1 /* 2 Routines to compute overlapping regions of a parallel MPI matrix 3 and to find submatrices that were shared across processors. 4 */ 5 #include "src/mat/impls/aij/mpi/mpiaij.h" 6 #include "petscbt.h" 7 8 static int MatIncreaseOverlap_MPIAIJ_Once(Mat,int,IS *); 9 static int MatIncreaseOverlap_MPIAIJ_Local(Mat,int,char **,int*,int**); 10 static int MatIncreaseOverlap_MPIAIJ_Receive(Mat,int,int **,int**,int*); 11 EXTERN int MatGetRow_MPIAIJ(Mat,int,int*,int**,PetscScalar**); 12 EXTERN int MatRestoreRow_MPIAIJ(Mat,int,int*,int**,PetscScalar**); 13 14 #undef __FUNCT__ 15 #define __FUNCT__ "MatIncreaseOverlap_MPIAIJ" 16 int MatIncreaseOverlap_MPIAIJ(Mat C,int imax,IS is[],int ov) 17 { 18 int i,ierr; 19 20 PetscFunctionBegin; 21 if (ov < 0) SETERRQ(PETSC_ERR_ARG_OUTOFRANGE,"Negative overlap specified"); 22 for (i=0; i<ov; ++i) { 23 ierr = MatIncreaseOverlap_MPIAIJ_Once(C,imax,is);CHKERRQ(ierr); 24 } 25 PetscFunctionReturn(0); 26 } 27 28 /* 29 Sample message format: 30 If a processor A wants processor B to process some elements corresponding 31 to index sets is[1],is[5] 32 mesg [0] = 2 (no of index sets in the mesg) 33 ----------- 34 mesg [1] = 1 => is[1] 35 mesg [2] = sizeof(is[1]); 36 ----------- 37 mesg [3] = 5 => is[5] 38 mesg [4] = sizeof(is[5]); 39 ----------- 40 mesg [5] 41 mesg [n] datas[1] 42 ----------- 43 mesg[n+1] 44 mesg[m] data(is[5]) 45 ----------- 46 47 Notes: 48 nrqs - no of requests sent (or to be sent out) 49 nrqr - no of requests recieved (which have to be or which have been processed 50 */ 51 #undef __FUNCT__ 52 #define __FUNCT__ "MatIncreaseOverlap_MPIAIJ_Once" 53 static int MatIncreaseOverlap_MPIAIJ_Once(Mat C,int imax,IS is[]) 54 { 55 Mat_MPIAIJ *c = (Mat_MPIAIJ*)C->data; 56 int **idx,*n,*w1,*w2,*w3,*w4,*rtable,**data,len,*idx_i; 57 int size,rank,m,i,j,k,ierr,**rbuf,row,proc,nrqs,msz,**outdat,**ptr; 58 int *ctr,*pa,*tmp,nrqr,*isz,*isz1,**xdata,**rbuf2; 59 int *onodes1,*olengths1,tag1,tag2,*onodes2,*olengths2; 60 PetscBT *table; 61 MPI_Comm comm; 62 MPI_Request *s_waits1,*r_waits1,*s_waits2,*r_waits2; 63 MPI_Status *s_status,*recv_status; 64 65 PetscFunctionBegin; 66 comm = C->comm; 67 size = c->size; 68 rank = c->rank; 69 m = C->M; 70 71 ierr = PetscObjectGetNewTag((PetscObject)C,&tag1);CHKERRQ(ierr); 72 ierr = PetscObjectGetNewTag((PetscObject)C,&tag2);CHKERRQ(ierr); 73 74 len = (imax+1)*sizeof(int*)+ (imax + m)*sizeof(int); 75 ierr = PetscMalloc(len,&idx);CHKERRQ(ierr); 76 n = (int*)(idx + imax); 77 rtable = n + imax; 78 79 for (i=0; i<imax; i++) { 80 ierr = ISGetIndices(is[i],&idx[i]);CHKERRQ(ierr); 81 ierr = ISGetLocalSize(is[i],&n[i]);CHKERRQ(ierr); 82 } 83 84 /* Create hash table for the mapping :row -> proc*/ 85 for (i=0,j=0; i<size; i++) { 86 len = c->rowners[i+1]; 87 for (; j<len; j++) { 88 rtable[j] = i; 89 } 90 } 91 92 /* evaluate communication - mesg to who,length of mesg, and buffer space 93 required. Based on this, buffers are allocated, and data copied into them*/ 94 ierr = PetscMalloc(size*4*sizeof(int),&w1);CHKERRQ(ierr);/* mesg size */ 95 w2 = w1 + size; /* if w2[i] marked, then a message to proc i*/ 96 w3 = w2 + size; /* no of IS that needs to be sent to proc i */ 97 w4 = w3 + size; /* temp work space used in determining w1, w2, w3 */ 98 ierr = PetscMemzero(w1,size*3*sizeof(int));CHKERRQ(ierr); /* initialise work vector*/ 99 for (i=0; i<imax; i++) { 100 ierr = PetscMemzero(w4,size*sizeof(int));CHKERRQ(ierr); /* initialise work vector*/ 101 idx_i = idx[i]; 102 len = n[i]; 103 for (j=0; j<len; j++) { 104 row = idx_i[j]; 105 if (row < 0) { 106 SETERRQ(1,"Index set cannot have negative entries"); 107 } 108 proc = rtable[row]; 109 w4[proc]++; 110 } 111 for (j=0; j<size; j++){ 112 if (w4[j]) { w1[j] += w4[j]; w3[j]++;} 113 } 114 } 115 116 nrqs = 0; /* no of outgoing messages */ 117 msz = 0; /* total mesg length (for all proc */ 118 w1[rank] = 0; /* no mesg sent to intself */ 119 w3[rank] = 0; 120 for (i=0; i<size; i++) { 121 if (w1[i]) {w2[i] = 1; nrqs++;} /* there exists a message to proc i */ 122 } 123 /* pa - is list of processors to communicate with */ 124 ierr = PetscMalloc((nrqs+1)*sizeof(int),&pa);CHKERRQ(ierr); 125 for (i=0,j=0; i<size; i++) { 126 if (w1[i]) {pa[j] = i; j++;} 127 } 128 129 /* Each message would have a header = 1 + 2*(no of IS) + data */ 130 for (i=0; i<nrqs; i++) { 131 j = pa[i]; 132 w1[j] += w2[j] + 2*w3[j]; 133 msz += w1[j]; 134 } 135 136 /* Determine the number of messages to expect, their lengths, from from-ids */ 137 ierr = PetscGatherNumberOfMessages(comm,w2,w1,&nrqr);CHKERRQ(ierr); 138 ierr = PetscGatherMessageLengths(comm,nrqs,nrqr,w1,&onodes1,&olengths1);CHKERRQ(ierr); 139 140 /* Now post the Irecvs corresponding to these messages */ 141 ierr = PetscPostIrecvInt(comm,tag1,nrqr,onodes1,olengths1,&rbuf,&r_waits1);CHKERRQ(ierr); 142 143 /* Allocate Memory for outgoing messages */ 144 len = 2*size*sizeof(int*) + (size+msz)*sizeof(int); 145 ierr = PetscMalloc(len,&outdat);CHKERRQ(ierr); 146 ptr = outdat + size; /* Pointers to the data in outgoing buffers */ 147 ierr = PetscMemzero(outdat,2*size*sizeof(int*));CHKERRQ(ierr); 148 tmp = (int*)(outdat + 2*size); 149 ctr = tmp + msz; 150 151 { 152 int *iptr = tmp,ict = 0; 153 for (i=0; i<nrqs; i++) { 154 j = pa[i]; 155 iptr += ict; 156 outdat[j] = iptr; 157 ict = w1[j]; 158 } 159 } 160 161 /* Form the outgoing messages */ 162 /*plug in the headers*/ 163 for (i=0; i<nrqs; i++) { 164 j = pa[i]; 165 outdat[j][0] = 0; 166 ierr = PetscMemzero(outdat[j]+1,2*w3[j]*sizeof(int));CHKERRQ(ierr); 167 ptr[j] = outdat[j] + 2*w3[j] + 1; 168 } 169 170 /* Memory for doing local proc's work*/ 171 { 172 int *d_p; 173 char *t_p; 174 175 len = (imax)*(sizeof(PetscBT) + sizeof(int*)+ sizeof(int)) + 176 (m)*imax*sizeof(int) + (m/PETSC_BITS_PER_BYTE+1)*imax*sizeof(char) + 1; 177 ierr = PetscMalloc(len,&table);CHKERRQ(ierr); 178 ierr = PetscMemzero(table,len);CHKERRQ(ierr); 179 data = (int **)(table + imax); 180 isz = (int *)(data + imax); 181 d_p = (int *)(isz + imax); 182 t_p = (char *)(d_p + m*imax); 183 for (i=0; i<imax; i++) { 184 table[i] = t_p + (m/PETSC_BITS_PER_BYTE+1)*i; 185 data[i] = d_p + (m)*i; 186 } 187 } 188 189 /* Parse the IS and update local tables and the outgoing buf with the data*/ 190 { 191 int n_i,*data_i,isz_i,*outdat_j,ctr_j; 192 PetscBT table_i; 193 194 for (i=0; i<imax; i++) { 195 ierr = PetscMemzero(ctr,size*sizeof(int));CHKERRQ(ierr); 196 n_i = n[i]; 197 table_i = table[i]; 198 idx_i = idx[i]; 199 data_i = data[i]; 200 isz_i = isz[i]; 201 for (j=0; j<n_i; j++) { /* parse the indices of each IS */ 202 row = idx_i[j]; 203 proc = rtable[row]; 204 if (proc != rank) { /* copy to the outgoing buffer */ 205 ctr[proc]++; 206 *ptr[proc] = row; 207 ptr[proc]++; 208 } else { /* Update the local table */ 209 if (!PetscBTLookupSet(table_i,row)) { data_i[isz_i++] = row;} 210 } 211 } 212 /* Update the headers for the current IS */ 213 for (j=0; j<size; j++) { /* Can Optimise this loop by using pa[] */ 214 if ((ctr_j = ctr[j])) { 215 outdat_j = outdat[j]; 216 k = ++outdat_j[0]; 217 outdat_j[2*k] = ctr_j; 218 outdat_j[2*k-1] = i; 219 } 220 } 221 isz[i] = isz_i; 222 } 223 } 224 225 226 227 /* Now post the sends */ 228 ierr = PetscMalloc((nrqs+1)*sizeof(MPI_Request),&s_waits1);CHKERRQ(ierr); 229 for (i=0; i<nrqs; ++i) { 230 j = pa[i]; 231 ierr = MPI_Isend(outdat[j],w1[j],MPI_INT,j,tag1,comm,s_waits1+i);CHKERRQ(ierr); 232 } 233 234 /* No longer need the original indices*/ 235 for (i=0; i<imax; ++i) { 236 ierr = ISRestoreIndices(is[i],idx+i);CHKERRQ(ierr); 237 } 238 ierr = PetscFree(idx);CHKERRQ(ierr); 239 240 for (i=0; i<imax; ++i) { 241 ierr = ISDestroy(is[i]);CHKERRQ(ierr); 242 } 243 244 /* Do Local work*/ 245 ierr = MatIncreaseOverlap_MPIAIJ_Local(C,imax,table,isz,data);CHKERRQ(ierr); 246 247 /* Receive messages*/ 248 ierr = PetscMalloc((nrqr+1)*sizeof(MPI_Status),&recv_status);CHKERRQ(ierr); 249 ierr = MPI_Waitall(nrqr,r_waits1,recv_status);CHKERRQ(ierr); 250 251 ierr = PetscMalloc((nrqs+1)*sizeof(MPI_Status),&s_status);CHKERRQ(ierr); 252 ierr = MPI_Waitall(nrqs,s_waits1,s_status);CHKERRQ(ierr); 253 254 /* Phase 1 sends are complete - deallocate buffers */ 255 ierr = PetscFree(outdat);CHKERRQ(ierr); 256 ierr = PetscFree(w1);CHKERRQ(ierr); 257 258 ierr = PetscMalloc((nrqr+1)*sizeof(int*),&xdata);CHKERRQ(ierr); 259 ierr = PetscMalloc((nrqr+1)*sizeof(int),&isz1);CHKERRQ(ierr); 260 ierr = MatIncreaseOverlap_MPIAIJ_Receive(C,nrqr,rbuf,xdata,isz1);CHKERRQ(ierr); 261 ierr = PetscFree(rbuf);CHKERRQ(ierr); 262 263 264 /* Send the data back*/ 265 /* Do a global reduction to know the buffer space req for incoming messages*/ 266 { 267 int *rw1; 268 269 ierr = PetscMalloc(size*sizeof(int),&rw1);CHKERRQ(ierr); 270 ierr = PetscMemzero(rw1,size*sizeof(int));CHKERRQ(ierr); 271 272 for (i=0; i<nrqr; ++i) { 273 proc = recv_status[i].MPI_SOURCE; 274 if (proc != onodes1[i]) SETERRQ(1,"MPI_SOURCE mismatch"); 275 rw1[proc] = isz1[i]; 276 } 277 ierr = PetscFree(onodes1);CHKERRQ(ierr); 278 ierr = PetscFree(olengths1);CHKERRQ(ierr); 279 280 /* Determine the number of messages to expect, their lengths, from from-ids */ 281 ierr = PetscGatherMessageLengths(comm,nrqr,nrqs,rw1,&onodes2,&olengths2);CHKERRQ(ierr); 282 PetscFree(rw1); 283 } 284 /* Now post the Irecvs corresponding to these messages */ 285 ierr = PetscPostIrecvInt(comm,tag2,nrqs,onodes2,olengths2,&rbuf2,&r_waits2);CHKERRQ(ierr); 286 287 /* Now post the sends */ 288 ierr = PetscMalloc((nrqr+1)*sizeof(MPI_Request),&s_waits2);CHKERRQ(ierr); 289 for (i=0; i<nrqr; ++i) { 290 j = recv_status[i].MPI_SOURCE; 291 ierr = MPI_Isend(xdata[i],isz1[i],MPI_INT,j,tag2,comm,s_waits2+i);CHKERRQ(ierr); 292 } 293 294 /* receive work done on other processors*/ 295 { 296 int idex,is_no,ct1,max,*rbuf2_i,isz_i,*data_i,jmax; 297 PetscBT table_i; 298 MPI_Status *status2; 299 300 ierr = PetscMalloc((PetscMax(nrqr,nrqs)+1)*sizeof(MPI_Status),&status2);CHKERRQ(ierr); 301 for (i=0; i<nrqs; ++i) { 302 ierr = MPI_Waitany(nrqs,r_waits2,&idex,status2+i);CHKERRQ(ierr); 303 /* Process the message*/ 304 rbuf2_i = rbuf2[idex]; 305 ct1 = 2*rbuf2_i[0]+1; 306 jmax = rbuf2[idex][0]; 307 for (j=1; j<=jmax; j++) { 308 max = rbuf2_i[2*j]; 309 is_no = rbuf2_i[2*j-1]; 310 isz_i = isz[is_no]; 311 data_i = data[is_no]; 312 table_i = table[is_no]; 313 for (k=0; k<max; k++,ct1++) { 314 row = rbuf2_i[ct1]; 315 if (!PetscBTLookupSet(table_i,row)) { data_i[isz_i++] = row;} 316 } 317 isz[is_no] = isz_i; 318 } 319 } 320 321 ierr = MPI_Waitall(nrqr,s_waits2,status2);CHKERRQ(ierr); 322 ierr = PetscFree(status2);CHKERRQ(ierr); 323 } 324 325 for (i=0; i<imax; ++i) { 326 ierr = ISCreateGeneral(PETSC_COMM_SELF,isz[i],data[i],is+i);CHKERRQ(ierr); 327 } 328 329 ierr = PetscFree(onodes2);CHKERRQ(ierr); 330 ierr = PetscFree(olengths2);CHKERRQ(ierr); 331 332 ierr = PetscFree(pa);CHKERRQ(ierr); 333 ierr = PetscFree(rbuf2);CHKERRQ(ierr); 334 ierr = PetscFree(s_waits1);CHKERRQ(ierr); 335 ierr = PetscFree(r_waits1);CHKERRQ(ierr); 336 ierr = PetscFree(s_waits2);CHKERRQ(ierr); 337 ierr = PetscFree(r_waits2);CHKERRQ(ierr); 338 ierr = PetscFree(table);CHKERRQ(ierr); 339 ierr = PetscFree(s_status);CHKERRQ(ierr); 340 ierr = PetscFree(recv_status);CHKERRQ(ierr); 341 ierr = PetscFree(xdata[0]);CHKERRQ(ierr); 342 ierr = PetscFree(xdata);CHKERRQ(ierr); 343 ierr = PetscFree(isz1);CHKERRQ(ierr); 344 PetscFunctionReturn(0); 345 } 346 347 #undef __FUNCT__ 348 #define __FUNCT__ "MatIncreaseOverlap_MPIAIJ_Local" 349 /* 350 MatIncreaseOverlap_MPIAIJ_Local - Called by MatincreaseOverlap, to do 351 the work on the local processor. 