1 /*$Id: mpisbaij.c,v 1.54 2001/06/21 23:38:52 buschelm Exp buschelm $*/ 2 3 #include "src/mat/impls/baij/mpi/mpibaij.h" /*I "petscmat.h" I*/ 4 #include "src/vec/vecimpl.h" 5 #include "mpisbaij.h" 6 #include "src/mat/impls/sbaij/seq/sbaij.h" 7 8 extern int MatSetUpMultiply_MPISBAIJ(Mat); 9 extern int DisAssemble_MPISBAIJ(Mat); 10 extern int MatIncreaseOverlap_MPISBAIJ(Mat,int,IS *,int); 11 extern int MatGetSubMatrices_MPISBAIJ(Mat,int,IS *,IS *,MatReuse,Mat **); 12 extern int MatGetValues_SeqSBAIJ(Mat,int,int *,int,int *,Scalar *); 13 extern int MatSetValues_SeqSBAIJ(Mat,int,int *,int,int *,Scalar *,InsertMode); 14 extern int MatSetValuesBlocked_SeqSBAIJ(Mat,int,int*,int,int*,Scalar*,InsertMode); 15 extern int MatGetRow_SeqSBAIJ(Mat,int,int*,int**,Scalar**); 16 extern int MatRestoreRow_SeqSBAIJ(Mat,int,int*,int**,Scalar**); 17 extern int MatPrintHelp_SeqSBAIJ(Mat); 18 extern int MatZeroRows_SeqSBAIJ(Mat,IS,Scalar*); 19 extern int MatZeroRows_SeqBAIJ(Mat,IS,Scalar *); 20 extern int MatGetRowMax_MPISBAIJ(Mat,Vec); 21 22 /* UGLY, ugly, ugly 23 When MatScalar == Scalar the function MatSetValuesBlocked_MPIBAIJ_MatScalar() does 24 not exist. Otherwise ..._MatScalar() takes matrix elements in single precision and 25 inserts them into the single precision data structure. The function MatSetValuesBlocked_MPIBAIJ() 26 converts the entries into single precision and then calls ..._MatScalar() to put them 27 into the single precision data structures. 28 */ 29 #if defined(PETSC_USE_MAT_SINGLE) 30 extern int MatSetValuesBlocked_SeqSBAIJ_MatScalar(Mat,int,int*,int,int*,MatScalar*,InsertMode); 31 extern int MatSetValues_MPISBAIJ_MatScalar(Mat,int,int*,int,int*,MatScalar*,InsertMode); 32 extern int MatSetValuesBlocked_MPISBAIJ_MatScalar(Mat,int,int*,int,int*,MatScalar*,InsertMode); 33 extern int MatSetValues_MPISBAIJ_HT_MatScalar(Mat,int,int*,int,int*,MatScalar*,InsertMode); 34 extern int MatSetValuesBlocked_MPISBAIJ_HT_MatScalar(Mat,int,int*,int,int*,MatScalar*,InsertMode); 35 #else 36 #define MatSetValuesBlocked_SeqSBAIJ_MatScalar MatSetValuesBlocked_SeqSBAIJ 37 #define MatSetValues_MPISBAIJ_MatScalar MatSetValues_MPISBAIJ 38 #define MatSetValuesBlocked_MPISBAIJ_MatScalar MatSetValuesBlocked_MPISBAIJ 39 #define MatSetValues_MPISBAIJ_HT_MatScalar MatSetValues_MPISBAIJ_HT 40 #define MatSetValuesBlocked_MPISBAIJ_HT_MatScalar MatSetValuesBlocked_MPISBAIJ_HT 41 #endif 42 43 EXTERN_C_BEGIN 44 #undef __FUNCT__ 45 #define __FUNCT__ "MatStoreValues_MPISBAIJ" 46 int MatStoreValues_MPISBAIJ(Mat mat) 47 { 48 Mat_MPISBAIJ *aij = (Mat_MPISBAIJ *)mat->data; 49 int ierr; 50 51 PetscFunctionBegin; 52 ierr = MatStoreValues(aij->A);CHKERRQ(ierr); 53 ierr = MatStoreValues(aij->B);CHKERRQ(ierr); 54 PetscFunctionReturn(0); 55 } 56 EXTERN_C_END 57 58 EXTERN_C_BEGIN 59 #undef __FUNCT__ 60 #define __FUNCT__ "MatRetrieveValues_MPISBAIJ" 61 int MatRetrieveValues_MPISBAIJ(Mat mat) 62 { 63 Mat_MPISBAIJ *aij = (Mat_MPISBAIJ *)mat->data; 64 int ierr; 65 66 PetscFunctionBegin; 67 ierr = MatRetrieveValues(aij->A);CHKERRQ(ierr); 68 ierr = MatRetrieveValues(aij->B);CHKERRQ(ierr); 69 PetscFunctionReturn(0); 70 } 71 EXTERN_C_END 72 73 /* 74 Local utility routine that creates a mapping from the global column 75 number to the local number in the off-diagonal part of the local 76 storage of the matrix. This is done in a non scable way since the 77 length of colmap equals the global matrix length. 78 */ 79 #undef __FUNCT__ 80 #define __FUNCT__ "CreateColmap_MPISBAIJ_Private" 81 static int CreateColmap_MPISBAIJ_Private(Mat mat) 82 { 83 PetscFunctionBegin; 84 SETERRQ(1,"Function not yet written for SBAIJ format"); 85 /* PetscFunctionReturn(0); */ 86 } 87 88 #define CHUNKSIZE 10 89 90 #define MatSetValues_SeqSBAIJ_A_Private(row,col,value,addv) \ 91 { \ 92 \ 93 brow = row/bs; \ 94 rp = aj + ai[brow]; ap = aa + bs2*ai[brow]; \ 95 rmax = aimax[brow]; nrow = ailen[brow]; \ 96 bcol = col/bs; \ 97 ridx = row % bs; cidx = col % bs; \ 98 low = 0; high = nrow; \ 99 while (high-low > 3) { \ 100 t = (low+high)/2; \ 101 if (rp[t] > bcol) high = t; \ 102 else low = t; \ 103 } \ 104 for (_i=low; _i<high; _i++) { \ 105 if (rp[_i] > bcol) break; \ 106 if (rp[_i] == bcol) { \ 107 bap = ap + bs2*_i + bs*cidx + ridx; \ 108 if (addv == ADD_VALUES) *bap += value; \ 109 else *bap = value; \ 110 goto a_noinsert; \ 111 } \ 112 } \ 113 if (a->nonew == 1) goto a_noinsert; \ 114 else if (a->nonew == -1) SETERRQ(PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new nonzero into matrix"); \ 115 if (nrow >= rmax) { \ 116 /* there is no extra room in row, therefore enlarge */ \ 117 int new_nz = ai[a->mbs] + CHUNKSIZE,len,*new_i,*new_j; \ 118 MatScalar *new_a; \ 119 \ 120 if (a->nonew == -2) SETERRQ(PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new nonzero in the matrix"); \ 121 \ 122 /* malloc new storage space */ \ 123 len = new_nz*(sizeof(int)+bs2*sizeof(MatScalar))+(a->mbs+1)*sizeof(int); \ 124 ierr = PetscMalloc(len,&new_a);CHKERRQ(ierr); \ 125 new_j = (int*)(new_a + bs2*new_nz); \ 126 new_i = new_j + new_nz; \ 127 \ 128 /* copy over old data into new slots */ \ 129 for (ii=0; ii<brow+1; ii++) {new_i[ii] = ai[ii];} \ 130 for (ii=brow+1; ii<a->mbs+1; ii++) {new_i[ii] = ai[ii]+CHUNKSIZE;} \ 131 ierr = PetscMemcpy(new_j,aj,(ai[brow]+nrow)*sizeof(int));CHKERRQ(ierr); \ 132 len = (new_nz - CHUNKSIZE - ai[brow] - nrow); \ 133 ierr = PetscMemcpy(new_j+ai[brow]+nrow+CHUNKSIZE,aj+ai[brow]+nrow,len*sizeof(int));CHKERRQ(ierr); \ 134 ierr = PetscMemcpy(new_a,aa,(ai[brow]+nrow)*bs2*sizeof(MatScalar));CHKERRQ(ierr); \ 135 ierr = PetscMemzero(new_a+bs2*(ai[brow]+nrow),bs2*CHUNKSIZE*sizeof(Scalar));CHKERRQ(ierr); \ 136 ierr = PetscMemcpy(new_a+bs2*(ai[brow]+nrow+CHUNKSIZE), \ 137 aa+bs2*(ai[brow]+nrow),bs2*len*sizeof(MatScalar));CHKERRQ(ierr); \ 138 /* free up old matrix storage */ \ 139 ierr = PetscFree(a->a);CHKERRQ(ierr); \ 140 if (!a->singlemalloc) { \ 141 ierr = PetscFree(a->i);CHKERRQ(ierr); \ 142 ierr = PetscFree(a->j);CHKERRQ(ierr);\ 143 } \ 144 aa = a->a = new_a; ai = a->i = new_i; aj = a->j = new_j; \ 145 a->singlemalloc = PETSC_TRUE; \ 146 \ 147 rp = aj + ai[brow]; ap = aa + bs2*ai[brow]; \ 148 rmax = aimax[brow] = aimax[brow] + CHUNKSIZE; \ 149 PetscLogObjectMemory(A,CHUNKSIZE*(sizeof(int) + bs2*sizeof(MatScalar))); \ 150 a->s_maxnz += bs2*CHUNKSIZE; \ 151 a->reallocs++; \ 152 a->s_nz++; \ 153 } \ 154 N = nrow++ - 1; \ 155 /* shift up all the later entries in this row */ \ 156 for (ii=N; ii>=_i; ii--) { \ 157 rp[ii+1] = rp[ii]; \ 158 ierr = PetscMemcpy(ap+bs2*(ii+1),ap+bs2*(ii),bs2*sizeof(MatScalar));CHKERRQ(ierr); \ 159 } \ 160 if (N>=_i) { ierr = PetscMemzero(ap+bs2*_i,bs2*sizeof(MatScalar));CHKERRQ(ierr); } \ 161 rp[_i] = bcol; \ 162 ap[bs2*_i + bs*cidx + ridx] = value; \ 163 a_noinsert:; \ 164 ailen[brow] = nrow; \ 165 } 166 #ifndef MatSetValues_SeqBAIJ_B_Private 167 #define MatSetValues_SeqSBAIJ_B_Private(row,col,value,addv) \ 168 { \ 169 brow = row/bs; \ 170 rp = bj + bi[brow]; ap = ba + bs2*bi[brow]; \ 171 rmax = bimax[brow]; nrow = bilen[brow]; \ 172 bcol = col/bs; \ 173 ridx = row % bs; cidx = col % bs; \ 174 low = 0; high = nrow; \ 175 while (high-low > 3) { \ 176 t = (low+high)/2; \ 177 if (rp[t] > bcol) high = t; \ 178 else low = t; \ 179 } \ 180 for (_i=low; _i<high; _i++) { \ 181 if (rp[_i] > bcol) break; \ 182 if (rp[_i] == bcol) { \ 183 bap = ap + bs2*_i + bs*cidx + ridx; \ 184 if (addv == ADD_VALUES) *bap += value; \ 185 else *bap = value; \ 186 goto b_noinsert; \ 187 } \ 188 } \ 189 if (b->nonew == 1) goto b_noinsert; \ 190 else if (b->nonew == -1) SETERRQ(PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new nonzero into matrix"); \ 191 if (nrow >= rmax) { \ 192 /* there is no extra room in row, therefore enlarge */ \ 193 int new_nz = bi[b->mbs] + CHUNKSIZE,len,*new_i,*new_j; \ 194 MatScalar *new_a; \ 195 \ 196 if (b->nonew == -2) SETERRQ(PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new nonzero in the matrix"); \ 197 \ 198 /* malloc new storage space */ \ 199 len = new_nz*(sizeof(int)+bs2*sizeof(MatScalar))+(b->mbs+1)*sizeof(int); \ 200 ierr = PetscMalloc(len,&new_a);CHKERRQ(ierr); \ 201 new_j = (int*)(new_a + bs2*new_nz); \ 202 new_i = new_j + new_nz; \ 203 \ 204 /* copy over old data into new slots */ \ 205 for (ii=0; ii<brow+1; ii++) {new_i[ii] = bi[ii];} \ 206 for (ii=brow+1; ii<b->mbs+1; ii++) {new_i[ii] = bi[ii]+CHUNKSIZE;} \ 207 ierr = PetscMemcpy(new_j,bj,(bi[brow]+nrow)*sizeof(int));CHKERRQ(ierr); \ 208 len = (new_nz - CHUNKSIZE - bi[brow] - nrow); \ 209 ierr = PetscMemcpy(new_j+bi[brow]+nrow+CHUNKSIZE,bj+bi[brow]+nrow,len*sizeof(int));CHKERRQ(ierr); \ 210 ierr = PetscMemcpy(new_a,ba,(bi[brow]+nrow)*bs2*sizeof(MatScalar));CHKERRQ(ierr); \ 211 ierr = PetscMemzero(new_a+bs2*(bi[brow]+nrow),bs2*CHUNKSIZE*sizeof(MatScalar));CHKERRQ(ierr); \ 212 ierr = PetscMemcpy(new_a+bs2*(bi[brow]+nrow+CHUNKSIZE), \ 213 ba+bs2*(bi[brow]+nrow),bs2*len*sizeof(MatScalar));CHKERRQ(ierr); \ 214 /* free up old matrix storage */ \ 215 ierr = PetscFree(b->a);CHKERRQ(ierr); \ 216 if (!b->singlemalloc) { \ 217 ierr = PetscFree(b->i);CHKERRQ(ierr); \ 218 ierr = PetscFree(b->j);CHKERRQ(ierr); \ 219 } \ 220 ba = b->a = new_a; bi = b->i = new_i; bj = b->j = new_j; \ 221 b->singlemalloc = PETSC_TRUE; \ 222 \ 223 rp = bj + bi[brow]; ap = ba + bs2*bi[brow]; \ 224 rmax = bimax[brow] = bimax[brow] + CHUNKSIZE; \ 225 PetscLogObjectMemory(B,CHUNKSIZE*(sizeof(int) + bs2*sizeof(MatScalar))); \ 226 b->maxnz += bs2*CHUNKSIZE; \ 227 b->reallocs++; \ 228 b->nz++; \ 229 } \ 230 N = nrow++ - 1; \ 231 /* shift up all the later entries in this row */ \ 232 for (ii=N; ii>=_i; ii--) { \ 233 rp[ii+1] = rp[ii]; \ 234 ierr = PetscMemcpy(ap+bs2*(ii+1),ap+bs2*(ii),bs2*sizeof(MatScalar));CHKERRQ(ierr); \ 235 } \ 236 if (N>=_i) { ierr = PetscMemzero(ap+bs2*_i,bs2*sizeof(MatScalar));CHKERRQ(ierr);} \ 237 rp[_i] = bcol; \ 238 ap[bs2*_i + bs*cidx + ridx] = value; \ 239 b_noinsert:; \ 240 bilen[brow] = nrow; \ 241 } 242 #endif 243 244 #if defined(PETSC_USE_MAT_SINGLE) 245 #undef __FUNCT__ 246 #define __FUNCT__ "MatSetValues_MPISBAIJ" 247 int MatSetValues_MPISBAIJ(Mat mat,int m,int *im,int n,int *in,Scalar *v,InsertMode addv) 248 { 249 Mat_MPISBAIJ *b = (Mat_MPISBAIJ*)mat->data; 250 int ierr,i,N = m*n; 251 MatScalar *vsingle; 252 253 PetscFunctionBegin; 254 if (N > b->setvalueslen) { 255 if (b->setvaluescopy) {ierr = PetscFree(b->setvaluescopy);CHKERRQ(ierr);} 256 ierr = PetscMalloc(N*sizeof(MatScalar),&b->setvaluescopy);CHKERRQ(ierr); 257 b->setvalueslen = N; 258 } 259 vsingle = b->setvaluescopy; 260 261 for (i=0; i<N; i++) { 262 vsingle[i] = v[i]; 263 } 264 ierr = MatSetValues_MPISBAIJ_MatScalar(mat,m,im,n,in,vsingle,addv);CHKERRQ(ierr); 265 PetscFunctionReturn(0); 266 } 267 268 #undef __FUNCT__ 269 #define __FUNCT__ "MatSetValuesBlocked_MPISBAIJ" 270 int MatSetValuesBlocked_MPISBAIJ(Mat mat,int m,int *im,int n,int *in,Scalar *v,InsertMode addv) 271 { 272 Mat_MPIBAIJ *b = (Mat_MPIBAIJ*)mat->data; 273 int ierr,i,N = m*n*b->bs2; 274 MatScalar *vsingle; 275 276 PetscFunctionBegin; 277 if (N > b->setvalueslen) { 278 if (b->setvaluescopy) {ierr = PetscFree(b->setvaluescopy);CHKERRQ(ierr);} 279 ierr = PetscMalloc(N*sizeof(MatScalar),&b->setvaluescopy);CHKERRQ(ierr); 280 b->setvalueslen = N; 281 } 282 vsingle = b->setvaluescopy; 283 for (i=0; i<N; i++) { 284 vsingle[i] = v[i]; 285 } 286 ierr = MatSetValuesBlocked_MPISBAIJ_MatScalar(mat,m,im,n,in,vsingle,addv);CHKERRQ(ierr); 287 PetscFunctionReturn(0); 288 } 289 290 #undef __FUNCT__ 291 #define __FUNCT__ "MatSetValues_MPISBAIJ_HT" 292 int MatSetValues_MPISBAIJ_HT(Mat mat,int m,int *im,int n,int *in,Scalar *v,InsertMode addv) 293 { 294 Mat_MPIBAIJ *b = (Mat_MPIBAIJ*)mat->data; 295 int ierr,i,N = m*n; 296 MatScalar *vsingle; 297 298 PetscFunctionBegin; 299 SETERRQ(1,"Function not yet written for SBAIJ format"); 300 /* PetscFunctionReturn(0); */ 301 } 302 303 #undef __FUNCT__ 304 #define __FUNCT__ "MatSetValuesBlocked_MPISBAIJ_HT" 305 int MatSetValuesBlocked_MPISBAIJ_HT(Mat mat,int m,int *im,int n,int *in,Scalar *v,InsertMode addv) 306 { 307 Mat_MPIBAIJ *b = (Mat_MPIBAIJ*)mat->data; 308 int ierr,i,N = m*n*b->bs2; 309 MatScalar *vsingle; 310 311 PetscFunctionBegin; 312 SETERRQ(1,"Function not yet written for SBAIJ format"); 313 /* PetscFunctionReturn(0); */ 314 } 315 #endif 316 317 /* Only add/insert a(i,j) with i<=j (blocks). 