1 2 #include <../src/mat/impls/baij/mpi/mpibaij.h> /*I "petscmat.h" I*/ 3 #include <petscblaslapack.h> 4 5 extern PetscErrorCode MatSetUpMultiply_MPIBAIJ(Mat); 6 extern PetscErrorCode MatDisAssemble_MPIBAIJ(Mat); 7 extern PetscErrorCode MatGetValues_SeqBAIJ(Mat,PetscInt,const PetscInt[],PetscInt,const PetscInt [],PetscScalar []); 8 extern PetscErrorCode MatSetValues_SeqBAIJ(Mat,PetscInt,const PetscInt[],PetscInt,const PetscInt [],const PetscScalar [],InsertMode); 9 extern PetscErrorCode MatSetValuesBlocked_SeqBAIJ(Mat,PetscInt,const PetscInt[],PetscInt,const PetscInt[],const PetscScalar[],InsertMode); 10 extern PetscErrorCode MatGetRow_SeqBAIJ(Mat,PetscInt,PetscInt*,PetscInt*[],PetscScalar*[]); 11 extern PetscErrorCode MatRestoreRow_SeqBAIJ(Mat,PetscInt,PetscInt*,PetscInt*[],PetscScalar*[]); 12 extern PetscErrorCode MatZeroRows_SeqBAIJ(Mat,PetscInt,const PetscInt[],PetscScalar,Vec,Vec); 13 14 #undef __FUNCT__ 15 #define __FUNCT__ "MatGetRowMaxAbs_MPIBAIJ" 16 PetscErrorCode MatGetRowMaxAbs_MPIBAIJ(Mat A,Vec v,PetscInt idx[]) 17 { 18 Mat_MPIBAIJ *a = (Mat_MPIBAIJ*)A->data; 19 PetscErrorCode ierr; 20 PetscInt i,*idxb = 0; 21 PetscScalar *va,*vb; 22 Vec vtmp; 23 24 PetscFunctionBegin; 25 ierr = MatGetRowMaxAbs(a->A,v,idx);CHKERRQ(ierr); 26 ierr = VecGetArray(v,&va);CHKERRQ(ierr); 27 if (idx) { 28 for (i=0; i<A->rmap->n; i++) {if (PetscAbsScalar(va[i])) idx[i] += A->cmap->rstart;} 29 } 30 31 ierr = VecCreateSeq(PETSC_COMM_SELF,A->rmap->n,&vtmp);CHKERRQ(ierr); 32 if (idx) {ierr = PetscMalloc(A->rmap->n*sizeof(PetscInt),&idxb);CHKERRQ(ierr);} 33 ierr = MatGetRowMaxAbs(a->B,vtmp,idxb);CHKERRQ(ierr); 34 ierr = VecGetArray(vtmp,&vb);CHKERRQ(ierr); 35 36 for (i=0; i<A->rmap->n; i++){ 37 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);} 38 } 39 40 ierr = VecRestoreArray(v,&va);CHKERRQ(ierr); 41 ierr = VecRestoreArray(vtmp,&vb);CHKERRQ(ierr); 42 ierr = PetscFree(idxb);CHKERRQ(ierr); 43 ierr = VecDestroy(&vtmp);CHKERRQ(ierr); 44 PetscFunctionReturn(0); 45 } 46 47 EXTERN_C_BEGIN 48 #undef __FUNCT__ 49 #define __FUNCT__ "MatStoreValues_MPIBAIJ" 50 PetscErrorCode MatStoreValues_MPIBAIJ(Mat mat) 51 { 52 Mat_MPIBAIJ *aij = (Mat_MPIBAIJ *)mat->data; 53 PetscErrorCode ierr; 54 55 PetscFunctionBegin; 56 ierr = MatStoreValues(aij->A);CHKERRQ(ierr); 57 ierr = MatStoreValues(aij->B);CHKERRQ(ierr); 58 PetscFunctionReturn(0); 59 } 60 EXTERN_C_END 61 62 EXTERN_C_BEGIN 63 #undef __FUNCT__ 64 #define __FUNCT__ "MatRetrieveValues_MPIBAIJ" 65 PetscErrorCode MatRetrieveValues_MPIBAIJ(Mat mat) 66 { 67 Mat_MPIBAIJ *aij = (Mat_MPIBAIJ *)mat->data; 68 PetscErrorCode ierr; 69 70 PetscFunctionBegin; 71 ierr = MatRetrieveValues(aij->A);CHKERRQ(ierr); 72 ierr = MatRetrieveValues(aij->B);CHKERRQ(ierr); 73 PetscFunctionReturn(0); 74 } 75 EXTERN_C_END 76 77 /* 78 Local utility routine that creates a mapping from the global column 79 number to the local number in the off-diagonal part of the local 80 storage of the matrix. This is done in a non scalable way since the 81 length of colmap equals the global matrix length. 82 */ 83 #undef __FUNCT__ 84 #define __FUNCT__ "MatCreateColmap_MPIBAIJ_Private" 85 PetscErrorCode MatCreateColmap_MPIBAIJ_Private(Mat mat) 86 { 87 Mat_MPIBAIJ *baij = (Mat_MPIBAIJ*)mat->data; 88 Mat_SeqBAIJ *B = (Mat_SeqBAIJ*)baij->B->data; 89 PetscErrorCode ierr; 90 PetscInt nbs = B->nbs,i,bs=mat->rmap->bs; 91 92 PetscFunctionBegin; 93 #if defined (PETSC_USE_CTABLE) 94 ierr = PetscTableCreate(baij->nbs,baij->Nbs+1,&baij->colmap);CHKERRQ(ierr); 95 for (i=0; i<nbs; i++){ 96 ierr = PetscTableAdd(baij->colmap,baij->garray[i]+1,i*bs+1,INSERT_VALUES);CHKERRQ(ierr); 97 } 98 #else 99 ierr = PetscMalloc((baij->Nbs+1)*sizeof(PetscInt),&baij->colmap);CHKERRQ(ierr); 100 ierr = PetscLogObjectMemory(mat,baij->Nbs*sizeof(PetscInt));CHKERRQ(ierr); 101 ierr = PetscMemzero(baij->colmap,baij->Nbs*sizeof(PetscInt));CHKERRQ(ierr); 102 for (i=0; i<nbs; i++) baij->colmap[baij->garray[i]] = i*bs+1; 103 #endif 104 PetscFunctionReturn(0); 105 } 106 107 #define MatSetValues_SeqBAIJ_A_Private(row,col,value,addv) \ 108 { \ 109 \ 110 brow = row/bs; \ 111 rp = aj + ai[brow]; ap = aa + bs2*ai[brow]; \ 112 rmax = aimax[brow]; nrow = ailen[brow]; \ 113 bcol = col/bs; \ 114 ridx = row % bs; cidx = col % bs; \ 115 low = 0; high = nrow; \ 116 while (high-low > 3) { \ 117 t = (low+high)/2; \ 118 if (rp[t] > bcol) high = t; \ 119 else low = t; \ 120 } \ 121 for (_i=low; _i<high; _i++) { \ 122 if (rp[_i] > bcol) break; \ 123 if (rp[_i] == bcol) { \ 124 bap = ap + bs2*_i + bs*cidx + ridx; \ 125 if (addv == ADD_VALUES) *bap += value; \ 126 else *bap = value; \ 127 goto a_noinsert; \ 128 } \ 129 } \ 130 if (a->nonew == 1) goto a_noinsert; \ 131 if (a->nonew == -1) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new nonzero (%D, %D) into matrix", row, col); \ 132 MatSeqXAIJReallocateAIJ(A,a->mbs,bs2,nrow,brow,bcol,rmax,aa,ai,aj,rp,ap,aimax,a->nonew,MatScalar); \ 133 N = nrow++ - 1; \ 134 /* shift up all the later entries in this row */ \ 135 for (ii=N; ii>=_i; ii--) { \ 136 rp[ii+1] = rp[ii]; \ 137 ierr = PetscMemcpy(ap+bs2*(ii+1),ap+bs2*(ii),bs2*sizeof(MatScalar));CHKERRQ(ierr); \ 138 } \ 139 if (N>=_i) { ierr = PetscMemzero(ap+bs2*_i,bs2*sizeof(MatScalar));CHKERRQ(ierr); } \ 140 rp[_i] = bcol; \ 141 ap[bs2*_i + bs*cidx + ridx] = value; \ 142 a_noinsert:; \ 143 ailen[brow] = nrow; \ 144 } 145 146 #define MatSetValues_SeqBAIJ_B_Private(row,col,value,addv) \ 147 { \ 148 brow = row/bs; \ 149 rp = bj + bi[brow]; ap = ba + bs2*bi[brow]; \ 150 rmax = bimax[brow]; nrow = bilen[brow]; \ 151 bcol = col/bs; \ 152 ridx = row % bs; cidx = col % bs; \ 153 low = 0; high = nrow; \ 154 while (high-low > 3) { \ 155 t = (low+high)/2; \ 156 if (rp[t] > bcol) high = t; \ 157 else low = t; \ 158 } \ 159 for (_i=low; _i<high; _i++) { \ 160 if (rp[_i] > bcol) break; \ 161 if (rp[_i] == bcol) { \ 162 bap = ap + bs2*_i + bs*cidx + ridx; \ 163 if (addv == ADD_VALUES) *bap += value; \ 164 else *bap = value; \ 165 goto b_noinsert; \ 166 } \ 167 } \ 168 if (b->nonew == 1) goto b_noinsert; \ 169 if (b->nonew == -1) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new nonzero (%D, %D) into matrix", row, col); \ 170 MatSeqXAIJReallocateAIJ(B,b->mbs,bs2,nrow,brow,bcol,rmax,ba,bi,bj,rp,ap,bimax,b->nonew,MatScalar); \ 171 CHKMEMQ;\ 172 N = nrow++ - 1; \ 173 /* shift up all the later entries in this row */ \ 174 for (ii=N; ii>=_i; ii--) { \ 175 rp[ii+1] = rp[ii]; \ 176 ierr = PetscMemcpy(ap+bs2*(ii+1),ap+bs2*(ii),bs2*sizeof(MatScalar));CHKERRQ(ierr); \ 177 } \ 178 if (N>=_i) { ierr = PetscMemzero(ap+bs2*_i,bs2*sizeof(MatScalar));CHKERRQ(ierr);} \ 179 rp[_i] = bcol; \ 180 ap[bs2*_i + bs*cidx + ridx] = value; \ 181 b_noinsert:; \ 182 bilen[brow] = nrow; \ 183 } 184 185 #undef __FUNCT__ 186 #define __FUNCT__ "MatSetValues_MPIBAIJ" 187 PetscErrorCode MatSetValues_MPIBAIJ(Mat mat,PetscInt m,const PetscInt im[],PetscInt n,const PetscInt in[],const PetscScalar v[],InsertMode addv) 188 { 189 Mat_MPIBAIJ *baij = (Mat_MPIBAIJ*)mat->data; 190 MatScalar value; 191 PetscBool roworiented = baij->roworiented; 192 PetscErrorCode ierr; 193 PetscInt i,j,row,col; 194 PetscInt rstart_orig=mat->rmap->rstart; 195 PetscInt rend_orig=mat->rmap->rend,cstart_orig=mat->cmap->rstart; 196 PetscInt cend_orig=mat->cmap->rend,bs=mat->rmap->bs; 197 198 /* Some Variables required in the macro */ 199 Mat A = baij->A; 200 Mat_SeqBAIJ *a = (Mat_SeqBAIJ*)(A)->data; 201 PetscInt *aimax=a->imax,*ai=a->i,*ailen=a->ilen,*aj=a->j; 202 MatScalar *aa=a->a; 203 204 Mat B = baij->B; 205 Mat_SeqBAIJ *b = (Mat_SeqBAIJ*)(B)->data; 206 PetscInt *bimax=b->imax,*bi=b->i,*bilen=b->ilen,*bj=b->j; 207 MatScalar *ba=b->a; 208 209 PetscInt *rp,ii,nrow,_i,rmax,N,brow,bcol; 210 PetscInt low,high,t,ridx,cidx,bs2=a->bs2; 211 MatScalar *ap,*bap; 212 213 PetscFunctionBegin; 214 if (v) PetscValidScalarPointer(v,6); 215 for (i=0; i<m; i++) { 216 if (im[i] < 0) continue; 217 #if defined(PETSC_USE_DEBUG) 218 if (im[i] >= mat->rmap->N) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Row too large: row %D max %D",im[i],mat->rmap->N-1); 219 #endif 220 if (im[i] >= rstart_orig && im[i] < rend_orig) { 221 row = im[i] - rstart_orig; 222 for (j=0; j<n; j++) { 223 if (in[j] >= cstart_orig && in[j] < cend_orig){ 224 col = in[j] - cstart_orig; 225 if (roworiented) value = v[i*n+j]; else value = v[i+j*m]; 226 MatSetValues_SeqBAIJ_A_Private(row,col,value,addv); 227 /* ierr = MatSetValues_SeqBAIJ(baij->A,1,&row,1,&col,&value,addv);CHKERRQ(ierr); */ 228 } else if (in[j] < 0) continue; 229 #if defined(PETSC_USE_DEBUG) 230 else if (in[j] >= mat->cmap->N) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Column too large: col %D max %D",in[j],mat->cmap->N-1); 231 #endif 232 else { 233 if (mat->was_assembled) { 234 if (!baij->colmap) { 235 ierr = MatCreateColmap_MPIBAIJ_Private(mat);CHKERRQ(ierr); 236 } 237 #if defined (PETSC_USE_CTABLE) 238 ierr = PetscTableFind(baij->colmap,in[j]/bs + 1,&col);CHKERRQ(ierr); 239 col = col - 1; 240 #else 241 col = baij->colmap[in[j]/bs] - 1; 242 #endif 243 if (col < 0 && !((Mat_SeqBAIJ*)(baij->B->data))->nonew) { 244 ierr = MatDisAssemble_MPIBAIJ(mat);CHKERRQ(ierr); 245 col = in[j]; 246 /* Reinitialize the variables required by MatSetValues_SeqBAIJ_B_Private() */ 247 B = baij->B; 248 b = (Mat_SeqBAIJ*)(B)->data; 249 bimax=b->imax;bi=b->i;bilen=b->ilen;bj=b->j; 250 ba=b->a; 251 } else if (col < 0) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new nonzero (%D, %D) into matrix", im[i], in[j]); 252 else col += in[j]%bs; 253 } else col = in[j]; 254 if (roworiented) value = v[i*n+j]; else value = v[i+j*m]; 255 MatSetValues_SeqBAIJ_B_Private(row,col,value,addv); 256 /* ierr = MatSetValues_SeqBAIJ(baij->B,1,&row,1,&col,&value,addv);CHKERRQ(ierr); */ 257 } 258 } 259 } else { 260 if (mat->nooffprocentries) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Setting off process row %D even though MatSetOption(,MAT_NO_OFF_PROC_ENTRIES,PETSC_TRUE) was set",im[i]); 261 if (!baij->donotstash) { 262 mat->assembled = PETSC_FALSE; 263 if (roworiented) { 264 ierr = MatStashValuesRow_Private(&mat->stash,im[i],n,in,v+i*n,PETSC_FALSE);CHKERRQ(ierr); 265 } else { 266 ierr = MatStashValuesCol_Private(&mat->stash,im[i],n,in,v+i,m,PETSC_FALSE);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 PetscBool 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_COMM_SELF,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) + j)*bs; 314 } else { 315 value = v + (j*(stepval+bs) + i)*bs; 316 } 317 for (ii=0; ii<bs; ii++,value+=bs+stepval) { 318 for (jj=0; jj<bs; jj++) { 319 barray[jj] = value[jj]; 320 } 321 barray += bs; 322 } 323 barray -= bs2; 324 } 325 326 if (in[j] >= cstart && in[j] < cend){ 327 col = in[j] - cstart; 328 ierr = MatSetValuesBlocked_SeqBAIJ(baij->A,1,&row,1,&col,barray,addv);CHKERRQ(ierr); 329 } 330 else if (in[j] < 0) continue; 331 #if defined(PETSC_USE_DEBUG) 332 else if (in[j] >= baij->Nbs) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Column too large, col %D max %D",in[j],baij->Nbs-1); 333 #endif 334 else { 335 if (mat->was_assembled) { 336 if (!baij->colmap) { 337 ierr = MatCreateColmap_MPIBAIJ_Private(mat);CHKERRQ(ierr); 338 } 339 340 #if defined(PETSC_USE_DEBUG) 341 #if defined (PETSC_USE_CTABLE) 342 { PetscInt data; 343 ierr = PetscTableFind(baij->colmap,in[j]+1,&data);CHKERRQ(ierr); 344 if ((data - 1) % bs) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Incorrect colmap"); 345 } 346 #else 347 if ((baij->colmap[in[j]] - 1) % bs) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Incorrect colmap"); 348 #endif 349 #endif 350 #if defined (PETSC_USE_CTABLE) 351 ierr = PetscTableFind(baij->colmap,in[j]+1,&col);CHKERRQ(ierr); 352 col = (col - 1)/bs; 353 #else 354 col = (baij->colmap[in[j]] - 1)/bs; 355 #endif 356 if (col < 0 && !((Mat_SeqBAIJ*)(baij->B->data))->nonew) { 357 ierr = MatDisAssemble_MPIBAIJ(mat);CHKERRQ(ierr); 358 col = in[j]; 359 } else if (col < 0) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new nonzero (%D, %D) into matrix", bs*im[i], bs*in[j]); 360 } 361 else col = in[j]; 362 ierr = MatSetValuesBlocked_SeqBAIJ(baij->B,1,&row,1,&col,barray,addv);CHKERRQ(ierr); 363 } 364 } 365 } else { 366 if (mat->nooffprocentries) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Setting off process row %D even though MatSetOption(,MAT_NO_OFF_PROC_ENTRIES,PETSC_TRUE) was set",im[i]); 367 if (!baij->donotstash) { 368 if (roworiented) { 369 ierr = MatStashValuesRowBlocked_Private(&mat->bstash,im[i],n,in,v,m,n,i);CHKERRQ(ierr); 370 } else { 371 ierr = MatStashValuesColBlocked_Private(&mat->bstash,im[i],n,in,v,m,n,i);CHKERRQ(ierr); 372 } 373 } 374 } 375 } 376 PetscFunctionReturn(0); 377 } 378 379 #define HASH_KEY 0.6180339887 380 #define HASH(size,key,tmp) (tmp = (key)*HASH_KEY,(PetscInt)((size)*(tmp-(PetscInt)tmp))) 381 /* #define HASH(size,key) ((PetscInt)((size)*fmod(((key)*HASH_KEY),1))) */ 382 /* #define HASH(size,key,tmp) ((PetscInt)((size)*fmod(((key)*HASH_KEY),1))) */ 383 #undef __FUNCT__ 384 #define __FUNCT__ "MatSetValues_MPIBAIJ_HT" 385 PetscErrorCode MatSetValues_MPIBAIJ_HT(Mat mat,PetscInt m,const PetscInt im[],PetscInt n,const PetscInt in[],const PetscScalar v[],InsertMode addv) 386 { 387 Mat_MPIBAIJ *baij = (Mat_MPIBAIJ*)mat->data; 388 PetscBool roworiented = baij->roworiented; 389 PetscErrorCode ierr; 390 PetscInt i,j,row,col; 391 PetscInt rstart_orig=mat->rmap->rstart; 392 PetscInt rend_orig=mat->rmap->rend,Nbs=baij->Nbs; 393 PetscInt h1,key,size=baij->ht_size,bs=mat->rmap->bs,*HT=baij->ht,idx; 394 PetscReal tmp; 395 MatScalar **HD = baij->hd,value; 396 #if defined(PETSC_USE_DEBUG) 397 PetscInt total_ct=baij->ht_total_ct,insert_ct=baij->ht_insert_ct; 398 #endif 399 400 PetscFunctionBegin; 401 if (v) PetscValidScalarPointer(v,6); 402 for (i=0; i<m; i++) { 403 #if defined(PETSC_USE_DEBUG) 404 if (im[i] < 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Negative row"); 405 if (im[i] >= mat->rmap->N) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Row too large: row %D max %D",im[i],mat->rmap->N-1); 406 #endif 407 row = im[i]; 408 if (row >= rstart_orig && row < rend_orig) { 409 for (j=0; j<n; j++) { 410 col = in[j]; 411 if (roworiented) value = v[i*n+j]; else value = v[i+j*m]; 412 /* Look up PetscInto the Hash Table */ 413 key = (row/bs)*Nbs+(col/bs)+1; 414 h1 = HASH(size,key,tmp); 415 416 417 idx = h1; 418 #if defined(PETSC_USE_DEBUG) 419 insert_ct++; 420 total_ct++; 421 if (HT[idx] != key) { 422 for (idx=h1; (idx<size) && (HT[idx]!=key); idx++,total_ct++); 423 if (idx == size) { 424 for (idx=0; (idx<h1) && (HT[idx]!=key); idx++,total_ct++); 425 if (idx == h1) { 426 SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"(%D,%D) has no entry in the hash table", row, col); 427 } 428 } 429 } 430 #else 431 if (HT[idx] != key) { 432 for (idx=h1; (idx<size) && (HT[idx]!=key); idx++); 433 if (idx == size) { 434 for (idx=0; (idx<h1) && (HT[idx]!=key); idx++); 435 if (idx == h1) { 436 SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"(%D,%D) has no entry in the hash table", row, col); 437 } 438 } 439 } 440 #endif 441 /* A HASH table entry is found, so insert the values at the correct address */ 442 if (addv == ADD_VALUES) *(HD[idx]+ (col % bs)*bs + (row % bs)) += value; 443 else *(HD[idx]+ (col % bs)*bs + (row % bs)) = value; 444 } 445 } else { 446 if (!baij->donotstash) { 447 if (roworiented) { 448 ierr = MatStashValuesRow_Private(&mat->stash,im[i],n,in,v+i*n,PETSC_FALSE);CHKERRQ(ierr); 449 } else { 450 ierr = MatStashValuesCol_Private(&mat->stash,im[i],n,in,v+i,m,PETSC_FALSE);CHKERRQ(ierr); 451 } 452 } 453 } 454 } 455 #if defined(PETSC_USE_DEBUG) 456 baij->ht_total_ct = total_ct; 457 baij->ht_insert_ct = insert_ct; 458 #endif 459 PetscFunctionReturn(0); 460 } 461 462 #undef __FUNCT__ 463 #define __FUNCT__ "MatSetValuesBlocked_MPIBAIJ_HT" 464 PetscErrorCode MatSetValuesBlocked_MPIBAIJ_HT(Mat mat,PetscInt m,const PetscInt im[],PetscInt n,const PetscInt in[],const PetscScalar v[],InsertMode addv) 465 { 466 Mat_MPIBAIJ *baij = (Mat_MPIBAIJ*)mat->data; 467 PetscBool roworiented = baij->roworiented; 468 PetscErrorCode ierr; 469 PetscInt i,j,ii,jj,row,col; 470 PetscInt rstart=baij->rstartbs; 471 PetscInt rend=mat->rmap->rend,stepval,bs=mat->rmap->bs,bs2=baij->bs2,nbs2=n*bs2; 472 PetscInt h1,key,size=baij->ht_size,idx,*HT=baij->ht,Nbs=baij->Nbs; 473 PetscReal tmp; 474 MatScalar **HD = baij->hd,*baij_a; 475 const PetscScalar *v_t,*value; 476 #if defined(PETSC_USE_DEBUG) 477 PetscInt total_ct=baij->ht_total_ct,insert_ct=baij->ht_insert_ct; 478 #endif 479 480 PetscFunctionBegin; 481 482 if (roworiented) { 483 stepval = (n-1)*bs; 484 } else { 485 stepval = (m-1)*bs; 486 } 487 for (i=0; i<m; i++) { 488 #if defined(PETSC_USE_DEBUG) 489 if (im[i] < 0) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Negative row: %D",im[i]); 490 if (im[i] >= baij->Mbs) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Row too large: row %D max %D",im[i],baij->Mbs-1); 491 #endif 492 row = im[i]; 493 v_t = v + i*nbs2; 494 if (row >= rstart && row < rend) { 495 for (j=0; j<n; j++) { 496 col = in[j]; 497 498 /* Look up into the Hash Table */ 499 key = row*Nbs+col+1; 500 h1 = HASH(size,key,tmp); 501 502 idx = h1; 503 #if defined(PETSC_USE_DEBUG) 504 total_ct++; 505 insert_ct++; 506 if (HT[idx] != key) { 507 for (idx=h1; (idx<size) && (HT[idx]!=key); idx++,total_ct++); 508 if (idx == size) { 509 for (idx=0; (idx<h1) && (HT[idx]!=key); idx++,total_ct++); 510 if (idx == h1) { 511 SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"(%D,%D) has no entry in the hash table", row, col); 512 } 513 } 514 } 515 #else 516 if (HT[idx] != key) { 517 for (idx=h1; (idx<size) && (HT[idx]!=key); idx++); 518 if (idx == size) { 519 for (idx=0; (idx<h1) && (HT[idx]!