1 2 /* 3 Routines to compute overlapping regions of a parallel MPI matrix 4 and to find submatrices that were shared across processors. 5 */ 6 #include <../src/mat/impls/baij/mpi/mpibaij.h> 7 #include <petscbt.h> 8 9 static PetscErrorCode MatIncreaseOverlap_MPIBAIJ_Local(Mat,PetscInt,char**,PetscInt*,PetscInt**); 10 static PetscErrorCode MatIncreaseOverlap_MPIBAIJ_Receive(Mat,PetscInt,PetscInt**,PetscInt**,PetscInt*); 11 extern PetscErrorCode MatGetRow_MPIBAIJ(Mat,PetscInt,PetscInt*,PetscInt**,PetscScalar**); 12 extern PetscErrorCode MatRestoreRow_MPIBAIJ(Mat,PetscInt,PetscInt*,PetscInt**,PetscScalar**); 13 14 PetscErrorCode MatIncreaseOverlap_MPIBAIJ(Mat C,PetscInt imax,IS is[],PetscInt ov) 15 { 16 PetscErrorCode ierr; 17 PetscInt i,N=C->cmap->N, bs=C->rmap->bs; 18 IS *is_new; 19 20 PetscFunctionBegin; 21 ierr = PetscMalloc1(imax,&is_new);CHKERRQ(ierr); 22 /* Convert the indices into block format */ 23 ierr = ISCompressIndicesGeneral(N,C->rmap->n,bs,imax,is,is_new);CHKERRQ(ierr); 24 if (ov < 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Negative overlap specified\n"); 25 for (i=0; i<ov; ++i) { 26 ierr = MatIncreaseOverlap_MPIBAIJ_Once(C,imax,is_new);CHKERRQ(ierr); 27 } 28 for (i=0; i<imax; i++) {ierr = ISDestroy(&is[i]);CHKERRQ(ierr);} 29 ierr = ISExpandIndicesGeneral(N,N,bs,imax,is_new,is);CHKERRQ(ierr); 30 for (i=0; i<imax; i++) {ierr = ISDestroy(&is_new[i]);CHKERRQ(ierr);} 31 ierr = PetscFree(is_new);CHKERRQ(ierr); 32 PetscFunctionReturn(0); 33 } 34 35 /* 36 Sample message format: 37 If a processor A wants processor B to process some elements corresponding 38 to index sets is[1], is[5] 39 mesg [0] = 2 (no of index sets in the mesg) 40 ----------- 41 mesg [1] = 1 => is[1] 42 mesg [2] = sizeof(is[1]); 43 ----------- 44 mesg [5] = 5 => is[5] 45 mesg [6] = sizeof(is[5]); 46 ----------- 47 mesg [7] 48 mesg [n] data(is[1]) 49 ----------- 50 mesg[n+1] 51 mesg[m] data(is[5]) 52 ----------- 53 54 Notes: 55 nrqs - no of requests sent (or to be sent out) 56 nrqr - no of requests recieved (which have to be or which have been processed 57 */ 58 PetscErrorCode MatIncreaseOverlap_MPIBAIJ_Once(Mat C,PetscInt imax,IS is[]) 59 { 60 Mat_MPIBAIJ *c = (Mat_MPIBAIJ*)C->data; 61 const PetscInt **idx,*idx_i; 62 PetscInt *n,*w3,*w4,**data,len; 63 PetscErrorCode ierr; 64 PetscMPIInt size,rank,tag1,tag2,*w2,*w1,nrqr; 65 PetscInt Mbs,i,j,k,**rbuf,row,proc=-1,nrqs,msz,**outdat,**ptr; 66 PetscInt *ctr,*pa,*tmp,*isz,*isz1,**xdata,**rbuf2,*d_p; 67 PetscMPIInt *onodes1,*olengths1,*onodes2,*olengths2; 68 PetscBT *table; 69 MPI_Comm comm; 70 MPI_Request *s_waits1,*r_waits1,*s_waits2,*r_waits2; 71 MPI_Status *s_status,*recv_status; 72 char *t_p; 73 74 PetscFunctionBegin; 75 ierr = PetscObjectGetComm((PetscObject)C,&comm);CHKERRQ(ierr); 76 size = c->size; 77 rank = c->rank; 78 Mbs = c->Mbs; 79 80 ierr = PetscObjectGetNewTag((PetscObject)C,&tag1);CHKERRQ(ierr); 81 ierr = PetscObjectGetNewTag((PetscObject)C,&tag2);CHKERRQ(ierr); 82 83 ierr = PetscMalloc2(imax+1,&idx,imax,&n);CHKERRQ(ierr); 84 85 for (i=0; i<imax; i++) { 86 ierr = ISGetIndices(is[i],&idx[i]);CHKERRQ(ierr); 87 ierr = ISGetLocalSize(is[i],&n[i]);CHKERRQ(ierr); 88 } 89 90 /* evaluate communication - mesg to who,length of mesg, and buffer space 91 required. Based on this, buffers are allocated, and data copied into them*/ 92 ierr = PetscCalloc4(size,&w1,size,&w2,size,&w3,size,&w4);CHKERRQ(ierr); 93 for (i=0; i<imax; i++) { 94 ierr = PetscMemzero(w4,size*sizeof(PetscInt));CHKERRQ(ierr); /* initialise work vector*/ 95 idx_i = idx[i]; 96 len = n[i]; 97 for (j=0; j<len; j++) { 98 row = idx_i[j]; 99 if (row < 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Index set cannot have negative entries"); 100 ierr = PetscLayoutFindOwner(C->rmap,row*C->rmap->bs,&proc);CHKERRQ(ierr); 101 w4[proc]++; 102 } 103 for (j=0; j<size; j++) { 104 if (w4[j]) { w1[j] += w4[j]; w3[j]++;} 105 } 106 } 107 108 nrqs = 0; /* no of outgoing messages */ 109 msz = 0; /* total mesg length (for all proc */ 110 w1[rank] = 0; /* no mesg sent to itself */ 111 w3[rank] = 0; 112 for (i=0; i<size; i++) { 113 if (w1[i]) {w2[i] = 1; nrqs++;} /* there exists a message to proc i */ 114 } 115 /* pa - is list of processors to communicate with */ 116 ierr = PetscMalloc1(nrqs+1,&pa);CHKERRQ(ierr); 117 for (i=0,j=0; i<size; i++) { 118 if (w1[i]) {pa[j] = i; j++;} 119 } 120 121 /* Each message would have a header = 1 + 2*(no of IS) + data */ 122 for (i=0; i<nrqs; i++) { 123 j = pa[i]; 124 w1[j] += w2[j] + 2*w3[j]; 125 msz += w1[j]; 126 } 127 128 /* Determine the number of messages to expect, their lengths, from from-ids */ 129 ierr = PetscGatherNumberOfMessages(comm,w2,w1,&nrqr);CHKERRQ(ierr); 130 ierr = PetscGatherMessageLengths(comm,nrqs,nrqr,w1,&onodes1,&olengths1);CHKERRQ(ierr); 131 132 /* Now post the Irecvs corresponding to these messages */ 133 ierr = PetscPostIrecvInt(comm,tag1,nrqr,onodes1,olengths1,&rbuf,&r_waits1);CHKERRQ(ierr); 134 135 /* Allocate Memory for outgoing messages */ 136 ierr = PetscMalloc4(size,&outdat,size,&ptr,msz,&tmp,size,&ctr);CHKERRQ(ierr); 137 ierr = PetscMemzero(outdat,size*sizeof(PetscInt*));CHKERRQ(ierr); 138 ierr = PetscMemzero(ptr,size*sizeof(PetscInt*));CHKERRQ(ierr); 139 { 140 PetscInt *iptr = tmp,ict = 0; 141 for (i=0; i<nrqs; i++) { 142 j = pa[i]; 143 iptr += ict; 144 outdat[j] = iptr; 145 ict = w1[j]; 146 } 147 } 148 149 /* Form the outgoing messages */ 150 /*plug in the headers*/ 151 for (i=0; i<nrqs; i++) { 152 j = pa[i]; 153 outdat[j][0] = 0; 154 ierr = PetscMemzero(outdat[j]+1,2*w3[j]*sizeof(PetscInt));CHKERRQ(ierr); 155 ptr[j] = outdat[j] + 2*w3[j] + 1; 156 } 157 158 /* Memory for doing local proc's work*/ 159 { 160 ierr = PetscCalloc5(imax,&table, imax,&data, imax,&isz, Mbs*imax,&d_p, (Mbs/PETSC_BITS_PER_BYTE+1)*imax,&t_p);CHKERRQ(ierr); 161 162 for (i=0; i<imax; i++) { 163 table[i] = t_p + (Mbs/PETSC_BITS_PER_BYTE+1)*i; 164 data[i] = d_p + (Mbs)*i; 165 } 166 } 167 168 /* Parse the IS and update local tables and the outgoing buf with the data*/ 169 { 170 PetscInt n_i,*data_i,isz_i,*outdat_j,ctr_j; 171 PetscBT table_i; 172 173 for (i=0; i<imax; i++) { 174 ierr = PetscMemzero(ctr,size*sizeof(PetscInt));CHKERRQ(ierr); 175 n_i = n[i]; 176 table_i = table[i]; 177 idx_i = idx[i]; 178 data_i = data[i]; 179 isz_i = isz[i]; 180 for (j=0; j<n_i; j++) { /* parse the indices of each IS */ 181 row = idx_i[j]; 182 ierr = PetscLayoutFindOwner(C->rmap,row*C->rmap->bs,&proc);CHKERRQ(ierr); 183 if (proc != rank) { /* copy to the outgoing buffer */ 184 ctr[proc]++; 185 *ptr[proc] = row; 186 ptr[proc]++; 187 } else { /* Update the local table */ 188 if (!PetscBTLookupSet(table_i,row)) data_i[isz_i++] = row; 189 } 190 } 191 /* Update the headers for the current IS */ 192 for (j=0; j<size; j++) { /* Can Optimise this loop by using pa[] */ 193 if ((ctr_j = ctr[j])) { 194 outdat_j = outdat[j]; 195 k = ++outdat_j[0]; 196 outdat_j[2*k] = ctr_j; 197 outdat_j[2*k-1] = i; 198 } 199 } 200 isz[i] = isz_i; 201 } 202 } 203 204 /* Now post the sends */ 205 ierr = PetscMalloc1(nrqs+1,&s_waits1);CHKERRQ(ierr); 206 for (i=0; i<nrqs; ++i) { 207 j = pa[i]; 208 ierr = MPI_Isend(outdat[j],w1[j],MPIU_INT,j,tag1,comm,s_waits1+i);CHKERRQ(ierr); 209 } 210 211 /* No longer need the original indices*/ 212 for (i=0; i<imax; ++i) { 213 ierr = ISRestoreIndices(is[i],idx+i);CHKERRQ(ierr); 214 } 215 ierr = PetscFree2(idx,n);CHKERRQ(ierr); 216 217 for (i=0; i<imax; ++i) { 218 ierr = ISDestroy(&is[i]);CHKERRQ(ierr); 219 } 220 221 /* Do Local work*/ 222 ierr = MatIncreaseOverlap_MPIBAIJ_Local(C,imax,table,isz,data);CHKERRQ(ierr); 223 224 /* Receive messages*/ 225 ierr = PetscMalloc1(nrqr+1,&recv_status);CHKERRQ(ierr); 226 if (nrqr) {ierr = MPI_Waitall(nrqr,r_waits1,recv_status);CHKERRQ(ierr);} 227 228 ierr = PetscMalloc1(nrqs+1,&s_status);CHKERRQ(ierr); 229 if (nrqs) {ierr = MPI_Waitall(nrqs,s_waits1,s_status);CHKERRQ(ierr);} 230 231 /* Phase 1 sends are complete - deallocate buffers */ 232 ierr = PetscFree4(outdat,ptr,tmp,ctr);CHKERRQ(ierr); 233 ierr = PetscFree4(w1,w2,w3,w4);CHKERRQ(ierr); 234 235 ierr = PetscMalloc1(nrqr+1,&xdata);CHKERRQ(ierr); 236 ierr = PetscMalloc1(nrqr+1,&isz1);CHKERRQ(ierr); 237 ierr = MatIncreaseOverlap_MPIBAIJ_Receive(C,nrqr,rbuf,xdata,isz1);CHKERRQ(ierr); 238 ierr = PetscFree(rbuf[0]);CHKERRQ(ierr); 239 ierr = PetscFree(rbuf);CHKERRQ(ierr); 240 241 /* Send the data back*/ 242 /* Do a global reduction to know the buffer space req for incoming messages*/ 243 { 244 PetscMPIInt *rw1; 245 246 ierr = PetscCalloc1(size,&rw1);CHKERRQ(ierr); 247 248 for (i=0; i<nrqr; ++i) { 249 proc = recv_status[i].MPI_SOURCE; 250 if (proc != onodes1[i]) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"MPI_SOURCE mismatch"); 251 rw1[proc] = isz1[i]; 252 } 253 254 ierr = PetscFree(onodes1);CHKERRQ(ierr); 255 ierr = PetscFree(olengths1);CHKERRQ(ierr); 256 257 /* Determine the number of messages to expect, their lengths, from from-ids */ 258 ierr = PetscGatherMessageLengths(comm,nrqr,nrqs,rw1,&onodes2,&olengths2);CHKERRQ(ierr); 259 ierr = PetscFree(rw1);CHKERRQ(ierr); 260 } 261 /* Now post the Irecvs corresponding to these messages */ 262 ierr = PetscPostIrecvInt(comm,tag2,nrqs,onodes2,olengths2,&rbuf2,&r_waits2);CHKERRQ(ierr); 263 264 /* Now post the sends */ 265 ierr = PetscMalloc1(nrqr+1,&s_waits2);CHKERRQ(ierr); 266 for (i=0; i<nrqr; ++i) { 267 j = recv_status[i].MPI_SOURCE; 268 ierr = MPI_Isend(xdata[i],isz1[i],MPIU_INT,j,tag2,comm,s_waits2+i);CHKERRQ(ierr); 269 } 270 271 /* receive work done on other processors*/ 272 { 273 PetscMPIInt idex; 274 PetscInt is_no,ct1,max,*rbuf2_i,isz_i,*data_i,jmax; 275 PetscBT table_i; 276 MPI_Status *status2; 277 278 ierr = PetscMalloc1(PetscMax(nrqr,nrqs)+1,&status2);CHKERRQ(ierr); 279 for (i=0; i<nrqs; ++i) { 280 ierr = MPI_Waitany(nrqs,r_waits2,&idex,status2+i);CHKERRQ(ierr); 281 /* Process the message*/ 282 rbuf2_i = rbuf2[idex]; 283 ct1 = 2*rbuf2_i[0]+1; 284 jmax = rbuf2[idex][0]; 285 for (j=1; j<=jmax; j++) { 286 max = rbuf2_i[2*j]; 287 is_no = rbuf2_i[2*j-1]; 288 isz_i = isz[is_no]; 289 data_i = data[is_no]; 290 table_i = table[is_no]; 291 for (k=0; k<max; k++,ct1++) { 292 row = rbuf2_i[ct1]; 293 if (!PetscBTLookupSet(table_i,row)) data_i[isz_i++] = row; 294 } 295 isz[is_no] = isz_i; 296 } 297 } 298 if (nrqr) {ierr = MPI_Waitall(nrqr,s_waits2,status2);CHKERRQ(ierr);} 299 ierr = PetscFree(status2);CHKERRQ(ierr); 300 } 301 302 for (i=0; i<imax; ++i) { 303 ierr = ISCreateGeneral(PETSC_COMM_SELF,isz[i],data[i],PETSC_COPY_VALUES,is+i);CHKERRQ(ierr); 304 } 305 306 307 ierr = PetscFree(onodes2);CHKERRQ(ierr); 308 ierr = PetscFree(olengths2);CHKERRQ(ierr); 309 310 ierr = PetscFree(pa);CHKERRQ(ierr); 311 ierr = PetscFree(rbuf2[0]);CHKERRQ(ierr); 312 ierr = PetscFree(rbuf2);CHKERRQ(ierr); 313 ierr = PetscFree(s_waits1);CHKERRQ(ierr); 314 ierr = PetscFree(r_waits1);CHKERRQ(ierr); 315 ierr = PetscFree(s_waits2);CHKERRQ(ierr); 316 ierr = PetscFree(r_waits2);CHKERRQ(ierr); 317 ierr = PetscFree5(table,data,isz,d_p,t_p);CHKERRQ(ierr); 318 ierr = PetscFree(s_status);CHKERRQ(ierr); 319 ierr = PetscFree(recv_status);CHKERRQ(ierr); 320 ierr = PetscFree(xdata[0]);CHKERRQ(ierr); 321 ierr = PetscFree(xdata);CHKERRQ(ierr); 322 ierr = PetscFree(isz1);CHKERRQ(ierr); 323 PetscFunctionReturn(0); 324 } 325 326 /* 327 MatIncreaseOverlap_MPIBAIJ_Local - Called by MatincreaseOverlap, to do 328 the work on the local processor. 