1 2 #include <../src/mat/impls/baij/mpi/mpibaij.h> /*I "petscmat.h" I*/ 3 4 #include <petscblaslapack.h> 5 #include <petscsf.h> 6 7 #undef __FUNCT__ 8 #define __FUNCT__ "MatGetRowMaxAbs_MPIBAIJ" 9 PetscErrorCode MatGetRowMaxAbs_MPIBAIJ(Mat A,Vec v,PetscInt idx[]) 10 { 11 Mat_MPIBAIJ *a = (Mat_MPIBAIJ*)A->data; 12 PetscErrorCode ierr; 13 PetscInt i,*idxb = 0; 14 PetscScalar *va,*vb; 15 Vec vtmp; 16 17 PetscFunctionBegin; 18 ierr = MatGetRowMaxAbs(a->A,v,idx);CHKERRQ(ierr); 19 ierr = VecGetArray(v,&va);CHKERRQ(ierr); 20 if (idx) { 21 for (i=0; i<A->rmap->n; i++) { 22 if (PetscAbsScalar(va[i])) idx[i] += A->cmap->rstart; 23 } 24 } 25 26 ierr = VecCreateSeq(PETSC_COMM_SELF,A->rmap->n,&vtmp);CHKERRQ(ierr); 27 if (idx) {ierr = PetscMalloc1(A->rmap->n,&idxb);CHKERRQ(ierr);} 28 ierr = MatGetRowMaxAbs(a->B,vtmp,idxb);CHKERRQ(ierr); 29 ierr = VecGetArray(vtmp,&vb);CHKERRQ(ierr); 30 31 for (i=0; i<A->rmap->n; i++) { 32 if (PetscAbsScalar(va[i]) < PetscAbsScalar(vb[i])) { 33 va[i] = vb[i]; 34 if (idx) idx[i] = A->cmap->bs*a->garray[idxb[i]/A->cmap->bs] + (idxb[i] % A->cmap->bs); 35 } 36 } 37 38 ierr = VecRestoreArray(v,&va);CHKERRQ(ierr); 39 ierr = VecRestoreArray(vtmp,&vb);CHKERRQ(ierr); 40 ierr = PetscFree(idxb);CHKERRQ(ierr); 41 ierr = VecDestroy(&vtmp);CHKERRQ(ierr); 42 PetscFunctionReturn(0); 43 } 44 45 #undef __FUNCT__ 46 #define __FUNCT__ "MatStoreValues_MPIBAIJ" 47 PetscErrorCode MatStoreValues_MPIBAIJ(Mat mat) 48 { 49 Mat_MPIBAIJ *aij = (Mat_MPIBAIJ*)mat->data; 50 PetscErrorCode ierr; 51 52 PetscFunctionBegin; 53 ierr = MatStoreValues(aij->A);CHKERRQ(ierr); 54 ierr = MatStoreValues(aij->B);CHKERRQ(ierr); 55 PetscFunctionReturn(0); 56 } 57 58 #undef __FUNCT__ 59 #define __FUNCT__ "MatRetrieveValues_MPIBAIJ" 60 PetscErrorCode MatRetrieveValues_MPIBAIJ(Mat mat) 61 { 62 Mat_MPIBAIJ *aij = (Mat_MPIBAIJ*)mat->data; 63 PetscErrorCode ierr; 64 65 PetscFunctionBegin; 66 ierr = MatRetrieveValues(aij->A);CHKERRQ(ierr); 67 ierr = MatRetrieveValues(aij->B);CHKERRQ(ierr); 68 PetscFunctionReturn(0); 69 } 70 71 /* 72 Local utility routine that creates a mapping from the global column 73 number to the local number in the off-diagonal part of the local 74 storage of the matrix. This is done in a non scalable way since the 75 length of colmap equals the global matrix length. 76 */ 77 #undef __FUNCT__ 78 #define __FUNCT__ "MatCreateColmap_MPIBAIJ_Private" 79 PetscErrorCode MatCreateColmap_MPIBAIJ_Private(Mat mat) 80 { 81 Mat_MPIBAIJ *baij = (Mat_MPIBAIJ*)mat->data; 82 Mat_SeqBAIJ *B = (Mat_SeqBAIJ*)baij->B->data; 83 PetscErrorCode ierr; 84 PetscInt nbs = B->nbs,i,bs=mat->rmap->bs; 85 86 PetscFunctionBegin; 87 #if defined(PETSC_USE_CTABLE) 88 ierr = PetscTableCreate(baij->nbs,baij->Nbs+1,&baij->colmap);CHKERRQ(ierr); 89 for (i=0; i<nbs; i++) { 90 ierr = PetscTableAdd(baij->colmap,baij->garray[i]+1,i*bs+1,INSERT_VALUES);CHKERRQ(ierr); 91 } 92 #else 93 ierr = PetscMalloc1(baij->Nbs+1,&baij->colmap);CHKERRQ(ierr); 94 ierr = PetscLogObjectMemory((PetscObject)mat,baij->Nbs*sizeof(PetscInt));CHKERRQ(ierr); 95 ierr = PetscMemzero(baij->colmap,baij->Nbs*sizeof(PetscInt));CHKERRQ(ierr); 96 for (i=0; i<nbs; i++) baij->colmap[baij->garray[i]] = i*bs+1; 97 #endif 98 PetscFunctionReturn(0); 99 } 100 101 #define MatSetValues_SeqBAIJ_A_Private(row,col,value,addv,orow,ocol) \ 102 { \ 103 \ 104 brow = row/bs; \ 105 rp = aj + ai[brow]; ap = aa + bs2*ai[brow]; \ 106 rmax = aimax[brow]; nrow = ailen[brow]; \ 107 bcol = col/bs; \ 108 ridx = row % bs; cidx = col % bs; \ 109 low = 0; high = nrow; \ 110 while (high-low > 3) { \ 111 t = (low+high)/2; \ 112 if (rp[t] > bcol) high = t; \ 113 else low = t; \ 114 } \ 115 for (_i=low; _i<high; _i++) { \ 116 if (rp[_i] > bcol) break; \ 117 if (rp[_i] == bcol) { \ 118 bap = ap + bs2*_i + bs*cidx + ridx; \ 119 if (addv == ADD_VALUES) *bap += value; \ 120 else *bap = value; \ 121 goto a_noinsert; \ 122 } \ 123 } \ 124 if (a->nonew == 1) goto a_noinsert; \ 125 if (a->nonew == -1) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new nonzero at global row/column (%D, %D) into matrix", orow, ocol); \ 126 MatSeqXAIJReallocateAIJ(A,a->mbs,bs2,nrow,brow,bcol,rmax,aa,ai,aj,rp,ap,aimax,a->nonew,MatScalar); \ 127 N = nrow++ - 1; \ 128 /* shift up all the later entries in this row */ \ 129 for (ii=N; ii>=_i; ii--) { \ 130 rp[ii+1] = rp[ii]; \ 131 ierr = PetscMemcpy(ap+bs2*(ii+1),ap+bs2*(ii),bs2*sizeof(MatScalar));CHKERRQ(ierr); \ 132 } \ 133 if (N>=_i) { ierr = PetscMemzero(ap+bs2*_i,bs2*sizeof(MatScalar));CHKERRQ(ierr); } \ 134 rp[_i] = bcol; \ 135 ap[bs2*_i + bs*cidx + ridx] = value; \ 136 a_noinsert:; \ 137 ailen[brow] = nrow; \ 138 } 139 140 #define MatSetValues_SeqBAIJ_B_Private(row,col,value,addv,orow,ocol) \ 141 { \ 142 brow = row/bs; \ 143 rp = bj + bi[brow]; ap = ba + bs2*bi[brow]; \ 144 rmax = bimax[brow]; nrow = bilen[brow]; \ 145 bcol = col/bs; \ 146 ridx = row % bs; cidx = col % bs; \ 147 low = 0; high = nrow; \ 148 while (high-low > 3) { \ 149 t = (low+high)/2; \ 150 if (rp[t] > bcol) high = t; \ 151 else low = t; \ 152 } \ 153 for (_i=low; _i<high; _i++) { \ 154 if (rp[_i] > bcol) break; \ 155 if (rp[_i] == bcol) { \ 156 bap = ap + bs2*_i + bs*cidx + ridx; \ 157 if (addv == ADD_VALUES) *bap += value; \ 158 else *bap = value; \ 159 goto b_noinsert; \ 160 } \ 161 } \ 162 if (b->nonew == 1) goto b_noinsert; \ 163 if (b->nonew == -1) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new nonzero at global row/column (%D, %D) into matrix", orow, ocol); \ 164 MatSeqXAIJReallocateAIJ(B,b->mbs,bs2,nrow,brow,bcol,rmax,ba,bi,bj,rp,ap,bimax,b->nonew,MatScalar); \ 165 N = nrow++ - 1; \ 166 /* shift up all the later entries in this row */ \ 167 for (ii=N; ii>=_i; ii--) { \ 168 rp[ii+1] = rp[ii]; \ 169 ierr = PetscMemcpy(ap+bs2*(ii+1),ap+bs2*(ii),bs2*sizeof(MatScalar));CHKERRQ(ierr); \ 170 } \ 171 if (N>=_i) { ierr = PetscMemzero(ap+bs2*_i,bs2*sizeof(MatScalar));CHKERRQ(ierr);} \ 172 rp[_i] = bcol; \ 173 ap[bs2*_i + bs*cidx + ridx] = value; \ 174 b_noinsert:; \ 175 bilen[brow] = nrow; \ 176 } 177 178 #undef __FUNCT__ 179 #define __FUNCT__ "MatSetValues_MPIBAIJ" 180 PetscErrorCode MatSetValues_MPIBAIJ(Mat mat,PetscInt m,const PetscInt im[],PetscInt n,const PetscInt in[],const PetscScalar v[],InsertMode addv) 181 { 182 Mat_MPIBAIJ *baij = (Mat_MPIBAIJ*)mat->data; 183 MatScalar value; 184 PetscBool roworiented = baij->roworiented; 185 PetscErrorCode ierr; 186 PetscInt i,j,row,col; 187 PetscInt rstart_orig=mat->rmap->rstart; 188 PetscInt rend_orig =mat->rmap->rend,cstart_orig=mat->cmap->rstart; 189 PetscInt cend_orig =mat->cmap->rend,bs=mat->rmap->bs; 190 191 /* Some Variables required in the macro */ 192 Mat A = baij->A; 193 Mat_SeqBAIJ *a = (Mat_SeqBAIJ*)(A)->data; 194 PetscInt *aimax=a->imax,*ai=a->i,*ailen=a->ilen,*aj=a->j; 195 MatScalar *aa =a->a; 196 197 Mat B = baij->B; 198 Mat_SeqBAIJ *b = (Mat_SeqBAIJ*)(B)->data; 199 PetscInt *bimax=b->imax,*bi=b->i,*bilen=b->ilen,*bj=b->j; 200 MatScalar *ba =b->a; 201 202 PetscInt *rp,ii,nrow,_i,rmax,N,brow,bcol; 203 PetscInt low,high,t,ridx,cidx,bs2=a->bs2; 204 MatScalar *ap,*bap; 205 206 PetscFunctionBegin; 207 for (i=0; i<m; i++) { 208 if (im[i] < 0) continue; 209 #if defined(PETSC_USE_DEBUG) 210 if (im[i] >= mat->rmap->N) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Row too large: row %D max %D",im[i],mat->rmap->N-1); 211 #endif 212 if (im[i] >= rstart_orig && im[i] < rend_orig) { 213 row = im[i] - rstart_orig; 214 for (j=0; j<n; j++) { 215 if (in[j] >= cstart_orig && in[j] < cend_orig) { 216 col = in[j] - cstart_orig; 217 if (roworiented) value = v[i*n+j]; 218 else value = v[i+j*m]; 219 MatSetValues_SeqBAIJ_A_Private(row,col,value,addv,im[i],in[j]); 220 /* ierr = MatSetValues_SeqBAIJ(baij->A,1,&row,1,&col,&value,addv);CHKERRQ(ierr); */ 221 } else if (in[j] < 0) continue; 222 #if defined(PETSC_USE_DEBUG) 223 else if (in[j] >= mat->cmap->N) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Column too large: col %D max %D",in[j],mat->cmap->N-1); 224 #endif 225 else { 226 if (mat->was_assembled) { 227 if (!baij->colmap) { 228 ierr = MatCreateColmap_MPIBAIJ_Private(mat);CHKERRQ(ierr); 229 } 230 #if defined(PETSC_USE_CTABLE) 231 ierr = PetscTableFind(baij->colmap,in[j]/bs + 1,&col);CHKERRQ(ierr); 232 col = col - 1; 233 #else 234 col = baij->colmap[in[j]/bs] - 1; 235 #endif 236 if (col < 0 && !((Mat_SeqBAIJ*)(baij->B->data))->nonew) { 237 ierr = MatDisAssemble_MPIBAIJ(mat);CHKERRQ(ierr); 238 col = in[j]; 239 /* Reinitialize the variables required by MatSetValues_SeqBAIJ_B_Private() */ 240 B = baij->B; 241 b = (Mat_SeqBAIJ*)(B)->data; 242 bimax=b->imax;bi=b->i;bilen=b->ilen;bj=b->j; 243 ba =b->a; 244 } else if (col < 0) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new nonzero (%D, %D) into matrix", im[i], in[j]); 245 else col += in[j]%bs; 246 } else col = in[j]; 247 if (roworiented) value = v[i*n+j]; 248 else value = v[i+j*m]; 249 MatSetValues_SeqBAIJ_B_Private(row,col,value,addv,im[i],in[j]); 250 /* ierr = MatSetValues_SeqBAIJ(baij->B,1,&row,1,&col,&value,addv);CHKERRQ(ierr); */ 251 } 252 } 253 } else { 254 if (mat->nooffprocentries) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Setting off process row %D even though MatSetOption(,MAT_NO_OFF_PROC_ENTRIES,PETSC_TRUE) was set",im[i]); 255 if (!baij->donotstash) { 256 mat->assembled = PETSC_FALSE; 257 if (roworiented) { 258 ierr = MatStashValuesRow_Private(&mat->stash,im[i],n,in,v+i*n,PETSC_FALSE);CHKERRQ(ierr); 259 } else { 260 ierr = MatStashValuesCol_Private(&mat->stash,im[i],n,in,v+i,m,PETSC_FALSE);CHKERRQ(ierr); 261 } 262 } 263 } 264 } 265 PetscFunctionReturn(0); 266 } 267 268 #undef __FUNCT__ 269 #define __FUNCT__ "MatSetValuesBlocked_SeqBAIJ_Inlined" 270 PETSC_STATIC_INLINE PetscErrorCode MatSetValuesBlocked_SeqBAIJ_Inlined(Mat A,PetscInt row,PetscInt col,const PetscScalar v[],InsertMode is,PetscInt orow,PetscInt ocol) 271 { 272 Mat_SeqBAIJ *a = (Mat_SeqBAIJ*)A->data; 273 PetscInt *rp,low,high,t,ii,jj,nrow,i,rmax,N; 274 PetscInt *imax=a->imax,*ai=a->i,*ailen=a->ilen; 275 PetscErrorCode ierr; 276 PetscInt *aj =a->j,nonew=a->nonew,bs2=a->bs2,bs=A->rmap->bs; 277 PetscBool roworiented=a->roworiented; 278 const PetscScalar *value = v; 279 MatScalar *ap,*aa = a->a,*bap; 280 281 PetscFunctionBegin; 282 rp = aj + ai[row]; 283 ap = aa + bs2*ai[row]; 284 rmax = imax[row]; 285 nrow = ailen[row]; 286 low = 0; 287 high = nrow; 288 value = v; 289 low = 0; 290 high = nrow; 291 while (high-low > 7) { 292 t = (low+high)/2; 293 if (rp[t] > col) high = t; 294 else low = t; 295 } 296 for (i=low; i<high; i++) { 297 if (rp[i] > col) break; 298 if (rp[i] == col) { 299 bap = ap + bs2*i; 300 if (roworiented) { 301 if (is == ADD_VALUES) { 302 for (ii=0; ii<bs; ii++) { 303 for (jj=ii; jj<bs2; jj+=bs) { 304 bap[jj] += *value++; 305 } 306 } 307 } else { 308 for (ii=0; ii<bs; ii++) { 309 for (jj=ii; jj<bs2; jj+=bs) { 310 bap[jj] = *value++; 311 } 312 } 313 } 314 } else { 315 if (is == ADD_VALUES) { 316 for (ii=0; ii<bs; ii++,value+=bs) { 317 for (jj=0; jj<bs; jj++) { 318 bap[jj] += value[jj]; 319 } 320 bap += bs; 321 } 322 } else { 323 for (ii=0; ii<bs; ii++,value+=bs) { 324 for (jj=0; jj<bs; jj++) { 325 bap[jj] = value[jj]; 326 } 327 bap += bs; 328 } 329 } 330 } 331 goto noinsert2; 332 } 333 } 334 if (nonew == 1) goto noinsert2; 335 if (nonew == -1) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new global block indexed nonzero block (%D, %D) in the matrix", orow, ocol); 336 MatSeqXAIJReallocateAIJ(A,a->mbs,bs2,nrow,row,col,rmax,aa,ai,aj,rp,ap,imax,nonew,MatScalar); 337 N = nrow++ - 1; high++; 338 /* shift up all the later entries in this row */ 339 for (ii=N; ii>=i; ii--) { 340 rp[ii+1] = rp[ii]; 341 ierr = PetscMemcpy(ap+bs2*(ii+1),ap+bs2*(ii),bs2*sizeof(MatScalar));CHKERRQ(ierr); 342 } 343 if (N >= i) { 344 ierr = PetscMemzero(ap+bs2*i,bs2*sizeof(MatScalar));CHKERRQ(ierr); 345 } 346 rp[i] = col; 347 bap = ap + bs2*i; 348 if (roworiented) { 349 for (ii=0; ii<bs; ii++) { 350 for (jj=ii; jj<bs2; jj+=bs) { 351 bap[jj] = *value++; 352 } 353 } 354 } else { 355 for (ii=0; ii<bs; ii++) { 356 for (jj=0; jj<bs; jj++) { 357 *bap++ = *value++; 358 } 359 } 360 } 361 noinsert2:; 362 ailen[row] = nrow; 363 PetscFunctionReturn(0); 364 } 365 366 #undef __FUNCT__ 367 #define __FUNCT__ "MatSetValuesBlocked_MPIBAIJ" 368 /* 369 This routine should be optimized so that the block copy at ** Here a copy is required ** below is not needed 370 by passing additional stride information into the MatSetValuesBlocked_SeqBAIJ_Inlined() routine 371 */ 372 PetscErrorCode MatSetValuesBlocked_MPIBAIJ(Mat mat,PetscInt m,const PetscInt im[],PetscInt n,const PetscInt in[],const PetscScalar v[],InsertMode addv) 373 { 374 Mat_MPIBAIJ *baij = (Mat_MPIBAIJ*)mat->data; 375 const PetscScalar *value; 376 MatScalar *barray = baij->barray; 377 PetscBool roworiented = baij->roworiented; 378 PetscErrorCode ierr; 379 PetscInt i,j,ii,jj,row,col,rstart=baij->rstartbs; 380 PetscInt rend=baij->rendbs,cstart=baij->cstartbs,stepval; 381 PetscInt cend=baij->cendbs,bs=mat->rmap->bs,bs2=baij->bs2; 382 383 PetscFunctionBegin; 384 if (!barray) { 385 ierr = PetscMalloc1(bs2,&barray);CHKERRQ(ierr); 386 baij->barray = barray; 387 } 388 389 if (roworiented) stepval = (n-1)*bs; 390 else stepval = (m-1)*bs; 391 392 for (i=0; i<m; i++) { 393 if (im[i] < 0) continue; 394 #if defined(PETSC_USE_DEBUG) 395 if (im[i] >= baij->Mbs) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Block indexed row too large %D max %D",im[i],baij->Mbs-1); 396 #endif 397 if (im[i] >= rstart && im[i] < rend) { 398 row = im[i] - rstart; 399 for (j=0; j<n; j++) { 400 /* If NumCol = 1 then a copy is not required */ 401 if ((roworiented) && (n == 1)) { 402 barray = (MatScalar*)v + i*bs2; 403 } else if ((!roworiented) && (m == 1)) { 404 barray = (MatScalar*)v + j*bs2; 405 } else { /* Here a copy is required */ 406 if (roworiented) { 407 value = v + (i*(stepval+bs) + j)*bs; 408 } else { 409 value = v + (j*(stepval+bs) + i)*bs; 410 } 411 for (ii=0; ii<bs; ii++,value+=bs+stepval) { 412 for (jj=0; jj<bs; jj++) barray[jj] = value[jj]; 413 barray += bs; 414 } 415 barray -= bs2; 416 } 417 418 if (in[j] >= cstart && in[j] < cend) { 419 col = in[j] - cstart; 420 ierr = MatSetValuesBlocked_SeqBAIJ_Inlined(baij->A,row,col,barray,addv,im[i],in[j]);CHKERRQ(ierr); 421 } else if (in[j] < 0) continue; 422 #if defined(PETSC_USE_DEBUG) 423 else if (in[j] >= baij->Nbs) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Block indexed column too large %D max %D",in[j],baij->Nbs-1); 424 #endif 425 else { 426 if (mat->was_assembled) { 427 if (!baij->colmap) { 428 ierr = MatCreateColmap_MPIBAIJ_Private(mat);CHKERRQ(ierr); 429 } 430 431 #if defined(PETSC_USE_DEBUG) 432 #if defined(PETSC_USE_CTABLE) 433 { PetscInt data; 434 ierr = PetscTableFind(baij->colmap,in[j]+1,&data);CHKERRQ(ierr); 435 if ((data - 1) % bs) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Incorrect colmap"); 436 } 437 #else 438 if ((baij->colmap[in[j]] - 1) % bs) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Incorrect colmap"); 439 #endif 440 #endif 441 #if defined(PETSC_USE_CTABLE) 442 ierr = PetscTableFind(baij->colmap,in[j]+1,&col);CHKERRQ(ierr); 443 col = (col - 1)/bs; 444 #else 445 col = (baij->colmap[in[j]] - 1)/bs; 446 #endif 447 if (col < 0 && !((Mat_SeqBAIJ*)(baij->B->data))->nonew) { 448 ierr = MatDisAssemble_MPIBAIJ(mat);CHKERRQ(ierr); 449 col = in[j]; 450 } else if (col < 0) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new blocked indexed nonzero block (%D, %D) into matrix",im[i],in[j]); 451 } else col = in[j]; 452 ierr = MatSetValuesBlocked_SeqBAIJ_Inlined(baij->B,row,col,barray,addv,im[i],in[j]);CHKERRQ(ierr); 453 } 454 } 455 } else { 456 if (mat->nooffprocentries) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Setting off process block indexed row %D even though MatSetOption(,MAT_NO_OFF_PROC_ENTRIES,PETSC_TRUE) was set",im[i]); 457 if (!baij->donotstash) { 458 if (roworiented) { 459 ierr = MatStashValuesRowBlocked_Private(&mat->bstash,im[i],n,in,v,m,n,i);CHKERRQ(ierr); 460 } else { 461 ierr = MatStashValuesColBlocked_Private(&mat->bstash,im[i],n,in,v,m,n,i);CHKERRQ(ierr); 462 } 463 } 464 } 465 } 466 PetscFunctionReturn(0); 467 } 468 469 #define HASH_KEY 0.6180339887 470 #define HASH(size,key,tmp) (tmp = (key)*HASH_KEY,(PetscInt)((size)*(tmp-(PetscInt)tmp))) 471 /* #define HASH(size,key) ((PetscInt)((size)*fmod(((key)*HASH_KEY),1))) */ 472 /* #define HASH(size,key,tmp) ((PetscInt)((size)*fmod(((key)*HASH_KEY),1))) */ 473 #undef __FUNCT__ 474 #define __FUNCT__ "MatSetValues_MPIBAIJ_HT" 475 PetscErrorCode MatSetValues_MPIBAIJ_HT(Mat mat,PetscInt m,const PetscInt im[],PetscInt n,const PetscInt in[],const PetscScalar v[],InsertMode addv) 476 { 477 Mat_MPIBAIJ *baij = (Mat_MPIBAIJ*)mat->data; 478 PetscBool roworiented = baij->roworiented; 479 PetscErrorCode ierr; 480 PetscInt i,j,row,col; 481 PetscInt rstart_orig=mat->rmap->rstart; 482 PetscInt rend_orig =mat->rmap->rend,Nbs=baij->Nbs; 483 PetscInt h1,key,size=baij->ht_size,bs=mat->rmap->bs,*HT=baij->ht,idx; 484 PetscReal tmp; 485 MatScalar **HD = baij->hd,value; 486 #if defined(PETSC_USE_DEBUG) 487 PetscInt total_ct=baij->ht_total_ct,insert_ct=baij->ht_insert_ct; 488 #endif 489 490 PetscFunctionBegin; 491 for (i=0; i<m; i++) { 492 #if defined(PETSC_USE_DEBUG) 493 if (im[i] < 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Negative row"); 494 if (im[i] >= mat->rmap->N) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Row too large: row %D max %D",im[i],mat->rmap->N-1); 495 #endif 496 row = im[i]; 497 if (row >= rstart_orig && row < rend_orig) { 498 for (j=0; j<n; j++) { 499 col = in[j]; 500 if (roworiented) value = v[i*n+j]; 501 else value = v[i+j*m]; 502 /* Look up PetscInto the Hash Table */ 503 key = (row/bs)*Nbs+(col/bs)+1; 504 h1 = HASH(size,key,tmp); 505 506 507 idx = h1; 508 #if defined(PETSC_USE_DEBUG) 509 insert_ct++; 510 total_ct++; 511 if (HT[idx] != key) { 512 for (idx=h1; (idx<size) && (HT[idx]!=key); idx++,total_ct++) ; 513 if (idx == size) { 514 for (idx=0; (idx<h1) && (HT[idx]!=key); idx++,total_ct++) ; 515 if (idx == h1) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"(%D,%D) has no entry in the hash table", row, col); 516 } 517 } 518 #else 519 if (HT[idx] != key) { 520 for (idx=h1; (idx<size) && (HT[idx]!=key); idx++) ; 521 if (idx == size) { 522 for (idx=0; (idx<h1) && (HT[idx]!=key); idx++) ; 523 if (idx == h1) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"(%D,%D) has no entry in the hash table", row, col); 524 } 525 } 526 #endif 527 /* A HASH table entry is found, so insert the values at the correct address */ 528 if (addv == ADD_VALUES) *(HD[idx]+ (col % bs)*bs + (row % bs)) += value; 529 else *(HD[idx]+ (col % bs)*bs + (row % bs)) = value; 530 } 531 } else if (!baij->donotstash) { 532 if (roworiented) { 533 ierr = MatStashValuesRow_Private(&mat->stash,im[i],n,in,v+i*n,PETSC_FALSE);CHKERRQ(ierr); 534 } else { 535 ierr = MatStashValuesCol_Private(&mat->stash,im[i],n,in,v+i,m,PETSC_FALSE);CHKERRQ(ierr); 536 } 537 } 538 } 539 #if defined(PETSC_USE_DEBUG) 540 baij->ht_total_ct = total_ct; 541 baij->ht_insert_ct = insert_ct; 542 #endif 543 PetscFunctionReturn(0); 544 } 545 546 #undef __FUNCT__ 547 #define __FUNCT__ "MatSetValuesBlocked_MPIBAIJ_HT" 548 PetscErrorCode MatSetValuesBlocked_MPIBAIJ_HT(Mat mat,PetscInt m,const PetscInt im[],PetscInt n,const PetscInt in[],const PetscScalar v[],InsertMode addv) 549 { 550 Mat_MPIBAIJ *baij = (Mat_MPIBAIJ*)mat->data; 551 PetscBool roworiented = baij->roworiented; 552 PetscErrorCode ierr; 553 PetscInt i,j,ii,jj,row,col; 554 PetscInt rstart=baij->rstartbs; 555 PetscInt rend =mat->rmap->rend,stepval,bs=mat->rmap->bs,bs2=baij->bs2,nbs2=n*bs2; 556 PetscInt h1,key,size=baij->ht_size,idx,*HT=baij->ht,Nbs=baij->Nbs; 557 PetscReal tmp; 558 MatScalar **HD = baij->hd,*baij_a; 559 const PetscScalar *v_t,*value; 560 #if defined(PETSC_USE_DEBUG) 561 PetscInt total_ct=baij->ht_total_ct,insert_ct=baij->ht_insert_ct; 562 #endif 563 564 PetscFunctionBegin; 565 if (roworiented) stepval = (n-1)*bs; 566 else stepval = (m-1)*bs; 567 568 for (i=0; i<m; i++) { 569 #if defined(PETSC_USE_DEBUG) 570 if (im[i] < 0) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Negative row: %D",im[i]); 571 if (im[i] >= baij->Mbs) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Row too large: row %D max %D",im[i],baij->Mbs-1); 572 #endif 573 row = im[i]; 574 v_t = v + i*nbs2; 575 if (row >= rstart && row < rend) { 576 for (j=0; j<n; j++) { 577 col = in[j]; 578 579 /* Look up into the Hash Table */ 580 key = row*Nbs+col+1; 581 h1 = HASH(size,key,tmp); 582 583 idx = h1; 584 #if defined(PETSC_USE_DEBUG) 585 total_ct++; 586 insert_ct++; 587 if (HT[idx] != key) { 588 for (idx=h1; (idx<size) && (HT[idx]!