xref: /petsc/src/mat/impls/aij/seq/fdaij.c (revision a1cb98fac0cdf0eb4d3e8a0c8b58f3fe8f800bc6)
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     brows = 1000 / bcols;
39     if (bcols > nis) bcols = nis;
40     if (brows == 0 || brows > m) brows = m;
41     c->brows = brows;
42     c->bcols = bcols;
43   }
44 
45   c->M       = mat->rmap->N / bs; /* set total rows, columns and local rows */
46   c->N       = mat->cmap->N / bs;
47   c->m       = mat->rmap->N / bs;
48   c->rstart  = 0;
49   c->ncolors = nis;
50   c->ctype   = iscoloring->ctype;
51   PetscFunctionReturn(0);
52 }
53 
54 /*
55  Reorder Jentry such that blocked brows*bols of entries from dense matrix are inserted into Jacobian for improved cache performance
56    Input Parameters:
57 +  mat - the matrix containing the nonzero structure of the Jacobian
58 .  color - the coloring context
59 -  nz - number of local non-zeros in mat
60 */
61 PetscErrorCode MatFDColoringSetUpBlocked_AIJ_Private(Mat mat, MatFDColoring c, PetscInt nz)
62 {
63   PetscInt  i, j, nrows, nbcols, brows = c->brows, bcols = c->bcols, mbs = c->m, nis = c->ncolors;
64   PetscInt *color_start, *row_start, *nrows_new, nz_new, row_end;
65 
66   PetscFunctionBegin;
67   if (brows < 1 || brows > mbs) brows = mbs;
68   PetscCall(PetscMalloc2(bcols + 1, &color_start, bcols, &row_start));
69   PetscCall(PetscCalloc1(nis, &nrows_new));
70   PetscCall(PetscMalloc1(bcols * mat->rmap->n, &c->dy));
71 
72   nz_new             = 0;
73   nbcols             = 0;
74   color_start[bcols] = 0;
75 
76   if (c->htype[0] == 'd') { /* ----  c->htype == 'ds', use MatEntry --------*/
77     MatEntry *Jentry_new, *Jentry = c->matentry;
78 
79     PetscCall(PetscMalloc1(nz, &Jentry_new));
80     for (i = 0; i < nis; i += bcols) { /* loop over colors */
81       if (i + bcols > nis) {
82         color_start[nis - i] = color_start[bcols];
83         bcols                = nis - i;
84       }
85 
86       color_start[0] = color_start[bcols];
87       for (j = 0; j < bcols; j++) {
88         color_start[j + 1] = c->nrows[i + j] + color_start[j];
89         row_start[j]       = 0;
90       }
91 
92       row_end = brows;
93       if (row_end > mbs) row_end = mbs;
94 
95       while (row_end <= mbs) {        /* loop over block rows */
96         for (j = 0; j < bcols; j++) { /* loop over block columns */
97           nrows = c->nrows[i + j];
98           nz    = color_start[j];
99           while (row_start[j] < nrows) {
100             if (Jentry[nz].row >= row_end) {
101               color_start[j] = nz;
102               break;
103             } else {                                                 /* copy Jentry[nz] to Jentry_new[nz_new] */
104               Jentry_new[nz_new].row     = Jentry[nz].row + j * mbs; /* index in dy-array */
105               Jentry_new[nz_new].col     = Jentry[nz].col;
106               Jentry_new[nz_new].valaddr = Jentry[nz].valaddr;
107               nz_new++;
108               nz++;
109               row_start[j]++;
110             }
111           }
112         }
113         if (row_end == mbs) break;
114         row_end += brows;
115         if (row_end > mbs) row_end = mbs;
116       }
117       nrows_new[nbcols++] = nz_new;
118     }
119     PetscCall(PetscFree(Jentry));
120     c->matentry = Jentry_new;
121   } else { /* ---------  c->htype == 'wp', use MatEntry2 ------------------*/
122     MatEntry2 *Jentry2_new, *Jentry2 = c->matentry2;
123 
124     PetscCall(PetscMalloc1(nz, &Jentry2_new));
125     for (i = 0; i < nis; i += bcols) { /* loop over colors */
