xref: /petsc/src/mat/impls/aij/mpi/fdmpiaij.c (revision 2205254efee3a00a594e5e2a3a70f74dcb40bc03)
1 
2 #include <../src/mat/impls/aij/mpi/mpiaij.h>
3 
4 extern PetscErrorCode MatCreateColmap_MPIAIJ_Private(Mat);
5 
6 #undef __FUNCT__
7 #define __FUNCT__ "MatFDColoringCreate_MPIAIJ"
8 PetscErrorCode MatFDColoringCreate_MPIAIJ(Mat mat,ISColoring iscoloring,MatFDColoring c)
9 {
10   Mat_MPIAIJ             *aij = (Mat_MPIAIJ*)mat->data;
11   PetscErrorCode         ierr;
12   PetscMPIInt            size,*ncolsonproc,*disp,nn;
13   PetscInt               i,n,nrows,j,k,m,ncols,col;
14   const PetscInt         *is,*A_ci,*A_cj,*B_ci,*B_cj,*rows = 0;
15   PetscInt               nis = iscoloring->n,nctot,*cols;
16   PetscInt               *rowhit,M,cstart,cend,colb;
17   PetscInt               *columnsforrow,l;
18   IS                     *isa;
19   PetscBool              done,flg;
20   ISLocalToGlobalMapping map   = mat->cmap->mapping;
21   PetscInt               *ltog = (map ? map->indices : (PetscInt*) PETSC_NULL),ctype=c->ctype;
22 
23   PetscFunctionBegin;
24   if (!mat->assembled) SETERRQ(((PetscObject)mat)->comm,PETSC_ERR_ARG_WRONGSTATE,"Matrix must be assembled first; MatAssemblyBegin/End();");
25   if (ctype == IS_COLORING_GHOSTED && !map) SETERRQ(((PetscObject)mat)->comm,PETSC_ERR_ARG_INCOMP,"When using ghosted differencing matrix must have local to global mapping provided with MatSetLocalToGlobalMapping");
26 
27   ierr = ISColoringGetIS(iscoloring,PETSC_IGNORE,&isa);CHKERRQ(ierr);
28 
29   M         = mat->rmap->n;
30   cstart    = mat->cmap->rstart;
31   cend      = mat->cmap->rend;
32   c->M      = mat->rmap->N;         /* set the global rows and columns and local rows */
33   c->N      = mat->cmap->N;
34   c->m      = mat->rmap->n;
35   c->rstart = mat->rmap->rstart;
36 
37   c->ncolors = nis;
38   ierr       = PetscMalloc(nis*sizeof(PetscInt),&c->ncolumns);CHKERRQ(ierr);
39   ierr       = PetscMalloc(nis*sizeof(PetscInt*),&c->columns);CHKERRQ(ierr);
40   ierr       = PetscMalloc(nis*sizeof(PetscInt),&c->nrows);CHKERRQ(ierr);
41   ierr       = PetscMalloc(nis*sizeof(PetscInt*),&c->rows);CHKERRQ(ierr);
42   ierr       = PetscMalloc(nis*sizeof(PetscInt*),&c->columnsforrow);CHKERRQ(ierr);
43   ierr       = PetscLogObjectMemory(c,5*nis*sizeof(PetscInt));CHKERRQ(ierr);
44 
45   /* Allow access to data structures of local part of matrix */
46   if (!aij->colmap) {
47     ierr = MatCreateColmap_MPIAIJ_Private(mat);CHKERRQ(ierr);
48   }
49   ierr = MatGetColumnIJ(aij->A,0,PETSC_FALSE,PETSC_FALSE,&ncols,&A_ci,&A_cj,&done);CHKERRQ(ierr);
50   ierr = MatGetColumnIJ(aij->B,0,PETSC_FALSE,PETSC_FALSE,&ncols,&B_ci,&B_cj,&done);CHKERRQ(ierr);
51 
52   ierr = PetscMalloc((M+1)*sizeof(PetscInt),&rowhit);CHKERRQ(ierr);
53   ierr = PetscMalloc((M+1)*sizeof(PetscInt),&columnsforrow);CHKERRQ(ierr);
54 
55   for (i=0; i<nis; i++) {
56     ierr = ISGetLocalSize(isa[i],&n);CHKERRQ(ierr);
57     ierr = ISGetIndices(isa[i],&is);CHKERRQ(ierr);
58 
59     c->ncolumns[i] = n;
60     if (n) {
61       ierr = PetscMalloc(n*sizeof(PetscInt),&c->columns[i]);CHKERRQ(ierr);
