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 c->ncolumns[i] = n; 59 if (n) { 60 ierr = PetscMalloc(n*sizeof(PetscInt),&c->columns[i]);CHKERRQ(ierr); 61 ierr = PetscLogObjectMemory(c,n*sizeof(PetscInt));CHKERRQ(ierr); 62 ierr = PetscMemcpy(c->columns[i],is,n*sizeof(PetscInt));CHKERRQ(ierr); 63 } else { 64 c->columns[i] = 0; 65 } 66 67 if (ctype == IS_COLORING_GLOBAL) { 68 /* Determine the total (parallel) number of columns of this color */ 69 ierr = MPI_Comm_size(((PetscObject)mat)->comm,&size);CHKERRQ(ierr); 70 ierr = PetscMalloc2(size,PetscMPIInt,&ncolsonproc,size,PetscMPIInt,&disp);CHKERRQ(ierr); 71 72 nn = PetscMPIIntCast(n); 73 ierr = MPI_Allgather(&nn,1,MPI_INT,ncolsonproc,1,MPI_INT,((PetscObject)mat)->comm);CHKERRQ(ierr); 74 nctot = 0; for (j=0; j<size; j++) {nctot += ncolsonproc[j];} 75 if (!nctot) { 76 ierr = PetscInfo(mat,"Coloring of matrix has some unneeded colors with no corresponding rows\n");CHKERRQ(ierr); 77 } 78 79 disp[0] = 0; 80 for (j=1; j<size; j++) { 81 disp[j] = disp[j-1] + ncolsonproc[j-1]; 82 } 83 84 /* Get complete list of columns for color on each processor */ 85 ierr = PetscMalloc((nctot+1)*sizeof(PetscInt),&cols);CHKERRQ(ierr); 86 ierr = MPI_Allgatherv((void*)is,n,MPIU_INT,cols,ncolsonproc,disp,MPIU_INT,((PetscObject)mat)->comm);CHKERRQ(ierr); 87 ierr = PetscFree2(ncolsonproc,disp);CHKERRQ(ierr); 88 } else if (ctype == IS_COLORING_GHOSTED) { 89 /* Determine local number of columns of this color on this process, including ghost points */ 90 nctot = n; 91 ierr = PetscMalloc((nctot+1)*sizeof(PetscInt),&cols);CHKERRQ(ierr); 92 ierr = PetscMemcpy(cols,is,n*sizeof(PetscInt));CHKERRQ(ierr); 93 } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Not provided for this MatFDColoring type"); 94 95 /* 96 Mark all rows affect by these columns 97 */ 98 /* Temporary option to allow for debugging/testing */ 99 flg = PETSC_FALSE; 100 ierr = PetscOptionsGetBool(PETSC_NULL,"-matfdcoloring_slow",&flg,PETSC_NULL);CHKERRQ(ierr); 101 if (!flg) {/*-----------------------------------------------------------------------------*/ 102 /* crude, fast version */ 103 ierr = PetscMemzero(rowhit,M*sizeof(PetscInt));CHKERRQ(ierr); 104 /* loop over columns*/ 105 for (j=0; j<nctot; j++) { 106 if (ctype == IS_COLORING_GHOSTED) { 107 col = ltog[cols[j]]; 108 } else { 109 col = cols[j]; 110 } 111 if (col >= cstart && col < cend) { 112 /* column is in diagonal block of matrix */ 113 rows = A_cj + A_ci[col-cstart]; 114 m = A_ci[col-cstart+1] - A_ci[col-cstart]; 115 } else { 116 #if defined (PETSC_USE_CTABLE) 117 ierr = PetscTableFind(aij->colmap,col+1,&colb);CHKERRQ(ierr); 118 colb --; 119 #else 120 colb = aij->colmap[col] - 1; 121 #endif 122 if (colb == -1) { 123 m = 0; 124 } else { 125 rows = B_cj + B_ci[colb]; 126 m = B_ci[colb+1] - B_ci[colb]; 127 } 128 } 129 /* loop over columns marking them in rowhit */ 130 for (k=0; k<m; k++) { 131 rowhit[*rows++] = col + 1; 132 } 133 } 134 135 /* count the number of hits */ 136 nrows = 0; 137 for (j=0; j<M; j++) { 138 if (rowhit[j]) nrows++; 139 } 140 c->nrows[i] = nrows; 141 ierr = PetscMalloc((nrows+1)*sizeof(PetscInt),&c->rows[i]);CHKERRQ(ierr); 142 ierr = PetscMalloc((nrows+1)*sizeof(PetscInt),&c->columnsforrow[i]);CHKERRQ(ierr); 143 ierr = PetscLogObjectMemory(c,2*(nrows+1)*sizeof(PetscInt));CHKERRQ(ierr); 144 nrows = 0; 145 for (j=0; j<M; j++) { 146 if (rowhit[j]) { 147 c->rows[i][nrows] = j; 148 c->columnsforrow[i][nrows] = rowhit[j] - 1; 149 nrows++; 150 } 151 } 152 } else {/*-------------------------------------------------------------------------------*/ 153 /* slow version, using rowhit as a linked list */ 154 PetscInt currentcol,fm,mfm; 155 rowhit[M] = M; 156 nrows = 0; 157 /* loop over columns*/ 158 for (j=0; j<nctot; j++) { 159 if (ctype == IS_COLORING_GHOSTED) { 160 col = ltog[cols[j]]; 161 } else { 162 col = cols[j]; 163 } 164 if (col >= cstart && col < cend) { 165 /* column is in diagonal block of matrix */ 166 rows = A_cj + A_ci[col-cstart]; 