1 /*$Id: mpisbaij.c,v 1.43 2001/01/20 03:35:02 bsmith Exp bsmith $*/ 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 __FUNC__ 45 #define __FUNC__ "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 __FUNC__ 60 #define __FUNC__ "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 __FUNC__ 80 #define __FUNC__ "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 __FUNC__ 246 #define __FUNC__ "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 __FUNC__ 269 #define __FUNC__ "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 __FUNC__ 291 #define __FUNC__ "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 __FUNC__ 304 #define __FUNC__ "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 __FUNC__ 321 #define __FUNC__ "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 __FUNC__ 430 #define __FUNC__ "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 __FUNC__ 443 #define __FUNC__ "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 __FUNC__ 452 #define __FUNC__ "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 __FUNC__ 461 #define __FUNC__ "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 __FUNC__ 505 #define __FUNC__ "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 __FUNC__ 550 #define __FUNC__ "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 __FUNC__ 559 #define __FUNC__ "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 __FUNC__ 588 #define __FUNC__ "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 __FUNC__ 696 #define __FUNC__ "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 = MatView(baij->A,viewer);CHKERRQ(ierr); 739 } else { 740 /* assemble the entire matrix onto first processor. */ 741 Mat A; 742 Mat_SeqSBAIJ *Aloc; 743 Mat_SeqBAIJ *Bloc; 744 int M = mat->M,N = mat->N,*ai,*aj,col,i,j,k,*rvals,mbs = baij->mbs; 745 MatScalar *a; 746 747 if (!rank) { 748 ierr = MatCreateMPISBAIJ(mat->comm,baij->bs,M,N,M,N,0,PETSC_NULL,0,PETSC_NULL,&A);CHKERRQ(ierr); 749 } else { 750 ierr = MatCreateMPISBAIJ(mat->comm,baij->bs,0,0,M,N,0,PETSC_NULL,0,PETSC_NULL,&A);CHKERRQ(ierr); 751 } 752 PetscLogObjectParent(mat,A); 753 754 /* copy over the A part */ 755 Aloc = (Mat_SeqSBAIJ*)baij->A->data; 756 ai = Aloc->i; aj = Aloc->j; a = Aloc->a; 757 ierr = PetscMalloc(bs*sizeof(int),&rvals);CHKERRQ(ierr); 758 759 for (i=0; i<mbs; i++) { 760 rvals[0] = bs*(baij->rstart + i); 761 for (j=1; j<bs; j++) { rvals[j] = rvals[j-1] + 1; } 762 for (j=ai[i]; j<ai[i+1]; j++) { 763 col = (baij->cstart+aj[j])*bs; 764 for (k=0; k<bs; k++) { 765 ierr = MatSetValues_MPISBAIJ_MatScalar(A,bs,rvals,1,&col,a,INSERT_VALUES);CHKERRQ(ierr); 766 col++; a += bs; 767 } 768 } 769 } 770 /* copy over the B part */ 771 Bloc = (Mat_SeqBAIJ*)baij->B->data; 772 ai = Bloc->i; aj = Bloc->j; a = Bloc->a; 773 for (i=0; i<mbs; i++) { 774 rvals[0] = bs*(baij->rstart + i); 775 for (j=1; j<bs; j++) { rvals[j] = rvals[j-1] + 1; } 776 for (j=ai[i]; j<ai[i+1]; j++) { 777 col = baij->garray[aj[j]]*bs; 778 for (k=0; k<bs; k++) { 779 ierr = MatSetValues_MPISBAIJ_MatScalar(A,bs,rvals,1,&col,a,INSERT_VALUES);CHKERRQ(ierr); 780 col++; a += bs; 781 } 782 } 783 } 784 ierr = PetscFree(rvals);CHKERRQ(ierr); 785 ierr = MatAssemblyBegin(A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 786 ierr = MatAssemblyEnd(A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 787 /* 788 Everyone has to call to draw the matrix since the graphics waits are 789 synchronized across all processors that share the PetscDraw object 790 */ 791 ierr = PetscViewerGetSingleton(viewer,&sviewer);CHKERRQ(ierr); 792 if (!rank) { 793 ierr = MatView(((Mat_MPISBAIJ*)(A->data))->A,sviewer);CHKERRQ(ierr); 794 } 795 ierr = PetscViewerRestoreSingleton(viewer,&sviewer);CHKERRQ(ierr); 796 ierr = MatDestroy(A);CHKERRQ(ierr); 797 } 798 PetscFunctionReturn(0); 799 } 800 801 #undef __FUNC__ 802 #define __FUNC__ "MatView_MPISBAIJ" 803 int MatView_MPISBAIJ(Mat mat,PetscViewer viewer) 804 { 805 int ierr; 806 PetscTruth isascii,isdraw,issocket,isbinary; 807 808 PetscFunctionBegin; 809 ierr = PetscTypeCompare((PetscObject)viewer,PETSC_VIEWER_ASCII,&isascii);CHKERRQ(ierr); 810 ierr = PetscTypeCompare((PetscObject)viewer,PETSC_VIEWER_DRAW,&isdraw);CHKERRQ(ierr); 811 ierr = PetscTypeCompare((PetscObject)viewer,PETSC_VIEWER_SOCKET,&issocket);CHKERRQ(ierr); 812 ierr = PetscTypeCompare((PetscObject)viewer,PETSC_VIEWER_BINARY,&isbinary);CHKERRQ(ierr); 813 if (isascii || isdraw || issocket || isbinary) { 814 ierr = MatView_MPISBAIJ_ASCIIorDraworSocket(mat,viewer);CHKERRQ(ierr); 815 } else { 816 SETERRQ1(1,"Viewer type %s not supported by MPISBAIJ matrices",((PetscObject)viewer)->type_name); 817 } 818 PetscFunctionReturn(0); 819 } 820 821 #undef __FUNC__ 822 #define __FUNC__ "MatDestroy_MPISBAIJ" 823 int MatDestroy_MPISBAIJ(Mat mat) 824 { 825 Mat_MPISBAIJ *baij = (Mat_MPISBAIJ*)mat->data; 826 int ierr; 827 828 PetscFunctionBegin; 829 #if defined(PETSC_USE_LOG) 830 PetscLogObjectState((PetscObject)mat,"Rows=%d,Cols=%d",mat->M,mat->N); 831 #endif 832 ierr = MatStashDestroy_Private(&mat->stash);CHKERRQ(ierr); 833 ierr = MatStashDestroy_Private(&mat->bstash);CHKERRQ(ierr); 834 ierr = PetscFree(baij->rowners);CHKERRQ(ierr); 835 ierr = MatDestroy(baij->A);CHKERRQ(ierr); 836 ierr = MatDestroy(baij->B);CHKERRQ(ierr); 837 #if defined (PETSC_USE_CTABLE) 838 if (baij->colmap) {ierr = PetscTableDelete(baij->colmap);CHKERRQ(ierr);} 839 #else 840 if (baij->colmap) {ierr = PetscFree(baij->colmap);CHKERRQ(ierr);} 841 #endif 842 if (baij->garray) {ierr = PetscFree(baij->garray);CHKERRQ(ierr);} 843 if (baij->lvec) {ierr = VecDestroy(baij->lvec);CHKERRQ(ierr);} 844 if (baij->Mvctx) {ierr = VecScatterDestroy(baij->Mvctx);CHKERRQ(ierr);} 845 if (baij->rowvalues) {ierr = PetscFree(baij->rowvalues);CHKERRQ(ierr);} 846 if (baij->barray) {ierr = PetscFree(baij->barray);CHKERRQ(ierr);} 847 if (baij->hd) {ierr = PetscFree(baij->hd);CHKERRQ(ierr);} 848 #if defined(PETSC_USE_MAT_SINGLE) 849 if (baij->setvaluescopy) {ierr = PetscFree(baij->setvaluescopy);CHKERRQ(ierr);} 850 #endif 851 ierr = PetscFree(baij);CHKERRQ(ierr); 852 PetscFunctionReturn(0); 853 } 854 855 #undef __FUNC__ 856 #define __FUNC__ "MatMult_MPISBAIJ" 857 int MatMult_MPISBAIJ(Mat A,Vec xx,Vec yy) 858 { 859 Mat_MPISBAIJ *a = (Mat_MPISBAIJ*)A->data; 860 int ierr,nt; 861 862 PetscFunctionBegin; 863 ierr = VecGetLocalSize(xx,&nt);CHKERRQ(ierr); 864 if (nt != A->n) { 865 SETERRQ(PETSC_ERR_ARG_SIZ,"Incompatible partition of A and xx"); 866 } 867 ierr = VecGetLocalSize(yy,&nt);CHKERRQ(ierr); 868 if (nt != A->m) { 869 SETERRQ(PETSC_ERR_ARG_SIZ,"Incompatible parition of A and yy"); 870 } 871 872 ierr = VecScatterBegin(xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD,a->Mvctx);CHKERRQ(ierr); 873 /* do diagonal part */ 874 ierr = (*a->A->ops->mult)(a->A,xx,yy);CHKERRQ(ierr); 875 /* do supperdiagonal part */ 876 ierr = VecScatterEnd(xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD,a->Mvctx);CHKERRQ(ierr); 877 ierr = (*a->B->ops->multadd)(a->B,a->lvec,yy,yy);CHKERRQ(ierr); 878 /* do subdiagonal part */ 879 ierr = (*a->B->ops->multtranspose)(a->B,xx,a->lvec);CHKERRQ(ierr); 880 ierr = VecScatterBegin(a->lvec,yy,ADD_VALUES,SCATTER_REVERSE,a->Mvctx);CHKERRQ(ierr); 881 ierr = VecScatterEnd(a->lvec,yy,ADD_VALUES,SCATTER_REVERSE,a->Mvctx);CHKERRQ(ierr); 882 883 PetscFunctionReturn(0); 884 } 885 886 #undef __FUNC__ 887 #define __FUNC__ "MatMultAdd_MPISBAIJ" 888 int MatMultAdd_MPISBAIJ(Mat A,Vec xx,Vec yy,Vec zz) 889 { 890 Mat_MPISBAIJ *a = (Mat_MPISBAIJ*)A->data; 