xref: /petsc/src/mat/impls/aij/seq/fdaij.c (revision 21e3ffae2f3b73c0bd738cf6d0a809700fc04bb0)
1 #include <../src/mat/impls/aij/seq/aij.h>
2 #include <../src/mat/impls/baij/seq/baij.h>
3 #include <../src/mat/impls/sell/seq/sell.h>
4 #include <petsc/private/isimpl.h>
5 
6 /*
7     This routine is shared by SeqAIJ and SeqBAIJ matrices,
8     since it operators only on the nonzero structure of the elements or blocks.
9 */
10 PetscErrorCode MatFDColoringCreate_SeqXAIJ(Mat mat, ISColoring iscoloring, MatFDColoring c)
11 {
12   PetscInt  bs, nis = iscoloring->n, m = mat->rmap->n;
13   PetscBool isBAIJ, isSELL;
14 
15   PetscFunctionBegin;
16   /* set default brows and bcols for speedup inserting the dense matrix into sparse Jacobian */
17   PetscCall(MatGetBlockSize(mat, &bs));
18   PetscCall(PetscObjectBaseTypeCompare((PetscObject)mat, MATSEQBAIJ, &isBAIJ));
19   PetscCall(PetscObjectTypeCompare((PetscObject)mat, MATSEQSELL, &isSELL));
20   if (isBAIJ) {
21     c->brows = m;
22     c->bcols = 1;
23   } else { /* seqaij matrix */
24     /* bcols is chosen s.t. dy-array takes 50% of memory space as mat */
25     PetscReal mem;
26     PetscInt  nz, brows, bcols;
27     if (isSELL) {
28       Mat_SeqSELL *spA = (Mat_SeqSELL *)mat->data;
29       nz               = spA->nz;
30     } else {
31       Mat_SeqAIJ *spA = (Mat_SeqAIJ *)mat->data;
32       nz              = spA->nz;
33     }
34 
35     bs    = 1; /* only bs=1 is supported for SeqAIJ matrix */
36     mem   = nz * (sizeof(PetscScalar) + sizeof(PetscInt)) + 3 * m * sizeof(PetscInt);
37     bcols = (PetscInt)(0.5 * mem / (m * sizeof(PetscScalar)));
38     if (!bcols) bcols = 1;
39     brows = 1000 / bcols;
40     if (bcols > nis) bcols = nis;
41     if (brows == 0 || brows > m) brows = m;
42     c->brows = brows;
43     c->bcols = bcols;
44   }
45 
46   c->M       = mat->rmap->N / bs; /* set total rows, columns and local rows */
47   c->N       = mat->cmap->N / bs;
48   c->m       = mat->rmap->N / bs;
49   c->rstart  = 0;
50   c->ncolors = nis;
51   c->ctype   = iscoloring->ctype;
52   PetscFunctionReturn(PETSC_SUCCESS);
53 }
54 
55 /*
56  Reorder Jentry such that blocked brows*bols of entries from dense matrix are inserted into Jacobian for improved cache performance
57    Input Parameters:
58 +  mat - the matrix containing the nonzero structure of the Jacobian
59 .  color - the coloring context
60 -  nz - number of local non-zeros in mat
61 */
62 PetscErrorCode MatFDColoringSetUpBlocked_AIJ_Private(Mat mat, MatFDColoring c, PetscInt nz)
63 {
64   PetscInt  i, j, nrows, nbcols, brows = c->brows, bcols = c->bcols, mbs = c->m, nis = c->ncolors;
65   PetscInt *color_start, *row_start, *nrows_new, nz_new, row_end;
66 
67   PetscFunctionBegin;
68   if (brows < 1 || brows > mbs) brows = mbs;
69   PetscCall(PetscMalloc2(bcols + 1, &color_start, bcols, &row_start));
70   PetscCall(PetscCalloc1(nis, &nrows_new));
71   PetscCall(PetscMalloc1(bcols * mat->rmap->n, &c->dy));
72 
73   nz_new             = 0;
74   nbcols             = 0;
75   color_start[bcols] = 0;
76 
77   if (c->htype[0] == 'd') { /* ----  c->htype == 'ds', use MatEntry --------*/
78     MatEntry *Jentry_new, *Jentry = c->matentry;
79 
80     PetscCall(PetscMalloc1(nz, &Jentry_new));
81     for (i = 0; i < nis; i += bcols) { /* loop over colors */
82       if (i + bcols > nis) {
83         color_start[nis - i] = color_start[bcols];
84         bcols                = nis - i;
85       }
86 
87       color_start[0] = color_start[bcols];
88       for (j = 0; j < bcols; j++) {
89         color_start[j + 1] = c->nrows[i + j] + color_start[j];
90         row_start[j]       = 0;
91       }
92 
93       row_end = brows;
94       if (row_end > mbs) row_end = mbs;
95 
96       while (row_end <= mbs) {        /* loop over block rows */
97         for (j = 0; j < bcols; j++) { /* loop over block columns */
98           nrows = c->nrows[i + j];
99           nz    = color_start[j];
100           while (row_start[j] < nrows) {
101             if (Jentry[nz].row >= row_end) {
102               color_start[j] = nz;
