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++) { 29 if (PetscAbsScalar(va[i])) idx[i] += A->cmap->rstart; 30 } 31 } 32 33 ierr = VecCreateSeq(PETSC_COMM_SELF,A->rmap->n,&vtmp);CHKERRQ(ierr); 34 if (idx) {ierr = PetscMalloc(A->rmap->n*sizeof(PetscInt),&idxb);CHKERRQ(ierr);} 35 ierr = MatGetRowMaxAbs(a->B,vtmp,idxb);CHKERRQ(ierr); 36 ierr = VecGetArray(vtmp,&vb);CHKERRQ(ierr); 37 38 for (i=0; i<A->rmap->n; i++) { 39 if (PetscAbsScalar(va[i]) < PetscAbsScalar(vb[i])) { 40 va[i] = vb[i]; 41 if (idx) idx[i] = A->cmap->bs*a->garray[idxb[i]/A->cmap->bs] + (idxb[i] % A->cmap->bs); 42 } 43 } 44 45 ierr = VecRestoreArray(v,&va);CHKERRQ(ierr); 46 ierr = VecRestoreArray(vtmp,&vb);CHKERRQ(ierr); 47 ierr = PetscFree(idxb);CHKERRQ(ierr); 48 ierr = VecDestroy(&vtmp);CHKERRQ(ierr); 49 PetscFunctionReturn(0); 50 } 51 52 #undef __FUNCT__ 53 #define __FUNCT__ "MatStoreValues_MPIBAIJ" 54 PetscErrorCode MatStoreValues_MPIBAIJ(Mat mat) 55 { 56 Mat_MPIBAIJ *aij = (Mat_MPIBAIJ*)mat->data; 57 PetscErrorCode ierr; 58 59 PetscFunctionBegin; 60 ierr = MatStoreValues(aij->A);CHKERRQ(ierr); 61 ierr = MatStoreValues(aij->B);CHKERRQ(ierr); 62 PetscFunctionReturn(0); 63 } 64 65 #undef __FUNCT__ 66 #define __FUNCT__ "MatRetrieveValues_MPIBAIJ" 67 PetscErrorCode MatRetrieveValues_MPIBAIJ(Mat mat) 68 { 69 Mat_MPIBAIJ *aij = (Mat_MPIBAIJ*)mat->data; 70 PetscErrorCode ierr; 71 72 PetscFunctionBegin; 73 ierr = MatRetrieveValues(aij->A);CHKERRQ(ierr); 74 ierr = MatRetrieveValues(aij->B);CHKERRQ(ierr); 75 PetscFunctionReturn(0); 76 } 77 78 /* 79 Local utility routine that creates a mapping from the global column 80 number to the local number in the off-diagonal part of the local 81 storage of the matrix. This is done in a non scalable way since the 82 length of colmap equals the global matrix length. 83 */ 84 #undef __FUNCT__ 85 #define __FUNCT__ "MatCreateColmap_MPIBAIJ_Private" 86 PetscErrorCode MatCreateColmap_MPIBAIJ_Private(Mat mat) 87 { 88 Mat_MPIBAIJ *baij = (Mat_MPIBAIJ*)mat->data; 89 Mat_SeqBAIJ *B = (Mat_SeqBAIJ*)baij->B->data; 90 PetscErrorCode ierr; 91 PetscInt nbs = B->nbs,i,bs=mat->rmap->bs; 92 93 PetscFunctionBegin; 94 #if defined(PETSC_USE_CTABLE) 95 ierr = PetscTableCreate(baij->nbs,baij->Nbs+1,&baij->colmap);CHKERRQ(ierr); 96 for (i=0; i<nbs; i++) { 97 ierr = PetscTableAdd(baij->colmap,baij->garray[i]+1,i*bs+1,INSERT_VALUES);CHKERRQ(ierr); 98 } 99 #else 100 ierr = PetscMalloc((baij->Nbs+1)*sizeof(PetscInt),&baij->colmap);CHKERRQ(ierr); 101 ierr = PetscLogObjectMemory((PetscObject)mat,baij->Nbs*sizeof(PetscInt));CHKERRQ(ierr); 102 ierr = PetscMemzero(baij->colmap,baij->Nbs*sizeof(PetscInt));CHKERRQ(ierr); 103 for (i=0; i<nbs; i++) baij->colmap[baij->garray[i]] = i*bs+1; 104 #endif 105 PetscFunctionReturn(0); 106 } 107 108 #define MatSetValues_SeqBAIJ_A_Private(row,col,value,addv) \ 109 { \ 110 \ 111 brow = row/bs; \ 112 rp = aj + ai[brow]; ap = aa + bs2*ai[brow]; \ 113 rmax = aimax[brow]; nrow = ailen[brow]; \ 114 bcol = col/bs; \ 115 ridx = row % bs; cidx = col % bs; \ 116 low = 0; high = nrow; \ 117 while (high-low > 3) { \ 118 t = (low+high)/2; \ 119 if (rp[t] > bcol) high = t; \ 120 else low = t; \ 121 } \ 122 for (_i=low; _i<high; _i++) { \ 123 if (rp[_i] > bcol) break; \ 124 if (rp[_i] == bcol) { \ 125 bap = ap + bs2*_i + bs*cidx + ridx; \ 126 if (addv == ADD_VALUES) *bap += value; \ 127 else *bap = value; \ 128 goto a_noinsert; \ 129 } \ 130 } \ 131 if (a->nonew == 1) goto a_noinsert; \ 132 if (a->nonew == -1) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new nonzero (%D, %D) into matrix", row, col); \ 133 MatSeqXAIJReallocateAIJ(A,a->mbs,bs2,nrow,brow,bcol,rmax,aa,ai,aj,rp,ap,aimax,a->nonew,MatScalar); \ 134 N = nrow++ - 1; \ 135 /* shift up all the later entries in this row */ \ 136 for (ii=N; ii>=_i; ii--) { \ 137 rp[ii+1] = rp[ii]; \ 138 ierr = PetscMemcpy(ap+bs2*(ii+1),ap+bs2*(ii),bs2*sizeof(MatScalar));CHKERRQ(ierr); \ 139 } \ 140 if (N>=_i) { ierr = PetscMemzero(ap+bs2*_i,bs2*sizeof(MatScalar));CHKERRQ(ierr); } \ 141 rp[_i] = bcol; \ 142 ap[bs2*_i + bs*cidx + ridx] = value; \ 143 a_noinsert:; \ 144 ailen[brow] = nrow; \ 145 } 146 147 #define MatSetValues_SeqBAIJ_B_Private(row,col,value,addv) \ 148 { \ 149 brow = row/bs; \ 150 rp = bj + bi[brow]; ap = ba + bs2*bi[brow]; \ 151 rmax = bimax[brow]; nrow = bilen[brow]; \ 152 bcol = col/bs; \ 153 ridx = row % bs; cidx = col % bs; \ 154 low = 0; high = nrow; \ 155 while (high-low > 3) { \ 156 t = (low+high)/2; \ 157 if (rp[t] > bcol) high = t; \ 158 else low = t; \ 159 } \ 160 for (_i=low; _i<high; _i++) { \ 161 if (rp[_i] > bcol) break; \ 162 if (rp[_i] == bcol) { \ 163 bap = ap + bs2*_i + bs*cidx + ridx; \ 164 if (addv == ADD_VALUES) *bap += value; \ 165 else *bap = value; \ 166 goto b_noinsert; \ 167 } \ 168 } \ 169 if (b->nonew == 1) goto b_noinsert; \ 170 if (b->nonew == -1) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new nonzero (%D, %D) into matrix", row, col); \ 171 MatSeqXAIJReallocateAIJ(B,b->mbs,bs2,nrow,brow,bcol,rmax,ba,bi,bj,rp,ap,bimax,b->nonew,MatScalar); \ 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]; 226 else value = v[i+j*m]; 227 MatSetValues_SeqBAIJ_A_Private(row,col,value,addv); 228 /* ierr = MatSetValues_SeqBAIJ(baij->A,1,&row,1,&col,&value,addv);CHKERRQ(ierr); */ 229 } else if (in[j] < 0) continue; 230 #if defined(PETSC_USE_DEBUG) 231 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); 232 #endif 233 else { 234 if (mat->was_assembled) { 235 if (!baij->colmap) { 236 ierr = MatCreateColmap_MPIBAIJ_Private(mat);CHKERRQ(ierr); 237 } 238 #if defined(PETSC_USE_CTABLE) 239 ierr = PetscTableFind(baij->colmap,in[j]/bs + 1,&col);CHKERRQ(ierr); 240 col = col - 1; 241 #else 242 col = baij->colmap[in[j]/bs] - 1; 243 #endif 244 if (col < 0 && !((Mat_SeqBAIJ*)(baij->B->data))->nonew) { 245 ierr = MatDisAssemble_MPIBAIJ(mat);CHKERRQ(ierr); 246 col = in[j]; 247 /* Reinitialize the variables required by MatSetValues_SeqBAIJ_B_Private() */ 248 B = baij->B; 249 b = (Mat_SeqBAIJ*)(B)->data; 250 bimax=b->imax;bi=b->i;bilen=b->ilen;bj=b->j; 251 ba =b->a; 252 } else if (col < 0) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new nonzero (%D, %D) into matrix", im[i], in[j]); 253 else col += in[j]%bs; 254 } else col = in[j]; 255 if (roworiented) value = v[i*n+j]; 256 else value = v[i+j*m]; 257 MatSetValues_SeqBAIJ_B_Private(row,col,value,addv); 258 /* ierr = MatSetValues_SeqBAIJ(baij->B,1,&row,1,&col,&value,addv);CHKERRQ(ierr); */ 259 } 260 } 261 } else { 262 if (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]); 263 if (!baij->donotstash) { 264 mat->assembled = PETSC_FALSE; 265 if (roworiented) { 266 ierr = MatStashValuesRow_Private(&mat->stash,im[i],n,in,v+i*n,PETSC_FALSE);CHKERRQ(ierr); 267 } else { 268 ierr = MatStashValuesCol_Private(&mat->stash,im[i],n,in,v+i,m,PETSC_FALSE);CHKERRQ(ierr); 269 } 270 } 271 } 272 } 273 PetscFunctionReturn(0); 274 } 275 276 #undef __FUNCT__ 277 #define __FUNCT__ "MatSetValuesBlocked_MPIBAIJ" 278 PetscErrorCode MatSetValuesBlocked_MPIBAIJ(Mat mat,PetscInt m,const PetscInt im[],PetscInt n,const PetscInt in[],const PetscScalar v[],InsertMode addv) 279 { 280 Mat_MPIBAIJ *baij = (Mat_MPIBAIJ*)mat->data; 281 const PetscScalar *value; 282 MatScalar *barray = baij->barray; 283 PetscBool roworiented = baij->roworiented; 284 PetscErrorCode ierr; 285 PetscInt i,j,ii,jj,row,col,rstart=baij->rstartbs; 286 PetscInt rend=baij->rendbs,cstart=baij->cstartbs,stepval; 287 PetscInt cend=baij->cendbs,bs=mat->rmap->bs,bs2=baij->bs2; 288 289 PetscFunctionBegin; 290 if (!barray) { 291 ierr = PetscMalloc(bs2*sizeof(MatScalar),&barray);CHKERRQ(ierr); 292 baij->barray = barray; 293 } 294 295 if (roworiented) stepval = (n-1)*bs; 296 else 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++) barray[jj] = value[jj]; 319 barray += bs; 320 } 321 barray -= bs2; 322 } 323 324 if (in[j] >= cstart && in[j] < cend) { 325 col = in[j] - cstart; 326 ierr = MatSetValuesBlocked_SeqBAIJ(baij->A,1,&row,1,&col,barray,addv);CHKERRQ(ierr); 327 } else if (in[j] < 0) continue; 328 #if defined(PETSC_USE_DEBUG) 329 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); 330 #endif 331 else { 332 if (mat->was_assembled) { 333 if (!baij->colmap) { 334 ierr = MatCreateColmap_MPIBAIJ_Private(mat);CHKERRQ(ierr); 335 } 336 337 #if defined(PETSC_USE_DEBUG) 338 #if defined(PETSC_USE_CTABLE) 339 { PetscInt data; 340 ierr = PetscTableFind(baij->colmap,in[j]+1,&data);CHKERRQ(ierr); 341 if ((data - 1) % bs) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Incorrect colmap"); 342 } 343 #else 344 if ((baij->colmap[in[j]] - 1) % bs) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Incorrect colmap"); 345 #endif 346 #endif 347 #if defined(PETSC_USE_CTABLE) 348 ierr = PetscTableFind(baij->colmap,in[j]+1,&col);CHKERRQ(ierr); 349 col = (col - 1)/bs; 350 #else 351 col = (baij->colmap[in[j]] - 1)/bs; 352 #endif 353 if (col < 0 && !((Mat_SeqBAIJ*)(baij->B->data))->nonew) { 354 ierr = MatDisAssemble_MPIBAIJ(mat);CHKERRQ(ierr); 355 col = in[j]; 356 } 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]); 357 } else col = in[j]; 358 ierr = MatSetValuesBlocked_SeqBAIJ(baij->B,1,&row,1,&col,barray,addv);CHKERRQ(ierr); 359 } 360 } 361 } else { 362 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]); 363 if (!baij->donotstash) { 364 if (roworiented) { 365 ierr = MatStashValuesRowBlocked_Private(&mat->bstash,im[i],n,in,v,m,n,i);CHKERRQ(ierr); 366 } else { 367 ierr = MatStashValuesColBlocked_Private(&mat->bstash,im[i],n,in,v,m,n,i);CHKERRQ(ierr); 368 } 369 } 370 } 371 } 372 PetscFunctionReturn(0); 373 } 374 375 #define HASH_KEY 0.6180339887 376 #define HASH(size,key,tmp) (tmp = (key)*HASH_KEY,(PetscInt)((size)*(tmp-(PetscInt)tmp))) 377 /* #define HASH(size,key) ((PetscInt)((size)*fmod(((key)*HASH_KEY),1))) */ 378 /* #define HASH(size,key,tmp) ((PetscInt)((size)*fmod(((key)*HASH_KEY),1))) */ 379 #undef __FUNCT__ 380 #define __FUNCT__ "MatSetValues_MPIBAIJ_HT" 381 PetscErrorCode MatSetValues_MPIBAIJ_HT(Mat mat,PetscInt m,const PetscInt im[],PetscInt n,const PetscInt in[],const PetscScalar v[],InsertMode addv) 382 { 383 Mat_MPIBAIJ *baij = (Mat_MPIBAIJ*)mat->data; 384 PetscBool roworiented = baij->roworiented; 385 PetscErrorCode ierr; 386 PetscInt i,j,row,col; 387 PetscInt rstart_orig=mat->rmap->rstart; 388 PetscInt rend_orig =mat->rmap->rend,Nbs=baij->Nbs; 389 PetscInt h1,key,size=baij->ht_size,bs=mat->rmap->bs,*HT=baij->ht,idx; 390 PetscReal tmp; 391 MatScalar **HD = baij->hd,value; 392 #if defined(PETSC_USE_DEBUG) 393 PetscInt total_ct=baij->ht_total_ct,insert_ct=baij->ht_insert_ct; 394 #endif 395 396 PetscFunctionBegin; 397 if (v) PetscValidScalarPointer(v,6); 398 for (i=0; i<m; i++) { 399 #if defined(PETSC_USE_DEBUG) 400 if (im[i] < 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Negative row"); 401 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); 402 #endif 403 row = im[i]; 404 if (row >= rstart_orig && row < rend_orig) { 405 for (j=0; j<n; j++) { 406 col = in[j]; 407 if (roworiented) value = v[i*n+j]; 408 else value = v[i+j*m]; 409 /* Look up PetscInto the Hash Table */ 410 key = (row/bs)*Nbs+(col/bs)+1; 411 h1 = HASH(size,key,tmp); 412 413 414 idx = h1; 415 #if defined(PETSC_USE_DEBUG) 416 insert_ct++; 417 total_ct++; 418 if (HT[idx] != key) { 419 for (idx=h1; (idx<size) && (HT[idx]!=key); idx++,total_ct++) ; 420 if (idx == size) { 421 for (idx=0; (idx<h1) && (HT[idx]!=key); idx++,total_ct++) ; 422 if (idx == h1) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"(%D,%D) has no entry in the hash table", row, col); 423 } 424 } 425 #else 426 if (HT[idx] != key) { 427 for (idx=h1; (idx<size) && (HT[idx]!=key); idx++) ; 428 if (idx == size) { 429 for (idx=0; (idx<h1) && (HT[idx]!=key); idx++) ; 430 if (idx == h1) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"(%D,%D) has no entry in the hash table", row, col); 431 } 432 } 433 #endif 434 /* A HASH table entry is found, so insert the values at the correct address */ 435 if (addv == ADD_VALUES) *(HD[idx]+ (col % bs)*bs + (row % bs)) += value; 436 else *(HD[idx]+ (col % bs)*bs + (row % bs)) = value; 437 } 438 } else if (!baij->donotstash) { 439 if (roworiented) { 440 ierr = MatStashValuesRow_Private(&mat->stash,im[i],n,in,v+i*n,PETSC_FALSE);CHKERRQ(ierr); 441 } else { 442 ierr = MatStashValuesCol_Private(&mat->stash,im[i],n,in,v+i,m,PETSC_FALSE);CHKERRQ(ierr); 443 } 444 } 445 } 446 #if defined(PETSC_USE_DEBUG) 447 baij->ht_total_ct = total_ct; 448 baij->ht_insert_ct = insert_ct; 449 #endif 450 PetscFunctionReturn(0); 451 } 452 453 #undef __FUNCT__ 454 #define __FUNCT__ "MatSetValuesBlocked_MPIBAIJ_HT" 455 PetscErrorCode MatSetValuesBlocked_MPIBAIJ_HT(Mat mat,PetscInt m,const PetscInt im[],PetscInt n,const PetscInt in[],const PetscScalar v[],InsertMode addv) 456 { 457 Mat_MPIBAIJ *baij = (Mat_MPIBAIJ*)mat->data; 458 PetscBool roworiented = baij->roworiented; 459 PetscErrorCode ierr; 460 PetscInt i,j,ii,jj,row,col; 461 PetscInt rstart=baij->rstartbs; 462 PetscInt rend =mat->rmap->rend,stepval,bs=mat->rmap->bs,bs2=baij->bs2,nbs2=n*bs2; 463 PetscInt h1,key,size=baij->ht_size,idx,*HT=baij->ht,Nbs=baij->Nbs; 464 PetscReal tmp; 465 MatScalar **HD = baij->hd,*baij_a; 466 const PetscScalar *v_t,*value; 467 #if defined(PETSC_USE_DEBUG) 468 PetscInt total_ct=baij->ht_total_ct,insert_ct=baij->ht_insert_ct; 469 #endif 470 471 PetscFunctionBegin; 472 if (roworiented) stepval = (n-1)*bs; 473 else stepval = (m-1)*bs; 474 475 for (i=0; i<m; i++) { 476 #if defined(PETSC_USE_DEBUG) 477 if (im[i] < 0) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Negative row: %D",im[i]); 478 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); 479 #endif 480 row = im[i]; 481 v_t = v + i*nbs2; 482 if (row >= rstart && row < rend) { 483 for (j=0; j<n; j++) { 484 col = in[j]; 485 486 /* Look up into the Hash Table */ 487 key = row*Nbs+col+1; 488 h1 = HASH(size,key,tmp); 489 490 idx = h1; 491 #if defined(PETSC_USE_DEBUG) 492 total_ct++; 493 insert_ct++; 494 if (HT[idx] != key) { 495 for (idx=h1; (idx<size) && (HT[idx]!=key); idx++,total_ct++) ; 496 if (idx == size) { 497 for (idx=0; (idx<h1) && (HT[idx]!=key); idx++,total_ct++) ; 498 if (idx == h1) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"(%D,%D) has no entry in the hash table", row, col); 499 } 500 } 501 #else 502 if (HT[idx] != key) { 503 for (idx=h1; (idx<size) && (HT[idx]!=key); idx++) ; 504 if (idx == size) { 505 for (idx=0; (idx<h1) && (HT[idx]!=key); idx++) ; 506 if (idx == h1) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"(%D,%D) has no entry in the hash table", row, col); 507 } 508 } 509 #endif 510 baij_a = HD[idx]; 511 if (roworiented) { 512 /*value = v + i*(stepval+bs)*bs + j*bs;*/ 513 /* value = v + (i*(stepval+bs)+j)*bs; */ 514 value = v_t; 515 v_t += bs; 516 if (addv == ADD_VALUES) { 517 for (ii=0; ii<bs; ii++,value+=stepval) { 518 for (jj=ii; jj<bs2; jj+=bs) { 519 baij_a[jj] += *value++; 520 } 521 } 522 } else { 523 for (ii=0; ii<bs; ii++,value+=stepval) { 524 for (jj=ii; jj<bs2; jj+=bs) { 525 baij_a[jj] = *value++; 526 } 527 } 528 } 529 } else { 530 value = v + j*(stepval+bs)*bs + i*bs; 531 if (addv == ADD_VALUES) { 532 for (ii=0; ii<bs; ii++,value+=stepval,baij_a+=bs) { 533 for (jj=0; jj<bs; jj++) { 534 baij_a[jj] += *value++; 535 } 536 } 537 } else { 538 for (ii=0; ii<bs; ii++,value+=stepval,baij_a+=bs) { 539 for (jj=0; jj<bs; jj++) { 540 baij_a[jj] = *value++; 541 } 542 } 543 } 544 } 545 } 546 } else { 547 if (!baij->donotstash) { 548 if (roworiented) { 549 ierr = MatStashValuesRowBlocked_Private(&mat->bstash,im[i],n,in,v,m,n,i);CHKERRQ(ierr); 550 } else { 551 ierr = MatStashValuesColBlocked_Private(&mat->bstash,im[i],n,in,v,m,n,i);CHKERRQ(ierr); 552 } 553 } 554 } 555 } 556 #if defined(PETSC_USE_DEBUG) 557 baij->ht_total_ct = total_ct; 558 baij->ht_insert_ct = insert_ct; 559 #endif 560 PetscFunctionReturn(0); 561 } 562 563 #undef __FUNCT__ 564 #define __FUNCT__ "MatGetValues_MPIBAIJ" 565 PetscErrorCode MatGetValues_MPIBAIJ(Mat mat,PetscInt m,const PetscInt idxm[],PetscInt n,const PetscInt idxn[],PetscScalar v[]) 566 { 567 Mat_MPIBAIJ *baij = (Mat_MPIBAIJ*)mat->data; 568 PetscErrorCode ierr; 569 PetscInt bs = mat->rmap->bs,i,j,bsrstart = mat->rmap->rstart,bsrend = mat->rmap->rend; 570 PetscInt bscstart = mat->cmap->rstart,bscend = mat->cmap->rend,row,col,data; 571 572 PetscFunctionBegin; 573 for (i=0; i<m; i++) { 574 if (idxm[i] < 0) continue; /* SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Negative row: %D",idxm[i]);*/ 575 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); 576 if (idxm[i] >= bsrstart && idxm[i] < bsrend) { 577 row = idxm[i] - bsrstart; 578 for (j=0; j<n; j++) { 579 if (idxn[j] < 0) continue; /* SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Negative column: %D",idxn[j]); */ 580 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); 581 if (idxn[j] >= bscstart && idxn[j] < bscend) { 582 col = idxn[j] - bscstart; 583 ierr = MatGetValues_SeqBAIJ(baij->A,1,&row,1,&col,v+i*n+j);CHKERRQ(ierr); 584 } else { 585 if (!baij->colmap) { 586 ierr = MatCreateColmap_MPIBAIJ_Private(mat);CHKERRQ(ierr); 587 } 588 #if defined(PETSC_USE_CTABLE) 589 ierr = PetscTableFind(baij->colmap,idxn[j]/bs+1,&data);CHKERRQ(ierr); 590 data--; 591 #else 592 data = baij->colmap[idxn[j]/bs]-1; 593 #endif 594 if ((data < 0) || (baij->garray[data/bs] != idxn[j]/bs)) *(v+i*n+j) = 0.0; 595 else { 596 col = data + idxn[j]%bs; 597 ierr = MatGetValues_SeqBAIJ(baij->B,1,&row,1,&col,v+i*n+j);CHKERRQ(ierr); 598 } 599 } 600 } 601 } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Only local values currently supported"); 602 } 603 PetscFunctionReturn(0); 604 } 605 606 #undef __FUNCT__ 607 #define __FUNCT__ "MatNorm_MPIBAIJ" 608 PetscErrorCode MatNorm_MPIBAIJ(Mat mat,NormType type,PetscReal *nrm) 609 { 610 Mat_MPIBAIJ *baij = (Mat_MPIBAIJ*)mat->data; 611 Mat_SeqBAIJ *amat = (Mat_SeqBAIJ*)baij->A->data,*bmat = (Mat_SeqBAIJ*)baij->B->data; 612 PetscErrorCode ierr; 613 PetscInt i,j,bs2=baij->bs2,bs=baij->A->rmap->bs,nz,row,col; 614 PetscReal sum = 0.0; 615 MatScalar *v; 616 617 PetscFunctionBegin; 618 if (baij->size == 1) { 619 ierr = MatNorm(baij->A,type,nrm);CHKERRQ(ierr); 620 } else { 621 if (type == NORM_FROBENIUS) { 622 v = amat->a; 623 nz = amat->nz*bs2; 624 for (i=0; i<nz; i++) { 625 sum += PetscRealPart(PetscConj(*v)*(*v)); v++; 626 } 627 v = bmat->a; 628 nz = bmat->nz*bs2; 629 for (i=0; i<nz; i++) { 630 sum += PetscRealPart(PetscConj(*v)*(*v)); v++; 631 } 632 ierr = MPI_Allreduce(&sum,nrm,1,MPIU_REAL,MPIU_SUM,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 633 *nrm = PetscSqrtReal(*nrm); 634 } else if (type == NORM_1) { /* max column sum */ 635 PetscReal *tmp,*tmp2; 636 PetscInt *jj,*garray=baij->garray,cstart=baij->rstartbs; 637 ierr = PetscMalloc2(mat->cmap->N,PetscReal,&tmp,mat->cmap->N,PetscReal,&tmp2);CHKERRQ(ierr); 638 ierr = PetscMemzero(tmp,mat->cmap->N*sizeof(PetscReal));CHKERRQ(ierr); 639 v = amat->a; jj = amat->j; 640 for (i=0; i<amat->nz; i++) { 641 for (j=0; j<bs; j++) { 642 col = bs*(cstart + *jj) + j; /* column index */ 643 for (row=0; row<bs; row++) { 644 tmp[col] += PetscAbsScalar(*v); v++; 645 } 646 } 647 jj++; 648 } 649 v = bmat->a; jj = bmat->j; 650 for (i=0; i<bmat->nz; i++) { 651 for (j=0; j<bs; j++) { 652 col = bs*garray[*jj] + j; 653 for (row=0; row<bs; row++) { 654 tmp[col] += PetscAbsScalar(*v); v++; 655 } 656 } 657 jj++; 658 } 659 ierr = MPI_Allreduce(tmp,tmp2,mat->cmap->N,MPIU_REAL,MPIU_SUM,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 660 *nrm = 0.0; 661 for (j=0; j<mat->cmap->N; j++) { 662 if (tmp2[j] > *nrm) *nrm = tmp2[j]; 663 } 664 ierr = PetscFree2(tmp,tmp2);CHKERRQ(ierr); 665 } else if (type == NORM_INFINITY) { /* max row sum */ 666 PetscReal *sums; 667 ierr = PetscMalloc(bs*sizeof(PetscReal),&sums);CHKERRQ(ierr); 668 sum = 0.0; 669 for (j=0; j<amat->mbs; j++) { 670 for (row=0; row<bs; row++) sums[row] = 0.0; 671 v = amat->a + bs2*amat->i[j]; 672 nz = amat->i[j+1]-amat->i[j]; 673 for (i=0; i<nz; i++) { 674 for (col=0; col<bs; col++) { 675 for (row=0; row<bs; row++) { 676 sums[row] += PetscAbsScalar(*v); v++; 677 } 678 } 679 } 680 v = bmat->a + bs2*bmat->i[j]; 681 nz = bmat->i[j+1]-bmat->i[j]; 682 for (i=0; i<nz; i++) { 683 for (col=0; col<bs; col++) { 684 for (row=0; row<bs; row++) { 685 sums[row] += PetscAbsScalar(*v); v++; 686 } 687 } 688 } 689 for (row=0; row<bs; row++) { 690 if (sums[row] > sum) sum = sums[row]; 691 } 692 } 693 ierr = MPI_Allreduce(&sum,nrm,1,MPIU_REAL,MPIU_MAX,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 694 ierr = PetscFree(sums);CHKERRQ(ierr); 695 } else SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"No support for this norm yet"); 696 } 697 PetscFunctionReturn(0); 698 } 699 700 /* 701 Creates the hash table, and sets the table 702 This table is created only once. 703 If new entried need to be added to the matrix 704 then the hash table has to be destroyed and 705 recreated. 