352 353 Inputs: 354 C - MAT_MPIAIJ; 355 imax - total no of index sets processed at a time; 356 table - an array of char - size = m bits. 357 358 Output: 359 isz - array containing the count of the solution elements correspondign 360 to each index set; 361 data - pointer to the solutions 362 */ 363 static int MatIncreaseOverlap_MPIAIJ_Local(Mat C,int imax,PetscBT *table,int *isz,int **data) 364 { 365 Mat_MPIAIJ *c = (Mat_MPIAIJ*)C->data; 366 Mat A = c->A,B = c->B; 367 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data,*b = (Mat_SeqAIJ*)B->data; 368 int start,end,val,max,rstart,cstart,*ai,*aj; 369 int *bi,*bj,*garray,i,j,k,row,*data_i,isz_i; 370 PetscBT table_i; 371 372 PetscFunctionBegin; 373 rstart = c->rstart; 374 cstart = c->cstart; 375 ai = a->i; 376 aj = a->j; 377 bi = b->i; 378 bj = b->j; 379 garray = c->garray; 380 381 382 for (i=0; i<imax; i++) { 383 data_i = data[i]; 384 table_i = table[i]; 385 isz_i = isz[i]; 386 for (j=0,max=isz[i]; j<max; j++) { 387 row = data_i[j] - rstart; 388 start = ai[row]; 389 end = ai[row+1]; 390 for (k=start; k<end; k++) { /* Amat */ 391 val = aj[k] + cstart; 392 if (!PetscBTLookupSet(table_i,val)) { data_i[isz_i++] = val;} 393 } 394 start = bi[row]; 395 end = bi[row+1]; 396 for (k=start; k<end; k++) { /* Bmat */ 397 val = garray[bj[k]]; 398 if (!PetscBTLookupSet(table_i,val)) { data_i[isz_i++] = val;} 399 } 400 } 401 isz[i] = isz_i; 402 } 403 PetscFunctionReturn(0); 404 } 405 406 #undef __FUNCT__ 407 #define __FUNCT__ "MatIncreaseOverlap_MPIAIJ_Receive" 408 /* 409 MatIncreaseOverlap_MPIAIJ_Receive - Process the recieved messages, 410 and return the output 411 412 Input: 413 C - the matrix 414 nrqr - no of messages being processed. 415 rbuf - an array of pointers to the recieved requests 416 417 Output: 418 xdata - array of messages to be sent back 419 isz1 - size of each message 420 421 For better efficiency perhaps we should malloc seperately each xdata[i], 422 then if a remalloc is required we need only copy the data for that one row 423 rather then all previous rows as it is now where a single large chunck of 424 memory is used. 425 426 */ 427 static int MatIncreaseOverlap_MPIAIJ_Receive(Mat C,int nrqr,int **rbuf,int **xdata,int * isz1) 428 { 429 Mat_MPIAIJ *c = (Mat_MPIAIJ*)C->data; 430 Mat A = c->A,B = c->B; 431 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data,*b = (Mat_SeqAIJ*)B->data; 432 int rstart,cstart,*ai,*aj,*bi,*bj,*garray,i,j,k; 433 int row,total_sz,ct,ct1,ct2,ct3,mem_estimate,oct2,l,start,end; 434 int val,max1,max2,rank,m,no_malloc =0,*tmp,new_estimate,ctr; 435 int *rbuf_i,kmax,rbuf_0,ierr; 436 PetscBT xtable; 437 438 PetscFunctionBegin; 439 rank = c->rank; 440 m = C->M; 441 rstart = c->rstart; 442 cstart = c->cstart; 443 ai = a->i; 444 aj = a->j; 445 bi = b->i; 446 bj = b->j; 447 garray = c->garray; 448 449 450 for (i=0,ct=0,total_sz=0; i<nrqr; ++i) { 451 rbuf_i = rbuf[i]; 452 rbuf_0 = rbuf_i[0]; 453 ct += rbuf_0; 454 for (j=1; j<=rbuf_0; j++) { total_sz += rbuf_i[2*j]; } 455 } 456 457 if (C->m) max1 = ct*(a->nz + b->nz)/C->m; 458 else max1 = 1; 459 mem_estimate = 3*((total_sz > max1 ? total_sz : max1)+1); 460 ierr = PetscMalloc(mem_estimate*sizeof(int),&xdata[0]);CHKERRQ(ierr); 461 ++no_malloc; 462 ierr = PetscBTCreate(m,xtable);CHKERRQ(ierr); 463 ierr = PetscMemzero(isz1,nrqr*sizeof(int));CHKERRQ(ierr); 464 465 ct3 = 0; 466 for (i=0; i<nrqr; i++) { /* for easch mesg from proc i */ 467 rbuf_i = rbuf[i]; 468 rbuf_0 = rbuf_i[0]; 469 ct1 = 2*rbuf_0+1; 470 ct2 = ct1; 471 ct3 += ct1; 472 for (j=1; j<=rbuf_0; j++) { /* for each IS from proc i*/ 473 ierr = PetscBTMemzero(m,xtable);CHKERRQ(ierr); 474 oct2 = ct2; 475 kmax = rbuf_i[2*j]; 476 for (k=0; k<kmax; k++,ct1++) { 477 row = rbuf_i[ct1]; 478 if (!PetscBTLookupSet(xtable,row)) { 479 if (!(ct3 < mem_estimate)) { 480 new_estimate = (int)(1.5*mem_estimate)+1; 481 ierr = PetscMalloc(new_estimate*sizeof(int),&tmp);CHKERRQ(ierr); 482 ierr = PetscMemcpy(tmp,xdata[0],mem_estimate*sizeof(int));CHKERRQ(ierr); 483 ierr = PetscFree(xdata[0]);CHKERRQ(ierr); 484 xdata[0] = tmp; 485 mem_estimate = new_estimate; ++no_malloc; 486 for (ctr=1; ctr<=i; ctr++) { xdata[ctr] = xdata[ctr-1] + isz1[ctr-1];} 487 } 488 xdata[i][ct2++] = row; 489 ct3++; 490 } 491 } 492 for (k=oct2,max2=ct2; k<max2; k++) { 493 row = xdata[i][k] - rstart; 494 start = ai[row]; 495 end = ai[row+1]; 496 for (l=start; l<end; l++) { 497 val = aj[l] + cstart; 498 if (!PetscBTLookupSet(xtable,val)) { 499 if (!(ct3 < mem_estimate)) { 500 new_estimate = (int)(1.5*mem_estimate)+1; 501 ierr = PetscMalloc(new_estimate*sizeof(int),&tmp);CHKERRQ(ierr); 502 ierr = PetscMemcpy(tmp,xdata[0],mem_estimate*sizeof(int));CHKERRQ(ierr); 503 ierr = PetscFree(xdata[0]);CHKERRQ(ierr); 504 xdata[0] = tmp; 505 mem_estimate = new_estimate; ++no_malloc; 506 for (ctr=1; ctr<=i; ctr++) { xdata[ctr] = xdata[ctr-1] + isz1[ctr-1];} 507 } 508 xdata[i][ct2++] = val; 509 ct3++; 510 } 511 } 512 start = bi[row]; 513 end = bi[row+1]; 514 for (l=start; l<end; l++) { 515 val = garray[bj[l]]; 516 if (!PetscBTLookupSet(xtable,val)) { 517 if (!(ct3 < mem_estimate)) { 518 new_estimate = (int)(1.5*mem_estimate)+1; 519 ierr = PetscMalloc(new_estimate*sizeof(int),&tmp);CHKERRQ(ierr); 520 ierr = PetscMemcpy(tmp,xdata[0],mem_estimate*sizeof(int));CHKERRQ(ierr); 521 ierr = PetscFree(xdata[0]);CHKERRQ(ierr); 522 xdata[0] = tmp; 523 mem_estimate = new_estimate; ++no_malloc; 524 for (ctr =1; ctr <=i; ctr++) { xdata[ctr] = xdata[ctr-1] + isz1[ctr-1];} 525 } 526 xdata[i][ct2++] = val; 527 ct3++; 528 } 529 } 530 } 531 /* Update the header*/ 532 xdata[i][2*j] = ct2 - oct2; /* Undo the vector isz1 and use only a var*/ 533 xdata[i][2*j-1] = rbuf_i[2*j-1]; 534 } 535 xdata[i][0] = rbuf_0; 536 xdata[i+1] = xdata[i] + ct2; 537 isz1[i] = ct2; /* size of each message */ 538 } 539 ierr = PetscBTDestroy(xtable);CHKERRQ(ierr); 540 PetscLogInfo(0,"MatIncreaseOverlap_MPIAIJ:[%d] Allocated %d bytes, required %d bytes, no of mallocs = %d\n",rank,mem_estimate, ct3,no_malloc); 541 PetscFunctionReturn(0); 542 } 543 /* -------------------------------------------------------------------------*/ 544 EXTERN int MatGetSubMatrices_MPIAIJ_Local(Mat,int,const