1 2 /* 3 Defines the basic matrix operations for the AIJ (compressed row) 4 matrix storage format. 5 */ 6 7 8 #include <../src/mat/impls/aij/seq/aij.h> /*I "petscmat.h" I*/ 9 #include <petscblaslapack.h> 10 #include <petscbt.h> 11 #include <petsc/private/kernels/blocktranspose.h> 12 13 PetscErrorCode MatSeqAIJSetTypeFromOptions(Mat A) 14 { 15 PetscErrorCode ierr; 16 PetscBool flg; 17 char type[256]; 18 19 PetscFunctionBegin; 20 ierr = PetscObjectOptionsBegin((PetscObject)A); 21 ierr = PetscOptionsFList("-mat_seqaij_type","Matrix SeqAIJ type","MatSeqAIJSetType",MatSeqAIJList,"seqaij",type,256,&flg);CHKERRQ(ierr); 22 if (flg) { 23 ierr = MatSeqAIJSetType(A,type);CHKERRQ(ierr); 24 } 25 ierr = PetscOptionsEnd();CHKERRQ(ierr); 26 PetscFunctionReturn(0); 27 } 28 29 PetscErrorCode MatGetColumnNorms_SeqAIJ(Mat A,NormType type,PetscReal *norms) 30 { 31 PetscErrorCode ierr; 32 PetscInt i,m,n; 33 Mat_SeqAIJ *aij = (Mat_SeqAIJ*)A->data; 34 35 PetscFunctionBegin; 36 ierr = MatGetSize(A,&m,&n);CHKERRQ(ierr); 37 ierr = PetscArrayzero(norms,n);CHKERRQ(ierr); 38 if (type == NORM_2) { 39 for (i=0; i<aij->i[m]; i++) { 40 norms[aij->j[i]] += PetscAbsScalar(aij->a[i]*aij->a[i]); 41 } 42 } else if (type == NORM_1) { 43 for (i=0; i<aij->i[m]; i++) { 44 norms[aij->j[i]] += PetscAbsScalar(aij->a[i]); 45 } 46 } else if (type == NORM_INFINITY) { 47 for (i=0; i<aij->i[m]; i++) { 48 norms[aij->j[i]] = PetscMax(PetscAbsScalar(aij->a[i]),norms[aij->j[i]]); 49 } 50 } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Unknown NormType"); 51 52 if (type == NORM_2) { 53 for (i=0; i<n; i++) norms[i] = PetscSqrtReal(norms[i]); 54 } 55 PetscFunctionReturn(0); 56 } 57 58 PetscErrorCode MatFindOffBlockDiagonalEntries_SeqAIJ(Mat A,IS *is) 59 { 60 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 61 PetscInt i,m=A->rmap->n,cnt = 0, bs = A->rmap->bs; 62 const PetscInt *jj = a->j,*ii = a->i; 63 PetscInt *rows; 64 PetscErrorCode ierr; 65 66 PetscFunctionBegin; 67 for (i=0; i<m; i++) { 68 if ((ii[i] != ii[i+1]) && ((jj[ii[i]] < bs*(i/bs)) || (jj[ii[i+1]-1] > bs*((i+bs)/bs)-1))) { 69 cnt++; 70 } 71 } 72 ierr = PetscMalloc1(cnt,&rows);CHKERRQ(ierr); 73 cnt = 0; 74 for (i=0; i<m; i++) { 75 if ((ii[i] != ii[i+1]) && ((jj[ii[i]] < bs*(i/bs)) || (jj[ii[i+1]-1] > bs*((i+bs)/bs)-1))) { 76 rows[cnt] = i; 77 cnt++; 78 } 79 } 80 ierr = ISCreateGeneral(PETSC_COMM_SELF,cnt,rows,PETSC_OWN_POINTER,is);CHKERRQ(ierr); 81 PetscFunctionReturn(0); 82 } 83 84 PetscErrorCode MatFindZeroDiagonals_SeqAIJ_Private(Mat A,PetscInt *nrows,PetscInt **zrows) 85 { 86 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 87 const MatScalar *aa = a->a; 88 PetscInt i,m=A->rmap->n,cnt = 0; 89 const PetscInt *ii = a->i,*jj = a->j,*diag; 90 PetscInt *rows; 91 PetscErrorCode ierr; 92 93 PetscFunctionBegin; 94 ierr = MatMarkDiagonal_SeqAIJ(A);CHKERRQ(ierr); 95 diag = a->diag; 96 for (i=0; i<m; i++) { 97 if ((diag[i] >= ii[i+1]) || (jj[diag[i]] != i) || (aa[diag[i]] == 0.0)) { 98 cnt++; 99 } 100 } 101 ierr = PetscMalloc1(cnt,&rows);CHKERRQ(ierr); 102 cnt = 0; 103 for (i=0; i<m; i++) { 104 if ((diag[i] >= ii[i+1]) || (jj[diag[i]] != i) || (aa[diag[i]] == 0.0)) { 105 rows[cnt++] = i; 106 } 107 } 108 *nrows = cnt; 109 *zrows = rows; 110 PetscFunctionReturn(0); 111 } 112 113 PetscErrorCode MatFindZeroDiagonals_SeqAIJ(Mat A,IS *zrows) 114 { 115 PetscInt nrows,*rows; 116 PetscErrorCode ierr; 117 118 PetscFunctionBegin; 119 *zrows = NULL; 120 ierr = MatFindZeroDiagonals_SeqAIJ_Private(A,&nrows,&rows);CHKERRQ(ierr); 121 ierr = ISCreateGeneral(PetscObjectComm((PetscObject)A),nrows,rows,PETSC_OWN_POINTER,zrows);CHKERRQ(ierr); 122 PetscFunctionReturn(0); 123 } 124 125 PetscErrorCode MatFindNonzeroRows_SeqAIJ(Mat A,IS *keptrows) 126 { 127 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 128 const MatScalar *aa; 129 PetscInt m=A->rmap->n,cnt = 0; 130 const PetscInt *ii; 131 PetscInt n,i,j,*rows; 132 PetscErrorCode ierr; 133 134 PetscFunctionBegin; 135 *keptrows = 0; 136 ii = a->i; 137 for (i=0; i<m; i++) { 138 n = ii[i+1] - ii[i]; 139 if (!n) { 140 cnt++; 141 goto ok1; 142 } 143 aa = a->a + ii[i]; 144 for (j=0; j<n; j++) { 145 if (aa[j] != 0.0) goto ok1; 146 } 147 cnt++; 148 ok1:; 149 } 150 if (!cnt) PetscFunctionReturn(0); 151 ierr = PetscMalloc1(A->rmap->n-cnt,&rows);CHKERRQ(ierr); 152 cnt = 0; 153 for (i=0; i<m; i++) { 154 n = ii[i+1] - ii[i]; 155 if (!n) continue; 156 aa = a->a + ii[i]; 157 for (j=0; j<n; j++) { 158 if (aa[j] != 0.0) { 159 rows[cnt++] = i; 160 break; 161 } 162 } 163 } 164 ierr = ISCreateGeneral(PETSC_COMM_SELF,cnt,rows,PETSC_OWN_POINTER,keptrows);CHKERRQ(ierr); 165 PetscFunctionReturn(0); 166 } 167 168 PetscErrorCode MatDiagonalSet_SeqAIJ(Mat Y,Vec D,InsertMode is) 169 { 170 PetscErrorCode ierr; 171 Mat_SeqAIJ *aij = (Mat_SeqAIJ*) Y->data; 172 PetscInt i,m = Y->rmap->n; 173 const PetscInt *diag; 174 MatScalar *aa = aij->a; 175 const PetscScalar *v; 176 PetscBool missing; 177 178 PetscFunctionBegin; 179 if (Y->assembled) { 180 ierr = MatMissingDiagonal_SeqAIJ(Y,&missing,NULL);CHKERRQ(ierr); 181 if (!missing) { 182 diag = aij->diag; 183 ierr = VecGetArrayRead(D,&v);CHKERRQ(ierr); 184 if (is == INSERT_VALUES) { 185 for (i=0; i<m; i++) { 186 aa[diag[i]] = v[i]; 187 } 188 } else { 189 for (i=0; i<m; i++) { 190 aa[diag[i]] += v[i]; 191 } 192 } 193 ierr = VecRestoreArrayRead(D,&v);CHKERRQ(ierr); 194 PetscFunctionReturn(0); 195 } 196 ierr = MatSeqAIJInvalidateDiagonal(Y);CHKERRQ(ierr); 197 } 198 ierr = MatDiagonalSet_Default(Y,D,is);CHKERRQ(ierr); 199 PetscFunctionReturn(0); 200 } 201 202 PetscErrorCode MatGetRowIJ_SeqAIJ(Mat A,PetscInt oshift,PetscBool symmetric,PetscBool inodecompressed,PetscInt *m,const PetscInt *ia[],const PetscInt *ja[],PetscBool *done) 203 { 204 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 205 PetscErrorCode ierr; 206 PetscInt i,ishift; 207 208 PetscFunctionBegin; 209 *m = A->rmap->n; 210 if (!ia) PetscFunctionReturn(0); 211 ishift = 0; 212 if (symmetric && !A->structurally_symmetric) { 213 ierr = MatToSymmetricIJ_SeqAIJ(A->rmap->n,a->i,a->j,PETSC_TRUE,ishift,oshift,(PetscInt**)ia,(PetscInt**)ja);CHKERRQ(ierr); 214 } else if (oshift == 1) { 215 PetscInt *tia; 216 PetscInt nz = a->i[A->rmap->n]; 217 /* malloc space and add 1 to i and j indices */ 218 ierr = PetscMalloc1(A->rmap->n+1,&tia);CHKERRQ(ierr); 219 for (i=0; i<A->rmap->n+1; i++) tia[i] = a->i[i] + 1; 220 *ia = tia; 221 if (ja) { 222 PetscInt *tja; 223 ierr = PetscMalloc1(nz+1,&tja);CHKERRQ(ierr); 224 for (i=0; i<nz; i++) tja[i] = a->j[i] + 1; 225 *ja = tja; 226 } 227 } else { 228 *ia = a->i; 229 if (ja) *ja = a->j; 230 } 231 PetscFunctionReturn(0); 232 } 233 234 PetscErrorCode MatRestoreRowIJ_SeqAIJ(Mat A,PetscInt oshift,PetscBool symmetric,PetscBool inodecompressed,PetscInt *n,const PetscInt *ia[],const PetscInt *ja[],PetscBool *done) 235 { 236 PetscErrorCode ierr; 237 238 PetscFunctionBegin; 239 if (!ia) PetscFunctionReturn(0); 240 if ((symmetric && !A->structurally_symmetric) || oshift == 1) { 241 ierr = PetscFree(*ia);CHKERRQ(ierr); 242 if (ja) {ierr = PetscFree(*ja);CHKERRQ(ierr);} 243 } 244 PetscFunctionReturn(0); 245 } 246 247 PetscErrorCode MatGetColumnIJ_SeqAIJ(Mat A,PetscInt oshift,PetscBool symmetric,PetscBool inodecompressed,PetscInt *nn,const PetscInt *ia[],const PetscInt *ja[],PetscBool *done) 248 { 249 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 250 PetscErrorCode ierr; 251 PetscInt i,*collengths,*cia,*cja,n = A->cmap->n,m = A->rmap->n; 252 PetscInt nz = a->i[m],row,*jj,mr,col; 253 254 PetscFunctionBegin; 255 *nn = n; 256 if (!ia) PetscFunctionReturn(0); 257 if (symmetric) { 258 ierr = MatToSymmetricIJ_SeqAIJ(A->rmap->n,a->i,a->j,PETSC_TRUE,0,oshift,(PetscInt**)ia,(PetscInt**)ja);CHKERRQ(ierr); 259 } else { 260 ierr = PetscCalloc1(n,&collengths);CHKERRQ(ierr); 261 ierr = PetscMalloc1(n+1,&cia);CHKERRQ(ierr); 262 ierr = PetscMalloc1(nz,&cja);CHKERRQ(ierr); 263 jj = a->j; 264 for (i=0; i<nz; i++) { 265 collengths[jj[i]]++; 266 } 267 cia[0] = oshift; 268 for (i=0; i<n; i++) { 269 cia[i+1] = cia[i] + collengths[i]; 270 } 271 ierr = PetscArrayzero(collengths,n);CHKERRQ(ierr); 272 jj = a->j; 273 for (row=0; row<m; row++) { 274 mr = a->i[row+1] - a->i[row]; 275 for (i=0; i<mr; i++) { 276 col = *jj++; 277 278 cja[cia[col] + collengths[col]++ - oshift] = row + oshift; 279 } 280 } 281 ierr = PetscFree(collengths);CHKERRQ(ierr); 282 *ia = cia; *ja = cja; 283 } 284 PetscFunctionReturn(0); 285 } 286 287 PetscErrorCode MatRestoreColumnIJ_SeqAIJ(Mat A,PetscInt oshift,PetscBool symmetric,PetscBool inodecompressed,PetscInt *n,const PetscInt *ia[],const PetscInt *ja[],PetscBool *done) 288 { 289 PetscErrorCode ierr; 290 291 PetscFunctionBegin; 292 if (!ia) PetscFunctionReturn(0); 293 294 ierr = PetscFree(*ia);CHKERRQ(ierr); 295 ierr = PetscFree(*ja);CHKERRQ(ierr); 296 PetscFunctionReturn(0); 297 } 298 299 /* 300 MatGetColumnIJ_SeqAIJ_Color() and MatRestoreColumnIJ_SeqAIJ_Color() are customized from 301 MatGetColumnIJ_SeqAIJ() and MatRestoreColumnIJ_SeqAIJ() by adding an output 302 spidx[], index of a->a, to be used in MatTransposeColoringCreate_SeqAIJ() and MatFDColoringCreate_SeqXAIJ() 303 */ 304 PetscErrorCode MatGetColumnIJ_SeqAIJ_Color(Mat A,PetscInt oshift,PetscBool symmetric,PetscBool inodecompressed,PetscInt *nn,const PetscInt *ia[],const PetscInt *ja[],PetscInt *spidx[],PetscBool *done) 305 { 306 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 307 PetscErrorCode ierr; 308 PetscInt i,*collengths,*cia,*cja,n = A->cmap->n,m = A->rmap->n; 309 PetscInt nz = a->i[m],row,mr,col,tmp; 310 PetscInt *cspidx; 311 const PetscInt *jj; 312 313 PetscFunctionBegin; 314 *nn = n; 315 if (!ia) PetscFunctionReturn(0); 316 317 ierr = PetscCalloc1(n,&collengths);CHKERRQ(ierr); 318 ierr = PetscMalloc1(n+1,&cia);CHKERRQ(ierr); 319 ierr = PetscMalloc1(nz,&cja);CHKERRQ(ierr); 320 ierr = PetscMalloc1(nz,&cspidx);CHKERRQ(ierr); 321 jj = a->j; 322 for (i=0; i<nz; i++) { 323 collengths[jj[i]]++; 324 } 325 cia[0] = oshift; 326 for (i=0; i<n; i++) { 327 cia[i+1] = cia[i] + collengths[i]; 328 } 329 ierr = PetscArrayzero(collengths,n);CHKERRQ(ierr); 330 jj = a->j; 331 for (row=0; row<m; row++) { 332 mr = a->i[row+1] - a->i[row]; 333 for (i=0; i<mr; i++) { 334 col = *jj++; 335 tmp = cia[col] + collengths[col]++ - oshift; 336 cspidx[tmp] = a->i[row] + i; /* index of a->j */ 337 cja[tmp] = row + oshift; 338 } 339 } 340 ierr = PetscFree(collengths);CHKERRQ(ierr); 341 *ia = cia; 342 *ja = cja; 343 *spidx = cspidx; 344 PetscFunctionReturn(0); 345 } 346 347 PetscErrorCode MatRestoreColumnIJ_SeqAIJ_Color(Mat A,PetscInt oshift,PetscBool symmetric,PetscBool inodecompressed,PetscInt *n,const PetscInt *ia[],const PetscInt *ja[],PetscInt *spidx[],PetscBool *done) 348 { 349 PetscErrorCode ierr; 350 351 PetscFunctionBegin; 352 ierr = MatRestoreColumnIJ_SeqAIJ(A,oshift,symmetric,inodecompressed,n,ia,ja,done);CHKERRQ(ierr); 353 ierr = PetscFree(*spidx);CHKERRQ(ierr); 354 PetscFunctionReturn(0); 355 } 356 357 PetscErrorCode MatSetValuesRow_SeqAIJ(Mat A,PetscInt row,const PetscScalar v[]) 358 { 359 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 360 PetscInt *ai = a->i; 361 PetscErrorCode ierr; 362 363 PetscFunctionBegin; 364 ierr = PetscArraycpy(a->a+ai[row],v,ai[row+1]-ai[row]);CHKERRQ(ierr); 365 #if defined(PETSC_HAVE_VIENNACL) || defined(PETSC_HAVE_CUDA) 366 if (A->offloadmask != PETSC_OFFLOAD_UNALLOCATED && ai[row+1]-ai[row]) A->offloadmask = PETSC_OFFLOAD_CPU; 367 #endif 368 PetscFunctionReturn(0); 369 } 370 371 /* 372 MatSeqAIJSetValuesLocalFast - An optimized version of MatSetValuesLocal() for SeqAIJ matrices with several assumptions 373 374 - a single row of values is set with each call 375 - no row or column indices are negative or (in error) larger than the number of rows or columns 376 - the values are always added to the matrix, not set 377 - no new locations are introduced in the nonzero structure of the matrix 378 379 This does NOT assume the global column indices are sorted 380 381 */ 382 383 #include <petsc/private/isimpl.h> 384 PetscErrorCode MatSeqAIJSetValuesLocalFast(Mat A,PetscInt m,const PetscInt im[],PetscInt n,const PetscInt in[],const PetscScalar v[],InsertMode is) 385 { 386 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 387 PetscInt low,high,t,row,nrow,i,col,l; 388 const PetscInt *rp,*ai = a->i,*ailen = a->ilen,*aj = a->j; 389 PetscInt lastcol = -1; 390 MatScalar *ap,value,*aa = a->a; 391 const PetscInt *ridx = A->rmap->mapping->indices,*cidx = A->cmap->mapping->indices; 392 393 row = ridx[im[0]]; 394 rp = aj + ai[row]; 395 ap = aa + ai[row]; 396 nrow = ailen[row]; 397 low = 0; 398 high = nrow; 399 for (l=0; l<n; l++) { /* loop over added columns */ 400 col = cidx[in[l]]; 401 value = v[l]; 402 403 if (col <= lastcol) low = 0; 404 else high = nrow; 405 lastcol = col; 406 while (high-low > 5) { 407 t = (low+high)/2; 408 if (rp[t] > col) high = t; 409 else low = t; 410 } 411 for (i=low; i<high; i++) { 412 if (rp[i] == col) { 413 ap[i] += value; 414 low = i + 1; 415 break; 416 } 417 } 418 } 419 #if defined(PETSC_HAVE_VIENNACL) || defined(PETSC_HAVE_CUDA) 420 if (A->offloadmask != PETSC_OFFLOAD_UNALLOCATED && m && n) A->offloadmask = PETSC_OFFLOAD_CPU; 421 #endif 422 return 0; 423 } 424 425 PetscErrorCode MatSetValues_SeqAIJ(Mat A,PetscInt m,const PetscInt im[],PetscInt n,const PetscInt in[],const PetscScalar v[],InsertMode is) 426 { 427 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 428 PetscInt *rp,k,low,high,t,ii,row,nrow,i,col,l,rmax,N; 429 PetscInt *imax = a->imax,*ai = a->i,*ailen = a->ilen; 430 PetscErrorCode ierr; 431 PetscInt *aj = a->j,nonew = a->nonew,lastcol = -1; 432 MatScalar *ap=NULL,value=0.0,*aa = a->a; 433 PetscBool ignorezeroentries = a->ignorezeroentries; 434 PetscBool roworiented = a->roworiented; 435 #if defined(PETSC_HAVE_VIENNACL) || defined(PETSC_HAVE_CUDA) 436 PetscBool inserted = PETSC_FALSE; 437 #endif 438 439 PetscFunctionBegin; 440 for (k=0; k<m; k++) { /* loop over added rows */ 441 row = im[k]; 442 if (row < 0) continue; 443 #if defined(PETSC_USE_DEBUG) 444 if (row >= A->rmap->n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Row too large: row %D max %D",row,A->rmap->n-1); 445 #endif 446 rp = aj + ai[row]; 447 if (!A->structure_only) ap = aa + ai[row]; 448 rmax = imax[row]; nrow = ailen[row]; 449 low = 0; 450 high = nrow; 451 for (l=0; l<n; l++) { /* loop over added columns */ 452 if (in[l] < 0) continue; 453 #if defined(PETSC_USE_DEBUG) 454 if (in[l] >= A->cmap->n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Column too large: col %D max %D",in[l],A->cmap->n-1); 455 #endif 456 col = in[l]; 457 if (v && !A->structure_only) value = roworiented ? v[l + k*n] : v[k + l*m]; 458 if (!A->structure_only && value == 0.0 && ignorezeroentries && is == ADD_VALUES && row != col) continue; 459 460 if (col <= lastcol) low = 0; 461 else high = nrow; 462 lastcol = col; 463 while (high-low > 5) { 464 t = (low+high)/2; 465 if (rp[t] > col) high = t; 466 else low = t; 467 } 468 for (i=low; i<high; i++) { 469 if (rp[i] > col) break; 470 if (rp[i] == col) { 471 if (!A->structure_only) { 472 if (is == ADD_VALUES) { 473 ap[i] += value; 474 (void)PetscLogFlops(1.0); 475 } 476 else ap[i] = value; 477 #if defined(PETSC_HAVE_VIENNACL) || defined(PETSC_HAVE_CUDA) 478 inserted = PETSC_TRUE; 479 #endif 480 } 481 low = i + 1; 482 goto noinsert; 483 } 484 } 485 if (value == 0.0 && ignorezeroentries && row != col) goto noinsert; 486 if (nonew == 1) goto noinsert; 487 if (nonew == -1) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new nonzero at (%D,%D) in the matrix",row,col); 488 if (A->structure_only) { 489 MatSeqXAIJReallocateAIJ_structure_only(A,A->rmap->n,1,nrow,row,col,rmax,ai,aj,rp,imax,nonew,MatScalar); 490 } else { 491 MatSeqXAIJReallocateAIJ(A,A->rmap->n,1,nrow,row,col,rmax,aa,ai,aj,rp,ap,imax,nonew,MatScalar); 492 } 493 N = nrow++ - 1; a->nz++; high++; 494 /* shift up all the later entries in this row */ 495 ierr = PetscArraymove(rp+i+1,rp+i,N-i+1);CHKERRQ(ierr); 496 rp[i] = col; 497 if (!A->structure_only){ 498 ierr = PetscArraymove(ap+i+1,ap+i,N-i+1);CHKERRQ(ierr); 499 ap[i] = value; 500 } 501 low = i + 1; 502 A->nonzerostate++; 503 #if defined(PETSC_HAVE_VIENNACL) || defined(PETSC_HAVE_CUDA) 504 inserted = PETSC_TRUE; 505 #endif 506 noinsert:; 507 } 508 ailen[row] = nrow; 509 } 510 #if defined(PETSC_HAVE_VIENNACL) || defined(PETSC_HAVE_CUDA) 511 if (A->offloadmask != PETSC_OFFLOAD_UNALLOCATED && inserted) A->offloadmask = PETSC_OFFLOAD_CPU; 512 #endif 513 PetscFunctionReturn(0); 514 } 515 516 PetscErrorCode MatSetValues_SeqAIJ_SortedFull(Mat A,PetscInt m,const PetscInt im[],PetscInt n,const PetscInt in[],const PetscScalar v[],InsertMode is) 517 { 518 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 519 PetscInt *rp,k,row; 520 PetscInt *ai = a->i,*ailen = a->ilen; 521 PetscErrorCode ierr; 522 PetscInt *aj = a->j; 523 MatScalar *aa = a->a,*ap; 524 525 PetscFunctionBegin; 526 for (k=0; k<m; k++) { /* loop over added rows */ 527 row = im[k]; 528 rp = aj + ai[row]; 529 ap = aa + ai[row]; 530 if (!A->was_assembled) { 531 ierr = PetscMemcpy(rp,in,n*sizeof(PetscInt));CHKERRQ(ierr); 532 } 533 if (!A->structure_only) { 534 if (v) { 535 ierr = PetscMemcpy(ap,v,n*sizeof(PetscScalar));CHKERRQ(ierr); 536 v += n; 537 } else { 538 ierr = PetscMemzero(ap,n*sizeof(PetscScalar));CHKERRQ(ierr); 539 } 540 } 541 ailen[row] = n; 542 a->nz += n; 543 } 544 #if defined(PETSC_HAVE_VIENNACL) || defined(PETSC_HAVE_CUDA) 545 if (A->offloadmask != PETSC_OFFLOAD_UNALLOCATED && m && n) A->offloadmask = PETSC_OFFLOAD_CPU; 546 #endif 547 PetscFunctionReturn(0); 548 } 549 550 551 PetscErrorCode MatGetValues_SeqAIJ(Mat A,PetscInt m,const PetscInt im[],PetscInt n,const PetscInt in[],PetscScalar v[]) 552 { 553 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 554 PetscInt *rp,k,low,high,t,row,nrow,i,col,l,*aj = a->j; 555 PetscInt *ai = a->i,*ailen = a->ilen; 556 MatScalar *ap,*aa = a->a; 557 558 PetscFunctionBegin; 559 for (k=0; k<m; k++) { /* loop over rows */ 560 row = im[k]; 561 if (row < 0) {v += n; continue;} /* SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Negative row: %D",row); */ 562 if (row >= A->rmap->n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Row too large: row %D max %D",row,A->rmap->n-1); 563 rp = aj + ai[row]; ap = aa + ai[row]; 564 nrow = ailen[row]; 565 for (l=0; l<n; l++) { /* loop over columns */ 566 if (in[l] < 0) {v++; continue;} /* SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Negative column: %D",in[l]); */ 567 if (in[l] >= A->cmap->n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Column too large: col %D max %D",in[l],A->cmap->n-1); 568 col = in[l]; 569 high = nrow; low = 0; /* assume unsorted */ 570 while (high-low > 5) { 571 t = (low+high)/2; 572 if (rp[t] > col) high = t; 573 else low = t; 574 } 575 for (i=low; i<high; i++) { 576 if (rp[i] > col) break; 577 if (rp[i] == col) { 578 *v++ = ap[i]; 579 goto finished; 580 } 581 } 582 *v++ = 0.0; 583 finished:; 584 } 585 } 586 PetscFunctionReturn(0); 587 } 588 589 590 PetscErrorCode MatView_SeqAIJ_Binary(Mat A,PetscViewer viewer) 591 { 592 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 593 PetscErrorCode ierr; 594 PetscInt i,*col_lens; 595 int fd; 596 FILE *file; 597 598 PetscFunctionBegin; 599 ierr = PetscViewerBinaryGetDescriptor(viewer,&fd);CHKERRQ(ierr); 600 ierr = PetscMalloc1(4+A->rmap->n,&col_lens);CHKERRQ(ierr); 601 602 col_lens[0] = MAT_FILE_CLASSID; 603 col_lens[1] = A->rmap->n; 604 col_lens[2] = A->cmap->n; 605 col_lens[3] = a->nz; 606 607 /* store lengths of each row and write (including header) to file */ 608 for (i=0; i<A->rmap->n; i++) { 609 col_lens[4+i] = a->i[i+1] - a->i[i]; 610 } 611 ierr = PetscBinaryWrite(fd,col_lens,4+A->rmap->n,PETSC_INT,PETSC_TRUE);CHKERRQ(ierr); 612 ierr = PetscFree(col_lens);CHKERRQ(ierr); 613 614 /* store column indices (zero start index) */ 615 ierr = PetscBinaryWrite(fd,a->j,a->nz,PETSC_INT,PETSC_FALSE);CHKERRQ(ierr); 616 617 /* store nonzero values */ 618 ierr = PetscBinaryWrite(fd,a->a,a->nz,PETSC_SCALAR,PETSC_FALSE);CHKERRQ(ierr); 619 620 ierr = PetscViewerBinaryGetInfoPointer(viewer,&file);CHKERRQ(ierr); 621 if (file) { 622 fprintf(file,"-matload_block_size %d\n",(int)PetscAbs(A->rmap->bs)); 623 } 624 PetscFunctionReturn(0); 625 } 626 627 static PetscErrorCode MatView_SeqAIJ_ASCII_structonly(Mat A,PetscViewer viewer) 628 { 629 PetscErrorCode ierr; 630 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 631 PetscInt i,k,m=A->rmap->N; 632 633 PetscFunctionBegin; 634 ierr = PetscViewerASCIIUseTabs(viewer,PETSC_FALSE);CHKERRQ(ierr); 635 for (i=0; i<m; i++) { 636 ierr = PetscViewerASCIIPrintf(viewer,"row %D:",i);CHKERRQ(ierr); 637 for (k=a->i[i]; k<a->i[i+1]; k++) { 638 ierr = PetscViewerASCIIPrintf(viewer," (%D) ",a->j[k]);CHKERRQ(ierr); 639 } 640 ierr = PetscViewerASCIIPrintf(viewer,"\n");CHKERRQ(ierr); 641 } 642 ierr = PetscViewerASCIIUseTabs(viewer,PETSC_TRUE);CHKERRQ(ierr); 643 PetscFunctionReturn(0); 644 } 645 646 extern PetscErrorCode MatSeqAIJFactorInfo_Matlab(Mat,PetscViewer); 647 648 PetscErrorCode MatView_SeqAIJ_ASCII(Mat A,PetscViewer viewer) 649 { 650 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 651 PetscErrorCode ierr; 652 PetscInt i,j,m = A->rmap->n; 653 const char *name; 654 PetscViewerFormat format; 655 656 PetscFunctionBegin; 657 if (A->structure_only) { 658 ierr = MatView_SeqAIJ_ASCII_structonly(A,viewer);CHKERRQ(ierr); 659 PetscFunctionReturn(0); 660 } 661 662 ierr = PetscViewerGetFormat(viewer,&format);CHKERRQ(ierr); 663 if (format == PETSC_VIEWER_ASCII_MATLAB) { 664 PetscInt nofinalvalue = 0; 665 if (m && ((a->i[m] == a->i[m-1]) || (a->j[a->nz-1] != A->cmap->n-1))) { 666 /* Need a dummy value to ensure the dimension of the matrix. */ 667 nofinalvalue = 1; 668 } 669 ierr = PetscViewerASCIIUseTabs(viewer,PETSC_FALSE);CHKERRQ(ierr); 670 ierr = PetscViewerASCIIPrintf(viewer,"%% Size = %D %D \n",m,A->cmap->n);CHKERRQ(ierr); 671 ierr = PetscViewerASCIIPrintf(viewer,"%% Nonzeros = %D \n",a->nz);CHKERRQ(ierr); 672 #if defined(PETSC_USE_COMPLEX) 673 ierr = PetscViewerASCIIPrintf(viewer,"zzz = zeros(%D,4);\n",a->nz+nofinalvalue);CHKERRQ(ierr); 674 #else 675 ierr = PetscViewerASCIIPrintf(viewer,"zzz = zeros(%D,3);\n",a->nz+nofinalvalue);CHKERRQ(ierr); 676 #endif 677 ierr = PetscViewerASCIIPrintf(viewer,"zzz = [\n");CHKERRQ(ierr); 678 679 for (i=0; i<m; i++) { 680 for (j=a->i[i]; j<a->i[i+1]; j++) { 681 #if defined(PETSC_USE_COMPLEX) 682 ierr = PetscViewerASCIIPrintf(viewer,"%D %D %18.16e %18.16e\n",i+1,a->j[j]+1,(double)PetscRealPart(a->a[j]),(double)PetscImaginaryPart(a->a[j]));CHKERRQ(ierr); 683 #else 684 ierr = PetscViewerASCIIPrintf(viewer,"%D %D %18.16e\n",i+1,a->j[j]+1,(double)a->a[j]);CHKERRQ(ierr); 685 #endif 686 } 687 } 688 if (nofinalvalue) { 689 #if defined(PETSC_USE_COMPLEX) 690 ierr = PetscViewerASCIIPrintf(viewer,"%D %D %18.16e %18.16e\n",m,A->cmap->n,0.,0.);CHKERRQ(ierr); 691 #else 692 ierr = PetscViewerASCIIPrintf(viewer,"%D %D %18.16e\n",m,A->cmap->n,0.0);CHKERRQ(ierr); 693 #endif 694 } 695 ierr = PetscObjectGetName((PetscObject)A,&name);CHKERRQ(ierr); 696 ierr = PetscViewerASCIIPrintf(viewer,"];\n %s = spconvert(zzz);\n",name);CHKERRQ(ierr); 697 ierr = PetscViewerASCIIUseTabs(viewer,PETSC_TRUE);CHKERRQ(ierr); 698 } else if (format == PETSC_VIEWER_ASCII_FACTOR_INFO || format == PETSC_VIEWER_ASCII_INFO || format == PETSC_VIEWER_ASCII_INFO_DETAIL) { 699 PetscFunctionReturn(0); 700 } else if (format == PETSC_VIEWER_ASCII_COMMON) { 701 ierr = PetscViewerASCIIUseTabs(viewer,PETSC_FALSE);CHKERRQ(ierr); 702 for (i=0; i<m; i++) { 703 ierr = PetscViewerASCIIPrintf(viewer,"row %D:",i);CHKERRQ(ierr); 704 for (j=a->i[i]; j<a->i[i+1]; j++) { 705 #if defined(PETSC_USE_COMPLEX) 706 if (PetscImaginaryPart(a->a[j]) > 0.0 && PetscRealPart(a->a[j]) != 0.0) { 707 ierr = PetscViewerASCIIPrintf(viewer," (%D, %g + %g i)",a->j[j],(double)PetscRealPart(a->a[j]),(double)PetscImaginaryPart(a->a[j]));CHKERRQ(ierr); 708 } else if (PetscImaginaryPart(a->a[j]) < 0.0 && PetscRealPart(a->a[j]) != 0.0) { 709 ierr = PetscViewerASCIIPrintf(viewer," (%D, %g - %g i)",a->j[j],(double)PetscRealPart(a->a[j]),(double)-PetscImaginaryPart(a->a[j]));CHKERRQ(ierr); 710 } else if (PetscRealPart(a->a[j]) != 0.0) { 711 ierr = PetscViewerASCIIPrintf(viewer," (%D, %g) ",a->j[j],(double)PetscRealPart(a->a[j]));CHKERRQ(ierr); 712 } 713 #else 714 if (a->a[j] != 0.0) {ierr = PetscViewerASCIIPrintf(viewer," (%D, %g) ",a->j[j],(double)a->a[j]);CHKERRQ(ierr);} 715 #endif 716 } 717 ierr = PetscViewerASCIIPrintf(viewer,"\n");CHKERRQ(ierr); 718 } 719 ierr = PetscViewerASCIIUseTabs(viewer,PETSC_TRUE);CHKERRQ(ierr); 720 } else if (format == PETSC_VIEWER_ASCII_SYMMODU) { 721 PetscInt nzd=0,fshift=1,*sptr; 722 ierr = PetscViewerASCIIUseTabs(viewer,PETSC_FALSE);CHKERRQ(ierr); 723 ierr = PetscMalloc1(m+1,&sptr);CHKERRQ(ierr); 724 for (i=0; i<m; i++) { 725 sptr[i] = nzd+1; 726 for (j=a->i[i]; j<a->i[i+1]; j++) { 727 if (a->j[j] >= i) { 728 #if defined(PETSC_USE_COMPLEX) 729 if (PetscImaginaryPart(a->a[j]) != 0.0 || PetscRealPart(a->a[j]) != 0.0) nzd++; 730 #else 731 if (a->a[j] != 0.0) nzd++; 732 #endif 733 } 734 } 735 } 736 sptr[m] = nzd+1; 737 ierr = PetscViewerASCIIPrintf(viewer," %D %D\n\n",m,nzd);CHKERRQ(ierr); 738 for (i=0; i<m+1; i+=6) { 739 if (i+4<m) { 740 ierr = PetscViewerASCIIPrintf(viewer," %D %D %D %D %D %D\n",sptr[i],sptr[i+1],sptr[i+2],sptr[i+3],sptr[i+4],sptr[i+5]);CHKERRQ(ierr); 741 } else if (i+3<m) { 742 ierr = PetscViewerASCIIPrintf(viewer," %D %D %D %D %D\n",sptr[i],sptr[i+1],sptr[i+2],sptr[i+3],sptr[i+4]);CHKERRQ(ierr); 743 } else if (i+2<m) { 744 ierr = PetscViewerASCIIPrintf(viewer," %D %D %D %D\n",sptr[i],sptr[i+1],sptr[i+2],sptr[i+3]);CHKERRQ(ierr); 745 } else if (i+1<m) { 746 ierr = PetscViewerASCIIPrintf(viewer," %D %D %D\n",sptr[i],sptr[i+1],sptr[i+2]);CHKERRQ(ierr); 747 } else if (i<m) { 748 ierr = PetscViewerASCIIPrintf(viewer," %D %D\n",sptr[i],sptr[i+1]);CHKERRQ(ierr); 749 } else { 750 ierr = PetscViewerASCIIPrintf(viewer," %D\n",sptr[i]);CHKERRQ(ierr); 751 } 752 } 753 ierr = PetscViewerASCIIPrintf(viewer,"\n");CHKERRQ(ierr); 754 ierr = PetscFree(sptr);CHKERRQ(ierr); 755 for (i=0; i<m; i++) { 756 for (j=a->i[i]; j<a->i[i+1]; j++) { 757 if (a->j[j] >= i) {ierr = PetscViewerASCIIPrintf(viewer," %D ",a->j[j]+fshift);CHKERRQ(ierr);} 758 } 759 ierr = PetscViewerASCIIPrintf(viewer,"\n");CHKERRQ(ierr); 760 } 761 ierr = PetscViewerASCIIPrintf(viewer,"\n");CHKERRQ(ierr); 762 for (i=0; i<m; i++) { 763 for (j=a->i[i]; j<a->i[i+1]; j++) { 764 if (a->j[j] >= i) { 765 #if defined(PETSC_USE_COMPLEX) 766 if (PetscImaginaryPart(a->a[j]) != 0.0 || PetscRealPart(a->a[j]) != 0.0) { 767 ierr = PetscViewerASCIIPrintf(viewer," %18.16e %18.16e ",(double)PetscRealPart(a->a[j]),(double)PetscImaginaryPart(a->a[j]));CHKERRQ(ierr); 768 } 769 #else 770 if (a->a[j] != 0.0) {ierr = PetscViewerASCIIPrintf(viewer," %18.16e ",(double)a->a[j]);CHKERRQ(ierr);} 771 #endif 772 } 773 } 774 ierr = PetscViewerASCIIPrintf(viewer,"\n");CHKERRQ(ierr); 775 } 776 ierr = PetscViewerASCIIUseTabs(viewer,PETSC_TRUE);CHKERRQ(ierr); 777 } else if (format == PETSC_VIEWER_ASCII_DENSE) { 778 PetscInt cnt = 0,jcnt; 779 PetscScalar value; 780 #if defined(PETSC_USE_COMPLEX) 781 PetscBool realonly = PETSC_TRUE; 782 783 for (i=0; i<a->i[m]; i++) { 784 if (PetscImaginaryPart(a->a[i]) != 0.0) { 785 realonly = PETSC_FALSE; 786 break; 787 } 788 } 789 #endif 790 791 ierr = PetscViewerASCIIUseTabs(viewer,PETSC_FALSE);CHKERRQ(ierr); 792 for (i=0; i<m; i++) { 793 jcnt = 0; 794 for (j=0; j<A->cmap->n; j++) { 795 if (jcnt < a->i[i+1]-a->i[i] && j == a->j[cnt]) { 796 value = a->a[cnt++]; 797 jcnt++; 798 } else { 799 value = 0.0; 800 } 801 #if defined(PETSC_USE_COMPLEX) 802 if (realonly) { 803 ierr = PetscViewerASCIIPrintf(viewer," %7.5e ",(double)PetscRealPart(value));CHKERRQ(ierr); 804 } else { 805 ierr = PetscViewerASCIIPrintf(viewer," %7.5e+%7.5e i ",(double)PetscRealPart(value),(double)PetscImaginaryPart(value));CHKERRQ(ierr); 806 } 807 #else 808 ierr = PetscViewerASCIIPrintf(viewer," %7.5e ",(double)value);CHKERRQ(ierr); 809 #endif 810 } 811 ierr = PetscViewerASCIIPrintf(viewer,"\n");CHKERRQ(ierr); 812 } 813 ierr = PetscViewerASCIIUseTabs(viewer,PETSC_TRUE);CHKERRQ(ierr); 814 } else if (format == PETSC_VIEWER_ASCII_MATRIXMARKET) { 815 PetscInt fshift=1; 816 ierr = PetscViewerASCIIUseTabs(viewer,PETSC_FALSE);CHKERRQ(ierr); 817 #if defined(PETSC_USE_COMPLEX) 818 ierr = PetscViewerASCIIPrintf(viewer,"%%%%MatrixMarket matrix coordinate complex general\n");CHKERRQ(ierr); 819 #else 820 ierr = PetscViewerASCIIPrintf(viewer,"%%%%MatrixMarket matrix coordinate real general\n");CHKERRQ(ierr); 821 #endif 822 ierr = PetscViewerASCIIPrintf(viewer,"%D %D %D\n", m, A->cmap->n, a->nz);CHKERRQ(ierr); 823 for (i=0; i<m; i++) { 824 for (j=a->i[i]; j<a->i[i+1]; j++) { 825 #if defined(PETSC_USE_COMPLEX) 826 ierr = PetscViewerASCIIPrintf(viewer,"%D %D %g %g\n", i+fshift,a->j[j]+fshift,(double)PetscRealPart(a->a[j]),(double)PetscImaginaryPart(a->a[j]));CHKERRQ(ierr); 827 #else 828 ierr = PetscViewerASCIIPrintf(viewer,"%D %D %g\n", i+fshift, a->j[j]+fshift, (double)a->a[j]);CHKERRQ(ierr); 829 #endif 830 } 831 } 832 ierr = PetscViewerASCIIUseTabs(viewer,PETSC_TRUE);CHKERRQ(ierr); 833 } else { 834 ierr = PetscViewerASCIIUseTabs(viewer,PETSC_FALSE);CHKERRQ(ierr); 835 if (A->factortype) { 836 for (i=0; i<m; i++) { 837 ierr = PetscViewerASCIIPrintf(viewer,"row %D:",i);CHKERRQ(ierr); 838 /* L part */ 839 for (j=a->i[i]; j<a->i[i+1]; j++) { 840 #if defined(PETSC_USE_COMPLEX) 841 if (PetscImaginaryPart(a->a[j]) > 0.0) { 842 ierr = PetscViewerASCIIPrintf(viewer," (%D, %g + %g i)",a->j[j],(double)PetscRealPart(a->a[j]),(double)PetscImaginaryPart(a->a[j]));CHKERRQ(ierr); 843 } else if (PetscImaginaryPart(a->a[j]) < 0.0) { 844 ierr = PetscViewerASCIIPrintf(viewer," (%D, %g - %g i)",a->j[j],(double)PetscRealPart(a->a[j]),(double)(-PetscImaginaryPart(a->a[j])));CHKERRQ(ierr); 845 } else { 846 ierr = PetscViewerASCIIPrintf(viewer," (%D, %g) ",a->j[j],(double)PetscRealPart(a->a[j]));CHKERRQ(ierr); 847 } 848 #else 849 ierr = PetscViewerASCIIPrintf(viewer," (%D, %g) ",a->j[j],(double)a->a[j]);CHKERRQ(ierr); 850 #endif 851 } 852 /* diagonal */ 853 j = a->diag[i]; 854 #if defined(PETSC_USE_COMPLEX) 855 if (PetscImaginaryPart(a->a[j]) > 0.0) { 856 ierr = PetscViewerASCIIPrintf(viewer," (%D, %g + %g i)",a->j[j],(double)PetscRealPart(1.0/a->a[j]),(double)PetscImaginaryPart(1.0/a->a[j]));CHKERRQ(ierr); 857 } else if (PetscImaginaryPart(a->a[j]) < 0.0) { 858 ierr = PetscViewerASCIIPrintf(viewer," (%D, %g - %g i)",a->j[j],(double)PetscRealPart(1.0/a->a[j]),(double)(-PetscImaginaryPart(1.0/a->a[j])));CHKERRQ(ierr); 859 } else { 860 ierr = PetscViewerASCIIPrintf(viewer," (%D, %g) ",a->j[j],(double)PetscRealPart(1.0/a->a[j]));CHKERRQ(ierr); 861 } 862 #else 863 ierr = PetscViewerASCIIPrintf(viewer," (%D, %g) ",a->j[j],(double)(1.0/a->a[j]));CHKERRQ(ierr); 864 #endif 865 866 /* U part */ 867 for (j=a->diag[i+1]+1; j<a->diag[i]; j++) { 868 #if defined(PETSC_USE_COMPLEX) 869 if (PetscImaginaryPart(a->a[j]) > 0.0) { 870 ierr = PetscViewerASCIIPrintf(viewer," (%D, %g + %g i)",a->j[j],(double)PetscRealPart(a->a[j]),(double)PetscImaginaryPart(a->a[j]));CHKERRQ(ierr); 871 } else if (PetscImaginaryPart(a->a[j]) < 0.0) { 872 ierr = PetscViewerASCIIPrintf(viewer," (%D, %g - %g i)",a->j[j],(double)PetscRealPart(a->a[j]),(double)(-PetscImaginaryPart(a->a[j])));CHKERRQ(ierr); 873 } else { 874 ierr = PetscViewerASCIIPrintf(viewer," (%D, %g) ",a->j[j],(double)PetscRealPart(a->a[j]));CHKERRQ(ierr); 875 } 876 #else 877 ierr = PetscViewerASCIIPrintf(viewer," (%D, %g) ",a->j[j],(double)a->a[j]);CHKERRQ(ierr); 878 #endif 879 } 880 ierr = PetscViewerASCIIPrintf(viewer,"\n");CHKERRQ(ierr); 881 } 882 } else { 883 for (i=0; i<m; i++) { 884 ierr = PetscViewerASCIIPrintf(viewer,"row %D:",i);CHKERRQ(ierr); 885 for (j=a->i[i]; j<a->i[i+1]; j++) { 886 #if defined(PETSC_USE_COMPLEX) 887 if (PetscImaginaryPart(a->a[j]) > 0.0) { 888 ierr = PetscViewerASCIIPrintf(viewer," (%D, %g + %g i)",a->j[j],(double)PetscRealPart(a->a[j]),(double)PetscImaginaryPart(a->a[j]));CHKERRQ(ierr); 889 } else if (PetscImaginaryPart(a->a[j]) < 0.0) { 890 ierr = PetscViewerASCIIPrintf(viewer," (%D, %g - %g i)",a->j[j],(double)PetscRealPart(a->a[j]),(double)-PetscImaginaryPart(a->a[j]));CHKERRQ(ierr); 891 } else { 892 ierr = PetscViewerASCIIPrintf(viewer," (%D, %g) ",a->j[j],(double)PetscRealPart(a->a[j]));CHKERRQ(ierr); 893 } 894 #else 895 ierr = PetscViewerASCIIPrintf(viewer," (%D, %g) ",a->j[j],(double)a->a[j]);CHKERRQ(ierr); 896 #endif 897 } 898 ierr = PetscViewerASCIIPrintf(viewer,"\n");CHKERRQ(ierr); 899 } 900 } 901 ierr = PetscViewerASCIIUseTabs(viewer,PETSC_TRUE);CHKERRQ(ierr); 902 } 903 ierr = PetscViewerFlush(viewer);CHKERRQ(ierr); 904 PetscFunctionReturn(0); 905 } 906 907 #include <petscdraw.h> 908 PetscErrorCode MatView_SeqAIJ_Draw_Zoom(PetscDraw draw,void *Aa) 909 { 910 Mat A = (Mat) Aa; 911 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 912 PetscErrorCode ierr; 913 PetscInt i,j,m = A->rmap->n; 914 int color; 915 PetscReal xl,yl,xr,yr,x_l,x_r,y_l,y_r; 916 PetscViewer viewer; 917 PetscViewerFormat format; 918 919 PetscFunctionBegin; 920 ierr = PetscObjectQuery((PetscObject)A,"Zoomviewer",(PetscObject*)&viewer);CHKERRQ(ierr); 921 ierr = PetscViewerGetFormat(viewer,&format);CHKERRQ(ierr); 922 ierr = PetscDrawGetCoordinates(draw,&xl,&yl,&xr,&yr);CHKERRQ(ierr); 923 924 /* loop over matrix elements drawing boxes */ 925 926 if (format != PETSC_VIEWER_DRAW_CONTOUR) { 927 ierr = PetscDrawCollectiveBegin(draw);CHKERRQ(ierr); 928 /* Blue for negative, Cyan for zero and Red for positive */ 929 color = PETSC_DRAW_BLUE; 930 for (i=0; i<m; i++) { 931 y_l = m - i - 1.0; y_r = y_l + 1.0; 932 for (j=a->i[i]; j<a->i[i+1]; j++) { 933 x_l = a->j[j]; x_r = x_l + 1.0; 934 if (PetscRealPart(a->a[j]) >= 0.) continue; 935 ierr = PetscDrawRectangle(draw,x_l,y_l,x_r,y_r,color,color,color,color);CHKERRQ(ierr); 936 } 937 } 938 color = PETSC_DRAW_CYAN; 939 for (i=0; i<m; i++) { 940 y_l = m - i - 1.0; y_r = y_l + 1.0; 941 for (j=a->i[i]; j<a->i[i+1]; j++) { 942 x_l = a->j[j]; x_r = x_l + 1.0; 943 if (a->a[j] != 0.) continue; 944 ierr = PetscDrawRectangle(draw,x_l,y_l,x_r,y_r,color,color,color,color);CHKERRQ(ierr); 945 } 946 } 947 color = PETSC_DRAW_RED; 948 for (i=0; i<m; i++) { 949 y_l = m - i - 1.0; y_r = y_l + 1.0; 950 for (j=a->i[i]; j<a->i[i+1]; j++) { 951 x_l = a->j[j]; x_r = x_l + 1.0; 952 if (PetscRealPart(a->a[j]) <= 0.) continue; 953 ierr = PetscDrawRectangle(draw,x_l,y_l,x_r,y_r,color,color,color,color);CHKERRQ(ierr); 954 } 955 } 956 ierr = PetscDrawCollectiveEnd(draw);CHKERRQ(ierr); 957 } else { 958 /* use contour shading to indicate magnitude of values */ 959 /* first determine max of all nonzero values */ 960 PetscReal minv = 0.0, maxv = 0.0; 961 PetscInt nz = a->nz, count = 0; 962 PetscDraw popup; 963 964 for (i=0; i<nz; i++) { 965 if (PetscAbsScalar(a->a[i]) > maxv) maxv = PetscAbsScalar(a->a[i]); 966 } 967 if (minv >= maxv) maxv = minv + PETSC_SMALL; 968 ierr = PetscDrawGetPopup(draw,&popup);CHKERRQ(ierr); 969 ierr = PetscDrawScalePopup(popup,minv,maxv);CHKERRQ(ierr); 970 971 ierr = PetscDrawCollectiveBegin(draw);CHKERRQ(ierr); 972 for (i=0; i<m; i++) { 973 y_l = m - i - 1.0; 974 y_r = y_l + 1.0; 975 for (j=a->i[i]; j<a->i[i+1]; j++) { 976 x_l = a->j[j]; 977 x_r = x_l + 1.0; 978 color = PetscDrawRealToColor(PetscAbsScalar(a->a[count]),minv,maxv); 979 ierr = PetscDrawRectangle(draw,x_l,y_l,x_r,y_r,color,color,color,color);CHKERRQ(ierr); 980 count++; 981 } 982 } 983 ierr = PetscDrawCollectiveEnd(draw);CHKERRQ(ierr); 984 } 985 PetscFunctionReturn(0); 986 } 987 988 #include <petscdraw.h> 989 PetscErrorCode MatView_SeqAIJ_Draw(Mat A,PetscViewer viewer) 990 { 991 PetscErrorCode ierr; 992 PetscDraw draw; 993 PetscReal xr,yr,xl,yl,h,w; 994 PetscBool isnull; 995 996 PetscFunctionBegin; 997 ierr = PetscViewerDrawGetDraw(viewer,0,&draw);CHKERRQ(ierr); 998 ierr = PetscDrawIsNull(draw,&isnull);CHKERRQ(ierr); 999 if (isnull) PetscFunctionReturn(0); 1000 1001 xr = A->cmap->n; yr = A->rmap->n; h = yr/10.0; w = xr/10.0; 1002 xr += w; yr += h; xl = -w; yl = -h; 1003 ierr = PetscDrawSetCoordinates(draw,xl,yl,xr,yr);CHKERRQ(ierr); 1004 ierr = PetscObjectCompose((PetscObject)A,"Zoomviewer",(PetscObject)viewer);CHKERRQ(ierr); 1005 ierr = PetscDrawZoom(draw,MatView_SeqAIJ_Draw_Zoom,A);CHKERRQ(ierr); 1006 ierr = PetscObjectCompose((PetscObject)A,"Zoomviewer",NULL);CHKERRQ(ierr); 1007 ierr = PetscDrawSave(draw);CHKERRQ(ierr); 1008 PetscFunctionReturn(0); 1009 } 1010 1011 PetscErrorCode MatView_SeqAIJ(Mat A,PetscViewer viewer) 1012 { 1013 PetscErrorCode ierr; 1014 PetscBool iascii,isbinary,isdraw; 1015 1016 PetscFunctionBegin; 1017 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERASCII,&iascii);CHKERRQ(ierr); 1018 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERBINARY,&isbinary);CHKERRQ(ierr); 1019 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERDRAW,&isdraw);CHKERRQ(ierr); 1020 if (iascii) { 1021 ierr = MatView_SeqAIJ_ASCII(A,viewer);CHKERRQ(ierr); 1022 } else if (isbinary) { 1023 ierr = MatView_SeqAIJ_Binary(A,viewer);CHKERRQ(ierr); 1024 } else if (isdraw) { 1025 ierr = MatView_SeqAIJ_Draw(A,viewer);CHKERRQ(ierr); 1026 } 1027 ierr = MatView_SeqAIJ_Inode(A,viewer);CHKERRQ(ierr); 1028 PetscFunctionReturn(0); 1029 } 1030 1031 PetscErrorCode MatAssemblyEnd_SeqAIJ(Mat A,MatAssemblyType mode) 1032 { 1033 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 1034 PetscErrorCode ierr; 1035 PetscInt fshift = 0,i,*ai = a->i,*aj = a->j,*imax = a->imax; 1036 PetscInt m = A->rmap->n,*ip,N,*ailen = a->ilen,rmax = 0; 1037 MatScalar *aa = a->a,*ap; 1038 PetscReal ratio = 0.6; 1039 1040 PetscFunctionBegin; 1041 if (mode == MAT_FLUSH_ASSEMBLY) PetscFunctionReturn(0); 1042 ierr = MatSeqAIJInvalidateDiagonal(A);CHKERRQ(ierr); 1043 if (A->was_assembled && A->ass_nonzerostate == A->nonzerostate) PetscFunctionReturn(0); 1044 1045 if (m) rmax = ailen[0]; /* determine row with most nonzeros */ 1046 for (i=1; i<m; i++) { 1047 /* move each row back by the amount of empty slots (fshift) before it*/ 1048 fshift += imax[i-1] - ailen[i-1]; 1049 rmax = PetscMax(rmax,ailen[i]); 1050 if (fshift) { 1051 ip = aj + ai[i]; 1052 ap = aa + ai[i]; 1053 N = ailen[i]; 1054 ierr = PetscArraymove(ip-fshift,ip,N);CHKERRQ(ierr); 1055 if (!A->structure_only) { 1056 ierr = PetscArraymove(ap-fshift,ap,N);CHKERRQ(ierr); 1057 } 1058 } 1059 ai[i] = ai[i-1] + ailen[i-1]; 1060 } 1061 if (m) { 1062 fshift += imax[m-1] - ailen[m-1]; 1063 ai[m] = ai[m-1] + ailen[m-1]; 1064 } 1065 1066 /* reset ilen and imax for each row */ 1067 a->nonzerorowcnt = 0; 1068 if (A->structure_only) { 1069 ierr = PetscFree(a->imax);CHKERRQ(ierr); 1070 ierr = PetscFree(a->ilen);CHKERRQ(ierr); 1071 } else { /* !A->structure_only */ 1072 for (i=0; i<m; i++) { 1073 ailen[i] = imax[i] = ai[i+1] - ai[i]; 1074 a->nonzerorowcnt += ((ai[i+1] - ai[i]) > 0); 1075 } 1076 } 1077 a->nz = ai[m]; 1078 if (fshift && a->nounused == -1) SETERRQ3(PETSC_COMM_SELF,PETSC_ERR_PLIB, "Unused space detected in matrix: %D X %D, %D unneeded", m, A->cmap->n, fshift); 1079 1080 ierr = MatMarkDiagonal_SeqAIJ(A);CHKERRQ(ierr); 1081 ierr = PetscInfo4(A,"Matrix size: %D X %D; storage space: %D unneeded,%D used\n",m,A->cmap->n,fshift,a->nz);CHKERRQ(ierr); 1082 ierr = PetscInfo1(A,"Number of mallocs during MatSetValues() is %D\n",a->reallocs);CHKERRQ(ierr); 1083 ierr = PetscInfo1(A,"Maximum nonzeros in any row is %D\n",rmax);CHKERRQ(ierr); 1084 1085 A->info.mallocs += a->reallocs; 1086 a->reallocs = 0; 1087 A->info.nz_unneeded = (PetscReal)fshift; 1088 a->rmax = rmax; 1089 1090 if (!A->structure_only) { 1091 ierr = MatCheckCompressedRow(A,a->nonzerorowcnt,&a->compressedrow,a->i,m,ratio);CHKERRQ(ierr); 1092 } 1093 ierr = MatAssemblyEnd_SeqAIJ_Inode(A,mode);CHKERRQ(ierr); 1094 PetscFunctionReturn(0); 1095 } 1096 1097 PetscErrorCode MatRealPart_SeqAIJ(Mat A) 1098 { 1099 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 1100 PetscInt i,nz = a->nz; 1101 MatScalar *aa = a->a; 1102 PetscErrorCode ierr; 1103 1104 PetscFunctionBegin; 1105 for (i=0; i<nz; i++) aa[i] = PetscRealPart(aa[i]); 1106 ierr = MatSeqAIJInvalidateDiagonal(A);CHKERRQ(ierr); 1107 #if defined(PETSC_HAVE_VIENNACL) || defined(PETSC_HAVE_CUDA) 1108 if (A->offloadmask != PETSC_OFFLOAD_UNALLOCATED) A->offloadmask = PETSC_OFFLOAD_CPU; 1109 #endif 1110 PetscFunctionReturn(0); 1111 } 1112 1113 PetscErrorCode MatImaginaryPart_SeqAIJ(Mat A) 1114 { 1115 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 1116 PetscInt i,nz = a->nz; 1117 MatScalar *aa = a->a; 1118 PetscErrorCode ierr; 1119 1120 PetscFunctionBegin; 1121 for (i=0; i<nz; i++) aa[i] = PetscImaginaryPart(aa[i]); 1122 ierr = MatSeqAIJInvalidateDiagonal(A);CHKERRQ(ierr); 1123 #if defined(PETSC_HAVE_VIENNACL) || defined(PETSC_HAVE_CUDA) 1124 if (A->offloadmask != PETSC_OFFLOAD_UNALLOCATED) A->offloadmask = PETSC_OFFLOAD_CPU; 1125 #endif 1126 PetscFunctionReturn(0); 1127 } 1128 1129 PetscErrorCode MatZeroEntries_SeqAIJ(Mat A) 1130 { 1131 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 1132 PetscErrorCode ierr; 1133 1134 PetscFunctionBegin; 1135 ierr = PetscArrayzero(a->a,a->i[A->rmap->n]);CHKERRQ(ierr); 1136 ierr = MatSeqAIJInvalidateDiagonal(A);CHKERRQ(ierr); 1137 #if defined(PETSC_HAVE_VIENNACL) || defined(PETSC_HAVE_CUDA) 1138 if (A->offloadmask != PETSC_OFFLOAD_UNALLOCATED) A->offloadmask = PETSC_OFFLOAD_CPU; 1139 #endif 1140 PetscFunctionReturn(0); 1141 } 1142 1143 PetscErrorCode MatDestroy_SeqAIJ(Mat A) 1144 { 1145 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 1146 PetscErrorCode ierr; 1147 1148 PetscFunctionBegin; 1149 #if defined(PETSC_USE_LOG) 1150 PetscLogObjectState((PetscObject)A,"Rows=%D, Cols=%D, NZ=%D",A->rmap->n,A->cmap->n,a->nz); 1151 #endif 1152 ierr = MatSeqXAIJFreeAIJ(A,&a->a,&a->j,&a->i);CHKERRQ(ierr); 1153 ierr = ISDestroy(&a->row);CHKERRQ(ierr); 1154 ierr = ISDestroy(&a->col);CHKERRQ(ierr); 1155 ierr = PetscFree(a->diag);CHKERRQ(ierr); 1156 ierr = PetscFree(a->ibdiag);CHKERRQ(ierr); 1157 ierr = PetscFree(a->imax);CHKERRQ(ierr); 1158 ierr = PetscFree(a->ilen);CHKERRQ(ierr); 1159 ierr = PetscFree(a->ipre);CHKERRQ(ierr); 1160 ierr = PetscFree3(a->idiag,a->mdiag,a->ssor_work);CHKERRQ(ierr); 1161 ierr = PetscFree(a->solve_work);CHKERRQ(ierr); 1162 ierr = ISDestroy(&a->icol);CHKERRQ(ierr); 1163 ierr = PetscFree(a->saved_values);CHKERRQ(ierr); 1164 ierr = ISColoringDestroy(&a->coloring);CHKERRQ(ierr); 1165 ierr = PetscFree2(a->compressedrow.i,a->compressedrow.rindex);CHKERRQ(ierr); 1166 ierr = PetscFree(a->matmult_abdense);CHKERRQ(ierr); 1167 1168 ierr = MatDestroy_SeqAIJ_Inode(A);CHKERRQ(ierr); 1169 ierr = PetscFree(A->data);CHKERRQ(ierr); 1170 1171 ierr = PetscObjectChangeTypeName((PetscObject)A,0);CHKERRQ(ierr); 1172 ierr = PetscObjectComposeFunction((PetscObject)A,"MatSeqAIJSetColumnIndices_C",NULL);CHKERRQ(ierr); 1173 ierr = PetscObjectComposeFunction((PetscObject)A,"MatStoreValues_C",NULL);CHKERRQ(ierr); 1174 ierr = PetscObjectComposeFunction((PetscObject)A,"MatRetrieveValues_C",NULL);CHKERRQ(ierr); 1175 ierr = PetscObjectComposeFunction((PetscObject)A,"MatConvert_seqaij_seqsbaij_C",NULL);CHKERRQ(ierr); 1176 ierr = PetscObjectComposeFunction((PetscObject)A,"MatConvert_seqaij_seqbaij_C",NULL);CHKERRQ(ierr); 1177 ierr = PetscObjectComposeFunction((PetscObject)A,"MatConvert_seqaij_seqaijperm_C",NULL);CHKERRQ(ierr); 1178 #if defined(PETSC_HAVE_ELEMENTAL) 1179 ierr = PetscObjectComposeFunction((PetscObject)A,"MatConvert_seqaij_elemental_C",NULL);CHKERRQ(ierr); 1180 #endif 1181 #if defined(PETSC_HAVE_HYPRE) 1182 ierr = PetscObjectComposeFunction((PetscObject)A,"MatConvert_seqaij_hypre_C",NULL);CHKERRQ(ierr); 1183 ierr = PetscObjectComposeFunction((PetscObject)A,"MatMatMatMult_transpose_seqaij_seqaij_C",NULL);CHKERRQ(ierr); 1184 #endif 1185 ierr = PetscObjectComposeFunction((PetscObject)A,"MatConvert_seqaij_seqdense_C",NULL);CHKERRQ(ierr); 1186 ierr = PetscObjectComposeFunction((PetscObject)A,"MatConvert_seqaij_seqsell_C",NULL);CHKERRQ(ierr); 1187 ierr = PetscObjectComposeFunction((PetscObject)A,"MatConvert_seqaij_is_C",NULL);CHKERRQ(ierr); 1188 ierr = PetscObjectComposeFunction((PetscObject)A,"MatIsTranspose_C",NULL);CHKERRQ(ierr); 1189 ierr = PetscObjectComposeFunction((PetscObject)A,"MatSeqAIJSetPreallocation_C",NULL);CHKERRQ(ierr); 1190 ierr = PetscObjectComposeFunction((PetscObject)A,"MatResetPreallocation_C",NULL);CHKERRQ(ierr); 1191 ierr = PetscObjectComposeFunction((PetscObject)A,"MatSeqAIJSetPreallocationCSR_C",NULL);CHKERRQ(ierr); 1192 ierr = PetscObjectComposeFunction((PetscObject)A,"MatReorderForNonzeroDiagonal_C",NULL);CHKERRQ(ierr); 1193 ierr = PetscObjectComposeFunction((PetscObject)A,"MatPtAP_is_seqaij_C",NULL);CHKERRQ(ierr); 1194 PetscFunctionReturn(0); 1195 } 1196 1197 PetscErrorCode MatSetOption_SeqAIJ(Mat A,MatOption op,PetscBool flg) 1198 { 1199 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 1200 PetscErrorCode ierr; 1201 1202 PetscFunctionBegin; 1203 switch (op) { 1204 case MAT_ROW_ORIENTED: 1205 a->roworiented = flg; 1206 break; 1207 case MAT_KEEP_NONZERO_PATTERN: 1208 a->keepnonzeropattern = flg; 1209 break; 1210 case MAT_NEW_NONZERO_LOCATIONS: 1211 a->nonew = (flg ? 