1 /*$Id: mpisbaij.c,v 1.61 2001/08/10 03:31:37 bsmith Exp $*/ 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 *,PetscScalar *); 13 extern int MatSetValues_SeqSBAIJ(Mat,int,int *,int,int *,PetscScalar *,InsertMode); 14 extern int MatSetValuesBlocked_SeqSBAIJ(Mat,int,int*,int,int*,PetscScalar*,InsertMode); 15 extern int MatGetRow_SeqSBAIJ(Mat,int,int*,int**,PetscScalar**); 16 extern int MatRestoreRow_SeqSBAIJ(Mat,int,int*,int**,PetscScalar**); 17 extern int MatPrintHelp_SeqSBAIJ(Mat); 18 extern int MatZeroRows_SeqSBAIJ(Mat,IS,PetscScalar*); 19 extern int MatZeroRows_SeqBAIJ(Mat,IS,PetscScalar *); 20 extern int MatGetRowMax_MPISBAIJ(Mat,Vec); 21 extern int MatRelax_MPISBAIJ(Mat,Vec,PetscReal,MatSORType,PetscReal,int,int,Vec); 22 23 /* UGLY, ugly, ugly 24 When MatScalar == PetscScalar the function MatSetValuesBlocked_MPIBAIJ_MatScalar() does 25 not exist. Otherwise ..._MatScalar() takes matrix elements in single precision and 26 inserts them into the single precision data structure. The function MatSetValuesBlocked_MPIBAIJ() 27 converts the entries into single precision and then calls ..._MatScalar() to put them 28 into the single precision data structures. 29 */ 30 #if defined(PETSC_USE_MAT_SINGLE) 31 extern int MatSetValuesBlocked_SeqSBAIJ_MatScalar(Mat,int,int*,int,int*,MatScalar*,InsertMode); 32 extern int MatSetValues_MPISBAIJ_MatScalar(Mat,int,int*,int,int*,MatScalar*,InsertMode); 33 extern int MatSetValuesBlocked_MPISBAIJ_MatScalar(Mat,int,int*,int,int*,MatScalar*,InsertMode); 34 extern int MatSetValues_MPISBAIJ_HT_MatScalar(Mat,int,int*,int,int*,MatScalar*,InsertMode); 35 extern int MatSetValuesBlocked_MPISBAIJ_HT_MatScalar(Mat,int,int*,int,int*,MatScalar*,InsertMode); 36 #else 37 #define MatSetValuesBlocked_SeqSBAIJ_MatScalar MatSetValuesBlocked_SeqSBAIJ 38 #define MatSetValues_MPISBAIJ_MatScalar MatSetValues_MPISBAIJ 39 #define MatSetValuesBlocked_MPISBAIJ_MatScalar MatSetValuesBlocked_MPISBAIJ 40 #define MatSetValues_MPISBAIJ_HT_MatScalar MatSetValues_MPISBAIJ_HT 41 #define MatSetValuesBlocked_MPISBAIJ_HT_MatScalar MatSetValuesBlocked_MPISBAIJ_HT 42 #endif 43 44 EXTERN_C_BEGIN 45 #undef __FUNCT__ 46 #define __FUNCT__ "MatStoreValues_MPISBAIJ" 47 int MatStoreValues_MPISBAIJ(Mat mat) 48 { 49 Mat_MPISBAIJ *aij = (Mat_MPISBAIJ *)mat->data; 50 int ierr; 51 52 PetscFunctionBegin; 53 ierr = MatStoreValues(aij->A);CHKERRQ(ierr); 54 ierr = MatStoreValues(aij->B);CHKERRQ(ierr); 55 PetscFunctionReturn(0); 56 } 57 EXTERN_C_END 58 59 EXTERN_C_BEGIN 60 #undef __FUNCT__ 61 #define __FUNCT__ "MatRetrieveValues_MPISBAIJ" 62 int MatRetrieveValues_MPISBAIJ(Mat mat) 63 { 64 Mat_MPISBAIJ *aij = (Mat_MPISBAIJ *)mat->data; 65 int ierr; 66 67 PetscFunctionBegin; 68 ierr = MatRetrieveValues(aij->A);CHKERRQ(ierr); 69 ierr = MatRetrieveValues(aij->B);CHKERRQ(ierr); 70 PetscFunctionReturn(0); 71 } 72 EXTERN_C_END 73 74 /* 75 Local utility routine that creates a mapping from the global column 76 number to the local number in the off-diagonal part of the local 77 storage of the matrix. This is done in a non scable way since the 78 length of colmap equals the global matrix length. 79 */ 80 #undef __FUNCT__ 81 #define __FUNCT__ "CreateColmap_MPISBAIJ_Private" 82 static int CreateColmap_MPISBAIJ_Private(Mat mat) 83 { 84 PetscFunctionBegin; 85 SETERRQ(1,"Function not yet written for SBAIJ format"); 86 /* PetscFunctionReturn(0); */ 87 } 88 89 #define CHUNKSIZE 10 90 91 #define MatSetValues_SeqSBAIJ_A_Private(row,col,value,addv) \ 92 { \ 93 \ 94 brow = row/bs; \ 95 rp = aj + ai[brow]; ap = aa + bs2*ai[brow]; \ 96 rmax = aimax[brow]; nrow = ailen[brow]; \ 97 bcol = col/bs; \ 98 ridx = row % bs; cidx = col % bs; \ 99 low = 0; high = nrow; \ 100 while (high-low > 3) { \ 101 t = (low+high)/2; \ 102 if (rp[t] > bcol) high = t; \ 103 else low = t; \ 104 } \ 105 for (_i=low; _i<high; _i++) { \ 106 if (rp[_i] > bcol) break; \ 107 if (rp[_i] == bcol) { \ 108 bap = ap + bs2*_i + bs*cidx + ridx; \ 109 if (addv == ADD_VALUES) *bap += value; \ 110 else *bap = value; \ 111 goto a_noinsert; \ 112 } \ 113 } \ 114 if (a->nonew == 1) goto a_noinsert; \ 115 else if (a->nonew == -1) SETERRQ(PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new nonzero into matrix"); \ 116 if (nrow >= rmax) { \ 117 /* there is no extra room in row, therefore enlarge */ \ 118 int new_nz = ai[a->mbs] + CHUNKSIZE,len,*new_i,*new_j; \ 119 MatScalar *new_a; \ 120 \ 121 if (a->nonew == -2) SETERRQ(PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new nonzero in the matrix"); \ 122 \ 123 /* malloc new storage space */ \ 124 len = new_nz*(sizeof(int)+bs2*sizeof(MatScalar))+(a->mbs+1)*sizeof(int); \ 125 ierr = PetscMalloc(len,&new_a);CHKERRQ(ierr); \ 126 new_j = (int*)(new_a + bs2*new_nz); \ 127 new_i = new_j + new_nz; \ 128 \ 129 /* copy over old data into new slots */ \ 130 for (ii=0; ii<brow+1; ii++) {new_i[ii] = ai[ii];} \ 131 for (ii=brow+1; ii<a->mbs+1; ii++) {new_i[ii] = ai[ii]+CHUNKSIZE;} \ 132 ierr = PetscMemcpy(new_j,aj,(ai[brow]+nrow)*sizeof(int));CHKERRQ(ierr); \ 133 len = (new_nz - CHUNKSIZE - ai[brow] - nrow); \ 134 ierr = PetscMemcpy(new_j+ai[brow]+nrow+CHUNKSIZE,aj+ai[brow]+nrow,len*sizeof(int));CHKERRQ(ierr); \ 135 ierr = PetscMemcpy(new_a,aa,(ai[brow]+nrow)*bs2*sizeof(MatScalar));CHKERRQ(ierr); \ 136 ierr = PetscMemzero(new_a+bs2*(ai[brow]+nrow),bs2*CHUNKSIZE*sizeof(PetscScalar));CHKERRQ(ierr); \ 137 ierr = PetscMemcpy(new_a+bs2*(ai[brow]+nrow+CHUNKSIZE), \ 138 aa+bs2*(ai[brow]+nrow),bs2*len*sizeof(MatScalar));CHKERRQ(ierr); \ 139 /* free up old matrix storage */ \ 140 ierr = PetscFree(a->a);CHKERRQ(ierr); \ 141 if (!a->singlemalloc) { \ 142 ierr = PetscFree(a->i);CHKERRQ(ierr); \ 143 ierr = PetscFree(a->j);CHKERRQ(ierr);\ 144 } \ 145 aa = a->a = new_a; ai = a->i = new_i; aj = a->j = new_j; \ 146 a->singlemalloc = PETSC_TRUE; \ 147 \ 148 rp = aj + ai[brow]; ap = aa + bs2*ai[brow]; \ 149 rmax = aimax[brow] = aimax[brow] + CHUNKSIZE; \ 150 PetscLogObjectMemory(A,CHUNKSIZE*(sizeof(int) + bs2*sizeof(MatScalar))); \ 151 a->s_maxnz += bs2*CHUNKSIZE; \ 152 a->reallocs++; \ 153 a->s_nz++; \ 154 } \ 155 N = nrow++ - 1; \ 156 /* shift up all the later entries in this row */ \ 157 for (ii=N; ii>=_i; ii--) { \ 158 rp[ii+1] = rp[ii]; \ 159 ierr = PetscMemcpy(ap+bs2*(ii+1),ap+bs2*(ii),bs2*sizeof(MatScalar));CHKERRQ(ierr); \ 160 } \ 161 if (N>=_i) { ierr = PetscMemzero(ap+bs2*_i,bs2*sizeof(MatScalar));CHKERRQ(ierr); } \ 162 rp[_i] = bcol; \ 163 ap[bs2*_i + bs*cidx + ridx] = value; \ 164 a_noinsert:; \ 165 ailen[brow] = nrow; \ 166 } 167 #ifndef MatSetValues_SeqBAIJ_B_Private 168 #define MatSetValues_SeqSBAIJ_B_Private(row,col,value,addv) \ 169 { \ 170 brow = row/bs; \ 171 rp = bj + bi[brow]; ap = ba + bs2*bi[brow]; \ 172 rmax = bimax[brow]; nrow = bilen[brow]; \ 173 bcol = col/bs; \ 174 ridx = row % bs; cidx = col % bs; \ 175 low = 0; high = nrow; \ 176 while (high-low > 3) { \ 177 t = (low+high)/2; \ 178 if (rp[t] > bcol) high = t; \ 179 else low = t; \ 180 } \ 181 for (_i=low; _i<high; _i++) { \ 182 if (rp[_i] > bcol) break; \ 183 if (rp[_i] == bcol) { \ 184 bap = ap + bs2*_i + bs*cidx + ridx; \ 185 if (addv == ADD_VALUES) *bap += value; \ 186 else *bap = value; \ 187 goto b_noinsert; \ 188 } \ 189 } \ 190 if (b->nonew == 1) goto b_noinsert; \ 191 else if (b->nonew == -1) SETERRQ(PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new nonzero into matrix"); \ 192 if (nrow >= rmax) { \ 193 /* there is no extra room in row, therefore enlarge */ \ 194 int new_nz = bi[b->mbs] + CHUNKSIZE,len,*new_i,*new_j; \ 195 MatScalar *new_a; \ 196 \ 197 if (b->nonew == -2) SETERRQ(PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new nonzero in the matrix"); \ 198 \ 199 /* malloc new storage space */ \ 200 len = new_nz*(sizeof(int)+bs2*sizeof(MatScalar))+(b->mbs+1)*sizeof(int); \ 201 ierr = PetscMalloc(len,&new_a);CHKERRQ(ierr); \ 202 new_j = (int*)(new_a + bs2*new_nz); \ 203 new_i = new_j + new_nz; \ 204 \ 205 /* copy over old data into new slots */ \ 206 for (ii=0; ii<brow+1; ii++) {new_i[ii] = bi[ii];} \ 207 for (ii=brow+1; ii<b->mbs+1; ii++) {new_i[ii] = bi[ii]+CHUNKSIZE;} \ 208 ierr = PetscMemcpy(new_j,bj,(bi[brow]+nrow)*sizeof(int));CHKERRQ(ierr); \ 209 len = (new_nz - CHUNKSIZE - bi[brow] - nrow); \ 210 ierr = PetscMemcpy(new_j+bi[brow]+nrow+CHUNKSIZE,bj+bi[brow]+nrow,len*sizeof(int));CHKERRQ(ierr); \ 211 ierr = PetscMemcpy(new_a,ba,(bi[brow]+nrow)*bs2*sizeof(MatScalar));CHKERRQ(ierr); \ 212 ierr = PetscMemzero(new_a+bs2*(bi[brow]+nrow),bs2*CHUNKSIZE*sizeof(MatScalar));CHKERRQ(ierr); \ 213 ierr = PetscMemcpy(new_a+bs2*(bi[brow]+nrow+CHUNKSIZE), \ 214 ba+bs2*(bi[brow]+nrow),bs2*len*sizeof(MatScalar));CHKERRQ(ierr); \ 215 /* free up old matrix storage */ \ 216 ierr = PetscFree(b->a);CHKERRQ(ierr); \ 217 if (!b->singlemalloc) { \ 218 ierr = PetscFree(b->i);CHKERRQ(ierr); \ 219 ierr = PetscFree(b->j);CHKERRQ(ierr); \ 220 } \ 221 ba = b->a = new_a; bi = b->i = new_i; bj = b->j = new_j; \ 222 b->singlemalloc = PETSC_TRUE; \ 223 \ 224 rp = bj + bi[brow]; ap = ba + bs2*bi[brow]; \ 225 rmax = bimax[brow] = bimax[brow] + CHUNKSIZE; \ 226 PetscLogObjectMemory(B,CHUNKSIZE*(sizeof(int) + bs2*sizeof(MatScalar))); \ 227 b->maxnz += bs2*CHUNKSIZE; \ 228 b->reallocs++; \ 229 b->nz++; \ 230 } \ 231 N = nrow++ - 1; \ 232 /* shift up all the later entries in this row */ \ 233 for (ii=N; ii>=_i; ii--) { \ 234 rp[ii+1] = rp[ii]; \ 235 ierr = PetscMemcpy(ap+bs2*(ii+1),ap+bs2*(ii),bs2*sizeof(MatScalar));CHKERRQ(ierr); \ 236 } \ 237 if (N>=_i) { ierr = PetscMemzero(ap+bs2*_i,bs2*sizeof(MatScalar));CHKERRQ(ierr);} \ 238 rp[_i] = bcol; \ 239 ap[bs2*_i + bs*cidx + ridx] = value; \ 240 b_noinsert:; \ 241 bilen[brow] = nrow; \ 242 } 243 #endif 244 245 #if defined(PETSC_USE_MAT_SINGLE) 246 #undef __FUNCT__ 247 #define __FUNCT__ "MatSetValues_MPISBAIJ" 248 int MatSetValues_MPISBAIJ(Mat mat,int m,int *im,int n,int *in,PetscScalar *v,InsertMode addv) 249 { 250 Mat_MPISBAIJ *b = (Mat_MPISBAIJ*)mat->data; 251 int ierr,i,N = m*n; 252 MatScalar *vsingle; 253 254 PetscFunctionBegin; 255 if (N > b->setvalueslen) { 256 if (b->setvaluescopy) {ierr = PetscFree(b->setvaluescopy);CHKERRQ(ierr);} 257 ierr = PetscMalloc(N*sizeof(MatScalar),&b->setvaluescopy);CHKERRQ(ierr); 258 b->setvalueslen = N; 259 } 260 vsingle = b->setvaluescopy; 261 262 for (i=0; i<N; i++) { 263 vsingle[i] = v[i]; 264 } 265 ierr = MatSetValues_MPISBAIJ_MatScalar(mat,m,im,n,in,vsingle,addv);CHKERRQ(ierr); 266 PetscFunctionReturn(0); 267 } 268 269 #undef __FUNCT__ 270 #define __FUNCT__ "MatSetValuesBlocked_MPISBAIJ" 271 int MatSetValuesBlocked_MPISBAIJ(Mat mat,int m,int *im,int n,int *in,PetscScalar *v,InsertMode addv) 272 { 273 Mat_MPIBAIJ *b = (Mat_MPIBAIJ*)mat->data; 274 int ierr,i,N = m*n*b->bs2; 275 MatScalar *vsingle; 276 277 PetscFunctionBegin; 278 if (N > b->setvalueslen) { 279 if (b->setvaluescopy) {ierr = PetscFree(b->setvaluescopy);CHKERRQ(ierr);} 280 ierr = PetscMalloc(N*sizeof(MatScalar),&b->setvaluescopy);CHKERRQ(ierr); 281 b->setvalueslen = N; 282 } 283 vsingle = b->setvaluescopy; 284 for (i=0; i<N; i++) { 285 vsingle[i] = v[i]; 286 } 287 ierr = MatSetValuesBlocked_MPISBAIJ_MatScalar(mat,m,im,n,in,vsingle,addv);CHKERRQ(ierr); 288 PetscFunctionReturn(0); 289 } 290 291 #undef __FUNCT__ 292 #define __FUNCT__ "MatSetValues_MPISBAIJ_HT" 293 int MatSetValues_MPISBAIJ_HT(Mat mat,int m,int *im,int n,int *in,PetscScalar *v,InsertMode addv) 294 { 295 Mat_MPIBAIJ *b = (Mat_MPIBAIJ*)mat->data; 296 int ierr,i,N = m*n; 297 MatScalar *vsingle; 298 299 PetscFunctionBegin; 300 SETERRQ(1,"Function not yet written for SBAIJ format"); 301 /* PetscFunctionReturn(0); */ 302 } 303 304 #undef __FUNCT__ 305 #define __FUNCT__ "MatSetValuesBlocked_MPISBAIJ_HT" 306 int MatSetValuesBlocked_MPISBAIJ_HT(Mat mat,int m,int *im,int n,int *in,PetscScalar *v,InsertMode addv) 307 { 308 Mat_MPIBAIJ *b = (Mat_MPIBAIJ*)mat->data; 309 int ierr,i,N = m*n*b->bs2; 310 MatScalar *vsingle; 311 312 PetscFunctionBegin; 313 SETERRQ(1,"Function not yet written for SBAIJ format"); 314 /* PetscFunctionReturn(0); */ 315 } 316 #endif 317 318 /* Only add/insert a(i,j) with i<=j (blocks). 