706 */ 707 #undef __FUNCT__ 708 #define __FUNCT__ "MatCreateHashTable_MPIBAIJ_Private" 709 PetscErrorCode MatCreateHashTable_MPIBAIJ_Private(Mat mat,PetscReal factor) 710 { 711 Mat_MPIBAIJ *baij = (Mat_MPIBAIJ*)mat->data; 712 Mat A = baij->A,B=baij->B; 713 Mat_SeqBAIJ *a = (Mat_SeqBAIJ*)A->data,*b=(Mat_SeqBAIJ*)B->data; 714 PetscInt i,j,k,nz=a->nz+b->nz,h1,*ai=a->i,*aj=a->j,*bi=b->i,*bj=b->j; 715 PetscErrorCode ierr; 716 PetscInt ht_size,bs2=baij->bs2,rstart=baij->rstartbs; 717 PetscInt cstart=baij->cstartbs,*garray=baij->garray,row,col,Nbs=baij->Nbs; 718 PetscInt *HT,key; 719 MatScalar **HD; 720 PetscReal tmp; 721 #if defined(PETSC_USE_INFO) 722 PetscInt ct=0,max=0; 723 #endif 724 725 PetscFunctionBegin; 726 if (baij->ht) PetscFunctionReturn(0); 727 728 baij->ht_size = (PetscInt)(factor*nz); 729 ht_size = baij->ht_size; 730 731 /* Allocate Memory for Hash Table */ 732 ierr = PetscMalloc2(ht_size,MatScalar*,&baij->hd,ht_size,PetscInt,&baij->ht);CHKERRQ(ierr); 733 ierr = PetscMemzero(baij->hd,ht_size*sizeof(MatScalar*));CHKERRQ(ierr); 734 ierr = PetscMemzero(baij->ht,ht_size*sizeof(PetscInt));CHKERRQ(ierr); 735 HD = baij->hd; 736 HT = baij->ht; 737 738 /* Loop Over A */ 739 for (i=0; i<a->mbs; i++) { 740 for (j=ai[i]; j<ai[i+1]; j++) { 741 row = i+rstart; 742 col = aj[j]+cstart; 743 744 key = row*Nbs + col + 1; 745 h1 = HASH(ht_size,key,tmp); 746 for (k=0; k<ht_size; k++) { 747 if (!HT[(h1+k)%ht_size]) { 748 HT[(h1+k)%ht_size] = key; 749 HD[(h1+k)%ht_size] = a->a + j*bs2; 750 break; 751 #if defined(PETSC_USE_INFO) 752 } else { 753 ct++; 754 #endif 755 } 756 } 757 #if defined(PETSC_USE_INFO) 758 if (k> max) max = k; 759 #endif 760 } 761 } 762 /* Loop Over B */ 763 for (i=0; i<b->mbs; i++) { 764 for (j=bi[i]; j<bi[i+1]; j++) { 765 row = i+rstart; 766 col = garray[bj[j]]; 767 key = row*Nbs + col + 1; 768 h1 = HASH(ht_size,key,tmp); 769 for (k=0; k<ht_size; k++) { 770 if (!HT[(h1+k)%ht_size]) { 771 HT[(h1+k)%ht_size] = key; 772 HD[(h1+k)%ht_size] = b->a + j*bs2; 773 break; 774 #if defined(PETSC_USE_INFO) 775 } else { 776 ct++; 777 #endif 778 } 779 } 780 #if defined(PETSC_USE_INFO) 781 if (k> max) max = k; 782 #endif 783 } 784 } 785 786 /* Print Summary */ 787 #if defined(PETSC_USE_INFO) 788 for (i=0,j=0; i<ht_size; i++) { 789 if (HT[i]) j++; 790 } 791 ierr = PetscInfo2(mat,"Average Search = %5.2f,max search = %D\n",(!j)? 0.0:((PetscReal)(ct+j))/j,max);CHKERRQ(ierr); 792 #endif 793 PetscFunctionReturn(0); 794 } 795 796 #undef __FUNCT__ 797 #define __FUNCT__ "MatAssemblyBegin_MPIBAIJ" 798 PetscErrorCode MatAssemblyBegin_MPIBAIJ(Mat mat,MatAssemblyType mode) 799 { 800 Mat_MPIBAIJ *baij = (Mat_MPIBAIJ*)mat->data; 801 PetscErrorCode ierr; 802 PetscInt nstash,reallocs; 803 InsertMode addv; 804 805 PetscFunctionBegin; 806 if (baij->donotstash || mat->nooffprocentries) PetscFunctionReturn(0); 807 808 /* make sure all processors are either in INSERTMODE or ADDMODE */ 809 ierr = MPI_Allreduce((PetscEnum*)&mat->insertmode,(PetscEnum*)&addv,1,MPIU_ENUM,MPI_BOR,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 810 if (addv == (ADD_VALUES|INSERT_VALUES)) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_ARG_WRONGSTATE,"Some processors inserted others added"); 811 mat->insertmode = addv; /* in case this processor had no cache */ 812 813 ierr = MatStashScatterBegin_Private(mat,&mat->stash,mat->rmap->range);CHKERRQ(ierr); 814 ierr = MatStashScatterBegin_Private(mat,&mat->bstash,baij->rangebs);CHKERRQ(ierr); 815 ierr = MatStashGetInfo_Private(&mat->stash,&nstash,&reallocs);CHKERRQ(ierr); 816 ierr = PetscInfo2(mat,"Stash has %D entries,uses %D mallocs.\n",nstash,reallocs);CHKERRQ(ierr); 817 ierr = MatStashGetInfo_Private(&mat->bstash,&nstash,&reallocs);CHKERRQ(ierr); 818 ierr = PetscInfo2(mat,"Block-Stash has %D entries, uses %D mallocs.\n",nstash,reallocs);CHKERRQ(ierr); 819 PetscFunctionReturn(0); 820 } 821 822 #undef __FUNCT__ 823 #define __FUNCT__ "MatAssemblyEnd_MPIBAIJ" 824 PetscErrorCode MatAssemblyEnd_MPIBAIJ(Mat mat,MatAssemblyType mode) 825 { 826 Mat_MPIBAIJ *baij=(Mat_MPIBAIJ*)mat->data; 827 Mat_SeqBAIJ *a =(Mat_SeqBAIJ*)baij->A->data; 828 PetscErrorCode ierr; 829 PetscInt i,j,rstart,ncols,flg,bs2=baij->bs2; 830 PetscInt *row,*col; 831 PetscBool r1,r2,r3,other_disassembled; 832 MatScalar *val; 833 InsertMode addv = mat->insertmode; 834 PetscMPIInt n; 835 836 PetscFunctionBegin; 837 /* do not use 'b=(Mat_SeqBAIJ*)baij->B->data' as B can be reset in disassembly */ 838 if (!baij->donotstash && !mat->nooffprocentries) { 839 while (1) { 840 ierr = MatStashScatterGetMesg_Private(&mat->stash,&n,&row,&col,&val,&flg);CHKERRQ(ierr); 841 if (!flg) break; 842 843 for (i=0; i<n;) { 844 /* Now identify the consecutive vals belonging to the same row */ 845 for (j=i,rstart=row[j]; j<n; j++) { 846 if (row[j] != rstart) break; 847 } 848 if (j < n) ncols = j-i; 849 else ncols = n-i; 850 /* Now assemble all these values with a single function call */ 851 ierr = MatSetValues_MPIBAIJ(mat,1,row+i,ncols,col+i,val+i,addv);CHKERRQ(ierr); 852 i = j; 853 } 854 } 855 ierr = MatStashScatterEnd_Private(&mat->stash);CHKERRQ(ierr); 856 /* Now process the block-stash. Since the values are stashed column-oriented, 857 set the roworiented flag to column oriented, and after MatSetValues() 858 restore the original flags */ 859 r1 = baij->roworiented; 860 r2 = a->roworiented; 861 r3 = ((Mat_SeqBAIJ*)baij->B->data)->roworiented; 862 863 baij->roworiented = PETSC_FALSE; 864 a->roworiented = PETSC_FALSE; 865 866 (((Mat_SeqBAIJ*)baij->B->data))->roworiented = PETSC_FALSE; /* b->roworiented */ 867 while (1) { 868 ierr = MatStashScatterGetMesg_Private(&mat->bstash,&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++) { 874 if (row[j] != rstart) break; 875 } 876 if (j < n) ncols = j-i; 877 else ncols = n-i; 878 ierr = MatSetValuesBlocked_MPIBAIJ(mat,1,row+i,ncols,col+i,val+i*bs2,addv);CHKERRQ(ierr); 879 i = j; 880 } 881 } 882 ierr = MatStashScatterEnd_Private(&mat->bstash);CHKERRQ(ierr); 883 884 baij->roworiented = r1; 885 a->roworiented = r2; 886 887 ((Mat_SeqBAIJ*)baij->B->data)->roworiented = r3; /* b->roworiented */ 888 } 889 890 ierr = MatAssemblyBegin(baij->A,mode);CHKERRQ(ierr); 891 ierr = MatAssemblyEnd(baij->A,mode);CHKERRQ(ierr); 892 893 /* determine if any processor has disassembled, if so we must 894 also disassemble ourselfs, in order that we may reassemble. */ 895 /* 896 if nonzero structure of submatrix B cannot change then we know that 897 no processor disassembled thus we can skip this stuff 898 */ 899 if (!((Mat_SeqBAIJ*)baij->B->data)->nonew) { 900 ierr = MPI_Allreduce(&mat->was_assembled,&other_disassembled,1,MPIU_BOOL,MPI_PROD,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 901 if (mat->was_assembled && !other_disassembled) { 902 ierr = MatDisAssemble_MPIBAIJ(mat);CHKERRQ(ierr); 903 } 904 } 905 906 if (!mat->was_assembled && mode == MAT_FINAL_ASSEMBLY) { 907 ierr = MatSetUpMultiply_MPIBAIJ(mat);CHKERRQ(ierr); 908 } 909 ierr = MatAssemblyBegin(baij->B,mode);CHKERRQ(ierr); 910 ierr = MatAssemblyEnd(baij->B,mode);CHKERRQ(ierr); 911 912 #if defined(PETSC_USE_INFO) 913 if (baij->ht && mode== MAT_FINAL_ASSEMBLY) { 914 ierr = PetscInfo1(mat,"Average Hash Table Search in MatSetValues = %5.2f\n",((PetscReal)baij->ht_total_ct)/baij->ht_insert_ct);CHKERRQ(ierr); 915 916 baij->ht_total_ct = 0; 917 baij->ht_insert_ct = 0; 918 } 919 #endif 920 if (baij->ht_flag && !baij->ht && mode == MAT_FINAL_ASSEMBLY) { 921 ierr = MatCreateHashTable_MPIBAIJ_Private(mat,baij->ht_fact);CHKERRQ(ierr); 922 923 mat->ops->setvalues = MatSetValues_MPIBAIJ_HT; 924 mat->ops->setvaluesblocked = MatSetValuesBlocked_MPIBAIJ_HT; 925 } 926 927 ierr = PetscFree2(baij->rowvalues,baij->rowindices);CHKERRQ(ierr); 928 929 baij->rowvalues = 0; 930 PetscFunctionReturn(0); 931 } 932 933 #include <petscdraw.h> 934 #undef __FUNCT__ 935 #define __FUNCT__ "MatView_MPIBAIJ_ASCIIorDraworSocket" 936 static PetscErrorCode MatView_MPIBAIJ_ASCIIorDraworSocket(Mat mat,PetscViewer viewer) 937 { 938 Mat_MPIBAIJ *baij = (Mat_MPIBAIJ*)mat->data; 939 PetscErrorCode ierr; 940 PetscMPIInt size = baij->size,rank = baij->rank; 941 PetscInt bs = mat->rmap->bs; 942 PetscBool iascii,isdraw; 943 PetscViewer sviewer; 944 PetscViewerFormat format; 945 946 PetscFunctionBegin; 947 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERASCII,&iascii);CHKERRQ(ierr); 948 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERDRAW,&isdraw);CHKERRQ(ierr); 949 if (iascii) { 950 ierr = PetscViewerGetFormat(viewer,&format);CHKERRQ(ierr); 951 if (format == PETSC_VIEWER_ASCII_INFO_DETAIL) { 952 MatInfo info; 953 ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)mat),&rank);CHKERRQ(ierr); 954 ierr = MatGetInfo(mat,MAT_LOCAL,&info);CHKERRQ(ierr); 955 ierr = PetscViewerASCIISynchronizedAllow(viewer,PETSC_TRUE);CHKERRQ(ierr); 956 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"[%d] Local rows %D nz %D nz alloced %D bs %D mem %D\n", 957 rank,mat->rmap->n,(PetscInt)info.nz_used,(PetscInt)info.nz_allocated,mat->rmap->bs,(PetscInt)info.memory);CHKERRQ(ierr); 958 ierr = MatGetInfo(baij->A,MAT_LOCAL,&info);CHKERRQ(ierr); 959 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"[%d] on-diagonal part: nz %D \n",rank,(PetscInt)info.nz_used);CHKERRQ(ierr); 960 ierr = MatGetInfo(baij->B,MAT_LOCAL,&info);CHKERRQ(ierr); 961 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"[%d] off-diagonal part: nz %D \n",rank,(PetscInt)info.nz_used);CHKERRQ(ierr); 962 ierr = PetscViewerFlush(viewer);CHKERRQ(ierr); 963 ierr = PetscViewerASCIISynchronizedAllow(viewer,PETSC_FALSE);CHKERRQ(ierr); 964 ierr = PetscViewerASCIIPrintf(viewer,"Information on VecScatter used in matrix-vector product: \n");CHKERRQ(ierr); 965 ierr = VecScatterView(baij->Mvctx,viewer);CHKERRQ(ierr); 966 PetscFunctionReturn(0); 967 } else if (format == PETSC_VIEWER_ASCII_INFO) { 968 ierr = PetscViewerASCIIPrintf(viewer," block size is %D\n",bs);CHKERRQ(ierr); 969 PetscFunctionReturn(0); 970 } else if (format == PETSC_VIEWER_ASCII_FACTOR_INFO) { 971 PetscFunctionReturn(0); 972 } 973 } 974 975 if (isdraw) { 976 PetscDraw draw; 977 PetscBool isnull; 978 ierr = PetscViewerDrawGetDraw(viewer,0,&draw);CHKERRQ(ierr); 979 ierr = PetscDrawIsNull(draw,&isnull);CHKERRQ(ierr); if (isnull) PetscFunctionReturn(0); 980 } 981 982 if (size == 1) { 983 ierr = PetscObjectSetName((PetscObject)baij->A,((PetscObject)mat)->name);CHKERRQ(ierr); 984 ierr = MatView(baij->A,viewer);CHKERRQ(ierr); 985 } else { 986 /* assemble the entire matrix onto first processor. */ 987 Mat A; 988 Mat_SeqBAIJ *Aloc; 989 PetscInt M = mat->rmap->N,N = mat->cmap->N,*ai,*aj,col,i,j,k,*rvals,mbs = baij->mbs; 990 MatScalar *a; 991 992 /* Here we are creating a temporary matrix, so will assume MPIBAIJ is acceptable */ 993 /* Perhaps this should be the type of mat? */ 994 ierr = MatCreate(PetscObjectComm((PetscObject)mat),&A);CHKERRQ(ierr); 995 if (!rank) { 996 ierr = MatSetSizes(A,M,N,M,N);CHKERRQ(ierr); 997 } else { 998 ierr = MatSetSizes(A,0,0,M,N);CHKERRQ(ierr); 999 } 1000 ierr = MatSetType(A,MATMPIBAIJ);CHKERRQ(ierr); 1001 ierr = MatMPIBAIJSetPreallocation(A,mat->rmap->bs,0,NULL,0,NULL);CHKERRQ(ierr); 1002 ierr = MatSetOption(A,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 1003 ierr = PetscLogObjectParent((PetscObject)mat,(PetscObject)A);CHKERRQ(ierr); 1004 1005 /* copy over the A part */ 1006 Aloc = (Mat_SeqBAIJ*)baij->A->data; 1007 ai = Aloc->i; aj = Aloc->j; a = Aloc->a; 1008 ierr = PetscMalloc(bs*sizeof(PetscInt),&rvals);CHKERRQ(ierr); 1009 1010 for (i=0; i<mbs; i++) { 1011 rvals[0] = bs*(baij->rstartbs + i); 1012 for (j=1; j<bs; j++) rvals[j] = rvals[j-1] + 1; 1013 for (j=ai[i]; j<ai[i+1]; j++) { 1014 col = (baij->cstartbs+aj[j])*bs; 1015 for (k=0; k<bs; k++) { 1016 ierr = MatSetValues_MPIBAIJ(A,bs,rvals,1,&col,a,INSERT_VALUES);CHKERRQ(ierr); 1017 col++; a += bs; 1018 } 1019 } 1020 } 1021 /* copy over the B part */ 1022 Aloc = (Mat_SeqBAIJ*)baij->B->data; 1023 ai = Aloc->i; aj = Aloc->j; a = Aloc->a; 1024 for (i=0; i<mbs; i++) { 1025 rvals[0] = bs*(baij->rstartbs + i); 1026 for (j=1; j<bs; j++) rvals[j] = rvals[j-1] + 1; 1027 for (j=ai[i]; j<ai[i+1]; j++) { 1028 col = baij->garray[aj[j]]*bs; 1029 for (k=0; k<bs; k++) { 1030 ierr = MatSetValues_MPIBAIJ(A,bs,rvals,1,&col,a,INSERT_VALUES);CHKERRQ(ierr); 1031 col++; a += bs; 1032 } 1033 } 1034 } 1035 ierr = PetscFree(rvals);CHKERRQ(ierr); 1036 ierr = MatAssemblyBegin(A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1037 ierr = MatAssemblyEnd(A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1038 /* 1039 Everyone has to call to draw the matrix since the graphics waits are 1040 synchronized across all processors that share the PetscDraw object 1041 */ 1042 ierr = PetscViewerGetSingleton(viewer,&sviewer);CHKERRQ(ierr); 1043 if (!rank) { 1044 ierr = PetscObjectSetName((PetscObject)((Mat_MPIBAIJ*)(A->data))->A,((PetscObject)mat)->name);CHKERRQ(ierr); 1045 /* Set the type name to MATMPIBAIJ so that the correct type can be printed out by PetscObjectPrintClassNamePrefixType() in MatView_SeqBAIJ_ASCII()*/ 1046 PetscStrcpy(((PetscObject)((Mat_MPIBAIJ*)(A->data))->A)->type_name,MATMPIBAIJ); 1047 ierr = MatView(((Mat_MPIBAIJ*)(A->data))->A,sviewer);CHKERRQ(ierr); 1048 } 1049 ierr = PetscViewerRestoreSingleton(viewer,&sviewer);CHKERRQ(ierr); 1050 ierr = MatDestroy(&A);CHKERRQ(ierr); 1051 } 1052 PetscFunctionReturn(0); 1053 } 1054 1055 #undef __FUNCT__ 1056 #define __FUNCT__ "MatView_MPIBAIJ_Binary" 1057 static PetscErrorCode MatView_MPIBAIJ_Binary(Mat mat,PetscViewer viewer) 1058 { 1059 Mat_MPIBAIJ *a = (Mat_MPIBAIJ*)mat->data; 1060 Mat_SeqBAIJ *A = (Mat_SeqBAIJ*)a->A->data; 1061 Mat_SeqBAIJ *B = (Mat_SeqBAIJ*)a->B->data; 1062 PetscErrorCode ierr; 1063 PetscInt i,*row_lens,*crow_lens,bs = mat->rmap->bs,j,k,bs2=a->bs2,header[4],nz,rlen; 1064 PetscInt *range=0,nzmax,*column_indices,cnt,col,*garray = a->garray,cstart = mat->cmap->rstart/bs,len,pcnt,l,ll; 1065 int fd; 1066 PetscScalar *column_values; 1067 FILE *file; 1068 PetscMPIInt rank,size,tag = ((PetscObject)viewer)->tag; 1069 PetscInt message_count,flowcontrolcount; 1070 1071 PetscFunctionBegin; 1072 ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)mat),&rank);CHKERRQ(ierr); 1073 ierr = MPI_Comm_size(PetscObjectComm((PetscObject)mat),&size);CHKERRQ(ierr); 1074 nz = bs2*(A->nz + B->nz); 1075 rlen = mat->rmap->n; 1076 if (!rank) { 1077 header[0] = MAT_FILE_CLASSID; 1078 header[1] = mat->rmap->N; 1079 header[2] = mat->cmap->N; 1080 1081 ierr = MPI_Reduce(&nz,&header[3],1,MPIU_INT,MPI_SUM,0,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1082 ierr = PetscViewerBinaryGetDescriptor(viewer,&fd);CHKERRQ(ierr); 1083 ierr = PetscBinaryWrite(fd,header,4,PETSC_INT,PETSC_TRUE);CHKERRQ(ierr); 1084 /* get largest number of rows any processor has */ 1085 range = mat->rmap->range; 1086 for (i=1; i<size; i++) { 1087 rlen = PetscMax(rlen,range[i+1] - range[i]); 1088 } 1089 } else { 1090 ierr = MPI_Reduce(&nz,0,1,MPIU_INT,MPI_SUM,0,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1091 } 1092 1093 ierr = PetscMalloc((rlen/bs)*sizeof(PetscInt),&crow_lens);CHKERRQ(ierr); 1094 /* compute lengths of each row */ 1095 for (i=0; i<a->mbs; i++) { 1096 crow_lens[i] = A->i[i+1] - A->i[i] + B->i[i+1] - B->i[i]; 1097 } 1098 /* store the row lengths to the file */ 1099 ierr = PetscViewerFlowControlStart(viewer,&message_count,&flowcontrolcount);CHKERRQ(ierr); 1100 if (!rank) { 1101 MPI_Status status; 1102 ierr = PetscMalloc(rlen*sizeof(PetscInt),&row_lens);CHKERRQ(ierr); 1103 rlen = (range[1] - range[0])/bs; 1104 for (i=0; i<rlen; i++) { 1105 for (j=0; j<bs; j++) { 1106 row_lens[i*bs+j] = bs*crow_lens[i]; 1107 } 1108 } 1109 ierr = PetscBinaryWrite(fd,row_lens,bs*rlen,PETSC_INT,PETSC_TRUE);CHKERRQ(ierr); 1110 for (i=1; i<size; i++) { 1111 rlen = (range[i+1] - range[i])/bs; 1112 ierr = PetscViewerFlowControlStepMaster(viewer,i,&message_count,flowcontrolcount);CHKERRQ(ierr); 1113 ierr = MPI_Recv(crow_lens,rlen,MPIU_INT,i,tag,PetscObjectComm((PetscObject)mat),&status);CHKERRQ(ierr); 1114 for (k=0; k<rlen; k++) { 1115 for (j=0; j<bs; j++) { 1116 row_lens[k*bs+j] = bs*crow_lens[k]; 1117 } 1118 } 1119 ierr = PetscBinaryWrite(fd,row_lens,bs*rlen,PETSC_INT,PETSC_TRUE);CHKERRQ(ierr); 1120 } 1121 ierr = PetscViewerFlowControlEndMaster(viewer,&message_count);CHKERRQ(ierr); 1122 ierr = PetscFree(row_lens);CHKERRQ(ierr); 1123 } else { 1124 ierr = PetscViewerFlowControlStepWorker(viewer,rank,&message_count);CHKERRQ(ierr); 1125 ierr = MPI_Send(crow_lens,mat->rmap->n/bs,MPIU_INT,0,tag,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1126 ierr = PetscViewerFlowControlEndWorker(viewer,&message_count);CHKERRQ(ierr); 1127 } 1128 ierr = PetscFree(crow_lens);CHKERRQ(ierr); 1129 1130 /* load up the local column indices. Include for all rows not just one for each block row since process 0 does not have the 1131 information needed to make it for each row from a block row. This does require more communication but still not more than 1132 the communication needed for the nonzero values */ 1133 nzmax = nz; /* space a largest processor needs */ 1134 ierr = MPI_Reduce(&nz,&nzmax,1,MPIU_INT,MPI_MAX,0,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1135 ierr = PetscMalloc(nzmax*sizeof(PetscInt),&column_indices);CHKERRQ(ierr); 1136 cnt = 0; 1137 for (i=0; i<a->mbs; i++) { 1138 pcnt = cnt; 1139 for (j=B->i[i]; j<B->i[i+1]; j++) { 1140 if ((col = garray[B->j[j]]) > cstart) break; 1141 for (l=0; l<bs; l++) { 1142 column_indices[cnt++] = bs*col+l; 1143 } 1144 } 1145 for (k=A->i[i]; k<A->i[i+1]; k++) { 1146 for (l=0; l<bs; l++) { 1147 column_indices[cnt++] = bs*(A->j[k] + cstart)+l; 1148 } 1149 } 1150 for (; j<B->i[i+1]; j++) { 1151 for (l=0; l<bs; l++) { 1152 column_indices[cnt++] = bs*garray[B->j[j]]+l; 1153 } 1154 } 1155 len = cnt - pcnt; 1156 for (k=1; k<bs; k++) { 1157 ierr = PetscMemcpy(&column_indices[cnt],&column_indices[pcnt],len*sizeof(PetscInt));CHKERRQ(ierr); 1158 cnt += len; 1159 } 1160 } 1161 if (cnt != nz) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_LIB,"Internal PETSc error: cnt = %D nz = %D",cnt,nz); 1162 1163 /* store the columns to the file */ 1164 ierr = PetscViewerFlowControlStart(viewer,&message_count,&flowcontrolcount);CHKERRQ(ierr); 1165 if (!rank) { 1166 MPI_Status status; 1167 ierr = PetscBinaryWrite(fd,column_indices,nz,PETSC_INT,PETSC_TRUE);CHKERRQ(ierr); 1168 for (i=1; i<size; i++) { 1169 ierr = PetscViewerFlowControlStepMaster(viewer,i,&message_count,flowcontrolcount);CHKERRQ(ierr); 1170 ierr = MPI_Recv(&cnt,1,MPIU_INT,i,tag,PetscObjectComm((PetscObject)mat),&status);CHKERRQ(ierr); 1171 ierr = MPI_Recv(column_indices,cnt,MPIU_INT,i,tag,PetscObjectComm((PetscObject)mat),&status);CHKERRQ(ierr); 1172 ierr = PetscBinaryWrite(fd,column_indices,cnt,PETSC_INT,PETSC_TRUE);CHKERRQ(ierr); 1173 } 1174 ierr = PetscViewerFlowControlEndMaster(viewer,&message_count);CHKERRQ(ierr); 1175 } else { 1176 ierr = PetscViewerFlowControlStepWorker(viewer,rank,&message_count);CHKERRQ(ierr); 1177 ierr = MPI_Send(&cnt,1,MPIU_INT,0,tag,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1178 ierr = MPI_Send(column_indices,cnt,MPIU_INT,0,tag,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1179 ierr = PetscViewerFlowControlEndWorker(viewer,&message_count);CHKERRQ(ierr); 1180 } 1181 ierr = PetscFree(column_indices);CHKERRQ(ierr); 1182 1183 /* load up the numerical values */ 1184 ierr = PetscMalloc(nzmax*sizeof(PetscScalar),&column_values);CHKERRQ(ierr); 1185 cnt = 0; 1186 for (i=0; i<a->mbs; i++) { 1187 rlen = bs*(B->i[i+1] - B->i[i] + A->i[i+1] - A->i[i]); 1188 for (j=B->i[i]; j<B->i[i+1]; j++) { 1189 if (garray[B->j[j]] > cstart) break; 1190 for (l=0; l<bs; l++) { 1191 for (ll=0; ll<bs; ll++) { 1192 column_values[cnt + l*rlen + ll] = B->a[bs2*j+l+bs*ll]; 1193 } 1194 } 1195 cnt += bs; 1196 } 1197 for (k=A->i[i]; k<A->i[i+1]; k++) { 1198 for (l=0; l<bs; l++) { 1199 for (ll=0; ll<bs; ll++) { 1200 column_values[cnt + l*rlen + ll] = A->a[bs2*k+l+bs*ll]; 1201 } 1202 } 1203 cnt += bs; 1204 } 1205 for (; j<B->i[i+1]; j++) { 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 cnt += (bs-1)*rlen; 1214 } 1215 if (cnt != nz) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Internal PETSc error: cnt = %D nz = %D",cnt,nz); 1216 1217 /* store the column values to the file */ 1218 ierr = PetscViewerFlowControlStart(viewer,&message_count,&flowcontrolcount);CHKERRQ(ierr); 1219 if (!rank) { 1220 MPI_Status status; 1221 ierr = PetscBinaryWrite(fd,column_values,nz,PETSC_SCALAR,PETSC_TRUE);CHKERRQ(ierr); 1222 for (i=1; i<size; i++) { 1223 ierr = PetscViewerFlowControlStepMaster(viewer,i,&message_count,flowcontrolcount);CHKERRQ(ierr); 1224 ierr = MPI_Recv(&cnt,1,MPIU_INT,i,tag,PetscObjectComm((PetscObject)mat),&status);CHKERRQ(ierr); 1225 ierr = MPI_Recv(column_values,cnt,MPIU_SCALAR,i,tag,PetscObjectComm((PetscObject)mat),&status);CHKERRQ(ierr); 1226 ierr = PetscBinaryWrite(fd,column_values,cnt,PETSC_SCALAR,PETSC_TRUE);CHKERRQ(ierr); 1227 } 1228 ierr = PetscViewerFlowControlEndMaster(viewer,&message_count);CHKERRQ(ierr); 1229 } else { 1230 ierr = PetscViewerFlowControlStepWorker(viewer,rank,&message_count);CHKERRQ(ierr); 1231 ierr = MPI_Send(&nz,1,MPIU_INT,0,tag,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1232 ierr = MPI_Send(column_values,nz,MPIU_SCALAR,0,tag,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1233 ierr = PetscViewerFlowControlEndWorker(viewer,&message_count);CHKERRQ(ierr); 1234 } 1235 ierr = PetscFree(column_values);CHKERRQ(ierr); 1236 1237 ierr = PetscViewerBinaryGetInfoPointer(viewer,&file);CHKERRQ(ierr); 1238 if (file) { 1239 fprintf(file,"-matload_block_size %d\n",(int)mat->rmap->bs); 1240 } 1241 PetscFunctionReturn(0); 1242 } 1243 1244 #undef __FUNCT__ 1245 #define __FUNCT__ "MatView_MPIBAIJ" 1246 PetscErrorCode MatView_MPIBAIJ(Mat mat,PetscViewer viewer) 1247 { 1248 PetscErrorCode ierr; 1249 PetscBool iascii,isdraw,issocket,isbinary; 1250 1251 PetscFunctionBegin; 1252 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERASCII,&iascii);CHKERRQ(ierr); 1253 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERDRAW,&isdraw);CHKERRQ(ierr); 1254 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERSOCKET,&issocket);CHKERRQ(ierr); 1255 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERBINARY,&isbinary);CHKERRQ(ierr); 1256 if (iascii || isdraw || issocket) { 1257 ierr = MatView_MPIBAIJ_ASCIIorDraworSocket(mat,viewer);CHKERRQ(ierr); 1258 } else if (isbinary) { 1259 ierr = MatView_MPIBAIJ_Binary(mat,viewer);CHKERRQ(ierr); 1260 } 1261 PetscFunctionReturn(0); 1262 } 1263 1264 #undef __FUNCT__ 1265 #define __FUNCT__ "MatDestroy_MPIBAIJ" 1266 PetscErrorCode MatDestroy_MPIBAIJ(Mat mat) 1267 { 1268 Mat_MPIBAIJ *baij = (Mat_MPIBAIJ*)mat->data; 1269 PetscErrorCode ierr; 1270 1271 PetscFunctionBegin; 1272 #if defined(PETSC_USE_LOG) 1273 PetscLogObjectState((PetscObject)mat,"Rows=%D,Cols=%D",mat->rmap->N,mat->cmap->N); 1274 #endif 1275 ierr = MatStashDestroy_Private(&mat->stash);CHKERRQ(ierr); 1276 ierr = MatStashDestroy_Private(&mat->bstash);CHKERRQ(ierr); 1277 ierr = MatDestroy(&baij->A);CHKERRQ(ierr); 1278 ierr = MatDestroy(&baij->B);CHKERRQ(ierr); 1279 #if defined(PETSC_USE_CTABLE) 1280 ierr = PetscTableDestroy(&baij->colmap);CHKERRQ(ierr); 1281 #else 1282 ierr = PetscFree(baij->colmap);CHKERRQ(ierr); 1283 #endif 1284 ierr = PetscFree(baij->garray);CHKERRQ(ierr); 1285 ierr = VecDestroy(&baij->lvec);CHKERRQ(ierr); 1286 ierr = VecScatterDestroy(&baij->Mvctx);CHKERRQ(ierr); 1287 ierr = PetscFree2(baij->rowvalues,baij->rowindices);CHKERRQ(ierr); 1288 ierr = PetscFree(baij->barray);CHKERRQ(ierr); 1289 ierr = PetscFree2(baij->hd,baij->ht);CHKERRQ(ierr); 1290 ierr = PetscFree(baij->rangebs);CHKERRQ(ierr); 1291 ierr = PetscFree(mat->data);CHKERRQ(ierr); 1292 1293 ierr = PetscObjectChangeTypeName((PetscObject)mat,0);CHKERRQ(ierr); 1294 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatStoreValues_C",NULL);CHKERRQ(ierr); 1295 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatRetrieveValues_C",NULL);CHKERRQ(ierr); 1296 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatGetDiagonalBlock_C",NULL);CHKERRQ(ierr); 1297 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatMPIBAIJSetPreallocation_C",NULL);CHKERRQ(ierr); 1298 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatMPIBAIJSetPreallocationCSR_C",NULL);CHKERRQ(ierr); 1299 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatDiagonalScaleLocal_C",NULL);CHKERRQ(ierr); 1300 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatSetHashTableFactor_C",NULL);CHKERRQ(ierr); 1301 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatConvert_mpibaij_mpisbaij_C",NULL);CHKERRQ(ierr); 1302 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatConvert_mpibaij_mpibstrm_C",NULL);CHKERRQ(ierr); 1303 PetscFunctionReturn(0); 1304 } 1305 1306 #undef __FUNCT__ 1307 #define __FUNCT__ "MatMult_MPIBAIJ" 1308 PetscErrorCode MatMult_MPIBAIJ(Mat A,Vec xx,Vec yy) 1309 { 1310 Mat_MPIBAIJ *a = (Mat_MPIBAIJ*)A->data; 1311 PetscErrorCode ierr; 1312 PetscInt nt; 1313 1314 PetscFunctionBegin; 1315 ierr = VecGetLocalSize(xx,&nt);CHKERRQ(ierr); 1316 if (nt != A->cmap->n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Incompatible partition of A and xx"); 1317 ierr = VecGetLocalSize(yy,&nt);CHKERRQ(ierr); 1318 if (nt != A->rmap->n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Incompatible parition of A and yy"); 1319 ierr = VecScatterBegin(a->Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1320 ierr = (*a->A->ops->mult)(a->A,xx,yy);CHKERRQ(ierr); 1321 ierr = VecScatterEnd(a->Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1322 ierr = (*a->B->ops->multadd)(a->B,a->lvec,yy,yy);CHKERRQ(ierr); 1323 PetscFunctionReturn(0); 1324 } 1325 1326 #undef __FUNCT__ 1327 #define __FUNCT__ "MatMultAdd_MPIBAIJ" 1328 PetscErrorCode MatMultAdd_MPIBAIJ(Mat A,Vec xx,Vec yy,Vec zz) 1329 { 1330 Mat_MPIBAIJ *a = (Mat_MPIBAIJ*)A->data; 1331 PetscErrorCode ierr; 1332 1333 PetscFunctionBegin; 1334 ierr = VecScatterBegin(a->Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1335 ierr = (*a->A->ops->multadd)(a->A,xx,yy,zz);CHKERRQ(ierr); 1336 ierr = VecScatterEnd(a->Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1337 ierr = (*a->B->ops->multadd)(a->B,a->lvec,zz,zz);CHKERRQ(ierr); 1338 PetscFunctionReturn(0); 1339 } 1340 1341 #undef __FUNCT__ 1342 #define __FUNCT__ "MatMultTranspose_MPIBAIJ" 1343 PetscErrorCode MatMultTranspose_MPIBAIJ(Mat A,Vec xx,Vec yy) 1344 { 1345 Mat_MPIBAIJ *a = (Mat_MPIBAIJ*)A->data; 1346 PetscErrorCode ierr; 1347 PetscBool merged; 1348 1349 PetscFunctionBegin; 1350 ierr = VecScatterGetMerged(a->Mvctx,&merged);CHKERRQ(ierr); 1351 /* do nondiagonal part */ 1352 ierr = (*a->B->ops->multtranspose)(a->B,xx,a->lvec);CHKERRQ(ierr); 1353 if (!merged) { 1354 /* send it on its way */ 1355 ierr = VecScatterBegin(a->Mvctx,a->lvec,yy,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1356 /* do local part */ 1357 ierr = (*a->A->ops->multtranspose)(a->A,xx,yy);CHKERRQ(ierr); 1358 /* receive remote parts: note this assumes the values are not actually */ 1359 /* inserted in yy until the next line */ 1360 ierr = VecScatterEnd(a->Mvctx,a->lvec,yy,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1361 } else { 1362 /* do local part */ 1363 ierr = (*a->A->ops->multtranspose)(a->A,xx,yy);CHKERRQ(ierr); 1364 /* send it on its way */ 1365 ierr = VecScatterBegin(a->Mvctx,a->lvec,yy,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1366 /* values actually were received in the Begin() but we need to call this nop */ 1367 ierr = VecScatterEnd(a->Mvctx,a->lvec,yy,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1368 } 1369 PetscFunctionReturn(0); 1370 } 1371 1372 #undef __FUNCT__ 1373 #define __FUNCT__ "MatMultTransposeAdd_MPIBAIJ" 1374 PetscErrorCode MatMultTransposeAdd_MPIBAIJ(Mat A,Vec xx,Vec yy,Vec zz) 1375 { 1376 Mat_MPIBAIJ *a = (Mat_MPIBAIJ*)A->data; 1377 PetscErrorCode ierr; 1378 1379 PetscFunctionBegin; 1380 /* do nondiagonal part */ 1381 ierr = (*a->B->ops->multtranspose)(a->B,xx,a->lvec);CHKERRQ(ierr); 1382 /* send it on its way */ 1383 ierr = VecScatterBegin(a->Mvctx,a->lvec,zz,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1384 /* do local part */ 1385 ierr = (*a->A->ops->multtransposeadd)(a->A,xx,yy,zz);CHKERRQ(ierr); 1386 /* receive remote parts: note this assumes the values are not actually */ 1387 /* inserted in yy until the next line, which is true for my implementation*/ 1388 /* but is not perhaps always true. */ 1389 ierr = VecScatterEnd(a->Mvctx,a->lvec,zz,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1390 PetscFunctionReturn(0); 1391 } 1392 1393 /* 1394 This only works correctly for square matrices where the subblock A->A is the 1395 diagonal block 1396 */ 1397 #undef __FUNCT__ 1398 #define __FUNCT__ "MatGetDiagonal_MPIBAIJ" 1399 PetscErrorCode MatGetDiagonal_MPIBAIJ(Mat A,Vec v) 1400 { 1401 Mat_MPIBAIJ *a = (Mat_MPIBAIJ*)A->data; 1402 PetscErrorCode ierr; 1403 1404 PetscFunctionBegin; 1405 if (A->rmap->N != A->cmap->N) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Supports only square matrix where A->A is diag block"); 1406 ierr = MatGetDiagonal(a->A,v);CHKERRQ(ierr); 1407 PetscFunctionReturn(0); 1408 } 1409 1410 #undef __FUNCT__ 1411 #define __FUNCT__ "MatScale_MPIBAIJ" 1412 PetscErrorCode MatScale_MPIBAIJ(Mat A,PetscScalar aa) 1413 { 1414 Mat_MPIBAIJ *a = (Mat_MPIBAIJ*)A->data; 1415 PetscErrorCode ierr; 1416 1417 PetscFunctionBegin; 1418 ierr = MatScale(a->A,aa);CHKERRQ(ierr); 1419 ierr = MatScale(a->B,aa);CHKERRQ(ierr); 1420 PetscFunctionReturn(0); 1421 } 1422 1423 #undef __FUNCT__ 1424 #define __FUNCT__ "MatGetRow_MPIBAIJ" 1425 PetscErrorCode MatGetRow_MPIBAIJ(Mat matin,PetscInt row,PetscInt *nz,PetscInt **idx,PetscScalar **v) 1426 { 1427 Mat_MPIBAIJ *mat = (Mat_MPIBAIJ*)matin->data; 1428 PetscScalar *vworkA,*vworkB,**pvA,**pvB,*v_p; 1429 PetscErrorCode ierr; 1430 PetscInt bs = matin->rmap->bs,bs2 = mat->bs2,i,*cworkA,*cworkB,**pcA,**pcB; 1431 PetscInt nztot,nzA,nzB,lrow,brstart = matin->rmap->rstart,brend = matin->rmap->rend; 1432 PetscInt *cmap,*idx_p,cstart = mat->cstartbs; 1433 1434 PetscFunctionBegin; 1435 if (row < brstart || row >= brend) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Only local rows"); 1436 if (mat->getrowactive) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Already active"); 1437 mat->getrowactive = PETSC_TRUE; 1438 1439 if (!mat->rowvalues && (idx || v)) { 1440 /* 1441 allocate enough space to hold information from the longest row. 1442 */ 1443 Mat_SeqBAIJ *Aa = (Mat_SeqBAIJ*)mat->A->data,*Ba = (Mat_SeqBAIJ*)mat->B->data; 1444 PetscInt max = 1,mbs = mat->mbs,tmp; 1445 for (i=0; i<mbs; i++) { 1446 tmp = Aa->i[i+1] - Aa->i[i] + Ba->i[i+1] - Ba->i[i]; 1447 if (max < tmp) max = tmp; 1448 } 1449 ierr = PetscMalloc2(max*bs2,PetscScalar,&mat->rowvalues,max*bs2,PetscInt,&mat->rowindices);CHKERRQ(ierr); 1450 } 1451 lrow = row - brstart; 1452 1453 pvA = &vworkA; pcA = &cworkA; pvB = &vworkB; pcB = &cworkB; 1454 if (!v) {pvA = 0; pvB = 0;} 1455 if (!idx) {pcA = 0; if (!v) pcB = 0;} 1456 ierr = (*mat->A->ops->getrow)(mat->A,lrow,&nzA,pcA,pvA);CHKERRQ(ierr); 1457 ierr = (*mat->B->ops->getrow)(mat->B,lrow,&nzB,pcB,pvB);CHKERRQ(ierr); 1458 nztot = nzA + nzB; 1459 1460 cmap = mat->garray; 1461 if (v || idx) { 1462 if (nztot) { 1463 /* Sort by increasing column numbers, assuming A and B already sorted */ 1464 PetscInt imark = -1; 1465 if (v) { 1466 *v = v_p = mat->rowvalues; 1467 for (i=0; i<nzB; i++) { 1468 if (cmap[cworkB[i]/bs] < cstart) v_p[i] = vworkB[i]; 1469 else break; 1470 } 1471 imark = i; 1472 for (i=0; i<nzA; i++) v_p[imark+i] = vworkA[i]; 1473 for (i=imark; i<nzB; i++) v_p[nzA+i] = vworkB[i]; 1474 } 1475 if (idx) { 1476 *idx = idx_p = mat->rowindices; 1477 if (imark > -1) { 1478 for (i=0; i<imark; i++) { 1479 idx_p[i] = cmap[cworkB[i]/bs]*bs + cworkB[i]%bs; 1480 } 1481 } else { 1482 for (i=0; i<nzB; i++) { 1483 if (cmap[cworkB[i]/bs] < cstart) idx_p[i] = cmap[cworkB[i]/bs]*bs + cworkB[i]%bs; 1484 else break; 1485 } 1486 imark = i; 1487 } 1488 for (i=0; i<nzA; i++) idx_p[imark+i] = cstart*bs + cworkA[i]; 1489 for (i=imark; i<nzB; i++) idx_p[nzA+i] = cmap[cworkB[i]/bs]*bs + cworkB[i]%bs ; 1490 } 1491 } else { 1492 if (idx) *idx = 0; 1493 if (v) *v = 0; 1494 } 1495 } 1496 *nz = nztot; 1497 ierr = (*mat->A->ops->restorerow)(mat->A,lrow,&nzA,pcA,pvA);CHKERRQ(ierr); 1498 ierr = (*mat->B->ops->restorerow)(mat->B,lrow,&nzB,pcB,pvB);CHKERRQ(ierr); 1499 PetscFunctionReturn(0); 1500 } 1501 1502 #undef __FUNCT__ 1503 #define __FUNCT__ "MatRestoreRow_MPIBAIJ" 1504 PetscErrorCode MatRestoreRow_MPIBAIJ(Mat mat,PetscInt row,PetscInt *nz,PetscInt **idx,PetscScalar **v) 1505 { 1506 Mat_MPIBAIJ *baij = (Mat_MPIBAIJ*)mat->data; 1507 1508 PetscFunctionBegin; 1509 if (!baij->getrowactive) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"MatGetRow not called"); 1510 baij->getrowactive = PETSC_FALSE; 1511 PetscFunctionReturn(0); 1512 } 1513 1514 #undef __FUNCT__ 1515 #define __FUNCT__ "MatZeroEntries_MPIBAIJ" 1516 PetscErrorCode MatZeroEntries_MPIBAIJ(Mat A) 1517 { 1518 Mat_MPIBAIJ *l = (Mat_MPIBAIJ*)A->data; 1519 PetscErrorCode ierr; 1520 1521 PetscFunctionBegin; 1522 ierr = MatZeroEntries(l->A);CHKERRQ(ierr); 1523 ierr = MatZeroEntries(l->B);CHKERRQ(ierr); 1524 PetscFunctionReturn(0); 1525 } 1526 1527 #undef __FUNCT__ 1528 #define __FUNCT__ "MatGetInfo_MPIBAIJ" 1529 PetscErrorCode MatGetInfo_MPIBAIJ(Mat matin,MatInfoType flag,MatInfo *info) 1530 { 1531 Mat_MPIBAIJ *a = (Mat_MPIBAIJ*)matin->data; 1532 Mat A = a->A,B = a->B; 1533 PetscErrorCode ierr; 1534 PetscReal isend[5],irecv[5]; 1535 1536 PetscFunctionBegin; 1537 info->block_size = (PetscReal)matin->rmap->bs; 1538 1539 ierr = MatGetInfo(A,MAT_LOCAL,info);CHKERRQ(ierr); 1540 1541 isend[0] = info->nz_used; isend[1] = info->nz_allocated; isend[2] = info->nz_unneeded; 1542 isend[3] = info->memory; isend[4] = info->mallocs; 1543 1544 ierr = MatGetInfo(B,MAT_LOCAL,info);CHKERRQ(ierr); 1545 1546 isend[0] += info->nz_used; isend[1] += info->nz_allocated; isend[2] += info->nz_unneeded; 1547 isend[3] += info->memory; isend[4] += info->mallocs; 1548 1549 if (flag == MAT_LOCAL) { 1550 info->nz_used = isend[0]; 1551 info->nz_allocated = isend[1]; 1552 info->nz_unneeded = isend[2]; 1553 info->memory = isend[3]; 1554 info->mallocs = isend[4]; 1555 } else if (flag == MAT_GLOBAL_MAX) { 1556 ierr = MPI_Allreduce(isend,irecv,5,MPIU_REAL,MPIU_MAX,PetscObjectComm((PetscObject)matin));CHKERRQ(ierr); 1557 1558 info->nz_used = irecv[0]; 1559 info->nz_allocated = irecv[1]; 1560 info->nz_unneeded = irecv[2]; 1561 info->memory = irecv[3]; 1562 info->mallocs = irecv[4]; 1563 } else if (flag == MAT_GLOBAL_SUM) { 1564 ierr = MPI_Allreduce(isend,irecv,5,MPIU_REAL,MPIU_SUM,PetscObjectComm((PetscObject)matin));CHKERRQ(ierr); 1565 1566 info->nz_used = irecv[0]; 1567 info->nz_allocated = irecv[1]; 1568 info->nz_unneeded = irecv[2]; 1569 info->memory = irecv[3]; 1570 info->mallocs = irecv[4]; 1571 } else SETERRQ1(PetscObjectComm((PetscObject)matin),PETSC_ERR_ARG_WRONG,"Unknown MatInfoType argument %d",(int)flag); 1572 info->fill_ratio_given = 0; /* no parallel LU/ILU/Cholesky */ 1573 info->fill_ratio_needed = 0; 1574 info->factor_mallocs = 0; 1575 PetscFunctionReturn(0); 1576 } 1577 1578 #undef __FUNCT__ 1579 #define __FUNCT__ "MatSetOption_MPIBAIJ" 1580 PetscErrorCode MatSetOption_MPIBAIJ(Mat A,MatOption op,PetscBool flg) 1581 { 1582 Mat_MPIBAIJ *a = (Mat_MPIBAIJ*)A->data; 1583 PetscErrorCode ierr; 1584 1585 PetscFunctionBegin; 1586 switch (op) { 1587 case MAT_NEW_NONZERO_LOCATIONS: 1588 case MAT_NEW_NONZERO_ALLOCATION_ERR: 1589 case MAT_UNUSED_NONZERO_LOCATION_ERR: 1590 case MAT_KEEP_NONZERO_PATTERN: 1591 case MAT_NEW_NONZERO_LOCATION_ERR: 1592 ierr = MatSetOption(a->A,op,flg);CHKERRQ(ierr); 1593 ierr = MatSetOption(a->B,op,flg);CHKERRQ(ierr); 1594 break; 1595 case MAT_ROW_ORIENTED: 1596 a->roworiented = flg; 1597 1598 ierr = MatSetOption(a->A,op,flg);CHKERRQ(ierr); 1599 ierr = MatSetOption(a->B,op,flg);CHKERRQ(ierr); 1600 break; 1601 case MAT_NEW_DIAGONALS: 1602 ierr = PetscInfo1(A,"Option %s ignored\n",MatOptions[op]);CHKERRQ(ierr); 1603 break; 1604 case MAT_IGNORE_OFF_PROC_ENTRIES: 1605 a->donotstash = flg; 1606 break; 1607 case MAT_USE_HASH_TABLE: 1608 a->ht_flag = flg; 1609 break; 1610 case MAT_SYMMETRIC: 1611 case MAT_STRUCTURALLY_SYMMETRIC: 1612 case MAT_HERMITIAN: 1613 case MAT_SYMMETRY_ETERNAL: 1614 ierr = MatSetOption(a->A,op,flg);CHKERRQ(ierr); 1615 break; 1616 default: 1617 SETERRQ1(PetscObjectComm((PetscObject)A),PETSC_ERR_SUP,"unknown option %d",op); 1618 } 1619 PetscFunctionReturn(0); 1620 } 1621 1622 #undef __FUNCT__ 1623 #define __FUNCT__ "MatTranspose_MPIBAIJ" 1624 PetscErrorCode MatTranspose_MPIBAIJ(Mat A,MatReuse reuse,Mat *matout) 1625 { 1626 Mat_MPIBAIJ *baij = (Mat_MPIBAIJ*)A->data; 1627 Mat_SeqBAIJ *Aloc; 1628 Mat B; 1629 PetscErrorCode ierr; 1630 PetscInt M =A->rmap->N,N=A->cmap->N,*ai,*aj,i,*rvals,j,k,col; 1631 PetscInt bs=A->rmap->bs,mbs=baij->mbs; 1632 MatScalar *a; 1633 1634 PetscFunctionBegin; 1635 if (reuse == MAT_REUSE_MATRIX && A == *matout && M != N) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_ARG_SIZ,"Square matrix only for in-place"); 1636 if (reuse == MAT_INITIAL_MATRIX || *matout == A) { 1637 ierr = MatCreate(PetscObjectComm((PetscObject)A),&B);CHKERRQ(ierr); 1638 ierr = MatSetSizes(B,A->cmap->n,A->rmap->n,N,M);CHKERRQ(ierr); 1639 ierr = MatSetType(B,((PetscObject)A)->type_name);CHKERRQ(ierr); 1640 /* Do not know preallocation information, but must set block size */ 1641 ierr = MatMPIBAIJSetPreallocation(B,A->rmap->bs,PETSC_DECIDE,NULL,PETSC_DECIDE,NULL);CHKERRQ(ierr); 1642 } else { 1643 B = *matout; 1644 } 1645 1646 /* copy over the A part */ 1647 Aloc = (Mat_SeqBAIJ*)baij->A->data; 1648 ai = Aloc->i; aj = Aloc->j; a = Aloc->a; 1649 ierr = PetscMalloc(bs*sizeof(PetscInt),&rvals);CHKERRQ(ierr); 1650 1651 for (i=0; i<mbs; i++) { 1652 rvals[0] = bs*(baij->rstartbs + i); 1653 for (j=1; j<bs; j++) rvals[j] = rvals[j-1] + 1; 1654 for (j=ai[i]; j<ai[i+1]; j++) { 1655 col = (baij->cstartbs+aj[j])*bs; 1656 for (k=0; k<bs; k++) { 1657 ierr = MatSetValues_MPIBAIJ(B,1,&col,bs,rvals,a,INSERT_VALUES);CHKERRQ(ierr); 1658 1659 col++; a += bs; 1660 } 1661 } 1662 } 1663 /* copy over the B part */ 1664 Aloc = (Mat_SeqBAIJ*)baij->B->data; 1665 ai = Aloc->i; aj = Aloc->j; a = Aloc->a; 1666 for (i=0; i<mbs; i++) { 1667 rvals[0] = bs*(baij->rstartbs + i); 1668 for (j=1; j<bs; j++) rvals[j] = rvals[j-1] + 1; 1669 for (j=ai[i]; j<ai[i+1]; j++) { 1670 col = baij->garray[aj[j]]*bs; 1671 for (k=0; k<bs; k++) { 1672 ierr = MatSetValues_MPIBAIJ(B,1,&col,bs,rvals,a,INSERT_VALUES);CHKERRQ(ierr); 1673 col++; 1674 a += bs; 1675 } 1676 } 1677 } 1678 ierr = PetscFree(rvals);CHKERRQ(ierr); 1679 ierr = MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1680 ierr = MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1681 1682 if (reuse == MAT_INITIAL_MATRIX || *matout != A) *matout = B; 1683 else { 1684 ierr = MatHeaderMerge(A,B);CHKERRQ(ierr); 1685 } 1686 PetscFunctionReturn(0); 1687 } 1688 1689 #undef __FUNCT__ 1690 #define __FUNCT__ "MatDiagonalScale_MPIBAIJ" 1691 PetscErrorCode MatDiagonalScale_MPIBAIJ(Mat mat,Vec ll,Vec rr) 1692 { 1693 Mat_MPIBAIJ *baij = (Mat_MPIBAIJ*)mat->data; 1694 Mat a = baij->A,b = baij->B; 1695 PetscErrorCode ierr; 1696 PetscInt s1,s2,s3; 1697 1698 PetscFunctionBegin; 1699 ierr = MatGetLocalSize(mat,&s2,&s3);CHKERRQ(ierr); 1700 if (rr) { 1701 ierr = VecGetLocalSize(rr,&s1);CHKERRQ(ierr); 1702 if (s1!