xref: /petsc/src/mat/impls/aij/mpi/fdmpiaij.c (revision b3e1f37b5be90d2fa371e91c89aba21b3af857c2)
1 
2 #include <../src/mat/impls/aij/mpi/mpiaij.h>
3 #include <../src/mat/impls/baij/mpi/mpibaij.h>
4 
5 #undef __FUNCT__
6 #define __FUNCT__ "MatFDColoringApply_BAIJ"
7 PetscErrorCode  MatFDColoringApply_BAIJ(Mat J,MatFDColoring coloring,Vec x1,MatStructure *flag,void *sctx)
8 {
9   PetscErrorCode (*f)(void*,Vec,Vec,void*) = (PetscErrorCode (*)(void*,Vec,Vec,void*))coloring->f;
10   PetscErrorCode ierr;
11   PetscInt       k,cstart,cend,l,row,col,nz,spidx,i,j;
12   PetscScalar    dx=0.0,*xx,*w3_array,*dy_i,*dy=coloring->dy;
13   PetscScalar    *vscale_array;
14   PetscReal      epsilon=coloring->error_rel,umin=coloring->umin,unorm;
15   Vec            w1=coloring->w1,w2=coloring->w2,w3,vscale=coloring->vscale;
16   void           *fctx=coloring->fctx;
17   PetscInt       ctype=coloring->ctype,nxloc,nrows_k;
18   PetscScalar    *valaddr;
19   MatEntry       *Jentry=coloring->matentry;
20   const PetscInt ncolors=coloring->ncolors,*ncolumns=coloring->ncolumns,*nrows=coloring->nrows;
21   PetscInt       bs=J->rmap->bs;
22 
23   PetscFunctionBegin;
24   /* create vscale for storing dx */
25   if (!vscale) {
26     if (ctype == IS_COLORING_GLOBAL && coloring->htype[0] == 'd') {
27       /* contain the "diagonal" on processor scalings followed by the off processor* - garray must be non-bloced! */
28       Mat_MPIBAIJ    *baij=(Mat_MPIBAIJ*)J->data;
29       PetscInt       *garray;
30       ierr = PetscMalloc(baij->B->cmap->n*sizeof(PetscInt),&garray);CHKERRQ(ierr);
31       for (i=0; i<baij->B->cmap->n/bs; i++) {
32         for (j=0; j<bs; j++) {
33           garray[i*bs+j] = bs*baij->garray[i]+j;
34         }
35       }
36       ierr = VecCreateGhost(PetscObjectComm((PetscObject)J),J->cmap->n,PETSC_DETERMINE,baij->B->cmap->n,garray,&vscale);CHKERRQ(ierr);
37       ierr = PetscFree(garray);CHKERRQ(ierr);
38     } else if (ctype == IS_COLORING_GHOSTED) {
39       ierr = VecDuplicate(x1,&vscale);CHKERRQ(ierr);
40     }
41     coloring->vscale = vscale;
42   }
43 
44   /* (1) Set w1 = F(x1) */
45   if (!coloring->fset) {
46     ierr = PetscLogEventBegin(MAT_FDColoringFunction,0,0,0,0);CHKERRQ(ierr);
47     ierr = (*f)(sctx,x1,w1,fctx);CHKERRQ(ierr);
48     ierr = PetscLogEventEnd(MAT_FDColoringFunction,0,0,0,0);CHKERRQ(ierr);
49   } else {
50     coloring->fset = PETSC_FALSE;
51   }
52 
53   /* (2) Compute vscale = 1./dx - the local scale factors, including ghost points */
54   ierr = VecGetLocalSize(x1,&nxloc);CHKERRQ(ierr);
55   if (coloring->htype[0] == 'w') {
56     /* vscale = dx is a constant scalar */
57     ierr = VecNorm(x1,NORM_2,&unorm);CHKERRQ(ierr);
58     dx = 1.0/(PetscSqrtReal(1.0 + unorm)*epsilon);
59   } else {
60     ierr = VecGetArray(x1,&xx);CHKERRQ(ierr);
61     ierr = VecGetArray(vscale,&vscale_array);CHKERRQ(ierr);
62     for (col=0; col<nxloc; col++) {
63       dx = xx[col];
64       if (PetscAbsScalar(dx) < umin) {
65         if (PetscRealPart(dx) >= 0.0)      dx = umin;
66         else if (PetscRealPart(dx) < 0.0 ) dx = -umin;
67       }
68       dx               *= epsilon;
69       vscale_array[col] = 1.0/dx;
70     }
71     ierr = VecRestoreArray(x1,&xx);CHKERRQ(ierr);
72     ierr = VecRestoreArray(vscale,&vscale_array);CHKERRQ(ierr);
73   }
74   if (ctype == IS_COLORING_GLOBAL && coloring->htype[0] == 'd') {
75     ierr = VecGhostUpdateBegin(vscale,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
76     ierr = VecGhostUpdateEnd(vscale,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
77   }
78 
79   /* (3) Loop over each color */
80   if (!coloring->w3) {
81     ierr = VecDuplicate(x1,&coloring->w3);CHKERRQ(ierr);
82     ierr = PetscLogObjectParent((PetscObject)coloring,(PetscObject)coloring->w3);CHKERRQ(ierr);
83   }
84   w3 = coloring->w3;
85 
86   ierr = VecGetOwnershipRange(x1,&cstart,&cend);CHKERRQ(ierr); /* used by ghosted vscale */
87   if (vscale) {
88     ierr = VecGetArray(vscale,&vscale_array);CHKERRQ(ierr);
89   }
90   nz   = 0;
91   for (k=0; k<ncolors; k++) {
92     coloring->currentcolor = k;
93 
94     /*
95       (3-1) Loop over each column associated with color
96       adding the perturbation to the vector w3 = x1 + dx.