318 Any a(i,j) with i>j input by user is ingored. 319 */ 320 #undef __FUNCT__ 321 #define __FUNCT__ "MatSetValues_MPIBAIJ" 322 int MatSetValues_MPISBAIJ_MatScalar(Mat mat,int m,int *im,int n,int *in,MatScalar *v,InsertMode addv) 323 { 324 Mat_MPISBAIJ *baij = (Mat_MPISBAIJ*)mat->data; 325 MatScalar value; 326 PetscTruth roworiented = baij->roworiented; 327 int ierr,i,j,row,col; 328 int rstart_orig=baij->rstart_bs; 329 int rend_orig=baij->rend_bs,cstart_orig=baij->cstart_bs; 330 int cend_orig=baij->cend_bs,bs=baij->bs; 331 332 /* Some Variables required in the macro */ 333 Mat A = baij->A; 334 Mat_SeqSBAIJ *a = (Mat_SeqSBAIJ*)(A)->data; 335 int *aimax=a->imax,*ai=a->i,*ailen=a->ilen,*aj=a->j; 336 MatScalar *aa=a->a; 337 338 Mat B = baij->B; 339 Mat_SeqBAIJ *b = (Mat_SeqBAIJ*)(B)->data; 340 int *bimax=b->imax,*bi=b->i,*bilen=b->ilen,*bj=b->j; 341 MatScalar *ba=b->a; 342 343 int *rp,ii,nrow,_i,rmax,N,brow,bcol; 344 int low,high,t,ridx,cidx,bs2=a->bs2; 345 MatScalar *ap,*bap; 346 347 /* for stash */ 348 int n_loc, *in_loc=0; 349 MatScalar *v_loc=0; 350 351 PetscFunctionBegin; 352 353 if(!baij->donotstash){ 354 ierr = PetscMalloc(n*sizeof(int),&in_loc);CHKERRQ(ierr); 355 ierr = PetscMalloc(n*sizeof(MatScalar),&v_loc);CHKERRQ(ierr); 356 } 357 358 for (i=0; i<m; i++) { 359 if (im[i] < 0) continue; 360 #if defined(PETSC_USE_BOPT_g) 361 if (im[i] >= mat->M) SETERRQ(PETSC_ERR_ARG_OUTOFRANGE,"Row too large"); 362 #endif 363 if (im[i] >= rstart_orig && im[i] < rend_orig) { /* this processor entry */ 364 row = im[i] - rstart_orig; /* local row index */ 365 for (j=0; j<n; j++) { 366 if (im[i]/bs > in[j]/bs) continue; /* ignore lower triangular blocks */ 367 if (in[j] >= cstart_orig && in[j] < cend_orig){ /* diag entry (A) */ 368 col = in[j] - cstart_orig; /* local col index */ 369 brow = row/bs; bcol = col/bs; 370 if (brow > bcol) continue; /* ignore lower triangular blocks of A */ 371 if (roworiented) value = v[i*n+j]; else value = v[i+j*m]; 372 MatSetValues_SeqSBAIJ_A_Private(row,col,value,addv); 373 /* ierr = MatSetValues_SeqBAIJ(baij->A,1,&row,1,&col,&value,addv);CHKERRQ(ierr); */ 374 } else if (in[j] < 0) continue; 375 #if defined(PETSC_USE_BOPT_g) 376 else if (in[j] >= mat->N) {SETERRQ(PETSC_ERR_ARG_OUTOFRANGE,"Col too large");} 377 #endif 378 else { /* off-diag entry (B) */ 379 if (mat->was_assembled) { 380 if (!baij->colmap) { 381 ierr = CreateColmap_MPISBAIJ_Private(mat);CHKERRQ(ierr); 382 } 383 #if defined (PETSC_USE_CTABLE) 384 ierr = PetscTableFind(baij->colmap,in[j]/bs + 1,&col);CHKERRQ(ierr); 385 col = col - 1 + in[j]%bs; 386 #else 387 col = baij->colmap[in[j]/bs] - 1 + in[j]%bs; 388 #endif 389 if (col < 0 && !((Mat_SeqSBAIJ*)(baij->A->data))->nonew) { 390 ierr = DisAssemble_MPISBAIJ(mat);CHKERRQ(ierr); 391 col = in[j]; 392 /* Reinitialize the variables required by MatSetValues_SeqBAIJ_B_Private() */ 393 B = baij->B; 394 b = (Mat_SeqBAIJ*)(B)->data; 395 bimax=b->imax;bi=b->i;bilen=b->ilen;bj=b->j; 396 ba=b->a; 397 } 398 } else col = in[j]; 399 if (roworiented) value = v[i*n+j]; else value = v[i+j*m]; 400 MatSetValues_SeqSBAIJ_B_Private(row,col,value,addv); 401 /* ierr = MatSetValues_SeqBAIJ(baij->B,1,&row,1,&col,&value,addv);CHKERRQ(ierr); */ 402 } 403 } 404 } else { /* off processor entry */ 405 if (!baij->donotstash) { 406 n_loc = 0; 407 for (j=0; j<n; j++){ 408 if (im[i]/bs > in[j]/bs) continue; /* ignore lower triangular blocks */ 409 in_loc[n_loc] = in[j]; 410 if (roworiented) { 411 v_loc[n_loc] = v[i*n+j]; 412 } else { 413 v_loc[n_loc] = v[j*m+i]; 414 } 415 n_loc++; 416 } 417 ierr = MatStashValuesRow_Private(&mat->stash,im[i],n_loc,in_loc,v_loc);CHKERRQ(ierr); 418 } 419 } 420 } 421 422 if(!baij->donotstash){ 423 ierr = PetscFree(in_loc);CHKERRQ(ierr); 424 ierr = PetscFree(v_loc);CHKERRQ(ierr); 425 } 426 PetscFunctionReturn(0); 427 } 428 429 #undef __FUNCT__ 430 #define __FUNCT__ "MatSetValuesBlocked_MPISBAIJ" 431 int MatSetValuesBlocked_MPISBAIJ_MatScalar(Mat mat,int m,int *im,int n,int *in,MatScalar *v,InsertMode addv) 432 { 433 PetscFunctionBegin; 434 SETERRQ(1,"Function not yet written for SBAIJ format"); 435 /* PetscFunctionReturn(0); */ 436 } 437 438 #define HASH_KEY 0.6180339887 439 #define HASH(size,key,tmp) (tmp = (key)*HASH_KEY,(int)((size)*(tmp-(int)tmp))) 440 /* #define HASH(size,key) ((int)((size)*fmod(((key)*HASH_KEY),1))) */ 441 /* #define HASH(size,key,tmp) ((int)((size)*fmod(((key)*HASH_KEY),1))) */ 442 #undef __FUNCT__ 443 #define __FUNCT__ "MatSetValues_MPISBAIJ_HT_MatScalar" 444 int MatSetValues_MPISBAIJ_HT_MatScalar(Mat mat,int m,int *im,int n,int *in,MatScalar *v,InsertMode addv) 445 { 446 PetscFunctionBegin; 447 SETERRQ(1,"Function not yet written for SBAIJ format"); 448 /* PetscFunctionReturn(0); */ 449 } 450 451 #undef __FUNCT__ 452 #define __FUNCT__ "MatSetValuesBlocked_MPISBAIJ_HT_MatScalar" 453 int MatSetValuesBlocked_MPISBAIJ_HT_MatScalar(Mat mat,int m,int *im,int n,int *in,MatScalar *v,InsertMode addv) 454 { 455 PetscFunctionBegin; 456 SETERRQ(1,"Function not yet written for SBAIJ format"); 457 /* PetscFunctionReturn(0); */ 458 } 459 460 #undef __FUNCT__ 461 #define __FUNCT__ "MatGetValues_MPISBAIJ" 462 int MatGetValues_MPISBAIJ(Mat mat,int m,int *idxm,int n,int *idxn,Scalar *v) 463 { 464 Mat_MPISBAIJ *baij = (Mat_MPISBAIJ*)mat->data; 465 int bs=baij->bs,ierr,i,j,bsrstart = baij->rstart*bs,bsrend = baij->rend*bs; 466 int bscstart = baij->cstart*bs,bscend = baij->cend*bs,row,col,data; 467 468 PetscFunctionBegin; 469 for (i=0; i<m; i++) { 470 if (idxm[i] < 0) SETERRQ(PETSC_ERR_ARG_OUTOFRANGE,"Negative row"); 471 if (idxm[i] >= mat->M) SETERRQ(PETSC_ERR_ARG_OUTOFRANGE,"Row too large"); 472 if (idxm[i] >= bsrstart && idxm[i] < bsrend) { 473 row = idxm[i] - bsrstart; 474 for (j=0; j<n; j++) { 475 if (idxn[j] < 0) SETERRQ(PETSC_ERR_ARG_OUTOFRANGE,"Negative column"); 476 if (idxn[j] >= mat->N) SETERRQ(PETSC_ERR_ARG_OUTOFRANGE,"Column too large"); 477 if (idxn[j] >= bscstart && idxn[j] < bscend){ 478 col = idxn[j] - bscstart; 479 ierr = MatGetValues_SeqSBAIJ(baij->A,1,&row,1,&col,v+i*n+j);CHKERRQ(ierr); 480 } else { 481 if (!baij->colmap) { 482 ierr = CreateColmap_MPISBAIJ_Private(mat);CHKERRQ(ierr); 483 } 484 #if defined (PETSC_USE_CTABLE) 485 ierr = PetscTableFind(baij->colmap,idxn[j]/bs+1,&data);CHKERRQ(ierr); 486 data --; 487 #else 488 data = baij->colmap[idxn[j]/bs]-1; 489 #endif 490 if((data < 0) || (baij->garray[data/bs] != idxn[j]/bs)) *(v+i*n+j) = 0.0; 491 else { 492 col = data + idxn[j]%bs; 493 ierr = MatGetValues_SeqSBAIJ(baij->B,1,&row,1,&col,v+i*n+j);CHKERRQ(ierr); 494 } 495 } 496 } 497 } else { 498 SETERRQ(PETSC_ERR_SUP,"Only local values currently supported"); 499 } 500 } 501 PetscFunctionReturn(0); 502 } 503 504 #undef __FUNCT__ 505 #define __FUNCT__ "MatNorm_MPISBAIJ" 506 int MatNorm_MPISBAIJ(Mat mat,NormType type,PetscReal *norm) 507 { 508 Mat_MPISBAIJ *baij = (Mat_MPISBAIJ*)mat->data; 509 /* Mat_SeqSBAIJ *amat = (Mat_SeqSBAIJ*)baij->A->data; */ 510 /* Mat_SeqBAIJ *bmat = (Mat_SeqBAIJ*)baij->B->data; */ 511 int ierr; 512 PetscReal sum[2],*lnorm2; 513 514 PetscFunctionBegin; 515 if (baij->size == 1) { 516 ierr = MatNorm(baij->A,type,norm);CHKERRQ(ierr); 517 } else { 518 if (type == NORM_FROBENIUS) { 519 ierr = PetscMalloc(2*sizeof(double),&lnorm2);CHKERRQ(ierr); 520 ierr = MatNorm(baij->A,type,lnorm2);CHKERRQ(ierr); 521 *lnorm2 = (*lnorm2)*(*lnorm2); lnorm2++; /* squar power of norm(A) */ 522 ierr = MatNorm(baij->B,type,lnorm2);CHKERRQ(ierr); 523 *lnorm2 = (*lnorm2)*(*lnorm2); lnorm2--; /* squar power of norm(B) */ 524 /* 525 ierr = MPI_Comm_rank(PETSC_COMM_WORLD,&rank);CHKERRQ(ierr); 526 PetscSynchronizedPrintf(PETSC_COMM_WORLD,"[%d], lnorm2=%g, %g\n",rank,lnorm2[0],lnorm2[1]); 527 */ 528 ierr = MPI_Allreduce(lnorm2,&sum,2,MPI_DOUBLE,MPI_SUM,mat->comm);CHKERRQ(ierr); 529 /* 530 PetscSynchronizedPrintf(PETSC_COMM_WORLD,"[%d], sum=%g, %g\n",rank,sum[0],sum[1]); 531 PetscSynchronizedFlush(PETSC_COMM_WORLD); */ 532 533 *norm = sqrt(sum[0] + 2*sum[1]); 534 ierr = PetscFree(lnorm2);CHKERRQ(ierr); 535 } else { 536 SETERRQ(PETSC_ERR_SUP,"No support for this norm yet"); 537 } 538 } 539 PetscFunctionReturn(0); 540 } 541 542 /* 543 Creates the hash table, and sets the table 544 This table is created only once. 545 If new entried need to be added to the matrix 546 then the hash table has to be destroyed and 547 recreated. 548 */ 549 #undef __FUNCT__ 550 #define __FUNCT__ "MatCreateHashTable_MPISBAIJ_Private" 551 int MatCreateHashTable_MPISBAIJ_Private(Mat mat,PetscReal factor) 552 { 553 PetscFunctionBegin; 554 SETERRQ(1,"Function not yet written for SBAIJ format"); 555 /* PetscFunctionReturn(0); */ 556 } 557 558 #undef __FUNCT__ 559 #define __FUNCT__ "MatAssemblyBegin_MPISBAIJ" 560 int MatAssemblyBegin_MPISBAIJ(Mat mat,MatAssemblyType mode) 561 { 562 Mat_MPISBAIJ *baij = (Mat_MPISBAIJ*)mat->data; 563 int ierr,nstash,reallocs; 564 InsertMode addv; 565 566 PetscFunctionBegin; 567 if (baij->donotstash) { 568 PetscFunctionReturn(0); 569 } 570 571 /* make sure all processors are either in INSERTMODE or ADDMODE */ 572 ierr = MPI_Allreduce(&mat->insertmode,&addv,1,MPI_INT,MPI_BOR,mat->comm);CHKERRQ(ierr); 573 if (addv == (ADD_VALUES|INSERT_VALUES)) { 574 SETERRQ(PETSC_ERR_ARG_WRONGSTATE,"Some processors inserted others added"); 575 } 576 mat->insertmode = addv; /* in case this processor had no cache */ 577 578 ierr = MatStashScatterBegin_Private(&mat->stash,baij->rowners_bs);CHKERRQ(ierr); 579 ierr = MatStashScatterBegin_Private(&mat->bstash,baij->rowners);CHKERRQ(ierr); 580 ierr = MatStashGetInfo_Private(&mat->stash,&nstash,&reallocs);CHKERRQ(ierr); 581 PetscLogInfo(0,"MatAssemblyBegin_MPISBAIJ:Stash has %d entries,uses %d mallocs.