=key); idx++); 520 if (idx == h1) { 521 SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"(%D,%D) has no entry in the hash table", row, col); 522 } 523 } 524 } 525 #endif 526 baij_a = HD[idx]; 527 if (roworiented) { 528 /*value = v + i*(stepval+bs)*bs + j*bs;*/ 529 /* value = v + (i*(stepval+bs)+j)*bs; */ 530 value = v_t; 531 v_t += bs; 532 if (addv == ADD_VALUES) { 533 for (ii=0; ii<bs; ii++,value+=stepval) { 534 for (jj=ii; jj<bs2; jj+=bs) { 535 baij_a[jj] += *value++; 536 } 537 } 538 } else { 539 for (ii=0; ii<bs; ii++,value+=stepval) { 540 for (jj=ii; jj<bs2; jj+=bs) { 541 baij_a[jj] = *value++; 542 } 543 } 544 } 545 } else { 546 value = v + j*(stepval+bs)*bs + i*bs; 547 if (addv == ADD_VALUES) { 548 for (ii=0; ii<bs; ii++,value+=stepval,baij_a+=bs) { 549 for (jj=0; jj<bs; jj++) { 550 baij_a[jj] += *value++; 551 } 552 } 553 } else { 554 for (ii=0; ii<bs; ii++,value+=stepval,baij_a+=bs) { 555 for (jj=0; jj<bs; jj++) { 556 baij_a[jj] = *value++; 557 } 558 } 559 } 560 } 561 } 562 } else { 563 if (!baij->donotstash) { 564 if (roworiented) { 565 ierr = MatStashValuesRowBlocked_Private(&mat->bstash,im[i],n,in,v,m,n,i);CHKERRQ(ierr); 566 } else { 567 ierr = MatStashValuesColBlocked_Private(&mat->bstash,im[i],n,in,v,m,n,i);CHKERRQ(ierr); 568 } 569 } 570 } 571 } 572 #if defined(PETSC_USE_DEBUG) 573 baij->ht_total_ct = total_ct; 574 baij->ht_insert_ct = insert_ct; 575 #endif 576 PetscFunctionReturn(0); 577 } 578 579 #undef __FUNCT__ 580 #define __FUNCT__ "MatGetValues_MPIBAIJ" 581 PetscErrorCode MatGetValues_MPIBAIJ(Mat mat,PetscInt m,const PetscInt idxm[],PetscInt n,const PetscInt idxn[],PetscScalar v[]) 582 { 583 Mat_MPIBAIJ *baij = (Mat_MPIBAIJ*)mat->data; 584 PetscErrorCode ierr; 585 PetscInt bs=mat->rmap->bs,i,j,bsrstart = mat->rmap->rstart,bsrend = mat->rmap->rend; 586 PetscInt bscstart = mat->cmap->rstart,bscend = mat->cmap->rend,row,col,data; 587 588 PetscFunctionBegin; 589 for (i=0; i<m; i++) { 590 if (idxm[i] < 0) continue; /* SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Negative row: %D",idxm[i]);*/ 591 if (idxm[i] >= mat->rmap->N) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Row too large: row %D max %D",idxm[i],mat->rmap->N-1); 592 if (idxm[i] >= bsrstart && idxm[i] < bsrend) { 593 row = idxm[i] - bsrstart; 594 for (j=0; j<n; j++) { 595 if (idxn[j] < 0) continue; /* SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Negative column: %D",idxn[j]); */ 596 if (idxn[j] >= mat->cmap->N) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Column too large: col %D max %D",idxn[j],mat->cmap->N-1); 597 if (idxn[j] >= bscstart && idxn[j] < bscend){ 598 col = idxn[j] - bscstart; 599 ierr = MatGetValues_SeqBAIJ(baij->A,1,&row,1,&col,v+i*n+j);CHKERRQ(ierr); 600 } else { 601 if (!baij->colmap) { 602 ierr = MatCreateColmap_MPIBAIJ_Private(mat);CHKERRQ(ierr); 603 } 604 #if defined (PETSC_USE_CTABLE) 605 ierr = PetscTableFind(baij->colmap,idxn[j]/bs+1,&data);CHKERRQ(ierr); 606 data --; 607 #else 608 data = baij->colmap[idxn[j]/bs]-1; 609 #endif 610 if ((data < 0) || (baij->garray[data/bs] != idxn[j]/bs)) *(v+i*n+j) = 0.0; 611 else { 612 col = data + idxn[j]%bs; 613 ierr = MatGetValues_SeqBAIJ(baij->B,1,&row,1,&col,v+i*n+j);CHKERRQ(ierr); 614 } 615 } 616 } 617 } else { 618 SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Only local values currently supported"); 619 } 620 } 621 PetscFunctionReturn(0); 622 } 623 624 #undef __FUNCT__ 625 #define __FUNCT__ "MatNorm_MPIBAIJ" 626 PetscErrorCode MatNorm_MPIBAIJ(Mat mat,NormType type,PetscReal *nrm) 627 { 628 Mat_MPIBAIJ *baij = (Mat_MPIBAIJ*)mat->data; 629 Mat_SeqBAIJ *amat = (Mat_SeqBAIJ*)baij->A->data,*bmat = (Mat_SeqBAIJ*)baij->B->data; 630 PetscErrorCode ierr; 631 PetscInt i,j,bs2=baij->bs2,bs=baij->A->rmap->bs,nz,row,col; 632 PetscReal sum = 0.0; 633 MatScalar *v; 634 635 PetscFunctionBegin; 636 if (baij->size == 1) { 637 ierr = MatNorm(baij->A,type,nrm);CHKERRQ(ierr); 638 } else { 639 if (type == NORM_FROBENIUS) { 640 v = amat->a; 641 nz = amat->nz*bs2; 642 for (i=0; i<nz; i++) { 643 #if defined(PETSC_USE_COMPLEX) 644 sum += PetscRealPart(PetscConj(*v)*(*v)); v++; 645 #else 646 sum += (*v)*(*v); v++; 647 #endif 648 } 649 v = bmat->a; 650 nz = bmat->nz*bs2; 651 for (i=0; i<nz; i++) { 652 #if defined(PETSC_USE_COMPLEX) 653 sum += PetscRealPart(PetscConj(*v)*(*v)); v++; 654 #else 655 sum += (*v)*(*v); v++; 656 #endif 657 } 658 ierr = MPI_Allreduce(&sum,nrm,1,MPIU_REAL,MPIU_SUM,((PetscObject)mat)->comm);CHKERRQ(ierr); 659 *nrm = PetscSqrtReal(*nrm); 660 } else if (type == NORM_1) { /* max column sum */ 661 PetscReal *tmp,*tmp2; 662 PetscInt *jj,*garray=baij->garray,cstart=baij->rstartbs; 663 ierr = PetscMalloc2(mat->cmap->N,PetscReal,&tmp,mat->cmap->N,PetscReal,&tmp2);CHKERRQ(ierr); 664 ierr = PetscMemzero(tmp,mat->cmap->N*sizeof(PetscReal));CHKERRQ(ierr); 665 v = amat->a; jj = amat->j; 666 for (i=0; i<amat->nz; i++) { 667 for (j=0; j<bs; j++){ 668 col = bs*(cstart + *jj) + j; /* column index */ 669 for (row=0; row<bs; row++){ 670 tmp[col] += PetscAbsScalar(*v); v++; 671 } 672 } 673 jj++; 674 } 675 v = bmat->a; jj = bmat->j; 676 for (i=0; i<bmat->nz; i++) { 677 for (j=0; j<bs; j++){ 678 col = bs*garray[*jj] + j; 679 for (row=0; row<bs; row++){ 680 tmp[col] += PetscAbsScalar(*v); v++; 681 } 682 } 683 jj++; 684 } 685 ierr = MPI_Allreduce(tmp,tmp2,mat->cmap->N,MPIU_REAL,MPIU_SUM,((PetscObject)mat)->comm);CHKERRQ(ierr); 686 *nrm = 0.0; 687 for (j=0; j<mat->cmap->N; j++) { 688 if (tmp2[j] > *nrm) *nrm = tmp2[j]; 689 } 690 ierr = PetscFree2(tmp,tmp2);CHKERRQ(ierr); 691 } else if (type == NORM_INFINITY) { /* max row sum */ 692 PetscReal *sums; 693 ierr = PetscMalloc(bs*sizeof(PetscReal),&sums);CHKERRQ(ierr); 694 sum = 0.0; 695 for (j=0; j<amat->mbs; j++) { 696 for (row=0; row<bs; row++) sums[row] = 0.0; 697 v = amat->a + bs2*amat->i[j]; 698 nz = amat->i[j+1]-amat->i[j]; 699 for (i=0; i<nz; i++) { 700 for (col=0; col<bs; col++){ 701 for (row=0; row<bs; row++){ 702 sums[row] += PetscAbsScalar(*v); v++; 703 } 704 } 705 } 706 v = bmat->a + bs2*bmat->i[j]; 707 nz = bmat->i[j+1]-bmat->i[j]; 708 for (i=0; i<nz; i++) { 709 for (col=0; col<bs; col++){ 710 for (row=0; row<bs; row++){ 711 sums[row] += PetscAbsScalar(*v); v++; 712 } 713 } 714 } 715 for (row=0; row<bs; row++){ 716 if (sums[row] > sum) sum = sums[row]; 717 } 718 } 719 ierr = MPI_Allreduce(&sum,nrm,1,MPIU_REAL,MPIU_MAX,((PetscObject)mat)->comm);CHKERRQ(ierr); 720 ierr = PetscFree(sums);CHKERRQ(ierr); 721 } else SETERRQ(((PetscObject)mat)->comm,PETSC_ERR_SUP,"No support for this norm yet"); 722 } 723 PetscFunctionReturn(0); 724 } 725 726 /* 727 Creates the hash table, and sets the table 728 This table is created only once. 729 If new entried need to be added to the matrix 730 then the hash table has to be destroyed and 731 recreated. 732 */ 733 #undef __FUNCT__ 734 #define __FUNCT__ "MatCreateHashTable_MPIBAIJ_Private" 735 PetscErrorCode MatCreateHashTable_MPIBAIJ_Private(Mat mat,PetscReal factor) 736 { 737 Mat_MPIBAIJ *baij = (Mat_MPIBAIJ*)mat->data; 738 Mat A = baij->A,B=baij->B; 739 Mat_SeqBAIJ *a=(Mat_SeqBAIJ *)A->data,*b=(Mat_SeqBAIJ *)B->data; 740 PetscInt i,j,k,nz=a->nz+b->nz,h1,*ai=a->i,*aj=a->j,*bi=b->i,*bj=b->j; 741 PetscErrorCode ierr; 742 PetscInt ht_size,bs2=baij->bs2,rstart=baij->rstartbs; 743 PetscInt cstart=baij->cstartbs,*garray=baij->garray,row,col,Nbs=baij->Nbs; 744 PetscInt *HT,key; 745 MatScalar **HD; 746 PetscReal tmp; 747 #if defined(PETSC_USE_INFO) 748 PetscInt ct=0,max=0; 749 #endif 750 751 PetscFunctionBegin; 752 if (baij->ht) PetscFunctionReturn(0); 753 754 baij->ht_size = (PetscInt)(factor*nz); 755 ht_size = baij->ht_size; 756 757 /* Allocate Memory for Hash Table */ 758 ierr = PetscMalloc2(ht_size,MatScalar*,&baij->hd,ht_size,PetscInt,&baij->ht);CHKERRQ(ierr); 759 ierr = PetscMemzero(baij->hd,ht_size*sizeof(MatScalar*));CHKERRQ(ierr); 760 ierr = PetscMemzero(baij->ht,ht_size*sizeof(PetscInt));CHKERRQ(ierr); 761 HD = baij->hd; 762 HT = baij->ht; 763 764 /* Loop Over A */ 765 for (i=0; i<a->mbs; i++) { 766 for (j=ai[i]; j<ai[i+1]; j++) { 767 row = i+rstart; 768 col = aj[j]+cstart; 769 770 key = row*Nbs + col + 1; 771 h1 = HASH(ht_size,key,tmp); 772 for (k=0; k<ht_size; k++){ 773 if (!HT[(h1+k)%ht_size]) { 774 HT[(h1+k)%ht_size] = key; 775 HD[(h1+k)%ht_size] = a->a + j*bs2; 776 break; 777 #if defined(PETSC_USE_INFO) 778 } else { 779 ct++; 780 #endif 781 } 782 } 783 #if defined(PETSC_USE_INFO) 784 if (k> max) max = k; 785 #endif 786 } 787 } 788 /* Loop Over B */ 789 for (i=0; i<b->mbs; i++) { 790 for (j=bi[i]; j<bi[i+1]; j++) { 791 row = i+rstart; 792 col = garray[bj[j]]; 793 key = row*Nbs + col + 1; 794 h1 = HASH(ht_size,key,tmp); 795 for (k=0; k<ht_size; k++){ 796 if (!HT[(h1+k)%ht_size]) { 797 HT[(h1+k)%ht_size] = key; 798 HD[(h1+k)%ht_size] = b->a + j*bs2; 799 break; 800 #if defined(PETSC_USE_INFO) 801 } else { 802 ct++; 803 #endif 804 } 805 } 806 #if defined(PETSC_USE_INFO) 807 if (k> max) max = k; 808 #endif 809 } 810 } 811 812 /* Print Summary */ 813 #if defined(PETSC_USE_INFO) 814 for (i=0,j=0; i<ht_size; i++) { 815 if (HT[i]) {j++;} 816 } 817 ierr = PetscInfo2(mat,"Average Search = %5.2f,max search = %D\n",(!j)? 0.0:((PetscReal)(ct+j))/j,max);CHKERRQ(ierr); 818 #endif 819 PetscFunctionReturn(0); 820 } 821 822 #undef __FUNCT__ 823 #define __FUNCT__ "MatAssemblyBegin_MPIBAIJ" 824 PetscErrorCode MatAssemblyBegin_MPIBAIJ(Mat mat,MatAssemblyType mode) 825 { 826 Mat_MPIBAIJ *baij = (Mat_MPIBAIJ*)mat->data; 827 PetscErrorCode ierr; 828 PetscInt nstash,reallocs; 829 InsertMode addv; 830 831 PetscFunctionBegin; 832 if (baij->donotstash || mat->nooffprocentries) { 833 PetscFunctionReturn(0); 834 } 835 836 /* make sure all processors are either in INSERTMODE or ADDMODE */ 837 ierr = MPI_Allreduce(&mat->insertmode,&addv,1,MPI_INT,MPI_BOR,((PetscObject)mat)->comm);CHKERRQ(ierr); 838 if (addv == (ADD_VALUES|INSERT_VALUES)) SETERRQ(((PetscObject)mat)->comm,PETSC_ERR_ARG_WRONGSTATE,"Some processors inserted others added"); 839 mat->insertmode = addv; /* in case this processor had no cache */ 840 841 ierr = MatStashScatterBegin_Private(mat,&mat->stash,mat->rmap->range);CHKERRQ(ierr); 842 ierr = MatStashScatterBegin_Private(mat,&mat->bstash,baij->rangebs);CHKERRQ(ierr); 843 ierr = MatStashGetInfo_Private(&mat->stash,&nstash,&reallocs);CHKERRQ(ierr); 844 ierr = PetscInfo2(mat,"Stash has %D entries,uses %D mallocs.\n",nstash,reallocs);CHKERRQ(ierr); 845 ierr = MatStashGetInfo_Private(&mat->bstash,&nstash,&reallocs);CHKERRQ(ierr); 846 ierr = PetscInfo2(mat,"Block-Stash has %D entries, uses %D mallocs.\n",nstash,reallocs);CHKERRQ(ierr); 847 PetscFunctionReturn(0); 848 } 849 850 #undef __FUNCT__ 851 #define __FUNCT__ "MatAssemblyEnd_MPIBAIJ" 852 PetscErrorCode MatAssemblyEnd_MPIBAIJ(Mat mat,MatAssemblyType mode) 853 { 854 Mat_MPIBAIJ *baij=(Mat_MPIBAIJ*)mat->data; 855 Mat_SeqBAIJ *a=(Mat_SeqBAIJ*)baij->A->data; 856 PetscErrorCode ierr; 857 PetscInt i,j,rstart,ncols,flg,bs2=baij->bs2; 858 PetscInt *row,*col; 859 PetscBool r1,r2,r3,other_disassembled; 860 MatScalar *val; 861 InsertMode addv = mat->insertmode; 862 PetscMPIInt n; 863 864 /* do not use 'b=(Mat_SeqBAIJ*)baij->B->data' as B can be reset in disassembly */ 865 PetscFunctionBegin; 866 if (!baij->donotstash && !mat->nooffprocentries) { 867 while (1) { 868 ierr = MatStashScatterGetMesg_Private(&mat->stash,&n,&row,&col,&val,&flg);CHKERRQ(ierr); 869 if (!flg) break; 870 871 for (i=0; i<n;) { 872 /* Now identify the consecutive vals belonging to the same row */ 873 for (j=i,rstart=row[j]; j<n; j++) { if (row[j] != rstart) break; } 874 if (j < n) ncols = j-i; 875 else ncols = n-i; 876 /* Now assemble all these values with a single function call */ 877 ierr = MatSetValues_MPIBAIJ(mat,1,row+i,ncols,col+i,val+i,addv);CHKERRQ(ierr); 878 i = j; 879 } 880 } 881 ierr = MatStashScatterEnd_Private(&mat->stash);CHKERRQ(ierr); 882 /* Now process the block-stash. Since the values are stashed column-oriented, 883 set the roworiented flag to column oriented, and after MatSetValues() 884 restore the original flags */ 885 r1 = baij->roworiented; 886 r2 = a->roworiented; 887 r3 = ((Mat_SeqBAIJ*)baij->B->data)->roworiented; 888 baij->roworiented = PETSC_FALSE; 889 a->roworiented = PETSC_FALSE; 890 (((Mat_SeqBAIJ*)baij->B->data))->roworiented = PETSC_FALSE; /* b->roworiented */ 891 while (1) { 892 ierr = MatStashScatterGetMesg_Private(&mat->bstash,&n,&row,&col,&val,&flg);CHKERRQ(ierr); 893 if (!flg) break; 894 895 for (i=0; i<n;) { 896 /* Now identify the consecutive vals belonging to the same row */ 897 for (j=i,rstart=row[j]; j<n; j++) { if (row[j] != rstart) break; } 898 if (j < n) ncols = j-i; 899 else ncols = n-i; 900 ierr = MatSetValuesBlocked_MPIBAIJ(mat,1,row+i,ncols,col+i,val+i*bs2,addv);CHKERRQ(ierr); 901 i = j; 902 } 903 } 904 ierr = MatStashScatterEnd_Private(&mat->bstash);CHKERRQ(ierr); 905 baij->roworiented = r1; 906 a->roworiented = r2; 907 ((Mat_SeqBAIJ*)baij->B->data)->roworiented = r3; /* b->roworiented */ 908 } 909 910 ierr = MatAssemblyBegin(baij->A,mode);CHKERRQ(ierr); 911 ierr = MatAssemblyEnd(baij->A,mode);CHKERRQ(ierr); 912 913 /* determine if any processor has disassembled, if so we must 914 also disassemble ourselfs, in order that we may reassemble. */ 915 /* 916 if nonzero structure of submatrix B cannot change then we know that 917 no processor disassembled thus we can skip this stuff 918 */ 919 if (!((Mat_SeqBAIJ*)baij->B->data)->nonew) { 920 ierr = MPI_Allreduce(&mat->was_assembled,&other_disassembled,1,MPI_INT,MPI_PROD,((PetscObject)mat)->comm);CHKERRQ(ierr); 921 if (mat->was_assembled && !other_disassembled) { 922 ierr = MatDisAssemble_MPIBAIJ(mat);CHKERRQ(ierr); 923 } 924 } 925 926 if (!mat->was_assembled && mode == MAT_FINAL_ASSEMBLY) { 927 ierr = MatSetUpMultiply_MPIBAIJ(mat);CHKERRQ(ierr); 928 } 929 ierr = MatSetOption(baij->B,MAT_CHECK_COMPRESSED_ROW,PETSC_FALSE);CHKERRQ(ierr); 930 ierr = MatAssemblyBegin(baij->B,mode);CHKERRQ(ierr); 931 ierr = MatAssemblyEnd(baij->B,mode);CHKERRQ(ierr); 932 933 #if defined(PETSC_USE_INFO) 934 if (baij->ht && mode== MAT_FINAL_ASSEMBLY) { 935 ierr = PetscInfo1(mat,"Average Hash Table Search in MatSetValues = %5.2f\n",((PetscReal)baij->ht_total_ct)/baij->ht_insert_ct);CHKERRQ(ierr); 936 baij->ht_total_ct = 0; 937 baij->ht_insert_ct = 0; 938 } 939 #endif 940 if (baij->ht_flag && !baij->ht && mode == MAT_FINAL_ASSEMBLY) { 941 ierr = MatCreateHashTable_MPIBAIJ_Private(mat,baij->ht_fact);CHKERRQ(ierr); 942 mat->ops->setvalues = MatSetValues_MPIBAIJ_HT; 943 mat->ops->setvaluesblocked = MatSetValuesBlocked_MPIBAIJ_HT; 944 } 945 946 ierr = PetscFree2(baij->rowvalues,baij->rowindices);CHKERRQ(ierr); 947 baij->rowvalues = 0; 948 PetscFunctionReturn(0); 949 } 950 951 #undef __FUNCT__ 952 #define __FUNCT__ "MatView_MPIBAIJ_ASCIIorDraworSocket" 953 static PetscErrorCode MatView_MPIBAIJ_ASCIIorDraworSocket(Mat mat,PetscViewer viewer) 954 { 955 Mat_MPIBAIJ *baij = (Mat_MPIBAIJ*)mat->data; 956 PetscErrorCode ierr; 957 PetscMPIInt size = baij->size,rank = baij->rank; 958 PetscInt bs = mat->rmap->bs; 959 PetscBool iascii,isdraw; 960 PetscViewer sviewer; 961 PetscViewerFormat format; 962 963 PetscFunctionBegin; 964 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERASCII,&iascii);CHKERRQ(ierr); 965 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERDRAW,&isdraw);CHKERRQ(ierr); 966 if (iascii) { 967 ierr = PetscViewerGetFormat(viewer,&format);CHKERRQ(ierr); 968 if (format == PETSC_VIEWER_ASCII_INFO_DETAIL) { 969 MatInfo info; 970 ierr = MPI_Comm_rank(((PetscObject)mat)->comm,&rank);CHKERRQ(ierr); 971 ierr = MatGetInfo(mat,MAT_LOCAL,&info);CHKERRQ(ierr); 972 ierr = PetscViewerASCIISynchronizedAllow(viewer,PETSC_TRUE);CHKERRQ(ierr); 973 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"[%d] Local rows %D nz %D nz alloced %D bs %D mem %D\n", 974 rank,mat->rmap->n,(PetscInt)info.nz_used,(PetscInt)info.nz_allocated,mat->rmap->bs,(PetscInt)info.memory);CHKERRQ(ierr); 975 ierr = MatGetInfo(baij->A,MAT_LOCAL,&info);CHKERRQ(ierr); 976 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"[%d] on-diagonal part: nz %D \n",rank,(PetscInt)info.nz_used);CHKERRQ(ierr); 977 ierr = MatGetInfo(baij->B,MAT_LOCAL,&info);CHKERRQ(ierr); 978 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"[%d] off-diagonal part: nz %D \n",rank,(PetscInt)info.nz_used);CHKERRQ(ierr); 979 ierr = PetscViewerFlush(viewer);CHKERRQ(ierr); 980 ierr = PetscViewerASCIISynchronizedAllow(viewer,PETSC_FALSE);CHKERRQ(ierr); 981 ierr = PetscViewerASCIIPrintf(viewer,"Information on VecScatter used in matrix-vector product: \n");CHKERRQ(ierr); 982 ierr = VecScatterView(baij->Mvctx,viewer);CHKERRQ(ierr); 983 PetscFunctionReturn(0); 984 } else if (format == PETSC_VIEWER_ASCII_INFO) { 985 ierr = PetscViewerASCIIPrintf(viewer," block size is %D\n",bs);CHKERRQ(ierr); 986 PetscFunctionReturn(0); 987 } else if (format == PETSC_VIEWER_ASCII_FACTOR_INFO) { 988 PetscFunctionReturn(0); 989 } 990 } 991 992 if (isdraw) { 993 PetscDraw draw; 994 PetscBool isnull; 995 ierr = PetscViewerDrawGetDraw(viewer,0,&draw);CHKERRQ(ierr); 996 ierr = PetscDrawIsNull(draw,&isnull);CHKERRQ(ierr); if (isnull) PetscFunctionReturn(0); 997 } 998 999 if (size == 1) { 1000 ierr = PetscObjectSetName((PetscObject)baij->A,((PetscObject)mat)->name);CHKERRQ(ierr); 1001 ierr = MatView(baij->A,viewer);CHKERRQ(ierr); 1002 } else { 1003 /* assemble the entire matrix onto first processor. */ 1004 Mat A; 1005 Mat_SeqBAIJ *Aloc; 1006 PetscInt M = mat->rmap->N,N = mat->cmap->N,*ai,*aj,col,i,j,k,*rvals,mbs = baij->mbs; 1007 MatScalar *a; 1008 1009 /* Here we are creating a temporary matrix, so will assume MPIBAIJ is acceptable */ 1010 /* Perhaps this should be the type of mat? */ 1011 ierr = MatCreate(((PetscObject)mat)->comm,&A);CHKERRQ(ierr); 1012 if (!rank) { 1013 ierr = MatSetSizes(A,M,N,M,N);CHKERRQ(ierr); 1014 } else { 1015 ierr = MatSetSizes(A,0,0,M,N);CHKERRQ(ierr); 1016 } 1017 ierr = MatSetType(A,MATMPIBAIJ);CHKERRQ(ierr); 1018 ierr = MatMPIBAIJSetPreallocation(A,mat->rmap->bs,0,PETSC_NULL,0,PETSC_NULL);CHKERRQ(ierr); 1019 ierr = MatSetOption(A,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 1020 ierr = PetscLogObjectParent(mat,A);CHKERRQ(ierr); 1021 1022 /* copy over the A part */ 1023 Aloc = (Mat_SeqBAIJ*)baij->A->data; 1024 ai = Aloc->i; aj = Aloc->j; a = Aloc->a; 1025 ierr = PetscMalloc(bs*sizeof(PetscInt),&rvals);CHKERRQ(ierr); 1026 1027 for (i=0; i<mbs; i++) { 1028 rvals[0] = bs*(baij->rstartbs + i); 1029 for (j=1; j<bs; j++) { rvals[j] = rvals[j-1] + 1; } 1030 for (j=ai[i]; j<ai[i+1]; j++) { 1031 col = (baij->cstartbs+aj[j])*bs; 1032 for (k=0; k<bs; k++) { 1033 ierr = MatSetValues_MPIBAIJ(A,bs,rvals,1,&col,a,INSERT_VALUES);CHKERRQ(ierr); 1034 col++; a += bs; 1035 } 1036 } 1037 } 1038 /* copy over the B part */ 1039 Aloc = (Mat_SeqBAIJ*)baij->B->data; 1040 ai = Aloc->i; aj = Aloc->j; a = Aloc->a; 1041 for (i=0; i<mbs; i++) { 1042 rvals[0] = bs*(baij->rstartbs + i); 1043 for (j=1; j<bs; j++) { rvals[j] = rvals[j-1] + 1; } 1044 for (j=ai[i]; j<ai[i+1]; j++) { 1045 col = baij->garray[aj[j]]*bs; 1046 for (k=0; k<bs; k++) { 1047 ierr = MatSetValues_MPIBAIJ(A,bs,rvals,1,&col,a,INSERT_VALUES);CHKERRQ(ierr); 1048 col++; a += bs; 1049 } 1050 } 1051 } 1052 ierr = PetscFree(rvals);CHKERRQ(ierr); 1053 ierr = MatAssemblyBegin(A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1054 ierr = MatAssemblyEnd(A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1055 /* 1056 Everyone has to call to draw the matrix since the graphics waits are 1057 synchronized across all processors that share the PetscDraw object 1058 */ 1059 ierr = PetscViewerGetSingleton(viewer,&sviewer);CHKERRQ(ierr); 1060 if (!rank) { 1061 ierr = PetscObjectSetName((PetscObject)((Mat_MPIBAIJ*)(A->data))->A,((PetscObject)mat)->name);CHKERRQ(ierr); 1062 /* Set the type name to MATMPIBAIJ so that the correct type can be printed out by PetscObjectPrintClassNamePrefixType() in MatView_SeqBAIJ_ASCII()*/ 1063 PetscStrcpy(((PetscObject)((Mat_MPIBAIJ*)(A->data))->A)->type_name,MATMPIBAIJ); 1064 ierr = MatView(((Mat_MPIBAIJ*)(A->data))->A,sviewer);CHKERRQ(ierr); 1065 } 1066 ierr = PetscViewerRestoreSingleton(viewer,&sviewer);CHKERRQ(ierr); 1067 ierr = MatDestroy(&A);CHKERRQ(ierr); 1068 } 1069 PetscFunctionReturn(0); 1070 } 1071 1072 #undef __FUNCT__ 1073 #define __FUNCT__ "MatView_MPIBAIJ_Binary" 1074 static PetscErrorCode MatView_MPIBAIJ_Binary(Mat mat,PetscViewer viewer) 1075 { 1076 Mat_MPIBAIJ *a = (Mat_MPIBAIJ*)mat->data; 1077 Mat_SeqBAIJ* A = (Mat_SeqBAIJ*)a->A->data; 1078 Mat_SeqBAIJ* B = (Mat_SeqBAIJ*)a->B->data; 1079 PetscErrorCode ierr; 1080 PetscInt i,*row_lens,*crow_lens,bs = mat->rmap->bs,j,k,bs2=a->bs2,header[4],nz,rlen; 1081 PetscInt *range=0,nzmax,*column_indices,cnt,col,*garray = a->garray,cstart = mat->cmap->rstart/bs,len,pcnt,l,ll; 1082 int fd; 1083 PetscScalar *column_values; 1084 FILE *file; 1085 PetscMPIInt rank,size,tag = ((PetscObject)viewer)->tag; 1086 PetscInt message_count,flowcontrolcount; 1087 1088 PetscFunctionBegin; 1089 ierr = MPI_Comm_rank(((PetscObject)mat)->comm,&rank);CHKERRQ(ierr); 1090 ierr = MPI_Comm_size(((PetscObject)mat)->comm,&size);CHKERRQ(ierr); 1091 nz = bs2*(A->nz + B->nz); 1092 rlen = mat->rmap->n; 1093 if (!rank) { 1094 header[0] = MAT_FILE_CLASSID; 1095 header[1] = mat->rmap->N; 1096 header[2] = mat->cmap->N; 1097 ierr = MPI_Reduce(&nz,&header[3],1,MPIU_INT,MPI_SUM,0,((PetscObject)mat)->comm);CHKERRQ(ierr); 1098 ierr = PetscViewerBinaryGetDescriptor(viewer,&fd);CHKERRQ(ierr); 1099 ierr = PetscBinaryWrite(fd,header,4,PETSC_INT,PETSC_TRUE);CHKERRQ(ierr); 1100 /* get largest number of rows any processor has */ 1101 range = mat->rmap->range; 1102 for (i=1; i<size; i++) { 1103 rlen = PetscMax(rlen,range[i+1] - range[i]); 1104 } 1105 } else { 1106 ierr = MPI_Reduce(&nz,0,1,MPIU_INT,MPI_SUM,0,((PetscObject)mat)->comm);CHKERRQ(ierr); 