97     */
98     ierr = VecCopy(x1,w3);CHKERRQ(ierr);
99     dy_i = dy;
100     for (i=0; i<bs; i++) {     /* Loop over a block of columns */
101       ierr = VecGetArray(w3,&w3_array);CHKERRQ(ierr);
102       if (ctype == IS_COLORING_GLOBAL) w3_array -= cstart; /* shift pointer so global index can be used */
103       if (coloring->htype[0] == 'w') {
104         for (l=0; l<ncolumns[k]; l++) {
105           col            = i + bs*coloring->columns[k][l];  /* local column (in global index!) of the matrix we are probing for */
106           w3_array[col] += 1.0/dx;
107           if (i) w3_array[col-1] -= 1.0/dx; /* resume original w3[col-1] */
108         }
109       } else { /* htype == 'ds' */
110         vscale_array -= cstart; /* shift pointer so global index can be used */
111         for (l=0; l<ncolumns[k]; l++) {
112           col = i + bs*coloring->columns[k][l]; /* local column (in global index!) of the matrix we are probing for */
113           w3_array[col] += 1.0/vscale_array[col];
114           if (i) w3_array[col-1] -=  1.0/vscale_array[col-1]; /* resume original w3[col-1] */
115         }
116         vscale_array += cstart;
117       }
118       if (ctype == IS_COLORING_GLOBAL) w3_array += cstart;
119       ierr = VecRestoreArray(w3,&w3_array);CHKERRQ(ierr);
120 
121       /*
122        (3-2) Evaluate function at w3 = x1 + dx (here dx is a vector of perturbations)
123                            w2 = F(x1 + dx) - F(x1)
124        */
125       ierr = PetscLogEventBegin(MAT_FDColoringFunction,0,0,0,0);CHKERRQ(ierr);
126       ierr = VecPlaceArray(w2,dy_i);CHKERRQ(ierr); /* place w2 to the array dy_i */
127       ierr = (*f)(sctx,w3,w2,fctx);CHKERRQ(ierr);
128       ierr = PetscLogEventEnd(MAT_FDColoringFunction,0,0,0,0);CHKERRQ(ierr);
129       ierr = VecAXPY(w2,-1.0,w1);CHKERRQ(ierr);
130       ierr = VecResetArray(w2);CHKERRQ(ierr);
131       dy_i += nxloc; /* points to dy+i*nxloc */
132     }
133 
134     /*
135      (3-3) Loop over rows of vector, putting results into Jacobian matrix
136     */
137     nrows_k = nrows[k];
138     if (coloring->htype[0] == 'w') {
139       for (l=0; l<nrows_k; l++) {
140         row     = bs*Jentry[nz].row;   /* local row index */
141         valaddr = Jentry[nz++].valaddr;
142         spidx   = 0;
143         dy_i    = dy;
144         for (i=0; i<bs; i++) {   /* column of the block */
145           for (j=0; j<bs; j++) { /* row of the block */
146             valaddr[spidx++] = dy_i[row+j]*dx;
147           }
148           dy_i += nxloc; /* points to dy+i*nxloc */
149         }
150       }
151     } else { /* htype == 'ds' */
152       for (l=0; l<nrows_k; l++) {
153         row     = bs*Jentry[nz].row;   /* local row index */
154         col     = bs*Jentry[nz].col;   /* local column index */
155         valaddr = Jentry[nz++].valaddr;
156         spidx   = 0;
157         dy_i    = dy;
158         for (i=0; i<bs; i++) {   /* column of the block */
159           for (j=0; j<bs; j++) { /* row of the block */
160             valaddr[spidx++] = dy_i[row+j]*vscale_array[col+i];
161           }
162           dy_i += nxloc; /* points to dy+i*nxloc */
163         }
164       }
165     }
166   }
167   ierr = MatAssemblyBegin(J,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
168   ierr = MatAssemblyEnd(J,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
169   if (vscale) {
170     ierr = VecRestoreArray(vscale,&vscale_array);CHKERRQ(ierr);
171   }
172 
173   coloring->currentcolor = -1;
174   PetscFunctionReturn(0);
175 }
176 
177 #undef __FUNCT__
178 #define __FUNCT__ "MatFDColoringApply_AIJ"
179 PetscErrorCode  MatFDColoringApply_AIJ(Mat J,MatFDColoring coloring,Vec x1,MatStructure *flag,void *sctx)
