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 #undef __FUNCT__ 17 #define __FUNCT__ "MatDestroy_MPIAIJ_PtAP" 18 PetscErrorCode MatDestroy_MPIAIJ_PtAP(Mat A) 19 { 20 PetscErrorCode ierr; 21 Mat_MPIAIJ *a =(Mat_MPIAIJ*)A->data; 22 Mat_PtAPMPI *ptap=a->ptap; 23 24 PetscFunctionBegin; 25 if (ptap) { 26 Mat_Merge_SeqsToMPI *merge=ptap->merge; 27 ierr = PetscFree2(ptap->startsj_s,ptap->startsj_r);CHKERRQ(ierr); 28 ierr = PetscFree(ptap->bufa);CHKERRQ(ierr); 29 ierr = MatDestroy(&ptap->P_loc);CHKERRQ(ierr); 30 ierr = MatDestroy(&ptap->P_oth);CHKERRQ(ierr); 31 ierr = MatDestroy(&ptap->A_loc);CHKERRQ(ierr); /* used by MatTransposeMatMult() */ 32 33 ierr = MatDestroy(&ptap->Rd);CHKERRQ(ierr); 34 ierr = MatDestroy(&ptap->Ro);CHKERRQ(ierr); 35 ierr = MatDestroy(&ptap->AP_loc);CHKERRQ(ierr); 36 ierr = MatDestroy(&ptap->C_loc);CHKERRQ(ierr); 37 ierr = MatDestroy(&ptap->C_oth);CHKERRQ(ierr); 38 39 if (ptap->api) {ierr = PetscFree(ptap->api);CHKERRQ(ierr);} 40 if (ptap->apj) {ierr = PetscFree(ptap->apj);CHKERRQ(ierr);} 41 if (ptap->apa) {ierr = PetscFree(ptap->apa);CHKERRQ(ierr);} 42 if (merge) { 43 ierr = PetscFree(merge->id_r);CHKERRQ(ierr); 44 ierr = PetscFree(merge->len_s);CHKERRQ(ierr); 45 ierr = PetscFree(merge->len_r);CHKERRQ(ierr); 46 ierr = PetscFree(merge->bi);CHKERRQ(ierr); 47 ierr = PetscFree(merge->bj);CHKERRQ(ierr); 48 ierr = PetscFree(merge->buf_ri[0]);CHKERRQ(ierr); 49 ierr = PetscFree(merge->buf_ri);CHKERRQ(ierr); 50 ierr = PetscFree(merge->buf_rj[0]);CHKERRQ(ierr); 51 ierr = PetscFree(merge->buf_rj);CHKERRQ(ierr); 52 if (merge->coi) {ierr = PetscFree(merge->coi);CHKERRQ(ierr);} 53 if (merge->coj) {ierr = PetscFree(merge->coj);CHKERRQ(ierr);} 54 ierr = PetscFree(merge->owners_co);CHKERRQ(ierr); 55 ierr = PetscLayoutDestroy(&merge->rowmap);CHKERRQ(ierr); 56 ierr = merge->destroy(A);CHKERRQ(ierr); 57 ierr = PetscFree(ptap->merge);CHKERRQ(ierr); 58 } 59 ierr = PetscFree(ptap);CHKERRQ(ierr); 60 } 61 PetscFunctionReturn(0); 62 } 63 64 #undef __FUNCT__ 65 #define __FUNCT__ "MatDuplicate_MPIAIJ_MatPtAP" 66 PetscErrorCode MatDuplicate_MPIAIJ_MatPtAP(Mat A, MatDuplicateOption op, Mat *M) 67 { 68 PetscErrorCode ierr; 69 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 70 Mat_PtAPMPI *ptap = a->ptap; 71 Mat_Merge_SeqsToMPI *merge = ptap->merge; 72 73 PetscFunctionBegin; 74 ierr = (*merge->duplicate)(A,op,M);CHKERRQ(ierr); 75 76 (*M)->ops->destroy = merge->destroy; 77 (*M)->ops->duplicate = merge->duplicate; 78 PetscFunctionReturn(0); 79 } 80 81 #undef __FUNCT__ 82 #define __FUNCT__ "MatPtAP_MPIAIJ_MPIAIJ" 83 PetscErrorCode MatPtAP_MPIAIJ_MPIAIJ(Mat A,Mat P,MatReuse scall,PetscReal fill,Mat *C) 84 { 85 PetscErrorCode ierr; 86 PetscBool newalg=PETSC_FALSE; 87 88 PetscFunctionBegin; 89 ierr = PetscOptionsGetBool(NULL,"-matptap_new",&newalg,NULL);CHKERRQ(ierr); 90 if (scall == MAT_INITIAL_MATRIX) { 91 ierr = PetscLogEventBegin(MAT_PtAPSymbolic,A,P,0,0);CHKERRQ(ierr); 92 if (newalg) { 93 ierr = MatPtAPSymbolic_MPIAIJ_MPIAIJ_new(A,P,fill,C);CHKERRQ(ierr); 94 } else { 95 ierr = MatPtAPSymbolic_MPIAIJ_MPIAIJ(A,P,fill,C);CHKERRQ(ierr); 96 } 97 ierr = PetscLogEventEnd(MAT_PtAPSymbolic,A,P,0,0);CHKERRQ(ierr); 98 } 99 ierr = PetscLogEventBegin(MAT_PtAPNumeric,A,P,0,0);CHKERRQ(ierr); 100 if (newalg) { 101 ierr = MatPtAPNumeric_MPIAIJ_MPIAIJ_new(A,P,*C);CHKERRQ(ierr); 102 } else { 103 ierr = MatPtAPNumeric_MPIAIJ_MPIAIJ(A,P,*C);CHKERRQ(ierr); 104 } 105 ierr = PetscLogEventEnd(MAT_PtAPNumeric,A,P,0,0);CHKERRQ(ierr); 106 PetscFunctionReturn(0); 107 } 108 109 #undef __FUNCT__ 110 #define __FUNCT__ "MatPtAPSymbolic_MPIAIJ_MPIAIJ_new" 111 PetscErrorCode MatPtAPSymbolic_MPIAIJ_MPIAIJ_new(Mat A,Mat P,PetscReal fill,Mat *C) 112 { 113 PetscErrorCode ierr; 114 Mat_PtAPMPI *ptap; 115 Mat_MPIAIJ *p=(Mat_MPIAIJ*)P->data; 116 Mat AP; 117 Mat_MPIAIJ *c; 118 MPI_Comm comm; 119 PetscMPIInt size,rank; 120 PetscLogDouble t0,t1,t2,t3,t4; 121 122 PetscFunctionBegin; 123 ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr); 124 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 125 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 126 127 /* check if matrix local sizes are compatible -- MV! */ 128 if (A->rmap->rstart != P->rmap->rstart || A->rmap->rend != P->rmap->rend) { 129 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); 130 } 131 if (A->cmap->rstart != P->rmap->rstart || A->cmap->rend != P->rmap->rend) { 132 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); 133 } 134 135 //ierr = MatPtAPSymbolic_MPIAIJ_MPIAIJ(A,P,fill,C);CHKERRQ(ierr); //rm later !!! 136 //============================================================================= 137 Mat Cmpi; 138 PetscFreeSpaceList free_space=NULL,current_space=NULL; 139 Mat_SeqAIJ *p_loc; 140 PetscInt *pi_loc; 141 PetscInt *pdti,*pdtj,*poti,*potj,*ptJ,nnz; 142 PetscInt *lnk,*owners_co,*coi,*coj,i,k,pnz,row; 143 PetscInt am=A->rmap->n,pN=P->cmap->N,pm=P->rmap->n,pn=P->cmap->n; 144 PetscBT lnkbt; 145 PetscMPIInt tagi,tagj,*len_si,*len_s,*len_ri,icompleted=0; 146 PetscInt **buf_rj,**buf_ri,**buf_ri_k; 147 PetscInt len,proc,*dnz,*onz,*owners; 148 PetscInt nzi,*pti,*ptj; 149 PetscInt nrows,*buf_s,*buf_si,*buf_si_i,**nextrow,**nextci; 150 MPI_Request *swaits,*rwaits; 151 MPI_Status *sstatus,rstatus; 152 Mat_Merge_SeqsToMPI *merge; 153 PetscInt *api,*apj,*Jptr,apnz,*prmap=p->garray,pon,nspacedouble=0,j,ap_rmax=0; 154 PetscReal afill=1.0,afill_tmp; 155 PetscInt rmax; 156 157 ierr = PetscTime(&t0);CHKERRQ(ierr); 158 /* create struct Mat_PtAPMPI and attached it to C later */ 159 ierr = PetscNew(&ptap);CHKERRQ(ierr); 160 ierr = PetscNew(&merge);CHKERRQ(ierr); 161 ptap->merge = merge; 162 ptap->reuse = MAT_INITIAL_MATRIX; 163 164 /* get P_oth by taking rows of P (= non-zero cols of local A) from other processors */ 165 ierr = MatGetBrowsOfAoCols_MPIAIJ(A,P,MAT_INITIAL_MATRIX,&ptap->startsj_s,&ptap->startsj_r,&ptap->bufa,&ptap->P_oth);CHKERRQ(ierr); 166 167 /* get P_loc by taking all local rows of P */ 168 ierr = MatMPIAIJGetLocalMat(P,MAT_INITIAL_MATRIX,&ptap->P_loc);CHKERRQ(ierr); 169 170 ptap->reuse = MAT_INITIAL_MATRIX; 171 172 /* (1) compute symbolic AP = A*P, then get AP_loc */ 173 /*--------------------------------------------------------------------------*/ 174 ierr = MatTranspose_SeqAIJ(p->A,MAT_INITIAL_MATRIX,&ptap->Rd);CHKERRQ(ierr); 175 ierr = MatTranspose_SeqAIJ(p->B,MAT_INITIAL_MATRIX,&ptap->Ro);CHKERRQ(ierr); 176 177 ierr = MatMatMult(A,P,MAT_INITIAL_MATRIX,2.0,&AP);CHKERRQ(ierr); 178 