0 : 1); 1212 break; 1213 case MAT_NEW_NONZERO_LOCATION_ERR: 1214 a->nonew = (flg ? -1 : 0); 1215 break; 1216 case MAT_NEW_NONZERO_ALLOCATION_ERR: 1217 a->nonew = (flg ? -2 : 0); 1218 break; 1219 case MAT_UNUSED_NONZERO_LOCATION_ERR: 1220 a->nounused = (flg ? -1 : 0); 1221 break; 1222 case MAT_IGNORE_ZERO_ENTRIES: 1223 a->ignorezeroentries = flg; 1224 break; 1225 case MAT_SPD: 1226 case MAT_SYMMETRIC: 1227 case MAT_STRUCTURALLY_SYMMETRIC: 1228 case MAT_HERMITIAN: 1229 case MAT_SYMMETRY_ETERNAL: 1230 case MAT_STRUCTURE_ONLY: 1231 /* These options are handled directly by MatSetOption() */ 1232 break; 1233 case MAT_NEW_DIAGONALS: 1234 case MAT_IGNORE_OFF_PROC_ENTRIES: 1235 case MAT_USE_HASH_TABLE: 1236 ierr = PetscInfo1(A,"Option %s ignored\n",MatOptions[op]);CHKERRQ(ierr); 1237 break; 1238 case MAT_USE_INODES: 1239 /* Not an error because MatSetOption_SeqAIJ_Inode handles this one */ 1240 break; 1241 case MAT_SUBMAT_SINGLEIS: 1242 A->submat_singleis = flg; 1243 break; 1244 case MAT_SORTED_FULL: 1245 if (flg) A->ops->setvalues = MatSetValues_SeqAIJ_SortedFull; 1246 else A->ops->setvalues = MatSetValues_SeqAIJ; 1247 break; 1248 default: 1249 SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_SUP,"unknown option %d",op); 1250 } 1251 ierr = MatSetOption_SeqAIJ_Inode(A,op,flg);CHKERRQ(ierr); 1252 PetscFunctionReturn(0); 1253 } 1254 1255 PetscErrorCode MatGetDiagonal_SeqAIJ(Mat A,Vec v) 1256 { 1257 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 1258 PetscErrorCode ierr; 1259 PetscInt i,j,n,*ai=a->i,*aj=a->j; 1260 PetscScalar *aa=a->a,*x; 1261 1262 PetscFunctionBegin; 1263 ierr = VecGetLocalSize(v,&n);CHKERRQ(ierr); 1264 if (n != A->rmap->n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Nonconforming matrix and vector"); 1265 1266 if (A->factortype == MAT_FACTOR_ILU || A->factortype == MAT_FACTOR_LU) { 1267 PetscInt *diag=a->diag; 1268 ierr = VecGetArrayWrite(v,&x);CHKERRQ(ierr); 1269 for (i=0; i<n; i++) x[i] = 1.0/aa[diag[i]]; 1270 ierr = VecRestoreArrayWrite(v,&x);CHKERRQ(ierr); 1271 PetscFunctionReturn(0); 1272 } 1273 1274 ierr = VecGetArrayWrite(v,&x);CHKERRQ(ierr); 1275 for (i=0; i<n; i++) { 1276 x[i] = 0.0; 1277 for (j=ai[i]; j<ai[i+1]; j++) { 1278 if (aj[j] == i) { 1279 x[i] = aa[j]; 1280 break; 1281 } 1282 } 1283 } 1284 ierr = VecRestoreArrayWrite(v,&x);CHKERRQ(ierr); 1285 PetscFunctionReturn(0); 1286 } 1287 1288 #include <../src/mat/impls/aij/seq/ftn-kernels/fmult.h> 1289 PetscErrorCode MatMultTransposeAdd_SeqAIJ(Mat A,Vec xx,Vec zz,Vec yy) 1290 { 1291 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 1292 PetscScalar *y; 1293 const PetscScalar *x; 1294 PetscErrorCode ierr; 1295 PetscInt m = A->rmap->n; 1296 #if !defined(PETSC_USE_FORTRAN_KERNEL_MULTTRANSPOSEAIJ) 1297 const MatScalar *v; 1298 PetscScalar alpha; 1299 PetscInt n,i,j; 1300 const PetscInt *idx,*ii,*ridx=NULL; 1301 Mat_CompressedRow cprow = a->compressedrow; 1302 PetscBool usecprow = cprow.use; 1303 #endif 1304 1305 PetscFunctionBegin; 1306 if (zz != yy) {ierr = VecCopy(zz,yy);CHKERRQ(ierr);} 1307 ierr = VecGetArrayRead(xx,&x);CHKERRQ(ierr); 1308 ierr = VecGetArray(yy,&y);CHKERRQ(ierr); 1309 1310 #if defined(PETSC_USE_FORTRAN_KERNEL_MULTTRANSPOSEAIJ) 1311 fortranmulttransposeaddaij_(&m,x,a->i,a->j,a->a,y); 1312 #else 1313 if (usecprow) { 1314 m = cprow.nrows; 1315 ii = cprow.i; 1316 ridx = cprow.rindex; 1317 } else { 1318 ii = a->i; 1319 } 1320 for (i=0; i<m; i++) { 1321 idx = a->j + ii[i]; 1322 v = a->a + ii[i]; 1323 n = ii[i+1] - ii[i]; 1324 if (usecprow) { 1325 alpha = x[ridx[i]]; 1326 } else { 1327 alpha = x[i]; 1328 } 1329 for (j=0; j<n; j++) y[idx[j]] += alpha*v[j]; 1330 } 1331 #endif 1332 ierr = PetscLogFlops(2.0*a->nz);CHKERRQ(ierr); 1333 ierr = VecRestoreArrayRead(xx,&x);CHKERRQ(ierr); 1334 ierr = VecRestoreArray(yy,&y);CHKERRQ(ierr); 1335 PetscFunctionReturn(0); 1336 } 1337 1338 PetscErrorCode MatMultTranspose_SeqAIJ(Mat A,Vec xx,Vec yy) 1339 { 1340 PetscErrorCode ierr; 1341 1342 PetscFunctionBegin; 1343 ierr = VecSet(yy,0.0);CHKERRQ(ierr); 1344 ierr = MatMultTransposeAdd_SeqAIJ(A,xx,yy,yy);CHKERRQ(ierr); 1345 PetscFunctionReturn(0); 1346 } 1347 1348 #include <../src/mat/impls/aij/seq/ftn-kernels/fmult.h> 1349 1350 PetscErrorCode MatMult_SeqAIJ(Mat A,Vec xx,Vec yy) 1351 { 1352 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 1353 PetscScalar *y; 1354 const PetscScalar *x; 1355 const MatScalar *aa; 1356 PetscErrorCode ierr; 1357 PetscInt m=A->rmap->n; 1358 const PetscInt *aj,*ii,*ridx=NULL; 1359 PetscInt n,i; 1360 PetscScalar sum; 1361 PetscBool usecprow=a->compressedrow.use; 1362 1363 #if defined(PETSC_HAVE_PRAGMA_DISJOINT) 1364 #pragma disjoint(*x,*y,*aa) 1365 #endif 1366 1367 PetscFunctionBegin; 1368 ierr = VecGetArrayRead(xx,&x);CHKERRQ(ierr); 1369 ierr = VecGetArray(yy,&y);CHKERRQ(ierr); 1370 ii = a->i; 1371 if (usecprow) { /* use compressed row format */ 1372 ierr = PetscArrayzero(y,m);CHKERRQ(ierr); 1373 m = a->compressedrow.nrows; 1374 ii = a->compressedrow.i; 1375 ridx = a->compressedrow.rindex; 1376 for (i=0; i<m; i++) { 1377 n = ii[i+1] - ii[i]; 1378 aj = a->j + ii[i]; 1379 aa = a->a + ii[i]; 1380 sum = 0.0; 1381 PetscSparseDensePlusDot(sum,x,aa,aj,n); 1382 /* for (j=0; j<n; j++) sum += (*aa++)*x[*aj++]; */ 1383 y[*ridx++] = sum; 1384 } 1385 } else { /* do not use compressed row format */ 1386 #if defined(PETSC_USE_FORTRAN_KERNEL_MULTAIJ) 1387 aj = a->j; 1388 aa = a->a; 1389 fortranmultaij_(&m,x,ii,aj,aa,y); 1390 #else 1391 for (i=0; i<m; i++) { 1392 n = ii[i+1] - ii[i]; 1393 aj = a->j + ii[i]; 1394 aa = a->a + ii[i]; 1395 sum = 0.0; 1396 PetscSparseDensePlusDot(sum,x,aa,aj,n); 1397 y[i] = sum; 1398 } 1399 #endif 1400 } 1401 ierr = PetscLogFlops(2.0*a->nz - a->nonzerorowcnt);CHKERRQ(ierr); 1402 ierr = VecRestoreArrayRead(xx,&x);CHKERRQ(ierr); 1403 ierr = VecRestoreArray(yy,&y);CHKERRQ(ierr); 1404 PetscFunctionReturn(0); 1405 } 1406 1407 PetscErrorCode MatMultMax_SeqAIJ(Mat A,Vec xx,Vec yy) 1408 { 1409 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 1410 PetscScalar *y; 1411 const PetscScalar *x; 1412 const MatScalar *aa; 1413 PetscErrorCode ierr; 1414 PetscInt m=A->rmap->n; 1415 const PetscInt *aj,*ii,*ridx=NULL; 1416 PetscInt n,i,nonzerorow=0; 1417 PetscScalar sum; 1418 PetscBool usecprow=a->compressedrow.use; 1419 1420 #if defined(PETSC_HAVE_PRAGMA_DISJOINT) 1421 #pragma disjoint(*x,*y,*aa) 1422 #endif 1423 1424 PetscFunctionBegin; 1425 ierr = VecGetArrayRead(xx,&x);CHKERRQ(ierr); 1426 ierr = VecGetArray(yy,&y);CHKERRQ(ierr); 1427 if (usecprow) { /* use compressed row format */ 1428 m = a->compressedrow.nrows; 1429 ii = a->compressedrow.i; 1430 ridx = a->compressedrow.rindex; 1431 for (i=0; i<m; i++) { 1432 n = ii[i+1] - ii[i]; 1433 aj = a->j + ii[i]; 1434 aa = a->a + ii[i]; 1435 sum = 0.0; 1436 nonzerorow += (n>0); 1437 PetscSparseDenseMaxDot(sum,x,aa,aj,n); 1438 /* for (j=0; j<n; j++) sum += (*aa++)*x[*aj++]; */ 1439 y[*ridx++] = sum; 1440 } 1441 } else { /* do not use compressed row format */ 1442 ii = a->i; 1443 for (i=0; i<m; i++) { 1444 n = ii[i+1] - ii[i]; 1445 aj = a->j + ii[i]; 1446 aa = a->a + ii[i]; 1447 sum = 0.0; 1448 nonzerorow += (n>0); 1449 PetscSparseDenseMaxDot(sum,x,aa,aj,n); 1450 y[i] = sum; 1451 } 1452 } 1453 ierr = PetscLogFlops(2.0*a->nz - nonzerorow);CHKERRQ(ierr); 1454 ierr = VecRestoreArrayRead(xx,&x);CHKERRQ(ierr); 1455 ierr = VecRestoreArray(yy,&y);CHKERRQ(ierr); 1456 PetscFunctionReturn(0); 1457 } 1458 1459 PetscErrorCode MatMultAddMax_SeqAIJ(Mat A,Vec xx,Vec yy,Vec zz) 1460 { 1461 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 1462 PetscScalar *y,*z; 1463 const PetscScalar *x; 1464 const MatScalar *aa; 1465 PetscErrorCode ierr; 1466 PetscInt m = A->rmap->n,*aj,*ii; 1467 PetscInt n,i,*ridx=NULL; 1468 PetscScalar sum; 1469 PetscBool usecprow=a->compressedrow.use; 1470 1471 PetscFunctionBegin; 1472 ierr = VecGetArrayRead(xx,&x);CHKERRQ(ierr); 1473 ierr = VecGetArrayPair(yy,zz,&y,&z);CHKERRQ(ierr); 1474 if (usecprow) { /* use compressed row format */ 1475 if (zz != yy) { 1476 ierr = PetscArraycpy(z,y,m);CHKERRQ(ierr); 1477 } 1478 m = a->compressedrow.nrows; 1479 ii = a->compressedrow.i; 1480 ridx = a->compressedrow.rindex; 1481 for (i=0; i<m; i++) { 1482 n = ii[i+1] - ii[i]; 1483 aj = a->j + ii[i]; 1484 aa = a->a + ii[i]; 1485 sum = y[*ridx]; 1486 PetscSparseDenseMaxDot(sum,x,aa,aj,n); 1487 z[*ridx++] = sum; 1488 } 1489 } else { /* do not use compressed row format */ 1490 ii = a->i; 1491 for (i=0; i<m; i++) { 1492 n = ii[i+1] - ii[i]; 1493 aj = a->j + ii[i]; 1494 aa = a->a + ii[i]; 1495 sum = y[i]; 1496 PetscSparseDenseMaxDot(sum,x,aa,aj,n); 1497 z[i] = sum; 1498 } 1499 } 1500 ierr = PetscLogFlops(2.0*a->nz);CHKERRQ(ierr); 1501 ierr = VecRestoreArrayRead(xx,&x);CHKERRQ(ierr); 1502 ierr = VecRestoreArrayPair(yy,zz,&y,&z);CHKERRQ(ierr); 1503 PetscFunctionReturn(0); 1504 } 1505 1506 #include <../src/mat/impls/aij/seq/ftn-kernels/fmultadd.h> 1507 PetscErrorCode MatMultAdd_SeqAIJ(Mat A,Vec xx,Vec yy,Vec zz) 1508 { 1509 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 1510 PetscScalar *y,*z; 1511 const PetscScalar *x; 1512 const MatScalar *aa; 1513 PetscErrorCode ierr; 1514 const PetscInt *aj,*ii,*ridx=NULL; 1515 PetscInt m = A->rmap->n,n,i; 1516 PetscScalar sum; 1517 PetscBool usecprow=a->compressedrow.use; 1518 1519 PetscFunctionBegin; 1520 ierr = VecGetArrayRead(xx,&x);CHKERRQ(ierr); 1521 ierr = VecGetArrayPair(yy,zz,&y,&z);CHKERRQ(ierr); 1522 if (usecprow) { /* use compressed row format */ 1523 if (zz != yy) { 1524 ierr = PetscArraycpy(z,y,m);CHKERRQ(ierr); 1525 } 1526 m = a->compressedrow.nrows; 1527 ii = a->compressedrow.i; 1528 ridx = a->compressedrow.rindex; 1529 for (i=0; i<m; i++) { 1530 n = ii[i+1] - ii[i]; 1531 aj = a->j + ii[i]; 1532 aa = a->a + ii[i]; 1533 sum = y[*ridx]; 1534 PetscSparseDensePlusDot(sum,x,aa,aj,n); 1535 z[*ridx++] = sum; 1536 } 1537 } else { /* do not use compressed row format */ 1538 ii = a->i; 1539 #if defined(PETSC_USE_FORTRAN_KERNEL_MULTADDAIJ) 1540 aj = a->j; 1541 aa = a->a; 1542 fortranmultaddaij_(&m,x,ii,aj,aa,y,z); 1543 #else 1544 for (i=0; i<m; i++) { 1545 n = ii[i+1] - ii[i]; 1546 aj = a->j + ii[i]; 1547 aa = a->a + ii[i]; 1548 sum = y[i]; 1549 PetscSparseDensePlusDot(sum,x,aa,aj,n); 1550 z[i] = sum; 1551 } 1552 #endif 1553 } 1554 ierr = PetscLogFlops(2.0*a->nz);CHKERRQ(ierr); 1555 ierr = VecRestoreArrayRead(xx,&x);CHKERRQ(ierr); 1556 ierr = VecRestoreArrayPair(yy,zz,&y,&z);CHKERRQ(ierr); 1557 PetscFunctionReturn(0); 1558 } 1559 1560 /* 1561 Adds diagonal pointers to sparse matrix structure. 1562 */ 1563 PetscErrorCode MatMarkDiagonal_SeqAIJ(Mat A) 1564 { 1565 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 1566 PetscErrorCode ierr; 1567 PetscInt i,j,m = A->rmap->n; 1568 1569 PetscFunctionBegin; 1570 if (!a->diag) { 1571 ierr = PetscMalloc1(m,&a->diag);CHKERRQ(ierr); 1572 ierr = PetscLogObjectMemory((PetscObject)A, m*sizeof(PetscInt));CHKERRQ(ierr); 1573 } 1574 for (i=0; i<A->rmap->n; i++) { 1575 a->diag[i] = a->i[i+1]; 1576 for (j=a->i[i]; j<a->i[i+1]; j++) { 1577 if (a->j[j] == i) { 1578 a->diag[i] = j; 1579 break; 1580 } 1581 } 1582 } 1583 PetscFunctionReturn(0); 1584 } 1585 1586 PetscErrorCode MatShift_SeqAIJ(Mat A,PetscScalar v) 1587 { 1588 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 1589 const PetscInt *diag = (const PetscInt*)a->diag; 1590 const PetscInt *ii = (const PetscInt*) a->i; 1591 PetscInt i,*mdiag = NULL; 1592 PetscErrorCode ierr; 1593 PetscInt cnt = 0; /* how many diagonals are missing */ 1594 1595 PetscFunctionBegin; 1596 if (!A->preallocated || !a->nz) { 1597 ierr = MatSeqAIJSetPreallocation(A,1,NULL);CHKERRQ(ierr); 1598 ierr = MatShift_Basic(A,v);CHKERRQ(ierr); 1599 PetscFunctionReturn(0); 1600 } 1601 1602 if (a->diagonaldense) { 1603 cnt = 0; 1604 } else { 1605 ierr = PetscCalloc1(A->rmap->n,&mdiag);CHKERRQ(ierr); 1606 for (i=0; i<A->rmap->n; i++) { 1607 if (diag[i] >= ii[i+1]) { 1608 cnt++; 1609 mdiag[i] = 1; 1610 } 1611 } 1612 } 1613 if (!cnt) { 1614 ierr = MatShift_Basic(A,v);CHKERRQ(ierr); 1615 } else { 1616 PetscScalar *olda = a->a; /* preserve pointers to current matrix nonzeros structure and values */ 1617 PetscInt *oldj = a->j, *oldi = a->i; 1618 PetscBool singlemalloc = a->singlemalloc,free_a = a->free_a,free_ij = a->free_ij; 1619 1620 a->a = NULL; 1621 a->j = NULL; 1622 a->i = NULL; 1623 /* increase the values in imax for each row where a diagonal is being inserted then reallocate the matrix data structures */ 1624 for (i=0; i<A->rmap->n; i++) { 1625 a->imax[i] += mdiag[i]; 1626 a->imax[i] = PetscMin(a->imax[i],A->cmap->n); 1627 } 1628 ierr = MatSeqAIJSetPreallocation_SeqAIJ(A,0,a->imax);CHKERRQ(ierr); 1629 1630 /* copy old values into new matrix data structure */ 1631 for (i=0; i<A->rmap->n; i++) { 1632 ierr = MatSetValues(A,1,&i,a->imax[i] - mdiag[i],&oldj[oldi[i]],&olda[oldi[i]],ADD_VALUES);CHKERRQ(ierr); 1633 if (i < A->cmap->n) { 1634 ierr = MatSetValue(A,i,i,v,ADD_VALUES);CHKERRQ(ierr); 1635 } 1636 } 1637 ierr = MatAssemblyBegin(A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1638 ierr = MatAssemblyEnd(A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1639 if (singlemalloc) { 1640 ierr = PetscFree3(olda,oldj,oldi);CHKERRQ(ierr); 1641 } else { 1642 if (free_a) {ierr = PetscFree(olda);CHKERRQ(ierr);} 1643 if (free_ij) {ierr = PetscFree(oldj);CHKERRQ(ierr);} 1644 if (free_ij) {ierr = PetscFree(oldi);CHKERRQ(ierr);} 1645 } 1646 } 1647 ierr = PetscFree(mdiag);CHKERRQ(ierr); 1648 a->diagonaldense = PETSC_TRUE; 1649 PetscFunctionReturn(0); 1650 } 1651 1652 /* 1653 Checks for missing diagonals 1654 */ 1655 PetscErrorCode MatMissingDiagonal_SeqAIJ(Mat A,PetscBool *missing,PetscInt *d) 1656 { 1657 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 1658 PetscInt *diag,*ii = a->i,i; 1659 PetscErrorCode ierr; 1660 1661 PetscFunctionBegin; 1662 *missing = PETSC_FALSE; 1663 if (A->rmap->n > 0 && !ii) { 1664 *missing = PETSC_TRUE; 1665 if (d) *d = 0; 1666 ierr = PetscInfo(A,"Matrix has no entries therefore is missing diagonal\n");CHKERRQ(ierr); 1667 } else { 1668 PetscInt n; 1669 n = PetscMin(A->rmap->n, A->cmap->n); 1670 diag = a->diag; 1671 for (i=0; i<n; i++) { 1672 if (diag[i] >= ii[i+1]) { 1673 *missing = PETSC_TRUE; 1674 if (d) *d = i; 1675 ierr = PetscInfo1(A,"Matrix is missing diagonal number %D\n",i);CHKERRQ(ierr); 1676 break; 1677 } 1678 } 1679 } 1680 PetscFunctionReturn(0); 1681 } 1682 1683 #include <petscblaslapack.h> 1684 #include <petsc/private/kernels/blockinvert.h> 1685 1686 /* 1687 Note that values is allocated externally by the PC and then passed into this routine 1688 */ 1689 PetscErrorCode MatInvertVariableBlockDiagonal_SeqAIJ(Mat A,PetscInt nblocks,const PetscInt *bsizes,PetscScalar *diag) 1690 { 1691 PetscErrorCode ierr; 1692 PetscInt n = A->rmap->n, i, ncnt = 0, *indx,j,bsizemax = 0,*v_pivots; 1693 PetscBool allowzeropivot,zeropivotdetected=PETSC_FALSE; 1694 const PetscReal shift = 0.0; 1695 PetscInt ipvt[5]; 1696 PetscScalar work[25],*v_work; 1697 1698 PetscFunctionBegin; 1699 allowzeropivot = PetscNot(A->erroriffailure); 1700 for (i=0; i<nblocks; i++) ncnt += bsizes[i]; 1701 if (ncnt != n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Total blocksizes %D doesn't match number matrix rows %D",ncnt,n); 1702 for (i=0; i<nblocks; i++) { 1703 bsizemax = PetscMax(bsizemax,bsizes[i]); 1704 } 1705 ierr = PetscMalloc1(bsizemax,&indx);CHKERRQ(ierr); 1706 if (bsizemax > 7) { 1707 ierr = PetscMalloc2(bsizemax,&v_work,bsizemax,&v_pivots);CHKERRQ(ierr); 1708 } 1709 ncnt = 0; 1710 for (i=0; i<nblocks; i++) { 1711 for (j=0; j<bsizes[i]; j++) indx[j] = ncnt+j; 1712 ierr = MatGetValues(A,bsizes[i],indx,bsizes[i],indx,diag);CHKERRQ(ierr); 1713 switch (bsizes[i]) { 1714 case 1: 1715 *diag = 1.0/(*diag); 1716 break; 1717 case 2: 1718 ierr = PetscKernel_A_gets_inverse_A_2(diag,shift,allowzeropivot,&zeropivotdetected);CHKERRQ(ierr); 1719 if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT; 1720 ierr = PetscKernel_A_gets_transpose_A_2(diag);CHKERRQ(ierr); 1721 break; 1722 case 3: 1723 ierr = PetscKernel_A_gets_inverse_A_3(diag,shift,allowzeropivot,&zeropivotdetected);CHKERRQ(ierr); 1724 if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT; 1725 ierr = PetscKernel_A_gets_transpose_A_3(diag);CHKERRQ(ierr); 1726 break; 1727 case 4: 1728 ierr = PetscKernel_A_gets_inverse_A_4(diag,shift,allowzeropivot,&zeropivotdetected);CHKERRQ(ierr); 1729 if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT; 1730 ierr = PetscKernel_A_gets_transpose_A_4(diag);CHKERRQ(ierr); 1731 break; 1732 case 5: 1733 ierr = PetscKernel_A_gets_inverse_A_5(diag,ipvt,work,shift,allowzeropivot,&zeropivotdetected);CHKERRQ(ierr); 1734 if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT; 1735 ierr = PetscKernel_A_gets_transpose_A_5(diag);CHKERRQ(ierr); 1736 break; 1737 case 6: 1738 ierr = PetscKernel_A_gets_inverse_A_6(diag,shift,allowzeropivot,&zeropivotdetected);CHKERRQ(ierr); 1739 if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT; 1740 ierr = PetscKernel_A_gets_transpose_A_6(diag);CHKERRQ(ierr); 1741 break; 1742 case 7: 1743 ierr = PetscKernel_A_gets_inverse_A_7(diag,shift,allowzeropivot,&zeropivotdetected);CHKERRQ(ierr); 1744 if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT; 1745 ierr = PetscKernel_A_gets_transpose_A_7(diag);CHKERRQ(ierr); 1746 break; 1747 default: 1748 ierr = PetscKernel_A_gets_inverse_A(bsizes[i],diag,v_pivots,v_work,allowzeropivot,&zeropivotdetected);CHKERRQ(ierr); 1749 if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT; 1750 ierr = PetscKernel_A_gets_transpose_A_N(diag,bsizes[i]);CHKERRQ(ierr); 1751 } 1752 ncnt += bsizes[i]; 1753 diag += bsizes[i]*bsizes[i]; 1754 } 1755 if (bsizemax > 7) { 1756 ierr = PetscFree2(v_work,v_pivots);CHKERRQ(ierr); 1757 } 1758 ierr = PetscFree(indx);CHKERRQ(ierr); 1759 PetscFunctionReturn(0); 1760 } 1761 1762 /* 1763 Negative shift indicates do not generate an error if there is a zero diagonal, just invert it anyways 1764 */ 1765 PetscErrorCode MatInvertDiagonal_SeqAIJ(Mat A,PetscScalar omega,PetscScalar fshift) 1766 { 1767 Mat_SeqAIJ *a = (Mat_SeqAIJ*) A->data; 1768 PetscErrorCode ierr; 1769 PetscInt i,*diag,m = A->rmap->n; 1770 MatScalar *v = a->a; 1771 PetscScalar *idiag,*mdiag; 1772 1773 PetscFunctionBegin; 1774 if (a->idiagvalid) PetscFunctionReturn(0); 1775 ierr = MatMarkDiagonal_SeqAIJ(A);CHKERRQ(ierr); 1776 diag = a->diag; 1777 if (!a->idiag) { 1778 ierr = PetscMalloc3(m,&a->idiag,m,&a->mdiag,m,&a->ssor_work);CHKERRQ(ierr); 1779 ierr = PetscLogObjectMemory((PetscObject)A, 3*m*sizeof(PetscScalar));CHKERRQ(ierr); 1780 v = a->a; 1781 } 1782 mdiag = a->mdiag; 1783 idiag = a->idiag; 1784 1785 if (omega == 1.0 && PetscRealPart(fshift) <= 0.0) { 1786 for (i=0; i<m; i++) { 1787 mdiag[i] = v[diag[i]]; 1788 if (!PetscAbsScalar(mdiag[i])) { /* zero diagonal */ 1789 if (PetscRealPart(fshift)) { 1790 ierr = PetscInfo1(A,"Zero diagonal on row %D\n",i);CHKERRQ(ierr); 1791 A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT; 1792 A->factorerror_zeropivot_value = 0.0; 1793 A->factorerror_zeropivot_row = i; 1794 } else SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_INCOMP,"Zero diagonal on row %D",i); 1795 } 1796 idiag[i] = 1.0/v[diag[i]]; 1797 } 1798 ierr = PetscLogFlops(m);CHKERRQ(ierr); 1799 } else { 1800 for (i=0; i<m; i++) { 1801 mdiag[i] = v[diag[i]]; 1802 idiag[i] = omega/(fshift + v[diag[i]]); 1803 } 1804 ierr = PetscLogFlops(2.0*m);CHKERRQ(ierr); 1805 } 1806 a->idiagvalid = PETSC_TRUE; 1807 PetscFunctionReturn(0); 1808 } 1809 1810 #include <../src/mat/impls/aij/seq/ftn-kernels/frelax.h> 1811 PetscErrorCode MatSOR_SeqAIJ(Mat A,Vec bb,PetscReal omega,MatSORType flag,PetscReal fshift,PetscInt its,PetscInt lits,Vec xx) 1812 { 1813 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 1814 PetscScalar *x,d,sum,*t,scale; 1815 const MatScalar *v,*idiag=0,*mdiag; 1816 const PetscScalar *b, *bs,*xb, *ts; 1817 PetscErrorCode ierr; 1818 PetscInt n,m = A->rmap->n,i; 1819 const PetscInt *idx,*diag; 1820 1821 PetscFunctionBegin; 1822 its = its*lits; 1823 1824 if (fshift != a->fshift || omega != a->omega) a->idiagvalid = PETSC_FALSE; /* must recompute idiag[] */ 1825 if (!a->idiagvalid) {ierr = MatInvertDiagonal_SeqAIJ(A,omega,fshift);CHKERRQ(ierr);} 1826 a->fshift = fshift; 1827 a->omega = omega; 1828 1829 diag = a->diag; 1830 t = a->ssor_work; 1831 idiag = a->idiag; 1832 mdiag = a->mdiag; 1833 1834 ierr = VecGetArray(xx,&x);CHKERRQ(ierr); 1835 ierr = VecGetArrayRead(bb,&b);CHKERRQ(ierr); 1836 /* We count flops by assuming the upper triangular and lower triangular parts have the same number of nonzeros */ 1837 if (flag == SOR_APPLY_UPPER) { 1838 /* apply (U + D/omega) to the vector */ 1839 bs = b; 1840 for (i=0; i<m; i++) { 1841 d = fshift + mdiag[i]; 1842 n = a->i[i+1] - diag[i] - 1; 1843 idx = a->j + diag[i] + 1; 1844 v = a->a + diag[i] + 1; 1845 sum = b[i]*d/omega; 1846 PetscSparseDensePlusDot(sum,bs,v,idx,n); 1847 x[i] = sum; 1848 } 1849 ierr = VecRestoreArray(xx,&x);CHKERRQ(ierr); 1850 ierr = VecRestoreArrayRead(bb,&b);CHKERRQ(ierr); 1851 ierr = PetscLogFlops(a->nz);CHKERRQ(ierr); 1852 PetscFunctionReturn(0); 1853 } 1854 1855 if (flag == SOR_APPLY_LOWER) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"SOR_APPLY_LOWER is not implemented"); 1856 else if (flag & SOR_EISENSTAT) { 1857 /* Let A = L + U + D; where L is lower triangular, 1858 U is upper triangular, E = D/omega; This routine applies 1859 1860 (L + E)^{-1} A (U + E)^{-1} 1861 1862 to a vector efficiently using Eisenstat's trick. 