=s3) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"right vector non-conforming local size"); 1703 /* Overlap communication with computation. */ 1704 ierr = VecScatterBegin(baij->Mvctx,rr,baij->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1705 } 1706 if (ll) { 1707 ierr = VecGetLocalSize(ll,&s1);CHKERRQ(ierr); 1708 if (s1!=s2) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"left vector non-conforming local size"); 1709 ierr = (*b->ops->diagonalscale)(b,ll,NULL);CHKERRQ(ierr); 1710 } 1711 /* scale the diagonal block */ 1712 ierr = (*a->ops->diagonalscale)(a,ll,rr);CHKERRQ(ierr); 1713 1714 if (rr) { 1715 /* Do a scatter end and then right scale the off-diagonal block */ 1716 ierr = VecScatterEnd(baij->Mvctx,rr,baij->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1717 ierr = (*b->ops->diagonalscale)(b,NULL,baij->lvec);CHKERRQ(ierr); 1718 } 1719 PetscFunctionReturn(0); 1720 } 1721 1722 #undef __FUNCT__ 1723 #define __FUNCT__ "MatZeroRows_MPIBAIJ" 1724 PetscErrorCode MatZeroRows_MPIBAIJ(Mat A,PetscInt N,const PetscInt rows[],PetscScalar diag,Vec x,Vec b) 1725 { 1726 Mat_MPIBAIJ *l = (Mat_MPIBAIJ*)A->data; 1727 PetscErrorCode ierr; 1728 PetscMPIInt imdex,size = l->size,n,rank = l->rank; 1729 PetscInt i,*owners = A->rmap->range; 1730 PetscInt *nprocs,j,idx,nsends,row; 1731 PetscInt nmax,*svalues,*starts,*owner,nrecvs; 1732 PetscInt *rvalues,tag = ((PetscObject)A)->tag,count,base,slen,*source,lastidx = -1; 1733 PetscInt *lens,*lrows,*values,rstart_bs=A->rmap->rstart; 1734 MPI_Comm comm; 1735 MPI_Request *send_waits,*recv_waits; 1736 MPI_Status recv_status,*send_status; 1737 const PetscScalar *xx; 1738 PetscScalar *bb; 1739 #if defined(PETSC_DEBUG) 1740 PetscBool found = PETSC_FALSE; 1741 #endif 1742 1743 PetscFunctionBegin; 1744 ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr); 1745 /* first count number of contributors to each processor */ 1746 ierr = PetscMalloc(2*size*sizeof(PetscInt),&nprocs);CHKERRQ(ierr); 1747 ierr = PetscMemzero(nprocs,2*size*sizeof(PetscInt));CHKERRQ(ierr); 1748 ierr = PetscMalloc((N+1)*sizeof(PetscInt),&owner);CHKERRQ(ierr); /* see note*/ 1749 j = 0; 1750 for (i=0; i<N; i++) { 1751 if (lastidx > (idx = rows[i])) j = 0; 1752 lastidx = idx; 1753 for (; j<size; j++) { 1754 if (idx >= owners[j] && idx < owners[j+1]) { 1755 nprocs[2*j]++; 1756 nprocs[2*j+1] = 1; 1757 owner[i] = j; 1758 #if defined(PETSC_DEBUG) 1759 found = PETSC_TRUE; 1760 #endif 1761 break; 1762 } 1763 } 1764 #if defined(PETSC_DEBUG) 1765 if (!found) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Index out of range"); 1766 found = PETSC_FALSE; 1767 #endif 1768 } 1769 nsends = 0; for (i=0; i<size; i++) nsends += nprocs[2*i+1]; 1770 1771 if (A->nooffproczerorows) { 1772 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"); 1773 nrecvs = nsends; 1774 nmax = N; 1775 } else { 1776 /* inform other processors of number of messages and max length*/ 1777 ierr = PetscMaxSum(comm,nprocs,&nmax,&nrecvs);CHKERRQ(ierr); 1778 } 1779 1780 /* post receives: */ 1781 ierr = PetscMalloc((nrecvs+1)*(nmax+1)*sizeof(PetscInt),&rvalues);CHKERRQ(ierr); 1782 ierr = PetscMalloc((nrecvs+1)*sizeof(MPI_Request),&recv_waits);CHKERRQ(ierr); 1783 for (i=0; i<nrecvs; i++) { 1784 ierr = MPI_Irecv(rvalues+nmax*i,nmax,MPIU_INT,MPI_ANY_SOURCE,tag,comm,recv_waits+i);CHKERRQ(ierr); 1785 } 1786 1787 /* do sends: 1788 1) starts[i] gives the starting index in svalues for stuff going to 1789 the ith processor 1790 */ 1791 ierr = PetscMalloc((N+1)*sizeof(PetscInt),&svalues);CHKERRQ(ierr); 1792 ierr = PetscMalloc((nsends+1)*sizeof(MPI_Request),&send_waits);CHKERRQ(ierr); 1793 ierr = PetscMalloc((size+1)*sizeof(PetscInt),&starts);CHKERRQ(ierr); 1794 starts[0] = 0; 1795 for (i=1; i<size; i++) starts[i] = starts[i-1] + nprocs[2*i-2]; 1796 for (i=0; i<N; i++) { 1797 svalues[starts[owner[i]]++] = rows[i]; 1798 } 1799 1800 starts[0] = 0; 1801 for (i=1; i<size+1; i++) starts[i] = starts[i-1] + nprocs[2*i-2]; 1802 count = 0; 1803 for (i=0; i<size; i++) { 1804 if (nprocs[2*i+1]) { 1805 ierr = MPI_Isend(svalues+starts[i],nprocs[2*i],MPIU_INT,i,tag,comm,send_waits+count++);CHKERRQ(ierr); 1806 } 1807 } 1808 ierr = PetscFree(starts);CHKERRQ(ierr); 1809 1810 base = owners[rank]; 1811 1812 /* wait on receives */ 1813 ierr = PetscMalloc2(nrecvs+1,PetscInt,&lens,nrecvs+1,PetscInt,&source);CHKERRQ(ierr); 1814 count = nrecvs; 1815 slen = 0; 1816 while (count) { 1817 ierr = MPI_Waitany(nrecvs,recv_waits,&imdex,&recv_status);CHKERRQ(ierr); 1818 /* unpack receives into our local space */ 1819 ierr = MPI_Get_count(&recv_status,MPIU_INT,&n);CHKERRQ(ierr); 1820 1821 source[imdex] = recv_status.MPI_SOURCE; 1822 lens[imdex] = n; 1823 slen += n; 1824 count--; 1825 } 1826 ierr = PetscFree(recv_waits);CHKERRQ(ierr); 1827 1828 /* move the data into the send scatter */ 1829 ierr = PetscMalloc((slen+1)*sizeof(PetscInt),&lrows);CHKERRQ(ierr); 1830 count = 0; 1831 for (i=0; i<nrecvs; i++) { 1832 values = rvalues + i*nmax; 1833 for (j=0; j<lens[i]; j++) { 1834 lrows[count++] = values[j] - base; 1835 } 1836 } 1837 ierr = PetscFree(rvalues);CHKERRQ(ierr); 1838 ierr = PetscFree2(lens,source);CHKERRQ(ierr); 1839 ierr = PetscFree(owner);CHKERRQ(ierr); 1840 ierr = PetscFree(nprocs);CHKERRQ(ierr); 1841 1842 /* fix right hand side if needed */ 1843 if (x && b) { 1844 ierr = VecGetArrayRead(x,&xx);CHKERRQ(ierr); 1845 ierr = VecGetArray(b,&bb);CHKERRQ(ierr); 1846 for (i=0; i<slen; i++) { 1847 bb[lrows[i]] = diag*xx[lrows[i]]; 1848 } 1849 ierr = VecRestoreArrayRead(x,&xx);CHKERRQ(ierr); 1850 ierr = VecRestoreArray(b,&bb);CHKERRQ(ierr); 1851 } 1852 1853 /* actually zap the local rows */ 1854 /* 1855 Zero the required rows. If the "diagonal block" of the matrix 1856 is square and the user wishes to set the diagonal we use separate 1857 code so that MatSetValues() is not called for each diagonal allocating 1858 new memory, thus calling lots of mallocs and slowing things down. 1859 1860 */ 1861 /* must zero l->B before l->A because the (diag) case below may put values into l->B*/ 1862 ierr = MatZeroRows_SeqBAIJ(l->B,slen,lrows,0.0,0,0);CHKERRQ(ierr); 1863 if ((diag != 0.0) && (l->A->rmap->N == l->A->cmap->N)) { 1864 ierr = MatZeroRows_SeqBAIJ(l->A,slen,lrows,diag,0,0);CHKERRQ(ierr); 1865 } else if (diag != 0.0) { 1866 ierr = MatZeroRows_SeqBAIJ(l->A,slen,lrows,0.0,0,0);CHKERRQ(ierr); 1867 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\ 1868 MAT_NEW_NONZERO_LOCATIONS,MAT_NEW_NONZERO_LOCATION_ERR,MAT_NEW_NONZERO_ALLOCATION_ERR"); 1869 for (i=0; i<slen; i++) { 1870 row = lrows[i] + rstart_bs; 1871 ierr = MatSetValues(A,1,&row,1,&row,&diag,INSERT_VALUES);CHKERRQ(ierr); 1872 } 1873 ierr = MatAssemblyBegin(A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1874 ierr = MatAssemblyEnd(A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1875 } else { 1876 ierr = MatZeroRows_SeqBAIJ(l->A,slen,lrows,0.0,0,0);CHKERRQ(ierr); 1877 } 1878 1879 ierr = PetscFree(lrows);CHKERRQ(ierr); 1880 1881 /* wait on sends */ 1882 if (nsends) { 1883 ierr = PetscMalloc(nsends*sizeof(MPI_Status),&send_status);CHKERRQ(ierr); 1884 ierr = MPI_Waitall(nsends,send_waits,send_status);CHKERRQ(ierr); 1885 ierr = PetscFree(send_status);CHKERRQ(ierr); 1886 } 1887 ierr = PetscFree(send_waits);CHKERRQ(ierr); 1888 ierr = PetscFree(svalues);CHKERRQ(ierr); 1889 PetscFunctionReturn(0); 1890 } 1891 1892 #undef __FUNCT__ 1893 #define __FUNCT__ "MatSetUnfactored_MPIBAIJ" 1894 PetscErrorCode MatSetUnfactored_MPIBAIJ(Mat A) 1895 { 1896 Mat_MPIBAIJ *a = (Mat_MPIBAIJ*)A->data; 1897 PetscErrorCode ierr; 1898 1899 PetscFunctionBegin; 1900 ierr = MatSetUnfactored(a->A);CHKERRQ(ierr); 1901 PetscFunctionReturn(0); 1902 } 1903 1904 static PetscErrorCode MatDuplicate_MPIBAIJ(Mat,MatDuplicateOption,Mat*); 1905 1906 #undef __FUNCT__ 1907 #define __FUNCT__ "MatEqual_MPIBAIJ" 1908 PetscErrorCode MatEqual_MPIBAIJ(Mat A,Mat B,PetscBool *flag) 1909 { 1910 Mat_MPIBAIJ *matB = (Mat_MPIBAIJ*)B->data,*matA = (Mat_MPIBAIJ*)A->data; 1911 Mat a,b,c,d; 1912 PetscBool flg; 1913 PetscErrorCode ierr; 1914 1915 PetscFunctionBegin; 1916 a = matA->A; b = matA->B; 1917 c = matB->A; d = matB->B; 1918 1919 ierr = MatEqual(a,c,&flg);CHKERRQ(ierr); 1920 if (flg) { 1921 ierr = MatEqual(b,d,&flg);CHKERRQ(ierr); 1922 } 1923 ierr = MPI_Allreduce(&flg,flag,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)A));CHKERRQ(ierr); 1924 PetscFunctionReturn(0); 1925 } 1926 1927 #undef __FUNCT__ 1928 #define __FUNCT__ "MatCopy_MPIBAIJ" 1929 PetscErrorCode MatCopy_MPIBAIJ(Mat A,Mat B,MatStructure str) 1930 { 1931 PetscErrorCode ierr; 1932 Mat_MPIBAIJ *a = (Mat_MPIBAIJ*)A->data; 1933 Mat_MPIBAIJ *b = (Mat_MPIBAIJ*)B->data; 1934 1935 PetscFunctionBegin; 1936 /* If the two matrices don't have the same copy implementation, they aren't compatible for fast copy. */ 1937 if ((str != SAME_NONZERO_PATTERN) || (A->ops->copy != B->ops->copy)) { 1938 ierr = MatCopy_Basic(A,B,str);CHKERRQ(ierr); 1939 } else { 1940 ierr = MatCopy(a->A,b->A,str);CHKERRQ(ierr); 1941 ierr = MatCopy(a->B,b->B,str);CHKERRQ(ierr); 1942 } 1943 PetscFunctionReturn(0); 1944 } 1945 1946 #undef __FUNCT__ 1947 #define __FUNCT__ "MatSetUp_MPIBAIJ" 1948 PetscErrorCode MatSetUp_MPIBAIJ(Mat A) 1949 { 1950 PetscErrorCode ierr; 1951 1952 PetscFunctionBegin; 1953 ierr = MatMPIBAIJSetPreallocation(A,A->rmap->bs,PETSC_DEFAULT,0,PETSC_DEFAULT,0);CHKERRQ(ierr); 1954 PetscFunctionReturn(0); 1955 } 1956 1957 #undef __FUNCT__ 1958 #define __FUNCT__ "MatAXPY_MPIBAIJ" 1959 PetscErrorCode MatAXPY_MPIBAIJ(Mat Y,PetscScalar a,Mat X,MatStructure str) 1960 { 1961 PetscErrorCode ierr; 1962 Mat_MPIBAIJ *xx=(Mat_MPIBAIJ*)X->data,*yy=(Mat_MPIBAIJ*)Y->data; 1963 PetscBLASInt bnz,one=1; 1964 Mat_SeqBAIJ *x,*y; 1965 1966 PetscFunctionBegin; 1967 if (str == SAME_NONZERO_PATTERN) { 1968 PetscScalar alpha = a; 1969 x = (Mat_SeqBAIJ*)xx->A->data; 1970 y = (Mat_SeqBAIJ*)yy->A->data; 1971 ierr = PetscBLASIntCast(x->nz,&bnz);CHKERRQ(ierr); 1972 PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&bnz,&alpha,x->a,&one,y->a,&one)); 1973 x = (Mat_SeqBAIJ*)xx->B->data; 1974 y = (Mat_SeqBAIJ*)yy->B->data; 1975 ierr = PetscBLASIntCast(x->nz,&bnz);CHKERRQ(ierr); 1976 PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&bnz,&alpha,x->a,&one,y->a,&one)); 1977 } else { 1978 ierr = MatAXPY_Basic(Y,a,X,str);CHKERRQ(ierr); 1979 } 1980 PetscFunctionReturn(0); 1981 } 1982 1983 #undef __FUNCT__ 1984 #define __FUNCT__ "MatRealPart_MPIBAIJ" 1985 PetscErrorCode MatRealPart_MPIBAIJ(Mat A) 1986 { 1987 Mat_MPIBAIJ *a = (Mat_MPIBAIJ*)A->data; 1988 PetscErrorCode ierr; 1989 1990 PetscFunctionBegin; 1991 ierr = MatRealPart(a->A);CHKERRQ(ierr); 1992 ierr = MatRealPart(a->B);CHKERRQ(ierr); 1993 PetscFunctionReturn(0); 1994 } 1995 1996 #undef __FUNCT__ 1997 #define __FUNCT__ "MatImaginaryPart_MPIBAIJ" 1998 PetscErrorCode MatImaginaryPart_MPIBAIJ(Mat A) 1999 { 2000 Mat_MPIBAIJ *a = (Mat_MPIBAIJ*)A->data; 2001 PetscErrorCode ierr; 2002 2003 PetscFunctionBegin; 2004 ierr = MatImaginaryPart(a->A);CHKERRQ(ierr); 2005 ierr = MatImaginaryPart(a->B);CHKERRQ(ierr); 2006 PetscFunctionReturn(0); 2007 } 2008 2009 #undef __FUNCT__ 2010 #define __FUNCT__ "MatGetSubMatrix_MPIBAIJ" 2011 PetscErrorCode MatGetSubMatrix_MPIBAIJ(Mat mat,IS isrow,IS iscol,MatReuse call,Mat *newmat) 2012 { 2013 PetscErrorCode ierr; 2014 IS iscol_local; 2015 PetscInt csize; 2016 2017 PetscFunctionBegin; 2018 ierr = ISGetLocalSize(iscol,&csize);CHKERRQ(ierr); 2019 if (call == MAT_REUSE_MATRIX) { 2020 ierr = PetscObjectQuery((PetscObject)*newmat,"ISAllGather",(PetscObject*)&iscol_local);CHKERRQ(ierr); 2021 if (!iscol_local) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Submatrix passed in was not used before, cannot reuse"); 2022 } else { 2023 ierr = ISAllGather(iscol,&iscol_local);CHKERRQ(ierr); 2024 } 2025 ierr = MatGetSubMatrix_MPIBAIJ_Private(mat,isrow,iscol_local,csize,call,newmat);CHKERRQ(ierr); 2026 if (call == MAT_INITIAL_MATRIX) { 2027 ierr = PetscObjectCompose((PetscObject)*newmat,"ISAllGather",(PetscObject)iscol_local);CHKERRQ(ierr); 2028 ierr = ISDestroy(&iscol_local);CHKERRQ(ierr); 2029 } 2030 PetscFunctionReturn(0); 2031 } 2032 extern PetscErrorCode MatGetSubMatrices_MPIBAIJ_local(Mat,PetscInt,const IS[],const IS[],MatReuse,PetscBool*,PetscBool*,Mat*); 2033 #undef __FUNCT__ 2034 #define __FUNCT__ "MatGetSubMatrix_MPIBAIJ_Private" 2035 /* 2036 Not great since it makes two copies of the submatrix, first an SeqBAIJ 2037 in local and then by concatenating the local matrices the end result. 2038 Writing it directly would be much like MatGetSubMatrices_MPIBAIJ() 2039 */ 2040 PetscErrorCode MatGetSubMatrix_MPIBAIJ_Private(Mat mat,IS isrow,IS iscol,PetscInt csize,MatReuse call,Mat *newmat) 2041 { 2042 PetscErrorCode ierr; 2043 PetscMPIInt rank,size; 2044 PetscInt i,m,n,rstart,row,rend,nz,*cwork,j,bs; 2045 PetscInt *ii,*jj,nlocal,*dlens,*olens,dlen,olen,jend,mglobal,ncol,nrow; 2046 Mat M,Mreuse; 2047 MatScalar *vwork,*aa; 2048 MPI_Comm comm; 2049 IS isrow_new, iscol_new; 2050 PetscBool idflag,allrows, allcols; 2051 Mat_SeqBAIJ *aij; 2052 2053 PetscFunctionBegin; 2054 ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr); 2055 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 2056 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 2057 /* The compression and expansion should be avoided. Doesn't point 2058 out errors, might change the indices, hence buggey */ 2059 ierr = ISCompressIndicesGeneral(mat->rmap->N,mat->rmap->n,mat->rmap->bs,1,&isrow,&isrow_new);CHKERRQ(ierr); 2060 ierr = ISCompressIndicesGeneral(mat->cmap->N,mat->cmap->n,mat->cmap->bs,1,&iscol,&iscol_new);CHKERRQ(ierr); 2061 2062 /* Check for special case: each processor gets entire matrix columns */ 2063 ierr = ISIdentity(iscol,&idflag);CHKERRQ(ierr); 2064 ierr = ISGetLocalSize(iscol,&ncol);CHKERRQ(ierr); 2065 if (idflag && ncol == mat->cmap->N) allcols = PETSC_TRUE; 2066 else allcols = PETSC_FALSE; 2067 2068 ierr = ISIdentity(isrow,&idflag);CHKERRQ(ierr); 2069 ierr = ISGetLocalSize(isrow,&nrow);CHKERRQ(ierr); 2070 if (idflag && nrow == mat->rmap->N) allrows = PETSC_TRUE; 2071 else allrows = PETSC_FALSE; 2072 2073 if (call == MAT_REUSE_MATRIX) { 2074 ierr = PetscObjectQuery((PetscObject)*newmat,"SubMatrix",(PetscObject*)&Mreuse);CHKERRQ(ierr); 2075 if (!Mreuse) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Submatrix passed in was not used before, cannot reuse"); 2076 ierr = MatGetSubMatrices_MPIBAIJ_local(mat,1,&isrow_new,&iscol_new,MAT_REUSE_MATRIX,&allrows,&allcols,&Mreuse);CHKERRQ(ierr); 2077 } else { 2078 ierr = MatGetSubMatrices_MPIBAIJ_local(mat,1,&isrow_new,&iscol_new,MAT_INITIAL_MATRIX,&allrows,&allcols,&Mreuse);CHKERRQ(ierr); 2079 } 2080 ierr = ISDestroy(&isrow_new);CHKERRQ(ierr); 2081 ierr = ISDestroy(&iscol_new);CHKERRQ(ierr); 2082 /* 2083 m - number of local rows 2084 n - number of columns (same on all processors) 2085 rstart - first row in new global matrix generated 2086 */ 2087 ierr = MatGetBlockSize(mat,&bs);CHKERRQ(ierr); 2088 ierr = MatGetSize(Mreuse,&m,&n);CHKERRQ(ierr); 2089 m = m/bs; 2090 n = n/bs; 2091 2092 if (call == MAT_INITIAL_MATRIX) { 2093 aij = (Mat_SeqBAIJ*)(Mreuse)->data; 2094 ii = aij->i; 2095 jj = aij->j; 2096 2097 /* 2098 Determine the number of non-zeros in the diagonal and off-diagonal 2099 portions of the matrix in order to do correct preallocation 2100 */ 2101 2102 /* first get start and end of "diagonal" columns */ 2103 if (csize == PETSC_DECIDE) { 2104 ierr = ISGetSize(isrow,&mglobal);CHKERRQ(ierr); 2105 if (mglobal == n*bs) { /* square matrix */ 2106 nlocal = m; 2107 } else { 2108 nlocal = n/size + ((n % size) > rank); 2109 } 2110 } else { 2111 nlocal = csize/bs; 2112 } 2113 ierr = MPI_Scan(&nlocal,&rend,1,MPIU_INT,MPI_SUM,comm);CHKERRQ(ierr); 2114 rstart = rend - nlocal; 2115 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); 2116 2117 /* next, compute all the lengths */ 2118 ierr = PetscMalloc2(m+1,PetscInt,&dlens,m+1,PetscInt,&olens);CHKERRQ(ierr); 2119 for (i=0; i<m; i++) { 2120 jend = ii[i+1] - ii[i]; 2121 olen = 0; 2122 dlen = 0; 2123 for (j=0; j<jend; j++) { 2124 if (*jj < rstart || *jj >= rend) olen++; 2125 else dlen++; 2126 jj++; 2127 } 2128 olens[i] = olen; 2129 dlens[i] = dlen; 2130 } 2131 ierr = MatCreate(comm,&M);CHKERRQ(ierr); 2132 ierr = MatSetSizes(M,bs*m,bs*nlocal,PETSC_DECIDE,bs*n);CHKERRQ(ierr); 2133 ierr = MatSetType(M,((PetscObject)mat)->type_name);CHKERRQ(ierr); 2134 ierr = MatMPIBAIJSetPreallocation(M,bs,0,dlens,0,olens);CHKERRQ(ierr); 2135 ierr = PetscFree2(dlens,olens);CHKERRQ(ierr); 2136 } else { 2137 PetscInt ml,nl; 2138 2139 M = *newmat; 2140 ierr = MatGetLocalSize(M,&ml,&nl);CHKERRQ(ierr); 2141 if (ml != m) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Previous matrix must be same size/layout as request"); 2142 ierr = MatZeroEntries(M);CHKERRQ(ierr); 2143 /* 2144 The next two lines are needed so we may call MatSetValues_MPIAIJ() below directly, 2145 rather than the slower MatSetValues(). 