329 330 Inputs: 331 C - MAT_MPIBAIJ; 332 imax - total no of index sets processed at a time; 333 table - an array of char - size = Mbs bits. 334 335 Output: 336 isz - array containing the count of the solution elements corresponding 337 to each index set; 338 data - pointer to the solutions 339 */ 340 static PetscErrorCode MatIncreaseOverlap_MPIBAIJ_Local(Mat C,PetscInt imax,PetscBT *table,PetscInt *isz,PetscInt **data) 341 { 342 Mat_MPIBAIJ *c = (Mat_MPIBAIJ*)C->data; 343 Mat A = c->A,B = c->B; 344 Mat_SeqBAIJ *a = (Mat_SeqBAIJ*)A->data,*b = (Mat_SeqBAIJ*)B->data; 345 PetscInt start,end,val,max,rstart,cstart,*ai,*aj; 346 PetscInt *bi,*bj,*garray,i,j,k,row,*data_i,isz_i; 347 PetscBT table_i; 348 349 PetscFunctionBegin; 350 rstart = c->rstartbs; 351 cstart = c->cstartbs; 352 ai = a->i; 353 aj = a->j; 354 bi = b->i; 355 bj = b->j; 356 garray = c->garray; 357 358 359 for (i=0; i<imax; i++) { 360 data_i = data[i]; 361 table_i = table[i]; 362 isz_i = isz[i]; 363 for (j=0,max=isz[i]; j<max; j++) { 364 row = data_i[j] - rstart; 365 start = ai[row]; 366 end = ai[row+1]; 367 for (k=start; k<end; k++) { /* Amat */ 368 val = aj[k] + cstart; 369 if (!PetscBTLookupSet(table_i,val)) data_i[isz_i++] = val; 370 } 371 start = bi[row]; 372 end = bi[row+1]; 373 for (k=start; k<end; k++) { /* Bmat */ 374 val = garray[bj[k]]; 375 if (!PetscBTLookupSet(table_i,val)) data_i[isz_i++] = val; 376 } 377 } 378 isz[i] = isz_i; 379 } 380 PetscFunctionReturn(0); 381 } 382 /* 383 MatIncreaseOverlap_MPIBAIJ_Receive - Process the recieved messages, 384 and return the output 385 386 Input: 387 C - the matrix 388 nrqr - no of messages being processed. 389 rbuf - an array of pointers to the recieved requests 390 391 Output: 392 xdata - array of messages to be sent back 393 isz1 - size of each message 394 395 For better efficiency perhaps we should malloc separately each xdata[i], 396 then if a remalloc is required we need only copy the data for that one row 397 rather than all previous rows as it is now where a single large chunck of 398 memory is used. 399 400 */ 401 static PetscErrorCode MatIncreaseOverlap_MPIBAIJ_Receive(Mat C,PetscInt nrqr,PetscInt **rbuf,PetscInt **xdata,PetscInt * isz1) 402 { 403 Mat_MPIBAIJ *c = (Mat_MPIBAIJ*)C->data; 404 Mat A = c->A,B = c->B; 405 Mat_SeqBAIJ *a = (Mat_SeqBAIJ*)A->data,*b = (Mat_SeqBAIJ*)B->data; 406 PetscErrorCode ierr; 407 PetscInt rstart,cstart,*ai,*aj,*bi,*bj,*garray,i,j,k; 408 PetscInt row,total_sz,ct,ct1,ct2,ct3,mem_estimate,oct2,l,start,end; 409 PetscInt val,max1,max2,Mbs,no_malloc =0,*tmp,new_estimate,ctr; 410 PetscInt *rbuf_i,kmax,rbuf_0; 411 PetscBT xtable; 412 413 PetscFunctionBegin; 414 Mbs = c->Mbs; 415 rstart = c->rstartbs; 416 cstart = c->cstartbs; 417 ai = a->i; 418 aj = a->j; 419 bi = b->i; 420 bj = b->j; 421 garray = c->garray; 422 423 424 for (i=0,ct=0,total_sz=0; i<nrqr; ++i) { 425 rbuf_i = rbuf[i]; 426 rbuf_0 = rbuf_i[0]; 427 ct += rbuf_0; 428 for (j=1; j<=rbuf_0; j++) total_sz += rbuf_i[2*j]; 429 } 430 431 if (c->Mbs) max1 = ct*(a->nz +b->nz)/c->Mbs; 432 else max1 = 1; 433 mem_estimate = 3*((total_sz > max1 ? total_sz : max1)+1); 434 ierr = PetscMalloc1(mem_estimate,&xdata[0]);CHKERRQ(ierr); 435 ++no_malloc; 436 ierr = PetscBTCreate(Mbs,&xtable);CHKERRQ(ierr); 437 ierr = PetscMemzero(isz1,nrqr*sizeof(PetscInt));CHKERRQ(ierr); 438 439 ct3 = 0; 440 for (i=0; i<nrqr; i++) { /* for easch mesg from proc i */ 441 rbuf_i = rbuf[i]; 442 rbuf_0 = rbuf_i[0]; 443 ct1 = 2*rbuf_0+1; 444 ct2 = ct1; 445 ct3 += ct1; 446 for (j=1; j<=rbuf_0; j++) { /* for each IS from proc i*/ 447 ierr = PetscBTMemzero(Mbs,xtable);CHKERRQ(ierr); 448 oct2 = ct2; 449 kmax = rbuf_i[2*j]; 450 for (k=0; k<kmax; k++,ct1++) { 451 row = rbuf_i[ct1]; 452 if (!PetscBTLookupSet(xtable,row)) { 453 if (!(ct3 < mem_estimate)) { 454 new_estimate = (PetscInt)(1.5*mem_estimate)+1; 455 ierr = PetscMalloc1(new_estimate,&tmp);CHKERRQ(ierr); 456 ierr = PetscMemcpy(tmp,xdata[0],mem_estimate*sizeof(PetscInt));CHKERRQ(ierr); 457 ierr = PetscFree(xdata[0]);CHKERRQ(ierr); 458 xdata[0] = tmp; 459 mem_estimate = new_estimate; ++no_malloc; 460 for (ctr=1; ctr<=i; ctr++) xdata[ctr] = xdata[ctr-1] + isz1[ctr-1]; 461 } 462 xdata[i][ct2++] = row; 463 ct3++; 464 } 465 } 466 for (k=oct2,max2=ct2; k<max2; k++) { 467 row = xdata[i][k] - rstart; 468 start = ai[row]; 469 end = ai[row+1]; 470 for (l=start; l<end; l++) { 471 val = aj[l] + cstart; 472 if (!PetscBTLookupSet(xtable,val)) { 473 if (!(ct3 < mem_estimate)) { 474 new_estimate = (PetscInt)(1.5*mem_estimate)+1; 475 ierr = PetscMalloc1(new_estimate,&tmp);CHKERRQ(ierr); 476 ierr = PetscMemcpy(tmp,xdata[0],mem_estimate*sizeof(PetscInt));CHKERRQ(ierr); 477 ierr = PetscFree(xdata[0]);CHKERRQ(ierr); 478 xdata[0] = tmp; 479 mem_estimate = new_estimate; ++no_malloc; 480 for (ctr=1; ctr<=i; ctr++) xdata[ctr] = xdata[ctr-1] + isz1[ctr-1]; 481 } 482 xdata[i][ct2++] = val; 483 ct3++; 484 } 485 } 486 start = bi[row]; 487 end = bi[row+1]; 488 for (l=start; l<end; l++) { 489 val = garray[bj[l]]; 490 if (!PetscBTLookupSet(xtable,val)) { 491 if (!(ct3 < mem_estimate)) { 492 new_estimate = (PetscInt)(1.5*mem_estimate)+1; 493 ierr = PetscMalloc1(new_estimate,&tmp);CHKERRQ(ierr); 494 ierr = PetscMemcpy(tmp,xdata[0],mem_estimate*sizeof(PetscInt));CHKERRQ(ierr); 495 ierr = PetscFree(xdata[0]);CHKERRQ(ierr); 496 xdata[0] = tmp; 497 mem_estimate = new_estimate; ++no_malloc; 498 for (ctr =1; ctr <=i; ctr++) xdata[ctr] = xdata[ctr-1] + isz1[ctr-1]; 499 } 500 xdata[i][ct2++] = val; 501 ct3++; 502 } 503 } 504 } 505 /* Update the header*/ 506 xdata[i][2*j] = ct2 - oct2; /* Undo the vector isz1 and use only a var*/ 507 xdata[i][2*j-1] = rbuf_i[2*j-1]; 508 } 509 xdata[i][0] = rbuf_0; 510 xdata[i+1] = xdata[i] + ct2; 511 isz1[i] = ct2; /* size of each message */ 512 } 513 ierr = PetscBTDestroy(&xtable);CHKERRQ(ierr); 514 ierr = PetscInfo3(C,"Allocated %D bytes, required %D, no of mallocs = %D\n",mem_estimate,ct3,no_malloc);CHKERRQ(ierr); 515 PetscFunctionReturn(0); 516 } 517 518 PetscErrorCode MatGetSubMatrices_MPIBAIJ(Mat C,PetscInt ismax,const IS isrow[],const IS iscol[],MatReuse scall,Mat *submat[]) 519 { 520 IS *isrow_block,*iscol_block; 521 Mat_MPIBAIJ *c = (Mat_MPIBAIJ*)C->data; 522 PetscErrorCode ierr; 523 PetscInt nmax,nstages_local,nstages,i,pos,max_no,ncol,nrow,N=C->cmap->N,bs=C->rmap->bs; 524 PetscBool colflag,*allcolumns,*allrows; 525 526 PetscFunctionBegin; 527 //printf("MatGetSubMatrices_MPIBAIJ scall %d, bs %d\n",scall,bs); 528 /* The compression and expansion should be avoided. Doesn't point 529 out errors, might change the indices, hence buggey */ 530 ierr = PetscMalloc2(ismax+1,&isrow_block,ismax+1,&iscol_block);CHKERRQ(ierr); 531 ierr = ISCompressIndicesGeneral(N,C->rmap->n,bs,ismax,isrow,isrow_block);CHKERRQ(ierr); 532 ierr = ISCompressIndicesGeneral(N,C->cmap->n,bs,ismax,iscol,iscol_block);CHKERRQ(ierr); 533 534 /* Check for special case: each processor gets entire matrix columns */ 535 ierr = PetscMalloc2(ismax+1,&allcolumns,ismax+1,&allrows);CHKERRQ(ierr); 536 for (i=0; i<ismax; i++) { 537 ierr = ISIdentity(iscol[i],&colflag);CHKERRQ(ierr); 538 ierr = ISGetLocalSize(iscol[i],&ncol);CHKERRQ(ierr); 539 if (colflag && ncol == C->cmap->N) allcolumns[i] = PETSC_TRUE; 540 else allcolumns[i] = PETSC_FALSE; 541 542 ierr = ISIdentity(isrow[i],&colflag);CHKERRQ(ierr); 543 ierr = ISGetLocalSize(isrow[i],&nrow);CHKERRQ(ierr); 544 if (colflag && nrow == C->rmap->N) allrows[i] = PETSC_TRUE; 545 else allrows[i] = PETSC_FALSE; 546 } 547 548 /* Allocate memory to hold all the submatrices */ 549 if (scall == MAT_INITIAL_MATRIX) { 550 ierr = PetscMalloc1(ismax+1,submat);CHKERRQ(ierr); 551 } 552 /* Determine the number of stages through which submatrices are done */ 553 nmax = 20*1000000 / (c->Nbs * sizeof(PetscInt)); 554 if (!nmax) nmax = 1; 555 nstages_local = ismax/nmax + ((ismax % nmax) ? 1 : 0); 556 557 /* Make sure every processor loops through the nstages */ 558 ierr = MPIU_Allreduce(&nstages_local,&nstages,1,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)C));CHKERRQ(ierr); 559 for (i=0,pos=0; i<nstages; i++) { 560 if (pos+nmax <= ismax) max_no = nmax; 561 else if (pos == ismax) max_no = 0; 562 else max_no = ismax-pos; 563 564 PetscBool isbaij; 565 ierr = PetscObjectTypeCompare((PetscObject)C,MATMPIBAIJ,&isbaij);CHKERRQ(ierr); 566 if (isbaij) { 567 ierr = MatGetSubMatrices_MPIBAIJ_local_new(C,max_no,isrow_block+pos,iscol_block+pos,scall,allrows+pos,allcolumns+pos,*submat+pos);CHKERRQ(ierr); 568 } else { 569 ierr = MatGetSubMatrices_MPIBAIJ_local(C,max_no,isrow_block+pos,iscol_block+pos,scall,allrows+pos,allcolumns+pos,*submat+pos);CHKERRQ(ierr); 570 } 571 pos += max_no; 572 } 573 574 for (i=0; i<ismax; i++) { 575 ierr = ISDestroy(&isrow_block[i]);CHKERRQ(ierr); 576 ierr = ISDestroy(&iscol_block[i]);CHKERRQ(ierr); 577 } 578 ierr = PetscFree2(isrow_block,iscol_block);CHKERRQ(ierr); 579 ierr = PetscFree2(allcolumns,allrows);CHKERRQ(ierr); 580 PetscFunctionReturn(0); 581 } 582 583 #if defined(PETSC_USE_CTABLE) 584 PetscErrorCode PetscGetProc(const PetscInt row, const PetscMPIInt size, const PetscInt proc_gnode[], PetscMPIInt *rank) 585 { 586 PetscInt nGlobalNd = proc_gnode[size]; 587 PetscMPIInt fproc; 588 PetscErrorCode ierr; 589 590 PetscFunctionBegin; 591 ierr = PetscMPIIntCast((PetscInt)(((float)row * (float)size / (float)nGlobalNd + 0.5)),&fproc);CHKERRQ(ierr); 592 if (fproc > size) fproc = size; 593 while (row < proc_gnode[fproc] || row >= proc_gnode[fproc+1]) { 594 if (row < proc_gnode[fproc]) fproc--; 595 else fproc++; 596 } 597 *rank = fproc; 598 PetscFunctionReturn(0); 599 } 600 #endif 601 602 /* -------------------------------------------------------------------------*/ 603 /* This code is used for BAIJ and SBAIJ matrices (unfortunate dependency) */ 604 605 PetscErrorCode MatDestroy_MPIBAIJ_MatGetSubmatrices(Mat C) 606 { 607 PetscErrorCode ierr; 608 Mat_SeqBAIJ *c = (Mat_SeqBAIJ*)C->data; 609 Mat_SubMat *submatj = c->submatis1; 610 PetscInt i; 611 612 PetscFunctionBegin; 613 if (!submatj->id) { /* delete data that are linked only to submats[id=0] */ 614 ierr = PetscFree4(submatj->sbuf1,submatj->ptr,submatj->tmp,submatj->ctr);CHKERRQ(ierr); 615 616 for (i=0; i<submatj->nrqr; ++i) { 617 ierr = PetscFree(submatj->sbuf2[i]);CHKERRQ(ierr); 618 } 619 ierr = PetscFree3(submatj->sbuf2,submatj->req_size,submatj->req_source1);CHKERRQ(ierr); 620 621 if (submatj->rbuf1) { 622 ierr = PetscFree(submatj->rbuf1[0]);CHKERRQ(ierr); 623 ierr = PetscFree(submatj->rbuf1);CHKERRQ(ierr); 624 } 625 626 for (i=0; i<submatj->nrqs; ++i) { 627 ierr = PetscFree(submatj->rbuf3[i]);CHKERRQ(ierr); 628 } 629 ierr = PetscFree3(submatj->req_source2,submatj->rbuf2,submatj->rbuf3);CHKERRQ(ierr); 630 ierr = PetscFree(submatj->pa);CHKERRQ(ierr); 631 } 632 633 #if defined(PETSC_USE_CTABLE) 634 ierr = PetscTableDestroy((PetscTable*)&submatj->rmap);CHKERRQ(ierr); 635 if (submatj->cmap_loc) {ierr = PetscFree(submatj->cmap_loc);CHKERRQ(ierr);} 636 ierr = PetscFree(submatj->rmap_loc);CHKERRQ(ierr); 637 #else 638 ierr = PetscFree(submatj->rmap);CHKERRQ(ierr); 639 #endif 640 641 if (!submatj->allcolumns) { 642 #if defined(PETSC_USE_CTABLE) 643 ierr = PetscTableDestroy((PetscTable*)&submatj->cmap);CHKERRQ(ierr); 644 #else 645 ierr = PetscFree(submatj->cmap);CHKERRQ(ierr); 646 #endif 647 } 648 ierr = submatj->destroy(C);CHKERRQ(ierr); 649 ierr = PetscFree(submatj->row2proc);CHKERRQ(ierr); 650 651 ierr = PetscFree(submatj);CHKERRQ(ierr); 652 PetscFunctionReturn(0); 653 } 654 655 PetscErrorCode MatGetSubMatrices_MPIBAIJ_local_new(Mat C,PetscInt ismax,const IS isrow[],const IS iscol[],MatReuse scall,PetscBool *allrows,PetscBool *allcolumns,Mat *submats) 656 { 657 Mat_MPIBAIJ *c = (Mat_MPIBAIJ*)C->data; 658 Mat A = c->A; 659 Mat_SeqBAIJ *a = (Mat_SeqBAIJ*)A->data,*b = (Mat_SeqBAIJ*)c->B->data,*subc; 660 const PetscInt **icol,**irow; 661 PetscInt *nrow,*ncol,start; 662 PetscErrorCode ierr; 663 PetscMPIInt rank,size,tag0,tag2,tag3,tag4,*w1,*w2,*w3,*w4,nrqr; 664 PetscInt **sbuf1,**sbuf2,i,j,k,l,ct1,ct2,**rbuf1,row,proc=-1; 665 PetscInt nrqs=0,msz,**ptr=NULL,*req_size=NULL,*ctr=NULL,*pa,*tmp=NULL,tcol; 666 PetscInt **rbuf3=NULL,*req_source1=NULL,*req_source2,**sbuf_aj,**rbuf2=NULL,max1,max2; 667 PetscInt **lens,is_no,ncols,*cols,mat_i,*mat_j,tmp2,jmax; 668 #if defined(PETSC_USE_CTABLE) 669 PetscTable *cmap,cmap_i=NULL,*rmap,rmap_i; 670 #else 671 PetscInt **cmap,*cmap_i=NULL,**rmap,*rmap_i; 672 #endif 673 const PetscInt *irow_i; 674 PetscInt ctr_j,*sbuf1_j,*sbuf_aj_i,*rbuf1_i,kmax,*lens_i; 675 MPI_Request *s_waits1,*r_waits1,*s_waits2,*r_waits2,*r_waits3; 676 MPI_Request *r_waits4,*s_waits3,*s_waits4; 677 MPI_Status *r_status1,*r_status2,*s_status1,*s_status3,*s_status2; 678 MPI_Status *r_status3,*r_status4,*s_status4; 679 MPI_Comm comm; 680 PetscScalar **rbuf4,*rbuf4_i,**sbuf_aa,*vals,*mat_a,*imat_a,*sbuf_aa_i; 681 PetscMPIInt *onodes1,*olengths1,end; 682 PetscInt **row2proc,*row2proc_i,ilen_row,*imat_ilen,*imat_j,*imat_i,old_row; 683 Mat_SubMat **smats,*smat_i; 684 PetscBool *issorted,colflag,iscsorted=PETSC_TRUE; 685 PetscInt *sbuf1_i,*rbuf2_i,*rbuf3_i,ilen; 686 687 PetscInt bs=C->rmap->bs,bs2=c->bs2,rstart = c->rstartbs; 688 PetscBool ijonly=c->ijonly; /* private flag indicates only matrix data structures are requested */ 689 PetscInt nzA,nzB,*a_i=a->i,*b_i=b->i,*a_j = a->j,*b_j = b->j,ctmp,imark,*cworkA,*cworkB; 690 PetscScalar *vworkA,*vworkB,*a_a = a->a,*b_a = b->a; 691 PetscInt cstart = c->cstartbs,*bmap = c->garray; 692 693 PetscFunctionBegin; 694 ierr = PetscObjectGetComm((PetscObject)C,&comm);CHKERRQ(ierr); 695 size = c->size; 696 rank = c->rank; 697 if (!rank) { 698 printf("MatGetSubMatrices_MPIBAIJ_new scall %d, bs %d\n",scall,bs); 699 } 700 701 ierr = PetscMalloc5(ismax,&smats,ismax,&row2proc,ismax,&cmap,ismax,&rmap,ismax,&allcolumns);CHKERRQ(ierr); 702 ierr = PetscMalloc5(ismax,&irow,ismax,&icol,ismax,&nrow,ismax,&ncol,ismax,&issorted);CHKERRQ(ierr); 703 704 for (i=0; i<ismax; i++) { 705 ierr = ISSorted(iscol[i],&issorted[i]);CHKERRQ(ierr); 706 if (!issorted[i]) iscsorted = issorted[i]; 707 708 ierr = ISSorted(isrow[i],&issorted[i]);CHKERRQ(ierr); 709 710 ierr = ISGetIndices(isrow[i],&irow[i]);CHKERRQ(ierr); 711 ierr = ISGetLocalSize(isrow[i],&nrow[i]);CHKERRQ(ierr); 712 713 /* Check for special case: allcolumn */ 714 ierr = ISIdentity(iscol[i],&colflag);CHKERRQ(ierr); 715 ierr = ISGetLocalSize(iscol[i],&ncol[i]);CHKERRQ(ierr); 716 if (colflag && ncol[i] == C->cmap->N) { 717 allcolumns[i] = PETSC_TRUE; 718 icol[i] = NULL; 719 } else { 720 allcolumns[i] = PETSC_FALSE; 721 ierr = ISGetIndices(iscol[i],&icol[i]);CHKERRQ(ierr); 722 } 723 } 724 725 if (scall == MAT_REUSE_MATRIX) { 726 /* Assumes new rows are same length as the old rows */ 727 for (i=0; i<ismax; i++) { 728 subc = (Mat_SeqBAIJ*)(submats[i]->data); 729 if ((submats[i]->rmap->n != nrow[i]) || (submats[i]->cmap->n != ncol[i])) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Cannot reuse matrix. wrong size"); 730 731 /* Initial matrix as if empty */ 732 ierr = PetscMemzero(subc->ilen,submats[i]->rmap->n*sizeof(PetscInt));CHKERRQ(ierr); 733 734 /* Initial matrix as if empty */ 735 submats[i]->factortype = C->factortype; 736 737 smat_i = subc->submatis1; 738 smats[i] = smat_i; 739 740 nrqs = smat_i->nrqs; 741 nrqr = smat_i->nrqr; 742 rbuf1 = smat_i->rbuf1; 743 rbuf2 = smat_i->rbuf2; 744 rbuf3 = smat_i->rbuf3; 745 req_source2 = smat_i->req_source2; 746 747 sbuf1 = smat_i->sbuf1; 748 sbuf2 = smat_i->sbuf2; 749 ptr = smat_i->ptr; 750 tmp = smat_i->tmp; 751 ctr = smat_i->ctr; 752 753 pa = smat_i->pa; 754 req_size = smat_i->req_size; 755 req_source1 = smat_i->req_source1; 756 757 allcolumns[i] = smat_i->allcolumns; 758 row2proc[i] = smat_i->row2proc; 759 rmap[i] = smat_i->rmap; 760 cmap[i] = smat_i->cmap; 761 } 762 } else { /* scall == MAT_INITIAL_MATRIX */ 763 /* Get some new tags to keep the communication clean */ 764 ierr = PetscObjectGetNewTag((PetscObject)C,&tag2);CHKERRQ(ierr); 765 ierr = PetscObjectGetNewTag((PetscObject)C,&tag3);CHKERRQ(ierr); 766 767 /* evaluate communication - mesg to who, length of mesg, and buffer space 768 required. Based on this, buffers are allocated, and data copied into them*/ 769 ierr = PetscCalloc4(size,&w1,size,&w2,size,&w3,size,&w4);CHKERRQ(ierr); /* mesg size, initialize work vectors */ 770 771 for (i=0; i<ismax; i++) { 772 jmax = nrow[i]; 773 irow_i = irow[i]; 774 775 ierr = PetscMalloc1(jmax,&row2proc_i);CHKERRQ(ierr); 776 row2proc[i] = row2proc_i; 777 778 if (issorted[i]) proc = 0; 779 for (j=0; j<jmax; j++) { 780 if (!issorted[i]) proc = 0; 781 row = irow_i[j]; 782 //while (row >= C->rmap->range[proc+1]) proc++; 783 while (row >= c->rangebs[proc+1]) proc++; 784 w4[proc]++; 785 row2proc_i[j] = proc; /* map row index to proc */ 786 } 787 for (j=0; j<size; j++) { 788 if (w4[j]) { w1[j] += w4[j]; w3[j]++; w4[j] = 0;} 789 } 790 } 791 792 nrqs = 0; /* no of outgoing messages */ 793 msz = 0; /* total mesg length (for all procs) */ 794 w1[rank] = 0; /* no mesg sent to self */ 795 w3[rank] = 0; 796 for (i=0; i<size; i++) { 797 if (w1[i]) { w2[i] = 1; nrqs++;} /* there exists a message to proc i */ 798 } 799 ierr = PetscMalloc1(nrqs+1,&pa);CHKERRQ(ierr); /*(proc -array)*/ 800 for (i=0,j=0; i<size; i++) { 801 if (w1[i]) { pa[j] = i; j++; } 802 } 803 804 /* Each message would have a header = 1 + 2*(no of IS) + data */ 805 for (i=0; i<nrqs; i++) { 806 j = pa[i]; 807 w1[j] += w2[j] + 2* w3[j]; 808 msz += w1[j]; 809 } 810 ierr = PetscInfo2(0,"Number of outgoing messages %D Total message length %D\n",nrqs,msz);CHKERRQ(ierr); 811 812 /* Determine the number of messages to expect, their lengths, from from-ids */ 813 ierr = PetscGatherNumberOfMessages(comm,w2,w1,&nrqr);CHKERRQ(ierr); 814 ierr = PetscGatherMessageLengths(comm,nrqs,nrqr,w1,&onodes1,&olengths1);CHKERRQ(ierr); 815 printf("[%d] nrqs %d, nrqr %d\n",rank,nrqs,nrqr); 816 817 /* Now post the Irecvs corresponding to these messages */ 818 tag0 = ((PetscObject)C)->tag; 819 ierr = PetscPostIrecvInt(comm,tag0,nrqr,onodes1,olengths1,&rbuf1,&r_waits1);CHKERRQ(ierr); 820 if (rank == 1) { 821 for (i=0; i<nrqr; i++) printf(" onodes1 [%d] olengths1 %d\n",onodes1[i],olengths1[i]); 822 } 823 824 ierr = PetscFree(onodes1);CHKERRQ(ierr); 825 ierr = PetscFree(olengths1);CHKERRQ(ierr); 826 827 /* Allocate Memory for outgoing messages */ 828 ierr = PetscMalloc4(size,&sbuf1,size,&ptr,2*msz,&tmp,size,&ctr);CHKERRQ(ierr); 829 ierr = PetscMemzero(sbuf1,size*sizeof(PetscInt*));CHKERRQ(ierr); 830 ierr = PetscMemzero(ptr,size*sizeof(PetscInt*));CHKERRQ(ierr); 831 832 { 833 PetscInt *iptr = tmp; 834 k = 0; 835 for (i=0; i<nrqs; i++) { 836 j = pa[i]; 837 iptr += k; 838 sbuf1[j] = iptr; 839 k = w1[j]; 840 } 841 } 842 843 /* Form the outgoing messages. Initialize the header space */ 844 for (i=0; i<nrqs; i++) { 845 j = pa[i]; 846 sbuf1[j][0] = 0; 847 ierr = PetscMemzero(sbuf1[j]+1,2*w3[j]*sizeof(PetscInt));CHKERRQ(ierr); 848 ptr[j] = sbuf1[j] + 2*w3[j] + 1; 849 } 850 851 /* Parse the isrow and copy data into outbuf */ 852 for (i=0; i<ismax; i++) { 853 row2proc_i = row2proc[i]; 854 ierr = PetscMemzero(ctr,size*sizeof(PetscInt));CHKERRQ(ierr); 855 irow_i = irow[i]; 856 jmax = nrow[i]; 857 for (j=0; j<jmax; j++) { /* parse the indices of each IS */ 858 proc = row2proc_i[j]; 859 if (proc != rank) { /* copy to the outgoing buf*/ 860 ctr[proc]++; 861 *ptr[proc] = irow_i[j]; 862 ptr[proc]++; 863 } 864 } 865 /* Update the headers for the current IS */ 866 for (j=0; j<size; j++) { /* Can Optimise this loop too */ 867 if ((ctr_j = ctr[j])) { 868 sbuf1_j = sbuf1[j]; 869 k = ++sbuf1_j[0]; 870 sbuf1_j[2*k] = ctr_j; 871 sbuf1_j[2*k-1] = i; 872 } 873 } 874 } 875 876 /* Now post the sends */ 877 ierr = PetscMalloc1(nrqs+1,&s_waits1);CHKERRQ(ierr); 878 for (i=0; i<nrqs; ++i) { 879 j = pa[i]; 880 printf("[%d] sends sbuf1 %d to [%d]\n",rank,w1[j],j); 881 ierr = MPI_Isend(sbuf1[j],w1[j],MPIU_INT,j,tag0,comm,s_waits1+i);CHKERRQ(ierr); 882 } 883 884 /* Post Receives to capture the buffer size */ 885 ierr = PetscMalloc1(nrqs+1,&r_waits2);CHKERRQ(ierr); 886 ierr = PetscMalloc3(nrqs+1,&req_source2,nrqs+1,&rbuf2,nrqs+1,&rbuf3);CHKERRQ(ierr); 887 rbuf2[0] = tmp + msz; 888 for (i=1; i<nrqs; ++i) { 889 rbuf2[i] = rbuf2[i-1]+w1[pa[i-1]]; 890 } 891 for (i=0; i<nrqs; ++i) { 892 j = pa[i]; 893 ierr = MPI_Irecv(rbuf2[i],w1[j],MPIU_INT,j,tag2,comm,r_waits2+i);CHKERRQ(ierr); 894 } 895 896 /* Send to other procs the buf size they should allocate */ 897 /* Receive messages*/ 898 ierr = PetscMalloc1(nrqr+1,&s_waits2);CHKERRQ(ierr); 899 ierr = PetscMalloc1(nrqr+1,&r_status1);CHKERRQ(ierr); 900 ierr = PetscMalloc3(nrqr,&sbuf2,nrqr,&req_size,nrqr,&req_source1);CHKERRQ(ierr); 901 { 902 PetscInt *sAi = a->i,*sBi = b->i,id; 903 PetscInt *sbuf2_i; 904 905 ierr = MPI_Waitall(nrqr,r_waits1,r_status1);CHKERRQ(ierr); 906 for (i=0; i<nrqr; ++i) { 907 req_size[i] = 0; 908 rbuf1_i = rbuf1[i]; 909 start = 2*rbuf1_i[0] + 1; 910 ierr = MPI_Get_count(r_status1+i,MPIU_INT,&end);CHKERRQ(ierr); 911 ierr = PetscMalloc1(end+1,&sbuf2[i]);CHKERRQ(ierr); 912 sbuf2_i = sbuf2[i]; 913 for (j=start; j<end; j++) { 914 id = rbuf1_i[j] - rstart; 915 ncols = sAi[id+1] - sAi[id] + sBi[id+1] - sBi[id]; 916 sbuf2_i[j] = ncols; 917 req_size[i] += ncols; 918 } 919 req_source1[i] = r_status1[i].MPI_SOURCE; 920 /* form the header */ 921 sbuf2_i[0] = req_size[i]; 922 for (j=1; j<start; j++) sbuf2_i[j] = rbuf1_i[j]; 923 924 ierr = MPI_Isend(sbuf2_i,end,MPIU_INT,req_source1[i],tag2,comm,s_waits2+i);CHKERRQ(ierr); 925 } 926 } 927 ierr = PetscFree(r_status1);CHKERRQ(ierr); 928 ierr = PetscFree(r_waits1);CHKERRQ(ierr); 929 ierr = PetscFree4(w1,w2,w3,w4);CHKERRQ(ierr); 930 931 /* Receive messages*/ 932 ierr = PetscMalloc1(nrqs+1,&r_waits3);CHKERRQ(ierr); 933 ierr = PetscMalloc1(nrqs+1,&r_status2);CHKERRQ(ierr); 934 935 ierr = MPI_Waitall(nrqs,r_waits2,r_status2);CHKERRQ(ierr); 936 for (i=0; i<nrqs; ++i) { 937 ierr = PetscMalloc1(rbuf2[i][0]+1,&rbuf3[i]);CHKERRQ(ierr); 938 req_source2[i] = r_status2[i].MPI_SOURCE; 939 ierr = MPI_Irecv(rbuf3[i],rbuf2[i][0],MPIU_INT,req_source2[i],tag3,comm,r_waits3+i);CHKERRQ(ierr); 940 } 941 ierr = PetscFree(r_status2);CHKERRQ(ierr); 942 ierr = PetscFree(r_waits2);CHKERRQ(ierr); 943 944 /* Wait on sends1 and sends2 */ 945 ierr = PetscMalloc1(nrqs+1,&s_status1);CHKERRQ(ierr); 946 ierr = PetscMalloc1(nrqr+1,&s_status2);CHKERRQ(ierr); 947 948 if (nrqs) {ierr = MPI_Waitall(nrqs,s_waits1,s_status1);CHKERRQ(ierr);} 949 if (nrqr) {ierr = MPI_Waitall(nrqr,s_waits2,s_status2);CHKERRQ(ierr);} 950 ierr = PetscFree(s_status1);CHKERRQ(ierr); 951 ierr = PetscFree(s_status2);CHKERRQ(ierr); 952 ierr = PetscFree(s_waits1);CHKERRQ(ierr); 953 ierr = PetscFree(s_waits2);CHKERRQ(ierr); 954 955 /* Now allocate sending buffers for a->j, and send them off */ 956 ierr = PetscMalloc1(nrqr+1,&sbuf_aj);CHKERRQ(ierr); 957 for (i=0,j=0; i<nrqr; i++) j += req_size[i]; 958 ierr = PetscMalloc1(j+1,&sbuf_aj[0]);CHKERRQ(ierr); 959 for (i=1; i<nrqr; i++) sbuf_aj[i] = sbuf_aj[i-1] + req_size[i-1]; 960 961 ierr = PetscMalloc1(nrqr+1,&s_waits3);CHKERRQ(ierr); 962 { 963 //PetscInt *a_i = a->i,*b_i = b->i; 964 //PetscInt cstart = c->cstartbs,*bmap = c->garray; 965 //PetscInt cend = c->cendbs; 966 //PetscInt *a_j = a->j,*b_j = b->j,ctmp,imark; 967 968 for (i=0; i<nrqr; i++) { 969 rbuf1_i = rbuf1[i]; 970 sbuf_aj_i = sbuf_aj[i]; 971 ct1 = 2*rbuf1_i[0] + 1; 972 ct2 = 0; 973 for (j=1,max1=rbuf1_i[0]; j<=max1; j++) { 974 kmax = rbuf1[i][2*j]; 975 for (k=0; k<kmax; k++,ct1++) { 976 row = rbuf1_i[ct1] - rstart; 977 nzA = a_i[row+1] - a_i[row]; nzB = b_i[row+1] - b_i[row]; 978 ncols = nzA + nzB; 979 cworkA = a_j + a_i[row]; cworkB = b_j + b_i[row]; 980 981 /* load the column indices for this row into cols */ 982 cols = sbuf_aj_i + ct2; 983 for (l=0; l<nzB; l++) { 984 if ((ctmp = bmap[cworkB[l]]) < cstart) cols[l] = ctmp; 985 else break; 986 } 987 imark = l; 988 for (l=0; l<nzA; l++) {cols[imark+l] = cstart + cworkA[l];} 989 for (l=imark; l<nzB; l++) cols[nzA+l] = bmap[cworkB[l]]; 990 ct2 += ncols; 991 } 992 } 993 ierr = MPI_Isend(sbuf_aj_i,req_size[i],MPIU_INT,req_source1[i],tag3,comm,s_waits3+i);CHKERRQ(ierr); 994 } 995 } 996 ierr = PetscMalloc2(nrqs+1,&r_status3,nrqr+1,&s_status3);CHKERRQ(ierr); 997 998 /* create col map: global col of C -> local col of submatrices */ 999 { 1000 const PetscInt *icol_i; 1001 #if defined(PETSC_USE_CTABLE) 1002 for (i=0; i<ismax; i++) { 1003 if (!allcolumns[i]) { 1004 //ierr = PetscTableCreate(ncol[i]+1,C->cmap->N+1,&cmap[i]);CHKERRQ(ierr); 1005 ierr = PetscTableCreate(ncol[i]+1,c->Nbs+1,&cmap[i]);CHKERRQ(ierr); 1006 1007 jmax = ncol[i]; 1008 icol_i = icol[i]; 1009 cmap_i = cmap[i]; 1010 for (j=0; j<jmax; j++) { 1011 ierr = PetscTableAdd(cmap[i],icol_i[j]+1,j+1,INSERT_VALUES);CHKERRQ(ierr); 1012 } 1013 } else cmap[i] = NULL; 1014 } 1015 #else 1016 for (i=0; i<ismax; i++) { 1017 if (!allcolumns[i]) { 1018 ierr = PetscCalloc1(c->Nbs,&cmap[i]);CHKERRQ(ierr); 1019 jmax = ncol[i]; 1020 icol_i = icol[i]; 1021 cmap_i = cmap[i]; 1022 for (j=0; j<jmax; j++) cmap_i[icol_i[j]] = j+1; 1023 } else cmap[i] = NULL; 1024 } 1025 #endif 1026 } 1027 1028 /* Create lens which is required for MatCreate... */ 1029 for (i=0,j=0; i<ismax; i++) j += nrow[i]; 1030 ierr = PetscMalloc1(ismax,&lens);CHKERRQ(ierr); 1031 1032 if (ismax) { 1033 ierr = PetscCalloc1(j,&lens[0]);CHKERRQ(ierr); 1034 } 1035 for (i=1; i<ismax; i++) lens[i] = lens[i-1] + nrow[i-1]; 1036 1037 /* Update lens from local data */ 1038 for (i=0; i<ismax; i++) { 1039 row2proc_i = row2proc[i]; 1040 jmax = nrow[i]; 1041 if (!allcolumns[i]) cmap_i = cmap[i]; 1042 irow_i = irow[i]; 1043 lens_i = lens[i]; 1044 for (j=0; j<jmax; j++) { 1045 row = irow_i[j]; /* global blocked row of C */ 1046 proc = row2proc_i[j]; 1047 //if (!rank) printf("j %d, proc %d\n",j,proc); 1048 if (proc == rank) { 1049 1050 /* Get indices from matA and then from matB */ 1051 PetscInt nzA,nzB,*cworkA,*cworkB,cstart = c->cstartbs,*bmap = c->garray; 1052 PetscInt *a_i = a->i,*b_i=b->i,*a_j = a->j,*b_j = b->j; 1053 #if defined(PETSC_USE_CTABLE) 1054 PetscInt tt; 1055 #endif 1056 row = row - rstart; 1057 nzA = a_i[row+1] - a_i[row]; 1058 nzB = b_i[row+1] - b_i[row]; 1059 cworkA = a_j + a_i[row]; 1060 cworkB = b_j + b_i[row]; 1061 1062 //printf("[%d] row_loc %d\n",rank,row); 1063 if (!allcolumns[i]) { 1064 #if defined(PETSC_USE_CTABLE) 1065 for (k=0; k<nzA; k++) { 1066 ierr = PetscTableFind(cmap_i,cstart+cworkA[k]+1,&tt);CHKERRQ(ierr); 1067 if (tt) lens_i[j]++; 1068 } 1069 for (k=0; k<nzB; k++) { 1070 ierr = PetscTableFind(cmap_i,bmap[cworkB[k]]+1,&tt);CHKERRQ(ierr); 1071 if (tt) lens_i[j]++; 1072 } 1073 1074 #else 1075 for (k=0; k<nzA; k++) { 1076 if (cmap_i[cstart + cworkA[k]]) lens_i[j]++; 1077 } 1078 for (k=0; k<nzB; k++) { 1079 if (cmap_i[bmap[cworkB[k]]]) lens_i[j]++; 1080 } 1081 #endif 1082 } else { /* allcolumns */ 1083 lens_i[j] = nzA + nzB; 1084 } 1085 1086 #if 0 1087 //------------------------------------------ 1088 /* Get indices from matA and then from matB */ 1089 PetscInt *a_i = a->i,*b_i=b->i; 1090 PetscInt row_loc = row - rstart; /* local row of matA and matB */ 1091 PetscInt nzA = a_i[row_loc+1] - a_i[row_loc]; 1092 PetscInt nzB = b_i[row_loc+1] - b_i[row_loc]; 1093 1094 printf("[%d] row %d, row_loc %d\n",rank,row,row_loc); 1095 ierr = MatGetRow_MPIBAIJ(C,row*bs,&ncols,&cols,0);CHKERRQ(ierr); 1096 if (nzA + nzB != ncols/bs) SETERRQ3(PETSC_COMM_SELF,0,"nzA %d + nzB %d != ncols/bs %d",nzA,nzB,ncols/bs); 1097 1098 if (!allcolumns[i]) { 1099 //for (k=0; k<ncols; k++) { 1100 for (k=0; k<ncols; k+=bs) { 1101 #if defined(PETSC_USE_CTABLE) 1102 ierr = PetscTableFind(cmap_i,cols[k]/bs+1,&tcol);CHKERRQ(ierr); 1103 #else 1104 tcol = cmap_i[cols[k]]; 1105 #endif 1106 if (tcol) lens_i[j]++; 1107 } 1108 } else { /* allcolumns */ 1109 lens_i[j] = ncols; 1110 } 1111 ierr = MatRestoreRow_MPIBAIJ(C,row*bs,&ncols,&cols,0);CHKERRQ(ierr); 1112 #endif //------------------------------------------------------ 1113 } 1114 } 1115 } 1116 1117 /* Create row map: global row of C -> local row of submatrices */ 1118 #if defined(PETSC_USE_CTABLE) 1119 for (i=0; i<ismax; i++) { 1120 //ierr = PetscTableCreate(nrow[i]+1,C->rmap->N+1,&rmap[i]);CHKERRQ(ierr); 1121 ierr = PetscTableCreate(nrow[i]+1,c->Mbs+1,&rmap[i]);CHKERRQ(ierr); 1122 irow_i = irow[i]; 1123 jmax = nrow[i]; 1124 for (j=0; j<jmax; j++) { 1125 ierr = PetscTableAdd(rmap[i],irow_i[j]+1,j+1,INSERT_VALUES);CHKERRQ(ierr); 1126 } 1127 } 1128 #else 1129 for (i=0; i<ismax; i++) { 1130 ierr = PetscCalloc1(c->Mbs,&rmap[i]);CHKERRQ(ierr); 1131 rmap_i = rmap[i]; 1132 irow_i = irow[i]; 1133 jmax = nrow[i]; 1134 for (j=0; j<jmax; j++) { 1135 rmap_i[irow_i[j]] = j; 1136 } 1137 } 1138 #endif 1139 1140 /* Update lens from offproc data */ 1141 { 1142 PetscInt *rbuf2_i,*rbuf3_i,*sbuf1_i; 1143 1144 ierr = MPI_Waitall(nrqs,r_waits3,r_status3);CHKERRQ(ierr); 1145 for (tmp2=0; tmp2<nrqs; tmp2++) { 1146 sbuf1_i = sbuf1[pa[tmp2]]; 1147 jmax = sbuf1_i[0]; 1148 ct1 = 2*jmax+1; 1149 ct2 = 0; 1150 rbuf2_i = rbuf2[tmp2]; 1151 rbuf3_i = rbuf3[tmp2]; 1152 for (j=1; j<=jmax; j++) { 1153 is_no = sbuf1_i[2*j-1]; 1154 max1 = sbuf1_i[2*j]; 1155 lens_i = lens[is_no]; 1156 if (!allcolumns[is_no]) cmap_i = cmap[is_no]; 1157 rmap_i = rmap[is_no]; 1158 for (k=0; k<max1; k++,ct1++) { 1159 #if defined(PETSC_USE_CTABLE) 1160 ierr = PetscTableFind(rmap_i,sbuf1_i[ct1]+1,&row);CHKERRQ(ierr); 1161 row--; 1162 if (row < 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"row not found in table"); 1163 #else 1164 row = rmap_i[sbuf1_i[ct1]]; /* the val in the new matrix to be */ 1165 #endif 1166 max2 = rbuf2_i[ct1]; 1167 for (l=0; l<max2; l++,ct2++) { 1168 if (!allcolumns[is_no]) { 1169 #if defined(PETSC_USE_CTABLE) 1170 ierr = PetscTableFind(cmap_i,rbuf3_i[ct2]+1,&tcol);CHKERRQ(ierr); 1171 #else 1172 tcol = cmap_i[rbuf3_i[ct2]]; 1173 #endif 1174 if (tcol) lens_i[row]++; 1175 } else { /* allcolumns */ 1176 lens_i[row]++; /* lens_i[row] += max2 ? */ 1177 } 1178 } 1179 } 1180 } 1181 } 1182 } 1183 ierr = PetscFree(r_waits3);CHKERRQ(ierr); 1184 if (nrqr) {ierr = MPI_Waitall(nrqr,s_waits3,s_status3);CHKERRQ(ierr);} 1185 ierr = PetscFree2(r_status3,s_status3);CHKERRQ(ierr); 1186 ierr = PetscFree(s_waits3);CHKERRQ(ierr); 1187 1188 /* Create the submatrices */ 1189 for (i=0; i<ismax; i++) { 1190 PetscInt bs_tmp; 1191 if (ijonly) bs_tmp = 1; 1192 else bs_tmp = bs; 1193 1194 ierr = MatCreate(PETSC_COMM_SELF,submats+i);CHKERRQ(ierr); 1195 ierr = MatSetSizes(submats[i],nrow[i]*bs_tmp,ncol[i]*bs_tmp,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 1196 1197 ierr = MatSetType(submats[i],((PetscObject)A)->type_name);CHKERRQ(ierr); 1198 ierr = MatSeqBAIJSetPreallocation(submats[i],bs_tmp,0,lens[i]);CHKERRQ(ierr); 