\n",nstash,reallocs); 582 ierr = MatStashGetInfo_Private(&mat->stash,&nstash,&reallocs);CHKERRQ(ierr); 583 PetscLogInfo(0,"MatAssemblyBegin_MPISBAIJ:Block-Stash has %d entries, uses %d mallocs.\n",nstash,reallocs); 584 PetscFunctionReturn(0); 585 } 586 587 #undef __FUNCT__ 588 #define __FUNCT__ "MatAssemblyEnd_MPISBAIJ" 589 int MatAssemblyEnd_MPISBAIJ(Mat mat,MatAssemblyType mode) 590 { 591 Mat_MPISBAIJ *baij=(Mat_MPISBAIJ*)mat->data; 592 Mat_SeqSBAIJ *a=(Mat_SeqSBAIJ*)baij->A->data; 593 Mat_SeqBAIJ *b=(Mat_SeqBAIJ*)baij->B->data; 594 int i,j,rstart,ncols,n,ierr,flg,bs2=baij->bs2; 595 int *row,*col,other_disassembled; 596 PetscTruth r1,r2,r3; 597 MatScalar *val; 598 InsertMode addv = mat->insertmode; 599 /* int rank;*/ 600 601 PetscFunctionBegin; 602 /* remove 2 line below later */ 603 /*ierr = MPI_Comm_rank(PETSC_COMM_WORLD, &rank);CHKERRQ(ierr); */ 604 605 if (!baij->donotstash) { 606 while (1) { 607 ierr = MatStashScatterGetMesg_Private(&mat->stash,&n,&row,&col,&val,&flg);CHKERRQ(ierr); 608 /* 609 PetscSynchronizedPrintf(PETSC_COMM_WORLD,"[%d]: in AssemblyEnd, stash, flg=%d\n",rank,flg); 610 PetscSynchronizedFlush(PETSC_COMM_WORLD); 611 */ 612 if (!flg) break; 613 614 for (i=0; i<n;) { 615 /* Now identify the consecutive vals belonging to the same row */ 616 for (j=i,rstart=row[j]; j<n; j++) { if (row[j] != rstart) break; } 617 if (j < n) ncols = j-i; 618 else ncols = n-i; 619 /* Now assemble all these values with a single function call */ 620 ierr = MatSetValues_MPISBAIJ_MatScalar(mat,1,row+i,ncols,col+i,val+i,addv);CHKERRQ(ierr); 621 i = j; 622 } 623 } 624 ierr = MatStashScatterEnd_Private(&mat->stash);CHKERRQ(ierr); 625 /* Now process the block-stash. Since the values are stashed column-oriented, 626 set the roworiented flag to column oriented, and after MatSetValues() 627 restore the original flags */ 628 r1 = baij->roworiented; 629 r2 = a->roworiented; 630 r3 = b->roworiented; 631 baij->roworiented = PETSC_FALSE; 632 a->roworiented = PETSC_FALSE; 633 b->roworiented = PETSC_FALSE; 634 while (1) { 635 ierr = MatStashScatterGetMesg_Private(&mat->bstash,&n,&row,&col,&val,&flg);CHKERRQ(ierr); 636 if (!flg) break; 637 638 for (i=0; i<n;) { 639 /* Now identify the consecutive vals belonging to the same row */ 640 for (j=i,rstart=row[j]; j<n; j++) { if (row[j] != rstart) break; } 641 if (j < n) ncols = j-i; 642 else ncols = n-i; 643 ierr = MatSetValuesBlocked_MPISBAIJ_MatScalar(mat,1,row+i,ncols,col+i,val+i*bs2,addv);CHKERRQ(ierr); 644 i = j; 645 } 646 } 647 ierr = MatStashScatterEnd_Private(&mat->bstash);CHKERRQ(ierr); 648 baij->roworiented = r1; 649 a->roworiented = r2; 650 b->roworiented = r3; 651 } 652 653 ierr = MatAssemblyBegin(baij->A,mode);CHKERRQ(ierr); 654 ierr = MatAssemblyEnd(baij->A,mode);CHKERRQ(ierr); 655 656 /* determine if any processor has disassembled, if so we must 657 also disassemble ourselfs, in order that we may reassemble. */ 658 /* 659 if nonzero structure of submatrix B cannot change then we know that 660 no processor disassembled thus we can skip this stuff 661 */ 662 if (!((Mat_SeqBAIJ*)baij->B->data)->nonew) { 663 ierr = MPI_Allreduce(&mat->was_assembled,&other_disassembled,1,MPI_INT,MPI_PROD,mat->comm);CHKERRQ(ierr); 664 if (mat->was_assembled && !other_disassembled) { 665 ierr = DisAssemble_MPISBAIJ(mat);CHKERRQ(ierr); 666 } 667 } 668 669 if (!mat->was_assembled && mode == MAT_FINAL_ASSEMBLY) { 670 ierr = MatSetUpMultiply_MPISBAIJ(mat);CHKERRQ(ierr); 671 } 672 ierr = MatAssemblyBegin(baij->B,mode);CHKERRQ(ierr); 673 ierr = MatAssemblyEnd(baij->B,mode);CHKERRQ(ierr); 674 675 #if defined(PETSC_USE_BOPT_g) 676 if (baij->ht && mode== MAT_FINAL_ASSEMBLY) { 677 PetscLogInfo(0,"MatAssemblyEnd_MPISBAIJ:Average Hash Table Search in MatSetValues = %5.2f\n",((double)baij->ht_total_ct)/baij->ht_insert_ct); 678 baij->ht_total_ct = 0; 679 baij->ht_insert_ct = 0; 680 } 681 #endif 682 if (baij->ht_flag && !baij->ht && mode == MAT_FINAL_ASSEMBLY) { 683 ierr = MatCreateHashTable_MPISBAIJ_Private(mat,baij->ht_fact);CHKERRQ(ierr); 684 mat->ops->setvalues = MatSetValues_MPISBAIJ_HT; 685 mat->ops->setvaluesblocked = MatSetValuesBlocked_MPISBAIJ_HT; 686 } 687 688 if (baij->rowvalues) { 689 ierr = PetscFree(baij->rowvalues);CHKERRQ(ierr); 690 baij->rowvalues = 0; 691 } 692 PetscFunctionReturn(0); 693 } 694 695 #undef __FUNCT__ 696 #define __FUNCT__ "MatView_MPISBAIJ_ASCIIorDraworSocket" 697 static int MatView_MPISBAIJ_ASCIIorDraworSocket(Mat mat,PetscViewer viewer) 698 { 699 Mat_MPISBAIJ *baij = (Mat_MPISBAIJ*)mat->data; 700 int ierr,bs = baij->bs,size = baij->size,rank = baij->rank; 701 PetscTruth isascii,isdraw; 702 PetscViewer sviewer; 703 PetscViewerFormat format; 704 705 PetscFunctionBegin; 706 ierr = PetscTypeCompare((PetscObject)viewer,PETSC_VIEWER_ASCII,&isascii);CHKERRQ(ierr); 707 ierr = PetscTypeCompare((PetscObject)viewer,PETSC_VIEWER_DRAW,&isdraw);CHKERRQ(ierr); 708 if (isascii) { 709 ierr = PetscViewerGetFormat(viewer,&format);CHKERRQ(ierr); 710 if (format == PETSC_VIEWER_ASCII_INFO_LONG) { 711 MatInfo info; 712 ierr = MPI_Comm_rank(mat->comm,&rank);CHKERRQ(ierr); 713 ierr = MatGetInfo(mat,MAT_LOCAL,&info);CHKERRQ(ierr); 714 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"[%d] Local rows %d nz %d nz alloced %d bs %d mem %d\n", 715 rank,mat->m,(int)info.nz_used*bs,(int)info.nz_allocated*bs, 716 baij->bs,(int)info.memory);CHKERRQ(ierr); 717 ierr = MatGetInfo(baij->A,MAT_LOCAL,&info);CHKERRQ(ierr); 718 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"[%d] on-diagonal part: nz %d \n",rank,(int)info.nz_used*bs);CHKERRQ(ierr); 719 ierr = MatGetInfo(baij->B,MAT_LOCAL,&info);CHKERRQ(ierr); 720 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"[%d] off-diagonal part: nz %d \n",rank,(int)info.nz_used*bs);CHKERRQ(ierr); 721 ierr = PetscViewerFlush(viewer);CHKERRQ(ierr); 722 ierr = VecScatterView(baij->Mvctx,viewer);CHKERRQ(ierr); 723 PetscFunctionReturn(0); 724 } else if (format == PETSC_VIEWER_ASCII_INFO) { 725 ierr = PetscViewerASCIIPrintf(viewer," block size is %d\n",bs);CHKERRQ(ierr); 726 PetscFunctionReturn(0); 727 } 728 } 729 730 if (isdraw) { 731 PetscDraw draw; 732 PetscTruth isnull; 733 ierr = PetscViewerDrawGetDraw(viewer,0,&draw);CHKERRQ(ierr); 734 ierr = PetscDrawIsNull(draw,&isnull);CHKERRQ(ierr); if (isnull) PetscFunctionReturn(0); 735 } 736 737 if (size == 1) { 738 ierr = PetscObjectSetName((PetscObject)baij->A,mat->name);CHKERRQ(ierr); 739 ierr = MatView(baij->A,viewer);CHKERRQ(ierr); 740 } else { 741 /* assemble the entire matrix onto first processor. */ 742 Mat A; 743 Mat_SeqSBAIJ *Aloc; 744 Mat_SeqBAIJ *Bloc; 745 int M = mat->M,N = mat->N,*ai,*aj,col,i,j,k,*rvals,mbs = baij->mbs; 746 MatScalar *a; 747 748 if (!rank) { 749 ierr = MatCreateMPISBAIJ(mat->comm,baij->bs,M,N,M,N,0,PETSC_NULL,0,PETSC_NULL,&A);CHKERRQ(ierr); 750 } else { 751 ierr = MatCreateMPISBAIJ(mat->comm,baij->bs,0,0,M,N,0,PETSC_NULL,0,PETSC_NULL,&A);CHKERRQ(ierr); 752 } 753 PetscLogObjectParent(mat,A); 754 755 /* copy over the A part */ 756 Aloc = (Mat_SeqSBAIJ*)baij->A->data; 757 ai = Aloc->i; aj = Aloc->j; a = Aloc->a; 758 ierr = PetscMalloc(bs*sizeof(int),&rvals);CHKERRQ(ierr); 759 760 for (i=0; i<mbs; i++) { 761 rvals[0] = bs*(baij->rstart + i); 762 for (j=1; j<bs; j++) { rvals[j] = rvals[j-1] + 1; } 763 for (j=ai[i]; j<ai[i+1]; j++) { 764 col = (baij->cstart+aj[j])*bs; 765 for (k=0; k<bs; k++) { 766 ierr = MatSetValues_MPISBAIJ_MatScalar(A,bs,rvals,1,&col,a,INSERT_VALUES);CHKERRQ(ierr); 767 col++; a += bs; 768 } 769 } 770 } 771 /* copy over the B part */ 772 Bloc = (Mat_SeqBAIJ*)baij->B->data; 773 ai = Bloc->i; aj = Bloc->j; a = Bloc->a; 774 for (i=0; i<mbs; i++) { 775 rvals[0] = bs*(baij->rstart + i); 776 for (j=1; j<bs; j++) { rvals[j] = rvals[j-1] + 1; } 777 for (j=ai[i]; j<ai[i+1]; j++) { 778 col = baij->garray[aj[j]]*bs; 779 for (k=0; k<bs; k++) { 780 ierr = MatSetValues_MPISBAIJ_MatScalar(A,bs,rvals,1,&col,a,INSERT_VALUES);CHKERRQ(ierr); 781 col++; a += bs; 782 } 783 } 784 } 785 ierr = PetscFree(rvals);CHKERRQ(ierr); 786 ierr = MatAssemblyBegin(A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 787 ierr = MatAssemblyEnd(A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 788 /* 789 Everyone has to call to draw the matrix since the graphics waits are 790 synchronized across all processors that share the PetscDraw object 791 */ 792 ierr = PetscViewerGetSingleton(viewer,&sviewer);CHKERRQ(ierr); 793 if (!rank) { 794 ierr = PetscObjectSetName((PetscObject)((Mat_MPISBAIJ*)(A->data))->A,mat->name);CHKERRQ(ierr); 795 ierr = MatView(((Mat_MPISBAIJ*)(A->data))->A,sviewer);CHKERRQ(ierr); 796 } 797 ierr = PetscViewerRestoreSingleton(viewer,&sviewer);CHKERRQ(ierr); 798 ierr = MatDestroy(A);CHKERRQ(ierr); 799 } 800 PetscFunctionReturn(0); 801 } 802 803 #undef __FUNCT__ 804 #define __FUNCT__ "MatView_MPISBAIJ" 805 int MatView_MPISBAIJ(Mat mat,PetscViewer viewer) 806 { 807 int ierr; 808 PetscTruth isascii,isdraw,issocket,isbinary; 809 810 PetscFunctionBegin; 811 ierr = PetscTypeCompare((PetscObject)viewer,PETSC_VIEWER_ASCII,&isascii);CHKERRQ(ierr); 812 ierr = PetscTypeCompare((PetscObject)viewer,PETSC_VIEWER_DRAW,&isdraw);CHKERRQ(ierr); 813 ierr = PetscTypeCompare((PetscObject)viewer,PETSC_VIEWER_SOCKET,&issocket);CHKERRQ(ierr); 814 ierr = PetscTypeCompare((PetscObject)viewer,PETSC_VIEWER_BINARY,&isbinary);CHKERRQ(ierr); 815 if (isascii || isdraw || issocket || isbinary) { 816 ierr = MatView_MPISBAIJ_ASCIIorDraworSocket(mat,viewer);CHKERRQ(ierr); 817 } else { 818 SETERRQ1(1,"Viewer type %s not supported by MPISBAIJ matrices",((PetscObject)viewer)->type_name); 819 } 820 PetscFunctionReturn(0); 821 } 822 823 #undef __FUNCT__ 824 #define __FUNCT__ "MatDestroy_MPISBAIJ" 825 int MatDestroy_MPISBAIJ(Mat mat) 826 { 827 Mat_MPISBAIJ *baij = (Mat_MPISBAIJ*)mat->data; 828 int ierr; 829 830 PetscFunctionBegin; 831 #if defined(PETSC_USE_LOG) 832 PetscLogObjectState((PetscObject)mat,"Rows=%d,Cols=%d",mat->M,mat->N); 833 #endif 834 ierr = MatStashDestroy_Private(&mat->stash);CHKERRQ(ierr); 835 ierr = MatStashDestroy_Private(&mat->bstash);CHKERRQ(ierr); 836 ierr = PetscFree(baij->rowners);CHKERRQ(ierr); 837 ierr = MatDestroy(baij->A);CHKERRQ(ierr); 838 ierr = MatDestroy(baij->B);CHKERRQ(ierr); 839 #if defined (PETSC_USE_CTABLE) 840 if (baij->colmap) {ierr = PetscTableDelete(baij->colmap);CHKERRQ(ierr);} 841 #else 842 if (baij->colmap) {ierr = PetscFree(baij->colmap);CHKERRQ(ierr);} 843 #endif 844 if (baij->garray) {ierr = PetscFree(baij->garray);CHKERRQ(ierr);} 845 if (baij->lvec) {ierr = VecDestroy(baij->lvec);CHKERRQ(ierr);} 846 if (baij->Mvctx) {ierr = VecScatterDestroy(baij->Mvctx);CHKERRQ(ierr);} 847 if (baij->rowvalues) {ierr = PetscFree(baij->rowvalues);CHKERRQ(ierr);} 848 if (baij->barray) {ierr = PetscFree(baij->barray);CHKERRQ(ierr);} 849 if (baij->hd) {ierr = PetscFree(baij->hd);CHKERRQ(ierr);} 850 #if defined(PETSC_USE_MAT_SINGLE) 851 if (baij->setvaluescopy) {ierr = PetscFree(baij->setvaluescopy);CHKERRQ(ierr);} 852 #endif 853 ierr = PetscFree(baij);CHKERRQ(ierr); 854 PetscFunctionReturn(0); 855 } 856 857 #undef __FUNCT__ 858 #define __FUNCT__ "MatMult_MPISBAIJ" 859 int MatMult_MPISBAIJ(Mat A,Vec xx,Vec yy) 860 { 861 Mat_MPISBAIJ *a = (Mat_MPISBAIJ*)A->data; 862 int ierr,nt; 863 864 PetscFunctionBegin; 865 ierr = VecGetLocalSize(xx,&nt);CHKERRQ(ierr); 866 if (nt != A->n) { 867 SETERRQ(PETSC_ERR_ARG_SIZ,"Incompatible partition of A and xx"); 868 } 869 ierr = VecGetLocalSize(yy,&nt);CHKERRQ(ierr); 870 if (nt != A->m) { 871 SETERRQ(PETSC_ERR_ARG_SIZ,"Incompatible parition of A and yy"); 872 } 873 874 ierr = VecScatterBegin(xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD,a->Mvctx);CHKERRQ(ierr); 875 /* do diagonal part */ 876 ierr = (*a->A->ops->mult)(a->A,xx,yy);CHKERRQ(ierr); 877 /* do supperdiagonal part */ 878 ierr = VecScatterEnd(xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD,a->Mvctx);CHKERRQ(ierr); 879 ierr = (*a->B->ops->multadd)(a->B,a->lvec,yy,yy);CHKERRQ(ierr); 880 /* do subdiagonal part */ 881 ierr = (*a->B->ops->multtranspose)(a->B,xx,a->lvec);CHKERRQ(ierr); 882 ierr = VecScatterBegin(a->lvec,yy,ADD_VALUES,SCATTER_REVERSE,a->Mvctx);CHKERRQ(ierr); 883 ierr = VecScatterEnd(a->lvec,yy,ADD_VALUES,SCATTER_REVERSE,a->Mvctx);CHKERRQ(ierr); 884 885 PetscFunctionReturn(0); 886 } 887 888 #undef __FUNCT__ 889 #define __FUNCT__ "MatMultAdd_MPISBAIJ" 890 int MatMultAdd_MPISBAIJ(Mat A,Vec xx,Vec yy,Vec zz) 891 { 892 Mat_MPISBAIJ *a = (Mat_MPISBAIJ*)A->data; 893 int ierr; 894 895 PetscFunctionBegin; 896 ierr = VecScatterBegin(xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD,a->Mvctx);CHKERRQ(ierr); 897 /* do diagonal part */ 898 ierr = (*a->A->ops->multadd)(a->A,xx,yy,zz);CHKERRQ(ierr); 899 /* do supperdiagonal part */ 900 ierr = VecScatterEnd(xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD,a->Mvctx);CHKERRQ(ierr); 901 ierr = (*a->B->ops->multadd)(a->B,a->lvec,zz,zz);CHKERRQ(ierr); 902 903 /* do subdiagonal part */ 904 ierr = (*a->B->ops->multtranspose)(a->B,xx,a->lvec);CHKERRQ(ierr); 905 ierr = VecScatterBegin(a->lvec,zz,ADD_VALUES,SCATTER_REVERSE,a->Mvctx);CHKERRQ(ierr); 906 ierr = VecScatterEnd(a->lvec,zz,ADD_VALUES,SCATTER_REVERSE,a->Mvctx);CHKERRQ(ierr); 907 908 PetscFunctionReturn(0); 909 } 910 911 #undef __FUNCT__ 912 #define __FUNCT__ "MatMultTranspose_MPISBAIJ" 913 int MatMultTranspose_MPISBAIJ(Mat A,Vec xx,Vec yy) 914 { 915 PetscFunctionBegin; 916 SETERRQ(1,"Matrix is symmetric. Call MatMult()."); 917 /* PetscFunctionReturn(0); */ 918 } 919 920 #undef __FUNCT__ 921 #define __FUNCT__ "MatMultTransposeAdd_MPISBAIJ" 922 int MatMultTransposeAdd_MPISBAIJ(Mat A,Vec xx,Vec yy,Vec zz) 923 { 924 PetscFunctionBegin; 925 SETERRQ(1,"Matrix is symmetric. Call MatMultAdd()."); 926 /* PetscFunctionReturn(0); */ 927 } 928 929 /* 930 This only works correctly for square matrices where the subblock A->A is the 931 diagonal block 932 */ 933 #undef __FUNCT__ 934 #define __FUNCT__ "MatGetDiagonal_MPISBAIJ" 935 int MatGetDiagonal_MPISBAIJ(Mat A,Vec v) 936 { 937 Mat_MPISBAIJ *a = (Mat_MPISBAIJ*)A->data; 938 int ierr; 939 940 PetscFunctionBegin; 941 /* if (a->M != a->N) SETERRQ(PETSC_ERR_SUP,"Supports only square matrix where A->A is diag block"); */ 942 ierr = MatGetDiagonal(a->A,v);CHKERRQ(ierr); 943 PetscFunctionReturn(0); 944 } 945 946 #undef __FUNCT__ 947 #define __FUNCT__ "MatScale_MPISBAIJ" 948 int MatScale_MPISBAIJ(Scalar *aa,Mat A) 949 { 950 Mat_MPISBAIJ *a = (Mat_MPISBAIJ*)A->data; 951 int ierr; 952 953 PetscFunctionBegin; 954 ierr = MatScale(aa,a->A);CHKERRQ(ierr); 955 ierr = MatScale(aa,a->B);CHKERRQ(ierr); 956 PetscFunctionReturn(0); 957 } 958 959 #undef __FUNCT__ 960 #define __FUNCT__ "MatGetOwnershipRange_MPISBAIJ" 961 int MatGetOwnershipRange_MPISBAIJ(Mat matin,int *m,int *n) 962 { 963 Mat_MPISBAIJ *mat = (Mat_MPISBAIJ*)matin->data; 964 965 PetscFunctionBegin; 966 if (m) *m = mat->rstart*mat->bs; 967 if (n) *n = mat->rend*mat->bs; 968 PetscFunctionReturn(0); 969 } 970 971 #undef __FUNCT__ 972 #define __FUNCT__ "MatGetRow_MPISBAIJ" 973 int MatGetRow_MPISBAIJ(Mat matin,int row,int *nz,int **idx,Scalar **v) 974 { 975 Mat_MPISBAIJ *mat = (Mat_MPISBAIJ*)matin->data; 976 Scalar *vworkA,*vworkB,**pvA,**pvB,*v_p; 977 int bs = mat->bs,bs2 = mat->bs2,i,ierr,*cworkA,*cworkB,**pcA,**pcB; 978 int nztot,nzA,nzB,lrow,brstart = mat->rstart*bs,brend = mat->rend*bs; 979 int *cmap,*idx_p,cstart = mat->cstart; 980 981 PetscFunctionBegin; 982 if (mat->getrowactive == PETSC_TRUE) SETERRQ(PETSC_ERR_ARG_WRONGSTATE,"Already active"); 983 mat->getrowactive = PETSC_TRUE; 984 985 if (!mat->rowvalues && (idx || v)) { 986 /* 987 allocate enough space to hold information from the longest row. 988 */ 989 Mat_SeqSBAIJ *Aa = (Mat_SeqSBAIJ*)mat->A->data; 990 Mat_SeqBAIJ *Ba = (Mat_SeqBAIJ*)mat->B->data; 991 int max = 1,mbs = mat->mbs,tmp; 992 for (i=0; i<mbs; i++) { 993 tmp = Aa->i[i+1] - Aa->i[i] + Ba->i[i+1] - Ba->i[i]; /* row length */ 994 if (max < tmp) { max = tmp; } 995 } 996 ierr = PetscMalloc(max*bs2*(sizeof(int)+sizeof(Scalar)),&mat->rowvalues);CHKERRQ(ierr); 997 mat->rowindices = (int*)(mat->rowvalues + max*bs2); 998 } 999 1000 if (row < brstart || row >= brend) SETERRQ(PETSC_ERR_SUP,"Only local rows") 1001 lrow = row - brstart; /* local row index */ 1002 1003 pvA = &vworkA; pcA = &cworkA; pvB = &vworkB; pcB = &cworkB; 1004 if (!v) {pvA = 0; pvB = 0;} 1005 if (!idx) {pcA = 0; if (!v) pcB = 0;} 1006 ierr = (*mat->A->ops->getrow)(mat->A,lrow,&nzA,pcA,pvA);CHKERRQ(ierr); 1007 ierr = (*mat->B->ops->getrow)(mat->B,lrow,&nzB,pcB,pvB);CHKERRQ(ierr); 1008 nztot = nzA + nzB; 1009 1010 cmap = mat->garray; 1011 if (v || idx) { 1012 if (nztot) { 1013 /* Sort by increasing column numbers, assuming A and B already sorted */ 1014 int imark = -1; 1015 if (v) { 1016 *v = v_p = mat->rowvalues; 1017 for (i=0; i<nzB; i++) { 1018 if (cmap[cworkB[i]/bs] < cstart) v_p[i] = vworkB[i]; 1019 else break; 1020 } 1021 imark = i; 1022 for (i=0; i<nzA; i++) v_p[imark+i] = vworkA[i]; 1023 for (i=imark; i<nzB; i++) v_p[nzA+i] = vworkB[i]; 1024 } 1025 if (idx) { 1026 *idx = idx_p = mat->rowindices; 1027 if (imark > -1) { 1028 for (i=0; i<imark; i++) { 1029 idx_p[i] = cmap[cworkB[i]/bs]*bs + cworkB[i]%bs; 1030 } 1031 } else { 1032 for (i=0; i<nzB; i++) { 1033 if (cmap[cworkB[i]/bs] < cstart) 1034 idx_p[i] = cmap[cworkB[i]/bs]*bs + cworkB[i]%bs ; 1035 else break; 1036 } 1037 imark = i; 1038 } 1039 for (i=0; i<nzA; i++) idx_p[imark+i] = cstart*bs + cworkA[i]; 1040 for (i=imark; i<nzB; i++) idx_p[nzA+i] = cmap[cworkB[i]/bs]*bs + cworkB[i]%bs ; 1041 } 1042 } else { 1043 if (idx) *idx = 0; 1044 if (v) *v = 0; 1045 } 1046 } 1047 *nz = nztot; 1048 ierr = (*mat->A->ops->restorerow)(mat->A,lrow,&nzA,pcA,pvA);CHKERRQ(ierr); 1049 ierr = (*mat->B->ops->restorerow)(mat->B,lrow,&nzB,pcB,pvB);CHKERRQ(ierr); 1050 PetscFunctionReturn(0); 1051 } 1052 1053 #undef __FUNCT__ 1054 #define __FUNCT__ "MatRestoreRow_MPISBAIJ" 1055 int MatRestoreRow_MPISBAIJ(Mat mat,int row,int *nz,int **idx,Scalar **v) 1056 { 1057 Mat_MPISBAIJ *baij = (Mat_MPISBAIJ*)mat->data; 1058 1059 PetscFunctionBegin; 1060 if (baij->getrowactive == PETSC_FALSE) { 1061 SETERRQ(PETSC_ERR_ARG_WRONGSTATE,"MatGetRow not called"); 1062 } 1063 baij->getrowactive = PETSC_FALSE; 1064 PetscFunctionReturn(0); 1065 } 1066 1067 #undef __FUNCT__ 1068 #define __FUNCT__ "MatGetBlockSize_MPISBAIJ" 1069 int MatGetBlockSize_MPISBAIJ(Mat mat,int *bs) 1070 { 1071 Mat_MPISBAIJ *baij = (Mat_MPISBAIJ*)mat->data; 1072 1073 PetscFunctionBegin; 1074 *bs = baij->bs; 1075 PetscFunctionReturn(0); 1076 } 1077 1078 #undef __FUNCT__ 1079 #define __FUNCT__ "MatZeroEntries_MPISBAIJ" 1080 int MatZeroEntries_MPISBAIJ(Mat A) 1081 { 1082 Mat_MPISBAIJ *l = (Mat_MPISBAIJ*)A->data; 1083 int ierr; 1084 1085 PetscFunctionBegin; 1086 ierr = MatZeroEntries(l->A);CHKERRQ(ierr); 1087 ierr = MatZeroEntries(l->B);CHKERRQ(ierr); 1088 PetscFunctionReturn(0); 1089 } 1090 1091 #undef __FUNCT__ 1092 #define __FUNCT__ "MatGetInfo_MPISBAIJ" 1093 int MatGetInfo_MPISBAIJ(Mat matin,MatInfoType flag,MatInfo *info) 1094 { 1095 Mat_MPISBAIJ *a = (Mat_MPISBAIJ*)matin->data; 1096 Mat A = a->A,B = a->B; 1097 int ierr; 1098 PetscReal isend[5],irecv[5]; 1099 1100 PetscFunctionBegin; 1101 info->block_size = (double)a->bs; 1102 ierr = MatGetInfo(A,MAT_LOCAL,info);CHKERRQ(ierr); 1103 isend[0] = info->nz_used; isend[1] = info->nz_allocated; isend[2] = info->nz_unneeded; 1104 isend[3] = info->memory; isend[4] = info->mallocs; 1105 ierr = MatGetInfo(B,MAT_LOCAL,info);CHKERRQ(ierr); 1106 isend[0] += info->nz_used; isend[1] += info->nz_allocated; isend[2] += info->nz_unneeded; 1107 isend[3] += info->memory; isend[4] += info->mallocs; 1108 if (flag == MAT_LOCAL) { 1109 info->nz_used = isend[0]; 1110 info->nz_allocated = isend[1]; 1111 info->nz_unneeded = isend[2]; 1112 info->memory = isend[3]; 1113 info->mallocs = isend[4]; 1114 } else if (flag == MAT_GLOBAL_MAX) { 1115 ierr = MPI_Allreduce(isend,irecv,5,MPI_DOUBLE,MPI_MAX,matin->comm);CHKERRQ(ierr); 1116 info->nz_used = irecv[0]; 1117 info->nz_allocated = irecv[1]; 1118 info->nz_unneeded = irecv[2]; 1119 info->memory = irecv[3]; 1120 info->mallocs = irecv[4]; 1121 } else if (flag == MAT_GLOBAL_SUM) { 1122 ierr = MPI_Allreduce(isend,irecv,5,MPI_DOUBLE,MPI_SUM,matin->comm);CHKERRQ(ierr); 1123 info->nz_used = irecv[0]; 1124 info->nz_allocated = irecv[1]; 1125 info->nz_unneeded = irecv[2]; 1126 info->memory = irecv[3]; 1127 info->mallocs = irecv[4]; 1128 } else { 1129 SETERRQ1(1,"Unknown MatInfoType argument %d",flag); 1130 } 1131 info->rows_global = (double)A->M; 1132 info->columns_global = (double)A->N; 1133 info->rows_local = (double)A->m; 1134 info->columns_local = (double)A->N; 1135 info->fill_ratio_given = 0; /* no parallel LU/ILU/Cholesky */ 1136 info->fill_ratio_needed = 0; 1137 info->factor_mallocs = 0; 1138 PetscFunctionReturn(0); 1139 } 1140 1141 #undef __FUNCT__ 1142 #define __FUNCT__ "MatSetOption_MPISBAIJ" 1143 int MatSetOption_MPISBAIJ(Mat A,MatOption op) 1144 { 1145 Mat_MPISBAIJ *a = (Mat_MPISBAIJ*)A->data; 1146 int ierr; 1147 1148 PetscFunctionBegin; 1149 switch (op) { 1150 case MAT_NO_NEW_NONZERO_LOCATIONS: 1151 case MAT_YES_NEW_NONZERO_LOCATIONS: 1152 case MAT_COLUMNS_UNSORTED: 1153 case MAT_COLUMNS_SORTED: 1154 case MAT_NEW_NONZERO_ALLOCATION_ERR: 1155 case MAT_KEEP_ZEROED_ROWS: 1156 case MAT_NEW_NONZERO_LOCATION_ERR: 1157 ierr = MatSetOption(a->A,op);CHKERRQ(ierr); 1158 ierr = MatSetOption(a->B,op);CHKERRQ(ierr); 1159 break; 1160 case MAT_ROW_ORIENTED: 1161 a->roworiented = PETSC_TRUE; 1162 ierr = MatSetOption(a->A,op);CHKERRQ(ierr); 1163 ierr = MatSetOption(a->B,op);CHKERRQ(ierr); 1164 break; 1165 case MAT_ROWS_SORTED: 1166 case MAT_ROWS_UNSORTED: 1167 case MAT_YES_NEW_DIAGONALS: 1168 case MAT_USE_HASH_TABLE: 1169 case MAT_USE_SINGLE_PRECISION_SOLVES: 1170 PetscLogInfo(A,"Info:MatSetOption_MPIBAIJ:Option ignored\n"); 1171 break; 1172 case MAT_COLUMN_ORIENTED: 1173 a->roworiented = PETSC_FALSE; 1174 ierr = MatSetOption(a->A,op);CHKERRQ(ierr); 1175 ierr = MatSetOption(a->B,op);CHKERRQ(ierr); 1176 break; 1177 case MAT_IGNORE_OFF_PROC_ENTRIES: 1178 a->donotstash = PETSC_TRUE; 1179 break; 1180 case MAT_NO_NEW_DIAGONALS: 1181 SETERRQ(PETSC_ERR_SUP,"MAT_NO_NEW_DIAGONALS"); 1182 break; 1183 case MAT_USE_HASH_TABLE: 1184 a->ht_flag = PETSC_TRUE; 1185 break; 1186 default: 1187 SETERRQ(PETSC_ERR_SUP,"unknown option"); 1188 break; 1189 } 1190 PetscFunctionReturn(0); 1191 } 1192 1193 #undef __FUNCT__ 1194 #define __FUNCT__ "MatTranspose_MPISBAIJ(" 1195 int MatTranspose_MPISBAIJ(Mat A,Mat *matout) 1196 { 1197 PetscFunctionBegin; 1198 SETERRQ(1,"Matrix is symmetric. MatTranspose() should not be called"); 1199 /* PetscFunctionReturn(0); */ 1200 } 1201 1202 #undef __FUNCT__ 1203 #define __FUNCT__ "MatDiagonalScale_MPISBAIJ" 1204 int MatDiagonalScale_MPISBAIJ(Mat mat,Vec ll,Vec rr) 1205 { 1206 Mat_MPISBAIJ *baij = (Mat_MPISBAIJ*)mat->data; 1207 Mat a = baij->A,b = baij->B; 1208 int ierr,s1,s2,s3; 1209 1210 PetscFunctionBegin; 1211 if (ll != rr) { 1212 SETERRQ(PETSC_ERR_ARG_OUTOFRANGE,"For symmetric format, left and right scaling vectors must be same\n"); 1213 } 1214 ierr = MatGetLocalSize(mat,&s2,&s3);CHKERRQ(ierr); 1215 if (rr) { 1216 ierr = VecGetLocalSize(rr,&s1);CHKERRQ(ierr); 1217 if (s1!