ierr = MatMPIAIJGetLocalMat(AP,MAT_INITIAL_MATRIX,&ptap->AP_loc);CHKERRQ(ierr); 179 ierr = MatDestroy(&AP);CHKERRQ(ierr); 180 181 /* (2) compute C_loc=Rd*AP_loc, Co=Ro*AP_loc */ 182 ierr = MatMatMult_SeqAIJ_SeqAIJ(ptap->Rd,ptap->AP_loc,MAT_INITIAL_MATRIX,2.0,&ptap->C_loc);CHKERRQ(ierr); 183 ierr = MatMatMult_SeqAIJ_SeqAIJ(ptap->Ro,ptap->AP_loc,MAT_INITIAL_MATRIX,2.0,&ptap->C_oth);CHKERRQ(ierr); 184 ierr = PetscTime(&t1);CHKERRQ(ierr); 185 186 p_loc = (Mat_SeqAIJ*)(ptap->P_loc)->data; 187 pi_loc = p_loc->i; 188 189 Mat_SeqAIJ *ap = (Mat_SeqAIJ*)ptap->AP_loc->data; 190 Mat_SeqAIJ *c_oth = (Mat_SeqAIJ*)ptap->C_oth->data; 191 api = ap->i; apj = ap->j; 192 193 /* (2) determine symbolic Co=(p->B)^T*AP - send to others (coi,coj)*/ 194 /*-----------------------------------------------------------------*/ 195 coi = c_oth->i; coj = c_oth->j; 196 197 Mat_SeqAIJ *po = (Mat_SeqAIJ *)ptap->Ro->data; 198 poti = po->i; potj = po->j; 199 200 /* then, compute symbolic Co = (p->B)^T*AP */ 201 pon = (p->B)->cmap->n; /* total num of rows to be sent to other processors 202 >= (num of nonzero rows of C_seq) - pn */ 203 204 /* create and initialize a linked list */ 205 ierr = PetscLLCondensedCreate(pN,pN,&lnk,&lnkbt);CHKERRQ(ierr); 206 207 afill_tmp = (PetscReal)coi[pon]/(poti[pon] + api[am]+1); 208 if (afill_tmp > afill) afill = afill_tmp; 209 210 /* (3) send j-array (coj) of Co to other processors */ 211 /*--------------------------------------------------*/ 212 ierr = PetscCalloc1(size,&merge->len_s);CHKERRQ(ierr); 213 len_s = merge->len_s; 214 merge->nsend = 0; 215 216 /* determine row ownership */ 217 ierr = PetscLayoutCreate(comm,&merge->rowmap);CHKERRQ(ierr); 218 merge->rowmap->n = pn; 219 merge->rowmap->bs = 1; 220 221 ierr = PetscLayoutSetUp(merge->rowmap);CHKERRQ(ierr); 222 owners = merge->rowmap->range; 223 224 /* determine the number of messages to send, their lengths */ 225 ierr = PetscMalloc2(size,&len_si,size,&sstatus);CHKERRQ(ierr); 226 ierr = PetscMemzero(len_si,size*sizeof(PetscMPIInt));CHKERRQ(ierr); 227 ierr = PetscMalloc1(size+2,&owners_co);CHKERRQ(ierr); 228 229 proc = 0; 230 for (i=0; i<pon; i++) { 231 while (prmap[i] >= owners[proc+1]) proc++; 232 len_si[proc]++; /* num of rows in Co(=Pt*AP) to be sent to [proc] */ 233 len_s[proc] += coi[i+1] - coi[i]; /* num of nonzeros in Co to be sent to [proc] */ 234 } 235 236 len = 0; /* max length of buf_si[], see (4) */ 237 owners_co[0] = 0; 238 for (proc=0; proc<size; proc++) { 239 owners_co[proc+1] = owners_co[proc] + len_si[proc]; 240 if (len_s[proc]) { 241 merge->nsend++; 242 len_si[proc] = 2*(len_si[proc] + 1); /* length of buf_si to be sent to [proc] */ 243 len += len_si[proc]; 244 } 245 } 246 247 /* determine the number and length of messages to receive for coi and coj */ 248 ierr = PetscGatherNumberOfMessages(comm,NULL,len_s,&merge->nrecv);CHKERRQ(ierr); 249 ierr = PetscGatherMessageLengths2(comm,merge->nsend,merge->nrecv,len_s,len_si,&merge->id_r,&merge->len_r,&len_ri);CHKERRQ(ierr); 250 251 /* post the Irecv and Isend of coj */ 252 ierr = PetscCommGetNewTag(comm,&tagj);CHKERRQ(ierr); 253 ierr = PetscPostIrecvInt(comm,tagj,merge->nrecv,merge->id_r,merge->len_r,&buf_rj,&rwaits);CHKERRQ(ierr); 254 ierr = PetscMalloc1(merge->nsend+1,&swaits);CHKERRQ(ierr); 255 for (proc=0, k=0; proc<size; proc++) { 256 if (!len_s[proc]) continue; 257 i = owners_co[proc]; 258 ierr = MPI_Isend(coj+coi[i],len_s[proc],MPIU_INT,proc,tagj,comm,swaits+k);CHKERRQ(ierr); 259 k++; 260 } 261 262 /* receives and sends of coj are complete */ 263 for (i=0; i<merge->nrecv; i++) { 264 ierr = MPI_Waitany(merge->nrecv,rwaits,&icompleted,&rstatus);CHKERRQ(ierr); 265 } 266 ierr = PetscFree(rwaits);CHKERRQ(ierr); 267 if (merge->nsend) {ierr = MPI_Waitall(merge->nsend,swaits,sstatus);CHKERRQ(ierr);} 268 269 /* (4) send and recv coi */ 270 /*-----------------------*/ 271 ierr = PetscCommGetNewTag(comm,&tagi);CHKERRQ(ierr); 272 ierr = PetscPostIrecvInt(comm,tagi,merge->nrecv,merge->id_r,len_ri,&buf_ri,&rwaits);CHKERRQ(ierr); 273 ierr = PetscMalloc1(len+1,&buf_s);CHKERRQ(ierr); 274 buf_si = buf_s; /* points to the beginning of k-th msg to be sent */ 275 for (proc=0,k=0; proc<size; proc++) { 276 if (!len_s[proc]) continue; 277 /* form outgoing message for i-structure: 278 buf_si[0]: nrows to be sent 279 [1:nrows]: row index (global) 280 [nrows+1:2*nrows+1]: i-structure index 281 */ 282 /*-------------------------------------------*/ 283 nrows = len_si[proc]/2 - 1; /* num of rows in Co to be sent to [proc] */ 284 buf_si_i = buf_si + nrows+1; 285 buf_si[0] = nrows; 286 buf_si_i[0] = 0; 287 nrows = 0; 288 for (i=owners_co[proc]; i<owners_co[proc+1]; i++) { 289 nzi = coi[i+1] - coi[i]; 290 buf_si_i[nrows+1] = buf_si_i[nrows] + nzi; /* i-structure */ 291 buf_si[nrows+1] = prmap[i] -owners[proc]; /* local row index */ 292 nrows++; 293 } 294 ierr = MPI_Isend(buf_si,len_si[proc],MPIU_INT,proc,tagi,comm,swaits+k);CHKERRQ(ierr); 295 k++; 296 buf_si += len_si[proc]; 297 } 298 i = merge->nrecv; 299 while (i--) { 300 ierr = MPI_Waitany(merge->nrecv,rwaits,&icompleted,&rstatus);CHKERRQ(ierr); 301 } 302 ierr = PetscFree(rwaits);CHKERRQ(ierr); 303 if (merge->nsend) {ierr = MPI_Waitall(merge->nsend,swaits,sstatus);CHKERRQ(ierr);} 304 305 ierr = PetscFree2(len_si,sstatus);CHKERRQ(ierr); 306 ierr = PetscFree(len_ri);CHKERRQ(ierr); 307 ierr = PetscFree(swaits);CHKERRQ(ierr); 308 ierr = PetscFree(buf_s);CHKERRQ(ierr); 309 310 ierr = PetscTime(&t2);CHKERRQ(ierr); 311 312 //ierr = MPI_Barrier(comm); 313 314 /* (5) compute the local portion of C (mpi mat) */ 315 /*----------------------------------------------*/ 316 317 Mat_SeqAIJ *pd = (Mat_SeqAIJ *)ptap->Rd->data; 318 pdti = pd->i; pdtj = pd->j; 319 320 /* allocate pti array and free space for accumulating nonzero column info */ 321 ierr = PetscMalloc1(pn+1,&pti);CHKERRQ(ierr); 322 pti[0] = 0; 323 324 /* set initial free space to be fill*(nnz(P) + nnz(AP)) */ 325 nnz = fill*(pi_loc[pm] + api[am]); 326 ierr = PetscFreeSpaceGet(nnz,&free_space);CHKERRQ(ierr); 327 current_space = free_space; 328 329 ierr = PetscMalloc3(merge->nrecv,&buf_ri_k,merge->nrecv,&nextrow,merge->nrecv,&nextci);CHKERRQ(ierr); 330 for (k=0; k<merge->nrecv; k++) { 331 buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */ 332 nrows = *buf_ri_k[k]; 333 nextrow[k] = buf_ri_k[k] + 1; /* next row number of k-th recved i-structure */ 334 nextci[k] = buf_ri_k[k] + (nrows + 1); /* poins to the next i-structure of k-th recved