180 {
181   PetscErrorCode (*f)(void*,Vec,Vec,void*) = (PetscErrorCode (*)(void*,Vec,Vec,void*))coloring->f;
182   PetscErrorCode ierr;
183   PetscInt       k,cstart,cend,l,row,col,nz;
184   PetscScalar    dx=0.0,*y,*xx,*w3_array;
185   PetscScalar    *vscale_array;
186   PetscReal      epsilon=coloring->error_rel,umin=coloring->umin,unorm;
187   Vec            w1=coloring->w1,w2=coloring->w2,w3,vscale=coloring->vscale;
188   void           *fctx=coloring->fctx;
189   PetscInt       ctype=coloring->ctype,nxloc,nrows_k;
190   MatEntry       *Jentry=coloring->matentry;
191   const PetscInt ncolors=coloring->ncolors,*ncolumns=coloring->ncolumns,*nrows=coloring->nrows;
192 
193   PetscFunctionBegin;
194   /* create vscale for storing dx */
195   if (!vscale) {
196     if (ctype == IS_COLORING_GLOBAL && coloring->htype[0] == 'd') {
197       Mat_MPIAIJ     *aij=(Mat_MPIAIJ*)J->data;
198       ierr = VecCreateGhost(PetscObjectComm((PetscObject)J),J->cmap->n,PETSC_DETERMINE,aij->B->cmap->n,aij->garray,&vscale);CHKERRQ(ierr);
199     } else if (ctype == IS_COLORING_GHOSTED) {
200       ierr = VecDuplicate(x1,&vscale);CHKERRQ(ierr);
201     }
202     coloring->vscale = vscale;
203   }
204 
205   /* (1) Set w1 = F(x1) */
206   if (!coloring->fset) {
207     ierr = PetscLogEventBegin(MAT_FDColoringFunction,0,0,0,0);CHKERRQ(ierr);
208     ierr = (*f)(sctx,x1,w1,fctx);CHKERRQ(ierr);
209     ierr = PetscLogEventEnd(MAT_FDColoringFunction,0,0,0,0);CHKERRQ(ierr);
210   } else {
211     coloring->fset = PETSC_FALSE;
212   }
213 
214   /* (2) Compute vscale = 1./dx - the local scale factors, including ghost points */
215   if (coloring->htype[0] == 'w') {
216     /* vscale = 1./dx is a constant scalar */
217     ierr = VecNorm(x1,NORM_2,&unorm);CHKERRQ(ierr);
218     dx = 1.0/(PetscSqrtReal(1.0 + unorm)*epsilon);
219   } else {
220     ierr = VecGetLocalSize(x1,&nxloc);CHKERRQ(ierr);
221     ierr = VecGetArray(x1,&xx);CHKERRQ(ierr);
222     ierr = VecGetArray(vscale,&vscale_array);CHKERRQ(ierr);
223     for (col=0; col<nxloc; col++) {
224       dx = xx[col];
225       if (PetscAbsScalar(dx) < umin) {
226         if (PetscRealPart(dx) >= 0.0)      dx = umin;
227         else if (PetscRealPart(dx) < 0.0 ) dx = -umin;
228       }
229       dx               *= epsilon;
230       vscale_array[col] = 1.0/dx;
231     }
232     ierr = VecRestoreArray(x1,&xx);CHKERRQ(ierr);
233     ierr = VecRestoreArray(vscale,&vscale_array);CHKERRQ(ierr);
234   }
235   if (ctype == IS_COLORING_GLOBAL && coloring->htype[0] == 'd') {
236     ierr = VecGhostUpdateBegin(vscale,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
237     ierr = VecGhostUpdateEnd(vscale,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
238   }
239 
240   /* (3) Loop over each color */
241   if (!coloring->w3) {
242     ierr = VecDuplicate(x1,&coloring->w3);CHKERRQ(ierr);
243     ierr = PetscLogObjectParent((PetscObject)coloring,(PetscObject)coloring->w3);CHKERRQ(ierr);
244   }
245   w3 = coloring->w3;
246 
247   ierr = VecGetOwnershipRange(x1,&cstart,&cend);CHKERRQ(ierr); /* used by ghosted vscale */
248   if (vscale) {
249     ierr = VecGetArray(vscale,&vscale_array);CHKERRQ(ierr);
250   }
251   nz   = 0;
252 
253   if (coloring->bcols > 1) { /* use blocked insertion of Jentry */
254     PetscInt    i,m=J->rmap->n,nbcols,bcols=coloring->bcols;
255     PetscScalar *dy=coloring->dy,*dy_k;
256 
257     nbcols = 0;
258     for (k=0; k<ncolors; k+=bcols) {
259       coloring->currentcolor = k;
260 
261       /*
262        (3-1) Loop over each column associated with color
263        adding the perturbation to the vector w3 = x1 + dx.
