1 2 /* 3 Defines projective product routines where A is a MPIAIJ matrix 4 C = P^T * A * P 5 */ 6 7 #include <../src/mat/impls/aij/seq/aij.h> /*I "petscmat.h" I*/ 8 #include <../src/mat/utils/freespace.h> 9 #include <../src/mat/impls/aij/mpi/mpiaij.h> 10 #include <petscbt.h> 11 #include <petsctime.h> 12 13 /* #define PTAP_PROFILE */ 14 15 extern PetscErrorCode MatDestroy_MPIAIJ(Mat); 16 PetscErrorCode MatDestroy_MPIAIJ_PtAP(Mat A) 17 { 18 PetscErrorCode ierr; 19 Mat_MPIAIJ *a=(Mat_MPIAIJ*)A->data; 20 Mat_PtAPMPI *ptap=a->ptap; 21 22 PetscFunctionBegin; 23 if (ptap) { 24 Mat_Merge_SeqsToMPI *merge=ptap->merge; 25 ierr = PetscFree2(ptap->startsj_s,ptap->startsj_r);CHKERRQ(ierr); 26 ierr = PetscFree(ptap->bufa);CHKERRQ(ierr); 27 ierr = MatDestroy(&ptap->P_loc);CHKERRQ(ierr); 28 ierr = MatDestroy(&ptap->P_oth);CHKERRQ(ierr); 29 ierr = MatDestroy(&ptap->A_loc);CHKERRQ(ierr); /* used by MatTransposeMatMult() */ 30 ierr = MatDestroy(&ptap->Rd);CHKERRQ(ierr); 31 ierr = MatDestroy(&ptap->Ro);CHKERRQ(ierr); 32 if (ptap->AP_loc) { /* used by alg_rap */ 33 Mat_SeqAIJ *ap = (Mat_SeqAIJ*)(ptap->AP_loc)->data; 34 ierr = PetscFree(ap->i);CHKERRQ(ierr); 35 ierr = PetscFree2(ap->j,ap->a);CHKERRQ(ierr); 36 ierr = MatDestroy(&ptap->AP_loc);CHKERRQ(ierr); 37 } else { /* used by alg_ptap */ 38 ierr = PetscFree(ptap->api);CHKERRQ(ierr); 39 ierr = PetscFree(ptap->apj);CHKERRQ(ierr); 40 } 41 ierr = MatDestroy(&ptap->C_loc);CHKERRQ(ierr); 42 ierr = MatDestroy(&ptap->C_oth);CHKERRQ(ierr); 43 if (ptap->apa) {ierr = PetscFree(ptap->apa);CHKERRQ(ierr);} 44 45 if (merge) { /* used by alg_ptap */ 46 ierr = PetscFree(merge->id_r);CHKERRQ(ierr); 47 ierr = PetscFree(merge->len_s);CHKERRQ(ierr); 48 ierr = PetscFree(merge->len_r);CHKERRQ(ierr); 49 ierr = PetscFree(merge->bi);CHKERRQ(ierr); 50 ierr = PetscFree(merge->bj);CHKERRQ(ierr); 51 ierr = PetscFree(merge->buf_ri[0]);CHKERRQ(ierr); 52 ierr = PetscFree(merge->buf_ri);CHKERRQ(ierr); 53 ierr = PetscFree(merge->buf_rj[0]);CHKERRQ(ierr); 54 ierr = PetscFree(merge->buf_rj);CHKERRQ(ierr); 55 ierr = PetscFree(merge->coi);CHKERRQ(ierr); 56 ierr = PetscFree(merge->coj);CHKERRQ(ierr); 57 ierr = PetscFree(merge->owners_co);CHKERRQ(ierr); 58 ierr = PetscLayoutDestroy(&merge->rowmap);CHKERRQ(ierr); 59 ierr = PetscFree(ptap->merge);CHKERRQ(ierr); 60 } 61 62 ierr = ptap->destroy(A);CHKERRQ(ierr); 63 ierr = PetscFree(ptap);CHKERRQ(ierr); 64 } 65 PetscFunctionReturn(0); 66 } 67 68 PetscErrorCode MatDuplicate_MPIAIJ_MatPtAP(Mat A, MatDuplicateOption op, Mat *M) 69 { 70 PetscErrorCode ierr; 71 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 72 Mat_PtAPMPI *ptap = a->ptap; 73 74 PetscFunctionBegin; 75 ierr = (*ptap->duplicate)(A,op,M);CHKERRQ(ierr); 76 (*M)->ops->destroy = ptap->destroy; 77 (*M)->ops->duplicate = ptap->duplicate; 78 PetscFunctionReturn(0); 79 } 80 81 PETSC_INTERN PetscErrorCode MatPtAP_MPIAIJ_MPIAIJ(Mat A,Mat P,MatReuse scall,PetscReal fill,Mat *C) 82 { 83 PetscErrorCode ierr; 84 PetscBool rap=PETSC_TRUE; /* do R=P^T locally, then C=R*A*P */ 85 MPI_Comm comm; 86 87 PetscFunctionBegin; 88 /* check if matrix local sizes are compatible */ 89 ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr); 90 if (A->rmap->rstart != P->rmap->rstart || A->rmap->rend != P->rmap->rend) SETERRQ4(comm,PETSC_ERR_ARG_SIZ,"Matrix local dimensions are incompatible, Arow (%D, %D) != Prow (%D,%D)",A->rmap->rstart,A->rmap->rend,P->rmap->rstart,P->rmap->rend); 91 if (A->cmap->rstart != P->rmap->rstart || A->cmap->rend != P->rmap->rend) SETERRQ4(comm,PETSC_ERR_ARG_SIZ,"Matrix local dimensions are incompatible, Acol (%D, %D) != Prow (%D,%D)",A->cmap->rstart,A->cmap->rend,P->rmap->rstart,P->rmap->rend); 92 93 ierr = PetscOptionsGetBool(NULL,NULL,"-matrap",&rap,NULL);CHKERRQ(ierr); 94 if (scall == MAT_INITIAL_MATRIX) { 95 ierr = PetscLogEventBegin(MAT_PtAPSymbolic,A,P,0,0);CHKERRQ(ierr); 96 if (rap) { /* do R=P^T locally, then C=R*A*P */ 97 ierr = MatPtAPSymbolic_MPIAIJ_MPIAIJ(A,P,fill,C);CHKERRQ(ierr); 98 } else { /* do P^T*A*P */ 99 ierr = MatPtAPSymbolic_MPIAIJ_MPIAIJ_ptap(A,P,fill,C);CHKERRQ(ierr); 100 } 101 ierr = PetscLogEventEnd(MAT_PtAPSymbolic,A,P,0,0);CHKERRQ(ierr); 102 } 103 ierr = PetscLogEventBegin(MAT_PtAPNumeric,A,P,0,0);CHKERRQ(ierr); 104 ierr = (*(*C)->ops->ptapnumeric)(A,P,*C);CHKERRQ(ierr); 105 ierr = PetscLogEventEnd(MAT_PtAPNumeric,A,P,0,0);CHKERRQ(ierr); 106 PetscFunctionReturn(0); 107 } 108 109 #if defined(PETSC_HAVE_HYPRE) 110 PETSC_INTERN PetscErrorCode MatPtAPSymbolic_AIJ_AIJ_wHYPRE(Mat,Mat,PetscReal,Mat*); 111 #endif 112 113 PetscErrorCode MatPtAPSymbolic_MPIAIJ_MPIAIJ_scalable(Mat A,Mat P,PetscReal fill,Mat *C) 114 { 115 PetscErrorCode ierr; 116 Mat_PtAPMPI *ptap; 117 Mat_MPIAIJ *a=(Mat_MPIAIJ*)A->data,*p=(Mat_MPIAIJ*)P->data,*c; 118 MPI_Comm comm; 119 PetscMPIInt size,rank; 120 Mat Cmpi; 121 PetscFreeSpaceList free_space=NULL,current_space=NULL; 122 PetscInt am=A->rmap->n,pm=P->rmap->n,pN=P->cmap->N,pn=P->cmap->n; 123 PetscInt *lnk,i,k,pnz,row,nsend; 124 PetscBT lnkbt; 125 PetscMPIInt tagi,tagj,*len_si,*len_s,*len_ri,icompleted=0,nrecv; 126 PetscInt **buf_rj,**buf_ri,**buf_ri_k; 127 PetscInt len,proc,*dnz,*onz,*owners,nzi,nspacedouble; 128 PetscInt nrows,*buf_s,*buf_si,*buf_si_i,**nextrow,**nextci; 129 MPI_Request *swaits,*rwaits; 130 MPI_Status *sstatus,rstatus; 131 PetscLayout rowmap; 132 PetscInt *owners_co,*coi,*coj; /* i and j array of (p->B)^T*A*P - used in the communication */ 133 PetscMPIInt *len_r,*id_r; /* array of length of comm->size, store send/recv matrix values */ 134 PetscInt *api,*apj,*Jptr,apnz,*prmap=p->garray,con,j,ap_rmax=0,Crmax,*aj,*ai,*pi; 135 Mat_SeqAIJ *p_loc,*p_oth,*ad=(Mat_SeqAIJ*)(a->A)->data,*ao=(Mat_SeqAIJ*)(a->B)->data,*c_loc,*c_oth; 136 PetscScalar *apv; 137 PetscTable ta; 138 #if defined(PETSC_HAVE_HYPRE) 139 const char *algTypes[3] = {"scalable","nonscalable","hypre"}; 140 PetscInt nalg = 3; 141 #else 142 const char *algTypes[2] = {"scalable","nonscalable"}; 143 PetscInt nalg = 2; 144 #endif 145 PetscInt alg = 1; /* set default algorithm */ 146 #if defined(PETSC_USE_INFO) 147 PetscReal apfill; 148 #endif 149 #if defined(PTAP_PROFILE) 150 PetscLogDouble t0,t1,t11,t12,t2,t3,t4; 151 #endif 152 153 PetscFunctionBegin; 154 ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr); 155 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 156 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 157 158 /* create symbolic parallel matrix Cmpi */ 159 ierr = MatCreate(comm,&Cmpi);CHKERRQ(ierr); 160 ierr = MatSetType(Cmpi,MATMPIAIJ);CHKERRQ(ierr); 161 162 #if defined(PTAP_PROFILE) 163 ierr = PetscTime(&t0);CHKERRQ(ierr); 164 #endif 165 166 /* create struct Mat_PtAPMPI and attached it to C later */ 167 ierr = PetscNew(&ptap);CHKERRQ(ierr); 168 ptap->reuse = MAT_INITIAL_MATRIX; 169 170 /* get P_oth by taking rows of P (= non-zero cols of local A) from other processors */ 171 ierr = MatGetBrowsOfAoCols_MPIAIJ(A,P,MAT_INITIAL_MATRIX,&ptap->startsj_s,&ptap->startsj_r,&ptap->bufa,&ptap->P_oth);CHKERRQ(ierr); 172 /* get P_loc by taking all local rows of P */ 173 ierr = MatMPIAIJGetLocalMat(P,MAT_INITIAL_MATRIX,&ptap->P_loc);CHKERRQ(ierr); 174 175 /* (0) compute Rd = Pd^T, Ro = Po^T */ 176 /* --------------------------------- */ 177 ierr = MatTranspose_SeqAIJ(p->A,MAT_INITIAL_MATRIX,&ptap->Rd);CHKERRQ(ierr); 178 ierr = MatTranspose_SeqAIJ(p->B,MAT_INITIAL_MATRIX,&ptap->Ro);CHKERRQ(ierr); 179 #if defined(PTAP_PROFILE) 180 ierr = PetscTime(&t11);CHKERRQ(ierr); 181 #endif 182 183 /* (1) compute symbolic AP = A_loc*P = Ad*P_loc + Ao*P_oth (api,apj) */ 184 /* ----------------------------------------------------------------- */ 185 p_loc = (Mat_SeqAIJ*)(ptap->P_loc)->data; 186 p_oth = (Mat_SeqAIJ*)(ptap->P_oth)->data; 187 188 /* create and initialize a linked list */ 189 ierr = PetscTableCreate(pn,pN,&ta);CHKERRQ(ierr); /* for compute AP_loc and Cmpi */ 190 MatRowMergeMax_SeqAIJ(p_loc,ptap->P_loc->rmap->N,ta); 191 MatRowMergeMax_SeqAIJ(p_oth,ptap->P_oth->rmap->N,ta); 192 ierr = PetscTableGetCount(ta,&Crmax);CHKERRQ(ierr); /* Crmax = nnz(sum of Prows) */ 193 /* printf("[%d] est %d, Crmax %d; pN %d\n",rank,5*(p_loc->rmax+p_oth->rmax + (PetscInt)(1.e-2*pN)),Crmax,pN); */ 194 195 ierr = PetscLLCondensedCreate(Crmax,pN,&lnk,&lnkbt);CHKERRQ(ierr); 196 197 /* Initial FreeSpace size is fill*(nnz(A) + nnz(P)) */ 198 ierr = PetscFreeSpaceGet(PetscRealIntMultTruncate(fill,PetscIntSumTruncate(ad->i[am],PetscIntSumTruncate(ao->i[am],p_loc->i[pm]))),&free_space);CHKERRQ(ierr); 199 current_space = free_space; 200 