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