1863 */ 1864 scale = (2.0/omega) - 1.0; 1865 1866 /* x = (E + U)^{-1} b */ 1867 for (i=m-1; i>=0; i--) { 1868 n = a->i[i+1] - diag[i] - 1; 1869 idx = a->j + diag[i] + 1; 1870 v = a->a + diag[i] + 1; 1871 sum = b[i]; 1872 PetscSparseDenseMinusDot(sum,x,v,idx,n); 1873 x[i] = sum*idiag[i]; 1874 } 1875 1876 /* t = b - (2*E - D)x */ 1877 v = a->a; 1878 for (i=0; i<m; i++) t[i] = b[i] - scale*(v[*diag++])*x[i]; 1879 1880 /* t = (E + L)^{-1}t */ 1881 ts = t; 1882 diag = a->diag; 1883 for (i=0; i<m; i++) { 1884 n = diag[i] - a->i[i]; 1885 idx = a->j + a->i[i]; 1886 v = a->a + a->i[i]; 1887 sum = t[i]; 1888 PetscSparseDenseMinusDot(sum,ts,v,idx,n); 1889 t[i] = sum*idiag[i]; 1890 /* x = x + t */ 1891 x[i] += t[i]; 1892 } 1893 1894 ierr = PetscLogFlops(6.0*m-1 + 2.0*a->nz);CHKERRQ(ierr); 1895 ierr = VecRestoreArray(xx,&x);CHKERRQ(ierr); 1896 ierr = VecRestoreArrayRead(bb,&b);CHKERRQ(ierr); 1897 PetscFunctionReturn(0); 1898 } 1899 if (flag & SOR_ZERO_INITIAL_GUESS) { 1900 if (flag & SOR_FORWARD_SWEEP || flag & SOR_LOCAL_FORWARD_SWEEP) { 1901 for (i=0; i<m; i++) { 1902 n = diag[i] - a->i[i]; 1903 idx = a->j + a->i[i]; 1904 v = a->a + a->i[i]; 1905 sum = b[i]; 1906 PetscSparseDenseMinusDot(sum,x,v,idx,n); 1907 t[i] = sum; 1908 x[i] = sum*idiag[i]; 1909 } 1910 xb = t; 1911 ierr = PetscLogFlops(a->nz);CHKERRQ(ierr); 1912 } else xb = b; 1913 if (flag & SOR_BACKWARD_SWEEP || flag & SOR_LOCAL_BACKWARD_SWEEP) { 1914 for (i=m-1; i>=0; i--) { 1915 n = a->i[i+1] - diag[i] - 1; 1916 idx = a->j + diag[i] + 1; 1917 v = a->a + diag[i] + 1; 1918 sum = xb[i]; 1919 PetscSparseDenseMinusDot(sum,x,v,idx,n); 1920 if (xb == b) { 1921 x[i] = sum*idiag[i]; 1922 } else { 1923 x[i] = (1-omega)*x[i] + sum*idiag[i]; /* omega in idiag */ 1924 } 1925 } 1926 ierr = PetscLogFlops(a->nz);CHKERRQ(ierr); /* assumes 1/2 in upper */ 1927 } 1928 its--; 1929 } 1930 while (its--) { 1931 if (flag & SOR_FORWARD_SWEEP || flag & SOR_LOCAL_FORWARD_SWEEP) { 1932 for (i=0; i<m; i++) { 1933 /* lower */ 1934 n = diag[i] - a->i[i]; 1935 idx = a->j + a->i[i]; 1936 v = a->a + a->i[i]; 1937 sum = b[i]; 1938 PetscSparseDenseMinusDot(sum,x,v,idx,n); 1939 t[i] = sum; /* save application of the lower-triangular part */ 1940 /* upper */ 1941 n = a->i[i+1] - diag[i] - 1; 1942 idx = a->j + diag[i] + 1; 1943 v = a->a + diag[i] + 1; 1944 PetscSparseDenseMinusDot(sum,x,v,idx,n); 1945 x[i] = (1. - omega)*x[i] + sum*idiag[i]; /* omega in idiag */ 1946 } 1947 xb = t; 1948 ierr = PetscLogFlops(2.0*a->nz);CHKERRQ(ierr); 1949 } else xb = b; 1950 if (flag & SOR_BACKWARD_SWEEP || flag & SOR_LOCAL_BACKWARD_SWEEP) { 1951 for (i=m-1; i>=0; i--) { 1952 sum = xb[i]; 1953 if (xb == b) { 1954 /* whole matrix (no checkpointing available) */ 1955 n = a->i[i+1] - a->i[i]; 1956 idx = a->j + a->i[i]; 1957 v = a->a + a->i[i]; 1958 PetscSparseDenseMinusDot(sum,x,v,idx,n); 1959 x[i] = (1. - omega)*x[i] + (sum + mdiag[i]*x[i])*idiag[i]; 1960 } else { /* lower-triangular part has been saved, so only apply upper-triangular */ 1961 n = a->i[i+1] - diag[i] - 1; 1962 idx = a->j + diag[i] + 1; 1963 v = a->a + diag[i] + 1; 1964 PetscSparseDenseMinusDot(sum,x,v,idx,n); 1965 x[i] = (1. - omega)*x[i] + sum*idiag[i]; /* omega in idiag */ 1966 } 1967 } 1968 if (xb == b) { 1969 ierr = PetscLogFlops(2.0*a->nz);CHKERRQ(ierr); 1970 } else { 1971 ierr = PetscLogFlops(a->nz);CHKERRQ(ierr); /* assumes 1/2 in upper */ 1972 } 1973 } 1974 } 1975 ierr = VecRestoreArray(xx,&x);CHKERRQ(ierr); 1976 ierr = VecRestoreArrayRead(bb,&b);CHKERRQ(ierr); 1977 PetscFunctionReturn(0); 1978 } 1979 1980 1981 PetscErrorCode MatGetInfo_SeqAIJ(Mat A,MatInfoType flag,MatInfo *info) 1982 { 1983 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 1984 1985 PetscFunctionBegin; 1986 info->block_size = 1.0; 1987 info->nz_allocated = a->maxnz; 1988 info->nz_used = a->nz; 1989 info->nz_unneeded = (a->maxnz - a->nz); 1990 info->assemblies = A->num_ass; 1991 info->mallocs = A->info.mallocs; 1992 info->memory = ((PetscObject)A)->mem; 1993 if (A->factortype) { 1994 info->fill_ratio_given = A->info.fill_ratio_given; 1995 info->fill_ratio_needed = A->info.fill_ratio_needed; 1996 info->factor_mallocs = A->info.factor_mallocs; 1997 } else { 1998 info->fill_ratio_given = 0; 1999 info->fill_ratio_needed = 0; 2000 info->factor_mallocs = 0; 2001 } 2002 PetscFunctionReturn(0); 2003 } 2004 2005 PetscErrorCode MatZeroRows_SeqAIJ(Mat A,PetscInt N,const PetscInt rows[],PetscScalar diag,Vec x,Vec b) 2006 { 2007 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 2008 PetscInt i,m = A->rmap->n - 1; 2009 PetscErrorCode ierr; 2010 const PetscScalar *xx; 2011 PetscScalar *bb; 2012 PetscInt d = 0; 2013 2014 PetscFunctionBegin; 2015 if (x && b) { 2016 ierr = VecGetArrayRead(x,&xx);CHKERRQ(ierr); 2017 ierr = VecGetArray(b,&bb);CHKERRQ(ierr); 2018 for (i=0; i<N; i++) { 2019 if (rows[i] < 0 || rows[i] > m) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"row %D out of range", rows[i]); 2020 if (rows[i] >= A->cmap->n) continue; 2021 bb[rows[i]] = diag*xx[rows[i]]; 2022 } 2023 ierr = VecRestoreArrayRead(x,&xx);CHKERRQ(ierr); 2024 ierr = VecRestoreArray(b,&bb);CHKERRQ(ierr); 2025 } 2026 2027 if (a->keepnonzeropattern) { 2028 for (i=0; i<N; i++) { 2029 if (rows[i] < 0 || rows[i] > m) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"row %D out of range", rows[i]); 2030 ierr = PetscArrayzero(&a->a[a->i[rows[i]]],a->ilen[rows[i]]);CHKERRQ(ierr); 2031 } 2032 if (diag != 0.0) { 2033 for (i=0; i<N; i++) { 2034 d = rows[i]; 2035 if (rows[i] >= A->cmap->n) continue; 2036 if (a->diag[d] >= a->i[d+1]) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Matrix is missing diagonal entry in the zeroed row %D",d); 2037 } 2038 for (i=0; i<N; i++) { 2039 if (rows[i] >= A->cmap->n) continue; 2040 a->a[a->diag[rows[i]]] = diag; 2041 } 2042 } 2043 } else { 2044 if (diag != 0.0) { 2045 for (i=0; i<N; i++) { 2046 if (rows[i] < 0 || rows[i] > m) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"row %D out of range", rows[i]); 2047 if (a->ilen[rows[i]] > 0) { 2048 if (rows[i] >= A->cmap->n) { 2049 a->ilen[rows[i]] = 0; 2050 } else { 2051 a->ilen[rows[i]] = 1; 2052 a->a[a->i[rows[i]]] = diag; 2053 a->j[a->i[rows[i]]] = rows[i]; 2054 } 2055 } else if (rows[i] < A->cmap->n) { /* in case row was completely empty */ 2056 ierr = MatSetValues_SeqAIJ(A,1,&rows[i],1,&rows[i],&diag,INSERT_VALUES);CHKERRQ(ierr); 2057 } 2058 } 2059 } else { 2060 for (i=0; i<N; i++) { 2061 if (rows[i] < 0 || rows[i] > m) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"row %D out of range", rows[i]); 2062 a->ilen[rows[i]] = 0; 2063 } 2064 } 2065 A->nonzerostate++; 2066 } 2067 #if defined(PETSC_HAVE_VIENNACL) || defined(PETSC_HAVE_CUDA) 2068 if (A->offloadmask != PETSC_OFFLOAD_UNALLOCATED) A->offloadmask = PETSC_OFFLOAD_CPU; 2069 #endif 2070 ierr = (*A->ops->assemblyend)(A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2071 PetscFunctionReturn(0); 2072 } 2073 2074 PetscErrorCode MatZeroRowsColumns_SeqAIJ(Mat A,PetscInt N,const PetscInt rows[],PetscScalar diag,Vec x,Vec b) 2075 { 2076 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 2077 PetscInt i,j,m = A->rmap->n - 1,d = 0; 2078 PetscErrorCode ierr; 2079 PetscBool missing,*zeroed,vecs = PETSC_FALSE; 2080 const PetscScalar *xx; 2081 PetscScalar *bb; 2082 2083 PetscFunctionBegin; 2084 if (x && b) { 2085 ierr = VecGetArrayRead(x,&xx);CHKERRQ(ierr); 2086 ierr = VecGetArray(b,&bb);CHKERRQ(ierr); 2087 vecs = PETSC_TRUE; 2088 } 2089 ierr = PetscCalloc1(A->rmap->n,&zeroed);CHKERRQ(ierr); 2090 for (i=0; i<N; i++) { 2091 if (rows[i] < 0 || rows[i] > m) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"row %D out of range", rows[i]); 2092 ierr = PetscArrayzero(&a->a[a->i[rows[i]]],a->ilen[rows[i]]);CHKERRQ(ierr); 2093 2094 zeroed[rows[i]] = PETSC_TRUE; 2095 } 2096 for (i=0; i<A->rmap->n; i++) { 2097 if (!zeroed[i]) { 2098 for (j=a->i[i]; j<a->i[i+1]; j++) { 2099 if (a->j[j] < A->rmap->n && zeroed[a->j[j]]) { 2100 if (vecs) bb[i] -= a->a[j]*xx[a->j[j]]; 2101 a->a[j] = 0.0; 2102 } 2103 } 2104 } else if (vecs && i < A->cmap->N) bb[i] = diag*xx[i]; 2105 } 2106 if (x && b) { 2107 ierr = VecRestoreArrayRead(x,&xx);CHKERRQ(ierr); 2108 ierr = VecRestoreArray(b,&bb);CHKERRQ(ierr); 2109 } 2110 ierr = PetscFree(zeroed);CHKERRQ(ierr); 2111 if (diag != 0.0) { 2112 ierr = MatMissingDiagonal_SeqAIJ(A,&missing,&d);CHKERRQ(ierr); 2113 if (missing) { 2114 for (i=0; i<N; i++) { 2115 if (rows[i] >= A->cmap->N) continue; 2116 if (a->nonew && rows[i] >= d) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Matrix is missing diagonal entry in row %D (%D)",d,rows[i]); 2117 ierr = MatSetValues_SeqAIJ(A,1,&rows[i],1,&rows[i],&diag,INSERT_VALUES);CHKERRQ(ierr); 2118 } 2119 } else { 2120 for (i=0; i<N; i++) { 2121 a->a[a->diag[rows[i]]] = diag; 2122 } 2123 } 2124 } 2125 #if defined(PETSC_HAVE_VIENNACL) || defined(PETSC_HAVE_CUDA) 2126 if (A->offloadmask != PETSC_OFFLOAD_UNALLOCATED) A->offloadmask = PETSC_OFFLOAD_CPU; 2127 #endif 2128 ierr = (*A->ops->assemblyend)(A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2129 PetscFunctionReturn(0); 2130 } 2131 2132 PetscErrorCode MatGetRow_SeqAIJ(Mat A,PetscInt row,PetscInt *nz,PetscInt **idx,PetscScalar **v) 2133 { 2134 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 2135 PetscInt *itmp; 2136 2137 PetscFunctionBegin; 2138 if (row < 0 || row >= A->rmap->n) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Row %D out of range",row); 2139 2140 *nz = a->i[row+1] - a->i[row]; 2141 if (v) *v = a->a + a->i[row]; 2142 if (idx) { 2143 itmp = a->j + a->i[row]; 2144 if (*nz) *idx = itmp; 2145 else *idx = 0; 2146 } 2147 PetscFunctionReturn(0); 2148 } 2149 2150 /* remove this function? */ 2151 PetscErrorCode MatRestoreRow_SeqAIJ(Mat A,PetscInt row,PetscInt *nz,PetscInt **idx,PetscScalar **v) 2152 { 2153 PetscFunctionBegin; 2154 PetscFunctionReturn(0); 2155 } 2156 2157 PetscErrorCode MatNorm_SeqAIJ(Mat A,NormType type,PetscReal *nrm) 2158 { 2159 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 2160 MatScalar *v = a->a; 2161 PetscReal sum = 0.0; 2162 PetscErrorCode ierr; 2163 PetscInt i,j; 2164 2165 PetscFunctionBegin; 2166 if (type == NORM_FROBENIUS) { 2167 #if defined(PETSC_USE_REAL___FP16) 2168 PetscBLASInt one = 1,nz = a->nz; 2169 *nrm = BLASnrm2_(&nz,v,&one); 2170 #else 2171 for (i=0; i<a->nz; i++) { 2172 sum += PetscRealPart(PetscConj(*v)*(*v)); v++; 2173 } 2174 *nrm = PetscSqrtReal(sum); 2175 #endif 2176 ierr = PetscLogFlops(2*a->nz);CHKERRQ(ierr); 2177 } else if (type == NORM_1) { 2178 PetscReal *tmp; 2179 PetscInt *jj = a->j; 2180 ierr = PetscCalloc1(A->cmap->n+1,&tmp);CHKERRQ(ierr); 2181 *nrm = 0.0; 2182 for (j=0; j<a->nz; j++) { 2183 tmp[*jj++] += PetscAbsScalar(*v); v++; 2184 } 2185 for (j=0; j<A->cmap->n; j++) { 2186 if (tmp[j] > *nrm) *nrm = tmp[j]; 2187 } 2188 ierr = PetscFree(tmp);CHKERRQ(ierr); 2189 ierr = PetscLogFlops(PetscMax(a->nz-1,0));CHKERRQ(ierr); 2190 } else if (type == NORM_INFINITY) { 2191 *nrm = 0.0; 2192 for (j=0; j<A->rmap->n; j++) { 2193 v = a->a + a->i[j]; 2194 sum = 0.0; 2195 for (i=0; i<a->i[j+1]-a->i[j]; i++) { 2196 sum += PetscAbsScalar(*v); v++; 2197 } 2198 if (sum > *nrm) *nrm = sum; 2199 } 2200 ierr = PetscLogFlops(PetscMax(a->nz-1,0));CHKERRQ(ierr); 2201 } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"No support for two norm"); 2202 PetscFunctionReturn(0); 2203 } 2204 2205 /* Merged from MatGetSymbolicTranspose_SeqAIJ() - replace MatGetSymbolicTranspose_SeqAIJ()? */ 2206 PetscErrorCode MatTransposeSymbolic_SeqAIJ(Mat A,Mat *B) 2207 { 2208 PetscErrorCode ierr; 2209 PetscInt i,j,anzj; 2210 Mat_SeqAIJ *a=(Mat_SeqAIJ*)A->data,*b; 2211 PetscInt an=A->cmap->N,am=A->rmap->N; 2212 PetscInt *ati,*atj,*atfill,*ai=a->i,*aj=a->j; 2213 2214 PetscFunctionBegin; 2215 /* Allocate space for symbolic transpose info and work array */ 2216 ierr = PetscCalloc1(an+1,&ati);CHKERRQ(ierr); 2217 ierr = PetscMalloc1(ai[am],&atj);CHKERRQ(ierr); 2218 ierr = PetscMalloc1(an,&atfill);CHKERRQ(ierr); 2219 2220 /* Walk through aj and count ## of non-zeros in each row of A^T. */ 2221 /* Note: offset by 1 for fast conversion into csr format. */ 2222 for (i=0;i<ai[am];i++) ati[aj[i]+1] += 1; 2223 /* Form ati for csr format of A^T. */ 2224 for (i=0;i<an;i++) ati[i+1] += ati[i]; 2225 2226 /* Copy ati into atfill so we have locations of the next free space in atj */ 2227 ierr = PetscArraycpy(atfill,ati,an);CHKERRQ(ierr); 2228 2229 /* Walk through A row-wise and mark nonzero entries of A^T. */ 2230 for (i=0;i<am;i++) { 2231 anzj = ai[i+1] - ai[i]; 2232 for (j=0;j<anzj;j++) { 2233 atj[atfill[*aj]] = i; 2234 atfill[*aj++] += 1; 2235 } 2236 } 2237 2238 /* Clean up temporary space and complete requests. */ 2239 ierr = PetscFree(atfill);CHKERRQ(ierr); 2240 ierr = MatCreateSeqAIJWithArrays(PetscObjectComm((PetscObject)A),an,am,ati,atj,NULL,B);CHKERRQ(ierr); 2241 ierr = MatSetBlockSizes(*B,PetscAbs(A->cmap->bs),PetscAbs(A->rmap->bs));CHKERRQ(ierr); 2242 ierr = MatSetType(*B,((PetscObject)A)->type_name);CHKERRQ(ierr); 2243 2244 b = (Mat_SeqAIJ*)((*B)->data); 2245 b->free_a = PETSC_FALSE; 2246 b->free_ij = PETSC_TRUE; 2247 b->nonew = 0; 2248 PetscFunctionReturn(0); 2249 } 2250 2251 PetscErrorCode MatIsTranspose_SeqAIJ(Mat A,Mat B,PetscReal tol,PetscBool *f) 2252 { 2253 Mat_SeqAIJ *aij = (Mat_SeqAIJ*) A->data,*bij = (Mat_SeqAIJ*) B->data; 2254 PetscInt *adx,*bdx,*aii,*bii,*aptr,*bptr; 2255 MatScalar *va,*vb; 2256 PetscErrorCode ierr; 2257 PetscInt ma,na,mb,nb, i; 2258 2259 PetscFunctionBegin; 2260 ierr = MatGetSize(A,&ma,&na);CHKERRQ(ierr); 2261 ierr = MatGetSize(B,&mb,&nb);CHKERRQ(ierr); 2262 if (ma!=nb || na!=mb) { 2263 *f = PETSC_FALSE; 2264 PetscFunctionReturn(0); 2265 } 2266 aii = aij->i; bii = bij->i; 2267 adx = aij->j; bdx = bij->j; 2268 va = aij->a; vb = bij->a; 2269 ierr = PetscMalloc1(ma,&aptr);CHKERRQ(ierr); 2270 ierr = PetscMalloc1(mb,&bptr);CHKERRQ(ierr); 2271 for (i=0; i<ma; i++) aptr[i] = aii[i]; 2272 for (i=0; i<mb; i++) bptr[i] = bii[i]; 2273 2274 *f = PETSC_TRUE; 2275 for (i=0; i<ma; i++) { 2276 while (aptr[i]<aii[i+1]) { 2277 PetscInt idc,idr; 2278 PetscScalar vc,vr; 2279 /* column/row index/value */ 2280 idc = adx[aptr[i]]; 2281 idr = bdx[bptr[idc]]; 2282 vc = va[aptr[i]]; 2283 vr = vb[bptr[idc]]; 2284 if (i!=idr || PetscAbsScalar(vc-vr) > tol) { 2285 *f = PETSC_FALSE; 2286 goto done; 2287 } else { 2288 aptr[i]++; 2289 if (B || i!=idc) bptr[idc]++; 2290 } 2291 } 2292 } 2293 done: 2294 ierr = PetscFree(aptr);CHKERRQ(ierr); 2295 ierr = PetscFree(bptr);CHKERRQ(ierr); 2296 PetscFunctionReturn(0); 2297 } 2298 2299 PetscErrorCode MatIsHermitianTranspose_SeqAIJ(Mat A,Mat B,PetscReal tol,PetscBool *f) 2300 { 2301 Mat_SeqAIJ *aij = (Mat_SeqAIJ*) A->data,*bij = (Mat_SeqAIJ*) B->data; 2302 PetscInt *adx,*bdx,*aii,*bii,*aptr,*bptr; 2303 MatScalar *va,*vb; 2304 PetscErrorCode ierr; 2305 PetscInt ma,na,mb,nb, i; 2306 2307 PetscFunctionBegin; 2308 ierr = MatGetSize(A,&ma,&na);CHKERRQ(ierr); 2309 ierr = MatGetSize(B,&mb,&nb);CHKERRQ(ierr); 2310 if (ma!=nb || na!=mb) { 2311 *f = PETSC_FALSE; 2312 PetscFunctionReturn(0); 2313 } 2314 aii = aij->i; bii = bij->i; 2315 adx = aij->j; bdx = bij->j; 2316 va = aij->a; vb = bij->a; 2317 ierr = PetscMalloc1(ma,&aptr);CHKERRQ(ierr); 2318 ierr = PetscMalloc1(mb,&bptr);CHKERRQ(ierr); 2319 for (i=0; i<ma; i++) aptr[i] = aii[i]; 2320 for (i=0; i<mb; i++) bptr[i] = bii[i]; 2321 2322 *f = PETSC_TRUE; 2323 for (i=0; i<ma; i++) { 2324 while (aptr[i]<aii[i+1]) { 2325 PetscInt idc,idr; 2326 PetscScalar vc,vr; 2327 /* column/row index/value */ 2328 idc = adx[aptr[i]]; 2329 idr = bdx[bptr[idc]]; 2330 vc = va[aptr[i]]; 2331 vr = vb[bptr[idc]]; 2332 if (i!=idr || PetscAbsScalar(vc-PetscConj(vr)) > tol) { 2333 *f = PETSC_FALSE; 2334 goto done; 2335 } else { 2336 aptr[i]++; 2337 if (B || i!=idc) bptr[idc]++; 2338 } 2339 } 2340 } 2341 done: 2342 ierr = PetscFree(aptr);CHKERRQ(ierr); 2343 ierr = PetscFree(bptr);CHKERRQ(ierr); 2344 PetscFunctionReturn(0); 2345 } 2346 2347 PetscErrorCode MatIsSymmetric_SeqAIJ(Mat A,PetscReal tol,PetscBool *f) 2348 { 2349 PetscErrorCode ierr; 2350 2351 PetscFunctionBegin; 2352 ierr = MatIsTranspose_SeqAIJ(A,A,tol,f);CHKERRQ(ierr); 2353 PetscFunctionReturn(0); 2354 } 2355 2356 PetscErrorCode MatIsHermitian_SeqAIJ(Mat A,PetscReal tol,PetscBool *f) 2357 { 2358 PetscErrorCode ierr; 2359 2360 PetscFunctionBegin; 2361 ierr = MatIsHermitianTranspose_SeqAIJ(A,A,tol,f);CHKERRQ(ierr); 2362 PetscFunctionReturn(0); 2363 } 2364 2365 PetscErrorCode MatDiagonalScale_SeqAIJ(Mat A,Vec ll,Vec rr) 2366 { 2367 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 2368 const PetscScalar *l,*r; 2369 PetscScalar x; 2370 MatScalar *v; 2371 PetscErrorCode ierr; 2372 PetscInt i,j,m = A->rmap->n,n = A->cmap->n,M,nz = a->nz; 2373 const PetscInt *jj; 2374 2375 PetscFunctionBegin; 2376 if (ll) { 2377 /* The local size is used so that VecMPI can be passed to this routine 2378 by MatDiagonalScale_MPIAIJ */ 2379 ierr = VecGetLocalSize(ll,&m);CHKERRQ(ierr); 2380 if (m != A->rmap->n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Left scaling vector wrong length"); 2381 ierr = VecGetArrayRead(ll,&l);CHKERRQ(ierr); 2382 v = a->a; 2383 for (i=0; i<m; i++) { 2384 x = l[i]; 2385 M = a->i[i+1] - a->i[i]; 2386 for (j=0; j<M; j++) (*v++) *= x; 2387 } 2388 ierr = VecRestoreArrayRead(ll,&l);CHKERRQ(ierr); 2389 ierr = PetscLogFlops(nz);CHKERRQ(ierr); 2390 } 2391 if (rr) { 2392 ierr = VecGetLocalSize(rr,&n);CHKERRQ(ierr); 2393 if (n != A->cmap->n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Right scaling vector wrong length"); 2394 ierr = VecGetArrayRead(rr,&r);CHKERRQ(ierr); 2395 v = a->a; jj = a->j; 2396 for (i=0; i<nz; i++) (*v++) *= r[*jj++]; 2397 ierr = VecRestoreArrayRead(rr,&r);CHKERRQ(ierr); 2398 ierr = PetscLogFlops(nz);CHKERRQ(ierr); 2399 } 2400 ierr = MatSeqAIJInvalidateDiagonal(A);CHKERRQ(ierr); 2401 #if defined(PETSC_HAVE_VIENNACL) || defined(PETSC_HAVE_CUDA) 2402 if (A->offloadmask != PETSC_OFFLOAD_UNALLOCATED) A->offloadmask = PETSC_OFFLOAD_CPU; 2403 #endif 2404 PetscFunctionReturn(0); 2405 } 2406 2407 PetscErrorCode MatCreateSubMatrix_SeqAIJ(Mat A,IS isrow,IS iscol,PetscInt csize,MatReuse scall,Mat *B) 2408 { 2409 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data,*c; 2410 PetscErrorCode ierr; 2411 PetscInt *smap,i,k,kstart,kend,oldcols = A->cmap->n,*lens; 2412 PetscInt row,mat_i,*mat_j,tcol,first,step,*mat_ilen,sum,lensi; 2413 const PetscInt *irow,*icol; 2414 PetscInt nrows,ncols; 2415 PetscInt *starts,*j_new,*i_new,*aj = a->j,*ai = a->i,ii,*ailen = a->ilen; 2416 MatScalar *a_new,*mat_a; 2417 Mat C; 2418 PetscBool stride; 2419 2420 PetscFunctionBegin; 2421 2422 ierr = ISGetIndices(isrow,&irow);CHKERRQ(ierr); 2423 ierr = ISGetLocalSize(isrow,&nrows);CHKERRQ(ierr); 2424 ierr = ISGetLocalSize(iscol,&ncols);CHKERRQ(ierr); 2425 2426 ierr = PetscObjectTypeCompare((PetscObject)iscol,ISSTRIDE,&stride);CHKERRQ(ierr); 2427 if (stride) { 2428 ierr = ISStrideGetInfo(iscol,&first,&step);CHKERRQ(ierr); 2429 } else { 2430 first = 0; 2431 step = 0; 2432 } 2433 if (stride && step == 1) { 2434 /* special case of contiguous rows */ 2435 ierr = PetscMalloc2(nrows,&lens,nrows,&starts);CHKERRQ(ierr); 2436 /* loop over new rows determining lens and starting points */ 2437 for (i=0; i<nrows; i++) { 2438 kstart = ai[irow[i]]; 2439 kend = kstart + ailen[irow[i]]; 2440 starts[i] = kstart; 2441 for (k=kstart; k<kend; k++) { 2442 if (aj[k] >= first) { 2443 starts[i] = k; 2444 break; 2445 } 2446 } 2447 sum = 0; 2448 while (k < kend) { 2449 if (aj[k++] >= first+ncols) break; 2450 sum++; 2451 } 2452 lens[i] = sum; 2453 } 2454 /* create submatrix */ 2455 if (scall == MAT_REUSE_MATRIX) { 2456 PetscInt n_cols,n_rows; 2457 ierr = MatGetSize(*B,&n_rows,&n_cols);CHKERRQ(ierr); 2458 if (n_rows != nrows || n_cols != ncols) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Reused submatrix wrong size"); 2459 ierr = MatZeroEntries(*B);CHKERRQ(ierr); 2460 C = *B; 2461 } else { 2462 PetscInt rbs,cbs; 2463 ierr = MatCreate(PetscObjectComm((PetscObject)A),&C);CHKERRQ(ierr); 2464 ierr = MatSetSizes(C,nrows,ncols,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 2465 ierr = ISGetBlockSize(isrow,&rbs);CHKERRQ(ierr); 2466 ierr = ISGetBlockSize(iscol,&cbs);CHKERRQ(ierr); 2467 ierr = MatSetBlockSizes(C,rbs,cbs);CHKERRQ(ierr); 2468 ierr = MatSetType(C,((PetscObject)A)->type_name);CHKERRQ(ierr); 2469 ierr = MatSeqAIJSetPreallocation_SeqAIJ(C,0,lens);CHKERRQ(ierr); 2470 } 2471 c = (Mat_SeqAIJ*)C->data; 2472 2473 /* loop over rows inserting into submatrix */ 2474 a_new = c->a; 2475 j_new = c->j; 2476 i_new = c->i; 2477 2478 for (i=0; i<nrows; i++) { 2479 ii = starts[i]; 2480 lensi = lens[i]; 2481 for (k=0; k<lensi; k++) { 2482 *j_new++ = aj[ii+k] - first; 2483 } 2484 ierr = PetscArraycpy(a_new,a->a + starts[i],lensi);CHKERRQ(ierr); 2485 a_new += lensi; 2486 i_new[i+1] = i_new[i] + lensi; 2487 c->ilen[i] = lensi; 2488 } 2489 ierr = PetscFree2(lens,starts);CHKERRQ(ierr); 2490 } else { 2491 ierr = ISGetIndices(iscol,&icol);CHKERRQ(ierr); 2492 ierr = PetscCalloc1(oldcols,&smap);CHKERRQ(ierr); 2493 ierr = PetscMalloc1(1+nrows,&lens);CHKERRQ(ierr); 2494 for (i=0; i<ncols; i++) { 2495 #if defined(PETSC_USE_DEBUG) 2496 if (icol[i] >= oldcols) SETERRQ3(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Requesting column beyond largest column icol[%D] %D <= A->cmap->n %D",i,icol[i],oldcols); 2497 #endif 2498 smap[icol[i]] = i+1; 2499 } 2500 2501 /* determine lens of each row */ 2502 for (i=0; i<nrows; i++) { 2503 kstart = ai[irow[i]]; 2504 kend = kstart + a->ilen[irow[i]]; 2505 lens[i] = 0; 2506 for (k=kstart; k<kend; k++) { 2507 if (smap[aj[k]]) { 2508 lens[i]++; 2509 } 2510 } 2511 } 2512 /* Create and fill new matrix */ 2513 if (scall == MAT_REUSE_MATRIX) { 2514 PetscBool equal; 2515 2516 c = (Mat_SeqAIJ*)((*B)->data); 2517 if ((*B)->rmap->n != nrows || (*B)->cmap->n != ncols) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Cannot reuse matrix. wrong size"); 2518 ierr = PetscArraycmp(c->ilen,lens,(*B)->rmap->n,&equal);CHKERRQ(ierr); 2519 if (!equal) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Cannot reuse matrix. wrong no of nonzeros"); 2520 ierr = PetscArrayzero(c->ilen,(*B)->rmap->n);CHKERRQ(ierr); 2521 C = *B; 2522 } else { 2523 PetscInt rbs,cbs; 2524 ierr = MatCreate(PetscObjectComm((PetscObject)A),&C);CHKERRQ(ierr); 2525 ierr = MatSetSizes(C,nrows,ncols,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 2526 ierr = ISGetBlockSize(isrow,&rbs);CHKERRQ(ierr); 2527 ierr = ISGetBlockSize(iscol,&cbs);CHKERRQ(ierr); 2528 ierr = MatSetBlockSizes(C,rbs,cbs);CHKERRQ(ierr); 2529 ierr = MatSetType(C,((PetscObject)A)->type_name);CHKERRQ(ierr); 2530 ierr = MatSeqAIJSetPreallocation_SeqAIJ(C,0,lens);CHKERRQ(ierr); 2531 } 2532 c = (Mat_SeqAIJ*)(C->data); 2533 for (i=0; i<nrows; i++) { 2534 row = irow[i]; 2535 kstart = ai[row]; 2536 kend = kstart + a->ilen[row]; 2537 mat_i = c->i[i]; 2538 mat_j = c->j + mat_i; 2539 mat_a = c->a + mat_i; 2540 mat_ilen = c->ilen + i; 2541 for (k=kstart; k<kend; k++) { 2542 if ((tcol=smap[a->j[k]])) { 2543 *mat_j++ = tcol - 1; 2544 *mat_a++ = a->a[k]; 2545 (*mat_ilen)++; 2546 2547 } 2548 } 2549 } 2550 /* Free work space */ 2551 ierr = ISRestoreIndices(iscol,&icol);CHKERRQ(ierr); 2552 ierr = PetscFree(smap);CHKERRQ(ierr); 2553 ierr = PetscFree(lens);CHKERRQ(ierr); 2554 /* sort */ 2555 for (i = 0; i < nrows; i++) { 2556 PetscInt ilen; 2557 2558 mat_i = c->i[i]; 2559 mat_j = c->j + mat_i; 2560 mat_a = c->a + mat_i; 2561 ilen = c->ilen[i]; 2562 ierr = PetscSortIntWithScalarArray(ilen,mat_j,mat_a);CHKERRQ(ierr); 2563 } 2564 } 2565 #if defined(PETSC_HAVE_VIENNACL) || defined(PETSC_HAVE_CUDA) 2566 ierr = MatPinToCPU(C,A->pinnedtocpu);CHKERRQ(ierr); 2567 #endif 2568 ierr = MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2569 ierr = MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2570 2571 ierr = ISRestoreIndices(isrow,&irow);CHKERRQ(ierr); 2572 *B = C; 2573 PetscFunctionReturn(0); 2574 } 2575 2576 PetscErrorCode MatGetMultiProcBlock_SeqAIJ(Mat mat,MPI_Comm subComm,MatReuse scall,Mat *subMat) 2577 { 2578 PetscErrorCode ierr; 2579 Mat B; 2580 2581 PetscFunctionBegin; 2582 if (scall == MAT_INITIAL_MATRIX) { 2583 ierr = MatCreate(subComm,&B);CHKERRQ(ierr); 2584 ierr = MatSetSizes(B,mat->rmap->n,mat->cmap->n,mat->rmap->n,mat->cmap->n);CHKERRQ(ierr); 2585 ierr = MatSetBlockSizesFromMats(B,mat,mat);CHKERRQ(ierr); 2586 ierr = MatSetType(B,MATSEQAIJ);CHKERRQ(ierr); 2587 ierr = MatDuplicateNoCreate_SeqAIJ(B,mat,MAT_COPY_VALUES,PETSC_TRUE);CHKERRQ(ierr); 2588 *subMat = B; 2589 } else { 2590 ierr = MatCopy_SeqAIJ(mat,*subMat,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 2591 } 2592 PetscFunctionReturn(0); 2593 } 2594 2595 PetscErrorCode MatILUFactor_SeqAIJ(Mat inA,IS row,IS col,const MatFactorInfo *info) 2596 { 2597 Mat_SeqAIJ *a = (Mat_SeqAIJ*)inA->data; 2598 PetscErrorCode ierr; 2599 Mat outA; 2600 PetscBool row_identity,col_identity; 2601 2602 PetscFunctionBegin; 2603 if (info->levels != 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Only levels=0 supported for in-place ilu"); 2604 2605 ierr = ISIdentity(row,&row_identity);CHKERRQ(ierr); 2606 ierr = ISIdentity(col,&col_identity);CHKERRQ(ierr); 2607 2608 outA = inA; 2609 outA->factortype = MAT_FACTOR_LU; 2610 ierr = PetscFree(inA->solvertype);CHKERRQ(ierr); 2611 ierr = PetscStrallocpy(MATSOLVERPETSC,&inA->solvertype);CHKERRQ(ierr); 2612 2613 ierr = PetscObjectReference((PetscObject)row);CHKERRQ(ierr); 2614 ierr = ISDestroy(&a->row);CHKERRQ(ierr); 2615 2616 a->row = row; 2617 2618 ierr = PetscObjectReference((PetscObject)col);CHKERRQ(ierr); 2619 ierr = ISDestroy(&a->col);CHKERRQ(ierr); 2620 2621 a->col = col; 2622 2623 /* Create the inverse permutation so that it can be used in MatLUFactorNumeric() */ 2624 ierr = ISDestroy(&a->icol);CHKERRQ(ierr); 2625 ierr = ISInvertPermutation(col,PETSC_DECIDE,&a->icol);CHKERRQ(ierr); 2626 ierr = PetscLogObjectParent((PetscObject)inA,(PetscObject)a->icol);CHKERRQ(ierr); 2627 2628 if (!a->solve_work) { /* this matrix may have been factored before */ 2629 ierr = PetscMalloc1(inA->rmap->n+1,&a->solve_work);CHKERRQ(ierr); 2630 ierr = PetscLogObjectMemory((PetscObject)inA, (inA->rmap->n+1)*sizeof(PetscScalar));CHKERRQ(ierr); 2631 } 2632 2633 ierr = MatMarkDiagonal_SeqAIJ(inA);CHKERRQ(ierr); 2634 if (row_identity && col_identity) { 2635 ierr = MatLUFactorNumeric_SeqAIJ_inplace(outA,inA,info);CHKERRQ(ierr); 2636 } else { 2637 ierr = MatLUFactorNumeric_SeqAIJ_InplaceWithPerm(outA,inA,info);CHKERRQ(ierr); 2638 } 2639 PetscFunctionReturn(0); 2640 } 2641 2642 PetscErrorCode MatScale_SeqAIJ(Mat inA,PetscScalar alpha) 2643 { 2644 Mat_SeqAIJ *a = (Mat_SeqAIJ*)inA->data; 2645 PetscScalar oalpha = alpha; 2646 PetscErrorCode ierr; 2647 PetscBLASInt one = 1,bnz; 2648 2649 PetscFunctionBegin; 2650 ierr = PetscBLASIntCast(a->nz,&bnz);CHKERRQ(ierr); 2651 PetscStackCallBLAS("BLASscal",BLASscal_(&bnz,&oalpha,a->a,&one)); 2652 ierr = PetscLogFlops(a->nz);CHKERRQ(ierr); 2653 ierr = MatSeqAIJInvalidateDiagonal(inA);CHKERRQ(ierr); 2654 #if defined(PETSC_HAVE_VIENNACL) || defined(PETSC_HAVE_CUDA) 2655 if (inA->offloadmask != PETSC_OFFLOAD_UNALLOCATED) inA->offloadmask = PETSC_OFFLOAD_CPU; 2656 #endif 2657 PetscFunctionReturn(0); 2658 } 2659 2660 PetscErrorCode MatDestroySubMatrix_Private(Mat_SubSppt *submatj) 2661 { 2662 PetscErrorCode ierr; 2663 PetscInt i; 2664 2665 PetscFunctionBegin; 2666 if (!submatj->id) { /* delete data that are linked only to submats[id=0] */ 2667 ierr = PetscFree4(submatj->sbuf1,submatj->ptr,submatj->tmp,submatj->ctr);CHKERRQ(ierr); 2668 2669 for (i=0; i<submatj->nrqr; ++i) { 2670 ierr = PetscFree(submatj->sbuf2[i]);CHKERRQ(ierr); 2671 } 2672 ierr = PetscFree3(submatj->sbuf2,submatj->req_size,submatj->req_source1);CHKERRQ(ierr); 2673 2674 if (submatj->rbuf1) { 2675 ierr = PetscFree(submatj->rbuf1[0]);CHKERRQ(ierr); 2676 ierr = PetscFree(submatj->rbuf1);CHKERRQ(ierr); 2677 } 2678 2679 for (i=0; i<submatj->nrqs; ++i) { 2680 ierr = PetscFree(submatj->rbuf3[i]);CHKERRQ(ierr); 2681 } 2682 ierr = PetscFree3(submatj->req_source2,submatj->rbuf2,submatj->rbuf3);CHKERRQ(ierr); 2683 ierr = PetscFree(submatj->pa);CHKERRQ(ierr); 2684 } 2685 2686 #if defined(PETSC_USE_CTABLE) 2687 ierr = PetscTableDestroy((PetscTable*)&submatj->rmap);CHKERRQ(ierr); 2688 if (submatj->cmap_loc) {ierr = PetscFree(submatj->cmap_loc);CHKERRQ(ierr);} 2689 ierr = PetscFree(submatj->rmap_loc);CHKERRQ(ierr); 2690 #else 2691 ierr = PetscFree(submatj->rmap);CHKERRQ(ierr); 2692 #endif 2693 2694 if (!submatj->allcolumns) { 2695 #if defined(PETSC_USE_CTABLE) 2696 ierr = PetscTableDestroy((PetscTable*)&submatj->cmap);CHKERRQ(ierr); 2697 #else 2698 ierr = PetscFree(submatj->cmap);CHKERRQ(ierr); 2699 #endif 2700 } 2701 ierr = PetscFree(submatj->row2proc);CHKERRQ(ierr); 2702 2703 ierr = PetscFree(submatj);CHKERRQ(ierr); 2704 PetscFunctionReturn(0); 2705 } 2706 2707 PetscErrorCode MatDestroySubMatrix_SeqAIJ(Mat C) 2708 { 2709 PetscErrorCode ierr; 2710 Mat_SeqAIJ *c = (Mat_SeqAIJ*)C->data; 2711 Mat_SubSppt *submatj = c->submatis1; 2712 2713 PetscFunctionBegin; 2714 ierr = (*submatj->destroy)(C);CHKERRQ(ierr); 2715 ierr = MatDestroySubMatrix_Private(submatj);CHKERRQ(ierr); 2716 PetscFunctionReturn(0); 2717 } 2718 2719 PetscErrorCode MatDestroySubMatrices_SeqAIJ(PetscInt n,Mat *mat[]) 2720 { 2721 PetscErrorCode ierr; 2722 PetscInt i; 2723 Mat C; 2724 Mat_SeqAIJ *c; 2725 Mat_SubSppt *submatj; 2726 2727 PetscFunctionBegin; 2728 for (i=0; i<n; i++) { 2729 C = (*mat)[i]; 2730 c = (Mat_SeqAIJ*)C->data; 2731 submatj = c->submatis1; 2732 if (submatj) { 2733 if (--((PetscObject)C)->refct <= 0) { 2734 ierr = (*submatj->destroy)(C);CHKERRQ(ierr); 2735 ierr = MatDestroySubMatrix_Private(submatj);CHKERRQ(ierr); 2736 ierr = PetscFree(C->defaultvectype);CHKERRQ(ierr); 2737 ierr = PetscLayoutDestroy(&C->rmap);CHKERRQ(ierr); 2738 ierr = PetscLayoutDestroy(&C->cmap);CHKERRQ(ierr); 2739 ierr = PetscHeaderDestroy(&C);CHKERRQ(ierr); 2740 } 2741 } else { 2742 ierr = MatDestroy(&C);CHKERRQ(ierr); 2743 } 2744 } 2745 2746 /* Destroy Dummy submatrices created for reuse */ 2747 ierr = MatDestroySubMatrices_Dummy(n,mat);CHKERRQ(ierr); 2748 2749 ierr = PetscFree(*mat);CHKERRQ(ierr); 2750 PetscFunctionReturn(0); 2751 } 2752 2753 PetscErrorCode MatCreateSubMatrices_SeqAIJ(Mat A,PetscInt n,const IS irow[],const IS icol[],MatReuse scall,Mat *B[]) 2754 { 2755 PetscErrorCode ierr; 2756 PetscInt i; 2757 2758 PetscFunctionBegin; 2759 if (scall == MAT_INITIAL_MATRIX) { 2760 ierr = PetscCalloc1(n+1,B);CHKERRQ(ierr); 2761 } 2762 2763 for (i=0; i<n; i++) { 2764 ierr = MatCreateSubMatrix_SeqAIJ(A,irow[i],icol[i],PETSC_DECIDE,scall,&(*B)[i]);CHKERRQ(ierr); 2765 } 2766 PetscFunctionReturn(0); 2767 } 2768 2769 PetscErrorCode MatIncreaseOverlap_SeqAIJ(Mat A,PetscInt is_max,IS is[],PetscInt ov) 2770 { 2771 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 2772 PetscErrorCode ierr; 2773 PetscInt row,i,j,k,l,m,n,*nidx,isz,val; 2774 const PetscInt *idx; 2775 PetscInt start,end,*ai,*aj; 2776 PetscBT table; 2777 2778 PetscFunctionBegin; 2779 m = A->rmap->n; 2780 ai = a->i; 2781 aj = a->j; 2782 2783 if (ov < 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"illegal negative overlap value used"); 2784 2785 ierr = PetscMalloc1(m+1,&nidx);CHKERRQ(ierr); 2786 ierr = PetscBTCreate(m,&table);CHKERRQ(ierr); 2787 2788 for (i=0; i<is_max; i++) { 2789 /* Initialize the two local arrays */ 2790 isz = 0; 2791 ierr = PetscBTMemzero(m,table);CHKERRQ(ierr); 2792 2793 /* Extract the indices, assume there can be duplicate entries */ 2794 ierr = ISGetIndices(is[i],&idx);CHKERRQ(ierr); 2795 ierr = ISGetLocalSize(is[i],&n);CHKERRQ(ierr); 2796 2797 /* Enter these into the temp arrays. I.e., mark table[row], enter row into new index */ 2798 for (j=0; j<n; ++j) { 2799 if (!PetscBTLookupSet(table,idx[j])) nidx[isz++] = idx[j]; 2800 } 2801 ierr = ISRestoreIndices(is[i],&idx);CHKERRQ(ierr); 2802 ierr = ISDestroy(&is[i]);CHKERRQ(ierr); 2803 2804 k = 0; 2805 for (j=0; j<ov; j++) { /* for each overlap */ 2806 n = isz; 2807 for (; k<n; k++) { /* do only those rows in nidx[k], which are not done yet */ 2808 row = nidx[k]; 2809 start = ai[row]; 2810 end = ai[row+1]; 2811 for (l = start; l<end; l++) { 2812 val = aj[l]; 2813 if (!PetscBTLookupSet(table,val)) nidx[isz++] = val; 2814 } 2815 } 2816 } 2817 ierr = ISCreateGeneral(PETSC_COMM_SELF,isz,nidx,PETSC_COPY_VALUES,(is+i));CHKERRQ(ierr); 2818 } 2819 ierr = PetscBTDestroy(&table);CHKERRQ(ierr); 2820 ierr = PetscFree(nidx);CHKERRQ(ierr); 2821 PetscFunctionReturn(0); 2822 } 2823 2824 /* -------------------------------------------------------------- */ 2825 PetscErrorCode MatPermute_SeqAIJ(Mat A,IS rowp,IS colp,Mat *B) 2826 { 2827 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 2828 PetscErrorCode ierr; 2829 PetscInt i,nz = 0,m = A->rmap->n,n = A->cmap->n; 2830 const PetscInt *row,*col; 2831 PetscInt *cnew,j,*lens; 2832 IS icolp,irowp; 2833 PetscInt *cwork = NULL; 2834 PetscScalar *vwork = NULL; 2835 2836 PetscFunctionBegin; 2837 ierr = ISInvertPermutation(rowp,PETSC_DECIDE,&irowp);CHKERRQ(ierr); 2838 ierr = ISGetIndices(irowp,&row);CHKERRQ(ierr); 2839 ierr = ISInvertPermutation(colp,PETSC_DECIDE,&icolp);CHKERRQ(ierr); 2840 ierr = ISGetIndices(icolp,&col);CHKERRQ(ierr); 2841 2842 /* determine lengths of permuted rows */ 2843 ierr = PetscMalloc1(m+1,&lens);CHKERRQ(ierr); 2844 for (i=0; i<m; i++) lens[row[i]] = a->i[i+1] - a->i[i]; 2845 ierr = MatCreate(PetscObjectComm((PetscObject)A),B);CHKERRQ(ierr); 2846 ierr = MatSetSizes(*B,m,n,m,n);CHKERRQ(ierr); 2847 ierr = MatSetBlockSizesFromMats(*B,A,A);CHKERRQ(ierr); 2848 ierr = MatSetType(*B,((PetscObject)A)->type_name);CHKERRQ(ierr); 2849 ierr = MatSeqAIJSetPreallocation_SeqAIJ(*B,0,lens);CHKERRQ(ierr); 2850 ierr = PetscFree(lens);CHKERRQ(ierr); 2851 2852 ierr = PetscMalloc1(n,&cnew);CHKERRQ(ierr); 2853 for (i=0; i<m; i++) { 2854 ierr = MatGetRow_SeqAIJ(A,i,&nz,&cwork,&vwork);CHKERRQ(ierr); 2855 for (j=0; j<nz; j++) cnew[j] = col[cwork[j]]; 2856 ierr = MatSetValues_SeqAIJ(*B,1,&row[i],nz,cnew,vwork,INSERT_VALUES);CHKERRQ(ierr); 2857 ierr = MatRestoreRow_SeqAIJ(A,i,&nz,&cwork,&vwork);CHKERRQ(ierr); 2858 } 2859 ierr = PetscFree(cnew);CHKERRQ(ierr); 2860 2861 (*B)->assembled = PETSC_FALSE; 2862 2863 #if defined(PETSC_HAVE_VIENNACL) || defined(PETSC_HAVE_CUDA) 2864 ierr = MatPinToCPU(*B,A->pinnedtocpu);CHKERRQ(ierr); 2865 #endif 2866 ierr = MatAssemblyBegin(*B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2867 ierr = MatAssemblyEnd(*B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2868 ierr = ISRestoreIndices(irowp,&row);CHKERRQ(ierr); 2869 ierr = ISRestoreIndices(icolp,&col);CHKERRQ(ierr); 2870 ierr = ISDestroy(&irowp);CHKERRQ(ierr); 2871 ierr = ISDestroy(&icolp);CHKERRQ(ierr); 2872 PetscFunctionReturn(0); 2873 } 2874 2875 PetscErrorCode MatCopy_SeqAIJ(Mat A,Mat B,MatStructure str) 2876 { 2877 PetscErrorCode ierr; 2878 2879 PetscFunctionBegin; 2880 /* If the two matrices have the same copy implementation, use fast copy. */ 2881 if (str == SAME_NONZERO_PATTERN && (A->ops->copy == B->ops->copy)) { 2882 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 2883 Mat_SeqAIJ *b = (Mat_SeqAIJ*)B->data; 2884 2885 if (a->i[A->rmap->n] != b->i[B->rmap->n]) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_INCOMP,"Number of nonzeros in two matrices are different"); 2886 ierr = PetscArraycpy(b->a,a->a,a->i[A->rmap->n]);CHKERRQ(ierr); 2887 ierr = PetscObjectStateIncrease((PetscObject)B);CHKERRQ(ierr); 2888 } else { 2889 ierr = MatCopy_Basic(A,B,str);CHKERRQ(ierr); 2890 } 2891 PetscFunctionReturn(0); 2892 } 2893 2894 PetscErrorCode MatSetUp_SeqAIJ(Mat A) 2895 { 2896 PetscErrorCode ierr; 2897 2898 PetscFunctionBegin; 2899 ierr = MatSeqAIJSetPreallocation_SeqAIJ(A,PETSC_DEFAULT,0);CHKERRQ(ierr); 2900 PetscFunctionReturn(0); 2901 } 2902 2903 PETSC_INTERN PetscErrorCode MatSeqAIJGetArray_SeqAIJ(Mat A,PetscScalar *array[]) 2904 { 2905 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 2906 2907 PetscFunctionBegin; 2908 *array = a->a; 2909 PetscFunctionReturn(0); 2910 } 2911 2912 PETSC_INTERN PetscErrorCode MatSeqAIJRestoreArray_SeqAIJ(Mat A,PetscScalar *array[]) 2913 { 2914 PetscFunctionBegin; 2915 *array = NULL; 2916 PetscFunctionReturn(0); 2917 } 2918 2919 /* 2920 Computes the number of nonzeros per row needed for preallocation when X and Y 2921 have different nonzero structure. 2922 */ 2923 PetscErrorCode MatAXPYGetPreallocation_SeqX_private(PetscInt m,const PetscInt *xi,const PetscInt *xj,const PetscInt *yi,const PetscInt *yj,PetscInt *nnz) 2924 { 2925 PetscInt i,j,k,nzx,nzy; 2926 2927 PetscFunctionBegin; 2928 /* Set the number of nonzeros in the new matrix */ 2929 for (i=0; i<m; i++) { 2930 const PetscInt *xjj = xj+xi[i],*yjj = yj+yi[i]; 2931 nzx = xi[i+1] - xi[i]; 2932 nzy = yi[i+1] - yi[i]; 2933 nnz[i] = 0; 2934 for (j=0,k=0; j<nzx; j++) { /* Point in X */ 2935 for (; k<nzy && yjj[k]<xjj[j]; k++) nnz[i]++; /* Catch up to X */ 2936 if (k<nzy && yjj[k]==xjj[j]) k++; /* Skip duplicate */ 2937 nnz[i]++; 2938 } 2939 for (; k<nzy; k++) nnz[i]++; 2940 } 2941 PetscFunctionReturn(0); 2942 } 2943 2944 PetscErrorCode MatAXPYGetPreallocation_SeqAIJ(Mat Y,Mat X,PetscInt *nnz) 2945 { 2946 PetscInt m = Y->rmap->N; 2947 Mat_SeqAIJ *x = (Mat_SeqAIJ*)X->data; 2948 Mat_SeqAIJ *y = (Mat_SeqAIJ*)Y->data; 2949 PetscErrorCode ierr; 2950 2951 PetscFunctionBegin; 2952 /* Set the number of nonzeros in the new matrix */ 2953 ierr = MatAXPYGetPreallocation_SeqX_private(m,x->i,x->j,y->i,y->j,nnz);CHKERRQ(ierr); 2954 PetscFunctionReturn(0); 2955 } 2956 2957 PetscErrorCode MatAXPY_SeqAIJ(Mat Y,PetscScalar a,Mat X,MatStructure str) 2958 { 2959 PetscErrorCode ierr; 2960 Mat_SeqAIJ *x = (Mat_SeqAIJ*)X->data,*y = (Mat_SeqAIJ*)Y->data; 2961 PetscBLASInt one=1,bnz; 2962 2963 PetscFunctionBegin; 2964 ierr = PetscBLASIntCast(x->nz,&bnz);CHKERRQ(ierr); 2965 if (str == SAME_NONZERO_PATTERN) { 2966 PetscScalar alpha = a; 2967 PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&bnz,&alpha,x->a,&one,y->a,&one)); 2968 ierr = MatSeqAIJInvalidateDiagonal(Y);CHKERRQ(ierr); 2969 ierr = PetscObjectStateIncrease((PetscObject)Y);CHKERRQ(ierr); 2970 /* the MatAXPY_Basic* subroutines calls MatAssembly, so the matrix on the GPU 2971 will be updated */ 2972 #if defined(PETSC_HAVE_VIENNACL) || defined(PETSC_HAVE_CUDA) 2973 if (Y->offloadmask != PETSC_OFFLOAD_UNALLOCATED) { 2974 Y->offloadmask = PETSC_OFFLOAD_CPU; 2975 } 2976 #endif 2977 } else if (str == SUBSET_NONZERO_PATTERN) { /* nonzeros of X is a subset of Y's */ 2978 ierr = MatAXPY_Basic(Y,a,X,str);CHKERRQ(ierr); 2979 } else { 2980 Mat B; 2981 PetscInt *nnz; 2982 ierr = PetscMalloc1(Y->rmap->N,&nnz);CHKERRQ(ierr); 2983 ierr = MatCreate(PetscObjectComm((PetscObject)Y),&B);CHKERRQ(ierr); 2984 ierr = PetscObjectSetName((PetscObject)B,((PetscObject)Y)->name);CHKERRQ(ierr); 2985 ierr = MatSetSizes(B,Y->rmap->n,Y->cmap->n,Y->rmap->N,Y->cmap->N);CHKERRQ(ierr); 2986 ierr = MatSetBlockSizesFromMats(B,Y,Y);CHKERRQ(ierr); 2987 ierr = MatSetType(B,(MatType) ((PetscObject)Y)->type_name);CHKERRQ(ierr); 2988 ierr = MatAXPYGetPreallocation_SeqAIJ(Y,X,nnz);CHKERRQ(ierr); 2989 ierr = MatSeqAIJSetPreallocation(B,0,nnz);CHKERRQ(ierr); 2990 ierr = MatAXPY_BasicWithPreallocation(B,Y,a,X,str);CHKERRQ(ierr); 2991 ierr = MatHeaderReplace(Y,&B);CHKERRQ(ierr); 2992 ierr = PetscFree(nnz);CHKERRQ(ierr); 2993 } 2994 PetscFunctionReturn(0); 2995 } 2996 2997 PetscErrorCode MatConjugate_SeqAIJ(Mat mat) 2998 { 2999 #if defined(PETSC_USE_COMPLEX) 3000 Mat_SeqAIJ *aij = (Mat_SeqAIJ*)mat->data; 3001 PetscInt i,nz; 3002 PetscScalar *a; 3003 3004 PetscFunctionBegin; 3005 nz = aij->nz; 3006 a = aij->a; 3007 for (i=0; i<nz; i++) a[i] = PetscConj(a[i]); 3008 #if defined(PETSC_HAVE_VIENNACL) || defined(PETSC_HAVE_CUDA) 3009 if (mat->offloadmask != PETSC_OFFLOAD_UNALLOCATED) mat->offloadmask = PETSC_OFFLOAD_CPU; 3010 #endif 3011 #else 3012 PetscFunctionBegin; 3013 #endif 3014 PetscFunctionReturn(0); 3015 } 3016 3017 PetscErrorCode MatGetRowMaxAbs_SeqAIJ(Mat A,Vec v,PetscInt idx[]) 3018 { 3019 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 3020 PetscErrorCode ierr; 3021 PetscInt i,j,m = A->rmap->n,*ai,*aj,ncols,n; 3022 PetscReal atmp; 3023 PetscScalar *x; 3024 MatScalar *aa; 3025 3026 PetscFunctionBegin; 3027 if (A->factortype) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Not for factored matrix"); 3028 aa = a->a; 3029 ai = a->i; 3030 aj = a->j; 3031 3032 ierr = VecSet(v,0.0);CHKERRQ(ierr); 3033 ierr = VecGetArray(v,&x);CHKERRQ(ierr); 3034 ierr = VecGetLocalSize(v,&n);CHKERRQ(ierr); 3035 if (n != A->rmap->n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Nonconforming matrix and vector"); 3036 for (i=0; i<m; i++) { 3037 ncols = ai[1] - ai[0]; ai++; 3038 x[i] = 0.0; 3039 for (j=0; j<ncols; j++) { 3040 atmp = PetscAbsScalar(*aa); 3041 if (PetscAbsScalar(x[i]) < atmp) {x[i] = atmp; if (idx) idx[i] = *aj;} 3042 aa++; aj++; 3043 } 3044 } 3045 ierr = VecRestoreArray(v,&x);CHKERRQ(ierr); 3046 PetscFunctionReturn(0); 3047 } 3048 3049 PetscErrorCode MatGetRowMax_SeqAIJ(Mat A,Vec v,PetscInt idx[]) 3050 { 3051 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 3052 PetscErrorCode ierr; 3053 PetscInt i,j,m = A->rmap->n,*ai,*aj,ncols,n; 3054 PetscScalar *x; 3055 MatScalar *aa; 3056 3057 PetscFunctionBegin; 3058 if (A->factortype) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Not for factored matrix"); 3059 aa = a->a; 3060 ai = a->i; 3061 aj = a->j; 3062 3063 ierr = VecSet(v,0.0);CHKERRQ(ierr); 3064 ierr = VecGetArray(v,&x);CHKERRQ(ierr); 3065 ierr = VecGetLocalSize(v,&n);CHKERRQ(ierr); 3066 if (n != A->rmap->n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Nonconforming matrix and vector"); 3067 for (i=0; i<m; i++) { 3068 ncols = ai[1] - ai[0]; ai++; 3069 if (ncols == A->cmap->n) { /* row is dense */ 3070 x[i] = *aa; if (idx) idx[i] = 0; 3071 } else { /* row is sparse so already KNOW maximum is 0.0 or higher */ 3072 x[i] = 0.0; 3073 if (idx) { 3074 idx[i] = 0; /* in case ncols is zero */ 3075 for (j=0;j<ncols;j++) { /* find first implicit 0.0 in the row */ 3076 if (aj[j] > j) { 3077 idx[i] = j; 3078 break; 3079 } 3080 } 3081 } 3082 } 3083 for (j=0; j<ncols; j++) { 3084 if (PetscRealPart(x[i]) < PetscRealPart(*aa)) {x[i] = *aa; if (idx) idx[i] = *aj;} 3085 aa++; aj++; 3086 } 3087 } 3088 ierr = VecRestoreArray(v,&x);CHKERRQ(ierr); 3089 PetscFunctionReturn(0); 3090 } 3091 3092 PetscErrorCode MatGetRowMinAbs_SeqAIJ(Mat A,Vec v,PetscInt idx[]) 3093 { 3094 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 3095 PetscErrorCode ierr; 3096 PetscInt i,j,m = A->rmap->n,*ai,*aj,ncols,n; 3097 PetscReal atmp; 3098 PetscScalar *x; 3099 MatScalar *aa; 3100 3101 PetscFunctionBegin; 3102 if (A->factortype) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Not for factored matrix"); 3103 aa = a->a; 3104 ai = a->i; 3105 aj = a->j; 3106 3107 ierr = VecSet(v,0.0);CHKERRQ(ierr); 3108 ierr = VecGetArray(v,&x);CHKERRQ(ierr); 3109 ierr = VecGetLocalSize(v,&n);CHKERRQ(ierr); 3110 if (n != A->rmap->n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Nonconforming matrix and vector, %D vs. %D rows", A->rmap->n, n); 3111 for (i=0; i<m; i++) { 3112 ncols = ai[1] - ai[0]; ai++; 3113 if (ncols) { 3114 /* Get first nonzero */ 3115 for (j = 0; j < ncols; j++) { 3116 atmp = PetscAbsScalar(aa[j]); 3117 if (atmp > 1.0e-12) { 3118 x[i] = atmp; 3119 if (idx) idx[i] = aj[j]; 3120 break; 3121 } 3122 } 3123 if (j == ncols) {x[i] = PetscAbsScalar(*aa); if (idx) idx[i] = *aj;} 3124 } else { 3125 x[i] = 0.0; if (idx) idx[i] = 0; 3126 } 3127 for (j = 0; j < ncols; j++) { 3128 atmp = PetscAbsScalar(*aa); 3129 if (atmp > 1.0e-12 && PetscAbsScalar(x[i]) > atmp) {x[i] = atmp; if (idx) idx[i] = *aj;} 3130 aa++; aj++; 3131 } 3132 } 3133 ierr = VecRestoreArray(v,&x);CHKERRQ(ierr); 3134 PetscFunctionReturn(0); 3135 } 3136 3137 PetscErrorCode MatGetRowMin_SeqAIJ(Mat A,Vec v,PetscInt idx[]) 3138 { 3139 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 3140 PetscErrorCode ierr; 3141 PetscInt i,j,m = A->rmap->n,ncols,n; 3142 const PetscInt *ai,*aj; 3143 PetscScalar *x; 3144 const MatScalar *aa; 3145 3146 PetscFunctionBegin; 3147 if (A->factortype) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Not for factored matrix"); 3148 aa = a->a; 3149 ai = a->i; 3150 aj = a->j; 3151 3152 ierr = VecSet(v,0.0);CHKERRQ(ierr); 3153 ierr = VecGetArray(v,&x);CHKERRQ(ierr); 3154 ierr = VecGetLocalSize(v,&n);CHKERRQ(ierr); 3155 if (n != A->rmap->n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Nonconforming matrix and vector"); 3156 for (i=0; i<m; i++) { 3157 ncols = ai[1] - ai[0]; ai++; 3158 if (ncols == A->cmap->n) { /* row is dense */ 3159 x[i] = *aa; if (idx) idx[i] = 0; 3160 } else { /* row is sparse so already KNOW minimum is 0.0 or lower */ 3161 x[i] = 0.0; 3162 if (idx) { /* find first implicit 0.0 in the row */ 3163 idx[i] = 0; /* in case ncols is zero */ 3164 for (j=0; j<ncols; j++) { 3165 if (aj[j] > j) { 3166 idx[i] = j; 3167 break; 3168 } 3169 } 3170 } 3171 } 3172 for (j=0; j<ncols; j++) { 3173 if (PetscRealPart(x[i]) > PetscRealPart(*aa)) {x[i] = *aa; if (idx) idx[i] = *aj;} 3174 aa++; aj++; 3175 } 3176 } 3177 ierr = VecRestoreArray(v,&x);CHKERRQ(ierr); 3178 PetscFunctionReturn(0); 3179 } 3180 3181 PetscErrorCode MatInvertBlockDiagonal_SeqAIJ(Mat A,const PetscScalar **values) 3182 { 3183 Mat_SeqAIJ *a = (Mat_SeqAIJ*) A->data; 3184 PetscErrorCode ierr; 3185 PetscInt i,bs = PetscAbs(A->rmap->bs),mbs = A->rmap->n/bs,ipvt[5],bs2 = bs*bs,*v_pivots,ij[7],*IJ,j; 3186 MatScalar *diag,work[25],*v_work; 3187 const PetscReal shift = 0.0; 3188 PetscBool allowzeropivot,zeropivotdetected=PETSC_FALSE; 3189 3190 PetscFunctionBegin; 3191 allowzeropivot = PetscNot(A->erroriffailure); 3192 if (a->ibdiagvalid) { 3193 if (values) *values = a->ibdiag; 3194 PetscFunctionReturn(0); 3195 } 3196 ierr = MatMarkDiagonal_SeqAIJ(A);CHKERRQ(ierr); 3197 if (!a->ibdiag) { 3198 ierr = PetscMalloc1(bs2*mbs,&a->ibdiag);CHKERRQ(ierr); 3199 ierr = PetscLogObjectMemory((PetscObject)A,bs2*mbs*sizeof(PetscScalar));CHKERRQ(ierr); 3200 } 3201 diag = a->ibdiag; 3202 if (values) *values = a->ibdiag; 3203 /* factor and invert each block */ 3204 switch (bs) { 3205 case 1: 3206 for (i=0; i<mbs; i++) { 3207 ierr = MatGetValues(A,1,&i,1,&i,diag+i);CHKERRQ(ierr); 3208 if (PetscAbsScalar(diag[i] + shift) < PETSC_MACHINE_EPSILON) { 3209 if (allowzeropivot) { 3210 A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT; 3211 A->factorerror_zeropivot_value = PetscAbsScalar(diag[i]); 3212 A->factorerror_zeropivot_row = i; 3213 ierr = PetscInfo3(A,"Zero pivot, row %D pivot %g tolerance %g\n",i,(double)PetscAbsScalar(diag[i]),(double)PETSC_MACHINE_EPSILON);CHKERRQ(ierr); 3214 } else SETERRQ3(PETSC_COMM_SELF,PETSC_ERR_MAT_LU_ZRPVT,"Zero pivot, row %D pivot %g tolerance %g",i,(double)PetscAbsScalar(diag[i]),(double)PETSC_MACHINE_EPSILON); 3215 } 3216 diag[i] = (PetscScalar)1.0 / (diag[i] + shift); 3217 } 3218 break; 3219 case 2: 3220 for (i=0; i<mbs; i++) { 3221 ij[0] = 2*i; ij[1] = 2*i + 1; 3222 ierr = MatGetValues(A,2,ij,2,ij,diag);CHKERRQ(ierr); 3223 ierr = PetscKernel_A_gets_inverse_A_2(diag,shift,allowzeropivot,&zeropivotdetected);CHKERRQ(ierr); 3224 if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT; 3225 ierr = PetscKernel_A_gets_transpose_A_2(diag);CHKERRQ(ierr); 3226 diag += 4; 3227 } 3228 break; 3229 case 3: 3230 for (i=0; i<mbs; i++) { 3231 ij[0] = 3*i; ij[1] = 3*i + 1; ij[2] = 3*i + 2; 3232 ierr = MatGetValues(A,3,ij,3,ij,diag);CHKERRQ(ierr); 3233 ierr = PetscKernel_A_gets_inverse_A_3(diag,shift,allowzeropivot,&zeropivotdetected);CHKERRQ(ierr); 3234 if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT; 3235 ierr = PetscKernel_A_gets_transpose_A_3(diag);CHKERRQ(ierr); 3236 diag += 9; 3237 } 3238 break; 3239 case 4: 3240 for (i=0; i<mbs; i++) { 3241 ij[0] = 4*i; ij[1] = 4*i + 1; ij[2] = 4*i + 2; ij[3] = 4*i + 3; 3242 ierr = MatGetValues(A,4,ij,4,ij,diag);CHKERRQ(ierr); 3243 ierr = PetscKernel_A_gets_inverse_A_4(diag,shift,allowzeropivot,&zeropivotdetected);CHKERRQ(ierr); 3244 if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT; 3245 ierr = PetscKernel_A_gets_transpose_A_4(diag);CHKERRQ(ierr); 3246 diag += 16; 3247 } 3248 break; 3249 case 5: 3250 for (i=0; i<mbs; i++) { 3251 ij[0] = 5*i; ij[1] = 5*i + 1; ij[2] = 5*i + 2; ij[3] = 5*i + 3; ij[4] = 5*i + 4; 3252 ierr = MatGetValues(A,5,ij,5,ij,diag);CHKERRQ(ierr); 3253 ierr = PetscKernel_A_gets_inverse_A_5(diag,ipvt,work,shift,allowzeropivot,&zeropivotdetected);CHKERRQ(ierr); 3254 if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT; 3255 ierr = PetscKernel_A_gets_transpose_A_5(diag);CHKERRQ(ierr); 3256 diag += 25; 3257 } 3258 break; 3259 case 6: 3260 for (i=0; i<mbs; i++) { 3261 ij[0] = 6*i; ij[1] = 6*i + 1; ij[2] = 6*i + 2; ij[3] = 6*i + 3; ij[4] = 6*i + 4; ij[5] = 6*i + 5; 3262 ierr = MatGetValues(A,6,ij,6,ij,diag);CHKERRQ(ierr); 3263 ierr = PetscKernel_A_gets_inverse_A_6(diag,shift,allowzeropivot,&zeropivotdetected);CHKERRQ(ierr); 3264 if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT; 3265 ierr = PetscKernel_A_gets_transpose_A_6(diag);CHKERRQ(ierr); 3266 diag += 36; 3267 } 3268 break; 3269 case 7: 3270 for (i=0; i<mbs; i++) { 3271 ij[0] = 7*i; ij[1] = 7*i + 1; ij[2] = 7*i + 2; ij[3] = 7*i + 3; ij[4] = 7*i + 4; ij[5] = 7*i + 5; ij[5] = 7*i + 6; 3272 ierr = MatGetValues(A,7,ij,7,ij,diag);CHKERRQ(ierr); 3273 ierr = PetscKernel_A_gets_inverse_A_7(diag,shift,allowzeropivot,&zeropivotdetected);CHKERRQ(ierr); 3274 if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT; 3275 ierr = PetscKernel_A_gets_transpose_A_7(diag);CHKERRQ(ierr); 3276 diag += 49; 3277 } 3278 break; 3279 default: 3280 ierr = PetscMalloc3(bs,&v_work,bs,&v_pivots,bs,&IJ);CHKERRQ(ierr); 