891 int ierr; 892 893 PetscFunctionBegin; 894 ierr = VecScatterBegin(xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD,a->Mvctx);CHKERRQ(ierr); 895 /* do diagonal part */ 896 ierr = (*a->A->ops->multadd)(a->A,xx,yy,zz);CHKERRQ(ierr); 897 /* do supperdiagonal part */ 898 ierr = VecScatterEnd(xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD,a->Mvctx);CHKERRQ(ierr); 899 ierr = (*a->B->ops->multadd)(a->B,a->lvec,zz,zz);CHKERRQ(ierr); 900 901 /* do subdiagonal part */ 902 ierr = (*a->B->ops->multtranspose)(a->B,xx,a->lvec);CHKERRQ(ierr); 903 ierr = VecScatterBegin(a->lvec,zz,ADD_VALUES,SCATTER_REVERSE,a->Mvctx);CHKERRQ(ierr); 904 ierr = VecScatterEnd(a->lvec,zz,ADD_VALUES,SCATTER_REVERSE,a->Mvctx);CHKERRQ(ierr); 905 906 PetscFunctionReturn(0); 907 } 908 909 #undef __FUNC__ 910 #define __FUNC__ "MatMultTranspose_MPISBAIJ" 911 int MatMultTranspose_MPISBAIJ(Mat A,Vec xx,Vec yy) 912 { 913 PetscFunctionBegin; 914 SETERRQ(1,"Matrix is symmetric. Call MatMult()."); 915 /* PetscFunctionReturn(0); */ 916 } 917 918 #undef __FUNC__ 919 #define __FUNC__ "MatMultTransposeAdd_MPISBAIJ" 920 int MatMultTransposeAdd_MPISBAIJ(Mat A,Vec xx,Vec yy,Vec zz) 921 { 922 PetscFunctionBegin; 923 SETERRQ(1,"Matrix is symmetric. Call MatMultAdd()."); 924 /* PetscFunctionReturn(0); */ 925 } 926 927 /* 928 This only works correctly for square matrices where the subblock A->A is the 929 diagonal block 930 */ 931 #undef __FUNC__ 932 #define __FUNC__ "MatGetDiagonal_MPISBAIJ" 933 int MatGetDiagonal_MPISBAIJ(Mat A,Vec v) 934 { 935 Mat_MPISBAIJ *a = (Mat_MPISBAIJ*)A->data; 936 int ierr; 937 938 PetscFunctionBegin; 939 /* if (a->M != a->N) SETERRQ(PETSC_ERR_SUP,"Supports only square matrix where A->A is diag block"); */ 940 ierr = MatGetDiagonal(a->A,v);CHKERRQ(ierr); 941 PetscFunctionReturn(0); 942 } 943 944 #undef __FUNC__ 945 #define __FUNC__ "MatScale_MPISBAIJ" 946 int MatScale_MPISBAIJ(Scalar *aa,Mat A) 947 { 948 Mat_MPISBAIJ *a = (Mat_MPISBAIJ*)A->data; 949 int ierr; 950 951 PetscFunctionBegin; 952 ierr = MatScale(aa,a->A);CHKERRQ(ierr); 953 ierr = MatScale(aa,a->B);CHKERRQ(ierr); 954 PetscFunctionReturn(0); 955 } 956 957 #undef __FUNC__ 958 #define __FUNC__ "MatGetOwnershipRange_MPISBAIJ" 959 int MatGetOwnershipRange_MPISBAIJ(Mat matin,int *m,int *n) 960 { 961 Mat_MPISBAIJ *mat = (Mat_MPISBAIJ*)matin->data; 962 963 PetscFunctionBegin; 964 if (m) *m = mat->rstart*mat->bs; 965 if (n) *n = mat->rend*mat->bs; 966 PetscFunctionReturn(0); 967 } 968 969 #undef __FUNC__ 970 #define __FUNC__ "MatGetRow_MPISBAIJ" 971 int MatGetRow_MPISBAIJ(Mat matin,int row,int *nz,int **idx,Scalar **v) 972 { 973 Mat_MPISBAIJ *mat = (Mat_MPISBAIJ*)matin->data; 974 Scalar *vworkA,*vworkB,**pvA,**pvB,*v_p; 975 int bs = mat->bs,bs2 = mat->bs2,i,ierr,*cworkA,*cworkB,**pcA,**pcB; 976 int nztot,nzA,nzB,lrow,brstart = mat->rstart*bs,brend = mat->rend*bs; 977 int *cmap,*idx_p,cstart = mat->cstart; 978 979 PetscFunctionBegin; 980 if (mat->getrowactive == PETSC_TRUE) SETERRQ(PETSC_ERR_ARG_WRONGSTATE,"Already active"); 981 mat->getrowactive = PETSC_TRUE; 982 983 if (!mat->rowvalues && (idx || v)) { 984 /* 985 allocate enough space to hold information from the longest row. 986 */ 987 Mat_SeqSBAIJ *Aa = (Mat_SeqSBAIJ*)mat->A->data; 988 Mat_SeqBAIJ *Ba = (Mat_SeqBAIJ*)mat->B->data; 989 int max = 1,mbs = mat->mbs,tmp; 990 for (i=0; i<mbs; i++) { 991 tmp = Aa->i[i+1] - Aa->i[i] + Ba->i[i+1] - Ba->i[i]; /* row length */ 992 if (max < tmp) { max = tmp; } 993 } 994 ierr = PetscMalloc(max*bs2*(sizeof(int)+sizeof(Scalar)),&mat->rowvalues);CHKERRQ(ierr); 995 mat->rowindices = (int*)(mat->rowvalues + max*bs2); 996 } 997 998 if (row < brstart || row >= brend) SETERRQ(PETSC_ERR_SUP,"Only local rows") 999 lrow = row - brstart; /* local row index */ 1000 1001 pvA = &vworkA; pcA = &cworkA; pvB = &vworkB; pcB = &cworkB; 1002 if (!v) {pvA = 0; pvB = 0;} 1003 if (!idx) {pcA = 0; if (!v) pcB = 0;} 1004 ierr = (*mat->A->ops->getrow)(mat->A,lrow,&nzA,pcA,pvA);CHKERRQ(ierr); 1005 ierr = (*mat->B->ops->getrow)(mat->B,lrow,&nzB,pcB,pvB);CHKERRQ(ierr); 1006 nztot = nzA + nzB; 1007 1008 cmap = mat->garray; 1009 if (v || idx) { 1010 if (nztot) { 1011 /* Sort by increasing column numbers, assuming A and B already sorted */ 1012 int imark = -1; 1013 if (v) { 1014 *v = v_p = mat->rowvalues; 1015 for (i=0; i<nzB; i++) { 1016 if (cmap[cworkB[i]/bs] < cstart) v_p[i] = vworkB[i]; 1017 else break; 1018 } 1019 imark = i; 1020 for (i=0; i<nzA; i++) v_p[imark+i] = vworkA[i]; 1021 for (i=imark; i<nzB; i++) v_p[nzA+i] = vworkB[i]; 1022 } 1023 if (idx) { 1024 *idx = idx_p = mat->rowindices; 1025 if (imark > -1) { 1026 for (i=0; i<imark; i++) { 1027 idx_p[i] = cmap[cworkB[i]/bs]*bs + cworkB[i]%bs; 1028 } 1029 } else { 1030 for (i=0; i<nzB; i++) { 1031 if (cmap[cworkB[i]/bs] < cstart) 1032 idx_p[i] = cmap[cworkB[i]/bs]*bs + cworkB[i]%bs ; 1033 else break; 1034 } 1035 imark = i; 1036 } 1037 for (i=0; i<nzA; i++) idx_p[imark+i] = cstart*bs + cworkA[i]; 1038 for (i=imark; i<nzB; i++) idx_p[nzA+i] = cmap[cworkB[i]/bs]*bs + cworkB[i]%bs ; 1039 } 1040 } else { 1041 if (idx) *idx = 0; 1042 if (v) *v = 0; 1043 } 1044 } 1045 *nz = nztot; 1046 ierr = (*mat->A->ops->restorerow)(mat->A,lrow,&nzA,pcA,pvA);CHKERRQ(ierr); 1047 ierr = (*mat->B->ops->restorerow)(mat->B,lrow,&nzB,pcB,pvB);CHKERRQ(ierr); 1048 PetscFunctionReturn(0); 1049 } 1050 1051 #undef __FUNC__ 1052 #define __FUNC__ "MatRestoreRow_MPISBAIJ" 1053 int MatRestoreRow_MPISBAIJ(Mat mat,int row,int *nz,int **idx,Scalar **v) 1054 { 1055 Mat_MPISBAIJ *baij = (Mat_MPISBAIJ*)mat->data; 1056 1057 PetscFunctionBegin; 1058 if (baij->getrowactive == PETSC_FALSE) { 1059 SETERRQ(PETSC_ERR_ARG_WRONGSTATE,"MatGetRow not called"); 1060 } 1061 baij->getrowactive = PETSC_FALSE; 1062 PetscFunctionReturn(0); 1063 } 1064 1065 #undef __FUNC__ 1066 #define __FUNC__ "MatGetBlockSize_MPISBAIJ" 1067 int MatGetBlockSize_MPISBAIJ(Mat mat,int *bs) 1068 { 1069 Mat_MPISBAIJ *baij = (Mat_MPISBAIJ*)mat->data; 1070 1071 PetscFunctionBegin; 1072 *bs = baij->bs; 1073 PetscFunctionReturn(0); 1074 } 1075 1076 #undef __FUNC__ 1077 #define __FUNC__ "MatZeroEntries_MPISBAIJ" 1078 int MatZeroEntries_MPISBAIJ(Mat A) 1079 { 1080 Mat_MPISBAIJ *l = (Mat_MPISBAIJ*)A->data; 1081 int ierr; 1082 1083 PetscFunctionBegin; 1084 ierr = MatZeroEntries(l->A);CHKERRQ(ierr); 1085 ierr = MatZeroEntries(l->B);CHKERRQ(ierr); 1086 PetscFunctionReturn(0); 1087 } 1088 1089 #undef __FUNC__ 1090 #define __FUNC__ "MatGetInfo_MPISBAIJ" 1091 int MatGetInfo_MPISBAIJ(Mat matin,MatInfoType flag,MatInfo *info) 1092 { 1093 Mat_MPISBAIJ *a = (Mat_MPISBAIJ*)matin->data; 1094 Mat A = a->A,B = a->B; 1095 int ierr; 1096 PetscReal isend[5],irecv[5]; 1097 1098 PetscFunctionBegin; 1099 info->block_size = (double)a->bs; 1100 ierr = MatGetInfo(A,MAT_LOCAL,info);CHKERRQ(ierr); 1101 isend[0] = info->nz_used; isend[1] = info->nz_allocated; isend[2] = info->nz_unneeded; 1102 isend[3] = info->memory; isend[4] = info->mallocs; 1103 ierr = MatGetInfo(B,MAT_LOCAL,info);CHKERRQ(ierr); 1104 isend[0] += info->nz_used; isend[1] += info->nz_allocated; isend[2] += info->nz_unneeded; 1105 isend[3] += info->memory; isend[4] += info->mallocs; 1106 if (flag == MAT_LOCAL) { 1107 info->nz_used = isend[0]; 1108 info->nz_allocated = isend[1]; 1109 info->nz_unneeded = isend[2]; 1110 info->memory = isend[3]; 1111 info->mallocs = isend[4]; 1112 } else if (flag == MAT_GLOBAL_MAX) { 1113 ierr = MPI_Allreduce(isend,irecv,5,MPI_DOUBLE,MPI_MAX,matin->comm);CHKERRQ(ierr); 1114 