=s3) SETERRQ(PETSC_ERR_ARG_SIZ,"right vector non-conforming local size"); 1218 /* Overlap communication with computation. */ 1219 ierr = VecScatterBegin(rr,baij->lvec,INSERT_VALUES,SCATTER_FORWARD,baij->Mvctx);CHKERRQ(ierr); 1220 /*} if (ll) { */ 1221 ierr = VecGetLocalSize(ll,&s1);CHKERRQ(ierr); 1222 if (s1!=s2) SETERRQ(PETSC_ERR_ARG_SIZ,"left vector non-conforming local size"); 1223 ierr = (*b->ops->diagonalscale)(b,ll,PETSC_NULL);CHKERRQ(ierr); 1224 /* } */ 1225 /* scale the diagonal block */ 1226 ierr = (*a->ops->diagonalscale)(a,ll,rr);CHKERRQ(ierr); 1227 1228 /* if (rr) { */ 1229 /* Do a scatter end and then right scale the off-diagonal block */ 1230 ierr = VecScatterEnd(rr,baij->lvec,INSERT_VALUES,SCATTER_FORWARD,baij->Mvctx);CHKERRQ(ierr); 1231 ierr = (*b->ops->diagonalscale)(b,PETSC_NULL,baij->lvec);CHKERRQ(ierr); 1232 } 1233 1234 PetscFunctionReturn(0); 1235 } 1236 1237 #undef __FUNCT__ 1238 #define __FUNCT__ "MatZeroRows_MPISBAIJ" 1239 int MatZeroRows_MPISBAIJ(Mat A,IS is,Scalar *diag) 1240 { 1241 Mat_MPISBAIJ *l = (Mat_MPISBAIJ*)A->data; 1242 int i,ierr,N,*rows,*owners = l->rowners,size = l->size; 1243 int *procs,*nprocs,j,idx,nsends,*work,row; 1244 int nmax,*svalues,*starts,*owner,nrecvs,rank = l->rank; 1245 int *rvalues,tag = A->tag,count,base,slen,n,*source; 1246 int *lens,imdex,*lrows,*values,bs=l->bs,rstart_bs=l->rstart_bs; 1247 MPI_Comm comm = A->comm; 1248 MPI_Request *send_waits,*recv_waits; 1249 MPI_Status recv_status,*send_status; 1250 IS istmp; 1251 PetscTruth found; 1252 1253 PetscFunctionBegin; 1254 ierr = ISGetSize(is,&N);CHKERRQ(ierr); 1255 ierr = ISGetIndices(is,&rows);CHKERRQ(ierr); 1256 1257 /* first count number of contributors to each processor */ 1258 ierr = PetscMalloc(2*size*sizeof(int),&nprocs);CHKERRQ(ierr); 1259 ierr = PetscMemzero(nprocs,2*size*sizeof(int));CHKERRQ(ierr); 1260 procs = nprocs + size; 1261 ierr = PetscMalloc((N+1)*sizeof(int),&owner);CHKERRQ(ierr); /* see note*/ 1262 for (i=0; i<N; i++) { 1263 idx = rows[i]; 1264 found = PETSC_FALSE; 1265 for (j=0; j<size; j++) { 1266 if (idx >= owners[j]*bs && idx < owners[j+1]*bs) { 1267 nprocs[j]++; procs[j] = 1; owner[i] = j; found = PETSC_TRUE; break; 1268 } 1269 } 1270 if (!found) SETERRQ(PETSC_ERR_ARG_OUTOFRANGE,"Index out of range"); 1271 } 1272 nsends = 0; for (i=0; i<size; i++) { nsends += procs[i];} 1273 1274 /* inform other processors of number of messages and max length*/ 1275 ierr = PetscMalloc(2*size*sizeof(int),&work);CHKERRQ(ierr); 1276 ierr = MPI_Allreduce(nprocs,work,2*size,MPI_INT,PetscMaxSum_Op,comm);CHKERRQ(ierr); 1277 nmax = work[rank]; 1278 nrecvs = work[size+rank]; 1279 ierr = PetscFree(work);CHKERRQ(ierr); 1280 1281 /* post receives: */ 1282 ierr = PetscMalloc((nrecvs+1)*(nmax+1)*sizeof(int),&rvalues);CHKERRQ(ierr); 1283 ierr = PetscMalloc((nrecvs+1)*sizeof(MPI_Request),&recv_waits);CHKERRQ(ierr); 1284 for (i=0; i<nrecvs; i++) { 1285 ierr = MPI_Irecv(rvalues+nmax*i,nmax,MPI_INT,MPI_ANY_SOURCE,tag,comm,recv_waits+i);CHKERRQ(ierr); 1286 } 1287 1288 /* do sends: 1289 1) starts[i] gives the starting index in svalues for stuff going to 1290 the ith processor 1291 */ 1292 ierr = PetscMalloc((N+1)*sizeof(int),&svalues);CHKERRQ(ierr); 1293 ierr = PetscMalloc((nsends+1)*sizeof(MPI_Request),&send_waits);CHKERRQ(ierr); 1294 ierr = PetscMalloc((size+1)*sizeof(int),&starts);CHKERRQ(ierr); 1295 starts[0] = 0; 1296 for (i=1; i<size; i++) { starts[i] = starts[i-1] + nprocs[i-1];} 1297 for (i=0; i<N; i++) { 1298 svalues[starts[owner[i]]++] = rows[i]; 1299 } 1300 ierr = ISRestoreIndices(is,&rows);CHKERRQ(ierr); 1301 1302 starts[0] = 0; 1303 for (i=1; i<size+1; i++) { starts[i] = starts[i-1] + nprocs[i-1];} 1304 count = 0; 1305 for (i=0; i<size; i++) { 1306 if (procs[i]) { 1307 ierr = MPI_Isend(svalues+starts[i],nprocs[i],MPI_INT,i,tag,comm,send_waits+count++);CHKERRQ(ierr); 1308 } 1309 } 1310 ierr = PetscFree(starts);CHKERRQ(ierr); 1311 1312 base = owners[rank]*bs; 1313 1314 /* wait on receives */ 1315 ierr = PetscMalloc(2*(nrecvs+1)*sizeof(int),&lens);CHKERRQ(ierr); 1316 source = lens + nrecvs; 1317 count = nrecvs; slen = 0; 1318 while (count) { 1319 ierr = MPI_Waitany(nrecvs,recv_waits,&imdex,&recv_status);CHKERRQ(ierr); 1320 /* unpack receives into our local space */ 1321 ierr = MPI_Get_count(&recv_status,MPI_INT,&n);CHKERRQ(ierr); 1322 source[imdex] = recv_status.MPI_SOURCE; 1323 lens[imdex] = n; 1324 slen += n; 1325 count--; 1326 } 1327 ierr = PetscFree(recv_waits);CHKERRQ(ierr); 1328 1329 /* move the data into the send scatter */ 1330 ierr = PetscMalloc((slen+1)*sizeof(int),&lrows);CHKERRQ(ierr); 1331 count = 0; 1332 for (i=0; i<nrecvs; i++) { 1333 values = rvalues + i*nmax; 1334 for (j=0; j<lens[i]; j++) { 1335 lrows[count++] = values[j] - base; 1336 } 1337 } 1338 ierr = PetscFree(rvalues);CHKERRQ(ierr); 1339 ierr = PetscFree(lens);CHKERRQ(ierr); 1340 ierr = PetscFree(owner);CHKERRQ(ierr); 1341 ierr = PetscFree(nprocs);CHKERRQ(ierr); 1342 1343 /* actually zap the local rows */ 1344 ierr = ISCreateGeneral(PETSC_COMM_SELF,slen,lrows,&istmp);CHKERRQ(ierr); 1345 PetscLogObjectParent(A,istmp); 1346 1347 /* 1348 Zero the required rows. If the "diagonal block" of the matrix 1349 is square and the user wishes to set the diagonal we use seperate 1350 code so that MatSetValues() is not called for each diagonal allocating 1351 new memory, thus calling lots of mallocs and slowing things down. 1352 1353 Contributed by: Mathew Knepley 1354 */ 1355 /* must zero l->B before l->A because the (diag) case below may put values into l->B*/ 1356 ierr = MatZeroRows_SeqBAIJ(l->B,istmp,0);CHKERRQ(ierr); 1357 if (diag && (l->A->M == l->A->N)) { 1358 ierr = MatZeroRows_SeqSBAIJ(l->A,istmp,diag);CHKERRQ(ierr); 1359 } else if (diag) { 1360 ierr = MatZeroRows_SeqSBAIJ(l->A,istmp,0);CHKERRQ(ierr); 1361 if (((Mat_SeqSBAIJ*)l->A->data)->nonew) { 1362 SETERRQ(PETSC_ERR_SUP,"MatZeroRows() on rectangular matrices cannot be used with the Mat options \n\ 1363 MAT_NO_NEW_NONZERO_LOCATIONS,MAT_NEW_NONZERO_LOCATION_ERR,MAT_NEW_NONZERO_ALLOCATION_ERR"); 1364 } 1365 for (i=0; i<slen; i++) { 1366 row = lrows[i] + rstart_bs; 1367 ierr = MatSetValues(A,1,&row,1,&row,diag,INSERT_VALUES);CHKERRQ(ierr); 1368 } 1369 ierr = MatAssemblyBegin(A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1370 ierr = MatAssemblyEnd(A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1371 } else { 1372 ierr = MatZeroRows_SeqSBAIJ(l->A,istmp,0);CHKERRQ(ierr); 1373 } 1374 1375 ierr = ISDestroy(istmp);CHKERRQ(ierr); 1376 ierr = PetscFree(lrows);CHKERRQ(ierr); 1377 1378 /* wait on sends */ 1379 if (nsends) { 1380 ierr = PetscMalloc(nsends*sizeof(MPI_Status),&send_status);CHKERRQ(ierr); 1381 ierr = MPI_Waitall(nsends,send_waits,send_status);CHKERRQ(ierr); 1382 ierr = PetscFree(send_status);CHKERRQ(ierr); 1383 } 1384 ierr = PetscFree(send_waits);CHKERRQ(ierr); 1385 ierr = PetscFree(svalues);CHKERRQ(ierr); 1386 1387 PetscFunctionReturn(0); 1388 } 1389 1390 #undef __FUNCT__ 1391 #define __FUNCT__ "MatPrintHelp_MPISBAIJ" 1392 int MatPrintHelp_MPISBAIJ(Mat A) 1393 { 1394 Mat_MPISBAIJ *a = (Mat_MPISBAIJ*)A->data; 1395 MPI_Comm comm = A->comm; 1396 static int called = 0; 1397 int ierr; 1398 1399 PetscFunctionBegin; 1400 if (!a->rank) { 1401 ierr = MatPrintHelp_SeqSBAIJ(a->A);CHKERRQ(ierr); 1402 } 1403 if (called) {PetscFunctionReturn(0);} else called = 1; 1404 ierr = (*PetscHelpPrintf)(comm," Options for MATMPISBAIJ matrix format (the defaults):\n");CHKERRQ(ierr); 1405 ierr = (*PetscHelpPrintf)(comm," -mat_use_hash_table <factor>: Use hashtable for efficient matrix assembly\n");CHKERRQ(ierr); 1406 PetscFunctionReturn(0); 1407 } 1408 1409 #undef __FUNCT__ 1410 #define __FUNCT__ "MatSetUnfactored_MPISBAIJ" 1411 int MatSetUnfactored_MPISBAIJ(Mat A) 1412 { 1413 Mat_MPISBAIJ *a = (Mat_MPISBAIJ*)A->data; 1414 int ierr; 1415 1416 PetscFunctionBegin; 1417 ierr = MatSetUnfactored(a->A);CHKERRQ(ierr); 1418 PetscFunctionReturn(0); 1419 } 1420 1421 static int MatDuplicate_MPISBAIJ(Mat,MatDuplicateOption,Mat *); 1422 1423 #undef __FUNCT__ 1424 #define __FUNCT__ "MatEqual_MPISBAIJ" 1425 int MatEqual_MPISBAIJ(Mat A,Mat B,PetscTruth *flag) 1426 { 1427 Mat_MPISBAIJ *matB = (Mat_MPISBAIJ*)B->data,*matA = (Mat_MPISBAIJ*)A->data; 1428 Mat a,b,c,d; 1429 PetscTruth flg; 1430 int ierr; 1431 1432 PetscFunctionBegin; 1433 ierr = PetscTypeCompare((PetscObject)B,MATMPISBAIJ,&flg);CHKERRQ(ierr); 1434 if (!flg) SETERRQ(PETSC_ERR_ARG_INCOMP,"Matrices must be same type"); 1435 a = matA->A; b = matA->B; 1436 c = matB->A; d = matB->B; 1437 1438 ierr = MatEqual(a,c,&flg);CHKERRQ(ierr); 1439 if (flg == PETSC_TRUE) { 1440 ierr = MatEqual(b,d,&flg);CHKERRQ(ierr); 1441 } 1442 ierr = MPI_Allreduce(&flg,flag,1,MPI_INT,MPI_LAND,A->comm);CHKERRQ(ierr); 1443 PetscFunctionReturn(0); 1444 } 1445 1446 #undef __FUNCT__ 1447 #define __FUNCT__ "MatSetUpPreallocation_MPISBAIJ" 1448 int MatSetUpPreallocation_MPISBAIJ(Mat A) 1449 { 1450 int ierr; 1451 1452 PetscFunctionBegin; 1453 ierr = MatMPISBAIJSetPreallocation(A,1,PETSC_DEFAULT,0,PETSC_DEFAULT,0);CHKERRQ(ierr); 1454 PetscFunctionReturn(0); 1455 } 1456 /* -------------------------------------------------------------------*/ 1457 static struct _MatOps MatOps_Values = { 1458 MatSetValues_MPISBAIJ, 1459 MatGetRow_MPISBAIJ, 1460 MatRestoreRow_MPISBAIJ, 1461 MatMult_MPISBAIJ, 1462 MatMultAdd_MPISBAIJ, 1463 