264        */
265 
266       dy_k = dy;
267       if (k + bcols > ncolors) bcols = ncolors - k;
268       for (i=0; i<bcols; i++) {
269 
270         ierr = VecCopy(x1,w3);CHKERRQ(ierr);
271         ierr = VecGetArray(w3,&w3_array);CHKERRQ(ierr);
272         if (ctype == IS_COLORING_GLOBAL) w3_array -= cstart; /* shift pointer so global index can be used */
273         if (coloring->htype[0] == 'w') {
274           for (l=0; l<ncolumns[k+i]; l++) {
275             col = coloring->columns[k+i][l]; /* local column (in global index!) of the matrix we are probing for */
276             w3_array[col] += 1.0/dx;
277           }
278         } else { /* htype == 'ds' */
279           vscale_array -= cstart; /* shift pointer so global index can be used */
280           for (l=0; l<ncolumns[k+i]; l++) {
281             col = coloring->columns[k+i][l]; /* local column (in global index!) of the matrix we are probing for */
282             w3_array[col] += 1.0/vscale_array[col];
283           }
284           vscale_array += cstart;
285         }
286         if (ctype == IS_COLORING_GLOBAL) w3_array += cstart;
287         ierr = VecRestoreArray(w3,&w3_array);CHKERRQ(ierr);
288 
289         /*
290          (3-2) Evaluate function at w3 = x1 + dx (here dx is a vector of perturbations)
291                            w2 = F(x1 + dx) - F(x1)
292          */
293         ierr = PetscLogEventBegin(MAT_FDColoringFunction,0,0,0,0);CHKERRQ(ierr);
294         ierr = VecPlaceArray(w2,dy_k);CHKERRQ(ierr); /* place w2 to the array dy_i */
295         ierr = (*f)(sctx,w3,w2,fctx);CHKERRQ(ierr);
296         ierr = PetscLogEventEnd(MAT_FDColoringFunction,0,0,0,0);CHKERRQ(ierr);
297         ierr = VecAXPY(w2,-1.0,w1);CHKERRQ(ierr);
298         ierr = VecResetArray(w2);CHKERRQ(ierr);
299         dy_k += m; /* points to dy+i*nxloc */
300 
301       }
302 
303       /*
304        (3-3) Loop over block rows of vector, putting results into Jacobian matrix
305        */
306       nrows_k = nrows[nbcols++];
307       ierr = VecGetArray(w2,&y);CHKERRQ(ierr);
308 
309       if (coloring->htype[0] == 'w') {
310         for (l=0; l<nrows_k; l++) {
311           row                     = Jentry[nz].row;   /* local row index */
312           *(Jentry[nz++].valaddr) = dy[row]*dx;
313         }
314       } else { /* htype == 'ds' */
315         for (l=0; l<nrows_k; l++) {
316           row                   = Jentry[nz].row;   /* local row index */
317           *(Jentry[nz].valaddr) = dy[row]*vscale_array[Jentry[nz].col];
318           nz++;
319         }
320       }
321       ierr = VecRestoreArray(w2,&y);CHKERRQ(ierr);
322     }
323   } else { /* bcols == 1 */
324     for (k=0; k<ncolors; k++) {
325       coloring->currentcolor = k;
326 
327       /*
328        (3-1) Loop over each column associated with color
329        adding the perturbation to the vector w3 = x1 + dx.