=key); idx++,total_ct++) ; 589 if (idx == size) { 590 for (idx=0; (idx<h1) && (HT[idx]!=key); idx++,total_ct++) ; 591 if (idx == h1) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"(%D,%D) has no entry in the hash table", row, col); 592 } 593 } 594 #else 595 if (HT[idx] != key) { 596 for (idx=h1; (idx<size) && (HT[idx]!=key); idx++) ; 597 if (idx == size) { 598 for (idx=0; (idx<h1) && (HT[idx]!=key); idx++) ; 599 if (idx == h1) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"(%D,%D) has no entry in the hash table", row, col); 600 } 601 } 602 #endif 603 baij_a = HD[idx]; 604 if (roworiented) { 605 /*value = v + i*(stepval+bs)*bs + j*bs;*/ 606 /* value = v + (i*(stepval+bs)+j)*bs; */ 607 value = v_t; 608 v_t += bs; 609 if (addv == ADD_VALUES) { 610 for (ii=0; ii<bs; ii++,value+=stepval) { 611 for (jj=ii; jj<bs2; jj+=bs) { 612 baij_a[jj] += *value++; 613 } 614 } 615 } else { 616 for (ii=0; ii<bs; ii++,value+=stepval) { 617 for (jj=ii; jj<bs2; jj+=bs) { 618 baij_a[jj] = *value++; 619 } 620 } 621 } 622 } else { 623 value = v + j*(stepval+bs)*bs + i*bs; 624 if (addv == ADD_VALUES) { 625 for (ii=0; ii<bs; ii++,value+=stepval,baij_a+=bs) { 626 for (jj=0; jj<bs; jj++) { 627 baij_a[jj] += *value++; 628 } 629 } 630 } else { 631 for (ii=0; ii<bs; ii++,value+=stepval,baij_a+=bs) { 632 for (jj=0; jj<bs; jj++) { 633 baij_a[jj] = *value++; 634 } 635 } 636 } 637 } 638 } 639 } else { 640 if (!baij->donotstash) { 641 if (roworiented) { 642 ierr = MatStashValuesRowBlocked_Private(&mat->bstash,im[i],n,in,v,m,n,i);CHKERRQ(ierr); 643 } else { 644 ierr = MatStashValuesColBlocked_Private(&mat->bstash,im[i],n,in,v,m,n,i);CHKERRQ(ierr); 645 } 646 } 647 } 648 } 649 #if defined(PETSC_USE_DEBUG) 650 baij->ht_total_ct = total_ct; 651 baij->ht_insert_ct = insert_ct; 652 #endif 653 PetscFunctionReturn(0); 654 } 655 656 #undef __FUNCT__ 657 #define __FUNCT__ "MatGetValues_MPIBAIJ" 658 PetscErrorCode MatGetValues_MPIBAIJ(Mat mat,PetscInt m,const PetscInt idxm[],PetscInt n,const PetscInt idxn[],PetscScalar v[]) 659 { 660 Mat_MPIBAIJ *baij = (Mat_MPIBAIJ*)mat->data; 661 PetscErrorCode ierr; 662 PetscInt bs = mat->rmap->bs,i,j,bsrstart = mat->rmap->rstart,bsrend = mat->rmap->rend; 663 PetscInt bscstart = mat->cmap->rstart,bscend = mat->cmap->rend,row,col,data; 664 665 PetscFunctionBegin; 666 for (i=0; i<m; i++) { 667 if (idxm[i] < 0) continue; /* SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Negative row: %D",idxm[i]);*/ 668 if (idxm[i] >= mat->rmap->N) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Row too large: row %D max %D",idxm[i],mat->rmap->N-1); 669 if (idxm[i] >= bsrstart && idxm[i] < bsrend) { 670 row = idxm[i] - bsrstart; 671 for (j=0; j<n; j++) { 672 if (idxn[j] < 0) continue; /* SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Negative column: %D",idxn[j]); */ 673 if (idxn[j] >= mat->cmap->N) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Column too large: col %D max %D",idxn[j],mat->cmap->N-1); 674 if (idxn[j] >= bscstart && idxn[j] < bscend) { 675 col = idxn[j] - bscstart; 676 ierr = MatGetValues_SeqBAIJ(baij->A,1,&row,1,&col,v+i*n+j);CHKERRQ(ierr); 677 } else { 678 if (!baij->colmap) { 679 ierr = MatCreateColmap_MPIBAIJ_Private(mat);CHKERRQ(ierr); 680 } 681 #if defined(PETSC_USE_CTABLE) 682 ierr = PetscTableFind(baij->colmap,idxn[j]/bs+1,&data);CHKERRQ(ierr); 683 data--; 684 #else 685 data = baij->colmap[idxn[j]/bs]-1; 686 #endif 687 if ((data < 0) || (baij->garray[data/bs] != idxn[j]/bs)) *(v+i*n+j) = 0.0; 688 else { 689 col = data + idxn[j]%bs; 690 ierr = MatGetValues_SeqBAIJ(baij->B,1,&row,1,&col,v+i*n+j);CHKERRQ(ierr); 691 } 692 } 693 } 694 } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Only local values currently supported"); 695 } 696 PetscFunctionReturn(0); 697 } 698 699 #undef __FUNCT__ 700 #define __FUNCT__ "MatNorm_MPIBAIJ" 701 PetscErrorCode MatNorm_MPIBAIJ(Mat mat,NormType type,PetscReal *nrm) 702 { 703 Mat_MPIBAIJ *baij = (Mat_MPIBAIJ*)mat->data; 704 Mat_SeqBAIJ *amat = (Mat_SeqBAIJ*)baij->A->data,*bmat = (Mat_SeqBAIJ*)baij->B->data; 705 PetscErrorCode ierr; 706 PetscInt i,j,bs2=baij->bs2,bs=baij->A->rmap->bs,nz,row,col; 707 PetscReal sum = 0.0; 708 MatScalar *v; 709 710 PetscFunctionBegin; 711 if (baij->size == 1) { 712 ierr = MatNorm(baij->A,type,nrm);CHKERRQ(ierr); 713 } else { 714 if (type == NORM_FROBENIUS) { 715 v = amat->a; 716 nz = amat->nz*bs2; 717 for (i=0; i<nz; i++) { 718 sum += PetscRealPart(PetscConj(*v)*(*v)); v++; 719 } 720 v = bmat->a; 721 nz = bmat->nz*bs2; 722 for (i=0; i<nz; i++) { 723 sum += PetscRealPart(PetscConj(*v)*(*v)); v++; 724 } 725 ierr = MPI_Allreduce(&sum,nrm,1,MPIU_REAL,MPIU_SUM,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 726 *nrm = PetscSqrtReal(*nrm); 727 } else if (type == NORM_1) { /* max column sum */ 728 PetscReal *tmp,*tmp2; 729 PetscInt *jj,*garray=baij->garray,cstart=baij->rstartbs; 730 ierr = PetscMalloc2(mat->cmap->N,&tmp,mat->cmap->N,&tmp2);CHKERRQ(ierr); 731 ierr = PetscMemzero(tmp,mat->cmap->N*sizeof(PetscReal));CHKERRQ(ierr); 732 v = amat->a; jj = amat->j; 733 for (i=0; i<amat->nz; i++) { 734 for (j=0; j<bs; j++) { 735 col = bs*(cstart + *jj) + j; /* column index */ 736 for (row=0; row<bs; row++) { 737 tmp[col] += PetscAbsScalar(*v); v++; 738 } 739 } 740 jj++; 741 } 742 v = bmat->a; jj = bmat->j; 743 for (i=0; i<bmat->nz; i++) { 744 for (j=0; j<bs; j++) { 745 col = bs*garray[*jj] + j; 746 for (row=0; row<bs; row++) { 747 tmp[col] += PetscAbsScalar(*v); v++; 748 } 749 } 750 jj++; 751 } 752 ierr = MPI_Allreduce(tmp,tmp2,mat->cmap->N,MPIU_REAL,MPIU_SUM,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 753 *nrm = 0.0; 754 for (j=0; j<mat->cmap->N; j++) { 755 if (tmp2[j] > *nrm) *nrm = tmp2[j]; 756 } 757 ierr = PetscFree2(tmp,tmp2);CHKERRQ(ierr); 758 } else if (type == NORM_INFINITY) { /* max row sum */ 759 PetscReal *sums; 760 ierr = PetscMalloc1(bs,&sums);CHKERRQ(ierr); 761 sum = 0.0; 762 for (j=0; j<amat->mbs; j++) { 763 for (row=0; row<bs; row++) sums[row] = 0.0; 764 v = amat->a + bs2*amat->i[j]; 765 nz = amat->i[j+1]-amat->i[j]; 766 for (i=0; i<nz; i++) { 767 for (col=0; col<bs; col++) { 768 for (row=0; row<bs; row++) { 769 sums[row] += PetscAbsScalar(*v); v++; 770 } 771 } 772 } 773 v = bmat->a + bs2*bmat->i[j]; 774 nz = bmat->i[j+1]-bmat->i[j]; 775 for (i=0; i<nz; i++) { 776 for (col=0; col<bs; col++) { 777 for (row=0; row<bs; row++) { 778 sums[row] += PetscAbsScalar(*v); v++; 779 } 780 } 781 } 782 for (row=0; row<bs; row++) { 783 if (sums[row] > sum) sum = sums[row]; 784 } 785 } 786 ierr = MPI_Allreduce(&sum,nrm,1,MPIU_REAL,MPIU_MAX,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 787 ierr = PetscFree(sums);CHKERRQ(ierr); 788 } else SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"No support for this norm yet"); 789 } 790 PetscFunctionReturn(0); 791 } 792 793 /* 794 Creates the hash table, and sets the table 795 This table is created only once. 796 If new entried need to be added to the matrix 797 then the hash table has to be destroyed and 798 recreated. 799 */ 800 #undef __FUNCT__ 801 #define __FUNCT__ "MatCreateHashTable_MPIBAIJ_Private" 802 PetscErrorCode MatCreateHashTable_MPIBAIJ_Private(Mat mat,PetscReal factor) 803 { 804 Mat_MPIBAIJ *baij = (Mat_MPIBAIJ*)mat->data; 805 Mat A = baij->A,B=baij->B; 806 Mat_SeqBAIJ *a = (Mat_SeqBAIJ*)A->data,*b=(Mat_SeqBAIJ*)B->data; 807 PetscInt i,j,k,nz=a->nz+b->nz,h1,*ai=a->i,*aj=a->j,*bi=b->i,*bj=b->j; 808 PetscErrorCode ierr; 809 PetscInt ht_size,bs2=baij->bs2,rstart=baij->rstartbs; 810 PetscInt cstart=baij->cstartbs,*garray=baij->garray,row,col,Nbs=baij->Nbs; 811 PetscInt *HT,key; 812 MatScalar **HD; 813 PetscReal tmp; 814 #if defined(PETSC_USE_INFO) 815 PetscInt ct=0,max=0; 816 #endif 817 818 PetscFunctionBegin; 819 if (baij->ht) PetscFunctionReturn(0); 820 821 baij->ht_size = (PetscInt)(factor*nz); 822 ht_size = baij->ht_size; 823 824 /* Allocate Memory for Hash Table */ 825 ierr = PetscCalloc2(ht_size,&baij->hd,ht_size,&baij->ht);CHKERRQ(ierr); 826 HD = baij->hd; 827 HT = baij->ht; 828 829 /* Loop Over A */ 830 for (i=0; i<a->mbs; i++) { 831 for (j=ai[i]; j<ai[i+1]; j++) { 832 row = i+rstart; 833 col = aj[j]+cstart; 834 835 key = row*Nbs + col + 1; 836 h1 = HASH(ht_size,key,tmp); 837 for (k=0; k<ht_size; k++) { 838 if (!HT[(h1+k)%ht_size]) { 839 HT[(h1+k)%ht_size] = key; 840 HD[(h1+k)%ht_size] = a->a + j*bs2; 841 break; 842 #if defined(PETSC_USE_INFO) 843 } else { 844 ct++; 845 #endif 846 } 847 } 848 #if defined(PETSC_USE_INFO) 849 if (k> max) max = k; 850 #endif 851 } 852 } 853 /* Loop Over B */ 854 for (i=0; i<b->mbs; i++) { 855 for (j=bi[i]; j<bi[i+1]; j++) { 856 row = i+rstart; 857 col = garray[bj[j]]; 858 key = row*Nbs + col + 1; 859 h1 = HASH(ht_size,key,tmp); 860 for (k=0; k<ht_size; k++) { 861 if (!HT[(h1+k)%ht_size]) { 862 HT[(h1+k)%ht_size] = key; 863 HD[(h1+k)%ht_size] = b->a + j*bs2; 864 break; 865 #if defined(PETSC_USE_INFO) 866 } else { 867 ct++; 868 #endif 869 } 870 } 871 #if defined(PETSC_USE_INFO) 872 if (k> max) max = k; 873 #endif 874 } 875 } 876 877 /* Print Summary */ 878 #if defined(PETSC_USE_INFO) 879 for (i=0,j=0; i<ht_size; i++) { 880 if (HT[i]) j++; 881 } 882 ierr = PetscInfo2(mat,"Average Search = %5.2f,max search = %D\n",(!j)? 0.0:((PetscReal)(ct+j))/j,max);CHKERRQ(ierr); 883 #endif 884 PetscFunctionReturn(0); 885 } 886 887 #undef __FUNCT__ 888 #define __FUNCT__ "MatAssemblyBegin_MPIBAIJ" 889 PetscErrorCode MatAssemblyBegin_MPIBAIJ(Mat mat,MatAssemblyType mode) 890 { 891 Mat_MPIBAIJ *baij = (Mat_MPIBAIJ*)mat->data; 892 PetscErrorCode ierr; 893 PetscInt nstash,reallocs; 894 895 PetscFunctionBegin; 896 if (baij->donotstash || mat->nooffprocentries) PetscFunctionReturn(0); 897 898 ierr = MatStashScatterBegin_Private(mat,&mat->stash,mat->rmap->range);CHKERRQ(ierr); 899 ierr = MatStashScatterBegin_Private(mat,&mat->bstash,baij->rangebs);CHKERRQ(ierr); 900 ierr = MatStashGetInfo_Private(&mat->stash,&nstash,&reallocs);CHKERRQ(ierr); 901 ierr = PetscInfo2(mat,"Stash has %D entries,uses %D mallocs.\n",nstash,reallocs);CHKERRQ(ierr); 902 ierr = MatStashGetInfo_Private(&mat->bstash,&nstash,&reallocs);CHKERRQ(ierr); 903 ierr = PetscInfo2(mat,"Block-Stash has %D entries, uses %D mallocs.\n",nstash,reallocs);CHKERRQ(ierr); 904 PetscFunctionReturn(0); 905 } 906 907 #undef __FUNCT__ 908 #define __FUNCT__ "MatAssemblyEnd_MPIBAIJ" 909 PetscErrorCode MatAssemblyEnd_MPIBAIJ(Mat mat,MatAssemblyType mode) 910 { 911 Mat_MPIBAIJ *baij=(Mat_MPIBAIJ*)mat->data; 912 Mat_SeqBAIJ *a =(Mat_SeqBAIJ*)baij->A->data; 913 PetscErrorCode ierr; 914 PetscInt i,j,rstart,ncols,flg,bs2=baij->bs2; 915 PetscInt *row,*col; 916 PetscBool r1,r2,r3,other_disassembled; 917 MatScalar *val; 918 PetscMPIInt n; 919 920 PetscFunctionBegin; 921 /* do not use 'b=(Mat_SeqBAIJ*)baij->B->data' as B can be reset in disassembly */ 922 if (!baij->donotstash && !mat->nooffprocentries) { 923 while (1) { 924 ierr = MatStashScatterGetMesg_Private(&mat->stash,&n,&row,&col,&val,&flg);CHKERRQ(ierr); 925 if (!flg) break; 926 927 for (i=0; i<n;) { 928 /* Now identify the consecutive vals belonging to the same row */ 929 for (j=i,rstart=row[j]; j<n; j++) { 930 if (row[j] != rstart) break; 931 } 932 if (j < n) ncols = j-i; 933 else ncols = n-i; 934 /* Now assemble all these values with a single function call */ 935 ierr = MatSetValues_MPIBAIJ(mat,1,row+i,ncols,col+i,val+i,mat->insertmode);CHKERRQ(ierr); 936 i = j; 937 } 938 } 939 ierr = MatStashScatterEnd_Private(&mat->stash);CHKERRQ(ierr); 940 /* Now process the block-stash. Since the values are stashed column-oriented, 941 set the roworiented flag to column oriented, and after MatSetValues() 942 restore the original flags */ 943 r1 = baij->roworiented; 944 r2 = a->roworiented; 945 r3 = ((Mat_SeqBAIJ*)baij->B->data)->roworiented; 946 947 baij->roworiented = PETSC_FALSE; 948 a->roworiented = PETSC_FALSE; 949 950 (((Mat_SeqBAIJ*)baij->B->data))->roworiented = PETSC_FALSE; /* b->roworiented */ 951 while (1) { 952 ierr = MatStashScatterGetMesg_Private(&mat->bstash,&n,&row,&col,&val,&flg);CHKERRQ(ierr); 953 if (!flg) break; 954 955 for (i=0; i<n;) { 956 /* Now identify the consecutive vals belonging to the same row */ 957 for (j=i,rstart=row[j]; j<n; j++) { 958 if (row[j] != rstart) break; 959 } 960 if (j < n) ncols = j-i; 961 else ncols = n-i; 962 ierr = MatSetValuesBlocked_MPIBAIJ(mat,1,row+i,ncols,col+i,val+i*bs2,mat->insertmode);CHKERRQ(ierr); 963 i = j; 964 } 965 } 966 ierr = MatStashScatterEnd_Private(&mat->bstash);CHKERRQ(ierr); 967 968 baij->roworiented = r1; 969 a->roworiented = r2; 970 971 ((Mat_SeqBAIJ*)baij->B->data)->roworiented = r3; /* b->roworiented */ 972 } 973 974 ierr = MatAssemblyBegin(baij->A,mode);CHKERRQ(ierr); 975 ierr = MatAssemblyEnd(baij->A,mode);CHKERRQ(ierr); 976 977 /* determine if any processor has disassembled, if so we must 978 also disassemble ourselfs, in order that we may reassemble. */ 979 /* 980 if nonzero structure of submatrix B cannot change then we know that 981 no processor disassembled thus we can skip this stuff 982 */ 983 if (!((Mat_SeqBAIJ*)baij->B->data)->nonew) { 984 ierr = MPI_Allreduce(&mat->was_assembled,&other_disassembled,1,MPIU_BOOL,MPI_PROD,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 985 if (mat->was_assembled && !other_disassembled) { 986 ierr = MatDisAssemble_MPIBAIJ(mat);CHKERRQ(ierr); 987 } 988 } 989 990 if (!mat->was_assembled && mode == MAT_FINAL_ASSEMBLY) { 991 ierr = MatSetUpMultiply_MPIBAIJ(mat);CHKERRQ(ierr); 992 } 993 ierr = MatAssemblyBegin(baij->B,mode);CHKERRQ(ierr); 994 ierr = MatAssemblyEnd(baij->B,mode);CHKERRQ(ierr); 995 996 #if defined(PETSC_USE_INFO) 997 if (baij->ht && mode== MAT_FINAL_ASSEMBLY) { 998 ierr = PetscInfo1(mat,"Average Hash Table Search in MatSetValues = %5.2f\n",((PetscReal)baij->ht_total_ct)/baij->ht_insert_ct);CHKERRQ(ierr); 999 1000 baij->ht_total_ct = 0; 1001 baij->ht_insert_ct = 0; 1002 } 1003 #endif 1004 if (baij->ht_flag && !baij->ht && mode == MAT_FINAL_ASSEMBLY) { 1005 ierr = MatCreateHashTable_MPIBAIJ_Private(mat,baij->ht_fact);CHKERRQ(ierr); 1006 1007 mat->ops->setvalues = MatSetValues_MPIBAIJ_HT; 1008 mat->ops->setvaluesblocked = MatSetValuesBlocked_MPIBAIJ_HT; 1009 } 1010 1011 ierr = PetscFree2(baij->rowvalues,baij->rowindices);CHKERRQ(ierr); 1012 1013 baij->rowvalues = 0; 1014 1015 /* if no new nonzero locations are allowed in matrix then only set the matrix state the first time through */ 1016 if ((!mat->was_assembled && mode == MAT_FINAL_ASSEMBLY) || !((Mat_SeqBAIJ*)(baij->A->data))->nonew) { 1017 PetscObjectState state = baij->A->nonzerostate + baij->B->nonzerostate; 1018 ierr = MPI_Allreduce(&state,&mat->nonzerostate,1,MPIU_INT64,MPI_SUM,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1019 } 1020 PetscFunctionReturn(0); 1021 } 1022 1023 extern PetscErrorCode MatView_SeqBAIJ(Mat,PetscViewer); 1024 #include <petscdraw.h> 1025 #undef __FUNCT__ 1026 #define __FUNCT__ "MatView_MPIBAIJ_ASCIIorDraworSocket" 1027 static PetscErrorCode MatView_MPIBAIJ_ASCIIorDraworSocket(Mat mat,PetscViewer viewer) 1028 { 1029 Mat_MPIBAIJ *baij = (Mat_MPIBAIJ*)mat->data; 1030 PetscErrorCode ierr; 1031 PetscMPIInt rank = baij->rank; 1032 PetscInt bs = mat->rmap->bs; 1033 PetscBool iascii,isdraw; 1034 PetscViewer sviewer; 1035 PetscViewerFormat format; 1036 1037 PetscFunctionBegin; 1038 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERASCII,&iascii);CHKERRQ(ierr); 1039 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERDRAW,&isdraw);CHKERRQ(ierr); 1040 if (iascii) { 1041 ierr = PetscViewerGetFormat(viewer,&format);CHKERRQ(ierr); 1042 if (format == PETSC_VIEWER_ASCII_INFO_DETAIL) { 1043 MatInfo info; 1044 ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)mat),&rank);CHKERRQ(ierr); 1045 ierr = MatGetInfo(mat,MAT_LOCAL,&info);CHKERRQ(ierr); 1046 ierr = PetscViewerASCIIPushSynchronized(viewer);CHKERRQ(ierr); 1047 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"[%d] Local rows %D nz %D nz alloced %D bs %D mem %D\n", 1048 rank,mat->rmap->n,(PetscInt)info.nz_used,(PetscInt)info.nz_allocated,mat->rmap->bs,(PetscInt)info.memory);CHKERRQ(ierr); 1049 ierr = MatGetInfo(baij->A,MAT_LOCAL,&info);CHKERRQ(ierr); 1050 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"[%d] on-diagonal part: nz %D \n",rank,(PetscInt)info.nz_used);CHKERRQ(ierr); 1051 ierr = MatGetInfo(baij->B,MAT_LOCAL,&info);CHKERRQ(ierr); 1052 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"[%d] off-diagonal part: nz %D \n",rank,(PetscInt)info.nz_used);CHKERRQ(ierr); 1053 ierr = PetscViewerFlush(viewer);CHKERRQ(ierr); 1054 ierr = PetscViewerASCIIPopSynchronized(viewer);CHKERRQ(ierr); 1055 ierr = PetscViewerASCIIPrintf(viewer,"Information on VecScatter used in matrix-vector product: \n");CHKERRQ(ierr); 1056 ierr = VecScatterView(baij->Mvctx,viewer);CHKERRQ(ierr); 1057 PetscFunctionReturn(0); 1058 } else if (format == PETSC_VIEWER_ASCII_INFO) { 1059 ierr = PetscViewerASCIIPrintf(viewer," block size is %D\n",bs);CHKERRQ(ierr); 1060 PetscFunctionReturn(0); 1061 } else if (format == PETSC_VIEWER_ASCII_FACTOR_INFO) { 1062 PetscFunctionReturn(0); 1063 } 1064 } 1065 1066 if (isdraw) { 1067 PetscDraw draw; 1068 PetscBool isnull; 1069 ierr = PetscViewerDrawGetDraw(viewer,0,&draw);CHKERRQ(ierr); 1070 ierr = PetscDrawIsNull(draw,&isnull);CHKERRQ(ierr); if (isnull) PetscFunctionReturn(0); 1071 } 1072 1073 { 1074 /* assemble the entire matrix onto first processor. */ 1075 Mat A; 1076 Mat_SeqBAIJ *Aloc; 1077 PetscInt M = mat->rmap->N,N = mat->cmap->N,*ai,*aj,col,i,j,k,*rvals,mbs = baij->mbs; 1078 MatScalar *a; 1079 const char *matname; 1080 1081 /* Here we are creating a temporary matrix, so will assume MPIBAIJ is acceptable */ 1082 /* Perhaps this should be the type of mat? */ 1083 ierr = MatCreate(PetscObjectComm((PetscObject)mat),&A);CHKERRQ(ierr); 1084 if (!rank) { 1085 ierr = MatSetSizes(A,M,N,M,N);CHKERRQ(ierr); 1086 } else { 1087 ierr = MatSetSizes(A,0,0,M,N);CHKERRQ(ierr); 1088 } 1089 ierr = MatSetType(A,MATMPIBAIJ);CHKERRQ(ierr); 1090 ierr = MatMPIBAIJSetPreallocation(A,mat->rmap->bs,0,NULL,0,NULL);CHKERRQ(ierr); 1091 ierr = MatSetOption(A,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 1092 ierr = PetscLogObjectParent((PetscObject)mat,(PetscObject)A);CHKERRQ(ierr); 1093 1094 /* copy over the A part */ 1095 Aloc = (Mat_SeqBAIJ*)baij->A->data; 1096 ai = Aloc->i; aj = Aloc->j; a = Aloc->a; 1097 ierr = PetscMalloc1(bs,&rvals);CHKERRQ(ierr); 1098 1099 for (i=0; i<mbs; i++) { 1100 rvals[0] = bs*(baij->rstartbs + i); 1101 for (j=1; j<bs; j++) rvals[j] = rvals[j-1] + 1; 1102 for (j=ai[i]; j<ai[i+1]; j++) { 1103 col = (baij->cstartbs+aj[j])*bs; 1104 for (k=0; k<bs; k++) { 1105 ierr = MatSetValues_MPIBAIJ(A,bs,rvals,1,&col,a,INSERT_VALUES);CHKERRQ(ierr); 1106 col++; a += bs; 1107 } 1108 } 1109 } 1110 /* copy over the B part */ 1111 Aloc = (Mat_SeqBAIJ*)baij->B->data; 1112 ai = Aloc->i; aj = Aloc->j; a = Aloc->a; 1113 for (i=0; i<mbs; i++) { 1114 rvals[0] = bs*(baij->rstartbs + i); 1115 for (j=1; j<bs; j++) rvals[j] = rvals[j-1] + 1; 1116 for (j=ai[i]; j<ai[i+1]; j++) { 1117 col = baij->garray[aj[j]]*bs; 1118 for (k=0; k<bs; k++) { 1119 ierr = MatSetValues_MPIBAIJ(A,bs,rvals,1,&col,a,INSERT_VALUES);CHKERRQ(ierr); 1120 col++; a += bs; 1121 } 1122 } 1123 } 1124 ierr = PetscFree(rvals);CHKERRQ(ierr); 1125 ierr = MatAssemblyBegin(A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1126 ierr = MatAssemblyEnd(A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1127 /* 1128 Everyone has to call to draw the matrix since the graphics waits are 1129 synchronized across all processors that share the PetscDraw object 1130 */ 1131 ierr = PetscViewerGetSubViewer(viewer,PETSC_COMM_SELF,&sviewer);CHKERRQ(ierr); 1132 ierr = PetscObjectGetName((PetscObject)mat,&matname);CHKERRQ(ierr); 1133 if (!rank) { 1134 ierr = PetscObjectSetName((PetscObject)((Mat_MPIBAIJ*)(A->data))->A,matname);CHKERRQ(ierr); 1135 ierr = MatView_SeqBAIJ(((Mat_MPIBAIJ*)(A->data))->A,sviewer);CHKERRQ(ierr); 1136 } 1137 ierr = PetscViewerRestoreSubViewer(viewer,PETSC_COMM_SELF,&sviewer);CHKERRQ(ierr); 1138 ierr = PetscViewerFlush(viewer);CHKERRQ(ierr); 1139 ierr = MatDestroy(&A);CHKERRQ(ierr); 1140 } 1141 PetscFunctionReturn(0); 1142 } 1143 1144 #undef __FUNCT__ 1145 #define __FUNCT__ "MatView_MPIBAIJ_Binary" 1146 static PetscErrorCode MatView_MPIBAIJ_Binary(Mat mat,PetscViewer viewer) 1147 { 1148 Mat_MPIBAIJ *a = (Mat_MPIBAIJ*)mat->data; 1149 Mat_SeqBAIJ *A = (Mat_SeqBAIJ*)a->A->data; 1150 Mat_SeqBAIJ *B = (Mat_SeqBAIJ*)a->B->data; 1151 PetscErrorCode ierr; 1152 PetscInt i,*row_lens,*crow_lens,bs = mat->rmap->bs,j,k,bs2=a->bs2,header[4],nz,rlen; 1153 PetscInt *range=0,nzmax,*column_indices,cnt,col,*garray = a->garray,cstart = mat->cmap->rstart/bs,len,pcnt,l,ll; 1154 int fd; 1155 PetscScalar *column_values; 1156 FILE *file; 1157 PetscMPIInt rank,size,tag = ((PetscObject)viewer)->tag; 1158 PetscInt message_count,flowcontrolcount; 1159 1160 PetscFunctionBegin; 1161 ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)mat),&rank);CHKERRQ(ierr); 1162 ierr = MPI_Comm_size(PetscObjectComm((PetscObject)mat),&size);CHKERRQ(ierr); 1163 nz = bs2*(A->nz + B->nz); 1164 rlen = mat->rmap->n; 1165 ierr = PetscViewerBinaryGetDescriptor(viewer,&fd);CHKERRQ(ierr); 1166 if (!rank) { 1167 header[0] = MAT_FILE_CLASSID; 1168 header[1] = mat->rmap->N; 1169 header[2] = mat->cmap->N; 1170 1171 ierr = MPI_Reduce(&nz,&header[3],1,MPIU_INT,MPI_SUM,0,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1172 ierr = PetscBinaryWrite(fd,header,4,PETSC_INT,PETSC_TRUE);CHKERRQ(ierr); 1173 /* get largest number of rows any processor has */ 1174 range = mat->rmap->range; 1175 for (i=1; i<size; i++) { 1176 rlen = PetscMax(rlen,range[i+1] - range[i]); 1177 } 1178 } else { 1179 ierr = MPI_Reduce(&nz,0,1,MPIU_INT,MPI_SUM,0,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1180 } 1181 1182 ierr = PetscMalloc1(rlen/bs,&crow_lens);CHKERRQ(ierr); 1183 /* compute lengths of each row */ 1184 for (i=0; i<a->mbs; i++) { 1185 crow_lens[i] = A->i[i+1] - A->i[i] + B->i[i+1] - B->i[i]; 1186 } 1187 /* store the row lengths to the file */ 1188 ierr = PetscViewerFlowControlStart(viewer,&message_count,&flowcontrolcount);CHKERRQ(ierr); 1189 if (!rank) { 1190 MPI_Status status; 1191 ierr = PetscMalloc1(rlen,&row_lens);CHKERRQ(ierr); 1192 rlen = (range[1] - range[0])/bs; 1193 for (i=0; i<rlen; i++) { 1194 for (j=0; j<bs; j++) { 1195 row_lens[i*bs+j] = bs*crow_lens[i]; 1196 } 1197 } 1198 ierr = PetscBinaryWrite(fd,row_lens,bs*rlen,PETSC_INT,PETSC_TRUE);CHKERRQ(ierr); 1199 for (i=1; i<size; i++) { 1200 rlen = (range[i+1] - range[i])/bs; 1201 ierr = PetscViewerFlowControlStepMaster(viewer,i,&message_count,flowcontrolcount);CHKERRQ(ierr); 1202 ierr = MPI_Recv(crow_lens,rlen,MPIU_INT,i,tag,PetscObjectComm((PetscObject)mat),&status);CHKERRQ(ierr); 1203 for (k=0; k<rlen; k++) { 1204 for (j=0; j<bs; j++) { 1205 row_lens[k*bs+j] = bs*crow_lens[k]; 1206 } 1207 } 1208 ierr = PetscBinaryWrite(fd,row_lens,bs*rlen,PETSC_INT,PETSC_TRUE);CHKERRQ(ierr); 1209 } 1210 ierr = PetscViewerFlowControlEndMaster(viewer,&message_count);CHKERRQ(ierr); 1211 ierr = PetscFree(row_lens);CHKERRQ(ierr); 1212 } else { 1213 ierr = PetscViewerFlowControlStepWorker(viewer,rank,&message_count);CHKERRQ(ierr); 1214 ierr = MPI_Send(crow_lens,mat->rmap->n/bs,MPIU_INT,0,tag,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1215 ierr = PetscViewerFlowControlEndWorker(viewer,&message_count);CHKERRQ(ierr); 1216 } 1217 ierr = PetscFree(crow_lens);CHKERRQ(ierr); 1218 1219 /* load up the local column indices. Include for all rows not just one for each block row since process 0 does not have the 1220 information needed to make it for each row from a block row. This does require more communication but still not more than 1221 the communication needed for the nonzero values */ 1222 nzmax = nz; /* space a largest processor needs */ 1223 ierr = MPI_Reduce(&nz,&nzmax,1,MPIU_INT,MPI_MAX,0,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1224 ierr = PetscMalloc1(nzmax,&column_indices);CHKERRQ(ierr); 1225 cnt = 0; 1226 for (i=0; i<a->mbs; i++) { 1227 pcnt = cnt; 1228 for (j=B->i[i]; j<B->i[i+1]; j++) { 1229 if ((col = garray[B->j[j]]) > cstart) break; 1230 for (l=0; l<bs; l++) { 1231 column_indices[cnt++] = bs*col+l; 1232 } 1233 } 1234 for (k=A->i[i]; k<A->i[i+1]; k++) { 1235 for (l=0; l<bs; l++) { 1236 column_indices[cnt++] = bs*(A->j[k] + cstart)+l; 1237 } 1238 } 1239 for (; j<B->i[i+1]; j++) { 1240 for (l=0; l<bs; l++) { 1241 column_indices[cnt++] = bs*garray[B->j[j]]+l; 1242 } 1243 } 1244 len = cnt - pcnt; 1245 for (k=1; k<bs; k++) { 1246 ierr = PetscMemcpy(&column_indices[cnt],&column_indices[pcnt],len*sizeof(PetscInt));CHKERRQ(ierr); 1247 cnt += len; 1248 } 1249 } 1250 if (cnt != nz) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_LIB,"Internal PETSc error: cnt = %D nz = %D",cnt,nz); 1251 1252 /* store the columns to the file */ 1253 ierr = PetscViewerFlowControlStart(viewer,&message_count,&flowcontrolcount);CHKERRQ(ierr); 1254 if (!rank) { 1255 MPI_Status status; 1256 ierr = PetscBinaryWrite(fd,column_indices,nz,PETSC_INT,PETSC_TRUE);CHKERRQ(ierr); 1257 for (i=1; i<size; i++) { 1258 ierr = PetscViewerFlowControlStepMaster(viewer,i,&message_count,flowcontrolcount);CHKERRQ(ierr); 1259 ierr = MPI_Recv(&cnt,1,MPIU_INT,i,tag,PetscObjectComm((PetscObject)mat),&status);CHKERRQ(ierr); 1260 ierr = MPI_Recv(column_indices,cnt,MPIU_INT,i,tag,PetscObjectComm((PetscObject)mat),&status);CHKERRQ(ierr); 1261 ierr = PetscBinaryWrite(fd,column_indices,cnt,PETSC_INT,PETSC_TRUE);CHKERRQ(ierr); 1262 } 1263 ierr = PetscViewerFlowControlEndMaster(viewer,&message_count);CHKERRQ(ierr); 1264 } else { 1265 ierr = PetscViewerFlowControlStepWorker(viewer,rank,&message_count);CHKERRQ(ierr); 1266 ierr = MPI_Send(&cnt,1,MPIU_INT,0,tag,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1267 ierr = MPI_Send(column_indices,cnt,MPIU_INT,0,tag,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1268 ierr = PetscViewerFlowControlEndWorker(viewer,&message_count);CHKERRQ(ierr); 1269 } 1270 ierr = PetscFree(column_indices);CHKERRQ(ierr); 1271 1272 /* load up the numerical values */ 1273 ierr = PetscMalloc1(nzmax,&column_values);CHKERRQ(ierr); 1274 cnt = 0; 1275 for (i=0; i<a->mbs; i++) { 1276 rlen = bs*(B->i[i+1] - B->i[i] + A->i[i+1] - A->i[i]); 1277 for (j=B->i[i]; j<B->i[i+1]; j++) { 1278 if (garray[B->j[j]] > cstart) break; 1279 for (l=0; l<bs; l++) { 1280 for (ll=0; ll<bs; ll++) { 1281 column_values[cnt + l*rlen + ll] = B->a[bs2*j+l+bs*ll]; 1282 } 1283 } 1284 cnt += bs; 1285 } 1286 for (k=A->i[i]; k<A->i[i+1]; k++) { 1287 for (l=0; l<bs; l++) { 1288 for (ll=0; ll<bs; ll++) { 1289 column_values[cnt + l*rlen + ll] = A->a[bs2*k+l+bs*ll]; 1290 } 1291 } 1292 cnt += bs; 1293 } 1294 for (; j<B->i[i+1]; j++) { 1295 for (l=0; l<bs; l++) { 1296 for (ll=0; ll<bs; ll++) { 1297 column_values[cnt + l*rlen + ll] = B->a[bs2*j+l+bs*ll]; 1298 } 1299 } 1300 cnt += bs; 1301 } 1302 cnt += (bs-1)*rlen; 1303 } 1304 if (cnt != nz) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Internal PETSc error: cnt = %D nz = %D",cnt,nz); 1305 1306 /* store the column values to the file */ 1307 ierr = PetscViewerFlowControlStart(viewer,&message_count,&flowcontrolcount);CHKERRQ(ierr); 1308 if (!rank) { 1309 MPI_Status status; 1310 ierr = PetscBinaryWrite(fd,column_values,nz,PETSC_SCALAR,PETSC_TRUE);CHKERRQ(ierr); 1311 for (i=1; i<size; i++) { 1312 ierr = PetscViewerFlowControlStepMaster(viewer,i,&message_count,flowcontrolcount);CHKERRQ(ierr); 1313 ierr = MPI_Recv(&cnt,1,MPIU_INT,i,tag,PetscObjectComm((PetscObject)mat),&status);CHKERRQ(ierr); 1314 ierr = MPI_Recv(column_values,cnt,MPIU_SCALAR,i,tag,PetscObjectComm((PetscObject)mat),&status);CHKERRQ(ierr); 1315 ierr = PetscBinaryWrite(fd,column_values,cnt,PETSC_SCALAR,PETSC_TRUE);CHKERRQ(ierr); 1316 } 1317 ierr = PetscViewerFlowControlEndMaster(viewer,&message_count);CHKERRQ(ierr); 1318 } else { 1319 ierr = PetscViewerFlowControlStepWorker(viewer,rank,&message_count);CHKERRQ(ierr); 1320 ierr = MPI_Send(&nz,1,MPIU_INT,0,tag,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1321 ierr = MPI_Send(column_values,nz,MPIU_SCALAR,0,tag,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1322 ierr = PetscViewerFlowControlEndWorker(viewer,&message_count);CHKERRQ(ierr); 1323 } 1324 ierr = PetscFree(column_values);CHKERRQ(ierr); 1325 1326 ierr = PetscViewerBinaryGetInfoPointer(viewer,&file);CHKERRQ(ierr); 1327 if (file) { 1328 fprintf(file,"-matload_block_size %d\n",(int)mat->rmap->bs); 1329 } 1330 PetscFunctionReturn(0); 1331 } 1332 1333 #undef __FUNCT__ 1334 #define __FUNCT__ "MatView_MPIBAIJ" 1335 PetscErrorCode MatView_MPIBAIJ(Mat mat,PetscViewer viewer) 1336 { 1337 PetscErrorCode ierr; 1338 PetscBool iascii,isdraw,issocket,isbinary; 1339 1340 PetscFunctionBegin; 1341 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERASCII,&iascii);CHKERRQ(ierr); 1342 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERDRAW,&isdraw);CHKERRQ(ierr); 1343 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERSOCKET,&issocket);CHKERRQ(ierr); 1344 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERBINARY,&isbinary);CHKERRQ(ierr); 1345 if (iascii || isdraw || issocket) { 1346 ierr = MatView_MPIBAIJ_ASCIIorDraworSocket(mat,viewer);CHKERRQ(ierr); 1347 } else if (isbinary) { 1348 ierr = MatView_MPIBAIJ_Binary(mat,viewer);CHKERRQ(ierr); 1349 } 1350 PetscFunctionReturn(0); 1351 } 1352 1353 #undef __FUNCT__ 1354 #define __FUNCT__ "MatDestroy_MPIBAIJ" 1355 PetscErrorCode MatDestroy_MPIBAIJ(Mat mat) 1356 { 1357 Mat_MPIBAIJ *baij = (Mat_MPIBAIJ*)mat->data; 1358 PetscErrorCode ierr; 1359 1360 PetscFunctionBegin; 1361 #if defined(PETSC_USE_LOG) 1362 PetscLogObjectState((PetscObject)mat,"Rows=%D,Cols=%D",mat->rmap->N,mat->cmap->N); 1363 #endif 1364 ierr = MatStashDestroy_Private(&mat->stash);CHKERRQ(ierr); 1365 ierr = MatStashDestroy_Private(&mat->bstash);CHKERRQ(ierr); 1366 ierr = MatDestroy(&baij->A);CHKERRQ(ierr); 1367 ierr = MatDestroy(&baij->B);CHKERRQ(ierr); 1368 #if defined(PETSC_USE_CTABLE) 1369 ierr = PetscTableDestroy(&baij->colmap);CHKERRQ(ierr); 1370 #else 1371 ierr = PetscFree(baij->colmap);CHKERRQ(ierr); 1372 #endif 1373 ierr = PetscFree(baij->garray);CHKERRQ(ierr); 1374 ierr = VecDestroy(&baij->lvec);CHKERRQ(ierr); 1375 ierr = VecScatterDestroy(&baij->Mvctx);CHKERRQ(ierr); 1376 ierr = PetscFree2(baij->rowvalues,baij->rowindices);CHKERRQ(ierr); 1377 ierr = PetscFree(baij->barray);CHKERRQ(ierr); 1378 ierr = PetscFree2(baij->hd,baij->ht);CHKERRQ(ierr); 1379 ierr = PetscFree(baij->rangebs);CHKERRQ(ierr); 1380 ierr = PetscFree(mat->data);CHKERRQ(ierr); 1381 1382 ierr = PetscObjectChangeTypeName((PetscObject)mat,0);CHKERRQ(ierr); 1383 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatStoreValues_C",NULL);CHKERRQ(ierr); 1384 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatRetrieveValues_C",NULL);CHKERRQ(ierr); 1385 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatGetDiagonalBlock_C",NULL);CHKERRQ(ierr); 1386 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatMPIBAIJSetPreallocation_C",NULL);CHKERRQ(ierr); 1387 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatMPIBAIJSetPreallocationCSR_C",NULL);CHKERRQ(ierr); 1388 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatDiagonalScaleLocal_C",NULL);CHKERRQ(ierr); 1389 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatSetHashTableFactor_C",NULL);CHKERRQ(ierr); 1390 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatConvert_mpibaij_mpisbaij_C",NULL);CHKERRQ(ierr); 1391 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatConvert_mpibaij_mpibstrm_C",NULL);CHKERRQ(ierr); 1392 PetscFunctionReturn(0); 1393 } 1394 1395 #undef __FUNCT__ 1396 #define __FUNCT__ "MatMult_MPIBAIJ" 1397 PetscErrorCode MatMult_MPIBAIJ(Mat A,Vec xx,Vec yy) 1398 { 1399 Mat_MPIBAIJ *a = (Mat_MPIBAIJ*)A->data; 1400 PetscErrorCode ierr; 1401 PetscInt nt; 1402 1403 PetscFunctionBegin; 1404 ierr = VecGetLocalSize(xx,&nt);CHKERRQ(ierr); 1405 if (nt != A->cmap->n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Incompatible partition of A and xx"); 1406 ierr = VecGetLocalSize(yy,&nt);CHKERRQ(ierr); 1407 if (nt != A->rmap->n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Incompatible parition of A and yy"); 1408 ierr = VecScatterBegin(a->Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1409 ierr = (*a->A->ops->mult)(a->A,xx,yy);CHKERRQ(ierr); 1410 ierr = VecScatterEnd(a->Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1411 ierr = (*a->B->ops->multadd)(a->B,a->lvec,yy,yy);CHKERRQ(ierr); 1412 PetscFunctionReturn(0); 1413 } 1414 1415 #undef __FUNCT__ 1416 #define __FUNCT__ "MatMultAdd_MPIBAIJ" 1417 PetscErrorCode MatMultAdd_MPIBAIJ(Mat A,Vec xx,Vec yy,Vec zz) 1418 { 1419 Mat_MPIBAIJ *a = (Mat_MPIBAIJ*)A->data; 1420 PetscErrorCode ierr; 1421 1422 PetscFunctionBegin; 1423 ierr = VecScatterBegin(a->Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1424 ierr = (*a->A->ops->multadd)(a->A,xx,yy,zz);CHKERRQ(ierr); 1425 ierr = VecScatterEnd(a->Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1426 ierr = (*a->B->ops->multadd)(a->B,a->lvec,zz,zz);CHKERRQ(ierr); 1427 PetscFunctionReturn(0); 1428 } 1429 1430 #undef __FUNCT__ 1431 #define __FUNCT__ "MatMultTranspose_MPIBAIJ" 1432 PetscErrorCode MatMultTranspose_MPIBAIJ(Mat A,Vec xx,Vec yy) 1433 { 1434 Mat_MPIBAIJ *a = (Mat_MPIBAIJ*)A->data; 1435 PetscErrorCode ierr; 1436 PetscBool merged; 1437 1438 PetscFunctionBegin; 1439 ierr = VecScatterGetMerged(a->Mvctx,&merged);CHKERRQ(ierr); 1440 /* do nondiagonal part */ 1441 ierr = (*a->B->ops->multtranspose)(a->B,xx,a->lvec);CHKERRQ(ierr); 1442 if (!merged) { 1443 /* send it on its way */ 1444 ierr = VecScatterBegin(a->Mvctx,a->lvec,yy,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1445 /* do local part */ 1446 ierr = (*a->A->ops->multtranspose)(a->A,xx,yy);CHKERRQ(ierr); 1447 /* receive remote parts: note this assumes the values are not actually */ 1448 /* inserted in yy until the next line */ 1449 ierr = VecScatterEnd(a->Mvctx,a->lvec,yy,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1450 } else { 1451 /* do local part */ 1452 ierr = (*a->A->ops->multtranspose)(a->A,xx,yy);CHKERRQ(ierr); 1453 /* send it on its way */ 1454 ierr = VecScatterBegin(a->Mvctx,a->lvec,yy,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1455 /* values actually were received in the Begin() but we need to call this nop */ 1456 ierr = VecScatterEnd(a->Mvctx,a->lvec,yy,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1457 } 1458 PetscFunctionReturn(0); 1459 } 1460 1461 #undef __FUNCT__ 1462 #define __FUNCT__ "MatMultTransposeAdd_MPIBAIJ" 1463 PetscErrorCode MatMultTransposeAdd_MPIBAIJ(Mat A,Vec xx,Vec yy,Vec zz) 1464 { 1465 Mat_MPIBAIJ *a = (Mat_MPIBAIJ*)A->data; 1466 PetscErrorCode ierr; 1467 1468 PetscFunctionBegin; 1469 /* do nondiagonal part */ 1470 ierr = (*a->B->ops->multtranspose)(a->B,xx,a->lvec);CHKERRQ(ierr); 1471 /* send it on its way */ 1472 ierr = VecScatterBegin(a->Mvctx,a->lvec,zz,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1473 /* do local part */ 1474 ierr = (*a->A->ops->multtransposeadd)(a->A,xx,yy,zz);CHKERRQ(ierr); 1475 /* receive remote parts: note this assumes the values are not actually */ 1476 /* inserted in yy until the next line, which is true for my implementation*/ 1477 /* but is not perhaps always true. */ 1478 ierr = VecScatterEnd(a->Mvctx,a->lvec,zz,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1479 PetscFunctionReturn(0); 1480 } 1481 1482 /* 1483 This only works correctly for square matrices where the subblock A->A is the 1484 diagonal block 1485 */ 1486 #undef __FUNCT__ 1487 #define __FUNCT__ "MatGetDiagonal_MPIBAIJ" 1488 PetscErrorCode MatGetDiagonal_MPIBAIJ(Mat A,Vec v) 1489 { 1490 Mat_MPIBAIJ *a = (Mat_MPIBAIJ*)A->data; 1491 PetscErrorCode ierr; 1492 1493 PetscFunctionBegin; 1494 if (A->rmap->N != A->cmap->N) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Supports only square matrix where A->A is diag block"); 1495 ierr = MatGetDiagonal(a->A,v);CHKERRQ(ierr); 1496 PetscFunctionReturn(0); 1497 } 1498 1499 #undef __FUNCT__ 1500 #define __FUNCT__ "MatScale_MPIBAIJ" 1501 PetscErrorCode MatScale_MPIBAIJ(Mat A,PetscScalar aa) 1502 { 1503 Mat_MPIBAIJ *a = (Mat_MPIBAIJ*)A->data; 1504 PetscErrorCode ierr; 1505 1506 PetscFunctionBegin; 1507 ierr = MatScale(a->A,aa);CHKERRQ(ierr); 1508 ierr = MatScale(a->B,aa);CHKERRQ(ierr); 1509 PetscFunctionReturn(0); 1510 } 1511 1512 #undef __FUNCT__ 1513 #define __FUNCT__ "MatGetRow_MPIBAIJ" 1514 PetscErrorCode MatGetRow_MPIBAIJ(Mat matin,PetscInt row,PetscInt *nz,PetscInt **idx,PetscScalar **v) 1515 { 1516 Mat_MPIBAIJ *mat = (Mat_MPIBAIJ*)matin->data; 1517 PetscScalar *vworkA,*vworkB,**pvA,**pvB,*v_p; 1518 PetscErrorCode ierr; 1519 PetscInt bs = matin->rmap->bs,bs2 = mat->bs2,i,*cworkA,*cworkB,**pcA,**pcB; 1520 PetscInt nztot,nzA,nzB,lrow,brstart = matin->rmap->rstart,brend = matin->rmap->rend; 1521 PetscInt *cmap,*idx_p,cstart = mat->cstartbs; 1522 1523 PetscFunctionBegin; 1524 if (row < brstart || row >= brend) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Only local rows"); 1525 if (mat->getrowactive) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Already active"); 1526 mat->getrowactive = PETSC_TRUE; 1527 1528 if (!mat->rowvalues && (idx || v)) { 1529 /* 1530 allocate enough space to hold information from the longest row. 1531 */ 1532 Mat_SeqBAIJ *Aa = (Mat_SeqBAIJ*)mat->A->data,*Ba = (Mat_SeqBAIJ*)mat->B->data; 1533 PetscInt max = 1,mbs = mat->mbs,tmp; 1534 for (i=0; i<mbs; i++) { 1535 tmp = Aa->i[i+1] - Aa->i[i] + Ba->i[i+1] - Ba->i[i]; 1536 if (max < tmp) max = tmp; 1537 } 1538 ierr = PetscMalloc2(max*bs2,&mat->rowvalues,max*bs2,&mat->rowindices);CHKERRQ(ierr); 1539 } 1540 lrow = row - brstart; 1541 1542 pvA = &vworkA; pcA = &cworkA; pvB = &vworkB; pcB = &cworkB; 1543 if (!v) {pvA = 0; pvB = 0;} 1544 if (!idx) {pcA = 0; if (!v) pcB = 0;} 1545 ierr = (*mat->A->ops->getrow)(mat->A,lrow,&nzA,pcA,pvA);CHKERRQ(ierr); 1546 ierr = (*mat->B->ops->getrow)(mat->B,lrow,&nzB,pcB,pvB);CHKERRQ(ierr); 1547 nztot = nzA + nzB; 1548 1549 cmap = mat->garray; 1550 if (v || idx) { 1551 if (nztot) { 1552 /* Sort by increasing column numbers, assuming A and B already sorted */ 1553 PetscInt imark = -1; 1554 if (v) { 1555 *v = v_p = mat->rowvalues; 1556 for (i=0; i<nzB; i++) { 1557 if (cmap[cworkB[i]/bs] < cstart) v_p[i] = vworkB[i]; 1558 else break; 1559 } 1560 imark = i; 1561 for (i=0; i<nzA; i++) v_p[imark+i] = vworkA[i]; 1562 for (i=imark; i<nzB; i++) v_p[nzA+i] = vworkB[i]; 1563 } 1564 if (idx) { 1565 *idx = idx_p = mat->rowindices; 1566 if (imark > -1) { 1567 for (i=0; i<imark; i++) { 1568 idx_p[i] = cmap[cworkB[i]/bs]*bs + cworkB[i]%bs; 1569 } 1570 } else { 1571 for (i=0; i<nzB; i++) { 1572 if (cmap[cworkB[i]/bs] < cstart) idx_p[i] = cmap[cworkB[i]/bs]*bs + cworkB[i]%bs; 1573 else break; 1574 } 1575 imark = i; 1576 } 1577 for (i=0; i<nzA; i++) idx_p[imark+i] = cstart*bs + cworkA[i]; 1578 for (i=imark; i<nzB; i++) idx_p[nzA+i] = cmap[cworkB[i]/bs]*bs + cworkB[i]%bs ; 1579 } 1580 } else { 1581 if (idx) *idx = 0; 1582 if (v) *v = 0; 1583 } 1584 } 1585 *nz = nztot; 1586 ierr = (*mat->A->ops->restorerow)(mat->A,lrow,&nzA,pcA,pvA);CHKERRQ(ierr); 1587 ierr = (*mat->B->ops->restorerow)(mat->B,lrow,&nzB,pcB,pvB);CHKERRQ(ierr); 1588 PetscFunctionReturn(0); 1589 } 1590 1591 #undef __FUNCT__ 1592 #define __FUNCT__ "MatRestoreRow_MPIBAIJ" 1593 PetscErrorCode MatRestoreRow_MPIBAIJ(Mat mat,PetscInt row,PetscInt *nz,PetscInt **idx,PetscScalar **v) 1594 { 1595 Mat_MPIBAIJ *baij = (Mat_MPIBAIJ*)mat->data; 1596 1597 PetscFunctionBegin; 1598 if (!baij->getrowactive) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"MatGetRow not called"); 1599 baij->getrowactive = PETSC_FALSE; 1600 PetscFunctionReturn(0); 1601 } 1602 1603 #undef __FUNCT__ 1604 #define __FUNCT__ "MatZeroEntries_MPIBAIJ" 1605 PetscErrorCode MatZeroEntries_MPIBAIJ(Mat A) 1606 { 1607 Mat_MPIBAIJ *l = (Mat_MPIBAIJ*)A->data; 1608 PetscErrorCode ierr; 1609 1610 PetscFunctionBegin; 1611 ierr = MatZeroEntries(l->A);CHKERRQ(ierr); 1612 ierr = MatZeroEntries(l->B);CHKERRQ(ierr); 1613 PetscFunctionReturn(0); 1614 } 1615 1616 #undef __FUNCT__ 1617 #define __FUNCT__ "MatGetInfo_MPIBAIJ" 1618 PetscErrorCode MatGetInfo_MPIBAIJ(Mat matin,MatInfoType flag,MatInfo *info) 1619 { 1620 Mat_MPIBAIJ *a = (Mat_MPIBAIJ*)matin->data; 1621 Mat A = a->A,B = a->B; 1622 PetscErrorCode ierr; 1623 PetscReal isend[5],irecv[5]; 1624 1625 PetscFunctionBegin; 1626 info->block_size = (PetscReal)matin->rmap->bs; 1627 1628 ierr = MatGetInfo(A,MAT_LOCAL,info);CHKERRQ(ierr); 1629 1630 isend[0] = info->nz_used; isend[1] = info->nz_allocated; isend[2] = info->nz_unneeded; 1631 isend[3] = info->memory; isend[4] = info->mallocs; 1632 1633 ierr = MatGetInfo(B,MAT_LOCAL,info);CHKERRQ(ierr); 1634 1635 isend[0] += info->nz_used; isend[1] += info->nz_allocated; isend[2] += info->nz_unneeded; 1636 isend[3] += info->memory; isend[4] += info->mallocs; 1637 1638 if (flag == MAT_LOCAL) { 1639 info->nz_used = isend[0]; 1640 info->nz_allocated = isend[1]; 1641 info->nz_unneeded = isend[2]; 1642 info->memory = isend[3]; 1643 info->mallocs = isend[4]; 1644 } else if (flag == MAT_GLOBAL_MAX) { 1645 ierr = MPI_Allreduce(isend,irecv,5,MPIU_REAL,MPIU_MAX,PetscObjectComm((PetscObject)matin));CHKERRQ(ierr); 1646 1647 info->nz_used = irecv[0]; 1648 info->nz_allocated = irecv[1]; 1649 info->nz_unneeded = irecv[2]; 1650 info->memory = irecv[3]; 1651 info->mallocs = irecv[4]; 1652 } else if (flag == MAT_GLOBAL_SUM) { 1653 ierr = MPI_Allreduce(isend,irecv,5,MPIU_REAL,MPIU_SUM,PetscObjectComm((PetscObject)matin));CHKERRQ(ierr); 1654 1655 info->nz_used = irecv[0]; 1656 info->nz_allocated = irecv[1]; 1657 info->nz_unneeded = irecv[2]; 1658 info->memory = irecv[3]; 1659 info->mallocs = irecv[4]; 1660 } else SETERRQ1(PetscObjectComm((PetscObject)matin),PETSC_ERR_ARG_WRONG,"Unknown MatInfoType argument %d",(int)flag); 1661 info->fill_ratio_given = 0; /* no parallel LU/ILU/Cholesky */ 1662 info->fill_ratio_needed = 0; 1663 info->factor_mallocs = 0; 1664 PetscFunctionReturn(0); 1665 } 1666 1667 #undef __FUNCT__ 1668 #define __FUNCT__ "MatSetOption_MPIBAIJ" 1669 PetscErrorCode MatSetOption_MPIBAIJ(Mat A,MatOption op,PetscBool flg) 1670 { 1671 Mat_MPIBAIJ *a = (Mat_MPIBAIJ*)A->data; 1672 PetscErrorCode ierr; 1673 1674 PetscFunctionBegin; 1675 switch (op) { 1676 case MAT_NEW_NONZERO_LOCATIONS: 1677 case MAT_NEW_NONZERO_ALLOCATION_ERR: 1678 case MAT_UNUSED_NONZERO_LOCATION_ERR: 1679 case MAT_KEEP_NONZERO_PATTERN: 1680 case MAT_NEW_NONZERO_LOCATION_ERR: 1681 ierr = MatSetOption(a->A,op,flg);CHKERRQ(ierr); 1682 ierr = MatSetOption(a->B,op,flg);CHKERRQ(ierr); 1683 break; 1684 case MAT_ROW_ORIENTED: 1685 a->roworiented = flg; 1686 1687 ierr = MatSetOption(a->A,op,flg);CHKERRQ(ierr); 1688 ierr = MatSetOption(a->B,op,flg);CHKERRQ(ierr); 1689 break; 1690 case MAT_NEW_DIAGONALS: 1691 ierr = PetscInfo1(A,"Option %s ignored\n",MatOptions[op]);CHKERRQ(ierr); 1692 break; 1693 case MAT_IGNORE_OFF_PROC_ENTRIES: 1694 a->donotstash = flg; 1695 break; 1696 case MAT_USE_HASH_TABLE: 1697 a->ht_flag = flg; 1698 break; 1699 case MAT_SYMMETRIC: 1700 case MAT_STRUCTURALLY_SYMMETRIC: 1701 case MAT_HERMITIAN: 1702 case MAT_SYMMETRY_ETERNAL: 1703 ierr = MatSetOption(a->A,op,flg);CHKERRQ(ierr); 1704 break; 1705 default: 1706 SETERRQ1(PetscObjectComm((PetscObject)A),PETSC_ERR_SUP,"unknown option %d",op); 1707 } 1708 PetscFunctionReturn(0); 1709 } 1710 1711 #undef __FUNCT__ 1712 #define __FUNCT__ "MatTranspose_MPIBAIJ" 1713 PetscErrorCode MatTranspose_MPIBAIJ(Mat A,MatReuse reuse,Mat *matout) 1714 { 1715 Mat_MPIBAIJ *baij = (Mat_MPIBAIJ*)A->data; 1716 Mat_SeqBAIJ *Aloc; 1717 Mat B; 1718 PetscErrorCode ierr; 1719 PetscInt M =A->rmap->N,N=A->cmap->N,*ai,*aj,i,*rvals,j,k,col; 1720 PetscInt bs=A->rmap->bs,mbs=baij->mbs; 1721 MatScalar *a; 1722 1723 PetscFunctionBegin; 1724 if (reuse == MAT_REUSE_MATRIX && A == *matout && M != N) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_ARG_SIZ,"Square matrix only for in-place"); 1725 if (reuse == MAT_INITIAL_MATRIX || *matout == A) { 1726 ierr = MatCreate(PetscObjectComm((PetscObject)A),&B);CHKERRQ(ierr); 1727 ierr = MatSetSizes(B,A->cmap->n,A->rmap->n,N,M);CHKERRQ(ierr); 1728 ierr = MatSetType(B,((PetscObject)A)->type_name);CHKERRQ(ierr); 1729 /* Do not know preallocation information, but must set block size */ 1730 ierr = MatMPIBAIJSetPreallocation(B,A->rmap->bs,PETSC_DECIDE,NULL,PETSC_DECIDE,NULL);CHKERRQ(ierr); 1731 } else { 1732 B = *matout; 1733 } 1734 1735 /* copy over the A part */ 1736 Aloc = (Mat_SeqBAIJ*)baij->A->data; 1737 ai = Aloc->i; aj = Aloc->j; a = Aloc->a; 1738 ierr = PetscMalloc1(bs,&rvals);CHKERRQ(ierr); 1739 1740 for (i=0; i<mbs; i++) { 1741 rvals[0] = bs*(baij->rstartbs + i); 1742 for (j=1; j<bs; j++) rvals[j] = rvals[j-1] + 1; 1743 for (j=ai[i]; j<ai[i+1]; j++) { 1744 col = (baij->cstartbs+aj[j])*bs; 1745 for (k=0; k<bs; k++) { 1746 ierr = MatSetValues_MPIBAIJ(B,1,&col,bs,rvals,a,INSERT_VALUES);CHKERRQ(ierr); 1747 1748 col++; a += bs; 1749 } 1750 } 1751 } 1752 /* copy over the B part */ 1753 Aloc = (Mat_SeqBAIJ*)baij->B->data; 1754 ai = Aloc->i; aj = Aloc->j; a = Aloc->a; 1755 for (i=0; i<mbs; i++) { 1756 rvals[0] = bs*(baij->rstartbs + i); 1757 for (j=1; j<bs; j++) rvals[j] = rvals[j-1] + 1; 1758 for (j=ai[i]; j<ai[i+1]; j++) { 1759 col = baij->garray[aj[j]]*bs; 1760 for (k=0; k<bs; k++) { 1761 ierr = MatSetValues_MPIBAIJ(B,1,&col,bs,rvals,a,INSERT_VALUES);CHKERRQ(ierr); 1762 col++; 1763 a += bs; 1764 } 1765 } 1766 } 1767 ierr = PetscFree(rvals);CHKERRQ(ierr); 1768 ierr = MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1769 ierr = MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1770 1771 if (reuse == MAT_INITIAL_MATRIX || *matout != A) *matout = B; 1772 else { 1773 ierr = MatHeaderMerge(A,B);CHKERRQ(ierr); 1774 } 1775 PetscFunctionReturn(0); 1776 } 1777 1778 #undef __FUNCT__ 1779 #define __FUNCT__ "MatDiagonalScale_MPIBAIJ" 1780 PetscErrorCode MatDiagonalScale_MPIBAIJ(Mat mat,Vec ll,Vec rr) 1781 { 1782 Mat_MPIBAIJ *baij = (Mat_MPIBAIJ*)mat->data; 1783 Mat a = baij->A,b = baij->B; 1784 PetscErrorCode ierr; 1785 PetscInt s1,s2,s3; 1786 1787 PetscFunctionBegin; 1788 ierr = MatGetLocalSize(mat,&s2,&s3);CHKERRQ(ierr); 1789 if (rr) { 1790 ierr = VecGetLocalSize(rr,&s1);CHKERRQ(ierr); 1791 if (s1!=s3) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"right vector non-conforming local size"); 1792 /* Overlap communication with computation. */ 1793 ierr = VecScatterBegin(baij->Mvctx,rr,baij->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1794 } 1795 if (ll) { 1796 ierr = VecGetLocalSize(ll,&s1);CHKERRQ(ierr); 1797 if (s1!=s2) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"left vector non-conforming local size"); 1798 ierr = (*b->ops->diagonalscale)(b,ll,NULL);CHKERRQ(ierr); 1799 } 1800 /* scale the diagonal block */ 1801 ierr = (*a->ops->diagonalscale)(a,ll,rr);CHKERRQ(ierr); 1802 1803 if (rr) { 1804 /* Do a scatter end and then right scale the off-diagonal block */ 1805 ierr = VecScatterEnd(baij->Mvctx,rr,baij->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1806 ierr = (*b->ops->diagonalscale)(b,NULL,baij->lvec);CHKERRQ(ierr); 1807 } 1808 PetscFunctionReturn(0); 1809 } 1810 1811 #undef __FUNCT__ 1812 #define __FUNCT__ "MatZeroRows_MPIBAIJ" 1813 PetscErrorCode MatZeroRows_MPIBAIJ(Mat A,PetscInt N,const PetscInt rows[],PetscScalar diag,Vec x,Vec b) 1814 { 1815 Mat_MPIBAIJ *l = (Mat_MPIBAIJ *) A->data; 1816 PetscInt *owners = A->rmap->range; 1817 PetscInt n = A->rmap->n; 1818 PetscSF sf; 1819 PetscInt *lrows; 1820 PetscSFNode *rrows; 1821 PetscInt r, p = 0, len = 0; 1822 PetscErrorCode ierr; 1823 1824 PetscFunctionBegin; 1825 /* Create SF where leaves are input rows and roots are owned rows */ 1826 ierr = PetscMalloc1(n, &lrows);CHKERRQ(ierr); 1827 for (r = 0; r < n; ++r) lrows[r] = -1; 1828 if (!A->nooffproczerorows) {ierr = PetscMalloc1(N, &rrows);CHKERRQ(ierr);} 1829 for (r = 0; r < N; ++r) { 1830 const PetscInt idx = rows[r]; 1831 if (idx < 0 || A->rmap->N <= idx) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Row %D out of range [0,%D)",idx,A->rmap->N); 1832 if (idx < owners[p] || owners[p+1] <= idx) { /* short-circuit the search if the last p owns this row too */ 1833 ierr = PetscLayoutFindOwner(A->rmap,idx,&p);CHKERRQ(ierr); 1834 } 1835 if (A->nooffproczerorows) { 1836 if (p != l->rank) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"MAT_NO_OFF_PROC_ZERO_ROWS set, but row %D is not owned by rank %d",idx,l->rank); 1837 lrows[len++] = idx - owners[p]; 1838 } else { 1839 rrows[r].rank = p; 1840 rrows[r].index = rows[r] - owners[p]; 1841 } 1842 } 1843 if (!A->nooffproczerorows) { 1844 ierr = PetscSFCreate(PetscObjectComm((PetscObject) A), &sf);CHKERRQ(ierr); 1845 ierr = PetscSFSetGraph(sf, n, N, NULL, PETSC_OWN_POINTER, rrows, PETSC_OWN_POINTER);CHKERRQ(ierr); 1846 /* Collect flags for rows to be zeroed */ 1847 ierr = PetscSFReduceBegin(sf, MPIU_INT, (PetscInt *) rows, lrows, MPI_LOR);CHKERRQ(ierr); 1848 ierr = PetscSFReduceEnd(sf, MPIU_INT, (PetscInt *) rows, lrows, MPI_LOR);CHKERRQ(ierr); 1849 ierr = PetscSFDestroy(&sf);CHKERRQ(ierr); 1850 /* Compress and put in row numbers */ 1851 for (r = 0; r < n; ++r) if (lrows[r] >= 0) lrows[len++] = r; 1852 } 1853 /* fix right hand side if needed */ 1854 if (x && b) { 1855 const PetscScalar *xx; 1856 PetscScalar *bb; 1857 1858 ierr = VecGetArrayRead(x,&xx);CHKERRQ(ierr); 1859 ierr = VecGetArray(b,&bb);CHKERRQ(ierr); 1860 for (r = 0; r < len; ++r) bb[lrows[r]] = diag*xx[lrows[r]]; 1861 ierr = VecRestoreArrayRead(x,&xx);CHKERRQ(ierr); 1862 ierr = VecRestoreArray(b,&bb);CHKERRQ(ierr); 1863 } 1864 1865 /* actually zap the local rows */ 1866 /* 1867 Zero the required rows. If the "diagonal block" of the matrix 1868 is square and the user wishes to set the diagonal we use separate 1869 code so that MatSetValues() is not called for each diagonal allocating 1870 new memory, thus calling lots of mallocs and slowing things down. 1871 1872 */ 1873 /* must zero l->B before l->A because the (diag) case below may put values into l->B*/ 1874 ierr = MatZeroRows_SeqBAIJ(l->B,len,lrows,0.0,NULL,NULL);CHKERRQ(ierr); 1875 if ((diag != 0.0) && (l->A->rmap->N == l->A->cmap->N)) { 1876 ierr = MatZeroRows_SeqBAIJ(l->A,len,lrows,diag,NULL,NULL);CHKERRQ(ierr); 1877 } else if (diag != 0.0) { 1878 ierr = MatZeroRows_SeqBAIJ(l->A,len,lrows,0.0,0,0);CHKERRQ(ierr); 1879 if (((Mat_SeqBAIJ*)l->A->data)->nonew) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"MatZeroRows() on rectangular matrices cannot be used with the Mat options \n\ 1880 MAT_NEW_NONZERO_LOCATIONS,MAT_NEW_NONZERO_LOCATION_ERR,MAT_NEW_NONZERO_ALLOCATION_ERR"); 1881 for (r = 0; r < len; ++r) { 1882 const PetscInt row = lrows[r] + A->rmap->rstart; 1883 ierr = MatSetValues(A,1,&row,1,&row,&diag,INSERT_VALUES);CHKERRQ(ierr); 1884 } 1885 ierr = MatAssemblyBegin(A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1886 ierr = MatAssemblyEnd(A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1887 } else { 1888 ierr = MatZeroRows_SeqBAIJ(l->A,len,lrows,0.0,NULL,NULL);CHKERRQ(ierr); 1889 } 1890 ierr = PetscFree(lrows);CHKERRQ(ierr); 1891 1892 /* only change matrix nonzero state if pattern was allowed to be changed */ 1893 if (!((Mat_SeqBAIJ*)(l->A->data))->keepnonzeropattern) { 1894 PetscObjectState state = l->A->nonzerostate + l->B->nonzerostate; 1895 ierr = MPI_Allreduce(&state,&A->nonzerostate,1,MPIU_INT64,MPI_SUM,PetscObjectComm((PetscObject)A));CHKERRQ(ierr); 1896 } 1897 PetscFunctionReturn(0); 1898 } 1899 1900 #undef __FUNCT__ 1901 #define __FUNCT__ "MatZeroRowsColumns_MPIBAIJ" 1902 PetscErrorCode MatZeroRowsColumns_MPIBAIJ(Mat A,PetscInt N,const PetscInt rows[],PetscScalar diag,Vec x,Vec b) 1903 { 1904 Mat_MPIBAIJ *l = (Mat_MPIBAIJ*)A->data; 1905 PetscErrorCode ierr; 1906 PetscMPIInt n = A->rmap->n; 1907 PetscInt i,j,k,r,p = 0,len = 0,row,col,count; 1908 PetscInt *lrows,*owners = A->rmap->range; 1909 PetscSFNode *rrows; 1910 PetscSF sf; 1911 const PetscScalar *xx; 1912 PetscScalar *bb,*mask; 1913 Vec xmask,lmask; 1914 Mat_SeqBAIJ *baij = (Mat_SeqBAIJ*)l->B->data; 1915 PetscInt bs = A->rmap->bs, bs2 = baij->bs2; 1916 PetscScalar *aa; 1917 1918 PetscFunctionBegin; 1919 /* Create SF where leaves are input rows and roots are owned rows */ 1920 ierr = PetscMalloc1(n, &lrows);CHKERRQ(ierr); 1921 for (r = 0; r < n; ++r) lrows[r] = -1; 1922 ierr = PetscMalloc1(N, &rrows);CHKERRQ(ierr); 1923 for (r = 0; r < N; ++r) { 1924 const PetscInt idx = rows[r]; 1925 if (idx < 0 || A->rmap->N <= idx) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Row %D out of range [0,%D)",idx,A->rmap->N); 1926 if (idx < owners[p] || owners[p+1] <= idx) { /* short-circuit the search if the last p owns this row too */ 1927 ierr = PetscLayoutFindOwner(A->rmap,idx,&p);CHKERRQ(ierr); 1928 } 1929 rrows[r].rank = p; 1930 rrows[r].index = rows[r] - owners[p]; 1931 } 1932 ierr = PetscSFCreate(PetscObjectComm((PetscObject) A), &sf);CHKERRQ(ierr); 1933 ierr = PetscSFSetGraph(sf, n, N, NULL, PETSC_OWN_POINTER, rrows, PETSC_OWN_POINTER);CHKERRQ(ierr); 1934 /* Collect flags for rows to be zeroed */ 1935 ierr = PetscSFReduceBegin(sf, MPIU_INT, (PetscInt *) rows, lrows, MPI_LOR);CHKERRQ(ierr); 1936 ierr = PetscSFReduceEnd(sf, MPIU_INT, (PetscInt *) rows, lrows, MPI_LOR);CHKERRQ(ierr); 1937 ierr = PetscSFDestroy(&sf);CHKERRQ(ierr); 1938 /* Compress and put in row numbers */ 1939 for (r = 0; r < n; ++r) if (lrows[r] >= 0) lrows[len++] = r; 1940 /* zero diagonal part of matrix */ 1941 ierr = MatZeroRowsColumns(l->A,len,lrows,diag,x,b);CHKERRQ(ierr); 1942 /* handle off diagonal part of matrix */ 1943 ierr = MatCreateVecs(A,&xmask,NULL);CHKERRQ(ierr); 1944 ierr = VecDuplicate(l->lvec,&lmask);CHKERRQ(ierr); 1945 ierr = VecGetArray(xmask,&bb);CHKERRQ(ierr); 1946 for (i=0; i<len; i++) bb[lrows[i]] = 1; 1947 ierr = VecRestoreArray(xmask,&bb);CHKERRQ(ierr); 1948 ierr = VecScatterBegin(l->Mvctx,xmask,lmask,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1949 ierr = VecScatterEnd(l->Mvctx,xmask,lmask,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1950 ierr = VecDestroy(&xmask);CHKERRQ(ierr); 1951 if (x) { 1952 ierr = VecScatterBegin(l->Mvctx,x,l->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1953 ierr = VecScatterEnd(l->Mvctx,x,l->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1954 ierr = VecGetArrayRead(l->lvec,&xx);CHKERRQ(ierr); 1955 ierr = VecGetArray(b,&bb);CHKERRQ(ierr); 1956 } 1957 ierr = VecGetArray(lmask,&mask);CHKERRQ(ierr); 1958 /* remove zeroed rows of off diagonal matrix */ 1959 for (i = 0; i < len; ++i) { 1960 row = lrows[i]; 1961 count = (baij->i[row/bs +1] - baij->i[row/bs])*bs; 1962 aa = ((MatScalar*)(baij->a)) + baij->i[row/bs]*bs2 + (row%bs); 1963 for (k = 0; k < count; ++k) { 1964 aa[0] = 0.0; 1965 aa += bs; 1966 } 1967 } 1968 /* loop over all elements of off process part of matrix zeroing removed columns*/ 1969 for (i = 0; i < l->B->rmap->N; ++i) { 1970 row = i/bs; 1971 for (j = baij->i[row]; j < baij->i[row+1]; ++j) { 1972 for (k = 0; k < bs; ++k) { 1973 col = bs*baij->j[j] + k; 1974 if (PetscAbsScalar(mask[col])) { 1975 aa = ((MatScalar*)(baij->a)) + j*bs2 + (i%bs) + bs*k; 1976 if (b) bb[i] -= aa[0]*xx[col]; 1977 aa[0] = 0.0; 1978 } 1979 } 1980 } 1981 } 1982 if (x) { 1983 ierr = VecRestoreArray(b,&bb);CHKERRQ(ierr); 1984 ierr = VecRestoreArrayRead(l->lvec,&xx);CHKERRQ(ierr); 1985 } 1986 ierr = VecRestoreArray(lmask,&mask);CHKERRQ(ierr); 1987 ierr = VecDestroy(&lmask);CHKERRQ(ierr); 1988 ierr = PetscFree(lrows);CHKERRQ(ierr); 1989 1990 /* only change matrix nonzero state if pattern was allowed to be changed */ 1991 if (!