1107 } 1108 1109 ierr = PetscMalloc((rlen/bs)*sizeof(PetscInt),&crow_lens);CHKERRQ(ierr); 1110 /* compute lengths of each row */ 1111 for (i=0; i<a->mbs; i++) { 1112 crow_lens[i] = A->i[i+1] - A->i[i] + B->i[i+1] - B->i[i]; 1113 } 1114 /* store the row lengths to the file */ 1115 ierr = PetscViewerFlowControlStart(viewer,&message_count,&flowcontrolcount);CHKERRQ(ierr); 1116 if (!rank) { 1117 MPI_Status status; 1118 ierr = PetscMalloc(rlen*sizeof(PetscInt),&row_lens);CHKERRQ(ierr); 1119 rlen = (range[1] - range[0])/bs; 1120 for (i=0; i<rlen; i++) { 1121 for (j=0; j<bs; j++) { 1122 row_lens[i*bs+j] = bs*crow_lens[i]; 1123 } 1124 } 1125 ierr = PetscBinaryWrite(fd,row_lens,bs*rlen,PETSC_INT,PETSC_TRUE);CHKERRQ(ierr); 1126 for (i=1; i<size; i++) { 1127 rlen = (range[i+1] - range[i])/bs; 1128 ierr = PetscViewerFlowControlStepMaster(viewer,i,message_count,flowcontrolcount);CHKERRQ(ierr); 1129 ierr = MPI_Recv(crow_lens,rlen,MPIU_INT,i,tag,((PetscObject)mat)->comm,&status);CHKERRQ(ierr); 1130 for (k=0; k<rlen; k++) { 1131 for (j=0; j<bs; j++) { 1132 row_lens[k*bs+j] = bs*crow_lens[k]; 1133 } 1134 } 1135 ierr = PetscBinaryWrite(fd,row_lens,bs*rlen,PETSC_INT,PETSC_TRUE);CHKERRQ(ierr); 1136 } 1137 ierr = PetscViewerFlowControlEndMaster(viewer,message_count);CHKERRQ(ierr); 1138 ierr = PetscFree(row_lens);CHKERRQ(ierr); 1139 } else { 1140 ierr = PetscViewerFlowControlStepWorker(viewer,rank,message_count);CHKERRQ(ierr); 1141 ierr = MPI_Send(crow_lens,mat->rmap->n/bs,MPIU_INT,0,tag,((PetscObject)mat)->comm);CHKERRQ(ierr); 1142 ierr = PetscViewerFlowControlEndWorker(viewer,message_count);CHKERRQ(ierr); 1143 } 1144 ierr = PetscFree(crow_lens);CHKERRQ(ierr); 1145 1146 /* load up the local column indices. Include for all rows not just one for each block row since process 0 does not have the 1147 information needed to make it for each row from a block row. This does require more communication but still not more than 1148 the communication needed for the nonzero values */ 1149 nzmax = nz; /* space a largest processor needs */ 1150 ierr = MPI_Reduce(&nz,&nzmax,1,MPIU_INT,MPI_MAX,0,((PetscObject)mat)->comm);CHKERRQ(ierr); 1151 ierr = PetscMalloc(nzmax*sizeof(PetscInt),&column_indices);CHKERRQ(ierr); 1152 cnt = 0; 1153 for (i=0; i<a->mbs; i++) { 1154 pcnt = cnt; 1155 for (j=B->i[i]; j<B->i[i+1]; j++) { 1156 if ( (col = garray[B->j[j]]) > cstart) break; 1157 for (l=0; l<bs; l++) { 1158 column_indices[cnt++] = bs*col+l; 1159 } 1160 } 1161 for (k=A->i[i]; k<A->i[i+1]; k++) { 1162 for (l=0; l<bs; l++) { 1163 column_indices[cnt++] = bs*(A->j[k] + cstart)+l; 1164 } 1165 } 1166 for (; j<B->i[i+1]; j++) { 1167 for (l=0; l<bs; l++) { 1168 column_indices[cnt++] = bs*garray[B->j[j]]+l; 1169 } 1170 } 1171 len = cnt - pcnt; 1172 for (k=1; k<bs; k++) { 1173 ierr = PetscMemcpy(&column_indices[cnt],&column_indices[pcnt],len*sizeof(PetscInt));CHKERRQ(ierr); 1174 cnt += len; 1175 } 1176 } 1177 if (cnt != nz) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_LIB,"Internal PETSc error: cnt = %D nz = %D",cnt,nz); 1178 1179 /* store the columns to the file */ 1180 ierr = PetscViewerFlowControlStart(viewer,&message_count,&flowcontrolcount);CHKERRQ(ierr); 1181 if (!rank) { 1182 MPI_Status status; 1183 ierr = PetscBinaryWrite(fd,column_indices,nz,PETSC_INT,PETSC_TRUE);CHKERRQ(ierr); 1184 for (i=1; i<size; i++) { 1185 ierr = PetscViewerFlowControlStepMaster(viewer,i,message_count,flowcontrolcount);CHKERRQ(ierr); 1186 ierr = MPI_Recv(&cnt,1,MPIU_INT,i,tag,((PetscObject)mat)->comm,&status);CHKERRQ(ierr); 1187 ierr = MPI_Recv(column_indices,cnt,MPIU_INT,i,tag,((PetscObject)mat)->comm,&status);CHKERRQ(ierr); 1188 ierr = PetscBinaryWrite(fd,column_indices,cnt,PETSC_INT,PETSC_TRUE);CHKERRQ(ierr); 1189 } 1190 ierr = PetscViewerFlowControlEndMaster(viewer,message_count);CHKERRQ(ierr); 1191 } else { 1192 ierr = PetscViewerFlowControlStepWorker(viewer,rank,message_count);CHKERRQ(ierr); 1193 ierr = MPI_Send(&cnt,1,MPIU_INT,0,tag,((PetscObject)mat)->comm);CHKERRQ(ierr); 1194 ierr = MPI_Send(column_indices,cnt,MPIU_INT,0,tag,((PetscObject)mat)->comm);CHKERRQ(ierr); 1195 ierr = PetscViewerFlowControlEndWorker(viewer,message_count);CHKERRQ(ierr); 1196 } 1197 ierr = PetscFree(column_indices);CHKERRQ(ierr); 1198 1199 /* load up the numerical values */ 1200 ierr = PetscMalloc(nzmax*sizeof(PetscScalar),&column_values);CHKERRQ(ierr); 1201 cnt = 0; 1202 for (i=0; i<a->mbs; i++) { 1203 rlen = bs*(B->i[i+1] - B->i[i] + A->i[i+1] - A->i[i]); 1204 for (j=B->i[i]; j<B->i[i+1]; j++) { 1205 if ( garray[B->j[j]] > cstart) break; 1206 for (l=0; l<bs; l++) { 1207 for (ll=0; ll<bs; ll++) { 1208 column_values[cnt + l*rlen + ll] = B->a[bs2*j+l+bs*ll]; 1209 } 1210 } 1211 cnt += bs; 1212 } 1213 for (k=A->i[i]; k<A->i[i+1]; k++) { 1214 for (l=0; l<bs; l++) { 1215 for (ll=0; ll<bs; ll++) { 1216 column_values[cnt + l*rlen + ll] = A->a[bs2*k+l+bs*ll]; 1217 } 1218 } 1219 cnt += bs; 1220 } 1221 for (; j<B->i[i+1]; j++) { 1222 for (l=0; l<bs; l++) { 1223 for (ll=0; ll<bs; ll++) { 1224 column_values[cnt + l*rlen + ll] = B->a[bs2*j+l+bs*ll]; 1225 } 1226 } 1227 cnt += bs; 1228 } 1229 cnt += (bs-1)*rlen; 1230 } 1231 if (cnt != nz) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Internal PETSc error: cnt = %D nz = %D",cnt,nz); 1232 1233 /* store the column values to the file */ 1234 ierr = PetscViewerFlowControlStart(viewer,&message_count,&flowcontrolcount);CHKERRQ(ierr); 1235 if (!rank) { 1236 MPI_Status status; 1237 ierr = PetscBinaryWrite(fd,column_values,nz,PETSC_SCALAR,PETSC_TRUE);CHKERRQ(ierr); 1238 for (i=1; i<size; i++) { 1239 ierr = PetscViewerFlowControlStepMaster(viewer,i,message_count,flowcontrolcount);CHKERRQ(ierr); 1240 ierr = MPI_Recv(&cnt,1,MPIU_INT,i,tag,((PetscObject)mat)->comm,&status);CHKERRQ(ierr); 1241 ierr = MPI_Recv(column_values,cnt,MPIU_SCALAR,i,tag,((PetscObject)mat)->comm,&status);CHKERRQ(ierr); 1242 ierr = PetscBinaryWrite(fd,column_values,cnt,PETSC_SCALAR,PETSC_TRUE);CHKERRQ(ierr); 1243 } 1244 ierr = PetscViewerFlowControlEndMaster(viewer,message_count);CHKERRQ(ierr); 1245 } else { 1246 ierr = PetscViewerFlowControlStepWorker(viewer,rank,message_count);CHKERRQ(ierr); 1247 ierr = MPI_Send(&nz,1,MPIU_INT,0,tag,((PetscObject)mat)->comm);CHKERRQ(ierr); 1248 ierr = MPI_Send(column_values,nz,MPIU_SCALAR,0,tag,((PetscObject)mat)->comm);CHKERRQ(ierr); 1249 ierr = PetscViewerFlowControlEndWorker(viewer,message_count);CHKERRQ(ierr); 1250 } 1251 ierr = PetscFree(column_values);CHKERRQ(ierr); 1252 1253 ierr = PetscViewerBinaryGetInfoPointer(viewer,&file);CHKERRQ(ierr); 1254 if (file) { 1255 fprintf(file,"-matload_block_size %d\n",(int)mat->rmap->bs); 1256 } 1257 PetscFunctionReturn(0); 1258 } 1259 1260 #undef __FUNCT__ 1261 #define __FUNCT__ "MatView_MPIBAIJ" 1262 PetscErrorCode MatView_MPIBAIJ(Mat mat,PetscViewer viewer) 1263 { 1264 PetscErrorCode ierr; 1265 PetscBool iascii,isdraw,issocket,isbinary; 1266 1267 PetscFunctionBegin; 1268 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERASCII,&iascii);CHKERRQ(ierr); 1269 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERDRAW,&isdraw);CHKERRQ(ierr); 1270 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERSOCKET,&issocket);CHKERRQ(ierr); 1271 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERBINARY,&isbinary);CHKERRQ(ierr); 1272 if (iascii || isdraw || issocket) { 1273 ierr = MatView_MPIBAIJ_ASCIIorDraworSocket(mat,viewer);CHKERRQ(ierr); 1274 } else if (isbinary) { 1275 ierr = MatView_MPIBAIJ_Binary(mat,viewer);CHKERRQ(ierr); 1276 } else { 1277 SETERRQ1(((PetscObject)mat)->comm,PETSC_ERR_SUP,"Viewer type %s not supported by MPIBAIJ matrices",((PetscObject)viewer)->type_name); 1278 } 1279 PetscFunctionReturn(0); 1280 } 1281 1282 #undef __FUNCT__ 1283 #define __FUNCT__ "MatDestroy_MPIBAIJ" 1284 PetscErrorCode MatDestroy_MPIBAIJ(Mat mat) 1285 { 1286 Mat_MPIBAIJ *baij = (Mat_MPIBAIJ*)mat->data; 1287 PetscErrorCode ierr; 1288 1289 PetscFunctionBegin; 1290 #if defined(PETSC_USE_LOG) 1291 PetscLogObjectState((PetscObject)mat,"Rows=%D,Cols=%D",mat->rmap->N,mat->cmap->N); 1292 #endif 1293 ierr = MatStashDestroy_Private(&mat->stash);CHKERRQ(ierr); 1294 ierr = MatStashDestroy_Private(&mat->bstash);CHKERRQ(ierr); 1295 ierr = MatDestroy(&baij->A);CHKERRQ(ierr); 1296 ierr = MatDestroy(&baij->B);CHKERRQ(ierr); 1297 #if defined (PETSC_USE_CTABLE) 1298 ierr = PetscTableDestroy(&baij->colmap);CHKERRQ(ierr); 1299 #else 1300 ierr = PetscFree(baij->colmap);CHKERRQ(ierr); 1301 #endif 1302 ierr = PetscFree(baij->garray);CHKERRQ(ierr); 1303 ierr = VecDestroy(&baij->lvec);CHKERRQ(ierr); 1304 ierr = VecScatterDestroy(&baij->Mvctx);CHKERRQ(ierr); 1305 ierr = PetscFree2(baij->rowvalues,baij->rowindices);CHKERRQ(ierr); 1306 ierr = PetscFree(baij->barray);CHKERRQ(ierr); 1307 ierr = PetscFree2(baij->hd,baij->ht);CHKERRQ(ierr); 1308 ierr = PetscFree(baij->rangebs);CHKERRQ(ierr); 1309 ierr = PetscFree(mat->data);CHKERRQ(ierr); 1310 1311 ierr = PetscObjectChangeTypeName((PetscObject)mat,0);CHKERRQ(ierr); 1312 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatStoreValues_C","",PETSC_NULL);CHKERRQ(ierr); 1313 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatRetrieveValues_C","",PETSC_NULL);CHKERRQ(ierr); 1314 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatGetDiagonalBlock_C","",PETSC_NULL);CHKERRQ(ierr); 1315 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatMPIBAIJSetPreallocation_C","",PETSC_NULL);CHKERRQ(ierr); 1316 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatMPIBAIJSetPreallocationCSR_C","",PETSC_NULL);CHKERRQ(ierr); 1317 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatDiagonalScaleLocal_C","",PETSC_NULL);CHKERRQ(ierr); 1318 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatSetHashTableFactor_C","",PETSC_NULL);CHKERRQ(ierr); 1319 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatConvert_mpibaij_mpisbaij_C","",PETSC_NULL);CHKERRQ(ierr); 1320 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatConvert_mpibaij_mpibstrm_C","",PETSC_NULL);CHKERRQ(ierr); 1321 PetscFunctionReturn(0); 1322 } 1323 1324 #undef __FUNCT__ 1325 #define __FUNCT__ "MatMult_MPIBAIJ" 1326 PetscErrorCode MatMult_MPIBAIJ(Mat A,Vec xx,Vec yy) 1327 { 1328 Mat_MPIBAIJ *a = (Mat_MPIBAIJ*)A->data; 1329 PetscErrorCode ierr; 1330 PetscInt nt; 1331 1332 PetscFunctionBegin; 1333 ierr = VecGetLocalSize(xx,&nt);CHKERRQ(ierr); 1334 if (nt != A->cmap->n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Incompatible partition of A and xx"); 1335 ierr = VecGetLocalSize(yy,&nt);CHKERRQ(ierr); 1336 if (nt != A->rmap->n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Incompatible parition of A and yy"); 1337 ierr = VecScatterBegin(a->Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1338 ierr = (*a->A->ops->mult)(a->A,xx,yy);CHKERRQ(ierr); 1339 ierr = VecScatterEnd(a->Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1340 ierr = (*a->B->ops->multadd)(a->B,a->lvec,yy,yy);CHKERRQ(ierr); 1341 PetscFunctionReturn(0); 1342 } 1343 1344 #undef __FUNCT__ 1345 #define __FUNCT__ "MatMultAdd_MPIBAIJ" 1346 PetscErrorCode MatMultAdd_MPIBAIJ(Mat A,Vec xx,Vec yy,Vec zz) 1347 { 1348 Mat_MPIBAIJ *a = (Mat_MPIBAIJ*)A->data; 1349 PetscErrorCode ierr; 1350 1351 PetscFunctionBegin; 1352 ierr = VecScatterBegin(a->Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1353 ierr = (*a->A->ops->multadd)(a->A,xx,yy,zz);CHKERRQ(ierr); 1354 ierr = VecScatterEnd(a->Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1355 ierr = (*a->B->ops->multadd)(a->B,a->lvec,zz,zz);CHKERRQ(ierr); 1356 PetscFunctionReturn(0); 1357 } 1358 1359 #undef __FUNCT__ 1360 #define __FUNCT__ "MatMultTranspose_MPIBAIJ" 1361 PetscErrorCode MatMultTranspose_MPIBAIJ(Mat A,Vec xx,Vec yy) 1362 { 1363 Mat_MPIBAIJ *a = (Mat_MPIBAIJ*)A->data; 1364 PetscErrorCode ierr; 1365 PetscBool merged; 1366 1367 PetscFunctionBegin; 1368 ierr = VecScatterGetMerged(a->Mvctx,&merged);CHKERRQ(ierr); 1369 /* do nondiagonal part */ 1370 ierr = (*a->B->ops->multtranspose)(a->B,xx,a->lvec);CHKERRQ(ierr); 1371 if (!merged) { 1372 /* send it on its way */ 1373 ierr = VecScatterBegin(a->Mvctx,a->lvec,yy,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1374 /* do local part */ 1375 ierr = (*a->A->ops->multtranspose)(a->A,xx,yy);CHKERRQ(ierr); 1376 /* receive remote parts: note this assumes the values are not actually */ 1377 /* inserted in yy until the next line */ 1378 ierr = VecScatterEnd(a->Mvctx,a->lvec,yy,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1379 } else { 1380 /* do local part */ 1381 ierr = (*a->A->ops->multtranspose)(a->A,xx,yy);CHKERRQ(ierr); 1382 /* send it on its way */ 1383 ierr = VecScatterBegin(a->Mvctx,a->lvec,yy,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1384 /* values actually were received in the Begin() but we need to call this nop */ 1385 ierr = VecScatterEnd(a->Mvctx,a->lvec,yy,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1386 } 1387 PetscFunctionReturn(0); 1388 } 1389 1390 #undef __FUNCT__ 1391 #define __FUNCT__ "MatMultTransposeAdd_MPIBAIJ" 1392 PetscErrorCode MatMultTransposeAdd_MPIBAIJ(Mat A,Vec xx,Vec yy,Vec zz) 1393 { 1394 Mat_MPIBAIJ *a = (Mat_MPIBAIJ*)A->data; 1395 PetscErrorCode ierr; 1396 1397 PetscFunctionBegin; 1398 /* do nondiagonal part */ 1399 ierr = (*a->B->ops->multtranspose)(a->B,xx,a->lvec);CHKERRQ(ierr); 1400 /* send it on its way */ 1401 ierr = VecScatterBegin(a->Mvctx,a->lvec,zz,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1402 /* do local part */ 1403 ierr = (*a->A->ops->multtransposeadd)(a->A,xx,yy,zz);CHKERRQ(ierr); 1404 /* receive remote parts: note this assumes the values are not actually */ 1405 /* inserted in yy until the next line, which is true for my implementation*/ 1406 /* but is not perhaps always true. */ 1407 ierr = VecScatterEnd(a->Mvctx,a->lvec,zz,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1408 PetscFunctionReturn(0); 1409 } 1410 1411 /* 1412 This only works correctly for square matrices where the subblock A->A is the 1413 diagonal block 1414 */ 1415 #undef __FUNCT__ 1416 #define __FUNCT__ "MatGetDiagonal_MPIBAIJ" 1417 PetscErrorCode MatGetDiagonal_MPIBAIJ(Mat A,Vec v) 1418 { 1419 Mat_MPIBAIJ *a = (Mat_MPIBAIJ*)A->data; 1420 PetscErrorCode ierr; 1421 1422 PetscFunctionBegin; 1423 if (A->rmap->N != A->cmap->N) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Supports only square matrix where A->A is diag block"); 1424 ierr = MatGetDiagonal(a->A,v);CHKERRQ(ierr); 1425 PetscFunctionReturn(0); 1426 } 1427 1428 #undef __FUNCT__ 1429 #define __FUNCT__ "MatScale_MPIBAIJ" 1430 PetscErrorCode MatScale_MPIBAIJ(Mat A,PetscScalar aa) 1431 { 1432 Mat_MPIBAIJ *a = (Mat_MPIBAIJ*)A->data; 1433 PetscErrorCode ierr; 1434 1435 PetscFunctionBegin; 1436 ierr = MatScale(a->A,aa);CHKERRQ(ierr); 1437 ierr = MatScale(a->B,aa);CHKERRQ(ierr); 1438 PetscFunctionReturn(0); 1439 } 1440 1441 #undef __FUNCT__ 1442 #define __FUNCT__ "MatGetRow_MPIBAIJ" 1443 PetscErrorCode MatGetRow_MPIBAIJ(Mat matin,PetscInt row,PetscInt *nz,PetscInt **idx,PetscScalar **v) 1444 { 1445 Mat_MPIBAIJ *mat = (Mat_MPIBAIJ*)matin->data; 1446 PetscScalar *vworkA,*vworkB,**pvA,**pvB,*v_p; 1447 PetscErrorCode ierr; 1448 PetscInt bs = matin->rmap->bs,bs2 = mat->bs2,i,*cworkA,*cworkB,**pcA,**pcB; 1449 PetscInt nztot,nzA,nzB,lrow,brstart = matin->rmap->rstart,brend = matin->rmap->rend; 1450 PetscInt *cmap,*idx_p,cstart = mat->cstartbs; 1451 1452 PetscFunctionBegin; 1453 if (row < brstart || row >= brend) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Only local rows"); 1454 if (mat->getrowactive) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Already active"); 1455 mat->getrowactive = PETSC_TRUE; 1456 1457 if (!mat->rowvalues && (idx || v)) { 1458 /* 1459 allocate enough space to hold information from the longest row. 1460 */ 1461 Mat_SeqBAIJ *Aa = (Mat_SeqBAIJ*)mat->A->data,*Ba = (Mat_SeqBAIJ*)mat->B->data; 1462 PetscInt max = 1,mbs = mat->mbs,tmp; 1463 for (i=0; i<mbs; i++) { 1464 tmp = Aa->i[i+1] - Aa->i[i] + Ba->i[i+1] - Ba->i[i]; 1465 if (max < tmp) { max = tmp; } 1466 } 1467 ierr = PetscMalloc2(max*bs2,PetscScalar,&mat->rowvalues,max*bs2,PetscInt,&mat->rowindices);CHKERRQ(ierr); 1468 } 1469 lrow = row - brstart; 1470 1471 pvA = &vworkA; pcA = &cworkA; pvB = &vworkB; pcB = &cworkB; 1472 if (!v) {pvA = 0; pvB = 0;} 1473 if (!idx) {pcA = 0; if (!v) pcB = 0;} 1474 ierr = (*mat->A->ops->getrow)(mat->A,lrow,&nzA,pcA,pvA);CHKERRQ(ierr); 1475 ierr = (*mat->B->ops->getrow)(mat->B,lrow,&nzB,pcB,pvB);CHKERRQ(ierr); 1476 nztot = nzA + nzB; 1477 1478 cmap = mat->garray; 1479 if (v || idx) { 1480 if (nztot) { 1481 /* Sort by increasing column numbers, assuming A and B already sorted */ 1482 PetscInt imark = -1; 1483 if (v) { 1484 *v = v_p = mat->rowvalues; 1485 for (i=0; i<nzB; i++) { 1486 if (cmap[cworkB[i]/bs] < cstart) v_p[i] = vworkB[i]; 1487 else break; 1488 } 1489 imark = i; 1490 for (i=0; i<nzA; i++) v_p[imark+i] = vworkA[i]; 1491 for (i=imark; i<nzB; i++) v_p[nzA+i] = vworkB[i]; 1492 } 1493 if (idx) { 1494 *idx = idx_p = mat->rowindices; 1495 if (imark > -1) { 1496 for (i=0; i<imark; i++) { 1497 idx_p[i] = cmap[cworkB[i]/bs]*bs + cworkB[i]%bs; 1498 } 1499 } else { 1500 for (i=0; i<nzB; i++) { 1501 if (cmap[cworkB[i]/bs] < cstart) 1502 idx_p[i] = cmap[cworkB[i]/bs]*bs + cworkB[i]%bs ; 1503 else break; 1504 } 1505 imark = i; 1506 } 1507 for (i=0; i<nzA; i++) idx_p[imark+i] = cstart*bs + cworkA[i]; 1508 for (i=imark; i<nzB; i++) idx_p[nzA+i] = cmap[cworkB[i]/bs]*bs + cworkB[i]%bs ; 1509 } 1510 } else { 1511 if (idx) *idx = 0; 1512 if (v) *v = 0; 1513 } 1514 } 1515 *nz = nztot; 1516 ierr = (*mat->A->ops->restorerow)(mat->A,lrow,&nzA,pcA,pvA);CHKERRQ(ierr); 1517 ierr = (*mat->B->ops->restorerow)(mat->B,lrow,&nzB,pcB,pvB);CHKERRQ(ierr); 1518 PetscFunctionReturn(0); 1519 } 1520 1521 #undef __FUNCT__ 1522 #define __FUNCT__ "MatRestoreRow_MPIBAIJ" 1523 PetscErrorCode MatRestoreRow_MPIBAIJ(Mat mat,PetscInt row,PetscInt *nz,PetscInt **idx,PetscScalar **v) 1524 { 1525 Mat_MPIBAIJ *baij = (Mat_MPIBAIJ*)mat->data; 1526 1527 PetscFunctionBegin; 1528 if (!baij->getrowactive) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"MatGetRow not called"); 1529 baij->getrowactive = PETSC_FALSE; 1530 PetscFunctionReturn(0); 1531 } 1532 1533 #undef __FUNCT__ 1534 #define __FUNCT__ "MatZeroEntries_MPIBAIJ" 1535 PetscErrorCode MatZeroEntries_MPIBAIJ(Mat A) 1536 { 1537 Mat_MPIBAIJ *l = (Mat_MPIBAIJ*)A->data; 1538 PetscErrorCode ierr; 1539 1540 PetscFunctionBegin; 1541 ierr = MatZeroEntries(l->A);CHKERRQ(ierr); 1542 ierr = MatZeroEntries(l->B);CHKERRQ(ierr); 1543 PetscFunctionReturn(0); 1544 } 1545 1546 #undef __FUNCT__ 1547 #define __FUNCT__ "MatGetInfo_MPIBAIJ" 1548 PetscErrorCode MatGetInfo_MPIBAIJ(Mat matin,MatInfoType flag,MatInfo *info) 1549 { 1550 Mat_MPIBAIJ *a = (Mat_MPIBAIJ*)matin->data; 1551 Mat A = a->A,B = a->B; 1552 PetscErrorCode ierr; 1553 PetscReal isend[5],irecv[5]; 1554 1555 PetscFunctionBegin; 1556 info->block_size = (PetscReal)matin->rmap->bs; 1557 ierr = MatGetInfo(A,MAT_LOCAL,info);CHKERRQ(ierr); 1558 isend[0] = info->nz_used; isend[1] = info->nz_allocated; isend[2] = info->nz_unneeded; 1559 isend[3] = info->memory; isend[4] = info->mallocs; 1560 ierr = MatGetInfo(B,MAT_LOCAL,info);CHKERRQ(ierr); 1561 isend[0] += info->nz_used; isend[1] += info->nz_allocated; isend[2] += info->nz_unneeded; 1562 isend[3] += info->memory; isend[4] += info->mallocs; 1563 if (flag == MAT_LOCAL) { 1564 info->nz_used = isend[0]; 1565 info->nz_allocated = isend[1]; 1566 info->nz_unneeded = isend[2]; 1567 info->memory = isend[3]; 1568 info->mallocs = isend[4]; 1569 } else if (flag == MAT_GLOBAL_MAX) { 1570 ierr = MPI_Allreduce(isend,irecv,5,MPIU_REAL,MPIU_MAX,((PetscObject)matin)->comm);CHKERRQ(ierr); 1571 info->nz_used = irecv[0]; 1572 info->nz_allocated = irecv[1]; 1573 