Actual source code: mpihashmat.h
1: /*
2: used by MPIAIJ, BAIJ and SBAIJ to reduce code duplication
4: define TYPE to AIJ BAIJ or SBAIJ
5: TYPE_SBAIJ for SBAIJ matrix
7: */
9: static PetscErrorCode MatSetValues_MPI_Hash(Mat A, PetscInt m, const PetscInt *rows, PetscInt n, const PetscInt *cols, const PetscScalar *values, InsertMode addv)
10: {
11: PetscConcat(Mat_MPI, TYPE) *a = (PetscConcat(Mat_MPI, TYPE) *)A->data;
12: const PetscInt rStart = A->rmap->rstart;
13: const PetscInt rEnd = A->rmap->rend;
14: const PetscInt cStart = A->cmap->rstart;
15: const PetscInt cEnd = A->cmap->rend;
16: #if defined(TYPE_SBAIJ)
17: const PetscInt bs = A->rmap->bs;
18: #endif
19: const PetscBool ignorezeroentries = ((Mat_SeqAIJ *)a->A->data)->ignorezeroentries;
21: PetscFunctionBegin;
22: for (PetscInt r = 0; r < m; ++r) {
23: PetscScalar value;
24: if (rows[r] < 0) continue;
25: if (rows[r] < rStart || rows[r] >= rEnd) {
26: PetscCheck(!A->nooffprocentries, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Setting off process row %" PetscInt_FMT " even though MatSetOption(,MAT_NO_OFF_PROC_ENTRIES,PETSC_TRUE) was set", rows[r]);
27: if (!a->donotstash) {
28: A->assembled = PETSC_FALSE;
29: if (a->roworiented) {
30: PetscCall(MatStashValuesRow_Private(&A->stash, rows[r], n, cols, PetscSafePointerPlusOffset(values, r * n), (PetscBool)(ignorezeroentries && (addv == ADD_VALUES))));
31: } else {
32: PetscCall(MatStashValuesCol_Private(&A->stash, rows[r], n, cols, PetscSafePointerPlusOffset(values, r), m, (PetscBool)(ignorezeroentries && (addv == ADD_VALUES))));
33: }
34: }
35: } else {
36: for (PetscInt c = 0; c < n; ++c) {
37: #if defined(TYPE_SBAIJ)
38: if (cols[c] / bs < rows[r] / bs) continue;
39: #else
40: if (cols[c] < 0) continue;
41: #endif
42: value = values ? (a->roworiented ? values[r * n + c] : values[r + m * c]) : 0;
43: if (ignorezeroentries && value == 0.0 && (addv == ADD_VALUES) && rows[r] != cols[c]) continue;
44: if (cols[c] >= cStart && cols[c] < cEnd) {
45: PetscCall(MatSetValue(a->A, rows[r] - rStart, cols[c] - cStart, value, addv));
46: } else if (!ignorezeroentries || value != 0.0) {
47: /* MPIAIJ never inserts in B if it is 0 */
48: PetscCall(MatSetValue(a->B, rows[r] - rStart, cols[c], value, addv));
49: }
50: }
51: }
52: }
53: PetscFunctionReturn(PETSC_SUCCESS);
54: }
56: static PetscErrorCode MatAssemblyBegin_MPI_Hash(Mat A, PETSC_UNUSED MatAssemblyType type)
57: {
58: PetscConcat(Mat_MPI, TYPE) *a = (PetscConcat(Mat_MPI, TYPE) *)A->data;
59: PetscInt nstash, reallocs;
61: PetscFunctionBegin;
62: if (a->donotstash || A->nooffprocentries) PetscFunctionReturn(PETSC_SUCCESS);
63: PetscCall(MatStashScatterBegin_Private(A, &A->stash, A->rmap->range));
64: PetscCall(MatStashGetInfo_Private(&A->stash, &nstash, &reallocs));
65: PetscCall(PetscInfo(A, "Stash has %" PetscInt_FMT " entries, uses %" PetscInt_FMT " mallocs.\n", nstash, reallocs));
66: PetscFunctionReturn(PETSC_SUCCESS);
67: }
69: static PetscErrorCode MatFinishScatterAndSetValues_MPI_Hash(Mat A)
70: {
71: PetscConcat(Mat_MPI, TYPE) *a = (PetscConcat(Mat_MPI, TYPE) *)A->data;
72: PetscMPIInt n;
73: PetscScalar *val;
74: PetscInt *row, *col;
