1 2 #include <../src/mat/impls/aij/mpi/mpiaij.h> 3 #include <../src/mat/impls/baij/mpi/mpibaij.h> 4 #include <petsc/private/isimpl.h> 5 6 #undef __FUNCT__ 7 #define __FUNCT__ "MatFDColoringApply_BAIJ" 8 PetscErrorCode MatFDColoringApply_BAIJ(Mat J,MatFDColoring coloring,Vec x1,void *sctx) 9 { 10 PetscErrorCode (*f)(void*,Vec,Vec,void*) = (PetscErrorCode (*)(void*,Vec,Vec,void*))coloring->f; 11 PetscErrorCode ierr; 12 PetscInt k,cstart,cend,l,row,col,nz,spidx,i,j; 13 PetscScalar dx=0.0,*w3_array,*dy_i,*dy=coloring->dy; 14 PetscScalar *vscale_array; 15 const PetscScalar *xx; 16 PetscReal epsilon=coloring->error_rel,umin=coloring->umin,unorm; 17 Vec w1=coloring->w1,w2=coloring->w2,w3,vscale=coloring->vscale; 18 void *fctx=coloring->fctx; 19 PetscInt ctype=coloring->ctype,nxloc,nrows_k; 20 PetscScalar *valaddr; 21 MatEntry *Jentry=coloring->matentry; 22 MatEntry2 *Jentry2=coloring->matentry2; 23 const PetscInt ncolors=coloring->ncolors,*ncolumns=coloring->ncolumns,*nrows=coloring->nrows; 24 PetscInt bs=J->rmap->bs; 25 26 PetscFunctionBegin; 27 /* (1) Set w1 = F(x1) */ 28 if (!coloring->fset) { 29 ierr = PetscLogEventBegin(MAT_FDColoringFunction,0,0,0,0);CHKERRQ(ierr); 30 ierr = (*f)(sctx,x1,w1,fctx);CHKERRQ(ierr); 31 ierr = PetscLogEventEnd(MAT_FDColoringFunction,0,0,0,0);CHKERRQ(ierr); 32 } else { 33 coloring->fset = PETSC_FALSE; 34 } 35 36 /* (2) Compute vscale = 1./dx - the local scale factors, including ghost points */ 37 ierr = VecGetLocalSize(x1,&nxloc);CHKERRQ(ierr); 38 if (coloring->htype[0] == 'w') { 39 /* vscale = dx is a constant scalar */ 40 ierr = VecNorm(x1,NORM_2,&unorm);CHKERRQ(ierr); 41 dx = 1.0/(PetscSqrtReal(1.0 + unorm)*epsilon); 42 } else { 43 ierr = VecGetArrayRead(x1,&xx);CHKERRQ(ierr); 44 ierr = VecGetArray(vscale,&vscale_array);CHKERRQ(ierr); 45 for (col=0; col<nxloc; col++) { 46 dx = xx[col]; 47 if (PetscAbsScalar(dx) < umin) { 48 if (PetscRealPart(dx) >= 0.0) dx = umin; 49 else if (PetscRealPart(dx) < 0.0 ) dx = -umin; 50 } 51 dx *= epsilon; 52 vscale_array[col] = 1.0/dx; 53 } 54 ierr = VecRestoreArrayRead(x1,&xx);CHKERRQ(ierr); 55 ierr = VecRestoreArray(vscale,&vscale_array);CHKERRQ(ierr); 56 } 57 if (ctype == IS_COLORING_GLOBAL && coloring->htype[0] == 'd') { 58 ierr = VecGhostUpdateBegin(vscale,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 59 ierr = VecGhostUpdateEnd(vscale,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 60 } 61 62 /* (3) Loop over each color */ 63 if (!coloring->w3) { 64 ierr = VecDuplicate(x1,&coloring->w3);CHKERRQ(ierr); 65 ierr = PetscLogObjectParent((PetscObject)coloring,(PetscObject)coloring->w3);CHKERRQ(ierr); 66 } 67 w3 = coloring->w3; 68 69 ierr = VecGetOwnershipRange(x1,&cstart,&cend);CHKERRQ(ierr); /* used by ghosted vscale */ 70 if (vscale) { 71 ierr = VecGetArray(vscale,&vscale_array);CHKERRQ(ierr); 72 } 73 nz = 0; 74 for (k=0; k<ncolors; k++) { 75 coloring->currentcolor = k; 76 77 /* 78 (3-1) Loop over each column associated with color 79 adding the perturbation to the vector w3 = x1 + dx. 80 */ 81 ierr = VecCopy(x1,w3);CHKERRQ(ierr); 82 dy_i = dy; 83 for (i=0; i<bs; i++) { /* Loop over a block of columns */ 84 ierr = VecGetArray(w3,&w3_array);CHKERRQ(ierr); 85 if (ctype == IS_COLORING_GLOBAL) w3_array -= cstart; /* shift pointer so global index can be used */ 86 if (coloring->htype[0] == 