1199 ierr = MatSeqSBAIJSetPreallocation(submats[i],bs_tmp,0,lens[i]);CHKERRQ(ierr); /* this subroutine is used by SBAIJ routines */ 1200 1201 /* create struct Mat_SubMat and attached it to submat */ 1202 ierr = PetscNew(&smat_i);CHKERRQ(ierr); 1203 subc = (Mat_SeqBAIJ*)submats[i]->data; 1204 subc->submatis1 = smat_i; 1205 smats[i] = smat_i; 1206 1207 smat_i->destroy = submats[i]->ops->destroy; 1208 submats[i]->ops->destroy = MatDestroy_MPIBAIJ_MatGetSubmatrices; 1209 submats[i]->factortype = C->factortype; 1210 1211 smat_i->id = i; 1212 smat_i->nrqs = nrqs; 1213 smat_i->nrqr = nrqr; 1214 smat_i->rbuf1 = rbuf1; 1215 smat_i->rbuf2 = rbuf2; 1216 smat_i->rbuf3 = rbuf3; 1217 smat_i->sbuf2 = sbuf2; 1218 smat_i->req_source2 = req_source2; 1219 1220 smat_i->sbuf1 = sbuf1; 1221 smat_i->ptr = ptr; 1222 smat_i->tmp = tmp; 1223 smat_i->ctr = ctr; 1224 1225 smat_i->pa = pa; 1226 smat_i->req_size = req_size; 1227 smat_i->req_source1 = req_source1; 1228 1229 smat_i->allcolumns = allcolumns[i]; 1230 smat_i->singleis = PETSC_FALSE; 1231 smat_i->row2proc = row2proc[i]; 1232 smat_i->rmap = rmap[i]; 1233 smat_i->cmap = cmap[i]; 1234 } 1235 1236 if (ismax) {ierr = PetscFree(lens[0]);CHKERRQ(ierr);} 1237 ierr = PetscFree(lens);CHKERRQ(ierr); 1238 ierr = PetscFree(sbuf_aj[0]);CHKERRQ(ierr); 1239 ierr = PetscFree(sbuf_aj);CHKERRQ(ierr); 1240 1241 } /* endof scall == MAT_INITIAL_MATRIX */ 1242 1243 /* Post recv matrix values */ 1244 ierr = PetscObjectGetNewTag((PetscObject)C,&tag4);CHKERRQ(ierr); 1245 ierr = PetscMalloc1(nrqs+1,&rbuf4);CHKERRQ(ierr); 1246 ierr = PetscMalloc1(nrqs+1,&r_waits4);CHKERRQ(ierr); 1247 ierr = PetscMalloc1(nrqs+1,&r_status4);CHKERRQ(ierr); 1248 ierr = PetscMalloc1(nrqr+1,&s_status4);CHKERRQ(ierr); 1249 for (i=0; i<nrqs; ++i) { 1250 ierr = PetscMalloc1(rbuf2[i][0]*bs2,&rbuf4[i]);CHKERRQ(ierr); 1251 ierr = MPI_Irecv(rbuf4[i],rbuf2[i][0]*bs2,MPIU_SCALAR,req_source2[i],tag4,comm,r_waits4+i);CHKERRQ(ierr); 1252 } 1253 1254 /* Allocate sending buffers for a->a, and send them off */ 1255 ierr = PetscMalloc1(nrqr+1,&sbuf_aa);CHKERRQ(ierr); 1256 for (i=0,j=0; i<nrqr; i++) j += req_size[i]; 1257 //ierr = PetscMalloc1(j+1,&sbuf_aa[0]);CHKERRQ(ierr); 1258 //for (i=1; i<nrqr; i++) sbuf_aa[i] = sbuf_aa[i-1] + req_size[i-1]; 1259 ierr = PetscMalloc1((j+1)*bs2,&sbuf_aa[0]);CHKERRQ(ierr); 1260 for (i=1; i<nrqr; i++) sbuf_aa[i] = sbuf_aa[i-1] + req_size[i-1]*bs2; 1261 1262 ierr = PetscMalloc1(nrqr+1,&s_waits4);CHKERRQ(ierr); 1263 { 1264 //PetscInt nzA,nzB,*a_i = a->i,*b_i = b->i, *cworkB; 1265 PetscInt cstart = c->cstartbs,*bmap = c->garray; 1266 PetscInt *b_j = b->j; 1267 //PetscScalar *vworkA,*vworkB,*a_a = a->a,*b_a = b->a; 1268 PetscInt imark; 1269 1270 for (i=0; i<nrqr; i++) { 1271 rbuf1_i = rbuf1[i]; 1272 sbuf_aa_i = sbuf_aa[i]; 1273 ct1 = 2*rbuf1_i[0]+1; 1274 ct2 = 0; 1275 for (j=1,max1=rbuf1_i[0]; j<=max1; j++) { 1276 kmax = rbuf1_i[2*j]; 1277 for (k=0; k<kmax; k++,ct1++) { 1278 row = rbuf1_i[ct1] - rstart; 1279 nzA = a_i[row+1] - a_i[row]; 1280 nzB = b_i[row+1] - b_i[row]; 1281 ncols = nzA + nzB; 1282 cworkB = b_j + b_i[row]; 1283 vworkA = a_a + a_i[row]*bs2; 1284 vworkB = b_a + b_i[row]*bs2; 1285 //if (rank==1) printf("i %d, j %d, k %d, ncols %d\n",i,j,k,ncols); 1286 1287 /* load the column values for this row into vals*/ 1288 vals = sbuf_aa_i+ct2*bs2; 1289 for (l=0; l<nzB; l++) { 1290 if ((bmap[cworkB[l]]) < cstart) { 1291 ierr = PetscMemcpy(vals+l*bs2,vworkB+l*bs2,bs2*sizeof(MatScalar));CHKERRQ(ierr); 1292 } else break; 1293 } 1294 imark = l; 1295 for (l=0; l<nzA; l++) { 1296 ierr = PetscMemcpy(vals+(imark+l)*bs2,vworkA+l*bs2,bs2*sizeof(MatScalar));CHKERRQ(ierr);//error in proc[1]??? 1297 } 1298 for (l=imark; l<nzB; l++) { 1299 ierr = PetscMemcpy(vals+(nzA+l)*bs2,vworkB+l*bs2,bs2*sizeof(MatScalar));CHKERRQ(ierr); 1300 } 1301 #if 0 1302 vals = sbuf_aa_i+ct2; 1303 1304 lwrite = 0; 1305 for (l=0; l<nzB; l++) { 1306 if ((bmap[cworkB[l]]) < cstart) vals[lwrite++] = vworkB[l]; 1307 } 1308 for (l=0; l<nzA; l++) vals[lwrite++] = vworkA[l]; 1309 for (l=0; l<nzB; l++) { 1310 if ((bmap[cworkB[l]]) >= cend) vals[lwrite++] = vworkB[l]; 1311 } 1312 #endif 1313 ct2 += ncols; 1314 } 1315 } 1316 printf("[%d] send %d aa to [%d] \n",rank,req_size[i]*bs2,req_source1[i]); 1317 ierr = MPI_Isend(sbuf_aa_i,req_size[i]*bs2,MPIU_SCALAR,req_source1[i],tag4,comm,s_waits4+i);CHKERRQ(ierr); 1318 } 1319 } 1320 ierr = MPI_Barrier(comm);CHKERRQ(ierr); 1321 1322 if (!ismax) { 1323 ierr = PetscFree(rbuf1[0]);CHKERRQ(ierr); 1324 ierr = PetscFree(rbuf1);CHKERRQ(ierr); 1325 } 1326 1327 /* Assemble the matrices */ 1328 /* First assemble the local rows */ 1329 for (i=0; i<ismax; i++) { 1330 row2proc_i = row2proc[i]; 1331 subc = (Mat_SeqBAIJ*)submats[i]->data; 1332 imat_ilen = subc->ilen; 1333 imat_j = subc->j; 1334 imat_i = subc->i; 1335 imat_a = subc->a; 1336 1337 if (!allcolumns[i]) cmap_i = cmap[i]; 1338 rmap_i = rmap[i]; 1339 irow_i = irow[i]; 1340 jmax = nrow[i]; 1341 for (j=0; j<jmax; j++) { 1342 row = irow_i[j]; 1343 proc = row2proc_i[j]; 1344 1345 if (proc == rank) { 1346 1347 row = row - rstart; 1348 nzA = a_i[row+1] - a_i[row]; 1349 nzB = b_i[row+1] - b_i[row]; 1350 cworkA = a_j + a_i[row]; 1351 cworkB = b_j + b_i[row]; 1352 if (!ijonly) { 1353 vworkA = a_a + a_i[row]*bs2; 1354 vworkB = b_a + b_i[row]*bs2; 1355 } 1356 #if defined(PETSC_USE_CTABLE) 1357 ierr = PetscTableFind(rmap_i,row+rstart+1,&row);CHKERRQ(ierr); 1358 row--; 1359 if (row < 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"row not found in table"); 1360 #else 1361 row = rmap_i[row + rstart]; 1362 #endif 1363 mat_i = imat_i[row]; 1364 if (!ijonly) mat_a = imat_a + mat_i*bs2; 1365 mat_j = imat_j + mat_i; 1366 ilen_row = imat_ilen[row]; 1367 1368 #if 1 1369 /* load the column indices for this row into cols*/ 1370 if (!allcolumns[i]) { 1371 for (l=0; l<nzB; l++) { 1372 if ((ctmp = bmap[cworkB[l]]) < cstart) { 1373 #if defined(PETSC_USE_CTABLE) 1374 ierr = PetscTableFind(cmap_i,ctmp+1,&tcol);CHKERRQ(ierr); 1375 if (tcol) { 1376 #else 1377 if ((tcol = cmap_i[ctmp])) { 1378 #endif 1379 *mat_j++ = tcol - 1; 1380 ierr = PetscMemcpy(mat_a,vworkB+l*bs2,bs2*sizeof(MatScalar));CHKERRQ(ierr); 1381 mat_a += bs2; 1382 ilen_row++; 1383 } 1384 } else break; 1385 } 1386 imark = l; 1387 for (l=0; l<nzA; l++) { 1388 #if defined(PETSC_USE_CTABLE) 1389 ierr = PetscTableFind(cmap_i,cstart+cworkA[l]+1,&tcol);CHKERRQ(ierr); 1390 if (tcol) { 1391 #else 1392 if ((tcol = cmap_i[cstart + cworkA[l]])) { 1393 #endif 1394 *mat_j++ = tcol - 1; 1395 if (!ijonly) { 1396 ierr = PetscMemcpy(mat_a,vworkA+l*bs2,bs2*sizeof(MatScalar));CHKERRQ(ierr); 1397 mat_a += bs2; 1398 } 1399 ilen_row++; 1400 } 1401 } 1402 for (l=imark; l<nzB; l++) { 1403 #if defined(PETSC_USE_CTABLE) 1404 ierr = PetscTableFind(cmap_i,bmap[cworkB[l]]+1,&tcol);CHKERRQ(ierr); 1405 if (tcol) { 1406 #else 1407 if ((tcol = cmap_i[bmap[cworkB[l]]])) { 1408 #endif 1409 *mat_j++ = tcol - 1; 1410 if (!ijonly) { 1411 ierr = PetscMemcpy(mat_a,vworkB+l*bs2,bs2*sizeof(MatScalar));CHKERRQ(ierr); 1412 mat_a += bs2; 1413 } 1414 ilen_row++; 1415 } 1416 } 1417 } else { /* allcolumns */ 1418 for (l=0; l<nzB; l++) { 1419 if ((ctmp = bmap[cworkB[l]]) < cstart) { 1420 *mat_j++ = ctmp; 1421 ierr = PetscMemcpy(mat_a,vworkB+l*bs2,bs2*sizeof(MatScalar));CHKERRQ(ierr); 1422 mat_a += bs2; 1423 ilen_row++; 1424 } else break; 1425 } 1426 imark = l; 1427 for (l=0; l<nzA; l++) { 1428 *mat_j++ = cstart+cworkA[l]; 1429 if (!ijonly) { 1430 ierr = PetscMemcpy(mat_a,vworkA+l*bs2,bs2*sizeof(MatScalar));CHKERRQ(ierr); 1431 mat_a += bs2; 1432 } 1433 ilen_row++; 1434 } 1435 for (l=imark; l<nzB; l++) { 1436 *mat_j++ = bmap[cworkB[l]]; 1437 if (!ijonly) { 1438 ierr = PetscMemcpy(mat_a,vworkB+l*bs2,bs2*sizeof(MatScalar));CHKERRQ(ierr); 1439 mat_a += bs2; 1440 } 1441 ilen_row++; 1442 } 1443 } 1444 //imat_ilen[row] = ilen_row; 1445 //} 1446 #endif 1447 1448 #if 0 // ------- old 1449 old_row = row; 1450 #if defined(PETSC_USE_CTABLE) 1451 ierr = PetscTableFind(rmap_i,row+1,&row);CHKERRQ(ierr); 1452 row--; 1453 #else 1454 row = rmap_i[row]; 1455 #endif 1456 ilen_row = imat_ilen[row]; 1457 ierr = MatGetRow_MPIBAIJ(C,old_row,&ncols,&cols,&vals);CHKERRQ(ierr); 1458 mat_i = imat_i[row]; 1459 mat_a = imat_a + mat_i; 1460 mat_j = imat_j + mat_i; 1461 if (!allcolumns[i]) { 1462 for (k=0; k<ncols; k++) { 1463 #if defined(PETSC_USE_CTABLE) 1464 ierr = PetscTableFind(cmap_i,cols[k]+1,&tcol);CHKERRQ(ierr); 1465 #else 1466 tcol = cmap_i[cols[k]]; 1467 #endif 1468 if (tcol) { 1469 *mat_j++ = tcol - 1; 1470 *mat_a++ = vals[k]; 1471 ilen_row++; 1472 } 1473 } 1474 } else { /* allcolumns */ 1475 for (k=0; k<ncols; k++) { 1476 *mat_j++ = cols[k]; /* global col index! */ 1477 *mat_a++ = vals[k]; 1478 ilen_row++; 1479 } 1480 } 1481 ierr = MatRestoreRow_MPIBAIJ(C,old_row,&ncols,&cols,&vals);CHKERRQ(ierr); 1482 #endif //------ end of old 1483 imat_ilen[row] = ilen_row; 1484 } 1485 } 1486 } 1487 1488 /* Now assemble the off proc rows */ 1489 ierr = MPI_Waitall(nrqs,r_waits4,r_status4);CHKERRQ(ierr); 1490 for (tmp2=0; tmp2<nrqs; tmp2++) { 1491 sbuf1_i = sbuf1[pa[tmp2]]; 1492 jmax = sbuf1_i[0]; 1493 ct1 = 2*jmax + 1; 1494 ct2 = 0; 1495 rbuf2_i = rbuf2[tmp2]; 1496 rbuf3_i = rbuf3[tmp2]; 1497 rbuf4_i = rbuf4[tmp2]; 1498 for (j=1; j<=jmax; j++) { 1499 is_no = sbuf1_i[2*j-1]; 1500 rmap_i = rmap[is_no]; 1501 if (!allcolumns[is_no]) cmap_i = cmap[is_no]; 1502 subc = (Mat_SeqBAIJ*)submats[is_no]->data; 1503 imat_ilen = subc->ilen; 1504 imat_j = subc->j; 1505 imat_i = subc->i; 1506 imat_a = subc->a; 1507 max1 = sbuf1_i[2*j]; 1508 for (k=0; k<max1; k++,ct1++) { 1509 row = sbuf1_i[ct1]; 1510 #if defined(PETSC_USE_CTABLE) 1511 ierr = PetscTableFind(rmap_i,row+1,&row);CHKERRQ(ierr); 1512 row--; 1513 #else 1514 row = rmap_i[row]; 1515 #endif 1516 ilen = imat_ilen[row]; 1517 mat_i = imat_i[row]; 1518 mat_a = imat_a + mat_i; 1519 mat_j = imat_j + mat_i; 1520 max2 = rbuf2_i[ct1]; 1521 if (!allcolumns[is_no]) { 1522 for (l=0; l<max2; l++,ct2++) { 1523 #if defined(PETSC_USE_CTABLE) 1524 ierr = PetscTableFind(cmap_i,rbuf3_i[ct2]+1,&tcol);CHKERRQ(ierr); 1525 #else 1526 tcol = cmap_i[rbuf3_i[ct2]]; 1527 #endif 1528 if (tcol) { 1529 *mat_j++ = tcol - 1; 1530 *mat_a++ = rbuf4_i[ct2]; 1531 ilen++; 1532 } 1533 } 1534 } else { /* allcolumns */ 1535 for (l=0; l<max2; l++,ct2++) { 1536 *mat_j++ = rbuf3_i[ct2]; /* same global column index of C */ 1537 *mat_a++ = rbuf4_i[ct2]; 1538 ilen++; 1539 } 1540 } 1541 imat_ilen[row] = ilen; 1542 } 1543 } 1544 } 1545 1546 if (!iscsorted) { /* sort