info->nz_used = irecv[0]; 1115 info->nz_allocated = irecv[1]; 1116 info->nz_unneeded = irecv[2]; 1117 info->memory = irecv[3]; 1118 info->mallocs = irecv[4]; 1119 } else if (flag == MAT_GLOBAL_SUM) { 1120 ierr = MPI_Allreduce(isend,irecv,5,MPI_DOUBLE,MPI_SUM,matin->comm);CHKERRQ(ierr); 1121 info->nz_used = irecv[0]; 1122 info->nz_allocated = irecv[1]; 1123 info->nz_unneeded = irecv[2]; 1124 info->memory = irecv[3]; 1125 info->mallocs = irecv[4]; 1126 } else { 1127 SETERRQ1(1,"Unknown MatInfoType argument %d",flag); 1128 } 1129 info->rows_global = (double)A->M; 1130 info->columns_global = (double)A->N; 1131 info->rows_local = (double)A->m; 1132 info->columns_local = (double)A->N; 1133 info->fill_ratio_given = 0; /* no parallel LU/ILU/Cholesky */ 1134 info->fill_ratio_needed = 0; 1135 info->factor_mallocs = 0; 1136 PetscFunctionReturn(0); 1137 } 1138 1139 #undef __FUNC__ 1140 #define __FUNC__ "MatSetOption_MPISBAIJ" 1141 int MatSetOption_MPISBAIJ(Mat A,MatOption op) 1142 { 1143 Mat_MPISBAIJ *a = (Mat_MPISBAIJ*)A->data; 1144 int ierr; 1145 1146 PetscFunctionBegin; 1147 if (op == MAT_NO_NEW_NONZERO_LOCATIONS || 1148 op == MAT_YES_NEW_NONZERO_LOCATIONS || 1149 op == MAT_COLUMNS_UNSORTED || 1150 op == MAT_COLUMNS_SORTED || 1151 op == MAT_NEW_NONZERO_ALLOCATION_ERR || 1152 op == MAT_KEEP_ZEROED_ROWS || 1153 op == MAT_NEW_NONZERO_LOCATION_ERR) { 1154 ierr = MatSetOption(a->A,op);CHKERRQ(ierr); 1155 ierr = MatSetOption(a->B,op);CHKERRQ(ierr); 1156 } else if (op == MAT_ROW_ORIENTED) { 1157 a->roworiented = PETSC_TRUE; 1158 ierr = MatSetOption(a->A,op);CHKERRQ(ierr); 1159 ierr = MatSetOption(a->B,op);CHKERRQ(ierr); 1160 } else if (op == MAT_ROWS_SORTED || 1161 op == MAT_ROWS_UNSORTED || 1162 op == MAT_SYMMETRIC || 1163 op == MAT_STRUCTURALLY_SYMMETRIC || 1164 op == MAT_YES_NEW_DIAGONALS || 1165 op == MAT_USE_HASH_TABLE) { 1166 PetscLogInfo(A,"Info:MatSetOption_MPIBAIJ:Option ignored\n"); 1167 } else if (op == MAT_COLUMN_ORIENTED) { 1168 a->roworiented = PETSC_FALSE; 1169 ierr = MatSetOption(a->A,op);CHKERRQ(ierr); 1170 ierr = MatSetOption(a->B,op);CHKERRQ(ierr); 1171 } else if (op == MAT_IGNORE_OFF_PROC_ENTRIES) { 1172 a->donotstash = PETSC_TRUE; 1173 } else if (op == MAT_NO_NEW_DIAGONALS) { 1174 SETERRQ(PETSC_ERR_SUP,"MAT_NO_NEW_DIAGONALS"); 1175 } else if (op == MAT_USE_HASH_TABLE) { 1176 a->ht_flag = PETSC_TRUE; 1177 } else { 1178 SETERRQ(PETSC_ERR_SUP,"unknown option"); 1179 } 1180 PetscFunctionReturn(0); 1181 } 1182 1183 #undef __FUNC__ 1184 #define __FUNC__ "MatTranspose_MPISBAIJ(" 1185 int MatTranspose_MPISBAIJ(Mat A,Mat *matout) 1186 { 1187 PetscFunctionBegin; 1188 SETERRQ(1,"Matrix is symmetric. MatTranspose() should not be called"); 1189 /* PetscFunctionReturn(0); */ 1190 } 1191 1192 #undef __FUNC__ 1193 #define __FUNC__ "MatDiagonalScale_MPISBAIJ" 1194 int MatDiagonalScale_MPISBAIJ(Mat mat,Vec ll,Vec rr) 1195 { 1196 Mat_MPISBAIJ *baij = (Mat_MPISBAIJ*)mat->data; 1197 Mat a = baij->A,b = baij->B; 1198 int ierr,s1,s2,s3; 1199 1200 PetscFunctionBegin; 1201 if (ll != rr) { 1202 SETERRQ(PETSC_ERR_ARG_OUTOFRANGE,"For symmetric format, left and right scaling vectors must be same\n"); 1203 } 1204 ierr = MatGetLocalSize(mat,&s2,&s3);CHKERRQ(ierr); 1205 if (rr) { 1206 ierr = VecGetLocalSize(rr,&s1);CHKERRQ(ierr); 1207 if (s1!=s3) SETERRQ(PETSC_ERR_ARG_SIZ,"right vector non-conforming local size"); 1208 /* Overlap communication with computation. */ 1209 ierr = VecScatterBegin(rr,baij->lvec,INSERT_VALUES,SCATTER_FORWARD,baij->Mvctx);CHKERRQ(ierr); 1210 /*} if (ll) { */ 1211 ierr = VecGetLocalSize(ll,&s1);CHKERRQ(ierr); 1212 if (s1!=s2) SETERRQ(PETSC_ERR_ARG_SIZ,"left vector non-conforming local size"); 1213 ierr = (*b->ops->diagonalscale)(b,ll,PETSC_NULL);CHKERRQ(ierr); 1214 /* } */ 1215 /* scale the diagonal block */ 1216 ierr = (*a->ops->diagonalscale)(a,ll,rr);CHKERRQ(ierr); 1217 1218 /* if (rr) { */ 1219 /* Do a scatter end and then right scale the off-diagonal block */ 1220 ierr = VecScatterEnd(rr,baij->lvec,INSERT_VALUES,SCATTER_FORWARD,baij->Mvctx);CHKERRQ(ierr); 1221 ierr = (*b->ops->diagonalscale)(b,PETSC_NULL,baij->lvec);CHKERRQ(ierr); 1222 } 1223 1224 PetscFunctionReturn(0); 1225 } 1226 1227 #undef __FUNC__ 1228 #define __FUNC__ "MatZeroRows_MPISBAIJ" 1229 int MatZeroRows_MPISBAIJ(Mat A,IS is,Scalar *diag) 1230 { 1231 Mat_MPISBAIJ *l = (Mat_MPISBAIJ*)A->data; 1232 int i,ierr,N,*rows,*owners = l->rowners,size = l->size; 1233 int *procs,*nprocs,j,idx,nsends,*work,row; 1234 int nmax,*svalues,*starts,*owner,nrecvs,rank = l->rank; 1235 int *rvalues,tag = A->tag,count,base,slen,n,*source; 1236 int *lens,imdex,*lrows,*values,bs=l->bs,rstart_bs=l->rstart_bs; 1237 MPI_Comm comm = A->comm; 1238 MPI_Request *send_waits,*recv_waits; 1239 MPI_Status recv_status,*send_status; 1240 IS istmp; 1241 PetscTruth found; 1242 1243 PetscFunctionBegin; 1244 ierr = ISGetSize(is,&N);CHKERRQ(ierr); 1245 ierr = ISGetIndices(is,&rows);CHKERRQ(ierr); 1246 1247 /* first count number of contributors to each processor */ 1248 ierr = PetscMalloc(2*size*sizeof(int),&nprocs);CHKERRQ(ierr); 1249 ierr = PetscMemzero(nprocs,2*size*sizeof(int));CHKERRQ(ierr); 1250 procs = nprocs + size; 1251 ierr = PetscMalloc((N+1)*sizeof(int),&owner);CHKERRQ(ierr); /* see note*/ 1252 for (i=0; i<N; i++) { 1253 idx = rows[i]; 1254 found = PETSC_FALSE; 1255 for (j=0; j<size; j++) { 1256 if (idx >= owners[j]*bs && idx < owners[j+1]*bs) { 1257 nprocs[j]++; procs[j] = 1; owner[i] = j; found = PETSC_TRUE; break; 1258 } 1259 } 1260 if (!found) SETERRQ(PETSC_ERR_ARG_OUTOFRANGE,"Index out of range"); 1261 } 1262 nsends = 0; for (i=0; i<size; i++) { nsends += procs[i];} 1263 1264 /* inform other processors of number of messages and max length*/ 1265 ierr = PetscMalloc(2*size*sizeof(int),&work);CHKERRQ(ierr); 1266 ierr = MPI_Allreduce(nprocs,work,2*size,MPI_INT,PetscMaxSum_Op,comm);CHKERRQ(ierr); 1267 nmax = work[rank]; 1268 nrecvs = work[size+rank]; 1269 ierr = PetscFree(work);CHKERRQ(ierr); 1270 1271 /* post receives: */ 1272 ierr = PetscMalloc((nrecvs+1)*(nmax+1)*sizeof(int),&rvalues);CHKERRQ(ierr); 1273 ierr = PetscMalloc((nrecvs+1)*sizeof(MPI_Request),&recv_waits);CHKERRQ(ierr); 1274 for (i=0; i<nrecvs; i++) { 1275 ierr = MPI_Irecv(rvalues+nmax*i,nmax,MPI_INT,MPI_ANY_SOURCE,tag,comm,recv_waits+i);CHKERRQ(ierr); 1276 } 1277 1278 /* do sends: 1279 1) starts[i] gives the starting index in svalues for stuff going to 1280 the ith processor 1281 */ 1282 ierr = PetscMalloc((N+1)*sizeof(int),&svalues);CHKERRQ(ierr); 1283 ierr = PetscMalloc((nsends+1)*sizeof(MPI_Request),&send_waits);CHKERRQ(ierr); 1284 ierr = PetscMalloc((size+1)*sizeof(int),&starts);CHKERRQ(ierr); 1285 starts[0] = 0; 1286 for (i=1; i<size; i++) { starts[i] = starts[i-1] + nprocs[i-1];} 1287 for (i=0; i<N; i++) { 1288 svalues[starts[owner[i]]++] = rows[i]; 1289 } 1290 ierr = ISRestoreIndices(is,&rows);CHKERRQ(ierr); 1291 1292 starts[0] = 0; 1293 for (i=1; i<size+1; i++) { starts[i] = starts[i-1] + nprocs[i-1];} 1294 count = 0; 1295 for (i=0; i<size; i++) { 1296 if (procs[i]) { 1297 ierr = MPI_Isend(svalues+starts[i],nprocs[i],MPI_INT,i,tag,comm,send_waits+count++);CHKERRQ(ierr); 1298 } 1299 } 1300 ierr = PetscFree(starts);CHKERRQ(ierr); 1301 1302 base = owners[rank]*bs; 1303 1304 /* wait on receives */ 1305 ierr = PetscMalloc(2*(nrecvs+1)*sizeof(int),&lens);CHKERRQ(ierr); 1306 source = lens + nrecvs; 1307 count = nrecvs; slen = 0; 1308 while (count) { 1309 ierr = MPI_Waitany(nrecvs,recv_waits,&imdex,&recv_status);CHKERRQ(ierr); 