i-structure */ 335 } 336 ierr = MatPreallocateInitialize(comm,pn,pn,dnz,onz);CHKERRQ(ierr); 337 rmax = 0; 338 for (i=0; i<pn; i++) { 339 /* add pdt[i,:]*AP into lnk */ 340 pnz = pdti[i+1] - pdti[i]; 341 ptJ = pdtj + pdti[i]; 342 for (j=0; j<pnz; j++) { 343 row = ptJ[j]; /* row of AP == col of Pt */ 344 apnz = api[row+1] - api[row]; 345 Jptr = apj + api[row]; 346 /* add non-zero cols of AP into the sorted linked list lnk */ 347 ierr = PetscLLCondensedAddSorted(apnz,Jptr,lnk,lnkbt);CHKERRQ(ierr); 348 } 349 350 /* add received col data into lnk */ 351 for (k=0; k<merge->nrecv; k++) { /* k-th received message */ 352 if (i == *nextrow[k]) { /* i-th row */ 353 nzi = *(nextci[k]+1) - *nextci[k]; 354 Jptr = buf_rj[k] + *nextci[k]; 355 ierr = PetscLLCondensedAddSorted(nzi,Jptr,lnk,lnkbt);CHKERRQ(ierr); 356 nextrow[k]++; nextci[k]++; 357 } 358 } 359 nnz = lnk[0]; 360 361 /* if free space is not available, make more free space */ 362 if (current_space->local_remaining<nnz) { 363 ierr = PetscFreeSpaceGet(nnz+current_space->total_array_size,¤t_space);CHKERRQ(ierr); 364 nspacedouble++; 365 } 366 /* copy data into free space, then initialize lnk */ 367 ierr = PetscLLCondensedClean(pN,nnz,current_space->array,lnk,lnkbt);CHKERRQ(ierr); 368 ierr = MatPreallocateSet(i+owners[rank],nnz,current_space->array,dnz,onz);CHKERRQ(ierr); 369 370 current_space->array += nnz; 371 current_space->local_used += nnz; 372 current_space->local_remaining -= nnz; 373 374 pti[i+1] = pti[i] + nnz; 375 if (nnz > rmax) rmax = nnz; 376 } 377 ierr = PetscFree3(buf_ri_k,nextrow,nextci);CHKERRQ(ierr); 378 379 ierr = PetscMalloc1(pti[pn]+1,&ptj);CHKERRQ(ierr); 380 ierr = PetscFreeSpaceContiguous(&free_space,ptj);CHKERRQ(ierr); 381 afill_tmp = (PetscReal)pti[pn]/(pi_loc[pm] + api[am]+1); 382 if (afill_tmp > afill) afill = afill_tmp; 383 ierr = PetscLLDestroy(lnk,lnkbt);CHKERRQ(ierr); 384 385 ierr = PetscTime(&t3);CHKERRQ(ierr); 386 387 /* (6) create symbolic parallel matrix Cmpi */ 388 /*------------------------------------------*/ 389 ierr = MatCreate(comm,&Cmpi);CHKERRQ(ierr); 390 ierr = MatSetSizes(Cmpi,pn,pn,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 391 ierr = MatSetBlockSizes(Cmpi,PetscAbs(P->cmap->bs),PetscAbs(P->cmap->bs));CHKERRQ(ierr); 392 ierr = MatSetType(Cmpi,MATMPIAIJ);CHKERRQ(ierr); 393 ierr = MatMPIAIJSetPreallocation(Cmpi,0,dnz,0,onz);CHKERRQ(ierr); 394 ierr = MatPreallocateFinalize(dnz,onz);CHKERRQ(ierr); 395 396 merge->bi = pti; /* Cseq->i */ 397 merge->bj = ptj; /* Cseq->j */ 398 merge->coi = NULL; //merge->coi = coi; /* Co->i */ 399 merge->coj = NULL; //merge->coj = coj; /* Co->j */ 400 merge->buf_ri = buf_ri; 401 merge->buf_rj = buf_rj; 402 merge->owners_co = owners_co; 403 merge->destroy = Cmpi->ops->destroy; 404 merge->duplicate = Cmpi->ops->duplicate; 405 406 /* Cmpi is not ready for use - assembly will be done by MatPtAPNumeric() */ 407 Cmpi->assembled = PETSC_FALSE; 408 Cmpi->ops->destroy = MatDestroy_MPIAIJ_PtAP; 409 Cmpi->ops->duplicate = MatDuplicate_MPIAIJ_MatPtAP; 410 Cmpi->ops->ptapnumeric = MatPtAPNumeric_MPIAIJ_MPIAIJ_new; 411 412 /* attach the supporting struct to Cmpi for reuse */ 413 c = (Mat_MPIAIJ*)Cmpi->data; 414 c->ptap = ptap; 415 ptap->api = NULL; 416 ptap->apj = NULL; 417 ptap->rmax = ap_rmax; 418 *C = Cmpi; 419 420 /* flag 'scalable' determines which implementations to be used: 421 0: do dense axpy in MatPtAPNumeric() - fast, but requires storage of a nonscalable dense array apa; 422 1: do sparse axpy in MatPtAPNumeric() - might slow, uses a sparse array apa */ 423 /* set default scalable */ 424 ptap->scalable = PETSC_FALSE; //PETSC_TRUE; 425 426 ierr = PetscOptionsGetBool(((PetscObject)(*C))->prefix,"-matptap_scalable",&ptap->scalable,NULL);CHKERRQ(ierr); 427 if (!ptap->scalable) { /* Do dense axpy */ 428 ierr = PetscCalloc1(P->cmap->N,&ptap->apa);CHKERRQ(ierr); 429 } else { 430 //ierr = PetscCalloc1(ap_rmax+1,&ptap->apa);CHKERRQ(ierr); 431 } 432 433 ierr = PetscTime(&t4);CHKERRQ(ierr); 434 if (rank == 1) printf("PtAPSym: %g + %g + %g = %g\n",t1-t0,t2-t1,t3-t2,t4-t3); 435 PetscFunctionReturn(0); 436 } 437 438 #undef __FUNCT__ 439 #define __FUNCT__ "MatPtAPNumeric_MPIAIJ_MPIAIJ_new" 440 PetscErrorCode MatPtAPNumeric_MPIAIJ_MPIAIJ_new(Mat A,Mat P,Mat C) 441 { 442 PetscErrorCode ierr; 443 Mat_MPIAIJ *a=(Mat_MPIAIJ*)A->data,*p=(Mat_MPIAIJ*)P->data,*c=(Mat_MPIAIJ*)C->data; 444 Mat_SeqAIJ *ad=(Mat_SeqAIJ*)(a->A)->data,*ao=(Mat_SeqAIJ*)(a->B)->data; 445 Mat_PtAPMPI *ptap = c->ptap; 446 Mat AP_loc,C_loc,C_oth; 447 PetscInt i,rstart,rend,cm,ncols,row; 448 PetscMPIInt rank; 449 MPI_Comm comm; 450 const PetscInt *cols; 451 const PetscScalar *vals; 452 PetscLogDouble t0,t1,t2,t3,t4,eR,eAP,eCseq,eCmpi; 453 454 PetscFunctionBegin; 455 ierr = PetscObjectGetComm((PetscObject)C,&comm);CHKERRQ(ierr); 456 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 457 458 ierr = MatZeroEntries(C);CHKERRQ(ierr); 459 460 /* 1) get R = Pd^T,Ro = Po^T */ 461 ierr = PetscTime(&t0);CHKERRQ(ierr); 462 if (ptap->reuse == MAT_REUSE_MATRIX) { 463 ierr = MatTranspose_SeqAIJ(p->A,MAT_REUSE_MATRIX,&ptap->Rd);CHKERRQ(ierr); 464 ierr = MatTranspose_SeqAIJ(p->B,MAT_REUSE_MATRIX,&ptap->Ro);CHKERRQ(ierr); 465 } 466 ierr = PetscTime(&t1);CHKERRQ(ierr); 467 eR = t1 - t0; 468 469 /* 2) get AP_loc */ 470 AP_loc = ptap->AP_loc; 471 Mat_SeqAIJ *ap=(Mat_SeqAIJ*)AP_loc->data; 472 473 /* 2-1) get P_oth = ptap->P_oth and P_loc = ptap->P_loc */ 474 /*-----------------------------------------------------*/ 475 if (ptap->reuse == MAT_REUSE_MATRIX) { 476 /* P_oth and P_loc are obtained in MatPtASymbolic() when reuse == MAT_INITIAL_MATRIX */ 477 ierr = MatGetBrowsOfAoCols_MPIAIJ(A,P,MAT_REUSE_MATRIX,&ptap->startsj_s,&ptap->startsj_r,&ptap->bufa,&ptap->P_oth);CHKERRQ(ierr); 478 ierr = MatMPIAIJGetLocalMat(P,MAT_REUSE_MATRIX,&ptap->P_loc);CHKERRQ(ierr); 479 } 480 481 /* 2-2) compute numeric C_seq = P_loc^T*A_loc*P - dominating part */ 482 /*--------------------------------------------------------------*/ 483 /* get data from symbolic products */ 484 Mat_SeqAIJ *p_loc = (Mat_SeqAIJ*)(ptap->P_loc)->data; 485 Mat_SeqAIJ *p_oth = (Mat_SeqAIJ*)(ptap->P_oth)->data; 486 PetscInt *api,*apj,am = A->rmap->n,j,col,apnz; 487 PetscScalar *apa = ptap->apa; 488 if (ptap->reuse == MAT_REUSE_MATRIX) { 489 api = ap->i; 490 apj = ap->j; 491 for (i=0; i<am; i++) { 492 /* AP[i,:] = A[i,:]*P = Ad*P_loc Ao*P_oth */ 493 AProw_nonscalable(i,ad,ao,p_loc,p_oth,apa); 494 apnz = api[i+1] - api[i]; 