319 Any a(i,j) with i>j input by user is ingored. 320 */ 321 #undef __FUNCT__ 322 #define __FUNCT__ "MatSetValues_MPIBAIJ" 323 int MatSetValues_MPISBAIJ_MatScalar(Mat mat,int m,int *im,int n,int *in,MatScalar *v,InsertMode addv) 324 { 325 Mat_MPISBAIJ *baij = (Mat_MPISBAIJ*)mat->data; 326 MatScalar value; 327 PetscTruth roworiented = baij->roworiented; 328 int ierr,i,j,row,col; 329 int rstart_orig=baij->rstart_bs; 330 int rend_orig=baij->rend_bs,cstart_orig=baij->cstart_bs; 331 int cend_orig=baij->cend_bs,bs=baij->bs; 332 333 /* Some Variables required in the macro */ 334 Mat A = baij->A; 335 Mat_SeqSBAIJ *a = (Mat_SeqSBAIJ*)(A)->data; 336 int *aimax=a->imax,*ai=a->i,*ailen=a->ilen,*aj=a->j; 337 MatScalar *aa=a->a; 338 339 Mat B = baij->B; 340 Mat_SeqBAIJ *b = (Mat_SeqBAIJ*)(B)->data; 341 int *bimax=b->imax,*bi=b->i,*bilen=b->ilen,*bj=b->j; 342 MatScalar *ba=b->a; 343 344 int *rp,ii,nrow,_i,rmax,N,brow,bcol; 345 int low,high,t,ridx,cidx,bs2=a->bs2; 346 MatScalar *ap,*bap; 347 348 /* for stash */ 349 int n_loc, *in_loc=0; 350 MatScalar *v_loc=0; 351 352 PetscFunctionBegin; 353 354 if(!baij->donotstash){ 355 ierr = PetscMalloc(n*sizeof(int),&in_loc);CHKERRQ(ierr); 356 ierr = PetscMalloc(n*sizeof(MatScalar),&v_loc);CHKERRQ(ierr); 357 } 358 359 for (i=0; i<m; i++) { 360 if (im[i] < 0) continue; 361 #if defined(PETSC_USE_BOPT_g) 362 if (im[i] >= mat->M) SETERRQ(PETSC_ERR_ARG_OUTOFRANGE,"Row too large"); 363 #endif 364 if (im[i] >= rstart_orig && im[i] < rend_orig) { /* this processor entry */ 365 row = im[i] - rstart_orig; /* local row index */ 366 for (j=0; j<n; j++) { 367 if (im[i]/bs > in[j]/bs) continue; /* ignore lower triangular blocks */ 368 if (in[j] >= cstart_orig && in[j] < cend_orig){ /* diag entry (A) */ 369 col = in[j] - cstart_orig; /* local col index */ 370 brow = row/bs; bcol = col/bs; 371 if (brow > bcol) continue; /* ignore lower triangular blocks of A */ 372 if (roworiented) value = v[i*n+j]; else value = v[i+j*m]; 373 MatSetValues_SeqSBAIJ_A_Private(row,col,value,addv); 374 /* ierr = MatSetValues_SeqBAIJ(baij->A,1,&row,1,&col,&value,addv);CHKERRQ(ierr); */ 375 } else if (in[j] < 0) continue; 376 #if defined(PETSC_USE_BOPT_g) 377 else if (in[j] >= mat->N) {SETERRQ(PETSC_ERR_ARG_OUTOFRANGE,"Col too large");} 378 #endif 379 else { /* off-diag entry (B) */ 380 if (mat->was_assembled) { 381 if (!baij->colmap) { 382 ierr = CreateColmap_MPISBAIJ_Private(mat);CHKERRQ(ierr); 383 } 384 #if defined (PETSC_USE_CTABLE) 385 ierr = PetscTableFind(baij->colmap,in[j]/bs + 1,&col);CHKERRQ(ierr); 386 col = col - 1; 387 #else 388 col = baij->colmap[in[j]/bs] - 1; 389 #endif 390 if (col < 0 && !((Mat_SeqSBAIJ*)(baij->A->data))->nonew) { 391 ierr = DisAssemble_MPISBAIJ(mat);CHKERRQ(ierr); 392 col = in[j]; 393 /* Reinitialize the variables required by MatSetValues_SeqBAIJ_B_Private() */ 394 B = baij->B; 395 b = (Mat_SeqBAIJ*)(B)->data; 396 bimax=b->imax;bi=b->i;bilen=b->ilen;bj=b->j; 397 ba=b->a; 398 } else col += in[j]%bs; 399 } else col = in[j]; 400 if (roworiented) value = v[i*n+j]; else value = v[i+j*m]; 401 MatSetValues_SeqSBAIJ_B_Private(row,col,value,addv); 402 /* ierr = MatSetValues_SeqBAIJ(baij->B,1,&row,1,&col,&value,addv);CHKERRQ(ierr); */ 403 } 404 } 405 } else { /* off processor entry */ 406 if (!baij->donotstash) { 407 n_loc = 0; 408 for (j=0; j<n; j++){ 409 if (im[i]/bs > in[j]/bs) continue; /* ignore lower triangular blocks */ 410 in_loc[n_loc] = in[j]; 411 if (roworiented) { 412 v_loc[n_loc] = v[i*n+j]; 413 } else { 414 v_loc[n_loc] = v[j*m+i]; 415 } 416 n_loc++; 417 } 418 ierr = MatStashValuesRow_Private(&mat->stash,im[i],n_loc,in_loc,v_loc);CHKERRQ(ierr); 419 } 420 } 421 } 422 423 if(!baij->donotstash){ 424 ierr = PetscFree(in_loc);CHKERRQ(ierr); 425 ierr = PetscFree(v_loc);CHKERRQ(ierr); 426 } 427 PetscFunctionReturn(0); 428 } 429 430 #undef __FUNCT__ 431 #define __FUNCT__ "MatSetValuesBlocked_MPISBAIJ" 432 int MatSetValuesBlocked_MPISBAIJ_MatScalar(Mat mat,int m,int *im,int n,int *in,MatScalar *v,InsertMode addv) 433 { 434 PetscFunctionBegin; 435 SETERRQ(1,"Function not yet written for SBAIJ format"); 436 /* PetscFunctionReturn(0); */ 437 } 438 439 #define HASH_KEY 0.6180339887 440 #define HASH(size,key,tmp) (tmp = (key)*HASH_KEY,(int)((size)*(tmp-(int)tmp))) 441 /* #define HASH(size,key) ((int)((size)*fmod(((key)*HASH_KEY),1))) */ 442 /* #define HASH(size,key,tmp) ((int)((size)*fmod(((key)*HASH_KEY),1))) */ 443 #undef __FUNCT__ 444 #define __FUNCT__ "MatSetValues_MPISBAIJ_HT_MatScalar" 445 int MatSetValues_MPISBAIJ_HT_MatScalar(Mat mat,int m,int *im,int n,int *in,MatScalar *v,InsertMode addv) 446 { 447 PetscFunctionBegin; 448 SETERRQ(1,"Function not yet written for SBAIJ format"); 449 /* PetscFunctionReturn(0); */ 450 } 451 452 #undef __FUNCT__ 453 #define __FUNCT__ "MatSetValuesBlocked_MPISBAIJ_HT_MatScalar" 454 int MatSetValuesBlocked_MPISBAIJ_HT_MatScalar(Mat mat,int m,int *im,int n,int *in,MatScalar *v,InsertMode addv) 455 { 456 PetscFunctionBegin; 457 SETERRQ(1,"Function not yet written for SBAIJ format"); 458 /* PetscFunctionReturn(0); */ 459 } 460 461 #undef __FUNCT__ 462 #define __FUNCT__ "MatGetValues_MPISBAIJ" 463 int MatGetValues_MPISBAIJ(Mat mat,int m,int *idxm,int n,int *idxn,PetscScalar *v) 464 { 465 Mat_MPISBAIJ *baij = (Mat_MPISBAIJ*)mat->data; 466 int bs=baij->bs,ierr,i,j,bsrstart = baij->rstart*bs,bsrend = baij->rend*bs; 467 int bscstart = baij->cstart*bs,bscend = baij->cend*bs,row,col,data; 468 469 PetscFunctionBegin; 470 for (i=0; i<m; i++) { 471 if (idxm[i] < 0) SETERRQ(PETSC_ERR_ARG_OUTOFRANGE,"Negative row"); 472 if (idxm[i] >= mat->M) SETERRQ(PETSC_ERR_ARG_OUTOFRANGE,"Row too large"); 473 if (idxm[i] >= bsrstart && idxm[i] < bsrend) { 474 row = idxm[i] - bsrstart; 475 for (j=0; j<n; j++) { 476 if (idxn[j] < 0) SETERRQ(PETSC_ERR_ARG_OUTOFRANGE,"Negative column"); 477 if (idxn[j] >= mat->N) SETERRQ(PETSC_ERR_ARG_OUTOFRANGE,"Column too large"); 478 if (idxn[j] >= bscstart && idxn[j] < bscend){ 479 col = idxn[j] - bscstart; 480 ierr = MatGetValues_SeqSBAIJ(baij->A,1,&row,1,&col,v+i*n+j);CHKERRQ(ierr); 481 } else { 482 if (!baij->colmap) { 483 ierr = CreateColmap_MPISBAIJ_Private(mat);CHKERRQ(ierr); 484 } 485 #if defined (PETSC_USE_CTABLE) 486 ierr = PetscTableFind(baij->colmap,idxn[j]/bs+1,&data);CHKERRQ(ierr); 487 data --; 488 #else 489 data = baij->colmap[idxn[j]/bs]-1; 490 #endif 491 if((data < 0) || (baij->garray[data/bs] != idxn[j]/bs)) *(v+i*n+j) = 0.0; 492 else { 493 col = data + idxn[j]%bs; 494 ierr = MatGetValues_SeqSBAIJ(baij->B,1,&row,1,&col,v+i*n+j);CHKERRQ(ierr); 495 } 496 } 497 } 498 } else { 499 SETERRQ(PETSC_ERR_SUP,"Only local values currently supported"); 500 } 501 } 502 PetscFunctionReturn(0); 503 } 504 505 #undef __FUNCT__ 506 #define __FUNCT__ "MatNorm_MPISBAIJ" 507 int MatNorm_MPISBAIJ(Mat mat,NormType type,PetscReal *norm) 508 { 509 Mat_MPISBAIJ *baij = (Mat_MPISBAIJ*)mat->data; 510 /* Mat_SeqSBAIJ *amat = (Mat_SeqSBAIJ*)baij->A->data; */ 511 /* Mat_SeqBAIJ *bmat = (Mat_SeqBAIJ*)baij->B->data; */ 512 int ierr; 513 PetscReal sum[2],*lnorm2; 514 515 PetscFunctionBegin; 516 if (baij->size == 1) { 517 ierr = MatNorm(baij->A,type,norm);CHKERRQ(ierr); 518 } else { 519 if (type == NORM_FROBENIUS) { 520 ierr = PetscMalloc(2*sizeof(PetscReal),&lnorm2);CHKERRQ(ierr); 521 ierr = MatNorm(baij->A,type,lnorm2);CHKERRQ(ierr); 522 *lnorm2 = (*lnorm2)*(*lnorm2); lnorm2++; /* squar power of norm(A) */ 523 ierr = MatNorm(baij->B,type,lnorm2);CHKERRQ(ierr); 524 *lnorm2 = (*lnorm2)*(*lnorm2); lnorm2--; /* squar power of norm(B) */ 525 /* 526 ierr = MPI_Comm_rank(PETSC_COMM_WORLD,&rank);CHKERRQ(ierr); 527 PetscSynchronizedPrintf(PETSC_COMM_WORLD,"[%d], lnorm2=%g, %g\n",rank,lnorm2[0],lnorm2[1]); 528 */ 529 ierr = MPI_Allreduce(lnorm2,&sum,2,MPIU_REAL,MPI_SUM,mat->comm);CHKERRQ(ierr); 530 /* 531 PetscSynchronizedPrintf(PETSC_COMM_WORLD,"[%d], sum=%g, %g\n",rank,sum[0],sum[1]); 532 PetscSynchronizedFlush(PETSC_COMM_WORLD); */ 533 534 *norm = sqrt(sum[0] + 2*sum[1]); 535 ierr = PetscFree(lnorm2);CHKERRQ(ierr); 536 } else { 537 SETERRQ(PETSC_ERR_SUP,"No support for this norm yet"); 538 } 539 } 540 PetscFunctionReturn(0); 541 } 542 543 /* 544 Creates the hash table, and sets the table 545 This table is created only once. 546 If new entried need to be added to the matrix 547 then the hash table has to be destroyed and 548 recreated. 549 */ 550 #undef __FUNCT__ 551 #define __FUNCT__ "MatCreateHashTable_MPISBAIJ_Private" 552 int MatCreateHashTable_MPISBAIJ_Private(Mat mat,PetscReal factor) 553 { 554 PetscFunctionBegin; 555 SETERRQ(1,"Function not yet written for SBAIJ format"); 556 /* PetscFunctionReturn(0); */ 557 } 558 559 #undef __FUNCT__ 560 #define __FUNCT__ "MatAssemblyBegin_MPISBAIJ" 561 int MatAssemblyBegin_MPISBAIJ(Mat mat,MatAssemblyType mode) 562 { 563 Mat_MPISBAIJ *baij = (Mat_MPISBAIJ*)mat->data; 564 int ierr,nstash,reallocs; 565 InsertMode addv; 566 567 PetscFunctionBegin; 568 if (baij->donotstash) { 569 PetscFunctionReturn(0); 570 } 571 572 /* make sure all processors are either in INSERTMODE or ADDMODE */ 573 ierr = MPI_Allreduce(&mat->insertmode,&addv,1,MPI_INT,MPI_BOR,mat->comm);CHKERRQ(ierr); 574 if (addv == (ADD_VALUES|INSERT_VALUES)) { 575 SETERRQ(PETSC_ERR_ARG_WRONGSTATE,"Some processors inserted others added"); 576 } 577 mat->insertmode = addv; /* in case this processor had no cache */ 578 579 ierr = MatStashScatterBegin_Private(&mat->stash,baij->rowners_bs);CHKERRQ(ierr); 580 ierr = MatStashScatterBegin_Private(&mat->bstash,baij->rowners);CHKERRQ(ierr); 581 ierr = MatStashGetInfo_Private(&mat->stash,&nstash,&reallocs);CHKERRQ(ierr); 582 PetscLogInfo(0,"MatAssemblyBegin_MPISBAIJ:Stash has %d entries,uses %d mallocs.