1310 /* unpack receives into our local space */ 1311 ierr = MPI_Get_count(&recv_status,MPI_INT,&n);CHKERRQ(ierr); 1312 source[imdex] = recv_status.MPI_SOURCE; 1313 lens[imdex] = n; 1314 slen += n; 1315 count--; 1316 } 1317 ierr = PetscFree(recv_waits);CHKERRQ(ierr); 1318 1319 /* move the data into the send scatter */ 1320 ierr = PetscMalloc((slen+1)*sizeof(int),&lrows);CHKERRQ(ierr); 1321 count = 0; 1322 for (i=0; i<nrecvs; i++) { 1323 values = rvalues + i*nmax; 1324 for (j=0; j<lens[i]; j++) { 1325 lrows[count++] = values[j] - base; 1326 } 1327 } 1328 ierr = PetscFree(rvalues);CHKERRQ(ierr); 1329 ierr = PetscFree(lens);CHKERRQ(ierr); 1330 ierr = PetscFree(owner);CHKERRQ(ierr); 1331 ierr = PetscFree(nprocs);CHKERRQ(ierr); 1332 1333 /* actually zap the local rows */ 1334 ierr = ISCreateGeneral(PETSC_COMM_SELF,slen,lrows,&istmp);CHKERRQ(ierr); 1335 PetscLogObjectParent(A,istmp); 1336 1337 /* 1338 Zero the required rows. If the "diagonal block" of the matrix 1339 is square and the user wishes to set the diagonal we use seperate 1340 code so that MatSetValues() is not called for each diagonal allocating 1341 new memory, thus calling lots of mallocs and slowing things down. 1342 1343 Contributed by: Mathew Knepley 1344 */ 1345 /* must zero l->B before l->A because the (diag) case below may put values into l->B*/ 1346 ierr = MatZeroRows_SeqBAIJ(l->B,istmp,0);CHKERRQ(ierr); 1347 if (diag && (l->A->M == l->A->N)) { 1348 ierr = MatZeroRows_SeqSBAIJ(l->A,istmp,diag);CHKERRQ(ierr); 1349 } else if (diag) { 1350 ierr = MatZeroRows_SeqSBAIJ(l->A,istmp,0);CHKERRQ(ierr); 1351 if (((Mat_SeqSBAIJ*)l->A->data)->nonew) { 1352 SETERRQ(PETSC_ERR_SUP,"MatZeroRows() on rectangular matrices cannot be used with the Mat options \n\ 1353 MAT_NO_NEW_NONZERO_LOCATIONS,MAT_NEW_NONZERO_LOCATION_ERR,MAT_NEW_NONZERO_ALLOCATION_ERR"); 1354 } 1355 for (i=0; i<slen; i++) { 1356 row = lrows[i] + rstart_bs; 1357 ierr = MatSetValues(A,1,&row,1,&row,diag,INSERT_VALUES);CHKERRQ(ierr); 1358 } 1359 ierr = MatAssemblyBegin(A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1360 ierr = MatAssemblyEnd(A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1361 } else { 1362 ierr = MatZeroRows_SeqSBAIJ(l->A,istmp,0);CHKERRQ(ierr); 1363 } 1364 1365 ierr = ISDestroy(istmp);CHKERRQ(ierr); 1366 ierr = PetscFree(lrows);CHKERRQ(ierr); 1367 1368 /* wait on sends */ 1369 if (nsends) { 1370 ierr = PetscMalloc(nsends*sizeof(MPI_Status),&send_status);CHKERRQ(ierr); 1371 ierr = MPI_Waitall(nsends,send_waits,send_status);CHKERRQ(ierr); 1372 ierr = PetscFree(send_status);CHKERRQ(ierr); 1373 } 1374 ierr = PetscFree(send_waits);CHKERRQ(ierr); 1375 ierr = PetscFree(svalues);CHKERRQ(ierr); 1376 1377 PetscFunctionReturn(0); 1378 } 1379 1380 #undef __FUNC__ 1381 #define __FUNC__ "MatPrintHelp_MPISBAIJ" 1382 int MatPrintHelp_MPISBAIJ(Mat A) 1383 { 1384 Mat_MPISBAIJ *a = (Mat_MPISBAIJ*)A->data; 1385 MPI_Comm comm = A->comm; 1386 static int called = 0; 1387 int ierr; 1388 1389 PetscFunctionBegin; 1390 if (!a->rank) { 1391 ierr = MatPrintHelp_SeqSBAIJ(a->A);CHKERRQ(ierr); 1392 } 1393 if (called) {PetscFunctionReturn(0);} else called = 1; 1394 ierr = (*PetscHelpPrintf)(comm," Options for MATMPISBAIJ matrix format (the defaults):\n");CHKERRQ(ierr); 1395 ierr = (*PetscHelpPrintf)(comm," -mat_use_hash_table <factor>: Use hashtable for efficient matrix assembly\n");CHKERRQ(ierr); 1396 PetscFunctionReturn(0); 1397 } 1398 1399 #undef __FUNC__ 1400 #define __FUNC__ "MatSetUnfactored_MPISBAIJ" 1401 int MatSetUnfactored_MPISBAIJ(Mat A) 1402 { 1403 Mat_MPISBAIJ *a = (Mat_MPISBAIJ*)A->data; 1404 int ierr; 1405 1406 PetscFunctionBegin; 1407 ierr = MatSetUnfactored(a->A);CHKERRQ(ierr); 1408 PetscFunctionReturn(0); 1409 } 1410 1411 static int MatDuplicate_MPISBAIJ(Mat,MatDuplicateOption,Mat *); 1412 1413 #undef __FUNC__ 1414 #define __FUNC__ "MatEqual_MPISBAIJ" 1415 int MatEqual_MPISBAIJ(Mat A,Mat B,PetscTruth *flag) 1416 { 1417 Mat_MPISBAIJ *matB = (Mat_MPISBAIJ*)B->data,*matA = (Mat_MPISBAIJ*)A->data; 1418 Mat a,b,c,d; 1419 PetscTruth flg; 1420 int ierr; 1421 1422 PetscFunctionBegin; 1423 ierr = PetscTypeCompare((PetscObject)B,MATMPISBAIJ,&flg);CHKERRQ(ierr); 1424 if (!flg) SETERRQ(PETSC_ERR_ARG_INCOMP,"Matrices must be same type"); 1425 a = matA->A; b = matA->B; 1426 c = matB->A; d = matB->B; 1427 1428 ierr = MatEqual(a,c,&flg);CHKERRQ(ierr); 1429 if (flg == PETSC_TRUE) { 1430 ierr = MatEqual(b,d,&flg);CHKERRQ(ierr); 1431 } 1432 ierr = MPI_Allreduce(&flg,flag,1,MPI_INT,MPI_LAND,A->comm);CHKERRQ(ierr); 1433 PetscFunctionReturn(0); 1434 } 1435 1436 #undef __FUNC__ 1437 #define __FUNC__ "MatSetUpPreallocation_MPISBAIJ" 1438 int MatSetUpPreallocation_MPISBAIJ(Mat A) 1439 { 1440 int ierr; 1441 1442 PetscFunctionBegin; 1443 ierr = MatMPISBAIJSetPreallocation(A,1,PETSC_DEFAULT,0,PETSC_DEFAULT,0);CHKERRQ(ierr); 1444 PetscFunctionReturn(0); 1445 } 1446 /* -------------------------------------------------------------------*/ 1447 static struct _MatOps MatOps_Values = { 1448 MatSetValues_MPISBAIJ, 1449 MatGetRow_MPISBAIJ, 1450 MatRestoreRow_MPISBAIJ, 1451 MatMult_MPISBAIJ, 1452 MatMultAdd_MPISBAIJ, 1453 MatMultTranspose_MPISBAIJ, 1454 MatMultTransposeAdd_MPISBAIJ, 1455 0, 1456 0, 1457 0, 1458 0, 1459 0, 1460 0, 1461 0, 1462 MatTranspose_MPISBAIJ, 1463 MatGetInfo_MPISBAIJ, 1464 MatEqual_MPISBAIJ, 1465 MatGetDiagonal_MPISBAIJ, 1466 MatDiagonalScale_MPISBAIJ, 1467 MatNorm_MPISBAIJ, 1468 MatAssemblyBegin_MPISBAIJ, 1469 MatAssemblyEnd_MPISBAIJ, 1470 0, 1471 MatSetOption_MPISBAIJ, 1472 MatZeroEntries_MPISBAIJ, 1473 MatZeroRows_MPISBAIJ, 1474 0, 1475 0, 1476 0, 1477 0, 1478 MatSetUpPreallocation_MPISBAIJ, 1479 0, 1480 MatGetOwnershipRange_MPISBAIJ, 1481 0, 1482 0, 1483 0, 1484 0, 1485 MatDuplicate_MPISBAIJ, 1486 0, 1487 0, 1488 0, 1489 0, 1490 0, 1491 MatGetSubMatrices_MPISBAIJ, 1492 MatIncreaseOverlap_MPISBAIJ, 1493 MatGetValues_MPISBAIJ, 1494 0, 1495 MatPrintHelp_MPISBAIJ, 1496 MatScale_MPISBAIJ, 1497 0, 1498 0, 1499 0, 1500 MatGetBlockSize_MPISBAIJ, 1501 0, 1502 0, 1503 0, 1504 0, 1505 0, 1506 0, 1507 MatSetUnfactored_MPISBAIJ, 1508 0, 1509 MatSetValuesBlocked_MPISBAIJ, 1510 0, 1511 0, 1512 0, 1513 MatGetMaps_Petsc, 1514 0, 1515 0, 1516 0, 1517 0, 1518 0, 1519 0, 1520 MatGetRowMax_MPISBAIJ}; 1521 1522 1523 EXTERN_C_BEGIN 1524 #undef __FUNC__ 1525 #define __FUNC__ "MatGetDiagonalBlock_MPISBAIJ" 1526 int MatGetDiagonalBlock_MPISBAIJ(Mat A,PetscTruth *iscopy,MatReuse reuse,Mat *a) 1527 { 1528 PetscFunctionBegin; 1529 *a = ((Mat_MPISBAIJ *)A->data)->A; 1530 *iscopy = PETSC_FALSE; 1531 PetscFunctionReturn(0); 1532 } 1533 EXTERN_C_END 1534 1535 EXTERN_C_BEGIN 1536 #undef __FUNC__ 1537 #define __FUNC__ "MatCreate_MPISBAIJ" 1538 int MatCreate_MPISBAIJ(Mat B) 1539 { 1540 Mat_MPISBAIJ *b; 1541 int ierr; 1542 PetscTruth flg; 1543 1544 PetscFunctionBegin; 1545 1546 ierr = PetscNew(Mat_MPISBAIJ,&b);CHKERRQ(ierr); 1547 B->data = (void*)b; 1548 ierr = PetscMemzero(b,sizeof(Mat_MPISBAIJ));CHKERRQ(ierr); 1549 ierr = PetscMemcpy(B->ops,&MatOps_Values,sizeof(struct _MatOps));CHKERRQ(ierr); 1550 1551 B->ops->destroy = MatDestroy_MPISBAIJ; 1552 B->ops->view = MatView_MPISBAIJ; 1553 B->mapping = 0; 1554 B->factor = 0; 1555 B->assembled = PETSC_FALSE; 1556 1557 B->insertmode = NOT_SET_VALUES; 1558 ierr = MPI_Comm_rank(B->comm,&b->rank);CHKERRQ(ierr); 1559 ierr = MPI_Comm_size(B->comm,&b->size);CHKERRQ(ierr); 1560 1561 /* build local table of row and column ownerships */ 1562 ierr = PetscMalloc(3*(b->size+2)*sizeof(int),&b->rowners);CHKERRQ(ierr); 