((Mat_SeqBAIJ*)(l->A->data))->keepnonzeropattern) { 1992 PetscObjectState state = l->A->nonzerostate + l->B->nonzerostate; 1993 ierr = MPI_Allreduce(&state,&A->nonzerostate,1,MPIU_INT64,MPI_SUM,PetscObjectComm((PetscObject)A));CHKERRQ(ierr); 1994 } 1995 PetscFunctionReturn(0); 1996 } 1997 1998 #undef __FUNCT__ 1999 #define __FUNCT__ "MatSetUnfactored_MPIBAIJ" 2000 PetscErrorCode MatSetUnfactored_MPIBAIJ(Mat A) 2001 { 2002 Mat_MPIBAIJ *a = (Mat_MPIBAIJ*)A->data; 2003 PetscErrorCode ierr; 2004 2005 PetscFunctionBegin; 2006 ierr = MatSetUnfactored(a->A);CHKERRQ(ierr); 2007 PetscFunctionReturn(0); 2008 } 2009 2010 static PetscErrorCode MatDuplicate_MPIBAIJ(Mat,MatDuplicateOption,Mat*); 2011 2012 #undef __FUNCT__ 2013 #define __FUNCT__ "MatEqual_MPIBAIJ" 2014 PetscErrorCode MatEqual_MPIBAIJ(Mat A,Mat B,PetscBool *flag) 2015 { 2016 Mat_MPIBAIJ *matB = (Mat_MPIBAIJ*)B->data,*matA = (Mat_MPIBAIJ*)A->data; 2017 Mat a,b,c,d; 2018 PetscBool flg; 2019 PetscErrorCode ierr; 2020 2021 PetscFunctionBegin; 2022 a = matA->A; b = matA->B; 2023 c = matB->A; d = matB->B; 2024 2025 ierr = MatEqual(a,c,&flg);CHKERRQ(ierr); 2026 if (flg) { 2027 ierr = MatEqual(b,d,&flg);CHKERRQ(ierr); 2028 } 2029 ierr = MPI_Allreduce(&flg,flag,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)A));CHKERRQ(ierr); 2030 PetscFunctionReturn(0); 2031 } 2032 2033 #undef __FUNCT__ 2034 #define __FUNCT__ "MatCopy_MPIBAIJ" 2035 PetscErrorCode MatCopy_MPIBAIJ(Mat A,Mat B,MatStructure str) 2036 { 2037 PetscErrorCode ierr; 2038 Mat_MPIBAIJ *a = (Mat_MPIBAIJ*)A->data; 2039 Mat_MPIBAIJ *b = (Mat_MPIBAIJ*)B->data; 2040 2041 PetscFunctionBegin; 2042 /* If the two matrices don't have the same copy implementation, they aren't compatible for fast copy. */ 2043 if ((str != SAME_NONZERO_PATTERN) || (A->ops->copy != B->ops->copy)) { 2044 ierr = MatCopy_Basic(A,B,str);CHKERRQ(ierr); 2045 } else { 2046 ierr = MatCopy(a->A,b->A,str);CHKERRQ(ierr); 2047 ierr = MatCopy(a->B,b->B,str);CHKERRQ(ierr); 2048 } 2049 PetscFunctionReturn(0); 2050 } 2051 2052 #undef __FUNCT__ 2053 #define __FUNCT__ "MatSetUp_MPIBAIJ" 2054 PetscErrorCode MatSetUp_MPIBAIJ(Mat A) 2055 { 2056 PetscErrorCode ierr; 2057 2058 PetscFunctionBegin; 2059 ierr = MatMPIBAIJSetPreallocation(A,A->rmap->bs,PETSC_DEFAULT,0,PETSC_DEFAULT,0);CHKERRQ(ierr); 2060 PetscFunctionReturn(0); 2061 } 2062 2063 #undef __FUNCT__ 2064 #define __FUNCT__ "MatAXPYGetPreallocation_MPIBAIJ" 2065 PetscErrorCode MatAXPYGetPreallocation_MPIBAIJ(Mat Y,const PetscInt *yltog,Mat X,const PetscInt *xltog,PetscInt *nnz) 2066 { 2067 PetscErrorCode ierr; 2068 PetscInt bs = Y->rmap->bs,m = Y->rmap->N/bs; 2069 Mat_SeqBAIJ *x = (Mat_SeqBAIJ*)X->data; 2070 Mat_SeqBAIJ *y = (Mat_SeqBAIJ*)Y->data; 2071 2072 PetscFunctionBegin; 2073 ierr = MatAXPYGetPreallocation_MPIX_private(m,x->i,x->j,xltog,y->i,y->j,yltog,nnz);CHKERRQ(ierr); 2074 PetscFunctionReturn(0); 2075 } 2076 2077 #undef __FUNCT__ 2078 #define __FUNCT__ "MatAXPY_MPIBAIJ" 2079 PetscErrorCode MatAXPY_MPIBAIJ(Mat Y,PetscScalar a,Mat X,MatStructure str) 2080 { 2081 PetscErrorCode ierr; 2082 Mat_MPIBAIJ *xx=(Mat_MPIBAIJ*)X->data,*yy=(Mat_MPIBAIJ*)Y->data; 2083 PetscBLASInt bnz,one=1; 2084 Mat_SeqBAIJ *x,*y; 2085 2086 PetscFunctionBegin; 2087 if (str == SAME_NONZERO_PATTERN) { 2088 PetscScalar alpha = a; 2089 x = (Mat_SeqBAIJ*)xx->A->data; 2090 y = (Mat_SeqBAIJ*)yy->A->data; 2091 ierr = PetscBLASIntCast(x->nz,&bnz);CHKERRQ(ierr); 2092 PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&bnz,&alpha,x->a,&one,y->a,&one)); 2093 x = (Mat_SeqBAIJ*)xx->B->data; 2094 y = (Mat_SeqBAIJ*)yy->B->data; 2095 ierr = PetscBLASIntCast(x->nz,&bnz);CHKERRQ(ierr); 2096 PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&bnz,&alpha,x->a,&one,y->a,&one)); 2097 ierr = PetscObjectStateIncrease((PetscObject)Y);CHKERRQ(ierr); 2098 } else if (str == SUBSET_NONZERO_PATTERN) { /* nonzeros of X is a subset of Y's */ 2099 ierr = MatAXPY_Basic(Y,a,X,str);CHKERRQ(ierr); 2100 } else { 2101 Mat B; 2102 PetscInt *nnz_d,*nnz_o,bs=Y->rmap->bs; 2103 ierr = PetscMalloc1(yy->A->rmap->N,&nnz_d);CHKERRQ(ierr); 2104 ierr = PetscMalloc1(yy->B->rmap->N,&nnz_o);CHKERRQ(ierr); 2105 ierr = MatCreate(PetscObjectComm((PetscObject)Y),&B);CHKERRQ(ierr); 2106 ierr = PetscObjectSetName((PetscObject)B,((PetscObject)Y)->name);CHKERRQ(ierr); 2107 ierr = MatSetSizes(B,Y->rmap->n,Y->cmap->n,Y->rmap->N,Y->cmap->N);CHKERRQ(ierr); 2108 ierr = MatSetBlockSizesFromMats(B,Y,Y);CHKERRQ(ierr); 2109 ierr = MatSetType(B,MATMPIBAIJ);CHKERRQ(ierr); 2110 ierr = MatAXPYGetPreallocation_SeqBAIJ(yy->A,xx->A,nnz_d);CHKERRQ(ierr); 2111 ierr = MatAXPYGetPreallocation_MPIBAIJ(yy->B,yy->garray,xx->B,xx->garray,nnz_o);CHKERRQ(ierr); 2112 ierr = MatMPIBAIJSetPreallocation(B,bs,0,nnz_d,0,nnz_o);CHKERRQ(ierr); 2113 /* MatAXPY_BasicWithPreallocation() for BAIJ matrix is much slower than AIJ, even for bs=1 ! */ 2114 ierr = MatAXPY_BasicWithPreallocation(B,Y,a,X,str);CHKERRQ(ierr); 2115 ierr = MatHeaderReplace(Y,B);CHKERRQ(ierr); 2116 ierr = PetscFree(nnz_d);CHKERRQ(ierr); 2117 ierr = PetscFree(nnz_o);CHKERRQ(ierr); 2118 } 2119 PetscFunctionReturn(0); 2120 } 2121 2122 #undef __FUNCT__ 2123 #define __FUNCT__ "MatRealPart_MPIBAIJ" 2124 PetscErrorCode MatRealPart_MPIBAIJ(Mat A) 2125 { 2126 Mat_MPIBAIJ *a = (Mat_MPIBAIJ*)A->data; 2127 PetscErrorCode ierr; 2128 2129 PetscFunctionBegin; 2130 ierr = MatRealPart(a->A);CHKERRQ(ierr); 2131 ierr = MatRealPart(a->B);CHKERRQ(ierr); 2132 PetscFunctionReturn(0); 2133 } 2134 2135 #undef __FUNCT__ 2136 #define __FUNCT__ "MatImaginaryPart_MPIBAIJ" 2137 PetscErrorCode MatImaginaryPart_MPIBAIJ(Mat A) 2138 { 2139 Mat_MPIBAIJ *a = (Mat_MPIBAIJ*)A->data; 2140 PetscErrorCode ierr; 2141 2142 PetscFunctionBegin; 2143 ierr = MatImaginaryPart(a->A);CHKERRQ(ierr); 2144 ierr = MatImaginaryPart(a->B);CHKERRQ(ierr); 2145 PetscFunctionReturn(0); 2146 } 2147 2148 #undef __FUNCT__ 2149 #define __FUNCT__ "MatGetSubMatrix_MPIBAIJ" 2150 PetscErrorCode MatGetSubMatrix_MPIBAIJ(Mat mat,IS isrow,IS iscol,MatReuse call,Mat *newmat) 2151 { 2152 PetscErrorCode ierr; 2153 IS iscol_local; 2154 PetscInt csize; 2155 2156 PetscFunctionBegin; 2157 ierr = ISGetLocalSize(iscol,&csize);CHKERRQ(ierr); 2158 if (call == MAT_REUSE_MATRIX) { 2159 ierr = PetscObjectQuery((PetscObject)*newmat,"ISAllGather",(PetscObject*)&iscol_local);CHKERRQ(ierr); 2160 if (!iscol_local) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Submatrix passed in was not used before, cannot reuse"); 2161 } else { 2162 ierr = ISAllGather(iscol,&iscol_local);CHKERRQ(ierr); 2163 } 2164 ierr = MatGetSubMatrix_MPIBAIJ_Private(mat,isrow,iscol_local,csize,call,newmat);CHKERRQ(ierr); 2165 if (call == MAT_INITIAL_MATRIX) { 2166 ierr = PetscObjectCompose((PetscObject)*newmat,"ISAllGather",(PetscObject)iscol_local);CHKERRQ(ierr); 2167 ierr = ISDestroy(&iscol_local);CHKERRQ(ierr); 2168 } 2169 PetscFunctionReturn(0); 2170 } 2171 extern PetscErrorCode MatGetSubMatrices_MPIBAIJ_local(Mat,PetscInt,const IS[],const IS[],MatReuse,PetscBool*,PetscBool*,Mat*); 2172 #undef __FUNCT__ 2173 #define __FUNCT__ "MatGetSubMatrix_MPIBAIJ_Private" 2174 /* 2175 Not great since it makes two copies of the submatrix, first an SeqBAIJ 2176 in local and then by concatenating the local matrices the end result. 2177 Writing it directly would be much like MatGetSubMatrices_MPIBAIJ(). 2178 This routine is used for BAIJ and SBAIJ matrices (unfortunate dependency). 2179 */ 2180 PetscErrorCode MatGetSubMatrix_MPIBAIJ_Private(Mat mat,IS isrow,IS iscol,PetscInt csize,MatReuse call,Mat *newmat) 2181 { 2182 PetscErrorCode ierr; 2183 PetscMPIInt rank,size; 2184 PetscInt i,m,n,rstart,row,rend,nz,*cwork,j,bs; 2185 PetscInt *ii,*jj,nlocal,*dlens,*olens,dlen,olen,jend,mglobal,ncol,nrow; 2186 Mat M,Mreuse; 2187 MatScalar *vwork,*aa; 2188 MPI_Comm comm; 2189 IS isrow_new, iscol_new; 2190 PetscBool idflag,allrows, allcols; 2191 Mat_SeqBAIJ *aij; 2192 2193 PetscFunctionBegin; 2194 ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr); 2195 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 2196 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 2197 /* The compression and expansion should be avoided. Doesn't point 2198 out errors, might change the indices, hence buggey */ 2199 ierr = ISCompressIndicesGeneral(mat->rmap->N,mat->rmap->n,mat->rmap->bs,1,&isrow,&isrow_new);CHKERRQ(ierr); 2200 ierr = ISCompressIndicesGeneral(mat->cmap->N,mat->cmap->n,mat->cmap->bs,1,&iscol,&iscol_new);CHKERRQ(ierr); 2201 2202 /* Check for special case: each processor gets entire matrix columns */ 2203 ierr = ISIdentity(iscol,&idflag);CHKERRQ(ierr); 2204 ierr = ISGetLocalSize(iscol,&ncol);CHKERRQ(ierr); 2205 if (idflag && ncol == mat->cmap->N) allcols = PETSC_TRUE; 2206 else allcols = PETSC_FALSE; 2207 2208 ierr = ISIdentity(isrow,&idflag);CHKERRQ(ierr); 2209 ierr = ISGetLocalSize(isrow,&nrow);CHKERRQ(ierr); 2210 if (idflag && nrow == mat->rmap->N) allrows = PETSC_TRUE; 2211 else allrows = PETSC_FALSE; 2212 2213 if (call == MAT_REUSE_MATRIX) { 2214 ierr = PetscObjectQuery((PetscObject)*newmat,"SubMatrix",(PetscObject*)&Mreuse);CHKERRQ(ierr); 2215 if (!Mreuse) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Submatrix passed in was not used before, cannot reuse"); 2216 ierr = MatGetSubMatrices_MPIBAIJ_local(mat,1,&isrow_new,&iscol_new,MAT_REUSE_MATRIX,&allrows,&allcols,&Mreuse);CHKERRQ(ierr); 2217 } else { 2218 ierr = MatGetSubMatrices_MPIBAIJ_local(mat,1,&isrow_new,&iscol_new,MAT_INITIAL_MATRIX,&allrows,&allcols,&Mreuse);CHKERRQ(ierr); 2219 } 2220 ierr = ISDestroy(&isrow_new);CHKERRQ(ierr); 2221 ierr = ISDestroy(&iscol_new);CHKERRQ(ierr); 2222 /* 2223 m - number of local rows 2224 n - number of columns (same on all processors) 2225 rstart - first row in new global matrix generated 2226 */ 2227 ierr = MatGetBlockSize(mat,&bs);CHKERRQ(ierr); 2228 ierr = MatGetSize(Mreuse,&m,&n);CHKERRQ(ierr); 2229 m = m/bs; 2230 n = n/bs; 2231 2232 if (call == MAT_INITIAL_MATRIX) { 2233 aij = (Mat_SeqBAIJ*)(Mreuse)->data; 2234 ii = aij->i; 2235 jj = aij->j; 2236 2237 /* 2238 Determine the number of non-zeros in the diagonal and off-diagonal 2239 portions of the matrix in order to do correct preallocation 2240 */ 2241 2242 /* first get start and end of "diagonal" columns */ 2243 if (csize == PETSC_DECIDE) { 2244 ierr = ISGetSize(isrow,&mglobal);CHKERRQ(ierr); 2245 if (mglobal == n*bs) { /* square matrix */ 2246 nlocal = m; 2247 } else { 2248 nlocal = n/size + ((n % size) > rank); 2249 } 2250 } else { 2251 nlocal = csize/bs; 2252 } 2253 ierr = MPI_Scan(&nlocal,&rend,1,MPIU_INT,MPI_SUM,comm);CHKERRQ(ierr); 2254 rstart = rend - nlocal; 2255 if (rank == size - 1 && rend != n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Local column sizes %D do not add up to total number of columns %D",rend,n); 2256 2257 /* next, compute all the lengths */ 2258 ierr = PetscMalloc2(m+1,&dlens,m+1,&olens);CHKERRQ(ierr); 2259 for (i=0; i<m; i++) { 2260 jend = ii[i+1] - ii[i]; 2261 olen = 0; 2262 dlen = 0; 2263 for (j=0; j<jend; j++) { 2264 if (*jj < rstart || *jj >= rend) olen++; 2265 else dlen++; 2266 jj++; 2267 } 2268 olens[i] = olen; 2269 dlens[i] = dlen; 2270 } 2271 ierr = MatCreate(comm,&M);CHKERRQ(ierr); 2272 ierr = MatSetSizes(M,bs*m,bs*nlocal,PETSC_DECIDE,bs*n);CHKERRQ(ierr); 2273 ierr = MatSetType(M,((PetscObject)mat)->type_name);CHKERRQ(ierr); 2274 ierr = MatMPIBAIJSetPreallocation(M,bs,0,dlens,0,olens);CHKERRQ(ierr); 2275 ierr = MatMPISBAIJSetPreallocation(M,bs,0,dlens,0,olens);CHKERRQ(ierr); 2276 ierr = PetscFree2(dlens,olens);CHKERRQ(ierr); 2277 } else { 2278 PetscInt ml,nl; 2279 2280 M = *newmat; 2281 ierr = MatGetLocalSize(M,&ml,&nl);CHKERRQ(ierr); 2282 if (ml != m) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Previous matrix must be same size/layout as request"); 2283 ierr = MatZeroEntries(M);CHKERRQ(ierr); 2284 /* 2285 The next two lines are needed so we may call MatSetValues_MPIAIJ() below directly, 2286 rather than the slower MatSetValues(). 2287 */ 2288 M->was_assembled = PETSC_TRUE; 2289 M->assembled = PETSC_FALSE; 2290 } 2291 ierr = MatSetOption(M,MAT_ROW_ORIENTED,PETSC_FALSE);CHKERRQ(ierr); 2292 ierr = MatGetOwnershipRange(M,&rstart,&rend);CHKERRQ(ierr); 2293 aij = (Mat_SeqBAIJ*)(Mreuse)->data; 2294 ii = aij->i; 2295 jj = aij->j; 2296 aa = aij->a; 2297 for (i=0; i<m; i++) { 2298 row = rstart/bs + i; 2299 nz = ii[i+1] - ii[i]; 2300 cwork = jj; jj += nz; 2301 vwork = aa; aa += nz*bs*bs; 2302 ierr = MatSetValuesBlocked_MPIBAIJ(M,1,&row,nz,cwork,vwork,INSERT_VALUES);CHKERRQ(ierr); 2303 } 2304 2305 ierr = MatAssemblyBegin(M,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2306 ierr = MatAssemblyEnd(M,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2307 *newmat = M; 2308 2309 /* save submatrix used in processor for next request */ 2310 if (call == MAT_INITIAL_MATRIX) { 2311 ierr = PetscObjectCompose((PetscObject)M,"SubMatrix",(PetscObject)Mreuse);CHKERRQ(ierr); 2312 ierr = PetscObjectDereference((PetscObject)Mreuse);CHKERRQ(ierr); 2313 } 2314 PetscFunctionReturn(0); 2315 } 2316 2317 #undef __FUNCT__ 2318 #define __FUNCT__ "MatPermute_MPIBAIJ" 2319 PetscErrorCode MatPermute_MPIBAIJ(Mat A,IS rowp,IS colp,Mat *B) 2320 { 2321 MPI_Comm comm,pcomm; 2322 PetscInt clocal_size,nrows; 2323 const PetscInt *rows; 2324 PetscMPIInt size; 2325 IS crowp,lcolp; 2326 PetscErrorCode ierr; 2327 2328 PetscFunctionBegin; 2329 ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr); 2330 /* make a collective version of 'rowp' */ 2331 ierr = PetscObjectGetComm((PetscObject)rowp,&pcomm);CHKERRQ(ierr); 2332 if (pcomm==comm) { 2333 crowp = rowp; 2334 } else { 2335 ierr = ISGetSize(rowp,&nrows);CHKERRQ(ierr); 2336 ierr = ISGetIndices(rowp,&rows);CHKERRQ(ierr); 2337 ierr = ISCreateGeneral(comm,nrows,rows,PETSC_COPY_VALUES,&crowp);CHKERRQ(ierr); 2338 ierr = ISRestoreIndices(rowp,&rows);CHKERRQ(ierr); 2339 } 2340 ierr = ISSetPermutation(crowp);CHKERRQ(ierr); 2341 /* make a local version of 'colp' */ 2342 ierr = PetscObjectGetComm((PetscObject)colp,&pcomm);CHKERRQ(ierr); 2343 ierr = MPI_Comm_size(pcomm,&size);CHKERRQ(ierr); 2344 if (size==1) { 2345 lcolp = colp; 2346 } else { 2347 ierr = ISAllGather(colp,&lcolp);CHKERRQ(ierr); 2348 } 2349 ierr = ISSetPermutation(lcolp);CHKERRQ(ierr); 2350 /* now we just get the submatrix */ 2351 ierr = MatGetLocalSize(A,NULL,&clocal_size);CHKERRQ(ierr); 2352 ierr = MatGetSubMatrix_MPIBAIJ_Private(A,crowp,lcolp,clocal_size,MAT_INITIAL_MATRIX,B);CHKERRQ(ierr); 2353 /* clean up */ 2354 if (pcomm!=comm) { 2355 ierr = ISDestroy(&crowp);CHKERRQ(ierr); 2356 } 2357 if (size>1) { 2358 ierr = ISDestroy(&lcolp);CHKERRQ(ierr); 2359 } 2360 PetscFunctionReturn(0); 2361 } 2362 2363 #undef __FUNCT__ 2364 #define __FUNCT__ "MatGetGhosts_MPIBAIJ" 2365 PetscErrorCode MatGetGhosts_MPIBAIJ(Mat mat,PetscInt *nghosts,const PetscInt *ghosts[]) 2366 { 2367 Mat_MPIBAIJ *baij = (Mat_MPIBAIJ*) mat->data; 2368 Mat_SeqBAIJ *B = (Mat_SeqBAIJ*)baij->B->data; 2369 2370 PetscFunctionBegin; 2371 if (nghosts) *nghosts = B->nbs; 2372 if (ghosts) *ghosts = baij->garray; 2373 PetscFunctionReturn(0); 2374 } 2375 2376 #undef __FUNCT__ 2377 #define __FUNCT__ "MatGetSeqNonzeroStructure_MPIBAIJ" 2378 PetscErrorCode MatGetSeqNonzeroStructure_MPIBAIJ(Mat A,Mat *newmat) 2379 { 2380 Mat B; 2381 Mat_MPIBAIJ *a = (Mat_MPIBAIJ*)A->data; 2382 Mat_SeqBAIJ *ad = (Mat_SeqBAIJ*)a->A->data,*bd = (Mat_SeqBAIJ*)a->B->data; 2383 Mat_SeqAIJ *b; 2384 PetscErrorCode ierr; 2385 PetscMPIInt size,rank,*recvcounts = 0,*displs = 0; 2386 PetscInt sendcount,i,*rstarts = A->rmap->range,n,cnt,j,bs = A->rmap->bs; 2387 PetscInt m,*garray = a->garray,*lens,*jsendbuf,*a_jsendbuf,*b_jsendbuf; 2388 2389 PetscFunctionBegin; 2390 ierr = MPI_Comm_size(PetscObjectComm((PetscObject)A),&size);CHKERRQ(ierr); 2391 ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)A),&rank);CHKERRQ(ierr); 2392 2393 /* ---------------------------------------------------------------- 2394 Tell every processor the number of nonzeros per row 2395 */ 2396 ierr = PetscMalloc1(A->rmap->N/bs,&lens);CHKERRQ(ierr); 2397 for (i=A->rmap->rstart/bs; i<A->rmap->rend/bs; i++) { 2398 lens[i] = ad->i[i-A->rmap->rstart/bs+1] - ad->i[i-A->rmap->rstart/bs] + bd->i[i-A->rmap->rstart/bs+1] - bd->i[i-A->rmap->rstart/bs]; 2399 } 2400 sendcount = A->rmap->rend/bs - A->rmap->rstart/bs; 2401 ierr = PetscMalloc1(2*size,&recvcounts);CHKERRQ(ierr); 2402 displs = recvcounts + size; 2403 for (i=0; i<size; i++) { 2404 recvcounts[i] = A->rmap->range[i+1]/bs - A->rmap->range[i]/bs; 2405 displs[i] = A->rmap->range[i]/bs; 2406 } 2407 #if defined(PETSC_HAVE_MPI_IN_PLACE) 2408 ierr = MPI_Allgatherv(MPI_IN_PLACE,0,MPI_DATATYPE_NULL,lens,recvcounts,displs,MPIU_INT,PetscObjectComm((PetscObject)A));CHKERRQ(ierr); 2409 #else 2410 ierr = MPI_Allgatherv(lens+A->rmap->rstart/bs,sendcount,MPIU_INT,lens,recvcounts,displs,MPIU_INT,PetscObjectComm((PetscObject)A));CHKERRQ(ierr); 2411 #endif 2412 /* --------------------------------------------------------------- 2413 Create the sequential matrix of the same type as the local block diagonal 2414 */ 2415 ierr = MatCreate(PETSC_COMM_SELF,&B);CHKERRQ(ierr); 2416 ierr = MatSetSizes(B,A->rmap->N/bs,A->cmap->N/bs,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 2417 ierr = MatSetType(B,MATSEQAIJ);CHKERRQ(ierr); 2418 ierr = MatSeqAIJSetPreallocation(B,0,lens);CHKERRQ(ierr); 2419 b = (Mat_SeqAIJ*)B->data; 2420 2421 /*-------------------------------------------------------------------- 2422 Copy my part of matrix column indices over 2423 */ 2424 sendcount = ad->nz + bd->nz; 2425 jsendbuf = b->j + b->i[rstarts[rank]/bs]; 2426 a_jsendbuf = ad->j; 2427 b_jsendbuf = bd->j; 2428 n = A->rmap->rend/bs - A->rmap->rstart/bs; 2429 cnt = 0; 2430 for (i=0; i<n; i++) { 2431 2432 /* put in lower diagonal portion */ 2433 m = bd->i[i+1] - bd->i[i]; 2434 while (m > 0) { 2435 /* is it above diagonal (in bd (compressed) numbering) */ 2436 if (garray[*b_jsendbuf] > A->rmap->rstart/bs + i) break; 2437 jsendbuf[cnt++] = garray[*b_jsendbuf++]; 2438 m--; 2439 } 2440 2441 /* put in diagonal portion */ 2442 for (j=ad->i[i]; j<ad->i[i+1]; j++) { 2443 jsendbuf[cnt++] = A->rmap->rstart/bs + *a_jsendbuf++; 2444 } 2445 2446 /* put in upper diagonal portion */ 2447 while (m-- > 0) { 2448 jsendbuf[cnt++] = garray[*b_jsendbuf++]; 2449 } 2450 } 2451 if (cnt != sendcount) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Corrupted PETSc matrix: nz given %D actual nz %D",sendcount,cnt); 2452 2453 /*-------------------------------------------------------------------- 2454 Gather all column indices to all processors 2455 */ 2456 for (i=0; i<size; i++) { 2457 recvcounts[i] = 0; 2458 for (j=A->rmap->range[i]/bs; j<A->rmap->range[i+1]/bs; j++) { 2459 recvcounts[i] += lens[j]; 2460 } 2461 } 2462 displs[0] = 0; 2463 for (i=1; i<size; i++) { 2464 displs[i] = displs[i-1] + recvcounts[i-1]; 2465 } 2466 #if defined(PETSC_HAVE_MPI_IN_PLACE) 2467 ierr = MPI_Allgatherv(MPI_IN_PLACE,0,MPI_DATATYPE_NULL,b->j,recvcounts,displs,MPIU_INT,PetscObjectComm((PetscObject)A));CHKERRQ(ierr); 2468 #else 2469 ierr = MPI_Allgatherv(jsendbuf,sendcount,MPIU_INT,b->j,recvcounts,displs,MPIU_INT,PetscObjectComm((PetscObject)A));CHKERRQ(ierr); 2470 #endif 2471 /*-------------------------------------------------------------------- 2472 Assemble the matrix into useable form (note numerical values not yet set) 2473 */ 2474 /* set the b->ilen (length of each row) values */ 2475 ierr = PetscMemcpy(b->ilen,lens,(A->rmap->N/bs)*sizeof(PetscInt));CHKERRQ(ierr); 2476 /* set the b->i indices */ 2477 b->i[0] = 0; 2478 for (i=1; i<=A->rmap->N/bs; i++) { 2479 b->i[i] = b->i[i-1] + lens[i-1]; 2480 } 2481 ierr = PetscFree(lens);CHKERRQ(ierr); 2482 ierr = MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2483 ierr = MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2484 ierr = PetscFree(recvcounts);CHKERRQ(ierr); 2485 2486 if (A->symmetric) { 2487 ierr = MatSetOption(B,MAT_SYMMETRIC,PETSC_TRUE);CHKERRQ(ierr); 2488 } else if (A->hermitian) { 2489 ierr = MatSetOption(B,MAT_HERMITIAN,PETSC_TRUE);CHKERRQ(ierr); 2490 } else if (A->structurally_symmetric) { 2491 ierr = MatSetOption(B,MAT_STRUCTURALLY_SYMMETRIC,PETSC_TRUE);CHKERRQ(ierr); 2492 } 2493 *newmat = B; 2494 PetscFunctionReturn(0); 2495 } 2496 2497 #undef __FUNCT__ 2498 #define __FUNCT__ "MatSOR_MPIBAIJ" 2499 PetscErrorCode MatSOR_MPIBAIJ(Mat matin,Vec bb,PetscReal omega,MatSORType flag,PetscReal fshift,PetscInt its,PetscInt lits,Vec xx) 2500 { 2501 Mat_MPIBAIJ *mat = (Mat_MPIBAIJ*)matin->data; 2502 PetscErrorCode ierr; 2503 Vec bb1 = 0; 2504 2505 PetscFunctionBegin; 2506 if (flag == SOR_APPLY_UPPER) { 2507 ierr = (*mat->A->ops->sor)(mat->A,bb,omega,flag,fshift,lits,1,xx);CHKERRQ(ierr); 2508 PetscFunctionReturn(0); 2509 } 2510 2511 if (its > 1 || ~flag & SOR_ZERO_INITIAL_GUESS) { 2512 ierr = VecDuplicate(bb,&bb1);CHKERRQ(ierr); 2513 } 2514 2515 if ((flag & SOR_LOCAL_SYMMETRIC_SWEEP) == SOR_LOCAL_SYMMETRIC_SWEEP) { 2516 if (flag & SOR_ZERO_INITIAL_GUESS) { 2517 ierr = (*mat->A->ops->sor)(mat->A,bb,omega,flag,fshift,lits,1,xx);CHKERRQ(ierr); 2518 its--; 2519 } 2520 2521 while (its--) { 2522 ierr = VecScatterBegin(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2523 ierr = VecScatterEnd(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2524 2525 /* update rhs: bb1 = bb - B*x */ 2526 ierr = VecScale(mat->lvec,-1.0);CHKERRQ(ierr); 2527 ierr = (*mat->B->ops->multadd)(mat->B,mat->lvec,bb,bb1);CHKERRQ(ierr); 2528 2529 /* local sweep */ 2530 ierr = (*mat->A->ops->sor)(mat->A,bb1,omega,SOR_SYMMETRIC_SWEEP,fshift,lits,1,xx);CHKERRQ(ierr); 2531 } 2532 } else if (flag & SOR_LOCAL_FORWARD_SWEEP) { 2533 if (flag & SOR_ZERO_INITIAL_GUESS) { 2534 ierr = (*mat->A->ops->sor)(mat->A,bb,omega,flag,fshift,lits,1,xx);CHKERRQ(ierr); 2535 its--; 2536 } 2537 while (its--) { 2538 ierr = VecScatterBegin(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2539 ierr = VecScatterEnd(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2540 2541 /* update rhs: bb1 = bb - B*x */ 2542 ierr = VecScale(mat->lvec,-1.0);CHKERRQ(ierr); 2543 ierr = (*mat->B->ops->multadd)(mat->B,mat->lvec,bb,bb1);CHKERRQ(ierr); 2544 2545 /* local sweep */ 2546 ierr = (*mat->A->ops->sor)(mat->A,bb1,omega,SOR_FORWARD_SWEEP,fshift,lits,1,xx);CHKERRQ(ierr); 2547 } 2548 } else if (flag & SOR_LOCAL_BACKWARD_SWEEP) { 2549 if (flag & SOR_ZERO_INITIAL_GUESS) { 2550 ierr = (*mat->A->ops->sor)(mat->A,bb,omega,flag,fshift,lits,1,xx);CHKERRQ(ierr); 2551 its--; 2552 } 2553 while (its--) { 2554 ierr = VecScatterBegin(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2555 ierr = VecScatterEnd(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2556 2557 /* update rhs: bb1 = bb - B*x */ 2558 ierr = VecScale(mat->lvec,-1.0);CHKERRQ(ierr); 2559 ierr = (*mat->B->ops->multadd)(mat->B,mat->lvec,bb,bb1);CHKERRQ(ierr); 2560 2561 /* local sweep */ 2562 ierr = (*mat->A->ops->sor)(mat->A,bb1,omega,SOR_BACKWARD_SWEEP,fshift,lits,1,xx);CHKERRQ(ierr); 2563 } 2564 } else SETERRQ(PetscObjectComm((PetscObject)matin),PETSC_ERR_SUP,"Parallel version of SOR requested not supported"); 2565 2566 ierr = VecDestroy(&bb1);CHKERRQ(ierr); 2567 PetscFunctionReturn(0); 2568 } 2569 2570 #undef __FUNCT__ 2571 #define __FUNCT__ "MatGetColumnNorms_MPIBAIJ" 2572 PetscErrorCode MatGetColumnNorms_MPIBAIJ(Mat A,NormType type,PetscReal *norms) 2573 { 2574 PetscErrorCode ierr; 2575 Mat_MPIBAIJ *aij = (Mat_MPIBAIJ*)A->data; 2576 PetscInt N,i,*garray = aij->garray; 2577 PetscInt ib,jb,bs = A->rmap->bs; 2578 Mat_SeqBAIJ *a_aij = (Mat_SeqBAIJ*) aij->A->data; 2579 MatScalar *a_val = a_aij->a; 2580 Mat_SeqBAIJ *b_aij = (Mat_SeqBAIJ*) aij->B->data; 2581 MatScalar *b_val = b_aij->a; 2582 PetscReal *work; 2583 2584 PetscFunctionBegin; 2585 ierr = MatGetSize(A,NULL,&N);CHKERRQ(ierr); 2586 ierr = PetscCalloc1(N,&work);CHKERRQ(ierr); 2587 if (type == NORM_2) { 2588 for (i=a_aij->i[0]; i<a_aij->i[aij->A->rmap->n/bs]; i++) { 2589 for (jb=0; jb<bs; jb++) { 2590 for (ib=0; ib<bs; ib++) { 2591 work[A->cmap->rstart + a_aij->j[i] * bs + jb] += PetscAbsScalar(*a_val * *a_val); 2592 a_val++; 2593 } 2594 } 2595 } 2596 for (i=b_aij->i[0]; i<b_aij->i[aij->B->rmap->n/bs]; i++) { 2597 for (jb=0; jb<bs; jb++) { 2598 for (ib=0; ib<bs; ib++) { 2599 work[garray[b_aij->j[i]] * bs + jb] += PetscAbsScalar(*b_val * *b_val); 2600 b_val++; 2601 } 2602 } 2603 } 2604 } else if (type == NORM_1) { 2605 for (i=a_aij->i[0]; i<a_aij->i[aij->A->rmap->n/bs]; i++) { 2606 for (jb=0; jb<bs; jb++) { 2607 for (ib=0; ib<bs; ib++) { 2608 work[A->cmap->rstart + a_aij->j[i] * bs + jb] += PetscAbsScalar(*a_val); 2609 a_val++; 2610 } 2611 } 2612 } 2613 for (i=b_aij->i[0]; i<b_aij->i[aij->B->rmap->n/bs]; i++) { 2614 for (jb=0; jb<bs; jb++) { 2615 for (ib=0; ib<bs; ib++) { 2616 work[garray[b_aij->j[i]] * bs + jb] += PetscAbsScalar(*b_val); 2617 b_val++; 2618 } 2619 } 2620 } 2621 } else if (type == NORM_INFINITY) { 2622 for (i=a_aij->i[0]; i<a_aij->i[aij->A->rmap->n/bs]; i++) { 2623 for (jb=0; jb<bs; jb++) { 2624 for (ib=0; ib<bs; ib++) { 2625 int col = A->cmap->rstart + a_aij->j[i] * bs + jb; 2626 work[col] = PetscMax(PetscAbsScalar(*a_val), work[col]); 2627 a_val++; 2628 } 2629 } 2630 } 2631 for (i=b_aij->i[0]; i<b_aij->i[aij->B->rmap->n/bs]; i++) { 2632 for (jb=0; jb<bs; jb++) { 2633 for (ib=0; ib<bs; ib++) { 2634 int col = garray[b_aij->j[i]] * bs + jb; 2635 work[col] = PetscMax(PetscAbsScalar(*b_val), work[col]); 2636 b_val++; 2637 } 2638 } 2639 } 2640 } else SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_ARG_WRONG,"Unknown NormType"); 2641 if (type == NORM_INFINITY) { 2642 ierr = MPI_Allreduce(work,norms,N,MPIU_REAL,MPIU_MAX,PetscObjectComm((PetscObject)A));CHKERRQ(ierr); 2643 } else { 2644 ierr = MPI_Allreduce(work,norms,N,MPIU_REAL,MPIU_SUM,PetscObjectComm((PetscObject)A));CHKERRQ(ierr); 2645 } 2646 ierr = PetscFree(work);CHKERRQ(ierr); 2647 if (type == NORM_2) { 2648 for (i=0; i<N; i++) norms[i] = PetscSqrtReal(norms[i]); 2649 } 2650 PetscFunctionReturn(0); 2651 } 2652 2653 #undef __FUNCT__ 2654 #define __FUNCT__ "MatInvertBlockDiagonal_MPIBAIJ" 2655 PetscErrorCode MatInvertBlockDiagonal_MPIBAIJ(Mat A,const PetscScalar **values) 2656 { 2657 Mat_MPIBAIJ *a = (Mat_MPIBAIJ*) A->data; 2658 PetscErrorCode ierr; 2659 2660 PetscFunctionBegin; 2661 ierr = MatInvertBlockDiagonal(a->A,values);CHKERRQ(ierr); 2662 PetscFunctionReturn(0); 2663 } 2664 2665 #undef __FUNCT__ 2666 #define __FUNCT__ "MatShift_MPIBAIJ" 2667 PetscErrorCode MatShift_MPIBAIJ(Mat Y,PetscScalar a) 2668 { 2669 PetscErrorCode ierr; 2670 Mat_MPIBAIJ *maij = (Mat_MPIBAIJ*)Y->data; 2671 Mat_SeqBAIJ *aij = (Mat_SeqBAIJ*)maij->A->data; 2672 2673 PetscFunctionBegin; 2674 if (!Y->preallocated) { 2675 ierr = MatMPIBAIJSetPreallocation(Y,Y->rmap->bs,1,NULL,0,NULL);CHKERRQ(ierr); 2676 } else if (!aij->nz) { 2677 ierr = MatSeqBAIJSetPreallocation(maij->A,Y->rmap->bs,1,NULL);CHKERRQ(ierr); 2678 } 2679 ierr = MatShift_Basic(Y,a);CHKERRQ(ierr); 2680 PetscFunctionReturn(0); 2681 } 2682 2683 /* -------------------------------------------------------------------*/ 2684 static struct _MatOps MatOps_Values = {MatSetValues_MPIBAIJ, 2685 MatGetRow_MPIBAIJ, 2686 MatRestoreRow_MPIBAIJ, 2687 MatMult_MPIBAIJ, 2688 /* 4*/ MatMultAdd_MPIBAIJ, 2689 MatMultTranspose_MPIBAIJ, 2690 MatMultTransposeAdd_MPIBAIJ, 2691 0, 2692 0, 2693 0, 2694 /*10*/ 0, 2695 0, 2696 0, 2697 MatSOR_MPIBAIJ, 2698 MatTranspose_MPIBAIJ, 2699 /*15*/ MatGetInfo_MPIBAIJ, 2700 MatEqual_MPIBAIJ, 2701 MatGetDiagonal_MPIBAIJ, 2702 MatDiagonalScale_MPIBAIJ, 2703 MatNorm_MPIBAIJ, 2704 /*20*/ MatAssemblyBegin_MPIBAIJ, 2705 MatAssemblyEnd_MPIBAIJ, 2706 MatSetOption_MPIBAIJ, 2707 MatZeroEntries_MPIBAIJ, 2708 /*24*/ MatZeroRows_MPIBAIJ, 2709 0, 2710 0, 2711 0, 2712 0, 2713 /*29*/ MatSetUp_MPIBAIJ, 2714 0, 2715 0, 2716 0, 2717 0, 2718 /*34*/ MatDuplicate_MPIBAIJ, 2719 0, 2720 0, 2721 0, 2722 0, 2723 /*39*/ MatAXPY_MPIBAIJ, 2724 MatGetSubMatrices_MPIBAIJ, 2725 MatIncreaseOverlap_MPIBAIJ, 2726 MatGetValues_MPIBAIJ, 2727 MatCopy_MPIBAIJ, 2728 /*44*/ 0, 2729 MatScale_MPIBAIJ, 2730 MatShift_MPIBAIJ, 2731 0, 2732 MatZeroRowsColumns_MPIBAIJ, 2733 /*49*/ 0, 2734 0, 2735 0, 2736 0, 2737 0, 2738 /*54*/ MatFDColoringCreate_MPIXAIJ, 2739 0, 2740 MatSetUnfactored_MPIBAIJ, 2741 MatPermute_MPIBAIJ, 2742 MatSetValuesBlocked_MPIBAIJ, 2743 /*59*/ MatGetSubMatrix_MPIBAIJ, 2744 MatDestroy_MPIBAIJ, 2745 MatView_MPIBAIJ, 2746 0, 2747 0, 2748 /*64*/ 0, 2749 0, 2750 0, 2751 0, 2752 0, 2753 /*69*/ MatGetRowMaxAbs_MPIBAIJ, 2754 0, 2755 0, 2756 0, 2757 0, 2758 /*74*/ 0, 2759 MatFDColoringApply_BAIJ, 2760 0, 2761 0, 2762 0, 2763 /*79*/ 0, 2764 0, 2765 0, 2766 0, 2767 MatLoad_MPIBAIJ, 2768 /*84*/ 0, 2769 0, 2770 0, 2771 0, 2772 0, 2773 /*89*/ 0, 2774 0, 2775 0, 2776 0, 2777 0, 2778 /*94*/ 0, 2779 0, 2780 0, 2781 0, 2782 0, 2783 /*99*/ 0, 2784 0, 2785 0, 2786 0, 2787 0, 2788 /*104*/0, 2789 MatRealPart_MPIBAIJ, 2790 MatImaginaryPart_MPIBAIJ, 2791 0, 2792 0, 2793 /*109*/0, 2794 0, 2795 0, 2796 0, 2797 0, 2798 /*114*/MatGetSeqNonzeroStructure_MPIBAIJ, 2799 0, 2800 MatGetGhosts_MPIBAIJ, 2801 0, 2802 0, 2803 /*119*/0, 2804 0, 2805 0, 2806 0, 2807 MatGetMultiProcBlock_MPIBAIJ, 2808 /*124*/0, 2809 MatGetColumnNorms_MPIBAIJ, 2810 MatInvertBlockDiagonal_MPIBAIJ, 2811 0, 2812 0, 2813 /*129*/ 0, 2814 0, 2815 0, 2816 0, 2817 0, 2818 /*134*/ 0, 2819 0, 2820 0, 2821 0, 2822 0, 2823 /*139*/ 0, 2824 0, 2825 0, 2826 MatFDColoringSetUp_MPIXAIJ, 2827 0, 2828 /*144*/MatCreateMPIMatConcatenateSeqMat_MPIBAIJ 2829 }; 2830 2831 #undef __FUNCT__ 2832 #define __FUNCT__ "MatGetDiagonalBlock_MPIBAIJ" 2833 PetscErrorCode MatGetDiagonalBlock_MPIBAIJ(Mat A,Mat *a) 2834 { 2835 PetscFunctionBegin; 2836 *a = ((Mat_MPIBAIJ*)A->data)->A; 2837 PetscFunctionReturn(0); 2838 } 2839 2840 PETSC_EXTERN PetscErrorCode MatConvert_MPIBAIJ_MPISBAIJ(Mat, MatType,MatReuse,Mat*); 2841 2842 #undef __FUNCT__ 2843 #define __FUNCT__ "MatMPIBAIJSetPreallocationCSR_MPIBAIJ" 2844 PetscErrorCode MatMPIBAIJSetPreallocationCSR_MPIBAIJ(Mat B,PetscInt bs,const PetscInt ii[],const PetscInt jj[],const PetscScalar V[]) 2845 { 2846 PetscInt m,rstart,cstart,cend; 2847 PetscInt i,j,d,nz,nz_max=0,*d_nnz=0,*o_nnz=0; 2848 const PetscInt *JJ =0; 2849 PetscScalar *values=0; 2850 PetscBool roworiented = ((Mat_MPIBAIJ*)B->data)->roworiented; 2851 PetscErrorCode ierr; 2852 2853 PetscFunctionBegin; 2854 ierr = PetscLayoutSetBlockSize(B->rmap,bs);CHKERRQ(ierr); 2855 ierr = PetscLayoutSetBlockSize(B->cmap,bs);CHKERRQ(ierr); 2856 ierr = PetscLayoutSetUp(B->rmap);CHKERRQ(ierr); 2857 ierr = PetscLayoutSetUp(B->cmap);CHKERRQ(ierr); 2858 ierr = PetscLayoutGetBlockSize(B->rmap,&bs);CHKERRQ(ierr); 2859 m = B->rmap->n/bs; 2860 rstart = B->rmap->rstart/bs; 2861 cstart = B->cmap->rstart/bs; 2862 cend = B->cmap->rend/bs; 2863 2864 if (ii[0]) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"ii[0] must be 0 but it is %D",ii[0]); 2865 ierr = PetscMalloc2(m,&d_nnz,m,&o_nnz);CHKERRQ(ierr); 2866 for (i=0; i<m; i++) { 2867 nz = ii[i+1] - ii[i]; 2868 if (nz < 0) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Local row %D has a negative number of columns %D",i,nz); 2869 nz_max = PetscMax(nz_max,nz); 2870 JJ = jj + ii[i]; 2871 for (j=0; j<nz; j++) { 2872 if (*JJ >= cstart) break; 2873 JJ++; 2874 } 2875 d = 0; 2876 for (; j<nz; j++) { 2877 if (*JJ++ >= cend) break; 2878 d++; 2879 } 2880 d_nnz[i] = d; 2881 o_nnz[i] = nz - d; 2882 } 2883 ierr = MatMPIBAIJSetPreallocation(B,bs,0,d_nnz,0,o_nnz);CHKERRQ(ierr); 2884 ierr = PetscFree2(d_nnz,o_nnz);CHKERRQ(ierr); 2885 2886 values = (PetscScalar*)V; 2887 if (!values) { 2888 ierr = PetscMalloc1(bs*bs*nz_max,&values);CHKERRQ(ierr); 2889 ierr = PetscMemzero(values,bs*bs*nz_max*sizeof(PetscScalar));CHKERRQ(ierr); 2890 } 2891 for (i=0; i<m; i++) { 2892 PetscInt row = i + rstart; 2893 PetscInt ncols = ii[i+1] - ii[i]; 2894 const PetscInt *icols = jj + ii[i]; 2895 if (!roworiented) { /* block ordering matches the non-nested layout of MatSetValues so we can insert entire rows */ 2896 const PetscScalar *svals = values + (V ? (bs*bs*ii[i]) : 0); 2897 ierr = MatSetValuesBlocked_MPIBAIJ(B,1,&row,ncols,icols,svals,INSERT_VALUES);CHKERRQ(ierr); 2898 } else { /* block ordering does not match so we can only insert one block at a time. */ 2899 PetscInt j; 2900 for (j=0; j<ncols; j++) { 2901 const PetscScalar *svals = values + (V ? (bs*bs*(ii[i]+j)) : 0); 2902 ierr = MatSetValuesBlocked_MPIBAIJ(B,1,&row,1,&icols[j],svals,INSERT_VALUES);CHKERRQ(ierr); 2903 } 2904 } 2905 } 2906 2907 if (!V) { ierr = PetscFree(values);CHKERRQ(ierr); } 2908 ierr = MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2909 ierr = MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2910 ierr = MatSetOption(B,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_TRUE);CHKERRQ(ierr); 2911 PetscFunctionReturn(0); 2912 } 2913 2914 #undef __FUNCT__ 2915 #define __FUNCT__ "MatMPIBAIJSetPreallocationCSR" 2916 /*@C 2917 MatMPIBAIJSetPreallocationCSR - Allocates memory for a sparse parallel matrix in BAIJ format 2918 (the default parallel PETSc format). 2919 2920 Collective on MPI_Comm 2921 2922 Input Parameters: 2923 + B - the matrix 2924 . bs - the block size 2925 . i - the indices into j for the start of each local row (starts with zero) 2926 . j - the column indices for each local row (starts with zero) these must be sorted for each row 2927 - v - optional values in the matrix 2928 2929 Level: developer 2930 2931 Notes: The order of the entries in values is specified by the MatOption MAT_ROW_ORIENTED. For example, C programs 2932 may want to use the default MAT_ROW_ORIENTED=PETSC_TRUE and use an array v[nnz][bs][bs] where the second index is 2933 over rows within a block and the last index is over columns within a block row. Fortran programs will likely set 2934 MAT_ROW_ORIENTED=PETSC_FALSE and use a Fortran array v(bs,bs,nnz) in which the first index is over rows within a 2935 block column and the second index is over columns within a block. 2936 2937 .keywords: matrix, aij, compressed row, sparse, parallel 2938 2939 .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatMPIBAIJSetPreallocation(), MatCreateAIJ(), MPIAIJ, MatCreateMPIBAIJWithArrays(), MPIBAIJ 2940 @*/ 2941 PetscErrorCode MatMPIBAIJSetPreallocationCSR(Mat B,PetscInt bs,const PetscInt i[],const PetscInt j[], const PetscScalar v[]) 2942 { 2943 PetscErrorCode ierr; 2944 2945 PetscFunctionBegin; 2946 PetscValidHeaderSpecific(B,MAT_CLASSID,1); 2947 PetscValidType(B,1); 2948 PetscValidLogicalCollectiveInt(B,bs,2); 2949 ierr = PetscTryMethod(B,"MatMPIBAIJSetPreallocationCSR_C",(Mat,PetscInt,const PetscInt[],const PetscInt[],const PetscScalar[]),(B,bs,i,j,v));CHKERRQ(ierr); 2950 PetscFunctionReturn(0); 2951 } 2952 2953 #undef __FUNCT__ 2954 #define __FUNCT__ "MatMPIBAIJSetPreallocation_MPIBAIJ" 2955 PetscErrorCode MatMPIBAIJSetPreallocation_MPIBAIJ(Mat B,PetscInt bs,PetscInt d_nz,const PetscInt *d_nnz,PetscInt o_nz,const PetscInt *o_nnz) 2956 { 2957 Mat_MPIBAIJ *b; 2958 PetscErrorCode ierr; 2959 PetscInt i; 2960 2961 PetscFunctionBegin; 2962 ierr = MatSetBlockSize(B,PetscAbs(bs));CHKERRQ(ierr); 2963 ierr = PetscLayoutSetUp(B->rmap);CHKERRQ(ierr); 2964 ierr = PetscLayoutSetUp(B->cmap);CHKERRQ(ierr); 2965 ierr = PetscLayoutGetBlockSize(B->rmap,&bs);CHKERRQ(ierr); 2966 2967 if (d_nnz) { 2968 for (i=0; i<B->rmap->n/bs; i++) { 2969 if (d_nnz[i] < 0) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"d_nnz cannot be less than -1: local row %D value %D",i,d_nnz[i]); 2970 } 2971 } 2972 if (o_nnz) { 2973 for (i=0; i<B->rmap->n/bs; i++) { 2974 if (o_nnz[i] < 0) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"o_nnz cannot be less than -1: local row %D value %D",i,o_nnz[i]); 2975 } 2976 } 2977 2978 b = (Mat_MPIBAIJ*)B->data; 2979 b->bs2 = bs*bs; 2980 b->mbs = B->rmap->n/bs; 2981 b->nbs = B->cmap->n/bs; 2982 b->Mbs = B->rmap->N/bs; 2983 b->Nbs = B->cmap->N/bs; 2984 2985 for (i=0; i<=b->size; i++) { 2986 b->rangebs[i] = B->rmap->range[i]/bs; 2987 } 2988 b->rstartbs = B->rmap->rstart/bs; 2989 b->rendbs = B->rmap->rend/bs; 2990 b->cstartbs = B->cmap->rstart/bs; 2991 b->cendbs = B->cmap->rend/bs; 2992 2993 if (!B->preallocated) { 2994 ierr = MatCreate(PETSC_COMM_SELF,&b->A);CHKERRQ(ierr); 2995 ierr = MatSetSizes(b->A,B->rmap->n,B->cmap->n,B->rmap->n,B->cmap->n);CHKERRQ(ierr); 2996 ierr = MatSetType(b->A,MATSEQBAIJ);CHKERRQ(ierr); 2997 ierr = PetscLogObjectParent((PetscObject)B,(PetscObject)b->A);CHKERRQ(ierr); 2998 ierr = MatCreate(PETSC_COMM_SELF,&b->B);CHKERRQ(ierr); 2999 ierr = MatSetSizes(b->B,B->rmap->n,B->cmap->N,B->rmap->n,B->cmap->N);CHKERRQ(ierr); 3000 ierr = MatSetType(b->B,MATSEQBAIJ);CHKERRQ(ierr); 3001 ierr = PetscLogObjectParent((PetscObject)B,(PetscObject)b->B);CHKERRQ(ierr); 3002 ierr = MatStashCreate_Private(PetscObjectComm((PetscObject)B),bs,&B->bstash);CHKERRQ(ierr); 3003 } 3004 3005 ierr = MatSeqBAIJSetPreallocation(b->A,bs,d_nz,d_nnz);CHKERRQ(ierr); 3006 ierr = MatSeqBAIJSetPreallocation(b->B,bs,o_nz,o_nnz);CHKERRQ(ierr); 3007 B->preallocated = PETSC_TRUE; 3008 PetscFunctionReturn(0); 3009 } 3010 3011 extern PetscErrorCode MatDiagonalScaleLocal_MPIBAIJ(Mat,Vec); 3012 extern PetscErrorCode MatSetHashTableFactor_MPIBAIJ(Mat,PetscReal); 3013 3014 #undef __FUNCT__ 3015 #define __FUNCT__ "MatConvert_MPIBAIJ_MPIAdj" 3016 PETSC_EXTERN PetscErrorCode MatConvert_MPIBAIJ_MPIAdj(Mat B, MatType newtype,MatReuse reuse,Mat *adj) 3017 { 3018 Mat_MPIBAIJ *b = (Mat_MPIBAIJ*)B->data; 3019 PetscErrorCode ierr; 3020 Mat_SeqBAIJ *d = (Mat_SeqBAIJ*) b->A->data,*o = (Mat_SeqBAIJ*) b->B->data; 3021 PetscInt M = B->rmap->n/B->rmap->bs,i,*ii,*jj,cnt,j,k,rstart = B->rmap->rstart/B->rmap->bs; 3022 const PetscInt *id = d->i, *jd = d->j, *io = o->i, *jo = o->j, *garray = b->garray; 3023 3024 PetscFunctionBegin; 3025 ierr = PetscMalloc1(M+1,&ii);CHKERRQ(ierr); 3026 ii[0] = 0; 3027 for (i=0; i<M; i++) { 3028 if ((id[i+1] - id[i]) < 0) SETERRQ3(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Indices wrong %D %D %D",i,id[i],id[i+1]); 3029 if ((io[i+1] - io[i]) < 0) SETERRQ3(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Indices wrong %D %D %D",i,io[i],io[i+1]); 3030 ii[i+1] = ii[i] + id[i+1] - id[i] + io[i+1] - io[i]; 3031 /* remove one from count of matrix has diagonal */ 3032 for (j=id[i]; j<id[i+1]; j++) { 3033 if (jd[j] == i) {ii[i+1]--;break;} 3034 } 3035 } 3036 ierr = PetscMalloc1(ii[M],&jj);CHKERRQ(ierr); 3037 cnt = 0; 3038 for (i=0; i<M; i++) { 3039 for (j=io[i]; j<io[i+1]; j++) { 3040 if (garray[jo[j]] > rstart) break; 3041 jj[cnt++] = garray[jo[j]]; 3042 } 3043 for (k=id[i]; k<id[i+1]; k++) { 3044 if (jd[k] != i) { 3045 jj[cnt++] = rstart + jd[k]; 3046 } 3047 } 3048 for (; j<io[i+1]; j++) { 3049 jj[cnt++] = garray[jo[j]]; 3050 } 3051 } 3052 ierr = MatCreateMPIAdj(PetscObjectComm((PetscObject)B),M,B->cmap->N/B->rmap->bs,ii,jj,NULL,adj);CHKERRQ(ierr); 3053 PetscFunctionReturn(0); 3054 } 3055 3056 #include <../src/mat/impls/aij/mpi/mpiaij.h> 3057 3058 PETSC_EXTERN PetscErrorCode MatConvert_SeqBAIJ_SeqAIJ(Mat,MatType,MatReuse,Mat*); 3059 3060 #undef __FUNCT__ 3061 #define __FUNCT__ "MatConvert_MPIBAIJ_MPIAIJ" 3062 PETSC_EXTERN PetscErrorCode MatConvert_MPIBAIJ_MPIAIJ(Mat A,MatType newtype,MatReuse reuse,Mat *newmat) 3063 { 3064 PetscErrorCode ierr; 3065 Mat_MPIBAIJ *a = (Mat_MPIBAIJ*)A->data; 3066 Mat B; 3067 Mat_MPIAIJ *b; 3068 3069 PetscFunctionBegin; 3070 if (!A->assembled) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_SUP,"Matrix must be assembled"); 3071 3072 ierr = MatCreate(PetscObjectComm((PetscObject)A),&B);CHKERRQ(ierr); 3073 ierr = MatSetType(B,MATMPIAIJ);CHKERRQ(ierr); 3074 ierr = MatSetSizes(B,A->rmap->n,A->cmap->n,A->rmap->N,A->cmap->N);CHKERRQ(ierr); 3075 ierr = MatSetBlockSizes(B,A->rmap->bs,A->cmap->bs);CHKERRQ(ierr); 3076 ierr = MatSeqAIJSetPreallocation(B,0,NULL);CHKERRQ(ierr); 3077 ierr = MatMPIAIJSetPreallocation(B,0,NULL,0,NULL);CHKERRQ(ierr); 3078 b = (Mat_MPIAIJ*) B->data; 3079 3080 ierr = MatDestroy(&b->A);CHKERRQ(ierr); 3081 ierr = MatDestroy(&b->B);CHKERRQ(ierr); 3082 ierr = MatDisAssemble_MPIBAIJ(A);CHKERRQ(ierr); 3083 ierr = MatConvert_SeqBAIJ_SeqAIJ(a->A, MATSEQAIJ, MAT_INITIAL_MATRIX, &b->A);CHKERRQ(ierr); 3084 ierr = MatConvert_SeqBAIJ_SeqAIJ(a->B, MATSEQAIJ, MAT_INITIAL_MATRIX, &b->B);CHKERRQ(ierr); 3085 ierr = MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3086 ierr = MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3087 ierr = MatAssemblyBegin(A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3088 ierr = MatAssemblyEnd(A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3089 if (reuse == MAT_REUSE_MATRIX) { 3090 ierr = MatHeaderReplace(A,B);CHKERRQ(ierr); 3091 } else { 3092 *newmat = B; 3093 } 3094 PetscFunctionReturn(0); 3095 } 3096 3097 /*MC 3098 MATMPIBAIJ - MATMPIBAIJ = "mpibaij" - A matrix type to be used for distributed block sparse matrices. 