2146 */ 2147 M->was_assembled = PETSC_TRUE; 2148 M->assembled = PETSC_FALSE; 2149 } 2150 ierr = MatSetOption(M,MAT_ROW_ORIENTED,PETSC_FALSE);CHKERRQ(ierr); 2151 ierr = MatGetOwnershipRange(M,&rstart,&rend);CHKERRQ(ierr); 2152 aij = (Mat_SeqBAIJ*)(Mreuse)->data; 2153 ii = aij->i; 2154 jj = aij->j; 2155 aa = aij->a; 2156 for (i=0; i<m; i++) { 2157 row = rstart/bs + i; 2158 nz = ii[i+1] - ii[i]; 2159 cwork = jj; jj += nz; 2160 vwork = aa; aa += nz*bs*bs; 2161 ierr = MatSetValuesBlocked_MPIBAIJ(M,1,&row,nz,cwork,vwork,INSERT_VALUES);CHKERRQ(ierr); 2162 } 2163 2164 ierr = MatAssemblyBegin(M,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2165 ierr = MatAssemblyEnd(M,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2166 *newmat = M; 2167 2168 /* save submatrix used in processor for next request */ 2169 if (call == MAT_INITIAL_MATRIX) { 2170 ierr = PetscObjectCompose((PetscObject)M,"SubMatrix",(PetscObject)Mreuse);CHKERRQ(ierr); 2171 ierr = PetscObjectDereference((PetscObject)Mreuse);CHKERRQ(ierr); 2172 } 2173 PetscFunctionReturn(0); 2174 } 2175 2176 #undef __FUNCT__ 2177 #define __FUNCT__ "MatPermute_MPIBAIJ" 2178 PetscErrorCode MatPermute_MPIBAIJ(Mat A,IS rowp,IS colp,Mat *B) 2179 { 2180 MPI_Comm comm,pcomm; 2181 PetscInt clocal_size,nrows; 2182 const PetscInt *rows; 2183 PetscMPIInt size; 2184 IS crowp,lcolp; 2185 PetscErrorCode ierr; 2186 2187 PetscFunctionBegin; 2188 ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr); 2189 /* make a collective version of 'rowp' */ 2190 ierr = PetscObjectGetComm((PetscObject)rowp,&pcomm);CHKERRQ(ierr); 2191 if (pcomm==comm) { 2192 crowp = rowp; 2193 } else { 2194 ierr = ISGetSize(rowp,&nrows);CHKERRQ(ierr); 2195 ierr = ISGetIndices(rowp,&rows);CHKERRQ(ierr); 2196 ierr = ISCreateGeneral(comm,nrows,rows,PETSC_COPY_VALUES,&crowp);CHKERRQ(ierr); 2197 ierr = ISRestoreIndices(rowp,&rows);CHKERRQ(ierr); 2198 } 2199 ierr = ISSetPermutation(crowp);CHKERRQ(ierr); 2200 /* make a local version of 'colp' */ 2201 ierr = PetscObjectGetComm((PetscObject)colp,&pcomm);CHKERRQ(ierr); 2202 ierr = MPI_Comm_size(pcomm,&size);CHKERRQ(ierr); 2203 if (size==1) { 2204 lcolp = colp; 2205 } else { 2206 ierr = ISAllGather(colp,&lcolp);CHKERRQ(ierr); 2207 } 2208 ierr = ISSetPermutation(lcolp);CHKERRQ(ierr); 2209 /* now we just get the submatrix */ 2210 ierr = MatGetLocalSize(A,PETSC_NULL,&clocal_size);CHKERRQ(ierr); 2211 ierr = MatGetSubMatrix_MPIBAIJ_Private(A,crowp,lcolp,clocal_size,MAT_INITIAL_MATRIX,B);CHKERRQ(ierr); 2212 /* clean up */ 2213 if (pcomm!=comm) { 2214 ierr = ISDestroy(&crowp);CHKERRQ(ierr); 2215 } 2216 if (size>1) { 2217 ierr = ISDestroy(&lcolp);CHKERRQ(ierr); 2218 } 2219 PetscFunctionReturn(0); 2220 } 2221 2222 #undef __FUNCT__ 2223 #define __FUNCT__ "MatGetGhosts_MPIBAIJ" 2224 PetscErrorCode MatGetGhosts_MPIBAIJ(Mat mat,PetscInt *nghosts,const PetscInt *ghosts[]) 2225 { 2226 Mat_MPIBAIJ *baij = (Mat_MPIBAIJ*) mat->data; 2227 Mat_SeqBAIJ *B = (Mat_SeqBAIJ*)baij->B->data; 2228 2229 PetscFunctionBegin; 2230 if (nghosts) *nghosts = B->nbs; 2231 if (ghosts) *ghosts = baij->garray; 2232 PetscFunctionReturn(0); 2233 } 2234 2235 extern PetscErrorCode MatCreateColmap_MPIBAIJ_Private(Mat); 2236 2237 #undef __FUNCT__ 2238 #define __FUNCT__ "MatFDColoringCreate_MPIBAIJ" 2239 /* 2240 This routine is almost identical to MatFDColoringCreate_MPIBAIJ()! 2241 */ 2242 PetscErrorCode MatFDColoringCreate_MPIBAIJ(Mat mat,ISColoring iscoloring,MatFDColoring c) 2243 { 2244 Mat_MPIBAIJ *baij = (Mat_MPIBAIJ*)mat->data; 2245 PetscErrorCode ierr; 2246 PetscMPIInt size,*ncolsonproc,*disp,nn; 2247 PetscInt bs,i,n,nrows,j,k,m,ncols,col; 2248 const PetscInt *is,*rows = 0,*A_ci,*A_cj,*B_ci,*B_cj,*ltog; 2249 PetscInt nis = iscoloring->n,nctot,*cols; 2250 PetscInt *rowhit,M,cstart,cend,colb; 2251 PetscInt *columnsforrow,l; 2252 IS *isa; 2253 PetscBool done,flg; 2254 ISLocalToGlobalMapping map = mat->cmap->bmapping; 2255 PetscInt ctype=c->ctype; 2256 2257 PetscFunctionBegin; 2258 if (!mat->assembled) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_ARG_WRONGSTATE,"Matrix must be assembled first; MatAssemblyBegin/End();"); 2259 if (ctype == IS_COLORING_GHOSTED && !map) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_ARG_INCOMP,"When using ghosted differencing matrix must have local to global mapping provided with MatSetLocalToGlobalMappingBlock"); 2260 2261 if (map) {ierr = ISLocalToGlobalMappingGetIndices(map,<og);CHKERRQ(ierr);} 2262 else ltog = NULL; 2263 ierr = ISColoringGetIS(iscoloring,PETSC_IGNORE,&isa);CHKERRQ(ierr); 2264 ierr = MatGetBlockSize(mat,&bs);CHKERRQ(ierr); 2265 2266 M = mat->rmap->n/bs; 2267 cstart = mat->cmap->rstart/bs; 2268 cend = mat->cmap->rend/bs; 2269 c->M = mat->rmap->N/bs; /* set the global rows and columns and local rows */ 2270 c->N = mat->cmap->N/bs; 2271 c->m = mat->rmap->n/bs; 2272 c->rstart = mat->rmap->rstart/bs; 2273 2274 c->ncolors = nis; 2275 ierr = PetscMalloc(nis*sizeof(PetscInt),&c->ncolumns);CHKERRQ(ierr); 2276 ierr = PetscMalloc(nis*sizeof(PetscInt*),&c->columns);CHKERRQ(ierr); 2277 ierr = PetscMalloc(nis*sizeof(PetscInt),&c->nrows);CHKERRQ(ierr); 2278 ierr = PetscMalloc(nis*sizeof(PetscInt*),&c->rows);CHKERRQ(ierr); 2279 ierr = PetscMalloc(nis*sizeof(PetscInt*),&c->columnsforrow);CHKERRQ(ierr); 2280 ierr = PetscLogObjectMemory((PetscObject)c,5*nis*sizeof(PetscInt));CHKERRQ(ierr); 2281 2282 /* Allow access to data structures of local part of matrix */ 2283 if (!baij->colmap) { 2284 ierr = MatCreateColmap_MPIBAIJ_Private(mat);CHKERRQ(ierr); 2285 } 2286 ierr = MatGetColumnIJ(baij->A,0,PETSC_FALSE,PETSC_FALSE,&ncols,&A_ci,&A_cj,&done);CHKERRQ(ierr); 2287 ierr = MatGetColumnIJ(baij->B,0,PETSC_FALSE,PETSC_FALSE,&ncols,&B_ci,&B_cj,&done);CHKERRQ(ierr); 2288 2289 ierr = PetscMalloc((M+1)*sizeof(PetscInt),&rowhit);CHKERRQ(ierr); 2290 ierr = PetscMalloc((M+1)*sizeof(PetscInt),&columnsforrow);CHKERRQ(ierr); 2291 2292 for (i=0; i<nis; i++) { 2293 ierr = ISGetLocalSize(isa[i],&n);CHKERRQ(ierr); 2294 ierr = ISGetIndices(isa[i],&is);CHKERRQ(ierr); 2295 2296 c->ncolumns[i] = n; 2297 if (n) { 2298 ierr = PetscMalloc(n*sizeof(PetscInt),&c->columns[i]);CHKERRQ(ierr); 2299 ierr = PetscLogObjectMemory((PetscObject)c,n*sizeof(PetscInt));CHKERRQ(ierr); 2300 ierr = PetscMemcpy(c->columns[i],is,n*sizeof(PetscInt));CHKERRQ(ierr); 2301 } else { 2302 c->columns[i] = 0; 2303 } 2304 2305 if (ctype == IS_COLORING_GLOBAL) { 2306 /* Determine the total (parallel) number of columns of this color */ 2307 ierr = MPI_Comm_size(PetscObjectComm((PetscObject)mat),&size);CHKERRQ(ierr); 2308 ierr = PetscMalloc2(size,PetscMPIInt,&ncolsonproc,size,PetscMPIInt,&disp);CHKERRQ(ierr); 2309 2310 ierr = PetscMPIIntCast(n,&nn);CHKERRQ(ierr); 2311 ierr = MPI_Allgather(&nn,1,MPI_INT,ncolsonproc,1,MPI_INT,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 2312 nctot = 0; for (j=0; j<size; j++) nctot += ncolsonproc[j]; 2313 if (!nctot) { 2314 ierr = PetscInfo(mat,"Coloring of matrix has some unneeded colors with no corresponding rows\n");CHKERRQ(ierr); 2315 } 2316 2317 disp[0] = 0; 2318 for (j=1; j<size; j++) { 2319 disp[j] = disp[j-1] + ncolsonproc[j-1]; 2320 } 2321 2322 /* Get complete list of columns for color on each processor */ 2323 ierr = PetscMalloc((nctot+1)*sizeof(PetscInt),&cols);CHKERRQ(ierr); 2324 ierr = MPI_Allgatherv((void*)is,n,MPIU_INT,cols,ncolsonproc,disp,MPIU_INT,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 2325 ierr = PetscFree2(ncolsonproc,disp);CHKERRQ(ierr); 2326 } else if (ctype == IS_COLORING_GHOSTED) { 2327 /* Determine local number of columns of this color on this process, including ghost points */ 2328 nctot = n; 2329 ierr = PetscMalloc((nctot+1)*sizeof(PetscInt),&cols);CHKERRQ(ierr); 2330 ierr = PetscMemcpy(cols,is,n*sizeof(PetscInt));CHKERRQ(ierr); 2331 } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Not provided for this MatFDColoring type"); 2332 2333 /* 2334 Mark all rows affect by these columns 2335 */ 2336 /* Temporary option to allow for debugging/testing */ 2337 flg = PETSC_FALSE; 2338 ierr = PetscOptionsGetBool(NULL,"-matfdcoloring_slow",&flg,NULL);CHKERRQ(ierr); 2339 if (!flg) { /*-----------------------------------------------------------------------------*/ 2340 /* crude, fast version */ 2341 ierr = PetscMemzero(rowhit,M*sizeof(PetscInt));CHKERRQ(ierr); 2342 /* loop over columns*/ 2343 for (j=0; j<nctot; j++) { 2344 if (ctype == IS_COLORING_GHOSTED) { 2345 col = ltog[cols[j]]; 2346 } else { 2347 col = cols[j]; 2348 } 2349 if (col >= cstart && col < cend) { 2350 /* column is in diagonal block of matrix */ 2351 rows = A_cj + A_ci[col-cstart]; 2352 m = A_ci[col-cstart+1] - A_ci[col-cstart]; 2353 } else { 2354 #if defined(PETSC_USE_CTABLE) 2355 ierr = PetscTableFind(baij->colmap,col+1,&colb);CHKERRQ(ierr); 2356 colb--; 2357 #else 2358 colb = baij->colmap[col] - 1; 2359 #endif 2360 if (colb == -1) { 2361 m = 0; 2362 } else { 2363 colb = colb/bs; 2364 rows = B_cj + B_ci[colb]; 2365 m = B_ci[colb+1] - B_ci[colb]; 2366 } 2367 } 2368 /* loop over columns marking them in rowhit */ 2369 for (k=0; k<m; k++) { 2370 rowhit[*rows++] = col + 1; 2371 } 2372 } 2373 2374 /* count the number of hits */ 2375 nrows = 0; 2376 for (j=0; j<M; j++) { 2377 if (rowhit[j]) nrows++; 2378 } 2379 c->nrows[i] = nrows; 2380 ierr = PetscMalloc((nrows+1)*sizeof(PetscInt),&c->rows[i]);CHKERRQ(ierr); 2381 ierr = PetscMalloc((nrows+1)*sizeof(PetscInt),&c->columnsforrow[i]);CHKERRQ(ierr); 2382 ierr = PetscLogObjectMemory((PetscObject)c,2*(nrows+1)*sizeof(PetscInt));CHKERRQ(ierr); 2383 nrows = 0; 2384 for (j=0; j<M; j++) { 2385 if (rowhit[j]) { 2386 c->rows[i][nrows] = j; 2387 c->columnsforrow[i][nrows] = rowhit[j] - 1; 2388 nrows++; 2389 } 2390 } 2391 } else { /*-------------------------------------------------------------------------------*/ 2392 /* slow version, using rowhit as a linked list */ 2393 PetscInt currentcol,fm,mfm; 2394 rowhit[M] = M; 2395 nrows = 0; 2396 /* loop over columns*/ 2397 for (j=0; j<nctot; j++) { 2398 if (ctype == IS_COLORING_GHOSTED) { 2399 col = ltog[cols[j]]; 2400 } else { 2401 col = cols[j]; 2402 } 2403 if (col >= cstart && col < cend) { 2404 /* column is in diagonal block of matrix */ 2405 rows = A_cj + A_ci[col-cstart]; 2406 m = A_ci[col-cstart+1] - A_ci[col-cstart]; 2407 } else { 2408 #if defined(PETSC_USE_CTABLE) 2409 ierr = PetscTableFind(baij->colmap,col+1,&colb);CHKERRQ(ierr); 2410 colb--; 2411 #else 2412 colb = baij->colmap[col] - 1; 2413 #endif 2414 if (colb == -1) { 2415 m = 0; 2416 } else { 2417 colb = colb/bs; 2418 rows = B_cj + B_ci[colb]; 2419 m = B_ci[colb+1] - B_ci[colb]; 2420 } 2421 } 2422 2423 /* loop over columns marking them in rowhit */ 2424 fm = M; /* fm points to first entry in linked list */ 2425 for (k=0; k<m; k++) { 2426 currentcol = *rows++; 2427 /* is it already in the list? */ 2428 do { 2429 mfm = fm; 2430 fm = rowhit[fm]; 2431 } while (fm < currentcol); 2432 /* not in list so add it */ 2433 if (fm != currentcol) { 2434 nrows++; 2435 columnsforrow[currentcol] = col; 2436 /* next three lines insert new entry into linked list */ 2437 rowhit[mfm] = currentcol; 2438 rowhit[currentcol] = fm; 2439 fm = currentcol; 2440 /* fm points to present position in list since we know the columns are sorted */ 2441 } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Invalid coloring of matrix detected"); 2442 } 2443 } 2444 c->nrows[i] = nrows; 2445 ierr = PetscMalloc((nrows+1)*sizeof(PetscInt),&c->rows[i]);CHKERRQ(ierr); 2446 ierr = PetscMalloc((nrows+1)*sizeof(PetscInt),&c->columnsforrow[i]);CHKERRQ(ierr); 2447 ierr = PetscLogObjectMemory((PetscObject)c,(nrows+1)*sizeof(PetscInt));CHKERRQ(ierr); 2448 /* now store the linked list of rows into c->rows[i] */ 2449 nrows = 0; 2450 fm = rowhit[M]; 2451 do { 2452 c->rows[i][nrows] = fm; 2453 c->columnsforrow[i][nrows++] = columnsforrow[fm]; 2454 fm = rowhit[fm]; 2455 } while (fm < M); 2456 } /* ---------------------------------------------------------------------------------------*/ 2457 ierr = PetscFree(cols);CHKERRQ(ierr); 2458 } 2459 2460 /* Optimize by adding the vscale, and scaleforrow[][] fields */ 2461 /* 2462 vscale will contain the "diagonal" on processor scalings followed by the off processor 2463 */ 2464 if (ctype == IS_COLORING_GLOBAL) { 2465 PetscInt *garray; 2466 ierr = PetscMalloc(baij->B->cmap->n*sizeof(PetscInt),&garray);CHKERRQ(ierr); 2467 for (i=0; i<baij->B->cmap->n/bs; i++) { 2468 for (j=0; j<bs; j++) { 2469 garray[i*bs+j] = bs*baij->garray[i]+j; 2470 } 2471 } 2472 ierr = VecCreateGhost(PetscObjectComm((PetscObject)mat),baij->A->rmap->n,PETSC_DETERMINE,baij->B->cmap->n,garray,&c->vscale);CHKERRQ(ierr); 2473 ierr = PetscFree(garray);CHKERRQ(ierr); 2474 ierr = PetscMalloc(c->ncolors*sizeof(PetscInt*),&c->vscaleforrow);CHKERRQ(ierr); 2475 for (k=0; k<c->ncolors; k++) { 2476 ierr = PetscMalloc((c->nrows[k]+1)*sizeof(PetscInt),&c->vscaleforrow[k]);CHKERRQ(ierr); 2477 for (l=0; l<c->nrows[k]; l++) { 2478 col = c->columnsforrow[k][l]; 2479 if (col >= cstart && col < cend) { 2480 /* column is in diagonal block of matrix */ 2481 colb = col - cstart; 2482 } else { 2483 /* column is in "off-processor" part */ 2484 #if defined(PETSC_USE_CTABLE) 2485 ierr = PetscTableFind(baij->colmap,col+1,&colb);CHKERRQ(ierr); 2486 colb--; 2487 #else 2488 colb = baij->colmap[col] - 1; 2489 #endif 2490 colb = colb/bs; 2491 colb += cend - cstart; 2492 } 2493 c->vscaleforrow[k][l] = colb; 2494 } 2495 } 2496 } else if (ctype == IS_COLORING_GHOSTED) { 2497 /* Get gtol mapping */ 2498 PetscInt N = mat->cmap->N,nlocal,*gtol; 2499 ierr = PetscMalloc((N+1)*sizeof(PetscInt),>ol);CHKERRQ(ierr); 2500 for (i=0; i<N; i++) gtol[i] = -1; 2501 ierr = ISLocalToGlobalMappingGetSize(map,&nlocal);CHKERRQ(ierr); 2502 for (i=0; i<nlocal; i++) gtol[ltog[i]] = i; 2503 2504 c->vscale = 0; /* will be created in MatFDColoringApply() */ 2505 ierr = PetscMalloc(c->ncolors*sizeof(PetscInt*),&c->vscaleforrow);CHKERRQ(ierr); 2506 for (k=0; k<c->ncolors; k++) { 2507 ierr = PetscMalloc((c->nrows[k]+1)*sizeof(PetscInt),&c->vscaleforrow[k]);CHKERRQ(ierr); 2508 for (l=0; l<c->nrows[k]; l++) { 2509 col = c->columnsforrow[k][l]; /* global column index */ 2510 2511 c->vscaleforrow[k][l] = gtol[col]; /* local column index */ 2512 } 2513 } 2514 ierr = PetscFree(gtol);CHKERRQ(ierr); 2515 } 2516 ierr = ISColoringRestoreIS(iscoloring,&isa);CHKERRQ(ierr); 2517 2518 ierr = PetscFree(rowhit);CHKERRQ(ierr); 2519 ierr = PetscFree(columnsforrow);CHKERRQ(ierr); 2520 ierr = MatRestoreColumnIJ(baij->A,0,PETSC_FALSE,PETSC_FALSE,&ncols,&A_ci,&A_cj,&done);CHKERRQ(ierr); 2521 ierr = MatRestoreColumnIJ(baij->B,0,PETSC_FALSE,PETSC_FALSE,&ncols,&B_ci,&B_cj,&done);CHKERRQ(ierr); 2522 if (map) {ierr = ISLocalToGlobalMappingRestoreIndices(map,<og);CHKERRQ(ierr);} 2523 PetscFunctionReturn(0); 2524 } 2525 2526 #undef __FUNCT__ 2527 #define __FUNCT__ "MatGetSeqNonzeroStructure_MPIBAIJ" 2528 PetscErrorCode MatGetSeqNonzeroStructure_MPIBAIJ(Mat A,Mat *newmat) 2529 { 2530 Mat B; 2531 Mat_MPIBAIJ *a = (Mat_MPIBAIJ*)A->data; 2532 Mat_SeqBAIJ *ad = (Mat_SeqBAIJ*)a->A->data,*bd = (Mat_SeqBAIJ*)a->B->data; 2533 Mat_SeqAIJ *b; 2534 PetscErrorCode ierr; 2535 PetscMPIInt size,rank,*recvcounts = 0,*displs = 0; 2536 PetscInt sendcount,i,*rstarts = A->rmap->range,n,cnt,j,bs = A->rmap->bs; 2537 PetscInt m,*garray = a->garray,*lens,*jsendbuf,*a_jsendbuf,*b_jsendbuf; 2538 2539 PetscFunctionBegin; 2540 ierr = MPI_Comm_size(PetscObjectComm((PetscObject)A),&size);CHKERRQ(ierr); 2541 ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)A),&rank);CHKERRQ(ierr); 2542 2543 /* ---------------------------------------------------------------- 2544 Tell every processor the number of nonzeros per row 2545 */ 2546 ierr = PetscMalloc((A->rmap->N/bs)*sizeof(PetscInt),&lens);CHKERRQ(ierr); 2547 for (i=A->rmap->rstart/bs; i<A->rmap->rend/bs; i++) { 2548 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]; 2549 } 2550 sendcount = A->rmap->rend/bs - A->rmap->rstart/bs; 2551 ierr = PetscMalloc(2*size*sizeof(PetscMPIInt),&recvcounts);CHKERRQ(ierr); 2552 displs = recvcounts + size; 2553 for (i=0; i<size; i++) { 2554 recvcounts[i] = A->rmap->range[i+1]/bs - A->rmap->range[i]/bs; 2555 displs[i] = A->rmap->range[i]/bs; 2556 } 2557 #if defined(PETSC_HAVE_MPI_IN_PLACE) 2558 ierr = MPI_Allgatherv(MPI_IN_PLACE,0,MPI_DATATYPE_NULL,lens,recvcounts,displs,MPIU_INT,PetscObjectComm((PetscObject)A));CHKERRQ(ierr); 2559 #else 2560 ierr = MPI_Allgatherv(lens+A->rmap->rstart/bs,sendcount,MPIU_INT,lens,recvcounts,displs,MPIU_INT,PetscObjectComm((PetscObject)A));CHKERRQ(ierr); 2561 #endif 2562 /* --------------------------------------------------------------- 2563 Create the sequential matrix of the same type as the local block diagonal 2564 */ 2565 ierr = MatCreate(PETSC_COMM_SELF,&B);CHKERRQ(ierr); 2566 ierr = MatSetSizes(B,A->rmap->N/bs,A->cmap->N/bs,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 2567 ierr = MatSetType(B,MATSEQAIJ);CHKERRQ(ierr); 2568 ierr = MatSeqAIJSetPreallocation(B,0,lens);CHKERRQ(ierr); 2569 b = (Mat_SeqAIJ*)B->data; 2570 2571 /*-------------------------------------------------------------------- 2572 Copy my part of matrix column indices over 2573 */ 2574 sendcount = ad->nz + bd->nz; 2575 jsendbuf = b->j + b->i[rstarts[rank]/bs]; 2576 a_jsendbuf = ad->j; 2577 b_jsendbuf = bd->j; 2578 n = A->rmap->rend/bs - A->rmap->rstart/bs; 2579 cnt = 0; 2580 for (i=0; i<n; i++) { 2581 2582 /* put in lower diagonal portion */ 2583 m = bd->i[i+1] - bd->i[i]; 2584 while (m > 0) { 2585 /* is it above diagonal (in bd (compressed) numbering) */ 2586 if (garray[*b_jsendbuf] > A->rmap->rstart/bs + i) break; 2587 jsendbuf[cnt++] = garray[*b_jsendbuf++]; 2588 m--; 2589 } 2590 2591 /* put in diagonal portion */ 2592 for (j=ad->i[i]; j<ad->i[i+1]; j++) { 2593 jsendbuf[cnt++] = A->rmap->rstart/bs + *a_jsendbuf++; 2594 } 2595 2596 /* put in upper diagonal portion */ 2597 while (m-- > 0) { 2598 jsendbuf[cnt++] = garray[*b_jsendbuf++]; 2599 } 2600 } 2601 if (cnt != sendcount) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Corrupted PETSc matrix: nz given %D actual nz %D",sendcount,cnt); 2602 2603 /*-------------------------------------------------------------------- 2604 Gather all column indices to all processors 2605 */ 2606 for (i=0; i<size; i++) { 2607 recvcounts[i] = 0; 2608 for (j=A->rmap->range[i]/bs; j<A->rmap->range[i+1]/bs; j++) { 2609 recvcounts[i] += lens[j]; 2610 } 2611 } 2612 displs[0] = 0; 2613 for (i=1; i<size; i++) { 2614 displs[i] = displs[i-1] + recvcounts[i-1]; 2615 } 2616 #if defined(PETSC_HAVE_MPI_IN_PLACE) 2617 ierr = MPI_Allgatherv(MPI_IN_PLACE,0,MPI_DATATYPE_NULL,b->j,recvcounts,displs,MPIU_INT,PetscObjectComm((PetscObject)A));CHKERRQ(ierr); 2618 #else 2619 ierr = MPI_Allgatherv(jsendbuf,sendcount,MPIU_INT,b->j,recvcounts,displs,MPIU_INT,PetscObjectComm((PetscObject)A));CHKERRQ(ierr); 2620 #endif 2621 /*-------------------------------------------------------------------- 2622 Assemble the matrix into useable form (note numerical values not yet set) 2623 */ 2624 /* set the b->ilen (length of each row) values */ 2625 ierr = PetscMemcpy(b->ilen,lens,(A->rmap->N/bs)*sizeof(PetscInt));CHKERRQ(ierr); 2626 /* set the b->i indices */ 2627 b->i[0] = 0; 2628 for (i=1; i<=A->rmap->N/bs; i++) { 2629 b->i[i] = b->i[i-1] + lens[i-1]; 2630 } 2631 ierr = PetscFree(lens);CHKERRQ(ierr); 2632 ierr = MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2633 ierr = MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2634 ierr = PetscFree(recvcounts);CHKERRQ(ierr); 2635 2636 if (A->symmetric) { 2637 ierr = MatSetOption(B,MAT_SYMMETRIC,PETSC_TRUE);CHKERRQ(ierr); 2638 } else if (A->hermitian) { 2639 ierr = MatSetOption(B,MAT_HERMITIAN,PETSC_TRUE);CHKERRQ(ierr); 2640 } else if (A->structurally_symmetric) { 2641 ierr = MatSetOption(B,MAT_STRUCTURALLY_SYMMETRIC,PETSC_TRUE);CHKERRQ(ierr); 2642 } 2643 *newmat = B; 2644 PetscFunctionReturn(0); 2645 } 2646 2647 #undef __FUNCT__ 2648 #define __FUNCT__ "MatSOR_MPIBAIJ" 2649 PetscErrorCode MatSOR_MPIBAIJ(Mat matin,Vec bb,PetscReal omega,MatSORType flag,PetscReal fshift,PetscInt its,PetscInt lits,Vec xx) 2650 { 2651 Mat_MPIBAIJ *mat = (Mat_MPIBAIJ*)matin->data; 2652 PetscErrorCode ierr; 2653 Vec bb1 = 0; 2654 2655 PetscFunctionBegin; 2656 if (flag == SOR_APPLY_UPPER) { 2657 ierr = (*mat->A->ops->sor)(mat->A,bb,omega,flag,fshift,lits,1,xx);CHKERRQ(ierr); 2658 PetscFunctionReturn(0); 2659 } 2660 2661 if (its > 1 || ~flag & SOR_ZERO_INITIAL_GUESS) { 2662 ierr = VecDuplicate(bb,&bb1);CHKERRQ(ierr); 2663 } 2664 2665 if ((flag & SOR_LOCAL_SYMMETRIC_SWEEP) == SOR_LOCAL_SYMMETRIC_SWEEP) { 2666 if (flag & SOR_ZERO_INITIAL_GUESS) { 2667 ierr = (*mat->A->ops->sor)(mat->A,bb,omega,flag,fshift,lits,1,xx);CHKERRQ(ierr); 2668 its--; 2669 } 2670 2671 while (its--) { 2672 ierr = VecScatterBegin(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2673 ierr = VecScatterEnd(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2674 2675 /* update rhs: bb1 = bb - B*x */ 2676 ierr = VecScale(mat->lvec,-1.0);CHKERRQ(ierr); 2677 ierr = (*mat->B->ops->multadd)(mat->B,mat->lvec,bb,bb1);CHKERRQ(ierr); 2678 2679 /* local sweep */ 2680 ierr = (*mat->A->ops->sor)(mat->A,bb1,omega,SOR_SYMMETRIC_SWEEP,fshift,lits,1,xx);CHKERRQ(ierr); 2681 } 2682 } else if (flag & SOR_LOCAL_FORWARD_SWEEP) { 2683 if (flag & SOR_ZERO_INITIAL_GUESS) { 2684 ierr = (*mat->A->ops->sor)(mat->A,bb,omega,flag,fshift,lits,1,xx);CHKERRQ(ierr); 2685 its--; 2686 } 2687 while (its--) { 2688 ierr = VecScatterBegin(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2689 ierr = VecScatterEnd(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2690 2691 /* update rhs: bb1 = bb - B*x */ 2692 ierr = VecScale(mat->lvec,-1.0);CHKERRQ(ierr); 2693 ierr = (*mat->B->ops->multadd)(mat->B,mat->lvec,bb,bb1);CHKERRQ(ierr); 2694 2695 /* local sweep */ 2696 ierr = (*mat->A->ops->sor)(mat->A,bb1,omega,SOR_FORWARD_SWEEP,fshift,lits,1,xx);CHKERRQ(ierr); 2697 } 2698 } else if (flag & SOR_LOCAL_BACKWARD_SWEEP) { 2699 if (flag & SOR_ZERO_INITIAL_GUESS) { 2700 ierr = (*mat->A->ops->sor)(mat->A,bb,omega,flag,fshift,lits,1,xx);CHKERRQ(ierr); 2701 its--; 2702 } 2703 while (its--) { 2704 ierr = VecScatterBegin(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2705 ierr = VecScatterEnd(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2706 2707 /* update rhs: bb1 = bb - B*x */ 2708 ierr = VecScale(mat->lvec,-1.0);CHKERRQ(ierr); 2709 ierr = (*mat->B->ops->multadd)(mat->B,mat->lvec,bb,bb1);CHKERRQ(ierr); 2710 2711 /* local sweep */ 2712 ierr = (*mat->A->ops->sor)(mat->A,bb1,omega,SOR_BACKWARD_SWEEP,fshift,lits,1,xx);CHKERRQ(ierr); 2713 } 2714 } else SETERRQ(PetscObjectComm((PetscObject)matin),PETSC_ERR_SUP,"Parallel version of SOR requested not supported"); 2715 2716 ierr = VecDestroy(&bb1);CHKERRQ(ierr); 2717 PetscFunctionReturn(0); 2718 } 2719 2720 extern PetscErrorCode MatFDColoringApply_BAIJ(Mat,MatFDColoring,Vec,MatStructure*,void*); 2721 2722 #undef __FUNCT__ 2723 #define __FUNCT__ "MatGetColumnNorms_MPIBAIJ" 2724 PetscErrorCode MatGetColumnNorms_MPIBAIJ(Mat A,NormType type,PetscReal *norms) 2725 { 2726 PetscErrorCode ierr; 2727 Mat_MPIBAIJ *aij = (Mat_MPIBAIJ*)A->data; 2728 PetscInt N,i,*garray = aij->garray; 2729 PetscInt ib,jb,bs = A->rmap->bs; 2730 Mat_SeqBAIJ *a_aij = (Mat_SeqBAIJ*) aij->A->data; 2731 MatScalar *a_val = a_aij->a; 2732 Mat_SeqBAIJ *b_aij = (Mat_SeqBAIJ*) aij->B->data; 2733 MatScalar *b_val = b_aij->a; 2734 PetscReal *work; 2735 2736 PetscFunctionBegin; 2737 ierr = MatGetSize(A,NULL,&N);CHKERRQ(ierr); 2738 ierr = PetscMalloc(N*sizeof(PetscReal),&work);CHKERRQ(ierr); 2739 ierr = PetscMemzero(work,N*sizeof(PetscReal));CHKERRQ(ierr); 2740 if (type == NORM_2) { 2741 for (i=a_aij->i[0]; i<a_aij->i[aij->A->rmap->n/bs]; i++) { 2742 for (jb=0; jb<bs; jb++) { 2743 for (ib=0; ib<bs; ib++) { 2744 work[A->cmap->rstart + a_aij->j[i] * bs + jb] += PetscAbsScalar(*a_val * *a_val); 2745 a_val++; 2746 } 2747 } 2748 } 2749 for (i=b_aij->i[0]; i<b_aij->i[aij->B->rmap->n/bs]; i++) { 2750 for (jb=0; jb<bs; jb++) { 2751 for (ib=0; ib<bs; ib++) { 2752 work[garray[b_aij->j[i]] * bs + jb] += PetscAbsScalar(*b_val * *b_val); 2753 b_val++; 2754 } 2755 } 2756 } 2757 } else if (type == NORM_1) { 2758 for (i=a_aij->i[0]; i<a_aij->i[aij->A->rmap->n/bs]; i++) { 2759 for (jb=0; jb<bs; jb++) { 2760 for (ib=0; ib<bs; ib++) { 2761 work[A->cmap->rstart + a_aij->j[i] * bs + jb] += PetscAbsScalar(*a_val); 2762 a_val++; 2763 } 2764 } 2765 } 2766 for (i=b_aij->i[0]; i<b_aij->i[aij->B->rmap->n/bs]; i++) { 2767 for (jb=0; jb<bs; jb++) { 2768 for (ib=0; ib<bs; ib++) { 2769 work[garray[b_aij->j[i]] * bs + jb] += PetscAbsScalar(*b_val); 2770 b_val++; 2771 } 2772 } 2773 } 2774 } else if (type == NORM_INFINITY) { 2775 for (i=a_aij->i[0]; i<a_aij->i[aij->A->rmap->n/bs]; i++) { 2776 for (jb=0; jb<bs; jb++) { 2777 for (ib=0; ib<bs; ib++) { 2778 int col = A->cmap->rstart + a_aij->j[i] * bs + jb; 2779 work[col] = PetscMax(PetscAbsScalar(*a_val), work[col]); 2780 a_val++; 2781 } 2782 } 2783 } 2784 for (i=b_aij->i[0]; i<b_aij->i[aij->B->rmap->n/bs]; i++) { 2785 for (jb=0; jb<bs; jb++) { 2786 for (ib=0; ib<bs; ib++) { 2787 int col = garray[b_aij->j[i]] * bs + jb; 2788 work[col] = PetscMax(PetscAbsScalar(*b_val), work[col]); 2789 b_val++; 2790 } 2791 } 2792 } 2793 } else SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_ARG_WRONG,"Unknown NormType"); 2794 if (type == NORM_INFINITY) { 2795 ierr = MPI_Allreduce(work,norms,N,MPIU_REAL,MPIU_MAX,PetscObjectComm((PetscObject)A));CHKERRQ(ierr); 2796 } else { 2797 ierr = MPI_Allreduce(work,norms,N,MPIU_REAL,MPIU_SUM,PetscObjectComm((PetscObject)A));CHKERRQ(ierr); 2798 } 2799 ierr = PetscFree(work);CHKERRQ(ierr); 2800 if (type == NORM_2) { 2801 for (i=0; i<N; i++) norms[i] = PetscSqrtReal(norms[i]); 2802 } 2803 PetscFunctionReturn(0); 2804 } 2805 2806 #undef __FUNCT__ 2807 #define __FUNCT__ "MatInvertBlockDiagonal_MPIBAIJ" 2808 PetscErrorCode MatInvertBlockDiagonal_MPIBAIJ(Mat A,const PetscScalar **values) 2809 { 2810 Mat_MPIBAIJ *a = (Mat_MPIBAIJ*) A->data; 2811 PetscErrorCode ierr; 2812 2813 PetscFunctionBegin; 2814 ierr = MatInvertBlockDiagonal(a->A,values);CHKERRQ(ierr); 2815 PetscFunctionReturn(0); 2816 } 2817 2818 2819 /* -------------------------------------------------------------------*/ 2820 static struct _MatOps MatOps_Values = {MatSetValues_MPIBAIJ, 2821 MatGetRow_MPIBAIJ, 2822 MatRestoreRow_MPIBAIJ, 2823 MatMult_MPIBAIJ, 2824 /* 4*/ MatMultAdd_MPIBAIJ, 2825 MatMultTranspose_MPIBAIJ, 2826 MatMultTransposeAdd_MPIBAIJ, 2827 0, 2828 0, 2829 0, 2830 /*10*/ 0, 2831 0, 2832 0, 2833 MatSOR_MPIBAIJ, 2834 MatTranspose_MPIBAIJ, 2835 /*15*/ MatGetInfo_MPIBAIJ, 2836 MatEqual_MPIBAIJ, 2837 MatGetDiagonal_MPIBAIJ, 2838 MatDiagonalScale_MPIBAIJ, 2839 MatNorm_MPIBAIJ, 2840 /*20*/ MatAssemblyBegin_MPIBAIJ, 2841 MatAssemblyEnd_MPIBAIJ, 2842 MatSetOption_MPIBAIJ, 2843 MatZeroEntries_MPIBAIJ, 2844 /*24*/ MatZeroRows_MPIBAIJ, 2845 0, 2846 0, 2847 0, 2848 0, 2849 /*29*/ MatSetUp_MPIBAIJ, 2850 0, 2851 0, 2852 0, 2853 0, 2854 /*34*/ MatDuplicate_MPIBAIJ, 2855 0, 2856 0, 2857 0, 2858 0, 2859 /*39*/ MatAXPY_MPIBAIJ, 2860 MatGetSubMatrices_MPIBAIJ, 2861 MatIncreaseOverlap_MPIBAIJ, 2862 MatGetValues_MPIBAIJ, 2863 MatCopy_MPIBAIJ, 2864 /*44*/ 0, 2865 MatScale_MPIBAIJ, 2866 0, 2867 0, 2868 0, 2869 /*49*/ 0, 2870 0, 2871 0, 2872 0, 2873 0, 2874 /*54*/ MatFDColoringCreate_MPIBAIJ, 2875 0, 2876 MatSetUnfactored_MPIBAIJ, 2877 MatPermute_MPIBAIJ, 2878 MatSetValuesBlocked_MPIBAIJ, 2879 /*59*/ MatGetSubMatrix_MPIBAIJ, 2880 MatDestroy_MPIBAIJ, 2881 MatView_MPIBAIJ, 2882 0, 2883 0, 2884 /*64*/ 0, 2885 0, 2886 0, 2887 0, 2888 0, 2889 /*69*/ MatGetRowMaxAbs_MPIBAIJ, 2890 0, 2891 0, 2892 0, 2893 0, 2894 /*74*/ 0, 2895 MatFDColoringApply_BAIJ, 2896 0, 2897 0, 2898 0, 2899 /*79*/ 0, 2900 0, 2901 0, 2902 0, 2903 MatLoad_MPIBAIJ, 2904 /*84*/ 0, 2905 0, 2906 0, 2907 0, 2908 0, 2909 /*89*/ 0, 2910 0, 2911 0, 2912 0, 2913 0, 2914 /*94*/ 0, 2915 0, 2916 0, 2917 0, 2918 0, 2919 /*99*/ 0, 2920 0, 2921 0, 2922 0, 2923 0, 2924 /*104*/0, 2925 MatRealPart_MPIBAIJ, 2926 MatImaginaryPart_MPIBAIJ, 2927 0, 2928 0, 2929 /*109*/0, 2930 0, 2931 0, 2932 0, 2933 0, 2934 /*114*/MatGetSeqNonzeroStructure_MPIBAIJ, 2935 0, 2936 MatGetGhosts_MPIBAIJ, 2937 0, 2938 0, 2939 /*119*/0, 2940 0, 2941 0, 2942 0, 2943 MatGetMultiProcBlock_MPIBAIJ, 2944 /*124*/0, 2945 MatGetColumnNorms_MPIBAIJ, 2946 MatInvertBlockDiagonal_MPIBAIJ, 2947 0, 2948 0, 2949 /*129*/ 0, 2950 0, 2951 0, 2952 0, 2953 0, 2954 /*134*/ 0, 2955 0, 2956 0, 2957 0, 2958 0, 2959 /*139*/ 0, 2960 0, 2961 0 2962 }; 2963 2964 #undef __FUNCT__ 2965 #define __FUNCT__ "MatGetDiagonalBlock_MPIBAIJ" 2966 PetscErrorCode MatGetDiagonalBlock_MPIBAIJ(Mat A,Mat *a) 2967 { 2968 PetscFunctionBegin; 2969 *a = ((Mat_MPIBAIJ*)A->data)->A; 2970 PetscFunctionReturn(0); 2971 } 2972 2973 PETSC_EXTERN PetscErrorCode MatConvert_MPIBAIJ_MPISBAIJ(Mat, MatType,MatReuse,Mat*); 2974 2975 #undef __FUNCT__ 2976 #define __FUNCT__ "MatMPIBAIJSetPreallocationCSR_MPIBAIJ" 2977 PetscErrorCode MatMPIBAIJSetPreallocationCSR_MPIBAIJ(Mat B,PetscInt bs,const PetscInt ii[],const PetscInt jj[],const PetscScalar V[]) 2978 { 2979 PetscInt m,rstart,cstart,cend; 2980 PetscInt i,j,d,nz,nz_max=0,*d_nnz=0,*o_nnz=0; 2981 const PetscInt *JJ =0; 2982 PetscScalar *values=0; 2983 PetscBool roworiented = ((Mat_MPIBAIJ*)B->data)->roworiented; 2984 PetscErrorCode ierr; 2985 2986 PetscFunctionBegin; 2987 ierr = PetscLayoutSetBlockSize(B->rmap,bs);CHKERRQ(ierr); 2988 ierr = PetscLayoutSetBlockSize(B->cmap,bs);CHKERRQ(ierr); 2989 ierr = PetscLayoutSetUp(B->rmap);CHKERRQ(ierr); 2990 ierr = PetscLayoutSetUp(B->cmap);CHKERRQ(ierr); 2991 ierr = PetscLayoutGetBlockSize(B->rmap,&bs);CHKERRQ(ierr); 2992 m = B->rmap->n/bs; 2993 rstart = B->rmap->rstart/bs; 2994 cstart = B->cmap->rstart/bs; 2995 cend = B->cmap->rend/bs; 2996 2997 if (ii[0]) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"ii[0] must be 0 but it is %D",ii[0]); 2998 ierr = PetscMalloc2(m,PetscInt,&d_nnz,m,PetscInt,&o_nnz);CHKERRQ(ierr); 2999 for (i=0; i<m; i++) { 3000 nz = ii[i+1] - ii[i]; 3001 if (nz < 0) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Local row %D has a negative number of columns %D",i,nz); 3002 nz_max = PetscMax(nz_max,nz); 3003 JJ = jj + ii[i]; 3004 for (j=0; j<nz; j++) { 3005 if (*JJ >= cstart) break; 3006 JJ++; 3007 } 3008 d = 0; 3009 for (; j<nz; j++) { 3010 if (*JJ++ >= cend) break; 3011 d++; 3012 } 3013 d_nnz[i] = d; 3014 o_nnz[i] = nz - d; 3015 } 3016 ierr = MatMPIBAIJSetPreallocation(B,bs,0,d_nnz,0,o_nnz);CHKERRQ(ierr); 3017 ierr = PetscFree2(d_nnz,o_nnz);CHKERRQ(ierr); 3018 3019 values = (PetscScalar*)V; 3020 if (!values) { 3021 ierr = PetscMalloc(bs*bs*nz_max*sizeof(PetscScalar),&values);CHKERRQ(ierr); 3022 ierr = PetscMemzero(values,bs*bs*nz_max*sizeof(PetscScalar));CHKERRQ(ierr); 3023 } 3024 for (i=0; i<m; i++) { 3025 PetscInt row = i + rstart; 3026 PetscInt ncols = ii[i+1] - ii[i]; 3027 const PetscInt *icols = jj + ii[i]; 3028 if (!roworiented) { /* block ordering matches the non-nested layout of MatSetValues so we can insert entire rows */ 3029 const PetscScalar *svals = values + (V ? (bs*bs*ii[i]) : 0); 3030 ierr = MatSetValuesBlocked_MPIBAIJ(B,1,&row,ncols,icols,svals,INSERT_VALUES);CHKERRQ(ierr); 3031 } else { /* block ordering does not match so we can only insert one block at a time. */ 3032 PetscInt j; 3033 for (j=0; j<ncols; j++) { 3034 const PetscScalar *svals = values + (V ? (bs*bs*(ii[i]+j)) : 0); 3035 ierr = MatSetValuesBlocked_MPIBAIJ(B,1,&row,1,&icols[j],svals,INSERT_VALUES);CHKERRQ(ierr); 3036 } 3037 } 3038 } 3039 3040 if (!V) { ierr = PetscFree(values);CHKERRQ(ierr); } 3041 ierr = MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3042 ierr = MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3043 ierr = MatSetOption(B,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_TRUE);CHKERRQ(ierr); 3044 PetscFunctionReturn(0); 3045 } 3046 3047 #undef __FUNCT__ 3048 #define __FUNCT__ "MatMPIBAIJSetPreallocationCSR" 3049 /*@C 3050 MatMPIBAIJSetPreallocationCSR - Allocates memory for a sparse parallel matrix in BAIJ format 3051 (the default parallel PETSc format). 3052 3053 Collective on MPI_Comm 3054 3055 Input Parameters: 3056 + A - the matrix 3057 . bs - the block size 3058 . i - the indices into j for the start of each local row (starts with zero) 3059 . j - the column indices for each local row (starts with zero) these must be sorted for each row 3060 - v - optional values in the matrix 3061 3062 Level: developer 3063 3064 Notes: The order of the entries in values is specified by the MatOption MAT_ROW_ORIENTED. For example, C programs 3065 may want to use the default MAT_ROW_ORIENTED=PETSC_TRUE and use an array v[nnz][bs][bs] where the second index is 3066 over rows within a block and the last index is over columns within a block row. Fortran programs will likely set 3067 MAT_ROW_ORIENTED=PETSC_FALSE and use a Fortran array v(bs,bs,nnz) in which the first index is over rows within a 3068 block column and the second index is over columns within a block. 3069 3070 .keywords: matrix, aij, compressed row, sparse, parallel 3071 3072 .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatMPIBAIJSetPreallocation(), MatCreateAIJ(), MPIAIJ, MatCreateMPIBAIJWithArrays(), MPIBAIJ 3073 @*/ 3074 PetscErrorCode MatMPIBAIJSetPreallocationCSR(Mat B,PetscInt bs,const PetscInt i[],const PetscInt j[], const PetscScalar v[]) 3075 { 3076 PetscErrorCode ierr; 3077 3078 PetscFunctionBegin; 3079 PetscValidHeaderSpecific(B,MAT_CLASSID,1); 3080 PetscValidType(B,1); 3081 PetscValidLogicalCollectiveInt(B,bs,2); 3082 ierr = PetscTryMethod(B,"MatMPIBAIJSetPreallocationCSR_C",(Mat,PetscInt,const PetscInt[],const PetscInt[],const PetscScalar[]),(B,bs,i,j,v));CHKERRQ(ierr); 3083 PetscFunctionReturn(0); 3084 } 3085 3086 #undef __FUNCT__ 3087 #define __FUNCT__ "MatMPIBAIJSetPreallocation_MPIBAIJ" 3088 PetscErrorCode MatMPIBAIJSetPreallocation_MPIBAIJ(Mat B,PetscInt bs,PetscInt d_nz,const PetscInt *d_nnz,PetscInt o_nz,const PetscInt *o_nnz) 3089 { 3090 Mat_MPIBAIJ *b; 3091 PetscErrorCode ierr; 3092 PetscInt i; 3093 3094 PetscFunctionBegin; 3095 ierr = PetscLayoutSetBlockSize(B->rmap,bs);CHKERRQ(ierr); 3096 ierr = PetscLayoutSetBlockSize(B->cmap,bs);CHKERRQ(ierr); 3097 ierr = PetscLayoutSetUp(B->rmap);CHKERRQ(ierr); 3098 ierr = PetscLayoutSetUp(B->cmap);CHKERRQ(ierr); 3099 ierr = PetscLayoutGetBlockSize(B->rmap,&bs);CHKERRQ(ierr); 3100 3101 if (d_nnz) { 3102 for (i=0; i<B->rmap->n/bs; i++) { 3103 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]); 3104 } 3105 } 3106 if (o_nnz) { 3107 for (i=0; i<B->rmap->n/bs; i++) { 3108 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]); 3109 } 3110 } 3111 3112 b = (Mat_MPIBAIJ*)B->data; 3113 b->bs2 = bs*bs; 3114 b->mbs = B->rmap->n/bs; 3115 b->nbs = B->cmap->n/bs; 3116 b->Mbs = B->rmap->N/bs; 3117 b->Nbs = B->cmap->N/bs; 3118 3119 for (i=0; i<=b->size; i++) { 3120 b->rangebs[i] = B->rmap->range[i]/bs; 3121 } 3122 b->rstartbs = B->rmap->rstart/bs; 3123 b->rendbs = B->rmap->rend/bs; 3124 b->cstartbs = B->cmap->rstart/bs; 3125 b->cendbs = B->cmap->rend/bs; 3126 3127 if (!B->preallocated) { 3128 ierr = MatCreate(PETSC_COMM_SELF,&b->A);CHKERRQ(ierr); 3129 ierr = MatSetSizes(b->A,B->rmap->n,B->cmap->n,B->rmap->n,B->cmap->n);CHKERRQ(ierr); 3130 ierr = MatSetType(b->A,MATSEQBAIJ);CHKERRQ(ierr); 3131 ierr = PetscLogObjectParent((PetscObject)B,(PetscObject)b->A);CHKERRQ(ierr); 3132 ierr = MatCreate(PETSC_COMM_SELF,&b->B);CHKERRQ(ierr); 3133 ierr = MatSetSizes(b->B,B->rmap->n,B->cmap->N,B->rmap->n,B->cmap->N);CHKERRQ(ierr); 3134 ierr = MatSetType(b->B,MATSEQBAIJ);CHKERRQ(ierr); 3135 ierr = PetscLogObjectParent((PetscObject)B,(PetscObject)b->B);CHKERRQ(ierr); 3136 ierr = MatStashCreate_Private(PetscObjectComm((PetscObject)B),bs,&B->bstash);CHKERRQ(ierr); 3137 } 3138 3139 ierr = MatSeqBAIJSetPreallocation(b->A,bs,d_nz,d_nnz);CHKERRQ(ierr); 3140 ierr = MatSeqBAIJSetPreallocation(b->B,bs,o_nz,o_nnz);CHKERRQ(ierr); 3141 B->preallocated = PETSC_TRUE; 3142 PetscFunctionReturn(0); 3143 } 3144 3145 extern PetscErrorCode MatDiagonalScaleLocal_MPIBAIJ(Mat,Vec); 3146 extern PetscErrorCode MatSetHashTableFactor_MPIBAIJ(Mat,PetscReal); 3147 3148 #undef __FUNCT__ 3149 #define __FUNCT__ "MatConvert_MPIBAIJ_MPIAdj" 3150 PETSC_EXTERN PetscErrorCode MatConvert_MPIBAIJ_MPIAdj(Mat B, MatType newtype,MatReuse reuse,Mat *adj) 3151 { 3152 Mat_MPIBAIJ *b = (Mat_MPIBAIJ*)B->data; 3153 PetscErrorCode ierr; 3154 Mat_SeqBAIJ *d = (Mat_SeqBAIJ*) b->A->data,*o = (Mat_SeqBAIJ*) b->B->data; 3155 PetscInt M = B->rmap->n/B->rmap->bs,i,*ii,*jj,cnt,j,k,rstart = B->rmap->rstart/B->rmap->bs; 3156 const PetscInt *id = d->i, *jd = d->j, *io = o->i, *jo = o->j, *garray = b->garray; 3157 3158 PetscFunctionBegin; 3159 ierr = PetscMalloc((M+1)*sizeof(PetscInt),&ii);CHKERRQ(ierr); 3160 ii[0] = 0; 3161 for (i=0; i<M; i++) { 3162 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]); 3163 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]); 3164 ii[i+1] = ii[i] + id[i+1] - id[i] + io[i+1] - io[i]; 3165 /* remove one from count of matrix has diagonal */ 3166 for (j=id[i]; j<id[i+1]; j++) { 3167 if (jd[j] == i) {ii[i+1]--;break;} 3168 } 3169 } 3170 ierr = PetscMalloc(ii[M]*sizeof(PetscInt),&jj);CHKERRQ(ierr); 3171 cnt = 0; 3172 for (i=0; i<M; i++) { 3173 for (j=io[i]; j<io[i+1]; j++) { 3174 if (garray[jo[j]] > rstart) break; 3175 jj[cnt++] = garray[jo[j]]; 3176 } 3177 for (k=id[i]; k<id[i+1]; k++) { 3178 if (jd[k] != i) { 3179 jj[cnt++] = rstart + jd[k]; 3180 } 3181 } 3182 for (; j<io[i+1]; j++) { 3183 jj[cnt++] = garray[jo[j]]; 3184 } 3185 } 3186 ierr = MatCreateMPIAdj(PetscObjectComm((PetscObject)B),M,B->cmap->N/B->rmap->bs,ii,jj,NULL,adj);CHKERRQ(ierr); 3187 PetscFunctionReturn(0); 3188 } 3189 3190 #include <../src/mat/impls/aij/mpi/mpiaij.h> 3191 3192 PETSC_EXTERN PetscErrorCode MatConvert_SeqBAIJ_SeqAIJ(Mat,MatType,MatReuse,Mat*); 3193 3194 #undef __FUNCT__ 3195 #define __FUNCT__ "MatConvert_MPIBAIJ_MPIAIJ" 3196 PETSC_EXTERN PetscErrorCode MatConvert_MPIBAIJ_MPIAIJ(Mat A,MatType newtype,MatReuse reuse,Mat *newmat) 3197 { 3198 PetscErrorCode ierr; 3199 Mat_MPIBAIJ *a = (Mat_MPIBAIJ*)A->data; 3200 Mat B; 3201 Mat_MPIAIJ *b; 3202 3203 PetscFunctionBegin; 3204 if (!A->assembled) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_SUP,"Matrix must be assembled"); 3205 3206 ierr = MatCreate(PetscObjectComm((PetscObject)A),&B);CHKERRQ(ierr); 3207 ierr = MatSetSizes(B,A->rmap->n,A->cmap->n,A->rmap->N,A->cmap->N);CHKERRQ(ierr); 3208 ierr = MatSetType(B,MATMPIAIJ);CHKERRQ(ierr); 3209 ierr = MatSeqAIJSetPreallocation(B,0,NULL);CHKERRQ(ierr); 3210 ierr = MatMPIAIJSetPreallocation(B,0,NULL,0,NULL);CHKERRQ(ierr); 3211 b = (Mat_MPIAIJ*) B->data; 3212 3213 ierr = MatDestroy(&b->A);CHKERRQ(ierr); 3214 ierr = MatDestroy(&b->B);CHKERRQ(ierr); 3215 ierr = MatDisAssemble_MPIBAIJ(A);CHKERRQ(ierr); 3216 ierr = MatConvert_SeqBAIJ_SeqAIJ(a->A, MATSEQAIJ, MAT_INITIAL_MATRIX, &b->A);CHKERRQ(ierr); 3217 ierr = MatConvert_SeqBAIJ_SeqAIJ(a->B, MATSEQAIJ, MAT_INITIAL_MATRIX, &b->B);CHKERRQ(ierr); 3218 ierr = MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3219 ierr = MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3220 ierr = MatAssemblyBegin(A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3221 ierr = MatAssemblyEnd(A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3222 if (reuse == MAT_REUSE_MATRIX) { 3223 ierr = MatHeaderReplace(A,B);CHKERRQ(ierr); 3224 } else { 3225 *newmat = B; 3226 } 3227 PetscFunctionReturn(0); 3228 } 3229 3230 #if defined(PETSC_HAVE_MUMPS) 3231 PETSC_EXTERN PetscErrorCode MatGetFactor_baij_mumps(Mat,MatFactorType,Mat*); 3232 #endif 3233 3234 /*MC 3235 MATMPIBAIJ - MATMPIBAIJ = "mpibaij" - A matrix type to be used for distributed block sparse matrices. 