {
+
+    if ( !( (gdest == MPI_UNDEFINED) || Stencil_force_mpi ) ) {
+      // Intranode
+      void *shm = (void *) this->ShmBufferTranslate(dest,recv);
+      GRID_ASSERT(shm!=NULL);
+
+      CommsRequest_t srq;
+      
+      srq.ev = acceleratorCopyDeviceToDeviceAsynch(xmit,shm,xbytes);
+
+      srq.PacketType = IntraNodeXmit;
+      srq.bytes      = xbytes;
+      //      srq.req        = xrq;
+      srq.host_buf   = NULL;
+      srq.device_buf = xmit;
+      srq.tag        = -1;
+      srq.dest       = dest;
+      srq.commdir    = dir;
+      list.push_back(srq);
+      
+    }
+  }
+#endif
+  return off_node_bytes;
+}
 void CartesianCommunicator::StencilSendToRecvFromComplete(std::vector<CommsRequest_t> &list,int dir)
 {
-  int nreq=list.size();
+  acceleratorCopySynchronise(); // Complete all pending copy transfers D2D
 
-  acceleratorCopySynchronise();
+  std::vector<MPI_Status> status;
+  std::vector<MPI_Request> MpiRequests;
     
-  if (nreq==0) return;
+  for(int r=0;r<list.size();r++){
-
+    // Must check each Send buf is clear to reuse
-  std::vector<MPI_Status> status(nreq);
+    if ( list[r].PacketType == InterNodeXmitISend ) MpiRequests.push_back(list[r].req);
-  int ierr = MPI_Waitall(nreq,&list[0],&status[0]);
+    //    if ( list[r].PacketType == InterNodeRecv ) MpiRequests.push_back(list[r].req); // Already "Test" passed
-  assert(ierr==0);
-  list.resize(0);
   }
+
+  int nreq=MpiRequests.size();
+
+  if (nreq>0) {
+    status.resize(MpiRequests.size());
+    int ierr = MPI_Waitall(MpiRequests.size(),&MpiRequests[0],&status[0]); // Sends are guaranteed in order. No harm in not completing.
+    GRID_ASSERT(ierr==0);
+  }
+  
+  //  for(int r=0;r<nreq;r++){
+  //    if ( list[r].PacketType==InterNodeRecv ) {
+  //      acceleratorCopyToDeviceAsynch(list[r].host_buf,list[r].device_buf,list[r].bytes);
+  //    }
+  //  }
+#ifdef GRID_CHECKSUM_COMMS
+  for(int r=0;r<list.size();r++){
+    if ( list[r].PacketType == InterNodeReceiveHtoD ) {
+      uint64_t rbytes_data = list[r].bytes - 8;
+      uint64_t expected_cs = *(uint64_t*)(((char*)list[r].host_buf) + rbytes_data);
+      uint64_t computed_cs = checksum_gpu((uint64_t*)list[r].device_buf, rbytes_data / 8) ^ (checksum_index + 1 + 1000 * list[r].tag); //
+      if (expected_cs != computed_cs) {
+	// TODO: error message, backtrace, quit
+
+	fprintf(stderr, "GRID_CHECKSUM_COMMS error:\n");
+	fprintf(stderr, " processor = %d\n", (int)_processor);
+	for(int d=0;d<_processors.size();d++)
+	  fprintf(stderr, " processor_coord[%d] = %d\n", d, _processor_coor[d]);
+	fprintf(stderr, " hostname: %s\n", GridHostname());
+	fprintf(stderr, " expected_cs: %ld\n", expected_cs);
+	fprintf(stderr, " computed_cs: %ld\n", computed_cs);
+	fprintf(stderr, " dest: %d\n", list[r].dest);
+	fprintf(stderr, " tag: %d\n", list[r].tag);
+	fprintf(stderr, " commdir: %d\n", list[r].commdir);
+	fprintf(stderr, " bytes: %ld\n", (uint64_t)list[r].bytes);
+
+	fflush(stderr);
+
+	// backtrace
+	int symbols = backtrace(Grid_backtrace_buffer,_NBACKTRACE);
+	backtrace_symbols_fd(Grid_backtrace_buffer,symbols, 2);
+
+	exit(1);
+      }
+    }
+  }
+
+  checksum_index += 1;
+#endif
+  
+  list.resize(0);               // Delete the list
+  this->HostBufferFreeAll();    // Clean up the buffer allocs
+#ifndef NVLINK_GET
+  this->StencilBarrier(); // if PUT must check our nbrs have filled our receive buffers.
+#endif   
+}
+#endif
+////////////////////////////////////////////
+// END PIPELINE MODE / NO CUDA AWARE MPI
+////////////////////////////////////////////
+
 void CartesianCommunicator::StencilBarrier(void)
 {
+  FlightRecorder::StepLog("NodeBarrier");
   MPI_Barrier  (ShmComm);
 }
 //void CartesianCommunicator::SendToRecvFromComplete(std::vector<CommsRequest_t> &list)
@@ -468,17 +879,19 @@ void CartesianCommunicator::StencilBarrier(void)
 //}
 void CartesianCommunicator::Barrier(void)
 {
+  FlightRecorder::StepLog("GridBarrier");
   int ierr = MPI_Barrier(communicator);
-  assert(ierr==0);
+  GRID_ASSERT(ierr==0);
 }
-void CartesianCommunicator::Broadcast(int root,void* data, int bytes)
+void CartesianCommunicator::Broadcast(int root,void* data,uint64_t bytes)
 {
+  FlightRecorder::StepLog("Broadcast");
   int ierr=MPI_Bcast(data,
-		     bytes,
+		     (int)bytes,
 		     MPI_BYTE,
 		     root,
 		     communicator);
-  assert(ierr==0);
+  GRID_ASSERT(ierr==0);
 }
 int CartesianCommunicator::RankWorld(void){
   int r;
@@ -486,23 +899,25 @@ int CartesianCommunicator::RankWorld(void){
   return r;
 }
 void CartesianCommunicator::BarrierWorld(void){
+  FlightRecorder::StepLog("BarrierWorld");
   int ierr = MPI_Barrier(communicator_world);
-  assert(ierr==0);
+  GRID_ASSERT(ierr==0);
 }
-void CartesianCommunicator::BroadcastWorld(int root,void* data, int bytes)
+void CartesianCommunicator::BroadcastWorld(int root,void* data, uint64_t bytes)
 {
+  FlightRecorder::StepLog("BroadcastWorld");
   int ierr= MPI_Bcast(data,
-		      bytes,
+		      (int)bytes,
 		      MPI_BYTE,
 		      root,
 		      communicator_world);
-  assert(ierr==0);
+  GRID_ASSERT(ierr==0);
 }
 
 void CartesianCommunicator::AllToAll(int dim,void  *in,void *out,uint64_t words,uint64_t bytes)
 {
   Coordinate row(_ndimension,1);
-  assert(dim>=0 && dim<_ndimension);
+  GRID_ASSERT(dim>=0 && dim<_ndimension);
 
   //  Split the communicator
   row[dim] = _processors[dim];
@@ -513,6 +928,7 @@ void CartesianCommunicator::AllToAll(int dim,void  *in,void *out,uint64_t words,
 }
 void CartesianCommunicator::AllToAll(void  *in,void *out,uint64_t words,uint64_t bytes)
 {
+  FlightRecorder::StepLog("AllToAll");
   // MPI is a pain and uses "int" arguments
   // 64*64*64*128*16 == 500Million elements of data.
   // When 24*4 bytes multiples get 50x 10^9 >>> 2x10^9 Y2K bug.
@@ -522,8 +938,8 @@ void CartesianCommunicator::AllToAll(void  *in,void *out,uint64_t words,uint64_t
   int ibytes;
   iwords = words;
   ibytes = bytes;
-  assert(words == iwords); // safe to cast to int ?
+  GRID_ASSERT(words == iwords); // safe to cast to int ?
-  assert(bytes == ibytes); // safe to cast to int ?
+  GRID_ASSERT(bytes == ibytes); // safe to cast to int ?
   MPI_Type_contiguous(ibytes,MPI_BYTE,&object);
   MPI_Type_commit(&object);
   MPI_Alltoall(in,iwords,object,out,iwords,object,communicator);

Grid/communicator/Communicator_none.cc

@@ -27,6 +27,8 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 /*  END LEGAL */
 #include <Grid/GridCore.h>
 
+void GridAbort(void) { abort(); }
+
 NAMESPACE_BEGIN(Grid);
 
 ///////////////////////////////////////////////////////////////////////////////////////////////////
@@ -34,6 +36,7 @@ NAMESPACE_BEGIN(Grid);
 ///////////////////////////////////////////////////////////////////////////////////////////////////
 Grid_MPI_Comm       CartesianCommunicator::communicator_world;
 
+
 void CartesianCommunicator::Init(int *argc, char *** arv)
 {
   GlobalSharedMemory::Init(communicator_world);
@@ -54,14 +57,14 @@ CartesianCommunicator::CartesianCommunicator(const Coordinate &processors)
 {
   _shm_processors = Coordinate(processors.size(),1);
   _processors = processors;
-  _ndimension = processors.size();  assert(_ndimension>=1);
+  _ndimension = processors.size();  GRID_ASSERT(_ndimension>=1);
   _processor_coor.resize(_ndimension);
   
   // Require 1^N processor grid for fake
   _Nprocessors=1;
   _processor = 0;
   for(int d=0;d<_ndimension;d++) {
-    assert(_processors[d]==1);
+    GRID_ASSERT(_processors[d]==1);
     _processor_coor[d] = 0;
   }
   SetCommunicator(communicator_world);
@@ -87,19 +90,19 @@ void CartesianCommunicator::SendToRecvFrom(void *xmit,
 					   int dest,
 					   void *recv,
 					   int from,
-					   int bytes)
+					   uint64_t bytes)
 {
-  assert(0);
+  GRID_ASSERT(0);
 }
-void CartesianCommunicator::CommsComplete(std::vector<CommsRequest_t> &list){ assert(0);}
+void CartesianCommunicator::CommsComplete(std::vector<CommsRequest_t> &list){ GRID_ASSERT(list.size()==0);}
 void CartesianCommunicator::SendToRecvFromBegin(std::vector<CommsRequest_t> &list,
 						void *xmit,
 						int dest,
 						void *recv,
 						int from,
-						int bytes,int dir)
+						uint64_t bytes,int dir)
 {
-  assert(0);
+  GRID_ASSERT(0);
 }
 
 void CartesianCommunicator::AllToAll(int dim,void  *in,void *out,uint64_t words,uint64_t bytes)
@@ -113,8 +116,8 @@ void CartesianCommunicator::AllToAll(void  *in,void *out,uint64_t words,uint64_t
 
 int  CartesianCommunicator::RankWorld(void){return 0;}
 void CartesianCommunicator::Barrier(void){}
-void CartesianCommunicator::Broadcast(int root,void* data, int bytes) {}
+void CartesianCommunicator::Broadcast(int root,void* data, uint64_t bytes) {}
-void CartesianCommunicator::BroadcastWorld(int root,void* data, int bytes) { }
+void CartesianCommunicator::BroadcastWorld(int root,void* data, uint64_t bytes) { }
 void CartesianCommunicator::BarrierWorld(void) { }
 int  CartesianCommunicator::RankFromProcessorCoor(Coordinate &coor) {  return 0;}
 void CartesianCommunicator::ProcessorCoorFromRank(int rank, Coordinate &coor){  coor = _processor_coor; }
@@ -124,20 +127,33 @@ void CartesianCommunicator::ShiftedRanks(int dim,int shift,int &source,int &dest
   dest=0;
 }
 
+int CartesianCommunicator::IsOffNode(int rank) { return false; }
+
 double CartesianCommunicator::StencilSendToRecvFrom( void *xmit,
 						     int xmit_to_rank,int dox,
 						     void *recv,
 						     int recv_from_rank,int dor,
-						     int bytes, int dir)
+						     uint64_t bytes, int dir)
 {
   return 2.0*bytes;
 }
-double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list,
+void CartesianCommunicator::StencilSendToRecvFromPollIRecv(std::vector<CommsRequest_t> &list) {};
+void CartesianCommunicator::StencilSendToRecvFromPollDtoH(std::vector<CommsRequest_t> &list) {};
+double CartesianCommunicator::StencilSendToRecvFromPrepare(std::vector<CommsRequest_t> &list,
 							   void *xmit,
 							   int xmit_to_rank,int dox,
 							   void *recv,
 							   int recv_from_rank,int dor,
-							 int xbytes,int rbytes, int dir)
+							   uint64_t xbytes,uint64_t rbytes, int dir)
+{
+  return 0.0;
+}
+double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list,
+							 void *xmit, void *xmit_comp,
+							 int xmit_to_rank,int dox,
+							 void *recv, void *recv_comp,
+							 int recv_from_rank,int dor,
+							 uint64_t xbytes,uint64_t rbytes, int dir)
 {
   return xbytes+rbytes;
 }

Grid/communicator/SharedMemory.cc

@@ -58,8 +58,8 @@ int                 GlobalSharedMemory::WorldNode;
 
 void GlobalSharedMemory::SharedMemoryFree(void)
 {
-  assert(_ShmAlloc);
+  GRID_ASSERT(_ShmAlloc);
-  assert(_ShmAllocBytes>0);
+  GRID_ASSERT(_ShmAllocBytes>0);
   for(int r=0;r<WorldShmSize;r++){
     munmap(WorldShmCommBufs[r],_ShmAllocBytes);
   }
@@ -80,7 +80,7 @@ void *SharedMemory::HostBufferMalloc(size_t bytes){
     std::cout<< " Current alloc is " << (bytes/(1024*1024)) <<"MB"<<std::endl;
     std::cout<< " Current bytes is " << (host_heap_bytes/(1024*1024)) <<"MB"<<std::endl;
     std::cout<< " Current heap  is " << (host_heap_size/(1024*1024)) <<"MB"<<std::endl;
-    assert(host_heap_bytes<host_heap_size);
+    GRID_ASSERT(host_heap_bytes<host_heap_size);
   }
   return ptr;
 }
@@ -100,7 +100,7 @@ void *SharedMemory::ShmBufferMalloc(size_t bytes){
     std::cout<< " Current alloc is " << (bytes/(1024*1024)) <<"MB"<<std::endl;
     std::cout<< " Current bytes is " << (heap_bytes/(1024*1024)) <<"MB"<<std::endl;
     std::cout<< " Current heap  is " << (heap_size/(1024*1024)) <<"MB"<<std::endl;
-    assert(heap_bytes<heap_size);
+    GRID_ASSERT(heap_bytes<heap_size);
   }
   //std::cerr << "ShmBufferMalloc "<<std::hex<< ptr<<" - "<<((uint64_t)ptr+bytes)<<std::dec<<std::endl;
   return ptr;
@@ -127,13 +127,13 @@ void GlobalSharedMemory::GetShmDims(const Coordinate &WorldDims,Coordinate &ShmD
   if ( str ) {
     std::vector<int> IntShmDims;
     GridCmdOptionIntVector(std::string(str),IntShmDims);
-    assert(IntShmDims.size() == WorldDims.size());
+    GRID_ASSERT(IntShmDims.size() == WorldDims.size());
     long ShmSize = 1;
     for (int dim=0;dim<WorldDims.size();dim++) {
       ShmSize *= (ShmDims[dim] = IntShmDims[dim]);
-      assert(divides(ShmDims[dim],WorldDims[dim]));
+      GRID_ASSERT(divides(ShmDims[dim],WorldDims[dim]));
     }
-    assert(ShmSize == WorldShmSize);
+    GRID_ASSERT(ShmSize == WorldShmSize);
     return;
   }
   

Grid/communicator/SharedMemory.h

@@ -46,8 +46,40 @@ NAMESPACE_BEGIN(Grid);
 