62       ierr = PetscLogObjectMemory(c,n*sizeof(PetscInt));CHKERRQ(ierr);
63       ierr = PetscMemcpy(c->columns[i],is,n*sizeof(PetscInt));CHKERRQ(ierr);
64     } else {
65       c->columns[i] = 0;
66     }
67 
68     if (ctype == IS_COLORING_GLOBAL) {
69       /* Determine the total (parallel) number of columns of this color */
70       ierr = MPI_Comm_size(((PetscObject)mat)->comm,&size);CHKERRQ(ierr);
71       ierr = PetscMalloc2(size,PetscMPIInt,&ncolsonproc,size,PetscMPIInt,&disp);CHKERRQ(ierr);
72 
73       ierr = PetscMPIIntCast(n,&nn);CHKERRQ(ierr);
74       ierr = MPI_Allgather(&nn,1,MPI_INT,ncolsonproc,1,MPI_INT,((PetscObject)mat)->comm);CHKERRQ(ierr);
75       nctot = 0; for (j=0; j<size; j++) nctot += ncolsonproc[j];
76       if (!nctot) {
77         ierr = PetscInfo(mat,"Coloring of matrix has some unneeded colors with no corresponding rows\n");CHKERRQ(ierr);
78       }
79 
80       disp[0] = 0;
81       for (j=1; j<size; j++) {
82         disp[j] = disp[j-1] + ncolsonproc[j-1];
83       }
84 
85       /* Get complete list of columns for color on each processor */
86       ierr = PetscMalloc((nctot+1)*sizeof(PetscInt),&cols);CHKERRQ(ierr);
87       ierr = MPI_Allgatherv((void*)is,n,MPIU_INT,cols,ncolsonproc,disp,MPIU_INT,((PetscObject)mat)->comm);CHKERRQ(ierr);
88       ierr = PetscFree2(ncolsonproc,disp);CHKERRQ(ierr);
89     } else if (ctype == IS_COLORING_GHOSTED) {
90       /* Determine local number of columns of this color on this process, including ghost points */
91       nctot = n;
92       ierr  = PetscMalloc((nctot+1)*sizeof(PetscInt),&cols);CHKERRQ(ierr);
93       ierr  = PetscMemcpy(cols,is,n*sizeof(PetscInt));CHKERRQ(ierr);
94     } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Not provided for this MatFDColoring type");
95 
96     /*
97        Mark all rows affect by these columns
98     */
99     /* Temporary option to allow for debugging/testing */
100     flg  = PETSC_FALSE;
101     ierr = PetscOptionsGetBool(PETSC_NULL,"-matfdcoloring_slow",&flg,PETSC_NULL);CHKERRQ(ierr);
102     if (!flg) { /*-----------------------------------------------------------------------------*/
103       /* crude, fast version */
104       ierr = PetscMemzero(rowhit,M*sizeof(PetscInt));CHKERRQ(ierr);
105       /* loop over columns*/
106       for (j=0; j<nctot; j++) {
107         if (ctype == IS_COLORING_GHOSTED) {
108           col = ltog[cols[j]];
109         } else {
110           col  = cols[j];
111         }
112         if (col >= cstart && col < cend) {
113           /* column is in diagonal block of matrix */
114           rows = A_cj + A_ci[col-cstart];
115           m    = A_ci[col-cstart+1] - A_ci[col-cstart];
116         } else {
117 #if defined(PETSC_USE_CTABLE)
118           ierr = PetscTableFind(aij->colmap,col+1,&colb);CHKERRQ(ierr);
119           colb--;
120 #else
121           colb = aij->colmap[col] - 1;
122 #endif
123           if (colb == -1) {
124             m = 0;
125           } else {
126             rows = B_cj + B_ci[colb];
127             m    = B_ci[colb+1] - B_ci[colb];
128           }
129         }
