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