\n",nstash,reallocs); 583 ierr = MatStashGetInfo_Private(&mat->stash,&nstash,&reallocs);CHKERRQ(ierr); 584 PetscLogInfo(0,"MatAssemblyBegin_MPISBAIJ:Block-Stash has %d entries, uses %d mallocs.\n",nstash,reallocs); 585 PetscFunctionReturn(0); 586 } 587 588 #undef __FUNCT__ 589 #define __FUNCT__ "MatAssemblyEnd_MPISBAIJ" 590 int MatAssemblyEnd_MPISBAIJ(Mat mat,MatAssemblyType mode) 591 { 592 Mat_MPISBAIJ *baij=(Mat_MPISBAIJ*)mat->data; 593 Mat_SeqSBAIJ *a=(Mat_SeqSBAIJ*)baij->A->data; 594 Mat_SeqBAIJ *b=(Mat_SeqBAIJ*)baij->B->data; 595 int i,j,rstart,ncols,n,ierr,flg,bs2=baij->bs2; 596 int *row,*col,other_disassembled; 597 PetscTruth r1,r2,r3; 598 MatScalar *val; 599 InsertMode addv = mat->insertmode; 600 /* int rank;*/ 601 602 PetscFunctionBegin; 603 /* remove 2 line below later */ 604 /*ierr = MPI_Comm_rank(PETSC_COMM_WORLD, &rank);CHKERRQ(ierr); */ 605 606 if (!baij->donotstash) { 607 while (1) { 608 ierr = MatStashScatterGetMesg_Private(&mat->stash,&n,&row,&col,&val,&flg);CHKERRQ(ierr); 609 /* 610 PetscSynchronizedPrintf(PETSC_COMM_WORLD,"[%d]: in AssemblyEnd, stash, flg=%d\n",rank,flg); 611 PetscSynchronizedFlush(PETSC_COMM_WORLD); 612 */ 613 if (!flg) break; 614 615 for (i=0; i<n;) { 616 /* Now identify the consecutive vals belonging to the same row */ 617 for (j=i,rstart=row[j]; j<n; j++) { if (row[j] != rstart) break; } 618 if (j < n) ncols = j-i; 619 else ncols = n-i; 620 /* Now assemble all these values with a single function call */ 621 ierr = MatSetValues_MPISBAIJ_MatScalar(mat,1,row+i,ncols,col+i,val+i,addv);CHKERRQ(ierr); 622 i = j; 623 } 624 } 625 ierr = MatStashScatterEnd_Private(&mat->stash);CHKERRQ(ierr); 626 /* Now process the block-stash. Since the values are stashed column-oriented, 627 set the roworiented flag to column oriented, and after MatSetValues() 628 restore the original flags */ 629 r1 = baij->roworiented; 630 r2 = a->roworiented; 631 r3 = b->roworiented; 632 baij->roworiented = PETSC_FALSE; 633 a->roworiented = PETSC_FALSE; 634 b->roworiented = PETSC_FALSE; 635 while (1) { 636 ierr = MatStashScatterGetMesg_Private(&mat->bstash,&n,&row,&col,&val,&flg);CHKERRQ(ierr); 637 if (!flg) break; 638 639 for (i=0; i<n;) { 640 /* Now identify the consecutive vals belonging to the same row */ 641 for (j=i,rstart=row[j]; j<n; j++) { if (row[j] != rstart) break; } 642 if (j < n) ncols = j-i; 643 else ncols = n-i; 644 ierr = MatSetValuesBlocked_MPISBAIJ_MatScalar(mat,1,row+i,ncols,col+i,val+i*bs2,addv);CHKERRQ(ierr); 645 i = j; 646 } 647 } 648 ierr = MatStashScatterEnd_Private(&mat->bstash);CHKERRQ(ierr); 649 baij->roworiented = r1; 650 a->roworiented = r2; 651 b->roworiented = r3; 652 } 653 654 ierr = MatAssemblyBegin(baij->A,mode);CHKERRQ(ierr); 655 ierr = MatAssemblyEnd(baij->A,mode);CHKERRQ(ierr); 656 657 /* determine if any processor has disassembled, if so we must 658 also disassemble ourselfs, in order that we may reassemble. */ 659 /* 660 if nonzero structure of submatrix B cannot change then we know that 661 no processor disassembled thus we can skip this stuff 662 */ 663 if (!((Mat_SeqBAIJ*)baij->B->data)->nonew) { 664 ierr = MPI_Allreduce(&mat->was_assembled,&other_disassembled,1,MPI_INT,MPI_PROD,mat->comm);CHKERRQ(ierr); 665 if (mat->was_assembled && !other_disassembled) { 666 ierr = DisAssemble_MPISBAIJ(mat);CHKERRQ(ierr); 667 } 668 } 669 670 if (!mat->was_assembled && mode == MAT_FINAL_ASSEMBLY) { 671 ierr = MatSetUpMultiply_MPISBAIJ(mat);CHKERRQ(ierr); 672 } 673 ierr = MatAssemblyBegin(baij->B,mode);CHKERRQ(ierr); 674 ierr = MatAssemblyEnd(baij->B,mode);CHKERRQ(ierr); 675 676 #if defined(PETSC_USE_BOPT_g) 677 if (baij->ht && mode== MAT_FINAL_ASSEMBLY) { 678 PetscLogInfo(0,"MatAssemblyEnd_MPISBAIJ:Average Hash Table Search in MatSetValues = %5.2f\n",((PetscReal)baij->ht_total_ct)/baij->ht_insert_ct); 679 baij->ht_total_ct = 0; 680 baij->ht_insert_ct = 0; 681 } 682 #endif 683 if (baij->ht_flag && !baij->ht && mode == MAT_FINAL_ASSEMBLY) { 684 ierr = MatCreateHashTable_MPISBAIJ_Private(mat,baij->ht_fact);CHKERRQ(ierr); 685 mat->ops->setvalues = MatSetValues_MPISBAIJ_HT; 686 mat->ops->setvaluesblocked = MatSetValuesBlocked_MPISBAIJ_HT; 687 } 688 689 if (baij->rowvalues) { 690 ierr = PetscFree(baij->rowvalues);CHKERRQ(ierr); 691 baij->rowvalues = 0; 692 } 693 PetscFunctionReturn(0); 694 } 695 696 #undef __FUNCT__ 697 #define __FUNCT__ "MatView_MPISBAIJ_ASCIIorDraworSocket" 698 static int MatView_MPISBAIJ_ASCIIorDraworSocket(Mat mat,PetscViewer viewer) 699 { 700 Mat_MPISBAIJ *baij = (Mat_MPISBAIJ*)mat->data; 701 int ierr,bs = baij->bs,size = baij->size,rank = baij->rank; 702 PetscTruth isascii,isdraw; 703 PetscViewer sviewer; 704 PetscViewerFormat format; 705 706 PetscFunctionBegin; 707 ierr = PetscTypeCompare((PetscObject)viewer,PETSC_VIEWER_ASCII,&isascii);CHKERRQ(ierr); 708 ierr = PetscTypeCompare((PetscObject)viewer,PETSC_VIEWER_DRAW,&isdraw);CHKERRQ(ierr); 709 if (isascii) { 710 ierr = PetscViewerGetFormat(viewer,&format);CHKERRQ(ierr); 711 if (format == PETSC_VIEWER_ASCII_INFO_LONG) { 712 MatInfo info; 713 ierr = MPI_Comm_rank(mat->comm,&rank);CHKERRQ(ierr); 714 ierr = MatGetInfo(mat,MAT_LOCAL,&info);CHKERRQ(ierr); 715 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"[%d] Local rows %d nz %d nz alloced %d bs %d mem %d\n", 716 rank,mat->m,(int)info.nz_used*bs,(int)info.nz_allocated*bs, 717 baij->bs,(int)info.memory);CHKERRQ(ierr); 718 ierr = MatGetInfo(baij->A,MAT_LOCAL,&info);CHKERRQ(ierr); 719 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"[%d] on-diagonal part: nz %d \n",rank,(int)info.nz_used*bs);CHKERRQ(ierr); 720 ierr = MatGetInfo(baij->B,MAT_LOCAL,&info);CHKERRQ(ierr); 721 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"[%d] off-diagonal part: nz %d \n",rank,(int)info.nz_used*bs);CHKERRQ(ierr); 722 ierr = PetscViewerFlush(viewer);CHKERRQ(ierr); 723 ierr = VecScatterView(baij->Mvctx,viewer);CHKERRQ(ierr); 724 PetscFunctionReturn(0); 725 } else if (format == PETSC_VIEWER_ASCII_INFO) { 726 ierr = PetscViewerASCIIPrintf(viewer," block size is %d\n",bs);CHKERRQ(ierr); 727 PetscFunctionReturn(0); 728 } 729 } 730 731 if (isdraw) { 732 PetscDraw draw; 733 PetscTruth isnull; 734 ierr = PetscViewerDrawGetDraw(viewer,0,&draw);CHKERRQ(ierr); 735 ierr = PetscDrawIsNull(draw,&isnull);CHKERRQ(ierr); if (isnull) PetscFunctionReturn(0); 736 } 737 738 if (size == 1) { 739 ierr = PetscObjectSetName((PetscObject)baij->A,mat->name);CHKERRQ(ierr); 740 ierr = MatView(baij->A,viewer);CHKERRQ(ierr); 741 } else { 742 /* assemble the entire matrix onto first processor. */ 743 Mat A; 744 Mat_SeqSBAIJ *Aloc; 745 Mat_SeqBAIJ *Bloc; 746 int M = mat->M,N = mat->N,*ai,*aj,col,i,j,k,*rvals,mbs = baij->mbs; 747 MatScalar *a; 748 749 if (!rank) { 750 ierr = MatCreateMPISBAIJ(mat->comm,baij->bs,M,N,M,N,0,PETSC_NULL,0,PETSC_NULL,&A);CHKERRQ(ierr); 751 } else { 752 ierr = MatCreateMPISBAIJ(mat->comm,baij->bs,0,0,M,N,0,PETSC_NULL,0,PETSC_NULL,&A);CHKERRQ(ierr); 753 } 754 PetscLogObjectParent(mat,A); 755 756 /* copy over the A part */ 757 Aloc = (Mat_SeqSBAIJ*)baij->A->data; 758 ai = Aloc->i; aj = Aloc->j; a = Aloc->a; 759 ierr = PetscMalloc(bs*sizeof(int),&rvals);CHKERRQ(ierr); 760 761 for (i=0; i<mbs; i++) { 762 rvals[0] = bs*(baij->rstart + i); 763 for (j=1; j<bs; j++) { rvals[j] = rvals[j-1] + 1; } 764 for (j=ai[i]; j<ai[i+1]; j++) { 765 col = (baij->cstart+aj[j])*bs; 766 for (k=0; k<bs; k++) { 767 ierr = MatSetValues_MPISBAIJ_MatScalar(A,bs,rvals,1,&col,a,INSERT_VALUES);CHKERRQ(ierr); 768 col++; a += bs; 769 } 770 } 771 } 772 /* copy over the B part */ 773 Bloc = (Mat_SeqBAIJ*)baij->B->data; 774 ai = Bloc->i; aj = Bloc->j; a = Bloc->a; 775 for (i=0; i<mbs; i++) { 776 rvals[0] = bs*(baij->rstart + i); 777 for (j=1; j<bs; j++) { rvals[j] = rvals[j-1] + 1; } 778 for (j=ai[i]; j<ai[i+1]; j++) { 779 col = baij->garray[aj[j]]*bs; 780 for (k=0; k<bs; k++) { 781 ierr = MatSetValues_MPISBAIJ_MatScalar(A,bs,rvals,1,&col,a,INSERT_VALUES);CHKERRQ(ierr); 782 col++; a += bs; 783 } 784 } 785 } 786 ierr = PetscFree(rvals);CHKERRQ(ierr); 787 ierr = MatAssemblyBegin(A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 788 ierr = MatAssemblyEnd(A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 789 /* 790 Everyone has to call to draw the matrix since the graphics waits are 791 synchronized across all processors that share the PetscDraw object 792 */ 793 ierr = PetscViewerGetSingleton(viewer,&sviewer);CHKERRQ(ierr); 794 if (!rank) { 795 ierr = PetscObjectSetName((PetscObject)((Mat_MPISBAIJ*)(A->data))->A,mat->name);CHKERRQ(ierr); 796 ierr = MatView(((Mat_MPISBAIJ*)(A->data))->A,sviewer);CHKERRQ(ierr); 797 } 798 ierr = PetscViewerRestoreSingleton(viewer,&sviewer);CHKERRQ(ierr); 799 ierr = MatDestroy(A);CHKERRQ(ierr); 800 } 801 PetscFunctionReturn(0); 802 } 803 804 #undef __FUNCT__ 805 #define __FUNCT__ "MatView_MPISBAIJ" 806 int MatView_MPISBAIJ(Mat mat,PetscViewer viewer) 807 { 808 int ierr; 809 PetscTruth isascii,isdraw,issocket,isbinary; 810 811 PetscFunctionBegin; 812 ierr = PetscTypeCompare((PetscObject)viewer,PETSC_VIEWER_ASCII,&isascii);CHKERRQ(ierr); 813 ierr = PetscTypeCompare((PetscObject)viewer,PETSC_VIEWER_DRAW,&isdraw);CHKERRQ(ierr); 814 ierr = PetscTypeCompare((PetscObject)viewer,PETSC_VIEWER_SOCKET,&issocket);CHKERRQ(ierr); 815 ierr = PetscTypeCompare((PetscObject)viewer,PETSC_VIEWER_BINARY,&isbinary);CHKERRQ(ierr); 816 if (isascii || isdraw || issocket || isbinary) { 817 ierr = MatView_MPISBAIJ_ASCIIorDraworSocket(mat,viewer);CHKERRQ(ierr); 818 } else { 819 SETERRQ1(1,"Viewer type %s not supported by MPISBAIJ matrices",((PetscObject)viewer)->type_name); 820 } 821 PetscFunctionReturn(0); 822 } 823 824 #undef __FUNCT__ 825 #define __FUNCT__ "MatDestroy_MPISBAIJ" 826 int MatDestroy_MPISBAIJ(Mat mat) 827 { 828 Mat_MPISBAIJ *baij = (Mat_MPISBAIJ*)mat->data; 829 int ierr; 830 831 PetscFunctionBegin; 832 #if defined(PETSC_USE_LOG) 833 PetscLogObjectState((PetscObject)mat,"Rows=%d,Cols=%d",mat->M,mat->N); 834 #endif 835 ierr = MatStashDestroy_Private(&mat->stash);CHKERRQ(ierr); 836 ierr = MatStashDestroy_Private(&mat->bstash);CHKERRQ(ierr); 837 ierr = PetscFree(baij->rowners);CHKERRQ(ierr); 838 ierr = MatDestroy(baij->A);CHKERRQ(ierr); 839 ierr = MatDestroy(baij->B);CHKERRQ(ierr); 840 #if defined (PETSC_USE_CTABLE) 841 if (baij->colmap) {ierr = PetscTableDelete(baij->colmap);CHKERRQ(ierr);} 842 #else 843 if (baij->colmap) {ierr = PetscFree(baij->colmap);CHKERRQ(ierr);} 844 #endif 845 if (baij->garray) {ierr = PetscFree(baij->garray);CHKERRQ(ierr);} 846 if (baij->lvec) {ierr = VecDestroy(baij->lvec);CHKERRQ(ierr);} 847 if (baij->Mvctx) {ierr = VecScatterDestroy(baij->Mvctx);CHKERRQ(ierr);} 848 if (baij->slvec0) {ierr = VecDestroy(baij->slvec0);CHKERRQ(ierr);} 849 if (baij->slvec1) {ierr = VecDestroy(baij->slvec1);CHKERRQ(ierr);} 850 if (baij->sMvctx) {ierr = VecScatterDestroy(baij->sMvctx);CHKERRQ(ierr);} 851 if (baij->rowvalues) {ierr = PetscFree(baij->rowvalues);CHKERRQ(ierr);} 852 if (baij->barray) {ierr = PetscFree(baij->barray);CHKERRQ(ierr);} 853 if (baij->hd) {ierr = PetscFree(baij->hd);CHKERRQ(ierr);} 854 #if defined(PETSC_USE_MAT_SINGLE) 855 if (baij->setvaluescopy) {ierr = PetscFree(baij->setvaluescopy);CHKERRQ(ierr);} 856 #endif 857 ierr = PetscFree(baij);CHKERRQ(ierr); 858 PetscFunctionReturn(0); 859 } 860 861 #undef __FUNCT__ 862 #define __FUNCT__ "MatMult_MPISBAIJ" 863 int MatMult_MPISBAIJ(Mat A,Vec xx,Vec yy) 864 { 865 Mat_MPISBAIJ *a = (Mat_MPISBAIJ*)A->data; 866 int ierr,nt; 867 868 PetscFunctionBegin; 869 ierr = VecGetLocalSize(xx,&nt);CHKERRQ(ierr); 870 if (nt != A->n) { 871 SETERRQ(PETSC_ERR_ARG_SIZ,"Incompatible partition of A and xx"); 872 } 873 ierr = VecGetLocalSize(yy,&nt);CHKERRQ(ierr); 874 if (nt != A->m) { 875 SETERRQ(PETSC_ERR_ARG_SIZ,"Incompatible parition of A and yy"); 876 } 877 878 ierr = VecScatterBegin(xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD,a->Mvctx);CHKERRQ(ierr); 879 /* do diagonal part */ 880 ierr = (*a->A->ops->mult)(a->A,xx,yy);CHKERRQ(ierr); 881 /* do supperdiagonal part */ 882 ierr = VecScatterEnd(xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD,a->Mvctx);CHKERRQ(ierr); 883 ierr = (*a->B->ops->multadd)(a->B,a->lvec,yy,yy);CHKERRQ(ierr); 884 /* do subdiagonal part */ 885 ierr = (*a->B->ops->multtranspose)(a->B,xx,a->lvec);CHKERRQ(ierr); 886 ierr = VecScatterBegin(a->lvec,yy,ADD_VALUES,SCATTER_REVERSE,a->Mvctx);CHKERRQ(ierr); 887 ierr = VecScatterEnd(a->lvec,yy,ADD_VALUES,SCATTER_REVERSE,a->Mvctx);CHKERRQ(ierr); 888 889 PetscFunctionReturn(0); 890 } 891 892 #undef __FUNCT__ 893 #define __FUNCT__ "MatMultAdd_MPISBAIJ" 894 int MatMultAdd_MPISBAIJ(Mat A,Vec xx,Vec yy,Vec zz) 895 { 896 Mat_MPISBAIJ *a = (Mat_MPISBAIJ*)A->data; 897 int ierr; 898 899 PetscFunctionBegin; 900 ierr = VecScatterBegin(xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD,a->Mvctx);CHKERRQ(ierr); 901 /* do diagonal part */ 902 ierr = (*a->A->ops->multadd)(a->A,xx,yy,zz);CHKERRQ(ierr); 903 /* do supperdiagonal part */ 904 ierr = VecScatterEnd(xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD,a->Mvctx);CHKERRQ(ierr); 905 ierr = (*a->B->ops->multadd)(a->B,a->lvec,zz,zz);CHKERRQ(ierr); 906 907 /* do subdiagonal part */ 908 ierr = (*a->B->ops->multtranspose)(a->B,xx,a->lvec);CHKERRQ(ierr); 909 ierr = VecScatterBegin(a->lvec,zz,ADD_VALUES,SCATTER_REVERSE,a->Mvctx);CHKERRQ(ierr); 910 ierr = VecScatterEnd(a->lvec,zz,ADD_VALUES,SCATTER_REVERSE,a->Mvctx);CHKERRQ(ierr); 911 912 PetscFunctionReturn(0); 913 } 914 915 #undef __FUNCT__ 916 #define __FUNCT__ "MatMultTranspose_MPISBAIJ" 917 int MatMultTranspose_MPISBAIJ(Mat A,Vec xx,Vec yy) 918 { 919 PetscFunctionBegin; 920 SETERRQ(1,"Matrix is symmetric. Call MatMult()."); 921 /* PetscFunctionReturn(0); */ 922 } 923 924 #undef __FUNCT__ 925 #define __FUNCT__ "MatMultTransposeAdd_MPISBAIJ" 926 int MatMultTransposeAdd_MPISBAIJ(Mat A,Vec xx,Vec yy,Vec zz) 927 { 928 PetscFunctionBegin; 929 SETERRQ(1,"Matrix is symmetric. Call MatMultAdd()."); 930 /* PetscFunctionReturn(0); */ 931 } 932 933 /* 934 This only works correctly for square matrices where the subblock A->A is the 935 diagonal block 936 */ 937 #undef __FUNCT__ 938 #define __FUNCT__ "MatGetDiagonal_MPISBAIJ" 939 int MatGetDiagonal_MPISBAIJ(Mat A,Vec v) 940 { 941 Mat_MPISBAIJ *a = (Mat_MPISBAIJ*)A->data; 942 int ierr; 943 944 PetscFunctionBegin; 945 /* if (a->M != a->N) SETERRQ(PETSC_ERR_SUP,"Supports only square matrix where A->A is diag block"); */ 946 ierr = MatGetDiagonal(a->A,v);CHKERRQ(ierr); 947 PetscFunctionReturn(0); 948 } 949 950 #undef __FUNCT__ 951 #define __FUNCT__ "MatScale_MPISBAIJ" 952 int MatScale_MPISBAIJ(PetscScalar *aa,Mat A) 953 { 954 Mat_MPISBAIJ *a = (Mat_MPISBAIJ*)A->data; 955 int ierr; 956 957 PetscFunctionBegin; 958 ierr = MatScale(aa,a->A);CHKERRQ(ierr); 959 ierr = MatScale(aa,a->B);CHKERRQ(ierr); 960 PetscFunctionReturn(0); 961 } 962 963 #undef __FUNCT__ 964 #define __FUNCT__ "MatGetRow_MPISBAIJ" 965 int MatGetRow_MPISBAIJ(Mat matin,int row,int *nz,int **idx,PetscScalar **v) 966 { 967 Mat_MPISBAIJ *mat = (Mat_MPISBAIJ*)matin->data; 968 PetscScalar *vworkA,*vworkB,**pvA,**pvB,*v_p; 969 int bs = mat->bs,bs2 = mat->bs2,i,ierr,*cworkA,*cworkB,**pcA,**pcB; 970 int nztot,nzA,nzB,lrow,brstart = mat->rstart*bs,brend = mat->rend*bs; 971 int *cmap,*idx_p,cstart = mat->cstart; 972 973 PetscFunctionBegin; 974 if (mat->getrowactive == PETSC_TRUE) SETERRQ(PETSC_ERR_ARG_WRONGSTATE,"Already active"); 975 mat->getrowactive = PETSC_TRUE; 976 977 if (!mat->rowvalues && (idx || v)) { 978 /* 979 allocate enough space to hold information from the longest row. 980 */ 981 Mat_SeqSBAIJ *Aa = (Mat_SeqSBAIJ*)mat->A->data; 982 Mat_SeqBAIJ *Ba = (Mat_SeqBAIJ*)mat->B->data; 983 int max = 1,mbs = mat->mbs,tmp; 984 for (i=0; i<mbs; i++) { 985 tmp = Aa->i[i+1] - Aa->i[i] + Ba->i[i+1] - Ba->i[i]; /* row length */ 986 if (max < tmp) { max = tmp; } 987 } 988 ierr = PetscMalloc(max*bs2*(sizeof(int)+sizeof(PetscScalar)),&mat->rowvalues);CHKERRQ(ierr); 989 mat->rowindices = (int*)(mat->rowvalues + max*bs2); 990 } 991 992 if (row < brstart || row >= brend) SETERRQ(PETSC_ERR_SUP,"Only local rows") 993 lrow = row - brstart; /* local row index */ 994 995 pvA = &vworkA; pcA = &cworkA; pvB = &vworkB; pcB = &cworkB; 996 if (!v) {pvA = 0; pvB = 0;} 997 if (!idx) {pcA = 0; if (!v) pcB = 0;} 998 ierr = (*mat->A->ops->getrow)(mat->A,lrow,&nzA,pcA,pvA);CHKERRQ(ierr); 999 ierr = (*mat->B->ops->getrow)(mat->B,lrow,&nzB,pcB,pvB);CHKERRQ(ierr); 1000 nztot = nzA + nzB; 1001 1002 cmap = mat->garray; 1003 if (v || idx) { 1004 if (nztot) { 1005 /* Sort by increasing column numbers, assuming A and B already sorted */ 1006 int imark = -1; 1007 if (v) { 1008 *v = v_p = mat->rowvalues; 1009 for (i=0; i<nzB; i++) { 1010 if (cmap[cworkB[i]/bs] < cstart) v_p[i] = vworkB[i]; 1011 else break; 1012 } 1013 imark = i; 1014 for (i=0; i<nzA; i++) v_p[imark+i] = vworkA[i]; 1015 for (i=imark; i<nzB; i++) v_p[nzA+i] = vworkB[i]; 1016 } 1017 if (idx) { 1018 *idx = idx_p = mat->rowindices; 1019 if (imark > -1) { 1020 for (i=0; i<imark; i++) { 1021 idx_p[i] = cmap[cworkB[i]/bs]*bs + cworkB[i]%bs; 1022 } 1023 } else { 1024 for (i=0; i<nzB; i++) { 1025 if (cmap[cworkB[i]/bs] < cstart) 1026 idx_p[i] = cmap[cworkB[i]/bs]*bs + cworkB[i]%bs ; 1027 else break; 1028 } 1029 imark = i; 1030 } 1031 for (i=0; i<nzA; i++) idx_p[imark+i] = cstart*bs + cworkA[i]; 1032 for (i=imark; i<nzB; i++) idx_p[nzA+i] = cmap[cworkB[i]/bs]*bs + cworkB[i]%bs ; 1033 } 1034 } else { 1035 if (idx) *idx = 0; 1036 if (v) *v = 0; 1037 } 1038 } 1039 *nz = nztot; 1040 ierr = (*mat->A->ops->restorerow)(mat->A,lrow,&nzA,pcA,pvA);CHKERRQ(ierr); 1041 ierr = (*mat->B->ops->restorerow)(mat->B,lrow,&nzB,pcB,pvB);CHKERRQ(ierr); 1042 PetscFunctionReturn(0); 1043 } 1044 1045 #undef __FUNCT__ 1046 #define __FUNCT__ "MatRestoreRow_MPISBAIJ" 1047 int MatRestoreRow_MPISBAIJ(Mat mat,int row,int *nz,int **idx,PetscScalar **v) 1048 { 1049 Mat_MPISBAIJ *baij = (Mat_MPISBAIJ*)mat->data; 1050 1051 PetscFunctionBegin; 1052 if (baij->getrowactive == PETSC_FALSE) { 1053 SETERRQ(PETSC_ERR_ARG_WRONGSTATE,"MatGetRow not called"); 1054 } 1055 baij->getrowactive = PETSC_FALSE; 1056 PetscFunctionReturn(0); 1057 } 1058 1059 #undef __FUNCT__ 1060 #define __FUNCT__ "MatGetBlockSize_MPISBAIJ" 1061 int MatGetBlockSize_MPISBAIJ(Mat mat,int *bs) 1062 { 1063 Mat_MPISBAIJ *baij = (Mat_MPISBAIJ*)mat->data; 1064 1065 PetscFunctionBegin; 1066 *bs = baij->bs; 1067 PetscFunctionReturn(0); 1068 } 1069 1070 #undef __FUNCT__ 1071 #define __FUNCT__ "MatZeroEntries_MPISBAIJ" 1072 int MatZeroEntries_MPISBAIJ(Mat A) 1073 { 1074 Mat_MPISBAIJ *l = (Mat_MPISBAIJ*)A->data; 1075 int ierr; 1076 1077 PetscFunctionBegin; 1078 ierr = MatZeroEntries(l->A);CHKERRQ(ierr); 1079 ierr = MatZeroEntries(l->B);CHKERRQ(ierr); 1080 PetscFunctionReturn(0); 1081 } 1082 1083 #undef __FUNCT__ 1084 #define __FUNCT__ "MatGetInfo_MPISBAIJ" 1085 int MatGetInfo_MPISBAIJ(Mat matin,MatInfoType flag,MatInfo *info) 1086 { 1087 Mat_MPISBAIJ *a = (Mat_MPISBAIJ*)matin->data; 1088 Mat A = a->A,B = a->B; 1089 int ierr; 1090 PetscReal isend[5],irecv[5]; 1091 1092 PetscFunctionBegin; 1093 info->block_size = (PetscReal)a->bs; 1094 ierr = MatGetInfo(A,MAT_LOCAL,info);CHKERRQ(ierr); 1095 isend[0] = info->nz_used; isend[1] = info->nz_allocated; isend[2] = info->nz_unneeded; 1096 isend[3] = info->memory; isend[4] = info->mallocs; 1097 ierr = MatGetInfo(B,MAT_LOCAL,info);CHKERRQ(ierr); 1098 isend[0] += info->nz_used; isend[1] += info->nz_allocated; isend[2] += info->nz_unneeded; 1099 isend[3] += info->memory; isend[4] += info->mallocs; 1100 if (flag == MAT_LOCAL) { 1101 info->nz_used = isend[0]; 1102 info->nz_allocated = isend[1]; 1103 info->nz_unneeded = isend[2]; 1104 info->memory = isend[3]; 1105 info->mallocs = isend[4]; 1106 } else if (flag == MAT_GLOBAL_MAX) { 1107 ierr = MPI_Allreduce(isend,irecv,5,MPIU_REAL,MPI_MAX,matin->comm);CHKERRQ(ierr); 1108 info->nz_used = irecv[0]; 1109 info->nz_allocated = irecv[1]; 1110 info->nz_unneeded = irecv[2]; 1111 info->memory = irecv[3]; 1112 info->mallocs = irecv[4]; 1113 } else if (flag == MAT_GLOBAL_SUM) { 1114 ierr = MPI_Allreduce(isend,irecv,5,MPIU_REAL,MPI_SUM,matin->comm);CHKERRQ(ierr); 1115 info->nz_used = irecv[0]; 1116 info->nz_allocated = irecv[1]; 1117 info->nz_unneeded = irecv[2]; 1118 info->memory = irecv[3]; 1119 info->mallocs = irecv[4]; 1120 } else { 1121 SETERRQ1(1,"Unknown MatInfoType argument %d",flag); 1122 } 1123 info->rows_global = (PetscReal)A->M; 1124 info->columns_global = (PetscReal)A->N; 1125 info->rows_local = (PetscReal)A->m; 1126 info->columns_local = (PetscReal)A->N; 1127 info->fill_ratio_given = 0; /* no parallel LU/ILU/Cholesky */ 1128 info->fill_ratio_needed = 0; 1129 info->factor_mallocs = 0; 1130 PetscFunctionReturn(0); 1131 } 1132 1133 #undef __FUNCT__ 1134 #define __FUNCT__ "MatSetOption_MPISBAIJ" 1135 int MatSetOption_MPISBAIJ(Mat A,MatOption op) 1136 { 1137 Mat_MPISBAIJ *a = (Mat_MPISBAIJ*)A->data; 1138 int ierr; 1139 1140 PetscFunctionBegin; 1141 switch (op) { 1142 case MAT_NO_NEW_NONZERO_LOCATIONS: 1143 case MAT_YES_NEW_NONZERO_LOCATIONS: 1144 case MAT_COLUMNS_UNSORTED: 1145 case MAT_COLUMNS_SORTED: 1146 case MAT_NEW_NONZERO_ALLOCATION_ERR: 1147 case MAT_KEEP_ZEROED_ROWS: 