75: PetscInt j, ncols, flg, rstart;
77: PetscFunctionBegin;
78: if (!a->donotstash && !A->nooffprocentries) {
79: while (1) {
80: PetscCall(MatStashScatterGetMesg_Private(&A->stash, &n, &row, &col, &val, &flg));
81: if (!flg) break;
83: for (PetscInt i = 0; i < n;) {
84: /* Now identify the consecutive vals belonging to the same row */
85: for (j = i, rstart = row[j]; j < n; j++) {
86: if (row[j] != rstart) break;
87: }
88: if (j < n) ncols = j - i;
89: else ncols = n - i;
90: /* Now assemble all these values with a single function call */
91: PetscCall(MatSetValues_MPI_Hash(A, 1, row + i, ncols, col + i, val + i, A->insertmode));
92: i = j;
93: }
94: }
95: PetscCall(MatStashScatterEnd_Private(&A->stash));
96: }
97: PetscFunctionReturn(PETSC_SUCCESS);
98: }
100: static PetscErrorCode MatAssemblyEnd_MPI_Hash(Mat A, MatAssemblyType type)
101: {
102: PetscConcat(Mat_MPI, TYPE) *a = (PetscConcat(Mat_MPI, TYPE) *)A->data;
103: PetscBool nooffprocentries = A->nooffprocentries;
105: PetscFunctionBegin;
106: PetscCall(MatFinishScatterAndSetValues_MPI_Hash(A));
107: if (type != MAT_FINAL_ASSEMBLY) PetscFunctionReturn(PETSC_SUCCESS);
109: A->insertmode = NOT_SET_VALUES; /* this was set by the previous calls to MatSetValues() */
111: A->ops[0] = a->cops;
112: A->hash_active = PETSC_FALSE;
114: /* a->B must be in CSR before MatSetUpMultiply_MPIAIJ() runs inside MatAssemblyEnd_MPIAIJ().
115: a->A is assembled by MatAssemblyEnd_MPIAIJ() itself and does not need a separate call here. */
116: PetscCall(MatAssemblyBegin(a->B, MAT_FINAL_ASSEMBLY));
117: PetscCall(MatAssemblyEnd(a->B, MAT_FINAL_ASSEMBLY));
118: /* Off-process entries were already scattered in MatAssemblyBegin_MPI_Hash().
119: Suppress the stash scatter in MatAssemblyBegin_MPIAIJ() to avoid a second
120: PetscCommBuildTwoSidedFReq() call on an empty stash. */
121: A->nooffprocentries = PETSC_TRUE;
122: PetscCall(MatAssemblyBegin(A, MAT_FINAL_ASSEMBLY));
123: PetscCall(MatAssemblyEnd(A, MAT_FINAL_ASSEMBLY));
124: A->nooffprocentries = nooffprocentries;
125: PetscFunctionReturn(PETSC_SUCCESS);
126: }
128: static PetscErrorCode MatCopyHashToXAIJ_MPI_Hash(Mat A, Mat B)
129: {
130: PetscConcat(Mat_MPI, TYPE) *a = (PetscConcat(Mat_MPI, TYPE) *)A->data, *b = (PetscConcat(Mat_MPI, TYPE) *)B->data;
132: PetscFunctionBegin;
133: /* Let's figure there's no harm done in doing the scatters for A now even if A != B */
134: PetscCall(MatAssemblyBegin_MPI_Hash(A, /*unused*/ MAT_FINAL_ASSEMBLY));
135: PetscCall(MatFinishScatterAndSetValues_MPI_Hash(A));
137: PetscCall(MatCopyHashToXAIJ(a->A, b->A));
138: PetscCall(MatCopyHashToXAIJ(a->B, b->B));
139: PetscCall(MatAssemblyBegin(B, MAT_FINAL_ASSEMBLY));
140: PetscCall(MatAssemblyEnd(B, MAT_FINAL_ASSEMBLY));
141: PetscFunctionReturn(PETSC_SUCCESS);
142: }
144: static PetscErrorCode MatDestroy_MPI_Hash(Mat A)
145: {
146: PetscConcat(Mat_MPI, TYPE) *a = (PetscConcat(Mat_MPI, TYPE) *)A->data;
148: PetscFunctionBegin;
149: PetscCall(MatStashDestroy_Private(&A->stash));
150: PetscCall(MatDestroy(&a->A));
151: PetscCall(MatDestroy(&a->B));
152: PetscCall((*a->cops.destroy)(A));
153: PetscFunctionReturn(PETSC_SUCCESS);
154: }