 #if defined (GRID_COMMS_MPI3) 
 typedef MPI_Comm    Grid_MPI_Comm;
+typedef MPI_Request MpiCommsRequest_t;
+#ifdef ACCELERATOR_AWARE_MPI
 typedef MPI_Request CommsRequest_t;
 #else
+/*
+ * Enable state transitions as each packet flows.
+ */
+enum PacketType_t {
+  FaceGather,
+  InterNodeXmit,
+  InterNodeRecv,
+  IntraNodeXmit,
+  IntraNodeRecv,
+  InterNodeXmitISend,
+  InterNodeReceiveHtoD
+};
+/*
+ *Package arguments needed for various actions along packet flow
+ */
+typedef struct {
+  PacketType_t PacketType;
+  void *host_buf;
+  void *device_buf;
+  int dest;
+  int tag;
+  int commdir;
+  unsigned long bytes;
+  acceleratorEvent_t ev;
+  MpiCommsRequest_t req;
+} CommsRequest_t;
+#endif
+
+#else 
+typedef int MpiCommsRequest_t;
 typedef int CommsRequest_t;
 typedef int Grid_MPI_Comm;
 #endif
@@ -105,7 +137,7 @@ public:
   ///////////////////////////////////////////////////
   static void SharedMemoryAllocate(uint64_t bytes, int flags);
   static void SharedMemoryFree(void);
-  static void SharedMemoryCopy(void *dest,void *src,size_t bytes);
+  //  static void SharedMemoryCopy(void *dest,void *src,size_t bytes);
   static void SharedMemoryZero(void *dest,size_t bytes);
 
 };

Grid/communicator/SharedMemoryMPI.cc

@@ -42,6 +42,7 @@ Author: Christoph Lehner <christoph@lhnr.de>
 #ifdef ACCELERATOR_AWARE_MPI
 #define GRID_SYCL_LEVEL_ZERO_IPC
 #define SHM_SOCKETS
+#else
 #endif 
 #include <syscall.h>
 #endif
@@ -66,7 +67,7 @@ public:
   {
     int errnum;
 
-    sock = socket(AF_UNIX, SOCK_DGRAM, 0);  assert(sock>0);
+    sock = socket(AF_UNIX, SOCK_DGRAM, 0);  GRID_ASSERT(sock>0);
 
     struct sockaddr_un sa_un = { 0 };
     sa_un.sun_family = AF_UNIX;
@@ -157,7 +158,7 @@ public:
 /*Construct from an MPI communicator*/
 void GlobalSharedMemory::Init(Grid_MPI_Comm comm)
 {
-  assert(_ShmSetup==0);
+  GRID_ASSERT(_ShmSetup==0);
   WorldComm = comm;
   MPI_Comm_rank(WorldComm,&WorldRank);
   MPI_Comm_size(WorldComm,&WorldSize);
@@ -183,7 +184,7 @@ void GlobalSharedMemory::Init(Grid_MPI_Comm comm)
 
   // WorldNodes
   WorldNodes = WorldSize/WorldShmSize;
-  assert( (WorldNodes * WorldShmSize) == WorldSize );
+  GRID_ASSERT( (WorldNodes * WorldShmSize) == WorldSize );
 
 
   // FIXME: Check all WorldShmSize are the same ?
@@ -208,7 +209,7 @@ void GlobalSharedMemory::Init(Grid_MPI_Comm comm)
   MyGroup.resize(WorldShmSize);
   for(int rank=0;rank<WorldSize;rank++){
     if(WorldShmRanks[rank]!=MPI_UNDEFINED){
-      assert(g<WorldShmSize);
+      GRID_ASSERT(g<WorldShmSize);
       MyGroup[g++] = rank;
     }
   }
@@ -224,7 +225,7 @@ void GlobalSharedMemory::Init(Grid_MPI_Comm comm)
   // global sum leaders over comm world
   ///////////////////////////////////////////////////////////////////
   int ierr=MPI_Allreduce(MPI_IN_PLACE,&leaders_1hot[0],WorldSize,MPI_INT,MPI_SUM,WorldComm);
-  assert(ierr==0);
+  GRID_ASSERT(ierr==0);
 
   ///////////////////////////////////////////////////////////////////
   // find the group leaders world rank
@@ -245,7 +246,7 @@ void GlobalSharedMemory::Init(Grid_MPI_Comm comm)
       WorldNode=g;
     }
   }
-  assert(WorldNode!=-1);
+  GRID_ASSERT(WorldNode!=-1);
   _ShmSetup=1;
 }
 // Gray encode support 
@@ -287,7 +288,7 @@ void GlobalSharedMemory::OptimalCommunicatorHypercube(const Coordinate &processo
   // Assert power of two shm_size.
   ////////////////////////////////////////////////////////////////
   int log2size = Log2Size(WorldShmSize,MAXLOG2RANKSPERNODE);
-  assert(log2size != -1);
+  GRID_ASSERT(log2size != -1);
 
   ////////////////////////////////////////////////////////////////
   // Identify the hypercube coordinate of this node using hostname
@@ -308,7 +309,7 @@ void GlobalSharedMemory::OptimalCommunicatorHypercube(const Coordinate &processo
   // Parse ICE-XA hostname to get hypercube location
   gethostname(name,namelen);
   int nscan = sscanf(name,"r%di%dn%d",&R,&I,&N) ;
-  assert(nscan==3);
+  GRID_ASSERT(nscan==3);
 
   int nlo = N%9;
   int nhi = N/9;
@@ -332,8 +333,8 @@ void GlobalSharedMemory::OptimalCommunicatorHypercube(const Coordinate &processo
   //////////////////////////////////////////////////////////////////
   MPI_Bcast(&rootcoor, sizeof(rootcoor), MPI_BYTE, 0, WorldComm); 
   hypercoor=hypercoor-rootcoor;
-  assert(hypercoor<WorldSize);
+  GRID_ASSERT(hypercoor<WorldSize);
-  assert(hypercoor>=0);
+  GRID_ASSERT(hypercoor>=0);
 
   //////////////////////////////////////
   // Printing
@@ -381,7 +382,7 @@ void GlobalSharedMemory::OptimalCommunicatorHypercube(const Coordinate &processo
   for(int i=0;i<ndimension;i++){
     Nprocessors*=processors[i];
   }
-  assert(WorldSize==Nprocessors);
+  GRID_ASSERT(WorldSize==Nprocessors);
 
   ////////////////////////////////////////////////////////////////
   // Establish mapping between lexico physics coord and WorldRank
@@ -400,7 +401,7 @@ void GlobalSharedMemory::OptimalCommunicatorHypercube(const Coordinate &processo
   // Build the new communicator
   /////////////////////////////////////////////////////////////////
   int ierr= MPI_Comm_split(WorldComm,0,rank,&optimal_comm);
-  assert(ierr==0);
+  GRID_ASSERT(ierr==0);
 }
 void GlobalSharedMemory::OptimalCommunicatorSharedMemory(const Coordinate &processors,Grid_MPI_Comm & optimal_comm,Coordinate &SHM)
 {
@@ -430,7 +431,8 @@ void GlobalSharedMemory::OptimalCommunicatorSharedMemory(const Coordinate &proce
   for(int i=0;i<ndimension;i++){
     Nprocessors*=processors[i];
   }
-  assert(WorldSize==Nprocessors);
+  //  std::cerr << " WorldSize "<<WorldSize << " Nprocessors "<<Nprocessors<<" "<<processors<<std::endl; 
+  GRID_ASSERT(WorldSize==Nprocessors);
 
   ////////////////////////////////////////////////////////////////
   // Establish mapping between lexico physics coord and WorldRank
@@ -446,7 +448,7 @@ void GlobalSharedMemory::OptimalCommunicatorSharedMemory(const Coordinate &proce
   // Build the new communicator
   /////////////////////////////////////////////////////////////////
   int ierr= MPI_Comm_split(WorldComm,0,rank,&optimal_comm);
-  assert(ierr==0);
+  GRID_ASSERT(ierr==0);
 }
 ////////////////////////////////////////////////////////////////////////////////////////////
 // SHMGET
@@ -455,8 +457,8 @@ void GlobalSharedMemory::OptimalCommunicatorSharedMemory(const Coordinate &proce
 void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 {
   std::cout << Mheader "SharedMemoryAllocate "<< bytes<< " shmget implementation "<<std::endl;
-  assert(_ShmSetup==1);
+  GRID_ASSERT(_ShmSetup==1);
-  assert(_ShmAlloc==0);
+  GRID_ASSERT(_ShmAlloc==0);
 
   //////////////////////////////////////////////////////////////////////////////////////////////////////////
   // allocate the shared windows for our group
@@ -517,8 +519,8 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 {
   void * ShmCommBuf ; 
-  assert(_ShmSetup==1);
+  GRID_ASSERT(_ShmSetup==1);
-  assert(_ShmAlloc==0);
+  GRID_ASSERT(_ShmAlloc==0);
 
   //////////////////////////////////////////////////////////////////////////////////////////////////////////
   // allocate the pointer array for shared windows for our group
@@ -537,7 +539,8 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
   // Each MPI rank should allocate our own buffer
   ///////////////////////////////////////////////////////////////////////////////////////////////////////////
 #ifndef ACCELERATOR_AWARE_MPI
-  HostCommBuf= malloc(bytes);
+  // printf("Host buffer allocate for GPU non-aware MPI\n");
+  HostCommBuf= malloc(bytes); /// CHANGE THIS TO malloc_host
 #endif  
   ShmCommBuf = acceleratorAllocDevice(bytes);
   if (ShmCommBuf == (void *)NULL ) {
@@ -545,11 +548,13 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
     exit(EXIT_FAILURE);  
   }
   if ( WorldRank == 0 ){
-    std::cout << WorldRank << Mheader " SharedMemoryMPI.cc acceleratorAllocDevice "<< bytes 
+    std::cout << Mheader " acceleratorAllocDevice "<< bytes 
 	      << "bytes at "<< std::hex<< ShmCommBuf << " - "<<(bytes-1+(uint64_t)ShmCommBuf) <<std::dec<<" for comms buffers " <<std::endl;
   }
   SharedMemoryZero(ShmCommBuf,bytes);
-  std::cout<< "Setting up IPC"<<std::endl;
+  if ( WorldRank == 0 ){
+    std::cout<< Mheader "Setting up IPC"<<std::endl;
+  }
   ///////////////////////////////////////////////////////////////////////////////////////////////////////////
   // Loop over ranks/gpu's on our node
   ///////////////////////////////////////////////////////////////////////////////////////////////////////////
@@ -569,8 +574,8 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 #ifdef GRID_SYCL_LEVEL_ZERO_IPC
     typedef struct { int fd; pid_t pid ; ze_ipc_mem_handle_t ze; } clone_mem_t;
 
-    auto zeDevice    = cl::sycl::get_native<cl::sycl::backend::ext_oneapi_level_zero>(theGridAccelerator->get_device());
+    auto zeDevice    = sycl::get_native<sycl::backend::ext_oneapi_level_zero>(theGridAccelerator->get_device());
-    auto zeContext   = cl::sycl::get_native<cl::sycl::backend::ext_oneapi_level_zero>(theGridAccelerator->get_context());
+    auto zeContext   = sycl::get_native<sycl::backend::ext_oneapi_level_zero>(theGridAccelerator->get_context());
       
     ze_ipc_mem_handle_t ihandle;
     clone_mem_t handle;
@@ -580,8 +585,6 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
       if ( err != ZE_RESULT_SUCCESS ) {
 	std::cerr << "SharedMemoryMPI.cc zeMemGetIpcHandle failed for rank "<<r<<" "<<std::hex<<err<<std::dec<<std::endl;
 	exit(EXIT_FAILURE);
-      } else {
-	std::cout << "SharedMemoryMPI.cc zeMemGetIpcHandle succeeded for rank "<<r<<" "<<std::hex<<err<<std::dec<<std::endl;
       }
       memcpy((void *)&handle.fd,(void *)&ihandle,sizeof(int));
       handle.pid = getpid();
@@ -626,7 +629,7 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 			 MPI_BYTE,
 			 r,
 			 WorldShmComm);
-      assert(ierr==0);
+      GRID_ASSERT(ierr==0);
     }
     
     ///////////////////////////////////////////////////////////////
@@ -640,12 +643,12 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 #ifdef SHM_SOCKETS
       myfd=UnixSockets::RecvFileDescriptor();
 #else
-      std::cout<<"mapping seeking remote pid/fd "
+      //      std::cout<<"mapping seeking remote pid/fd "
-	       <<handle.pid<<"/"
+      //	       <<handle.pid<<"/"
-	       <<handle.fd<<std::endl;
+      //	       <<handle.fd<<std::endl;
 
       int pidfd = syscall(SYS_pidfd_open,handle.pid,0);
-      std::cout<<"Using IpcHandle pidfd "<<pidfd<<"\n";
+      //      std::cout<<"Using IpcHandle pidfd "<<pidfd<<"\n";
       //      int myfd  = syscall(SYS_pidfd_getfd,pidfd,handle.fd,0);
       myfd  = syscall(438,pidfd,handle.fd,0);
       int err_t = errno;
@@ -655,7 +658,7 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 	assert(0);
       }
 #endif
-      std::cout<<"Using IpcHandle mapped remote pid "<<handle.pid <<" FD "<<handle.fd <<" to myfd "<<myfd<<"\n";
+      //      std::cout<<"Using IpcHandle mapped remote pid "<<handle.pid <<" FD "<<handle.fd <<" to myfd "<<myfd<<"\n";
       memcpy((void *)&ihandle,(void *)&handle.ze,sizeof(ihandle));
       memcpy((void *)&ihandle,(void *)&myfd,sizeof(int));
 
@@ -664,11 +667,8 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 	std::cerr << "SharedMemoryMPI.cc "<<zeContext<<" "<<zeDevice<<std::endl;
 	std::cerr << "SharedMemoryMPI.cc zeMemOpenIpcHandle failed for rank "<<r<<" "<<std::hex<<err<<std::dec<<std::endl; 
 	exit(EXIT_FAILURE);
-      } else {
-	std::cout << "SharedMemoryMPI.cc zeMemOpenIpcHandle succeeded for rank "<<r<<std::endl;
-	std::cout << "SharedMemoryMPI.cc zeMemOpenIpcHandle pointer is "<<std::hex<<thisBuf<<std::dec<<std::endl;
       }
-      assert(thisBuf!=nullptr);
+      GRID_ASSERT(thisBuf!=nullptr);
     }
 #endif
 #ifdef GRID_CUDA
@@ -709,8 +709,8 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 {
   std::cout << Mheader "SharedMemoryAllocate "<< bytes<< " MMAP implementation "<< GRID_SHM_PATH <<std::endl;
-  assert(_ShmSetup==1);
+  GRID_ASSERT(_ShmSetup==1);
-  assert(_ShmAlloc==0);
+  GRID_ASSERT(_ShmAlloc==0);
   //////////////////////////////////////////////////////////////////////////////////////////////////////////
   // allocate the shared windows for our group
   //////////////////////////////////////////////////////////////////////////////////////////////////////////
@@ -740,13 +740,14 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
     void *ptr = (void *) mmap(NULL, bytes, PROT_READ | PROT_WRITE, mmap_flag,fd, 0); 
     if ( ptr == (void *)MAP_FAILED ) {    
       printf("mmap %s failed\n",shm_name);
-      perror("failed mmap");      assert(0);    
+      perror("failed mmap");      GRID_ASSERT(0);    
     }
-    assert(((uint64_t)ptr&0x3F)==0);
+    GRID_ASSERT(((uint64_t)ptr&0x3F)==0);
     close(fd);
     WorldShmCommBufs[r] =ptr;
     //    std::cout << Mheader "Set WorldShmCommBufs["<<r<<"]="<<ptr<< "("<< bytes<< "bytes)"<<std::endl;
   }
+  std::cout<< Mheader " Intra-node IPC setup is complete "<<std::endl;
   _ShmAlloc=1;
   _ShmAllocBytes  = bytes;
 };
@@ -756,8 +757,8 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 {
   std::cout << Mheader "SharedMemoryAllocate "<< bytes<< " MMAP anonymous implementation "<<std::endl;
-  assert(_ShmSetup==1);
+  GRID_ASSERT(_ShmSetup==1);
-  assert(_ShmAlloc==0);
+  GRID_ASSERT(_ShmAlloc==0);
   //////////////////////////////////////////////////////////////////////////////////////////////////////////
   // allocate the shared windows for our group
   //////////////////////////////////////////////////////////////////////////////////////////////////////////
@@ -768,7 +769,7 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
   // Hugetlbf and others map filesystems as mappable huge pages
   ////////////////////////////////////////////////////////////////////////////////////////////
   char shm_name [NAME_MAX];
-  assert(WorldShmSize == 1);
+  GRID_ASSERT(WorldShmSize == 1);
   for(int r=0;r<WorldShmSize;r++){
     
     int fd=-1;
@@ -782,9 +783,9 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
     void *ptr = (void *) mmap(NULL, bytes, PROT_READ | PROT_WRITE, mmap_flag,fd, 0); 
     if ( ptr == (void *)MAP_FAILED ) {    
       printf("mmap %s failed\n",shm_name);
-      perror("failed mmap");      assert(0);    
+      perror("failed mmap");      GRID_ASSERT(0);    
     }
-    assert(((uint64_t)ptr&0x3F)==0);
+    GRID_ASSERT(((uint64_t)ptr&0x3F)==0);
     close(fd);
     WorldShmCommBufs[r] =ptr;
     //    std::cout << "Set WorldShmCommBufs["<<r<<"]="<<ptr<< "("<< bytes<< "bytes)"<<std::endl;
@@ -803,8 +804,8 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 { 
   std::cout << Mheader "SharedMemoryAllocate "<< bytes<< " SHMOPEN implementation "<<std::endl;
-  assert(_ShmSetup==1);
+  GRID_ASSERT(_ShmSetup==1);
-  assert(_ShmAlloc==0); 
+  GRID_ASSERT(_ShmAlloc==0); 
   MPI_Barrier(WorldShmComm);
   WorldShmCommBufs.resize(WorldShmSize);
 