103               break;
104             } else {                                                 /* copy Jentry[nz] to Jentry_new[nz_new] */
105               Jentry_new[nz_new].row     = Jentry[nz].row + j * mbs; /* index in dy-array */
106               Jentry_new[nz_new].col     = Jentry[nz].col;
107               Jentry_new[nz_new].valaddr = Jentry[nz].valaddr;
108               nz_new++;
109               nz++;
110               row_start[j]++;
111             }
112           }
113         }
114         if (row_end == mbs) break;
115         row_end += brows;
116         if (row_end > mbs) row_end = mbs;
117       }
118       nrows_new[nbcols++] = nz_new;
119     }
120     PetscCall(PetscFree(Jentry));
121     c->matentry = Jentry_new;
122   } else { /* ---------  c->htype == 'wp', use MatEntry2 ------------------*/
123     MatEntry2 *Jentry2_new, *Jentry2 = c->matentry2;
124 
125     PetscCall(PetscMalloc1(nz, &Jentry2_new));
126     for (i = 0; i < nis; i += bcols) { /* loop over colors */
127       if (i + bcols > nis) {
128         color_start[nis - i] = color_start[bcols];
129         bcols                = nis - i;
130       }
131 
132       color_start[0] = color_start[bcols];
133       for (j = 0; j < bcols; j++) {
134         color_start[j + 1] = c->nrows[i + j] + color_start[j];
135         row_start[j]       = 0;
136       }
137 
138       row_end = brows;
139       if (row_end > mbs) row_end = mbs;
140 
141       while (row_end <= mbs) {        /* loop over block rows */
142         for (j = 0; j < bcols; j++) { /* loop over block columns */
143           nrows = c->nrows[i + j];
144           nz    = color_start[j];
145           while (row_start[j] < nrows) {
146             if (Jentry2[nz].row >= row_end) {
147               color_start[j] = nz;
148               break;
149             } else {                                                   /* copy Jentry2[nz] to Jentry2_new[nz_new] */
150               Jentry2_new[nz_new].row     = Jentry2[nz].row + j * mbs; /* index in dy-array */
151               Jentry2_new[nz_new].valaddr = Jentry2[nz].valaddr;
152               nz_new++;
153               nz++;
154               row_start[j]++;
155             }
156           }
157         }
158         if (row_end == mbs) break;
159         row_end += brows;
160         if (row_end > mbs) row_end = mbs;
161       }
162       nrows_new[nbcols++] = nz_new;
163     }
164     PetscCall(PetscFree(Jentry2));
165     c->matentry2 = Jentry2_new;
166   } /* ---------------------------------------------*/
167 
168   PetscCall(PetscFree2(color_start, row_start));
169 
170   for (i = nbcols - 1; i > 0; i--) nrows_new[i] -= nrows_new[i - 1];
171   PetscCall(PetscFree(c->nrows));
172   c->nrows = nrows_new;
173   PetscFunctionReturn(PETSC_SUCCESS);
174 }
175 
176 PetscErrorCode MatFDColoringSetUp_SeqXAIJ(Mat mat, ISColoring iscoloring, MatFDColoring c)
177 {
178   PetscInt           i, n, nrows, mbs = c->m, j, k, m, ncols, col, nis = iscoloring->n, *rowhit, bs, bs2, *spidx, nz, tmp;
179   const PetscInt    *is, *row, *ci, *cj;
180   PetscBool          isBAIJ, isSELL;
181   const PetscScalar *A_val;
182   PetscScalar      **valaddrhit;
183   MatEntry          *Jentry;
184   MatEntry2         *Jentry2;
185 
186   PetscFunctionBegin;
187   PetscCall(ISColoringGetIS(iscoloring, PETSC_OWN_POINTER, PETSC_IGNORE, &c->isa));
188 
189   PetscCall(MatGetBlockSize(mat, &bs));
190   PetscCall(PetscObjectBaseTypeCompare((PetscObject)mat, MATSEQBAIJ, &isBAIJ));
191   PetscCall(PetscObjectTypeCompare((PetscObject)mat, MATSEQSELL, &isSELL));
192   if (isBAIJ) {
193     Mat_SeqBAIJ *spA = (Mat_SeqBAIJ *)mat->data;
194 
195     A_val = spA->a;
196     nz    = spA->nz;
197   } else if (isSELL) {
198     Mat_SeqSELL *spA = (Mat_SeqSELL *)mat->data;
199 
200     A_val = spA->val;
201     nz    = spA->nz;
202     bs    = 1; /* only bs=1 is supported for SeqSELL matrix */
203   } else {
204     Mat_SeqAIJ *spA = (Mat_SeqAIJ *)mat->data;
205 
206     A_val = spA->a;
207     nz    = spA->nz;
208     bs    = 1; /* only bs=1 is supported for SeqAIJ matrix */
