1 2 /* 3 Defines a matrix-vector product for the MATMPIAIJCRL matrix class. 4 This class is derived from the MATMPIAIJ class and retains the 5 compressed row storage (aka Yale sparse matrix format) but augments 6 it with a column oriented storage that is more efficient for 7 matrix vector products on Vector machines. 8 9 CRL stands for constant row length (that is the same number of columns 10 is kept (padded with zeros) for each row of the sparse matrix. 11 12 See src/mat/impls/aij/seq/crl/crl.c for the sequential version 13 */ 14 15 #include <../src/mat/impls/aij/mpi/mpiaij.h> 16 #include <../src/mat/impls/aij/seq/crl/crl.h> 17 18 PetscErrorCode MatDestroy_MPIAIJCRL(Mat A) 19 { 20 PetscErrorCode ierr; 21 Mat_AIJCRL *aijcrl = (Mat_AIJCRL*) A->spptr; 22 23 /* Free everything in the Mat_AIJCRL data structure. */ 24 if (aijcrl) { 25 ierr = PetscFree2(aijcrl->acols,aijcrl->icols);CHKERRQ(ierr); 26 ierr = VecDestroy(&aijcrl->fwork);CHKERRQ(ierr); 27 ierr = VecDestroy(&aijcrl->xwork);CHKERRQ(ierr); 28 ierr = PetscFree(aijcrl->array);CHKERRQ(ierr); 29 } 30 ierr = PetscFree(A->spptr);CHKERRQ(ierr); 31 32 ierr = PetscObjectChangeTypeName((PetscObject)A, MATMPIAIJ);CHKERRQ(ierr); 33 ierr = MatDestroy_MPIAIJ(A);CHKERRQ(ierr); 34 PetscFunctionReturn(0); 35 } 36 37 PetscErrorCode MatMPIAIJCRL_create_aijcrl(Mat A) 38 { 39 Mat_MPIAIJ *a = (Mat_MPIAIJ*)(A)->data; 40 Mat_SeqAIJ *Aij = (Mat_SeqAIJ*)(a->A->data), *Bij = (Mat_SeqAIJ*)(a->B->data); 41 Mat_AIJCRL *aijcrl = (Mat_AIJCRL*) A->spptr; 42 PetscInt m = A->rmap->n; /* Number of rows in the matrix. */ 43 PetscInt nd = a->A->cmap->n; /* number of columns in diagonal portion */ 44 PetscInt *aj = Aij->j,*bj = Bij->j; /* From the CSR representation; points to the beginning of each row. */ 45 PetscInt i, j,rmax = 0,*icols, *ailen = Aij->ilen, *bilen = Bij->ilen; 46 PetscScalar *aa = Aij->a,*ba = Bij->a,*acols,*array; 47 PetscErrorCode ierr; 48 49 PetscFunctionBegin; 50 /* determine the row with the most columns */ 51 for (i=0; i<m; i++) { 52 rmax = PetscMax(rmax,ailen[i]+bilen[i]); 53 } 54 aijcrl->nz = Aij->nz+Bij->nz; 55 aijcrl->m = A->rmap->n; 56 aijcrl->rmax = rmax; 57 58 ierr = PetscFree2(aijcrl->acols,aijcrl->icols);CHKERRQ(ierr); 59 ierr = PetscMalloc2(rmax*m,&aijcrl->acols,rmax*m,&aijcrl->icols);CHKERRQ(ierr); 60 acols = aijcrl->acols; 61 icols = aijcrl->icols; 62 for (i=0; i<m; i++) { 63 for (j=0; j<ailen[i]; j++) { 64 acols[j*m+i] = *aa++; 65 icols[j*m+i] = *aj++; 66 } 67 for (; j<ailen[i]+bilen[i]; j++) { 68 acols[j*m+i] = *ba++; 69 icols[j*m+i] = nd + *bj++; 70 } 71 for (; j<rmax; j++) { /* empty column entries */ 72 acols[j*m+i] = 0.0; 73 icols[j*m+i] = (j) ? icols[(j-1)*m+i] : 0; /* handle case where row is EMPTY */ 74 } 75 } 76 ierr = PetscInfo1(A,"Percentage of 0's introduced for vectorized multiply %g\n",1.0-((double)(aijcrl->nz))/((double)(rmax*m)));CHKERRQ(ierr); 77 78 ierr = PetscFree(aijcrl->array);CHKERRQ(ierr); 79 ierr = PetscMalloc1(a->B->cmap->n+nd,&array);CHKERRQ(ierr); 80 /* xwork array is actually B->n+nd long, but we define xwork this length so can copy into it */ 81 ierr = VecDestroy(&aijcrl->xwork);CHKERRQ(ierr); 82 ierr = VecCreateMPIWithArray(PetscObjectComm((PetscObject)A),1,nd,PETSC_DECIDE,array,&aijcrl->xwork);CHKERRQ(ierr); 83 ierr = VecDestroy(&aijcrl->fwork);CHKERRQ(ierr); 84 ierr = VecCreateSeqWithArray(PETSC_COMM_SELF,1,a->B->cmap->n,array+nd,&aijcrl->fwork);CHKERRQ(ierr); 85 86 aijcrl->array = array; 87 aijcrl->xscat = a->Mvctx; 88 PetscFunctionReturn(0); 89 } 90 91 extern PetscErrorCode MatAssemblyEnd_MPIAIJ(Mat,MatAssemblyType); 92 93 PetscErrorCode MatAssemblyEnd_MPIAIJCRL(Mat A, MatAssemblyType mode) 94 { 95 PetscErrorCode ierr; 96 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 97 Mat_SeqAIJ *Aij = (Mat_SeqAIJ*)(a->A->data), *Bij = (Mat_SeqAIJ*)(a->A->data); 98 99 PetscFunctionBegin; 100 Aij->inode.use = PETSC_FALSE; 101 Bij->inode.use = PETSC_FALSE; 102 103 ierr = MatAssemblyEnd_MPIAIJ(A,mode);CHKERRQ(ierr); 104 if (mode == MAT_FLUSH_ASSEMBLY) PetscFunctionReturn(0); 105 106 /* Now calculate the permutation and grouping information. */ 107 ierr = MatMPIAIJCRL_create_aijcrl(A);CHKERRQ(ierr); 108 PetscFunctionReturn(0); 109 } 110 111 extern PetscErrorCode MatMult_AIJCRL(Mat,Vec,Vec); 112 extern PetscErrorCode MatDuplicate_AIJCRL(Mat,MatDuplicateOption,Mat*); 113 114 /* MatConvert_MPIAIJ_MPIAIJCRL converts a MPIAIJ matrix into a 115 * MPIAIJCRL matrix. This routine is called by the MatCreate_MPIAIJCRL() 116 * routine, but can also be used to convert an assembled MPIAIJ matrix 117 * into a MPIAIJCRL one. */ 118 119 PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJCRL(Mat A,MatType type,MatReuse reuse,Mat *newmat) 120 { 121 PetscErrorCode ierr; 122 Mat B = *newmat; 123 Mat_AIJCRL *aijcrl; 124 125 PetscFunctionBegin; 126 if (reuse == MAT_INITIAL_MATRIX) { 127 ierr = MatDuplicate(A,MAT_COPY_VALUES,&B);CHKERRQ(ierr); 128 } 129 130 ierr = PetscNewLog(B,&aijcrl);CHKERRQ(ierr); 131 B->spptr = (void*) aijcrl; 132 133 /* Set function pointers for methods that we inherit from AIJ but override. */ 134 B->ops->duplicate = MatDuplicate_AIJCRL; 135 B->ops->assemblyend = MatAssemblyEnd_MPIAIJCRL; 136 B->ops->destroy = MatDestroy_MPIAIJCRL; 137 B->ops->mult = MatMult_AIJCRL; 138 139 /* If A has already been assembled, compute the permutation. */ 140 if (A->assembled) { 141 ierr = MatMPIAIJCRL_create_aijcrl(B);CHKERRQ(ierr); 142 } 143 ierr = PetscObjectChangeTypeName((PetscObject)B,MATMPIAIJCRL);CHKERRQ(ierr); 144 *newmat = B; 145 PetscFunctionReturn(0); 146 } 147 148 /*@C 149 MatCreateMPIAIJCRL - Creates a sparse matrix of type MPIAIJCRL. 150 This type inherits from AIJ, but stores some additional 151 information that is used to allow better vectorization of 152 the matrix-vector product. At the cost of increased storage, the AIJ formatted 153 matrix can be copied to a format in which pieces of the matrix are 154 stored in ELLPACK format, allowing the vectorized matrix multiply 155 routine to use stride-1 memory accesses. As with the AIJ type, it is 156 important to preallocate matrix storage in order to get good assembly 157 performance. 158 159 Collective on MPI_Comm 160 161 Input Parameters: 162 + comm - MPI communicator, set to PETSC_COMM_SELF 163 . m - number of rows 164 . n - number of columns 165 . nz - number of nonzeros per row (same for all rows) 166 - nnz - array containing the number of nonzeros in the various rows 167 (possibly different for each row) or NULL 168 169 Output Parameter: 170 . A - the matrix 171 172 Notes: 173 If nnz is given then nz is ignored 174 175 Level: intermediate 176 177 .keywords: matrix, cray, sparse, parallel 178 179 .seealso: MatCreate(), MatCreateMPIAIJPERM(), MatSetValues() 180 @*/ 181 PetscErrorCode MatCreateMPIAIJCRL(MPI_Comm comm,PetscInt m,PetscInt n,PetscInt nz,const PetscInt nnz[],PetscInt onz,const PetscInt onnz[],Mat *A) 182 { 183 PetscErrorCode ierr; 184 185 PetscFunctionBegin; 186 ierr = MatCreate(comm,A);CHKERRQ(ierr); 187 ierr = MatSetSizes(*A,m,n,m,n);CHKERRQ(ierr); 188 ierr = MatSetType(*A,MATMPIAIJCRL);CHKERRQ(ierr); 189 ierr = MatMPIAIJSetPreallocation_MPIAIJ(*A,nz,(PetscInt*)nnz,onz,(PetscInt*)onnz);CHKERRQ(ierr); 190 PetscFunctionReturn(0); 191 } 192 193 PETSC_EXTERN PetscErrorCode MatCreate_MPIAIJCRL(Mat A) 194 { 195 PetscErrorCode ierr; 196 197 PetscFunctionBegin; 198 ierr = MatSetType(A,MATMPIAIJ);CHKERRQ(ierr); 199 ierr = MatConvert_MPIAIJ_MPIAIJCRL(A,MATMPIAIJCRL,MAT_INPLACE_MATRIX,&A);CHKERRQ(ierr); 200 PetscFunctionReturn(0); 201 } 202 203