1148 case MAT_NEW_NONZERO_LOCATION_ERR: 1149 ierr = MatSetOption(a->A,op);CHKERRQ(ierr); 1150 ierr = MatSetOption(a->B,op);CHKERRQ(ierr); 1151 break; 1152 case MAT_ROW_ORIENTED: 1153 a->roworiented = PETSC_TRUE; 1154 ierr = MatSetOption(a->A,op);CHKERRQ(ierr); 1155 ierr = MatSetOption(a->B,op);CHKERRQ(ierr); 1156 break; 1157 case MAT_ROWS_SORTED: 1158 case MAT_ROWS_UNSORTED: 1159 case MAT_YES_NEW_DIAGONALS: 1160 case MAT_USE_SINGLE_PRECISION_SOLVES: 1161 PetscLogInfo(A,"Info:MatSetOption_MPIBAIJ:Option ignored\n"); 1162 break; 1163 case MAT_COLUMN_ORIENTED: 1164 a->roworiented = PETSC_FALSE; 1165 ierr = MatSetOption(a->A,op);CHKERRQ(ierr); 1166 ierr = MatSetOption(a->B,op);CHKERRQ(ierr); 1167 break; 1168 case MAT_IGNORE_OFF_PROC_ENTRIES: 1169 a->donotstash = PETSC_TRUE; 1170 break; 1171 case MAT_NO_NEW_DIAGONALS: 1172 SETERRQ(PETSC_ERR_SUP,"MAT_NO_NEW_DIAGONALS"); 1173 break; 1174 case MAT_USE_HASH_TABLE: 1175 a->ht_flag = PETSC_TRUE; 1176 break; 1177 default: 1178 SETERRQ(PETSC_ERR_SUP,"unknown option"); 1179 break; 1180 } 1181 PetscFunctionReturn(0); 1182 } 1183 1184 #undef __FUNCT__ 1185 #define __FUNCT__ "MatTranspose_MPISBAIJ(" 1186 int MatTranspose_MPISBAIJ(Mat A,Mat *matout) 1187 { 1188 PetscFunctionBegin; 1189 SETERRQ(1,"Matrix is symmetric. MatTranspose() should not be called"); 1190 /* PetscFunctionReturn(0); */ 1191 } 1192 1193 #undef __FUNCT__ 1194 #define __FUNCT__ "MatDiagonalScale_MPISBAIJ" 1195 int MatDiagonalScale_MPISBAIJ(Mat mat,Vec ll,Vec rr) 1196 { 1197 Mat_MPISBAIJ *baij = (Mat_MPISBAIJ*)mat->data; 1198 Mat a = baij->A,b = baij->B; 1199 int ierr,s1,s2,s3; 1200 1201 PetscFunctionBegin; 1202 if (ll != rr) { 1203 SETERRQ(PETSC_ERR_ARG_OUTOFRANGE,"For symmetric format, left and right scaling vectors must be same\n"); 1204 } 1205 ierr = MatGetLocalSize(mat,&s2,&s3);CHKERRQ(ierr); 1206 if (rr) { 1207 ierr = VecGetLocalSize(rr,&s1);CHKERRQ(ierr); 1208 if (s1!=s3) SETERRQ(PETSC_ERR_ARG_SIZ,"right vector non-conforming local size"); 1209 /* Overlap communication with computation. */ 1210 ierr = VecScatterBegin(rr,baij->lvec,INSERT_VALUES,SCATTER_FORWARD,baij->Mvctx);CHKERRQ(ierr); 1211 /*} if (ll) { */ 1212 ierr = VecGetLocalSize(ll,&s1);CHKERRQ(ierr); 1213 if (s1!=s2) SETERRQ(PETSC_ERR_ARG_SIZ,"left vector non-conforming local size"); 1214 ierr = (*b->ops->diagonalscale)(b,ll,PETSC_NULL);CHKERRQ(ierr); 1215 /* } */ 1216 /* scale the diagonal block */ 1217 ierr = (*a->ops->diagonalscale)(a,ll,rr);CHKERRQ(ierr); 1218 1219 /* if (rr) { */ 1220 /* Do a scatter end and then right scale the off-diagonal block */ 1221 ierr = VecScatterEnd(rr,baij->lvec,INSERT_VALUES,SCATTER_FORWARD,baij->Mvctx);CHKERRQ(ierr); 1222 ierr = (*b->ops->diagonalscale)(b,PETSC_NULL,baij->lvec);CHKERRQ(ierr); 1223 } 1224 1225 PetscFunctionReturn(0); 1226 } 1227 1228 #undef __FUNCT__ 1229 #define __FUNCT__ "MatZeroRows_MPISBAIJ" 1230 int MatZeroRows_MPISBAIJ(Mat A,IS is,PetscScalar *diag) 1231 { 1232 Mat_MPISBAIJ *l = (Mat_MPISBAIJ*)A->data; 1233 int i,ierr,N,*rows,*owners = l->rowners,size = l->size; 1234 int *procs,*nprocs,j,idx,nsends,*work,row; 1235 int nmax,*svalues,*starts,*owner,nrecvs,rank = l->rank; 1236 int *rvalues,tag = A->tag,count,base,slen,n,*source; 1237 int *lens,imdex,*lrows,*values,bs=l->bs,rstart_bs=l->rstart_bs; 1238 MPI_Comm comm = A->comm; 1239 MPI_Request *send_waits,*recv_waits; 1240 MPI_Status recv_status,*send_status; 1241 IS istmp; 1242 PetscTruth found; 1243 1244 PetscFunctionBegin; 1245 ierr = ISGetSize(is,&N);CHKERRQ(ierr); 1246 ierr = ISGetIndices(is,&rows);CHKERRQ(ierr); 1247 1248 /* first count number of contributors to each processor */ 1249 ierr = PetscMalloc(2*size*sizeof(int),&nprocs);CHKERRQ(ierr); 1250 ierr = PetscMemzero(nprocs,2*size*sizeof(int));CHKERRQ(ierr); 1251 procs = nprocs + size; 1252 ierr = PetscMalloc((N+1)*sizeof(int),&owner);CHKERRQ(ierr); /* see note*/ 1253 for (i=0; i<N; i++) { 1254 idx = rows[i]; 1255 found = PETSC_FALSE; 1256 for (j=0; j<size; j++) { 1257 if (idx >= owners[j]*bs && idx < owners[j+1]*bs) { 1258 nprocs[j]++; procs[j] = 1; owner[i] = j; found = PETSC_TRUE; break; 1259 } 1260 } 1261 if (!found) SETERRQ(PETSC_ERR_ARG_OUTOFRANGE,"Index out of range"); 1262 } 1263 nsends = 0; for (i=0; i<size; i++) { nsends += procs[i];} 1264 1265 /* inform other processors of number of messages and max length*/ 1266 ierr = PetscMalloc(2*size*sizeof(int),&work);CHKERRQ(ierr); 1267 ierr = MPI_Allreduce(nprocs,work,2*size,MPI_INT,PetscMaxSum_Op,comm);CHKERRQ(ierr); 1268 nmax = work[rank]; 1269 nrecvs = work[size+rank]; 1270 ierr = PetscFree(work);CHKERRQ(ierr); 1271 1272 /* post receives: */ 1273 ierr = PetscMalloc((nrecvs+1)*(nmax+1)*sizeof(int),&rvalues);CHKERRQ(ierr); 1274 ierr = PetscMalloc((nrecvs+1)*sizeof(MPI_Request),&recv_waits);CHKERRQ(ierr); 1275 for (i=0; i<nrecvs; i++) { 1276 ierr = MPI_Irecv(rvalues+nmax*i,nmax,MPI_INT,MPI_ANY_SOURCE,tag,comm,recv_waits+i);CHKERRQ(ierr); 1277 } 1278 1279 /* do sends: 1280 1) starts[i] gives the starting index in svalues for stuff going to 1281 the ith processor 1282 */ 1283 ierr = PetscMalloc((N+1)*sizeof(int),&svalues);CHKERRQ(ierr); 1284 ierr = PetscMalloc((nsends+1)*sizeof(MPI_Request),&send_waits);CHKERRQ(ierr); 1285 ierr = PetscMalloc((size+1)*sizeof(int),&starts);CHKERRQ(ierr); 1286 starts[0] = 0; 1287 for (i=1; i<size; i++) { starts[i] = starts[i-1] + nprocs[i-1];} 1288 for (i=0; i<N; i++) { 1289 svalues[starts[owner[i]]++] = rows[i]; 1290 } 1291 ierr = ISRestoreIndices(is,&rows);CHKERRQ(ierr); 1292 1293 starts[0] = 0; 1294 for (i=1; i<size+1; i++) { starts[i] = starts[i-1] + nprocs[i-1];} 1295 count = 0; 1296 for (i=0; i<size; i++) { 1297 if (procs[i]) { 1298 ierr = MPI_Isend(svalues+starts[i],nprocs[i],MPI_INT,i,tag,comm,send_waits+count++);CHKERRQ(ierr); 1299 } 1300 } 1301 ierr = PetscFree(starts);CHKERRQ(ierr); 1302 1303 base = owners[rank]*bs; 1304 1305 /* wait on receives */ 1306 ierr = PetscMalloc(2*(nrecvs+1)*sizeof(int),&lens);CHKERRQ(ierr); 1307 source = lens + nrecvs; 1308 count = nrecvs; slen = 0; 1309 while (count) { 1310 ierr = MPI_Waitany(nrecvs,recv_waits,&imdex,&recv_status);CHKERRQ(ierr); 1311 /* unpack receives into our local space */ 1312 ierr = MPI_Get_count(&recv_status,MPI_INT,&n);CHKERRQ(ierr); 1313 source[imdex] = recv_status.MPI_SOURCE; 1314 lens[imdex] = n; 1315 slen += n; 1316 count--; 1317 } 1318 ierr = PetscFree(recv_waits);CHKERRQ(ierr); 1319 1320 /* move the data into the send scatter */ 1321 ierr = PetscMalloc((slen+1)*sizeof(int),&lrows);CHKERRQ(ierr); 1322 count = 0; 1323 for (i=0; i<nrecvs; i++) { 1324 values = rvalues + i*nmax; 1325 for (j=0; j<lens[i]; j++) { 1326 lrows[count++] = values[j] - base; 1327 } 1328 } 1329 ierr = PetscFree(rvalues);CHKERRQ(ierr); 1330 ierr = PetscFree(lens);CHKERRQ(ierr); 1331 ierr = PetscFree(owner);CHKERRQ(ierr); 1332 ierr = PetscFree(nprocs);CHKERRQ(ierr); 1333 1334 /* actually zap the local rows */ 1335 ierr = ISCreateGeneral(PETSC_COMM_SELF,slen,lrows,&istmp);CHKERRQ(ierr); 1336 PetscLogObjectParent(A,istmp); 1337 1338 /* 1339 Zero the required rows. If the "diagonal block" of the matrix 1340 is square and the user wishes to set the diagonal we use seperate 1341 code so that MatSetValues() is not called for each diagonal allocating 1342 new memory, thus calling lots of mallocs and slowing things down. 1343 1344 Contributed by: Mathew Knepley 1345 */ 1346 /* must zero l->B before l->A because the (diag) case below may put values into l->B*/ 1347 ierr = MatZeroRows_SeqBAIJ(l->B,istmp,0);CHKERRQ(ierr); 1348 if (diag && (l->A->M == l->A->N)) { 1349 ierr = MatZeroRows_SeqSBAIJ(l->A,istmp,diag);CHKERRQ(ierr); 1350 } else if (diag) { 1351 ierr = MatZeroRows_SeqSBAIJ(l->A,istmp,0);CHKERRQ(ierr); 1352 if (((Mat_SeqSBAIJ*)l->A->data)->nonew) { 1353 SETERRQ(PETSC_ERR_SUP,"MatZeroRows() on rectangular matrices cannot be used with the Mat options \n\ 1354 MAT_NO_NEW_NONZERO_LOCATIONS,MAT_NEW_NONZERO_LOCATION_ERR,MAT_NEW_NONZERO_ALLOCATION_ERR"); 1355 } 1356 for (i=0; i<slen; i++) { 1357 row = lrows[i] + rstart_bs; 1358 ierr = MatSetValues(A,1,&row,1,&row,diag,INSERT_VALUES);CHKERRQ(ierr); 1359 } 1360 ierr = MatAssemblyBegin(A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1361 ierr = MatAssemblyEnd(A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1362 } else { 1363 ierr = MatZeroRows_SeqSBAIJ(l->A,istmp,0);CHKERRQ(ierr); 1364 } 1365 1366 ierr = ISDestroy(istmp);CHKERRQ(ierr); 1367 ierr = PetscFree(lrows);CHKERRQ(ierr); 1368 1369 /* wait on sends */ 1370 if (nsends) { 1371 ierr = PetscMalloc(nsends*sizeof(MPI_Status),&send_status);CHKERRQ(ierr); 1372 ierr = MPI_Waitall(nsends,send_waits,send_status);CHKERRQ(ierr); 1373 ierr = PetscFree(send_status);CHKERRQ(ierr); 1374 } 1375 ierr = PetscFree(send_waits);CHKERRQ(ierr); 1376 ierr = PetscFree(svalues);CHKERRQ(ierr); 1377 1378 PetscFunctionReturn(0); 1379 } 1380 1381 #undef __FUNCT__ 1382 #define __FUNCT__ "MatPrintHelp_MPISBAIJ" 1383 int MatPrintHelp_MPISBAIJ(Mat A) 1384 { 1385 Mat_MPISBAIJ *a = (Mat_MPISBAIJ*)A->data; 1386 MPI_Comm comm = A->comm; 1387 static int called = 0; 1388 int ierr; 1389 1390 PetscFunctionBegin; 1391 if (!a->rank) { 1392 ierr = MatPrintHelp_SeqSBAIJ(a->A);CHKERRQ(ierr); 1393 } 1394 if (called) {PetscFunctionReturn(0);} else called = 1; 1395 ierr = (*PetscHelpPrintf)(comm," Options for MATMPISBAIJ matrix format (the defaults):\n");CHKERRQ(ierr); 1396 ierr = (*PetscHelpPrintf)(comm," -mat_use_hash_table <factor>: Use hashtable for efficient matrix assembly\n");CHKERRQ(ierr); 1397 PetscFunctionReturn(0); 1398 } 1399 1400 #undef __FUNCT__ 1401 #define __FUNCT__ "MatSetUnfactored_MPISBAIJ" 1402 int MatSetUnfactored_MPISBAIJ(Mat A) 1403 { 1404 Mat_MPISBAIJ *a = (Mat_MPISBAIJ*)A->data; 1405 int ierr; 1406 1407 PetscFunctionBegin; 1408 ierr = MatSetUnfactored(a->A);CHKERRQ(ierr); 1409 PetscFunctionReturn(0); 1410 } 1411 1412 static int MatDuplicate_MPISBAIJ(Mat,MatDuplicateOption,Mat *); 1413 1414 #undef __FUNCT__ 1415 #define __FUNCT__ "MatEqual_MPISBAIJ" 1416 int MatEqual_MPISBAIJ(Mat A,Mat B,PetscTruth *flag) 1417 { 1418 Mat_MPISBAIJ *matB = (Mat_MPISBAIJ*)B->data,*matA = (Mat_MPISBAIJ*)A->data; 1419 Mat a,b,c,d; 1420 PetscTruth flg; 1421 int ierr; 1422 1423 PetscFunctionBegin; 1424 ierr = PetscTypeCompare((PetscObject)B,MATMPISBAIJ,&flg);CHKERRQ(ierr); 1425 if (!flg) SETERRQ(PETSC_ERR_ARG_INCOMP,"Matrices must be same type"); 1426 a = matA->A; b = matA->B; 1427 c = matB->A; d = matB->B; 1428 1429 ierr = MatEqual(a,c,&flg);CHKERRQ(ierr); 1430 if (flg == PETSC_TRUE) { 1431 ierr = MatEqual(b,d,&flg);CHKERRQ(ierr); 1432 } 1433 ierr = MPI_Allreduce(&flg,flag,1,MPI_INT,MPI_LAND,A->comm);CHKERRQ(ierr); 1434 PetscFunctionReturn(0); 1435 } 1436 1437 #undef __FUNCT__ 1438 #define __FUNCT__ "MatSetUpPreallocation_MPISBAIJ" 1439 int MatSetUpPreallocation_MPISBAIJ(Mat A) 1440 { 1441 int ierr; 1442 1443 PetscFunctionBegin; 1444 ierr = MatMPISBAIJSetPreallocation(A,1,PETSC_DEFAULT,0,PETSC_DEFAULT,0);CHKERRQ(ierr); 1445 PetscFunctionReturn(0); 1446 } 1447 /* -------------------------------------------------------------------*/ 1448 static struct _MatOps MatOps_Values = { 1449 MatSetValues_MPISBAIJ, 1450 MatGetRow_MPISBAIJ, 1451 MatRestoreRow_MPISBAIJ, 1452 MatMult_MPISBAIJ, 1453 MatMultAdd_MPISBAIJ, 1454 MatMultTranspose_MPISBAIJ, 1455 MatMultTransposeAdd_MPISBAIJ, 1456 0, 1457 0, 1458 0, 1459 0, 1460 0, 1461 0, 1462 MatRelax_MPISBAIJ, 1463 MatTranspose_MPISBAIJ, 1464 MatGetInfo_MPISBAIJ, 1465 MatEqual_MPISBAIJ, 1466 MatGetDiagonal_MPISBAIJ, 1467 MatDiagonalScale_MPISBAIJ, 1468 MatNorm_MPISBAIJ, 1469 MatAssemblyBegin_MPISBAIJ, 1470 MatAssemblyEnd_MPISBAIJ, 1471 0, 1472 MatSetOption_MPISBAIJ, 1473 MatZeroEntries_MPISBAIJ, 1474 MatZeroRows_MPISBAIJ, 1475 0, 1476 0, 1477 0, 1478 0, 1479 MatSetUpPreallocation_MPISBAIJ, 1480 0, 1481 0, 1482 0, 1483 0, 1484 MatDuplicate_MPISBAIJ, 1485 0, 1486 0, 1487 0, 1488 0, 1489 0, 1490 MatGetSubMatrices_MPISBAIJ, 1491 MatIncreaseOverlap_MPISBAIJ, 1492 MatGetValues_MPISBAIJ, 1493 0, 1494 MatPrintHelp_MPISBAIJ, 1495 MatScale_MPISBAIJ, 1496 0, 1497 0, 1498 0, 1499 MatGetBlockSize_MPISBAIJ, 1500 0, 1501 0, 1502 0, 1503 0, 1504 0, 1505 0, 1506 MatSetUnfactored_MPISBAIJ, 1507 0, 1508 MatSetValuesBlocked_MPISBAIJ, 1509 0, 1510 0, 1511 0, 1512 MatGetPetscMaps_Petsc, 1513 0, 1514 0, 1515 0, 1516 0, 1517 0, 1518 0, 1519 MatGetRowMax_MPISBAIJ}; 1520 1521 1522 EXTERN_C_BEGIN 1523 #undef __FUNCT__ 1524 #define __FUNCT__ "MatGetDiagonalBlock_MPISBAIJ" 1525 int MatGetDiagonalBlock_MPISBAIJ(Mat A,PetscTruth *iscopy,MatReuse reuse,Mat *a) 1526 { 1527 PetscFunctionBegin; 1528 *a = ((Mat_MPISBAIJ *)A->data)->A; 1529 *iscopy = PETSC_FALSE; 1530 PetscFunctionReturn(0); 1531 } 1532 EXTERN_C_END 1533 1534 EXTERN_C_BEGIN 1535 #undef __FUNCT__ 1536 #define __FUNCT__ "MatCreate_MPISBAIJ" 1537 int MatCreate_MPISBAIJ(Mat B) 1538 { 1539 Mat_MPISBAIJ *b; 1540 int ierr; 1541 PetscTruth flg; 1542 1543 PetscFunctionBegin; 1544 1545 ierr = PetscNew(Mat_MPISBAIJ,&b);CHKERRQ(ierr); 1546 B->data = (void*)b; 1547 ierr = PetscMemzero(b,sizeof(Mat_MPISBAIJ));CHKERRQ(ierr); 1548 ierr = PetscMemcpy(B->ops,&MatOps_Values,sizeof(struct _MatOps));CHKERRQ(ierr); 1549 1550 B->ops->destroy = MatDestroy_MPISBAIJ; 1551 B->ops->view = MatView_MPISBAIJ; 1552 B->mapping = 0; 1553 B->factor = 0; 1554 B->assembled = PETSC_FALSE; 1555 1556 B->insertmode = NOT_SET_VALUES; 1557 ierr = MPI_Comm_rank(B->comm,&b->rank);CHKERRQ(ierr); 1558 ierr = MPI_Comm_size(B->comm,&b->size);CHKERRQ(ierr); 1559 1560 /* build local table of row and column ownerships */ 1561 ierr = PetscMalloc(3*(b->size+2)*sizeof(int),&b->rowners);CHKERRQ(ierr); 1562 b->cowners = b->rowners + b->size + 2; 1563 b->rowners_bs = b->cowners + b->size + 2; 1564 PetscLogObjectMemory(B,3*(b->size+2)*sizeof(int)+sizeof(struct _p_Mat)+sizeof(Mat_MPISBAIJ)); 1565 1566 /* build cache for off array entries formed */ 1567 ierr = MatStashCreate_Private(B->comm,1,&B->stash);CHKERRQ(ierr); 1568 b->donotstash = PETSC_FALSE; 1569 b->colmap = PETSC_NULL; 1570 b->garray = PETSC_NULL; 1571 b->roworiented = PETSC_TRUE; 1572 1573 #if defined(PETSC_USE_MAT_SINGLE) 1574 /* stuff for MatSetValues_XXX in single precision */ 1575 b->setvalueslen = 0; 1576 b->setvaluescopy = PETSC_NULL; 1577 #endif 1578 1579 /* stuff used in block assembly */ 1580 b->barray = 0; 1581 1582 /* stuff used for matrix vector multiply */ 1583 b->lvec = 0; 1584 b->Mvctx = 0; 1585 1586 /* stuff for MatGetRow() */ 1587 b->rowindices = 0; 1588 b->rowvalues = 0; 1589 b->getrowactive = PETSC_FALSE; 1590 1591 /* hash table stuff */ 1592 b->ht = 0; 1593 b->hd = 0; 1594 b->ht_size = 0; 1595 b->ht_flag = PETSC_FALSE; 1596 b->ht_fact = 0; 1597 b->ht_total_ct = 0; 1598 b->ht_insert_ct = 0; 1599 1600 ierr = PetscOptionsHasName(PETSC_NULL,"-mat_use_hash_table",&flg);CHKERRQ(ierr); 1601 if (flg) { 1602 PetscReal fact = 1.39; 1603 ierr = MatSetOption(B,MAT_USE_HASH_TABLE);CHKERRQ(ierr); 1604 ierr = PetscOptionsGetReal(PETSC_NULL,"-mat_use_hash_table",&fact,PETSC_NULL);CHKERRQ(ierr); 1605 if (fact <= 1.0) fact = 1.39; 1606 ierr = MatMPIBAIJSetHashTableFactor(B,fact);CHKERRQ(ierr); 1607 PetscLogInfo(0,"MatCreateMPISBAIJ:Hash table Factor used %5.2f\n",fact); 1608 } 1609 ierr = PetscObjectComposeFunctionDynamic((PetscObject)B,"MatStoreValues_C", 1610 "MatStoreValues_MPISBAIJ", 1611 MatStoreValues_MPISBAIJ);CHKERRQ(ierr); 1612 ierr = PetscObjectComposeFunctionDynamic((PetscObject)B,"MatRetrieveValues_C", 1613 "MatRetrieveValues_MPISBAIJ", 1614 MatRetrieveValues_MPISBAIJ);CHKERRQ(ierr); 1615 ierr = PetscObjectComposeFunctionDynamic((PetscObject)B,"MatGetDiagonalBlock_C", 1616 "MatGetDiagonalBlock_MPISBAIJ", 1617 MatGetDiagonalBlock_MPISBAIJ);CHKERRQ(ierr); 1618 PetscFunctionReturn(0); 1619 } 1620 EXTERN_C_END 1621 1622 #undef __FUNCT__ 1623 #define __FUNCT__ "MatMPISBAIJSetPreallocation" 1624 /*@C 1625 MatMPISBAIJSetPreallocation - For good matrix assembly performance 1626 the user should preallocate the matrix storage by setting the parameters 1627 d_nz (or d_nnz) and o_nz (or o_nnz). By setting these parameters accurately, 1628 performance can be increased by more than a factor of 50. 1629 1630 Collective on Mat 1631 1632 Input Parameters: 1633 + A - the matrix 1634 . bs - size of blockk 1635 . d_nz - number of block nonzeros per block row in diagonal portion of local 1636 submatrix (same for all local rows) 1637 . d_nnz - array containing the number of block nonzeros in the various block rows 1638 of the in diagonal portion of the local (possibly different for each block 1639 row) or PETSC_NULL. You must leave room for the diagonal entry even if it is zero. 1640 . o_nz - number of block nonzeros per block row in the off-diagonal portion of local 1641 submatrix (same for all local rows). 1642 - o_nnz - array containing the number of nonzeros in the various block rows of the 1643 off-diagonal portion of the local submatrix (possibly different for 1644 each block row) or PETSC_NULL. 1645 1646 1647 Options Database Keys: 1648 . -mat_no_unroll - uses code that does not unroll the loops in the 1649 block calculations (much slower) 1650 . -mat_block_size - size of the blocks to use 1651 1652 Notes: 1653 1654 If PETSC_DECIDE or PETSC_DETERMINE is used for a particular argument on one processor 1655 than it must be used on all processors that share the object for that argument. 1656 1657 Storage Information: 1658 For a square global matrix we define each processor's diagonal portion 1659 to be its local rows and the corresponding columns (a square submatrix); 1660 each processor's off-diagonal portion encompasses the remainder of the 1661 local matrix (a rectangular submatrix). 1662 1663 The user can specify preallocated storage for the diagonal part of 1664 the local submatrix with either d_nz or d_nnz (not both). Set 1665 d_nz=PETSC_DEFAULT and d_nnz=PETSC_NULL for PETSc to control dynamic 1666 memory allocation. Likewise, specify preallocated storage for the 1667 off-diagonal part of the local submatrix with o_nz or o_nnz (not both). 1668 1669 Consider a processor that owns rows 3, 4 and 5 of a parallel matrix. In 1670 the figure below we depict these three local rows and all columns (0-11). 1671 1672 .vb 1673 0 1 2 3 4 5 6 7 8 9 10 11 1674 ------------------- 1675 row 3 | o o o d d d o o o o o o 1676 row 4 | o o o d d d o o o o o o 1677 row 5 | o o o d d d o o o o o o 1678 ------------------- 1679 .ve 1680 1681 Thus, any entries in the d locations are stored in the d (diagonal) 1682 submatrix, and any entries in the o locations are stored in the 1683 o (off-diagonal) submatrix. Note that the d and the o submatrices are 1684 stored simply in the MATSEQBAIJ format for compressed row storage. 1685 1686 Now d_nz should indicate the number of block nonzeros per row in the d matrix, 1687 and o_nz should indicate the number of block nonzeros per row in the o matrix. 1688 In general, for PDE problems in which most nonzeros are near the diagonal, 1689 one expects d_nz >> o_nz. For large problems you MUST preallocate memory 1690 or you will get TERRIBLE performance; see the users' manual chapter on 1691 matrices. 1692 1693 Level: intermediate 1694 1695 .keywords: matrix, block, aij, compressed row, sparse, parallel 1696 1697 .seealso: MatCreate(), MatCreateSeqSBAIJ(), MatSetValues(), MatCreateMPIBAIJ() 1698 @*/ 1699 1700 int MatMPISBAIJSetPreallocation(Mat B,int bs,int d_nz,int *d_nnz,int o_nz,int *o_nnz) 1701 { 1702 Mat_MPISBAIJ *b; 1703 int ierr,i,mbs,Mbs; 1704 PetscTruth flg2; 1705 1706 PetscFunctionBegin; 1707 ierr = PetscTypeCompare((PetscObject)B,MATMPISBAIJ,&flg2);CHKERRQ(ierr); 1708 if (!flg2) PetscFunctionReturn(0); 1709 1710 ierr = PetscOptionsGetInt(PETSC_NULL,"-mat_block_size",&bs,PETSC_NULL);CHKERRQ(ierr); 1711 1712 if (bs < 1) SETERRQ(PETSC_ERR_ARG_OUTOFRANGE,"Invalid block size specified, must be positive"); 1713 if (d_nz == PETSC_DECIDE || d_nz == PETSC_DEFAULT) d_nz = 3; 1714 if (o_nz == PETSC_DECIDE || o_nz == PETSC_DEFAULT) o_nz = 1; 1715 if (d_nz < 0) SETERRQ1(PETSC_ERR_ARG_OUTOFRANGE,"d_nz cannot be less than 0: value %d",d_nz); 1716 if (o_nz < 0) SETERRQ1(PETSC_ERR_ARG_OUTOFRANGE,"o_nz cannot be less than 0: value %d",o_nz); 1717 if (d_nnz) { 1718 for (i=0; i<B->m/bs; i++) { 1719 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]); 1720 } 1721 } 1722 if (o_nnz) { 1723 for (i=0; i<B->m/bs; i++) { 1724 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]); 1725 } 1726 } 1727 B->preallocated = PETSC_TRUE; 1728 ierr = PetscSplitOwnershipBlock(B->comm,bs,&B->m,&B->M);CHKERRQ(ierr); 1729 ierr = PetscSplitOwnershipBlock(B->comm,bs,&B->n,&B->N);CHKERRQ(ierr); 1730 ierr = PetscMapCreateMPI(B->comm,B->m,B->M,&B->rmap);CHKERRQ(ierr); 1731 ierr = PetscMapCreateMPI(B->comm,B->m,B->M,&B->cmap);CHKERRQ(ierr); 1732 1733 b = (Mat_MPISBAIJ*)B->data; 1734 mbs = B->m/bs; 1735 Mbs = B->M/bs; 1736 if (mbs*bs != B->m) { 1737 SETERRQ2(PETSC_ERR_ARG_SIZ,"No of local rows %d must be divisible by blocksize %d",B->m,bs); 1738 } 1739 1740 b->bs = bs; 1741 b->bs2 = bs*bs; 1742 b->mbs = mbs; 1743 b->nbs = mbs; 1744 b->Mbs = Mbs; 1745 b->Nbs = Mbs; 1746 1747 ierr = MPI_Allgather(&b->mbs,1,MPI_INT,b->rowners+1,1,MPI_INT,B->comm);CHKERRQ(ierr); 1748 b->rowners[0] = 0; 1749 for (i=2; i<=b->size; i++) { 1750 b->rowners[i] += b->rowners[i-1]; 1751 } 1752 b->rstart = b->rowners[b->rank]; 1753 b->rend = b->rowners[b->rank+1]; 1754 b->cstart = b->rstart; 1755 b->cend = b->rend; 1756 for (i=0; i<=b->size; i++) { 1757 b->rowners_bs[i] = b->rowners[i]*bs; 1758 } 1759 b->rstart_bs = b-> rstart*bs; 1760 b->rend_bs = b->rend*bs; 1761 1762 b->cstart_bs = b->cstart*bs; 1763 b->cend_bs = b->cend*bs; 1764 1765 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 ierr = MatCreateSeqBAIJ(PETSC_COMM_SELF,bs,B->m,B->M,o_nz,o_nnz,&b->B);CHKERRQ(ierr); 1769 PetscLogObjectParent(B,b->B); 1770 1771 /* build cache for off array entries formed */ 1772 ierr = MatStashCreate_Private(B->comm,bs,&B->bstash);CHKERRQ(ierr); 1773 1774 PetscFunctionReturn(0); 1775 } 1776 1777 #undef __FUNCT__ 1778 #define __FUNCT__ "MatCreateMPISBAIJ" 1779 /*@C 1780 MatCreateMPISBAIJ - Creates a sparse parallel matrix in symmetric block AIJ format 1781 (block compressed row). For good matrix assembly performance 1782 the user should preallocate the matrix storage by setting the parameters 1783 d_nz (or d_nnz) and o_nz (or o_nnz). By setting these parameters accurately, 1784 performance can be increased by more than a factor of 50. 