3236 3237 Options Database Keys: 3238 + -mat_type mpibaij - sets the matrix type to "mpibaij" during a call to MatSetFromOptions() 3239 . -mat_block_size <bs> - set the blocksize used to store the matrix 3240 - -mat_use_hash_table <fact> 3241 3242 Level: beginner 3243 3244 .seealso: MatCreateMPIBAIJ 3245 M*/ 3246 3247 PETSC_EXTERN PetscErrorCode MatConvert_MPIBAIJ_MPIBSTRM(Mat,MatType,MatReuse,Mat*); 3248 3249 #undef __FUNCT__ 3250 #define __FUNCT__ "MatCreate_MPIBAIJ" 3251 PETSC_EXTERN PetscErrorCode MatCreate_MPIBAIJ(Mat B) 3252 { 3253 Mat_MPIBAIJ *b; 3254 PetscErrorCode ierr; 3255 PetscBool flg; 3256 3257 PetscFunctionBegin; 3258 ierr = PetscNewLog(B,Mat_MPIBAIJ,&b);CHKERRQ(ierr); 3259 B->data = (void*)b; 3260 3261 ierr = PetscMemcpy(B->ops,&MatOps_Values,sizeof(struct _MatOps));CHKERRQ(ierr); 3262 B->assembled = PETSC_FALSE; 3263 3264 B->insertmode = NOT_SET_VALUES; 3265 ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)B),&b->rank);CHKERRQ(ierr); 3266 ierr = MPI_Comm_size(PetscObjectComm((PetscObject)B),&b->size);CHKERRQ(ierr); 3267 3268 /* build local table of row and column ownerships */ 3269 ierr = PetscMalloc((b->size+1)*sizeof(PetscInt),&b->rangebs);CHKERRQ(ierr); 3270 3271 /* build cache for off array entries formed */ 3272 ierr = MatStashCreate_Private(PetscObjectComm((PetscObject)B),1,&B->stash);CHKERRQ(ierr); 3273 3274 b->donotstash = PETSC_FALSE; 3275 b->colmap = NULL; 3276 b->garray = NULL; 3277 b->roworiented = PETSC_TRUE; 3278 3279 /* stuff used in block assembly */ 3280 b->barray = 0; 3281 3282 /* stuff used for matrix vector multiply */ 3283 b->lvec = 0; 3284 b->Mvctx = 0; 3285 3286 /* stuff for MatGetRow() */ 3287 b->rowindices = 0; 3288 b->rowvalues = 0; 3289 b->getrowactive = PETSC_FALSE; 3290 3291 /* hash table stuff */ 3292 b->ht = 0; 3293 b->hd = 0; 3294 b->ht_size = 0; 3295 b->ht_flag = PETSC_FALSE; 3296 b->ht_fact = 0; 3297 b->ht_total_ct = 0; 3298 b->ht_insert_ct = 0; 3299 3300 /* stuff for MatGetSubMatrices_MPIBAIJ_local() */ 3301 b->ijonly = PETSC_FALSE; 3302 3303 ierr = PetscOptionsBegin(PetscObjectComm((PetscObject)B),NULL,"Options for loading MPIBAIJ matrix 1","Mat");CHKERRQ(ierr); 3304 ierr = PetscOptionsBool("-mat_use_hash_table","Use hash table to save memory in constructing matrix","MatSetOption",PETSC_FALSE,&flg,NULL);CHKERRQ(ierr); 3305 if (flg) { 3306 PetscReal fact = 1.39; 3307 ierr = MatSetOption(B,MAT_USE_HASH_TABLE,PETSC_TRUE);CHKERRQ(ierr); 3308 ierr = PetscOptionsReal("-mat_use_hash_table","Use hash table factor","MatMPIBAIJSetHashTableFactor",fact,&fact,NULL);CHKERRQ(ierr); 3309 if (fact <= 1.0) fact = 1.39; 3310 ierr = MatMPIBAIJSetHashTableFactor(B,fact);CHKERRQ(ierr); 3311 ierr = PetscInfo1(B,"Hash table Factor used %5.2f\n",fact);CHKERRQ(ierr); 3312 } 3313 ierr = PetscOptionsEnd();CHKERRQ(ierr); 3314 3315 #if defined(PETSC_HAVE_MUMPS) 3316 ierr = PetscObjectComposeFunction((PetscObject)B,"MatGetFactor_mumps_C",MatGetFactor_baij_mumps);CHKERRQ(ierr); 3317 #endif 3318 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpibaij_mpiadj_C",MatConvert_MPIBAIJ_MPIAdj);CHKERRQ(ierr); 3319 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpibaij_mpiaij_C",MatConvert_MPIBAIJ_MPIAIJ);CHKERRQ(ierr); 3320 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpibaij_mpisbaij_C",MatConvert_MPIBAIJ_MPISBAIJ);CHKERRQ(ierr); 3321 ierr = PetscObjectComposeFunction((PetscObject)B,"MatStoreValues_C",MatStoreValues_MPIBAIJ);CHKERRQ(ierr); 3322 ierr = PetscObjectComposeFunction((PetscObject)B,"MatRetrieveValues_C",MatRetrieveValues_MPIBAIJ);CHKERRQ(ierr); 3323 ierr = PetscObjectComposeFunction((PetscObject)B,"MatGetDiagonalBlock_C",MatGetDiagonalBlock_MPIBAIJ);CHKERRQ(ierr); 3324 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMPIBAIJSetPreallocation_C",MatMPIBAIJSetPreallocation_MPIBAIJ);CHKERRQ(ierr); 3325 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMPIBAIJSetPreallocationCSR_C",MatMPIBAIJSetPreallocationCSR_MPIBAIJ);CHKERRQ(ierr); 3326 ierr = PetscObjectComposeFunction((PetscObject)B,"MatDiagonalScaleLocal_C",MatDiagonalScaleLocal_MPIBAIJ);CHKERRQ(ierr); 3327 ierr = PetscObjectComposeFunction((PetscObject)B,"MatSetHashTableFactor_C",MatSetHashTableFactor_MPIBAIJ);CHKERRQ(ierr); 3328 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpibaij_mpibstrm_C",MatConvert_MPIBAIJ_MPIBSTRM);CHKERRQ(ierr); 3329 ierr = PetscObjectChangeTypeName((PetscObject)B,MATMPIBAIJ);CHKERRQ(ierr); 3330 PetscFunctionReturn(0); 3331 } 3332 3333 /*MC 3334 MATBAIJ - MATBAIJ = "baij" - A matrix type to be used for block sparse matrices. 3335 3336 This matrix type is identical to MATSEQBAIJ when constructed with a single process communicator, 3337 and MATMPIBAIJ otherwise. 3338 3339 Options Database Keys: 3340 . -mat_type baij - sets the matrix type to "baij" during a call to MatSetFromOptions() 3341 3342 Level: beginner 3343 3344 .seealso: MatCreateBAIJ(),MATSEQBAIJ,MATMPIBAIJ, MatMPIBAIJSetPreallocation(), MatMPIBAIJSetPreallocationCSR() 3345 M*/ 3346 3347 #undef __FUNCT__ 3348 #define __FUNCT__ "MatMPIBAIJSetPreallocation" 3349 /*@C 3350 MatMPIBAIJSetPreallocation - Allocates memory for a sparse parallel matrix in block AIJ format 3351 (block compressed row). For good matrix assembly performance 3352 the user should preallocate the matrix storage by setting the parameters 3353 d_nz (or d_nnz) and o_nz (or o_nnz). By setting these parameters accurately, 3354 performance can be increased by more than a factor of 50. 3355 3356 Collective on Mat 3357 3358 Input Parameters: 3359 + A - the matrix 3360 . bs - size of block 3361 . d_nz - number of block nonzeros per block row in diagonal portion of local 3362 submatrix (same for all local rows) 3363 . d_nnz - array containing the number of block nonzeros in the various block rows 3364 of the in diagonal portion of the local (possibly different for each block 3365 row) or NULL. If you plan to factor the matrix you must leave room for the diagonal entry and 3366 set it even if it is zero. 3367 . o_nz - number of block nonzeros per block row in the off-diagonal portion of local 3368 submatrix (same for all local rows). 3369 - o_nnz - array containing the number of nonzeros in the various block rows of the 3370 off-diagonal portion of the local submatrix (possibly different for 3371 each block row) or NULL. 3372 3373 If the *_nnz parameter is given then the *_nz parameter is ignored 3374 3375 Options Database Keys: 3376 + -mat_block_size - size of the blocks to use 3377 - -mat_use_hash_table <fact> 3378 3379 Notes: 3380 If PETSC_DECIDE or PETSC_DETERMINE is used for a particular argument on one processor 3381 than it must be used on all processors that share the object for that argument. 3382 3383 Storage Information: 3384 For a square global matrix we define each processor's diagonal portion 3385 to be its local rows and the corresponding columns (a square submatrix); 3386 each processor's off-diagonal portion encompasses the remainder of the 3387 local matrix (a rectangular submatrix). 3388 3389 The user can specify preallocated storage for the diagonal part of 3390 the local submatrix with either d_nz or d_nnz (not both). Set 3391 d_nz=PETSC_DEFAULT and d_nnz=NULL for PETSc to control dynamic 3392 memory allocation. Likewise, specify preallocated storage for the 3393 off-diagonal part of the local submatrix with o_nz or o_nnz (not both). 3394 3395 Consider a processor that owns rows 3, 4 and 5 of a parallel matrix. In 3396 the figure below we depict these three local rows and all columns (0-11). 3397 3398 .vb 3399 0 1 2 3 4 5 6 7 8 9 10 11 3400 -------------------------- 3401 row 3 |o o o d d d o o o o o o 3402 row 4 |o o o d d d o o o o o o 3403 row 5 |o o o d d d o o o o o o 3404 -------------------------- 3405 .ve 3406 3407 Thus, any entries in the d locations are stored in the d (diagonal) 3408 submatrix, and any entries in the o locations are stored in the 3409 o (off-diagonal) submatrix. Note that the d and the o submatrices are 3410 stored simply in the MATSEQBAIJ format for compressed row storage. 3411 3412 Now d_nz should indicate the number of block nonzeros per row in the d matrix, 3413 and o_nz should indicate the number of block nonzeros per row in the o matrix. 3414 In general, for PDE problems in which most nonzeros are near the diagonal, 3415 one expects d_nz >> o_nz. For large problems you MUST preallocate memory 3416 or you will get TERRIBLE performance; see the users' manual chapter on 3417 matrices. 3418 3419 You can call MatGetInfo() to get information on how effective the preallocation was; 3420 for example the fields mallocs,nz_allocated,nz_used,nz_unneeded; 3421 You can also run with the option -info and look for messages with the string 3422 malloc in them to see if additional memory allocation was needed. 3423 3424 Level: intermediate 3425 3426 .keywords: matrix, block, aij, compressed row, sparse, parallel 3427 3428 .seealso: MatCreate(), MatCreateSeqBAIJ(), MatSetValues(), MatCreateBAIJ(), MatMPIBAIJSetPreallocationCSR(), PetscSplitOwnership() 3429 @*/ 3430 PetscErrorCode MatMPIBAIJSetPreallocation(Mat B,PetscInt bs,PetscInt d_nz,const PetscInt d_nnz[],PetscInt o_nz,const PetscInt o_nnz[]) 3431 { 3432 PetscErrorCode ierr; 3433 3434 PetscFunctionBegin; 3435 PetscValidHeaderSpecific(B,MAT_CLASSID,1); 3436 PetscValidType(B,1); 3437 PetscValidLogicalCollectiveInt(B,bs,2); 3438 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); 3439 PetscFunctionReturn(0); 3440 } 3441 3442 #undef __FUNCT__ 3443 #define __FUNCT__ "MatCreateBAIJ" 3444 /*@C 3445 MatCreateBAIJ - Creates a sparse parallel matrix in block AIJ format 3446 (block compressed row). For good matrix assembly performance 3447 the user should preallocate the matrix storage by setting the parameters 3448 d_nz (or d_nnz) and o_nz (or o_nnz). By setting these parameters accurately, 3449 performance can be increased by more than a factor of 50. 3450 3451 Collective on MPI_Comm 3452 3453 Input Parameters: 3454 + comm - MPI communicator 3455 . bs - size of blockk 3456 . m - number of local rows (or PETSC_DECIDE to have calculated if M is given) 3457 This value should be the same as the local size used in creating the 3458 y vector for the matrix-vector product y = Ax. 3459 . n - number of local columns (or PETSC_DECIDE to have calculated if N is given) 3460 This value should be the same as the local size used in creating the 3461 x vector for the matrix-vector product y = Ax. 3462 . M - number of global rows (or PETSC_DETERMINE to have calculated if m is given) 3463 . N - number of global columns (or PETSC_DETERMINE to have calculated if n is given) 3464 . d_nz - number of nonzero blocks per block row in diagonal portion of local 3465 submatrix (same for all local rows) 3466 . d_nnz - array containing the number of nonzero blocks in the various block rows 3467 of the in diagonal portion of the local (possibly different for each block 3468 row) or NULL. If you plan to factor the matrix you must leave room for the diagonal entry 3469 and set it even if it is zero. 3470 . o_nz - number of nonzero blocks per block row in the off-diagonal portion of local 3471 submatrix (same for all local rows). 3472 - o_nnz - array containing the number of nonzero blocks in the various block rows of the 3473 off-diagonal portion of the local submatrix (possibly different for 3474 each block row) or NULL. 3475 3476 Output Parameter: 3477 . A - the matrix 3478 3479 Options Database Keys: 3480 + -mat_block_size - size of the blocks to use 3481 - -mat_use_hash_table <fact> 3482 3483 It is recommended that one use the MatCreate(), MatSetType() and/or MatSetFromOptions(), 3484 MatXXXXSetPreallocation() paradgm instead of this routine directly. 3485 [MatXXXXSetPreallocation() is, for example, MatSeqAIJSetPreallocation] 3486 3487 Notes: 3488 If the *_nnz parameter is given then the *_nz parameter is ignored 3489 3490 A nonzero block is any block that as 1 or more nonzeros in it 3491 3492 The user MUST specify either the local or global matrix dimensions 3493 (possibly both). 3494 3495 If PETSC_DECIDE or PETSC_DETERMINE is used for a particular argument on one processor 3496 than it must be used on all processors that share the object for that argument. 3497 3498 Storage Information: 3499 For a square global matrix we define each processor's diagonal portion 3500 to be its local rows and the corresponding columns (a square submatrix); 3501 each processor's off-diagonal portion encompasses the remainder of the 3502 local matrix (a rectangular submatrix). 3503 3504 The user can specify preallocated storage for the diagonal part of 3505 the local submatrix with either d_nz or d_nnz (not both). Set 3506 d_nz=PETSC_DEFAULT and d_nnz=NULL for PETSc to control dynamic 3507 memory allocation. Likewise, specify preallocated storage for the 3508 off-diagonal part of the local submatrix with o_nz or o_nnz (not both). 3509 3510 Consider a processor that owns rows 3, 4 and 5 of a parallel matrix. In 3511 the figure below we depict these three local rows and all columns (0-11). 3512 3513 .vb 3514 0 1 2 3 4 5 6 7 8 9 10 11 3515 -------------------------- 3516 row 3 |o o o d d d o o o o o o 3517 row 4 |o o o d d d o o o o o o 3518 row 5 |o o o d d d o o o o o o 3519 -------------------------- 3520 .ve 3521 3522 Thus, any entries in the d locations are stored in the d (diagonal) 3523 submatrix, and any entries in the o locations are stored in the 3524 o (off-diagonal) submatrix. Note that the d and the o submatrices are 3525 stored simply in the MATSEQBAIJ format for compressed row storage. 3526 3527 Now d_nz should indicate the number of block nonzeros per row in the d matrix, 3528 and o_nz should indicate the number of block nonzeros per row in the o matrix. 3529 In general, for PDE problems in which most nonzeros are near the diagonal, 3530 one expects d_nz >> o_nz. For large problems you MUST preallocate memory 3531 or you will get TERRIBLE performance; see the users' manual chapter on 3532 matrices. 3533 3534 Level: intermediate 3535 3536 .keywords: matrix, block, aij, compressed row, sparse, parallel 3537 3538 .seealso: MatCreate(), MatCreateSeqBAIJ(), MatSetValues(), MatCreateBAIJ(), MatMPIBAIJSetPreallocation(), MatMPIBAIJSetPreallocationCSR() 3539 @*/ 3540 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) 3541 { 3542 PetscErrorCode ierr; 3543 PetscMPIInt size; 3544 3545 PetscFunctionBegin; 3546 ierr = MatCreate(comm,A);CHKERRQ(ierr); 3547 ierr = MatSetSizes(*A,m,n,M,N);CHKERRQ(ierr); 3548 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 3549 if (size > 1) { 3550 ierr = MatSetType(*A,MATMPIBAIJ);CHKERRQ(ierr); 3551 ierr = MatMPIBAIJSetPreallocation(*A,bs,d_nz,d_nnz,o_nz,o_nnz);CHKERRQ(ierr); 3552 } else { 3553 ierr = MatSetType(*A,MATSEQBAIJ);CHKERRQ(ierr); 3554 ierr = MatSeqBAIJSetPreallocation(*A,bs,d_nz,d_nnz);CHKERRQ(ierr); 3555 } 3556 PetscFunctionReturn(0); 3557 } 3558 3559 #undef __FUNCT__ 3560 #define __FUNCT__ "MatDuplicate_MPIBAIJ" 3561 static PetscErrorCode MatDuplicate_MPIBAIJ(Mat matin,MatDuplicateOption cpvalues,Mat *newmat) 3562 { 3563 Mat mat; 3564 Mat_MPIBAIJ *a,*oldmat = (Mat_MPIBAIJ*)matin->data; 3565 PetscErrorCode ierr; 3566 PetscInt len=0; 3567 3568 PetscFunctionBegin; 3569 *newmat = 0; 3570 ierr = MatCreate(PetscObjectComm((PetscObject)matin),&mat);CHKERRQ(ierr); 3571 ierr = MatSetSizes(mat,matin->rmap->n,matin->cmap->n,matin->rmap->N,matin->cmap->N);CHKERRQ(ierr); 3572 ierr = MatSetType(mat,((PetscObject)matin)->type_name);CHKERRQ(ierr); 3573 ierr = PetscMemcpy(mat->ops,matin->ops,sizeof(struct _MatOps));CHKERRQ(ierr); 3574 3575 mat->factortype = matin->factortype; 3576 mat->preallocated = PETSC_TRUE; 3577 mat->assembled = PETSC_TRUE; 3578 mat->insertmode = NOT_SET_VALUES; 3579 3580 a = (Mat_MPIBAIJ*)mat->data; 3581 mat->rmap->bs = matin->rmap->bs; 3582 a->bs2 = oldmat->bs2; 3583 a->mbs = oldmat->mbs; 3584 a->nbs = oldmat->nbs; 3585 a->Mbs = oldmat->Mbs; 3586 a->Nbs = oldmat->Nbs; 3587 3588 ierr = PetscLayoutReference(matin->rmap,&mat->rmap);CHKERRQ(ierr); 3589 ierr = PetscLayoutReference(matin->cmap,&mat->cmap);CHKERRQ(ierr); 3590 3591 a->size = oldmat->size; 3592 a->rank = oldmat->rank; 3593 a->donotstash = oldmat->donotstash; 3594 a->roworiented = oldmat->roworiented; 3595 a->rowindices = 0; 3596 a->rowvalues = 0; 3597 a->getrowactive = PETSC_FALSE; 3598 a->barray = 0; 3599 a->rstartbs = oldmat->rstartbs; 3600 a->rendbs = oldmat->rendbs; 3601 a->cstartbs = oldmat->cstartbs; 3602 a->cendbs = oldmat->cendbs; 3603 3604 /* hash table stuff */ 3605 a->ht = 0; 3606 a->hd = 0; 3607 a->ht_size = 0; 3608 a->ht_flag = oldmat->ht_flag; 3609 a->ht_fact = oldmat->ht_fact; 3610 a->ht_total_ct = 0; 3611 a->ht_insert_ct = 0; 3612 3613 ierr = PetscMemcpy(a->rangebs,oldmat->rangebs,(a->size+1)*sizeof(PetscInt));CHKERRQ(ierr); 3614 if (oldmat->colmap) { 3615 #if defined(PETSC_USE_CTABLE) 3616 ierr = PetscTableCreateCopy(oldmat->colmap,&a->colmap);CHKERRQ(ierr); 3617 #else 3618 ierr = PetscMalloc((a->Nbs)*sizeof(PetscInt),&a->colmap);CHKERRQ(ierr); 3619 ierr = PetscLogObjectMemory((PetscObject)mat,(a->Nbs)*sizeof(PetscInt));CHKERRQ(ierr); 3620 ierr = PetscMemcpy(a->colmap,oldmat->colmap,(a->Nbs)*sizeof(PetscInt));CHKERRQ(ierr); 3621 #endif 3622 } else a->colmap = 0; 3623 3624 if (oldmat->garray && (len = ((Mat_SeqBAIJ*)(oldmat->B->data))->nbs)) { 3625 ierr = PetscMalloc(len*sizeof(PetscInt),&a->garray);CHKERRQ(ierr); 3626 ierr = PetscLogObjectMemory((PetscObject)mat,len*sizeof(PetscInt));CHKERRQ(ierr); 3627 ierr = PetscMemcpy(a->garray,oldmat->garray,len*sizeof(PetscInt));CHKERRQ(ierr); 3628 } else a->garray = 0; 3629 3630 ierr = MatStashCreate_Private(PetscObjectComm((PetscObject)matin),matin->rmap->bs,&mat->bstash);CHKERRQ(ierr); 3631 ierr = VecDuplicate(oldmat->lvec,&a->lvec);CHKERRQ(ierr); 3632 ierr = PetscLogObjectParent((PetscObject)mat,(PetscObject)a->lvec);CHKERRQ(ierr); 3633 ierr = VecScatterCopy(oldmat->Mvctx,&a->Mvctx);CHKERRQ(ierr); 3634 ierr = PetscLogObjectParent((PetscObject)mat,(PetscObject)a->Mvctx);CHKERRQ(ierr); 3635 3636 ierr = MatDuplicate(oldmat->A,cpvalues,&a->A);CHKERRQ(ierr); 3637 ierr = PetscLogObjectParent((PetscObject)mat,(PetscObject)a->A);CHKERRQ(ierr); 3638 ierr = MatDuplicate(oldmat->B,cpvalues,&a->B);CHKERRQ(ierr); 3639 ierr = PetscLogObjectParent((PetscObject)mat,(PetscObject)a->B);CHKERRQ(ierr); 3640 ierr = PetscFunctionListDuplicate(((PetscObject)matin)->qlist,&((PetscObject)mat)->qlist);CHKERRQ(ierr); 3641 *newmat = mat; 3642 PetscFunctionReturn(0); 3643 } 3644 3645 #undef __FUNCT__ 3646 #define __FUNCT__ "MatLoad_MPIBAIJ" 3647 PetscErrorCode MatLoad_MPIBAIJ(Mat newmat,PetscViewer viewer) 3648 { 3649 PetscErrorCode ierr; 3650 int fd; 3651 PetscInt i,nz,j,rstart,rend; 3652 PetscScalar *vals,*buf; 3653 MPI_Comm comm; 3654 MPI_Status status; 3655 PetscMPIInt rank,size,maxnz; 3656 PetscInt header[4],*rowlengths = 0,M,N,m,*rowners,*cols; 3657 PetscInt *locrowlens = NULL,*procsnz = NULL,*browners = NULL; 3658 PetscInt jj,*mycols,*ibuf,bs=1,Mbs,mbs,extra_rows,mmax; 