nspacedouble = 0; 201 202 ierr = PetscMalloc1(am+1,&api);CHKERRQ(ierr); 203 api[0] = 0; 204 for (i=0; i<am; i++) { 205 /* diagonal portion: Ad[i,:]*P */ 206 ai = ad->i; pi = p_loc->i; 207 nzi = ai[i+1] - ai[i]; 208 aj = ad->j + ai[i]; 209 for (j=0; j<nzi; j++) { 210 row = aj[j]; 211 pnz = pi[row+1] - pi[row]; 212 Jptr = p_loc->j + pi[row]; 213 /* add non-zero cols of P into the sorted linked list lnk */ 214 ierr = PetscLLCondensedAddSorted(pnz,Jptr,lnk,lnkbt);CHKERRQ(ierr); 215 } 216 /* off-diagonal portion: Ao[i,:]*P */ 217 ai = ao->i; pi = p_oth->i; 218 nzi = ai[i+1] - ai[i]; 219 aj = ao->j + ai[i]; 220 for (j=0; j<nzi; j++) { 221 row = aj[j]; 222 pnz = pi[row+1] - pi[row]; 223 Jptr = p_oth->j + pi[row]; 224 ierr = PetscLLCondensedAddSorted(pnz,Jptr,lnk,lnkbt);CHKERRQ(ierr); 225 } 226 apnz = lnk[0]; 227 api[i+1] = api[i] + apnz; 228 if (ap_rmax < apnz) ap_rmax = apnz; 229 230 /* if free space is not available, double the total space in the list */ 231 if (current_space->local_remaining<apnz) { 232 ierr = PetscFreeSpaceGet(PetscIntSumTruncate(apnz,current_space->total_array_size),¤t_space);CHKERRQ(ierr); 233 nspacedouble++; 234 } 235 236 /* Copy data into free space, then initialize lnk */ 237 ierr = PetscLLCondensedClean(pN,apnz,current_space->array,lnk,lnkbt);CHKERRQ(ierr); 238 239 current_space->array += apnz; 240 current_space->local_used += apnz; 241 current_space->local_remaining -= apnz; 242 } 243 /* Allocate space for apj and apv, initialize apj, and */ 244 /* destroy list of free space and other temporary array(s) */ 245 ierr = PetscMalloc2(api[am],&apj,api[am],&apv);CHKERRQ(ierr); 246 ierr = PetscFreeSpaceContiguous(&free_space,apj);CHKERRQ(ierr); 247 ierr = PetscLLDestroy(lnk,lnkbt);CHKERRQ(ierr); 248 249 /* Create AP_loc for reuse */ 250 ierr = MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,am,pN,api,apj,apv,&ptap->AP_loc);CHKERRQ(ierr); 251 252 #if defined(PETSC_USE_INFO) 253 apfill = (PetscReal)api[am]/(ad->i[am]+ao->i[am]+p_loc->i[pm]+1); 254 ptap->AP_loc->info.mallocs = nspacedouble; 255 ptap->AP_loc->info.fill_ratio_given = fill; 256 ptap->AP_loc->info.fill_ratio_needed = apfill; 257 258 if (api[am]) { 259 ierr = PetscInfo3(ptap->AP_loc,"AP_loc reallocs %D; Fill ratio: given %g needed %g.\n",nspacedouble,(double)fill,(double)apfill);CHKERRQ(ierr); 260 ierr = PetscInfo1(ptap->AP_loc,"Use MatPtAP(A,B,MatReuse,%g,&C) for best AP_loc performance.;\n",(double)apfill);CHKERRQ(ierr); 261 } else { 262 ierr = PetscInfo(ptap->AP_loc,"AP_loc is empty \n");CHKERRQ(ierr); 263 } 264 #endif 265 266 #if defined(PTAP_PROFILE) 267 ierr = PetscTime(&t12);CHKERRQ(ierr); 268 #endif 269 270 /* (2-1) compute symbolic Co = Ro*AP_loc */ 271 /* ------------------------------------ */ 272 ierr = MatMatMultSymbolic_SeqAIJ_SeqAIJ(ptap->Ro,ptap->AP_loc,fill,&ptap->C_oth);CHKERRQ(ierr); 273 #if defined(PTAP_PROFILE) 274 ierr = PetscTime(&t1);CHKERRQ(ierr); 275 #endif 276 277 /* (3) send coj of C_oth to other processors */ 278 /* ------------------------------------------ */ 279 /* determine row ownership */ 280 ierr = PetscLayoutCreate(comm,&rowmap);CHKERRQ(ierr); 281 rowmap->n = pn; 282 rowmap->bs = 1; 283 ierr = PetscLayoutSetUp(rowmap);CHKERRQ(ierr); 284 owners = rowmap->range; 285 286 /* determine the number of messages to send, their lengths */ 287 ierr = PetscMalloc4(size,&len_s,size,&len_si,size,&sstatus,size+2,&owners_co);CHKERRQ(ierr); 288 ierr = PetscMemzero(len_s,size*sizeof(PetscMPIInt));CHKERRQ(ierr); 289 ierr = PetscMemzero(len_si,size*sizeof(PetscMPIInt));CHKERRQ(ierr); 290 291 c_oth = (Mat_SeqAIJ*)ptap->C_oth->data; 292 coi = c_oth->i; coj = c_oth->j; 293 con = ptap->C_oth->rmap->n; 294 proc = 0; 295 for (i=0; i<con; i++) { 296 while (prmap[i] >= owners[proc+1]) proc++; 297 len_si[proc]++; /* num of rows in Co(=Pt*AP) to be sent to [proc] */ 298 len_s[proc] += coi[i+1] - coi[i]; /* num of nonzeros in Co to be sent to [proc] */ 299 } 300 301 len = 0; /* max length of buf_si[], see (4) */ 302 owners_co[0] = 0; 303 nsend = 0; 304 for (proc=0; proc<size; proc++) { 305 owners_co[proc+1] = owners_co[proc] + len_si[proc]; 306 if (len_s[proc]) { 307 nsend++; 308 len_si[proc] = 2*(len_si[proc] + 1); /* length of buf_si to be sent to [proc] */ 309 len += len_si[proc]; 310 } 311 } 312 313 /* determine the number and length of messages to receive for coi and coj */ 314 ierr = PetscGatherNumberOfMessages(comm,NULL,len_s,&nrecv);CHKERRQ(ierr); 315 ierr = PetscGatherMessageLengths2(comm,nsend,nrecv,len_s,len_si,&id_r,&len_r,&len_ri);CHKERRQ(ierr); 316 317 /* post the Irecv and Isend of coj */ 318 ierr = PetscCommGetNewTag(comm,&tagj);CHKERRQ(ierr); 319 ierr = PetscPostIrecvInt(comm,tagj,nrecv,id_r,len_r,&buf_rj,&rwaits);CHKERRQ(ierr); 320 ierr = PetscMalloc1(nsend+1,&swaits);CHKERRQ(ierr); 321 for (proc=0, k=0; proc<size; proc++) { 322 if (!len_s[proc]) continue; 323 i = owners_co[proc]; 324 ierr = MPI_Isend(coj+coi[i],len_s[proc],MPIU_INT,proc,tagj,comm,swaits+k);CHKERRQ(ierr); 325 k++; 326 } 327 328 /* (2-2) compute symbolic C_loc = Rd*AP_loc */ 329 /* ---------------------------------------- */ 330 ierr = MatMatMultSymbolic_SeqAIJ_SeqAIJ(ptap->Rd,ptap->AP_loc,fill,&ptap->C_loc);CHKERRQ(ierr); 331 c_loc = (Mat_SeqAIJ*)ptap->C_loc->data; 332 333 /* receives coj are complete */ 334 for (i=0; i<nrecv; i++) { 335 ierr = MPI_Waitany(nrecv,rwaits,&icompleted,&rstatus);CHKERRQ(ierr); 336 } 337 ierr = PetscFree(rwaits);CHKERRQ(ierr); 338 if (nsend) {ierr = MPI_Waitall(nsend,swaits,sstatus);CHKERRQ(ierr);} 339 340 /* add received column indices into ta to update Crmax */ 341 for (k=0; k<nrecv; k++) {/* k-th received message */ 342 Jptr = buf_rj[k]; 343 for (j=0; j<len_r[k]; j++) { 344 ierr = PetscTableAdd(ta,*(Jptr+j)+1,1,INSERT_VALUES);CHKERRQ(ierr); 345 } 346 } 347 ierr = PetscTableGetCount(ta,&Crmax);CHKERRQ(ierr); 348 ierr = PetscTableDestroy(&ta);CHKERRQ(ierr); 349 350 /* (4) send and recv coi */ 351 /*-----------------------*/ 352 ierr = PetscCommGetNewTag(comm,&tagi);CHKERRQ(ierr); 353 ierr = PetscPostIrecvInt(comm,tagi,nrecv,id_r,len_ri,&buf_ri,&rwaits);CHKERRQ(ierr); 354 ierr = PetscMalloc1(len+1,&buf_s);CHKERRQ(ierr); 355 buf_si = buf_s; /* points to the beginning of k-th msg to be sent */ 356 for (proc=0,k=0; proc<size; proc++) { 357 if (!len_s[proc]) continue; 358 /* form outgoing message for i-structure: 359 buf_si[0]: nrows to be sent 360 [1:nrows]: row index (global) 361 [nrows+1:2*nrows+1]: i-structure index 362 */ 363 /*-------------------------------------------*/ 364 nrows = len_si[proc]/2 - 1; /* num of rows in Co to be sent to [proc] */ 365 buf_si_i = buf_si + nrows+1; 366 buf_si[0] = nrows; 367 buf_si_i[0] = 0; 368 nrows = 0; 369 for (i=owners_co[proc]; i<owners_co[proc+1]; i++) { 370 nzi = coi[i+1] - coi[i]; 371 buf_si_i[nrows+1] = buf_si_i[nrows] + nzi; /* i-structure */ 372 buf_si[nrows+1] = prmap[i] -owners[proc]; /* local row index */ 373 nrows++; 374 } 375 ierr = MPI_Isend(buf_si,len_si[proc],MPIU_INT,proc,tagi,comm,swaits+k);CHKERRQ(ierr); 376 k++; 377 buf_si += len_si[proc]; 378 } 379 for (i=0; i<nrecv; i++) { 380 ierr = MPI_Waitany(nrecv,rwaits,&icompleted,&rstatus);CHKERRQ(ierr); 381 } 382 ierr = PetscFree(rwaits);CHKERRQ(ierr); 383 if (nsend) {ierr = MPI_Waitall(nsend,swaits,sstatus);CHKERRQ(ierr);} 384 385 ierr = PetscFree4(len_s,len_si,sstatus,owners_co);CHKERRQ(ierr); 386 ierr = PetscFree(len_ri);CHKERRQ(ierr); 387 ierr = PetscFree(swaits);CHKERRQ(ierr); 388 ierr = PetscFree(buf_s);CHKERRQ(ierr); 389 #if defined(PTAP_PROFILE) 390 ierr = PetscTime(&t2);CHKERRQ(ierr); 391 #endif 392 /* (5) compute the local portion of Cmpi */ 393 /* ------------------------------------------ */ 394 /* set initial free space to be Crmax, sufficient for holding nozeros in each row of Cmpi */ 395 ierr = PetscFreeSpaceGet(Crmax,&free_space);CHKERRQ(ierr); 396 current_space = free_space; 397 398 ierr = PetscMalloc3(nrecv,&buf_ri_k,nrecv,&nextrow,nrecv,&nextci);CHKERRQ(ierr); 399 for (k=0; k<nrecv; k++) { 400 buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */ 401 nrows = *buf_ri_k[k]; 402 nextrow[k] = buf_ri_k[k] + 1; /* next row number of k-th recved i-structure */ 403 nextci[k] = buf_ri_k[k] + (nrows + 1); /* poins to the next i-structure of k-th recved i-structure */ 404 } 405 406 ierr = MatPreallocateInitialize(comm,pn,pn,dnz,onz);CHKERRQ(ierr); 407 ierr = PetscLLCondensedCreate(Crmax,pN,&lnk,&lnkbt);CHKERRQ(ierr); 408 for (i=0; i<pn; i++) { 409 /* add C_loc into Cmpi */ 410 nzi = c_loc->i[i+1] - c_loc->i[i]; 411 