'w') { 87 for (l=0; l<ncolumns[k]; l++) { 88 col = i + bs*coloring->columns[k][l]; /* local column (in global index!) of the matrix we are probing for */ 89 w3_array[col] += 1.0/dx; 90 if (i) w3_array[col-1] -= 1.0/dx; /* resume original w3[col-1] */ 91 } 92 } else { /* htype == 'ds' */ 93 vscale_array -= cstart; /* shift pointer so global index can be used */ 94 for (l=0; l<ncolumns[k]; l++) { 95 col = i + bs*coloring->columns[k][l]; /* local column (in global index!) of the matrix we are probing for */ 96 w3_array[col] += 1.0/vscale_array[col]; 97 if (i) w3_array[col-1] -= 1.0/vscale_array[col-1]; /* resume original w3[col-1] */ 98 } 99 vscale_array += cstart; 100 } 101 if (ctype == IS_COLORING_GLOBAL) w3_array += cstart; 102 ierr = VecRestoreArray(w3,&w3_array);CHKERRQ(ierr); 103 104 /* 105 (3-2) Evaluate function at w3 = x1 + dx (here dx is a vector of perturbations) 106 w2 = F(x1 + dx) - F(x1) 107 */ 108 ierr = PetscLogEventBegin(MAT_FDColoringFunction,0,0,0,0);CHKERRQ(ierr); 109 ierr = VecPlaceArray(w2,dy_i);CHKERRQ(ierr); /* place w2 to the array dy_i */ 110 ierr = (*f)(sctx,w3,w2,fctx);CHKERRQ(ierr); 111 ierr = PetscLogEventEnd(MAT_FDColoringFunction,0,0,0,0);CHKERRQ(ierr); 112 ierr = VecAXPY(w2,-1.0,w1);CHKERRQ(ierr); 113 ierr = VecResetArray(w2);CHKERRQ(ierr); 114 dy_i += nxloc; /* points to dy+i*nxloc */ 115 } 116 117 /* 118 (3-3) Loop over rows of vector, putting results into Jacobian matrix 119 */ 120 nrows_k = nrows[k]; 121 if (coloring->htype[0] == 'w') { 122 for (l=0; l<nrows_k; l++) { 123 row = bs*Jentry2[nz].row; /* local row index */ 124 valaddr = Jentry2[nz++].valaddr; 125 spidx = 0; 126 dy_i = dy; 127 for (i=0; i<bs; i++) { /* column of the block */ 128 for (j=0; j<bs; j++) { /* row of the block */ 129 valaddr[spidx++] = dy_i[row+j]*dx; 130 } 131 dy_i += nxloc; /* points to dy+i*nxloc */ 132 } 133 } 134 } else { /* htype == 'ds' */ 135 for (l=0; l<nrows_k; l++) { 136 row = bs*Jentry[nz].row; /* local row index */ 137 col = bs*Jentry[nz].col; /* local column index */ 138 valaddr = Jentry[nz++].valaddr; 139 spidx = 0; 140 dy_i = dy; 141 for (i=0; i<bs; i++) { /* column of the block */ 142 for (j=0; j<bs; j++) { /* row of the block */ 143 valaddr[spidx++] = dy_i[row+j]*vscale_array[col+i]; 144 } 145 dy_i += nxloc; /* points to dy+i*nxloc */ 146 } 147 } 148 } 149 } 150 ierr = MatAssemblyBegin(J,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 151 ierr = MatAssemblyEnd(J,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 152 if (vscale) { 153 ierr = VecRestoreArray(vscale,&vscale_array);CHKERRQ(ierr); 154 } 155 156 coloring->currentcolor = -1; 157 PetscFunctionReturn(0); 158 } 159 160 #undef __FUNCT__ 161 #define __FUNCT__ "MatFDColoringApply_AIJ" 162 PetscErrorCode MatFDColoringApply_AIJ(Mat J,MatFDColoring coloring,Vec x1,void *sctx) 163 { 164 PetscErrorCode (*f)(void*,Vec,Vec,void*) = (PetscErrorCode (*)(void*,Vec,Vec,void*))coloring->f; 165 PetscErrorCode ierr; 166 PetscInt k,cstart,cend,l,row,col,nz; 167 PetscScalar dx=0.0,*y,*w3_array; 168 const PetscScalar *xx; 169 PetscScalar *vscale_array; 170 PetscReal epsilon=coloring->error_rel,umin=coloring->umin,unorm; 171 Vec w1=coloring->w1,w2=coloring->w2,w3,vscale=coloring->vscale; 172 void *fctx=coloring->fctx; 173 PetscInt ctype=coloring->ctype,nxloc,nrows_k; 