1 #define PETSCMAT_DLL 2 3 #include "../src/mat/impls/baij/mpi/mpibaij.h" /*I "petscmat.h" I*/ 4 5 EXTERN PetscErrorCode MatSetUpMultiply_MPIBAIJ(Mat); 6 EXTERN PetscErrorCode DisAssemble_MPIBAIJ(Mat); 7 EXTERN PetscErrorCode MatIncreaseOverlap_MPIBAIJ(Mat,PetscInt,IS[],PetscInt); 8 EXTERN PetscErrorCode MatGetSubMatrices_MPIBAIJ(Mat,PetscInt,const IS[],const IS[],MatReuse,Mat *[]); 9 EXTERN PetscErrorCode MatGetValues_SeqBAIJ(Mat,PetscInt,const PetscInt[],PetscInt,const PetscInt [],PetscScalar []); 10 EXTERN PetscErrorCode MatSetValues_SeqBAIJ(Mat,PetscInt,const PetscInt[],PetscInt,const PetscInt [],const PetscScalar [],InsertMode); 11 EXTERN PetscErrorCode MatSetValuesBlocked_SeqBAIJ(Mat,PetscInt,const PetscInt[],PetscInt,const PetscInt[],const PetscScalar[],InsertMode); 12 EXTERN PetscErrorCode MatGetRow_SeqBAIJ(Mat,PetscInt,PetscInt*,PetscInt*[],PetscScalar*[]); 13 EXTERN PetscErrorCode MatRestoreRow_SeqBAIJ(Mat,PetscInt,PetscInt*,PetscInt*[],PetscScalar*[]); 14 EXTERN PetscErrorCode MatZeroRows_SeqBAIJ(Mat,PetscInt,const PetscInt[],PetscScalar); 15 16 #undef __FUNCT__ 17 #define __FUNCT__ "MatGetRowMaxAbs_MPIBAIJ" 18 PetscErrorCode MatGetRowMaxAbs_MPIBAIJ(Mat A,Vec v,PetscInt idx[]) 19 { 20 Mat_MPIBAIJ *a = (Mat_MPIBAIJ*)A->data; 21 PetscErrorCode ierr; 22 PetscInt i,*idxb = 0; 23 PetscScalar *va,*vb; 24 Vec vtmp; 25 26 PetscFunctionBegin; 27 ierr = MatGetRowMaxAbs(a->A,v,idx);CHKERRQ(ierr); 28 ierr = VecGetArray(v,&va);CHKERRQ(ierr); 29 if (idx) { 30 for (i=0; i<A->rmap->n; i++) {if (PetscAbsScalar(va[i])) idx[i] += A->cmap->rstart;} 31 } 32 33 ierr = VecCreateSeq(PETSC_COMM_SELF,A->rmap->n,&vtmp);CHKERRQ(ierr); 34 if (idx) {ierr = PetscMalloc(A->rmap->n*sizeof(PetscInt),&idxb);CHKERRQ(ierr);} 35 ierr = MatGetRowMaxAbs(a->B,vtmp,idxb);CHKERRQ(ierr); 36 ierr = VecGetArray(vtmp,&vb);CHKERRQ(ierr); 37 38 for (i=0; i<A->rmap->n; i++){ 39 if (PetscAbsScalar(va[i]) < PetscAbsScalar(vb[i])) {va[i] = vb[i]; if (idx) idx[i] = A->cmap->bs*a->garray[idxb[i]/A->cmap->bs] + (idxb[i] % A->cmap->bs);} 40 } 41 42 ierr = VecRestoreArray(v,&va);CHKERRQ(ierr); 43 ierr = VecRestoreArray(vtmp,&vb);CHKERRQ(ierr); 44 if (idxb) {ierr = PetscFree(idxb);CHKERRQ(ierr);} 45 ierr = VecDestroy(vtmp);CHKERRQ(ierr); 46 PetscFunctionReturn(0); 47 } 48 49 EXTERN_C_BEGIN 50 #undef __FUNCT__ 51 #define __FUNCT__ "MatStoreValues_MPIBAIJ" 52 PetscErrorCode PETSCMAT_DLLEXPORT MatStoreValues_MPIBAIJ(Mat mat) 53 { 54 Mat_MPIBAIJ *aij = (Mat_MPIBAIJ *)mat->data; 55 PetscErrorCode ierr; 56 57 PetscFunctionBegin; 58 ierr = MatStoreValues(aij->A);CHKERRQ(ierr); 59 ierr = MatStoreValues(aij->B);CHKERRQ(ierr); 60 PetscFunctionReturn(0); 61 } 62 EXTERN_C_END 63 64 EXTERN_C_BEGIN 65 #undef __FUNCT__ 66 #define __FUNCT__ "MatRetrieveValues_MPIBAIJ" 67 PetscErrorCode PETSCMAT_DLLEXPORT MatRetrieveValues_MPIBAIJ(Mat mat) 68 { 69 Mat_MPIBAIJ *aij = (Mat_MPIBAIJ *)mat->data; 70 PetscErrorCode ierr; 71 72 PetscFunctionBegin; 73 ierr = MatRetrieveValues(aij->A);CHKERRQ(ierr); 74 ierr = MatRetrieveValues(aij->B);CHKERRQ(ierr); 75 PetscFunctionReturn(0); 76 } 77 EXTERN_C_END 78 79 /* 80 Local utility routine that creates a mapping from the global column 81 number to the local number in the off-diagonal part of the local 82 storage of the matrix. This is done in a non scable way since the 83 length of colmap equals the global matrix length. 84 */ 85 #undef __FUNCT__ 86 #define __FUNCT__ "CreateColmap_MPIBAIJ_Private" 87 PetscErrorCode CreateColmap_MPIBAIJ_Private(Mat mat) 88 { 89 Mat_MPIBAIJ *baij = (Mat_MPIBAIJ*)mat->data; 90 Mat_SeqBAIJ *B = (Mat_SeqBAIJ*)baij->B->data; 91 PetscErrorCode ierr; 92 PetscInt nbs = B->nbs,i,bs=mat->rmap->bs; 93 94 PetscFunctionBegin; 95 #if defined (PETSC_USE_CTABLE) 96 ierr = PetscTableCreate(baij->nbs,&baij->colmap);CHKERRQ(ierr); 97 for (i=0; i<nbs; i++){ 98 ierr = PetscTableAdd(baij->colmap,baij->garray[i]+1,i*bs+1);CHKERRQ(ierr); 99 } 100 #else 101 ierr = PetscMalloc((baij->Nbs+1)*sizeof(PetscInt),&baij->colmap);CHKERRQ(ierr); 102 ierr = PetscLogObjectMemory(mat,baij->Nbs*sizeof(PetscInt));CHKERRQ(ierr); 103 ierr = PetscMemzero(baij->colmap,baij->Nbs*sizeof(PetscInt));CHKERRQ(ierr); 104 for (i=0; i<nbs; i++) baij->colmap[baij->garray[i]] = i*bs+1; 105 #endif 106 PetscFunctionReturn(0); 107 } 108 109 #define CHUNKSIZE 10 110 111 #define MatSetValues_SeqBAIJ_A_Private(row,col,value,addv) \ 112 { \ 113 \ 114 brow = row/bs; \ 115 rp = aj + ai[brow]; ap = aa + bs2*ai[brow]; \ 116 rmax = aimax[brow]; nrow = ailen[brow]; \ 117 bcol = col/bs; \ 118 ridx = row % bs; cidx = col % bs; \ 119 low = 0; high = nrow; \ 120 while (high-low > 3) { \ 121 t = (low+high)/2; \ 122 if (rp[t] > bcol) high = t; \ 123 else low = t; \ 124 } \ 125 for (_i=low; _i<high; _i++) { \ 126 if (rp[_i] > bcol) break; \ 127 if (rp[_i] == bcol) { \ 128 bap = ap + bs2*_i + bs*cidx + ridx; \ 129 if (addv == ADD_VALUES) *bap += value; \ 130 else *bap = value; \ 131 goto a_noinsert; \ 132 } \ 133 } \ 134 if (a->nonew == 1) goto a_noinsert; \ 135 if (a->nonew == -1) SETERRQ2(PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new nonzero (%D, %D) into matrix", row, col); \ 136 MatSeqXAIJReallocateAIJ(A,a->mbs,bs2,nrow,brow,bcol,rmax,aa,ai,aj,rp,ap,aimax,a->nonew,MatScalar); \ 137 N = nrow++ - 1; \ 138 /* shift up all the later entries in this row */ \ 139 for (ii=N; ii>=_i; ii--) { \ 140 rp[ii+1] = rp[ii]; \ 141 ierr = PetscMemcpy(ap+bs2*(ii+1),ap+bs2*(ii),bs2*sizeof(MatScalar));CHKERRQ(ierr); \ 142 } \ 143 if (N>=_i) { ierr = PetscMemzero(ap+bs2*_i,bs2*sizeof(MatScalar));CHKERRQ(ierr); } \ 144 rp[_i] = bcol; \ 145 ap[bs2*_i + bs*cidx + ridx] = value; \ 146 a_noinsert:; \ 147 ailen[brow] = nrow; \ 148 } 149 150 #define MatSetValues_SeqBAIJ_B_Private(row,col,value,addv) \ 151 { \ 152 brow = row/bs; \ 153 rp = bj + bi[brow]; ap = ba + bs2*bi[brow]; \ 154 rmax = bimax[brow]; nrow = bilen[brow]; \ 155 bcol = col/bs; \ 156 ridx = row % bs; cidx = col % bs; \ 157 low = 0; high = nrow; \ 158 while (high-low > 3) { \ 159 t = (low+high)/2; \ 160 if (rp[t] > bcol) high = t; \ 161 else low = t; \ 162 } \ 163 for (_i=low; _i<high; _i++) { \ 164 if (rp[_i] > bcol) break; \ 165 if (rp[_i] == bcol) { \ 166 bap = ap + bs2*_i + bs*cidx + ridx; \ 167 if (addv == ADD_VALUES) *bap += value; \ 168 else *bap = value; \ 169 goto b_noinsert; \ 170 } \ 171 } \ 172 if (b->nonew == 1) goto b_noinsert; \ 173 if (b->nonew == -1) SETERRQ2(PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new nonzero (%D, %D) into matrix", row, col); \ 174 MatSeqXAIJReallocateAIJ(B,b->mbs,bs2,nrow,brow,bcol,rmax,ba,bi,bj,rp,ap,bimax,b->nonew,MatScalar); \ 175 CHKMEMQ;\ 176 N = nrow++ - 1; \ 177 /* shift up all the later entries in this row */ \ 178 for (ii=N; ii>=_i; ii--) { \ 179 rp[ii+1] = rp[ii]; \ 180 ierr = PetscMemcpy(ap+bs2*(ii+1),ap+bs2*(ii),bs2*sizeof(MatScalar));CHKERRQ(ierr); \ 181 } \ 182 if (N>=_i) { ierr = PetscMemzero(ap+bs2*_i,bs2*sizeof(MatScalar));CHKERRQ(ierr);} \ 183 rp[_i] = bcol; \ 184 ap[bs2*_i + bs*cidx + ridx] = value; \ 185 b_noinsert:; \ 186 bilen[brow] = nrow; \ 187 } 188 189 #undef __FUNCT__ 190 #define __FUNCT__ "MatSetValues_MPIBAIJ" 191 PetscErrorCode MatSetValues_MPIBAIJ(Mat mat,PetscInt m,const PetscInt im[],PetscInt n,const PetscInt in[],const PetscScalar v[],InsertMode addv) 192 { 193 Mat_MPIBAIJ *baij = (Mat_MPIBAIJ*)mat->data; 194 MatScalar value; 195 PetscTruth roworiented = baij->roworiented; 196 PetscErrorCode ierr; 197 PetscInt i,j,row,col; 198 PetscInt rstart_orig=mat->rmap->rstart; 199 PetscInt rend_orig=mat->rmap->rend,cstart_orig=mat->cmap->rstart; 200 PetscInt cend_orig=mat->cmap->rend,bs=mat->rmap->bs; 201 202 /* Some Variables required in the macro */ 203 Mat A = baij->A; 204 Mat_SeqBAIJ *a = (Mat_SeqBAIJ*)(A)->data; 205 PetscInt *aimax=a->imax,*ai=a->i,*ailen=a->ilen,*aj=a->j; 206 MatScalar *aa=a->a; 207 208 Mat B = baij->B; 209 Mat_SeqBAIJ *b = (Mat_SeqBAIJ*)(B)->data; 210 PetscInt *bimax=b->imax,*bi=b->i,*bilen=b->ilen,*bj=b->j; 211 MatScalar *ba=b->a; 212 213 PetscInt *rp,ii,nrow,_i,rmax,N,brow,bcol; 214 PetscInt low,high,t,ridx,cidx,bs2=a->bs2; 215 MatScalar *ap,*bap; 216 217 PetscFunctionBegin; 218 for (i=0; i<m; i++) { 219 if (im[i] < 0) continue; 220 #if defined(PETSC_USE_DEBUG) 221 if (im[i] >= mat->rmap->N) SETERRQ2(PETSC_ERR_ARG_OUTOFRANGE,"Row too large: row %D max %D",im[i],mat->rmap->N-1); 222 #endif 223 if (im[i] >= rstart_orig && im[i] < rend_orig) { 224 row = im[i] - rstart_orig; 225 for (j=0; j<n; j++) { 226 if (in[j] >= cstart_orig && in[j] < cend_orig){ 227 col = in[j] - cstart_orig; 228 if (roworiented) value = v[i*n+j]; else value = v[i+j*m]; 229 MatSetValues_SeqBAIJ_A_Private(row,col,value,addv); 230 /* ierr = MatSetValues_SeqBAIJ(baij->A,1,&row,1,&col,&value,addv);CHKERRQ(ierr); */ 231 } else if (in[j] < 0) continue; 232 #if defined(PETSC_USE_DEBUG) 233 else if (in[j] >= mat->cmap->N) {SETERRQ2(PETSC_ERR_ARG_OUTOFRANGE,"Column too large: col %D max %D",in[i],mat->cmap->N-1);} 234 #endif 235 else { 236 if (mat->was_assembled) { 237 if (!baij->colmap) { 238 ierr = CreateColmap_MPIBAIJ_Private(mat);CHKERRQ(ierr); 239 } 240 #if defined (PETSC_USE_CTABLE) 241 ierr = PetscTableFind(baij->colmap,in[j]/bs + 1,&col);CHKERRQ(ierr); 242 col = col - 1; 243 #else 244 col = baij->colmap[in[j]/bs] - 1; 245 #endif 246 if (col < 0 && !((Mat_SeqBAIJ*)(baij->A->data))->nonew) { 247 ierr = DisAssemble_MPIBAIJ(mat);CHKERRQ(ierr); 248 col = in[j]; 249 /* Reinitialize the variables required by MatSetValues_SeqBAIJ_B_Private() */ 250 B = baij->B; 251 b = (Mat_SeqBAIJ*)(B)->data; 252 bimax=b->imax;bi=b->i;bilen=b->ilen;bj=b->j; 253 ba=b->a; 254 } else col += in[j]%bs; 255 } else col = in[j]; 256 if (roworiented) value = v[i*n+j]; else value = v[i+j*m]; 257 MatSetValues_SeqBAIJ_B_Private(row,col,value,addv); 258 /* ierr = MatSetValues_SeqBAIJ(baij->B,1,&row,1,&col,&value,addv);CHKERRQ(ierr); */ 259 } 260 } 261 } else { 262 if (!baij->donotstash) { 263 if (roworiented) { 264 ierr = MatStashValuesRow_Private(&mat->stash,im[i],n,in,v+i*n);CHKERRQ(ierr); 265 } else { 266 ierr = MatStashValuesCol_Private(&mat->stash,im[i],n,in,v+i,m);CHKERRQ(ierr); 267 } 268 } 269 } 270 } 271 PetscFunctionReturn(0); 272 } 273 274 #undef __FUNCT__ 275 #define __FUNCT__ "MatSetValuesBlocked_MPIBAIJ" 276 PetscErrorCode MatSetValuesBlocked_MPIBAIJ(Mat mat,PetscInt m,const PetscInt im[],PetscInt n,const PetscInt in[],const PetscScalar v[],InsertMode addv) 277 { 278 Mat_MPIBAIJ *baij = (Mat_MPIBAIJ*)mat->data; 279 const PetscScalar *value; 280 MatScalar *barray=baij->barray; 281 PetscTruth roworiented = baij->roworiented; 282 PetscErrorCode ierr; 283 PetscInt i,j,ii,jj,row,col,rstart=baij->rstartbs; 284 PetscInt rend=baij->rendbs,cstart=baij->cstartbs,stepval; 285 PetscInt cend=baij->cendbs,bs=mat->rmap->bs,bs2=baij->bs2; 286 287 PetscFunctionBegin; 288 if(!barray) { 289 ierr = PetscMalloc(bs2*sizeof(MatScalar),&barray);CHKERRQ(ierr); 290 baij->barray = barray; 291 } 292 293 if (roworiented) { 294 stepval = (n-1)*bs; 295 } else { 296 stepval = (m-1)*bs; 297 } 298 for (i=0; i<m; i++) { 299 if (im[i] < 0) continue; 300 #if defined(PETSC_USE_DEBUG) 301 if (im[i] >= baij->Mbs) SETERRQ2(PETSC_ERR_ARG_OUTOFRANGE,"Row too large, row %D max %D",im[i],baij->Mbs-1); 302 #endif 303 if (im[i] >= rstart && im[i] < rend) { 304 row = im[i] - rstart; 305 for (j=0; j<n; j++) { 306 /* If NumCol = 1 then a copy is not required */ 307 if ((roworiented) && (n == 1)) { 308 barray = (MatScalar*)v + i*bs2; 309 } else if((!roworiented) && (m == 1)) { 310 barray = (MatScalar*)v + j*bs2; 311 } else { /* Here a copy is required */ 312 if (roworiented) { 313 value = v + i*(stepval+bs)*bs + j*bs; 314 } else { 315 value = v + j*(stepval+bs)*bs + i*bs; 316 } 317 for (ii=0; ii<bs; ii++,value+=stepval) { 318 for (jj=0; jj<bs; jj++) { 319 *barray++ = *value++; 320 } 321 } 322 barray -=bs2; 323 } 324 325 if (in[j] >= cstart && in[j] < cend){ 326 col = in[j] - cstart; 327 ierr = MatSetValuesBlocked_SeqBAIJ(baij->A,1,&row,1,&col,barray,addv);CHKERRQ(ierr); 328 } 329 else if (in[j] < 0) continue; 330 #if defined(PETSC_USE_DEBUG) 331 else if (in[j] >= baij->Nbs) {SETERRQ2(PETSC_ERR_ARG_OUTOFRANGE,"Column too large, col %D max %D",in[j],baij->Nbs-1);} 332 #endif 333 else { 334 if (mat->was_assembled) { 335 if (!baij->colmap) { 336 ierr = CreateColmap_MPIBAIJ_Private(mat);CHKERRQ(ierr); 337 } 338 339 #if defined(PETSC_USE_DEBUG) 340 #if defined (PETSC_USE_CTABLE) 341 { PetscInt data; 342 ierr = PetscTableFind(baij->colmap,in[j]+1,&data);CHKERRQ(ierr); 343 if ((data - 1) % bs) SETERRQ(PETSC_ERR_PLIB,"Incorrect colmap"); 344 } 345 #else 346 if ((baij->colmap[in[j]] - 1) % bs) SETERRQ(PETSC_ERR_PLIB,"Incorrect colmap"); 347 #endif 348 #endif 349 #if defined (PETSC_USE_CTABLE) 350 ierr = PetscTableFind(baij->colmap,in[j]+1,&col);CHKERRQ(ierr); 351 col = (col - 1)/bs; 352 #else 353 col = (baij->colmap[in[j]] - 1)/bs; 354 #endif 355 if (col < 0 && !((Mat_SeqBAIJ*)(baij->A->data))->nonew) { 356 ierr = DisAssemble_MPIBAIJ(mat);CHKERRQ(ierr); 357 col = in[j]; 358 } 359 } 360 else col = in[j]; 361 ierr = MatSetValuesBlocked_SeqBAIJ(baij->B,1,&row,1,&col,barray,addv);CHKERRQ(ierr); 362 } 363 } 364 } else { 365 if (!baij->donotstash) { 366 if (roworiented) { 367 ierr = MatStashValuesRowBlocked_Private(&mat->bstash,im[i],n,in,v,m,n,i);CHKERRQ(ierr); 368 } else { 369 ierr = MatStashValuesColBlocked_Private(&mat->bstash,im[i],n,in,v,m,n,i);CHKERRQ(ierr); 370 } 371 } 372 } 373 } 374 PetscFunctionReturn(0); 375 } 376 377 #define HASH_KEY 0.6180339887 378 #define HASH(size,key,tmp) (tmp = (key)*HASH_KEY,(PetscInt)((size)*(tmp-(PetscInt)tmp))) 379 /* #define HASH(size,key) ((PetscInt)((size)*fmod(((key)*HASH_KEY),1))) */ 380 /* #define HASH(size,key,tmp) ((PetscInt)((size)*fmod(((key)*HASH_KEY),1))) */ 381 #undef __FUNCT__ 382 #define __FUNCT__ "MatSetValues_MPIBAIJ_HT" 383 PetscErrorCode MatSetValues_MPIBAIJ_HT(Mat mat,PetscInt m,const PetscInt im[],PetscInt n,const PetscInt in[],const PetscScalar v[],InsertMode addv) 384 { 385 Mat_MPIBAIJ *baij = (Mat_MPIBAIJ*)mat->data; 386 PetscTruth roworiented = baij->roworiented; 387 PetscErrorCode ierr; 388 PetscInt i,j,row,col; 389 PetscInt rstart_orig=mat->rmap->rstart; 390 PetscInt rend_orig=mat->rmap->rend,Nbs=baij->Nbs; 391 PetscInt h1,key,size=baij->ht_size,bs=mat->rmap->bs,*HT=baij->ht,idx; 392 PetscReal tmp; 393 MatScalar **HD = baij->hd,value; 394 #if defined(PETSC_USE_DEBUG) 395 PetscInt total_ct=baij->ht_total_ct,insert_ct=baij->ht_insert_ct; 396 #endif 397 398 PetscFunctionBegin; 399 400 for (i=0; i<m; i++) { 401 #if defined(PETSC_USE_DEBUG) 402 if (im[i] < 0) SETERRQ(PETSC_ERR_ARG_OUTOFRANGE,"Negative row"); 403 if (im[i] >= mat->rmap->N) SETERRQ2(PETSC_ERR_ARG_OUTOFRANGE,"Row too large: row %D max %D",im[i],mat->rmap->N-1); 404 #endif 405 row = im[i]; 406 if (row >= rstart_orig && row < rend_orig) { 407 for (j=0; j<n; j++) { 408 col = in[j]; 409 if (roworiented) value = v[i*n+j]; else value = v[i+j*m]; 410 /* Look up PetscInto the Hash Table */ 411 key = (row/bs)*Nbs+(col/bs)+1; 412 h1 = HASH(size,key,tmp); 413 414 415 idx = h1; 416 #if defined(PETSC_USE_DEBUG) 417 insert_ct++; 418 total_ct++; 419 if (HT[idx] != key) { 420 for (idx=h1; (idx<size) && (HT[idx]!