330        */
331       ierr = VecCopy(x1,w3);CHKERRQ(ierr);
332       ierr = VecGetArray(w3,&w3_array);CHKERRQ(ierr);
333       if (ctype == IS_COLORING_GLOBAL) w3_array -= cstart; /* shift pointer so global index can be used */
334       if (coloring->htype[0] == 'w') {
335         for (l=0; l<ncolumns[k]; l++) {
336           col = coloring->columns[k][l]; /* local column (in global index!) of the matrix we are probing for */
337           w3_array[col] += 1.0/dx;
338         }
339       } else { /* htype == 'ds' */
340         vscale_array -= cstart; /* shift pointer so global index can be used */
341         for (l=0; l<ncolumns[k]; l++) {
342           col = coloring->columns[k][l]; /* local column (in global index!) of the matrix we are probing for */
343           w3_array[col] += 1.0/vscale_array[col];
344         }
345         vscale_array += cstart;
346       }
347       if (ctype == IS_COLORING_GLOBAL) w3_array += cstart;
348       ierr = VecRestoreArray(w3,&w3_array);CHKERRQ(ierr);
349 
350       /*
351        (3-2) Evaluate function at w3 = x1 + dx (here dx is a vector of perturbations)
352                            w2 = F(x1 + dx) - F(x1)
353        */
354       ierr = PetscLogEventBegin(MAT_FDColoringFunction,0,0,0,0);CHKERRQ(ierr);
355       ierr = (*f)(sctx,w3,w2,fctx);CHKERRQ(ierr);
356       ierr = PetscLogEventEnd(MAT_FDColoringFunction,0,0,0,0);CHKERRQ(ierr);
357       ierr = VecAXPY(w2,-1.0,w1);CHKERRQ(ierr);
358 
359       /*
360        (3-3) Loop over rows of vector, putting results into Jacobian matrix
361        */
362       nrows_k = nrows[k];
363       ierr = VecGetArray(w2,&y);CHKERRQ(ierr);
364       if (coloring->htype[0] == 'w') {
365         for (l=0; l<nrows_k; l++) {
366           row                     = Jentry[nz].row;   /* local row index */
367           *(Jentry[nz++].valaddr) = y[row]*dx;
368         }
369       } else { /* htype == 'ds' */
370         for (l=0; l<nrows_k; l++) {
371           row                   = Jentry[nz].row;   /* local row index */
372           *(Jentry[nz].valaddr) = y[row]*vscale_array[Jentry[nz].col];
373           nz++;
374         }
375       }
376       ierr = VecRestoreArray(w2,&y);CHKERRQ(ierr);
377     }
378   }
379 
380   ierr = MatAssemblyBegin(J,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
381   ierr = MatAssemblyEnd(J,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
382   if (vscale) {
383     ierr = VecRestoreArray(vscale,&vscale_array);CHKERRQ(ierr);
384   }
385   coloring->currentcolor = -1;
386   PetscFunctionReturn(0);
387 }
388 
389 #undef __FUNCT__
390 #define __FUNCT__ "MatFDColoringSetUp_MPIXAIJ"
391 PetscErrorCode MatFDColoringSetUp_MPIXAIJ(Mat mat,ISColoring iscoloring,MatFDColoring c)
392 {
393   PetscErrorCode         ierr;
394   PetscMPIInt            size,*ncolsonproc,*disp,nn;
395   PetscInt               i,n,nrows,nrows_i,j,k,m,ncols,col,*rowhit,cstart,cend,colb;
396   const PetscInt         *is,*A_ci,*A_cj,*B_ci,*B_cj,*row=NULL,*ltog=NULL;
397   PetscInt               nis=iscoloring->n,nctot,*cols;
398   IS                     *isa;
399   ISLocalToGlobalMapping map=mat->cmap->mapping;
400   PetscInt               ctype=c->ctype,*spidxA,*spidxB,nz,bs,bs2,spidx;
401   Mat                    A,B;
402   PetscScalar            *A_val,*B_val,**valaddrhit;
403   MatEntry               *Jentry;
404   PetscBool              isBAIJ;
405   PetscInt               bcols=c->bcols;
406 #if defined(PETSC_USE_CTABLE)
407   PetscTable             colmap=NULL;
408 #else
