1 #include <petscconf.h>
2 #include <petscdevice.h>
3 #include <../src/mat/impls/sell/mpi/mpisell.h> /*I "petscmat.h" I*/
4
MatMPISELLSetPreallocation_MPISELLCUDA(Mat B,PetscInt d_rlenmax,const PetscInt d_rlen[],PetscInt o_rlenmax,const PetscInt o_rlen[])5 static PetscErrorCode MatMPISELLSetPreallocation_MPISELLCUDA(Mat B, PetscInt d_rlenmax, const PetscInt d_rlen[], PetscInt o_rlenmax, const PetscInt o_rlen[])
6 {
7 Mat_MPISELL *b = (Mat_MPISELL *)B->data;
8
9 PetscFunctionBegin;
10 PetscCall(PetscLayoutSetUp(B->rmap));
11 PetscCall(PetscLayoutSetUp(B->cmap));
12
13 if (!B->preallocated) {
14 /* Explicitly create 2 MATSEQSELLCUDA matrices. */
15 PetscCall(MatCreate(PETSC_COMM_SELF, &b->A));
16 PetscCall(MatBindToCPU(b->A, B->boundtocpu));
17 PetscCall(MatSetSizes(b->A, B->rmap->n, B->cmap->n, B->rmap->n, B->cmap->n));
18 PetscCall(MatSetType(b->A, MATSEQSELLCUDA));
19 PetscCall(MatCreate(PETSC_COMM_SELF, &b->B));
20 PetscCall(MatBindToCPU(b->B, B->boundtocpu));
21 PetscCall(MatSetSizes(b->B, B->rmap->n, B->cmap->N, B->rmap->n, B->cmap->N));
22 PetscCall(MatSetType(b->B, MATSEQSELLCUDA));
23 }
24 PetscCall(MatSeqSELLSetPreallocation(b->A, d_rlenmax, d_rlen));
25 PetscCall(MatSeqSELLSetPreallocation(b->B, o_rlenmax, o_rlen));
26 B->preallocated = PETSC_TRUE;
27 B->was_assembled = PETSC_FALSE;
28 B->assembled = PETSC_FALSE;
29 PetscFunctionReturn(PETSC_SUCCESS);
30 }
31
MatSetFromOptions_MPISELLCUDA(Mat,PetscOptionItems)32 static PetscErrorCode MatSetFromOptions_MPISELLCUDA(Mat, PetscOptionItems)
33 {
34 return PETSC_SUCCESS;
35 }
36
MatAssemblyEnd_MPISELLCUDA(Mat A,MatAssemblyType mode)37 static PetscErrorCode MatAssemblyEnd_MPISELLCUDA(Mat A, MatAssemblyType mode)
38 {
39 PetscFunctionBegin;
40 PetscCall(MatAssemblyEnd_MPISELL(A, mode));
41 if (!A->was_assembled && mode == MAT_FINAL_ASSEMBLY) PetscCall(VecSetType(((Mat_MPISELL *)A->data)->lvec, VECSEQCUDA));
42 PetscFunctionReturn(PETSC_SUCCESS);
43 }
44
MatDestroy_MPISELLCUDA(Mat A)45 static PetscErrorCode MatDestroy_MPISELLCUDA(Mat A)
46 {
47 PetscFunctionBegin;
48 PetscCall(MatDestroy_MPISELL(A));
49 PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatMPISELLSetPreallocation_C", NULL));
50 PetscFunctionReturn(PETSC_SUCCESS);
51 }
52
MatConvert_MPISELL_MPISELLCUDA(Mat B,MatType,MatReuse reuse,Mat * newmat)53 PETSC_INTERN PetscErrorCode MatConvert_MPISELL_MPISELLCUDA(Mat B, MatType, MatReuse reuse, Mat *newmat)
54 {
55 Mat_MPISELL *a;
56 Mat A;
57
58 PetscFunctionBegin;
59 PetscCall(PetscDeviceInitialize(PETSC_DEVICE_CUDA));
60 if (reuse == MAT_INITIAL_MATRIX) PetscCall(MatDuplicate(B, MAT_COPY_VALUES, newmat));
61 else if (reuse == MAT_REUSE_MATRIX) PetscCall(MatCopy(B, *newmat, SAME_NONZERO_PATTERN));
62 A = *newmat;
63 A->boundtocpu = PETSC_FALSE;
64 PetscCall(PetscFree(A->defaultvectype));
65 PetscCall(PetscStrallocpy(VECCUDA, &A->defaultvectype));
66
67 a = (Mat_MPISELL *)A->data;
68 if (a->A) PetscCall(MatSetType(a->A, MATSEQSELLCUDA));
69 if (a->B) PetscCall(MatSetType(a->B, MATSEQSELLCUDA));
70 if (a->lvec) PetscCall(VecSetType(a->lvec, VECSEQCUDA));
71
72 A->ops->assemblyend = MatAssemblyEnd_MPISELLCUDA;
73 A->ops->setfromoptions = MatSetFromOptions_MPISELLCUDA;
74 A->ops->destroy = MatDestroy_MPISELLCUDA;
75
76 PetscCall(PetscObjectChangeTypeName((PetscObject)A, MATMPISELLCUDA));
77 PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatMPISELLSetPreallocation_C", MatMPISELLSetPreallocation_MPISELLCUDA));
78 PetscFunctionReturn(PETSC_SUCCESS);
79 }
80
MatCreate_MPISELLCUDA(Mat A)81 PETSC_EXTERN PetscErrorCode MatCreate_MPISELLCUDA(Mat A)