@@ -835,7 +836,7 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 	perror("failed mmap");     
 	assert(0);    
       }
-      assert(((uint64_t)ptr&0x3F)==0);
+      GRID_ASSERT(((uint64_t)ptr&0x3F)==0);
       
       WorldShmCommBufs[r] =ptr;
       close(fd);
@@ -856,8 +857,8 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
       if ( fd<0 ) {	perror("failed shm_open");	assert(0);      }
       
       void * ptr =  mmap(NULL,size, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
-      if ( ptr == MAP_FAILED ) {       perror("failed mmap");      assert(0);    }
+      if ( ptr == MAP_FAILED ) {       perror("failed mmap");      GRID_ASSERT(0);    }
-      assert(((uint64_t)ptr&0x3F)==0);
+      GRID_ASSERT(((uint64_t)ptr&0x3F)==0);
       WorldShmCommBufs[r] =ptr;
 
       close(fd);
@@ -880,14 +881,14 @@ void GlobalSharedMemory::SharedMemoryZero(void *dest,size_t bytes)
   bzero(dest,bytes);
 #endif
 }
-void GlobalSharedMemory::SharedMemoryCopy(void *dest,void *src,size_t bytes)
+//void GlobalSharedMemory::SharedMemoryCopy(void *dest,void *src,size_t bytes)
-{
+//{
-#if defined(GRID_CUDA) || defined(GRID_HIP) || defined(GRID_SYCL)
+//#if defined(GRID_CUDA) || defined(GRID_HIP) || defined(GRID_SYCL)
-  acceleratorCopyToDevice(src,dest,bytes);
+//  acceleratorCopyToDevice(src,dest,bytes);
-#else   
+//#else   
-  bcopy(src,dest,bytes);
+//  bcopy(src,dest,bytes);
-#endif
+//#endif
-}
+//}
 ////////////////////////////////////////////////////////
 // Global shared functionality finished
 // Now move to per communicator functionality
@@ -914,7 +915,7 @@ void SharedMemory::SetCommunicator(Grid_MPI_Comm comm)
   //////////////////////////////////////////////////////////////////////
   // Map ShmRank to WorldShmRank and use the right buffer
   //////////////////////////////////////////////////////////////////////
-  assert (GlobalSharedMemory::ShmAlloc()==1);
+  GRID_ASSERT (GlobalSharedMemory::ShmAlloc()==1);
   heap_size = GlobalSharedMemory::ShmAllocBytes();
   for(int r=0;r<ShmSize;r++){
 
@@ -923,6 +924,7 @@ void SharedMemory::SetCommunicator(Grid_MPI_Comm comm)
     MPI_Allreduce(MPI_IN_PLACE,&wsr,1,MPI_UINT32_T,MPI_SUM,ShmComm);
 
     ShmCommBufs[r] = GlobalSharedMemory::WorldShmCommBufs[wsr];
+    //    std::cerr << " SetCommunicator rank "<<r<<" comm "<<ShmCommBufs[r] <<std::endl;
   }
   ShmBufferFreeAll();
 
@@ -975,19 +977,18 @@ void SharedMemory::SharedMemoryTest(void)
        check[0]=GlobalSharedMemory::WorldNode;
        check[1]=r;
        check[2]=magic;
-       GlobalSharedMemory::SharedMemoryCopy( ShmCommBufs[r], check, 3*sizeof(uint64_t));
+       acceleratorCopyToDevice(check,ShmCommBufs[r],3*sizeof(uint64_t));
     }
   }
   ShmBarrier();
   for(uint64_t r=0;r<ShmSize;r++){
-    ShmBarrier();
+    acceleratorCopyFromDevice(ShmCommBufs[r],check,3*sizeof(uint64_t));
-    GlobalSharedMemory::SharedMemoryCopy(check,ShmCommBufs[r], 3*sizeof(uint64_t));
+    GRID_ASSERT(check[0]==GlobalSharedMemory::WorldNode);
-    ShmBarrier();
+    GRID_ASSERT(check[1]==r);
-    assert(check[0]==GlobalSharedMemory::WorldNode);
+    GRID_ASSERT(check[2]==magic);
-    assert(check[1]==r);
-    assert(check[2]==magic);
-    ShmBarrier();
   }
+  ShmBarrier();
+  std::cout << GridLogDebug << " SharedMemoryTest has passed "<<std::endl;
 }
 
 void *SharedMemory::ShmBuffer(int rank)
@@ -1002,12 +1003,14 @@ void *SharedMemory::ShmBuffer(int rank)
 void *SharedMemory::ShmBufferTranslate(int rank,void * local_p)
 {
   int gpeer = ShmRanks[rank];
-  assert(gpeer!=ShmRank); // never send to self
+  GRID_ASSERT(gpeer!=ShmRank); // never send to self
+  //  std::cout << "ShmBufferTranslate for rank " << rank<<" peer "<<gpeer<<std::endl;
   if (gpeer == MPI_UNDEFINED){
     return NULL;
   } else { 
     uint64_t offset = (uint64_t)local_p - (uint64_t)ShmCommBufs[ShmRank];
     uint64_t remote = (uint64_t)ShmCommBufs[gpeer]+offset;
+    //    std::cout << "ShmBufferTranslate : local,offset,remote "<<std::hex<<local_p<<" "<<offset<<" "<<remote<<std::dec<<std::endl;
     return (void *) remote;
   }
 }

Grid/communicator/SharedMemoryNone.cc

@@ -34,7 +34,7 @@ NAMESPACE_BEGIN(Grid);
 /*Construct from an MPI communicator*/
 void GlobalSharedMemory::Init(Grid_MPI_Comm comm)
 {
-  assert(_ShmSetup==0);
+  GRID_ASSERT(_ShmSetup==0);
   WorldComm = 0;
   WorldRank = 0;
   WorldSize = 1;
@@ -62,8 +62,8 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 {
   std::cout << header "SharedMemoryAllocate "<< bytes<< " GPU implementation "<<std::endl;
   void * ShmCommBuf ; 
-  assert(_ShmSetup==1);
+  GRID_ASSERT(_ShmSetup==1);
-  assert(_ShmAlloc==0);
+  GRID_ASSERT(_ShmAlloc==0);
 
   ///////////////////////////////////////////////////////////////////////////////////////////////////////////
   // Each MPI rank should allocate our own buffer
@@ -92,8 +92,8 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 {
   void * ShmCommBuf ; 
-  assert(_ShmSetup==1);
+  GRID_ASSERT(_ShmSetup==1);
-  assert(_ShmAlloc==0);
+  GRID_ASSERT(_ShmAlloc==0);
   int mmap_flag =0;
 #ifdef MAP_ANONYMOUS
   mmap_flag = mmap_flag| MAP_SHARED | MAP_ANONYMOUS;
@@ -122,17 +122,17 @@ void GlobalSharedMemory::SharedMemoryZero(void *dest,size_t bytes)
 {
   acceleratorMemSet(dest,0,bytes);
 }
-void GlobalSharedMemory::SharedMemoryCopy(void *dest,void *src,size_t bytes)
+//void GlobalSharedMemory::SharedMemoryCopy(void *dest,void *src,size_t bytes)
-{
+//{
-  acceleratorCopyToDevice(src,dest,bytes);
+//  acceleratorCopyToDevice(src,dest,bytes);
-}
+//}
 ////////////////////////////////////////////////////////
 // Global shared functionality finished
 // Now move to per communicator functionality
 ////////////////////////////////////////////////////////
 void SharedMemory::SetCommunicator(Grid_MPI_Comm comm)
 {
-  assert(GlobalSharedMemory::ShmAlloc()==1);
+  GRID_ASSERT(GlobalSharedMemory::ShmAlloc()==1);
   ShmRanks.resize(1);
   ShmCommBufs.resize(1);
   ShmRanks[0] = 0;

Grid/cshift/Cshift.h

@@ -51,7 +51,6 @@ 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

@@ -30,12 +30,11 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 NAMESPACE_BEGIN(Grid);
 
 extern std::vector<std::pair<int,int> > Cshift_table; 
-extern commVector<std::pair<int,int> > Cshift_table_device; 
+extern deviceVector<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);
@@ -45,16 +44,13 @@ 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,cshiftVector<vobj> &buffer,int dimension,int plane,int cbmask, int off=0)
+Gather_plane_simple (const Lattice<vobj> &rhs,deviceVector<vobj> &buffer,int dimension,int plane,int cbmask, int off=0)
 {
   int rd = rhs.Grid()->_rdimensions[dimension];
 
@@ -94,17 +90,10 @@ Gather_plane_simple (const Lattice<vobj> &rhs,cshiftVector<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
   }
 }
 
@@ -129,7 +118,6 @@ 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;
@@ -140,21 +128,10 @@ 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;
@@ -175,33 +152,13 @@ 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,cshiftVector<vobj> &buffer, int dimension,int plane,int cbmask)
+template<class vobj> void Scatter_plane_simple (Lattice<vobj> &rhs,deviceVector<vobj> &buffer, int dimension,int plane,int cbmask)
 {
   int rd = rhs.Grid()->_rdimensions[dimension];
 
@@ -245,17 +202,10 @@ template<class vobj> void Scatter_plane_simple (Lattice<vobj> &rhs,cshiftVector<
   {
     auto buffer_p = & buffer[0];
     auto table = MapCshiftTable();
-#ifdef ACCELERATOR_CSHIFT    
+    autoView( rhs_v, rhs, AcceleratorWriteDiscard);
-    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
   }
 }
 
@@ -278,8 +228,7 @@ 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, AcceleratorWriteDiscard);
-    autoView( rhs_v , rhs, AcceleratorWrite);
     accelerator_for(nn,e1*e2,1,{
 	int n = nn%e1;
 	int b = nn/e1;
@@ -287,21 +236,13 @@ 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 
     // Test_cshift_red_black code.
-    std::cout << "Scatter_plane merge assert(0); think this is buggy FIXME "<< std::endl;// think this is buggy FIXME
+    std::cout << "Scatter_plane merge GRID_ASSERT(0); think this is buggy FIXME "<< std::endl;// think this is buggy FIXME
     std::cout<<" Unthreaded warning -- buffer is not densely packed ??"<<std::endl;
-    assert(0); // This will fail if hit on GPU
+    GRID_ASSERT(0); // This will fail if hit on GPU
     autoView( rhs_v, rhs, CpuWrite);
     for(int n=0;n<e1;n++){
       for(int b=0;b<e2;b++){
@@ -360,19 +301,11 @@ 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);
+    autoView(lhs_v , lhs, AcceleratorWriteDiscard);
     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
   }
 }
 
@@ -412,19 +345,11 @@ 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

@@ -29,9 +29,13 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 #ifndef _GRID_CSHIFT_MPI_H_
 #define _GRID_CSHIFT_MPI_H_
 
-
 NAMESPACE_BEGIN(Grid); 
 
+#ifdef GRID_CHECKSUM_COMMS
+extern uint64_t checksum_index;
+#endif
+
+const int Cshift_verbose=0;
 template<class vobj> Lattice<vobj> Cshift(const Lattice<vobj> &rhs,int dimension,int shift)
 {
   typedef typename vobj::vector_type vector_type;
@@ -45,6 +49,20 @@ template<class vobj> Lattice<vobj> Cshift(const Lattice<vobj> &rhs,int dimension
   // Map to always positive shift modulo global full dimension.
   shift = (shift+fd)%fd;
 
+  if( shift ==0 ) {
+    ret = rhs;
+    return ret;
+  }
+  //
+  // Potential easy fast cases:
+  // Shift is a multiple of the local lattice extent.
+  // Then need only to shift whole subvolumes
+  int L = rhs.Grid()->_ldimensions[dimension];
+  if ( (shift%L )==0 && !rhs.Grid()->CheckerBoarded(dimension) ) {
+    Cshift_simple(ret,rhs,dimension,shift);
+    return ret;
+  }
+  
   ret.Checkerboard() = rhs.Grid()->CheckerBoardDestination(rhs.Checkerboard(),shift,dimension);
         
   // the permute type
@@ -65,10 +83,59 @@ template<class vobj> Lattice<vobj> Cshift(const Lattice<vobj> &rhs,int dimension
     Cshift_comms(ret,rhs,dimension,shift);
   }
   t1=usecond();
-  //  std::cout << GridLogPerformance << "Cshift took "<< (t1-t0)/1e3 << " ms"<<std::endl;
+  if(Cshift_verbose) std::cout << GridLogPerformance << "Cshift took "<< (t1-t0)/1e3 << " ms"<<std::endl;
   return ret;
 }
 
+template<class vobj> void Cshift_simple(Lattice<vobj>& ret,const Lattice<vobj> &rhs,int dimension,int shift)
+{
+  GridBase *grid=rhs.Grid();
+  int comm_proc, xmit_to_rank, recv_from_rank;
+  
+  int fd              = rhs.Grid()->_fdimensions[dimension];
+  int rd              = rhs.Grid()->_rdimensions[dimension];
+  int ld              = rhs.Grid()->_ldimensions[dimension];
+  int pd              = rhs.Grid()->_processors[dimension];
+  int simd_layout     = rhs.Grid()->_simd_layout[dimension];
+  int comm_dim        = rhs.Grid()->_processors[dimension] >1 ;
+
+  comm_proc = ((shift)/ld)%pd;
+
+  grid->ShiftedRanks(dimension,comm_proc,xmit_to_rank,recv_from_rank);
+  if(comm_dim) {
+
+    int64_t bytes = sizeof(vobj) * grid->oSites();
+
+    autoView(rhs_v , rhs, AcceleratorRead);
+    autoView(ret_v , ret, AcceleratorWrite);
+    void *send_buf  = (void *)&rhs_v[0];
+    void *recv_buf  = (void *)&ret_v[0];
+
+#ifdef ACCELERATOR_AWARE_MPI
+    grid->SendToRecvFrom(send_buf,
+			 xmit_to_rank,
+			 recv_buf,
+			 recv_from_rank,
+			 bytes);
+#else
+    static hostVector<vobj> hrhs; hrhs.resize(grid->oSites());
+    static hostVector<vobj> hret; hret.resize(grid->oSites());
+
+    void *hsend_buf = (void *)&hrhs[0];
+    void *hrecv_buf = (void *)&hret[0];
+
+    acceleratorCopyFromDevice(send_buf,hsend_buf,bytes);
+
+    grid->SendToRecvFrom(hsend_buf,
+			 xmit_to_rank,
+			 hrecv_buf,
+			 recv_from_rank,
+			 bytes);
+
+    acceleratorCopyToDevice(hrecv_buf,recv_buf,bytes);
+#endif
+  }
+}
 template<class vobj> void Cshift_comms(Lattice<vobj>& ret,const Lattice<vobj> &rhs,int dimension,int shift)
 {
   int sshift[2];
@@ -94,7 +161,7 @@ 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);
@@ -104,8 +171,6 @@ 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;
@@ -119,14 +184,19 @@ template<class vobj> void Cshift_comms(Lattice<vobj> &ret,const Lattice<vobj> &r
   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);
+  GRID_ASSERT(simd_layout==1);
-  assert(comm_dim==1);
+  GRID_ASSERT(comm_dim==1);
-  assert(shift>=0);
+  GRID_ASSERT(shift>=0);
-  assert(shift<fd);
+  GRID_ASSERT(shift<fd);
   
   int buffer_size = rhs.Grid()->_slice_nblock[dimension]*rhs.Grid()->_slice_block[dimension];
-  static cshiftVector<vobj> send_buf; send_buf.resize(buffer_size);
+  static deviceVector<vobj> send_buf; send_buf.resize(buffer_size);
-  static cshiftVector<vobj> recv_buf; recv_buf.resize(buffer_size);
+  static deviceVector<vobj> recv_buf; recv_buf.resize(buffer_size);
+#ifndef ACCELERATOR_AWARE_MPI
+  int pad = (8 + sizeof(vobj) - 1) / sizeof(vobj);
+  static hostVector<vobj> hsend_buf; hsend_buf.resize(buffer_size+pad);
+  static hostVector<vobj> hrecv_buf; hrecv_buf.resize(buffer_size+pad);
+#endif
   
   int cb= (cbmask==0x2)? Odd : Even;
   int sshift= rhs.Grid()->CheckerBoardShiftForCB(rhs.Checkerboard(),dimension,shift,cb);
@@ -141,9 +211,11 @@ template<class vobj> void Cshift_comms(Lattice<vobj> &ret,const Lattice<vobj> &r
     int comm_proc = ((x+sshift)/rd)%pd;
     
     if (comm_proc==0) {
+      FlightRecorder::StepLog("Cshift_Copy_plane");
       tcopy-=usecond();
       Copy_plane(ret,rhs,dimension,x,sx,cbmask); 
       tcopy+=usecond();
+      FlightRecorder::StepLog("Cshift_Copy_plane_complete");
     } else {
 
       int words = buffer_size;
@@ -151,39 +223,84 @@ template<class vobj> void Cshift_comms(Lattice<vobj> &ret,const Lattice<vobj> &r
 
       int bytes = words * sizeof(vobj);
 