Jptr = c_loc->j + c_loc->i[i]; 412 ierr = PetscLLCondensedAddSorted(nzi,Jptr,lnk,lnkbt);CHKERRQ(ierr); 413 414 /* add received col data into lnk */ 415 for (k=0; k<nrecv; k++) { /* k-th received message */ 416 if (i == *nextrow[k]) { /* i-th row */ 417 nzi = *(nextci[k]+1) - *nextci[k]; 418 Jptr = buf_rj[k] + *nextci[k]; 419 ierr = PetscLLCondensedAddSorted(nzi,Jptr,lnk,lnkbt);CHKERRQ(ierr); 420 nextrow[k]++; nextci[k]++; 421 } 422 } 423 nzi = lnk[0]; 424 425 /* copy data into free space, then initialize lnk */ 426 ierr = PetscLLCondensedClean(pN,nzi,current_space->array,lnk,lnkbt);CHKERRQ(ierr); 427 ierr = MatPreallocateSet(i+owners[rank],nzi,current_space->array,dnz,onz);CHKERRQ(ierr); 428 } 429 ierr = PetscFree3(buf_ri_k,nextrow,nextci);CHKERRQ(ierr); 430 ierr = PetscLLDestroy(lnk,lnkbt);CHKERRQ(ierr); 431 ierr = PetscFreeSpaceDestroy(free_space);CHKERRQ(ierr); 432 #if defined(PTAP_PROFILE) 433 ierr = PetscTime(&t3);CHKERRQ(ierr); 434 #endif 435 436 /* local sizes and preallocation */ 437 ierr = MatSetSizes(Cmpi,pn,pn,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 438 ierr = MatSetBlockSizes(Cmpi,PetscAbs(P->cmap->bs),PetscAbs(P->cmap->bs));CHKERRQ(ierr); 439 ierr = MatMPIAIJSetPreallocation(Cmpi,0,dnz,0,onz);CHKERRQ(ierr); 440 ierr = MatPreallocateFinalize(dnz,onz);CHKERRQ(ierr); 441 442 /* members in merge */ 443 ierr = PetscFree(id_r);CHKERRQ(ierr); 444 ierr = PetscFree(len_r);CHKERRQ(ierr); 445 ierr = PetscFree(buf_ri[0]);CHKERRQ(ierr); 446 ierr = PetscFree(buf_ri);CHKERRQ(ierr); 447 ierr = PetscFree(buf_rj[0]);CHKERRQ(ierr); 448 ierr = PetscFree(buf_rj);CHKERRQ(ierr); 449 ierr = PetscLayoutDestroy(&rowmap);CHKERRQ(ierr); 450 451 /* attach the supporting struct to Cmpi for reuse */ 452 c = (Mat_MPIAIJ*)Cmpi->data; 453 c->ptap = ptap; 454 ptap->duplicate = Cmpi->ops->duplicate; 455 ptap->destroy = Cmpi->ops->destroy; 456 457 if (alg == 1) { 458 ierr = PetscCalloc1(pN,&ptap->apa);CHKERRQ(ierr); 459 } 460 461 /* Cmpi is not ready for use - assembly will be done by MatPtAPNumeric() */ 462 Cmpi->assembled = PETSC_FALSE; 463 Cmpi->ops->destroy = MatDestroy_MPIAIJ_PtAP; 464 Cmpi->ops->duplicate = MatDuplicate_MPIAIJ_MatPtAP; 465 *C = Cmpi; 466 467 #if defined(PTAP_PROFILE) 468 ierr = PetscTime(&t4);CHKERRQ(ierr); 469 if (rank == 1) { 470 printf("PtAPSym: %g + %g + %g + %g + %g + %g = %g\n",t11-t0,t1-t11,t12-t11,t2-t2,t3-t2,t4-t3,t4-t0); 471 } 472 #endif 473 PetscFunctionReturn(0); 474 } 475 476 PetscErrorCode MatPtAPSymbolic_MPIAIJ_MPIAIJ(Mat A,Mat P,PetscReal fill,Mat *C) 477 { 478 PetscErrorCode ierr; 479 Mat_PtAPMPI *ptap; 480 Mat_MPIAIJ *a=(Mat_MPIAIJ*)A->data,*p=(Mat_MPIAIJ*)P->data,*c; 481 MPI_Comm comm; 482 PetscMPIInt size,rank; 483 Mat Cmpi; 484 PetscFreeSpaceList free_space=NULL,current_space=NULL; 485 PetscInt am=A->rmap->n,pm=P->rmap->n,pN=P->cmap->N,pn=P->cmap->n; 486 PetscInt *lnk,i,k,pnz,row,nsend; 487 PetscBT lnkbt; 488 PetscMPIInt tagi,tagj,*len_si,*len_s,*len_ri,icompleted=0,nrecv; 489 PetscInt **buf_rj,**buf_ri,**buf_ri_k; 490 PetscInt len,proc,*dnz,*onz,*owners,nzi,nspacedouble; 491 PetscInt nrows,*buf_s,*buf_si,*buf_si_i,**nextrow,**nextci; 492 MPI_Request *swaits,*rwaits; 493 MPI_Status *sstatus,rstatus; 494 PetscLayout rowmap; 495 PetscInt *owners_co,*coi,*coj; /* i and j array of (p->B)^T*A*P - used in the communication */ 496 PetscMPIInt *len_r,*id_r; /* array of length of comm->size, store send/recv matrix values */ 497 PetscInt *api,*apj,*Jptr,apnz,*prmap=p->garray,con,j,ap_rmax=0,Crmax,*aj,*ai,*pi; 498 Mat_SeqAIJ *p_loc,*p_oth,*ad=(Mat_SeqAIJ*)(a->A)->data,*ao=(Mat_SeqAIJ*)(a->B)->data,*c_loc,*c_oth; 499 PetscScalar *apv; 500 PetscTable ta; 501 #if defined(PETSC_HAVE_HYPRE) 502 const char *algTypes[3] = {"scalable","nonscalable","hypre"}; 503 PetscInt nalg = 3; 504 #else 505 const char *algTypes[2] = {"scalable","nonscalable"}; 506 PetscInt nalg = 2; 507 #endif 508 PetscInt alg = 1; /* set default algorithm */ 509 #if defined(PETSC_USE_INFO) 510 PetscReal apfill; 511 #endif 512 #if defined(PTAP_PROFILE) 513 PetscLogDouble t0,t1,t11,t12,t2,t3,t4; 514 #endif 515 516 PetscFunctionBegin; 517 /* pick an algorithm */ 518 ierr = PetscObjectOptionsBegin((PetscObject)A);CHKERRQ(ierr); 519 ierr = PetscOptionsEList("-matptap_via","Algorithmic approach","MatPtAP",algTypes,nalg,algTypes[1],&alg,NULL);CHKERRQ(ierr); 520 ierr = PetscOptionsEnd();CHKERRQ(ierr); 521 522 if (alg == 0) { 523 ierr = MatPtAPSymbolic_MPIAIJ_MPIAIJ_scalable(A,P,fill,C);CHKERRQ(ierr); 524 (*C)->ops->ptapnumeric = MatPtAPNumeric_MPIAIJ_MPIAIJ; /* _scalable; not done yet */ 525 PetscFunctionReturn(0); 526 527 #if defined(PETSC_HAVE_HYPRE) 528 } else if (alg == 2) { 529 /* Use boomerAMGBuildCoarseOperator */ 530 ierr = MatPtAPSymbolic_AIJ_AIJ_wHYPRE(A,P,fill,C);CHKERRQ(ierr); 531 PetscFunctionReturn(0); 532 #endif 533 } 534 535 #if defined(PTAP_PROFILE) 536 ierr = PetscTime(&t0);CHKERRQ(ierr); 537 #endif 538 539 ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr); 540 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 541 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 542 543 /* create symbolic parallel matrix Cmpi */ 544 ierr = MatCreate(comm,&Cmpi);CHKERRQ(ierr); 545 ierr = MatSetType(Cmpi,MATMPIAIJ);CHKERRQ(ierr); 546 547 /* Do dense axpy in MatPtAPNumeric_MPIAIJ_MPIAIJ() */ 548 Cmpi->ops->ptapnumeric = MatPtAPNumeric_MPIAIJ_MPIAIJ; 549 550 /* create struct Mat_PtAPMPI and attached it to C later */ 551 ierr = PetscNew(&ptap);CHKERRQ(ierr); 552 ptap->reuse = MAT_INITIAL_MATRIX; 553 554 /* get P_oth by taking rows of P (= non-zero cols of local A) from other processors */ 555 ierr = MatGetBrowsOfAoCols_MPIAIJ(A,P,MAT_INITIAL_MATRIX,&ptap->startsj_s,&ptap->startsj_r,&ptap->bufa,&ptap->P_oth);CHKERRQ(ierr); 556 /* get P_loc by taking all local rows of P */ 557 ierr = MatMPIAIJGetLocalMat(P,MAT_INITIAL_MATRIX,&ptap->P_loc);CHKERRQ(ierr); 558 559 /* (0) compute Rd = Pd^T, Ro = Po^T */ 560 /* --------------------------------- */ 561 ierr = MatTranspose_SeqAIJ(p->A,MAT_INITIAL_MATRIX,&ptap->Rd);CHKERRQ(ierr); 562 ierr = MatTranspose_SeqAIJ(p->B,MAT_INITIAL_MATRIX,&ptap->Ro);CHKERRQ(ierr); 563 #if defined(PTAP_PROFILE) 564 ierr = PetscTime(&t11);CHKERRQ(ierr); 565 #endif 566 567 /* (1) compute symbolic AP = A_loc*P = Ad*P_loc + Ao*P_oth (api,apj) */ 568 /* ----------------------------------------------------------------- */ 569 p_loc = (Mat_SeqAIJ*)(ptap->P_loc)->data; 570 p_oth = (Mat_SeqAIJ*)(ptap->P_oth)->data; 571 572 /* create and initialize a linked list */ 573 ierr = PetscTableCreate(pn,pN,&ta);CHKERRQ(ierr); /* for compute AP_loc and Cmpi */ 574 MatRowMergeMax_SeqAIJ(p_loc,ptap->P_loc->rmap->N,ta); 575 MatRowMergeMax_SeqAIJ(p_oth,ptap->P_oth->rmap->N,ta); 576 ierr = PetscTableGetCount(ta,&Crmax);CHKERRQ(ierr); /* Crmax = nnz(sum of Prows) */ 577 /* printf("[%d] est %d, Crmax %d; pN %d\n",rank,5*(p_loc->rmax+p_oth->rmax + (PetscInt)(1.e-2*pN)),Crmax,pN); */ 578 579 ierr = PetscLLCondensedCreate(Crmax,pN,&lnk,&lnkbt);CHKERRQ(ierr); 580 581 /* Initial FreeSpace size is fill*(nnz(A) + nnz(P)) */ 582 ierr = PetscFreeSpaceGet(PetscRealIntMultTruncate(fill,PetscIntSumTruncate(ad->i[am],PetscIntSumTruncate(ao->i[am],p_loc->i[pm]))),&free_space);CHKERRQ(ierr); 583 current_space = free_space; 584 nspacedouble = 0; 585 586 ierr = PetscMalloc1(am+1,&api);CHKERRQ(ierr); 587 api[0] = 0; 588 for (i=0; i<am; i++) { 589 /* diagonal portion: Ad[i,:]*P */ 590 ai = ad->i; pi = p_loc->i; 591 nzi = ai[i+1] - ai[i]; 592 aj = ad->j + ai[i]; 593 for (j=0; j<nzi; j++) { 594 row = aj[j]; 595 pnz = pi[row+1] - pi[row]; 596 Jptr = p_loc->j + pi[row]; 597 /* add non-zero cols of P into the sorted linked list lnk */ 598 ierr = PetscLLCondensedAddSorted(pnz,Jptr,lnk,lnkbt);CHKERRQ(ierr); 599 } 600 /* off-diagonal portion: Ao[i,:]*P */ 601 ai = ao->i; pi = p_oth->i; 602 nzi = ai[i+1] - ai[i]; 603 aj = ao->j + ai[i]; 604 for (j=0; j<nzi; j++) { 605 row = aj[j]; 606 pnz = pi[row+1] - pi[row]; 607 Jptr = p_oth->j + pi[row]; 608 ierr = PetscLLCondensedAddSorted(pnz,Jptr,lnk,lnkbt);CHKERRQ(ierr); 609 } 610 apnz = lnk[0]; 611 api[i+1] = api[i] + apnz; 612 if (ap_rmax < apnz) ap_rmax = apnz; 613 614 /* if free space is not available, double the total space in the list */ 615 if (current_space->local_remaining<apnz) { 616 ierr = PetscFreeSpaceGet(PetscIntSumTruncate(apnz,current_space->total_array_size),¤t_space);CHKERRQ(ierr); 617 nspacedouble++; 618 } 619 620 /* Copy data into free space, then initialize lnk */ 621 ierr = PetscLLCondensedClean(pN,apnz,current_space->array,lnk,lnkbt);CHKERRQ(ierr); 622 623 current_space->array += apnz; 624 current_space->local_used += apnz; 625 current_space->local_remaining -= apnz; 626 } 627 /* Allocate space for apj and apv, initialize apj, and */ 628 /* destroy list of free space and other temporary array(s) */ 629 ierr = PetscMalloc2(api[am],&apj,api[am],&apv);CHKERRQ(ierr); 630 ierr = PetscFreeSpaceContiguous(&free_space,apj);CHKERRQ(ierr); 631 ierr = PetscLLDestroy(lnk,lnkbt);CHKERRQ(ierr); 632 633 /* Create AP_loc for reuse */ 634 ierr = MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,am,pN,api,apj,apv,&ptap->AP_loc);CHKERRQ(ierr); 635 636 #if defined(PETSC_USE_INFO) 637 apfill = (PetscReal)api[am]/(ad->i[am]+ao->i[am]+p_loc->i[pm]+1); 638 ptap->AP_loc->info.mallocs = nspacedouble; 639 ptap->AP_loc->info.fill_ratio_given = fill; 640 ptap->AP_loc->info.fill_ratio_needed = apfill; 641 642 if (api[am]) { 643 ierr = PetscInfo3(ptap->AP_loc,"AP_loc reallocs %D; Fill ratio: given %g needed %g.