409   PetscInt               *colmap=NULL;     /* local col number of off-diag col */
410 #endif
411 
412   PetscFunctionBegin;
413   if (ctype == IS_COLORING_GHOSTED) {
414     if (!map) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_ARG_INCOMP,"When using ghosted differencing matrix must have local to global mapping provided with MatSetLocalToGlobalMapping");
415     ierr = ISLocalToGlobalMappingGetIndices(map,&ltog);CHKERRQ(ierr);
416   }
417 
418   ierr = MatGetBlockSize(mat,&bs);CHKERRQ(ierr);
419   ierr = PetscObjectTypeCompare((PetscObject)mat,MATMPIBAIJ,&isBAIJ);CHKERRQ(ierr);
420   if (isBAIJ) {
421     Mat_MPIBAIJ *aij=(Mat_MPIBAIJ*)mat->data;
422     Mat_SeqBAIJ *spA,*spB;
423     A = aij->A;  spA = (Mat_SeqBAIJ*)A->data; A_val = spA->a;
424     B = aij->B;  spB = (Mat_SeqBAIJ*)B->data; B_val = spB->a;
425     nz = spA->nz + spB->nz; /* total nonzero entries of mat */
426     if (!aij->colmap) {
427       ierr = MatCreateColmap_MPIBAIJ_Private(mat);CHKERRQ(ierr);
428       colmap = aij->colmap;
429     }
430     ierr = MatGetColumnIJ_SeqBAIJ_Color(A,0,PETSC_FALSE,PETSC_FALSE,&ncols,&A_ci,&A_cj,&spidxA,NULL);CHKERRQ(ierr);
431     ierr = MatGetColumnIJ_SeqBAIJ_Color(B,0,PETSC_FALSE,PETSC_FALSE,&ncols,&B_ci,&B_cj,&spidxB,NULL);CHKERRQ(ierr);
432   } else {
433     Mat_MPIAIJ *aij=(Mat_MPIAIJ*)mat->data;
434     Mat_SeqAIJ *spA,*spB;
435     A = aij->A;  spA = (Mat_SeqAIJ*)A->data; A_val = spA->a;
436     B = aij->B;  spB = (Mat_SeqAIJ*)B->data; B_val = spB->a;
437     nz = spA->nz + spB->nz; /* total nonzero entries of mat */
438     if (!aij->colmap) {
439       /* Allow access to data structures of local part of matrix
440        - creates aij->colmap which maps global column number to local number in part B */
441       ierr = MatCreateColmap_MPIAIJ_Private(mat);CHKERRQ(ierr);
442       colmap = aij->colmap;
443     }
444     ierr = MatGetColumnIJ_SeqAIJ_Color(A,0,PETSC_FALSE,PETSC_FALSE,&ncols,&A_ci,&A_cj,&spidxA,NULL);CHKERRQ(ierr);
445     ierr = MatGetColumnIJ_SeqAIJ_Color(B,0,PETSC_FALSE,PETSC_FALSE,&ncols,&B_ci,&B_cj,&spidxB,NULL);CHKERRQ(ierr);
446 
447     bs = 1; /* only bs=1 is supported for non MPIBAIJ matrix */
448   }
449 
450   m         = mat->rmap->n/bs;
451   cstart    = mat->cmap->rstart/bs;
452   cend      = mat->cmap->rend/bs;
453 
454   ierr       = PetscMalloc(nis*sizeof(PetscInt),&c->ncolumns);CHKERRQ(ierr);
455   ierr       = PetscMalloc(nis*sizeof(PetscInt*),&c->columns);CHKERRQ(ierr);
456   ierr       = PetscMalloc(nis*sizeof(PetscInt),&c->nrows);CHKERRQ(ierr);
457   ierr       = PetscLogObjectMemory((PetscObject)c,3*nis*sizeof(PetscInt));CHKERRQ(ierr);
458 
459   ierr       = PetscMalloc(nz*sizeof(MatEntry),&Jentry);CHKERRQ(ierr);
460   ierr       = PetscLogObjectMemory((PetscObject)c,nz*sizeof(MatEntry));CHKERRQ(ierr);
461   c->matentry = Jentry;
462 
463   ierr = PetscMalloc2(m+1,PetscInt,&rowhit,m+1,PetscScalar*,&valaddrhit);CHKERRQ(ierr);
464   nz = 0;
465   ierr = ISColoringGetIS(iscoloring,PETSC_IGNORE,&isa);CHKERRQ(ierr);
466   for (i=0; i<nis; i++) { /* for each local color */
467     ierr = ISGetLocalSize(isa[i],&n);CHKERRQ(ierr);
468     ierr = ISGetIndices(isa[i],&is);CHKERRQ(ierr);
469 
470     c->ncolumns[i] = n; /* local number of columns of this color on this process */
471     if (n) {
472       ierr = PetscMalloc(n*sizeof(PetscInt),&c->columns[i]);CHKERRQ(ierr);
473       ierr = PetscLogObjectMemory((PetscObject)c,n*sizeof(PetscInt));CHKERRQ(ierr);
474       ierr = PetscMemcpy(c->columns[i],is,n*sizeof(PetscInt));CHKERRQ(ierr);
475     } else {