209   }
210 
211   PetscCall(PetscMalloc2(nis, &c->ncolumns, nis, &c->columns));
212   PetscCall(PetscMalloc1(nis, &c->nrows)); /* nrows is freed separately from ncolumns and columns */
213 
214   if (c->htype[0] == 'd') {
215     PetscCall(PetscMalloc1(nz, &Jentry));
216     c->matentry = Jentry;
217   } else if (c->htype[0] == 'w') {
218     PetscCall(PetscMalloc1(nz, &Jentry2));
219     c->matentry2 = Jentry2;
220   } else SETERRQ(PetscObjectComm((PetscObject)mat), PETSC_ERR_SUP, "htype is not supported");
221 
222   if (isBAIJ) {
223     PetscCall(MatGetColumnIJ_SeqBAIJ_Color(mat, 0, PETSC_FALSE, PETSC_FALSE, &ncols, &ci, &cj, &spidx, NULL));
224   } else if (isSELL) {
225     PetscCall(MatGetColumnIJ_SeqSELL_Color(mat, 0, PETSC_FALSE, PETSC_FALSE, &ncols, &ci, &cj, &spidx, NULL));
226   } else {
227     PetscCall(MatGetColumnIJ_SeqAIJ_Color(mat, 0, PETSC_FALSE, PETSC_FALSE, &ncols, &ci, &cj, &spidx, NULL));
228   }
229 
230   PetscCall(PetscCalloc1(c->m, &rowhit));
231   PetscCall(PetscMalloc1(c->m, &valaddrhit));
232 
233   nz = 0;
234   for (i = 0; i < nis; i++) { /* loop over colors */
235     PetscCall(ISGetLocalSize(c->isa[i], &n));
236     PetscCall(ISGetIndices(c->isa[i], &is));
237 
238     c->ncolumns[i] = n;
239     c->columns[i]  = (PetscInt *)is;
240     /* note: we know that c->isa is going to be around as long at the c->columns values */
241     PetscCall(ISRestoreIndices(c->isa[i], &is));
242 
243     /* fast, crude version requires O(N*N) work */
244     bs2   = bs * bs;
245     nrows = 0;
246     for (j = 0; j < n; j++) { /* loop over columns */
247       col = is[j];
248       tmp = ci[col];
249       row = cj + tmp;
250       m   = ci[col + 1] - tmp;
251       nrows += m;
252       for (k = 0; k < m; k++) { /* loop over columns marking them in rowhit */
253         rowhit[*row]       = col + 1;
254         valaddrhit[*row++] = (PetscScalar *)&A_val[bs2 * spidx[tmp + k]];
255       }
256     }
257     c->nrows[i] = nrows; /* total num of rows for this color */
258 
259     if (c->htype[0] == 'd') {
260       for (j = 0; j < mbs; j++) { /* loop over rows */
261         if (rowhit[j]) {
262           Jentry[nz].row     = j;             /* local row index */
263           Jentry[nz].col     = rowhit[j] - 1; /* local column index */
264           Jentry[nz].valaddr = valaddrhit[j]; /* address of mat value for this entry */
265           nz++;
266           rowhit[j] = 0.0; /* zero rowhit for reuse */
267         }
268       }
269     } else {                      /* c->htype == 'wp' */
270       for (j = 0; j < mbs; j++) { /* loop over rows */
271         if (rowhit[j]) {
272           Jentry2[nz].row     = j;             /* local row index */
273           Jentry2[nz].valaddr = valaddrhit[j]; /* address of mat value for this entry */
274           nz++;
275           rowhit[j] = 0.0; /* zero rowhit for reuse */
276         }
277       }
278     }
279   }
280 
281   if (c->bcols > 1) { /* reorder Jentry for faster MatFDColoringApply() */
282     PetscCall(MatFDColoringSetUpBlocked_AIJ_Private(mat, c, nz));
283   }
284 
285   if (isBAIJ) {
286     PetscCall(MatRestoreColumnIJ_SeqBAIJ_Color(mat, 0, PETSC_FALSE, PETSC_FALSE, &ncols, &ci, &cj, &spidx, NULL));
287     PetscCall(PetscMalloc1(bs * mat->rmap->n, &c->dy));
288   } else if (isSELL) {
289     PetscCall(MatRestoreColumnIJ_SeqSELL_Color(mat, 0, PETSC_FALSE, PETSC_FALSE, &ncols, &ci, &cj, &spidx, NULL));
290   } else {
291     PetscCall(MatRestoreColumnIJ_SeqAIJ_Color(mat, 0, PETSC_FALSE, PETSC_FALSE, &ncols, &ci, &cj, &spidx, NULL));
292   }
293   PetscCall(PetscFree(rowhit));
294   PetscCall(PetscFree(valaddrhit));
295   PetscCall(ISColoringRestoreIS(iscoloring, PETSC_OWN_POINTER, &c->isa));
296 
297   PetscCall(VecCreateGhost(PetscObjectComm((PetscObject)mat), mat->rmap->n, PETSC_DETERMINE, 0, NULL, &c->vscale));
298   PetscCall(PetscInfo(c, "ncolors %" PetscInt_FMT ", brows %" PetscInt_FMT " and bcols %" PetscInt_FMT " are used.\n", c->ncolors, c->brows, c->bcols));
299   PetscFunctionReturn(PETSC_SUCCESS);
300 }
301