126       if (i + bcols > nis) {
127         color_start[nis - i] = color_start[bcols];
128         bcols                = nis - i;
129       }
130 
131       color_start[0] = color_start[bcols];
132       for (j = 0; j < bcols; j++) {
133         color_start[j + 1] = c->nrows[i + j] + color_start[j];
134         row_start[j]       = 0;
135       }
136 
137       row_end = brows;
138       if (row_end > mbs) row_end = mbs;
139 
140       while (row_end <= mbs) {        /* loop over block rows */
141         for (j = 0; j < bcols; j++) { /* loop over block columns */
142           nrows = c->nrows[i + j];
143           nz    = color_start[j];
144           while (row_start[j] < nrows) {
145             if (Jentry2[nz].row >= row_end) {
146               color_start[j] = nz;
147               break;
148             } else {                                                   /* copy Jentry2[nz] to Jentry2_new[nz_new] */
149               Jentry2_new[nz_new].row     = Jentry2[nz].row + j * mbs; /* index in dy-array */
150               Jentry2_new[nz_new].valaddr = Jentry2[nz].valaddr;
151               nz_new++;
152               nz++;
153               row_start[j]++;
154             }
155           }
156         }
157         if (row_end == mbs) break;
158         row_end += brows;
159         if (row_end > mbs) row_end = mbs;
160       }
161       nrows_new[nbcols++] = nz_new;
162     }
163     PetscCall(PetscFree(Jentry2));
164     c->matentry2 = Jentry2_new;
165   } /* ---------------------------------------------*/
166 
167   PetscCall(PetscFree2(color_start, row_start));
168 
169   for (i = nbcols - 1; i > 0; i--) nrows_new[i] -= nrows_new[i - 1];
170   PetscCall(PetscFree(c->nrows));
171   c->nrows = nrows_new;
172   PetscFunctionReturn(0);
173 }
174 
175 PetscErrorCode MatFDColoringSetUp_SeqXAIJ(Mat mat, ISColoring iscoloring, MatFDColoring c)
176 {
177   PetscInt           i, n, nrows, mbs = c->m, j, k, m, ncols, col, nis = iscoloring->n, *rowhit, bs, bs2, *spidx, nz, tmp;
178   const PetscInt    *is, *row, *ci, *cj;
179   PetscBool          isBAIJ, isSELL;
180   const PetscScalar *A_val;
181   PetscScalar      **valaddrhit;
182   MatEntry          *Jentry;
183   MatEntry2         *Jentry2;
184 
185   PetscFunctionBegin;
186   PetscCall(ISColoringGetIS(iscoloring, PETSC_OWN_POINTER, PETSC_IGNORE, &c->isa));
187 
188   PetscCall(MatGetBlockSize(mat, &bs));
189   PetscCall(PetscObjectBaseTypeCompare((PetscObject)mat, MATSEQBAIJ, &isBAIJ));
190   PetscCall(PetscObjectTypeCompare((PetscObject)mat, MATSEQSELL, &isSELL));
191   if (isBAIJ) {
192     Mat_SeqBAIJ *spA = (Mat_SeqBAIJ *)mat->data;
193 
194     A_val = spA->a;
195     nz    = spA->nz;
196   } else if (isSELL) {
197     Mat_SeqSELL *spA = (Mat_SeqSELL *)mat->data;
198 
199     A_val = spA->val;
200     nz    = spA->nz;
201     bs    = 1; /* only bs=1 is supported for SeqSELL matrix */
202   } else {
203     Mat_SeqAIJ *spA = (Mat_SeqAIJ *)mat->data;
204 
205     A_val = spA->a;
206     nz    = spA->nz;
207     bs    = 1; /* only bs=1 is supported for SeqAIJ matrix */
208   }
209 
210   PetscCall(PetscMalloc2(nis, &c->ncolumns, nis, &c->columns));
211   PetscCall(PetscMalloc1(nis, &c->nrows)); /* nrows is freed separately from ncolumns and columns */
212 
213   if (c->htype[0] == 'd') {
214     PetscCall(PetscMalloc1(nz, &Jentry));
215     c->matentry = Jentry;
216   } else if (c->htype[0] == 'w') {
217     PetscCall(PetscMalloc1(nz, &Jentry2));
218     c->matentry2 = Jentry2;
219   } else SETERRQ(PetscObjectComm((PetscObject)mat), PETSC_ERR_SUP, "htype is not supported");