info->nz_unneeded = irecv[2]; 1574 info->memory = irecv[3]; 1575 info->mallocs = irecv[4]; 1576 } else if (flag == MAT_GLOBAL_SUM) { 1577 ierr = MPI_Allreduce(isend,irecv,5,MPIU_REAL,MPIU_SUM,((PetscObject)matin)->comm);CHKERRQ(ierr); 1578 info->nz_used = irecv[0]; 1579 info->nz_allocated = irecv[1]; 1580 info->nz_unneeded = irecv[2]; 1581 info->memory = irecv[3]; 1582 info->mallocs = irecv[4]; 1583 } else { 1584 SETERRQ1(((PetscObject)matin)->comm,PETSC_ERR_ARG_WRONG,"Unknown MatInfoType argument %d",(int)flag); 1585 } 1586 info->fill_ratio_given = 0; /* no parallel LU/ILU/Cholesky */ 1587 info->fill_ratio_needed = 0; 1588 info->factor_mallocs = 0; 1589 PetscFunctionReturn(0); 1590 } 1591 1592 #undef __FUNCT__ 1593 #define __FUNCT__ "MatSetOption_MPIBAIJ" 1594 PetscErrorCode MatSetOption_MPIBAIJ(Mat A,MatOption op,PetscBool flg) 1595 { 1596 Mat_MPIBAIJ *a = (Mat_MPIBAIJ*)A->data; 1597 PetscErrorCode ierr; 1598 1599 PetscFunctionBegin; 1600 switch (op) { 1601 case MAT_NEW_NONZERO_LOCATIONS: 1602 case MAT_NEW_NONZERO_ALLOCATION_ERR: 1603 case MAT_UNUSED_NONZERO_LOCATION_ERR: 1604 case MAT_KEEP_NONZERO_PATTERN: 1605 case MAT_NEW_NONZERO_LOCATION_ERR: 1606 ierr = MatSetOption(a->A,op,flg);CHKERRQ(ierr); 1607 ierr = MatSetOption(a->B,op,flg);CHKERRQ(ierr); 1608 break; 1609 case MAT_ROW_ORIENTED: 1610 a->roworiented = flg; 1611 ierr = MatSetOption(a->A,op,flg);CHKERRQ(ierr); 1612 ierr = MatSetOption(a->B,op,flg);CHKERRQ(ierr); 1613 break; 1614 case MAT_NEW_DIAGONALS: 1615 ierr = PetscInfo1(A,"Option %s ignored\n",MatOptions[op]);CHKERRQ(ierr); 1616 break; 1617 case MAT_IGNORE_OFF_PROC_ENTRIES: 1618 a->donotstash = flg; 1619 break; 1620 case MAT_USE_HASH_TABLE: 1621 a->ht_flag = flg; 1622 break; 1623 case MAT_SYMMETRIC: 1624 case MAT_STRUCTURALLY_SYMMETRIC: 1625 case MAT_HERMITIAN: 1626 case MAT_SYMMETRY_ETERNAL: 1627 ierr = MatSetOption(a->A,op,flg);CHKERRQ(ierr); 1628 break; 1629 default: 1630 SETERRQ1(((PetscObject)A)->comm,PETSC_ERR_SUP,"unknown option %d",op); 1631 } 1632 PetscFunctionReturn(0); 1633 } 1634 1635 #undef __FUNCT__ 1636 #define __FUNCT__ "MatTranspose_MPIBAIJ" 1637 PetscErrorCode MatTranspose_MPIBAIJ(Mat A,MatReuse reuse,Mat *matout) 1638 { 1639 Mat_MPIBAIJ *baij = (Mat_MPIBAIJ*)A->data; 1640 Mat_SeqBAIJ *Aloc; 1641 Mat B; 1642 PetscErrorCode ierr; 1643 PetscInt M=A->rmap->N,N=A->cmap->N,*ai,*aj,i,*rvals,j,k,col; 1644 PetscInt bs=A->rmap->bs,mbs=baij->mbs; 1645 MatScalar *a; 1646 1647 PetscFunctionBegin; 1648 if (reuse == MAT_REUSE_MATRIX && A == *matout && M != N) SETERRQ(((PetscObject)A)->comm,PETSC_ERR_ARG_SIZ,"Square matrix only for in-place"); 1649 if (reuse == MAT_INITIAL_MATRIX || *matout == A) { 1650 ierr = MatCreate(((PetscObject)A)->comm,&B);CHKERRQ(ierr); 1651 ierr = MatSetSizes(B,A->cmap->n,A->rmap->n,N,M);CHKERRQ(ierr); 1652 ierr = MatSetType(B,((PetscObject)A)->type_name);CHKERRQ(ierr); 1653 /* Do not know preallocation information, but must set block size */ 1654 ierr = MatMPIBAIJSetPreallocation(B,A->rmap->bs,PETSC_DECIDE,PETSC_NULL,PETSC_DECIDE,PETSC_NULL);CHKERRQ(ierr); 1655 } else { 1656 B = *matout; 1657 } 1658 1659 /* copy over the A part */ 1660 Aloc = (Mat_SeqBAIJ*)baij->A->data; 1661 ai = Aloc->i; aj = Aloc->j; a = Aloc->a; 1662 ierr = PetscMalloc(bs*sizeof(PetscInt),&rvals);CHKERRQ(ierr); 1663 1664 for (i=0; i<mbs; i++) { 1665 rvals[0] = bs*(baij->rstartbs + i); 1666 for (j=1; j<bs; j++) { rvals[j] = rvals[j-1] + 1; } 1667 for (j=ai[i]; j<ai[i+1]; j++) { 1668 col = (baij->cstartbs+aj[j])*bs; 1669 for (k=0; k<bs; k++) { 1670 ierr = MatSetValues_MPIBAIJ(B,1,&col,bs,rvals,a,INSERT_VALUES);CHKERRQ(ierr); 1671 col++; a += bs; 1672 } 1673 } 1674 } 1675 /* copy over the B part */ 1676 Aloc = (Mat_SeqBAIJ*)baij->B->data; 1677 ai = Aloc->i; aj = Aloc->j; a = Aloc->a; 1678 for (i=0; i<mbs; i++) { 1679 rvals[0] = bs*(baij->rstartbs + i); 1680 for (j=1; j<bs; j++) { rvals[j] = rvals[j-1] + 1; } 1681 for (j=ai[i]; j<ai[i+1]; j++) { 1682 col = baij->garray[aj[j]]*bs; 1683 for (k=0; k<bs; k++) { 1684 ierr = MatSetValues_MPIBAIJ(B,1,&col,bs,rvals,a,INSERT_VALUES);CHKERRQ(ierr); 1685 col++; a += bs; 1686 } 1687 } 1688 } 1689 ierr = PetscFree(rvals);CHKERRQ(ierr); 1690 ierr = MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1691 ierr = MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1692 1693 if (reuse == MAT_INITIAL_MATRIX || *matout != A) { 1694 *matout = B; 1695 } else { 1696 ierr = MatHeaderMerge(A,B);CHKERRQ(ierr); 1697 } 1698 PetscFunctionReturn(0); 1699 } 1700 1701 #undef __FUNCT__ 1702 #define __FUNCT__ "MatDiagonalScale_MPIBAIJ" 1703 PetscErrorCode MatDiagonalScale_MPIBAIJ(Mat mat,Vec ll,Vec rr) 1704 { 1705 Mat_MPIBAIJ *baij = (Mat_MPIBAIJ*)mat->data; 1706 Mat a = baij->A,b = baij->B; 1707 PetscErrorCode ierr; 1708 PetscInt s1,s2,s3; 1709 1710 PetscFunctionBegin; 1711 ierr = MatGetLocalSize(mat,&s2,&s3);CHKERRQ(ierr); 1712 if (rr) { 1713 ierr = VecGetLocalSize(rr,&s1);CHKERRQ(ierr); 1714 if (s1!=s3) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"right vector non-conforming local size"); 1715 /* Overlap communication with computation. */ 1716 ierr = VecScatterBegin(baij->Mvctx,rr,baij->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1717 } 1718 if (ll) { 1719 ierr = VecGetLocalSize(ll,&s1);CHKERRQ(ierr); 1720 if (s1!=s2) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"left vector non-conforming local size"); 1721 ierr = (*b->ops->diagonalscale)(b,ll,PETSC_NULL);CHKERRQ(ierr); 1722 } 1723 /* scale the diagonal block */ 1724 ierr = (*a->ops->diagonalscale)(a,ll,rr);CHKERRQ(ierr); 1725 1726 if (rr) { 1727 /* Do a scatter end and then right scale the off-diagonal block */ 1728 ierr = VecScatterEnd(baij->Mvctx,rr,baij->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1729 ierr = (*b->ops->diagonalscale)(b,PETSC_NULL,baij->lvec);CHKERRQ(ierr); 1730 } 1731 1732 PetscFunctionReturn(0); 1733 } 1734 1735 #undef __FUNCT__ 1736 #define __FUNCT__ "MatZeroRows_MPIBAIJ" 1737 PetscErrorCode MatZeroRows_MPIBAIJ(Mat A,PetscInt N,const PetscInt rows[],PetscScalar diag,Vec x,Vec b) 1738 { 1739 Mat_MPIBAIJ *l = (Mat_MPIBAIJ*)A->data; 1740 PetscErrorCode ierr; 1741 PetscMPIInt imdex,size = l->size,n,rank = l->rank; 1742 PetscInt i,*owners = A->rmap->range; 1743 PetscInt *nprocs,j,idx,nsends,row; 1744 PetscInt nmax,*svalues,*starts,*owner,nrecvs; 1745 PetscInt *rvalues,tag = ((PetscObject)A)->tag,count,base,slen,*source,lastidx = -1; 1746 PetscInt *lens,*lrows,*values,rstart_bs=A->rmap->rstart; 1747 MPI_Comm comm = ((PetscObject)A)->comm; 1748 MPI_Request *send_waits,*recv_waits; 1749 MPI_Status recv_status,*send_status; 1750 const PetscScalar *xx; 1751 PetscScalar *bb; 1752 #if defined(PETSC_DEBUG) 1753 PetscBool found = PETSC_FALSE; 1754 #endif 1755 1756 PetscFunctionBegin; 1757 /* first count number of contributors to each processor */ 1758 ierr = PetscMalloc(2*size*sizeof(PetscInt),&nprocs);CHKERRQ(ierr); 1759 ierr = PetscMemzero(nprocs,2*size*sizeof(PetscInt));CHKERRQ(ierr); 1760 ierr = PetscMalloc((N+1)*sizeof(PetscInt),&owner);CHKERRQ(ierr); /* see note*/ 1761 j = 0; 1762 for (i=0; i<N; i++) { 1763 if (lastidx > (idx = rows[i])) j = 0; 1764 lastidx = idx; 1765 for (; j<size; j++) { 1766 if (idx >= owners[j] && idx < owners[j+1]) { 1767 nprocs[2*j]++; 1768 nprocs[2*j+1] = 1; 1769 owner[i] = j; 1770 #if defined(PETSC_DEBUG) 1771 found = PETSC_TRUE; 1772 #endif 1773 break; 1774 } 1775 } 1776 #if defined(PETSC_DEBUG) 1777 if (!found) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Index out of range"); 1778 found = PETSC_FALSE; 1779 #endif 1780 } 1781 nsends = 0; for (i=0; i<size; i++) { nsends += nprocs[2*i+1];} 1782 1783 if (A->nooffproczerorows) { 1784 if (nsends > 1) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"You called MatSetOption(,MAT_NO_OFF_PROC_ZERO_ROWS,PETSC_TRUE) but set an off process zero row"); 1785 nrecvs = nsends; 1786 nmax = N; 1787 } else { 1788 /* inform other processors of number of messages and max length*/ 1789 ierr = PetscMaxSum(comm,nprocs,&nmax,&nrecvs);CHKERRQ(ierr); 1790 } 1791 1792 /* post receives: */ 1793 ierr = PetscMalloc((nrecvs+1)*(nmax+1)*sizeof(PetscInt),&rvalues);CHKERRQ(ierr); 1794 ierr = PetscMalloc((nrecvs+1)*sizeof(MPI_Request),&recv_waits);CHKERRQ(ierr); 1795 for (i=0; i<nrecvs; i++) { 1796 ierr = MPI_Irecv(rvalues+nmax*i,nmax,MPIU_INT,MPI_ANY_SOURCE,tag,comm,recv_waits+i);CHKERRQ(ierr); 1797 } 1798 1799 /* do sends: 1800 1) starts[i] gives the starting index in svalues for stuff going to 1801 the ith processor 1802 */ 1803 ierr = PetscMalloc((N+1)*sizeof(PetscInt),&svalues);CHKERRQ(ierr); 1804 ierr = PetscMalloc((nsends+1)*sizeof(MPI_Request),&send_waits);CHKERRQ(ierr); 1805 ierr = PetscMalloc((size+1)*sizeof(PetscInt),&starts);CHKERRQ(ierr); 1806 starts[0] = 0; 1807 for (i=1; i<size; i++) { starts[i] = starts[i-1] + nprocs[2*i-2];} 1808 for (i=0; i<N; i++) { 1809 svalues[starts[owner[i]]++] = rows[i]; 1810 } 1811 1812 starts[0] = 0; 1813 for (i=1; i<size+1; i++) { starts[i] = starts[i-1] + nprocs[2*i-2];} 1814 count = 0; 1815 for (i=0; i<size; i++) { 1816 if (nprocs[2*i+1]) { 1817 ierr = MPI_Isend(svalues+starts[i],nprocs[2*i],MPIU_INT,i,tag,comm,send_waits+count++);CHKERRQ(ierr); 1818 } 1819 } 1820 ierr = PetscFree(starts);CHKERRQ(ierr); 1821 1822 base = owners[rank]; 1823 1824 /* wait on receives */ 1825 ierr = PetscMalloc2(nrecvs+1,PetscInt,&lens,nrecvs+1,PetscInt,&source);CHKERRQ(ierr); 1826 count = nrecvs; 1827 slen = 0; 1828 while (count) { 1829 ierr = MPI_Waitany(nrecvs,recv_waits,&imdex,&recv_status);CHKERRQ(ierr); 1830 /* unpack receives into our local space */ 1831 ierr = MPI_Get_count(&recv_status,MPIU_INT,&n);CHKERRQ(ierr); 1832 source[imdex] = recv_status.MPI_SOURCE; 1833 lens[imdex] = n; 1834 slen += n; 1835 count--; 1836 } 1837 ierr = PetscFree(recv_waits);CHKERRQ(ierr); 1838 1839 /* move the data into the send scatter */ 1840 ierr = PetscMalloc((slen+1)*sizeof(PetscInt),&lrows);CHKERRQ(ierr); 1841 count = 0; 1842 for (i=0; i<nrecvs; i++) { 1843 values = rvalues + i*nmax; 1844 for (j=0; j<lens[i]; j++) { 1845 lrows[count++] = values[j] - base; 1846 } 1847 } 1848 ierr = PetscFree(rvalues);CHKERRQ(ierr); 1849 ierr = PetscFree2(lens,source);CHKERRQ(ierr); 1850 ierr = PetscFree(owner);CHKERRQ(ierr); 1851 ierr = PetscFree(nprocs);CHKERRQ(ierr); 1852 1853 /* fix right hand side if needed */ 1854 if (x && b) { 1855 ierr = VecGetArrayRead(x,&xx);CHKERRQ(ierr); 1856 ierr = VecGetArray(b,&bb);CHKERRQ(ierr); 1857 for (i=0; i<slen; i++) { 1858 bb[lrows[i]] = diag*xx[lrows[i]]; 1859 } 1860 ierr = VecRestoreArrayRead(x,&xx);CHKERRQ(ierr); 1861 ierr = VecRestoreArray(b,&bb);CHKERRQ(ierr); 1862 } 1863 1864 /* actually zap the local rows */ 1865 /* 1866 Zero the required rows. If the "diagonal block" of the matrix 1867 is square and the user wishes to set the diagonal we use separate 1868 code so that MatSetValues() is not called for each diagonal allocating 1869 new memory, thus calling lots of mallocs and slowing things down. 1870 1871 */ 1872 /* must zero l->B before l->A because the (diag) case below may put values into l->B*/ 1873 ierr = MatZeroRows_SeqBAIJ(l->B,slen,lrows,0.0,0,0);CHKERRQ(ierr); 1874 if ((diag != 0.0) && (l->A->rmap->N == l->A->cmap->N)) { 1875 ierr = MatZeroRows_SeqBAIJ(l->A,slen,lrows,diag,0,0);CHKERRQ(ierr); 1876 } else if (diag != 0.0) { 1877 ierr = MatZeroRows_SeqBAIJ(l->A,slen,lrows,0.0,0,0);CHKERRQ(ierr); 1878 if (((Mat_SeqBAIJ*)l->A->data)->nonew) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"MatZeroRows() on rectangular matrices cannot be used with the Mat options \n\ 1879 MAT_NEW_NONZERO_LOCATIONS,MAT_NEW_NONZERO_LOCATION_ERR,MAT_NEW_NONZERO_ALLOCATION_ERR"); 1880 for (i=0; i<slen; i++) { 1881 row = lrows[i] + rstart_bs; 1882 ierr = MatSetValues(A,1,&row,1,&row,&diag,INSERT_VALUES);CHKERRQ(ierr); 1883 } 1884 ierr = MatAssemblyBegin(A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1885 ierr = MatAssemblyEnd(A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1886 } else { 1887 ierr = MatZeroRows_SeqBAIJ(l->A,slen,lrows,0.0,0,0);CHKERRQ(ierr); 1888 } 1889 1890 ierr = PetscFree(lrows);CHKERRQ(ierr); 1891 1892 /* wait on sends */ 1893 if (nsends) { 1894 ierr = PetscMalloc(nsends*sizeof(MPI_Status),&send_status);CHKERRQ(ierr); 1895 ierr = MPI_Waitall(nsends,send_waits,send_status);CHKERRQ(ierr); 1896 ierr = PetscFree(send_status);CHKERRQ(ierr); 1897 } 1898 ierr = PetscFree(send_waits);CHKERRQ(ierr); 1899 ierr = PetscFree(svalues);CHKERRQ(ierr); 1900 1901 PetscFunctionReturn(0); 1902 } 1903 1904 #undef __FUNCT__ 1905 #define __FUNCT__ "MatSetUnfactored_MPIBAIJ" 1906 PetscErrorCode MatSetUnfactored_MPIBAIJ(Mat A) 1907 { 1908 Mat_MPIBAIJ *a = (Mat_MPIBAIJ*)A->data; 1909 PetscErrorCode ierr; 1910 1911 PetscFunctionBegin; 1912 ierr = MatSetUnfactored(a->A);CHKERRQ(ierr); 1913 PetscFunctionReturn(0); 1914 } 1915 1916 static PetscErrorCode MatDuplicate_MPIBAIJ(Mat,MatDuplicateOption,Mat *); 1917 1918 #undef __FUNCT__ 1919 #define __FUNCT__ "MatEqual_MPIBAIJ" 1920 PetscErrorCode MatEqual_MPIBAIJ(Mat A,Mat B,PetscBool *flag) 1921 { 1922 Mat_MPIBAIJ *matB = (Mat_MPIBAIJ*)B->data,*matA = (Mat_MPIBAIJ*)A->data; 1923 Mat a,b,c,d; 1924 PetscBool flg; 1925 PetscErrorCode ierr; 1926 1927 PetscFunctionBegin; 1928 a = matA->A; b = matA->B; 1929 c = matB->A; d = matB->B; 1930 1931 ierr = MatEqual(a,c,&flg);CHKERRQ(ierr); 1932 if (flg) { 1933 ierr = MatEqual(b,d,&flg);CHKERRQ(ierr); 1934 } 1935 ierr = MPI_Allreduce(&flg,flag,1,MPI_INT,MPI_LAND,((PetscObject)A)->comm);CHKERRQ(ierr); 1936 PetscFunctionReturn(0); 1937 } 1938 1939 #undef __FUNCT__ 1940 #define __FUNCT__ "MatCopy_MPIBAIJ" 1941 PetscErrorCode MatCopy_MPIBAIJ(Mat A,Mat B,MatStructure str) 1942 { 1943 PetscErrorCode ierr; 1944 Mat_MPIBAIJ *a = (Mat_MPIBAIJ *)A->data; 1945 Mat_MPIBAIJ *b = (Mat_MPIBAIJ *)B->data; 1946 1947 PetscFunctionBegin; 1948 /* If the two matrices don't have the same copy implementation, they aren't compatible for fast copy. */ 1949 if ((str != SAME_NONZERO_PATTERN) || (A->ops->copy != B->ops->copy)) { 1950 ierr = MatCopy_Basic(A,B,str);CHKERRQ(ierr); 1951 } else { 1952 ierr = MatCopy(a->A,b->A,str);CHKERRQ(ierr); 1953 ierr = MatCopy(a->B,b->B,str);CHKERRQ(ierr); 1954 } 1955 PetscFunctionReturn(0); 1956 } 1957 1958 #undef __FUNCT__ 1959 #define __FUNCT__ "MatSetUp_MPIBAIJ" 1960 PetscErrorCode MatSetUp_MPIBAIJ(Mat A) 1961 { 1962 PetscErrorCode ierr; 1963 1964 PetscFunctionBegin; 1965 ierr = MatMPIBAIJSetPreallocation(A,A->rmap->bs,PETSC_DEFAULT,0,PETSC_DEFAULT,0);CHKERRQ(ierr); 1966 PetscFunctionReturn(0); 1967 } 1968 1969 #undef __FUNCT__ 1970 #define __FUNCT__ "MatAXPY_MPIBAIJ" 1971 PetscErrorCode MatAXPY_MPIBAIJ(Mat Y,PetscScalar a,Mat X,MatStructure str) 1972 { 1973 PetscErrorCode ierr; 1974 Mat_MPIBAIJ *xx=(Mat_MPIBAIJ *)X->data,*yy=(Mat_MPIBAIJ *)Y->data; 1975 PetscBLASInt bnz,one=1; 1976 Mat_SeqBAIJ *x,*y; 1977 1978 PetscFunctionBegin; 1979 if (str == SAME_NONZERO_PATTERN) { 1980 PetscScalar alpha = a; 1981 x = (Mat_SeqBAIJ *)xx->A->data; 1982 y = (Mat_SeqBAIJ *)yy->A->data; 1983 bnz = PetscBLASIntCast(x->nz); 1984 BLASaxpy_(&bnz,&alpha,x->a,&one,y->a,&one); 1985 x = (Mat_SeqBAIJ *)xx->B->data; 1986 y = (Mat_SeqBAIJ *)yy->B->data; 1987 bnz = PetscBLASIntCast(x->nz); 1988 BLASaxpy_(&bnz,&alpha,x->a,&one,y->a,&one); 1989 } else { 1990 ierr = MatAXPY_Basic(Y,a,X,str);CHKERRQ(ierr); 1991 } 1992 PetscFunctionReturn(0); 1993 } 1994 1995 #undef __FUNCT__ 1996 #define __FUNCT__ "MatRealPart_MPIBAIJ" 1997 PetscErrorCode MatRealPart_MPIBAIJ(Mat A) 1998 { 1999 Mat_MPIBAIJ *a = (Mat_MPIBAIJ*)A->data; 2000 PetscErrorCode ierr; 2001 2002 PetscFunctionBegin; 2003 ierr = MatRealPart(a->A);CHKERRQ(ierr); 2004 ierr = MatRealPart(a->B);CHKERRQ(ierr); 2005 PetscFunctionReturn(0); 2006 } 2007 2008 #undef __FUNCT__ 2009 #define __FUNCT__ "MatImaginaryPart_MPIBAIJ" 2010 PetscErrorCode MatImaginaryPart_MPIBAIJ(Mat A) 2011 { 2012 Mat_MPIBAIJ *a = (Mat_MPIBAIJ*)A->data; 2013 PetscErrorCode ierr; 2014 2015 PetscFunctionBegin; 2016 ierr = MatImaginaryPart(a->A);CHKERRQ(ierr); 2017 ierr = MatImaginaryPart(a->B);CHKERRQ(ierr); 2018 PetscFunctionReturn(0); 2019 } 2020 2021 #undef __FUNCT__ 2022 #define __FUNCT__ "MatGetSubMatrix_MPIBAIJ" 2023 PetscErrorCode MatGetSubMatrix_MPIBAIJ(Mat mat,IS isrow,IS iscol,MatReuse call,Mat *newmat) 2024 { 2025 PetscErrorCode ierr; 2026 IS iscol_local; 2027 PetscInt csize; 2028 2029 PetscFunctionBegin; 2030 ierr = ISGetLocalSize(iscol,&csize);CHKERRQ(ierr); 2031 if (call == MAT_REUSE_MATRIX) { 2032 ierr = PetscObjectQuery((PetscObject)*newmat,"ISAllGather",(PetscObject*)&iscol_local);CHKERRQ(ierr); 2033 if (!iscol_local) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Submatrix passed in was not used before, cannot reuse"); 2034 } else { 2035 ierr = ISAllGather(iscol,&iscol_local);CHKERRQ(ierr); 2036 } 2037 ierr = MatGetSubMatrix_MPIBAIJ_Private(mat,isrow,iscol_local,csize,call,newmat);CHKERRQ(ierr); 2038 if (call == MAT_INITIAL_MATRIX) { 2039 ierr = PetscObjectCompose((PetscObject)*newmat,"ISAllGather",(PetscObject)iscol_local);CHKERRQ(ierr); 2040 ierr = ISDestroy(&iscol_local);CHKERRQ(ierr); 2041 } 2042 PetscFunctionReturn(0); 2043 } 2044 extern PetscErrorCode MatGetSubMatrices_MPIBAIJ_local(Mat,PetscInt,const IS[],const IS[],MatReuse,PetscBool*,PetscBool*,Mat*); 2045 #undef __FUNCT__ 2046 #define __FUNCT__ "MatGetSubMatrix_MPIBAIJ_Private" 2047 /* 2048 Not great since it makes two copies of the submatrix, first an SeqBAIJ 2049 in local and then by concatenating the local matrices the end result. 