476       c->columns[i] = 0;
477     }
478 
479     if (ctype == IS_COLORING_GLOBAL) {
480       /* Determine nctot, the total (parallel) number of columns of this color */
481       ierr = MPI_Comm_size(PetscObjectComm((PetscObject)mat),&size);CHKERRQ(ierr);
482       ierr = PetscMalloc2(size,PetscMPIInt,&ncolsonproc,size,PetscMPIInt,&disp);CHKERRQ(ierr);
483 
484       /* ncolsonproc[j]: local ncolumns on proc[j] of this color */
485       ierr  = PetscMPIIntCast(n,&nn);CHKERRQ(ierr);
486       ierr  = MPI_Allgather(&nn,1,MPI_INT,ncolsonproc,1,MPI_INT,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr);
487       nctot = 0; for (j=0; j<size; j++) nctot += ncolsonproc[j];
488       if (!nctot) {
489         ierr = PetscInfo(mat,"Coloring of matrix has some unneeded colors with no corresponding rows\n");CHKERRQ(ierr);
490       }
491 
492       disp[0] = 0;
493       for (j=1; j<size; j++) {
494         disp[j] = disp[j-1] + ncolsonproc[j-1];
495       }
496 
497       /* Get cols, the complete list of columns for this color on each process */
498       ierr = PetscMalloc((nctot+1)*sizeof(PetscInt),&cols);CHKERRQ(ierr);
499       ierr = MPI_Allgatherv((void*)is,n,MPIU_INT,cols,ncolsonproc,disp,MPIU_INT,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr);
500       ierr = PetscFree2(ncolsonproc,disp);CHKERRQ(ierr);
501     } else if (ctype == IS_COLORING_GHOSTED) {
502       /* Determine local number of columns of this color on this process, including ghost points */
503       nctot = n;
504       ierr  = PetscMalloc((nctot+1)*sizeof(PetscInt),&cols);CHKERRQ(ierr);
505       ierr  = PetscMemcpy(cols,is,n*sizeof(PetscInt));CHKERRQ(ierr);
506     } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Not provided for this MatFDColoring type");
507 
508     /* Mark all rows affect by these columns */
509     ierr    = PetscMemzero(rowhit,m*sizeof(PetscInt));CHKERRQ(ierr);
510     bs2     = bs*bs;
511     nrows_i = 0;
512     for (j=0; j<nctot; j++) { /* loop over columns*/
513       if (ctype == IS_COLORING_GHOSTED) {
514         col = ltog[cols[j]];
515       } else {
516         col = cols[j];
517       }
518       if (col >= cstart && col < cend) { /* column is in A, diagonal block of mat */
519         row      = A_cj + A_ci[col-cstart];
520         nrows    = A_ci[col-cstart+1] - A_ci[col-cstart];
521         nrows_i += nrows;
522         /* loop over columns of A marking them in rowhit */
523         for (k=0; k<nrows; k++) {
524           /* set valaddrhit for part A */
525           spidx            = bs2*spidxA[A_ci[col-cstart] + k];
526           valaddrhit[*row] = &A_val[spidx];
527           rowhit[*row++]   = col - cstart + 1; /* local column index */
528         }
529       } else { /* column is in B, off-diagonal block of mat */
530 #if defined(PETSC_USE_CTABLE)
531         ierr = PetscTableFind(colmap,col+1,&colb);CHKERRQ(ierr);
532         colb--;
533 #else
534         colb = colmap[col] - 1; /* local column index */
535 #endif
536         if (colb == -1) {
537           nrows = 0;
538         } else {
539           colb  = colb/bs;
540           row   = B_cj + B_ci[colb];
541           nrows = B_ci[colb+1] - B_ci[colb];
542         }
543         nrows_i += nrows;
544         /* loop over columns of B marking them in rowhit */
545         for (k=0; k<nrows; k++) {
546           /* set valaddrhit for part B */
547           spidx            = bs2*spidxB[B_ci[colb] + k];
548           valaddrhit[*row] = &B_val[spidx];
549           rowhit[*row++]   = colb + 1 + cend - cstart; /* local column index */