=key); idx++,total_ct++); 421 if (idx == size) { 422 for (idx=0; (idx<h1) && (HT[idx]!=key); idx++,total_ct++); 423 if (idx == h1) { 424 SETERRQ2(PETSC_ERR_ARG_OUTOFRANGE,"(%D,%D) has no entry in the hash table", row, col); 425 } 426 } 427 } 428 #else 429 if (HT[idx] != key) { 430 for (idx=h1; (idx<size) && (HT[idx]!=key); idx++); 431 if (idx == size) { 432 for (idx=0; (idx<h1) && (HT[idx]!=key); idx++); 433 if (idx == h1) { 434 SETERRQ2(PETSC_ERR_ARG_OUTOFRANGE,"(%D,%D) has no entry in the hash table", row, col); 435 } 436 } 437 } 438 #endif 439 /* A HASH table entry is found, so insert the values at the correct address */ 440 if (addv == ADD_VALUES) *(HD[idx]+ (col % bs)*bs + (row % bs)) += value; 441 else *(HD[idx]+ (col % bs)*bs + (row % bs)) = value; 442 } 443 } else { 444 if (!baij->donotstash) { 445 if (roworiented) { 446 ierr = MatStashValuesRow_Private(&mat->stash,im[i],n,in,v+i*n);CHKERRQ(ierr); 447 } else { 448 ierr = MatStashValuesCol_Private(&mat->stash,im[i],n,in,v+i,m);CHKERRQ(ierr); 449 } 450 } 451 } 452 } 453 #if defined(PETSC_USE_DEBUG) 454 baij->ht_total_ct = total_ct; 455 baij->ht_insert_ct = insert_ct; 456 #endif 457 PetscFunctionReturn(0); 458 } 459 460 #undef __FUNCT__ 461 #define __FUNCT__ "MatSetValuesBlocked_MPIBAIJ_HT" 462 PetscErrorCode MatSetValuesBlocked_MPIBAIJ_HT(Mat mat,PetscInt m,const PetscInt im[],PetscInt n,const PetscInt in[],const PetscScalar v[],InsertMode addv) 463 { 464 Mat_MPIBAIJ *baij = (Mat_MPIBAIJ*)mat->data; 465 PetscTruth roworiented = baij->roworiented; 466 PetscErrorCode ierr; 467 PetscInt i,j,ii,jj,row,col; 468 PetscInt rstart=baij->rstartbs; 469 PetscInt rend=mat->rmap->rend,stepval,bs=mat->rmap->bs,bs2=baij->bs2,nbs2=n*bs2; 470 PetscInt h1,key,size=baij->ht_size,idx,*HT=baij->ht,Nbs=baij->Nbs; 471 PetscReal tmp; 472 MatScalar **HD = baij->hd,*baij_a; 473 const PetscScalar *v_t,*value; 474 #if defined(PETSC_USE_DEBUG) 475 PetscInt total_ct=baij->ht_total_ct,insert_ct=baij->ht_insert_ct; 476 #endif 477 478 PetscFunctionBegin; 479 480 if (roworiented) { 481 stepval = (n-1)*bs; 482 } else { 483 stepval = (m-1)*bs; 484 } 485 for (i=0; i<m; i++) { 486 #if defined(PETSC_USE_DEBUG) 487 if (im[i] < 0) SETERRQ1(PETSC_ERR_ARG_OUTOFRANGE,"Negative row: %D",im[i]); 488 if (im[i] >= baij->Mbs) SETERRQ2(PETSC_ERR_ARG_OUTOFRANGE,"Row too large: row %D max %D",im[i],baij->Mbs-1); 489 #endif 490 row = im[i]; 491 v_t = v + i*nbs2; 492 if (row >= rstart && row < rend) { 493 for (j=0; j<n; j++) { 494 col = in[j]; 495 496 /* Look up into the Hash Table */ 497 key = row*Nbs+col+1; 498 h1 = HASH(size,key,tmp); 499 500 idx = h1; 501 #if defined(PETSC_USE_DEBUG) 502 total_ct++; 503 insert_ct++; 504 if (HT[idx] != key) { 505 for (idx=h1; (idx<size) && (HT[idx]!=key); idx++,total_ct++); 506 if (idx == size) { 507 for (idx=0; (idx<h1) && (HT[idx]!=key); idx++,total_ct++); 508 if (idx == h1) { 509 SETERRQ2(PETSC_ERR_ARG_OUTOFRANGE,"(%D,%D) has no entry in the hash table", row, col); 510 } 511 } 512 } 513 #else 514 if (HT[idx] != key) { 515 for (idx=h1; (idx<size) && (HT[idx]!=key); idx++); 516 if (idx == size) { 517 for (idx=0; (idx<h1) && (HT[idx]!=key); idx++); 518 if (idx == h1) { 519 SETERRQ2(PETSC_ERR_ARG_OUTOFRANGE,"(%D,%D) has no entry in the hash table", row, col); 520 } 521 } 522 } 523 #endif 524 baij_a = HD[idx]; 525 if (roworiented) { 526 /*value = v + i*(stepval+bs)*bs + j*bs;*/ 527 /* value = v + (i*(stepval+bs)+j)*bs; */ 528 value = v_t; 529 v_t += bs; 530 if (addv == ADD_VALUES) { 531 for (ii=0; ii<bs; ii++,value+=stepval) { 532 for (jj=ii; jj<bs2; jj+=bs) { 533 baij_a[jj] += *value++; 534 } 535 } 536 } else { 537 for (ii=0; ii<bs; ii++,value+=stepval) { 538 for (jj=ii; jj<bs2; jj+=bs) { 539 baij_a[jj] = *value++; 540 } 541 } 542 } 543 } else { 544 value = v + j*(stepval+bs)*bs + i*bs; 545 if (addv == ADD_VALUES) { 546 for (ii=0; ii<bs; ii++,value+=stepval,baij_a+=bs) { 547 for (jj=0; jj<bs; jj++) { 548 baij_a[jj] += *value++; 549 } 550 } 551 } else { 552 for (ii=0; ii<bs; ii++,value+=stepval,baij_a+=bs) { 553 for (jj=0; jj<bs; jj++) { 554 baij_a[jj] = *value++; 555 } 556 } 557 } 558 } 559 } 560 } else { 561 if (!baij->donotstash) { 562 if (roworiented) { 563 ierr = MatStashValuesRowBlocked_Private(&mat->bstash,im[i],n,in,v,m,n,i);CHKERRQ(ierr); 564 } else { 565 ierr = MatStashValuesColBlocked_Private(&mat->bstash,im[i],n,in,v,m,n,i);CHKERRQ(ierr); 566 } 567 } 568 } 569 } 570 #if defined(PETSC_USE_DEBUG) 571 baij->ht_total_ct = total_ct; 572 baij->ht_insert_ct = insert_ct; 573 #endif 574 PetscFunctionReturn(0); 575 } 576 577 #undef __FUNCT__ 578 #define __FUNCT__ "MatGetValues_MPIBAIJ" 579 PetscErrorCode MatGetValues_MPIBAIJ(Mat mat,PetscInt m,const PetscInt idxm[],PetscInt n,const PetscInt idxn[],PetscScalar v[]) 580 { 581 Mat_MPIBAIJ *baij = (Mat_MPIBAIJ*)mat->data; 582 PetscErrorCode ierr; 583 PetscInt bs=mat->rmap->bs,i,j,bsrstart = mat->rmap->rstart,bsrend = mat->rmap->rend; 584 PetscInt bscstart = mat->cmap->rstart,bscend = mat->cmap->rend,row,col,data; 585 586 PetscFunctionBegin; 587 for (i=0; i<m; i++) { 588 if (idxm[i] < 0) continue; /* SETERRQ1(PETSC_ERR_ARG_OUTOFRANGE,"Negative row: %D",idxm[i]);*/ 589 if (idxm[i] >= mat->rmap->N) SETERRQ2(PETSC_ERR_ARG_OUTOFRANGE,"Row too large: row %D max %D",idxm[i],mat->rmap->N-1); 590 if (idxm[i] >= bsrstart && idxm[i] < bsrend) { 591 row = idxm[i] - bsrstart; 592 for (j=0; j<n; j++) { 593 if (idxn[j] < 0) continue; /* SETERRQ1(PETSC_ERR_ARG_OUTOFRANGE,"Negative column: %D",idxn[j]); */ 594 if (idxn[j] >= mat->cmap->N) SETERRQ2(PETSC_ERR_ARG_OUTOFRANGE,"Column too large: col %D max %D",idxn[j],mat->cmap->N-1); 595 if (idxn[j] >= bscstart && idxn[j] < bscend){ 596 col = idxn[j] - bscstart; 597 ierr = MatGetValues_SeqBAIJ(baij->A,1,&row,1,&col,v+i*n+j);CHKERRQ(ierr); 598 } else { 599 if (!baij->colmap) { 600 ierr = CreateColmap_MPIBAIJ_Private(mat);CHKERRQ(ierr); 601 } 602 #if defined (PETSC_USE_CTABLE) 603 ierr = PetscTableFind(baij->colmap,idxn[j]/bs+1,&data);CHKERRQ(ierr); 604 data --; 605 #else 606 data = baij->colmap[idxn[j]/bs]-1; 607 #endif 608 if((data < 0) || (baij->garray[data/bs] != idxn[j]/bs)) *(v+i*n+j) = 0.0; 609 else { 610 col = data + idxn[j]%bs; 611 ierr = MatGetValues_SeqBAIJ(baij->B,1,&row,1,&col,v+i*n+j);CHKERRQ(ierr); 612 } 613 } 614 } 615 } else { 616 SETERRQ(PETSC_ERR_SUP,"Only local values currently supported"); 617 } 618 } 619 PetscFunctionReturn(0); 620 } 621 622 #undef __FUNCT__ 623 #define __FUNCT__ "MatNorm_MPIBAIJ" 624 PetscErrorCode MatNorm_MPIBAIJ(Mat mat,NormType type,PetscReal *nrm) 625 { 626 Mat_MPIBAIJ *baij = (Mat_MPIBAIJ*)mat->data; 627 Mat_SeqBAIJ *amat = (Mat_SeqBAIJ*)baij->A->data,*bmat = (Mat_SeqBAIJ*)baij->B->data; 628 PetscErrorCode ierr; 629 PetscInt i,j,bs2=baij->bs2,bs=baij->A->rmap->bs,nz,row,col; 630 PetscReal sum = 0.0; 631 MatScalar *v; 632 633 PetscFunctionBegin; 634 if (baij->size == 1) { 635 ierr = MatNorm(baij->A,type,nrm);CHKERRQ(ierr); 636 } else { 637 if (type == NORM_FROBENIUS) { 638 v = amat->a; 639 nz = amat->nz*bs2; 640 for (i=0; i<nz; i++) { 641 #if defined(PETSC_USE_COMPLEX) 642 sum += PetscRealPart(PetscConj(*v)*(*v)); v++; 643 #else 644 sum += (*v)*(*v); v++; 645 #endif 646 } 647 v = bmat->a; 648 nz = bmat->nz*bs2; 649 for (i=0; i<nz; i++) { 650 #if defined(PETSC_USE_COMPLEX) 651 sum += PetscRealPart(PetscConj(*v)*(*v)); v++; 652 #else 653 sum += (*v)*(*v); v++; 654 #endif 655 } 656 ierr = MPI_Allreduce(&sum,nrm,1,MPIU_REAL,MPI_SUM,((PetscObject)mat)->comm);CHKERRQ(ierr); 657 *nrm = sqrt(*nrm); 658 } else if (type == NORM_1) { /* max column sum */ 659 PetscReal *tmp,*tmp2; 660 PetscInt *jj,*garray=baij->garray,cstart=baij->rstartbs; 661 ierr = PetscMalloc((2*mat->cmap->N+1)*sizeof(PetscReal),&tmp);CHKERRQ(ierr); 662 tmp2 = tmp + mat->cmap->N; 663 ierr = PetscMemzero(tmp,mat->cmap->N*sizeof(PetscReal));CHKERRQ(ierr); 664 v = amat->a; jj = amat->j; 665 for (i=0; i<amat->nz; i++) { 666 for (j=0; j<bs; j++){ 667 col = bs*(cstart + *jj) + j; /* column index */ 668 for (row=0; row<bs; row++){ 669 tmp[col] += PetscAbsScalar(*v); v++; 670 } 671 } 672 jj++; 673 } 674 v = bmat->a; jj = bmat->j; 675 for (i=0; i<bmat->nz; i++) { 676 for (j=0; j<bs; j++){ 677 col = bs*garray[*jj] + j; 678 for (row=0; row<bs; row++){ 679 tmp[col] += PetscAbsScalar(*v); v++; 680 } 681 } 682 jj++; 683 } 684 ierr = MPI_Allreduce(tmp,tmp2,mat->cmap->N,MPIU_REAL,MPI_SUM,((PetscObject)mat)->comm);CHKERRQ(ierr); 685 *nrm = 0.0; 686 for (j=0; j<mat->cmap->N; j++) { 687 if (tmp2[j] > *nrm) *nrm = tmp2[j]; 688 } 689 ierr = PetscFree(tmp);CHKERRQ(ierr); 690 } else if (type == NORM_INFINITY) { /* max row sum */ 691 PetscReal *sums; 692 ierr = PetscMalloc(bs*sizeof(PetscReal),&sums);CHKERRQ(ierr) 693 sum = 0.0; 694 for (j=0; j<amat->mbs; j++) { 695 for (row=0; row<bs; row++) sums[row] = 0.0; 696 v = amat->a + bs2*amat->i[j]; 697 nz = amat->i[j+1]-amat->i[j]; 698 for (i=0; i<nz; i++) { 699 for (col=0; col<bs; col++){ 700 for (row=0; row<bs; row++){ 701 sums[row] += PetscAbsScalar(*v); v++; 702 } 703 } 704 } 705 v = bmat->a + bs2*bmat->i[j]; 706 nz = bmat->i[j+1]-bmat->i[j]; 707 for (i=0; i<nz; i++) { 708 for (col=0; col<bs; col++){ 709 for (row=0; row<bs; row++){ 710 sums[row] += PetscAbsScalar(*v); v++; 711 } 712 } 713 } 714 for (row=0; row<bs; row++){ 715 if (sums[row] > sum) sum = sums[row]; 716 } 717 } 718 ierr = MPI_Allreduce(&sum,nrm,1,MPIU_REAL,MPI_MAX,((PetscObject)mat)->comm);CHKERRQ(ierr); 719 ierr = PetscFree(sums);CHKERRQ(ierr); 720 } else { 721 SETERRQ(PETSC_ERR_SUP,"No support for this norm yet"); 722 } 723 } 724 PetscFunctionReturn(0); 725 } 726 727 /* 728 Creates the hash table, and sets the table 729 This table is created only once. 730 If new entried need to be added to the matrix 731 then the hash table has to be destroyed and 732 recreated. 733 */ 734 #undef __FUNCT__ 735 #define __FUNCT__ "MatCreateHashTable_MPIBAIJ_Private" 736 PetscErrorCode MatCreateHashTable_MPIBAIJ_Private(Mat mat,PetscReal factor) 737 { 738 Mat_MPIBAIJ *baij = (Mat_MPIBAIJ*)mat->data; 739 Mat A = baij->A,B=baij->B; 740 Mat_SeqBAIJ *a=(Mat_SeqBAIJ *)A->data,*b=(Mat_SeqBAIJ *)B->data; 741 PetscInt i,j,k,nz=a->nz+b->nz,h1,*ai=a->i,*aj=a->j,*bi=b->i,*bj=b->j; 742 PetscErrorCode ierr; 743 PetscInt size,bs2=baij->bs2,rstart=baij->rstartbs; 744 PetscInt cstart=baij->cstartbs,*garray=baij->garray,row,col,Nbs=baij->Nbs; 745 PetscInt *HT,key; 746 MatScalar **HD; 747 PetscReal tmp; 748 #if defined(PETSC_USE_INFO) 749 PetscInt ct=0,max=0; 750 #endif 751 752 PetscFunctionBegin; 753 baij->ht_size=(PetscInt)(factor*nz); 754 size = baij->ht_size; 755 756 if (baij->ht) { 757 PetscFunctionReturn(0); 758 } 759 760 /* Allocate Memory for Hash Table */ 761 ierr = PetscMalloc((size)*(sizeof(PetscInt)+sizeof(MatScalar*))+1,&baij->hd);CHKERRQ(ierr); 762 baij->ht = (PetscInt*)(baij->hd + size); 763 HD = baij->hd; 764 HT = baij->ht; 765 766 767 ierr = PetscMemzero(HD,size*(sizeof(PetscInt)+sizeof(PetscScalar*)));CHKERRQ(ierr); 768 769 770 /* Loop Over A */ 771 for (i=0; i<a->mbs; i++) { 772 for (j=ai[i]; j<ai[i+1]; j++) { 773 row = i+rstart; 774 col = aj[j]+cstart; 775 776 key = row*Nbs + col + 1; 777 h1 = HASH(size,key,tmp); 778 for (k=0; k<size; k++){ 779 if (!HT[(h1+k)%size]) { 780 HT[(h1+k)%size] = key; 781 HD[(h1+k)%size] = a->a + j*bs2; 782 break; 783 #if defined(PETSC_USE_INFO) 784 } else { 785 ct++; 786 #endif 787 } 788 } 789 #if defined(PETSC_USE_INFO) 790 if (k> max) max = k; 791 #endif 792 } 793 } 794 /* Loop Over B */ 795 for (i=0; i<b->mbs; i++) { 796 for (j=bi[i]; j<bi[i+1]; j++) { 797 row = i+rstart; 798 col = garray[bj[j]]; 799 key = row*Nbs + col + 1; 800 h1 = HASH(size,key,tmp); 801 for (k=0; k<size; k++){ 802 if (!HT[(h1+k)%size]) { 803 HT[(h1+k)%size] = key; 804 HD[(h1+k)%size] = b->a + j*bs2; 805 break; 806 #if defined(PETSC_USE_INFO) 807 } else { 808 ct++; 809 #endif 810 } 811 } 812 #if defined(PETSC_USE_INFO) 813 if (k> max) max = k; 814 #endif 815 } 816 } 817 818 /* Print Summary */ 819 #if defined(PETSC_USE_INFO) 820 for (i=0,j=0; i<size; i++) { 821 if (HT[i]) {j++;} 822 } 823 ierr = PetscInfo2(mat,"Average Search = %5.2f,max search = %D\n",(!j)? 0.0:((PetscReal)(ct+j))/j,max);CHKERRQ(ierr); 824 #endif 825 PetscFunctionReturn(0); 826 } 827 828 #undef __FUNCT__ 829 #define __FUNCT__ "MatAssemblyBegin_MPIBAIJ" 830 PetscErrorCode MatAssemblyBegin_MPIBAIJ(Mat mat,MatAssemblyType mode) 831 { 832 Mat_MPIBAIJ *baij = (Mat_MPIBAIJ*)mat->data; 833 PetscErrorCode ierr; 834 PetscInt nstash,reallocs; 835 InsertMode addv; 836 837 PetscFunctionBegin; 838 if (baij->donotstash) { 839 PetscFunctionReturn(0); 840 } 841 842 /* make sure all processors are either in INSERTMODE or ADDMODE */ 843 ierr = MPI_Allreduce(&mat->insertmode,&addv,1,MPI_INT,MPI_BOR,((PetscObject)mat)->comm);CHKERRQ(ierr); 844 if (addv == (ADD_VALUES|INSERT_VALUES)) { 845 SETERRQ(PETSC_ERR_ARG_WRONGSTATE,"Some processors inserted others added"); 846 } 847 mat->insertmode = addv; /* in case this processor had no cache */ 848 849 ierr = MatStashScatterBegin_Private(mat,&mat->stash,mat->rmap->range);CHKERRQ(ierr); 850 ierr = MatStashScatterBegin_Private(mat,&mat->bstash,baij->rangebs);CHKERRQ(ierr); 851 ierr = MatStashGetInfo_Private(&mat->stash,&nstash,&reallocs);CHKERRQ(ierr); 852 ierr = PetscInfo2(mat,"Stash has %D entries,uses %D mallocs.\n",nstash,reallocs);CHKERRQ(ierr); 853 ierr = MatStashGetInfo_Private(&mat->bstash,&nstash,&reallocs);CHKERRQ(ierr); 854 ierr = PetscInfo2(mat,"Block-Stash has %D entries, uses %D mallocs.\n",nstash,reallocs);CHKERRQ(ierr); 855 PetscFunctionReturn(0); 856 } 857 858 #undef __FUNCT__ 859 #define __FUNCT__ "MatAssemblyEnd_MPIBAIJ" 860 PetscErrorCode MatAssemblyEnd_MPIBAIJ(Mat mat,MatAssemblyType mode) 861 { 862 Mat_MPIBAIJ *baij=(Mat_MPIBAIJ*)mat->data; 863 Mat_SeqBAIJ *a=(Mat_SeqBAIJ*)baij->A->data; 864 PetscErrorCode ierr; 865 PetscInt i,j,rstart,ncols,flg,bs2=baij->bs2; 866 PetscInt *row,*col; 867 PetscTruth r1,r2,r3,other_disassembled; 868 MatScalar *val; 869 InsertMode addv = mat->insertmode; 870 PetscMPIInt n; 871 872 /* do not use 'b=(Mat_SeqBAIJ*)baij->B->data' as B can be reset in disassembly */ 873 PetscFunctionBegin; 874 if (!baij->donotstash) { 875 while (1) { 876 ierr = MatStashScatterGetMesg_Private(&mat->stash,&n,&row,&col,&val,&flg);CHKERRQ(ierr); 877 if (!flg) break; 878 879 for (i=0; i<n;) { 880 /* Now identify the consecutive vals belonging to the same row */ 881 for (j=i,rstart=row[j]; j<n; j++) { if (row[j] != rstart) break; } 882 if (j < n) ncols = j-i; 883 else ncols = n-i; 884 /* Now assemble all these values with a single function call */ 885 ierr = MatSetValues_MPIBAIJ(mat,1,row+i,ncols,col+i,val+i,addv);CHKERRQ(ierr); 886 i = j; 887 } 888 } 889 ierr = MatStashScatterEnd_Private(&mat->stash);CHKERRQ(ierr); 890 /* Now process the block-stash. Since the values are stashed column-oriented, 891 set the roworiented flag to column oriented, and after MatSetValues() 892 restore the original flags */ 893 r1 = baij->roworiented; 894 r2 = a->roworiented; 895 r3 = ((Mat_SeqBAIJ*)baij->B->data)->roworiented; 896 baij->roworiented = PETSC_FALSE; 897 a->roworiented = PETSC_FALSE; 898 (((Mat_SeqBAIJ*)baij->B->data))->roworiented = PETSC_FALSE; /* b->roworiented */ 899 while (1) { 900 ierr = MatStashScatterGetMesg_Private(&mat->bstash,&n,&row,&col,&val,&flg);CHKERRQ(ierr); 901 if (!flg) break; 902 903 for (i=0; i<n;) { 904 /* Now identify the consecutive vals belonging to the same row */ 905 for (j=i,rstart=row[j]; j<n; j++) { if (row[j] != rstart) break; } 906 if (j < n) ncols = j-i; 907 else ncols = n-i; 908 ierr = MatSetValuesBlocked_MPIBAIJ(mat,1,row+i,ncols,col+i,val+i*bs2,addv);CHKERRQ(ierr); 909 i = j; 910 } 911 } 912 ierr = MatStashScatterEnd_Private(&mat->bstash);CHKERRQ(ierr); 913 baij->roworiented = r1; 914 a->roworiented = r2; 915 ((Mat_SeqBAIJ*)baij->B->data)->roworiented = r3; /* b->roworiented */ 916 } 917 918 ierr = MatAssemblyBegin(baij->A,mode);CHKERRQ(ierr); 919 ierr = MatAssemblyEnd(baij->A,mode);CHKERRQ(ierr); 920 921 /* determine if any processor has disassembled, if so we must 922 also disassemble ourselfs, in order that we may reassemble. */ 923 /* 924 if nonzero structure of submatrix B cannot change then we know that 925 no processor disassembled thus we can skip this stuff 926 */ 927 if (!