1563 b->cowners = b->rowners + b->size + 2; 1564 b->rowners_bs = b->cowners + b->size + 2; 1565 PetscLogObjectMemory(B,3*(b->size+2)*sizeof(int)+sizeof(struct _p_Mat)+sizeof(Mat_MPISBAIJ)); 1566 1567 /* build cache for off array entries formed */ 1568 ierr = MatStashCreate_Private(B->comm,1,&B->stash);CHKERRQ(ierr); 1569 b->donotstash = PETSC_FALSE; 1570 b->colmap = PETSC_NULL; 1571 b->garray = PETSC_NULL; 1572 b->roworiented = PETSC_TRUE; 1573 1574 #if defined(PEYSC_USE_MAT_SINGLE) 1575 /* stuff for MatSetValues_XXX in single precision */ 1576 b->lensetvalues = 0; 1577 b->setvaluescopy = PETSC_NULL; 1578 #endif 1579 1580 /* stuff used in block assembly */ 1581 b->barray = 0; 1582 1583 /* stuff used for matrix vector multiply */ 1584 b->lvec = 0; 1585 b->Mvctx = 0; 1586 1587 /* stuff for MatGetRow() */ 1588 b->rowindices = 0; 1589 b->rowvalues = 0; 1590 b->getrowactive = PETSC_FALSE; 1591 1592 /* hash table stuff */ 1593 b->ht = 0; 1594 b->hd = 0; 1595 b->ht_size = 0; 1596 b->ht_flag = PETSC_FALSE; 1597 b->ht_fact = 0; 1598 b->ht_total_ct = 0; 1599 b->ht_insert_ct = 0; 1600 1601 ierr = PetscOptionsHasName(PETSC_NULL,"-mat_use_hash_table",&flg);CHKERRQ(ierr); 1602 if (flg) { 1603 double fact = 1.39; 1604 ierr = MatSetOption(B,MAT_USE_HASH_TABLE);CHKERRQ(ierr); 1605 ierr = PetscOptionsGetDouble(PETSC_NULL,"-mat_use_hash_table",&fact,PETSC_NULL);CHKERRQ(ierr); 1606 if (fact <= 1.0) fact = 1.39; 1607 ierr = MatMPIBAIJSetHashTableFactor(B,fact);CHKERRQ(ierr); 1608 PetscLogInfo(0,"MatCreateMPISBAIJ:Hash table Factor used %5.2f\n",fact); 1609 } 1610 ierr = PetscObjectComposeFunctionDynamic((PetscObject)B,"MatStoreValues_C", 1611 "MatStoreValues_MPISBAIJ", 1612 MatStoreValues_MPISBAIJ);CHKERRQ(ierr); 1613 ierr = PetscObjectComposeFunctionDynamic((PetscObject)B,"MatRetrieveValues_C", 1614 "MatRetrieveValues_MPISBAIJ", 1615 MatRetrieveValues_MPISBAIJ);CHKERRQ(ierr); 1616 ierr = PetscObjectComposeFunctionDynamic((PetscObject)B,"MatGetDiagonalBlock_C", 1617 "MatGetDiagonalBlock_MPISBAIJ", 1618 MatGetDiagonalBlock_MPISBAIJ);CHKERRQ(ierr); 1619 PetscFunctionReturn(0); 1620 } 1621 EXTERN_C_END 1622 1623 #undef __FUNC__ 1624 #define __FUNC__ "MatMPISBAIJSetPreallocation" 1625 /*@C 1626 MatMPISBAIJSetPreallocation - For good matrix assembly performance 1627 the user should preallocate the matrix storage by setting the parameters 1628 d_nz (or d_nnz) and o_nz (or o_nnz). By setting these parameters accurately, 1629 performance can be increased by more than a factor of 50. 1630 1631 Collective on Mat 1632 1633 Input Parameters: 1634 + A - the matrix 1635 . bs - size of blockk 1636 . d_nz - number of block nonzeros per block row in diagonal portion of local 1637 submatrix (same for all local rows) 1638 . d_nnz - array containing the number of block nonzeros in the various block rows 1639 of the in diagonal portion of the local (possibly different for each block 1640 row) or PETSC_NULL. You must leave room for the diagonal entry even if it is zero. 1641 . o_nz - number of block nonzeros per block row in the off-diagonal portion of local 1642 submatrix (same for all local rows). 1643 - o_nnz - array containing the number of nonzeros in the various block rows of the 1644 off-diagonal portion of the local submatrix (possibly different for 1645 each block row) or PETSC_NULL. 1646 1647 1648 Options Database Keys: 1649 . -mat_no_unroll - uses code that does not unroll the loops in the 1650 block calculations (much slower) 1651 . -mat_block_size - size of the blocks to use 1652 1653 Notes: 1654 1655 If PETSC_DECIDE or PETSC_DETERMINE is used for a particular argument on one processor 1656 than it must be used on all processors that share the object for that argument. 1657 1658 Storage Information: 1659 For a square global matrix we define each processor's diagonal portion 1660 to be its local rows and the corresponding columns (a square submatrix); 1661 each processor's off-diagonal portion encompasses the remainder of the 1662 local matrix (a rectangular submatrix). 1663 1664 The user can specify preallocated storage for the diagonal part of 1665 the local submatrix with either d_nz or d_nnz (not both). Set 1666 d_nz=PETSC_DEFAULT and d_nnz=PETSC_NULL for PETSc to control dynamic 1667 memory allocation. Likewise, specify preallocated storage for the 1668 off-diagonal part of the local submatrix with o_nz or o_nnz (not both). 1669 1670 Consider a processor that owns rows 3, 4 and 5 of a parallel matrix. In 1671 the figure below we depict these three local rows and all columns (0-11). 1672 1673 .vb 1674 0 1 2 3 4 5 6 7 8 9 10 11 1675 ------------------- 1676 row 3 | o o o d d d o o o o o o 1677 row 4 | o o o d d d o o o o o o 1678 row 5 | o o o d d d o o o o o o 1679 ------------------- 1680 .ve 1681 1682 Thus, any entries in the d locations are stored in the d (diagonal) 1683 submatrix, and any entries in the o locations are stored in the 1684 o (off-diagonal) submatrix. Note that the d and the o submatrices are 1685 stored simply in the MATSEQBAIJ format for compressed row storage. 1686 1687 Now d_nz should indicate the number of block nonzeros per row in the d matrix, 1688 and o_nz should indicate the number of block nonzeros per row in the o matrix. 1689 In general, for PDE problems in which most nonzeros are near the diagonal, 1690 one expects d_nz >> o_nz. For large problems you MUST preallocate memory 1691 or you will get TERRIBLE performance; see the users' manual chapter on 1692 matrices. 1693 1694 Level: intermediate 1695 1696 .keywords: matrix, block, aij, compressed row, sparse, parallel 1697 1698 .seealso: MatCreate(), MatCreateSeqSBAIJ(), MatSetValues(), MatCreateMPIBAIJ() 1699 @*/ 1700 1701 int MatMPISBAIJSetPreallocation(Mat B,int bs,int d_nz,int *d_nnz,int o_nz,int *o_nnz) 1702 { 1703 Mat_MPISBAIJ *b; 1704 int ierr,i,mbs,Mbs=PETSC_DECIDE; 1705 PetscTruth flg2; 1706 1707 PetscFunctionBegin; 1708 ierr = PetscTypeCompare((PetscObject)B,MATMPISBAIJ,&flg2);CHKERRQ(ierr); 1709 if (!flg2) PetscFunctionReturn(0); 1710 1711 ierr = PetscOptionsGetInt(PETSC_NULL,"-mat_block_size",&bs,PETSC_NULL);CHKERRQ(ierr); 1712 1713 if (bs < 1) SETERRQ(PETSC_ERR_ARG_OUTOFRANGE,"Invalid block size specified, must be positive"); 1714 if (d_nz < -2) SETERRQ1(PETSC_ERR_ARG_OUTOFRANGE,"d_nz cannot be less than -2: value %d",d_nz); 1715 if (o_nz < -2) SETERRQ1(PETSC_ERR_ARG_OUTOFRANGE,"o_nz cannot be less than -2: value %d",o_nz); 1716 if (d_nnz) { 1717 for (i=0; i<B->m/bs; i++) { 1718 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]); 1719 } 1720 } 1721 if (o_nnz) { 1722 for (i=0; i<B->m/bs; i++) { 1723 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]); 1724 } 1725 } 1726 B->preallocated = PETSC_TRUE; 1727 ierr = PetscSplitOwnershipBlock(B->comm,bs,&B->m,&B->M);CHKERRQ(ierr); 1728 ierr = PetscSplitOwnershipBlock(B->comm,bs,&B->n,&B->N);CHKERRQ(ierr); 1729 ierr = MapCreateMPI(B->comm,B->m,B->M,&B->rmap);CHKERRQ(ierr); 1730 ierr = MapCreateMPI(B->comm,B->m,B->M,&B->cmap);CHKERRQ(ierr); 1731 1732 b = (Mat_MPISBAIJ*)B->data; 1733 mbs = B->m/bs; 1734 Mbs = B->M/bs; 1735 if (mbs*bs != B->m) { 1736 SETERRQ2(PETSC_ERR_ARG_SIZ,"No of local rows %d must be divisible by blocksize %d",B->m,bs); 1737 } 1738 1739 b->bs = bs; 1740 b->bs2 = bs*bs; 1741 b->mbs = mbs; 1742 b->nbs = mbs; 1743 b->Mbs = Mbs; 1744 b->Nbs = Mbs; 1745 1746 ierr = MPI_Allgather(&b->mbs,1,MPI_INT,b->rowners+1,1,MPI_INT,B->comm);CHKERRQ(ierr); 1747 b->rowners[0] = 0; 1748 for (i=2; i<=b->size; i++) { 1749 b->rowners[i] += b->rowners[i-1]; 1750 } 1751 b->rstart = b->rowners[b->rank]; 1752 b->rend = b->rowners[b->rank+1]; 1753 b->cstart = b->rstart; 1754 b->cend = b->rend; 1755 for (i=0; i<=b->size; i++) { 1756 b->rowners_bs[i] = b->rowners[i]*bs; 1757 } 1758 b->rstart_bs = b-> rstart*bs; 1759 b->rend_bs = b->rend*bs; 1760 1761 b->cstart_bs = b->cstart*bs; 1762 b->cend_bs = b->cend*bs; 1763 1764 1765 if (d_nz == PETSC_DEFAULT) d_nz = 5; 1766 ierr = MatCreateSeqSBAIJ(PETSC_COMM_SELF,bs,B->m,B->m,d_nz,d_nnz,&b->A);CHKERRQ(ierr); 1767 PetscLogObjectParent(B,b->A); 1768 if (o_nz == PETSC_DEFAULT) o_nz = 0; 1769 ierr = MatCreateSeqBAIJ(PETSC_COMM_SELF,bs,B->m,B->M,o_nz,o_nnz,&b->B);CHKERRQ(ierr); 1770 PetscLogObjectParent(B,b->B); 1771 1772 /* build cache for off array entries formed */ 1773 ierr = MatStashCreate_Private(B->comm,bs,&B->bstash);CHKERRQ(ierr); 1774 1775 PetscFunctionReturn(0); 1776 } 1777 1778 #undef __FUNC__ 1779 #define __FUNC__ "MatCreateMPISBAIJ" 1780 /*@C 1781 MatCreateMPISBAIJ - Creates a sparse parallel matrix in symmetric block AIJ format 1782 (block compressed row). For good matrix assembly performance 1783 the user should preallocate the matrix storage by setting the parameters 1784 d_nz (or d_nnz) and o_nz (or o_nnz). By setting these parameters accurately, 1785 performance can be increased by more than a factor of 50. 1786 1787 Collective on MPI_Comm 1788 1789 Input Parameters: 1790 + comm - MPI communicator 1791 . bs - size of blockk 1792 . m - number of local rows (or PETSC_DECIDE to have calculated if M is given) 1793 This value should be the same as the local size used in creating the 1794 y vector for the matrix-vector product y = Ax. 1795 . n - number of local columns (or PETSC_DECIDE to have calculated if N is given) 1796 This value should be the same as the local size used in creating the 1797 x vector for the matrix-vector product y = Ax. 1798 . M - number of global rows (or PETSC_DETERMINE to have calculated if m is given) 1799 . N - number of global columns (or PETSC_DETERMINE to have calculated if n is given) 1800 . d_nz - number of block nonzeros per block row in diagonal portion of local 1801 submatrix (same for all local rows) 1802 . d_nnz - array containing the number of block nonzeros in the various block rows 1803 of the in diagonal portion of the local (possibly different for each block 1804 row) or PETSC_NULL. You must leave room for the diagonal entry even if it is zero. 1805 . o_nz - number of block nonzeros per block row in the off-diagonal portion of local 1806 submatrix (same for all local rows). 1807 - o_nnz - array containing the number of nonzeros in the various block rows of the 1808 off-diagonal portion of the local submatrix (possibly different for 1809 each block row) or PETSC_NULL. 1810 1811 Output Parameter: 1812 . A - the matrix 1813 1814 Options Database Keys: 1815 . -mat_no_unroll - uses code that does not unroll the loops in the 1816 block calculations (much slower) 1817 . -mat_block_size - size of the blocks to use 1818 . -mat_mpi - use the parallel matrix data structures even on one processor 1819 (defaults to using SeqBAIJ format on one processor) 1820 1821 Notes: 1822 The user MUST specify either the local or global matrix dimensions 1823 (possibly both). 1824 1825 If PETSC_DECIDE or PETSC_DETERMINE is used for a particular argument on one processor 1826 than it must be used on all processors that share the object for that argument. 1827 1828 Storage Information: 1829 For a square global matrix we define each processor's diagonal portion 1830 to be its local rows and the corresponding columns (a square submatrix); 1831 each processor's off-diagonal portion encompasses the remainder of the 1832 local matrix (a rectangular submatrix). 1833 1834 The user can specify preallocated storage for the diagonal part of 1835 the local submatrix with either d_nz or d_nnz (not both). Set 1836 d_nz=PETSC_DEFAULT and d_nnz=PETSC_NULL for PETSc to control dynamic 1837 memory allocation. Likewise, specify preallocated storage for the 1838 off-diagonal part of the local submatrix with o_nz or o_nnz (not both). 1839 1840 Consider a processor that owns rows 3, 4 and 5 of a parallel matrix. In 1841 the figure below we depict these three local rows and all columns (0-11). 1842 1843 .vb 1844 0 1 2 3 4 5 6 7 8 9 10 11 1845 ------------------- 1846 row 3 | o o o d d d o o o o o o 1847 row 4 | o o o d d d o o o o o o 1848 row 5 | o o o d d d o o o o o o 1849 ------------------- 1850 .ve 1851 1852 Thus, any entries in the d locations are stored in the d (diagonal) 1853 submatrix, and any entries in the o locations are stored in the 1854 o (off-diagonal) submatrix. Note that the d and the o submatrices are 1855 stored simply in the MATSEQBAIJ format for compressed row storage. 1856 1857 Now d_nz should indicate the number of block nonzeros per row in the d matrix, 1858 and o_nz should indicate the number of block nonzeros per row in the o matrix. 1859 In general, for PDE problems in which most nonzeros are near the diagonal, 1860 one expects d_nz >> o_nz. For large problems you MUST preallocate memory 1861 or you will get TERRIBLE performance; see the users' manual chapter on 1862 matrices. 1863 1864 Level: intermediate 1865 1866 .keywords: matrix, block, aij, compressed row, sparse, parallel 1867 1868 .seealso: MatCreate(), MatCreateSeqSBAIJ(), MatSetValues(), MatCreateMPIBAIJ() 1869 @*/ 1870 1871 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) 1872 { 1873 int ierr,size; 1874 1875 PetscFunctionBegin; 1876 ierr = MatCreate(comm,m,n,M,N,A);CHKERRQ(ierr); 1877 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 1878 if (size > 1) { 1879 ierr = MatSetType(*A,MATMPISBAIJ);CHKERRQ(ierr); 1880 ierr = MatMPISBAIJSetPreallocation(*A,bs,d_nz,d_nnz,o_nz,o_nnz);CHKERRQ(ierr); 1881 } else { 1882 ierr = MatSetType(*A,MATSEQSBAIJ);CHKERRQ(ierr); 1883 ierr = MatSeqSBAIJSetPreallocation(*A,bs,d_nz,d_nnz);CHKERRQ(ierr); 1884 } 1885 PetscFunctionReturn(0); 1886 } 1887 1888 1889 #undef __FUNC__ 1890 #define __FUNC__ "MatDuplicate_MPISBAIJ" 1891 static int MatDuplicate_MPISBAIJ(Mat matin,MatDuplicateOption cpvalues,Mat *newmat) 1892 { 1893 Mat mat; 1894 Mat_MPISBAIJ *a,*oldmat = (Mat_MPISBAIJ*)matin->data; 1895 int ierr,len=0; 1896 1897 PetscFunctionBegin; 1898 *newmat = 0; 1899 ierr = MatCreate(matin->comm,matin->m,matin->n,matin->M,matin->N,&mat);CHKERRQ(ierr); 1900 ierr = MatSetType(mat,MATMPISBAIJ);CHKERRQ(ierr); 1901 mat->preallocated = PETSC_TRUE; 1902 a = (Mat_MPISBAIJ*)mat->data; 1903 a->bs = oldmat->bs; 1904 a->bs2 = oldmat->bs2; 1905 a->mbs = oldmat->mbs; 1906 a->nbs = oldmat->nbs; 1907 a->Mbs = oldmat->Mbs; 1908 a->Nbs = oldmat->Nbs; 1909 1910 a->rstart = oldmat->rstart; 1911 a->rend = oldmat->rend; 1912 a->cstart = oldmat->cstart; 1913 a->cend = oldmat->cend; 1914 a->size = oldmat->size; 1915 a->rank = oldmat->rank; 1916 a->donotstash = oldmat->donotstash; 1917 a->roworiented = oldmat->roworiented; 1918 a->rowindices = 0; 1919 a->rowvalues = 0; 1920 a->getrowactive = PETSC_FALSE; 1921 a->barray = 0; 1922 a->rstart_bs = oldmat->rstart_bs; 1923 a->rend_bs = oldmat->rend_bs; 1924 a->cstart_bs = oldmat->cstart_bs; 1925 a->cend_bs = oldmat->cend_bs; 1926 1927 /* hash table stuff */ 1928 a->ht = 0; 1929 a->hd = 0; 1930 a->ht_size = 0; 1931 a->ht_flag = oldmat->ht_flag; 1932 a->ht_fact = oldmat->ht_fact; 1933 a->ht_total_ct = 0; 1934 a->ht_insert_ct = 0; 1935 1936 