+      FlightRecorder::StepLog("Cshift_Gather_plane");
       tgather-=usecond();
       Gather_plane_simple (rhs,send_buf,dimension,sx,cbmask);
       tgather+=usecond();
+      FlightRecorder::StepLog("Cshift_Gather_plane_complete");
 
       //      int rank           = grid->_processor;
       int recv_from_rank;
       int xmit_to_rank;
+
       grid->ShiftedRanks(dimension,comm_proc,xmit_to_rank,recv_from_rank);
       
       tcomms-=usecond();
-      //      grid->Barrier();
+      grid->Barrier();
 
+      FlightRecorder::StepLog("Cshift_SendRecv");
+#ifdef ACCELERATOR_AWARE_MPI
       grid->SendToRecvFrom((void *)&send_buf[0],
 			   xmit_to_rank,
 			   (void *)&recv_buf[0],
 			   recv_from_rank,
 			   bytes);
+#else
+      // bouncy bouncy
+      acceleratorCopyFromDevice(&send_buf[0],&hsend_buf[0],bytes);
+
+#ifdef GRID_CHECKSUM_COMMS
+      GRID_ASSERT(bytes % 8 == 0);
+      checksum_index++;
+      uint64_t xsum = checksum_gpu((uint64_t*)&send_buf[0], bytes / 8) ^ (1 + checksum_index);
+      *(uint64_t*)(((char*)&hsend_buf[0]) + bytes) = xsum;
+      bytes += 8;
+#endif
+
+      grid->SendToRecvFrom((void *)&hsend_buf[0],
+			   xmit_to_rank,
+			   (void *)&hrecv_buf[0],
+			   recv_from_rank,
+			   bytes);
+
+#ifdef GRID_CHECKSUM_COMMS
+      bytes -= 8;
+      acceleratorCopyToDevice(&hrecv_buf[0],&recv_buf[0],bytes);
+      uint64_t expected_cs = *(uint64_t*)(((char*)&hrecv_buf[0]) + bytes);
+      uint64_t computed_cs = checksum_gpu((uint64_t*)&recv_buf[0], bytes / 8) ^ (1 + checksum_index);
+      std::cout << GridLogComms<< " Cshift: "
+		<<" dim"<<dimension
+		<<" shift "<<shift
+		<< " rank "<< grid->ThisRank()
+		<<" Coor "<<grid->ThisProcessorCoor()
+		<<" send "<<xsum<<" to   "<<xmit_to_rank
+		<<" recv "<<computed_cs<<" from "<<recv_from_rank
+		<<std::endl;
+      GRID_ASSERT(expected_cs == computed_cs);
+#else
+      acceleratorCopyToDevice(&hrecv_buf[0],&recv_buf[0],bytes);
+#endif
+
+#endif
+      FlightRecorder::StepLog("Cshift_SendRecv_complete");
+
       xbytes+=bytes;
-      //      grid->Barrier();
+      grid->Barrier();
       tcomms+=usecond();
+      FlightRecorder::StepLog("Cshift_barrier_complete");
 
       tscatter-=usecond();
       Scatter_plane_simple (ret,recv_buf,dimension,x,cbmask);
       tscatter+=usecond();
     }
   }
-  /*
+  if (Cshift_verbose){
     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)
@@ -205,10 +322,10 @@ template<class vobj> void  Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vo
   //	    << " ld "<<ld<<" pd " << pd<<" simd_layout "<<simd_layout 
   //	    << " comm_dim " << comm_dim << " cbmask " << cbmask <<std::endl;
 
-  assert(comm_dim==1);
+  GRID_ASSERT(comm_dim==1);
-  assert(simd_layout==2);
+  GRID_ASSERT(simd_layout==2);
-  assert(shift>=0);
+  GRID_ASSERT(shift>=0);
-  assert(shift<fd);
+  GRID_ASSERT(shift<fd);
 
   RealD tcopy=0.0;
   RealD tgather=0.0;
@@ -224,8 +341,8 @@ 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<cshiftVector<scalar_object> >  send_buf_extract; send_buf_extract.resize(Nsimd);
+  static std::vector<deviceVector<scalar_object> >  send_buf_extract; send_buf_extract.resize(Nsimd);
-  static std::vector<cshiftVector<scalar_object> >  recv_buf_extract; recv_buf_extract.resize(Nsimd);
+  static std::vector<deviceVector<scalar_object> >  recv_buf_extract; recv_buf_extract.resize(Nsimd);
   scalar_object *  recv_buf_extract_mpi;
   scalar_object *  send_buf_extract_mpi;
 
@@ -233,6 +350,18 @@ template<class vobj> void  Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vo
     send_buf_extract[s].resize(buffer_size);
     recv_buf_extract[s].resize(buffer_size);
   }
+#ifndef ACCELERATOR_AWARE_MPI
+#ifdef GRID_CHECKSUM_COMMS
+  buffer_size += (8 + sizeof(vobj) - 1) / sizeof(vobj);
+#endif
+
+  static hostVector<vobj> hsend_buf; hsend_buf.resize(buffer_size);
+  static hostVector<vobj> hrecv_buf; hrecv_buf.resize(buffer_size);
+
+#ifdef GRID_CHECKSUM_COMMS
+  buffer_size -= (8 + sizeof(vobj) - 1) / sizeof(vobj);
+#endif
+#endif
   
   int bytes = buffer_size*sizeof(scalar_object);
 
@@ -275,252 +404,62 @@ template<class vobj> void  Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vo
 
       if (nbr_ic) nbr_lane|=inner_bit;
 
-      assert (sx == nbr_ox);
+      GRID_ASSERT (sx == nbr_ox);
 
       if(nbr_proc){
 	grid->ShiftedRanks(dimension,nbr_proc,xmit_to_rank,recv_from_rank); 
 
 	tcomms-=usecond();
-	//	grid->Barrier();
+	grid->Barrier();
 
 	send_buf_extract_mpi = &send_buf_extract[nbr_lane][0];
 	recv_buf_extract_mpi = &recv_buf_extract[i][0];
+#ifdef ACCELERATOR_AWARE_MPI
 	grid->SendToRecvFrom((void *)send_buf_extract_mpi,
 			     xmit_to_rank,
 			     (void *)recv_buf_extract_mpi,
 			     recv_from_rank,
 			     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 "<<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)
+      // bouncy bouncy
-{
+	acceleratorCopyFromDevice((void *)send_buf_extract_mpi,(void *)&hsend_buf[0],bytes);
-  typedef typename vobj::vector_type vector_type;
+#ifdef GRID_CHECKSUM_COMMS
-  typedef typename vobj::scalar_type scalar_type;
+	assert(bytes % 8 == 0);
-
+	checksum_index++;
-  GridBase *grid=rhs.Grid();
+	uint64_t xsum = checksum_gpu((uint64_t*)send_buf_extract_mpi, bytes / 8) ^ (1 + checksum_index);
-  Lattice<vobj> temp(rhs.Grid());
+	*(uint64_t*)(((char*)&hsend_buf[0]) + bytes) = xsum;
-
+	bytes += 8;
-  int fd              = rhs.Grid()->_fdimensions[dimension];
+#endif
-  int rd              = rhs.Grid()->_rdimensions[dimension];
+	grid->SendToRecvFrom((void *)&hsend_buf[0],
-  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],
+			     (void *)&hrecv_buf[0],
 			     recv_from_rank,
 			     bytes);
-      xbytes+=bytes;
+#ifdef GRID_CHECKSUM_COMMS
-      acceleratorCopyDeviceToDevice((void *)&recv_buf[0],(void *)&recv_buf_v[0],bytes);
+	bytes -= 8;
+	acceleratorCopyToDevice((void *)&hrecv_buf[0],(void *)recv_buf_extract_mpi,bytes);
+	uint64_t expected_cs = *(uint64_t*)(((char*)&hrecv_buf[0]) + bytes);
+	uint64_t computed_cs = checksum_gpu((uint64_t*)recv_buf_extract_mpi, bytes / 8) ^ (1 + checksum_index);
 
-      //      grid->Barrier();
+	std::cout << GridLogComms<< " Cshift_comms_simd: "
+		<<" dim"<<dimension
+		<<" shift "<<shift
+		<< " rank "<< grid->ThisRank()
+		<<" Coor "<<grid->ThisProcessorCoor()
+		<<" send "<<xsum<<" to   "<<xmit_to_rank
+		<<" recv "<<computed_cs<<" from "<<recv_from_rank
+		<<std::endl;
+	assert(expected_cs == computed_cs);
+#else
+	acceleratorCopyToDevice((void *)&hrecv_buf[0],(void *)recv_buf_extract_mpi,bytes);
+#endif
+
+#endif
+
+	xbytes+=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];
@@ -530,17 +469,16 @@ template<class vobj> void  Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vo
     tscatter-=usecond();
     Scatter_plane_merge(ret,rpointers,dimension,x,cbmask);
     tscatter+=usecond();
-
   }
-  /*
+  if(Cshift_verbose){
     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;
+    std::cout << GridLogPerformance << " Cshift BW      "<<(2.0*xbytes)/tcomms<<" MB/s "<<2*xbytes<< " Bytes "<<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; 
-commVector<std::pair<int,int> > Cshift_table_device; 
+deviceVector<std::pair<int,int> > Cshift_table_device; 
 NAMESPACE_END(Grid);

Grid/lattice/Lattice_ET.h

@@ -245,7 +245,7 @@ template <class T1,typename std::enable_if<is_lattice<T1>::value, T1>::type * =
 inline void CBFromExpression(int &cb, const T1 &lat)  // Lattice leaf
 {
   if ((cb == Odd) || (cb == Even)) {
-    assert(cb == lat.Checkerboard());
+    GRID_ASSERT(cb == lat.Checkerboard());
   }
   cb = lat.Checkerboard();
 }

Grid/lattice/Lattice_arith.h

@@ -257,17 +257,30 @@ void axpby(Lattice<vobj> &ret,sobj a,sobj b,const Lattice<vobj> &x,const Lattice
   });
 }
 
+#define FAST_AXPY_NORM
 template<class sobj,class vobj> inline
 RealD axpy_norm(Lattice<vobj> &ret,sobj a,const Lattice<vobj> &x,const Lattice<vobj> &y)
 {
   GRID_TRACE("axpy_norm");
+#ifdef FAST_AXPY_NORM
   return axpy_norm_fast(ret,a,x,y);
+#else
+  ret = a*x+y;
+  RealD nn=norm2(ret);
+  return nn;
+#endif
 }
 template<class sobj,class vobj> inline
 RealD axpby_norm(Lattice<vobj> &ret,sobj a,sobj b,const Lattice<vobj> &x,const Lattice<vobj> &y)
 {
   GRID_TRACE("axpby_norm");
+#ifdef FAST_AXPY_NORM
   return axpby_norm_fast(ret,a,b,x,y);
+#else
+  ret = a*x+b*y;
+  RealD nn=norm2(ret);
+  return nn;
+#endif
 }
 
 /// Trace product

Grid/lattice/Lattice_base.h

@@ -120,12 +120,12 @@ public:
     GRID_TRACE("ExpressionTemplateEval");
     GridBase *egrid(nullptr);
     GridFromExpression(egrid,expr);
-    assert(egrid!=nullptr);
+    GRID_ASSERT(egrid!=nullptr);
     conformable(this->_grid,egrid);
 
     int cb=-1;
     CBFromExpression(cb,expr);
-    assert( (cb==Odd) || (cb==Even));
+    GRID_ASSERT( (cb==Odd) || (cb==Even));
     this->checkerboard=cb;
     
     auto exprCopy = expr;
@@ -144,12 +144,12 @@ public:
     GRID_TRACE("ExpressionTemplateEval");
     GridBase *egrid(nullptr);
     GridFromExpression(egrid,expr);
-    assert(egrid!=nullptr);
+    GRID_ASSERT(egrid!=nullptr);
     conformable(this->_grid,egrid);
 
     int cb=-1;
     CBFromExpression(cb,expr);
-    assert( (cb==Odd) || (cb==Even));
+    GRID_ASSERT( (cb==Odd) || (cb==Even));
     this->checkerboard=cb;
 
     auto exprCopy = expr;
@@ -168,12 +168,12 @@ public:
     GRID_TRACE("ExpressionTemplateEval");
     GridBase *egrid(nullptr);
     GridFromExpression(egrid,expr);
-    assert(egrid!=nullptr);
+    GRID_ASSERT(egrid!=nullptr);
     conformable(this->_grid,egrid);
 
     int cb=-1;
     CBFromExpression(cb,expr);
-    assert( (cb==Odd) || (cb==Even));
+    GRID_ASSERT( (cb==Odd) || (cb==Even));
     this->checkerboard=cb;
     auto exprCopy = expr;
     ExpressionViewOpen(exprCopy);
@@ -191,11 +191,11 @@ public:
   Lattice(const LatticeUnaryExpression<Op,T1> & expr) {
     this->_grid = nullptr;
     GridFromExpression(this->_grid,expr);
-    assert(this->_grid!=nullptr);
+    GRID_ASSERT(this->_grid!=nullptr);
 
     int cb=-1;
     CBFromExpression(cb,expr);
-    assert( (cb==Odd) || (cb==Even));
+    GRID_ASSERT( (cb==Odd) || (cb==Even));
     this->checkerboard=cb;
 
     resize(this->_grid->oSites());
@@ -206,11 +206,11 @@ public:
   Lattice(const LatticeBinaryExpression<Op,T1,T2> & expr) {
     this->_grid = nullptr;
     GridFromExpression(this->_grid,expr);
-    assert(this->_grid!=nullptr);
+    GRID_ASSERT(this->_grid!=nullptr);
 
     int cb=-1;
     CBFromExpression(cb,expr);
-    assert( (cb==Odd) || (cb==Even));
+    GRID_ASSERT( (cb==Odd) || (cb==Even));
     this->checkerboard=cb;
 
     resize(this->_grid->oSites());
@@ -221,11 +221,11 @@ public:
   Lattice(const LatticeTrinaryExpression<Op,T1,T2,T3> & expr) {
     this->_grid = nullptr;
     GridFromExpression(this->_grid,expr);
-    assert(this->_grid!=nullptr);
+    GRID_ASSERT(this->_grid!=nullptr);
 
     int cb=-1;
     CBFromExpression(cb,expr);
-    assert( (cb==Odd) || (cb==Even));
+    GRID_ASSERT( (cb==Odd) || (cb==Even));
     this->checkerboard=cb;
 
     resize(this->_grid->oSites());
@@ -237,16 +237,19 @@ public:
     vobj vtmp;
     vtmp = r;
 #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;
@@ -261,7 +264,7 @@ public:
   Lattice(GridBase *grid,ViewMode mode=AcceleratorWriteDiscard) { 
     this->_grid = grid;
     resize(this->_grid->oSites());
-    assert((((uint64_t)&this->_odata[0])&0xF) ==0);
+    GRID_ASSERT((((uint64_t)&this->_odata[0])&0xF) ==0);
     this->checkerboard=0;
     SetViewMode(mode);
   }