((Mat_SeqBAIJ*)baij->B->data)->nonew) { 928 ierr = MPI_Allreduce(&mat->was_assembled,&other_disassembled,1,MPI_INT,MPI_PROD,((PetscObject)mat)->comm);CHKERRQ(ierr); 929 if (mat->was_assembled && !other_disassembled) { 930 ierr = DisAssemble_MPIBAIJ(mat);CHKERRQ(ierr); 931 } 932 } 933 934 if (!mat->was_assembled && mode == MAT_FINAL_ASSEMBLY) { 935 ierr = MatSetUpMultiply_MPIBAIJ(mat);CHKERRQ(ierr); 936 } 937 ((Mat_SeqBAIJ*)baij->B->data)->compressedrow.use = PETSC_TRUE; /* b->compressedrow.use */ 938 ierr = MatAssemblyBegin(baij->B,mode);CHKERRQ(ierr); 939 ierr = MatAssemblyEnd(baij->B,mode);CHKERRQ(ierr); 940 941 #if defined(PETSC_USE_INFO) 942 if (baij->ht && mode== MAT_FINAL_ASSEMBLY) { 943 ierr = PetscInfo1(mat,"Average Hash Table Search in MatSetValues = %5.2f\n",((PetscReal)baij->ht_total_ct)/baij->ht_insert_ct);CHKERRQ(ierr); 944 baij->ht_total_ct = 0; 945 baij->ht_insert_ct = 0; 946 } 947 #endif 948 if (baij->ht_flag && !baij->ht && mode == MAT_FINAL_ASSEMBLY) { 949 ierr = MatCreateHashTable_MPIBAIJ_Private(mat,baij->ht_fact);CHKERRQ(ierr); 950 mat->ops->setvalues = MatSetValues_MPIBAIJ_HT; 951 mat->ops->setvaluesblocked = MatSetValuesBlocked_MPIBAIJ_HT; 952 } 953 954 ierr = PetscFree(baij->rowvalues);CHKERRQ(ierr); 955 baij->rowvalues = 0; 956 PetscFunctionReturn(0); 957 } 958 959 #undef __FUNCT__ 960 #define __FUNCT__ "MatView_MPIBAIJ_ASCIIorDraworSocket" 961 static PetscErrorCode MatView_MPIBAIJ_ASCIIorDraworSocket(Mat mat,PetscViewer viewer) 962 { 963 Mat_MPIBAIJ *baij = (Mat_MPIBAIJ*)mat->data; 964 PetscErrorCode ierr; 965 PetscMPIInt size = baij->size,rank = baij->rank; 966 PetscInt bs = mat->rmap->bs; 967 PetscTruth iascii,isdraw; 968 PetscViewer sviewer; 969 PetscViewerFormat format; 970 971 PetscFunctionBegin; 972 ierr = PetscTypeCompare((PetscObject)viewer,PETSC_VIEWER_ASCII,&iascii);CHKERRQ(ierr); 973 ierr = PetscTypeCompare((PetscObject)viewer,PETSC_VIEWER_DRAW,&isdraw);CHKERRQ(ierr); 974 if (iascii) { 975 ierr = PetscViewerGetFormat(viewer,&format);CHKERRQ(ierr); 976 if (format == PETSC_VIEWER_ASCII_INFO_DETAIL) { 977 MatInfo info; 978 ierr = MPI_Comm_rank(((PetscObject)mat)->comm,&rank);CHKERRQ(ierr); 979 ierr = MatGetInfo(mat,MAT_LOCAL,&info);CHKERRQ(ierr); 980 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"[%d] Local rows %D nz %D nz alloced %D bs %D mem %D\n", 981 rank,mat->rmap->N,(PetscInt)info.nz_used*bs,(PetscInt)info.nz_allocated*bs, 982 mat->rmap->bs,(PetscInt)info.memory);CHKERRQ(ierr); 983 ierr = MatGetInfo(baij->A,MAT_LOCAL,&info);CHKERRQ(ierr); 984 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"[%d] on-diagonal part: nz %D \n",rank,(PetscInt)info.nz_used*bs);CHKERRQ(ierr); 985 ierr = MatGetInfo(baij->B,MAT_LOCAL,&info);CHKERRQ(ierr); 986 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"[%d] off-diagonal part: nz %D \n",rank,(PetscInt)info.nz_used*bs);CHKERRQ(ierr); 987 ierr = PetscViewerFlush(viewer);CHKERRQ(ierr); 988 ierr = PetscViewerASCIIPrintf(viewer,"Information on VecScatter used in matrix-vector product: \n");CHKERRQ(ierr); 989 ierr = VecScatterView(baij->Mvctx,viewer);CHKERRQ(ierr); 990 PetscFunctionReturn(0); 991 } else if (format == PETSC_VIEWER_ASCII_INFO) { 992 ierr = PetscViewerASCIIPrintf(viewer," block size is %D\n",bs);CHKERRQ(ierr); 993 PetscFunctionReturn(0); 994 } else if (format == PETSC_VIEWER_ASCII_FACTOR_INFO) { 995 PetscFunctionReturn(0); 996 } 997 } 998 999 if (isdraw) { 1000 PetscDraw draw; 1001 PetscTruth isnull; 1002 ierr = PetscViewerDrawGetDraw(viewer,0,&draw);CHKERRQ(ierr); 1003 ierr = PetscDrawIsNull(draw,&isnull);CHKERRQ(ierr); if (isnull) PetscFunctionReturn(0); 1004 } 1005 1006 if (size == 1) { 1007 ierr = PetscObjectSetName((PetscObject)baij->A,((PetscObject)mat)->name);CHKERRQ(ierr); 1008 ierr = MatView(baij->A,viewer);CHKERRQ(ierr); 1009 } else { 1010 /* assemble the entire matrix onto first processor. */ 1011 Mat A; 1012 Mat_SeqBAIJ *Aloc; 1013 PetscInt M = mat->rmap->N,N = mat->cmap->N,*ai,*aj,col,i,j,k,*rvals,mbs = baij->mbs; 1014 MatScalar *a; 1015 1016 /* Here we are creating a temporary matrix, so will assume MPIBAIJ is acceptable */ 1017 /* Perhaps this should be the type of mat? */ 1018 ierr = MatCreate(((PetscObject)mat)->comm,&A);CHKERRQ(ierr); 1019 if (!rank) { 1020 ierr = MatSetSizes(A,M,N,M,N);CHKERRQ(ierr); 1021 } else { 1022 ierr = MatSetSizes(A,0,0,M,N);CHKERRQ(ierr); 1023 } 1024 ierr = MatSetType(A,MATMPIBAIJ);CHKERRQ(ierr); 1025 ierr = MatMPIBAIJSetPreallocation(A,mat->rmap->bs,0,PETSC_NULL,0,PETSC_NULL);CHKERRQ(ierr); 1026 ierr = PetscLogObjectParent(mat,A);CHKERRQ(ierr); 1027 1028 /* copy over the A part */ 1029 Aloc = (Mat_SeqBAIJ*)baij->A->data; 1030 ai = Aloc->i; aj = Aloc->j; a = Aloc->a; 1031 ierr = PetscMalloc(bs*sizeof(PetscInt),&rvals);CHKERRQ(ierr); 1032 1033 for (i=0; i<mbs; i++) { 1034 rvals[0] = bs*(baij->rstartbs + i); 1035 for (j=1; j<bs; j++) { rvals[j] = rvals[j-1] + 1; } 1036 for (j=ai[i]; j<ai[i+1]; j++) { 1037 col = (baij->cstartbs+aj[j])*bs; 1038 for (k=0; k<bs; k++) { 1039 ierr = MatSetValues_MPIBAIJ(A,bs,rvals,1,&col,a,INSERT_VALUES);CHKERRQ(ierr); 1040 col++; a += bs; 1041 } 1042 } 1043 } 1044 /* copy over the B part */ 1045 Aloc = (Mat_SeqBAIJ*)baij->B->data; 1046 ai = Aloc->i; aj = Aloc->j; a = Aloc->a; 1047 for (i=0; i<mbs; i++) { 1048 rvals[0] = bs*(baij->rstartbs + i); 1049 for (j=1; j<bs; j++) { rvals[j] = rvals[j-1] + 1; } 1050 for (j=ai[i]; j<ai[i+1]; j++) { 1051 col = baij->garray[aj[j]]*bs; 1052 for (k=0; k<bs; k++) { 1053 ierr = MatSetValues_MPIBAIJ(A,bs,rvals,1,&col,a,INSERT_VALUES);CHKERRQ(ierr); 1054 col++; a += bs; 1055 } 1056 } 1057 } 1058 ierr = PetscFree(rvals);CHKERRQ(ierr); 1059 ierr = MatAssemblyBegin(A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1060 ierr = MatAssemblyEnd(A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1061 /* 1062 Everyone has to call to draw the matrix since the graphics waits are 1063 synchronized across all processors that share the PetscDraw object 1064 */ 1065 ierr = PetscViewerGetSingleton(viewer,&sviewer);CHKERRQ(ierr); 1066 if (!rank) { 1067 ierr = PetscObjectSetName((PetscObject)((Mat_MPIBAIJ*)(A->data))->A,((PetscObject)mat)->name);CHKERRQ(ierr); 1068 ierr = MatView(((Mat_MPIBAIJ*)(A->data))->A,sviewer);CHKERRQ(ierr); 1069 } 1070 ierr = PetscViewerRestoreSingleton(viewer,&sviewer);CHKERRQ(ierr); 1071 ierr = MatDestroy(A);CHKERRQ(ierr); 1072 } 1073 PetscFunctionReturn(0); 1074 } 1075 1076 #undef __FUNCT__ 1077 #define __FUNCT__ "MatView_MPIBAIJ" 1078 PetscErrorCode MatView_MPIBAIJ(Mat mat,PetscViewer viewer) 1079 { 1080 PetscErrorCode ierr; 1081 PetscTruth iascii,isdraw,issocket,isbinary; 1082 1083 PetscFunctionBegin; 1084 ierr = PetscTypeCompare((PetscObject)viewer,PETSC_VIEWER_ASCII,&iascii);CHKERRQ(ierr); 1085 ierr = PetscTypeCompare((PetscObject)viewer,PETSC_VIEWER_DRAW,&isdraw);CHKERRQ(ierr); 1086 ierr = PetscTypeCompare((PetscObject)viewer,PETSC_VIEWER_SOCKET,&issocket);CHKERRQ(ierr); 1087 ierr = PetscTypeCompare((PetscObject)viewer,PETSC_VIEWER_BINARY,&isbinary);CHKERRQ(ierr); 1088 if (iascii || isdraw || issocket || isbinary) { 1089 ierr = MatView_MPIBAIJ_ASCIIorDraworSocket(mat,viewer);CHKERRQ(ierr); 1090 } else { 1091 SETERRQ1(PETSC_ERR_SUP,"Viewer type %s not supported by MPIBAIJ matrices",((PetscObject)viewer)->type_name); 1092 } 1093 PetscFunctionReturn(0); 1094 } 1095 1096 #undef __FUNCT__ 1097 #define __FUNCT__ "MatDestroy_MPIBAIJ" 1098 PetscErrorCode MatDestroy_MPIBAIJ(Mat mat) 1099 { 1100 Mat_MPIBAIJ *baij = (Mat_MPIBAIJ*)mat->data; 1101 PetscErrorCode ierr; 1102 1103 PetscFunctionBegin; 1104 #if defined(PETSC_USE_LOG) 1105 PetscLogObjectState((PetscObject)mat,"Rows=%D,Cols=%D",mat->rmap->N,mat->cmap->N); 1106 #endif 1107 ierr = MatStashDestroy_Private(&mat->stash);CHKERRQ(ierr); 1108 ierr = MatStashDestroy_Private(&mat->bstash);CHKERRQ(ierr); 1109 ierr = MatDestroy(baij->A);CHKERRQ(ierr); 1110 ierr = MatDestroy(baij->B);CHKERRQ(ierr); 1111 #if defined (PETSC_USE_CTABLE) 1112 if (baij->colmap) {ierr = PetscTableDestroy(baij->colmap);CHKERRQ(ierr);} 1113 #else 1114 ierr = PetscFree(baij->colmap);CHKERRQ(ierr); 1115 #endif 1116 ierr = PetscFree(baij->garray);CHKERRQ(ierr); 1117 if (baij->lvec) {ierr = VecDestroy(baij->lvec);CHKERRQ(ierr);} 1118 if (baij->Mvctx) {ierr = VecScatterDestroy(baij->Mvctx);CHKERRQ(ierr);} 1119 ierr = PetscFree(baij->rowvalues);CHKERRQ(ierr); 1120 ierr = PetscFree(baij->barray);CHKERRQ(ierr); 1121 ierr = PetscFree(baij->hd);CHKERRQ(ierr); 1122 ierr = PetscFree(baij->rangebs);CHKERRQ(ierr); 1123 ierr = PetscFree(baij);CHKERRQ(ierr); 1124 1125 ierr = PetscObjectChangeTypeName((PetscObject)mat,0);CHKERRQ(ierr); 1126 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatStoreValues_C","",PETSC_NULL);CHKERRQ(ierr); 1127 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatRetrieveValues_C","",PETSC_NULL);CHKERRQ(ierr); 1128 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatGetDiagonalBlock_C","",PETSC_NULL);CHKERRQ(ierr); 1129 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatMPIBAIJSetPreallocation_C","",PETSC_NULL);CHKERRQ(ierr); 1130 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatMPIBAIJSetPreallocationCSR_C","",PETSC_NULL);CHKERRQ(ierr); 1131 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatDiagonalScaleLocal_C","",PETSC_NULL);CHKERRQ(ierr); 1132 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatSetHashTableFactor_C","",PETSC_NULL);CHKERRQ(ierr); 1133 PetscFunctionReturn(0); 1134 } 1135 1136 #undef __FUNCT__ 1137 #define __FUNCT__ "MatMult_MPIBAIJ" 1138 PetscErrorCode MatMult_MPIBAIJ(Mat A,Vec xx,Vec yy) 1139 { 1140 Mat_MPIBAIJ *a = (Mat_MPIBAIJ*)A->data; 1141 PetscErrorCode ierr; 1142 PetscInt nt; 1143 1144 PetscFunctionBegin; 1145 ierr = VecGetLocalSize(xx,&nt);CHKERRQ(ierr); 1146 if (nt != A->cmap->n) { 1147 SETERRQ(PETSC_ERR_ARG_SIZ,"Incompatible partition of A and xx"); 1148 } 1149 ierr = VecGetLocalSize(yy,&nt);CHKERRQ(ierr); 1150 if (nt != A->rmap->n) { 1151 SETERRQ(PETSC_ERR_ARG_SIZ,"Incompatible parition of A and yy"); 1152 } 1153 ierr = VecScatterBegin(a->Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1154 ierr = (*a->A->ops->mult)(a->A,xx,yy);CHKERRQ(ierr); 1155 ierr = VecScatterEnd(a->Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1156 ierr = (*a->B->ops->multadd)(a->B,a->lvec,yy,yy);CHKERRQ(ierr); 1157 PetscFunctionReturn(0); 1158 } 1159 1160 #undef __FUNCT__ 1161 #define __FUNCT__ "MatMultAdd_MPIBAIJ" 1162 PetscErrorCode MatMultAdd_MPIBAIJ(Mat A,Vec xx,Vec yy,Vec zz) 1163 { 1164 Mat_MPIBAIJ *a = (Mat_MPIBAIJ*)A->data; 1165 PetscErrorCode ierr; 1166 1167 PetscFunctionBegin; 1168 ierr = VecScatterBegin(a->Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1169 ierr = (*a->A->ops->multadd)(a->A,xx,yy,zz);CHKERRQ(ierr); 1170 ierr = VecScatterEnd(a->Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1171 ierr = (*a->B->ops->multadd)(a->B,a->lvec,zz,zz);CHKERRQ(ierr); 1172 PetscFunctionReturn(0); 1173 } 1174 1175 #undef __FUNCT__ 1176 #define __FUNCT__ "MatMultTranspose_MPIBAIJ" 1177 PetscErrorCode MatMultTranspose_MPIBAIJ(Mat A,Vec xx,Vec yy) 1178 { 1179 Mat_MPIBAIJ *a = (Mat_MPIBAIJ*)A->data; 1180 PetscErrorCode ierr; 1181 PetscTruth merged; 1182 1183 PetscFunctionBegin; 1184 ierr = VecScatterGetMerged(a->Mvctx,&merged);CHKERRQ(ierr); 1185 /* do nondiagonal part */ 1186 ierr = (*a->B->ops->multtranspose)(a->B,xx,a->lvec);CHKERRQ(ierr); 1187 if (!merged) { 1188 /* send it on its way */ 1189 ierr = VecScatterBegin(a->Mvctx,a->lvec,yy,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1190 /* do local part */ 1191 ierr = (*a->A->ops->multtranspose)(a->A,xx,yy);CHKERRQ(ierr); 1192 /* receive remote parts: note this assumes the values are not actually */ 1193 /* inserted in yy until the next line */ 1194 ierr = VecScatterEnd(a->Mvctx,a->lvec,yy,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1195 } else { 1196 /* do local part */ 1197 ierr = (*a->A->ops->multtranspose)(a->A,xx,yy);CHKERRQ(ierr); 1198 /* send it on its way */ 1199 ierr = VecScatterBegin(a->Mvctx,a->lvec,yy,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1200 /* values actually were received in the Begin() but we need to call this nop */ 1201 ierr = VecScatterEnd(a->Mvctx,a->lvec,yy,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1202 } 1203 PetscFunctionReturn(0); 1204 } 1205 1206 #undef __FUNCT__ 1207 #define __FUNCT__ "MatMultTransposeAdd_MPIBAIJ" 1208 PetscErrorCode MatMultTransposeAdd_MPIBAIJ(Mat A,Vec xx,Vec yy,Vec zz) 1209 { 1210 Mat_MPIBAIJ *a = (Mat_MPIBAIJ*)A->data; 1211 PetscErrorCode ierr; 1212 1213 PetscFunctionBegin; 1214 /* do nondiagonal part */ 1215 ierr = (*a->B->ops->multtranspose)(a->B,xx,a->lvec);CHKERRQ(ierr); 1216 /* send it on its way */ 1217 ierr = VecScatterBegin(a->Mvctx,a->lvec,zz,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1218 /* do local part */ 1219 ierr = (*a->A->ops->multtransposeadd)(a->A,xx,yy,zz);CHKERRQ(ierr); 1220 /* receive remote parts: note this assumes the values are not actually */ 1221 /* inserted in yy until the next line, which is true for my implementation*/ 1222 /* but is not perhaps always true. */ 1223 ierr = VecScatterEnd(a->Mvctx,a->lvec,zz,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1224 PetscFunctionReturn(0); 1225 } 1226 1227 /* 1228 This only works correctly for square matrices where the subblock A->A is the 1229 diagonal block 1230 */ 1231 #undef __FUNCT__ 1232 #define __FUNCT__ "MatGetDiagonal_MPIBAIJ" 1233 PetscErrorCode MatGetDiagonal_MPIBAIJ(Mat A,Vec v) 1234 { 1235 Mat_MPIBAIJ *a = (Mat_MPIBAIJ*)A->data; 1236 PetscErrorCode ierr; 1237 1238 PetscFunctionBegin; 1239 if (A->rmap->N != A->cmap->N) SETERRQ(PETSC_ERR_SUP,"Supports