ierr = PetscMalloc(3*(a->size+2)*sizeof(int),&a->rowners);CHKERRQ(ierr); 1937 PetscLogObjectMemory(mat,3*(a->size+2)*sizeof(int)+sizeof(struct _p_Mat)+sizeof(Mat_MPISBAIJ)); 1938 a->cowners = a->rowners + a->size + 2; 1939 a->rowners_bs = a->cowners + a->size + 2; 1940 ierr = PetscMemcpy(a->rowners,oldmat->rowners,3*(a->size+2)*sizeof(int));CHKERRQ(ierr); 1941 ierr = MatStashCreate_Private(matin->comm,1,&mat->stash);CHKERRQ(ierr); 1942 ierr = MatStashCreate_Private(matin->comm,oldmat->bs,&mat->bstash);CHKERRQ(ierr); 1943 if (oldmat->colmap) { 1944 #if defined (PETSC_USE_CTABLE) 1945 ierr = PetscTableCreateCopy(oldmat->colmap,&a->colmap);CHKERRQ(ierr); 1946 #else 1947 ierr = PetscMalloc((a->Nbs)*sizeof(int),&a->colmap);CHKERRQ(ierr); 1948 PetscLogObjectMemory(mat,(a->Nbs)*sizeof(int)); 1949 ierr = PetscMemcpy(a->colmap,oldmat->colmap,(a->Nbs)*sizeof(int));CHKERRQ(ierr); 1950 #endif 1951 } else a->colmap = 0; 1952 if (oldmat->garray && (len = ((Mat_SeqBAIJ*)(oldmat->B->data))->nbs)) { 1953 ierr = PetscMalloc(len*sizeof(int),&a->garray);CHKERRQ(ierr); 1954 PetscLogObjectMemory(mat,len*sizeof(int)); 1955 ierr = PetscMemcpy(a->garray,oldmat->garray,len*sizeof(int));CHKERRQ(ierr); 1956 } else a->garray = 0; 1957 1958 ierr = VecDuplicate(oldmat->lvec,&a->lvec);CHKERRQ(ierr); 1959 PetscLogObjectParent(mat,a->lvec); 1960 ierr = VecScatterCopy(oldmat->Mvctx,&a->Mvctx);CHKERRQ(ierr); 1961 1962 PetscLogObjectParent(mat,a->Mvctx); 1963 ierr = MatDuplicate(oldmat->A,cpvalues,&a->A);CHKERRQ(ierr); 1964 PetscLogObjectParent(mat,a->A); 1965 ierr = MatDuplicate(oldmat->B,cpvalues,&a->B);CHKERRQ(ierr); 1966 PetscLogObjectParent(mat,a->B); 1967 ierr = PetscFListDuplicate(mat->qlist,&matin->qlist);CHKERRQ(ierr); 1968 *newmat = mat; 1969 PetscFunctionReturn(0); 1970 } 1971 1972 #include "petscsys.h" 1973 1974 EXTERN_C_BEGIN 1975 #undef __FUNC__ 1976 #define __FUNC__ "MatLoad_MPISBAIJ" 1977 int MatLoad_MPISBAIJ(PetscViewer viewer,MatType type,Mat *newmat) 1978 { 1979 Mat A; 1980 int i,nz,ierr,j,rstart,rend,fd; 1981 Scalar *vals,*buf; 1982 MPI_Comm comm = ((PetscObject)viewer)->comm; 1983 MPI_Status status; 1984 int header[4],rank,size,*rowlengths = 0,M,N,m,*rowners,*browners,maxnz,*cols; 1985 int *locrowlens,*sndcounts = 0,*procsnz = 0,jj,*mycols,*ibuf; 1986 int tag = ((PetscObject)viewer)->tag,bs=1,Mbs,mbs,extra_rows; 1987 int *dlens,*odlens,*mask,*masked1,*masked2,rowcount,odcount; 1988 int dcount,kmax,k,nzcount,tmp; 1989 1990 PetscFunctionBegin; 1991 ierr = PetscOptionsGetInt(PETSC_NULL,"-matload_block_size",&bs,PETSC_NULL);CHKERRQ(ierr); 1992 1993 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 1994 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 1995 if (!rank) { 1996 ierr = PetscViewerBinaryGetDescriptor(viewer,&fd);CHKERRQ(ierr); 1997 ierr = PetscBinaryRead(fd,(char *)header,4,PETSC_INT);CHKERRQ(ierr); 1998 if (header[0] != MAT_COOKIE) SETERRQ(PETSC_ERR_FILE_UNEXPECTED,"not matrix object"); 1999 if (header[3] < 0) { 2000 SETERRQ(PETSC_ERR_FILE_UNEXPECTED,"Matrix stored in special format, cannot load as MPISBAIJ"); 2001 } 2002 } 2003 2004 ierr = MPI_Bcast(header+1,3,MPI_INT,0,comm);CHKERRQ(ierr); 2005 M = header[1]; N = header[2]; 2006 2007 if (M != N) SETERRQ(PETSC_ERR_SUP,"Can only do square matrices"); 2008 2009 /* 2010 This code adds extra rows to make sure the number of rows is 2011 divisible by the blocksize 2012 */ 2013 Mbs = M/bs; 2014 extra_rows = bs - M + bs*(Mbs); 2015 if (extra_rows == bs) extra_rows = 0; 2016 else Mbs++; 2017 if (extra_rows &&!rank) { 2018 PetscLogInfo(0,"MatLoad_MPISBAIJ:Padding loaded matrix to match blocksize\n"); 2019 } 2020 2021 /* determine ownership of all rows */ 2022 mbs = Mbs/size + ((Mbs % size) > rank); 2023 m = mbs*bs; 2024 ierr = PetscMalloc(2*(size+2)*sizeof(int),&rowners);CHKERRQ(ierr); 2025 browners = rowners + size + 1; 2026 ierr = MPI_Allgather(&mbs,1,MPI_INT,rowners+1,1,MPI_INT,comm);CHKERRQ(ierr); 2027 rowners[0] = 0; 2028 for (i=2; i<=size; i++) rowners[i] += rowners[i-1]; 2029 for (i=0; i<=size; i++) browners[i] = rowners[i]*bs; 2030 rstart = rowners[rank]; 2031 rend = rowners[rank+1]; 2032 2033 /* distribute row lengths to all processors */ 2034 ierr = PetscMalloc((rend-rstart)*bs*sizeof(int),&locrowlens);CHKERRQ(ierr); 2035 if (!rank) { 2036 ierr = PetscMalloc((M+extra_rows)*sizeof(int),&rowlengths);CHKERRQ(ierr); 2037 ierr = PetscBinaryRead(fd,rowlengths,M,PETSC_INT);CHKERRQ(ierr); 2038 for (i=0; i<extra_rows; i++) rowlengths[M+i] = 1; 2039 ierr = PetscMalloc(size*sizeof(int),&sndcounts);CHKERRQ(ierr); 2040 for (i=0; i<size; i++) sndcounts[i] = browners[i+1] - browners[i]; 2041 ierr = MPI_Scatterv(rowlengths,sndcounts,browners,MPI_INT,locrowlens,(rend-rstart)*bs,MPI_INT,0,comm);CHKERRQ(ierr); 2042 ierr = PetscFree(sndcounts);CHKERRQ(ierr); 2043 } else { 2044 ierr = MPI_Scatterv(0,0,0,MPI_INT,locrowlens,(rend-rstart)*bs,MPI_INT,0,comm);CHKERRQ(ierr); 2045 } 2046 2047 if (!rank) { /* procs[0] */ 2048 /* calculate the number of nonzeros on each processor */ 2049 ierr = PetscMalloc(size*sizeof(int),&procsnz);CHKERRQ(ierr); 2050 ierr = PetscMemzero(procsnz,size*sizeof(int));CHKERRQ(ierr); 2051 for (i=0; i<size; i++) { 2052 for (j=rowners[i]*bs; j< rowners[i+1]*bs; j++) { 2053 procsnz[i] += rowlengths[j]; 2054 } 2055 } 2056 ierr = PetscFree(rowlengths);CHKERRQ(ierr); 2057 2058 /* determine max buffer needed and allocate it */ 2059 maxnz = 0; 2060 for (i=0; i<size; i++) { 2061 maxnz = PetscMax(maxnz,procsnz[i]); 2062 } 2063 ierr = PetscMalloc(maxnz*sizeof(int),&cols);CHKERRQ(ierr); 2064 2065 /* read in my part of the matrix column indices */ 2066 nz = procsnz[0]; 2067 ierr = PetscMalloc(nz*sizeof(int),&ibuf);CHKERRQ(ierr); 2068 mycols = ibuf; 2069 if (size == 1) nz -= extra_rows; 2070 ierr = PetscBinaryRead(fd,mycols,nz,PETSC_INT);CHKERRQ(ierr); 2071 if (size == 1) for (i=0; i< extra_rows; i++) { mycols[nz+i] = M+i; } 2072 2073 /* read in every ones (except the last) and ship off */ 2074 for (i=1; i<size-1; i++) { 2075 nz = procsnz[i]; 2076 ierr = PetscBinaryRead(fd,cols,nz,PETSC_INT);CHKERRQ(ierr); 2077 ierr = MPI_Send(cols,nz,MPI_INT,i,tag,comm);CHKERRQ(ierr); 2078 } 2079 /* read in the stuff for the last proc */ 2080 if (size != 1) { 2081 nz = procsnz[size-1] - extra_rows; /* the extra rows are not on the disk */ 2082 ierr = PetscBinaryRead(fd,cols,nz,PETSC_INT);CHKERRQ(ierr); 2083 for (i=0; i<extra_rows; i++) cols[nz+i] = M+i; 2084 ierr = MPI_Send(cols,nz+extra_rows,MPI_INT,size-1,tag,comm);CHKERRQ(ierr); 2085 } 2086 ierr = PetscFree(cols);CHKERRQ(ierr); 2087 } else { /* procs[i], i>0 */ 2088 /* determine buffer space needed for message */ 2089 nz = 0; 2090 for (i=0; i<m; i++) { 2091 nz += locrowlens[i]; 2092 } 2093 ierr = PetscMalloc(nz*sizeof(int),&ibuf);CHKERRQ(ierr); 2094 mycols = ibuf; 2095 /* receive message of column indices*/ 2096 ierr = MPI_Recv(mycols,nz,MPI_INT,0,tag,comm,&status);CHKERRQ(ierr); 2097 ierr = MPI_Get_count(&status,MPI_INT,&maxnz);CHKERRQ(ierr); 2098 if (maxnz != nz) SETERRQ(PETSC_ERR_FILE_UNEXPECTED,"something is wrong with file"); 2099 } 2100 2101 /* loop over local rows, determining number of off diagonal entries */ 2102 ierr = PetscMalloc(2*(rend-rstart+1)*sizeof(int),&dlens);CHKERRQ(ierr); 2103 odlens = dlens + (rend-rstart); 2104 ierr = PetscMalloc(3*Mbs*sizeof(int),&mask);CHKERRQ(ierr); 2105 ierr = PetscMemzero(mask,3*Mbs*sizeof(int));CHKERRQ(ierr); 2106 masked1 = mask + Mbs; 2107 masked2 = masked1 + Mbs; 2108 rowcount = 0; nzcount = 0; 2109 