Grid/lattice/Lattice_basis.h

@@ -53,36 +53,19 @@ 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;
 
-  Vector<View> basis_v; basis_v.reserve(basis.size());
+  hostVector<View>  h_basis_v(basis.size());
-  typedef typename std::remove_reference<decltype(basis_v[0][0])>::type vobj;
+  deviceVector<View> d_basis_v(basis.size());
+  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++){
-    basis_v.push_back(basis[k].View(AcceleratorWrite));
+    h_basis_v[k] = basis[k].View(AcceleratorWrite);
+    acceleratorPut(d_basis_v[k],h_basis_v[k]);
   }
 
-#if ( !(defined(GRID_CUDA) || defined(GRID_HIP) || defined(GRID_SYCL)) )
+  View *basis_vp = &d_basis_v[0];
-  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
@@ -91,17 +74,19 @@ 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
 
-  Vector <vobj> Bt(siteBlock * nrot); 
+  deviceVector <vobj> Bt(siteBlock * nrot); 
   auto Bp=&Bt[0];
 
   // GPU readable copy of matrix
-  Vector<Coeff_t> Qt_jv(Nm*Nm);
+  hostVector<Coeff_t> h_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;
-      Qt_p[i]=Qt(j,k);
+      h_Qt_jv[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){
@@ -137,9 +122,8 @@ 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++) basis_v[k].ViewClose();
+  for(int k=0;k<basis.size();k++) h_basis_v[k].ViewClose();
 }
 
 // Extract a single rotated vector
@@ -152,16 +136,19 @@ void basisRotateJ(Field &result,std::vector<Field> &basis,Eigen::MatrixXd& Qt,in
 
   result.Checkerboard() = basis[0].Checkerboard();
 
-  Vector<View> basis_v; basis_v.reserve(basis.size());
+  hostVector<View>  h_basis_v(basis.size());
+  deviceVector<View> d_basis_v(basis.size());
   for(int k=0;k<basis.size();k++){
-    basis_v.push_back(basis[k].View(AcceleratorRead));
+    h_basis_v[k]=basis[k].View(AcceleratorRead);
+    acceleratorPut(d_basis_v[k],h_basis_v[k]);
   }
-  vobj zz=Zero();
-  Vector<double> Qt_jv(Nm);
-  double * Qt_j = & Qt_jv[0];
-  for(int k=0;k<Nm;++k) Qt_j[k]=Qt(j,k);
 
-  auto basis_vp=& basis_v[0];
+  vobj zz=Zero();
+  deviceVector<double> Qt_jv(Nm);
+  double * Qt_j = & Qt_jv[0];
+  for(int k=0;k<Nm;++k) acceleratorPut(Qt_j[k],Qt(j,k));
+
+  auto basis_vp=& d_basis_v[0];
   autoView(result_v,result,AcceleratorWrite);
   accelerator_for(ss, grid->oSites(),vobj::Nsimd(),{
     vobj zzz=Zero();
@@ -171,7 +158,7 @@ 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++) basis_v[k].ViewClose();
+  for(int k=0;k<basis.size();k++) h_basis_v[k].ViewClose();
 }
 
 template<class Field>
@@ -179,9 +166,9 @@ void basisReorderInPlace(std::vector<Field> &_v,std::vector<RealD>& sort_vals, s
 {
   int vlen = idx.size();
 
-  assert(vlen>=1);
+  GRID_ASSERT(vlen>=1);
-  assert(vlen<=sort_vals.size());
+  GRID_ASSERT(vlen<=sort_vals.size());
-  assert(vlen<=_v.size());
+  GRID_ASSERT(vlen<=_v.size());
 
   for (size_t i=0;i<vlen;i++) {
 
@@ -199,7 +186,7 @@ void basisReorderInPlace(std::vector<Field> &_v,std::vector<RealD>& sort_vals, s
 	if (idx[j]==i)
 	  break;
 
-      assert(idx[i] > i);     assert(j!=idx.size());      assert(idx[j]==i);
+      GRID_ASSERT(idx[i] > i);     GRID_ASSERT(j!=idx.size());      GRID_ASSERT(idx[j]==i);
 
       swap(_v[i],_v[idx[i]]); // should use vector move constructor, no data copy
       std::swap(sort_vals[i],sort_vals[idx[i]]);
@@ -237,7 +224,7 @@ void basisSortInPlace(std::vector<Field> & _v,std::vector<RealD>& sort_vals, boo
 template<class Field>
 void basisDeflate(const std::vector<Field> &_v,const std::vector<RealD>& eval,const Field& src_orig,Field& result) {
   result = Zero();
-  assert(_v.size()==eval.size());
+  GRID_ASSERT(_v.size()==eval.size());
   int N = (int)_v.size();
   for (int i=0;i<N;i++) {
     Field& tmp = _v[i];

Grid/lattice/Lattice_conformable.h

@@ -32,8 +32,8 @@ NAMESPACE_BEGIN(Grid);
 
 template<class obj1,class obj2> void conformable(const Lattice<obj1> &lhs,const Lattice<obj2> &rhs)
 {
-  assert(lhs.Grid() == rhs.Grid());
+  GRID_ASSERT(lhs.Grid() == rhs.Grid());
-  assert(lhs.Checkerboard() == rhs.Checkerboard());
+  GRID_ASSERT(lhs.Checkerboard() == rhs.Checkerboard());
 }
 
 NAMESPACE_END(Grid);

Grid/lattice/Lattice_matrix_reduction.h

@@ -42,7 +42,7 @@ static void sliceMaddMatrix (Lattice<vobj> &R,Eigen::MatrixXcd &aa,const Lattice
   //  Lattice<vobj> Xslice(SliceGrid);
   //  Lattice<vobj> Rslice(SliceGrid);
 
-  assert( FullGrid->_simd_layout[Orthog]==1);
+  GRID_ASSERT( FullGrid->_simd_layout[Orthog]==1);
 
   //FIXME package in a convenient iterator
   //Should loop over a plane orthogonal to direction "Orthog"
@@ -86,7 +86,7 @@ static void sliceMulMatrix (Lattice<vobj> &R,Eigen::MatrixXcd &aa,const Lattice<
   int Nblock = X.Grid()->GlobalDimensions()[Orthog];
 
   GridBase *FullGrid  = X.Grid();
-  assert( FullGrid->_simd_layout[Orthog]==1);
+  GRID_ASSERT( FullGrid->_simd_layout[Orthog]==1);
 
   //FIXME package in a convenient iterator
   //Should loop over a plane orthogonal to direction "Orthog"
@@ -140,7 +140,7 @@ static void sliceInnerProductMatrix(  Eigen::MatrixXcd &mat, const Lattice<vobj>
   
   mat = Eigen::MatrixXcd::Zero(Nblock,Nblock);
 
-  assert( FullGrid->_simd_layout[Orthog]==1);
+  GRID_ASSERT( FullGrid->_simd_layout[Orthog]==1);
   //  int nh =  FullGrid->_ndimension;
   //  int nl = SliceGrid->_ndimension;
   //  int nl = nh-1;

Grid/lattice/Lattice_peekpoke.h

@@ -98,8 +98,8 @@ void pokeSite(const sobj &s,Lattice<vobj> &l,const Coordinate &site){
 
   int Nsimd = grid->Nsimd();
 
-  assert( l.Checkerboard()== l.Grid()->CheckerBoard(site));
+  GRID_ASSERT( l.Checkerboard()== l.Grid()->CheckerBoard(site));
-  assert( sizeof(sobj)*Nsimd == sizeof(vobj));
+  GRID_ASSERT( sizeof(sobj)*Nsimd == sizeof(vobj));
 
   int rank,odx,idx;
   // Optional to broadcast from node 0.
@@ -135,7 +135,7 @@ void peekSite(sobj &s,const Lattice<vobj> &l,const Coordinate &site){
 
   int Nsimd = grid->Nsimd();
 
-  assert( l.Checkerboard() == l.Grid()->CheckerBoard(site));
+  GRID_ASSERT( l.Checkerboard() == l.Grid()->CheckerBoard(site));
 
   int rank,odx,idx;
   grid->GlobalCoorToRankIndex(rank,odx,idx,site);
@@ -159,14 +159,14 @@ template<class vobj,class sobj>
 inline void peekLocalSite(sobj &s,const LatticeView<vobj> &l,Coordinate &site)
 {
   GridBase *grid = l.getGrid();
-  assert(l.mode==CpuRead);
+  GRID_ASSERT(l.mode==CpuRead);
   typedef typename vobj::scalar_type scalar_type;
   typedef typename vobj::vector_type vector_type;
 
   int Nsimd = grid->Nsimd();
 
-  assert( l.Checkerboard()== grid->CheckerBoard(site));
+  //  GRID_ASSERT( l.Checkerboard()== grid->CheckerBoard(site));
-  assert( sizeof(sobj)*Nsimd == sizeof(vobj));
+  GRID_ASSERT( sizeof(sobj)*Nsimd == sizeof(vobj));
 
   static const int words=sizeof(vobj)/sizeof(vector_type);
   int odx,idx;
@@ -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>
@@ -195,15 +195,15 @@ template<class vobj,class sobj>
 inline void pokeLocalSite(const sobj &s,LatticeView<vobj> &l,Coordinate &site)
 {
   GridBase *grid=l.getGrid();
-  assert(l.mode==CpuWrite);
+  GRID_ASSERT(l.mode==CpuWrite);
 
   typedef typename vobj::scalar_type scalar_type;
   typedef typename vobj::vector_type vector_type;
 
   int Nsimd = grid->Nsimd();
 
-  assert( l.Checkerboard()== grid->CheckerBoard(site));
+  //  GRID_ASSERT( l.Checkerboard()== grid->CheckerBoard(site));
-  assert( sizeof(sobj)*Nsimd == sizeof(vobj));
+  GRID_ASSERT( sizeof(sobj)*Nsimd == sizeof(vobj));
 
   static const int words=sizeof(vobj)/sizeof(vector_type);
   int odx,idx;

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();
 
-  Vector<sobj> sumarray(nthread);
+  std::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();
 
-  Vector<sobj> sumarray(nthread);
+  std::vector<sobj> sumarray(nthread);
   for(int i=0;i<nthread;i++){
     sumarray[i]=Zero();
   }
@@ -290,23 +290,45 @@ template<class vobj>
 inline ComplexD innerProduct(const Lattice<vobj> &left,const Lattice<vobj> &right) {
   GridBase *grid = left.Grid();
 
+  bool ok;
 #ifdef GRID_SYCL
-  uint64_t csum=0;
+  //  uint64_t csum=0;
-  if ( FlightRecorder::LoggingMode != FlightRecorder::LoggingModeNone)
+  //  uint64_t csum2=0;
-  {
+  //  if ( FlightRecorder::LoggingMode != FlightRecorder::LoggingModeNone)
+  //  {
   // Hack
   // Fast integer xor checksum. Can also be used in comms now.
-    autoView(l_v,left,AcceleratorRead);
+  //    autoView(l_v,left,AcceleratorRead);
-    Integer words = left.Grid()->oSites()*sizeof(vobj)/sizeof(uint64_t);
+  //    Integer words = left.Grid()->oSites()*sizeof(vobj)/sizeof(uint64_t);
-    uint64_t *base= (uint64_t *)&l_v[0];
+  //    uint64_t *base= (uint64_t *)&l_v[0];
-    csum=svm_xor(base,words);
+  //    csum=svm_xor(base,words);
-  }
+  //    ok = FlightRecorder::CsumLog(csum);
-  FlightRecorder::CsumLog(csum);
+  //    if ( !ok ) {
+  //      csum2=svm_xor(base,words);
+  //      std::cerr<< " Bad CSUM " << std::hex<< csum << " recomputed as "<<csum2<<std::dec<<std::endl;
+  //    } else {
+  //      csum2=svm_xor(base,words);
+  //      std::cerr<< " ok CSUM " << std::hex<< csum << " recomputed as "<<csum2<<std::dec<<std::endl;
+  //    }
+  //    GRID_ASSERT(ok);
+  // }
 #endif
+  FlightRecorder::StepLog("rank inner product");
   ComplexD nrm = rankInnerProduct(left,right);
+  //  ComplexD nrmck=nrm;
   RealD local = real(nrm);
-  FlightRecorder::NormLog(real(nrm)); 
+  ok = FlightRecorder::NormLog(real(nrm));
-  grid->GlobalSum(nrm);
+  if ( !ok ) {
+    ComplexD nrm2 = rankInnerProduct(left,right);
+    RealD local2 = real(nrm2);
+    std::cerr<< " Bad NORM " << local << " recomputed as "<<local2<<std::endl;
+    GRID_ASSERT(ok);
+  }
+  FlightRecorder::StepLog("Start global sum");
+  grid->GlobalSumP2P(nrm);
+  //  grid->GlobalSum(nrm);
+  FlightRecorder::StepLog("Finished global sum");
+  //  std::cout << " norm "<< nrm << " p2p norm "<<nrmck<<std::endl;
   FlightRecorder::ReductionLog(local,real(nrm)); 
   return nrm;
 }
@@ -343,18 +365,6 @@ 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);
@@ -365,9 +375,13 @@ axpby_norm_fast(Lattice<vobj> &z,sobj a,sobj b,const Lattice<vobj> &x,const Latt
       coalescedWrite(inner_tmp_v[ss],innerProduct(tmp,tmp));
       coalescedWrite(z_v[ss],tmp);
   });
+  bool ok;
   nrm = real(TensorRemove(sumD(inner_tmp_v,sites)));
-#endif
+  ok = FlightRecorder::NormLog(real(nrm));
+  GRID_ASSERT(ok);
+  RealD local = real(nrm);
   grid->GlobalSum(nrm);
+  FlightRecorder::ReductionLog(local,real(nrm));
   return nrm; 
 }
  
@@ -377,7 +391,7 @@ innerProductNorm(ComplexD& ip, RealD &nrm, const Lattice<vobj> &left,const Latti
   conformable(left,right);
 
   typedef typename vobj::vector_typeD vector_type;
-  Vector<ComplexD> tmp(2);
+  std::vector<ComplexD> tmp(2);
 
   GridBase *grid = left.Grid();
 
@@ -387,8 +401,8 @@ 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;
-  Vector<inner_t> inner_tmp(sites);
+  deviceVector<inner_t> inner_tmp(sites);
-  Vector<norm_t>  norm_tmp(sites);
+  deviceVector<norm_t>  norm_tmp(sites);
   auto inner_tmp_v = &inner_tmp[0];
   auto norm_tmp_v = &norm_tmp[0];
   {
@@ -438,7 +452,9 @@ 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,std::vector<typename vobj::scalar_object> &result,int orthogdim)
+template<class vobj> inline void sliceSum(const Lattice<vobj> &Data,
+					  std::vector<typename vobj::scalar_object> &result,
+					  int orthogdim)
 {
   ///////////////////////////////////////////////////////
   // FIXME precision promoted summation
@@ -448,20 +464,20 @@ template<class vobj> inline void sliceSum(const Lattice<vobj> &Data,std::vector<
   typedef typename vobj::scalar_object sobj;
   typedef typename vobj::scalar_object::scalar_type scalar_type;
   GridBase  *grid = Data.Grid();
-  assert(grid!=NULL);
+  GRID_ASSERT(grid!=NULL);
 
   const int    Nd = grid->_ndimension;
   const int Nsimd = grid->Nsimd();
 
-  assert(orthogdim >= 0);
+  GRID_ASSERT(orthogdim >= 0);
-  assert(orthogdim < Nd);
+  GRID_ASSERT(orthogdim < Nd);
 
   int fd=grid->_fdimensions[orthogdim];
   int ld=grid->_ldimensions[orthogdim];
   int rd=grid->_rdimensions[orthogdim];
 