156: static PetscErrorCode MatZeroEntries_MPI_Hash(PETSC_UNUSED Mat A)
157: {
158: PetscFunctionBegin;
159: PetscFunctionReturn(PETSC_SUCCESS);
160: }
162: static PetscErrorCode MatSetRandom_MPI_Hash(Mat A, PETSC_UNUSED PetscRandom r)
163: {
164: SETERRQ(PetscObjectComm((PetscObject)A), PETSC_ERR_ARG_WRONGSTATE, "Must set preallocation first");
165: }
167: static PetscErrorCode MatSetUp_MPI_Hash(Mat A)
168: {
169: PetscConcat(Mat_MPI, TYPE) *a = (PetscConcat(Mat_MPI, TYPE) *)A->data;
170: PetscMPIInt size;
171: #if !defined(TYPE_AIJ)
172: PetscInt bs;
173: #endif
175: PetscFunctionBegin;
176: PetscCall(PetscInfo(A, "Using hash-based MatSetValues() for MATMPI" PetscStringize(TYPE) " because no preallocation provided\n"));
177: PetscCall(PetscLayoutSetUp(A->rmap));
178: PetscCall(PetscLayoutSetUp(A->cmap));
179: if (A->rmap->bs < 1) A->rmap->bs = 1;
180: if (A->cmap->bs < 1) A->cmap->bs = 1;
181: PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)A), &size));
183: #if !defined(TYPE_AIJ)
184: PetscCall(MatGetBlockSize(A, &bs));
185: /* these values are set in MatMPISBAIJSetPreallocation() */
186: a->bs2 = bs * bs;
187: a->mbs = A->rmap->n / bs;
188: a->nbs = A->cmap->n / bs;
189: a->Mbs = A->rmap->N / bs;
190: a->Nbs = A->cmap->N / bs;
192: for (PetscInt i = 0; i <= a->size; i++) a->rangebs[i] = A->rmap->range[i] / bs;
193: a->rstartbs = A->rmap->rstart / bs;
194: a->rendbs = A->rmap->rend / bs;
195: a->cstartbs = A->cmap->rstart / bs;
196: a->cendbs = A->cmap->rend / bs;
197: PetscCall(MatStashCreate_Private(PetscObjectComm((PetscObject)A), A->rmap->bs, &A->bstash));
198: #endif
200: PetscCall(MatCreate(PETSC_COMM_SELF, &a->A));
201: PetscCall(MatSetSizes(a->A, A->rmap->n, A->cmap->n, A->rmap->n, A->cmap->n));
202: PetscCall(MatSetBlockSizesFromMats(a->A, A, A));
203: #if defined(SUB_TYPE_CUSPARSE)
204: PetscCall(MatSetType(a->A, MATSEQAIJCUSPARSE));
205: #else
206: PetscCall(MatSetType(a->A, PetscConcat(MATSEQ, TYPE)));
207: #endif
208: PetscCall(MatSetUp(a->A));
210: PetscCall(MatCreate(PETSC_COMM_SELF, &a->B));
211: PetscCall(MatSetSizes(a->B, A->rmap->n, size > 1 ? A->cmap->N : 0, A->rmap->n, size > 1 ? A->cmap->N : 0));
212: PetscCall(MatSetBlockSizesFromMats(a->B, A, A));
213: #if defined(TYPE_SBAIJ)
214: PetscCall(MatSetType(a->B, MATSEQBAIJ));
215: #else
216: #if defined(SUB_TYPE_CUSPARSE)
217: PetscCall(MatSetType(a->B, MATSEQAIJCUSPARSE));
218: #else
219: PetscCall(MatSetType(a->B, PetscConcat(MATSEQ, TYPE)));
220: #endif
221: #endif
222: PetscCall(MatSetUp(a->B));
224: /* keep a record of the operations so they can be reset when the hash handling is complete */
225: a->cops = A->ops[0];
226: A->ops->assemblybegin = MatAssemblyBegin_MPI_Hash;
227: A->ops->assemblyend = MatAssemblyEnd_MPI_Hash;
228: A->ops->setvalues = MatSetValues_MPI_Hash;
229: A->ops->destroy = MatDestroy_MPI_Hash;
230: A->ops->zeroentries = MatZeroEntries_MPI_Hash;
231: A->ops->setrandom = MatSetRandom_MPI_Hash;
232: A->ops->copyhashtoxaij = MatCopyHashToXAIJ_MPI_Hash;
233: A->ops->setvaluesblocked = NULL;
235: A->preallocated = PETSC_TRUE;
236: A->hash_active = PETSC_TRUE;
237: PetscFunctionReturn(PETSC_SUCCESS);
238: }