for (i=0; i<mbs; i++) { 2110 dcount = 0; 2111 odcount = 0; 2112 for (j=0; j<bs; j++) { 2113 kmax = locrowlens[rowcount]; 2114 for (k=0; k<kmax; k++) { 2115 tmp = mycols[nzcount++]/bs; /* block col. index */ 2116 if (!mask[tmp]) { 2117 mask[tmp] = 1; 2118 if (tmp < rstart || tmp >= rend) masked2[odcount++] = tmp; /* entry in off-diag portion */ 2119 else masked1[dcount++] = tmp; /* entry in diag portion */ 2120 } 2121 } 2122 rowcount++; 2123 } 2124 2125 dlens[i] = dcount; /* d_nzz[i] */ 2126 odlens[i] = odcount; /* o_nzz[i] */ 2127 2128 /* zero out the mask elements we set */ 2129 for (j=0; j<dcount; j++) mask[masked1[j]] = 0; 2130 for (j=0; j<odcount; j++) mask[masked2[j]] = 0; 2131 } 2132 2133 /* create our matrix */ 2134 ierr = MatCreateMPISBAIJ(comm,bs,m,m,PETSC_DETERMINE,PETSC_DETERMINE,0,dlens,0,odlens,newmat); 2135 CHKERRQ(ierr); 2136 A = *newmat; 2137 ierr = MatSetOption(A,MAT_COLUMNS_SORTED);CHKERRQ(ierr); 2138 2139 if (!rank) { 2140 ierr = PetscMalloc(maxnz*sizeof(Scalar),&buf);CHKERRQ(ierr); 2141 /* read in my part of the matrix numerical values */ 2142 nz = procsnz[0]; 2143 vals = buf; 2144 mycols = ibuf; 2145 if (size == 1) nz -= extra_rows; 2146 ierr = PetscBinaryRead(fd,vals,nz,PETSC_SCALAR);CHKERRQ(ierr); 2147 if (size == 1) for (i=0; i< extra_rows; i++) { vals[nz+i] = 1.0; } 2148 2149 /* insert into matrix */ 2150 jj = rstart*bs; 2151 for (i=0; i<m; i++) { 2152 ierr = MatSetValues(A,1,&jj,locrowlens[i],mycols,vals,INSERT_VALUES);CHKERRQ(ierr); 2153 mycols += locrowlens[i]; 2154 vals += locrowlens[i]; 2155 jj++; 2156 } 2157 2158 /* read in other processors (except the last one) and ship out */ 2159 for (i=1; i<size-1; i++) { 2160 nz = procsnz[i]; 2161 vals = buf; 2162 ierr = PetscBinaryRead(fd,vals,nz,PETSC_SCALAR);CHKERRQ(ierr); 2163 ierr = MPI_Send(vals,nz,MPIU_SCALAR,i,A->tag,comm);CHKERRQ(ierr); 2164 } 2165 /* the last proc */ 2166 if (size != 1){ 2167 nz = procsnz[i] - extra_rows; 2168 vals = buf; 2169 ierr = PetscBinaryRead(fd,vals,nz,PETSC_SCALAR);CHKERRQ(ierr); 2170 for (i=0; i<extra_rows; i++) vals[nz+i] = 1.0; 2171 ierr = MPI_Send(vals,nz+extra_rows,MPIU_SCALAR,size-1,A->tag,comm);CHKERRQ(ierr); 2172 } 2173 ierr = PetscFree(procsnz);CHKERRQ(ierr); 2174 2175 } else { 2176 /* receive numeric values */ 2177 ierr = PetscMalloc(nz*sizeof(Scalar),&buf);CHKERRQ(ierr); 2178 2179 /* receive message of values*/ 2180 vals = buf; 2181 mycols = ibuf; 2182 ierr = MPI_Recv(vals,nz,MPIU_SCALAR,0,A->tag,comm,&status);CHKERRQ(ierr); 2183 ierr = MPI_Get_count(&status,MPIU_SCALAR,&maxnz);CHKERRQ(ierr); 2184 if (maxnz != nz) SETERRQ(PETSC_ERR_FILE_UNEXPECTED,"something is wrong with file"); 2185 2186 /* insert into matrix */ 2187 jj = rstart*bs; 2188 for (i=0; i<m; i++) { 2189 ierr = MatSetValues_MPISBAIJ(A,1,&jj,locrowlens[i],mycols,vals,INSERT_VALUES);CHKERRQ(ierr); 2190 mycols += locrowlens[i]; 2191 vals += locrowlens[i]; 2192 jj++; 2193 } 2194 } 2195 2196 ierr = PetscFree(locrowlens);CHKERRQ(ierr); 2197 ierr = PetscFree(buf);CHKERRQ(ierr); 2198 ierr = PetscFree(ibuf);CHKERRQ(ierr); 2199 ierr = PetscFree(rowners);CHKERRQ(ierr); 2200 ierr = PetscFree(dlens);CHKERRQ(ierr); 2201 ierr = PetscFree(mask);CHKERRQ(ierr); 2202 ierr = MatAssemblyBegin(A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2203 ierr = MatAssemblyEnd(A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2204 PetscFunctionReturn(0); 2205 } 2206 EXTERN_C_END 2207 2208 #undef __FUNC__ 2209 #define __FUNC__ "MatMPISBAIJSetHashTableFactor" 2210 /*@ 2211 MatMPISBAIJSetHashTableFactor - Sets the factor required to compute the size of the HashTable. 2212 2213 Input Parameters: 2214 . mat - the matrix 2215 . fact - factor 2216 2217 Collective on Mat 2218 2219 Level: advanced 2220 2221 Notes: 2222 This can also be set by the command line option: -mat_use_hash_table fact 2223 2224 .keywords: matrix, hashtable, factor, HT 2225 2226 .seealso: MatSetOption() 2227 @*/ 2228 int MatMPISBAIJSetHashTableFactor(Mat mat,PetscReal fact) 2229 { 2230 PetscFunctionBegin; 2231 SETERRQ(1,"Function not yet written for SBAIJ format"); 2232 /* PetscFunctionReturn(0); */ 2233 } 2234 2235 #undef __FUNC__ 2236 #define __FUNC__ "MatGetRowMax_MPISBAIJ" 2237 int MatGetRowMax_MPISBAIJ(Mat A,Vec v) 2238 { 2239 Mat_MPISBAIJ *a = (Mat_MPISBAIJ*)A->data; 2240 Mat_SeqBAIJ *b = (Mat_SeqBAIJ*)(a->B)->data; 2241 PetscReal atmp; 2242 double *work,*svalues,*rvalues; 2243 int ierr,i,bs,mbs,*bi,*bj,brow,j,ncols,krow,kcol,col,row,Mbs,bcol; 2244 int rank,size,*rowners_bs,dest,count,source; 2245 Scalar *ba,*va; 2246 MPI_Status stat; 2247 2248 PetscFunctionBegin; 2249 ierr = MatGetRowMax(a->A,v);CHKERRQ(ierr); 2250 ierr = VecGetArray(v,&va);CHKERRQ(ierr); 2251 2252 ierr = MPI_Comm_size(PETSC_COMM_WORLD,&size);CHKERRQ(ierr); 2253 ierr = MPI_Comm_rank(PETSC_COMM_WORLD,&rank);CHKERRQ(ierr); 2254 2255 bs = a->bs; 2256 mbs = a->mbs; 2257 Mbs = a->Mbs; 2258 ba = b->a; 2259 bi = b->i; 2260 bj = b->j; 2261 /* 2262 PetscSynchronizedPrintf(PETSC_COMM_WORLD,"[%d] M: %d, bs: %d, mbs: %d \n",rank,bs*Mbs,bs,mbs); 2263 PetscSynchronizedFlush(PETSC_COMM_WORLD); 2264 */ 2265 2266 /* find ownerships */ 2267 rowners_bs = a->rowners_bs; 2268 /* 2269 if (!rank){ 2270 for (i=0; i<size+1; i++) PetscPrintf(PETSC_COMM_SELF," rowners_bs[%d]: %d\n",i,rowners_bs[i]); 2271 } 2272 */ 2273 2274 /* each proc creates an array to be distributed */ 2275 ierr = PetscMalloc(bs*Mbs*sizeof(PetscReal),&work);CHKERRQ(ierr); 2276 ierr = PetscMemzero(work,bs*Mbs*sizeof(PetscReal));CHKERRQ(ierr); 2277 2278 /* row_max for B */ 2279 if (rank != size-1){ 2280 for (i=0; i<mbs; i++) { 2281 ncols = bi[1] - bi[0]; bi++; 2282 brow = bs*i; 2283 for (j=0; j<ncols; j++){ 2284 bcol = bs*(*bj); 2285 for (kcol=0; kcol<bs; kcol++){ 2286 col = bcol + kcol; /* local col index */ 2287 col += rowners_bs[rank+1]; /* global col index */ 2288 /* PetscPrintf(PETSC_COMM_SELF,"[%d], col: %d\n",rank,col); */ 2289 for (krow=0; krow<bs; krow++){ 2290 atmp = PetscAbsScalar(*ba); ba++; 2291 row = brow + krow; /* local row index */ 2292 /* printf("val[%d,%d]: %g\n",row,col,atmp); */ 2293 if (PetscRealPart(va[row]) < atmp) va[row] = atmp; 2294 if (work[col] < atmp) work[col] = atmp; 2295 } 2296 } 2297 bj++; 2298 } 2299 } 2300 /* 2301 PetscPrintf(PETSC_COMM_SELF,"[%d], work: ",rank); 2302 for (i=0; i<bs*Mbs; i++) PetscPrintf(PETSC_COMM_SELF,"%g ",work[i]); 2303 PetscPrintf(PETSC_COMM_SELF,"[%d]: \n"); 2304 */ 2305 2306 /* send values to its owners */ 2307 for (dest=rank+1; dest<size; dest++){ 2308 svalues = work + rowners_bs[dest]; 2309 count = rowners_bs[dest+1]-rowners_bs[dest]; 2310 ierr = MPI_Send(svalues,count,MPI_DOUBLE,dest,rank,PETSC_COMM_WORLD);CHKERRQ(ierr); 2311 /* 2312 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]); 2313 PetscSynchronizedFlush(PETSC_COMM_WORLD); 2314 */ 2315 } 2316 } 2317 2318 /* receive values */ 2319 if (rank){ 2320 rvalues = work; 2321 count = rowners_bs[rank+1]-rowners_bs[rank]; 2322 for (source=0; source<rank; source++){ 2323 ierr = MPI_Recv(rvalues,count,MPI_DOUBLE,MPI_ANY_SOURCE,MPI_ANY_TAG,PETSC_COMM_WORLD,&stat);CHKERRQ(ierr); 2324 /* process values */ 2325 for (i=0; i<count; i++){ 2326 if (PetscRealPart(va[i]) < rvalues[i]) va[i] = rvalues[i]; 2327 } 2328 /* 2329 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]); 2330 PetscSynchronizedFlush(PETSC_COMM_WORLD); 2331 */ 2332 } 2333 } 2334 2335 ierr = VecRestoreArray(v,&va);CHKERRQ(ierr); 2336 ierr = PetscFree(work);CHKERRQ(ierr); 2337 PetscFunctionReturn(0); 2338 } 2339