550         }
551       }
552     }
553     c->nrows[i] = nrows_i;
554 
555     for (j=0; j<m; j++) {
556       if (rowhit[j]) {
557         Jentry[nz].row     = j;              /* local row index */
558         Jentry[nz].col     = rowhit[j] - 1;  /* local column index */
559         Jentry[nz].valaddr = valaddrhit[j];  /* address of mat value for this entry */
560         nz++;
561       }
562     }
563     ierr = PetscFree(cols);CHKERRQ(ierr);
564   }
565 
566   if (bcols > 1) { /* reorder Jentry for faster MatFDColoringApply() */
567     ierr = MatFDColoringSetUpBlocked_AIJ_Private(mat,c,nz);CHKERRQ(ierr);
568   }
569 
570   if (isBAIJ) {
571     ierr = MatRestoreColumnIJ_SeqBAIJ_Color(A,0,PETSC_FALSE,PETSC_FALSE,&ncols,&A_ci,&A_cj,&spidxA,NULL);CHKERRQ(ierr);
572     ierr = MatRestoreColumnIJ_SeqBAIJ_Color(B,0,PETSC_FALSE,PETSC_FALSE,&ncols,&B_ci,&B_cj,&spidxB,NULL);CHKERRQ(ierr);
573     ierr = PetscMalloc(bs*mat->rmap->n*sizeof(PetscScalar),&c->dy);CHKERRQ(ierr);
574   } else {
575     ierr = MatRestoreColumnIJ_SeqAIJ_Color(A,0,PETSC_FALSE,PETSC_FALSE,&ncols,&A_ci,&A_cj,&spidxA,NULL);CHKERRQ(ierr);
576     ierr = MatRestoreColumnIJ_SeqAIJ_Color(B,0,PETSC_FALSE,PETSC_FALSE,&ncols,&B_ci,&B_cj,&spidxB,NULL);CHKERRQ(ierr);
577   }
578 
579   ierr = ISColoringRestoreIS(iscoloring,&isa);CHKERRQ(ierr);
580   ierr = PetscFree2(rowhit,valaddrhit);CHKERRQ(ierr);
581 
582   if (ctype == IS_COLORING_GHOSTED) {
583     ierr = ISLocalToGlobalMappingRestoreIndices(map,&ltog);CHKERRQ(ierr);
584   }
585 #if defined(PETSC_USE_INFO)
586   ierr = PetscInfo3(c,"ncolors %D, brows %D and bcols %D are used.\n",c->ncolors,c->brows,c->bcols);CHKERRQ(ierr);
587 #endif
588   PetscFunctionReturn(0);
589 }
590 
591 #undef __FUNCT__
592 #define __FUNCT__ "MatFDColoringCreate_MPIXAIJ"
593 PetscErrorCode MatFDColoringCreate_MPIXAIJ(Mat mat,ISColoring iscoloring,MatFDColoring c)
594 {
595   PetscErrorCode ierr;
596   PetscInt       bs,nis=iscoloring->n,m=mat->rmap->n;
597   PetscBool      isBAIJ;
598 
599   PetscFunctionBegin;
600   /* set default brows and bcols for speedup inserting the dense matrix into sparse Jacobian;
601    bcols is chosen s.t. dy-array takes 50% of memory space as mat */
602   ierr = MatGetBlockSize(mat,&bs);CHKERRQ(ierr);
603   ierr = PetscObjectTypeCompare((PetscObject)mat,MATMPIBAIJ,&isBAIJ);CHKERRQ(ierr);
604   if (isBAIJ) {
605     c->brows = m;
606     c->bcols = 1;
607   } else { /* mpiaij matrix */
608     /* bcols is chosen s.t. dy-array takes 50% of local memory space as mat */
609     Mat_MPIAIJ *aij=(Mat_MPIAIJ*)mat->data;
610     Mat_SeqAIJ *spA,*spB;
611     Mat        A,B;
612     PetscInt   nz,brows,bcols;
613     PetscReal  mem;
614 
615     bs    = 1; /* only bs=1 is supported for MPIAIJ matrix */
616 
617     A = aij->A;  spA = (Mat_SeqAIJ*)A->data;
618     B = aij->B;  spB = (Mat_SeqAIJ*)B->data;
619     nz = spA->nz + spB->nz; /* total local nonzero entries of mat */
620     mem = nz*(sizeof(PetscScalar) + sizeof(PetscInt)) + 3*m*sizeof(PetscInt);
621     bcols = (PetscInt)(0.5*mem /(m*sizeof(PetscScalar)));
622     brows = 1000/bcols;
623     if (bcols > nis) bcols = nis;
624     if (brows == 0 || brows > m) brows = m;
625     c->brows = brows;
626     c->bcols = bcols;
627   }
628 
629   c->M       = mat->rmap->N/bs;         /* set the global rows and columns and local rows */
630   c->N       = mat->cmap->N/bs;
631   c->m       = mat->rmap->n/bs;
632   c->rstart  = mat->rmap->rstart/bs;
633   c->ncolors = nis;
634   PetscFunctionReturn(0);
635 }
636