-  Vector<vobj> lvSum(rd); // will locally sum vectors first
+  std::vector<vobj> lvSum(rd); // will locally sum vectors first
-  Vector<sobj> lsSum(ld,Zero());                    // sum across these down to scalars
+  std::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 
@@ -509,6 +525,8 @@ template<class vobj> inline void sliceSum(const Lattice<vobj> &Data,std::vector<
   scalar_type * ptr = (scalar_type *) &result[0];
   int words = fd*sizeof(sobj)/sizeof(scalar_type);
   grid->GlobalSumVector(ptr, words);
+  //  std::cout << GridLogMessage << " sliceSum local"<<t_sum<<" us, host+mpi "<<t_rest<<std::endl;
+  
 }
 template<class vobj> inline
 std::vector<typename vobj::scalar_object> 
@@ -519,28 +537,41 @@ 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) 
 {
   typedef typename vobj::vector_type   vector_type;
   typedef typename vobj::scalar_type   scalar_type;
   GridBase  *grid = lhs.Grid();
-  assert(grid!=NULL);
+  GRID_ASSERT(grid!=NULL);
   conformable(grid,rhs.Grid());
 
   const int    Nd = grid->_ndimension;
   const int Nsimd = grid->Nsimd();
 
-  assert(orthogdim >= 0);
+  GRID_ASSERT(orthogdim >= 0);
-  assert(orthogdim < Nd);
+  GRID_ASSERT(orthogdim < Nd);
 
   int fd=grid->_fdimensions[orthogdim];
   int ld=grid->_ldimensions[orthogdim];
   int rd=grid->_rdimensions[orthogdim];
 
-  Vector<vector_type> lvSum(rd); // will locally sum vectors first
+  std::vector<vector_type> lvSum(rd); // will locally sum vectors first
-  Vector<scalar_type > lsSum(ld,scalar_type(0.0));                    // sum across these down to scalars
+  std::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
@@ -670,203 +701,96 @@ 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(0);
+  std::vector<int> latt_phys(NN-1);
-  std::vector<int> simd_phys(0);
+  Coordinate simd_phys;
-  std::vector<int>  mpi_phys(0);
+  std::vector<int>  mpi_phys(NN-1);
+  Coordinate checker_dim_mask(NN-1);
+  int checker_dim=-1;
 
+  int dd;
   for(int d=0;d<NN;d++){
     if( d!=Orthog ) { 
-      latt_phys.push_back(BlockSolverGrid->_fdimensions[d]);
+      latt_phys[dd]=BlockSolverGrid->_fdimensions[d];
-      simd_phys.push_back(BlockSolverGrid->_simd_layout[d]);
+      mpi_phys[dd] =BlockSolverGrid->_processors[d];
-      mpi_phys.push_back(BlockSolverGrid->_processors[d]);
+      checker_dim_mask[dd] = BlockSolverGrid->_checker_dim_mask[d];
+      if ( d == BlockSolverGrid->_checker_dim ) checker_dim = dd;
+      dd++;
     }
   }
-  return (GridBase *)new GridCartesian(latt_phys,simd_phys,mpi_phys); 
+  simd_phys=GridDefaultSimd(latt_phys.size(),nsimd);
+  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];
-  int Nblock = X.Grid()->GlobalDimensions()[Orthog];
+  for(int i=0;i<Nslice;i++){
-
+    ExtractSlice(Ys,Y,i,Orthog);
-  GridBase *FullGrid  = X.Grid();
+    ExtractSlice(Rs,R,i,Orthog);
-  //  GridBase *SliceGrid = makeSubSliceGrid(FullGrid,Orthog);
+    Rs=Ys;
-
+    for(int j=0;j<Nslice;j++){
-  //  Lattice<vobj> Xslice(SliceGrid);
+      ExtractSlice(Xs,X,j,Orthog);
-  //  Lattice<vobj> Rslice(SliceGrid);
+      Rs = Rs + Xs*(scale*aa(j,i));
-
-  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)
 {
-  typedef typename vobj::scalar_object sobj;
+  R=Zero();
-  typedef typename vobj::vector_type vector_type;
+  sliceMaddMatrix(R,aa,X,R,Orthog,scale);
-
-  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];
-  GridBase *FullGrid  = lhs.Grid();
+  mat = Eigen::MatrixXcd::Zero(Nslice,Nslice);
-  //  GridBase *SliceGrid = makeSubSliceGrid(FullGrid,Orthog);
+  for(int s=0;s<Nslice;s++){
-  
+    ExtractSlice(ls,lhs,s,Orthog);
-  int Nblock = FullGrid->GlobalDimensions()[Orthog];
+    for(int ss=0;ss<Nslice;ss++){
-  
+      ExtractSlice(rs,rhs,ss,Orthog);
-  //  Lattice<vobj> Lslice(SliceGrid);
+      mat(s,ss) = innerProduct(ls,rs);
-  //  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

@@ -208,28 +208,18 @@ inline typename vobj::scalar_objectD sumD_gpu_small(const vobj *lat, Integer osi
 
   Integer numThreads, numBlocks;
   int ok = getNumBlocksAndThreads(size, sizeof(sobj), numThreads, numBlocks);
-  assert(ok);
+  GRID_ASSERT(ok);
 
   Integer smemSize = numThreads * sizeof(sobj);
   // Move out of UVM
   // Turns out I had messed up the synchronise after move to compute stream
   // as running this on the default stream fools the synchronise
-#undef UVM_BLOCK_BUFFER  
+  deviceVector<sobj> buffer(numBlocks);
-#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;
 }
 
@@ -244,7 +234,7 @@ inline typename vobj::scalar_objectD sumD_gpu_large(const vobj *lat, Integer osi
   
   const int words = sizeof(vobj)/sizeof(vector);
 
-  Vector<vector> buffer(osites);
+  deviceVector<vector> buffer(osites);
   vector *dat = (vector *)lat;
   vector *buf = &buffer[0];
   iScalar<vector> *tbuf =(iScalar<vector> *)  &buffer[0];

Grid/lattice/Lattice_reduction_sycl.h

@@ -4,33 +4,28 @@ 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);
-  mysum[0] = identity;
+  sobj ret; zeroit(ret);
-  sobj ret ; 
-
   Integer nsimd= vobj::Nsimd();
-
+  { 
-  const cl::sycl::property_list PropList ({ cl::sycl::property::reduction::initialize_to_identity() });
+    sycl::buffer<sobj, 1> abuff(&ret, {1});
-  theGridAccelerator->submit([&](cl::sycl::handler &cgh) {
+    theGridAccelerator->submit([&](sycl::handler &cgh) {
-    auto Reduction = cl::sycl::reduction(mysum_p,identity,std::plus<>(),PropList);
+      auto Reduction = sycl::reduction(abuff,cgh,identity,std::plus<>());
-     cgh.parallel_for(cl::sycl::range<1>{osites},
+      cgh.parallel_for(sycl::range<1>{osites},
                       Reduction,
-		      [=] (cl::sycl::id<1> item, auto &sum) {
+                      [=] (sycl::id<1> item, auto &sum) {
                         auto osite   = item[0];
                         sum +=Reduce(lat[osite]);
                       });
     });
-  theGridAccelerator->wait();
+  }
-  ret = mysum[0];
-  //  free(mysum,*theGridAccelerator);
   sobjD dret; convertType(dret,ret);
   return dret;
 }
@@ -76,59 +71,41 @@ 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;
-  d_sum[0] = identity;
+  Word ret = 0;
-  const cl::sycl::property_list PropList ({ cl::sycl::property::reduction::initialize_to_identity() });
+  { 
-  theGridAccelerator->submit([&](cl::sycl::handler &cgh) {
+    sycl::buffer<Word, 1> abuff(&ret, {1});
-    auto Reduction = cl::sycl::reduction(d_sum_p,identity,std::bit_xor<>(),PropList);
+    theGridAccelerator->submit([&](sycl::handler &cgh) {
-     cgh.parallel_for(cl::sycl::range<1>{L},
+      auto Reduction = sycl::reduction(abuff,cgh,identity,std::bit_xor<>());
+      cgh.parallel_for(sycl::range<1>{L},
                       Reduction,
-		      [=] (cl::sycl::id<1> index, auto &sum) {
+                      [=] (sycl::id<1> index, auto &sum) {
                         sum ^=vec[index];
                       });
     });
+  }
+  theGridAccelerator->wait();
+  return ret;
+}
+template<class Word> Word checksum_gpu(Word *vec,uint64_t L)
+{
+  Word identity;  identity=0;
+  Word ret = 0;
+  { 
+    sycl::buffer<Word, 1> abuff(&ret, {1});
+    theGridAccelerator->submit([&](sycl::handler &cgh) {
+      auto Reduction = sycl::reduction(abuff,cgh,identity,std::bit_xor<>());
+      cgh.parallel_for(sycl::range<1>{L},
+                       Reduction,
+                       [=] (sycl::id<1> index, auto &sum) {
+			 auto l = index % 61;
+                         sum ^= vec[index]<<l | vec[index]>>(64-l);
+                       });
+    });
+  }
   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_rng.h

@@ -53,10 +53,10 @@ inline int RNGfillable(GridBase *coarse,GridBase *fine)
 
   // trivially extended in higher dims, with locality guaranteeing RNG state is local to node
   int lowerdims   = fine->_ndimension - coarse->_ndimension;
-  assert(lowerdims >= 0);
+  GRID_ASSERT(lowerdims >= 0);
   for(int d=0;d<lowerdims;d++){
-    assert(fine->_simd_layout[d]==1);
+    GRID_ASSERT(fine->_simd_layout[d]==1);
-    assert(fine->_processors[d]==1);
+    GRID_ASSERT(fine->_processors[d]==1);
   }
 
   int multiplicity=1;
@@ -66,9 +66,9 @@ inline int RNGfillable(GridBase *coarse,GridBase *fine)
   // local and global volumes subdivide cleanly after SIMDization
   for(int d=0;d<rngdims;d++){
     int fd= d+lowerdims;
-    assert(coarse->_processors[d]  == fine->_processors[fd]);
+    GRID_ASSERT(coarse->_processors[d]  == fine->_processors[fd]);
-    assert(coarse->_simd_layout[d] == fine->_simd_layout[fd]);
+    GRID_ASSERT(coarse->_simd_layout[d] == fine->_simd_layout[fd]);
-    assert(((fine->_rdimensions[fd] / coarse->_rdimensions[d])* coarse->_rdimensions[d])==fine->_rdimensions[fd]); 
+    GRID_ASSERT(((fine->_rdimensions[fd] / coarse->_rdimensions[d])* coarse->_rdimensions[d])==fine->_rdimensions[fd]); 
 
     multiplicity = multiplicity *fine->_rdimensions[fd] / coarse->_rdimensions[d]; 
   }
@@ -83,18 +83,18 @@ inline int RNGfillable_general(GridBase *coarse,GridBase *fine)
   int rngdims = coarse->_ndimension;
     
   // trivially extended in higher dims, with locality guaranteeing RNG state is local to node
-  int lowerdims   = fine->_ndimension - coarse->_ndimension;  assert(lowerdims >= 0);
+  int lowerdims   = fine->_ndimension - coarse->_ndimension;  GRID_ASSERT(lowerdims >= 0);
   // assumes that the higher dimensions are not using more processors
   // all further divisions are local
-  for(int d=0;d<lowerdims;d++) assert(fine->_processors[d]==1);
+  for(int d=0;d<lowerdims;d++) GRID_ASSERT(fine->_processors[d]==1);
-  for(int d=0;d<rngdims;d++) assert(coarse->_processors[d] == fine->_processors[d+lowerdims]);
+  for(int d=0;d<rngdims;d++) GRID_ASSERT(coarse->_processors[d] == fine->_processors[d+lowerdims]);
 
   // then divide the number of local sites
   // check that the total number of sims agree, meanse the iSites are the same
-  assert(fine->Nsimd() == coarse->Nsimd());
+  GRID_ASSERT(fine->Nsimd() == coarse->Nsimd());
 
   // check that the two grids divide cleanly
-  assert( (fine->lSites() / coarse->lSites() ) * coarse->lSites() == fine->lSites() );
+  GRID_ASSERT( (fine->lSites() / coarse->lSites() ) * coarse->lSites() == fine->lSites() );
 
   return fine->lSites() / coarse->lSites();
 }
@@ -177,7 +177,7 @@ public:
 
     skip = skip<<shift;
 
-    assert((skip >> shift)==site); // check for overflow
+    GRID_ASSERT((skip >> shift)==site); // check for overflow
 
     eng.discard(skip);
 #else
@@ -218,7 +218,7 @@ public:
     GetState(saved,_generators[gen]);
   }
   void SetState(std::vector<RngStateType> & saved,RngEngine &eng){
-    assert(saved.size()==RngStateCount);
+    GRID_ASSERT(saved.size()==RngStateCount);
     std::stringstream ss;
     for(int i=0;i<RngStateCount;i++){
       ss<< saved[i]<<" ";

Grid/lattice/Lattice_slicesum_core.h

@@ -21,9 +21,18 @@ NAMESPACE_BEGIN(Grid);
 
 
 #if defined(GRID_CUDA) || defined(GRID_HIP)
-template<class vobj> inline void sliceSumReduction_cub_small(const vobj *Data, Vector<vobj> &lvSum, const int rd, const int e1, const int e2, const int stride, const int ostride, const int Nsimd) {
+template<class vobj>
+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;
-  commVector<vobj> reduction_buffer(rd*subvol_size);
+  deviceVector<vobj> reduction_buffer(rd*subvol_size);
   auto rb_p = &reduction_buffer[0];
   vobj zero_init;
   zeroit(zero_init);
@@ -46,7 +55,7 @@ template<class vobj> inline void sliceSumReduction_cub_small(const vobj *Data, V
   d_offsets = static_cast<int*>(acceleratorAllocDevice((rd+1)*sizeof(int)));
   
   //copy offsets to device
-  acceleratorCopyToDeviceAsync(&offsets[0],d_offsets,sizeof(int)*(rd+1),computeStream);
+  acceleratorCopyToDeviceAsynch(&offsets[0],d_offsets,sizeof(int)*(rd+1),computeStream);
   
   
   gpuError_t gpuErr = gpucub::DeviceSegmentedReduce::Reduce(temp_storage_array, temp_storage_bytes, rb_p,d_out, rd, d_offsets, d_offsets+1, ::gpucub::Sum(), zero_init, computeStream);
@@ -79,7 +88,7 @@ template<class vobj> inline void sliceSumReduction_cub_small(const vobj *Data, V
     exit(EXIT_FAILURE);
   }
   
-  acceleratorCopyFromDeviceAsync(d_out,&lvSum[0],rd*sizeof(vobj),computeStream);
+  acceleratorCopyFromDeviceAsynch(d_out,&lvSum[0],rd*sizeof(vobj),computeStream);
   
   //sync after copy
   accelerator_barrier();
@@ -94,7 +103,15 @@ template<class vobj> inline void sliceSumReduction_cub_small(const vobj *Data, V
 
 
 #if defined(GRID_SYCL)
-template<class vobj> inline void sliceSumReduction_sycl_small(const vobj *Data, Vector <vobj> &lvSum, const int  &rd, const int &e1, const int &e2, const int &stride, const int &ostride, const int &Nsimd)
+template<class vobj>
+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;
 
@@ -105,7 +122,7 @@ template<class vobj> inline void sliceSumReduction_sycl_small(const vobj *Data,
     mysum[r] = vobj_zero; 
   }
 
-  commVector<vobj> reduction_buffer(rd*subvol_size);    
+  deviceVector<vobj> reduction_buffer(rd*subvol_size);    
 
   auto rb_p = &reduction_buffer[0];
 