MatMultTranspose_MPISBAIJ, 1464 MatMultTransposeAdd_MPISBAIJ, 1465 0, 1466 0, 1467 0, 1468 0, 1469 0, 1470 0, 1471 0, 1472 MatTranspose_MPISBAIJ, 1473 MatGetInfo_MPISBAIJ, 1474 MatEqual_MPISBAIJ, 1475 MatGetDiagonal_MPISBAIJ, 1476 MatDiagonalScale_MPISBAIJ, 1477 MatNorm_MPISBAIJ, 1478 MatAssemblyBegin_MPISBAIJ, 1479 MatAssemblyEnd_MPISBAIJ, 1480 0, 1481 MatSetOption_MPISBAIJ, 1482 MatZeroEntries_MPISBAIJ, 1483 MatZeroRows_MPISBAIJ, 1484 0, 1485 0, 1486 0, 1487 0, 1488 MatSetUpPreallocation_MPISBAIJ, 1489 0, 1490 MatGetOwnershipRange_MPISBAIJ, 1491 0, 1492 0, 1493 0, 1494 0, 1495 MatDuplicate_MPISBAIJ, 1496 0, 1497 0, 1498 0, 1499 0, 1500 0, 1501 MatGetSubMatrices_MPISBAIJ, 1502 MatIncreaseOverlap_MPISBAIJ, 1503 MatGetValues_MPISBAIJ, 1504 0, 1505 MatPrintHelp_MPISBAIJ, 1506 MatScale_MPISBAIJ, 1507 0, 1508 0, 1509 0, 1510 MatGetBlockSize_MPISBAIJ, 1511 0, 1512 0, 1513 0, 1514 0, 1515 0, 1516 0, 1517 MatSetUnfactored_MPISBAIJ, 1518 0, 1519 MatSetValuesBlocked_MPISBAIJ, 1520 0, 1521 0, 1522 0, 1523 MatGetMaps_Petsc, 1524 0, 1525 0, 1526 0, 1527 0, 1528 0, 1529 0, 1530 MatGetRowMax_MPISBAIJ}; 1531 1532 1533 EXTERN_C_BEGIN 1534 #undef __FUNCT__ 1535 #define __FUNCT__ "MatGetDiagonalBlock_MPISBAIJ" 1536 int MatGetDiagonalBlock_MPISBAIJ(Mat A,PetscTruth *iscopy,MatReuse reuse,Mat *a) 1537 { 1538 PetscFunctionBegin; 1539 *a = ((Mat_MPISBAIJ *)A->data)->A; 1540 *iscopy = PETSC_FALSE; 1541 PetscFunctionReturn(0); 1542 } 1543 EXTERN_C_END 1544 1545 EXTERN_C_BEGIN 1546 #undef __FUNCT__ 1547 #define __FUNCT__ "MatCreate_MPISBAIJ" 1548 int MatCreate_MPISBAIJ(Mat B) 1549 { 1550 Mat_MPISBAIJ *b; 1551 int ierr; 1552 PetscTruth flg; 1553 1554 PetscFunctionBegin; 1555 1556 ierr = PetscNew(Mat_MPISBAIJ,&b);CHKERRQ(ierr); 1557 B->data = (void*)b; 1558 ierr = PetscMemzero(b,sizeof(Mat_MPISBAIJ));CHKERRQ(ierr); 1559 ierr = PetscMemcpy(B->ops,&MatOps_Values,sizeof(struct _MatOps));CHKERRQ(ierr); 1560 1561 B->ops->destroy = MatDestroy_MPISBAIJ; 1562 B->ops->view = MatView_MPISBAIJ; 1563 B->mapping = 0; 1564 B->factor = 0; 1565 B->assembled = PETSC_FALSE; 1566 1567 B->insertmode = NOT_SET_VALUES; 1568 ierr = MPI_Comm_rank(B->comm,&b->rank);CHKERRQ(ierr); 1569 ierr = MPI_Comm_size(B->comm,&b->size);CHKERRQ(ierr); 1570 1571 /* build local table of row and column ownerships */ 1572 ierr = PetscMalloc(3*(b->size+2)*sizeof(int),&b->rowners);CHKERRQ(ierr); 1573 b->cowners = b->rowners + b->size + 2; 1574 b->rowners_bs = b->cowners + b->size + 2; 1575 PetscLogObjectMemory(B,3*(b->size+2)*sizeof(int)+sizeof(struct _p_Mat)+sizeof(Mat_MPISBAIJ)); 1576 1577 /* build cache for off array entries formed */ 1578 ierr = MatStashCreate_Private(B->comm,1,&B->stash);CHKERRQ(ierr); 1579 b->donotstash = PETSC_FALSE; 1580 b->colmap = PETSC_NULL; 1581 b->garray = PETSC_NULL; 1582 b->roworiented = PETSC_TRUE; 1583 1584 #if defined(PETSC_USE_MAT_SINGLE) 1585 /* stuff for MatSetValues_XXX in single precision */ 1586 b->setvalueslen = 0; 1587 b->setvaluescopy = PETSC_NULL; 1588 #endif 1589 1590 /* stuff used in block assembly */ 1591 b->barray = 0; 1592 1593 /* stuff used for matrix vector multiply */ 1594 b->lvec = 0; 1595 b->Mvctx = 0; 1596 1597 /* stuff for MatGetRow() */ 1598 b->rowindices = 0; 1599 b->rowvalues = 0; 1600 b->getrowactive = PETSC_FALSE; 1601 1602 /* hash table stuff */ 1603 b->ht = 0; 1604 b->hd = 0; 1605 b->ht_size = 0; 1606 b->ht_flag = PETSC_FALSE; 1607 b->ht_fact = 0; 1608 b->ht_total_ct = 0; 1609 b->ht_insert_ct = 0; 1610 1611 ierr = PetscOptionsHasName(PETSC_NULL,"-mat_use_hash_table",&flg);CHKERRQ(ierr); 1612 if (flg) { 1613 double fact = 1.39; 1614 ierr = MatSetOption(B,MAT_USE_HASH_TABLE);CHKERRQ(ierr); 1615 ierr = PetscOptionsGetDouble(PETSC_NULL,"-mat_use_hash_table",&fact,PETSC_NULL);CHKERRQ(ierr); 1616 if (fact <= 1.0) fact = 1.39; 1617 ierr = MatMPIBAIJSetHashTableFactor(B,fact);CHKERRQ(ierr); 1618 PetscLogInfo(0,"MatCreateMPISBAIJ:Hash table Factor used %5.2f\n",fact); 1619 } 1620 ierr = PetscObjectComposeFunctionDynamic((PetscObject)B,"MatStoreValues_C", 1621 "MatStoreValues_MPISBAIJ", 1622 MatStoreValues_MPISBAIJ);CHKERRQ(ierr); 1623 ierr = PetscObjectComposeFunctionDynamic((PetscObject)B,"MatRetrieveValues_C", 1624 "MatRetrieveValues_MPISBAIJ", 1625 MatRetrieveValues_MPISBAIJ);CHKERRQ(ierr); 1626 ierr = PetscObjectComposeFunctionDynamic((PetscObject)B,"MatGetDiagonalBlock_C", 1627 "MatGetDiagonalBlock_MPISBAIJ", 1628 MatGetDiagonalBlock_MPISBAIJ);CHKERRQ(ierr); 1629 PetscFunctionReturn(0); 1630 } 1631 EXTERN_C_END 1632 1633 #undef __FUNCT__ 1634 #define __FUNCT__ "MatMPISBAIJSetPreallocation" 1635 /*@C 1636 MatMPISBAIJSetPreallocation - For good matrix assembly performance 1637 the user should preallocate the matrix storage by setting the parameters 1638 d_nz (or d_nnz) and o_nz (or o_nnz). By setting these parameters accurately, 1639 performance can be increased by more than a factor of 50. 1640 1641 Collective on Mat 1642 1643 Input Parameters: 1644 + A - the matrix 1645 . bs - size of blockk 1646 . d_nz - number of block nonzeros per block row in diagonal portion of local 1647 submatrix (same for all local rows) 1648 . d_nnz - array containing the number of block nonzeros in the various block rows 1649 of the in diagonal portion of the local (possibly different for each block 1650 row) or PETSC_NULL. You must leave room for the diagonal entry even if it is zero. 1651 . o_nz - number of block nonzeros per block row in the off-diagonal portion of local 1652 submatrix (same for all local rows). 1653 - o_nnz - array containing the number of nonzeros in the various block rows of the 1654 off-diagonal portion of the local submatrix (possibly different for 1655 each block row) or PETSC_NULL. 1656 1657 1658 Options Database Keys: 1659 . -mat_no_unroll - uses code that does not unroll the loops in the 1660 block calculations (much slower) 1661 . -mat_block_size - size of the blocks to use 1662 1663 Notes: 1664 1665 If PETSC_DECIDE or PETSC_DETERMINE is used for a particular argument on one processor 1666 than it must be used on all processors that share the object for that argument. 1667 1668 Storage Information: 1669 For a square global matrix we define each processor's diagonal portion 1670 to be its local rows and the corresponding columns (a square submatrix); 1671 each processor's off-diagonal portion encompasses the remainder of the 1672 local matrix (a rectangular submatrix). 1673 1674 The user can specify preallocated storage for the diagonal part of 1675 the local submatrix with either d_nz or d_nnz (not both). Set 1676 d_nz=PETSC_DEFAULT and d_nnz=PETSC_NULL for PETSc to control dynamic 1677 memory allocation. Likewise, specify preallocated storage for the 1678 off-diagonal part of the local submatrix with o_nz or o_nnz (not both). 1679 1680 Consider a processor that owns rows 3, 4 and 5 of a parallel matrix. In 1681 the figure below we depict these three local rows and all columns (0-11). 1682 1683 .vb 1684 0 1 2 3 4 5 6 7 8 9 10 11 1685 ------------------- 1686 row 3 | o o o d d d o o o o o o 1687 row 4 | o o o d d d o o o o o o 1688 row 5 | o o o d d d o o o o o o 1689 ------------------- 1690 .ve 1691 1692 Thus, any entries in the d locations are stored in the d (diagonal) 1693 submatrix, and any entries in the o locations are stored in the 1694 o (off-diagonal) submatrix. Note that the d and the o submatrices are 1695 stored simply in the MATSEQBAIJ format for compressed row storage. 1696 1697 Now d_nz should indicate the number of block nonzeros per row in the d matrix, 1698 and o_nz should indicate the number of block nonzeros per row in the o matrix. 1699 In general, for PDE problems in which most nonzeros are near the diagonal, 1700 one expects d_nz >> o_nz. For large problems you MUST preallocate memory 1701 or you will get TERRIBLE performance; see the users' manual chapter on 1702 matrices. 1703 1704 Level: intermediate 1705 1706 .keywords: matrix, block, aij, compressed row, sparse, parallel 1707 1708 .seealso: MatCreate(), MatCreateSeqSBAIJ(), MatSetValues(), MatCreateMPIBAIJ() 1709 @*/ 1710 1711 int MatMPISBAIJSetPreallocation(Mat B,int bs,int d_nz,int *d_nnz,int o_nz,int *o_nnz) 1712 { 1713 Mat_MPISBAIJ *b; 1714 int ierr,i,mbs,Mbs; 1715 PetscTruth flg2; 1716 1717 PetscFunctionBegin; 1718 ierr = PetscTypeCompare((PetscObject)B,MATMPISBAIJ,&flg2);CHKERRQ(ierr); 1719 if (!flg2) PetscFunctionReturn(0); 1720 1721 ierr = PetscOptionsGetInt(PETSC_NULL,"-mat_block_size",&bs,PETSC_NULL);CHKERRQ(ierr); 1722 1723 if (bs < 1) SETERRQ(PETSC_ERR_ARG_OUTOFRANGE,"Invalid block size specified, must be positive"); 1724 if (d_nz == PETSC_DECIDE || d_nz == PETSC_DEFAULT) d_nz = 3; 1725 if (o_nz == PETSC_DECIDE || o_nz == PETSC_DEFAULT) o_nz = 1; 1726 if (d_nz < 0) SETERRQ1(PETSC_ERR_ARG_OUTOFRANGE,"d_nz cannot be less than 0: value %d",d_nz); 1727 if (o_nz < 0) SETERRQ1(PETSC_ERR_ARG_OUTOFRANGE,"o_nz cannot be less than 0: value %d",o_nz); 1728 if (d_nnz) { 1729 for (i=0; i<B->m/bs; i++) { 1730 if (d_nnz[i] < 0) SETERRQ2(PETSC_ERR_ARG_OUTOFRANGE,"d_nnz cannot be less than -1: local row %d value %d",i,d_nnz[i]); 1731 } 1732 } 1733 if (o_nnz) { 1734 for (i=0; i<B->m/bs; i++) { 1735 if (o_nnz[i] < 0) SETERRQ2(PETSC_ERR_ARG_OUTOFRANGE,"o_nnz cannot be less than -1: local row %d value %d",i,o_nnz[i]); 1736 } 1737 } 1738 B->preallocated = PETSC_TRUE; 1739 ierr = PetscSplitOwnershipBlock(B->comm,bs,&B->m,&B->M);CHKERRQ(ierr); 1740 ierr = PetscSplitOwnershipBlock(B->comm,bs,&B->n,&B->N);CHKERRQ(ierr); 1741 ierr = MapCreateMPI(B->comm,B->m,B->M,&B->rmap);CHKERRQ(ierr); 1742 ierr = MapCreateMPI(B->comm,B->m,B->M,&B->cmap);CHKERRQ(ierr); 1743 1744 b = (Mat_MPISBAIJ*)B->data; 1745 mbs = B->m/bs; 1746 Mbs = B->M/bs; 1747 if (mbs*bs != B->m) { 1748 SETERRQ2(PETSC_ERR_ARG_SIZ,"No of local rows %d must be divisible by blocksize %d",B->m,bs); 1749 } 1750 1751 b->bs = bs; 1752 b->bs2 = bs*bs; 1753 b->mbs = mbs; 1754 b->nbs = mbs; 1755 b->Mbs = Mbs; 1756 b->Nbs = Mbs; 1757 1758 ierr = MPI_Allgather(&b->mbs,1,MPI_INT,b->rowners+1,1,MPI_INT,B->comm);CHKERRQ(ierr); 1759 b->rowners[0] = 0; 1760 for (i=2; i<=b->size; i++) { 1761 b->rowners[i] += b->rowners[i-1]; 1762 } 1763 b->rstart = b->rowners[b->rank]; 1764 b->rend = b->rowners[b->rank+1]; 1765 b->cstart = b->rstart; 1766 b->cend = b->rend; 1767 for (i=0; i<=b->size; i++) { 1768 b->rowners_bs[i] = b->rowners[i]*bs; 1769 } 1770 b->rstart_bs = b-> rstart*bs; 1771 b->rend_bs = b->rend*bs; 1772 1773 b->cstart_bs = b->cstart*bs; 1774 b->cend_bs = b->cend*bs; 1775 1776 1777 ierr = MatCreateSeqSBAIJ(PETSC_COMM_SELF,bs,B->m,B->m,d_nz,d_nnz,&b->A);CHKERRQ(ierr); 1778 PetscLogObjectParent(B,b->A); 1779 ierr = MatCreateSeqBAIJ(PETSC_COMM_SELF,bs,B->m,B->M,o_nz,o_nnz,&b->B);CHKERRQ(ierr); 1780 PetscLogObjectParent(B,b->B); 1781 1782 /* build cache for off array entries formed */ 1783 ierr = MatStashCreate_Private(B->comm,bs,&B->bstash);CHKERRQ(ierr); 1784 1785 PetscFunctionReturn(0); 1786 } 1787 1788 #undef __FUNCT__ 1789 #define __FUNCT__ "MatCreateMPISBAIJ" 1790 /*@C 1791 MatCreateMPISBAIJ - Creates a sparse parallel matrix in symmetric block AIJ format 1792 (block compressed row). For good matrix assembly performance 1793 the user should preallocate the matrix storage by setting the parameters 1794 d_nz (or d_nnz) and o_nz (or o_nnz). By setting these parameters accurately, 1795 performance can be increased by more than a factor of 50. 