column indices of the rows */ 1547 for (i=0; i<ismax; i++) { 1548 subc = (Mat_SeqBAIJ*)submats[i]->data; 1549 imat_j = subc->j; 1550 imat_i = subc->i; 1551 imat_a = subc->a; 1552 imat_ilen = subc->ilen; 1553 1554 if (allcolumns[i]) continue; 1555 jmax = nrow[i]; 1556 for (j=0; j<jmax; j++) { 1557 PetscInt ilen; 1558 1559 mat_i = imat_i[j]; 1560 mat_a = imat_a + mat_i; 1561 mat_j = imat_j + mat_i; 1562 ilen = imat_ilen[j]; 1563 ierr = PetscSortIntWithScalarArray(ilen,mat_j,mat_a);CHKERRQ(ierr); 1564 } 1565 } 1566 } 1567 1568 ierr = PetscFree(r_status4);CHKERRQ(ierr); 1569 ierr = PetscFree(r_waits4);CHKERRQ(ierr); 1570 if (nrqr) {ierr = MPI_Waitall(nrqr,s_waits4,s_status4);CHKERRQ(ierr);} 1571 ierr = PetscFree(s_waits4);CHKERRQ(ierr); 1572 ierr = PetscFree(s_status4);CHKERRQ(ierr); 1573 1574 /* Restore the indices */ 1575 for (i=0; i<ismax; i++) { 1576 ierr = ISRestoreIndices(isrow[i],irow+i);CHKERRQ(ierr); 1577 if (!allcolumns[i]) { 1578 ierr = ISRestoreIndices(iscol[i],icol+i);CHKERRQ(ierr); 1579 } 1580 } 1581 1582 for (i=0; i<ismax; i++) { 1583 ierr = MatAssemblyBegin(submats[i],MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1584 ierr = MatAssemblyEnd(submats[i],MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1585 } 1586 1587 /* Destroy allocated memory */ 1588 if (!ismax) { 1589 ierr = PetscFree(pa);CHKERRQ(ierr); 1590 1591 ierr = PetscFree4(sbuf1,ptr,tmp,ctr);CHKERRQ(ierr); 1592 for (i=0; i<nrqr; ++i) { 1593 ierr = PetscFree(sbuf2[i]);CHKERRQ(ierr); 1594 } 1595 for (i=0; i<nrqs; ++i) { 1596 ierr = PetscFree(rbuf3[i]);CHKERRQ(ierr); 1597 } 1598 1599 ierr = PetscFree3(sbuf2,req_size,req_source1);CHKERRQ(ierr); 1600 ierr = PetscFree3(req_source2,rbuf2,rbuf3);CHKERRQ(ierr); 1601 } 1602 1603 ierr = PetscFree(sbuf_aa[0]);CHKERRQ(ierr); 1604 ierr = PetscFree(sbuf_aa);CHKERRQ(ierr); 1605 ierr = PetscFree5(irow,icol,nrow,ncol,issorted);CHKERRQ(ierr); 1606 1607 for (i=0; i<nrqs; ++i) { 1608 ierr = PetscFree(rbuf4[i]);CHKERRQ(ierr); 1609 } 1610 ierr = PetscFree(rbuf4);CHKERRQ(ierr); 1611 1612 ierr = PetscFree5(smats,row2proc,cmap,rmap,allcolumns);CHKERRQ(ierr); 1613 PetscFunctionReturn(0); 1614 } 1615 1616 //------------ endof new ------------- 1617 1618 PetscErrorCode MatGetSubMatrices_MPIBAIJ_local(Mat C,PetscInt ismax,const IS isrow[],const IS iscol[],MatReuse scall,PetscBool *allrows,PetscBool *allcolumns,Mat *submats) 1619 { 1620 Mat_MPIBAIJ *c = (Mat_MPIBAIJ*)C->data; 1621 Mat A = c->A; 1622 Mat_SeqBAIJ *a = (Mat_SeqBAIJ*)A->data,*b = (Mat_SeqBAIJ*)c->B->data,*mat; 1623 const PetscInt **irow,**icol,*irow_i; 1624 PetscInt *nrow,*ncol,*w3,*w4,start; 1625 PetscErrorCode ierr; 1626 PetscMPIInt size,tag0,tag1,tag2,tag3,*w1,*w2,nrqr,idex,end,proc; 1627 PetscInt **sbuf1,**sbuf2,rank,i,j,k,l,ct1,ct2,**rbuf1,row; 1628 PetscInt nrqs,msz,**ptr,*req_size,*ctr,*pa,*tmp,tcol; 1629 PetscInt **rbuf3,*req_source,**sbuf_aj,**rbuf2,max1,max2; 1630 PetscInt **lens,is_no,ncols,*cols,mat_i,*mat_j,tmp2,jmax; 1631 PetscInt ctr_j,*sbuf1_j,*sbuf_aj_i,*rbuf1_i,kmax,*lens_i; 1632 PetscInt bs =C->rmap->bs,bs2=c->bs2,*a_j=a->j,*b_j=b->j,*cworkA,*cworkB; 1633 PetscInt cstart = c->cstartbs,nzA,nzB,*a_i=a->i,*b_i=b->i,imark; 1634 PetscInt *bmap = c->garray,ctmp,rstart=c->rstartbs; 1635 MPI_Request *s_waits1,*r_waits1,*s_waits2,*r_waits2,*r_waits3,*s_waits3; 1636 MPI_Status *r_status1,*r_status2,*s_status1,*s_status3,*s_status2,*r_status3; 1637 MPI_Comm comm; 1638 PetscBool flag; 1639 PetscMPIInt *onodes1,*olengths1; 1640 PetscBool ijonly=c->ijonly; /* private flag indicates only matrix data structures are requested */ 1641 1642 /* variables below are used for the matrix numerical values - case of !ijonly */ 1643 MPI_Request *r_waits4,*s_waits4; 1644 MPI_Status *r_status4,*s_status4; 1645 MatScalar **rbuf4,**sbuf_aa,*vals,*mat_a = NULL,*sbuf_aa_i,*vworkA = NULL,*vworkB = NULL; 1646 MatScalar *a_a=a->a,*b_a=b->a; 1647 1648 #if defined(PETSC_USE_CTABLE) 1649 PetscInt tt; 1650 PetscTable *rmap,*cmap,rmap_i,cmap_i=NULL; 1651 #else 1652 PetscInt **cmap,*cmap_i=NULL,*rtable,*rmap_i,**rmap, Mbs = c->Mbs; 1653 #endif 1654 1655 PetscFunctionBegin; 1656 ierr = PetscObjectGetComm((PetscObject)C,&comm);CHKERRQ(ierr); 1657 tag0 = ((PetscObject)C)->tag; 1658 size = c->size; 1659 rank = c->rank; 1660 if (!rank) printf("MatGetSubMatrices_MPIBAIJ scall %d, bs %d\n",scall,bs); 1661 1662 /* Get some new tags to keep the communication clean */ 1663 ierr = PetscObjectGetNewTag((PetscObject)C,&tag1);CHKERRQ(ierr); 1664 ierr = PetscObjectGetNewTag((PetscObject)C,&tag2);CHKERRQ(ierr); 1665 ierr = PetscObjectGetNewTag((PetscObject)C,&tag3);CHKERRQ(ierr); 1666 1667 #if defined(PETSC_USE_CTABLE) 1668 ierr = PetscMalloc4(ismax,&irow,ismax,&icol,ismax,&nrow,ismax,&ncol);CHKERRQ(ierr); 1669 #else 1670 ierr = PetscMalloc5(ismax,&irow,ismax,&icol,ismax,&nrow,ismax,&ncol,Mbs+1,&rtable);CHKERRQ(ierr); 1671 /* Create hash table for the mapping :row -> proc*/ 1672 for (i=0,j=0; i<size; i++) { 1673 jmax = C->rmap->range[i+1]/bs; 1674 for (; j<jmax; j++) rtable[j] = i; 1675 } 1676 #endif 1677 1678 for (i=0; i<ismax; i++) { 1679 if (allrows[i]) { 1680 irow[i] = NULL; 1681 nrow[i] = C->rmap->N/bs; 1682 } else { 1683 ierr = ISGetIndices(isrow[i],&irow[i]);CHKERRQ(ierr); 1684 ierr = ISGetLocalSize(isrow[i],&nrow[i]);CHKERRQ(ierr); 1685 } 1686 1687 if (allcolumns[i]) { 1688 icol[i] = NULL; 1689 ncol[i] = C->cmap->N/bs; 1690 } else { 1691 ierr = ISGetIndices(iscol[i],&icol[i]);CHKERRQ(ierr); 1692 ierr = ISGetLocalSize(iscol[i],&ncol[i]);CHKERRQ(ierr); 1693 } 1694 } 1695 1696 /* evaluate communication - mesg to who,length of mesg,and buffer space 1697 required. Based on this, buffers are allocated, and data copied into them*/ 1698 ierr = PetscCalloc4(size,&w1,size,&w2,size,&w3,size,&w4);CHKERRQ(ierr); 1699 for (i=0; i<ismax; i++) { 1700 ierr = PetscMemzero(w4,size*sizeof(PetscInt));CHKERRQ(ierr); /* initialise work vector*/ 1701 jmax = nrow[i]; 1702 irow_i = irow[i]; 1703 for (j=0; j<jmax; j++) { 1704 if (allrows[i]) row = j; 1705 else row = irow_i[j]; 1706 1707 #if defined(PETSC_USE_CTABLE) 1708 ierr = PetscGetProc(row,size,c->rangebs,&proc);CHKERRQ(ierr); 1709 #else 1710 proc = rtable[row]; 1711 #endif 1712 w4[proc]++; 1713 } 1714 for (j=0; j<size; j++) { 1715 if (w4[j]) { w1[j] += w4[j]; w3[j]++;} 1716 } 1717 } 1718 1719 nrqs = 0; /* no of outgoing messages */ 1720 msz = 0; /* total mesg length for all proc */ 1721 w1[rank] = 0; /* no mesg sent to intself */ 1722 w3[rank] = 0; 1723 for (i=0; i<size; i++) { 1724 if (w1[i]) { w2[i] = 1; nrqs++;} /* there exists a message to proc i */ 1725 } 1726 ierr = PetscMalloc1(nrqs+1,&pa);CHKERRQ(ierr); /*(proc -array)*/ 1727 for (i=0,j=0; i<size; i++) { 1728 if (w1[i]) { pa[j] = i; j++; } 1729 } 1730 1731 /* Each message would have a header = 1 + 2*(no of IS) + data */ 1732 for (i=0; i<nrqs; i++) { 1733 j = pa[i]; 1734 w1[j] += w2[j] + 2* w3[j]; 1735 msz += w1[j]; 1736 } 1737 1738 /* Determine the number of messages to expect, their lengths, from from-ids */ 1739 ierr = PetscGatherNumberOfMessages(comm,w2,w1,&nrqr);CHKERRQ(ierr); 1740 ierr = PetscGatherMessageLengths(comm,nrqs,nrqr,w1,&onodes1,&olengths1);CHKERRQ(ierr); 1741 1742 /* Now post the Irecvs corresponding to these messages */ 1743 ierr = PetscPostIrecvInt(comm,tag0,nrqr,onodes1,olengths1,&rbuf1,&r_waits1);CHKERRQ(ierr); 1744 1745 ierr = PetscFree(onodes1);CHKERRQ(ierr); 1746 ierr = PetscFree(olengths1);CHKERRQ(ierr); 1747 1748 /* Allocate Memory for outgoing messages */ 1749 ierr = PetscMalloc4(size,&sbuf1,size,&ptr,2*msz,&tmp,size,&ctr);CHKERRQ(ierr); 1750 ierr = PetscMemzero(sbuf1,size*sizeof(PetscInt*));CHKERRQ(ierr); 1751 ierr = PetscMemzero(ptr,size*sizeof(PetscInt*));CHKERRQ(ierr); 1752 { 1753 PetscInt *iptr = tmp,ict = 0; 1754 for (i=0; i<nrqs; i++) { 1755 j = pa[i]; 1756 iptr += ict; 1757 sbuf1[j] = iptr; 1758 ict = w1[j]; 1759 } 1760 } 1761 1762 /* Form the outgoing messages */ 1763 /* Initialise the header space */ 1764 for (i=0; i<nrqs; i++) { 1765 j = pa[i]; 1766 sbuf1[j][0] = 0; 1767 ierr = PetscMemzero(sbuf1[j]+1,2*w3[j]*sizeof(PetscInt));CHKERRQ(ierr); 1768 ptr[j] = sbuf1[j] + 2*w3[j] + 1; 1769 } 1770 1771 /* Parse the isrow and copy data into outbuf */ 1772 for (i=0; i<ismax; i++) { 1773 ierr = PetscMemzero(ctr,size*sizeof(PetscInt));CHKERRQ(ierr); 1774 irow_i = irow[i]; 1775 jmax = nrow[i]; 1776 for (j=0; j<jmax; j++) { /* parse the indices of each IS */ 1777 if (allrows[i]) row = j; 1778 else row = irow_i[j]; 1779 1780 #if defined(PETSC_USE_CTABLE) 1781 ierr = PetscGetProc(row,size,c->rangebs,&proc);CHKERRQ(ierr); 1782 #else 1783 proc = rtable[row]; 1784 #endif 1785 if (proc != rank) { /* copy to the outgoing buf*/ 1786 ctr[proc]++; 1787 *ptr[proc] = row; 1788 ptr[proc]++; 1789 } 1790 } 1791 /* Update the headers for the current IS */ 1792 for (j=0; j<size; j++) { /* Can Optimise this loop too */ 1793 if ((ctr_j = ctr[j])) { 1794 sbuf1_j = sbuf1[j]; 1795 k = ++sbuf1_j[0]; 1796 sbuf1_j[2*k] = ctr_j; 1797 sbuf1_j[2*k-1] = i; 1798 } 1799 } 1800 } 1801 1802 /* Now post the sends */ 1803 ierr = PetscMalloc1(nrqs+1,&s_waits1);CHKERRQ(ierr); 1804 for (i=0; i<nrqs; ++i) { 1805 j = pa[i]; 1806 ierr = MPI_Isend(sbuf1[j],w1[j],MPIU_INT,j,tag0,comm,s_waits1+i);CHKERRQ(ierr); 1807 } 1808 1809 /* Post Recieves to capture the buffer size */ 1810 ierr = PetscMalloc1(nrqs+1,&r_waits2);CHKERRQ(ierr); 1811 ierr = PetscMalloc1(nrqs+1,&rbuf2);CHKERRQ(ierr); 1812 rbuf2[0] = tmp + msz; 1813 for (i=1; i<nrqs; ++i) { 1814 rbuf2[i] = rbuf2[i-1]+w1[pa[i-1]]; 1815 } 1816 for (i=0; i<nrqs; ++i) { 1817 j = pa[i]; 1818 ierr = MPI_Irecv(rbuf2[i],w1[j],MPIU_INT,j,tag1,comm,r_waits2+i);CHKERRQ(ierr); 1819 } 1820 1821 /* Send to other procs the buf size they should allocate */ 1822 1823 /* Receive messages*/ 1824 ierr = PetscMalloc1(nrqr+1,&s_waits2);CHKERRQ(ierr); 1825 ierr = PetscMalloc1(nrqr+1,&r_status1);CHKERRQ(ierr); 1826 ierr = PetscMalloc3(nrqr+1,&sbuf2,nrqr,&req_size,nrqr,&req_source);CHKERRQ(ierr); 1827 { 1828 Mat_SeqBAIJ *sA = (Mat_SeqBAIJ*)c->A->data,*sB = (Mat_SeqBAIJ*)c->B->data; 1829 PetscInt *sAi = sA->i,*sBi = sB->i,id,*sbuf2_i; 1830 1831 for (i=0; i<nrqr; ++i) { 1832 ierr = MPI_Waitany(nrqr,r_waits1,&idex,r_status1+i);CHKERRQ(ierr); 1833 1834 req_size[idex] = 0; 1835 rbuf1_i = rbuf1[idex]; 1836 start = 2*rbuf1_i[0] + 1; 1837 ierr = MPI_Get_count(r_status1+i,MPIU_INT,&end);CHKERRQ(ierr); 1838 ierr = PetscMalloc1(end,&sbuf2[idex]);CHKERRQ(ierr); 1839 sbuf2_i = sbuf2[idex]; 1840 for (j=start; j<end; j++) { 1841 id = rbuf1_i[j] - rstart; 