1 2 #ifdef PETSC_RCS_HEADER 3 static char vcid[] = "$Id: fdmpiaij.c,v 1.15 1997/11/28 16:19:52 bsmith Exp bsmith $"; 4 #endif 5 6 #include "src/mat/impls/aij/mpi/mpiaij.h" 7 #include "src/vec/vecimpl.h" 8 #include "petsc.h" 9 10 extern int CreateColmap_MPIAIJ_Private(Mat); 11 extern int MatGetColumnIJ_SeqAIJ(Mat,int,PetscTruth,int*,int**,int**,PetscTruth*); 12 extern int MatRestoreColumnIJ_SeqAIJ(Mat,int,PetscTruth,int*,int**,int**,PetscTruth*); 13 14 #undef __FUNC__ 15 #define __FUNC__ "MatFDColoringCreate_MPIAIJ" 16 int MatFDColoringCreate_MPIAIJ(Mat mat,ISColoring iscoloring,MatFDColoring c) 17 { 18 Mat_MPIAIJ *aij = (Mat_MPIAIJ *) mat->data; 19 int i,*is,n,nrows,j,k,m,*rows = 0,ierr,*A_ci,*A_cj,ncols,col,flg; 20 int nis = iscoloring->n,*ncolsonproc,size,nctot,*cols,*disp,*B_ci,*B_cj; 21 int *rowhit, M = mat->m,cstart = aij->cstart, cend = aij->cend,colb; 22 int *columnsforrow; 23 IS *isa = iscoloring->is; 24 PetscTruth done; 25 26 PetscFunctionBegin; 27 if (!mat->assembled) { 28 SETERRQ(PETSC_ERR_ARG_WRONGSTATE,1,"Matrix must be assembled first; MatAssemblyBegin/End();"); 29 } 30 31 c->M = mat->M; /* set the global rows and columns and local rows */ 32 c->N = mat->N; 33 c->m = mat->m; 34 c->rstart = aij->rstart; 35 36 c->ncolors = nis; 37 c->ncolumns = (int *) PetscMalloc( nis*sizeof(int) ); CHKPTRQ(c->ncolumns); 38 c->columns = (int **) PetscMalloc( nis*sizeof(int *)); CHKPTRQ(c->columns); 39 c->nrows = (int *) PetscMalloc( nis*sizeof(int) ); CHKPTRQ(c->nrows); 40 c->rows = (int **) PetscMalloc( nis*sizeof(int *)); CHKPTRQ(c->rows); 41 c->columnsforrow = (int **) PetscMalloc( nis*sizeof(int *)); CHKPTRQ(c->columnsforrow); 42 PLogObjectMemory(c,5*nis*sizeof(int)); 43 44 /* Allow access to data structures of local part of matrix */ 45 if (!aij->colmap) { 46 ierr = CreateColmap_MPIAIJ_Private(mat);CHKERRQ(ierr); 47 } 48 /* 49 Calls the _SeqAIJ() version of these routines to make sure it does not 50 get the reduced (by inodes) version of I and J 51 */ 52 ierr = MatGetColumnIJ_SeqAIJ(aij->A,0,PETSC_FALSE,&ncols,&A_ci,&A_cj,&done); CHKERRQ(ierr); 53 ierr = MatGetColumnIJ_SeqAIJ(aij->B,0,PETSC_FALSE,&ncols,&B_ci,&B_cj,&done); CHKERRQ(ierr); 54 55 MPI_Comm_size(mat->comm,&size); 56 ncolsonproc = (int *) PetscMalloc( 2*size*sizeof(int *) ); CHKPTRQ(ncolsonproc); 57 disp = ncolsonproc + size; 58 59 rowhit = (int *) PetscMalloc( (M+1)*sizeof(int) ); CHKPTRQ(rowhit); 60 columnsforrow = (int *) PetscMalloc( (M+1)*sizeof(int) );CHKPTRQ(columnsforrow); 61 62 /* 63 Temporary option to allow for debugging/testing 64 */ 65 ierr = OptionsHasName(0,"-matfdcoloring_slow",&flg); 66 67 for ( i=0; i<nis; i++ ) { 68 ierr = ISGetSize(isa[i],&n); CHKERRQ(ierr); 69 ierr = ISGetIndices(isa[i],&is); CHKERRQ(ierr); 70 c->ncolumns[i] = n; 71 c->ncolumns[i] = n; 72 if (n) { 73 c->columns[i] = (int *) PetscMalloc( n*sizeof(int) ); CHKPTRQ(c->columns[i]); 74 PLogObjectMemory(c,n*sizeof(int)); 75 PetscMemcpy(c->columns[i],is,n*sizeof(int)); 76 } else { 77 c->columns[i] = 0; 78 } 79 80 /* Determine the total (parallel) number of columns of this color */ 81 ierr = MPI_Allgather(&n,1,MPI_INT,ncolsonproc,1,MPI_INT,mat->comm);CHKERRQ(ierr); 82 nctot = 0; for ( j=0; j<size; j++ ) {nctot += ncolsonproc[j];} 83 if (!nctot) SETERRQ(PETSC_ERR_PLIB,0,"Invalid coloring of matrix detected"); 84 85 disp[0] = 0; 86 for ( j=1; j<size; j++ ) { 87 disp[j] = disp[j-1] + ncolsonproc[j-1]; 88 } 89 90 /* Get complete list of columns for color on each processor */ 91 cols = (int *) PetscMalloc( nctot*sizeof(int) ); CHKPTRQ(cols); 92 ierr = MPI_Allgatherv(is,n,MPI_INT,cols,ncolsonproc,disp,MPI_INT,mat->comm);CHKERRQ(ierr); 93 94 /* 95 for ( j=0; j<nctot; j++ ) { 96 printf("color %d %d col %d\n",i,j,cols[j]); 97 } 98 */ 99 100 /* 101 Mark all rows affect by these columns 102 */ 103 if (flg) {/*-----------------------------------------------------------------------------*/ 104 /* crude, slow version */ 105 PetscMemzero(rowhit,M*sizeof(int)); 106 /* loop over columns*/ 107 for ( j=0; j<nctot; j++ ) { 108 col = cols[j]; 109 if (col >= cstart && col < cend) { 110 /* column is in diagonal block of matrix */ 111 rows = A_cj + A_ci[col-cstart]; 112 m = A_ci[col-cstart+1] - A_ci[col-cstart]; 113 } else { 114 colb = aij->colmap[col] - 1; 115 if (colb == -1) { 116 m = 0; 117 } else { 118 rows = B_cj + B_ci[colb]; 119 m = B_ci[colb+1] - B_ci[colb]; 120 } 121 } 122 /* loop over columns marking them in rowhit */ 123 for ( k=0; k<m; k++ ) { 124 rowhit[*rows++] = col + 1; 125 } 126 } 127 128 /* 129 printf("for col %d found rows \n",i); 130 for ( j=0; j<M; j++ ) printf("rhow hit %d %d\n",j,rowhit[j]); 131 */ 132 133 /* count the number of hits */ 134 nrows = 0; 135 for ( j=0; j<M; j++ ) { 136 if (rowhit[j]) nrows++; 137 } 138 c->nrows[i] = nrows; 139 c->rows[i] = (int *) PetscMalloc((nrows+1)*sizeof(int)); CHKPTRQ(c->rows[i]); 140 c->columnsforrow[i] = (int *) PetscMalloc((nrows+1)*sizeof(int)); CHKPTRQ(c->columnsforrow[i]); 141 PLogObjectMemory(c,2*(nrows+1)*sizeof(int)); 142 nrows = 0; 143 for ( j=0; j<M; j++ ) { 144 if (rowhit[j]) { 145 c->rows[i][nrows] = j; 146 c->columnsforrow[i][nrows] = rowhit[j] - 1; 147 nrows++; 148 } 149 } 150 } else {/*-------------------------------------------------------------------------------*/ 151 /* efficient version, using rowhit as a linked list */ 152 int currentcol,fm,mfm; 153 rowhit[M] = M; 154 nrows = 0; 155 /* loop over columns*/ 156 for ( j=0; j<nctot; j++ ) { 157 col = cols[j]; 158 if (col >= cstart && col < cend) { 159 /* column is in diagonal block of matrix */ 160 rows = A_cj + A_ci[col-cstart]; 161 m = A_ci[col-cstart+1] - A_ci[col-cstart]; 162 } else { 163 colb = aij->colmap[col] - 1; 164 if (colb == -1) { 165 m = 0; 166 } else { 167 rows = B_cj + B_ci[colb]; 168 m = B_ci[colb+1] - B_ci[colb]; 169 } 170 } 171 /* loop over columns marking them in rowhit */ 172 fm = M; /* fm points to first entry in linked list */ 173 for ( k=0; k<m; k++ ) { 174 currentcol = *rows++; 175 /* is it already in the list? */ 176 do { 177 mfm = fm; 178 fm = rowhit[fm]; 179 } while (fm < currentcol); 180 /* not in list so add it */ 181 if (fm != currentcol) { 182 nrows++; 183 columnsforrow[currentcol] = col; 184 /* next three lines insert new entry into linked list */ 185 rowhit[mfm] = currentcol; 186 rowhit[currentcol] = fm; 187 fm = currentcol; 188 /* fm points to present position in list since we know the columns are sorted */ 189 } else { 190 SETERRQ(PETSC_ERR_PLIB,0,"Invalid coloring of matrix detected"); 191 } 192 } 193 } 194 c->nrows[i] = nrows; 195 c->rows[i] = (int *)PetscMalloc((nrows+1)*sizeof(int));CHKPTRQ(c->rows[i]); 196 c->columnsforrow[i] = (int *)PetscMalloc((nrows+1)*sizeof(int));CHKPTRQ(c->columnsforrow[i]); 197 PLogObjectMemory(c,(nrows+1)*sizeof(int)); 198 /* now store the linked list of rows into c->rows[i] */ 199 nrows = 0; 200 fm = rowhit[M]; 201 do { 202 c->rows[i][nrows] = fm; 203 c->columnsforrow[i][nrows++] = columnsforrow[fm]; 204 fm = rowhit[fm]; 205 } while (fm < M); 206 } /* ---------------------------------------------------------------------------------------*/ 207 PetscFree(cols); 208 } 209 PetscFree(rowhit); 210 PetscFree(columnsforrow); 211 PetscFree(ncolsonproc); 212 ierr = MatRestoreColumnIJ_SeqAIJ(aij->A,0,PETSC_FALSE,&ncols,&A_ci,&A_cj,&done); CHKERRQ(ierr); 213 ierr = MatRestoreColumnIJ_SeqAIJ(aij->B,0,PETSC_FALSE,&ncols,&B_ci,&B_cj,&done); CHKERRQ(ierr); 214 215 c->scale = (Scalar *) PetscMalloc( 2*mat->N*sizeof(Scalar) ); CHKPTRQ(c->scale); 216 PLogObjectMemory(c,2*mat->N*sizeof(Scalar)); 217 c->wscale = c->scale + mat->N; 218 PetscFunctionReturn(0); 219 } 220 221