1796 1797 Collective on MPI_Comm 1798 1799 Input Parameters: 1800 + comm - MPI communicator 1801 . bs - size of blockk 1802 . m - number of local rows (or PETSC_DECIDE to have calculated if M is given) 1803 This value should be the same as the local size used in creating the 1804 y vector for the matrix-vector product y = Ax. 1805 . n - number of local columns (or PETSC_DECIDE to have calculated if N is given) 1806 This value should be the same as the local size used in creating the 1807 x vector for the matrix-vector product y = Ax. 1808 . M - number of global rows (or PETSC_DETERMINE to have calculated if m is given) 1809 . N - number of global columns (or PETSC_DETERMINE to have calculated if n is given) 1810 . d_nz - number of block nonzeros per block row in diagonal portion of local 1811 submatrix (same for all local rows) 1812 . d_nnz - array containing the number of block nonzeros in the various block rows 1813 of the in diagonal portion of the local (possibly different for each block 1814 row) or PETSC_NULL. You must leave room for the diagonal entry even if it is zero. 1815 . o_nz - number of block nonzeros per block row in the off-diagonal portion of local 1816 submatrix (same for all local rows). 1817 - o_nnz - array containing the number of nonzeros in the various block rows of the 1818 off-diagonal portion of the local submatrix (possibly different for 1819 each block row) or PETSC_NULL. 1820 1821 Output Parameter: 1822 . A - the matrix 1823 1824 Options Database Keys: 1825 . -mat_no_unroll - uses code that does not unroll the loops in the 1826 block calculations (much slower) 1827 . -mat_block_size - size of the blocks to use 1828 . -mat_mpi - use the parallel matrix data structures even on one processor 1829 (defaults to using SeqBAIJ format on one processor) 1830 1831 Notes: 1832 The user MUST specify either the local or global matrix dimensions 1833 (possibly both). 1834 1835 If PETSC_DECIDE or PETSC_DETERMINE is used for a particular argument on one processor 1836 than it must be used on all processors that share the object for that argument. 1837 1838 Storage Information: 1839 For a square global matrix we define each processor's diagonal portion 1840 to be its local rows and the corresponding columns (a square submatrix); 1841 each processor's off-diagonal portion encompasses the remainder of the 1842 local matrix (a rectangular submatrix). 1843 1844 The user can specify preallocated storage for the diagonal part of 1845 the local submatrix with either d_nz or d_nnz (not both). Set 1846 d_nz=PETSC_DEFAULT and d_nnz=PETSC_NULL for PETSc to control dynamic 1847 memory allocation. Likewise, specify preallocated storage for the 1848 off-diagonal part of the local submatrix with o_nz or o_nnz (not both). 1849 1850 Consider a processor that owns rows 3, 4 and 5 of a parallel matrix. In 1851 the figure below we depict these three local rows and all columns (0-11). 1852 1853 .vb 1854 0 1 2 3 4 5 6 7 8 9 10 11 1855 ------------------- 1856 row 3 | o o o d d d o o o o o o 1857 row 4 | o o o d d d o o o o o o 1858 row 5 | o o o d d d o o o o o o 1859 ------------------- 1860 .ve 1861 1862 Thus, any entries in the d locations are stored in the d (diagonal) 1863 submatrix, and any entries in the o locations are stored in the 1864 o (off-diagonal) submatrix. Note that the d and the o submatrices are 1865 stored simply in the MATSEQBAIJ format for compressed row storage. 1866 1867 Now d_nz should indicate the number of block nonzeros per row in the d matrix, 1868 and o_nz should indicate the number of block nonzeros per row in the o matrix. 1869 In general, for PDE problems in which most nonzeros are near the diagonal, 1870 one expects d_nz >> o_nz. For large problems you MUST preallocate memory 1871 or you will get TERRIBLE performance; see the users' manual chapter on 1872 matrices. 1873 1874 Level: intermediate 1875 1876 .keywords: matrix, block, aij, compressed row, sparse, parallel 1877 1878 .seealso: MatCreate(), MatCreateSeqSBAIJ(), MatSetValues(), MatCreateMPIBAIJ() 1879 @*/ 1880 1881 int MatCreateMPISBAIJ(MPI_Comm comm,int bs,int m,int n,int M,int N,int d_nz,int *d_nnz,int o_nz,int *o_nnz,Mat *A) 1882 { 1883 int ierr,size; 1884 1885 PetscFunctionBegin; 1886 ierr = MatCreate(comm,m,n,M,N,A);CHKERRQ(ierr); 1887 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 1888 if (size > 1) { 1889 ierr = MatSetType(*A,MATMPISBAIJ);CHKERRQ(ierr); 1890 ierr = MatMPISBAIJSetPreallocation(*A,bs,d_nz,d_nnz,o_nz,o_nnz);CHKERRQ(ierr); 1891 } else { 1892 ierr = MatSetType(*A,MATSEQSBAIJ);CHKERRQ(ierr); 1893 ierr = MatSeqSBAIJSetPreallocation(*A,bs,d_nz,d_nnz);CHKERRQ(ierr); 1894 } 1895 PetscFunctionReturn(0); 1896 } 1897 1898 1899 #undef __FUNCT__ 1900 #define __FUNCT__ "MatDuplicate_MPISBAIJ" 1901 static int MatDuplicate_MPISBAIJ(Mat matin,MatDuplicateOption cpvalues,Mat *newmat) 1902 { 1903 Mat mat; 1904 Mat_MPISBAIJ *a,*oldmat = (Mat_MPISBAIJ*)matin->data; 1905 int ierr,len=0; 1906 1907 PetscFunctionBegin; 1908 *newmat = 0; 1909 ierr = MatCreate(matin->comm,matin->m,matin->n,matin->M,matin->N,&mat);CHKERRQ(ierr); 1910 ierr = MatSetType(mat,MATMPISBAIJ);CHKERRQ(ierr); 1911 mat->preallocated = PETSC_TRUE; 1912 a = (Mat_MPISBAIJ*)mat->data; 1913 a->bs = oldmat->bs; 1914 a->bs2 = oldmat->bs2; 1915 a->mbs = oldmat->mbs; 1916 a->nbs = oldmat->nbs; 1917 a->Mbs = oldmat->Mbs; 1918 a->Nbs = oldmat->Nbs; 1919 1920 a->rstart = oldmat->rstart; 1921 a->rend = oldmat->rend; 1922 a->cstart = oldmat->cstart; 1923 a->cend = oldmat->cend; 1924 a->size = oldmat->size; 1925 a->rank = oldmat->rank; 1926 a->donotstash = oldmat->donotstash; 1927 a->roworiented = oldmat->roworiented; 1928 a->rowindices = 0; 1929 a->rowvalues = 0; 1930 a->getrowactive = PETSC_FALSE; 1931 a->barray = 0; 1932 a->rstart_bs = oldmat->rstart_bs; 1933 a->rend_bs = oldmat->rend_bs; 1934 a->cstart_bs = oldmat->cstart_bs; 1935 a->cend_bs = oldmat->cend_bs; 1936 1937 /* hash table stuff */ 1938 a->ht = 0; 1939 a->hd = 0; 1940 a->ht_size = 0; 1941 a->ht_flag = oldmat->ht_flag; 1942 a->ht_fact = oldmat->ht_fact; 1943 a->ht_total_ct = 0; 1944 a->ht_insert_ct = 0; 1945 1946 ierr = PetscMalloc(3*(a->size+2)*sizeof(int),&a->rowners);CHKERRQ(ierr); 1947 PetscLogObjectMemory(mat,3*(a->size+2)*sizeof(int)+sizeof(struct _p_Mat)+sizeof(Mat_MPISBAIJ)); 1948 a->cowners = a->rowners + a->size + 2; 1949 a->rowners_bs = a->cowners + a->size + 2; 1950 ierr = PetscMemcpy(a->rowners,oldmat->rowners,3*(a->size+2)*sizeof(int));CHKERRQ(ierr); 1951 ierr = MatStashCreate_Private(matin->comm,1,&mat->stash);CHKERRQ(ierr); 1952 ierr = MatStashCreate_Private(matin->comm,oldmat->bs,&mat->bstash);CHKERRQ(ierr); 1953 if (oldmat->colmap) { 1954 #if defined (PETSC_USE_CTABLE) 1955 ierr = PetscTableCreateCopy(oldmat->colmap,&a->colmap);CHKERRQ(ierr); 1956 #else 1957 ierr = PetscMalloc((a->Nbs)*sizeof(int),&a->colmap);CHKERRQ(ierr); 1958 PetscLogObjectMemory(mat,(a->Nbs)*sizeof(int)); 1959 ierr = PetscMemcpy(a->colmap,oldmat->colmap,(a->Nbs)*sizeof(int));CHKERRQ(ierr); 1960 #endif 1961 } else a->colmap = 0; 1962 if (oldmat->garray && (len = ((Mat_SeqBAIJ*)(oldmat->B->data))->nbs)) { 1963 ierr = PetscMalloc(len*sizeof(int),&a->garray);CHKERRQ(ierr); 1964 PetscLogObjectMemory(mat,len*sizeof(int)); 1965 ierr = PetscMemcpy(a->garray,oldmat->garray,len*sizeof(int));CHKERRQ(ierr); 1966 } else a->garray = 0; 1967 1968 ierr = VecDuplicate(oldmat->lvec,&a->lvec);CHKERRQ(ierr); 1969 PetscLogObjectParent(mat,a->lvec); 1970 ierr = VecScatterCopy(oldmat->Mvctx,&a->Mvctx);CHKERRQ(ierr); 1971 1972 PetscLogObjectParent(mat,a->Mvctx); 1973 ierr = MatDuplicate(oldmat->A,cpvalues,&a->A);CHKERRQ(ierr); 1974 PetscLogObjectParent(mat,a->A); 1975 ierr = MatDuplicate(oldmat->B,cpvalues,&a->B);CHKERRQ(ierr); 1976 PetscLogObjectParent(mat,a->B); 1977 ierr = PetscFListDuplicate(mat->qlist,&matin->qlist);CHKERRQ(ierr); 1978 *newmat = mat; 1979 PetscFunctionReturn(0); 1980 } 1981 1982 #include "petscsys.h" 1983 1984 EXTERN_C_BEGIN 1985 #undef __FUNCT__ 1986 #define __FUNCT__ "MatLoad_MPISBAIJ" 1987 int MatLoad_MPISBAIJ(PetscViewer viewer,MatType type,Mat *newmat) 1988 { 1989 Mat A; 1990 int i,nz,ierr,j,rstart,rend,fd; 1991 Scalar *vals,*buf; 1992 MPI_Comm comm = ((PetscObject)viewer)->comm; 1993 MPI_Status status; 1994 int header[4],rank,size,*rowlengths = 0,M,N,m,*rowners,*browners,maxnz,*cols; 1995 int *locrowlens,*sndcounts = 0,*procsnz = 0,jj,*mycols,*ibuf; 1996 int tag = ((PetscObject)viewer)->tag,bs=1,Mbs,mbs,extra_rows; 1997 int *dlens,*odlens,*mask,*masked1,*masked2,rowcount,odcount; 1998 int dcount,kmax,k,nzcount,tmp; 1999 2000 PetscFunctionBegin; 2001 ierr = PetscOptionsGetInt(PETSC_NULL,"-matload_block_size",&bs,PETSC_NULL);CHKERRQ(ierr); 2002 2003 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 2004 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 2005 if (!rank) { 2006 ierr = PetscViewerBinaryGetDescriptor(viewer,&fd);CHKERRQ(ierr); 2007 ierr = PetscBinaryRead(fd,(char *)header,4,PETSC_INT);CHKERRQ(ierr); 2008 if (header[0] != MAT_COOKIE) SETERRQ(PETSC_ERR_FILE_UNEXPECTED,"not matrix object"); 2009 if (header[3] < 0) { 2010 SETERRQ(PETSC_ERR_FILE_UNEXPECTED,"Matrix stored in special format, cannot load as MPISBAIJ"); 2011 } 2012 } 2013 2014 ierr = MPI_Bcast(header+1,3,MPI_INT,0,comm);CHKERRQ(ierr); 2015 M = header[1]; N = header[2]; 2016 2017 if (M != N) SETERRQ(PETSC_ERR_SUP,"Can only do square matrices"); 2018 2019 /* 2020 This code adds extra rows to make sure the number of rows is 2021 divisible by the blocksize 2022 */ 2023 Mbs = M/bs; 2024 extra_rows = bs - M + bs*(Mbs); 2025 if (extra_rows == bs) extra_rows = 0; 2026 else Mbs++; 2027 if (extra_rows &&!rank) { 2028 PetscLogInfo(0,"MatLoad_MPISBAIJ:Padding loaded matrix to match blocksize\n"); 2029 } 2030 2031 /* determine ownership of all rows */ 2032 mbs = Mbs/size + ((Mbs % size) > rank); 2033 m = mbs*bs; 2034 ierr = PetscMalloc(2*(size+2)*sizeof(int),&rowners);CHKERRQ(ierr); 2035 browners = rowners + size + 1; 2036 ierr = MPI_Allgather(&mbs,1,MPI_INT,rowners+1,1,MPI_INT,comm);CHKERRQ(ierr); 2037 rowners[0] = 0; 2038 for (i=2; i<=size; i++) rowners[i] += rowners[i-1]; 2039 for (i=0; i<=size; i++) browners[i] = rowners[i]*bs; 2040 rstart = rowners[rank]; 2041 rend = rowners[rank+1]; 2042 2043 /* distribute row lengths to all processors */ 2044 ierr = PetscMalloc((rend-rstart)*bs*sizeof(int),&locrowlens);CHKERRQ(ierr); 2045 if (!rank) { 2046 