130         /* loop over columns marking them in rowhit */
131         for (k=0; k<m; k++) {
132           rowhit[*rows++] = col + 1;
133         }
134       }
135 
136       /* count the number of hits */
137       nrows = 0;
138       for (j=0; j<M; j++) {
139         if (rowhit[j]) nrows++;
140       }
141       c->nrows[i] = nrows;
142       ierr        = PetscMalloc((nrows+1)*sizeof(PetscInt),&c->rows[i]);CHKERRQ(ierr);
143       ierr        = PetscMalloc((nrows+1)*sizeof(PetscInt),&c->columnsforrow[i]);CHKERRQ(ierr);
144       ierr        = PetscLogObjectMemory(c,2*(nrows+1)*sizeof(PetscInt));CHKERRQ(ierr);
145       nrows       = 0;
146       for (j=0; j<M; j++) {
147         if (rowhit[j]) {
148           c->rows[i][nrows]          = j;
149           c->columnsforrow[i][nrows] = rowhit[j] - 1;
150           nrows++;
151         }
152       }
153     } else { /*-------------------------------------------------------------------------------*/
154       /* slow version, using rowhit as a linked list */
155       PetscInt currentcol,fm,mfm;
156       rowhit[M] = M;
157       nrows     = 0;
158       /* loop over columns*/
159       for (j=0; j<nctot; j++) {
160         if (ctype == IS_COLORING_GHOSTED) {
161           col = ltog[cols[j]];
162         } else {
163           col = cols[j];
164         }
165         if (col >= cstart && col < cend) {
166           /* column is in diagonal block of matrix */
167           rows = A_cj + A_ci[col-cstart];
168           m    = A_ci[col-cstart+1] - A_ci[col-cstart];
169         } else {
170 #if defined(PETSC_USE_CTABLE)
171           ierr = PetscTableFind(aij->colmap,col+1,&colb);CHKERRQ(ierr);
172           colb--;
173 #else
174           colb = aij->colmap[col] - 1;
175 #endif
176           if (colb == -1) {
177             m = 0;
178           } else {
179             rows = B_cj + B_ci[colb];
180             m    = B_ci[colb+1] - B_ci[colb];
181           }
182         }
183 
184         /* loop over columns marking them in rowhit */
185         fm = M;    /* fm points to first entry in linked list */
186         for (k=0; k<m; k++) {
187           currentcol = *rows++;
188           /* is it already in the list? */
189           do {
190             mfm = fm;
191             fm  = rowhit[fm];
192           } while (fm < currentcol);
193           /* not in list so add it */
194           if (fm != currentcol) {
195             nrows++;
196             columnsforrow[currentcol] = col;
197             /* next three lines insert new entry into linked list */
198             rowhit[mfm]        = currentcol;
199             rowhit[currentcol] = fm;
200             fm                 = currentcol;
201             /* fm points to present position in list since we know the columns are sorted */
202           } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Invalid coloring of matrix detected");
203         }
204       }
205       c->nrows[i] = nrows;
206 
207       ierr = PetscMalloc((nrows+1)*sizeof(PetscInt),&c->rows[i]);CHKERRQ(ierr);
208       ierr = PetscMalloc((nrows+1)*sizeof(PetscInt),&c->columnsforrow[i]);CHKERRQ(ierr);