174 MatEntry *Jentry=coloring->matentry; 175 MatEntry2 *Jentry2=coloring->matentry2; 176 const PetscInt ncolors=coloring->ncolors,*ncolumns=coloring->ncolumns,*nrows=coloring->nrows; 177 178 PetscFunctionBegin; 179 /* (1) Set w1 = F(x1) */ 180 if (!coloring->fset) { 181 ierr = PetscLogEventBegin(MAT_FDColoringFunction,0,0,0,0);CHKERRQ(ierr); 182 ierr = (*f)(sctx,x1,w1,fctx);CHKERRQ(ierr); 183 ierr = PetscLogEventEnd(MAT_FDColoringFunction,0,0,0,0);CHKERRQ(ierr); 184 } else { 185 coloring->fset = PETSC_FALSE; 186 } 187 188 /* (2) Compute vscale = 1./dx - the local scale factors, including ghost points */ 189 if (coloring->htype[0] == 'w') { 190 /* vscale = 1./dx is a constant scalar */ 191 ierr = VecNorm(x1,NORM_2,&unorm);CHKERRQ(ierr); 192 dx = 1.0/(PetscSqrtReal(1.0 + unorm)*epsilon); 193 } else { 194 ierr = VecGetLocalSize(x1,&nxloc);CHKERRQ(ierr); 195 ierr = VecGetArrayRead(x1,&xx);CHKERRQ(ierr); 196 ierr = VecGetArray(vscale,&vscale_array);CHKERRQ(ierr); 197 for (col=0; col<nxloc; col++) { 198 dx = xx[col]; 199 if (PetscAbsScalar(dx) < umin) { 200 if (PetscRealPart(dx) >= 0.0) dx = umin; 201 else if (PetscRealPart(dx) < 0.0 ) dx = -umin; 202 } 203 dx *= epsilon; 204 vscale_array[col] = 1.0/dx; 205 } 206 ierr = VecRestoreArrayRead(x1,&xx);CHKERRQ(ierr); 207 ierr = VecRestoreArray(vscale,&vscale_array);CHKERRQ(ierr); 208 } 209 if (ctype == IS_COLORING_GLOBAL && coloring->htype[0] == 'd') { 210 ierr = VecGhostUpdateBegin(vscale,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 211 ierr = VecGhostUpdateEnd(vscale,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 212 } 213 214 /* (3) Loop over each color */ 215 if (!coloring->w3) { 216 ierr = VecDuplicate(x1,&coloring->w3);CHKERRQ(ierr); 217 ierr = PetscLogObjectParent((PetscObject)coloring,(PetscObject)coloring->w3);CHKERRQ(ierr); 218 } 219 w3 = coloring->w3; 220 221 ierr = VecGetOwnershipRange(x1,&cstart,&cend);CHKERRQ(ierr); /* used by ghosted vscale */ 222 if (vscale) { 223 ierr = VecGetArray(vscale,&vscale_array);CHKERRQ(ierr); 224 } 225 nz = 0; 226 227 if (coloring->bcols > 1) { /* use blocked insertion of Jentry */ 228 PetscInt i,m=J->rmap->n,nbcols,bcols=coloring->bcols; 229 PetscScalar *dy=coloring->dy,*dy_k; 230 231 nbcols = 0; 232 for (k=0; k<ncolors; k+=bcols) { 233 coloring->currentcolor = k; 234 235 /* 236 (3-1) Loop over each column associated with color 237 adding the perturbation to the vector w3 = x1 + dx. 238 */ 239 240 dy_k = dy; 241 if (k + bcols > ncolors) bcols = ncolors - k; 242 for (i=0; i<bcols; i++) { 243 244 ierr = VecCopy(x1,w3);CHKERRQ(ierr); 245 ierr = VecGetArray(w3,&w3_array);CHKERRQ(ierr); 246 if (ctype == IS_COLORING_GLOBAL) w3_array -= cstart; /* shift pointer so global index can be used */ 247 if (coloring->htype[0] == 'w') { 248 for (l=0; l<ncolumns[k+i]; l++) { 249 col = coloring->columns[k+i][l]; /* local column (in global index!) of the matrix we are probing for */ 250 w3_array[col] += 1.0/dx; 251 } 252 } else { /* htype == 'ds' */ 253 vscale_array -= cstart; /* shift pointer so global index can be used */ 254 for (l=0; l<ncolumns[k+i]; l++) { 255 col = coloring->columns[k+i][l]; /* local column (in global index!) of the matrix we are probing for */ 256 w3_array[col] += 1.0/vscale_array[col]; 257 } 258 vscale_array += cstart; 259 } 260 if (ctype == IS_COLORING_GLOBAL) w3_array += cstart; 261 ierr = VecRestoreArray(w3,&w3_array);CHKERRQ(ierr); 262 263 /* 264 (3-2) Evaluate function at w3 = x1 + dx (here dx is a vector of perturbations) 265 w2 = F(x1 + dx) - F(x1) 266 */ 267 ierr = PetscLogEventBegin(MAT_FDColoringFunction,0,0,0,0);CHKERRQ(ierr); 268 ierr = VecPlaceArray(w2,dy_k);CHKERRQ(ierr); /* place w2 to the array dy_i */ 269 ierr = (*f)(sctx,w3,w2,fctx);CHKERRQ(ierr); 270 ierr = PetscLogEventEnd(MAT_FDColoringFunction,0,0,0,0);CHKERRQ(ierr); 271 ierr = VecAXPY(w2,-1.0,w1);CHKERRQ(ierr); 272 ierr = VecResetArray(w2);CHKERRQ(ierr); 273 dy_k += m; /* points to dy+i*nxloc */ 274 } 275 276 /* 277 (3-3) Loop over block rows of vector, putting results into Jacobian matrix 278 */ 279 nrows_k = nrows[nbcols++]; 280 ierr = VecGetArray(w2,&y);CHKERRQ(ierr); 281 282 if (coloring->htype[0] == 'w') { 283 for (l=0; l<nrows_k; l++) { 284 row = Jentry2[nz].row; /* local row index */ 285 *(Jentry2[nz++].valaddr) = dy[row]*dx; 286 } 287 } else { /* htype == 'ds' */ 288 for (l=0; l<nrows_k; l++) { 289 row = Jentry[nz].row; /* local row index */ 290 *(Jentry[nz].valaddr) = dy[row]*vscale_array[Jentry[nz].col]; 291 nz++; 292 } 293 } 294 ierr = VecRestoreArray(w2,&y);CHKERRQ(ierr); 295 } 296 } else { /* bcols == 1 */ 297 for (k=0; k<ncolors; k++) { 298 coloring->currentcolor = k; 299 300 /* 301 (3-1) Loop over each column associated with color 302 adding the perturbation to the vector w3 = x1 + dx. 303 */ 304 ierr = VecCopy(x1,w3);CHKERRQ(ierr); 305 ierr = VecGetArray(w3,&w3_array);CHKERRQ(ierr); 306 if (ctype == IS_COLORING_GLOBAL) w3_array -= cstart; /* shift pointer so global index can be used */ 307 if (coloring->htype[0] == 'w') { 308 for (l=0; l<ncolumns[k]; l++) { 309 col = coloring->columns[k][l]; /* local column (in global index!) of the matrix we are probing for */ 310 w3_array[col] += 1.0/dx; 311 } 312 } else { /* htype == 'ds' */ 313 vscale_array -= cstart; /* shift pointer so global index can be used */ 314 for (l=0; l<ncolumns[k]; l++) { 315 col = coloring->columns[k][l]; /* local column (in global index!) of the matrix we are probing for */ 316 w3_array[col] += 1.0/vscale_array[col]; 317 } 318 vscale_array += cstart; 319 } 320 if (ctype == IS_COLORING_GLOBAL) w3_array += cstart; 321 ierr = VecRestoreArray(w3,&w3_array);CHKERRQ(ierr); 322 323 /* 324 (3-2) Evaluate function at w3 = x1 + dx (here dx is a vector of perturbations) 325 w2 = F(x1 + dx) - F(x1) 326 */ 327 ierr = PetscLogEventBegin(MAT_FDColoringFunction,0,0,0,0);CHKERRQ(ierr); 328 ierr = (*f)(sctx,w3,w2,fctx);CHKERRQ(ierr); 329 ierr = PetscLogEventEnd(MAT_FDColoringFunction,0,0,0,0);CHKERRQ(ierr); 330 ierr = VecAXPY(w2,-1.0,w1);CHKERRQ(ierr); 331 332 /* 333 (3-3) Loop over rows of vector, putting results into Jacobian matrix 334 */ 335 nrows_k = nrows[k]; 336 ierr = VecGetArray(w2,&y);CHKERRQ(ierr); 337 if (coloring->htype[0] == 'w') { 338 for (l=0; l<nrows_k; l++) { 339 row = Jentry2[nz].row; /* local row index */ 340 *(Jentry2[nz++].valaddr) = y[row]*dx; 341 } 342 } else { /* htype == 'ds' */ 343 for (l=0; l<nrows_k; l++) { 344 row = Jentry[nz].row; /* local row index */ 345 *(Jentry[nz].valaddr) = y[row]*vscale_array[Jentry[nz].col]; 346 nz++; 347 } 348 } 349 ierr = VecRestoreArray(w2,&y);CHKERRQ(ierr); 350 } 351 } 352 353 ierr = MatAssemblyBegin(J,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 354 ierr = MatAssemblyEnd(J,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 355 if (vscale) { 356 ierr = VecRestoreArray(vscale,&vscale_array);CHKERRQ(ierr); 357 } 358 coloring->currentcolor = -1; 359 PetscFunctionReturn(0); 360 } 361 362 #undef __FUNCT__ 363 #define __FUNCT__ "MatFDColoringSetUp_MPIXAIJ" 364 PetscErrorCode MatFDColoringSetUp_MPIXAIJ(Mat mat,ISColoring iscoloring,MatFDColoring c) 365 { 366 PetscErrorCode ierr; 367 PetscMPIInt size,*ncolsonproc,*disp,nn,rank; 368 PetscInt i,n,nrows,nrows_i,j,k,m,ncols,col,*rowhit,cstart,cend,colb; 369 const PetscInt *is,*A_ci,*A_cj,*B_ci,*B_cj,*row=NULL,*ltog=NULL; 370 PetscInt nis=iscoloring->n,nctot,*cols,*cols_new; 371 IS *isa; 372 ISLocalToGlobalMapping map=mat->cmap->mapping; 373 PetscInt ctype=c->ctype,*spidxA,*spidxB,nz,bs,bs2,spidx; 374 Mat A,B; 375 PetscScalar *A_val,*B_val,**valaddrhit; 376 MatEntry *Jentry; 377 MatEntry2 *Jentry2; 378 PetscBool isBAIJ; 379 PetscInt bcols=c->bcols; 380 MPI_Comm comm; 381 PetscMPIInt tag,nrecvs,nsends,proc; 382 MPI_Request *rwaits = NULL,*swaits = NULL; 383 MPI_Status status; 384 #if defined(PETSC_USE_CTABLE) 385 PetscTable colmap=NULL; 386 #else 387 PetscInt *colmap=NULL; /* local col number of off-diag col */ 388 #endif 389 390 PetscFunctionBegin; 391 ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr); 392 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 393 394 if (ctype == IS_COLORING_GHOSTED) { 395 if (!map) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_ARG_INCOMP,"When using ghosted differencing matrix must have local to global mapping provided with MatSetLocalToGlobalMapping"); 396 ierr = ISLocalToGlobalMappingGetIndices(map,<og);CHKERRQ(ierr); 397 } 398 399 ierr = MatGetBlockSize(mat,&bs);CHKERRQ(ierr); 400 ierr = PetscObjectTypeCompare((PetscObject)mat,MATMPIBAIJ,&isBAIJ);CHKERRQ(ierr); 401 if (isBAIJ) { 402 Mat_MPIBAIJ *baij=(Mat_MPIBAIJ*)mat->data; 403 Mat_SeqBAIJ *spA,*spB; 404 A = baij->A; spA = (Mat_SeqBAIJ*)A->data; A_val = spA->a; 405 B = baij->B; spB = (Mat_SeqBAIJ*)B->data; B_val = spB->a; 406 nz = spA->nz + spB->nz; /* total nonzero entries of mat */ 407 if (!baij->colmap) { 408 ierr = MatCreateColmap_MPIBAIJ_Private(mat);CHKERRQ(ierr); 409 } 410 colmap = baij->colmap; 411 ierr = MatGetColumnIJ_SeqBAIJ_Color(A,0,PETSC_FALSE,PETSC_FALSE,&ncols,&A_ci,&A_cj,&spidxA,NULL);CHKERRQ(ierr); 412 ierr = MatGetColumnIJ_SeqBAIJ_Color(B,0,PETSC_FALSE,PETSC_FALSE,&ncols,&B_ci,&B_cj,&spidxB,NULL);CHKERRQ(ierr); 413 414 if (ctype == IS_COLORING_GLOBAL && c->htype[0] == 'd') { /* create vscale for storing dx */ 415 PetscInt *garray; 416 ierr = PetscMalloc1(B->cmap->n,&garray);CHKERRQ(ierr); 417 for (i=0; i<baij->B->cmap->n/bs; i++) { 418 for (j=0; j<bs; j++) { 419 garray[i*bs+j] = bs*baij->garray[i]+j; 420 } 421 } 422 ierr = VecCreateGhost(PetscObjectComm((PetscObject)mat),mat->cmap->n,PETSC_DETERMINE,B->cmap->n,garray,&c->vscale);CHKERRQ(ierr); 423 ierr = PetscFree(garray);CHKERRQ(ierr); 424 } 425 } else { 426 Mat_MPIAIJ *aij=(Mat_MPIAIJ*)mat->data; 427 Mat_SeqAIJ *spA,*spB; 428 A = aij->A; spA = (Mat_SeqAIJ*)A->data; A_val = spA->a; 429 B = aij->B; spB = (Mat_SeqAIJ*)B->data; B_val = spB->a; 430 nz = spA->nz + spB->nz; /* total nonzero entries of mat */ 431 if (!aij->colmap) { 432 /* Allow access to data structures of local part of matrix 433 - creates aij->colmap which maps global column number to local number in part B */ 434 ierr = MatCreateColmap_MPIAIJ_Private(mat);CHKERRQ(ierr); 435 } 436 colmap = aij->colmap; 437 ierr = MatGetColumnIJ_SeqAIJ_Color(A,0,PETSC_FALSE,PETSC_FALSE,&ncols,&A_ci,&A_cj,&spidxA,NULL);CHKERRQ(ierr); 438 ierr = MatGetColumnIJ_SeqAIJ_Color(B,0,PETSC_FALSE,PETSC_FALSE,&ncols,&B_ci,&B_cj,&spidxB,NULL);CHKERRQ(ierr); 439 440 bs = 1; /* only bs=1 is supported for non MPIBAIJ matrix */ 441 442 if (ctype == IS_COLORING_GLOBAL && c->htype[0] == 'd') { /* create vscale for storing dx */ 443 ierr = VecCreateGhost(PetscObjectComm((PetscObject)mat),mat->cmap->n,PETSC_DETERMINE,B->cmap->n,aij->garray,&c->vscale);CHKERRQ(ierr); 444 } 445 } 446 447 m = mat->rmap->n/bs; 448 cstart = mat->cmap->rstart/bs; 449 cend = mat->cmap->rend/bs; 450 451 ierr = PetscMalloc1(nis,&c->ncolumns);CHKERRQ(ierr); 452 ierr = PetscMalloc1(nis,&c->columns);CHKERRQ(ierr); 453 ierr = PetscMalloc1(nis,&c->nrows);CHKERRQ(ierr); 454 ierr = PetscLogObjectMemory((PetscObject)c,3*nis*sizeof(PetscInt));CHKERRQ(ierr); 455 456 if (c->htype[0] == 'd') { 457 ierr = PetscMalloc1(nz,&Jentry);CHKERRQ(ierr); 458 ierr = PetscLogObjectMemory((PetscObject)c,nz*sizeof(MatEntry));CHKERRQ(ierr); 459 c->matentry = Jentry; 460 } else if (c->htype[0] == 'w') { 461 ierr = PetscMalloc1(nz,&Jentry2);CHKERRQ(ierr); 462 ierr = PetscLogObjectMemory((PetscObject)c,nz*sizeof(MatEntry2));CHKERRQ(ierr); 463 c->matentry2 = Jentry2; 464 } else SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"htype is not supported"); 465 466 ierr = PetscMalloc2(m+1,&rowhit,m+1,&valaddrhit);CHKERRQ(ierr); 467 nz = 0; 468 ierr = ISColoringGetIS(iscoloring,PETSC_IGNORE,&isa);CHKERRQ(ierr); 469 for (i=0; i<nis; i++) { /* for each local color */ 470 ierr = ISGetLocalSize(isa[i],&n);CHKERRQ(ierr); 471 ierr = ISGetIndices(isa[i],&is);CHKERRQ(ierr); 472 473 c->ncolumns[i] = n; /* local number of columns of this color on this process */ 474 if (n) { 475 ierr = PetscMalloc1(n,&c->columns[i]);CHKERRQ(ierr); 476 ierr = PetscLogObjectMemory((PetscObject)c,n*sizeof(PetscInt));CHKERRQ(ierr); 477 ierr = PetscMemcpy(c->columns[i],is,n*sizeof(PetscInt));CHKERRQ(ierr); 478 } else { 479 c->columns[i] = 0; 480 } 481 482 if (ctype == IS_COLORING_GLOBAL) { 483 /* Determine nctot, the total (parallel) number of columns of this color */ 484 ierr = MPI_Comm_size(PetscObjectComm((PetscObject)mat),&size);CHKERRQ(ierr); 485 ierr = PetscMalloc2(size,&ncolsonproc,size,&disp);CHKERRQ(ierr); 486 487 /* ncolsonproc[j]: local ncolumns on proc[j] of this color */ 488 ierr = PetscMPIIntCast(n,&nn);CHKERRQ(ierr); 489 ierr = MPI_Allgather(&nn,1,MPI_INT,ncolsonproc,1,MPI_INT,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 490 nctot = 0; 491 for (j=0; j<size; j++) nctot += ncolsonproc[j]; 492 if (!nctot) { 493 ierr = PetscInfo(mat,"Coloring of matrix has some unneeded colors with no corresponding rows\n");CHKERRQ(ierr); 494 } 495 496 disp[0] = 0; 497 for (j=1; j<size; j++) { 498 disp[j] = disp[j-1] + ncolsonproc[j-1]; 499 } 500 501 /* Get cols, the complete list of columns for this color on each process */ 502 ierr = PetscMalloc1(nctot+1,&cols);CHKERRQ(ierr); 503 504 /****************** non-scalable !!! *********************/ 505 /* ierr = MPI_Allgatherv((void*)is,n,MPIU_INT,cols,ncolsonproc,disp,MPIU_INT,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); */ 506 cols_new = cols; 507 508 nrecvs = size-1; 509 ierr = PetscCommGetNewTag(comm,&tag);CHKERRQ(ierr); 510 ierr = PetscMalloc2(nrecvs,&rwaits,nrecvs,&swaits);CHKERRQ(ierr); 511 512 nrecvs = 0; 513 for (proc=0; proc<size; proc++) { 514 if (proc == rank || ncolsonproc[proc] == 0 ) continue; 515 /* printf("[%d] Irecv %d from [%d] \n",rank,ncolsonproc[proc],proc); */ 516 ierr = MPI_Irecv(cols_new+disp[proc],ncolsonproc[proc],MPIU_INT,proc,tag,comm,rwaits+nrecvs);CHKERRQ(ierr); 517 nrecvs++; 518 } 519 520 nsends = 0; 521 for (proc=0; proc<size; proc++) { 522 if (proc == rank || n == 0) continue; 523 /* printf("[%d] Isend %d is to [%d] \n",rank,n,proc); */ 524 ierr = MPI_Isend((void*)is,n,MPIU_INT,proc,tag,comm,swaits+nsends);CHKERRQ(ierr); 525 nsends++; 526 } 527 528 for (j=0; j<ncolsonproc[rank]; j++) { 529 cols_new[disp[rank] + j] = is[j]; 530 } 531 532 j = nrecvs; 533 while (j--) { 534 ierr = MPI_Waitany(nrecvs,rwaits,&k,&status);CHKERRQ(ierr); 535 } 536 if (nsends) {ierr = MPI_Waitall(nrecvs,swaits,MPI_STATUSES_IGNORE);CHKERRQ(ierr);} 537 ierr = PetscFree2(rwaits,swaits);CHKERRQ(ierr); 538 539 /******************/ 540 ierr = PetscFree2(ncolsonproc,disp);CHKERRQ(ierr); 541 } else if (ctype == IS_COLORING_GHOSTED) { 542 /* Determine local number of columns of this color on this process, including ghost points */ 543 nctot = n; 544 ierr = PetscMalloc1(nctot+1,&cols);CHKERRQ(ierr); 545 ierr = PetscMemcpy(cols,is,n*sizeof(PetscInt));CHKERRQ(ierr); 546 } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Not provided for this MatFDColoring type"); 547 548 /* Mark all rows affect by these columns */ 549 ierr = PetscMemzero(rowhit,m*sizeof(PetscInt));CHKERRQ(ierr); 550 bs2 = bs*bs; 551 nrows_i = 0; 552 for (j=0; j<nctot; j++) { /* loop over columns*/ 553 if (ctype == IS_COLORING_GHOSTED) { 554 col = ltog[cols[j]]; 555 } else { 556 col = cols[j]; 557 } 558 if (col >= cstart && col < cend) { /* column is in A, diagonal block of mat */ 559 row = A_cj + A_ci[col-cstart]; 560 nrows = A_ci[col-cstart+1] - A_ci[col-cstart]; 561 nrows_i += nrows; 562 /* loop over columns of A marking them in rowhit */ 563 for (k=0; k<nrows; k++) { 564 /* set valaddrhit for part A */ 565 spidx = bs2*spidxA[A_ci[col-cstart] + k]; 566 valaddrhit[*row] = &A_val[spidx]; 567 rowhit[*row++] = col - cstart + 1; /* local column index */ 568 } 569 } else { /* column is in B, off-diagonal block of mat */ 570 #if defined(PETSC_USE_CTABLE) 571 ierr = PetscTableFind(colmap,col+1,&colb);CHKERRQ(ierr); 572 colb--; 573 #else 574 colb = colmap[col] - 1; /* local column index */ 575 #endif 576 if (colb == -1) { 577 nrows = 0; 578 } else { 579 colb = colb/bs; 580 row = B_cj + B_ci[colb]; 581 nrows = B_ci[colb+1] - B_ci[colb]; 582 } 583 nrows_i += nrows; 584 /* loop over columns of B marking them in rowhit */ 585 for (k=0; k<nrows; k++) { 586 /* set valaddrhit for part B */ 587 spidx = bs2*spidxB[B_ci[colb] + k]; 588 valaddrhit[*row] = &B_val[spidx]; 589 rowhit[*row++] = colb + 1 + cend - cstart; /* local column index */ 590 } 591 } 592 } 593 ierr = PetscFree(cols);CHKERRQ(ierr); 594 c->nrows[i] = nrows_i; 595 596 if (c->htype[0] == 