IS[],const IS[],MatReuse,Mat*); 545 EXTERN int MatAssemblyEnd_SeqAIJ(Mat,MatAssemblyType); 546 /* 547 Every processor gets the entire matrix 548 */ 549 #undef __FUNCT__ 550 #define __FUNCT__ "MatGetSubMatrix_MPIAIJ_All" 551 int MatGetSubMatrix_MPIAIJ_All(Mat A,MatReuse scall,Mat *Bin[]) 552 { 553 Mat B; 554 Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data; 555 Mat_SeqAIJ *b,*ad = (Mat_SeqAIJ*)a->A->data,*bd = (Mat_SeqAIJ*)a->B->data; 556 int ierr,sendcount,*recvcounts = 0,*displs = 0,size,i,*rstarts = a->rowners,rank,n,cnt,j; 557 int m,*b_sendj,*garray = a->garray,*lens,*jsendbuf,*a_jsendbuf,*b_jsendbuf; 558 PetscScalar *sendbuf,*recvbuf,*a_sendbuf,*b_sendbuf; 559 560 PetscFunctionBegin; 561 ierr = MPI_Comm_size(A->comm,&size);CHKERRQ(ierr); 562 ierr = MPI_Comm_rank(A->comm,&rank);CHKERRQ(ierr); 563 564 if (scall == MAT_INITIAL_MATRIX) { 565 /* ---------------------------------------------------------------- 566 Tell every processor the number of nonzeros per row 567 */ 568 ierr = PetscMalloc(A->M*sizeof(int),&lens);CHKERRQ(ierr); 569 for (i=a->rstart; i<a->rend; i++) { 570 lens[i] = ad->i[i-a->rstart+1] - ad->i[i-a->rstart] + bd->i[i-a->rstart+1] - bd->i[i-a->rstart]; 571 } 572 sendcount = a->rend - a->rstart; 573 ierr = PetscMalloc(2*size*sizeof(int),&recvcounts);CHKERRQ(ierr); 574 displs = recvcounts + size; 575 for (i=0; i<size; i++) { 576 recvcounts[i] = a->rowners[i+1] - a->rowners[i]; 577 displs[i] = a->rowners[i]; 578 } 579 ierr = MPI_Allgatherv(lens+a->rstart,sendcount,MPI_INT,lens,recvcounts,displs,MPI_INT,A->comm);CHKERRQ(ierr); 580 581 /* --------------------------------------------------------------- 582 Create the sequential matrix of the same type as the local block diagonal 583 */ 584 ierr = MatCreate(PETSC_COMM_SELF,A->M,A->N,PETSC_DETERMINE,PETSC_DETERMINE,&B);CHKERRQ(ierr); 585 ierr = MatSetType(B,a->A->type_name);CHKERRQ(ierr); 586 ierr = MatSeqAIJSetPreallocation(B,0,lens);CHKERRQ(ierr); 587 ierr = PetscMalloc(sizeof(Mat),Bin);CHKERRQ(ierr); 588 **Bin = B; 589 b = (Mat_SeqAIJ *)B->data; 590 591 /*-------------------------------------------------------------------- 592 Copy my part of matrix column indices over 593 */ 594 sendcount = ad->nz + bd->nz; 595 jsendbuf = b->j + b->i[rstarts[rank]]; 596 a_jsendbuf = ad->j; 597 b_jsendbuf = bd->j; 598 n = a->rend - a->rstart; 599 cnt = 0; 600 for (i=0; i<n; i++) { 601 602 /* put in lower diagonal portion */ 603 m = bd->i[i+1] - bd->i[i]; 604 while (m > 0) { 605 /* is it above diagonal (in bd (compressed) numbering) */ 606 if (garray[*b_jsendbuf] > a->rstart + i) break; 607 jsendbuf[cnt++] = garray[*b_jsendbuf++]; 608 m--; 609 } 610 611 /* put in diagonal portion */ 612 for (j=ad->i[i]; j<ad->i[i+1]; j++) { 613 jsendbuf[cnt++] = a->rstart + *a_jsendbuf++; 614 } 615 616 /* put in upper diagonal portion */ 617 while (m-- > 0) { 618 jsendbuf[cnt++] = garray[*b_jsendbuf++]; 619 } 620 } 621 if (cnt != sendcount) SETERRQ2(1,"Corrupted PETSc matrix: nz given %d actual nz %d",sendcount,cnt); 622 623 /*-------------------------------------------------------------------- 624 Gather all column indices to all processors 625 */ 626 for (i=0; i<size; i++) { 627 recvcounts[i] = 0; 628 for (j=a->rowners[i]; j<a->rowners[i+1]; j++) { 629 recvcounts[i] += lens[j]; 630 } 631 } 632 displs[0] = 0; 633 for (i=1; i<size; i++) { 634 displs[i] = displs[i-1] + recvcounts[i-1]; 635 } 636 ierr = MPI_Allgatherv(jsendbuf,sendcount,MPI_INT,b->j,recvcounts,displs,MPI_INT,A->comm);CHKERRQ(ierr); 637 638 /*-------------------------------------------------------------------- 639 Assemble the matrix into useable form (note numerical values not yet set) 640 */ 641 /* set the b->ilen (length of each row) values */ 642 ierr = PetscMemcpy(b->ilen,lens,A->M*sizeof(int));CHKERRQ(ierr); 643 /* set the b->i indices */ 644 b->i[0] = 0; 645 for (i=1; i<=A->M; i++) { 646 b->i[i] = b->i[i-1] + lens[i-1]; 647 } 648 ierr = PetscFree(lens);CHKERRQ(ierr); 649 ierr = MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 650 ierr = MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 651 652 } else { 653 B = **Bin; 654 b = (Mat_SeqAIJ *)B->data; 655 } 656 657 /*-------------------------------------------------------------------- 658 Copy my part of matrix numerical values into the values location 659 */ 660 sendcount = ad->nz + bd->nz; 661 sendbuf = b->a + b->i[rstarts[rank]]; 662 a_sendbuf = ad->a; 663 b_sendbuf = bd->a; 664 b_sendj = bd->j; 665 n = a->rend - a->rstart; 666 cnt = 0; 667 for (i=0; i<n; i++) { 668 669 /* put in lower diagonal portion */ 670 m = bd->i[i+1] - bd->i[i]; 671 while (m > 0) { 672 /* is it above diagonal (in bd (compressed) numbering) */ 673 if (garray[*b_sendj] > a->rstart + i) break; 674 sendbuf[cnt++] = *b_sendbuf++; 675 m--; 676 b_sendj++; 677 } 678 679 /* put in diagonal portion */ 680 for (j=ad->i[i]; j<ad->i[i+1]; j++) { 681 sendbuf[cnt++] = *a_sendbuf++; 682 } 683 684 /* put in upper diagonal portion */ 685 while (m-- > 0) { 686 sendbuf[cnt++] = *b_sendbuf++; 687 b_sendj++; 688 } 689 } 690 if (cnt != sendcount) SETERRQ2(1,"Corrupted PETSc matrix: nz given %d actual nz %d",sendcount,cnt); 691 692 /* ----------------------------------------------------------------- 693 Gather all numerical values to all processors 694 */ 695 if (!recvcounts) { 696 ierr = PetscMalloc(2*size*sizeof(int),&recvcounts);CHKERRQ(ierr); 697 displs = recvcounts + size; 698 } 699 for (i=0; i<size; i++) { 700 recvcounts[i] = b->i[rstarts[i+1]] - b->i[rstarts[i]]; 701 } 702 displs[0] = 0; 703 for (i=1; i<size; i++) { 704 displs[i] = displs[i-1] + recvcounts[i-1]; 705 } 706 recvbuf = b->a; 707 ierr = MPI_Allgatherv(sendbuf,sendcount,MPIU_SCALAR,recvbuf,recvcounts,displs,MPIU_SCALAR,A->comm);CHKERRQ(ierr); 708 ierr = PetscFree(recvcounts);CHKERRQ(ierr); 709 if (A->symmetric){ 710 ierr = MatSetOption(B,MAT_SYMMETRIC);CHKERRQ(ierr); 711 } else if (A->hermitian) { 712 ierr = MatSetOption(B,MAT_HERMITIAN);CHKERRQ(ierr); 713 } else if (A->structurally_symmetric) { 714 ierr = MatSetOption(B,MAT_STRUCTURALLY_SYMMETRIC);CHKERRQ(ierr); 715 } 716 717 PetscFunctionReturn(0); 718 } 719 720 #undef __FUNCT__ 721 #define __FUNCT__ "MatGetSubMatrices_MPIAIJ" 722 int MatGetSubMatrices_MPIAIJ(Mat C,int ismax,const IS isrow[],const IS iscol[],MatReuse scall,Mat *submat[]) 723 { 724 int nmax,nstages_local,nstages,i,pos,max_no,ierr,nrow,ncol; 725 PetscTruth rowflag,colflag,wantallmatrix = PETSC_FALSE,twantallmatrix; 726 727 PetscFunctionBegin; 728 /* 729 Check for special case each processor gets entire matrix 730 */ 731 if (ismax == 1 && C->M == C->N) { 732 ierr = ISIdentity(*isrow,&rowflag);CHKERRQ(ierr); 733 ierr = ISIdentity(*iscol,&colflag);CHKERRQ(ierr); 734 ierr = ISGetLocalSize(*isrow,&nrow);CHKERRQ(ierr); 735 ierr = ISGetLocalSize(*iscol,&ncol);CHKERRQ(ierr); 736 if (rowflag && colflag && nrow == C->M && ncol == C->N) { 737 wantallmatrix = PETSC_TRUE; 738 ierr = PetscOptionsGetLogical(C->prefix,"-use_fast_submatrix",&wantallmatrix,PETSC_NULL);CHKERRQ(ierr); 739 } 740 } 741 ierr = MPI_Allreduce(&wantallmatrix,&twantallmatrix,1,MPI_INT,MPI_MIN,C->comm);CHKERRQ(ierr); 742 if (twantallmatrix) { 743 ierr = MatGetSubMatrix_MPIAIJ_All(C,scall,submat);CHKERRQ(ierr); 744 PetscFunctionReturn(0); 745 } 746 747 /* Allocate memory to hold all the submatrices */ 748 if (scall != MAT_REUSE_MATRIX) { 749 ierr = PetscMalloc((ismax+1)*sizeof(Mat),submat);CHKERRQ(ierr); 750 } 751 /* Determine the number of stages through which submatrices are done */ 752 nmax = 20*1000000 / (C->N * sizeof(int)); 753 if (!nmax) nmax = 1; 754 nstages_local = ismax/nmax + ((ismax % nmax)?1:0); 755 756 /* Make sure every processor loops through the nstages */ 757 ierr = MPI_Allreduce(&nstages_local,&nstages,1,MPI_INT,MPI_MAX,C->comm);CHKERRQ(ierr); 758 759 for (i=0,pos=0; i<nstages; i++) { 760 if (pos+nmax <= ismax) max_no = nmax; 761 else if (pos == ismax) max_no = 0; 762 else max_no = ismax-pos; 763 ierr = MatGetSubMatrices_MPIAIJ_Local(C,max_no,isrow+pos,iscol+pos,scall,*submat+pos);CHKERRQ(ierr); 764 pos += max_no; 765 } 766 PetscFunctionReturn(0); 767 } 768 /* -------------------------------------------------------------------------*/ 769 #undef __FUNCT__ 770 #define __FUNCT__ "MatGetSubMatrices_MPIAIJ_Local" 771 int MatGetSubMatrices_MPIAIJ_Local(Mat C,int ismax,const IS isrow[],const IS iscol[],MatReuse scall,Mat *submats) 772 { 773 Mat_MPIAIJ *c = (Mat_MPIAIJ*)C->data; 774 Mat A = c->A; 775 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data,*b = (Mat_SeqAIJ*)c->B->data,*mat; 776 int **irow,**icol,*nrow,*ncol,*w1,*w2,*w3,*w4,*rtable,start,end,size; 777 int **sbuf1,**sbuf2,rank,m,i,j,k,l,ct1,ct2,ierr,**rbuf1,row,proc; 778 int nrqs,msz,**ptr,idex,*req_size,*ctr,*pa,*tmp,tcol,nrqr; 779 int **rbuf3,*req_source,**sbuf_aj,**rbuf2,max1,max2,**rmap; 780 int **cmap,**lens,is_no,ncols,*cols,mat_i,*mat_j,tmp2,jmax,*irow_i; 781 int len,ctr_j,*sbuf1_j,*sbuf_aj_i,*rbuf1_i,kmax,*cmap_i,*lens_i; 782 int *rmap_i,tag0,tag1,tag2,tag3; 783 MPI_Request *s_waits1,*r_waits1,*s_waits2,*r_waits2,*r_waits3; 784 MPI_Request *r_waits4,*s_waits3,*s_waits4; 785 MPI_Status *r_status1,*r_status2,*s_status1,*s_status3,*s_status2; 786 MPI_Status *r_status3,*r_status4,*s_status4; 787 MPI_Comm comm; 788 PetscScalar **rbuf4,**sbuf_aa,*vals,*mat_a,*sbuf_aa_i; 789 PetscTruth sorted; 790 int *onodes1,*olengths1; 791 792 PetscFunctionBegin; 793 comm = C->comm; 794 tag0 = C->tag; 795 size = c->size; 796 rank = c->rank; 797 m = C->M; 798 799 /* Get some new tags to keep the communication clean */ 800 ierr = PetscObjectGetNewTag((PetscObject)C,&tag1);CHKERRQ(ierr); 801 ierr = PetscObjectGetNewTag((PetscObject)C,&tag2);CHKERRQ(ierr); 802 ierr = PetscObjectGetNewTag((PetscObject)C,&tag3);CHKERRQ(ierr); 803 804 /* Check if the col indices are sorted */ 805 for (i=0; i<ismax; i++) { 806 ierr = ISSorted(isrow[i],&sorted);CHKERRQ(ierr); 807 if (!sorted) SETERRQ(PETSC_ERR_ARG_WRONGSTATE,"ISrow is not sorted"); 808 ierr = ISSorted(iscol[i],&sorted);CHKERRQ(ierr); 809 /* if (!sorted) SETERRQ(PETSC_ERR_ARG_WRONGSTATE,"IScol is not sorted"); */ 810 } 811 812 len = (2*ismax+1)*(sizeof(int*)+ sizeof(int)) + (m+1)*sizeof(int); 813 ierr = PetscMalloc(len,&irow);CHKERRQ(ierr); 814 icol = irow + ismax; 815 nrow = (int*)(icol + ismax); 816 ncol = nrow + ismax; 817 rtable = ncol + ismax; 818 819 for (i=0; i<ismax; i++) { 820 ierr = ISGetIndices(isrow[i],&irow[i]);CHKERRQ(ierr); 821 ierr = ISGetIndices(iscol[i],&icol[i]);CHKERRQ(ierr); 822 ierr = ISGetLocalSize(isrow[i],&nrow[i]);CHKERRQ(ierr); 823 ierr = ISGetLocalSize(iscol[i],&ncol[i]);CHKERRQ(ierr); 824 } 825 826 /* Create hash table for the mapping :row -> proc*/ 827 for (i=0,j=0; i<size; i++) { 828 jmax = c->rowners[i+1]; 829 for (; j<jmax; j++) { 830 rtable[j] = i; 831 } 832 } 833 834 /* evaluate communication - mesg to who, length of mesg, and buffer space 835 required. Based on this, buffers are allocated, and data copied into them*/ 836 ierr = PetscMalloc(size*4*sizeof(int),&w1);CHKERRQ(ierr); /* mesg size */ 837 w2 = w1 + size; /* if w2[i] marked, then a message to proc i*/ 838 w3 = w2 + size; /* no of IS that needs to be sent to proc i */ 839 w4 = w3 + size; /* temp work space used in determining w1, w2, w3 */ 840 ierr = PetscMemzero(w1,size*3*sizeof(int));CHKERRQ(ierr); /* initialize work vector*/ 841 for (i=0; i<ismax; i++) { 842 ierr = PetscMemzero(w4,size*sizeof(int));CHKERRQ(ierr); /* initialize work vector*/ 843 jmax = nrow[i]; 844 irow_i = irow[i]; 845 for (j=0; j<jmax; j++) { 846 row = irow_i[j]; 847 proc = rtable[row]; 848 w4[proc]++; 849 } 850 for (j=0; j<size; j++) { 851 if (w4[j]) { w1[j] += w4[j]; w3[j]++;} 852 } 853 } 854 855 nrqs = 0; /* no of outgoing messages */ 856 msz = 0; /* total mesg length (for all procs) */ 857 w1[rank] = 