xref: /petsc/src/mat/impls/aij/mpi/aijmkl/mpiaijmkl.c (revision 72a35d6d63ea98b60d8ef4c5aaefda5ff933c0fe)
1 #include <../src/mat/impls/aij/mpi/mpiaij.h>
2 /*@C
3    MatCreateMPIAIJMKL - Creates a sparse parallel matrix whose local
4    portions are stored as `MATSEQAIJMKL` matrices (a matrix class that inherits
5    from `MATSEQAIJ` but uses some operations provided by Intel MKL).  The same
6    guidelines that apply to `MATMPIAIJ` matrices for preallocating the matrix
7    storage apply here as well.
8 
9       Collective
10 
11    Input Parameters:
12 +  comm - MPI communicator
13 .  m - number of local rows (or `PETSC_DECIDE` to have calculated if `M` is given)
14            This value should be the same as the local size used in creating the
15            y vector for the matrix-vector product y = Ax.
16 .  n - This value should be the same as the local size used in creating the
17        x vector for the matrix-vector product y = Ax. (or `PETSC_DECIDE` to have
18        calculated if N is given) For square matrices n is almost always `m`.
19 .  M - number of global rows (or `PETSC_DETERMINE` to have calculated if `m` is given)
20 .  N - number of global columns (or `PETSC_DETERMINE` to have calculated if `n` is given)
21 .  d_nz  - number of nonzeros per row in DIAGONAL portion of local submatrix
22            (same value is used for all local rows)
23 .  d_nnz - array containing the number of nonzeros in the various rows of the
24            DIAGONAL portion of the local submatrix (possibly different for each row)
25            or `NULL`, if `d_nz` is used to specify the nonzero structure.
26            The size of this array is equal to the number of local rows, i.e `m`.
27            For matrices you plan to factor you must leave room for the diagonal entry and
28            put in the entry even if it is zero.
29 .  o_nz  - number of nonzeros per row in the OFF-DIAGONAL portion of local
30            submatrix (same value is used for all local rows).
31 -  o_nnz - array containing the number of nonzeros in the various rows of the
32            OFF-DIAGONAL portion of the local submatrix (possibly different for
33            each row) or `NULL`, if `o_nz` is used to specify the nonzero
34            structure. The size of this array is equal to the number
35            of local rows, i.e `m`.
36 
37    Output Parameter:
38 .  A - the matrix
39 
40    Options Database Key:
41 .  -mat_aijmkl_no_spmv2 - disables use of the SpMV2 inspector-executor routines
42 
43    Level: intermediate
44 
45    Notes:
46    If the *_nnz parameter is given then the *_nz parameter is ignored
47 
48    `m`,`n`,`M`,`N` parameters specify the size of the matrix, and its partitioning across
49    processors, while `d_nz`,`d_nnz`,`o_nz`,`o_nnz` parameters specify the approximate
50    storage requirements for this matrix.
51 
52    If `PETSC_DECIDE` or `PETSC_DETERMINE` is used for a particular argument on one
53    processor than it must be used on all processors that share the object for
54    that argument.
55 
56    The user MUST specify either the local or global matrix dimensions
57    (possibly both).
58 
59    The parallel matrix is partitioned such that the first m0 rows belong to
60    process 0, the next m1 rows belong to process 1, the next m2 rows belong
61    to process 2 etc.. where m0,m1,m2... are the input parameter `m`.
62 
63    The DIAGONAL portion of the local submatrix of a processor can be defined
64    as the submatrix which is obtained by extraction the part corresponding
65    to the rows r1-r2 and columns r1-r2 of the global matrix, where r1 is the
66    first row that belongs to the processor, and r2 is the last row belonging
67    to the this processor. This is a square mxm matrix. The remaining portion
68    of the local submatrix (mxN) constitute the OFF-DIAGONAL portion.