2050 Writing it directly would be much like MatGetSubMatrices_MPIBAIJ() 2051 */ 2052 PetscErrorCode MatGetSubMatrix_MPIBAIJ_Private(Mat mat,IS isrow,IS iscol,PetscInt csize,MatReuse call,Mat *newmat) 2053 { 2054 PetscErrorCode ierr; 2055 PetscMPIInt rank,size; 2056 PetscInt i,m,n,rstart,row,rend,nz,*cwork,j,bs; 2057 PetscInt *ii,*jj,nlocal,*dlens,*olens,dlen,olen,jend,mglobal,ncol,nrow; 2058 Mat M,Mreuse; 2059 MatScalar *vwork,*aa; 2060 MPI_Comm comm = ((PetscObject)mat)->comm; 2061 IS isrow_new, iscol_new; 2062 PetscBool idflag,allrows, allcols; 2063 Mat_SeqBAIJ *aij; 2064 2065 2066 PetscFunctionBegin; 2067 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 2068 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 2069 /* The compression and expansion should be avoided. Doesn't point 2070 out errors, might change the indices, hence buggey */ 2071 ierr = ISCompressIndicesGeneral(mat->rmap->N,mat->rmap->n,mat->rmap->bs,1,&isrow,&isrow_new);CHKERRQ(ierr); 2072 ierr = ISCompressIndicesGeneral(mat->cmap->N,mat->cmap->n,mat->cmap->bs,1,&iscol,&iscol_new);CHKERRQ(ierr); 2073 2074 /* Check for special case: each processor gets entire matrix columns */ 2075 ierr = ISIdentity(iscol,&idflag);CHKERRQ(ierr); 2076 ierr = ISGetLocalSize(iscol,&ncol);CHKERRQ(ierr); 2077 if (idflag && ncol == mat->cmap->N){ 2078 allcols = PETSC_TRUE; 2079 } else { 2080 allcols = PETSC_FALSE; 2081 } 2082 2083 ierr = ISIdentity(isrow,&idflag);CHKERRQ(ierr); 2084 ierr = ISGetLocalSize(isrow,&nrow);CHKERRQ(ierr); 2085 if (idflag && nrow == mat->rmap->N){ 2086 allrows = PETSC_TRUE; 2087 } else { 2088 allrows = PETSC_FALSE; 2089 } 2090 if (call == MAT_REUSE_MATRIX) { 2091 ierr = PetscObjectQuery((PetscObject)*newmat,"SubMatrix",(PetscObject *)&Mreuse);CHKERRQ(ierr); 2092 if (!Mreuse) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Submatrix passed in was not used before, cannot reuse"); 2093 ierr = MatGetSubMatrices_MPIBAIJ_local(mat,1,&isrow_new,&iscol_new,MAT_REUSE_MATRIX,&allrows,&allcols,&Mreuse);CHKERRQ(ierr); 2094 } else { 2095 ierr = MatGetSubMatrices_MPIBAIJ_local(mat,1,&isrow_new,&iscol_new,MAT_INITIAL_MATRIX,&allrows,&allcols,&Mreuse);CHKERRQ(ierr); 2096 } 2097 ierr = ISDestroy(&isrow_new); CHKERRQ(ierr); 2098 ierr = ISDestroy(&iscol_new); CHKERRQ(ierr); 2099 /* 2100 m - number of local rows 2101 n - number of columns (same on all processors) 2102 rstart - first row in new global matrix generated 2103 */ 2104 ierr = MatGetBlockSize(mat,&bs);CHKERRQ(ierr); 2105 ierr = MatGetSize(Mreuse,&m,&n);CHKERRQ(ierr); 2106 m = m/bs; 2107 n = n/bs; 2108 2109 if (call == MAT_INITIAL_MATRIX) { 2110 aij = (Mat_SeqBAIJ*)(Mreuse)->data; 2111 ii = aij->i; 2112 jj = aij->j; 2113 2114 /* 2115 Determine the number of non-zeros in the diagonal and off-diagonal 2116 portions of the matrix in order to do correct preallocation 2117 */ 2118 2119 /* first get start and end of "diagonal" columns */ 2120 if (csize == PETSC_DECIDE) { 2121 ierr = ISGetSize(isrow,&mglobal);CHKERRQ(ierr); 2122 if (mglobal == n*bs) { /* square matrix */ 2123 nlocal = m; 2124 } else { 2125 nlocal = n/size + ((n % size) > rank); 2126 } 2127 } else { 2128 nlocal = csize/bs; 2129 } 2130 ierr = MPI_Scan(&nlocal,&rend,1,MPIU_INT,MPI_SUM,comm);CHKERRQ(ierr); 2131 rstart = rend - nlocal; 2132 if (rank == size - 1 && rend != n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Local column sizes %D do not add up to total number of columns %D",rend,n); 2133 2134 /* next, compute all the lengths */ 2135 ierr = PetscMalloc((2*m+1)*sizeof(PetscInt),&dlens);CHKERRQ(ierr); 2136 olens = dlens + m; 2137 for (i=0; i<m; i++) { 2138 jend = ii[i+1] - ii[i]; 2139 olen = 0; 2140 dlen = 0; 2141 for (j=0; j<jend; j++) { 2142 if (*jj < rstart || *jj >= rend) olen++; 2143 else dlen++; 2144 jj++; 2145 } 2146 olens[i] = olen; 2147 dlens[i] = dlen; 2148 } 2149 ierr = MatCreate(comm,&M);CHKERRQ(ierr); 2150 ierr = MatSetSizes(M,bs*m,bs*nlocal,PETSC_DECIDE,bs*n);CHKERRQ(ierr); 2151 ierr = MatSetType(M,((PetscObject)mat)->type_name);CHKERRQ(ierr); 2152 ierr = MatMPIBAIJSetPreallocation(M,bs,0,dlens,0,olens);CHKERRQ(ierr); 2153 ierr = PetscFree(dlens);CHKERRQ(ierr); 2154 } else { 2155 PetscInt ml,nl; 2156 2157 M = *newmat; 2158 ierr = MatGetLocalSize(M,&ml,&nl);CHKERRQ(ierr); 2159 if (ml != m) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Previous matrix must be same size/layout as request"); 2160 ierr = MatZeroEntries(M);CHKERRQ(ierr); 2161 /* 2162 The next two lines are needed so we may call MatSetValues_MPIAIJ() below directly, 2163 rather than the slower MatSetValues(). 2164 */ 2165 M->was_assembled = PETSC_TRUE; 2166 M->assembled = PETSC_FALSE; 2167 } 2168 ierr = MatSetOption(M,MAT_ROW_ORIENTED,PETSC_FALSE);CHKERRQ(ierr); 2169 ierr = MatGetOwnershipRange(M,&rstart,&rend);CHKERRQ(ierr); 2170 aij = (Mat_SeqBAIJ*)(Mreuse)->data; 2171 ii = aij->i; 2172 jj = aij->j; 2173 aa = aij->a; 2174 for (i=0; i<m; i++) { 2175 row = rstart/bs + i; 2176 nz = ii[i+1] - ii[i]; 2177 cwork = jj; jj += nz; 2178 vwork = aa; aa += nz*bs*bs; 2179 ierr = MatSetValuesBlocked_MPIBAIJ(M,1,&row,nz,cwork,vwork,INSERT_VALUES);CHKERRQ(ierr); 2180 } 2181 2182 ierr = MatAssemblyBegin(M,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2183 ierr = MatAssemblyEnd(M,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2184 *newmat = M; 2185 2186 /* save submatrix used in processor for next request */ 2187 if (call == MAT_INITIAL_MATRIX) { 2188 ierr = PetscObjectCompose((PetscObject)M,"SubMatrix",(PetscObject)Mreuse);CHKERRQ(ierr); 2189 ierr = PetscObjectDereference((PetscObject)Mreuse);CHKERRQ(ierr); 2190 } 2191 2192 PetscFunctionReturn(0); 2193 } 2194 2195 #undef __FUNCT__ 2196 #define __FUNCT__ "MatPermute_MPIBAIJ" 2197 PetscErrorCode MatPermute_MPIBAIJ(Mat A,IS rowp,IS colp,Mat *B) 2198 { 2199 MPI_Comm comm,pcomm; 2200 PetscInt first,rlocal_size,clocal_size,nrows; 2201 const PetscInt *rows; 2202 PetscMPIInt size; 2203 IS crowp,growp,irowp,lrowp,lcolp; 2204 PetscErrorCode ierr; 2205 2206 PetscFunctionBegin; 2207 ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr); 2208 /* make a collective version of 'rowp' */ 2209 ierr = PetscObjectGetComm((PetscObject)rowp,&pcomm);CHKERRQ(ierr); 2210 if (pcomm==comm) { 2211 crowp = rowp; 2212 } else { 2213 ierr = ISGetSize(rowp,&nrows);CHKERRQ(ierr); 2214 ierr = ISGetIndices(rowp,&rows);CHKERRQ(ierr); 2215 ierr = ISCreateGeneral(comm,nrows,rows,PETSC_COPY_VALUES,&crowp);CHKERRQ(ierr); 2216 ierr = ISRestoreIndices(rowp,&rows);CHKERRQ(ierr); 2217 } 2218 /* collect the global row permutation and invert it */ 2219 ierr = ISAllGather(crowp,&growp);CHKERRQ(ierr); 2220 ierr = ISSetPermutation(growp);CHKERRQ(ierr); 2221 if (pcomm!=comm) { 2222 ierr = ISDestroy(&crowp);CHKERRQ(ierr); 2223 } 2224 ierr = ISInvertPermutation(growp,PETSC_DECIDE,&irowp);CHKERRQ(ierr); 2225 ierr = ISDestroy(&growp);CHKERRQ(ierr); 2226 /* get the local target indices */ 2227 ierr = MatGetOwnershipRange(A,&first,PETSC_NULL);CHKERRQ(ierr); 2228 ierr = MatGetLocalSize(A,&rlocal_size,&clocal_size);CHKERRQ(ierr); 2229 ierr = ISGetIndices(irowp,&rows);CHKERRQ(ierr); 2230 ierr = ISCreateGeneral(MPI_COMM_SELF,rlocal_size,rows+first,PETSC_COPY_VALUES,&lrowp);CHKERRQ(ierr); 2231 ierr = ISRestoreIndices(irowp,&rows);CHKERRQ(ierr); 2232 ierr = ISDestroy(&irowp);CHKERRQ(ierr); 2233 /* the column permutation is so much easier; 2234 make a local version of 'colp' and invert it */ 2235 ierr = PetscObjectGetComm((PetscObject)colp,&pcomm);CHKERRQ(ierr); 2236 ierr = MPI_Comm_size(pcomm,&size);CHKERRQ(ierr); 2237 if (size==1) { 2238 lcolp = colp; 2239 } else { 2240 ierr = ISAllGather(colp,&lcolp);CHKERRQ(ierr); 2241 } 2242 ierr = ISSetPermutation(lcolp);CHKERRQ(ierr); 2243 /* now we just get the submatrix */ 2244 ierr = MatGetSubMatrix_MPIBAIJ_Private(A,lrowp,lcolp,clocal_size,MAT_INITIAL_MATRIX,B);CHKERRQ(ierr); 2245 if (size>1) { 2246 ierr = ISDestroy(&lcolp);CHKERRQ(ierr); 2247 } 2248 /* clean up */ 2249 ierr = ISDestroy(&lrowp);CHKERRQ(ierr); 2250 PetscFunctionReturn(0); 2251 } 2252 2253 #undef __FUNCT__ 2254 #define __FUNCT__ "MatGetGhosts_MPIBAIJ" 2255 PetscErrorCode MatGetGhosts_MPIBAIJ(Mat mat,PetscInt *nghosts,const PetscInt *ghosts[]) 2256 { 2257 Mat_MPIBAIJ *baij = (Mat_MPIBAIJ*) mat->data; 2258 Mat_SeqBAIJ *B = (Mat_SeqBAIJ*)baij->B->data; 2259 2260 PetscFunctionBegin; 2261 if (nghosts) { *nghosts = B->nbs;} 2262 if (ghosts) {*ghosts = baij->garray;} 2263 PetscFunctionReturn(0); 2264 } 2265 2266 extern PetscErrorCode MatCreateColmap_MPIBAIJ_Private(Mat); 2267 2268 #undef __FUNCT__ 2269 #define __FUNCT__ "MatFDColoringCreate_MPIBAIJ" 2270 /* 2271 This routine is almost identical to MatFDColoringCreate_MPIBAIJ()! 2272 */ 2273 PetscErrorCode MatFDColoringCreate_MPIBAIJ(Mat mat,ISColoring iscoloring,MatFDColoring c) 2274 { 2275 Mat_MPIBAIJ *baij = (Mat_MPIBAIJ*)mat->data; 2276 PetscErrorCode ierr; 2277 PetscMPIInt size,*ncolsonproc,*disp,nn; 2278 PetscInt bs,i,n,nrows,j,k,m,ncols,col; 2279 const PetscInt *is,*rows = 0,*A_ci,*A_cj,*B_ci,*B_cj; 2280 PetscInt nis = iscoloring->n,nctot,*cols; 2281 PetscInt *rowhit,M,cstart,cend,colb; 2282 PetscInt *columnsforrow,l; 2283 IS *isa; 2284 PetscBool done,flg; 2285 ISLocalToGlobalMapping map = mat->cmap->bmapping; 2286 PetscInt *ltog = (map ? map->indices : (PetscInt*) PETSC_NULL) ,ctype=c->ctype; 2287 2288 PetscFunctionBegin; 2289 if (!mat->assembled) SETERRQ(((PetscObject)mat)->comm,PETSC_ERR_ARG_WRONGSTATE,"Matrix must be assembled first; MatAssemblyBegin/End();"); 2290 if (ctype == IS_COLORING_GHOSTED && !map) SETERRQ(((PetscObject)mat)->comm,PETSC_ERR_ARG_INCOMP,"When using ghosted differencing matrix must have local to global mapping provided with MatSetLocalToGlobalMappingBlock"); 2291 2292 ierr = ISColoringGetIS(iscoloring,PETSC_IGNORE,&isa);CHKERRQ(ierr); 2293 ierr = MatGetBlockSize(mat,&bs);CHKERRQ(ierr); 2294 M = mat->rmap->n/bs; 2295 cstart = mat->cmap->rstart/bs; 2296 cend = mat->cmap->rend/bs; 2297 c->M = mat->rmap->N/bs; /* set the global rows and columns and local rows */ 2298 c->N = mat->cmap->N/bs; 2299 c->m = mat->rmap->n/bs; 2300 c->rstart = mat->rmap->rstart/bs; 2301 2302 c->ncolors = nis; 2303 ierr = PetscMalloc(nis*sizeof(PetscInt),&c->ncolumns);CHKERRQ(ierr); 2304 ierr = PetscMalloc(nis*sizeof(PetscInt*),&c->columns);CHKERRQ(ierr); 2305 ierr = PetscMalloc(nis*sizeof(PetscInt),&c->nrows);CHKERRQ(ierr); 2306 ierr = PetscMalloc(nis*sizeof(PetscInt*),&c->rows);CHKERRQ(ierr); 2307 ierr = PetscMalloc(nis*sizeof(PetscInt*),&c->columnsforrow);CHKERRQ(ierr); 2308 ierr = PetscLogObjectMemory(c,5*nis*sizeof(PetscInt));CHKERRQ(ierr); 2309 2310 /* Allow access to data structures of local part of matrix */ 2311 if (!baij->colmap) { 2312 ierr = MatCreateColmap_MPIBAIJ_Private(mat);CHKERRQ(ierr); 2313 } 2314 ierr = MatGetColumnIJ(baij->A,0,PETSC_FALSE,PETSC_FALSE,&ncols,&A_ci,&A_cj,&done);CHKERRQ(ierr); 2315 ierr = MatGetColumnIJ(baij->B,0,PETSC_FALSE,PETSC_FALSE,&ncols,&B_ci,&B_cj,&done);CHKERRQ(ierr); 2316 2317 ierr = PetscMalloc((M+1)*sizeof(PetscInt),&rowhit);CHKERRQ(ierr); 2318 ierr = PetscMalloc((M+1)*sizeof(PetscInt),&columnsforrow);CHKERRQ(ierr); 2319 2320 for (i=0; i<nis; i++) { 2321 ierr = ISGetLocalSize(isa[i],&n);CHKERRQ(ierr); 2322 ierr = ISGetIndices(isa[i],&is);CHKERRQ(ierr); 2323 c->ncolumns[i] = n; 2324 if (n) { 2325 ierr = PetscMalloc(n*sizeof(PetscInt),&c->columns[i]);CHKERRQ(ierr); 2326 ierr = PetscLogObjectMemory(c,n*sizeof(PetscInt));CHKERRQ(ierr); 2327 ierr = PetscMemcpy(c->columns[i],is,n*sizeof(PetscInt));CHKERRQ(ierr); 2328 } else { 2329 c->columns[i] = 0; 2330 } 2331 2332 if (ctype == IS_COLORING_GLOBAL){ 2333 /* Determine the total (parallel) number of columns of this color */ 2334 ierr = MPI_Comm_size(((PetscObject)mat)->comm,&size);CHKERRQ(ierr); 2335 ierr = PetscMalloc2(size,PetscMPIInt,&ncolsonproc,size,PetscMPIInt,&disp);CHKERRQ(ierr); 2336 2337 nn = PetscMPIIntCast(n); 2338 ierr = MPI_Allgather(&nn,1,MPI_INT,ncolsonproc,1,MPI_INT,((PetscObject)mat)->comm);CHKERRQ(ierr); 2339 nctot = 0; for (j=0; j<size; j++) {nctot += ncolsonproc[j];} 2340 if (!nctot) { 2341 ierr = PetscInfo(mat,"Coloring of matrix has some unneeded colors with no corresponding rows\n");CHKERRQ(ierr); 2342 } 2343 2344 disp[0] = 0; 2345 for (j=1; j<size; j++) { 2346 disp[j] = disp[j-1] + ncolsonproc[j-1]; 2347 } 2348 2349 /* Get complete list of columns for color on each processor */ 2350 ierr = PetscMalloc((nctot+1)*sizeof(PetscInt),&cols);CHKERRQ(ierr); 2351 ierr = MPI_Allgatherv((void*)is,n,MPIU_INT,cols,ncolsonproc,disp,MPIU_INT,((PetscObject)mat)->comm);CHKERRQ(ierr); 2352 ierr = PetscFree2(ncolsonproc,disp);CHKERRQ(ierr); 2353 } else if (ctype == IS_COLORING_GHOSTED){ 2354 /* Determine local number of columns of this color on this process, including ghost points */ 2355 nctot = n; 2356 ierr = PetscMalloc((nctot+1)*sizeof(PetscInt),&cols);CHKERRQ(ierr); 2357 ierr = PetscMemcpy(cols,is,n*sizeof(PetscInt));CHKERRQ(ierr); 2358 } else { 2359 SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Not provided for this MatFDColoring type"); 2360 } 2361 2362 /* 2363 Mark all rows affect by these columns 2364 */ 2365 /* Temporary option to allow for debugging/testing */ 2366 flg = PETSC_FALSE; 2367 ierr = PetscOptionsGetBool(PETSC_NULL,"-matfdcoloring_slow",&flg,PETSC_NULL);CHKERRQ(ierr); 2368 if (!flg) {/*-----------------------------------------------------------------------------*/ 2369 /* crude, fast version */ 2370 ierr = PetscMemzero(rowhit,M*sizeof(PetscInt));CHKERRQ(ierr); 2371 /* loop over columns*/ 2372 for (j=0; j<nctot; j++) { 2373 if (ctype == IS_COLORING_GHOSTED) { 2374 col = ltog[cols[j]]; 2375 } else { 2376 col = cols[j]; 2377 } 2378 if (col >= cstart && col < cend) { 2379 /* column is in diagonal block of matrix */ 2380 rows = A_cj + A_ci[col-cstart]; 2381 m = A_ci[col-cstart+1] - A_ci[col-cstart]; 2382 } else { 2383 #if defined (PETSC_USE_CTABLE) 2384 ierr = PetscTableFind(baij->colmap,col+1,&colb);CHKERRQ(ierr); 2385 colb --; 2386 #else 2387 colb = baij->colmap[col] - 1; 2388 #endif 2389 if (colb == -1) { 2390 m = 0; 2391 } else { 2392 colb = colb/bs; 2393 rows = B_cj + B_ci[colb]; 2394 m = B_ci[colb+1] - B_ci[colb]; 2395 } 2396 } 2397 /* loop over columns marking them in rowhit */ 2398 for (k=0; k<m; k++) { 2399 rowhit[*rows++] = col + 1; 2400 } 2401 } 2402 2403 /* count the number of hits */ 2404 nrows = 0; 2405 for (j=0; j<M; j++) { 2406 if (rowhit[j]) nrows++; 2407 } 2408 c->nrows[i] = nrows; 2409 ierr = PetscMalloc((nrows+1)*sizeof(PetscInt),&c->rows[i]);CHKERRQ(ierr); 2410 ierr = PetscMalloc((nrows+1)*sizeof(PetscInt),&c->columnsforrow[i]);CHKERRQ(ierr); 2411 ierr = PetscLogObjectMemory(c,2*(nrows+1)*sizeof(PetscInt));CHKERRQ(ierr); 2412 nrows = 0; 2413 for (j=0; j<M; j++) { 2414 if (rowhit[j]) { 2415 c->rows[i][nrows] = j; 2416 c->columnsforrow[i][nrows] = rowhit[j] - 1; 2417 nrows++; 2418 } 2419 } 2420 } else {/*-------------------------------------------------------------------------------*/ 2421 /* slow version, using rowhit as a linked list */ 2422 PetscInt currentcol,fm,mfm; 2423 rowhit[M] = M; 2424 nrows = 0; 2425 /* loop over columns*/ 2426 for (j=0; j<nctot; j++) { 2427 if (ctype == IS_COLORING_GHOSTED) { 2428 col = ltog[cols[j]]; 2429 } else { 2430 col = cols[j]; 2431 } 2432 if (col >= cstart && col < cend) { 2433 /* column is in diagonal block of matrix */ 2434 rows = A_cj + A_ci[col-cstart]; 2435 m = A_ci[col-cstart+1] - A_ci[col-cstart]; 2436 } else { 2437 #if defined (PETSC_USE_CTABLE) 2438 ierr = PetscTableFind(baij->colmap,col+1,&colb);CHKERRQ(ierr); 2439 colb --; 2440 #else 2441 colb = baij->colmap[col] - 1; 2442 #endif 2443 if (colb == -1) { 2444 m = 0; 2445 } else { 2446 colb = colb/bs; 2447 rows = B_cj + B_ci[colb]; 2448 m = B_ci[colb+1] - B_ci[colb]; 2449 } 2450 } 2451 2452 /* loop over columns marking them in rowhit */ 2453 fm = M; /* fm points to first entry in linked list */ 2454 for (k=0; k<m; k++) { 2455 currentcol = *rows++; 2456 /* is it already in the list? */ 2457 do { 2458 mfm = fm; 2459 fm = rowhit[fm]; 2460 } while (fm < currentcol); 2461 /* not in list so add it */ 2462 if (fm != currentcol) { 2463 nrows++; 2464 columnsforrow[currentcol] = col; 2465 /* next three lines insert new entry into linked list */ 2466 rowhit[mfm] = currentcol; 2467 rowhit[currentcol] = fm; 2468 fm = currentcol; 2469 /* fm points to present position in list since we know the columns are sorted */ 2470 } else { 2471 SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Invalid coloring of matrix detected"); 2472 } 2473 } 2474 } 2475 c->nrows[i] = nrows; 2476 ierr = PetscMalloc((nrows+1)*sizeof(PetscInt),&c->rows[i]);CHKERRQ(ierr); 2477 ierr = PetscMalloc((nrows+1)*sizeof(PetscInt),&c->columnsforrow[i]);CHKERRQ(ierr); 2478 ierr = PetscLogObjectMemory(c,(nrows+1)*sizeof(PetscInt));CHKERRQ(ierr); 2479 /* now store the linked list of rows into c->rows[i] */ 2480 nrows = 0; 2481 fm = rowhit[M]; 2482 do { 2483 c->rows[i][nrows] = fm; 2484 c->columnsforrow[i][nrows++] = columnsforrow[fm]; 2485 fm = rowhit[fm]; 2486 } while (fm < M); 2487 } /* ---------------------------------------------------------------------------------------*/ 2488 ierr = PetscFree(cols);CHKERRQ(ierr); 2489 } 2490 2491 /* Optimize by adding the vscale, and scaleforrow[][] fields */ 2492 /* 2493 vscale will contain the "diagonal" on processor scalings followed by the off processor 2494 */ 2495 if (ctype == IS_COLORING_GLOBAL) { 2496 PetscInt *garray; 2497 ierr = PetscMalloc(baij->B->cmap->n*sizeof(PetscInt),&garray);CHKERRQ(ierr); 2498 for (i=0; i<baij->B->cmap->n/bs; i++) { 2499 for (j=0; j<bs; j++) { 2500 garray[i*bs+j] = bs*baij->garray[i]+j; 2501 } 2502 } 2503 ierr = VecCreateGhost(((PetscObject)mat)->comm,baij->A->rmap->n,PETSC_DETERMINE,baij->B->cmap->n,garray,&c->vscale);CHKERRQ(ierr); 2504 ierr = PetscFree(garray);CHKERRQ(ierr); 2505 CHKMEMQ; 2506 ierr = PetscMalloc(c->ncolors*sizeof(PetscInt*),&c->vscaleforrow);CHKERRQ(ierr); 2507 for (k=0; k<c->ncolors; k++) { 2508 ierr = PetscMalloc((c->nrows[k]+1)*sizeof(PetscInt),&c->vscaleforrow[k]);CHKERRQ(ierr); 2509 for (l=0; l<c->nrows[k]; l++) { 2510 col = c->columnsforrow[k][l]; 2511 if (col >= cstart && col < cend) { 2512 /* column is in diagonal block of matrix */ 2513 colb = col - cstart; 2514 } else { 2515 /* column is in "off-processor" part */ 2516 #if defined (PETSC_USE_CTABLE) 2517 ierr = PetscTableFind(baij->colmap,col+1,&colb);CHKERRQ(ierr); 2518 colb --; 2519 #else 2520 colb = baij->colmap[col] - 1; 2521 #endif 2522 colb = colb/bs; 2523 colb += cend - cstart; 2524 } 2525 c->vscaleforrow[k][l] = colb; 2526 } 2527 } 2528 } else if (ctype == IS_COLORING_GHOSTED) { 2529 /* Get gtol mapping */ 2530 PetscInt N = mat->cmap->N, *gtol; 2531 ierr = PetscMalloc((N+1)*sizeof(PetscInt),>ol);CHKERRQ(ierr); 2532 for (i=0; i<N; i++) gtol[i] = -1; 2533 for (i=0; i<map->n; i++) gtol[ltog[i]] = i; 2534 2535 c->vscale = 0; /* will be created in MatFDColoringApply() */ 2536 ierr = PetscMalloc(c->ncolors*sizeof(PetscInt*),&c->vscaleforrow);CHKERRQ(ierr); 2537 for (k=0; k<c->ncolors; k++) { 2538 ierr = PetscMalloc((c->nrows[k]+1)*sizeof(PetscInt),&c->vscaleforrow[k]);CHKERRQ(ierr); 2539 for (l=0; l<c->nrows[k]; l++) { 2540 col = c->columnsforrow[k][l]; /* global column index */ 2541 c->vscaleforrow[k][l] = gtol[col]; /* local column index */ 2542 } 2543 } 2544 ierr = PetscFree(gtol);CHKERRQ(ierr); 2545 } 2546 ierr = ISColoringRestoreIS(iscoloring,&isa);CHKERRQ(ierr); 2547 2548 ierr = PetscFree(rowhit);CHKERRQ(ierr); 2549 ierr = PetscFree(columnsforrow);CHKERRQ(ierr); 2550 ierr = MatRestoreColumnIJ(baij->A,0,PETSC_FALSE,PETSC_FALSE,&ncols,&A_ci,&A_cj,&done);CHKERRQ(ierr); 2551 ierr = MatRestoreColumnIJ(baij->B,0,PETSC_FALSE,PETSC_FALSE,&ncols,&B_ci,&B_cj,&done);CHKERRQ(ierr); 2552 CHKMEMQ; 2553 PetscFunctionReturn(0); 2554 } 2555 2556 #undef __FUNCT__ 2557 #define __FUNCT__ "MatGetSeqNonzeroStructure_MPIBAIJ" 2558 PetscErrorCode MatGetSeqNonzeroStructure_MPIBAIJ(Mat A,Mat *newmat) 2559 { 2560 Mat B; 2561 Mat_MPIBAIJ *a = (Mat_MPIBAIJ *)A->data; 2562 Mat_SeqBAIJ *ad = (Mat_SeqBAIJ*)a->A->data,*bd = (Mat_SeqBAIJ*)a->B->data; 2563 Mat_SeqAIJ *b; 2564 PetscErrorCode ierr; 2565 PetscMPIInt size,rank,*recvcounts = 0,*displs = 0; 2566 PetscInt sendcount,i,*rstarts = A->rmap->range,n,cnt,j,bs = A->rmap->bs; 2567 PetscInt m,*garray = a->garray,*lens,*jsendbuf,*a_jsendbuf,*b_jsendbuf; 2568 2569 PetscFunctionBegin; 2570 ierr = MPI_Comm_size(((PetscObject)A)->comm,&size);CHKERRQ(ierr); 2571 ierr = MPI_Comm_rank(((PetscObject)A)->comm,&rank);CHKERRQ(ierr); 2572 2573 /* ---------------------------------------------------------------- 2574 Tell every processor the number of nonzeros per row 2575 */ 2576 ierr = PetscMalloc((A->rmap->N/bs)*sizeof(PetscInt),&lens);CHKERRQ(ierr); 2577 for (i=A->rmap->rstart/bs; i<A->rmap->rend/bs; i++) { 2578 lens[i] = ad->i[i-A->rmap->rstart/bs+1] - ad->i[i-A->rmap->rstart/bs] + bd->i[i-A->rmap->rstart/bs+1] - bd->i[i-A->rmap->rstart/bs]; 2579 } 2580 sendcount = A->rmap->rend/bs - A->rmap->rstart/bs; 2581 ierr = PetscMalloc(2*size*sizeof(PetscMPIInt),&recvcounts);CHKERRQ(ierr); 2582 displs = recvcounts + size; 2583 for (i=0; i<size; i++) { 2584 recvcounts[i] = A->rmap->range[i+1]/bs - A->rmap->range[i]/bs; 2585 displs[i] = A->rmap->range[i]/bs; 2586 } 2587 #if defined(PETSC_HAVE_MPI_IN_PLACE) 2588 ierr = MPI_Allgatherv(MPI_IN_PLACE,0,MPI_DATATYPE_NULL,lens,recvcounts,displs,MPIU_INT,((PetscObject)A)->comm);CHKERRQ(ierr); 2589 #else 2590 ierr = MPI_Allgatherv(lens+A->rmap->rstart/bs,sendcount,MPIU_INT,lens,recvcounts,displs,MPIU_INT,((PetscObject)A)->comm);CHKERRQ(ierr); 2591 #endif 2592 /* --------------------------------------------------------------- 2593 Create the sequential matrix of the same type as the local block diagonal 2594 */ 2595 ierr = MatCreate(PETSC_COMM_SELF,&B);CHKERRQ(ierr); 2596 ierr = MatSetSizes(B,A->rmap->N/bs,A->cmap->N/bs,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 2597 ierr = MatSetType(B,MATSEQAIJ);CHKERRQ(ierr); 2598 ierr = MatSeqAIJSetPreallocation(B,0,lens);CHKERRQ(ierr); 2599 b = (Mat_SeqAIJ *)B->data; 2600 2601 /*-------------------------------------------------------------------- 2602 Copy my part of matrix column indices over 2603 */ 2604 sendcount = ad->nz + bd->nz; 2605 jsendbuf = b->j + b->i[rstarts[rank]/bs]; 2606 a_jsendbuf = ad->j; 2607 b_jsendbuf = bd->j; 2608 n = A->rmap->rend/bs - A->rmap->rstart/bs; 2609 cnt = 0; 2610 for (i=0; i<n; i++) { 2611 2612 /* put in lower diagonal portion */ 2613 m = bd->i[i+1] - bd->i[i]; 2614 while (m > 0) { 2615 /* is it above diagonal (in bd (compressed) numbering) */ 2616 if (garray[*b_jsendbuf] > A->rmap->rstart/bs + i) break; 2617 jsendbuf[cnt++] = garray[*b_jsendbuf++]; 2618 m--; 2619 } 2620 2621 /* put in diagonal portion */ 2622 for (j=ad->i[i]; j<ad->i[i+1]; j++) { 2623 jsendbuf[cnt++] = A->rmap->rstart/bs + *a_jsendbuf++; 2624 } 2625 2626 /* put in upper diagonal portion */ 2627 while (m-- > 0) { 2628 jsendbuf[cnt++] = garray[*b_jsendbuf++]; 2629 } 2630 } 2631 if (cnt != sendcount) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Corrupted PETSc matrix: nz given %D actual nz %D",sendcount,cnt); 2632 2633 /*-------------------------------------------------------------------- 2634 Gather all column indices to all processors 2635 */ 2636 for (i=0; i<size; i++) { 2637 recvcounts[i] = 0; 2638 for (j=A->rmap->range[i]/bs; j<A->rmap->range[i+1]/bs; j++) { 2639 recvcounts[i] += lens[j]; 2640 } 2641 } 2642 displs[0] = 0; 2643 for (i=1; i<size; i++) { 2644 displs[i] = displs[i-1] + recvcounts[i-1]; 2645 } 2646 #if defined(PETSC_HAVE_MPI_IN_PLACE) 2647 ierr = MPI_Allgatherv(MPI_IN_PLACE,0,MPI_DATATYPE_NULL,b->j,recvcounts,displs,MPIU_INT,((PetscObject)A)->comm);CHKERRQ(ierr); 2648 #else 2649 ierr = MPI_Allgatherv(jsendbuf,sendcount,MPIU_INT,b->j,recvcounts,displs,MPIU_INT,((PetscObject)A)->comm);CHKERRQ(ierr); 2650 #endif 2651 /*-------------------------------------------------------------------- 2652 Assemble the matrix into useable form (note numerical values not yet set) 2653 */ 2654 /* set the b->ilen (length of each row) values */ 2655 ierr = PetscMemcpy(b->ilen,lens,(A->rmap->N/bs)*sizeof(PetscInt));CHKERRQ(ierr); 2656 /* set the b->i indices */ 2657 b->i[0] = 0; 2658 for (i=1; i<=A->rmap->N/bs; i++) { 2659 b->i[i] = b->i[i-1] + lens[i-1]; 2660 } 2661 ierr = PetscFree(lens);CHKERRQ(ierr); 2662 ierr = MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2663 ierr = MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2664 ierr = PetscFree(recvcounts);CHKERRQ(ierr); 2665 2666 if (A->symmetric){ 2667 ierr = MatSetOption(B,MAT_SYMMETRIC,PETSC_TRUE);CHKERRQ(ierr); 2668 } else if (A->hermitian) { 2669 ierr = MatSetOption(B,MAT_HERMITIAN,PETSC_TRUE);CHKERRQ(ierr); 2670 } else if (A->structurally_symmetric) { 2671 ierr = MatSetOption(B,MAT_STRUCTURALLY_SYMMETRIC,PETSC_TRUE);CHKERRQ(ierr); 2672 } 2673 *newmat = B; 2674 PetscFunctionReturn(0); 2675 } 2676 2677 #undef __FUNCT__ 2678 #define __FUNCT__ "MatSOR_MPIBAIJ" 2679 PetscErrorCode MatSOR_MPIBAIJ(Mat matin,Vec bb,PetscReal omega,MatSORType flag,PetscReal fshift,PetscInt its,PetscInt lits,Vec xx) 2680 { 2681 Mat_MPIBAIJ *mat = (Mat_MPIBAIJ*)matin->data; 2682 PetscErrorCode ierr; 2683 Vec bb1 = 0; 2684 2685 PetscFunctionBegin; 2686 if (flag == SOR_APPLY_UPPER) { 2687 ierr = (*mat->A->ops->sor)(mat->A,bb,omega,flag,fshift,lits,1,xx);CHKERRQ(ierr); 2688 PetscFunctionReturn(0); 2689 } 2690 2691 if (its > 1 || ~flag & SOR_ZERO_INITIAL_GUESS) { 2692 ierr = VecDuplicate(bb,&bb1);CHKERRQ(ierr); 2693 } 2694 2695 if ((flag & SOR_LOCAL_SYMMETRIC_SWEEP) == SOR_LOCAL_SYMMETRIC_SWEEP){ 2696 if (flag & SOR_ZERO_INITIAL_GUESS) { 2697 ierr = (*mat->A->ops->sor)(mat->A,bb,omega,flag,fshift,lits,1,xx);CHKERRQ(ierr); 2698 its--; 2699 } 2700 2701 while (its--) { 2702 ierr = VecScatterBegin(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2703 ierr = VecScatterEnd(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2704 2705 /* update rhs: bb1 = bb - B*x */ 2706 ierr = VecScale(mat->lvec,-1.0);CHKERRQ(ierr); 2707 ierr = (*mat->B->ops->multadd)(mat->B,mat->lvec,bb,bb1);CHKERRQ(ierr); 2708 2709 /* local sweep */ 2710 ierr = (*mat->A->ops->sor)(mat->A,bb1,omega,SOR_SYMMETRIC_SWEEP,fshift,lits,1,xx);CHKERRQ(ierr); 2711 } 2712 } else if (flag & SOR_LOCAL_FORWARD_SWEEP){ 2713 if (flag & SOR_ZERO_INITIAL_GUESS) { 2714 ierr = (*mat->A->ops->sor)(mat->A,bb,omega,flag,fshift,lits,1,xx);CHKERRQ(ierr); 2715 its--; 2716 } 2717 while (its--) { 2718 ierr = VecScatterBegin(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2719 ierr = VecScatterEnd(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2720 2721 /* update rhs: bb1 = bb - B*x */ 2722 ierr = VecScale(mat->lvec,-1.0);CHKERRQ(ierr); 2723 ierr = (*mat->B->ops->multadd)(mat->B,mat->lvec,bb,bb1);CHKERRQ(ierr); 2724 2725 /* local sweep */ 2726 ierr = (*mat->A->ops->sor)(mat->A,bb1,omega,SOR_FORWARD_SWEEP,fshift,lits,1,xx);CHKERRQ(ierr); 2727 } 2728 } else if (flag & SOR_LOCAL_BACKWARD_SWEEP){ 2729 if (flag & SOR_ZERO_INITIAL_GUESS) { 2730 ierr = (*mat->A->ops->sor)(mat->A,bb,omega,flag,fshift,lits,1,xx);CHKERRQ(ierr); 2731 its--; 2732 } 2733 while (its--) { 2734 ierr = VecScatterBegin(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2735 ierr = VecScatterEnd(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2736 2737 /* update rhs: bb1 = bb - B*x */ 2738 ierr = VecScale(mat->lvec,-1.0);CHKERRQ(ierr); 2739 ierr = (*mat->B->ops->multadd)(mat->B,mat->lvec,bb,bb1);CHKERRQ(ierr); 2740 2741 /* local sweep */ 2742 ierr = (*mat->A->ops->sor)(mat->A,bb1,omega,SOR_BACKWARD_SWEEP,fshift,lits,1,xx);CHKERRQ(ierr); 2743 } 2744 } else SETERRQ(((PetscObject)matin)->comm,PETSC_ERR_SUP,"Parallel version of SOR requested not supported"); 2745 2746 ierr = VecDestroy(&bb1);CHKERRQ(ierr); 2747 PetscFunctionReturn(0); 2748 } 2749 2750 extern PetscErrorCode MatFDColoringApply_BAIJ(Mat,MatFDColoring,Vec,MatStructure*,void*); 2751 2752 #undef __FUNCT__ 2753 #define __FUNCT__ "MatInvertBlockDiagonal_MPIBAIJ" 2754 PetscErrorCode MatInvertBlockDiagonal_MPIBAIJ(Mat A,const PetscScalar **values) 2755 { 2756 Mat_MPIBAIJ *a = (Mat_MPIBAIJ*) A->data; 2757 PetscErrorCode ierr; 2758 2759 PetscFunctionBegin; 2760 ierr = MatInvertBlockDiagonal(a->A,values);CHKERRQ(ierr); 2761 PetscFunctionReturn(0); 2762 } 2763 2764 2765 /* -------------------------------------------------------------------*/ 2766 static struct _MatOps MatOps_Values = { 2767 MatSetValues_MPIBAIJ, 2768 MatGetRow_MPIBAIJ, 2769 MatRestoreRow_MPIBAIJ, 2770 MatMult_MPIBAIJ, 2771 /* 4*/ MatMultAdd_MPIBAIJ, 2772 MatMultTranspose_MPIBAIJ, 2773 MatMultTransposeAdd_MPIBAIJ, 2774 0, 2775 0, 2776 0, 2777 /*10*/ 0, 2778 0, 2779 0, 2780 MatSOR_MPIBAIJ, 2781 MatTranspose_MPIBAIJ, 2782 /*15*/ MatGetInfo_MPIBAIJ, 2783 MatEqual_MPIBAIJ, 2784 MatGetDiagonal_MPIBAIJ, 2785 MatDiagonalScale_MPIBAIJ, 2786 MatNorm_MPIBAIJ, 2787 /*20*/ MatAssemblyBegin_MPIBAIJ, 2788 MatAssemblyEnd_MPIBAIJ, 2789 MatSetOption_MPIBAIJ, 2790 MatZeroEntries_MPIBAIJ, 2791 /*24*/ MatZeroRows_MPIBAIJ, 2792 0, 2793 0, 2794 0, 2795 0, 2796 /*29*/ MatSetUp_MPIBAIJ, 2797 0, 2798 0, 2799 0, 2800 0, 2801 /*34*/ MatDuplicate_MPIBAIJ, 2802 0, 2803 0, 2804 0, 2805 0, 2806 /*39*/ MatAXPY_MPIBAIJ, 2807 MatGetSubMatrices_MPIBAIJ, 2808 MatIncreaseOverlap_MPIBAIJ, 2809 MatGetValues_MPIBAIJ, 2810 MatCopy_MPIBAIJ, 2811 /*44*/ 0, 2812 MatScale_MPIBAIJ, 2813 0, 2814 0, 2815 0, 2816 /*49*/ 0, 2817 0, 2818 0, 2819 0, 2820 0, 2821 /*54*/ MatFDColoringCreate_MPIBAIJ, 2822 0, 2823 MatSetUnfactored_MPIBAIJ, 2824 MatPermute_MPIBAIJ, 2825 MatSetValuesBlocked_MPIBAIJ, 2826 /*59*/ MatGetSubMatrix_MPIBAIJ, 2827 MatDestroy_MPIBAIJ, 2828 MatView_MPIBAIJ, 2829 0, 2830 0, 2831 /*64*/ 0, 2832 0, 2833 0, 2834 0, 2835 0, 2836 /*69*/ MatGetRowMaxAbs_MPIBAIJ, 2837 0, 2838 0, 2839 0, 2840 0, 2841 /*74*/ 0, 2842 MatFDColoringApply_BAIJ, 2843 0, 2844 0, 2845 0, 2846 /*79*/ 0, 2847 0, 2848 0, 2849 0, 2850 MatLoad_MPIBAIJ, 2851 /*84*/ 0, 2852 0, 2853 0, 2854 0, 2855 0, 2856 /*89*/ 0, 2857 0, 2858 0, 2859 0, 2860 0, 2861 /*94*/ 0, 2862 0, 2863 0, 2864 0, 2865 0, 2866 /*99*/ 0, 2867 0, 2868 0, 2869 0, 2870 0, 2871 /*104*/0, 2872 MatRealPart_MPIBAIJ, 2873 MatImaginaryPart_MPIBAIJ, 2874 0, 2875 0, 2876 /*109*/0, 2877 0, 2878 0, 2879 0, 2880 0, 2881 /*114*/MatGetSeqNonzeroStructure_MPIBAIJ, 2882 0, 2883 MatGetGhosts_MPIBAIJ, 2884 0, 2885 0, 2886 /*119*/0, 2887 0, 2888 0, 2889 0, 2890 0, 2891 /*124*/0, 2892 0, 2893 MatInvertBlockDiagonal_MPIBAIJ 2894 }; 2895 2896 EXTERN_C_BEGIN 2897 #undef __FUNCT__ 2898 #define __FUNCT__ "MatGetDiagonalBlock_MPIBAIJ" 2899 PetscErrorCode MatGetDiagonalBlock_MPIBAIJ(Mat A,Mat *a) 2900 { 2901 PetscFunctionBegin; 2902 *a = ((Mat_MPIBAIJ *)A->data)->A; 2903 PetscFunctionReturn(0); 2904 } 2905 EXTERN_C_END 2906 2907 EXTERN_C_BEGIN 2908 extern PetscErrorCode MatConvert_MPIBAIJ_MPISBAIJ(Mat, MatType,MatReuse,Mat*); 2909 EXTERN_C_END 2910 2911 EXTERN_C_BEGIN 2912 #undef __FUNCT__ 2913 #define __FUNCT__ "MatMPIBAIJSetPreallocationCSR_MPIBAIJ" 2914 PetscErrorCode MatMPIBAIJSetPreallocationCSR_MPIBAIJ(Mat B,PetscInt bs,const PetscInt ii[],const PetscInt jj[],const PetscScalar V[]) 2915 { 2916 PetscInt m,rstart,cstart,cend; 2917 PetscInt i,j,d,nz,nz_max=0,*d_nnz=0,*o_nnz=0; 2918 const PetscInt *JJ=0; 2919 PetscScalar *values=0; 2920 PetscErrorCode ierr; 2921 2922 PetscFunctionBegin; 2923 ierr = PetscLayoutSetBlockSize(B->rmap,bs);CHKERRQ(ierr); 2924 ierr = PetscLayoutSetBlockSize(B->cmap,bs);CHKERRQ(ierr); 2925 ierr = PetscLayoutSetUp(B->rmap);CHKERRQ(ierr); 2926 ierr = PetscLayoutSetUp(B->cmap);CHKERRQ(ierr); 2927 ierr = PetscLayoutGetBlockSize(B->rmap,&bs);CHKERRQ(ierr); 2928 m = B->rmap->n/bs; 2929 rstart = B->rmap->rstart/bs; 2930 cstart = B->cmap->rstart/bs; 2931 cend = B->cmap->rend/bs; 2932 2933 if (ii[0]) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"ii[0] must be 0 but it is %D",ii[0]); 2934 ierr = PetscMalloc2(m,PetscInt,&d_nnz,m,PetscInt,&o_nnz);CHKERRQ(ierr); 2935 for (i=0; i<m; i++) { 2936 nz = ii[i+1] - ii[i]; 2937 if (nz < 0) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Local row %D has a negative number of columns %D",i,nz); 2938 nz_max = PetscMax(nz_max,nz); 2939 JJ = jj + ii[i]; 2940 for (j=0; j<nz; j++) { 2941 if (*JJ >= cstart) break; 2942 JJ++; 2943 } 2944 d = 0; 2945 for (; j<nz; j++) { 2946 if (*JJ++ >= cend) break; 2947 d++; 2948 } 2949 d_nnz[i] = d; 2950 o_nnz[i] = nz - d; 2951 } 2952 ierr = MatMPIBAIJSetPreallocation(B,bs,0,d_nnz,0,o_nnz);CHKERRQ(ierr); 2953 ierr = PetscFree2(d_nnz,o_nnz);CHKERRQ(ierr); 2954 2955 values = (PetscScalar*)V; 2956 if (!values) { 2957 ierr = PetscMalloc(bs*bs*nz_max*sizeof(PetscScalar),&values);CHKERRQ(ierr); 2958 ierr = PetscMemzero(values,bs*bs*nz_max*sizeof(PetscScalar));CHKERRQ(ierr); 2959 } 2960 for (i=0; i<m; i++) { 2961 PetscInt row = i + rstart; 2962 PetscInt ncols = ii[i+1] - ii[i]; 2963 const PetscInt *icols = jj + ii[i]; 2964 const PetscScalar *svals = values + (V ? (bs*bs*ii[i]) : 0); 2965 ierr = MatSetValuesBlocked_MPIBAIJ(B,1,&row,ncols,icols,svals,INSERT_VALUES);CHKERRQ(ierr); 2966 } 2967 2968 if (!V) { ierr = PetscFree(values);CHKERRQ(ierr); } 2969 ierr = MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2970 ierr = MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2971 ierr = MatSetOption(B,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_TRUE);CHKERRQ(ierr); 2972 PetscFunctionReturn(0); 2973 } 2974 EXTERN_C_END 2975 2976 #undef __FUNCT__ 2977 #define __FUNCT__ "MatMPIBAIJSetPreallocationCSR" 2978 /*@C 2979 MatMPIBAIJSetPreallocationCSR - Allocates memory for a sparse parallel matrix in BAIJ format 2980 (the default parallel PETSc format). 2981 2982 Collective on MPI_Comm 2983 2984 Input Parameters: 2985 + A - the matrix 2986 . bs - the block size 2987 . i - the indices into j for the start of each local row (starts with zero) 2988 . j - the column indices for each local row (starts with zero) these must be sorted for each row 2989 - v - optional values in the matrix 2990 2991 Level: developer 2992 2993 .keywords: matrix, aij, compressed row, sparse, parallel 2994 2995 .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatMPIBAIJSetPreallocation(), MatCreateAIJ(), MPIAIJ 2996 @*/ 2997 PetscErrorCode MatMPIBAIJSetPreallocationCSR(Mat B,PetscInt bs,const PetscInt i[],const PetscInt j[], const PetscScalar v[]) 2998 { 2999 PetscErrorCode ierr; 3000 3001 PetscFunctionBegin; 3002 PetscValidHeaderSpecific(B,MAT_CLASSID,1); 3003 PetscValidType(B,1); 3004 PetscValidLogicalCollectiveInt(B,bs,2); 3005 ierr = PetscTryMethod(B,"MatMPIBAIJSetPreallocationCSR_C",(Mat,PetscInt,const PetscInt[],const PetscInt[],const PetscScalar[]),(B,bs,i,j,v));CHKERRQ(ierr); 3006 PetscFunctionReturn(0); 3007 } 3008 3009 EXTERN_C_BEGIN 3010 #undef __FUNCT__ 3011 #define __FUNCT__ "MatMPIBAIJSetPreallocation_MPIBAIJ" 3012 PetscErrorCode MatMPIBAIJSetPreallocation_MPIBAIJ(Mat B,PetscInt bs,PetscInt d_nz,PetscInt *d_nnz,PetscInt o_nz,PetscInt *o_nnz) 3013 { 3014 Mat_MPIBAIJ *b; 3015 PetscErrorCode ierr; 3016 PetscInt i; 3017 PetscBool d_realalloc = PETSC_FALSE,o_realalloc = PETSC_FALSE; 3018 3019 PetscFunctionBegin; 3020 if (d_nz >= 0 || d_nnz) d_realalloc = PETSC_TRUE; 3021 if (o_nz >= 0 || o_nnz) o_realalloc = PETSC_TRUE; 3022 3023 if (d_nz == PETSC_DEFAULT || d_nz == PETSC_DECIDE) d_nz = 5; 3024 if (o_nz == PETSC_DEFAULT || o_nz == PETSC_DECIDE) o_nz = 2; 3025 if (d_nz < 0) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"d_nz cannot be less than 0: value %D",d_nz); 3026 if (o_nz < 0) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"o_nz cannot be less than 0: value %D",o_nz); 3027 3028 ierr = PetscLayoutSetBlockSize(B->rmap,bs);CHKERRQ(ierr); 3029 ierr = PetscLayoutSetBlockSize(B->cmap,bs);CHKERRQ(ierr); 3030 ierr = PetscLayoutSetUp(B->rmap);CHKERRQ(ierr); 3031 ierr = PetscLayoutSetUp(B->cmap);CHKERRQ(ierr); 3032 ierr = PetscLayoutGetBlockSize(B->rmap,&bs);CHKERRQ(ierr); 3033 3034 if (d_nnz) { 3035 for (i=0; i<B->rmap->n/bs; i++) { 3036 if (d_nnz[i] < 0) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"d_nnz cannot be less than -1: local row %D value %D",i,d_nnz[i]); 3037 } 3038 } 3039 if (o_nnz) { 3040 for (i=0; i<B->rmap->n/bs; i++) { 3041 if (o_nnz[i] < 0) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"o_nnz cannot be less than -1: local row %D value %D",i,o_nnz[i]); 3042 } 3043 } 3044 3045 b = (Mat_MPIBAIJ*)B->data; 3046 b->bs2 = bs*bs; 3047 b->mbs = B->rmap->n/bs; 3048 b->nbs = B->cmap->n/bs; 3049 b->Mbs = B->rmap->N/bs; 3050 b->Nbs = B->cmap->N/bs; 3051 3052 for (i=0; i<=b->size; i++) { 3053 b->rangebs[i] = B->rmap->range[i]/bs; 3054 } 3055 b->rstartbs = B->rmap->rstart/bs; 3056 b->rendbs = B->rmap->rend/bs; 3057 b->cstartbs = B->cmap->rstart/bs; 3058 b->cendbs = B->cmap->rend/bs; 3059 3060 if (!B->preallocated) { 3061 ierr = MatCreate(PETSC_COMM_SELF,&b->A);CHKERRQ(ierr); 3062 ierr = MatSetSizes(b->A,B->rmap->n,B->cmap->n,B->rmap->n,B->cmap->n);CHKERRQ(ierr); 3063 ierr = MatSetType(b->A,MATSEQBAIJ);CHKERRQ(ierr); 3064 ierr = PetscLogObjectParent(B,b->A);CHKERRQ(ierr); 3065 ierr = MatCreate(PETSC_COMM_SELF,&b->B);CHKERRQ(ierr); 3066 ierr = MatSetSizes(b->B,B->rmap->n,B->cmap->N,B->rmap->n,B->cmap->N);CHKERRQ(ierr); 3067 ierr = MatSetType(b->B,MATSEQBAIJ);CHKERRQ(ierr); 3068 ierr = PetscLogObjectParent(B,b->B);CHKERRQ(ierr); 3069 ierr = MatStashCreate_Private(((PetscObject)B)->comm,bs,&B->bstash);CHKERRQ(ierr); 3070 } 3071 3072 ierr = MatSeqBAIJSetPreallocation(b->A,bs,d_nz,d_nnz);CHKERRQ(ierr); 3073 ierr = MatSeqBAIJSetPreallocation(b->B,bs,o_nz,o_nnz);CHKERRQ(ierr); 3074 /* Do not error if the user did not give real preallocation information. Ugly because this would overwrite a previous user call to MatSetOption(). */ 3075 if (!d_realalloc) {ierr = MatSetOption(b->A,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr);} 3076 if (!o_realalloc) {ierr = MatSetOption(b->B,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr);} 3077 B->preallocated = PETSC_TRUE; 3078 PetscFunctionReturn(0); 3079 } 3080 EXTERN_C_END 3081 3082 EXTERN_C_BEGIN 3083 extern PetscErrorCode MatDiagonalScaleLocal_MPIBAIJ(Mat,Vec); 3084 extern PetscErrorCode MatSetHashTableFactor_MPIBAIJ(Mat,PetscReal); 3085 EXTERN_C_END 3086 3087 3088 EXTERN_C_BEGIN 3089 #undef __FUNCT__ 3090 #define __FUNCT__ "MatConvert_MPIBAIJ_MPIAdj" 3091 PetscErrorCode MatConvert_MPIBAIJ_MPIAdj(Mat B, MatType newtype,MatReuse reuse,Mat *adj) 3092 { 3093 Mat_MPIBAIJ *b = (Mat_MPIBAIJ*)B->data; 3094 PetscErrorCode ierr; 3095 Mat_SeqBAIJ *d = (Mat_SeqBAIJ*) b->A->data,*o = (Mat_SeqBAIJ*) b->B->data; 3096 PetscInt M = B->rmap->n/B->rmap->bs,i,*ii,*jj,cnt,j,k,rstart = B->rmap->rstart/B->rmap->bs; 3097 const PetscInt *id = d->i, *jd = d->j, *io = o->i, *jo = o->j, *garray = b->garray; 3098 3099 PetscFunctionBegin; 3100 ierr = PetscMalloc((M+1)*sizeof(PetscInt),&ii);CHKERRQ(ierr); 3101 ii[0] = 0; 3102 CHKMEMQ; 3103 for (i=0; i<M; i++) { 3104 if ((id[i+1] - id[i]) < 0) SETERRQ3(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Indices wrong %D %D %D",i,id[i],id[i+1]); 3105 if ((io[i+1] - io[i]) < 0) SETERRQ3(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Indices wrong %D %D %D",i,io[i],io[i+1]); 3106 ii[i+1] = ii[i] + id[i+1] - id[i] + io[i+1] - io[i]; 3107 CHKMEMQ; 3108 } 3109 ierr = PetscMalloc(ii[M]*sizeof(PetscInt),&jj);CHKERRQ(ierr); 3110 cnt = 0; 3111 for (i=0; i<M; i++) { 3112 for (j=io[i]; j<io[i+1]; j++) { 3113 if (garray[jo[j]] > rstart) break; 3114 jj[cnt++] = garray[jo[j]]; 3115 CHKMEMQ; 3116 } 3117 for (k=id[i]; k<id[i+1]; k++) { 3118 jj[cnt++] = rstart + jd[k]; 3119 CHKMEMQ; 3120 } 3121 for (;j<io[i+1]; j++) { 3122 jj[cnt++] = garray[jo[j]]; 3123 CHKMEMQ; 3124 } 3125 } 3126 ierr = MatCreateMPIAdj(((PetscObject)B)->comm,M,B->cmap->N/B->rmap->bs,ii,jj,PETSC_NULL,adj);CHKERRQ(ierr); 3127 PetscFunctionReturn(0); 3128 } 3129 EXTERN_C_END 3130 3131 #include <../src/mat/impls/aij/mpi/mpiaij.h> 3132 EXTERN_C_BEGIN 3133 PetscErrorCode MatConvert_SeqBAIJ_SeqAIJ(Mat,MatType,MatReuse,Mat*); 3134 EXTERN_C_END 3135 3136 EXTERN_C_BEGIN 3137 #undef __FUNCT__ 3138 #define __FUNCT__ "MatConvert_MPIBAIJ_MPIAIJ" 3139 PetscErrorCode MatConvert_MPIBAIJ_MPIAIJ(Mat A,MatType newtype,MatReuse reuse,Mat *newmat) 3140 { 3141 PetscErrorCode ierr; 3142 Mat_MPIBAIJ *a = (Mat_MPIBAIJ*)A->data; 3143 Mat B; 3144 Mat_MPIAIJ *b; 3145 3146 PetscFunctionBegin; 3147 if (!A->assembled) SETERRQ(((PetscObject)A)->comm,PETSC_ERR_SUP,"Matrix must be assembled"); 3148 3149 ierr = MatCreate(((PetscObject)A)->comm,&B);CHKERRQ(ierr); 3150 ierr = MatSetSizes(B,A->rmap->n,A->cmap->n,A->rmap->N,A->cmap->N);CHKERRQ(ierr); 3151 ierr = MatSetType(B,MATMPIAIJ);CHKERRQ(ierr); 3152 ierr = MatSeqAIJSetPreallocation(B,0,PETSC_NULL);CHKERRQ(ierr); 3153 ierr = MatMPIAIJSetPreallocation(B,0,PETSC_NULL,0,PETSC_NULL);CHKERRQ(ierr); 3154 b = (Mat_MPIAIJ*) B->data; 3155 3156 ierr = MatDestroy(&b->A);CHKERRQ(ierr); 3157 ierr = MatDestroy(&b->B);CHKERRQ(ierr); 3158 ierr = MatDisAssemble_MPIBAIJ(A);CHKERRQ(ierr); 3159 ierr = MatConvert_SeqBAIJ_SeqAIJ(a->A, MATSEQAIJ, MAT_INITIAL_MATRIX, &b->A);CHKERRQ(ierr); 3160 ierr = MatConvert_SeqBAIJ_SeqAIJ(a->B, MATSEQAIJ, MAT_INITIAL_MATRIX, &b->B);CHKERRQ(ierr); 3161 ierr = MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3162 ierr = MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3163 ierr = MatAssemblyBegin(A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3164 ierr = MatAssemblyEnd(A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3165 if (reuse == MAT_REUSE_MATRIX) { 3166 ierr = MatHeaderReplace(A,B);CHKERRQ(ierr); 3167 } else { 3168 *newmat = B; 3169 } 3170 PetscFunctionReturn(0); 3171 } 3172 EXTERN_C_END 3173 3174 EXTERN_C_BEGIN 3175 #if defined(PETSC_HAVE_MUMPS) 3176 extern PetscErrorCode MatGetFactor_baij_mumps(Mat,MatFactorType,Mat*); 3177 #endif 3178 EXTERN_C_END 3179 3180 /*MC 3181 MATMPIBAIJ - MATMPIBAIJ = "mpibaij" - A matrix type to be used for distributed block sparse matrices. 