0; /* no mesg sent to self */ 858 w3[rank] = 0; 859 for (i=0; i<size; i++) { 860 if (w1[i]) { w2[i] = 1; nrqs++;} /* there exists a message to proc i */ 861 } 862 ierr = PetscMalloc((nrqs+1)*sizeof(int),&pa);CHKERRQ(ierr); /*(proc -array)*/ 863 for (i=0,j=0; i<size; i++) { 864 if (w1[i]) { pa[j] = i; j++; } 865 } 866 867 /* Each message would have a header = 1 + 2*(no of IS) + data */ 868 for (i=0; i<nrqs; i++) { 869 j = pa[i]; 870 w1[j] += w2[j] + 2* w3[j]; 871 msz += w1[j]; 872 } 873 874 /* Determine the number of messages to expect, their lengths, from from-ids */ 875 ierr = PetscGatherNumberOfMessages(comm,w2,w1,&nrqr);CHKERRQ(ierr); 876 ierr = PetscGatherMessageLengths(comm,nrqs,nrqr,w1,&onodes1,&olengths1);CHKERRQ(ierr); 877 878 /* Now post the Irecvs corresponding to these messages */ 879 ierr = PetscPostIrecvInt(comm,tag0,nrqr,onodes1,olengths1,&rbuf1,&r_waits1);CHKERRQ(ierr); 880 881 ierr = PetscFree(onodes1);CHKERRQ(ierr); 882 ierr = PetscFree(olengths1);CHKERRQ(ierr); 883 884 /* Allocate Memory for outgoing messages */ 885 len = 2*size*sizeof(int*) + 2*msz*sizeof(int) + size*sizeof(int); 886 ierr = PetscMalloc(len,&sbuf1);CHKERRQ(ierr); 887 ptr = sbuf1 + size; /* Pointers to the data in outgoing buffers */ 888 ierr = PetscMemzero(sbuf1,2*size*sizeof(int*));CHKERRQ(ierr); 889 /* allocate memory for outgoing data + buf to receive the first reply */ 890 tmp = (int*)(ptr + size); 891 ctr = tmp + 2*msz; 892 893 { 894 int *iptr = tmp,ict = 0; 895 for (i=0; i<nrqs; i++) { 896 j = pa[i]; 897 iptr += ict; 898 sbuf1[j] = iptr; 899 ict = w1[j]; 900 } 901 } 902 903 /* Form the outgoing messages */ 904 /* Initialize the header space */ 905 for (i=0; i<nrqs; i++) { 906 j = pa[i]; 907 sbuf1[j][0] = 0; 908 ierr = PetscMemzero(sbuf1[j]+1,2*w3[j]*sizeof(int));CHKERRQ(ierr); 909 ptr[j] = sbuf1[j] + 2*w3[j] + 1; 910 } 911 912 /* Parse the isrow and copy data into outbuf */ 913 for (i=0; i<ismax; i++) { 914 ierr = PetscMemzero(ctr,size*sizeof(int));CHKERRQ(ierr); 915 irow_i = irow[i]; 916 jmax = nrow[i]; 917 for (j=0; j<jmax; j++) { /* parse the indices of each IS */ 918 row = irow_i[j]; 919 proc = rtable[row]; 920 if (proc != rank) { /* copy to the outgoing buf*/ 921 ctr[proc]++; 922 *ptr[proc] = row; 923 ptr[proc]++; 924 } 925 } 926 /* Update the headers for the current IS */ 927 for (j=0; j<size; j++) { /* Can Optimise this loop too */ 928 if ((ctr_j = ctr[j])) { 929 sbuf1_j = sbuf1[j]; 930 k = ++sbuf1_j[0]; 931 sbuf1_j[2*k] = ctr_j; 932 sbuf1_j[2*k-1] = i; 933 } 934 } 935 } 936 937 /* Now post the sends */ 938 ierr = PetscMalloc((nrqs+1)*sizeof(MPI_Request),&s_waits1);CHKERRQ(ierr); 939 for (i=0; i<nrqs; ++i) { 940 j = pa[i]; 941 ierr = MPI_Isend(sbuf1[j],w1[j],MPI_INT,j,tag0,comm,s_waits1+i);CHKERRQ(ierr); 942 } 943 944 /* Post Receives to capture the buffer size */ 945 ierr = PetscMalloc((nrqs+1)*sizeof(MPI_Request),&r_waits2);CHKERRQ(ierr); 946 ierr = PetscMalloc((nrqs+1)*sizeof(int*),&rbuf2);CHKERRQ(ierr); 947 rbuf2[0] = tmp + msz; 948 for (i=1; i<nrqs; ++i) { 949 rbuf2[i] = rbuf2[i-1]+w1[pa[i-1]]; 950 } 951 for (i=0; i<nrqs; ++i) { 952 j = pa[i]; 953 ierr = MPI_Irecv(rbuf2[i],w1[j],MPI_INT,j,tag1,comm,r_waits2+i);CHKERRQ(ierr); 954 } 955 956 /* Send to other procs the buf size they should allocate */ 957 958 959 /* Receive messages*/ 960 ierr = PetscMalloc((nrqr+1)*sizeof(MPI_Request),&s_waits2);CHKERRQ(ierr); 961 ierr = PetscMalloc((nrqr+1)*sizeof(MPI_Status),&r_status1);CHKERRQ(ierr); 962 len = 2*nrqr*sizeof(int) + (nrqr+1)*sizeof(int*); 963 ierr = PetscMalloc(len,&sbuf2);CHKERRQ(ierr); 964 req_size = (int*)(sbuf2 + nrqr); 965 req_source = req_size + nrqr; 966 967 { 968 Mat_SeqAIJ *sA = (Mat_SeqAIJ*)c->A->data,*sB = (Mat_SeqAIJ*)c->B->data; 969 int *sAi = sA->i,*sBi = sB->i,id,rstart = c->rstart; 970 int *sbuf2_i; 971 972 for (i=0; i<nrqr; ++i) { 973 ierr = MPI_Waitany(nrqr,r_waits1,&idex,r_status1+i);CHKERRQ(ierr); 974 req_size[idex] = 0; 975 rbuf1_i = rbuf1[idex]; 976 start = 2*rbuf1_i[0] + 1; 977 ierr = MPI_Get_count(r_status1+i,MPI_INT,&end);CHKERRQ(ierr); 978 ierr = PetscMalloc((end+1)*sizeof(int),&sbuf2[idex]);CHKERRQ(ierr); 979 sbuf2_i = sbuf2[idex]; 980 for (j=start; j<end; j++) { 981 id = rbuf1_i[j] - rstart; 982 ncols = sAi[id+1] - sAi[id] + sBi[id+1] - sBi[id]; 983 sbuf2_i[j] = ncols; 984 req_size[idex] += ncols; 985 } 986 req_source[idex] = r_status1[i].MPI_SOURCE; 987 /* form the header */ 988 sbuf2_i[0] = req_size[idex]; 989 for (j=1; j<start; j++) { sbuf2_i[j] = rbuf1_i[j]; } 990 ierr = MPI_Isend(sbuf2_i,end,MPI_INT,req_source[idex],tag1,comm,s_waits2+i);CHKERRQ(ierr); 991 } 992 } 993 ierr = PetscFree(r_status1);CHKERRQ(ierr); 994 ierr = PetscFree(r_waits1);CHKERRQ(ierr); 995 996 /* recv buffer sizes */ 997 /* Receive messages*/ 998 999 ierr = PetscMalloc((nrqs+1)*sizeof(int*),&rbuf3);CHKERRQ(ierr); 1000 ierr = PetscMalloc((nrqs+1)*sizeof(PetscScalar*),&rbuf4);CHKERRQ(ierr); 1001 ierr = PetscMalloc((nrqs+1)*sizeof(MPI_Request),&r_waits3);CHKERRQ(ierr); 1002 ierr = PetscMalloc((nrqs+1)*sizeof(MPI_Request),&r_waits4);CHKERRQ(ierr); 1003 ierr = PetscMalloc((nrqs+1)*sizeof(MPI_Status),&r_status2);CHKERRQ(ierr); 1004 1005 for (i=0; i<nrqs; ++i) { 1006 ierr = MPI_Waitany(nrqs,r_waits2,&idex,r_status2+i);CHKERRQ(ierr); 1007 ierr = PetscMalloc((rbuf2[idex][0]+1)*sizeof(int),&rbuf3[idex]);CHKERRQ(ierr); 1008 ierr = PetscMalloc((rbuf2[idex][0]+1)*sizeof(PetscScalar),&rbuf4[idex]);CHKERRQ(ierr); 1009 ierr = MPI_Irecv(rbuf3[idex],rbuf2[idex][0],MPI_INT,r_status2[i].MPI_SOURCE,tag2,comm,r_waits3+idex);CHKERRQ(ierr); 1010 ierr = MPI_Irecv(rbuf4[idex],rbuf2[idex][0],MPIU_SCALAR,r_status2[i].MPI_SOURCE,tag3,comm,r_waits4+idex);CHKERRQ(ierr); 1011 } 1012 ierr = PetscFree(r_status2);CHKERRQ(ierr); 1013 ierr = PetscFree(r_waits2);CHKERRQ(ierr); 1014 1015 /* Wait on sends1 and sends2 */ 1016 ierr = PetscMalloc((nrqs+1)*sizeof(MPI_Status),&s_status1);CHKERRQ(ierr); 1017 ierr = PetscMalloc((nrqr+1)*sizeof(MPI_Status),&s_status2);CHKERRQ(ierr); 1018 1019 ierr = MPI_Waitall(nrqs,s_waits1,s_status1);CHKERRQ(ierr); 1020 ierr = MPI_Waitall(nrqr,s_waits2,s_status2);CHKERRQ(ierr); 1021 ierr = PetscFree(s_status1);CHKERRQ(ierr); 1022 ierr = PetscFree(s_status2);CHKERRQ(ierr); 1023 ierr = PetscFree(s_waits1);CHKERRQ(ierr); 1024 ierr = PetscFree(s_waits2);CHKERRQ(ierr); 1025 1026 /* Now allocate buffers for a->j, and send them off */ 1027 ierr = PetscMalloc((nrqr+1)*sizeof(int*),&sbuf_aj);CHKERRQ(ierr); 1028 for (i=0,j=0; i<nrqr; i++) j += req_size[i]; 1029 ierr = PetscMalloc((j+1)*sizeof(int),&sbuf_aj[0]);CHKERRQ(ierr); 1030 for (i=1; i<nrqr; i++) sbuf_aj[i] = sbuf_aj[i-1] + req_size[i-1]; 1031 1032 ierr = PetscMalloc((nrqr+1)*sizeof(MPI_Request),&s_waits3);CHKERRQ(ierr); 1033 { 1034 int nzA,nzB,*a_i = a->i,*b_i = b->i,imark; 1035 int *cworkA,*cworkB,cstart = c->cstart,rstart = c->rstart,*bmap = c->garray; 1036 int *a_j = a->j,*b_j = b->j,ctmp; 1037 1038 for (i=0; i<nrqr; i++) { 1039 rbuf1_i = rbuf1[i]; 1040 sbuf_aj_i = sbuf_aj[i]; 1041 ct1 = 2*rbuf1_i[0] + 1; 1042 ct2 = 0; 1043 for (j=1,max1=rbuf1_i[0]; j<=max1; j++) { 1044 kmax = rbuf1[i][2*j]; 1045 for (k=0; k<kmax; k++,ct1++) { 1046 row = rbuf1_i[ct1] - rstart; 1047 nzA = a_i[row+1] - a_i[row]; nzB = b_i[row+1] - b_i[row]; 1048 ncols = nzA + nzB; 1049 cworkA = a_j + a_i[row]; cworkB = b_j + b_i[row]; 1050 1051 /* load the column indices for this row into cols*/ 1052 cols = sbuf_aj_i + ct2; 1053 1054 for (l=0; l<nzB; l++) { 1055 if ((ctmp = bmap[cworkB[l]]) < cstart) cols[l] = ctmp; 1056 else break; 1057 } 1058 imark = l; 1059 for (l=0; l<nzA; l++) cols[imark+l] = cstart + cworkA[l]; 1060 for (l=imark; l<nzB; l++) cols[nzA+l] = bmap[cworkB[l]]; 1061 1062 ct2 += ncols; 1063 } 1064 } 1065 ierr = MPI_Isend(sbuf_aj_i,req_size[i],MPI_INT,req_source[i],tag2,comm,s_waits3+i);CHKERRQ(ierr); 1066 } 1067 } 1068 ierr = PetscMalloc((nrqs+1)*sizeof(MPI_Status),&r_status3);CHKERRQ(ierr); 1069 ierr = PetscMalloc((nrqr+1)*sizeof(MPI_Status),&s_status3);CHKERRQ(ierr); 1070 1071 /* Allocate buffers for a->a, and send them off */ 1072 ierr = PetscMalloc((nrqr+1)*sizeof(PetscScalar*),&sbuf_aa);CHKERRQ(ierr); 1073 for (i=0,j=0; i<nrqr; i++) j += req_size[i]; 1074 ierr = PetscMalloc((j+1)*sizeof(PetscScalar),&sbuf_aa[0]);CHKERRQ(ierr); 1075 for (i=1; i<nrqr; i++) sbuf_aa[i] = sbuf_aa[i-1] + req_size[i-1]; 1076 1077 ierr = PetscMalloc((nrqr+1)*sizeof(MPI_Request),&s_waits4);CHKERRQ(ierr); 1078 { 1079 int nzA,nzB,*a_i = a->i,*b_i = b->i, *cworkB,imark; 1080 int cstart = c->cstart,rstart = c->rstart,*bmap = c->garray; 1081 int *b_j = b->j; 1082 PetscScalar *vworkA,*vworkB,*a_a = a->a,*b_a = b->a; 1083 1084 for (i=0; i<nrqr; i++) { 1085 rbuf1_i = rbuf1[i]; 1086 sbuf_aa_i = sbuf_aa[i]; 1087 ct1 = 2*rbuf1_i[0]+1; 1088 ct2 = 0; 1089 for (j=1,max1=rbuf1_i[0]; j<=max1; j++) { 1090 kmax = rbuf1_i[2*j]; 1091 for (k=0; k<kmax; k++,ct1++) { 1092 row = rbuf1_i[ct1] - rstart; 1093 nzA = a_i[row+1] - a_i[row]; nzB = b_i[row+1] - b_i[row]; 1094 ncols = nzA + nzB; 1095 cworkB = b_j + b_i[row]; 1096 vworkA = a_a + a_i[row]; 1097 vworkB = b_a + b_i[row]; 1098 1099 /* load the column values for this row into vals*/ 1100 vals = sbuf_aa_i+ct2; 1101 1102 for (l=0; l<nzB; l++) { 1103 if ((bmap[cworkB[l]]) < cstart) vals[l] = vworkB[l]; 1104 else break; 1105 } 1106 imark = l; 1107 for (l=0; l<nzA; l++) vals[imark+l] = vworkA[l]; 1108 for (l=imark; l<nzB; l++) vals[nzA+l] = vworkB[l]; 1109 1110 ct2 += ncols; 1111 } 1112 } 1113 ierr = MPI_Isend(sbuf_aa_i,req_size[i],MPIU_SCALAR,req_source[i],tag3,comm,s_waits4+i);CHKERRQ(ierr); 1114 } 1115 } 1116 ierr = PetscMalloc((nrqs+1)*sizeof(MPI_Status),&r_status4);CHKERRQ(ierr); 1117 ierr = PetscMalloc((nrqr+1)*sizeof(MPI_Status),&s_status4);CHKERRQ(ierr); 1118 ierr = PetscFree(rbuf1);CHKERRQ(ierr); 1119 1120 /* Form the matrix */ 1121 /* create col map */ 1122 { 1123 int *icol_i; 1124 1125 len = (1+ismax)*sizeof(int*)+ ismax*C->N*sizeof(int); 1126 ierr = PetscMalloc(len,&cmap);CHKERRQ(ierr); 1127 cmap[0] = (int*)(cmap + ismax); 1128 ierr = PetscMemzero(cmap[0],(1+ismax*C->N)*sizeof(int));CHKERRQ(ierr); 1129 for (i=1; i<ismax; i++) { cmap[i] = cmap[i-1] + C->N; } 1130 for (i=0; i<ismax; i++) { 1131 jmax = ncol[i]; 1132 icol_i = icol[i]; 1133 cmap_i = cmap[i]; 1134 for (j=0; j<jmax; j++) { 1135 cmap_i[icol_i[j]] = j+1; 1136 } 1137 } 1138 } 1139 1140 /* Create lens which is required for MatCreate... */ 1141 for (i=0,j=0; i<ismax; i++) { j += nrow[i]; } 1142 len = (1+ismax)*sizeof(int*)+ j*sizeof(int); 1143 ierr = PetscMalloc(len,&lens);CHKERRQ(ierr); 1144 lens[0] = (int*)(lens + ismax); 1145 ierr = PetscMemzero(lens[0],j*sizeof(int));CHKERRQ(ierr); 1146 for (i=1; i<ismax; i++) { lens[i] = lens[i-1] + nrow[i-1]; } 1147 1148 /* Update lens from local data */ 1149 for (i=0; i<ismax; i++) { 1150 jmax = nrow[i]; 1151 cmap_i = cmap[i]; 1152 irow_i = irow[i]; 1153 lens_i = lens[i]; 1154 for (j=0; j<jmax; j++) { 1155 row = irow_i[j]; 1156 proc = rtable[row]; 1157 if (proc == rank) { 1158 ierr = MatGetRow_MPIAIJ(C,row,&ncols,&cols,0);CHKERRQ(ierr); 1159 for (k=0; k<ncols; k++) { 1160 if (cmap_i[cols[k]]) { lens_i[j]++;} 1161 } 1162 ierr = MatRestoreRow_MPIAIJ(C,row,&ncols,&cols,0);CHKERRQ(ierr); 1163 } 1164 } 1165 } 1166 1167 /* Create row map*/ 1168 len = (1+ismax)*sizeof(int*)+ ismax*C->M*sizeof(int); 1169 ierr = PetscMalloc(len,&rmap);CHKERRQ(ierr); 1170 rmap[0] = (int*)(rmap + ismax); 1171 ierr = PetscMemzero(rmap[0],ismax*C->M*sizeof(int));CHKERRQ(ierr); 1172 for (i=1; i<ismax; i++) { rmap[i] = rmap[i-1] + C->M;} 1173 for (i=0; i<ismax; i++) { 1174 rmap_i = rmap[i]; 1175 irow_i = irow[i]; 1176 jmax = nrow[i]; 1177 for (j=0; j<jmax; j++) { 1178 rmap_i[irow_i[j]] = j; 1179 } 1180 } 1181 1182 /* Update lens from offproc data */ 1183 { 1184 int *rbuf2_i,*rbuf3_i,*sbuf1_i; 1185 1186 for (tmp2=0; tmp2<nrqs; tmp2++) { 1187 ierr = MPI_Waitany(nrqs,r_waits3,&i,r_status3+tmp2);CHKERRQ(ierr); 1188 idex = pa[i]; 1189 sbuf1_i = sbuf1[idex]; 1190 jmax = sbuf1_i[0]; 1191 ct1 = 2*jmax+1; 1192 ct2 = 0; 1193 rbuf2_i = rbuf2[i]; 1194 rbuf3_i = rbuf3[i]; 1195 for (j=1; j<=jmax; j++) { 1196 is_no = sbuf1_i[2*j-1]; 1197 max1 = sbuf1_i[2*j]; 1198 lens_i = lens[is_no]; 1199 cmap_i = cmap[is_no]; 1200 rmap_i = rmap[is_no]; 1201 for (k=0; k<max1; k++,ct1++) { 1202 row = rmap_i[sbuf1_i[ct1]]; /* the val in the new matrix to be */ 1203 max2 = rbuf2_i[ct1]; 1204 for (l=0; l<max2; l++,ct2++) { 1205 if (cmap_i[rbuf3_i[ct2]]) { 1206 lens_i[row]++; 1207 } 1208 } 1209 } 1210 } 1211 } 1212 } 1213 ierr = PetscFree(r_status3);CHKERRQ(ierr); 1214 ierr = PetscFree(r_waits3);CHKERRQ(ierr); 1215 ierr = MPI_Waitall(nrqr,s_waits3,s_status3);CHKERRQ(ierr); 1216 ierr = PetscFree(s_status3);CHKERRQ(ierr); 1217 ierr = PetscFree(s_waits3);CHKERRQ(ierr); 1218 1219 /* Create the submatrices */ 1220 if (scall == MAT_REUSE_MATRIX) { 1221 PetscTruth flag; 1222 1223 /* 1224 Assumes new rows are same length as the old rows,hence bug! 1225 */ 1226 for (i=0; i<ismax; i++) { 1227 mat = (Mat_SeqAIJ *)(submats[i]->data); 1228 if ((submats[i]->m != nrow[i]) || (submats[i]->n != ncol[i])) { 1229 SETERRQ(PETSC_ERR_ARG_SIZ,"Cannot reuse matrix. wrong size"); 1230 } 1231 ierr = PetscMemcmp(mat->ilen,lens[i],submats[i]->m*sizeof(int),&flag);CHKERRQ(ierr); 1232 if (flag == PETSC_FALSE) { 1233 SETERRQ(PETSC_ERR_ARG_SIZ,"Cannot reuse matrix. wrong no of nonzeros"); 1234 } 1235 /* Initial matrix as if empty */ 1236 ierr = PetscMemzero(mat->ilen,submats[i]->m*sizeof(int));CHKERRQ(ierr); 1237 submats[i]->factor = C->factor; 1238 } 1239 } else { 1240 for (i=0; i<ismax; i++) { 1241 ierr = MatCreate(PETSC_COMM_SELF,nrow[i],ncol[i],PETSC_DETERMINE,PETSC_DETERMINE,submats+i);CHKERRQ(ierr); 1242 ierr = MatSetType(submats[i],A->type_name);CHKERRQ(ierr); 1243 ierr = MatSeqAIJSetPreallocation(submats[i],0,lens[i]);CHKERRQ(ierr); 1244 } 1245 } 1246 1247 /* Assemble the matrices */ 1248 /* First assemble the local rows */ 1249 { 1250 int ilen_row,*imat_ilen,*imat_j,*imat_i,old_row; 1251 PetscScalar *imat_a; 1252 1253 for (i=0; i<ismax; i++) { 1254 mat = (Mat_SeqAIJ*)submats[i]->data; 1255 imat_ilen = mat->ilen; 1256 imat_j = mat->j; 1257 imat_i = mat->i; 1258 imat_a = mat->a; 1259 cmap_i = cmap[i]; 1260 rmap_i = rmap[i]; 1261 irow_i = irow[i]; 1262 jmax = nrow[i]; 1263 for (j=0; j<jmax; j++) { 1264 row = irow_i[j]; 1265 proc = rtable[row]; 1266 if (proc == rank) { 1267 old_row = row; 1268 row = rmap_i[row]; 1269 ilen_row = imat_ilen[row]; 1270 ierr = MatGetRow_MPIAIJ(C,old_row,&ncols,&cols,&vals);CHKERRQ(ierr); 1271 mat_i = imat_i[row] ; 1272 mat_a = imat_a + mat_i; 1273 mat_j = imat_j + mat_i; 1274 for (k=0; k<ncols; k++) { 1275 if ((tcol = cmap_i[cols[k]])) { 1276 *mat_j++ = tcol - 1; 1277 *mat_a++ = vals[k]; 1278 ilen_row++; 1279 } 1280 } 1281 ierr = MatRestoreRow_MPIAIJ(C,old_row,&ncols,&cols,&vals);CHKERRQ(ierr); 1282 imat_ilen[row] = ilen_row; 1283 } 1284 } 1285 } 1286 } 1287 1288 /* Now assemble the off proc rows*/ 1289 { 1290 int *sbuf1_i,*rbuf2_i,*rbuf3_i,*imat_ilen,ilen; 1291 int *imat_j,*imat_i; 1292 PetscScalar *imat_a,*rbuf4_i; 1293 1294 for (tmp2=0; tmp2<nrqs; tmp2++) { 1295 ierr = MPI_Waitany(nrqs,r_waits4,&i,r_status4+tmp2);CHKERRQ(ierr); 1296 idex = pa[i]; 1297 sbuf1_i = sbuf1[idex]; 1298 jmax = sbuf1_i[0]; 1299 ct1 = 2*jmax + 1; 1300 ct2 = 0; 1301 rbuf2_i = rbuf2[i]; 1302 rbuf3_i = rbuf3[i]; 1303 rbuf4_i = rbuf4[i]; 1304 for (j=1; j<=jmax; j++) { 1305 is_no = sbuf1_i[2*j-1]; 1306 rmap_i = rmap[is_no]; 1307 cmap_i = cmap[is_no]; 1308 mat = (Mat_SeqAIJ*)submats[is_no]->data; 1309 imat_ilen = mat->ilen; 1310 imat_j = mat->j; 1311 imat_i = mat->i; 1312 imat_a = mat->a; 1313 max1 = sbuf1_i[2*j]; 1314 for (k=0; k<max1; k++,ct1++) { 1315 row = sbuf1_i[ct1]; 1316 row = rmap_i[row]; 1317 ilen = imat_ilen[row]; 1318 mat_i = imat_i[row] ; 1319 mat_a = imat_a + mat_i; 1320 mat_j = imat_j + mat_i; 1321 max2 = rbuf2_i[ct1]; 1322 for (l=0; l<max2; l++,ct2++) { 1323 if ((tcol = cmap_i[rbuf3_i[ct2]])) { 1324 *mat_j++ = tcol - 1; 1325 *mat_a++ = rbuf4_i[ct2]; 1326 ilen++; 1327 } 1328 } 1329 imat_ilen[row] = ilen; 1330 } 1331 } 1332 } 1333 } 1334 ierr = PetscFree(r_status4);CHKERRQ(ierr); 1335 ierr = PetscFree(r_waits4);CHKERRQ(ierr); 1336 ierr = MPI_Waitall(nrqr,s_waits4,s_status4);CHKERRQ(ierr); 1337 ierr = PetscFree(s_waits4);CHKERRQ(ierr); 1338 ierr = PetscFree(s_status4);CHKERRQ(ierr); 1339 1340 /* Restore the indices */ 1341 for (i=0; i<ismax; i++) { 1342 ierr = ISRestoreIndices(isrow[i],irow+i);CHKERRQ(ierr); 1343 ierr = ISRestoreIndices(iscol[i],icol+i);CHKERRQ(ierr); 1344 } 1345 1346 /* Destroy allocated memory */ 1347 ierr = PetscFree(irow);CHKERRQ(ierr); 1348 ierr = PetscFree(w1);CHKERRQ(ierr); 1349 ierr = PetscFree(pa);CHKERRQ(ierr); 1350 1351 ierr = PetscFree(sbuf1);CHKERRQ(ierr); 1352 ierr = PetscFree(rbuf2);CHKERRQ(ierr); 1353 for (i=0; i<nrqr; ++i) { 1354 ierr = PetscFree(sbuf2[i]);CHKERRQ(ierr); 1355 } 1356 for (i=0; i<nrqs; ++i) { 1357 ierr = PetscFree(rbuf3[i]);CHKERRQ(ierr); 1358 ierr = PetscFree(rbuf4[i]);CHKERRQ(ierr); 1359 } 1360 1361 ierr = PetscFree(sbuf2);CHKERRQ(ierr); 1362 ierr = PetscFree(rbuf3);CHKERRQ(ierr); 1363 ierr = PetscFree(rbuf4);CHKERRQ(ierr); 1364 ierr = PetscFree(sbuf_aj[0]);CHKERRQ(ierr); 1365 ierr = PetscFree(sbuf_aj);CHKERRQ(ierr); 1366 ierr = PetscFree(sbuf_aa[0]);CHKERRQ(ierr); 1367 ierr = PetscFree(sbuf_aa);CHKERRQ(ierr); 1368 1369 ierr = PetscFree(cmap);CHKERRQ(ierr); 1370 ierr = PetscFree(rmap);CHKERRQ(ierr); 1371 ierr = PetscFree(lens);CHKERRQ(ierr); 1372 1373 for (i=0; i<ismax; i++) { 1374 ierr = MatAssemblyBegin(submats[i],MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1375 ierr = MatAssemblyEnd(submats[i],MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1376 } 1377 PetscFunctionReturn(0); 1378 } 1379 1380