1842 ncols = sAi[id+1] - sAi[id] + sBi[id+1] - sBi[id]; 1843 sbuf2_i[j] = ncols; 1844 req_size[idex] += ncols; 1845 } 1846 req_source[idex] = r_status1[i].MPI_SOURCE; 1847 /* form the header */ 1848 sbuf2_i[0] = req_size[idex]; 1849 for (j=1; j<start; j++) sbuf2_i[j] = rbuf1_i[j]; 1850 ierr = MPI_Isend(sbuf2_i,end,MPIU_INT,req_source[idex],tag1,comm,s_waits2+i);CHKERRQ(ierr); 1851 } 1852 } 1853 ierr = PetscFree(r_status1);CHKERRQ(ierr); 1854 ierr = PetscFree(r_waits1);CHKERRQ(ierr); 1855 1856 /* recv buffer sizes */ 1857 /* Receive messages*/ 1858 ierr = PetscMalloc1(nrqs+1,&rbuf3);CHKERRQ(ierr); 1859 ierr = PetscMalloc1(nrqs+1,&r_waits3);CHKERRQ(ierr); 1860 ierr = PetscMalloc1(nrqs+1,&r_status2);CHKERRQ(ierr); 1861 if (!ijonly) { 1862 ierr = PetscMalloc1(nrqs+1,&rbuf4);CHKERRQ(ierr); 1863 ierr = PetscMalloc1(nrqs+1,&r_waits4);CHKERRQ(ierr); 1864 } 1865 1866 for (i=0; i<nrqs; ++i) { 1867 ierr = MPI_Waitany(nrqs,r_waits2,&idex,r_status2+i);CHKERRQ(ierr); 1868 ierr = PetscMalloc1(rbuf2[idex][0],&rbuf3[idex]);CHKERRQ(ierr); 1869 ierr = MPI_Irecv(rbuf3[idex],rbuf2[idex][0],MPIU_INT,r_status2[i].MPI_SOURCE,tag2,comm,r_waits3+idex);CHKERRQ(ierr); 1870 if (!ijonly) { 1871 ierr = PetscMalloc1(rbuf2[idex][0]*bs2,&rbuf4[idex]);CHKERRQ(ierr); 1872 ierr = MPI_Irecv(rbuf4[idex],rbuf2[idex][0]*bs2,MPIU_MATSCALAR,r_status2[i].MPI_SOURCE,tag3,comm,r_waits4+idex);CHKERRQ(ierr); 1873 } 1874 } 1875 ierr = PetscFree(r_status2);CHKERRQ(ierr); 1876 ierr = PetscFree(r_waits2);CHKERRQ(ierr); 1877 1878 /* Wait on sends1 and sends2 */ 1879 ierr = PetscMalloc1(nrqs+1,&s_status1);CHKERRQ(ierr); 1880 ierr = PetscMalloc1(nrqr+1,&s_status2);CHKERRQ(ierr); 1881 1882 if (nrqs) {ierr = MPI_Waitall(nrqs,s_waits1,s_status1);CHKERRQ(ierr);} 1883 if (nrqr) {ierr = MPI_Waitall(nrqr,s_waits2,s_status2);CHKERRQ(ierr);} 1884 ierr = PetscFree(s_status1);CHKERRQ(ierr); 1885 ierr = PetscFree(s_status2);CHKERRQ(ierr); 1886 ierr = PetscFree(s_waits1);CHKERRQ(ierr); 1887 ierr = PetscFree(s_waits2);CHKERRQ(ierr); 1888 1889 /* Now allocate buffers for a->j, and send them off */ 1890 ierr = PetscMalloc1(nrqr+1,&sbuf_aj);CHKERRQ(ierr); 1891 for (i=0,j=0; i<nrqr; i++) j += req_size[i]; 1892 ierr = PetscMalloc1(j+1,&sbuf_aj[0]);CHKERRQ(ierr); 1893 for (i=1; i<nrqr; i++) sbuf_aj[i] = sbuf_aj[i-1] + req_size[i-1]; 1894 1895 ierr = PetscMalloc1(nrqr+1,&s_waits3);CHKERRQ(ierr); 1896 { 1897 for (i=0; i<nrqr; i++) { 1898 rbuf1_i = rbuf1[i]; 1899 sbuf_aj_i = sbuf_aj[i]; 1900 ct1 = 2*rbuf1_i[0] + 1; 1901 ct2 = 0; 1902 for (j=1,max1=rbuf1_i[0]; j<=max1; j++) { 1903 kmax = rbuf1[i][2*j]; 1904 for (k=0; k<kmax; k++,ct1++) { 1905 row = rbuf1_i[ct1] - rstart; 1906 nzA = a_i[row+1] - a_i[row]; 1907 nzB = b_i[row+1] - b_i[row]; 1908 ncols = nzA + nzB; 1909 cworkA = a_j + a_i[row]; 1910 cworkB = b_j + b_i[row]; 1911 1912 /* load the column indices for this row into cols*/ 1913 cols = sbuf_aj_i + ct2; 1914 for (l=0; l<nzB; l++) { 1915 if ((ctmp = bmap[cworkB[l]]) < cstart) cols[l] = ctmp; 1916 else break; 1917 } 1918 imark = l; 1919 for (l=0; l<nzA; l++) cols[imark+l] = cstart + cworkA[l]; 1920 for (l=imark; l<nzB; l++) cols[nzA+l] = bmap[cworkB[l]]; 1921 ct2 += ncols; 1922 } 1923 } 1924 ierr = MPI_Isend(sbuf_aj_i,req_size[i],MPIU_INT,req_source[i],tag2,comm,s_waits3+i);CHKERRQ(ierr); 1925 } 1926 } 1927 ierr = PetscMalloc1(nrqs+1,&r_status3);CHKERRQ(ierr); 1928 ierr = PetscMalloc1(nrqr+1,&s_status3);CHKERRQ(ierr); 1929 1930 /* Allocate buffers for a->a, and send them off */ 1931 if (!ijonly) { 1932 ierr = PetscMalloc1(nrqr+1,&sbuf_aa);CHKERRQ(ierr); 1933 for (i=0,j=0; i<nrqr; i++) j += req_size[i]; 1934 ierr = PetscMalloc1((j+1)*bs2,&sbuf_aa[0]);CHKERRQ(ierr); 1935 for (i=1; i<nrqr; i++) sbuf_aa[i] = sbuf_aa[i-1] + req_size[i-1]*bs2; 1936 1937 ierr = PetscMalloc1(nrqr+1,&s_waits4);CHKERRQ(ierr); 1938 { 1939 for (i=0; i<nrqr; i++) { 1940 rbuf1_i = rbuf1[i]; 1941 sbuf_aa_i = sbuf_aa[i]; 1942 ct1 = 2*rbuf1_i[0]+1; 1943 ct2 = 0; 1944 for (j=1,max1=rbuf1_i[0]; j<=max1; j++) { 1945 kmax = rbuf1_i[2*j]; 1946 for (k=0; k<kmax; k++,ct1++) { 1947 row = rbuf1_i[ct1] - rstart; 1948 nzA = a_i[row+1] - a_i[row]; 1949 nzB = b_i[row+1] - b_i[row]; 1950 ncols = nzA + nzB; 1951 cworkB = b_j + b_i[row]; 1952 vworkA = a_a + a_i[row]*bs2; 1953 vworkB = b_a + b_i[row]*bs2; 1954 1955 /* load the column values for this row into vals*/ 1956 vals = sbuf_aa_i+ct2*bs2; 1957 for (l=0; l<nzB; l++) { 1958 if ((bmap[cworkB[l]]) < cstart) { 1959 ierr = PetscMemcpy(vals+l*bs2,vworkB+l*bs2,bs2*sizeof(MatScalar));CHKERRQ(ierr); 1960 } else break; 1961 } 1962 imark = l; 1963 for (l=0; l<nzA; l++) { 1964 ierr = PetscMemcpy(vals+(imark+l)*bs2,vworkA+l*bs2,bs2*sizeof(MatScalar));CHKERRQ(ierr); 1965 } 1966 for (l=imark; l<nzB; l++) { 1967 ierr = PetscMemcpy(vals+(nzA+l)*bs2,vworkB+l*bs2,bs2*sizeof(MatScalar));CHKERRQ(ierr); 1968 } 1969 ct2 += ncols; 1970 } 1971 } 1972 ierr = MPI_Isend(sbuf_aa_i,req_size[i]*bs2,MPIU_MATSCALAR,req_source[i],tag3,comm,s_waits4+i);CHKERRQ(ierr); 1973 } 1974 } 1975 ierr = PetscMalloc1(nrqs+1,&r_status4);CHKERRQ(ierr); 1976 ierr = PetscMalloc1(nrqr+1,&s_status4);CHKERRQ(ierr); 1977 } 1978 ierr = PetscFree(rbuf1[0]);CHKERRQ(ierr); 1979 ierr = PetscFree(rbuf1);CHKERRQ(ierr); 1980 1981 /* Form the matrix */ 1982 /* create col map: global col of C -> local col of submatrices */ 1983 { 1984 const PetscInt *icol_i; 1985 #if defined(PETSC_USE_CTABLE) 1986 ierr = PetscMalloc1(1+ismax,&cmap);CHKERRQ(ierr); 1987 for (i=0; i<ismax; i++) { 1988 if (!allcolumns[i]) { 1989 ierr = PetscTableCreate(ncol[i]+1,c->Nbs+1,&cmap[i]);CHKERRQ(ierr); 1990 jmax = ncol[i]; 1991 icol_i = icol[i]; 1992 cmap_i = cmap[i]; 1993 for (j=0; j<jmax; j++) { 1994 ierr = PetscTableAdd(cmap_i,icol_i[j]+1,j+1,INSERT_VALUES);CHKERRQ(ierr); 1995 } 1996 } else { 1997 cmap[i] = NULL; 1998 } 1999 } 2000 #else 2001 ierr = PetscMalloc1(ismax,&cmap);CHKERRQ(ierr); 2002 for (i=0; i<ismax; i++) { 2003 if (!allcolumns[i]) { 2004 ierr = PetscCalloc1(c->Nbs,&cmap[i]);CHKERRQ(ierr); 2005 jmax = ncol[i]; 2006 icol_i = icol[i]; 2007 cmap_i = cmap[i]; 2008 for (j=0; j<jmax; j++) { 2009 cmap_i[icol_i[j]] = j+1; 2010 } 2011 } else { /* allcolumns[i] */ 2012 cmap[i] = NULL; 2013 } 2014 } 2015 #endif 2016 } 2017 2018 /* Create lens which is required for MatCreate... */ 2019 for (i=0,j=0; i<ismax; i++) j += nrow[i]; 2020 ierr = PetscMalloc((1+ismax)*sizeof(PetscInt*)+ j*sizeof(PetscInt),&lens);CHKERRQ(ierr); 2021 lens[0] = (PetscInt*)(lens + ismax); 2022 ierr = PetscMemzero(lens[0],j*sizeof(PetscInt));CHKERRQ(ierr); 2023 for (i=1; i<ismax; i++) lens[i] = lens[i-1] + nrow[i-1]; 2024 2025 /* Update lens from local data */ 2026 for (i=0; i<ismax; i++) { 2027 jmax = nrow[i]; 2028 if (!allcolumns[i]) cmap_i = cmap[i]; 2029 irow_i = irow[i]; 2030 lens_i = lens[i]; 2031 for (j=0; j<jmax; j++) { 2032 if (allrows[i]) row = j; 2033 else row = irow_i[j]; 2034 2035 #if defined(PETSC_USE_CTABLE) 2036 ierr = PetscGetProc(row,size,c->rangebs,&proc);CHKERRQ(ierr); 2037 #else 2038 proc = rtable[row]; 2039 #endif 2040 if (proc == rank) { 2041 /* Get indices from matA and then from matB */ 2042 row = row - rstart; 2043 nzA = a_i[row+1] - a_i[row]; 2044 nzB = b_i[row+1] - b_i[row]; 2045 cworkA = a_j + a_i[row]; 2046 cworkB = b_j + b_i[row]; 2047 if (!allcolumns[i]) { 2048 #if defined(PETSC_USE_CTABLE) 2049 for (k=0; k<nzA; k++) { 2050 ierr = PetscTableFind(cmap_i,cstart+cworkA[k]+1,&tt);CHKERRQ(ierr); 2051 if (tt) lens_i[j]++; 2052 } 2053 for (k=0; k<nzB; k++) { 2054 ierr = PetscTableFind(cmap_i,bmap[cworkB[k]]+1,&tt);CHKERRQ(ierr); 2055 if (tt) lens_i[j]++; 2056 } 2057 2058 #else 2059 for (k=0; k<nzA; k++) { 2060 if (cmap_i[cstart + cworkA[k]]) lens_i[j]++; 2061 } 2062 for (k=0; k<nzB; k++) { 2063 if (cmap_i[bmap[cworkB[k]]]) lens_i[j]++; 2064 } 2065 #endif 2066 } else { /* allcolumns */ 2067 lens_i[j] = nzA + nzB; 2068 } 2069 } 2070 } 2071 } 2072 #if defined(PETSC_USE_CTABLE) 2073 /* Create row map*/ 2074 ierr = PetscMalloc1(1+ismax,&rmap);CHKERRQ(ierr); 2075 for (i=0; i<ismax; i++) { 2076 ierr = PetscTableCreate(nrow[i]+1,c->Mbs+1,&rmap[i]);CHKERRQ(ierr); 2077 } 2078 #else 2079 /* Create row map*/ 2080 ierr = PetscMalloc((1+ismax)*sizeof(PetscInt*)+ ismax*Mbs*sizeof(PetscInt),&rmap);CHKERRQ(ierr); 2081 rmap[0] = (PetscInt*)(rmap + ismax); 2082 ierr = PetscMemzero(rmap[0],ismax*Mbs*sizeof(PetscInt));CHKERRQ(ierr); 2083 for (i=1; i<ismax; i++) rmap[i] = rmap[i-1] + Mbs; 2084 #endif 2085 for (i=0; i<ismax; i++) { 2086 irow_i = irow[i]; 2087 jmax = nrow[i]; 2088 #if defined(PETSC_USE_CTABLE) 2089 rmap_i = rmap[i]; 2090 for (j=0; j<jmax; j++) { 2091 if (allrows[i]) { 2092 ierr = PetscTableAdd(rmap_i,j+1,j+1,INSERT_VALUES);CHKERRQ(ierr); 2093 } else { 2094 ierr = PetscTableAdd(rmap_i,irow_i[j]+1,j+1,INSERT_VALUES);CHKERRQ(ierr); 2095 } 2096 } 2097 #else 2098 rmap_i = rmap[i]; 2099 for (j=0; j<jmax; j++) { 2100 if (allrows[i]) rmap_i[j] = j; 2101 else rmap_i[irow_i[j]] = j; 2102 } 2103 #endif 2104 } 2105 2106 /* Update lens from offproc data */ 2107 { 2108 PetscInt *rbuf2_i,*rbuf3_i,*sbuf1_i; 2109 PetscMPIInt ii; 2110 2111 for (tmp2=0; tmp2<nrqs; tmp2++) { 2112 ierr = MPI_Waitany(nrqs,r_waits3,&ii,r_status3+tmp2);CHKERRQ(ierr); 2113 idex = pa[ii]; 2114 sbuf1_i = sbuf1[idex]; 2115 jmax = sbuf1_i[0]; 2116 ct1 = 2*jmax+1; 2117 ct2 = 0; 2118 rbuf2_i = rbuf2[ii]; 2119 rbuf3_i = rbuf3[ii]; 2120 for (j=1; j<=jmax; j++) { 2121 is_no = sbuf1_i[2*j-1]; 2122 max1 = sbuf1_i[2*j]; 2123 lens_i = lens[is_no]; 2124 if (!allcolumns[is_no]) cmap_i = cmap[is_no]; 2125 rmap_i = rmap[is_no]; 2126 for (k=0; k<max1; k++,ct1++) { 2127 #if defined(PETSC_USE_CTABLE) 2128 ierr = PetscTableFind(rmap_i,sbuf1_i[ct1]+1,&row);CHKERRQ(ierr); 2129 row--; 2130 if (row < 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"row not found in table"); 2131 #else 2132 row = rmap_i[sbuf1_i[ct1]]; /* the val in the new matrix to be */ 2133 #endif 2134 max2 = rbuf2_i[ct1]; 2135 for (l=0; l<max2; l++,ct2++) { 2136 if (!allcolumns[is_no]) { 2137 #if defined(PETSC_USE_CTABLE) 2138 ierr = PetscTableFind(cmap_i,rbuf3_i[ct2]+1,&tt);CHKERRQ(ierr); 2139 if (tt) lens_i[row]++; 2140 #else 2141 if (cmap_i[rbuf3_i[ct2]]) lens_i[row]++; 2142 #endif 2143 } else { /* allcolumns */ 2144 lens_i[row]++; 2145 } 2146 } 2147 } 2148 } 2149 } 2150 } 2151 ierr = PetscFree(r_status3);CHKERRQ(ierr); 2152 ierr = PetscFree(r_waits3);CHKERRQ(ierr); 2153 if (nrqr) {ierr = MPI_Waitall(nrqr,s_waits3,s_status3);CHKERRQ(ierr);} 2154 ierr = PetscFree(s_status3);CHKERRQ(ierr); 2155 ierr = PetscFree(s_waits3);CHKERRQ(ierr); 