only square matrix where A->A is diag block"); 1240 ierr = MatGetDiagonal(a->A,v);CHKERRQ(ierr); 1241 PetscFunctionReturn(0); 1242 } 1243 1244 #undef __FUNCT__ 1245 #define __FUNCT__ "MatScale_MPIBAIJ" 1246 PetscErrorCode MatScale_MPIBAIJ(Mat A,PetscScalar aa) 1247 { 1248 Mat_MPIBAIJ *a = (Mat_MPIBAIJ*)A->data; 1249 PetscErrorCode ierr; 1250 1251 PetscFunctionBegin; 1252 ierr = MatScale(a->A,aa);CHKERRQ(ierr); 1253 ierr = MatScale(a->B,aa);CHKERRQ(ierr); 1254 PetscFunctionReturn(0); 1255 } 1256 1257 #undef __FUNCT__ 1258 #define __FUNCT__ "MatGetRow_MPIBAIJ" 1259 PetscErrorCode MatGetRow_MPIBAIJ(Mat matin,PetscInt row,PetscInt *nz,PetscInt **idx,PetscScalar **v) 1260 { 1261 Mat_MPIBAIJ *mat = (Mat_MPIBAIJ*)matin->data; 1262 PetscScalar *vworkA,*vworkB,**pvA,**pvB,*v_p; 1263 PetscErrorCode ierr; 1264 PetscInt bs = matin->rmap->bs,bs2 = mat->bs2,i,*cworkA,*cworkB,**pcA,**pcB; 1265 PetscInt nztot,nzA,nzB,lrow,brstart = matin->rmap->rstart,brend = matin->rmap->rend; 1266 PetscInt *cmap,*idx_p,cstart = mat->cstartbs; 1267 1268 PetscFunctionBegin; 1269 if (mat->getrowactive) SETERRQ(PETSC_ERR_ARG_WRONGSTATE,"Already active"); 1270 mat->getrowactive = PETSC_TRUE; 1271 1272 if (!mat->rowvalues && (idx || v)) { 1273 /* 1274 allocate enough space to hold information from the longest row. 1275 */ 1276 Mat_SeqBAIJ *Aa = (Mat_SeqBAIJ*)mat->A->data,*Ba = (Mat_SeqBAIJ*)mat->B->data; 1277 PetscInt max = 1,mbs = mat->mbs,tmp; 1278 for (i=0; i<mbs; i++) { 1279 tmp = Aa->i[i+1] - Aa->i[i] + Ba->i[i+1] - Ba->i[i]; 1280 if (max < tmp) { max = tmp; } 1281 } 1282 ierr = PetscMalloc(max*bs2*(sizeof(PetscInt)+sizeof(PetscScalar)),&mat->rowvalues);CHKERRQ(ierr); 1283 mat->rowindices = (PetscInt*)(mat->rowvalues + max*bs2); 1284 } 1285 1286 if (row < brstart || row >= brend) SETERRQ(PETSC_ERR_SUP,"Only local rows") 1287 lrow = row - brstart; 1288 1289 pvA = &vworkA; pcA = &cworkA; pvB = &vworkB; pcB = &cworkB; 1290 if (!v) {pvA = 0; pvB = 0;} 1291 if (!idx) {pcA = 0; if (!v) pcB = 0;} 1292 ierr = (*mat->A->ops->getrow)(mat->A,lrow,&nzA,pcA,pvA);CHKERRQ(ierr); 1293 ierr = (*mat->B->ops->getrow)(mat->B,lrow,&nzB,pcB,pvB);CHKERRQ(ierr); 1294 nztot = nzA + nzB; 1295 1296 cmap = mat->garray; 1297 if (v || idx) { 1298 if (nztot) { 1299 /* Sort by increasing column numbers, assuming A and B already sorted */ 1300 PetscInt imark = -1; 1301 if (v) { 1302 *v = v_p = mat->rowvalues; 1303 for (i=0; i<nzB; i++) { 1304 if (cmap[cworkB[i]/bs] < cstart) v_p[i] = vworkB[i]; 1305 else break; 1306 } 1307 imark = i; 1308 for (i=0; i<nzA; i++) v_p[imark+i] = vworkA[i]; 1309 for (i=imark; i<nzB; i++) v_p[nzA+i] = vworkB[i]; 1310 } 1311 if (idx) { 1312 *idx = idx_p = mat->rowindices; 1313 if (imark > -1) { 1314 for (i=0; i<imark; i++) { 1315 idx_p[i] = cmap[cworkB[i]/bs]*bs + cworkB[i]%bs; 1316 } 1317 } else { 1318 for (i=0; i<nzB; i++) { 1319 if (cmap[cworkB[i]/bs] < cstart) 1320 idx_p[i] = cmap[cworkB[i]/bs]*bs + cworkB[i]%bs ; 1321 else break; 1322 } 1323 imark = i; 1324 } 1325 for (i=0; i<nzA; i++) idx_p[imark+i] = cstart*bs + cworkA[i]; 1326 for (i=imark; i<nzB; i++) idx_p[nzA+i] = cmap[cworkB[i]/bs]*bs + cworkB[i]%bs ; 1327 } 1328 } else { 1329 if (idx) *idx = 0; 1330 if (v) *v = 0; 1331 } 1332 } 1333 *nz = nztot; 1334 ierr = (*mat->A->ops->restorerow)(mat->A,lrow,&nzA,pcA,pvA);CHKERRQ(ierr); 1335 ierr = (*mat->B->ops->restorerow)(mat->B,lrow,&nzB,pcB,pvB);CHKERRQ(ierr); 1336 PetscFunctionReturn(0); 1337 } 1338 1339 #undef __FUNCT__ 1340 #define __FUNCT__ "MatRestoreRow_MPIBAIJ" 1341 PetscErrorCode MatRestoreRow_MPIBAIJ(Mat mat,PetscInt row,PetscInt *nz,PetscInt **idx,PetscScalar **v) 1342 { 1343 Mat_MPIBAIJ *baij = (Mat_MPIBAIJ*)mat->data; 1344 1345 PetscFunctionBegin; 1346 if (!baij->getrowactive) { 1347 SETERRQ(PETSC_ERR_ARG_WRONGSTATE,"MatGetRow not called"); 1348 } 1349 baij->getrowactive = PETSC_FALSE; 1350 PetscFunctionReturn(0); 1351 } 1352 1353 #undef __FUNCT__ 1354 #define __FUNCT__ "MatZeroEntries_MPIBAIJ" 1355 PetscErrorCode MatZeroEntries_MPIBAIJ(Mat A) 1356 { 1357 Mat_MPIBAIJ *l = (Mat_MPIBAIJ*)A->data; 1358 PetscErrorCode ierr; 1359 1360 PetscFunctionBegin; 1361 ierr = MatZeroEntries(l->A);CHKERRQ(ierr); 1362 ierr = MatZeroEntries(l->B);CHKERRQ(ierr); 1363 PetscFunctionReturn(0); 1364 } 1365 1366 #undef __FUNCT__ 1367 #define __FUNCT__ "MatGetInfo_MPIBAIJ" 1368 PetscErrorCode MatGetInfo_MPIBAIJ(Mat matin,MatInfoType flag,MatInfo *info) 1369 { 1370 Mat_MPIBAIJ *a = (Mat_MPIBAIJ*)matin->data; 1371 Mat A = a->A,B = a->B; 1372 PetscErrorCode ierr; 1373 PetscReal isend[5],irecv[5]; 1374 1375 PetscFunctionBegin; 1376 info->block_size = (PetscReal)matin->rmap->bs; 1377 ierr = MatGetInfo(A,MAT_LOCAL,info);CHKERRQ(ierr); 1378 isend[0] = info->nz_used; isend[1] = info->nz_allocated; isend[2] = info->nz_unneeded; 1379 isend[3] = info->memory; isend[4] = info->mallocs; 1380 ierr = MatGetInfo(B,MAT_LOCAL,info);CHKERRQ(ierr); 1381 isend[0] += info->nz_used; isend[1] += info->nz_allocated; isend[2] += info->nz_unneeded; 1382 isend[3] += info->memory; isend[4] += info->mallocs; 1383 if (flag == MAT_LOCAL) { 1384 info->nz_used = isend[0]; 1385 info->nz_allocated = isend[1]; 1386 info->nz_unneeded = isend[2]; 1387 info->memory = isend[3]; 1388 info->mallocs = isend[4]; 1389 } else if (flag == MAT_GLOBAL_MAX) { 1390 ierr = MPI_Allreduce(isend,irecv,5,MPIU_REAL,MPI_MAX,((PetscObject)matin)->comm);CHKERRQ(ierr); 1391 info->nz_used = irecv[0]; 1392 info->nz_allocated = irecv[1]; 1393 info->nz_unneeded = irecv[2]; 1394 info->memory = irecv[3]; 1395 info->mallocs = irecv[4]; 1396 } else if (flag == MAT_GLOBAL_SUM) { 1397 ierr = MPI_Allreduce(isend,irecv,5,MPIU_REAL,MPI_SUM,((PetscObject)matin)->comm);CHKERRQ(ierr); 1398 info->nz_used = irecv[0]; 1399 info->nz_allocated = irecv[1]; 1400 info->nz_unneeded = irecv[2]; 1401 info->memory = irecv[3]; 1402 info->mallocs = irecv[4]; 1403 } else { 1404 SETERRQ1(PETSC_ERR_ARG_WRONG,"Unknown MatInfoType argument %d",(int)flag); 1405 } 1406 info->fill_ratio_given = 0; /* no parallel LU/ILU/Cholesky */ 1407 info->fill_ratio_needed = 0; 1408 info->factor_mallocs = 0; 1409 PetscFunctionReturn(0); 1410 } 1411 1412 #undef __FUNCT__ 1413 #define __FUNCT__ "MatSetOption_MPIBAIJ" 1414 PetscErrorCode MatSetOption_MPIBAIJ(Mat A,MatOption op,PetscTruth flg) 1415 { 1416 Mat_MPIBAIJ *a = (Mat_MPIBAIJ*)A->data; 1417 PetscErrorCode ierr; 1418 1419 PetscFunctionBegin; 1420 switch (op) { 1421 case MAT_NEW_NONZERO_LOCATIONS: 1422 case MAT_NEW_NONZERO_ALLOCATION_ERR: 1423 case MAT_UNUSED_NONZERO_LOCATION_ERR: 1424 case MAT_KEEP_ZEROED_ROWS: 1425 case MAT_NEW_NONZERO_LOCATION_ERR: 1426 ierr = MatSetOption(a->A,op,flg);CHKERRQ(ierr); 1427 ierr = MatSetOption(a->B,op,flg);CHKERRQ(ierr); 1428 break; 1429 case MAT_ROW_ORIENTED: 1430 a->roworiented = flg; 1431 ierr = MatSetOption(a->A,op,flg);CHKERRQ(ierr); 1432 ierr = MatSetOption(a->B,op,flg);CHKERRQ(ierr); 1433 break; 1434 case MAT_NEW_DIAGONALS: 1435 ierr = PetscInfo1(A,"Option %s ignored\n",MatOptions[op]);CHKERRQ(ierr); 1436 break; 1437 case MAT_IGNORE_OFF_PROC_ENTRIES: 1438 a->donotstash = flg; 1439 break; 1440 case MAT_USE_HASH_TABLE: 1441 a->ht_flag = flg; 1442 break; 1443 case MAT_SYMMETRIC: 1444 case MAT_STRUCTURALLY_SYMMETRIC: 1445 case MAT_HERMITIAN: 1446 case MAT_SYMMETRY_ETERNAL: 1447 ierr = MatSetOption(a->A,op,flg);CHKERRQ(ierr); 1448 break; 1449 default: 1450 SETERRQ1(PETSC_ERR_SUP,"unknown option %d",op); 1451 } 1452 PetscFunctionReturn(0); 1453 } 1454 1455 #undef __FUNCT__ 1456 #define __FUNCT__ "MatTranspose_MPIBAIJ(" 1457 PetscErrorCode MatTranspose_MPIBAIJ(Mat A,MatReuse reuse,Mat *matout) 1458 { 1459 Mat_MPIBAIJ *baij = (Mat_MPIBAIJ*)A->data; 1460 Mat_SeqBAIJ *Aloc; 1461 Mat B; 1462 PetscErrorCode ierr; 1463 PetscInt M=A->rmap->N,N=A->cmap->N,*ai,*aj,i,*rvals,j,k,col; 1464 PetscInt bs=A->rmap->bs,mbs=baij->mbs; 1465 MatScalar *a; 1466 1467 PetscFunctionBegin; 1468 if (reuse == MAT_REUSE_MATRIX && A == *matout && M != N) SETERRQ(PETSC_ERR_ARG_SIZ,"Square matrix only for in-place"); 1469 if (reuse == MAT_INITIAL_MATRIX || *matout == A) { 1470 ierr = MatCreate(((PetscObject)A)->comm,&B);CHKERRQ(ierr); 1471 ierr = MatSetSizes(B,A->cmap->n,A->rmap->n,N,M);CHKERRQ(ierr); 1472 ierr = MatSetType(B,((PetscObject)A)->type_name);CHKERRQ(ierr); 1473 ierr = MatMPIBAIJSetPreallocation(B,A->rmap->bs,0,PETSC_NULL,0,PETSC_NULL);CHKERRQ(ierr); 1474 } else { 1475 B = *matout; 1476 } 1477 1478 /* copy over the A part */ 1479 Aloc = (Mat_SeqBAIJ*)baij->A->data; 1480 ai = Aloc->i; aj = Aloc->j; a = Aloc->a; 1481 ierr = PetscMalloc(bs*sizeof(PetscInt),&rvals);CHKERRQ(ierr); 1482 1483 for (i=0; i<mbs; i++) { 1484 rvals[0] = bs*(baij->rstartbs + i); 1485 for (j=1; j<bs; j++) { rvals[j] = rvals[j-1] + 1; } 1486 for (j=ai[i]; j<ai[i+1]; j++) { 1487 col = (baij->cstartbs+aj[j])*bs; 1488 for (k=0; k<bs; k++) { 1489 ierr = MatSetValues_MPIBAIJ(B,1,&col,bs,rvals,a,INSERT_VALUES);CHKERRQ(ierr); 1490 col++; a += bs; 1491 } 1492 } 1493 } 1494 /* copy over the B part */ 1495 Aloc = (Mat_SeqBAIJ*)baij->B->data; 1496 ai = Aloc->i; aj = Aloc->j; a = Aloc->a; 1497 for (i=0; i<mbs; i++) { 1498 rvals[0] = bs*(baij->rstartbs + i); 1499 for (j=1; j<bs; j++) { rvals[j] = rvals[j-1] + 1; } 1500 for (j=ai[i]; j<ai[i+1]; j++) { 1501 col = baij->garray[aj[j]]*bs; 1502 for (k=0; k<bs; k++) { 1503 ierr = MatSetValues_MPIBAIJ(B,1,&col,bs,rvals,a,INSERT_VALUES);CHKERRQ(ierr); 1504 col++; a += bs; 1505 } 1506 } 1507 } 1508 ierr = PetscFree(rvals);CHKERRQ(ierr); 1509 ierr = MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1510 ierr = MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1511 1512 if (reuse == MAT_INITIAL_MATRIX || *matout != A) { 1513 *matout = B; 1514 } else { 1515 ierr = MatHeaderCopy(A,B);CHKERRQ(ierr); 1516 } 1517 PetscFunctionReturn(0); 1518 } 1519 1520 #undef __FUNCT__ 1521 #define __FUNCT__ "MatDiagonalScale_MPIBAIJ" 1522 PetscErrorCode MatDiagonalScale_MPIBAIJ(Mat mat,Vec ll,Vec rr) 1523 { 1524 Mat_MPIBAIJ *baij = (Mat_MPIBAIJ*)mat->data; 1525 Mat a = baij->A,b = baij->B; 1526 PetscErrorCode ierr; 1527 PetscInt s1,s2,s3; 1528 1529 PetscFunctionBegin; 1530 ierr = MatGetLocalSize(mat,&s2,&s3);CHKERRQ(ierr); 1531 if (rr) { 1532 ierr = VecGetLocalSize(rr,&s1);CHKERRQ(ierr); 1533 if (s1!=s3) SETERRQ(PETSC_ERR_ARG_SIZ,"right vector non-conforming local size"); 1534 /* Overlap communication with computation. */ 1535 ierr = VecScatterBegin(baij->Mvctx,rr,baij->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1536 } 1537 if (ll) { 1538 ierr = VecGetLocalSize(ll,&s1);CHKERRQ(ierr); 1539 if (s1!=s2) SETERRQ(PETSC_ERR_ARG_SIZ,"left vector non-conforming local size"); 1540 ierr = (*b->ops->diagonalscale)(b,ll,PETSC_NULL);CHKERRQ(ierr); 1541 } 1542 /* scale the diagonal block */ 1543 ierr = (*a->ops->diagonalscale)(a,ll,rr);CHKERRQ(ierr); 1544 1545 if (rr) { 1546 /* Do a scatter end and then right scale the off-diagonal block */ 1547 ierr = VecScatterEnd(baij->Mvctx,rr,baij->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1548 ierr = (*b->ops->diagonalscale)(b,PETSC_NULL,baij->lvec);CHKERRQ(ierr); 1549 } 1550 1551 PetscFunctionReturn(0); 1552 } 1553 1554 #undef __FUNCT__ 1555 #define __FUNCT__ "MatZeroRows_MPIBAIJ" 1556 PetscErrorCode MatZeroRows_MPIBAIJ(Mat A,PetscInt N,const PetscInt rows[],PetscScalar diag) 1557 { 1558 Mat_MPIBAIJ *l = (Mat_MPIBAIJ*)A->data; 1559 PetscErrorCode ierr; 1560 PetscMPIInt imdex,size = l->size,n,rank = l->rank; 1561 PetscInt i,*owners = A->rmap->range; 1562 PetscInt *nprocs,j,idx,nsends,row; 1563 PetscInt nmax,*svalues,*starts,*owner,nrecvs; 1564 PetscInt *rvalues,tag = ((PetscObject)A)->tag,count,base,slen,*source,lastidx = -1; 1565 PetscInt *lens,*lrows,*values,rstart_bs=A->rmap->rstart; 1566 MPI_Comm comm = ((PetscObject)A)->comm; 1567 MPI_Request *send_waits,*recv_waits; 1568 MPI_Status recv_status,*send_status; 1569 #if defined(PETSC_DEBUG) 1570 PetscTruth found = PETSC_FALSE; 1571 #endif 1572 1573 PetscFunctionBegin; 1574 /* first count number of contributors to each processor */ 1575 ierr = PetscMalloc(2*size*sizeof(PetscInt),&nprocs);CHKERRQ(ierr); 1576 ierr = PetscMemzero(nprocs,2*size*sizeof(PetscInt));CHKERRQ(ierr); 1577 ierr = PetscMalloc((N+1)*sizeof(PetscInt),&owner);CHKERRQ(ierr); /* see note*/ 1578 j = 0; 1579 for (i=0; i<N; i++) { 1580 if (lastidx > (idx = rows[i])) j = 0; 1581 lastidx = idx; 1582 for (; j<size; j++) { 1583 if (idx >= owners[j] && idx < owners[j+1]) { 1584 nprocs[2*j]++; 1585 nprocs[2*j+1] = 1; 1586 owner[i] = j; 1587 #if defined(PETSC_DEBUG) 1588 found = PETSC_TRUE; 1589 #endif 1590 break; 1591 } 1592 } 1593 #if defined(PETSC_DEBUG) 1594 if (!found) SETERRQ(PETSC_ERR_ARG_OUTOFRANGE,"Index out of range"); 1595 found = PETSC_FALSE; 1596 #endif 1597 } 1598 nsends = 0; for (i=0; i<size; i++) { nsends += nprocs[2*i+1];} 1599 1600 /* inform other processors of number of messages and max length*/ 1601 ierr = PetscMaxSum(comm,nprocs,&nmax,&nrecvs);CHKERRQ(ierr); 1602 1603 /* post receives: */ 1604 ierr = PetscMalloc((nrecvs+1)*(nmax+1)*sizeof(PetscInt),&rvalues);CHKERRQ(ierr); 1605 ierr = PetscMalloc((nrecvs+1)*sizeof(MPI_Request),&recv_waits);CHKERRQ(ierr); 1606 for (i=0; i<nrecvs; i++) { 1607 ierr = MPI_Irecv(rvalues+nmax*i,nmax,MPIU_INT,MPI_ANY_SOURCE,tag,comm,recv_waits+i);CHKERRQ(ierr); 1608 } 1609 1610 /* do sends: 1611 1) starts[i] gives the starting index in svalues for stuff going to 1612 the ith processor 1613 */ 1614 ierr = PetscMalloc((N+1)*sizeof(PetscInt),&svalues);CHKERRQ(ierr); 1615 ierr = PetscMalloc((nsends+1)*sizeof(MPI_Request),&send_waits);CHKERRQ(ierr); 1616 ierr = PetscMalloc((size+1)*sizeof(PetscInt),&starts);CHKERRQ(ierr); 1617 starts[0] = 0; 1618 for (i=1; i<size; i++) { starts[i] = starts[i-1] + nprocs[2*i-2];} 1619 for (i=0; i<N; i++) { 1620 svalues[starts[owner[i]]++] = rows[i]; 1621 } 1622 1623 starts[0] = 0; 1624 for (i=1; i<size+1; i++) { starts[i] = starts[i-1] + nprocs[2*i-2];} 1625 count = 0; 1626 for (i=0; i<size; i++) { 1627 if (nprocs[2*i+1]) { 1628 ierr = MPI_Isend(svalues+starts[i],nprocs[2*i],MPIU_INT,i,tag,comm,send_waits+count++);CHKERRQ(ierr); 1629 } 1630 } 1631 ierr = PetscFree(starts);CHKERRQ(ierr); 1632 1633 base = owners[rank]; 1634 1635 /* wait on receives */ 1636 ierr = PetscMalloc(2*(nrecvs+1)*sizeof(PetscInt),&lens);CHKERRQ(ierr); 1637 source = lens + nrecvs; 1638 count = nrecvs; slen = 0; 1639 while (count) { 1640 ierr = MPI_Waitany(nrecvs,recv_waits,&imdex,&recv_status);CHKERRQ(ierr); 1641 /* unpack receives into our local space */ 1642 ierr = MPI_Get_count(&recv_status,MPIU_INT,&n);CHKERRQ(ierr); 1643 source[imdex] = recv_status.MPI_SOURCE; 1644 lens[imdex] = n; 1645 slen += n; 1646 count--; 1647 } 1648 ierr = PetscFree(recv_waits);CHKERRQ(ierr); 1649 1650 /* move the data into the send scatter */ 1651 ierr = PetscMalloc((slen+1)*sizeof(PetscInt),&lrows);CHKERRQ(ierr); 1652 count = 0; 1653 for (i=0; i<nrecvs; i++) { 1654 values = rvalues + i*nmax; 1655 for (j=0; j<lens[i]; j++) { 1656 lrows[count++] = values[j] - base; 1657 } 1658 } 1659 ierr = PetscFree(rvalues);CHKERRQ(ierr); 1660 ierr = PetscFree(lens);CHKERRQ(ierr); 1661 ierr = PetscFree(owner);CHKERRQ(ierr); 1662 ierr = PetscFree(nprocs);CHKERRQ(ierr); 1663 1664 /* actually zap the local rows */ 1665 /* 1666 Zero the required rows. If the "diagonal block" of the matrix 1667 is square and the user wishes to set the diagonal we use separate 1668 code so that MatSetValues() is not called for each diagonal allocating 1669 new memory, thus calling lots of mallocs and slowing things down. 1670 1671 Contributed by: Matthew Knepley 1672 */ 1673 /* must zero l->B before l->A because the (diag) case below may put values into l->B*/ 1674 ierr = MatZeroRows_SeqBAIJ(l->B,slen,lrows,0.0);CHKERRQ(ierr); 1675 if ((diag != 0.0) && (l->A->rmap->N == l->A->cmap->N)) { 1676 ierr = MatZeroRows_SeqBAIJ(l->A,slen,lrows,diag);CHKERRQ(ierr); 1677 } else if (diag != 0.0) { 1678 ierr = MatZeroRows_SeqBAIJ(l->A,slen,lrows,0.0);CHKERRQ(ierr); 1679 if (((Mat_SeqBAIJ*)l->A->data)->nonew) { 1680 SETERRQ(PETSC_ERR_SUP,"MatZeroRows() on rectangular matrices cannot be used with the Mat options \n\ 1681 MAT_NEW_NONZERO_LOCATIONS,MAT_NEW_NONZERO_LOCATION_ERR,MAT_NEW_NONZERO_ALLOCATION_ERR"); 1682 } 1683 for (i=0; i<slen; i++) { 1684 row = lrows[i] + rstart_bs; 1685 ierr = MatSetValues(A,1,&row,1,&row,&diag,INSERT_VALUES);CHKERRQ(ierr); 1686 } 1687 ierr = MatAssemblyBegin(A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1688 ierr = MatAssemblyEnd(A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1689 } else { 1690 ierr = MatZeroRows_SeqBAIJ(l->A,slen,lrows,0.0);CHKERRQ(ierr); 1691 } 1692 1693 ierr = PetscFree(lrows);CHKERRQ(ierr); 1694 1695 /* wait on sends */ 1696 if (nsends) { 1697 ierr = PetscMalloc(nsends*sizeof(MPI_Status),&send_status);CHKERRQ(ierr); 1698 ierr = MPI_Waitall(nsends,send_waits,send_status);CHKERRQ(ierr); 1699 ierr = PetscFree(send_status);CHKERRQ(ierr); 1700 } 1701 ierr = PetscFree(send_waits);CHKERRQ(ierr); 1702 ierr = PetscFree(svalues);CHKERRQ(ierr); 1703 1704 PetscFunctionReturn(0); 1705 } 1706 1707 #undef __FUNCT__ 1708 #define __FUNCT__ "MatSetUnfactored_MPIBAIJ" 1709 PetscErrorCode MatSetUnfactored_MPIBAIJ(Mat A) 1710 { 1711 Mat_MPIBAIJ *a = (Mat_MPIBAIJ*)A->data; 1712 PetscErrorCode ierr; 1713 1714 PetscFunctionBegin; 1715 ierr = MatSetUnfactored(a->A);CHKERRQ(ierr); 1716 PetscFunctionReturn(0); 1717 } 1718 1719 static PetscErrorCode MatDuplicate_MPIBAIJ(Mat,MatDuplicateOption,Mat *); 1720 1721 #undef __FUNCT__ 1722 #define __FUNCT__ "MatEqual_MPIBAIJ" 1723 PetscErrorCode MatEqual_MPIBAIJ(Mat A,Mat B,PetscTruth *flag) 1724 { 1725 Mat_MPIBAIJ *matB = (Mat_MPIBAIJ*)B->data,*matA = (Mat_MPIBAIJ*)A->data; 1726 Mat a,b,c,d; 1727 PetscTruth flg; 1728 PetscErrorCode ierr; 1729 1730 PetscFunctionBegin; 1731 a = matA->A; b = matA->B; 1732 c = matB->A; d = matB->B; 1733 1734 ierr = MatEqual(a,c,&flg);CHKERRQ(ierr); 1735 if (flg) { 1736 ierr = MatEqual(b,d,&flg);CHKERRQ(ierr); 1737 } 1738 ierr = MPI_Allreduce(&flg,flag,1,MPI_INT,MPI_LAND,((PetscObject)A)->comm);CHKERRQ(ierr); 1739 PetscFunctionReturn(0); 1740 } 1741 1742 #undef __FUNCT__ 1743 #define __FUNCT__ "MatCopy_MPIBAIJ" 1744 PetscErrorCode MatCopy_MPIBAIJ(Mat A,Mat B,MatStructure str) 1745 { 1746 PetscErrorCode ierr; 1747 Mat_MPIBAIJ *a = (Mat_MPIBAIJ *)A->data; 1748 Mat_MPIBAIJ *b = (Mat_MPIBAIJ *)B->data; 1749 1750 PetscFunctionBegin; 1751 /* If the two matrices don't have the same copy implementation, they aren't compatible for fast copy. */ 1752 if ((str != SAME_NONZERO_PATTERN) || (A->ops->copy != B->ops->copy)) { 1753 ierr = MatCopy_Basic(A,B,str);CHKERRQ(ierr); 1754 } else { 1755 ierr = MatCopy(a->A,b->A,str);CHKERRQ(ierr); 1756 ierr = MatCopy(a->B,b->B,str);CHKERRQ(ierr); 1757 } 1758 PetscFunctionReturn(0); 1759 } 1760 1761 #undef __FUNCT__ 1762 #define __FUNCT__ "MatSetUpPreallocation_MPIBAIJ" 1763 PetscErrorCode MatSetUpPreallocation_MPIBAIJ(Mat A) 1764 { 1765 PetscErrorCode ierr; 1766 1767 PetscFunctionBegin; 1768 ierr = MatMPIBAIJSetPreallocation(A,-PetscMax(A->rmap->bs,1),PETSC_DEFAULT,0,PETSC_DEFAULT,0);CHKERRQ(ierr); 1769 PetscFunctionReturn(0); 1770 } 1771 1772 #include "petscblaslapack.h" 1773 #undef __FUNCT__ 1774 #define __FUNCT__ "MatAXPY_MPIBAIJ" 1775 PetscErrorCode MatAXPY_MPIBAIJ(Mat Y,PetscScalar a,Mat X,MatStructure str) 1776 { 1777 PetscErrorCode ierr; 1778 Mat_MPIBAIJ *xx=(Mat_MPIBAIJ *)X->data,*yy=(Mat_MPIBAIJ *)Y->data; 1779 PetscBLASInt bnz,one=1; 1780 Mat_SeqBAIJ *x,*y; 1781 1782 PetscFunctionBegin; 1783 if (str == SAME_NONZERO_PATTERN) { 1784 PetscScalar alpha = a; 1785 x = (Mat_SeqBAIJ *)xx->A->data; 1786 y = (Mat_SeqBAIJ *)yy->A->data; 1787 bnz = PetscBLASIntCast(x->nz); 1788 BLASaxpy_(&bnz,&alpha,x->a,&one,y->a,&one); 1789 x = (Mat_SeqBAIJ *)xx->B->data; 1790 y = (Mat_SeqBAIJ *)yy->B->data; 1791 bnz = PetscBLASIntCast(x->nz); 1792 BLASaxpy_(&bnz,&alpha,x->a,&one,y->a,&one); 1793 } else { 1794 ierr = MatAXPY_Basic(Y,a,X,str);CHKERRQ(ierr); 1795 } 1796 PetscFunctionReturn(0); 1797 } 1798 1799 #undef __FUNCT__ 1800 #define __FUNCT__ "MatRealPart_MPIBAIJ" 1801 PetscErrorCode MatRealPart_MPIBAIJ(Mat A) 1802 { 1803 Mat_MPIBAIJ *a = (Mat_MPIBAIJ*)A->data; 1804 PetscErrorCode ierr; 1805 1806 PetscFunctionBegin; 1807 ierr = MatRealPart(a->A);CHKERRQ(ierr); 1808 ierr = MatRealPart(a->B);CHKERRQ(ierr); 1809 PetscFunctionReturn(0); 1810 } 1811 1812 #undef __FUNCT__ 1813 #define __FUNCT__ "MatImaginaryPart_MPIBAIJ" 1814 PetscErrorCode MatImaginaryPart_MPIBAIJ(Mat A) 1815 { 1816 Mat_MPIBAIJ *a = (Mat_MPIBAIJ*)A->data; 1817 PetscErrorCode ierr; 1818 1819 PetscFunctionBegin; 1820 ierr = MatImaginaryPart(a->A);CHKERRQ(ierr); 1821 ierr = MatImaginaryPart(a->B);CHKERRQ(ierr); 1822 PetscFunctionReturn(0); 1823 } 1824 1825 #undef __FUNCT__ 1826 #define __FUNCT__ "MatGetSubMatrix_MPIBAIJ" 1827 /* 1828 Not great since it makes two copies of the submatrix, first an SeqBAIJ 1829 in local and then by concatenating the local matrices the end result. 1830 Writing it directly would be much like MatGetSubMatrices_MPIBAIJ() 1831 */ 1832 PetscErrorCode MatGetSubMatrix_MPIBAIJ(Mat mat,IS isrow,IS iscol,PetscInt csize,MatReuse call,Mat *newmat) 1833 { 1834 PetscErrorCode ierr; 1835 PetscMPIInt rank,size; 1836 PetscInt i,m,n,rstart,row,rend,nz,*cwork,j,bs; 1837 PetscInt *ii,*jj,nlocal,*dlens,*olens,dlen,olen,jend,mglobal; 1838 Mat *local,M,Mreuse; 1839 MatScalar *vwork,*aa; 1840 MPI_Comm comm = ((PetscObject)mat)->comm; 1841 Mat_SeqBAIJ *aij; 1842 1843 1844 PetscFunctionBegin; 1845 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 1846 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 1847 1848 if (call == MAT_REUSE_MATRIX) { 1849 ierr = PetscObjectQuery((PetscObject)*newmat,"SubMatrix",(PetscObject *)&Mreuse);CHKERRQ(ierr); 1850 if (!Mreuse) SETERRQ(PETSC_ERR_ARG_WRONGSTATE,"Submatrix passed in was not used before, cannot reuse"); 1851 local = &Mreuse; 1852 ierr = MatGetSubMatrices(mat,1,&isrow,&iscol,MAT_REUSE_MATRIX,&local);CHKERRQ(ierr); 1853 } else { 1854 ierr = MatGetSubMatrices(mat,1,&isrow,&iscol,MAT_INITIAL_MATRIX,&local);CHKERRQ(ierr); 1855 Mreuse = *local; 1856 ierr = PetscFree(local);CHKERRQ(ierr); 1857 } 1858 1859 /* 1860 m - number of local rows 1861 n - number of columns (same on all processors) 1862 rstart - first row in new global matrix generated 1863 */ 1864 ierr = MatGetBlockSize(mat,&bs);CHKERRQ(ierr); 1865 ierr = MatGetSize(Mreuse,&m,&n);CHKERRQ(ierr); 1866 m = m/bs; 1867 n = n/bs; 1868 1869 if (call == MAT_INITIAL_MATRIX) { 1870 aij = (Mat_SeqBAIJ*)(Mreuse)->data; 1871 ii = aij->i; 1872 jj = aij->j; 1873 1874 /* 1875 Determine the number of non-zeros in the diagonal and off-diagonal 1876 portions of the matrix in order to do correct preallocation 1877 */ 1878 1879 /* first get start and end of "diagonal" columns */ 1880 if (csize == PETSC_DECIDE) { 1881 ierr = ISGetSize(isrow,&mglobal);CHKERRQ(ierr); 1882 if (mglobal == n*bs) { /* square matrix */ 1883 nlocal = m; 1884 } else { 1885 nlocal = n/size + ((n % size) > rank); 1886 } 1887 } else { 1888 nlocal = csize/bs; 1889 } 1890 ierr = MPI_Scan(&nlocal,&rend,1,MPIU_INT,MPI_SUM,comm);CHKERRQ(ierr); 1891 rstart = rend - nlocal; 1892 if (rank == size - 1 && rend != n) { 1893 SETERRQ2(PETSC_ERR_ARG_SIZ,"Local column sizes %D do not add up to total number of columns %D",rend,n); 1894 } 1895 1896 /* next, compute all the lengths */ 1897 ierr = PetscMalloc((2*m+1)*sizeof(PetscInt),&dlens);CHKERRQ(ierr); 1898 olens = dlens + m; 1899 for (i=0; i<m; i++) { 1900 jend = ii[i+1] - ii[i]; 1901 olen = 0; 1902 dlen = 0; 1903 for (j=0; j<jend; j++) { 1904 if (*jj < rstart || *jj >= rend) olen++; 1905 else dlen++; 1906 jj++; 1907 } 1908 olens[i] = olen; 1909 dlens[i] = dlen; 1910 } 1911 ierr = MatCreate(comm,&M);CHKERRQ(ierr); 1912 ierr = MatSetSizes(M,bs*m,bs*nlocal,PETSC_DECIDE,bs*n);CHKERRQ(ierr); 1913 ierr = MatSetType(M,((PetscObject)mat)->type_name);CHKERRQ(ierr); 1914 ierr = MatMPIBAIJSetPreallocation(M,bs,0,dlens,0,olens);CHKERRQ(ierr); 1915 ierr = PetscFree(dlens);CHKERRQ(ierr); 1916 } else { 1917 PetscInt ml,nl; 1918 1919 M = *newmat; 1920 ierr = MatGetLocalSize(M,&ml,&nl);CHKERRQ(ierr); 1921 if (ml != m) SETERRQ(PETSC_ERR_ARG_SIZ,"Previous matrix must be same size/layout as request"); 1922 ierr = MatZeroEntries(M);CHKERRQ(ierr); 1923 /* 1924 The next two lines are needed so we may call MatSetValues_MPIAIJ() below directly, 1925 rather than the slower MatSetValues(). 1926 */ 1927 M->was_assembled = PETSC_TRUE; 1928 M->assembled = PETSC_FALSE; 1929 } 1930 ierr = MatSetOption(M,MAT_ROW_ORIENTED,PETSC_FALSE);CHKERRQ(ierr); 1931 ierr = MatGetOwnershipRange(M,&rstart,&rend);CHKERRQ(ierr); 1932 aij = (Mat_SeqBAIJ*)(Mreuse)->data; 1933 ii = aij->i; 1934 jj = aij->j; 1935 aa = aij->a; 1936 for (i=0; i<m; i++) { 1937 row = rstart/bs + i; 1938 nz = ii[i+1] - ii[i]; 1939 cwork = jj; jj += nz; 1940 vwork = aa; aa += nz; 1941 ierr = MatSetValuesBlocked_MPIBAIJ(M,1,&row,nz,cwork,vwork,INSERT_VALUES);CHKERRQ(ierr); 1942 } 1943 1944 ierr = MatAssemblyBegin(M,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1945 ierr = MatAssemblyEnd(M,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1946 *newmat = M; 1947 1948 /* save submatrix used in processor for next request */ 1949 if (call == MAT_INITIAL_MATRIX) { 1950 ierr = PetscObjectCompose((PetscObject)M,"SubMatrix",(PetscObject)Mreuse);CHKERRQ(ierr); 1951 ierr = PetscObjectDereference((PetscObject)Mreuse);CHKERRQ(ierr); 1952 } 1953 1954 PetscFunctionReturn(0); 1955 } 1956 1957 #undef __FUNCT__ 1958 #define __FUNCT__ "MatPermute_MPIBAIJ" 1959 PetscErrorCode MatPermute_MPIBAIJ(Mat A,IS rowp,IS colp,Mat *B) 1960 { 1961 MPI_Comm comm,pcomm; 1962 PetscInt first,local_size,nrows; 1963 const PetscInt *rows; 1964 int ntids; 1965 IS crowp,growp,irowp,lrowp,lcolp,icolp; 1966 PetscErrorCode ierr; 1967 1968 PetscFunctionBegin; 1969 ierr = PetscObjectGetComm((PetscObject)A,&comm); CHKERRQ(ierr); 1970 /* make a collective version of 'rowp' */ 1971 ierr = PetscObjectGetComm((PetscObject)rowp,&pcomm); CHKERRQ(ierr); 1972 if (pcomm==comm) { 1973 crowp = rowp; 1974 } else { 1975 ierr = ISGetSize(rowp,&nrows); CHKERRQ(ierr); 1976 ierr = ISGetIndices(rowp,&rows); CHKERRQ(ierr); 1977 ierr = ISCreateGeneral(comm,nrows,rows,&crowp); CHKERRQ(ierr); 1978 ierr = ISRestoreIndices(rowp,&rows); CHKERRQ(ierr); 1979 } 1980 /* collect the global row permutation and invert it */ 1981 ierr = ISAllGather(crowp,&growp); CHKERRQ(ierr); 1982 ierr = ISSetPermutation(growp); CHKERRQ(ierr); 1983 if (pcomm!=comm) { 1984 ierr = ISDestroy(crowp); CHKERRQ(ierr); 1985 } 1986 ierr = ISInvertPermutation(growp,PETSC_DECIDE,&irowp);CHKERRQ(ierr); 1987 /* get the local target indices */ 1988 ierr = MatGetOwnershipRange(A,&first,PETSC_NULL); CHKERRQ(ierr); 1989 ierr = MatGetLocalSize(A,&local_size,PETSC_NULL); CHKERRQ(ierr); 1990 ierr = ISGetIndices(irowp,&rows); CHKERRQ(ierr); 1991 ierr = ISCreateGeneral(MPI_COMM_SELF,local_size,rows+first,&lrowp); CHKERRQ(ierr); 1992 ierr = ISRestoreIndices(irowp,&rows); CHKERRQ(ierr); 1993 ierr = ISDestroy(irowp); CHKERRQ(ierr); 1994 /* the column permutation is so much easier; 1995 make a local version of 'colp' and invert it */ 1996 ierr = PetscObjectGetComm((PetscObject)colp,&pcomm); CHKERRQ(ierr); 1997 ierr = MPI_Comm_size(pcomm,&ntids); CHKERRQ(ierr); 1998 if (ntids==1) { 1999 lcolp = colp; 2000 } else { 2001 ierr = ISGetSize(colp,&nrows); CHKERRQ(ierr); 2002 ierr = ISGetIndices(colp,&rows); CHKERRQ(ierr); 2003 ierr = ISCreateGeneral(MPI_COMM_SELF,nrows,rows,&lcolp); CHKERRQ(ierr); 2004 } 2005 ierr = ISInvertPermutation(lcolp,PETSC_DECIDE,&icolp); CHKERRQ(ierr); 2006 ierr = ISSetPermutation(lcolp); CHKERRQ(ierr); 2007 if (ntids>1) { 2008 ierr = ISRestoreIndices(colp,&rows); CHKERRQ(ierr); 2009 ierr = ISDestroy(lcolp); CHKERRQ(ierr); 2010 } 2011 /* now we just get the submatrix */ 2012 ierr = MatGetSubMatrix(A,lrowp,icolp,local_size,MAT_INITIAL_MATRIX,B); CHKERRQ(ierr); 2013 /* clean up */ 2014 ierr = ISDestroy(lrowp); CHKERRQ(ierr); 2015 ierr = ISDestroy(icolp); CHKERRQ(ierr); 2016 PetscFunctionReturn(0); 2017 } 2018 2019 /* -------------------------------------------------------------------*/ 2020 static struct _MatOps MatOps_Values = { 2021 MatSetValues_MPIBAIJ, 2022 MatGetRow_MPIBAIJ, 2023 MatRestoreRow_MPIBAIJ, 2024 MatMult_MPIBAIJ, 2025 /* 4*/ MatMultAdd_MPIBAIJ, 2026 MatMultTranspose_MPIBAIJ, 2027 MatMultTransposeAdd_MPIBAIJ, 2028 0, 2029 0, 2030 0, 2031 /*10*/ 0, 2032 0, 2033 0, 2034 0, 2035 MatTranspose_MPIBAIJ, 2036 /*15*/ MatGetInfo_MPIBAIJ, 2037 MatEqual_MPIBAIJ, 2038 MatGetDiagonal_MPIBAIJ, 2039 MatDiagonalScale_MPIBAIJ, 2040 MatNorm_MPIBAIJ, 2041 /*20*/ MatAssemblyBegin_MPIBAIJ, 2042 MatAssemblyEnd_MPIBAIJ, 2043 0, 2044 MatSetOption_MPIBAIJ, 2045 MatZeroEntries_MPIBAIJ, 2046 /*25*/ MatZeroRows_MPIBAIJ, 2047 0, 2048 0, 2049 0, 2050 0, 2051 /*30*/ MatSetUpPreallocation_MPIBAIJ, 2052 0, 2053 0, 2054 0, 2055 0, 2056 /*35*/ MatDuplicate_MPIBAIJ, 2057 0, 2058 0, 2059 0, 2060 0, 2061 /*40*/ MatAXPY_MPIBAIJ, 2062 MatGetSubMatrices_MPIBAIJ, 2063 MatIncreaseOverlap_MPIBAIJ, 2064 MatGetValues_MPIBAIJ, 2065 MatCopy_MPIBAIJ, 2066 /*45*/ 0, 2067 MatScale_MPIBAIJ, 2068 0, 2069 0, 2070 0, 2071 /*50*/ 0, 2072 0, 2073 0, 2074 0, 2075 0, 2076 /*55*/ 0, 2077 0, 2078 MatSetUnfactored_MPIBAIJ, 2079 MatPermute_MPIBAIJ, 2080 MatSetValuesBlocked_MPIBAIJ, 2081 /*60*/ MatGetSubMatrix_MPIBAIJ, 2082 MatDestroy_MPIBAIJ, 2083 MatView_MPIBAIJ, 2084 0, 2085 0, 2086 /*65*/ 0, 2087 0, 2088 0, 2089 0, 2090 0, 2091 /*70*/ MatGetRowMaxAbs_MPIBAIJ, 2092 0, 2093 0, 2094 0, 2095 0, 2096 /*75*/ 0, 2097 0, 2098 0, 2099 0, 2100 0, 2101 /*80*/ 0, 2102 0, 2103 0, 2104 0, 2105 MatLoad_MPIBAIJ, 2106 /*85*/ 0, 2107 0, 2108 0, 2109 0, 2110 0, 2111 /*90*/ 0, 2112 0, 2113 0, 2114 0, 2115 0, 2116 /*95*/ 0, 2117 0, 2118 0, 2119 0, 2120 0, 2121 /*100*/0, 2122 0, 2123 0, 2124 0, 2125 0, 2126 /*105*/0, 2127 MatRealPart_MPIBAIJ, 2128 MatImaginaryPart_MPIBAIJ}; 2129 2130 2131 EXTERN_C_BEGIN 2132 #undef __FUNCT__ 2133 #define __FUNCT__ "MatGetDiagonalBlock_MPIBAIJ" 2134 PetscErrorCode PETSCMAT_DLLEXPORT MatGetDiagonalBlock_MPIBAIJ(Mat A,PetscTruth *iscopy,MatReuse reuse,Mat *a) 2135 { 2136 PetscFunctionBegin; 2137 *a = ((Mat_MPIBAIJ *)A->data)->A; 2138 *iscopy = PETSC_FALSE; 2139 PetscFunctionReturn(0); 2140 } 2141 EXTERN_C_END 2142 2143 EXTERN_C_BEGIN 2144 extern PetscErrorCode PETSCMAT_DLLEXPORT MatConvert_MPIBAIJ_MPISBAIJ(Mat, MatType,MatReuse,Mat*); 2145 EXTERN_C_END 2146 2147 EXTERN_C_BEGIN 2148 #undef __FUNCT__ 2149 #define __FUNCT__ "MatMPIBAIJSetPreallocationCSR_MPIBAIJ" 2150 PetscErrorCode MatMPIBAIJSetPreallocationCSR_MPIBAIJ(Mat B,PetscInt bs,const PetscInt ii[],const PetscInt jj[],const PetscScalar V[]) 2151 { 2152 PetscInt m,rstart,cstart,cend; 2153 PetscInt i,j,d,nz,nz_max=0,*d_nnz=0,*o_nnz=0; 2154 const PetscInt *JJ=0; 2155 PetscScalar *values=0; 2156 PetscErrorCode ierr; 2157 2158 PetscFunctionBegin; 2159 2160 if (bs < 1) SETERRQ1(PETSC_ERR_ARG_OUTOFRANGE,"Invalid block size specified, must be positive but it is %D",bs); 2161 ierr = PetscMapSetBlockSize(B->rmap,bs);CHKERRQ(ierr); 2162 ierr = PetscMapSetBlockSize(B->cmap,bs);CHKERRQ(ierr); 2163 ierr = PetscMapSetUp(B->rmap);CHKERRQ(ierr); 2164 ierr = PetscMapSetUp(B->cmap);CHKERRQ(ierr); 2165 m = B->rmap->n/bs; 2166 rstart = B->rmap->rstart/bs; 2167 cstart = B->cmap->rstart/bs; 2168 cend = B->cmap->rend/bs; 2169 2170 if (ii[0]) SETERRQ1(PETSC_ERR_ARG_OUTOFRANGE,"ii[0] must be 0 but it is %D",ii[0]); 2171 ierr = PetscMalloc((2*m+1)*sizeof(PetscInt),&d_nnz);CHKERRQ(ierr); 2172 o_nnz = d_nnz + m; 2173 for (i=0; i<m; i++) { 2174 nz = ii[i+1] - ii[i]; 2175 if (nz < 0) SETERRQ2(PETSC_ERR_ARG_OUTOFRANGE,"Local row %D has a negative number of columns %D",i,nz); 2176 nz_max = PetscMax(nz_max,nz); 2177 JJ = jj + ii[i]; 2178 for (j=0; j<nz; j++) { 2179 if (*JJ >= cstart) break; 2180 JJ++; 2181 } 2182 d = 0; 2183 for (; j<nz; j++) { 2184 if (*JJ++ >= cend) break; 2185 d++; 2186 } 2187 d_nnz[i] = d; 2188 o_nnz[i] = nz - d; 2189 } 2190 ierr = MatMPIBAIJSetPreallocation(B,bs,0,d_nnz,0,o_nnz);CHKERRQ(ierr); 2191 ierr = PetscFree(d_nnz);CHKERRQ(ierr); 2192 2193 values = (PetscScalar*)V; 2194 if (!values) { 2195 ierr = PetscMalloc(bs*bs*(nz_max+1)*sizeof(PetscScalar),&values);CHKERRQ(ierr); 2196 ierr = PetscMemzero(values,bs*bs*nz_max*sizeof(PetscScalar));CHKERRQ(ierr); 2197 } 2198 for (i=0; i<m; i++) { 2199 PetscInt row = i + rstart; 2200 PetscInt ncols = ii[i+1] - ii[i]; 2201 const PetscInt *icols = jj + ii[i]; 2202 const PetscScalar *svals = values + (V ? (bs*bs*ii[i]) : 0); 2203 ierr = MatSetValuesBlocked_MPIBAIJ(B,1,&row,ncols,icols,svals,INSERT_VALUES);CHKERRQ(ierr); 2204 } 2205 2206 if (!V) { ierr = PetscFree(values);CHKERRQ(ierr); } 2207 ierr = MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2208 ierr = MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2209 2210 PetscFunctionReturn(0); 2211 } 2212 EXTERN_C_END 2213 2214 #undef __FUNCT__ 2215 #define __FUNCT__ "MatMPIBAIJSetPreallocationCSR" 2216 /*@C 2217 MatMPIBAIJSetPreallocationCSR - Allocates memory for a sparse parallel matrix in AIJ format 2218 (the default parallel PETSc format). 2219 2220 Collective on MPI_Comm 2221 2222 Input Parameters: 2223 + A - the matrix 2224 . i - the indices into j for the start of each local row (starts with zero) 2225 . j - the column indices for each local row (starts with zero) these must be sorted for each row 2226 - v - optional values in the matrix 2227 2228 Level: developer 2229 2230 .keywords: matrix, aij, compressed row, sparse, parallel 2231 2232 .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatMPIBAIJSetPreallocation(), MatCreateMPIAIJ(), MPIAIJ 2233 @*/ 2234 PetscErrorCode PETSCMAT_DLLEXPORT MatMPIBAIJSetPreallocationCSR(Mat B,PetscInt bs,const PetscInt i[],const PetscInt j[], const PetscScalar v[]) 2235 { 2236 PetscErrorCode ierr,(*f)(Mat,PetscInt,const PetscInt[],const PetscInt[],const PetscScalar[]); 2237 2238 PetscFunctionBegin; 2239 ierr = PetscObjectQueryFunction((PetscObject)B,"MatMPIBAIJSetPreallocationCSR_C",(void (**)(void))&f);CHKERRQ(ierr); 2240 if (f) { 2241 ierr = (*f)(B,bs,i,j,v);CHKERRQ(ierr); 2242 } 2243 PetscFunctionReturn(0); 2244 } 2245 2246 EXTERN_C_BEGIN 2247 #undef __FUNCT__ 2248 #define __FUNCT__ "MatMPIBAIJSetPreallocation_MPIBAIJ" 2249 PetscErrorCode PETSCMAT_DLLEXPORT MatMPIBAIJSetPreallocation_MPIBAIJ(Mat B,PetscInt bs,PetscInt d_nz,PetscInt *d_nnz,PetscInt o_nz,PetscInt *o_nnz) 2250 { 2251 Mat_MPIBAIJ *b; 2252 PetscErrorCode ierr; 2253 PetscInt i, newbs = PetscAbs(bs); 2254 2255 PetscFunctionBegin; 2256 B->preallocated = PETSC_TRUE; 2257 if (bs < 0) { 2258 ierr = PetscOptionsBegin(((PetscObject)B)->comm,((PetscObject)B)->prefix,"Options for MPIBAIJ matrix","Mat");CHKERRQ(ierr); 2259 ierr = PetscOptionsInt("-mat_block_size","Set the blocksize used to store the matrix","MatMPIBAIJSetPreallocation",newbs,&newbs,PETSC_NULL);CHKERRQ(ierr); 2260 ierr = PetscOptionsEnd();CHKERRQ(ierr); 2261 bs = PetscAbs(bs); 2262 } 2263 if ((d_nnz || o_nnz) && newbs != bs) { 2264 SETERRQ(PETSC_ERR_ARG_WRONG,"Cannot change blocksize from command line if setting d_nnz or o_nnz"); 2265 } 2266 bs = newbs; 2267 2268 2269 if (bs < 1) SETERRQ(PETSC_ERR_ARG_OUTOFRANGE,"Invalid block size specified, must be positive"); 2270 if (d_nz == PETSC_DEFAULT || d_nz == PETSC_DECIDE) d_nz = 5; 2271 if (o_nz == PETSC_DEFAULT || o_nz == PETSC_DECIDE) o_nz = 2; 2272 if (d_nz < 0) SETERRQ1(PETSC_ERR_ARG_OUTOFRANGE,"d_nz cannot be less than 0: value %D",d_nz); 2273 if (o_nz < 0) SETERRQ1(PETSC_ERR_ARG_OUTOFRANGE,"o_nz cannot be less than 0: value %D",o_nz); 2274 2275 ierr = PetscMapSetBlockSize(B->rmap,bs);CHKERRQ(ierr); 2276 ierr = PetscMapSetBlockSize(B->cmap,bs);CHKERRQ(ierr); 2277 ierr = PetscMapSetUp(B->rmap);CHKERRQ(ierr); 2278 ierr = PetscMapSetUp(B->cmap);CHKERRQ(ierr); 2279 2280 if (d_nnz) { 2281 for (i=0; i<B->rmap->n/bs; i++) { 2282 if (d_nnz[i] < 0) SETERRQ2(PETSC_ERR_ARG_OUTOFRANGE,"d_nnz cannot be less than -1: local row %D value %D",i,d_nnz[i]); 2283 } 2284 } 2285 if (o_nnz) { 2286 for (i=0; i<B->rmap->n/bs; i++) { 2287 if (o_nnz[i] < 0) SETERRQ2(PETSC_ERR_ARG_OUTOFRANGE,"o_nnz cannot be less than -1: local row %D value %D",i,o_nnz[i]); 2288 } 2289 } 2290 2291 b = (Mat_MPIBAIJ*)B->data; 2292 b->bs2 = bs*bs; 2293 b->mbs = B->rmap->n/bs; 2294 b->nbs = B->cmap->n/bs; 2295 b->Mbs = B->rmap->N/bs; 2296 b->Nbs = B->cmap->N/bs; 2297 2298 for (i=0; i<=b->size; i++) { 2299 b->rangebs[i] = B->rmap->range[i]/bs; 2300 } 2301 b->rstartbs = B->rmap->rstart/bs; 2302 b->rendbs = B->rmap->rend/bs; 2303 b->cstartbs = B->cmap->rstart/bs; 2304 b->cendbs = B->cmap->rend/bs; 2305 2306 ierr = MatCreate(PETSC_COMM_SELF,&b->A);CHKERRQ(ierr); 2307 ierr = MatSetSizes(b->A,B->rmap->n,B->cmap->n,B->rmap->n,B->cmap->n);CHKERRQ(ierr); 2308 ierr = MatSetType(b->A,MATSEQBAIJ);CHKERRQ(ierr); 2309 ierr = MatSeqBAIJSetPreallocation(b->A,bs,d_nz,d_nnz);CHKERRQ(ierr); 2310 ierr = PetscLogObjectParent(B,b->A);CHKERRQ(ierr); 2311 ierr = MatCreate(PETSC_COMM_SELF,&b->B);CHKERRQ(ierr); 2312 ierr = MatSetSizes(b->B,B->rmap->n,B->cmap->N,B->rmap->n,B->cmap->N);CHKERRQ(ierr); 2313 ierr = MatSetType(b->B,MATSEQBAIJ);CHKERRQ(ierr); 2314 ierr = MatSeqBAIJSetPreallocation(b->B,bs,o_nz,o_nnz);CHKERRQ(ierr); 2315 ierr = PetscLogObjectParent(B,b->B);CHKERRQ(ierr); 2316 2317 ierr = MatStashCreate_Private(((PetscObject)B)->comm,bs,&B->bstash);CHKERRQ(ierr); 2318 2319 PetscFunctionReturn(0); 2320 } 2321 EXTERN_C_END 2322 2323 EXTERN_C_BEGIN 2324 EXTERN PetscErrorCode PETSCMAT_DLLEXPORT MatDiagonalScaleLocal_MPIBAIJ(Mat,Vec); 2325 EXTERN PetscErrorCode PETSCMAT_DLLEXPORT MatSetHashTableFactor_MPIBAIJ(Mat,PetscReal); 2326 EXTERN_C_END 2327 2328 2329 EXTERN_C_BEGIN 2330 #undef __FUNCT__ 2331 #define __FUNCT__ "MatConvert_MPIBAIJ_MPIAdj" 2332 PetscErrorCode PETSCMAT_DLLEXPORT MatConvert_MPIBAIJ_MPIAdj(Mat B, const MatType newtype,MatReuse reuse,Mat *adj) 2333 { 2334 Mat_MPIBAIJ *b = (Mat_MPIBAIJ*)B->data; 2335 PetscErrorCode ierr; 2336 Mat_SeqBAIJ *d = (Mat_SeqBAIJ*) b->A->data,*o = (Mat_SeqBAIJ*) b->B->data; 2337 PetscInt M = B->rmap->n/B->rmap->bs,i,*ii,*jj,cnt,j,k,rstart = B->rmap->rstart/B->rmap->bs; 2338 const PetscInt *id = d->i, *jd = d->j, *io = o->i, *jo = o->j, *garray = b->garray; 2339 2340 PetscFunctionBegin; 2341 ierr = PetscMalloc((M+1)*sizeof(PetscInt),&ii);CHKERRQ(ierr); 2342 ii[0] = 0; 2343 CHKMEMQ; 2344 for (i=0; i<M; i++) { 2345 if ((id[i+1] - id[i]) < 0) SETERRQ3(PETSC_ERR_PLIB,"Indices wrong %D %D %D",i,id[i],id[i+1]); 2346 if ((io[i+1] - io[i]) < 0) SETERRQ3(PETSC_ERR_PLIB,"Indices wrong %D %D %D",i,io[i],io[i+1]); 2347 ii[i+1] = ii[i] + id[i+1] - id[i] + io[i+1] - io[i]; 2348 /* remove one from count of matrix has diagonal */ 2349 for (j=id[i]; j<id[i+1]; j++) { 2350 if (jd[j] == i) {ii[i+1]--;break;} 2351 } 2352 CHKMEMQ; 2353 } 2354 PetscIntView(M+1,id,0); 2355 PetscIntView(M+1,io,0); 2356 PetscIntView(M+1,ii,0); 2357 ierr = PetscMalloc(ii[M]*sizeof(PetscInt),&jj);CHKERRQ(ierr); 2358 cnt = 0; 2359 for (i=0; i<M; i++) { 2360 for (j=io[i]; j<io[i+1]; j++) { 2361 if (garray[jo[j]] > rstart) break; 2362 jj[cnt++] = garray[jo[j]]; 2363 CHKMEMQ; 2364 } 2365 for (k=id[i]; k<id[i+1]; k++) { 2366 if (jd[k] != i) { 2367 jj[cnt++] = rstart + jd[k]; 2368 CHKMEMQ; 2369 } 2370 } 2371 for (;j<io[i+1]; j++) { 2372 jj[cnt++] = garray[jo[j]]; 2373 CHKMEMQ; 2374 } 2375 } 2376 ierr = MatCreateMPIAdj(((PetscObject)B)->comm,M,B->cmap->N/B->rmap->bs,ii,jj,PETSC_NULL,adj);CHKERRQ(ierr); 2377 PetscFunctionReturn(0); 2378 } 2379 EXTERN_C_END 2380 2381 2382 /*MC 2383 MATMPIBAIJ - MATMPIBAIJ = "mpibaij" - A matrix type to be used for distributed block sparse matrices. 2384 2385 Options Database Keys: 2386 + -mat_type mpibaij - sets the matrix type to "mpibaij" during a call to MatSetFromOptions() 2387 . -mat_block_size <bs> - set the blocksize used to store the matrix 2388 - -mat_use_hash_table <fact> 2389 2390 Level: beginner 2391 2392 .seealso: MatCreateMPIBAIJ 2393 M*/ 2394 2395 EXTERN_C_BEGIN 2396 #undef __FUNCT__ 2397 #define __FUNCT__ "MatCreate_MPIBAIJ" 2398 PetscErrorCode PETSCMAT_DLLEXPORT MatCreate_MPIBAIJ(Mat B) 2399 { 2400 Mat_MPIBAIJ *b; 2401 PetscErrorCode ierr; 2402 PetscTruth flg; 2403 2404 PetscFunctionBegin; 2405 ierr = PetscNewLog(B,Mat_MPIBAIJ,&b);CHKERRQ(ierr); 2406 B->data = (void*)b; 2407 2408 2409 ierr = PetscMemcpy(B->ops,&MatOps_Values,sizeof(struct _MatOps));CHKERRQ(ierr); 2410 B->mapping = 0; 2411 B->assembled = PETSC_FALSE; 2412 2413 B->insertmode = NOT_SET_VALUES; 2414 ierr = MPI_Comm_rank(((PetscObject)B)->comm,&b->rank);CHKERRQ(ierr); 2415 ierr = MPI_Comm_size(((PetscObject)B)->comm,&b->size);CHKERRQ(ierr); 2416 2417 /* build local table of row and column ownerships */ 2418 ierr = PetscMalloc((b->size+1)*sizeof(PetscInt),&b->rangebs);CHKERRQ(ierr); 2419 2420 /* build cache for off array entries formed */ 2421 ierr = MatStashCreate_Private(((PetscObject)B)->comm,1,&B->stash);CHKERRQ(ierr); 2422 b->donotstash = PETSC_FALSE; 2423 b->colmap = PETSC_NULL; 2424 b->garray = PETSC_NULL; 2425 b->roworiented = PETSC_TRUE; 2426 2427 /* stuff used in block assembly */ 2428 b->barray = 0; 2429 2430 /* stuff used for matrix vector multiply */ 2431 b->lvec = 0; 2432 b->Mvctx = 0; 2433 2434 /* stuff for MatGetRow() */ 2435 b->rowindices = 0; 2436 b->rowvalues = 0; 2437 b->getrowactive = PETSC_FALSE; 2438 2439 /* hash table stuff */ 2440 b->ht = 0; 2441 b->hd = 0; 2442 b->ht_size = 0; 2443 b->ht_flag = PETSC_FALSE; 2444 b->ht_fact = 0; 2445 b->ht_total_ct = 0; 2446 b->ht_insert_ct = 0; 2447 2448 ierr = PetscOptionsBegin(((PetscObject)B)->comm,PETSC_NULL,"Options for loading MPIBAIJ matrix 1","Mat");CHKERRQ(ierr); 2449 ierr = PetscOptionsTruth("-mat_use_hash_table","Use hash table to save memory in constructing matrix","MatSetOption",PETSC_FALSE,&flg,PETSC_NULL);CHKERRQ(ierr); 2450 if (flg) { 2451 PetscReal fact = 1.39; 2452 ierr = MatSetOption(B,MAT_USE_HASH_TABLE,PETSC_TRUE);CHKERRQ(ierr); 2453 ierr = PetscOptionsReal("-mat_use_hash_table","Use hash table factor","MatMPIBAIJSetHashTableFactor",fact,&fact,PETSC_NULL);CHKERRQ(ierr); 2454 if (fact <= 1.0) fact = 1.39; 2455 ierr = MatMPIBAIJSetHashTableFactor(B,fact);CHKERRQ(ierr); 2456 ierr = PetscInfo1(B,"Hash table Factor used %5.2f\n",fact);CHKERRQ(ierr); 2457 } 2458 ierr = PetscOptionsEnd();CHKERRQ(ierr); 2459 2460 ierr = PetscObjectComposeFunctionDynamic((PetscObject)B,"MatConvert_mpibaij_mpiadj_C", 2461 "MatConvert_MPIBAIJ_MPIAdj", 2462 MatConvert_MPIBAIJ_MPIAdj);CHKERRQ(ierr); 2463 ierr = PetscObjectComposeFunctionDynamic((PetscObject)B,"MatStoreValues_C", 2464 "MatStoreValues_MPIBAIJ", 2465 MatStoreValues_MPIBAIJ);CHKERRQ(ierr); 2466 ierr = PetscObjectComposeFunctionDynamic((PetscObject)B,"MatRetrieveValues_C", 2467 "MatRetrieveValues_MPIBAIJ", 2468 MatRetrieveValues_MPIBAIJ);CHKERRQ(ierr); 2469 ierr = PetscObjectComposeFunctionDynamic((PetscObject)B,"MatGetDiagonalBlock_C", 2470 "MatGetDiagonalBlock_MPIBAIJ", 2471 MatGetDiagonalBlock_MPIBAIJ);CHKERRQ(ierr); 2472 ierr = PetscObjectComposeFunctionDynamic((PetscObject)B,"MatMPIBAIJSetPreallocation_C", 2473 "MatMPIBAIJSetPreallocation_MPIBAIJ", 2474 MatMPIBAIJSetPreallocation_MPIBAIJ);CHKERRQ(ierr); 2475 ierr = PetscObjectComposeFunctionDynamic((PetscObject)B,"MatMPIBAIJSetPreallocationCSR_C", 2476 "MatMPIBAIJSetPreallocationCSR_MPIBAIJ", 2477 MatMPIBAIJSetPreallocationCSR_MPIBAIJ);CHKERRQ(ierr); 2478 ierr = PetscObjectComposeFunctionDynamic((PetscObject)B,"MatDiagonalScaleLocal_C", 2479 "MatDiagonalScaleLocal_MPIBAIJ", 2480 MatDiagonalScaleLocal_MPIBAIJ);CHKERRQ(ierr); 2481 ierr = PetscObjectComposeFunctionDynamic((PetscObject)B,"MatSetHashTableFactor_C", 2482 "MatSetHashTableFactor_MPIBAIJ", 2483 MatSetHashTableFactor_MPIBAIJ);CHKERRQ(ierr); 2484 ierr = PetscObjectChangeTypeName((PetscObject)B,MATMPIBAIJ);CHKERRQ(ierr); 2485 PetscFunctionReturn(0); 2486 } 2487 EXTERN_C_END 2488 2489 /*MC 2490 MATBAIJ - MATBAIJ = "baij" - A matrix type to be used for block sparse matrices. 2491 2492 This matrix type is identical to MATSEQBAIJ when constructed with a single process communicator, 2493 and MATMPIBAIJ otherwise. 2494 2495 Options Database Keys: 2496 . -mat_type baij - sets the matrix type to "baij" during a call to MatSetFromOptions() 2497 2498 Level: beginner 2499 2500 .seealso: MatCreateMPIBAIJ(),MATSEQBAIJ,MATMPIBAIJ, MatMPIBAIJSetPreallocation(), MatMPIBAIJSetPreallocationCSR() 2501 M*/ 2502 2503 EXTERN_C_BEGIN 2504 #undef __FUNCT__ 2505 #define __FUNCT__ "MatCreate_BAIJ" 2506 PetscErrorCode PETSCMAT_DLLEXPORT MatCreate_BAIJ(Mat A) 2507 { 2508 PetscErrorCode ierr; 2509 PetscMPIInt size; 2510 2511 PetscFunctionBegin; 2512 ierr = MPI_Comm_size(((PetscObject)A)->comm,&size);CHKERRQ(ierr); 2513 if (size == 1) { 2514 ierr = MatSetType(A,MATSEQBAIJ);CHKERRQ(ierr); 2515 } else { 2516 ierr = MatSetType(A,MATMPIBAIJ);CHKERRQ(ierr); 2517 } 2518 PetscFunctionReturn(0); 2519 } 2520 EXTERN_C_END 2521 2522 #undef __FUNCT__ 2523 #define __FUNCT__ "MatMPIBAIJSetPreallocation" 2524 /*@C 2525 MatMPIBAIJSetPreallocation - Allocates memory for a sparse parallel matrix in block AIJ format 2526 (block compressed row). For good matrix assembly performance 2527 the user should preallocate the matrix storage by setting the parameters 2528 d_nz (or d_nnz) and o_nz (or o_nnz). By setting these parameters accurately, 2529 performance can be increased by more than a factor of 50. 