\n",nspacedouble,(double)fill,(double)apfill);CHKERRQ(ierr); 644 ierr = PetscInfo1(ptap->AP_loc,"Use MatPtAP(A,B,MatReuse,%g,&C) for best AP_loc performance.;\n",(double)apfill);CHKERRQ(ierr); 645 } else { 646 ierr = PetscInfo(ptap->AP_loc,"AP_loc is empty \n");CHKERRQ(ierr); 647 } 648 #endif 649 650 #if defined(PTAP_PROFILE) 651 ierr = PetscTime(&t12);CHKERRQ(ierr); 652 #endif 653 654 /* (2-1) compute symbolic Co = Ro*AP_loc */ 655 /* ------------------------------------ */ 656 ierr = MatMatMultSymbolic_SeqAIJ_SeqAIJ(ptap->Ro,ptap->AP_loc,fill,&ptap->C_oth);CHKERRQ(ierr); 657 #if defined(PTAP_PROFILE) 658 ierr = PetscTime(&t1);CHKERRQ(ierr); 659 #endif 660 661 /* (3) send coj of C_oth to other processors */ 662 /* ------------------------------------------ */ 663 /* determine row ownership */ 664 ierr = PetscLayoutCreate(comm,&rowmap);CHKERRQ(ierr); 665 rowmap->n = pn; 666 rowmap->bs = 1; 667 ierr = PetscLayoutSetUp(rowmap);CHKERRQ(ierr); 668 owners = rowmap->range; 669 670 /* determine the number of messages to send, their lengths */ 671 ierr = PetscMalloc4(size,&len_s,size,&len_si,size,&sstatus,size+2,&owners_co);CHKERRQ(ierr); 672 ierr = PetscMemzero(len_s,size*sizeof(PetscMPIInt));CHKERRQ(ierr); 673 ierr = PetscMemzero(len_si,size*sizeof(PetscMPIInt));CHKERRQ(ierr); 674 675 c_oth = (Mat_SeqAIJ*)ptap->C_oth->data; 676 coi = c_oth->i; coj = c_oth->j; 677 con = ptap->C_oth->rmap->n; 678 proc = 0; 679 for (i=0; i<con; i++) { 680 while (prmap[i] >= owners[proc+1]) proc++; 681 len_si[proc]++; /* num of rows in Co(=Pt*AP) to be sent to [proc] */ 682 len_s[proc] += coi[i+1] - coi[i]; /* num of nonzeros in Co to be sent to [proc] */ 683 } 684 685 len = 0; /* max length of buf_si[], see (4) */ 686 owners_co[0] = 0; 687 nsend = 0; 688 for (proc=0; proc<size; proc++) { 689 owners_co[proc+1] = owners_co[proc] + len_si[proc]; 690 if (len_s[proc]) { 691 nsend++; 692 len_si[proc] = 2*(len_si[proc] + 1); /* length of buf_si to be sent to [proc] */ 693 len += len_si[proc]; 694 } 695 } 696 697 /* determine the number and length of messages to receive for coi and coj */ 698 ierr = PetscGatherNumberOfMessages(comm,NULL,len_s,&nrecv);CHKERRQ(ierr); 699 ierr = PetscGatherMessageLengths2(comm,nsend,nrecv,len_s,len_si,&id_r,&len_r,&len_ri);CHKERRQ(ierr); 700 701 /* post the Irecv and Isend of coj */ 702 ierr = PetscCommGetNewTag(comm,&tagj);CHKERRQ(ierr); 703 ierr = PetscPostIrecvInt(comm,tagj,nrecv,id_r,len_r,&buf_rj,&rwaits);CHKERRQ(ierr); 704 ierr = PetscMalloc1(nsend+1,&swaits);CHKERRQ(ierr); 705 for (proc=0, k=0; proc<size; proc++) { 706 if (!len_s[proc]) continue; 707 i = owners_co[proc]; 708 ierr = MPI_Isend(coj+coi[i],len_s[proc],MPIU_INT,proc,tagj,comm,swaits+k);CHKERRQ(ierr); 709 k++; 710 } 711 712 /* (2-2) compute symbolic C_loc = Rd*AP_loc */ 713 /* ---------------------------------------- */ 714 ierr = MatMatMultSymbolic_SeqAIJ_SeqAIJ(ptap->Rd,ptap->AP_loc,fill,&ptap->C_loc);CHKERRQ(ierr); 715 c_loc = (Mat_SeqAIJ*)ptap->C_loc->data; 716 717 /* receives coj are complete */ 718 for (i=0; i<nrecv; i++) { 719 ierr = MPI_Waitany(nrecv,rwaits,&icompleted,&rstatus);CHKERRQ(ierr); 720 } 721 ierr = PetscFree(rwaits);CHKERRQ(ierr); 722 if (nsend) {ierr = MPI_Waitall(nsend,swaits,sstatus);CHKERRQ(ierr);} 723 724 /* add received column indices into ta to update Crmax */ 725 for (k=0; k<nrecv; k++) {/* k-th received message */ 726 Jptr = buf_rj[k]; 727 for (j=0; j<len_r[k]; j++) { 728 ierr = PetscTableAdd(ta,*(Jptr+j)+1,1,INSERT_VALUES);CHKERRQ(ierr); 729 } 730 } 731 ierr = PetscTableGetCount(ta,&Crmax);CHKERRQ(ierr); 732 ierr = PetscTableDestroy(&ta);CHKERRQ(ierr); 733 734 /* (4) send and recv coi */ 735 /*-----------------------*/ 736 ierr = PetscCommGetNewTag(comm,&tagi);CHKERRQ(ierr); 737 ierr = PetscPostIrecvInt(comm,tagi,nrecv,id_r,len_ri,&buf_ri,&rwaits);CHKERRQ(ierr); 738 ierr = PetscMalloc1(len+1,&buf_s);CHKERRQ(ierr); 739 buf_si = buf_s; /* points to the beginning of k-th msg to be sent */ 740 for (proc=0,k=0; proc<size; proc++) { 741 if (!len_s[proc]) continue; 742 /* form outgoing message for i-structure: 743 buf_si[0]: nrows to be sent 744 [1:nrows]: row index (global) 745 [nrows+1:2*nrows+1]: i-structure index 746 */ 747 /*-------------------------------------------*/ 748 nrows = len_si[proc]/2 - 1; /* num of rows in Co to be sent to [proc] */ 749 buf_si_i = buf_si + nrows+1; 750 buf_si[0] = nrows; 751 buf_si_i[0] = 0; 752 nrows = 0; 753 for (i=owners_co[proc]; i<owners_co[proc+1]; i++) { 754 nzi = coi[i+1] - coi[i]; 755 buf_si_i[nrows+1] = buf_si_i[nrows] + nzi; /* i-structure */ 756 buf_si[nrows+1] = prmap[i] -owners[proc]; /* local row index */ 757 nrows++; 758 } 759 ierr = MPI_Isend(buf_si,len_si[proc],MPIU_INT,proc,tagi,comm,swaits+k);CHKERRQ(ierr); 760 k++; 761 buf_si += len_si[proc]; 762 } 763 for (i=0; i<nrecv; i++) { 764 ierr = MPI_Waitany(nrecv,rwaits,&icompleted,&rstatus);CHKERRQ(ierr); 765 } 766 ierr = PetscFree(rwaits);CHKERRQ(ierr); 767 if (nsend) {ierr = MPI_Waitall(nsend,swaits,sstatus);CHKERRQ(ierr);} 768 769 ierr = PetscFree4(len_s,len_si,sstatus,owners_co);CHKERRQ(ierr); 770 ierr = PetscFree(len_ri);CHKERRQ(ierr); 771 ierr = PetscFree(swaits);CHKERRQ(ierr); 772 ierr = PetscFree(buf_s);CHKERRQ(ierr); 773 #if defined(PTAP_PROFILE) 774 ierr = PetscTime(&t2);CHKERRQ(ierr); 775 #endif 776 /* (5) compute the local portion of Cmpi */ 777 /* ------------------------------------------ */ 778 /* set initial free space to be Crmax, sufficient for holding nozeros in each row of Cmpi */ 779 ierr = PetscFreeSpaceGet(Crmax,&free_space);CHKERRQ(ierr); 780 current_space = free_space; 781 782 ierr = PetscMalloc3(nrecv,&buf_ri_k,nrecv,&nextrow,nrecv,&nextci);CHKERRQ(ierr); 783 for (k=0; k<nrecv; k++) { 784 buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */ 785 nrows = *buf_ri_k[k]; 786 nextrow[k] = buf_ri_k[k] + 1; /* next row number of k-th recved i-structure */ 787 nextci[k] = buf_ri_k[k] + (nrows + 1); /* poins to the next i-structure of k-th recved i-structure */ 788 } 789 790 ierr = MatPreallocateInitialize(comm,pn,pn,dnz,onz);CHKERRQ(ierr); 791 ierr = PetscLLCondensedCreate(Crmax,pN,&lnk,&lnkbt);CHKERRQ(ierr); 792 for (i=0; i<pn; i++) { 793 /* add C_loc into Cmpi */ 794 nzi = c_loc->i[i+1] - c_loc->i[i]; 795 Jptr = c_loc->j + c_loc->i[i]; 796 ierr = PetscLLCondensedAddSorted(nzi,Jptr,lnk,lnkbt);CHKERRQ(ierr); 797 798 /* add received col data into lnk */ 799 for (k=0; k<nrecv; k++) { /* k-th received message */ 800 if (i == *nextrow[k]) { /* i-th row */ 801 nzi = *(nextci[k]+1) - *nextci[k]; 802 Jptr = buf_rj[k] + *nextci[k]; 803 ierr = PetscLLCondensedAddSorted(nzi,Jptr,lnk,lnkbt);CHKERRQ(ierr); 804 nextrow[k]++; nextci[k]++; 805 } 806 } 807 nzi = lnk[0]; 808 809 /* copy data into free space, then initialize lnk */ 810 ierr = PetscLLCondensedClean(pN,nzi,current_space->array,lnk,lnkbt);CHKERRQ(ierr); 811 ierr = MatPreallocateSet(i+owners[rank],nzi,current_space->array,dnz,onz);CHKERRQ(ierr); 812 } 813 ierr = PetscFree3(buf_ri_k,nextrow,nextci);CHKERRQ(ierr); 814 ierr = PetscLLDestroy(lnk,lnkbt);CHKERRQ(ierr); 815 ierr = PetscFreeSpaceDestroy(free_space);CHKERRQ(ierr); 816 #if defined(PTAP_PROFILE) 817 ierr = PetscTime(&t3);CHKERRQ(ierr); 818 #endif 819 820 /* local sizes and preallocation */ 821 ierr = MatSetSizes(Cmpi,pn,pn,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 822 ierr = MatSetBlockSizes(Cmpi,PetscAbs(P->cmap->bs),PetscAbs(P->cmap->bs));CHKERRQ(ierr); 823 ierr = MatMPIAIJSetPreallocation(Cmpi,0,dnz,0,onz);CHKERRQ(ierr); 824 