2156 2157 /* Create the submatrices */ 2158 if (scall == MAT_REUSE_MATRIX) { 2159 if (ijonly) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP," MAT_REUSE_MATRIX and ijonly is not supported yet"); 2160 /* 2161 Assumes new rows are same length as the old rows, hence bug! 2162 */ 2163 for (i=0; i<ismax; i++) { 2164 mat = (Mat_SeqBAIJ*)(submats[i]->data); 2165 if ((mat->mbs != nrow[i]) || (mat->nbs != ncol[i] || C->rmap->bs != bs)) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Cannot reuse matrix. wrong size"); 2166 ierr = PetscMemcmp(mat->ilen,lens[i],mat->mbs *sizeof(PetscInt),&flag);CHKERRQ(ierr); 2167 if (!flag) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_INCOMP,"Cannot reuse matrix. wrong no of nonzeros"); 2168 /* Initial matrix as if empty */ 2169 ierr = PetscMemzero(mat->ilen,mat->mbs*sizeof(PetscInt));CHKERRQ(ierr); 2170 2171 submats[i]->factortype = C->factortype; 2172 } 2173 } else { 2174 PetscInt bs_tmp; 2175 if (ijonly) bs_tmp = 1; 2176 else bs_tmp = bs; 2177 for (i=0; i<ismax; i++) { 2178 ierr = MatCreate(PETSC_COMM_SELF,submats+i);CHKERRQ(ierr); 2179 ierr = MatSetSizes(submats[i],nrow[i]*bs_tmp,ncol[i]*bs_tmp,nrow[i]*bs_tmp,ncol[i]*bs_tmp);CHKERRQ(ierr); 2180 ierr = MatSetType(submats[i],((PetscObject)A)->type_name);CHKERRQ(ierr); 2181 ierr = MatSeqBAIJSetPreallocation(submats[i],bs_tmp,0,lens[i]);CHKERRQ(ierr); 2182 ierr = MatSeqSBAIJSetPreallocation(submats[i],bs_tmp,0,lens[i]);CHKERRQ(ierr); /* this subroutine is used by SBAIJ routines */ 2183 } 2184 } 2185 2186 /* Assemble the matrices */ 2187 /* First assemble the local rows */ 2188 { 2189 PetscInt ilen_row,*imat_ilen,*imat_j,*imat_i; 2190 MatScalar *imat_a = NULL; 2191 2192 for (i=0; i<ismax; i++) { 2193 mat = (Mat_SeqBAIJ*)submats[i]->data; 2194 imat_ilen = mat->ilen; 2195 imat_j = mat->j; 2196 imat_i = mat->i; 2197 if (!ijonly) imat_a = mat->a; 2198 if (!allcolumns[i]) cmap_i = cmap[i]; 2199 rmap_i = rmap[i]; 2200 irow_i = irow[i]; 2201 jmax = nrow[i]; 2202 for (j=0; j<jmax; j++) { 2203 if (allrows[i]) row = j; 2204 else row = irow_i[j]; 2205 2206 #if defined(PETSC_USE_CTABLE) 2207 ierr = PetscGetProc(row,size,c->rangebs,&proc);CHKERRQ(ierr); 2208 #else 2209 proc = rtable[row]; 2210 #endif 2211 if (proc == rank) { 2212 row = row - rstart; 2213 nzA = a_i[row+1] - a_i[row]; 2214 nzB = b_i[row+1] - b_i[row]; 2215 cworkA = a_j + a_i[row]; 2216 cworkB = b_j + b_i[row]; 2217 if (!ijonly) { 2218 vworkA = a_a + a_i[row]*bs2; 2219 vworkB = b_a + b_i[row]*bs2; 2220 } 2221 #if defined(PETSC_USE_CTABLE) 2222 ierr = PetscTableFind(rmap_i,row+rstart+1,&row);CHKERRQ(ierr); 2223 row--; 2224 if (row < 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"row not found in table"); 2225 #else 2226 row = rmap_i[row + rstart]; 2227 #endif 2228 mat_i = imat_i[row]; 2229 if (!ijonly) mat_a = imat_a + mat_i*bs2; 2230 mat_j = imat_j + mat_i; 2231 ilen_row = imat_ilen[row]; 2232 2233 /* load the column indices for this row into cols*/ 2234 if (!allcolumns[i]) { 2235 for (l=0; l<nzB; l++) { 2236 if ((ctmp = bmap[cworkB[l]]) < cstart) { 2237 #if defined(PETSC_USE_CTABLE) 2238 ierr = PetscTableFind(cmap_i,ctmp+1,&tcol);CHKERRQ(ierr); 2239 if (tcol) { 2240 #else 2241 if ((tcol = cmap_i[ctmp])) { 2242 #endif 2243 *mat_j++ = tcol - 1; 2244 ierr = PetscMemcpy(mat_a,vworkB+l*bs2,bs2*sizeof(MatScalar));CHKERRQ(ierr); 2245 mat_a += bs2; 2246 ilen_row++; 2247 } 2248 } else break; 2249 } 2250 imark = l; 2251 for (l=0; l<nzA; l++) { 2252 #if defined(PETSC_USE_CTABLE) 2253 ierr = PetscTableFind(cmap_i,cstart+cworkA[l]+1,&tcol);CHKERRQ(ierr); 2254 if (tcol) { 2255 #else 2256 if ((tcol = cmap_i[cstart + cworkA[l]])) { 2257 #endif 2258 *mat_j++ = tcol - 1; 2259 if (!ijonly) { 2260 ierr = PetscMemcpy(mat_a,vworkA+l*bs2,bs2*sizeof(MatScalar));CHKERRQ(ierr); 2261 mat_a += bs2; 2262 } 2263 ilen_row++; 2264 } 2265 } 2266 for (l=imark; l<nzB; l++) { 2267 #if defined(PETSC_USE_CTABLE) 2268 ierr = PetscTableFind(cmap_i,bmap[cworkB[l]]+1,&tcol);CHKERRQ(ierr); 2269 if (tcol) { 2270 #else 2271 if ((tcol = cmap_i[bmap[cworkB[l]]])) { 2272 #endif 2273 *mat_j++ = tcol - 1; 2274 if (!ijonly) { 2275 ierr = PetscMemcpy(mat_a,vworkB+l*bs2,bs2*sizeof(MatScalar));CHKERRQ(ierr); 2276 mat_a += bs2; 2277 } 2278 ilen_row++; 2279 } 2280 } 2281 } else { /* allcolumns */ 2282 for (l=0; l<nzB; l++) { 2283 if ((ctmp = bmap[cworkB[l]]) < cstart) { 2284 *mat_j++ = ctmp; 2285 ierr = PetscMemcpy(mat_a,vworkB+l*bs2,bs2*sizeof(MatScalar));CHKERRQ(ierr); 2286 mat_a += bs2; 2287 ilen_row++; 2288 } else break; 2289 } 2290 imark = l; 2291 for (l=0; l<nzA; l++) { 2292 *mat_j++ = cstart+cworkA[l]; 2293 if (!ijonly) { 2294 ierr = PetscMemcpy(mat_a,vworkA+l*bs2,bs2*sizeof(MatScalar));CHKERRQ(ierr); 2295 mat_a += bs2; 2296 } 2297 ilen_row++; 2298 } 2299 for (l=imark; l<nzB; l++) { 2300 *mat_j++ = bmap[cworkB[l]]; 2301 if (!ijonly) { 2302 ierr = PetscMemcpy(mat_a,vworkB+l*bs2,bs2*sizeof(MatScalar));CHKERRQ(ierr); 2303 mat_a += bs2; 2304 } 2305 ilen_row++; 2306 } 2307 } 2308 imat_ilen[row] = ilen_row; 2309 } 2310 } 2311 } 2312 } 2313 2314 /* Now assemble the off proc rows*/ 2315 { 2316 PetscInt *sbuf1_i,*rbuf2_i,*rbuf3_i,*imat_ilen,ilen; 2317 PetscInt *imat_j,*imat_i; 2318 MatScalar *imat_a = NULL,*rbuf4_i = NULL; 2319 PetscMPIInt ii; 2320 2321 for (tmp2=0; tmp2<nrqs; tmp2++) { 2322 if (ijonly) ii = tmp2; 2323 else { 2324 ierr = MPI_Waitany(nrqs,r_waits4,&ii,r_status4+tmp2);CHKERRQ(ierr); 2325 } 2326 idex = pa[ii]; 2327 sbuf1_i = sbuf1[idex]; 2328 jmax = sbuf1_i[0]; 2329 ct1 = 2*jmax + 1; 2330 ct2 = 0; 2331 rbuf2_i = rbuf2[ii]; 2332 rbuf3_i = rbuf3[ii]; 2333 if (!ijonly) rbuf4_i = rbuf4[ii]; 2334 for (j=1; j<=jmax; j++) { 2335 is_no = sbuf1_i[2*j-1]; 2336 if (!allcolumns[is_no]) cmap_i = cmap[is_no]; 2337 rmap_i = rmap[is_no]; 2338 mat = (Mat_SeqBAIJ*)submats[is_no]->data; 2339 imat_ilen = mat->ilen; 2340 imat_j = mat->j; 2341 imat_i = mat->i; 2342 if (!ijonly) imat_a = mat->a; 2343 max1 = sbuf1_i[2*j]; 2344 for (k=0; k<max1; k++,ct1++) { 2345 row = sbuf1_i[ct1]; 2346 #if defined(PETSC_USE_CTABLE) 2347 ierr = PetscTableFind(rmap_i,row+1,&row);CHKERRQ(ierr); 2348 row--; 2349 if (row < 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"row not found in table"); 2350 #else 2351 row = rmap_i[row]; 2352 #endif 2353 ilen = imat_ilen[row]; 2354 mat_i = imat_i[row]; 2355 if (!ijonly) mat_a = imat_a + mat_i*bs2; 2356 mat_j = imat_j + mat_i; 2357 max2 = rbuf2_i[ct1]; 2358 2359 if (!allcolumns[is_no]) { 2360 for (l=0; l<max2; l++,ct2++) { 2361 #if defined(PETSC_USE_CTABLE) 2362 ierr = PetscTableFind(cmap_i,rbuf3_i[ct2]+1,&tcol);CHKERRQ(ierr); 2363 if (tcol) { 2364 #else 2365 if ((tcol = cmap_i[rbuf3_i[ct2]])) { 2366 #endif 2367 *mat_j++ = tcol - 1; 2368 if (!ijonly) { 2369 ierr = PetscMemcpy(mat_a,rbuf4_i+ct2*bs2,bs2*sizeof(MatScalar));CHKERRQ(ierr); 2370 mat_a += bs2; 2371 } 2372 ilen++; 2373 } 2374 } 2375 } else { /* allcolumns */ 2376 for (l=0; l<max2; l++,ct2++) { 2377 *mat_j++ = rbuf3_i[ct2]; 2378 if (!ijonly) { 2379 ierr = PetscMemcpy(mat_a,rbuf4_i+ct2*bs2,bs2*sizeof(MatScalar));CHKERRQ(ierr); 2380 mat_a += bs2; 2381 } 2382 ilen++; 2383 } 2384 } 2385 imat_ilen[row] = ilen; 2386 } 2387 } 2388 } 2389 } 2390 /* sort the rows */ 2391 { 2392 PetscInt ilen_row,*imat_ilen,*imat_j,*imat_i; 2393 MatScalar *imat_a = NULL; 2394 MatScalar *work; 2395 2396 ierr = PetscMalloc1(bs2,&work);CHKERRQ(ierr); 2397 for (i=0; i<ismax; i++) { 2398 mat = (Mat_SeqBAIJ*)submats[i]->data; 2399 imat_ilen = mat->ilen; 2400 imat_j = mat->j; 2401 imat_i = mat->i; 2402 if (!ijonly) imat_a = mat->a; 2403 if (allcolumns[i]) continue; 2404 jmax = nrow[i]; 2405 for (j=0; j<jmax; j++) { 2406 mat_i = imat_i[j]; 2407 if (!ijonly) mat_a = imat_a + mat_i*bs2; 2408 mat_j = imat_j + mat_i; 2409 ilen_row = imat_ilen[j]; 2410 if (!ijonly) {ierr = PetscSortIntWithDataArray(ilen_row,mat_j,mat_a,bs2*sizeof(MatScalar),work);CHKERRQ(ierr);} 2411 else {ierr = PetscSortInt(ilen_row,mat_j);CHKERRQ(ierr);} 2412 } 2413 } 2414 ierr = PetscFree(work);CHKERRQ(ierr); 2415 } 2416 if (!ijonly) { 2417 ierr = PetscFree(r_status4);CHKERRQ(ierr); 2418 ierr = PetscFree(r_waits4);CHKERRQ(ierr); 2419 if (nrqr) {ierr = MPI_Waitall(nrqr,s_waits4,s_status4);CHKERRQ(ierr);} 2420 ierr = PetscFree(s_waits4);CHKERRQ(ierr); 2421 ierr = PetscFree(s_status4);CHKERRQ(ierr); 2422 } 2423 2424 /* Restore the indices */ 2425 for (i=0; i<ismax; i++) { 2426 if (!allrows[i]) { 2427 ierr = ISRestoreIndices(isrow[i],irow+i);CHKERRQ(ierr); 2428 } 2429 if (!allcolumns[i]) { 2430 ierr = ISRestoreIndices(iscol[i],icol+i);CHKERRQ(ierr); 2431 } 2432 } 2433 2434 /* Destroy allocated memory */ 2435 #if defined(PETSC_USE_CTABLE) 2436 ierr = PetscFree4(irow,icol,nrow,ncol);CHKERRQ(ierr); 2437 #else 2438 ierr = PetscFree5(irow,icol,nrow,ncol,rtable);CHKERRQ(ierr); 2439 #endif 2440 ierr = PetscFree4(w1,w2,w3,w4);CHKERRQ(ierr); 2441 ierr = PetscFree(pa);CHKERRQ(ierr); 2442 2443 ierr = PetscFree4(sbuf1,ptr,tmp,ctr);CHKERRQ(ierr); 2444 ierr = PetscFree(sbuf1);CHKERRQ(ierr); 2445 ierr = PetscFree(rbuf2);CHKERRQ(ierr); 2446 for (i=0; i<nrqr; ++i) { 2447 ierr = PetscFree(sbuf2[i]);CHKERRQ(ierr); 2448 } 2449 for (i=0; i<nrqs; ++i) { 2450 ierr = PetscFree(rbuf3[i]);CHKERRQ(ierr); 2451 } 2452 ierr = PetscFree3(sbuf2,req_size,req_source);CHKERRQ(ierr); 2453 ierr = PetscFree(rbuf3);CHKERRQ(ierr); 2454 ierr = PetscFree(sbuf_aj[0]);CHKERRQ(ierr); 2455 ierr = PetscFree(sbuf_aj);CHKERRQ(ierr); 2456 if (!ijonly) { 2457 for (i=0; i<nrqs; ++i) {ierr = PetscFree(rbuf4[i]);CHKERRQ(ierr);} 2458 ierr = PetscFree(rbuf4);CHKERRQ(ierr); 2459 ierr = PetscFree(sbuf_aa[0]);CHKERRQ(ierr); 2460 ierr = PetscFree(sbuf_aa);CHKERRQ(ierr); 2461 } 2462 2463 #if defined(PETSC_USE_CTABLE) 2464 for (i=0; i<ismax; i++) { 2465 ierr = PetscTableDestroy((PetscTable*)&rmap[i]);CHKERRQ(ierr); 2466 } 2467 #endif 2468 ierr = PetscFree(rmap);CHKERRQ(ierr); 2469 2470 for (i=0; i<ismax; i++) { 2471 if (!allcolumns[i]) { 2472 #if defined(PETSC_USE_CTABLE) 2473 ierr = PetscTableDestroy((PetscTable*)&cmap[i]);CHKERRQ(ierr); 2474 #else 2475 ierr = PetscFree(cmap[i]);CHKERRQ(ierr); 2476 #endif 2477 } 2478 } 2479 ierr = PetscFree(cmap);CHKERRQ(ierr); 2480 ierr = PetscFree(lens);CHKERRQ(ierr); 2481 2482 for (i=0; i<ismax; i++) { 2483 ierr = MatAssemblyBegin(submats[i],MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2484 ierr = MatAssemblyEnd(submats[i],MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2485 } 2486 2487 c->ijonly = PETSC_FALSE; /* set back to the default */ 2488 PetscFunctionReturn(0); 2489 } 2490 2491