495 for (j=0; j<apnz; j++) { 496 col = apj[j+api[i]]; 497 ap->a[j+ap->i[i]] = apa[col]; 498 apa[col] = 0.0; 499 } 500 } 501 } 502 503 ierr = PetscTime(&t2);CHKERRQ(ierr); 504 eAP = t2 - t1; 505 506 /* 3) C_loc = R*AP_loc, Co = Ro*AP_loc */ 507 if (ptap->reuse == MAT_REUSE_MATRIX) { 508 ierr = MatMatMult_SeqAIJ_SeqAIJ(ptap->Rd,AP_loc,MAT_REUSE_MATRIX,2.0,&ptap->C_loc);CHKERRQ(ierr); 509 ierr = MatMatMult_SeqAIJ_SeqAIJ(ptap->Ro,AP_loc,MAT_REUSE_MATRIX,2.0,&ptap->C_oth);CHKERRQ(ierr); 510 } 511 C_loc = ptap->C_loc; 512 C_oth = ptap->C_oth; 513 //printf("[%d] Co %d, %d\n", rank,Co->rmap->N,Co->cmap->N); 514 ierr = PetscTime(&t3);CHKERRQ(ierr); 515 eCseq = t3 - t2; 516 517 /* add C_loc and Co to to C */ 518 ierr = MatGetOwnershipRange(C,&rstart,&rend);CHKERRQ(ierr); 519 520 /* C_loc -> C */ 521 cm = C_loc->rmap->N; 522 Mat_SeqAIJ *c_seq; 523 c_seq = (Mat_SeqAIJ*)C_loc->data; 524 for (i=0; i<cm; i++) { 525 ncols = c_seq->i[i+1] - c_seq->i[i]; 526 row = rstart + i; 527 cols = c_seq->j + c_seq->i[i]; 528 vals = c_seq->a + c_seq->i[i]; 529 ierr = MatSetValues(C,1,&row,ncols,cols,vals,ADD_VALUES);CHKERRQ(ierr); 530 } 531 532 /* Co -> C, off-processor part */ 533 //printf("[%d] p->B %d, %d\n",rank,p->B->rmap->N,p->B->cmap->N); 534 cm = C_oth->rmap->N; 535 c_seq = (Mat_SeqAIJ*)C_oth->data; 536 for (i=0; i<cm; i++) { 537 ncols = c_seq->i[i+1] - c_seq->i[i]; 538 row = p->garray[i]; 539 cols = c_seq->j + c_seq->i[i]; 540 vals = c_seq->a + c_seq->i[i]; 541 //printf("[%d] row[%d] = %d\n",rank,i,row); 542 ierr = MatSetValues(C,1,&row,ncols,cols,vals,ADD_VALUES);CHKERRQ(ierr); 543 } 544 ierr = MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 545 ierr = MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 546 ierr = PetscTime(&t4);CHKERRQ(ierr); 547 eCmpi = t4 - t3; 548 549 if (rank==1) { 550 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); 551 } 552 ptap->reuse = MAT_REUSE_MATRIX; 553 PetscFunctionReturn(0); 554 } 555 556 #undef __FUNCT__ 557 #define __FUNCT__ "MatPtAPSymbolic_MPIAIJ_MPIAIJ" 558 PetscErrorCode MatPtAPSymbolic_MPIAIJ_MPIAIJ(Mat A,Mat P,PetscReal fill,Mat *C) 559 { 560 PetscErrorCode ierr; 561 Mat Cmpi; 562 Mat_PtAPMPI *ptap; 563 PetscFreeSpaceList free_space=NULL,current_space=NULL; 564 Mat_MPIAIJ *a =(Mat_MPIAIJ*)A->data,*p=(Mat_MPIAIJ*)P->data,*c; 565 Mat_SeqAIJ *ad =(Mat_SeqAIJ*)(a->A)->data,*ao=(Mat_SeqAIJ*)(a->B)->data; 566 Mat_SeqAIJ *p_loc,*p_oth; 567 PetscInt *pi_loc,*pj_loc,*pi_oth,*pj_oth,*pdti,*pdtj,*poti,*potj,*ptJ; 568 PetscInt *adi=ad->i,*aj,*aoi=ao->i,nnz; 569 PetscInt *lnk,*owners_co,*coi,*coj,i,k,pnz,row; 570 PetscInt am=A->rmap->n,pN=P->cmap->N,pm=P->rmap->n,pn=P->cmap->n; 571 PetscBT lnkbt; 572 MPI_Comm comm; 573 PetscMPIInt size,rank,tagi,tagj,*len_si,*len_s,*len_ri,icompleted=0; 574 PetscInt **buf_rj,**buf_ri,**buf_ri_k; 575 PetscInt len,proc,*dnz,*onz,*owners; 576 PetscInt nzi,*pti,*ptj; 577 PetscInt nrows,*buf_s,*buf_si,*buf_si_i,**nextrow,**nextci; 578 MPI_Request *swaits,*rwaits; 579 MPI_Status *sstatus,rstatus; 580 Mat_Merge_SeqsToMPI *merge; 581 PetscInt *api,*apj,*Jptr,apnz,*prmap=p->garray,pon,nspacedouble=0,j,ap_rmax=0; 582 PetscReal afill=1.0,afill_tmp; 583 PetscInt rmax; 584 585 PetscFunctionBegin; 586 ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr); 587 588 /* check if matrix local sizes are compatible */ 589 if (A->rmap->rstart != P->rmap->rstart || A->rmap->rend != P->rmap->rend) { 590 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); 591 } 592 if (A->cmap->rstart != P->rmap->rstart || A->cmap->rend != P->rmap->rend) { 593 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); 594 } 595 596 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 597 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 598 599 /* create struct Mat_PtAPMPI and attached it to C later */ 600 ierr = PetscNew(&ptap);CHKERRQ(ierr); 601 ierr = PetscNew(&merge);CHKERRQ(ierr); 602 ptap->merge = merge; 603 ptap->reuse = MAT_INITIAL_MATRIX; 604 605 /* get P_oth by taking rows of P (= non-zero cols of local A) from other processors */ 606 ierr = MatGetBrowsOfAoCols_MPIAIJ(A,P,MAT_INITIAL_MATRIX,&ptap->startsj_s,&ptap->startsj_r,&ptap->bufa,&ptap->P_oth);CHKERRQ(ierr); 607 608 /* get P_loc by taking all local rows of P */ 609 ierr = MatMPIAIJGetLocalMat(P,MAT_INITIAL_MATRIX,&ptap->P_loc);CHKERRQ(ierr); 610 611 p_loc = (Mat_SeqAIJ*)(ptap->P_loc)->data; 612 p_oth = (Mat_SeqAIJ*)(ptap->P_oth)->data; 613 pi_loc = p_loc->i; pj_loc = p_loc->j; 614 pi_oth = p_oth->i; pj_oth = p_oth->j; 615 616 /* (1) compute symbolic AP = A_loc*P = A_diag*P_loc + A_off*P_oth (api,apj) */ 617 /*--------------------------------------------------------------------------*/ 618 ierr = PetscMalloc1(am+1,&api);CHKERRQ(ierr); 619 api[0] = 0; 620 621 /* create and initialize a linked list */ 622 ierr = PetscLLCondensedCreate(pN,pN,&lnk,&lnkbt);CHKERRQ(ierr); 623 624 /* Initial FreeSpace size is fill*(nnz(A) + nnz(P)) -OOM for ex56, np=8k on Intrepid! */ 625 ierr = PetscFreeSpaceGet((PetscInt)(fill*(adi[am]+aoi[am]+pi_loc[pm])),&free_space);CHKERRQ(ierr); 626 627 current_space = free_space; 628 629 for (i=0; i<am; i++) { 630 /* diagonal portion of A */ 631 nzi = adi[i+1] - adi[i]; 632 aj = ad->j + adi[i]; 633 for (j=0; j<nzi; j++) { 634 row = aj[j]; 635 pnz = pi_loc[row+1] - pi_loc[row]; 636 Jptr = pj_loc + pi_loc[row]; 637 /* add non-zero cols of P into the sorted linked list lnk */ 638 ierr = PetscLLCondensedAddSorted(pnz,Jptr,lnk,lnkbt);CHKERRQ(ierr); 639 } 640 /* off-diagonal portion of A */ 641 nzi = aoi[i+1] - aoi[i]; 642 aj = ao->j + aoi[i]; 643 for (j=0; j<nzi; j++) { 644 row = aj[j]; 645 pnz = pi_oth[row+1] - pi_oth[row]; 646 Jptr = pj_oth + pi_oth[row]; 647 ierr = PetscLLCondensedAddSorted(pnz,Jptr,lnk,lnkbt);CHKERRQ(ierr); 648 } 649 apnz = lnk[0]; 650 api[i+1] = api[i] + apnz; 651 if (ap_rmax < apnz) ap_rmax = apnz; 652 653 /* if free space is not available, double the total space in the list */ 654 if (current_space->local_remaining<apnz) { 655 ierr = PetscFreeSpaceGet(apnz+current_space->total_array_size,¤t_space);CHKERRQ(ierr); 656 nspacedouble++; 657 } 658 659 /* Copy data into free space, then initialize lnk */ 660 ierr = PetscLLCondensedClean(pN,apnz,current_space->array,lnk,lnkbt);CHKERRQ(ierr); 661 662 current_space->array += apnz; 663 current_space->local_used += apnz; 664 current_space->local_remaining -= apnz; 665 } 666 667 /* Allocate space for apj, initialize apj, and */ 668 /* destroy list of free