ierr = MatPreallocateFinalize(dnz,onz);CHKERRQ(ierr); 825 826 /* members in merge */ 827 ierr = PetscFree(id_r);CHKERRQ(ierr); 828 ierr = PetscFree(len_r);CHKERRQ(ierr); 829 ierr = PetscFree(buf_ri[0]);CHKERRQ(ierr); 830 ierr = PetscFree(buf_ri);CHKERRQ(ierr); 831 ierr = PetscFree(buf_rj[0]);CHKERRQ(ierr); 832 ierr = PetscFree(buf_rj);CHKERRQ(ierr); 833 ierr = PetscLayoutDestroy(&rowmap);CHKERRQ(ierr); 834 835 /* attach the supporting struct to Cmpi for reuse */ 836 c = (Mat_MPIAIJ*)Cmpi->data; 837 c->ptap = ptap; 838 ptap->duplicate = Cmpi->ops->duplicate; 839 ptap->destroy = Cmpi->ops->destroy; 840 841 if (alg == 1) { 842 ierr = PetscCalloc1(pN,&ptap->apa);CHKERRQ(ierr); 843 } 844 845 /* Cmpi is not ready for use - assembly will be done by MatPtAPNumeric() */ 846 Cmpi->assembled = PETSC_FALSE; 847 Cmpi->ops->destroy = MatDestroy_MPIAIJ_PtAP; 848 Cmpi->ops->duplicate = MatDuplicate_MPIAIJ_MatPtAP; 849 *C = Cmpi; 850 851 #if defined(PTAP_PROFILE) 852 ierr = PetscTime(&t4);CHKERRQ(ierr); 853 if (rank == 1) { 854 printf("PtAPSym: %g + %g + %g + %g + %g + %g = %g\n",t11-t0,t1-t11,t12-t11,t2-t2,t3-t2,t4-t3,t4-t0); 855 } 856 #endif 857 PetscFunctionReturn(0); 858 } 859 860 PetscErrorCode MatPtAPNumeric_MPIAIJ_MPIAIJ(Mat A,Mat P,Mat C) 861 { 862 PetscErrorCode ierr; 863 Mat_MPIAIJ *a=(Mat_MPIAIJ*)A->data,*p=(Mat_MPIAIJ*)P->data,*c=(Mat_MPIAIJ*)C->data; 864 Mat_SeqAIJ *ad=(Mat_SeqAIJ*)(a->A)->data,*ao=(Mat_SeqAIJ*)(a->B)->data; 865 Mat_SeqAIJ *ap,*p_loc,*p_oth,*c_seq; 866 Mat_PtAPMPI *ptap = c->ptap; 867 Mat AP_loc,C_loc,C_oth; 868 PetscInt i,rstart,rend,cm,ncols,row; 869 PetscInt *api,*apj,am = A->rmap->n,j,col,apnz; 870 PetscScalar *apa; 871 const PetscInt *cols; 872 const PetscScalar *vals; 873 #if defined(PTAP_PROFILE) 874 PetscMPIInt rank; 875 MPI_Comm comm; 876 PetscLogDouble t0,t1,t2,t3,t4,eR,eAP,eCseq,eCmpi; 877 #endif 878 879 PetscFunctionBegin; 880 ierr = MatZeroEntries(C);CHKERRQ(ierr); 881 882 /* 1) get R = Pd^T,Ro = Po^T */ 883 #if defined(PTAP_PROFILE) 884 ierr = PetscTime(&t0);CHKERRQ(ierr); 885 #endif 886 if (ptap->reuse == MAT_REUSE_MATRIX) { 887 ierr = MatTranspose_SeqAIJ(p->A,MAT_REUSE_MATRIX,&ptap->Rd);CHKERRQ(ierr); 888 ierr = MatTranspose_SeqAIJ(p->B,MAT_REUSE_MATRIX,&ptap->Ro);CHKERRQ(ierr); 889 } 890 #if defined(PTAP_PROFILE) 891 ierr = PetscTime(&t1);CHKERRQ(ierr); 892 eR = t1 - t0; 893 #endif 894 895 /* 2) get AP_loc */ 896 AP_loc = ptap->AP_loc; 897 ap = (Mat_SeqAIJ*)AP_loc->data; 898 899 /* 2-1) get P_oth = ptap->P_oth and P_loc = ptap->P_loc */ 900 /*-----------------------------------------------------*/ 901 if (ptap->reuse == MAT_REUSE_MATRIX) { 902 /* P_oth and P_loc are obtained in MatPtASymbolic() when reuse == MAT_INITIAL_MATRIX */ 903 ierr = MatGetBrowsOfAoCols_MPIAIJ(A,P,MAT_REUSE_MATRIX,&ptap->startsj_s,&ptap->startsj_r,&ptap->bufa,&ptap->P_oth);CHKERRQ(ierr); 904 ierr = MatMPIAIJGetLocalMat(P,MAT_REUSE_MATRIX,&ptap->P_loc);CHKERRQ(ierr); 905 } 906 907 /* 2-2) compute numeric A_loc*P - dominating part */ 908 /* ---------------------------------------------- */ 909 /* get data from symbolic products */ 910 p_loc = (Mat_SeqAIJ*)(ptap->P_loc)->data; 911 p_oth = (Mat_SeqAIJ*)(ptap->P_oth)->data; 912 apa = ptap->apa; 913 api = ap->i; 914 apj = ap->j; 915 for (i=0; i<am; i++) { 916 /* AP[i,:] = A[i,:]*P = Ad*P_loc Ao*P_oth */ 917 AProw_nonscalable(i,ad,ao,p_loc,p_oth,apa); 918 apnz = api[i+1] - api[i]; 919 for (j=0; j<apnz; j++) { 920 col = apj[j+api[i]]; 921 ap->a[j+ap->i[i]] = apa[col]; 922 apa[col] = 0.0; 923 } 924 ierr = PetscLogFlops(2.0*apnz);CHKERRQ(ierr); 925 } 926 #if defined(PTAP_PROFILE) 927 ierr = PetscTime(&t2);CHKERRQ(ierr); 928 eAP = t2 - t1; 929 #endif 930 931 /* 3) C_loc = Rd*AP_loc, C_oth = Ro*AP_loc */ 932 ierr = ((ptap->C_loc)->ops->matmultnumeric)(ptap->Rd,AP_loc,ptap->C_loc);CHKERRQ(ierr); 933 ierr = ((ptap->C_oth)->ops->matmultnumeric)(ptap->Ro,AP_loc,ptap->C_oth);CHKERRQ(ierr); 934 C_loc = ptap->C_loc; 935 C_oth = ptap->C_oth; 936 #if defined(PTAP_PROFILE) 937 ierr = PetscTime(&t3);CHKERRQ(ierr); 938 eCseq = t3 - t2; 939 #endif 940 941 /* add C_loc and Co to to C */ 942 ierr = MatGetOwnershipRange(C,&rstart,&rend);CHKERRQ(ierr); 943 944 /* C_loc -> C */ 945 cm = C_loc->rmap->N; 946 c_seq = (Mat_SeqAIJ*)C_loc->data; 947 cols = c_seq->j; 948 vals = c_seq->a; 949 for (i=0; i<cm; i++) { 950 ncols = c_seq->i[i+1] - c_seq->i[i]; 951 row = rstart + i; 952 ierr = MatSetValues(C,1,&row,ncols,cols,vals,ADD_VALUES);CHKERRQ(ierr); 953 cols += ncols; vals += ncols; 954 } 955 956 /* Co -> C, off-processor part */ 957 cm = C_oth->rmap->N; 958 c_seq = (Mat_SeqAIJ*)C_oth->data; 959 cols = c_seq->j; 960 vals = c_seq->a; 961 for (i=0; i<cm; i++) { 962 ncols = c_seq->i[i+1] - c_seq->i[i]; 963 row = p->garray[i]; 964 ierr = MatSetValues(C,1,&row,ncols,cols,vals,ADD_VALUES);CHKERRQ(ierr); 965 cols += ncols; vals += ncols; 966 } 967 ierr = MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 968 ierr = MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 969 #if defined(PTAP_PROFILE) 970 ierr = PetscTime(&t4);CHKERRQ(ierr); 971 eCmpi = t4 - t3; 972 973 ierr = PetscObjectGetComm((PetscObject)C,&comm);CHKERRQ(ierr); 974 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 975 if (rank==1) { 976 ierr = PetscPrintf(MPI_COMM_SELF," R %g, AP %g, Cseq %g, Cmpi %g = %g\n", eR,eAP,eCseq,eCmpi,eR+eAP+eCseq+eCmpi);CHKERRQ(ierr); 977 } 978 #endif 979 ptap->reuse = MAT_REUSE_MATRIX; 980 PetscFunctionReturn(0); 981 } 982 983 PetscErrorCode MatPtAPSymbolic_MPIAIJ_MPIAIJ_ptap(Mat A,Mat P,PetscReal fill,Mat *C) 984 { 985 PetscErrorCode ierr; 986 Mat Cmpi; 987 Mat_PtAPMPI *ptap; 988 PetscFreeSpaceList free_space=NULL,current_space=NULL; 989 Mat_MPIAIJ *a =(Mat_MPIAIJ*)A->data,*p=(Mat_MPIAIJ*)P->data,*c; 990 Mat_SeqAIJ *ad =(Mat_SeqAIJ*)(a->A)->data,*ao=(Mat_SeqAIJ*)(a->B)->data; 991 Mat_SeqAIJ *p_loc,*p_oth; 992 PetscInt *pi_loc,*pj_loc,*pi_oth,*pj_oth,*pdti,*pdtj,*poti,*potj,*ptJ; 993 PetscInt *adi=ad->i,*aj,*aoi=ao->i,nnz; 994 PetscInt *lnk,*owners_co,*coi,*coj,i,k,pnz,row; 995 PetscInt am=A->rmap->n,pN=P->cmap->N,pm=P->rmap->n,pn=P->cmap->n; 996 PetscBT lnkbt; 997 MPI_Comm comm; 998 PetscMPIInt size,rank,tagi,tagj,*len_si,*len_s,*len_ri,icompleted=0; 999 PetscInt **buf_rj,**buf_ri,**buf_ri_k; 1000 PetscInt len,proc,*dnz,*onz,*owners; 1001 PetscInt nzi,*pti,*ptj; 1002 PetscInt nrows,*buf_s,*buf_si,*buf_si_i,**nextrow,**nextci; 1003 MPI_Request *swaits,*rwaits; 1004 MPI_Status *sstatus,rstatus; 1005 Mat_Merge_SeqsToMPI *merge; 1006 PetscInt *api,*apj,*Jptr,apnz,*prmap=p->garray,pon,nspacedouble=0,j,ap_rmax=0; 1007 PetscReal afill=1.0,afill_tmp; 1008 1009 PetscFunctionBegin; 1010 ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr); 1011 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 1012 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 1013 1014 /* create struct Mat_PtAPMPI and attached it to C later */ 1015 ierr = PetscNew(&ptap);CHKERRQ(ierr); 1016 ierr = PetscNew(&merge);CHKERRQ(ierr); 1017 ptap->merge = merge; 1018 ptap->reuse = MAT_INITIAL_MATRIX; 1019 1020 /* get P_oth by taking rows of P (= non-zero cols of local A) from other processors */ 1021 ierr = MatGetBrowsOfAoCols_MPIAIJ(A,P,MAT_INITIAL_MATRIX,&ptap->startsj_s,&ptap->startsj_r,&ptap->bufa,&ptap->P_oth);CHKERRQ(ierr); 1022 1023 /* get P_loc by taking all local rows of P */ 1024 ierr = MatMPIAIJGetLocalMat(P,MAT_INITIAL_MATRIX,&ptap->P_loc);CHKERRQ(ierr); 1025 1026 p_loc = (Mat_SeqAIJ*)(ptap->P_loc)->data; 1027 p_oth = (Mat_SeqAIJ*)(ptap->P_oth)->data; 1028 pi_loc = p_loc->i; pj_loc = p_loc->j; 1029 pi_oth = p_oth->i; pj_oth = p_oth->j; 1030 1031 /* (1) compute symbolic AP = A_loc*P = A_diag*P_loc + A_off*P_oth (api,apj) */ 1032 /*--------------------------------------------------------------------------*/ 1033 ierr = PetscMalloc1(am+1,&api);CHKERRQ(ierr); 1034 api[0] = 0; 1035 1036 /* create and initialize a linked list */ 1037 ierr = PetscLLCondensedCreate(pN,pN,&lnk,&lnkbt);CHKERRQ(ierr); 1038 1039 /* Initial FreeSpace size is fill*(nnz(A) + nnz(P)) -OOM for ex56, np=8k on Intrepid! */ 1040 ierr = PetscFreeSpaceGet(PetscRealIntMultTruncate(fill,PetscIntSumTruncate(adi[am],PetscIntSumTruncate(aoi[am],pi_loc[pm]))),&free_space);CHKERRQ(ierr); 1041 current_space = free_space; 1042 1043 for (i=0; i<am; i++) { 1044 /* diagonal portion of A */ 1045 nzi = adi[i+1] - adi[i]; 1046 aj = ad->j + adi[i]; 1047 for (j=0; j<nzi; j++) { 1048 row = aj[j]; 1049 pnz = pi_loc[row+1] - pi_loc[row]; 1050 Jptr = pj_loc + pi_loc[row]; 1051 /* add non-zero cols