2530 2531 Collective on Mat 2532 2533 Input Parameters: 2534 + A - the matrix 2535 . bs - size of blockk 2536 . d_nz - number of block nonzeros per block row in diagonal portion of local 2537 submatrix (same for all local rows) 2538 . d_nnz - array containing the number of block nonzeros in the various block rows 2539 of the in diagonal portion of the local (possibly different for each block 2540 row) or PETSC_NULL. You must leave room for the diagonal entry even if it is zero. 2541 . o_nz - number of block nonzeros per block row in the off-diagonal portion of local 2542 submatrix (same for all local rows). 2543 - o_nnz - array containing the number of nonzeros in the various block rows of the 2544 off-diagonal portion of the local submatrix (possibly different for 2545 each block row) or PETSC_NULL. 2546 2547 If the *_nnz parameter is given then the *_nz parameter is ignored 2548 2549 Options Database Keys: 2550 + -mat_block_size - size of the blocks to use 2551 - -mat_use_hash_table <fact> 2552 2553 Notes: 2554 If PETSC_DECIDE or PETSC_DETERMINE is used for a particular argument on one processor 2555 than it must be used on all processors that share the object for that argument. 2556 2557 Storage Information: 2558 For a square global matrix we define each processor's diagonal portion 2559 to be its local rows and the corresponding columns (a square submatrix); 2560 each processor's off-diagonal portion encompasses the remainder of the 2561 local matrix (a rectangular submatrix). 2562 2563 The user can specify preallocated storage for the diagonal part of 2564 the local submatrix with either d_nz or d_nnz (not both). Set 2565 d_nz=PETSC_DEFAULT and d_nnz=PETSC_NULL for PETSc to control dynamic 2566 memory allocation. Likewise, specify preallocated storage for the 2567 off-diagonal part of the local submatrix with o_nz or o_nnz (not both). 2568 2569 Consider a processor that owns rows 3, 4 and 5 of a parallel matrix. In 2570 the figure below we depict these three local rows and all columns (0-11). 2571 2572 .vb 2573 0 1 2 3 4 5 6 7 8 9 10 11 2574 ------------------- 2575 row 3 | o o o d d d o o o o o o 2576 row 4 | o o o d d d o o o o o o 2577 row 5 | o o o d d d o o o o o o 2578 ------------------- 2579 .ve 2580 2581 Thus, any entries in the d locations are stored in the d (diagonal) 2582 submatrix, and any entries in the o locations are stored in the 2583 o (off-diagonal) submatrix. Note that the d and the o submatrices are 2584 stored simply in the MATSEQBAIJ format for compressed row storage. 2585 2586 Now d_nz should indicate the number of block nonzeros per row in the d matrix, 2587 and o_nz should indicate the number of block nonzeros per row in the o matrix. 2588 In general, for PDE problems in which most nonzeros are near the diagonal, 2589 one expects d_nz >> o_nz. For large problems you MUST preallocate memory 2590 or you will get TERRIBLE performance; see the users' manual chapter on 2591 matrices. 2592 2593 You can call MatGetInfo() to get information on how effective the preallocation was; 2594 for example the fields mallocs,nz_allocated,nz_used,nz_unneeded; 2595 You can also run with the option -info and look for messages with the string 2596 malloc in them to see if additional memory allocation was needed. 2597 2598 Level: intermediate 2599 2600 .keywords: matrix, block, aij, compressed row, sparse, parallel 2601 2602 .seealso: MatCreate(), MatCreateSeqBAIJ(), MatSetValues(), MatCreateMPIBAIJ(), MatMPIBAIJSetPreallocationCSR() 2603 @*/ 2604 PetscErrorCode PETSCMAT_DLLEXPORT MatMPIBAIJSetPreallocation(Mat B,PetscInt bs,PetscInt d_nz,const PetscInt d_nnz[],PetscInt o_nz,const PetscInt o_nnz[]) 2605 { 2606 PetscErrorCode ierr,(*f)(Mat,PetscInt,PetscInt,const PetscInt[],PetscInt,const PetscInt[]); 2607 2608 PetscFunctionBegin; 2609 ierr = PetscObjectQueryFunction((PetscObject)B,"MatMPIBAIJSetPreallocation_C",(void (**)(void))&f);CHKERRQ(ierr); 2610 if (f) { 2611 ierr = (*f)(B,bs,d_nz,d_nnz,o_nz,o_nnz);CHKERRQ(ierr); 2612 } 2613 PetscFunctionReturn(0); 2614 } 2615 2616 #undef __FUNCT__ 2617 #define __FUNCT__ "MatCreateMPIBAIJ" 2618 /*@C 2619 MatCreateMPIBAIJ - Creates a sparse parallel matrix in block AIJ format 2620 (block compressed row). For good matrix assembly performance 2621 the user should preallocate the matrix storage by setting the parameters 2622 d_nz (or d_nnz) and o_nz (or o_nnz). By setting these parameters accurately, 2623 performance can be increased by more than a factor of 50. 2624 2625 Collective on MPI_Comm 2626 2627 Input Parameters: 2628 + comm - MPI communicator 2629 . bs - size of blockk 2630 . m - number of local rows (or PETSC_DECIDE to have calculated if M is given) 2631 This value should be the same as the local size used in creating the 2632 y vector for the matrix-vector product y = Ax. 2633 . n - number of local columns (or PETSC_DECIDE to have calculated if N is given) 2634 This value should be the same as the local size used in creating the 2635 x vector for the matrix-vector product y = Ax. 2636 . M - number of global rows (or PETSC_DETERMINE to have calculated if m is given) 2637 . N - number of global columns (or PETSC_DETERMINE to have calculated if n is given) 2638 . d_nz - number of nonzero blocks per block row in diagonal portion of local 2639 submatrix (same for all local rows) 2640 . d_nnz - array containing the number of nonzero blocks in the various block rows 2641 of the in diagonal portion of the local (possibly different for each block 2642 row) or PETSC_NULL. You must leave room for the diagonal entry even if it is zero. 2643 . o_nz - number of nonzero blocks per block row in the off-diagonal portion of local 2644 submatrix (same for all local rows). 2645 - o_nnz - array containing the number of nonzero blocks in the various block rows of the 2646 off-diagonal portion of the local submatrix (possibly different for 2647 each block row) or PETSC_NULL. 2648 2649 Output Parameter: 2650 . A - the matrix 2651 2652 Options Database Keys: 2653 + -mat_block_size - size of the blocks to use 2654 - -mat_use_hash_table <fact> 2655 2656 It is recommended that one use the MatCreate(), MatSetType() and/or MatSetFromOptions(), 2657 MatXXXXSetPreallocation() paradgm instead of this routine directly. 2658 [MatXXXXSetPreallocation() is, for example, MatSeqAIJSetPreallocation] 2659 2660 Notes: 2661 If the *_nnz parameter is given then the *_nz parameter is ignored 2662 2663 A nonzero block is any block that as 1 or more nonzeros in it 2664 2665 The user MUST specify either the local or global matrix dimensions 2666 (possibly both). 2667 2668 If PETSC_DECIDE or PETSC_DETERMINE is used for a particular argument on one processor 2669 than it must be used on all processors that share the object for that argument. 2670 2671 Storage Information: 2672 For a square global matrix we define each processor's diagonal portion 2673 to be its local rows and the corresponding columns (a square submatrix); 2674 each processor's off-diagonal portion encompasses the remainder of the 2675 local matrix (a rectangular submatrix). 2676 2677 The user can specify preallocated storage for the diagonal part of 2678 the local submatrix with either d_nz or d_nnz (not both). Set 2679 d_nz=PETSC_DEFAULT and d_nnz=PETSC_NULL for PETSc to control dynamic 2680 memory allocation. Likewise, specify preallocated storage for the 2681 off-diagonal part of the local submatrix with o_nz or o_nnz (not both). 2682 2683 Consider a processor that owns rows 3, 4 and 5 of a parallel matrix. In 2684 the figure below we depict these three local rows and all columns (0-11). 2685 2686 .vb 2687 0 1 2 3 4 5 6 7 8 9 10 11 2688 ------------------- 2689 row 3 | o o o d d d o o o o o o 2690 row 4 | o o o d d d o o o o o o 2691 row 5 | o o o d d d o o o o o o 2692 ------------------- 2693 .ve 2694 2695 Thus, any entries in the d locations are stored in the d (diagonal) 2696 submatrix, and any entries in the o locations are stored in the 2697 o (off-diagonal) submatrix. Note that the d and the o submatrices are 2698 stored simply in the MATSEQBAIJ format for compressed row storage. 2699 2700 Now d_nz should indicate the number of block nonzeros per row in the d matrix, 2701 and o_nz should indicate the number of block nonzeros per row in the o matrix. 2702 In general, for PDE problems in which most nonzeros are near the diagonal, 2703 one expects d_nz >> o_nz. For large problems you MUST preallocate memory 2704 or you will get TERRIBLE performance; see the users' manual chapter on 2705 matrices. 2706 2707 Level: intermediate 2708 2709 .keywords: matrix, block, aij, compressed row, sparse, parallel 2710 2711 .seealso: MatCreate(), MatCreateSeqBAIJ(), MatSetValues(), MatCreateMPIBAIJ(), MatMPIBAIJSetPreallocation(), MatMPIBAIJSetPreallocationCSR() 2712 @*/ 2713 PetscErrorCode PETSCMAT_DLLEXPORT MatCreateMPIBAIJ(MPI_Comm comm,PetscInt bs,PetscInt m,PetscInt n,PetscInt M,PetscInt N,PetscInt d_nz,const PetscInt d_nnz[],PetscInt o_nz,const PetscInt o_nnz[],Mat *A) 2714 { 2715 PetscErrorCode ierr; 2716 PetscMPIInt size; 2717 2718 PetscFunctionBegin; 2719 ierr = MatCreate(comm,A);CHKERRQ(ierr); 2720 ierr = MatSetSizes(*A,m,n,M,N);CHKERRQ(ierr); 2721 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 2722 if (size > 1) { 2723 ierr = MatSetType(*A,MATMPIBAIJ);CHKERRQ(ierr); 2724 ierr = MatMPIBAIJSetPreallocation(*A,bs,d_nz,d_nnz,o_nz,o_nnz);CHKERRQ(ierr); 2725 } else { 2726 ierr = MatSetType(*A,MATSEQBAIJ);CHKERRQ(ierr); 2727 ierr = MatSeqBAIJSetPreallocation(*A,bs,d_nz,d_nnz);CHKERRQ(ierr); 2728 } 2729 PetscFunctionReturn(0); 2730 } 2731 2732 #undef __FUNCT__ 2733 #define __FUNCT__ "MatDuplicate_MPIBAIJ" 2734 static PetscErrorCode MatDuplicate_MPIBAIJ(Mat matin,MatDuplicateOption cpvalues,Mat *newmat) 2735 { 2736 Mat mat; 2737 Mat_MPIBAIJ *a,*oldmat = (Mat_MPIBAIJ*)matin->data; 2738 PetscErrorCode ierr; 2739 PetscInt len=0; 2740 2741 PetscFunctionBegin; 2742 *newmat = 0; 2743 ierr = MatCreate(((PetscObject)matin)->comm,&mat);CHKERRQ(ierr); 2744 ierr = MatSetSizes(mat,matin->rmap->n,matin->cmap->n,matin->rmap->N,matin->cmap->N);CHKERRQ(ierr); 2745 ierr = MatSetType(mat,((PetscObject)matin)->type_name);CHKERRQ(ierr); 2746 ierr = PetscMemcpy(mat->ops,matin->ops,sizeof(struct _MatOps));CHKERRQ(ierr); 2747 2748 mat->factor = matin->factor; 2749 mat->preallocated = PETSC_TRUE; 2750 mat->assembled = PETSC_TRUE; 2751 mat->insertmode = NOT_SET_VALUES; 2752 2753 a = (Mat_MPIBAIJ*)mat->data; 2754 mat->rmap->bs = matin->rmap->bs; 2755 a->bs2 = oldmat->bs2; 2756 a->mbs = oldmat->mbs; 2757 a->nbs = oldmat->nbs; 2758 a->Mbs = oldmat->Mbs; 2759 a->Nbs = oldmat->Nbs; 2760 2761 ierr = PetscMapCopy(((PetscObject)matin)->comm,matin->rmap,mat->rmap);CHKERRQ(ierr); 2762 ierr = PetscMapCopy(((PetscObject)matin)->comm,matin->cmap,mat->cmap);CHKERRQ(ierr); 2763 2764 a->size = oldmat->size; 2765 a->rank = oldmat->rank; 2766 a->donotstash = oldmat->donotstash; 2767 a->roworiented = oldmat->roworiented; 2768 a->rowindices = 0; 2769 a->rowvalues = 0; 2770 a->getrowactive = PETSC_FALSE; 2771 a->barray = 0; 2772 a->rstartbs = oldmat->rstartbs; 2773 a->rendbs = oldmat->rendbs; 2774 a->cstartbs = oldmat->cstartbs; 2775 a->cendbs = oldmat->cendbs; 2776 2777 /* hash table stuff */ 2778 a->ht = 0; 2779 a->hd = 0; 2780 a->ht_size = 0; 2781 a->ht_flag = oldmat->ht_flag; 2782 a->ht_fact = oldmat->ht_fact; 2783 a->ht_total_ct = 0; 2784 a->ht_insert_ct = 0; 2785 2786 ierr = PetscMemcpy(a->rangebs,oldmat->rangebs,(a->size+1)*sizeof(PetscInt));CHKERRQ(ierr); 2787 ierr = MatStashCreate_Private(((PetscObject)matin)->comm,1,&mat->stash);CHKERRQ(ierr); 2788 ierr = MatStashCreate_Private(((PetscObject)matin)->comm,matin->rmap->bs,&mat->bstash);CHKERRQ(ierr); 2789 if (oldmat->colmap) { 2790 #if defined (PETSC_USE_CTABLE) 2791 ierr = PetscTableCreateCopy(oldmat->colmap,&a->colmap);CHKERRQ(ierr); 2792 #else 2793 ierr = PetscMalloc((a->Nbs)*sizeof(PetscInt),&a->colmap);CHKERRQ(ierr); 2794 ierr = PetscLogObjectMemory(mat,(a->Nbs)*sizeof(PetscInt));CHKERRQ(ierr); 2795 ierr = PetscMemcpy(a->colmap,oldmat->colmap,(a->Nbs)*sizeof(PetscInt));CHKERRQ(ierr); 2796 #endif 2797 } else a->colmap = 0; 2798 2799 if (oldmat->garray && (len = ((Mat_SeqBAIJ*)(oldmat->B->data))->nbs)) { 2800 ierr = PetscMalloc(len*sizeof(PetscInt),&a->garray);CHKERRQ(ierr); 2801 ierr = PetscLogObjectMemory(mat,len*sizeof(PetscInt));CHKERRQ(ierr); 2802 ierr = PetscMemcpy(a->garray,oldmat->garray,len*sizeof(PetscInt));CHKERRQ(ierr); 2803 } else a->garray = 0; 2804 2805 ierr = VecDuplicate(oldmat->lvec,&a->lvec);CHKERRQ(ierr); 2806 ierr = PetscLogObjectParent(mat,a->lvec);CHKERRQ(ierr); 2807 ierr = VecScatterCopy(oldmat->Mvctx,&a->Mvctx);CHKERRQ(ierr); 2808 ierr = PetscLogObjectParent(mat,a->Mvctx);CHKERRQ(ierr); 2809 2810 ierr = MatDuplicate(oldmat->A,cpvalues,&a->A);CHKERRQ(ierr); 2811 ierr = PetscLogObjectParent(mat,a->A);CHKERRQ(ierr); 2812 ierr = MatDuplicate(oldmat->B,cpvalues,&a->B);CHKERRQ(ierr); 2813 ierr = PetscLogObjectParent(mat,a->B);CHKERRQ(ierr); 2814 ierr = PetscFListDuplicate(((PetscObject)matin)->qlist,&((PetscObject)mat)->qlist);CHKERRQ(ierr); 2815 *newmat = mat; 2816 2817 PetscFunctionReturn(0); 2818 } 2819 2820 #include "petscsys.h" 2821 2822 #undef __FUNCT__ 2823 #define __FUNCT__ "MatLoad_MPIBAIJ" 2824 PetscErrorCode MatLoad_MPIBAIJ(PetscViewer viewer, const MatType type,Mat *newmat) 2825 { 2826 Mat A; 2827 PetscErrorCode ierr; 2828 int fd; 2829 PetscInt i,nz,j,rstart,rend; 2830 PetscScalar *vals,*buf; 2831 MPI_Comm comm = ((PetscObject)viewer)->comm; 2832 MPI_Status status; 2833 PetscMPIInt rank,size,maxnz; 2834 PetscInt header[4],*rowlengths = 0,M,N,m,*rowners,*cols; 2835 PetscInt *locrowlens = PETSC_NULL,*procsnz = PETSC_NULL,*browners = PETSC_NULL; 2836 PetscInt jj,*mycols,*ibuf,bs=1,Mbs,mbs,extra_rows,mmax; 2837 PetscMPIInt tag = ((PetscObject)viewer)->tag; 2838 PetscInt *dlens = PETSC_NULL,*odlens = PETSC_NULL,*mask = PETSC_NULL,*masked1 = PETSC_NULL,*masked2 = PETSC_NULL,rowcount,odcount; 2839 PetscInt dcount,kmax,k,nzcount,tmp,mend; 2840 2841 PetscFunctionBegin; 2842 ierr = PetscOptionsBegin(comm,PETSC_NULL,"Options for loading MPIBAIJ matrix 2","Mat");CHKERRQ(ierr); 2843 ierr = PetscOptionsInt("-matload_block_size","Set the blocksize used to store the matrix","MatLoad",bs,&bs,PETSC_NULL);CHKERRQ(ierr); 2844 ierr = PetscOptionsEnd();CHKERRQ(ierr); 2845 2846 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 2847 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 2848 if (!rank) { 2849 ierr = PetscViewerBinaryGetDescriptor(viewer,&fd);CHKERRQ(ierr); 2850 ierr = PetscBinaryRead(fd,(char *)header,4,PETSC_INT);CHKERRQ(ierr); 2851 if (header[0] != MAT_FILE_COOKIE) SETERRQ(PETSC_ERR_FILE_UNEXPECTED,"not matrix object"); 2852 } 2853 2854 ierr = MPI_Bcast(header+1,3,MPIU_INT,0,comm);CHKERRQ(ierr); 