'd') { 597 for (j=0; j<m; j++) { 598 if (rowhit[j]) { 599 Jentry[nz].row = j; /* local row index */ 600 Jentry[nz].col = rowhit[j] - 1; /* local column index */ 601 Jentry[nz].valaddr = valaddrhit[j]; /* address of mat value for this entry */ 602 nz++; 603 } 604 } 605 } else { /* c->htype == 'wp' */ 606 for (j=0; j<m; j++) { 607 if (rowhit[j]) { 608 Jentry2[nz].row = j; /* local row index */ 609 Jentry2[nz].valaddr = valaddrhit[j]; /* address of mat value for this entry */ 610 nz++; 611 } 612 } 613 } 614 } 615 616 if (bcols > 1) { /* reorder Jentry for faster MatFDColoringApply() */ 617 ierr = MatFDColoringSetUpBlocked_AIJ_Private(mat,c,nz);CHKERRQ(ierr); 618 } 619 620 if (isBAIJ) { 621 ierr = MatRestoreColumnIJ_SeqBAIJ_Color(A,0,PETSC_FALSE,PETSC_FALSE,&ncols,&A_ci,&A_cj,&spidxA,NULL);CHKERRQ(ierr); 622 ierr = MatRestoreColumnIJ_SeqBAIJ_Color(B,0,PETSC_FALSE,PETSC_FALSE,&ncols,&B_ci,&B_cj,&spidxB,NULL);CHKERRQ(ierr); 623 ierr = PetscMalloc1(bs*mat->rmap->n,&c->dy);CHKERRQ(ierr); 624 } else { 625 ierr = MatRestoreColumnIJ_SeqAIJ_Color(A,0,PETSC_FALSE,PETSC_FALSE,&ncols,&A_ci,&A_cj,&spidxA,NULL);CHKERRQ(ierr); 626 ierr = MatRestoreColumnIJ_SeqAIJ_Color(B,0,PETSC_FALSE,PETSC_FALSE,&ncols,&B_ci,&B_cj,&spidxB,NULL);CHKERRQ(ierr); 627 } 628 629 ierr = ISColoringRestoreIS(iscoloring,&isa);CHKERRQ(ierr); 630 ierr = PetscFree2(rowhit,valaddrhit);CHKERRQ(ierr); 631 632 if (ctype == IS_COLORING_GHOSTED) { 633 ierr = ISLocalToGlobalMappingRestoreIndices(map,<og);CHKERRQ(ierr); 634 } 635 ierr = PetscInfo3(c,"ncolors %D, brows %D and bcols %D are used.\n",c->ncolors,c->brows,c->bcols);CHKERRQ(ierr); 636 PetscFunctionReturn(0); 637 } 638 639 #undef __FUNCT__ 640 #define __FUNCT__ "MatFDColoringCreate_MPIXAIJ" 641 PetscErrorCode MatFDColoringCreate_MPIXAIJ(Mat mat,ISColoring iscoloring,MatFDColoring c) 642 { 643 PetscErrorCode ierr; 644 PetscInt bs,nis=iscoloring->n,m=mat->rmap->n; 645 PetscBool isBAIJ; 646 647 PetscFunctionBegin; 648 /* set default brows and bcols for speedup inserting the dense matrix into sparse Jacobian; 649 bcols is chosen s.t. dy-array takes 50% of memory space as mat */ 650 ierr = MatGetBlockSize(mat,&bs);CHKERRQ(ierr); 651 ierr = PetscObjectTypeCompare((PetscObject)mat,MATMPIBAIJ,&isBAIJ);CHKERRQ(ierr); 652 if (isBAIJ || m == 0) { 653 c->brows = m; 654 c->bcols = 1; 655 } else { /* mpiaij matrix */ 656 /* bcols is chosen s.t. dy-array takes 50% of local memory space as mat */ 657 Mat_MPIAIJ *aij=(Mat_MPIAIJ*)mat->data; 658 Mat_SeqAIJ *spA,*spB; 659 Mat A,B; 660 PetscInt nz,brows,bcols; 661 PetscReal mem; 662 663 bs = 1; /* only bs=1 is supported for MPIAIJ matrix */ 664 665 A = aij->A; spA = (Mat_SeqAIJ*)A->data; 666 B = aij->B; spB = (Mat_SeqAIJ*)B->data; 667 nz = spA->nz + spB->nz; /* total local nonzero entries of mat */ 668 mem = nz*(sizeof(PetscScalar) + sizeof(PetscInt)) + 3*m*sizeof(PetscInt); 669 bcols = (PetscInt)(0.5*mem /(m*sizeof(PetscScalar))); 670 brows = 1000/bcols; 671 if (bcols > nis) bcols = nis; 672 if (brows == 0 || brows > m) brows = m; 673 c->brows = brows; 674 c->bcols = bcols; 675 } 676 677 c->M = mat->rmap->N/bs; /* set the global rows and columns and local rows */ 678 c->N = mat->cmap->N/bs; 679 c->m = mat->rmap->n/bs; 680 c->rstart = mat->rmap->rstart/bs; 681 c->ncolors = nis; 682 PetscFunctionReturn(0); 683 } 684