of P into the sorted linked list lnk */ 1052 ierr = PetscLLCondensedAddSorted(pnz,Jptr,lnk,lnkbt);CHKERRQ(ierr); 1053 } 1054 /* off-diagonal portion of A */ 1055 nzi = aoi[i+1] - aoi[i]; 1056 aj = ao->j + aoi[i]; 1057 for (j=0; j<nzi; j++) { 1058 row = aj[j]; 1059 pnz = pi_oth[row+1] - pi_oth[row]; 1060 Jptr = pj_oth + pi_oth[row]; 1061 ierr = PetscLLCondensedAddSorted(pnz,Jptr,lnk,lnkbt);CHKERRQ(ierr); 1062 } 1063 apnz = lnk[0]; 1064 api[i+1] = api[i] + apnz; 1065 if (ap_rmax < apnz) ap_rmax = apnz; 1066 1067 /* if free space is not available, double the total space in the list */ 1068 if (current_space->local_remaining<apnz) { 1069 ierr = PetscFreeSpaceGet(PetscIntSumTruncate(apnz,current_space->total_array_size),¤t_space);CHKERRQ(ierr); 1070 nspacedouble++; 1071 } 1072 1073 /* Copy data into free space, then initialize lnk */ 1074 ierr = PetscLLCondensedClean(pN,apnz,current_space->array,lnk,lnkbt);CHKERRQ(ierr); 1075 1076 current_space->array += apnz; 1077 current_space->local_used += apnz; 1078 current_space->local_remaining -= apnz; 1079 } 1080 1081 /* Allocate space for apj, initialize apj, and */ 1082 /* destroy list of free space and other temporary array(s) */ 1083 ierr = PetscMalloc1(api[am]+1,&apj);CHKERRQ(ierr); 1084 ierr = PetscFreeSpaceContiguous(&free_space,apj);CHKERRQ(ierr); 1085 afill_tmp = (PetscReal)api[am]/(adi[am]+aoi[am]+pi_loc[pm]+1); 1086 if (afill_tmp > afill) afill = afill_tmp; 1087 1088 /* (2) determine symbolic Co=(p->B)^T*AP - send to others (coi,coj)*/ 1089 /*-----------------------------------------------------------------*/ 1090 ierr = MatGetSymbolicTranspose_SeqAIJ(p->B,&poti,&potj);CHKERRQ(ierr); 1091 1092 /* then, compute symbolic Co = (p->B)^T*AP */ 1093 pon = (p->B)->cmap->n; /* total num of rows to be sent to other processors 1094 >= (num of nonzero rows of C_seq) - pn */ 1095 ierr = PetscMalloc1(pon+1,&coi);CHKERRQ(ierr); 1096 coi[0] = 0; 1097 1098 /* set initial free space to be fill*(nnz(p->B) + nnz(AP)) */ 1099 nnz = PetscRealIntMultTruncate(fill,PetscIntSumTruncate(poti[pon],api[am])); 1100 ierr = PetscFreeSpaceGet(nnz,&free_space);CHKERRQ(ierr); 1101 current_space = free_space; 1102 1103 for (i=0; i<pon; i++) { 1104 pnz = poti[i+1] - poti[i]; 1105 ptJ = potj + poti[i]; 1106 for (j=0; j<pnz; j++) { 1107 row = ptJ[j]; /* row of AP == col of Pot */ 1108 apnz = api[row+1] - api[row]; 1109 Jptr = apj + api[row]; 1110 /* add non-zero cols of AP into the sorted linked list lnk */ 1111 ierr = PetscLLCondensedAddSorted(apnz,Jptr,lnk,lnkbt);CHKERRQ(ierr); 1112 } 1113 nnz = lnk[0]; 1114 1115 /* If free space is not available, double the total space in the list */ 1116 if (current_space->local_remaining<nnz) { 1117 ierr = PetscFreeSpaceGet(PetscIntSumTruncate(nnz,current_space->total_array_size),¤t_space);CHKERRQ(ierr); 1118 nspacedouble++; 1119 } 1120 1121 /* Copy data into free space, and zero out denserows */ 1122 ierr = PetscLLCondensedClean(pN,nnz,current_space->array,lnk,lnkbt);CHKERRQ(ierr); 1123 1124 current_space->array += nnz; 1125 current_space->local_used += nnz; 1126 current_space->local_remaining -= nnz; 1127 1128 coi[i+1] = coi[i] + nnz; 1129 } 1130 1131 ierr = PetscMalloc1(coi[pon],&coj);CHKERRQ(ierr); 1132 ierr = PetscFreeSpaceContiguous(&free_space,coj);CHKERRQ(ierr); 1133 afill_tmp = (PetscReal)coi[pon]/(poti[pon] + api[am]+1); 1134 if (afill_tmp > afill) afill = afill_tmp; 1135 ierr = MatRestoreSymbolicTranspose_SeqAIJ(p->B,&poti,&potj);CHKERRQ(ierr); 1136 1137 /* (3) send j-array (coj) of Co to other processors */ 1138 /*--------------------------------------------------*/ 1139 ierr = PetscCalloc1(size,&merge->len_s);CHKERRQ(ierr); 1140 len_s = merge->len_s; 1141 merge->nsend = 0; 1142 1143 1144 /* determine row ownership */ 1145 ierr = PetscLayoutCreate(comm,&merge->rowmap);CHKERRQ(ierr); 1146 merge->rowmap->n = pn; 1147 merge->rowmap->bs = 1; 1148 1149 ierr = PetscLayoutSetUp(merge->rowmap);CHKERRQ(ierr); 1150 owners = merge->rowmap->range; 1151 1152 /* determine the number of messages to send, their lengths */ 1153 ierr = PetscMalloc2(size,&len_si,size,&sstatus);CHKERRQ(ierr); 1154 ierr = PetscMemzero(len_si,size*sizeof(PetscMPIInt));CHKERRQ(ierr); 1155 ierr = PetscMalloc1(size+2,&owners_co);CHKERRQ(ierr); 1156 1157 proc = 0; 1158 for (i=0; i<pon; i++) { 1159 while (prmap[i] >= owners[proc+1]) proc++; 1160 len_si[proc]++; /* num of rows in Co(=Pt*AP) to be sent to [proc] */ 1161 len_s[proc] += coi[i+1] - coi[i]; /* num of nonzeros in Co to be sent to [proc] */ 1162 } 1163 1164 len = 0; /* max length of buf_si[], see (4) */ 1165 owners_co[0] = 0; 1166 for (proc=0; proc<size; proc++) { 1167 owners_co[proc+1] = owners_co[proc] + len_si[proc]; 1168 if (len_s[proc]) { 1169 merge->nsend++; 1170 len_si[proc] = 2*(len_si[proc] + 1); /* length of buf_si to be sent to [proc] */ 1171 len += len_si[proc]; 1172 } 1173 } 1174 1175 /* determine the number and length of messages to receive for coi and coj */ 1176 ierr = PetscGatherNumberOfMessages(comm,NULL,len_s,&merge->nrecv);CHKERRQ(ierr); 1177 ierr = PetscGatherMessageLengths2(comm,merge->nsend,merge->nrecv,len_s,len_si,&merge->id_r,&merge->len_r,&len_ri);CHKERRQ(ierr); 1178 1179 /* post the Irecv and Isend of coj */ 1180 ierr = PetscCommGetNewTag(comm,&tagj);CHKERRQ(ierr); 1181 ierr = PetscPostIrecvInt(comm,tagj,merge->nrecv,merge->id_r,merge->len_r,&buf_rj,&rwaits);CHKERRQ(ierr); 1182 ierr = PetscMalloc1(merge->nsend+1,&swaits);CHKERRQ(ierr); 1183 for (proc=0, k=0; proc<size; proc++) { 1184 if (!len_s[proc]) continue; 1185 i = owners_co[proc]; 1186 ierr = MPI_Isend(coj+coi[i],len_s[proc],MPIU_INT,proc,tagj,comm,swaits+k);CHKERRQ(ierr); 1187 k++; 1188 } 1189 1190 /* receives and sends of coj are complete */ 1191 for (i=0; i<merge->nrecv; i++) { 1192 ierr = MPI_Waitany(merge->nrecv,rwaits,&icompleted,&rstatus);CHKERRQ(ierr); 1193 } 1194 ierr = PetscFree(rwaits);CHKERRQ(ierr); 1195 if (merge->nsend) {ierr = MPI_Waitall(merge->nsend,swaits,sstatus);CHKERRQ(ierr);} 1196 1197 /* (4) send and recv coi */ 1198 /*-----------------------*/ 1199 ierr = PetscCommGetNewTag(comm,&tagi);CHKERRQ(ierr); 1200 ierr = PetscPostIrecvInt(comm,tagi,merge->nrecv,merge->id_r,len_ri,&buf_ri,&rwaits);CHKERRQ(ierr); 1201 ierr = PetscMalloc1(len+1,&buf_s);CHKERRQ(ierr); 1202 buf_si = buf_s; /* points to the beginning of k-th msg to be sent */ 1203 for (proc=0,k=0; proc<size; proc++) { 1204 if (!len_s[proc]) continue; 1205 /* form outgoing message for i-structure: 1206 buf_si[0]: nrows to be sent 1207 [1:nrows]: row index (global) 1208 [nrows+1:2*nrows+1]: i-structure index 1209 */ 1210 /*-------------------------------------------*/ 1211 nrows = len_si[proc]/2 - 1; /* num of rows in Co to be sent to [proc] */ 1212 buf_si_i = buf_si + nrows+1; 1213 buf_si[0] = nrows; 1214 buf_si_i[0] = 0; 1215 nrows = 0; 1216 for (i=owners_co[proc]; i<owners_co[proc+1]; i++) { 1217 nzi = coi[i+1] - coi[i]; 1218 buf_si_i[nrows+1] = buf_si_i[nrows] + nzi; /* i-structure */ 1219 buf_si[nrows+1] = prmap[i] -owners[proc]; /* local row index */ 1220 nrows++; 1221 } 1222 ierr = MPI_Isend(buf_si,len_si[proc],MPIU_INT,proc,tagi,comm,swaits+k);CHKERRQ(ierr); 1223 k++; 1224 buf_si += len_si[proc]; 1225 } 1226 i = merge->nrecv; 1227 while (i--) { 1228 ierr = MPI_Waitany(merge->nrecv,rwaits,&icompleted,&rstatus);CHKERRQ(ierr); 1229 } 1230 ierr = PetscFree(rwaits);CHKERRQ(ierr); 1231 if (merge->nsend) {ierr = MPI_Waitall(merge->nsend,swaits,sstatus);CHKERRQ(ierr);} 1232 1233 ierr = PetscFree2(len_si,sstatus);CHKERRQ(ierr); 1234 ierr = PetscFree(len_ri);CHKERRQ(ierr); 1235 ierr = PetscFree(swaits);CHKERRQ(ierr); 1236 ierr = PetscFree(buf_s);CHKERRQ(ierr); 1237 1238 /* (5) compute the local portion of C (mpi mat) */ 1239 /*----------------------------------------------*/ 1240 ierr = MatGetSymbolicTranspose_SeqAIJ(p->A,&pdti,&pdtj);CHKERRQ(ierr); 1241 1242 /* allocate pti array and free space for accumulating nonzero column info */ 1243 ierr = PetscMalloc1(pn+1,&pti);CHKERRQ(ierr); 1244 pti[0] = 0; 1245 1246 /* set initial free space to be fill*(nnz(P) + nnz(AP)) */ 1247 nnz = PetscRealIntMultTruncate(fill,PetscIntSumTruncate(pi_loc[pm],api[am])); 1248 ierr = PetscFreeSpaceGet(nnz,&free_space);CHKERRQ(ierr); 1249 current_space = free_space; 1250 1251 ierr = PetscMalloc3(merge->nrecv,&buf_ri_k,merge->nrecv,&nextrow,merge->nrecv,&nextci);CHKERRQ(ierr); 1252 for (k=0; k<merge->nrecv; k++) { 1253 buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */ 1254 nrows = *buf_ri_k[k]; 1255 