69 
70    If `o_nnz`, `d_nnz` are specified, then `o_nz`, and `d_nz` are ignored.
71 
72    When calling this routine with a single process communicator, a matrix of
73    type `MATSEQAIJMKL` is returned.  If a matrix of type `MATMPIAIJMKL` is desired
74    for this type of communicator, use the construction mechanism:
75      `MatCreate`(...,&A); `MatSetType`(A,MPIAIJMKL); `MatMPIAIJSetPreallocation`(A,...);
76 
77 .seealso: [](chapter_matrices), `Mat`, [Sparse Matrix Creation](sec_matsparse), `MATMPIAIJMKL`, `MatCreate()`, `MatCreateSeqAIJMKL()`, `MatSetValues()`
78 @*/
79 PetscErrorCode MatCreateMPIAIJMKL(MPI_Comm comm, PetscInt m, PetscInt n, PetscInt M, PetscInt N, PetscInt d_nz, const PetscInt d_nnz[], PetscInt o_nz, const PetscInt o_nnz[], Mat *A)
80 {
81   PetscMPIInt size;
82 
83   PetscFunctionBegin;
84   PetscCall(MatCreate(comm, A));
85   PetscCall(MatSetSizes(*A, m, n, M, N));
86   PetscCallMPI(MPI_Comm_size(comm, &size));
87   if (size > 1) {
88     PetscCall(MatSetType(*A, MATMPIAIJMKL));
89     PetscCall(MatMPIAIJSetPreallocation(*A, d_nz, d_nnz, o_nz, o_nnz));
90   } else {
91     PetscCall(MatSetType(*A, MATSEQAIJMKL));
92     PetscCall(MatSeqAIJSetPreallocation(*A, d_nz, d_nnz));
93   }
94   PetscFunctionReturn(PETSC_SUCCESS);
95 }
96 
97 PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_SeqAIJMKL(Mat, MatType, MatReuse, Mat *);
98 
99 PetscErrorCode MatMPIAIJSetPreallocation_MPIAIJMKL(Mat B, PetscInt d_nz, const PetscInt d_nnz[], PetscInt o_nz, const PetscInt o_nnz[])
100 {
101   Mat_MPIAIJ *b = (Mat_MPIAIJ *)B->data;
102 
103   PetscFunctionBegin;
104   PetscCall(MatMPIAIJSetPreallocation_MPIAIJ(B, d_nz, d_nnz, o_nz, o_nnz));
105   PetscCall(MatConvert_SeqAIJ_SeqAIJMKL(b->A, MATSEQAIJMKL, MAT_INPLACE_MATRIX, &b->A));
106   PetscCall(MatConvert_SeqAIJ_SeqAIJMKL(b->B, MATSEQAIJMKL, MAT_INPLACE_MATRIX, &b->B));
107   PetscFunctionReturn(PETSC_SUCCESS);
108 }
109 
110 PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJMKL(Mat A, MatType type, MatReuse reuse, Mat *newmat)
111 {
112   Mat B = *newmat;
113 
114   PetscFunctionBegin;
115   if (reuse == MAT_INITIAL_MATRIX) PetscCall(MatDuplicate(A, MAT_COPY_VALUES, &B));
116   PetscCall(PetscObjectChangeTypeName((PetscObject)B, MATMPIAIJMKL));
117   PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatMPIAIJSetPreallocation_C", MatMPIAIJSetPreallocation_MPIAIJMKL));
118   *newmat = B;
119   PetscFunctionReturn(PETSC_SUCCESS);
120 }
121 
122 PETSC_EXTERN PetscErrorCode MatCreate_MPIAIJMKL(Mat A)
123 {
124   PetscFunctionBegin;
125   PetscCall(MatSetType(A, MATMPIAIJ));
126   PetscCall(MatConvert_MPIAIJ_MPIAIJMKL(A, MATMPIAIJMKL, MAT_INPLACE_MATRIX, &A));
127   PetscFunctionReturn(PETSC_SUCCESS);
128 }
129 
130 /*MC
131    MATAIJMKL - MATAIJMKL = "AIJMKL" - A matrix type to be used for sparse matrices.
132 
133    This matrix type is identical to `MATSEQAIJMKL` when constructed with a single process communicator,
134    and `MATMPIAIJMKL` otherwise.  As a result, for single process communicators,
135   MatSeqAIJSetPreallocation() is supported, and similarly `MatMPIAIJSetPreallocation()` is supported
136   for communicators controlling multiple processes.  It is recommended that you call both of
137   the above preallocation routines for simplicity.
138 
139    Options Database Key:
140 . -mat_type aijmkl - sets the matrix type to `MATAIJMKL` during a call to `MatSetFromOptions()`
141 
142   Level: beginner
143 
144 .seealso: [](chapter_matrices), `Mat`, `MATMPIAIJMKL`, `MATSEQAIJMKL`, `MatCreateMPIAIJMKL()`, `MATSEQAIJMKL`, `MATMPIAIJMKL`, `MATSEQAIJSELL`, `MATMPIAIJSELL`, `MATSEQAIJPERM`, `MATMPIAIJPERM`
145 M*/
146