3281 for (i=0; i<mbs; i++) { 3282 for (j=0; j<bs; j++) { 3283 IJ[j] = bs*i + j; 3284 } 3285 ierr = MatGetValues(A,bs,IJ,bs,IJ,diag);CHKERRQ(ierr); 3286 ierr = PetscKernel_A_gets_inverse_A(bs,diag,v_pivots,v_work,allowzeropivot,&zeropivotdetected);CHKERRQ(ierr); 3287 if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT; 3288 ierr = PetscKernel_A_gets_transpose_A_N(diag,bs);CHKERRQ(ierr); 3289 diag += bs2; 3290 } 3291 ierr = PetscFree3(v_work,v_pivots,IJ);CHKERRQ(ierr); 3292 } 3293 a->ibdiagvalid = PETSC_TRUE; 3294 PetscFunctionReturn(0); 3295 } 3296 3297 static PetscErrorCode MatSetRandom_SeqAIJ(Mat x,PetscRandom rctx) 3298 { 3299 PetscErrorCode ierr; 3300 Mat_SeqAIJ *aij = (Mat_SeqAIJ*)x->data; 3301 PetscScalar a; 3302 PetscInt m,n,i,j,col; 3303 3304 PetscFunctionBegin; 3305 if (!x->assembled) { 3306 ierr = MatGetSize(x,&m,&n);CHKERRQ(ierr); 3307 for (i=0; i<m; i++) { 3308 for (j=0; j<aij->imax[i]; j++) { 3309 ierr = PetscRandomGetValue(rctx,&a);CHKERRQ(ierr); 3310 col = (PetscInt)(n*PetscRealPart(a)); 3311 ierr = MatSetValues(x,1,&i,1,&col,&a,ADD_VALUES);CHKERRQ(ierr); 3312 } 3313 } 3314 } else { 3315 for (i=0; i<aij->nz; i++) {ierr = PetscRandomGetValue(rctx,aij->a+i);CHKERRQ(ierr);} 3316 } 3317 ierr = MatAssemblyBegin(x,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3318 ierr = MatAssemblyEnd(x,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3319 PetscFunctionReturn(0); 3320 } 3321 3322 /* Like MatSetRandom_SeqAIJ, but do not set values on columns in range of [low, high) */ 3323 PetscErrorCode MatSetRandomSkipColumnRange_SeqAIJ_Private(Mat x,PetscInt low,PetscInt high,PetscRandom rctx) 3324 { 3325 PetscErrorCode ierr; 3326 Mat_SeqAIJ *aij = (Mat_SeqAIJ*)x->data; 3327 PetscScalar a; 3328 PetscInt m,n,i,j,col,nskip; 3329 3330 PetscFunctionBegin; 3331 nskip = high - low; 3332 ierr = MatGetSize(x,&m,&n);CHKERRQ(ierr); 3333 n -= nskip; /* shrink number of columns where nonzeros can be set */ 3334 for (i=0; i<m; i++) { 3335 for (j=0; j<aij->imax[i]; j++) { 3336 ierr = PetscRandomGetValue(rctx,&a);CHKERRQ(ierr); 3337 col = (PetscInt)(n*PetscRealPart(a)); 3338 if (col >= low) col += nskip; /* shift col rightward to skip the hole */ 3339 ierr = MatSetValues(x,1,&i,1,&col,&a,ADD_VALUES);CHKERRQ(ierr); 3340 } 3341 } 3342 ierr = MatAssemblyBegin(x,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3343 ierr = MatAssemblyEnd(x,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3344 PetscFunctionReturn(0); 3345 } 3346 3347 3348 /* -------------------------------------------------------------------*/ 3349 static struct _MatOps MatOps_Values = { MatSetValues_SeqAIJ, 3350 MatGetRow_SeqAIJ, 3351 MatRestoreRow_SeqAIJ, 3352 MatMult_SeqAIJ, 3353 /* 4*/ MatMultAdd_SeqAIJ, 3354 MatMultTranspose_SeqAIJ, 3355 MatMultTransposeAdd_SeqAIJ, 3356 0, 3357 0, 3358 0, 3359 /* 10*/ 0, 3360 MatLUFactor_SeqAIJ, 3361 0, 3362 MatSOR_SeqAIJ, 3363 MatTranspose_SeqAIJ, 3364 /*1 5*/ MatGetInfo_SeqAIJ, 3365 MatEqual_SeqAIJ, 3366 MatGetDiagonal_SeqAIJ, 3367 MatDiagonalScale_SeqAIJ, 3368 MatNorm_SeqAIJ, 3369 /* 20*/ 0, 3370 MatAssemblyEnd_SeqAIJ, 3371 MatSetOption_SeqAIJ, 3372 MatZeroEntries_SeqAIJ, 3373 /* 24*/ MatZeroRows_SeqAIJ, 3374 0, 3375 0, 3376 0, 3377 0, 3378 /* 29*/ MatSetUp_SeqAIJ, 3379 0, 3380 0, 3381 0, 3382 0, 3383 /* 34*/ MatDuplicate_SeqAIJ, 3384 0, 3385 0, 3386 MatILUFactor_SeqAIJ, 3387 0, 3388 /* 39*/ MatAXPY_SeqAIJ, 3389 MatCreateSubMatrices_SeqAIJ, 3390 MatIncreaseOverlap_SeqAIJ, 3391 MatGetValues_SeqAIJ, 3392 MatCopy_SeqAIJ, 3393 /* 44*/ MatGetRowMax_SeqAIJ, 3394 MatScale_SeqAIJ, 3395 MatShift_SeqAIJ, 3396 MatDiagonalSet_SeqAIJ, 3397 MatZeroRowsColumns_SeqAIJ, 3398 /* 49*/ MatSetRandom_SeqAIJ, 3399 MatGetRowIJ_SeqAIJ, 3400 MatRestoreRowIJ_SeqAIJ, 3401 MatGetColumnIJ_SeqAIJ, 3402 MatRestoreColumnIJ_SeqAIJ, 3403 /* 54*/ MatFDColoringCreate_SeqXAIJ, 3404 0, 3405 0, 3406 MatPermute_SeqAIJ, 3407 0, 3408 /* 59*/ 0, 3409 MatDestroy_SeqAIJ, 3410 MatView_SeqAIJ, 3411 0, 3412 MatMatMatMult_SeqAIJ_SeqAIJ_SeqAIJ, 3413 /* 64*/ MatMatMatMultSymbolic_SeqAIJ_SeqAIJ_SeqAIJ, 3414 MatMatMatMultNumeric_SeqAIJ_SeqAIJ_SeqAIJ, 3415 0, 3416 0, 3417 0, 3418 /* 69*/ MatGetRowMaxAbs_SeqAIJ, 3419 MatGetRowMinAbs_SeqAIJ, 3420 0, 3421 0, 3422 0, 3423 /* 74*/ 0, 3424 MatFDColoringApply_AIJ, 3425 0, 3426 0, 3427 0, 3428 /* 79*/ MatFindZeroDiagonals_SeqAIJ, 3429 0, 3430 0, 3431 0, 3432 MatLoad_SeqAIJ, 3433 /* 84*/ MatIsSymmetric_SeqAIJ, 3434 MatIsHermitian_SeqAIJ, 3435 0, 3436 0, 3437 0, 3438 /* 89*/ MatMatMult_SeqAIJ_SeqAIJ, 3439 MatMatMultSymbolic_SeqAIJ_SeqAIJ, 3440 MatMatMultNumeric_SeqAIJ_SeqAIJ, 3441 MatPtAP_SeqAIJ_SeqAIJ, 3442 MatPtAPSymbolic_SeqAIJ_SeqAIJ_SparseAxpy, 3443 /* 94*/ MatPtAPNumeric_SeqAIJ_SeqAIJ_SparseAxpy, 3444 MatMatTransposeMult_SeqAIJ_SeqAIJ, 3445 MatMatTransposeMultSymbolic_SeqAIJ_SeqAIJ, 3446 MatMatTransposeMultNumeric_SeqAIJ_SeqAIJ, 3447 0, 3448 /* 99*/ 0, 3449 0, 3450 0, 3451 MatConjugate_SeqAIJ, 3452 0, 3453 /*104*/ MatSetValuesRow_SeqAIJ, 3454 MatRealPart_SeqAIJ, 3455 MatImaginaryPart_SeqAIJ, 3456 0, 3457 0, 3458 /*109*/ MatMatSolve_SeqAIJ, 3459 0, 3460 MatGetRowMin_SeqAIJ, 3461 0, 3462 MatMissingDiagonal_SeqAIJ, 3463 /*114*/ 0, 3464 0, 3465 0, 3466 0, 3467 0, 3468 /*119*/ 0, 3469 0, 3470 0, 3471 0, 3472 MatGetMultiProcBlock_SeqAIJ, 3473 /*124*/ MatFindNonzeroRows_SeqAIJ, 3474 MatGetColumnNorms_SeqAIJ, 3475 MatInvertBlockDiagonal_SeqAIJ, 3476 MatInvertVariableBlockDiagonal_SeqAIJ, 3477 0, 3478 /*129*/ 0, 3479 MatTransposeMatMult_SeqAIJ_SeqAIJ, 3480 MatTransposeMatMultSymbolic_SeqAIJ_SeqAIJ, 3481 MatTransposeMatMultNumeric_SeqAIJ_SeqAIJ, 3482 MatTransposeColoringCreate_SeqAIJ, 3483 /*134*/ MatTransColoringApplySpToDen_SeqAIJ, 3484 MatTransColoringApplyDenToSp_SeqAIJ, 3485 MatRARt_SeqAIJ_SeqAIJ, 3486 MatRARtSymbolic_SeqAIJ_SeqAIJ, 3487 MatRARtNumeric_SeqAIJ_SeqAIJ, 3488 /*139*/0, 3489 0, 3490 0, 3491 MatFDColoringSetUp_SeqXAIJ, 3492 MatFindOffBlockDiagonalEntries_SeqAIJ, 3493 /*144*/MatCreateMPIMatConcatenateSeqMat_SeqAIJ, 3494 MatDestroySubMatrices_SeqAIJ 3495 }; 3496 3497 PetscErrorCode MatSeqAIJSetColumnIndices_SeqAIJ(Mat mat,PetscInt *indices) 3498 { 3499 Mat_SeqAIJ *aij = (Mat_SeqAIJ*)mat->data; 3500 PetscInt i,nz,n; 3501 3502 PetscFunctionBegin; 3503 nz = aij->maxnz; 3504 n = mat->rmap->n; 3505 for (i=0; i<nz; i++) { 3506 aij->j[i] = indices[i]; 3507 } 3508 aij->nz = nz; 3509 for (i=0; i<n; i++) { 3510 aij->ilen[i] = aij->imax[i]; 3511 } 3512 PetscFunctionReturn(0); 3513 } 3514 3515 /* 3516 * When a sparse matrix has many zero columns, we should compact them out to save the space 3517 * This happens in MatPtAPSymbolic_MPIAIJ_MPIAIJ_scalable() 3518 * */ 3519 PetscErrorCode MatSeqAIJCompactOutExtraColumns_SeqAIJ(Mat mat, ISLocalToGlobalMapping *mapping) 3520 { 3521 Mat_SeqAIJ *aij = (Mat_SeqAIJ*)mat->data; 3522 PetscTable gid1_lid1; 3523 PetscTablePosition tpos; 3524 PetscInt gid,lid,i,j,ncols,ec; 3525 PetscInt *garray; 3526 PetscErrorCode ierr; 3527 3528 PetscFunctionBegin; 3529 PetscValidHeaderSpecific(mat,MAT_CLASSID,1); 3530 PetscValidPointer(mapping,2); 3531 /* use a table */ 3532 ierr = PetscTableCreate(mat->rmap->n,mat->cmap->N+1,&gid1_lid1);CHKERRQ(ierr); 3533 ec = 0; 3534 for (i=0; i<mat->rmap->n; i++) { 3535 ncols = aij->i[i+1] - aij->i[i]; 3536 for (j=0; j<ncols; j++) { 3537 PetscInt data,gid1 = aij->j[aij->i[i] + j] + 1; 3538 ierr = PetscTableFind(gid1_lid1,gid1,&data);CHKERRQ(ierr); 3539 if (!data) { 3540 /* one based table */ 3541 ierr = PetscTableAdd(gid1_lid1,gid1,++ec,INSERT_VALUES);CHKERRQ(ierr); 3542 } 3543 } 3544 } 3545 /* form array of columns we need */ 3546 ierr = PetscMalloc1(ec+1,&garray);CHKERRQ(ierr); 3547 ierr = PetscTableGetHeadPosition(gid1_lid1,&tpos);CHKERRQ(ierr); 3548 while (tpos) { 3549 ierr = PetscTableGetNext(gid1_lid1,&tpos,&gid,&lid);CHKERRQ(ierr); 3550 gid--; 3551 lid--; 3552 garray[lid] = gid; 3553 } 3554 ierr = PetscSortInt(ec,garray);CHKERRQ(ierr); /* sort, and rebuild */ 3555 ierr = PetscTableRemoveAll(gid1_lid1);CHKERRQ(ierr); 3556 for (i=0; i<ec; i++) { 3557 ierr = PetscTableAdd(gid1_lid1,garray[i]+1,i+1,INSERT_VALUES);CHKERRQ(ierr); 3558 } 3559 /* compact out the extra columns in B */ 3560 for (i=0; i<mat->rmap->n; i++) { 3561 ncols = aij->i[i+1] - aij->i[i]; 3562 for (j=0; j<ncols; j++) { 3563 PetscInt gid1 = aij->j[aij->i[i] + j] + 1; 3564 ierr = PetscTableFind(gid1_lid1,gid1,&lid);CHKERRQ(ierr); 3565 lid--; 3566 aij->j[aij->i[i] + j] = lid; 3567 } 3568 } 3569 ierr = PetscLayoutDestroy(&mat->cmap);CHKERRQ(ierr); 3570 ierr = PetscLayoutCreateFromSizes(PetscObjectComm((PetscObject)mat),ec,ec,1,&mat->cmap);CHKERRQ(ierr); 3571 ierr = PetscTableDestroy(&gid1_lid1);CHKERRQ(ierr); 3572 ierr = ISLocalToGlobalMappingCreate(PETSC_COMM_SELF,mat->cmap->bs,mat->cmap->n,garray,PETSC_OWN_POINTER,mapping);CHKERRQ(ierr); 3573 ierr = ISLocalToGlobalMappingSetType(*mapping,ISLOCALTOGLOBALMAPPINGHASH);CHKERRQ(ierr); 3574 PetscFunctionReturn(0); 3575 } 3576 3577 /*@ 3578 MatSeqAIJSetColumnIndices - Set the column indices for all the rows 3579 in the matrix. 3580 3581 Input Parameters: 3582 + mat - the SeqAIJ matrix 3583 - indices - the column indices 3584 3585 Level: advanced 3586 3587 Notes: 3588 This can be called if you have precomputed the nonzero structure of the 3589 matrix and want to provide it to the matrix object to improve the performance 3590 of the MatSetValues() operation. 3591 3592 You MUST have set the correct numbers of nonzeros per row in the call to 3593 MatCreateSeqAIJ(), and the columns indices MUST be sorted. 3594 3595 MUST be called before any calls to MatSetValues(); 3596 3597 The indices should start with zero, not one. 3598 3599 @*/ 3600 PetscErrorCode MatSeqAIJSetColumnIndices(Mat mat,PetscInt *indices) 3601 { 3602 PetscErrorCode ierr; 3603 3604 PetscFunctionBegin; 3605 PetscValidHeaderSpecific(mat,MAT_CLASSID,1); 3606 PetscValidPointer(indices,2); 3607 ierr = PetscUseMethod(mat,"MatSeqAIJSetColumnIndices_C",(Mat,PetscInt*),(mat,indices));CHKERRQ(ierr); 3608 PetscFunctionReturn(0); 3609 } 3610 3611 /* ----------------------------------------------------------------------------------------*/ 3612 3613 PetscErrorCode MatStoreValues_SeqAIJ(Mat mat) 3614 { 3615 Mat_SeqAIJ *aij = (Mat_SeqAIJ*)mat->data; 3616 PetscErrorCode ierr; 3617 size_t nz = aij->i[mat->rmap->n]; 3618 3619 PetscFunctionBegin; 3620 if (!aij->nonew) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ORDER,"Must call MatSetOption(A,MAT_NEW_NONZERO_LOCATIONS,PETSC_FALSE);first"); 3621 3622 /* allocate space for values if not already there */ 3623 if (!aij->saved_values) { 3624 ierr = PetscMalloc1(nz+1,&aij->saved_values);CHKERRQ(ierr); 3625 ierr = PetscLogObjectMemory((PetscObject)mat,(nz+1)*sizeof(PetscScalar));CHKERRQ(ierr); 3626 } 3627 3628 /* copy values over */ 3629 ierr = PetscArraycpy(aij->saved_values,aij->a,nz);CHKERRQ(ierr); 3630 PetscFunctionReturn(0); 3631 } 3632 3633 /*@ 3634 MatStoreValues - Stashes a copy of the matrix values; this allows, for 3635 example, reuse of the linear part of a Jacobian, while recomputing the 3636 nonlinear portion. 3637 3638 Collect on Mat 3639 3640 Input Parameters: 3641 . mat - the matrix (currently only AIJ matrices support this option) 3642 3643 Level: advanced 3644 3645 Common Usage, with SNESSolve(): 3646 $ Create Jacobian matrix 3647 $ Set linear terms into matrix 3648 $ Apply boundary conditions to matrix, at this time matrix must have 3649 $ final nonzero structure (i.e. setting the nonlinear terms and applying 3650 $ boundary conditions again will not change the nonzero structure 3651 $ ierr = MatSetOption(mat,MAT_NEW_NONZERO_LOCATIONS,PETSC_FALSE); 3652 $ ierr = MatStoreValues(mat); 3653 $ Call SNESSetJacobian() with matrix 3654 $ In your Jacobian routine 3655 $ ierr = MatRetrieveValues(mat); 3656 $ Set nonlinear terms in matrix 3657 3658 Common Usage without SNESSolve(), i.e. when you handle nonlinear solve yourself: 3659 $ // build linear portion of Jacobian 3660 $ ierr = MatSetOption(mat,MAT_NEW_NONZERO_LOCATIONS,PETSC_FALSE); 3661 $ ierr = MatStoreValues(mat); 3662 $ loop over nonlinear iterations 3663 $ ierr = MatRetrieveValues(mat); 3664 $ // call MatSetValues(mat,...) to set nonliner portion of Jacobian 3665 $ // call MatAssemblyBegin/End() on matrix 3666 $ Solve linear system with Jacobian 3667 $ endloop 3668 3669 Notes: 3670 Matrix must already be assemblied before calling this routine 3671 Must set the matrix option MatSetOption(mat,MAT_NEW_NONZERO_LOCATIONS,PETSC_FALSE); before 3672 calling this routine. 3673 3674 When this is called multiple times it overwrites the previous set of stored values 3675 and does not allocated additional space. 3676 3677 .seealso: MatRetrieveValues() 3678 3679 @*/ 3680 PetscErrorCode MatStoreValues(Mat mat) 3681 { 3682 PetscErrorCode ierr; 3683 3684 PetscFunctionBegin; 3685 PetscValidHeaderSpecific(mat,MAT_CLASSID,1); 3686 if (!mat->assembled) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Not for unassembled matrix"); 3687 if (mat->factortype) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Not for factored matrix"); 3688 ierr = PetscUseMethod(mat,"MatStoreValues_C",(Mat),(mat));CHKERRQ(ierr); 3689 PetscFunctionReturn(0); 3690 } 3691 3692 PetscErrorCode MatRetrieveValues_SeqAIJ(Mat mat) 3693 { 3694 Mat_SeqAIJ *aij = (Mat_SeqAIJ*)mat->data; 3695 PetscErrorCode ierr; 3696 PetscInt nz = aij->i[mat->rmap->n]; 3697 3698 PetscFunctionBegin; 3699 if (!aij->nonew) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ORDER,"Must call MatSetOption(A,MAT_NEW_NONZERO_LOCATIONS,PETSC_FALSE);first"); 3700 if (!aij->saved_values) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ORDER,"Must call MatStoreValues(A);first"); 3701 /* copy values over */ 3702 ierr = PetscArraycpy(aij->a,aij->saved_values,nz);CHKERRQ(ierr); 3703 PetscFunctionReturn(0); 3704 } 3705 3706 /*@ 3707 MatRetrieveValues - Retrieves the copy of the matrix values; this allows, for 3708 example, reuse of the linear part of a Jacobian, while recomputing the 3709 nonlinear portion. 3710 3711 Collect on Mat 3712 3713 Input Parameters: 3714 . mat - the matrix (currently only AIJ matrices support this option) 3715 3716 Level: advanced 3717 3718 .seealso: MatStoreValues() 3719 3720 @*/ 3721 PetscErrorCode MatRetrieveValues(Mat mat) 3722 { 3723 PetscErrorCode ierr; 3724 3725 PetscFunctionBegin; 3726 PetscValidHeaderSpecific(mat,MAT_CLASSID,1); 3727 if (!mat->assembled) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Not for unassembled matrix"); 3728 if (mat->factortype) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Not for factored matrix"); 3729 ierr = PetscUseMethod(mat,"MatRetrieveValues_C",(Mat),(mat));CHKERRQ(ierr); 3730 PetscFunctionReturn(0); 3731 } 3732 3733 3734 /* --------------------------------------------------------------------------------*/ 3735 /*@C 3736 MatCreateSeqAIJ - Creates a sparse matrix in AIJ (compressed row) format 3737 (the default parallel PETSc format). For good matrix assembly performance 3738 the user should preallocate the matrix storage by setting the parameter nz 3739 (or the array nnz). By setting these parameters accurately, performance 3740 during matrix assembly can be increased by more than a factor of 50. 3741 3742 Collective 3743 3744 Input Parameters: 3745 + comm - MPI communicator, set to PETSC_COMM_SELF 3746 . m - number of rows 3747 . n - number of columns 3748 . nz - number of nonzeros per row (same for all rows) 3749 - nnz - array containing the number of nonzeros in the various rows 3750 (possibly different for each row) or NULL 3751 3752 Output Parameter: 3753 . A - the matrix 3754 3755 It is recommended that one use the MatCreate(), MatSetType() and/or MatSetFromOptions(), 3756 MatXXXXSetPreallocation() paradigm instead of this routine directly. 3757 [MatXXXXSetPreallocation() is, for example, MatSeqAIJSetPreallocation] 3758 3759 Notes: 3760 If nnz is given then nz is ignored 3761 3762 The AIJ format (also called the Yale sparse matrix format or 3763 compressed row storage), is fully compatible with standard Fortran 77 3764 storage. That is, the stored row and column indices can begin at 3765 either one (as in Fortran) or zero. See the users' manual for details. 3766 3767 Specify the preallocated storage with either nz or nnz (not both). 3768 Set nz=PETSC_DEFAULT and nnz=NULL for PETSc to control dynamic memory 3769 allocation. For large problems you MUST preallocate memory or you 3770 will get TERRIBLE performance, see the users' manual chapter on matrices. 3771 3772 By default, this format uses inodes (identical nodes) when possible, to 3773 improve numerical efficiency of matrix-vector products and solves. We 3774 search for consecutive rows with the same nonzero structure, thereby 3775 reusing matrix information to achieve increased efficiency. 3776 3777 Options Database Keys: 3778 + -mat_no_inode - Do not use inodes 3779 - -mat_inode_limit <limit> - Sets inode limit (max limit=5) 3780 3781 Level: intermediate 3782 3783 .seealso: MatCreate(), MatCreateAIJ(), MatSetValues(), MatSeqAIJSetColumnIndices(), MatCreateSeqAIJWithArrays() 3784 3785 @*/ 3786 PetscErrorCode MatCreateSeqAIJ(MPI_Comm comm,PetscInt m,PetscInt n,PetscInt nz,const PetscInt nnz[],Mat *A) 3787 { 3788 PetscErrorCode ierr; 3789 3790 PetscFunctionBegin; 3791 ierr = MatCreate(comm,A);CHKERRQ(ierr); 3792 ierr = MatSetSizes(*A,m,n,m,n);CHKERRQ(ierr); 3793 ierr = MatSetType(*A,MATSEQAIJ);CHKERRQ(ierr); 3794 ierr = MatSeqAIJSetPreallocation_SeqAIJ(*A,nz,nnz);CHKERRQ(ierr); 3795 PetscFunctionReturn(0); 3796 } 3797 3798 /*@C 3799 MatSeqAIJSetPreallocation - For good matrix assembly performance 3800 the user should preallocate the matrix storage by setting the parameter nz 3801 (or the array nnz). By setting these parameters accurately, performance 3802 during matrix assembly can be increased by more than a factor of 50. 3803 3804 Collective 3805 3806 Input Parameters: 3807 + B - The matrix 3808 . nz - number of nonzeros per row (same for all rows) 3809 - nnz - array containing the number of nonzeros in the various rows 3810 (possibly different for each row) or NULL 3811 3812 Notes: 3813 If nnz is given then nz is ignored 3814 3815 The AIJ format (also called the Yale sparse matrix format or 3816 compressed row storage), is fully compatible with standard Fortran 77 3817 storage. That is, the stored row and column indices can begin at 3818 either one (as in Fortran) or zero. See the users' manual for details. 3819 3820 Specify the preallocated storage with either nz or nnz (not both). 3821 Set nz=PETSC_DEFAULT and nnz=NULL for PETSc to control dynamic memory 3822 allocation. For large problems you MUST preallocate memory or you 3823 will get TERRIBLE performance, see the users' manual chapter on matrices. 3824 3825 You can call MatGetInfo() to get information on how effective the preallocation was; 3826 for example the fields mallocs,nz_allocated,nz_used,nz_unneeded; 3827 You can also run with the option -info and look for messages with the string 3828 malloc in them to see if additional memory allocation was needed. 3829 3830 Developers: Use nz of MAT_SKIP_ALLOCATION to not allocate any space for the matrix 3831 entries or columns indices 3832 3833 By default, this format uses inodes (identical nodes) when possible, to 3834 improve numerical efficiency of matrix-vector products and solves. We 3835 search for consecutive rows with the same nonzero structure, thereby 3836 reusing matrix information to achieve increased efficiency. 3837 3838 Options Database Keys: 3839 + -mat_no_inode - Do not use inodes 3840 - -mat_inode_limit <limit> - Sets inode limit (max limit=5) 3841 3842 Level: intermediate 3843 3844 .seealso: MatCreate(), MatCreateAIJ(), MatSetValues(), MatSeqAIJSetColumnIndices(), MatCreateSeqAIJWithArrays(), MatGetInfo() 3845 3846 @*/ 3847 PetscErrorCode MatSeqAIJSetPreallocation(Mat B,PetscInt nz,const PetscInt nnz[]) 3848 { 3849 PetscErrorCode ierr; 3850 3851 PetscFunctionBegin; 3852 PetscValidHeaderSpecific(B,MAT_CLASSID,1); 3853 PetscValidType(B,1); 3854 ierr = PetscTryMethod(B,"MatSeqAIJSetPreallocation_C",(Mat,PetscInt,const PetscInt[]),(B,nz,nnz));CHKERRQ(ierr); 3855 PetscFunctionReturn(0); 3856 } 3857 3858 PetscErrorCode MatSeqAIJSetPreallocation_SeqAIJ(Mat B,PetscInt nz,const PetscInt *nnz) 3859 { 3860 Mat_SeqAIJ *b; 3861 PetscBool skipallocation = PETSC_FALSE,realalloc = PETSC_FALSE; 3862 PetscErrorCode ierr; 3863 PetscInt i; 3864 3865 PetscFunctionBegin; 3866 if (nz >= 0 || nnz) realalloc = PETSC_TRUE; 3867 if (nz == MAT_SKIP_ALLOCATION) { 3868 skipallocation = PETSC_TRUE; 3869 nz = 0; 3870 } 3871 ierr = PetscLayoutSetUp(B->rmap);CHKERRQ(ierr); 3872 ierr = PetscLayoutSetUp(B->cmap);CHKERRQ(ierr); 3873 3874 if (nz == PETSC_DEFAULT || nz == PETSC_DECIDE) nz = 5; 3875 if (nz < 0) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"nz cannot be less than 0: value %D",nz); 3876 #if defined(PETSC_USE_DEBUG) 3877 if (nnz) { 3878 for (i=0; i<B->rmap->n; i++) { 3879 if (nnz[i] < 0) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"nnz cannot be less than 0: local row %D value %D",i,nnz[i]); 3880 if (nnz[i] > B->cmap->n) SETERRQ3(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"nnz cannot be greater than row length: local row %D value %d rowlength %D",i,nnz[i],B->cmap->n); 3881 } 3882 } 3883 #endif 3884 3885 B->preallocated = PETSC_TRUE; 3886 3887 b = (Mat_SeqAIJ*)B->data; 3888 3889 if (!skipallocation) { 3890 if (!b->imax) { 3891 ierr = PetscMalloc1(B->rmap->n,&b->imax);CHKERRQ(ierr); 3892 ierr = PetscLogObjectMemory((PetscObject)B,B->rmap->n*sizeof(PetscInt));CHKERRQ(ierr); 3893 } 3894 if (!b->ilen) { 3895 /* b->ilen will count nonzeros in each row so far. */ 3896 ierr = PetscCalloc1(B->rmap->n,&b->ilen);CHKERRQ(ierr); 3897 ierr = PetscLogObjectMemory((PetscObject)B,B->rmap->n*sizeof(PetscInt));CHKERRQ(ierr); 3898 } else { 3899 ierr = PetscMemzero(b->ilen,B->rmap->n*sizeof(PetscInt));CHKERRQ(ierr); 3900 } 3901 if (!b->ipre) { 3902 ierr = PetscMalloc1(B->rmap->n,&b->ipre);CHKERRQ(ierr); 3903 ierr = PetscLogObjectMemory((PetscObject)B,B->rmap->n*sizeof(PetscInt));CHKERRQ(ierr); 3904 } 3905 if (!nnz) { 3906 if (nz == PETSC_DEFAULT || nz == PETSC_DECIDE) nz = 10; 3907 else if (nz < 0) nz = 1; 3908 nz = PetscMin(nz,B->cmap->n); 3909 for (i=0; i<B->rmap->n; i++) b->imax[i] = nz; 3910 nz = nz*B->rmap->n; 3911 } else { 3912 PetscInt64 nz64 = 0; 3913 for (i=0; i<B->rmap->n; i++) {b->imax[i] = nnz[i]; nz64 += nnz[i];} 3914 ierr = PetscIntCast(nz64,&nz);CHKERRQ(ierr); 3915 } 3916 3917 /* allocate the matrix space */ 3918 /* FIXME: should B's old memory be unlogged? */ 3919 ierr = MatSeqXAIJFreeAIJ(B,&b->a,&b->j,&b->i);CHKERRQ(ierr); 3920 if (B->structure_only) { 3921 ierr = PetscMalloc1(nz,&b->j);CHKERRQ(ierr); 3922 ierr = PetscMalloc1(B->rmap->n+1,&b->i);CHKERRQ(ierr); 3923 ierr = PetscLogObjectMemory((PetscObject)B,(B->rmap->n+1)*sizeof(PetscInt)+nz*sizeof(PetscInt));CHKERRQ(ierr); 3924 } else { 3925 ierr = PetscMalloc3(nz,&b->a,nz,&b->j,B->rmap->n+1,&b->i);CHKERRQ(ierr); 3926 ierr = PetscLogObjectMemory((PetscObject)B,(B->rmap->n+1)*sizeof(PetscInt)+nz*(sizeof(PetscScalar)+sizeof(PetscInt)));CHKERRQ(ierr); 3927 } 3928 b->i[0] = 0; 3929 for (i=1; i<B->rmap->n+1; i++) { 3930 b->i[i] = b->i[i-1] + b->imax[i-1]; 3931 } 3932 if (B->structure_only) { 3933 b->singlemalloc = PETSC_FALSE; 3934 b->free_a = PETSC_FALSE; 3935 } else { 3936 b->singlemalloc = PETSC_TRUE; 3937 b->free_a = PETSC_TRUE; 3938 } 3939 b->free_ij = PETSC_TRUE; 3940 } else { 3941 b->free_a = PETSC_FALSE; 3942 b->free_ij = PETSC_FALSE; 3943 } 3944 3945 if (b->ipre && nnz != b->ipre && b->imax) { 3946 /* reserve user-requested sparsity */ 3947 ierr = PetscArraycpy(b->ipre,b->imax,B->rmap->n);CHKERRQ(ierr); 3948 } 3949 3950 3951 b->nz = 0; 3952 b->maxnz = nz; 3953 B->info.nz_unneeded = (double)b->maxnz; 3954 if (realalloc) { 3955 ierr = MatSetOption(B,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_TRUE);CHKERRQ(ierr); 3956 } 3957 B->was_assembled = PETSC_FALSE; 3958 B->assembled = PETSC_FALSE; 3959 PetscFunctionReturn(0); 3960 } 3961 3962 3963 PetscErrorCode MatResetPreallocation_SeqAIJ(Mat A) 3964 { 3965 Mat_SeqAIJ *a; 3966 PetscInt i; 3967 PetscErrorCode ierr; 3968 3969 PetscFunctionBegin; 3970 PetscValidHeaderSpecific(A,MAT_CLASSID,1); 3971 3972 /* Check local size. If zero, then return */ 3973 if (!A->rmap->n) PetscFunctionReturn(0); 3974 3975 a = (Mat_SeqAIJ*)A->data; 3976 /* if no saved info, we error out */ 3977 if (!a->ipre) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_NULL,"No saved preallocation info \n"); 3978 3979 if (!a->i || !a->j || !a->a || !a->imax || !a->ilen) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_NULL,"Memory info is incomplete, and can not reset preallocation \n"); 3980 3981 ierr = PetscArraycpy(a->imax,a->ipre,A->rmap->n);CHKERRQ(ierr); 3982 ierr = PetscArrayzero(a->ilen,A->rmap->n);CHKERRQ(ierr); 3983 a->i[0] = 0; 3984 for (i=1; i<A->rmap->n+1; i++) { 3985 a->i[i] = a->i[i-1] + a->imax[i-1]; 3986 } 3987 A->preallocated = PETSC_TRUE; 3988 a->nz = 0; 3989 a->maxnz = a->i[A->rmap->n]; 3990 A->info.nz_unneeded = (double)a->maxnz; 3991 A->was_assembled = PETSC_FALSE; 3992 A->assembled = PETSC_FALSE; 3993 PetscFunctionReturn(0); 3994 } 3995 3996 /*@ 3997 MatSeqAIJSetPreallocationCSR - Allocates memory for a sparse sequential matrix in AIJ format. 