167 m = A_ci[col-cstart+1] - A_ci[col-cstart]; 168 } else { 169 #if defined (PETSC_USE_CTABLE) 170 ierr = PetscTableFind(aij->colmap,col+1,&colb);CHKERRQ(ierr); 171 colb --; 172 #else 173 colb = aij->colmap[col] - 1; 174 #endif 175 if (colb == -1) { 176 m = 0; 177 } else { 178 rows = B_cj + B_ci[colb]; 179 m = B_ci[colb+1] - B_ci[colb]; 180 } 181 } 182 183 /* loop over columns marking them in rowhit */ 184 fm = M; /* fm points to first entry in linked list */ 185 for (k=0; k<m; k++) { 186 currentcol = *rows++; 187 /* is it already in the list? */ 188 do { 189 mfm = fm; 190 fm = rowhit[fm]; 191 } while (fm < currentcol); 192 /* not in list so add it */ 193 if (fm != currentcol) { 194 nrows++; 195 columnsforrow[currentcol] = col; 196 /* next three lines insert new entry into linked list */ 197 rowhit[mfm] = currentcol; 198 rowhit[currentcol] = fm; 199 fm = currentcol; 200 /* fm points to present position in list since we know the columns are sorted */ 201 } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Invalid coloring of matrix detected"); 202 } 203 } 204 c->nrows[i] = nrows; 205 ierr = PetscMalloc((nrows+1)*sizeof(PetscInt),&c->rows[i]);CHKERRQ(ierr); 206 ierr = PetscMalloc((nrows+1)*sizeof(PetscInt),&c->columnsforrow[i]);CHKERRQ(ierr); 207 ierr = PetscLogObjectMemory(c,(nrows+1)*sizeof(PetscInt));CHKERRQ(ierr); 208 /* now store the linked list of rows into c->rows[i] */ 209 nrows = 0; 210 fm = rowhit[M]; 211 do { 212 c->rows[i][nrows] = fm; 213 c->columnsforrow[i][nrows++] = columnsforrow[fm]; 214 fm = rowhit[fm]; 215 } while (fm < M); 216 } /* ---------------------------------------------------------------------------------------*/ 217 ierr = PetscFree(cols);CHKERRQ(ierr); 218 } 219 220 /* Optimize by adding the vscale, and scaleforrow[][] fields */ 221 /* 222 vscale will contain the "diagonal" on processor scalings followed by the off processor 223 */ 224 if (ctype == IS_COLORING_GLOBAL) { 225 ierr = VecCreateGhost(((PetscObject)mat)->comm,aij->A->rmap->n,PETSC_DETERMINE,aij->B->cmap->n,aij->garray,&c->vscale);CHKERRQ(ierr); 226 ierr = PetscMalloc(c->ncolors*sizeof(PetscInt*),&c->vscaleforrow);CHKERRQ(ierr); 227 for (k=0; k<c->ncolors; k++) { 228 ierr = PetscMalloc((c->nrows[k]+1)*sizeof(PetscInt),&c->vscaleforrow[k]);CHKERRQ(ierr); 229 for (l=0; l<c->nrows[k]; l++) { 230 col = c->columnsforrow[k][l]; 231 if (col >= cstart && col < cend) { 232 /* column is in diagonal block of matrix */ 233 colb = col - cstart; 234 } else { 235 /* column is in "off-processor" part */ 236 #if defined (PETSC_USE_CTABLE) 237 ierr = PetscTableFind(aij->colmap,col+1,&colb);CHKERRQ(ierr); 238 colb --; 239 #else 240 colb = aij->colmap[col] - 1; 241 #endif 242 colb += cend - cstart; 243 } 244 c->vscaleforrow[k][l] = colb; 245 } 246 } 247 } else if (ctype == IS_COLORING_GHOSTED) { 248 /* Get gtol mapping */ 249 PetscInt N = mat->cmap->N, *gtol; 250 ierr = PetscMalloc((N+1)*sizeof(PetscInt),>ol);CHKERRQ(ierr); 251 for (i=0; i<N; i++) gtol[i] = -1; 252 for (i=0; i<map->n; i++) gtol[ltog[i]] = i; 253 254 c->vscale = 0; /* will be created in MatFDColoringApply() */ 255 ierr = PetscMalloc(c->ncolors*sizeof(PetscInt*),&c->vscaleforrow);CHKERRQ(ierr); 256 for (k=0; k<c->ncolors; k++) { 257 ierr = PetscMalloc((c->nrows[k]+1)*sizeof(PetscInt),&c->vscaleforrow[k]);CHKERRQ(ierr); 258 for (l=0; l<c->nrows[k]; l++) { 259 col = c->columnsforrow[k][l]; /* global column index */ 260 c->vscaleforrow[k][l] = gtol[col]; /* local column index */ 261 } 262 } 263 ierr = PetscFree(gtol);CHKERRQ(ierr); 264 } 265 ierr = ISColoringRestoreIS(iscoloring,&isa);CHKERRQ(ierr); 266 267 ierr = PetscFree(rowhit);CHKERRQ(ierr); 268 ierr = PetscFree(columnsforrow);CHKERRQ(ierr); 269 ierr = MatRestoreColumnIJ(aij->A,0,PETSC_FALSE,PETSC_FALSE,&ncols,&A_ci,&A_cj,&done);CHKERRQ(ierr); 270 ierr = MatRestoreColumnIJ(aij->B,0,PETSC_FALSE,PETSC_FALSE,&ncols,&B_ci,&B_cj,&done);CHKERRQ(ierr); 271 PetscFunctionReturn(0); 272 } 273 274 275 276 277 278 279