3099 3100 Options Database Keys: 3101 + -mat_type mpibaij - sets the matrix type to "mpibaij" during a call to MatSetFromOptions() 3102 . -mat_block_size <bs> - set the blocksize used to store the matrix 3103 - -mat_use_hash_table <fact> 3104 3105 Level: beginner 3106 3107 .seealso: MatCreateMPIBAIJ 3108 M*/ 3109 3110 PETSC_EXTERN PetscErrorCode MatConvert_MPIBAIJ_MPIBSTRM(Mat,MatType,MatReuse,Mat*); 3111 3112 #undef __FUNCT__ 3113 #define __FUNCT__ "MatCreate_MPIBAIJ" 3114 PETSC_EXTERN PetscErrorCode MatCreate_MPIBAIJ(Mat B) 3115 { 3116 Mat_MPIBAIJ *b; 3117 PetscErrorCode ierr; 3118 PetscBool flg = PETSC_FALSE; 3119 3120 PetscFunctionBegin; 3121 ierr = PetscNewLog(B,&b);CHKERRQ(ierr); 3122 B->data = (void*)b; 3123 3124 ierr = PetscMemcpy(B->ops,&MatOps_Values,sizeof(struct _MatOps));CHKERRQ(ierr); 3125 B->assembled = PETSC_FALSE; 3126 3127 B->insertmode = NOT_SET_VALUES; 3128 ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)B),&b->rank);CHKERRQ(ierr); 3129 ierr = MPI_Comm_size(PetscObjectComm((PetscObject)B),&b->size);CHKERRQ(ierr); 3130 3131 /* build local table of row and column ownerships */ 3132 ierr = PetscMalloc1(b->size+1,&b->rangebs);CHKERRQ(ierr); 3133 3134 /* build cache for off array entries formed */ 3135 ierr = MatStashCreate_Private(PetscObjectComm((PetscObject)B),1,&B->stash);CHKERRQ(ierr); 3136 3137 b->donotstash = PETSC_FALSE; 3138 b->colmap = NULL; 3139 b->garray = NULL; 3140 b->roworiented = PETSC_TRUE; 3141 3142 /* stuff used in block assembly */ 3143 b->barray = 0; 3144 3145 /* stuff used for matrix vector multiply */ 3146 b->lvec = 0; 3147 b->Mvctx = 0; 3148 3149 /* stuff for MatGetRow() */ 3150 b->rowindices = 0; 3151 b->rowvalues = 0; 3152 b->getrowactive = PETSC_FALSE; 3153 3154 /* hash table stuff */ 3155 b->ht = 0; 3156 b->hd = 0; 3157 b->ht_size = 0; 3158 b->ht_flag = PETSC_FALSE; 3159 b->ht_fact = 0; 3160 b->ht_total_ct = 0; 3161 b->ht_insert_ct = 0; 3162 3163 /* stuff for MatGetSubMatrices_MPIBAIJ_local() */ 3164 b->ijonly = PETSC_FALSE; 3165 3166 3167 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpibaij_mpiadj_C",MatConvert_MPIBAIJ_MPIAdj);CHKERRQ(ierr); 3168 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpibaij_mpiaij_C",MatConvert_MPIBAIJ_MPIAIJ);CHKERRQ(ierr); 3169 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpibaij_mpisbaij_C",MatConvert_MPIBAIJ_MPISBAIJ);CHKERRQ(ierr); 3170 ierr = PetscObjectComposeFunction((PetscObject)B,"MatStoreValues_C",MatStoreValues_MPIBAIJ);CHKERRQ(ierr); 3171 ierr = PetscObjectComposeFunction((PetscObject)B,"MatRetrieveValues_C",MatRetrieveValues_MPIBAIJ);CHKERRQ(ierr); 3172 ierr = PetscObjectComposeFunction((PetscObject)B,"MatGetDiagonalBlock_C",MatGetDiagonalBlock_MPIBAIJ);CHKERRQ(ierr); 3173 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMPIBAIJSetPreallocation_C",MatMPIBAIJSetPreallocation_MPIBAIJ);CHKERRQ(ierr); 3174 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMPIBAIJSetPreallocationCSR_C",MatMPIBAIJSetPreallocationCSR_MPIBAIJ);CHKERRQ(ierr); 3175 ierr = PetscObjectComposeFunction((PetscObject)B,"MatDiagonalScaleLocal_C",MatDiagonalScaleLocal_MPIBAIJ);CHKERRQ(ierr); 3176 ierr = PetscObjectComposeFunction((PetscObject)B,"MatSetHashTableFactor_C",MatSetHashTableFactor_MPIBAIJ);CHKERRQ(ierr); 3177 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpibaij_mpibstrm_C",MatConvert_MPIBAIJ_MPIBSTRM);CHKERRQ(ierr); 3178 ierr = PetscObjectChangeTypeName((PetscObject)B,MATMPIBAIJ);CHKERRQ(ierr); 3179 3180 ierr = PetscOptionsBegin(PetscObjectComm((PetscObject)B),NULL,"Options for loading MPIBAIJ matrix 1","Mat");CHKERRQ(ierr); 3181 ierr = PetscOptionsBool("-mat_use_hash_table","Use hash table to save memory in constructing matrix","MatSetOption",flg,&flg,NULL);CHKERRQ(ierr); 3182 if (flg) { 3183 PetscReal fact = 1.39; 3184 ierr = MatSetOption(B,MAT_USE_HASH_TABLE,PETSC_TRUE);CHKERRQ(ierr); 3185 ierr = PetscOptionsReal("-mat_use_hash_table","Use hash table factor","MatMPIBAIJSetHashTableFactor",fact,&fact,NULL);CHKERRQ(ierr); 3186 if (fact <= 1.0) fact = 1.39; 3187 ierr = MatMPIBAIJSetHashTableFactor(B,fact);CHKERRQ(ierr); 3188 ierr = PetscInfo1(B,"Hash table Factor used %5.2f\n",fact);CHKERRQ(ierr); 3189 } 3190 ierr = PetscOptionsEnd();CHKERRQ(ierr); 3191 PetscFunctionReturn(0); 3192 } 3193 3194 /*MC 3195 MATBAIJ - MATBAIJ = "baij" - A matrix type to be used for block sparse matrices. 3196 3197 This matrix type is identical to MATSEQBAIJ when constructed with a single process communicator, 3198 and MATMPIBAIJ otherwise. 3199 3200 Options Database Keys: 3201 . -mat_type baij - sets the matrix type to "baij" during a call to MatSetFromOptions() 3202 3203 Level: beginner 3204 3205 .seealso: MatCreateBAIJ(),MATSEQBAIJ,MATMPIBAIJ, MatMPIBAIJSetPreallocation(), MatMPIBAIJSetPreallocationCSR() 3206 M*/ 3207 3208 #undef __FUNCT__ 3209 #define __FUNCT__ "MatMPIBAIJSetPreallocation" 3210 /*@C 3211 MatMPIBAIJSetPreallocation - Allocates memory for a sparse parallel matrix in block AIJ format 3212 (block compressed row). For good matrix assembly performance 3213 the user should preallocate the matrix storage by setting the parameters 3214 d_nz (or d_nnz) and o_nz (or o_nnz). By setting these parameters accurately, 3215 performance can be increased by more than a factor of 50. 3216 3217 Collective on Mat 3218 3219 Input Parameters: 3220 + B - the matrix 3221 . bs - size of block, the blocks are ALWAYS square. One can use MatSetBlockSizes() to set a different row and column blocksize but the row 3222 blocksize always defines the size of the blocks. The column blocksize sets the blocksize of the vectors obtained with MatCreateVecs() 3223 . d_nz - number of block nonzeros per block row in diagonal portion of local 3224 submatrix (same for all local rows) 3225 . d_nnz - array containing the number of block nonzeros in the various block rows 3226 of the in diagonal portion of the local (possibly different for each block 3227 row) or NULL. If you plan to factor the matrix you must leave room for the diagonal entry and 3228 set it even if it is zero. 3229 . o_nz - number of block nonzeros per block row in the off-diagonal portion of local 3230 submatrix (same for all local rows). 3231 - o_nnz - array containing the number of nonzeros in the various block rows of the 3232 off-diagonal portion of the local submatrix (possibly different for 3233 each block row) or NULL. 3234 3235 If the *_nnz parameter is given then the *_nz parameter is ignored 3236 3237 Options Database Keys: 3238 + -mat_block_size - size of the blocks to use 3239 - -mat_use_hash_table <fact> 3240 3241 Notes: 3242 If PETSC_DECIDE or PETSC_DETERMINE is used for a particular argument on one processor 3243 than it must be used on all processors that share the object for that argument. 3244 3245 Storage Information: 3246 For a square global matrix we define each processor's diagonal portion 3247 to be its local rows and the corresponding columns (a square submatrix); 3248 each processor's off-diagonal portion encompasses the remainder of the 3249 local matrix (a rectangular submatrix). 3250 3251 The user can specify preallocated storage for the diagonal part of 3252 the local submatrix with either d_nz or d_nnz (not both). Set 3253 d_nz=PETSC_DEFAULT and d_nnz=NULL for PETSc to control dynamic 3254 memory allocation. Likewise, specify preallocated storage for the 3255 off-diagonal part of the local submatrix with o_nz or o_nnz (not both). 3256 3257 Consider a processor that owns rows 3, 4 and 5 of a parallel matrix. In 3258 the figure below we depict these three local rows and all columns (0-11). 3259 3260 .vb 3261 0 1 2 3 4 5 6 7 8 9 10 11 3262 -------------------------- 3263 row 3 |o o o d d d o o o o o o 3264 row 4 |o o o d d d o o o o o o 3265 row 5 |o o o d d d o o o o o o 3266 -------------------------- 3267 .ve 3268 3269 Thus, any entries in the d locations are stored in the d (diagonal) 3270 submatrix, and any entries in the o locations are stored in the 3271 o (off-diagonal) submatrix. Note that the d and the o submatrices are 3272 stored simply in the MATSEQBAIJ format for compressed row storage. 3273 3274 Now d_nz should indicate the number of block nonzeros per row in the d matrix, 3275 and o_nz should indicate the number of block nonzeros per row in the o matrix. 3276 In general, for PDE problems in which most nonzeros are near the diagonal, 3277 one expects d_nz >> o_nz. For large problems you MUST preallocate memory 3278 or you will get TERRIBLE performance; see the users' manual chapter on 3279 matrices. 3280 3281 You can call MatGetInfo() to get information on how effective the preallocation was; 3282 for example the fields mallocs,nz_allocated,nz_used,nz_unneeded; 3283 You can also run with the option -info and look for messages with the string 3284 malloc in them to see if additional memory allocation was needed. 3285 3286 Level: intermediate 3287 3288 .keywords: matrix, block, aij, compressed row, sparse, parallel 3289 3290 .seealso: MatCreate(), MatCreateSeqBAIJ(), MatSetValues(), MatCreateBAIJ(), MatMPIBAIJSetPreallocationCSR(), PetscSplitOwnership() 3291 @*/ 3292 PetscErrorCode MatMPIBAIJSetPreallocation(Mat B,PetscInt bs,PetscInt d_nz,const PetscInt d_nnz[],PetscInt o_nz,const PetscInt o_nnz[]) 3293 { 3294 PetscErrorCode ierr; 3295 3296 PetscFunctionBegin; 3297 PetscValidHeaderSpecific(B,MAT_CLASSID,1); 3298 PetscValidType(B,1); 3299 PetscValidLogicalCollectiveInt(B,bs,2); 3300 ierr = PetscTryMethod(B,"MatMPIBAIJSetPreallocation_C",(Mat,PetscInt,PetscInt,const PetscInt[],PetscInt,const PetscInt[]),(B,bs,d_nz,d_nnz,o_nz,o_nnz));CHKERRQ(ierr); 3301 PetscFunctionReturn(0); 3302 } 3303 3304 #undef __FUNCT__ 3305 #define __FUNCT__ "MatCreateBAIJ" 3306 /*@C 3307 MatCreateBAIJ - Creates a sparse parallel matrix in block AIJ format 3308 (block compressed row). For good matrix assembly performance 3309 the user should preallocate the matrix storage by setting the parameters 3310 d_nz (or d_nnz) and o_nz (or o_nnz). By setting these parameters accurately, 3311 performance can be increased by more than a factor of 50. 3312 3313 Collective on MPI_Comm 3314 3315 Input Parameters: 3316 + comm - MPI communicator 3317 . bs - size of block, the blocks are ALWAYS square. One can use MatSetBlockSizes() to set a different row and column blocksize but the row 3318 blocksize always defines the size of the blocks. The column blocksize sets the blocksize of the vectors obtained with MatCreateVecs() 3319 . m - number of local rows (or PETSC_DECIDE to have calculated if M is given) 3320 This value should be the same as the local size used in creating the 3321 y vector for the matrix-vector product y = Ax. 3322 . n - number of local columns (or PETSC_DECIDE to have calculated if N is given) 3323 This value should be the same as the local size used in creating the 3324 x vector for the matrix-vector product y = Ax. 3325 . M - number of global rows (or PETSC_DETERMINE to have calculated if m is given) 3326 . N - number of global columns (or PETSC_DETERMINE to have calculated if n is given) 3327 . d_nz - number of nonzero blocks per block row in diagonal portion of local 3328 submatrix (same for all local rows) 3329 . d_nnz - array containing the number of nonzero blocks in the various block rows 3330 of the in diagonal portion of the local (possibly different for each block 3331 row) or NULL. If you plan to factor the matrix you must leave room for the diagonal entry 3332 and set it even if it is zero. 3333 . o_nz - number of nonzero blocks per block row in the off-diagonal portion of local 3334 submatrix (same for all local rows). 3335 - o_nnz - array containing the number of nonzero blocks in the various block rows of the 3336 off-diagonal portion of the local submatrix (possibly different for 3337 each block row) or NULL. 3338 3339 Output Parameter: 3340 . A - the matrix 3341 3342 Options Database Keys: 3343 + -mat_block_size - size of the blocks to use 3344 - -mat_use_hash_table <fact> 3345 3346 It is recommended that one use the MatCreate(), MatSetType() and/or MatSetFromOptions(), 3347 MatXXXXSetPreallocation() paradgm instead of this routine directly. 3348 [MatXXXXSetPreallocation() is, for example, MatSeqAIJSetPreallocation] 3349 3350 Notes: 3351 If the *_nnz parameter is given then the *_nz parameter is ignored 3352 3353 A nonzero block is any block that as 1 or more nonzeros in it 3354 3355 The user MUST specify either the local or global matrix dimensions 3356 (possibly both). 3357 3358 If PETSC_DECIDE or PETSC_DETERMINE is used for a particular argument on one processor 3359 than it must be used on all processors that share the object for that argument. 3360 3361 Storage Information: 3362 For a square global matrix we define each processor's diagonal portion 3363 to be its local rows and the corresponding columns (a square submatrix); 3364 each processor's off-diagonal portion encompasses the remainder of the 3365 local matrix (a rectangular submatrix). 3366 3367 The user can specify preallocated storage for the diagonal part of 3368 the local submatrix with either d_nz or d_nnz (not both). Set 3369 d_nz=PETSC_DEFAULT and d_nnz=NULL for PETSc to control dynamic 3370 memory allocation. Likewise, specify preallocated storage for the 3371 off-diagonal part of the local submatrix with o_nz or o_nnz (not both). 3372 3373 Consider a processor that owns rows 3, 4 and 5 of a parallel matrix. In 3374 the figure below we depict these three local rows and all columns (0-11). 3375 3376 .vb 3377 0 1 2 3 4 5 6 7 8 9 10 11 3378 -------------------------- 3379 row 3 |o o o d d d o o o o o o 3380 row 4 |o o o d d d o o o o o o 3381 row 5 |o o o d d d o o o o o o 3382 -------------------------- 3383 .ve 3384 3385 Thus, any entries in the d locations are stored in the d (diagonal) 3386 submatrix, and any entries in the o locations are stored in the 3387 o (off-diagonal) submatrix. Note that the d and the o submatrices are 3388 stored simply in the MATSEQBAIJ format for compressed row storage. 3389 3390 Now d_nz should indicate the number of block nonzeros per row in the d matrix, 3391 and o_nz should indicate the number of block nonzeros per row in the o matrix. 3392 In general, for PDE problems in which most nonzeros are near the diagonal, 3393 one expects d_nz >> o_nz. For large problems you MUST preallocate memory 3394 or you will get TERRIBLE performance; see the users' manual chapter on 3395 matrices. 3396 3397 Level: intermediate 3398 3399 .keywords: matrix, block, aij, compressed row, sparse, parallel 3400 3401 .seealso: MatCreate(), MatCreateSeqBAIJ(), MatSetValues(), MatCreateBAIJ(), MatMPIBAIJSetPreallocation(), MatMPIBAIJSetPreallocationCSR() 3402 @*/ 3403 PetscErrorCode MatCreateBAIJ(MPI_Comm comm,PetscInt bs,PetscInt m,PetscInt n,PetscInt M,PetscInt N,PetscInt d_nz,const PetscInt d_nnz[],PetscInt o_nz,const PetscInt o_nnz[],Mat *A) 3404 { 3405 PetscErrorCode ierr; 3406 PetscMPIInt size; 3407 3408 PetscFunctionBegin; 3409 ierr = MatCreate(comm,A);CHKERRQ(ierr); 3410 ierr = MatSetSizes(*A,m,n,M,N);CHKERRQ(ierr); 3411 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 3412 if (size > 1) { 3413 ierr = MatSetType(*A,MATMPIBAIJ);CHKERRQ(ierr); 3414 ierr = MatMPIBAIJSetPreallocation(*A,bs,d_nz,d_nnz,o_nz,o_nnz);CHKERRQ(ierr); 3415 } else { 3416 ierr = MatSetType(*A,MATSEQBAIJ);CHKERRQ(ierr); 3417 ierr = MatSeqBAIJSetPreallocation(*A,bs,d_nz,d_nnz);CHKERRQ(ierr); 3418 } 3419 PetscFunctionReturn(0); 3420 } 3421 3422 #undef __FUNCT__ 3423 #define __FUNCT__ "MatDuplicate_MPIBAIJ" 3424 static PetscErrorCode MatDuplicate_MPIBAIJ(Mat matin,MatDuplicateOption cpvalues,Mat *newmat) 3425 { 3426 Mat mat; 3427 Mat_MPIBAIJ *a,*oldmat = (Mat_MPIBAIJ*)matin->data; 3428 PetscErrorCode ierr; 3429 PetscInt len=0; 3430 3431 PetscFunctionBegin; 3432 *newmat = 0; 3433 ierr = MatCreate(PetscObjectComm((PetscObject)matin),&mat);CHKERRQ(ierr); 3434 ierr = MatSetSizes(mat,matin->rmap->n,matin->cmap->n,matin->rmap->N,matin->cmap->N);CHKERRQ(ierr); 3435 ierr = MatSetType(mat,((PetscObject)matin)->type_name);CHKERRQ(ierr); 3436 ierr = PetscMemcpy(mat->ops,matin->ops,sizeof(struct _MatOps));CHKERRQ(ierr); 3437 3438 mat->factortype = matin->factortype; 3439 mat->preallocated = PETSC_TRUE; 3440 mat->assembled = PETSC_TRUE; 3441 mat->insertmode = NOT_SET_VALUES; 3442 3443 a = (Mat_MPIBAIJ*)mat->data; 3444 mat->rmap->bs = matin->rmap->bs; 3445 a->bs2 = oldmat->bs2; 3446 a->mbs = oldmat->mbs; 3447 a->nbs = oldmat->nbs; 3448 a->Mbs = oldmat->Mbs; 3449 a->Nbs = oldmat->Nbs; 3450 3451 ierr = PetscLayoutReference(matin->rmap,&mat->rmap);CHKERRQ(ierr); 3452 ierr = PetscLayoutReference(matin->cmap,&mat->cmap);CHKERRQ(ierr); 3453 3454 a->size = oldmat->size; 3455 a->rank = oldmat->rank; 3456 a->donotstash = oldmat->donotstash; 3457 a->roworiented = oldmat->roworiented; 3458 a->rowindices = 0; 3459 a->rowvalues = 0; 3460 a->getrowactive = PETSC_FALSE; 3461 a->barray = 0; 3462 a->rstartbs = oldmat->rstartbs; 3463 a->rendbs = oldmat->rendbs; 3464 a->cstartbs = oldmat->cstartbs; 3465 a->cendbs = oldmat->cendbs; 3466 3467 /* hash table stuff */ 3468 a->ht = 0; 3469 a->hd = 0; 3470 a->ht_size = 0; 3471 a->ht_flag = oldmat->ht_flag; 3472 a->ht_fact = oldmat->ht_fact; 3473 a->ht_total_ct = 0; 3474 a->ht_insert_ct = 0; 3475 3476 ierr = PetscMemcpy(a->rangebs,oldmat->rangebs,(a->size+1)*sizeof(PetscInt));CHKERRQ(ierr); 3477 if (oldmat->colmap) { 3478 #if defined(PETSC_USE_CTABLE) 3479 ierr = PetscTableCreateCopy(oldmat->colmap,&a->colmap);CHKERRQ(ierr); 3480 #else 3481 ierr = PetscMalloc1(a->Nbs,&a->colmap);CHKERRQ(ierr); 3482 ierr = PetscLogObjectMemory((PetscObject)mat,(a->Nbs)*sizeof(PetscInt));CHKERRQ(ierr); 3483 ierr = PetscMemcpy(a->colmap,oldmat->colmap,(a->Nbs)*sizeof(PetscInt));CHKERRQ(ierr); 3484 #endif 3485 } else a->colmap = 0; 3486 3487 if (oldmat->garray && (len = ((Mat_SeqBAIJ*)(oldmat->B->data))->nbs)) { 3488 ierr = PetscMalloc1(len,&a->garray);CHKERRQ(ierr); 3489 ierr = PetscLogObjectMemory((PetscObject)mat,len*sizeof(PetscInt));CHKERRQ(ierr); 3490 ierr = PetscMemcpy(a->garray,oldmat->garray,len*sizeof(PetscInt));CHKERRQ(ierr); 3491 } else a->garray = 0; 3492 3493 ierr = MatStashCreate_Private(PetscObjectComm((PetscObject)matin),matin->rmap->bs,&mat->bstash);CHKERRQ(ierr); 3494 ierr = VecDuplicate(oldmat->lvec,&a->lvec);CHKERRQ(ierr); 3495 ierr = PetscLogObjectParent((PetscObject)mat,(PetscObject)a->lvec);CHKERRQ(ierr); 3496 ierr = VecScatterCopy(oldmat->Mvctx,&a->Mvctx);CHKERRQ(ierr); 3497 ierr = PetscLogObjectParent((PetscObject)mat,(PetscObject)a->Mvctx);CHKERRQ(ierr); 3498 3499 ierr = MatDuplicate(oldmat->A,cpvalues,&a->A);CHKERRQ(ierr); 3500 ierr = PetscLogObjectParent((PetscObject)mat,(PetscObject)a->A);CHKERRQ(ierr); 3501 ierr = MatDuplicate(oldmat->B,cpvalues,&a->B);CHKERRQ(ierr); 3502 ierr = PetscLogObjectParent((PetscObject)mat,(PetscObject)a->B);CHKERRQ(ierr); 3503 ierr = PetscFunctionListDuplicate(((PetscObject)matin)->qlist,&((PetscObject)mat)->qlist);CHKERRQ(ierr); 3504 *newmat = mat; 3505 PetscFunctionReturn(0); 3506 } 3507 3508 #undef __FUNCT__ 3509 #define __FUNCT__ "MatLoad_MPIBAIJ" 3510 PetscErrorCode MatLoad_MPIBAIJ(Mat newmat,PetscViewer viewer) 3511 { 3512 PetscErrorCode ierr; 3513 int fd; 3514 PetscInt i,nz,j,rstart,rend; 3515 PetscScalar *vals,*buf; 3516 MPI_Comm comm; 3517 MPI_Status status; 3518 PetscMPIInt rank,size,maxnz; 3519 PetscInt header[4],*rowlengths = 0,M,N,m,*rowners,*cols; 3520 PetscInt *locrowlens = NULL,*procsnz = NULL,*browners = NULL; 3521 PetscInt jj,*mycols,*ibuf,bs = newmat->rmap->bs,Mbs,mbs,extra_rows,mmax; 3522 PetscMPIInt tag = ((PetscObject)viewer)->tag; 3523 PetscInt *dlens = NULL,*odlens = NULL,*mask = NULL,*masked1 = NULL,*masked2 = NULL,rowcount,odcount; 3524 PetscInt dcount,kmax,k,nzcount,tmp,mend; 3525 3526 PetscFunctionBegin; 3527 /* force binary viewer to load .info file if it has not yet done so */ 3528 ierr = PetscViewerSetUp(viewer);CHKERRQ(ierr); 3529 ierr = PetscObjectGetComm((PetscObject)viewer,&comm);CHKERRQ(ierr); 3530 ierr = PetscOptionsBegin(comm,NULL,"Options for loading MPIBAIJ matrix 2","Mat");CHKERRQ(ierr); 3531 ierr = PetscOptionsInt("-matload_block_size","Set the blocksize used to store the matrix","MatLoad",bs,&bs,NULL);CHKERRQ(ierr); 3532 ierr = PetscOptionsEnd();CHKERRQ(ierr); 3533 if (bs < 0) bs = 1; 3534 3535 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 3536 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 3537 ierr = PetscViewerBinaryGetDescriptor(viewer,&fd);CHKERRQ(ierr); 3538 if (!rank) { 3539 ierr = PetscBinaryRead(fd,(char*)header,4,PETSC_INT);CHKERRQ(ierr); 3540 if (header[0] != MAT_FILE_CLASSID) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED,"not matrix object"); 3541 } 3542 ierr = MPI_Bcast(header+1,3,MPIU_INT,0,comm);CHKERRQ(ierr); 3543 M = header[1]; N = header[2]; 3544 3545 /* If global sizes are set, check if they are consistent with that given in the file */ 3546 if (newmat->rmap->N >= 0 && newmat->rmap->N != M) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED,"Inconsistent # of rows:Matrix in file has (%D) and input matrix has (%D)",newmat->rmap->N,M); 3547 if (newmat->cmap->N >= 0 && newmat->cmap->N != N) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED,"Inconsistent # of cols:Matrix in file has (%D) and input matrix has (%D)",newmat->cmap->N,N); 3548 3549 if (M != N) SETERRQ(PetscObjectComm((PetscObject)viewer),PETSC_ERR_SUP,"Can only do square matrices"); 3550 3551 /* 3552 This code adds extra rows to make sure the number of rows is 3553 divisible by the blocksize 3554 */ 3555 Mbs = M/bs; 3556 extra_rows = bs - M + bs*Mbs; 3557 if (extra_rows == bs) extra_rows = 0; 3558 else Mbs++; 3559 if (extra_rows && !rank) { 3560 ierr = PetscInfo(viewer,"Padding loaded matrix to match blocksize\n");CHKERRQ(ierr); 3561 } 3562 3563 /* determine ownership of all rows */ 3564 if (newmat->rmap->n < 0) { /* PETSC_DECIDE */ 3565 mbs = Mbs/size + ((Mbs % size) > rank); 3566 m = mbs*bs; 3567 } else { /* User set */ 3568 m = newmat->rmap->n; 3569 mbs = m/bs; 3570 } 3571 ierr = PetscMalloc2(size+1,&rowners,size+1,&browners);CHKERRQ(ierr); 3572 ierr = MPI_Allgather(&mbs,1,MPIU_INT,rowners+1,1,MPIU_INT,comm);CHKERRQ(ierr); 3573 3574 /* process 0 needs enough room for process with most rows */ 3575 if (!rank) { 3576 mmax = rowners[1]; 3577 for (i=2; i<=size; i++) { 3578 mmax = PetscMax(mmax,rowners[i]); 3579 } 3580 mmax*=bs; 3581 } else mmax = -1; /* unused, but compiler warns anyway */ 3582 3583 rowners[0] = 0; 3584 for (i=2; i<=size; i++) rowners[i] += rowners[i-1]; 3585 for (i=0; i<=size; i++) browners[i] = rowners[i]*bs; 3586 rstart = rowners[rank]; 3587 rend = rowners[rank+1]; 3588 3589 /* distribute row lengths to all processors */ 3590 ierr = PetscMalloc1(m,&locrowlens);CHKERRQ(ierr); 3591 if (!rank) { 3592 mend = m; 3593 if (size == 1) mend = mend - extra_rows; 3594 ierr = PetscBinaryRead(fd,locrowlens,mend,PETSC_INT);CHKERRQ(ierr); 3595 for (j=mend; j<m; j++) locrowlens[j] = 1; 3596 ierr = PetscMalloc1(mmax,&rowlengths);CHKERRQ(ierr); 3597 ierr = PetscCalloc1(size,&procsnz);CHKERRQ(ierr); 3598 for (j=0; j<m; j++) { 3599 procsnz[0] += locrowlens[j]; 3600 } 3601 for (i=1; i<size; i++) { 3602 mend = browners[i+1] - browners[i]; 3603 if (i == size-1) mend = mend - extra_rows; 3604 ierr = PetscBinaryRead(fd,rowlengths,mend,PETSC_INT);CHKERRQ(ierr); 3605 for (j=mend; j<browners[i+1] - browners[i]; j++) rowlengths[j] = 1; 3606 /* calculate the number of nonzeros on each processor */ 3607 for (j=0; j<browners[i+1]-browners[i]; j++) { 3608 procsnz[i] += rowlengths[j]; 3609 } 3610 ierr = MPI_Send(rowlengths,browners[i+1]-browners[i],MPIU_INT,i,tag,comm);CHKERRQ(ierr); 3611 } 3612 ierr = PetscFree(rowlengths);CHKERRQ(ierr); 3613 } else { 3614 ierr = MPI_Recv(locrowlens,m,MPIU_INT,0,tag,comm,&status);CHKERRQ(ierr); 3615 } 3616 3617 if (!rank) { 3618 /* determine max buffer needed and allocate it */ 3619 maxnz = procsnz[0]; 3620 for (i=1; i<size; i++) { 3621 maxnz = PetscMax(maxnz,procsnz[i]); 3622 } 3623 ierr = PetscMalloc1(maxnz,&cols);CHKERRQ(ierr); 3624 3625 /* read in my part of the matrix column indices */ 3626 nz = procsnz[0]; 3627 ierr = PetscMalloc1(nz+1,&ibuf);CHKERRQ(ierr); 3628 mycols = ibuf; 3629 if (size == 1) nz -= extra_rows; 3630 ierr = PetscBinaryRead(fd,mycols,nz,PETSC_INT);CHKERRQ(ierr); 3631 if (size == 1) { 3632 for (i=0; i< extra_rows; i++) mycols[nz+i] = M+i; 3633 } 3634 3635 /* read in every ones (except the last) and ship off */ 3636 for (i=1; i<size-1; i++) { 3637 nz = procsnz[i]; 3638 ierr = PetscBinaryRead(fd,cols,nz,PETSC_INT);CHKERRQ(ierr); 3639 ierr = MPI_Send(cols,nz,MPIU_INT,i,tag,comm);CHKERRQ(ierr); 3640 } 3641 /* read in the stuff for the last proc */ 3642 if (size != 1) { 3643 nz = procsnz[size-1] - extra_rows; /* the extra rows are not on the disk */ 3644 ierr = PetscBinaryRead(fd,cols,nz,PETSC_INT);CHKERRQ(ierr); 3645 for (i=0; i<extra_rows; i++) cols[nz+i] = M+i; 3646 ierr = MPI_Send(cols,nz+extra_rows,MPIU_INT,size-1,tag,comm);CHKERRQ(ierr); 3647 } 3648 ierr = PetscFree(cols);CHKERRQ(ierr); 3649 } else { 3650 /* determine buffer space needed for message */ 3651 nz = 0; 3652 for (i=0; i<m; i++) { 3653 nz += locrowlens[i]; 3654 } 3655 ierr = PetscMalloc1(nz+1,&ibuf);CHKERRQ(ierr); 3656 mycols = ibuf; 3657 /* receive message of column indices*/ 3658 ierr = MPI_Recv(mycols,nz,MPIU_INT,0,tag,comm,&status);CHKERRQ(ierr); 3659 ierr = MPI_Get_count(&status,MPIU_INT,&maxnz);CHKERRQ(ierr); 3660 if (maxnz != nz) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED,"something is wrong with file"); 3661 } 3662 3663 /* loop over local rows, determining number of off diagonal entries */ 3664 ierr = PetscMalloc2(rend-rstart,&dlens,rend-rstart,&odlens);CHKERRQ(ierr); 3665 ierr = PetscCalloc3(Mbs,&mask,Mbs,&masked1,Mbs,&masked2);CHKERRQ(ierr); 3666 rowcount = 0; nzcount = 0; 3667 for (i=0; i<mbs; i++) { 3668 dcount = 0; 3669 odcount = 0; 3670 for (j=0; j<bs; j++) { 3671 kmax = locrowlens[rowcount]; 3672 for (k=0; k<kmax; k++) { 3673 tmp = mycols[nzcount++]/bs; 3674 if (!mask[tmp]) { 3675 mask[tmp] = 1; 3676 if (tmp < rstart || tmp >= rend) masked2[odcount++] = tmp; 3677 else masked1[dcount++] = tmp; 3678 } 3679 } 3680 rowcount++; 3681 } 3682 3683 dlens[i] = dcount; 3684 odlens[i] = odcount; 3685 3686 /* zero out the mask elements we set */ 3687 for (j=0; j<dcount; j++) mask[masked1[j]] = 0; 3688 for (j=0; j<odcount; j++) mask[masked2[j]] = 0; 3689 } 3690 3691 ierr = MatSetSizes(newmat,m,m,M+extra_rows,N+extra_rows);CHKERRQ(ierr); 3692 ierr = MatMPIBAIJSetPreallocation(newmat,bs,0,dlens,0,odlens);CHKERRQ(ierr); 3693 3694 if (!rank) { 3695 ierr = PetscMalloc1(maxnz+1,&buf);CHKERRQ(ierr); 3696 /* read in my part of the matrix numerical values */ 3697 nz = procsnz[0]; 3698 vals = buf; 3699 mycols = ibuf; 3700 if (size == 1) nz -= extra_rows; 3701 ierr = PetscBinaryRead(fd,vals,nz,PETSC_SCALAR);CHKERRQ(ierr); 3702 if (size == 1) { 3703 for (i=0; i< extra_rows; i++) vals[nz+i] = 1.0; 3704 } 3705 3706 /* insert into matrix */ 3707 jj = rstart*bs; 3708 for (i=0; i<m; i++) { 3709 ierr = MatSetValues_MPIBAIJ(newmat,1,&jj,locrowlens[i],mycols,vals,INSERT_VALUES);CHKERRQ(ierr); 3710 mycols += locrowlens[i]; 3711 vals += locrowlens[i]; 3712 jj++; 3713 } 3714 /* read in other processors (except the last one) and ship out */ 3715 for (i=1; i<size-1; i++) { 3716 nz = procsnz[i]; 3717 vals = buf; 3718 ierr = PetscBinaryRead(fd,vals,nz,PETSC_SCALAR);CHKERRQ(ierr); 3719 ierr = MPIULong_Send(vals,nz,MPIU_SCALAR,i,((PetscObject)newmat)->tag,comm);CHKERRQ(ierr); 3720 } 3721 /* the last proc */ 3722 if (size != 1) { 3723 nz = procsnz[i] - extra_rows; 3724 vals = buf; 3725 ierr = PetscBinaryRead(fd,vals,nz,PETSC_SCALAR);CHKERRQ(ierr); 3726 for (i=0; i<extra_rows; i++) vals[nz+i] = 1.0; 3727 ierr = MPIULong_Send(vals,nz+extra_rows,MPIU_SCALAR,size-1,((PetscObject)newmat)->tag,comm);CHKERRQ(ierr); 3728 } 3729 ierr = PetscFree(procsnz);CHKERRQ(ierr); 3730 } else { 3731 /* receive numeric values */ 3732 ierr = PetscMalloc1(nz+1,&buf);CHKERRQ(ierr); 3733 3734 /* receive message of values*/ 3735 vals = buf; 3736 mycols = ibuf; 3737 ierr = MPIULong_Recv(vals,nz,MPIU_SCALAR,0,((PetscObject)newmat)->tag,comm);CHKERRQ(ierr); 3738 3739 /* insert into matrix */ 3740 jj = rstart*bs; 3741 for (i=0; i<m; i++) { 3742 ierr = MatSetValues_MPIBAIJ(newmat,1,&jj,locrowlens[i],mycols,vals,INSERT_VALUES);CHKERRQ(ierr); 3743 mycols += locrowlens[i]; 3744 vals += locrowlens[i]; 3745 jj++; 3746 } 3747 } 3748 ierr = PetscFree(locrowlens);CHKERRQ(ierr); 3749 ierr = PetscFree(buf);CHKERRQ(ierr); 3750 ierr = PetscFree(ibuf);CHKERRQ(ierr); 3751 ierr = PetscFree2(rowners,browners);CHKERRQ(ierr); 3752 ierr = PetscFree2(dlens,odlens);CHKERRQ(ierr); 3753 ierr = PetscFree3(mask,masked1,masked2);CHKERRQ(ierr); 3754 ierr = MatAssemblyBegin(newmat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3755 ierr = MatAssemblyEnd(newmat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3756 PetscFunctionReturn(0); 3757 } 3758 3759 #undef __FUNCT__ 3760 #define __FUNCT__ "MatMPIBAIJSetHashTableFactor" 3761 /*@ 3762 MatMPIBAIJSetHashTableFactor - Sets the factor required to compute the size of the HashTable. 3763 3764 Input Parameters: 3765 . mat - the matrix 3766 . fact - factor 3767 3768 Not Collective, each process can use a different factor 3769 3770 Level: advanced 3771 3772 Notes: 3773 This can also be set by the command line option: -mat_use_hash_table <fact> 3774 3775 .keywords: matrix, hashtable, factor, HT 3776 3777 .seealso: MatSetOption() 3778 @*/ 3779 PetscErrorCode MatMPIBAIJSetHashTableFactor(Mat mat,PetscReal fact) 3780 { 3781 PetscErrorCode ierr; 3782 3783 PetscFunctionBegin; 3784 ierr = PetscTryMethod(mat,"MatSetHashTableFactor_C",(Mat,PetscReal),(mat,fact));CHKERRQ(ierr); 3785 PetscFunctionReturn(0); 3786 } 3787 3788 #undef __FUNCT__ 3789 #define __FUNCT__ "MatSetHashTableFactor_MPIBAIJ" 3790 PetscErrorCode MatSetHashTableFactor_MPIBAIJ(Mat mat,PetscReal fact) 3791 { 3792 Mat_MPIBAIJ *baij; 3793 3794 PetscFunctionBegin; 3795 baij = (Mat_MPIBAIJ*)mat->data; 3796 baij->ht_fact = fact; 3797 PetscFunctionReturn(0); 3798 } 3799 3800 #undef __FUNCT__ 3801 #define __FUNCT__ "MatMPIBAIJGetSeqBAIJ" 3802 PetscErrorCode MatMPIBAIJGetSeqBAIJ(Mat A,Mat *Ad,Mat *Ao,const PetscInt *colmap[]) 3803 { 3804 Mat_MPIBAIJ *a = (Mat_MPIBAIJ*)A->data; 3805 3806 PetscFunctionBegin; 3807 if (Ad) *Ad = a->A; 3808 if (Ao) *Ao = a->B; 3809 if (colmap) *colmap = a->garray; 3810 PetscFunctionReturn(0); 3811 } 3812 3813 /* 3814 Special version for direct calls from Fortran (to eliminate two function call overheads 3815 */ 3816 #if defined(PETSC_HAVE_FORTRAN_CAPS) 3817 #define matmpibaijsetvaluesblocked_ MATMPIBAIJSETVALUESBLOCKED 3818 #elif !defined(PETSC_HAVE_FORTRAN_UNDERSCORE) 3819 #define matmpibaijsetvaluesblocked_ matmpibaijsetvaluesblocked 3820 #endif 3821 3822 #undef __FUNCT__ 3823 #define __FUNCT__ "matmpibiajsetvaluesblocked" 3824 /*@C 3825 MatMPIBAIJSetValuesBlocked - Direct Fortran call to replace call to MatSetValuesBlocked() 3826 3827 Collective on Mat 3828 3829 Input Parameters: 3830 + mat - the matrix 3831 . min - number of input rows 3832 . im - input rows 3833 . nin - number of input columns 3834 . in - input columns 3835 . v - numerical values input 3836 - addvin - INSERT_VALUES or ADD_VALUES 3837 3838 Notes: This has a complete copy of MatSetValuesBlocked_MPIBAIJ() which is terrible code un-reuse. 3839 3840 Level: advanced 3841 3842 .seealso: MatSetValuesBlocked() 3843 @*/ 3844 PetscErrorCode matmpibaijsetvaluesblocked_(Mat *matin,PetscInt *min,const PetscInt im[],PetscInt *nin,const PetscInt in[],const MatScalar v[],InsertMode *addvin) 3845 { 3846 /* convert input arguments to C version */ 3847 Mat mat = *matin; 3848 PetscInt m = *min, n = *nin; 3849 InsertMode addv = *addvin; 3850 3851 Mat_MPIBAIJ *baij = (Mat_MPIBAIJ*)mat->data; 3852 const MatScalar *value; 3853 MatScalar *barray = baij->barray; 3854 PetscBool roworiented = baij->roworiented; 3855 PetscErrorCode ierr; 3856 PetscInt i,j,ii,jj,row,col,rstart=baij->rstartbs; 3857 PetscInt rend=baij->rendbs,cstart=baij->cstartbs,stepval; 3858 PetscInt cend=baij->cendbs,bs=mat->rmap->bs,bs2=baij->bs2; 3859 3860 PetscFunctionBegin; 3861 /* tasks normally handled by MatSetValuesBlocked() */ 3862 if (mat->insertmode == NOT_SET_VALUES) mat->insertmode = addv; 3863 #if defined(PETSC_USE_DEBUG) 3864 else if (mat->insertmode != addv) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Cannot mix add values and insert values"); 3865 if (mat->factortype) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Not for factored matrix"); 3866 #endif 3867 if (mat->assembled) { 3868 mat->was_assembled = PETSC_TRUE; 3869 mat->assembled = PETSC_FALSE; 3870 } 3871 ierr = PetscLogEventBegin(MAT_SetValues,mat,0,0,0);CHKERRQ(ierr); 3872 3873 3874 if (!barray) { 3875 ierr = PetscMalloc1(bs2,&barray);CHKERRQ(ierr); 3876 baij->barray = barray; 3877 } 3878 3879 if (roworiented) stepval = (n-1)*bs; 3880 else stepval = (m-1)*bs; 3881 3882 for (i=0; i<m; i++) { 3883 if (im[i] < 0) continue; 3884 #if defined(PETSC_USE_DEBUG) 3885 if (im[i] >= baij->Mbs) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Row too large, row %D max %D",im[i],baij->Mbs-1); 3886 #endif 3887 if (im[i] >= rstart && im[i] < rend) { 3888 row = im[i] - rstart; 3889 for (j=0; j<n; j++) { 3890 /* If NumCol = 1 then a copy is not required */ 3891 if ((roworiented) && (n == 1)) { 3892 barray = (MatScalar*)v + i*bs2; 3893 } else if ((!roworiented) && (m == 1)) { 3894 barray = (MatScalar*)v + j*bs2; 3895 } else { /* Here a copy is required */ 3896 if (roworiented) { 3897 value = v + i*(stepval+bs)*bs + j*bs; 3898 } else { 3899 value = v + j*(stepval+bs)*bs + i*bs; 3900 } 3901 for (ii=0; ii<bs; ii++,value+=stepval) { 3902 for (jj=0; jj<bs; jj++) { 3903 *barray++ = *value++; 3904 } 3905 } 3906 barray -=bs2; 3907 } 3908 3909 if (in[j] >= cstart && in[j] < cend) { 3910 col = in[j] - cstart; 3911 ierr = MatSetValuesBlocked_SeqBAIJ_Inlined(baij->A,row,col,barray,addv,im[i],in[j]);CHKERRQ(ierr); 3912 } else if (in[j] < 0) continue; 3913 #if defined(PETSC_USE_DEBUG) 3914 else if (in[j] >= baij->Nbs) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Column too large, col %D max %D",in[j],baij->Nbs-1); 3915 #endif 3916 else { 3917 if (mat->was_assembled) { 3918 if (!baij->colmap) { 3919 ierr = MatCreateColmap_MPIBAIJ_Private(mat);CHKERRQ(ierr); 3920 } 3921 3922 #if defined(PETSC_USE_DEBUG) 3923 #if defined(PETSC_USE_CTABLE) 3924 { PetscInt data; 3925 ierr = PetscTableFind(baij->colmap,in[j]+1,&data);CHKERRQ(ierr); 3926 if ((data - 1) % bs) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Incorrect colmap"); 3927 } 3928 #else 3929 if ((baij->colmap[in[j]] - 1) % bs) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Incorrect colmap"); 3930 #endif 3931 #endif 3932 #if defined(PETSC_USE_CTABLE) 3933 ierr = PetscTableFind(baij->colmap,in[j]+1,&col);CHKERRQ(ierr); 3934 col = (col - 1)/bs; 3935 #else 3936 col = (baij->colmap[in[j]] - 1)/bs; 3937 #endif 3938 if (col < 0 && !((Mat_SeqBAIJ*)(baij->A->data))->nonew) { 3939 ierr = MatDisAssemble_MPIBAIJ(mat);CHKERRQ(ierr); 3940 col = in[j]; 3941 } 3942 } else col = in[j]; 3943 ierr = MatSetValuesBlocked_SeqBAIJ_Inlined(baij->B,row,col,barray,addv,im[i],in[j]);CHKERRQ(ierr); 3944 } 3945 } 3946 } else { 3947 if (!baij->donotstash) { 3948 if (roworiented) { 3949 ierr = MatStashValuesRowBlocked_Private(&mat->bstash,im[i],n,in,v,m,n,i);CHKERRQ(ierr); 3950 } else { 3951 ierr = MatStashValuesColBlocked_Private(&mat->bstash,im[i],n,in,v,m,n,i);CHKERRQ(ierr); 3952 } 3953 } 3954 } 3955 } 3956 3957 /* task normally handled by MatSetValuesBlocked() */ 3958 ierr = PetscLogEventEnd(MAT_SetValues,mat,0,0,0);CHKERRQ(ierr); 3959 PetscFunctionReturn(0); 3960 } 3961 3962 #undef __FUNCT__ 3963 #define __FUNCT__ "MatCreateMPIBAIJWithArrays" 3964 /*@ 3965 MatCreateMPIBAIJWithArrays - creates a MPI BAIJ matrix using arrays that contain in standard 3966 CSR format the local rows. 3967 3968 Collective on MPI_Comm 3969 3970 Input Parameters: 3971 + comm - MPI communicator 3972 . bs - the block size, only a block size of 1 is supported 3973 . m - number of local rows (Cannot be PETSC_DECIDE) 3974 . n - This value should be the same as the local size used in creating the 3975 x vector for the matrix-vector product y = Ax. (or PETSC_DECIDE to have 3976 calculated if N is given) For square matrices n is almost always m. 3977 . M - number of global rows (or PETSC_DETERMINE to have calculated if m is given) 3978 . N - number of global columns (or PETSC_DETERMINE to have calculated if n is given) 3979 . i - row indices 3980 . j - column indices 3981 - a - matrix values 3982 3983 Output Parameter: 3984 . mat - the matrix 3985 3986 Level: intermediate 3987 3988 Notes: 3989 The i, j, and a arrays ARE copied by this routine into the internal format used by PETSc; 3990 thus you CANNOT change the matrix entries by changing the values of a[] after you have 3991 called this routine. Use MatCreateMPIAIJWithSplitArrays() to avoid needing to copy the arrays. 3992 3993 The order of the entries in values is the same as the block compressed sparse row storage format; that is, it is 3994 the same as a three dimensional array in Fortran values(bs,bs,nnz) that contains the first column of the first 3995 block, followed by the second column of the first block etc etc. That is, the blocks are contiguous in memory 3996 with column-major ordering within blocks. 3997 3998 The i and j indices are 0 based, and i indices are indices corresponding to the local j array. 3999 4000 .keywords: matrix, aij, compressed row, sparse, parallel 4001 4002 .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatMPIAIJSetPreallocation(), MatMPIAIJSetPreallocationCSR(), 4003 MPIAIJ, MatCreateAIJ(), MatCreateMPIAIJWithSplitArrays() 4004 @*/ 4005 PetscErrorCode MatCreateMPIBAIJWithArrays(MPI_Comm comm,PetscInt bs,PetscInt m,PetscInt n,PetscInt M,PetscInt N,const PetscInt i[],const PetscInt j[],const PetscScalar a[],Mat *mat) 4006 { 4007 PetscErrorCode ierr; 4008 4009 PetscFunctionBegin; 4010 if (i[0]) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"i (row indices) must start with 0"); 4011 if (m < 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"local number of rows (m) cannot be PETSC_DECIDE, or negative"); 4012 ierr = MatCreate(comm,mat);CHKERRQ(ierr); 4013 ierr = MatSetSizes(*mat,m,n,M,N);CHKERRQ(ierr); 4014 ierr = MatSetType(*mat,MATMPISBAIJ);CHKERRQ(ierr); 4015 ierr = MatSetOption(*mat,MAT_ROW_ORIENTED,PETSC_FALSE);CHKERRQ(ierr); 4016 ierr = MatMPIBAIJSetPreallocationCSR(*mat,bs,i,j,a);CHKERRQ(ierr); 4017 ierr = MatSetOption(*mat,MAT_ROW_ORIENTED,PETSC_TRUE);CHKERRQ(ierr); 4018 PetscFunctionReturn(0); 4019 } 4020 4021 #undef __FUNCT__ 4022 #define __FUNCT__ "MatCreateMPIMatConcatenateSeqMat_MPIBAIJ" 4023 PetscErrorCode MatCreateMPIMatConcatenateSeqMat_MPIBAIJ(MPI_Comm comm,Mat inmat,PetscInt n,MatReuse scall,Mat *outmat) 4024 { 4025 PetscErrorCode ierr; 4026 PetscInt m,N,i,rstart,nnz,Ii,bs,cbs; 4027 PetscInt *indx; 4028 PetscScalar *values; 4029 4030 PetscFunctionBegin; 4031 ierr = MatGetSize(inmat,&m,&N);CHKERRQ(ierr); 4032 if (scall == MAT_INITIAL_MATRIX) { /* symbolic phase */ 4033 Mat_SeqBAIJ *a = (Mat_SeqBAIJ*)inmat->data; 4034 PetscInt *dnz,*onz,sum,mbs,Nbs; 4035 PetscInt *bindx,rmax=a->rmax,j; 4036 4037 ierr = MatGetBlockSizes(inmat,&bs,&cbs);CHKERRQ(ierr); 4038 mbs = m/bs; Nbs = N/cbs; 4039 if (n == PETSC_DECIDE) { 4040 ierr = PetscSplitOwnership(comm,&n,&Nbs);CHKERRQ(ierr); 4041 } 4042 /* Check sum(n) = Nbs */ 4043 ierr = MPI_Allreduce(&n,&sum,1,MPIU_INT,MPI_SUM,comm);CHKERRQ(ierr); 4044 if (sum != Nbs) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_INCOMP,"Sum of local columns != global columns %d",Nbs); 4045 4046 ierr = MPI_Scan(&mbs, &rstart,1,MPIU_INT,MPI_SUM,comm);CHKERRQ(ierr); 4047 rstart -= mbs; 4048 4049 ierr = PetscMalloc1(rmax,&bindx);CHKERRQ(ierr); 4050 ierr = MatPreallocateInitialize(comm,mbs,n,dnz,onz);CHKERRQ(ierr); 4051 for (i=0; i<mbs; i++) { 4052 ierr = MatGetRow_SeqBAIJ(inmat,i*bs,&nnz,&indx,NULL);CHKERRQ(ierr); /* non-blocked nnz and indx */ 4053 nnz = nnz/bs; 4054 for (j=0; j<nnz; j++) bindx[j] = indx[j*bs]/bs; 4055 ierr = MatPreallocateSet(i+rstart,nnz,bindx,dnz,onz);CHKERRQ(ierr); 4056 ierr = MatRestoreRow_SeqBAIJ(inmat,i*bs,&nnz,&indx,NULL);CHKERRQ(ierr); 4057 } 4058 ierr = PetscFree(bindx);CHKERRQ(ierr); 4059 4060 ierr = MatCreate(comm,outmat);CHKERRQ(ierr); 4061 ierr = MatSetSizes(*outmat,m,n*bs,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 4062 ierr = MatSetBlockSizes(*outmat,bs,cbs);CHKERRQ(ierr); 4063 ierr = MatSetType(*outmat,MATMPIBAIJ);CHKERRQ(ierr); 4064 ierr = MatMPIBAIJSetPreallocation(*outmat,bs,0,dnz,0,onz);CHKERRQ(ierr); 4065 ierr = MatPreallocateFinalize(dnz,onz);CHKERRQ(ierr); 4066 } 4067 4068 /* numeric phase */ 4069 ierr = MatGetBlockSizes(inmat,&bs,&cbs);CHKERRQ(ierr); 4070 ierr = MatGetOwnershipRange(*outmat,&rstart,NULL);CHKERRQ(ierr); 4071 4072 for (i=0; i<m; i++) { 4073 ierr = MatGetRow_SeqBAIJ(inmat,i,&nnz,&indx,&values);CHKERRQ(ierr); 4074 Ii = i + rstart; 4075 ierr = MatSetValues(*outmat,1,&Ii,nnz,indx,values,INSERT_VALUES);CHKERRQ(ierr); 4076 ierr = MatRestoreRow_SeqBAIJ(inmat,i,&nnz,&indx,&values);CHKERRQ(ierr); 4077 } 4078 ierr = MatAssemblyBegin(*outmat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4079 ierr = MatAssemblyEnd(*outmat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4080 PetscFunctionReturn(0); 4081 } 4082