1 #include <../src/mat/impls/baij/mpi/mpibaij.h> 2 3 PETSC_INTERN PetscErrorCode MatConvert_SeqBAIJ_SeqBAIJMKL(Mat,MatType,MatReuse,Mat*); 4 5 static PetscErrorCode MatMPIBAIJSetPreallocation_MPIBAIJMKL(Mat B,PetscInt bs,PetscInt d_nz,const PetscInt *d_nnz,PetscInt o_nz,const PetscInt *o_nnz) 6 { 7 Mat_MPIBAIJ *b = (Mat_MPIBAIJ*)B->data; 8 9 PetscFunctionBegin; 10 PetscCall(MatMPIBAIJSetPreallocation_MPIBAIJ(B,bs,d_nz,d_nnz,o_nz,o_nnz)); 11 PetscCall(MatConvert_SeqBAIJ_SeqBAIJMKL(b->A,MATSEQBAIJMKL,MAT_INPLACE_MATRIX,&b->A)); 12 PetscCall(MatConvert_SeqBAIJ_SeqBAIJMKL(b->B,MATSEQBAIJMKL,MAT_INPLACE_MATRIX,&b->B)); 13 PetscFunctionReturn(0); 14 } 15 16 static PetscErrorCode MatConvert_MPIBAIJ_MPIBAIJMKL(Mat A,MatType type,MatReuse reuse,Mat *newmat) 17 { 18 Mat B = *newmat; 19 20 PetscFunctionBegin; 21 if (reuse == MAT_INITIAL_MATRIX) { 22 PetscCall(MatDuplicate(A,MAT_COPY_VALUES,&B)); 23 } 24 25 PetscCall(PetscObjectChangeTypeName((PetscObject) B, MATMPIBAIJMKL)); 26 PetscCall(PetscObjectComposeFunction((PetscObject)B,"MatMPIBAIJSetPreallocation_C",MatMPIBAIJSetPreallocation_MPIBAIJMKL)); 27 *newmat = B; 28 PetscFunctionReturn(0); 29 } 30 31 /*@C 32 MatCreateBAIJMKL - Creates a sparse parallel matrix in block AIJ format 33 (block compressed row). 34 This type inherits from BAIJ and is largely identical, but uses sparse BLAS 35 routines from Intel MKL whenever possible. 36 MatMult, MatMultAdd, MatMultTranspose, and MatMultTransposeAdd 37 operations are currently supported. 38 If the installed version of MKL supports the "SpMV2" sparse 39 inspector-executor routines, then those are used by default. 40 Default PETSc kernels are used otherwise. 41 For good matrix assembly performance the user should preallocate the matrix 42 storage by setting the parameters d_nz (or d_nnz) and o_nz (or o_nnz). 43 By setting these parameters accurately, performance can be increased by more 44 than a factor of 50. 45 46 Collective 47 48 Input Parameters: 49 + comm - MPI communicator 50 . bs - size of block, the blocks are ALWAYS square. One can use MatSetBlockSizes() to set a different row and column blocksize but the row 51 blocksize always defines the size of the blocks. The column blocksize sets the blocksize of the vectors obtained with MatCreateVecs() 52 . m - number of local rows (or PETSC_DECIDE to have calculated if M is given) 53 This value should be the same as the local size used in creating the 54 y vector for the matrix-vector product y = Ax. 55 . n - number of local columns (or PETSC_DECIDE to have calculated if N is given) 56 This value should be the same as the local size used in creating the 57 x vector for the matrix-vector product y = Ax. 58 . M - number of global rows (or PETSC_DETERMINE to have calculated if m is given) 59 . N - number of global columns (or PETSC_DETERMINE to have calculated if n is given) 60 . d_nz - number of nonzero blocks per block row in diagonal portion of local 61 submatrix (same for all local rows) 62 . d_nnz - array containing the number of nonzero blocks in the various block rows 63 of the in diagonal portion of the local (possibly different for each block 64 row) or NULL. If you plan to factor the matrix you must leave room for the diagonal entry 65 and set it even if it is zero. 66 . o_nz - number of nonzero blocks per block row in the off-diagonal portion of local 67 submatrix (same for all local rows). 68 - o_nnz - array containing the number of nonzero blocks in the various block rows of the 69 off-diagonal portion of the local submatrix (possibly different for 70 each block row) or NULL. 71 72 Output Parameter: 73 . A - the matrix 74 75 Options Database Keys: 76 + -mat_block_size - size of the blocks to use 77 - -mat_use_hash_table <fact> - set hash table factor 78 79 It is recommended that one use the MatCreate(), MatSetType() and/or MatSetFromOptions(), 80 MatXXXXSetPreallocation() paradigm instead of this routine directly. 