3998 3999 Input Parameters: 4000 + B - the matrix 4001 . i - the indices into j for the start of each row (starts with zero) 4002 . j - the column indices for each row (starts with zero) these must be sorted for each row 4003 - v - optional values in the matrix 4004 4005 Level: developer 4006 4007 The i,j,v values are COPIED with this routine; to avoid the copy use MatCreateSeqAIJWithArrays() 4008 4009 .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatSeqAIJSetPreallocation(), MatCreateSeqAIJ(), MATSEQAIJ 4010 @*/ 4011 PetscErrorCode MatSeqAIJSetPreallocationCSR(Mat B,const PetscInt i[],const PetscInt j[],const PetscScalar v[]) 4012 { 4013 PetscErrorCode ierr; 4014 4015 PetscFunctionBegin; 4016 PetscValidHeaderSpecific(B,MAT_CLASSID,1); 4017 PetscValidType(B,1); 4018 ierr = PetscTryMethod(B,"MatSeqAIJSetPreallocationCSR_C",(Mat,const PetscInt[],const PetscInt[],const PetscScalar[]),(B,i,j,v));CHKERRQ(ierr); 4019 PetscFunctionReturn(0); 4020 } 4021 4022 PetscErrorCode MatSeqAIJSetPreallocationCSR_SeqAIJ(Mat B,const PetscInt Ii[],const PetscInt J[],const PetscScalar v[]) 4023 { 4024 PetscInt i; 4025 PetscInt m,n; 4026 PetscInt nz; 4027 PetscInt *nnz, nz_max = 0; 4028 PetscErrorCode ierr; 4029 4030 PetscFunctionBegin; 4031 if (Ii[0]) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE, "Ii[0] must be 0 it is %D", Ii[0]); 4032 4033 ierr = PetscLayoutSetUp(B->rmap);CHKERRQ(ierr); 4034 ierr = PetscLayoutSetUp(B->cmap);CHKERRQ(ierr); 4035 4036 ierr = MatGetSize(B, &m, &n);CHKERRQ(ierr); 4037 ierr = PetscMalloc1(m+1, &nnz);CHKERRQ(ierr); 4038 for (i = 0; i < m; i++) { 4039 nz = Ii[i+1]- Ii[i]; 4040 nz_max = PetscMax(nz_max, nz); 4041 if (nz < 0) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE, "Local row %D has a negative number of columns %D", i, nnz); 4042 nnz[i] = nz; 4043 } 4044 ierr = MatSeqAIJSetPreallocation(B, 0, nnz);CHKERRQ(ierr); 4045 ierr = PetscFree(nnz);CHKERRQ(ierr); 4046 4047 for (i = 0; i < m; i++) { 4048 ierr = MatSetValues_SeqAIJ(B, 1, &i, Ii[i+1] - Ii[i], J+Ii[i], v ? v + Ii[i] : NULL, INSERT_VALUES);CHKERRQ(ierr); 4049 } 4050 4051 ierr = MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4052 ierr = MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4053 4054 ierr = MatSetOption(B,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_TRUE);CHKERRQ(ierr); 4055 PetscFunctionReturn(0); 4056 } 4057 4058 #include <../src/mat/impls/dense/seq/dense.h> 4059 #include <petsc/private/kernels/petscaxpy.h> 4060 4061 /* 4062 Computes (B'*A')' since computing B*A directly is untenable 4063 4064 n p p 4065 ( ) ( ) ( ) 4066 m ( A ) * n ( B ) = m ( C ) 4067 ( ) ( ) ( ) 4068 4069 */ 4070 PetscErrorCode MatMatMultNumeric_SeqDense_SeqAIJ(Mat A,Mat B,Mat C) 4071 { 4072 PetscErrorCode ierr; 4073 Mat_SeqDense *sub_a = (Mat_SeqDense*)A->data; 4074 Mat_SeqAIJ *sub_b = (Mat_SeqAIJ*)B->data; 4075 Mat_SeqDense *sub_c = (Mat_SeqDense*)C->data; 4076 PetscInt i,n,m,q,p; 4077 const PetscInt *ii,*idx; 4078 const PetscScalar *b,*a,*a_q; 4079 PetscScalar *c,*c_q; 4080 4081 PetscFunctionBegin; 4082 m = A->rmap->n; 4083 n = A->cmap->n; 4084 p = B->cmap->n; 4085 a = sub_a->v; 4086 b = sub_b->a; 4087 c = sub_c->v; 4088 ierr = PetscArrayzero(c,m*p);CHKERRQ(ierr); 4089 4090 ii = sub_b->i; 4091 idx = sub_b->j; 4092 for (i=0; i<n; i++) { 4093 q = ii[i+1] - ii[i]; 4094 while (q-->0) { 4095 c_q = c + m*(*idx); 4096 a_q = a + m*i; 4097 PetscKernelAXPY(c_q,*b,a_q,m); 4098 idx++; 4099 b++; 4100 } 4101 } 4102 PetscFunctionReturn(0); 4103 } 4104 4105 PetscErrorCode MatMatMultSymbolic_SeqDense_SeqAIJ(Mat A,Mat B,PetscReal fill,Mat *C) 4106 { 4107 PetscErrorCode ierr; 4108 PetscInt m=A->rmap->n,n=B->cmap->n; 4109 Mat Cmat; 4110 4111 PetscFunctionBegin; 4112 if (A->cmap->n != B->rmap->n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"A->cmap->n %D != B->rmap->n %D\n",A->cmap->n,B->rmap->n); 4113 ierr = MatCreate(PetscObjectComm((PetscObject)A),&Cmat);CHKERRQ(ierr); 4114 ierr = MatSetSizes(Cmat,m,n,m,n);CHKERRQ(ierr); 4115 ierr = MatSetBlockSizesFromMats(Cmat,A,B);CHKERRQ(ierr); 4116 ierr = MatSetType(Cmat,MATSEQDENSE);CHKERRQ(ierr); 4117 ierr = MatSeqDenseSetPreallocation(Cmat,NULL);CHKERRQ(ierr); 4118 4119 Cmat->ops->matmultnumeric = MatMatMultNumeric_SeqDense_SeqAIJ; 4120 4121 *C = Cmat; 4122 PetscFunctionReturn(0); 4123 } 4124 4125 /* ----------------------------------------------------------------*/ 4126 PETSC_INTERN PetscErrorCode MatMatMult_SeqDense_SeqAIJ(Mat A,Mat B,MatReuse scall,PetscReal fill,Mat *C) 4127 { 4128 PetscErrorCode ierr; 4129 4130 PetscFunctionBegin; 4131 if (scall == MAT_INITIAL_MATRIX) { 4132 ierr = PetscLogEventBegin(MAT_MatMultSymbolic,A,B,0,0);CHKERRQ(ierr); 4133 ierr = MatMatMultSymbolic_SeqDense_SeqAIJ(A,B,fill,C);CHKERRQ(ierr); 4134 ierr = PetscLogEventEnd(MAT_MatMultSymbolic,A,B,0,0);CHKERRQ(ierr); 4135 } 4136 ierr = PetscLogEventBegin(MAT_MatMultNumeric,A,B,0,0);CHKERRQ(ierr); 4137 ierr = MatMatMultNumeric_SeqDense_SeqAIJ(A,B,*C);CHKERRQ(ierr); 4138 ierr = PetscLogEventEnd(MAT_MatMultNumeric,A,B,0,0);CHKERRQ(ierr); 4139 PetscFunctionReturn(0); 4140 } 4141 4142 4143 /*MC 4144 MATSEQAIJ - MATSEQAIJ = "seqaij" - A matrix type to be used for sequential sparse matrices, 4145 based on compressed sparse row format. 4146 4147 Options Database Keys: 4148 . -mat_type seqaij - sets the matrix type to "seqaij" during a call to MatSetFromOptions() 4149 4150 Level: beginner 4151 4152 Notes: 4153 MatSetValues() may be called for this matrix type with a NULL argument for the numerical values, 4154 in this case the values associated with the rows and columns one passes in are set to zero 4155 in the matrix 4156 4157 MatSetOptions(,MAT_STRUCTURE_ONLY,PETSC_TRUE) may be called for this matrix type. In this no 4158 space is allocated for the nonzero entries and any entries passed with MatSetValues() are ignored 4159 4160 Developer Notes: 4161 It would be nice if all matrix formats supported passing NULL in for the numerical values 4162 4163 .seealso: MatCreateSeqAIJ(), MatSetFromOptions(), MatSetType(), MatCreate(), MatType 4164 M*/ 4165 4166 /*MC 4167 MATAIJ - MATAIJ = "aij" - A matrix type to be used for sparse matrices. 4168 4169 This matrix type is identical to MATSEQAIJ when constructed with a single process communicator, 4170 and MATMPIAIJ otherwise. As a result, for single process communicators, 4171 MatSeqAIJSetPreallocation is supported, and similarly MatMPIAIJSetPreallocation() is supported 4172 for communicators controlling multiple processes. It is recommended that you call both of 4173 the above preallocation routines for simplicity. 4174 4175 Options Database Keys: 4176 . -mat_type aij - sets the matrix type to "aij" during a call to MatSetFromOptions() 4177 4178 Developer Notes: 4179 Subclasses include MATAIJCUSPARSE, MATAIJPERM, MATAIJSELL, MATAIJMKL, MATAIJCRL, and also automatically switches over to use inodes when 4180 enough exist. 4181 4182 Level: beginner 4183 4184 .seealso: MatCreateAIJ(), MatCreateSeqAIJ(), MATSEQAIJ,MATMPIAIJ 4185 M*/ 4186 4187 /*MC 4188 MATAIJCRL - MATAIJCRL = "aijcrl" - A matrix type to be used for sparse matrices. 4189 4190 This matrix type is identical to MATSEQAIJCRL when constructed with a single process communicator, 4191 and MATMPIAIJCRL otherwise. As a result, for single process communicators, 4192 MatSeqAIJSetPreallocation() is supported, and similarly MatMPIAIJSetPreallocation() is supported 4193 for communicators controlling multiple processes. It is recommended that you call both of 4194 the above preallocation routines for simplicity. 4195 4196 Options Database Keys: 4197 . -mat_type aijcrl - sets the matrix type to "aijcrl" during a call to MatSetFromOptions() 4198 4199 Level: beginner 4200 4201 .seealso: MatCreateMPIAIJCRL,MATSEQAIJCRL,MATMPIAIJCRL, MATSEQAIJCRL, MATMPIAIJCRL 4202 M*/ 4203 4204 PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_SeqAIJCRL(Mat,MatType,MatReuse,Mat*); 4205 #if defined(PETSC_HAVE_ELEMENTAL) 4206 PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_Elemental(Mat,MatType,MatReuse,Mat*); 4207 #endif 4208 #if defined(PETSC_HAVE_HYPRE) 4209 PETSC_INTERN PetscErrorCode MatConvert_AIJ_HYPRE(Mat A,MatType,MatReuse,Mat*); 4210 PETSC_INTERN PetscErrorCode MatMatMatMult_Transpose_AIJ_AIJ(Mat,Mat,Mat,MatReuse,PetscReal,Mat*); 4211 #endif 4212 PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_SeqDense(Mat,MatType,MatReuse,Mat*); 4213 4214 PETSC_EXTERN PetscErrorCode MatConvert_SeqAIJ_SeqSELL(Mat,MatType,MatReuse,Mat*); 4215 PETSC_INTERN PetscErrorCode MatConvert_XAIJ_IS(Mat,MatType,MatReuse,Mat*); 4216 PETSC_INTERN PetscErrorCode MatPtAP_IS_XAIJ(Mat,Mat,MatReuse,PetscReal,Mat*); 4217 4218 /*@C 4219 MatSeqAIJGetArray - gives read/write access to the array where the data for a MATSEQAIJ matrix is stored 4220 4221 Not Collective 4222 4223 Input Parameter: 4224 . mat - a MATSEQAIJ matrix 4225 4226 Output Parameter: 4227 . array - pointer to the data 4228 4229 Level: intermediate 4230 4231 .seealso: MatSeqAIJRestoreArray(), MatSeqAIJGetArrayF90() 4232 @*/ 4233 PetscErrorCode MatSeqAIJGetArray(Mat A,PetscScalar **array) 4234 { 4235 PetscErrorCode ierr; 4236 4237 PetscFunctionBegin; 4238 ierr = PetscUseMethod(A,"MatSeqAIJGetArray_C",(Mat,PetscScalar**),(A,array));CHKERRQ(ierr); 4239 PetscFunctionReturn(0); 4240 } 4241 4242 /*@C 4243 MatSeqAIJGetArrayRead - gives read-only access to the array where the data for a MATSEQAIJ matrix is stored 4244 4245 Not Collective 4246 4247 Input Parameter: 4248 . mat - a MATSEQAIJ matrix 4249 4250 Output Parameter: 4251 . array - pointer to the data 4252 4253 Level: intermediate 4254 4255 .seealso: MatSeqAIJGetArray(), MatSeqAIJRestoreArrayRead() 4256 @*/ 4257 PetscErrorCode MatSeqAIJGetArrayRead(Mat A,const PetscScalar **array) 4258 { 4259 #if defined(PETSC_HAVE_CUDA) || defined(PETSC_HAVE_VIENNACL) 4260 PetscOffloadMask oval; 4261 #endif 4262 PetscErrorCode ierr; 4263 4264 PetscFunctionBegin; 4265 #if defined(PETSC_HAVE_CUDA) || defined(PETSC_HAVE_VIENNACL) 4266 oval = A->offloadmask; 4267 #endif 4268 ierr = MatSeqAIJGetArray(A,(PetscScalar**)array);CHKERRQ(ierr); 4269 #if defined(PETSC_HAVE_CUDA) || defined(PETSC_HAVE_VIENNACL) 4270 if (oval == PETSC_OFFLOAD_GPU || oval == PETSC_OFFLOAD_BOTH) A->offloadmask = PETSC_OFFLOAD_BOTH; 4271 #endif 4272 PetscFunctionReturn(0); 4273 } 4274 4275 /*@C 4276 MatSeqAIJRestoreArrayRead - restore the read-only access array obtained from MatSeqAIJGetArrayRead 4277 4278 Not Collective 4279 4280 Input Parameter: 4281 . mat - a MATSEQAIJ matrix 4282 4283 Output Parameter: 4284 . array - pointer to the data 4285 4286 Level: intermediate 4287 4288 .seealso: MatSeqAIJGetArray(), MatSeqAIJGetArrayRead() 4289 @*/ 4290 PetscErrorCode MatSeqAIJRestoreArrayRead(Mat A,const PetscScalar **array) 4291 { 4292 #if defined(PETSC_HAVE_CUDA) || defined(PETSC_HAVE_VIENNACL) 4293 PetscOffloadMask oval; 4294 #endif 4295 PetscErrorCode ierr; 4296 4297 PetscFunctionBegin; 4298 #if defined(PETSC_HAVE_CUDA) || defined(PETSC_HAVE_VIENNACL) 4299 oval = A->offloadmask; 4300 #endif 4301 ierr = MatSeqAIJRestoreArray(A,(PetscScalar**)array);CHKERRQ(ierr); 4302 #if defined(PETSC_HAVE_CUDA) || defined(PETSC_HAVE_VIENNACL) 4303 A->offloadmask = oval; 4304 #endif 4305 PetscFunctionReturn(0); 4306 } 4307 4308 /*@C 4309 MatSeqAIJGetMaxRowNonzeros - returns the maximum number of nonzeros in any row 4310 4311 Not Collective 4312 4313 Input Parameter: 4314 . mat - a MATSEQAIJ matrix 4315 4316 Output Parameter: 4317 . nz - the maximum number of nonzeros in any row 4318 4319 Level: intermediate 4320 4321 .seealso: MatSeqAIJRestoreArray(), MatSeqAIJGetArrayF90() 4322 @*/ 4323 PetscErrorCode MatSeqAIJGetMaxRowNonzeros(Mat A,PetscInt *nz) 4324 { 4325 Mat_SeqAIJ *aij = (Mat_SeqAIJ*)A->data; 4326 4327 PetscFunctionBegin; 4328 *nz = aij->rmax; 4329 PetscFunctionReturn(0); 4330 } 4331 4332 /*@C 4333 MatSeqAIJRestoreArray - returns access to the array where the data for a MATSEQAIJ matrix is stored obtained by MatSeqAIJGetArray() 4334 4335 Not Collective 4336 4337 Input Parameters: 4338 + mat - a MATSEQAIJ matrix 4339 - array - pointer to the data 4340 4341 Level: intermediate 4342 4343 .seealso: MatSeqAIJGetArray(), MatSeqAIJRestoreArrayF90() 4344 @*/ 4345 PetscErrorCode MatSeqAIJRestoreArray(Mat A,PetscScalar **array) 4346 { 4347 PetscErrorCode ierr; 4348 4349 PetscFunctionBegin; 4350 ierr = PetscUseMethod(A,"MatSeqAIJRestoreArray_C",(Mat,PetscScalar**),(A,array));CHKERRQ(ierr); 4351 PetscFunctionReturn(0); 4352 } 4353 4354 #if defined(PETSC_HAVE_CUDA) 4355 PETSC_EXTERN PetscErrorCode MatConvert_SeqAIJ_SeqAIJCUSPARSE(Mat); 4356 #endif 4357 4358 PETSC_EXTERN PetscErrorCode MatCreate_SeqAIJ(Mat B) 4359 { 4360 Mat_SeqAIJ *b; 4361 PetscErrorCode ierr; 4362 PetscMPIInt size; 4363 4364 PetscFunctionBegin; 4365 ierr = MPI_Comm_size(PetscObjectComm((PetscObject)B),&size);CHKERRQ(ierr); 4366 if (size > 1) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Comm must be of size 1"); 4367 4368 ierr = PetscNewLog(B,&b);CHKERRQ(ierr); 4369 4370 B->data = (void*)b; 4371 4372 ierr = PetscMemcpy(B->ops,&MatOps_Values,sizeof(struct _MatOps));CHKERRQ(ierr); 4373 if (B->sortedfull) B->ops->setvalues = MatSetValues_SeqAIJ_SortedFull; 4374 4375 b->row = 0; 4376 b->col = 0; 4377 b->icol = 0; 4378 b->reallocs = 0; 4379 b->ignorezeroentries = PETSC_FALSE; 4380 b->roworiented = PETSC_TRUE; 4381 b->nonew = 0; 4382 b->diag = 0; 4383 b->solve_work = 0; 4384 B->spptr = 0; 4385 b->saved_values = 0; 4386 b->idiag = 0; 4387 b->mdiag = 0; 4388 b->ssor_work = 0; 4389 b->omega = 1.0; 4390 b->fshift = 0.0; 4391 b->idiagvalid = PETSC_FALSE; 4392 b->ibdiagvalid = PETSC_FALSE; 4393 b->keepnonzeropattern = PETSC_FALSE; 4394 4395 ierr = PetscObjectChangeTypeName((PetscObject)B,MATSEQAIJ);CHKERRQ(ierr); 4396 ierr = PetscObjectComposeFunction((PetscObject)B,"MatSeqAIJGetArray_C",MatSeqAIJGetArray_SeqAIJ);CHKERRQ(ierr); 4397 ierr = PetscObjectComposeFunction((PetscObject)B,"MatSeqAIJRestoreArray_C",MatSeqAIJRestoreArray_SeqAIJ);CHKERRQ(ierr); 4398 4399 #if defined(PETSC_HAVE_MATLAB_ENGINE) 4400 ierr = PetscObjectComposeFunction((PetscObject)B,"PetscMatlabEnginePut_C",MatlabEnginePut_SeqAIJ);CHKERRQ(ierr); 4401 ierr = PetscObjectComposeFunction((PetscObject)B,"PetscMatlabEngineGet_C",MatlabEngineGet_SeqAIJ);CHKERRQ(ierr); 4402 #endif 4403 4404 ierr = PetscObjectComposeFunction((PetscObject)B,"MatSeqAIJSetColumnIndices_C",MatSeqAIJSetColumnIndices_SeqAIJ);CHKERRQ(ierr); 4405 ierr = PetscObjectComposeFunction((PetscObject)B,"MatStoreValues_C",MatStoreValues_SeqAIJ);CHKERRQ(ierr); 4406 ierr = PetscObjectComposeFunction((PetscObject)B,"MatRetrieveValues_C",MatRetrieveValues_SeqAIJ);CHKERRQ(ierr); 4407 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_seqaij_seqsbaij_C",MatConvert_SeqAIJ_SeqSBAIJ);CHKERRQ(ierr); 4408 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_seqaij_seqbaij_C",MatConvert_SeqAIJ_SeqBAIJ);CHKERRQ(ierr); 4409 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_seqaij_seqaijperm_C",MatConvert_SeqAIJ_SeqAIJPERM);CHKERRQ(ierr); 4410 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_seqaij_seqaijsell_C",MatConvert_SeqAIJ_SeqAIJSELL);CHKERRQ(ierr); 4411 #if defined(PETSC_HAVE_MKL_SPARSE) 4412 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_seqaij_seqaijmkl_C",MatConvert_SeqAIJ_SeqAIJMKL);CHKERRQ(ierr); 4413 #endif 4414 #if defined(PETSC_HAVE_CUDA) 4415 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_seqaij_seqaijcusparse_C",MatConvert_SeqAIJ_SeqAIJCUSPARSE);CHKERRQ(ierr); 4416 #endif 4417 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_seqaij_seqaijcrl_C",MatConvert_SeqAIJ_SeqAIJCRL);CHKERRQ(ierr); 4418 #if defined(PETSC_HAVE_ELEMENTAL) 4419 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_seqaij_elemental_C",MatConvert_SeqAIJ_Elemental);CHKERRQ(ierr); 4420 #endif 4421 #if defined(PETSC_HAVE_HYPRE) 4422 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_seqaij_hypre_C",MatConvert_AIJ_HYPRE);CHKERRQ(ierr); 4423 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMatMatMult_transpose_seqaij_seqaij_C",MatMatMatMult_Transpose_AIJ_AIJ);CHKERRQ(ierr); 4424 #endif 4425 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_seqaij_seqdense_C",MatConvert_SeqAIJ_SeqDense);CHKERRQ(ierr); 4426 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_seqaij_seqsell_C",MatConvert_SeqAIJ_SeqSELL);CHKERRQ(ierr); 4427 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_seqaij_is_C",MatConvert_XAIJ_IS);CHKERRQ(ierr); 4428 ierr = PetscObjectComposeFunction((PetscObject)B,"MatIsTranspose_C",MatIsTranspose_SeqAIJ);CHKERRQ(ierr); 4429 ierr = PetscObjectComposeFunction((PetscObject)B,"MatIsHermitianTranspose_C",MatIsTranspose_SeqAIJ);CHKERRQ(ierr); 4430 ierr = PetscObjectComposeFunction((PetscObject)B,"MatSeqAIJSetPreallocation_C",MatSeqAIJSetPreallocation_SeqAIJ);CHKERRQ(ierr); 4431 ierr = PetscObjectComposeFunction((PetscObject)B,"MatResetPreallocation_C",MatResetPreallocation_SeqAIJ);CHKERRQ(ierr); 4432 ierr = PetscObjectComposeFunction((PetscObject)B,"MatSeqAIJSetPreallocationCSR_C",MatSeqAIJSetPreallocationCSR_SeqAIJ);CHKERRQ(ierr); 4433 ierr = PetscObjectComposeFunction((PetscObject)B,"MatReorderForNonzeroDiagonal_C",MatReorderForNonzeroDiagonal_SeqAIJ);CHKERRQ(ierr); 4434 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMatMult_seqdense_seqaij_C",MatMatMult_SeqDense_SeqAIJ);CHKERRQ(ierr); 4435 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMatMultSymbolic_seqdense_seqaij_C",MatMatMultSymbolic_SeqDense_SeqAIJ);CHKERRQ(ierr); 4436 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMatMultNumeric_seqdense_seqaij_C",MatMatMultNumeric_SeqDense_SeqAIJ);CHKERRQ(ierr); 4437 ierr = PetscObjectComposeFunction((PetscObject)B,"MatPtAP_is_seqaij_C",MatPtAP_IS_XAIJ);CHKERRQ(ierr); 4438 ierr = MatCreate_SeqAIJ_Inode(B);CHKERRQ(ierr); 4439 ierr = PetscObjectChangeTypeName((PetscObject)B,MATSEQAIJ);CHKERRQ(ierr); 4440 ierr = MatSeqAIJSetTypeFromOptions(B);CHKERRQ(ierr); /* this allows changing the matrix subtype to say MATSEQAIJPERM */ 4441 PetscFunctionReturn(0); 4442 } 4443 4444 /* 4445 Given a matrix generated with MatGetFactor() duplicates all the information in A into B 4446 */ 4447 PetscErrorCode MatDuplicateNoCreate_SeqAIJ(Mat C,Mat A,MatDuplicateOption cpvalues,PetscBool mallocmatspace) 4448 { 4449 Mat_SeqAIJ *c,*a = (Mat_SeqAIJ*)A->data; 4450 PetscErrorCode ierr; 4451 PetscInt m = A->rmap->n,i; 4452 4453 PetscFunctionBegin; 4454 c = (Mat_SeqAIJ*)C->data; 4455 4456 C->factortype = A->factortype; 4457 c->row = 0; 4458 c->col = 0; 4459 c->icol = 0; 4460 c->reallocs = 0; 4461 4462 C->assembled = PETSC_TRUE; 4463 4464 ierr = PetscLayoutReference(A->rmap,&C->rmap);CHKERRQ(ierr); 4465 ierr = PetscLayoutReference(A->cmap,&C->cmap);CHKERRQ(ierr); 4466 4467 ierr = PetscMalloc1(m,&c->imax);CHKERRQ(ierr); 4468 ierr = PetscMemcpy(c->imax,a->imax,m*sizeof(PetscInt));CHKERRQ(ierr); 4469 ierr = PetscMalloc1(m,&c->ilen);CHKERRQ(ierr); 4470 ierr = PetscMemcpy(c->ilen,a->ilen,m*sizeof(PetscInt));CHKERRQ(ierr); 4471 ierr = PetscLogObjectMemory((PetscObject)C, 2*m*sizeof(PetscInt));CHKERRQ(ierr); 4472 4473 /* allocate the matrix space */ 4474 if (mallocmatspace) { 4475 ierr = PetscMalloc3(a->i[m],&c->a,a->i[m],&c->j,m+1,&c->i);CHKERRQ(ierr); 4476 ierr = PetscLogObjectMemory((PetscObject)C, a->i[m]*(sizeof(PetscScalar)+sizeof(PetscInt))+(m+1)*sizeof(PetscInt));CHKERRQ(ierr); 4477 4478 c->singlemalloc = PETSC_TRUE; 4479 4480 ierr = PetscArraycpy(c->i,a->i,m+1);CHKERRQ(ierr); 4481 if (m > 0) { 4482 ierr = PetscArraycpy(c->j,a->j,a->i[m]);CHKERRQ(ierr); 4483 if (cpvalues == MAT_COPY_VALUES) { 4484 ierr = PetscArraycpy(c->a,a->a,a->i[m]);CHKERRQ(ierr); 4485 } else { 4486 ierr = PetscArrayzero(c->a,a->i[m]);CHKERRQ(ierr); 4487 } 4488 } 4489 } 4490 4491 c->ignorezeroentries = a->ignorezeroentries; 4492 c->roworiented = a->roworiented; 4493 c->nonew = a->nonew; 4494 if (a->diag) { 4495 ierr = PetscMalloc1(m+1,&c->diag);CHKERRQ(ierr); 4496 ierr = PetscMemcpy(c->diag,a->diag,m*sizeof(PetscInt));CHKERRQ(ierr); 4497 ierr = PetscLogObjectMemory((PetscObject)C,(m+1)*sizeof(PetscInt));CHKERRQ(ierr); 4498 } else c->diag = NULL; 4499 4500 c->solve_work = 0; 4501 c->saved_values = 0; 4502 c->idiag = 0; 4503 c->ssor_work = 0; 4504 c->keepnonzeropattern = a->keepnonzeropattern; 4505 c->free_a = PETSC_TRUE; 4506 c->free_ij = PETSC_TRUE; 4507 4508 c->rmax = a->rmax; 4509 c->nz = a->nz; 4510 c->maxnz = a->nz; /* Since we allocate exactly the right amount */ 4511 C->preallocated = PETSC_TRUE; 4512 4513 c->compressedrow.use = a->compressedrow.use; 4514 c->compressedrow.nrows = a->compressedrow.nrows; 4515 if (a->compressedrow.use) { 4516 i = a->compressedrow.nrows; 4517 ierr = PetscMalloc2(i+1,&c->compressedrow.i,i,&c->compressedrow.rindex);CHKERRQ(ierr); 4518 ierr = PetscArraycpy(c->compressedrow.i,a->compressedrow.i,i+1);CHKERRQ(ierr); 4519 ierr = PetscArraycpy(c->compressedrow.rindex,a->compressedrow.rindex,i);CHKERRQ(ierr); 4520 } else { 4521 c->compressedrow.use = PETSC_FALSE; 4522 c->compressedrow.i = NULL; 4523 c->compressedrow.rindex = NULL; 4524 } 4525 c->nonzerorowcnt = a->nonzerorowcnt; 4526 C->nonzerostate = A->nonzerostate; 4527 4528 ierr = MatDuplicate_SeqAIJ_Inode(A,cpvalues,&C);CHKERRQ(ierr); 4529 ierr = PetscFunctionListDuplicate(((PetscObject)A)->qlist,&((PetscObject)C)->qlist);CHKERRQ(ierr); 4530 PetscFunctionReturn(0); 4531 } 4532 4533 PetscErrorCode MatDuplicate_SeqAIJ(Mat A,MatDuplicateOption cpvalues,Mat *B) 4534 { 4535 PetscErrorCode ierr; 4536 4537 PetscFunctionBegin; 4538 ierr = MatCreate(PetscObjectComm((PetscObject)A),B);CHKERRQ(ierr); 4539 ierr = MatSetSizes(*B,A->rmap->n,A->cmap->n,A->rmap->n,A->cmap->n);CHKERRQ(ierr); 4540 if (!