ierr = PetscMalloc((M+extra_rows)*sizeof(int),&rowlengths);CHKERRQ(ierr); 2047 ierr = PetscBinaryRead(fd,rowlengths,M,PETSC_INT);CHKERRQ(ierr); 2048 for (i=0; i<extra_rows; i++) rowlengths[M+i] = 1; 2049 ierr = PetscMalloc(size*sizeof(int),&sndcounts);CHKERRQ(ierr); 2050 for (i=0; i<size; i++) sndcounts[i] = browners[i+1] - browners[i]; 2051 ierr = MPI_Scatterv(rowlengths,sndcounts,browners,MPI_INT,locrowlens,(rend-rstart)*bs,MPI_INT,0,comm);CHKERRQ(ierr); 2052 ierr = PetscFree(sndcounts);CHKERRQ(ierr); 2053 } else { 2054 ierr = MPI_Scatterv(0,0,0,MPI_INT,locrowlens,(rend-rstart)*bs,MPI_INT,0,comm);CHKERRQ(ierr); 2055 } 2056 2057 if (!rank) { /* procs[0] */ 2058 /* calculate the number of nonzeros on each processor */ 2059 ierr = PetscMalloc(size*sizeof(int),&procsnz);CHKERRQ(ierr); 2060 ierr = PetscMemzero(procsnz,size*sizeof(int));CHKERRQ(ierr); 2061 for (i=0; i<size; i++) { 2062 for (j=rowners[i]*bs; j< rowners[i+1]*bs; j++) { 2063 procsnz[i] += rowlengths[j]; 2064 } 2065 } 2066 ierr = PetscFree(rowlengths);CHKERRQ(ierr); 2067 2068 /* determine max buffer needed and allocate it */ 2069 maxnz = 0; 2070 for (i=0; i<size; i++) { 2071 maxnz = PetscMax(maxnz,procsnz[i]); 2072 } 2073 ierr = PetscMalloc(maxnz*sizeof(int),&cols);CHKERRQ(ierr); 2074 2075 /* read in my part of the matrix column indices */ 2076 nz = procsnz[0]; 2077 ierr = PetscMalloc(nz*sizeof(int),&ibuf);CHKERRQ(ierr); 2078 mycols = ibuf; 2079 if (size == 1) nz -= extra_rows; 2080 ierr = PetscBinaryRead(fd,mycols,nz,PETSC_INT);CHKERRQ(ierr); 2081 if (size == 1) for (i=0; i< extra_rows; i++) { mycols[nz+i] = M+i; } 2082 2083 /* read in every ones (except the last) and ship off */ 2084 for (i=1; i<size-1; i++) { 2085 nz = procsnz[i]; 2086 ierr = PetscBinaryRead(fd,cols,nz,PETSC_INT);CHKERRQ(ierr); 2087 ierr = MPI_Send(cols,nz,MPI_INT,i,tag,comm);CHKERRQ(ierr); 2088 } 2089 /* read in the stuff for the last proc */ 2090 if (size != 1) { 2091 nz = procsnz[size-1] - extra_rows; /* the extra rows are not on the disk */ 2092 ierr = PetscBinaryRead(fd,cols,nz,PETSC_INT);CHKERRQ(ierr); 2093 for (i=0; i<extra_rows; i++) cols[nz+i] = M+i; 2094 ierr = MPI_Send(cols,nz+extra_rows,MPI_INT,size-1,tag,comm);CHKERRQ(ierr); 2095 } 2096 ierr = PetscFree(cols);CHKERRQ(ierr); 2097 } else { /* procs[i], i>0 */ 2098 /* determine buffer space needed for message */ 2099 nz = 0; 2100 for (i=0; i<m; i++) { 2101 nz += locrowlens[i]; 2102 } 2103 ierr = PetscMalloc(nz*sizeof(int),&ibuf);CHKERRQ(ierr); 2104 mycols = ibuf; 2105 /* receive message of column indices*/ 2106 ierr = MPI_Recv(mycols,nz,MPI_INT,0,tag,comm,&status);CHKERRQ(ierr); 2107 ierr = MPI_Get_count(&status,MPI_INT,&maxnz);CHKERRQ(ierr); 2108 if (maxnz != nz) SETERRQ(PETSC_ERR_FILE_UNEXPECTED,"something is wrong with file"); 2109 } 2110 2111 /* loop over local rows, determining number of off diagonal entries */ 2112 ierr = PetscMalloc(2*(rend-rstart+1)*sizeof(int),&dlens);CHKERRQ(ierr); 2113 odlens = dlens + (rend-rstart); 2114 ierr = PetscMalloc(3*Mbs*sizeof(int),&mask);CHKERRQ(ierr); 2115 ierr = PetscMemzero(mask,3*Mbs*sizeof(int));CHKERRQ(ierr); 2116 masked1 = mask + Mbs; 2117 masked2 = masked1 + Mbs; 2118 rowcount = 0; nzcount = 0; 2119 for (i=0; i<mbs; i++) { 2120 dcount = 0; 2121 odcount = 0; 2122 for (j=0; j<bs; j++) { 2123 kmax = locrowlens[rowcount]; 2124 for (k=0; k<kmax; k++) { 2125 tmp = mycols[nzcount++]/bs; /* block col. index */ 2126 if (!mask[tmp]) { 2127 mask[tmp] = 1; 2128 if (tmp < rstart || tmp >= rend) masked2[odcount++] = tmp; /* entry in off-diag portion */ 2129 else masked1[dcount++] = tmp; /* entry in diag portion */ 2130 } 2131 } 2132 rowcount++; 2133 } 2134 2135 dlens[i] = dcount; /* d_nzz[i] */ 2136 odlens[i] = odcount; /* o_nzz[i] */ 2137 2138 /* zero out the mask elements we set */ 2139 for (j=0; j<dcount; j++) mask[masked1[j]] = 0; 2140 for (j=0; j<odcount; j++) mask[masked2[j]] = 0; 2141 } 2142 2143 /* create our matrix */ 2144 ierr = MatCreateMPISBAIJ(comm,bs,m,m,PETSC_DETERMINE,PETSC_DETERMINE,0,dlens,0,odlens,newmat); 2145 CHKERRQ(ierr); 2146 A = *newmat; 2147 ierr = MatSetOption(A,MAT_COLUMNS_SORTED);CHKERRQ(ierr); 2148 2149 if (!rank) { 2150 ierr = PetscMalloc(maxnz*sizeof(Scalar),&buf);CHKERRQ(ierr); 2151 /* read in my part of the matrix numerical values */ 2152 nz = procsnz[0]; 2153 vals = buf; 2154 mycols = ibuf; 2155 if (size == 1) nz -= extra_rows; 2156 ierr = PetscBinaryRead(fd,vals,nz,PETSC_SCALAR);CHKERRQ(ierr); 2157 if (size == 1) for (i=0; i< extra_rows; i++) { vals[nz+i] = 1.0; } 2158 2159 /* insert into matrix */ 2160 jj = rstart*bs; 2161 for (i=0; i<m; i++) { 2162 ierr = MatSetValues(A,1,&jj,locrowlens[i],mycols,vals,INSERT_VALUES);CHKERRQ(ierr); 2163 mycols += locrowlens[i]; 2164 vals += locrowlens[i]; 2165 jj++; 2166 } 2167 2168 /* read in other processors (except the last one) and ship out */ 2169 for (i=1; i<size-1; i++) { 2170 nz = procsnz[i]; 2171 vals = buf; 2172 ierr = PetscBinaryRead(fd,vals,nz,PETSC_SCALAR);CHKERRQ(ierr); 2173 ierr = MPI_Send(vals,nz,MPIU_SCALAR,i,A->tag,comm);CHKERRQ(ierr); 2174 } 2175 /* the last proc */ 2176 if (size != 1){ 2177 nz = procsnz[i] - extra_rows; 2178 vals = buf; 2179 ierr = PetscBinaryRead(fd,vals,nz,PETSC_SCALAR);CHKERRQ(ierr); 2180 for (i=0; i<extra_rows; i++) vals[nz+i] = 1.0; 2181 ierr = MPI_Send(vals,nz+extra_rows,MPIU_SCALAR,size-1,A->tag,comm);CHKERRQ(ierr); 2182 } 2183 ierr = PetscFree(procsnz);CHKERRQ(ierr); 2184 2185 } else { 2186 /* receive numeric values */ 2187 ierr = PetscMalloc(nz*sizeof(Scalar),&buf);CHKERRQ(ierr); 2188 2189 /* receive message of values*/ 2190 vals = buf; 2191 mycols = ibuf; 2192 ierr = MPI_Recv(vals,nz,MPIU_SCALAR,0,A->tag,comm,&status);CHKERRQ(ierr); 2193 ierr = MPI_Get_count(&status,MPIU_SCALAR,&maxnz);CHKERRQ(ierr); 2194 if (maxnz != nz) SETERRQ(PETSC_ERR_FILE_UNEXPECTED,"something is wrong with file"); 2195 2196 /* insert into matrix */ 2197 jj = rstart*bs; 2198 for (i=0; i<m; i++) { 2199 ierr = MatSetValues_MPISBAIJ(A,1,&jj,locrowlens[i],mycols,vals,INSERT_VALUES);CHKERRQ(ierr); 2200 mycols += locrowlens[i]; 2201 vals += locrowlens[i]; 2202 jj++; 2203 } 2204 } 2205 2206 ierr = PetscFree(locrowlens);CHKERRQ(ierr); 2207 ierr = PetscFree(buf);CHKERRQ(ierr); 2208 ierr = PetscFree(ibuf);CHKERRQ(ierr); 2209 ierr = PetscFree(rowners);CHKERRQ(ierr); 2210 ierr = PetscFree(dlens);CHKERRQ(ierr); 2211 ierr = PetscFree(mask);CHKERRQ(ierr); 2212 ierr = MatAssemblyBegin(A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2213 ierr = MatAssemblyEnd(A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2214 PetscFunctionReturn(0); 2215 } 2216 EXTERN_C_END 2217 2218 #undef __FUNCT__ 2219 #define __FUNCT__ "MatMPISBAIJSetHashTableFactor" 2220 /*@ 2221 MatMPISBAIJSetHashTableFactor - Sets the factor required to compute the size of the HashTable. 2222 2223 Input Parameters: 2224 . mat - the matrix 2225 . fact - factor 2226 2227 Collective on Mat 2228 2229 Level: advanced 2230 2231 Notes: 2232 This can also be set by the command line option: -mat_use_hash_table fact 2233 2234 .keywords: matrix, hashtable, factor, HT 2235 2236 .seealso: MatSetOption() 2237 @*/ 2238 int MatMPISBAIJSetHashTableFactor(Mat mat,PetscReal fact) 2239 { 2240 PetscFunctionBegin; 2241 SETERRQ(1,"Function not yet written for SBAIJ format"); 2242 /* PetscFunctionReturn(0); */ 2243 } 2244 2245 #undef __FUNCT__ 2246 #define __FUNCT__ "MatGetRowMax_MPISBAIJ" 2247 int MatGetRowMax_MPISBAIJ(Mat A,Vec v) 2248 { 2249 Mat_MPISBAIJ *a = (Mat_MPISBAIJ*)A->data; 2250 Mat_SeqBAIJ *b = (Mat_SeqBAIJ*)(a->B)->data; 2251 PetscReal atmp; 2252 double *work,*svalues,*rvalues; 2253 int ierr,i,bs,mbs,*bi,*bj,brow,j,ncols,krow,kcol,col,row,Mbs,bcol; 2254 int rank,size,*rowners_bs,dest,count,source; 2255 Scalar *va; 2256 MatScalar *ba; 2257 MPI_Status stat; 2258 2259 PetscFunctionBegin; 2260 ierr = MatGetRowMax(a->A,v);CHKERRQ(ierr); 2261 ierr = VecGetArray(v,&va);CHKERRQ(ierr); 2262 2263 ierr = MPI_Comm_size(PETSC_COMM_WORLD,&size);CHKERRQ(ierr); 2264 ierr = MPI_Comm_rank(PETSC_COMM_WORLD,&rank);CHKERRQ(ierr); 2265 2266 bs = a->bs; 2267 mbs = a->mbs; 2268 Mbs = a->Mbs; 2269 ba = b->a; 2270 bi = b->i; 2271 bj = b->j; 2272 /* 2273 PetscSynchronizedPrintf(PETSC_COMM_WORLD,"[%d] M: %d, bs: %d, mbs: %d \n",rank,bs*Mbs,bs,mbs); 2274 PetscSynchronizedFlush(PETSC_COMM_WORLD); 2275 */ 2276 2277 /* find ownerships */ 2278 rowners_bs = a->rowners_bs; 2279 /* 2280 if (!rank){ 2281 for (i=0; i<size+1; i++) PetscPrintf(PETSC_COMM_SELF," rowners_bs[%d]: %d\n",i,rowners_bs[i]); 2282 } 2283 */ 2284 2285 /* each proc creates an array to be distributed */ 2286 ierr = PetscMalloc(bs*Mbs*sizeof(PetscReal),&work);CHKERRQ(ierr); 2287 ierr = PetscMemzero(work,bs*Mbs*sizeof(PetscReal));CHKERRQ(ierr); 2288 2289 /* row_max for B */ 2290 if (rank != size-1){ 2291 for (i=0; i<mbs; i++) { 2292 ncols = bi[1] - bi[0]; bi++; 2293 brow = bs*i; 2294 for (j=0; j<ncols; j++){ 2295 bcol = bs*(*bj); 2296 for (kcol=0; kcol<bs; kcol++){ 2297 col = bcol + kcol; /* local col index */ 2298 col += rowners_bs[rank+1]; /* global col index */ 2299 /* PetscPrintf(PETSC_COMM_SELF,"[%d], col: %d\n",rank,col); */ 2300 for (krow=0; krow<bs; krow++){ 2301 atmp = PetscAbsScalar(*ba); ba++; 2302 row = brow + krow; /* local row index */ 2303 /* printf("val[%d,%d]: %g\n",row,col,atmp); */ 2304 if (PetscRealPart(va[row]) < atmp) va[row] = atmp; 2305 if (work[col] < atmp) work[col] = atmp; 2306 } 2307 } 2308 bj++; 2309 } 2310 } 2311 /* 2312 PetscPrintf(PETSC_COMM_SELF,"[%d], work: ",rank); 2313 for (i=0; i<bs*Mbs; i++) PetscPrintf(PETSC_COMM_SELF,"%g ",work[i]); 2314 PetscPrintf(PETSC_COMM_SELF,"[%d]: \n"); 2315 */ 2316 2317 /* send values to its owners */ 2318 for (dest=rank+1; dest<size; dest++){ 2319 svalues = work + rowners_bs[dest]; 2320 count = rowners_bs[dest+1]-rowners_bs[dest]; 2321 ierr = MPI_Send(svalues,count,MPI_DOUBLE,dest,rank,PETSC_COMM_WORLD);CHKERRQ(ierr); 2322 /* 2323 PetscSynchronizedPrintf(PETSC_COMM_WORLD,"[%d] sends %d values to [%d]: %g, %g, %g, %g\n",rank,count,dest,svalues[0],svalues[1],svalues[2],svalues[3]); 2324 PetscSynchronizedFlush(PETSC_COMM_WORLD); 2325 */ 2326 } 2327 } 2328 2329 /* receive values */ 2330 if (rank){ 2331 rvalues = work; 2332 count = rowners_bs[rank+1]-rowners_bs[rank]; 2333 for (source=0; source<rank; source++){ 2334 ierr = MPI_Recv(rvalues,count,MPI_DOUBLE,MPI_ANY_SOURCE,MPI_ANY_TAG,PETSC_COMM_WORLD,&stat);CHKERRQ(ierr); 2335 /* process values */ 2336 for (i=0; i<count; i++){ 2337 if (PetscRealPart(va[i]) < rvalues[i]) va[i] = rvalues[i]; 2338 } 2339 /* 2340 PetscSynchronizedPrintf(PETSC_COMM_WORLD,"[%d] received %d values from [%d]: %g, %g, %g, %g \n",rank,count,stat.MPI_SOURCE,rvalues[0],rvalues[1],rvalues[2],rvalues[3]); 2341 PetscSynchronizedFlush(PETSC_COMM_WORLD); 2342 */ 2343 } 2344 } 2345 2346 ierr = VecRestoreArray(v,&va);CHKERRQ(ierr); 2347 ierr = PetscFree(work);CHKERRQ(ierr); 2348 PetscFunctionReturn(0); 2349 } 2350