81 [MatXXXXSetPreallocation() is, for example, MatSeqAIJSetPreallocation] 82 83 Notes: 84 If the *_nnz parameter is given then the *_nz parameter is ignored 85 86 A nonzero block is any block that as 1 or more nonzeros in it 87 88 The user MUST specify either the local or global matrix dimensions 89 (possibly both). 90 91 If PETSC_DECIDE or PETSC_DETERMINE is used for a particular argument on one processor 92 than it must be used on all processors that share the object for that argument. 93 94 Storage Information: 95 For a square global matrix we define each processor's diagonal portion 96 to be its local rows and the corresponding columns (a square submatrix); 97 each processor's off-diagonal portion encompasses the remainder of the 98 local matrix (a rectangular submatrix). 99 100 The user can specify preallocated storage for the diagonal part of 101 the local submatrix with either d_nz or d_nnz (not both). Set 102 d_nz=PETSC_DEFAULT and d_nnz=NULL for PETSc to control dynamic 103 memory allocation. Likewise, specify preallocated storage for the 104 off-diagonal part of the local submatrix with o_nz or o_nnz (not both). 105 106 Consider a processor that owns rows 3, 4 and 5 of a parallel matrix. In 107 the figure below we depict these three local rows and all columns (0-11). 108 109 .vb 110 0 1 2 3 4 5 6 7 8 9 10 11 111 -------------------------- 112 row 3 |o o o d d d o o o o o o 113 row 4 |o o o d d d o o o o o o 114 row 5 |o o o d d d o o o o o o 115 -------------------------- 116 .ve 117 118 Thus, any entries in the d locations are stored in the d (diagonal) 119 submatrix, and any entries in the o locations are stored in the 120 o (off-diagonal) submatrix. Note that the d and the o submatrices are 121 stored simply in the MATSEQBAIJMKL format for compressed row storage. 122 123 Now d_nz should indicate the number of block nonzeros per row in the d matrix, 124 and o_nz should indicate the number of block nonzeros per row in the o matrix. 125 In general, for PDE problems in which most nonzeros are near the diagonal, 126 one expects d_nz >> o_nz. For large problems you MUST preallocate memory 127 or you will get TERRIBLE performance; see the users' manual chapter on 128 matrices. 129 130 Level: intermediate 131 132 .seealso: MatCreate(), MatCreateSeqBAIJMKL(), MatSetValues(), MatCreateBAIJMKL(), MatMPIBAIJSetPreallocation(), MatMPIBAIJSetPreallocationCSR() 133 @*/ 134 135 PetscErrorCode MatCreateBAIJMKL(MPI_Comm comm,PetscInt bs,PetscInt m,PetscInt n,PetscInt M,PetscInt N,PetscInt d_nz,const PetscInt d_nnz[],PetscInt o_nz,const PetscInt o_nnz[],Mat *A) 136 { 137 PetscMPIInt size; 138 139 PetscFunctionBegin; 140 PetscCall(MatCreate(comm,A)); 141 PetscCall(MatSetSizes(*A,m,n,M,N)); 142 PetscCallMPI(MPI_Comm_size(comm,&size)); 143 if (size > 1) { 144 PetscCall(MatSetType(*A,MATMPIBAIJMKL)); 145 PetscCall(MatMPIBAIJSetPreallocation(*A,bs,d_nz,d_nnz,o_nz,o_nnz)); 146 } else { 147 PetscCall(MatSetType(*A,MATSEQBAIJMKL)); 148 PetscCall(MatSeqBAIJSetPreallocation(*A,bs,d_nz,d_nnz)); 149 } 150 PetscFunctionReturn(0); 151 } 152 153 PETSC_EXTERN PetscErrorCode MatCreate_MPIBAIJMKL(Mat A) 154 { 155 PetscFunctionBegin; 156 PetscCall(MatSetType(A,MATMPIBAIJ)); 157 PetscCall(MatConvert_MPIBAIJ_MPIBAIJMKL(A,MATMPIBAIJMKL,MAT_INPLACE_MATRIX,&A)); 158 PetscFunctionReturn(0); 159 } 160 161 /*MC 162 MATBAIJMKL - MATBAIJMKL = "BAIJMKL" - A matrix type to be used for sparse matrices. 163 164 This matrix type is identical to MATSEQBAIJMKL when constructed with a single process communicator, 165 and MATMPIBAIJMKL otherwise. As a result, for single process communicators, 166 MatSeqBAIJSetPreallocation() is supported, and similarly MatMPIBAIJSetPreallocation() is supported 167 for communicators controlling multiple processes. It is recommended that you call both of 168 the above preallocation routines for simplicity. 169 170 Options Database Keys: 171 . -mat_type baijmkl - sets the matrix type to "BAIJMKL" during a call to MatSetFromOptions() 172 173 Level: beginner 174 175 .seealso: MatCreateBAIJMKL(), MATSEQBAIJMKL, MATMPIBAIJMKL 176 M*/ 177