nextrow[k] = buf_ri_k[k] + 1; /* next row number of k-th recved i-structure */ 1256 nextci[k] = buf_ri_k[k] + (nrows + 1); /* poins to the next i-structure of k-th recved i-structure */ 1257 } 1258 ierr = MatPreallocateInitialize(comm,pn,pn,dnz,onz);CHKERRQ(ierr); 1259 1260 for (i=0; i<pn; i++) { 1261 /* add pdt[i,:]*AP into lnk */ 1262 pnz = pdti[i+1] - pdti[i]; 1263 ptJ = pdtj + pdti[i]; 1264 for (j=0; j<pnz; j++) { 1265 row = ptJ[j]; /* row of AP == col of Pt */ 1266 apnz = api[row+1] - api[row]; 1267 Jptr = apj + api[row]; 1268 /* add non-zero cols of AP into the sorted linked list lnk */ 1269 ierr = PetscLLCondensedAddSorted(apnz,Jptr,lnk,lnkbt);CHKERRQ(ierr); 1270 } 1271 1272 /* add received col data into lnk */ 1273 for (k=0; k<merge->nrecv; k++) { /* k-th received message */ 1274 if (i == *nextrow[k]) { /* i-th row */ 1275 nzi = *(nextci[k]+1) - *nextci[k]; 1276 Jptr = buf_rj[k] + *nextci[k]; 1277 ierr = PetscLLCondensedAddSorted(nzi,Jptr,lnk,lnkbt);CHKERRQ(ierr); 1278 nextrow[k]++; nextci[k]++; 1279 } 1280 } 1281 nnz = lnk[0]; 1282 1283 /* if free space is not available, make more free space */ 1284 if (current_space->local_remaining<nnz) { 1285 ierr = PetscFreeSpaceGet(PetscIntSumTruncate(nnz,current_space->total_array_size),¤t_space);CHKERRQ(ierr); 1286 nspacedouble++; 1287 } 1288 /* copy data into free space, then initialize lnk */ 1289 ierr = PetscLLCondensedClean(pN,nnz,current_space->array,lnk,lnkbt);CHKERRQ(ierr); 1290 ierr = MatPreallocateSet(i+owners[rank],nnz,current_space->array,dnz,onz);CHKERRQ(ierr); 1291 1292 current_space->array += nnz; 1293 current_space->local_used += nnz; 1294 current_space->local_remaining -= nnz; 1295 1296 pti[i+1] = pti[i] + nnz; 1297 } 1298 ierr = MatRestoreSymbolicTranspose_SeqAIJ(p->A,&pdti,&pdtj);CHKERRQ(ierr); 1299 ierr = PetscFree3(buf_ri_k,nextrow,nextci);CHKERRQ(ierr); 1300 1301 ierr = PetscMalloc1(pti[pn]+1,&ptj);CHKERRQ(ierr); 1302 ierr = PetscFreeSpaceContiguous(&free_space,ptj);CHKERRQ(ierr); 1303 afill_tmp = (PetscReal)pti[pn]/(pi_loc[pm] + api[am]+1); 1304 if (afill_tmp > afill) afill = afill_tmp; 1305 ierr = PetscLLDestroy(lnk,lnkbt);CHKERRQ(ierr); 1306 1307 /* (6) create symbolic parallel matrix Cmpi */ 1308 /*------------------------------------------*/ 1309 ierr = MatCreate(comm,&Cmpi);CHKERRQ(ierr); 1310 ierr = MatSetSizes(Cmpi,pn,pn,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 1311 ierr = MatSetBlockSizes(Cmpi,PetscAbs(P->cmap->bs),PetscAbs(P->cmap->bs));CHKERRQ(ierr); 1312 ierr = MatSetType(Cmpi,MATMPIAIJ);CHKERRQ(ierr); 1313 ierr = MatMPIAIJSetPreallocation(Cmpi,0,dnz,0,onz);CHKERRQ(ierr); 1314 ierr = MatPreallocateFinalize(dnz,onz);CHKERRQ(ierr); 1315 1316 merge->bi = pti; /* Cseq->i */ 1317 merge->bj = ptj; /* Cseq->j */ 1318 merge->coi = coi; /* Co->i */ 1319 merge->coj = coj; /* Co->j */ 1320 merge->buf_ri = buf_ri; 1321 merge->buf_rj = buf_rj; 1322 merge->owners_co = owners_co; 1323 merge->destroy = Cmpi->ops->destroy; 1324 1325 /* attach the supporting struct to Cmpi for reuse */ 1326 c = (Mat_MPIAIJ*)Cmpi->data; 1327 c->ptap = ptap; 1328 ptap->api = api; 1329 ptap->apj = apj; 1330 ptap->duplicate = Cmpi->ops->duplicate; 1331 ptap->destroy = Cmpi->ops->destroy; 1332 1333 /* Cmpi is not ready for use - assembly will be done by MatPtAPNumeric() */ 1334 Cmpi->assembled = PETSC_FALSE; 1335 Cmpi->ops->destroy = MatDestroy_MPIAIJ_PtAP; 1336 Cmpi->ops->duplicate = MatDuplicate_MPIAIJ_MatPtAP; 1337 Cmpi->ops->ptapnumeric = MatPtAPNumeric_MPIAIJ_MPIAIJ_ptap; 1338 *C = Cmpi; 1339 1340 /* flag 'scalable' determines which implementations to be used: 1341 0: do dense axpy in MatPtAPNumeric() - fast, but requires storage of a nonscalable dense array apa; 1342 1: do sparse axpy in MatPtAPNumeric() - might slow, uses a sparse array apa */ 1343 /* set default scalable */ 1344 ptap->scalable = PETSC_FALSE; /* PETSC_TRUE; */ 1345 1346 ierr = PetscOptionsGetBool(((PetscObject)Cmpi)->options,((PetscObject)Cmpi)->prefix,"-matptap_scalable",&ptap->scalable,NULL);CHKERRQ(ierr); 1347 if (!ptap->scalable) { /* Do dense axpy */ 1348 ierr = PetscCalloc1(pN,&ptap->apa);CHKERRQ(ierr); 1349 } else { 1350 ierr = PetscCalloc1(ap_rmax+1,&ptap->apa);CHKERRQ(ierr); 1351 } 1352 1353 #if defined(PETSC_USE_INFO) 1354 if (pti[pn] != 0) { 1355 ierr = PetscInfo3(Cmpi,"Reallocs %D; Fill ratio: given %g needed %g.\n",nspacedouble,(double)fill,(double)afill);CHKERRQ(ierr); 1356 ierr = PetscInfo1(Cmpi,"Use MatPtAP(A,P,MatReuse,%g,&C) for best performance.\n",(double)afill);CHKERRQ(ierr); 1357 } else { 1358 ierr = PetscInfo(Cmpi,"Empty matrix product\n");CHKERRQ(ierr); 1359 } 1360 #endif 1361 PetscFunctionReturn(0); 1362 } 1363 1364 PetscErrorCode MatPtAPNumeric_MPIAIJ_MPIAIJ_ptap(Mat A,Mat P,Mat C) 1365 { 1366 PetscErrorCode ierr; 1367 Mat_MPIAIJ *a =(Mat_MPIAIJ*)A->data,*p=(Mat_MPIAIJ*)P->data,*c=(Mat_MPIAIJ*)C->data; 1368 Mat_SeqAIJ *ad=(Mat_SeqAIJ*)(a->A)->data,*ao=(Mat_SeqAIJ*)(a->B)->data; 1369 Mat_SeqAIJ *pd=(Mat_SeqAIJ*)(p->A)->data,*po=(Mat_SeqAIJ*)(p->B)->data; 1370 Mat_SeqAIJ *p_loc,*p_oth; 1371 Mat_PtAPMPI *ptap; 1372 Mat_Merge_SeqsToMPI *merge; 1373 PetscInt *adi=ad->i,*aoi=ao->i,*adj,*aoj,*apJ,nextp; 1374 PetscInt *pi_loc,*pj_loc,*pi_oth,*pj_oth,*pJ,*pj; 1375 PetscInt i,j,k,anz,pnz,apnz,nextap,row,*cj; 1376 MatScalar *ada,*aoa,*apa,*pa,*ca,*pa_loc,*pa_oth,valtmp; 1377 PetscInt am =A->rmap->n,cm=C->rmap->n,pon=(p->B)->cmap->n; 1378 MPI_Comm comm; 1379 PetscMPIInt size,rank,taga,*len_s; 1380 PetscInt *owners,proc,nrows,**buf_ri_k,**nextrow,**nextci; 1381 PetscInt **buf_ri,**buf_rj; 1382 PetscInt cnz=0,*bj_i,*bi,*bj,bnz,nextcj; /* bi,bj,ba: local array of C(mpi mat) */ 1383 MPI_Request *s_waits,*r_waits; 1384 MPI_Status *status; 1385 MatScalar **abuf_r,*ba_i,*pA,*coa,*ba; 1386 PetscInt *api,*apj,*coi,*coj; 1387 PetscInt *poJ=po->j,*pdJ=pd->j,pcstart=P->cmap->rstart,pcend=P->cmap->rend; 1388 PetscBool scalable; 1389 #if defined(PTAP_PROFILE) 1390 PetscLogDouble t0,t1,t2,eP,t3,t4,et2_AP=0.0,ePtAP=0.0,t2_0,t2_1,t2_2; 1391 #endif 1392 1393 PetscFunctionBegin; 1394 ierr = PetscObjectGetComm((PetscObject)C,&comm);CHKERRQ(ierr); 1395 #if defined(PTAP_PROFILE) 1396 ierr = PetscTime(&t0);CHKERRQ(ierr); 1397 #endif 1398 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 1399 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 1400 1401 ptap = c->ptap; 1402 if (!ptap) SETERRQ(PetscObjectComm((PetscObject)C),PETSC_ERR_ARG_INCOMP,"MatPtAP() has not been called to create matrix C yet, cannot use MAT_REUSE_MATRIX"); 1403 merge = ptap->merge; 1404 apa = ptap->apa; 1405 scalable = ptap->scalable; 1406 1407 /* 1) get P_oth = ptap->P_oth and P_loc = ptap->P_loc */ 1408 /*-----------------------------------------------------*/ 1409 if (ptap->reuse == MAT_INITIAL_MATRIX) { 1410 /* P_oth and P_loc are obtained in MatPtASymbolic(), skip calling MatGetBrowsOfAoCols() and MatMPIAIJGetLocalMat() */ 1411 ptap->reuse = MAT_REUSE_MATRIX; 1412 } else { /* update numerical values of P_oth and P_loc */ 1413 ierr = MatGetBrowsOfAoCols_MPIAIJ(A,P,MAT_REUSE_MATRIX,&ptap->startsj_s,&ptap->startsj_r,&ptap->bufa,&ptap->P_oth);CHKERRQ(ierr); 1414 ierr = MatMPIAIJGetLocalMat(P,MAT_REUSE_MATRIX,&ptap->P_loc);CHKERRQ(ierr); 1415 } 1416 #if defined(PTAP_PROFILE) 1417 ierr = PetscTime(&t1);CHKERRQ(ierr); 1418 eP = t1-t0; 1419 #endif 1420 /* 1421 printf("[%d] Ad: %d, %d; Ao: %d, %d; P_loc: %d, %d; P_oth %d, %d;\n",rank, 1422 a->A->rmap->N,a->A->cmap->N,a->B->rmap->N,a->B->cmap->N, 1423 ptap->P_loc->rmap->N,ptap->P_loc->cmap->N, 1424 ptap->P_oth->rmap->N,ptap->P_oth->cmap->N); 1425 */ 1426 1427 /* 2) compute numeric C_seq = P_loc^T*A_loc*P - dominating part */ 1428 /*--------------------------------------------------------------*/ 1429 /* get data from symbolic products */ 1430 p_loc = (Mat_SeqAIJ*)(ptap->P_loc)->data; 1431 p_oth = (Mat_SeqAIJ*)(ptap->P_oth)->data; 1432 pi_loc=p_loc->i; pj_loc=p_loc->j; pa_loc=p_loc->a; 1433 pi_oth=p_oth->i; pj_oth=p_oth->j; pa_oth=p_oth->a; 1434 1435 coi = merge->coi; coj = merge->coj; 1436 ierr = PetscCalloc1(coi[pon]+1,&coa);CHKERRQ(ierr); 1437 1438 bi = merge->bi; bj = merge->bj; 1439 owners = merge->rowmap->range; 1440 ierr = PetscCalloc1(bi[cm]+1,&ba);CHKERRQ(ierr); /* ba: Cseq->a */ 1441 1442 api = ptap->api; apj = ptap->apj; 1443 1444 if (!scalable) { /* Do dense axpy on apa (length of pN, stores A[i,:]*P) - nonscalable, but faster (could take 1/3 scalable time) */ 1445 ierr = PetscInfo(C,"Using non-scalable dense axpy\n");CHKERRQ(ierr); 1446 #if defined(PTAP_PROFILE) 1447 ierr = PetscTime(&t1);CHKERRQ(ierr); 1448 #endif 1449 for (i=0; i<am; i++) { 1450 #if defined(PTAP_PROFILE) 1451 ierr = PetscTime(&t2_0);CHKERRQ(ierr); 1452 #endif 1453 /* 2-a) form i-th sparse row of A_loc*P = Ad*P_loc + Ao*P_oth */ 1454 /*------------------------------------------------------------*/ 1455 apJ = apj + api[i]; 1456 1457 /* diagonal portion of A */ 1458 anz = adi[i+1] - adi[i]; 1459 adj = ad->j + adi[i]; 1460 ada = ad->a + adi[i]; 1461 for (j=0; j<anz; j++) { 1462 row = adj[j]; 1463 pnz = pi_loc[row+1] - pi_loc[row]; 1464 pj = pj_loc + pi_loc[row]; 1465 pa = pa_loc + pi_loc[row]; 1466 1467 /* perform dense axpy */ 1468 valtmp = ada[j]; 1469 for (k=0; k<pnz; k++) { 1470 apa[pj[k]] += valtmp*pa[k]; 1471 } 1472 ierr = PetscLogFlops(2.0*pnz);CHKERRQ(ierr); 1473 } 1474 1475 /* off-diagonal portion of A */ 1476 anz = aoi[i+1] - aoi[i]; 1477 aoj = ao->j + aoi[i]; 1478 aoa = ao->a + aoi[i]; 1479 for (j=0; j<anz; j++) { 1480 row = aoj[j]; 1481 pnz = pi_oth[row+1] - pi_oth[row]; 1482 pj = pj_oth + pi_oth[row]; 1483 pa = pa_oth + pi_oth[row]; 1484 1485 /* perform dense axpy */ 1486 valtmp = aoa[j]; 1487 for (k=0; k<pnz; k++) { 1488 apa[pj[k]] += valtmp*pa[k]; 1489 } 1490 ierr = PetscLogFlops(2.0*pnz);CHKERRQ(ierr); 1491 } 1492 #if defined(PTAP_PROFILE) 1493 ierr = PetscTime(&t2_1);CHKERRQ(ierr); 1494 et2_AP += t2_1 - t2_0; 1495 #endif 1496 1497 /* 2-b) Compute Cseq = P_loc[i,:]^T*AP[i,:] using outer product */ 1498 /*--------------------------------------------------------------*/ 1499 apnz = api[i+1] - api[i]; 1500 /* put the value into Co=(p->B)^T*AP (off-diagonal part, send to others) */ 1501 pnz = po->i[i+1] - po->i[i]; 1502 poJ = po->j + po->i[i]; 1503 pA = po->a + po->i[i]; 1504 for (j=0; j<pnz; j++) { 1505 row = poJ[j]; 1506 cnz = coi[row+1] - coi[row]; 1507 cj = coj + coi[row]; 1508 ca = coa + coi[row]; 1509 /* perform dense axpy */ 1510 valtmp = pA[j]; 1511 for (k=0; k<cnz; k++) { 1512 ca[k] += valtmp*apa[cj[k]]; 1513 } 1514 ierr = PetscLogFlops(2.0*cnz);CHKERRQ(ierr); 1515 } 1516 /* put the value into Cd (diagonal part) */ 1517 pnz = pd->i[i+1] - pd->i[i]; 1518 pdJ = pd->j + pd->i[i]; 1519 pA = pd->a + pd->i[i]; 1520 for (j=0; j<pnz; j++) { 1521 row = pdJ[j]; 1522 cnz = bi[row+1] - bi[row]; 1523 cj = bj + bi[row]; 1524 ca = ba + bi[row]; 1525 /* perform dense axpy */ 1526 valtmp = pA[j]; 1527 for (k=0; k<cnz; k++) { 1528 ca[k] += valtmp*apa[cj[k]]; 1529 } 1530 ierr = PetscLogFlops(2.0*cnz);CHKERRQ(ierr); 1531 } 1532 1533 /* zero the current row of A*P */ 1534 for (k=0; k<apnz; k++) apa[apJ[k]] = 0.0; 1535 #if defined(PTAP_PROFILE) 1536 ierr = PetscTime(&t2_2);CHKERRQ(ierr); 1537 ePtAP += t2_2 - t2_1; 1538 #endif 1539 } 1540 } else { /* Do sparse axpy on apa (length of ap_rmax, stores A[i,:]*P) - scalable, but slower */ 1541 ierr = PetscInfo(C,"Using scalable sparse axpy\n");CHKERRQ(ierr); 1542 /*-----------------------------------------------------------------------------------------*/ 1543 pA=pa_loc; 1544 for (i=0; i<am; i++) { 1545 #if defined(PTAP_PROFILE) 1546 ierr = PetscTime(&t2_0);CHKERRQ(ierr); 1547 #endif 1548 /* form i-th sparse row of A*P */ 1549 apnz = api[i+1] - api[i]; 1550 apJ = apj + api[i]; 1551 /* diagonal portion of A */ 1552 anz = adi[i+1] - adi[i]; 1553 adj = ad->j + adi[i]; 1554 ada = ad->a + adi[i]; 1555 for (j=0; j<anz; j++) { 1556 row = adj[j]; 1557 pnz = pi_loc[row+1] - pi_loc[row]; 1558 pj = pj_loc + pi_loc[row]; 1559 pa = pa_loc + pi_loc[row]; 1560 valtmp = ada[j]; 1561 nextp = 0; 1562 for (k=0; nextp<pnz; k++) { 1563 if (apJ[k] == pj[nextp]) { /* col of AP == col of P */ 1564 apa[k] += valtmp*pa[nextp++]; 1565 } 1566 } 1567 ierr = PetscLogFlops(2.0*pnz);CHKERRQ(ierr); 1568 } 1569 /* off-diagonal portion of A */ 1570 anz = aoi[i+1] - aoi[i]; 1571 aoj = ao->j + aoi[i]; 1572 aoa = ao->a + aoi[i]; 1573 for (j=0; j<anz; j++) { 1574 row = aoj[j]; 1575 pnz = pi_oth[row+1] - pi_oth[row]; 1576 pj = pj_oth + pi_oth[row]; 1577 pa = pa_oth + pi_oth[row]; 1578 valtmp = aoa[j]; 1579 nextp = 0; 1580 for (k=0; nextp<pnz; k++) { 1581 if (apJ[k] == pj[nextp]) { /* col of AP == col of P */ 1582 apa[k] += valtmp*pa[nextp++]; 1583 } 1584 } 1585 ierr = PetscLogFlops(2.0*pnz);CHKERRQ(ierr); 1586 } 1587 #if defined(PTAP_PROFILE) 1588 ierr = PetscTime(&t2_1);CHKERRQ(ierr); 1589 et2_AP += t2_1 - t2_0; 1590 #endif 1591 1592 /* 2-b) Compute Cseq = P_loc[i,:]^T*AP[i,:] using outer product */ 1593 /*--------------------------------------------------------------*/ 1594 pnz = pi_loc[i+1] - pi_loc[i]; 1595 pJ = pj_loc + pi_loc[i]; 1596 for (j=0; j<pnz; j++) { 1597 nextap = 0; 1598 row = pJ[j]; /* global index */ 1599 if (row < pcstart || row >=pcend) { /* put the value into Co */ 1600 row = *poJ; 1601 cj = coj + coi[row]; 1602 ca = coa + coi[row]; poJ++; 1603 } else { /* put the value into Cd */ 1604 row = *pdJ; 1605 cj = bj + bi[row]; 1606 ca = ba + bi[row]; pdJ++; 1607 } 1608 valtmp = pA[j]; 1609 for (k=0; nextap<apnz; k++) { 1610 if (cj[k]==apJ[nextap]) ca[k] += valtmp*apa[nextap++]; 1611 } 1612 ierr = PetscLogFlops(2.0*apnz);CHKERRQ(ierr); 1613 } 1614 pA += pnz; 1615 /* zero the current row info for A*P */ 1616 ierr = PetscMemzero(apa,apnz*sizeof(MatScalar));CHKERRQ(ierr); 1617 #if defined(PTAP_PROFILE) 1618 ierr = PetscTime(&t2_2);CHKERRQ(ierr); 1619 ePtAP += t2_2 - t2_1; 1620 #endif 1621 } 1622 } 1623 #if defined(PTAP_PROFILE) 1624 ierr = PetscTime(&t2);CHKERRQ(ierr); 1625 #endif 1626 1627 /* 3) send and recv matrix values coa */ 1628 /*------------------------------------*/ 1629 buf_ri = merge->buf_ri; 1630 buf_rj = merge->buf_rj; 1631 len_s = merge->len_s; 1632 ierr = PetscCommGetNewTag(comm,&taga);CHKERRQ(ierr); 1633 ierr = PetscPostIrecvScalar(comm,taga,merge->nrecv,merge->id_r,merge->len_r,&abuf_r,&r_waits);CHKERRQ(ierr); 1634 1635 ierr = PetscMalloc2(merge->nsend+1,&s_waits,size,&status);CHKERRQ(ierr); 1636 for (proc=0,k=0; proc<size; proc++) { 1637 if (!len_s[proc]) continue; 1638 i = merge->owners_co[proc]; 1639 ierr = MPI_Isend(coa+coi[i],len_s[proc],MPIU_MATSCALAR,proc,taga,comm,s_waits+k);CHKERRQ(ierr); 1640 k++; 1641 } 1642 if (merge->nrecv) {ierr = MPI_Waitall(merge->nrecv,r_waits,status);CHKERRQ(ierr);} 1643 if (merge->nsend) {ierr = MPI_Waitall(merge->nsend,s_waits,status);CHKERRQ(ierr);} 1644 1645 ierr = PetscFree2(s_waits,status);CHKERRQ(ierr); 1646 ierr = PetscFree(r_waits);CHKERRQ(ierr); 1647 ierr = PetscFree(coa);CHKERRQ(ierr); 1648 #if defined(PTAP_PROFILE) 1649 ierr = PetscTime(&t3);CHKERRQ(ierr); 1650 #endif 1651 1652 /* 4) insert local Cseq and received values into Cmpi */ 1653 /*------------------------------------------------------*/ 1654 ierr = PetscMalloc3(merge->nrecv,&buf_ri_k,merge->nrecv,&nextrow,merge->nrecv,&nextci);CHKERRQ(ierr); 1655 for (k=0; k<merge->nrecv; k++) { 1656 buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */ 1657 nrows = *(buf_ri_k[k]); 1658 nextrow[k] = buf_ri_k[k]+1; /* next row number of k-th recved i-structure */ 1659 nextci[k] = buf_ri_k[k] + (nrows + 1); /* poins to the next i-structure of k-th recved i-structure */ 1660 } 1661 1662 for (i=0; i<cm; i++) { 1663 row = owners[rank] + i; /* global row index of C_seq */ 1664 bj_i = bj + bi[i]; /* col indices of the i-th row of C */ 1665 ba_i = ba + bi[i]; 1666 bnz = bi[i+1] - bi[i]; 1667 /* add received vals into ba */ 1668 for (k=0; k<merge->nrecv; k++) { /* k-th received message */ 1669 /* i-th row */ 1670 if (i == *nextrow[k]) { 1671 cnz = *(nextci[k]+1) - *nextci[k]; 1672 cj = buf_rj[k] + *(nextci[k]); 1673 ca = abuf_r[k] + *(nextci[k]); 1674 nextcj = 0; 1675 for (j=0; nextcj<cnz; j++) { 1676 if (bj_i[j] == cj[nextcj]) { /* bcol == ccol */ 1677 ba_i[j] += ca[nextcj++]; 1678 } 1679 } 1680 nextrow[k]++; nextci[k]++; 1681 ierr = PetscLogFlops(2.0*cnz);CHKERRQ(ierr); 1682 } 1683 } 1684 ierr = MatSetValues(C,1,&row,bnz,bj_i,ba_i,INSERT_VALUES);CHKERRQ(ierr); 1685 } 1686 ierr = MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1687 ierr = MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1688 1689 ierr = PetscFree(ba);CHKERRQ(ierr); 1690 ierr = PetscFree(abuf_r[0]);CHKERRQ(ierr); 1691 ierr = PetscFree(abuf_r);CHKERRQ(ierr); 1692 ierr = PetscFree3(buf_ri_k,nextrow,nextci);CHKERRQ(ierr); 1693 #if defined(PTAP_PROFILE) 1694 ierr = PetscTime(&t4);CHKERRQ(ierr); 1695 if (rank==1) { 1696 ierr = PetscPrintf(MPI_COMM_SELF," [%d] PtAPNum %g/P + %g/PtAP( %g/A*P + %g/Pt*AP ) + %g/comm + %g/Cloc = %g\n\n",rank,eP,t2-t1,et2_AP,ePtAP,t3-t2,t4-t3,t4-t0);CHKERRQ(ierr); 1697 } 1698 #endif 1699 PetscFunctionReturn(0); 1700 } 1701