@@ -124,11 +141,11 @@ template<class vobj> inline void sliceSumReduction_sycl_small(const vobj *Data,
   });
 
   for (int r = 0; r < rd; r++) {
-      theGridAccelerator->submit([&](cl::sycl::handler &cgh) {
+      theGridAccelerator->submit([&](sycl::handler &cgh) {
-          auto Reduction = cl::sycl::reduction(&mysum[r],std::plus<>());
+          auto Reduction = sycl::reduction(&mysum[r],std::plus<>());
-          cgh.parallel_for(cl::sycl::range<1>{subvol_size},
+          cgh.parallel_for(sycl::range<1>{subvol_size},
           Reduction,
-          [=](cl::sycl::id<1> item, auto &sum) {
+          [=](sycl::id<1> item, auto &sum) {
               auto s = item[0];
               sum += rb_p[r*subvol_size+s];
           });
@@ -144,14 +161,23 @@ template<class vobj> inline void sliceSumReduction_sycl_small(const vobj *Data,
 }
 #endif
 
-template<class vobj> inline void sliceSumReduction_large(const vobj *Data, Vector<vobj> &lvSum, const int rd, const int e1, const int e2, const int stride, const int ostride, const int Nsimd) {
+template<class vobj>
+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;
-  commVector<vector>buffer(osites);
+  deviceVector<vector>buffer(osites);
   vector *dat = (vector *)Data;
   vector *buf = &buffer[0];
-  Vector<vector> lvSum_small(rd);
+  std::vector<vector> lvSum_small(rd);
   vector *lvSum_ptr = (vector *)&lvSum[0];
 
   for (int w = 0; w < words; w++) {
@@ -168,13 +194,18 @@ template<class vobj> inline void sliceSumReduction_large(const vobj *Data, Vecto
     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,
-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 rd,
+				  const int e1,
+				  const int e2,
+				  const int stride,
+				  const int ostride,
+				  const int Nsimd)
 {
   autoView(Data_v, Data, AcceleratorRead); //reduction libraries cannot deal with large vobjs so we split into small/large case.
     if constexpr (sizeof(vobj) <= 256) { 
@@ -192,7 +223,15 @@ template<class vobj> inline void sliceSumReduction_gpu(const Lattice<vobj> &Data
 }
 
 
-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)
+template<class vobj>
+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
@@ -208,15 +247,19 @@ template<class vobj> inline void sliceSumReduction_cpu(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) 
+template<class vobj> inline void sliceSumReduction(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) 
 {
 #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

@@ -31,32 +31,61 @@ NAMESPACE_BEGIN(Grid);
 
 inline void subdivides(GridBase *coarse,GridBase *fine)
 {
-  assert(coarse->_ndimension == fine->_ndimension);
+  GRID_ASSERT(coarse->_ndimension == fine->_ndimension);
 
   int _ndimension = coarse->_ndimension;
 
   // local and global volumes subdivide cleanly after SIMDization
   for(int d=0;d<_ndimension;d++){
-    assert(coarse->_processors[d]  == fine->_processors[d]);
+    GRID_ASSERT(coarse->_processors[d]  == fine->_processors[d]);
-    assert(coarse->_simd_layout[d] == fine->_simd_layout[d]);
+    GRID_ASSERT(coarse->_simd_layout[d] == fine->_simd_layout[d]);
-    assert((fine->_rdimensions[d] / coarse->_rdimensions[d])* coarse->_rdimensions[d]==fine->_rdimensions[d]); 
+    GRID_ASSERT((fine->_rdimensions[d] / coarse->_rdimensions[d])* coarse->_rdimensions[d]==fine->_rdimensions[d]); 
   }
 }
 
+ 
 ////////////////////////////////////////////////////////////////////////////////////////////
 // remove and insert a half checkerboard
 ////////////////////////////////////////////////////////////////////////////////////////////
-
 template<class vobj> inline void pickCheckerboard(int cb,Lattice<vobj> &half,const Lattice<vobj> &full)
 {
-  acceleratorPickCheckerboard(cb,half,full);
+  half.Checkerboard() = cb;
+
+  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)
 {
-  acceleratorSetCheckerboard(full,half);
+  int cb = half.Checkerboard();
+  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 dummy=0)
+template<class vobj> inline void acceleratorPickCheckerboard(int cb,Lattice<vobj> &half,const Lattice<vobj> &full, int checker_dim_half=0)
 {
   half.Checkerboard() = cb;
   autoView(half_v, half, AcceleratorWrite);
@@ -66,7 +95,6 @@ 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;
@@ -91,7 +119,7 @@ template<class vobj> inline void acceleratorPickCheckerboard(int cb,Lattice<vobj
     }
   });
 }
-template<class vobj> inline void acceleratorSetCheckerboard(Lattice<vobj> &full,const Lattice<vobj> &half, int dummy=0)
+template<class vobj> inline void acceleratorSetCheckerboard(Lattice<vobj> &full,const Lattice<vobj> &half, int checker_dim_half=0)
 {
   int cb = half.Checkerboard();
   autoView(half_v , half, AcceleratorRead);
@@ -101,7 +129,6 @@ 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;
@@ -282,7 +309,7 @@ inline void batchBlockProject(std::vector<Lattice<iVector<CComplex,nbasis>>> &co
                                const VLattice &Basis)
 {
   int NBatch = fineData.size();
-  assert(coarseData.size() == NBatch);
+  GRID_ASSERT(coarseData.size() == NBatch);
 
   GridBase * fine  = fineData[0].Grid();
   GridBase * coarse= coarseData[0].Grid();
@@ -317,7 +344,7 @@ template<class vobj,class vobj2,class CComplex>
   GridBase * coarse= coarseA.Grid();
 
   fineZ.Checkerboard()=fineX.Checkerboard();
-  assert(fineX.Checkerboard()==fineY.Checkerboard());
+  GRID_ASSERT(fineX.Checkerboard()==fineY.Checkerboard());
   subdivides(coarse,fine); // require they map
   conformable(fineX,fineY);
   conformable(fineX,fineZ);
@@ -329,7 +356,7 @@ template<class vobj,class vobj2,class CComplex>
   // FIXME merge with subdivide checking routine as this is redundant
   for(int d=0 ; d<_ndimension;d++){
     block_r[d] = fine->_rdimensions[d] / coarse->_rdimensions[d];
-    assert(block_r[d]*coarse->_rdimensions[d]==fine->_rdimensions[d]);
+    GRID_ASSERT(block_r[d]*coarse->_rdimensions[d]==fine->_rdimensions[d]);
   }
 
   autoView( fineZ_  , fineZ, AcceleratorWrite);
@@ -586,7 +613,7 @@ inline void blockPromote(const Lattice<iVector<CComplex,nbasis > > &coarseData,
   int  _ndimension = coarse->_ndimension;
 
   // checks
-  assert( nbasis == Basis.size() );
+  GRID_ASSERT( nbasis == Basis.size() );
   subdivides(coarse,fine); 
   for(int i=0;i<nbasis;i++){
     conformable(Basis[i].Grid(),fine);
@@ -660,7 +687,7 @@ inline void batchBlockPromote(const std::vector<Lattice<iVector<CComplex,nbasis>
                                const VLattice &Basis)
 {
   int NBatch = coarseData.size();
-  assert(fineData.size() == NBatch);
+  GRID_ASSERT(fineData.size() == NBatch);
 
   GridBase * fine   = fineData[0].Grid();
   GridBase * coarse = coarseData[0].Grid();
@@ -688,12 +715,12 @@ void localConvert(const Lattice<vobj> &in,Lattice<vvobj> &out)
   int ni = ig->_ndimension;
   int no = og->_ndimension;
 
-  assert(ni == no);
+  GRID_ASSERT(ni == no);
 
   for(int d=0;d<no;d++){
-    assert(ig->_processors[d]  == og->_processors[d]);
+    GRID_ASSERT(ig->_processors[d]  == og->_processors[d]);
-    assert(ig->_ldimensions[d] == og->_ldimensions[d]);
+    GRID_ASSERT(ig->_ldimensions[d] == og->_ldimensions[d]);
-    assert(ig->lSites() == og->lSites());
+    GRID_ASSERT(ig->lSites() == og->lSites());
   }
 
   autoView(in_v,in,CpuRead);
@@ -725,16 +752,16 @@ void localCopyRegion(const Lattice<vobj> &From,Lattice<vobj> & To,Coordinate Fro
 
   GridBase *Fg = From.Grid();
   GridBase *Tg = To.Grid();
-  assert(!Fg->_isCheckerBoarded);
+  GRID_ASSERT(!Fg->_isCheckerBoarded);
-  assert(!Tg->_isCheckerBoarded);
+  GRID_ASSERT(!Tg->_isCheckerBoarded);
   int Nsimd = Fg->Nsimd();
   int nF = Fg->_ndimension;
   int nT = Tg->_ndimension;
   int nd = nF;
-  assert(nF == nT);
+  GRID_ASSERT(nF == nT);
 
   for(int d=0;d<nd;d++){
-    assert(Fg->_processors[d]  == Tg->_processors[d]);
+    GRID_ASSERT(Fg->_processors[d]  == Tg->_processors[d]);
   }
 
   ///////////////////////////////////////////////////////////
@@ -794,12 +821,12 @@ void InsertSliceFast(const Lattice<vobj> &From,Lattice<vobj> & To,int slice, int
   //////////////////////////////////////////////////////////////////////////////////////////
   GridBase *Fg = From.Grid();
   GridBase *Tg = To.Grid();
-  assert(!Fg->_isCheckerBoarded);
+  GRID_ASSERT(!Fg->_isCheckerBoarded);
-  assert(!Tg->_isCheckerBoarded);
+  GRID_ASSERT(!Tg->_isCheckerBoarded);
   int Nsimd = Fg->Nsimd();
   int nF = Fg->_ndimension;
   int nT = Tg->_ndimension;
-  assert(nF+1 == nT);
+  GRID_ASSERT(nF+1 == nT);
 
   ///////////////////////////////////////////////////////////
   // do the index calc on the GPU
@@ -863,12 +890,12 @@ void ExtractSliceFast(Lattice<vobj> &To,const Lattice<vobj> & From,int slice, in
   //////////////////////////////////////////////////////////////////////////////////////////
   GridBase *Fg = From.Grid();
   GridBase *Tg = To.Grid();
-  assert(!Fg->_isCheckerBoarded);
+  GRID_ASSERT(!Fg->_isCheckerBoarded);
-  assert(!Tg->_isCheckerBoarded);
+  GRID_ASSERT(!Tg->_isCheckerBoarded);
   int Nsimd = Fg->Nsimd();
   int nF = Fg->_ndimension;
   int nT = Tg->_ndimension;
-  assert(nT+1 == nF);
+  GRID_ASSERT(nT+1 == nF);
 
   ///////////////////////////////////////////////////////////
   // do the index calc on the GPU
@@ -928,16 +955,16 @@ void InsertSlice(const Lattice<vobj> &lowDim,Lattice<vobj> & higherDim,int slice
   int nl = lg->_ndimension;
   int nh = hg->_ndimension;
 
-  assert(nl+1 == nh);
+  GRID_ASSERT(nl+1 == nh);
-  assert(orthog<nh);
+  GRID_ASSERT(orthog<nh);
-  assert(orthog>=0);
+  GRID_ASSERT(orthog>=0);
-  assert(hg->_processors[orthog]==1);
+  GRID_ASSERT(hg->_processors[orthog]==1);
 
   int dl; dl = 0;
   for(int d=0;d<nh;d++){
     if ( d != orthog) {
-      assert(lg->_processors[dl]  == hg->_processors[d]);
+      GRID_ASSERT(lg->_processors[dl]  == hg->_processors[d]);
-      assert(lg->_ldimensions[dl] == hg->_ldimensions[d]);
+      GRID_ASSERT(lg->_ldimensions[dl] == hg->_ldimensions[d]);
       dl++;
     }
   }
@@ -954,8 +981,14 @@ 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);
@@ -972,16 +1005,17 @@ void ExtractSlice(Lattice<vobj> &lowDim,const Lattice<vobj> & higherDim,int slic
   int nl = lg->_ndimension;
   int nh = hg->_ndimension;
 
-  assert(nl+1 == nh);
+  GRID_ASSERT(nl+1 == nh);
-  assert(orthog<nh);
+  GRID_ASSERT(orthog<nh);
-  assert(orthog>=0);
+  GRID_ASSERT(orthog>=0);
-  assert(hg->_processors[orthog]==1);
+  GRID_ASSERT(hg->_processors[orthog]==1);
+  lowDim.Checkerboard() = higherDim.Checkerboard();
 
   int dl; dl = 0;
   for(int d=0;d<nh;d++){
     if ( d != orthog) {
-      assert(lg->_processors[dl]  == hg->_processors[d]);
+      GRID_ASSERT(lg->_processors[dl]  == hg->_processors[d]);
-      assert(lg->_ldimensions[dl] == hg->_ldimensions[d]);
+      GRID_ASSERT(lg->_ldimensions[dl] == hg->_ldimensions[d]);
       dl++;
     }
   }
@@ -993,11 +1027,16 @@ void ExtractSlice(Lattice<vobj> &lowDim,const Lattice<vobj> & higherDim,int slic
     Coordinate lcoor(nl);
     Coordinate hcoor(nh);
     lg->LocalIndexToLocalCoor(idx,lcoor);
-    int ddl=0;
     hcoor[orthog] = slice;
+    int ddl=0;
     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);
@@ -1017,14 +1056,14 @@ void InsertSliceLocal(const Lattice<vobj> &lowDim, Lattice<vobj> & higherDim,int
   int nl = lg->_ndimension;
   int nh = hg->_ndimension;
 
-  assert(nl == nh);
+  GRID_ASSERT(nl == nh);
-  assert(orthog<nh);
+  GRID_ASSERT(orthog<nh);
-  assert(orthog>=0);
+  GRID_ASSERT(orthog>=0);
 
   for(int d=0;d<nh;d++){
     if ( d!=orthog ) {
-      assert(lg->_processors[d]  == hg->_processors[d]);
+      GRID_ASSERT(lg->_processors[d]  == hg->_processors[d]);
-      assert(lg->_ldimensions[d] == hg->_ldimensions[d]);
+      GRID_ASSERT(lg->_ldimensions[d] == hg->_ldimensions[d]);
     }
   }
   Coordinate sz = lg->_ldimensions;
@@ -1054,7 +1093,7 @@ void Replicate(const Lattice<vobj> &coarse,Lattice<vobj> & fine)
 
   subdivides(cg,fg); 
 
-  assert(cg->_ndimension==fg->_ndimension);
+  GRID_ASSERT(cg->_ndimension==fg->_ndimension);
 
   Coordinate ratio(cg->_ndimension);
 
@@ -1118,7 +1157,7 @@ unvectorizeToLexOrdArray(std::vector<sobj> &out, const Lattice<vobj> &in)
 
       int lex;
       Lexicographic::IndexFromCoor(lcoor, lex, in_grid->_ldimensions);
-      assert(lex < out.size());
+      GRID_ASSERT(lex < out.size());
       out_ptrs[lane] = &out[lex];
     }
     
@@ -1182,7 +1221,7 @@ vectorizeFromLexOrdArray( std::vector<sobj> &in, Lattice<vobj> &out)
   typedef typename vobj::vector_type vtype;
   
   GridBase* grid = out.Grid();
-  assert(in.size()==grid->lSites());
+  GRID_ASSERT(in.size()==grid->lSites());
   
   const int ndim     = grid->Nd();
   constexpr int nsimd    = vtype::Nsimd();
@@ -1229,7 +1268,7 @@ vectorizeFromRevLexOrdArray( std::vector<sobj> &in, Lattice<vobj> &out)
   typedef typename vobj::vector_type vtype;
   
   GridBase* grid = out._grid;
-  assert(in.size()==grid->lSites());
+  GRID_ASSERT(in.size()==grid->lSites());
   
   int ndim     = grid->Nd();
   int nsimd    = vtype::Nsimd();
@@ -1290,9 +1329,9 @@ void precisionChangeFast(Lattice<VobjOut> &out, const Lattice<VobjIn> &in)
 template<class VobjOut, class VobjIn>
 void precisionChangeOrig(Lattice<VobjOut> &out, const Lattice<VobjIn> &in)
 {
-  assert(out.Grid()->Nd() == in.Grid()->Nd());
+  GRID_ASSERT(out.Grid()->Nd() == in.Grid()->Nd());
   for(int d=0;d<out.Grid()->Nd();d++){
-    assert(out.Grid()->FullDimensions()[d] == in.Grid()->FullDimensions()[d]);
+    GRID_ASSERT(out.Grid()->FullDimensions()[d] == in.Grid()->FullDimensions()[d]);
   }
   out.Checkerboard() = in.Checkerboard();
   GridBase *in_grid=in.Grid();
@@ -1343,9 +1382,9 @@ class precisionChangeWorkspace{
 public:
   precisionChangeWorkspace(GridBase *out_grid, GridBase *in_grid): _out_grid(out_grid), _in_grid(in_grid){
     //Build a map between the sites and lanes of the output field and the input field as we cannot use the Grids on the device
-    assert(out_grid->Nd() == in_grid->Nd());
+    GRID_ASSERT(out_grid->Nd() == in_grid->Nd());
     for(int d=0;d<out_grid->Nd();d++){
-      assert(out_grid->FullDimensions()[d] == in_grid->FullDimensions()[d]);
+      GRID_ASSERT(out_grid->FullDimensions()[d] == in_grid->FullDimensions()[d]);
     }
     int Nsimd_out = out_grid->Nsimd();
 