1785 1786 Collective on MPI_Comm 1787 1788 Input Parameters: 1789 + comm - MPI communicator 1790 . bs - size of blockk 1791 . m - number of local rows (or PETSC_DECIDE to have calculated if M is given) 1792 This value should be the same as the local size used in creating the 1793 y vector for the matrix-vector product y = Ax. 1794 . n - number of local columns (or PETSC_DECIDE to have calculated if N is given) 1795 This value should be the same as the local size used in creating the 1796 x vector for the matrix-vector product y = Ax. 1797 . M - number of global rows (or PETSC_DETERMINE to have calculated if m is given) 1798 . N - number of global columns (or PETSC_DETERMINE to have calculated if n is given) 1799 . d_nz - number of block nonzeros per block row in diagonal portion of local 1800 submatrix (same for all local rows) 1801 . d_nnz - array containing the number of block nonzeros in the various block rows 1802 of the in diagonal portion of the local (possibly different for each block 1803 row) or PETSC_NULL. You must leave room for the diagonal entry even if it is zero. 1804 . o_nz - number of block nonzeros per block row in the off-diagonal portion of local 1805 submatrix (same for all local rows). 1806 - o_nnz - array containing the number of nonzeros in the various block rows of the 1807 off-diagonal portion of the local submatrix (possibly different for 1808 each block row) or PETSC_NULL. 1809 1810 Output Parameter: 1811 . A - the matrix 1812 1813 Options Database Keys: 1814 . -mat_no_unroll - uses code that does not unroll the loops in the 1815 block calculations (much slower) 1816 . -mat_block_size - size of the blocks to use 1817 . -mat_mpi - use the parallel matrix data structures even on one processor 1818 (defaults to using SeqBAIJ format on one processor) 1819 1820 Notes: 1821 The user MUST specify either the local or global matrix dimensions 1822 (possibly both). 1823 1824 If PETSC_DECIDE or PETSC_DETERMINE is used for a particular argument on one processor 1825 than it must be used on all processors that share the object for that argument. 1826 1827 Storage Information: 1828 For a square global matrix we define each processor's diagonal portion 1829 to be its local rows and the corresponding columns (a square submatrix); 1830 each processor's off-diagonal portion encompasses the remainder of the 1831 local matrix (a rectangular submatrix). 1832 1833 The user can specify preallocated storage for the diagonal part of 1834 the local submatrix with either d_nz or d_nnz (not both). Set 1835 d_nz=PETSC_DEFAULT and d_nnz=PETSC_NULL for PETSc to control dynamic 1836 memory allocation. Likewise, specify preallocated storage for the 1837 off-diagonal part of the local submatrix with o_nz or o_nnz (not both). 1838 1839 Consider a processor that owns rows 3, 4 and 5 of a parallel matrix. In 1840 the figure below we depict these three local rows and all columns (0-11). 1841 1842 .vb 1843 0 1 2 3 4 5 6 7 8 9 10 11 1844 ------------------- 1845 row 3 | o o o d d d o o o o o o 1846 row 4 | o o o d d d o o o o o o 1847 row 5 | o o o d d d o o o o o o 1848 ------------------- 1849 .ve 1850 1851 Thus, any entries in the d locations are stored in the d (diagonal) 1852 submatrix, and any entries in the o locations are stored in the 1853 o (off-diagonal) submatrix. Note that the d and the o submatrices are 1854 stored simply in the MATSEQBAIJ format for compressed row storage. 1855 1856 Now d_nz should indicate the number of block nonzeros per row in the d matrix, 1857 and o_nz should indicate the number of block nonzeros per row in the o matrix. 1858 In general, for PDE problems in which most nonzeros are near the diagonal, 1859 one expects d_nz >> o_nz. For large problems you MUST preallocate memory 1860 or you will get TERRIBLE performance; see the users' manual chapter on 1861 matrices. 1862 1863 Level: intermediate 1864 1865 .keywords: matrix, block, aij, compressed row, sparse, parallel 1866 1867 .seealso: MatCreate(), MatCreateSeqSBAIJ(), MatSetValues(), MatCreateMPIBAIJ() 1868 @*/ 1869 1870 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) 1871 { 1872 int ierr,size; 1873 1874 PetscFunctionBegin; 1875 ierr = MatCreate(comm,m,n,M,N,A);CHKERRQ(ierr); 1876 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 1877 if (size > 1) { 1878 ierr = MatSetType(*A,MATMPISBAIJ);CHKERRQ(ierr); 1879 ierr = MatMPISBAIJSetPreallocation(*A,bs,d_nz,d_nnz,o_nz,o_nnz);CHKERRQ(ierr); 1880 } else { 1881 ierr = MatSetType(*A,MATSEQSBAIJ);CHKERRQ(ierr); 1882 ierr = MatSeqSBAIJSetPreallocation(*A,bs,d_nz,d_nnz);CHKERRQ(ierr); 1883 } 1884 PetscFunctionReturn(0); 1885 } 1886 1887 1888 #undef __FUNCT__ 1889 #define __FUNCT__ "MatDuplicate_MPISBAIJ" 1890 static int MatDuplicate_MPISBAIJ(Mat matin,MatDuplicateOption cpvalues,Mat *newmat) 1891 { 1892 Mat mat; 1893 Mat_MPISBAIJ *a,*oldmat = (Mat_MPISBAIJ*)matin->data; 1894 int ierr,len=0; 1895 1896 PetscFunctionBegin; 1897 *newmat = 0; 1898 ierr = MatCreate(matin->comm,matin->m,matin->n,matin->M,matin->N,&mat);CHKERRQ(ierr); 1899 ierr = MatSetType(mat,MATMPISBAIJ);CHKERRQ(ierr); 1900 mat->preallocated = PETSC_TRUE; 1901 a = (Mat_MPISBAIJ*)mat->data; 1902 a->bs = oldmat->bs; 1903 a->bs2 = oldmat->bs2; 1904 a->mbs = oldmat->mbs; 1905 a->nbs = oldmat->nbs; 1906 a->Mbs = oldmat->Mbs; 1907 a->Nbs = oldmat->Nbs; 1908 1909 a->rstart = oldmat->rstart; 1910 a->rend = oldmat->rend; 1911 a->cstart = oldmat->cstart; 1912 a->cend = oldmat->cend; 1913 a->size = oldmat->size; 1914 a->rank = oldmat->rank; 1915 a->donotstash = oldmat->donotstash; 1916 a->roworiented = oldmat->roworiented; 1917 a->rowindices = 0; 1918 a->rowvalues = 0; 1919 a->getrowactive = PETSC_FALSE; 1920 a->barray = 0; 1921 a->rstart_bs = oldmat->rstart_bs; 1922 a->rend_bs = oldmat->rend_bs; 1923 a->cstart_bs = oldmat->cstart_bs; 1924 a->cend_bs = oldmat->cend_bs; 1925 1926 /* hash table stuff */ 1927 a->ht = 0; 1928 a->hd = 0; 1929 a->ht_size = 0; 1930 a->ht_flag = oldmat->ht_flag; 1931 a->ht_fact = oldmat->ht_fact; 1932 a->ht_total_ct = 0; 1933 a->ht_insert_ct = 0; 1934 1935 ierr = PetscMalloc(3*(a->size+2)*sizeof(int),&a->rowners);CHKERRQ(ierr); 1936 PetscLogObjectMemory(mat,3*(a->size+2)*sizeof(int)+sizeof(struct _p_Mat)+sizeof(Mat_MPISBAIJ)); 1937 a->cowners = a->rowners + a->size + 2; 1938 a->rowners_bs = a->cowners + a->size + 2; 1939 ierr = PetscMemcpy(a->rowners,oldmat->rowners,3*(a->size+2)*sizeof(int));CHKERRQ(ierr); 1940 ierr = MatStashCreate_Private(matin->comm,1,&mat->stash);CHKERRQ(ierr); 1941 ierr = MatStashCreate_Private(matin->comm,oldmat->bs,&mat->bstash);CHKERRQ(ierr); 1942 if (oldmat->colmap) { 1943 #if defined (PETSC_USE_CTABLE) 1944 ierr = PetscTableCreateCopy(oldmat->colmap,&a->colmap);CHKERRQ(ierr); 1945 #else 1946 ierr = PetscMalloc((a->Nbs)*sizeof(int),&a->colmap);CHKERRQ(ierr); 1947 PetscLogObjectMemory(mat,(a->Nbs)*sizeof(int)); 1948 ierr = PetscMemcpy(a->colmap,oldmat->colmap,(a->Nbs)*sizeof(int));CHKERRQ(ierr); 1949 #endif 1950 } else a->colmap = 0; 1951 if (oldmat->garray && (len = ((Mat_SeqBAIJ*)(oldmat->B->data))->nbs)) { 1952 ierr = PetscMalloc(len*sizeof(int),&a->garray);CHKERRQ(ierr); 1953 PetscLogObjectMemory(mat,len*sizeof(int)); 1954 ierr = PetscMemcpy(a->garray,oldmat->garray,len*sizeof(int));CHKERRQ(ierr); 1955 } else a->garray = 0; 1956 1957 ierr = VecDuplicate(oldmat->lvec,&a->lvec);CHKERRQ(ierr); 1958 PetscLogObjectParent(mat,a->lvec); 1959 ierr = VecScatterCopy(oldmat->Mvctx,&a->Mvctx);CHKERRQ(ierr); 1960 1961 PetscLogObjectParent(mat,a->Mvctx); 1962 ierr = MatDuplicate(oldmat->A,cpvalues,&a->A);CHKERRQ(ierr); 1963 PetscLogObjectParent(mat,a->A); 1964 ierr = MatDuplicate(oldmat->B,cpvalues,&a->B);CHKERRQ(ierr); 1965 PetscLogObjectParent(mat,a->B); 1966 ierr = PetscFListDuplicate(mat->qlist,&matin->qlist);CHKERRQ(ierr); 1967 *newmat = mat; 1968 PetscFunctionReturn(0); 1969 } 1970 1971 #include "petscsys.h" 1972 1973 EXTERN_C_BEGIN 1974 #undef __FUNCT__ 1975 #define __FUNCT__ "MatLoad_MPISBAIJ" 1976 int MatLoad_MPISBAIJ(PetscViewer viewer,MatType type,Mat *newmat) 1977 { 1978 Mat A; 1979 int i,nz,ierr,j,rstart,rend,fd; 1980 PetscScalar *vals,*buf; 1981 MPI_Comm comm = ((PetscObject)viewer)->comm; 1982 MPI_Status status; 1983 int header[4],rank,size,*rowlengths = 0,M,N,m,*rowners,*browners,maxnz,*cols; 1984 int *locrowlens,*sndcounts = 0,*procsnz = 0,jj,*mycols,*ibuf; 1985 int tag = ((PetscObject)viewer)->tag,bs=1,Mbs,mbs,extra_rows; 1986 int *dlens,*odlens,*mask,*masked1,*masked2,rowcount,odcount; 1987 int dcount,kmax,k,nzcount,tmp; 1988 1989 PetscFunctionBegin; 1990 ierr = PetscOptionsGetInt(PETSC_NULL,"-matload_block_size",&bs,PETSC_NULL);CHKERRQ(ierr); 1991 1992 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 1993 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 1994 if (!rank) { 1995 ierr = PetscViewerBinaryGetDescriptor(viewer,&fd);CHKERRQ(ierr); 1996 ierr = PetscBinaryRead(fd,(char *)header,4,PETSC_INT);CHKERRQ(ierr); 1997 if (header[0] != MAT_FILE_COOKIE) SETERRQ(PETSC_ERR_FILE_UNEXPECTED,"not matrix object"); 1998 if (header[3] < 0) { 1999 SETERRQ(PETSC_ERR_FILE_UNEXPECTED,"Matrix stored in special format, cannot load as MPISBAIJ"); 2000 } 2001 } 2002 2003 ierr = MPI_Bcast(header+1,3,MPI_INT,0,comm);CHKERRQ(ierr); 2004 M = header[1]; N = header[2]; 2005 2006 if (M != N) SETERRQ(PETSC_ERR_SUP,"Can only do square matrices"); 2007 2008 /* 2009 This code adds extra rows to make sure the number of rows is 2010 divisible by the blocksize 2011 */ 2012 Mbs = M/bs; 2013 extra_rows = bs - M + bs*(Mbs); 2014 if (extra_rows == bs) extra_rows = 0; 2015 else Mbs++; 2016 if (extra_rows &&!rank) { 2017 PetscLogInfo(0,"MatLoad_MPISBAIJ:Padding loaded matrix to match blocksize\n"); 2018 } 2019 2020 /* determine ownership of