220 
221   if (isBAIJ) {
222     PetscCall(MatGetColumnIJ_SeqBAIJ_Color(mat, 0, PETSC_FALSE, PETSC_FALSE, &ncols, &ci, &cj, &spidx, NULL));
223   } else if (isSELL) {
224     PetscCall(MatGetColumnIJ_SeqSELL_Color(mat, 0, PETSC_FALSE, PETSC_FALSE, &ncols, &ci, &cj, &spidx, NULL));
225   } else {
226     PetscCall(MatGetColumnIJ_SeqAIJ_Color(mat, 0, PETSC_FALSE, PETSC_FALSE, &ncols, &ci, &cj, &spidx, NULL));
227   }
228 
229   PetscCall(PetscCalloc1(c->m, &rowhit));
230   PetscCall(PetscMalloc1(c->m, &valaddrhit));
231 
232   nz = 0;
233   for (i = 0; i < nis; i++) { /* loop over colors */
234     PetscCall(ISGetLocalSize(c->isa[i], &n));
235     PetscCall(ISGetIndices(c->isa[i], &is));
236 
237     c->ncolumns[i] = n;
238     c->columns[i]  = (PetscInt *)is;
239     /* note: we know that c->isa is going to be around as long at the c->columns values */
240     PetscCall(ISRestoreIndices(c->isa[i], &is));
241 
242     /* fast, crude version requires O(N*N) work */
243     bs2   = bs * bs;
244     nrows = 0;
245     for (j = 0; j < n; j++) { /* loop over columns */
246       col = is[j];
247       tmp = ci[col];
248       row = cj + tmp;
249       m   = ci[col + 1] - tmp;
250       nrows += m;
251       for (k = 0; k < m; k++) { /* loop over columns marking them in rowhit */
252         rowhit[*row]       = col + 1;
253         valaddrhit[*row++] = (PetscScalar *)&A_val[bs2 * spidx[tmp + k]];
254       }
255     }
256     c->nrows[i] = nrows; /* total num of rows for this color */
257 
258     if (c->htype[0] == 'd') {
259       for (j = 0; j < mbs; j++) { /* loop over rows */
260         if (rowhit[j]) {
261           Jentry[nz].row     = j;             /* local row index */
262           Jentry[nz].col     = rowhit[j] - 1; /* local column index */
263           Jentry[nz].valaddr = valaddrhit[j]; /* address of mat value for this entry */
264           nz++;
265           rowhit[j] = 0.0; /* zero rowhit for reuse */
266         }
267       }
268     } else {                      /* c->htype == 'wp' */
269       for (j = 0; j < mbs; j++) { /* loop over rows */
270         if (rowhit[j]) {
271           Jentry2[nz].row     = j;             /* local row index */
272           Jentry2[nz].valaddr = valaddrhit[j]; /* address of mat value for this entry */
273           nz++;
274           rowhit[j] = 0.0; /* zero rowhit for reuse */
275         }
276       }
277     }
278   }
279 
280   if (c->bcols > 1) { /* reorder Jentry for faster MatFDColoringApply() */
281     PetscCall(MatFDColoringSetUpBlocked_AIJ_Private(mat, c, nz));
282   }
283 
284   if (isBAIJ) {
285     PetscCall(MatRestoreColumnIJ_SeqBAIJ_Color(mat, 0, PETSC_FALSE, PETSC_FALSE, &ncols, &ci, &cj, &spidx, NULL));
286     PetscCall(PetscMalloc1(bs * mat->rmap->n, &c->dy));
287   } else if (isSELL) {
288     PetscCall(MatRestoreColumnIJ_SeqSELL_Color(mat, 0, PETSC_FALSE, PETSC_FALSE, &ncols, &ci, &cj, &spidx, NULL));
289   } else {
290     PetscCall(MatRestoreColumnIJ_SeqAIJ_Color(mat, 0, PETSC_FALSE, PETSC_FALSE, &ncols, &ci, &cj, &spidx, NULL));
291   }
292   PetscCall(PetscFree(rowhit));
293   PetscCall(PetscFree(valaddrhit));
294   PetscCall(ISColoringRestoreIS(iscoloring, PETSC_OWN_POINTER, &c->isa));
295 
296   PetscCall(VecCreateGhost(PetscObjectComm((PetscObject)mat), mat->rmap->n, PETSC_DETERMINE, 0, NULL, &c->vscale));
297   PetscCall(PetscInfo(c, "ncolors %" PetscInt_FMT ", brows %" PetscInt_FMT " and bcols %" PetscInt_FMT " are used.\n", c->ncolors, c->brows, c->bcols));
298   PetscFunctionReturn(0);
299 }
300