3182 3183 Options Database Keys: 3184 + -mat_type mpibaij - sets the matrix type to "mpibaij" during a call to MatSetFromOptions() 3185 . -mat_block_size <bs> - set the blocksize used to store the matrix 3186 - -mat_use_hash_table <fact> 3187 3188 Level: beginner 3189 3190 .seealso: MatCreateMPIBAIJ 3191 M*/ 3192 3193 EXTERN_C_BEGIN 3194 extern PetscErrorCode MatConvert_MPIBAIJ_MPIBSTRM(Mat,MatType,MatReuse,Mat*); 3195 EXTERN_C_END 3196 3197 EXTERN_C_BEGIN 3198 #undef __FUNCT__ 3199 #define __FUNCT__ "MatCreate_MPIBAIJ" 3200 PetscErrorCode MatCreate_MPIBAIJ(Mat B) 3201 { 3202 Mat_MPIBAIJ *b; 3203 PetscErrorCode ierr; 3204 PetscBool flg; 3205 3206 PetscFunctionBegin; 3207 ierr = PetscNewLog(B,Mat_MPIBAIJ,&b);CHKERRQ(ierr); 3208 B->data = (void*)b; 3209 3210 ierr = PetscMemcpy(B->ops,&MatOps_Values,sizeof(struct _MatOps));CHKERRQ(ierr); 3211 B->assembled = PETSC_FALSE; 3212 3213 B->insertmode = NOT_SET_VALUES; 3214 ierr = MPI_Comm_rank(((PetscObject)B)->comm,&b->rank);CHKERRQ(ierr); 3215 ierr = MPI_Comm_size(((PetscObject)B)->comm,&b->size);CHKERRQ(ierr); 3216 3217 /* build local table of row and column ownerships */ 3218 ierr = PetscMalloc((b->size+1)*sizeof(PetscInt),&b->rangebs);CHKERRQ(ierr); 3219 3220 /* build cache for off array entries formed */ 3221 ierr = MatStashCreate_Private(((PetscObject)B)->comm,1,&B->stash);CHKERRQ(ierr); 3222 b->donotstash = PETSC_FALSE; 3223 b->colmap = PETSC_NULL; 3224 b->garray = PETSC_NULL; 3225 b->roworiented = PETSC_TRUE; 3226 3227 /* stuff used in block assembly */ 3228 b->barray = 0; 3229 3230 /* stuff used for matrix vector multiply */ 3231 b->lvec = 0; 3232 b->Mvctx = 0; 3233 3234 /* stuff for MatGetRow() */ 3235 b->rowindices = 0; 3236 b->rowvalues = 0; 3237 b->getrowactive = PETSC_FALSE; 3238 3239 /* hash table stuff */ 3240 b->ht = 0; 3241 b->hd = 0; 3242 b->ht_size = 0; 3243 b->ht_flag = PETSC_FALSE; 3244 b->ht_fact = 0; 3245 b->ht_total_ct = 0; 3246 b->ht_insert_ct = 0; 3247 3248 /* stuff for MatGetSubMatrices_MPIBAIJ_local() */ 3249 b->ijonly = PETSC_FALSE; 3250 3251 ierr = PetscOptionsBegin(((PetscObject)B)->comm,PETSC_NULL,"Options for loading MPIBAIJ matrix 1","Mat");CHKERRQ(ierr); 3252 ierr = PetscOptionsBool("-mat_use_hash_table","Use hash table to save memory in constructing matrix","MatSetOption",PETSC_FALSE,&flg,PETSC_NULL);CHKERRQ(ierr); 3253 if (flg) { 3254 PetscReal fact = 1.39; 3255 ierr = MatSetOption(B,MAT_USE_HASH_TABLE,PETSC_TRUE);CHKERRQ(ierr); 3256 ierr = PetscOptionsReal("-mat_use_hash_table","Use hash table factor","MatMPIBAIJSetHashTableFactor",fact,&fact,PETSC_NULL);CHKERRQ(ierr); 3257 if (fact <= 1.0) fact = 1.39; 3258 ierr = MatMPIBAIJSetHashTableFactor(B,fact);CHKERRQ(ierr); 3259 ierr = PetscInfo1(B,"Hash table Factor used %5.2f\n",fact);CHKERRQ(ierr); 3260 } 3261 ierr = PetscOptionsEnd();CHKERRQ(ierr); 3262 3263 #if defined(PETSC_HAVE_MUMPS) 3264 ierr = PetscObjectComposeFunctionDynamic((PetscObject)B,"MatGetFactor_mumps_C", "MatGetFactor_baij_mumps",MatGetFactor_baij_mumps);CHKERRQ(ierr); 3265 #endif 3266 ierr = PetscObjectComposeFunctionDynamic((PetscObject)B,"MatConvert_mpibaij_mpiadj_C", 3267 "MatConvert_MPIBAIJ_MPIAdj", 3268 MatConvert_MPIBAIJ_MPIAdj);CHKERRQ(ierr); 3269 ierr = PetscObjectComposeFunctionDynamic((PetscObject)B,"MatConvert_mpibaij_mpiaij_C", 3270 "MatConvert_MPIBAIJ_MPIAIJ", 3271 MatConvert_MPIBAIJ_MPIAIJ);CHKERRQ(ierr); 3272 ierr = PetscObjectComposeFunctionDynamic((PetscObject)B,"MatConvert_mpibaij_mpisbaij_C", 3273 "MatConvert_MPIBAIJ_MPISBAIJ", 3274 MatConvert_MPIBAIJ_MPISBAIJ);CHKERRQ(ierr); 3275 ierr = PetscObjectComposeFunctionDynamic((PetscObject)B,"MatStoreValues_C", 3276 "MatStoreValues_MPIBAIJ", 3277 MatStoreValues_MPIBAIJ);CHKERRQ(ierr); 3278 ierr = PetscObjectComposeFunctionDynamic((PetscObject)B,"MatRetrieveValues_C", 3279 "MatRetrieveValues_MPIBAIJ", 3280 MatRetrieveValues_MPIBAIJ);CHKERRQ(ierr); 3281 ierr = PetscObjectComposeFunctionDynamic((PetscObject)B,"MatGetDiagonalBlock_C", 3282 "MatGetDiagonalBlock_MPIBAIJ", 3283 MatGetDiagonalBlock_MPIBAIJ);CHKERRQ(ierr); 3284 ierr = PetscObjectComposeFunctionDynamic((PetscObject)B,"MatMPIBAIJSetPreallocation_C", 3285 "MatMPIBAIJSetPreallocation_MPIBAIJ", 3286 MatMPIBAIJSetPreallocation_MPIBAIJ);CHKERRQ(ierr); 3287 ierr = PetscObjectComposeFunctionDynamic((PetscObject)B,"MatMPIBAIJSetPreallocationCSR_C", 3288 "MatMPIBAIJSetPreallocationCSR_MPIBAIJ", 3289 MatMPIBAIJSetPreallocationCSR_MPIBAIJ);CHKERRQ(ierr); 3290 ierr = PetscObjectComposeFunctionDynamic((PetscObject)B,"MatDiagonalScaleLocal_C", 3291 "MatDiagonalScaleLocal_MPIBAIJ", 3292 MatDiagonalScaleLocal_MPIBAIJ);CHKERRQ(ierr); 3293 ierr = PetscObjectComposeFunctionDynamic((PetscObject)B,"MatSetHashTableFactor_C", 3294 "MatSetHashTableFactor_MPIBAIJ", 3295 MatSetHashTableFactor_MPIBAIJ);CHKERRQ(ierr); 3296 ierr = PetscObjectComposeFunctionDynamic((PetscObject)B,"MatConvert_mpibaij_mpibstrm_C", 3297 "MatConvert_MPIBAIJ_MPIBSTRM", 3298 MatConvert_MPIBAIJ_MPIBSTRM);CHKERRQ(ierr); 3299 ierr = PetscObjectChangeTypeName((PetscObject)B,MATMPIBAIJ);CHKERRQ(ierr); 3300 PetscFunctionReturn(0); 3301 } 3302 EXTERN_C_END 3303 3304 /*MC 3305 MATBAIJ - MATBAIJ = "baij" - A matrix type to be used for block sparse matrices. 3306 3307 This matrix type is identical to MATSEQBAIJ when constructed with a single process communicator, 3308 and MATMPIBAIJ otherwise. 3309 3310 Options Database Keys: 3311 . -mat_type baij - sets the matrix type to "baij" during a call to MatSetFromOptions() 3312 3313 Level: beginner 3314 3315 .seealso: MatCreateBAIJ(),MATSEQBAIJ,MATMPIBAIJ, MatMPIBAIJSetPreallocation(), MatMPIBAIJSetPreallocationCSR() 3316 M*/ 3317 3318 #undef __FUNCT__ 3319 #define __FUNCT__ "MatMPIBAIJSetPreallocation" 3320 /*@C 3321 MatMPIBAIJSetPreallocation - Allocates memory for a sparse parallel matrix in block AIJ format 3322 (block compressed row). For good matrix assembly performance 3323 the user should preallocate the matrix storage by setting the parameters 3324 d_nz (or d_nnz) and o_nz (or o_nnz). By setting these parameters accurately, 3325 performance can be increased by more than a factor of 50. 3326 3327 Collective on Mat 3328 3329 Input Parameters: 3330 + A - the matrix 3331 . bs - size of blockk 3332 . d_nz - number of block nonzeros per block row in diagonal portion of local 3333 submatrix (same for all local rows) 3334 . d_nnz - array containing the number of block nonzeros in the various block rows 3335 of the in diagonal portion of the local (possibly different for each block 3336 row) or PETSC_NULL. If you plan to factor the matrix you must leave room for the diagonal entry and 3337 set it even if it is zero. 3338 . o_nz - number of block nonzeros per block row in the off-diagonal portion of local 3339 submatrix (same for all local rows). 3340 - o_nnz - array containing the number of nonzeros in the various block rows of the 3341 off-diagonal portion of the local submatrix (possibly different for 3342 each block row) or PETSC_NULL. 3343 3344 If the *_nnz parameter is given then the *_nz parameter is ignored 3345 3346 Options Database Keys: 3347 + -mat_block_size - size of the blocks to use 3348 - -mat_use_hash_table <fact> 3349 3350 Notes: 3351 If PETSC_DECIDE or PETSC_DETERMINE is used for a particular argument on one processor 3352 than it must be used on all processors that share the object for that argument. 3353 3354 Storage Information: 3355 For a square global matrix we define each processor's diagonal portion 3356 to be its local rows and the corresponding columns (a square submatrix); 3357 each processor's off-diagonal portion encompasses the remainder of the 3358 local matrix (a rectangular submatrix). 3359 3360 The user can specify preallocated storage for the diagonal part of 3361 the local submatrix with either d_nz or d_nnz (not both). Set 3362 d_nz=PETSC_DEFAULT and d_nnz=PETSC_NULL for PETSc to control dynamic 3363 memory allocation. Likewise, specify preallocated storage for the 3364 off-diagonal part of the local submatrix with o_nz or o_nnz (not both). 3365 3366 Consider a processor that owns rows 3, 4 and 5 of a parallel matrix. In 3367 the figure below we depict these three local rows and all columns (0-11). 3368 3369 .vb 3370 0 1 2 3 4 5 6 7 8 9 10 11 3371 ------------------- 3372 row 3 | o o o d d d o o o o o o 3373 row 4 | o o o d d d o o o o o o 3374 row 5 | o o o d d d o o o o o o 3375 ------------------- 3376 .ve 3377 3378 Thus, any entries in the d locations are stored in the d (diagonal) 3379 submatrix, and any entries in the o locations are stored in the 3380 o (off-diagonal) submatrix. Note that the d and the o submatrices are 3381 stored simply in the MATSEQBAIJ format for compressed row storage. 3382 3383 Now d_nz should indicate the number of block nonzeros per row in the d matrix, 3384 and o_nz should indicate the number of block nonzeros per row in the o matrix. 3385 In general, for PDE problems in which most nonzeros are near the diagonal, 3386 one expects d_nz >> o_nz. For large problems you MUST preallocate memory 3387 or you will get TERRIBLE performance; see the users' manual chapter on 3388 matrices. 3389 3390 You can call MatGetInfo() to get information on how effective the preallocation was; 3391 for example the fields mallocs,nz_allocated,nz_used,nz_unneeded; 3392 You can also run with the option -info and look for messages with the string 3393 malloc in them to see if additional memory allocation was needed. 3394 3395 Level: intermediate 3396 3397 .keywords: matrix, block, aij, compressed row, sparse, parallel 3398 3399 .seealso: MatCreate(), MatCreateSeqBAIJ(), MatSetValues(), MatCreateBAIJ(), MatMPIBAIJSetPreallocationCSR(), PetscSplitOwnership() 3400 @*/ 3401 PetscErrorCode MatMPIBAIJSetPreallocation(Mat B,PetscInt bs,PetscInt d_nz,const PetscInt d_nnz[],PetscInt o_nz,const PetscInt o_nnz[]) 3402 { 3403 PetscErrorCode ierr; 3404 3405 PetscFunctionBegin; 3406 PetscValidHeaderSpecific(B,MAT_CLASSID,1); 3407 PetscValidType(B,1); 3408 PetscValidLogicalCollectiveInt(B,bs,2); 3409 ierr = PetscTryMethod(B,"MatMPIBAIJSetPreallocation_C",(Mat,PetscInt,PetscInt,const PetscInt[],PetscInt,const PetscInt[]),(B,bs,d_nz,d_nnz,o_nz,o_nnz));CHKERRQ(ierr); 3410 PetscFunctionReturn(0); 3411 } 3412 3413 #undef __FUNCT__ 3414 #define __FUNCT__ "MatCreateBAIJ" 3415 /*@C 3416 MatCreateBAIJ - Creates a sparse parallel matrix in block AIJ format 3417 (block compressed row). For good matrix assembly performance 3418 the user should preallocate the matrix storage by setting the parameters 3419 d_nz (or d_nnz) and o_nz (or o_nnz). By setting these parameters accurately, 3420 performance can be increased by more than a factor of 50. 3421 3422 Collective on MPI_Comm 3423 3424 Input Parameters: 3425 + comm - MPI communicator 3426 . bs - size of blockk 3427 . m - number of local rows (or PETSC_DECIDE to have calculated if M is given) 3428 This value should be the same as the local size used in creating the 3429 y vector for the matrix-vector product y = Ax. 3430 . n - number of local columns (or PETSC_DECIDE to have calculated if N is given) 3431 This value should be the same as the local size used in creating the 3432 x vector for the matrix-vector product y = Ax. 3433 . M - number of global rows (or PETSC_DETERMINE to have calculated if m is given) 3434 . N - number of global columns (or PETSC_DETERMINE to have calculated if n is given) 3435 . d_nz - number of nonzero blocks per block row in diagonal portion of local 3436 submatrix (same for all local rows) 3437 . d_nnz - array containing the number of nonzero blocks in the various block rows 3438 of the in diagonal portion of the local (possibly different for each block 3439 row) or PETSC_NULL. If you plan to factor the matrix you must leave room for the diagonal entry 3440 and set it even if it is zero. 3441 . o_nz - number of nonzero blocks per block row in the off-diagonal portion of local 3442 submatrix (same for all local rows). 3443 - o_nnz - array containing the number of nonzero blocks in the various block rows of the 3444 off-diagonal portion of the local submatrix (possibly different for 3445 each block row) or PETSC_NULL. 3446 3447 Output Parameter: 3448 . A - the matrix 3449 3450 Options Database Keys: 3451 + -mat_block_size - size of the blocks to use 3452 - -mat_use_hash_table <fact> 3453 3454 It is recommended that one use the MatCreate(), MatSetType() and/or MatSetFromOptions(), 3455 MatXXXXSetPreallocation() paradgm instead of this routine directly. 3456 [MatXXXXSetPreallocation() is, for example, MatSeqAIJSetPreallocation] 3457 3458 Notes: 3459 If the *_nnz parameter is given then the *_nz parameter is ignored 3460 3461 A nonzero block is any block that as 1 or more nonzeros in it 3462 3463 The user MUST specify either the local or global matrix dimensions 3464 (possibly both). 3465 3466 If PETSC_DECIDE or PETSC_DETERMINE is used for a particular argument on one processor 3467 than it must be used on all processors that share the object for that argument. 3468 3469 Storage Information: 3470 For a square global matrix we define each processor's diagonal portion 3471 to be its local rows and the corresponding columns (a square submatrix); 3472 each processor's off-diagonal portion encompasses the remainder of the 3473 local matrix (a rectangular submatrix). 3474 3475 The user can specify preallocated storage for the diagonal part of 3476 the local submatrix with either d_nz or d_nnz (not both). Set 3477 d_nz=PETSC_DEFAULT and d_nnz=PETSC_NULL for PETSc to control dynamic 3478 memory allocation. Likewise, specify preallocated storage for the 3479 off-diagonal part of the local submatrix with o_nz or o_nnz (not both). 3480 3481 Consider a processor that owns rows 3, 4 and 5 of a parallel matrix. In 3482 the figure below we depict these three local rows and all columns (0-11). 3483 3484 .vb 3485 0 1 2 3 4 5 6 7 8 9 10 11 3486 ------------------- 3487 row 3 | o o o d d d o o o o o o 3488 row 4 | o o o d d d o o o o o o 3489 row 5 | o o o d d d o o o o o o 3490 ------------------- 3491 .ve 3492 3493 Thus, any entries in the d locations are stored in the d (diagonal) 3494 submatrix, and any entries in the o locations are stored in the 3495 o (off-diagonal) submatrix. Note that the d and the o submatrices are 3496 stored simply in the MATSEQBAIJ format for compressed row storage. 3497 3498 Now d_nz should indicate the number of block nonzeros per row in the d matrix, 3499 and o_nz should indicate the number of block nonzeros per row in the o matrix. 3500 In general, for PDE problems in which most nonzeros are near the diagonal, 3501 one expects d_nz >> o_nz. For large problems you MUST preallocate memory 3502 or you will get TERRIBLE performance; see the users' manual chapter on 3503 matrices. 3504 3505 Level: intermediate 3506 3507 .keywords: matrix, block, aij, compressed row, sparse, parallel 3508 3509 .seealso: MatCreate(), MatCreateSeqBAIJ(), MatSetValues(), MatCreateBAIJ(), MatMPIBAIJSetPreallocation(), MatMPIBAIJSetPreallocationCSR() 3510 @*/ 3511 PetscErrorCode MatCreateBAIJ(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) 3512 { 3513 PetscErrorCode ierr; 3514 PetscMPIInt size; 3515 3516 PetscFunctionBegin; 3517 ierr = MatCreate(comm,A);CHKERRQ(ierr); 3518 ierr = MatSetSizes(*A,m,n,M,N);CHKERRQ(ierr); 3519 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 3520 if (size > 1) { 3521 ierr = MatSetType(*A,MATMPIBAIJ);CHKERRQ(ierr); 3522 ierr = MatMPIBAIJSetPreallocation(*A,bs,d_nz,d_nnz,o_nz,o_nnz);CHKERRQ(ierr); 3523 } else { 3524 ierr = MatSetType(*A,MATSEQBAIJ);CHKERRQ(ierr); 3525 ierr = MatSeqBAIJSetPreallocation(*A,bs,d_nz,d_nnz);CHKERRQ(ierr); 3526 } 3527 PetscFunctionReturn(0); 3528 } 3529 3530 #undef __FUNCT__ 3531 #define __FUNCT__ "MatDuplicate_MPIBAIJ" 3532 static PetscErrorCode MatDuplicate_MPIBAIJ(Mat matin,MatDuplicateOption cpvalues,Mat *newmat) 3533 { 3534 Mat mat; 3535 Mat_MPIBAIJ *a,*oldmat = (Mat_MPIBAIJ*)matin->data; 3536 PetscErrorCode ierr; 3537 PetscInt len=0; 3538 3539 PetscFunctionBegin; 3540 *newmat = 0; 3541 ierr = MatCreate(((PetscObject)matin)->comm,&mat);CHKERRQ(ierr); 3542 ierr = MatSetSizes(mat,matin->rmap->n,matin->cmap->n,matin->rmap->N,matin->cmap->N);CHKERRQ(ierr); 3543 ierr = MatSetType(mat,((PetscObject)matin)->type_name);CHKERRQ(ierr); 3544 ierr = PetscMemcpy(mat->ops,matin->ops,sizeof(struct _MatOps));CHKERRQ(ierr); 3545 3546 mat->factortype = matin->factortype; 3547 mat->preallocated = PETSC_TRUE; 3548 mat->assembled = PETSC_TRUE; 3549 mat->insertmode = NOT_SET_VALUES; 3550 3551 a = (Mat_MPIBAIJ*)mat->data; 3552 mat->rmap->bs = matin->rmap->bs; 3553 a->bs2 = oldmat->bs2; 3554 a->mbs = oldmat->mbs; 3555 a->nbs = oldmat->nbs; 3556 a->Mbs = oldmat->Mbs; 3557 a->Nbs = oldmat->Nbs; 3558 3559 ierr = PetscLayoutReference(matin->rmap,&mat->rmap);CHKERRQ(ierr); 3560 ierr = PetscLayoutReference(matin->cmap,&mat->cmap);CHKERRQ(ierr); 3561 3562 a->size = oldmat->size; 3563 a->rank = oldmat->rank; 3564 a->donotstash = oldmat->donotstash; 3565 a->roworiented = oldmat->roworiented; 3566 a->rowindices = 0; 3567 a->rowvalues = 0; 3568 a->getrowactive = PETSC_FALSE; 3569 a->barray = 0; 3570 a->rstartbs = oldmat->rstartbs; 3571 a->rendbs = oldmat->rendbs; 3572 a->cstartbs = oldmat->cstartbs; 3573 a->cendbs = oldmat->cendbs; 3574 3575 /* hash table stuff */ 3576 a->ht = 0; 3577 a->hd = 0; 3578 a->ht_size = 0; 3579 a->ht_flag = oldmat->ht_flag; 3580 a->ht_fact = oldmat->ht_fact; 3581 a->ht_total_ct = 0; 3582 a->ht_insert_ct = 0; 3583 3584 ierr = PetscMemcpy(a->rangebs,oldmat->rangebs,(a->size+1)*sizeof(PetscInt));CHKERRQ(ierr); 3585 if (oldmat->colmap) { 3586 #if defined (PETSC_USE_CTABLE) 3587 ierr = PetscTableCreateCopy(oldmat->colmap,&a->colmap);CHKERRQ(ierr); 3588 #else 3589 ierr = PetscMalloc((a->Nbs)*sizeof(PetscInt),&a->colmap);CHKERRQ(ierr); 3590 ierr = PetscLogObjectMemory(mat,(a->Nbs)*sizeof(PetscInt));CHKERRQ(ierr); 3591 ierr = PetscMemcpy(a->colmap,oldmat->colmap,(a->Nbs)*sizeof(PetscInt));CHKERRQ(ierr); 3592 #endif 3593 } else a->colmap = 0; 3594 3595 if (oldmat->garray && (len = ((Mat_SeqBAIJ*)(oldmat->B->data))->nbs)) { 3596 ierr = PetscMalloc(len*sizeof(PetscInt),&a->garray);CHKERRQ(ierr); 3597 ierr = PetscLogObjectMemory(mat,len*sizeof(PetscInt));CHKERRQ(ierr); 3598 ierr = PetscMemcpy(a->garray,oldmat->garray,len*sizeof(PetscInt));CHKERRQ(ierr); 3599 } else a->garray = 0; 3600 3601 ierr = MatStashCreate_Private(((PetscObject)matin)->comm,matin->rmap->bs,&mat->bstash);CHKERRQ(ierr); 3602 ierr = VecDuplicate(oldmat->lvec,&a->lvec);CHKERRQ(ierr); 3603 ierr = PetscLogObjectParent(mat,a->lvec);CHKERRQ(ierr); 3604 ierr = VecScatterCopy(oldmat->Mvctx,&a->Mvctx);CHKERRQ(ierr); 3605 ierr = PetscLogObjectParent(mat,a->Mvctx);CHKERRQ(ierr); 3606 3607 ierr = MatDuplicate(oldmat->A,cpvalues,&a->A);CHKERRQ(ierr); 3608 ierr = PetscLogObjectParent(mat,a->A);CHKERRQ(ierr); 3609 ierr = MatDuplicate(oldmat->B,cpvalues,&a->B);CHKERRQ(ierr); 3610 ierr = PetscLogObjectParent(mat,a->B);CHKERRQ(ierr); 3611 ierr = PetscFListDuplicate(((PetscObject)matin)->qlist,&((PetscObject)mat)->qlist);CHKERRQ(ierr); 3612 *newmat = mat; 3613 3614 PetscFunctionReturn(0); 3615 } 3616 3617 #undef __FUNCT__ 3618 #define __FUNCT__ "MatLoad_MPIBAIJ" 3619 PetscErrorCode MatLoad_MPIBAIJ(Mat newmat,PetscViewer viewer) 3620 { 3621 PetscErrorCode ierr; 3622 int fd; 3623 PetscInt i,nz,j,rstart,rend; 3624 PetscScalar *vals,*buf; 3625 MPI_Comm comm = ((PetscObject)viewer)->comm; 3626 MPI_Status status; 3627 PetscMPIInt rank,size,maxnz; 3628 PetscInt header[4],*rowlengths = 