@@ -1510,7 +1549,7 @@ void Grid_split(std::vector<Lattice<Vobj> > & full,Lattice<Vobj>   & split)
 
   int full_vecs   = full.size();
 
-  assert(full_vecs>=1);
+  GRID_ASSERT(full_vecs>=1);
 
   GridBase * full_grid = full[0].Grid();
   GridBase *split_grid = split.Grid();
@@ -1528,18 +1567,18 @@ void Grid_split(std::vector<Lattice<Vobj> > & full,Lattice<Vobj>   & split)
   //////////////////////////////
   // Checks
   //////////////////////////////
-  assert(full_grid->_ndimension==split_grid->_ndimension);
+  GRID_ASSERT(full_grid->_ndimension==split_grid->_ndimension);
   for(int n=0;n<full_vecs;n++){
-    assert(full[n].Checkerboard() == cb);
+    GRID_ASSERT(full[n].Checkerboard() == cb);
     for(int d=0;d<ndim;d++){
-      assert(full[n].Grid()->_gdimensions[d]==split.Grid()->_gdimensions[d]);
+      GRID_ASSERT(full[n].Grid()->_gdimensions[d]==split.Grid()->_gdimensions[d]);
-      assert(full[n].Grid()->_fdimensions[d]==split.Grid()->_fdimensions[d]);
+      GRID_ASSERT(full[n].Grid()->_fdimensions[d]==split.Grid()->_fdimensions[d]);
     }
   }
 
   int   nvector   =full_nproc/split_nproc; 
-  assert(nvector*split_nproc==full_nproc);
+  GRID_ASSERT(nvector*split_nproc==full_nproc);
-  assert(nvector == full_vecs);
+  GRID_ASSERT(nvector == full_vecs);
 
   Coordinate ratio(ndim);
   for(int d=0;d<ndim;d++){
@@ -1583,7 +1622,7 @@ void Grid_split(std::vector<Lattice<Vobj> > & full,Lattice<Vobj>   & split)
 
       int fvol   = lsites;
       
-      int chunk  = (nvec*fvol)/sP;          assert(chunk*sP == nvec*fvol);
+      int chunk  = (nvec*fvol)/sP;          GRID_ASSERT(chunk*sP == nvec*fvol);
 
       // Loop over reordered data post A2A
       thread_for(c, chunk, {
@@ -1636,7 +1675,7 @@ void Grid_unsplit(std::vector<Lattice<Vobj> > & full,Lattice<Vobj>   & split)
 
   int full_vecs   = full.size();
 
-  assert(full_vecs>=1);
+  GRID_ASSERT(full_vecs>=1);
 
   GridBase * full_grid = full[0].Grid();
   GridBase *split_grid = split.Grid();
@@ -1654,18 +1693,18 @@ void Grid_unsplit(std::vector<Lattice<Vobj> > & full,Lattice<Vobj>   & split)
   //////////////////////////////
   // Checks
   //////////////////////////////
-  assert(full_grid->_ndimension==split_grid->_ndimension);
+  GRID_ASSERT(full_grid->_ndimension==split_grid->_ndimension);
   for(int n=0;n<full_vecs;n++){
-    assert(full[n].Checkerboard() == cb);
+    GRID_ASSERT(full[n].Checkerboard() == cb);
     for(int d=0;d<ndim;d++){
-      assert(full[n].Grid()->_gdimensions[d]==split.Grid()->_gdimensions[d]);
+      GRID_ASSERT(full[n].Grid()->_gdimensions[d]==split.Grid()->_gdimensions[d]);
-      assert(full[n].Grid()->_fdimensions[d]==split.Grid()->_fdimensions[d]);
+      GRID_ASSERT(full[n].Grid()->_fdimensions[d]==split.Grid()->_fdimensions[d]);
     }
   }
 
   int   nvector   =full_nproc/split_nproc; 
-  assert(nvector*split_nproc==full_nproc);
+  GRID_ASSERT(nvector*split_nproc==full_nproc);
-  assert(nvector == full_vecs);
+  GRID_ASSERT(nvector == full_vecs);
 
   Coordinate ratio(ndim);
   for(int d=0;d<ndim;d++){
@@ -1701,7 +1740,7 @@ void Grid_unsplit(std::vector<Lattice<Vobj> > & full,Lattice<Vobj>   & split)
       auto lsites= rsites/M;                // Decreases rsites by M
       
       int fvol   = lsites;
-      int chunk  = (nvec*fvol)/sP;          assert(chunk*sP == nvec*fvol);
+      int chunk  = (nvec*fvol)/sP;          GRID_ASSERT(chunk*sP == nvec*fvol);
 	
       {
 	// Loop over reordered data post A2A

Grid/lattice/Lattice_view.h

@@ -106,6 +106,47 @@ public:
   }
 
 };
+
+#ifdef GRID_LOG_VIEWS
+// Little autoscope assister
+template<class View> 
+class ViewCloser
+{
+  View v;  // Take a copy of view and call view close when I go out of scope automatically
+  const char* filename; int line, mode;
+public:
+  ViewCloser(View &_v, const char* _filename, int _line, int _mode) :
+    v(_v), filename(_filename), line(_line), mode(_mode) {
+    
+    switch (mode){
+    case AcceleratorRead:
+    case AcceleratorWrite:
+    case CpuRead:
+    case CpuWrite:
+      ViewLogger::LogOpen(filename, line, 1, mode, &v[0], v.size() * sizeof(v[0]));
+      break;
+    } 
+    
+  };
+  ~ViewCloser() {
+    
+    switch (mode) {
+    case AcceleratorWriteDiscard:
+    case AcceleratorWrite:
+    case CpuWrite:
+      ViewLogger::LogClose(filename, line, -1, mode, &v[0], v.size() * sizeof(v[0]));
+      break;
+    }
+    
+    v.ViewClose();
+  }
+};
+
+#define autoView(l_v,l,mode)				\
+	  auto l_v = l.View(mode);			\
+	  ViewCloser<decltype(l_v)> _autoView##l_v(l_v,__FILE__,__LINE__,mode);
+
+#else
 // Little autoscope assister
 template<class View> 
 class ViewCloser
@@ -119,6 +160,7 @@ class ViewCloser
 #define autoView(l_v,l,mode)				\
 	  auto l_v = l.View(mode);			\
 	  ViewCloser<decltype(l_v)> _autoView##l_v(l_v);
+#endif
 
 /////////////////////////////////////////////////////////////////////////////////////////
 // Lattice expression types used by ET to assemble the AST

Grid/lattice/PaddedCell.h

@@ -54,7 +54,7 @@ struct CshiftImplGauge: public CshiftImplBase<typename Gimpl::GaugeLinkField::ve
  *
  */
 
-template<class vobj> inline void ScatterSlice(const cshiftVector<vobj> &buf,
+template<class vobj> inline void ScatterSlice(const deviceVector<vobj> &buf,
 					      Lattice<vobj> &lat,
 					      int x,
 					      int dim,
@@ -82,10 +82,10 @@ template<class vobj> inline void ScatterSlice(const cshiftVector<vobj> &buf,
 
   int rNsimd = 1; for(int d=0;d<Nd;d++) rNsimd*=rsimd[d];
   int rNsimda= Nsimd/simd[dim]; // should be equal
-  assert(rNsimda==rNsimd);
+  GRID_ASSERT(rNsimda==rNsimd);
   int face_ovol=block*nblock;
 
-  //  assert(buf.size()==face_ovol*rNsimd);
+  //  GRID_ASSERT(buf.size()==face_ovol*rNsimd);
 
   /*This will work GPU ONLY unless rNsimd is put in the lexico index*/
   //Let's make it work on GPU and then make a special accelerator_for that
@@ -140,7 +140,7 @@ template<class vobj> inline void ScatterSlice(const cshiftVector<vobj> &buf,
   });
 }
 
-template<class vobj> inline void GatherSlice(cshiftVector<vobj> &buf,
+template<class vobj> inline void GatherSlice(deviceVector<vobj> &buf,
 					     const Lattice<vobj> &lat,
 					     int x,
 					     int dim,
@@ -172,7 +172,7 @@ template<class vobj> inline void GatherSlice(cshiftVector<vobj> &buf,
   
   int face_ovol=block*nblock;
 
-  //  assert(buf.size()==face_ovol*rNsimd);
+  //  GRID_ASSERT(buf.size()==face_ovol*rNsimd);
 
   /*This will work GPU ONLY unless rNsimd is put in the lexico index*/
   //Let's make it work on GPU and then make a special accelerator_for that
@@ -247,7 +247,7 @@ public:
     Coordinate local     =unpadded_grid->LocalDimensions();
     Coordinate procs     =unpadded_grid->ProcessorGrid();
     for(int d=0;d<dims;d++){
-      if ( procs[d] > 1 ) assert(local[d]>=depth);
+      if ( procs[d] > 1 ) GRID_ASSERT(local[d]>=depth);
     }
   }
   void DeleteGrids(void)
@@ -448,9 +448,9 @@ public:
     int nld   = to.Grid()->_ldimensions[dimension];
     const int Nsimd = vobj::Nsimd();
 
-    assert(depth<=lds[dimension]); // A must be on neighbouring node
+    GRID_ASSERT(depth<=lds[dimension]); // A must be on neighbouring node
-    assert(depth>0);   // A caller bug if zero
+    GRID_ASSERT(depth>0);   // A caller bug if zero
-    assert(ld+2*depth==nld);
+    GRID_ASSERT(ld+2*depth==nld);
     ////////////////////////////////////////////////////////////////////////////
     // Face size and byte calculations
     ////////////////////////////////////////////////////////////////////////////
@@ -460,15 +460,21 @@ public:
     }
     buffer_size = buffer_size  / Nsimd;
     int rNsimd = Nsimd / simd[dimension];
-    assert( buffer_size == from.Grid()->_slice_nblock[dimension]*from.Grid()->_slice_block[dimension] / simd[dimension]);
+    GRID_ASSERT( buffer_size == from.Grid()->_slice_nblock[dimension]*from.Grid()->_slice_block[dimension] / simd[dimension]);
 
-    static cshiftVector<vobj> send_buf; 
+    static deviceVector<vobj> send_buf; 
-    static cshiftVector<vobj> recv_buf;
+    static deviceVector<vobj> recv_buf;
     send_buf.resize(buffer_size*2*depth);    
     recv_buf.resize(buffer_size*2*depth);
+#ifndef ACCELERATOR_AWARE_MPI
+    static hostVector<vobj> hsend_buf; 
+    static hostVector<vobj> hrecv_buf;
+    hsend_buf.resize(buffer_size*2*depth);    
+    hrecv_buf.resize(buffer_size*2*depth);
+#endif    
 
-    std::vector<CommsRequest_t> fwd_req;   
+    std::vector<MpiCommsRequest_t> fwd_req;   
-    std::vector<CommsRequest_t> bwd_req;   
+    std::vector<MpiCommsRequest_t> bwd_req;   
 
     int words = buffer_size;
     int bytes = words * sizeof(vobj);
@@ -495,9 +501,16 @@ public:
       t_gather+=usecond()-t;
 
       t=usecond();
+#ifdef ACCELERATOR_AWARE_MPI
       grid->SendToRecvFromBegin(fwd_req,
 				(void *)&send_buf[d*buffer_size], xmit_to_rank,
 				(void *)&recv_buf[d*buffer_size], recv_from_rank, bytes, tag);
+#else
+      acceleratorCopyFromDevice(&send_buf[d*buffer_size],&hsend_buf[d*buffer_size],bytes);
+      grid->SendToRecvFromBegin(fwd_req,
+				(void *)&hsend_buf[d*buffer_size], xmit_to_rank,
+				(void *)&hrecv_buf[d*buffer_size], recv_from_rank, bytes, tag);
+#endif
       t_comms+=usecond()-t;
      }
     for ( int d=0;d < depth ; d ++ ) {
@@ -508,9 +521,16 @@ public:
       t_gather+= usecond() - t;
 
       t=usecond();
+#ifdef ACCELERATOR_AWARE_MPI
       grid->SendToRecvFromBegin(bwd_req,
 				(void *)&send_buf[(d+depth)*buffer_size], recv_from_rank,
 				(void *)&recv_buf[(d+depth)*buffer_size], xmit_to_rank, bytes,tag);
+#else
+      acceleratorCopyFromDevice(&send_buf[(d+depth)*buffer_size],&hsend_buf[(d+depth)*buffer_size],bytes);
+      grid->SendToRecvFromBegin(bwd_req,
+				(void *)&hsend_buf[(d+depth)*buffer_size], recv_from_rank,
+				(void *)&hrecv_buf[(d+depth)*buffer_size], xmit_to_rank, bytes,tag);
+#endif      
       t_comms+=usecond()-t;
     }
 
@@ -533,6 +553,11 @@ public:
 
     t=usecond();
     grid->CommsComplete(fwd_req);
+#ifndef ACCELERATOR_AWARE_MPI
+    for ( int d=0;d < depth ; d ++ ) {
+      acceleratorCopyToDevice(&hrecv_buf[d*buffer_size],&recv_buf[d*buffer_size],bytes);
+    }
+#endif
     t_comms+= usecond() - t;
     
     t=usecond();
@@ -543,6 +568,11 @@ public:
 
     t=usecond();
     grid->CommsComplete(bwd_req);
+#ifndef ACCELERATOR_AWARE_MPI
+    for ( int d=0;d < depth ; d ++ ) {
+      acceleratorCopyToDevice(&hrecv_buf[(d+depth)*buffer_size],&recv_buf[(d+depth)*buffer_size],bytes);
+    }
+#endif
     t_comms+= usecond() - t;
     
     t=usecond();

Grid/log/Log.cc

@@ -69,6 +69,7 @@ GridLogger GridLogMemory (1, "Memory", GridLogColours, "NORMAL");
 GridLogger GridLogTracing(1, "Tracing", GridLogColours, "NORMAL");
 GridLogger GridLogDebug  (1, "Debug", GridLogColours, "PURPLE");
 GridLogger GridLogPerformance(1, "Performance", GridLogColours, "GREEN");
+GridLogger GridLogComms      (1, "Comms",  GridLogColours, "BLUE");
 GridLogger GridLogDslash     (1, "Dslash", GridLogColours, "BLUE");
 GridLogger GridLogIterative  (1, "Iterative", GridLogColours, "BLUE");
 GridLogger GridLogIntegrator (1, "Integrator", GridLogColours, "BLUE");
@@ -84,6 +85,7 @@ void GridLogConfigure(std::vector<std::string> &logstreams) {
   GridLogDebug.Active(0);
   GridLogPerformance.Active(0);
   GridLogDslash.Active(0);
+  GridLogComms.Active(0);
   GridLogIntegrator.Active(1);
   GridLogColours.Active(0);
   GridLogHMC.Active(1);
@@ -97,6 +99,7 @@ void GridLogConfigure(std::vector<std::string> &logstreams) {
     if (logstreams[i] == std::string("Debug"))       GridLogDebug.Active(1);
     if (logstreams[i] == std::string("Performance")) GridLogPerformance.Active(1);
     if (logstreams[i] == std::string("Dslash"))      GridLogDslash.Active(1);
+    if (logstreams[i] == std::string("Comms"))       GridLogComms.Active(1);
     if (logstreams[i] == std::string("NoIntegrator"))GridLogIntegrator.Active(0);
     if (logstreams[i] == std::string("NoHMC"))       GridLogHMC.Active(0);
     if (logstreams[i] == std::string("Colours"))     GridLogColours.Active(1);