3659 PetscMPIInt tag = ((PetscObject)viewer)->tag; 3660 PetscInt *dlens = NULL,*odlens = NULL,*mask = NULL,*masked1 = NULL,*masked2 = NULL,rowcount,odcount; 3661 PetscInt dcount,kmax,k,nzcount,tmp,mend,sizesset=1,grows,gcols; 3662 3663 PetscFunctionBegin; 3664 ierr = PetscObjectGetComm((PetscObject)viewer,&comm);CHKERRQ(ierr); 3665 ierr = PetscOptionsBegin(comm,NULL,"Options for loading MPIBAIJ matrix 2","Mat");CHKERRQ(ierr); 3666 ierr = PetscOptionsInt("-matload_block_size","Set the blocksize used to store the matrix","MatLoad",bs,&bs,NULL);CHKERRQ(ierr); 3667 ierr = PetscOptionsEnd();CHKERRQ(ierr); 3668 3669 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 3670 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 3671 if (!rank) { 3672 ierr = PetscViewerBinaryGetDescriptor(viewer,&fd);CHKERRQ(ierr); 3673 ierr = PetscBinaryRead(fd,(char*)header,4,PETSC_INT);CHKERRQ(ierr); 3674 if (header[0] != MAT_FILE_CLASSID) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED,"not matrix object"); 3675 } 3676 3677 if (newmat->rmap->n < 0 && newmat->rmap->N < 0 && newmat->cmap->n < 0 && newmat->cmap->N < 0) sizesset = 0; 3678 3679 ierr = MPI_Bcast(header+1,3,MPIU_INT,0,comm);CHKERRQ(ierr); 3680 M = header[1]; N = header[2]; 3681 3682 /* If global rows/cols are set to PETSC_DECIDE, set it to the sizes given in the file */ 3683 if (sizesset && newmat->rmap->N < 0) newmat->rmap->N = M; 3684 if (sizesset && newmat->cmap->N < 0) newmat->cmap->N = N; 3685 3686 /* If global sizes are set, check if they are consistent with that given in the file */ 3687 if (sizesset) { 3688 ierr = MatGetSize(newmat,&grows,&gcols);CHKERRQ(ierr); 3689 } 3690 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); 3691 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); 3692 3693 if (M != N) SETERRQ(PetscObjectComm((PetscObject)viewer),PETSC_ERR_SUP,"Can only do square matrices"); 3694 3695 /* 3696 This code adds extra rows to make sure the number of rows is 3697 divisible by the blocksize 3698 */ 3699 Mbs = M/bs; 3700 extra_rows = bs - M + bs*Mbs; 3701 if (extra_rows == bs) extra_rows = 0; 3702 else Mbs++; 3703 if (extra_rows && !rank) { 3704 ierr = PetscInfo(viewer,"Padding loaded matrix to match blocksize\n");CHKERRQ(ierr); 3705 } 3706 3707 /* determine ownership of all rows */ 3708 if (newmat->rmap->n < 0) { /* PETSC_DECIDE */ 3709 mbs = Mbs/size + ((Mbs % size) > rank); 3710 m = mbs*bs; 3711 } else { /* User set */ 3712 m = newmat->rmap->n; 3713 mbs = m/bs; 3714 } 3715 ierr = PetscMalloc2(size+1,PetscInt,&rowners,size+1,PetscInt,&browners);CHKERRQ(ierr); 3716 ierr = MPI_Allgather(&mbs,1,MPIU_INT,rowners+1,1,MPIU_INT,comm);CHKERRQ(ierr); 3717 3718 /* process 0 needs enough room for process with most rows */ 3719 if (!rank) { 3720 mmax = rowners[1]; 3721 for (i=2; i<=size; i++) { 3722 mmax = PetscMax(mmax,rowners[i]); 3723 } 3724 mmax*=bs; 3725 } else mmax = -1; /* unused, but compiler warns anyway */ 3726 3727 rowners[0] = 0; 3728 for (i=2; i<=size; i++) rowners[i] += rowners[i-1]; 3729 for (i=0; i<=size; i++) browners[i] = rowners[i]*bs; 3730 rstart = rowners[rank]; 3731 rend = rowners[rank+1]; 3732 3733 /* distribute row lengths to all processors */ 3734 ierr = PetscMalloc(m*sizeof(PetscInt),&locrowlens);CHKERRQ(ierr); 3735 if (!rank) { 3736 mend = m; 3737 if (size == 1) mend = mend - extra_rows; 3738 ierr = PetscBinaryRead(fd,locrowlens,mend,PETSC_INT);CHKERRQ(ierr); 3739 for (j=mend; j<m; j++) locrowlens[j] = 1; 3740 ierr = PetscMalloc(mmax*sizeof(PetscInt),&rowlengths);CHKERRQ(ierr); 3741 ierr = PetscMalloc(size*sizeof(PetscInt),&procsnz);CHKERRQ(ierr); 3742 ierr = PetscMemzero(procsnz,size*sizeof(PetscInt));CHKERRQ(ierr); 3743 for (j=0; j<m; j++) { 3744 procsnz[0] += locrowlens[j]; 3745 } 3746 for (i=1; i<size; i++) { 3747 mend = browners[i+1] - browners[i]; 3748 if (i == size-1) mend = mend - extra_rows; 3749 ierr = PetscBinaryRead(fd,rowlengths,mend,PETSC_INT);CHKERRQ(ierr); 3750 for (j=mend; j<browners[i+1] - browners[i]; j++) rowlengths[j] = 1; 3751 /* calculate the number of nonzeros on each processor */ 3752 for (j=0; j<browners[i+1]-browners[i]; j++) { 3753 procsnz[i] += rowlengths[j]; 3754 } 3755 ierr = MPI_Send(rowlengths,browners[i+1]-browners[i],MPIU_INT,i,tag,comm);CHKERRQ(ierr); 3756 } 3757 ierr = PetscFree(rowlengths);CHKERRQ(ierr); 3758 } else { 3759 ierr = MPI_Recv(locrowlens,m,MPIU_INT,0,tag,comm,&status);CHKERRQ(ierr); 3760 } 3761 3762 if (!rank) { 3763 /* determine max buffer needed and allocate it */ 3764 maxnz = procsnz[0]; 3765 for (i=1; i<size; i++) { 3766 maxnz = PetscMax(maxnz,procsnz[i]); 3767 } 3768 ierr = PetscMalloc(maxnz*sizeof(PetscInt),&cols);CHKERRQ(ierr); 3769 3770 /* read in my part of the matrix column indices */ 3771 nz = procsnz[0]; 3772 ierr = PetscMalloc((nz+1)*sizeof(PetscInt),&ibuf);CHKERRQ(ierr); 3773 mycols = ibuf; 3774 if (size == 1) nz -= extra_rows; 3775 ierr = PetscBinaryRead(fd,mycols,nz,PETSC_INT);CHKERRQ(ierr); 3776 if (size == 1) { 3777 for (i=0; i< extra_rows; i++) mycols[nz+i] = M+i; 3778 } 3779 3780 /* read in every ones (except the last) and ship off */ 3781 for (i=1; i<size-1; i++) { 3782 nz = procsnz[i]; 3783 ierr = PetscBinaryRead(fd,cols,nz,PETSC_INT);CHKERRQ(ierr); 3784 ierr = MPI_Send(cols,nz,MPIU_INT,i,tag,comm);CHKERRQ(ierr); 3785 } 3786 /* read in the stuff for the last proc */ 3787 if (size != 1) { 3788 nz = procsnz[size-1] - extra_rows; /* the extra rows are not on the disk */ 3789 ierr = PetscBinaryRead(fd,cols,nz,PETSC_INT);CHKERRQ(ierr); 3790 for (i=0; i<extra_rows; i++) cols[nz+i] = M+i; 3791 ierr = MPI_Send(cols,nz+extra_rows,MPIU_INT,size-1,tag,comm);CHKERRQ(ierr); 3792 } 3793 ierr = PetscFree(cols);CHKERRQ(ierr); 3794 } else { 3795 /* determine buffer space needed for message */ 3796 nz = 0; 3797 for (i=0; i<m; i++) { 3798 nz += locrowlens[i]; 3799 } 3800 ierr = PetscMalloc((nz+1)*sizeof(PetscInt),&ibuf);CHKERRQ(ierr); 3801 mycols = ibuf; 3802 /* receive message of column indices*/ 3803 ierr = MPI_Recv(mycols,nz,MPIU_INT,0,tag,comm,&status);CHKERRQ(ierr); 3804 ierr = MPI_Get_count(&status,MPIU_INT,&maxnz);CHKERRQ(ierr); 3805 if (maxnz != nz) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED,"something is wrong with file"); 3806 } 3807 3808 /* loop over local rows, determining number of off diagonal entries */ 3809 ierr = PetscMalloc2(rend-rstart,PetscInt,&dlens,rend-rstart,PetscInt,&odlens);CHKERRQ(ierr); 3810 ierr = PetscMalloc3(Mbs,PetscInt,&mask,Mbs,PetscInt,&masked1,Mbs,PetscInt,&masked2);CHKERRQ(ierr); 3811 ierr = PetscMemzero(mask,Mbs*sizeof(PetscInt));CHKERRQ(ierr); 3812 ierr = PetscMemzero(masked1,Mbs*sizeof(PetscInt));CHKERRQ(ierr); 3813 ierr = PetscMemzero(masked2,Mbs*sizeof(PetscInt));CHKERRQ(ierr); 3814 rowcount = 0; nzcount = 0; 3815 for (i=0; i<mbs; i++) { 3816 dcount = 0; 3817 odcount = 0; 3818 for (j=0; j<bs; j++) { 3819 kmax = locrowlens[rowcount]; 3820 for (k=0; k<kmax; k++) { 3821 tmp = mycols[nzcount++]/bs; 3822 if (!mask[tmp]) { 3823 mask[tmp] = 1; 3824 if (tmp < rstart || tmp >= rend) masked2[odcount++] = tmp; 3825 else masked1[dcount++] = tmp; 3826 } 3827 } 3828 rowcount++; 3829 } 3830 3831 dlens[i] = dcount; 3832 odlens[i] = odcount; 3833 3834 /* zero out the mask elements we set */ 3835 for (j=0; j<dcount; j++) mask[masked1[j]] = 0; 3836 for (j=0; j<odcount; j++) mask[masked2[j]] = 0; 3837 } 3838 3839 3840 if (!sizesset) { 3841 ierr = MatSetSizes(newmat,m,m,M+extra_rows,N+extra_rows);CHKERRQ(ierr); 3842 } 3843 ierr = MatMPIBAIJSetPreallocation(newmat,bs,0,dlens,0,odlens);CHKERRQ(ierr); 3844 3845 if (!rank) { 3846 ierr = PetscMalloc((maxnz+1)*sizeof(PetscScalar),&buf);CHKERRQ(ierr); 3847 /* read in my part of the matrix numerical values */ 3848 nz = procsnz[0]; 3849 vals = buf; 3850 mycols = ibuf; 3851 if (size == 1) nz -= extra_rows; 3852 ierr = PetscBinaryRead(fd,vals,nz,PETSC_SCALAR);CHKERRQ(ierr); 3853 if (size == 1) { 3854 for (i=0; i< extra_rows; i++) vals[nz+i] = 1.0; 3855 } 3856 3857 /* insert into matrix */ 3858 jj = rstart*bs; 3859 for (i=0; i<m; i++) { 3860 ierr = MatSetValues_MPIBAIJ(newmat,1,&jj,locrowlens[i],mycols,vals,INSERT_VALUES);CHKERRQ(ierr); 3861 mycols += locrowlens[i]; 3862 vals += locrowlens[i]; 3863 jj++; 3864 } 3865 /* read in other processors (except the last one) and ship out */ 3866 for (i=1; i<size-1; i++) { 3867 nz = procsnz[i]; 3868 vals = buf; 3869 ierr = PetscBinaryRead(fd,vals,nz,PETSC_SCALAR);CHKERRQ(ierr); 3870 ierr = MPIULong_Send(vals,nz,MPIU_SCALAR,i,((PetscObject)newmat)->tag,comm);CHKERRQ(ierr); 3871 } 3872 /* the last proc */ 3873 if (size != 1) { 3874 nz = procsnz[i] - extra_rows; 3875 vals = buf; 3876 ierr = PetscBinaryRead(fd,vals,nz,PETSC_SCALAR);CHKERRQ(ierr); 3877 for (i=0; i<extra_rows; i++) vals[nz+i] = 1.0; 3878 ierr = MPIULong_Send(vals,nz+extra_rows,MPIU_SCALAR,size-1,((PetscObject)newmat)->tag,comm);CHKERRQ(ierr); 3879 } 3880 ierr = PetscFree(procsnz);CHKERRQ(ierr); 3881 } else { 3882 /* receive numeric values */ 3883 ierr = PetscMalloc((nz+1)*sizeof(PetscScalar),&buf);CHKERRQ(ierr); 3884 3885 /* receive message of values*/ 3886 vals = buf; 3887 mycols = ibuf; 3888 ierr = MPIULong_Recv(vals,nz,MPIU_SCALAR,0,((PetscObject)newmat)->tag,comm);CHKERRQ(ierr); 3889 3890 /* insert into matrix */ 3891 jj = rstart*bs; 3892 for (i=0; i<m; i++) { 3893 ierr = MatSetValues_MPIBAIJ(newmat,1,&jj,locrowlens[i],mycols,vals,INSERT_VALUES);CHKERRQ(ierr); 3894 mycols += locrowlens[i]; 3895 vals += locrowlens[i]; 3896 jj++; 3897 } 3898 } 3899 ierr = PetscFree(locrowlens);CHKERRQ(ierr); 3900 ierr = PetscFree(buf);CHKERRQ(ierr); 3901 ierr = PetscFree(ibuf);CHKERRQ(ierr); 3902 ierr = PetscFree2(rowners,browners);CHKERRQ(ierr); 3903 ierr = PetscFree2(dlens,odlens);CHKERRQ(ierr); 3904 ierr = PetscFree3(mask,masked1,masked2);CHKERRQ(ierr); 3905 ierr = MatAssemblyBegin(newmat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3906 ierr = MatAssemblyEnd(newmat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3907 PetscFunctionReturn(0); 3908 } 3909 3910 #undef __FUNCT__ 3911 #define __FUNCT__ "MatMPIBAIJSetHashTableFactor" 3912 /*@ 3913 MatMPIBAIJSetHashTableFactor - Sets the factor required to compute the size of the HashTable. 3914 3915 Input Parameters: 3916 . mat - the matrix 3917 . fact - factor 3918 3919 Not Collective, each process can use a different factor 3920 3921 Level: advanced 3922 3923 Notes: 3924 This can also be set by the command line option: -mat_use_hash_table <fact> 3925 3926 .keywords: matrix, hashtable, factor, HT 3927 3928 .seealso: MatSetOption() 3929 @*/ 3930 PetscErrorCode MatMPIBAIJSetHashTableFactor(Mat mat,PetscReal fact) 3931 { 3932 PetscErrorCode ierr; 3933 3934 PetscFunctionBegin; 3935 ierr = PetscTryMethod(mat,"MatSetHashTableFactor_C",(Mat,PetscReal),(mat,fact));CHKERRQ(ierr); 3936 PetscFunctionReturn(0); 3937 } 3938 3939 #undef __FUNCT__ 3940 #define __FUNCT__ "MatSetHashTableFactor_MPIBAIJ" 3941 PetscErrorCode MatSetHashTableFactor_MPIBAIJ(Mat mat,PetscReal fact) 3942 { 3943 Mat_MPIBAIJ *baij; 3944 3945 PetscFunctionBegin; 3946 baij = (Mat_MPIBAIJ*)mat->data; 3947 baij->ht_fact = fact; 3948 PetscFunctionReturn(0); 3949 } 3950 3951 #undef __FUNCT__ 3952 #define __FUNCT__ "MatMPIBAIJGetSeqBAIJ" 3953 PetscErrorCode MatMPIBAIJGetSeqBAIJ(Mat A,Mat *Ad,Mat *Ao,const PetscInt *colmap[]) 3954 { 3955 Mat_MPIBAIJ *a = (Mat_MPIBAIJ*)A->data; 3956 3957 PetscFunctionBegin; 3958 *Ad = a->A; 3959 *Ao = a->B; 3960 *colmap = a->garray; 3961 PetscFunctionReturn(0); 3962 } 3963 3964 /* 3965 Special version for direct calls from Fortran (to eliminate two function call overheads 3966 */ 3967 #if defined(PETSC_HAVE_FORTRAN_CAPS) 3968 #define matmpibaijsetvaluesblocked_ MATMPIBAIJSETVALUESBLOCKED 3969 #elif !defined(PETSC_HAVE_FORTRAN_UNDERSCORE) 3970 #define matmpibaijsetvaluesblocked_ matmpibaijsetvaluesblocked 3971 #endif 3972 3973 #undef __FUNCT__ 3974 #define __FUNCT__ "matmpibiajsetvaluesblocked" 3975 /*@C 3976 MatMPIBAIJSetValuesBlocked - Direct Fortran call to replace call to MatSetValuesBlocked() 3977 3978 Collective on Mat 3979 3980 Input Parameters: 3981 + mat - the matrix 3982 . min - number of input rows 3983 . im - input rows 3984 . nin - number of input columns 3985 . in - input columns 3986 . v - numerical values input 3987 - addvin - INSERT_VALUES or ADD_VALUES 3988 3989 Notes: This has a complete copy of MatSetValuesBlocked_MPIBAIJ() which is terrible code un-reuse. 3990 3991 Level: advanced 3992 3993 .seealso: MatSetValuesBlocked() 3994 @*/ 3995 PetscErrorCode matmpibaijsetvaluesblocked_(Mat *matin,PetscInt *min,const PetscInt im[],PetscInt *nin,const PetscInt in[],const MatScalar v[],InsertMode *addvin) 3996 { 3997 /* convert input arguments to C version */ 3998 Mat mat = *matin; 3999 PetscInt m = *min, n = *nin; 4000 InsertMode addv = *addvin; 4001 4002 Mat_MPIBAIJ *baij = (Mat_MPIBAIJ*)mat->data; 4003 const MatScalar *value; 4004 MatScalar *barray = baij->barray; 4005 PetscBool roworiented = baij->roworiented; 4006 PetscErrorCode ierr; 4007 PetscInt i,j,ii,jj,row,col,rstart=baij->rstartbs; 4008 PetscInt rend=baij->rendbs,cstart=baij->cstartbs,stepval; 4009 PetscInt cend=baij->cendbs,bs=mat->rmap->bs,bs2=baij->bs2; 4010 4011 PetscFunctionBegin; 4012 /* tasks normally handled by MatSetValuesBlocked() */ 4013 if (mat->insertmode == NOT_SET_VALUES) mat->insertmode = addv; 4014 #if defined(PETSC_USE_DEBUG) 4015 else if (mat->insertmode != addv) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Cannot mix add values and insert values"); 4016 if (mat->factortype) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Not for factored matrix"); 4017 #endif 4018 if (mat->assembled) { 4019 mat->was_assembled = PETSC_TRUE; 4020 mat->assembled = PETSC_FALSE; 4021 } 4022 ierr = PetscLogEventBegin(MAT_SetValues,mat,0,0,0);CHKERRQ(ierr); 4023 4024 4025 if (!barray) { 4026 ierr = PetscMalloc(bs2*sizeof(MatScalar),&barray);CHKERRQ(ierr); 4027 baij->barray = barray; 4028 } 4029 4030 if (roworiented) stepval = (n-1)*bs; 4031 else stepval = (m-1)*bs; 4032 4033 for (i=0; i<m; i++) { 4034 if (im[i] < 0) continue; 4035 #if defined(PETSC_USE_DEBUG) 4036 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); 4037 #endif 4038 if (im[i] >= rstart && im[i] < rend) { 4039 row = im[i] - rstart; 4040 for (j=0; j<n; j++) { 4041 /* If NumCol = 1 then a copy is not required */ 4042 if ((roworiented) && (n == 1)) { 4043 barray = (MatScalar*)v + i*bs2; 4044 } else if ((!roworiented) && (m == 1)) { 4045 barray = (MatScalar*)v + j*bs2; 4046 } else { /* Here a copy is required */ 4047 if (roworiented) { 4048 value = v + i*(stepval+bs)*bs + j*bs; 4049 } else { 4050 value = v + j*(stepval+bs)*bs + i*bs; 4051 } 4052 for (ii=0; ii<bs; ii++,value+=stepval) { 4053 for (jj=0; jj<bs; jj++) { 4054 *barray++ = *value++; 4055 } 4056 } 4057 barray -=bs2; 4058 } 4059 4060 if (in[j] >= cstart && in[j] < cend) { 4061 col = in[j] - cstart; 4062 ierr = MatSetValuesBlocked_SeqBAIJ(baij->A,1,&row,1,&col,barray,addv);CHKERRQ(ierr); 4063 } else if (in[j] < 0) continue; 4064 #if defined(PETSC_USE_DEBUG) 4065 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); 4066 #endif 4067 else { 4068 if (mat->was_assembled) { 4069 if (!baij->colmap) { 4070 ierr = MatCreateColmap_MPIBAIJ_Private(mat);CHKERRQ(ierr); 4071 } 4072 4073 #if defined(PETSC_USE_DEBUG) 4074 #if defined(PETSC_USE_CTABLE) 4075 { PetscInt data; 4076 ierr = PetscTableFind(baij->colmap,in[j]+1,&data);CHKERRQ(ierr); 4077 if ((data - 1) % bs) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Incorrect colmap"); 4078 } 4079 #else 4080 if ((baij->colmap[in[j]] - 1) % bs) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Incorrect colmap"); 4081 #endif 4082 #endif 4083 #if defined(PETSC_USE_CTABLE) 4084 ierr = PetscTableFind(baij->colmap,in[j]+1,&col);CHKERRQ(ierr); 4085 col = (col - 1)/bs; 4086 #else 4087 col = (baij->colmap[in[j]] - 1)/bs; 4088 #endif 4089 if (col < 0 && !((Mat_SeqBAIJ*)(baij->A->data))->nonew) { 4090 ierr = MatDisAssemble_MPIBAIJ(mat);CHKERRQ(ierr); 4091 col = in[j]; 4092 } 4093 } else col = in[j]; 4094 ierr = MatSetValuesBlocked_SeqBAIJ(baij->B,1,&row,1,&col,barray,addv);CHKERRQ(ierr); 4095 } 4096 } 4097 } else { 4098 if (!baij->donotstash) { 4099 if (roworiented) { 4100 ierr = MatStashValuesRowBlocked_Private(&mat->bstash,im[i],n,in,v,m,n,i);CHKERRQ(ierr); 4101 } else { 4102 ierr = MatStashValuesColBlocked_Private(&mat->bstash,im[i],n,in,v,m,n,i);CHKERRQ(ierr); 4103 } 4104 } 4105 } 4106 } 4107 4108 /* task normally handled by MatSetValuesBlocked() */ 4109 ierr = PetscLogEventEnd(MAT_SetValues,mat,0,0,0);CHKERRQ(ierr); 4110 PetscFunctionReturn(0); 4111 } 4112 4113 #undef __FUNCT__ 4114 #define __FUNCT__ "MatCreateMPIBAIJWithArrays" 4115 /*@ 4116 MatCreateMPIBAIJWithArrays - creates a MPI BAIJ matrix using arrays that contain in standard 4117 CSR format the local rows. 4118 4119 Collective on MPI_Comm 4120 4121 Input Parameters: 4122 + comm - MPI communicator 4123 . bs - the block size, only a block size of 1 is supported 4124 . m - number of local rows (Cannot be PETSC_DECIDE) 4125 . n - This value should be the same as the local size used in creating the 4126 x vector for the matrix-vector product y = Ax. (or PETSC_DECIDE to have 4127 calculated if N is given) For square matrices n is almost always m. 4128 . M - number of global rows (or PETSC_DETERMINE to have calculated if m is given) 4129 . N - number of global columns (or PETSC_DETERMINE to have calculated if n is given) 4130 . i - row indices 4131 . j - column indices 4132 - a - matrix values 4133 4134 Output Parameter: 4135 . mat - the matrix 4136 4137 Level: intermediate 4138 4139 Notes: 4140 The i, j, and a arrays ARE copied by this routine into the internal format used by PETSc; 4141 thus you CANNOT change the matrix entries by changing the values of a[] after you have 4142 called this routine. Use MatCreateMPIAIJWithSplitArrays() to avoid needing to copy the arrays. 4143 4144 The order of the entries in values is the same as the block compressed sparse row storage format; that is, it is 4145 the same as a three dimensional array in Fortran values(bs,bs,nnz) that contains the first column of the first 4146 block, followed by the second column of the first block etc etc. That is, the blocks are contiguous in memory 4147 with column-major ordering within blocks. 4148 4149 The i and j indices are 0 based, and i indices are indices corresponding to the local j array. 4150 4151 .keywords: matrix, aij, compressed row, sparse, parallel 4152 4153 .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatMPIAIJSetPreallocation(), MatMPIAIJSetPreallocationCSR(), 4154 MPIAIJ, MatCreateAIJ(), MatCreateMPIAIJWithSplitArrays() 4155 @*/ 4156 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) 4157 { 4158 PetscErrorCode ierr; 4159 4160 PetscFunctionBegin; 4161 if (i[0]) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"i (row indices) must start with 0"); 4162 if (m < 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"local number of rows (m) cannot be PETSC_DECIDE, or negative"); 4163 ierr = MatCreate(comm,mat);CHKERRQ(ierr); 4164 ierr = MatSetSizes(*mat,m,n,M,N);CHKERRQ(ierr); 4165 ierr = MatSetType(*mat,MATMPISBAIJ);CHKERRQ(ierr); 4166 ierr = MatSetOption(*mat,MAT_ROW_ORIENTED,PETSC_FALSE);CHKERRQ(ierr); 4167 ierr = MatMPIBAIJSetPreallocationCSR(*mat,bs,i,j,a);CHKERRQ(ierr); 4168 ierr = MatSetOption(*mat,MAT_ROW_ORIENTED,PETSC_TRUE);CHKERRQ(ierr); 4169 PetscFunctionReturn(0); 4170 } 4171