0,M,N,m,*rowners,*cols; 3629 PetscInt *locrowlens = PETSC_NULL,*procsnz = PETSC_NULL,*browners = PETSC_NULL; 3630 PetscInt jj,*mycols,*ibuf,bs=1,Mbs,mbs,extra_rows,mmax; 3631 PetscMPIInt tag = ((PetscObject)viewer)->tag; 3632 PetscInt *dlens = PETSC_NULL,*odlens = PETSC_NULL,*mask = PETSC_NULL,*masked1 = PETSC_NULL,*masked2 = PETSC_NULL,rowcount,odcount; 3633 PetscInt dcount,kmax,k,nzcount,tmp,mend,sizesset=1,grows,gcols; 3634 3635 PetscFunctionBegin; 3636 ierr = PetscOptionsBegin(comm,PETSC_NULL,"Options for loading MPIBAIJ matrix 2","Mat");CHKERRQ(ierr); 3637 ierr = PetscOptionsInt("-matload_block_size","Set the blocksize used to store the matrix","MatLoad",bs,&bs,PETSC_NULL);CHKERRQ(ierr); 3638 ierr = PetscOptionsEnd();CHKERRQ(ierr); 3639 3640 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 3641 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 3642 if (!rank) { 3643 ierr = PetscViewerBinaryGetDescriptor(viewer,&fd);CHKERRQ(ierr); 3644 ierr = PetscBinaryRead(fd,(char *)header,4,PETSC_INT);CHKERRQ(ierr); 3645 if (header[0] != MAT_FILE_CLASSID) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED,"not matrix object"); 3646 } 3647 3648 if (newmat->rmap->n < 0 && newmat->rmap->N < 0 && newmat->cmap->n < 0 && newmat->cmap->N < 0) sizesset = 0; 3649 3650 ierr = MPI_Bcast(header+1,3,MPIU_INT,0,comm);CHKERRQ(ierr); 3651 M = header[1]; N = header[2]; 3652 3653 /* If global rows/cols are set to PETSC_DECIDE, set it to the sizes given in the file */ 3654 if (sizesset && newmat->rmap->N < 0) newmat->rmap->N = M; 3655 if (sizesset && newmat->cmap->N < 0) newmat->cmap->N = N; 3656 3657 /* If global sizes are set, check if they are consistent with that given in the file */ 3658 if (sizesset) { 3659 ierr = MatGetSize(newmat,&grows,&gcols);CHKERRQ(ierr); 3660 } 3661 if (sizesset && newmat->rmap->N != grows) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED, "Inconsistent # of rows:Matrix in file has (%d) and input matrix has (%d)",M,grows); 3662 if (sizesset && newmat->cmap->N != gcols) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED, "Inconsistent # of cols:Matrix in file has (%d) and input matrix has (%d)",N,gcols); 3663 3664 if (M != N) SETERRQ(((PetscObject)viewer)->comm,PETSC_ERR_SUP,"Can only do square matrices"); 3665 3666 /* 3667 This code adds extra rows to make sure the number of rows is 3668 divisible by the blocksize 3669 */ 3670 Mbs = M/bs; 3671 extra_rows = bs - M + bs*Mbs; 3672 if (extra_rows == bs) extra_rows = 0; 3673 else Mbs++; 3674 if (extra_rows && !rank) { 3675 ierr = PetscInfo(viewer,"Padding loaded matrix to match blocksize\n");CHKERRQ(ierr); 3676 } 3677 3678 /* determine ownership of all rows */ 3679 if (newmat->rmap->n < 0) { /* PETSC_DECIDE */ 3680 mbs = Mbs/size + ((Mbs % size) > rank); 3681 m = mbs*bs; 3682 } else { /* User set */ 3683 m = newmat->rmap->n; 3684 mbs = m/bs; 3685 } 3686 ierr = PetscMalloc2(size+1,PetscInt,&rowners,size+1,PetscInt,&browners);CHKERRQ(ierr); 3687 ierr = MPI_Allgather(&mbs,1,MPIU_INT,rowners+1,1,MPIU_INT,comm);CHKERRQ(ierr); 3688 3689 /* process 0 needs enough room for process with most rows */ 3690 if (!rank) { 3691 mmax = rowners[1]; 3692 for (i=2; i<=size; i++) { 3693 mmax = PetscMax(mmax,rowners[i]); 3694 } 3695 mmax*=bs; 3696 } else mmax = m; 3697 3698 rowners[0] = 0; 3699 for (i=2; i<=size; i++) rowners[i] += rowners[i-1]; 3700 for (i=0; i<=size; i++) browners[i] = rowners[i]*bs; 3701 rstart = rowners[rank]; 3702 rend = rowners[rank+1]; 3703 3704 /* distribute row lengths to all processors */ 3705 ierr = PetscMalloc((mmax+1)*sizeof(PetscInt),&locrowlens);CHKERRQ(ierr); 3706 if (!rank) { 3707 mend = m; 3708 if (size == 1) mend = mend - extra_rows; 3709 ierr = PetscBinaryRead(fd,locrowlens,mend,PETSC_INT);CHKERRQ(ierr); 3710 for (j=mend; j<m; j++) locrowlens[j] = 1; 3711 ierr = PetscMalloc(m*sizeof(PetscInt),&rowlengths);CHKERRQ(ierr); 3712 ierr = PetscMalloc(size*sizeof(PetscInt),&procsnz);CHKERRQ(ierr); 3713 ierr = PetscMemzero(procsnz,size*sizeof(PetscInt));CHKERRQ(ierr); 3714 for (j=0; j<m; j++) { 3715 procsnz[0] += locrowlens[j]; 3716 } 3717 for (i=1; i<size; i++) { 3718 mend = browners[i+1] - browners[i]; 3719 if (i == size-1) mend = mend - extra_rows; 3720 ierr = PetscBinaryRead(fd,rowlengths,mend,PETSC_INT);CHKERRQ(ierr); 3721 for (j=mend; j<browners[i+1] - browners[i]; j++) rowlengths[j] = 1; 3722 /* calculate the number of nonzeros on each processor */ 3723 for (j=0; j<browners[i+1]-browners[i]; j++) { 3724 procsnz[i] += rowlengths[j]; 3725 } 3726 ierr = MPI_Send(rowlengths,browners[i+1]-browners[i],MPIU_INT,i,tag,comm);CHKERRQ(ierr); 3727 } 3728 ierr = PetscFree(rowlengths);CHKERRQ(ierr); 3729 } else { 3730 ierr = MPI_Recv(locrowlens,m,MPIU_INT,0,tag,comm,&status);CHKERRQ(ierr); 3731 } 3732 3733 if (!rank) { 3734 /* determine max buffer needed and allocate it */ 3735 maxnz = procsnz[0]; 3736 for (i=1; i<size; i++) { 3737 maxnz = PetscMax(maxnz,procsnz[i]); 3738 } 3739 ierr = PetscMalloc(maxnz*sizeof(PetscInt),&cols);CHKERRQ(ierr); 3740 3741 /* read in my part of the matrix column indices */ 3742 nz = procsnz[0]; 3743 ierr = PetscMalloc((nz+1)*sizeof(PetscInt),&ibuf);CHKERRQ(ierr); 3744 mycols = ibuf; 3745 if (size == 1) nz -= extra_rows; 3746 ierr = PetscBinaryRead(fd,mycols,nz,PETSC_INT);CHKERRQ(ierr); 3747 if (size == 1) for (i=0; i< extra_rows; i++) { mycols[nz+i] = M+i; } 3748 3749 /* read in every ones (except the last) and ship off */ 3750 for (i=1; i<size-1; i++) { 3751 nz = procsnz[i]; 3752 ierr = PetscBinaryRead(fd,cols,nz,PETSC_INT);CHKERRQ(ierr); 3753 ierr = MPI_Send(cols,nz,MPIU_INT,i,tag,comm);CHKERRQ(ierr); 3754 } 3755 /* read in the stuff for the last proc */ 3756 if (size != 1) { 3757 nz = procsnz[size-1] - extra_rows; /* the extra rows are not on the disk */ 3758 ierr = PetscBinaryRead(fd,cols,nz,PETSC_INT);CHKERRQ(ierr); 3759 for (i=0; i<extra_rows; i++) cols[nz+i] = M+i; 3760 ierr = MPI_Send(cols,nz+extra_rows,MPIU_INT,size-1,tag,comm);CHKERRQ(ierr); 3761 } 3762 ierr = PetscFree(cols);CHKERRQ(ierr); 3763 } else { 3764 /* determine buffer space needed for message */ 3765 nz = 0; 3766 for (i=0; i<m; i++) { 3767 nz += locrowlens[i]; 3768 } 3769 ierr = PetscMalloc((nz+1)*sizeof(PetscInt),&ibuf);CHKERRQ(ierr); 3770 mycols = ibuf; 3771 /* receive message of column indices*/ 3772 ierr = MPI_Recv(mycols,nz,MPIU_INT,0,tag,comm,&status);CHKERRQ(ierr); 3773 ierr = MPI_Get_count(&status,MPIU_INT,&maxnz);CHKERRQ(ierr); 3774 if (maxnz != nz) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED,"something is wrong with file"); 3775 } 3776 3777 /* loop over local rows, determining number of off diagonal entries */ 3778 ierr = PetscMalloc2(rend-rstart,PetscInt,&dlens,rend-rstart,PetscInt,&odlens);CHKERRQ(ierr); 3779 ierr = PetscMalloc3(Mbs,PetscInt,&mask,Mbs,PetscInt,&masked1,Mbs,PetscInt,&masked2);CHKERRQ(ierr); 3780 ierr = PetscMemzero(mask,Mbs*sizeof(PetscInt));CHKERRQ(ierr); 3781 ierr = PetscMemzero(masked1,Mbs*sizeof(PetscInt));CHKERRQ(ierr); 3782 ierr = PetscMemzero(masked2,Mbs*sizeof(PetscInt));CHKERRQ(ierr); 3783 rowcount = 0; nzcount = 0; 3784 for (i=0; i<mbs; i++) { 3785 dcount = 0; 3786 odcount = 0; 3787 for (j=0; j<bs; j++) { 3788 kmax = locrowlens[rowcount]; 3789 for (k=0; k<kmax; k++) { 3790 tmp = mycols[nzcount++]/bs; 3791 if (!mask[tmp]) { 3792 mask[tmp] = 1; 3793 if (tmp < rstart || tmp >= rend) masked2[odcount++] = tmp; 3794 else masked1[dcount++] = tmp; 3795 } 3796 } 3797 rowcount++; 3798 } 3799 3800 dlens[i] = dcount; 3801 odlens[i] = odcount; 3802 3803 /* zero out the mask elements we set */ 3804 for (j=0; j<dcount; j++) mask[masked1[j]] = 0; 3805 for (j=0; j<odcount; j++) mask[masked2[j]] = 0; 3806 } 3807 3808 3809 if (!sizesset) { 3810 ierr = MatSetSizes(newmat,m,m,M+extra_rows,N+extra_rows);CHKERRQ(ierr); 3811 } 3812 ierr = MatMPIBAIJSetPreallocation(newmat,bs,0,dlens,0,odlens);CHKERRQ(ierr); 3813 3814 if (!rank) { 3815 ierr = PetscMalloc((maxnz+1)*sizeof(PetscScalar),&buf);CHKERRQ(ierr); 3816 /* read in my part of the matrix numerical values */ 3817 nz = procsnz[0]; 3818 vals = buf; 3819 mycols = ibuf; 3820 if (size == 1) nz -= extra_rows; 3821 ierr = PetscBinaryRead(fd,vals,nz,PETSC_SCALAR);CHKERRQ(ierr); 3822 if (size == 1) for (i=0; i< extra_rows; i++) { vals[nz+i] = 1.0; } 3823 3824 /* insert into matrix */ 3825 jj = rstart*bs; 3826 for (i=0; i<m; i++) { 3827 ierr = MatSetValues_MPIBAIJ(newmat,1,&jj,locrowlens[i],mycols,vals,INSERT_VALUES);CHKERRQ(ierr); 3828 mycols += locrowlens[i]; 3829 vals += locrowlens[i]; 3830 jj++; 3831 } 3832 /* read in other processors (except the last one) and ship out */ 3833 for (i=1; i<size-1; i++) { 3834 nz = procsnz[i]; 3835 vals = buf; 3836 ierr = PetscBinaryRead(fd,vals,nz,PETSC_SCALAR);CHKERRQ(ierr); 3837 ierr = MPI_Send(vals,nz,MPIU_SCALAR,i,((PetscObject)newmat)->tag,comm);CHKERRQ(ierr); 3838 } 3839 /* the last proc */ 3840 if (size != 1){ 3841 nz = procsnz[i] - extra_rows; 3842 vals = buf; 3843 ierr = PetscBinaryRead(fd,vals,nz,PETSC_SCALAR);CHKERRQ(ierr); 3844 for (i=0; i<extra_rows; i++) vals[nz+i] = 1.0; 3845 ierr = MPI_Send(vals,nz+extra_rows,MPIU_SCALAR,size-1,((PetscObject)newmat)->tag,comm);CHKERRQ(ierr); 3846 } 3847 ierr = PetscFree(procsnz);CHKERRQ(ierr); 3848 } else { 3849 /* receive numeric values */ 3850 ierr = PetscMalloc((nz+1)*sizeof(PetscScalar),&buf);CHKERRQ(ierr); 3851 3852 /* receive message of values*/ 3853 vals = buf; 3854 mycols = ibuf; 3855 ierr = MPI_Recv(vals,nz,MPIU_SCALAR,0,((PetscObject)newmat)->tag,comm,&status);CHKERRQ(ierr); 3856 ierr = MPI_Get_count(&status,MPIU_SCALAR,&maxnz);CHKERRQ(ierr); 3857 if (maxnz != nz) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED,"something is wrong with file"); 3858 3859 /* insert into matrix */ 3860 jj = rstart*bs; 3861 for (i=0; i<m; i++) { 3862 ierr = MatSetValues_MPIBAIJ(newmat,1,&jj,locrowlens[i],mycols,vals,INSERT_VALUES);CHKERRQ(ierr); 3863 mycols += locrowlens[i]; 3864 vals += locrowlens[i]; 3865 jj++; 3866 } 3867 } 3868 ierr = PetscFree(locrowlens);CHKERRQ(ierr); 3869 ierr = PetscFree(buf);CHKERRQ(ierr); 3870 ierr = PetscFree(ibuf);CHKERRQ(ierr); 3871 ierr = PetscFree2(rowners,browners);CHKERRQ(ierr); 3872 ierr = PetscFree2(dlens,odlens);CHKERRQ(ierr); 3873 ierr = PetscFree3(mask,masked1,masked2);CHKERRQ(ierr); 3874 ierr = MatAssemblyBegin(newmat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3875 ierr = MatAssemblyEnd(newmat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3876 3877 PetscFunctionReturn(0); 3878 } 3879 3880 #undef __FUNCT__ 3881 #define __FUNCT__ "MatMPIBAIJSetHashTableFactor" 3882 /*@ 3883 MatMPIBAIJSetHashTableFactor - Sets the factor required to compute the size of the HashTable. 3884 3885 Input Parameters: 3886 . mat - the matrix 3887 . fact - factor 3888 3889 Not Collective, each process can use a different factor 3890 3891 Level: advanced 3892 3893 Notes: 3894 This can also be set by the command line option: -mat_use_hash_table <fact> 3895 3896 .keywords: matrix, hashtable, factor, HT 3897 3898 .seealso: MatSetOption() 3899 @*/ 3900 PetscErrorCode MatMPIBAIJSetHashTableFactor(Mat mat,PetscReal fact) 3901 { 3902 PetscErrorCode ierr; 3903 3904 PetscFunctionBegin; 3905 ierr = PetscTryMethod(mat,"MatSetHashTableFactor_C",(Mat,PetscReal),(mat,fact));CHKERRQ(ierr); 3906 PetscFunctionReturn(0); 3907 } 3908 3909 EXTERN_C_BEGIN 3910 #undef __FUNCT__ 3911 #define __FUNCT__ "MatSetHashTableFactor_MPIBAIJ" 3912 PetscErrorCode MatSetHashTableFactor_MPIBAIJ(Mat mat,PetscReal fact) 3913 { 3914 Mat_MPIBAIJ *baij; 3915 3916 PetscFunctionBegin; 3917 baij = (Mat_MPIBAIJ*)mat->data; 3918 baij->ht_fact = fact; 3919 PetscFunctionReturn(0); 3920 } 3921 EXTERN_C_END 3922 3923 #undef __FUNCT__ 3924 #define __FUNCT__ "MatMPIBAIJGetSeqBAIJ" 3925 PetscErrorCode MatMPIBAIJGetSeqBAIJ(Mat A,Mat *Ad,Mat *Ao,const PetscInt *colmap[]) 3926 { 3927 Mat_MPIBAIJ *a = (Mat_MPIBAIJ *)A->data; 3928 PetscFunctionBegin; 3929 *Ad = a->A; 3930 *Ao = a->B; 3931 *colmap = a->garray; 3932 PetscFunctionReturn(0); 3933 } 3934 3935 /* 3936 Special version for direct calls from Fortran (to eliminate two function call overheads 3937 */ 3938 #if defined(PETSC_HAVE_FORTRAN_CAPS) 3939 #define matmpibaijsetvaluesblocked_ MATMPIBAIJSETVALUESBLOCKED 3940 #elif !defined(PETSC_HAVE_FORTRAN_UNDERSCORE) 3941 #define matmpibaijsetvaluesblocked_ matmpibaijsetvaluesblocked 3942 #endif 3943 3944 #undef __FUNCT__ 3945 #define __FUNCT__ "matmpibiajsetvaluesblocked" 3946 /*@C 3947 MatMPIBAIJSetValuesBlocked - Direct Fortran call to replace call to MatSetValuesBlocked() 3948 3949 Collective on Mat 3950 3951 Input Parameters: 3952 + mat - the matrix 3953 . min - number of input rows 3954 . im - input rows 3955 . nin - number of input columns 3956 . in - input columns 3957 . v - numerical values input 3958 - addvin - INSERT_VALUES or ADD_VALUES 3959 3960 Notes: This has a complete copy of MatSetValuesBlocked_MPIBAIJ() which is terrible code un-reuse. 3961 3962 Level: advanced 3963 3964 .seealso: MatSetValuesBlocked() 3965 @*/ 3966 PetscErrorCode matmpibaijsetvaluesblocked_(Mat *matin,PetscInt *min,const PetscInt im[],PetscInt *nin,const PetscInt in[],const MatScalar v[],InsertMode *addvin) 3967 { 3968 /* convert input arguments to C version */ 3969 Mat mat = *matin; 3970 PetscInt m = *min, n = *nin; 3971 InsertMode addv = *addvin; 3972 3973 Mat_MPIBAIJ *baij = (Mat_MPIBAIJ*)mat->data; 3974 const MatScalar *value; 3975 MatScalar *barray=baij->barray; 3976 PetscBool roworiented = baij->roworiented; 3977 PetscErrorCode ierr; 3978 PetscInt i,j,ii,jj,row,col,rstart=baij->rstartbs; 3979 PetscInt rend=baij->rendbs,cstart=baij->cstartbs,stepval; 3980 PetscInt cend=baij->cendbs,bs=mat->rmap->bs,bs2=baij->bs2; 3981 3982 PetscFunctionBegin; 3983 /* tasks normally handled by MatSetValuesBlocked() */ 3984 if (mat->insertmode == NOT_SET_VALUES) { 3985 mat->insertmode = addv; 3986 } 3987 #if defined(PETSC_USE_DEBUG) 3988 else if (mat->insertmode != addv) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Cannot mix add values and insert values"); 3989 if (mat->factortype) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Not for factored matrix"); 3990 #endif 3991 if (mat->assembled) { 3992 mat->was_assembled = PETSC_TRUE; 3993 mat->assembled = PETSC_FALSE; 3994 } 3995 ierr = PetscLogEventBegin(MAT_SetValues,mat,0,0,0);CHKERRQ(ierr); 3996 3997 3998 if (!barray) { 3999 ierr = PetscMalloc(bs2*sizeof(MatScalar),&barray);CHKERRQ(ierr); 4000 baij->barray = barray; 4001 } 4002 4003 if (roworiented) { 4004 stepval = (n-1)*bs; 4005 } else { 4006 stepval = (m-1)*bs; 4007 } 4008 for (i=0; i<m; i++) { 4009 if (im[i] < 0) continue; 4010 #if defined(PETSC_USE_DEBUG) 4011 if (im[i] >= baij->Mbs) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Row too large, row %D max %D",im[i],baij->Mbs-1); 4012 #endif 4013 if (im[i] >= rstart && im[i] < rend) { 4014 row = im[i] - rstart; 4015 for (j=0; j<n; j++) { 4016 /* If NumCol = 1 then a copy is not required */ 4017 if ((roworiented) && (n == 1)) { 4018 barray = (MatScalar*)v + i*bs2; 4019 } else if ((!roworiented) && (m == 1)) { 4020 barray = (MatScalar*)v + j*bs2; 4021 } else { /* Here a copy is required */ 4022 if (roworiented) { 4023 value = v + i*(stepval+bs)*bs + j*bs; 4024 } else { 4025 value = v + j*(stepval+bs)*bs + i*bs; 4026 } 4027 for (ii=0; ii<bs; ii++,value+=stepval) { 4028 for (jj=0; jj<bs; jj++) { 4029 *barray++ = *value++; 4030 } 4031 } 4032 barray -=bs2; 4033 } 4034 4035 if (in[j] >= cstart && in[j] < cend){ 4036 col = in[j] - cstart; 4037 ierr = MatSetValuesBlocked_SeqBAIJ(baij->A,1,&row,1,&col,barray,addv);CHKERRQ(ierr); 4038 } 4039 else if (in[j] < 0) continue; 4040 #if defined(PETSC_USE_DEBUG) 4041 else if (in[j] >= baij->Nbs) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Column too large, col %D max %D",in[j],baij->Nbs-1); 4042 #endif 4043 else { 4044 if (mat->was_assembled) { 4045 if (!baij->colmap) { 4046 ierr = MatCreateColmap_MPIBAIJ_Private(mat);CHKERRQ(ierr); 4047 } 4048 4049 #if defined(PETSC_USE_DEBUG) 4050 #if defined (PETSC_USE_CTABLE) 4051 { PetscInt data; 4052 ierr = PetscTableFind(baij->colmap,in[j]+1,&data);CHKERRQ(ierr); 4053 if ((data - 1) % bs) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Incorrect colmap"); 4054 } 4055 #else 4056 if ((baij->colmap[in[j]] - 1) % bs) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Incorrect colmap"); 4057 #endif 4058 #endif 4059 #if defined (PETSC_USE_CTABLE) 4060 ierr = PetscTableFind(baij->colmap,in[j]+1,&col);CHKERRQ(ierr); 4061 col = (col - 1)/bs; 4062 #else 4063 col = (baij->colmap[in[j]] - 1)/bs; 4064 #endif 4065 if (col < 0 && !((Mat_SeqBAIJ*)(baij->A->data))->nonew) { 4066 ierr = MatDisAssemble_MPIBAIJ(mat);CHKERRQ(ierr); 4067 col = in[j]; 4068 } 4069 } 4070 else col = in[j]; 4071 ierr = MatSetValuesBlocked_SeqBAIJ(baij->B,1,&row,1,&col,barray,addv);CHKERRQ(ierr); 4072 } 4073 } 4074 } else { 4075 if (!baij->donotstash) { 4076 if (roworiented) { 4077 ierr = MatStashValuesRowBlocked_Private(&mat->bstash,im[i],n,in,v,m,n,i);CHKERRQ(ierr); 4078 } else { 4079 ierr = MatStashValuesColBlocked_Private(&mat->bstash,im[i],n,in,v,m,n,i);CHKERRQ(ierr); 4080 } 4081 } 4082 } 4083 } 4084 4085 /* task normally handled by MatSetValuesBlocked() */ 4086 ierr = PetscLogEventEnd(MAT_SetValues,mat,0,0,0);CHKERRQ(ierr); 4087 PetscFunctionReturn(0); 4088 } 4089 4090 #undef __FUNCT__ 4091 #define __FUNCT__ "MatCreateMPIBAIJWithArrays" 4092 /*@ 4093 MatCreateMPIBAIJWithArrays - creates a MPI BAIJ matrix using arrays that contain in standard 4094 CSR format the local rows. 4095 4096 Collective on MPI_Comm 4097 4098 Input Parameters: 4099 + comm - MPI communicator 4100 . bs - the block size, only a block size of 1 is supported 4101 . m - number of local rows (Cannot be PETSC_DECIDE) 4102 . n - This value should be the same as the local size used in creating the 4103 x vector for the matrix-vector product y = Ax. (or PETSC_DECIDE to have 4104 calculated if N is given) For square matrices n is almost always m. 4105 . M - number of global rows (or PETSC_DETERMINE to have calculated if m is given) 4106 . N - number of global columns (or PETSC_DETERMINE to have calculated if n is given) 4107 . i - row indices 4108 . j - column indices 4109 - a - matrix values 4110 4111 Output Parameter: 4112 . mat - the matrix 4113 4114 Level: intermediate 4115 4116 Notes: 4117 The i, j, and a arrays ARE copied by this routine into the internal format used by PETSc; 4118 thus you CANNOT change the matrix entries by changing the values of a[] after you have 4119 called this routine. Use MatCreateMPIAIJWithSplitArrays() to avoid needing to copy the arrays. 4120 4121 The i and j indices are 0 based, and i indices are indices corresponding to the local j array. 4122 4123 .keywords: matrix, aij, compressed row, sparse, parallel 4124 4125 .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatMPIAIJSetPreallocation(), MatMPIAIJSetPreallocationCSR(), 4126 MPIAIJ, MatCreateAIJ(), MatCreateMPIAIJWithSplitArrays() 4127 @*/ 4128 PetscErrorCode MatCreateMPIBAIJWithArrays(MPI_Comm comm,PetscInt bs,PetscInt m,PetscInt n,PetscInt M,PetscInt N,const PetscInt i[],const PetscInt j[],const PetscScalar a[],Mat *mat) 4129 { 4130 PetscErrorCode ierr; 4131 4132 4133 PetscFunctionBegin; 4134 if (i[0]) { 4135 SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"i (row indices) must start with 0"); 4136 } 4137 if (m < 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"local number of rows (m) cannot be PETSC_DECIDE, or negative"); 4138 ierr = MatCreate(comm,mat);CHKERRQ(ierr); 4139 ierr = MatSetSizes(*mat,m,n,M,N);CHKERRQ(ierr); 4140 ierr = MatSetType(*mat,MATMPISBAIJ);CHKERRQ(ierr); 4141 ierr = MatMPIBAIJSetPreallocationCSR(*mat,bs,i,j,a);CHKERRQ(ierr); 4142 PetscFunctionReturn(0); 4143 } 4144