space and other temporary array(s) */ 669 ierr = PetscMalloc1(api[am]+1,&apj);CHKERRQ(ierr); 670 ierr = PetscFreeSpaceContiguous(&free_space,apj);CHKERRQ(ierr); 671 afill_tmp = (PetscReal)api[am]/(adi[am]+aoi[am]+pi_loc[pm]+1); 672 if (afill_tmp > afill) afill = afill_tmp; 673 674 /* (2) determine symbolic Co=(p->B)^T*AP - send to others (coi,coj)*/ 675 /*-----------------------------------------------------------------*/ 676 ierr = MatGetSymbolicTranspose_SeqAIJ(p->B,&poti,&potj);CHKERRQ(ierr); 677 678 /* then, compute symbolic Co = (p->B)^T*AP */ 679 pon = (p->B)->cmap->n; /* total num of rows to be sent to other processors 680 >= (num of nonzero rows of C_seq) - pn */ 681 ierr = PetscMalloc1(pon+1,&coi);CHKERRQ(ierr); 682 coi[0] = 0; 683 684 /* set initial free space to be fill*(nnz(p->B) + nnz(AP)) */ 685 nnz = fill*(poti[pon] + api[am]); 686 ierr = PetscFreeSpaceGet(nnz,&free_space);CHKERRQ(ierr); 687 current_space = free_space; 688 689 for (i=0; i<pon; i++) { 690 pnz = poti[i+1] - poti[i]; 691 ptJ = potj + poti[i]; 692 for (j=0; j<pnz; j++) { 693 row = ptJ[j]; /* row of AP == col of Pot */ 694 apnz = api[row+1] - api[row]; 695 Jptr = apj + api[row]; 696 /* add non-zero cols of AP into the sorted linked list lnk */ 697 ierr = PetscLLCondensedAddSorted(apnz,Jptr,lnk,lnkbt);CHKERRQ(ierr); 698 } 699 nnz = lnk[0]; 700 701 /* If free space is not available, double the total space in the list */ 702 if (current_space->local_remaining<nnz) { 703 ierr = PetscFreeSpaceGet(nnz+current_space->total_array_size,¤t_space);CHKERRQ(ierr); 704 nspacedouble++; 705 } 706 707 /* Copy data into free space, and zero out denserows */ 708 ierr = PetscLLCondensedClean(pN,nnz,current_space->array,lnk,lnkbt);CHKERRQ(ierr); 709 710 current_space->array += nnz; 711 current_space->local_used += nnz; 712 current_space->local_remaining -= nnz; 713 714 coi[i+1] = coi[i] + nnz; 715 } 716 717 ierr = PetscMalloc1(coi[pon],&coj);CHKERRQ(ierr); 718 ierr = PetscFreeSpaceContiguous(&free_space,coj);CHKERRQ(ierr); 719 afill_tmp = (PetscReal)coi[pon]/(poti[pon] + api[am]+1); 720 if (afill_tmp > afill) afill = afill_tmp; 721 ierr = MatRestoreSymbolicTranspose_SeqAIJ(p->B,&poti,&potj);CHKERRQ(ierr); 722 723 /* (3) send j-array (coj) of Co to other processors */ 724 /*--------------------------------------------------*/ 725 ierr = PetscCalloc1(size,&merge->len_s);CHKERRQ(ierr); 726 len_s = merge->len_s; 727 merge->nsend = 0; 728 729 730 /* determine row ownership */ 731 ierr = PetscLayoutCreate(comm,&merge->rowmap);CHKERRQ(ierr); 732 merge->rowmap->n = pn; 733 merge->rowmap->bs = 1; 734 735 ierr = PetscLayoutSetUp(merge->rowmap);CHKERRQ(ierr); 736 owners = merge->rowmap->range; 737 738 /* determine the number of messages to send, their lengths */ 739 ierr = PetscMalloc2(size,&len_si,size,&sstatus);CHKERRQ(ierr); 740 ierr = PetscMemzero(len_si,size*sizeof(PetscMPIInt));CHKERRQ(ierr); 741 ierr = PetscMalloc1(size+2,&owners_co);CHKERRQ(ierr); 742 743 proc = 0; 744 for (i=0; i<pon; i++) { 745 while (prmap[i] >= owners[proc+1]) proc++; 746 len_si[proc]++; /* num of rows in Co(=Pt*AP) to be sent to [proc] */ 747 len_s[proc] += coi[i+1] - coi[i]; /* num of nonzeros in Co to be sent to [proc] */ 748 } 749 750 len = 0; /* max length of buf_si[], see (4) */ 751 owners_co[0] = 0; 752 for (proc=0; proc<size; proc++) { 753 owners_co[proc+1] = owners_co[proc] + len_si[proc]; 754 if (len_s[proc]) { 755 merge->nsend++; 756 len_si[proc] = 2*(len_si[proc] + 1); /* length of buf_si to be sent to [proc] */ 757 len += len_si[proc]; 758 } 759 } 760 761 /* determine the number and length of messages to receive for coi and coj */ 762 ierr = PetscGatherNumberOfMessages(comm,NULL,len_s,&merge->nrecv);CHKERRQ(ierr); 763 ierr = PetscGatherMessageLengths2(comm,merge->nsend,merge->nrecv,len_s,len_si,&merge->id_r,&merge->len_r,&len_ri);CHKERRQ(ierr); 764 765 /* post the Irecv and Isend of coj */ 766 ierr = PetscCommGetNewTag(comm,&tagj);CHKERRQ(ierr); 767 ierr = PetscPostIrecvInt(comm,tagj,merge->nrecv,merge->id_r,merge->len_r,&buf_rj,&rwaits);CHKERRQ(ierr); 768 ierr = PetscMalloc1(merge->nsend+1,&swaits);CHKERRQ(ierr); 769 for (proc=0, k=0; proc<size; proc++) { 770 if (!len_s[proc]) continue; 771 i = owners_co[proc]; 772 ierr = MPI_Isend(coj+coi[i],len_s[proc],MPIU_INT,proc,tagj,comm,swaits+k);CHKERRQ(ierr); 773 k++; 774 } 775 776 /* receives and sends of coj are complete */ 777 for (i=0; i<merge->nrecv; i++) { 778 ierr = MPI_Waitany(merge->nrecv,rwaits,&icompleted,&rstatus);CHKERRQ(ierr); 779 } 780 ierr = PetscFree(rwaits);CHKERRQ(ierr); 781 if (merge->nsend) {ierr = MPI_Waitall(merge->nsend,swaits,sstatus);CHKERRQ(ierr);} 782 783 /* (4) send and recv coi */ 784 /*-----------------------*/ 785 ierr = PetscCommGetNewTag(comm,&tagi);CHKERRQ(ierr); 786 ierr = PetscPostIrecvInt(comm,tagi,merge->nrecv,merge->id_r,len_ri,&buf_ri,&rwaits);CHKERRQ(ierr); 787 ierr = PetscMalloc1(len+1,&buf_s);CHKERRQ(ierr); 788 buf_si = buf_s; /* points to the beginning of k-th msg to be sent */ 789 for (proc=0,k=0; proc<size; proc++) { 790 if (!len_s[proc]) continue; 791 /* form outgoing message for i-structure: 792 buf_si[0]: nrows to be sent 793 [1:nrows]: row index (global) 794 [nrows+1:2*nrows+1]: i-structure index 795 */ 796 /*-------------------------------------------*/ 797 nrows = len_si[proc]/2 - 1; /* num of rows in Co to be sent to [proc] */ 798 buf_si_i = buf_si + nrows+1; 799 buf_si[0] = nrows; 800 buf_si_i[0] = 0; 801 nrows = 0; 802 for (i=owners_co[proc]; i<owners_co[proc+1]; i++) { 803 nzi = coi[i+1] - coi[i]; 804 buf_si_i[nrows+1] = buf_si_i[nrows] + nzi; /* i-structure */ 805 buf_si[nrows+1] = prmap[i] -owners[proc]; /* local row index */ 806 nrows++; 807 } 808 ierr = MPI_Isend(buf_si,len_si[proc],MPIU_INT,proc,tagi,comm,swaits+k);CHKERRQ(ierr); 809 k++; 810 buf_si += len_si[proc]; 811 } 812 i = merge->nrecv; 813 while (i--) { 814 ierr = MPI_Waitany(merge->nrecv,rwaits,&icompleted,&rstatus);CHKERRQ(ierr); 815 } 816 ierr = PetscFree(rwaits);CHKERRQ(ierr); 817 if (merge->nsend) {ierr = MPI_Waitall(merge->nsend,swaits,sstatus);CHKERRQ(ierr);} 818 819 ierr = PetscFree2(len_si,sstatus);CHKERRQ(ierr); 820 ierr = PetscFree(len_ri);CHKERRQ(ierr); 821 ierr = PetscFree(swaits);CHKERRQ(ierr); 822 ierr = PetscFree(buf_s);CHKERRQ(ierr); 823 824 /* (5) compute the local portion of C (mpi mat) */ 825 /*----------------------------------------------*/ 826 ierr = MatGetSymbolicTranspose_SeqAIJ(p->A,&pdti,&pdtj);CHKERRQ(ierr); 