209       ierr = PetscLogObjectMemory(c,(nrows+1)*sizeof(PetscInt));CHKERRQ(ierr);
210       /* now store the linked list of rows into c->rows[i] */
211       nrows = 0;
212       fm    = rowhit[M];
213       do {
214         c->rows[i][nrows]            = fm;
215         c->columnsforrow[i][nrows++] = columnsforrow[fm];
216         fm                           = rowhit[fm];
217       } while (fm < M);
218     } /* ---------------------------------------------------------------------------------------*/
219     ierr = PetscFree(cols);CHKERRQ(ierr);
220   }
221 
222   /* Optimize by adding the vscale, and scaleforrow[][] fields */
223   /*
224        vscale will contain the "diagonal" on processor scalings followed by the off processor
225   */
226   if (ctype == IS_COLORING_GLOBAL) {
227     ierr = VecCreateGhost(((PetscObject)mat)->comm,aij->A->rmap->n,PETSC_DETERMINE,aij->B->cmap->n,aij->garray,&c->vscale);CHKERRQ(ierr);
228     ierr = PetscMalloc(c->ncolors*sizeof(PetscInt*),&c->vscaleforrow);CHKERRQ(ierr);
229     for (k=0; k<c->ncolors; k++) {
230       ierr = PetscMalloc((c->nrows[k]+1)*sizeof(PetscInt),&c->vscaleforrow[k]);CHKERRQ(ierr);
231       for (l=0; l<c->nrows[k]; l++) {
232         col = c->columnsforrow[k][l];
233         if (col >= cstart && col < cend) {
234           /* column is in diagonal block of matrix */
235           colb = col - cstart;
236         } else {
237           /* column  is in "off-processor" part */
238 #if defined(PETSC_USE_CTABLE)
239           ierr = PetscTableFind(aij->colmap,col+1,&colb);CHKERRQ(ierr);
240           colb--;
241 #else
242           colb = aij->colmap[col] - 1;
243 #endif
244           colb += cend - cstart;
245         }
246         c->vscaleforrow[k][l] = colb;
247       }
248     }
249   } else if (ctype == IS_COLORING_GHOSTED) {
250     /* Get gtol mapping */
251     PetscInt N = mat->cmap->N, *gtol;
252     ierr = PetscMalloc((N+1)*sizeof(PetscInt),&gtol);CHKERRQ(ierr);
253     for (i=0; i<N; i++) gtol[i] = -1;
254     for (i=0; i<map->n; i++) gtol[ltog[i]] = i;
255 
256     c->vscale = 0; /* will be created in MatFDColoringApply() */
257     ierr      = PetscMalloc(c->ncolors*sizeof(PetscInt*),&c->vscaleforrow);CHKERRQ(ierr);
258     for (k=0; k<c->ncolors; k++) {
259       ierr = PetscMalloc((c->nrows[k]+1)*sizeof(PetscInt),&c->vscaleforrow[k]);CHKERRQ(ierr);
260       for (l=0; l<c->nrows[k]; l++) {
261         col = c->columnsforrow[k][l];      /* global column index */
262 
263         c->vscaleforrow[k][l] = gtol[col]; /* local column index */
264       }
265     }
266     ierr = PetscFree(gtol);CHKERRQ(ierr);
267   }
268   ierr = ISColoringRestoreIS(iscoloring,&isa);CHKERRQ(ierr);
269 
270   ierr = PetscFree(rowhit);CHKERRQ(ierr);
271   ierr = PetscFree(columnsforrow);CHKERRQ(ierr);
272   ierr = MatRestoreColumnIJ(aij->A,0,PETSC_FALSE,PETSC_FALSE,&ncols,&A_ci,&A_cj,&done);CHKERRQ(ierr);
273   ierr = MatRestoreColumnIJ(aij->B,0,PETSC_FALSE,PETSC_FALSE,&ncols,&B_ci,&B_cj,&done);CHKERRQ(ierr);
274   PetscFunctionReturn(0);
275 }
276 
277 
278 
279 
280 
281 
282