all rows */ 2021 mbs = Mbs/size + ((Mbs % size) > rank); 2022 m = mbs*bs; 2023 ierr = PetscMalloc(2*(size+2)*sizeof(int),&rowners);CHKERRQ(ierr); 2024 browners = rowners + size + 1; 2025 ierr = MPI_Allgather(&mbs,1,MPI_INT,rowners+1,1,MPI_INT,comm);CHKERRQ(ierr); 2026 rowners[0] = 0; 2027 for (i=2; i<=size; i++) rowners[i] += rowners[i-1]; 2028 for (i=0; i<=size; i++) browners[i] = rowners[i]*bs; 2029 rstart = rowners[rank]; 2030 rend = rowners[rank+1]; 2031 2032 /* distribute row lengths to all processors */ 2033 ierr = PetscMalloc((rend-rstart)*bs*sizeof(int),&locrowlens);CHKERRQ(ierr); 2034 if (!rank) { 2035 ierr = PetscMalloc((M+extra_rows)*sizeof(int),&rowlengths);CHKERRQ(ierr); 2036 ierr = PetscBinaryRead(fd,rowlengths,M,PETSC_INT);CHKERRQ(ierr); 2037 for (i=0; i<extra_rows; i++) rowlengths[M+i] = 1; 2038 ierr = PetscMalloc(size*sizeof(int),&sndcounts);CHKERRQ(ierr); 2039 for (i=0; i<size; i++) sndcounts[i] = browners[i+1] - browners[i]; 2040 ierr = MPI_Scatterv(rowlengths,sndcounts,browners,MPI_INT,locrowlens,(rend-rstart)*bs,MPI_INT,0,comm);CHKERRQ(ierr); 2041 ierr = PetscFree(sndcounts);CHKERRQ(ierr); 2042 } else { 2043 ierr = MPI_Scatterv(0,0,0,MPI_INT,locrowlens,(rend-rstart)*bs,MPI_INT,0,comm);CHKERRQ(ierr); 2044 } 2045 2046 if (!rank) { /* procs[0] */ 2047 /* calculate the number of nonzeros on each processor */ 2048 ierr = PetscMalloc(size*sizeof(int),&procsnz);CHKERRQ(ierr); 2049 ierr = PetscMemzero(procsnz,size*sizeof(int));CHKERRQ(ierr); 2050 for (i=0; i<size; i++) { 2051 for (j=rowners[i]*bs; j< rowners[i+1]*bs; j++) { 2052 procsnz[i] += rowlengths[j]; 2053 } 2054 } 2055 ierr = PetscFree(rowlengths);CHKERRQ(ierr); 2056 2057 /* determine max buffer needed and allocate it */ 2058 maxnz = 0; 2059 for (i=0; i<size; i++) { 2060 maxnz = PetscMax(maxnz,procsnz[i]); 2061 } 2062 ierr = PetscMalloc(maxnz*sizeof(int),&cols);CHKERRQ(ierr); 2063 2064 /* read in my part of the matrix column indices */ 2065 nz = procsnz[0]; 2066 ierr = PetscMalloc(nz*sizeof(int),&ibuf);CHKERRQ(ierr); 2067 mycols = ibuf; 2068 if (size == 1) nz -= extra_rows; 2069 ierr = PetscBinaryRead(fd,mycols,nz,PETSC_INT);CHKERRQ(ierr); 2070 if (size == 1) for (i=0; i< extra_rows; i++) { mycols[nz+i] = M+i; } 2071 2072 /* read in every ones (except the last) and ship off */ 2073 for (i=1; i<size-1; i++) { 2074 nz = procsnz[i]; 2075 ierr = PetscBinaryRead(fd,cols,nz,PETSC_INT);CHKERRQ(ierr); 2076 ierr = MPI_Send(cols,nz,MPI_INT,i,tag,comm);CHKERRQ(ierr); 2077 } 2078 /* read in the stuff for the last proc */ 2079 if (size != 1) { 2080 nz = procsnz[size-1] - extra_rows; /* the extra rows are not on the disk */ 2081 ierr = PetscBinaryRead(fd,cols,nz,PETSC_INT);CHKERRQ(ierr); 2082 for (i=0; i<extra_rows; i++) cols[nz+i] = M+i; 2083 ierr = MPI_Send(cols,nz+extra_rows,MPI_INT,size-1,tag,comm);CHKERRQ(ierr); 2084 } 2085 ierr = PetscFree(cols);CHKERRQ(ierr); 2086 } else { /* procs[i], i>0 */ 2087 /* determine buffer space needed for message */ 2088 nz = 0; 2089 for (i=0; i<m; i++) { 2090 nz += locrowlens[i]; 2091 } 2092 ierr = PetscMalloc(nz*sizeof(int),&ibuf);CHKERRQ(ierr); 2093 mycols = ibuf; 2094 /* receive message of column indices*/ 2095 ierr = MPI_Recv(mycols,nz,MPI_INT,0,tag,comm,&status);CHKERRQ(ierr); 2096 ierr = MPI_Get_count(&status,MPI_INT,&maxnz);CHKERRQ(ierr); 2097 if (maxnz != nz) SETERRQ(PETSC_ERR_FILE_UNEXPECTED,"something is wrong with file"); 2098 } 2099 2100 /* loop over local rows, determining number of off diagonal entries */ 2101 ierr = PetscMalloc(2*(rend-rstart+1)*sizeof(int),&dlens);CHKERRQ(ierr); 2102 odlens = dlens + (rend-rstart); 2103 ierr = PetscMalloc(3*Mbs*sizeof(int),&mask);CHKERRQ(ierr); 2104 ierr = PetscMemzero(mask,3*Mbs*sizeof(int));CHKERRQ(ierr); 2105 masked1 = mask + Mbs; 2106 masked2 = masked1 + Mbs; 2107 rowcount = 0; nzcount = 0; 2108 for (i=0; i<mbs; i++) { 2109 dcount = 0; 2110 odcount = 0; 2111 for (j=0; j<bs; j++) { 2112 kmax = locrowlens[rowcount]; 2113 for (k=0; k<kmax; k++) { 2114 tmp = mycols[nzcount++]/bs; /* block col. index */ 2115 if (!mask[tmp]) { 2116 mask[tmp] = 1; 2117 if (tmp < rstart || tmp >= rend) masked2[odcount++] = tmp; /* entry in off-diag portion */ 2118 else masked1[dcount++] = tmp; /* entry in diag portion */ 2119 } 2120 } 2121 rowcount++; 2122 } 2123 2124 dlens[i] = dcount; /* d_nzz[i] */ 2125 odlens[i] = odcount; /* o_nzz[i] */ 2126 2127 /* zero out the mask elements we set */ 2128 for (j=0; j<dcount; j++) mask[masked1[j]] = 0; 2129 for (j=0; j<odcount; j++) mask[masked2[j]] = 0; 2130 } 2131 2132 /* create our matrix */ 2133 ierr = MatCreateMPISBAIJ(comm,bs,m,m,PETSC_DETERMINE,PETSC_DETERMINE,0,dlens,0,odlens,newmat); 2134 CHKERRQ(ierr); 2135 A = *newmat; 2136 ierr = MatSetOption(A,MAT_COLUMNS_SORTED);CHKERRQ(ierr); 2137 2138 if (!rank) { 2139 ierr = PetscMalloc(maxnz*sizeof(PetscScalar),&buf);CHKERRQ(ierr); 2140 /* read in my part of the matrix numerical values */ 2141 nz = procsnz[0]; 2142 vals = buf; 2143 mycols = ibuf; 2144 if (size == 1) nz -= extra_rows; 2145 ierr = PetscBinaryRead(fd,vals,nz,PETSC_SCALAR);CHKERRQ(ierr); 2146 if (size == 1) for (i=0; i< extra_rows; i++) { vals[nz+i] = 1.0; } 2147 2148 /* insert into matrix */ 2149 jj = rstart*bs; 2150 for (i=0; i<m; i++) { 2151 ierr = MatSetValues(A,1,&jj,locrowlens[i],mycols,vals,INSERT_VALUES);CHKERRQ(ierr); 2152 mycols += locrowlens[i]; 2153 vals += locrowlens[i]; 2154 jj++; 2155 } 2156 2157 /* read in other processors (except the last one) and ship out */ 2158 for (i=1; i<size-1; i++) { 2159 nz = procsnz[i]; 2160 vals = buf; 2161 ierr = PetscBinaryRead(fd,vals,nz,PETSC_SCALAR);CHKERRQ(ierr); 2162 ierr = MPI_Send(vals,nz,MPIU_SCALAR,i,A->tag,comm);CHKERRQ(ierr); 2163 } 2164 /* the last proc */ 2165 if (size != 1){ 2166 nz = procsnz[i] - extra_rows; 2167 vals = buf; 2168 ierr = PetscBinaryRead(fd,vals,nz,PETSC_SCALAR);CHKERRQ(ierr); 2169 for (i=0; i<extra_rows; i++) vals[nz+i] = 1.0; 2170 ierr = MPI_Send(vals,nz+extra_rows,MPIU_SCALAR,size-1,A->tag,comm);CHKERRQ(ierr); 2171 } 2172 ierr = PetscFree(procsnz);CHKERRQ(ierr); 2173 2174 } else { 2175 /* receive numeric values */ 2176 ierr = PetscMalloc(nz*sizeof(PetscScalar),&buf);CHKERRQ(ierr); 2177 2178 /* receive message of values*/ 2179 vals = buf; 2180 mycols = ibuf; 2181 ierr = MPI_Recv(vals,nz,MPIU_SCALAR,0,A->tag,comm,&status);CHKERRQ(ierr); 2182 ierr = MPI_Get_count(&status,MPIU_SCALAR,&maxnz);CHKERRQ(ierr); 2183 if (maxnz != nz) SETERRQ(PETSC_ERR_FILE_UNEXPECTED,"something is wrong with file"); 2184 2185 /* insert into matrix */ 2186 jj = rstart*bs; 2187 for (i=0; i<m; i++) { 2188 ierr = MatSetValues_MPISBAIJ(A,1,&jj,locrowlens[i],mycols,vals,INSERT_VALUES);CHKERRQ(ierr); 2189 mycols += locrowlens[i]; 2190 vals += locrowlens[i]; 2191 jj++; 2192 } 2193 } 2194 2195 ierr = PetscFree(locrowlens);CHKERRQ(ierr); 2196 ierr = PetscFree(buf);CHKERRQ(ierr); 2197 ierr = PetscFree(ibuf);CHKERRQ(ierr); 2198 ierr = PetscFree(rowners);CHKERRQ(ierr); 2199 ierr = PetscFree(dlens);CHKERRQ(ierr); 2200 ierr = PetscFree(mask);CHKERRQ(ierr); 2201 ierr = MatAssemblyBegin(A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2202 ierr = MatAssemblyEnd(A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2203 PetscFunctionReturn(0); 2204 } 2205 EXTERN_C_END 2206 2207 #undef __FUNCT__ 2208 #define __FUNCT__ "MatMPISBAIJSetHashTableFactor" 2209 /*@ 2210 MatMPISBAIJSetHashTableFactor - Sets the factor required to compute the size of the HashTable. 2211 2212 Input Parameters: 2213 . mat - the matrix 2214 . fact - factor 2215 2216 Collective on Mat 2217 2218 Level: advanced 2219 2220 Notes: 2221 This can also be set by the command line option: -mat_use_hash_table fact 2222 2223 .keywords: matrix, hashtable, factor, HT 2224 2225 .seealso: MatSetOption() 2226 @*/ 2227 int MatMPISBAIJSetHashTableFactor(Mat mat,PetscReal fact) 2228 { 2229 PetscFunctionBegin; 2230 SETERRQ(1,"Function not yet written for SBAIJ format"); 2231 /* PetscFunctionReturn(0); */ 2232 } 2233 2234 #undef __FUNCT__ 2235 #define __FUNCT__ "MatGetRowMax_MPISBAIJ" 2236 int MatGetRowMax_MPISBAIJ(Mat A,Vec v) 2237 { 2238 Mat_MPISBAIJ *a = (Mat_MPISBAIJ*)A->data; 2239 Mat_SeqBAIJ *b = (Mat_SeqBAIJ*)(a->B)->data; 2240 PetscReal atmp; 2241 PetscReal *work,*svalues,*rvalues; 2242 int ierr,i,bs,mbs,*bi,*bj,brow,j,ncols,krow,kcol,col,row,Mbs,bcol; 2243 int rank,size,*rowners_bs,dest,count,source; 2244 PetscScalar *va; 2245 MatScalar *ba; 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,MPIU_REAL,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,MPIU_REAL,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 2340 #undef __FUNCT__ 2341 #define __FUNCT__ "MatRelax_MPISBAIJ" 2342 int MatRelax_MPISBAIJ(Mat matin,Vec bb,PetscReal omega,MatSORType flag,PetscReal fshift,int its,int lits,Vec xx) 2343 { 2344 Mat_MPISBAIJ *mat = (Mat_MPISBAIJ*)matin->data; 2345 int ierr; 2346 PetscScalar mone=-1.0; 2347 Vec lvec1,bb1; 2348 2349 PetscFunctionBegin; 2350 if (its <= 0 || lits <= 0) SETERRQ2(PETSC_ERR_ARG_WRONG,"Relaxation requires global its %d and local its %d both positive",its,lits); 2351 if (mat->bs > 1) 2352 SETERRQ(PETSC_ERR_SUP,"SSOR for block size > 1 is not yet implemented"); 2353 2354 if ((flag & SOR_LOCAL_SYMMETRIC_SWEEP) == SOR_LOCAL_SYMMETRIC_SWEEP){ 2355 if ( flag & SOR_ZERO_INITIAL_GUESS ) { 2356 ierr = (*mat->A->ops->relax)(mat->A,bb,omega,flag,fshift,lits,PETSC_NULL,xx);CHKERRQ(ierr); 2357 its--; 2358 } 2359 2360 ierr = VecDuplicate(mat->lvec,&lvec1);CHKERRQ(ierr); 2361 ierr = VecDuplicate(bb,&bb1);CHKERRQ(ierr); 2362 while (its--){ 2363 ierr = VecScatterBegin(xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD,mat->Mvctx);CHKERRQ(ierr); 2364 2365 /* lower diagonal part: bb1 = bb - B^T*xx */ 2366 ierr = (*mat->B->ops->multtranspose)(mat->B,xx,lvec1);CHKERRQ(ierr); 2367 ierr = VecScale(&mone,lvec1);CHKERRQ(ierr); 2368 2369 ierr = VecScatterEnd(xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD,mat->Mvctx);CHKERRQ(ierr); 2370 ierr = VecCopy(bb,bb1);CHKERRQ(ierr); 2371 ierr = VecScatterBegin(lvec1,bb1,ADD_VALUES,SCATTER_REVERSE,mat->Mvctx);CHKERRQ(ierr); 2372 2373 /* upper diagonal part: bb1 = bb1 - B*x */ 2374 ierr = VecScale(&mone,mat->lvec);CHKERRQ(ierr); 2375 ierr = (*mat->B->ops->multadd)(mat->B,mat->lvec,bb1,bb1);CHKERRQ(ierr); 2376 2377 ierr = VecScatterEnd(lvec1,bb1,ADD_VALUES,SCATTER_REVERSE,mat->Mvctx);CHKERRQ(ierr); 2378 2379 /* diagonal sweep */ 2380 ierr = (*mat->A->ops->relax)(mat->A,bb1,omega,SOR_SYMMETRIC_SWEEP,fshift,lits,PETSC_NULL,xx);CHKERRQ(ierr); 2381 } 2382 ierr = VecDestroy(lvec1);CHKERRQ(ierr); 2383 ierr = VecDestroy(bb1);CHKERRQ(ierr); 2384 } else { 2385 SETERRQ(PETSC_ERR_SUP,"MatSORType is not supported for SBAIJ matrix format"); 2386 } 2387 PetscFunctionReturn(0); 2388 } 2389 2390