827 828 /* allocate pti array and free space for accumulating nonzero column info */ 829 ierr = PetscMalloc1(pn+1,&pti);CHKERRQ(ierr); 830 pti[0] = 0; 831 832 /* set initial free space to be fill*(nnz(P) + nnz(AP)) */ 833 nnz = fill*(pi_loc[pm] + api[am]); 834 ierr = PetscFreeSpaceGet(nnz,&free_space);CHKERRQ(ierr); 835 current_space = free_space; 836 837 ierr = PetscMalloc3(merge->nrecv,&buf_ri_k,merge->nrecv,&nextrow,merge->nrecv,&nextci);CHKERRQ(ierr); 838 for (k=0; k<merge->nrecv; k++) { 839 buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */ 840 nrows = *buf_ri_k[k]; 841 nextrow[k] = buf_ri_k[k] + 1; /* next row number of k-th recved i-structure */ 842 nextci[k] = buf_ri_k[k] + (nrows + 1); /* poins to the next i-structure of k-th recved i-structure */ 843 } 844 ierr = MatPreallocateInitialize(comm,pn,pn,dnz,onz);CHKERRQ(ierr); 845 rmax = 0; 846 for (i=0; i<pn; i++) { 847 /* add pdt[i,:]*AP into lnk */ 848 pnz = pdti[i+1] - pdti[i]; 849 ptJ = pdtj + pdti[i]; 850 for (j=0; j<pnz; j++) { 851 row = ptJ[j]; /* row of AP == col of Pt */ 852 apnz = api[row+1] - api[row]; 853 Jptr = apj + api[row]; 854 /* add non-zero cols of AP into the sorted linked list lnk */ 855 ierr = PetscLLCondensedAddSorted(apnz,Jptr,lnk,lnkbt);CHKERRQ(ierr); 856 } 857 858 /* add received col data into lnk */ 859 for (k=0; k<merge->nrecv; k++) { /* k-th received message */ 860 if (i == *nextrow[k]) { /* i-th row */ 861 nzi = *(nextci[k]+1) - *nextci[k]; 862 Jptr = buf_rj[k] + *nextci[k]; 863 ierr = PetscLLCondensedAddSorted(nzi,Jptr,lnk,lnkbt);CHKERRQ(ierr); 864 nextrow[k]++; nextci[k]++; 865 } 866 } 867 nnz = lnk[0]; 868 869 /* if free space is not available, make more free space */ 870 if (current_space->local_remaining<nnz) { 871 ierr = PetscFreeSpaceGet(nnz+current_space->total_array_size,¤t_space);CHKERRQ(ierr); 872 nspacedouble++; 873 } 874 /* copy data into free space, then initialize lnk */ 875 ierr = PetscLLCondensedClean(pN,nnz,current_space->array,lnk,lnkbt);CHKERRQ(ierr); 876 ierr = MatPreallocateSet(i+owners[rank],nnz,current_space->array,dnz,onz);CHKERRQ(ierr); 877 878 current_space->array += nnz; 879 current_space->local_used += nnz; 880 current_space->local_remaining -= nnz; 881 882 pti[i+1] = pti[i] + nnz; 883 if (nnz > rmax) rmax = nnz; 884 } 885 ierr = MatRestoreSymbolicTranspose_SeqAIJ(p->A,&pdti,&pdtj);CHKERRQ(ierr); 886 ierr = PetscFree3(buf_ri_k,nextrow,nextci);CHKERRQ(ierr); 887 888 ierr = PetscMalloc1(pti[pn]+1,&ptj);CHKERRQ(ierr); 889 ierr = PetscFreeSpaceContiguous(&free_space,ptj);CHKERRQ(ierr); 890 afill_tmp = (PetscReal)pti[pn]/(pi_loc[pm] + api[am]+1); 891 if (afill_tmp > afill) afill = afill_tmp; 892 ierr = PetscLLDestroy(lnk,lnkbt);CHKERRQ(ierr); 893 894 /* (6) create symbolic parallel matrix Cmpi */ 895 /*------------------------------------------*/ 896 ierr = MatCreate(comm,&Cmpi);CHKERRQ(ierr); 897 ierr = MatSetSizes(Cmpi,pn,pn,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 898 ierr = MatSetBlockSizes(Cmpi,PetscAbs(P->cmap->bs),PetscAbs(P->cmap->bs));CHKERRQ(ierr); 899 ierr = MatSetType(Cmpi,MATMPIAIJ);CHKERRQ(ierr); 900 ierr = MatMPIAIJSetPreallocation(Cmpi,0,dnz,0,onz);CHKERRQ(ierr); 901 ierr = MatPreallocateFinalize(dnz,onz);CHKERRQ(ierr); 902 903 merge->bi = pti; /* Cseq->i */ 904 merge->bj = ptj; /* Cseq->j */ 905 merge->coi = coi; /* Co->i */ 906 merge->coj = coj; /* Co->j */ 907 merge->buf_ri = buf_ri; 908 merge->buf_rj = buf_rj; 909 merge->owners_co = owners_co; 910 merge->destroy = Cmpi->ops->destroy; 911 merge->duplicate = Cmpi->ops->duplicate; 912 913 /* Cmpi is not ready for use - assembly will be done by MatPtAPNumeric() */ 914 Cmpi->assembled = PETSC_FALSE; 915 Cmpi->ops->destroy = MatDestroy_MPIAIJ_PtAP; 916 Cmpi->ops->duplicate = MatDuplicate_MPIAIJ_MatPtAP; 917 918 /* attach the supporting struct to Cmpi for reuse */ 919 c = (Mat_MPIAIJ*)Cmpi->data; 920 c->ptap = ptap; 921 ptap->api = api; 922 ptap->apj = apj; 923 ptap->rmax = ap_rmax; 924 *C = Cmpi; 925 926 /* flag 'scalable' determines which implementations to be used: 927 0: do dense axpy in MatPtAPNumeric() - fast, but requires storage of a nonscalable dense array apa; 928 1: do sparse axpy in MatPtAPNumeric() - might slow, uses a sparse array apa */ 929 /* set default scalable */ 930 ptap->scalable = PETSC_FALSE; //PETSC_TRUE; 931 932 ierr = PetscOptionsGetBool(((PetscObject)Cmpi)->prefix,"-matptap_scalable",&ptap->scalable,NULL);CHKERRQ(ierr); 933 if (!ptap->scalable) { /* Do dense axpy */ 934 ierr = PetscCalloc1(pN,&ptap->apa);CHKERRQ(ierr); 935 } else { 936 ierr = PetscCalloc1(ap_rmax+1,&ptap->apa);CHKERRQ(ierr); 937 } 938 939 #if defined(PETSC_USE_INFO) 940 if (pti[pn] != 0) { 941 ierr = PetscInfo3(Cmpi,"Reallocs %D; Fill ratio: given %g needed %g.\n",nspacedouble,(double)fill,(double)afill);CHKERRQ(ierr); 942 ierr = PetscInfo1(Cmpi,"Use MatPtAP(A,P,MatReuse,%g,&C) for best performance.\n",(double)afill);CHKERRQ(ierr); 943 } else { 944 ierr = PetscInfo(Cmpi,"Empty matrix product\n");CHKERRQ(ierr); 945 } 946 #endif 947 PetscFunctionReturn(0); 948 } 949 950 #undef __FUNCT__ 951 #define __FUNCT__ "MatPtAPNumeric_MPIAIJ_MPIAIJ" 952 PetscErrorCode MatPtAPNumeric_MPIAIJ_MPIAIJ(Mat A,Mat P,Mat C) 953 { 954 PetscErrorCode ierr; 955 Mat_MPIAIJ *a =(Mat_MPIAIJ*)A->data,*p=(Mat_MPIAIJ*)P->data,*c=(Mat_MPIAIJ*)C->data; 956 Mat_SeqAIJ *ad=(Mat_SeqAIJ*)(a->A)->data,*ao=(Mat_SeqAIJ*)(a->B)->data; 957 Mat_SeqAIJ *pd=(Mat_SeqAIJ*)(p->A)->data,*po=(Mat_SeqAIJ*)(p->B)->data; 958 Mat_SeqAIJ *p_loc,*p_oth; 959 Mat_PtAPMPI *ptap; 960 Mat_Merge_SeqsToMPI *merge; 961 PetscInt *adi=ad->i,*aoi=ao->i,*adj,*aoj,*apJ,nextp; 962 PetscInt *pi_loc,*pj_loc,*pi_oth,*pj_oth,*pJ,*pj; 963 PetscInt i,j,k,anz,pnz,apnz,nextap,row,*cj; 964 MatScalar *ada,*aoa,*apa,*pa,*ca,*pa_loc,*pa_oth,valtmp; 965 PetscInt am =A->rmap->n,cm=C->rmap->n,pon=(p->B)->cmap->n; 966 MPI_Comm comm; 967 PetscMPIInt size,rank,taga,*len_s; 968 PetscInt *owners,proc,nrows,**buf_ri_k,**nextrow,**nextci; 969 PetscInt **buf_ri,**buf_rj; 970 PetscInt cnz=0,*bj_i,*bi,*bj,bnz,nextcj; /* bi,bj,ba: local array of C(mpi mat) */ 971 MPI_Request *s_waits,*r_waits; 972 MPI_Status *status; 973 MatScalar **abuf_r,*ba_i,*pA,*coa,*ba; 974 PetscInt *api,*apj,*coi,*coj; 975 PetscInt *poJ=po->j,*pdJ=pd->j,pcstart=P->cmap->rstart,pcend=P->cmap->rend; 976 PetscBool scalable; 977 #if defined(PTAP_PROFILE) 978 PetscLogDouble t0,t1,t2,eP,t3,t4,et2_AP=0.0,ePtAP=0.0,t2_0,t2_1,t2_2; 979 #endif 980 981 