2855 M = header[1]; N = header[2]; 2856 2857 if (M != N) SETERRQ(PETSC_ERR_SUP,"Can only do square matrices"); 2858 2859 /* 2860 This code adds extra rows to make sure the number of rows is 2861 divisible by the blocksize 2862 */ 2863 Mbs = M/bs; 2864 extra_rows = bs - M + bs*Mbs; 2865 if (extra_rows == bs) extra_rows = 0; 2866 else Mbs++; 2867 if (extra_rows && !rank) { 2868 ierr = PetscInfo(viewer,"Padding loaded matrix to match blocksize\n");CHKERRQ(ierr); 2869 } 2870 2871 /* determine ownership of all rows */ 2872 mbs = Mbs/size + ((Mbs % size) > rank); 2873 m = mbs*bs; 2874 ierr = PetscMalloc2(size+1,PetscInt,&rowners,size+1,PetscInt,&browners);CHKERRQ(ierr); 2875 ierr = MPI_Allgather(&mbs,1,MPIU_INT,rowners+1,1,MPIU_INT,comm);CHKERRQ(ierr); 2876 2877 /* process 0 needs enough room for process with most rows */ 2878 if (!rank) { 2879 mmax = rowners[1]; 2880 for (i=2; i<size; i++) { 2881 mmax = PetscMax(mmax,rowners[i]); 2882 } 2883 mmax*=bs; 2884 } else mmax = m; 2885 2886 rowners[0] = 0; 2887 for (i=2; i<=size; i++) rowners[i] += rowners[i-1]; 2888 for (i=0; i<=size; i++) browners[i] = rowners[i]*bs; 2889 rstart = rowners[rank]; 2890 rend = rowners[rank+1]; 2891 2892 /* distribute row lengths to all processors */ 2893 ierr = PetscMalloc((mmax+1)*sizeof(PetscInt),&locrowlens);CHKERRQ(ierr); 2894 if (!rank) { 2895 mend = m; 2896 if (size == 1) mend = mend - extra_rows; 2897 ierr = PetscBinaryRead(fd,locrowlens,mend,PETSC_INT);CHKERRQ(ierr); 2898 for (j=mend; j<m; j++) locrowlens[j] = 1; 2899 ierr = PetscMalloc(m*sizeof(PetscInt),&rowlengths);CHKERRQ(ierr); 2900 ierr = PetscMalloc(size*sizeof(PetscInt),&procsnz);CHKERRQ(ierr); 2901 ierr = PetscMemzero(procsnz,size*sizeof(PetscInt));CHKERRQ(ierr); 2902 for (j=0; j<m; j++) { 2903 procsnz[0] += locrowlens[j]; 2904 } 2905 for (i=1; i<size; i++) { 2906 mend = browners[i+1] - browners[i]; 2907 if (i == size-1) mend = mend - extra_rows; 2908 ierr = PetscBinaryRead(fd,rowlengths,mend,PETSC_INT);CHKERRQ(ierr); 2909 for (j=mend; j<browners[i+1] - browners[i]; j++) rowlengths[j] = 1; 2910 /* calculate the number of nonzeros on each processor */ 2911 for (j=0; j<browners[i+1]-browners[i]; j++) { 2912 procsnz[i] += rowlengths[j]; 2913 } 2914 ierr = MPI_Send(rowlengths,browners[i+1]-browners[i],MPIU_INT,i,tag,comm);CHKERRQ(ierr); 2915 } 2916 ierr = PetscFree(rowlengths);CHKERRQ(ierr); 2917 } else { 2918 ierr = MPI_Recv(locrowlens,m,MPIU_INT,0,tag,comm,&status);CHKERRQ(ierr); 2919 } 2920 2921 if (!rank) { 2922 /* determine max buffer needed and allocate it */ 2923 maxnz = procsnz[0]; 2924 for (i=1; i<size; i++) { 2925 maxnz = PetscMax(maxnz,procsnz[i]); 2926 } 2927 ierr = PetscMalloc(maxnz*sizeof(PetscInt),&cols);CHKERRQ(ierr); 2928 2929 /* read in my part of the matrix column indices */ 2930 nz = procsnz[0]; 2931 ierr = PetscMalloc((nz+1)*sizeof(PetscInt),&ibuf);CHKERRQ(ierr); 2932 mycols = ibuf; 2933 if (size == 1) nz -= extra_rows; 2934 ierr = PetscBinaryRead(fd,mycols,nz,PETSC_INT);CHKERRQ(ierr); 2935 if (size == 1) for (i=0; i< extra_rows; i++) { mycols[nz+i] = M+i; } 2936 2937 /* read in every ones (except the last) and ship off */ 2938 for (i=1; i<size-1; i++) { 2939 nz = procsnz[i]; 2940 ierr = PetscBinaryRead(fd,cols,nz,PETSC_INT);CHKERRQ(ierr); 2941 ierr = MPI_Send(cols,nz,MPIU_INT,i,tag,comm);CHKERRQ(ierr); 2942 } 2943 /* read in the stuff for the last proc */ 2944 if (size != 1) { 2945 nz = procsnz[size-1] - extra_rows; /* the extra rows are not on the disk */ 2946 ierr = PetscBinaryRead(fd,cols,nz,PETSC_INT);CHKERRQ(ierr); 2947 for (i=0; i<extra_rows; i++) cols[nz+i] = M+i; 2948 ierr = MPI_Send(cols,nz+extra_rows,MPIU_INT,size-1,tag,comm);CHKERRQ(ierr); 2949 } 2950 ierr = PetscFree(cols);CHKERRQ(ierr); 2951 } else { 2952 /* determine buffer space needed for message */ 2953 nz = 0; 2954 for (i=0; i<m; i++) { 2955 nz += locrowlens[i]; 2956 } 2957 ierr = PetscMalloc((nz+1)*sizeof(PetscInt),&ibuf);CHKERRQ(ierr); 2958 mycols = ibuf; 2959 /* receive message of column indices*/ 2960 ierr = MPI_Recv(mycols,nz,MPIU_INT,0,tag,comm,&status);CHKERRQ(ierr); 2961 ierr = MPI_Get_count(&status,MPIU_INT,&maxnz);CHKERRQ(ierr); 2962 if (maxnz != nz) SETERRQ(PETSC_ERR_FILE_UNEXPECTED,"something is wrong with file"); 2963 } 2964 2965 /* loop over local rows, determining number of off diagonal entries */ 2966 ierr = PetscMalloc2(rend-rstart,PetscInt,&dlens,rend-rstart,PetscInt,&odlens);CHKERRQ(ierr); 2967 ierr = PetscMalloc3(Mbs,PetscInt,&mask,Mbs,PetscInt,&masked1,Mbs,PetscInt,&masked2);CHKERRQ(ierr); 2968 ierr = PetscMemzero(mask,Mbs*sizeof(PetscInt));CHKERRQ(ierr); 2969 ierr = PetscMemzero(masked1,Mbs*sizeof(PetscInt));CHKERRQ(ierr); 2970 ierr = PetscMemzero(masked2,Mbs*sizeof(PetscInt));CHKERRQ(ierr); 2971 rowcount = 0; nzcount = 0; 2972 for (i=0; i<mbs; i++) { 2973 dcount = 0; 2974 odcount = 0; 2975 for (j=0; j<bs; j++) { 2976 kmax = locrowlens[rowcount]; 2977 for (k=0; k<kmax; k++) { 2978 tmp = mycols[nzcount++]/bs; 2979 if (!mask[tmp]) { 2980 mask[tmp] = 1; 2981 if (tmp < rstart || tmp >= rend) masked2[odcount++] = tmp; 2982 else masked1[dcount++] = tmp; 2983 } 2984 } 2985 rowcount++; 2986 } 2987 2988 dlens[i] = dcount; 2989 odlens[i] = odcount; 2990 2991 /* zero out the mask elements we set */ 2992 for (j=0; j<dcount; j++) mask[masked1[j]] = 0; 2993 for (j=0; j<odcount; j++) mask[masked2[j]] = 0; 2994 } 2995 2996 /* create our matrix */ 2997 ierr = MatCreate(comm,&A);CHKERRQ(ierr); 2998 ierr = MatSetSizes(A,m,m,M+extra_rows,N+extra_rows);CHKERRQ(ierr); 2999 ierr = MatSetType(A,type);CHKERRQ(ierr) 3000 ierr = MatMPIBAIJSetPreallocation(A,bs,0,dlens,0,odlens);CHKERRQ(ierr); 3001 3002 if (!rank) { 3003 ierr = PetscMalloc((maxnz+1)*sizeof(PetscScalar),&buf);CHKERRQ(ierr); 3004 /* read in my part of the matrix numerical values */ 3005 nz = procsnz[0]; 3006 vals = buf; 3007 mycols = ibuf; 3008 if (size == 1) nz -= extra_rows; 3009 ierr = PetscBinaryRead(fd,vals,nz,PETSC_SCALAR);CHKERRQ(ierr); 3010 if (size == 1) for (i=0; i< extra_rows; i++) { vals[nz+i] = 1.0; } 3011 3012 /* insert into matrix */ 3013 jj = rstart*bs; 3014 for (i=0; i<m; i++) { 3015 ierr = MatSetValues_MPIBAIJ(A,1,&jj,locrowlens[i],mycols,vals,INSERT_VALUES);CHKERRQ(ierr); 3016 mycols += locrowlens[i]; 3017 vals += locrowlens[i]; 3018 jj++; 3019 } 3020 /* read in other processors (except the last one) and ship out */ 3021 for (i=1; i<size-1; i++) { 3022 nz = procsnz[i]; 3023 vals = buf; 3024 ierr = PetscBinaryRead(fd,vals,nz,PETSC_SCALAR);CHKERRQ(ierr); 3025 ierr = MPI_Send(vals,nz,MPIU_SCALAR,i,((PetscObject)A)->tag,comm);CHKERRQ(ierr); 3026 } 3027 /* the last proc */ 3028 if (size != 1){ 3029 nz = procsnz[i] - extra_rows; 3030 vals = buf; 3031 ierr = PetscBinaryRead(fd,vals,nz,PETSC_SCALAR);CHKERRQ(ierr); 3032 for (i=0; i<extra_rows; i++) vals[nz+i] = 1.0; 3033 ierr = MPI_Send(vals,nz+extra_rows,MPIU_SCALAR,size-1,((PetscObject)A)->tag,comm);CHKERRQ(ierr); 3034 } 3035 ierr = PetscFree(procsnz);CHKERRQ(ierr); 3036 } else { 3037 /* receive numeric values */ 3038 ierr = PetscMalloc((nz+1)*sizeof(PetscScalar),&buf);CHKERRQ(ierr); 3039 3040 /* receive message of values*/ 3041 vals = buf; 3042 mycols = ibuf; 3043 ierr = MPI_Recv(vals,nz,MPIU_SCALAR,0,((PetscObject)A)->tag,comm,&status);CHKERRQ(ierr); 3044 ierr = MPI_Get_count(&status,MPIU_SCALAR,&maxnz);CHKERRQ(ierr); 3045 if (maxnz != nz) SETERRQ(PETSC_ERR_FILE_UNEXPECTED,"something is wrong with file"); 3046 3047 /* insert into matrix */ 3048 jj = rstart*bs; 3049 for (i=0; i<m; i++) { 3050 ierr = MatSetValues_MPIBAIJ(A,1,&jj,locrowlens[i],mycols,vals,INSERT_VALUES);CHKERRQ(ierr); 3051 mycols += locrowlens[i]; 3052 vals += locrowlens[i]; 3053 jj++; 3054 } 3055 } 3056 ierr = PetscFree(locrowlens);CHKERRQ(ierr); 3057 ierr = PetscFree(buf);CHKERRQ(ierr); 3058 ierr = PetscFree(ibuf);CHKERRQ(ierr); 3059 ierr = PetscFree2(rowners,browners);CHKERRQ(ierr); 3060 ierr = PetscFree2(dlens,odlens);CHKERRQ(ierr); 3061 ierr = PetscFree3(mask,masked1,masked2);CHKERRQ(ierr); 3062 ierr = MatAssemblyBegin(A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3063 ierr = MatAssemblyEnd(A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3064 3065 *newmat = A; 3066 PetscFunctionReturn(0); 3067 } 3068 3069 #undef __FUNCT__ 3070 #define __FUNCT__ "MatMPIBAIJSetHashTableFactor" 3071 /*@ 3072 MatMPIBAIJSetHashTableFactor - Sets the factor required to compute the size of the HashTable. 3073 3074 Input Parameters: 3075 . mat - the matrix 3076 . fact - factor 3077 3078 Collective on Mat 3079 3080 Level: advanced 3081 3082 Notes: 3083 This can also be set by the command line option: -mat_use_hash_table <fact> 3084 3085 .keywords: matrix, hashtable, factor, HT 3086 3087 .seealso: MatSetOption() 3088 @*/ 3089 PetscErrorCode PETSCMAT_DLLEXPORT MatMPIBAIJSetHashTableFactor(Mat mat,PetscReal fact) 3090 { 3091 PetscErrorCode ierr,(*f)(Mat,PetscReal); 3092 3093 PetscFunctionBegin; 3094 ierr = PetscObjectQueryFunction((PetscObject)mat,"MatSetHashTableFactor_C",(void (**)(void))&f);CHKERRQ(ierr); 3095 if (f) { 3096 ierr = (*f)(mat,fact);CHKERRQ(ierr); 3097 } 3098 PetscFunctionReturn(0); 3099 } 3100 3101 EXTERN_C_BEGIN 3102 #undef __FUNCT__ 3103 #define __FUNCT__ "MatSetHashTableFactor_MPIBAIJ" 3104 PetscErrorCode PETSCMAT_DLLEXPORT MatSetHashTableFactor_MPIBAIJ(Mat mat,PetscReal fact) 3105 { 3106 Mat_MPIBAIJ *baij; 3107 3108 PetscFunctionBegin; 3109 baij = (Mat_MPIBAIJ*)mat->data; 3110 baij->ht_fact = fact; 3111 PetscFunctionReturn(0); 3112 } 3113 EXTERN_C_END 3114 3115 #undef __FUNCT__ 3116 #define __FUNCT__ "MatMPIBAIJGetSeqBAIJ" 3117 PetscErrorCode PETSCMAT_DLLEXPORT MatMPIBAIJGetSeqBAIJ(Mat A,Mat *Ad,Mat *Ao,PetscInt *colmap[]) 3118 { 3119 Mat_MPIBAIJ *a = (Mat_MPIBAIJ *)A->data; 3120 PetscFunctionBegin; 3121 *Ad = a->A; 3122 *Ao = a->B; 3123 *colmap = a->garray; 3124 PetscFunctionReturn(0); 3125 } 3126 3127 /* 3128 Special version for direct calls from Fortran (to eliminate two function call overheads 3129 */ 3130 #if defined(PETSC_HAVE_FORTRAN_CAPS) 3131 #define matmpibaijsetvaluesblocked_ MATMPIBAIJSETVALUESBLOCKED 3132 #elif !defined(PETSC_HAVE_FORTRAN_UNDERSCORE) 3133 #define matmpibaijsetvaluesblocked_ matmpibaijsetvaluesblocked 3134 #endif 3135 3136 #undef __FUNCT__ 3137 #define __FUNCT__ "matmpibiajsetvaluesblocked" 3138 /*@C 3139 MatMPIBAIJSetValuesBlocked - Direct Fortran call to replace call to MatSetValuesBlocked() 3140 3141 Collective on Mat 3142 3143 Input Parameters: 3144 + mat - the matrix 3145 . min - number of input rows 3146 . im - input rows 3147 . nin - number of input columns 3148 . in - input columns 3149 . v - numerical values input 3150 - addvin - INSERT_VALUES or ADD_VALUES 3151 3152 Notes: This has a complete copy of MatSetValuesBlocked_MPIBAIJ() which is terrible code un-reuse. 3153 3154 Level: advanced 3155 3156 .seealso: MatSetValuesBlocked() 3157 @*/ 3158 PetscErrorCode matmpibaijsetvaluesblocked_(Mat *matin,PetscInt *min,const PetscInt im[],PetscInt *nin,const PetscInt in[],const MatScalar v[],InsertMode *addvin) 3159 { 3160 /* convert input arguments to C version */ 3161 Mat mat = *matin; 3162 PetscInt m = *min, n = *nin; 3163 InsertMode addv = *addvin; 3164 3165 Mat_MPIBAIJ *baij = (Mat_MPIBAIJ*)mat->data; 3166 const MatScalar *value; 3167 MatScalar *barray=baij->barray; 3168 PetscTruth roworiented = baij->roworiented; 3169 PetscErrorCode ierr; 3170 PetscInt i,j,ii,jj,row,col,rstart=baij->rstartbs; 3171 PetscInt rend=baij->rendbs,cstart=baij->cstartbs,stepval; 3172 PetscInt cend=baij->cendbs,bs=mat->rmap->bs,bs2=baij->bs2; 3173 3174 PetscFunctionBegin; 3175 /* tasks normally handled by MatSetValuesBlocked() */ 3176 if (mat->insertmode == NOT_SET_VALUES) { 3177 mat->insertmode = addv; 3178 } 3179 #if defined(PETSC_USE_DEBUG) 3180 else if (mat->insertmode != addv) { 3181 SETERRQ(PETSC_ERR_ARG_WRONGSTATE,"Cannot mix add values and insert values"); 3182 } 3183 if (mat->factor) SETERRQ(PETSC_ERR_ARG_WRONGSTATE,"Not for factored matrix"); 3184 #endif 3185 if (mat->assembled) { 3186 mat->was_assembled = PETSC_TRUE; 3187 mat->assembled = PETSC_FALSE; 3188 } 3189 ierr = PetscLogEventBegin(MAT_SetValues,mat,0,0,0);CHKERRQ(ierr); 3190 3191 3192 if(!barray) { 3193 ierr = PetscMalloc(bs2*sizeof(MatScalar),&barray);CHKERRQ(ierr); 3194 baij->barray = barray; 3195 } 3196 3197 if (roworiented) { 3198 stepval = (n-1)*bs; 3199 } else { 3200 stepval = (m-1)*bs; 3201 } 3202 for (i=0; i<m; i++) { 3203 if (im[i] < 0) continue; 3204 #if defined(PETSC_USE_DEBUG) 3205 if (im[i] >= baij->Mbs) SETERRQ2(PETSC_ERR_ARG_OUTOFRANGE,"Row too large, row %D max %D",im[i],baij->Mbs-1); 3206 #endif 3207 if (im[i] >= rstart && im[i] < rend) { 3208 row = im[i] - rstart; 3209 for (j=0; j<n; j++) { 3210 /* If NumCol = 1 then a copy is not required */ 3211 if ((roworiented) && (n == 1)) { 3212 barray = (MatScalar*)v + i*bs2; 3213 } else if((!roworiented) && (m == 1)) { 3214 barray = (MatScalar*)v + j*bs2; 3215 } else { /* Here a copy is required */ 3216 if (roworiented) { 3217 value = v + i*(stepval+bs)*bs + j*bs; 3218 } else { 3219 value = v + j*(stepval+bs)*bs + i*bs; 3220 } 3221 for (ii=0; ii<bs; ii++,value+=stepval) { 3222 for (jj=0; jj<bs; jj++) { 3223 *barray++ = *value++; 3224 } 3225 } 3226 barray -=bs2; 3227 } 3228 3229 if (in[j] >= cstart && in[j] < cend){ 3230 col = in[j] - cstart; 3231 ierr = MatSetValuesBlocked_SeqBAIJ(baij->A,1,&row,1,&col,barray,addv);CHKERRQ(ierr); 3232 } 3233 else if (in[j] < 0) continue; 3234 #if defined(PETSC_USE_DEBUG) 3235 else if (in[j] >= baij->Nbs) {SETERRQ2(PETSC_ERR_ARG_OUTOFRANGE,"Column too large, col %D max %D",in[j],baij->Nbs-1);} 3236 #endif 3237 else { 3238 if (mat->was_assembled) { 3239 if (!baij->colmap) { 3240 ierr = CreateColmap_MPIBAIJ_Private(mat);CHKERRQ(ierr); 3241 } 3242 3243 #if defined(PETSC_USE_DEBUG) 3244 #if defined (PETSC_USE_CTABLE) 3245 { PetscInt data; 3246 ierr = PetscTableFind(baij->colmap,in[j]+1,&data);CHKERRQ(ierr); 3247 if ((data - 1) % bs) SETERRQ(PETSC_ERR_PLIB,"Incorrect colmap"); 3248 } 3249 #else 3250 if ((baij->colmap[in[j]] - 1) % bs) SETERRQ(PETSC_ERR_PLIB,"Incorrect colmap"); 3251 #endif 3252 #endif 3253 #if defined (PETSC_USE_CTABLE) 3254 ierr = PetscTableFind(baij->colmap,in[j]+1,&col);CHKERRQ(ierr); 3255 col = (col - 1)/bs; 3256 #else 3257 col = (baij->colmap[in[j]] - 1)/bs; 3258 #endif 3259 if (col < 0 && !((Mat_SeqBAIJ*)(baij->A->data))->nonew) { 3260 ierr = DisAssemble_MPIBAIJ(mat);CHKERRQ(ierr); 3261 col = in[j]; 3262 } 3263 } 3264 else col = in[j]; 3265 ierr = MatSetValuesBlocked_SeqBAIJ(baij->B,1,&row,1,&col,barray,addv);CHKERRQ(ierr); 3266 } 3267 } 3268 } else { 3269 if (!baij->donotstash) { 3270 if (roworiented) { 3271 ierr = MatStashValuesRowBlocked_Private(&mat->bstash,im[i],n,in,v,m,n,i);CHKERRQ(ierr); 3272 } else { 3273 ierr = MatStashValuesColBlocked_Private(&mat->bstash,im[i],n,in,v,m,n,i);CHKERRQ(ierr); 3274 } 3275 } 3276 } 3277 } 3278 3279 /* task normally handled by MatSetValuesBlocked() */ 3280 ierr = PetscLogEventEnd(MAT_SetValues,mat,0,0,0);CHKERRQ(ierr); 3281 PetscFunctionReturn(0); 3282 } 3283