(A->rmap->n % A->rmap->bs) && !(A->cmap->n % A->cmap->bs)) { 4541 ierr = MatSetBlockSizesFromMats(*B,A,A);CHKERRQ(ierr); 4542 } 4543 ierr = MatSetType(*B,((PetscObject)A)->type_name);CHKERRQ(ierr); 4544 ierr = MatDuplicateNoCreate_SeqAIJ(*B,A,cpvalues,PETSC_TRUE);CHKERRQ(ierr); 4545 PetscFunctionReturn(0); 4546 } 4547 4548 PetscErrorCode MatLoad_SeqAIJ(Mat newMat, PetscViewer viewer) 4549 { 4550 PetscBool isbinary, ishdf5; 4551 PetscErrorCode ierr; 4552 4553 PetscFunctionBegin; 4554 PetscValidHeaderSpecific(newMat,MAT_CLASSID,1); 4555 PetscValidHeaderSpecific(viewer,PETSC_VIEWER_CLASSID,2); 4556 /* force binary viewer to load .info file if it has not yet done so */ 4557 ierr = PetscViewerSetUp(viewer);CHKERRQ(ierr); 4558 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERBINARY,&isbinary);CHKERRQ(ierr); 4559 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERHDF5, &ishdf5);CHKERRQ(ierr); 4560 if (isbinary) { 4561 ierr = MatLoad_SeqAIJ_Binary(newMat,viewer);CHKERRQ(ierr); 4562 } else if (ishdf5) { 4563 #if defined(PETSC_HAVE_HDF5) 4564 ierr = MatLoad_AIJ_HDF5(newMat,viewer);CHKERRQ(ierr); 4565 #else 4566 SETERRQ(PetscObjectComm((PetscObject)newMat),PETSC_ERR_SUP,"HDF5 not supported in this build.\nPlease reconfigure using --download-hdf5"); 4567 #endif 4568 } else { 4569 SETERRQ2(PetscObjectComm((PetscObject)newMat),PETSC_ERR_SUP,"Viewer type %s not yet supported for reading %s matrices",((PetscObject)viewer)->type_name,((PetscObject)newMat)->type_name); 4570 } 4571 PetscFunctionReturn(0); 4572 } 4573 4574 PetscErrorCode MatLoad_SeqAIJ_Binary(Mat newMat, PetscViewer viewer) 4575 { 4576 Mat_SeqAIJ *a; 4577 PetscErrorCode ierr; 4578 PetscInt i,sum,nz,header[4],*rowlengths = 0,M,N,rows,cols; 4579 int fd; 4580 PetscMPIInt size; 4581 MPI_Comm comm; 4582 PetscInt bs = newMat->rmap->bs; 4583 4584 PetscFunctionBegin; 4585 ierr = PetscObjectGetComm((PetscObject)viewer,&comm);CHKERRQ(ierr); 4586 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 4587 if (size > 1) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"view must have one processor"); 4588 4589 ierr = PetscOptionsBegin(comm,NULL,"Options for loading SEQAIJ matrix","Mat");CHKERRQ(ierr); 4590 ierr = PetscOptionsInt("-matload_block_size","Set the blocksize used to store the matrix","MatLoad",bs,&bs,NULL);CHKERRQ(ierr); 4591 ierr = PetscOptionsEnd();CHKERRQ(ierr); 4592 if (bs < 0) bs = 1; 4593 ierr = MatSetBlockSize(newMat,bs);CHKERRQ(ierr); 4594 4595 ierr = PetscViewerBinaryGetDescriptor(viewer,&fd);CHKERRQ(ierr); 4596 ierr = PetscBinaryRead(fd,header,4,NULL,PETSC_INT);CHKERRQ(ierr); 4597 if (header[0] != MAT_FILE_CLASSID) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED,"not matrix object in file"); 4598 M = header[1]; N = header[2]; nz = header[3]; 4599 4600 if (nz < 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED,"Matrix stored in special format on disk,cannot load as SeqAIJ"); 4601 4602 /* read in row lengths */ 4603 ierr = PetscMalloc1(M,&rowlengths);CHKERRQ(ierr); 4604 ierr = PetscBinaryRead(fd,rowlengths,M,NULL,PETSC_INT);CHKERRQ(ierr); 4605 4606 /* check if sum of rowlengths is same as nz */ 4607 for (i=0,sum=0; i< M; i++) sum +=rowlengths[i]; 4608 if (sum != nz) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_FILE_READ,"Inconsistant matrix data in file. no-nonzeros = %dD, sum-row-lengths = %D\n",nz,sum); 4609 4610 /* set global size if not set already*/ 4611 if (newMat->rmap->n < 0 && newMat->rmap->N < 0 && newMat->cmap->n < 0 && newMat->cmap->N < 0) { 4612 ierr = MatSetSizes(newMat,PETSC_DECIDE,PETSC_DECIDE,M,N);CHKERRQ(ierr); 4613 } else { 4614 /* if sizes and type are already set, check if the matrix global sizes are correct */ 4615 ierr = MatGetSize(newMat,&rows,&cols);CHKERRQ(ierr); 4616 if (rows < 0 && cols < 0) { /* user might provide local size instead of global size */ 4617 ierr = MatGetLocalSize(newMat,&rows,&cols);CHKERRQ(ierr); 4618 } 4619 if (M != rows || N != cols) SETERRQ4(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED, "Matrix in file of different length (%D, %D) than the input matrix (%D, %D)",M,N,rows,cols); 4620 } 4621 ierr = MatSeqAIJSetPreallocation_SeqAIJ(newMat,0,rowlengths);CHKERRQ(ierr); 4622 a = (Mat_SeqAIJ*)newMat->data; 4623 4624 ierr = PetscBinaryRead(fd,a->j,nz,NULL,PETSC_INT);CHKERRQ(ierr); 4625 4626 /* read in nonzero values */ 4627 ierr = PetscBinaryRead(fd,a->a,nz,NULL,PETSC_SCALAR);CHKERRQ(ierr); 4628 4629 /* set matrix "i" values */ 4630 a->i[0] = 0; 4631 for (i=1; i<= M; i++) { 4632 a->i[i] = a->i[i-1] + rowlengths[i-1]; 4633 a->ilen[i-1] = rowlengths[i-1]; 4634 } 4635 ierr = PetscFree(rowlengths);CHKERRQ(ierr); 4636 4637 ierr = MatAssemblyBegin(newMat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4638 ierr = MatAssemblyEnd(newMat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4639 PetscFunctionReturn(0); 4640 } 4641 4642 PetscErrorCode MatEqual_SeqAIJ(Mat A,Mat B,PetscBool * flg) 4643 { 4644 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data,*b = (Mat_SeqAIJ*)B->data; 4645 PetscErrorCode ierr; 4646 #if defined(PETSC_USE_COMPLEX) 4647 PetscInt k; 4648 #endif 4649 4650 PetscFunctionBegin; 4651 /* If the matrix dimensions are not equal,or no of nonzeros */ 4652 if ((A->rmap->n != B->rmap->n) || (A->cmap->n != B->cmap->n) ||(a->nz != b->nz)) { 4653 *flg = PETSC_FALSE; 4654 PetscFunctionReturn(0); 4655 } 4656 4657 /* if the a->i are the same */ 4658 ierr = PetscArraycmp(a->i,b->i,A->rmap->n+1,flg);CHKERRQ(ierr); 4659 if (!*flg) PetscFunctionReturn(0); 4660 4661 /* if a->j are the same */ 4662 ierr = PetscArraycmp(a->j,b->j,a->nz,flg);CHKERRQ(ierr); 4663 if (!*flg) PetscFunctionReturn(0); 4664 4665 /* if a->a are the same */ 4666 #if defined(PETSC_USE_COMPLEX) 4667 for (k=0; k<a->nz; k++) { 4668 if (PetscRealPart(a->a[k]) != PetscRealPart(b->a[k]) || PetscImaginaryPart(a->a[k]) != PetscImaginaryPart(b->a[k])) { 4669 *flg = PETSC_FALSE; 4670 PetscFunctionReturn(0); 4671 } 4672 } 4673 #else 4674 ierr = PetscArraycmp(a->a,b->a,a->nz,flg);CHKERRQ(ierr); 4675 #endif 4676 PetscFunctionReturn(0); 4677 } 4678 4679 /*@ 4680 MatCreateSeqAIJWithArrays - Creates an sequential AIJ matrix using matrix elements (in CSR format) 4681 provided by the user. 4682 4683 Collective 4684 4685 Input Parameters: 4686 + comm - must be an MPI communicator of size 1 4687 . m - number of rows 4688 . n - number of columns 4689 . i - row indices; that is i[0] = 0, i[row] = i[row-1] + number of elements in that row of the matrix 4690 . j - column indices 4691 - a - matrix values 4692 4693 Output Parameter: 4694 . mat - the matrix 4695 4696 Level: intermediate 4697 4698 Notes: 4699 The i, j, and a arrays are not copied by this routine, the user must free these arrays 4700 once the matrix is destroyed and not before 4701 4702 You cannot set new nonzero locations into this matrix, that will generate an error. 4703 4704 The i and j indices are 0 based 4705 4706 The format which is used for the sparse matrix input, is equivalent to a 4707 row-major ordering.. i.e for the following matrix, the input data expected is 4708 as shown 4709 4710 $ 1 0 0 4711 $ 2 0 3 4712 $ 4 5 6 4713 $ 4714 $ i = {0,1,3,6} [size = nrow+1 = 3+1] 4715 $ j = {0,0,2,0,1,2} [size = 6]; values must be sorted for each row 4716 $ v = {1,2,3,4,5,6} [size = 6] 4717 4718 4719 .seealso: MatCreate(), MatCreateAIJ(), MatCreateSeqAIJ(), MatCreateMPIAIJWithArrays(), MatMPIAIJSetPreallocationCSR() 4720 4721 @*/ 4722 PetscErrorCode MatCreateSeqAIJWithArrays(MPI_Comm comm,PetscInt m,PetscInt n,PetscInt i[],PetscInt j[],PetscScalar a[],Mat *mat) 4723 { 4724 PetscErrorCode ierr; 4725 PetscInt ii; 4726 Mat_SeqAIJ *aij; 4727 #if defined(PETSC_USE_DEBUG) 4728 PetscInt jj; 4729 #endif 4730 4731 PetscFunctionBegin; 4732 if (m > 0 && i[0]) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"i (row indices) must start with 0"); 4733 ierr = MatCreate(comm,mat);CHKERRQ(ierr); 4734 ierr = MatSetSizes(*mat,m,n,m,n);CHKERRQ(ierr); 4735 /* ierr = MatSetBlockSizes(*mat,,);CHKERRQ(ierr); */ 4736 ierr = MatSetType(*mat,MATSEQAIJ);CHKERRQ(ierr); 4737 ierr = MatSeqAIJSetPreallocation_SeqAIJ(*mat,MAT_SKIP_ALLOCATION,0);CHKERRQ(ierr); 4738 aij = (Mat_SeqAIJ*)(*mat)->data; 4739 ierr = PetscMalloc1(m,&aij->imax);CHKERRQ(ierr); 4740 ierr = PetscMalloc1(m,&aij->ilen);CHKERRQ(ierr); 4741 4742 aij->i = i; 4743 aij->j = j; 4744 aij->a = a; 4745 aij->singlemalloc = PETSC_FALSE; 4746 aij->nonew = -1; /*this indicates that inserting a new value in the matrix that generates a new nonzero is an error*/ 4747 aij->free_a = PETSC_FALSE; 4748 aij->free_ij = PETSC_FALSE; 4749 4750 for (ii=0; ii<m; ii++) { 4751 aij->ilen[ii] = aij->imax[ii] = i[ii+1] - i[ii]; 4752 #if defined(PETSC_USE_DEBUG) 4753 if (i[ii+1] - i[ii] < 0) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Negative row length in i (row indices) row = %D length = %D",ii,i[ii+1] - i[ii]); 4754 for (jj=i[ii]+1; jj<i[ii+1]; jj++) { 4755 if (j[jj] < j[jj-1]) SETERRQ3(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Column entry number %D (actual column %D) in row %D is not sorted",jj-i[ii],j[jj],ii); 4756 if (j[jj] == j[jj-1]) SETERRQ3(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Column entry number %D (actual column %D) in row %D is identical to previous entry",jj-i[ii],j[jj],ii); 4757 } 4758 #endif 4759 } 4760 #if defined(PETSC_USE_DEBUG) 4761 for (ii=0; ii<aij->i[m]; ii++) { 4762 if (j[ii] < 0) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Negative column index at location = %D index = %D",ii,j[ii]); 4763 if (j[ii] > n - 1) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Column index to large at location = %D index = %D",ii,j[ii]); 4764 } 4765 #endif 4766 4767 ierr = MatAssemblyBegin(*mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4768 ierr = MatAssemblyEnd(*mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4769 PetscFunctionReturn(0); 4770 } 4771 /*@C 4772 MatCreateSeqAIJFromTriple - Creates an sequential AIJ matrix using matrix elements (in COO format) 4773 provided by the user. 4774 4775 Collective 4776 4777 Input Parameters: 4778 + comm - must be an MPI communicator of size 1 4779 . m - number of rows 4780 . n - number of columns 4781 . i - row indices 4782 . j - column indices 4783 . a - matrix values 4784 . nz - number of nonzeros 4785 - idx - 0 or 1 based 4786 4787 Output Parameter: 4788 . mat - the matrix 4789 4790 Level: intermediate 4791 4792 Notes: 4793 The i and j indices are 0 based 4794 4795 The format which is used for the sparse matrix input, is equivalent to a 4796 row-major ordering.. i.e for the following matrix, the input data expected is 4797 as shown: 4798 4799 1 0 0 4800 2 0 3 4801 4 5 6 4802 4803 i = {0,1,1,2,2,2} 4804 j = {0,0,2,0,1,2} 4805 v = {1,2,3,4,5,6} 4806 4807 4808 .seealso: MatCreate(), MatCreateAIJ(), MatCreateSeqAIJ(), MatCreateSeqAIJWithArrays(), MatMPIAIJSetPreallocationCSR() 4809 4810 @*/ 4811 PetscErrorCode MatCreateSeqAIJFromTriple(MPI_Comm comm,PetscInt m,PetscInt n,PetscInt i[],PetscInt j[],PetscScalar a[],Mat *mat,PetscInt nz,PetscBool idx) 4812 { 4813 PetscErrorCode ierr; 4814 PetscInt ii, *nnz, one = 1,row,col; 4815 4816 4817 PetscFunctionBegin; 4818 ierr = PetscCalloc1(m,&nnz);CHKERRQ(ierr); 4819 for (ii = 0; ii < nz; ii++) { 4820 nnz[i[ii] - !!idx] += 1; 4821 } 4822 ierr = MatCreate(comm,mat);CHKERRQ(ierr); 4823 ierr = MatSetSizes(*mat,m,n,m,n);CHKERRQ(ierr); 4824 ierr = MatSetType(*mat,MATSEQAIJ);CHKERRQ(ierr); 4825 ierr = MatSeqAIJSetPreallocation_SeqAIJ(*mat,0,nnz);CHKERRQ(ierr); 4826 for (ii = 0; ii < nz; ii++) { 4827 if (idx) { 4828 row = i[ii] - 1; 4829 col = j[ii] - 1; 4830 } else { 4831 row = i[ii]; 4832 col = j[ii]; 4833 } 4834 ierr = MatSetValues(*mat,one,&row,one,&col,&a[ii],ADD_VALUES);CHKERRQ(ierr); 4835 } 4836 ierr = MatAssemblyBegin(*mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4837 ierr = MatAssemblyEnd(*mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4838 ierr = PetscFree(nnz);CHKERRQ(ierr); 4839 PetscFunctionReturn(0); 4840 } 4841 4842 PetscErrorCode MatSeqAIJInvalidateDiagonal(Mat A) 4843 { 4844 Mat_SeqAIJ *a=(Mat_SeqAIJ*)A->data; 4845 PetscErrorCode ierr; 4846 4847 PetscFunctionBegin; 4848 a->idiagvalid = PETSC_FALSE; 4849 a->ibdiagvalid = PETSC_FALSE; 4850 4851 ierr = MatSeqAIJInvalidateDiagonal_Inode(A);CHKERRQ(ierr); 4852 PetscFunctionReturn(0); 4853 } 4854 4855 PetscErrorCode MatCreateMPIMatConcatenateSeqMat_SeqAIJ(MPI_Comm comm,Mat inmat,PetscInt n,MatReuse scall,Mat *outmat) 4856 { 4857 PetscErrorCode ierr; 4858 PetscMPIInt size; 4859 4860 PetscFunctionBegin; 4861 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 4862 if (size == 1) { 4863 if (scall == MAT_INITIAL_MATRIX) { 4864 ierr = MatDuplicate(inmat,MAT_COPY_VALUES,outmat);CHKERRQ(ierr); 4865 } else { 4866 ierr = MatCopy(inmat,*outmat,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 4867 } 4868 } else { 4869 ierr = MatCreateMPIMatConcatenateSeqMat_MPIAIJ(comm,inmat,n,scall,outmat);CHKERRQ(ierr); 4870 } 4871 PetscFunctionReturn(0); 4872 } 4873 4874 /* 4875 Permute A into C's *local* index space using rowemb,colemb. 4876 The embedding are supposed to be injections and the above implies that the range of rowemb is a subset 4877 of [0,m), colemb is in [0,n). 4878 If pattern == DIFFERENT_NONZERO_PATTERN, C is preallocated according to A. 4879 */ 4880 PetscErrorCode MatSetSeqMat_SeqAIJ(Mat C,IS rowemb,IS colemb,MatStructure pattern,Mat B) 4881 { 4882 /* If making this function public, change the error returned in this function away from _PLIB. */ 4883 PetscErrorCode ierr; 4884 Mat_SeqAIJ *Baij; 4885 PetscBool seqaij; 4886 PetscInt m,n,*nz,i,j,count; 4887 PetscScalar v; 4888 const PetscInt *rowindices,*colindices; 4889 4890 PetscFunctionBegin; 4891 if (!B) PetscFunctionReturn(0); 4892 /* Check to make sure the target matrix (and embeddings) are compatible with C and each other. */ 4893 ierr = PetscObjectBaseTypeCompare((PetscObject)B,MATSEQAIJ,&seqaij);CHKERRQ(ierr); 4894 if (!seqaij) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Input matrix is of wrong type"); 4895 if (rowemb) { 4896 ierr = ISGetLocalSize(rowemb,&m);CHKERRQ(ierr); 4897 if (m != B->rmap->n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Row IS of size %D is incompatible with matrix row size %D",m,B->rmap->n); 4898 } else { 4899 if (C->rmap->n != B->rmap->n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Input matrix is row-incompatible with the target matrix"); 4900 } 4901 if (colemb) { 4902 ierr = ISGetLocalSize(colemb,&n);CHKERRQ(ierr); 4903 if (n != B->cmap->n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Diag col IS of size %D is incompatible with input matrix col size %D",n,B->cmap->n); 4904 } else { 4905 if (C->cmap->n != B->cmap->n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Input matrix is col-incompatible with the target matrix"); 4906 } 4907 4908 Baij = (Mat_SeqAIJ*)(B->data); 4909 if (pattern == DIFFERENT_NONZERO_PATTERN) { 4910 ierr = PetscMalloc1(B->rmap->n,&nz);CHKERRQ(ierr); 4911 for (i=0; i<B->rmap->n; i++) { 4912 nz[i] = Baij->i[i+1] - Baij->i[i]; 4913 } 4914 ierr = MatSeqAIJSetPreallocation(C,0,nz);CHKERRQ(ierr); 4915 ierr = PetscFree(nz);CHKERRQ(ierr); 4916 } 4917 if (pattern == SUBSET_NONZERO_PATTERN) { 4918 ierr = MatZeroEntries(C);CHKERRQ(ierr); 4919 } 4920 count = 0; 4921 rowindices = NULL; 4922 colindices = NULL; 4923 if (rowemb) { 4924 ierr = ISGetIndices(rowemb,&rowindices);CHKERRQ(ierr); 4925 } 4926 if (colemb) { 4927 ierr = ISGetIndices(colemb,&colindices);CHKERRQ(ierr); 4928 } 4929 for (i=0; i<B->rmap->n; i++) { 4930 PetscInt row; 4931 row = i; 4932 if (rowindices) row = rowindices[i]; 4933 for (j=Baij->i[i]; j<Baij->i[i+1]; j++) { 4934 PetscInt col; 4935 col = Baij->j[count]; 4936 if (colindices) col = colindices[col]; 4937 v = Baij->a[count]; 4938 ierr = MatSetValues(C,1,&row,1,&col,&v,INSERT_VALUES);CHKERRQ(ierr); 4939 ++count; 4940 } 4941 } 4942 /* FIXME: set C's nonzerostate correctly. */ 4943 /* Assembly for C is necessary. */ 4944 C->preallocated = PETSC_TRUE; 4945 C->assembled = PETSC_TRUE; 4946 C->was_assembled = PETSC_FALSE; 4947 PetscFunctionReturn(0); 4948 } 4949 4950 PetscFunctionList MatSeqAIJList = NULL; 4951 4952 /*@C 4953 MatSeqAIJSetType - Converts a MATSEQAIJ matrix to a subtype 4954 4955 Collective on Mat 4956 4957 Input Parameters: 4958 + mat - the matrix object 4959 - matype - matrix type 4960 4961 Options Database Key: 4962 . -mat_seqai_type <method> - for example seqaijcrl 4963 4964 4965 Level: intermediate 4966 4967 .seealso: PCSetType(), VecSetType(), MatCreate(), MatType, Mat 4968 @*/ 4969 PetscErrorCode MatSeqAIJSetType(Mat mat, MatType matype) 4970 { 4971 PetscErrorCode ierr,(*r)(Mat,MatType,MatReuse,Mat*); 4972 PetscBool sametype; 4973 4974 PetscFunctionBegin; 4975 PetscValidHeaderSpecific(mat,MAT_CLASSID,1); 4976 ierr = PetscObjectTypeCompare((PetscObject)mat,matype,&sametype);CHKERRQ(ierr); 4977 if (sametype) PetscFunctionReturn(0); 4978 4979 ierr = PetscFunctionListFind(MatSeqAIJList,matype,&r);CHKERRQ(ierr); 4980 if (!r) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_UNKNOWN_TYPE,"Unknown Mat type given: %s",matype); 4981 ierr = (*r)(mat,matype,MAT_INPLACE_MATRIX,&mat);CHKERRQ(ierr); 4982 PetscFunctionReturn(0); 4983 } 4984 4985 4986 /*@C 4987 MatSeqAIJRegister - - Adds a new sub-matrix type for sequential AIJ matrices 4988 4989 Not Collective 4990 4991 Input Parameters: 4992 + name - name of a new user-defined matrix type, for example MATSEQAIJCRL 4993 - function - routine to convert to subtype 4994 4995 Notes: 4996 MatSeqAIJRegister() may be called multiple times to add several user-defined solvers. 4997 4998 4999 Then, your matrix can be chosen with the procedural interface at runtime via the option 5000 $ -mat_seqaij_type my_mat 5001 5002 Level: advanced 5003 5004 .seealso: MatSeqAIJRegisterAll() 5005 5006 5007 Level: advanced 5008 @*/ 5009 PetscErrorCode MatSeqAIJRegister(const char sname[],PetscErrorCode (*function)(Mat,MatType,MatReuse,Mat *)) 5010 { 5011 PetscErrorCode ierr; 5012 5013 PetscFunctionBegin; 5014 ierr = MatInitializePackage();CHKERRQ(ierr); 5015 ierr = PetscFunctionListAdd(&MatSeqAIJList,sname,function);CHKERRQ(ierr); 5016 PetscFunctionReturn(0); 5017 } 5018 5019 PetscBool MatSeqAIJRegisterAllCalled = PETSC_FALSE; 5020 5021 /*@C 5022 MatSeqAIJRegisterAll - Registers all of the matrix subtypes of SeqAIJ 5023 5024 Not Collective 5025 5026 Level: advanced 5027 5028 Developers Note: CUSP and CUSPARSE do not yet support the MatConvert_SeqAIJ..() paradigm and thus cannot be registered here 5029 5030 .seealso: MatRegisterAll(), MatSeqAIJRegister() 5031 @*/ 5032 PetscErrorCode MatSeqAIJRegisterAll(void) 5033 { 5034 PetscErrorCode ierr; 5035 5036 PetscFunctionBegin; 5037 if (MatSeqAIJRegisterAllCalled) PetscFunctionReturn(0); 5038 MatSeqAIJRegisterAllCalled = PETSC_TRUE; 5039 5040 ierr = MatSeqAIJRegister(MATSEQAIJCRL, MatConvert_SeqAIJ_SeqAIJCRL);CHKERRQ(ierr); 5041 ierr = MatSeqAIJRegister(MATSEQAIJPERM, MatConvert_SeqAIJ_SeqAIJPERM);CHKERRQ(ierr); 5042 ierr = MatSeqAIJRegister(MATSEQAIJSELL, MatConvert_SeqAIJ_SeqAIJSELL);CHKERRQ(ierr); 5043 #if defined(PETSC_HAVE_MKL_SPARSE) 5044 ierr = MatSeqAIJRegister(MATSEQAIJMKL, MatConvert_SeqAIJ_SeqAIJMKL);CHKERRQ(ierr); 5045 #endif 5046 #if defined(PETSC_HAVE_VIENNACL) && defined(PETSC_HAVE_VIENNACL_NO_CUDA) 5047 ierr = MatSeqAIJRegister(MATMPIAIJVIENNACL, MatConvert_SeqAIJ_SeqAIJViennaCL);CHKERRQ(ierr); 5048 #endif 5049 PetscFunctionReturn(0); 5050 } 5051 5052 /* 5053 Special version for direct calls from Fortran 5054 */ 5055 #include <petsc/private/fortranimpl.h> 5056 #if defined(PETSC_HAVE_FORTRAN_CAPS) 5057 #define matsetvaluesseqaij_ MATSETVALUESSEQAIJ 5058 #elif !defined(PETSC_HAVE_FORTRAN_UNDERSCORE) 5059 #define matsetvaluesseqaij_ matsetvaluesseqaij 5060 #endif 5061 5062 /* Change these macros so can be used in void function */ 5063 #undef CHKERRQ 5064 #define CHKERRQ(ierr) CHKERRABORT(PetscObjectComm((PetscObject)A),ierr) 5065 #undef SETERRQ2 5066 #define SETERRQ2(comm,ierr,b,c,d) CHKERRABORT(comm,ierr) 5067 #undef SETERRQ3 5068 #define SETERRQ3(comm,ierr,b,c,d,e) CHKERRABORT(comm,ierr) 5069 5070 PETSC_EXTERN void PETSC_STDCALL matsetvaluesseqaij_(Mat *AA,PetscInt *mm,const PetscInt im[],PetscInt *nn,const PetscInt in[],const PetscScalar v[],InsertMode *isis, PetscErrorCode *_ierr) 5071 { 5072 Mat A = *AA; 5073 PetscInt m = *mm, n = *nn; 5074 InsertMode is = *isis; 5075 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 5076 PetscInt *rp,k,low,high,t,ii,row,nrow,i,col,l,rmax,N; 5077 PetscInt *imax,*ai,*ailen; 5078 PetscErrorCode ierr; 5079 PetscInt *aj,nonew = a->nonew,lastcol = -1; 5080 MatScalar *ap,value,*aa; 5081 PetscBool ignorezeroentries = a->ignorezeroentries; 5082 PetscBool roworiented = a->roworiented; 5083 5084 PetscFunctionBegin; 5085 MatCheckPreallocated(A,1); 5086 imax = a->imax; 5087 ai = a->i; 5088 ailen = a->ilen; 5089 aj = a->j; 5090 aa = a->a; 5091 5092 for (k=0; k<m; k++) { /* loop over added rows */ 5093 row = im[k]; 5094 if (row < 0) continue; 5095 #if defined(PETSC_USE_DEBUG) 5096 if (row >= A->rmap->n) SETERRABORT(PetscObjectComm((PetscObject)A),PETSC_ERR_ARG_OUTOFRANGE,"Row too large"); 5097 #endif 5098 rp = aj + ai[row]; ap = aa + ai[row]; 5099 rmax = imax[row]; nrow = ailen[row]; 5100 low = 0; 5101 high = nrow; 5102 for (l=0; l<n; l++) { /* loop over added columns */ 5103 if (in[l] < 0) continue; 5104 #if defined(PETSC_USE_DEBUG) 5105 if (in[l] >= A->cmap->n) SETERRABORT(PetscObjectComm((PetscObject)A),PETSC_ERR_ARG_OUTOFRANGE,"Column too large"); 5106 #endif 5107 col = in[l]; 5108 if (roworiented) value = v[l + k*n]; 5109 else value = v[k + l*m]; 5110 5111 if (value == 0.0 && ignorezeroentries && (is == ADD_VALUES)) continue; 5112 5113 if (col <= lastcol) low = 0; 5114 else high = nrow; 5115 lastcol = col; 5116 while (high-low > 5) { 5117 t = (low+high)/2; 5118 if (rp[t] > col) high = t; 5119 else low = t; 5120 } 5121 for (i=low; i<high; i++) { 5122 if (rp[i] > col) break; 5123 if (rp[i] == col) { 5124 if (is == ADD_VALUES) ap[i] += value; 5125 else ap[i] = value; 5126 goto noinsert; 5127 } 5128 } 5129 if (value == 0.0 && ignorezeroentries) goto noinsert; 5130 if (nonew == 1) goto noinsert; 5131 if (nonew == -1) SETERRABORT(PetscObjectComm((PetscObject)A),PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new nonzero in the matrix"); 5132 MatSeqXAIJReallocateAIJ(A,A->rmap->n,1,nrow,row,col,rmax,aa,ai,aj,rp,ap,imax,nonew,MatScalar); 5133 N = nrow++ - 1; a->nz++; high++; 5134 /* shift up all the later entries in this row */ 5135 for (ii=N; ii>=i; ii--) { 5136 rp[ii+1] = rp[ii]; 5137 ap[ii+1] = ap[ii]; 5138 } 5139 rp[i] = col; 5140 ap[i] = value; 5141 A->nonzerostate++; 5142 noinsert:; 5143 low = i + 1; 5144 } 5145 ailen[row] = nrow; 5146 } 5147 PetscFunctionReturnVoid(); 5148 } 5149