PetscFunctionBegin; 982 ierr = PetscObjectGetComm((PetscObject)C,&comm);CHKERRQ(ierr); 983 #if defined(PTAP_PROFILE) 984 ierr = PetscTime(&t0);CHKERRQ(ierr); 985 #endif 986 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 987 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 988 989 ptap = c->ptap; 990 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"); 991 merge = ptap->merge; 992 apa = ptap->apa; 993 scalable = ptap->scalable; 994 995 /* 1) get P_oth = ptap->P_oth and P_loc = ptap->P_loc */ 996 /*-----------------------------------------------------*/ 997 if (ptap->reuse == MAT_INITIAL_MATRIX) { 998 /* P_oth and P_loc are obtained in MatPtASymbolic(), skip calling MatGetBrowsOfAoCols() and MatMPIAIJGetLocalMat() */ 999 ptap->reuse = MAT_REUSE_MATRIX; 1000 } else { /* update numerical values of P_oth and P_loc */ 1001 ierr = MatGetBrowsOfAoCols_MPIAIJ(A,P,MAT_REUSE_MATRIX,&ptap->startsj_s,&ptap->startsj_r,&ptap->bufa,&ptap->P_oth);CHKERRQ(ierr); 1002 ierr = MatMPIAIJGetLocalMat(P,MAT_REUSE_MATRIX,&ptap->P_loc);CHKERRQ(ierr); 1003 } 1004 #if defined(PTAP_PROFILE) 1005 ierr = PetscTime(&t1);CHKERRQ(ierr); 1006 eP = t1-t0; 1007 #endif 1008 /* 1009 printf("[%d] Ad: %d, %d; Ao: %d, %d; P_loc: %d, %d; P_oth %d, %d;\n",rank, 1010 a->A->rmap->N,a->A->cmap->N,a->B->rmap->N,a->B->cmap->N, 1011 ptap->P_loc->rmap->N,ptap->P_loc->cmap->N, 1012 ptap->P_oth->rmap->N,ptap->P_oth->cmap->N); 1013 */ 1014 1015 /* 2) compute numeric C_seq = P_loc^T*A_loc*P - dominating part */ 1016 /*--------------------------------------------------------------*/ 1017 /* get data from symbolic products */ 1018 p_loc = (Mat_SeqAIJ*)(ptap->P_loc)->data; 1019 p_oth = (Mat_SeqAIJ*)(ptap->P_oth)->data; 1020 pi_loc=p_loc->i; pj_loc=p_loc->j; pJ=pj_loc; pa_loc=p_loc->a; 1021 pi_oth=p_oth->i; pj_oth=p_oth->j; pa_oth=p_oth->a; 1022 1023 coi = merge->coi; coj = merge->coj; 1024 ierr = PetscCalloc1(coi[pon]+1,&coa);CHKERRQ(ierr); 1025 1026 bi = merge->bi; bj = merge->bj; 1027 owners = merge->rowmap->range; 1028 ierr = PetscCalloc1(bi[cm]+1,&ba);CHKERRQ(ierr); /* ba: Cseq->a */ 1029 1030 api = ptap->api; apj = ptap->apj; 1031 1032 if (!scalable) { /* Do dense axpy on apa (length of pN, stores A[i,:]*P) - nonscalable, but faster (could take 1/3 scalable time) */ 1033 ierr = PetscInfo(C,"Using non-scalable dense axpy\n");CHKERRQ(ierr); 1034 #if 0 1035 /* ------ 10x slower -------------- */ 1036 /*==================================*/ 1037 Mat R = ptap->R; 1038 Mat_SeqAIJ *r = (Mat_SeqAIJ*)R->data; 1039 PetscInt *ri=r->i,*rj=r->j,rnz,arow,l,prow,pcol,pN=P->cmap->N; 1040 PetscScalar *ra=r->a,tmp,cdense[pN]; 1041 1042 ierr = PetscMemzero(cdense,pN*sizeof(PetscScalar));CHKERRQ(ierr); 1043 for (i=0; i<cm; i++) { /* each row of C or R */ 1044 rnz = ri[i+1] - ri[i]; 1045 1046 for (j=0; j<rnz; j++) { /* each nz of R */ 1047 arow = rj[ri[i] + j]; 1048 1049 /* diagonal portion of A */ 1050 anz = ad->i[arow+1] - ad->i[arow]; 1051 for (k=0; k<anz; k++) { /* each nz of Ad */ 1052 tmp = ra[ri[i] + j]*ad->a[ad->i[arow] + k]; 1053 prow = ad->j[ad->i[arow] + k]; 1054 pnz = pi_loc[prow+1] - pi_loc[prow]; 1055 1056 for (l=0; l<pnz; l++) { /* each nz of P_loc */ 1057 pcol = pj_loc[pi_loc[prow] + l]; 1058 cdense[pcol] += tmp*pa_loc[pi_loc[prow] + l]; 1059 } 1060 } 1061 1062 /* off-diagonal portion of A */ 1063 anz = ao->i[arow+1] - ao->i[arow]; 1064 for (k=0; k<anz; k++) { /* each nz of Ao */ 1065 tmp = ra[ri[i] + j]*ao->a[ao->i[arow] + k]; 1066 prow = ao->j[ao->i[arow] + k]; 1067 pnz = pi_oth[prow+1] - pi_oth[prow]; 1068 1069 for (l=0; l<pnz; l++) { /* each nz of P_oth */ 1070 pcol = pj_oth[pi_oth[prow] + l]; 1071 cdense[pcol] += tmp*pa_oth[pi_oth[prow] + l]; 1072 } 1073 } 1074 1075 } //for (j=0; j<rnz; j++) 1076 1077 /* copy cdense[] into ca; zero cdense[] */ 1078 cnz = bi[i+1] - bi[i]; 1079 cj = bj + bi[i]; 1080 ca = ba + bi[i]; 1081 for (j=0; j<cnz; j++) { 1082 ca[j] += cdense[cj[j]]; 1083 cdense[cj[j]] = 0.0; 1084 } 1085 #if 0 1086 if (rank == 0) { 1087 printf("[%d] row %d: ",rank,i); 1088 for (j=0; j<pN; j++) printf(" %g,",cdense[j]); 1089 printf("\n"); 1090 } 1091 for (j=0; j<pN; j++) cdense[j]=0.0; // zero cdnese[] 1092 #endif 1093 } //for (i=0; i<cm; i++) { 1094 #endif 1095 1096 //========================================== 1097 1098 ierr = PetscTime(&t1);CHKERRQ(ierr); 1099 for (i=0; i<am; i++) { 1100 #if defined(PTAP_PROFILE) 1101 ierr = PetscTime(&t2_0);CHKERRQ(ierr); 1102 #endif 1103 /* 2-a) form i-th sparse row of A_loc*P = Ad*P_loc + Ao*P_oth */ 1104 /*------------------------------------------------------------*/ 1105 apJ = apj + api[i]; 1106 1107 /* diagonal portion of A */ 1108 anz = adi[i+1] - adi[i]; 1109 adj = ad->j + adi[i]; 1110 ada = ad->a + adi[i]; 1111 for (j=0; j<anz; j++) { 1112 row = adj[j]; 1113 pnz = pi_loc[row+1] - pi_loc[row]; 1114 pj = pj_loc + pi_loc[row]; 1115 pa = pa_loc + pi_loc[row]; 1116 1117 /* perform dense axpy */ 1118 valtmp = ada[j]; 1119 for (k=0; k<pnz; k++) { 1120 apa[pj[k]] += valtmp*pa[k]; 1121 } 1122 ierr = PetscLogFlops(2.0*pnz);CHKERRQ(ierr); 1123 } 1124 1125 /* off-diagonal portion of A */ 1126 anz = aoi[i+1] - aoi[i]; 1127 aoj = ao->j + aoi[i]; 1128 aoa = ao->a + aoi[i]; 1129 for (j=0; j<anz; j++) { 1130 row = aoj[j]; 1131 pnz = pi_oth[row+1] - pi_oth[row]; 1132 pj = pj_oth + pi_oth[row]; 1133 pa = pa_oth + pi_oth[row]; 1134 1135 /* perform dense axpy */ 1136 valtmp = aoa[j]; 1137 for (k=0; k<pnz; k++) { 1138 apa[pj[k]] += valtmp*pa[k]; 1139 } 1140 ierr = PetscLogFlops(2.0*pnz);CHKERRQ(ierr); 1141 } 1142 #if defined(PTAP_PROFILE) 1143 ierr = PetscTime(&t2_1);CHKERRQ(ierr); 1144 et2_AP += t2_1 - t2_0; 1145 #endif 1146 1147 /* 2-b) Compute Cseq = P_loc[i,:]^T*AP[i,:] using outer product */ 1148 /*--------------------------------------------------------------*/ 1149 apnz = api[i+1] - api[i]; 1150 /* put the value into Co=(p->B)^T*AP (off-diagonal part, send to others) */ 1151 pnz = po->i[i+1] - po->i[i]; 1152 poJ = po->j + po->i[i]; 1153 pA = po->a + po->i[i]; 1154 for (j=0; j<pnz; j++) { 1155 row = poJ[j]; 1156 cnz = coi[row+1] - coi[row]; 1157 cj = coj + coi[row]; 1158 ca = coa + coi[row]; 1159 /* perform dense axpy */ 1160 valtmp = pA[j]; 1161 for (k=0; k<cnz; k++) { 1162 ca[k] += valtmp*apa[cj[k]]; 1163 } 1164 ierr = PetscLogFlops(2.0*cnz);CHKERRQ(ierr); 1165 } 1166 #if 1 1167 /* put the value into Cd (diagonal part) */ 1168 pnz = pd->i[i+1] - pd->i[i]; 1169 pdJ = pd->j + pd->i[i]; 1170 pA = pd->a + pd->i[i]; 1171 for (j=0; j<pnz; j++) { 1172 row = pdJ[j]; 1173 cnz = bi[row+1] - bi[row]; 1174 cj = bj + bi[row]; 1175 ca = ba + bi[row]; 1176 /* perform dense axpy */ 1177 valtmp = pA[j]; 1178 for (k=0; k<cnz; k++) { 1179 ca[k] += valtmp*apa[cj[k]]; 1180 } 1181 ierr = PetscLogFlops(2.0*cnz);CHKERRQ(ierr); 1182 } 1183 #endif 1184 /* zero the current row of A*P */ 1185 for (k=0; k<apnz; k++) apa[apJ[k]] = 0.0; 1186 #if defined(PTAP_PROFILE) 1187 ierr = PetscTime(&t2_2);CHKERRQ(ierr); 1188 ePtAP += t2_2 - t2_1; 1189 #endif 1190 } 1191 1192 if (rank == 100) { 1193 for (row=0; row<cm; row++) { 1194 printf("[%d] row %d: ",rank,row); 1195 cnz = bi[row+1] - bi[row]; 1196 for (j=0; j<cnz; j++) printf(" %g,",ba[bi[row]+j]); 1197 printf("\n"); 1198 } 1199 } 1200 1201 } else { /* Do sparse axpy on apa (length of ap_rmax, stores A[i,:]*P) - scalable, but slower */ 1202 ierr = PetscInfo(C,"Using scalable sparse axpy\n");CHKERRQ(ierr); 1203 /*-----------------------------------------------------------------------------------------*/ 1204 pA=pa_loc; 1205 for (i=0; i<am; i++) { 1206 #if defined(PTAP_PROFILE) 1207 ierr = PetscTime(&t2_0);CHKERRQ(ierr); 1208 #endif 1209 /* form i-th sparse row of A*P */ 1210 apnz = api[i+1] - api[i]; 1211 apJ = apj + api[i]; 1212 /* diagonal portion of A */ 1213 anz = adi[i+1] - adi[i]; 1214 adj = ad->j + adi[i]; 1215 ada = ad->a + adi[i]; 1216 for (j=0; j<anz; j++) { 1217 row = adj[j]; 1218 pnz = pi_loc[row+1] - pi_loc[row]; 1219 pj = pj_loc + pi_loc[row]; 1220 pa = pa_loc + pi_loc[row]; 1221 valtmp = ada[j]; 1222 nextp = 0; 1223 for (k=0; nextp<pnz; k++) { 1224 if (apJ[k] == pj[nextp]) { /* col of AP == col of P */ 1225 apa[k] += valtmp*pa[nextp++]; 1226 } 1227 } 1228 ierr = PetscLogFlops(2.0*pnz);CHKERRQ(ierr); 1229 } 1230 /* off-diagonal portion of A */ 1231 anz = aoi[i+1] - aoi[i]; 1232 aoj = ao->j + aoi[i]; 1233 aoa = ao->a + aoi[i]; 1234 for (j=0; j<anz; j++) { 1235 row = aoj[j]; 1236 pnz = pi_oth[row+1] - pi_oth[row]; 1237 pj = pj_oth + pi_oth[row]; 1238 pa = pa_oth + pi_oth[row]; 1239 valtmp = aoa[j]; 1240 nextp = 0; 1241 for (k=0; nextp<pnz; k++) { 1242 if (apJ[k] == pj[nextp]) { /* col of AP == col of P */ 1243 apa[k] += valtmp*pa[nextp++]; 1244 } 1245 } 1246 ierr = PetscLogFlops(2.0*pnz);CHKERRQ(ierr); 1247 } 1248 #if defined(PTAP_PROFILE) 1249 ierr = PetscTime(&t2_1);CHKERRQ(ierr); 1250 et2_AP += t2_1 - t2_0; 1251 #endif 1252 1253 /* 2-b) Compute Cseq = P_loc[i,:]^T*AP[i,:] using outer product */ 1254 /*--------------------------------------------------------------*/ 1255 pnz = pi_loc[i+1] - pi_loc[i]; 1256 pJ = pj_loc + pi_loc[i]; 1257 for (j=0; j<pnz; j++) { 1258 nextap = 0; 1259 row = pJ[j]; /* global index */ 1260 if (row < pcstart || row >=pcend) { /* put the value into Co */ 1261 row = *poJ; 1262 cj = coj + coi[row]; 1263 ca = coa + coi[row]; poJ++; 1264 } else { /* put the value into Cd */ 1265 row = *pdJ; 1266 cj = bj + bi[row]; 1267 ca = ba + bi[row]; pdJ++; 1268 } 1269 valtmp = pA[j]; 1270 for (k=0; nextap<apnz; k++) { 1271 if (cj[k]==apJ[nextap]) ca[k] += valtmp*apa[nextap++]; 1272 } 1273 ierr = PetscLogFlops(2.0*apnz);CHKERRQ(ierr); 1274 } 1275 pA += pnz; 1276 /* zero the current row info for A*P */ 1277 ierr = PetscMemzero(apa,apnz*sizeof(MatScalar));CHKERRQ(ierr); 1278 #if defined(PTAP_PROFILE) 1279 ierr = PetscTime(&t2_2);CHKERRQ(ierr); 1280 ePtAP += t2_2 - t2_1; 1281 #endif 1282 } 1283 } 1284 #if defined(PTAP_PROFILE) 1285 ierr = PetscTime(&t2);CHKERRQ(ierr); 1286 #endif 1287 1288 /* 3) send and recv matrix values coa */ 1289 /*------------------------------------*/ 1290 buf_ri = merge->buf_ri; 1291 buf_rj = merge->buf_rj; 1292 len_s = merge->len_s; 1293 ierr = PetscCommGetNewTag(comm,&taga);CHKERRQ(ierr); 1294 ierr = PetscPostIrecvScalar(comm,taga,merge->nrecv,merge->id_r,merge->len_r,&abuf_r,&r_waits);CHKERRQ(ierr); 1295 1296 ierr = PetscMalloc2(merge->nsend+1,&s_waits,size,&status);CHKERRQ(ierr); 1297 for (proc=0,k=0; proc<size; proc++) { 1298 if (!len_s[proc]) continue; 1299 i = merge->owners_co[proc]; 1300 ierr = MPI_Isend(coa+coi[i],len_s[proc],MPIU_MATSCALAR,proc,taga,comm,s_waits+k);CHKERRQ(ierr); 1301 k++; 1302 } 1303 if (merge->nrecv) {ierr = MPI_Waitall(merge->nrecv,r_waits,status);CHKERRQ(ierr);} 1304 if (merge->nsend) {ierr = MPI_Waitall(merge->nsend,s_waits,status);CHKERRQ(ierr);} 1305 1306 ierr = PetscFree2(s_waits,status);CHKERRQ(ierr); 1307 ierr = PetscFree(r_waits);CHKERRQ(ierr); 1308 ierr = PetscFree(coa);CHKERRQ(ierr); 1309 #if defined(PTAP_PROFILE) 1310 ierr = PetscTime(&t3);CHKERRQ(ierr); 1311 #endif 1312 1313 /* 4) insert local Cseq and received values into Cmpi */ 1314 /*------------------------------------------------------*/ 1315 ierr = PetscMalloc3(merge->nrecv,&buf_ri_k,merge->nrecv,&nextrow,merge->nrecv,&nextci);CHKERRQ(ierr); 1316 for (k=0; k<merge->nrecv; k++) { 1317 buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */ 1318 nrows = *(buf_ri_k[k]); 1319 nextrow[k] = buf_ri_k[k]+1; /* next row number of k-th recved i-structure */ 1320 nextci[k] = buf_ri_k[k] + (nrows + 1); /* poins to the next i-structure of k-th recved i-structure */ 1321 } 1322 1323 for (i=0; i<cm; i++) { 1324 row = owners[rank] + i; /* global row index of C_seq */ 1325 bj_i = bj + bi[i]; /* col indices of the i-th row of C */ 1326 ba_i = ba + bi[i]; 1327 bnz = bi[i+1] - bi[i]; 1328 /* add received vals into ba */ 1329 for (k=0; k<merge->nrecv; k++) { /* k-th received message */ 1330 /* i-th row */ 1331 if (i == *nextrow[k]) { 1332 cnz = *(nextci[k]+1) - *nextci[k]; 1333 cj = buf_rj[k] + *(nextci[k]); 1334 ca = abuf_r[k] + *(nextci[k]); 1335 nextcj = 0; 1336 for (j=0; nextcj<cnz; j++) { 1337 if (bj_i[j] == cj[nextcj]) { /* bcol == ccol */ 1338 ba_i[j] += ca[nextcj++]; 1339 } 1340 } 1341 nextrow[k]++; nextci[k]++; 1342 ierr = PetscLogFlops(2.0*cnz);CHKERRQ(ierr); 1343 } 1344 } 1345 ierr = MatSetValues(C,1,&row,bnz,bj_i,ba_i,INSERT_VALUES);CHKERRQ(ierr); 1346 } 1347 ierr = MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1348 ierr = MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1349 1350 ierr = PetscFree(ba);CHKERRQ(ierr); 1351 ierr = PetscFree(abuf_r[0]);CHKERRQ(ierr); 1352 ierr = PetscFree(abuf_r);CHKERRQ(ierr); 1353 ierr = PetscFree3(buf_ri_k,nextrow,nextci);CHKERRQ(ierr); 1354 #if defined(PTAP_PROFILE) 1355 ierr = PetscTime(&t4);CHKERRQ(ierr); 1356 if (rank==1) { 1357 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); 1358 } 1359 #endif 1360 PetscFunctionReturn(0); 1361 } 1362