82 {
83 PetscFunctionBegin;
84 PetscCall(PetscDeviceInitialize(PETSC_DEVICE_CUDA));
85 PetscCall(MatCreate_MPISELL(A));
86 PetscCall(MatConvert_MPISELL_MPISELLCUDA(A, MATMPISELLCUDA, MAT_INPLACE_MATRIX, &A));
87 PetscFunctionReturn(PETSC_SUCCESS);
88 }
89
90 /*@
91 MatCreateSELLCUDA - Creates a sparse matrix in SELL format.
92 This matrix will ultimately pushed down to NVIDIA GPUs.
93
94 Collective
95
96 Input Parameters:
97 + comm - MPI communicator, set to `PETSC_COMM_SELF`
98 . m - number of local rows (or `PETSC_DECIDE` to have calculated if `M` is given)
99 This value should be the same as the local size used in creating the
100 y vector for the matrix-vector product y = Ax.
101 . n - This value should be the same as the local size used in creating the
102 x vector for the matrix-vector product y = Ax. (or PETSC_DECIDE to have
103 calculated if `N` is given) For square matrices `n` is almost always `m`.
104 . M - number of global rows (or `PETSC_DETERMINE` to have calculated if `m` is given)
105 . N - number of global columns (or `PETSC_DETERMINE` to have calculated if `n` is given)
106 . d_nz - number of nonzeros per row in DIAGONAL portion of local submatrix
107 (same value is used for all local rows)
108 . d_nnz - array containing the number of nonzeros in the various rows of the
109 DIAGONAL portion of the local submatrix (possibly different for each row)
110 or `NULL`, if `d_nz` is used to specify the nonzero structure.
111 The size of this array is equal to the number of local rows, i.e `m`.
112 For matrices you plan to factor you must leave room for the diagonal entry and
113 put in the entry even if it is zero.
114 . o_nz - number of nonzeros per row in the OFF-DIAGONAL portion of local
115 submatrix (same value is used for all local rows).
116 - o_nnz - array containing the number of nonzeros in the various rows of the
117 OFF-DIAGONAL portion of the local submatrix (possibly different for
118 each row) or `NULL`, if `o_nz` is used to specify the nonzero
119 structure. The size of this array is equal to the number
120 of local rows, i.e `m`.
121
122 Output Parameter:
123 . A - the matrix
124
125 Level: intermediate
126
127 Notes:
128 If `nnz` is given then `nz` is ignored
129
130 Specify the preallocated storage with either `nz` or `nnz` (not both).
131 Set `nz` = `PETSC_DEFAULT` and `nnz` = `NULL` for PETSc to control dynamic memory
132 allocation.
133
134 .seealso: [](ch_matrices), `Mat`, `MatCreate()`, `MatCreateSELL()`, `MatSetValues()`, `MATMPISELLCUDA`, `MATSELLCUDA`
135 @*/
MatCreateSELLCUDA(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)136 PetscErrorCode MatCreateSELLCUDA(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)
137 {
138 PetscMPIInt size;
139
140 PetscFunctionBegin;
141 PetscCall(MatCreate(comm, A));
142 PetscCall(MatSetSizes(*A, m, n, M, N));
143 PetscCallMPI(MPI_Comm_size(comm, &size));
144 if (size > 1) {
145 PetscCall(MatSetType(*A, MATMPISELLCUDA));
146 PetscCall(MatMPISELLSetPreallocation(*A, d_nz, d_nnz, o_nz, o_nnz));
147 } else {
148 PetscCall(MatSetType(*A, MATSEQSELLCUDA));
149 PetscCall(MatSeqSELLSetPreallocation(*A, d_nz, d_nnz));
150 }
151 PetscFunctionReturn(PETSC_SUCCESS);
152 }
153
154 /*MC
155 MATSELLCUDA - "sellcuda" = "mpisellcuda" - A matrix type to be used for sparse matrices.
156
157 Sliced ELLPACK matrix type whose data resides on NVIDIA GPUs.
158
159 This matrix type is identical to `MATSEQSELLCUDA` when constructed with a single process communicator,
160 and `MATMPISELLCUDA` otherwise. As a result, for single process communicators,
161 `MatSeqSELLSetPreallocation()` is supported, and similarly `MatMPISELLSetPreallocation()` is supported
162 for communicators controlling multiple processes. It is recommended that you call both of
163 the above preallocation routines for simplicity.
164
165 Options Database Key:
166 . -mat_type mpisellcuda - sets the matrix type to `MATMPISELLCUDA` during a call to MatSetFromOptions()
167
168 Level: beginner
169
170 .seealso: `MatCreateSELLCUDA()`, `MATSEQSELLCUDA`, `MatCreateSeqSELLCUDA()`, `MatCUDAFormatOperation()`
171 M*/
172